1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Generic hugetlb support.
4 * (C) Nadia Yvette Chambers, April 2004
5 */
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/mm.h>
9 #include <linux/seq_file.h>
10 #include <linux/sysctl.h>
11 #include <linux/highmem.h>
12 #include <linux/mmu_notifier.h>
13 #include <linux/nodemask.h>
14 #include <linux/pagemap.h>
15 #include <linux/mempolicy.h>
16 #include <linux/compiler.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/memblock.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/sched/mm.h>
23 #include <linux/mmdebug.h>
24 #include <linux/sched/signal.h>
25 #include <linux/rmap.h>
26 #include <linux/string_helpers.h>
27 #include <linux/swap.h>
28 #include <linux/swapops.h>
29 #include <linux/jhash.h>
30 #include <linux/numa.h>
31 #include <linux/llist.h>
32 #include <linux/cma.h>
33 #include <linux/migrate.h>
34 #include <linux/nospec.h>
35 #include <linux/delayacct.h>
36 #include <linux/memory.h>
37 #include <linux/mm_inline.h>
38
39 #include <asm/page.h>
40 #include <asm/pgalloc.h>
41 #include <asm/tlb.h>
42
43 #include <linux/io.h>
44 #include <linux/hugetlb.h>
45 #include <linux/hugetlb_cgroup.h>
46 #include <linux/node.h>
47 #include <linux/page_owner.h>
48 #include "internal.h"
49 #include "hugetlb_vmemmap.h"
50
51 int hugetlb_max_hstate __read_mostly;
52 unsigned int default_hstate_idx;
53 struct hstate hstates[HUGE_MAX_HSTATE];
54
55 #ifdef CONFIG_CMA
56 static struct cma *hugetlb_cma[MAX_NUMNODES];
57 static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata;
hugetlb_cma_folio(struct folio * folio,unsigned int order)58 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
59 {
60 return cma_pages_valid(hugetlb_cma[folio_nid(folio)], &folio->page,
61 1 << order);
62 }
63 #else
hugetlb_cma_folio(struct folio * folio,unsigned int order)64 static bool hugetlb_cma_folio(struct folio *folio, unsigned int order)
65 {
66 return false;
67 }
68 #endif
69 static unsigned long hugetlb_cma_size __initdata;
70
71 __initdata LIST_HEAD(huge_boot_pages);
72
73 /* for command line parsing */
74 static struct hstate * __initdata parsed_hstate;
75 static unsigned long __initdata default_hstate_max_huge_pages;
76 static bool __initdata parsed_valid_hugepagesz = true;
77 static bool __initdata parsed_default_hugepagesz;
78 static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata;
79
80 /*
81 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
82 * free_huge_pages, and surplus_huge_pages.
83 */
84 DEFINE_SPINLOCK(hugetlb_lock);
85
86 /*
87 * Serializes faults on the same logical page. This is used to
88 * prevent spurious OOMs when the hugepage pool is fully utilized.
89 */
90 static int num_fault_mutexes;
91 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
92
93 /* Forward declaration */
94 static int hugetlb_acct_memory(struct hstate *h, long delta);
95 static void hugetlb_vma_lock_free(struct vm_area_struct *vma);
96 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma);
97 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma);
98 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
99 unsigned long start, unsigned long end);
100 static struct resv_map *vma_resv_map(struct vm_area_struct *vma);
101
subpool_is_free(struct hugepage_subpool * spool)102 static inline bool subpool_is_free(struct hugepage_subpool *spool)
103 {
104 if (spool->count)
105 return false;
106 if (spool->max_hpages != -1)
107 return spool->used_hpages == 0;
108 if (spool->min_hpages != -1)
109 return spool->rsv_hpages == spool->min_hpages;
110
111 return true;
112 }
113
unlock_or_release_subpool(struct hugepage_subpool * spool,unsigned long irq_flags)114 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
115 unsigned long irq_flags)
116 {
117 spin_unlock_irqrestore(&spool->lock, irq_flags);
118
119 /* If no pages are used, and no other handles to the subpool
120 * remain, give up any reservations based on minimum size and
121 * free the subpool */
122 if (subpool_is_free(spool)) {
123 if (spool->min_hpages != -1)
124 hugetlb_acct_memory(spool->hstate,
125 -spool->min_hpages);
126 kfree(spool);
127 }
128 }
129
hugepage_new_subpool(struct hstate * h,long max_hpages,long min_hpages)130 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
131 long min_hpages)
132 {
133 struct hugepage_subpool *spool;
134
135 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
136 if (!spool)
137 return NULL;
138
139 spin_lock_init(&spool->lock);
140 spool->count = 1;
141 spool->max_hpages = max_hpages;
142 spool->hstate = h;
143 spool->min_hpages = min_hpages;
144
145 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
146 kfree(spool);
147 return NULL;
148 }
149 spool->rsv_hpages = min_hpages;
150
151 return spool;
152 }
153
hugepage_put_subpool(struct hugepage_subpool * spool)154 void hugepage_put_subpool(struct hugepage_subpool *spool)
155 {
156 unsigned long flags;
157
158 spin_lock_irqsave(&spool->lock, flags);
159 BUG_ON(!spool->count);
160 spool->count--;
161 unlock_or_release_subpool(spool, flags);
162 }
163
164 /*
165 * Subpool accounting for allocating and reserving pages.
166 * Return -ENOMEM if there are not enough resources to satisfy the
167 * request. Otherwise, return the number of pages by which the
168 * global pools must be adjusted (upward). The returned value may
169 * only be different than the passed value (delta) in the case where
170 * a subpool minimum size must be maintained.
