blob: a4682bbdf3529edfed073f1917dfc095dcaab9ad [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>
Michal Hockoab5ac902018-01-31 16:20:48 -080042#include <linux/page_owner.h>
Nick Piggin7835e982006-03-22 00:08:40 -080043#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070045int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070046unsigned int default_hstate_idx;
47struct hstate hstates[HUGE_MAX_HSTATE];
Roman Gushchincf11e852020-04-10 14:32:45 -070048
Barry Songdbda8fe2020-07-23 21:15:30 -070049#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -070050static struct cma *hugetlb_cma[MAX_NUMNODES];
Barry Songdbda8fe2020-07-23 21:15:30 -070051#endif
52static unsigned long hugetlb_cma_size __initdata;
Roman Gushchincf11e852020-04-10 14:32:45 -070053
Naoya Horiguchi641844f2015-06-24 16:56:59 -070054/*
55 * Minimum page order among possible hugepage sizes, set to a proper value
56 * at boot time.
57 */
58static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070059
Jon Tollefson53ba51d2008-07-23 21:27:52 -070060__initdata LIST_HEAD(huge_boot_pages);
61
Andi Kleene5ff2152008-07-23 21:27:42 -070062/* for command line parsing */
63static struct hstate * __initdata parsed_hstate;
64static unsigned long __initdata default_hstate_max_huge_pages;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -070065static bool __initdata parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -070066static bool __initdata parsed_default_hugepagesz;
Andi Kleene5ff2152008-07-23 21:27:42 -070067
David Gibson3935baa2006-03-22 00:08:53 -080068/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070069 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
70 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080071 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070072DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080073
Davidlohr Bueso8382d912014-04-03 14:47:31 -070074/*
75 * Serializes faults on the same logical page. This is used to
76 * prevent spurious OOMs when the hugepage pool is fully utilized.
77 */
78static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070079struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070080
James Houghton43390f12023-01-04 23:19:10 +000081static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
82 unsigned long start, unsigned long end);
83
Muchun Songe334b1f2021-02-04 18:32:06 -080084static inline bool PageHugeFreed(struct page *head)
85{
86 return page_private(head + 4) == -1UL;
87}
88
89static inline void SetPageHugeFreed(struct page *head)
90{
91 set_page_private(head + 4, -1UL);
92}
93
94static inline void ClearPageHugeFreed(struct page *head)
95{
96 set_page_private(head + 4, 0);
97}
98
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070099/* Forward declaration */
100static int hugetlb_acct_memory(struct hstate *h, long delta);
101
David Gibson90481622012-03-21 16:34:12 -0700102static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
103{
104 bool free = (spool->count == 0) && (spool->used_hpages == 0);
105
106 spin_unlock(&spool->lock);
107
108 /* If no pages are used, and no other handles to the subpool
Ethon Paul7c8de352020-06-04 16:49:07 -0700109 * remain, give up any reservations based on minimum size and
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700110 * free the subpool */
111 if (free) {
112 if (spool->min_hpages != -1)
113 hugetlb_acct_memory(spool->hstate,
114 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -0700115 kfree(spool);
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700116 }
David Gibson90481622012-03-21 16:34:12 -0700117}
118
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700119struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
120 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -0700121{
122 struct hugepage_subpool *spool;
123
Mike Kravetzc6a91822015-04-15 16:13:36 -0700124 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -0700125 if (!spool)
126 return NULL;
127
128 spin_lock_init(&spool->lock);
129 spool->count = 1;
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700130 spool->max_hpages = max_hpages;
131 spool->hstate = h;
132 spool->min_hpages = min_hpages;
133
134 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
135 kfree(spool);
136 return NULL;
137 }
138 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700139
140 return spool;
141}
142
143void hugepage_put_subpool(struct hugepage_subpool *spool)
144{
145 spin_lock(&spool->lock);
146 BUG_ON(!spool->count);
147 spool->count--;
148 unlock_or_release_subpool(spool);
149}
150
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700151/*
152 * Subpool accounting for allocating and reserving pages.
153 * Return -ENOMEM if there are not enough resources to satisfy the
Randy Dunlap9e7ee402020-08-11 18:32:59 -0700154 * request. Otherwise, return the number of pages by which the
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700155 * global pools must be adjusted (upward). The returned value may
156 * only be different than the passed value (delta) in the case where
Ethon Paul7c8de352020-06-04 16:49:07 -0700157 * a subpool minimum size must be maintained.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700158 */
159static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700160 long delta)
161{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700162 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700163
164 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700165 return ret;
David Gibson90481622012-03-21 16:34:12 -0700166
167 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700168
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700169 if (spool->max_hpages != -1) { /* maximum size accounting */
170 if ((spool->used_hpages + delta) <= spool->max_hpages)
171 spool->used_hpages += delta;
172 else {
173 ret = -ENOMEM;
174 goto unlock_ret;
175 }
176 }
177
Mike Kravetz09a95e22016-05-19 17:11:01 -0700178 /* minimum size accounting */
179 if (spool->min_hpages != -1 && spool->rsv_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700180 if (delta > spool->rsv_hpages) {
181 /*
182 * Asking for more reserves than those already taken on
183 * behalf of subpool. Return difference.
184 */
185 ret = delta - spool->rsv_hpages;
186 spool->rsv_hpages = 0;
187 } else {
188 ret = 0; /* reserves already accounted for */
189 spool->rsv_hpages -= delta;
190 }
191 }
192
193unlock_ret:
194 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700195 return ret;
196}
197
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700198/*
199 * Subpool accounting for freeing and unreserving pages.
200 * Return the number of global page reservations that must be dropped.
201 * The return value may only be different than the passed value (delta)
202 * in the case where a subpool minimum size must be maintained.
203 */
204static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700205 long delta)
206{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700207 long ret = delta;
208
David Gibson90481622012-03-21 16:34:12 -0700209 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700210 return delta;
David Gibson90481622012-03-21 16:34:12 -0700211
212 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700213
214 if (spool->max_hpages != -1) /* maximum size accounting */
215 spool->used_hpages -= delta;
216
Mike Kravetz09a95e22016-05-19 17:11:01 -0700217 /* minimum size accounting */
218 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700219 if (spool->rsv_hpages + delta <= spool->min_hpages)
220 ret = 0;
221 else
222 ret = spool->rsv_hpages + delta - spool->min_hpages;
223
224 spool->rsv_hpages += delta;
225 if (spool->rsv_hpages > spool->min_hpages)
226 spool->rsv_hpages = spool->min_hpages;
227 }
228
229 /*
230 * If hugetlbfs_put_super couldn't free spool due to an outstanding
231 * quota reference, free it now.
232 */
David Gibson90481622012-03-21 16:34:12 -0700233 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700234
235 return ret;
David Gibson90481622012-03-21 16:34:12 -0700236}
237
238static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
239{
240 return HUGETLBFS_SB(inode->i_sb)->spool;
241}
242
243static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
244{
Al Viro496ad9a2013-01-23 17:07:38 -0500245 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700246}
247
Mina Almasry0db9d742020-04-01 21:11:25 -0700248/* Helper that removes a struct file_region from the resv_map cache and returns
249 * it for use.
250 */
251static struct file_region *
252get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
253{
254 struct file_region *nrg = NULL;
255
256 VM_BUG_ON(resv->region_cache_count <= 0);
257
258 resv->region_cache_count--;
259 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
Mina Almasry0db9d742020-04-01 21:11:25 -0700260 list_del(&nrg->link);
261
262 nrg->from = from;
263 nrg->to = to;
264
265 return nrg;
266}
267
Mina Almasry075a61d2020-04-01 21:11:28 -0700268static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
269 struct file_region *rg)
270{
271#ifdef CONFIG_CGROUP_HUGETLB
272 nrg->reservation_counter = rg->reservation_counter;
273 nrg->css = rg->css;
274 if (rg->css)
275 css_get(rg->css);
276#endif
277}
278
279/* Helper that records hugetlb_cgroup uncharge info. */
280static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
281 struct hstate *h,
282 struct resv_map *resv,
283 struct file_region *nrg)
284{
285#ifdef CONFIG_CGROUP_HUGETLB
286 if (h_cg) {
287 nrg->reservation_counter =
288 &h_cg->rsvd_hugepage[hstate_index(h)];
289 nrg->css = &h_cg->css;
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700290 /*
291 * The caller will hold exactly one h_cg->css reference for the
292 * whole contiguous reservation region. But this area might be
293 * scattered when there are already some file_regions reside in
294 * it. As a result, many file_regions may share only one css
295 * reference. In order to ensure that one file_region must hold
296 * exactly one h_cg->css reference, we should do css_get for
297 * each file_region and leave the reference held by caller
298 * untouched.
299 */
300 css_get(&h_cg->css);
Mina Almasry075a61d2020-04-01 21:11:28 -0700301 if (!resv->pages_per_hpage)
302 resv->pages_per_hpage = pages_per_huge_page(h);
303 /* pages_per_hpage should be the same for all entries in
304 * a resv_map.
305 */
306 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
307 } else {
308 nrg->reservation_counter = NULL;
309 nrg->css = NULL;
310 }
311#endif
312}
313
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700314static void put_uncharge_info(struct file_region *rg)
315{
316#ifdef CONFIG_CGROUP_HUGETLB
317 if (rg->css)
318 css_put(rg->css);
319#endif
320}
321
Mina Almasrya9b3f862020-04-01 21:11:35 -0700322static bool has_same_uncharge_info(struct file_region *rg,
323 struct file_region *org)
324{
325#ifdef CONFIG_CGROUP_HUGETLB
326 return rg && org &&
327 rg->reservation_counter == org->reservation_counter &&
328 rg->css == org->css;
329
330#else
331 return true;
332#endif
333}
334
335static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
336{
337 struct file_region *nrg = NULL, *prg = NULL;
338
339 prg = list_prev_entry(rg, link);
340 if (&prg->link != &resv->regions && prg->to == rg->from &&
341 has_same_uncharge_info(prg, rg)) {
342 prg->to = rg->to;
343
344 list_del(&rg->link);
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700345 put_uncharge_info(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700346 kfree(rg);
347
Wei Yang7db5e7b2020-10-13 16:56:20 -0700348 rg = prg;
Mina Almasrya9b3f862020-04-01 21:11:35 -0700349 }
350
351 nrg = list_next_entry(rg, link);
352 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
353 has_same_uncharge_info(nrg, rg)) {
354 nrg->from = rg->from;
355
356 list_del(&rg->link);
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700357 put_uncharge_info(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700358 kfree(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700359 }
360}
361
Wei Yang972a3da32020-10-13 16:56:30 -0700362/*
363 * Must be called with resv->lock held.
364 *
365 * Calling this with regions_needed != NULL will count the number of pages
366 * to be added but will not modify the linked list. And regions_needed will
367 * indicate the number of file_regions needed in the cache to carry out to add
368 * the regions for this range.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800369 */
370static long add_reservation_in_range(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700371 struct hugetlb_cgroup *h_cg,
Wei Yang972a3da32020-10-13 16:56:30 -0700372 struct hstate *h, long *regions_needed)
Mina Almasryd75c6af2019-11-30 17:56:59 -0800373{
Mina Almasry0db9d742020-04-01 21:11:25 -0700374 long add = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800375 struct list_head *head = &resv->regions;
Mina Almasry0db9d742020-04-01 21:11:25 -0700376 long last_accounted_offset = f;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800377 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
378
Mina Almasry0db9d742020-04-01 21:11:25 -0700379 if (regions_needed)
380 *regions_needed = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800381
Mina Almasry0db9d742020-04-01 21:11:25 -0700382 /* In this loop, we essentially handle an entry for the range
383 * [last_accounted_offset, rg->from), at every iteration, with some
384 * bounds checking.
385 */
386 list_for_each_entry_safe(rg, trg, head, link) {
387 /* Skip irrelevant regions that start before our range. */
388 if (rg->from < f) {
389 /* If this region ends after the last accounted offset,
390 * then we need to update last_accounted_offset.
391 */
392 if (rg->to > last_accounted_offset)
393 last_accounted_offset = rg->to;
394 continue;
395 }
Mina Almasryd75c6af2019-11-30 17:56:59 -0800396
Mina Almasry0db9d742020-04-01 21:11:25 -0700397 /* When we find a region that starts beyond our range, we've
398 * finished.
399 */
Mina Almasryd75c6af2019-11-30 17:56:59 -0800400 if (rg->from > t)
401 break;
402
Mina Almasry0db9d742020-04-01 21:11:25 -0700403 /* Add an entry for last_accounted_offset -> rg->from, and
404 * update last_accounted_offset.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800405 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700406 if (rg->from > last_accounted_offset) {
407 add += rg->from - last_accounted_offset;
Wei Yang972a3da32020-10-13 16:56:30 -0700408 if (!regions_needed) {
Mina Almasry0db9d742020-04-01 21:11:25 -0700409 nrg = get_file_region_entry_from_cache(
410 resv, last_accounted_offset, rg->from);
Mina Almasry075a61d2020-04-01 21:11:28 -0700411 record_hugetlb_cgroup_uncharge_info(h_cg, h,
412 resv, nrg);
Mina Almasry0db9d742020-04-01 21:11:25 -0700413 list_add(&nrg->link, rg->link.prev);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700414 coalesce_file_region(resv, nrg);
Wei Yang972a3da32020-10-13 16:56:30 -0700415 } else
Mina Almasry0db9d742020-04-01 21:11:25 -0700416 *regions_needed += 1;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800417 }
Mina Almasryd75c6af2019-11-30 17:56:59 -0800418
Mina Almasry0db9d742020-04-01 21:11:25 -0700419 last_accounted_offset = rg->to;
420 }
421
422 /* Handle the case where our range extends beyond
423 * last_accounted_offset.
424 */
425 if (last_accounted_offset < t) {
426 add += t - last_accounted_offset;
Wei Yang972a3da32020-10-13 16:56:30 -0700427 if (!regions_needed) {
Mina Almasry0db9d742020-04-01 21:11:25 -0700428 nrg = get_file_region_entry_from_cache(
429 resv, last_accounted_offset, t);
Mina Almasry075a61d2020-04-01 21:11:28 -0700430 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
Mina Almasry0db9d742020-04-01 21:11:25 -0700431 list_add(&nrg->link, rg->link.prev);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700432 coalesce_file_region(resv, nrg);
Wei Yang972a3da32020-10-13 16:56:30 -0700433 } else
Mina Almasry0db9d742020-04-01 21:11:25 -0700434 *regions_needed += 1;
435 }
436
437 VM_BUG_ON(add < 0);
438 return add;
439}
440
441/* Must be called with resv->lock acquired. Will drop lock to allocate entries.
442 */
443static int allocate_file_region_entries(struct resv_map *resv,
444 int regions_needed)
445 __must_hold(&resv->lock)
446{
447 struct list_head allocated_regions;
448 int to_allocate = 0, i = 0;
449 struct file_region *trg = NULL, *rg = NULL;
450
451 VM_BUG_ON(regions_needed < 0);
452
453 INIT_LIST_HEAD(&allocated_regions);
454
455 /*
456 * Check for sufficient descriptors in the cache to accommodate
457 * the number of in progress add operations plus regions_needed.
458 *
459 * This is a while loop because when we drop the lock, some other call
460 * to region_add or region_del may have consumed some region_entries,
461 * so we keep looping here until we finally have enough entries for
462 * (adds_in_progress + regions_needed).
463 */
464 while (resv->region_cache_count <
465 (resv->adds_in_progress + regions_needed)) {
466 to_allocate = resv->adds_in_progress + regions_needed -
467 resv->region_cache_count;
468
469 /* At this point, we should have enough entries in the cache
470 * for all the existings adds_in_progress. We should only be
471 * needing to allocate for regions_needed.
472 */
473 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
474
475 spin_unlock(&resv->lock);
476 for (i = 0; i < to_allocate; i++) {
477 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
478 if (!trg)
479 goto out_of_memory;
480 list_add(&trg->link, &allocated_regions);
481 }
482
483 spin_lock(&resv->lock);
484
Wei Yangd3ec7b62020-10-13 16:56:27 -0700485 list_splice(&allocated_regions, &resv->region_cache);
486 resv->region_cache_count += to_allocate;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800487 }
488
Mina Almasry0db9d742020-04-01 21:11:25 -0700489 return 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800490
Mina Almasry0db9d742020-04-01 21:11:25 -0700491out_of_memory:
492 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
493 list_del(&rg->link);
494 kfree(rg);
495 }
496 return -ENOMEM;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800497}
498
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700499/*
500 * Add the huge page range represented by [f, t) to the reserve
Mina Almasry0db9d742020-04-01 21:11:25 -0700501 * map. Regions will be taken from the cache to fill in this range.
502 * Sufficient regions should exist in the cache due to the previous
503 * call to region_chg with the same range, but in some cases the cache will not
504 * have sufficient entries due to races with other code doing region_add or
505 * region_del. The extra needed entries will be allocated.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700506 *
Mina Almasry0db9d742020-04-01 21:11:25 -0700507 * regions_needed is the out value provided by a previous call to region_chg.
508 *
509 * Return the number of new huge pages added to the map. This number is greater
510 * than or equal to zero. If file_region entries needed to be allocated for
Ethon Paul7c8de352020-06-04 16:49:07 -0700511 * this operation and we were not able to allocate, it returns -ENOMEM.
Mina Almasry0db9d742020-04-01 21:11:25 -0700512 * region_add of regions of length 1 never allocate file_regions and cannot
513 * fail; region_chg will always allocate at least 1 entry and a region_add for
514 * 1 page will only require at most 1 entry.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700515 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700516static long region_add(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700517 long in_regions_needed, struct hstate *h,
518 struct hugetlb_cgroup *h_cg)
Andy Whitcroft96822902008-07-23 21:27:29 -0700519{
Mina Almasry0db9d742020-04-01 21:11:25 -0700520 long add = 0, actual_regions_needed = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700521
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700522 spin_lock(&resv->lock);
Mina Almasry0db9d742020-04-01 21:11:25 -0700523retry:
524
525 /* Count how many regions are actually needed to execute this add. */
Wei Yang972a3da32020-10-13 16:56:30 -0700526 add_reservation_in_range(resv, f, t, NULL, NULL,
527 &actual_regions_needed);
Andy Whitcroft96822902008-07-23 21:27:29 -0700528
Mike Kravetz5e911372015-09-08 15:01:28 -0700529 /*
Mina Almasry0db9d742020-04-01 21:11:25 -0700530 * Check for sufficient descriptors in the cache to accommodate
531 * this add operation. Note that actual_regions_needed may be greater
532 * than in_regions_needed, as the resv_map may have been modified since
533 * the region_chg call. In this case, we need to make sure that we
534 * allocate extra entries, such that we have enough for all the
535 * existing adds_in_progress, plus the excess needed for this
536 * operation.
Mike Kravetz5e911372015-09-08 15:01:28 -0700537 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700538 if (actual_regions_needed > in_regions_needed &&
539 resv->region_cache_count <
540 resv->adds_in_progress +
541 (actual_regions_needed - in_regions_needed)) {
542 /* region_add operation of range 1 should never need to
543 * allocate file_region entries.
544 */
545 VM_BUG_ON(t - f <= 1);
Mike Kravetz5e911372015-09-08 15:01:28 -0700546
Mina Almasry0db9d742020-04-01 21:11:25 -0700547 if (allocate_file_region_entries(
548 resv, actual_regions_needed - in_regions_needed)) {
549 return -ENOMEM;
550 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700551
Mina Almasry0db9d742020-04-01 21:11:25 -0700552 goto retry;
Mike Kravetz5e911372015-09-08 15:01:28 -0700553 }
554
Wei Yang972a3da32020-10-13 16:56:30 -0700555 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700556
Mina Almasry0db9d742020-04-01 21:11:25 -0700557 resv->adds_in_progress -= in_regions_needed;
558
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700559 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700560 VM_BUG_ON(add < 0);
561 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700562}
563
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700564/*
565 * Examine the existing reserve map and determine how many
566 * huge pages in the specified range [f, t) are NOT currently
567 * represented. This routine is called before a subsequent
568 * call to region_add that will actually modify the reserve
569 * map to add the specified range [f, t). region_chg does
570 * not change the number of huge pages represented by the
Mina Almasry0db9d742020-04-01 21:11:25 -0700571 * map. A number of new file_region structures is added to the cache as a
572 * placeholder, for the subsequent region_add call to use. At least 1
573 * file_region structure is added.
574 *
575 * out_regions_needed is the number of regions added to the
576 * resv->adds_in_progress. This value needs to be provided to a follow up call
577 * to region_add or region_abort for proper accounting.
Mike Kravetz5e911372015-09-08 15:01:28 -0700578 *
579 * Returns the number of huge pages that need to be added to the existing
580 * reservation map for the range [f, t). This number is greater or equal to
581 * zero. -ENOMEM is returned if a new file_region structure or cache entry
582 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700583 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700584static long region_chg(struct resv_map *resv, long f, long t,
585 long *out_regions_needed)
Andy Whitcroft96822902008-07-23 21:27:29 -0700586{
Andy Whitcroft96822902008-07-23 21:27:29 -0700587 long chg = 0;
588
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700589 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700590
Wei Yang972a3da32020-10-13 16:56:30 -0700591 /* Count how many hugepages in this range are NOT represented. */
Mina Almasry075a61d2020-04-01 21:11:28 -0700592 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
Wei Yang972a3da32020-10-13 16:56:30 -0700593 out_regions_needed);
Mike Kravetz5e911372015-09-08 15:01:28 -0700594
Mina Almasry0db9d742020-04-01 21:11:25 -0700595 if (*out_regions_needed == 0)
596 *out_regions_needed = 1;
Mike Kravetz5e911372015-09-08 15:01:28 -0700597
Mina Almasry0db9d742020-04-01 21:11:25 -0700598 if (allocate_file_region_entries(resv, *out_regions_needed))
599 return -ENOMEM;
Mike Kravetz5e911372015-09-08 15:01:28 -0700600
Mina Almasry0db9d742020-04-01 21:11:25 -0700601 resv->adds_in_progress += *out_regions_needed;
Andy Whitcroft96822902008-07-23 21:27:29 -0700602
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700603 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700604 return chg;
605}
606
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700607/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700608 * Abort the in progress add operation. The adds_in_progress field
609 * of the resv_map keeps track of the operations in progress between
610 * calls to region_chg and region_add. Operations are sometimes
611 * aborted after the call to region_chg. In such cases, region_abort
Mina Almasry0db9d742020-04-01 21:11:25 -0700612 * is called to decrement the adds_in_progress counter. regions_needed
613 * is the value returned by the region_chg call, it is used to decrement
614 * the adds_in_progress counter.
Mike Kravetz5e911372015-09-08 15:01:28 -0700615 *
616 * NOTE: The range arguments [f, t) are not needed or used in this
617 * routine. They are kept to make reading the calling code easier as
618 * arguments will match the associated region_chg call.
619 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700620static void region_abort(struct resv_map *resv, long f, long t,
621 long regions_needed)
Mike Kravetz5e911372015-09-08 15:01:28 -0700622{
623 spin_lock(&resv->lock);
624 VM_BUG_ON(!resv->region_cache_count);
Mina Almasry0db9d742020-04-01 21:11:25 -0700625 resv->adds_in_progress -= regions_needed;
Mike Kravetz5e911372015-09-08 15:01:28 -0700626 spin_unlock(&resv->lock);
627}
628
629/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700630 * Delete the specified range [f, t) from the reserve map. If the
631 * t parameter is LONG_MAX, this indicates that ALL regions after f
632 * should be deleted. Locate the regions which intersect [f, t)
633 * and either trim, delete or split the existing regions.
634 *
635 * Returns the number of huge pages deleted from the reserve map.
636 * In the normal case, the return value is zero or more. In the
637 * case where a region must be split, a new region descriptor must
638 * be allocated. If the allocation fails, -ENOMEM will be returned.
639 * NOTE: If the parameter t == LONG_MAX, then we will never split
640 * a region and possibly return -ENOMEM. Callers specifying
641 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700642 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700643static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700644{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700645 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700646 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700647 struct file_region *nrg = NULL;
648 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700649
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700650retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700651 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700652 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800653 /*
654 * Skip regions before the range to be deleted. file_region
655 * ranges are normally of the form [from, to). However, there
656 * may be a "placeholder" entry in the map which is of the form
657 * (from, to) with from == to. Check for placeholder entries
658 * at the beginning of the range to be deleted.
659 */
660 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700661 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800662
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700663 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700664 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700665
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700666 if (f > rg->from && t < rg->to) { /* Must split region */
667 /*
668 * Check for an entry in the cache before dropping
669 * lock and attempting allocation.
670 */
671 if (!nrg &&
672 resv->region_cache_count > resv->adds_in_progress) {
673 nrg = list_first_entry(&resv->region_cache,
674 struct file_region,
675 link);
676 list_del(&nrg->link);
677 resv->region_cache_count--;
678 }
679
680 if (!nrg) {
681 spin_unlock(&resv->lock);
682 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
683 if (!nrg)
684 return -ENOMEM;
685 goto retry;
686 }
687
688 del += t - f;
Mike Kravetz79aa9252020-11-01 17:07:27 -0800689 hugetlb_cgroup_uncharge_file_region(
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700690 resv, rg, t - f, false);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700691
692 /* New entry for end of split region */
693 nrg->from = t;
694 nrg->to = rg->to;
Mina Almasry075a61d2020-04-01 21:11:28 -0700695
696 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
697
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700698 INIT_LIST_HEAD(&nrg->link);
699
700 /* Original entry is trimmed */
701 rg->to = f;
702
703 list_add(&nrg->link, &rg->link);
704 nrg = NULL;
705 break;
706 }
707
708 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
709 del += rg->to - rg->from;
Mina Almasry075a61d2020-04-01 21:11:28 -0700710 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700711 rg->to - rg->from, true);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700712 list_del(&rg->link);
713 kfree(rg);
714 continue;
715 }
716
717 if (f <= rg->from) { /* Trim beginning of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700718 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700719 t - rg->from, false);
Mina Almasry075a61d2020-04-01 21:11:28 -0700720
Mike Kravetz79aa9252020-11-01 17:07:27 -0800721 del += t - rg->from;
722 rg->from = t;
723 } else { /* Trim end of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700724 hugetlb_cgroup_uncharge_file_region(resv, rg,
Miaohe Linfe03ccc2021-03-24 21:37:17 -0700725 rg->to - f, false);
Mike Kravetz79aa9252020-11-01 17:07:27 -0800726
727 del += rg->to - f;
728 rg->to = f;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700729 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700730 }
731
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700732 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700733 kfree(nrg);
734 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700735}
736
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700737/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700738 * A rare out of memory error was encountered which prevented removal of
739 * the reserve map region for a page. The huge page itself was free'ed
740 * and removed from the page cache. This routine will adjust the subpool
741 * usage count, and the global reserve count if needed. By incrementing
742 * these counts, the reserve map entry which could not be deleted will
743 * appear as a "reserved" entry instead of simply dangling with incorrect
744 * counts.
