blob: 5a620f6909114792a7e1698adab010f1a59b4f6a [file] [log] [blame]
Thomas Gleixner457c8992019-05-19 13:08:55 +01001// SPDX-License-Identifier: GPL-2.0-only
Linus Torvalds1da177e2005-04-16 15:20:36 -07002/*
3 * Generic hugetlb support.
Nadia Yvette Chambers6d49e352012-12-06 10:39:54 +01004 * (C) Nadia Yvette Chambers, April 2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006#include <linux/list.h>
7#include <linux/init.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -07008#include <linux/mm.h>
Alexey Dobriyane1759c22008-10-15 23:50:22 +04009#include <linux/seq_file.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070010#include <linux/sysctl.h>
11#include <linux/highmem.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070012#include <linux/mmu_notifier.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/nodemask.h>
David Gibson63551ae2005-06-21 17:14:44 -070014#include <linux/pagemap.h>
Christoph Lameter5da7ca82006-01-06 00:10:46 -080015#include <linux/mempolicy.h>
Gideon Israel Dsouza3b321232014-04-07 15:37:26 -070016#include <linux/compiler.h>
Christoph Lameteraea47ff2006-01-08 01:00:57 -080017#include <linux/cpuset.h>
David Gibson3935baa2006-03-22 00:08:53 -080018#include <linux/mutex.h>
Mike Rapoport97ad1082018-10-30 15:09:44 -070019#include <linux/memblock.h>
Nishanth Aravamudana3437872008-07-23 21:27:44 -070020#include <linux/sysfs.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Joonsoo Kimbbe88752020-08-11 18:37:38 -070022#include <linux/sched/mm.h>
Mike Kravetz63489f82018-03-22 16:17:13 -070023#include <linux/mmdebug.h>
Ingo Molnar174cd4b2017-02-02 19:15:33 +010024#include <linux/sched/signal.h>
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +090025#include <linux/rmap.h>
Matthew Wilcoxc6247f72017-07-10 15:48:56 -070026#include <linux/string_helpers.h>
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +090027#include <linux/swap.h>
28#include <linux/swapops.h>
Davidlohr Bueso8382d912014-04-03 14:47:31 -070029#include <linux/jhash.h>
Anshuman Khandual98fa15f2019-03-05 15:42:58 -080030#include <linux/numa.h>
Waiman Longc77c0a82020-01-04 13:00:15 -080031#include <linux/llist.h>
Roman Gushchincf11e852020-04-10 14:32:45 -070032#include <linux/cma.h>
Linus Torvaldsd6606682008-08-06 12:04:54 -070033
David Gibson63551ae2005-06-21 17:14:44 -070034#include <asm/page.h>
Mike Rapoportca15ca42020-08-06 23:22:28 -070035#include <asm/pgalloc.h>
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070036#include <asm/tlb.h>
David Gibson63551ae2005-06-21 17:14:44 -070037
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -070038#include <linux/io.h>
David Gibson63551ae2005-06-21 17:14:44 -070039#include <linux/hugetlb.h>
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -070040#include <linux/hugetlb_cgroup.h>
Lee Schermerhorn9a3052302009-12-14 17:58:25 -080041#include <linux/node.h>
Mike Kravetz1a1aad82017-02-22 15:43:01 -080042#include <linux/userfaultfd_k.h>
Michal Hockoab5ac902018-01-31 16:20:48 -080043#include <linux/page_owner.h>
Nick Piggin7835e982006-03-22 00:08:40 -080044#include "internal.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070046int hugetlb_max_hstate __read_mostly;
Andi Kleene5ff2152008-07-23 21:27:42 -070047unsigned int default_hstate_idx;
48struct hstate hstates[HUGE_MAX_HSTATE];
Roman Gushchincf11e852020-04-10 14:32:45 -070049
Barry Songdbda8fe2020-07-23 21:15:30 -070050#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -070051static struct cma *hugetlb_cma[MAX_NUMNODES];
Barry Songdbda8fe2020-07-23 21:15:30 -070052#endif
53static unsigned long hugetlb_cma_size __initdata;
Roman Gushchincf11e852020-04-10 14:32:45 -070054
Naoya Horiguchi641844f2015-06-24 16:56:59 -070055/*
56 * Minimum page order among possible hugepage sizes, set to a proper value
57 * at boot time.
58 */
59static unsigned int minimum_order __read_mostly = UINT_MAX;
Andi Kleene5ff2152008-07-23 21:27:42 -070060
Jon Tollefson53ba51d2008-07-23 21:27:52 -070061__initdata LIST_HEAD(huge_boot_pages);
62
Andi Kleene5ff2152008-07-23 21:27:42 -070063/* for command line parsing */
64static struct hstate * __initdata parsed_hstate;
65static unsigned long __initdata default_hstate_max_huge_pages;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -070066static bool __initdata parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -070067static bool __initdata parsed_default_hugepagesz;
Andi Kleene5ff2152008-07-23 21:27:42 -070068
David Gibson3935baa2006-03-22 00:08:53 -080069/*
Naoya Horiguchi31caf662013-09-11 14:21:59 -070070 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
71 * free_huge_pages, and surplus_huge_pages.
David Gibson3935baa2006-03-22 00:08:53 -080072 */
Aneesh Kumar K.Vc3f38a32012-07-31 16:42:10 -070073DEFINE_SPINLOCK(hugetlb_lock);
Eric Paris0bd0f9f2005-11-21 21:32:28 -080074
Davidlohr Bueso8382d912014-04-03 14:47:31 -070075/*
76 * Serializes faults on the same logical page. This is used to
77 * prevent spurious OOMs when the hugepage pool is fully utilized.
78 */
79static int num_fault_mutexes;
Mike Kravetzc672c7f2015-09-08 15:01:35 -070080struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
Davidlohr Bueso8382d912014-04-03 14:47:31 -070081
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070082/* Forward declaration */
83static int hugetlb_acct_memory(struct hstate *h, long delta);
84
David Gibson90481622012-03-21 16:34:12 -070085static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
86{
87 bool free = (spool->count == 0) && (spool->used_hpages == 0);
88
89 spin_unlock(&spool->lock);
90
91 /* If no pages are used, and no other handles to the subpool
Ethon Paul7c8de352020-06-04 16:49:07 -070092 * remain, give up any reservations based on minimum size and
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070093 * free the subpool */
94 if (free) {
95 if (spool->min_hpages != -1)
96 hugetlb_acct_memory(spool->hstate,
97 -spool->min_hpages);
David Gibson90481622012-03-21 16:34:12 -070098 kfree(spool);
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -070099 }
David Gibson90481622012-03-21 16:34:12 -0700100}
101
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700102struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
103 long min_hpages)
David Gibson90481622012-03-21 16:34:12 -0700104{
105 struct hugepage_subpool *spool;
106
Mike Kravetzc6a91822015-04-15 16:13:36 -0700107 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
David Gibson90481622012-03-21 16:34:12 -0700108 if (!spool)
109 return NULL;
110
111 spin_lock_init(&spool->lock);
112 spool->count = 1;
Mike Kravetz7ca02d0a2015-04-15 16:13:42 -0700113 spool->max_hpages = max_hpages;
114 spool->hstate = h;
115 spool->min_hpages = min_hpages;
116
117 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
118 kfree(spool);
119 return NULL;
120 }
121 spool->rsv_hpages = min_hpages;
David Gibson90481622012-03-21 16:34:12 -0700122
123 return spool;
124}
125
126void hugepage_put_subpool(struct hugepage_subpool *spool)
127{
128 spin_lock(&spool->lock);
129 BUG_ON(!spool->count);
130 spool->count--;
131 unlock_or_release_subpool(spool);
132}
133
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700134/*
135 * Subpool accounting for allocating and reserving pages.
136 * Return -ENOMEM if there are not enough resources to satisfy the
Randy Dunlap9e7ee402020-08-11 18:32:59 -0700137 * request. Otherwise, return the number of pages by which the
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700138 * global pools must be adjusted (upward). The returned value may
139 * only be different than the passed value (delta) in the case where
Ethon Paul7c8de352020-06-04 16:49:07 -0700140 * a subpool minimum size must be maintained.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700141 */
142static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700143 long delta)
144{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700145 long ret = delta;
David Gibson90481622012-03-21 16:34:12 -0700146
147 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700148 return ret;
David Gibson90481622012-03-21 16:34:12 -0700149
150 spin_lock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700151
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700152 if (spool->max_hpages != -1) { /* maximum size accounting */
153 if ((spool->used_hpages + delta) <= spool->max_hpages)
154 spool->used_hpages += delta;
155 else {
156 ret = -ENOMEM;
157 goto unlock_ret;
158 }
159 }
160
Mike Kravetz09a95e22016-05-19 17:11:01 -0700161 /* minimum size accounting */
162 if (spool->min_hpages != -1 && spool->rsv_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700163 if (delta > spool->rsv_hpages) {
164 /*
165 * Asking for more reserves than those already taken on
166 * behalf of subpool. Return difference.
167 */
168 ret = delta - spool->rsv_hpages;
169 spool->rsv_hpages = 0;
170 } else {
171 ret = 0; /* reserves already accounted for */
172 spool->rsv_hpages -= delta;
173 }
174 }
175
176unlock_ret:
177 spin_unlock(&spool->lock);
David Gibson90481622012-03-21 16:34:12 -0700178 return ret;
179}
180
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700181/*
182 * Subpool accounting for freeing and unreserving pages.
183 * Return the number of global page reservations that must be dropped.
184 * The return value may only be different than the passed value (delta)
185 * in the case where a subpool minimum size must be maintained.
186 */
187static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
David Gibson90481622012-03-21 16:34:12 -0700188 long delta)
189{
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700190 long ret = delta;
191
David Gibson90481622012-03-21 16:34:12 -0700192 if (!spool)
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700193 return delta;
David Gibson90481622012-03-21 16:34:12 -0700194
195 spin_lock(&spool->lock);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700196
197 if (spool->max_hpages != -1) /* maximum size accounting */
198 spool->used_hpages -= delta;
199
Mike Kravetz09a95e22016-05-19 17:11:01 -0700200 /* minimum size accounting */
201 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700202 if (spool->rsv_hpages + delta <= spool->min_hpages)
203 ret = 0;
204 else
205 ret = spool->rsv_hpages + delta - spool->min_hpages;
206
207 spool->rsv_hpages += delta;
208 if (spool->rsv_hpages > spool->min_hpages)
209 spool->rsv_hpages = spool->min_hpages;
210 }
211
212 /*
213 * If hugetlbfs_put_super couldn't free spool due to an outstanding
214 * quota reference, free it now.
215 */
David Gibson90481622012-03-21 16:34:12 -0700216 unlock_or_release_subpool(spool);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -0700217
218 return ret;
David Gibson90481622012-03-21 16:34:12 -0700219}
220
221static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
222{
223 return HUGETLBFS_SB(inode->i_sb)->spool;
224}
225
226static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
227{
Al Viro496ad9a2013-01-23 17:07:38 -0500228 return subpool_inode(file_inode(vma->vm_file));
David Gibson90481622012-03-21 16:34:12 -0700229}
230
Mina Almasry0db9d742020-04-01 21:11:25 -0700231/* Helper that removes a struct file_region from the resv_map cache and returns
232 * it for use.
233 */
234static struct file_region *
235get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
236{
237 struct file_region *nrg = NULL;
238
239 VM_BUG_ON(resv->region_cache_count <= 0);
240
241 resv->region_cache_count--;
242 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
Mina Almasry0db9d742020-04-01 21:11:25 -0700243 list_del(&nrg->link);
244
245 nrg->from = from;
246 nrg->to = to;
247
248 return nrg;
249}
250
Mina Almasry075a61d2020-04-01 21:11:28 -0700251static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
252 struct file_region *rg)
253{
254#ifdef CONFIG_CGROUP_HUGETLB
255 nrg->reservation_counter = rg->reservation_counter;
256 nrg->css = rg->css;
257 if (rg->css)
258 css_get(rg->css);
259#endif
260}
261
262/* Helper that records hugetlb_cgroup uncharge info. */
263static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
264 struct hstate *h,
265 struct resv_map *resv,
266 struct file_region *nrg)
267{
268#ifdef CONFIG_CGROUP_HUGETLB
269 if (h_cg) {
270 nrg->reservation_counter =
271 &h_cg->rsvd_hugepage[hstate_index(h)];
272 nrg->css = &h_cg->css;
273 if (!resv->pages_per_hpage)
274 resv->pages_per_hpage = pages_per_huge_page(h);
275 /* pages_per_hpage should be the same for all entries in
276 * a resv_map.
277 */
278 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
279 } else {
280 nrg->reservation_counter = NULL;
281 nrg->css = NULL;
282 }
283#endif
284}
285
Mina Almasrya9b3f862020-04-01 21:11:35 -0700286static bool has_same_uncharge_info(struct file_region *rg,
287 struct file_region *org)
288{
289#ifdef CONFIG_CGROUP_HUGETLB
290 return rg && org &&
291 rg->reservation_counter == org->reservation_counter &&
292 rg->css == org->css;
293
294#else
295 return true;
296#endif
297}
298
299static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
300{
301 struct file_region *nrg = NULL, *prg = NULL;
302
303 prg = list_prev_entry(rg, link);
304 if (&prg->link != &resv->regions && prg->to == rg->from &&
305 has_same_uncharge_info(prg, rg)) {
306 prg->to = rg->to;
307
308 list_del(&rg->link);
309 kfree(rg);
310
Wei Yang7db5e7b2020-10-13 16:56:20 -0700311 rg = prg;
Mina Almasrya9b3f862020-04-01 21:11:35 -0700312 }
313
314 nrg = list_next_entry(rg, link);
315 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
316 has_same_uncharge_info(nrg, rg)) {
317 nrg->from = rg->from;
318
319 list_del(&rg->link);
320 kfree(rg);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700321 }
322}
323
Wei Yang972a3da32020-10-13 16:56:30 -0700324/*
325 * Must be called with resv->lock held.
326 *
327 * Calling this with regions_needed != NULL will count the number of pages
328 * to be added but will not modify the linked list. And regions_needed will
329 * indicate the number of file_regions needed in the cache to carry out to add
330 * the regions for this range.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800331 */
332static long add_reservation_in_range(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700333 struct hugetlb_cgroup *h_cg,
Wei Yang972a3da32020-10-13 16:56:30 -0700334 struct hstate *h, long *regions_needed)
Mina Almasryd75c6af2019-11-30 17:56:59 -0800335{
Mina Almasry0db9d742020-04-01 21:11:25 -0700336 long add = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800337 struct list_head *head = &resv->regions;
Mina Almasry0db9d742020-04-01 21:11:25 -0700338 long last_accounted_offset = f;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800339 struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
340
Mina Almasry0db9d742020-04-01 21:11:25 -0700341 if (regions_needed)
342 *regions_needed = 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800343
Mina Almasry0db9d742020-04-01 21:11:25 -0700344 /* In this loop, we essentially handle an entry for the range
345 * [last_accounted_offset, rg->from), at every iteration, with some
346 * bounds checking.
347 */
348 list_for_each_entry_safe(rg, trg, head, link) {
349 /* Skip irrelevant regions that start before our range. */
350 if (rg->from < f) {
351 /* If this region ends after the last accounted offset,
352 * then we need to update last_accounted_offset.
353 */
354 if (rg->to > last_accounted_offset)
355 last_accounted_offset = rg->to;
356 continue;
357 }
Mina Almasryd75c6af2019-11-30 17:56:59 -0800358
Mina Almasry0db9d742020-04-01 21:11:25 -0700359 /* When we find a region that starts beyond our range, we've
360 * finished.
361 */
Mina Almasryd75c6af2019-11-30 17:56:59 -0800362 if (rg->from > t)
363 break;
364
Mina Almasry0db9d742020-04-01 21:11:25 -0700365 /* Add an entry for last_accounted_offset -> rg->from, and
366 * update last_accounted_offset.
Mina Almasryd75c6af2019-11-30 17:56:59 -0800367 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700368 if (rg->from > last_accounted_offset) {
369 add += rg->from - last_accounted_offset;
Wei Yang972a3da32020-10-13 16:56:30 -0700370 if (!regions_needed) {
Mina Almasry0db9d742020-04-01 21:11:25 -0700371 nrg = get_file_region_entry_from_cache(
372 resv, last_accounted_offset, rg->from);
Mina Almasry075a61d2020-04-01 21:11:28 -0700373 record_hugetlb_cgroup_uncharge_info(h_cg, h,
374 resv, nrg);
Mina Almasry0db9d742020-04-01 21:11:25 -0700375 list_add(&nrg->link, rg->link.prev);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700376 coalesce_file_region(resv, nrg);
Wei Yang972a3da32020-10-13 16:56:30 -0700377 } else
Mina Almasry0db9d742020-04-01 21:11:25 -0700378 *regions_needed += 1;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800379 }
Mina Almasryd75c6af2019-11-30 17:56:59 -0800380
Mina Almasry0db9d742020-04-01 21:11:25 -0700381 last_accounted_offset = rg->to;
382 }
383
384 /* Handle the case where our range extends beyond
385 * last_accounted_offset.
386 */
387 if (last_accounted_offset < t) {
388 add += t - last_accounted_offset;
Wei Yang972a3da32020-10-13 16:56:30 -0700389 if (!regions_needed) {
Mina Almasry0db9d742020-04-01 21:11:25 -0700390 nrg = get_file_region_entry_from_cache(
391 resv, last_accounted_offset, t);
Mina Almasry075a61d2020-04-01 21:11:28 -0700392 record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg);
Mina Almasry0db9d742020-04-01 21:11:25 -0700393 list_add(&nrg->link, rg->link.prev);
Mina Almasrya9b3f862020-04-01 21:11:35 -0700394 coalesce_file_region(resv, nrg);
Wei Yang972a3da32020-10-13 16:56:30 -0700395 } else
Mina Almasry0db9d742020-04-01 21:11:25 -0700396 *regions_needed += 1;
397 }
398
399 VM_BUG_ON(add < 0);
400 return add;
401}
402
403/* Must be called with resv->lock acquired. Will drop lock to allocate entries.
404 */
405static int allocate_file_region_entries(struct resv_map *resv,
406 int regions_needed)
407 __must_hold(&resv->lock)
408{
409 struct list_head allocated_regions;
410 int to_allocate = 0, i = 0;
411 struct file_region *trg = NULL, *rg = NULL;
412
413 VM_BUG_ON(regions_needed < 0);
414
415 INIT_LIST_HEAD(&allocated_regions);
416
417 /*
418 * Check for sufficient descriptors in the cache to accommodate
419 * the number of in progress add operations plus regions_needed.
420 *
421 * This is a while loop because when we drop the lock, some other call
422 * to region_add or region_del may have consumed some region_entries,
423 * so we keep looping here until we finally have enough entries for
424 * (adds_in_progress + regions_needed).
425 */
426 while (resv->region_cache_count <
427 (resv->adds_in_progress + regions_needed)) {
428 to_allocate = resv->adds_in_progress + regions_needed -
429 resv->region_cache_count;
430
431 /* At this point, we should have enough entries in the cache
432 * for all the existings adds_in_progress. We should only be
433 * needing to allocate for regions_needed.
434 */
435 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
436
437 spin_unlock(&resv->lock);
438 for (i = 0; i < to_allocate; i++) {
439 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
440 if (!trg)
441 goto out_of_memory;
442 list_add(&trg->link, &allocated_regions);
443 }
444
445 spin_lock(&resv->lock);
446
Wei Yangd3ec7b62020-10-13 16:56:27 -0700447 list_splice(&allocated_regions, &resv->region_cache);
448 resv->region_cache_count += to_allocate;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800449 }
450
Mina Almasry0db9d742020-04-01 21:11:25 -0700451 return 0;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800452
Mina Almasry0db9d742020-04-01 21:11:25 -0700453out_of_memory:
454 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
455 list_del(&rg->link);
456 kfree(rg);
457 }
458 return -ENOMEM;
Mina Almasryd75c6af2019-11-30 17:56:59 -0800459}
460
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700461/*
462 * Add the huge page range represented by [f, t) to the reserve
Mina Almasry0db9d742020-04-01 21:11:25 -0700463 * map. Regions will be taken from the cache to fill in this range.
464 * Sufficient regions should exist in the cache due to the previous
465 * call to region_chg with the same range, but in some cases the cache will not
466 * have sufficient entries due to races with other code doing region_add or
467 * region_del. The extra needed entries will be allocated.
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700468 *
Mina Almasry0db9d742020-04-01 21:11:25 -0700469 * regions_needed is the out value provided by a previous call to region_chg.
470 *
471 * Return the number of new huge pages added to the map. This number is greater
472 * than or equal to zero. If file_region entries needed to be allocated for
Ethon Paul7c8de352020-06-04 16:49:07 -0700473 * this operation and we were not able to allocate, it returns -ENOMEM.
Mina Almasry0db9d742020-04-01 21:11:25 -0700474 * region_add of regions of length 1 never allocate file_regions and cannot
475 * fail; region_chg will always allocate at least 1 entry and a region_add for
476 * 1 page will only require at most 1 entry.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700477 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700478static long region_add(struct resv_map *resv, long f, long t,
Mina Almasry075a61d2020-04-01 21:11:28 -0700479 long in_regions_needed, struct hstate *h,
480 struct hugetlb_cgroup *h_cg)
Andy Whitcroft96822902008-07-23 21:27:29 -0700481{
Mina Almasry0db9d742020-04-01 21:11:25 -0700482 long add = 0, actual_regions_needed = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700483
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700484 spin_lock(&resv->lock);
Mina Almasry0db9d742020-04-01 21:11:25 -0700485retry:
486
487 /* Count how many regions are actually needed to execute this add. */
Wei Yang972a3da32020-10-13 16:56:30 -0700488 add_reservation_in_range(resv, f, t, NULL, NULL,
489 &actual_regions_needed);
Andy Whitcroft96822902008-07-23 21:27:29 -0700490
Mike Kravetz5e911372015-09-08 15:01:28 -0700491 /*
Mina Almasry0db9d742020-04-01 21:11:25 -0700492 * Check for sufficient descriptors in the cache to accommodate
493 * this add operation. Note that actual_regions_needed may be greater
494 * than in_regions_needed, as the resv_map may have been modified since
495 * the region_chg call. In this case, we need to make sure that we
496 * allocate extra entries, such that we have enough for all the
497 * existing adds_in_progress, plus the excess needed for this
498 * operation.
Mike Kravetz5e911372015-09-08 15:01:28 -0700499 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700500 if (actual_regions_needed > in_regions_needed &&
501 resv->region_cache_count <
502 resv->adds_in_progress +
503 (actual_regions_needed - in_regions_needed)) {
504 /* region_add operation of range 1 should never need to
505 * allocate file_region entries.
506 */
507 VM_BUG_ON(t - f <= 1);
Mike Kravetz5e911372015-09-08 15:01:28 -0700508
Mina Almasry0db9d742020-04-01 21:11:25 -0700509 if (allocate_file_region_entries(
510 resv, actual_regions_needed - in_regions_needed)) {
511 return -ENOMEM;
512 }
Mike Kravetz5e911372015-09-08 15:01:28 -0700513
Mina Almasry0db9d742020-04-01 21:11:25 -0700514 goto retry;
Mike Kravetz5e911372015-09-08 15:01:28 -0700515 }
516
Wei Yang972a3da32020-10-13 16:56:30 -0700517 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700518
Mina Almasry0db9d742020-04-01 21:11:25 -0700519 resv->adds_in_progress -= in_regions_needed;
520
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700521 spin_unlock(&resv->lock);
Mike Kravetzcf3ad202015-06-24 16:57:55 -0700522 VM_BUG_ON(add < 0);
523 return add;
Andy Whitcroft96822902008-07-23 21:27:29 -0700524}
525
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700526/*
527 * Examine the existing reserve map and determine how many
528 * huge pages in the specified range [f, t) are NOT currently
529 * represented. This routine is called before a subsequent
530 * call to region_add that will actually modify the reserve
531 * map to add the specified range [f, t). region_chg does
532 * not change the number of huge pages represented by the
Mina Almasry0db9d742020-04-01 21:11:25 -0700533 * map. A number of new file_region structures is added to the cache as a
534 * placeholder, for the subsequent region_add call to use. At least 1
535 * file_region structure is added.
536 *
537 * out_regions_needed is the number of regions added to the
538 * resv->adds_in_progress. This value needs to be provided to a follow up call
539 * to region_add or region_abort for proper accounting.
Mike Kravetz5e911372015-09-08 15:01:28 -0700540 *
541 * Returns the number of huge pages that need to be added to the existing
542 * reservation map for the range [f, t). This number is greater or equal to
543 * zero. -ENOMEM is returned if a new file_region structure or cache entry
544 * is needed and can not be allocated.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700545 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700546static long region_chg(struct resv_map *resv, long f, long t,
547 long *out_regions_needed)
Andy Whitcroft96822902008-07-23 21:27:29 -0700548{
Andy Whitcroft96822902008-07-23 21:27:29 -0700549 long chg = 0;
550
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700551 spin_lock(&resv->lock);
Mike Kravetz5e911372015-09-08 15:01:28 -0700552
Wei Yang972a3da32020-10-13 16:56:30 -0700553 /* Count how many hugepages in this range are NOT represented. */
Mina Almasry075a61d2020-04-01 21:11:28 -0700554 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
Wei Yang972a3da32020-10-13 16:56:30 -0700555 out_regions_needed);
Mike Kravetz5e911372015-09-08 15:01:28 -0700556
Mina Almasry0db9d742020-04-01 21:11:25 -0700557 if (*out_regions_needed == 0)
558 *out_regions_needed = 1;
Mike Kravetz5e911372015-09-08 15:01:28 -0700559
Mina Almasry0db9d742020-04-01 21:11:25 -0700560 if (allocate_file_region_entries(resv, *out_regions_needed))
561 return -ENOMEM;
Mike Kravetz5e911372015-09-08 15:01:28 -0700562
Mina Almasry0db9d742020-04-01 21:11:25 -0700563 resv->adds_in_progress += *out_regions_needed;
Andy Whitcroft96822902008-07-23 21:27:29 -0700564
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700565 spin_unlock(&resv->lock);
Andy Whitcroft96822902008-07-23 21:27:29 -0700566 return chg;
567}
568
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700569/*
Mike Kravetz5e911372015-09-08 15:01:28 -0700570 * Abort the in progress add operation. The adds_in_progress field
571 * of the resv_map keeps track of the operations in progress between
572 * calls to region_chg and region_add. Operations are sometimes
573 * aborted after the call to region_chg. In such cases, region_abort
Mina Almasry0db9d742020-04-01 21:11:25 -0700574 * is called to decrement the adds_in_progress counter. regions_needed
575 * is the value returned by the region_chg call, it is used to decrement
576 * the adds_in_progress counter.
Mike Kravetz5e911372015-09-08 15:01:28 -0700577 *
578 * NOTE: The range arguments [f, t) are not needed or used in this
579 * routine. They are kept to make reading the calling code easier as
580 * arguments will match the associated region_chg call.
581 */
Mina Almasry0db9d742020-04-01 21:11:25 -0700582static void region_abort(struct resv_map *resv, long f, long t,
583 long regions_needed)
Mike Kravetz5e911372015-09-08 15:01:28 -0700584{
585 spin_lock(&resv->lock);
586 VM_BUG_ON(!resv->region_cache_count);
Mina Almasry0db9d742020-04-01 21:11:25 -0700587 resv->adds_in_progress -= regions_needed;
Mike Kravetz5e911372015-09-08 15:01:28 -0700588 spin_unlock(&resv->lock);
589}
590
591/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700592 * Delete the specified range [f, t) from the reserve map. If the
593 * t parameter is LONG_MAX, this indicates that ALL regions after f
594 * should be deleted. Locate the regions which intersect [f, t)
595 * and either trim, delete or split the existing regions.
596 *
597 * Returns the number of huge pages deleted from the reserve map.
598 * In the normal case, the return value is zero or more. In the
599 * case where a region must be split, a new region descriptor must
600 * be allocated. If the allocation fails, -ENOMEM will be returned.
601 * NOTE: If the parameter t == LONG_MAX, then we will never split
602 * a region and possibly return -ENOMEM. Callers specifying
603 * t == LONG_MAX do not need to check for -ENOMEM error.
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700604 */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700605static long region_del(struct resv_map *resv, long f, long t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700606{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700607 struct list_head *head = &resv->regions;
Andy Whitcroft96822902008-07-23 21:27:29 -0700608 struct file_region *rg, *trg;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700609 struct file_region *nrg = NULL;
610 long del = 0;
Andy Whitcroft96822902008-07-23 21:27:29 -0700611
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700612retry:
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700613 spin_lock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700614 list_for_each_entry_safe(rg, trg, head, link) {
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800615 /*
616 * Skip regions before the range to be deleted. file_region
617 * ranges are normally of the form [from, to). However, there
618 * may be a "placeholder" entry in the map which is of the form
619 * (from, to) with from == to. Check for placeholder entries
620 * at the beginning of the range to be deleted.
621 */
622 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700623 continue;
Mike Kravetzdbe409e2015-12-11 13:40:52 -0800624
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700625 if (rg->from >= t)
Andy Whitcroft96822902008-07-23 21:27:29 -0700626 break;
Andy Whitcroft96822902008-07-23 21:27:29 -0700627
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700628 if (f > rg->from && t < rg->to) { /* Must split region */
629 /*
630 * Check for an entry in the cache before dropping
631 * lock and attempting allocation.
632 */
633 if (!nrg &&
634 resv->region_cache_count > resv->adds_in_progress) {
635 nrg = list_first_entry(&resv->region_cache,
636 struct file_region,
637 link);
638 list_del(&nrg->link);
639 resv->region_cache_count--;
640 }
641
642 if (!nrg) {
643 spin_unlock(&resv->lock);
644 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
645 if (!nrg)
646 return -ENOMEM;
647 goto retry;
648 }
649
650 del += t - f;
Mike Kravetz79aa9252020-11-01 17:07:27 -0800651 hugetlb_cgroup_uncharge_file_region(
652 resv, rg, t - f);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700653
654 /* New entry for end of split region */
655 nrg->from = t;
656 nrg->to = rg->to;
Mina Almasry075a61d2020-04-01 21:11:28 -0700657
658 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
659
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700660 INIT_LIST_HEAD(&nrg->link);
661
662 /* Original entry is trimmed */
663 rg->to = f;
664
665 list_add(&nrg->link, &rg->link);
666 nrg = NULL;
667 break;
668 }
669
670 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
671 del += rg->to - rg->from;
Mina Almasry075a61d2020-04-01 21:11:28 -0700672 hugetlb_cgroup_uncharge_file_region(resv, rg,
673 rg->to - rg->from);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700674 list_del(&rg->link);
675 kfree(rg);
676 continue;
677 }
678
679 if (f <= rg->from) { /* Trim beginning of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700680 hugetlb_cgroup_uncharge_file_region(resv, rg,
681 t - rg->from);
Mina Almasry075a61d2020-04-01 21:11:28 -0700682
Mike Kravetz79aa9252020-11-01 17:07:27 -0800683 del += t - rg->from;
684 rg->from = t;
685 } else { /* Trim end of region */
Mina Almasry075a61d2020-04-01 21:11:28 -0700686 hugetlb_cgroup_uncharge_file_region(resv, rg,
687 rg->to - f);
Mike Kravetz79aa9252020-11-01 17:07:27 -0800688
689 del += rg->to - f;
690 rg->to = f;
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700691 }
Andy Whitcroft96822902008-07-23 21:27:29 -0700692 }
693
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700694 spin_unlock(&resv->lock);
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700695 kfree(nrg);
696 return del;
Andy Whitcroft96822902008-07-23 21:27:29 -0700697}
698
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700699/*
Mike Kravetzb5cec282015-09-08 15:01:41 -0700700 * A rare out of memory error was encountered which prevented removal of
701 * the reserve map region for a page. The huge page itself was free'ed
702 * and removed from the page cache. This routine will adjust the subpool
703 * usage count, and the global reserve count if needed. By incrementing
704 * these counts, the reserve map entry which could not be deleted will
705 * appear as a "reserved" entry instead of simply dangling with incorrect
706 * counts.
