blob: 47f56e2cc8041821458c658dea6c24103efe15a0 [file] [log] [blame]
Thomas Gleixner457c8992019-05-19 13:08:55 +01001// SPDX-License-Identifier: GPL-2.0-only
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
3 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01004 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006#include <linux/list.h>
7#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Mike Rapoport97ad1082018-10-30 15:09:44 -070019#include <linux/memblock.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Joonsoo Kimbbe88752020-08-11 18:37:38 -070022#include <linux/sched/mm.h>
Mike Kravetz63489f82018-03-22 16:17:13 -070023#include <linux/mmdebug.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010024#include <linux/sched/signal.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090025#include <linux/rmap.h>
Matthew Wilcoxc6247f72017-07-10 15:48:56 -070026#include <linux/string_helpers.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090027#include <linux/swap.h>
28#include <linux/swapops.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070029#include <linux/jhash.h>
Anshuman Khandual98fa15f2019-03-05 15:42:58 -080030#include <linux/numa.h>
Waiman Longc77c0a82020-01-04 13:00:15 -080031#include <linux/llist.h>
Roman Gushchincf11e852020-04-10 14:32:45 -070032#include <linux/cma.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070033
David Gibson63551ae2005-06-21 17:14:44 -070034#include <asm/page.h>
Mike Rapoportca15ca42020-08-06 23:22:28 -070035#include <asm/pgalloc.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070036#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070037
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070038#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070039#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070040#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080041#include <linux/node.h>
Mike Kravetz1a1aad82017-02-22 15:43:01 -080042#include <linux/userfaultfd_k.h>
Michal Hockoab5ac902018-01-31 16:20:48 -080043#include <linux/page_owner.h>
Nick Piggin7835e982006-03-22 00:08:40 -080044#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070046int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070047unsigned int default_hstate_idx;
48struct hstate hstates[HUGE_MAX_HSTATE];
Roman Gushchincf11e852020-04-10 14:32:45 -070049
Barry Songdbda8fe2020-07-23 21:15:30 -070050#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -070051static struct cma *hugetlb_cma[MAX_NUMNODES];
Barry Songdbda8fe2020-07-23 21:15:30 -070052#endif
53static unsigned long hugetlb_cma_size __initdata;
Roman Gushchincf11e852020-04-10 14:32:45 -070054
Naoya Horiguchi641844f2015-06-24 16:56:59 -070055/*
56 * Minimum page order among possible hugepage sizes, set to a proper value
57 * at boot time.
58 */
59static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070060
Jon Tollefson53ba51d2008-07-23 21:27:52 -070061__initdata LIST_HEAD(huge_boot_pages);
62
Andi Kleene5ff2152008-07-23 21:27:42 -070063/* for command line parsing */
64static struct hstate * __initdata parsed_hstate;
65static unsigned long __initdata default_hstate_max_huge_pages;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -070066static bool __initdata parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -070067static bool __initdata parsed_default_hugepagesz;
Andi Kleene5ff2152008-07-23 21:27:42 -070068
David Gibson3935baa2006-03-22 00:08:53 -080069/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070070 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080072 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070073DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080074
Davidlohr Bueso8382d912014-04-03 14:47:31 -070075/*
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
78 */
79static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070080struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070081
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070082/* Forward declaration */
83static int hugetlb_acct_memory(struct hstate *h, long delta);
84
Miaohe Lin1d884332021-02-24 12:07:05 -080085static inline bool subpool_is_free(struct hugepage_subpool *spool)
86{
87 if (spool->count)
88 return false;
89 if (spool->max_hpages != -1)
90 return spool->used_hpages == 0;
91 if (spool->min_hpages != -1)
92 return spool->rsv_hpages == spool->min_hpages;
93
94 return true;
95}
96
Mike Kravetzdb71ef72021-05-04 18:35:07 -070097static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
98 unsigned long irq_flags)
David Gibson90481622012-03-21 16:34:12 -070099{
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700100 spin_unlock_irqrestore(&spool->lock, irq_flags);
David Gibson90481622012-03-21 16:34:12 -0700101
102 /* If no pages are used, and no other handles to the subpool
Ethon Paul7c8de352020-06-04 16:49:07 -0700103 * remain, give up any reservations based on minimum size and
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700104 * free the subpool */
Miaohe Lin1d884332021-02-24 12:07:05 -0800105 if (subpool_is_free(spool)) {
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700106 if (spool->min_hpages != -1)
107 hugetlb_acct_memory(spool->hstate,
108 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -0700109 kfree(spool);
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700110 }
David Gibson90481622012-03-21 16:34:12 -0700111}
112
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700113struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
114 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -0700115{
116 struct hugepage_subpool *spool;
117
Mike Kravetzc6a91822015-04-15 16:13:36 -0700118 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -0700119 if (!spool)
120 return NULL;
121
122 spin_lock_init(&spool->lock);
123 spool->count = 1;
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700124 spool->max_hpages = max_hpages;
125 spool->hstate = h;
126 spool->min_hpages = min_hpages;
127
128 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
129 kfree(spool);
130 return NULL;
131 }
132 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700133
134 return spool;
135}
136
137void hugepage_put_subpool(struct hugepage_subpool *spool)
138{
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700139 unsigned long flags;
140
141 spin_lock_irqsave(&spool->lock, flags);
David Gibson90481622012-03-21 16:34:12 -0700142 BUG_ON(!spool->count);
143 spool->count--;
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700144 unlock_or_release_subpool(spool, flags);
David Gibson90481622012-03-21 16:34:12 -0700145}
146
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700147/*
148 * Subpool accounting for allocating and reserving pages.
149 * Return -ENOMEM if there are not enough resources to satisfy the
Randy Dunlap9e7ee402020-08-11 18:32:59 -0700150 * request. Otherwise, return the number of pages by which the
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700151 * global pools must be adjusted (upward). The returned value may
152 * only be different than the passed value (delta) in the case where
Ethon Paul7c8de352020-06-04 16:49:07 -0700153 * a subpool minimum size must be maintained.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700154 */
155static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700156 long delta)
157{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700158 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700159
160 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700161 return ret;
David Gibson90481622012-03-21 16:34:12 -0700162
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700163 spin_lock_irq(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700164
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700165 if (spool->max_hpages != -1) { /* maximum size accounting */
166 if ((spool->used_hpages + delta) <= spool->max_hpages)
167 spool->used_hpages += delta;
168 else {
169 ret = -ENOMEM;
170 goto unlock_ret;
171 }
172 }
173
Mike Kravetz09a95e22016-05-19 17:11:01 -0700174 /* minimum size accounting */
175 if (spool->min_hpages != -1 && spool->rsv_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700176 if (delta > spool->rsv_hpages) {
177 /*
178 * Asking for more reserves than those already taken on
179 * behalf of subpool. Return difference.
180 */
181 ret = delta - spool->rsv_hpages;
182 spool->rsv_hpages = 0;
183 } else {
184 ret = 0; /* reserves already accounted for */
185 spool->rsv_hpages -= delta;
186 }
187 }
188
189unlock_ret:
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700190 spin_unlock_irq(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700191 return ret;
192}
193
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700194/*
195 * Subpool accounting for freeing and unreserving pages.
196 * Return the number of global page reservations that must be dropped.
197 * The return value may only be different than the passed value (delta)
198 * in the case where a subpool minimum size must be maintained.
199 */
200static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700201 long delta)
202{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700203 long ret = delta;
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700204 unsigned long flags;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700205
David Gibson90481622012-03-21 16:34:12 -0700206 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700207 return delta;
David Gibson90481622012-03-21 16:34:12 -0700208
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700209 spin_lock_irqsave(&spool->lock, flags);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700210
211 if (spool->max_hpages != -1) /* maximum size accounting */
212 spool->used_hpages -= delta;
213
Mike Kravetz09a95e22016-05-19 17:11:01 -0700214 /* minimum size accounting */
215 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700216 if (spool->rsv_hpages + delta <= spool->min_hpages)
217 ret = 0;
218 else
219 ret = spool->rsv_hpages + delta - spool->min_hpages;
220
221 spool->rsv_hpages += delta;
222 if (spool->rsv_hpages > spool->min_hpages)
223 spool->rsv_hpages = spool->min_hpages;
224 }
225
226 /*
227 * If hugetlbfs_put_super couldn't free spool due to an outstanding
228 * quota reference, free it now.
229 */
Mike Kravetzdb71ef72021-05-04 18:35:07 -0700230 unlock_or_release_subpool(spool, flags);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700231
232 return ret;
David Gibson90481622012-03-21 16:34:12 -0700233}
234
235static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
236{
237 return HUGETLBFS_SB(inode->i_sb)->spool;
238}
239
240static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
241{
Al Viro496ad9a2013-01-23 17:07:38 -0500242 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700243}
244
Mina Almasry0db9d742020-04-01 21:11:25 -0700245/* Helper that removes a struct file_region from the resv_map cache and returns
246 * it for use.
247 */
248static struct file_region *
249get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
250{
251 struct file_region *nrg = NULL;
252
253 VM_BUG_ON(resv->region_cache_count <= 0);
254
255 resv->region_cache_count--;
256 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
Mina Almasry0db9d742020-04-01 21:11:25 -0700257 list_del(&nrg->link);
258
259 nrg->from = from;
260 nrg->to = to;
261
262 return nrg;
263}
264
Mina Almasry075a61d2020-04-01 21:11:28 -0700265static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
266 struct file_region *rg)
267{
268#ifdef CONFIG_CGROUP_HUGETLB
269 nrg->reservation_counter = rg->reservation_counter;
270 nrg->css = rg->css;
271 if (rg->css)
272 css_get(rg->css);
273#endif
274}
275
276/* Helper that records hugetlb_cgroup uncharge info. */
277static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
278 struct hstate *h,
279 struct resv_map *resv,
280 struct file_region *nrg)
281{
282#ifdef CONFIG_CGROUP_HUGETLB
283 if (h_cg) {
284 nrg->reservation_counter =
285 &h_cg->rsvd_hugepage[hstate_index(h)];
286 nrg->css = &h_cg->css;
Miaohe Lind85aecf2021-03-24 21:37:17 -0700287 /*
288 * The caller will hold exactly one h_cg->css reference for the
289 * whole contiguous reservation region. But this area might be
290 * scattered when there are already some file_regions reside in
291 * it. As a result, many file_regions may share only one css
292 * reference. In order to ensure that one file_region must hold
293 * exactly one h_cg->css reference, we should do css_get for
294 * each file_region and leave the reference held by caller
295 * untouched.
296 */
297 css_get(&h_cg->css);
Mina Almasry075a61d2020-04-01 21:11:28 -0700298 if (!resv->pages_per_hpage)
299 resv->pages_per_hpage = pages_per_huge_page(h);
300 /* pages_per_hpage should be the same for all entries in
301 * a resv_map.
302 */
303 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
304 } else {
305 nrg->reservation_counter = NULL;
306 nrg->css = NULL;
307 }
308#endif
309}
310
Miaohe Lind85aecf2021-03-24 21:37:17 -0700311static void put_uncharge_info(struct file_region *rg)
312{
313#ifdef CONFIG_CGROUP_HUGETLB
314 if (rg->css)
315 css_put(rg->css);
316#endif
317}
318
Mina Almasrya9b3f862020-04-01 21:11:35 -0700319static bool has_same_uncharge_info(struct file_region *rg,
320 struct file_region *org)
321{
322#ifdef CONFIG_CGROUP_HUGETLB
323 return rg && org &&
324 rg->reservation_counter == org->reservation_counter &&
325 rg->css == org->css;
326
327#else
328 return true;
329#endif
330}
331
332static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
333{
334 struct file_region *nrg = NULL, *prg = NULL;
335
336 prg = list_prev_entry(rg, link);
337 if (&prg->link != &resv->regions && prg->to == rg->from &&
338 has_same_uncharge_info(prg, rg)) {
339 prg->to = rg->to;
340
341 list_del(&rg->link);
Miaohe Lind85aecf2021-03-24 21:37:17 -0700342 put_uncharge_info(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700343 kfree(rg);
344
Wei Yang7db5e7b2020-10-13 16:56:20 -0700345 rg = prg;
Mina Almasrya9b3f862020-04-01 21:11:35 -0700346 }
347
348 nrg = list_next_entry(rg, link);
349 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
350 has_same_uncharge_info(nrg, rg)) {
351 nrg->from = rg->from;
352
353 list_del(&rg->link);
Miaohe Lind85aecf2021-03-24 21:37:17 -0700354 put_uncharge_info(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700355 kfree(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700356 }
357}
358
Peter Xu2103cf92021-03-12 21:07:18 -0800359static inline long
360hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from,
361 long to, struct hstate *h, struct hugetlb_cgroup *cg,
362 long *regions_needed)
363{
364 struct file_region *nrg;
365
366 if (!regions_needed) {
367 nrg = get_file_region_entry_from_cache(map, from, to);
368 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
369 list_add(&nrg->link, rg->link.prev);
370 coalesce_file_region(map, nrg);
371 } else
372 *regions_needed += 1;
373
374 return to - from;
375}
376
Wei Yang972a3da32020-10-13 16:56:30 -0700377/*
378 * Must be called with resv->lock held.
379 *
380 * Calling this with regions_needed != NULL will count the number of pages
381 * to be added but will not modify the linked list. And regions_needed will
382 * indicate the number of file_regions needed in the cache to carry out to add
383 * the regions for this range.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800384 */
385static long add_reservation_in_range(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700386 struct hugetlb_cgroup *h_cg,
Wei Yang972a3da32020-10-13 16:56:30 -0700387 struct hstate *h, long *regions_needed)
Mina Almasryd75c6af2019-11-30 17:56:59 -0800388{
Mina Almasry0db9d742020-04-01 21:11:25 -0700389 long add = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800390 struct list_head *head = &resv->regions;
Mina Almasry0db9d742020-04-01 21:11:25 -0700391 long last_accounted_offset = f;
Peter Xu2103cf92021-03-12 21:07:18 -0800392 struct file_region *rg = NULL, *trg = NULL;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800393
Mina Almasry0db9d742020-04-01 21:11:25 -0700394 if (regions_needed)
395 *regions_needed = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800396
Mina Almasry0db9d742020-04-01 21:11:25 -0700397 /* In this loop, we essentially handle an entry for the range
398 * [last_accounted_offset, rg->from), at every iteration, with some
399 * bounds checking.
400 */
401 list_for_each_entry_safe(rg, trg, head, link) {
402 /* Skip irrelevant regions that start before our range. */
403 if (rg->from < f) {
404 /* If this region ends after the last accounted offset,
405 * then we need to update last_accounted_offset.
406 */
407 if (rg->to > last_accounted_offset)
408 last_accounted_offset = rg->to;
409 continue;
410 }
Mina Almasryd75c6af2019-11-30 17:56:59 -0800411
Mina Almasry0db9d742020-04-01 21:11:25 -0700412 /* When we find a region that starts beyond our range, we've
413 * finished.
414 */
Peter Xuca7e0452021-03-12 21:07:22 -0800415 if (rg->from >= t)
Mina Almasryd75c6af2019-11-30 17:56:59 -0800416 break;
417
Mina Almasry0db9d742020-04-01 21:11:25 -0700418 /* Add an entry for last_accounted_offset -> rg->from, and
419 * update last_accounted_offset.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800420 */
Peter Xu2103cf92021-03-12 21:07:18 -0800421 if (rg->from > last_accounted_offset)
422 add += hugetlb_resv_map_add(resv, rg,
423 last_accounted_offset,
424 rg->from, h, h_cg,
425 regions_needed);
Mina Almasryd75c6af2019-11-30 17:56:59 -0800426
Mina Almasry0db9d742020-04-01 21:11:25 -0700427 last_accounted_offset = rg->to;
428 }
429
430 /* Handle the case where our range extends beyond
431 * last_accounted_offset.
432 */
Peter Xu2103cf92021-03-12 21:07:18 -0800433 if (last_accounted_offset < t)
434 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
435 t, h, h_cg, regions_needed);
Mina Almasry0db9d742020-04-01 21:11:25 -0700436
437 VM_BUG_ON(add < 0);
438 return add;
439}
440
441/* Must be called with resv->lock acquired. Will drop lock to allocate entries.
442 */
443static int allocate_file_region_entries(struct resv_map *resv,
444 int regions_needed)
445 __must_hold(&resv->lock)
446{
447 struct list_head allocated_regions;
448 int to_allocate = 0, i = 0;
449 struct file_region *trg = NULL, *rg = NULL;
450
451 VM_BUG_ON(regions_needed < 0);
452
453 INIT_LIST_HEAD(&allocated_regions);
454
455 /*
456 * Check for sufficient descriptors in the cache to accommodate
457 * the number of in progress add operations plus regions_needed.
458 *
459 * This is a while loop because when we drop the lock, some other call
460 * to region_add or region_del may have consumed some region_entries,
461 * so we keep looping here until we finally have enough entries for
462 * (adds_in_progress + regions_needed).
463 */
464 while (resv->region_cache_count <
465 (resv->adds_in_progress + regions_needed)) {
466 to_allocate = resv->adds_in_progress + regions_needed -
467 resv->region_cache_count;
468
469 /* At this point, we should have enough entries in the cache
470 * for all the existings adds_in_progress. We should only be
471 * needing to allocate for regions_needed.
472 */
473 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
474
475 spin_unlock(&resv->lock);
476 for (i = 0; i < to_allocate; i++) {
477 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
478 if (!trg)
479 goto out_of_memory;
480 list_add(&trg->link, &allocated_regions);
481 }
482
483 spin_lock(&resv->lock);
484
Wei Yangd3ec7b62020-10-13 16:56:27 -0700485 list_splice(&allocated_regions, &resv->region_cache);
486 resv->region_cache_count += to_allocate;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800487 }
488
Mina Almasry0db9d742020-04-01 21:11:25 -0700489 return 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800490
Mina Almasry0db9d742020-04-01 21:11:25 -0700491out_of_memory:
492 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
493 list_del(&rg->link);
494 kfree(rg);
495 }
496 return -ENOMEM;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800497}
498
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700499/*
500 * Add the huge page range represented by [f, t) to the reserve
Mina Almasry0db9d742020-04-01 21:11:25 -0700501 * map. Regions will be taken from the cache to fill in this range.
502 * Sufficient regions should exist in the cache due to the previous
503 * call to region_chg with the same range, but in some cases the cache will not
504 * have sufficient entries due to races with other code doing region_add or
505 * region_del. The extra needed entries will be allocated.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700506 *
Mina Almasry0db9d742020-04-01 21:11:25 -0700507 * regions_needed is the out value provided by a previous call to region_chg.
508 *
509 * Return the number of new huge pages added to the map. This number is greater
510 * than or equal to zero. If file_region entries needed to be allocated for
Ethon Paul7c8de352020-06-04 16:49:07 -0700511 * this operation and we were not able to allocate, it returns -ENOMEM.
Mina Almasry0db9d742020-04-01 21:11:25 -0700512 * region_add of regions of length 1 never allocate file_regions and cannot
513 * fail; region_chg will always allocate at least 1 entry and a region_add for
514 * 1 page will only require at most 1 entry.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700515 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700516static long region_add(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700517 long in_regions_needed, struct hstate *h,
518 struct hugetlb_cgroup *h_cg)
Andy Whitcroft96822902008-07-23 21:27:29 -0700519{
Mina Almasry0db9d742020-04-01 21:11:25 -0700520 long add = 0, actual_regions_needed = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700521
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700522 spin_lock(&resv->lock);
Mina Almasry0db9d742020-04-01 21:11:25 -0700523retry:
524
525 /* Count how many regions are actually needed to execute this add. */
Wei Yang972a3da32020-10-13 16:56:30 -0700526 add_reservation_in_range(resv, f, t, NULL, NULL,
527 &actual_regions_needed);
Andy Whitcroft96822902008-07-23 21:27:29 -0700528
Mike Kravetz5e911372015-09-08 15:01:28 -0700529 /*
Mina Almasry0db9d742020-04-01 21:11:25 -0700530 * Check for sufficient descriptors in the cache to accommodate
531 * this add operation. Note that actual_regions_needed may be greater
532 * than in_regions_needed, as the resv_map may have been modified since
533 * the region_chg call. In this case, we need to make sure that we
534 * allocate extra entries, such that we have enough for all the
535 * existing adds_in_progress, plus the excess needed for this
536 * operation.
Mike Kravetz5e911372015-09-08 15:01:28 -0700537 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700538 if (actual_regions_needed > in_regions_needed &&
539 resv->region_cache_count <
540 resv->adds_in_progress +
541 (actual_regions_needed - in_regions_needed)) {
542 /* region_add operation of range 1 should never need to
543 * allocate file_region entries.
544 */
545 VM_BUG_ON(t - f <= 1);
Mike Kravetz5e911372015-09-08 15:01:28 -0700546
Mina Almasry0db9d742020-04-01 21:11:25 -0700547 if (allocate_file_region_entries(
548 resv, actual_regions_needed - in_regions_needed)) {
549 return -ENOMEM;
550 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700551
Mina Almasry0db9d742020-04-01 21:11:25 -0700552 goto retry;
Mike Kravetz5e911372015-09-08 15:01:28 -0700553 }
554
Wei Yang972a3da32020-10-13 16:56:30 -0700555 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700556
Mina Almasry0db9d742020-04-01 21:11:25 -0700557 resv->adds_in_progress -= in_regions_needed;
558
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700559 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700560 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700561}
562
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700563/*
564 * Examine the existing reserve map and determine how many
565 * huge pages in the specified range [f, t) are NOT currently
566 * represented. This routine is called before a subsequent
567 * call to region_add that will actually modify the reserve
568 * map to add the specified range [f, t). region_chg does
569 * not change the number of huge pages represented by the
Mina Almasry0db9d742020-04-01 21:11:25 -0700570 * map. A number of new file_region structures is added to the cache as a
571 * placeholder, for the subsequent region_add call to use. At least 1
572 * file_region structure is added.
573 *
574 * out_regions_needed is the number of regions added to the
575 * resv->adds_in_progress. This value needs to be provided to a follow up call
576 * to region_add or region_abort for proper accounting.
Mike Kravetz5e911372015-09-08 15:01:28 -0700577 *
578 * Returns the number of huge pages that need to be added to the existing
579 * reservation map for the range [f, t). This number is greater or equal to
580 * zero. -ENOMEM is returned if a new file_region structure or cache entry
581 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700582 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700583static long region_chg(struct resv_map *resv, long f, long t,
584 long *out_regions_needed)
Andy Whitcroft96822902008-07-23 21:27:29 -0700585{
Andy Whitcroft96822902008-07-23 21:27:29 -0700586 long chg = 0;
587
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700588 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700589
Wei Yang972a3da32020-10-13 16:56:30 -0700590 /* Count how many hugepages in this range are NOT represented. */
Mina Almasry075a61d2020-04-01 21:11:28 -0700591 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
Wei Yang972a3da32020-10-13 16:56:30 -0700592 out_regions_needed);
Mike Kravetz5e911372015-09-08 15:01:28 -0700593
Mina Almasry0db9d742020-04-01 21:11:25 -0700594 if (*out_regions_needed == 0)
595 *out_regions_needed = 1;
Mike Kravetz5e911372015-09-08 15:01:28 -0700596
Mina Almasry0db9d742020-04-01 21:11:25 -0700597 if (allocate_file_region_entries(resv, *out_regions_needed))
598 return -ENOMEM;
Mike Kravetz5e911372015-09-08 15:01:28 -0700599
Mina Almasry0db9d742020-04-01 21:11:25 -0700600 resv->adds_in_progress += *out_regions_needed;
Andy Whitcroft96822902008-07-23 21:27:29 -0700601
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700602 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700603 return chg;
604}
605
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700606/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700607 * Abort the in progress add operation. The adds_in_progress field
608 * of the resv_map keeps track of the operations in progress between
609 * calls to region_chg and region_add. Operations are sometimes
610 * aborted after the call to region_chg. In such cases, region_abort
Mina Almasry0db9d742020-04-01 21:11:25 -0700611 * is called to decrement the adds_in_progress counter. regions_needed
612 * is the value returned by the region_chg call, it is used to decrement
613 * the adds_in_progress counter.
Mike Kravetz5e911372015-09-08 15:01:28 -0700614 *
615 * NOTE: The range arguments [f, t) are not needed or used in this
616 * routine. They are kept to make reading the calling code easier as
617 * arguments will match the associated region_chg call.
618 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700619static void region_abort(struct resv_map *resv, long f, long t,
620 long regions_needed)
Mike Kravetz5e911372015-09-08 15:01:28 -0700621{
622 spin_lock(&resv->lock);
623 VM_BUG_ON(!resv->region_cache_count);
Mina Almasry0db9d742020-04-01 21:11:25 -0700624 resv->adds_in_progress -= regions_needed;
Mike Kravetz5e911372015-09-08 15:01:28 -0700625 spin_unlock(&resv->lock);
626}
627
628/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700629 * Delete the specified range [f, t) from the reserve map. If the
630 * t parameter is LONG_MAX, this indicates that ALL regions after f
631 * should be deleted. Locate the regions which intersect [f, t)
632 * and either trim, delete or split the existing regions.
633 *
634 * Returns the number of huge pages deleted from the reserve map.
635 * In the normal case, the return value is zero or more. In the
636 * case where a region must be split, a new region descriptor must
637 * be allocated. If the allocation fails, -ENOMEM will be returned.
638 * NOTE: If the parameter t == LONG_MAX, then we will never split
639 * a region and possibly return -ENOMEM. Callers specifying
640 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700641 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700642static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700643{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700644 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700645 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700646 struct file_region *nrg = NULL;
647 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700648
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700649retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700650 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700651 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800652 /*
653 * Skip regions before the range to be deleted. file_region
654 * ranges are normally of the form [from, to). However, there
655 * may be a "placeholder" entry in the map which is of the form
656 * (from, to) with from == to. Check for placeholder entries
657 * at the beginning of the range to be deleted.
658 */
659 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700660 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800661
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700662 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700663 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700664
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700665 if (f > rg->from && t < rg->to) { /* Must split region */
666 /*
667 * Check for an entry in the cache before dropping
668 * lock and attempting allocation.