745 */
zhong jiang72e29362016-10-07 17:02:01 -0700746void hugetlb_fix_reserve_counts(struct inode *inode)
Mike Kravetzb5cec282015-09-08 15:01:41 -0700747{
748 struct hugepage_subpool *spool = subpool_inode(inode);
749 long rsv_adjust;
Miaohe Lin9639a752021-05-04 18:34:38 -0700750 bool reserved = false;
Mike Kravetzb5cec282015-09-08 15:01:41 -0700751
752 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
Miaohe Lin9639a752021-05-04 18:34:38 -0700753 if (rsv_adjust > 0) {
Mike Kravetzb5cec282015-09-08 15:01:41 -0700754 struct hstate *h = hstate_inode(inode);
755
Miaohe Lin9639a752021-05-04 18:34:38 -0700756 if (!hugetlb_acct_memory(h, 1))
757 reserved = true;
758 } else if (!rsv_adjust) {
759 reserved = true;
Mike Kravetzb5cec282015-09-08 15:01:41 -0700760 }
Miaohe Lin9639a752021-05-04 18:34:38 -0700761
762 if (!reserved)
763 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
Mike Kravetzb5cec282015-09-08 15:01:41 -0700764}
765
766/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700767 * Count and return the number of huge pages in the reserve map
768 * that intersect with the range [f, t).
769 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700770static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700771{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700772 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700773 struct file_region *rg;
774 long chg = 0;
775
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700776 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700777 /* Locate each segment we overlap with, and count that overlap. */
778 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700779 long seg_from;
780 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700781
782 if (rg->to <= f)
783 continue;
784 if (rg->from >= t)
785 break;
786
787 seg_from = max(rg->from, f);
788 seg_to = min(rg->to, t);
789
790 chg += seg_to - seg_from;
791 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700792 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700793
794 return chg;
795}
796
Andy Whitcroft96822902008-07-23 21:27:29 -0700797/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700798 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700799 * the mapping, in pagecache page units; huge pages here.
800 */
Andi Kleena5516432008-07-23 21:27:41 -0700801static pgoff_t vma_hugecache_offset(struct hstate *h,
802 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700803{
Andi Kleena5516432008-07-23 21:27:41 -0700804 return ((address - vma->vm_start) >> huge_page_shift(h)) +
805 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700806}
807
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900808pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
809 unsigned long address)
810{
811 return vma_hugecache_offset(hstate_vma(vma), vma, address);
812}
Dan Williamsdee41072016-05-14 12:20:44 -0700813EXPORT_SYMBOL_GPL(linear_hugepage_index);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900814
Andy Whitcroft84afd992008-07-23 21:27:32 -0700815/*
Mel Gorman08fba692009-01-06 14:38:53 -0800816 * Return the size of the pages allocated when backing a VMA. In the majority
817 * cases this will be same size as used by the page table entries.
818 */
819unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
820{
Dan Williams05ea8862018-04-05 16:24:25 -0700821 if (vma->vm_ops && vma->vm_ops->pagesize)
822 return vma->vm_ops->pagesize(vma);
823 return PAGE_SIZE;
Mel Gorman08fba692009-01-06 14:38:53 -0800824}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200825EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800826
827/*
Mel Gorman33402892009-01-06 14:38:54 -0800828 * Return the page size being used by the MMU to back a VMA. In the majority
829 * of cases, the page size used by the kernel matches the MMU size. On
Dan Williams09135cc2018-04-05 16:24:21 -0700830 * architectures where it differs, an architecture-specific 'strong'
831 * version of this symbol is required.
Mel Gorman33402892009-01-06 14:38:54 -0800832 */
Dan Williams09135cc2018-04-05 16:24:21 -0700833__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
Mel Gorman33402892009-01-06 14:38:54 -0800834{
835 return vma_kernel_pagesize(vma);
836}
Mel Gorman33402892009-01-06 14:38:54 -0800837
838/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700839 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
840 * bits of the reservation map pointer, which are always clear due to
841 * alignment.
842 */
843#define HPAGE_RESV_OWNER (1UL << 0)
844#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700845#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700846
Mel Gormana1e78772008-07-23 21:27:23 -0700847/*
848 * These helpers are used to track how many pages are reserved for
849 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
850 * is guaranteed to have their future faults succeed.
851 *
852 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
853 * the reserve counters are updated with the hugetlb_lock held. It is safe
854 * to reset the VMA at fork() time as it is not in use yet and there is no
855 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700856 *
857 * The private mapping reservation is represented in a subtly different
858 * manner to a shared mapping. A shared mapping has a region map associated
859 * with the underlying file, this region map represents the backing file
860 * pages which have ever had a reservation assigned which this persists even
861 * after the page is instantiated. A private mapping has a region map
862 * associated with the original mmap which is attached to all VMAs which
863 * reference it, this region map represents those offsets which have consumed
864 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700865 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700866static unsigned long get_vma_private_data(struct vm_area_struct *vma)
867{
868 return (unsigned long)vma->vm_private_data;
869}
870
871static void set_vma_private_data(struct vm_area_struct *vma,
872 unsigned long value)
873{
874 vma->vm_private_data = (void *)value;
875}
876
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700877static void
878resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
879 struct hugetlb_cgroup *h_cg,
880 struct hstate *h)
881{
882#ifdef CONFIG_CGROUP_HUGETLB
883 if (!h_cg || !h) {
884 resv_map->reservation_counter = NULL;
885 resv_map->pages_per_hpage = 0;
886 resv_map->css = NULL;
887 } else {
888 resv_map->reservation_counter =
889 &h_cg->rsvd_hugepage[hstate_index(h)];
890 resv_map->pages_per_hpage = pages_per_huge_page(h);
891 resv_map->css = &h_cg->css;
892 }
893#endif
894}
895
Joonsoo Kim9119a412014-04-03 14:47:25 -0700896struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700897{
898 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700899 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
900
901 if (!resv_map || !rg) {
902 kfree(resv_map);
903 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700904 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700905 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700906
907 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700908 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700909 INIT_LIST_HEAD(&resv_map->regions);
910
Mike Kravetz5e911372015-09-08 15:01:28 -0700911 resv_map->adds_in_progress = 0;
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700912 /*
913 * Initialize these to 0. On shared mappings, 0's here indicate these
914 * fields don't do cgroup accounting. On private mappings, these will be
915 * re-initialized to the proper values, to indicate that hugetlb cgroup
916 * reservations are to be un-charged from here.
917 */
918 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700919
920 INIT_LIST_HEAD(&resv_map->region_cache);
921 list_add(&rg->link, &resv_map->region_cache);
922 resv_map->region_cache_count = 1;
923
Andy Whitcroft84afd992008-07-23 21:27:32 -0700924 return resv_map;
925}
926
Joonsoo Kim9119a412014-04-03 14:47:25 -0700927void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700928{
929 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700930 struct list_head *head = &resv_map->region_cache;
931 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700932
933 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700934 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700935
936 /* ... and any entries left in the cache */
937 list_for_each_entry_safe(rg, trg, head, link) {
938 list_del(&rg->link);
939 kfree(rg);
940 }
941
942 VM_BUG_ON(resv_map->adds_in_progress);
943
Andy Whitcroft84afd992008-07-23 21:27:32 -0700944 kfree(resv_map);
945}
946
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700947static inline struct resv_map *inode_resv_map(struct inode *inode)
948{
Mike Kravetzf27a5132019-05-13 17:22:55 -0700949 /*
950 * At inode evict time, i_mapping may not point to the original
951 * address space within the inode. This original address space
952 * contains the pointer to the resv_map. So, always use the
953 * address space embedded within the inode.
954 * The VERY common case is inode->mapping == &inode->i_data but,
955 * this may not be true for device special inodes.
956 */
957 return (struct resv_map *)(&inode->i_data)->private_data;
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700958}
959
Andy Whitcroft84afd992008-07-23 21:27:32 -0700960static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700961{
Sasha Levin81d1b092014-10-09 15:28:10 -0700962 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700963 if (vma->vm_flags & VM_MAYSHARE) {
964 struct address_space *mapping = vma->vm_file->f_mapping;
965 struct inode *inode = mapping->host;
966
967 return inode_resv_map(inode);
968
969 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700970 return (struct resv_map *)(get_vma_private_data(vma) &
971 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700972 }
Mel Gormana1e78772008-07-23 21:27:23 -0700973}
974
Andy Whitcroft84afd992008-07-23 21:27:32 -0700975static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700976{
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);
Mel Gormana1e78772008-07-23 21:27:23 -0700979
Andy Whitcroft84afd992008-07-23 21:27:32 -0700980 set_vma_private_data(vma, (get_vma_private_data(vma) &
981 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700982}
983
984static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
985{
Sasha Levin81d1b092014-10-09 15:28:10 -0700986 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
987 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700988
989 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700990}
991
992static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
993{
Sasha Levin81d1b092014-10-09 15:28:10 -0700994 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700995
996 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700997}
998
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700999/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -07001000void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
1001{
Sasha Levin81d1b092014-10-09 15:28:10 -07001002 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -07001003 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -07001004 vma->vm_private_data = (void *)0;
1005}
1006
1007/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001008static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -07001009{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001010 if (vma->vm_flags & VM_NORESERVE) {
1011 /*
1012 * This address is already reserved by other process(chg == 0),
1013 * so, we should decrement reserved count. Without decrementing,
1014 * reserve count remains after releasing inode, because this
1015 * allocated page will go into page cache and is regarded as
1016 * coming from reserved pool in releasing step. Currently, we
1017 * don't have any other solution to deal with this situation
1018 * properly, so add work-around here.
1019 */
1020 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001021 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001022 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001023 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001024 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001025
1026 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -07001027 if (vma->vm_flags & VM_MAYSHARE) {
1028 /*
1029 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1030 * be a region map for all pages. The only situation where
1031 * there is no region map is if a hole was punched via
Ethon Paul7c8de352020-06-04 16:49:07 -07001032 * fallocate. In this case, there really are no reserves to
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -07001033 * use. This situation is indicated if chg != 0.
1034 */
1035 if (chg)
1036 return false;
1037 else
1038 return true;
1039 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001040
1041 /*
1042 * Only the process that called mmap() has reserves for
1043 * private mappings.
1044 */
Mike Kravetz67961f92016-06-08 15:33:42 -07001045 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1046 /*
1047 * Like the shared case above, a hole punch or truncate
1048 * could have been performed on the private mapping.
1049 * Examine the value of chg to determine if reserves
1050 * actually exist or were previously consumed.
1051 * Very Subtle - The value of chg comes from a previous
1052 * call to vma_needs_reserves(). The reserve map for
1053 * private mappings has different (opposite) semantics
1054 * than that of shared mappings. vma_needs_reserves()
1055 * has already taken this difference in semantics into
1056 * account. Therefore, the meaning of chg is the same
1057 * as in the shared case above. Code could easily be
1058 * combined, but keeping it separate draws attention to
1059 * subtle differences.
1060 */
1061 if (chg)
1062 return false;
1063 else
1064 return true;
1065 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001066
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001067 return false;
Mel Gormana1e78772008-07-23 21:27:23 -07001068}
1069
Andi Kleena5516432008-07-23 21:27:41 -07001070static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071{
1072 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001073 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -07001074 h->free_huge_pages++;
1075 h->free_huge_pages_node[nid]++;
Muchun Songe334b1f2021-02-04 18:32:06 -08001076 SetPageHugeFreed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077}
1078
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001079static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001080{
1081 struct page *page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001082 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001083
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001084 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1085 if (nocma && is_migrate_cma_page(page))
1086 continue;
1087
Wei Yang6664bfc2020-10-13 16:56:39 -07001088 if (PageHWPoison(page))
1089 continue;
1090
1091 list_move(&page->lru, &h->hugepage_activelist);
1092 set_page_refcounted(page);
Muchun Songe334b1f2021-02-04 18:32:06 -08001093 ClearPageHugeFreed(page);
Wei Yang6664bfc2020-10-13 16:56:39 -07001094 h->free_huge_pages--;
1095 h->free_huge_pages_node[nid]--;
1096 return page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001097 }
1098
Wei Yang6664bfc2020-10-13 16:56:39 -07001099 return NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001100}
1101
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001102static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1103 nodemask_t *nmask)
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001104{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001105 unsigned int cpuset_mems_cookie;
1106 struct zonelist *zonelist;
1107 struct zone *zone;
1108 struct zoneref *z;
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001109 int node = NUMA_NO_NODE;
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001110
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001111 zonelist = node_zonelist(nid, gfp_mask);
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001112
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001113retry_cpuset:
1114 cpuset_mems_cookie = read_mems_allowed_begin();
1115 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1116 struct page *page;
1117
1118 if (!cpuset_zone_allowed(zone, gfp_mask))
1119 continue;
1120 /*
1121 * no need to ask again on the same node. Pool is node rather than
1122 * zone aware
1123 */
1124 if (zone_to_nid(zone) == node)
1125 continue;
1126 node = zone_to_nid(zone);
1127
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001128 page = dequeue_huge_page_node_exact(h, node);
1129 if (page)
1130 return page;
1131 }
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001132 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1133 goto retry_cpuset;
1134
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001135 return NULL;
1136}
1137
Andi Kleena5516432008-07-23 21:27:41 -07001138static struct page *dequeue_huge_page_vma(struct hstate *h,
1139 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001140 unsigned long address, int avoid_reserve,
1141 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001143 struct page *page;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001144 struct mempolicy *mpol;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001145 gfp_t gfp_mask;
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001146 nodemask_t *nodemask;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001147 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148
Mel Gormana1e78772008-07-23 21:27:23 -07001149 /*
1150 * A child process with MAP_PRIVATE mappings created by their parent
1151 * have no page reserves. This check ensures that reservations are
1152 * not "stolen". The child may still get SIGKILLed
1153 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001154 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -07001155 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -07001156 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -07001157
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001158 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -07001159 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -07001160 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001161
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001162 gfp_mask = htlb_alloc_mask(h);
1163 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001164 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1165 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1166 SetPagePrivate(page);
1167 h->resv_huge_pages--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07001169
1170 mpol_cond_put(mpol);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001171 return page;
1172
Miao Xiec0ff7452010-05-24 14:32:08 -07001173err:
Mel Gormancc9a6c82012-03-21 16:34:11 -07001174 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175}
1176
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001177/*
1178 * common helper functions for hstate_next_node_to_{alloc|free}.
1179 * We may have allocated or freed a huge page based on a different
1180 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1181 * be outside of *nodes_allowed. Ensure that we use an allowed
1182 * node for alloc or free.
1183 */
1184static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1185{
Andrew Morton0edaf862016-05-19 17:10:58 -07001186 nid = next_node_in(nid, *nodes_allowed);
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001187 VM_BUG_ON(nid >= MAX_NUMNODES);
1188
1189 return nid;
1190}
1191
1192static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1193{
1194 if (!node_isset(nid, *nodes_allowed))
1195 nid = next_node_allowed(nid, nodes_allowed);
1196 return nid;
1197}
1198
1199/*
1200 * returns the previously saved node ["this node"] from which to
1201 * allocate a persistent huge page for the pool and advance the
1202 * next node from which to allocate, handling wrap at end of node
1203 * mask.
1204 */
1205static int hstate_next_node_to_alloc(struct hstate *h,
1206 nodemask_t *nodes_allowed)
1207{
1208 int nid;
1209
1210 VM_BUG_ON(!nodes_allowed);
1211
1212 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1213 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1214
1215 return nid;
1216}
1217
1218/*
1219 * helper for free_pool_huge_page() - return the previously saved
1220 * node ["this node"] from which to free a huge page. Advance the
1221 * next node id whether or not we find a free huge page to free so
1222 * that the next attempt to free addresses the next node.
1223 */
1224static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1225{
1226 int nid;
1227
1228 VM_BUG_ON(!nodes_allowed);
1229
1230 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1231 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1232
1233 return nid;
1234}
1235
1236#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1237 for (nr_nodes = nodes_weight(*mask); \
1238 nr_nodes > 0 && \
1239 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1240 nr_nodes--)
1241
1242#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1243 for (nr_nodes = nodes_weight(*mask); \
1244 nr_nodes > 0 && \
1245 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1246 nr_nodes--)
1247
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001248#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001249static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001250 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001251{
1252 int i;
1253 int nr_pages = 1 << order;
1254 struct page *p = page + 1;
1255
Gerald Schaeferc8cc7082016-06-24 14:50:04 -07001256 atomic_set(compound_mapcount_ptr(page), 0);
Yanfei Xu2e16ad52021-02-24 12:07:22 -08001257 atomic_set(compound_pincount_ptr(page), 0);
John Hubbard47e29d32020-04-01 21:05:33 -07001258
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001259 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001260 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001261 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001262 }
1263
1264 set_compound_order(page, 0);
Gerald Schaeferba9c1202020-12-11 13:36:53 -08001265 page[1].compound_nr = 0;
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001266 __ClearPageHead(page);
1267}
1268
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001269static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001270{
Roman Gushchincf11e852020-04-10 14:32:45 -07001271 /*
1272 * If the page isn't allocated using the cma allocator,
1273 * cma_release() returns false.
1274 */
Barry Songdbda8fe2020-07-23 21:15:30 -07001275#ifdef CONFIG_CMA
1276 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
Roman Gushchincf11e852020-04-10 14:32:45 -07001277 return;
Barry Songdbda8fe2020-07-23 21:15:30 -07001278#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001279
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001280 free_contig_range(page_to_pfn(page), 1 << order);
1281}
1282
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001283#ifdef CONFIG_CONTIG_ALLOC
Michal Hockod9cc948f2018-01-31 16:20:44 -08001284static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1285 int nid, nodemask_t *nodemask)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001286{
Anshuman Khandual5e27a2d2019-11-30 17:55:06 -08001287 unsigned long nr_pages = 1UL << huge_page_order(h);
Li Xinhai953f0642020-09-04 16:36:10 -07001288 if (nid == NUMA_NO_NODE)
1289 nid = numa_mem_id();
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001290
Barry Songdbda8fe2020-07-23 21:15:30 -07001291#ifdef CONFIG_CMA
1292 {
Roman Gushchincf11e852020-04-10 14:32:45 -07001293 struct page *page;
1294 int node;
1295
Li Xinhai953f0642020-09-04 16:36:10 -07001296 if (hugetlb_cma[nid]) {
1297 page = cma_alloc(hugetlb_cma[nid], nr_pages,
Minchan Kim23ba9902021-01-21 12:09:34 -08001298 huge_page_order(h),
1299 GFP_KERNEL | __GFP_NOWARN);
Roman Gushchincf11e852020-04-10 14:32:45 -07001300 if (page)
1301 return page;
1302 }
Li Xinhai953f0642020-09-04 16:36:10 -07001303
1304 if (!(gfp_mask & __GFP_THISNODE)) {
1305 for_each_node_mask(node, *nodemask) {
1306 if (node == nid || !hugetlb_cma[node])
1307 continue;
1308
1309 page = cma_alloc(hugetlb_cma[node], nr_pages,
Minchan Kim23ba9902021-01-21 12:09:34 -08001310 huge_page_order(h),
1311 GFP_KERNEL | __GFP_NOWARN);
Li Xinhai953f0642020-09-04 16:36:10 -07001312 if (page)
1313 return page;
1314 }
1315 }
Roman Gushchincf11e852020-04-10 14:32:45 -07001316 }
Barry Songdbda8fe2020-07-23 21:15:30 -07001317#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001318
Anshuman Khandual5e27a2d2019-11-30 17:55:06 -08001319 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001320}
1321
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001322#else /* !CONFIG_CONTIG_ALLOC */
1323static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1324 int nid, nodemask_t *nodemask)
1325{
1326 return NULL;
1327}
1328#endif /* CONFIG_CONTIG_ALLOC */
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001329
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001330#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
Michal Hockod9cc948f2018-01-31 16:20:44 -08001331static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001332 int nid, nodemask_t *nodemask)
1333{
1334 return NULL;
1335}
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001336static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001337static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001338 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001339#endif
1340
Andi Kleena5516432008-07-23 21:27:41 -07001341static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001342{
1343 int i;
Mike Kravetz65f6dc32021-02-24 12:07:50 -08001344 struct page *subpage = page;
Andi Kleena5516432008-07-23 21:27:41 -07001345
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001346 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001347 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001348
Andi Kleena5516432008-07-23 21:27:41 -07001349 h->nr_huge_pages--;
1350 h->nr_huge_pages_node[page_to_nid(page)]--;
Mike Kravetz65f6dc32021-02-24 12:07:50 -08001351 for (i = 0; i < pages_per_huge_page(h);
1352 i++, subpage = mem_map_next(subpage, page, i)) {
1353 subpage->flags &= ~(1 << PG_locked | 1 << PG_error |
Chris Forbes32f84522011-07-25 17:12:14 -07001354 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001355 1 << PG_active | 1 << PG_private |
1356 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001357 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001358 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Mina Almasry1adc4d42020-04-01 21:11:15 -07001359 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001360 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001361 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001362 if (hstate_is_gigantic(h)) {
Roman Gushchincf11e852020-04-10 14:32:45 -07001363 /*
1364 * Temporarily drop the hugetlb_lock, because
1365 * we might block in free_gigantic_page().
1366 */
1367 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001368 destroy_compound_gigantic_page(page, huge_page_order(h));
1369 free_gigantic_page(page, huge_page_order(h));
Roman Gushchincf11e852020-04-10 14:32:45 -07001370 spin_lock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001371 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001372 __free_pages(page, huge_page_order(h));
1373 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001374}
1375
Andi Kleene5ff2152008-07-23 21:27:42 -07001376struct hstate *size_to_hstate(unsigned long size)
1377{
1378 struct hstate *h;
1379
1380 for_each_hstate(h) {
1381 if (huge_page_size(h) == size)
1382 return h;
1383 }
1384 return NULL;
1385}
1386
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001387/*
1388 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1389 * to hstate->hugepage_activelist.)
1390 *
1391 * This function can be called for tail pages, but never returns true for them.
1392 */
1393bool page_huge_active(struct page *page)
1394{
Muchun Songeca84eb2021-02-04 18:32:13 -08001395 return PageHeadHuge(page) && PagePrivate(&page[1]);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001396}
1397
1398/* never called for tail page */
Muchun Songafe6c312021-02-04 18:32:03 -08001399void set_page_huge_active(struct page *page)
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001400{
1401 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1402 SetPagePrivate(&page[1]);
1403}
1404
1405static void clear_page_huge_active(struct page *page)
1406{
1407 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1408 ClearPagePrivate(&page[1]);
1409}
1410
Michal Hockoab5ac902018-01-31 16:20:48 -08001411/*
1412 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1413 * code
1414 */
1415static inline bool PageHugeTemporary(struct page *page)
1416{
1417 if (!PageHuge(page))
1418 return false;
1419
1420 return (unsigned long)page[2].mapping == -1U;
1421}
1422
1423static inline void SetPageHugeTemporary(struct page *page)
1424{
1425 page[2].mapping = (void *)-1U;
1426}
1427
1428static inline void ClearPageHugeTemporary(struct page *page)
1429{
1430 page[2].mapping = NULL;
1431}
1432
Waiman Longc77c0a82020-01-04 13:00:15 -08001433static void __free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001434{
Andi Kleena5516432008-07-23 21:27:41 -07001435 /*
1436 * Can't pass hstate in here because it is called from the
1437 * compound page destructor.
1438 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001439 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001440 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001441 struct hugepage_subpool *spool =
1442 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001443 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001444
Mike Kravetzb4330af2016-02-05 15:36:38 -08001445 VM_BUG_ON_PAGE(page_count(page), page);
1446 VM_BUG_ON_PAGE(page_mapcount(page), page);
Yongkai Wu8ace22b2018-12-14 14:17:10 -08001447
1448 set_page_private(page, 0);
1449 page->mapping = NULL;
Joonsoo Kim07443a82013-09-11 14:21:58 -07001450 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001451 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001452
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001453 /*
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001454 * If PagePrivate() was set on page, page allocation consumed a
1455 * reservation. If the page was associated with a subpool, there
1456 * would have been a page reserved in the subpool before allocation
1457 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1458 * reservtion, do not call hugepage_subpool_put_pages() as this will
1459 * remove the reserved page from the subpool.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001460 */
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001461 if (!restore_reserve) {
1462 /*
1463 * A return code of zero implies that the subpool will be
1464 * under its minimum size if the reservation is not restored
1465 * after page is free. Therefore, force restore_reserve
1466 * operation.