171 */
hugepage_subpool_get_pages(struct hugepage_subpool * spool,long delta)172 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
173 long delta)
174 {
175 long ret = delta;
176
177 if (!spool)
178 return ret;
179
180 spin_lock_irq(&spool->lock);
181
182 if (spool->max_hpages != -1) { /* maximum size accounting */
183 if ((spool->used_hpages + delta) <= spool->max_hpages)
184 spool->used_hpages += delta;
185 else {
186 ret = -ENOMEM;
187 goto unlock_ret;
188 }
189 }
190
191 /* minimum size accounting */
192 if (spool->min_hpages != -1 && spool->rsv_hpages) {
193 if (delta > spool->rsv_hpages) {
194 /*
195 * Asking for more reserves than those already taken on
196 * behalf of subpool. Return difference.
197 */
198 ret = delta - spool->rsv_hpages;
199 spool->rsv_hpages = 0;
200 } else {
201 ret = 0; /* reserves already accounted for */
202 spool->rsv_hpages -= delta;
203 }
204 }
205
206 unlock_ret:
207 spin_unlock_irq(&spool->lock);
208 return ret;
209 }
210
211 /*
212 * Subpool accounting for freeing and unreserving pages.
213 * Return the number of global page reservations that must be dropped.
214 * The return value may only be different than the passed value (delta)
215 * in the case where a subpool minimum size must be maintained.
216 */
hugepage_subpool_put_pages(struct hugepage_subpool * spool,long delta)217 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
218 long delta)
219 {
220 long ret = delta;
221 unsigned long flags;
222
223 if (!spool)
224 return delta;
225
226 spin_lock_irqsave(&spool->lock, flags);
227
228 if (spool->max_hpages != -1) /* maximum size accounting */
229 spool->used_hpages -= delta;
230
231 /* minimum size accounting */
232 if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
233 if (spool->rsv_hpages + delta <= spool->min_hpages)
234 ret = 0;
235 else
236 ret = spool->rsv_hpages + delta - spool->min_hpages;
237
238 spool->rsv_hpages += delta;
239 if (spool->rsv_hpages > spool->min_hpages)
240 spool->rsv_hpages = spool->min_hpages;
241 }
242
243 /*
244 * If hugetlbfs_put_super couldn't free spool due to an outstanding
245 * quota reference, free it now.
246 */
247 unlock_or_release_subpool(spool, flags);
248
249 return ret;
250 }
251
subpool_inode(struct inode * inode)252 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
253 {
254 return HUGETLBFS_SB(inode->i_sb)->spool;
255 }
256
subpool_vma(struct vm_area_struct * vma)257 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
258 {
259 return subpool_inode(file_inode(vma->vm_file));
260 }
261
262 /*
263 * hugetlb vma_lock helper routines
264 */
hugetlb_vma_lock_read(struct vm_area_struct * vma)265 void hugetlb_vma_lock_read(struct vm_area_struct *vma)
266 {
267 if (__vma_shareable_lock(vma)) {
268 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
269
270 down_read(&vma_lock->rw_sema);
271 } else if (__vma_private_lock(vma)) {
272 struct resv_map *resv_map = vma_resv_map(vma);
273
274 down_read(&resv_map->rw_sema);
275 }
276 }
277
hugetlb_vma_unlock_read(struct vm_area_struct * vma)278 void hugetlb_vma_unlock_read(struct vm_area_struct *vma)
279 {
280 if (__vma_shareable_lock(vma)) {
281 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
282
283 up_read(&vma_lock->rw_sema);
284 } else if (__vma_private_lock(vma)) {
285 struct resv_map *resv_map = vma_resv_map(vma);
286
287 up_read(&resv_map->rw_sema);
288 }
289 }
290
hugetlb_vma_lock_write(struct vm_area_struct * vma)291 void hugetlb_vma_lock_write(struct vm_area_struct *vma)
292 {
293 if (__vma_shareable_lock(vma)) {
294 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
295
296 down_write(&vma_lock->rw_sema);
297 } else if (__vma_private_lock(vma)) {
298 struct resv_map *resv_map = vma_resv_map(vma);
299
300 down_write(&resv_map->rw_sema);
301 }
302 }
303
hugetlb_vma_unlock_write(struct vm_area_struct * vma)304 void hugetlb_vma_unlock_write(struct vm_area_struct *vma)
305 {
306 if (__vma_shareable_lock(vma)) {
307 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
308
309 up_write(&vma_lock->rw_sema);
310 } else if (__vma_private_lock(vma)) {
311 struct resv_map *resv_map = vma_resv_map(vma);
312
313 up_write(&resv_map->rw_sema);
314 }
315 }
316
hugetlb_vma_trylock_write(struct vm_area_struct * vma)317 int hugetlb_vma_trylock_write(struct vm_area_struct *vma)
318 {
319
320 if (__vma_shareable_lock(vma)) {
321 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
322
323 return down_write_trylock(&vma_lock->rw_sema);
324 } else if (__vma_private_lock(vma)) {
325 struct resv_map *resv_map = vma_resv_map(vma);
326
327 return down_write_trylock(&resv_map->rw_sema);
328 }
329
330 return 1;
331 }
332
hugetlb_vma_assert_locked(struct vm_area_struct * vma)333 void hugetlb_vma_assert_locked(struct vm_area_struct *vma)
334 {
335 if (__vma_shareable_lock(vma)) {
336 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
337
338 lockdep_assert_held(&vma_lock->rw_sema);
339 } else if (__vma_private_lock(vma)) {
340 struct resv_map *resv_map = vma_resv_map(vma);
341
342 lockdep_assert_held(&resv_map->rw_sema);
343 }
344 }
345
hugetlb_vma_lock_release(struct kref * kref)346 void hugetlb_vma_lock_release(struct kref *kref)
347 {
348 struct hugetlb_vma_lock *vma_lock = container_of(kref,
349 struct hugetlb_vma_lock, refs);
350
351 kfree(vma_lock);
352 }
353
__hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock * vma_lock)354 static void __hugetlb_vma_unlock_write_put(struct hugetlb_vma_lock *vma_lock)
355 {
356 struct vm_area_struct *vma = vma_lock->vma;
357
358 /*
359 * vma_lock structure may or not be released as a result of put,
360 * it certainly will no longer be attached to vma so clear pointer.