669 */
670 if (!nrg &&
671 resv->region_cache_count > resv->adds_in_progress) {
672 nrg = list_first_entry(&resv->region_cache,
673 struct file_region,
674 link);
675 list_del(&nrg->link);
676 resv->region_cache_count--;
677 }
678
679 if (!nrg) {
680 spin_unlock(&resv->lock);
681 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
682 if (!nrg)
683 return -ENOMEM;
684 goto retry;
685 }
686
687 del += t - f;
Mike Kravetz79aa9252020-11-01 17:07:27 -0800688 hugetlb_cgroup_uncharge_file_region(
Miaohe Lind85aecf2021-03-24 21:37:17 -0700689 resv, rg, t - f, false);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700690
691 /* New entry for end of split region */
692 nrg->from = t;
693 nrg->to = rg->to;
Mina Almasry075a61d2020-04-01 21:11:28 -0700694
695 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
696
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700697 INIT_LIST_HEAD(&nrg->link);
698
699 /* Original entry is trimmed */
700 rg->to = f;
701
702 list_add(&nrg->link, &rg->link);
703 nrg = NULL;
704 break;
705 }
706
707 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
708 del += rg->to - rg->from;
Mina Almasry075a61d2020-04-01 21:11:28 -0700709 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Lind85aecf2021-03-24 21:37:17 -0700710 rg->to - rg->from, true);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700711 list_del(&rg->link);
712 kfree(rg);
713 continue;
714 }
715
716 if (f <= rg->from) { /* Trim beginning of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700717 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Lind85aecf2021-03-24 21:37:17 -0700718 t - rg->from, false);
Mina Almasry075a61d2020-04-01 21:11:28 -0700719
Mike Kravetz79aa9252020-11-01 17:07:27 -0800720 del += t - rg->from;
721 rg->from = t;
722 } else { /* Trim end of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700723 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Lind85aecf2021-03-24 21:37:17 -0700724 rg->to - f, false);
Mike Kravetz79aa9252020-11-01 17:07:27 -0800725
726 del += rg->to - f;
727 rg->to = f;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700728 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700729 }
730
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700731 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700732 kfree(nrg);
733 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700734}
735
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700736/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700737 * A rare out of memory error was encountered which prevented removal of
738 * the reserve map region for a page. The huge page itself was free'ed
739 * and removed from the page cache. This routine will adjust the subpool
740 * usage count, and the global reserve count if needed. By incrementing
741 * these counts, the reserve map entry which could not be deleted will
742 * appear as a "reserved" entry instead of simply dangling with incorrect
743 * counts.
744 */
zhong jiang72e29362016-10-07 17:02:01 -0700745void hugetlb_fix_reserve_counts(struct inode *inode)
Mike Kravetzb5cec282015-09-08 15:01:41 -0700746{
747 struct hugepage_subpool *spool = subpool_inode(inode);
748 long rsv_adjust;
Miaohe Linda563882021-05-04 18:34:38 -0700749 bool reserved = false;
Mike Kravetzb5cec282015-09-08 15:01:41 -0700750
751 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
Miaohe Linda563882021-05-04 18:34:38 -0700752 if (rsv_adjust > 0) {
Mike Kravetzb5cec282015-09-08 15:01:41 -0700753 struct hstate *h = hstate_inode(inode);
754
Miaohe Linda563882021-05-04 18:34:38 -0700755 if (!hugetlb_acct_memory(h, 1))
756 reserved = true;
757 } else if (!rsv_adjust) {
758 reserved = true;
Mike Kravetzb5cec282015-09-08 15:01:41 -0700759 }
Miaohe Linda563882021-05-04 18:34:38 -0700760
761 if (!reserved)
762 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
Mike Kravetzb5cec282015-09-08 15:01:41 -0700763}
764
765/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700766 * Count and return the number of huge pages in the reserve map
767 * that intersect with the range [f, t).
768 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700769static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700770{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700771 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700772 struct file_region *rg;
773 long chg = 0;
774
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700775 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700776 /* Locate each segment we overlap with, and count that overlap. */
777 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700778 long seg_from;
779 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700780
781 if (rg->to <= f)
782 continue;
783 if (rg->from >= t)
784 break;
785
786 seg_from = max(rg->from, f);
787 seg_to = min(rg->to, t);
788
789 chg += seg_to - seg_from;
790 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700791 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700792
793 return chg;
794}
795
Andy Whitcroft96822902008-07-23 21:27:29 -0700796/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700797 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700798 * the mapping, in pagecache page units; huge pages here.
799 */
Andi Kleena5516432008-07-23 21:27:41 -0700800static pgoff_t vma_hugecache_offset(struct hstate *h,
801 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700802{
Andi Kleena5516432008-07-23 21:27:41 -0700803 return ((address - vma->vm_start) >> huge_page_shift(h)) +
804 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700805}
806
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900807pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
808 unsigned long address)
809{
810 return vma_hugecache_offset(hstate_vma(vma), vma, address);
811}
Dan Williamsdee41072016-05-14 12:20:44 -0700812EXPORT_SYMBOL_GPL(linear_hugepage_index);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900813
Andy Whitcroft84afd992008-07-23 21:27:32 -0700814/*
Mel Gorman08fba692009-01-06 14:38:53 -0800815 * Return the size of the pages allocated when backing a VMA. In the majority
816 * cases this will be same size as used by the page table entries.
817 */
818unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
819{
Dan Williams05ea8862018-04-05 16:24:25 -0700820 if (vma->vm_ops && vma->vm_ops->pagesize)
821 return vma->vm_ops->pagesize(vma);
822 return PAGE_SIZE;
Mel Gorman08fba692009-01-06 14:38:53 -0800823}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200824EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800825
826/*
Mel Gorman33402892009-01-06 14:38:54 -0800827 * Return the page size being used by the MMU to back a VMA. In the majority
828 * of cases, the page size used by the kernel matches the MMU size. On
Dan Williams09135cc2018-04-05 16:24:21 -0700829 * architectures where it differs, an architecture-specific 'strong'
830 * version of this symbol is required.
Mel Gorman33402892009-01-06 14:38:54 -0800831 */
Dan Williams09135cc2018-04-05 16:24:21 -0700832__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
Mel Gorman33402892009-01-06 14:38:54 -0800833{
834 return vma_kernel_pagesize(vma);
835}
Mel Gorman33402892009-01-06 14:38:54 -0800836
837/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700838 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
839 * bits of the reservation map pointer, which are always clear due to
840 * alignment.
841 */
842#define HPAGE_RESV_OWNER (1UL << 0)
843#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700844#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700845
Mel Gormana1e78772008-07-23 21:27:23 -0700846/*
847 * These helpers are used to track how many pages are reserved for
848 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
849 * is guaranteed to have their future faults succeed.
850 *
851 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
852 * the reserve counters are updated with the hugetlb_lock held. It is safe
853 * to reset the VMA at fork() time as it is not in use yet and there is no
854 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700855 *
856 * The private mapping reservation is represented in a subtly different
857 * manner to a shared mapping. A shared mapping has a region map associated
858 * with the underlying file, this region map represents the backing file
859 * pages which have ever had a reservation assigned which this persists even
860 * after the page is instantiated. A private mapping has a region map
861 * associated with the original mmap which is attached to all VMAs which
862 * reference it, this region map represents those offsets which have consumed
863 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700864 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700865static unsigned long get_vma_private_data(struct vm_area_struct *vma)
866{
867 return (unsigned long)vma->vm_private_data;
868}
869
870static void set_vma_private_data(struct vm_area_struct *vma,
871 unsigned long value)
872{
873 vma->vm_private_data = (void *)value;
874}
875
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700876static void
877resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
878 struct hugetlb_cgroup *h_cg,
879 struct hstate *h)
880{
881#ifdef CONFIG_CGROUP_HUGETLB
882 if (!h_cg || !h) {
883 resv_map->reservation_counter = NULL;
884 resv_map->pages_per_hpage = 0;
885 resv_map->css = NULL;
886 } else {
887 resv_map->reservation_counter =
888 &h_cg->rsvd_hugepage[hstate_index(h)];
889 resv_map->pages_per_hpage = pages_per_huge_page(h);
890 resv_map->css = &h_cg->css;
891 }
892#endif
893}
894
Joonsoo Kim9119a412014-04-03 14:47:25 -0700895struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700896{
897 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700898 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
899
900 if (!resv_map || !rg) {
901 kfree(resv_map);
902 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700903 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700904 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700905
906 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700907 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700908 INIT_LIST_HEAD(&resv_map->regions);
909
Mike Kravetz5e911372015-09-08 15:01:28 -0700910 resv_map->adds_in_progress = 0;
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700911 /*
912 * Initialize these to 0. On shared mappings, 0's here indicate these
913 * fields don't do cgroup accounting. On private mappings, these will be
914 * re-initialized to the proper values, to indicate that hugetlb cgroup
915 * reservations are to be un-charged from here.
916 */
917 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700918
919 INIT_LIST_HEAD(&resv_map->region_cache);
920 list_add(&rg->link, &resv_map->region_cache);
921 resv_map->region_cache_count = 1;
922
Andy Whitcroft84afd992008-07-23 21:27:32 -0700923 return resv_map;
924}
925
Joonsoo Kim9119a412014-04-03 14:47:25 -0700926void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700927{
928 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700929 struct list_head *head = &resv_map->region_cache;
930 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700931
932 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700933 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700934
935 /* ... and any entries left in the cache */
936 list_for_each_entry_safe(rg, trg, head, link) {
937 list_del(&rg->link);
938 kfree(rg);
939 }
940
941 VM_BUG_ON(resv_map->adds_in_progress);
942
Andy Whitcroft84afd992008-07-23 21:27:32 -0700943 kfree(resv_map);
944}
945
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700946static inline struct resv_map *inode_resv_map(struct inode *inode)
947{
Mike Kravetzf27a5132019-05-13 17:22:55 -0700948 /*
949 * At inode evict time, i_mapping may not point to the original
950 * address space within the inode. This original address space
951 * contains the pointer to the resv_map. So, always use the
952 * address space embedded within the inode.
953 * The VERY common case is inode->mapping == &inode->i_data but,
954 * this may not be true for device special inodes.
955 */
956 return (struct resv_map *)(&inode->i_data)->private_data;
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700957}
958
Andy Whitcroft84afd992008-07-23 21:27:32 -0700959static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700960{
Sasha Levin81d1b092014-10-09 15:28:10 -0700961 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700962 if (vma->vm_flags & VM_MAYSHARE) {
963 struct address_space *mapping = vma->vm_file->f_mapping;
964 struct inode *inode = mapping->host;
965
966 return inode_resv_map(inode);
967
968 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700969 return (struct resv_map *)(get_vma_private_data(vma) &
970 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700971 }
Mel Gormana1e78772008-07-23 21:27:23 -0700972}
973
Andy Whitcroft84afd992008-07-23 21:27:32 -0700974static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700975{
Sasha Levin81d1b092014-10-09 15:28:10 -0700976 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
977 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700978
Andy Whitcroft84afd992008-07-23 21:27:32 -0700979 set_vma_private_data(vma, (get_vma_private_data(vma) &
980 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700981}
982
983static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
984{
Sasha Levin81d1b092014-10-09 15:28:10 -0700985 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
986 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700987
988 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700989}
990
991static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
992{
Sasha Levin81d1b092014-10-09 15:28:10 -0700993 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700994
995 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700996}
997
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700998/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700999void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1000{
Sasha Levin81d1b092014-10-09 15:28:10 -07001001 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -07001002 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -07001003 vma->vm_private_data = (void *)0;
1004}
1005
1006/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001007static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001008{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001009 if (vma->vm_flags & VM_NORESERVE) {
1010 /*
1011 * This address is already reserved by other process(chg == 0),
1012 * so, we should decrement reserved count. Without decrementing,
1013 * reserve count remains after releasing inode, because this
1014 * allocated page will go into page cache and is regarded as
1015 * coming from reserved pool in releasing step. Currently, we
1016 * don't have any other solution to deal with this situation
1017 * properly, so add work-around here.
1018 */
1019 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001020 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001021 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001022 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001023 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001024
1025 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -07001026 if (vma->vm_flags & VM_MAYSHARE) {
1027 /*
1028 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1029 * be a region map for all pages. The only situation where
1030 * there is no region map is if a hole was punched via
Ethon Paul7c8de352020-06-04 16:49:07 -07001031 * fallocate. In this case, there really are no reserves to
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -07001032 * use. This situation is indicated if chg != 0.
1033 */
1034 if (chg)
1035 return false;
1036 else
1037 return true;
1038 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001039
1040 /*
1041 * Only the process that called mmap() has reserves for
1042 * private mappings.
1043 */
Mike Kravetz67961f92016-06-08 15:33:42 -07001044 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1045 /*
1046 * Like the shared case above, a hole punch or truncate
1047 * could have been performed on the private mapping.
1048 * Examine the value of chg to determine if reserves
1049 * actually exist or were previously consumed.
1050 * Very Subtle - The value of chg comes from a previous
1051 * call to vma_needs_reserves(). The reserve map for
1052 * private mappings has different (opposite) semantics
1053 * than that of shared mappings. vma_needs_reserves()
1054 * has already taken this difference in semantics into
1055 * account. Therefore, the meaning of chg is the same
1056 * as in the shared case above. Code could easily be
1057 * combined, but keeping it separate draws attention to
1058 * subtle differences.
1059 */
1060 if (chg)
1061 return false;
1062 else
1063 return true;
1064 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001065
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001066 return false;
Mel Gormana1e78772008-07-23 21:27:23 -07001067}
1068
Andi Kleena5516432008-07-23 21:27:41 -07001069static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070{
1071 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001072 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -07001073 h->free_huge_pages++;
1074 h->free_huge_pages_node[nid]++;
Mike Kravetz6c037142021-02-24 12:09:04 -08001075 SetHPageFreed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076}
1077
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001078static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001079{
1080 struct page *page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001081 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001082
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001083 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1084 if (nocma && is_migrate_cma_page(page))
1085 continue;
1086
Wei Yang6664bfc2020-10-13 16:56:39 -07001087 if (PageHWPoison(page))
1088 continue;
1089
1090 list_move(&page->lru, &h->hugepage_activelist);
1091 set_page_refcounted(page);
Mike Kravetz6c037142021-02-24 12:09:04 -08001092 ClearHPageFreed(page);
Wei Yang6664bfc2020-10-13 16:56:39 -07001093 h->free_huge_pages--;
1094 h->free_huge_pages_node[nid]--;
1095 return page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001096 }
1097
Wei Yang6664bfc2020-10-13 16:56:39 -07001098 return NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001099}
1100
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001101static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1102 nodemask_t *nmask)
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001103{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001104 unsigned int cpuset_mems_cookie;
1105 struct zonelist *zonelist;
1106 struct zone *zone;
1107 struct zoneref *z;
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001108 int node = NUMA_NO_NODE;
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001109
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001110 zonelist = node_zonelist(nid, gfp_mask);
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001111
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001112retry_cpuset:
1113 cpuset_mems_cookie = read_mems_allowed_begin();
1114 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1115 struct page *page;
1116
1117 if (!cpuset_zone_allowed(zone, gfp_mask))
1118 continue;
1119 /*
1120 * no need to ask again on the same node. Pool is node rather than
1121 * zone aware
1122 */
1123 if (zone_to_nid(zone) == node)
1124 continue;
1125 node = zone_to_nid(zone);
1126
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001127 page = dequeue_huge_page_node_exact(h, node);
1128 if (page)
1129 return page;
1130 }
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001131 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1132 goto retry_cpuset;
1133
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001134 return NULL;
1135}
1136
Andi Kleena5516432008-07-23 21:27:41 -07001137static struct page *dequeue_huge_page_vma(struct hstate *h,
1138 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001139 unsigned long address, int avoid_reserve,
1140 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001142 struct page *page;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001143 struct mempolicy *mpol;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001144 gfp_t gfp_mask;
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001145 nodemask_t *nodemask;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001146 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147
Mel Gormana1e78772008-07-23 21:27:23 -07001148 /*
1149 * A child process with MAP_PRIVATE mappings created by their parent
1150 * have no page reserves. This check ensures that reservations are
1151 * not "stolen". The child may still get SIGKILLed
1152 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001153 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -07001154 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -07001155 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -07001156
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001157 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -07001158 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -07001159 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001160
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001161 gfp_mask = htlb_alloc_mask(h);
1162 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001163 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1164 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
Mike Kravetzd6995da2021-02-24 12:08:51 -08001165 SetHPageRestoreReserve(page);
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001166 h->resv_huge_pages--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07001168
1169 mpol_cond_put(mpol);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001170 return page;
1171
Miao Xiec0ff7452010-05-24 14:32:08 -07001172err:
Mel Gormancc9a6c82012-03-21 16:34:11 -07001173 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174}
1175
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001176/*
1177 * common helper functions for hstate_next_node_to_{alloc|free}.
1178 * We may have allocated or freed a huge page based on a different
1179 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1180 * be outside of *nodes_allowed. Ensure that we use an allowed
1181 * node for alloc or free.
1182 */
1183static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1184{
Andrew Morton0edaf862016-05-19 17:10:58 -07001185 nid = next_node_in(nid, *nodes_allowed);
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001186 VM_BUG_ON(nid >= MAX_NUMNODES);
1187
1188 return nid;
1189}
1190
1191static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1192{
1193 if (!node_isset(nid, *nodes_allowed))
1194 nid = next_node_allowed(nid, nodes_allowed);
1195 return nid;
1196}
1197
1198/*
1199 * returns the previously saved node ["this node"] from which to
1200 * allocate a persistent huge page for the pool and advance the
1201 * next node from which to allocate, handling wrap at end of node
1202 * mask.
1203 */
1204static int hstate_next_node_to_alloc(struct hstate *h,
1205 nodemask_t *nodes_allowed)
1206{
1207 int nid;
1208
1209 VM_BUG_ON(!nodes_allowed);
1210
1211 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1212 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1213
1214 return nid;
1215}
1216
1217/*
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001218 * helper for remove_pool_huge_page() - return the previously saved
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001219 * node ["this node"] from which to free a huge page. Advance the
1220 * next node id whether or not we find a free huge page to free so
1221 * that the next attempt to free addresses the next node.
1222 */
1223static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1224{
1225 int nid;
1226
1227 VM_BUG_ON(!nodes_allowed);
1228
1229 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1230 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1231
1232 return nid;
1233}
1234
1235#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1236 for (nr_nodes = nodes_weight(*mask); \
1237 nr_nodes > 0 && \
1238 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1239 nr_nodes--)
1240
1241#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1242 for (nr_nodes = nodes_weight(*mask); \
1243 nr_nodes > 0 && \
1244 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1245 nr_nodes--)
1246
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001247#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001248static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001249 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001250{
1251 int i;
1252 int nr_pages = 1 << order;
1253 struct page *p = page + 1;
1254
Gerald Schaeferc8cc7082016-06-24 14:50:04 -07001255 atomic_set(compound_mapcount_ptr(page), 0);
Yanfei Xu5291c092021-02-24 12:07:22 -08001256 atomic_set(compound_pincount_ptr(page), 0);
John Hubbard47e29d32020-04-01 21:05:33 -07001257
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001258 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001259 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001260 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001261 }
1262
1263 set_compound_order(page, 0);
Gerald Schaeferba9c1202020-12-11 13:36:53 -08001264 page[1].compound_nr = 0;
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001265 __ClearPageHead(page);
1266}
1267
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001268static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001269{
Roman Gushchincf11e852020-04-10 14:32:45 -07001270 /*
1271 * If the page isn't allocated using the cma allocator,
1272 * cma_release() returns false.
1273 */
Barry Songdbda8fe2020-07-23 21:15:30 -07001274#ifdef CONFIG_CMA
1275 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
Roman Gushchincf11e852020-04-10 14:32:45 -07001276 return;
Barry Songdbda8fe2020-07-23 21:15:30 -07001277#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001278
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001279 free_contig_range(page_to_pfn(page), 1 << order);
1280}
1281
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001282#ifdef CONFIG_CONTIG_ALLOC
Michal Hockod9cc948f2018-01-31 16:20:44 -08001283static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1284 int nid, nodemask_t *nodemask)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001285{
Miaohe Lin04adbc32021-05-04 18:33:22 -07001286 unsigned long nr_pages = pages_per_huge_page(h);
Li Xinhai953f0642020-09-04 16:36:10 -07001287 if (nid == NUMA_NO_NODE)
1288 nid = numa_mem_id();
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001289
Barry Songdbda8fe2020-07-23 21:15:30 -07001290#ifdef CONFIG_CMA
1291 {
Roman Gushchincf11e852020-04-10 14:32:45 -07001292 struct page *page;
1293 int node;
1294
Li Xinhai953f0642020-09-04 16:36:10 -07001295 if (hugetlb_cma[nid]) {
1296 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1297 huge_page_order(h), true);
Roman Gushchincf11e852020-04-10 14:32:45 -07001298 if (page)
1299 return page;
1300 }
Li Xinhai953f0642020-09-04 16:36:10 -07001301
1302 if (!(gfp_mask & __GFP_THISNODE)) {
1303 for_each_node_mask(node, *nodemask) {
1304 if (node == nid || !hugetlb_cma[node])
1305 continue;
1306
1307 page = cma_alloc(hugetlb_cma[node], nr_pages,
1308 huge_page_order(h), true);
1309 if (page)
1310 return page;
1311 }
1312 }
Roman Gushchincf11e852020-04-10 14:32:45 -07001313 }
Barry Songdbda8fe2020-07-23 21:15:30 -07001314#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001315
Anshuman Khandual5e27a2d2019-11-30 17:55:06 -08001316 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001317}
1318
1319static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001320static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001321#else /* !CONFIG_CONTIG_ALLOC */
1322static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1323 int nid, nodemask_t *nodemask)
1324{
1325 return NULL;
1326}
1327#endif /* CONFIG_CONTIG_ALLOC */
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001328
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001329#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
Michal Hockod9cc948f2018-01-31 16:20:44 -08001330static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001331 int nid, nodemask_t *nodemask)
1332{
1333 return NULL;
1334}
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001335static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001336static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001337 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001338#endif
1339
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001340/*
1341 * Remove hugetlb page from lists, and update dtor so that page appears
1342 * as just a compound page. A reference is held on the page.
1343 *
1344 * Must be called with hugetlb lock held.
1345 */
1346static void remove_hugetlb_page(struct hstate *h, struct page *page,
1347 bool adjust_surplus)
1348{
1349 int nid = page_to_nid(page);
1350
1351 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1352 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
1353
1354 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1355 return;
1356
1357 list_del(&page->lru);
1358
1359 if (HPageFreed(page)) {
1360 h->free_huge_pages--;
1361 h->free_huge_pages_node[nid]--;
1362 }
1363 if (adjust_surplus) {
1364 h->surplus_huge_pages--;
1365 h->surplus_huge_pages_node[nid]--;
1366 }
1367
1368 set_page_refcounted(page);
1369 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1370
1371 h->nr_huge_pages--;
1372 h->nr_huge_pages_node[nid]--;
1373}
1374
Andi Kleena5516432008-07-23 21:27:41 -07001375static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001376{
1377 int i;
Mike Kravetzdbfee5a2021-02-24 12:07:50 -08001378 struct page *subpage = page;
Andi Kleena5516432008-07-23 21:27:41 -07001379
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001380 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001381 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001382
Mike Kravetzdbfee5a2021-02-24 12:07:50 -08001383 for (i = 0; i < pages_per_huge_page(h);
1384 i++, subpage = mem_map_next(subpage, page, i)) {
1385 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
Chris Forbes32f84522011-07-25 17:12:14 -07001386 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001387 1 << PG_active | 1 << PG_private |
1388 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001389 }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001390 if (hstate_is_gigantic(h)) {
1391 destroy_compound_gigantic_page(page, huge_page_order(h));
1392 free_gigantic_page(page, huge_page_order(h));
1393 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001394 __free_pages(page, huge_page_order(h));
1395 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001396}
1397
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001398static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1399{
1400 struct page *page, *t_page;
1401
1402 list_for_each_entry_safe(page, t_page, list, lru) {
1403 update_and_free_page(h, page);
1404 cond_resched();
1405 }
1406}
1407
Andi Kleene5ff2152008-07-23 21:27:42 -07001408struct hstate *size_to_hstate(unsigned long size)
1409{
1410 struct hstate *h;
1411
1412 for_each_hstate(h) {
1413 if (huge_page_size(h) == size)
1414 return h;
1415 }
1416 return NULL;
1417}
1418
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001419void free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001420{
Andi Kleena5516432008-07-23 21:27:41 -07001421 /*
1422 * Can't pass hstate in here because it is called from the
1423 * compound page destructor.
1424 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001425 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001426 int nid = page_to_nid(page);
Mike Kravetzd6995da2021-02-24 12:08:51 -08001427 struct hugepage_subpool *spool = hugetlb_page_subpool(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001428 bool restore_reserve;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001429 unsigned long flags;
David Gibson27a85ef2006-03-22 00:08:56 -08001430
Mike Kravetzb4330af2016-02-05 15:36:38 -08001431 VM_BUG_ON_PAGE(page_count(page), page);
1432 VM_BUG_ON_PAGE(page_mapcount(page), page);
Yongkai Wu8ace22b2018-12-14 14:17:10 -08001433
Mike Kravetzd6995da2021-02-24 12:08:51 -08001434 hugetlb_set_page_subpool(page, NULL);
Yongkai Wu8ace22b2018-12-14 14:17:10 -08001435 page->mapping = NULL;
Mike Kravetzd6995da2021-02-24 12:08:51 -08001436 restore_reserve = HPageRestoreReserve(page);
1437 ClearHPageRestoreReserve(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001438
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001439 /*
Mike Kravetzd6995da2021-02-24 12:08:51 -08001440 * If HPageRestoreReserve was set on page, page allocation consumed a
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001441 * reservation. If the page was associated with a subpool, there
1442 * would have been a page reserved in the subpool before allocation
1443 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
Miaohe Lin6c26d312021-02-24 12:07:19 -08001444 * reservation, do not call hugepage_subpool_put_pages() as this will
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001445 * remove the reserved page from the subpool.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001446 */
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001447 if (!restore_reserve) {
1448 /*
1449 * A return code of zero implies that the subpool will be
1450 * under its minimum size if the reservation is not restored
1451 * after page is free. Therefore, force restore_reserve
1452 * operation.