707 */
zhong jiang72e29362016-10-07 17:02:01 -0700708void hugetlb_fix_reserve_counts(struct inode *inode)
Mike Kravetzb5cec282015-09-08 15:01:41 -0700709{
710 struct hugepage_subpool *spool = subpool_inode(inode);
711 long rsv_adjust;
712
713 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
zhong jiang72e29362016-10-07 17:02:01 -0700714 if (rsv_adjust) {
Mike Kravetzb5cec282015-09-08 15:01:41 -0700715 struct hstate *h = hstate_inode(inode);
716
717 hugetlb_acct_memory(h, 1);
718 }
719}
720
721/*
Mike Kravetz1dd308a2015-06-24 16:57:52 -0700722 * Count and return the number of huge pages in the reserve map
723 * that intersect with the range [f, t).
724 */
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700725static long region_count(struct resv_map *resv, long f, long t)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700726{
Joonsoo Kim1406ec92014-04-03 14:47:26 -0700727 struct list_head *head = &resv->regions;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700728 struct file_region *rg;
729 long chg = 0;
730
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700731 spin_lock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700732 /* Locate each segment we overlap with, and count that overlap. */
733 list_for_each_entry(rg, head, link) {
Wang Sheng-Huif2135a42012-05-29 15:06:17 -0700734 long seg_from;
735 long seg_to;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700736
737 if (rg->to <= f)
738 continue;
739 if (rg->from >= t)
740 break;
741
742 seg_from = max(rg->from, f);
743 seg_to = min(rg->to, t);
744
745 chg += seg_to - seg_from;
746 }
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700747 spin_unlock(&resv->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700748
749 return chg;
750}
751
Andy Whitcroft96822902008-07-23 21:27:29 -0700752/*
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700753 * Convert the address within this vma to the page offset within
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700754 * the mapping, in pagecache page units; huge pages here.
755 */
Andi Kleena5516432008-07-23 21:27:41 -0700756static pgoff_t vma_hugecache_offset(struct hstate *h,
757 struct vm_area_struct *vma, unsigned long address)
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700758{
Andi Kleena5516432008-07-23 21:27:41 -0700759 return ((address - vma->vm_start) >> huge_page_shift(h)) +
760 (vma->vm_pgoff >> huge_page_order(h));
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700761}
762
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900763pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
764 unsigned long address)
765{
766 return vma_hugecache_offset(hstate_vma(vma), vma, address);
767}
Dan Williamsdee41072016-05-14 12:20:44 -0700768EXPORT_SYMBOL_GPL(linear_hugepage_index);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +0900769
Andy Whitcroft84afd992008-07-23 21:27:32 -0700770/*
Mel Gorman08fba692009-01-06 14:38:53 -0800771 * Return the size of the pages allocated when backing a VMA. In the majority
772 * cases this will be same size as used by the page table entries.
773 */
774unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
775{
Dan Williams05ea8862018-04-05 16:24:25 -0700776 if (vma->vm_ops && vma->vm_ops->pagesize)
777 return vma->vm_ops->pagesize(vma);
778 return PAGE_SIZE;
Mel Gorman08fba692009-01-06 14:38:53 -0800779}
Joerg Roedelf340ca02009-06-19 15:16:22 +0200780EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
Mel Gorman08fba692009-01-06 14:38:53 -0800781
782/*
Mel Gorman33402892009-01-06 14:38:54 -0800783 * Return the page size being used by the MMU to back a VMA. In the majority
784 * of cases, the page size used by the kernel matches the MMU size. On
Dan Williams09135cc2018-04-05 16:24:21 -0700785 * architectures where it differs, an architecture-specific 'strong'
786 * version of this symbol is required.
Mel Gorman33402892009-01-06 14:38:54 -0800787 */
Dan Williams09135cc2018-04-05 16:24:21 -0700788__weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
Mel Gorman33402892009-01-06 14:38:54 -0800789{
790 return vma_kernel_pagesize(vma);
791}
Mel Gorman33402892009-01-06 14:38:54 -0800792
793/*
Andy Whitcroft84afd992008-07-23 21:27:32 -0700794 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
795 * bits of the reservation map pointer, which are always clear due to
796 * alignment.
797 */
798#define HPAGE_RESV_OWNER (1UL << 0)
799#define HPAGE_RESV_UNMAPPED (1UL << 1)
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700800#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700801
Mel Gormana1e78772008-07-23 21:27:23 -0700802/*
803 * These helpers are used to track how many pages are reserved for
804 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
805 * is guaranteed to have their future faults succeed.
806 *
807 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
808 * the reserve counters are updated with the hugetlb_lock held. It is safe
809 * to reset the VMA at fork() time as it is not in use yet and there is no
810 * chance of the global counters getting corrupted as a result of the values.
Andy Whitcroft84afd992008-07-23 21:27:32 -0700811 *
812 * The private mapping reservation is represented in a subtly different
813 * manner to a shared mapping. A shared mapping has a region map associated
814 * with the underlying file, this region map represents the backing file
815 * pages which have ever had a reservation assigned which this persists even
816 * after the page is instantiated. A private mapping has a region map
817 * associated with the original mmap which is attached to all VMAs which
818 * reference it, this region map represents those offsets which have consumed
819 * reservation ie. where pages have been instantiated.
Mel Gormana1e78772008-07-23 21:27:23 -0700820 */
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700821static unsigned long get_vma_private_data(struct vm_area_struct *vma)
822{
823 return (unsigned long)vma->vm_private_data;
824}
825
826static void set_vma_private_data(struct vm_area_struct *vma,
827 unsigned long value)
828{
829 vma->vm_private_data = (void *)value;
830}
831
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700832static void
833resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
834 struct hugetlb_cgroup *h_cg,
835 struct hstate *h)
836{
837#ifdef CONFIG_CGROUP_HUGETLB
838 if (!h_cg || !h) {
839 resv_map->reservation_counter = NULL;
840 resv_map->pages_per_hpage = 0;
841 resv_map->css = NULL;
842 } else {
843 resv_map->reservation_counter =
844 &h_cg->rsvd_hugepage[hstate_index(h)];
845 resv_map->pages_per_hpage = pages_per_huge_page(h);
846 resv_map->css = &h_cg->css;
847 }
848#endif
849}
850
Joonsoo Kim9119a412014-04-03 14:47:25 -0700851struct resv_map *resv_map_alloc(void)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700852{
853 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700854 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
855
856 if (!resv_map || !rg) {
857 kfree(resv_map);
858 kfree(rg);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700859 return NULL;
Mike Kravetz5e911372015-09-08 15:01:28 -0700860 }
Andy Whitcroft84afd992008-07-23 21:27:32 -0700861
862 kref_init(&resv_map->refs);
Davidlohr Bueso7b24d862014-04-03 14:47:27 -0700863 spin_lock_init(&resv_map->lock);
Andy Whitcroft84afd992008-07-23 21:27:32 -0700864 INIT_LIST_HEAD(&resv_map->regions);
865
Mike Kravetz5e911372015-09-08 15:01:28 -0700866 resv_map->adds_in_progress = 0;
Mina Almasrye9fe92a2020-04-01 21:11:21 -0700867 /*
868 * Initialize these to 0. On shared mappings, 0's here indicate these
869 * fields don't do cgroup accounting. On private mappings, these will be
870 * re-initialized to the proper values, to indicate that hugetlb cgroup
871 * reservations are to be un-charged from here.
872 */
873 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
Mike Kravetz5e911372015-09-08 15:01:28 -0700874
875 INIT_LIST_HEAD(&resv_map->region_cache);
876 list_add(&rg->link, &resv_map->region_cache);
877 resv_map->region_cache_count = 1;
878
Andy Whitcroft84afd992008-07-23 21:27:32 -0700879 return resv_map;
880}
881
Joonsoo Kim9119a412014-04-03 14:47:25 -0700882void resv_map_release(struct kref *ref)
Andy Whitcroft84afd992008-07-23 21:27:32 -0700883{
884 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
Mike Kravetz5e911372015-09-08 15:01:28 -0700885 struct list_head *head = &resv_map->region_cache;
886 struct file_region *rg, *trg;
Andy Whitcroft84afd992008-07-23 21:27:32 -0700887
888 /* Clear out any active regions before we release the map. */
Mike Kravetzfeba16e2015-09-08 15:01:31 -0700889 region_del(resv_map, 0, LONG_MAX);
Mike Kravetz5e911372015-09-08 15:01:28 -0700890
891 /* ... and any entries left in the cache */
892 list_for_each_entry_safe(rg, trg, head, link) {
893 list_del(&rg->link);
894 kfree(rg);
895 }
896
897 VM_BUG_ON(resv_map->adds_in_progress);
898
Andy Whitcroft84afd992008-07-23 21:27:32 -0700899 kfree(resv_map);
900}
901
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700902static inline struct resv_map *inode_resv_map(struct inode *inode)
903{
Mike Kravetzf27a5132019-05-13 17:22:55 -0700904 /*
905 * At inode evict time, i_mapping may not point to the original
906 * address space within the inode. This original address space
907 * contains the pointer to the resv_map. So, always use the
908 * address space embedded within the inode.
909 * The VERY common case is inode->mapping == &inode->i_data but,
910 * this may not be true for device special inodes.
911 */
912 return (struct resv_map *)(&inode->i_data)->private_data;
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700913}
914
Andy Whitcroft84afd992008-07-23 21:27:32 -0700915static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
Mel Gormana1e78772008-07-23 21:27:23 -0700916{
Sasha Levin81d1b092014-10-09 15:28:10 -0700917 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700918 if (vma->vm_flags & VM_MAYSHARE) {
919 struct address_space *mapping = vma->vm_file->f_mapping;
920 struct inode *inode = mapping->host;
921
922 return inode_resv_map(inode);
923
924 } else {
Andy Whitcroft84afd992008-07-23 21:27:32 -0700925 return (struct resv_map *)(get_vma_private_data(vma) &
926 ~HPAGE_RESV_MASK);
Joonsoo Kim4e35f482014-04-03 14:47:30 -0700927 }
Mel Gormana1e78772008-07-23 21:27:23 -0700928}
929
Andy Whitcroft84afd992008-07-23 21:27:32 -0700930static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
Mel Gormana1e78772008-07-23 21:27:23 -0700931{
Sasha Levin81d1b092014-10-09 15:28:10 -0700932 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
933 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Mel Gormana1e78772008-07-23 21:27:23 -0700934
Andy Whitcroft84afd992008-07-23 21:27:32 -0700935 set_vma_private_data(vma, (get_vma_private_data(vma) &
936 HPAGE_RESV_MASK) | (unsigned long)map);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700937}
938
939static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
940{
Sasha Levin81d1b092014-10-09 15:28:10 -0700941 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
942 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700943
944 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700945}
946
947static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
948{
Sasha Levin81d1b092014-10-09 15:28:10 -0700949 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -0700950
951 return (get_vma_private_data(vma) & flag) != 0;
Mel Gormana1e78772008-07-23 21:27:23 -0700952}
953
Mel Gorman04f2cbe2008-07-23 21:27:25 -0700954/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
Mel Gormana1e78772008-07-23 21:27:23 -0700955void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
956{
Sasha Levin81d1b092014-10-09 15:28:10 -0700957 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
Mel Gormanf83a2752009-05-28 14:34:40 -0700958 if (!(vma->vm_flags & VM_MAYSHARE))
Mel Gormana1e78772008-07-23 21:27:23 -0700959 vma->vm_private_data = (void *)0;
960}
961
962/* Returns true if the VMA has associated reserve pages */
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700963static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
Mel Gormana1e78772008-07-23 21:27:23 -0700964{
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700965 if (vma->vm_flags & VM_NORESERVE) {
966 /*
967 * This address is already reserved by other process(chg == 0),
968 * so, we should decrement reserved count. Without decrementing,
969 * reserve count remains after releasing inode, because this
970 * allocated page will go into page cache and is regarded as
971 * coming from reserved pool in releasing step. Currently, we
972 * don't have any other solution to deal with this situation
973 * properly, so add work-around here.
974 */
975 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700976 return true;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700977 else
Nicholas Krause559ec2f2015-09-04 15:48:27 -0700978 return false;
Joonsoo Kimaf0ed732013-09-11 14:21:18 -0700979 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700980
981 /* Shared mappings always use reserves */
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700982 if (vma->vm_flags & VM_MAYSHARE) {
983 /*
984 * We know VM_NORESERVE is not set. Therefore, there SHOULD
985 * be a region map for all pages. The only situation where
986 * there is no region map is if a hole was punched via
Ethon Paul7c8de352020-06-04 16:49:07 -0700987 * fallocate. In this case, there really are no reserves to
Mike Kravetz1fb1b0e2015-09-08 15:01:44 -0700988 * use. This situation is indicated if chg != 0.
989 */
990 if (chg)
991 return false;
992 else
993 return true;
994 }
Joonsoo Kima63884e2013-09-11 14:21:07 -0700995
996 /*
997 * Only the process that called mmap() has reserves for
998 * private mappings.
999 */
Mike Kravetz67961f92016-06-08 15:33:42 -07001000 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1001 /*
1002 * Like the shared case above, a hole punch or truncate
1003 * could have been performed on the private mapping.
1004 * Examine the value of chg to determine if reserves
1005 * actually exist or were previously consumed.
1006 * Very Subtle - The value of chg comes from a previous
1007 * call to vma_needs_reserves(). The reserve map for
1008 * private mappings has different (opposite) semantics
1009 * than that of shared mappings. vma_needs_reserves()
1010 * has already taken this difference in semantics into
1011 * account. Therefore, the meaning of chg is the same
1012 * as in the shared case above. Code could easily be
1013 * combined, but keeping it separate draws attention to
1014 * subtle differences.
1015 */
1016 if (chg)
1017 return false;
1018 else
1019 return true;
1020 }
Joonsoo Kima63884e2013-09-11 14:21:07 -07001021
Nicholas Krause559ec2f2015-09-04 15:48:27 -07001022 return false;
Mel Gormana1e78772008-07-23 21:27:23 -07001023}
1024
Andi Kleena5516432008-07-23 21:27:41 -07001025static void enqueue_huge_page(struct hstate *h, struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026{
1027 int nid = page_to_nid(page);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001028 list_move(&page->lru, &h->hugepage_freelists[nid]);
Andi Kleena5516432008-07-23 21:27:41 -07001029 h->free_huge_pages++;
1030 h->free_huge_pages_node[nid]++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031}
1032
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001033static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001034{
1035 struct page *page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001036 bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA);
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001037
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001038 list_for_each_entry(page, &h->hugepage_freelists[nid], lru) {
1039 if (nocma && is_migrate_cma_page(page))
1040 continue;
1041
Wei Yang6664bfc2020-10-13 16:56:39 -07001042 if (PageHWPoison(page))
1043 continue;
1044
1045 list_move(&page->lru, &h->hugepage_activelist);
1046 set_page_refcounted(page);
1047 h->free_huge_pages--;
1048 h->free_huge_pages_node[nid]--;
1049 return page;
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001050 }
1051
Wei Yang6664bfc2020-10-13 16:56:39 -07001052 return NULL;
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09001053}
1054
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001055static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
1056 nodemask_t *nmask)
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001057{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001058 unsigned int cpuset_mems_cookie;
1059 struct zonelist *zonelist;
1060 struct zone *zone;
1061 struct zoneref *z;
Anshuman Khandual98fa15f2019-03-05 15:42:58 -08001062 int node = NUMA_NO_NODE;
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001063
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001064 zonelist = node_zonelist(nid, gfp_mask);
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001065
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001066retry_cpuset:
1067 cpuset_mems_cookie = read_mems_allowed_begin();
1068 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1069 struct page *page;
1070
1071 if (!cpuset_zone_allowed(zone, gfp_mask))
1072 continue;
1073 /*
1074 * no need to ask again on the same node. Pool is node rather than
1075 * zone aware
1076 */
1077 if (zone_to_nid(zone) == node)
1078 continue;
1079 node = zone_to_nid(zone);
1080
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001081 page = dequeue_huge_page_node_exact(h, node);
1082 if (page)
1083 return page;
1084 }
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001085 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1086 goto retry_cpuset;
1087
Anshuman Khandual94310cb2017-07-06 15:38:38 -07001088 return NULL;
1089}
1090
Andi Kleena5516432008-07-23 21:27:41 -07001091static struct page *dequeue_huge_page_vma(struct hstate *h,
1092 struct vm_area_struct *vma,
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001093 unsigned long address, int avoid_reserve,
1094 long chg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095{
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001096 struct page *page;
Lee Schermerhorn480eccf2007-09-18 22:46:47 -07001097 struct mempolicy *mpol;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001098 gfp_t gfp_mask;
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001099 nodemask_t *nodemask;
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001100 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101
Mel Gormana1e78772008-07-23 21:27:23 -07001102 /*
1103 * A child process with MAP_PRIVATE mappings created by their parent
1104 * have no page reserves. This check ensures that reservations are
1105 * not "stolen". The child may still get SIGKILLed
1106 */
Joonsoo Kimaf0ed732013-09-11 14:21:18 -07001107 if (!vma_has_reserves(vma, chg) &&
Andi Kleena5516432008-07-23 21:27:41 -07001108 h->free_huge_pages - h->resv_huge_pages == 0)
Miao Xiec0ff7452010-05-24 14:32:08 -07001109 goto err;
Mel Gormana1e78772008-07-23 21:27:23 -07001110
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001111 /* If reserves cannot be used, ensure enough pages are in the pool */
Andi Kleena5516432008-07-23 21:27:41 -07001112 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
Justin P. Mattock6eab04a2011-04-08 19:49:08 -07001113 goto err;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07001114
Vlastimil Babka04ec6262017-07-06 15:40:03 -07001115 gfp_mask = htlb_alloc_mask(h);
1116 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001117 page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
1118 if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
1119 SetPagePrivate(page);
1120 h->resv_huge_pages--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 }
Mel Gormancc9a6c82012-03-21 16:34:11 -07001122
1123 mpol_cond_put(mpol);
Mel Gormancc9a6c82012-03-21 16:34:11 -07001124 return page;
1125
Miao Xiec0ff7452010-05-24 14:32:08 -07001126err:
Mel Gormancc9a6c82012-03-21 16:34:11 -07001127 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128}
1129
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001130/*
1131 * common helper functions for hstate_next_node_to_{alloc|free}.
1132 * We may have allocated or freed a huge page based on a different
1133 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1134 * be outside of *nodes_allowed. Ensure that we use an allowed
1135 * node for alloc or free.
1136 */
1137static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1138{
Andrew Morton0edaf862016-05-19 17:10:58 -07001139 nid = next_node_in(nid, *nodes_allowed);
Luiz Capitulino1cac6f22014-06-04 16:07:11 -07001140 VM_BUG_ON(nid >= MAX_NUMNODES);
1141
1142 return nid;
1143}
1144
1145static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1146{
1147 if (!node_isset(nid, *nodes_allowed))
1148 nid = next_node_allowed(nid, nodes_allowed);
1149 return nid;
1150}
1151
1152/*
1153 * returns the previously saved node ["this node"] from which to
1154 * allocate a persistent huge page for the pool and advance the
1155 * next node from which to allocate, handling wrap at end of node
1156 * mask.
1157 */
1158static int hstate_next_node_to_alloc(struct hstate *h,
1159 nodemask_t *nodes_allowed)
1160{
1161 int nid;
1162
1163 VM_BUG_ON(!nodes_allowed);
1164
1165 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1166 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1167
1168 return nid;
1169}
1170
1171/*
1172 * helper for free_pool_huge_page() - return the previously saved
1173 * node ["this node"] from which to free a huge page. Advance the
1174 * next node id whether or not we find a free huge page to free so
1175 * that the next attempt to free addresses the next node.
1176 */
1177static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1178{
1179 int nid;
1180
1181 VM_BUG_ON(!nodes_allowed);
1182
1183 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1184 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1185
1186 return nid;
1187}
1188
1189#define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1190 for (nr_nodes = nodes_weight(*mask); \
1191 nr_nodes > 0 && \
1192 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1193 nr_nodes--)
1194
1195#define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1196 for (nr_nodes = nodes_weight(*mask); \
1197 nr_nodes > 0 && \
1198 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1199 nr_nodes--)
1200
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001201#ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001202static void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001203 unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001204{
1205 int i;
1206 int nr_pages = 1 << order;
1207 struct page *p = page + 1;
1208
Gerald Schaeferc8cc7082016-06-24 14:50:04 -07001209 atomic_set(compound_mapcount_ptr(page), 0);
John Hubbard47e29d32020-04-01 21:05:33 -07001210 if (hpage_pincount_available(page))
1211 atomic_set(compound_pincount_ptr(page), 0);
1212
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001213 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001214 clear_compound_head(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001215 set_page_refcounted(p);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001216 }
1217
1218 set_compound_order(page, 0);
1219 __ClearPageHead(page);
1220}
1221
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001222static void free_gigantic_page(struct page *page, unsigned int order)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001223{
Roman Gushchincf11e852020-04-10 14:32:45 -07001224 /*
1225 * If the page isn't allocated using the cma allocator,
1226 * cma_release() returns false.
1227 */
Barry Songdbda8fe2020-07-23 21:15:30 -07001228#ifdef CONFIG_CMA
1229 if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order))
Roman Gushchincf11e852020-04-10 14:32:45 -07001230 return;
Barry Songdbda8fe2020-07-23 21:15:30 -07001231#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001232
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001233 free_contig_range(page_to_pfn(page), 1 << order);
1234}
1235
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001236#ifdef CONFIG_CONTIG_ALLOC
Michal Hockod9cc948f2018-01-31 16:20:44 -08001237static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1238 int nid, nodemask_t *nodemask)
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001239{
Anshuman Khandual5e27a2d2019-11-30 17:55:06 -08001240 unsigned long nr_pages = 1UL << huge_page_order(h);
Li Xinhai953f0642020-09-04 16:36:10 -07001241 if (nid == NUMA_NO_NODE)
1242 nid = numa_mem_id();
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001243
Barry Songdbda8fe2020-07-23 21:15:30 -07001244#ifdef CONFIG_CMA
1245 {
Roman Gushchincf11e852020-04-10 14:32:45 -07001246 struct page *page;
1247 int node;
1248
Li Xinhai953f0642020-09-04 16:36:10 -07001249 if (hugetlb_cma[nid]) {
1250 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1251 huge_page_order(h), true);
Roman Gushchincf11e852020-04-10 14:32:45 -07001252 if (page)
1253 return page;
1254 }
Li Xinhai953f0642020-09-04 16:36:10 -07001255
1256 if (!(gfp_mask & __GFP_THISNODE)) {
1257 for_each_node_mask(node, *nodemask) {
1258 if (node == nid || !hugetlb_cma[node])
1259 continue;
1260
1261 page = cma_alloc(hugetlb_cma[node], nr_pages,
1262 huge_page_order(h), true);
1263 if (page)
1264 return page;
1265 }
1266 }
Roman Gushchincf11e852020-04-10 14:32:45 -07001267 }
Barry Songdbda8fe2020-07-23 21:15:30 -07001268#endif
Roman Gushchincf11e852020-04-10 14:32:45 -07001269
Anshuman Khandual5e27a2d2019-11-30 17:55:06 -08001270 return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001271}
1272
1273static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001274static void prep_compound_gigantic_page(struct page *page, unsigned int order);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001275#else /* !CONFIG_CONTIG_ALLOC */
1276static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1277 int nid, nodemask_t *nodemask)
1278{
1279 return NULL;
1280}
1281#endif /* CONFIG_CONTIG_ALLOC */
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001282
Aneesh Kumar K.Ve1073d12017-07-06 15:39:17 -07001283#else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
Michal Hockod9cc948f2018-01-31 16:20:44 -08001284static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001285 int nid, nodemask_t *nodemask)
1286{
1287 return NULL;
1288}
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001289static inline void free_gigantic_page(struct page *page, unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001290static inline void destroy_compound_gigantic_page(struct page *page,
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001291 unsigned int order) { }
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001292#endif
1293
Andi Kleena5516432008-07-23 21:27:41 -07001294static void update_and_free_page(struct hstate *h, struct page *page)
Adam Litke6af2acb2007-10-16 01:26:16 -07001295{
1296 int i;
Andi Kleena5516432008-07-23 21:27:41 -07001297
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07001298 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001299 return;
Andy Whitcroft18229df2008-11-06 12:53:27 -08001300
Andi Kleena5516432008-07-23 21:27:41 -07001301 h->nr_huge_pages--;
1302 h->nr_huge_pages_node[page_to_nid(page)]--;
1303 for (i = 0; i < pages_per_huge_page(h); i++) {
Chris Forbes32f84522011-07-25 17:12:14 -07001304 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1305 1 << PG_referenced | 1 << PG_dirty |
Luiz Capitulinoa7407a22014-06-04 16:07:09 -07001306 1 << PG_active | 1 << PG_private |
1307 1 << PG_writeback);
Adam Litke6af2acb2007-10-16 01:26:16 -07001308 }
Sasha Levin309381fea2014-01-23 15:52:54 -08001309 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
Mina Almasry1adc4d42020-04-01 21:11:15 -07001310 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001311 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
Adam Litke6af2acb2007-10-16 01:26:16 -07001312 set_page_refcounted(page);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001313 if (hstate_is_gigantic(h)) {
Roman Gushchincf11e852020-04-10 14:32:45 -07001314 /*
1315 * Temporarily drop the hugetlb_lock, because
1316 * we might block in free_gigantic_page().
1317 */
1318 spin_unlock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001319 destroy_compound_gigantic_page(page, huge_page_order(h));
1320 free_gigantic_page(page, huge_page_order(h));
Roman Gushchincf11e852020-04-10 14:32:45 -07001321 spin_lock(&hugetlb_lock);
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001322 } else {
Luiz Capitulino944d9fe2014-06-04 16:07:13 -07001323 __free_pages(page, huge_page_order(h));
1324 }
Adam Litke6af2acb2007-10-16 01:26:16 -07001325}
1326
Andi Kleene5ff2152008-07-23 21:27:42 -07001327struct hstate *size_to_hstate(unsigned long size)
1328{
1329 struct hstate *h;
1330
1331 for_each_hstate(h) {
1332 if (huge_page_size(h) == size)
1333 return h;
1334 }
1335 return NULL;
1336}
1337
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001338/*
1339 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1340 * to hstate->hugepage_activelist.)
1341 *
1342 * This function can be called for tail pages, but never returns true for them.
1343 */
1344bool page_huge_active(struct page *page)
1345{
1346 VM_BUG_ON_PAGE(!PageHuge(page), page);
1347 return PageHead(page) && PagePrivate(&page[1]);
1348}
1349
1350/* never called for tail page */
1351static void set_page_huge_active(struct page *page)
1352{
1353 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1354 SetPagePrivate(&page[1]);
1355}
1356
1357static void clear_page_huge_active(struct page *page)
1358{
1359 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1360 ClearPagePrivate(&page[1]);
1361}
1362
Michal Hockoab5ac902018-01-31 16:20:48 -08001363/*
1364 * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1365 * code
1366 */
1367static inline bool PageHugeTemporary(struct page *page)
1368{
1369 if (!PageHuge(page))
1370 return false;
1371
1372 return (unsigned long)page[2].mapping == -1U;
1373}
1374
1375static inline void SetPageHugeTemporary(struct page *page)
1376{
1377 page[2].mapping = (void *)-1U;
1378}
1379
1380static inline void ClearPageHugeTemporary(struct page *page)
1381{
1382 page[2].mapping = NULL;
1383}
1384
Waiman Longc77c0a82020-01-04 13:00:15 -08001385static void __free_huge_page(struct page *page)
David Gibson27a85ef2006-03-22 00:08:56 -08001386{
Andi Kleena5516432008-07-23 21:27:41 -07001387 /*
1388 * Can't pass hstate in here because it is called from the
1389 * compound page destructor.
1390 */
Andi Kleene5ff2152008-07-23 21:27:42 -07001391 struct hstate *h = page_hstate(page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001392 int nid = page_to_nid(page);
David Gibson90481622012-03-21 16:34:12 -07001393 struct hugepage_subpool *spool =
1394 (struct hugepage_subpool *)page_private(page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001395 bool restore_reserve;
David Gibson27a85ef2006-03-22 00:08:56 -08001396
Mike Kravetzb4330af2016-02-05 15:36:38 -08001397 VM_BUG_ON_PAGE(page_count(page), page);
1398 VM_BUG_ON_PAGE(page_mapcount(page), page);
Yongkai Wu8ace22b2018-12-14 14:17:10 -08001399
1400 set_page_private(page, 0);
1401 page->mapping = NULL;
Joonsoo Kim07443a82013-09-11 14:21:58 -07001402 restore_reserve = PagePrivate(page);
Joonsoo Kim16c794b2013-10-16 13:46:48 -07001403 ClearPagePrivate(page);
David Gibson27a85ef2006-03-22 00:08:56 -08001404
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001405 /*
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001406 * If PagePrivate() was set on page, page allocation consumed a
1407 * reservation. If the page was associated with a subpool, there
1408 * would have been a page reserved in the subpool before allocation
1409 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1410 * reservtion, do not call hugepage_subpool_put_pages() as this will
1411 * remove the reserved page from the subpool.
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001412 */
Mike Kravetz0919e1b2019-05-13 17:19:38 -07001413 if (!restore_reserve) {
1414 /*
1415 * A return code of zero implies that the subpool will be
1416 * under its minimum size if the reservation is not restored
1417 * after page is free. Therefore, force restore_reserve
1418 * operation.
1419 */
1420 if (hugepage_subpool_put_pages(spool, 1) == 0)
1421 restore_reserve = true;
1422 }
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07001423
David Gibson27a85ef2006-03-22 00:08:56 -08001424 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07001425 clear_page_huge_active(page);
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07001426 hugetlb_cgroup_uncharge_page(hstate_index(h),
1427 pages_per_huge_page(h), page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07001428 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
1429 pages_per_huge_page(h), page);
Joonsoo Kim07443a82013-09-11 14:21:58 -07001430 if (restore_reserve)
1431 h->resv_huge_pages++;
1432
Michal Hockoab5ac902018-01-31 16:20:48 -08001433 if (PageHugeTemporary(page)) {
1434 list_del(&page->lru);
1435 ClearPageHugeTemporary(page);
1436 update_and_free_page(h, page);
1437 } else if (h->surplus_huge_pages_node[nid]) {
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001438 /* remove the page from active list */
1439 list_del(&page->lru);
Andi Kleena5516432008-07-23 21:27:41 -07001440 update_and_free_page(h, page);
1441 h->surplus_huge_pages--;
1442 h->surplus_huge_pages_node[nid]--;
Adam Litke7893d1d2007-10-16 01:26:18 -07001443 } else {
Will Deacon5d3a5512012-10-08 16:29:32 -07001444 arch_clear_hugepage_flags(page);
Andi Kleena5516432008-07-23 21:27:41 -07001445 enqueue_huge_page(h, page);
Adam Litke7893d1d2007-10-16 01:26:18 -07001446 }
David Gibson27a85ef2006-03-22 00:08:56 -08001447 spin_unlock(&hugetlb_lock);
1448}
1449
Waiman Longc77c0a82020-01-04 13:00:15 -08001450/*
1451 * As free_huge_page() can be called from a non-task context, we have
1452 * to defer the actual freeing in a workqueue to prevent potential
1453 * hugetlb_lock deadlock.
1454 *
1455 * free_hpage_workfn() locklessly retrieves the linked list of pages to
1456 * be freed and frees them one-by-one. As the page->mapping pointer is
1457 * going to be cleared in __free_huge_page() anyway, it is reused as the
1458 * llist_node structure of a lockless linked list of huge pages to be freed.
