blob: e2919488fd123eafbfc44c6f5d748882d1a8bed3 [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
Muchun Songe334b1f2021-02-04 18:32:06 -080082static inline bool PageHugeFreed(struct page *head)
83{
84 return page_private(head + 4) == -1UL;
85}
86
87static inline void SetPageHugeFreed(struct page *head)
88{
89 set_page_private(head + 4, -1UL);
90}
91
92static inline void ClearPageHugeFreed(struct page *head)
93{
94 set_page_private(head + 4, 0);
95}
96
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070097/* Forward declaration */
98static int hugetlb_acct_memory(struct hstate *h, long delta);
99
David Gibson90481622012-03-21 16:34:12 -0700100static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
101{
102 bool free = (spool->count == 0) && (spool->used_hpages == 0);
103
104 spin_unlock(&spool->lock);
105
106 /* If no pages are used, and no other handles to the subpool
Ethon Paul7c8de352020-06-04 16:49:07 -0700107 * remain, give up any reservations based on minimum size and
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700108 * free the subpool */
109 if (free) {
110 if (spool->min_hpages != -1)
111 hugetlb_acct_memory(spool->hstate,
112 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -0700113 kfree(spool);
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700114 }
David Gibson90481622012-03-21 16:34:12 -0700115}
116
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700117struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
118 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -0700119{
120 struct hugepage_subpool *spool;
121
Mike Kravetzc6a91822015-04-15 16:13:36 -0700122 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -0700123 if (!spool)
124 return NULL;
125
126 spin_lock_init(&spool->lock);
127 spool->count = 1;
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700128 spool->max_hpages = max_hpages;
129 spool->hstate = h;
130 spool->min_hpages = min_hpages;
131
132 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
133 kfree(spool);
134 return NULL;
135 }
136 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700137
138 return spool;
139}
140
141void hugepage_put_subpool(struct hugepage_subpool *spool)
142{
143 spin_lock(&spool->lock);
144 BUG_ON(!spool->count);
145 spool->count--;
146 unlock_or_release_subpool(spool);
147}
148
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700149/*
150 * Subpool accounting for allocating and reserving pages.
151 * Return -ENOMEM if there are not enough resources to satisfy the
Randy Dunlap9e7ee402020-08-11 18:32:59 -0700152 * request. Otherwise, return the number of pages by which the
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700153 * global pools must be adjusted (upward). The returned value may
154 * only be different than the passed value (delta) in the case where
Ethon Paul7c8de352020-06-04 16:49:07 -0700155 * a subpool minimum size must be maintained.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700156 */
157static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700158 long delta)
159{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700160 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700161
162 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700163 return ret;
David Gibson90481622012-03-21 16:34:12 -0700164
165 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700166
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700167 if (spool->max_hpages != -1) { /* maximum size accounting */
168 if ((spool->used_hpages + delta) <= spool->max_hpages)
169 spool->used_hpages += delta;
170 else {
171 ret = -ENOMEM;
172 goto unlock_ret;
173 }
174 }
175
Mike Kravetz09a95e22016-05-19 17:11:01 -0700176 /* minimum size accounting */
177 if (spool->min_hpages != -1 && spool->rsv_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700178 if (delta > spool->rsv_hpages) {
179 /*
180 * Asking for more reserves than those already taken on
181 * behalf of subpool. Return difference.
182 */
183 ret = delta - spool->rsv_hpages;
184 spool->rsv_hpages = 0;
185 } else {
186 ret = 0; /* reserves already accounted for */
187 spool->rsv_hpages -= delta;
188 }
189 }
190
191unlock_ret:
192 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700193 return ret;
194}
195
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700196/*
197 * Subpool accounting for freeing and unreserving pages.
198 * Return the number of global page reservations that must be dropped.
199 * The return value may only be different than the passed value (delta)
200 * in the case where a subpool minimum size must be maintained.
201 */
202static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700203 long delta)
204{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700205 long ret = delta;
206
David Gibson90481622012-03-21 16:34:12 -0700207 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700208 return delta;
David Gibson90481622012-03-21 16:34:12 -0700209
210 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700211
212 if (spool->max_hpages != -1) /* maximum size accounting */
213 spool->used_hpages -= delta;
214
Mike Kravetz09a95e22016-05-19 17:11:01 -0700215 /* minimum size accounting */
216 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700217 if (spool->rsv_hpages + delta <= spool->min_hpages)
218 ret = 0;
219 else
220 ret = spool->rsv_hpages + delta - spool->min_hpages;
221
222 spool->rsv_hpages += delta;
223 if (spool->rsv_hpages > spool->min_hpages)
224 spool->rsv_hpages = spool->min_hpages;
225 }
226
227 /*
228 * If hugetlbfs_put_super couldn't free spool due to an outstanding
229 * quota reference, free it now.
230 */
David Gibson90481622012-03-21 16:34:12 -0700231 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700232
233 return ret;
David Gibson90481622012-03-21 16:34:12 -0700234}
235
236static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
237{
238 return HUGETLBFS_SB(inode->i_sb)->spool;
239}
240
241static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
242{
Al Viro496ad9a2013-01-23 17:07:38 -0500243 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700244}
245
Mina Almasry0db9d742020-04-01 21:11:25 -0700246/* Helper that removes a struct file_region from the resv_map cache and returns
247 * it for use.
248 */
249static struct file_region *
250get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
251{
252 struct file_region *nrg = NULL;
253
254 VM_BUG_ON(resv->region_cache_count <= 0);
255
256 resv->region_cache_count--;
257 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
Mina Almasry0db9d742020-04-01 21:11:25 -0700258 list_del(&nrg->link);
259
260 nrg->from = from;
261 nrg->to = to;
262
263 return nrg;
264}
265
Mina Almasry075a61d2020-04-01 21:11:28 -0700266static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
267 struct file_region *rg)
268{
269#ifdef CONFIG_CGROUP_HUGETLB
270 nrg->reservation_counter = rg->reservation_counter;
271 nrg->css = rg->css;
272 if (rg->css)
273 css_get(rg->css);
274#endif
275}
276
277/* Helper that records hugetlb_cgroup uncharge info. */
278static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
279 struct hstate *h,
280 struct resv_map *resv,
281 struct file_region *nrg)
282{
283#ifdef CONFIG_CGROUP_HUGETLB
284 if (h_cg) {
285 nrg->reservation_counter =
286 &h_cg->rsvd_hugepage[hstate_index(h)];
287 nrg->css = &h_cg->css;
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700288 /*
289 * The caller will hold exactly one h_cg->css reference for the
290 * whole contiguous reservation region. But this area might be
291 * scattered when there are already some file_regions reside in
292 * it. As a result, many file_regions may share only one css
293 * reference. In order to ensure that one file_region must hold
294 * exactly one h_cg->css reference, we should do css_get for
295 * each file_region and leave the reference held by caller
296 * untouched.
297 */
298 css_get(&h_cg->css);
Mina Almasry075a61d2020-04-01 21:11:28 -0700299 if (!resv->pages_per_hpage)
300 resv->pages_per_hpage = pages_per_huge_page(h);
301 /* pages_per_hpage should be the same for all entries in
302 * a resv_map.
303 */
304 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
305 } else {
306 nrg->reservation_counter = NULL;
307 nrg->css = NULL;
308 }
309#endif
310}
311
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700312static void put_uncharge_info(struct file_region *rg)
313{
314#ifdef CONFIG_CGROUP_HUGETLB
315 if (rg->css)
316 css_put(rg->css);
317#endif
318}
319
Mina Almasrya9b3f862020-04-01 21:11:35 -0700320static bool has_same_uncharge_info(struct file_region *rg,
321 struct file_region *org)
322{
323#ifdef CONFIG_CGROUP_HUGETLB
324 return rg && org &&
325 rg->reservation_counter == org->reservation_counter &&
326 rg->css == org->css;
327
328#else
329 return true;
330#endif
331}
332
333static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
334{
335 struct file_region *nrg = NULL, *prg = NULL;
336
337 prg = list_prev_entry(rg, link);
338 if (&prg->link != &resv->regions && prg->to == rg->from &&
339 has_same_uncharge_info(prg, rg)) {
340 prg->to = rg->to;
341
342 list_del(&rg->link);
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700343 put_uncharge_info(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700344 kfree(rg);
345
Wei Yang7db5e7b2020-10-13 16:56:20 -0700346 rg = prg;
Mina Almasrya9b3f862020-04-01 21:11:35 -0700347 }
348
349 nrg = list_next_entry(rg, link);
350 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
351 has_same_uncharge_info(nrg, rg)) {
352 nrg->from = rg->from;
353
354 list_del(&rg->link);
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700355 put_uncharge_info(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700356 kfree(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700357 }
358}
359
Wei Yang972a3da32020-10-13 16:56:30 -0700360/*
361 * Must be called with resv->lock held.
362 *
363 * Calling this with regions_needed != NULL will count the number of pages
364 * to be added but will not modify the linked list. And regions_needed will
365 * indicate the number of file_regions needed in the cache to carry out to add
366 * the regions for this range.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800367 */
368static long add_reservation_in_range(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700369 struct hugetlb_cgroup *h_cg,
Wei Yang972a3da32020-10-13 16:56:30 -0700370 struct hstate *h, long *regions_needed)
Mina Almasryd75c6af2019-11-30 17:56:59 -0800371{
Mina Almasry0db9d742020-04-01 21:11:25 -0700372 long add = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800373 struct list_head *head = &resv->regions;
Mina Almasry0db9d742020-04-01 21:11:25 -0700374 long last_accounted_offset = f;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800375 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
376
Mina Almasry0db9d742020-04-01 21:11:25 -0700377 if (regions_needed)
378 *regions_needed = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800379
Mina Almasry0db9d742020-04-01 21:11:25 -0700380 /* In this loop, we essentially handle an entry for the range
381 * [last_accounted_offset, rg->from), at every iteration, with some
382 * bounds checking.
383 */
384 list_for_each_entry_safe(rg, trg, head, link) {
385 /* Skip irrelevant regions that start before our range. */
386 if (rg->from < f) {
387 /* If this region ends after the last accounted offset,
388 * then we need to update last_accounted_offset.
389 */
390 if (rg->to > last_accounted_offset)
391 last_accounted_offset = rg->to;
392 continue;
393 }
Mina Almasryd75c6af2019-11-30 17:56:59 -0800394
Mina Almasry0db9d742020-04-01 21:11:25 -0700395 /* When we find a region that starts beyond our range, we've
396 * finished.
397 */
Mina Almasryd75c6af2019-11-30 17:56:59 -0800398 if (rg->from > t)
399 break;
400
Mina Almasry0db9d742020-04-01 21:11:25 -0700401 /* Add an entry for last_accounted_offset -> rg->from, and
402 * update last_accounted_offset.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800403 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700404 if (rg->from > last_accounted_offset) {
405 add += rg->from - last_accounted_offset;
Wei Yang972a3da32020-10-13 16:56:30 -0700406 if (!regions_needed) {
Mina Almasry0db9d742020-04-01 21:11:25 -0700407 nrg = get_file_region_entry_from_cache(
408 resv, last_accounted_offset, rg->from);
Mina Almasry075a61d2020-04-01 21:11:28 -0700409 record_hugetlb_cgroup_uncharge_info(h_cg, h,
410 resv, nrg);
Mina Almasry0db9d742020-04-01 21:11:25 -0700411 list_add(&nrg->link, rg->link.prev);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700412 coalesce_file_region(resv, nrg);
Wei Yang972a3da32020-10-13 16:56:30 -0700413 } else
Mina Almasry0db9d742020-04-01 21:11:25 -0700414 *regions_needed += 1;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800415 }
Mina Almasryd75c6af2019-11-30 17:56:59 -0800416
Mina Almasry0db9d742020-04-01 21:11:25 -0700417 last_accounted_offset = rg->to;
418 }
419
420 /* Handle the case where our range extends beyond
421 * last_accounted_offset.
422 */
423 if (last_accounted_offset < t) {
424 add += t - last_accounted_offset;
Wei Yang972a3da32020-10-13 16:56:30 -0700425 if (!regions_needed) {
Mina Almasry0db9d742020-04-01 21:11:25 -0700426 nrg = get_file_region_entry_from_cache(
427 resv, last_accounted_offset, t);
Mina Almasry075a61d2020-04-01 21:11:28 -0700428 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
Mina Almasry0db9d742020-04-01 21:11:25 -0700429 list_add(&nrg->link, rg->link.prev);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700430 coalesce_file_region(resv, nrg);
Wei Yang972a3da32020-10-13 16:56:30 -0700431 } else
Mina Almasry0db9d742020-04-01 21:11:25 -0700432 *regions_needed += 1;
433 }
434
435 VM_BUG_ON(add < 0);
436 return add;
437}
438
439/* Must be called with resv->lock acquired. Will drop lock to allocate entries.
440 */
441static int allocate_file_region_entries(struct resv_map *resv,
442 int regions_needed)
443 __must_hold(&resv->lock)
444{
445 struct list_head allocated_regions;
446 int to_allocate = 0, i = 0;
447 struct file_region *trg = NULL, *rg = NULL;
448
449 VM_BUG_ON(regions_needed < 0);
450
451 INIT_LIST_HEAD(&allocated_regions);
452
453 /*
454 * Check for sufficient descriptors in the cache to accommodate
455 * the number of in progress add operations plus regions_needed.
456 *
457 * This is a while loop because when we drop the lock, some other call
458 * to region_add or region_del may have consumed some region_entries,
459 * so we keep looping here until we finally have enough entries for
460 * (adds_in_progress + regions_needed).
461 */
462 while (resv->region_cache_count <
463 (resv->adds_in_progress + regions_needed)) {
464 to_allocate = resv->adds_in_progress + regions_needed -
465 resv->region_cache_count;
466
467 /* At this point, we should have enough entries in the cache
468 * for all the existings adds_in_progress. We should only be
469 * needing to allocate for regions_needed.
470 */
471 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
472
473 spin_unlock(&resv->lock);
474 for (i = 0; i < to_allocate; i++) {
475 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
476 if (!trg)
477 goto out_of_memory;
478 list_add(&trg->link, &allocated_regions);
479 }
480
481 spin_lock(&resv->lock);
482
Wei Yangd3ec7b62020-10-13 16:56:27 -0700483 list_splice(&allocated_regions, &resv->region_cache);
484 resv->region_cache_count += to_allocate;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800485 }
486
Mina Almasry0db9d742020-04-01 21:11:25 -0700487 return 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800488
Mina Almasry0db9d742020-04-01 21:11:25 -0700489out_of_memory:
490 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
491 list_del(&rg->link);
492 kfree(rg);
493 }
494 return -ENOMEM;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800495}
496
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700497/*
498 * Add the huge page range represented by [f, t) to the reserve
Mina Almasry0db9d742020-04-01 21:11:25 -0700499 * map. Regions will be taken from the cache to fill in this range.
500 * Sufficient regions should exist in the cache due to the previous
501 * call to region_chg with the same range, but in some cases the cache will not
502 * have sufficient entries due to races with other code doing region_add or
503 * region_del. The extra needed entries will be allocated.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700504 *
Mina Almasry0db9d742020-04-01 21:11:25 -0700505 * regions_needed is the out value provided by a previous call to region_chg.
506 *
507 * Return the number of new huge pages added to the map. This number is greater
508 * than or equal to zero. If file_region entries needed to be allocated for
Ethon Paul7c8de352020-06-04 16:49:07 -0700509 * this operation and we were not able to allocate, it returns -ENOMEM.
Mina Almasry0db9d742020-04-01 21:11:25 -0700510 * region_add of regions of length 1 never allocate file_regions and cannot
511 * fail; region_chg will always allocate at least 1 entry and a region_add for
512 * 1 page will only require at most 1 entry.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700513 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700514static long region_add(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700515 long in_regions_needed, struct hstate *h,
516 struct hugetlb_cgroup *h_cg)
Andy Whitcroft96822902008-07-23 21:27:29 -0700517{
Mina Almasry0db9d742020-04-01 21:11:25 -0700518 long add = 0, actual_regions_needed = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700519
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700520 spin_lock(&resv->lock);
Mina Almasry0db9d742020-04-01 21:11:25 -0700521retry:
522
523 /* Count how many regions are actually needed to execute this add. */
Wei Yang972a3da32020-10-13 16:56:30 -0700524 add_reservation_in_range(resv, f, t, NULL, NULL,
525 &actual_regions_needed);
Andy Whitcroft96822902008-07-23 21:27:29 -0700526
Mike Kravetz5e911372015-09-08 15:01:28 -0700527 /*
Mina Almasry0db9d742020-04-01 21:11:25 -0700528 * Check for sufficient descriptors in the cache to accommodate
529 * this add operation. Note that actual_regions_needed may be greater
530 * than in_regions_needed, as the resv_map may have been modified since
531 * the region_chg call. In this case, we need to make sure that we
532 * allocate extra entries, such that we have enough for all the
533 * existing adds_in_progress, plus the excess needed for this
534 * operation.
Mike Kravetz5e911372015-09-08 15:01:28 -0700535 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700536 if (actual_regions_needed > in_regions_needed &&
537 resv->region_cache_count <
538 resv->adds_in_progress +
539 (actual_regions_needed - in_regions_needed)) {
540 /* region_add operation of range 1 should never need to
541 * allocate file_region entries.
542 */
543 VM_BUG_ON(t - f <= 1);
Mike Kravetz5e911372015-09-08 15:01:28 -0700544
Mina Almasry0db9d742020-04-01 21:11:25 -0700545 if (allocate_file_region_entries(
546 resv, actual_regions_needed - in_regions_needed)) {
547 return -ENOMEM;
548 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700549
Mina Almasry0db9d742020-04-01 21:11:25 -0700550 goto retry;
Mike Kravetz5e911372015-09-08 15:01:28 -0700551 }
552
Wei Yang972a3da32020-10-13 16:56:30 -0700553 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700554
Mina Almasry0db9d742020-04-01 21:11:25 -0700555 resv->adds_in_progress -= in_regions_needed;
556
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700557 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700558 VM_BUG_ON(add < 0);
559 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700560}
561
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700562/*
563 * Examine the existing reserve map and determine how many
564 * huge pages in the specified range [f, t) are NOT currently
565 * represented. This routine is called before a subsequent
566 * call to region_add that will actually modify the reserve
567 * map to add the specified range [f, t). region_chg does
568 * not change the number of huge pages represented by the
Mina Almasry0db9d742020-04-01 21:11:25 -0700569 * map. A number of new file_region structures is added to the cache as a
570 * placeholder, for the subsequent region_add call to use. At least 1
571 * file_region structure is added.
572 *
573 * out_regions_needed is the number of regions added to the
574 * resv->adds_in_progress. This value needs to be provided to a follow up call
575 * to region_add or region_abort for proper accounting.
Mike Kravetz5e911372015-09-08 15:01:28 -0700576 *
577 * Returns the number of huge pages that need to be added to the existing
578 * reservation map for the range [f, t). This number is greater or equal to
579 * zero. -ENOMEM is returned if a new file_region structure or cache entry
580 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700581 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700582static long region_chg(struct resv_map *resv, long f, long t,
583 long *out_regions_needed)
Andy Whitcroft96822902008-07-23 21:27:29 -0700584{
Andy Whitcroft96822902008-07-23 21:27:29 -0700585 long chg = 0;
586
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700587 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700588
Wei Yang972a3da32020-10-13 16:56:30 -0700589 /* Count how many hugepages in this range are NOT represented. */
Mina Almasry075a61d2020-04-01 21:11:28 -0700590 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
Wei Yang972a3da32020-10-13 16:56:30 -0700591 out_regions_needed);
Mike Kravetz5e911372015-09-08 15:01:28 -0700592
Mina Almasry0db9d742020-04-01 21:11:25 -0700593 if (*out_regions_needed == 0)
594 *out_regions_needed = 1;
Mike Kravetz5e911372015-09-08 15:01:28 -0700595
Mina Almasry0db9d742020-04-01 21:11:25 -0700596 if (allocate_file_region_entries(resv, *out_regions_needed))
597 return -ENOMEM;
Mike Kravetz5e911372015-09-08 15:01:28 -0700598
Mina Almasry0db9d742020-04-01 21:11:25 -0700599 resv->adds_in_progress += *out_regions_needed;
Andy Whitcroft96822902008-07-23 21:27:29 -0700600
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700601 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700602 return chg;
603}
604
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700605/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700606 * Abort the in progress add operation. The adds_in_progress field
607 * of the resv_map keeps track of the operations in progress between
608 * calls to region_chg and region_add. Operations are sometimes
609 * aborted after the call to region_chg. In such cases, region_abort
Mina Almasry0db9d742020-04-01 21:11:25 -0700610 * is called to decrement the adds_in_progress counter. regions_needed
611 * is the value returned by the region_chg call, it is used to decrement
612 * the adds_in_progress counter.
Mike Kravetz5e911372015-09-08 15:01:28 -0700613 *
614 * NOTE: The range arguments [f, t) are not needed or used in this
615 * routine. They are kept to make reading the calling code easier as
616 * arguments will match the associated region_chg call.
617 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700618static void region_abort(struct resv_map *resv, long f, long t,
619 long regions_needed)
Mike Kravetz5e911372015-09-08 15:01:28 -0700620{
621 spin_lock(&resv->lock);
622 VM_BUG_ON(!resv->region_cache_count);
Mina Almasry0db9d742020-04-01 21:11:25 -0700623 resv->adds_in_progress -= regions_needed;
Mike Kravetz5e911372015-09-08 15:01:28 -0700624 spin_unlock(&resv->lock);
625}
626
627/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700628 * Delete the specified range [f, t) from the reserve map. If the
629 * t parameter is LONG_MAX, this indicates that ALL regions after f
630 * should be deleted. Locate the regions which intersect [f, t)
631 * and either trim, delete or split the existing regions.
632 *
633 * Returns the number of huge pages deleted from the reserve map.
634 * In the normal case, the return value is zero or more. In the
635 * case where a region must be split, a new region descriptor must
636 * be allocated. If the allocation fails, -ENOMEM will be returned.
637 * NOTE: If the parameter t == LONG_MAX, then we will never split
638 * a region and possibly return -ENOMEM. Callers specifying
639 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700640 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700641static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700642{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700643 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700644 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700645 struct file_region *nrg = NULL;
646 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700647
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700648retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700649 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700650 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800651 /*
652 * Skip regions before the range to be deleted. file_region
653 * ranges are normally of the form [from, to). However, there
654 * may be a "placeholder" entry in the map which is of the form
655 * (from, to) with from == to. Check for placeholder entries
656 * at the beginning of the range to be deleted.
657 */
658 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700659 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800660
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700661 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700662 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700663
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700664 if (f > rg->from && t < rg->to) { /* Must split region */
665 /*
666 * Check for an entry in the cache before dropping
667 * lock and attempting allocation.
668 */
669 if (!nrg &&
670 resv->region_cache_count > resv->adds_in_progress) {
671 nrg = list_first_entry(&resv->region_cache,
672 struct file_region,
673 link);
674 list_del(&nrg->link);
675 resv->region_cache_count--;
676 }
677
678 if (!nrg) {
679 spin_unlock(&resv->lock);
680 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
681 if (!nrg)
682 return -ENOMEM;
683 goto retry;
684 }
685
686 del += t - f;
Mike Kravetz79aa9252020-11-01 17:07:27 -0800687 hugetlb_cgroup_uncharge_file_region(
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700688 resv, rg, t - f, false);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700689
690 /* New entry for end of split region */
691 nrg->from = t;
692 nrg->to = rg->to;
Mina Almasry075a61d2020-04-01 21:11:28 -0700693
694 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
695
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700696 INIT_LIST_HEAD(&nrg->link);
697
698 /* Original entry is trimmed */
699 rg->to = f;
700
701 list_add(&nrg->link, &rg->link);
702 nrg = NULL;
703 break;
704 }
705
706 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
707 del += rg->to - rg->from;
Mina Almasry075a61d2020-04-01 21:11:28 -0700708 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700709 rg->to - rg->from, true);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700710 list_del(&rg->link);
711 kfree(rg);
712 continue;
713 }
714
715 if (f <= rg->from) { /* Trim beginning of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700716 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700717 t - rg->from, false);
Mina Almasry075a61d2020-04-01 21:11:28 -0700718
Mike Kravetz79aa9252020-11-01 17:07:27 -0800719 del += t - rg->from;
720 rg->from = t;
721 } else { /* Trim end of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700722 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700723 rg->to - f, false);
Mike Kravetz79aa9252020-11-01 17:07:27 -0800724
725 del += rg->to - f;
726 rg->to = f;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700727 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700728 }
729
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700730 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700731 kfree(nrg);
732 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700733}
734
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700735/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700736 * A rare out of memory error was encountered which prevented removal of
737 * the reserve map region for a page. The huge page itself was free'ed
738 * and removed from the page cache. This routine will adjust the subpool
739 * usage count, and the global reserve count if needed. By incrementing
740 * these counts, the reserve map entry which could not be deleted will
741 * appear as a "reserved" entry instead of simply dangling with incorrect
742 * counts.