361 * Semaphore synchronizes access to vma_lock->vma field.
362 */
363 vma_lock->vma = NULL;
364 vma->vm_private_data = NULL;
365 up_write(&vma_lock->rw_sema);
366 kref_put(&vma_lock->refs, hugetlb_vma_lock_release);
367 }
368
__hugetlb_vma_unlock_write_free(struct vm_area_struct * vma)369 static void __hugetlb_vma_unlock_write_free(struct vm_area_struct *vma)
370 {
371 if (__vma_shareable_lock(vma)) {
372 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
373
374 __hugetlb_vma_unlock_write_put(vma_lock);
375 } else if (__vma_private_lock(vma)) {
376 struct resv_map *resv_map = vma_resv_map(vma);
377
378 /* no free for anon vmas, but still need to unlock */
379 up_write(&resv_map->rw_sema);
380 }
381 }
382
hugetlb_vma_lock_free(struct vm_area_struct * vma)383 static void hugetlb_vma_lock_free(struct vm_area_struct *vma)
384 {
385 /*
386 * Only present in sharable vmas.
387 */
388 if (!vma || !__vma_shareable_lock(vma))
389 return;
390
391 if (vma->vm_private_data) {
392 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
393
394 down_write(&vma_lock->rw_sema);
395 __hugetlb_vma_unlock_write_put(vma_lock);
396 }
397 }
398
hugetlb_vma_lock_alloc(struct vm_area_struct * vma)399 static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma)
400 {
401 struct hugetlb_vma_lock *vma_lock;
402
403 /* Only establish in (flags) sharable vmas */
404 if (!vma || !(vma->vm_flags & VM_MAYSHARE))
405 return;
406
407 /* Should never get here with non-NULL vm_private_data */
408 if (vma->vm_private_data)
409 return;
410
411 vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL);
412 if (!vma_lock) {
413 /*
414 * If we can not allocate structure, then vma can not
415 * participate in pmd sharing. This is only a possible
416 * performance enhancement and memory saving issue.
417 * However, the lock is also used to synchronize page
418 * faults with truncation. If the lock is not present,
419 * unlikely races could leave pages in a file past i_size
420 * until the file is removed. Warn in the unlikely case of
421 * allocation failure.
422 */
423 pr_warn_once("HugeTLB: unable to allocate vma specific lock\n");
424 return;
425 }
426
427 kref_init(&vma_lock->refs);
428 init_rwsem(&vma_lock->rw_sema);
429 vma_lock->vma = vma;
430 vma->vm_private_data = vma_lock;
431 }
432
433 /* Helper that removes a struct file_region from the resv_map cache and returns
434 * it for use.
435 */
436 static struct file_region *
get_file_region_entry_from_cache(struct resv_map * resv,long from,long to)437 get_file_region_entry_from_cache(struct resv_map *resv, long from, long to)
438 {
439 struct file_region *nrg;
440
441 VM_BUG_ON(resv->region_cache_count <= 0);
442
443 resv->region_cache_count--;
444 nrg = list_first_entry(&resv->region_cache, struct file_region, link);
445 list_del(&nrg->link);
446
447 nrg->from = from;
448 nrg->to = to;
449
450 return nrg;
451 }
452
copy_hugetlb_cgroup_uncharge_info(struct file_region * nrg,struct file_region * rg)453 static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg,
454 struct file_region *rg)
455 {
456 #ifdef CONFIG_CGROUP_HUGETLB
457 nrg->reservation_counter = rg->reservation_counter;
458 nrg->css = rg->css;
459 if (rg->css)
460 css_get(rg->css);
461 #endif
462 }
463
464 /* Helper that records hugetlb_cgroup uncharge info. */
record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup * h_cg,struct hstate * h,struct resv_map * resv,struct file_region * nrg)465 static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg,
466 struct hstate *h,
467 struct resv_map *resv,
468 struct file_region *nrg)
469 {
470 #ifdef CONFIG_CGROUP_HUGETLB
471 if (h_cg) {
472 nrg->reservation_counter =
473 &h_cg->rsvd_hugepage[hstate_index(h)];
474 nrg->css = &h_cg->css;
475 /*
476 * The caller will hold exactly one h_cg->css reference for the
477 * whole contiguous reservation region. But this area might be
478 * scattered when there are already some file_regions reside in
479 * it. As a result, many file_regions may share only one css
480 * reference. In order to ensure that one file_region must hold
481 * exactly one h_cg->css reference, we should do css_get for
482 * each file_region and leave the reference held by caller
483 * untouched.
484 */
485 css_get(&h_cg->css);
486 if (!resv->pages_per_hpage)
487 resv->pages_per_hpage = pages_per_huge_page(h);
488 /* pages_per_hpage should be the same for all entries in
489 * a resv_map.