1453 */
1454 if (hugepage_subpool_put_pages(spool, 1) == 0)
1455 restore_reserve = true;
1456 }
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001457
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001458 spin_lock_irqsave(&hugetlb_lock, flags);
Mike Kravetz8f251a32021-02-24 12:08:56 -08001459 ClearHPageMigratable(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001460 hugetlb_cgroup_uncharge_page(hstate_index(h),
1461 pages_per_huge_page(h), page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07001462 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1463 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001464 if (restore_reserve)
1465 h->resv_huge_pages++;
1466
Mike Kravetz9157c3112021-02-24 12:09:00 -08001467 if (HPageTemporary(page)) {
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001468 remove_hugetlb_page(h, page, false);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001469 spin_unlock_irqrestore(&hugetlb_lock, flags);
Michal Hockoab5ac902018-01-31 16:20:48 -08001470 update_and_free_page(h, page);
1471 } else if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001472 /* remove the page from active list */
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001473 remove_hugetlb_page(h, page, true);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001474 spin_unlock_irqrestore(&hugetlb_lock, flags);
Andi Kleena5516432008-07-23 21:27:41 -07001475 update_and_free_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001476 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001477 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001478 enqueue_huge_page(h, page);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001479 spin_unlock_irqrestore(&hugetlb_lock, flags);
Adam Litke7893d1d2007-10-16 01:26:18 -07001480 }
David Gibson27a85ef2006-03-22 00:08:56 -08001481}
1482
Andi Kleena5516432008-07-23 21:27:41 -07001483static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001484{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001485 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001486 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Mike Kravetzff546112021-02-24 12:09:11 -08001487 hugetlb_set_page_subpool(page, NULL);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001488 set_hugetlb_cgroup(page, NULL);
Mina Almasry1adc4d42020-04-01 21:11:15 -07001489 set_hugetlb_cgroup_rsvd(page, NULL);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001490 spin_lock_irq(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001491 h->nr_huge_pages++;
1492 h->nr_huge_pages_node[nid]++;
Mike Kravetz6c037142021-02-24 12:09:04 -08001493 ClearHPageFreed(page);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001494 spin_unlock_irq(&hugetlb_lock);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001495}
1496
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001497static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001498{
1499 int i;
1500 int nr_pages = 1 << order;
1501 struct page *p = page + 1;
1502
1503 /* we rely on prep_new_huge_page to set the destructor */
1504 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001505 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001506 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001507 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001508 /*
1509 * For gigantic hugepages allocated through bootmem at
1510 * boot, it's safer to be consistent with the not-gigantic
1511 * hugepages and clear the PG_reserved bit from all tail pages
Ethon Paul7c8de352020-06-04 16:49:07 -07001512 * too. Otherwise drivers using get_user_pages() to access tail
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001513 * pages may get the reference counting wrong if they see
1514 * PG_reserved set on a tail page (despite the head page not
1515 * having PG_reserved set). Enforcing this consistency between
1516 * head and tail pages allows drivers to optimize away a check
1517 * on the head page when they need know if put_page() is needed
1518 * after get_user_pages().
1519 */
1520 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001521 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001522 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001523 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001524 atomic_set(compound_mapcount_ptr(page), -1);
Yanfei Xu5291c092021-02-24 12:07:22 -08001525 atomic_set(compound_pincount_ptr(page), 0);
Wu Fengguang20a03072009-06-16 15:32:22 -07001526}
1527
Andrew Morton77959122012-10-08 16:34:11 -07001528/*
1529 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1530 * transparent huge pages. See the PageTransHuge() documentation for more
1531 * details.
1532 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001533int PageHuge(struct page *page)
1534{
Wu Fengguang20a03072009-06-16 15:32:22 -07001535 if (!PageCompound(page))
1536 return 0;
1537
1538 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001539 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001540}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001541EXPORT_SYMBOL_GPL(PageHuge);
1542
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001543/*
1544 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1545 * normal or transparent huge pages.
1546 */
1547int PageHeadHuge(struct page *page_head)
1548{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001549 if (!PageHead(page_head))
1550 return 0;
1551
Vlastimil Babkad4af73e2020-04-01 21:11:48 -07001552 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001553}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001554
Mike Kravetzc0d03812020-04-01 21:11:05 -07001555/*
Mike Kravetzc0d03812020-04-01 21:11:05 -07001556 * Find and lock address space (mapping) in write mode.
1557 *
Mike Kravetz336bf302020-11-13 22:52:16 -08001558 * Upon entry, the page is locked which means that page_mapping() is
1559 * stable. Due to locking order, we can only trylock_write. If we can
1560 * not get the lock, simply return NULL to caller.
Mike Kravetzc0d03812020-04-01 21:11:05 -07001561 */
1562struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1563{
Mike Kravetz336bf302020-11-13 22:52:16 -08001564 struct address_space *mapping = page_mapping(hpage);
Mike Kravetzc0d03812020-04-01 21:11:05 -07001565
Mike Kravetzc0d03812020-04-01 21:11:05 -07001566 if (!mapping)
1567 return mapping;
1568
Mike Kravetzc0d03812020-04-01 21:11:05 -07001569 if (i_mmap_trylock_write(mapping))
1570 return mapping;
1571
Mike Kravetz336bf302020-11-13 22:52:16 -08001572 return NULL;
Mike Kravetzc0d03812020-04-01 21:11:05 -07001573}
1574
Zhang Yi13d60f42013-06-25 21:19:31 +08001575pgoff_t __basepage_index(struct page *page)
1576{
1577 struct page *page_head = compound_head(page);
1578 pgoff_t index = page_index(page_head);
1579 unsigned long compound_idx;
1580
1581 if (!PageHuge(page_head))
1582 return page_index(page);
1583
1584 if (compound_order(page_head) >= MAX_ORDER)
1585 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1586 else
1587 compound_idx = page - page_head;
1588
1589 return (index << compound_order(page_head)) + compound_idx;
1590}
1591
Michal Hocko0c397da2018-01-31 16:20:56 -08001592static struct page *alloc_buddy_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001593 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1594 nodemask_t *node_alloc_noretry)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595{
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001596 int order = huge_page_order(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 struct page *page;
Mike Kravetzf60858f2019-09-23 15:37:35 -07001598 bool alloc_try_hard = true;
Joe Jinf96efd52007-07-15 23:38:12 -07001599
Mike Kravetzf60858f2019-09-23 15:37:35 -07001600 /*
1601 * By default we always try hard to allocate the page with
1602 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1603 * a loop (to adjust global huge page counts) and previous allocation
1604 * failed, do not continue to try hard on the same node. Use the
1605 * node_alloc_noretry bitmap to manage this state information.
1606 */
1607 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1608 alloc_try_hard = false;
1609 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1610 if (alloc_try_hard)
1611 gfp_mask |= __GFP_RETRY_MAYFAIL;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001612 if (nid == NUMA_NO_NODE)
1613 nid = numa_mem_id();
Matthew Wilcox (Oracle)84172f42021-04-29 23:01:15 -07001614 page = __alloc_pages(gfp_mask, order, nid, nmask);
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001615 if (page)
1616 __count_vm_event(HTLB_BUDDY_PGALLOC);
1617 else
1618 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001619
Mike Kravetzf60858f2019-09-23 15:37:35 -07001620 /*
1621 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1622 * indicates an overall state change. Clear bit so that we resume
1623 * normal 'try hard' allocations.
1624 */
1625 if (node_alloc_noretry && page && !alloc_try_hard)
1626 node_clear(nid, *node_alloc_noretry);
1627
1628 /*
1629 * If we tried hard to get a page but failed, set bit so that
1630 * subsequent attempts will not try as hard until there is an
1631 * overall state change.
1632 */
1633 if (node_alloc_noretry && !page && alloc_try_hard)
1634 node_set(nid, *node_alloc_noretry);
1635
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001636 return page;
1637}
1638
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001639/*
Michal Hocko0c397da2018-01-31 16:20:56 -08001640 * Common helper to allocate a fresh hugetlb page. All specific allocators
1641 * should use this function to get new hugetlb pages
1642 */
1643static struct page *alloc_fresh_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001644 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1645 nodemask_t *node_alloc_noretry)
Michal Hocko0c397da2018-01-31 16:20:56 -08001646{
1647 struct page *page;
1648
1649 if (hstate_is_gigantic(h))
1650 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1651 else
1652 page = alloc_buddy_huge_page(h, gfp_mask,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001653 nid, nmask, node_alloc_noretry);
Michal Hocko0c397da2018-01-31 16:20:56 -08001654 if (!page)
1655 return NULL;
1656
1657 if (hstate_is_gigantic(h))
1658 prep_compound_gigantic_page(page, huge_page_order(h));
1659 prep_new_huge_page(h, page, page_to_nid(page));
1660
1661 return page;
1662}
1663
1664/*
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001665 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1666 * manner.
1667 */
Mike Kravetzf60858f2019-09-23 15:37:35 -07001668static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1669 nodemask_t *node_alloc_noretry)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001670{
1671 struct page *page;
1672 int nr_nodes, node;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001673 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001674
1675 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
Mike Kravetzf60858f2019-09-23 15:37:35 -07001676 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1677 node_alloc_noretry);
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001678 if (page)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001679 break;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001680 }
1681
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001682 if (!page)
1683 return 0;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001684
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001685 put_page(page); /* free it into the hugepage allocator */
1686
1687 return 1;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001688}
1689
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001690/*
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001691 * Remove huge page from pool from next node to free. Attempt to keep
1692 * persistent huge pages more or less balanced over allowed nodes.
1693 * This routine only 'removes' the hugetlb page. The caller must make
1694 * an additional call to free the page to low level allocators.
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001695 * Called with hugetlb_lock locked.
1696 */
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001697static struct page *remove_pool_huge_page(struct hstate *h,
1698 nodemask_t *nodes_allowed,
1699 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001700{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001701 int nr_nodes, node;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001702 struct page *page = NULL;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001703
Joonsoo Kimb2261022013-09-11 14:21:00 -07001704 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001705 /*
1706 * If we're returning unused surplus pages, only examine
1707 * nodes with surplus pages.
1708 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001709 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1710 !list_empty(&h->hugepage_freelists[node])) {
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001711 page = list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001712 struct page, lru);
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001713 remove_hugetlb_page(h, page, acct_surplus);
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001714 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001715 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001716 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001717
Mike Kravetz10c6ec42021-05-04 18:35:03 -07001718 return page;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001719}
1720
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001721/*
1722 * Dissolve a given free hugepage into free buddy pages. This function does
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001723 * nothing for in-use hugepages and non-hugepages.
1724 * This function returns values like below:
1725 *
1726 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1727 * (allocated or reserved.)
1728 * 0: successfully dissolved free hugepages or the page is not a
1729 * hugepage (considered as already dissolved)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001730 */
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001731int dissolve_free_huge_page(struct page *page)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001732{
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001733 int rc = -EBUSY;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001734
Muchun Song7ffddd42021-02-04 18:32:06 -08001735retry:
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001736 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1737 if (!PageHuge(page))
1738 return 0;
1739
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001740 spin_lock_irq(&hugetlb_lock);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001741 if (!PageHuge(page)) {
1742 rc = 0;
1743 goto out;
1744 }
1745
1746 if (!page_count(page)) {
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001747 struct page *head = compound_head(page);
1748 struct hstate *h = page_hstate(head);
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001749 if (h->free_huge_pages - h->resv_huge_pages == 0)
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001750 goto out;
Muchun Song7ffddd42021-02-04 18:32:06 -08001751
1752 /*
1753 * We should make sure that the page is already on the free list
1754 * when it is dissolved.
1755 */
Mike Kravetz6c037142021-02-24 12:09:04 -08001756 if (unlikely(!HPageFreed(head))) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001757 spin_unlock_irq(&hugetlb_lock);
Muchun Song7ffddd42021-02-04 18:32:06 -08001758 cond_resched();
1759
1760 /*
1761 * Theoretically, we should return -EBUSY when we
1762 * encounter this race. In fact, we have a chance
1763 * to successfully dissolve the page if we do a
1764 * retry. Because the race window is quite small.
1765 * If we seize this opportunity, it is an optimization
1766 * for increasing the success rate of dissolving page.
1767 */
1768 goto retry;
1769 }
1770
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001771 /*
1772 * Move PageHWPoison flag from head page to the raw error page,
1773 * which makes any subpages rather than the error page reusable.
1774 */
1775 if (PageHWPoison(head) && page != head) {
1776 SetPageHWPoison(page);
1777 ClearPageHWPoison(head);
1778 }
Mike Kravetz6eb4e882021-05-04 18:34:55 -07001779 remove_hugetlb_page(h, page, false);
zhong jiangc1470b32016-08-11 15:32:55 -07001780 h->max_huge_pages--;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001781 spin_unlock_irq(&hugetlb_lock);
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001782 update_and_free_page(h, head);
Mike Kravetz11218282021-05-04 18:34:59 -07001783 return 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001784 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001785out:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001786 spin_unlock_irq(&hugetlb_lock);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001787 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001788}
1789
1790/*
1791 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1792 * make specified memory blocks removable from the system.
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001793 * Note that this will dissolve a free gigantic hugepage completely, if any
1794 * part of it lies within the given range.
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001795 * Also note that if dissolve_free_huge_page() returns with an error, all
1796 * free hugepages that were dissolved before that error are lost.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001797 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001798int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001799{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001800 unsigned long pfn;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001801 struct page *page;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001802 int rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001803
Li Zhongd0177632014-08-06 16:07:56 -07001804 if (!hugepages_supported())
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001805 return rc;
Li Zhongd0177632014-08-06 16:07:56 -07001806
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001807 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1808 page = pfn_to_page(pfn);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001809 rc = dissolve_free_huge_page(page);
1810 if (rc)
1811 break;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001812 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001813
1814 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001815}
1816
Michal Hockoab5ac902018-01-31 16:20:48 -08001817/*
1818 * Allocates a fresh surplus page from the page allocator.
1819 */
Michal Hocko0c397da2018-01-31 16:20:56 -08001820static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
Michal Hockoaaf14e42017-07-10 15:49:08 -07001821 int nid, nodemask_t *nmask)
Adam Litke7893d1d2007-10-16 01:26:18 -07001822{
Michal Hocko9980d742018-01-31 16:20:52 -08001823 struct page *page = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -07001824
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001825 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001826 return NULL;
1827
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001828 spin_lock_irq(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001829 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1830 goto out_unlock;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001831 spin_unlock_irq(&hugetlb_lock);
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001832
Mike Kravetzf60858f2019-09-23 15:37:35 -07001833 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hocko9980d742018-01-31 16:20:52 -08001834 if (!page)
Michal Hocko0c397da2018-01-31 16:20:56 -08001835 return NULL;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001836
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001837 spin_lock_irq(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001838 /*
1839 * We could have raced with the pool size change.
1840 * Double check that and simply deallocate the new page
1841 * if we would end up overcommiting the surpluses. Abuse
1842 * temporary page to workaround the nasty free_huge_page
1843 * codeflow
1844 */
1845 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
Mike Kravetz9157c3112021-02-24 12:09:00 -08001846 SetHPageTemporary(page);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001847 spin_unlock_irq(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001848 put_page(page);
Kai Shen2bf753e2019-05-13 17:15:37 -07001849 return NULL;
Michal Hocko9980d742018-01-31 16:20:52 -08001850 } else {
Michal Hocko9980d742018-01-31 16:20:52 -08001851 h->surplus_huge_pages++;
Michal Hocko4704dea2018-03-09 15:50:55 -08001852 h->surplus_huge_pages_node[page_to_nid(page)]++;
Adam Litke7893d1d2007-10-16 01:26:18 -07001853 }
Michal Hocko9980d742018-01-31 16:20:52 -08001854
1855out_unlock:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001856 spin_unlock_irq(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001857
1858 return page;
1859}
1860
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001861static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
Aneesh Kumar K.V9a4e9f32019-03-05 15:47:44 -08001862 int nid, nodemask_t *nmask)
Michal Hockoab5ac902018-01-31 16:20:48 -08001863{
1864 struct page *page;
1865
1866 if (hstate_is_gigantic(h))
1867 return NULL;
1868
Mike Kravetzf60858f2019-09-23 15:37:35 -07001869 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hockoab5ac902018-01-31 16:20:48 -08001870 if (!page)
1871 return NULL;
1872
1873 /*
1874 * We do not account these pages as surplus because they are only
1875 * temporary and will be released properly on the last reference
1876 */
Mike Kravetz9157c3112021-02-24 12:09:00 -08001877 SetHPageTemporary(page);
Michal Hockoab5ac902018-01-31 16:20:48 -08001878
1879 return page;
1880}
1881
Adam Litkee4e574b2007-10-16 01:26:19 -07001882/*
Dave Hansen099730d2015-11-05 18:50:17 -08001883 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1884 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001885static
Michal Hocko0c397da2018-01-31 16:20:56 -08001886struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
Dave Hansen099730d2015-11-05 18:50:17 -08001887 struct vm_area_struct *vma, unsigned long addr)
1888{
Michal Hockoaaf14e42017-07-10 15:49:08 -07001889 struct page *page;
1890 struct mempolicy *mpol;
1891 gfp_t gfp_mask = htlb_alloc_mask(h);
1892 int nid;
1893 nodemask_t *nodemask;
1894
1895 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
Michal Hocko0c397da2018-01-31 16:20:56 -08001896 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
Michal Hockoaaf14e42017-07-10 15:49:08 -07001897 mpol_cond_put(mpol);
1898
1899 return page;
Dave Hansen099730d2015-11-05 18:50:17 -08001900}
1901
Michal Hockoab5ac902018-01-31 16:20:48 -08001902/* page migration callback function */
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001903struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07001904 nodemask_t *nmask, gfp_t gfp_mask)
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001905{
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001906 spin_lock_irq(&hugetlb_lock);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001907 if (h->free_huge_pages - h->resv_huge_pages > 0) {
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001908 struct page *page;
1909
1910 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1911 if (page) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001912 spin_unlock_irq(&hugetlb_lock);
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001913 return page;
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001914 }
1915 }
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001916 spin_unlock_irq(&hugetlb_lock);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001917
Michal Hocko0c397da2018-01-31 16:20:56 -08001918 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001919}
1920
Michal Hockoebd63722018-01-31 16:21:00 -08001921/* mempolicy aware migration callback */
Michal Hocko389c8172018-01-31 16:21:03 -08001922struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1923 unsigned long address)
Michal Hockoebd63722018-01-31 16:21:00 -08001924{
1925 struct mempolicy *mpol;
1926 nodemask_t *nodemask;
1927 struct page *page;
Michal Hockoebd63722018-01-31 16:21:00 -08001928 gfp_t gfp_mask;
1929 int node;
1930
Michal Hockoebd63722018-01-31 16:21:00 -08001931 gfp_mask = htlb_alloc_mask(h);
1932 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07001933 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
Michal Hockoebd63722018-01-31 16:21:00 -08001934 mpol_cond_put(mpol);
1935
1936 return page;
1937}
1938
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001939/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001940 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07001941 * of size 'delta'.
1942 */
Liu Xiang0a4f3d12020-12-14 19:12:05 -08001943static int gather_surplus_pages(struct hstate *h, long delta)
Jules Irenge1b2a1e72020-04-06 20:08:09 -07001944 __must_hold(&hugetlb_lock)
Adam Litkee4e574b2007-10-16 01:26:19 -07001945{
1946 struct list_head surplus_list;
1947 struct page *page, *tmp;
Liu Xiang0a4f3d12020-12-14 19:12:05 -08001948 int ret;
1949 long i;
1950 long needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07001951 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07001952
Andi Kleena5516432008-07-23 21:27:41 -07001953 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001954 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07001955 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07001956 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08001957 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001958
1959 allocated = 0;
1960 INIT_LIST_HEAD(&surplus_list);
1961
1962 ret = -ENOMEM;
1963retry:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001964 spin_unlock_irq(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07001965 for (i = 0; i < needed; i++) {
Michal Hocko0c397da2018-01-31 16:20:56 -08001966 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
Michal Hockoaaf14e42017-07-10 15:49:08 -07001967 NUMA_NO_NODE, NULL);
Hillf Danton28073b02012-03-21 16:34:00 -07001968 if (!page) {
1969 alloc_ok = false;
1970 break;
1971 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001972 list_add(&page->lru, &surplus_list);
David Rientjes69ed7792017-07-10 15:48:50 -07001973 cond_resched();
Adam Litkee4e574b2007-10-16 01:26:19 -07001974 }
Hillf Danton28073b02012-03-21 16:34:00 -07001975 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07001976
1977 /*
1978 * After retaking hugetlb_lock, we need to recalculate 'needed'
1979 * because either resv_huge_pages or free_huge_pages may have changed.
1980 */
Mike Kravetzdb71ef72021-05-04 18:35:07 -07001981 spin_lock_irq(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001982 needed = (h->resv_huge_pages + delta) -
1983 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07001984 if (needed > 0) {
1985 if (alloc_ok)
1986 goto retry;
1987 /*
1988 * We were not able to allocate enough pages to
1989 * satisfy the entire reservation so we free what
1990 * we've allocated so far.
1991 */
1992 goto free;
1993 }
Adam Litkee4e574b2007-10-16 01:26:19 -07001994 /*
1995 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03001996 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07001997 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08001998 * allocator. Commit the entire reservation here to prevent another
1999 * process from stealing the pages as they are added to the pool but
2000 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07002001 */
2002 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07002003 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002004 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002005
Adam Litke19fc3f02008-04-28 02:12:20 -07002006 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07002007 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Muchun Songe5584642021-02-04 18:33:00 -08002008 int zeroed;
2009
Adam Litke19fc3f02008-04-28 02:12:20 -07002010 if ((--needed) < 0)
2011 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002012 /*
2013 * This page is now managed by the hugetlb allocator and has
2014 * no users -- drop the buddy allocator's reference.
2015 */
Muchun Songe5584642021-02-04 18:33:00 -08002016 zeroed = put_page_testzero(page);
2017 VM_BUG_ON_PAGE(!zeroed, page);
Andi Kleena5516432008-07-23 21:27:41 -07002018 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07002019 }
Hillf Danton28073b02012-03-21 16:34:00 -07002020free:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002021 spin_unlock_irq(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07002022
2023 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07002024 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2025 put_page(page);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002026 spin_lock_irq(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002027
2028 return ret;
2029}
2030
2031/*
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002032 * This routine has two main purposes:
2033 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2034 * in unused_resv_pages. This corresponds to the prior adjustments made
2035 * to the associated reservation map.
2036 * 2) Free any unused surplus pages that may have been allocated to satisfy
2037 * the reservation. As many as unused_resv_pages may be freed.
Adam Litkee4e574b2007-10-16 01:26:19 -07002038 */
Andi Kleena5516432008-07-23 21:27:41 -07002039static void return_unused_surplus_pages(struct hstate *h,
2040 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07002041{
Adam Litkee4e574b2007-10-16 01:26:19 -07002042 unsigned long nr_pages;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002043 struct page *page;
2044 LIST_HEAD(page_list);
2045
2046 /* Uncommit the reservation */
2047 h->resv_huge_pages -= unused_resv_pages;
Adam Litkee4e574b2007-10-16 01:26:19 -07002048
Andi Kleenaa888a72008-07-23 21:27:47 -07002049 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002050 if (hstate_is_gigantic(h))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002051 goto out;
Andi Kleenaa888a72008-07-23 21:27:47 -07002052
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002053 /*
2054 * Part (or even all) of the reservation could have been backed
2055 * by pre-allocated pages. Only free surplus pages.
2056 */
Andi Kleena5516432008-07-23 21:27:41 -07002057 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07002058
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002059 /*
2060 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002061 * evenly across all nodes with memory. Iterate across these nodes
2062 * until we can no longer free unreserved surplus pages. This occurs
2063 * when the nodes with surplus pages have no free pages.
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002064 * remove_pool_huge_page() will balance the freed pages across the
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002065 * on-line nodes with memory and will handle the hstate accounting.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002066 */
2067 while (nr_pages--) {
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002068 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2069 if (!page)
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002070 goto out;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002071
2072 list_add(&page->lru, &page_list);
Adam Litkee4e574b2007-10-16 01:26:19 -07002073 }
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002074
2075out:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002076 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002077 update_and_free_pages_bulk(h, &page_list);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002078 spin_lock_irq(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002079}
2080
Mike Kravetz5e911372015-09-08 15:01:28 -07002081
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002082/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002083 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07002084 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002085 *
2086 * vma_needs_reservation is called to determine if the huge page at addr
2087 * within the vma has an associated reservation. If a reservation is
2088 * needed, the value 1 is returned. The caller is then responsible for
2089 * managing the global reservation and subpool usage counts. After
2090 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002091 * to add the page to the reservation map. If the page allocation fails,
2092 * the reservation must be ended instead of committed. vma_end_reservation
2093 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002094 *
2095 * In the normal case, vma_commit_reservation returns the same value
2096 * as the preceding vma_needs_reservation call. The only time this
2097 * is not the case is if a reserve map was changed between calls. It
2098 * is the responsibility of the caller to notice the difference and
2099 * take appropriate action.
Mike Kravetz96b96a92016-11-10 10:46:32 -08002100 *
2101 * vma_add_reservation is used in error paths where a reservation must
2102 * be restored when a newly allocated huge page must be freed. It is
2103 * to be called after calling vma_needs_reservation to determine if a
2104 * reservation exists.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002105 */
Mike Kravetz5e911372015-09-08 15:01:28 -07002106enum vma_resv_mode {
2107 VMA_NEEDS_RESV,
2108 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002109 VMA_END_RESV,
Mike Kravetz96b96a92016-11-10 10:46:32 -08002110 VMA_ADD_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07002111};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002112static long __vma_reservation_common(struct hstate *h,
2113 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07002114 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002115{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002116 struct resv_map *resv;
2117 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002118 long ret;
Mina Almasry0db9d742020-04-01 21:11:25 -07002119 long dummy_out_regions_needed;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002120
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002121 resv = vma_resv_map(vma);
2122 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07002123 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002124
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002125 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07002126 switch (mode) {
2127 case VMA_NEEDS_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002128 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2129 /* We assume that vma_reservation_* routines always operate on
2130 * 1 page, and that adding to resv map a 1 page entry can only
2131 * ever require 1 region.
2132 */
2133 VM_BUG_ON(dummy_out_regions_needed != 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002134 break;
2135 case VMA_COMMIT_RESV:
Mina Almasry075a61d2020-04-01 21:11:28 -07002136 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002137 /* region_add calls of range 1 should never fail. */
2138 VM_BUG_ON(ret < 0);
Mike Kravetz5e911372015-09-08 15:01:28 -07002139 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002140 case VMA_END_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002141 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002142 ret = 0;
2143 break;
Mike Kravetz96b96a92016-11-10 10:46:32 -08002144 case VMA_ADD_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002145 if (vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07002146 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002147 /* region_add calls of range 1 should never fail. */
2148 VM_BUG_ON(ret < 0);
2149 } else {
2150 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz96b96a92016-11-10 10:46:32 -08002151 ret = region_del(resv, idx, idx + 1);
2152 }
2153 break;
Mike Kravetz5e911372015-09-08 15:01:28 -07002154 default:
2155 BUG();
2156 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002157
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002158 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002159 return ret;
Miaohe Linbf3d12b2021-05-04 18:34:32 -07002160 /*
2161 * We know private mapping must have HPAGE_RESV_OWNER set.