743 */
zhong jiang72e29362016-10-07 17:02:01 -0700744void hugetlb_fix_reserve_counts(struct inode *inode)
Mike Kravetzb5cec282015-09-08 15:01:41 -0700745{
746 struct hugepage_subpool *spool = subpool_inode(inode);
747 long rsv_adjust;
748
749 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
zhong jiang72e29362016-10-07 17:02:01 -0700750 if (rsv_adjust) {
Mike Kravetzb5cec282015-09-08 15:01:41 -0700751 struct hstate *h = hstate_inode(inode);
752
753 hugetlb_acct_memory(h, 1);
754 }
755}
756
757/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700758 * Count and return the number of huge pages in the reserve map
759 * that intersect with the range [f, t).
760 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700761static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700762{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700763 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700764 struct file_region *rg;
765 long chg = 0;
766
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700767 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700768 /* Locate each segment we overlap with, and count that overlap. */
769 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700770 long seg_from;
771 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700772
773 if (rg->to <= f)
774 continue;
775 if (rg->from >= t)
776 break;
777
778 seg_from = max(rg->from, f);
779 seg_to = min(rg->to, t);
780
781 chg += seg_to - seg_from;
782 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700783 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700784
785 return chg;
786}
787
Andy Whitcroft96822902008-07-23 21:27:29 -0700788/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700789 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700790 * the mapping, in pagecache page units; huge pages here.
791 */
Andi Kleena5516432008-07-23 21:27:41 -0700792static pgoff_t vma_hugecache_offset(struct hstate *h,
793 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700794{
Andi Kleena5516432008-07-23 21:27:41 -0700795 return ((address - vma->vm_start) >> huge_page_shift(h)) +
796 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700797}
798
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900799pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
800 unsigned long address)
801{
802 return vma_hugecache_offset(hstate_vma(vma), vma, address);
803}
Dan Williamsdee41072016-05-14 12:20:44 -0700804EXPORT_SYMBOL_GPL(linear_hugepage_index);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900805
Andy Whitcroft84afd992008-07-23 21:27:32 -0700806/*
Mel Gorman08fba692009-01-06 14:38:53 -0800807 * Return the size of the pages allocated when backing a VMA. In the majority
808 * cases this will be same size as used by the page table entries.
809 */
810unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
811{
Dan Williams05ea8862018-04-05 16:24:25 -0700812 if (vma->vm_ops && vma->vm_ops->pagesize)
813 return vma->vm_ops->pagesize(vma);
814 return PAGE_SIZE;
Mel Gorman08fba692009-01-06 14:38:53 -0800815}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200816EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800817
818/*
Mel Gorman33402892009-01-06 14:38:54 -0800819 * Return the page size being used by the MMU to back a VMA. In the majority
820 * of cases, the page size used by the kernel matches the MMU size. On
Dan Williams09135cc2018-04-05 16:24:21 -0700821 * architectures where it differs, an architecture-specific 'strong'
822 * version of this symbol is required.
Mel Gorman33402892009-01-06 14:38:54 -0800823 */
Dan Williams09135cc2018-04-05 16:24:21 -0700824__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
Mel Gorman33402892009-01-06 14:38:54 -0800825{
826 return vma_kernel_pagesize(vma);
827}
Mel Gorman33402892009-01-06 14:38:54 -0800828
829/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700830 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
831 * bits of the reservation map pointer, which are always clear due to
832 * alignment.
833 */
834#define HPAGE_RESV_OWNER (1UL << 0)
835#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700836#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700837
Mel Gormana1e78772008-07-23 21:27:23 -0700838/*
839 * These helpers are used to track how many pages are reserved for
840 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
841 * is guaranteed to have their future faults succeed.
842 *
843 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
844 * the reserve counters are updated with the hugetlb_lock held. It is safe
845 * to reset the VMA at fork() time as it is not in use yet and there is no
846 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700847 *
848 * The private mapping reservation is represented in a subtly different
849 * manner to a shared mapping. A shared mapping has a region map associated
850 * with the underlying file, this region map represents the backing file
851 * pages which have ever had a reservation assigned which this persists even
852 * after the page is instantiated. A private mapping has a region map
853 * associated with the original mmap which is attached to all VMAs which
854 * reference it, this region map represents those offsets which have consumed
855 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700856 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700857static unsigned long get_vma_private_data(struct vm_area_struct *vma)
858{
859 return (unsigned long)vma->vm_private_data;
860}
861
862static void set_vma_private_data(struct vm_area_struct *vma,
863 unsigned long value)
864{
865 vma->vm_private_data = (void *)value;
866}
867
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700868static void
869resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
870 struct hugetlb_cgroup *h_cg,
871 struct hstate *h)
872{
873#ifdef CONFIG_CGROUP_HUGETLB
874 if (!h_cg || !h) {
875 resv_map->reservation_counter = NULL;
876 resv_map->pages_per_hpage = 0;
877 resv_map->css = NULL;
878 } else {
879 resv_map->reservation_counter =
880 &h_cg->rsvd_hugepage[hstate_index(h)];
881 resv_map->pages_per_hpage = pages_per_huge_page(h);
882 resv_map->css = &h_cg->css;
883 }
884#endif
885}
886
Joonsoo Kim9119a412014-04-03 14:47:25 -0700887struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700888{
889 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700890 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
891
892 if (!resv_map || !rg) {
893 kfree(resv_map);
894 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700895 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700896 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700897
898 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700899 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700900 INIT_LIST_HEAD(&resv_map->regions);
901
Mike Kravetz5e911372015-09-08 15:01:28 -0700902 resv_map->adds_in_progress = 0;
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700903 /*
904 * Initialize these to 0. On shared mappings, 0's here indicate these
905 * fields don't do cgroup accounting. On private mappings, these will be
906 * re-initialized to the proper values, to indicate that hugetlb cgroup
907 * reservations are to be un-charged from here.
908 */
909 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700910
911 INIT_LIST_HEAD(&resv_map->region_cache);
912 list_add(&rg->link, &resv_map->region_cache);
913 resv_map->region_cache_count = 1;
914
Andy Whitcroft84afd992008-07-23 21:27:32 -0700915 return resv_map;
916}
917
Joonsoo Kim9119a412014-04-03 14:47:25 -0700918void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700919{
920 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700921 struct list_head *head = &resv_map->region_cache;
922 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700923
924 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700925 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700926
927 /* ... and any entries left in the cache */
928 list_for_each_entry_safe(rg, trg, head, link) {
929 list_del(&rg->link);
930 kfree(rg);
931 }
932
933 VM_BUG_ON(resv_map->adds_in_progress);
934
Andy Whitcroft84afd992008-07-23 21:27:32 -0700935 kfree(resv_map);
936}
937
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700938static inline struct resv_map *inode_resv_map(struct inode *inode)
939{
Mike Kravetzf27a5132019-05-13 17:22:55 -0700940 /*
941 * At inode evict time, i_mapping may not point to the original
942 * address space within the inode. This original address space
943 * contains the pointer to the resv_map. So, always use the
944 * address space embedded within the inode.
945 * The VERY common case is inode->mapping == &inode->i_data but,
946 * this may not be true for device special inodes.
947 */
948 return (struct resv_map *)(&inode->i_data)->private_data;
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700949}
950
Andy Whitcroft84afd992008-07-23 21:27:32 -0700951static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700952{
Sasha Levin81d1b092014-10-09 15:28:10 -0700953 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700954 if (vma->vm_flags & VM_MAYSHARE) {
955 struct address_space *mapping = vma->vm_file->f_mapping;
956 struct inode *inode = mapping->host;
957
958 return inode_resv_map(inode);
959
960 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700961 return (struct resv_map *)(get_vma_private_data(vma) &
962 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700963 }
Mel Gormana1e78772008-07-23 21:27:23 -0700964}
965
Andy Whitcroft84afd992008-07-23 21:27:32 -0700966static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700967{
Sasha Levin81d1b092014-10-09 15:28:10 -0700968 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
969 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700970
Andy Whitcroft84afd992008-07-23 21:27:32 -0700971 set_vma_private_data(vma, (get_vma_private_data(vma) &
972 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700973}
974
975static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
976{
Sasha Levin81d1b092014-10-09 15:28:10 -0700977 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
978 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700979
980 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700981}
982
983static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
984{
Sasha Levin81d1b092014-10-09 15:28:10 -0700985 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700986
987 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700988}
989
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700990/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700991void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
992{
Sasha Levin81d1b092014-10-09 15:28:10 -0700993 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700994 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700995 vma->vm_private_data = (void *)0;
996}
997
998/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700999static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001000{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001001 if (vma->vm_flags & VM_NORESERVE) {
1002 /*
1003 * This address is already reserved by other process(chg == 0),
1004 * so, we should decrement reserved count. Without decrementing,
1005 * reserve count remains after releasing inode, because this
1006 * allocated page will go into page cache and is regarded as
1007 * coming from reserved pool in releasing step. Currently, we
1008 * don't have any other solution to deal with this situation
1009 * properly, so add work-around here.
1010 */
1011 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001012 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001013 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001014 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001015 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001016
1017 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -07001018 if (vma->vm_flags & VM_MAYSHARE) {
1019 /*
1020 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1021 * be a region map for all pages. The only situation where
1022 * there is no region map is if a hole was punched via
Ethon Paul7c8de352020-06-04 16:49:07 -07001023 * fallocate. In this case, there really are no reserves to
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -07001024 * use. This situation is indicated if chg != 0.
1025 */
1026 if (chg)
1027 return false;
1028 else
1029 return true;
1030 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001031
1032 /*
1033 * Only the process that called mmap() has reserves for
1034 * private mappings.
1035 */
Mike Kravetz67961f92016-06-08 15:33:42 -07001036 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1037 /*
1038 * Like the shared case above, a hole punch or truncate
1039 * could have been performed on the private mapping.
1040 * Examine the value of chg to determine if reserves
1041 * actually exist or were previously consumed.
1042 * Very Subtle - The value of chg comes from a previous
1043 * call to vma_needs_reserves(). The reserve map for
1044 * private mappings has different (opposite) semantics
1045 * than that of shared mappings. vma_needs_reserves()
1046 * has already taken this difference in semantics into
1047 * account. Therefore, the meaning of chg is the same
1048 * as in the shared case above. Code could easily be
1049 * combined, but keeping it separate draws attention to
1050 * subtle differences.
1051 */
1052 if (chg)
1053 return false;
1054 else
1055 return true;
1056 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001057
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001058 return false;
Mel Gormana1e78772008-07-23 21:27:23 -07001059}
1060
Andi Kleena5516432008-07-23 21:27:41 -07001061static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062{
1063 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001064 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -07001065 h->free_huge_pages++;
1066 h->free_huge_pages_node[nid]++;
Muchun Songe334b1f2021-02-04 18:32:06 -08001067 SetPageHugeFreed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068}
1069
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001070static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001071{
1072 struct page *page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001073 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001074
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001075 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1076 if (nocma && is_migrate_cma_page(page))
1077 continue;
1078
Wei Yang6664bfc2020-10-13 16:56:39 -07001079 if (PageHWPoison(page))
1080 continue;
1081
1082 list_move(&page->lru, &h->hugepage_activelist);
1083 set_page_refcounted(page);
Muchun Songe334b1f2021-02-04 18:32:06 -08001084 ClearPageHugeFreed(page);
Wei Yang6664bfc2020-10-13 16:56:39 -07001085 h->free_huge_pages--;
1086 h->free_huge_pages_node[nid]--;
1087 return page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001088 }
1089
Wei Yang6664bfc2020-10-13 16:56:39 -07001090 return NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001091}
1092
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001093static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1094 nodemask_t *nmask)
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001095{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001096 unsigned int cpuset_mems_cookie;
1097 struct zonelist *zonelist;
1098 struct zone *zone;
1099 struct zoneref *z;
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001100 int node = NUMA_NO_NODE;
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001101
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001102 zonelist = node_zonelist(nid, gfp_mask);
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001103
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001104retry_cpuset:
1105 cpuset_mems_cookie = read_mems_allowed_begin();
1106 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1107 struct page *page;
1108
1109 if (!cpuset_zone_allowed(zone, gfp_mask))
1110 continue;
1111 /*
1112 * no need to ask again on the same node. Pool is node rather than
1113 * zone aware
1114 */
1115 if (zone_to_nid(zone) == node)
1116 continue;
1117 node = zone_to_nid(zone);
1118
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001119 page = dequeue_huge_page_node_exact(h, node);
1120 if (page)
1121 return page;
1122 }
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001123 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1124 goto retry_cpuset;
1125
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001126 return NULL;
1127}
1128
Andi Kleena5516432008-07-23 21:27:41 -07001129static struct page *dequeue_huge_page_vma(struct hstate *h,
1130 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001131 unsigned long address, int avoid_reserve,
1132 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001134 struct page *page;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001135 struct mempolicy *mpol;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001136 gfp_t gfp_mask;
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001137 nodemask_t *nodemask;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001138 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139
Mel Gormana1e78772008-07-23 21:27:23 -07001140 /*
1141 * A child process with MAP_PRIVATE mappings created by their parent
1142 * have no page reserves. This check ensures that reservations are
1143 * not "stolen". The child may still get SIGKILLed
1144 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001145 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -07001146 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -07001147 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -07001148
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001149 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -07001150 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -07001151 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001152
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001153 gfp_mask = htlb_alloc_mask(h);
1154 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001155 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1156 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1157 SetPagePrivate(page);
1158 h->resv_huge_pages--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07001160
1161 mpol_cond_put(mpol);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001162 return page;
1163
Miao Xiec0ff7452010-05-24 14:32:08 -07001164err:
Mel Gormancc9a6c82012-03-21 16:34:11 -07001165 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166}
1167
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001168/*
1169 * common helper functions for hstate_next_node_to_{alloc|free}.
1170 * We may have allocated or freed a huge page based on a different
1171 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1172 * be outside of *nodes_allowed. Ensure that we use an allowed
1173 * node for alloc or free.
1174 */
1175static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1176{
Andrew Morton0edaf862016-05-19 17:10:58 -07001177 nid = next_node_in(nid, *nodes_allowed);
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001178 VM_BUG_ON(nid >= MAX_NUMNODES);
1179
1180 return nid;
1181}
1182
1183static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1184{
1185 if (!node_isset(nid, *nodes_allowed))
1186 nid = next_node_allowed(nid, nodes_allowed);
1187 return nid;
1188}
1189
1190/*
1191 * returns the previously saved node ["this node"] from which to
1192 * allocate a persistent huge page for the pool and advance the
1193 * next node from which to allocate, handling wrap at end of node
1194 * mask.
1195 */
1196static int hstate_next_node_to_alloc(struct hstate *h,
1197 nodemask_t *nodes_allowed)
1198{
1199 int nid;
1200
1201 VM_BUG_ON(!nodes_allowed);
1202
1203 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1204 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1205
1206 return nid;
1207}
1208
1209/*
1210 * helper for free_pool_huge_page() - return the previously saved
1211 * node ["this node"] from which to free a huge page. Advance the
1212 * next node id whether or not we find a free huge page to free so
1213 * that the next attempt to free addresses the next node.
1214 */
1215static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1216{
1217 int nid;
1218
1219 VM_BUG_ON(!nodes_allowed);
1220
1221 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1222 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1223
1224 return nid;
1225}
1226
1227#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1228 for (nr_nodes = nodes_weight(*mask); \
1229 nr_nodes > 0 && \
1230 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1231 nr_nodes--)
1232
1233#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1234 for (nr_nodes = nodes_weight(*mask); \
1235 nr_nodes > 0 && \
1236 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1237 nr_nodes--)
1238
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001239#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001240static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001241 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001242{
1243 int i;
1244 int nr_pages = 1 << order;
1245 struct page *p = page + 1;
1246
Gerald Schaeferc8cc7082016-06-24 14:50:04 -07001247 atomic_set(compound_mapcount_ptr(page), 0);
John Hubbard47e29d32020-04-01 21:05:33 -07001248 if (hpage_pincount_available(page))
1249 atomic_set(compound_pincount_ptr(page), 0);
1250
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001251 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001252 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001253 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001254 }
1255
1256 set_compound_order(page, 0);
Gerald Schaeferba9c1202020-12-11 13:36:53 -08001257 page[1].compound_nr = 0;
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001258 __ClearPageHead(page);
1259}
1260
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001261static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001262{
Roman Gushchincf11e852020-04-10 14:32:45 -07001263 /*
1264 * If the page isn't allocated using the cma allocator,
1265 * cma_release() returns false.
1266 */
Barry Songdbda8fe2020-07-23 21:15:30 -07001267#ifdef CONFIG_CMA
1268 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
Roman Gushchincf11e852020-04-10 14:32:45 -07001269 return;
Barry Songdbda8fe2020-07-23 21:15:30 -07001270#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001271
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001272 free_contig_range(page_to_pfn(page), 1 << order);
1273}
1274
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001275#ifdef CONFIG_CONTIG_ALLOC
Michal Hockod9cc948f2018-01-31 16:20:44 -08001276static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1277 int nid, nodemask_t *nodemask)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001278{
Anshuman Khandual5e27a2d2019-11-30 17:55:06 -08001279 unsigned long nr_pages = 1UL << huge_page_order(h);
Li Xinhai953f0642020-09-04 16:36:10 -07001280 if (nid == NUMA_NO_NODE)
1281 nid = numa_mem_id();
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001282
Barry Songdbda8fe2020-07-23 21:15:30 -07001283#ifdef CONFIG_CMA
1284 {
Roman Gushchincf11e852020-04-10 14:32:45 -07001285 struct page *page;
1286 int node;
1287
Li Xinhai953f0642020-09-04 16:36:10 -07001288 if (hugetlb_cma[nid]) {
1289 page = cma_alloc(hugetlb_cma[nid], nr_pages,
Minchan Kim23ba9902021-01-21 12:09:34 -08001290 huge_page_order(h),
1291 GFP_KERNEL | __GFP_NOWARN);
Roman Gushchincf11e852020-04-10 14:32:45 -07001292 if (page)
1293 return page;
1294 }
Li Xinhai953f0642020-09-04 16:36:10 -07001295
1296 if (!(gfp_mask & __GFP_THISNODE)) {
1297 for_each_node_mask(node, *nodemask) {
1298 if (node == nid || !hugetlb_cma[node])
1299 continue;
1300
1301 page = cma_alloc(hugetlb_cma[node], nr_pages,
Minchan Kim23ba9902021-01-21 12:09:34 -08001302 huge_page_order(h),
1303 GFP_KERNEL | __GFP_NOWARN);
Li Xinhai953f0642020-09-04 16:36:10 -07001304 if (page)
1305 return page;
1306 }
1307 }
Roman Gushchincf11e852020-04-10 14:32:45 -07001308 }
Barry Songdbda8fe2020-07-23 21:15:30 -07001309#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001310
Anshuman Khandual5e27a2d2019-11-30 17:55:06 -08001311 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001312}
1313
1314static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001315static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001316#else /* !CONFIG_CONTIG_ALLOC */
1317static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1318 int nid, nodemask_t *nodemask)
1319{
1320 return NULL;
1321}
1322#endif /* CONFIG_CONTIG_ALLOC */
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001323
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001324#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
Michal Hockod9cc948f2018-01-31 16:20:44 -08001325static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001326 int nid, nodemask_t *nodemask)
1327{
1328 return NULL;
1329}
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001330static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001331static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001332 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001333#endif
1334
Andi Kleena5516432008-07-23 21:27:41 -07001335static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001336{
1337 int i;
Mike Kravetz65f6dc32021-02-24 12:07:50 -08001338 struct page *subpage = page;
Andi Kleena5516432008-07-23 21:27:41 -07001339
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001340 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001341 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001342
Andi Kleena5516432008-07-23 21:27:41 -07001343 h->nr_huge_pages--;
1344 h->nr_huge_pages_node[page_to_nid(page)]--;
Mike Kravetz65f6dc32021-02-24 12:07:50 -08001345 for (i = 0; i < pages_per_huge_page(h);
1346 i++, subpage = mem_map_next(subpage, page, i)) {
1347 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
Chris Forbes32f84522011-07-25 17:12:14 -07001348 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001349 1 << PG_active | 1 << PG_private |
1350 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001351 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001352 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Mina Almasry1adc4d42020-04-01 21:11:15 -07001353 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001354 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001355 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001356 if (hstate_is_gigantic(h)) {
Roman Gushchincf11e852020-04-10 14:32:45 -07001357 /*
1358 * Temporarily drop the hugetlb_lock, because
1359 * we might block in free_gigantic_page().
1360 */
1361 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001362 destroy_compound_gigantic_page(page, huge_page_order(h));
1363 free_gigantic_page(page, huge_page_order(h));
Roman Gushchincf11e852020-04-10 14:32:45 -07001364 spin_lock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001365 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001366 __free_pages(page, huge_page_order(h));
1367 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001368}
1369
Andi Kleene5ff2152008-07-23 21:27:42 -07001370struct hstate *size_to_hstate(unsigned long size)
1371{
1372 struct hstate *h;
1373
1374 for_each_hstate(h) {
1375 if (huge_page_size(h) == size)
1376 return h;
1377 }
1378 return NULL;
1379}
1380
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001381/*
1382 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1383 * to hstate->hugepage_activelist.)
1384 *
1385 * This function can be called for tail pages, but never returns true for them.
1386 */
1387bool page_huge_active(struct page *page)
1388{
Muchun Songeca84eb2021-02-04 18:32:13 -08001389 return PageHeadHuge(page) && PagePrivate(&page[1]);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001390}
1391
1392/* never called for tail page */
Muchun Songafe6c312021-02-04 18:32:03 -08001393void set_page_huge_active(struct page *page)
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001394{
1395 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1396 SetPagePrivate(&page[1]);
1397}
1398
1399static void clear_page_huge_active(struct page *page)
1400{
1401 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1402 ClearPagePrivate(&page[1]);
1403}
1404
Michal Hockoab5ac902018-01-31 16:20:48 -08001405/*
1406 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1407 * code
1408 */
1409static inline bool PageHugeTemporary(struct page *page)
1410{
1411 if (!PageHuge(page))
1412 return false;
1413
1414 return (unsigned long)page[2].mapping == -1U;
1415}
1416
1417static inline void SetPageHugeTemporary(struct page *page)
1418{
1419 page[2].mapping = (void *)-1U;
1420}
1421
1422static inline void ClearPageHugeTemporary(struct page *page)
1423{
1424 page[2].mapping = NULL;
1425}
1426
Waiman Longc77c0a82020-01-04 13:00:15 -08001427static void __free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001428{
Andi Kleena5516432008-07-23 21:27:41 -07001429 /*
1430 * Can't pass hstate in here because it is called from the
1431 * compound page destructor.
1432 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001433 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001434 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001435 struct hugepage_subpool *spool =
1436 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001437 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001438
Mike Kravetzb4330af2016-02-05 15:36:38 -08001439 VM_BUG_ON_PAGE(page_count(page), page);
1440 VM_BUG_ON_PAGE(page_mapcount(page), page);
Yongkai Wu8ace22b2018-12-14 14:17:10 -08001441
1442 set_page_private(page, 0);
1443 page->mapping = NULL;
Joonsoo Kim07443a82013-09-11 14:21:58 -07001444 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001445 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001446
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001447 /*
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001448 * If PagePrivate() was set on page, page allocation consumed a
1449 * reservation. If the page was associated with a subpool, there
1450 * would have been a page reserved in the subpool before allocation
1451 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1452 * reservtion, do not call hugepage_subpool_put_pages() as this will
1453 * remove the reserved page from the subpool.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001454 */
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001455 if (!restore_reserve) {
1456 /*
1457 * A return code of zero implies that the subpool will be
1458 * under its minimum size if the reservation is not restored
1459 * after page is free. Therefore, force restore_reserve
1460 * operation.