490 */
491 VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h));
492 } else {
493 nrg->reservation_counter = NULL;
494 nrg->css = NULL;
495 }
496 #endif
497 }
498
put_uncharge_info(struct file_region * rg)499 static void put_uncharge_info(struct file_region *rg)
500 {
501 #ifdef CONFIG_CGROUP_HUGETLB
502 if (rg->css)
503 css_put(rg->css);
504 #endif
505 }
506
has_same_uncharge_info(struct file_region * rg,struct file_region * org)507 static bool has_same_uncharge_info(struct file_region *rg,
508 struct file_region *org)
509 {
510 #ifdef CONFIG_CGROUP_HUGETLB
511 return rg->reservation_counter == org->reservation_counter &&
512 rg->css == org->css;
513
514 #else
515 return true;
516 #endif
517 }
518
coalesce_file_region(struct resv_map * resv,struct file_region * rg)519 static void coalesce_file_region(struct resv_map *resv, struct file_region *rg)
520 {
521 struct file_region *nrg, *prg;
522
523 prg = list_prev_entry(rg, link);
524 if (&prg->link != &resv->regions && prg->to == rg->from &&
525 has_same_uncharge_info(prg, rg)) {
526 prg->to = rg->to;
527
528 list_del(&rg->link);
529 put_uncharge_info(rg);
530 kfree(rg);
531
532 rg = prg;
533 }
534
535 nrg = list_next_entry(rg, link);
536 if (&nrg->link != &resv->regions && nrg->from == rg->to &&
537 has_same_uncharge_info(nrg, rg)) {
538 nrg->from = rg->from;
539
540 list_del(&rg->link);
541 put_uncharge_info(rg);
542 kfree(rg);
543 }
544 }
545
546 static inline long
hugetlb_resv_map_add(struct resv_map * map,struct list_head * rg,long from,long to,struct hstate * h,struct hugetlb_cgroup * cg,long * regions_needed)547 hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from,
548 long to, struct hstate *h, struct hugetlb_cgroup *cg,
549 long *regions_needed)
550 {
551 struct file_region *nrg;
552
553 if (!regions_needed) {
554 nrg = get_file_region_entry_from_cache(map, from, to);
555 record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg);
556 list_add(&nrg->link, rg);
557 coalesce_file_region(map, nrg);
558 } else
559 *regions_needed += 1;
560
561 return to - from;
562 }
563
564 /*
565 * Must be called with resv->lock held.
566 *
567 * Calling this with regions_needed != NULL will count the number of pages
568 * to be added but will not modify the linked list. And regions_needed will
569 * indicate the number of file_regions needed in the cache to carry out to add
570 * the regions for this range.
571 */
add_reservation_in_range(struct resv_map * resv,long f,long t,struct hugetlb_cgroup * h_cg,struct hstate * h,long * regions_needed)572 static long add_reservation_in_range(struct resv_map *resv, long f, long t,
573 struct hugetlb_cgroup *h_cg,
574 struct hstate *h, long *regions_needed)
575 {
576 long add = 0;
577 struct list_head *head = &resv->regions;
578 long last_accounted_offset = f;
579 struct file_region *iter, *trg = NULL;
580 struct list_head *rg = NULL;
581
582 if (regions_needed)
583 *regions_needed = 0;
584
585 /* In this loop, we essentially handle an entry for the range
586 * [last_accounted_offset, iter->from), at every iteration, with some
587 * bounds checking.
588 */
589 list_for_each_entry_safe(iter, trg, head, link) {
590 /* Skip irrelevant regions that start before our range. */
591 if (iter->from < f) {
592 /* If this region ends after the last accounted offset,
593 * then we need to update last_accounted_offset.
594 */
595 if (iter->to > last_accounted_offset)
596 last_accounted_offset = iter->to;
597 continue;
598 }
599
600 /* When we find a region that starts beyond our range, we've
601 * finished.
602 */
603 if (iter->from >= t) {
604 rg = iter->link.prev;
605 break;
606 }
607
608 /* Add an entry for last_accounted_offset -> iter->from, and
609 * update last_accounted_offset.
610 */
611 if (iter->from > last_accounted_offset)
612 add += hugetlb_resv_map_add(resv, iter->link.prev,
613 last_accounted_offset,
614 iter->from, h, h_cg,
615 regions_needed);
616
617 last_accounted_offset = iter->to;
618 }
619
620 /* Handle the case where our range extends beyond
621 * last_accounted_offset.
622 */
623 if (!rg)
624 rg = head->prev;
625 if (last_accounted_offset < t)
626 add += hugetlb_resv_map_add(resv, rg, last_accounted_offset,
627 t, h, h_cg, regions_needed);
628
629 return add;
630 }
631
632 /* Must be called with resv->lock acquired. Will drop lock to allocate entries.
633 */
allocate_file_region_entries(struct resv_map * resv,int regions_needed)634 static int allocate_file_region_entries(struct resv_map *resv,
635 int regions_needed)
636 __must_hold(&resv->lock)
637 {
638 LIST_HEAD(allocated_regions);
639 int to_allocate = 0, i = 0;
640 struct file_region *trg = NULL, *rg = NULL;
641
642 VM_BUG_ON(regions_needed < 0);
643
644 /*
645 * Check for sufficient descriptors in the cache to accommodate
646 * the number of in progress add operations plus regions_needed.
647 *
648 * This is a while loop because when we drop the lock, some other call
649 * to region_add or region_del may have consumed some region_entries,
650 * so we keep looping here until we finally have enough entries for
651 * (adds_in_progress + regions_needed).