1467 */
1468 if (hugepage_subpool_put_pages(spool, 1) == 0)
1469 restore_reserve = true;
1470 }
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001471
David Gibson27a85ef2006-03-22 00:08:56 -08001472 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001473 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001474 hugetlb_cgroup_uncharge_page(hstate_index(h),
1475 pages_per_huge_page(h), page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07001476 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1477 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001478 if (restore_reserve)
1479 h->resv_huge_pages++;
1480
Michal Hockoab5ac902018-01-31 16:20:48 -08001481 if (PageHugeTemporary(page)) {
1482 list_del(&page->lru);
1483 ClearPageHugeTemporary(page);
1484 update_and_free_page(h, page);
1485 } else if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001486 /* remove the page from active list */
1487 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001488 update_and_free_page(h, page);
1489 h->surplus_huge_pages--;
1490 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001491 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001492 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001493 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001494 }
David Gibson27a85ef2006-03-22 00:08:56 -08001495 spin_unlock(&hugetlb_lock);
1496}
1497
Waiman Longc77c0a82020-01-04 13:00:15 -08001498/*
1499 * As free_huge_page() can be called from a non-task context, we have
1500 * to defer the actual freeing in a workqueue to prevent potential
1501 * hugetlb_lock deadlock.
1502 *
1503 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1504 * be freed and frees them one-by-one. As the page->mapping pointer is
1505 * going to be cleared in __free_huge_page() anyway, it is reused as the
1506 * llist_node structure of a lockless linked list of huge pages to be freed.
1507 */
1508static LLIST_HEAD(hpage_freelist);
1509
1510static void free_hpage_workfn(struct work_struct *work)
1511{
1512 struct llist_node *node;
1513 struct page *page;
1514
1515 node = llist_del_all(&hpage_freelist);
1516
1517 while (node) {
1518 page = container_of((struct address_space **)node,
1519 struct page, mapping);
1520 node = node->next;
1521 __free_huge_page(page);
1522 }
1523}
1524static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1525
1526void free_huge_page(struct page *page)
1527{
1528 /*
1529 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1530 */
1531 if (!in_task()) {
1532 /*
1533 * Only call schedule_work() if hpage_freelist is previously
1534 * empty. Otherwise, schedule_work() had been called but the
1535 * workfn hasn't retrieved the list yet.
1536 */
1537 if (llist_add((struct llist_node *)&page->mapping,
1538 &hpage_freelist))
1539 schedule_work(&free_hpage_work);
1540 return;
1541 }
1542
1543 __free_huge_page(page);
1544}
1545
Andi Kleena5516432008-07-23 21:27:41 -07001546static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001547{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001548 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001549 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001550 set_hugetlb_cgroup(page, NULL);
Mina Almasry1adc4d42020-04-01 21:11:15 -07001551 set_hugetlb_cgroup_rsvd(page, NULL);
Wei Yang2f375112020-10-13 16:56:36 -07001552 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001553 h->nr_huge_pages++;
1554 h->nr_huge_pages_node[nid]++;
Muchun Songe334b1f2021-02-04 18:32:06 -08001555 ClearPageHugeFreed(page);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001556 spin_unlock(&hugetlb_lock);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001557}
1558
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001559static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001560{
1561 int i;
1562 int nr_pages = 1 << order;
1563 struct page *p = page + 1;
1564
1565 /* we rely on prep_new_huge_page to set the destructor */
1566 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001567 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001568 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001569 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001570 /*
1571 * For gigantic hugepages allocated through bootmem at
1572 * boot, it's safer to be consistent with the not-gigantic
1573 * hugepages and clear the PG_reserved bit from all tail pages
Ethon Paul7c8de352020-06-04 16:49:07 -07001574 * too. Otherwise drivers using get_user_pages() to access tail
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001575 * pages may get the reference counting wrong if they see
1576 * PG_reserved set on a tail page (despite the head page not
1577 * having PG_reserved set). Enforcing this consistency between
1578 * head and tail pages allows drivers to optimize away a check
1579 * on the head page when they need know if put_page() is needed
1580 * after get_user_pages().
1581 */
1582 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001583 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001584 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001585 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001586 atomic_set(compound_mapcount_ptr(page), -1);
Yanfei Xu2e16ad52021-02-24 12:07:22 -08001587 atomic_set(compound_pincount_ptr(page), 0);
Wu Fengguang20a03072009-06-16 15:32:22 -07001588}
1589
Andrew Morton77959122012-10-08 16:34:11 -07001590/*
1591 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1592 * transparent huge pages. See the PageTransHuge() documentation for more
1593 * details.
1594 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001595int PageHuge(struct page *page)
1596{
Wu Fengguang20a03072009-06-16 15:32:22 -07001597 if (!PageCompound(page))
1598 return 0;
1599
1600 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001601 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001602}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001603EXPORT_SYMBOL_GPL(PageHuge);
1604
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001605/*
1606 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1607 * normal or transparent huge pages.
1608 */
1609int PageHeadHuge(struct page *page_head)
1610{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001611 if (!PageHead(page_head))
1612 return 0;
1613
Vlastimil Babkad4af73e2020-04-01 21:11:48 -07001614 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001615}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001616
Mike Kravetzc0d03812020-04-01 21:11:05 -07001617/*
Mike Kravetzc0d03812020-04-01 21:11:05 -07001618 * Find and lock address space (mapping) in write mode.
1619 *
Mike Kravetz336bf302020-11-13 22:52:16 -08001620 * Upon entry, the page is locked which means that page_mapping() is
1621 * stable. Due to locking order, we can only trylock_write. If we can
1622 * not get the lock, simply return NULL to caller.
Mike Kravetzc0d03812020-04-01 21:11:05 -07001623 */
1624struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1625{
Mike Kravetz336bf302020-11-13 22:52:16 -08001626 struct address_space *mapping = page_mapping(hpage);
Mike Kravetzc0d03812020-04-01 21:11:05 -07001627
Mike Kravetzc0d03812020-04-01 21:11:05 -07001628 if (!mapping)
1629 return mapping;
1630
Mike Kravetzc0d03812020-04-01 21:11:05 -07001631 if (i_mmap_trylock_write(mapping))
1632 return mapping;
1633
Mike Kravetz336bf302020-11-13 22:52:16 -08001634 return NULL;
Mike Kravetzc0d03812020-04-01 21:11:05 -07001635}
1636
Hugh Dickins377a7962021-06-24 18:39:52 -07001637pgoff_t hugetlb_basepage_index(struct page *page)
Zhang Yi13d60f42013-06-25 21:19:31 +08001638{
1639 struct page *page_head = compound_head(page);
1640 pgoff_t index = page_index(page_head);
1641 unsigned long compound_idx;
1642
Zhang Yi13d60f42013-06-25 21:19:31 +08001643 if (compound_order(page_head) >= MAX_ORDER)
1644 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1645 else
1646 compound_idx = page - page_head;
1647
1648 return (index << compound_order(page_head)) + compound_idx;
1649}
1650
Michal Hocko0c397da2018-01-31 16:20:56 -08001651static struct page *alloc_buddy_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001652 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1653 nodemask_t *node_alloc_noretry)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654{
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001655 int order = huge_page_order(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 struct page *page;
Mike Kravetzf60858f2019-09-23 15:37:35 -07001657 bool alloc_try_hard = true;
Joe Jinf96efd52007-07-15 23:38:12 -07001658
Mike Kravetzf60858f2019-09-23 15:37:35 -07001659 /*
1660 * By default we always try hard to allocate the page with
1661 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1662 * a loop (to adjust global huge page counts) and previous allocation
1663 * failed, do not continue to try hard on the same node. Use the
1664 * node_alloc_noretry bitmap to manage this state information.
1665 */
1666 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1667 alloc_try_hard = false;
1668 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1669 if (alloc_try_hard)
1670 gfp_mask |= __GFP_RETRY_MAYFAIL;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001671 if (nid == NUMA_NO_NODE)
1672 nid = numa_mem_id();
1673 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1674 if (page)
1675 __count_vm_event(HTLB_BUDDY_PGALLOC);
1676 else
1677 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001678
Mike Kravetzf60858f2019-09-23 15:37:35 -07001679 /*
1680 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1681 * indicates an overall state change. Clear bit so that we resume
1682 * normal 'try hard' allocations.
1683 */
1684 if (node_alloc_noretry && page && !alloc_try_hard)
1685 node_clear(nid, *node_alloc_noretry);
1686
1687 /*
1688 * If we tried hard to get a page but failed, set bit so that
1689 * subsequent attempts will not try as hard until there is an
1690 * overall state change.
1691 */
1692 if (node_alloc_noretry && !page && alloc_try_hard)
1693 node_set(nid, *node_alloc_noretry);
1694
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001695 return page;
1696}
1697
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001698/*
Michal Hocko0c397da2018-01-31 16:20:56 -08001699 * Common helper to allocate a fresh hugetlb page. All specific allocators
1700 * should use this function to get new hugetlb pages
1701 */
1702static struct page *alloc_fresh_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001703 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1704 nodemask_t *node_alloc_noretry)
Michal Hocko0c397da2018-01-31 16:20:56 -08001705{
1706 struct page *page;
1707
1708 if (hstate_is_gigantic(h))
1709 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1710 else
1711 page = alloc_buddy_huge_page(h, gfp_mask,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001712 nid, nmask, node_alloc_noretry);
Michal Hocko0c397da2018-01-31 16:20:56 -08001713 if (!page)
1714 return NULL;
1715
1716 if (hstate_is_gigantic(h))
1717 prep_compound_gigantic_page(page, huge_page_order(h));
1718 prep_new_huge_page(h, page, page_to_nid(page));
1719
1720 return page;
1721}
1722
1723/*
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001724 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1725 * manner.
1726 */
Mike Kravetzf60858f2019-09-23 15:37:35 -07001727static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1728 nodemask_t *node_alloc_noretry)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001729{
1730 struct page *page;
1731 int nr_nodes, node;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001732 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001733
1734 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
Mike Kravetzf60858f2019-09-23 15:37:35 -07001735 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1736 node_alloc_noretry);
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001737 if (page)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001738 break;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001739 }
1740
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001741 if (!page)
1742 return 0;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001743
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001744 put_page(page); /* free it into the hugepage allocator */
1745
1746 return 1;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001747}
1748
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001749/*
1750 * Free huge page from pool from next node to free.
1751 * Attempt to keep persistent huge pages more or less
1752 * balanced over allowed nodes.
1753 * Called with hugetlb_lock locked.
1754 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001755static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1756 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001757{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001758 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001759 int ret = 0;
1760
Joonsoo Kimb2261022013-09-11 14:21:00 -07001761 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001762 /*
1763 * If we're returning unused surplus pages, only examine
1764 * nodes with surplus pages.
1765 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001766 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1767 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001768 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001769 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001770 struct page, lru);
1771 list_del(&page->lru);
1772 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001773 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001774 if (acct_surplus) {
1775 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001776 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001777 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001778 update_and_free_page(h, page);
1779 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001780 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001781 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001782 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001783
1784 return ret;
1785}
1786
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001787/*
1788 * Dissolve a given free hugepage into free buddy pages. This function does
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001789 * nothing for in-use hugepages and non-hugepages.
1790 * This function returns values like below:
1791 *
1792 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1793 * (allocated or reserved.)
1794 * 0: successfully dissolved free hugepages or the page is not a
1795 * hugepage (considered as already dissolved)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001796 */
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001797int dissolve_free_huge_page(struct page *page)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001798{
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001799 int rc = -EBUSY;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001800
Muchun Songe334b1f2021-02-04 18:32:06 -08001801retry:
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001802 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1803 if (!PageHuge(page))
1804 return 0;
1805
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001806 spin_lock(&hugetlb_lock);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001807 if (!PageHuge(page)) {
1808 rc = 0;
1809 goto out;
1810 }
1811
1812 if (!page_count(page)) {
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001813 struct page *head = compound_head(page);
1814 struct hstate *h = page_hstate(head);
1815 int nid = page_to_nid(head);
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001816 if (h->free_huge_pages - h->resv_huge_pages == 0)
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001817 goto out;
Muchun Songe334b1f2021-02-04 18:32:06 -08001818
1819 /*
1820 * We should make sure that the page is already on the free list
1821 * when it is dissolved.
1822 */
1823 if (unlikely(!PageHugeFreed(head))) {
1824 spin_unlock(&hugetlb_lock);
1825 cond_resched();
1826
1827 /*
1828 * Theoretically, we should return -EBUSY when we
1829 * encounter this race. In fact, we have a chance
1830 * to successfully dissolve the page if we do a
1831 * retry. Because the race window is quite small.
1832 * If we seize this opportunity, it is an optimization
1833 * for increasing the success rate of dissolving page.
1834 */
1835 goto retry;
1836 }
1837
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001838 /*
1839 * Move PageHWPoison flag from head page to the raw error page,
1840 * which makes any subpages rather than the error page reusable.
1841 */
1842 if (PageHWPoison(head) && page != head) {
1843 SetPageHWPoison(page);
1844 ClearPageHWPoison(head);
1845 }
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001846 list_del(&head->lru);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001847 h->free_huge_pages--;
1848 h->free_huge_pages_node[nid]--;
zhong jiangc1470b32016-08-11 15:32:55 -07001849 h->max_huge_pages--;
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001850 update_and_free_page(h, head);
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001851 rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001852 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001853out:
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001854 spin_unlock(&hugetlb_lock);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001855 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001856}
1857
1858/*
1859 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1860 * make specified memory blocks removable from the system.
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001861 * Note that this will dissolve a free gigantic hugepage completely, if any
1862 * part of it lies within the given range.
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001863 * Also note that if dissolve_free_huge_page() returns with an error, all
1864 * free hugepages that were dissolved before that error are lost.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001865 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001866int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001867{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001868 unsigned long pfn;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001869 struct page *page;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001870 int rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001871
Li Zhongd0177632014-08-06 16:07:56 -07001872 if (!hugepages_supported())
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001873 return rc;
Li Zhongd0177632014-08-06 16:07:56 -07001874
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001875 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1876 page = pfn_to_page(pfn);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001877 rc = dissolve_free_huge_page(page);
1878 if (rc)
1879 break;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001880 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001881
1882 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001883}
1884
Michal Hockoab5ac902018-01-31 16:20:48 -08001885/*
1886 * Allocates a fresh surplus page from the page allocator.
1887 */
Michal Hocko0c397da2018-01-31 16:20:56 -08001888static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
Michal Hockoaaf14e42017-07-10 15:49:08 -07001889 int nid, nodemask_t *nmask)
Adam Litke7893d1d2007-10-16 01:26:18 -07001890{
Michal Hocko9980d742018-01-31 16:20:52 -08001891 struct page *page = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -07001892
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001893 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001894 return NULL;
1895
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001896 spin_lock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001897 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1898 goto out_unlock;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001899 spin_unlock(&hugetlb_lock);
1900
Mike Kravetzf60858f2019-09-23 15:37:35 -07001901 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hocko9980d742018-01-31 16:20:52 -08001902 if (!page)
Michal Hocko0c397da2018-01-31 16:20:56 -08001903 return NULL;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001904
1905 spin_lock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001906 /*
1907 * We could have raced with the pool size change.
1908 * Double check that and simply deallocate the new page
1909 * if we would end up overcommiting the surpluses. Abuse
1910 * temporary page to workaround the nasty free_huge_page
1911 * codeflow
1912 */
1913 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1914 SetPageHugeTemporary(page);
Kai Shen2bf753e2019-05-13 17:15:37 -07001915 spin_unlock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001916 put_page(page);
Kai Shen2bf753e2019-05-13 17:15:37 -07001917 return NULL;
Michal Hocko9980d742018-01-31 16:20:52 -08001918 } else {
Michal Hocko9980d742018-01-31 16:20:52 -08001919 h->surplus_huge_pages++;
Michal Hocko4704dea2018-03-09 15:50:55 -08001920 h->surplus_huge_pages_node[page_to_nid(page)]++;
Adam Litke7893d1d2007-10-16 01:26:18 -07001921 }
Michal Hocko9980d742018-01-31 16:20:52 -08001922
1923out_unlock:
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001924 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001925
1926 return page;
1927}
1928
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001929static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
Aneesh Kumar K.V9a4e9f32019-03-05 15:47:44 -08001930 int nid, nodemask_t *nmask)
Michal Hockoab5ac902018-01-31 16:20:48 -08001931{
1932 struct page *page;
1933
1934 if (hstate_is_gigantic(h))
1935 return NULL;
1936
Mike Kravetzf60858f2019-09-23 15:37:35 -07001937 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hockoab5ac902018-01-31 16:20:48 -08001938 if (!page)
1939 return NULL;
1940
1941 /*
1942 * We do not account these pages as surplus because they are only
1943 * temporary and will be released properly on the last reference
1944 */
Michal Hockoab5ac902018-01-31 16:20:48 -08001945 SetPageHugeTemporary(page);
1946
1947 return page;
1948}
1949
Adam Litkee4e574b2007-10-16 01:26:19 -07001950/*
Dave Hansen099730d2015-11-05 18:50:17 -08001951 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1952 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001953static
Michal Hocko0c397da2018-01-31 16:20:56 -08001954struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
Dave Hansen099730d2015-11-05 18:50:17 -08001955 struct vm_area_struct *vma, unsigned long addr)
1956{
Michal Hockoaaf14e42017-07-10 15:49:08 -07001957 struct page *page;
1958 struct mempolicy *mpol;
1959 gfp_t gfp_mask = htlb_alloc_mask(h);
1960 int nid;
1961 nodemask_t *nodemask;
1962
1963 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
Michal Hocko0c397da2018-01-31 16:20:56 -08001964 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
Michal Hockoaaf14e42017-07-10 15:49:08 -07001965 mpol_cond_put(mpol);
1966
1967 return page;
Dave Hansen099730d2015-11-05 18:50:17 -08001968}
1969
Michal Hockoab5ac902018-01-31 16:20:48 -08001970/* page migration callback function */
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001971struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07001972 nodemask_t *nmask, gfp_t gfp_mask)
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001973{
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001974 spin_lock(&hugetlb_lock);
1975 if (h->free_huge_pages - h->resv_huge_pages > 0) {
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001976 struct page *page;
1977
1978 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1979 if (page) {
1980 spin_unlock(&hugetlb_lock);
1981 return page;
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001982 }
1983 }
1984 spin_unlock(&hugetlb_lock);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001985
Michal Hocko0c397da2018-01-31 16:20:56 -08001986 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001987}
1988
Michal Hockoebd63722018-01-31 16:21:00 -08001989/* mempolicy aware migration callback */
Michal Hocko389c8172018-01-31 16:21:03 -08001990struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1991 unsigned long address)
Michal Hockoebd63722018-01-31 16:21:00 -08001992{
1993 struct mempolicy *mpol;
1994 nodemask_t *nodemask;
1995 struct page *page;
Michal Hockoebd63722018-01-31 16:21:00 -08001996 gfp_t gfp_mask;
1997 int node;
1998
Michal Hockoebd63722018-01-31 16:21:00 -08001999 gfp_mask = htlb_alloc_mask(h);
2000 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07002001 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
Michal Hockoebd63722018-01-31 16:21:00 -08002002 mpol_cond_put(mpol);
2003
2004 return page;
2005}
2006
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09002007/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002008 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07002009 * of size 'delta'.
2010 */
Andi Kleena5516432008-07-23 21:27:41 -07002011static int gather_surplus_pages(struct hstate *h, int delta)
Jules Irenge1b2a1e72020-04-06 20:08:09 -07002012 __must_hold(&hugetlb_lock)
Adam Litkee4e574b2007-10-16 01:26:19 -07002013{
2014 struct list_head surplus_list;
2015 struct page *page, *tmp;
2016 int ret, i;
2017 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07002018 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07002019
Andi Kleena5516432008-07-23 21:27:41 -07002020 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08002021 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002022 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002023 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08002024 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002025
2026 allocated = 0;
2027 INIT_LIST_HEAD(&surplus_list);
2028
2029 ret = -ENOMEM;
2030retry:
2031 spin_unlock(&hugetlb_lock);
2032 for (i = 0; i < needed; i++) {
Michal Hocko0c397da2018-01-31 16:20:56 -08002033 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
Michal Hockoaaf14e42017-07-10 15:49:08 -07002034 NUMA_NO_NODE, NULL);
Hillf Danton28073b02012-03-21 16:34:00 -07002035 if (!page) {
2036 alloc_ok = false;
2037 break;
2038 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002039 list_add(&page->lru, &surplus_list);
David Rientjes69ed7792017-07-10 15:48:50 -07002040 cond_resched();
Adam Litkee4e574b2007-10-16 01:26:19 -07002041 }
Hillf Danton28073b02012-03-21 16:34:00 -07002042 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07002043
2044 /*
2045 * After retaking hugetlb_lock, we need to recalculate 'needed'
2046 * because either resv_huge_pages or free_huge_pages may have changed.
2047 */
2048 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002049 needed = (h->resv_huge_pages + delta) -
2050 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07002051 if (needed > 0) {
2052 if (alloc_ok)
2053 goto retry;
2054 /*
2055 * We were not able to allocate enough pages to
2056 * satisfy the entire reservation so we free what
2057 * we've allocated so far.
2058 */
2059 goto free;
2060 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002061 /*
2062 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002063 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07002064 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08002065 * allocator. Commit the entire reservation here to prevent another
2066 * process from stealing the pages as they are added to the pool but
2067 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07002068 */
2069 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07002070 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002071 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002072
Adam Litke19fc3f02008-04-28 02:12:20 -07002073 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07002074 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07002075 if ((--needed) < 0)
2076 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002077 /*
2078 * This page is now managed by the hugetlb allocator and has
2079 * no users -- drop the buddy allocator's reference.
2080 */
2081 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08002082 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07002083 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07002084 }
Hillf Danton28073b02012-03-21 16:34:00 -07002085free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08002086 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07002087
2088 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07002089 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2090 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002091 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002092
2093 return ret;
2094}
2095
2096/*
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002097 * This routine has two main purposes:
2098 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2099 * in unused_resv_pages. This corresponds to the prior adjustments made
2100 * to the associated reservation map.
2101 * 2) Free any unused surplus pages that may have been allocated to satisfy
2102 * the reservation. As many as unused_resv_pages may be freed.
2103 *
2104 * Called with hugetlb_lock held. However, the lock could be dropped (and
2105 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2106 * we must make sure nobody else can claim pages we are in the process of
2107 * freeing. Do this by ensuring resv_huge_page always is greater than the
2108 * number of huge pages we plan to free when dropping the lock.
Adam Litkee4e574b2007-10-16 01:26:19 -07002109 */
Andi Kleena5516432008-07-23 21:27:41 -07002110static void return_unused_surplus_pages(struct hstate *h,
2111 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07002112{
Adam Litkee4e574b2007-10-16 01:26:19 -07002113 unsigned long nr_pages;
2114
Andi Kleenaa888a72008-07-23 21:27:47 -07002115 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002116 if (hstate_is_gigantic(h))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002117 goto out;
Andi Kleenaa888a72008-07-23 21:27:47 -07002118
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002119 /*
2120 * Part (or even all) of the reservation could have been backed
2121 * by pre-allocated pages. Only free surplus pages.
2122 */
Andi Kleena5516432008-07-23 21:27:41 -07002123 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07002124
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002125 /*
2126 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002127 * evenly across all nodes with memory. Iterate across these nodes
2128 * until we can no longer free unreserved surplus pages. This occurs
2129 * when the nodes with surplus pages have no free pages.
Randy Dunlap9e7ee402020-08-11 18:32:59 -07002130 * free_pool_huge_page() will balance the freed pages across the
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002131 * on-line nodes with memory and will handle the hstate accounting.
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002132 *
2133 * Note that we decrement resv_huge_pages as we free the pages. If
2134 * we drop the lock, resv_huge_pages will still be sufficiently large
2135 * to cover subsequent pages we may free.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002136 */
2137 while (nr_pages--) {
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002138 h->resv_huge_pages--;
2139 unused_resv_pages--;
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002140 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002141 goto out;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07002142 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002143 }
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002144
2145out:
2146 /* Fully uncommit the reservation */
2147 h->resv_huge_pages -= unused_resv_pages;
Adam Litkee4e574b2007-10-16 01:26:19 -07002148}
2149
Mike Kravetz5e911372015-09-08 15:01:28 -07002150
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002151/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002152 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07002153 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002154 *
2155 * vma_needs_reservation is called to determine if the huge page at addr
2156 * within the vma has an associated reservation. If a reservation is
2157 * needed, the value 1 is returned. The caller is then responsible for
2158 * managing the global reservation and subpool usage counts. After
2159 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002160 * to add the page to the reservation map. If the page allocation fails,
2161 * the reservation must be ended instead of committed. vma_end_reservation
2162 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002163 *
2164 * In the normal case, vma_commit_reservation returns the same value
2165 * as the preceding vma_needs_reservation call. The only time this
2166 * is not the case is if a reserve map was changed between calls. It
2167 * is the responsibility of the caller to notice the difference and
2168 * take appropriate action.
Mike Kravetz96b96a92016-11-10 10:46:32 -08002169 *
2170 * vma_add_reservation is used in error paths where a reservation must
2171 * be restored when a newly allocated huge page must be freed. It is
2172 * to be called after calling vma_needs_reservation to determine if a
2173 * reservation exists.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002174 */
Mike Kravetz5e911372015-09-08 15:01:28 -07002175enum vma_resv_mode {
2176 VMA_NEEDS_RESV,
2177 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002178 VMA_END_RESV,
Mike Kravetz96b96a92016-11-10 10:46:32 -08002179 VMA_ADD_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07002180};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002181static long __vma_reservation_common(struct hstate *h,
2182 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07002183 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002184{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002185 struct resv_map *resv;
2186 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002187 long ret;
Mina Almasry0db9d742020-04-01 21:11:25 -07002188 long dummy_out_regions_needed;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002189
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002190 resv = vma_resv_map(vma);
2191 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07002192 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002193
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002194 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07002195 switch (mode) {
2196 case VMA_NEEDS_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002197 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2198 /* We assume that vma_reservation_* routines always operate on
2199 * 1 page, and that adding to resv map a 1 page entry can only
2200 * ever require 1 region.