1461 */
1462 if (hugepage_subpool_put_pages(spool, 1) == 0)
1463 restore_reserve = true;
1464 }
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001465
David Gibson27a85ef2006-03-22 00:08:56 -08001466 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001467 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001468 hugetlb_cgroup_uncharge_page(hstate_index(h),
1469 pages_per_huge_page(h), page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07001470 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1471 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001472 if (restore_reserve)
1473 h->resv_huge_pages++;
1474
Michal Hockoab5ac902018-01-31 16:20:48 -08001475 if (PageHugeTemporary(page)) {
1476 list_del(&page->lru);
1477 ClearPageHugeTemporary(page);
1478 update_and_free_page(h, page);
1479 } else if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001480 /* remove the page from active list */
1481 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001482 update_and_free_page(h, page);
1483 h->surplus_huge_pages--;
1484 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001485 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001486 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001487 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001488 }
David Gibson27a85ef2006-03-22 00:08:56 -08001489 spin_unlock(&hugetlb_lock);
1490}
1491
Waiman Longc77c0a82020-01-04 13:00:15 -08001492/*
1493 * As free_huge_page() can be called from a non-task context, we have
1494 * to defer the actual freeing in a workqueue to prevent potential
1495 * hugetlb_lock deadlock.
1496 *
1497 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1498 * be freed and frees them one-by-one. As the page->mapping pointer is
1499 * going to be cleared in __free_huge_page() anyway, it is reused as the
1500 * llist_node structure of a lockless linked list of huge pages to be freed.
1501 */
1502static LLIST_HEAD(hpage_freelist);
1503
1504static void free_hpage_workfn(struct work_struct *work)
1505{
1506 struct llist_node *node;
1507 struct page *page;
1508
1509 node = llist_del_all(&hpage_freelist);
1510
1511 while (node) {
1512 page = container_of((struct address_space **)node,
1513 struct page, mapping);
1514 node = node->next;
1515 __free_huge_page(page);
1516 }
1517}
1518static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1519
1520void free_huge_page(struct page *page)
1521{
1522 /*
1523 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1524 */
1525 if (!in_task()) {
1526 /*
1527 * Only call schedule_work() if hpage_freelist is previously
1528 * empty. Otherwise, schedule_work() had been called but the
1529 * workfn hasn't retrieved the list yet.
1530 */
1531 if (llist_add((struct llist_node *)&page->mapping,
1532 &hpage_freelist))
1533 schedule_work(&free_hpage_work);
1534 return;
1535 }
1536
1537 __free_huge_page(page);
1538}
1539
Andi Kleena5516432008-07-23 21:27:41 -07001540static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001541{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001542 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001543 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001544 set_hugetlb_cgroup(page, NULL);
Mina Almasry1adc4d42020-04-01 21:11:15 -07001545 set_hugetlb_cgroup_rsvd(page, NULL);
Wei Yang2f375112020-10-13 16:56:36 -07001546 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001547 h->nr_huge_pages++;
1548 h->nr_huge_pages_node[nid]++;
Muchun Songe334b1f2021-02-04 18:32:06 -08001549 ClearPageHugeFreed(page);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001550 spin_unlock(&hugetlb_lock);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001551}
1552
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001553static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001554{
1555 int i;
1556 int nr_pages = 1 << order;
1557 struct page *p = page + 1;
1558
1559 /* we rely on prep_new_huge_page to set the destructor */
1560 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001561 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001562 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001563 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001564 /*
1565 * For gigantic hugepages allocated through bootmem at
1566 * boot, it's safer to be consistent with the not-gigantic
1567 * hugepages and clear the PG_reserved bit from all tail pages
Ethon Paul7c8de352020-06-04 16:49:07 -07001568 * too. Otherwise drivers using get_user_pages() to access tail
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001569 * pages may get the reference counting wrong if they see
1570 * PG_reserved set on a tail page (despite the head page not
1571 * having PG_reserved set). Enforcing this consistency between
1572 * head and tail pages allows drivers to optimize away a check
1573 * on the head page when they need know if put_page() is needed
1574 * after get_user_pages().
1575 */
1576 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001577 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001578 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001579 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001580 atomic_set(compound_mapcount_ptr(page), -1);
John Hubbard47e29d32020-04-01 21:05:33 -07001581
1582 if (hpage_pincount_available(page))
1583 atomic_set(compound_pincount_ptr(page), 0);
Wu Fengguang20a03072009-06-16 15:32:22 -07001584}
1585
Andrew Morton77959122012-10-08 16:34:11 -07001586/*
1587 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1588 * transparent huge pages. See the PageTransHuge() documentation for more
1589 * details.
1590 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001591int PageHuge(struct page *page)
1592{
Wu Fengguang20a03072009-06-16 15:32:22 -07001593 if (!PageCompound(page))
1594 return 0;
1595
1596 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001597 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001598}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001599EXPORT_SYMBOL_GPL(PageHuge);
1600
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001601/*
1602 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1603 * normal or transparent huge pages.
1604 */
1605int PageHeadHuge(struct page *page_head)
1606{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001607 if (!PageHead(page_head))
1608 return 0;
1609
Vlastimil Babkad4af73e2020-04-01 21:11:48 -07001610 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001611}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001612
Mike Kravetzc0d03812020-04-01 21:11:05 -07001613/*
Mike Kravetzc0d03812020-04-01 21:11:05 -07001614 * Find and lock address space (mapping) in write mode.
1615 *
Mike Kravetz336bf302020-11-13 22:52:16 -08001616 * Upon entry, the page is locked which means that page_mapping() is
1617 * stable. Due to locking order, we can only trylock_write. If we can
1618 * not get the lock, simply return NULL to caller.
Mike Kravetzc0d03812020-04-01 21:11:05 -07001619 */
1620struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1621{
Mike Kravetz336bf302020-11-13 22:52:16 -08001622 struct address_space *mapping = page_mapping(hpage);
Mike Kravetzc0d03812020-04-01 21:11:05 -07001623
Mike Kravetzc0d03812020-04-01 21:11:05 -07001624 if (!mapping)
1625 return mapping;
1626
Mike Kravetzc0d03812020-04-01 21:11:05 -07001627 if (i_mmap_trylock_write(mapping))
1628 return mapping;
1629
Mike Kravetz336bf302020-11-13 22:52:16 -08001630 return NULL;
Mike Kravetzc0d03812020-04-01 21:11:05 -07001631}
1632
Zhang Yi13d60f42013-06-25 21:19:31 +08001633pgoff_t __basepage_index(struct page *page)
1634{
1635 struct page *page_head = compound_head(page);
1636 pgoff_t index = page_index(page_head);
1637 unsigned long compound_idx;
1638
1639 if (!PageHuge(page_head))
1640 return page_index(page);
1641
1642 if (compound_order(page_head) >= MAX_ORDER)
1643 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1644 else
1645 compound_idx = page - page_head;
1646
1647 return (index << compound_order(page_head)) + compound_idx;
1648}
1649
Michal Hocko0c397da2018-01-31 16:20:56 -08001650static struct page *alloc_buddy_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001651 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1652 nodemask_t *node_alloc_noretry)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653{
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001654 int order = huge_page_order(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 struct page *page;
Mike Kravetzf60858f2019-09-23 15:37:35 -07001656 bool alloc_try_hard = true;
Joe Jinf96efd52007-07-15 23:38:12 -07001657
Mike Kravetzf60858f2019-09-23 15:37:35 -07001658 /*
1659 * By default we always try hard to allocate the page with
1660 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1661 * a loop (to adjust global huge page counts) and previous allocation
1662 * failed, do not continue to try hard on the same node. Use the
1663 * node_alloc_noretry bitmap to manage this state information.
1664 */
1665 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1666 alloc_try_hard = false;
1667 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1668 if (alloc_try_hard)
1669 gfp_mask |= __GFP_RETRY_MAYFAIL;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001670 if (nid == NUMA_NO_NODE)
1671 nid = numa_mem_id();
1672 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1673 if (page)
1674 __count_vm_event(HTLB_BUDDY_PGALLOC);
1675 else
1676 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001677
Mike Kravetzf60858f2019-09-23 15:37:35 -07001678 /*
1679 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1680 * indicates an overall state change. Clear bit so that we resume
1681 * normal 'try hard' allocations.
1682 */
1683 if (node_alloc_noretry && page && !alloc_try_hard)
1684 node_clear(nid, *node_alloc_noretry);
1685
1686 /*
1687 * If we tried hard to get a page but failed, set bit so that
1688 * subsequent attempts will not try as hard until there is an
1689 * overall state change.
1690 */
1691 if (node_alloc_noretry && !page && alloc_try_hard)
1692 node_set(nid, *node_alloc_noretry);
1693
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001694 return page;
1695}
1696
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001697/*
Michal Hocko0c397da2018-01-31 16:20:56 -08001698 * Common helper to allocate a fresh hugetlb page. All specific allocators
1699 * should use this function to get new hugetlb pages
1700 */
1701static struct page *alloc_fresh_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001702 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1703 nodemask_t *node_alloc_noretry)
Michal Hocko0c397da2018-01-31 16:20:56 -08001704{
1705 struct page *page;
1706
1707 if (hstate_is_gigantic(h))
1708 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1709 else
1710 page = alloc_buddy_huge_page(h, gfp_mask,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001711 nid, nmask, node_alloc_noretry);
Michal Hocko0c397da2018-01-31 16:20:56 -08001712 if (!page)
1713 return NULL;
1714
1715 if (hstate_is_gigantic(h))
1716 prep_compound_gigantic_page(page, huge_page_order(h));
1717 prep_new_huge_page(h, page, page_to_nid(page));
1718
1719 return page;
1720}
1721
1722/*
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001723 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1724 * manner.
1725 */
Mike Kravetzf60858f2019-09-23 15:37:35 -07001726static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1727 nodemask_t *node_alloc_noretry)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001728{
1729 struct page *page;
1730 int nr_nodes, node;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001731 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001732
1733 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
Mike Kravetzf60858f2019-09-23 15:37:35 -07001734 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1735 node_alloc_noretry);
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001736 if (page)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001737 break;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001738 }
1739
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001740 if (!page)
1741 return 0;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001742
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001743 put_page(page); /* free it into the hugepage allocator */
1744
1745 return 1;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001746}
1747
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001748/*
1749 * Free huge page from pool from next node to free.
1750 * Attempt to keep persistent huge pages more or less
1751 * balanced over allowed nodes.
1752 * Called with hugetlb_lock locked.
1753 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001754static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1755 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001756{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001757 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001758 int ret = 0;
1759
Joonsoo Kimb2261022013-09-11 14:21:00 -07001760 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001761 /*
1762 * If we're returning unused surplus pages, only examine
1763 * nodes with surplus pages.
1764 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001765 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1766 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001767 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001768 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001769 struct page, lru);
1770 list_del(&page->lru);
1771 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001772 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001773 if (acct_surplus) {
1774 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001775 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001776 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001777 update_and_free_page(h, page);
1778 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001779 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001780 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001781 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001782
1783 return ret;
1784}
1785
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001786/*
1787 * Dissolve a given free hugepage into free buddy pages. This function does
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001788 * nothing for in-use hugepages and non-hugepages.
1789 * This function returns values like below:
1790 *
1791 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1792 * (allocated or reserved.)
1793 * 0: successfully dissolved free hugepages or the page is not a
1794 * hugepage (considered as already dissolved)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001795 */
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001796int dissolve_free_huge_page(struct page *page)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001797{
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001798 int rc = -EBUSY;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001799
Muchun Songe334b1f2021-02-04 18:32:06 -08001800retry:
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001801 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1802 if (!PageHuge(page))
1803 return 0;
1804
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001805 spin_lock(&hugetlb_lock);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001806 if (!PageHuge(page)) {
1807 rc = 0;
1808 goto out;
1809 }
1810
1811 if (!page_count(page)) {
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001812 struct page *head = compound_head(page);
1813 struct hstate *h = page_hstate(head);
1814 int nid = page_to_nid(head);
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001815 if (h->free_huge_pages - h->resv_huge_pages == 0)
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001816 goto out;
Muchun Songe334b1f2021-02-04 18:32:06 -08001817
1818 /*
1819 * We should make sure that the page is already on the free list
1820 * when it is dissolved.
1821 */
1822 if (unlikely(!PageHugeFreed(head))) {
1823 spin_unlock(&hugetlb_lock);
1824 cond_resched();
1825
1826 /*
1827 * Theoretically, we should return -EBUSY when we
1828 * encounter this race. In fact, we have a chance
1829 * to successfully dissolve the page if we do a
1830 * retry. Because the race window is quite small.
1831 * If we seize this opportunity, it is an optimization
1832 * for increasing the success rate of dissolving page.
1833 */
1834 goto retry;
1835 }
1836
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001837 /*
1838 * Move PageHWPoison flag from head page to the raw error page,
1839 * which makes any subpages rather than the error page reusable.
1840 */
1841 if (PageHWPoison(head) && page != head) {
1842 SetPageHWPoison(page);
1843 ClearPageHWPoison(head);
1844 }
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001845 list_del(&head->lru);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001846 h->free_huge_pages--;
1847 h->free_huge_pages_node[nid]--;
zhong jiangc1470b32016-08-11 15:32:55 -07001848 h->max_huge_pages--;
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001849 update_and_free_page(h, head);
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001850 rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001851 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001852out:
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001853 spin_unlock(&hugetlb_lock);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001854 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001855}
1856
1857/*
1858 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1859 * make specified memory blocks removable from the system.
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001860 * Note that this will dissolve a free gigantic hugepage completely, if any
1861 * part of it lies within the given range.
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001862 * Also note that if dissolve_free_huge_page() returns with an error, all
1863 * free hugepages that were dissolved before that error are lost.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001864 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001865int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001866{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001867 unsigned long pfn;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001868 struct page *page;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001869 int rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001870
Li Zhongd0177632014-08-06 16:07:56 -07001871 if (!hugepages_supported())
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001872 return rc;
Li Zhongd0177632014-08-06 16:07:56 -07001873
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001874 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1875 page = pfn_to_page(pfn);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001876 rc = dissolve_free_huge_page(page);
1877 if (rc)
1878 break;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001879 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001880
1881 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001882}
1883
Michal Hockoab5ac902018-01-31 16:20:48 -08001884/*
1885 * Allocates a fresh surplus page from the page allocator.
1886 */
Michal Hocko0c397da2018-01-31 16:20:56 -08001887static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
Michal Hockoaaf14e42017-07-10 15:49:08 -07001888 int nid, nodemask_t *nmask)
Adam Litke7893d1d2007-10-16 01:26:18 -07001889{
Michal Hocko9980d742018-01-31 16:20:52 -08001890 struct page *page = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -07001891
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001892 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001893 return NULL;
1894
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001895 spin_lock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001896 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1897 goto out_unlock;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001898 spin_unlock(&hugetlb_lock);
1899
Mike Kravetzf60858f2019-09-23 15:37:35 -07001900 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hocko9980d742018-01-31 16:20:52 -08001901 if (!page)
Michal Hocko0c397da2018-01-31 16:20:56 -08001902 return NULL;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001903
1904 spin_lock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001905 /*
1906 * We could have raced with the pool size change.
1907 * Double check that and simply deallocate the new page
1908 * if we would end up overcommiting the surpluses. Abuse
1909 * temporary page to workaround the nasty free_huge_page
1910 * codeflow
1911 */
1912 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1913 SetPageHugeTemporary(page);
Kai Shen2bf753e2019-05-13 17:15:37 -07001914 spin_unlock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001915 put_page(page);
Kai Shen2bf753e2019-05-13 17:15:37 -07001916 return NULL;
Michal Hocko9980d742018-01-31 16:20:52 -08001917 } else {
Michal Hocko9980d742018-01-31 16:20:52 -08001918 h->surplus_huge_pages++;
Michal Hocko4704dea2018-03-09 15:50:55 -08001919 h->surplus_huge_pages_node[page_to_nid(page)]++;
Adam Litke7893d1d2007-10-16 01:26:18 -07001920 }
Michal Hocko9980d742018-01-31 16:20:52 -08001921
1922out_unlock:
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001923 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001924
1925 return page;
1926}
1927
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001928static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
Aneesh Kumar K.V9a4e9f32019-03-05 15:47:44 -08001929 int nid, nodemask_t *nmask)
Michal Hockoab5ac902018-01-31 16:20:48 -08001930{
1931 struct page *page;
1932
1933 if (hstate_is_gigantic(h))
1934 return NULL;
1935
Mike Kravetzf60858f2019-09-23 15:37:35 -07001936 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hockoab5ac902018-01-31 16:20:48 -08001937 if (!page)
1938 return NULL;
1939
1940 /*
1941 * We do not account these pages as surplus because they are only
1942 * temporary and will be released properly on the last reference
1943 */
Michal Hockoab5ac902018-01-31 16:20:48 -08001944 SetPageHugeTemporary(page);
1945
1946 return page;
1947}
1948
Adam Litkee4e574b2007-10-16 01:26:19 -07001949/*
Dave Hansen099730d2015-11-05 18:50:17 -08001950 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1951 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001952static
Michal Hocko0c397da2018-01-31 16:20:56 -08001953struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
Dave Hansen099730d2015-11-05 18:50:17 -08001954 struct vm_area_struct *vma, unsigned long addr)
1955{
Michal Hockoaaf14e42017-07-10 15:49:08 -07001956 struct page *page;
1957 struct mempolicy *mpol;
1958 gfp_t gfp_mask = htlb_alloc_mask(h);
1959 int nid;
1960 nodemask_t *nodemask;
1961
1962 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
Michal Hocko0c397da2018-01-31 16:20:56 -08001963 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
Michal Hockoaaf14e42017-07-10 15:49:08 -07001964 mpol_cond_put(mpol);
1965
1966 return page;
Dave Hansen099730d2015-11-05 18:50:17 -08001967}
1968
Michal Hockoab5ac902018-01-31 16:20:48 -08001969/* page migration callback function */
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001970struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07001971 nodemask_t *nmask, gfp_t gfp_mask)
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001972{
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001973 spin_lock(&hugetlb_lock);
1974 if (h->free_huge_pages - h->resv_huge_pages > 0) {
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001975 struct page *page;
1976
1977 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1978 if (page) {
1979 spin_unlock(&hugetlb_lock);
1980 return page;
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001981 }
1982 }
1983 spin_unlock(&hugetlb_lock);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001984
Michal Hocko0c397da2018-01-31 16:20:56 -08001985 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001986}
1987
Michal Hockoebd63722018-01-31 16:21:00 -08001988/* mempolicy aware migration callback */
Michal Hocko389c8172018-01-31 16:21:03 -08001989struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1990 unsigned long address)
Michal Hockoebd63722018-01-31 16:21:00 -08001991{
1992 struct mempolicy *mpol;
1993 nodemask_t *nodemask;
1994 struct page *page;
Michal Hockoebd63722018-01-31 16:21:00 -08001995 gfp_t gfp_mask;
1996 int node;
1997
Michal Hockoebd63722018-01-31 16:21:00 -08001998 gfp_mask = htlb_alloc_mask(h);
1999 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07002000 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
Michal Hockoebd63722018-01-31 16:21:00 -08002001 mpol_cond_put(mpol);
2002
2003 return page;
2004}
2005
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09002006/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002007 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07002008 * of size 'delta'.
2009 */
Andi Kleena5516432008-07-23 21:27:41 -07002010static int gather_surplus_pages(struct hstate *h, int delta)
Jules Irenge1b2a1e72020-04-06 20:08:09 -07002011 __must_hold(&hugetlb_lock)
Adam Litkee4e574b2007-10-16 01:26:19 -07002012{
2013 struct list_head surplus_list;
2014 struct page *page, *tmp;
2015 int ret, i;
2016 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07002017 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07002018
Andi Kleena5516432008-07-23 21:27:41 -07002019 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08002020 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002021 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002022 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08002023 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002024
2025 allocated = 0;
2026 INIT_LIST_HEAD(&surplus_list);
2027
2028 ret = -ENOMEM;
2029retry:
2030 spin_unlock(&hugetlb_lock);
2031 for (i = 0; i < needed; i++) {
Michal Hocko0c397da2018-01-31 16:20:56 -08002032 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
Michal Hockoaaf14e42017-07-10 15:49:08 -07002033 NUMA_NO_NODE, NULL);
Hillf Danton28073b02012-03-21 16:34:00 -07002034 if (!page) {
2035 alloc_ok = false;
2036 break;
2037 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002038 list_add(&page->lru, &surplus_list);
David Rientjes69ed7792017-07-10 15:48:50 -07002039 cond_resched();
Adam Litkee4e574b2007-10-16 01:26:19 -07002040 }
Hillf Danton28073b02012-03-21 16:34:00 -07002041 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07002042
2043 /*
2044 * After retaking hugetlb_lock, we need to recalculate 'needed'
2045 * because either resv_huge_pages or free_huge_pages may have changed.
2046 */
2047 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002048 needed = (h->resv_huge_pages + delta) -
2049 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07002050 if (needed > 0) {
2051 if (alloc_ok)
2052 goto retry;
2053 /*
2054 * We were not able to allocate enough pages to
2055 * satisfy the entire reservation so we free what
2056 * we've allocated so far.
2057 */
2058 goto free;
2059 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002060 /*
2061 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002062 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07002063 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08002064 * allocator. Commit the entire reservation here to prevent another
2065 * process from stealing the pages as they are added to the pool but
2066 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07002067 */
2068 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07002069 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002070 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002071
Adam Litke19fc3f02008-04-28 02:12:20 -07002072 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07002073 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07002074 if ((--needed) < 0)
2075 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002076 /*
2077 * This page is now managed by the hugetlb allocator and has
2078 * no users -- drop the buddy allocator's reference.
2079 */
2080 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08002081 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07002082 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07002083 }
Hillf Danton28073b02012-03-21 16:34:00 -07002084free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08002085 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07002086
2087 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07002088 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2089 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002090 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002091
2092 return ret;
2093}
2094
2095/*
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002096 * This routine has two main purposes:
2097 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2098 * in unused_resv_pages. This corresponds to the prior adjustments made
2099 * to the associated reservation map.
2100 * 2) Free any unused surplus pages that may have been allocated to satisfy
2101 * the reservation. As many as unused_resv_pages may be freed.
2102 *
2103 * Called with hugetlb_lock held. However, the lock could be dropped (and
2104 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2105 * we must make sure nobody else can claim pages we are in the process of
2106 * freeing. Do this by ensuring resv_huge_page always is greater than the
2107 * number of huge pages we plan to free when dropping the lock.
Adam Litkee4e574b2007-10-16 01:26:19 -07002108 */
Andi Kleena5516432008-07-23 21:27:41 -07002109static void return_unused_surplus_pages(struct hstate *h,
2110 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07002111{
Adam Litkee4e574b2007-10-16 01:26:19 -07002112 unsigned long nr_pages;
2113
Andi Kleenaa888a72008-07-23 21:27:47 -07002114 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002115 if (hstate_is_gigantic(h))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002116 goto out;
Andi Kleenaa888a72008-07-23 21:27:47 -07002117
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002118 /*
2119 * Part (or even all) of the reservation could have been backed
2120 * by pre-allocated pages. Only free surplus pages.
2121 */
Andi Kleena5516432008-07-23 21:27:41 -07002122 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07002123
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002124 /*
2125 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002126 * evenly across all nodes with memory. Iterate across these nodes
2127 * until we can no longer free unreserved surplus pages. This occurs
2128 * when the nodes with surplus pages have no free pages.
Randy Dunlap9e7ee402020-08-11 18:32:59 -07002129 * free_pool_huge_page() will balance the freed pages across the
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002130 * on-line nodes with memory and will handle the hstate accounting.
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002131 *
2132 * Note that we decrement resv_huge_pages as we free the pages. If
2133 * we drop the lock, resv_huge_pages will still be sufficiently large
2134 * to cover subsequent pages we may free.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002135 */
2136 while (nr_pages--) {
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002137 h->resv_huge_pages--;
2138 unused_resv_pages--;
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002139 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002140 goto out;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07002141 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002142 }
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002143
2144out:
2145 /* Fully uncommit the reservation */
2146 h->resv_huge_pages -= unused_resv_pages;
Adam Litkee4e574b2007-10-16 01:26:19 -07002147}
2148
Mike Kravetz5e911372015-09-08 15:01:28 -07002149
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002150/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002151 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07002152 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002153 *
2154 * vma_needs_reservation is called to determine if the huge page at addr
2155 * within the vma has an associated reservation. If a reservation is
2156 * needed, the value 1 is returned. The caller is then responsible for
2157 * managing the global reservation and subpool usage counts. After
2158 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002159 * to add the page to the reservation map. If the page allocation fails,
2160 * the reservation must be ended instead of committed. vma_end_reservation
2161 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002162 *
2163 * In the normal case, vma_commit_reservation returns the same value
2164 * as the preceding vma_needs_reservation call. The only time this
2165 * is not the case is if a reserve map was changed between calls. It
2166 * is the responsibility of the caller to notice the difference and
2167 * take appropriate action.