652 */
653 while (resv->region_cache_count <
654 (resv->adds_in_progress + regions_needed)) {
655 to_allocate = resv->adds_in_progress + regions_needed -
656 resv->region_cache_count;
657
658 /* At this point, we should have enough entries in the cache
659 * for all the existing adds_in_progress. We should only be
660 * needing to allocate for regions_needed.
661 */
662 VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress);
663
664 spin_unlock(&resv->lock);
665 for (i = 0; i < to_allocate; i++) {
666 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
667 if (!trg)
668 goto out_of_memory;
669 list_add(&trg->link, &allocated_regions);
670 }
671
672 spin_lock(&resv->lock);
673
674 list_splice(&allocated_regions, &resv->region_cache);
675 resv->region_cache_count += to_allocate;
676 }
677
678 return 0;
679
680 out_of_memory:
681 list_for_each_entry_safe(rg, trg, &allocated_regions, link) {
682 list_del(&rg->link);
683 kfree(rg);
684 }
685 return -ENOMEM;
686 }
687
688 /*
689 * Add the huge page range represented by [f, t) to the reserve
690 * map. Regions will be taken from the cache to fill in this range.
691 * Sufficient regions should exist in the cache due to the previous
692 * call to region_chg with the same range, but in some cases the cache will not
693 * have sufficient entries due to races with other code doing region_add or
694 * region_del. The extra needed entries will be allocated.
695 *
696 * regions_needed is the out value provided by a previous call to region_chg.
697 *
698 * Return the number of new huge pages added to the map. This number is greater
699 * than or equal to zero. If file_region entries needed to be allocated for
700 * this operation and we were not able to allocate, it returns -ENOMEM.
701 * region_add of regions of length 1 never allocate file_regions and cannot
702 * fail; region_chg will always allocate at least 1 entry and a region_add for
703 * 1 page will only require at most 1 entry.
704 */
region_add(struct resv_map * resv,long f,long t,long in_regions_needed,struct hstate * h,struct hugetlb_cgroup * h_cg)705 static long region_add(struct resv_map *resv, long f, long t,
706 long in_regions_needed, struct hstate *h,
707 struct hugetlb_cgroup *h_cg)
708 {
709 long add = 0, actual_regions_needed = 0;
710
711 spin_lock(&resv->lock);
712 retry:
713
714 /* Count how many regions are actually needed to execute this add. */
715 add_reservation_in_range(resv, f, t, NULL, NULL,
716 &actual_regions_needed);
717
718 /*
719 * Check for sufficient descriptors in the cache to accommodate
720 * this add operation. Note that actual_regions_needed may be greater
721 * than in_regions_needed, as the resv_map may have been modified since
722 * the region_chg call. In this case, we need to make sure that we
723 * allocate extra entries, such that we have enough for all the
724 * existing adds_in_progress, plus the excess needed for this
725 * operation.
726 */
727 if (actual_regions_needed > in_regions_needed &&
728 resv->region_cache_count <
729 resv->adds_in_progress +
730 (actual_regions_needed - in_regions_needed)) {
731 /* region_add operation of range 1 should never need to
732 * allocate file_region entries.
733 */
734 VM_BUG_ON(t - f <= 1);
735
736 if (allocate_file_region_entries(
737 resv, actual_regions_needed - in_regions_needed)) {
738 return -ENOMEM;
739 }
740
741 goto retry;
742 }
743
744 add = add_reservation_in_range(resv, f, t, h_cg, h, NULL);
745
746 resv->adds_in_progress -= in_regions_needed;
747
748 spin_unlock(&resv->lock);
749 return add;
750 }
751
752 /*
753 * Examine the existing reserve map and determine how many
754 * huge pages in the specified range [f, t) are NOT currently
755 * represented. This routine is called before a subsequent
756 * call to region_add that will actually modify the reserve
757 * map to add the specified range [f, t). region_chg does
758 * not change the number of huge pages represented by the
759 * map. A number of new file_region structures is added to the cache as a
760 * placeholder, for the subsequent region_add call to use. At least 1
761 * file_region structure is added.
762 *
763 * out_regions_needed is the number of regions added to the
764 * resv->adds_in_progress. This value needs to be provided to a follow up call
765 * to region_add or region_abort for proper accounting.
766 *
767 * Returns the number of huge pages that need to be added to the existing
768 * reservation map for the range [f, t). This number is greater or equal to
769 * zero. -ENOMEM is returned if a new file_region structure or cache entry
770 * is needed and can not be allocated.
771 */
region_chg(struct resv_map * resv,long f,long t,long * out_regions_needed)772 static long region_chg(struct resv_map *resv, long f, long t,
773 long *out_regions_needed)
774 {
775 long chg = 0;
776
777 spin_lock(&resv->lock);
778
779 /* Count how many hugepages in this range are NOT represented. */
780 chg = add_reservation_in_range(resv, f, t, NULL, NULL,
781 out_regions_needed);
782
783 if (*out_regions_needed == 0)
784 *out_regions_needed = 1;
785
786 if (allocate_file_region_entries(resv, *out_regions_needed))
787 return -ENOMEM;
788
789 resv->adds_in_progress += *out_regions_needed;
790
791 spin_unlock(&resv->lock);
792 return chg;
793 }
794
795 /*
796 * Abort the in progress add operation. The adds_in_progress field
797 * of the resv_map keeps track of the operations in progress between
798 * calls to region_chg and region_add. Operations are sometimes
799 * aborted after the call to region_chg. In such cases, region_abort
800 * is called to decrement the adds_in_progress counter. regions_needed
801 * is the value returned by the region_chg call, it is used to decrement
802 * the adds_in_progress counter.