1459 */
1460static LLIST_HEAD(hpage_freelist);
1461
1462static void free_hpage_workfn(struct work_struct *work)
1463{
1464 struct llist_node *node;
1465 struct page *page;
1466
1467 node = llist_del_all(&hpage_freelist);
1468
1469 while (node) {
1470 page = container_of((struct address_space **)node,
1471 struct page, mapping);
1472 node = node->next;
1473 __free_huge_page(page);
1474 }
1475}
1476static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1477
1478void free_huge_page(struct page *page)
1479{
1480 /*
1481 * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1482 */
1483 if (!in_task()) {
1484 /*
1485 * Only call schedule_work() if hpage_freelist is previously
1486 * empty. Otherwise, schedule_work() had been called but the
1487 * workfn hasn't retrieved the list yet.
1488 */
1489 if (llist_add((struct llist_node *)&page->mapping,
1490 &hpage_freelist))
1491 schedule_work(&free_hpage_work);
1492 return;
1493 }
1494
1495 __free_huge_page(page);
1496}
1497
Andi Kleena5516432008-07-23 21:27:41 -07001498static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001499{
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07001500 INIT_LIST_HEAD(&page->lru);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001501 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
Aneesh Kumar K.V9dd540e2012-07-31 16:42:15 -07001502 set_hugetlb_cgroup(page, NULL);
Mina Almasry1adc4d42020-04-01 21:11:15 -07001503 set_hugetlb_cgroup_rsvd(page, NULL);
Wei Yang2f375112020-10-13 16:56:36 -07001504 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07001505 h->nr_huge_pages++;
1506 h->nr_huge_pages_node[nid]++;
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001507 spin_unlock(&hugetlb_lock);
Andi Kleenb7ba30c2008-07-23 21:27:40 -07001508}
1509
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08001510static void prep_compound_gigantic_page(struct page *page, unsigned int order)
Wu Fengguang20a03072009-06-16 15:32:22 -07001511{
1512 int i;
1513 int nr_pages = 1 << order;
1514 struct page *p = page + 1;
1515
1516 /* we rely on prep_new_huge_page to set the destructor */
1517 set_compound_order(page, order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001518 __ClearPageReserved(page);
Kirill A. Shutemovde09d312016-01-15 16:51:42 -08001519 __SetPageHead(page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001520 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001521 /*
1522 * For gigantic hugepages allocated through bootmem at
1523 * boot, it's safer to be consistent with the not-gigantic
1524 * hugepages and clear the PG_reserved bit from all tail pages
Ethon Paul7c8de352020-06-04 16:49:07 -07001525 * too. Otherwise drivers using get_user_pages() to access tail
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07001526 * pages may get the reference counting wrong if they see
1527 * PG_reserved set on a tail page (despite the head page not
1528 * having PG_reserved set). Enforcing this consistency between
1529 * head and tail pages allows drivers to optimize away a check
1530 * on the head page when they need know if put_page() is needed
1531 * after get_user_pages().
1532 */
1533 __ClearPageReserved(p);
Youquan Song58a84aa2011-12-08 14:34:18 -08001534 set_page_count(p, 0);
Kirill A. Shutemov1d798ca2015-11-06 16:29:54 -08001535 set_compound_head(p, page);
Wu Fengguang20a03072009-06-16 15:32:22 -07001536 }
Mike Kravetzb4330af2016-02-05 15:36:38 -08001537 atomic_set(compound_mapcount_ptr(page), -1);
John Hubbard47e29d32020-04-01 21:05:33 -07001538
1539 if (hpage_pincount_available(page))
1540 atomic_set(compound_pincount_ptr(page), 0);
Wu Fengguang20a03072009-06-16 15:32:22 -07001541}
1542
Andrew Morton77959122012-10-08 16:34:11 -07001543/*
1544 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1545 * transparent huge pages. See the PageTransHuge() documentation for more
1546 * details.
1547 */
Wu Fengguang20a03072009-06-16 15:32:22 -07001548int PageHuge(struct page *page)
1549{
Wu Fengguang20a03072009-06-16 15:32:22 -07001550 if (!PageCompound(page))
1551 return 0;
1552
1553 page = compound_head(page);
Kirill A. Shutemovf1e61552015-11-06 16:29:50 -08001554 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
Wu Fengguang20a03072009-06-16 15:32:22 -07001555}
Naoya Horiguchi43131e12010-05-28 09:29:22 +09001556EXPORT_SYMBOL_GPL(PageHuge);
1557
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001558/*
1559 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1560 * normal or transparent huge pages.
1561 */
1562int PageHeadHuge(struct page *page_head)
1563{
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001564 if (!PageHead(page_head))
1565 return 0;
1566
Vlastimil Babkad4af73e2020-04-01 21:11:48 -07001567 return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR;
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001568}
Andrea Arcangeli27c73ae2013-11-21 14:32:02 -08001569
Mike Kravetzc0d03812020-04-01 21:11:05 -07001570/*
1571 * Find address_space associated with hugetlbfs page.
1572 * Upon entry page is locked and page 'was' mapped although mapped state
1573 * could change. If necessary, use anon_vma to find vma and associated
1574 * address space. The returned mapping may be stale, but it can not be
1575 * invalid as page lock (which is held) is required to destroy mapping.
1576 */
1577static struct address_space *_get_hugetlb_page_mapping(struct page *hpage)
1578{
1579 struct anon_vma *anon_vma;
1580 pgoff_t pgoff_start, pgoff_end;
1581 struct anon_vma_chain *avc;
1582 struct address_space *mapping = page_mapping(hpage);
1583
1584 /* Simple file based mapping */
1585 if (mapping)
1586 return mapping;
1587
1588 /*
1589 * Even anonymous hugetlbfs mappings are associated with an
1590 * underlying hugetlbfs file (see hugetlb_file_setup in mmap
1591 * code). Find a vma associated with the anonymous vma, and
1592 * use the file pointer to get address_space.
1593 */
1594 anon_vma = page_lock_anon_vma_read(hpage);
1595 if (!anon_vma)
1596 return mapping; /* NULL */
1597
1598 /* Use first found vma */
1599 pgoff_start = page_to_pgoff(hpage);
Mike Kravetz1139d332020-07-03 15:15:18 -07001600 pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1;
Mike Kravetzc0d03812020-04-01 21:11:05 -07001601 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root,
1602 pgoff_start, pgoff_end) {
1603 struct vm_area_struct *vma = avc->vma;
1604
1605 mapping = vma->vm_file->f_mapping;
1606 break;
1607 }
1608
1609 anon_vma_unlock_read(anon_vma);
1610 return mapping;
1611}
1612
1613/*
1614 * Find and lock address space (mapping) in write mode.
1615 *
1616 * Upon entry, the page is locked which allows us to find the mapping
1617 * even in the case of an anon page. However, locking order dictates
1618 * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs
1619 * specific. So, we first try to lock the sema while still holding the
1620 * page lock. If this works, great! If not, then we need to drop the
1621 * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of
1622 * course, need to revalidate state along the way.
1623 */
1624struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
1625{
1626 struct address_space *mapping, *mapping2;
1627
1628 mapping = _get_hugetlb_page_mapping(hpage);
1629retry:
1630 if (!mapping)
1631 return mapping;
1632
1633 /*
1634 * If no contention, take lock and return
1635 */
1636 if (i_mmap_trylock_write(mapping))
1637 return mapping;
1638
1639 /*
1640 * Must drop page lock and wait on mapping sema.
1641 * Note: Once page lock is dropped, mapping could become invalid.
1642 * As a hack, increase map count until we lock page again.
1643 */
1644 atomic_inc(&hpage->_mapcount);
1645 unlock_page(hpage);
1646 i_mmap_lock_write(mapping);
1647 lock_page(hpage);
1648 atomic_add_negative(-1, &hpage->_mapcount);
1649
1650 /* verify page is still mapped */
1651 if (!page_mapped(hpage)) {
1652 i_mmap_unlock_write(mapping);
1653 return NULL;
1654 }
1655
1656 /*
1657 * Get address space again and verify it is the same one
1658 * we locked. If not, drop lock and retry.
1659 */
1660 mapping2 = _get_hugetlb_page_mapping(hpage);
1661 if (mapping2 != mapping) {
1662 i_mmap_unlock_write(mapping);
1663 mapping = mapping2;
1664 goto retry;
1665 }
1666
1667 return mapping;
1668}
1669
Zhang Yi13d60f42013-06-25 21:19:31 +08001670pgoff_t __basepage_index(struct page *page)
1671{
1672 struct page *page_head = compound_head(page);
1673 pgoff_t index = page_index(page_head);
1674 unsigned long compound_idx;
1675
1676 if (!PageHuge(page_head))
1677 return page_index(page);
1678
1679 if (compound_order(page_head) >= MAX_ORDER)
1680 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1681 else
1682 compound_idx = page - page_head;
1683
1684 return (index << compound_order(page_head)) + compound_idx;
1685}
1686
Michal Hocko0c397da2018-01-31 16:20:56 -08001687static struct page *alloc_buddy_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001688 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1689 nodemask_t *node_alloc_noretry)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690{
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001691 int order = huge_page_order(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 struct page *page;
Mike Kravetzf60858f2019-09-23 15:37:35 -07001693 bool alloc_try_hard = true;
Joe Jinf96efd52007-07-15 23:38:12 -07001694
Mike Kravetzf60858f2019-09-23 15:37:35 -07001695 /*
1696 * By default we always try hard to allocate the page with
1697 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1698 * a loop (to adjust global huge page counts) and previous allocation
1699 * failed, do not continue to try hard on the same node. Use the
1700 * node_alloc_noretry bitmap to manage this state information.
1701 */
1702 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1703 alloc_try_hard = false;
1704 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1705 if (alloc_try_hard)
1706 gfp_mask |= __GFP_RETRY_MAYFAIL;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001707 if (nid == NUMA_NO_NODE)
1708 nid = numa_mem_id();
1709 page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1710 if (page)
1711 __count_vm_event(HTLB_BUDDY_PGALLOC);
1712 else
1713 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001714
Mike Kravetzf60858f2019-09-23 15:37:35 -07001715 /*
1716 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1717 * indicates an overall state change. Clear bit so that we resume
1718 * normal 'try hard' allocations.
1719 */
1720 if (node_alloc_noretry && page && !alloc_try_hard)
1721 node_clear(nid, *node_alloc_noretry);
1722
1723 /*
1724 * If we tried hard to get a page but failed, set bit so that
1725 * subsequent attempts will not try as hard until there is an
1726 * overall state change.
1727 */
1728 if (node_alloc_noretry && !page && alloc_try_hard)
1729 node_set(nid, *node_alloc_noretry);
1730
Nishanth Aravamudan63b46132007-10-16 01:26:24 -07001731 return page;
1732}
1733
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001734/*
Michal Hocko0c397da2018-01-31 16:20:56 -08001735 * Common helper to allocate a fresh hugetlb page. All specific allocators
1736 * should use this function to get new hugetlb pages
1737 */
1738static struct page *alloc_fresh_huge_page(struct hstate *h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001739 gfp_t gfp_mask, int nid, nodemask_t *nmask,
1740 nodemask_t *node_alloc_noretry)
Michal Hocko0c397da2018-01-31 16:20:56 -08001741{
1742 struct page *page;
1743
1744 if (hstate_is_gigantic(h))
1745 page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1746 else
1747 page = alloc_buddy_huge_page(h, gfp_mask,
Mike Kravetzf60858f2019-09-23 15:37:35 -07001748 nid, nmask, node_alloc_noretry);
Michal Hocko0c397da2018-01-31 16:20:56 -08001749 if (!page)
1750 return NULL;
1751
1752 if (hstate_is_gigantic(h))
1753 prep_compound_gigantic_page(page, huge_page_order(h));
1754 prep_new_huge_page(h, page, page_to_nid(page));
1755
1756 return page;
1757}
1758
1759/*
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001760 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1761 * manner.
1762 */
Mike Kravetzf60858f2019-09-23 15:37:35 -07001763static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1764 nodemask_t *node_alloc_noretry)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001765{
1766 struct page *page;
1767 int nr_nodes, node;
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001768 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001769
1770 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
Mike Kravetzf60858f2019-09-23 15:37:35 -07001771 page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1772 node_alloc_noretry);
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001773 if (page)
Joonsoo Kimb2261022013-09-11 14:21:00 -07001774 break;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001775 }
1776
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001777 if (!page)
1778 return 0;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001779
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08001780 put_page(page); /* free it into the hugepage allocator */
1781
1782 return 1;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001783}
1784
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001785/*
1786 * Free huge page from pool from next node to free.
1787 * Attempt to keep persistent huge pages more or less
1788 * balanced over allowed nodes.
1789 * Called with hugetlb_lock locked.
1790 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08001791static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1792 bool acct_surplus)
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001793{
Joonsoo Kimb2261022013-09-11 14:21:00 -07001794 int nr_nodes, node;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001795 int ret = 0;
1796
Joonsoo Kimb2261022013-09-11 14:21:00 -07001797 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001798 /*
1799 * If we're returning unused surplus pages, only examine
1800 * nodes with surplus pages.
1801 */
Joonsoo Kimb2261022013-09-11 14:21:00 -07001802 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1803 !list_empty(&h->hugepage_freelists[node])) {
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001804 struct page *page =
Joonsoo Kimb2261022013-09-11 14:21:00 -07001805 list_entry(h->hugepage_freelists[node].next,
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001806 struct page, lru);
1807 list_del(&page->lru);
1808 h->free_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001809 h->free_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001810 if (acct_surplus) {
1811 h->surplus_huge_pages--;
Joonsoo Kimb2261022013-09-11 14:21:00 -07001812 h->surplus_huge_pages_node[node]--;
Lee Schermerhorn685f3452009-09-21 17:01:23 -07001813 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001814 update_and_free_page(h, page);
1815 ret = 1;
Lee Schermerhorn9a76db02009-12-14 17:58:15 -08001816 break;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001817 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07001818 }
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07001819
1820 return ret;
1821}
1822
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001823/*
1824 * Dissolve a given free hugepage into free buddy pages. This function does
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001825 * nothing for in-use hugepages and non-hugepages.
1826 * This function returns values like below:
1827 *
1828 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1829 * (allocated or reserved.)
1830 * 0: successfully dissolved free hugepages or the page is not a
1831 * hugepage (considered as already dissolved)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001832 */
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001833int dissolve_free_huge_page(struct page *page)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001834{
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001835 int rc = -EBUSY;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001836
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001837 /* Not to disrupt normal path by vainly holding hugetlb_lock */
1838 if (!PageHuge(page))
1839 return 0;
1840
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001841 spin_lock(&hugetlb_lock);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001842 if (!PageHuge(page)) {
1843 rc = 0;
1844 goto out;
1845 }
1846
1847 if (!page_count(page)) {
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001848 struct page *head = compound_head(page);
1849 struct hstate *h = page_hstate(head);
1850 int nid = page_to_nid(head);
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001851 if (h->free_huge_pages - h->resv_huge_pages == 0)
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001852 goto out;
Anshuman Khandualc3114a82017-07-10 15:47:41 -07001853 /*
1854 * Move PageHWPoison flag from head page to the raw error page,
1855 * which makes any subpages rather than the error page reusable.
1856 */
1857 if (PageHWPoison(head) && page != head) {
1858 SetPageHWPoison(page);
1859 ClearPageHWPoison(head);
1860 }
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001861 list_del(&head->lru);
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001862 h->free_huge_pages--;
1863 h->free_huge_pages_node[nid]--;
zhong jiangc1470b32016-08-11 15:32:55 -07001864 h->max_huge_pages--;
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001865 update_and_free_page(h, head);
Naoya Horiguchi6bc9b562018-08-23 17:00:38 -07001866 rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001867 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001868out:
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001869 spin_unlock(&hugetlb_lock);
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001870 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001871}
1872
1873/*
1874 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1875 * make specified memory blocks removable from the system.
Gerald Schaefer2247bb32016-10-07 17:01:07 -07001876 * Note that this will dissolve a free gigantic hugepage completely, if any
1877 * part of it lies within the given range.
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001878 * Also note that if dissolve_free_huge_page() returns with an error, all
1879 * free hugepages that were dissolved before that error are lost.
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001880 */
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001881int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001882{
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001883 unsigned long pfn;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001884 struct page *page;
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001885 int rc = 0;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001886
Li Zhongd0177632014-08-06 16:07:56 -07001887 if (!hugepages_supported())
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001888 return rc;
Li Zhongd0177632014-08-06 16:07:56 -07001889
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001890 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1891 page = pfn_to_page(pfn);
Naoya Horiguchifaf53de2019-06-28 12:06:56 -07001892 rc = dissolve_free_huge_page(page);
1893 if (rc)
1894 break;
Gerald Schaefereb03aa02016-10-07 17:01:13 -07001895 }
Gerald Schaefer082d5b62016-10-07 17:01:10 -07001896
1897 return rc;
Naoya Horiguchic8721bb2013-09-11 14:22:09 -07001898}
1899
Michal Hockoab5ac902018-01-31 16:20:48 -08001900/*
1901 * Allocates a fresh surplus page from the page allocator.
1902 */
Michal Hocko0c397da2018-01-31 16:20:56 -08001903static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
Michal Hockoaaf14e42017-07-10 15:49:08 -07001904 int nid, nodemask_t *nmask)
Adam Litke7893d1d2007-10-16 01:26:18 -07001905{
Michal Hocko9980d742018-01-31 16:20:52 -08001906 struct page *page = NULL;
Adam Litke7893d1d2007-10-16 01:26:18 -07001907
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07001908 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07001909 return NULL;
1910
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001911 spin_lock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001912 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1913 goto out_unlock;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001914 spin_unlock(&hugetlb_lock);
1915
Mike Kravetzf60858f2019-09-23 15:37:35 -07001916 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hocko9980d742018-01-31 16:20:52 -08001917 if (!page)
Michal Hocko0c397da2018-01-31 16:20:56 -08001918 return NULL;
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001919
1920 spin_lock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001921 /*
1922 * We could have raced with the pool size change.
1923 * Double check that and simply deallocate the new page
1924 * if we would end up overcommiting the surpluses. Abuse
1925 * temporary page to workaround the nasty free_huge_page
1926 * codeflow
1927 */
1928 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1929 SetPageHugeTemporary(page);
Kai Shen2bf753e2019-05-13 17:15:37 -07001930 spin_unlock(&hugetlb_lock);
Michal Hocko9980d742018-01-31 16:20:52 -08001931 put_page(page);
Kai Shen2bf753e2019-05-13 17:15:37 -07001932 return NULL;
Michal Hocko9980d742018-01-31 16:20:52 -08001933 } else {
Michal Hocko9980d742018-01-31 16:20:52 -08001934 h->surplus_huge_pages++;
Michal Hocko4704dea2018-03-09 15:50:55 -08001935 h->surplus_huge_pages_node[page_to_nid(page)]++;
Adam Litke7893d1d2007-10-16 01:26:18 -07001936 }
Michal Hocko9980d742018-01-31 16:20:52 -08001937
1938out_unlock:
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08001939 spin_unlock(&hugetlb_lock);
Adam Litke7893d1d2007-10-16 01:26:18 -07001940
1941 return page;
1942}
1943
Joonsoo Kimbbe88752020-08-11 18:37:38 -07001944static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
Aneesh Kumar K.V9a4e9f32019-03-05 15:47:44 -08001945 int nid, nodemask_t *nmask)
Michal Hockoab5ac902018-01-31 16:20:48 -08001946{
1947 struct page *page;
1948
1949 if (hstate_is_gigantic(h))
1950 return NULL;
1951
Mike Kravetzf60858f2019-09-23 15:37:35 -07001952 page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
Michal Hockoab5ac902018-01-31 16:20:48 -08001953 if (!page)
1954 return NULL;
1955
1956 /*
1957 * We do not account these pages as surplus because they are only
1958 * temporary and will be released properly on the last reference
1959 */
Michal Hockoab5ac902018-01-31 16:20:48 -08001960 SetPageHugeTemporary(page);
1961
1962 return page;
1963}
1964
Adam Litkee4e574b2007-10-16 01:26:19 -07001965/*
Dave Hansen099730d2015-11-05 18:50:17 -08001966 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1967 */
Dave Hansene0ec90e2015-11-05 18:50:20 -08001968static
Michal Hocko0c397da2018-01-31 16:20:56 -08001969struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
Dave Hansen099730d2015-11-05 18:50:17 -08001970 struct vm_area_struct *vma, unsigned long addr)
1971{
Michal Hockoaaf14e42017-07-10 15:49:08 -07001972 struct page *page;
1973 struct mempolicy *mpol;
1974 gfp_t gfp_mask = htlb_alloc_mask(h);
1975 int nid;
1976 nodemask_t *nodemask;
1977
1978 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
Michal Hocko0c397da2018-01-31 16:20:56 -08001979 page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
Michal Hockoaaf14e42017-07-10 15:49:08 -07001980 mpol_cond_put(mpol);
1981
1982 return page;
Dave Hansen099730d2015-11-05 18:50:17 -08001983}
1984
Michal Hockoab5ac902018-01-31 16:20:48 -08001985/* page migration callback function */
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001986struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07001987 nodemask_t *nmask, gfp_t gfp_mask)
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001988{
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001989 spin_lock(&hugetlb_lock);
1990 if (h->free_huge_pages - h->resv_huge_pages > 0) {
Michal Hocko3e59fcb2017-07-10 15:49:11 -07001991 struct page *page;
1992
1993 page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1994 if (page) {
1995 spin_unlock(&hugetlb_lock);
1996 return page;
Michal Hocko4db9b2e2017-07-10 15:48:44 -07001997 }
1998 }
1999 spin_unlock(&hugetlb_lock);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07002000
Michal Hocko0c397da2018-01-31 16:20:56 -08002001 return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
Michal Hocko4db9b2e2017-07-10 15:48:44 -07002002}
2003
Michal Hockoebd63722018-01-31 16:21:00 -08002004/* mempolicy aware migration callback */
Michal Hocko389c8172018-01-31 16:21:03 -08002005struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
2006 unsigned long address)
Michal Hockoebd63722018-01-31 16:21:00 -08002007{
2008 struct mempolicy *mpol;
2009 nodemask_t *nodemask;
2010 struct page *page;
Michal Hockoebd63722018-01-31 16:21:00 -08002011 gfp_t gfp_mask;
2012 int node;
2013
Michal Hockoebd63722018-01-31 16:21:00 -08002014 gfp_mask = htlb_alloc_mask(h);
2015 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
Joonsoo Kimd92bbc22020-08-11 18:37:17 -07002016 page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask);
Michal Hockoebd63722018-01-31 16:21:00 -08002017 mpol_cond_put(mpol);
2018
2019 return page;
2020}
2021
Naoya Horiguchibf50bab2010-09-08 10:19:33 +09002022/*
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002023 * Increase the hugetlb pool such that it can accommodate a reservation
Adam Litkee4e574b2007-10-16 01:26:19 -07002024 * of size 'delta'.
2025 */
Andi Kleena5516432008-07-23 21:27:41 -07002026static int gather_surplus_pages(struct hstate *h, int delta)
Jules Irenge1b2a1e72020-04-06 20:08:09 -07002027 __must_hold(&hugetlb_lock)
Adam Litkee4e574b2007-10-16 01:26:19 -07002028{
2029 struct list_head surplus_list;
2030 struct page *page, *tmp;
2031 int ret, i;
2032 int needed, allocated;
Hillf Danton28073b02012-03-21 16:34:00 -07002033 bool alloc_ok = true;
Adam Litkee4e574b2007-10-16 01:26:19 -07002034
Andi Kleena5516432008-07-23 21:27:41 -07002035 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
Adam Litkeac09b3a2008-03-04 14:29:38 -08002036 if (needed <= 0) {
Andi Kleena5516432008-07-23 21:27:41 -07002037 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002038 return 0;
Adam Litkeac09b3a2008-03-04 14:29:38 -08002039 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002040
2041 allocated = 0;
2042 INIT_LIST_HEAD(&surplus_list);
2043
2044 ret = -ENOMEM;
2045retry:
2046 spin_unlock(&hugetlb_lock);
2047 for (i = 0; i < needed; i++) {
Michal Hocko0c397da2018-01-31 16:20:56 -08002048 page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
Michal Hockoaaf14e42017-07-10 15:49:08 -07002049 NUMA_NO_NODE, NULL);
Hillf Danton28073b02012-03-21 16:34:00 -07002050 if (!page) {
2051 alloc_ok = false;
2052 break;
2053 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002054 list_add(&page->lru, &surplus_list);
David Rientjes69ed7792017-07-10 15:48:50 -07002055 cond_resched();
Adam Litkee4e574b2007-10-16 01:26:19 -07002056 }
Hillf Danton28073b02012-03-21 16:34:00 -07002057 allocated += i;
Adam Litkee4e574b2007-10-16 01:26:19 -07002058
2059 /*
2060 * After retaking hugetlb_lock, we need to recalculate 'needed'
2061 * because either resv_huge_pages or free_huge_pages may have changed.
2062 */
2063 spin_lock(&hugetlb_lock);
Andi Kleena5516432008-07-23 21:27:41 -07002064 needed = (h->resv_huge_pages + delta) -
2065 (h->free_huge_pages + allocated);
Hillf Danton28073b02012-03-21 16:34:00 -07002066 if (needed > 0) {
2067 if (alloc_ok)
2068 goto retry;
2069 /*
2070 * We were not able to allocate enough pages to
2071 * satisfy the entire reservation so we free what
2072 * we've allocated so far.
2073 */
2074 goto free;
2075 }
Adam Litkee4e574b2007-10-16 01:26:19 -07002076 /*
2077 * The surplus_list now contains _at_least_ the number of extra pages
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002078 * needed to accommodate the reservation. Add the appropriate number
Adam Litkee4e574b2007-10-16 01:26:19 -07002079 * of pages to the hugetlb pool and free the extras back to the buddy
Adam Litkeac09b3a2008-03-04 14:29:38 -08002080 * allocator. Commit the entire reservation here to prevent another
2081 * process from stealing the pages as they are added to the pool but
2082 * before they are reserved.
Adam Litkee4e574b2007-10-16 01:26:19 -07002083 */
2084 needed += allocated;
Andi Kleena5516432008-07-23 21:27:41 -07002085 h->resv_huge_pages += delta;
Adam Litkee4e574b2007-10-16 01:26:19 -07002086 ret = 0;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002087
Adam Litke19fc3f02008-04-28 02:12:20 -07002088 /* Free the needed pages to the hugetlb pool */
Adam Litkee4e574b2007-10-16 01:26:19 -07002089 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
Adam Litke19fc3f02008-04-28 02:12:20 -07002090 if ((--needed) < 0)
2091 break;
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002092 /*
2093 * This page is now managed by the hugetlb allocator and has
2094 * no users -- drop the buddy allocator's reference.
2095 */
2096 put_page_testzero(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08002097 VM_BUG_ON_PAGE(page_count(page), page);
Andi Kleena5516432008-07-23 21:27:41 -07002098 enqueue_huge_page(h, page);
Adam Litke19fc3f02008-04-28 02:12:20 -07002099 }
Hillf Danton28073b02012-03-21 16:34:00 -07002100free:
Hillf Dantonb0365c82011-12-28 15:57:16 -08002101 spin_unlock(&hugetlb_lock);
Adam Litke19fc3f02008-04-28 02:12:20 -07002102
2103 /* Free unnecessary surplus pages to the buddy allocator */
Joonsoo Kimc0d934b2013-09-11 14:21:02 -07002104 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
2105 put_page(page);
Naoya Horiguchia9869b82010-09-08 10:19:37 +09002106 spin_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002107
2108 return ret;
2109}
2110
2111/*
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002112 * This routine has two main purposes:
2113 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2114 * in unused_resv_pages. This corresponds to the prior adjustments made
2115 * to the associated reservation map.
2116 * 2) Free any unused surplus pages that may have been allocated to satisfy
2117 * the reservation. As many as unused_resv_pages may be freed.
2118 *
2119 * Called with hugetlb_lock held. However, the lock could be dropped (and
2120 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
2121 * we must make sure nobody else can claim pages we are in the process of
2122 * freeing. Do this by ensuring resv_huge_page always is greater than the
2123 * number of huge pages we plan to free when dropping the lock.
Adam Litkee4e574b2007-10-16 01:26:19 -07002124 */
Andi Kleena5516432008-07-23 21:27:41 -07002125static void return_unused_surplus_pages(struct hstate *h,
2126 unsigned long unused_resv_pages)
Adam Litkee4e574b2007-10-16 01:26:19 -07002127{
Adam Litkee4e574b2007-10-16 01:26:19 -07002128 unsigned long nr_pages;
2129
Andi Kleenaa888a72008-07-23 21:27:47 -07002130 /* Cannot return gigantic pages currently */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002131 if (hstate_is_gigantic(h))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002132 goto out;
Andi Kleenaa888a72008-07-23 21:27:47 -07002133
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002134 /*
2135 * Part (or even all) of the reservation could have been backed
2136 * by pre-allocated pages. Only free surplus pages.
2137 */
Andi Kleena5516432008-07-23 21:27:41 -07002138 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
Adam Litkee4e574b2007-10-16 01:26:19 -07002139
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002140 /*
2141 * We want to release as many surplus pages as possible, spread
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002142 * evenly across all nodes with memory. Iterate across these nodes
2143 * until we can no longer free unreserved surplus pages. This occurs
2144 * when the nodes with surplus pages have no free pages.
Randy Dunlap9e7ee402020-08-11 18:32:59 -07002145 * free_pool_huge_page() will balance the freed pages across the
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08002146 * on-line nodes with memory and will handle the hstate accounting.
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002147 *
2148 * Note that we decrement resv_huge_pages as we free the pages. If
2149 * we drop the lock, resv_huge_pages will still be sufficiently large
2150 * to cover subsequent pages we may free.
Lee Schermerhorn685f3452009-09-21 17:01:23 -07002151 */
2152 while (nr_pages--) {
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002153 h->resv_huge_pages--;
2154 unused_resv_pages--;
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08002155 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002156 goto out;
Mizuma, Masayoshi7848a4b2014-04-18 15:07:18 -07002157 cond_resched_lock(&hugetlb_lock);
Adam Litkee4e574b2007-10-16 01:26:19 -07002158 }
Mike Kravetze5bbc8a2017-01-10 16:58:27 -08002159
2160out:
2161 /* Fully uncommit the reservation */
2162 h->resv_huge_pages -= unused_resv_pages;
Adam Litkee4e574b2007-10-16 01:26:19 -07002163}
2164
Mike Kravetz5e911372015-09-08 15:01:28 -07002165
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002166/*
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002167 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
Mike Kravetz5e911372015-09-08 15:01:28 -07002168 * are used by the huge page allocation routines to manage reservations.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002169 *
2170 * vma_needs_reservation is called to determine if the huge page at addr
2171 * within the vma has an associated reservation. If a reservation is
2172 * needed, the value 1 is returned. The caller is then responsible for
2173 * managing the global reservation and subpool usage counts. After
2174 * the huge page has been allocated, vma_commit_reservation is called
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002175 * to add the page to the reservation map. If the page allocation fails,
2176 * the reservation must be ended instead of committed. vma_end_reservation
2177 * is called in such cases.
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002178 *
2179 * In the normal case, vma_commit_reservation returns the same value
2180 * as the preceding vma_needs_reservation call. The only time this
2181 * is not the case is if a reserve map was changed between calls. It
2182 * is the responsibility of the caller to notice the difference and
2183 * take appropriate action.
Mike Kravetz96b96a92016-11-10 10:46:32 -08002184 *
2185 * vma_add_reservation is used in error paths where a reservation must
2186 * be restored when a newly allocated huge page must be freed. It is
2187 * to be called after calling vma_needs_reservation to determine if a
2188 * reservation exists.