Mike Kravetz96b96a92016-11-10 10:46:32 -08002168 *
2169 * vma_add_reservation is used in error paths where a reservation must
2170 * be restored when a newly allocated huge page must be freed. It is
2171 * to be called after calling vma_needs_reservation to determine if a
2172 * reservation exists.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002173 */
Mike Kravetz5e911372015-09-08 15:01:28 -07002174enum vma_resv_mode {
2175 VMA_NEEDS_RESV,
2176 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002177 VMA_END_RESV,
Mike Kravetz96b96a92016-11-10 10:46:32 -08002178 VMA_ADD_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07002179};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002180static long __vma_reservation_common(struct hstate *h,
2181 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07002182 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002183{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002184 struct resv_map *resv;
2185 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002186 long ret;
Mina Almasry0db9d742020-04-01 21:11:25 -07002187 long dummy_out_regions_needed;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002188
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002189 resv = vma_resv_map(vma);
2190 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07002191 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002192
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002193 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07002194 switch (mode) {
2195 case VMA_NEEDS_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002196 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2197 /* We assume that vma_reservation_* routines always operate on
2198 * 1 page, and that adding to resv map a 1 page entry can only
2199 * ever require 1 region.
2200 */
2201 VM_BUG_ON(dummy_out_regions_needed != 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002202 break;
2203 case VMA_COMMIT_RESV:
Mina Almasry075a61d2020-04-01 21:11:28 -07002204 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002205 /* region_add calls of range 1 should never fail. */
2206 VM_BUG_ON(ret < 0);
Mike Kravetz5e911372015-09-08 15:01:28 -07002207 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002208 case VMA_END_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002209 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002210 ret = 0;
2211 break;
Mike Kravetz96b96a92016-11-10 10:46:32 -08002212 case VMA_ADD_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002213 if (vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07002214 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002215 /* region_add calls of range 1 should never fail. */
2216 VM_BUG_ON(ret < 0);
2217 } else {
2218 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz96b96a92016-11-10 10:46:32 -08002219 ret = region_del(resv, idx, idx + 1);
2220 }
2221 break;
Mike Kravetz5e911372015-09-08 15:01:28 -07002222 default:
2223 BUG();
2224 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002225
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002226 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002227 return ret;
Mike Kravetz67961f92016-06-08 15:33:42 -07002228 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2229 /*
2230 * In most cases, reserves always exist for private mappings.
2231 * However, a file associated with mapping could have been
2232 * hole punched or truncated after reserves were consumed.
2233 * As subsequent fault on such a range will not use reserves.
2234 * Subtle - The reserve map for private mappings has the
2235 * opposite meaning than that of shared mappings. If NO
2236 * entry is in the reserve map, it means a reservation exists.
2237 * If an entry exists in the reserve map, it means the
2238 * reservation has already been consumed. As a result, the
2239 * return value of this routine is the opposite of the
2240 * value returned from reserve map manipulation routines above.
2241 */
2242 if (ret)
2243 return 0;
2244 else
2245 return 1;
2246 }
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002247 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002248 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002249}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002250
2251static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07002252 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002253{
Mike Kravetz5e911372015-09-08 15:01:28 -07002254 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002255}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002256
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002257static long vma_commit_reservation(struct hstate *h,
2258 struct vm_area_struct *vma, unsigned long addr)
2259{
Mike Kravetz5e911372015-09-08 15:01:28 -07002260 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2261}
2262
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002263static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07002264 struct vm_area_struct *vma, unsigned long addr)
2265{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002266 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002267}
2268
Mike Kravetz96b96a92016-11-10 10:46:32 -08002269static long vma_add_reservation(struct hstate *h,
2270 struct vm_area_struct *vma, unsigned long addr)
2271{
2272 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2273}
2274
2275/*
2276 * This routine is called to restore a reservation on error paths. In the
2277 * specific error paths, a huge page was allocated (via alloc_huge_page)
2278 * and is about to be freed. If a reservation for the page existed,
2279 * alloc_huge_page would have consumed the reservation and set PagePrivate
2280 * in the newly allocated page. When the page is freed via free_huge_page,
2281 * the global reservation count will be incremented if PagePrivate is set.
2282 * However, free_huge_page can not adjust the reserve map. Adjust the
2283 * reserve map here to be consistent with global reserve count adjustments
2284 * to be made by free_huge_page.
2285 */
2286static void restore_reserve_on_error(struct hstate *h,
2287 struct vm_area_struct *vma, unsigned long address,
2288 struct page *page)
2289{
2290 if (unlikely(PagePrivate(page))) {
2291 long rc = vma_needs_reservation(h, vma, address);
2292
2293 if (unlikely(rc < 0)) {
2294 /*
2295 * Rare out of memory condition in reserve map
2296 * manipulation. Clear PagePrivate so that
2297 * global reserve count will not be incremented
2298 * by free_huge_page. This will make it appear
2299 * as though the reservation for this page was
2300 * consumed. This may prevent the task from
2301 * faulting in the page at a later time. This
2302 * is better than inconsistent global huge page
2303 * accounting of reserve counts.
2304 */
2305 ClearPagePrivate(page);
2306 } else if (rc) {
2307 rc = vma_add_reservation(h, vma, address);
2308 if (unlikely(rc < 0))
2309 /*
2310 * See above comment about rare out of
2311 * memory condition.
2312 */
2313 ClearPagePrivate(page);
2314 } else
2315 vma_end_reservation(h, vma, address);
2316 }
2317}
2318
Mike Kravetz70c35472015-09-08 15:01:54 -07002319struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002320 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321{
David Gibson90481622012-03-21 16:34:12 -07002322 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07002323 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08002324 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002325 long map_chg, map_commit;
2326 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002327 int ret, idx;
2328 struct hugetlb_cgroup *h_cg;
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002329 bool deferred_reserve;
Adam Litke2fc39ce2007-11-14 16:59:39 -08002330
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002331 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07002332 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002333 * Examine the region/reserve map to determine if the process
2334 * has a reservation for the page to be allocated. A return
2335 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07002336 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002337 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2338 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002339 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002340
2341 /*
2342 * Processes that did not create the mapping will have no
2343 * reserves as indicated by the region/reserve map. Check
2344 * that the allocation will not exceed the subpool limit.
2345 * Allocations for MAP_NORESERVE mappings also need to be
2346 * checked against any subpool limit.
2347 */
2348 if (map_chg || avoid_reserve) {
2349 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2350 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002351 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002352 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07002353 }
Mel Gormana1e78772008-07-23 21:27:23 -07002354
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002355 /*
2356 * Even though there was no reservation in the region/reserve
2357 * map, there could be reservations associated with the
2358 * subpool that can be used. This would be indicated if the
2359 * return value of hugepage_subpool_get_pages() is zero.
2360 * However, if avoid_reserve is specified we still avoid even
2361 * the subpool reservations.
2362 */
2363 if (avoid_reserve)
2364 gbl_chg = 1;
2365 }
2366
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002367 /* If this allocation is not consuming a reservation, charge it now.
2368 */
2369 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2370 if (deferred_reserve) {
2371 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2372 idx, pages_per_huge_page(h), &h_cg);
2373 if (ret)
2374 goto out_subpool_put;
2375 }
2376
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002377 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002378 if (ret)
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002379 goto out_uncharge_cgroup_reservation;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002380
Mel Gormana1e78772008-07-23 21:27:23 -07002381 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002382 /*
2383 * glb_chg is passed to indicate whether or not a page must be taken
2384 * from the global free pool (global change). gbl_chg == 0 indicates
2385 * a reservation exists for the allocation.
2386 */
2387 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002388 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07002389 spin_unlock(&hugetlb_lock);
Michal Hocko0c397da2018-01-31 16:20:56 -08002390 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002391 if (!page)
2392 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002393 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2394 SetPagePrivate(page);
2395 h->resv_huge_pages--;
2396 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07002397 spin_lock(&hugetlb_lock);
Wei Yang15a8d682020-10-13 16:56:33 -07002398 list_add(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002399 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07002400 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002401 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002402 /* If allocation is not consuming a reservation, also store the
2403 * hugetlb_cgroup pointer on the page.
2404 */
2405 if (deferred_reserve) {
2406 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2407 h_cg, page);
2408 }
2409
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002410 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07002411
David Gibson90481622012-03-21 16:34:12 -07002412 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07002413
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002414 map_commit = vma_commit_reservation(h, vma, addr);
2415 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07002416 /*
2417 * The page was added to the reservation map between
2418 * vma_needs_reservation and vma_commit_reservation.
2419 * This indicates a race with hugetlb_reserve_pages.
2420 * Adjust for the subpool count incremented above AND
2421 * in hugetlb_reserve_pages for the same page. Also,
2422 * the reservation count added in hugetlb_reserve_pages
2423 * no longer applies.
2424 */
2425 long rsv_adjust;
2426
2427 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2428 hugetlb_acct_memory(h, -rsv_adjust);
Mike Kravetz79aa9252020-11-01 17:07:27 -08002429 if (deferred_reserve)
2430 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2431 pages_per_huge_page(h), page);
Mike Kravetz33039672015-06-24 16:57:58 -07002432 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08002433 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002434
2435out_uncharge_cgroup:
2436 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002437out_uncharge_cgroup_reservation:
2438 if (deferred_reserve)
2439 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2440 h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002441out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002442 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002443 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002444 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002445 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08002446}
2447
Aneesh Kumar K.Ve24a1302017-07-28 10:31:25 +05302448int alloc_bootmem_huge_page(struct hstate *h)
2449 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2450int __alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002451{
2452 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002453 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002454
Joonsoo Kimb2261022013-09-11 14:21:00 -07002455 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002456 void *addr;
2457
Mike Rapoporteb31d552018-10-30 15:08:04 -07002458 addr = memblock_alloc_try_nid_raw(
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002459 huge_page_size(h), huge_page_size(h),
Mike Rapoport97ad1082018-10-30 15:09:44 -07002460 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002461 if (addr) {
2462 /*
2463 * Use the beginning of the huge page to store the
2464 * huge_bootmem_page struct (until gather_bootmem
2465 * puts them into the mem_map).
2466 */
2467 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002468 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002469 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002470 }
2471 return 0;
2472
2473found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002474 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002475 /* Put them into a private list first because mem_map is not up yet */
Cannon Matthews330d6e42018-08-17 15:49:17 -07002476 INIT_LIST_HEAD(&m->list);
Andi Kleenaa888a72008-07-23 21:27:47 -07002477 list_add(&m->list, &huge_boot_pages);
2478 m->hstate = h;
2479 return 1;
2480}
2481
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002482static void __init prep_compound_huge_page(struct page *page,
2483 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08002484{
2485 if (unlikely(order > (MAX_ORDER - 1)))
2486 prep_compound_gigantic_page(page, order);
2487 else
2488 prep_compound_page(page, order);
2489}
2490
Andi Kleenaa888a72008-07-23 21:27:47 -07002491/* Put bootmem huge pages into the standard lists after mem_map is up */
2492static void __init gather_bootmem_prealloc(void)
2493{
2494 struct huge_bootmem_page *m;
2495
2496 list_for_each_entry(m, &huge_boot_pages, list) {
Mike Kravetz40d18eb2018-08-17 15:49:07 -07002497 struct page *page = virt_to_page(m);
Andi Kleenaa888a72008-07-23 21:27:47 -07002498 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002499
Andi Kleenaa888a72008-07-23 21:27:47 -07002500 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08002501 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002502 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002503 prep_new_huge_page(h, page, page_to_nid(page));
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08002504 put_page(page); /* free it into the hugepage allocator */
2505
Rafael Aquinib0320c72011-06-15 15:08:39 -07002506 /*
2507 * If we had gigantic hugepages allocated at boot time, we need
2508 * to restore the 'stolen' pages to totalram_pages in order to
2509 * fix confusing memory reports from free(1) and another
2510 * side-effects, like CommitLimit going negative.
2511 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002512 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002513 adjust_managed_page_count(page, 1 << h->order);
Cannon Matthews520495f2018-07-03 17:02:43 -07002514 cond_resched();
Andi Kleenaa888a72008-07-23 21:27:47 -07002515 }
2516}
2517
Andi Kleen8faa8b02008-07-23 21:27:48 -07002518static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519{
2520 unsigned long i;
Mike Kravetzf60858f2019-09-23 15:37:35 -07002521 nodemask_t *node_alloc_noretry;
2522
2523 if (!hstate_is_gigantic(h)) {
2524 /*
2525 * Bit mask controlling how hard we retry per-node allocations.
2526 * Ignore errors as lower level routines can deal with
2527 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2528 * time, we are likely in bigger trouble.
2529 */
2530 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2531 GFP_KERNEL);
2532 } else {
2533 /* allocations done at boot time */
2534 node_alloc_noretry = NULL;
2535 }
2536
2537 /* bit mask controlling how hard we retry per-node allocations */
2538 if (node_alloc_noretry)
2539 nodes_clear(*node_alloc_noretry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540
Andi Kleene5ff2152008-07-23 21:27:42 -07002541 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002542 if (hstate_is_gigantic(h)) {
Barry Songdbda8fe2020-07-23 21:15:30 -07002543 if (hugetlb_cma_size) {
Roman Gushchincf11e852020-04-10 14:32:45 -07002544 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
Chen Wandunc9ea7712021-02-24 12:07:58 -08002545 goto free;
Roman Gushchincf11e852020-04-10 14:32:45 -07002546 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002547 if (!alloc_bootmem_huge_page(h))
2548 break;
Michal Hocko0c397da2018-01-31 16:20:56 -08002549 } else if (!alloc_pool_huge_page(h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07002550 &node_states[N_MEMORY],
2551 node_alloc_noretry))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 break;
David Rientjes69ed7792017-07-10 15:48:50 -07002553 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 }
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002555 if (i < h->max_huge_pages) {
2556 char buf[32];
2557
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002558 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002559 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2560 h->max_huge_pages, buf, i);
2561 h->max_huge_pages = i;
2562 }
Chen Wandunc9ea7712021-02-24 12:07:58 -08002563free:
Mike Kravetzf60858f2019-09-23 15:37:35 -07002564 kfree(node_alloc_noretry);
Andi Kleene5ff2152008-07-23 21:27:42 -07002565}
2566
2567static void __init hugetlb_init_hstates(void)
2568{
2569 struct hstate *h;
2570
2571 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002572 if (minimum_order > huge_page_order(h))
2573 minimum_order = huge_page_order(h);
2574
Andi Kleen8faa8b02008-07-23 21:27:48 -07002575 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002576 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002577 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002578 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002579 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002580}
2581
2582static void __init report_hugepages(void)
2583{
2584 struct hstate *h;
2585
2586 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002587 char buf[32];
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002588
2589 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Andrew Mortonffb22af2013-02-22 16:32:08 -08002590 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002591 buf, h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002592 }
2593}
2594
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002596static void try_to_free_low(struct hstate *h, unsigned long count,
2597 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002599 int i;
2600
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002601 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002602 return;
2603
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002604 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002606 struct list_head *freel = &h->hugepage_freelists[i];
2607 list_for_each_entry_safe(page, next, freel, lru) {
2608 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002609 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 if (PageHighMem(page))
2611 continue;
2612 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002613 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002614 h->free_huge_pages--;
2615 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 }
2617 }
2618}
2619#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002620static inline void try_to_free_low(struct hstate *h, unsigned long count,
2621 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622{
2623}
2624#endif
2625
Wu Fengguang20a03072009-06-16 15:32:22 -07002626/*
2627 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2628 * balanced by operating on them in a round-robin fashion.
2629 * Returns 1 if an adjustment was made.
2630 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002631static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2632 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002633{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002634 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002635
2636 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002637
Joonsoo Kimb2261022013-09-11 14:21:00 -07002638 if (delta < 0) {
2639 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2640 if (h->surplus_huge_pages_node[node])
2641 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002642 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002643 } else {
2644 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2645 if (h->surplus_huge_pages_node[node] <
2646 h->nr_huge_pages_node[node])
2647 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002648 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002649 }
2650 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002651
Joonsoo Kimb2261022013-09-11 14:21:00 -07002652found:
2653 h->surplus_huge_pages += delta;
2654 h->surplus_huge_pages_node[node] += delta;
2655 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002656}
2657
Andi Kleena5516432008-07-23 21:27:41 -07002658#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002659static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002660 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661{
Adam Litke7893d1d2007-10-16 01:26:18 -07002662 unsigned long min_count, ret;
Mike Kravetzf60858f2019-09-23 15:37:35 -07002663 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2664
2665 /*
2666 * Bit mask controlling how hard we retry per-node allocations.
2667 * If we can not allocate the bit mask, do not attempt to allocate
2668 * the requested huge pages.
2669 */
2670 if (node_alloc_noretry)
2671 nodes_clear(*node_alloc_noretry);
2672 else
2673 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002675 spin_lock(&hugetlb_lock);
2676
2677 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002678 * Check for a node specific request.
2679 * Changing node specific huge page count may require a corresponding
2680 * change to the global count. In any case, the passed node mask
2681 * (nodes_allowed) will restrict alloc/free to the specified node.
2682 */
2683 if (nid != NUMA_NO_NODE) {
2684 unsigned long old_count = count;
2685
2686 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2687 /*
2688 * User may have specified a large count value which caused the
2689 * above calculation to overflow. In this case, they wanted
2690 * to allocate as many huge pages as possible. Set count to
2691 * largest possible value to align with their intention.
2692 */
2693 if (count < old_count)
2694 count = ULONG_MAX;
2695 }
2696
2697 /*
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002698 * Gigantic pages runtime allocation depend on the capability for large
2699 * page range allocation.
2700 * If the system does not provide this feature, return an error when
2701 * the user tries to allocate gigantic pages but let the user free the
2702 * boottime allocated gigantic pages.
2703 */
2704 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2705 if (count > persistent_huge_pages(h)) {
2706 spin_unlock(&hugetlb_lock);
Mike Kravetzf60858f2019-09-23 15:37:35 -07002707 NODEMASK_FREE(node_alloc_noretry);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002708 return -EINVAL;
2709 }
2710 /* Fall through to decrease pool */
2711 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002712
Adam Litke7893d1d2007-10-16 01:26:18 -07002713 /*
2714 * Increase the pool size
2715 * First take pages out of surplus state. Then make up the
2716 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002717 *
Michal Hocko0c397da2018-01-31 16:20:56 -08002718 * We might race with alloc_surplus_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002719 * to convert a surplus huge page to a normal huge page. That is
2720 * not critical, though, it just means the overall size of the
2721 * pool might be one hugepage larger than it needs to be, but
2722 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002723 */
Andi Kleena5516432008-07-23 21:27:41 -07002724 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002725 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002726 break;
2727 }
2728
Andi Kleena5516432008-07-23 21:27:41 -07002729 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002730 /*
2731 * If this allocation races such that we no longer need the
2732 * page, free_huge_page will handle it by freeing the page
2733 * and reducing the surplus.
2734 */
2735 spin_unlock(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07002736
2737 /* yield cpu to avoid soft lockup */
2738 cond_resched();
2739
Mike Kravetzf60858f2019-09-23 15:37:35 -07002740 ret = alloc_pool_huge_page(h, nodes_allowed,
2741 node_alloc_noretry);
Adam Litke7893d1d2007-10-16 01:26:18 -07002742 spin_lock(&hugetlb_lock);
2743 if (!ret)
2744 goto out;
2745
Mel Gorman536240f22009-12-14 17:59:56 -08002746 /* Bail for signals. Probably ctrl-c from user */
2747 if (signal_pending(current))
2748 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002749 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002750
2751 /*
2752 * Decrease the pool size
2753 * First return free pages to the buddy allocator (being careful
2754 * to keep enough around to satisfy reservations). Then place
2755 * pages into surplus state as needed so the pool will shrink
2756 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002757 *
2758 * By placing pages into the surplus state independent of the
2759 * overcommit value, we are allowing the surplus pool size to
2760 * exceed overcommit. There are few sane options here. Since
Michal Hocko0c397da2018-01-31 16:20:56 -08002761 * alloc_surplus_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002762 * though, we'll note that we're not allowed to exceed surplus
2763 * and won't grow the pool anywhere else. Not until one of the
2764 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002765 */
Andi Kleena5516432008-07-23 21:27:41 -07002766 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002767 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002768 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002769 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002770 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002772 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 }
Andi Kleena5516432008-07-23 21:27:41 -07002774 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002775 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002776 break;
2777 }
2778out:
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002779 h->max_huge_pages = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 spin_unlock(&hugetlb_lock);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002781
Mike Kravetzf60858f2019-09-23 15:37:35 -07002782 NODEMASK_FREE(node_alloc_noretry);
2783
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002784 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785}
2786
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002787#define HSTATE_ATTR_RO(_name) \
2788 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2789
2790#define HSTATE_ATTR(_name) \
2791 static struct kobj_attribute _name##_attr = \
2792 __ATTR(_name, 0644, _name##_show, _name##_store)
2793
2794static struct kobject *hugepages_kobj;
2795static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2796
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002797static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2798
2799static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002800{
2801 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002802
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002803 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002804 if (hstate_kobjs[i] == kobj) {
2805 if (nidp)
2806 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002807 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002808 }
2809
2810 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002811}
2812
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002813static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002814 struct kobj_attribute *attr, char *buf)
2815{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002816 struct hstate *h;
2817 unsigned long nr_huge_pages;
2818 int nid;
2819
2820 h = kobj_to_hstate(kobj, &nid);
2821 if (nid == NUMA_NO_NODE)
2822 nr_huge_pages = h->nr_huge_pages;
2823 else
2824 nr_huge_pages = h->nr_huge_pages_node[nid];
2825
2826 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002827}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002828
David Rientjes238d3c12014-08-06 16:06:51 -07002829static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2830 struct hstate *h, int nid,
2831 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002832{
2833 int err;
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002834 nodemask_t nodes_allowed, *n_mask;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002835
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002836 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2837 return -EINVAL;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002838
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002839 if (nid == NUMA_NO_NODE) {
2840 /*
2841 * global hstate attribute
2842 */
2843 if (!(obey_mempolicy &&
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002844 init_nodemask_of_mempolicy(&nodes_allowed)))
2845 n_mask = &node_states[N_MEMORY];
2846 else
2847 n_mask = &nodes_allowed;
2848 } else {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002849 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002850 * Node specific request. count adjustment happens in
2851 * set_max_huge_pages() after acquiring hugetlb_lock.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002852 */
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002853 init_nodemask_of_node(&nodes_allowed, nid);
2854 n_mask = &nodes_allowed;
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002855 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002856
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002857 err = set_max_huge_pages(h, count, nid, n_mask);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002858
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002859 return err ? err : len;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002860}
2861
David Rientjes238d3c12014-08-06 16:06:51 -07002862static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2863 struct kobject *kobj, const char *buf,
2864 size_t len)
2865{
2866 struct hstate *h;
2867 unsigned long count;
2868 int nid;
2869 int err;
2870
2871 err = kstrtoul(buf, 10, &count);
2872 if (err)
2873 return err;
2874
2875 h = kobj_to_hstate(kobj, &nid);
2876 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2877}
2878
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002879static ssize_t nr_hugepages_show(struct kobject *kobj,
2880 struct kobj_attribute *attr, char *buf)
2881{
2882 return nr_hugepages_show_common(kobj, attr, buf);
2883}
2884
2885static ssize_t nr_hugepages_store(struct kobject *kobj,
2886 struct kobj_attribute *attr, const char *buf, size_t len)
2887{
David Rientjes238d3c12014-08-06 16:06:51 -07002888 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002889}
2890HSTATE_ATTR(nr_hugepages);
2891
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002892#ifdef CONFIG_NUMA
2893
2894/*
2895 * hstate attribute for optionally mempolicy-based constraint on persistent
2896 * huge page alloc/free.