803 *
804 * NOTE: The range arguments [f, t) are not needed or used in this
805 * routine. They are kept to make reading the calling code easier as
806 * arguments will match the associated region_chg call.
807 */
region_abort(struct resv_map * resv,long f,long t,long regions_needed)808 static void region_abort(struct resv_map *resv, long f, long t,
809 long regions_needed)
810 {
811 spin_lock(&resv->lock);
812 VM_BUG_ON(!resv->region_cache_count);
813 resv->adds_in_progress -= regions_needed;
814 spin_unlock(&resv->lock);
815 }
816
817 /*
818 * Delete the specified range [f, t) from the reserve map. If the
819 * t parameter is LONG_MAX, this indicates that ALL regions after f
820 * should be deleted. Locate the regions which intersect [f, t)
821 * and either trim, delete or split the existing regions.
822 *
823 * Returns the number of huge pages deleted from the reserve map.
824 * In the normal case, the return value is zero or more. In the
825 * case where a region must be split, a new region descriptor must
826 * be allocated. If the allocation fails, -ENOMEM will be returned.
827 * NOTE: If the parameter t == LONG_MAX, then we will never split
828 * a region and possibly return -ENOMEM. Callers specifying
829 * t == LONG_MAX do not need to check for -ENOMEM error.
830 */
region_del(struct resv_map * resv,long f,long t)831 static long region_del(struct resv_map *resv, long f, long t)
832 {
833 struct list_head *head = &resv->regions;
834 struct file_region *rg, *trg;
835 struct file_region *nrg = NULL;
836 long del = 0;
837
838 retry:
839 spin_lock(&resv->lock);
840 list_for_each_entry_safe(rg, trg, head, link) {
841 /*
842 * Skip regions before the range to be deleted. file_region
843 * ranges are normally of the form [from, to). However, there
844 * may be a "placeholder" entry in the map which is of the form
845 * (from, to) with from == to. Check for placeholder entries
846 * at the beginning of the range to be deleted.
847 */
848 if (rg->to <= f && (rg->to != rg->from || rg->to != f))
849 continue;
850
851 if (rg->from >= t)
852 break;
853
854 if (f > rg->from && t < rg->to) { /* Must split region */
855 /*
856 * Check for an entry in the cache before dropping
857 * lock and attempting allocation.
858 */
859 if (!nrg &&
860 resv->region_cache_count > resv->adds_in_progress) {
861 nrg = list_first_entry(&resv->region_cache,
862 struct file_region,
863 link);
864 list_del(&nrg->link);
865 resv->region_cache_count--;
866 }
867
868 if (!nrg) {
869 spin_unlock(&resv->lock);
870 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
871 if (!nrg)
872 return -ENOMEM;
873 goto retry;
874 }
875
876 del += t - f;
877 hugetlb_cgroup_uncharge_file_region(
878 resv, rg, t - f, false);
879
880 /* New entry for end of split region */
881 nrg->from = t;
882 nrg->to = rg->to;
883
884 copy_hugetlb_cgroup_uncharge_info(nrg, rg);
885
886 INIT_LIST_HEAD(&nrg->link);
887
888 /* Original entry is trimmed */
889 rg->to = f;
890
891 list_add(&nrg->link, &rg->link);
892 nrg = NULL;
893 break;
894 }
895
896 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
897 del += rg->to - rg->from;
898 hugetlb_cgroup_uncharge_file_region(resv, rg,
899 rg->to - rg->from, true);
900 list_del(&rg->link);
901 kfree(rg);
902 continue;
903 }
904
905 if (f <= rg->from) { /* Trim beginning of region */
906 hugetlb_cgroup_uncharge_file_region(resv, rg,
907 t - rg->from, false);
908
909 del += t - rg->from;
910 rg->from = t;
911 } else { /* Trim end of region */
912 hugetlb_cgroup_uncharge_file_region(resv, rg,
913 rg->to - f, false);
914
915 del += rg->to - f;
916 rg->to = f;
917 }
918 }
919
920 spin_unlock(&resv->lock);
921 kfree(nrg);
922 return del;
923 }
924
925 /*
926 * A rare out of memory error was encountered which prevented removal of
927 * the reserve map region for a page. The huge page itself was free'ed
928 * and removed from the page cache. This routine will adjust the subpool
929 * usage count, and the global reserve count if needed. By incrementing
930 * these counts, the reserve map entry which could not be deleted will
931 * appear as a "reserved" entry instead of simply dangling with incorrect
932 * counts.
933 */
hugetlb_fix_reserve_counts(struct inode * inode)934 void hugetlb_fix_reserve_counts(struct inode *inode)
935 {
936 struct hugepage_subpool *spool = subpool_inode(inode);
937 long rsv_adjust;
938 bool reserved = false;
939
940 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
941 if (rsv_adjust > 0) {
942 struct hstate *h = hstate_inode(inode);
943
944 if (!hugetlb_acct_memory(h, 1))
945 reserved = true;
946 } else if (!rsv_adjust) {
947 reserved = true;
948 }
949
950 if (!reserved)
951 pr_warn("hugetlb: Huge Page Reserved count may go negative.\n");
952 }
953
954 /*
955 * Count and return the number of huge pages in the reserve map
956 * that intersect with the range [f, t).