2201 */
2202 VM_BUG_ON(dummy_out_regions_needed != 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002203 break;
2204 case VMA_COMMIT_RESV:
Mina Almasry075a61d2020-04-01 21:11:28 -07002205 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002206 /* region_add calls of range 1 should never fail. */
2207 VM_BUG_ON(ret < 0);
Mike Kravetz5e911372015-09-08 15:01:28 -07002208 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002209 case VMA_END_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002210 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002211 ret = 0;
2212 break;
Mike Kravetz96b96a92016-11-10 10:46:32 -08002213 case VMA_ADD_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002214 if (vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07002215 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002216 /* region_add calls of range 1 should never fail. */
2217 VM_BUG_ON(ret < 0);
2218 } else {
2219 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz96b96a92016-11-10 10:46:32 -08002220 ret = region_del(resv, idx, idx + 1);
2221 }
2222 break;
Mike Kravetz5e911372015-09-08 15:01:28 -07002223 default:
2224 BUG();
2225 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002226
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002227 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002228 return ret;
Mike Kravetz67961f92016-06-08 15:33:42 -07002229 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2230 /*
2231 * In most cases, reserves always exist for private mappings.
2232 * However, a file associated with mapping could have been
2233 * hole punched or truncated after reserves were consumed.
2234 * As subsequent fault on such a range will not use reserves.
2235 * Subtle - The reserve map for private mappings has the
2236 * opposite meaning than that of shared mappings. If NO
2237 * entry is in the reserve map, it means a reservation exists.
2238 * If an entry exists in the reserve map, it means the
2239 * reservation has already been consumed. As a result, the
2240 * return value of this routine is the opposite of the
2241 * value returned from reserve map manipulation routines above.
2242 */
2243 if (ret)
2244 return 0;
2245 else
2246 return 1;
2247 }
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002248 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002249 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002250}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002251
2252static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07002253 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002254{
Mike Kravetz5e911372015-09-08 15:01:28 -07002255 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002256}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002257
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002258static long vma_commit_reservation(struct hstate *h,
2259 struct vm_area_struct *vma, unsigned long addr)
2260{
Mike Kravetz5e911372015-09-08 15:01:28 -07002261 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2262}
2263
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002264static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07002265 struct vm_area_struct *vma, unsigned long addr)
2266{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002267 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002268}
2269
Mike Kravetz96b96a92016-11-10 10:46:32 -08002270static long vma_add_reservation(struct hstate *h,
2271 struct vm_area_struct *vma, unsigned long addr)
2272{
2273 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2274}
2275
2276/*
2277 * This routine is called to restore a reservation on error paths. In the
2278 * specific error paths, a huge page was allocated (via alloc_huge_page)
2279 * and is about to be freed. If a reservation for the page existed,
2280 * alloc_huge_page would have consumed the reservation and set PagePrivate
2281 * in the newly allocated page. When the page is freed via free_huge_page,
2282 * the global reservation count will be incremented if PagePrivate is set.
2283 * However, free_huge_page can not adjust the reserve map. Adjust the
2284 * reserve map here to be consistent with global reserve count adjustments
2285 * to be made by free_huge_page.
2286 */
2287static void restore_reserve_on_error(struct hstate *h,
2288 struct vm_area_struct *vma, unsigned long address,
2289 struct page *page)
2290{
2291 if (unlikely(PagePrivate(page))) {
2292 long rc = vma_needs_reservation(h, vma, address);
2293
2294 if (unlikely(rc < 0)) {
2295 /*
2296 * Rare out of memory condition in reserve map
2297 * manipulation. Clear PagePrivate so that
2298 * global reserve count will not be incremented
2299 * by free_huge_page. This will make it appear
2300 * as though the reservation for this page was
2301 * consumed. This may prevent the task from
2302 * faulting in the page at a later time. This
2303 * is better than inconsistent global huge page
2304 * accounting of reserve counts.
2305 */
2306 ClearPagePrivate(page);
2307 } else if (rc) {
2308 rc = vma_add_reservation(h, vma, address);
2309 if (unlikely(rc < 0))
2310 /*
2311 * See above comment about rare out of
2312 * memory condition.
2313 */
2314 ClearPagePrivate(page);
2315 } else
2316 vma_end_reservation(h, vma, address);
2317 }
2318}
2319
Mike Kravetz70c35472015-09-08 15:01:54 -07002320struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002321 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322{
David Gibson90481622012-03-21 16:34:12 -07002323 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07002324 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08002325 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002326 long map_chg, map_commit;
2327 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002328 int ret, idx;
2329 struct hugetlb_cgroup *h_cg;
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002330 bool deferred_reserve;
Adam Litke2fc39ce2007-11-14 16:59:39 -08002331
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002332 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07002333 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002334 * Examine the region/reserve map to determine if the process
2335 * has a reservation for the page to be allocated. A return
2336 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07002337 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002338 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2339 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002340 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002341
2342 /*
2343 * Processes that did not create the mapping will have no
2344 * reserves as indicated by the region/reserve map. Check
2345 * that the allocation will not exceed the subpool limit.
2346 * Allocations for MAP_NORESERVE mappings also need to be
2347 * checked against any subpool limit.
2348 */
2349 if (map_chg || avoid_reserve) {
2350 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2351 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002352 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002353 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07002354 }
Mel Gormana1e78772008-07-23 21:27:23 -07002355
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002356 /*
2357 * Even though there was no reservation in the region/reserve
2358 * map, there could be reservations associated with the
2359 * subpool that can be used. This would be indicated if the
2360 * return value of hugepage_subpool_get_pages() is zero.
2361 * However, if avoid_reserve is specified we still avoid even
2362 * the subpool reservations.
2363 */
2364 if (avoid_reserve)
2365 gbl_chg = 1;
2366 }
2367
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002368 /* If this allocation is not consuming a reservation, charge it now.
2369 */
2370 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2371 if (deferred_reserve) {
2372 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2373 idx, pages_per_huge_page(h), &h_cg);
2374 if (ret)
2375 goto out_subpool_put;
2376 }
2377
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002378 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002379 if (ret)
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002380 goto out_uncharge_cgroup_reservation;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002381
Mel Gormana1e78772008-07-23 21:27:23 -07002382 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002383 /*
2384 * glb_chg is passed to indicate whether or not a page must be taken
2385 * from the global free pool (global change). gbl_chg == 0 indicates
2386 * a reservation exists for the allocation.
2387 */
2388 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002389 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07002390 spin_unlock(&hugetlb_lock);
Michal Hocko0c397da2018-01-31 16:20:56 -08002391 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002392 if (!page)
2393 goto out_uncharge_cgroup;
Rik van Riel568e3812022-10-17 20:25:05 -04002394 spin_lock(&hugetlb_lock);
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002395 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2396 SetPagePrivate(page);
2397 h->resv_huge_pages--;
2398 }
Wei Yang15a8d682020-10-13 16:56:33 -07002399 list_add(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002400 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07002401 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002402 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002403 /* If allocation is not consuming a reservation, also store the
2404 * hugetlb_cgroup pointer on the page.
2405 */
2406 if (deferred_reserve) {
2407 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2408 h_cg, page);
2409 }
2410
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002411 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07002412
David Gibson90481622012-03-21 16:34:12 -07002413 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07002414
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002415 map_commit = vma_commit_reservation(h, vma, addr);
2416 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07002417 /*
2418 * The page was added to the reservation map between
2419 * vma_needs_reservation and vma_commit_reservation.
2420 * This indicates a race with hugetlb_reserve_pages.
2421 * Adjust for the subpool count incremented above AND
2422 * in hugetlb_reserve_pages for the same page. Also,
2423 * the reservation count added in hugetlb_reserve_pages
2424 * no longer applies.
2425 */
2426 long rsv_adjust;
2427
2428 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2429 hugetlb_acct_memory(h, -rsv_adjust);
Mike Kravetz79aa9252020-11-01 17:07:27 -08002430 if (deferred_reserve)
2431 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2432 pages_per_huge_page(h), page);
Mike Kravetz33039672015-06-24 16:57:58 -07002433 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08002434 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002435
2436out_uncharge_cgroup:
2437 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002438out_uncharge_cgroup_reservation:
2439 if (deferred_reserve)
2440 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2441 h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002442out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002443 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002444 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002445 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002446 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08002447}
2448
Aneesh Kumar K.Ve24a1302017-07-28 10:31:25 +05302449int alloc_bootmem_huge_page(struct hstate *h)
2450 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2451int __alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002452{
2453 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002454 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002455
Joonsoo Kimb2261022013-09-11 14:21:00 -07002456 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002457 void *addr;
2458
Mike Rapoporteb31d552018-10-30 15:08:04 -07002459 addr = memblock_alloc_try_nid_raw(
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002460 huge_page_size(h), huge_page_size(h),
Mike Rapoport97ad1082018-10-30 15:09:44 -07002461 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002462 if (addr) {
2463 /*
2464 * Use the beginning of the huge page to store the
2465 * huge_bootmem_page struct (until gather_bootmem
2466 * puts them into the mem_map).
2467 */
2468 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002469 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002470 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002471 }
2472 return 0;
2473
2474found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002475 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002476 /* Put them into a private list first because mem_map is not up yet */
Cannon Matthews330d6e42018-08-17 15:49:17 -07002477 INIT_LIST_HEAD(&m->list);
Andi Kleenaa888a72008-07-23 21:27:47 -07002478 list_add(&m->list, &huge_boot_pages);
2479 m->hstate = h;
2480 return 1;
2481}
2482
Mike Kravetzebd6a292021-06-30 18:48:31 -07002483/*
2484 * Put bootmem huge pages into the standard lists after mem_map is up.
2485 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
2486 */
Andi Kleenaa888a72008-07-23 21:27:47 -07002487static void __init gather_bootmem_prealloc(void)
2488{
2489 struct huge_bootmem_page *m;
2490
2491 list_for_each_entry(m, &huge_boot_pages, list) {
Mike Kravetz40d18eb2018-08-17 15:49:07 -07002492 struct page *page = virt_to_page(m);
Andi Kleenaa888a72008-07-23 21:27:47 -07002493 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002494
Mike Kravetzebd6a292021-06-30 18:48:31 -07002495 VM_BUG_ON(!hstate_is_gigantic(h));
Andi Kleenaa888a72008-07-23 21:27:47 -07002496 WARN_ON(page_count(page) != 1);
Mike Kravetzebd6a292021-06-30 18:48:31 -07002497 prep_compound_gigantic_page(page, huge_page_order(h));
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002498 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002499 prep_new_huge_page(h, page, page_to_nid(page));
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08002500 put_page(page); /* free it into the hugepage allocator */
2501
Rafael Aquinib0320c72011-06-15 15:08:39 -07002502 /*
Mike Kravetzebd6a292021-06-30 18:48:31 -07002503 * We need to restore the 'stolen' pages to totalram_pages
2504 * in order to fix confusing memory reports from free(1) and
2505 * other side-effects, like CommitLimit going negative.
Rafael Aquinib0320c72011-06-15 15:08:39 -07002506 */
Mike Kravetzebd6a292021-06-30 18:48:31 -07002507 adjust_managed_page_count(page, pages_per_huge_page(h));
Cannon Matthews520495f2018-07-03 17:02:43 -07002508 cond_resched();
Andi Kleenaa888a72008-07-23 21:27:47 -07002509 }
2510}
2511
Andi Kleen8faa8b02008-07-23 21:27:48 -07002512static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513{
2514 unsigned long i;
Mike Kravetzf60858f2019-09-23 15:37:35 -07002515 nodemask_t *node_alloc_noretry;
2516
2517 if (!hstate_is_gigantic(h)) {
2518 /*
2519 * Bit mask controlling how hard we retry per-node allocations.
2520 * Ignore errors as lower level routines can deal with
2521 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2522 * time, we are likely in bigger trouble.
2523 */
2524 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2525 GFP_KERNEL);
2526 } else {
2527 /* allocations done at boot time */
2528 node_alloc_noretry = NULL;
2529 }
2530
2531 /* bit mask controlling how hard we retry per-node allocations */
2532 if (node_alloc_noretry)
2533 nodes_clear(*node_alloc_noretry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534
Andi Kleene5ff2152008-07-23 21:27:42 -07002535 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002536 if (hstate_is_gigantic(h)) {
Barry Songdbda8fe2020-07-23 21:15:30 -07002537 if (hugetlb_cma_size) {
Roman Gushchincf11e852020-04-10 14:32:45 -07002538 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
Chen Wandunc9ea7712021-02-24 12:07:58 -08002539 goto free;
Roman Gushchincf11e852020-04-10 14:32:45 -07002540 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002541 if (!alloc_bootmem_huge_page(h))
2542 break;
Michal Hocko0c397da2018-01-31 16:20:56 -08002543 } else if (!alloc_pool_huge_page(h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07002544 &node_states[N_MEMORY],
2545 node_alloc_noretry))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 break;
David Rientjes69ed7792017-07-10 15:48:50 -07002547 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 }
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002549 if (i < h->max_huge_pages) {
2550 char buf[32];
2551
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002552 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002553 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2554 h->max_huge_pages, buf, i);
2555 h->max_huge_pages = i;
2556 }
Chen Wandunc9ea7712021-02-24 12:07:58 -08002557free:
Mike Kravetzf60858f2019-09-23 15:37:35 -07002558 kfree(node_alloc_noretry);
Andi Kleene5ff2152008-07-23 21:27:42 -07002559}
2560
2561static void __init hugetlb_init_hstates(void)
2562{
2563 struct hstate *h;
2564
2565 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002566 if (minimum_order > huge_page_order(h))
2567 minimum_order = huge_page_order(h);
2568
Andi Kleen8faa8b02008-07-23 21:27:48 -07002569 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002570 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002571 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002572 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002573 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002574}
2575
2576static void __init report_hugepages(void)
2577{
2578 struct hstate *h;
2579
2580 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002581 char buf[32];
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002582
2583 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Andrew Mortonffb22af2013-02-22 16:32:08 -08002584 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002585 buf, h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002586 }
2587}
2588
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002590static void try_to_free_low(struct hstate *h, unsigned long count,
2591 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002593 int i;
2594
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002595 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002596 return;
2597
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002598 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002600 struct list_head *freel = &h->hugepage_freelists[i];
2601 list_for_each_entry_safe(page, next, freel, lru) {
2602 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002603 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604 if (PageHighMem(page))
2605 continue;
2606 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002607 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002608 h->free_huge_pages--;
2609 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 }
2611 }
2612}
2613#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002614static inline void try_to_free_low(struct hstate *h, unsigned long count,
2615 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616{
2617}
2618#endif
2619
Wu Fengguang20a03072009-06-16 15:32:22 -07002620/*
2621 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2622 * balanced by operating on them in a round-robin fashion.
2623 * Returns 1 if an adjustment was made.
2624 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002625static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2626 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002627{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002628 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002629
2630 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002631
Joonsoo Kimb2261022013-09-11 14:21:00 -07002632 if (delta < 0) {
2633 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2634 if (h->surplus_huge_pages_node[node])
2635 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002636 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002637 } else {
2638 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2639 if (h->surplus_huge_pages_node[node] <
2640 h->nr_huge_pages_node[node])
2641 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002642 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002643 }
2644 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002645
Joonsoo Kimb2261022013-09-11 14:21:00 -07002646found:
2647 h->surplus_huge_pages += delta;
2648 h->surplus_huge_pages_node[node] += delta;
2649 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002650}
2651
Andi Kleena5516432008-07-23 21:27:41 -07002652#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002653static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002654 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655{
Adam Litke7893d1d2007-10-16 01:26:18 -07002656 unsigned long min_count, ret;
Mike Kravetzf60858f2019-09-23 15:37:35 -07002657 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2658
2659 /*
2660 * Bit mask controlling how hard we retry per-node allocations.
2661 * If we can not allocate the bit mask, do not attempt to allocate
2662 * the requested huge pages.
2663 */
2664 if (node_alloc_noretry)
2665 nodes_clear(*node_alloc_noretry);
2666 else
2667 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002669 spin_lock(&hugetlb_lock);
2670
2671 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002672 * Check for a node specific request.
2673 * Changing node specific huge page count may require a corresponding
2674 * change to the global count. In any case, the passed node mask
2675 * (nodes_allowed) will restrict alloc/free to the specified node.
2676 */
2677 if (nid != NUMA_NO_NODE) {
2678 unsigned long old_count = count;
2679
2680 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2681 /*
2682 * User may have specified a large count value which caused the
2683 * above calculation to overflow. In this case, they wanted
2684 * to allocate as many huge pages as possible. Set count to
2685 * largest possible value to align with their intention.
2686 */
2687 if (count < old_count)
2688 count = ULONG_MAX;
2689 }
2690
2691 /*
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002692 * Gigantic pages runtime allocation depend on the capability for large
2693 * page range allocation.
2694 * If the system does not provide this feature, return an error when
2695 * the user tries to allocate gigantic pages but let the user free the
2696 * boottime allocated gigantic pages.
2697 */
2698 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2699 if (count > persistent_huge_pages(h)) {
2700 spin_unlock(&hugetlb_lock);
Mike Kravetzf60858f2019-09-23 15:37:35 -07002701 NODEMASK_FREE(node_alloc_noretry);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002702 return -EINVAL;
2703 }
2704 /* Fall through to decrease pool */
2705 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002706
Adam Litke7893d1d2007-10-16 01:26:18 -07002707 /*
2708 * Increase the pool size
2709 * First take pages out of surplus state. Then make up the
2710 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002711 *
Michal Hocko0c397da2018-01-31 16:20:56 -08002712 * We might race with alloc_surplus_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002713 * to convert a surplus huge page to a normal huge page. That is
2714 * not critical, though, it just means the overall size of the
2715 * pool might be one hugepage larger than it needs to be, but
2716 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002717 */
Andi Kleena5516432008-07-23 21:27:41 -07002718 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002719 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002720 break;
2721 }
2722
Andi Kleena5516432008-07-23 21:27:41 -07002723 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002724 /*
2725 * If this allocation races such that we no longer need the
2726 * page, free_huge_page will handle it by freeing the page
2727 * and reducing the surplus.
2728 */
2729 spin_unlock(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07002730
2731 /* yield cpu to avoid soft lockup */
2732 cond_resched();
2733
Mike Kravetzf60858f2019-09-23 15:37:35 -07002734 ret = alloc_pool_huge_page(h, nodes_allowed,
2735 node_alloc_noretry);
Adam Litke7893d1d2007-10-16 01:26:18 -07002736 spin_lock(&hugetlb_lock);
2737 if (!ret)
2738 goto out;
2739
Mel Gorman536240f22009-12-14 17:59:56 -08002740 /* Bail for signals. Probably ctrl-c from user */
2741 if (signal_pending(current))
2742 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002743 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002744
2745 /*
2746 * Decrease the pool size
2747 * First return free pages to the buddy allocator (being careful
2748 * to keep enough around to satisfy reservations). Then place
2749 * pages into surplus state as needed so the pool will shrink
2750 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002751 *
2752 * By placing pages into the surplus state independent of the
2753 * overcommit value, we are allowing the surplus pool size to
2754 * exceed overcommit. There are few sane options here. Since
Michal Hocko0c397da2018-01-31 16:20:56 -08002755 * alloc_surplus_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002756 * though, we'll note that we're not allowed to exceed surplus
2757 * and won't grow the pool anywhere else. Not until one of the
2758 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002759 */
Andi Kleena5516432008-07-23 21:27:41 -07002760 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002761 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002762 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002763 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002764 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002766 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 }
Andi Kleena5516432008-07-23 21:27:41 -07002768 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002769 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002770 break;
2771 }
2772out:
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002773 h->max_huge_pages = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 spin_unlock(&hugetlb_lock);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002775
Mike Kravetzf60858f2019-09-23 15:37:35 -07002776 NODEMASK_FREE(node_alloc_noretry);
2777
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002778 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779}
2780
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002781#define HSTATE_ATTR_RO(_name) \
2782 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2783
2784#define HSTATE_ATTR(_name) \
2785 static struct kobj_attribute _name##_attr = \
2786 __ATTR(_name, 0644, _name##_show, _name##_store)
2787
2788static struct kobject *hugepages_kobj;
2789static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2790
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002791static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2792
2793static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002794{
2795 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002796
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002797 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002798 if (hstate_kobjs[i] == kobj) {
2799 if (nidp)
2800 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002801 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002802 }
2803
2804 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002805}
2806
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002807static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002808 struct kobj_attribute *attr, char *buf)
2809{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002810 struct hstate *h;
2811 unsigned long nr_huge_pages;
2812 int nid;
2813
2814 h = kobj_to_hstate(kobj, &nid);
2815 if (nid == NUMA_NO_NODE)
2816 nr_huge_pages = h->nr_huge_pages;
2817 else
2818 nr_huge_pages = h->nr_huge_pages_node[nid];
2819
2820 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002821}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002822
David Rientjes238d3c12014-08-06 16:06:51 -07002823static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2824 struct hstate *h, int nid,
2825 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002826{
2827 int err;
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002828 nodemask_t nodes_allowed, *n_mask;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002829
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002830 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2831 return -EINVAL;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002832
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002833 if (nid == NUMA_NO_NODE) {
2834 /*
2835 * global hstate attribute
2836 */
2837 if (!(obey_mempolicy &&
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002838 init_nodemask_of_mempolicy(&nodes_allowed)))
2839 n_mask = &node_states[N_MEMORY];
2840 else
2841 n_mask = &nodes_allowed;
2842 } else {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002843 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002844 * Node specific request. count adjustment happens in
2845 * set_max_huge_pages() after acquiring hugetlb_lock.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002846 */
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002847 init_nodemask_of_node(&nodes_allowed, nid);
2848 n_mask = &nodes_allowed;
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002849 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002850
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002851 err = set_max_huge_pages(h, count, nid, n_mask);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002852
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002853 return err ? err : len;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002854}
2855
David Rientjes238d3c12014-08-06 16:06:51 -07002856static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2857 struct kobject *kobj, const char *buf,
2858 size_t len)
2859{
2860 struct hstate *h;
2861 unsigned long count;
2862 int nid;
2863 int err;
2864
2865 err = kstrtoul(buf, 10, &count);
2866 if (err)
2867 return err;
2868
2869 h = kobj_to_hstate(kobj, &nid);
2870 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2871}
2872
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002873static ssize_t nr_hugepages_show(struct kobject *kobj,
2874 struct kobj_attribute *attr, char *buf)
2875{
2876 return nr_hugepages_show_common(kobj, attr, buf);
2877}
2878
2879static ssize_t nr_hugepages_store(struct kobject *kobj,
2880 struct kobj_attribute *attr, const char *buf, size_t len)
2881{
David Rientjes238d3c12014-08-06 16:06:51 -07002882 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002883}
2884HSTATE_ATTR(nr_hugepages);
2885
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002886#ifdef CONFIG_NUMA
2887
2888/*
2889 * hstate attribute for optionally mempolicy-based constraint on persistent
2890 * huge page alloc/free.
2891 */
2892static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2893 struct kobj_attribute *attr, char *buf)
2894{
2895 return nr_hugepages_show_common(kobj, attr, buf);
2896}
2897
2898static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2899 struct kobj_attribute *attr, const char *buf, size_t len)
2900{
David Rientjes238d3c12014-08-06 16:06:51 -07002901 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002902}
2903HSTATE_ATTR(nr_hugepages_mempolicy);
2904#endif
2905
2906
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002907static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2908 struct kobj_attribute *attr, char *buf)
2909{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002910 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002911 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2912}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002913
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002914static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2915 struct kobj_attribute *attr, const char *buf, size_t count)
2916{
2917 int err;
2918 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002919 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002920
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002921 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002922 return -EINVAL;
2923
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002924 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002925 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002926 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002927
2928 spin_lock(&hugetlb_lock);
2929 h->nr_overcommit_huge_pages = input;
2930 spin_unlock(&hugetlb_lock);
2931
2932 return count;
2933}
2934HSTATE_ATTR(nr_overcommit_hugepages);
2935
2936static ssize_t free_hugepages_show(struct kobject *kobj,
2937 struct kobj_attribute *attr, char *buf)
2938{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002939 struct hstate *h;
2940 unsigned long free_huge_pages;
2941 int nid;
2942
2943 h = kobj_to_hstate(kobj, &nid);
2944 if (nid == NUMA_NO_NODE)
2945 free_huge_pages = h->free_huge_pages;
2946 else
2947 free_huge_pages = h->free_huge_pages_node[nid];
2948
2949 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002950}
2951HSTATE_ATTR_RO(free_hugepages);
2952
2953static ssize_t resv_hugepages_show(struct kobject *kobj,
2954 struct kobj_attribute *attr, char *buf)
2955{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002956 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002957 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2958}
2959HSTATE_ATTR_RO(resv_hugepages);
2960
2961static ssize_t surplus_hugepages_show(struct kobject *kobj,
2962 struct kobj_attribute *attr, char *buf)
2963{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002964 struct hstate *h;
2965 unsigned long surplus_huge_pages;
2966 int nid;
2967
2968 h = kobj_to_hstate(kobj, &nid);
2969 if (nid == NUMA_NO_NODE)
2970 surplus_huge_pages = h->surplus_huge_pages;
2971 else
2972 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2973
2974 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002975}
2976HSTATE_ATTR_RO(surplus_hugepages);
2977
2978static struct attribute *hstate_attrs[] = {
2979 &nr_hugepages_attr.attr,
2980 &nr_overcommit_hugepages_attr.attr,
2981 &free_hugepages_attr.attr,
2982 &resv_hugepages_attr.attr,
2983 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002984#ifdef CONFIG_NUMA
2985 &nr_hugepages_mempolicy_attr.attr,
2986#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002987 NULL,
2988};
2989
Arvind Yadav67e5ed92017-09-06 16:22:06 -07002990static const struct attribute_group hstate_attr_group = {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002991 .attrs = hstate_attrs,
2992};
2993
Jeff Mahoney094e9532010-02-02 13:44:14 -08002994static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2995 struct kobject **hstate_kobjs,
Arvind Yadav67e5ed92017-09-06 16:22:06 -07002996 const struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002997{
2998 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07002999 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003000
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003001 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3002 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003003 return -ENOMEM;
3004
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003005 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Miaohe Lin89b2dbd2021-02-24 12:06:50 -08003006 if (retval) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003007 kobject_put(hstate_kobjs[hi]);
Miaohe Lin89b2dbd2021-02-24 12:06:50 -08003008 hstate_kobjs[hi] = NULL;
3009 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003010
3011 return retval;
3012}
3013
3014static void __init hugetlb_sysfs_init(void)
3015{
3016 struct hstate *h;
3017 int err;
3018
3019 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3020 if (!hugepages_kobj)
3021 return;
3022
3023 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003024 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3025 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003026 if (err)
Mike Kravetz282f4212020-06-03 16:00:46 -07003027 pr_err("HugeTLB: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003028 }
3029}
3030
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003031#ifdef CONFIG_NUMA
3032
3033/*
3034 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08003035 * with node devices in node_devices[] using a parallel array. The array
3036 * index of a node device or _hstate == node id.