2897 */
2898static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2899 struct kobj_attribute *attr, char *buf)
2900{
2901 return nr_hugepages_show_common(kobj, attr, buf);
2902}
2903
2904static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2905 struct kobj_attribute *attr, const char *buf, size_t len)
2906{
David Rientjes238d3c12014-08-06 16:06:51 -07002907 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002908}
2909HSTATE_ATTR(nr_hugepages_mempolicy);
2910#endif
2911
2912
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002913static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2914 struct kobj_attribute *attr, char *buf)
2915{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002916 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002917 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2918}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002919
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002920static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2921 struct kobj_attribute *attr, const char *buf, size_t count)
2922{
2923 int err;
2924 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002925 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002926
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002927 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002928 return -EINVAL;
2929
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002930 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002931 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002932 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002933
2934 spin_lock(&hugetlb_lock);
2935 h->nr_overcommit_huge_pages = input;
2936 spin_unlock(&hugetlb_lock);
2937
2938 return count;
2939}
2940HSTATE_ATTR(nr_overcommit_hugepages);
2941
2942static ssize_t free_hugepages_show(struct kobject *kobj,
2943 struct kobj_attribute *attr, char *buf)
2944{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002945 struct hstate *h;
2946 unsigned long free_huge_pages;
2947 int nid;
2948
2949 h = kobj_to_hstate(kobj, &nid);
2950 if (nid == NUMA_NO_NODE)
2951 free_huge_pages = h->free_huge_pages;
2952 else
2953 free_huge_pages = h->free_huge_pages_node[nid];
2954
2955 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002956}
2957HSTATE_ATTR_RO(free_hugepages);
2958
2959static ssize_t resv_hugepages_show(struct kobject *kobj,
2960 struct kobj_attribute *attr, char *buf)
2961{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002962 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002963 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2964}
2965HSTATE_ATTR_RO(resv_hugepages);
2966
2967static ssize_t surplus_hugepages_show(struct kobject *kobj,
2968 struct kobj_attribute *attr, char *buf)
2969{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002970 struct hstate *h;
2971 unsigned long surplus_huge_pages;
2972 int nid;
2973
2974 h = kobj_to_hstate(kobj, &nid);
2975 if (nid == NUMA_NO_NODE)
2976 surplus_huge_pages = h->surplus_huge_pages;
2977 else
2978 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2979
2980 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002981}
2982HSTATE_ATTR_RO(surplus_hugepages);
2983
2984static struct attribute *hstate_attrs[] = {
2985 &nr_hugepages_attr.attr,
2986 &nr_overcommit_hugepages_attr.attr,
2987 &free_hugepages_attr.attr,
2988 &resv_hugepages_attr.attr,
2989 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002990#ifdef CONFIG_NUMA
2991 &nr_hugepages_mempolicy_attr.attr,
2992#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002993 NULL,
2994};
2995
Arvind Yadav67e5ed92017-09-06 16:22:06 -07002996static const struct attribute_group hstate_attr_group = {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002997 .attrs = hstate_attrs,
2998};
2999
Jeff Mahoney094e9532010-02-02 13:44:14 -08003000static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3001 struct kobject **hstate_kobjs,
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003002 const struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003003{
3004 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003005 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003006
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003007 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3008 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003009 return -ENOMEM;
3010
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003011 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Miaohe Lin89b2dbd2021-02-24 12:06:50 -08003012 if (retval) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003013 kobject_put(hstate_kobjs[hi]);
Miaohe Lin89b2dbd2021-02-24 12:06:50 -08003014 hstate_kobjs[hi] = NULL;
3015 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003016
3017 return retval;
3018}
3019
3020static void __init hugetlb_sysfs_init(void)
3021{
3022 struct hstate *h;
3023 int err;
3024
3025 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3026 if (!hugepages_kobj)
3027 return;
3028
3029 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003030 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3031 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003032 if (err)
Mike Kravetz282f4212020-06-03 16:00:46 -07003033 pr_err("HugeTLB: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003034 }
3035}
3036
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003037#ifdef CONFIG_NUMA
3038
3039/*
3040 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08003041 * with node devices in node_devices[] using a parallel array. The array
3042 * index of a node device or _hstate == node id.
3043 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003044 * the base kernel, on the hugetlb module.
3045 */
3046struct node_hstate {
3047 struct kobject *hugepages_kobj;
3048 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3049};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08003050static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003051
3052/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003053 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003054 */
3055static struct attribute *per_node_hstate_attrs[] = {
3056 &nr_hugepages_attr.attr,
3057 &free_hugepages_attr.attr,
3058 &surplus_hugepages_attr.attr,
3059 NULL,
3060};
3061
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003062static const struct attribute_group per_node_hstate_attr_group = {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003063 .attrs = per_node_hstate_attrs,
3064};
3065
3066/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003067 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003068 * Returns node id via non-NULL nidp.
3069 */
3070static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3071{
3072 int nid;
3073
3074 for (nid = 0; nid < nr_node_ids; nid++) {
3075 struct node_hstate *nhs = &node_hstates[nid];
3076 int i;
3077 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3078 if (nhs->hstate_kobjs[i] == kobj) {
3079 if (nidp)
3080 *nidp = nid;
3081 return &hstates[i];
3082 }
3083 }
3084
3085 BUG();
3086 return NULL;
3087}
3088
3089/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003090 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003091 * No-op if no hstate attributes attached.
3092 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003093static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003094{
3095 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003096 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003097
3098 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003099 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003100
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003101 for_each_hstate(h) {
3102 int idx = hstate_index(h);
3103 if (nhs->hstate_kobjs[idx]) {
3104 kobject_put(nhs->hstate_kobjs[idx]);
3105 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003106 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003107 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003108
3109 kobject_put(nhs->hugepages_kobj);
3110 nhs->hugepages_kobj = NULL;
3111}
3112
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003113
3114/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003115 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003116 * No-op if attributes already registered.
3117 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003118static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003119{
3120 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003121 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003122 int err;
3123
3124 if (nhs->hugepages_kobj)
3125 return; /* already allocated */
3126
3127 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08003128 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003129 if (!nhs->hugepages_kobj)
3130 return;
3131
3132 for_each_hstate(h) {
3133 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3134 nhs->hstate_kobjs,
3135 &per_node_hstate_attr_group);
3136 if (err) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003137 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003138 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003139 hugetlb_unregister_node(node);
3140 break;
3141 }
3142 }
3143}
3144
3145/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003146 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08003147 * devices of nodes that have memory. All on-line nodes should have
3148 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003149 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08003150static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003151{
3152 int nid;
3153
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08003154 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08003155 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08003156 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003157 hugetlb_register_node(node);
3158 }
3159
3160 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003161 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003162 * [un]register hstate attributes on node hotplug.
3163 */
3164 register_hugetlbfs_with_node(hugetlb_register_node,
3165 hugetlb_unregister_node);
3166}
3167#else /* !CONFIG_NUMA */
3168
3169static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3170{
3171 BUG();
3172 if (nidp)
3173 *nidp = -1;
3174 return NULL;
3175}
3176
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003177static void hugetlb_register_all_nodes(void) { }
3178
3179#endif
3180
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003181static int __init hugetlb_init(void)
3182{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003183 int i;
3184
Mike Kravetzc2833a52020-06-03 16:00:50 -07003185 if (!hugepages_supported()) {
3186 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3187 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07003188 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003189 }
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08003190
Mike Kravetz282f4212020-06-03 16:00:46 -07003191 /*
3192 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3193 * architectures depend on setup being done here.
3194 */
3195 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3196 if (!parsed_default_hugepagesz) {
3197 /*
3198 * If we did not parse a default huge page size, set
3199 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3200 * number of huge pages for this default size was implicitly
3201 * specified, set that here as well.
3202 * Note that the implicit setting will overwrite an explicit
3203 * setting. A warning will be printed in this case.
3204 */
3205 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3206 if (default_hstate_max_huge_pages) {
3207 if (default_hstate.max_huge_pages) {
3208 char buf[32];
Andi Kleenaa888a72008-07-23 21:27:47 -07003209
Mike Kravetz282f4212020-06-03 16:00:46 -07003210 string_get_size(huge_page_size(&default_hstate),
3211 1, STRING_UNITS_2, buf, 32);
3212 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3213 default_hstate.max_huge_pages, buf);
3214 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3215 default_hstate_max_huge_pages);
3216 }
3217 default_hstate.max_huge_pages =
3218 default_hstate_max_huge_pages;
3219 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003220 }
Andi Kleenaa888a72008-07-23 21:27:47 -07003221
Roman Gushchincf11e852020-04-10 14:32:45 -07003222 hugetlb_cma_check();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003223 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07003224 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003225 report_hugepages();
3226
3227 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003228 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08003229 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003230
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003231#ifdef CONFIG_SMP
3232 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3233#else
3234 num_fault_mutexes = 1;
3235#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003236 hugetlb_fault_mutex_table =
Kees Cook6da2ec52018-06-12 13:55:00 -07003237 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3238 GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003239 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003240
3241 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003242 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003243 return 0;
3244}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08003245subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003246
Mike Kravetzae94da82020-06-03 16:00:34 -07003247/* Overwritten by architectures with more huge page sizes */
3248bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003249{
Mike Kravetzae94da82020-06-03 16:00:34 -07003250 return size == HPAGE_SIZE;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003251}
3252
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08003253void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003254{
3255 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003256 unsigned long i;
3257
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003258 if (size_to_hstate(PAGE_SIZE << order)) {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003259 return;
3260 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003261 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003262 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003263 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003264 h->order = order;
3265 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003266 h->nr_huge_pages = 0;
3267 h->free_huge_pages = 0;
3268 for (i = 0; i < MAX_NUMNODES; ++i)
3269 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07003270 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07003271 h->next_nid_to_alloc = first_memory_node;
3272 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003273 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3274 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003275
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003276 parsed_hstate = h;
3277}
3278
Mike Kravetz282f4212020-06-03 16:00:46 -07003279/*
3280 * hugepages command line processing
3281 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3282 * specification. If not, ignore the hugepages value. hugepages can also
3283 * be the first huge page command line option in which case it implicitly
3284 * specifies the number of huge pages for the default size.
3285 */
3286static int __init hugepages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003287{
3288 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003289 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003290
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003291 if (!parsed_valid_hugepagesz) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003292 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003293 parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -07003294 return 0;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003295 }
Mike Kravetz282f4212020-06-03 16:00:46 -07003296
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003297 /*
Mike Kravetz282f4212020-06-03 16:00:46 -07003298 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3299 * yet, so this hugepages= parameter goes to the "default hstate".
3300 * Otherwise, it goes with the previously parsed hugepagesz or
3301 * default_hugepagesz.
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003302 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003303 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003304 mhp = &default_hstate_max_huge_pages;
3305 else
3306 mhp = &parsed_hstate->max_huge_pages;
3307
Andi Kleen8faa8b02008-07-23 21:27:48 -07003308 if (mhp == last_mhp) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003309 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3310 return 0;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003311 }
3312
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003313 if (sscanf(s, "%lu", mhp) <= 0)
3314 *mhp = 0;
3315
Andi Kleen8faa8b02008-07-23 21:27:48 -07003316 /*
3317 * Global state is always initialized later in hugetlb_init.
3318 * But we need to allocate >= MAX_ORDER hstates here early to still
3319 * use the bootmem allocator.
3320 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003321 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07003322 hugetlb_hstate_alloc_pages(parsed_hstate);
3323
3324 last_mhp = mhp;
3325
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003326 return 1;
3327}
Mike Kravetz282f4212020-06-03 16:00:46 -07003328__setup("hugepages=", hugepages_setup);
Nick Piggine11bfbf2008-07-23 21:27:52 -07003329
Mike Kravetz282f4212020-06-03 16:00:46 -07003330/*
3331 * hugepagesz command line processing
3332 * A specific huge page size can only be specified once with hugepagesz.
3333 * hugepagesz is followed by hugepages on the command line. The global
3334 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3335 * hugepagesz argument was valid.
3336 */
Mike Kravetz359f2542020-06-03 16:00:38 -07003337static int __init hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003338{
Mike Kravetz359f2542020-06-03 16:00:38 -07003339 unsigned long size;
Mike Kravetz282f4212020-06-03 16:00:46 -07003340 struct hstate *h;
3341
3342 parsed_valid_hugepagesz = false;
Mike Kravetz359f2542020-06-03 16:00:38 -07003343 size = (unsigned long)memparse(s, NULL);
3344
3345 if (!arch_hugetlb_valid_size(size)) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003346 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
Mike Kravetz359f2542020-06-03 16:00:38 -07003347 return 0;
3348 }
3349
Mike Kravetz282f4212020-06-03 16:00:46 -07003350 h = size_to_hstate(size);
3351 if (h) {
3352 /*
3353 * hstate for this size already exists. This is normally
3354 * an error, but is allowed if the existing hstate is the
3355 * default hstate. More specifically, it is only allowed if
3356 * the number of huge pages for the default hstate was not
3357 * previously specified.
3358 */
3359 if (!parsed_default_hugepagesz || h != &default_hstate ||
3360 default_hstate.max_huge_pages) {
3361 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3362 return 0;
3363 }
3364
3365 /*
3366 * No need to call hugetlb_add_hstate() as hstate already
3367 * exists. But, do set parsed_hstate so that a following
3368 * hugepages= parameter will be applied to this hstate.
3369 */
3370 parsed_hstate = h;
3371 parsed_valid_hugepagesz = true;
3372 return 1;
Mike Kravetz38237832020-06-03 16:00:42 -07003373 }
3374
Mike Kravetz359f2542020-06-03 16:00:38 -07003375 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
Mike Kravetz282f4212020-06-03 16:00:46 -07003376 parsed_valid_hugepagesz = true;
Nick Piggine11bfbf2008-07-23 21:27:52 -07003377 return 1;
3378}
Mike Kravetz359f2542020-06-03 16:00:38 -07003379__setup("hugepagesz=", hugepagesz_setup);
3380
Mike Kravetz282f4212020-06-03 16:00:46 -07003381/*
3382 * default_hugepagesz command line input
3383 * Only one instance of default_hugepagesz allowed on command line.
3384 */
Mike Kravetzae94da82020-06-03 16:00:34 -07003385static int __init default_hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003386{
Mike Kravetzae94da82020-06-03 16:00:34 -07003387 unsigned long size;
3388
Mike Kravetz282f4212020-06-03 16:00:46 -07003389 parsed_valid_hugepagesz = false;
Mike Kravetz282f4212020-06-03 16:00:46 -07003390 if (parsed_default_hugepagesz) {
3391 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3392 return 0;
3393 }
3394
3395 size = (unsigned long)memparse(s, NULL);
3396
3397 if (!arch_hugetlb_valid_size(size)) {
3398 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3399 return 0;
3400 }
3401
3402 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3403 parsed_valid_hugepagesz = true;
3404 parsed_default_hugepagesz = true;
3405 default_hstate_idx = hstate_index(size_to_hstate(size));
3406
3407 /*
3408 * The number of default huge pages (for this size) could have been
3409 * specified as the first hugetlb parameter: hugepages=X. If so,
3410 * then default_hstate_max_huge_pages is set. If the default huge
3411 * page size is gigantic (>= MAX_ORDER), then the pages must be
3412 * allocated here from bootmem allocator.
3413 */
3414 if (default_hstate_max_huge_pages) {
3415 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3416 if (hstate_is_gigantic(&default_hstate))
3417 hugetlb_hstate_alloc_pages(&default_hstate);
3418 default_hstate_max_huge_pages = 0;
3419 }
3420
Nick Piggine11bfbf2008-07-23 21:27:52 -07003421 return 1;
3422}
Mike Kravetzae94da82020-06-03 16:00:34 -07003423__setup("default_hugepagesz=", default_hugepagesz_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003424
Muchun Song8ca39e62020-08-11 18:30:32 -07003425static unsigned int allowed_mems_nr(struct hstate *h)
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003426{
3427 int node;
3428 unsigned int nr = 0;
Muchun Song8ca39e62020-08-11 18:30:32 -07003429 nodemask_t *mpol_allowed;
3430 unsigned int *array = h->free_huge_pages_node;
3431 gfp_t gfp_mask = htlb_alloc_mask(h);
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003432
Muchun Song8ca39e62020-08-11 18:30:32 -07003433 mpol_allowed = policy_nodemask_current(gfp_mask);
3434
3435 for_each_node_mask(node, cpuset_current_mems_allowed) {
3436 if (!mpol_allowed ||
3437 (mpol_allowed && node_isset(node, *mpol_allowed)))
3438 nr += array[node];
3439 }
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003440
3441 return nr;
3442}
3443
3444#ifdef CONFIG_SYSCTL
Muchun Song177437982020-09-04 16:36:13 -07003445static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3446 void *buffer, size_t *length,
3447 loff_t *ppos, unsigned long *out)
3448{
3449 struct ctl_table dup_table;
3450
3451 /*
3452 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3453 * can duplicate the @table and alter the duplicate of it.
3454 */
3455 dup_table = *table;
3456 dup_table.data = out;
3457
3458 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3459}
3460
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003461static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3462 struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003463 void *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464{
Andi Kleene5ff2152008-07-23 21:27:42 -07003465 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07003466 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08003467 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003468
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003469 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003470 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003471
Muchun Song177437982020-09-04 16:36:13 -07003472 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3473 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003474 if (ret)
3475 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003476
David Rientjes238d3c12014-08-06 16:06:51 -07003477 if (write)
3478 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3479 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08003480out:
3481 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482}
Mel Gorman396faf02007-07-17 04:03:13 -07003483
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003484int hugetlb_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003485 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003486{
3487
3488 return hugetlb_sysctl_handler_common(false, table, write,
3489 buffer, length, ppos);
3490}
3491
3492#ifdef CONFIG_NUMA
3493int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003494 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003495{
3496 return hugetlb_sysctl_handler_common(true, table, write,
3497 buffer, length, ppos);
3498}
3499#endif /* CONFIG_NUMA */
3500
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003501int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003502 void *buffer, size_t *length, loff_t *ppos)
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003503{
Andi Kleena5516432008-07-23 21:27:41 -07003504 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07003505 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08003506 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003507
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003508 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003509 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003510
Petr Holasekc033a932011-03-22 16:33:05 -07003511 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07003512
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07003513 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08003514 return -EINVAL;
3515
Muchun Song177437982020-09-04 16:36:13 -07003516 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3517 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003518 if (ret)
3519 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003520
3521 if (write) {
3522 spin_lock(&hugetlb_lock);
3523 h->nr_overcommit_huge_pages = tmp;
3524 spin_unlock(&hugetlb_lock);
3525 }
Michal Hocko08d4a242011-01-13 15:47:26 -08003526out:
3527 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003528}
3529
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530#endif /* CONFIG_SYSCTL */
3531
Alexey Dobriyane1759c22008-10-15 23:50:22 +04003532void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533{
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003534 struct hstate *h;
3535 unsigned long total = 0;
3536
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003537 if (!hugepages_supported())
3538 return;
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003539
3540 for_each_hstate(h) {
3541 unsigned long count = h->nr_huge_pages;
3542
3543 total += (PAGE_SIZE << huge_page_order(h)) * count;
3544
3545 if (h == &default_hstate)
3546 seq_printf(m,
3547 "HugePages_Total: %5lu\n"
3548 "HugePages_Free: %5lu\n"
3549 "HugePages_Rsvd: %5lu\n"
3550 "HugePages_Surp: %5lu\n"
3551 "Hugepagesize: %8lu kB\n",
3552 count,
3553 h->free_huge_pages,
3554 h->resv_huge_pages,
3555 h->surplus_huge_pages,
3556 (PAGE_SIZE << huge_page_order(h)) / 1024);
3557 }
3558
3559 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560}
3561
Joe Perches79815932020-09-16 13:40:43 -07003562int hugetlb_report_node_meminfo(char *buf, int len, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563{
Andi Kleena5516432008-07-23 21:27:41 -07003564 struct hstate *h = &default_hstate;
Joe Perches79815932020-09-16 13:40:43 -07003565
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003566 if (!hugepages_supported())
3567 return 0;
Joe Perches79815932020-09-16 13:40:43 -07003568
3569 return sysfs_emit_at(buf, len,
3570 "Node %d HugePages_Total: %5u\n"
3571 "Node %d HugePages_Free: %5u\n"
3572 "Node %d HugePages_Surp: %5u\n",
3573 nid, h->nr_huge_pages_node[nid],
3574 nid, h->free_huge_pages_node[nid],
3575 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576}
3577
David Rientjes949f7ec2013-04-29 15:07:48 -07003578void hugetlb_show_meminfo(void)
3579{
3580 struct hstate *h;
3581 int nid;
3582
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003583 if (!hugepages_supported())
3584 return;
3585
David Rientjes949f7ec2013-04-29 15:07:48 -07003586 for_each_node_state(nid, N_MEMORY)
3587 for_each_hstate(h)
3588 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3589 nid,
3590 h->nr_huge_pages_node[nid],
3591 h->free_huge_pages_node[nid],
3592 h->surplus_huge_pages_node[nid],
3593 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3594}
3595
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003596void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3597{
3598 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3599 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3600}
3601
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3603unsigned long hugetlb_total_pages(void)
3604{
Wanpeng Lid0028582013-03-22 15:04:40 -07003605 struct hstate *h;
3606 unsigned long nr_total_pages = 0;
3607
3608 for_each_hstate(h)
3609 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3610 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003612
Andi Kleena5516432008-07-23 21:27:41 -07003613static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003614{
3615 int ret = -ENOMEM;
3616
3617 spin_lock(&hugetlb_lock);
3618 /*
3619 * When cpuset is configured, it breaks the strict hugetlb page
3620 * reservation as the accounting is done on a global variable. Such
3621 * reservation is completely rubbish in the presence of cpuset because
3622 * the reservation is not checked against page availability for the
3623 * current cpuset. Application can still potentially OOM'ed by kernel
3624 * with lack of free htlb page in cpuset that the task is in.
3625 * Attempt to enforce strict accounting with cpuset is almost
3626 * impossible (or too ugly) because cpuset is too fluid that
3627 * task or memory node can be dynamically moved between cpusets.
3628 *
3629 * The change of semantics for shared hugetlb mapping with cpuset is
3630 * undesirable. However, in order to preserve some of the semantics,
3631 * we fall back to check against current free page availability as
3632 * a best attempt and hopefully to minimize the impact of changing
3633 * semantics that cpuset has.
Muchun Song8ca39e62020-08-11 18:30:32 -07003634 *
3635 * Apart from cpuset, we also have memory policy mechanism that
3636 * also determines from which node the kernel will allocate memory
3637 * in a NUMA system. So similar to cpuset, we also should consider
3638 * the memory policy of the current task. Similar to the description
3639 * above.
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003640 */
3641 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07003642 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003643 goto out;
3644
Muchun Song8ca39e62020-08-11 18:30:32 -07003645 if (delta > allowed_mems_nr(h)) {
Andi Kleena5516432008-07-23 21:27:41 -07003646 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003647 goto out;
3648 }
3649 }
3650
3651 ret = 0;
3652 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07003653 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003654
3655out:
3656 spin_unlock(&hugetlb_lock);
3657 return ret;
3658}
3659
Andy Whitcroft84afd992008-07-23 21:27:32 -07003660static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3661{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003662 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003663
3664 /*
3665 * This new VMA should share its siblings reservation map if present.
3666 * The VMA will only ever have a valid reservation map pointer where
3667 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003668 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07003669 * after this open call completes. It is therefore safe to take a
3670 * new reference here without additional locking.
3671 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003672 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003673 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003674}
3675
Mel Gormana1e78772008-07-23 21:27:23 -07003676static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3677{
Andi Kleena5516432008-07-23 21:27:41 -07003678 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003679 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07003680 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003681 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003682 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003683
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003684 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3685 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003686
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003687 start = vma_hugecache_offset(h, vma, vma->vm_start);
3688 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003689
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003690 reserve = (end - start) - region_count(resv, start, end);
Mina Almasrye9fe92a2020-04-01 21:11:21 -07003691 hugetlb_cgroup_uncharge_counter(resv, start, end);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003692 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003693 /*
3694 * Decrement reserve counts. The global reserve count may be
3695 * adjusted if the subpool has a minimum size.
3696 */
3697 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3698 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003699 }
Mina Almasrye9fe92a2020-04-01 21:11:21 -07003700
3701 kref_put(&resv->refs, resv_map_release);
Mel Gormana1e78772008-07-23 21:27:23 -07003702}
3703
Dan Williams31383c62017-11-29 16:10:28 -08003704static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3705{
3706 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3707 return -EINVAL;
3708 return 0;
3709}
3710
Dan Williams05ea8862018-04-05 16:24:25 -07003711static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3712{
3713 struct hstate *hstate = hstate_vma(vma);
3714
3715 return 1UL << huge_page_shift(hstate);
3716}
3717
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718/*
3719 * We cannot handle pagefaults against hugetlb pages at all. They cause
3720 * handle_mm_fault() to try to instantiate regular-sized pages in the
3721 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3722 * this far.
3723 */
Souptick Joarderb3ec9f32018-06-07 17:08:04 -07003724static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725{
3726 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003727 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003728}
3729
Jane Chueec36362018-08-02 15:36:05 -07003730/*
3731 * When a new function is introduced to vm_operations_struct and added
3732 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3733 * This is because under System V memory model, mappings created via
3734 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3735 * their original vm_ops are overwritten with shm_vm_ops.