2162 *
2163 * In most cases, reserves always exist for private mappings.
2164 * However, a file associated with mapping could have been
2165 * hole punched or truncated after reserves were consumed.
2166 * As subsequent fault on such a range will not use reserves.
2167 * Subtle - The reserve map for private mappings has the
2168 * opposite meaning than that of shared mappings. If NO
2169 * entry is in the reserve map, it means a reservation exists.
2170 * If an entry exists in the reserve map, it means the
2171 * reservation has already been consumed. As a result, the
2172 * return value of this routine is the opposite of the
2173 * value returned from reserve map manipulation routines above.
2174 */
2175 if (ret > 0)
2176 return 0;
2177 if (ret == 0)
2178 return 1;
2179 return ret;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002180}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002181
2182static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07002183 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002184{
Mike Kravetz5e911372015-09-08 15:01:28 -07002185 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002186}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002187
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002188static long vma_commit_reservation(struct hstate *h,
2189 struct vm_area_struct *vma, unsigned long addr)
2190{
Mike Kravetz5e911372015-09-08 15:01:28 -07002191 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2192}
2193
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002194static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07002195 struct vm_area_struct *vma, unsigned long addr)
2196{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002197 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002198}
2199
Mike Kravetz96b96a92016-11-10 10:46:32 -08002200static long vma_add_reservation(struct hstate *h,
2201 struct vm_area_struct *vma, unsigned long addr)
2202{
2203 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2204}
2205
2206/*
2207 * This routine is called to restore a reservation on error paths. In the
2208 * specific error paths, a huge page was allocated (via alloc_huge_page)
2209 * and is about to be freed. If a reservation for the page existed,
Mike Kravetzd6995da2021-02-24 12:08:51 -08002210 * alloc_huge_page would have consumed the reservation and set
2211 * HPageRestoreReserve in the newly allocated page. When the page is freed
2212 * via free_huge_page, the global reservation count will be incremented if
2213 * HPageRestoreReserve is set. However, free_huge_page can not adjust the
2214 * reserve map. Adjust the reserve map here to be consistent with global
2215 * reserve count adjustments to be made by free_huge_page.
Mike Kravetz96b96a92016-11-10 10:46:32 -08002216 */
2217static void restore_reserve_on_error(struct hstate *h,
2218 struct vm_area_struct *vma, unsigned long address,
2219 struct page *page)
2220{
Mike Kravetzd6995da2021-02-24 12:08:51 -08002221 if (unlikely(HPageRestoreReserve(page))) {
Mike Kravetz96b96a92016-11-10 10:46:32 -08002222 long rc = vma_needs_reservation(h, vma, address);
2223
2224 if (unlikely(rc < 0)) {
2225 /*
2226 * Rare out of memory condition in reserve map
Mike Kravetzd6995da2021-02-24 12:08:51 -08002227 * manipulation. Clear HPageRestoreReserve so that
Mike Kravetz96b96a92016-11-10 10:46:32 -08002228 * global reserve count will not be incremented
2229 * by free_huge_page. This will make it appear
2230 * as though the reservation for this page was
2231 * consumed. This may prevent the task from
2232 * faulting in the page at a later time. This
2233 * is better than inconsistent global huge page
2234 * accounting of reserve counts.
2235 */
Mike Kravetzd6995da2021-02-24 12:08:51 -08002236 ClearHPageRestoreReserve(page);
Mike Kravetz96b96a92016-11-10 10:46:32 -08002237 } else if (rc) {
2238 rc = vma_add_reservation(h, vma, address);
2239 if (unlikely(rc < 0))
2240 /*
2241 * See above comment about rare out of
2242 * memory condition.
2243 */
Mike Kravetzd6995da2021-02-24 12:08:51 -08002244 ClearHPageRestoreReserve(page);
Mike Kravetz96b96a92016-11-10 10:46:32 -08002245 } else
2246 vma_end_reservation(h, vma, address);
2247 }
2248}
2249
Mike Kravetz70c35472015-09-08 15:01:54 -07002250struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002251 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252{
David Gibson90481622012-03-21 16:34:12 -07002253 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07002254 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08002255 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002256 long map_chg, map_commit;
2257 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002258 int ret, idx;
2259 struct hugetlb_cgroup *h_cg;
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002260 bool deferred_reserve;
Adam Litke2fc39ce2007-11-14 16:59:39 -08002261
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002262 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07002263 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002264 * Examine the region/reserve map to determine if the process
2265 * has a reservation for the page to be allocated. A return
2266 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07002267 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002268 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2269 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002270 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002271
2272 /*
2273 * Processes that did not create the mapping will have no
2274 * reserves as indicated by the region/reserve map. Check
2275 * that the allocation will not exceed the subpool limit.
2276 * Allocations for MAP_NORESERVE mappings also need to be
2277 * checked against any subpool limit.
2278 */
2279 if (map_chg || avoid_reserve) {
2280 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2281 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002282 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002283 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07002284 }
Mel Gormana1e78772008-07-23 21:27:23 -07002285
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002286 /*
2287 * Even though there was no reservation in the region/reserve
2288 * map, there could be reservations associated with the
2289 * subpool that can be used. This would be indicated if the
2290 * return value of hugepage_subpool_get_pages() is zero.
2291 * However, if avoid_reserve is specified we still avoid even
2292 * the subpool reservations.
2293 */
2294 if (avoid_reserve)
2295 gbl_chg = 1;
2296 }
2297
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002298 /* If this allocation is not consuming a reservation, charge it now.
2299 */
Miaohe Lin6501fe52021-05-04 18:33:16 -07002300 deferred_reserve = map_chg || avoid_reserve;
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002301 if (deferred_reserve) {
2302 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2303 idx, pages_per_huge_page(h), &h_cg);
2304 if (ret)
2305 goto out_subpool_put;
2306 }
2307
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002308 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002309 if (ret)
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002310 goto out_uncharge_cgroup_reservation;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002311
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002312 spin_lock_irq(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002313 /*
2314 * glb_chg is passed to indicate whether or not a page must be taken
2315 * from the global free pool (global change). gbl_chg == 0 indicates
2316 * a reservation exists for the allocation.
2317 */
2318 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002319 if (!page) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002320 spin_unlock_irq(&hugetlb_lock);
Michal Hocko0c397da2018-01-31 16:20:56 -08002321 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002322 if (!page)
2323 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002324 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
Mike Kravetzd6995da2021-02-24 12:08:51 -08002325 SetHPageRestoreReserve(page);
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002326 h->resv_huge_pages--;
2327 }
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002328 spin_lock_irq(&hugetlb_lock);
Wei Yang15a8d682020-10-13 16:56:33 -07002329 list_add(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002330 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07002331 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002332 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002333 /* If allocation is not consuming a reservation, also store the
2334 * hugetlb_cgroup pointer on the page.
2335 */
2336 if (deferred_reserve) {
2337 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2338 h_cg, page);
2339 }
2340
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002341 spin_unlock_irq(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07002342
Mike Kravetzd6995da2021-02-24 12:08:51 -08002343 hugetlb_set_page_subpool(page, spool);
Mel Gormana1e78772008-07-23 21:27:23 -07002344
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002345 map_commit = vma_commit_reservation(h, vma, addr);
2346 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07002347 /*
2348 * The page was added to the reservation map between
2349 * vma_needs_reservation and vma_commit_reservation.
2350 * This indicates a race with hugetlb_reserve_pages.
2351 * Adjust for the subpool count incremented above AND
2352 * in hugetlb_reserve_pages for the same page. Also,
2353 * the reservation count added in hugetlb_reserve_pages
2354 * no longer applies.
2355 */
2356 long rsv_adjust;
2357
2358 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2359 hugetlb_acct_memory(h, -rsv_adjust);
Mike Kravetz79aa9252020-11-01 17:07:27 -08002360 if (deferred_reserve)
2361 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2362 pages_per_huge_page(h), page);
Mike Kravetz33039672015-06-24 16:57:58 -07002363 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08002364 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002365
2366out_uncharge_cgroup:
2367 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002368out_uncharge_cgroup_reservation:
2369 if (deferred_reserve)
2370 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2371 h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002372out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002373 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002374 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002375 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002376 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08002377}
2378
Aneesh Kumar K.Ve24a1302017-07-28 10:31:25 +05302379int alloc_bootmem_huge_page(struct hstate *h)
2380 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2381int __alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002382{
2383 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002384 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002385
Joonsoo Kimb2261022013-09-11 14:21:00 -07002386 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002387 void *addr;
2388
Mike Rapoporteb31d552018-10-30 15:08:04 -07002389 addr = memblock_alloc_try_nid_raw(
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002390 huge_page_size(h), huge_page_size(h),
Mike Rapoport97ad1082018-10-30 15:09:44 -07002391 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002392 if (addr) {
2393 /*
2394 * Use the beginning of the huge page to store the
2395 * huge_bootmem_page struct (until gather_bootmem
2396 * puts them into the mem_map).
2397 */
2398 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002399 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002400 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002401 }
2402 return 0;
2403
2404found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002405 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002406 /* Put them into a private list first because mem_map is not up yet */
Cannon Matthews330d6e42018-08-17 15:49:17 -07002407 INIT_LIST_HEAD(&m->list);
Andi Kleenaa888a72008-07-23 21:27:47 -07002408 list_add(&m->list, &huge_boot_pages);
2409 m->hstate = h;
2410 return 1;
2411}
2412
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002413static void __init prep_compound_huge_page(struct page *page,
2414 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08002415{
2416 if (unlikely(order > (MAX_ORDER - 1)))
2417 prep_compound_gigantic_page(page, order);
2418 else
2419 prep_compound_page(page, order);
2420}
2421
Andi Kleenaa888a72008-07-23 21:27:47 -07002422/* Put bootmem huge pages into the standard lists after mem_map is up */
2423static void __init gather_bootmem_prealloc(void)
2424{
2425 struct huge_bootmem_page *m;
2426
2427 list_for_each_entry(m, &huge_boot_pages, list) {
Mike Kravetz40d18eb2018-08-17 15:49:07 -07002428 struct page *page = virt_to_page(m);
Andi Kleenaa888a72008-07-23 21:27:47 -07002429 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002430
Andi Kleenaa888a72008-07-23 21:27:47 -07002431 WARN_ON(page_count(page) != 1);
Miaohe Linc78a7f32021-02-24 12:07:01 -08002432 prep_compound_huge_page(page, huge_page_order(h));
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002433 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002434 prep_new_huge_page(h, page, page_to_nid(page));
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08002435 put_page(page); /* free it into the hugepage allocator */
2436
Rafael Aquinib0320c72011-06-15 15:08:39 -07002437 /*
2438 * If we had gigantic hugepages allocated at boot time, we need
2439 * to restore the 'stolen' pages to totalram_pages in order to
2440 * fix confusing memory reports from free(1) and another
2441 * side-effects, like CommitLimit going negative.
2442 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002443 if (hstate_is_gigantic(h))
Miaohe Linc78a7f32021-02-24 12:07:01 -08002444 adjust_managed_page_count(page, pages_per_huge_page(h));
Cannon Matthews520495f2018-07-03 17:02:43 -07002445 cond_resched();
Andi Kleenaa888a72008-07-23 21:27:47 -07002446 }
2447}
2448
Andi Kleen8faa8b02008-07-23 21:27:48 -07002449static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450{
2451 unsigned long i;
Mike Kravetzf60858f2019-09-23 15:37:35 -07002452 nodemask_t *node_alloc_noretry;
2453
2454 if (!hstate_is_gigantic(h)) {
2455 /*
2456 * Bit mask controlling how hard we retry per-node allocations.
2457 * Ignore errors as lower level routines can deal with
2458 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2459 * time, we are likely in bigger trouble.
2460 */
2461 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2462 GFP_KERNEL);
2463 } else {
2464 /* allocations done at boot time */
2465 node_alloc_noretry = NULL;
2466 }
2467
2468 /* bit mask controlling how hard we retry per-node allocations */
2469 if (node_alloc_noretry)
2470 nodes_clear(*node_alloc_noretry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471
Andi Kleene5ff2152008-07-23 21:27:42 -07002472 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002473 if (hstate_is_gigantic(h)) {
Barry Songdbda8fe2020-07-23 21:15:30 -07002474 if (hugetlb_cma_size) {
Roman Gushchincf11e852020-04-10 14:32:45 -07002475 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
Chen Wandun7ecc9562021-02-24 12:07:58 -08002476 goto free;
Roman Gushchincf11e852020-04-10 14:32:45 -07002477 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002478 if (!alloc_bootmem_huge_page(h))
2479 break;
Michal Hocko0c397da2018-01-31 16:20:56 -08002480 } else if (!alloc_pool_huge_page(h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07002481 &node_states[N_MEMORY],
2482 node_alloc_noretry))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 break;
David Rientjes69ed7792017-07-10 15:48:50 -07002484 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 }
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002486 if (i < h->max_huge_pages) {
2487 char buf[32];
2488
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002489 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002490 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2491 h->max_huge_pages, buf, i);
2492 h->max_huge_pages = i;
2493 }
Chen Wandun7ecc9562021-02-24 12:07:58 -08002494free:
Mike Kravetzf60858f2019-09-23 15:37:35 -07002495 kfree(node_alloc_noretry);
Andi Kleene5ff2152008-07-23 21:27:42 -07002496}
2497
2498static void __init hugetlb_init_hstates(void)
2499{
2500 struct hstate *h;
2501
2502 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002503 if (minimum_order > huge_page_order(h))
2504 minimum_order = huge_page_order(h);
2505
Andi Kleen8faa8b02008-07-23 21:27:48 -07002506 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002507 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002508 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002509 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002510 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002511}
2512
2513static void __init report_hugepages(void)
2514{
2515 struct hstate *h;
2516
2517 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002518 char buf[32];
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002519
2520 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Andrew Mortonffb22af2013-02-22 16:32:08 -08002521 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002522 buf, h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002523 }
2524}
2525
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002527static void try_to_free_low(struct hstate *h, unsigned long count,
2528 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002530 int i;
Mike Kravetz11218282021-05-04 18:34:59 -07002531 LIST_HEAD(page_list);
Christoph Lameter4415cc82006-09-25 23:31:55 -07002532
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002533 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002534 return;
2535
Mike Kravetz11218282021-05-04 18:34:59 -07002536 /*
2537 * Collect pages to be freed on a list, and free after dropping lock
2538 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002539 for_each_node_mask(i, *nodes_allowed) {
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002540 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002541 struct list_head *freel = &h->hugepage_freelists[i];
2542 list_for_each_entry_safe(page, next, freel, lru) {
2543 if (count >= h->nr_huge_pages)
Mike Kravetz11218282021-05-04 18:34:59 -07002544 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 if (PageHighMem(page))
2546 continue;
Mike Kravetz6eb4e882021-05-04 18:34:55 -07002547 remove_hugetlb_page(h, page, false);
Mike Kravetz11218282021-05-04 18:34:59 -07002548 list_add(&page->lru, &page_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 }
2550 }
Mike Kravetz11218282021-05-04 18:34:59 -07002551
2552out:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002553 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002554 update_and_free_pages_bulk(h, &page_list);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002555 spin_lock_irq(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556}
2557#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002558static inline void try_to_free_low(struct hstate *h, unsigned long count,
2559 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560{
2561}
2562#endif
2563
Wu Fengguang20a03072009-06-16 15:32:22 -07002564/*
2565 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2566 * balanced by operating on them in a round-robin fashion.
2567 * Returns 1 if an adjustment was made.
2568 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002569static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2570 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002571{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002572 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002573
2574 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002575
Joonsoo Kimb2261022013-09-11 14:21:00 -07002576 if (delta < 0) {
2577 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2578 if (h->surplus_huge_pages_node[node])
2579 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002580 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002581 } else {
2582 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2583 if (h->surplus_huge_pages_node[node] <
2584 h->nr_huge_pages_node[node])
2585 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002586 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002587 }
2588 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002589
Joonsoo Kimb2261022013-09-11 14:21:00 -07002590found:
2591 h->surplus_huge_pages += delta;
2592 h->surplus_huge_pages_node[node] += delta;
2593 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002594}
2595
Andi Kleena5516432008-07-23 21:27:41 -07002596#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002597static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002598 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599{
Adam Litke7893d1d2007-10-16 01:26:18 -07002600 unsigned long min_count, ret;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002601 struct page *page;
2602 LIST_HEAD(page_list);
Mike Kravetzf60858f2019-09-23 15:37:35 -07002603 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2604
2605 /*
2606 * Bit mask controlling how hard we retry per-node allocations.
2607 * If we can not allocate the bit mask, do not attempt to allocate
2608 * the requested huge pages.
2609 */
2610 if (node_alloc_noretry)
2611 nodes_clear(*node_alloc_noretry);
2612 else
2613 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614
Mike Kravetz29383962021-05-04 18:34:52 -07002615 /*
2616 * resize_lock mutex prevents concurrent adjustments to number of
2617 * pages in hstate via the proc/sysfs interfaces.
2618 */
2619 mutex_lock(&h->resize_lock);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002620 spin_lock_irq(&hugetlb_lock);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002621
2622 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002623 * Check for a node specific request.
2624 * Changing node specific huge page count may require a corresponding
2625 * change to the global count. In any case, the passed node mask
2626 * (nodes_allowed) will restrict alloc/free to the specified node.
2627 */
2628 if (nid != NUMA_NO_NODE) {
2629 unsigned long old_count = count;
2630
2631 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2632 /*
2633 * User may have specified a large count value which caused the
2634 * above calculation to overflow. In this case, they wanted
2635 * to allocate as many huge pages as possible. Set count to
2636 * largest possible value to align with their intention.
2637 */
2638 if (count < old_count)
2639 count = ULONG_MAX;
2640 }
2641
2642 /*
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002643 * Gigantic pages runtime allocation depend on the capability for large
2644 * page range allocation.
2645 * If the system does not provide this feature, return an error when
2646 * the user tries to allocate gigantic pages but let the user free the
2647 * boottime allocated gigantic pages.
2648 */
2649 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2650 if (count > persistent_huge_pages(h)) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002651 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz29383962021-05-04 18:34:52 -07002652 mutex_unlock(&h->resize_lock);
Mike Kravetzf60858f2019-09-23 15:37:35 -07002653 NODEMASK_FREE(node_alloc_noretry);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002654 return -EINVAL;
2655 }
2656 /* Fall through to decrease pool */
2657 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002658
Adam Litke7893d1d2007-10-16 01:26:18 -07002659 /*
2660 * Increase the pool size
2661 * First take pages out of surplus state. Then make up the
2662 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002663 *
Michal Hocko0c397da2018-01-31 16:20:56 -08002664 * We might race with alloc_surplus_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002665 * to convert a surplus huge page to a normal huge page. That is
2666 * not critical, though, it just means the overall size of the
2667 * pool might be one hugepage larger than it needs to be, but
2668 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002669 */
Andi Kleena5516432008-07-23 21:27:41 -07002670 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002671 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002672 break;
2673 }
2674
Andi Kleena5516432008-07-23 21:27:41 -07002675 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002676 /*
2677 * If this allocation races such that we no longer need the
2678 * page, free_huge_page will handle it by freeing the page
2679 * and reducing the surplus.
2680 */
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002681 spin_unlock_irq(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07002682
2683 /* yield cpu to avoid soft lockup */
2684 cond_resched();
2685
Mike Kravetzf60858f2019-09-23 15:37:35 -07002686 ret = alloc_pool_huge_page(h, nodes_allowed,
2687 node_alloc_noretry);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002688 spin_lock_irq(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07002689 if (!ret)
2690 goto out;
2691
Mel Gorman536240f22009-12-14 17:59:56 -08002692 /* Bail for signals. Probably ctrl-c from user */
2693 if (signal_pending(current))
2694 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002695 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002696
2697 /*
2698 * Decrease the pool size
2699 * First return free pages to the buddy allocator (being careful
2700 * to keep enough around to satisfy reservations). Then place
2701 * pages into surplus state as needed so the pool will shrink
2702 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002703 *
2704 * By placing pages into the surplus state independent of the
2705 * overcommit value, we are allowing the surplus pool size to
2706 * exceed overcommit. There are few sane options here. Since
Michal Hocko0c397da2018-01-31 16:20:56 -08002707 * alloc_surplus_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002708 * though, we'll note that we're not allowed to exceed surplus
2709 * and won't grow the pool anywhere else. Not until one of the
2710 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002711 */
Andi Kleena5516432008-07-23 21:27:41 -07002712 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002713 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002714 try_to_free_low(h, min_count, nodes_allowed);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002715
2716 /*
2717 * Collect pages to be removed on list without dropping lock
2718 */
Andi Kleena5516432008-07-23 21:27:41 -07002719 while (min_count < persistent_huge_pages(h)) {
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002720 page = remove_pool_huge_page(h, nodes_allowed, 0);
2721 if (!page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 break;
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002723
2724 list_add(&page->lru, &page_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 }
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002726 /* free the pages after dropping lock */
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002727 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002728 update_and_free_pages_bulk(h, &page_list);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002729 spin_lock_irq(&hugetlb_lock);
Mike Kravetz10c6ec42021-05-04 18:35:03 -07002730
Andi Kleena5516432008-07-23 21:27:41 -07002731 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002732 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002733 break;
2734 }
2735out:
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002736 h->max_huge_pages = persistent_huge_pages(h);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002737 spin_unlock_irq(&hugetlb_lock);
Mike Kravetz29383962021-05-04 18:34:52 -07002738 mutex_unlock(&h->resize_lock);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002739
Mike Kravetzf60858f2019-09-23 15:37:35 -07002740 NODEMASK_FREE(node_alloc_noretry);
2741
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002742 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743}
2744
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002745#define HSTATE_ATTR_RO(_name) \
2746 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2747
2748#define HSTATE_ATTR(_name) \
2749 static struct kobj_attribute _name##_attr = \
2750 __ATTR(_name, 0644, _name##_show, _name##_store)
2751
2752static struct kobject *hugepages_kobj;
2753static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2754
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002755static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2756
2757static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002758{
2759 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002760
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002761 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002762 if (hstate_kobjs[i] == kobj) {
2763 if (nidp)
2764 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002765 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002766 }
2767
2768 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002769}
2770
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002771static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002772 struct kobj_attribute *attr, char *buf)
2773{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002774 struct hstate *h;
2775 unsigned long nr_huge_pages;
2776 int nid;
2777
2778 h = kobj_to_hstate(kobj, &nid);
2779 if (nid == NUMA_NO_NODE)
2780 nr_huge_pages = h->nr_huge_pages;
2781 else
2782 nr_huge_pages = h->nr_huge_pages_node[nid];
2783
Joe Perchesae7a9272020-12-14 19:14:42 -08002784 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002785}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002786
David Rientjes238d3c12014-08-06 16:06:51 -07002787static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2788 struct hstate *h, int nid,
2789 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002790{
2791 int err;
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002792 nodemask_t nodes_allowed, *n_mask;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002793
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002794 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2795 return -EINVAL;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002796
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002797 if (nid == NUMA_NO_NODE) {
2798 /*
2799 * global hstate attribute
2800 */
2801 if (!(obey_mempolicy &&
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002802 init_nodemask_of_mempolicy(&nodes_allowed)))
2803 n_mask = &node_states[N_MEMORY];
2804 else
2805 n_mask = &nodes_allowed;
2806 } else {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002807 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002808 * Node specific request. count adjustment happens in
2809 * set_max_huge_pages() after acquiring hugetlb_lock.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002810 */
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002811 init_nodemask_of_node(&nodes_allowed, nid);
2812 n_mask = &nodes_allowed;
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002813 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002814
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002815 err = set_max_huge_pages(h, count, nid, n_mask);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002816
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002817 return err ? err : len;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002818}
2819
David Rientjes238d3c12014-08-06 16:06:51 -07002820static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2821 struct kobject *kobj, const char *buf,
2822 size_t len)
2823{
2824 struct hstate *h;
2825 unsigned long count;
2826 int nid;
2827 int err;
2828
2829 err = kstrtoul(buf, 10, &count);
2830 if (err)
2831 return err;
2832
2833 h = kobj_to_hstate(kobj, &nid);
2834 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2835}
2836
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002837static ssize_t nr_hugepages_show(struct kobject *kobj,
2838 struct kobj_attribute *attr, char *buf)
2839{
2840 return nr_hugepages_show_common(kobj, attr, buf);
2841}
2842
2843static ssize_t nr_hugepages_store(struct kobject *kobj,
2844 struct kobj_attribute *attr, const char *buf, size_t len)
2845{
David Rientjes238d3c12014-08-06 16:06:51 -07002846 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002847}
2848HSTATE_ATTR(nr_hugepages);
2849
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002850#ifdef CONFIG_NUMA
2851
2852/*
2853 * hstate attribute for optionally mempolicy-based constraint on persistent
2854 * huge page alloc/free.