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002189 */
Mike Kravetz5e911372015-09-08 15:01:28 -07002190enum vma_resv_mode {
2191 VMA_NEEDS_RESV,
2192 VMA_COMMIT_RESV,
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002193 VMA_END_RESV,
Mike Kravetz96b96a92016-11-10 10:46:32 -08002194 VMA_ADD_RESV,
Mike Kravetz5e911372015-09-08 15:01:28 -07002195};
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002196static long __vma_reservation_common(struct hstate *h,
2197 struct vm_area_struct *vma, unsigned long addr,
Mike Kravetz5e911372015-09-08 15:01:28 -07002198 enum vma_resv_mode mode)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002199{
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002200 struct resv_map *resv;
2201 pgoff_t idx;
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002202 long ret;
Mina Almasry0db9d742020-04-01 21:11:25 -07002203 long dummy_out_regions_needed;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002204
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002205 resv = vma_resv_map(vma);
2206 if (!resv)
Andy Whitcroft84afd992008-07-23 21:27:32 -07002207 return 1;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002208
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002209 idx = vma_hugecache_offset(h, vma, addr);
Mike Kravetz5e911372015-09-08 15:01:28 -07002210 switch (mode) {
2211 case VMA_NEEDS_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002212 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2213 /* We assume that vma_reservation_* routines always operate on
2214 * 1 page, and that adding to resv map a 1 page entry can only
2215 * ever require 1 region.
2216 */
2217 VM_BUG_ON(dummy_out_regions_needed != 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002218 break;
2219 case VMA_COMMIT_RESV:
Mina Almasry075a61d2020-04-01 21:11:28 -07002220 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002221 /* region_add calls of range 1 should never fail. */
2222 VM_BUG_ON(ret < 0);
Mike Kravetz5e911372015-09-08 15:01:28 -07002223 break;
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002224 case VMA_END_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002225 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz5e911372015-09-08 15:01:28 -07002226 ret = 0;
2227 break;
Mike Kravetz96b96a92016-11-10 10:46:32 -08002228 case VMA_ADD_RESV:
Mina Almasry0db9d742020-04-01 21:11:25 -07002229 if (vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07002230 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
Mina Almasry0db9d742020-04-01 21:11:25 -07002231 /* region_add calls of range 1 should never fail. */
2232 VM_BUG_ON(ret < 0);
2233 } else {
2234 region_abort(resv, idx, idx + 1, 1);
Mike Kravetz96b96a92016-11-10 10:46:32 -08002235 ret = region_del(resv, idx, idx + 1);
2236 }
2237 break;
Mike Kravetz5e911372015-09-08 15:01:28 -07002238 default:
2239 BUG();
2240 }
Andy Whitcroft84afd992008-07-23 21:27:32 -07002241
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002242 if (vma->vm_flags & VM_MAYSHARE)
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002243 return ret;
Mike Kravetz67961f92016-06-08 15:33:42 -07002244 else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
2245 /*
2246 * In most cases, reserves always exist for private mappings.
2247 * However, a file associated with mapping could have been
2248 * hole punched or truncated after reserves were consumed.
2249 * As subsequent fault on such a range will not use reserves.
2250 * Subtle - The reserve map for private mappings has the
2251 * opposite meaning than that of shared mappings. If NO
2252 * entry is in the reserve map, it means a reservation exists.
2253 * If an entry exists in the reserve map, it means the
2254 * reservation has already been consumed. As a result, the
2255 * return value of this routine is the opposite of the
2256 * value returned from reserve map manipulation routines above.
2257 */
2258 if (ret)
2259 return 0;
2260 else
2261 return 1;
2262 }
Joonsoo Kim4e35f482014-04-03 14:47:30 -07002263 else
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002264 return ret < 0 ? ret : 0;
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002265}
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002266
2267static long vma_needs_reservation(struct hstate *h,
Andi Kleena5516432008-07-23 21:27:41 -07002268 struct vm_area_struct *vma, unsigned long addr)
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002269{
Mike Kravetz5e911372015-09-08 15:01:28 -07002270 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002271}
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002272
Mike Kravetzcf3ad202015-06-24 16:57:55 -07002273static long vma_commit_reservation(struct hstate *h,
2274 struct vm_area_struct *vma, unsigned long addr)
2275{
Mike Kravetz5e911372015-09-08 15:01:28 -07002276 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2277}
2278
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002279static void vma_end_reservation(struct hstate *h,
Mike Kravetz5e911372015-09-08 15:01:28 -07002280 struct vm_area_struct *vma, unsigned long addr)
2281{
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002282 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
Andy Whitcroftc37f9fb2008-07-23 21:27:30 -07002283}
2284
Mike Kravetz96b96a92016-11-10 10:46:32 -08002285static long vma_add_reservation(struct hstate *h,
2286 struct vm_area_struct *vma, unsigned long addr)
2287{
2288 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2289}
2290
2291/*
2292 * This routine is called to restore a reservation on error paths. In the
2293 * specific error paths, a huge page was allocated (via alloc_huge_page)
2294 * and is about to be freed. If a reservation for the page existed,
2295 * alloc_huge_page would have consumed the reservation and set PagePrivate
2296 * in the newly allocated page. When the page is freed via free_huge_page,
2297 * the global reservation count will be incremented if PagePrivate is set.
2298 * However, free_huge_page can not adjust the reserve map. Adjust the
2299 * reserve map here to be consistent with global reserve count adjustments
2300 * to be made by free_huge_page.
2301 */
2302static void restore_reserve_on_error(struct hstate *h,
2303 struct vm_area_struct *vma, unsigned long address,
2304 struct page *page)
2305{
2306 if (unlikely(PagePrivate(page))) {
2307 long rc = vma_needs_reservation(h, vma, address);
2308
2309 if (unlikely(rc < 0)) {
2310 /*
2311 * Rare out of memory condition in reserve map
2312 * manipulation. Clear PagePrivate so that
2313 * global reserve count will not be incremented
2314 * by free_huge_page. This will make it appear
2315 * as though the reservation for this page was
2316 * consumed. This may prevent the task from
2317 * faulting in the page at a later time. This
2318 * is better than inconsistent global huge page
2319 * accounting of reserve counts.
2320 */
2321 ClearPagePrivate(page);
2322 } else if (rc) {
2323 rc = vma_add_reservation(h, vma, address);
2324 if (unlikely(rc < 0))
2325 /*
2326 * See above comment about rare out of
2327 * memory condition.
2328 */
2329 ClearPagePrivate(page);
2330 } else
2331 vma_end_reservation(h, vma, address);
2332 }
2333}
2334
Mike Kravetz70c35472015-09-08 15:01:54 -07002335struct page *alloc_huge_page(struct vm_area_struct *vma,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07002336 unsigned long addr, int avoid_reserve)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337{
David Gibson90481622012-03-21 16:34:12 -07002338 struct hugepage_subpool *spool = subpool_vma(vma);
Andi Kleena5516432008-07-23 21:27:41 -07002339 struct hstate *h = hstate_vma(vma);
Adam Litke348ea202007-11-14 16:59:37 -08002340 struct page *page;
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002341 long map_chg, map_commit;
2342 long gbl_chg;
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002343 int ret, idx;
2344 struct hugetlb_cgroup *h_cg;
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002345 bool deferred_reserve;
Adam Litke2fc39ce2007-11-14 16:59:39 -08002346
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002347 idx = hstate_index(h);
Mel Gormana1e78772008-07-23 21:27:23 -07002348 /*
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002349 * Examine the region/reserve map to determine if the process
2350 * has a reservation for the page to be allocated. A return
2351 * code of zero indicates a reservation exists (no change).
Mel Gormana1e78772008-07-23 21:27:23 -07002352 */
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002353 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2354 if (map_chg < 0)
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002355 return ERR_PTR(-ENOMEM);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002356
2357 /*
2358 * Processes that did not create the mapping will have no
2359 * reserves as indicated by the region/reserve map. Check
2360 * that the allocation will not exceed the subpool limit.
2361 * Allocations for MAP_NORESERVE mappings also need to be
2362 * checked against any subpool limit.
2363 */
2364 if (map_chg || avoid_reserve) {
2365 gbl_chg = hugepage_subpool_get_pages(spool, 1);
2366 if (gbl_chg < 0) {
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002367 vma_end_reservation(h, vma, addr);
Aneesh Kumar K.V76dcee72012-07-31 16:41:57 -07002368 return ERR_PTR(-ENOSPC);
Mike Kravetz5e911372015-09-08 15:01:28 -07002369 }
Mel Gormana1e78772008-07-23 21:27:23 -07002370
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002371 /*
2372 * Even though there was no reservation in the region/reserve
2373 * map, there could be reservations associated with the
2374 * subpool that can be used. This would be indicated if the
2375 * return value of hugepage_subpool_get_pages() is zero.
2376 * However, if avoid_reserve is specified we still avoid even
2377 * the subpool reservations.
2378 */
2379 if (avoid_reserve)
2380 gbl_chg = 1;
2381 }
2382
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002383 /* If this allocation is not consuming a reservation, charge it now.
2384 */
2385 deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma);
2386 if (deferred_reserve) {
2387 ret = hugetlb_cgroup_charge_cgroup_rsvd(
2388 idx, pages_per_huge_page(h), &h_cg);
2389 if (ret)
2390 goto out_subpool_put;
2391 }
2392
Aneesh Kumar K.V6d76dcf2012-07-31 16:42:18 -07002393 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002394 if (ret)
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002395 goto out_uncharge_cgroup_reservation;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002396
Mel Gormana1e78772008-07-23 21:27:23 -07002397 spin_lock(&hugetlb_lock);
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002398 /*
2399 * glb_chg is passed to indicate whether or not a page must be taken
2400 * from the global free pool (global change). gbl_chg == 0 indicates
2401 * a reservation exists for the allocation.
2402 */
2403 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002404 if (!page) {
Aneesh Kumar K.V94ae8ba2012-07-31 16:42:35 -07002405 spin_unlock(&hugetlb_lock);
Michal Hocko0c397da2018-01-31 16:20:56 -08002406 page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002407 if (!page)
2408 goto out_uncharge_cgroup;
Naoya Horiguchia88c7692015-12-11 13:40:24 -08002409 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2410 SetPagePrivate(page);
2411 h->resv_huge_pages--;
2412 }
Aneesh Kumar K.V79dbb232012-07-31 16:42:32 -07002413 spin_lock(&hugetlb_lock);
Wei Yang15a8d682020-10-13 16:56:33 -07002414 list_add(&page->lru, &h->hugepage_activelist);
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002415 /* Fall through */
Mel Gormana1e78772008-07-23 21:27:23 -07002416 }
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002417 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002418 /* If allocation is not consuming a reservation, also store the
2419 * hugetlb_cgroup pointer on the page.
2420 */
2421 if (deferred_reserve) {
2422 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
2423 h_cg, page);
2424 }
2425
Joonsoo Kim81a6fca2013-09-11 14:20:58 -07002426 spin_unlock(&hugetlb_lock);
Mel Gormana1e78772008-07-23 21:27:23 -07002427
David Gibson90481622012-03-21 16:34:12 -07002428 set_page_private(page, (unsigned long)spool);
Mel Gormana1e78772008-07-23 21:27:23 -07002429
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002430 map_commit = vma_commit_reservation(h, vma, addr);
2431 if (unlikely(map_chg > map_commit)) {
Mike Kravetz33039672015-06-24 16:57:58 -07002432 /*
2433 * The page was added to the reservation map between
2434 * vma_needs_reservation and vma_commit_reservation.
2435 * This indicates a race with hugetlb_reserve_pages.
2436 * Adjust for the subpool count incremented above AND
2437 * in hugetlb_reserve_pages for the same page. Also,
2438 * the reservation count added in hugetlb_reserve_pages
2439 * no longer applies.
2440 */
2441 long rsv_adjust;
2442
2443 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2444 hugetlb_acct_memory(h, -rsv_adjust);
Mike Kravetz79aa9252020-11-01 17:07:27 -08002445 if (deferred_reserve)
2446 hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h),
2447 pages_per_huge_page(h), page);
Mike Kravetz33039672015-06-24 16:57:58 -07002448 }
Adam Litke90d8b7e2007-11-14 16:59:42 -08002449 return page;
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002450
2451out_uncharge_cgroup:
2452 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
Mina Almasry08cf9fa2020-04-01 21:11:31 -07002453out_uncharge_cgroup_reservation:
2454 if (deferred_reserve)
2455 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
2456 h_cg);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002457out_subpool_put:
Mike Kravetzd85f69b2015-09-08 15:01:47 -07002458 if (map_chg || avoid_reserve)
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002459 hugepage_subpool_put_pages(spool, 1);
Mike Kravetzfeba16e2015-09-08 15:01:31 -07002460 vma_end_reservation(h, vma, addr);
Jianyu Zhan8f34af62014-06-04 16:10:36 -07002461 return ERR_PTR(-ENOSPC);
David Gibsonb45b5bd2006-03-22 00:08:55 -08002462}
2463
Aneesh Kumar K.Ve24a1302017-07-28 10:31:25 +05302464int alloc_bootmem_huge_page(struct hstate *h)
2465 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2466int __alloc_bootmem_huge_page(struct hstate *h)
Andi Kleenaa888a72008-07-23 21:27:47 -07002467{
2468 struct huge_bootmem_page *m;
Joonsoo Kimb2261022013-09-11 14:21:00 -07002469 int nr_nodes, node;
Andi Kleenaa888a72008-07-23 21:27:47 -07002470
Joonsoo Kimb2261022013-09-11 14:21:00 -07002471 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
Andi Kleenaa888a72008-07-23 21:27:47 -07002472 void *addr;
2473
Mike Rapoporteb31d552018-10-30 15:08:04 -07002474 addr = memblock_alloc_try_nid_raw(
Grygorii Strashko8b89a112014-01-21 15:50:36 -08002475 huge_page_size(h), huge_page_size(h),
Mike Rapoport97ad1082018-10-30 15:09:44 -07002476 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
Andi Kleenaa888a72008-07-23 21:27:47 -07002477 if (addr) {
2478 /*
2479 * Use the beginning of the huge page to store the
2480 * huge_bootmem_page struct (until gather_bootmem
2481 * puts them into the mem_map).
2482 */
2483 m = addr;
Cyrill Gorcunov91f47662009-01-06 14:40:33 -08002484 goto found;
Andi Kleenaa888a72008-07-23 21:27:47 -07002485 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002486 }
2487 return 0;
2488
2489found:
Luiz Capitulinodf994ea2014-12-12 16:55:21 -08002490 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
Andi Kleenaa888a72008-07-23 21:27:47 -07002491 /* Put them into a private list first because mem_map is not up yet */
Cannon Matthews330d6e42018-08-17 15:49:17 -07002492 INIT_LIST_HEAD(&m->list);
Andi Kleenaa888a72008-07-23 21:27:47 -07002493 list_add(&m->list, &huge_boot_pages);
2494 m->hstate = h;
2495 return 1;
2496}
2497
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08002498static void __init prep_compound_huge_page(struct page *page,
2499 unsigned int order)
Andy Whitcroft18229df2008-11-06 12:53:27 -08002500{
2501 if (unlikely(order > (MAX_ORDER - 1)))
2502 prep_compound_gigantic_page(page, order);
2503 else
2504 prep_compound_page(page, order);
2505}
2506
Andi Kleenaa888a72008-07-23 21:27:47 -07002507/* Put bootmem huge pages into the standard lists after mem_map is up */
2508static void __init gather_bootmem_prealloc(void)
2509{
2510 struct huge_bootmem_page *m;
2511
2512 list_for_each_entry(m, &huge_boot_pages, list) {
Mike Kravetz40d18eb2018-08-17 15:49:07 -07002513 struct page *page = virt_to_page(m);
Andi Kleenaa888a72008-07-23 21:27:47 -07002514 struct hstate *h = m->hstate;
Becky Bruceee8f2482011-07-25 17:11:50 -07002515
Andi Kleenaa888a72008-07-23 21:27:47 -07002516 WARN_ON(page_count(page) != 1);
Andy Whitcroft18229df2008-11-06 12:53:27 -08002517 prep_compound_huge_page(page, h->order);
Andrea Arcangelief5a22b2013-10-16 13:46:56 -07002518 WARN_ON(PageReserved(page));
Andi Kleenaa888a72008-07-23 21:27:47 -07002519 prep_new_huge_page(h, page, page_to_nid(page));
Michal Hockoaf0fb9d2018-01-31 16:20:41 -08002520 put_page(page); /* free it into the hugepage allocator */
2521
Rafael Aquinib0320c72011-06-15 15:08:39 -07002522 /*
2523 * If we had gigantic hugepages allocated at boot time, we need
2524 * to restore the 'stolen' pages to totalram_pages in order to
2525 * fix confusing memory reports from free(1) and another
2526 * side-effects, like CommitLimit going negative.
2527 */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002528 if (hstate_is_gigantic(h))
Jiang Liu3dcc0572013-07-03 15:03:21 -07002529 adjust_managed_page_count(page, 1 << h->order);
Cannon Matthews520495f2018-07-03 17:02:43 -07002530 cond_resched();
Andi Kleenaa888a72008-07-23 21:27:47 -07002531 }
2532}
2533
Andi Kleen8faa8b02008-07-23 21:27:48 -07002534static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535{
2536 unsigned long i;
Mike Kravetzf60858f2019-09-23 15:37:35 -07002537 nodemask_t *node_alloc_noretry;
2538
2539 if (!hstate_is_gigantic(h)) {
2540 /*
2541 * Bit mask controlling how hard we retry per-node allocations.
2542 * Ignore errors as lower level routines can deal with
2543 * node_alloc_noretry == NULL. If this kmalloc fails at boot
2544 * time, we are likely in bigger trouble.
2545 */
2546 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2547 GFP_KERNEL);
2548 } else {
2549 /* allocations done at boot time */
2550 node_alloc_noretry = NULL;
2551 }
2552
2553 /* bit mask controlling how hard we retry per-node allocations */
2554 if (node_alloc_noretry)
2555 nodes_clear(*node_alloc_noretry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556
Andi Kleene5ff2152008-07-23 21:27:42 -07002557 for (i = 0; i < h->max_huge_pages; ++i) {
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002558 if (hstate_is_gigantic(h)) {
Barry Songdbda8fe2020-07-23 21:15:30 -07002559 if (hugetlb_cma_size) {
Roman Gushchincf11e852020-04-10 14:32:45 -07002560 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
2561 break;
2562 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002563 if (!alloc_bootmem_huge_page(h))
2564 break;
Michal Hocko0c397da2018-01-31 16:20:56 -08002565 } else if (!alloc_pool_huge_page(h,
Mike Kravetzf60858f2019-09-23 15:37:35 -07002566 &node_states[N_MEMORY],
2567 node_alloc_noretry))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568 break;
David Rientjes69ed7792017-07-10 15:48:50 -07002569 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570 }
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002571 if (i < h->max_huge_pages) {
2572 char buf[32];
2573
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002574 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Liam R. Howlettd715cf82017-07-10 15:48:15 -07002575 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2576 h->max_huge_pages, buf, i);
2577 h->max_huge_pages = i;
2578 }
Mike Kravetzf60858f2019-09-23 15:37:35 -07002579
2580 kfree(node_alloc_noretry);
Andi Kleene5ff2152008-07-23 21:27:42 -07002581}
2582
2583static void __init hugetlb_init_hstates(void)
2584{
2585 struct hstate *h;
2586
2587 for_each_hstate(h) {
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002588 if (minimum_order > huge_page_order(h))
2589 minimum_order = huge_page_order(h);
2590
Andi Kleen8faa8b02008-07-23 21:27:48 -07002591 /* oversize hugepages were init'ed in early boot */
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002592 if (!hstate_is_gigantic(h))
Andi Kleen8faa8b02008-07-23 21:27:48 -07002593 hugetlb_hstate_alloc_pages(h);
Andi Kleene5ff2152008-07-23 21:27:42 -07002594 }
Naoya Horiguchi641844f2015-06-24 16:56:59 -07002595 VM_BUG_ON(minimum_order == UINT_MAX);
Andi Kleene5ff2152008-07-23 21:27:42 -07002596}
2597
2598static void __init report_hugepages(void)
2599{
2600 struct hstate *h;
2601
2602 for_each_hstate(h) {
Andi Kleen4abd32d2008-07-23 21:27:49 -07002603 char buf[32];
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002604
2605 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
Andrew Mortonffb22af2013-02-22 16:32:08 -08002606 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
Matthew Wilcoxc6247f72017-07-10 15:48:56 -07002607 buf, h->free_huge_pages);
Andi Kleene5ff2152008-07-23 21:27:42 -07002608 }
2609}
2610
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611#ifdef CONFIG_HIGHMEM
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002612static void try_to_free_low(struct hstate *h, unsigned long count,
2613 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614{
Christoph Lameter4415cc82006-09-25 23:31:55 -07002615 int i;
2616
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002617 if (hstate_is_gigantic(h))
Andi Kleenaa888a72008-07-23 21:27:47 -07002618 return;
2619
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002620 for_each_node_mask(i, *nodes_allowed) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 struct page *page, *next;
Andi Kleena5516432008-07-23 21:27:41 -07002622 struct list_head *freel = &h->hugepage_freelists[i];
2623 list_for_each_entry_safe(page, next, freel, lru) {
2624 if (count >= h->nr_huge_pages)
Adam Litke6b0c8802007-10-16 01:26:23 -07002625 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 if (PageHighMem(page))
2627 continue;
2628 list_del(&page->lru);
Andi Kleene5ff2152008-07-23 21:27:42 -07002629 update_and_free_page(h, page);
Andi Kleena5516432008-07-23 21:27:41 -07002630 h->free_huge_pages--;
2631 h->free_huge_pages_node[page_to_nid(page)]--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 }
2633 }
2634}
2635#else
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002636static inline void try_to_free_low(struct hstate *h, unsigned long count,
2637 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638{
2639}
2640#endif
2641
Wu Fengguang20a03072009-06-16 15:32:22 -07002642/*
2643 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2644 * balanced by operating on them in a round-robin fashion.
2645 * Returns 1 if an adjustment was made.
2646 */
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002647static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2648 int delta)
Wu Fengguang20a03072009-06-16 15:32:22 -07002649{
Joonsoo Kimb2261022013-09-11 14:21:00 -07002650 int nr_nodes, node;
Wu Fengguang20a03072009-06-16 15:32:22 -07002651
2652 VM_BUG_ON(delta != -1 && delta != 1);
Wu Fengguang20a03072009-06-16 15:32:22 -07002653
Joonsoo Kimb2261022013-09-11 14:21:00 -07002654 if (delta < 0) {
2655 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2656 if (h->surplus_huge_pages_node[node])
2657 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002658 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002659 } else {
2660 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2661 if (h->surplus_huge_pages_node[node] <
2662 h->nr_huge_pages_node[node])
2663 goto found;
Lee Schermerhorne8c5c822009-09-21 17:01:22 -07002664 }
Joonsoo Kimb2261022013-09-11 14:21:00 -07002665 }
2666 return 0;
Wu Fengguang20a03072009-06-16 15:32:22 -07002667
Joonsoo Kimb2261022013-09-11 14:21:00 -07002668found:
2669 h->surplus_huge_pages += delta;
2670 h->surplus_huge_pages_node[node] += delta;
2671 return 1;
Wu Fengguang20a03072009-06-16 15:32:22 -07002672}
2673
Andi Kleena5516432008-07-23 21:27:41 -07002674#define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002675static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002676 nodemask_t *nodes_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677{
Adam Litke7893d1d2007-10-16 01:26:18 -07002678 unsigned long min_count, ret;
Mike Kravetzf60858f2019-09-23 15:37:35 -07002679 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2680
2681 /*
2682 * Bit mask controlling how hard we retry per-node allocations.
2683 * If we can not allocate the bit mask, do not attempt to allocate
2684 * the requested huge pages.
2685 */
2686 if (node_alloc_noretry)
2687 nodes_clear(*node_alloc_noretry);
2688 else
2689 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002691 spin_lock(&hugetlb_lock);
2692
2693 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002694 * Check for a node specific request.
2695 * Changing node specific huge page count may require a corresponding
2696 * change to the global count. In any case, the passed node mask
2697 * (nodes_allowed) will restrict alloc/free to the specified node.
2698 */
2699 if (nid != NUMA_NO_NODE) {
2700 unsigned long old_count = count;
2701
2702 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2703 /*
2704 * User may have specified a large count value which caused the
2705 * above calculation to overflow. In this case, they wanted
2706 * to allocate as many huge pages as possible. Set count to
2707 * largest possible value to align with their intention.
2708 */
2709 if (count < old_count)
2710 count = ULONG_MAX;
2711 }
2712
2713 /*
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002714 * Gigantic pages runtime allocation depend on the capability for large
2715 * page range allocation.
2716 * If the system does not provide this feature, return an error when
2717 * the user tries to allocate gigantic pages but let the user free the
2718 * boottime allocated gigantic pages.
2719 */
2720 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2721 if (count > persistent_huge_pages(h)) {
2722 spin_unlock(&hugetlb_lock);
Mike Kravetzf60858f2019-09-23 15:37:35 -07002723 NODEMASK_FREE(node_alloc_noretry);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002724 return -EINVAL;
2725 }
2726 /* Fall through to decrease pool */
2727 }
Andi Kleenaa888a72008-07-23 21:27:47 -07002728
Adam Litke7893d1d2007-10-16 01:26:18 -07002729 /*
2730 * Increase the pool size
2731 * First take pages out of surplus state. Then make up the
2732 * remaining difference by allocating fresh huge pages.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002733 *
Michal Hocko0c397da2018-01-31 16:20:56 -08002734 * We might race with alloc_surplus_huge_page() here and be unable
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002735 * to convert a surplus huge page to a normal huge page. That is
2736 * not critical, though, it just means the overall size of the
2737 * pool might be one hugepage larger than it needs to be, but
2738 * within all the constraints specified by the sysctls.
Adam Litke7893d1d2007-10-16 01:26:18 -07002739 */
Andi Kleena5516432008-07-23 21:27:41 -07002740 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002741 if (!adjust_pool_surplus(h, nodes_allowed, -1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002742 break;
2743 }
2744
Andi Kleena5516432008-07-23 21:27:41 -07002745 while (count > persistent_huge_pages(h)) {
Adam Litke7893d1d2007-10-16 01:26:18 -07002746 /*
2747 * If this allocation races such that we no longer need the
2748 * page, free_huge_page will handle it by freeing the page
2749 * and reducing the surplus.
2750 */
2751 spin_unlock(&hugetlb_lock);
Jia He649920c2016-08-02 14:02:31 -07002752
2753 /* yield cpu to avoid soft lockup */
2754 cond_resched();
2755
Mike Kravetzf60858f2019-09-23 15:37:35 -07002756 ret = alloc_pool_huge_page(h, nodes_allowed,
2757 node_alloc_noretry);
Adam Litke7893d1d2007-10-16 01:26:18 -07002758 spin_lock(&hugetlb_lock);
2759 if (!ret)
2760 goto out;
2761
Mel Gorman536240f22009-12-14 17:59:56 -08002762 /* Bail for signals. Probably ctrl-c from user */
2763 if (signal_pending(current))
2764 goto out;
Adam Litke7893d1d2007-10-16 01:26:18 -07002765 }
Adam Litke7893d1d2007-10-16 01:26:18 -07002766
2767 /*
2768 * Decrease the pool size
2769 * First return free pages to the buddy allocator (being careful
2770 * to keep enough around to satisfy reservations). Then place
2771 * pages into surplus state as needed so the pool will shrink
2772 * to the desired size as pages become free.
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002773 *
2774 * By placing pages into the surplus state independent of the
2775 * overcommit value, we are allowing the surplus pool size to
2776 * exceed overcommit. There are few sane options here. Since
Michal Hocko0c397da2018-01-31 16:20:56 -08002777 * alloc_surplus_huge_page() is checking the global counter,
Nishanth Aravamudand1c3fb12007-12-17 16:20:12 -08002778 * though, we'll note that we're not allowed to exceed surplus
2779 * and won't grow the pool anywhere else. Not until one of the
2780 * sysctls are changed, or the surplus pages go out of use.
Adam Litke7893d1d2007-10-16 01:26:18 -07002781 */
Andi Kleena5516432008-07-23 21:27:41 -07002782 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
Adam Litke6b0c8802007-10-16 01:26:23 -07002783 min_count = max(count, min_count);
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002784 try_to_free_low(h, min_count, nodes_allowed);
Andi Kleena5516432008-07-23 21:27:41 -07002785 while (min_count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002786 if (!free_pool_huge_page(h, nodes_allowed, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 break;
Mizuma, Masayoshi55f67142014-04-07 15:37:54 -07002788 cond_resched_lock(&hugetlb_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 }
Andi Kleena5516432008-07-23 21:27:41 -07002790 while (count < persistent_huge_pages(h)) {
Lee Schermerhorn6ae11b22009-12-14 17:58:16 -08002791 if (!adjust_pool_surplus(h, nodes_allowed, 1))
Adam Litke7893d1d2007-10-16 01:26:18 -07002792 break;
2793 }
2794out:
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002795 h->max_huge_pages = persistent_huge_pages(h);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 spin_unlock(&hugetlb_lock);
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002797
Mike Kravetzf60858f2019-09-23 15:37:35 -07002798 NODEMASK_FREE(node_alloc_noretry);
2799
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002800 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801}
2802
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002803#define HSTATE_ATTR_RO(_name) \
2804 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2805
2806#define HSTATE_ATTR(_name) \
2807 static struct kobj_attribute _name##_attr = \
2808 __ATTR(_name, 0644, _name##_show, _name##_store)
2809
2810static struct kobject *hugepages_kobj;
2811static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2812
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002813static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2814
2815static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002816{
2817 int i;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002818
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002819 for (i = 0; i < HUGE_MAX_HSTATE; i++)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002820 if (hstate_kobjs[i] == kobj) {
2821 if (nidp)
2822 *nidp = NUMA_NO_NODE;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002823 return &hstates[i];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002824 }
2825
2826 return kobj_to_node_hstate(kobj, nidp);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002827}
2828
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002829static ssize_t nr_hugepages_show_common(struct kobject *kobj,
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002830 struct kobj_attribute *attr, char *buf)
2831{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002832 struct hstate *h;
2833 unsigned long nr_huge_pages;
2834 int nid;
2835
2836 h = kobj_to_hstate(kobj, &nid);
2837 if (nid == NUMA_NO_NODE)
2838 nr_huge_pages = h->nr_huge_pages;
2839 else
2840 nr_huge_pages = h->nr_huge_pages_node[nid];
2841
2842 return sprintf(buf, "%lu\n", nr_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002843}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002844
David Rientjes238d3c12014-08-06 16:06:51 -07002845static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2846 struct hstate *h, int nid,
2847 unsigned long count, size_t len)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002848{
2849 int err;
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002850 nodemask_t nodes_allowed, *n_mask;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002851
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002852 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2853 return -EINVAL;
Eric B Munsonadbe8722011-01-13 15:47:27 -08002854
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002855 if (nid == NUMA_NO_NODE) {
2856 /*
2857 * global hstate attribute
2858 */
2859 if (!(obey_mempolicy &&
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002860 init_nodemask_of_mempolicy(&nodes_allowed)))
2861 n_mask = &node_states[N_MEMORY];
2862 else
2863 n_mask = &nodes_allowed;
2864 } else {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002865 /*
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002866 * Node specific request. count adjustment happens in
2867 * set_max_huge_pages() after acquiring hugetlb_lock.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002868 */
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002869 init_nodemask_of_node(&nodes_allowed, nid);
2870 n_mask = &nodes_allowed;
Mike Kravetzfd875dc2019-05-13 17:19:20 -07002871 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002872
Oscar Salvador2d0adf72019-05-13 17:19:23 -07002873 err = set_max_huge_pages(h, count, nid, n_mask);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002874
Alexandre Ghiti4eb07162019-05-13 17:19:04 -07002875 return err ? err : len;
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002876}
2877
David Rientjes238d3c12014-08-06 16:06:51 -07002878static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2879 struct kobject *kobj, const char *buf,
2880 size_t len)
2881{
2882 struct hstate *h;
2883 unsigned long count;
2884 int nid;
2885 int err;
2886
2887 err = kstrtoul(buf, 10, &count);
2888 if (err)
2889 return err;
2890
2891 h = kobj_to_hstate(kobj, &nid);
2892 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2893}
2894
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002895static ssize_t nr_hugepages_show(struct kobject *kobj,
2896 struct kobj_attribute *attr, char *buf)
2897{
2898 return nr_hugepages_show_common(kobj, attr, buf);
2899}
2900
2901static ssize_t nr_hugepages_store(struct kobject *kobj,
2902 struct kobj_attribute *attr, const char *buf, size_t len)
2903{
David Rientjes238d3c12014-08-06 16:06:51 -07002904 return nr_hugepages_store_common(false, kobj, buf, len);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002905}
2906HSTATE_ATTR(nr_hugepages);
2907
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002908#ifdef CONFIG_NUMA
2909
2910/*
2911 * hstate attribute for optionally mempolicy-based constraint on persistent
2912 * huge page alloc/free.