3736 */
Alexey Dobriyanf0f37e2f2009-09-27 22:29:37 +04003737const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003738 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003739 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003740 .close = hugetlb_vm_op_close,
Dan Williams31383c62017-11-29 16:10:28 -08003741 .split = hugetlb_vm_op_split,
Dan Williams05ea8862018-04-05 16:24:25 -07003742 .pagesize = hugetlb_vm_op_pagesize,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743};
3744
David Gibson1e8f8892006-01-06 00:10:44 -08003745static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3746 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003747{
3748 pte_t entry;
3749
David Gibson1e8f8892006-01-06 00:10:44 -08003750 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003751 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3752 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003753 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003754 entry = huge_pte_wrprotect(mk_huge_pte(page,
3755 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003756 }
3757 entry = pte_mkyoung(entry);
3758 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003759 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003760
3761 return entry;
3762}
3763
David Gibson1e8f8892006-01-06 00:10:44 -08003764static void set_huge_ptep_writable(struct vm_area_struct *vma,
3765 unsigned long address, pte_t *ptep)
3766{
3767 pte_t entry;
3768
Gerald Schaefer106c9922013-04-29 15:07:23 -07003769 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003770 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003771 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003772}
3773
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003774bool is_hugetlb_entry_migration(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003775{
3776 swp_entry_t swp;
3777
3778 if (huge_pte_none(pte) || pte_present(pte))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003779 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003780 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07003781 if (is_migration_entry(swp))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003782 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003783 else
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003784 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003785}
3786
Baoquan He3e5c3602020-10-13 16:56:10 -07003787static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003788{
3789 swp_entry_t swp;
3790
3791 if (huge_pte_none(pte) || pte_present(pte))
Baoquan He3e5c3602020-10-13 16:56:10 -07003792 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003793 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07003794 if (is_hwpoison_entry(swp))
Baoquan He3e5c3602020-10-13 16:56:10 -07003795 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003796 else
Baoquan He3e5c3602020-10-13 16:56:10 -07003797 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003798}
David Gibson1e8f8892006-01-06 00:10:44 -08003799
David Gibson63551ae2005-06-21 17:14:44 -07003800int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3801 struct vm_area_struct *vma)
3802{
Mike Kravetz5e415402018-11-16 15:08:04 -08003803 pte_t *src_pte, *dst_pte, entry, dst_entry;
David Gibson63551ae2005-06-21 17:14:44 -07003804 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003805 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003806 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003807 struct hstate *h = hstate_vma(vma);
3808 unsigned long sz = huge_page_size(h);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003809 struct address_space *mapping = vma->vm_file->f_mapping;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003810 struct mmu_notifier_range range;
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003811 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003812
3813 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003814
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003815 if (cow) {
Jérôme Glisse7269f992019-05-13 17:20:53 -07003816 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07003817 vma->vm_start,
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003818 vma->vm_end);
3819 mmu_notifier_invalidate_range_start(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003820 } else {
3821 /*
3822 * For shared mappings i_mmap_rwsem must be held to call
3823 * huge_pte_alloc, otherwise the returned ptep could go
3824 * away if part of a shared pmd and another thread calls
3825 * huge_pmd_unshare.
3826 */
3827 i_mmap_lock_read(mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003828 }
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003829
Andi Kleena5516432008-07-23 21:27:41 -07003830 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003831 spinlock_t *src_ptl, *dst_ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07003832 src_pte = huge_pte_offset(src, addr, sz);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003833 if (!src_pte)
3834 continue;
Peter Xu59caf932021-03-18 17:01:46 +11003835 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003836 if (!dst_pte) {
3837 ret = -ENOMEM;
3838 break;
3839 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003840
Mike Kravetz5e415402018-11-16 15:08:04 -08003841 /*
3842 * If the pagetables are shared don't copy or take references.
3843 * dst_pte == src_pte is the common case of src/dest sharing.
3844 *
3845 * However, src could have 'unshared' and dst shares with
3846 * another vma. If dst_pte !none, this implies sharing.
3847 * Check here before taking page table lock, and once again
3848 * after taking the lock below.
3849 */
3850 dst_entry = huge_ptep_get(dst_pte);
3851 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
Larry Woodmanc5c99422008-01-24 05:49:25 -08003852 continue;
3853
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003854 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3855 src_ptl = huge_pte_lockptr(h, src, src_pte);
3856 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003857 entry = huge_ptep_get(src_pte);
Mike Kravetz5e415402018-11-16 15:08:04 -08003858 dst_entry = huge_ptep_get(dst_pte);
3859 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3860 /*
3861 * Skip if src entry none. Also, skip in the
3862 * unlikely case dst entry !none as this implies
3863 * sharing with another vma.
3864 */
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003865 ;
3866 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3867 is_hugetlb_entry_hwpoisoned(entry))) {
3868 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3869
3870 if (is_write_migration_entry(swp_entry) && cow) {
3871 /*
3872 * COW mappings require pages in both
3873 * parent and child to be set to read.
3874 */
3875 make_migration_entry_read(&swp_entry);
3876 entry = swp_entry_to_pte(swp_entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07003877 set_huge_swap_pte_at(src, addr, src_pte,
3878 entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003879 }
Punit Agrawale5251fd2017-07-06 15:39:50 -07003880 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003881 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003882 if (cow) {
Jérôme Glisse0f108512017-11-15 17:34:07 -08003883 /*
3884 * No need to notify as we are downgrading page
3885 * table protection not changing it to point
3886 * to a new page.
3887 *
Mike Rapoportad56b732018-03-21 21:22:47 +02003888 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08003889 */
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003890 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003891 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003892 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003893 ptepage = pte_page(entry);
3894 get_page(ptepage);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003895 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003896 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003897 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003898 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003899 spin_unlock(src_ptl);
3900 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003901 }
David Gibson63551ae2005-06-21 17:14:44 -07003902
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003903 if (cow)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003904 mmu_notifier_invalidate_range_end(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003905 else
3906 i_mmap_unlock_read(mapping);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003907
3908 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003909}
3910
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003911void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3912 unsigned long start, unsigned long end,
3913 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003914{
3915 struct mm_struct *mm = vma->vm_mm;
3916 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003917 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003918 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003919 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003920 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003921 struct hstate *h = hstate_vma(vma);
3922 unsigned long sz = huge_page_size(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003923 struct mmu_notifier_range range;
Andi Kleena5516432008-07-23 21:27:41 -07003924
David Gibson63551ae2005-06-21 17:14:44 -07003925 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003926 BUG_ON(start & ~huge_page_mask(h));
3927 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003928
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08003929 /*
3930 * This is a hugetlb vma, all the pte entries should point
3931 * to huge page.
3932 */
Peter Zijlstraed6a7932018-08-31 14:46:08 +02003933 tlb_change_page_size(tlb, sz);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003934 tlb_start_vma(tlb, vma);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07003935
3936 /*
3937 * If sharing possible, alert mmu notifiers of worst case.
3938 */
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07003939 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3940 end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003941 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3942 mmu_notifier_invalidate_range_start(&range);
Hillf Danton569f48b82014-12-10 15:44:41 -08003943 address = start;
Hillf Danton569f48b82014-12-10 15:44:41 -08003944 for (; address < end; address += sz) {
Punit Agrawal7868a202017-07-06 15:39:42 -07003945 ptep = huge_pte_offset(mm, address, sz);
Adam Litke4c887262005-10-29 18:16:46 -07003946 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003947 continue;
3948
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003949 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae204f2020-08-11 18:31:38 -07003950 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003951 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07003952 /*
3953 * We just unmapped a page of PMDs by clearing a PUD.
3954 * The caller's TLB flush range should cover this area.
3955 */
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003956 continue;
3957 }
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003958
Hillf Danton66293262012-03-23 15:01:48 -07003959 pte = huge_ptep_get(ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003960 if (huge_pte_none(pte)) {
3961 spin_unlock(ptl);
3962 continue;
3963 }
Hillf Danton66293262012-03-23 15:01:48 -07003964
3965 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003966 * Migrating hugepage or HWPoisoned hugepage is already
3967 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003968 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003969 if (unlikely(!pte_present(pte))) {
Punit Agrawal9386fac2017-07-06 15:39:46 -07003970 huge_pte_clear(mm, address, ptep, sz);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003971 spin_unlock(ptl);
3972 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003973 }
Hillf Danton66293262012-03-23 15:01:48 -07003974
3975 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003976 /*
3977 * If a reference page is supplied, it is because a specific
3978 * page is being unmapped, not a range. Ensure the page we
3979 * are about to unmap is the actual page of interest.
3980 */
3981 if (ref_page) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003982 if (page != ref_page) {
3983 spin_unlock(ptl);
3984 continue;
3985 }
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003986 /*
3987 * Mark the VMA as having unmapped its page so that
3988 * future faults in this VMA will fail rather than
3989 * looking like data was lost
3990 */
3991 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3992 }
3993
David Gibsonc7546f82005-08-05 11:59:35 -07003994 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.Vb528e4b2016-12-12 16:42:37 -08003995 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07003996 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08003997 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07003998
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003999 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004000 page_remove_rmap(page, true);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004001
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004002 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07004003 tlb_remove_page_size(tlb, page, huge_page_size(h));
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004004 /*
4005 * Bail out after unmapping reference page if supplied
4006 */
4007 if (ref_page)
4008 break;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07004009 }
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004010 mmu_notifier_invalidate_range_end(&range);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004011 tlb_end_vma(tlb, vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012}
David Gibson63551ae2005-06-21 17:14:44 -07004013
Mel Gormand8333522012-07-31 16:46:20 -07004014void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4015 struct vm_area_struct *vma, unsigned long start,
4016 unsigned long end, struct page *ref_page)
4017{
4018 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4019
4020 /*
4021 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4022 * test will fail on a vma being torn down, and not grab a page table
4023 * on its way out. We're lucky that the flag has such an appropriate
4024 * name, and can in fact be safely cleared here. We could clear it
4025 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004026 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07004027 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004028 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07004029 */
4030 vma->vm_flags &= ~VM_MAYSHARE;
4031}
4032
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004033void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004034 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004035{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004036 struct mm_struct *mm;
4037 struct mmu_gather tlb;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004038 unsigned long tlb_start = start;
4039 unsigned long tlb_end = end;
4040
4041 /*
4042 * If shared PMDs were possibly used within this vma range, adjust
4043 * start/end for worst case tlb flushing.
4044 * Note that we can not be sure if PMDs are shared until we try to
4045 * unmap pages. However, we want to make sure TLB flushing covers
4046 * the largest possible range.
4047 */
4048 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004049
4050 mm = vma->vm_mm;
4051
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004052 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004053 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004054 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004055}
4056
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004057/*
4058 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4059 * mappping it owns the reserve page for. The intention is to unmap the page
4060 * from other VMAs and let the children be SIGKILLed if they are faulting the
4061 * same region.
4062 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004063static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4064 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004065{
Adam Litke75266742008-11-12 13:24:56 -08004066 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004067 struct vm_area_struct *iter_vma;
4068 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004069 pgoff_t pgoff;
4070
4071 /*
4072 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4073 * from page cache lookup which is in HPAGE_SIZE units.
4074 */
Adam Litke75266742008-11-12 13:24:56 -08004075 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07004076 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4077 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05004078 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004079
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08004080 /*
4081 * Take the mapping lock for the duration of the table walk. As
4082 * this mapping should be shared between all the VMAs,
4083 * __unmap_hugepage_range() is called as the lock is already held
4084 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004085 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07004086 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004087 /* Do not unmap the current VMA */
4088 if (iter_vma == vma)
4089 continue;
4090
4091 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07004092 * Shared VMAs have their own reserves and do not affect
4093 * MAP_PRIVATE accounting but it is possible that a shared
4094 * VMA is using the same page so check and skip such VMAs.
4095 */
4096 if (iter_vma->vm_flags & VM_MAYSHARE)
4097 continue;
4098
4099 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004100 * Unmap the page from other VMAs without their own reserves.
4101 * They get marked to be SIGKILLed if they fault in these
4102 * areas. This is because a future no-page fault on this VMA
4103 * could insert a zeroed page instead of the data existing
4104 * from the time of fork. This would look like data corruption
4105 */
4106 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004107 unmap_hugepage_range(iter_vma, address,
4108 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004109 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004110 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004111}
4112
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004113/*
4114 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08004115 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4116 * cannot race with other handlers or page migration.
4117 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004118 */
Souptick Joarder2b740302018-08-23 17:01:36 -07004119static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Huang Ying974e6d62018-08-17 15:45:57 -07004120 unsigned long address, pte_t *ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004121 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08004122{
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004123 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07004124 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08004125 struct page *old_page, *new_page;
Souptick Joarder2b740302018-08-23 17:01:36 -07004126 int outside_reserve = 0;
4127 vm_fault_t ret = 0;
Huang Ying974e6d62018-08-17 15:45:57 -07004128 unsigned long haddr = address & huge_page_mask(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004129 struct mmu_notifier_range range;
David Gibson1e8f8892006-01-06 00:10:44 -08004130
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004131 pte = huge_ptep_get(ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08004132 old_page = pte_page(pte);
4133
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004134retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08004135 /* If no-one else is actually using this page, avoid the copy
4136 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07004137 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07004138 page_move_anon_rmap(old_page, vma);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004139 set_huge_ptep_writable(vma, haddr, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004140 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004141 }
4142
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004143 /*
4144 * If the process that created a MAP_PRIVATE mapping is about to
4145 * perform a COW due to a shared page count, attempt to satisfy
4146 * the allocation without using the existing reserves. The pagecache
4147 * page is used to determine if the reserve at this address was
4148 * consumed or not. If reserves were used, a partial faulted mapping
4149 * at the time of fork() could consume its reserves on COW instead
4150 * of the full address range.
4151 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07004152 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004153 old_page != pagecache_page)
4154 outside_reserve = 1;
4155
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004156 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004157
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004158 /*
4159 * Drop page table lock as buddy allocator may be called. It will
4160 * be acquired again before returning to the caller, as expected.
4161 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004162 spin_unlock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004163 new_page = alloc_huge_page(vma, haddr, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08004164
Adam Litke2fc39ce2007-11-14 16:59:39 -08004165 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004166 /*
4167 * If a process owning a MAP_PRIVATE mapping fails to COW,
4168 * it is due to references held by a child and an insufficient
4169 * huge page pool. To guarantee the original mappers
4170 * reliability, unmap the page from child processes. The child
4171 * may get SIGKILLed if it later faults.
4172 */
4173 if (outside_reserve) {
Mike Kravetzdf73c802020-12-29 15:14:25 -08004174 struct address_space *mapping = vma->vm_file->f_mapping;
4175 pgoff_t idx;
4176 u32 hash;
4177
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004178 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004179 BUG_ON(huge_pte_none(pte));
Mike Kravetzdf73c802020-12-29 15:14:25 -08004180 /*
4181 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4182 * unmapping. unmapping needs to hold i_mmap_rwsem
4183 * in write mode. Dropping i_mmap_rwsem in read mode
4184 * here is OK as COW mappings do not interact with
4185 * PMD sharing.
4186 *
4187 * Reacquire both after unmap operation.
4188 */
4189 idx = vma_hugecache_offset(h, vma, haddr);
4190 hash = hugetlb_fault_mutex_hash(mapping, idx);
4191 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4192 i_mmap_unlock_read(mapping);
4193
Huang Ying5b7a1d42018-08-17 15:45:53 -07004194 unmap_ref_private(mm, vma, old_page, haddr);
Mike Kravetzdf73c802020-12-29 15:14:25 -08004195
4196 i_mmap_lock_read(mapping);
4197 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004198 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004199 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004200 if (likely(ptep &&
4201 pte_same(huge_ptep_get(ptep), pte)))
4202 goto retry_avoidcopy;
4203 /*
4204 * race occurs while re-acquiring page table
4205 * lock, and our job is done.
4206 */
4207 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004208 }
4209
Souptick Joarder2b740302018-08-23 17:01:36 -07004210 ret = vmf_error(PTR_ERR(new_page));
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004211 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08004212 }
4213
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004214 /*
4215 * When the original hugepage is shared one, it does not have
4216 * anon_vma prepared.
4217 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07004218 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004219 ret = VM_FAULT_OOM;
4220 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07004221 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004222
Huang Ying974e6d62018-08-17 15:45:57 -07004223 copy_user_huge_page(new_page, old_page, address, vma,
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004224 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004225 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004226
Jérôme Glisse7269f992019-05-13 17:20:53 -07004227 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07004228 haddr + huge_page_size(h));
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004229 mmu_notifier_invalidate_range_start(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004230
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004231 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004232 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004233 * before the page tables are altered
4234 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004235 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004236 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07004237 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07004238 ClearPagePrivate(new_page);
4239
David Gibson1e8f8892006-01-06 00:10:44 -08004240 /* Break COW */
Huang Ying5b7a1d42018-08-17 15:45:53 -07004241 huge_ptep_clear_flush(vma, haddr, ptep);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004242 mmu_notifier_invalidate_range(mm, range.start, range.end);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004243 set_huge_pte_at(mm, haddr, ptep,
David Gibson1e8f8892006-01-06 00:10:44 -08004244 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004245 page_remove_rmap(old_page, true);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004246 hugepage_add_new_anon_rmap(new_page, vma, haddr);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004247 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004248 /* Make the old page be freed below */
4249 new_page = old_page;
4250 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004251 spin_unlock(ptl);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004252 mmu_notifier_invalidate_range_end(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004253out_release_all:
Huang Ying5b7a1d42018-08-17 15:45:53 -07004254 restore_reserve_on_error(h, vma, haddr, new_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004255 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004256out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004257 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07004258
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004259 spin_lock(ptl); /* Caller expects lock to be held */
4260 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08004261}
4262
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004263/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07004264static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4265 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004266{
4267 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07004268 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004269
4270 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07004271 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004272
4273 return find_lock_page(mapping, idx);
4274}
4275
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004276/*
4277 * Return whether there is a pagecache page to back given address within VMA.
4278 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4279 */
4280static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004281 struct vm_area_struct *vma, unsigned long address)
4282{
4283 struct address_space *mapping;
4284 pgoff_t idx;
4285 struct page *page;
4286
4287 mapping = vma->vm_file->f_mapping;
4288 idx = vma_hugecache_offset(h, vma, address);
4289
4290 page = find_get_page(mapping, idx);
4291 if (page)
4292 put_page(page);
4293 return page != NULL;
4294}
4295
Mike Kravetzab76ad52015-09-08 15:01:50 -07004296int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4297 pgoff_t idx)
4298{
4299 struct inode *inode = mapping->host;
4300 struct hstate *h = hstate_inode(inode);
4301 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4302
4303 if (err)
4304 return err;
4305 ClearPagePrivate(page);
4306
Mike Kravetz22146c32018-10-26 15:10:58 -07004307 /*
4308 * set page dirty so that it will not be removed from cache/file
4309 * by non-hugetlbfs specific code paths.
4310 */
4311 set_page_dirty(page);
4312
Mike Kravetzab76ad52015-09-08 15:01:50 -07004313 spin_lock(&inode->i_lock);
4314 inode->i_blocks += blocks_per_huge_page(h);
4315 spin_unlock(&inode->i_lock);
4316 return 0;
4317}
4318
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004319static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4320 struct address_space *mapping,
4321 pgoff_t idx,
4322 unsigned int flags,
4323 unsigned long haddr,
4324 unsigned long reason)
4325{
4326 vm_fault_t ret;
4327 u32 hash;
4328 struct vm_fault vmf = {
4329 .vma = vma,
4330 .address = haddr,
4331 .flags = flags,
4332 .vma_flags = vma->vm_flags,
4333 .vma_page_prot = vma->vm_page_prot,
4334
4335 /*
4336 * Hard to debug if it ends up being
4337 * used by a callee that assumes
4338 * something about the other
4339 * uninitialized fields... same as in
4340 * memory.c
4341 */
4342 };
4343
4344 /*
4345 * hugetlb_fault_mutex and i_mmap_rwsem must be
4346 * dropped before handling userfault. Reacquire
4347 * after handling fault to make calling code simpler.
4348 */
4349 hash = hugetlb_fault_mutex_hash(mapping, idx);
4350 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4351 i_mmap_unlock_read(mapping);
4352 ret = handle_userfault(&vmf, reason);
4353 i_mmap_lock_read(mapping);
4354 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4355
4356 return ret;
4357}
4358
Souptick Joarder2b740302018-08-23 17:01:36 -07004359static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4360 struct vm_area_struct *vma,
4361 struct address_space *mapping, pgoff_t idx,
4362 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004363{
Andi Kleena5516432008-07-23 21:27:41 -07004364 struct hstate *h = hstate_vma(vma);
Souptick Joarder2b740302018-08-23 17:01:36 -07004365 vm_fault_t ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08004366 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004367 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07004368 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08004369 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004370 spinlock_t *ptl;
Huang Ying285b8dc2018-06-07 17:08:08 -07004371 unsigned long haddr = address & huge_page_mask(h);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004372 bool new_page = false;
Adam Litke4c887262005-10-29 18:16:46 -07004373
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004374 /*
4375 * Currently, we are forced to kill the process in the event the
4376 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004377 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004378 */
4379 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08004380 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08004381 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004382 return ret;
4383 }
4384
Adam Litke4c887262005-10-29 18:16:46 -07004385 /*
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004386 * We can not race with truncation due to holding i_mmap_rwsem.
4387 * i_size is modified when holding i_mmap_rwsem, so check here
4388 * once for faults beyond end of file.
Adam Litke4c887262005-10-29 18:16:46 -07004389 */
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004390 size = i_size_read(mapping->host) >> huge_page_shift(h);
4391 if (idx >= size)
4392 goto out;
4393
Christoph Lameter6bda6662006-01-06 00:10:49 -08004394retry:
4395 page = find_lock_page(mapping, idx);
4396 if (!page) {
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004397 /* Check for page in userfault range */
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004398 if (userfaultfd_missing(vma)) {
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004399 ret = hugetlb_handle_userfault(vma, mapping, idx,
4400 flags, haddr,
4401 VM_UFFD_MISSING);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004402 goto out;
4403 }
4404
Huang Ying285b8dc2018-06-07 17:08:08 -07004405 page = alloc_huge_page(vma, haddr, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08004406 if (IS_ERR(page)) {
Mike Kravetz4643d672019-08-13 15:38:00 -07004407 /*
4408 * Returning error will result in faulting task being
4409 * sent SIGBUS. The hugetlb fault mutex prevents two
4410 * tasks from racing to fault in the same page which
4411 * could result in false unable to allocate errors.
4412 * Page migration does not take the fault mutex, but
4413 * does a clear then write of pte's under page table
4414 * lock. Page fault code could race with migration,
4415 * notice the clear pte and try to allocate a page
4416 * here. Before returning error, get ptl and make
4417 * sure there really is no pte entry.
4418 */
4419 ptl = huge_pte_lock(h, mm, ptep);
4420 if (!huge_pte_none(huge_ptep_get(ptep))) {
4421 ret = 0;
4422 spin_unlock(ptl);
4423 goto out;
4424 }
4425 spin_unlock(ptl);
Souptick Joarder2b740302018-08-23 17:01:36 -07004426 ret = vmf_error(PTR_ERR(page));
Christoph Lameter6bda6662006-01-06 00:10:49 -08004427 goto out;
4428 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004429 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004430 __SetPageUptodate(page);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004431 new_page = true;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004432
Mel Gormanf83a2752009-05-28 14:34:40 -07004433 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07004434 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004435 if (err) {
4436 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004437 if (err == -EEXIST)
4438 goto retry;
4439 goto out;
4440 }
Mel Gorman23be7462010-04-23 13:17:56 -04004441 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08004442 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004443 if (unlikely(anon_vma_prepare(vma))) {
4444 ret = VM_FAULT_OOM;
4445 goto backout_unlocked;
4446 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08004447 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04004448 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004449 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004450 /*
4451 * If memory error occurs between mmap() and fault, some process
4452 * don't have hwpoisoned swap entry for errored virtual address.
4453 * So we need to block hugepage fault by PG_hwpoison bit check.
4454 */
4455 if (unlikely(PageHWPoison(page))) {
Miaohe Lind3e43af2021-01-12 15:49:24 -08004456 ret = VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004457 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004458 goto backout_unlocked;
4459 }
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004460
4461 /* Check for page in userfault range. */
4462 if (userfaultfd_minor(vma)) {
4463 unlock_page(page);
4464 ret = hugetlb_handle_userfault(vma, mapping, idx,
4465 flags, haddr,
4466 VM_UFFD_MINOR);
4467 goto out;
4468 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08004469 }
David Gibson1e8f8892006-01-06 00:10:44 -08004470
Andy Whitcroft57303d82008-08-12 15:08:47 -07004471 /*
4472 * If we are going to COW a private mapping later, we examine the
4473 * pending reservations for this page now. This will ensure that
4474 * any allocations necessary to record that reservation occur outside
4475 * the spinlock.