2855 */
2856static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
Joe Perchesae7a9272020-12-14 19:14:42 -08002857 struct kobj_attribute *attr,
2858 char *buf)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002859{
2860 return nr_hugepages_show_common(kobj, attr, buf);
2861}
2862
2863static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2864 struct kobj_attribute *attr, const char *buf, size_t len)
2865{
David Rientjes238d3c12014-08-06 16:06:51 -07002866 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002867}
2868HSTATE_ATTR(nr_hugepages_mempolicy);
2869#endif
2870
2871
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002872static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2873 struct kobj_attribute *attr, char *buf)
2874{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002875 struct hstate *h = kobj_to_hstate(kobj, NULL);
Joe Perchesae7a9272020-12-14 19:14:42 -08002876 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002877}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002878
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002879static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2880 struct kobj_attribute *attr, const char *buf, size_t count)
2881{
2882 int err;
2883 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002884 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002885
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002886 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002887 return -EINVAL;
2888
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002889 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002890 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002891 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002892
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002893 spin_lock_irq(&hugetlb_lock);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002894 h->nr_overcommit_huge_pages = input;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07002895 spin_unlock_irq(&hugetlb_lock);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002896
2897 return count;
2898}
2899HSTATE_ATTR(nr_overcommit_hugepages);
2900
2901static ssize_t free_hugepages_show(struct kobject *kobj,
2902 struct kobj_attribute *attr, char *buf)
2903{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002904 struct hstate *h;
2905 unsigned long free_huge_pages;
2906 int nid;
2907
2908 h = kobj_to_hstate(kobj, &nid);
2909 if (nid == NUMA_NO_NODE)
2910 free_huge_pages = h->free_huge_pages;
2911 else
2912 free_huge_pages = h->free_huge_pages_node[nid];
2913
Joe Perchesae7a9272020-12-14 19:14:42 -08002914 return sysfs_emit(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002915}
2916HSTATE_ATTR_RO(free_hugepages);
2917
2918static ssize_t resv_hugepages_show(struct kobject *kobj,
2919 struct kobj_attribute *attr, char *buf)
2920{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002921 struct hstate *h = kobj_to_hstate(kobj, NULL);
Joe Perchesae7a9272020-12-14 19:14:42 -08002922 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002923}
2924HSTATE_ATTR_RO(resv_hugepages);
2925
2926static ssize_t surplus_hugepages_show(struct kobject *kobj,
2927 struct kobj_attribute *attr, char *buf)
2928{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002929 struct hstate *h;
2930 unsigned long surplus_huge_pages;
2931 int nid;
2932
2933 h = kobj_to_hstate(kobj, &nid);
2934 if (nid == NUMA_NO_NODE)
2935 surplus_huge_pages = h->surplus_huge_pages;
2936 else
2937 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2938
Joe Perchesae7a9272020-12-14 19:14:42 -08002939 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002940}
2941HSTATE_ATTR_RO(surplus_hugepages);
2942
2943static struct attribute *hstate_attrs[] = {
2944 &nr_hugepages_attr.attr,
2945 &nr_overcommit_hugepages_attr.attr,
2946 &free_hugepages_attr.attr,
2947 &resv_hugepages_attr.attr,
2948 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002949#ifdef CONFIG_NUMA
2950 &nr_hugepages_mempolicy_attr.attr,
2951#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002952 NULL,
2953};
2954
Arvind Yadav67e5ed92017-09-06 16:22:06 -07002955static const struct attribute_group hstate_attr_group = {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002956 .attrs = hstate_attrs,
2957};
2958
Jeff Mahoney094e9532010-02-02 13:44:14 -08002959static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2960 struct kobject **hstate_kobjs,
Arvind Yadav67e5ed92017-09-06 16:22:06 -07002961 const struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002962{
2963 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002964 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002965
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002966 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2967 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002968 return -ENOMEM;
2969
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002970 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Miaohe Lincc2205a2021-02-24 12:06:50 -08002971 if (retval) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002972 kobject_put(hstate_kobjs[hi]);
Miaohe Lincc2205a2021-02-24 12:06:50 -08002973 hstate_kobjs[hi] = NULL;
2974 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002975
2976 return retval;
2977}
2978
2979static void __init hugetlb_sysfs_init(void)
2980{
2981 struct hstate *h;
2982 int err;
2983
2984 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2985 if (!hugepages_kobj)
2986 return;
2987
2988 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002989 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2990 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002991 if (err)
Mike Kravetz282f4212020-06-03 16:00:46 -07002992 pr_err("HugeTLB: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002993 }
2994}
2995
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002996#ifdef CONFIG_NUMA
2997
2998/*
2999 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08003000 * with node devices in node_devices[] using a parallel array. The array
3001 * index of a node device or _hstate == node id.
3002 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003003 * the base kernel, on the hugetlb module.
3004 */
3005struct node_hstate {
3006 struct kobject *hugepages_kobj;
3007 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3008};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08003009static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003010
3011/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003012 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003013 */
3014static struct attribute *per_node_hstate_attrs[] = {
3015 &nr_hugepages_attr.attr,
3016 &free_hugepages_attr.attr,
3017 &surplus_hugepages_attr.attr,
3018 NULL,
3019};
3020
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003021static const struct attribute_group per_node_hstate_attr_group = {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003022 .attrs = per_node_hstate_attrs,
3023};
3024
3025/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003026 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003027 * Returns node id via non-NULL nidp.
3028 */
3029static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3030{
3031 int nid;
3032
3033 for (nid = 0; nid < nr_node_ids; nid++) {
3034 struct node_hstate *nhs = &node_hstates[nid];
3035 int i;
3036 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3037 if (nhs->hstate_kobjs[i] == kobj) {
3038 if (nidp)
3039 *nidp = nid;
3040 return &hstates[i];
3041 }
3042 }
3043
3044 BUG();
3045 return NULL;
3046}
3047
3048/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003049 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003050 * No-op if no hstate attributes attached.
3051 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003052static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003053{
3054 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003055 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003056
3057 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003058 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003059
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003060 for_each_hstate(h) {
3061 int idx = hstate_index(h);
3062 if (nhs->hstate_kobjs[idx]) {
3063 kobject_put(nhs->hstate_kobjs[idx]);
3064 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003065 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003066 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003067
3068 kobject_put(nhs->hugepages_kobj);
3069 nhs->hugepages_kobj = NULL;
3070}
3071
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003072
3073/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003074 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003075 * No-op if attributes already registered.
3076 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003077static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003078{
3079 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003080 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003081 int err;
3082
3083 if (nhs->hugepages_kobj)
3084 return; /* already allocated */
3085
3086 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08003087 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003088 if (!nhs->hugepages_kobj)
3089 return;
3090
3091 for_each_hstate(h) {
3092 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3093 nhs->hstate_kobjs,
3094 &per_node_hstate_attr_group);
3095 if (err) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003096 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003097 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003098 hugetlb_unregister_node(node);
3099 break;
3100 }
3101 }
3102}
3103
3104/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003105 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08003106 * devices of nodes that have memory. All on-line nodes should have
3107 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003108 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08003109static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003110{
3111 int nid;
3112
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08003113 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08003114 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08003115 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003116 hugetlb_register_node(node);
3117 }
3118
3119 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003120 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003121 * [un]register hstate attributes on node hotplug.
3122 */
3123 register_hugetlbfs_with_node(hugetlb_register_node,
3124 hugetlb_unregister_node);
3125}
3126#else /* !CONFIG_NUMA */
3127
3128static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3129{
3130 BUG();
3131 if (nidp)
3132 *nidp = -1;
3133 return NULL;
3134}
3135
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003136static void hugetlb_register_all_nodes(void) { }
3137
3138#endif
3139
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003140static int __init hugetlb_init(void)
3141{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003142 int i;
3143
Mike Kravetzd6995da2021-02-24 12:08:51 -08003144 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
3145 __NR_HPAGEFLAGS);
3146
Mike Kravetzc2833a52020-06-03 16:00:50 -07003147 if (!hugepages_supported()) {
3148 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3149 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07003150 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003151 }
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08003152
Mike Kravetz282f4212020-06-03 16:00:46 -07003153 /*
3154 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3155 * architectures depend on setup being done here.
3156 */
3157 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3158 if (!parsed_default_hugepagesz) {
3159 /*
3160 * If we did not parse a default huge page size, set
3161 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3162 * number of huge pages for this default size was implicitly
3163 * specified, set that here as well.
3164 * Note that the implicit setting will overwrite an explicit
3165 * setting. A warning will be printed in this case.
3166 */
3167 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3168 if (default_hstate_max_huge_pages) {
3169 if (default_hstate.max_huge_pages) {
3170 char buf[32];
Andi Kleenaa888a72008-07-23 21:27:47 -07003171
Mike Kravetz282f4212020-06-03 16:00:46 -07003172 string_get_size(huge_page_size(&default_hstate),
3173 1, STRING_UNITS_2, buf, 32);
3174 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3175 default_hstate.max_huge_pages, buf);
3176 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3177 default_hstate_max_huge_pages);
3178 }
3179 default_hstate.max_huge_pages =
3180 default_hstate_max_huge_pages;
3181 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003182 }
Andi Kleenaa888a72008-07-23 21:27:47 -07003183
Roman Gushchincf11e852020-04-10 14:32:45 -07003184 hugetlb_cma_check();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003185 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07003186 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003187 report_hugepages();
3188
3189 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003190 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08003191 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003192
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003193#ifdef CONFIG_SMP
3194 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3195#else
3196 num_fault_mutexes = 1;
3197#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003198 hugetlb_fault_mutex_table =
Kees Cook6da2ec52018-06-12 13:55:00 -07003199 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3200 GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003201 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003202
3203 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003204 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003205 return 0;
3206}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08003207subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003208
Mike Kravetzae94da82020-06-03 16:00:34 -07003209/* Overwritten by architectures with more huge page sizes */
3210bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003211{
Mike Kravetzae94da82020-06-03 16:00:34 -07003212 return size == HPAGE_SIZE;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003213}
3214
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08003215void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003216{
3217 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003218 unsigned long i;
3219
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003220 if (size_to_hstate(PAGE_SIZE << order)) {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003221 return;
3222 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003223 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003224 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003225 h = &hstates[hugetlb_max_hstate++];
Mike Kravetz29383962021-05-04 18:34:52 -07003226 mutex_init(&h->resize_lock);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003227 h->order = order;
Miaohe Linaca78302021-02-24 12:07:46 -08003228 h->mask = ~(huge_page_size(h) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003229 for (i = 0; i < MAX_NUMNODES; ++i)
3230 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07003231 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07003232 h->next_nid_to_alloc = first_memory_node;
3233 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003234 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3235 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003236
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003237 parsed_hstate = h;
3238}
3239
Mike Kravetz282f4212020-06-03 16:00:46 -07003240/*
3241 * hugepages command line processing
3242 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3243 * specification. If not, ignore the hugepages value. hugepages can also
3244 * be the first huge page command line option in which case it implicitly
3245 * specifies the number of huge pages for the default size.
3246 */
3247static int __init hugepages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003248{
3249 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003250 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003251
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003252 if (!parsed_valid_hugepagesz) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003253 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003254 parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -07003255 return 0;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003256 }
Mike Kravetz282f4212020-06-03 16:00:46 -07003257
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003258 /*
Mike Kravetz282f4212020-06-03 16:00:46 -07003259 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3260 * yet, so this hugepages= parameter goes to the "default hstate".
3261 * Otherwise, it goes with the previously parsed hugepagesz or
3262 * default_hugepagesz.
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003263 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003264 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003265 mhp = &default_hstate_max_huge_pages;
3266 else
3267 mhp = &parsed_hstate->max_huge_pages;
3268
Andi Kleen8faa8b02008-07-23 21:27:48 -07003269 if (mhp == last_mhp) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003270 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3271 return 0;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003272 }
3273
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003274 if (sscanf(s, "%lu", mhp) <= 0)
3275 *mhp = 0;
3276
Andi Kleen8faa8b02008-07-23 21:27:48 -07003277 /*
3278 * Global state is always initialized later in hugetlb_init.
Miaohe Lin04adbc32021-05-04 18:33:22 -07003279 * But we need to allocate gigantic hstates here early to still
Andi Kleen8faa8b02008-07-23 21:27:48 -07003280 * use the bootmem allocator.
3281 */
Miaohe Lin04adbc32021-05-04 18:33:22 -07003282 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
Andi Kleen8faa8b02008-07-23 21:27:48 -07003283 hugetlb_hstate_alloc_pages(parsed_hstate);
3284
3285 last_mhp = mhp;
3286
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003287 return 1;
3288}
Mike Kravetz282f4212020-06-03 16:00:46 -07003289__setup("hugepages=", hugepages_setup);
Nick Piggine11bfbf2008-07-23 21:27:52 -07003290
Mike Kravetz282f4212020-06-03 16:00:46 -07003291/*
3292 * hugepagesz command line processing
3293 * A specific huge page size can only be specified once with hugepagesz.
3294 * hugepagesz is followed by hugepages on the command line. The global
3295 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3296 * hugepagesz argument was valid.
3297 */
Mike Kravetz359f2542020-06-03 16:00:38 -07003298static int __init hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003299{
Mike Kravetz359f2542020-06-03 16:00:38 -07003300 unsigned long size;
Mike Kravetz282f4212020-06-03 16:00:46 -07003301 struct hstate *h;
3302
3303 parsed_valid_hugepagesz = false;
Mike Kravetz359f2542020-06-03 16:00:38 -07003304 size = (unsigned long)memparse(s, NULL);
3305
3306 if (!arch_hugetlb_valid_size(size)) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003307 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
Mike Kravetz359f2542020-06-03 16:00:38 -07003308 return 0;
3309 }
3310
Mike Kravetz282f4212020-06-03 16:00:46 -07003311 h = size_to_hstate(size);
3312 if (h) {
3313 /*
3314 * hstate for this size already exists. This is normally
3315 * an error, but is allowed if the existing hstate is the
3316 * default hstate. More specifically, it is only allowed if
3317 * the number of huge pages for the default hstate was not
3318 * previously specified.
3319 */
3320 if (!parsed_default_hugepagesz || h != &default_hstate ||
3321 default_hstate.max_huge_pages) {
3322 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3323 return 0;
3324 }
3325
3326 /*
3327 * No need to call hugetlb_add_hstate() as hstate already
3328 * exists. But, do set parsed_hstate so that a following
3329 * hugepages= parameter will be applied to this hstate.
3330 */
3331 parsed_hstate = h;
3332 parsed_valid_hugepagesz = true;
3333 return 1;
Mike Kravetz38237832020-06-03 16:00:42 -07003334 }
3335
Mike Kravetz359f2542020-06-03 16:00:38 -07003336 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
Mike Kravetz282f4212020-06-03 16:00:46 -07003337 parsed_valid_hugepagesz = true;
Nick Piggine11bfbf2008-07-23 21:27:52 -07003338 return 1;
3339}
Mike Kravetz359f2542020-06-03 16:00:38 -07003340__setup("hugepagesz=", hugepagesz_setup);
3341
Mike Kravetz282f4212020-06-03 16:00:46 -07003342/*
3343 * default_hugepagesz command line input
3344 * Only one instance of default_hugepagesz allowed on command line.
3345 */
Mike Kravetzae94da82020-06-03 16:00:34 -07003346static int __init default_hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003347{
Mike Kravetzae94da82020-06-03 16:00:34 -07003348 unsigned long size;
3349
Mike Kravetz282f4212020-06-03 16:00:46 -07003350 parsed_valid_hugepagesz = false;
Mike Kravetz282f4212020-06-03 16:00:46 -07003351 if (parsed_default_hugepagesz) {
3352 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3353 return 0;
3354 }
3355
3356 size = (unsigned long)memparse(s, NULL);
3357
3358 if (!arch_hugetlb_valid_size(size)) {
3359 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3360 return 0;
3361 }
3362
3363 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3364 parsed_valid_hugepagesz = true;
3365 parsed_default_hugepagesz = true;
3366 default_hstate_idx = hstate_index(size_to_hstate(size));
3367
3368 /*
3369 * The number of default huge pages (for this size) could have been
3370 * specified as the first hugetlb parameter: hugepages=X. If so,
3371 * then default_hstate_max_huge_pages is set. If the default huge
3372 * page size is gigantic (>= MAX_ORDER), then the pages must be
3373 * allocated here from bootmem allocator.
3374 */
3375 if (default_hstate_max_huge_pages) {
3376 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3377 if (hstate_is_gigantic(&default_hstate))
3378 hugetlb_hstate_alloc_pages(&default_hstate);
3379 default_hstate_max_huge_pages = 0;
3380 }
3381
Nick Piggine11bfbf2008-07-23 21:27:52 -07003382 return 1;
3383}
Mike Kravetzae94da82020-06-03 16:00:34 -07003384__setup("default_hugepagesz=", default_hugepagesz_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003385
Muchun Song8ca39e62020-08-11 18:30:32 -07003386static unsigned int allowed_mems_nr(struct hstate *h)
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003387{
3388 int node;
3389 unsigned int nr = 0;
Muchun Song8ca39e62020-08-11 18:30:32 -07003390 nodemask_t *mpol_allowed;
3391 unsigned int *array = h->free_huge_pages_node;
3392 gfp_t gfp_mask = htlb_alloc_mask(h);
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003393
Muchun Song8ca39e62020-08-11 18:30:32 -07003394 mpol_allowed = policy_nodemask_current(gfp_mask);
3395
3396 for_each_node_mask(node, cpuset_current_mems_allowed) {
Jiapeng Zhongc93b0a92021-02-24 12:07:09 -08003397 if (!mpol_allowed || node_isset(node, *mpol_allowed))
Muchun Song8ca39e62020-08-11 18:30:32 -07003398 nr += array[node];
3399 }
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003400
3401 return nr;
3402}
3403
3404#ifdef CONFIG_SYSCTL
Muchun Song17743792020-09-04 16:36:13 -07003405static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3406 void *buffer, size_t *length,
3407 loff_t *ppos, unsigned long *out)
3408{
3409 struct ctl_table dup_table;
3410
3411 /*
3412 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3413 * can duplicate the @table and alter the duplicate of it.
3414 */
3415 dup_table = *table;
3416 dup_table.data = out;
3417
3418 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3419}
3420
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003421static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3422 struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003423 void *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424{
Andi Kleene5ff2152008-07-23 21:27:42 -07003425 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07003426 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08003427 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003428
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003429 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003430 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003431
Muchun Song17743792020-09-04 16:36:13 -07003432 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3433 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003434 if (ret)
3435 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003436
David Rientjes238d3c12014-08-06 16:06:51 -07003437 if (write)
3438 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3439 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08003440out:
3441 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442}
Mel Gorman396faf02007-07-17 04:03:13 -07003443
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003444int hugetlb_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003445 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003446{
3447
3448 return hugetlb_sysctl_handler_common(false, table, write,
3449 buffer, length, ppos);
3450}
3451
3452#ifdef CONFIG_NUMA
3453int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003454 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003455{
3456 return hugetlb_sysctl_handler_common(true, table, write,
3457 buffer, length, ppos);
3458}
3459#endif /* CONFIG_NUMA */
3460
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003461int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003462 void *buffer, size_t *length, loff_t *ppos)
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003463{
Andi Kleena5516432008-07-23 21:27:41 -07003464 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07003465 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08003466 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003467
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003468 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003469 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003470
Petr Holasekc033a932011-03-22 16:33:05 -07003471 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07003472
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07003473 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08003474 return -EINVAL;
3475
Muchun Song17743792020-09-04 16:36:13 -07003476 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3477 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003478 if (ret)
3479 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003480
3481 if (write) {
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003482 spin_lock_irq(&hugetlb_lock);
Andi Kleene5ff2152008-07-23 21:27:42 -07003483 h->nr_overcommit_huge_pages = tmp;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003484 spin_unlock_irq(&hugetlb_lock);
Andi Kleene5ff2152008-07-23 21:27:42 -07003485 }
Michal Hocko08d4a242011-01-13 15:47:26 -08003486out:
3487 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003488}
3489
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490#endif /* CONFIG_SYSCTL */
3491
Alexey Dobriyane1759c22008-10-15 23:50:22 +04003492void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493{
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003494 struct hstate *h;
3495 unsigned long total = 0;
3496
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003497 if (!hugepages_supported())
3498 return;
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003499
3500 for_each_hstate(h) {
3501 unsigned long count = h->nr_huge_pages;
3502
Miaohe Linaca78302021-02-24 12:07:46 -08003503 total += huge_page_size(h) * count;
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003504
3505 if (h == &default_hstate)
3506 seq_printf(m,
3507 "HugePages_Total: %5lu\n"
3508 "HugePages_Free: %5lu\n"
3509 "HugePages_Rsvd: %5lu\n"
3510 "HugePages_Surp: %5lu\n"
3511 "Hugepagesize: %8lu kB\n",
3512 count,
3513 h->free_huge_pages,
3514 h->resv_huge_pages,
3515 h->surplus_huge_pages,
Miaohe Linaca78302021-02-24 12:07:46 -08003516 huge_page_size(h) / SZ_1K);
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003517 }
3518
Miaohe Linaca78302021-02-24 12:07:46 -08003519 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520}
3521
Joe Perches79815932020-09-16 13:40:43 -07003522int hugetlb_report_node_meminfo(char *buf, int len, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523{
Andi Kleena5516432008-07-23 21:27:41 -07003524 struct hstate *h = &default_hstate;
Joe Perches79815932020-09-16 13:40:43 -07003525
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003526 if (!hugepages_supported())
3527 return 0;
Joe Perches79815932020-09-16 13:40:43 -07003528
3529 return sysfs_emit_at(buf, len,
3530 "Node %d HugePages_Total: %5u\n"
3531 "Node %d HugePages_Free: %5u\n"
3532 "Node %d HugePages_Surp: %5u\n",
3533 nid, h->nr_huge_pages_node[nid],
3534 nid, h->free_huge_pages_node[nid],
3535 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536}
3537
David Rientjes949f7ec2013-04-29 15:07:48 -07003538void hugetlb_show_meminfo(void)
3539{
3540 struct hstate *h;
3541 int nid;
3542
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003543 if (!hugepages_supported())
3544 return;
3545
David Rientjes949f7ec2013-04-29 15:07:48 -07003546 for_each_node_state(nid, N_MEMORY)
3547 for_each_hstate(h)
3548 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3549 nid,
3550 h->nr_huge_pages_node[nid],
3551 h->free_huge_pages_node[nid],
3552 h->surplus_huge_pages_node[nid],
Miaohe Linaca78302021-02-24 12:07:46 -08003553 huge_page_size(h) / SZ_1K);
David Rientjes949f7ec2013-04-29 15:07:48 -07003554}
3555
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003556void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3557{
3558 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3559 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3560}
3561
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3563unsigned long hugetlb_total_pages(void)
3564{
Wanpeng Lid0028582013-03-22 15:04:40 -07003565 struct hstate *h;
3566 unsigned long nr_total_pages = 0;
3567
3568 for_each_hstate(h)
3569 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3570 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572
Andi Kleena5516432008-07-23 21:27:41 -07003573static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003574{
3575 int ret = -ENOMEM;
3576
Miaohe Lin0aa7f352021-02-24 12:06:57 -08003577 if (!delta)
3578 return 0;
3579
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003580 spin_lock_irq(&hugetlb_lock);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003581 /*
3582 * When cpuset is configured, it breaks the strict hugetlb page
3583 * reservation as the accounting is done on a global variable. Such
3584 * reservation is completely rubbish in the presence of cpuset because
3585 * the reservation is not checked against page availability for the
3586 * current cpuset. Application can still potentially OOM'ed by kernel
3587 * with lack of free htlb page in cpuset that the task is in.
3588 * Attempt to enforce strict accounting with cpuset is almost
3589 * impossible (or too ugly) because cpuset is too fluid that
3590 * task or memory node can be dynamically moved between cpusets.
3591 *
3592 * The change of semantics for shared hugetlb mapping with cpuset is
3593 * undesirable. However, in order to preserve some of the semantics,
3594 * we fall back to check against current free page availability as
3595 * a best attempt and hopefully to minimize the impact of changing
3596 * semantics that cpuset has.
Muchun Song8ca39e62020-08-11 18:30:32 -07003597 *
3598 * Apart from cpuset, we also have memory policy mechanism that
3599 * also determines from which node the kernel will allocate memory
3600 * in a NUMA system. So similar to cpuset, we also should consider
3601 * the memory policy of the current task. Similar to the description
3602 * above.
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003603 */
3604 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07003605 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003606 goto out;
3607
Muchun Song8ca39e62020-08-11 18:30:32 -07003608 if (delta > allowed_mems_nr(h)) {
Andi Kleena5516432008-07-23 21:27:41 -07003609 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003610 goto out;
3611 }
3612 }
3613
3614 ret = 0;
3615 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07003616 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003617
3618out:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07003619 spin_unlock_irq(&hugetlb_lock);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003620 return ret;
3621}
3622
Andy Whitcroft84afd992008-07-23 21:27:32 -07003623static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3624{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003625 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003626
3627 /*
3628 * This new VMA should share its siblings reservation map if present.
3629 * The VMA will only ever have a valid reservation map pointer where
3630 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003631 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07003632 * after this open call completes. It is therefore safe to take a
3633 * new reference here without additional locking.
3634 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003635 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003636 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003637}
3638
Mel Gormana1e78772008-07-23 21:27:23 -07003639static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3640{
Andi Kleena5516432008-07-23 21:27:41 -07003641 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003642 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07003643 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003644 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003645 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003646
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003647 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3648 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003649
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003650 start = vma_hugecache_offset(h, vma, vma->vm_start);
3651 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003652
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003653 reserve = (end - start) - region_count(resv, start, end);
Mina Almasrye9fe92a2020-04-01 21:11:21 -07003654 hugetlb_cgroup_uncharge_counter(resv, start, end);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003655 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003656 /*
3657 * Decrement reserve counts. The global reserve count may be
3658 * adjusted if the subpool has a minimum size.
3659 */
3660 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3661 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003662 }
Mina Almasrye9fe92a2020-04-01 21:11:21 -07003663
3664 kref_put(&resv->refs, resv_map_release);
Mel Gormana1e78772008-07-23 21:27:23 -07003665}
3666
Dan Williams31383c62017-11-29 16:10:28 -08003667static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3668{
3669 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3670 return -EINVAL;
3671 return 0;
3672}
3673
Dan Williams05ea8862018-04-05 16:24:25 -07003674static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3675{
Miaohe Linaca78302021-02-24 12:07:46 -08003676 return huge_page_size(hstate_vma(vma));
Dan Williams05ea8862018-04-05 16:24:25 -07003677}
3678
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679/*
3680 * We cannot handle pagefaults against hugetlb pages at all. They cause
3681 * handle_mm_fault() to try to instantiate regular-sized pages in the
Miaohe Lin6c26d312021-02-24 12:07:19 -08003682 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683 * this far.
3684 */
Souptick Joarderb3ec9f32018-06-07 17:08:04 -07003685static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686{
3687 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003688 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689}
3690
Jane Chueec36362018-08-02 15:36:05 -07003691/*
3692 * When a new function is introduced to vm_operations_struct and added
3693 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3694 * This is because under System V memory model, mappings created via
3695 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3696 * their original vm_ops are overwritten with shm_vm_ops.