2913 */
2914static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2915 struct kobj_attribute *attr, char *buf)
2916{
2917 return nr_hugepages_show_common(kobj, attr, buf);
2918}
2919
2920static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2921 struct kobj_attribute *attr, const char *buf, size_t len)
2922{
David Rientjes238d3c12014-08-06 16:06:51 -07002923 return nr_hugepages_store_common(true, kobj, buf, len);
Lee Schermerhorn06808b02009-12-14 17:58:21 -08002924}
2925HSTATE_ATTR(nr_hugepages_mempolicy);
2926#endif
2927
2928
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002929static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2930 struct kobj_attribute *attr, char *buf)
2931{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002932 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002933 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2934}
Eric B Munsonadbe8722011-01-13 15:47:27 -08002935
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002936static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2937 struct kobj_attribute *attr, const char *buf, size_t count)
2938{
2939 int err;
2940 unsigned long input;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002941 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002942
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07002943 if (hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08002944 return -EINVAL;
2945
Jingoo Han3dbb95f2013-09-11 14:20:25 -07002946 err = kstrtoul(buf, 10, &input);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002947 if (err)
Eric B Munson73ae31e2011-01-13 15:47:28 -08002948 return err;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002949
2950 spin_lock(&hugetlb_lock);
2951 h->nr_overcommit_huge_pages = input;
2952 spin_unlock(&hugetlb_lock);
2953
2954 return count;
2955}
2956HSTATE_ATTR(nr_overcommit_hugepages);
2957
2958static ssize_t free_hugepages_show(struct kobject *kobj,
2959 struct kobj_attribute *attr, char *buf)
2960{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002961 struct hstate *h;
2962 unsigned long free_huge_pages;
2963 int nid;
2964
2965 h = kobj_to_hstate(kobj, &nid);
2966 if (nid == NUMA_NO_NODE)
2967 free_huge_pages = h->free_huge_pages;
2968 else
2969 free_huge_pages = h->free_huge_pages_node[nid];
2970
2971 return sprintf(buf, "%lu\n", free_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002972}
2973HSTATE_ATTR_RO(free_hugepages);
2974
2975static ssize_t resv_hugepages_show(struct kobject *kobj,
2976 struct kobj_attribute *attr, char *buf)
2977{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002978 struct hstate *h = kobj_to_hstate(kobj, NULL);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002979 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2980}
2981HSTATE_ATTR_RO(resv_hugepages);
2982
2983static ssize_t surplus_hugepages_show(struct kobject *kobj,
2984 struct kobj_attribute *attr, char *buf)
2985{
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08002986 struct hstate *h;
2987 unsigned long surplus_huge_pages;
2988 int nid;
2989
2990 h = kobj_to_hstate(kobj, &nid);
2991 if (nid == NUMA_NO_NODE)
2992 surplus_huge_pages = h->surplus_huge_pages;
2993 else
2994 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2995
2996 return sprintf(buf, "%lu\n", surplus_huge_pages);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07002997}
2998HSTATE_ATTR_RO(surplus_hugepages);
2999
3000static struct attribute *hstate_attrs[] = {
3001 &nr_hugepages_attr.attr,
3002 &nr_overcommit_hugepages_attr.attr,
3003 &free_hugepages_attr.attr,
3004 &resv_hugepages_attr.attr,
3005 &surplus_hugepages_attr.attr,
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003006#ifdef CONFIG_NUMA
3007 &nr_hugepages_mempolicy_attr.attr,
3008#endif
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003009 NULL,
3010};
3011
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003012static const struct attribute_group hstate_attr_group = {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003013 .attrs = hstate_attrs,
3014};
3015
Jeff Mahoney094e9532010-02-02 13:44:14 -08003016static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
3017 struct kobject **hstate_kobjs,
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003018 const struct attribute_group *hstate_attr_group)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003019{
3020 int retval;
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003021 int hi = hstate_index(h);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003022
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003023 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
3024 if (!hstate_kobjs[hi])
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003025 return -ENOMEM;
3026
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003027 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003028 if (retval)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003029 kobject_put(hstate_kobjs[hi]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003030
3031 return retval;
3032}
3033
3034static void __init hugetlb_sysfs_init(void)
3035{
3036 struct hstate *h;
3037 int err;
3038
3039 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
3040 if (!hugepages_kobj)
3041 return;
3042
3043 for_each_hstate(h) {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003044 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
3045 hstate_kobjs, &hstate_attr_group);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003046 if (err)
Mike Kravetz282f4212020-06-03 16:00:46 -07003047 pr_err("HugeTLB: Unable to add hstate %s", h->name);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003048 }
3049}
3050
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003051#ifdef CONFIG_NUMA
3052
3053/*
3054 * node_hstate/s - associate per node hstate attributes, via their kobjects,
Kay Sievers10fbcf42011-12-21 14:48:43 -08003055 * with node devices in node_devices[] using a parallel array. The array
3056 * index of a node device or _hstate == node id.
3057 * This is here to avoid any static dependency of the node device driver, in
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003058 * the base kernel, on the hugetlb module.
3059 */
3060struct node_hstate {
3061 struct kobject *hugepages_kobj;
3062 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3063};
Alexander Kuleshovb4e289a2015-11-05 18:50:14 -08003064static struct node_hstate node_hstates[MAX_NUMNODES];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003065
3066/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003067 * A subset of global hstate attributes for node devices
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003068 */
3069static struct attribute *per_node_hstate_attrs[] = {
3070 &nr_hugepages_attr.attr,
3071 &free_hugepages_attr.attr,
3072 &surplus_hugepages_attr.attr,
3073 NULL,
3074};
3075
Arvind Yadav67e5ed92017-09-06 16:22:06 -07003076static const struct attribute_group per_node_hstate_attr_group = {
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003077 .attrs = per_node_hstate_attrs,
3078};
3079
3080/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003081 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003082 * Returns node id via non-NULL nidp.
3083 */
3084static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3085{
3086 int nid;
3087
3088 for (nid = 0; nid < nr_node_ids; nid++) {
3089 struct node_hstate *nhs = &node_hstates[nid];
3090 int i;
3091 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3092 if (nhs->hstate_kobjs[i] == kobj) {
3093 if (nidp)
3094 *nidp = nid;
3095 return &hstates[i];
3096 }
3097 }
3098
3099 BUG();
3100 return NULL;
3101}
3102
3103/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003104 * Unregister hstate attributes from a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003105 * No-op if no hstate attributes attached.
3106 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003107static void hugetlb_unregister_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003108{
3109 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003110 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003111
3112 if (!nhs->hugepages_kobj)
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003113 return; /* no hstate attributes */
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003114
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003115 for_each_hstate(h) {
3116 int idx = hstate_index(h);
3117 if (nhs->hstate_kobjs[idx]) {
3118 kobject_put(nhs->hstate_kobjs[idx]);
3119 nhs->hstate_kobjs[idx] = NULL;
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003120 }
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07003121 }
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003122
3123 kobject_put(nhs->hugepages_kobj);
3124 nhs->hugepages_kobj = NULL;
3125}
3126
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003127
3128/*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003129 * Register hstate attributes for a single node device.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003130 * No-op if attributes already registered.
3131 */
Claudiu Ghioc3cd8b442013-03-04 12:46:15 +02003132static void hugetlb_register_node(struct node *node)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003133{
3134 struct hstate *h;
Kay Sievers10fbcf42011-12-21 14:48:43 -08003135 struct node_hstate *nhs = &node_hstates[node->dev.id];
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003136 int err;
3137
3138 if (nhs->hugepages_kobj)
3139 return; /* already allocated */
3140
3141 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
Kay Sievers10fbcf42011-12-21 14:48:43 -08003142 &node->dev.kobj);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003143 if (!nhs->hugepages_kobj)
3144 return;
3145
3146 for_each_hstate(h) {
3147 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
3148 nhs->hstate_kobjs,
3149 &per_node_hstate_attr_group);
3150 if (err) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003151 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08003152 h->name, node->dev.id);
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003153 hugetlb_unregister_node(node);
3154 break;
3155 }
3156 }
3157}
3158
3159/*
Lee Schermerhorn9b5e5d02009-12-14 17:58:32 -08003160 * hugetlb init time: register hstate attributes for all registered node
Kay Sievers10fbcf42011-12-21 14:48:43 -08003161 * devices of nodes that have memory. All on-line nodes should have
3162 * registered their associated device by this time.
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003163 */
Luiz Capitulino7d9ca002014-12-12 16:55:24 -08003164static void __init hugetlb_register_all_nodes(void)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003165{
3166 int nid;
3167
Lai Jiangshan8cebfcd2012-12-12 13:51:36 -08003168 for_each_node_state(nid, N_MEMORY) {
Wen Congyang87327942012-12-11 16:00:56 -08003169 struct node *node = node_devices[nid];
Kay Sievers10fbcf42011-12-21 14:48:43 -08003170 if (node->dev.id == nid)
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003171 hugetlb_register_node(node);
3172 }
3173
3174 /*
Kay Sievers10fbcf42011-12-21 14:48:43 -08003175 * Let the node device driver know we're here so it can
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003176 * [un]register hstate attributes on node hotplug.
3177 */
3178 register_hugetlbfs_with_node(hugetlb_register_node,
3179 hugetlb_unregister_node);
3180}
3181#else /* !CONFIG_NUMA */
3182
3183static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
3184{
3185 BUG();
3186 if (nidp)
3187 *nidp = -1;
3188 return NULL;
3189}
3190
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003191static void hugetlb_register_all_nodes(void) { }
3192
3193#endif
3194
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003195static int __init hugetlb_init(void)
3196{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003197 int i;
3198
Mike Kravetzc2833a52020-06-03 16:00:50 -07003199 if (!hugepages_supported()) {
3200 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
3201 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
Benjamin Herrenschmidt0ef89d22008-07-31 00:07:30 -07003202 return 0;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003203 }
Vaishali Thakkarf8b74812016-02-17 13:11:26 -08003204
Mike Kravetz282f4212020-06-03 16:00:46 -07003205 /*
3206 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
3207 * architectures depend on setup being done here.
3208 */
3209 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
3210 if (!parsed_default_hugepagesz) {
3211 /*
3212 * If we did not parse a default huge page size, set
3213 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
3214 * number of huge pages for this default size was implicitly
3215 * specified, set that here as well.
3216 * Note that the implicit setting will overwrite an explicit
3217 * setting. A warning will be printed in this case.
3218 */
3219 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
3220 if (default_hstate_max_huge_pages) {
3221 if (default_hstate.max_huge_pages) {
3222 char buf[32];
Andi Kleenaa888a72008-07-23 21:27:47 -07003223
Mike Kravetz282f4212020-06-03 16:00:46 -07003224 string_get_size(huge_page_size(&default_hstate),
3225 1, STRING_UNITS_2, buf, 32);
3226 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
3227 default_hstate.max_huge_pages, buf);
3228 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
3229 default_hstate_max_huge_pages);
3230 }
3231 default_hstate.max_huge_pages =
3232 default_hstate_max_huge_pages;
3233 }
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003234 }
Andi Kleenaa888a72008-07-23 21:27:47 -07003235
Roman Gushchincf11e852020-04-10 14:32:45 -07003236 hugetlb_cma_check();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003237 hugetlb_init_hstates();
Andi Kleenaa888a72008-07-23 21:27:47 -07003238 gather_bootmem_prealloc();
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003239 report_hugepages();
3240
3241 hugetlb_sysfs_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003242 hugetlb_register_all_nodes();
Jianguo Wu7179e7b2012-12-18 14:23:19 -08003243 hugetlb_cgroup_file_init();
Lee Schermerhorn9a3052302009-12-14 17:58:25 -08003244
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003245#ifdef CONFIG_SMP
3246 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
3247#else
3248 num_fault_mutexes = 1;
3249#endif
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003250 hugetlb_fault_mutex_table =
Kees Cook6da2ec52018-06-12 13:55:00 -07003251 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
3252 GFP_KERNEL);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003253 BUG_ON(!hugetlb_fault_mutex_table);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07003254
3255 for (i = 0; i < num_fault_mutexes; i++)
Mike Kravetzc672c7f2015-09-08 15:01:35 -07003256 mutex_init(&hugetlb_fault_mutex_table[i]);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003257 return 0;
3258}
Paul Gortmaker3e89e1c2016-01-14 15:21:52 -08003259subsys_initcall(hugetlb_init);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003260
Mike Kravetzae94da82020-06-03 16:00:34 -07003261/* Overwritten by architectures with more huge page sizes */
3262bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003263{
Mike Kravetzae94da82020-06-03 16:00:34 -07003264 return size == HPAGE_SIZE;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003265}
3266
Kirill A. Shutemovd00181b2015-11-06 16:29:57 -08003267void __init hugetlb_add_hstate(unsigned int order)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003268{
3269 struct hstate *h;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003270 unsigned long i;
3271
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003272 if (size_to_hstate(PAGE_SIZE << order)) {
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003273 return;
3274 }
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003275 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003276 BUG_ON(order == 0);
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003277 h = &hstates[hugetlb_max_hstate++];
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003278 h->order = order;
3279 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003280 h->nr_huge_pages = 0;
3281 h->free_huge_pages = 0;
3282 for (i = 0; i < MAX_NUMNODES; ++i)
3283 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
Aneesh Kumar K.V0edaecf2012-07-31 16:42:07 -07003284 INIT_LIST_HEAD(&h->hugepage_activelist);
Andrew Morton54f18d32016-05-19 17:11:40 -07003285 h->next_nid_to_alloc = first_memory_node;
3286 h->next_nid_to_free = first_memory_node;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003287 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
3288 huge_page_size(h)/1024);
Andi Kleen8faa8b02008-07-23 21:27:48 -07003289
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003290 parsed_hstate = h;
3291}
3292
Mike Kravetz282f4212020-06-03 16:00:46 -07003293/*
3294 * hugepages command line processing
3295 * hugepages normally follows a valid hugepagsz or default_hugepagsz
3296 * specification. If not, ignore the hugepages value. hugepages can also
3297 * be the first huge page command line option in which case it implicitly
3298 * specifies the number of huge pages for the default size.
3299 */
3300static int __init hugepages_setup(char *s)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003301{
3302 unsigned long *mhp;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003303 static unsigned long *last_mhp;
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003304
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003305 if (!parsed_valid_hugepagesz) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003306 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003307 parsed_valid_hugepagesz = true;
Mike Kravetz282f4212020-06-03 16:00:46 -07003308 return 0;
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003309 }
Mike Kravetz282f4212020-06-03 16:00:46 -07003310
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003311 /*
Mike Kravetz282f4212020-06-03 16:00:46 -07003312 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
3313 * yet, so this hugepages= parameter goes to the "default hstate".
3314 * Otherwise, it goes with the previously parsed hugepagesz or
3315 * default_hugepagesz.
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003316 */
Vaishali Thakkar9fee0212016-05-19 17:11:04 -07003317 else if (!hugetlb_max_hstate)
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003318 mhp = &default_hstate_max_huge_pages;
3319 else
3320 mhp = &parsed_hstate->max_huge_pages;
3321
Andi Kleen8faa8b02008-07-23 21:27:48 -07003322 if (mhp == last_mhp) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003323 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
3324 return 0;
Andi Kleen8faa8b02008-07-23 21:27:48 -07003325 }
3326
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003327 if (sscanf(s, "%lu", mhp) <= 0)
3328 *mhp = 0;
3329
Andi Kleen8faa8b02008-07-23 21:27:48 -07003330 /*
3331 * Global state is always initialized later in hugetlb_init.
3332 * But we need to allocate >= MAX_ORDER hstates here early to still
3333 * use the bootmem allocator.
3334 */
Aneesh Kumar K.V47d38342012-07-31 16:41:54 -07003335 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
Andi Kleen8faa8b02008-07-23 21:27:48 -07003336 hugetlb_hstate_alloc_pages(parsed_hstate);
3337
3338 last_mhp = mhp;
3339
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003340 return 1;
3341}
Mike Kravetz282f4212020-06-03 16:00:46 -07003342__setup("hugepages=", hugepages_setup);
Nick Piggine11bfbf2008-07-23 21:27:52 -07003343
Mike Kravetz282f4212020-06-03 16:00:46 -07003344/*
3345 * hugepagesz command line processing
3346 * A specific huge page size can only be specified once with hugepagesz.
3347 * hugepagesz is followed by hugepages on the command line. The global
3348 * variable 'parsed_valid_hugepagesz' is used to determine if prior
3349 * hugepagesz argument was valid.
3350 */
Mike Kravetz359f2542020-06-03 16:00:38 -07003351static int __init hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003352{
Mike Kravetz359f2542020-06-03 16:00:38 -07003353 unsigned long size;
Mike Kravetz282f4212020-06-03 16:00:46 -07003354 struct hstate *h;
3355
3356 parsed_valid_hugepagesz = false;
Mike Kravetz359f2542020-06-03 16:00:38 -07003357 size = (unsigned long)memparse(s, NULL);
3358
3359 if (!arch_hugetlb_valid_size(size)) {
Mike Kravetz282f4212020-06-03 16:00:46 -07003360 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
Mike Kravetz359f2542020-06-03 16:00:38 -07003361 return 0;
3362 }
3363
Mike Kravetz282f4212020-06-03 16:00:46 -07003364 h = size_to_hstate(size);
3365 if (h) {
3366 /*
3367 * hstate for this size already exists. This is normally
3368 * an error, but is allowed if the existing hstate is the
3369 * default hstate. More specifically, it is only allowed if
3370 * the number of huge pages for the default hstate was not
3371 * previously specified.
3372 */
3373 if (!parsed_default_hugepagesz || h != &default_hstate ||
3374 default_hstate.max_huge_pages) {
3375 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
3376 return 0;
3377 }
3378
3379 /*
3380 * No need to call hugetlb_add_hstate() as hstate already
3381 * exists. But, do set parsed_hstate so that a following
3382 * hugepages= parameter will be applied to this hstate.
3383 */
3384 parsed_hstate = h;
3385 parsed_valid_hugepagesz = true;
3386 return 1;
Mike Kravetz38237832020-06-03 16:00:42 -07003387 }
3388
Mike Kravetz359f2542020-06-03 16:00:38 -07003389 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
Mike Kravetz282f4212020-06-03 16:00:46 -07003390 parsed_valid_hugepagesz = true;
Nick Piggine11bfbf2008-07-23 21:27:52 -07003391 return 1;
3392}
Mike Kravetz359f2542020-06-03 16:00:38 -07003393__setup("hugepagesz=", hugepagesz_setup);
3394
Mike Kravetz282f4212020-06-03 16:00:46 -07003395/*
3396 * default_hugepagesz command line input
3397 * Only one instance of default_hugepagesz allowed on command line.
3398 */
Mike Kravetzae94da82020-06-03 16:00:34 -07003399static int __init default_hugepagesz_setup(char *s)
Nick Piggine11bfbf2008-07-23 21:27:52 -07003400{
Mike Kravetzae94da82020-06-03 16:00:34 -07003401 unsigned long size;
3402
Mike Kravetz282f4212020-06-03 16:00:46 -07003403 parsed_valid_hugepagesz = false;
Mike Kravetz282f4212020-06-03 16:00:46 -07003404 if (parsed_default_hugepagesz) {
3405 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
3406 return 0;
3407 }
3408
3409 size = (unsigned long)memparse(s, NULL);
3410
3411 if (!arch_hugetlb_valid_size(size)) {
3412 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
3413 return 0;
3414 }
3415
3416 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
3417 parsed_valid_hugepagesz = true;
3418 parsed_default_hugepagesz = true;
3419 default_hstate_idx = hstate_index(size_to_hstate(size));
3420
3421 /*
3422 * The number of default huge pages (for this size) could have been
3423 * specified as the first hugetlb parameter: hugepages=X. If so,
3424 * then default_hstate_max_huge_pages is set. If the default huge
3425 * page size is gigantic (>= MAX_ORDER), then the pages must be
3426 * allocated here from bootmem allocator.
3427 */
3428 if (default_hstate_max_huge_pages) {
3429 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
3430 if (hstate_is_gigantic(&default_hstate))
3431 hugetlb_hstate_alloc_pages(&default_hstate);
3432 default_hstate_max_huge_pages = 0;
3433 }
3434
Nick Piggine11bfbf2008-07-23 21:27:52 -07003435 return 1;
3436}
Mike Kravetzae94da82020-06-03 16:00:34 -07003437__setup("default_hugepagesz=", default_hugepagesz_setup);
Nishanth Aravamudana3437872008-07-23 21:27:44 -07003438
Muchun Song8ca39e62020-08-11 18:30:32 -07003439static unsigned int allowed_mems_nr(struct hstate *h)
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003440{
3441 int node;
3442 unsigned int nr = 0;
Muchun Song8ca39e62020-08-11 18:30:32 -07003443 nodemask_t *mpol_allowed;
3444 unsigned int *array = h->free_huge_pages_node;
3445 gfp_t gfp_mask = htlb_alloc_mask(h);
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003446
Muchun Song8ca39e62020-08-11 18:30:32 -07003447 mpol_allowed = policy_nodemask_current(gfp_mask);
3448
3449 for_each_node_mask(node, cpuset_current_mems_allowed) {
3450 if (!mpol_allowed ||
3451 (mpol_allowed && node_isset(node, *mpol_allowed)))
3452 nr += array[node];
3453 }
Nishanth Aravamudan8a213462008-07-25 19:44:37 -07003454
3455 return nr;
3456}
3457
3458#ifdef CONFIG_SYSCTL
Muchun Song17743792020-09-04 16:36:13 -07003459static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
3460 void *buffer, size_t *length,
3461 loff_t *ppos, unsigned long *out)
3462{
3463 struct ctl_table dup_table;
3464
3465 /*
3466 * In order to avoid races with __do_proc_doulongvec_minmax(), we
3467 * can duplicate the @table and alter the duplicate of it.
3468 */
3469 dup_table = *table;
3470 dup_table.data = out;
3471
3472 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
3473}
3474
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003475static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
3476 struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003477 void *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478{
Andi Kleene5ff2152008-07-23 21:27:42 -07003479 struct hstate *h = &default_hstate;
David Rientjes238d3c12014-08-06 16:06:51 -07003480 unsigned long tmp = h->max_huge_pages;
Michal Hocko08d4a242011-01-13 15:47:26 -08003481 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003482
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003483 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003484 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003485
Muchun Song17743792020-09-04 16:36:13 -07003486 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3487 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003488 if (ret)
3489 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003490
David Rientjes238d3c12014-08-06 16:06:51 -07003491 if (write)
3492 ret = __nr_hugepages_store_common(obey_mempolicy, h,
3493 NUMA_NO_NODE, tmp, *length);
Michal Hocko08d4a242011-01-13 15:47:26 -08003494out:
3495 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496}
Mel Gorman396faf02007-07-17 04:03:13 -07003497
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003498int hugetlb_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003499 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003500{
3501
3502 return hugetlb_sysctl_handler_common(false, table, write,
3503 buffer, length, ppos);
3504}
3505
3506#ifdef CONFIG_NUMA
3507int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003508 void *buffer, size_t *length, loff_t *ppos)
Lee Schermerhorn06808b02009-12-14 17:58:21 -08003509{
3510 return hugetlb_sysctl_handler_common(true, table, write,
3511 buffer, length, ppos);
3512}
3513#endif /* CONFIG_NUMA */
3514
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003515int hugetlb_overcommit_handler(struct ctl_table *table, int write,
Christoph Hellwig32927392020-04-24 08:43:38 +02003516 void *buffer, size_t *length, loff_t *ppos)
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003517{
Andi Kleena5516432008-07-23 21:27:41 -07003518 struct hstate *h = &default_hstate;
Andi Kleene5ff2152008-07-23 21:27:42 -07003519 unsigned long tmp;
Michal Hocko08d4a242011-01-13 15:47:26 -08003520 int ret;
Andi Kleene5ff2152008-07-23 21:27:42 -07003521
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003522 if (!hugepages_supported())
Jan Stancek86613622016-03-09 14:08:35 -08003523 return -EOPNOTSUPP;
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003524
Petr Holasekc033a932011-03-22 16:33:05 -07003525 tmp = h->nr_overcommit_huge_pages;
Andi Kleene5ff2152008-07-23 21:27:42 -07003526
Luiz Capitulinobae7f4a2014-06-04 16:07:08 -07003527 if (write && hstate_is_gigantic(h))
Eric B Munsonadbe8722011-01-13 15:47:27 -08003528 return -EINVAL;
3529
Muchun Song17743792020-09-04 16:36:13 -07003530 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
3531 &tmp);
Michal Hocko08d4a242011-01-13 15:47:26 -08003532 if (ret)
3533 goto out;
Andi Kleene5ff2152008-07-23 21:27:42 -07003534
3535 if (write) {
3536 spin_lock(&hugetlb_lock);
3537 h->nr_overcommit_huge_pages = tmp;
3538 spin_unlock(&hugetlb_lock);
3539 }
Michal Hocko08d4a242011-01-13 15:47:26 -08003540out:
3541 return ret;
Nishanth Aravamudana3d0c6a2008-02-08 04:18:18 -08003542}
3543
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544#endif /* CONFIG_SYSCTL */
3545
Alexey Dobriyane1759c22008-10-15 23:50:22 +04003546void hugetlb_report_meminfo(struct seq_file *m)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547{
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003548 struct hstate *h;
3549 unsigned long total = 0;
3550
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003551 if (!hugepages_supported())
3552 return;
Roman Gushchinfcb2b0c2018-01-31 16:16:22 -08003553
3554 for_each_hstate(h) {
3555 unsigned long count = h->nr_huge_pages;
3556
3557 total += (PAGE_SIZE << huge_page_order(h)) * count;
3558
3559 if (h == &default_hstate)
3560 seq_printf(m,
3561 "HugePages_Total: %5lu\n"
3562 "HugePages_Free: %5lu\n"
3563 "HugePages_Rsvd: %5lu\n"
3564 "HugePages_Surp: %5lu\n"
3565 "Hugepagesize: %8lu kB\n",
3566 count,
3567 h->free_huge_pages,
3568 h->resv_huge_pages,
3569 h->surplus_huge_pages,
3570 (PAGE_SIZE << huge_page_order(h)) / 1024);
3571 }
3572
3573 seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574}
3575
Joe Perches79815932020-09-16 13:40:43 -07003576int hugetlb_report_node_meminfo(char *buf, int len, int nid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577{
Andi Kleena5516432008-07-23 21:27:41 -07003578 struct hstate *h = &default_hstate;
Joe Perches79815932020-09-16 13:40:43 -07003579
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003580 if (!hugepages_supported())
3581 return 0;
Joe Perches79815932020-09-16 13:40:43 -07003582
3583 return sysfs_emit_at(buf, len,
3584 "Node %d HugePages_Total: %5u\n"
3585 "Node %d HugePages_Free: %5u\n"
3586 "Node %d HugePages_Surp: %5u\n",
3587 nid, h->nr_huge_pages_node[nid],
3588 nid, h->free_huge_pages_node[nid],
3589 nid, h->surplus_huge_pages_node[nid]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590}
3591
David Rientjes949f7ec2013-04-29 15:07:48 -07003592void hugetlb_show_meminfo(void)
3593{
3594 struct hstate *h;
3595 int nid;
3596
Nishanth Aravamudan457c1b22014-05-06 12:50:00 -07003597 if (!hugepages_supported())
3598 return;
3599
David Rientjes949f7ec2013-04-29 15:07:48 -07003600 for_each_node_state(nid, N_MEMORY)
3601 for_each_hstate(h)
3602 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3603 nid,
3604 h->nr_huge_pages_node[nid],
3605 h->free_huge_pages_node[nid],
3606 h->surplus_huge_pages_node[nid],
3607 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3608}
3609
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003610void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3611{
3612 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3613 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3614}
3615
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616/* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3617unsigned long hugetlb_total_pages(void)
3618{
Wanpeng Lid0028582013-03-22 15:04:40 -07003619 struct hstate *h;
3620 unsigned long nr_total_pages = 0;
3621
3622 for_each_hstate(h)
3623 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3624 return nr_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626
Andi Kleena5516432008-07-23 21:27:41 -07003627static int hugetlb_acct_memory(struct hstate *h, long delta)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003628{
3629 int ret = -ENOMEM;
3630
3631 spin_lock(&hugetlb_lock);
3632 /*
3633 * When cpuset is configured, it breaks the strict hugetlb page
3634 * reservation as the accounting is done on a global variable. Such
3635 * reservation is completely rubbish in the presence of cpuset because
3636 * the reservation is not checked against page availability for the
3637 * current cpuset. Application can still potentially OOM'ed by kernel
3638 * with lack of free htlb page in cpuset that the task is in.
3639 * Attempt to enforce strict accounting with cpuset is almost
3640 * impossible (or too ugly) because cpuset is too fluid that
3641 * task or memory node can be dynamically moved between cpusets.
3642 *
3643 * The change of semantics for shared hugetlb mapping with cpuset is
3644 * undesirable. However, in order to preserve some of the semantics,
3645 * we fall back to check against current free page availability as
3646 * a best attempt and hopefully to minimize the impact of changing
3647 * semantics that cpuset has.
Muchun Song8ca39e62020-08-11 18:30:32 -07003648 *
3649 * Apart from cpuset, we also have memory policy mechanism that
3650 * also determines from which node the kernel will allocate memory
3651 * in a NUMA system. So similar to cpuset, we also should consider
3652 * the memory policy of the current task. Similar to the description
3653 * above.
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003654 */
3655 if (delta > 0) {
Andi Kleena5516432008-07-23 21:27:41 -07003656 if (gather_surplus_pages(h, delta) < 0)
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003657 goto out;
3658
Muchun Song8ca39e62020-08-11 18:30:32 -07003659 if (delta > allowed_mems_nr(h)) {
Andi Kleena5516432008-07-23 21:27:41 -07003660 return_unused_surplus_pages(h, delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003661 goto out;
3662 }
3663 }
3664
3665 ret = 0;
3666 if (delta < 0)
Andi Kleena5516432008-07-23 21:27:41 -07003667 return_unused_surplus_pages(h, (unsigned long) -delta);
Mel Gormanfc1b8a72008-07-23 21:27:22 -07003668
3669out:
3670 spin_unlock(&hugetlb_lock);
3671 return ret;
3672}
3673
Andy Whitcroft84afd992008-07-23 21:27:32 -07003674static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3675{
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003676 struct resv_map *resv = vma_resv_map(vma);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003677
3678 /*
3679 * This new VMA should share its siblings reservation map if present.