3037 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003038 * the base kernel, on the hugetlb module.
3039 */
3040struct node_hstate {
3041 struct kobject *hugepages_kobj;
3042 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3043};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08003044static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003045
3046/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003047 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003048 */
3049static struct attribute *per_node_hstate_attrs[] = {
3050 &nr_hugepages_attr.attr,
3051 &free_hugepages_attr.attr,
3052 &surplus_hugepages_attr.attr,
3053 NULL,
3054};
3055
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003056static const struct attribute_group per_node_hstate_attr_group = {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003057 .attrs = per_node_hstate_attrs,
3058};
3059
3060/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003061 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003062 * Returns node id via non-NULL nidp.
3063 */
3064static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3065{
3066 int nid;
3067
3068 for (nid = 0; nid < nr_node_ids; nid++) {
3069 struct node_hstate *nhs = &node_hstates[nid];
3070 int i;
3071 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3072 if (nhs->hstate_kobjs[i] == kobj) {
3073 if (nidp)
3074 *nidp = nid;
3075 return &hstates[i];
3076 }
3077 }
3078
3079 BUG();
3080 return NULL;
3081}
3082
3083/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003084 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003085 * No-op if no hstate attributes attached.
3086 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003087static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003088{
3089 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003090 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003091
3092 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003093 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003094
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003095 for_each_hstate(h) {
3096 int idx = hstate_index(h);
3097 if (nhs->hstate_kobjs[idx]) {
3098 kobject_put(nhs->hstate_kobjs[idx]);
3099 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003100 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003101 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003102
3103 kobject_put(nhs->hugepages_kobj);
3104 nhs->hugepages_kobj = NULL;
3105}
3106
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003107
3108/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003109 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003110 * No-op if attributes already registered.
3111 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003112static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003113{
3114 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003115 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003116 int err;
3117
3118 if (nhs->hugepages_kobj)
3119 return; /* already allocated */
3120
3121 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08003122 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003123 if (!nhs->hugepages_kobj)
3124 return;
3125
3126 for_each_hstate(h) {
3127 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3128 nhs->hstate_kobjs,
3129 &per_node_hstate_attr_group);
3130 if (err) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003131 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003132 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003133 hugetlb_unregister_node(node);
3134 break;
3135 }
3136 }
3137}
3138
3139/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003140 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08003141 * devices of nodes that have memory. All on-line nodes should have
3142 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003143 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08003144static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003145{
3146 int nid;
3147
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08003148 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08003149 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08003150 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003151 hugetlb_register_node(node);
3152 }
3153
3154 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003155 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003156 * [un]register hstate attributes on node hotplug.
3157 */
3158 register_hugetlbfs_with_node(hugetlb_register_node,
3159 hugetlb_unregister_node);
3160}
3161#else /* !CONFIG_NUMA */
3162
3163static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3164{
3165 BUG();
3166 if (nidp)
3167 *nidp = -1;
3168 return NULL;
3169}
3170
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003171static void hugetlb_register_all_nodes(void) { }
3172
3173#endif
3174
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003175static int __init hugetlb_init(void)
3176{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003177 int i;
3178
Mike Kravetzc2833a52020-06-03 16:00:50 -07003179 if (!hugepages_supported()) {
3180 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3181 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07003182 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003183 }
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08003184
Mike Kravetz282f4212020-06-03 16:00:46 -07003185 /*
3186 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3187 * architectures depend on setup being done here.
3188 */
3189 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3190 if (!parsed_default_hugepagesz) {
3191 /*
3192 * If we did not parse a default huge page size, set
3193 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3194 * number of huge pages for this default size was implicitly
3195 * specified, set that here as well.
3196 * Note that the implicit setting will overwrite an explicit
3197 * setting. A warning will be printed in this case.
3198 */
3199 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3200 if (default_hstate_max_huge_pages) {
3201 if (default_hstate.max_huge_pages) {
3202 char buf[32];
Andi Kleenaa888a72008-07-23 21:27:47 -07003203
Mike Kravetz282f4212020-06-03 16:00:46 -07003204 string_get_size(huge_page_size(&default_hstate),
3205 1, STRING_UNITS_2, buf, 32);
3206 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3207 default_hstate.max_huge_pages, buf);
3208 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3209 default_hstate_max_huge_pages);
3210 }
3211 default_hstate.max_huge_pages =
3212 default_hstate_max_huge_pages;
3213 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003214 }
Andi Kleenaa888a72008-07-23 21:27:47 -07003215
Roman Gushchincf11e852020-04-10 14:32:45 -07003216 hugetlb_cma_check();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003217 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07003218 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003219 report_hugepages();
3220
3221 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003222 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08003223 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003224
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003225#ifdef CONFIG_SMP
3226 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3227#else
3228 num_fault_mutexes = 1;
3229#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003230 hugetlb_fault_mutex_table =
Kees Cook6da2ec52018-06-12 13:55:00 -07003231 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3232 GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003233 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003234
3235 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003236 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003237 return 0;
3238}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08003239subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003240
Mike Kravetzae94da82020-06-03 16:00:34 -07003241/* Overwritten by architectures with more huge page sizes */
3242bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003243{
Mike Kravetzae94da82020-06-03 16:00:34 -07003244 return size == HPAGE_SIZE;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003245}
3246
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08003247void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003248{
3249 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003250 unsigned long i;
3251
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003252 if (size_to_hstate(PAGE_SIZE << order)) {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003253 return;
3254 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003255 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003256 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003257 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003258 h->order = order;
3259 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003260 h->nr_huge_pages = 0;
3261 h->free_huge_pages = 0;
3262 for (i = 0; i < MAX_NUMNODES; ++i)
3263 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07003264 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07003265 h->next_nid_to_alloc = first_memory_node;
3266 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003267 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3268 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003269
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003270 parsed_hstate = h;
3271}
3272
Mike Kravetz282f4212020-06-03 16:00:46 -07003273/*
3274 * hugepages command line processing
3275 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3276 * specification. If not, ignore the hugepages value. hugepages can also
3277 * be the first huge page command line option in which case it implicitly
3278 * specifies the number of huge pages for the default size.
3279 */
3280static int __init hugepages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003281{
3282 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003283 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003284
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003285 if (!parsed_valid_hugepagesz) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003286 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003287 parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -07003288 return 0;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003289 }
Mike Kravetz282f4212020-06-03 16:00:46 -07003290
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003291 /*
Mike Kravetz282f4212020-06-03 16:00:46 -07003292 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3293 * yet, so this hugepages= parameter goes to the "default hstate".
3294 * Otherwise, it goes with the previously parsed hugepagesz or
3295 * default_hugepagesz.
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003296 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003297 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003298 mhp = &default_hstate_max_huge_pages;
3299 else
3300 mhp = &parsed_hstate->max_huge_pages;
3301
Andi Kleen8faa8b02008-07-23 21:27:48 -07003302 if (mhp == last_mhp) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003303 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3304 return 0;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003305 }
3306
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003307 if (sscanf(s, "%lu", mhp) <= 0)
3308 *mhp = 0;
3309
Andi Kleen8faa8b02008-07-23 21:27:48 -07003310 /*
3311 * Global state is always initialized later in hugetlb_init.
3312 * But we need to allocate >= MAX_ORDER hstates here early to still
3313 * use the bootmem allocator.
3314 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003315 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07003316 hugetlb_hstate_alloc_pages(parsed_hstate);
3317
3318 last_mhp = mhp;
3319
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003320 return 1;
3321}
Mike Kravetz282f4212020-06-03 16:00:46 -07003322__setup("hugepages=", hugepages_setup);
Nick Piggine11bfbf2008-07-23 21:27:52 -07003323
Mike Kravetz282f4212020-06-03 16:00:46 -07003324/*
3325 * hugepagesz command line processing
3326 * A specific huge page size can only be specified once with hugepagesz.
3327 * hugepagesz is followed by hugepages on the command line. The global
3328 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3329 * hugepagesz argument was valid.
3330 */
Mike Kravetz359f2542020-06-03 16:00:38 -07003331static int __init hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003332{
Mike Kravetz359f2542020-06-03 16:00:38 -07003333 unsigned long size;
Mike Kravetz282f4212020-06-03 16:00:46 -07003334 struct hstate *h;
3335
3336 parsed_valid_hugepagesz = false;
Mike Kravetz359f2542020-06-03 16:00:38 -07003337 size = (unsigned long)memparse(s, NULL);
3338
3339 if (!arch_hugetlb_valid_size(size)) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003340 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
Mike Kravetz359f2542020-06-03 16:00:38 -07003341 return 0;
3342 }
3343
Mike Kravetz282f4212020-06-03 16:00:46 -07003344 h = size_to_hstate(size);
3345 if (h) {
3346 /*
3347 * hstate for this size already exists. This is normally
3348 * an error, but is allowed if the existing hstate is the
3349 * default hstate. More specifically, it is only allowed if
3350 * the number of huge pages for the default hstate was not
3351 * previously specified.
3352 */
3353 if (!parsed_default_hugepagesz || h != &default_hstate ||
3354 default_hstate.max_huge_pages) {
3355 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3356 return 0;
3357 }
3358
3359 /*
3360 * No need to call hugetlb_add_hstate() as hstate already
3361 * exists. But, do set parsed_hstate so that a following
3362 * hugepages= parameter will be applied to this hstate.
3363 */
3364 parsed_hstate = h;
3365 parsed_valid_hugepagesz = true;
3366 return 1;
Mike Kravetz38237832020-06-03 16:00:42 -07003367 }
3368
Mike Kravetz359f2542020-06-03 16:00:38 -07003369 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
Mike Kravetz282f4212020-06-03 16:00:46 -07003370 parsed_valid_hugepagesz = true;
Nick Piggine11bfbf2008-07-23 21:27:52 -07003371 return 1;
3372}
Mike Kravetz359f2542020-06-03 16:00:38 -07003373__setup("hugepagesz=", hugepagesz_setup);
3374
Mike Kravetz282f4212020-06-03 16:00:46 -07003375/*
3376 * default_hugepagesz command line input
3377 * Only one instance of default_hugepagesz allowed on command line.
3378 */
Mike Kravetzae94da82020-06-03 16:00:34 -07003379static int __init default_hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003380{
Mike Kravetzae94da82020-06-03 16:00:34 -07003381 unsigned long size;
3382
Mike Kravetz282f4212020-06-03 16:00:46 -07003383 parsed_valid_hugepagesz = false;
Mike Kravetz282f4212020-06-03 16:00:46 -07003384 if (parsed_default_hugepagesz) {
3385 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3386 return 0;
3387 }
3388
3389 size = (unsigned long)memparse(s, NULL);
3390
3391 if (!arch_hugetlb_valid_size(size)) {
3392 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3393 return 0;
3394 }
3395
3396 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3397 parsed_valid_hugepagesz = true;
3398 parsed_default_hugepagesz = true;
3399 default_hstate_idx = hstate_index(size_to_hstate(size));
3400
3401 /*
3402 * The number of default huge pages (for this size) could have been
3403 * specified as the first hugetlb parameter: hugepages=X. If so,
3404 * then default_hstate_max_huge_pages is set. If the default huge
3405 * page size is gigantic (>= MAX_ORDER), then the pages must be
3406 * allocated here from bootmem allocator.
3407 */
3408 if (default_hstate_max_huge_pages) {
3409 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3410 if (hstate_is_gigantic(&default_hstate))
3411 hugetlb_hstate_alloc_pages(&default_hstate);
3412 default_hstate_max_huge_pages = 0;
3413 }
3414
Nick Piggine11bfbf2008-07-23 21:27:52 -07003415 return 1;
3416}
Mike Kravetzae94da82020-06-03 16:00:34 -07003417__setup("default_hugepagesz=", default_hugepagesz_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003418
Muchun Song8ca39e62020-08-11 18:30:32 -07003419static unsigned int allowed_mems_nr(struct hstate *h)
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003420{
3421 int node;
3422 unsigned int nr = 0;
Muchun Song8ca39e62020-08-11 18:30:32 -07003423 nodemask_t *mpol_allowed;
3424 unsigned int *array = h->free_huge_pages_node;
3425 gfp_t gfp_mask = htlb_alloc_mask(h);
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003426
Muchun Song8ca39e62020-08-11 18:30:32 -07003427 mpol_allowed = policy_nodemask_current(gfp_mask);
3428
3429 for_each_node_mask(node, cpuset_current_mems_allowed) {
3430 if (!mpol_allowed ||
3431 (mpol_allowed && node_isset(node, *mpol_allowed)))
3432 nr += array[node];
3433 }
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003434
3435 return nr;
3436}
3437
3438#ifdef CONFIG_SYSCTL
Muchun Song177437982020-09-04 16:36:13 -07003439static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3440 void *buffer, size_t *length,
3441 loff_t *ppos, unsigned long *out)
3442{
3443 struct ctl_table dup_table;
3444
3445 /*
3446 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3447 * can duplicate the @table and alter the duplicate of it.
3448 */
3449 dup_table = *table;
3450 dup_table.data = out;
3451
3452 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3453}
3454
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003455static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3456 struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003457 void *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458{
Andi Kleene5ff2152008-07-23 21:27:42 -07003459 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07003460 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08003461 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003462
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003463 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003464 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003465
Muchun Song177437982020-09-04 16:36:13 -07003466 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3467 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003468 if (ret)
3469 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003470
David Rientjes238d3c12014-08-06 16:06:51 -07003471 if (write)
3472 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3473 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08003474out:
3475 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476}
Mel Gorman396faf02007-07-17 04:03:13 -07003477
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003478int hugetlb_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003479 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003480{
3481
3482 return hugetlb_sysctl_handler_common(false, table, write,
3483 buffer, length, ppos);
3484}
3485
3486#ifdef CONFIG_NUMA
3487int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003488 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003489{
3490 return hugetlb_sysctl_handler_common(true, table, write,
3491 buffer, length, ppos);
3492}
3493#endif /* CONFIG_NUMA */
3494
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003495int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003496 void *buffer, size_t *length, loff_t *ppos)
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003497{
Andi Kleena5516432008-07-23 21:27:41 -07003498 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07003499 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08003500 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003501
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003502 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003503 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003504
Petr Holasekc033a932011-03-22 16:33:05 -07003505 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07003506
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07003507 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08003508 return -EINVAL;
3509
Muchun Song177437982020-09-04 16:36:13 -07003510 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3511 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003512 if (ret)
3513 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003514
3515 if (write) {
3516 spin_lock(&hugetlb_lock);
3517 h->nr_overcommit_huge_pages = tmp;
3518 spin_unlock(&hugetlb_lock);
3519 }
Michal Hocko08d4a242011-01-13 15:47:26 -08003520out:
3521 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003522}
3523
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524#endif /* CONFIG_SYSCTL */
3525
Alexey Dobriyane1759c22008-10-15 23:50:22 +04003526void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527{
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003528 struct hstate *h;
3529 unsigned long total = 0;
3530
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003531 if (!hugepages_supported())
3532 return;
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003533
3534 for_each_hstate(h) {
3535 unsigned long count = h->nr_huge_pages;
3536
3537 total += (PAGE_SIZE << huge_page_order(h)) * count;
3538
3539 if (h == &default_hstate)
3540 seq_printf(m,
3541 "HugePages_Total: %5lu\n"
3542 "HugePages_Free: %5lu\n"
3543 "HugePages_Rsvd: %5lu\n"
3544 "HugePages_Surp: %5lu\n"
3545 "Hugepagesize: %8lu kB\n",
3546 count,
3547 h->free_huge_pages,
3548 h->resv_huge_pages,
3549 h->surplus_huge_pages,
3550 (PAGE_SIZE << huge_page_order(h)) / 1024);
3551 }
3552
3553 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554}
3555
Joe Perches79815932020-09-16 13:40:43 -07003556int hugetlb_report_node_meminfo(char *buf, int len, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557{
Andi Kleena5516432008-07-23 21:27:41 -07003558 struct hstate *h = &default_hstate;
Joe Perches79815932020-09-16 13:40:43 -07003559
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003560 if (!hugepages_supported())
3561 return 0;
Joe Perches79815932020-09-16 13:40:43 -07003562
3563 return sysfs_emit_at(buf, len,
3564 "Node %d HugePages_Total: %5u\n"
3565 "Node %d HugePages_Free: %5u\n"
3566 "Node %d HugePages_Surp: %5u\n",
3567 nid, h->nr_huge_pages_node[nid],
3568 nid, h->free_huge_pages_node[nid],
3569 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570}
3571
David Rientjes949f7ec2013-04-29 15:07:48 -07003572void hugetlb_show_meminfo(void)
3573{
3574 struct hstate *h;
3575 int nid;
3576
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003577 if (!hugepages_supported())
3578 return;
3579
David Rientjes949f7ec2013-04-29 15:07:48 -07003580 for_each_node_state(nid, N_MEMORY)
3581 for_each_hstate(h)
3582 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3583 nid,
3584 h->nr_huge_pages_node[nid],
3585 h->free_huge_pages_node[nid],
3586 h->surplus_huge_pages_node[nid],
3587 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3588}
3589
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003590void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3591{
3592 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3593 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3594}
3595
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3597unsigned long hugetlb_total_pages(void)
3598{
Wanpeng Lid0028582013-03-22 15:04:40 -07003599 struct hstate *h;
3600 unsigned long nr_total_pages = 0;
3601
3602 for_each_hstate(h)
3603 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3604 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606
Andi Kleena5516432008-07-23 21:27:41 -07003607static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003608{
3609 int ret = -ENOMEM;
3610
3611 spin_lock(&hugetlb_lock);
3612 /*
3613 * When cpuset is configured, it breaks the strict hugetlb page
3614 * reservation as the accounting is done on a global variable. Such
3615 * reservation is completely rubbish in the presence of cpuset because
3616 * the reservation is not checked against page availability for the
3617 * current cpuset. Application can still potentially OOM'ed by kernel
3618 * with lack of free htlb page in cpuset that the task is in.
3619 * Attempt to enforce strict accounting with cpuset is almost
3620 * impossible (or too ugly) because cpuset is too fluid that
3621 * task or memory node can be dynamically moved between cpusets.
3622 *
3623 * The change of semantics for shared hugetlb mapping with cpuset is
3624 * undesirable. However, in order to preserve some of the semantics,
3625 * we fall back to check against current free page availability as
3626 * a best attempt and hopefully to minimize the impact of changing
3627 * semantics that cpuset has.
Muchun Song8ca39e62020-08-11 18:30:32 -07003628 *
3629 * Apart from cpuset, we also have memory policy mechanism that
3630 * also determines from which node the kernel will allocate memory
3631 * in a NUMA system. So similar to cpuset, we also should consider
3632 * the memory policy of the current task. Similar to the description
3633 * above.
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003634 */
3635 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07003636 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003637 goto out;
3638
Muchun Song8ca39e62020-08-11 18:30:32 -07003639 if (delta > allowed_mems_nr(h)) {
Andi Kleena5516432008-07-23 21:27:41 -07003640 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003641 goto out;
3642 }
3643 }
3644
3645 ret = 0;
3646 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07003647 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003648
3649out:
3650 spin_unlock(&hugetlb_lock);
3651 return ret;
3652}
3653
Andy Whitcroft84afd992008-07-23 21:27:32 -07003654static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3655{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003656 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003657
3658 /*
3659 * This new VMA should share its siblings reservation map if present.
3660 * The VMA will only ever have a valid reservation map pointer where
3661 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003662 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07003663 * after this open call completes. It is therefore safe to take a
3664 * new reference here without additional locking.
3665 */
Mike Kravetze1fa3b22021-09-02 14:58:53 -07003666 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
3667 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003668 kref_get(&resv->refs);
Mike Kravetze1fa3b22021-09-02 14:58:53 -07003669 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07003670}
3671
Mel Gormana1e78772008-07-23 21:27:23 -07003672static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3673{
Andi Kleena5516432008-07-23 21:27:41 -07003674 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003675 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07003676 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003677 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003678 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003679
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003680 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3681 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003682
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003683 start = vma_hugecache_offset(h, vma, vma->vm_start);
3684 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003685
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003686 reserve = (end - start) - region_count(resv, start, end);
Mina Almasrye9fe92a2020-04-01 21:11:21 -07003687 hugetlb_cgroup_uncharge_counter(resv, start, end);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003688 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003689 /*
3690 * Decrement reserve counts. The global reserve count may be
3691 * adjusted if the subpool has a minimum size.
3692 */
3693 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3694 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003695 }
Mina Almasrye9fe92a2020-04-01 21:11:21 -07003696
3697 kref_put(&resv->refs, resv_map_release);
Mel Gormana1e78772008-07-23 21:27:23 -07003698}
3699
Dan Williams31383c62017-11-29 16:10:28 -08003700static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3701{
3702 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3703 return -EINVAL;
James Houghton43390f12023-01-04 23:19:10 +00003704
3705 /*
3706 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
3707 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
3708 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
3709 */
3710 if (addr & ~PUD_MASK) {
3711 /*
3712 * hugetlb_vm_op_split is called right before we attempt to
3713 * split the VMA. We will need to unshare PMDs in the old and
3714 * new VMAs, so let's unshare before we split.
3715 */
3716 unsigned long floor = addr & PUD_MASK;
3717 unsigned long ceil = floor + PUD_SIZE;
3718
3719 if (floor >= vma->vm_start && ceil <= vma->vm_end)
3720 hugetlb_unshare_pmds(vma, floor, ceil);
3721 }
3722
Dan Williams31383c62017-11-29 16:10:28 -08003723 return 0;
3724}
3725
Dan Williams05ea8862018-04-05 16:24:25 -07003726static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3727{
3728 struct hstate *hstate = hstate_vma(vma);
3729
3730 return 1UL << huge_page_shift(hstate);
3731}
3732
Linus Torvalds1da177e2005-04-16 15:20:36 -07003733/*
3734 * We cannot handle pagefaults against hugetlb pages at all. They cause
3735 * handle_mm_fault() to try to instantiate regular-sized pages in the
3736 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3737 * this far.
3738 */
Souptick Joarderb3ec9f32018-06-07 17:08:04 -07003739static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740{
3741 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003742 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743}
3744
Jane Chueec36362018-08-02 15:36:05 -07003745/*
3746 * When a new function is introduced to vm_operations_struct and added
3747 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3748 * This is because under System V memory model, mappings created via
3749 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3750 * their original vm_ops are overwritten with shm_vm_ops.
3751 */
Alexey Dobriyanf0f37e2f2009-09-27 22:29:37 +04003752const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003753 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003754 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003755 .close = hugetlb_vm_op_close,
Dan Williams31383c62017-11-29 16:10:28 -08003756 .split = hugetlb_vm_op_split,
Dan Williams05ea8862018-04-05 16:24:25 -07003757 .pagesize = hugetlb_vm_op_pagesize,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758};
3759
David Gibson1e8f8892006-01-06 00:10:44 -08003760static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3761 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003762{
3763 pte_t entry;
3764
David Gibson1e8f8892006-01-06 00:10:44 -08003765 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003766 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3767 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003768 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003769 entry = huge_pte_wrprotect(mk_huge_pte(page,
3770 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003771 }
3772 entry = pte_mkyoung(entry);
3773 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003774 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003775
3776 return entry;
3777}
3778
David Gibson1e8f8892006-01-06 00:10:44 -08003779static void set_huge_ptep_writable(struct vm_area_struct *vma,
3780 unsigned long address, pte_t *ptep)
3781{
3782 pte_t entry;
3783
Gerald Schaefer106c9922013-04-29 15:07:23 -07003784 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003785 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003786 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003787}
3788
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003789bool is_hugetlb_entry_migration(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003790{
3791 swp_entry_t swp;
3792
3793 if (huge_pte_none(pte) || pte_present(pte))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003794 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003795 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07003796 if (is_migration_entry(swp))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003797 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003798 else
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003799 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003800}
3801
Baoquan He3e5c3602020-10-13 16:56:10 -07003802static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003803{
3804 swp_entry_t swp;
3805
3806 if (huge_pte_none(pte) || pte_present(pte))
Baoquan He3e5c3602020-10-13 16:56:10 -07003807 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003808 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07003809 if (is_hwpoison_entry(swp))
Baoquan He3e5c3602020-10-13 16:56:10 -07003810 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003811 else
Baoquan He3e5c3602020-10-13 16:56:10 -07003812 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003813}
David Gibson1e8f8892006-01-06 00:10:44 -08003814
David Gibson63551ae2005-06-21 17:14:44 -07003815int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3816 struct vm_area_struct *vma)
3817{
Mike Kravetz5e415402018-11-16 15:08:04 -08003818 pte_t *src_pte, *dst_pte, entry, dst_entry;
David Gibson63551ae2005-06-21 17:14:44 -07003819 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003820 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003821 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003822 struct hstate *h = hstate_vma(vma);
3823 unsigned long sz = huge_page_size(h);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003824 struct address_space *mapping = vma->vm_file->f_mapping;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003825 struct mmu_notifier_range range;
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003826 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003827
3828 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003829
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003830 if (cow) {
Jérôme Glisse7269f992019-05-13 17:20:53 -07003831 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07003832 vma->vm_start,
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003833 vma->vm_end);
3834 mmu_notifier_invalidate_range_start(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003835 } else {
3836 /*
3837 * For shared mappings i_mmap_rwsem must be held to call
3838 * huge_pte_alloc, otherwise the returned ptep could go
3839 * away if part of a shared pmd and another thread calls
3840 * huge_pmd_unshare.