3697 */
Alexey Dobriyanf0f37e2f2009-09-27 22:29:37 +04003698const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003699 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003700 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003701 .close = hugetlb_vm_op_close,
Dmitry Safonovdd3b6142020-12-14 19:08:17 -08003702 .may_split = hugetlb_vm_op_split,
Dan Williams05ea8862018-04-05 16:24:25 -07003703 .pagesize = hugetlb_vm_op_pagesize,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704};
3705
David Gibson1e8f8892006-01-06 00:10:44 -08003706static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3707 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003708{
3709 pte_t entry;
3710
David Gibson1e8f8892006-01-06 00:10:44 -08003711 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003712 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3713 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003714 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003715 entry = huge_pte_wrprotect(mk_huge_pte(page,
3716 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003717 }
3718 entry = pte_mkyoung(entry);
3719 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003720 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003721
3722 return entry;
3723}
3724
David Gibson1e8f8892006-01-06 00:10:44 -08003725static void set_huge_ptep_writable(struct vm_area_struct *vma,
3726 unsigned long address, pte_t *ptep)
3727{
3728 pte_t entry;
3729
Gerald Schaefer106c9922013-04-29 15:07:23 -07003730 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003731 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003732 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003733}
3734
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003735bool is_hugetlb_entry_migration(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003736{
3737 swp_entry_t swp;
3738
3739 if (huge_pte_none(pte) || pte_present(pte))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003740 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003741 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07003742 if (is_migration_entry(swp))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003743 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003744 else
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003745 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003746}
3747
Baoquan He3e5c3602020-10-13 16:56:10 -07003748static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003749{
3750 swp_entry_t swp;
3751
3752 if (huge_pte_none(pte) || pte_present(pte))
Baoquan He3e5c3602020-10-13 16:56:10 -07003753 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003754 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07003755 if (is_hwpoison_entry(swp))
Baoquan He3e5c3602020-10-13 16:56:10 -07003756 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003757 else
Baoquan He3e5c3602020-10-13 16:56:10 -07003758 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003759}
David Gibson1e8f8892006-01-06 00:10:44 -08003760
Peter Xu4eae4ef2021-03-12 21:07:33 -08003761static void
3762hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
3763 struct page *new_page)
3764{
3765 __SetPageUptodate(new_page);
3766 set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1));
3767 hugepage_add_new_anon_rmap(new_page, vma, addr);
3768 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
3769 ClearHPageRestoreReserve(new_page);
3770 SetHPageMigratable(new_page);
3771}
3772
David Gibson63551ae2005-06-21 17:14:44 -07003773int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3774 struct vm_area_struct *vma)
3775{
Mike Kravetz5e415402018-11-16 15:08:04 -08003776 pte_t *src_pte, *dst_pte, entry, dst_entry;
David Gibson63551ae2005-06-21 17:14:44 -07003777 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003778 unsigned long addr;
Peter Xuca6eb142021-03-12 21:07:30 -08003779 bool cow = is_cow_mapping(vma->vm_flags);
Andi Kleena5516432008-07-23 21:27:41 -07003780 struct hstate *h = hstate_vma(vma);
3781 unsigned long sz = huge_page_size(h);
Peter Xu4eae4ef2021-03-12 21:07:33 -08003782 unsigned long npages = pages_per_huge_page(h);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003783 struct address_space *mapping = vma->vm_file->f_mapping;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003784 struct mmu_notifier_range range;
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003785 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003786
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003787 if (cow) {
Jérôme Glisse7269f992019-05-13 17:20:53 -07003788 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07003789 vma->vm_start,
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003790 vma->vm_end);
3791 mmu_notifier_invalidate_range_start(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003792 } else {
3793 /*
3794 * For shared mappings i_mmap_rwsem must be held to call
3795 * huge_pte_alloc, otherwise the returned ptep could go
3796 * away if part of a shared pmd and another thread calls
3797 * huge_pmd_unshare.
3798 */
3799 i_mmap_lock_read(mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003800 }
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003801
Andi Kleena5516432008-07-23 21:27:41 -07003802 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003803 spinlock_t *src_ptl, *dst_ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07003804 src_pte = huge_pte_offset(src, addr, sz);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003805 if (!src_pte)
3806 continue;
Peter Xuaec44e02021-05-04 18:33:00 -07003807 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003808 if (!dst_pte) {
3809 ret = -ENOMEM;
3810 break;
3811 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003812
Mike Kravetz5e415402018-11-16 15:08:04 -08003813 /*
3814 * If the pagetables are shared don't copy or take references.
3815 * dst_pte == src_pte is the common case of src/dest sharing.
3816 *
3817 * However, src could have 'unshared' and dst shares with
3818 * another vma. If dst_pte !none, this implies sharing.
3819 * Check here before taking page table lock, and once again
3820 * after taking the lock below.
3821 */
3822 dst_entry = huge_ptep_get(dst_pte);
3823 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
Larry Woodmanc5c99422008-01-24 05:49:25 -08003824 continue;
3825
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003826 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3827 src_ptl = huge_pte_lockptr(h, src, src_pte);
3828 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003829 entry = huge_ptep_get(src_pte);
Mike Kravetz5e415402018-11-16 15:08:04 -08003830 dst_entry = huge_ptep_get(dst_pte);
Peter Xu4eae4ef2021-03-12 21:07:33 -08003831again:
Mike Kravetz5e415402018-11-16 15:08:04 -08003832 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3833 /*
3834 * Skip if src entry none. Also, skip in the
3835 * unlikely case dst entry !none as this implies
3836 * sharing with another vma.
3837 */
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003838 ;
3839 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3840 is_hugetlb_entry_hwpoisoned(entry))) {
3841 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3842
3843 if (is_write_migration_entry(swp_entry) && cow) {
3844 /*
3845 * COW mappings require pages in both
3846 * parent and child to be set to read.
3847 */
3848 make_migration_entry_read(&swp_entry);
3849 entry = swp_entry_to_pte(swp_entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07003850 set_huge_swap_pte_at(src, addr, src_pte,
3851 entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003852 }
Punit Agrawale5251fd2017-07-06 15:39:50 -07003853 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003854 } else {
Peter Xu4eae4ef2021-03-12 21:07:33 -08003855 entry = huge_ptep_get(src_pte);
3856 ptepage = pte_page(entry);
3857 get_page(ptepage);
3858
3859 /*
3860 * This is a rare case where we see pinned hugetlb
3861 * pages while they're prone to COW. We need to do the
3862 * COW earlier during fork.
3863 *
3864 * When pre-allocating the page or copying data, we
3865 * need to be without the pgtable locks since we could
3866 * sleep during the process.
3867 */
3868 if (unlikely(page_needs_cow_for_dma(vma, ptepage))) {
3869 pte_t src_pte_old = entry;
3870 struct page *new;
3871
3872 spin_unlock(src_ptl);
3873 spin_unlock(dst_ptl);
3874 /* Do not use reserve as it's private owned */
3875 new = alloc_huge_page(vma, addr, 1);
3876 if (IS_ERR(new)) {
3877 put_page(ptepage);
3878 ret = PTR_ERR(new);
3879 break;
3880 }
3881 copy_user_huge_page(new, ptepage, addr, vma,
3882 npages);
3883 put_page(ptepage);
3884
3885 /* Install the new huge page if src pte stable */
3886 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3887 src_ptl = huge_pte_lockptr(h, src, src_pte);
3888 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3889 entry = huge_ptep_get(src_pte);
3890 if (!pte_same(src_pte_old, entry)) {
3891 put_page(new);
3892 /* dst_entry won't change as in child */
3893 goto again;
3894 }
3895 hugetlb_install_page(vma, dst_pte, addr, new);
3896 spin_unlock(src_ptl);
3897 spin_unlock(dst_ptl);
3898 continue;
3899 }
3900
Joerg Roedel34ee6452014-11-13 13:46:09 +11003901 if (cow) {
Jérôme Glisse0f108512017-11-15 17:34:07 -08003902 /*
3903 * No need to notify as we are downgrading page
3904 * table protection not changing it to point
3905 * to a new page.
3906 *
Mike Rapoportad56b732018-03-21 21:22:47 +02003907 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08003908 */
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003909 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003910 }
Peter Xu4eae4ef2021-03-12 21:07:33 -08003911
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003912 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003913 set_huge_pte_at(dst, addr, dst_pte, entry);
Peter Xu4eae4ef2021-03-12 21:07:33 -08003914 hugetlb_count_add(npages, dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003915 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003916 spin_unlock(src_ptl);
3917 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003918 }
David Gibson63551ae2005-06-21 17:14:44 -07003919
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003920 if (cow)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003921 mmu_notifier_invalidate_range_end(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003922 else
3923 i_mmap_unlock_read(mapping);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003924
3925 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003926}
3927
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003928void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3929 unsigned long start, unsigned long end,
3930 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003931{
3932 struct mm_struct *mm = vma->vm_mm;
3933 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003934 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003935 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003936 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003937 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003938 struct hstate *h = hstate_vma(vma);
3939 unsigned long sz = huge_page_size(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003940 struct mmu_notifier_range range;
Andi Kleena5516432008-07-23 21:27:41 -07003941
David Gibson63551ae2005-06-21 17:14:44 -07003942 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003943 BUG_ON(start & ~huge_page_mask(h));
3944 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003945
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08003946 /*
3947 * This is a hugetlb vma, all the pte entries should point
3948 * to huge page.
3949 */
Peter Zijlstraed6a7932018-08-31 14:46:08 +02003950 tlb_change_page_size(tlb, sz);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003951 tlb_start_vma(tlb, vma);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07003952
3953 /*
3954 * If sharing possible, alert mmu notifiers of worst case.
3955 */
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07003956 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3957 end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003958 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3959 mmu_notifier_invalidate_range_start(&range);
Hillf Danton569f48b82014-12-10 15:44:41 -08003960 address = start;
Hillf Danton569f48b82014-12-10 15:44:41 -08003961 for (; address < end; address += sz) {
Punit Agrawal7868a202017-07-06 15:39:42 -07003962 ptep = huge_pte_offset(mm, address, sz);
Adam Litke4c887262005-10-29 18:16:46 -07003963 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003964 continue;
3965
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003966 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae2042020-08-11 18:31:38 -07003967 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003968 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07003969 /*
3970 * We just unmapped a page of PMDs by clearing a PUD.
3971 * The caller's TLB flush range should cover this area.
3972 */
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003973 continue;
3974 }
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003975
Hillf Danton66293262012-03-23 15:01:48 -07003976 pte = huge_ptep_get(ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003977 if (huge_pte_none(pte)) {
3978 spin_unlock(ptl);
3979 continue;
3980 }
Hillf Danton66293262012-03-23 15:01:48 -07003981
3982 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003983 * Migrating hugepage or HWPoisoned hugepage is already
3984 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003985 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003986 if (unlikely(!pte_present(pte))) {
Punit Agrawal9386fac2017-07-06 15:39:46 -07003987 huge_pte_clear(mm, address, ptep, sz);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003988 spin_unlock(ptl);
3989 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003990 }
Hillf Danton66293262012-03-23 15:01:48 -07003991
3992 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003993 /*
3994 * If a reference page is supplied, it is because a specific
3995 * page is being unmapped, not a range. Ensure the page we
3996 * are about to unmap is the actual page of interest.
3997 */
3998 if (ref_page) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003999 if (page != ref_page) {
4000 spin_unlock(ptl);
4001 continue;
4002 }
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004003 /*
4004 * Mark the VMA as having unmapped its page so that
4005 * future faults in this VMA will fail rather than
4006 * looking like data was lost
4007 */
4008 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4009 }
4010
David Gibsonc7546f82005-08-05 11:59:35 -07004011 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.Vb528e4b2016-12-12 16:42:37 -08004012 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07004013 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08004014 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07004015
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08004016 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004017 page_remove_rmap(page, true);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004018
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004019 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07004020 tlb_remove_page_size(tlb, page, huge_page_size(h));
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004021 /*
4022 * Bail out after unmapping reference page if supplied
4023 */
4024 if (ref_page)
4025 break;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07004026 }
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004027 mmu_notifier_invalidate_range_end(&range);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004028 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029}
David Gibson63551ae2005-06-21 17:14:44 -07004030
Mel Gormand8333522012-07-31 16:46:20 -07004031void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4032 struct vm_area_struct *vma, unsigned long start,
4033 unsigned long end, struct page *ref_page)
4034{
4035 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4036
4037 /*
4038 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4039 * test will fail on a vma being torn down, and not grab a page table
4040 * on its way out. We're lucky that the flag has such an appropriate
4041 * name, and can in fact be safely cleared here. We could clear it
4042 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004043 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07004044 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004045 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07004046 */
4047 vma->vm_flags &= ~VM_MAYSHARE;
4048}
4049
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004050void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004051 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004052{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004053 struct mmu_gather tlb;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004054
Will Deacona72afd82021-01-27 23:53:45 +00004055 tlb_gather_mmu(&tlb, vma->vm_mm);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004056 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
Will Deaconae8eba82021-01-27 23:53:43 +00004057 tlb_finish_mmu(&tlb);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004058}
4059
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004060/*
4061 * This is called when the original mapper is failing to COW a MAP_PRIVATE
Zhiyuan Dai578b7722021-02-24 12:07:26 -08004062 * mapping it owns the reserve page for. The intention is to unmap the page
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004063 * from other VMAs and let the children be SIGKILLed if they are faulting the
4064 * same region.
4065 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004066static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4067 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004068{
Adam Litke75266742008-11-12 13:24:56 -08004069 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004070 struct vm_area_struct *iter_vma;
4071 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004072 pgoff_t pgoff;
4073
4074 /*
4075 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4076 * from page cache lookup which is in HPAGE_SIZE units.
4077 */
Adam Litke75266742008-11-12 13:24:56 -08004078 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07004079 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4080 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05004081 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004082
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08004083 /*
4084 * Take the mapping lock for the duration of the table walk. As
4085 * this mapping should be shared between all the VMAs,
4086 * __unmap_hugepage_range() is called as the lock is already held
4087 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004088 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07004089 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004090 /* Do not unmap the current VMA */
4091 if (iter_vma == vma)
4092 continue;
4093
4094 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07004095 * Shared VMAs have their own reserves and do not affect
4096 * MAP_PRIVATE accounting but it is possible that a shared
4097 * VMA is using the same page so check and skip such VMAs.
4098 */
4099 if (iter_vma->vm_flags & VM_MAYSHARE)
4100 continue;
4101
4102 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004103 * Unmap the page from other VMAs without their own reserves.
4104 * They get marked to be SIGKILLed if they fault in these
4105 * areas. This is because a future no-page fault on this VMA
4106 * could insert a zeroed page instead of the data existing
4107 * from the time of fork. This would look like data corruption
4108 */
4109 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004110 unmap_hugepage_range(iter_vma, address,
4111 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004112 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004113 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004114}
4115
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004116/*
4117 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08004118 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4119 * cannot race with other handlers or page migration.
4120 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004121 */
Souptick Joarder2b740302018-08-23 17:01:36 -07004122static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Huang Ying974e6d62018-08-17 15:45:57 -07004123 unsigned long address, pte_t *ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004124 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08004125{
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004126 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07004127 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08004128 struct page *old_page, *new_page;
Souptick Joarder2b740302018-08-23 17:01:36 -07004129 int outside_reserve = 0;
4130 vm_fault_t ret = 0;
Huang Ying974e6d62018-08-17 15:45:57 -07004131 unsigned long haddr = address & huge_page_mask(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004132 struct mmu_notifier_range range;
David Gibson1e8f8892006-01-06 00:10:44 -08004133
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004134 pte = huge_ptep_get(ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08004135 old_page = pte_page(pte);
4136
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004137retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08004138 /* If no-one else is actually using this page, avoid the copy
4139 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07004140 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07004141 page_move_anon_rmap(old_page, vma);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004142 set_huge_ptep_writable(vma, haddr, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004143 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004144 }
4145
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004146 /*
4147 * If the process that created a MAP_PRIVATE mapping is about to
4148 * perform a COW due to a shared page count, attempt to satisfy
4149 * the allocation without using the existing reserves. The pagecache
4150 * page is used to determine if the reserve at this address was
4151 * consumed or not. If reserves were used, a partial faulted mapping
4152 * at the time of fork() could consume its reserves on COW instead
4153 * of the full address range.
4154 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07004155 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004156 old_page != pagecache_page)
4157 outside_reserve = 1;
4158
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004159 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004160
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004161 /*
4162 * Drop page table lock as buddy allocator may be called. It will
4163 * be acquired again before returning to the caller, as expected.
4164 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004165 spin_unlock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004166 new_page = alloc_huge_page(vma, haddr, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08004167
Adam Litke2fc39ce2007-11-14 16:59:39 -08004168 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004169 /*
4170 * If a process owning a MAP_PRIVATE mapping fails to COW,
4171 * it is due to references held by a child and an insufficient
4172 * huge page pool. To guarantee the original mappers
4173 * reliability, unmap the page from child processes. The child
4174 * may get SIGKILLed if it later faults.
4175 */
4176 if (outside_reserve) {
Mike Kravetze7dd91c2020-12-29 15:14:25 -08004177 struct address_space *mapping = vma->vm_file->f_mapping;
4178 pgoff_t idx;
4179 u32 hash;
4180
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004181 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004182 BUG_ON(huge_pte_none(pte));
Mike Kravetze7dd91c2020-12-29 15:14:25 -08004183 /*
4184 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4185 * unmapping. unmapping needs to hold i_mmap_rwsem
4186 * in write mode. Dropping i_mmap_rwsem in read mode
4187 * here is OK as COW mappings do not interact with
4188 * PMD sharing.
4189 *
4190 * Reacquire both after unmap operation.
4191 */
4192 idx = vma_hugecache_offset(h, vma, haddr);
4193 hash = hugetlb_fault_mutex_hash(mapping, idx);
4194 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4195 i_mmap_unlock_read(mapping);
4196
Huang Ying5b7a1d42018-08-17 15:45:53 -07004197 unmap_ref_private(mm, vma, old_page, haddr);
Mike Kravetze7dd91c2020-12-29 15:14:25 -08004198
4199 i_mmap_lock_read(mapping);
4200 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004201 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004202 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004203 if (likely(ptep &&
4204 pte_same(huge_ptep_get(ptep), pte)))
4205 goto retry_avoidcopy;
4206 /*
4207 * race occurs while re-acquiring page table
4208 * lock, and our job is done.
4209 */
4210 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004211 }
4212
Souptick Joarder2b740302018-08-23 17:01:36 -07004213 ret = vmf_error(PTR_ERR(new_page));
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004214 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08004215 }
4216
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004217 /*
4218 * When the original hugepage is shared one, it does not have
4219 * anon_vma prepared.
4220 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07004221 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004222 ret = VM_FAULT_OOM;
4223 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07004224 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004225
Huang Ying974e6d62018-08-17 15:45:57 -07004226 copy_user_huge_page(new_page, old_page, address, vma,
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004227 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004228 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004229
Jérôme Glisse7269f992019-05-13 17:20:53 -07004230 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07004231 haddr + huge_page_size(h));
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004232 mmu_notifier_invalidate_range_start(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004233
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004234 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004235 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004236 * before the page tables are altered
4237 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004238 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004239 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07004240 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Mike Kravetzd6995da2021-02-24 12:08:51 -08004241 ClearHPageRestoreReserve(new_page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07004242
David Gibson1e8f8892006-01-06 00:10:44 -08004243 /* Break COW */
Huang Ying5b7a1d42018-08-17 15:45:53 -07004244 huge_ptep_clear_flush(vma, haddr, ptep);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004245 mmu_notifier_invalidate_range(mm, range.start, range.end);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004246 set_huge_pte_at(mm, haddr, ptep,
David Gibson1e8f8892006-01-06 00:10:44 -08004247 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004248 page_remove_rmap(old_page, true);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004249 hugepage_add_new_anon_rmap(new_page, vma, haddr);
Mike Kravetz8f251a32021-02-24 12:08:56 -08004250 SetHPageMigratable(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004251 /* Make the old page be freed below */
4252 new_page = old_page;
4253 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004254 spin_unlock(ptl);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004255 mmu_notifier_invalidate_range_end(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004256out_release_all:
Huang Ying5b7a1d42018-08-17 15:45:53 -07004257 restore_reserve_on_error(h, vma, haddr, new_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004258 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004259out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004260 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07004261
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004262 spin_lock(ptl); /* Caller expects lock to be held */
4263 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08004264}
4265
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004266/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07004267static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4268 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004269{
4270 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07004271 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004272
4273 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07004274 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004275
4276 return find_lock_page(mapping, idx);
4277}
4278
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004279/*
4280 * Return whether there is a pagecache page to back given address within VMA.
4281 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4282 */
4283static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004284 struct vm_area_struct *vma, unsigned long address)
4285{
4286 struct address_space *mapping;
4287 pgoff_t idx;
4288 struct page *page;
4289
4290 mapping = vma->vm_file->f_mapping;
4291 idx = vma_hugecache_offset(h, vma, address);
4292
4293 page = find_get_page(mapping, idx);
4294 if (page)
4295 put_page(page);
4296 return page != NULL;
4297}
4298
Mike Kravetzab76ad52015-09-08 15:01:50 -07004299int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4300 pgoff_t idx)
4301{
4302 struct inode *inode = mapping->host;
4303 struct hstate *h = hstate_inode(inode);
4304 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4305
4306 if (err)
4307 return err;
Mike Kravetzd6995da2021-02-24 12:08:51 -08004308 ClearHPageRestoreReserve(page);
Mike Kravetzab76ad52015-09-08 15:01:50 -07004309
Mike Kravetz22146c32018-10-26 15:10:58 -07004310 /*
4311 * set page dirty so that it will not be removed from cache/file
4312 * by non-hugetlbfs specific code paths.
4313 */
4314 set_page_dirty(page);
4315
Mike Kravetzab76ad52015-09-08 15:01:50 -07004316 spin_lock(&inode->i_lock);
4317 inode->i_blocks += blocks_per_huge_page(h);
4318 spin_unlock(&inode->i_lock);
4319 return 0;
4320}
4321
Souptick Joarder2b740302018-08-23 17:01:36 -07004322static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4323 struct vm_area_struct *vma,
4324 struct address_space *mapping, pgoff_t idx,
4325 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004326{
Andi Kleena5516432008-07-23 21:27:41 -07004327 struct hstate *h = hstate_vma(vma);
Souptick Joarder2b740302018-08-23 17:01:36 -07004328 vm_fault_t ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08004329 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004330 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07004331 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08004332 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004333 spinlock_t *ptl;
Huang Ying285b8dc2018-06-07 17:08:08 -07004334 unsigned long haddr = address & huge_page_mask(h);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004335 bool new_page = false;
Adam Litke4c887262005-10-29 18:16:46 -07004336
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004337 /*
4338 * Currently, we are forced to kill the process in the event the
4339 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004340 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004341 */
4342 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08004343 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08004344 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004345 return ret;
4346 }
4347
Adam Litke4c887262005-10-29 18:16:46 -07004348 /*
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004349 * We can not race with truncation due to holding i_mmap_rwsem.
4350 * i_size is modified when holding i_mmap_rwsem, so check here
4351 * once for faults beyond end of file.
Adam Litke4c887262005-10-29 18:16:46 -07004352 */
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004353 size = i_size_read(mapping->host) >> huge_page_shift(h);
4354 if (idx >= size)
4355 goto out;
4356
Christoph Lameter6bda6662006-01-06 00:10:49 -08004357retry:
4358 page = find_lock_page(mapping, idx);
4359 if (!page) {
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004360 /*
4361 * Check for page in userfault range
4362 */
4363 if (userfaultfd_missing(vma)) {
4364 u32 hash;
4365 struct vm_fault vmf = {
4366 .vma = vma,
Huang Ying285b8dc2018-06-07 17:08:08 -07004367 .address = haddr,
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004368 .flags = flags,
4369 /*
4370 * Hard to debug if it ends up being
4371 * used by a callee that assumes
4372 * something about the other
4373 * uninitialized fields... same as in
4374 * memory.c
4375 */
4376 };
4377
4378 /*
Mike Kravetzc0d03812020-04-01 21:11:05 -07004379 * hugetlb_fault_mutex and i_mmap_rwsem must be
4380 * dropped before handling userfault. Reacquire
4381 * after handling fault to make calling code simpler.
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004382 */
Wei Yang188b04a2019-11-30 17:57:02 -08004383 hash = hugetlb_fault_mutex_hash(mapping, idx);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004384 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004385 i_mmap_unlock_read(mapping);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004386 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004387 i_mmap_lock_read(mapping);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004388 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4389 goto out;
4390 }
4391
Huang Ying285b8dc2018-06-07 17:08:08 -07004392 page = alloc_huge_page(vma, haddr, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08004393 if (IS_ERR(page)) {
Mike Kravetz4643d672019-08-13 15:38:00 -07004394 /*
4395 * Returning error will result in faulting task being
4396 * sent SIGBUS. The hugetlb fault mutex prevents two
4397 * tasks from racing to fault in the same page which
4398 * could result in false unable to allocate errors.
4399 * Page migration does not take the fault mutex, but
4400 * does a clear then write of pte's under page table
4401 * lock. Page fault code could race with migration,
4402 * notice the clear pte and try to allocate a page
4403 * here. Before returning error, get ptl and make
4404 * sure there really is no pte entry.
4405 */
4406 ptl = huge_pte_lock(h, mm, ptep);
Miaohe Lind83e6c8a2021-05-04 18:33:31 -07004407 ret = 0;
4408 if (huge_pte_none(huge_ptep_get(ptep)))
4409 ret = vmf_error(PTR_ERR(page));
Mike Kravetz4643d672019-08-13 15:38:00 -07004410 spin_unlock(ptl);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004411 goto out;
4412 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004413 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004414 __SetPageUptodate(page);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004415 new_page = true;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004416
Mel Gormanf83a2752009-05-28 14:34:40 -07004417 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07004418 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004419 if (err) {
4420 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004421 if (err == -EEXIST)
4422 goto retry;
4423 goto out;
4424 }
Mel Gorman23be7462010-04-23 13:17:56 -04004425 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08004426 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004427 if (unlikely(anon_vma_prepare(vma))) {
4428 ret = VM_FAULT_OOM;
4429 goto backout_unlocked;
4430 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08004431 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04004432 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004433 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004434 /*
4435 * If memory error occurs between mmap() and fault, some process
4436 * don't have hwpoisoned swap entry for errored virtual address.
4437 * So we need to block hugepage fault by PG_hwpoison bit check.
4438 */
4439 if (unlikely(PageHWPoison(page))) {
Miaohe Lin0eb98f12021-01-12 15:49:24 -08004440 ret = VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004441 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004442 goto backout_unlocked;
4443 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08004444 }
David Gibson1e8f8892006-01-06 00:10:44 -08004445
Andy Whitcroft57303d82008-08-12 15:08:47 -07004446 /*
4447 * If we are going to COW a private mapping later, we examine the
4448 * pending reservations for this page now. This will ensure that
4449 * any allocations necessary to record that reservation occur outside
4450 * the spinlock.