3680 * The VMA will only ever have a valid reservation map pointer where
3681 * it is being copied for another still existing VMA. As that VMA
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003682 * has a reference to the reservation map it cannot disappear until
Andy Whitcroft84afd992008-07-23 21:27:32 -07003683 * after this open call completes. It is therefore safe to take a
3684 * new reference here without additional locking.
3685 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003686 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003687 kref_get(&resv->refs);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003688}
3689
Mel Gormana1e78772008-07-23 21:27:23 -07003690static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3691{
Andi Kleena5516432008-07-23 21:27:41 -07003692 struct hstate *h = hstate_vma(vma);
Joonsoo Kimf522c3a2013-09-11 14:21:53 -07003693 struct resv_map *resv = vma_resv_map(vma);
David Gibson90481622012-03-21 16:34:12 -07003694 struct hugepage_subpool *spool = subpool_vma(vma);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003695 unsigned long reserve, start, end;
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003696 long gbl_reserve;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003697
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003698 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3699 return;
Andy Whitcroft84afd992008-07-23 21:27:32 -07003700
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003701 start = vma_hugecache_offset(h, vma, vma->vm_start);
3702 end = vma_hugecache_offset(h, vma, vma->vm_end);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003703
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003704 reserve = (end - start) - region_count(resv, start, end);
Mina Almasrye9fe92a2020-04-01 21:11:21 -07003705 hugetlb_cgroup_uncharge_counter(resv, start, end);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07003706 if (reserve) {
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07003707 /*
3708 * Decrement reserve counts. The global reserve count may be
3709 * adjusted if the subpool has a minimum size.
3710 */
3711 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3712 hugetlb_acct_memory(h, -gbl_reserve);
Andy Whitcroft84afd992008-07-23 21:27:32 -07003713 }
Mina Almasrye9fe92a2020-04-01 21:11:21 -07003714
3715 kref_put(&resv->refs, resv_map_release);
Mel Gormana1e78772008-07-23 21:27:23 -07003716}
3717
Dan Williams31383c62017-11-29 16:10:28 -08003718static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3719{
3720 if (addr & ~(huge_page_mask(hstate_vma(vma))))
3721 return -EINVAL;
3722 return 0;
3723}
3724
Dan Williams05ea8862018-04-05 16:24:25 -07003725static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3726{
3727 struct hstate *hstate = hstate_vma(vma);
3728
3729 return 1UL << huge_page_shift(hstate);
3730}
3731
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732/*
3733 * We cannot handle pagefaults against hugetlb pages at all. They cause
3734 * handle_mm_fault() to try to instantiate regular-sized pages in the
3735 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3736 * this far.
3737 */
Souptick Joarderb3ec9f32018-06-07 17:08:04 -07003738static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739{
3740 BUG();
Nick Piggind0217ac2007-07-19 01:47:03 -07003741 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003742}
3743
Jane Chueec36362018-08-02 15:36:05 -07003744/*
3745 * When a new function is introduced to vm_operations_struct and added
3746 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3747 * This is because under System V memory model, mappings created via
3748 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3749 * their original vm_ops are overwritten with shm_vm_ops.
3750 */
Alexey Dobriyanf0f37e2f2009-09-27 22:29:37 +04003751const struct vm_operations_struct hugetlb_vm_ops = {
Nick Piggind0217ac2007-07-19 01:47:03 -07003752 .fault = hugetlb_vm_op_fault,
Andy Whitcroft84afd992008-07-23 21:27:32 -07003753 .open = hugetlb_vm_op_open,
Mel Gormana1e78772008-07-23 21:27:23 -07003754 .close = hugetlb_vm_op_close,
Dan Williams31383c62017-11-29 16:10:28 -08003755 .split = hugetlb_vm_op_split,
Dan Williams05ea8862018-04-05 16:24:25 -07003756 .pagesize = hugetlb_vm_op_pagesize,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003757};
3758
David Gibson1e8f8892006-01-06 00:10:44 -08003759static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3760 int writable)
David Gibson63551ae2005-06-21 17:14:44 -07003761{
3762 pte_t entry;
3763
David Gibson1e8f8892006-01-06 00:10:44 -08003764 if (writable) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003765 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3766 vma->vm_page_prot)));
David Gibson63551ae2005-06-21 17:14:44 -07003767 } else {
Gerald Schaefer106c9922013-04-29 15:07:23 -07003768 entry = huge_pte_wrprotect(mk_huge_pte(page,
3769 vma->vm_page_prot));
David Gibson63551ae2005-06-21 17:14:44 -07003770 }
3771 entry = pte_mkyoung(entry);
3772 entry = pte_mkhuge(entry);
Chris Metcalfd9ed9fa2012-04-01 14:01:34 -04003773 entry = arch_make_huge_pte(entry, vma, page, writable);
David Gibson63551ae2005-06-21 17:14:44 -07003774
3775 return entry;
3776}
3777
David Gibson1e8f8892006-01-06 00:10:44 -08003778static void set_huge_ptep_writable(struct vm_area_struct *vma,
3779 unsigned long address, pte_t *ptep)
3780{
3781 pte_t entry;
3782
Gerald Schaefer106c9922013-04-29 15:07:23 -07003783 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
Chris Forbes32f84522011-07-25 17:12:14 -07003784 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
Russell King4b3073e2009-12-18 16:40:18 +00003785 update_mmu_cache(vma, address, ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08003786}
3787
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003788bool is_hugetlb_entry_migration(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003789{
3790 swp_entry_t swp;
3791
3792 if (huge_pte_none(pte) || pte_present(pte))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003793 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003794 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07003795 if (is_migration_entry(swp))
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003796 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003797 else
Aneesh Kumar K.Vd5ed7442017-07-06 15:38:47 -07003798 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003799}
3800
Baoquan He3e5c3602020-10-13 16:56:10 -07003801static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003802{
3803 swp_entry_t swp;
3804
3805 if (huge_pte_none(pte) || pte_present(pte))
Baoquan He3e5c3602020-10-13 16:56:10 -07003806 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003807 swp = pte_to_swp_entry(pte);
Baoquan Hed79d1762020-10-13 16:56:14 -07003808 if (is_hwpoison_entry(swp))
Baoquan He3e5c3602020-10-13 16:56:10 -07003809 return true;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003810 else
Baoquan He3e5c3602020-10-13 16:56:10 -07003811 return false;
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003812}
David Gibson1e8f8892006-01-06 00:10:44 -08003813
David Gibson63551ae2005-06-21 17:14:44 -07003814int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3815 struct vm_area_struct *vma)
3816{
Mike Kravetz5e415402018-11-16 15:08:04 -08003817 pte_t *src_pte, *dst_pte, entry, dst_entry;
David Gibson63551ae2005-06-21 17:14:44 -07003818 struct page *ptepage;
Hugh Dickins1c598272005-10-19 21:23:43 -07003819 unsigned long addr;
David Gibson1e8f8892006-01-06 00:10:44 -08003820 int cow;
Andi Kleena5516432008-07-23 21:27:41 -07003821 struct hstate *h = hstate_vma(vma);
3822 unsigned long sz = huge_page_size(h);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003823 struct address_space *mapping = vma->vm_file->f_mapping;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003824 struct mmu_notifier_range range;
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003825 int ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08003826
3827 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
David Gibson63551ae2005-06-21 17:14:44 -07003828
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003829 if (cow) {
Jérôme Glisse7269f992019-05-13 17:20:53 -07003830 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07003831 vma->vm_start,
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003832 vma->vm_end);
3833 mmu_notifier_invalidate_range_start(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003834 } else {
3835 /*
3836 * For shared mappings i_mmap_rwsem must be held to call
3837 * huge_pte_alloc, otherwise the returned ptep could go
3838 * away if part of a shared pmd and another thread calls
3839 * huge_pmd_unshare.
3840 */
3841 i_mmap_lock_read(mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003842 }
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003843
Andi Kleena5516432008-07-23 21:27:41 -07003844 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003845 spinlock_t *src_ptl, *dst_ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07003846 src_pte = huge_pte_offset(src, addr, sz);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003847 if (!src_pte)
3848 continue;
Andi Kleena5516432008-07-23 21:27:41 -07003849 dst_pte = huge_pte_alloc(dst, addr, sz);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003850 if (!dst_pte) {
3851 ret = -ENOMEM;
3852 break;
3853 }
Larry Woodmanc5c99422008-01-24 05:49:25 -08003854
Mike Kravetz5e415402018-11-16 15:08:04 -08003855 /*
3856 * If the pagetables are shared don't copy or take references.
3857 * dst_pte == src_pte is the common case of src/dest sharing.
3858 *
3859 * However, src could have 'unshared' and dst shares with
3860 * another vma. If dst_pte !none, this implies sharing.
3861 * Check here before taking page table lock, and once again
3862 * after taking the lock below.
3863 */
3864 dst_entry = huge_ptep_get(dst_pte);
3865 if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
Larry Woodmanc5c99422008-01-24 05:49:25 -08003866 continue;
3867
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003868 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3869 src_ptl = huge_pte_lockptr(h, src, src_pte);
3870 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003871 entry = huge_ptep_get(src_pte);
Mike Kravetz5e415402018-11-16 15:08:04 -08003872 dst_entry = huge_ptep_get(dst_pte);
3873 if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3874 /*
3875 * Skip if src entry none. Also, skip in the
3876 * unlikely case dst entry !none as this implies
3877 * sharing with another vma.
3878 */
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003879 ;
3880 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3881 is_hugetlb_entry_hwpoisoned(entry))) {
3882 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3883
3884 if (is_write_migration_entry(swp_entry) && cow) {
3885 /*
3886 * COW mappings require pages in both
3887 * parent and child to be set to read.
3888 */
3889 make_migration_entry_read(&swp_entry);
3890 entry = swp_entry_to_pte(swp_entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07003891 set_huge_swap_pte_at(src, addr, src_pte,
3892 entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003893 }
Punit Agrawale5251fd2017-07-06 15:39:50 -07003894 set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
Naoya Horiguchi4a705fe2014-06-23 13:22:03 -07003895 } else {
Joerg Roedel34ee6452014-11-13 13:46:09 +11003896 if (cow) {
Jérôme Glisse0f108512017-11-15 17:34:07 -08003897 /*
3898 * No need to notify as we are downgrading page
3899 * table protection not changing it to point
3900 * to a new page.
3901 *
Mike Rapoportad56b732018-03-21 21:22:47 +02003902 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08003903 */
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07003904 huge_ptep_set_wrprotect(src, addr, src_pte);
Joerg Roedel34ee6452014-11-13 13:46:09 +11003905 }
Naoya Horiguchi0253d632014-07-23 14:00:19 -07003906 entry = huge_ptep_get(src_pte);
Hugh Dickins1c598272005-10-19 21:23:43 -07003907 ptepage = pte_page(entry);
3908 get_page(ptepage);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08003909 page_dup_rmap(ptepage, true);
Hugh Dickins1c598272005-10-19 21:23:43 -07003910 set_huge_pte_at(dst, addr, dst_pte, entry);
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08003911 hugetlb_count_add(pages_per_huge_page(h), dst);
Hugh Dickins1c598272005-10-19 21:23:43 -07003912 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003913 spin_unlock(src_ptl);
3914 spin_unlock(dst_ptl);
David Gibson63551ae2005-06-21 17:14:44 -07003915 }
David Gibson63551ae2005-06-21 17:14:44 -07003916
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003917 if (cow)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003918 mmu_notifier_invalidate_range_end(&range);
Mike Kravetzc0d03812020-04-01 21:11:05 -07003919 else
3920 i_mmap_unlock_read(mapping);
Andreas Sandberge8569dd2014-01-21 15:49:09 -08003921
3922 return ret;
David Gibson63551ae2005-06-21 17:14:44 -07003923}
3924
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003925void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3926 unsigned long start, unsigned long end,
3927 struct page *ref_page)
David Gibson63551ae2005-06-21 17:14:44 -07003928{
3929 struct mm_struct *mm = vma->vm_mm;
3930 unsigned long address;
David Gibsonc7546f82005-08-05 11:59:35 -07003931 pte_t *ptep;
David Gibson63551ae2005-06-21 17:14:44 -07003932 pte_t pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003933 spinlock_t *ptl;
David Gibson63551ae2005-06-21 17:14:44 -07003934 struct page *page;
Andi Kleena5516432008-07-23 21:27:41 -07003935 struct hstate *h = hstate_vma(vma);
3936 unsigned long sz = huge_page_size(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003937 struct mmu_notifier_range range;
Andi Kleena5516432008-07-23 21:27:41 -07003938
David Gibson63551ae2005-06-21 17:14:44 -07003939 WARN_ON(!is_vm_hugetlb_page(vma));
Andi Kleena5516432008-07-23 21:27:41 -07003940 BUG_ON(start & ~huge_page_mask(h));
3941 BUG_ON(end & ~huge_page_mask(h));
David Gibson63551ae2005-06-21 17:14:44 -07003942
Aneesh Kumar K.V07e32662016-12-12 16:42:40 -08003943 /*
3944 * This is a hugetlb vma, all the pte entries should point
3945 * to huge page.
3946 */
Peter Zijlstraed6a7932018-08-31 14:46:08 +02003947 tlb_change_page_size(tlb, sz);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07003948 tlb_start_vma(tlb, vma);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07003949
3950 /*
3951 * If sharing possible, alert mmu notifiers of worst case.
3952 */
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07003953 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3954 end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08003955 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3956 mmu_notifier_invalidate_range_start(&range);
Hillf Danton569f48b82014-12-10 15:44:41 -08003957 address = start;
Hillf Danton569f48b82014-12-10 15:44:41 -08003958 for (; address < end; address += sz) {
Punit Agrawal7868a202017-07-06 15:39:42 -07003959 ptep = huge_pte_offset(mm, address, sz);
Adam Litke4c887262005-10-29 18:16:46 -07003960 if (!ptep)
David Gibsonc7546f82005-08-05 11:59:35 -07003961 continue;
3962
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08003963 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae2042020-08-11 18:31:38 -07003964 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Aneesh Kumar K.V31d49da2016-07-26 15:24:06 -07003965 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07003966 /*
3967 * We just unmapped a page of PMDs by clearing a PUD.
3968 * The caller's TLB flush range should cover this area.
3969 */
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);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026}
David Gibson63551ae2005-06-21 17:14:44 -07004027
Mel Gormand8333522012-07-31 16:46:20 -07004028void __unmap_hugepage_range_final(struct mmu_gather *tlb,
4029 struct vm_area_struct *vma, unsigned long start,
4030 unsigned long end, struct page *ref_page)
4031{
4032 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
4033
4034 /*
4035 * Clear this flag so that x86's huge_pmd_share page_table_shareable
4036 * test will fail on a vma being torn down, and not grab a page table
4037 * on its way out. We're lucky that the flag has such an appropriate
4038 * name, and can in fact be safely cleared here. We could clear it
4039 * before the __unmap_hugepage_range above, but all that's necessary
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004040 * is to clear it before releasing the i_mmap_rwsem. This works
Mel Gormand8333522012-07-31 16:46:20 -07004041 * because in the context this is called, the VMA is about to be
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08004042 * destroyed and the i_mmap_rwsem is held.
Mel Gormand8333522012-07-31 16:46:20 -07004043 */
4044 vma->vm_flags &= ~VM_MAYSHARE;
4045}
4046
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004047void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004048 unsigned long end, struct page *ref_page)
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004049{
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004050 struct mm_struct *mm;
4051 struct mmu_gather tlb;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004052 unsigned long tlb_start = start;
4053 unsigned long tlb_end = end;
4054
4055 /*
4056 * If shared PMDs were possibly used within this vma range, adjust
4057 * start/end for worst case tlb flushing.
4058 * Note that we can not be sure if PMDs are shared until we try to
4059 * unmap pages. However, we want to make sure TLB flushing covers
4060 * the largest possible range.
4061 */
4062 adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004063
4064 mm = vma->vm_mm;
4065
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004066 tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004067 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07004068 tlb_finish_mmu(&tlb, tlb_start, tlb_end);
Chen, Kenneth W502717f2006-10-11 01:20:46 -07004069}
4070
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004071/*
4072 * This is called when the original mapper is failing to COW a MAP_PRIVATE
4073 * mappping it owns the reserve page for. The intention is to unmap the page
4074 * from other VMAs and let the children be SIGKILLed if they are faulting the
4075 * same region.
4076 */
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004077static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
4078 struct page *page, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004079{
Adam Litke75266742008-11-12 13:24:56 -08004080 struct hstate *h = hstate_vma(vma);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004081 struct vm_area_struct *iter_vma;
4082 struct address_space *mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004083 pgoff_t pgoff;
4084
4085 /*
4086 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
4087 * from page cache lookup which is in HPAGE_SIZE units.
4088 */
Adam Litke75266742008-11-12 13:24:56 -08004089 address = address & huge_page_mask(h);
Michal Hocko36e4f202012-10-08 16:33:31 -07004090 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
4091 vma->vm_pgoff;
Al Viro93c76a32015-12-04 23:45:44 -05004092 mapping = vma->vm_file->f_mapping;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004093
Mel Gorman4eb2b1d2009-12-14 17:59:53 -08004094 /*
4095 * Take the mapping lock for the duration of the table walk. As
4096 * this mapping should be shared between all the VMAs,
4097 * __unmap_hugepage_range() is called as the lock is already held
4098 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004099 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07004100 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004101 /* Do not unmap the current VMA */
4102 if (iter_vma == vma)
4103 continue;
4104
4105 /*
Mel Gorman2f84a892015-10-01 15:36:57 -07004106 * Shared VMAs have their own reserves and do not affect
4107 * MAP_PRIVATE accounting but it is possible that a shared
4108 * VMA is using the same page so check and skip such VMAs.
4109 */
4110 if (iter_vma->vm_flags & VM_MAYSHARE)
4111 continue;
4112
4113 /*
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004114 * Unmap the page from other VMAs without their own reserves.
4115 * They get marked to be SIGKILLed if they fault in these
4116 * areas. This is because a future no-page fault on this VMA
4117 * could insert a zeroed page instead of the data existing
4118 * from the time of fork. This would look like data corruption
4119 */
4120 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07004121 unmap_hugepage_range(iter_vma, address,
4122 address + huge_page_size(h), page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004123 }
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08004124 i_mmap_unlock_write(mapping);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004125}
4126
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004127/*
4128 * Hugetlb_cow() should be called with page lock of the original hugepage held.
Michal Hockoef009b22012-01-10 15:07:21 -08004129 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
4130 * cannot race with other handlers or page migration.
4131 * Keep the pte_same checks anyway to make transition from the mutex easier.
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004132 */
Souptick Joarder2b740302018-08-23 17:01:36 -07004133static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
Huang Ying974e6d62018-08-17 15:45:57 -07004134 unsigned long address, pte_t *ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004135 struct page *pagecache_page, spinlock_t *ptl)
David Gibson1e8f8892006-01-06 00:10:44 -08004136{
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004137 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07004138 struct hstate *h = hstate_vma(vma);
David Gibson1e8f8892006-01-06 00:10:44 -08004139 struct page *old_page, *new_page;
Souptick Joarder2b740302018-08-23 17:01:36 -07004140 int outside_reserve = 0;
4141 vm_fault_t ret = 0;
Huang Ying974e6d62018-08-17 15:45:57 -07004142 unsigned long haddr = address & huge_page_mask(h);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004143 struct mmu_notifier_range range;
David Gibson1e8f8892006-01-06 00:10:44 -08004144
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004145 pte = huge_ptep_get(ptep);
David Gibson1e8f8892006-01-06 00:10:44 -08004146 old_page = pte_page(pte);
4147
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004148retry_avoidcopy:
David Gibson1e8f8892006-01-06 00:10:44 -08004149 /* If no-one else is actually using this page, avoid the copy
4150 * and just make the page writable */
Joonsoo Kim37a21402013-09-11 14:21:04 -07004151 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
Hugh Dickins5a499732016-07-14 12:07:38 -07004152 page_move_anon_rmap(old_page, vma);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004153 set_huge_ptep_writable(vma, haddr, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004154 return 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004155 }
4156
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004157 /*
4158 * If the process that created a MAP_PRIVATE mapping is about to
4159 * perform a COW due to a shared page count, attempt to satisfy
4160 * the allocation without using the existing reserves. The pagecache
4161 * page is used to determine if the reserve at this address was
4162 * consumed or not. If reserves were used, a partial faulted mapping
4163 * at the time of fork() could consume its reserves on COW instead
4164 * of the full address range.
4165 */
Joonsoo Kim5944d012013-09-11 14:21:55 -07004166 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004167 old_page != pagecache_page)
4168 outside_reserve = 1;
4169
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004170 get_page(old_page);
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004171
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004172 /*
4173 * Drop page table lock as buddy allocator may be called. It will
4174 * be acquired again before returning to the caller, as expected.
4175 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004176 spin_unlock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004177 new_page = alloc_huge_page(vma, haddr, outside_reserve);
David Gibson1e8f8892006-01-06 00:10:44 -08004178
Adam Litke2fc39ce2007-11-14 16:59:39 -08004179 if (IS_ERR(new_page)) {
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004180 /*
4181 * If a process owning a MAP_PRIVATE mapping fails to COW,
4182 * it is due to references held by a child and an insufficient
4183 * huge page pool. To guarantee the original mappers
4184 * reliability, unmap the page from child processes. The child
4185 * may get SIGKILLed if it later faults.
4186 */
4187 if (outside_reserve) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004188 put_page(old_page);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004189 BUG_ON(huge_pte_none(pte));
Huang Ying5b7a1d42018-08-17 15:45:53 -07004190 unmap_ref_private(mm, vma, old_page, haddr);
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004191 BUG_ON(huge_pte_none(pte));
4192 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004193 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Davidlohr Bueso2f4612a2014-08-06 16:06:45 -07004194 if (likely(ptep &&
4195 pte_same(huge_ptep_get(ptep), pte)))
4196 goto retry_avoidcopy;
4197 /*
4198 * race occurs while re-acquiring page table
4199 * lock, and our job is done.
4200 */
4201 return 0;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004202 }
4203
Souptick Joarder2b740302018-08-23 17:01:36 -07004204 ret = vmf_error(PTR_ERR(new_page));
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004205 goto out_release_old;
David Gibson1e8f8892006-01-06 00:10:44 -08004206 }
4207
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004208 /*
4209 * When the original hugepage is shared one, it does not have
4210 * anon_vma prepared.
4211 */
Dean Nelson44e2aa92010-10-26 14:22:08 -07004212 if (unlikely(anon_vma_prepare(vma))) {
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004213 ret = VM_FAULT_OOM;
4214 goto out_release_all;
Dean Nelson44e2aa92010-10-26 14:22:08 -07004215 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004216
Huang Ying974e6d62018-08-17 15:45:57 -07004217 copy_user_huge_page(new_page, old_page, address, vma,
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004218 pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004219 __SetPageUptodate(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004220
Jérôme Glisse7269f992019-05-13 17:20:53 -07004221 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
Jérôme Glisse6f4f13e2019-05-13 17:20:49 -07004222 haddr + huge_page_size(h));
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004223 mmu_notifier_invalidate_range_start(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004224
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004225 /*
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004226 * Retake the page table lock to check for racing updates
Larry Woodmanb76c8cf2009-12-14 17:59:37 -08004227 * before the page tables are altered
4228 */
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004229 spin_lock(ptl);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004230 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchia9af0c52014-04-07 15:36:54 -07004231 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
Joonsoo Kim07443a82013-09-11 14:21:58 -07004232 ClearPagePrivate(new_page);
4233
David Gibson1e8f8892006-01-06 00:10:44 -08004234 /* Break COW */
Huang Ying5b7a1d42018-08-17 15:45:53 -07004235 huge_ptep_clear_flush(vma, haddr, ptep);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004236 mmu_notifier_invalidate_range(mm, range.start, range.end);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004237 set_huge_pte_at(mm, haddr, ptep,
David Gibson1e8f8892006-01-06 00:10:44 -08004238 make_huge_pte(vma, new_page, 1));
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08004239 page_remove_rmap(old_page, true);
Huang Ying5b7a1d42018-08-17 15:45:53 -07004240 hugepage_add_new_anon_rmap(new_page, vma, haddr);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004241 set_page_huge_active(new_page);
David Gibson1e8f8892006-01-06 00:10:44 -08004242 /* Make the old page be freed below */
4243 new_page = old_page;
4244 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004245 spin_unlock(ptl);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08004246 mmu_notifier_invalidate_range_end(&range);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004247out_release_all:
Huang Ying5b7a1d42018-08-17 15:45:53 -07004248 restore_reserve_on_error(h, vma, haddr, new_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004249 put_page(new_page);
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004250out_release_old:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03004251 put_page(old_page);
Joonsoo Kim83120342013-09-11 14:21:57 -07004252
Davidlohr Buesoad4404a2014-08-06 16:06:47 -07004253 spin_lock(ptl); /* Caller expects lock to be held */
4254 return ret;
David Gibson1e8f8892006-01-06 00:10:44 -08004255}
4256
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004257/* Return the pagecache page at a given address within a VMA */
Andi Kleena5516432008-07-23 21:27:41 -07004258static struct page *hugetlbfs_pagecache_page(struct hstate *h,
4259 struct vm_area_struct *vma, unsigned long address)
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004260{
4261 struct address_space *mapping;
Andy Whitcrofte7c4b0b2008-07-23 21:27:26 -07004262 pgoff_t idx;
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004263
4264 mapping = vma->vm_file->f_mapping;
Andi Kleena5516432008-07-23 21:27:41 -07004265 idx = vma_hugecache_offset(h, vma, address);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004266
4267 return find_lock_page(mapping, idx);
4268}
4269
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004270/*
4271 * Return whether there is a pagecache page to back given address within VMA.
4272 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
4273 */
4274static bool hugetlbfs_pagecache_present(struct hstate *h,
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004275 struct vm_area_struct *vma, unsigned long address)
4276{
4277 struct address_space *mapping;
4278 pgoff_t idx;
4279 struct page *page;
4280
4281 mapping = vma->vm_file->f_mapping;
4282 idx = vma_hugecache_offset(h, vma, address);
4283
4284 page = find_get_page(mapping, idx);
4285 if (page)
4286 put_page(page);
4287 return page != NULL;
4288}
4289
Mike Kravetzab76ad52015-09-08 15:01:50 -07004290int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
4291 pgoff_t idx)
4292{
4293 struct inode *inode = mapping->host;
4294 struct hstate *h = hstate_inode(inode);
4295 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
4296
4297 if (err)
4298 return err;
4299 ClearPagePrivate(page);
4300
Mike Kravetz22146c32018-10-26 15:10:58 -07004301 /*
4302 * set page dirty so that it will not be removed from cache/file
4303 * by non-hugetlbfs specific code paths.
4304 */
4305 set_page_dirty(page);
4306
Mike Kravetzab76ad52015-09-08 15:01:50 -07004307 spin_lock(&inode->i_lock);
4308 inode->i_blocks += blocks_per_huge_page(h);
4309 spin_unlock(&inode->i_lock);
4310 return 0;
4311}
4312
Souptick Joarder2b740302018-08-23 17:01:36 -07004313static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
4314 struct vm_area_struct *vma,
4315 struct address_space *mapping, pgoff_t idx,
4316 unsigned long address, pte_t *ptep, unsigned int flags)
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004317{
Andi Kleena5516432008-07-23 21:27:41 -07004318 struct hstate *h = hstate_vma(vma);
Souptick Joarder2b740302018-08-23 17:01:36 -07004319 vm_fault_t ret = VM_FAULT_SIGBUS;
Hillf Danton409eb8c2012-01-20 14:34:13 -08004320 int anon_rmap = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004321 unsigned long size;
Adam Litke4c887262005-10-29 18:16:46 -07004322 struct page *page;
David Gibson1e8f8892006-01-06 00:10:44 -08004323 pte_t new_pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004324 spinlock_t *ptl;
Huang Ying285b8dc2018-06-07 17:08:08 -07004325 unsigned long haddr = address & huge_page_mask(h);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004326 bool new_page = false;
Adam Litke4c887262005-10-29 18:16:46 -07004327
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004328 /*
4329 * Currently, we are forced to kill the process in the event the
4330 * original mapper has unmapped pages from the child due to a failed
Lucas De Marchi25985ed2011-03-30 22:57:33 -03004331 * COW. Warn that such a situation has occurred as it may not be obvious
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004332 */
4333 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
Geoffrey Thomas910154d2016-03-09 14:08:04 -08004334 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
Andrew Mortonffb22af2013-02-22 16:32:08 -08004335 current->pid);
Mel Gorman04f2cbe2008-07-23 21:27:25 -07004336 return ret;
4337 }
4338
Adam Litke4c887262005-10-29 18:16:46 -07004339 /*
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004340 * We can not race with truncation due to holding i_mmap_rwsem.
4341 * i_size is modified when holding i_mmap_rwsem, so check here
4342 * once for faults beyond end of file.
Adam Litke4c887262005-10-29 18:16:46 -07004343 */
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004344 size = i_size_read(mapping->host) >> huge_page_shift(h);
4345 if (idx >= size)
4346 goto out;
4347
Christoph Lameter6bda6662006-01-06 00:10:49 -08004348retry:
4349 page = find_lock_page(mapping, idx);
4350 if (!page) {
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004351 /*
4352 * Check for page in userfault range
4353 */
4354 if (userfaultfd_missing(vma)) {
4355 u32 hash;
4356 struct vm_fault vmf = {
4357 .vma = vma,
Huang Ying285b8dc2018-06-07 17:08:08 -07004358 .address = haddr,
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004359 .flags = flags,
4360 /*
4361 * Hard to debug if it ends up being
4362 * used by a callee that assumes
4363 * something about the other
4364 * uninitialized fields... same as in
4365 * memory.c
4366 */
4367 };
4368
4369 /*
Mike Kravetzc0d03812020-04-01 21:11:05 -07004370 * hugetlb_fault_mutex and i_mmap_rwsem must be
4371 * dropped before handling userfault. Reacquire
4372 * after handling fault to make calling code simpler.
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004373 */
Wei Yang188b04a2019-11-30 17:57:02 -08004374 hash = hugetlb_fault_mutex_hash(mapping, idx);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004375 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004376 i_mmap_unlock_read(mapping);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004377 ret = handle_userfault(&vmf, VM_UFFD_MISSING);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004378 i_mmap_lock_read(mapping);
Mike Kravetz1a1aad82017-02-22 15:43:01 -08004379 mutex_lock(&hugetlb_fault_mutex_table[hash]);
4380 goto out;
4381 }
4382
Huang Ying285b8dc2018-06-07 17:08:08 -07004383 page = alloc_huge_page(vma, haddr, 0);
Adam Litke2fc39ce2007-11-14 16:59:39 -08004384 if (IS_ERR(page)) {
Mike Kravetz4643d672019-08-13 15:38:00 -07004385 /*
4386 * Returning error will result in faulting task being
4387 * sent SIGBUS. The hugetlb fault mutex prevents two
4388 * tasks from racing to fault in the same page which
4389 * could result in false unable to allocate errors.
4390 * Page migration does not take the fault mutex, but
4391 * does a clear then write of pte's under page table
4392 * lock. Page fault code could race with migration,
4393 * notice the clear pte and try to allocate a page
4394 * here. Before returning error, get ptl and make
4395 * sure there really is no pte entry.