957 */
region_count(struct resv_map * resv,long f,long t)958 static long region_count(struct resv_map *resv, long f, long t)
959 {
960 struct list_head *head = &resv->regions;
961 struct file_region *rg;
962 long chg = 0;
963
964 spin_lock(&resv->lock);
965 /* Locate each segment we overlap with, and count that overlap. */
966 list_for_each_entry(rg, head, link) {
967 long seg_from;
968 long seg_to;
969
970 if (rg->to <= f)
971 continue;
972 if (rg->from >= t)
973 break;
974
975 seg_from = max(rg->from, f);
976 seg_to = min(rg->to, t);
977
978 chg += seg_to - seg_from;
979 }
980 spin_unlock(&resv->lock);
981
982 return chg;
983 }
984
985 /*
986 * Convert the address within this vma to the page offset within
987 * the mapping, in pagecache page units; huge pages here.
988 */
vma_hugecache_offset(struct hstate * h,struct vm_area_struct * vma,unsigned long address)989 static pgoff_t vma_hugecache_offset(struct hstate *h,
990 struct vm_area_struct *vma, unsigned long address)
991 {
992 return ((address - vma->vm_start) >> huge_page_shift(h)) +
993 (vma->vm_pgoff >> huge_page_order(h));
994 }
995
linear_hugepage_index(struct vm_area_struct * vma,unsigned long address)996 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
997 unsigned long address)
998 {
999 return vma_hugecache_offset(hstate_vma(vma), vma, address);
1000 }
1001 EXPORT_SYMBOL_GPL(linear_hugepage_index);
1002
1003 /**
1004 * vma_kernel_pagesize - Page size granularity for this VMA.
1005 * @vma: The user mapping.
1006 *
1007 * Folios in this VMA will be aligned to, and at least the size of the
1008 * number of bytes returned by this function.
1009 *
1010 * Return: The default size of the folios allocated when backing a VMA.
1011 */
vma_kernel_pagesize(struct vm_area_struct * vma)1012 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
1013 {
1014 if (vma->vm_ops && vma->vm_ops->pagesize)
1015 return vma->vm_ops->pagesize(vma);
1016 return PAGE_SIZE;
1017 }
1018 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
1019
1020 /*
1021 * Return the page size being used by the MMU to back a VMA. In the majority
1022 * of cases, the page size used by the kernel matches the MMU size. On
1023 * architectures where it differs, an architecture-specific 'strong'
1024 * version of this symbol is required.
1025 */
vma_mmu_pagesize(struct vm_area_struct * vma)1026 __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
1027 {
1028 return vma_kernel_pagesize(vma);
1029 }
1030
1031 /*
1032 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
1033 * bits of the reservation map pointer, which are always clear due to
1034 * alignment.
1035 */
1036 #define HPAGE_RESV_OWNER (1UL << 0)
1037 #define HPAGE_RESV_UNMAPPED (1UL << 1)
1038 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
1039
1040 /*
1041 * These helpers are used to track how many pages are reserved for
1042 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
1043 * is guaranteed to have their future faults succeed.
1044 *
1045 * With the exception of hugetlb_dup_vma_private() which is called at fork(),
1046 * the reserve counters are updated with the hugetlb_lock held. It is safe
1047 * to reset the VMA at fork() time as it is not in use yet and there is no
1048 * chance of the global counters getting corrupted as a result of the values.
1049 *
1050 * The private mapping reservation is represented in a subtly different
1051 * manner to a shared mapping. A shared mapping has a region map associated
1052 * with the underlying file, this region map represents the backing file
1053 * pages which have ever had a reservation assigned which this persists even
1054 * after the page is instantiated. A private mapping has a region map
1055 * associated with the original mmap which is attached to all VMAs which
1056 * reference it, this region map represents those offsets which have consumed
1057 * reservation ie. where pages have been instantiated.
1058 */
get_vma_private_data(struct vm_area_struct * vma)1059 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
1060 {
1061 return (unsigned long)vma->vm_private_data;
1062 }
1063
set_vma_private_data(struct vm_area_struct * vma,unsigned long value)1064 static void set_vma_private_data(struct vm_area_struct *vma,
1065 unsigned long value)
1066 {
1067 vma->vm_private_data = (void *)value;
1068 }
1069
1070 static void
resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map * resv_map,struct hugetlb_cgroup * h_cg,struct hstate * h)1071 resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map,
1072 struct hugetlb_cgroup *h_cg,
1073 struct hstate *h)
1074 {
1075 #ifdef CONFIG_CGROUP_HUGETLB
1076 if (!h_cg || !h) {
1077 resv_map->reservation_counter = NULL;
1078 resv_map->pages_per_hpage = 0;
1079 resv_map->css = NULL;
1080 } else {
1081 resv_map->reservation_counter =
1082 &h_cg->rsvd_hugepage[hstate_index(h)];
1083 resv_map->pages_per_hpage = pages_per_huge_page(h);
1084 resv_map->css = &h_cg->css;
1085 }
1086 #endif
1087 }
1088
resv_map_alloc(void)1089 struct resv_map *resv_map_alloc(void)
1090 {
1091 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
1092 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
1093
1094 if (!resv_map || !rg) {
1095 kfree(resv_map);
1096 kfree(rg);
1097 return NULL;
1098 }
1099
1100 kref_init(&resv_map->refs);
1101 spin_lock_init(&resv_map->lock);
1102 INIT_LIST_HEAD(&resv_map->regions);
1103 init_rwsem(&resv_map->rw_sema);
1104
1105 resv_map->adds_in_progress = 0;
1106 /*
1107 * Initialize these to 0. On shared mappings, 0's here indicate these
1108 * fields don't do cgroup accounting. On private mappings, these will be
1109 * re-initialized to the proper values, to indicate that hugetlb cgroup
1110 * reservations are to be un-charged from here.