4451 */
Mike Kravetz5e911372015-09-08 15:01:28 -07004452 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004453 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004454 ret = VM_FAULT_OOM;
4455 goto backout_unlocked;
4456 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004457 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004458 vma_end_reservation(h, vma, haddr);
Mike Kravetz5e911372015-09-08 15:01:28 -07004459 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07004460
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08004461 ptl = huge_pte_lock(h, mm, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004462 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004463 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07004464 goto backout;
4465
Joonsoo Kim07443a82013-09-11 14:21:58 -07004466 if (anon_rmap) {
Mike Kravetzd6995da2021-02-24 12:08:51 -08004467 ClearHPageRestoreReserve(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004468 hugepage_add_new_anon_rmap(page, vma, haddr);
Choi Gi-yongac714902014-04-07 15:37:36 -07004469 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08004470 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08004471 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4472 && (vma->vm_flags & VM_SHARED)));
Huang Ying285b8dc2018-06-07 17:08:08 -07004473 set_huge_pte_at(mm, haddr, ptep, new_pte);
David Gibson1e8f8892006-01-06 00:10:44 -08004474
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08004475 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01004476 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08004477 /* Optimization, do the COW without a second fault */
Huang Ying974e6d62018-08-17 15:45:57 -07004478 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08004479 }
4480
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004481 spin_unlock(ptl);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004482
4483 /*
Mike Kravetz8f251a32021-02-24 12:08:56 -08004484 * Only set HPageMigratable in newly allocated pages. Existing pages
4485 * found in the pagecache may not have HPageMigratableset if they have
4486 * been isolated for migration.
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004487 */
4488 if (new_page)
Mike Kravetz8f251a32021-02-24 12:08:56 -08004489 SetHPageMigratable(page);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004490
Adam Litke4c887262005-10-29 18:16:46 -07004491 unlock_page(page);
4492out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004493 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07004494
4495backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004496 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07004497backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07004498 unlock_page(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004499 restore_reserve_on_error(h, vma, haddr, page);
Adam Litke4c887262005-10-29 18:16:46 -07004500 put_page(page);
4501 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004502}
4503
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004504#ifdef CONFIG_SMP
Wei Yang188b04a2019-11-30 17:57:02 -08004505u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004506{
4507 unsigned long key[2];
4508 u32 hash;
4509
Mike Kravetz1b426ba2019-05-13 17:19:41 -07004510 key[0] = (unsigned long) mapping;
4511 key[1] = idx;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004512
Mike Kravetz55254632019-11-30 17:56:30 -08004513 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004514
4515 return hash & (num_fault_mutexes - 1);
4516}
4517#else
4518/*
Miaohe Lin6c26d312021-02-24 12:07:19 -08004519 * For uniprocessor systems we always use a single mutex, so just
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004520 * return 0 and avoid the hashing overhead.
4521 */
Wei Yang188b04a2019-11-30 17:57:02 -08004522u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004523{
4524 return 0;
4525}
4526#endif
4527
Souptick Joarder2b740302018-08-23 17:01:36 -07004528vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01004529 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08004530{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004531 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004532 spinlock_t *ptl;
Souptick Joarder2b740302018-08-23 17:01:36 -07004533 vm_fault_t ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004534 u32 hash;
4535 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004536 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07004537 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07004538 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004539 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004540 int need_wait_lock = 0;
Huang Ying285b8dc2018-06-07 17:08:08 -07004541 unsigned long haddr = address & huge_page_mask(h);
Adam Litke86e52162006-01-06 00:10:43 -08004542
Huang Ying285b8dc2018-06-07 17:08:08 -07004543 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004544 if (ptep) {
Mike Kravetzc0d03812020-04-01 21:11:05 -07004545 /*
4546 * Since we hold no locks, ptep could be stale. That is
4547 * OK as we are only making decisions based on content and
4548 * not actually modifying content here.
4549 */
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004550 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09004551 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004552 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09004553 return 0;
4554 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07004555 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004556 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004557 }
4558
Mike Kravetzc0d03812020-04-01 21:11:05 -07004559 /*
4560 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004561 * until finished with ptep. This serves two purposes:
4562 * 1) It prevents huge_pmd_unshare from being called elsewhere
4563 * and making the ptep no longer valid.
4564 * 2) It synchronizes us with i_size modifications during truncation.
Mike Kravetzc0d03812020-04-01 21:11:05 -07004565 *
4566 * ptep could have already be assigned via huge_pte_offset. That
4567 * is OK, as huge_pte_alloc will return the same value unless
4568 * something has changed.
4569 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004570 mapping = vma->vm_file->f_mapping;
Mike Kravetzc0d03812020-04-01 21:11:05 -07004571 i_mmap_lock_read(mapping);
Peter Xuaec44e02021-05-04 18:33:00 -07004572 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
Mike Kravetzc0d03812020-04-01 21:11:05 -07004573 if (!ptep) {
4574 i_mmap_unlock_read(mapping);
4575 return VM_FAULT_OOM;
4576 }
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004577
David Gibson3935baa2006-03-22 00:08:53 -08004578 /*
4579 * Serialize hugepage allocation and instantiation, so that we don't
4580 * get spurious allocation failures if two CPUs race to instantiate
4581 * the same page in the page cache.
4582 */
Mike Kravetzc0d03812020-04-01 21:11:05 -07004583 idx = vma_hugecache_offset(h, vma, haddr);
Wei Yang188b04a2019-11-30 17:57:02 -08004584 hash = hugetlb_fault_mutex_hash(mapping, idx);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07004585 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004586
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004587 entry = huge_ptep_get(ptep);
4588 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004589 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07004590 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08004591 }
Adam Litke86e52162006-01-06 00:10:43 -08004592
Nick Piggin83c54072007-07-19 01:47:05 -07004593 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004594
Andy Whitcroft57303d82008-08-12 15:08:47 -07004595 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004596 * entry could be a migration/hwpoison entry at this point, so this
4597 * check prevents the kernel from going below assuming that we have
Ethon Paul7c8de352020-06-04 16:49:07 -07004598 * an active hugepage in pagecache. This goto expects the 2nd page
4599 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4600 * properly handle it.
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004601 */
4602 if (!pte_present(entry))
4603 goto out_mutex;
4604
4605 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07004606 * If we are going to COW the mapping later, we examine the pending
4607 * reservations for this page now. This will ensure that any
4608 * allocations necessary to record that reservation occur outside the
4609 * spinlock. For private mappings, we also lookup the pagecache
4610 * page now as it is used to determine if a reservation has been
4611 * consumed.
4612 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07004613 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004614 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004615 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07004616 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07004617 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004618 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004619 vma_end_reservation(h, vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004620
Mel Gormanf83a2752009-05-28 14:34:40 -07004621 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07004622 pagecache_page = hugetlbfs_pagecache_page(h,
Huang Ying285b8dc2018-06-07 17:08:08 -07004623 vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004624 }
4625
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004626 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004627
David Gibson1e8f8892006-01-06 00:10:44 -08004628 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07004629 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004630 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07004631
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004632 /*
4633 * hugetlb_cow() requires page locks of pte_page(entry) and
4634 * pagecache_page, so here we need take the former one
4635 * when page != pagecache_page or !pagecache_page.
4636 */
4637 page = pte_page(entry);
4638 if (page != pagecache_page)
4639 if (!trylock_page(page)) {
4640 need_wait_lock = 1;
4641 goto out_ptl;
4642 }
4643
4644 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07004645
Hugh Dickins788c7df2009-06-23 13:49:05 +01004646 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07004647 if (!huge_pte_write(entry)) {
Huang Ying974e6d62018-08-17 15:45:57 -07004648 ret = hugetlb_cow(mm, vma, address, ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004649 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004650 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07004651 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07004652 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07004653 }
4654 entry = pte_mkyoung(entry);
Huang Ying285b8dc2018-06-07 17:08:08 -07004655 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
Hugh Dickins788c7df2009-06-23 13:49:05 +01004656 flags & FAULT_FLAG_WRITE))
Huang Ying285b8dc2018-06-07 17:08:08 -07004657 update_mmu_cache(vma, haddr, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004658out_put_page:
4659 if (page != pagecache_page)
4660 unlock_page(page);
4661 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004662out_ptl:
4663 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004664
4665 if (pagecache_page) {
4666 unlock_page(pagecache_page);
4667 put_page(pagecache_page);
4668 }
David Gibsonb4d1d992008-10-15 22:01:11 -07004669out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07004670 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004671 i_mmap_unlock_read(mapping);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004672 /*
4673 * Generally it's safe to hold refcount during waiting page lock. But
4674 * here we just wait to defer the next page fault to avoid busy loop and
4675 * the page is not used after unlocked before returning from the current
4676 * page fault. So we are safe from accessing freed page, even if we wait
4677 * here without taking refcount.
4678 */
4679 if (need_wait_lock)
4680 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08004681 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08004682}
4683
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004684/*
4685 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4686 * modifications for huge pages.
4687 */
4688int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4689 pte_t *dst_pte,
4690 struct vm_area_struct *dst_vma,
4691 unsigned long dst_addr,
4692 unsigned long src_addr,
4693 struct page **pagep)
4694{
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004695 struct address_space *mapping;
4696 pgoff_t idx;
4697 unsigned long size;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004698 int vm_shared = dst_vma->vm_flags & VM_SHARED;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004699 struct hstate *h = hstate_vma(dst_vma);
4700 pte_t _dst_pte;
4701 spinlock_t *ptl;
4702 int ret;
4703 struct page *page;
4704
4705 if (!*pagep) {
4706 ret = -ENOMEM;
4707 page = alloc_huge_page(dst_vma, dst_addr, 0);
4708 if (IS_ERR(page))
4709 goto out;
4710
4711 ret = copy_huge_page_from_user(page,
4712 (const void __user *) src_addr,
Mike Kravetz810a56b2017-02-22 15:42:58 -08004713 pages_per_huge_page(h), false);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004714
Michel Lespinassec1e8d7c2020-06-08 21:33:54 -07004715 /* fallback to copy_from_user outside mmap_lock */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004716 if (unlikely(ret)) {
Andrea Arcangeli9e368252018-11-30 14:09:25 -08004717 ret = -ENOENT;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004718 *pagep = page;
4719 /* don't free the page */
4720 goto out;
4721 }
4722 } else {
4723 page = *pagep;
4724 *pagep = NULL;
4725 }
4726
4727 /*
4728 * The memory barrier inside __SetPageUptodate makes sure that
4729 * preceding stores to the page contents become visible before
4730 * the set_pte_at() write.
4731 */
4732 __SetPageUptodate(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004733
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004734 mapping = dst_vma->vm_file->f_mapping;
4735 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4736
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004737 /*
4738 * If shared, add to page cache
4739 */
4740 if (vm_shared) {
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004741 size = i_size_read(mapping->host) >> huge_page_shift(h);
4742 ret = -EFAULT;
4743 if (idx >= size)
4744 goto out_release_nounlock;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004745
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004746 /*
4747 * Serialization between remove_inode_hugepages() and
4748 * huge_add_to_page_cache() below happens through the
4749 * hugetlb_fault_mutex_table that here must be hold by
4750 * the caller.
4751 */
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004752 ret = huge_add_to_page_cache(page, mapping, idx);
4753 if (ret)
4754 goto out_release_nounlock;
4755 }
4756
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004757 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4758 spin_lock(ptl);
4759
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004760 /*
4761 * Recheck the i_size after holding PT lock to make sure not
4762 * to leave any page mapped (as page_mapped()) beyond the end
4763 * of the i_size (remove_inode_hugepages() is strict about
4764 * enforcing that). If we bail out here, we'll also leave a
4765 * page in the radix tree in the vm_shared case beyond the end
4766 * of the i_size, but remove_inode_hugepages() will take care
4767 * of it as soon as we drop the hugetlb_fault_mutex_table.
4768 */
4769 size = i_size_read(mapping->host) >> huge_page_shift(h);
4770 ret = -EFAULT;
4771 if (idx >= size)
4772 goto out_release_unlock;
4773
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004774 ret = -EEXIST;
4775 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4776 goto out_release_unlock;
4777
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004778 if (vm_shared) {
4779 page_dup_rmap(page, true);
4780 } else {
Mike Kravetzd6995da2021-02-24 12:08:51 -08004781 ClearHPageRestoreReserve(page);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004782 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4783 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004784
4785 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4786 if (dst_vma->vm_flags & VM_WRITE)
4787 _dst_pte = huge_pte_mkdirty(_dst_pte);
4788 _dst_pte = pte_mkyoung(_dst_pte);
4789
4790 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4791
4792 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4793 dst_vma->vm_flags & VM_WRITE);
4794 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4795
4796 /* No need to invalidate - it was non-present before */
4797 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4798
4799 spin_unlock(ptl);
Mike Kravetz8f251a32021-02-24 12:08:56 -08004800 SetHPageMigratable(page);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004801 if (vm_shared)
4802 unlock_page(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004803 ret = 0;
4804out:
4805 return ret;
4806out_release_unlock:
4807 spin_unlock(ptl);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004808 if (vm_shared)
4809 unlock_page(page);
Andrea Arcangeli5af10df2017-08-10 15:23:38 -07004810out_release_nounlock:
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004811 put_page(page);
4812 goto out;
4813}
4814
Joao Martins82e5d372021-02-24 12:07:16 -08004815static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma,
4816 int refs, struct page **pages,
4817 struct vm_area_struct **vmas)
4818{
4819 int nr;
4820
4821 for (nr = 0; nr < refs; nr++) {
4822 if (likely(pages))
4823 pages[nr] = mem_map_offset(page, nr);
4824 if (vmas)
4825 vmas[nr] = vma;
4826 }
4827}
4828
Michel Lespinasse28a35712013-02-22 16:35:55 -08004829long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4830 struct page **pages, struct vm_area_struct **vmas,
4831 unsigned long *position, unsigned long *nr_pages,
Peter Xu4f6da932020-04-01 21:07:58 -07004832 long i, unsigned int flags, int *locked)
David Gibson63551ae2005-06-21 17:14:44 -07004833{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004834 unsigned long pfn_offset;
4835 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08004836 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07004837 struct hstate *h = hstate_vma(vma);
Joao Martins0fa5bc42021-02-24 12:07:12 -08004838 int err = -EFAULT, refs;
David Gibson63551ae2005-06-21 17:14:44 -07004839
David Gibson63551ae2005-06-21 17:14:44 -07004840 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07004841 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004842 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004843 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07004844 struct page *page;
4845
4846 /*
David Rientjes02057962015-04-14 15:48:24 -07004847 * If we have a pending SIGKILL, don't keep faulting pages and
4848 * potentially allocating memory.
4849 */
Davidlohr Buesofa45f112019-01-03 15:28:55 -08004850 if (fatal_signal_pending(current)) {
David Rientjes02057962015-04-14 15:48:24 -07004851 remainder = 0;
4852 break;
4853 }
4854
4855 /*
Adam Litke4c887262005-10-29 18:16:46 -07004856 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004857 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07004858 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004859 *
4860 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07004861 */
Punit Agrawal7868a202017-07-06 15:39:42 -07004862 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4863 huge_page_size(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004864 if (pte)
4865 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004866 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07004867
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004868 /*
4869 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004870 * an error where there's an empty slot with no huge pagecache
4871 * to back it. This way, we avoid allocating a hugepage, and
4872 * the sparse dumpfile avoids allocating disk blocks, but its
4873 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004874 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004875 if (absent && (flags & FOLL_DUMP) &&
4876 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004877 if (pte)
4878 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004879 remainder = 0;
4880 break;
4881 }
4882
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07004883 /*
4884 * We need call hugetlb_fault for both hugepages under migration
4885 * (in which case hugetlb_fault waits for the migration,) and
4886 * hwpoisoned hugepages (in which case we need to prevent the
4887 * caller from accessing to them.) In order to do this, we use
4888 * here is_swap_pte instead of is_hugetlb_entry_migration and
4889 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4890 * both cases, and because we can't follow correct pages
4891 * directly from any kind of swap entries.
4892 */
4893 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07004894 ((flags & FOLL_WRITE) &&
4895 !huge_pte_write(huge_ptep_get(pte)))) {
Souptick Joarder2b740302018-08-23 17:01:36 -07004896 vm_fault_t ret;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004897 unsigned int fault_flags = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004898
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004899 if (pte)
4900 spin_unlock(ptl);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004901 if (flags & FOLL_WRITE)
4902 fault_flags |= FAULT_FLAG_WRITE;
Peter Xu4f6da932020-04-01 21:07:58 -07004903 if (locked)
Peter Xu71335f32020-04-01 21:08:53 -07004904 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4905 FAULT_FLAG_KILLABLE;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004906 if (flags & FOLL_NOWAIT)
4907 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4908 FAULT_FLAG_RETRY_NOWAIT;
4909 if (flags & FOLL_TRIED) {
Peter Xu4426e942020-04-01 21:08:49 -07004910 /*
4911 * Note: FAULT_FLAG_ALLOW_RETRY and
4912 * FAULT_FLAG_TRIED can co-exist
4913 */
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004914 fault_flags |= FAULT_FLAG_TRIED;
4915 }
4916 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4917 if (ret & VM_FAULT_ERROR) {
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004918 err = vm_fault_to_errno(ret, flags);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004919 remainder = 0;
4920 break;
4921 }
4922 if (ret & VM_FAULT_RETRY) {
Peter Xu4f6da932020-04-01 21:07:58 -07004923 if (locked &&
Andrea Arcangeli1ac25012019-02-01 14:20:16 -08004924 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
Peter Xu4f6da932020-04-01 21:07:58 -07004925 *locked = 0;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004926 *nr_pages = 0;
4927 /*
4928 * VM_FAULT_RETRY must not return an
4929 * error, it will return zero
4930 * instead.
4931 *
4932 * No need to update "position" as the
4933 * caller will not check it after
4934 * *nr_pages is set to 0.
4935 */
4936 return i;
4937 }
4938 continue;
Adam Litke4c887262005-10-29 18:16:46 -07004939 }
David Gibson63551ae2005-06-21 17:14:44 -07004940
Andi Kleena5516432008-07-23 21:27:41 -07004941 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004942 page = pte_page(huge_ptep_get(pte));
Linus Torvalds8fde12c2019-04-11 10:49:19 -07004943
4944 /*
Zhigang Luacbfb082019-11-30 17:57:06 -08004945 * If subpage information not requested, update counters
4946 * and skip the same_page loop below.
4947 */
4948 if (!pages && !vmas && !pfn_offset &&
4949 (vaddr + huge_page_size(h) < vma->vm_end) &&
4950 (remainder >= pages_per_huge_page(h))) {
4951 vaddr += huge_page_size(h);
4952 remainder -= pages_per_huge_page(h);
4953 i += pages_per_huge_page(h);
4954 spin_unlock(ptl);
4955 continue;
4956 }
4957
Joao Martins82e5d372021-02-24 12:07:16 -08004958 refs = min3(pages_per_huge_page(h) - pfn_offset,
4959 (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder);
Joao Martins0fa5bc42021-02-24 12:07:12 -08004960
Joao Martins82e5d372021-02-24 12:07:16 -08004961 if (pages || vmas)
4962 record_subpages_vmas(mem_map_offset(page, pfn_offset),
4963 vma, refs,
4964 likely(pages) ? pages + i : NULL,
4965 vmas ? vmas + i : NULL);
David Gibson63551ae2005-06-21 17:14:44 -07004966
Joao Martins82e5d372021-02-24 12:07:16 -08004967 if (pages) {
Joao Martins0fa5bc42021-02-24 12:07:12 -08004968 /*
4969 * try_grab_compound_head() should always succeed here,
4970 * because: a) we hold the ptl lock, and b) we've just
4971 * checked that the huge page is present in the page
4972 * tables. If the huge page is present, then the tail
4973 * pages must also be present. The ptl prevents the
4974 * head page and tail pages from being rearranged in
4975 * any way. So this page must be available at this
4976 * point, unless the page refcount overflowed:
4977 */
Joao Martins82e5d372021-02-24 12:07:16 -08004978 if (WARN_ON_ONCE(!try_grab_compound_head(pages[i],
Joao Martins0fa5bc42021-02-24 12:07:12 -08004979 refs,
4980 flags))) {
4981 spin_unlock(ptl);
4982 remainder = 0;
4983 err = -ENOMEM;
4984 break;
4985 }
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004986 }
Joao Martins82e5d372021-02-24 12:07:16 -08004987
4988 vaddr += (refs << PAGE_SHIFT);
4989 remainder -= refs;
4990 i += refs;
4991
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004992 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07004993 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08004994 *nr_pages = remainder;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004995 /*
4996 * setting position is actually required only if remainder is
4997 * not zero but it's faster not to add a "if (remainder)"
4998 * branch.
4999 */
David Gibson63551ae2005-06-21 17:14:44 -07005000 *position = vaddr;
5001
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07005002 return i ? i : err;
David Gibson63551ae2005-06-21 17:14:44 -07005003}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005004
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005005unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005006 unsigned long address, unsigned long end, pgprot_t newprot)
5007{
5008 struct mm_struct *mm = vma->vm_mm;
5009 unsigned long start = address;
5010 pte_t *ptep;
5011 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07005012 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005013 unsigned long pages = 0;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005014 bool shared_pmd = false;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005015 struct mmu_notifier_range range;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005016
5017 /*
5018 * In the case of shared PMDs, the area to flush could be beyond
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005019 * start/end. Set range.start/range.end to cover the maximum possible
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005020 * range if PMD sharing is possible.
5021 */
Jérôme Glisse7269f992019-05-13 17:20:53 -07005022 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5023 0, vma, mm, start, end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005024 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005025
5026 BUG_ON(address >= end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005027 flush_cache_range(vma, range.start, range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005028
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005029 mmu_notifier_invalidate_range_start(&range);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005030 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07005031 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005032 spinlock_t *ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07005033 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005034 if (!ptep)
5035 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005036 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae2042020-08-11 18:31:38 -07005037 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005038 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005039 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005040 shared_pmd = true;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08005041 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005042 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005043 pte = huge_ptep_get(ptep);
5044 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5045 spin_unlock(ptl);
5046 continue;
5047 }
5048 if (unlikely(is_hugetlb_entry_migration(pte))) {
5049 swp_entry_t entry = pte_to_swp_entry(pte);
5050
5051 if (is_write_migration_entry(entry)) {
5052 pte_t newpte;
5053
5054 make_migration_entry_read(&entry);
5055 newpte = swp_entry_to_pte(entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07005056 set_huge_swap_pte_at(mm, address, ptep,
5057 newpte, huge_page_size(h));
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005058 pages++;
5059 }
5060 spin_unlock(ptl);
5061 continue;
5062 }
5063 if (!huge_pte_none(pte)) {
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005064 pte_t old_pte;
5065
5066 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5067 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08005068 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005069 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005070 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005071 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005072 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005073 }
Mel Gormand8333522012-07-31 16:46:20 -07005074 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005075 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07005076 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005077 * once we release i_mmap_rwsem, another task can do the final put_page
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005078 * and that page table be reused and filled with junk. If we actually
5079 * did unshare a page of pmds, flush the range corresponding to the pud.
Mel Gormand8333522012-07-31 16:46:20 -07005080 */
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005081 if (shared_pmd)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005082 flush_hugetlb_tlb_range(vma, range.start, range.end);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005083 else
5084 flush_hugetlb_tlb_range(vma, start, end);
Jérôme Glisse0f108512017-11-15 17:34:07 -08005085 /*
5086 * No need to call mmu_notifier_invalidate_range() we are downgrading
5087 * page table protection not changing it to point to a new page.
5088 *
Mike Rapoportad56b732018-03-21 21:22:47 +02005089 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08005090 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005091 i_mmap_unlock_write(vma->vm_file->f_mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005092 mmu_notifier_invalidate_range_end(&range);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005093
5094 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005095}
5096
Mike Kravetz33b8f842021-02-24 12:09:54 -08005097/* Return true if reservation was successful, false otherwise. */
5098bool hugetlb_reserve_pages(struct inode *inode,
Mel Gormana1e78772008-07-23 21:27:23 -07005099 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00005100 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005101 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07005102{
Mike Kravetz33b8f842021-02-24 12:09:54 -08005103 long chg, add = -1;
Andi Kleena5516432008-07-23 21:27:41 -07005104 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07005105 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005106 struct resv_map *resv_map;
Mina Almasry075a61d2020-04-01 21:11:28 -07005107 struct hugetlb_cgroup *h_cg = NULL;
Mina Almasry0db9d742020-04-01 21:11:25 -07005108 long gbl_reserve, regions_needed = 0;
Adam Litkee4e574b2007-10-16 01:26:19 -07005109
Mike Kravetz63489f82018-03-22 16:17:13 -07005110 /* This should never happen */
5111 if (from > to) {
5112 VM_WARN(1, "%s called with a negative range\n", __func__);
Mike Kravetz33b8f842021-02-24 12:09:54 -08005113 return false;
Mike Kravetz63489f82018-03-22 16:17:13 -07005114 }
5115
Mel Gormana1e78772008-07-23 21:27:23 -07005116 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00005117 * Only apply hugepage reservation if asked. At fault time, an
5118 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07005119 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00005120 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005121 if (vm_flags & VM_NORESERVE)
Mike Kravetz33b8f842021-02-24 12:09:54 -08005122 return true;
Mel Gorman17c9d122009-02-11 16:34:16 +00005123
5124 /*
Mel Gormana1e78772008-07-23 21:27:23 -07005125 * Shared mappings base their reservation on the number of pages that
5126 * are already allocated on behalf of the file. Private mappings need
5127 * to reserve the full area even if read-only as mprotect() may be
5128 * called to make the mapping read-write. Assume !vma is a shm mapping
5129 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005130 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzf27a5132019-05-13 17:22:55 -07005131 /*
5132 * resv_map can not be NULL as hugetlb_reserve_pages is only
5133 * called for inodes for which resv_maps were created (see
5134 * hugetlbfs_get_inode).