4396 */
4397 ptl = huge_pte_lock(h, mm, ptep);
4398 if (!huge_pte_none(huge_ptep_get(ptep))) {
4399 ret = 0;
4400 spin_unlock(ptl);
4401 goto out;
4402 }
4403 spin_unlock(ptl);
Souptick Joarder2b740302018-08-23 17:01:36 -07004404 ret = vmf_error(PTR_ERR(page));
Christoph Lameter6bda6662006-01-06 00:10:49 -08004405 goto out;
4406 }
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08004407 clear_huge_page(page, address, pages_per_huge_page(h));
Nick Piggin0ed361d2008-02-04 22:29:34 -08004408 __SetPageUptodate(page);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004409 new_page = true;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004410
Mel Gormanf83a2752009-05-28 14:34:40 -07004411 if (vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzab76ad52015-09-08 15:01:50 -07004412 int err = huge_add_to_page_cache(page, mapping, idx);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004413 if (err) {
4414 put_page(page);
Christoph Lameter6bda6662006-01-06 00:10:49 -08004415 if (err == -EEXIST)
4416 goto retry;
4417 goto out;
4418 }
Mel Gorman23be7462010-04-23 13:17:56 -04004419 } else {
Christoph Lameter6bda6662006-01-06 00:10:49 -08004420 lock_page(page);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004421 if (unlikely(anon_vma_prepare(vma))) {
4422 ret = VM_FAULT_OOM;
4423 goto backout_unlocked;
4424 }
Hillf Danton409eb8c2012-01-20 14:34:13 -08004425 anon_rmap = 1;
Mel Gorman23be7462010-04-23 13:17:56 -04004426 }
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004427 } else {
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004428 /*
4429 * If memory error occurs between mmap() and fault, some process
4430 * don't have hwpoisoned swap entry for errored virtual address.
4431 * So we need to block hugepage fault by PG_hwpoison bit check.
4432 */
4433 if (unlikely(PageHWPoison(page))) {
Chris Forbes32f84522011-07-25 17:12:14 -07004434 ret = VM_FAULT_HWPOISON |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004435 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchi998b4382010-09-08 10:19:32 +09004436 goto backout_unlocked;
4437 }
Christoph Lameter6bda6662006-01-06 00:10:49 -08004438 }
David Gibson1e8f8892006-01-06 00:10:44 -08004439
Andy Whitcroft57303d82008-08-12 15:08:47 -07004440 /*
4441 * If we are going to COW a private mapping later, we examine the
4442 * pending reservations for this page now. This will ensure that
4443 * any allocations necessary to record that reservation occur outside
4444 * the spinlock.
4445 */
Mike Kravetz5e911372015-09-08 15:01:28 -07004446 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004447 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004448 ret = VM_FAULT_OOM;
4449 goto backout_unlocked;
4450 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004451 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004452 vma_end_reservation(h, vma, haddr);
Mike Kravetz5e911372015-09-08 15:01:28 -07004453 }
Andy Whitcroft57303d82008-08-12 15:08:47 -07004454
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08004455 ptl = huge_pte_lock(h, mm, ptep);
Nick Piggin83c54072007-07-19 01:47:05 -07004456 ret = 0;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004457 if (!huge_pte_none(huge_ptep_get(ptep)))
Adam Litke4c887262005-10-29 18:16:46 -07004458 goto backout;
4459
Joonsoo Kim07443a82013-09-11 14:21:58 -07004460 if (anon_rmap) {
4461 ClearPagePrivate(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004462 hugepage_add_new_anon_rmap(page, vma, haddr);
Choi Gi-yongac714902014-04-07 15:37:36 -07004463 } else
Kirill A. Shutemov53f92632016-01-15 16:53:42 -08004464 page_dup_rmap(page, true);
David Gibson1e8f8892006-01-06 00:10:44 -08004465 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
4466 && (vma->vm_flags & VM_SHARED)));
Huang Ying285b8dc2018-06-07 17:08:08 -07004467 set_huge_pte_at(mm, haddr, ptep, new_pte);
David Gibson1e8f8892006-01-06 00:10:44 -08004468
Naoya Horiguchi5d317b22015-11-05 18:47:14 -08004469 hugetlb_count_add(pages_per_huge_page(h), mm);
Hugh Dickins788c7df2009-06-23 13:49:05 +01004470 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
David Gibson1e8f8892006-01-06 00:10:44 -08004471 /* Optimization, do the COW without a second fault */
Huang Ying974e6d62018-08-17 15:45:57 -07004472 ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
David Gibson1e8f8892006-01-06 00:10:44 -08004473 }
4474
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004475 spin_unlock(ptl);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004476
4477 /*
4478 * Only make newly allocated pages active. Existing pages found
4479 * in the pagecache could be !page_huge_active() if they have been
4480 * isolated for migration.
4481 */
4482 if (new_page)
4483 set_page_huge_active(page);
4484
Adam Litke4c887262005-10-29 18:16:46 -07004485 unlock_page(page);
4486out:
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004487 return ret;
Adam Litke4c887262005-10-29 18:16:46 -07004488
4489backout:
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004490 spin_unlock(ptl);
Andy Whitcroft2b267362008-08-12 15:08:49 -07004491backout_unlocked:
Adam Litke4c887262005-10-29 18:16:46 -07004492 unlock_page(page);
Huang Ying285b8dc2018-06-07 17:08:08 -07004493 restore_reserve_on_error(h, vma, haddr, page);
Adam Litke4c887262005-10-29 18:16:46 -07004494 put_page(page);
4495 goto out;
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01004496}
4497
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004498#ifdef CONFIG_SMP
Wei Yang188b04a2019-11-30 17:57:02 -08004499u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004500{
4501 unsigned long key[2];
4502 u32 hash;
4503
Mike Kravetz1b426ba2019-05-13 17:19:41 -07004504 key[0] = (unsigned long) mapping;
4505 key[1] = idx;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004506
Mike Kravetz55254632019-11-30 17:56:30 -08004507 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004508
4509 return hash & (num_fault_mutexes - 1);
4510}
4511#else
4512/*
4513 * For uniprocesor systems we always use a single mutex, so just
4514 * return 0 and avoid the hashing overhead.
4515 */
Wei Yang188b04a2019-11-30 17:57:02 -08004516u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004517{
4518 return 0;
4519}
4520#endif
4521
Souptick Joarder2b740302018-08-23 17:01:36 -07004522vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins788c7df2009-06-23 13:49:05 +01004523 unsigned long address, unsigned int flags)
Adam Litke86e52162006-01-06 00:10:43 -08004524{
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004525 pte_t *ptep, entry;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004526 spinlock_t *ptl;
Souptick Joarder2b740302018-08-23 17:01:36 -07004527 vm_fault_t ret;
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004528 u32 hash;
4529 pgoff_t idx;
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004530 struct page *page = NULL;
Andy Whitcroft57303d82008-08-12 15:08:47 -07004531 struct page *pagecache_page = NULL;
Andi Kleena5516432008-07-23 21:27:41 -07004532 struct hstate *h = hstate_vma(vma);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004533 struct address_space *mapping;
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004534 int need_wait_lock = 0;
Huang Ying285b8dc2018-06-07 17:08:08 -07004535 unsigned long haddr = address & huge_page_mask(h);
Adam Litke86e52162006-01-06 00:10:43 -08004536
Huang Ying285b8dc2018-06-07 17:08:08 -07004537 ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004538 if (ptep) {
Mike Kravetzc0d03812020-04-01 21:11:05 -07004539 /*
4540 * Since we hold no locks, ptep could be stale. That is
4541 * OK as we are only making decisions based on content and
4542 * not actually modifying content here.
4543 */
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004544 entry = huge_ptep_get(ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09004545 if (unlikely(is_hugetlb_entry_migration(entry))) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004546 migration_entry_wait_huge(vma, mm, ptep);
Naoya Horiguchi290408d2010-09-08 10:19:35 +09004547 return 0;
4548 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
Chris Forbes32f84522011-07-25 17:12:14 -07004549 return VM_FAULT_HWPOISON_LARGE |
Aneesh Kumar K.V972dc4d2012-07-31 16:42:00 -07004550 VM_FAULT_SET_HINDEX(hstate_index(h));
Naoya Horiguchifd6a03e2010-05-28 09:29:21 +09004551 }
4552
Mike Kravetzc0d03812020-04-01 21:11:05 -07004553 /*
4554 * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold
Mike Kravetz87bf91d2020-04-01 21:11:08 -07004555 * until finished with ptep. This serves two purposes:
4556 * 1) It prevents huge_pmd_unshare from being called elsewhere
4557 * and making the ptep no longer valid.
4558 * 2) It synchronizes us with i_size modifications during truncation.
Mike Kravetzc0d03812020-04-01 21:11:05 -07004559 *
4560 * ptep could have already be assigned via huge_pte_offset. That
4561 * is OK, as huge_pte_alloc will return the same value unless
4562 * something has changed.
4563 */
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004564 mapping = vma->vm_file->f_mapping;
Mike Kravetzc0d03812020-04-01 21:11:05 -07004565 i_mmap_lock_read(mapping);
4566 ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4567 if (!ptep) {
4568 i_mmap_unlock_read(mapping);
4569 return VM_FAULT_OOM;
4570 }
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004571
David Gibson3935baa2006-03-22 00:08:53 -08004572 /*
4573 * Serialize hugepage allocation and instantiation, so that we don't
4574 * get spurious allocation failures if two CPUs race to instantiate
4575 * the same page in the page cache.
4576 */
Mike Kravetzc0d03812020-04-01 21:11:05 -07004577 idx = vma_hugecache_offset(h, vma, haddr);
Wei Yang188b04a2019-11-30 17:57:02 -08004578 hash = hugetlb_fault_mutex_hash(mapping, idx);
Mike Kravetzc672c7f2015-09-08 15:01:35 -07004579 mutex_lock(&hugetlb_fault_mutex_table[hash]);
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004580
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004581 entry = huge_ptep_get(ptep);
4582 if (huge_pte_none(entry)) {
Davidlohr Bueso8382d912014-04-03 14:47:31 -07004583 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
David Gibsonb4d1d992008-10-15 22:01:11 -07004584 goto out_mutex;
David Gibson3935baa2006-03-22 00:08:53 -08004585 }
Adam Litke86e52162006-01-06 00:10:43 -08004586
Nick Piggin83c54072007-07-19 01:47:05 -07004587 ret = 0;
David Gibson1e8f8892006-01-06 00:10:44 -08004588
Andy Whitcroft57303d82008-08-12 15:08:47 -07004589 /*
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004590 * entry could be a migration/hwpoison entry at this point, so this
4591 * check prevents the kernel from going below assuming that we have
Ethon Paul7c8de352020-06-04 16:49:07 -07004592 * an active hugepage in pagecache. This goto expects the 2nd page
4593 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
4594 * properly handle it.
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004595 */
4596 if (!pte_present(entry))
4597 goto out_mutex;
4598
4599 /*
Andy Whitcroft57303d82008-08-12 15:08:47 -07004600 * If we are going to COW the mapping later, we examine the pending
4601 * reservations for this page now. This will ensure that any
4602 * allocations necessary to record that reservation occur outside the
4603 * spinlock. For private mappings, we also lookup the pagecache
4604 * page now as it is used to determine if a reservation has been
4605 * consumed.
4606 */
Gerald Schaefer106c9922013-04-29 15:07:23 -07004607 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
Huang Ying285b8dc2018-06-07 17:08:08 -07004608 if (vma_needs_reservation(h, vma, haddr) < 0) {
Andy Whitcroft2b267362008-08-12 15:08:49 -07004609 ret = VM_FAULT_OOM;
David Gibsonb4d1d992008-10-15 22:01:11 -07004610 goto out_mutex;
Andy Whitcroft2b267362008-08-12 15:08:49 -07004611 }
Mike Kravetz5e911372015-09-08 15:01:28 -07004612 /* Just decrements count, does not deallocate */
Huang Ying285b8dc2018-06-07 17:08:08 -07004613 vma_end_reservation(h, vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004614
Mel Gormanf83a2752009-05-28 14:34:40 -07004615 if (!(vma->vm_flags & VM_MAYSHARE))
Andy Whitcroft57303d82008-08-12 15:08:47 -07004616 pagecache_page = hugetlbfs_pagecache_page(h,
Huang Ying285b8dc2018-06-07 17:08:08 -07004617 vma, haddr);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004618 }
4619
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004620 ptl = huge_pte_lock(h, mm, ptep);
Naoya Horiguchi0fe6e202010-05-28 09:29:16 +09004621
David Gibson1e8f8892006-01-06 00:10:44 -08004622 /* Check for a racing update before calling hugetlb_cow */
David Gibsonb4d1d992008-10-15 22:01:11 -07004623 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004624 goto out_ptl;
David Gibsonb4d1d992008-10-15 22:01:11 -07004625
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004626 /*
4627 * hugetlb_cow() requires page locks of pte_page(entry) and
4628 * pagecache_page, so here we need take the former one
4629 * when page != pagecache_page or !pagecache_page.
4630 */
4631 page = pte_page(entry);
4632 if (page != pagecache_page)
4633 if (!trylock_page(page)) {
4634 need_wait_lock = 1;
4635 goto out_ptl;
4636 }
4637
4638 get_page(page);
David Gibsonb4d1d992008-10-15 22:01:11 -07004639
Hugh Dickins788c7df2009-06-23 13:49:05 +01004640 if (flags & FAULT_FLAG_WRITE) {
Gerald Schaefer106c9922013-04-29 15:07:23 -07004641 if (!huge_pte_write(entry)) {
Huang Ying974e6d62018-08-17 15:45:57 -07004642 ret = hugetlb_cow(mm, vma, address, ptep,
Aneesh Kumar K.V3999f522016-12-12 16:41:56 -08004643 pagecache_page, ptl);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004644 goto out_put_page;
David Gibsonb4d1d992008-10-15 22:01:11 -07004645 }
Gerald Schaefer106c9922013-04-29 15:07:23 -07004646 entry = huge_pte_mkdirty(entry);
David Gibsonb4d1d992008-10-15 22:01:11 -07004647 }
4648 entry = pte_mkyoung(entry);
Huang Ying285b8dc2018-06-07 17:08:08 -07004649 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
Hugh Dickins788c7df2009-06-23 13:49:05 +01004650 flags & FAULT_FLAG_WRITE))
Huang Ying285b8dc2018-06-07 17:08:08 -07004651 update_mmu_cache(vma, haddr, ptep);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004652out_put_page:
4653 if (page != pagecache_page)
4654 unlock_page(page);
4655 put_page(page);
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004656out_ptl:
4657 spin_unlock(ptl);
Andy Whitcroft57303d82008-08-12 15:08:47 -07004658
4659 if (pagecache_page) {
4660 unlock_page(pagecache_page);
4661 put_page(pagecache_page);
4662 }
David Gibsonb4d1d992008-10-15 22:01:11 -07004663out_mutex:
Mike Kravetzc672c7f2015-09-08 15:01:35 -07004664 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
Mike Kravetzc0d03812020-04-01 21:11:05 -07004665 i_mmap_unlock_read(mapping);
Naoya Horiguchi0f792cf2015-02-11 15:25:25 -08004666 /*
4667 * Generally it's safe to hold refcount during waiting page lock. But
4668 * here we just wait to defer the next page fault to avoid busy loop and
4669 * the page is not used after unlocked before returning from the current
4670 * page fault. So we are safe from accessing freed page, even if we wait
4671 * here without taking refcount.
4672 */
4673 if (need_wait_lock)
4674 wait_on_page_locked(page);
David Gibson1e8f8892006-01-06 00:10:44 -08004675 return ret;
Adam Litke86e52162006-01-06 00:10:43 -08004676}
4677
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004678/*
4679 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4680 * modifications for huge pages.
4681 */
4682int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4683 pte_t *dst_pte,
4684 struct vm_area_struct *dst_vma,
4685 unsigned long dst_addr,
4686 unsigned long src_addr,
4687 struct page **pagep)
4688{
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004689 struct address_space *mapping;
4690 pgoff_t idx;
4691 unsigned long size;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004692 int vm_shared = dst_vma->vm_flags & VM_SHARED;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004693 struct hstate *h = hstate_vma(dst_vma);
4694 pte_t _dst_pte;
4695 spinlock_t *ptl;
4696 int ret;
4697 struct page *page;
4698
4699 if (!*pagep) {
4700 ret = -ENOMEM;
4701 page = alloc_huge_page(dst_vma, dst_addr, 0);
4702 if (IS_ERR(page))
4703 goto out;
4704
4705 ret = copy_huge_page_from_user(page,
4706 (const void __user *) src_addr,
Mike Kravetz810a56b2017-02-22 15:42:58 -08004707 pages_per_huge_page(h), false);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004708
Michel Lespinassec1e8d7c2020-06-08 21:33:54 -07004709 /* fallback to copy_from_user outside mmap_lock */
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004710 if (unlikely(ret)) {
Andrea Arcangeli9e368252018-11-30 14:09:25 -08004711 ret = -ENOENT;
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004712 *pagep = page;
4713 /* don't free the page */
4714 goto out;
4715 }
4716 } else {
4717 page = *pagep;
4718 *pagep = NULL;
4719 }
4720
4721 /*
4722 * The memory barrier inside __SetPageUptodate makes sure that
4723 * preceding stores to the page contents become visible before
4724 * the set_pte_at() write.
4725 */
4726 __SetPageUptodate(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004727
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004728 mapping = dst_vma->vm_file->f_mapping;
4729 idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4730
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004731 /*
4732 * If shared, add to page cache
4733 */
4734 if (vm_shared) {
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004735 size = i_size_read(mapping->host) >> huge_page_shift(h);
4736 ret = -EFAULT;
4737 if (idx >= size)
4738 goto out_release_nounlock;
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004739
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004740 /*
4741 * Serialization between remove_inode_hugepages() and
4742 * huge_add_to_page_cache() below happens through the
4743 * hugetlb_fault_mutex_table that here must be hold by
4744 * the caller.
4745 */
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004746 ret = huge_add_to_page_cache(page, mapping, idx);
4747 if (ret)
4748 goto out_release_nounlock;
4749 }
4750
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004751 ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4752 spin_lock(ptl);
4753
Andrea Arcangeli1e3921472017-11-02 15:59:29 -07004754 /*
4755 * Recheck the i_size after holding PT lock to make sure not
4756 * to leave any page mapped (as page_mapped()) beyond the end
4757 * of the i_size (remove_inode_hugepages() is strict about
4758 * enforcing that). If we bail out here, we'll also leave a
4759 * page in the radix tree in the vm_shared case beyond the end
4760 * of the i_size, but remove_inode_hugepages() will take care
4761 * of it as soon as we drop the hugetlb_fault_mutex_table.
4762 */
4763 size = i_size_read(mapping->host) >> huge_page_shift(h);
4764 ret = -EFAULT;
4765 if (idx >= size)
4766 goto out_release_unlock;
4767
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004768 ret = -EEXIST;
4769 if (!huge_pte_none(huge_ptep_get(dst_pte)))
4770 goto out_release_unlock;
4771
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004772 if (vm_shared) {
4773 page_dup_rmap(page, true);
4774 } else {
4775 ClearPagePrivate(page);
4776 hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4777 }
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004778
4779 _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4780 if (dst_vma->vm_flags & VM_WRITE)
4781 _dst_pte = huge_pte_mkdirty(_dst_pte);
4782 _dst_pte = pte_mkyoung(_dst_pte);
4783
4784 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4785
4786 (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4787 dst_vma->vm_flags & VM_WRITE);
4788 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4789
4790 /* No need to invalidate - it was non-present before */
4791 update_mmu_cache(dst_vma, dst_addr, dst_pte);
4792
4793 spin_unlock(ptl);
Mike Kravetzcb6acd02019-02-28 16:22:02 -08004794 set_page_huge_active(page);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004795 if (vm_shared)
4796 unlock_page(page);
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004797 ret = 0;
4798out:
4799 return ret;
4800out_release_unlock:
4801 spin_unlock(ptl);
Mike Kravetz1c9e8de2017-02-22 15:43:43 -08004802 if (vm_shared)
4803 unlock_page(page);
Andrea Arcangeli5af10df2017-08-10 15:23:38 -07004804out_release_nounlock:
Mike Kravetz8fb5deb2017-02-22 15:42:52 -08004805 put_page(page);
4806 goto out;
4807}
4808
Michel Lespinasse28a35712013-02-22 16:35:55 -08004809long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4810 struct page **pages, struct vm_area_struct **vmas,
4811 unsigned long *position, unsigned long *nr_pages,
Peter Xu4f6da932020-04-01 21:07:58 -07004812 long i, unsigned int flags, int *locked)
David Gibson63551ae2005-06-21 17:14:44 -07004813{
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004814 unsigned long pfn_offset;
4815 unsigned long vaddr = *position;
Michel Lespinasse28a35712013-02-22 16:35:55 -08004816 unsigned long remainder = *nr_pages;
Andi Kleena5516432008-07-23 21:27:41 -07004817 struct hstate *h = hstate_vma(vma);
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004818 int err = -EFAULT;
David Gibson63551ae2005-06-21 17:14:44 -07004819
David Gibson63551ae2005-06-21 17:14:44 -07004820 while (vaddr < vma->vm_end && remainder) {
Adam Litke4c887262005-10-29 18:16:46 -07004821 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004822 spinlock_t *ptl = NULL;
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004823 int absent;
Adam Litke4c887262005-10-29 18:16:46 -07004824 struct page *page;
4825
4826 /*
David Rientjes02057962015-04-14 15:48:24 -07004827 * If we have a pending SIGKILL, don't keep faulting pages and
4828 * potentially allocating memory.
4829 */
Davidlohr Buesofa45f112019-01-03 15:28:55 -08004830 if (fatal_signal_pending(current)) {
David Rientjes02057962015-04-14 15:48:24 -07004831 remainder = 0;
4832 break;
4833 }
4834
4835 /*
Adam Litke4c887262005-10-29 18:16:46 -07004836 * Some archs (sparc64, sh*) have multiple pte_ts to
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004837 * each hugepage. We have to make sure we get the
Adam Litke4c887262005-10-29 18:16:46 -07004838 * first, for the page indexing below to work.
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004839 *
4840 * Note that page table lock is not held when pte is null.
Adam Litke4c887262005-10-29 18:16:46 -07004841 */
Punit Agrawal7868a202017-07-06 15:39:42 -07004842 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4843 huge_page_size(h));
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004844 if (pte)
4845 ptl = huge_pte_lock(h, mm, pte);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004846 absent = !pte || huge_pte_none(huge_ptep_get(pte));
Adam Litke4c887262005-10-29 18:16:46 -07004847
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004848 /*
4849 * When coredumping, it suits get_dump_page if we just return
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004850 * an error where there's an empty slot with no huge pagecache
4851 * to back it. This way, we avoid allocating a hugepage, and
4852 * the sparse dumpfile avoids allocating disk blocks, but its
4853 * huge holes still show up with zeroes where they need to be.
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004854 */
Hugh Dickins3ae77f42009-09-21 17:03:33 -07004855 if (absent && (flags & FOLL_DUMP) &&
4856 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004857 if (pte)
4858 spin_unlock(ptl);
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004859 remainder = 0;
4860 break;
4861 }
4862
Naoya Horiguchi9cc3a5b2013-04-17 15:58:30 -07004863 /*
4864 * We need call hugetlb_fault for both hugepages under migration
4865 * (in which case hugetlb_fault waits for the migration,) and
4866 * hwpoisoned hugepages (in which case we need to prevent the
4867 * caller from accessing to them.) In order to do this, we use
4868 * here is_swap_pte instead of is_hugetlb_entry_migration and
4869 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4870 * both cases, and because we can't follow correct pages
4871 * directly from any kind of swap entries.
4872 */
4873 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
Gerald Schaefer106c9922013-04-29 15:07:23 -07004874 ((flags & FOLL_WRITE) &&
4875 !huge_pte_write(huge_ptep_get(pte)))) {
Souptick Joarder2b740302018-08-23 17:01:36 -07004876 vm_fault_t ret;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004877 unsigned int fault_flags = 0;
Adam Litke4c887262005-10-29 18:16:46 -07004878
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004879 if (pte)
4880 spin_unlock(ptl);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004881 if (flags & FOLL_WRITE)
4882 fault_flags |= FAULT_FLAG_WRITE;
Peter Xu4f6da932020-04-01 21:07:58 -07004883 if (locked)
Peter Xu71335f32020-04-01 21:08:53 -07004884 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4885 FAULT_FLAG_KILLABLE;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004886 if (flags & FOLL_NOWAIT)
4887 fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4888 FAULT_FLAG_RETRY_NOWAIT;
4889 if (flags & FOLL_TRIED) {
Peter Xu4426e942020-04-01 21:08:49 -07004890 /*
4891 * Note: FAULT_FLAG_ALLOW_RETRY and
4892 * FAULT_FLAG_TRIED can co-exist
4893 */
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004894 fault_flags |= FAULT_FLAG_TRIED;
4895 }
4896 ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4897 if (ret & VM_FAULT_ERROR) {
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004898 err = vm_fault_to_errno(ret, flags);
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004899 remainder = 0;
4900 break;
4901 }
4902 if (ret & VM_FAULT_RETRY) {
Peter Xu4f6da932020-04-01 21:07:58 -07004903 if (locked &&
Andrea Arcangeli1ac25012019-02-01 14:20:16 -08004904 !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
Peter Xu4f6da932020-04-01 21:07:58 -07004905 *locked = 0;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004906 *nr_pages = 0;
4907 /*
4908 * VM_FAULT_RETRY must not return an
4909 * error, it will return zero
4910 * instead.
4911 *
4912 * No need to update "position" as the
4913 * caller will not check it after
4914 * *nr_pages is set to 0.
4915 */
4916 return i;
4917 }
4918 continue;
Adam Litke4c887262005-10-29 18:16:46 -07004919 }
David Gibson63551ae2005-06-21 17:14:44 -07004920
Andi Kleena5516432008-07-23 21:27:41 -07004921 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
Gerald Schaefer7f2e9522008-04-28 02:13:29 -07004922 page = pte_page(huge_ptep_get(pte));
Linus Torvalds8fde12c2019-04-11 10:49:19 -07004923
4924 /*
Zhigang Luacbfb082019-11-30 17:57:06 -08004925 * If subpage information not requested, update counters
4926 * and skip the same_page loop below.
4927 */
4928 if (!pages && !vmas && !pfn_offset &&
4929 (vaddr + huge_page_size(h) < vma->vm_end) &&
4930 (remainder >= pages_per_huge_page(h))) {
4931 vaddr += huge_page_size(h);
4932 remainder -= pages_per_huge_page(h);
4933 i += pages_per_huge_page(h);
4934 spin_unlock(ptl);
4935 continue;
4936 }
4937
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004938same_page:
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004939 if (pages) {
Hugh Dickins2a15efc2009-09-21 17:03:27 -07004940 pages[i] = mem_map_offset(page, pfn_offset);
John Hubbard3faa52c2020-04-01 21:05:29 -07004941 /*
4942 * try_grab_page() should always succeed here, because:
4943 * a) we hold the ptl lock, and b) we've just checked
4944 * that the huge page is present in the page tables. If
4945 * the huge page is present, then the tail pages must
4946 * also be present. The ptl prevents the head page and
4947 * tail pages from being rearranged in any way. So this
4948 * page must be available at this point, unless the page
4949 * refcount overflowed:
4950 */
4951 if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) {
4952 spin_unlock(ptl);
4953 remainder = 0;
4954 err = -ENOMEM;
4955 break;
4956 }
Chen, Kenneth Wd6692182006-03-31 02:29:57 -08004957 }
David Gibson63551ae2005-06-21 17:14:44 -07004958
4959 if (vmas)
4960 vmas[i] = vma;
4961
4962 vaddr += PAGE_SIZE;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004963 ++pfn_offset;
David Gibson63551ae2005-06-21 17:14:44 -07004964 --remainder;
4965 ++i;
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004966 if (vaddr < vma->vm_end && remainder &&
Andi Kleena5516432008-07-23 21:27:41 -07004967 pfn_offset < pages_per_huge_page(h)) {
Chen, Kenneth Wd5d4b0a2006-03-22 00:09:03 -08004968 /*
4969 * We use pfn_offset to avoid touching the pageframes
4970 * of this compound page.
4971 */
4972 goto same_page;
4973 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08004974 spin_unlock(ptl);
David Gibson63551ae2005-06-21 17:14:44 -07004975 }
Michel Lespinasse28a35712013-02-22 16:35:55 -08004976 *nr_pages = remainder;
Andrea Arcangeli87ffc112017-02-22 15:43:13 -08004977 /*
4978 * setting position is actually required only if remainder is
4979 * not zero but it's faster not to add a "if (remainder)"
4980 * branch.
4981 */
David Gibson63551ae2005-06-21 17:14:44 -07004982 *position = vaddr;
4983
Daniel Jordan2be7cfe2017-08-02 13:31:47 -07004984 return i ? i : err;
David Gibson63551ae2005-06-21 17:14:44 -07004985}
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004986
Aneesh Kumar K.V5491ae72016-07-13 15:06:43 +05304987#ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4988/*
4989 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4990 * implement this.
4991 */
4992#define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4993#endif
4994
Peter Zijlstra7da4d642012-11-19 03:14:23 +01004995unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
Zhang, Yanmin8f860592006-03-22 00:08:50 -08004996 unsigned long address, unsigned long end, pgprot_t newprot)
4997{
4998 struct mm_struct *mm = vma->vm_mm;
4999 unsigned long start = address;
5000 pte_t *ptep;
5001 pte_t pte;
Andi Kleena5516432008-07-23 21:27:41 -07005002 struct hstate *h = hstate_vma(vma);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005003 unsigned long pages = 0;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005004 bool shared_pmd = false;
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005005 struct mmu_notifier_range range;
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005006
5007 /*
5008 * In the case of shared PMDs, the area to flush could be beyond
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005009 * start/end. Set range.start/range.end to cover the maximum possible
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005010 * range if PMD sharing is possible.
5011 */
Jérôme Glisse7269f992019-05-13 17:20:53 -07005012 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
5013 0, vma, mm, start, end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005014 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005015
5016 BUG_ON(address >= end);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005017 flush_cache_range(vma, range.start, range.end);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005018
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005019 mmu_notifier_invalidate_range_start(&range);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005020 i_mmap_lock_write(vma->vm_file->f_mapping);
Andi Kleena5516432008-07-23 21:27:41 -07005021 for (; address < end; address += huge_page_size(h)) {
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005022 spinlock_t *ptl;
Punit Agrawal7868a202017-07-06 15:39:42 -07005023 ptep = huge_pte_offset(mm, address, huge_page_size(h));
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005024 if (!ptep)
5025 continue;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005026 ptl = huge_pte_lock(h, mm, ptep);
Mike Kravetz34ae2042020-08-11 18:31:38 -07005027 if (huge_pmd_unshare(mm, vma, &address, ptep)) {
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005028 pages++;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005029 spin_unlock(ptl);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005030 shared_pmd = true;
Chen, Kenneth W39dde652006-12-06 20:32:03 -08005031 continue;
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005032 }
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005033 pte = huge_ptep_get(ptep);
5034 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
5035 spin_unlock(ptl);
5036 continue;
5037 }
5038 if (unlikely(is_hugetlb_entry_migration(pte))) {
5039 swp_entry_t entry = pte_to_swp_entry(pte);
5040
5041 if (is_write_migration_entry(entry)) {
5042 pte_t newpte;
5043
5044 make_migration_entry_read(&entry);
5045 newpte = swp_entry_to_pte(entry);
Punit Agrawale5251fd2017-07-06 15:39:50 -07005046 set_huge_swap_pte_at(mm, address, ptep,
5047 newpte, huge_page_size(h));
Naoya Horiguchia8bda282015-02-11 15:25:28 -08005048 pages++;
5049 }
5050 spin_unlock(ptl);
5051 continue;
5052 }
5053 if (!huge_pte_none(pte)) {
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005054 pte_t old_pte;
5055
5056 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
5057 pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
Tony Lube7517d2013-02-04 14:28:46 -08005058 pte = arch_make_huge_pte(pte, vma, NULL, 0);
Aneesh Kumar K.V023bdd02019-03-05 15:46:37 -08005059 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005060 pages++;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005061 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005062 spin_unlock(ptl);
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005063 }
Mel Gormand8333522012-07-31 16:46:20 -07005064 /*
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005065 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
Mel Gormand8333522012-07-31 16:46:20 -07005066 * may have cleared our pud entry and done put_page on the page table:
Davidlohr Buesoc8c06ef2014-12-12 16:54:24 -08005067 * once we release i_mmap_rwsem, another task can do the final put_page
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005068 * and that page table be reused and filled with junk. If we actually
5069 * did unshare a page of pmds, flush the range corresponding to the pud.