3841 */
3842 i_mmap_lock_read(mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003843 }
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003844
Andi Kleena5516432008-07-23 21:27:41 -07003845 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003846 spinlock_t *src_ptl, *dst_ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07003847 src_pte = huge_pte_offset(src, addr, sz);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003848 if (!src_pte)
3849 continue;
Peter Xu59caf932021-03-18 17:01:46 +11003850 dst_pte = huge_pte_alloc(dst, vma, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003851 if (!dst_pte) {
3852 ret = -ENOMEM;
3853 break;
3854 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003855
Mike Kravetz5e415402018-11-16 15:08:04 -08003856 /*
3857 * If the pagetables are shared don't copy or take references.
3858 * dst_pte == src_pte is the common case of src/dest sharing.
3859 *
3860 * However, src could have 'unshared' and dst shares with
3861 * another vma. If dst_pte !none, this implies sharing.
3862 * Check here before taking page table lock, and once again
3863 * after taking the lock below.
3864 */
3865 dst_entry = huge_ptep_get(dst_pte);
3866 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
Larry Woodmanc5c99422008-01-24 05:49:25 -08003867 continue;
3868
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003869 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3870 src_ptl = huge_pte_lockptr(h, src, src_pte);
3871 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003872 entry = huge_ptep_get(src_pte);
Mike Kravetz5e415402018-11-16 15:08:04 -08003873 dst_entry = huge_ptep_get(dst_pte);
3874 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3875 /*
3876 * Skip if src entry none. Also, skip in the
3877 * unlikely case dst entry !none as this implies
3878 * sharing with another vma.
3879 */
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003880 ;
3881 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3882 is_hugetlb_entry_hwpoisoned(entry))) {
3883 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3884
3885 if (is_write_migration_entry(swp_entry) && cow) {
3886 /*
3887 * COW mappings require pages in both
3888 * parent and child to be set to read.
3889 */
3890 make_migration_entry_read(&swp_entry);
3891 entry = swp_entry_to_pte(swp_entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07003892 set_huge_swap_pte_at(src, addr, src_pte,
3893 entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003894 }
Punit Agrawale5251fd2017-07-06 15:39:50 -07003895 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003896 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003897 if (cow) {
Jérôme Glisse0f108512017-11-15 17:34:07 -08003898 /*
3899 * No need to notify as we are downgrading page
3900 * table protection not changing it to point
3901 * to a new page.
3902 *
Mike Rapoportad56b732018-03-21 21:22:47 +02003903 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08003904 */
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003905 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003906 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003907 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003908 ptepage = pte_page(entry);
3909 get_page(ptepage);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003910 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003911 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003912 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003913 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003914 spin_unlock(src_ptl);
3915 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003916 }
David Gibson63551ae2005-06-21 17:14:44 -07003917
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003918 if (cow)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003919 mmu_notifier_invalidate_range_end(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003920 else
3921 i_mmap_unlock_read(mapping);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003922
3923 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003924}
3925
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003926void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3927 unsigned long start, unsigned long end,
3928 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003929{
3930 struct mm_struct *mm = vma->vm_mm;
3931 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003932 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003933 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003934 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003935 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003936 struct hstate *h = hstate_vma(vma);
3937 unsigned long sz = huge_page_size(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003938 struct mmu_notifier_range range;
Nadav Amit40bc8312021-11-21 12:40:07 -08003939 bool force_flush = false;
Andi Kleena5516432008-07-23 21:27:41 -07003940
David Gibson63551ae2005-06-21 17:14:44 -07003941 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003942 BUG_ON(start & ~huge_page_mask(h));
3943 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003944
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08003945 /*
3946 * This is a hugetlb vma, all the pte entries should point
3947 * to huge page.
3948 */
Peter Zijlstraed6a7932018-08-31 14:46:08 +02003949 tlb_change_page_size(tlb, sz);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003950 tlb_start_vma(tlb, vma);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07003951
3952 /*
3953 * If sharing possible, alert mmu notifiers of worst case.
3954 */
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07003955 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3956 end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003957 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3958 mmu_notifier_invalidate_range_start(&range);
Hillf Danton569f48b82014-12-10 15:44:41 -08003959 address = start;
Hillf Danton569f48b82014-12-10 15:44:41 -08003960 for (; address < end; address += sz) {
Punit Agrawal7868a202017-07-06 15:39:42 -07003961 ptep = huge_pte_offset(mm, address, sz);
Adam Litke4c887262005-10-29 18:16:46 -07003962 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003963 continue;
3964
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003965 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae204f2020-08-11 18:31:38 -07003966 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003967 spin_unlock(ptl);
Nadav Amit40bc8312021-11-21 12:40:07 -08003968 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
3969 force_flush = true;
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003970 continue;
3971 }
Chen, Kenneth W39dde652006-12-06 20:32:03 -08003972
Hillf Danton66293262012-03-23 15:01:48 -07003973 pte = huge_ptep_get(ptep);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003974 if (huge_pte_none(pte)) {
3975 spin_unlock(ptl);
3976 continue;
3977 }
Hillf Danton66293262012-03-23 15:01:48 -07003978
3979 /*
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003980 * Migrating hugepage or HWPoisoned hugepage is already
3981 * unmapped and its refcount is dropped, so just clear pte here.
Hillf Danton66293262012-03-23 15:01:48 -07003982 */
Naoya Horiguchi9fbc1f62015-02-11 15:25:32 -08003983 if (unlikely(!pte_present(pte))) {
Punit Agrawal9386fac2017-07-06 15:39:46 -07003984 huge_pte_clear(mm, address, ptep, sz);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003985 spin_unlock(ptl);
3986 continue;
Naoya Horiguchi8c4894c2012-12-12 13:52:28 -08003987 }
Hillf Danton66293262012-03-23 15:01:48 -07003988
3989 page = pte_page(pte);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07003990 /*
3991 * If a reference page is supplied, it is because a specific
3992 * page is being unmapped, not a range. Ensure the page we
3993 * are about to unmap is the actual page of interest.
3994 */
3995 if (ref_page) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003996 if (page != ref_page) {
3997 spin_unlock(ptl);
3998 continue;
3999 }
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004000 /*
4001 * Mark the VMA as having unmapped its page so that
4002 * future faults in this VMA will fail rather than
4003 * looking like data was lost
4004 */
4005 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
4006 }
4007
David Gibsonc7546f82005-08-05 11:59:35 -07004008 pte = huge_ptep_get_and_clear(mm, address, ptep);
Aneesh Kumar K.Vb528e4b2016-12-12 16:42:37 -08004009 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
Gerald Schaefer106c9922013-04-29 15:07:23 -07004010 if (huge_pte_dirty(pte))
Ken Chen6649a382007-02-08 14:20:27 -08004011 set_page_dirty(page);
Hillf Danton9e811302012-03-21 16:34:03 -07004012
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08004013 hugetlb_count_sub(pages_per_huge_page(h), mm);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004014 page_remove_rmap(page, true);
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004015
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004016 spin_unlock(ptl);
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -07004017 tlb_remove_page_size(tlb, page, huge_page_size(h));
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07004018 /*
4019 * Bail out after unmapping reference page if supplied
4020 */
4021 if (ref_page)
4022 break;
Chen, Kenneth Wfe1668a2006-10-04 02:15:24 -07004023 }
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004024 mmu_notifier_invalidate_range_end(&range);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004025 tlb_end_vma(tlb, vma);
Nadav Amit40bc8312021-11-21 12:40:07 -08004026
4027 /*
4028 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
4029 * could defer the flush until now, since by holding i_mmap_rwsem we
4030 * guaranteed that the last refernece would not be dropped. But we must
4031 * do the flushing before we return, as otherwise i_mmap_rwsem will be
4032 * dropped and the last reference to the shared PMDs page might be
4033 * dropped as well.
4034 *
4035 * In theory we could defer the freeing of the PMD pages as well, but
4036 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
4037 * detect sharing, so we cannot defer the release of the page either.
4038 * Instead, do flush now.
4039 */
4040 if (force_flush)
4041 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042}
David Gibson63551ae2005-06-21 17:14:44 -07004043
Mel Gormand8333522012-07-31 16:46:20 -07004044void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4045 struct vm_area_struct *vma, unsigned long start,
4046 unsigned long end, struct page *ref_page)
4047{
4048 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4049
4050 /*
4051 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4052 * test will fail on a vma being torn down, and not grab a page table
4053 * on its way out. We're lucky that the flag has such an appropriate
4054 * name, and can in fact be safely cleared here. We could clear it
4055 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004056 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07004057 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004058 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07004059 */
4060 vma->vm_flags &= ~VM_MAYSHARE;
4061}
4062
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004063void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004064 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004065{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004066 struct mm_struct *mm;
4067 struct mmu_gather tlb;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004068 unsigned long tlb_start = start;
4069 unsigned long tlb_end = end;
4070
4071 /*
4072 * If shared PMDs were possibly used within this vma range, adjust
4073 * start/end for worst case tlb flushing.
4074 * Note that we can not be sure if PMDs are shared until we try to
4075 * unmap pages. However, we want to make sure TLB flushing covers
4076 * the largest possible range.
4077 */
4078 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004079
4080 mm = vma->vm_mm;
4081
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004082 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004083 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004084 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004085}
4086
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004087/*
4088 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4089 * mappping it owns the reserve page for. The intention is to unmap the page
4090 * from other VMAs and let the children be SIGKILLed if they are faulting the
4091 * same region.
4092 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004093static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4094 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004095{
Adam Litke75266742008-11-12 13:24:56 -08004096 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004097 struct vm_area_struct *iter_vma;
4098 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004099 pgoff_t pgoff;
4100
4101 /*
4102 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4103 * from page cache lookup which is in HPAGE_SIZE units.
4104 */
Adam Litke75266742008-11-12 13:24:56 -08004105 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07004106 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4107 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05004108 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004109
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08004110 /*
4111 * Take the mapping lock for the duration of the table walk. As
4112 * this mapping should be shared between all the VMAs,
4113 * __unmap_hugepage_range() is called as the lock is already held
4114 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004115 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07004116 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004117 /* Do not unmap the current VMA */
4118 if (iter_vma == vma)
4119 continue;
4120
4121 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07004122 * Shared VMAs have their own reserves and do not affect
4123 * MAP_PRIVATE accounting but it is possible that a shared
4124 * VMA is using the same page so check and skip such VMAs.
4125 */
4126 if (iter_vma->vm_flags & VM_MAYSHARE)
4127 continue;
4128
4129 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004130 * Unmap the page from other VMAs without their own reserves.
4131 * They get marked to be SIGKILLed if they fault in these
4132 * areas. This is because a future no-page fault on this VMA
4133 * could insert a zeroed page instead of the data existing
4134 * from the time of fork. This would look like data corruption
4135 */
4136 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004137 unmap_hugepage_range(iter_vma, address,
4138 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004139 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004140 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004141}
4142
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004143/*
4144 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08004145 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4146 * cannot race with other handlers or page migration.
4147 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004148 */
Souptick Joarder2b740302018-08-23 17:01:36 -07004149static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Huang Ying974e6d62018-08-17 15:45:57 -07004150 unsigned long address, pte_t *ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004151 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08004152{
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004153 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07004154 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08004155 struct page *old_page, *new_page;
Souptick Joarder2b740302018-08-23 17:01:36 -07004156 int outside_reserve = 0;
4157 vm_fault_t ret = 0;
Huang Ying974e6d62018-08-17 15:45:57 -07004158 unsigned long haddr = address & huge_page_mask(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004159 struct mmu_notifier_range range;
David Gibson1e8f8892006-01-06 00:10:44 -08004160
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004161 pte = huge_ptep_get(ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08004162 old_page = pte_page(pte);
4163
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004164retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08004165 /* If no-one else is actually using this page, avoid the copy
4166 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07004167 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07004168 page_move_anon_rmap(old_page, vma);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004169 set_huge_ptep_writable(vma, haddr, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004170 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004171 }
4172
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004173 /*
4174 * If the process that created a MAP_PRIVATE mapping is about to
4175 * perform a COW due to a shared page count, attempt to satisfy
4176 * the allocation without using the existing reserves. The pagecache
4177 * page is used to determine if the reserve at this address was
4178 * consumed or not. If reserves were used, a partial faulted mapping
4179 * at the time of fork() could consume its reserves on COW instead
4180 * of the full address range.
4181 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07004182 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004183 old_page != pagecache_page)
4184 outside_reserve = 1;
4185
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004186 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004187
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004188 /*
4189 * Drop page table lock as buddy allocator may be called. It will
4190 * be acquired again before returning to the caller, as expected.
4191 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004192 spin_unlock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004193 new_page = alloc_huge_page(vma, haddr, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08004194
Adam Litke2fc39ce2007-11-14 16:59:39 -08004195 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004196 /*
4197 * If a process owning a MAP_PRIVATE mapping fails to COW,
4198 * it is due to references held by a child and an insufficient
4199 * huge page pool. To guarantee the original mappers
4200 * reliability, unmap the page from child processes. The child
4201 * may get SIGKILLed if it later faults.
4202 */
4203 if (outside_reserve) {
Mike Kravetzdf73c802020-12-29 15:14:25 -08004204 struct address_space *mapping = vma->vm_file->f_mapping;
4205 pgoff_t idx;
4206 u32 hash;
4207
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004208 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004209 BUG_ON(huge_pte_none(pte));
Mike Kravetzdf73c802020-12-29 15:14:25 -08004210 /*
4211 * Drop hugetlb_fault_mutex and i_mmap_rwsem before
4212 * unmapping. unmapping needs to hold i_mmap_rwsem
4213 * in write mode. Dropping i_mmap_rwsem in read mode
4214 * here is OK as COW mappings do not interact with
4215 * PMD sharing.
4216 *
4217 * Reacquire both after unmap operation.
4218 */
4219 idx = vma_hugecache_offset(h, vma, haddr);
4220 hash = hugetlb_fault_mutex_hash(mapping, idx);
4221 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4222 i_mmap_unlock_read(mapping);
4223
Huang Ying5b7a1d42018-08-17 15:45:53 -07004224 unmap_ref_private(mm, vma, old_page, haddr);
Mike Kravetzdf73c802020-12-29 15:14:25 -08004225
4226 i_mmap_lock_read(mapping);
4227 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004228 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004229 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004230 if (likely(ptep &&
4231 pte_same(huge_ptep_get(ptep), pte)))
4232 goto retry_avoidcopy;
4233 /*
4234 * race occurs while re-acquiring page table
4235 * lock, and our job is done.
4236 */
4237 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004238 }
4239
Souptick Joarder2b740302018-08-23 17:01:36 -07004240 ret = vmf_error(PTR_ERR(new_page));
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004241 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08004242 }
4243
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004244 /*
4245 * When the original hugepage is shared one, it does not have
4246 * anon_vma prepared.
4247 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07004248 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004249 ret = VM_FAULT_OOM;
4250 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07004251 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004252
Huang Ying974e6d62018-08-17 15:45:57 -07004253 copy_user_huge_page(new_page, old_page, address, vma,
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004254 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004255 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004256
Jérôme Glisse7269f992019-05-13 17:20:53 -07004257 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07004258 haddr + huge_page_size(h));
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004259 mmu_notifier_invalidate_range_start(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004260
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004261 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004262 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004263 * before the page tables are altered
4264 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004265 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004266 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07004267 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07004268 ClearPagePrivate(new_page);
4269
David Gibson1e8f8892006-01-06 00:10:44 -08004270 /* Break COW */
Huang Ying5b7a1d42018-08-17 15:45:53 -07004271 huge_ptep_clear_flush(vma, haddr, ptep);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004272 mmu_notifier_invalidate_range(mm, range.start, range.end);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004273 set_huge_pte_at(mm, haddr, ptep,
David Gibson1e8f8892006-01-06 00:10:44 -08004274 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004275 page_remove_rmap(old_page, true);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004276 hugepage_add_new_anon_rmap(new_page, vma, haddr);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004277 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004278 /* Make the old page be freed below */
4279 new_page = old_page;
4280 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004281 spin_unlock(ptl);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004282 mmu_notifier_invalidate_range_end(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004283out_release_all:
Huang Ying5b7a1d42018-08-17 15:45:53 -07004284 restore_reserve_on_error(h, vma, haddr, new_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004285 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004286out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004287 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07004288
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004289 spin_lock(ptl); /* Caller expects lock to be held */
4290 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08004291}
4292
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004293/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07004294static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4295 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004296{
4297 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07004298 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004299
4300 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07004301 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004302
4303 return find_lock_page(mapping, idx);
4304}
4305
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004306/*
4307 * Return whether there is a pagecache page to back given address within VMA.
4308 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4309 */
4310static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004311 struct vm_area_struct *vma, unsigned long address)
4312{
4313 struct address_space *mapping;
4314 pgoff_t idx;
4315 struct page *page;
4316
4317 mapping = vma->vm_file->f_mapping;
4318 idx = vma_hugecache_offset(h, vma, address);
4319
4320 page = find_get_page(mapping, idx);
4321 if (page)
4322 put_page(page);
4323 return page != NULL;
4324}
4325
Mike Kravetzab76ad52015-09-08 15:01:50 -07004326int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4327 pgoff_t idx)
4328{
4329 struct inode *inode = mapping->host;
4330 struct hstate *h = hstate_inode(inode);
4331 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4332
4333 if (err)
4334 return err;
4335 ClearPagePrivate(page);
4336
Mike Kravetz22146c32018-10-26 15:10:58 -07004337 /*
4338 * set page dirty so that it will not be removed from cache/file
4339 * by non-hugetlbfs specific code paths.
4340 */
4341 set_page_dirty(page);
4342
Mike Kravetzab76ad52015-09-08 15:01:50 -07004343 spin_lock(&inode->i_lock);
4344 inode->i_blocks += blocks_per_huge_page(h);
4345 spin_unlock(&inode->i_lock);
4346 return 0;
4347}
4348
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004349static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
4350 struct address_space *mapping,
4351 pgoff_t idx,
4352 unsigned int flags,
4353 unsigned long haddr,
4354 unsigned long reason)
4355{
Greg Kroah-Hartman0b500f52022-11-15 16:46:50 +00004356 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004357 struct vm_fault vmf = {
4358 .vma = vma,
4359 .address = haddr,
4360 .flags = flags,
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004361 /*
4362 * Hard to debug if it ends up being
4363 * used by a callee that assumes
4364 * something about the other
4365 * uninitialized fields... same as in
4366 * memory.c
4367 */
4368 };
4369
4370 /*
Greg Kroah-Hartman0b500f52022-11-15 16:46:50 +00004371 * vma_lock and hugetlb_fault_mutex must be dropped
4372 * before handling userfault. Also mmap_lock will
4373 * be dropped during handling userfault, any vma
4374 * operation should be careful from here.
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004375 */
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004376 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4377 i_mmap_unlock_read(mapping);
Greg Kroah-Hartman0b500f52022-11-15 16:46:50 +00004378 return handle_userfault(&vmf, VM_UFFD_MISSING);
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004379}
4380
Souptick Joarder2b740302018-08-23 17:01:36 -07004381static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4382 struct vm_area_struct *vma,
4383 struct address_space *mapping, pgoff_t idx,
4384 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004385{
Andi Kleena5516432008-07-23 21:27:41 -07004386 struct hstate *h = hstate_vma(vma);
Souptick Joarder2b740302018-08-23 17:01:36 -07004387 vm_fault_t ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08004388 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004389 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07004390 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08004391 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004392 spinlock_t *ptl;
Huang Ying285b8dc2018-06-07 17:08:08 -07004393 unsigned long haddr = address & huge_page_mask(h);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004394 bool new_page = false;
Liu Shixin45c33962022-09-23 12:21:13 +08004395 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
Adam Litke4c887262005-10-29 18:16:46 -07004396
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004397 /*
4398 * Currently, we are forced to kill the process in the event the
4399 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004400 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004401 */
4402 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08004403 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08004404 current->pid);
Liu Shixin45c33962022-09-23 12:21:13 +08004405 goto out;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004406 }
4407
Adam Litke4c887262005-10-29 18:16:46 -07004408 /*
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004409 * We can not race with truncation due to holding i_mmap_rwsem.
4410 * i_size is modified when holding i_mmap_rwsem, so check here
4411 * once for faults beyond end of file.
Adam Litke4c887262005-10-29 18:16:46 -07004412 */
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004413 size = i_size_read(mapping->host) >> huge_page_shift(h);
4414 if (idx >= size)
4415 goto out;
4416
Christoph Lameter6bda6662006-01-06 00:10:49 -08004417retry:
4418 page = find_lock_page(mapping, idx);
4419 if (!page) {
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004420 /* Check for page in userfault range */
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004421 if (userfaultfd_missing(vma)) {
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004422 ret = hugetlb_handle_userfault(vma, mapping, idx,
4423 flags, haddr,
4424 VM_UFFD_MISSING);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004425 goto out;
4426 }
4427
Huang Ying285b8dc2018-06-07 17:08:08 -07004428 page = alloc_huge_page(vma, haddr, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08004429 if (IS_ERR(page)) {
Mike Kravetz4643d672019-08-13 15:38:00 -07004430 /*
4431 * Returning error will result in faulting task being
4432 * sent SIGBUS. The hugetlb fault mutex prevents two
4433 * tasks from racing to fault in the same page which
4434 * could result in false unable to allocate errors.
4435 * Page migration does not take the fault mutex, but
4436 * does a clear then write of pte's under page table
4437 * lock. Page fault code could race with migration,
4438 * notice the clear pte and try to allocate a page
4439 * here. Before returning error, get ptl and make
4440 * sure there really is no pte entry.
4441 */
4442 ptl = huge_pte_lock(h, mm, ptep);
4443 if (!huge_pte_none(huge_ptep_get(ptep))) {
4444 ret = 0;
4445 spin_unlock(ptl);
4446 goto out;
4447 }
4448 spin_unlock(ptl);
Souptick Joarder2b740302018-08-23 17:01:36 -07004449 ret = vmf_error(PTR_ERR(page));
Christoph Lameter6bda6662006-01-06 00:10:49 -08004450 goto out;
4451 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004452 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004453 __SetPageUptodate(page);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004454 new_page = true;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004455
Mel Gormanf83a2752009-05-28 14:34:40 -07004456 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07004457 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004458 if (err) {
4459 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004460 if (err == -EEXIST)
4461 goto retry;
4462 goto out;
4463 }
Mel Gorman23be7462010-04-23 13:17:56 -04004464 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08004465 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004466 if (unlikely(anon_vma_prepare(vma))) {
4467 ret = VM_FAULT_OOM;
4468 goto backout_unlocked;
4469 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08004470 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04004471 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004472 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004473 /*
4474 * If memory error occurs between mmap() and fault, some process
4475 * don't have hwpoisoned swap entry for errored virtual address.
4476 * So we need to block hugepage fault by PG_hwpoison bit check.
4477 */
4478 if (unlikely(PageHWPoison(page))) {
Miaohe Lind3e43af2021-01-12 15:49:24 -08004479 ret = VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004480 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004481 goto backout_unlocked;
4482 }
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004483
4484 /* Check for page in userfault range. */
4485 if (userfaultfd_minor(vma)) {
4486 unlock_page(page);
Peter Xu26d1bb82021-03-22 13:51:32 -04004487 put_page(page);
Axel Rasmussen4d3dd332021-03-18 17:01:50 +11004488 ret = hugetlb_handle_userfault(vma, mapping, idx,
4489 flags, haddr,
4490 VM_UFFD_MINOR);
4491 goto out;
4492 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08004493 }
David Gibson1e8f8892006-01-06 00:10:44 -08004494
Andy Whitcroft57303d82008-08-12 15:08:47 -07004495 /*
4496 * If we are going to COW a private mapping later, we examine the
4497 * pending reservations for this page now. This will ensure that
4498 * any allocations necessary to record that reservation occur outside
4499 * the spinlock.
4500 */
Mike Kravetz5e911372015-09-08 15:01:28 -07004501 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004502 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004503 ret = VM_FAULT_OOM;
4504 goto backout_unlocked;
4505 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004506 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004507 vma_end_reservation(h, vma, haddr);
Mike Kravetz5e911372015-09-08 15:01:28 -07004508 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07004509
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08004510 ptl = huge_pte_lock(h, mm, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004511 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004512 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07004513 goto backout;
4514
Joonsoo Kim07443a82013-09-11 14:21:58 -07004515 if (anon_rmap) {
4516 ClearPagePrivate(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004517 hugepage_add_new_anon_rmap(page, vma, haddr);
Choi Gi-yongac714902014-04-07 15:37:36 -07004518 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08004519 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08004520 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4521 && (vma->vm_flags & VM_SHARED)));
Huang Ying285b8dc2018-06-07 17:08:08 -07004522 set_huge_pte_at(mm, haddr, ptep, new_pte);
David Gibson1e8f8892006-01-06 00:10:44 -08004523
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08004524 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01004525 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08004526 /* Optimization, do the COW without a second fault */
Huang Ying974e6d62018-08-17 15:45:57 -07004527 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08004528 }
4529
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004530 spin_unlock(ptl);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004531
4532 /*
4533 * Only make newly allocated pages active. Existing pages found
4534 * in the pagecache could be !page_huge_active() if they have been
4535 * isolated for migration.