5135 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005136 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005137
Mina Almasry0db9d742020-04-01 21:11:25 -07005138 chg = region_chg(resv_map, from, to, &regions_needed);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005139
5140 } else {
Mina Almasrye9fe92a2020-04-01 21:11:21 -07005141 /* Private mapping. */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005142 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00005143 if (!resv_map)
Mike Kravetz33b8f842021-02-24 12:09:54 -08005144 return false;
Mel Gorman5a6fe122009-02-10 14:02:27 +00005145
Mel Gorman17c9d122009-02-11 16:34:16 +00005146 chg = to - from;
5147
Mel Gorman5a6fe122009-02-10 14:02:27 +00005148 set_vma_resv_map(vma, resv_map);
5149 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5150 }
5151
Mike Kravetz33b8f842021-02-24 12:09:54 -08005152 if (chg < 0)
Dave Hansenc50ac052012-05-29 15:06:46 -07005153 goto out_err;
Mel Gorman17c9d122009-02-11 16:34:16 +00005154
Mike Kravetz33b8f842021-02-24 12:09:54 -08005155 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
5156 chg * pages_per_huge_page(h), &h_cg) < 0)
Mina Almasry075a61d2020-04-01 21:11:28 -07005157 goto out_err;
Mina Almasry075a61d2020-04-01 21:11:28 -07005158
5159 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5160 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5161 * of the resv_map.
5162 */
5163 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5164 }
5165
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005166 /*
5167 * There must be enough pages in the subpool for the mapping. If
5168 * the subpool has a minimum size, there may be some global
5169 * reservations already in place (gbl_reserve).
5170 */
5171 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
Mike Kravetz33b8f842021-02-24 12:09:54 -08005172 if (gbl_reserve < 0)
Mina Almasry075a61d2020-04-01 21:11:28 -07005173 goto out_uncharge_cgroup;
Mel Gorman17c9d122009-02-11 16:34:16 +00005174
5175 /*
5176 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07005177 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00005178 */
Mike Kravetz33b8f842021-02-24 12:09:54 -08005179 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
Mina Almasry075a61d2020-04-01 21:11:28 -07005180 goto out_put_pages;
Mel Gorman17c9d122009-02-11 16:34:16 +00005181
5182 /*
5183 * Account for the reservations made. Shared mappings record regions
5184 * that have reservations as they are shared by multiple VMAs.
5185 * When the last VMA disappears, the region map says how much
5186 * the reservation was and the page cache tells how much of
5187 * the reservation was consumed. Private mappings are per-VMA and
5188 * only the consumed reservations are tracked. When the VMA
5189 * disappears, the original reservation is the VMA size and the
5190 * consumed reservations are stored in the map. Hence, nothing
5191 * else has to be done for private mappings here
5192 */
Mike Kravetz33039672015-06-24 16:57:58 -07005193 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07005194 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005195
Mina Almasry0db9d742020-04-01 21:11:25 -07005196 if (unlikely(add < 0)) {
5197 hugetlb_acct_memory(h, -gbl_reserve);
Mina Almasry075a61d2020-04-01 21:11:28 -07005198 goto out_put_pages;
Mina Almasry0db9d742020-04-01 21:11:25 -07005199 } else if (unlikely(chg > add)) {
Mike Kravetz33039672015-06-24 16:57:58 -07005200 /*
5201 * pages in this range were added to the reserve
5202 * map between region_chg and region_add. This
5203 * indicates a race with alloc_huge_page. Adjust
5204 * the subpool and reserve counts modified above
5205 * based on the difference.
5206 */
5207 long rsv_adjust;
5208
Miaohe Lind85aecf2021-03-24 21:37:17 -07005209 /*
5210 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5211 * reference to h_cg->css. See comment below for detail.
5212 */
Mina Almasry075a61d2020-04-01 21:11:28 -07005213 hugetlb_cgroup_uncharge_cgroup_rsvd(
5214 hstate_index(h),
5215 (chg - add) * pages_per_huge_page(h), h_cg);
5216
Mike Kravetz33039672015-06-24 16:57:58 -07005217 rsv_adjust = hugepage_subpool_put_pages(spool,
5218 chg - add);
5219 hugetlb_acct_memory(h, -rsv_adjust);
Miaohe Lind85aecf2021-03-24 21:37:17 -07005220 } else if (h_cg) {
5221 /*
5222 * The file_regions will hold their own reference to
5223 * h_cg->css. So we should release the reference held
5224 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5225 * done.
5226 */
5227 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005228 }
5229 }
Mike Kravetz33b8f842021-02-24 12:09:54 -08005230 return true;
5231
Mina Almasry075a61d2020-04-01 21:11:28 -07005232out_put_pages:
5233 /* put back original number of pages, chg */
5234 (void)hugepage_subpool_put_pages(spool, chg);
5235out_uncharge_cgroup:
5236 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5237 chg * pages_per_huge_page(h), h_cg);
Dave Hansenc50ac052012-05-29 15:06:46 -07005238out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07005239 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mina Almasry0db9d742020-04-01 21:11:25 -07005240 /* Only call region_abort if the region_chg succeeded but the
5241 * region_add failed or didn't run.
5242 */
5243 if (chg >= 0 && add < 0)
5244 region_abort(resv_map, from, to, regions_needed);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07005245 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5246 kref_put(&resv_map->refs, resv_map_release);
Mike Kravetz33b8f842021-02-24 12:09:54 -08005247 return false;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005248}
5249
Mike Kravetzb5cec282015-09-08 15:01:41 -07005250long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5251 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005252{
Andi Kleena5516432008-07-23 21:27:41 -07005253 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005254 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005255 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07005256 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005257 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08005258
Mike Kravetzf27a5132019-05-13 17:22:55 -07005259 /*
5260 * Since this routine can be called in the evict inode path for all
5261 * hugetlbfs inodes, resv_map could be NULL.
5262 */
Mike Kravetzb5cec282015-09-08 15:01:41 -07005263 if (resv_map) {
5264 chg = region_del(resv_map, start, end);
5265 /*
5266 * region_del() can fail in the rare case where a region
5267 * must be split and another region descriptor can not be
5268 * allocated. If end == LONG_MAX, it will not fail.
5269 */
5270 if (chg < 0)
5271 return chg;
5272 }
5273
Ken Chen45c682a2007-11-14 16:59:44 -08005274 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07005275 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08005276 spin_unlock(&inode->i_lock);
5277
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005278 /*
5279 * If the subpool has a minimum size, the number of global
5280 * reservations to be released may be adjusted.
Miaohe Lindddf31a2021-05-04 18:34:35 -07005281 *
5282 * Note that !resv_map implies freed == 0. So (chg - freed)
5283 * won't go negative.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005284 */
5285 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5286 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07005287
5288 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005289}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09005290
Steve Capper3212b532013-04-23 12:35:02 +01005291#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5292static unsigned long page_table_shareable(struct vm_area_struct *svma,
5293 struct vm_area_struct *vma,
5294 unsigned long addr, pgoff_t idx)
5295{
5296 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5297 svma->vm_start;
5298 unsigned long sbase = saddr & PUD_MASK;
5299 unsigned long s_end = sbase + PUD_SIZE;
5300
5301 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08005302 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5303 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01005304
5305 /*
5306 * match the virtual addresses, permission and the alignment of the
5307 * page table page.
5308 */
5309 if (pmd_index(addr) != pmd_index(saddr) ||
5310 vm_flags != svm_flags ||
Miaohe Lin07e51ed2021-02-24 12:07:39 -08005311 !range_in_vma(svma, sbase, s_end))
Steve Capper3212b532013-04-23 12:35:02 +01005312 return 0;
5313
5314 return saddr;
5315}
5316
Nicholas Krause31aafb42015-09-04 15:47:58 -07005317static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01005318{
5319 unsigned long base = addr & PUD_MASK;
5320 unsigned long end = base + PUD_SIZE;
5321
5322 /*
5323 * check on proper vm_flags and page table alignment
5324 */
Mike Kravetz017b1662018-10-05 15:51:29 -07005325 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
Nicholas Krause31aafb42015-09-04 15:47:58 -07005326 return true;
5327 return false;
Steve Capper3212b532013-04-23 12:35:02 +01005328}
5329
Peter Xuc1991e02021-05-04 18:33:04 -07005330bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5331{
5332#ifdef CONFIG_USERFAULTFD
5333 if (uffd_disable_huge_pmd_share(vma))
5334 return false;
5335#endif
5336 return vma_shareable(vma, addr);
5337}
5338
Steve Capper3212b532013-04-23 12:35:02 +01005339/*
Mike Kravetz017b1662018-10-05 15:51:29 -07005340 * Determine if start,end range within vma could be mapped by shared pmd.
5341 * If yes, adjust start and end to cover range associated with possible
5342 * shared pmd mappings.
5343 */
5344void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5345 unsigned long *start, unsigned long *end)
5346{
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005347 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5348 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005349
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005350 /*
5351 * vma need span at least one aligned PUD size and the start,end range
5352 * must at least partialy within it.
5353 */
5354 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5355 (*end <= v_start) || (*start >= v_end))
Mike Kravetz017b1662018-10-05 15:51:29 -07005356 return;
5357
Peter Xu75802ca62020-08-06 23:26:11 -07005358 /* Extend the range to be PUD aligned for a worst case scenario */
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005359 if (*start > v_start)
5360 *start = ALIGN_DOWN(*start, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005361
Li Xinhaia1ba9da2021-02-24 12:06:54 -08005362 if (*end < v_end)
5363 *end = ALIGN(*end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005364}
5365
5366/*
Steve Capper3212b532013-04-23 12:35:02 +01005367 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5368 * and returns the corresponding pte. While this is not necessary for the
5369 * !shared pmd case because we can allocate the pmd later as well, it makes the
Mike Kravetzc0d03812020-04-01 21:11:05 -07005370 * code much cleaner.
5371 *
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005372 * This routine must be called with i_mmap_rwsem held in at least read mode if
5373 * sharing is possible. For hugetlbfs, this prevents removal of any page
5374 * table entries associated with the address space. This is important as we
5375 * are setting up sharing based on existing page table entries (mappings).
5376 *
5377 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5378 * huge_pte_alloc know that sharing is not possible and do not take
5379 * i_mmap_rwsem as a performance optimization. This is handled by the
5380 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5381 * only required for subsequent processing.
Steve Capper3212b532013-04-23 12:35:02 +01005382 */
Peter Xuaec44e02021-05-04 18:33:00 -07005383pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5384 unsigned long addr, pud_t *pud)
Steve Capper3212b532013-04-23 12:35:02 +01005385{
Steve Capper3212b532013-04-23 12:35:02 +01005386 struct address_space *mapping = vma->vm_file->f_mapping;
5387 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5388 vma->vm_pgoff;
5389 struct vm_area_struct *svma;
5390 unsigned long saddr;
5391 pte_t *spte = NULL;
5392 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005393 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01005394
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005395 i_mmap_assert_locked(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005396 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5397 if (svma == vma)
5398 continue;
5399
5400 saddr = page_table_shareable(svma, vma, addr, idx);
5401 if (saddr) {
Punit Agrawal7868a202017-07-06 15:39:42 -07005402 spte = huge_pte_offset(svma->vm_mm, saddr,
5403 vma_mmu_pagesize(svma));
Steve Capper3212b532013-04-23 12:35:02 +01005404 if (spte) {
5405 get_page(virt_to_page(spte));
5406 break;
5407 }
5408 }
5409 }
5410
5411 if (!spte)
5412 goto out;
5413
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08005414 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005415 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01005416 pud_populate(mm, pud,
5417 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07005418 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005419 } else {
Steve Capper3212b532013-04-23 12:35:02 +01005420 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005421 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005422 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01005423out:
5424 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Steve Capper3212b532013-04-23 12:35:02 +01005425 return pte;
5426}
5427
5428/*
5429 * unmap huge page backed by shared pte.
5430 *
5431 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5432 * indicated by page_count > 1, unmap is achieved by clearing pud and
5433 * decrementing the ref count. If count == 1, the pte page is not shared.
5434 *
Mike Kravetzc0d03812020-04-01 21:11:05 -07005435 * Called with page table lock held and i_mmap_rwsem held in write mode.
Steve Capper3212b532013-04-23 12:35:02 +01005436 *
5437 * returns: 1 successfully unmapped a shared pte page
5438 * 0 the underlying pte page is not shared, or it is the last user
5439 */
Mike Kravetz34ae2042020-08-11 18:31:38 -07005440int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5441 unsigned long *addr, pte_t *ptep)
Steve Capper3212b532013-04-23 12:35:02 +01005442{
5443 pgd_t *pgd = pgd_offset(mm, *addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005444 p4d_t *p4d = p4d_offset(pgd, *addr);
5445 pud_t *pud = pud_offset(p4d, *addr);
Steve Capper3212b532013-04-23 12:35:02 +01005446
Mike Kravetz34ae2042020-08-11 18:31:38 -07005447 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005448 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5449 if (page_count(virt_to_page(ptep)) == 1)
5450 return 0;
5451
5452 pud_clear(pud);
5453 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005454 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01005455 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5456 return 1;
5457}
Peter Xuc1991e02021-05-04 18:33:04 -07005458
Steve Capper9e5fc742013-04-30 08:02:03 +01005459#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
Peter Xuaec44e02021-05-04 18:33:00 -07005460pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5461 unsigned long addr, pud_t *pud)
Steve Capper9e5fc742013-04-30 08:02:03 +01005462{
5463 return NULL;
5464}
Zhang Zhene81f2d22015-06-24 16:56:13 -07005465
Mike Kravetz34ae2042020-08-11 18:31:38 -07005466int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5467 unsigned long *addr, pte_t *ptep)
Zhang Zhene81f2d22015-06-24 16:56:13 -07005468{
5469 return 0;
5470}
Mike Kravetz017b1662018-10-05 15:51:29 -07005471
5472void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5473 unsigned long *start, unsigned long *end)
5474{
5475}
Peter Xuc1991e02021-05-04 18:33:04 -07005476
5477bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5478{
5479 return false;
5480}
Steve Capper3212b532013-04-23 12:35:02 +01005481#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5482
Steve Capper9e5fc742013-04-30 08:02:03 +01005483#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
Peter Xuaec44e02021-05-04 18:33:00 -07005484pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
Steve Capper9e5fc742013-04-30 08:02:03 +01005485 unsigned long addr, unsigned long sz)
5486{
5487 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005488 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01005489 pud_t *pud;
5490 pte_t *pte = NULL;
5491
5492 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovf4f0a3d2017-11-29 16:11:30 -08005493 p4d = p4d_alloc(mm, pgd, addr);
5494 if (!p4d)
5495 return NULL;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005496 pud = pud_alloc(mm, p4d, addr);
Steve Capper9e5fc742013-04-30 08:02:03 +01005497 if (pud) {
5498 if (sz == PUD_SIZE) {
5499 pte = (pte_t *)pud;
5500 } else {
5501 BUG_ON(sz != PMD_SIZE);
Peter Xuc1991e02021-05-04 18:33:04 -07005502 if (want_pmd_share(vma, addr) && pud_none(*pud))
Peter Xuaec44e02021-05-04 18:33:00 -07005503 pte = huge_pmd_share(mm, vma, addr, pud);
Steve Capper9e5fc742013-04-30 08:02:03 +01005504 else
5505 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5506 }
5507 }
Michal Hocko4e666312016-08-02 14:02:34 -07005508 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01005509
5510 return pte;
5511}
5512
Punit Agrawal9b19df22017-09-06 16:21:01 -07005513/*
5514 * huge_pte_offset() - Walk the page table to resolve the hugepage
5515 * entry at address @addr
5516 *
Li Xinhai8ac0b812020-06-03 16:00:53 -07005517 * Return: Pointer to page table entry (PUD or PMD) for
5518 * address @addr, or NULL if a !p*d_present() entry is encountered and the
Punit Agrawal9b19df22017-09-06 16:21:01 -07005519 * size @sz doesn't match the hugepage size at this level of the page
5520 * table.
5521 */
Punit Agrawal7868a202017-07-06 15:39:42 -07005522pte_t *huge_pte_offset(struct mm_struct *mm,
5523 unsigned long addr, unsigned long sz)
Steve Capper9e5fc742013-04-30 08:02:03 +01005524{
5525 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005526 p4d_t *p4d;
Li Xinhai8ac0b812020-06-03 16:00:53 -07005527 pud_t *pud;
5528 pmd_t *pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01005529
5530 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005531 if (!pgd_present(*pgd))
5532 return NULL;
5533 p4d = p4d_offset(pgd, addr);
5534 if (!p4d_present(*p4d))
5535 return NULL;
Punit Agrawal9b19df22017-09-06 16:21:01 -07005536
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005537 pud = pud_offset(p4d, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07005538 if (sz == PUD_SIZE)
5539 /* must be pud huge, non-present or none */
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005540 return (pte_t *)pud;
Li Xinhai8ac0b812020-06-03 16:00:53 -07005541 if (!pud_present(*pud))
5542 return NULL;
5543 /* must have a valid entry and size to go further */
Punit Agrawal9b19df22017-09-06 16:21:01 -07005544
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005545 pmd = pmd_offset(pud, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07005546 /* must be pmd huge, non-present or none */
5547 return (pte_t *)pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01005548}
5549
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08005550#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5551
5552/*
5553 * These functions are overwritable if your architecture needs its own
5554 * behavior.
5555 */
5556struct page * __weak
5557follow_huge_addr(struct mm_struct *mm, unsigned long address,
5558 int write)
5559{
5560 return ERR_PTR(-EINVAL);
5561}
5562
5563struct page * __weak
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -07005564follow_huge_pd(struct vm_area_struct *vma,
5565 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5566{
5567 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5568 return NULL;
5569}
5570
5571struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01005572follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005573 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01005574{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005575 struct page *page = NULL;
5576 spinlock_t *ptl;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005577 pte_t pte;
John Hubbard3faa52c2020-04-01 21:05:29 -07005578
5579 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5580 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5581 (FOLL_PIN | FOLL_GET)))
5582 return NULL;
5583
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005584retry:
5585 ptl = pmd_lockptr(mm, pmd);
5586 spin_lock(ptl);
5587 /*
5588 * make sure that the address range covered by this pmd is not
5589 * unmapped from other threads.
5590 */
5591 if (!pmd_huge(*pmd))
5592 goto out;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005593 pte = huge_ptep_get((pte_t *)pmd);
5594 if (pte_present(pte)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07005595 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
John Hubbard3faa52c2020-04-01 21:05:29 -07005596 /*
5597 * try_grab_page() should always succeed here, because: a) we
5598 * hold the pmd (ptl) lock, and b) we've just checked that the
5599 * huge pmd (head) page is present in the page tables. The ptl
5600 * prevents the head page and tail pages from being rearranged
5601 * in any way. So this page must be available at this point,
5602 * unless the page refcount overflowed:
5603 */
5604 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5605 page = NULL;
5606 goto out;
5607 }
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005608 } else {
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005609 if (is_hugetlb_entry_migration(pte)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005610 spin_unlock(ptl);
5611 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5612 goto retry;
5613 }
5614 /*
5615 * hwpoisoned entry is treated as no_page_table in
5616 * follow_page_mask().
5617 */
5618 }
5619out:
5620 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01005621 return page;
5622}
5623
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08005624struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01005625follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005626 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01005627{
John Hubbard3faa52c2020-04-01 21:05:29 -07005628 if (flags & (FOLL_GET | FOLL_PIN))
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005629 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01005630
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005631 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01005632}
5633
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07005634struct page * __weak
5635follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5636{
John Hubbard3faa52c2020-04-01 21:05:29 -07005637 if (flags & (FOLL_GET | FOLL_PIN))
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07005638 return NULL;
5639
5640 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5641}
5642
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005643bool isolate_huge_page(struct page *page, struct list_head *list)
5644{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005645 bool ret = true;
5646
Mike Kravetzdb71ef72021-05-04 18:35:07 -07005647 spin_lock_irq(&hugetlb_lock);
Mike Kravetz8f251a32021-02-24 12:08:56 -08005648 if (!PageHeadHuge(page) ||
5649 !HPageMigratable(page) ||
Muchun Song0eb2df22021-02-04 18:32:10 -08005650 !get_page_unless_zero(page)) {
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005651 ret = false;
5652 goto unlock;
5653 }
Mike Kravetz8f251a32021-02-24 12:08:56 -08005654 ClearHPageMigratable(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005655 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005656unlock:
Mike Kravetzdb71ef72021-05-04 18:35:07 -07005657 spin_unlock_irq(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005658 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005659}
5660
5661void putback_active_hugepage(struct page *page)
5662{
Mike Kravetzdb71ef72021-05-04 18:35:07 -07005663 spin_lock_irq(&hugetlb_lock);
Mike Kravetz8f251a32021-02-24 12:08:56 -08005664 SetHPageMigratable(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005665 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
Mike Kravetzdb71ef72021-05-04 18:35:07 -07005666 spin_unlock_irq(&hugetlb_lock);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005667 put_page(page);
5668}
Michal Hockoab5ac902018-01-31 16:20:48 -08005669
5670void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5671{
5672 struct hstate *h = page_hstate(oldpage);
5673
5674 hugetlb_cgroup_migrate(oldpage, newpage);
5675 set_page_owner_migrate_reason(newpage, reason);
5676
5677 /*
5678 * transfer temporary state of the new huge page. This is
5679 * reverse to other transitions because the newpage is going to
5680 * be final while the old one will be freed so it takes over
5681 * the temporary status.
5682 *
5683 * Also note that we have to transfer the per-node surplus state
5684 * here as well otherwise the global surplus count will not match
5685 * the per-node's.
5686 */
Mike Kravetz9157c3112021-02-24 12:09:00 -08005687 if (HPageTemporary(newpage)) {
Michal Hockoab5ac902018-01-31 16:20:48 -08005688 int old_nid = page_to_nid(oldpage);
5689 int new_nid = page_to_nid(newpage);
5690
Mike Kravetz9157c3112021-02-24 12:09:00 -08005691 SetHPageTemporary(oldpage);
5692 ClearHPageTemporary(newpage);
Michal Hockoab5ac902018-01-31 16:20:48 -08005693
Miaohe Lin5af1ab12021-05-04 18:33:25 -07005694 /*
5695 * There is no need to transfer the per-node surplus state
5696 * when we do not cross the node.
5697 */
5698 if (new_nid == old_nid)
5699 return;
Mike Kravetzdb71ef72021-05-04 18:35:07 -07005700 spin_lock_irq(&hugetlb_lock);
Michal Hockoab5ac902018-01-31 16:20:48 -08005701 if (h->surplus_huge_pages_node[old_nid]) {
5702 h->surplus_huge_pages_node[old_nid]--;
5703 h->surplus_huge_pages_node[new_nid]++;
5704 }
Mike Kravetzdb71ef72021-05-04 18:35:07 -07005705 spin_unlock_irq(&hugetlb_lock);
Michal Hockoab5ac902018-01-31 16:20:48 -08005706 }
5707}
Roman Gushchincf11e852020-04-10 14:32:45 -07005708
Peter Xu6dfeaff2021-05-04 18:33:13 -07005709/*
5710 * This function will unconditionally remove all the shared pmd pgtable entries
5711 * within the specific vma for a hugetlbfs memory range.
5712 */
5713void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
5714{
5715 struct hstate *h = hstate_vma(vma);
5716 unsigned long sz = huge_page_size(h);
5717 struct mm_struct *mm = vma->vm_mm;
5718 struct mmu_notifier_range range;
5719 unsigned long address, start, end;
5720 spinlock_t *ptl;
5721 pte_t *ptep;
5722
5723 if (!(vma->vm_flags & VM_MAYSHARE))
5724 return;
5725
5726 start = ALIGN(vma->vm_start, PUD_SIZE);
5727 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5728
5729 if (start >= end)
5730 return;
5731
5732 /*
5733 * No need to call adjust_range_if_pmd_sharing_possible(), because
5734 * we have already done the PUD_SIZE alignment.
5735 */
5736 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
5737 start, end);
5738 mmu_notifier_invalidate_range_start(&range);
5739 i_mmap_lock_write(vma->vm_file->f_mapping);
5740 for (address = start; address < end; address += PUD_SIZE) {
5741 unsigned long tmp = address;
5742
5743 ptep = huge_pte_offset(mm, address, sz);
5744 if (!ptep)
5745 continue;
5746 ptl = huge_pte_lock(h, mm, ptep);
5747 /* We don't want 'address' to be changed */
5748 huge_pmd_unshare(mm, vma, &tmp, ptep);
5749 spin_unlock(ptl);
5750 }
5751 flush_hugetlb_tlb_range(vma, start, end);
5752 i_mmap_unlock_write(vma->vm_file->f_mapping);
5753 /*
5754 * No need to call mmu_notifier_invalidate_range(), see
5755 * Documentation/vm/mmu_notifier.rst.
5756 */
5757 mmu_notifier_invalidate_range_end(&range);
5758}
5759
Roman Gushchincf11e852020-04-10 14:32:45 -07005760#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -07005761static bool cma_reserve_called __initdata;
5762
5763static int __init cmdline_parse_hugetlb_cma(char *p)
5764{
5765 hugetlb_cma_size = memparse(p, &p);
5766 return 0;
5767}
5768
5769early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5770
5771void __init hugetlb_cma_reserve(int order)
5772{
5773 unsigned long size, reserved, per_node;
5774 int nid;
5775
5776 cma_reserve_called = true;
5777
5778 if (!hugetlb_cma_size)
5779 return;
5780
5781 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5782 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5783 (PAGE_SIZE << order) / SZ_1M);
5784 return;
5785 }
5786
5787 /*
5788 * If 3 GB area is requested on a machine with 4 numa nodes,
5789 * let's allocate 1 GB on first three nodes and ignore the last one.
5790 */
5791 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5792 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5793 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5794
5795 reserved = 0;
5796 for_each_node_state(nid, N_ONLINE) {
5797 int res;
Barry Song2281f792020-08-24 11:03:09 +12005798 char name[CMA_MAX_NAME];
Roman Gushchincf11e852020-04-10 14:32:45 -07005799
5800 size = min(per_node, hugetlb_cma_size - reserved);
5801 size = round_up(size, PAGE_SIZE << order);
5802
Barry Song2281f792020-08-24 11:03:09 +12005803 snprintf(name, sizeof(name), "hugetlb%d", nid);
Roman Gushchincf11e852020-04-10 14:32:45 -07005804 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
Barry Song29d0f412020-08-11 18:32:00 -07005805 0, false, name,
Roman Gushchincf11e852020-04-10 14:32:45 -07005806 &hugetlb_cma[nid], nid);
5807 if (res) {
5808 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5809 res, nid);
5810 continue;
5811 }
5812
5813 reserved += size;
5814 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5815 size / SZ_1M, nid);
5816
5817 if (reserved >= hugetlb_cma_size)
5818 break;
5819 }
5820}
5821
5822void __init hugetlb_cma_check(void)
5823{
5824 if (!hugetlb_cma_size || cma_reserve_called)
5825 return;
5826
5827 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5828}
5829
5830#endif /* CONFIG_CMA */