4476 */
Mike Kravetz5e911372015-09-08 15:01:28 -07004477 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004478 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004479 ret = VM_FAULT_OOM;
4480 goto backout_unlocked;
4481 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004482 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004483 vma_end_reservation(h, vma, haddr);
Mike Kravetz5e911372015-09-08 15:01:28 -07004484 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07004485
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08004486 ptl = huge_pte_lock(h, mm, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004487 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004488 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07004489 goto backout;
4490
Joonsoo Kim07443a82013-09-11 14:21:58 -07004491 if (anon_rmap) {
4492 ClearPagePrivate(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004493 hugepage_add_new_anon_rmap(page, vma, haddr);
Choi Gi-yongac714902014-04-07 15:37:36 -07004494 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08004495 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08004496 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4497 && (vma->vm_flags & VM_SHARED)));
Huang Ying285b8dc2018-06-07 17:08:08 -07004498 set_huge_pte_at(mm, haddr, ptep, new_pte);
David Gibson1e8f8892006-01-06 00:10:44 -08004499
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08004500 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01004501 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08004502 /* Optimization, do the COW without a second fault */
Huang Ying974e6d62018-08-17 15:45:57 -07004503 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08004504 }
4505
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004506 spin_unlock(ptl);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004507
4508 /*
4509 * Only make newly allocated pages active. Existing pages found
4510 * in the pagecache could be !page_huge_active() if they have been
4511 * isolated for migration.
4512 */
4513 if (new_page)
4514 set_page_huge_active(page);
4515
Adam Litke4c887262005-10-29 18:16:46 -07004516 unlock_page(page);
4517out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004518 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07004519
4520backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004521 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07004522backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07004523 unlock_page(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004524 restore_reserve_on_error(h, vma, haddr, page);
Adam Litke4c887262005-10-29 18:16:46 -07004525 put_page(page);
4526 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004527}
4528
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004529#ifdef CONFIG_SMP
Wei Yang188b04a2019-11-30 17:57:02 -08004530u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004531{
4532 unsigned long key[2];
4533 u32 hash;
4534
Mike Kravetz1b426ba2019-05-13 17:19:41 -07004535 key[0] = (unsigned long) mapping;
4536 key[1] = idx;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004537
Mike Kravetz55254632019-11-30 17:56:30 -08004538 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004539
4540 return hash & (num_fault_mutexes - 1);
4541}
4542#else
4543/*
4544 * For uniprocesor systems we always use a single mutex, so just
4545 * return 0 and avoid the hashing overhead.
4546 */
Wei Yang188b04a2019-11-30 17:57:02 -08004547u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004548{
4549 return 0;
4550}
4551#endif
4552
Souptick Joarder2b740302018-08-23 17:01:36 -07004553vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01004554 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08004555{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004556 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004557 spinlock_t *ptl;
Souptick Joarder2b740302018-08-23 17:01:36 -07004558 vm_fault_t ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004559 u32 hash;
4560 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004561 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07004562 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07004563 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004564 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004565 int need_wait_lock = 0;
Huang Ying285b8dc2018-06-07 17:08:08 -07004566 unsigned long haddr = address & huge_page_mask(h);
Adam Litke86e52162006-01-06 00:10:43 -08004567
Huang Ying285b8dc2018-06-07 17:08:08 -07004568 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004569 if (ptep) {
Mike Kravetzc0d03812020-04-01 21:11:05 -07004570 /*
4571 * Since we hold no locks, ptep could be stale. That is
4572 * OK as we are only making decisions based on content and
4573 * not actually modifying content here.
4574 */
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004575 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09004576 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004577 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09004578 return 0;
4579 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07004580 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004581 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004582 }
4583
Mike Kravetzc0d03812020-04-01 21:11:05 -07004584 /*
4585 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004586 * until finished with ptep. This serves two purposes:
4587 * 1) It prevents huge_pmd_unshare from being called elsewhere
4588 * and making the ptep no longer valid.
4589 * 2) It synchronizes us with i_size modifications during truncation.
Mike Kravetzc0d03812020-04-01 21:11:05 -07004590 *
4591 * ptep could have already be assigned via huge_pte_offset. That
4592 * is OK, as huge_pte_alloc will return the same value unless
4593 * something has changed.
4594 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004595 mapping = vma->vm_file->f_mapping;
Mike Kravetzc0d03812020-04-01 21:11:05 -07004596 i_mmap_lock_read(mapping);
Peter Xu59caf932021-03-18 17:01:46 +11004597 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
Mike Kravetzc0d03812020-04-01 21:11:05 -07004598 if (!ptep) {
4599 i_mmap_unlock_read(mapping);
4600 return VM_FAULT_OOM;
4601 }
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004602
David Gibson3935baa2006-03-22 00:08:53 -08004603 /*
4604 * Serialize hugepage allocation and instantiation, so that we don't
4605 * get spurious allocation failures if two CPUs race to instantiate
4606 * the same page in the page cache.
4607 */
Mike Kravetzc0d03812020-04-01 21:11:05 -07004608 idx = vma_hugecache_offset(h, vma, haddr);
Wei Yang188b04a2019-11-30 17:57:02 -08004609 hash = hugetlb_fault_mutex_hash(mapping, idx);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07004610 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004611
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004612 entry = huge_ptep_get(ptep);
4613 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004614 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07004615 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08004616 }
Adam Litke86e52162006-01-06 00:10:43 -08004617
Nick Piggin83c54072007-07-19 01:47:05 -07004618 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004619
Andy Whitcroft57303d82008-08-12 15:08:47 -07004620 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004621 * entry could be a migration/hwpoison entry at this point, so this
4622 * check prevents the kernel from going below assuming that we have
Ethon Paul7c8de352020-06-04 16:49:07 -07004623 * an active hugepage in pagecache. This goto expects the 2nd page
4624 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4625 * properly handle it.
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004626 */
4627 if (!pte_present(entry))
4628 goto out_mutex;
4629
4630 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07004631 * If we are going to COW the mapping later, we examine the pending
4632 * reservations for this page now. This will ensure that any
4633 * allocations necessary to record that reservation occur outside the
4634 * spinlock. For private mappings, we also lookup the pagecache
4635 * page now as it is used to determine if a reservation has been
4636 * consumed.
4637 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07004638 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004639 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004640 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07004641 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07004642 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004643 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004644 vma_end_reservation(h, vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004645
Mel Gormanf83a2752009-05-28 14:34:40 -07004646 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07004647 pagecache_page = hugetlbfs_pagecache_page(h,
Huang Ying285b8dc2018-06-07 17:08:08 -07004648 vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004649 }
4650
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004651 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004652
David Gibson1e8f8892006-01-06 00:10:44 -08004653 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07004654 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004655 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07004656
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004657 /*
4658 * hugetlb_cow() requires page locks of pte_page(entry) and
4659 * pagecache_page, so here we need take the former one
4660 * when page != pagecache_page or !pagecache_page.
4661 */
4662 page = pte_page(entry);
4663 if (page != pagecache_page)
4664 if (!trylock_page(page)) {
4665 need_wait_lock = 1;
4666 goto out_ptl;
4667 }
4668
4669 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07004670
Hugh Dickins788c7df2009-06-23 13:49:05 +01004671 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07004672 if (!huge_pte_write(entry)) {
Huang Ying974e6d62018-08-17 15:45:57 -07004673 ret = hugetlb_cow(mm, vma, address, ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004674 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004675 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07004676 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07004677 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07004678 }
4679 entry = pte_mkyoung(entry);
Huang Ying285b8dc2018-06-07 17:08:08 -07004680 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
Hugh Dickins788c7df2009-06-23 13:49:05 +01004681 flags & FAULT_FLAG_WRITE))
Huang Ying285b8dc2018-06-07 17:08:08 -07004682 update_mmu_cache(vma, haddr, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004683out_put_page:
4684 if (page != pagecache_page)
4685 unlock_page(page);
4686 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004687out_ptl:
4688 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004689
4690 if (pagecache_page) {
4691 unlock_page(pagecache_page);
4692 put_page(pagecache_page);
4693 }
David Gibsonb4d1d992008-10-15 22:01:11 -07004694out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07004695 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004696 i_mmap_unlock_read(mapping);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004697 /*
4698 * Generally it's safe to hold refcount during waiting page lock. But
4699 * here we just wait to defer the next page fault to avoid busy loop and
4700 * the page is not used after unlocked before returning from the current
4701 * page fault. So we are safe from accessing freed page, even if we wait
4702 * here without taking refcount.
4703 */
4704 if (need_wait_lock)
4705 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08004706 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08004707}
4708
Axel Rasmussene6bf0762021-03-18 17:01:50 +11004709#ifdef CONFIG_USERFAULTFD
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004710/*
4711 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4712 * modifications for huge pages.
4713 */
4714int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4715 pte_t *dst_pte,
4716 struct vm_area_struct *dst_vma,
4717 unsigned long dst_addr,
4718 unsigned long src_addr,
4719 struct page **pagep)
4720{
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004721 struct address_space *mapping;
4722 pgoff_t idx;
4723 unsigned long size;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004724 int vm_shared = dst_vma->vm_flags & VM_SHARED;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004725 struct hstate *h = hstate_vma(dst_vma);
4726 pte_t _dst_pte;
4727 spinlock_t *ptl;
4728 int ret;
4729 struct page *page;
4730
4731 if (!*pagep) {
4732 ret = -ENOMEM;
4733 page = alloc_huge_page(dst_vma, dst_addr, 0);
4734 if (IS_ERR(page))
4735 goto out;
4736
4737 ret = copy_huge_page_from_user(page,
4738 (const void __user *) src_addr,
Mike Kravetz810a56b2017-02-22 15:42:58 -08004739 pages_per_huge_page(h), false);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004740
Michel Lespinassec1e8d7c2020-06-08 21:33:54 -07004741 /* fallback to copy_from_user outside mmap_lock */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004742 if (unlikely(ret)) {
Andrea Arcangeli9e368252018-11-30 14:09:25 -08004743 ret = -ENOENT;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004744 *pagep = page;
4745 /* don't free the page */
4746 goto out;
4747 }
4748 } else {
4749 page = *pagep;
4750 *pagep = NULL;
4751 }
4752
4753 /*
4754 * The memory barrier inside __SetPageUptodate makes sure that
4755 * preceding stores to the page contents become visible before
4756 * the set_pte_at() write.
4757 */
4758 __SetPageUptodate(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004759
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004760 mapping = dst_vma->vm_file->f_mapping;
4761 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4762
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004763 /*
4764 * If shared, add to page cache
4765 */
4766 if (vm_shared) {
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004767 size = i_size_read(mapping->host) >> huge_page_shift(h);
4768 ret = -EFAULT;
4769 if (idx >= size)
4770 goto out_release_nounlock;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004771
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004772 /*
4773 * Serialization between remove_inode_hugepages() and
4774 * huge_add_to_page_cache() below happens through the
4775 * hugetlb_fault_mutex_table that here must be hold by
4776 * the caller.
4777 */
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004778 ret = huge_add_to_page_cache(page, mapping, idx);
4779 if (ret)
4780 goto out_release_nounlock;
4781 }
4782
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004783 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4784 spin_lock(ptl);
4785
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004786 /*
4787 * Recheck the i_size after holding PT lock to make sure not
4788 * to leave any page mapped (as page_mapped()) beyond the end
4789 * of the i_size (remove_inode_hugepages() is strict about
4790 * enforcing that). If we bail out here, we'll also leave a
4791 * page in the radix tree in the vm_shared case beyond the end
4792 * of the i_size, but remove_inode_hugepages() will take care
4793 * of it as soon as we drop the hugetlb_fault_mutex_table.
4794 */
4795 size = i_size_read(mapping->host) >> huge_page_shift(h);
4796 ret = -EFAULT;
4797 if (idx >= size)
4798 goto out_release_unlock;
4799
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004800 ret = -EEXIST;
4801 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4802 goto out_release_unlock;
4803
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004804 if (vm_shared) {
4805 page_dup_rmap(page, true);
4806 } else {
4807 ClearPagePrivate(page);
4808 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4809 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004810
4811 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4812 if (dst_vma->vm_flags & VM_WRITE)
4813 _dst_pte = huge_pte_mkdirty(_dst_pte);
4814 _dst_pte = pte_mkyoung(_dst_pte);
4815
4816 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4817
4818 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4819 dst_vma->vm_flags & VM_WRITE);
4820 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4821
4822 /* No need to invalidate - it was non-present before */
4823 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4824
4825 spin_unlock(ptl);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004826 set_page_huge_active(page);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004827 if (vm_shared)
4828 unlock_page(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004829 ret = 0;
4830out:
4831 return ret;
4832out_release_unlock:
4833 spin_unlock(ptl);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004834 if (vm_shared)
4835 unlock_page(page);
Andrea Arcangeli5af10df2017-08-10 15:23:38 -07004836out_release_nounlock:
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004837 put_page(page);
4838 goto out;
4839}
Axel Rasmussene6bf0762021-03-18 17:01:50 +11004840#endif /* CONFIG_USERFAULTFD */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004841
Michel Lespinasse28a35712013-02-22 16:35:55 -08004842long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4843 struct page **pages, struct vm_area_struct **vmas,
4844 unsigned long *position, unsigned long *nr_pages,
Peter Xu4f6da932020-04-01 21:07:58 -07004845 long i, unsigned int flags, int *locked)
David Gibson63551ae2005-06-21 17:14:44 -07004846{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004847 unsigned long pfn_offset;
4848 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08004849 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07004850 struct hstate *h = hstate_vma(vma);
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004851 int err = -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07004852
David Gibson63551ae2005-06-21 17:14:44 -07004853 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07004854 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004855 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004856 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07004857 struct page *page;
4858
4859 /*
David Rientjes02057962015-04-14 15:48:24 -07004860 * If we have a pending SIGKILL, don't keep faulting pages and
4861 * potentially allocating memory.
4862 */
Davidlohr Buesofa45f112019-01-03 15:28:55 -08004863 if (fatal_signal_pending(current)) {
David Rientjes02057962015-04-14 15:48:24 -07004864 remainder = 0;
4865 break;
4866 }
4867
4868 /*
Adam Litke4c887262005-10-29 18:16:46 -07004869 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004870 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07004871 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004872 *
4873 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07004874 */
Punit Agrawal7868a202017-07-06 15:39:42 -07004875 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4876 huge_page_size(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004877 if (pte)
4878 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004879 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07004880
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004881 /*
4882 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004883 * an error where there's an empty slot with no huge pagecache
4884 * to back it. This way, we avoid allocating a hugepage, and
4885 * the sparse dumpfile avoids allocating disk blocks, but its
4886 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004887 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004888 if (absent && (flags & FOLL_DUMP) &&
4889 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004890 if (pte)
4891 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004892 remainder = 0;
4893 break;
4894 }
4895
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07004896 /*
4897 * We need call hugetlb_fault for both hugepages under migration
4898 * (in which case hugetlb_fault waits for the migration,) and
4899 * hwpoisoned hugepages (in which case we need to prevent the
4900 * caller from accessing to them.) In order to do this, we use
4901 * here is_swap_pte instead of is_hugetlb_entry_migration and
4902 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4903 * both cases, and because we can't follow correct pages
4904 * directly from any kind of swap entries.
4905 */
4906 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07004907 ((flags & FOLL_WRITE) &&
4908 !huge_pte_write(huge_ptep_get(pte)))) {
Souptick Joarder2b740302018-08-23 17:01:36 -07004909 vm_fault_t ret;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004910 unsigned int fault_flags = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004911
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004912 if (pte)
4913 spin_unlock(ptl);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004914 if (flags & FOLL_WRITE)
4915 fault_flags |= FAULT_FLAG_WRITE;
Peter Xu4f6da932020-04-01 21:07:58 -07004916 if (locked)
Peter Xu71335f32020-04-01 21:08:53 -07004917 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4918 FAULT_FLAG_KILLABLE;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004919 if (flags & FOLL_NOWAIT)
4920 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4921 FAULT_FLAG_RETRY_NOWAIT;
4922 if (flags & FOLL_TRIED) {
Peter Xu4426e942020-04-01 21:08:49 -07004923 /*
4924 * Note: FAULT_FLAG_ALLOW_RETRY and
4925 * FAULT_FLAG_TRIED can co-exist
4926 */
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004927 fault_flags |= FAULT_FLAG_TRIED;
4928 }
4929 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4930 if (ret & VM_FAULT_ERROR) {
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004931 err = vm_fault_to_errno(ret, flags);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004932 remainder = 0;
4933 break;
4934 }
4935 if (ret & VM_FAULT_RETRY) {
Peter Xu4f6da932020-04-01 21:07:58 -07004936 if (locked &&
Andrea Arcangeli1ac25012019-02-01 14:20:16 -08004937 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
Peter Xu4f6da932020-04-01 21:07:58 -07004938 *locked = 0;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004939 *nr_pages = 0;
4940 /*
4941 * VM_FAULT_RETRY must not return an
4942 * error, it will return zero
4943 * instead.
4944 *
4945 * No need to update "position" as the
4946 * caller will not check it after
4947 * *nr_pages is set to 0.
4948 */
4949 return i;
4950 }
4951 continue;
Adam Litke4c887262005-10-29 18:16:46 -07004952 }
David Gibson63551ae2005-06-21 17:14:44 -07004953
Andi Kleena5516432008-07-23 21:27:41 -07004954 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004955 page = pte_page(huge_ptep_get(pte));
Linus Torvalds8fde12c2019-04-11 10:49:19 -07004956
4957 /*
Zhigang Luacbfb082019-11-30 17:57:06 -08004958 * If subpage information not requested, update counters
4959 * and skip the same_page loop below.
4960 */
4961 if (!pages && !vmas && !pfn_offset &&
4962 (vaddr + huge_page_size(h) < vma->vm_end) &&
4963 (remainder >= pages_per_huge_page(h))) {
4964 vaddr += huge_page_size(h);
4965 remainder -= pages_per_huge_page(h);
4966 i += pages_per_huge_page(h);
4967 spin_unlock(ptl);
4968 continue;
4969 }
4970
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004971same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004972 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004973 pages[i] = mem_map_offset(page, pfn_offset);
John Hubbard3faa52c2020-04-01 21:05:29 -07004974 /*
4975 * try_grab_page() should always succeed here, because:
4976 * a) we hold the ptl lock, and b) we've just checked
4977 * that the huge page is present in the page tables. If
4978 * the huge page is present, then the tail pages must
4979 * also be present. The ptl prevents the head page and
4980 * tail pages from being rearranged in any way. So this
4981 * page must be available at this point, unless the page
4982 * refcount overflowed:
4983 */
4984 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4985 spin_unlock(ptl);
4986 remainder = 0;
4987 err = -ENOMEM;
4988 break;
4989 }
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004990 }
David Gibson63551ae2005-06-21 17:14:44 -07004991
4992 if (vmas)
4993 vmas[i] = vma;
4994
4995 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004996 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07004997 --remainder;
4998 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004999 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07005000 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08005001 /*
5002 * We use pfn_offset to avoid touching the pageframes
5003 * of this compound page.
5004 */
5005 goto same_page;
5006 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005007 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07005008 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08005009 *nr_pages = remainder;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005010 /*
5011 * setting position is actually required only if remainder is
5012 * not zero but it's faster not to add a "if (remainder)"
5013 * branch.
5014 */
David Gibson63551ae2005-06-21 17:14:44 -07005015 *position = vaddr;
5016
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07005017 return i ? i : err;
David Gibson63551ae2005-06-21 17:14:44 -07005018}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005019
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005020unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005021 unsigned long address, unsigned long end, pgprot_t newprot)
5022{
5023 struct mm_struct *mm = vma->vm_mm;
5024 unsigned long start = address;
5025 pte_t *ptep;
5026 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07005027 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005028 unsigned long pages = 0;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005029 bool shared_pmd = false;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005030 struct mmu_notifier_range range;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005031
5032 /*
5033 * In the case of shared PMDs, the area to flush could be beyond
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005034 * start/end. Set range.start/range.end to cover the maximum possible
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005035 * range if PMD sharing is possible.
5036 */
Jérôme Glisse7269f992019-05-13 17:20:53 -07005037 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5038 0, vma, mm, start, end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005039 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005040
5041 BUG_ON(address >= end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005042 flush_cache_range(vma, range.start, range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005043
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005044 mmu_notifier_invalidate_range_start(&range);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005045 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07005046 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005047 spinlock_t *ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07005048 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005049 if (!ptep)
5050 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005051 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae204f2020-08-11 18:31:38 -07005052 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005053 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005054 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005055 shared_pmd = true;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08005056 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005057 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005058 pte = huge_ptep_get(ptep);
5059 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5060 spin_unlock(ptl);
5061 continue;
5062 }
5063 if (unlikely(is_hugetlb_entry_migration(pte))) {
5064 swp_entry_t entry = pte_to_swp_entry(pte);
5065
5066 if (is_write_migration_entry(entry)) {
5067 pte_t newpte;
5068
5069 make_migration_entry_read(&entry);
5070 newpte = swp_entry_to_pte(entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07005071 set_huge_swap_pte_at(mm, address, ptep,
5072 newpte, huge_page_size(h));
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005073 pages++;
5074 }
5075 spin_unlock(ptl);
5076 continue;
5077 }
5078 if (!huge_pte_none(pte)) {
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005079 pte_t old_pte;
5080
5081 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5082 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08005083 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005084 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005085 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005086 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005087 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005088 }
Mel Gormand8333522012-07-31 16:46:20 -07005089 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005090 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07005091 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005092 * once we release i_mmap_rwsem, another task can do the final put_page
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005093 * and that page table be reused and filled with junk. If we actually
5094 * did unshare a page of pmds, flush the range corresponding to the pud.
Mel Gormand8333522012-07-31 16:46:20 -07005095 */
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005096 if (shared_pmd)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005097 flush_hugetlb_tlb_range(vma, range.start, range.end);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005098 else
5099 flush_hugetlb_tlb_range(vma, start, end);
Jérôme Glisse0f108512017-11-15 17:34:07 -08005100 /*
5101 * No need to call mmu_notifier_invalidate_range() we are downgrading
5102 * page table protection not changing it to point to a new page.
5103 *
Mike Rapoportad56b732018-03-21 21:22:47 +02005104 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08005105 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005106 i_mmap_unlock_write(vma->vm_file->f_mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005107 mmu_notifier_invalidate_range_end(&range);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005108
5109 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005110}
5111
Mel Gormana1e78772008-07-23 21:27:23 -07005112int hugetlb_reserve_pages(struct inode *inode,
5113 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00005114 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005115 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07005116{
Mina Almasry0db9d742020-04-01 21:11:25 -07005117 long ret, chg, add = -1;
Andi Kleena5516432008-07-23 21:27:41 -07005118 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07005119 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005120 struct resv_map *resv_map;
Mina Almasry075a61d2020-04-01 21:11:28 -07005121 struct hugetlb_cgroup *h_cg = NULL;
Mina Almasry0db9d742020-04-01 21:11:25 -07005122 long gbl_reserve, regions_needed = 0;
Adam Litkee4e574b2007-10-16 01:26:19 -07005123
Mike Kravetz63489f82018-03-22 16:17:13 -07005124 /* This should never happen */
5125 if (from > to) {
5126 VM_WARN(1, "%s called with a negative range\n", __func__);
5127 return -EINVAL;
5128 }
5129
Mel Gormana1e78772008-07-23 21:27:23 -07005130 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00005131 * Only apply hugepage reservation if asked. At fault time, an
5132 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07005133 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00005134 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005135 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00005136 return 0;
5137
5138 /*
Mel Gormana1e78772008-07-23 21:27:23 -07005139 * Shared mappings base their reservation on the number of pages that
5140 * are already allocated on behalf of the file. Private mappings need
5141 * to reserve the full area even if read-only as mprotect() may be
5142 * called to make the mapping read-write. Assume !vma is a shm mapping
5143 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005144 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzf27a5132019-05-13 17:22:55 -07005145 /*
5146 * resv_map can not be NULL as hugetlb_reserve_pages is only
5147 * called for inodes for which resv_maps were created (see
5148 * hugetlbfs_get_inode).
5149 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005150 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005151
Mina Almasry0db9d742020-04-01 21:11:25 -07005152 chg = region_chg(resv_map, from, to, &regions_needed);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005153
5154 } else {
Mina Almasrye9fe92a2020-04-01 21:11:21 -07005155 /* Private mapping. */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005156 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00005157 if (!resv_map)
5158 return -ENOMEM;
5159
Mel Gorman17c9d122009-02-11 16:34:16 +00005160 chg = to - from;
5161
Mel Gorman5a6fe122009-02-10 14:02:27 +00005162 set_vma_resv_map(vma, resv_map);
5163 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5164 }
5165
Dave Hansenc50ac052012-05-29 15:06:46 -07005166 if (chg < 0) {
5167 ret = chg;
5168 goto out_err;
5169 }
Mel Gorman17c9d122009-02-11 16:34:16 +00005170
Mina Almasry075a61d2020-04-01 21:11:28 -07005171 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5172 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5173
5174 if (ret < 0) {
5175 ret = -ENOMEM;
5176 goto out_err;
5177 }
5178
5179 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5180 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5181 * of the resv_map.