4536 */
4537 if (new_page)
4538 set_page_huge_active(page);
4539
Adam Litke4c887262005-10-29 18:16:46 -07004540 unlock_page(page);
4541out:
Liu Shixin45c33962022-09-23 12:21:13 +08004542 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4543 i_mmap_unlock_read(mapping);
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004544 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07004545
4546backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004547 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07004548backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07004549 unlock_page(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004550 restore_reserve_on_error(h, vma, haddr, page);
Adam Litke4c887262005-10-29 18:16:46 -07004551 put_page(page);
4552 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004553}
4554
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004555#ifdef CONFIG_SMP
Wei Yang188b04a2019-11-30 17:57:02 -08004556u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004557{
4558 unsigned long key[2];
4559 u32 hash;
4560
Mike Kravetz1b426ba2019-05-13 17:19:41 -07004561 key[0] = (unsigned long) mapping;
4562 key[1] = idx;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004563
Mike Kravetz55254632019-11-30 17:56:30 -08004564 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004565
4566 return hash & (num_fault_mutexes - 1);
4567}
4568#else
4569/*
4570 * For uniprocesor systems we always use a single mutex, so just
4571 * return 0 and avoid the hashing overhead.
4572 */
Wei Yang188b04a2019-11-30 17:57:02 -08004573u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004574{
4575 return 0;
4576}
4577#endif
4578
Souptick Joarder2b740302018-08-23 17:01:36 -07004579vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01004580 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08004581{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004582 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004583 spinlock_t *ptl;
Souptick Joarder2b740302018-08-23 17:01:36 -07004584 vm_fault_t ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004585 u32 hash;
4586 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004587 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07004588 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07004589 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004590 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004591 int need_wait_lock = 0;
Huang Ying285b8dc2018-06-07 17:08:08 -07004592 unsigned long haddr = address & huge_page_mask(h);
Adam Litke86e52162006-01-06 00:10:43 -08004593
Huang Ying285b8dc2018-06-07 17:08:08 -07004594 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004595 if (ptep) {
Mike Kravetzc0d03812020-04-01 21:11:05 -07004596 /*
4597 * Since we hold no locks, ptep could be stale. That is
4598 * OK as we are only making decisions based on content and
4599 * not actually modifying content here.
4600 */
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004601 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09004602 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004603 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09004604 return 0;
4605 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07004606 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004607 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004608 }
4609
Mike Kravetzc0d03812020-04-01 21:11:05 -07004610 /*
4611 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004612 * until finished with ptep. This serves two purposes:
4613 * 1) It prevents huge_pmd_unshare from being called elsewhere
4614 * and making the ptep no longer valid.
4615 * 2) It synchronizes us with i_size modifications during truncation.
Mike Kravetzc0d03812020-04-01 21:11:05 -07004616 *
4617 * ptep could have already be assigned via huge_pte_offset. That
4618 * is OK, as huge_pte_alloc will return the same value unless
4619 * something has changed.
4620 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004621 mapping = vma->vm_file->f_mapping;
Mike Kravetzc0d03812020-04-01 21:11:05 -07004622 i_mmap_lock_read(mapping);
Peter Xu59caf932021-03-18 17:01:46 +11004623 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
Mike Kravetzc0d03812020-04-01 21:11:05 -07004624 if (!ptep) {
4625 i_mmap_unlock_read(mapping);
4626 return VM_FAULT_OOM;
4627 }
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004628
David Gibson3935baa2006-03-22 00:08:53 -08004629 /*
4630 * Serialize hugepage allocation and instantiation, so that we don't
4631 * get spurious allocation failures if two CPUs race to instantiate
4632 * the same page in the page cache.
4633 */
Mike Kravetzc0d03812020-04-01 21:11:05 -07004634 idx = vma_hugecache_offset(h, vma, haddr);
Wei Yang188b04a2019-11-30 17:57:02 -08004635 hash = hugetlb_fault_mutex_hash(mapping, idx);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07004636 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004637
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004638 entry = huge_ptep_get(ptep);
Liu Shixin45c33962022-09-23 12:21:13 +08004639 if (huge_pte_none(entry))
4640 /*
4641 * hugetlb_no_page will drop vma lock and hugetlb fault
4642 * mutex internally, which make us return immediately.
4643 */
4644 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
Adam Litke86e52162006-01-06 00:10:43 -08004645
Nick Piggin83c54072007-07-19 01:47:05 -07004646 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004647
Andy Whitcroft57303d82008-08-12 15:08:47 -07004648 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004649 * entry could be a migration/hwpoison entry at this point, so this
4650 * check prevents the kernel from going below assuming that we have
Ethon Paul7c8de352020-06-04 16:49:07 -07004651 * an active hugepage in pagecache. This goto expects the 2nd page
4652 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4653 * properly handle it.
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004654 */
4655 if (!pte_present(entry))
4656 goto out_mutex;
4657
4658 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07004659 * If we are going to COW the mapping later, we examine the pending
4660 * reservations for this page now. This will ensure that any
4661 * allocations necessary to record that reservation occur outside the
4662 * spinlock. For private mappings, we also lookup the pagecache
4663 * page now as it is used to determine if a reservation has been
4664 * consumed.
4665 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07004666 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004667 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004668 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07004669 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07004670 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004671 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004672 vma_end_reservation(h, vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004673
Mel Gormanf83a2752009-05-28 14:34:40 -07004674 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07004675 pagecache_page = hugetlbfs_pagecache_page(h,
Huang Ying285b8dc2018-06-07 17:08:08 -07004676 vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004677 }
4678
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004679 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004680
David Gibson1e8f8892006-01-06 00:10:44 -08004681 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07004682 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004683 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07004684
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004685 /*
4686 * hugetlb_cow() requires page locks of pte_page(entry) and
4687 * pagecache_page, so here we need take the former one
4688 * when page != pagecache_page or !pagecache_page.
4689 */
4690 page = pte_page(entry);
4691 if (page != pagecache_page)
4692 if (!trylock_page(page)) {
4693 need_wait_lock = 1;
4694 goto out_ptl;
4695 }
4696
4697 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07004698
Hugh Dickins788c7df2009-06-23 13:49:05 +01004699 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07004700 if (!huge_pte_write(entry)) {
Huang Ying974e6d62018-08-17 15:45:57 -07004701 ret = hugetlb_cow(mm, vma, address, ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004702 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004703 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07004704 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07004705 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07004706 }
4707 entry = pte_mkyoung(entry);
Huang Ying285b8dc2018-06-07 17:08:08 -07004708 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
Hugh Dickins788c7df2009-06-23 13:49:05 +01004709 flags & FAULT_FLAG_WRITE))
Huang Ying285b8dc2018-06-07 17:08:08 -07004710 update_mmu_cache(vma, haddr, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004711out_put_page:
4712 if (page != pagecache_page)
4713 unlock_page(page);
4714 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004715out_ptl:
4716 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004717
4718 if (pagecache_page) {
4719 unlock_page(pagecache_page);
4720 put_page(pagecache_page);
4721 }
David Gibsonb4d1d992008-10-15 22:01:11 -07004722out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07004723 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004724 i_mmap_unlock_read(mapping);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004725 /*
4726 * Generally it's safe to hold refcount during waiting page lock. But
4727 * here we just wait to defer the next page fault to avoid busy loop and
4728 * the page is not used after unlocked before returning from the current
4729 * page fault. So we are safe from accessing freed page, even if we wait
4730 * here without taking refcount.
4731 */
4732 if (need_wait_lock)
4733 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08004734 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08004735}
4736
Axel Rasmussene6bf0762021-03-18 17:01:50 +11004737#ifdef CONFIG_USERFAULTFD
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004738/*
4739 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4740 * modifications for huge pages.
4741 */
4742int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4743 pte_t *dst_pte,
4744 struct vm_area_struct *dst_vma,
4745 unsigned long dst_addr,
4746 unsigned long src_addr,
Axel Rasmussen4a5cf922021-03-18 17:01:50 +11004747 enum mcopy_atomic_mode mode,
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004748 struct page **pagep)
4749{
Axel Rasmussen4a5cf922021-03-18 17:01:50 +11004750 bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE);
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004751 struct address_space *mapping;
4752 pgoff_t idx;
4753 unsigned long size;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004754 int vm_shared = dst_vma->vm_flags & VM_SHARED;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004755 struct hstate *h = hstate_vma(dst_vma);
4756 pte_t _dst_pte;
4757 spinlock_t *ptl;
4758 int ret;
4759 struct page *page;
Axel Rasmussen4a5cf922021-03-18 17:01:50 +11004760 int writable;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004761
Axel Rasmussen4a5cf922021-03-18 17:01:50 +11004762 mapping = dst_vma->vm_file->f_mapping;
4763 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4764
4765 if (is_continue) {
4766 ret = -EFAULT;
4767 page = find_lock_page(mapping, idx);
4768 if (!page)
4769 goto out;
4770 } else if (!*pagep) {
Mina Almasry2eb4ec92021-06-04 20:01:36 -07004771 /* If a page already exists, then it's UFFDIO_COPY for
4772 * a non-missing case. Return -EEXIST.
4773 */
4774 if (vm_shared &&
4775 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
4776 ret = -EEXIST;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004777 goto out;
Mina Almasry2eb4ec92021-06-04 20:01:36 -07004778 }
4779
4780 page = alloc_huge_page(dst_vma, dst_addr, 0);
4781 if (IS_ERR(page)) {
4782 ret = -ENOMEM;
4783 goto out;
4784 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004785
4786 ret = copy_huge_page_from_user(page,
4787 (const void __user *) src_addr,
Mike Kravetz810a56b2017-02-22 15:42:58 -08004788 pages_per_huge_page(h), false);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004789
Michel Lespinassec1e8d7c2020-06-08 21:33:54 -07004790 /* fallback to copy_from_user outside mmap_lock */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004791 if (unlikely(ret)) {
Andrea Arcangeli9e368252018-11-30 14:09:25 -08004792 ret = -ENOENT;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004793 *pagep = page;
4794 /* don't free the page */
4795 goto out;
4796 }
4797 } else {
4798 page = *pagep;
4799 *pagep = NULL;
4800 }
4801
4802 /*
4803 * The memory barrier inside __SetPageUptodate makes sure that
4804 * preceding stores to the page contents become visible before
4805 * the set_pte_at() write.
4806 */
4807 __SetPageUptodate(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004808
Axel Rasmussen4a5cf922021-03-18 17:01:50 +11004809 /* Add shared, newly allocated pages to the page cache. */
4810 if (vm_shared && !is_continue) {
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004811 size = i_size_read(mapping->host) >> huge_page_shift(h);
4812 ret = -EFAULT;
4813 if (idx >= size)
4814 goto out_release_nounlock;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004815
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004816 /*
4817 * Serialization between remove_inode_hugepages() and
4818 * huge_add_to_page_cache() below happens through the
4819 * hugetlb_fault_mutex_table that here must be hold by
4820 * the caller.
4821 */
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004822 ret = huge_add_to_page_cache(page, mapping, idx);
4823 if (ret)
4824 goto out_release_nounlock;
4825 }
4826
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004827 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4828 spin_lock(ptl);
4829
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004830 /*
4831 * Recheck the i_size after holding PT lock to make sure not
4832 * to leave any page mapped (as page_mapped()) beyond the end
4833 * of the i_size (remove_inode_hugepages() is strict about
4834 * enforcing that). If we bail out here, we'll also leave a
4835 * page in the radix tree in the vm_shared case beyond the end
4836 * of the i_size, but remove_inode_hugepages() will take care
4837 * of it as soon as we drop the hugetlb_fault_mutex_table.
4838 */
4839 size = i_size_read(mapping->host) >> huge_page_shift(h);
4840 ret = -EFAULT;
4841 if (idx >= size)
4842 goto out_release_unlock;
4843
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004844 ret = -EEXIST;
4845 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4846 goto out_release_unlock;
4847
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004848 if (vm_shared) {
4849 page_dup_rmap(page, true);
4850 } else {
4851 ClearPagePrivate(page);
4852 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4853 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004854
Axel Rasmussen4a5cf922021-03-18 17:01:50 +11004855 /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */
4856 if (is_continue && !vm_shared)
4857 writable = 0;
4858 else
4859 writable = dst_vma->vm_flags & VM_WRITE;
4860
4861 _dst_pte = make_huge_pte(dst_vma, page, writable);
4862 if (writable)
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004863 _dst_pte = huge_pte_mkdirty(_dst_pte);
4864 _dst_pte = pte_mkyoung(_dst_pte);
4865
4866 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4867
4868 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4869 dst_vma->vm_flags & VM_WRITE);
4870 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4871
4872 /* No need to invalidate - it was non-present before */
4873 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4874
4875 spin_unlock(ptl);
Axel Rasmussen4a5cf922021-03-18 17:01:50 +11004876 if (!is_continue)
4877 set_page_huge_active(page);
4878 if (vm_shared || is_continue)
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004879 unlock_page(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004880 ret = 0;
4881out:
4882 return ret;
4883out_release_unlock:
4884 spin_unlock(ptl);
Axel Rasmussen4a5cf922021-03-18 17:01:50 +11004885 if (vm_shared || is_continue)
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004886 unlock_page(page);
Andrea Arcangeli5af10df2017-08-10 15:23:38 -07004887out_release_nounlock:
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004888 put_page(page);
4889 goto out;
4890}
Axel Rasmussene6bf0762021-03-18 17:01:50 +11004891#endif /* CONFIG_USERFAULTFD */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004892
Michel Lespinasse28a35712013-02-22 16:35:55 -08004893long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4894 struct page **pages, struct vm_area_struct **vmas,
4895 unsigned long *position, unsigned long *nr_pages,
Peter Xu4f6da932020-04-01 21:07:58 -07004896 long i, unsigned int flags, int *locked)
David Gibson63551ae2005-06-21 17:14:44 -07004897{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004898 unsigned long pfn_offset;
4899 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08004900 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07004901 struct hstate *h = hstate_vma(vma);
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004902 int err = -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07004903
David Gibson63551ae2005-06-21 17:14:44 -07004904 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07004905 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004906 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004907 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07004908 struct page *page;
4909
4910 /*
David Rientjes02057962015-04-14 15:48:24 -07004911 * If we have a pending SIGKILL, don't keep faulting pages and
4912 * potentially allocating memory.
4913 */
Davidlohr Buesofa45f112019-01-03 15:28:55 -08004914 if (fatal_signal_pending(current)) {
David Rientjes02057962015-04-14 15:48:24 -07004915 remainder = 0;
4916 break;
4917 }
4918
4919 /*
Adam Litke4c887262005-10-29 18:16:46 -07004920 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004921 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07004922 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004923 *
4924 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07004925 */
Punit Agrawal7868a202017-07-06 15:39:42 -07004926 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4927 huge_page_size(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004928 if (pte)
4929 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004930 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07004931
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004932 /*
4933 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004934 * an error where there's an empty slot with no huge pagecache
4935 * to back it. This way, we avoid allocating a hugepage, and
4936 * the sparse dumpfile avoids allocating disk blocks, but its
4937 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004938 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004939 if (absent && (flags & FOLL_DUMP) &&
4940 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004941 if (pte)
4942 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004943 remainder = 0;
4944 break;
4945 }
4946
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07004947 /*
4948 * We need call hugetlb_fault for both hugepages under migration
4949 * (in which case hugetlb_fault waits for the migration,) and
4950 * hwpoisoned hugepages (in which case we need to prevent the
4951 * caller from accessing to them.) In order to do this, we use
4952 * here is_swap_pte instead of is_hugetlb_entry_migration and
4953 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4954 * both cases, and because we can't follow correct pages
4955 * directly from any kind of swap entries.
4956 */
4957 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07004958 ((flags & FOLL_WRITE) &&
4959 !huge_pte_write(huge_ptep_get(pte)))) {
Souptick Joarder2b740302018-08-23 17:01:36 -07004960 vm_fault_t ret;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004961 unsigned int fault_flags = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004962
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004963 if (pte)
4964 spin_unlock(ptl);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004965 if (flags & FOLL_WRITE)
4966 fault_flags |= FAULT_FLAG_WRITE;
Peter Xu4f6da932020-04-01 21:07:58 -07004967 if (locked)
Peter Xu71335f32020-04-01 21:08:53 -07004968 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4969 FAULT_FLAG_KILLABLE;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004970 if (flags & FOLL_NOWAIT)
4971 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4972 FAULT_FLAG_RETRY_NOWAIT;
4973 if (flags & FOLL_TRIED) {
Peter Xu4426e942020-04-01 21:08:49 -07004974 /*
4975 * Note: FAULT_FLAG_ALLOW_RETRY and
4976 * FAULT_FLAG_TRIED can co-exist
4977 */
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004978 fault_flags |= FAULT_FLAG_TRIED;
4979 }
4980 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4981 if (ret & VM_FAULT_ERROR) {
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004982 err = vm_fault_to_errno(ret, flags);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004983 remainder = 0;
4984 break;
4985 }
4986 if (ret & VM_FAULT_RETRY) {
Peter Xu4f6da932020-04-01 21:07:58 -07004987 if (locked &&
Andrea Arcangeli1ac25012019-02-01 14:20:16 -08004988 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
Peter Xu4f6da932020-04-01 21:07:58 -07004989 *locked = 0;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004990 *nr_pages = 0;
4991 /*
4992 * VM_FAULT_RETRY must not return an
4993 * error, it will return zero
4994 * instead.
4995 *
4996 * No need to update "position" as the
4997 * caller will not check it after
4998 * *nr_pages is set to 0.
4999 */
5000 return i;
5001 }
5002 continue;
Adam Litke4c887262005-10-29 18:16:46 -07005003 }
David Gibson63551ae2005-06-21 17:14:44 -07005004
Andi Kleena5516432008-07-23 21:27:41 -07005005 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07005006 page = pte_page(huge_ptep_get(pte));
Linus Torvalds8fde12c2019-04-11 10:49:19 -07005007
5008 /*
Zhigang Luacbfb082019-11-30 17:57:06 -08005009 * If subpage information not requested, update counters
5010 * and skip the same_page loop below.
5011 */
5012 if (!pages && !vmas && !pfn_offset &&
5013 (vaddr + huge_page_size(h) < vma->vm_end) &&
5014 (remainder >= pages_per_huge_page(h))) {
5015 vaddr += huge_page_size(h);
5016 remainder -= pages_per_huge_page(h);
5017 i += pages_per_huge_page(h);
5018 spin_unlock(ptl);
5019 continue;
5020 }
5021
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08005022same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08005023 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07005024 pages[i] = mem_map_offset(page, pfn_offset);
John Hubbard3faa52c2020-04-01 21:05:29 -07005025 /*
5026 * try_grab_page() should always succeed here, because:
5027 * a) we hold the ptl lock, and b) we've just checked
5028 * that the huge page is present in the page tables. If
5029 * the huge page is present, then the tail pages must
5030 * also be present. The ptl prevents the head page and
5031 * tail pages from being rearranged in any way. So this
5032 * page must be available at this point, unless the page
5033 * refcount overflowed:
5034 */
5035 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
5036 spin_unlock(ptl);
5037 remainder = 0;
5038 err = -ENOMEM;
5039 break;
5040 }
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08005041 }
David Gibson63551ae2005-06-21 17:14:44 -07005042
5043 if (vmas)
5044 vmas[i] = vma;
5045
5046 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08005047 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07005048 --remainder;
5049 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08005050 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07005051 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08005052 /*
5053 * We use pfn_offset to avoid touching the pageframes
5054 * of this compound page.
5055 */
5056 goto same_page;
5057 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005058 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07005059 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08005060 *nr_pages = remainder;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08005061 /*
5062 * setting position is actually required only if remainder is
5063 * not zero but it's faster not to add a "if (remainder)"
5064 * branch.
5065 */
David Gibson63551ae2005-06-21 17:14:44 -07005066 *position = vaddr;
5067
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07005068 return i ? i : err;
David Gibson63551ae2005-06-21 17:14:44 -07005069}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005070
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005071unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005072 unsigned long address, unsigned long end, pgprot_t newprot)
5073{
5074 struct mm_struct *mm = vma->vm_mm;
5075 unsigned long start = address;
5076 pte_t *ptep;
5077 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07005078 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005079 unsigned long pages = 0;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005080 bool shared_pmd = false;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005081 struct mmu_notifier_range range;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005082
5083 /*
5084 * In the case of shared PMDs, the area to flush could be beyond
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005085 * start/end. Set range.start/range.end to cover the maximum possible
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005086 * range if PMD sharing is possible.
5087 */
Jérôme Glisse7269f992019-05-13 17:20:53 -07005088 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5089 0, vma, mm, start, end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005090 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005091
5092 BUG_ON(address >= end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005093 flush_cache_range(vma, range.start, range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005094
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005095 mmu_notifier_invalidate_range_start(&range);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005096 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07005097 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005098 spinlock_t *ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07005099 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005100 if (!ptep)
5101 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005102 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae204f2020-08-11 18:31:38 -07005103 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005104 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005105 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005106 shared_pmd = true;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08005107 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005108 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005109 pte = huge_ptep_get(ptep);
5110 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5111 spin_unlock(ptl);
5112 continue;
5113 }
5114 if (unlikely(is_hugetlb_entry_migration(pte))) {
5115 swp_entry_t entry = pte_to_swp_entry(pte);
5116
5117 if (is_write_migration_entry(entry)) {
5118 pte_t newpte;
5119
5120 make_migration_entry_read(&entry);
5121 newpte = swp_entry_to_pte(entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07005122 set_huge_swap_pte_at(mm, address, ptep,
5123 newpte, huge_page_size(h));
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005124 pages++;
5125 }
5126 spin_unlock(ptl);
5127 continue;
5128 }
5129 if (!huge_pte_none(pte)) {
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005130 pte_t old_pte;
5131
5132 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5133 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08005134 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005135 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005136 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005137 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005138 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005139 }
Mel Gormand8333522012-07-31 16:46:20 -07005140 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005141 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07005142 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005143 * once we release i_mmap_rwsem, another task can do the final put_page
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005144 * and that page table be reused and filled with junk. If we actually
5145 * did unshare a page of pmds, flush the range corresponding to the pud.
Mel Gormand8333522012-07-31 16:46:20 -07005146 */
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005147 if (shared_pmd)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005148 flush_hugetlb_tlb_range(vma, range.start, range.end);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005149 else
5150 flush_hugetlb_tlb_range(vma, start, end);
Jérôme Glisse0f108512017-11-15 17:34:07 -08005151 /*
5152 * No need to call mmu_notifier_invalidate_range() we are downgrading
5153 * page table protection not changing it to point to a new page.
5154 *
Mike Rapoportad56b732018-03-21 21:22:47 +02005155 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08005156 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005157 i_mmap_unlock_write(vma->vm_file->f_mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005158 mmu_notifier_invalidate_range_end(&range);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005159
5160 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005161}
5162
Mel Gormana1e78772008-07-23 21:27:23 -07005163int hugetlb_reserve_pages(struct inode *inode,
5164 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00005165 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005166 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07005167{
Mina Almasry0db9d742020-04-01 21:11:25 -07005168 long ret, chg, add = -1;
Andi Kleena5516432008-07-23 21:27:41 -07005169 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07005170 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005171 struct resv_map *resv_map;
Mina Almasry075a61d2020-04-01 21:11:28 -07005172 struct hugetlb_cgroup *h_cg = NULL;
Mina Almasry0db9d742020-04-01 21:11:25 -07005173 long gbl_reserve, regions_needed = 0;
Adam Litkee4e574b2007-10-16 01:26:19 -07005174
Mike Kravetz63489f82018-03-22 16:17:13 -07005175 /* This should never happen */
5176 if (from > to) {
5177 VM_WARN(1, "%s called with a negative range\n", __func__);
5178 return -EINVAL;
5179 }
5180
Mel Gormana1e78772008-07-23 21:27:23 -07005181 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00005182 * Only apply hugepage reservation if asked. At fault time, an
5183 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07005184 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00005185 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005186 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00005187 return 0;
5188
5189 /*
Mel Gormana1e78772008-07-23 21:27:23 -07005190 * Shared mappings base their reservation on the number of pages that
5191 * are already allocated on behalf of the file. Private mappings need
5192 * to reserve the full area even if read-only as mprotect() may be
5193 * called to make the mapping read-write. Assume !vma is a shm mapping
5194 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005195 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzf27a5132019-05-13 17:22:55 -07005196 /*
5197 * resv_map can not be NULL as hugetlb_reserve_pages is only
5198 * called for inodes for which resv_maps were created (see
5199 * hugetlbfs_get_inode).
5200 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005201 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005202
Mina Almasry0db9d742020-04-01 21:11:25 -07005203 chg = region_chg(resv_map, from, to, &regions_needed);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005204
5205 } else {
Mina Almasrye9fe92a2020-04-01 21:11:21 -07005206 /* Private mapping. */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005207 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00005208 if (!resv_map)
5209 return -ENOMEM;
5210
Mel Gorman17c9d122009-02-11 16:34:16 +00005211 chg = to - from;
5212
Mel Gorman5a6fe122009-02-10 14:02:27 +00005213 set_vma_resv_map(vma, resv_map);
5214 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5215 }
5216
Dave Hansenc50ac052012-05-29 15:06:46 -07005217 if (chg < 0) {
5218 ret = chg;
5219 goto out_err;
5220 }
Mel Gorman17c9d122009-02-11 16:34:16 +00005221
Mina Almasry075a61d2020-04-01 21:11:28 -07005222 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5223 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5224
5225 if (ret < 0) {
5226 ret = -ENOMEM;
5227 goto out_err;
5228 }
5229
5230 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5231 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5232 * of the resv_map.
5233 */
5234 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5235 }
5236
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005237 /*
5238 * There must be enough pages in the subpool for the mapping. If
5239 * the subpool has a minimum size, there may be some global
5240 * reservations already in place (gbl_reserve).