Mel Gormand8333522012-07-31 16:46:20 -07005070 */
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005071 if (shared_pmd)
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005072 flush_hugetlb_tlb_range(vma, range.start, range.end);
Mike Kravetzdff11ab2018-10-05 15:51:33 -07005073 else
5074 flush_hugetlb_tlb_range(vma, start, end);
Jérôme Glisse0f108512017-11-15 17:34:07 -08005075 /*
5076 * No need to call mmu_notifier_invalidate_range() we are downgrading
5077 * page table protection not changing it to point to a new page.
5078 *
Mike Rapoportad56b732018-03-21 21:22:47 +02005079 * See Documentation/vm/mmu_notifier.rst
Jérôme Glisse0f108512017-11-15 17:34:07 -08005080 */
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08005081 i_mmap_unlock_write(vma->vm_file->f_mapping);
Jérôme Glisseac46d4f2018-12-28 00:38:09 -08005082 mmu_notifier_invalidate_range_end(&range);
Peter Zijlstra7da4d642012-11-19 03:14:23 +01005083
5084 return pages << h->order;
Zhang, Yanmin8f860592006-03-22 00:08:50 -08005085}
5086
Mel Gormana1e78772008-07-23 21:27:23 -07005087int hugetlb_reserve_pages(struct inode *inode,
5088 long from, long to,
Mel Gorman5a6fe122009-02-10 14:02:27 +00005089 struct vm_area_struct *vma,
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005090 vm_flags_t vm_flags)
Adam Litkee4e574b2007-10-16 01:26:19 -07005091{
Mina Almasry0db9d742020-04-01 21:11:25 -07005092 long ret, chg, add = -1;
Andi Kleena5516432008-07-23 21:27:41 -07005093 struct hstate *h = hstate_inode(inode);
David Gibson90481622012-03-21 16:34:12 -07005094 struct hugepage_subpool *spool = subpool_inode(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005095 struct resv_map *resv_map;
Mina Almasry075a61d2020-04-01 21:11:28 -07005096 struct hugetlb_cgroup *h_cg = NULL;
Mina Almasry0db9d742020-04-01 21:11:25 -07005097 long gbl_reserve, regions_needed = 0;
Adam Litkee4e574b2007-10-16 01:26:19 -07005098
Mike Kravetz63489f82018-03-22 16:17:13 -07005099 /* This should never happen */
5100 if (from > to) {
5101 VM_WARN(1, "%s called with a negative range\n", __func__);
5102 return -EINVAL;
5103 }
5104
Mel Gormana1e78772008-07-23 21:27:23 -07005105 /*
Mel Gorman17c9d122009-02-11 16:34:16 +00005106 * Only apply hugepage reservation if asked. At fault time, an
5107 * attempt will be made for VM_NORESERVE to allocate a page
David Gibson90481622012-03-21 16:34:12 -07005108 * without using reserves
Mel Gorman17c9d122009-02-11 16:34:16 +00005109 */
KOSAKI Motohiroca16d142011-05-26 19:16:19 +09005110 if (vm_flags & VM_NORESERVE)
Mel Gorman17c9d122009-02-11 16:34:16 +00005111 return 0;
5112
5113 /*
Mel Gormana1e78772008-07-23 21:27:23 -07005114 * Shared mappings base their reservation on the number of pages that
5115 * are already allocated on behalf of the file. Private mappings need
5116 * to reserve the full area even if read-only as mprotect() may be
5117 * called to make the mapping read-write. Assume !vma is a shm mapping
5118 */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005119 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mike Kravetzf27a5132019-05-13 17:22:55 -07005120 /*
5121 * resv_map can not be NULL as hugetlb_reserve_pages is only
5122 * called for inodes for which resv_maps were created (see
5123 * hugetlbfs_get_inode).
5124 */
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005125 resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005126
Mina Almasry0db9d742020-04-01 21:11:25 -07005127 chg = region_chg(resv_map, from, to, &regions_needed);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005128
5129 } else {
Mina Almasrye9fe92a2020-04-01 21:11:21 -07005130 /* Private mapping. */
Joonsoo Kim9119a412014-04-03 14:47:25 -07005131 resv_map = resv_map_alloc();
Mel Gorman5a6fe122009-02-10 14:02:27 +00005132 if (!resv_map)
5133 return -ENOMEM;
5134
Mel Gorman17c9d122009-02-11 16:34:16 +00005135 chg = to - from;
5136
Mel Gorman5a6fe122009-02-10 14:02:27 +00005137 set_vma_resv_map(vma, resv_map);
5138 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
5139 }
5140
Dave Hansenc50ac052012-05-29 15:06:46 -07005141 if (chg < 0) {
5142 ret = chg;
5143 goto out_err;
5144 }
Mel Gorman17c9d122009-02-11 16:34:16 +00005145
Mina Almasry075a61d2020-04-01 21:11:28 -07005146 ret = hugetlb_cgroup_charge_cgroup_rsvd(
5147 hstate_index(h), chg * pages_per_huge_page(h), &h_cg);
5148
5149 if (ret < 0) {
5150 ret = -ENOMEM;
5151 goto out_err;
5152 }
5153
5154 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
5155 /* For private mappings, the hugetlb_cgroup uncharge info hangs
5156 * of the resv_map.
5157 */
5158 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
5159 }
5160
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005161 /*
5162 * There must be enough pages in the subpool for the mapping. If
5163 * the subpool has a minimum size, there may be some global
5164 * reservations already in place (gbl_reserve).
5165 */
5166 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
5167 if (gbl_reserve < 0) {
Dave Hansenc50ac052012-05-29 15:06:46 -07005168 ret = -ENOSPC;
Mina Almasry075a61d2020-04-01 21:11:28 -07005169 goto out_uncharge_cgroup;
Dave Hansenc50ac052012-05-29 15:06:46 -07005170 }
Mel Gorman17c9d122009-02-11 16:34:16 +00005171
5172 /*
5173 * Check enough hugepages are available for the reservation.
David Gibson90481622012-03-21 16:34:12 -07005174 * Hand the pages back to the subpool if there are not
Mel Gorman17c9d122009-02-11 16:34:16 +00005175 */
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005176 ret = hugetlb_acct_memory(h, gbl_reserve);
Mel Gorman17c9d122009-02-11 16:34:16 +00005177 if (ret < 0) {
Mina Almasry075a61d2020-04-01 21:11:28 -07005178 goto out_put_pages;
Mel Gorman17c9d122009-02-11 16:34:16 +00005179 }
5180
5181 /*
5182 * Account for the reservations made. Shared mappings record regions
5183 * that have reservations as they are shared by multiple VMAs.
5184 * When the last VMA disappears, the region map says how much
5185 * the reservation was and the page cache tells how much of
5186 * the reservation was consumed. Private mappings are per-VMA and
5187 * only the consumed reservations are tracked. When the VMA
5188 * disappears, the original reservation is the VMA size and the
5189 * consumed reservations are stored in the map. Hence, nothing
5190 * else has to be done for private mappings here
5191 */
Mike Kravetz33039672015-06-24 16:57:58 -07005192 if (!vma || vma->vm_flags & VM_MAYSHARE) {
Mina Almasry075a61d2020-04-01 21:11:28 -07005193 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
Mike Kravetz33039672015-06-24 16:57:58 -07005194
Mina Almasry0db9d742020-04-01 21:11:25 -07005195 if (unlikely(add < 0)) {
5196 hugetlb_acct_memory(h, -gbl_reserve);
Mina Almasry075a61d2020-04-01 21:11:28 -07005197 goto out_put_pages;
Mina Almasry0db9d742020-04-01 21:11:25 -07005198 } else if (unlikely(chg > add)) {
Mike Kravetz33039672015-06-24 16:57:58 -07005199 /*
5200 * pages in this range were added to the reserve
5201 * map between region_chg and region_add. This
5202 * indicates a race with alloc_huge_page. Adjust
5203 * the subpool and reserve counts modified above
5204 * based on the difference.
5205 */
5206 long rsv_adjust;
5207
Mina Almasry075a61d2020-04-01 21:11:28 -07005208 hugetlb_cgroup_uncharge_cgroup_rsvd(
5209 hstate_index(h),
5210 (chg - add) * pages_per_huge_page(h), h_cg);
5211
Mike Kravetz33039672015-06-24 16:57:58 -07005212 rsv_adjust = hugepage_subpool_put_pages(spool,
5213 chg - add);
5214 hugetlb_acct_memory(h, -rsv_adjust);
5215 }
5216 }
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005217 return 0;
Mina Almasry075a61d2020-04-01 21:11:28 -07005218out_put_pages:
5219 /* put back original number of pages, chg */
5220 (void)hugepage_subpool_put_pages(spool, chg);
5221out_uncharge_cgroup:
5222 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
5223 chg * pages_per_huge_page(h), h_cg);
Dave Hansenc50ac052012-05-29 15:06:46 -07005224out_err:
Mike Kravetz5e911372015-09-08 15:01:28 -07005225 if (!vma || vma->vm_flags & VM_MAYSHARE)
Mina Almasry0db9d742020-04-01 21:11:25 -07005226 /* Only call region_abort if the region_chg succeeded but the
5227 * region_add failed or didn't run.
5228 */
5229 if (chg >= 0 && add < 0)
5230 region_abort(resv_map, from, to, regions_needed);
Joonsoo Kimf031dd22014-04-03 14:47:28 -07005231 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
5232 kref_put(&resv_map->refs, resv_map_release);
Dave Hansenc50ac052012-05-29 15:06:46 -07005233 return ret;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005234}
5235
Mike Kravetzb5cec282015-09-08 15:01:41 -07005236long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
5237 long freed)
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005238{
Andi Kleena5516432008-07-23 21:27:41 -07005239 struct hstate *h = hstate_inode(inode);
Joonsoo Kim4e35f482014-04-03 14:47:30 -07005240 struct resv_map *resv_map = inode_resv_map(inode);
Joonsoo Kim9119a412014-04-03 14:47:25 -07005241 long chg = 0;
David Gibson90481622012-03-21 16:34:12 -07005242 struct hugepage_subpool *spool = subpool_inode(inode);
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005243 long gbl_reserve;
Ken Chen45c682a2007-11-14 16:59:44 -08005244
Mike Kravetzf27a5132019-05-13 17:22:55 -07005245 /*
5246 * Since this routine can be called in the evict inode path for all
5247 * hugetlbfs inodes, resv_map could be NULL.
5248 */
Mike Kravetzb5cec282015-09-08 15:01:41 -07005249 if (resv_map) {
5250 chg = region_del(resv_map, start, end);
5251 /*
5252 * region_del() can fail in the rare case where a region
5253 * must be split and another region descriptor can not be
5254 * allocated. If end == LONG_MAX, it will not fail.
5255 */
5256 if (chg < 0)
5257 return chg;
5258 }
5259
Ken Chen45c682a2007-11-14 16:59:44 -08005260 spin_lock(&inode->i_lock);
Eric Sandeene4c6f8b2009-07-29 15:02:16 -07005261 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
Ken Chen45c682a2007-11-14 16:59:44 -08005262 spin_unlock(&inode->i_lock);
5263
Mike Kravetz1c5ecae2015-04-15 16:13:39 -07005264 /*
5265 * If the subpool has a minimum size, the number of global
5266 * reservations to be released may be adjusted.
5267 */
5268 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
5269 hugetlb_acct_memory(h, -gbl_reserve);
Mike Kravetzb5cec282015-09-08 15:01:41 -07005270
5271 return 0;
Chen, Kenneth Wa43a8c32006-06-23 02:03:15 -07005272}
Naoya Horiguchi93f70f92010-05-28 09:29:20 +09005273
Steve Capper3212b532013-04-23 12:35:02 +01005274#ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
5275static unsigned long page_table_shareable(struct vm_area_struct *svma,
5276 struct vm_area_struct *vma,
5277 unsigned long addr, pgoff_t idx)
5278{
5279 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
5280 svma->vm_start;
5281 unsigned long sbase = saddr & PUD_MASK;
5282 unsigned long s_end = sbase + PUD_SIZE;
5283
5284 /* Allow segments to share if only one is marked locked */
Eric B Munsonde60f5f2015-11-05 18:51:36 -08005285 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
5286 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
Steve Capper3212b532013-04-23 12:35:02 +01005287
5288 /*
5289 * match the virtual addresses, permission and the alignment of the
5290 * page table page.
5291 */
5292 if (pmd_index(addr) != pmd_index(saddr) ||
5293 vm_flags != svm_flags ||
5294 sbase < svma->vm_start || svma->vm_end < s_end)
5295 return 0;
5296
5297 return saddr;
5298}
5299
Nicholas Krause31aafb42015-09-04 15:47:58 -07005300static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
Steve Capper3212b532013-04-23 12:35:02 +01005301{
5302 unsigned long base = addr & PUD_MASK;
5303 unsigned long end = base + PUD_SIZE;
5304
5305 /*
5306 * check on proper vm_flags and page table alignment
5307 */
Mike Kravetz017b1662018-10-05 15:51:29 -07005308 if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
Nicholas Krause31aafb42015-09-04 15:47:58 -07005309 return true;
5310 return false;
Steve Capper3212b532013-04-23 12:35:02 +01005311}
5312
5313/*
Mike Kravetz017b1662018-10-05 15:51:29 -07005314 * Determine if start,end range within vma could be mapped by shared pmd.
5315 * If yes, adjust start and end to cover range associated with possible
5316 * shared pmd mappings.
5317 */
5318void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5319 unsigned long *start, unsigned long *end)
5320{
Peter Xu75802ca62020-08-06 23:26:11 -07005321 unsigned long a_start, a_end;
Mike Kravetz017b1662018-10-05 15:51:29 -07005322
5323 if (!(vma->vm_flags & VM_MAYSHARE))
5324 return;
5325
Peter Xu75802ca62020-08-06 23:26:11 -07005326 /* Extend the range to be PUD aligned for a worst case scenario */
5327 a_start = ALIGN_DOWN(*start, PUD_SIZE);
5328 a_end = ALIGN(*end, PUD_SIZE);
Mike Kravetz017b1662018-10-05 15:51:29 -07005329
Peter Xu75802ca62020-08-06 23:26:11 -07005330 /*
5331 * Intersect the range with the vma range, since pmd sharing won't be
5332 * across vma after all
5333 */
5334 *start = max(vma->vm_start, a_start);
5335 *end = min(vma->vm_end, a_end);
Mike Kravetz017b1662018-10-05 15:51:29 -07005336}
5337
5338/*
Steve Capper3212b532013-04-23 12:35:02 +01005339 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
5340 * and returns the corresponding pte. While this is not necessary for the
5341 * !shared pmd case because we can allocate the pmd later as well, it makes the
Mike Kravetzc0d03812020-04-01 21:11:05 -07005342 * code much cleaner.
5343 *
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005344 * This routine must be called with i_mmap_rwsem held in at least read mode if
5345 * sharing is possible. For hugetlbfs, this prevents removal of any page
5346 * table entries associated with the address space. This is important as we
5347 * are setting up sharing based on existing page table entries (mappings).
5348 *
5349 * NOTE: This routine is only called from huge_pte_alloc. Some callers of
5350 * huge_pte_alloc know that sharing is not possible and do not take
5351 * i_mmap_rwsem as a performance optimization. This is handled by the
5352 * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is
5353 * only required for subsequent processing.
Steve Capper3212b532013-04-23 12:35:02 +01005354 */
5355pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5356{
5357 struct vm_area_struct *vma = find_vma(mm, addr);
5358 struct address_space *mapping = vma->vm_file->f_mapping;
5359 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
5360 vma->vm_pgoff;
5361 struct vm_area_struct *svma;
5362 unsigned long saddr;
5363 pte_t *spte = NULL;
5364 pte_t *pte;
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005365 spinlock_t *ptl;
Steve Capper3212b532013-04-23 12:35:02 +01005366
5367 if (!vma_shareable(vma, addr))
5368 return (pte_t *)pmd_alloc(mm, pud, addr);
5369
Mike Kravetz0bf7b642020-10-13 16:56:42 -07005370 i_mmap_assert_locked(mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005371 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
5372 if (svma == vma)
5373 continue;
5374
5375 saddr = page_table_shareable(svma, vma, addr, idx);
5376 if (saddr) {
Punit Agrawal7868a202017-07-06 15:39:42 -07005377 spte = huge_pte_offset(svma->vm_mm, saddr,
5378 vma_mmu_pagesize(svma));
Steve Capper3212b532013-04-23 12:35:02 +01005379 if (spte) {
5380 get_page(virt_to_page(spte));
5381 break;
5382 }
5383 }
5384 }
5385
5386 if (!spte)
5387 goto out;
5388
Aneesh Kumar K.V8bea8052016-12-12 16:41:59 -08005389 ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005390 if (pud_none(*pud)) {
Steve Capper3212b532013-04-23 12:35:02 +01005391 pud_populate(mm, pud,
5392 (pmd_t *)((unsigned long)spte & PAGE_MASK));
Kirill A. Shutemovc17b1f42016-06-24 14:49:51 -07005393 mm_inc_nr_pmds(mm);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005394 } else {
Steve Capper3212b532013-04-23 12:35:02 +01005395 put_page(virt_to_page(spte));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005396 }
Kirill A. Shutemovcb900f42013-11-14 14:31:02 -08005397 spin_unlock(ptl);
Steve Capper3212b532013-04-23 12:35:02 +01005398out:
5399 pte = (pte_t *)pmd_alloc(mm, pud, addr);
Steve Capper3212b532013-04-23 12:35:02 +01005400 return pte;
5401}
5402
5403/*
5404 * unmap huge page backed by shared pte.
5405 *
5406 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
5407 * indicated by page_count > 1, unmap is achieved by clearing pud and
5408 * decrementing the ref count. If count == 1, the pte page is not shared.
5409 *
Mike Kravetzc0d03812020-04-01 21:11:05 -07005410 * Called with page table lock held and i_mmap_rwsem held in write mode.
Steve Capper3212b532013-04-23 12:35:02 +01005411 *
5412 * returns: 1 successfully unmapped a shared pte page
5413 * 0 the underlying pte page is not shared, or it is the last user
5414 */
Mike Kravetz34ae2042020-08-11 18:31:38 -07005415int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5416 unsigned long *addr, pte_t *ptep)
Steve Capper3212b532013-04-23 12:35:02 +01005417{
5418 pgd_t *pgd = pgd_offset(mm, *addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005419 p4d_t *p4d = p4d_offset(pgd, *addr);
5420 pud_t *pud = pud_offset(p4d, *addr);
Steve Capper3212b532013-04-23 12:35:02 +01005421
Mike Kravetz34ae2042020-08-11 18:31:38 -07005422 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
Steve Capper3212b532013-04-23 12:35:02 +01005423 BUG_ON(page_count(virt_to_page(ptep)) == 0);
5424 if (page_count(virt_to_page(ptep)) == 1)
5425 return 0;
5426
5427 pud_clear(pud);
5428 put_page(virt_to_page(ptep));
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08005429 mm_dec_nr_pmds(mm);
Steve Capper3212b532013-04-23 12:35:02 +01005430 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
5431 return 1;
5432}
Steve Capper9e5fc742013-04-30 08:02:03 +01005433#define want_pmd_share() (1)
5434#else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5435pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
5436{
5437 return NULL;
5438}
Zhang Zhene81f2d22015-06-24 16:56:13 -07005439
Mike Kravetz34ae2042020-08-11 18:31:38 -07005440int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
5441 unsigned long *addr, pte_t *ptep)
Zhang Zhene81f2d22015-06-24 16:56:13 -07005442{
5443 return 0;
5444}
Mike Kravetz017b1662018-10-05 15:51:29 -07005445
5446void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
5447 unsigned long *start, unsigned long *end)
5448{
5449}
Steve Capper9e5fc742013-04-30 08:02:03 +01005450#define want_pmd_share() (0)
Steve Capper3212b532013-04-23 12:35:02 +01005451#endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
5452
Steve Capper9e5fc742013-04-30 08:02:03 +01005453#ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
5454pte_t *huge_pte_alloc(struct mm_struct *mm,
5455 unsigned long addr, unsigned long sz)
5456{
5457 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005458 p4d_t *p4d;
Steve Capper9e5fc742013-04-30 08:02:03 +01005459 pud_t *pud;
5460 pte_t *pte = NULL;
5461
5462 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovf4f0a3d2017-11-29 16:11:30 -08005463 p4d = p4d_alloc(mm, pgd, addr);
5464 if (!p4d)
5465 return NULL;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005466 pud = pud_alloc(mm, p4d, addr);
Steve Capper9e5fc742013-04-30 08:02:03 +01005467 if (pud) {
5468 if (sz == PUD_SIZE) {
5469 pte = (pte_t *)pud;
5470 } else {
5471 BUG_ON(sz != PMD_SIZE);
5472 if (want_pmd_share() && pud_none(*pud))
5473 pte = huge_pmd_share(mm, addr, pud);
5474 else
5475 pte = (pte_t *)pmd_alloc(mm, pud, addr);
5476 }
5477 }
Michal Hocko4e666312016-08-02 14:02:34 -07005478 BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
Steve Capper9e5fc742013-04-30 08:02:03 +01005479
5480 return pte;
5481}
5482
Punit Agrawal9b19df22017-09-06 16:21:01 -07005483/*
5484 * huge_pte_offset() - Walk the page table to resolve the hugepage
5485 * entry at address @addr
5486 *
Li Xinhai8ac0b812020-06-03 16:00:53 -07005487 * Return: Pointer to page table entry (PUD or PMD) for
5488 * address @addr, or NULL if a !p*d_present() entry is encountered and the
Punit Agrawal9b19df22017-09-06 16:21:01 -07005489 * size @sz doesn't match the hugepage size at this level of the page
5490 * table.
5491 */
Punit Agrawal7868a202017-07-06 15:39:42 -07005492pte_t *huge_pte_offset(struct mm_struct *mm,
5493 unsigned long addr, unsigned long sz)
Steve Capper9e5fc742013-04-30 08:02:03 +01005494{
5495 pgd_t *pgd;
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005496 p4d_t *p4d;
Li Xinhai8ac0b812020-06-03 16:00:53 -07005497 pud_t *pud;
5498 pmd_t *pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01005499
5500 pgd = pgd_offset(mm, addr);
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005501 if (!pgd_present(*pgd))
5502 return NULL;
5503 p4d = p4d_offset(pgd, addr);
5504 if (!p4d_present(*p4d))
5505 return NULL;
Punit Agrawal9b19df22017-09-06 16:21:01 -07005506
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005507 pud = pud_offset(p4d, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07005508 if (sz == PUD_SIZE)
5509 /* must be pud huge, non-present or none */
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005510 return (pte_t *)pud;
Li Xinhai8ac0b812020-06-03 16:00:53 -07005511 if (!pud_present(*pud))
5512 return NULL;
5513 /* must have a valid entry and size to go further */
Punit Agrawal9b19df22017-09-06 16:21:01 -07005514
Kirill A. Shutemovc2febaf2017-03-09 17:24:07 +03005515 pmd = pmd_offset(pud, addr);
Li Xinhai8ac0b812020-06-03 16:00:53 -07005516 /* must be pmd huge, non-present or none */
5517 return (pte_t *)pmd;
Steve Capper9e5fc742013-04-30 08:02:03 +01005518}
5519
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08005520#endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
5521
5522/*
5523 * These functions are overwritable if your architecture needs its own
5524 * behavior.
5525 */
5526struct page * __weak
5527follow_huge_addr(struct mm_struct *mm, unsigned long address,
5528 int write)
5529{
5530 return ERR_PTR(-EINVAL);
5531}
5532
5533struct page * __weak
Aneesh Kumar K.V4dc71452017-07-06 15:38:56 -07005534follow_huge_pd(struct vm_area_struct *vma,
5535 unsigned long address, hugepd_t hpd, int flags, int pdshift)
5536{
5537 WARN(1, "hugepd follow called with no support for hugepage directory format\n");
5538 return NULL;
5539}
5540
5541struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01005542follow_huge_pmd(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005543 pmd_t *pmd, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01005544{
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005545 struct page *page = NULL;
5546 spinlock_t *ptl;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005547 pte_t pte;
John Hubbard3faa52c2020-04-01 21:05:29 -07005548
5549 /* FOLL_GET and FOLL_PIN are mutually exclusive. */
5550 if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) ==
5551 (FOLL_PIN | FOLL_GET)))
5552 return NULL;
5553
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005554retry:
5555 ptl = pmd_lockptr(mm, pmd);
5556 spin_lock(ptl);
5557 /*
5558 * make sure that the address range covered by this pmd is not
5559 * unmapped from other threads.
5560 */
5561 if (!pmd_huge(*pmd))
5562 goto out;
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005563 pte = huge_ptep_get((pte_t *)pmd);
5564 if (pte_present(pte)) {
Gerald Schaefer97534122015-04-14 15:42:30 -07005565 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
John Hubbard3faa52c2020-04-01 21:05:29 -07005566 /*
5567 * try_grab_page() should always succeed here, because: a) we
5568 * hold the pmd (ptl) lock, and b) we've just checked that the
5569 * huge pmd (head) page is present in the page tables. The ptl
5570 * prevents the head page and tail pages from being rearranged
5571 * in any way. So this page must be available at this point,
5572 * unless the page refcount overflowed:
5573 */
5574 if (WARN_ON_ONCE(!try_grab_page(page, flags))) {
5575 page = NULL;
5576 goto out;
5577 }
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005578 } else {
Naoya Horiguchic9d398f2017-03-31 15:11:55 -07005579 if (is_hugetlb_entry_migration(pte)) {
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005580 spin_unlock(ptl);
5581 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
5582 goto retry;
5583 }
5584 /*
5585 * hwpoisoned entry is treated as no_page_table in
5586 * follow_page_mask().
5587 */
5588 }
5589out:
5590 spin_unlock(ptl);
Steve Capper9e5fc742013-04-30 08:02:03 +01005591 return page;
5592}
5593
Naoya Horiguchi61f77ed2015-02-11 15:25:15 -08005594struct page * __weak
Steve Capper9e5fc742013-04-30 08:02:03 +01005595follow_huge_pud(struct mm_struct *mm, unsigned long address,
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005596 pud_t *pud, int flags)
Steve Capper9e5fc742013-04-30 08:02:03 +01005597{
John Hubbard3faa52c2020-04-01 21:05:29 -07005598 if (flags & (FOLL_GET | FOLL_PIN))
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005599 return NULL;
Steve Capper9e5fc742013-04-30 08:02:03 +01005600
Naoya Horiguchie66f17f2015-02-11 15:25:22 -08005601 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
Steve Capper9e5fc742013-04-30 08:02:03 +01005602}
5603
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07005604struct page * __weak
5605follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5606{
John Hubbard3faa52c2020-04-01 21:05:29 -07005607 if (flags & (FOLL_GET | FOLL_PIN))
Anshuman Khandualfaaa5b62017-07-06 15:38:50 -07005608 return NULL;
5609
5610 return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5611}
5612
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005613bool isolate_huge_page(struct page *page, struct list_head *list)
5614{
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005615 bool ret = true;
5616
Sasha Levin309381fea2014-01-23 15:52:54 -08005617 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005618 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005619 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5620 ret = false;
5621 goto unlock;
5622 }
5623 clear_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005624 list_move_tail(&page->lru, list);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005625unlock:
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005626 spin_unlock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005627 return ret;
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005628}
5629
5630void putback_active_hugepage(struct page *page)
5631{
Sasha Levin309381fea2014-01-23 15:52:54 -08005632 VM_BUG_ON_PAGE(!PageHead(page), page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005633 spin_lock(&hugetlb_lock);
Naoya Horiguchibcc54222015-04-15 16:14:38 -07005634 set_page_huge_active(page);
Naoya Horiguchi31caf662013-09-11 14:21:59 -07005635 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5636 spin_unlock(&hugetlb_lock);
5637 put_page(page);
5638}
Michal Hockoab5ac902018-01-31 16:20:48 -08005639
5640void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5641{
5642 struct hstate *h = page_hstate(oldpage);
5643
5644 hugetlb_cgroup_migrate(oldpage, newpage);
5645 set_page_owner_migrate_reason(newpage, reason);
5646
5647 /*
5648 * transfer temporary state of the new huge page. This is
5649 * reverse to other transitions because the newpage is going to
5650 * be final while the old one will be freed so it takes over
5651 * the temporary status.
5652 *
5653 * Also note that we have to transfer the per-node surplus state
5654 * here as well otherwise the global surplus count will not match
5655 * the per-node's.
5656 */
5657 if (PageHugeTemporary(newpage)) {
5658 int old_nid = page_to_nid(oldpage);
5659 int new_nid = page_to_nid(newpage);
5660
5661 SetPageHugeTemporary(oldpage);
5662 ClearPageHugeTemporary(newpage);
5663
5664 spin_lock(&hugetlb_lock);
5665 if (h->surplus_huge_pages_node[old_nid]) {
5666 h->surplus_huge_pages_node[old_nid]--;
5667 h->surplus_huge_pages_node[new_nid]++;
5668 }
5669 spin_unlock(&hugetlb_lock);
5670 }
5671}
Roman Gushchincf11e852020-04-10 14:32:45 -07005672
5673#ifdef CONFIG_CMA
Roman Gushchincf11e852020-04-10 14:32:45 -07005674static bool cma_reserve_called __initdata;
5675
5676static int __init cmdline_parse_hugetlb_cma(char *p)
5677{
5678 hugetlb_cma_size = memparse(p, &p);
5679 return 0;
5680}
5681
5682early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
5683
5684void __init hugetlb_cma_reserve(int order)
5685{
5686 unsigned long size, reserved, per_node;
5687 int nid;
5688
5689 cma_reserve_called = true;
5690
5691 if (!hugetlb_cma_size)
5692 return;
5693
5694 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
5695 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
5696 (PAGE_SIZE << order) / SZ_1M);
5697 return;
5698 }
5699
5700 /*
5701 * If 3 GB area is requested on a machine with 4 numa nodes,
5702 * let's allocate 1 GB on first three nodes and ignore the last one.
5703 */
5704 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
5705 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
5706 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
5707
5708 reserved = 0;
5709 for_each_node_state(nid, N_ONLINE) {
5710 int res;
Barry Song2281f792020-08-24 11:03:09 +12005711 char name[CMA_MAX_NAME];
Roman Gushchincf11e852020-04-10 14:32:45 -07005712
5713 size = min(per_node, hugetlb_cma_size - reserved);
5714 size = round_up(size, PAGE_SIZE << order);
5715
Barry Song2281f792020-08-24 11:03:09 +12005716 snprintf(name, sizeof(name), "hugetlb%d", nid);
Roman Gushchincf11e852020-04-10 14:32:45 -07005717 res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order,
Barry Song29d0f412020-08-11 18:32:00 -07005718 0, false, name,
Roman Gushchincf11e852020-04-10 14:32:45 -07005719 &hugetlb_cma[nid], nid);
5720 if (res) {
5721 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
5722 res, nid);
5723 continue;
5724 }
5725
5726 reserved += size;
5727 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
5728 size / SZ_1M, nid);
5729
5730 if (reserved >= hugetlb_cma_size)
5731 break;
5732 }
5733}
5734
5735void __init hugetlb_cma_check(void)
5736{
5737 if (!hugetlb_cma_size || cma_reserve_called)
5738 return;
5739
5740 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
5741}
5742
5743#endif /* CONFIG_CMA */