5182 */
5183 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5184 }
5185
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005186 /*
5187 * There must be enough pages in the subpool for the mapping. If
5188 * the subpool has a minimum size, there may be some global
5189 * reservations already in place (gbl_reserve).
5190 */
5191 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5192 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07005193 ret = -ENOSPC;
Mina Almasry075a61d2020-04-01 21:11:28 -07005194 goto out_uncharge_cgroup;
Dave Hansenc50ac052012-05-29 15:06:46 -07005195 }
Mel Gorman17c9d122009-02-11 16:34:16 +00005196
5197 /*
5198 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07005199 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00005200 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005201 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00005202 if (ret < 0) {
Mina Almasry075a61d2020-04-01 21:11:28 -07005203 goto out_put_pages;
Mel Gorman17c9d122009-02-11 16:34:16 +00005204 }
5205
5206 /*
5207 * Account for the reservations made. Shared mappings record regions
5208 * that have reservations as they are shared by multiple VMAs.
5209 * When the last VMA disappears, the region map says how much
5210 * the reservation was and the page cache tells how much of
5211 * the reservation was consumed. Private mappings are per-VMA and
5212 * only the consumed reservations are tracked. When the VMA
5213 * disappears, the original reservation is the VMA size and the
5214 * consumed reservations are stored in the map. Hence, nothing
5215 * else has to be done for private mappings here
5216 */
Mike Kravetz33039672015-06-24 16:57:58 -07005217 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07005218 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005219
Mina Almasry0db9d742020-04-01 21:11:25 -07005220 if (unlikely(add < 0)) {
5221 hugetlb_acct_memory(h, -gbl_reserve);
Dan Carpenter9b52a372020-12-14 19:12:11 -08005222 ret = add;
Mina Almasry075a61d2020-04-01 21:11:28 -07005223 goto out_put_pages;
Mina Almasry0db9d742020-04-01 21:11:25 -07005224 } else if (unlikely(chg > add)) {
Mike Kravetz33039672015-06-24 16:57:58 -07005225 /*
5226 * pages in this range were added to the reserve
5227 * map between region_chg and region_add. This
5228 * indicates a race with alloc_huge_page. Adjust
5229 * the subpool and reserve counts modified above
5230 * based on the difference.
5231 */
5232 long rsv_adjust;
5233
Miaohe Linfe03ccc2021-03-24 21:37:17 -07005234 /*
5235 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5236 * reference to h_cg->css. See comment below for detail.
5237 */
Mina Almasry075a61d2020-04-01 21:11:28 -07005238 hugetlb_cgroup_uncharge_cgroup_rsvd(
5239 hstate_index(h),
5240 (chg - add) * pages_per_huge_page(h), h_cg);
5241
Mike Kravetz33039672015-06-24 16:57:58 -07005242 rsv_adjust = hugepage_subpool_put_pages(spool,
5243 chg - add);
5244 hugetlb_acct_memory(h, -rsv_adjust);
Miaohe Linfe03ccc2021-03-24 21:37:17 -07005245 } else if (h_cg) {
5246 /*
5247 * The file_regions will hold their own reference to
5248 * h_cg->css. So we should release the reference held
5249 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5250 * done.
5251 */
5252 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005253 }
5254 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005255 return 0;
Mina Almasry075a61d2020-04-01 21:11:28 -07005256out_put_pages:
5257 /* put back original number of pages, chg */
5258 (void)hugepage_subpool_put_pages(spool, chg);
5259out_uncharge_cgroup:
5260 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5261 chg * pages_per_huge_page(h), h_cg);
Dave Hansenc50ac052012-05-29 15:06:46 -07005262out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07005263 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mina Almasry0db9d742020-04-01 21:11:25 -07005264 /* Only call region_abort if the region_chg succeeded but the
5265 * region_add failed or didn't run.
5266 */
5267 if (chg >= 0 && add < 0)
5268 region_abort(resv_map, from, to, regions_needed);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07005269 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5270 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07005271 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005272}
5273
Mike Kravetzb5cec282015-09-08 15:01:41 -07005274long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5275 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005276{
Andi Kleena5516432008-07-23 21:27:41 -07005277 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005278 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005279 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07005280 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005281 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08005282
Mike Kravetzf27a5132019-05-13 17:22:55 -07005283 /*
5284 * Since this routine can be called in the evict inode path for all
5285 * hugetlbfs inodes, resv_map could be NULL.
5286 */
Mike Kravetzb5cec282015-09-08 15:01:41 -07005287 if (resv_map) {
5288 chg = region_del(resv_map, start, end);
5289 /*
5290 * region_del() can fail in the rare case where a region
5291 * must be split and another region descriptor can not be
5292 * allocated. If end == LONG_MAX, it will not fail.
5293 */
5294 if (chg < 0)
5295 return chg;
5296 }
5297
Ken Chen45c682a2007-11-14 16:59:44 -08005298 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07005299 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08005300 spin_unlock(&inode->i_lock);
5301
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005302 /*
5303 * If the subpool has a minimum size, the number of global
5304 * reservations to be released may be adjusted.
5305 */
5306 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5307 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07005308
5309 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005310}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09005311
Steve Capper3212b532013-04-23 12:35:02 +01005312#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5313static unsigned long page_table_shareable(struct vm_area_struct *svma,
5314 struct vm_area_struct *vma,
5315 unsigned long addr, pgoff_t idx)
5316{
5317 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5318 svma->vm_start;
5319 unsigned long sbase = saddr & PUD_MASK;
5320 unsigned long s_end = sbase + PUD_SIZE;
5321
5322 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08005323 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5324 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01005325
5326 /*
5327 * match the virtual addresses, permission and the alignment of the
5328 * page table page.
5329 */
5330 if (pmd_index(addr) != pmd_index(saddr) ||
5331 vm_flags != svm_flags ||
5332 sbase < svma->vm_start || svma->vm_end < s_end)
5333 return 0;
5334
5335 return saddr;
5336}
5337
Nicholas Krause31aafb42015-09-04 15:47:58 -07005338static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01005339{
5340 unsigned long base = addr & PUD_MASK;
5341 unsigned long end = base + PUD_SIZE;
5342
5343 /*
5344 * check on proper vm_flags and page table alignment
5345 */
Mike Kravetz017b1662018-10-05 15:51:29 -07005346 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
Nicholas Krause31aafb42015-09-04 15:47:58 -07005347 return true;
5348 return false;
Steve Capper3212b532013-04-23 12:35:02 +01005349}
5350
Peter Xuf1d8bd52021-03-18 17:01:46 +11005351bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5352{
Peter Xuf1d8bd52021-03-18 17:01:46 +11005353#ifdef CONFIG_USERFAULTFD
5354 if (uffd_disable_huge_pmd_share(vma))
5355 return false;
5356#endif
5357 return vma_shareable(vma, addr);
5358}
5359
Steve Capper3212b532013-04-23 12:35:02 +01005360/*
Mike Kravetz017b1662018-10-05 15:51:29 -07005361 * Determine if start,end range within vma could be mapped by shared pmd.
5362 * If yes, adjust start and end to cover range associated with possible
5363 * shared pmd mappings.
5364 */
5365void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5366 unsigned long *start, unsigned long *end)
5367{
Li Xinhaie3359522021-02-24 12:06:54 -08005368 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5369 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005370
Li Xinhaie3359522021-02-24 12:06:54 -08005371 /*
5372 * vma need span at least one aligned PUD size and the start,end range
5373 * must at least partialy within it.
5374 */
5375 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5376 (*end <= v_start) || (*start >= v_end))
Mike Kravetz017b1662018-10-05 15:51:29 -07005377 return;
5378
Peter Xu75802ca62020-08-06 23:26:11 -07005379 /* Extend the range to be PUD aligned for a worst case scenario */
Li Xinhaie3359522021-02-24 12:06:54 -08005380 if (*start > v_start)
5381 *start = ALIGN_DOWN(*start, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005382
Li Xinhaie3359522021-02-24 12:06:54 -08005383 if (*end < v_end)
5384 *end = ALIGN(*end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005385}
5386
5387/*
Steve Capper3212b532013-04-23 12:35:02 +01005388 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5389 * and returns the corresponding pte. While this is not necessary for the
5390 * !shared pmd case because we can allocate the pmd later as well, it makes the
Mike Kravetzc0d03812020-04-01 21:11:05 -07005391 * code much cleaner.
5392 *
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005393 * This routine must be called with i_mmap_rwsem held in at least read mode if
5394 * sharing is possible. For hugetlbfs, this prevents removal of any page
5395 * table entries associated with the address space. This is important as we
5396 * are setting up sharing based on existing page table entries (mappings).
5397 *
5398 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5399 * huge_pte_alloc know that sharing is not possible and do not take
5400 * i_mmap_rwsem as a performance optimization. This is handled by the
5401 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5402 * only required for subsequent processing.
Steve Capper3212b532013-04-23 12:35:02 +01005403 */
Peter Xu59caf932021-03-18 17:01:46 +11005404pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5405 unsigned long addr, pud_t *pud)
Steve Capper3212b532013-04-23 12:35:02 +01005406{
Steve Capper3212b532013-04-23 12:35:02 +01005407 struct address_space *mapping = vma->vm_file->f_mapping;
5408 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5409 vma->vm_pgoff;
5410 struct vm_area_struct *svma;
5411 unsigned long saddr;
5412 pte_t *spte = NULL;
5413 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005414 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01005415
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005416 i_mmap_assert_locked(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005417 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5418 if (svma == vma)
5419 continue;
5420
5421 saddr = page_table_shareable(svma, vma, addr, idx);
5422 if (saddr) {
Punit Agrawal7868a202017-07-06 15:39:42 -07005423 spte = huge_pte_offset(svma->vm_mm, saddr,
5424 vma_mmu_pagesize(svma));
Steve Capper3212b532013-04-23 12:35:02 +01005425 if (spte) {
5426 get_page(virt_to_page(spte));
5427 break;
5428 }
5429 }
5430 }
5431
5432 if (!spte)
5433 goto out;
5434
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08005435 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005436 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01005437 pud_populate(mm, pud,
5438 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07005439 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005440 } else {
Steve Capper3212b532013-04-23 12:35:02 +01005441 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005442 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005443 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01005444out:
5445 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Steve Capper3212b532013-04-23 12:35:02 +01005446 return pte;
5447}
5448
5449/*
5450 * unmap huge page backed by shared pte.
5451 *
5452 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5453 * indicated by page_count > 1, unmap is achieved by clearing pud and
5454 * decrementing the ref count. If count == 1, the pte page is not shared.
5455 *
Mike Kravetzc0d03812020-04-01 21:11:05 -07005456 * Called with page table lock held and i_mmap_rwsem held in write mode.
Steve Capper3212b532013-04-23 12:35:02 +01005457 *
5458 * returns: 1 successfully unmapped a shared pte page
5459 * 0 the underlying pte page is not shared, or it is the last user
5460 */
Mike Kravetz34ae204f2020-08-11 18:31:38 -07005461int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5462 unsigned long *addr, pte_t *ptep)
Steve Capper3212b532013-04-23 12:35:02 +01005463{
5464 pgd_t *pgd = pgd_offset(mm, *addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005465 p4d_t *p4d = p4d_offset(pgd, *addr);
5466 pud_t *pud = pud_offset(p4d, *addr);
Steve Capper3212b532013-04-23 12:35:02 +01005467
Mike Kravetz34ae204f2020-08-11 18:31:38 -07005468 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005469 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5470 if (page_count(virt_to_page(ptep)) == 1)
5471 return 0;
5472
5473 pud_clear(pud);
5474 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005475 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01005476 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5477 return 1;
5478}
Peter Xuf1d8bd52021-03-18 17:01:46 +11005479
Steve Capper9e5fc742013-04-30 08:02:03 +01005480#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
Peter Xu59caf932021-03-18 17:01:46 +11005481pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct vma,
5482 unsigned long addr, pud_t *pud)
Steve Capper9e5fc742013-04-30 08:02:03 +01005483{
5484 return NULL;
5485}
Zhang Zhene81f2d22015-06-24 16:56:13 -07005486
Mike Kravetz34ae204f2020-08-11 18:31:38 -07005487int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5488 unsigned long *addr, pte_t *ptep)
Zhang Zhene81f2d22015-06-24 16:56:13 -07005489{
5490 return 0;
5491}
Mike Kravetz017b1662018-10-05 15:51:29 -07005492
5493void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5494 unsigned long *start, unsigned long *end)
5495{
5496}
Peter Xu8b03b572021-03-18 17:01:47 +11005497
5498bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5499{
5500 return false;
5501}
Steve Capper3212b532013-04-23 12:35:02 +01005502#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5503
Steve Capper9e5fc742013-04-30 08:02:03 +01005504#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
Peter Xu59caf932021-03-18 17:01:46 +11005505pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
Steve Capper9e5fc742013-04-30 08:02:03 +01005506 unsigned long addr, unsigned long sz)
5507{
5508 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005509 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01005510 pud_t *pud;
5511 pte_t *pte = NULL;
5512
5513 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovf4f0a3d2017-11-29 16:11:30 -08005514 p4d = p4d_alloc(mm, pgd, addr);
5515 if (!p4d)
5516 return NULL;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005517 pud = pud_alloc(mm, p4d, addr);
Steve Capper9e5fc742013-04-30 08:02:03 +01005518 if (pud) {
5519 if (sz == PUD_SIZE) {
5520 pte = (pte_t *)pud;
5521 } else {
5522 BUG_ON(sz != PMD_SIZE);
Peter Xuf1d8bd52021-03-18 17:01:46 +11005523 if (want_pmd_share(vma, addr) && pud_none(*pud))
Peter Xu59caf932021-03-18 17:01:46 +11005524 pte = huge_pmd_share(mm, vma, addr, pud);
Steve Capper9e5fc742013-04-30 08:02:03 +01005525 else
5526 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5527 }
5528 }
Michal Hocko4e666312016-08-02 14:02:34 -07005529 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01005530
5531 return pte;
5532}
5533
Punit Agrawal9b19df22017-09-06 16:21:01 -07005534/*
5535 * huge_pte_offset() - Walk the page table to resolve the hugepage
5536 * entry at address @addr
5537 *
Li Xinhai8ac0b812020-06-03 16:00:53 -07005538 * Return: Pointer to page table entry (PUD or PMD) for
5539 * address @addr, or NULL if a !p*d_present() entry is encountered and the
Punit Agrawal9b19df22017-09-06 16:21:01 -07005540 * size @sz doesn't match the hugepage size at this level of the page
5541 * table.
5542 */
Punit Agrawal7868a202017-07-06 15:39:42 -07005543pte_t *huge_pte_offset(struct mm_struct *mm,
5544 unsigned long addr, unsigned long sz)
Steve Capper9e5fc742013-04-30 08:02:03 +01005545{
5546 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005547 p4d_t *p4d;
Li Xinhai8ac0b812020-06-03 16:00:53 -07005548 pud_t *pud;
5549 pmd_t *pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01005550
5551 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005552 if (!pgd_present(*pgd))
5553 return NULL;
5554 p4d = p4d_offset(pgd, addr);
5555 if (!p4d_present(*p4d))
5556 return NULL;
Punit Agrawal9b19df22017-09-06 16:21:01 -07005557
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005558 pud = pud_offset(p4d, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07005559 if (sz == PUD_SIZE)
5560 /* must be pud huge, non-present or none */
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005561 return (pte_t *)pud;
Li Xinhai8ac0b812020-06-03 16:00:53 -07005562 if (!pud_present(*pud))
5563 return NULL;
5564 /* must have a valid entry and size to go further */
Punit Agrawal9b19df22017-09-06 16:21:01 -07005565
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005566 pmd = pmd_offset(pud, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07005567 /* must be pmd huge, non-present or none */
5568 return (pte_t *)pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01005569}
5570
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08005571#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5572
5573/*
5574 * These functions are overwritable if your architecture needs its own
5575 * behavior.
5576 */
5577struct page * __weak
5578follow_huge_addr(struct mm_struct *mm, unsigned long address,
5579 int write)
5580{
5581 return ERR_PTR(-EINVAL);
5582}
5583
5584struct page * __weak
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -07005585follow_huge_pd(struct vm_area_struct *vma,
5586 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5587{
5588 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5589 return NULL;
5590}
5591
5592struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01005593follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005594 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01005595{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005596 struct page *page = NULL;
5597 spinlock_t *ptl;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005598 pte_t pte;
John Hubbard3faa52c2020-04-01 21:05:29 -07005599
5600 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5601 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5602 (FOLL_PIN | FOLL_GET)))
5603 return NULL;
5604
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005605retry:
5606 ptl = pmd_lockptr(mm, pmd);
5607 spin_lock(ptl);
5608 /*
5609 * make sure that the address range covered by this pmd is not
5610 * unmapped from other threads.
5611 */
5612 if (!pmd_huge(*pmd))
5613 goto out;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005614 pte = huge_ptep_get((pte_t *)pmd);
5615 if (pte_present(pte)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07005616 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
John Hubbard3faa52c2020-04-01 21:05:29 -07005617 /*
5618 * try_grab_page() should always succeed here, because: a) we
5619 * hold the pmd (ptl) lock, and b) we've just checked that the
5620 * huge pmd (head) page is present in the page tables. The ptl
5621 * prevents the head page and tail pages from being rearranged
5622 * in any way. So this page must be available at this point,
5623 * unless the page refcount overflowed:
5624 */
5625 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5626 page = NULL;
5627 goto out;
5628 }
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005629 } else {
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005630 if (is_hugetlb_entry_migration(pte)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005631 spin_unlock(ptl);
5632 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5633 goto retry;
5634 }
5635 /*
5636 * hwpoisoned entry is treated as no_page_table in
5637 * follow_page_mask().
5638 */
5639 }
5640out:
5641 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01005642 return page;
5643}
5644
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08005645struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01005646follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005647 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01005648{
John Hubbard3faa52c2020-04-01 21:05:29 -07005649 if (flags & (FOLL_GET | FOLL_PIN))
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005650 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01005651
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005652 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01005653}
5654
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07005655struct page * __weak
5656follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5657{
John Hubbard3faa52c2020-04-01 21:05:29 -07005658 if (flags & (FOLL_GET | FOLL_PIN))
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07005659 return NULL;
5660
5661 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5662}
5663
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005664bool isolate_huge_page(struct page *page, struct list_head *list)
5665{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005666 bool ret = true;
5667
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005668 spin_lock(&hugetlb_lock);
Muchun Song5b9631c2021-02-04 18:32:10 -08005669 if (!PageHeadHuge(page) || !page_huge_active(page) ||
5670 !get_page_unless_zero(page)) {
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005671 ret = false;
5672 goto unlock;
5673 }
5674 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005675 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005676unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005677 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005678 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005679}
5680
5681void putback_active_hugepage(struct page *page)
5682{
Sasha Levin309381fea2014-01-23 15:52:54 -08005683 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005684 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005685 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005686 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5687 spin_unlock(&hugetlb_lock);
5688 put_page(page);
5689}
Michal Hockoab5ac902018-01-31 16:20:48 -08005690
5691void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5692{
5693 struct hstate *h = page_hstate(oldpage);
5694
5695 hugetlb_cgroup_migrate(oldpage, newpage);
5696 set_page_owner_migrate_reason(newpage, reason);
5697
5698 /*
5699 * transfer temporary state of the new huge page. This is
5700 * reverse to other transitions because the newpage is going to
5701 * be final while the old one will be freed so it takes over
5702 * the temporary status.
5703 *
5704 * Also note that we have to transfer the per-node surplus state
5705 * here as well otherwise the global surplus count will not match
5706 * the per-node's.
5707 */
5708 if (PageHugeTemporary(newpage)) {
5709 int old_nid = page_to_nid(oldpage);
5710 int new_nid = page_to_nid(newpage);
5711
5712 SetPageHugeTemporary(oldpage);
5713 ClearPageHugeTemporary(newpage);
5714
5715 spin_lock(&hugetlb_lock);
5716 if (h->surplus_huge_pages_node[old_nid]) {
5717 h->surplus_huge_pages_node[old_nid]--;
5718 h->surplus_huge_pages_node[new_nid]++;
5719 }
5720 spin_unlock(&hugetlb_lock);
5721 }
5722}
Roman Gushchincf11e852020-04-10 14:32:45 -07005723
Peter Xu343cacf2021-03-18 17:01:47 +11005724/*
5725 * This function will unconditionally remove all the shared pmd pgtable entries
5726 * within the specific vma for a hugetlbfs memory range.
5727 */
5728void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
5729{
5730 struct hstate *h = hstate_vma(vma);
5731 unsigned long sz = huge_page_size(h);
5732 struct mm_struct *mm = vma->vm_mm;
5733 struct mmu_notifier_range range;
5734 unsigned long address, start, end;
5735 spinlock_t *ptl;
5736 pte_t *ptep;
5737
5738 if (!(vma->vm_flags & VM_MAYSHARE))
5739 return;
5740
5741 start = ALIGN(vma->vm_start, PUD_SIZE);
5742 end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
5743
5744 if (start >= end)
5745 return;
5746
5747 /*
5748 * No need to call adjust_range_if_pmd_sharing_possible(), because
5749 * we have already done the PUD_SIZE alignment.
5750 */
5751 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
5752 start, end);
5753 mmu_notifier_invalidate_range_start(&range);
5754 i_mmap_lock_write(vma->vm_file->f_mapping);
5755 for (address = start; address < end; address += PUD_SIZE) {
5756 unsigned long tmp = address;
5757
5758 ptep = huge_pte_offset(mm, address, sz);
5759 if (!ptep)
5760 continue;
5761 ptl = huge_pte_lock(h, mm, ptep);
5762 /* We don't want 'address' to be changed */
5763 huge_pmd_unshare(mm, vma, &tmp, ptep);
5764 spin_unlock(ptl);
5765 }
5766 flush_hugetlb_tlb_range(vma, start, end);
5767 i_mmap_unlock_write(vma->vm_file->f_mapping);
5768 /*
5769 * No need to call mmu_notifier_invalidate_range(), see
5770 * Documentation/vm/mmu_notifier.rst.
5771 */
5772 mmu_notifier_invalidate_range_end(&range);
5773}
5774
Roman Gushchincf11e852020-04-10 14:32:45 -07005775#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -07005776static bool cma_reserve_called __initdata;
5777
5778static int __init cmdline_parse_hugetlb_cma(char *p)
5779{
5780 hugetlb_cma_size = memparse(p, &p);
5781 return 0;
5782}
5783
5784early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5785
5786void __init hugetlb_cma_reserve(int order)
5787{
5788 unsigned long size, reserved, per_node;
5789 int nid;
5790
5791 cma_reserve_called = true;
5792
5793 if (!hugetlb_cma_size)
5794 return;
5795
5796 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5797 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5798 (PAGE_SIZE << order) / SZ_1M);
5799 return;
5800 }
5801
5802 /*
5803 * If 3 GB area is requested on a machine with 4 numa nodes,
5804 * let's allocate 1 GB on first three nodes and ignore the last one.
5805 */
5806 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5807 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5808 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5809
5810 reserved = 0;
5811 for_each_node_state(nid, N_ONLINE) {
5812 int res;
Barry Song2281f792020-08-24 11:03:09 +12005813 char name[CMA_MAX_NAME];
Roman Gushchincf11e852020-04-10 14:32:45 -07005814
5815 size = min(per_node, hugetlb_cma_size - reserved);
5816 size = round_up(size, PAGE_SIZE << order);
5817
Barry Song2281f792020-08-24 11:03:09 +12005818 snprintf(name, sizeof(name), "hugetlb%d", nid);
Roman Gushchincf11e852020-04-10 14:32:45 -07005819 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
Barry Song29d0f412020-08-11 18:32:00 -07005820 0, false, name,
Roman Gushchincf11e852020-04-10 14:32:45 -07005821 &hugetlb_cma[nid], nid);
5822 if (res) {
5823 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5824 res, nid);
5825 continue;
5826 }
5827
5828 reserved += size;
5829 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5830 size / SZ_1M, nid);
5831
5832 if (reserved >= hugetlb_cma_size)
5833 break;
5834 }
5835}
5836
5837void __init hugetlb_cma_check(void)
5838{
5839 if (!hugetlb_cma_size || cma_reserve_called)
5840 return;
5841
5842 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5843}
5844
5845#endif /* CONFIG_CMA */