5241 */
5242 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5243 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07005244 ret = -ENOSPC;
Mina Almasry075a61d2020-04-01 21:11:28 -07005245 goto out_uncharge_cgroup;
Dave Hansenc50ac052012-05-29 15:06:46 -07005246 }
Mel Gorman17c9d122009-02-11 16:34:16 +00005247
5248 /*
5249 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07005250 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00005251 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005252 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00005253 if (ret < 0) {
Mina Almasry075a61d2020-04-01 21:11:28 -07005254 goto out_put_pages;
Mel Gorman17c9d122009-02-11 16:34:16 +00005255 }
5256
5257 /*
5258 * Account for the reservations made. Shared mappings record regions
5259 * that have reservations as they are shared by multiple VMAs.
5260 * When the last VMA disappears, the region map says how much
5261 * the reservation was and the page cache tells how much of
5262 * the reservation was consumed. Private mappings are per-VMA and
5263 * only the consumed reservations are tracked. When the VMA
5264 * disappears, the original reservation is the VMA size and the
5265 * consumed reservations are stored in the map. Hence, nothing
5266 * else has to be done for private mappings here
5267 */
Mike Kravetz33039672015-06-24 16:57:58 -07005268 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07005269 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005270
Mina Almasry0db9d742020-04-01 21:11:25 -07005271 if (unlikely(add < 0)) {
5272 hugetlb_acct_memory(h, -gbl_reserve);
Dan Carpenter9b52a372020-12-14 19:12:11 -08005273 ret = add;
Mina Almasry075a61d2020-04-01 21:11:28 -07005274 goto out_put_pages;
Mina Almasry0db9d742020-04-01 21:11:25 -07005275 } else if (unlikely(chg > add)) {
Mike Kravetz33039672015-06-24 16:57:58 -07005276 /*
5277 * pages in this range were added to the reserve
5278 * map between region_chg and region_add. This
5279 * indicates a race with alloc_huge_page. Adjust
5280 * the subpool and reserve counts modified above
5281 * based on the difference.
5282 */
5283 long rsv_adjust;
5284
Miaohe Linfe03ccc2021-03-24 21:37:17 -07005285 /*
5286 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
5287 * reference to h_cg->css. See comment below for detail.
5288 */
Mina Almasry075a61d2020-04-01 21:11:28 -07005289 hugetlb_cgroup_uncharge_cgroup_rsvd(
5290 hstate_index(h),
5291 (chg - add) * pages_per_huge_page(h), h_cg);
5292
Mike Kravetz33039672015-06-24 16:57:58 -07005293 rsv_adjust = hugepage_subpool_put_pages(spool,
5294 chg - add);
5295 hugetlb_acct_memory(h, -rsv_adjust);
Miaohe Linfe03ccc2021-03-24 21:37:17 -07005296 } else if (h_cg) {
5297 /*
5298 * The file_regions will hold their own reference to
5299 * h_cg->css. So we should release the reference held
5300 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
5301 * done.
5302 */
5303 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005304 }
5305 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005306 return 0;
Mina Almasry075a61d2020-04-01 21:11:28 -07005307out_put_pages:
5308 /* put back original number of pages, chg */
5309 (void)hugepage_subpool_put_pages(spool, chg);
5310out_uncharge_cgroup:
5311 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5312 chg * pages_per_huge_page(h), h_cg);
Dave Hansenc50ac052012-05-29 15:06:46 -07005313out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07005314 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mina Almasry0db9d742020-04-01 21:11:25 -07005315 /* Only call region_abort if the region_chg succeeded but the
5316 * region_add failed or didn't run.
5317 */
5318 if (chg >= 0 && add < 0)
5319 region_abort(resv_map, from, to, regions_needed);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07005320 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5321 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07005322 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005323}
5324
Mike Kravetzb5cec282015-09-08 15:01:41 -07005325long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5326 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005327{
Andi Kleena5516432008-07-23 21:27:41 -07005328 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005329 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005330 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07005331 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005332 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08005333
Mike Kravetzf27a5132019-05-13 17:22:55 -07005334 /*
5335 * Since this routine can be called in the evict inode path for all
5336 * hugetlbfs inodes, resv_map could be NULL.
5337 */
Mike Kravetzb5cec282015-09-08 15:01:41 -07005338 if (resv_map) {
5339 chg = region_del(resv_map, start, end);
5340 /*
5341 * region_del() can fail in the rare case where a region
5342 * must be split and another region descriptor can not be
5343 * allocated. If end == LONG_MAX, it will not fail.
5344 */
5345 if (chg < 0)
5346 return chg;
5347 }
5348
Ken Chen45c682a2007-11-14 16:59:44 -08005349 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07005350 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08005351 spin_unlock(&inode->i_lock);
5352
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005353 /*
5354 * If the subpool has a minimum size, the number of global
5355 * reservations to be released may be adjusted.
5356 */
5357 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5358 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07005359
5360 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005361}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09005362
Steve Capper3212b532013-04-23 12:35:02 +01005363#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5364static unsigned long page_table_shareable(struct vm_area_struct *svma,
5365 struct vm_area_struct *vma,
5366 unsigned long addr, pgoff_t idx)
5367{
5368 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5369 svma->vm_start;
5370 unsigned long sbase = saddr & PUD_MASK;
5371 unsigned long s_end = sbase + PUD_SIZE;
5372
5373 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08005374 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5375 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01005376
5377 /*
5378 * match the virtual addresses, permission and the alignment of the
5379 * page table page.
5380 */
5381 if (pmd_index(addr) != pmd_index(saddr) ||
5382 vm_flags != svm_flags ||
5383 sbase < svma->vm_start || svma->vm_end < s_end)
5384 return 0;
5385
5386 return saddr;
5387}
5388
Nicholas Krause31aafb42015-09-04 15:47:58 -07005389static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01005390{
5391 unsigned long base = addr & PUD_MASK;
5392 unsigned long end = base + PUD_SIZE;
5393
5394 /*
5395 * check on proper vm_flags and page table alignment
5396 */
Mike Kravetz017b1662018-10-05 15:51:29 -07005397 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
Nicholas Krause31aafb42015-09-04 15:47:58 -07005398 return true;
5399 return false;
Steve Capper3212b532013-04-23 12:35:02 +01005400}
5401
Peter Xuf1d8bd52021-03-18 17:01:46 +11005402bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5403{
Peter Xuf1d8bd52021-03-18 17:01:46 +11005404#ifdef CONFIG_USERFAULTFD
5405 if (uffd_disable_huge_pmd_share(vma))
5406 return false;
5407#endif
5408 return vma_shareable(vma, addr);
5409}
5410
Steve Capper3212b532013-04-23 12:35:02 +01005411/*
Mike Kravetz017b1662018-10-05 15:51:29 -07005412 * Determine if start,end range within vma could be mapped by shared pmd.
5413 * If yes, adjust start and end to cover range associated with possible
5414 * shared pmd mappings.
5415 */
5416void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5417 unsigned long *start, unsigned long *end)
5418{
Li Xinhaie3359522021-02-24 12:06:54 -08005419 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
5420 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005421
Li Xinhaie3359522021-02-24 12:06:54 -08005422 /*
5423 * vma need span at least one aligned PUD size and the start,end range
5424 * must at least partialy within it.
5425 */
5426 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
5427 (*end <= v_start) || (*start >= v_end))
Mike Kravetz017b1662018-10-05 15:51:29 -07005428 return;
5429
Peter Xu75802ca62020-08-06 23:26:11 -07005430 /* Extend the range to be PUD aligned for a worst case scenario */
Li Xinhaie3359522021-02-24 12:06:54 -08005431 if (*start > v_start)
5432 *start = ALIGN_DOWN(*start, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005433
Li Xinhaie3359522021-02-24 12:06:54 -08005434 if (*end < v_end)
5435 *end = ALIGN(*end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005436}
5437
5438/*
Steve Capper3212b532013-04-23 12:35:02 +01005439 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5440 * and returns the corresponding pte. While this is not necessary for the
5441 * !shared pmd case because we can allocate the pmd later as well, it makes the
Mike Kravetzc0d03812020-04-01 21:11:05 -07005442 * code much cleaner.
5443 *
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005444 * This routine must be called with i_mmap_rwsem held in at least read mode if
5445 * sharing is possible. For hugetlbfs, this prevents removal of any page
5446 * table entries associated with the address space. This is important as we
5447 * are setting up sharing based on existing page table entries (mappings).
5448 *
5449 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5450 * huge_pte_alloc know that sharing is not possible and do not take
5451 * i_mmap_rwsem as a performance optimization. This is handled by the
5452 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5453 * only required for subsequent processing.
Steve Capper3212b532013-04-23 12:35:02 +01005454 */
Peter Xu59caf932021-03-18 17:01:46 +11005455pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
5456 unsigned long addr, pud_t *pud)
Steve Capper3212b532013-04-23 12:35:02 +01005457{
Steve Capper3212b532013-04-23 12:35:02 +01005458 struct address_space *mapping = vma->vm_file->f_mapping;
5459 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5460 vma->vm_pgoff;
5461 struct vm_area_struct *svma;
5462 unsigned long saddr;
5463 pte_t *spte = NULL;
5464 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005465 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01005466
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005467 i_mmap_assert_locked(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005468 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5469 if (svma == vma)
5470 continue;
5471
5472 saddr = page_table_shareable(svma, vma, addr, idx);
5473 if (saddr) {
Punit Agrawal7868a202017-07-06 15:39:42 -07005474 spte = huge_pte_offset(svma->vm_mm, saddr,
5475 vma_mmu_pagesize(svma));
Steve Capper3212b532013-04-23 12:35:02 +01005476 if (spte) {
5477 get_page(virt_to_page(spte));
5478 break;
5479 }
5480 }
5481 }
5482
5483 if (!spte)
5484 goto out;
5485
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08005486 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005487 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01005488 pud_populate(mm, pud,
5489 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07005490 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005491 } else {
Steve Capper3212b532013-04-23 12:35:02 +01005492 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005493 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005494 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01005495out:
5496 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Steve Capper3212b532013-04-23 12:35:02 +01005497 return pte;
5498}
5499
5500/*
5501 * unmap huge page backed by shared pte.
5502 *
5503 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5504 * indicated by page_count > 1, unmap is achieved by clearing pud and
5505 * decrementing the ref count. If count == 1, the pte page is not shared.
5506 *
Mike Kravetzc0d03812020-04-01 21:11:05 -07005507 * Called with page table lock held and i_mmap_rwsem held in write mode.
Steve Capper3212b532013-04-23 12:35:02 +01005508 *
5509 * returns: 1 successfully unmapped a shared pte page
5510 * 0 the underlying pte page is not shared, or it is the last user
5511 */
Mike Kravetz34ae204f2020-08-11 18:31:38 -07005512int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5513 unsigned long *addr, pte_t *ptep)
Steve Capper3212b532013-04-23 12:35:02 +01005514{
5515 pgd_t *pgd = pgd_offset(mm, *addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005516 p4d_t *p4d = p4d_offset(pgd, *addr);
5517 pud_t *pud = pud_offset(p4d, *addr);
Steve Capper3212b532013-04-23 12:35:02 +01005518
Mike Kravetz34ae204f2020-08-11 18:31:38 -07005519 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005520 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5521 if (page_count(virt_to_page(ptep)) == 1)
5522 return 0;
5523
5524 pud_clear(pud);
5525 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005526 mm_dec_nr_pmds(mm);
Mike Kravetz63758dd2022-05-24 13:50:03 -07005527 /*
5528 * This update of passed address optimizes loops sequentially
5529 * processing addresses in increments of huge page size (PMD_SIZE
5530 * in this case). By clearing the pud, a PUD_SIZE area is unmapped.
5531 * Update address to the 'last page' in the cleared area so that
5532 * calling loop can move to first page past this area.
5533 */
5534 *addr |= PUD_SIZE - PMD_SIZE;
Steve Capper3212b532013-04-23 12:35:02 +01005535 return 1;
5536}
Peter Xuf1d8bd52021-03-18 17:01:46 +11005537
Steve Capper9e5fc742013-04-30 08:02:03 +01005538#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
Lokesh Gidra167612c12021-05-21 10:17:36 -07005539pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
Peter Xu59caf932021-03-18 17:01:46 +11005540 unsigned long addr, pud_t *pud)
Steve Capper9e5fc742013-04-30 08:02:03 +01005541{
5542 return NULL;
5543}
Zhang Zhene81f2d22015-06-24 16:56:13 -07005544
Mike Kravetz34ae204f2020-08-11 18:31:38 -07005545int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5546 unsigned long *addr, pte_t *ptep)
Zhang Zhene81f2d22015-06-24 16:56:13 -07005547{
5548 return 0;
5549}
Mike Kravetz017b1662018-10-05 15:51:29 -07005550
5551void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5552 unsigned long *start, unsigned long *end)
5553{
5554}
Peter Xu8b03b572021-03-18 17:01:47 +11005555
5556bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
5557{
5558 return false;
5559}
Steve Capper3212b532013-04-23 12:35:02 +01005560#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5561
Steve Capper9e5fc742013-04-30 08:02:03 +01005562#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
Peter Xu59caf932021-03-18 17:01:46 +11005563pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
Steve Capper9e5fc742013-04-30 08:02:03 +01005564 unsigned long addr, unsigned long sz)
5565{
5566 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005567 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01005568 pud_t *pud;
5569 pte_t *pte = NULL;
5570
5571 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovf4f0a3d2017-11-29 16:11:30 -08005572 p4d = p4d_alloc(mm, pgd, addr);
5573 if (!p4d)
5574 return NULL;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005575 pud = pud_alloc(mm, p4d, addr);
Steve Capper9e5fc742013-04-30 08:02:03 +01005576 if (pud) {
5577 if (sz == PUD_SIZE) {
5578 pte = (pte_t *)pud;
5579 } else {
5580 BUG_ON(sz != PMD_SIZE);
Peter Xuf1d8bd52021-03-18 17:01:46 +11005581 if (want_pmd_share(vma, addr) && pud_none(*pud))
Peter Xu59caf932021-03-18 17:01:46 +11005582 pte = huge_pmd_share(mm, vma, addr, pud);
Steve Capper9e5fc742013-04-30 08:02:03 +01005583 else
5584 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5585 }
5586 }
Michal Hocko4e666312016-08-02 14:02:34 -07005587 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01005588
5589 return pte;
5590}
5591
Punit Agrawal9b19df22017-09-06 16:21:01 -07005592/*
5593 * huge_pte_offset() - Walk the page table to resolve the hugepage
5594 * entry at address @addr
5595 *
Li Xinhai8ac0b812020-06-03 16:00:53 -07005596 * Return: Pointer to page table entry (PUD or PMD) for
5597 * address @addr, or NULL if a !p*d_present() entry is encountered and the
Punit Agrawal9b19df22017-09-06 16:21:01 -07005598 * size @sz doesn't match the hugepage size at this level of the page
5599 * table.
5600 */
Punit Agrawal7868a202017-07-06 15:39:42 -07005601pte_t *huge_pte_offset(struct mm_struct *mm,
5602 unsigned long addr, unsigned long sz)
Steve Capper9e5fc742013-04-30 08:02:03 +01005603{
5604 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005605 p4d_t *p4d;
Li Xinhai8ac0b812020-06-03 16:00:53 -07005606 pud_t *pud;
5607 pmd_t *pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01005608
5609 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005610 if (!pgd_present(*pgd))
5611 return NULL;
5612 p4d = p4d_offset(pgd, addr);
5613 if (!p4d_present(*p4d))
5614 return NULL;
Punit Agrawal9b19df22017-09-06 16:21:01 -07005615
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005616 pud = pud_offset(p4d, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07005617 if (sz == PUD_SIZE)
5618 /* must be pud huge, non-present or none */
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005619 return (pte_t *)pud;
Li Xinhai8ac0b812020-06-03 16:00:53 -07005620 if (!pud_present(*pud))
5621 return NULL;
5622 /* must have a valid entry and size to go further */
Punit Agrawal9b19df22017-09-06 16:21:01 -07005623
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005624 pmd = pmd_offset(pud, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07005625 /* must be pmd huge, non-present or none */
5626 return (pte_t *)pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01005627}
5628
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08005629#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5630
5631/*
5632 * These functions are overwritable if your architecture needs its own
5633 * behavior.
5634 */
5635struct page * __weak
5636follow_huge_addr(struct mm_struct *mm, unsigned long address,
5637 int write)
5638{
5639 return ERR_PTR(-EINVAL);
5640}
5641
5642struct page * __weak
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -07005643follow_huge_pd(struct vm_area_struct *vma,
5644 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5645{
5646 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5647 return NULL;
5648}
5649
5650struct page * __weak
Baolin Wangfccee932022-09-01 18:41:31 +08005651follow_huge_pmd_pte(struct vm_area_struct *vma, unsigned long address, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01005652{
Baolin Wangfccee932022-09-01 18:41:31 +08005653 struct hstate *h = hstate_vma(vma);
5654 struct mm_struct *mm = vma->vm_mm;
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005655 struct page *page = NULL;
5656 spinlock_t *ptl;
Baolin Wangfccee932022-09-01 18:41:31 +08005657 pte_t *ptep, pte;
John Hubbard3faa52c2020-04-01 21:05:29 -07005658
5659 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5660 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5661 (FOLL_PIN | FOLL_GET)))
5662 return NULL;
5663
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005664retry:
Baolin Wangfccee932022-09-01 18:41:31 +08005665 ptep = huge_pte_offset(mm, address, huge_page_size(h));
5666 if (!ptep)
5667 return NULL;
5668
5669 ptl = huge_pte_lock(h, mm, ptep);
5670 pte = huge_ptep_get(ptep);
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005671 if (pte_present(pte)) {
Baolin Wangfccee932022-09-01 18:41:31 +08005672 page = pte_page(pte) +
5673 ((address & ~huge_page_mask(h)) >> PAGE_SHIFT);
John Hubbard3faa52c2020-04-01 21:05:29 -07005674 /*
5675 * try_grab_page() should always succeed here, because: a) we
5676 * hold the pmd (ptl) lock, and b) we've just checked that the
5677 * huge pmd (head) page is present in the page tables. The ptl
5678 * prevents the head page and tail pages from being rearranged
5679 * in any way. So this page must be available at this point,
5680 * unless the page refcount overflowed:
5681 */
5682 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5683 page = NULL;
5684 goto out;
5685 }
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005686 } else {
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005687 if (is_hugetlb_entry_migration(pte)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005688 spin_unlock(ptl);
Baolin Wangfccee932022-09-01 18:41:31 +08005689 __migration_entry_wait(mm, ptep, ptl);
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005690 goto retry;
5691 }
5692 /*
5693 * hwpoisoned entry is treated as no_page_table in
5694 * follow_page_mask().
5695 */
5696 }
5697out:
5698 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01005699 return page;
5700}
5701
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08005702struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01005703follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005704 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01005705{
John Hubbard3faa52c2020-04-01 21:05:29 -07005706 if (flags & (FOLL_GET | FOLL_PIN))
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005707 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01005708
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005709 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01005710}
5711
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07005712struct page * __weak
5713follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5714{
John Hubbard3faa52c2020-04-01 21:05:29 -07005715 if (flags & (FOLL_GET | FOLL_PIN))
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07005716 return NULL;
5717
5718 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5719}
5720
Miaohe Lin97a51042022-05-30 19:30:15 +08005721int isolate_hugetlb(struct page *page, struct list_head *list)
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005722{
Miaohe Lin97a51042022-05-30 19:30:15 +08005723 int ret = 0;
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005724
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005725 spin_lock(&hugetlb_lock);
Muchun Song5b9631c2021-02-04 18:32:10 -08005726 if (!PageHeadHuge(page) || !page_huge_active(page) ||
5727 !get_page_unless_zero(page)) {
Miaohe Lin97a51042022-05-30 19:30:15 +08005728 ret = -EBUSY;
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005729 goto unlock;
5730 }
5731 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005732 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005733unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005734 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005735 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005736}
5737
5738void putback_active_hugepage(struct page *page)
5739{
Sasha Levin309381fea2014-01-23 15:52:54 -08005740 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005741 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005742 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005743 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5744 spin_unlock(&hugetlb_lock);
5745 put_page(page);
5746}
Michal Hockoab5ac902018-01-31 16:20:48 -08005747
5748void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5749{
5750 struct hstate *h = page_hstate(oldpage);
5751
5752 hugetlb_cgroup_migrate(oldpage, newpage);
5753 set_page_owner_migrate_reason(newpage, reason);
5754
5755 /*
5756 * transfer temporary state of the new huge page. This is
5757 * reverse to other transitions because the newpage is going to
5758 * be final while the old one will be freed so it takes over
5759 * the temporary status.
5760 *
5761 * Also note that we have to transfer the per-node surplus state
5762 * here as well otherwise the global surplus count will not match
5763 * the per-node's.
5764 */
5765 if (PageHugeTemporary(newpage)) {
5766 int old_nid = page_to_nid(oldpage);
5767 int new_nid = page_to_nid(newpage);
5768
5769 SetPageHugeTemporary(oldpage);
5770 ClearPageHugeTemporary(newpage);
5771
5772 spin_lock(&hugetlb_lock);
5773 if (h->surplus_huge_pages_node[old_nid]) {
5774 h->surplus_huge_pages_node[old_nid]--;
5775 h->surplus_huge_pages_node[new_nid]++;
5776 }
5777 spin_unlock(&hugetlb_lock);
5778 }
5779}
Roman Gushchincf11e852020-04-10 14:32:45 -07005780
James Houghton43390f12023-01-04 23:19:10 +00005781static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
5782 unsigned long start,
5783 unsigned long end)
Peter Xu343cacf2021-03-18 17:01:47 +11005784{
5785 struct hstate *h = hstate_vma(vma);
5786 unsigned long sz = huge_page_size(h);
5787 struct mm_struct *mm = vma->vm_mm;
5788 struct mmu_notifier_range range;
James Houghton43390f12023-01-04 23:19:10 +00005789 unsigned long address;
Peter Xu343cacf2021-03-18 17:01:47 +11005790 spinlock_t *ptl;
5791 pte_t *ptep;
5792
5793 if (!(vma->vm_flags & VM_MAYSHARE))
5794 return;
5795
Peter Xu343cacf2021-03-18 17:01:47 +11005796 if (start >= end)
5797 return;
5798
Baolin Wangec3e1bd2022-04-29 14:36:58 -07005799 flush_cache_range(vma, start, end);
Peter Xu343cacf2021-03-18 17:01:47 +11005800 /*
5801 * No need to call adjust_range_if_pmd_sharing_possible(), because
5802 * we have already done the PUD_SIZE alignment.
5803 */
5804 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
5805 start, end);
5806 mmu_notifier_invalidate_range_start(&range);
5807 i_mmap_lock_write(vma->vm_file->f_mapping);
5808 for (address = start; address < end; address += PUD_SIZE) {
5809 unsigned long tmp = address;
5810
5811 ptep = huge_pte_offset(mm, address, sz);
5812 if (!ptep)
5813 continue;
5814 ptl = huge_pte_lock(h, mm, ptep);
5815 /* We don't want 'address' to be changed */
5816 huge_pmd_unshare(mm, vma, &tmp, ptep);
5817 spin_unlock(ptl);
5818 }
5819 flush_hugetlb_tlb_range(vma, start, end);
5820 i_mmap_unlock_write(vma->vm_file->f_mapping);
5821 /*
5822 * No need to call mmu_notifier_invalidate_range(), see
5823 * Documentation/vm/mmu_notifier.rst.
5824 */
5825 mmu_notifier_invalidate_range_end(&range);
5826}
5827
James Houghton43390f12023-01-04 23:19:10 +00005828/*
5829 * This function will unconditionally remove all the shared pmd pgtable entries
5830 * within the specific vma for a hugetlbfs memory range.
5831 */
5832void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
5833{
5834 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
5835 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
5836}
5837
Roman Gushchincf11e852020-04-10 14:32:45 -07005838#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -07005839static bool cma_reserve_called __initdata;
5840
5841static int __init cmdline_parse_hugetlb_cma(char *p)
5842{
5843 hugetlb_cma_size = memparse(p, &p);
5844 return 0;
5845}
5846
5847early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5848
5849void __init hugetlb_cma_reserve(int order)
5850{
5851 unsigned long size, reserved, per_node;
5852 int nid;
5853
5854 cma_reserve_called = true;
5855
5856 if (!hugetlb_cma_size)
5857 return;
5858
5859 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5860 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5861 (PAGE_SIZE << order) / SZ_1M);
5862 return;
5863 }
5864
5865 /*
5866 * If 3 GB area is requested on a machine with 4 numa nodes,
5867 * let's allocate 1 GB on first three nodes and ignore the last one.
5868 */
5869 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5870 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5871 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5872
5873 reserved = 0;
5874 for_each_node_state(nid, N_ONLINE) {
5875 int res;
Barry Song2281f792020-08-24 11:03:09 +12005876 char name[CMA_MAX_NAME];
Roman Gushchincf11e852020-04-10 14:32:45 -07005877
5878 size = min(per_node, hugetlb_cma_size - reserved);
5879 size = round_up(size, PAGE_SIZE << order);
5880
Barry Song2281f792020-08-24 11:03:09 +12005881 snprintf(name, sizeof(name), "hugetlb%d", nid);
Roman Gushchincf11e852020-04-10 14:32:45 -07005882 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
Barry Song29d0f412020-08-11 18:32:00 -07005883 0, false, name,
Roman Gushchincf11e852020-04-10 14:32:45 -07005884 &hugetlb_cma[nid], nid);
5885 if (res) {
5886 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5887 res, nid);
5888 continue;
5889 }
5890
5891 reserved += size;
5892 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5893 size / SZ_1M, nid);
5894
5895 if (reserved >= hugetlb_cma_size)
5896 break;
5897 }
5898}
5899
5900void __init hugetlb_cma_check(void)
5901{
5902 if (!hugetlb_cma_size || cma_reserve_called)
5903 return;
5904
5905 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5906}
5907
5908#endif /* CONFIG_CMA */