1111 */
1112 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL);
1113
1114 INIT_LIST_HEAD(&resv_map->region_cache);
1115 list_add(&rg->link, &resv_map->region_cache);
1116 resv_map->region_cache_count = 1;
1117
1118 return resv_map;
1119 }
1120
resv_map_release(struct kref * ref)1121 void resv_map_release(struct kref *ref)
1122 {
1123 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
1124 struct list_head *head = &resv_map->region_cache;
1125 struct file_region *rg, *trg;
1126
1127 /* Clear out any active regions before we release the map. */
1128 region_del(resv_map, 0, LONG_MAX);
1129
1130 /* ... and any entries left in the cache */
1131 list_for_each_entry_safe(rg, trg, head, link) {
1132 list_del(&rg->link);
1133 kfree(rg);
1134 }
1135
1136 VM_BUG_ON(resv_map->adds_in_progress);
1137
1138 kfree(resv_map);
1139 }
1140
inode_resv_map(struct inode * inode)1141 static inline struct resv_map *inode_resv_map(struct inode *inode)
1142 {
1143 /*
1144 * At inode evict time, i_mapping may not point to the original
1145 * address space within the inode. This original address space
1146 * contains the pointer to the resv_map. So, always use the
1147 * address space embedded within the inode.
1148 * The VERY common case is inode->mapping == &inode->i_data but,
1149 * this may not be true for device special inodes.
1150 */
1151 return (struct resv_map *)(&inode->i_data)->private_data;
1152 }
1153
vma_resv_map(struct vm_area_struct * vma)1154 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
1155 {
1156 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1157 if (vma->vm_flags & VM_MAYSHARE) {
1158 struct address_space *mapping = vma->vm_file->f_mapping;
1159 struct inode *inode = mapping->host;
1160
1161 return inode_resv_map(inode);
1162
1163 } else {
1164 return (struct resv_map *)(get_vma_private_data(vma) &
1165 ~HPAGE_RESV_MASK);
1166 }
1167 }
1168
set_vma_resv_map(struct vm_area_struct * vma,struct resv_map * map)1169 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
1170 {
1171 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1172 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1173
1174 set_vma_private_data(vma, (unsigned long)map);
1175 }
1176
set_vma_resv_flags(struct vm_area_struct * vma,unsigned long flags)1177 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
1178 {
1179 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1180 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
1181
1182 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
1183 }
1184
is_vma_resv_set(struct vm_area_struct * vma,unsigned long flag)1185 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
1186 {
1187 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1188
1189 return (get_vma_private_data(vma) & flag) != 0;
1190 }
1191
__vma_private_lock(struct vm_area_struct * vma)1192 bool __vma_private_lock(struct vm_area_struct *vma)
1193 {
1194 return !(vma->vm_flags & VM_MAYSHARE) &&
1195 get_vma_private_data(vma) & ~HPAGE_RESV_MASK &&
1196 is_vma_resv_set(vma, HPAGE_RESV_OWNER);
1197 }
1198
hugetlb_dup_vma_private(struct vm_area_struct * vma)1199 void hugetlb_dup_vma_private(struct vm_area_struct *vma)
1200 {
1201 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
1202 /*
1203 * Clear vm_private_data
1204 * - For shared mappings this is a per-vma semaphore that may be
1205 * allocated in a subsequent call to hugetlb_vm_op_open.
1206 * Before clearing, make sure pointer is not associated with vma
1207 * as this will leak the structure. This is the case when called
1208 * via clear_vma_resv_huge_pages() and hugetlb_vm_op_open has already
1209 * been called to allocate a new structure.
1210 * - For MAP_PRIVATE mappings, this is the reserve map which does
1211 * not apply to children. Faults generated by the children are
1212 * not guaranteed to succeed, even if read-only.
1213 */
1214 if (vma->vm_flags & VM_MAYSHARE) {
1215 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
1216
1217 if (vma_lock && vma_lock->vma != vma)
1218 vma->vm_private_data = NULL;
1219 } else
1220 vma->vm_private_data = NULL;
1221 }
1222
1223 /*
1224 * Reset and decrement one ref on hugepage private reservation.
1225 * Called with mm->mmap_lock writer semaphore held.
1226 * This function should be only used by move_vma() and operate on
1227 * same sized vma. It should never come here with last ref on the
1228 * reservation.
1229 */
clear_vma_resv_huge_pages(struct vm_area_struct * vma)1230 void clear_vma_resv_huge_pages(struct vm_area_struct *vma)
1231 {
1232 /*
1233 * Clear the old hugetlb private page reservation.
1234 * It has already been transferred to new_vma.
1235 *
1236 * During a mremap() operation of a hugetlb vma we call move_vma()
1237 * which copies vma into new_vma and unmaps vma. After the copy
1238 * operation both new_vma and vma share a reference to the resv_map
1239 * struct, and at that point vma is about to be unmapped. We don't
1240 * want to return the reservation to the pool at unmap of vma because
1241 * the reservation still lives on in new_vma, so simply decrement the
1242 * ref here and remove the resv_map reference from this vma.
1243 */
1244 struct resv_map *reservations = vma_resv_map(vma);
1245
1246 if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1247 resv_map_put_hugetlb_cgroup_uncharge_info(reservations);
1248 kref_put(&reservations->refs, resv_map_release);
1249 }
1250
1251 hugetlb_dup_vma_private(vma);
1252 }
1253
1254 /* Returns true if the VMA has associated reserve pages */
vma_has_reserves(struct vm_area_struct * vma,long chg)1255 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
1256 {
1257 if (vma->vm_flags & VM_NORESERVE) {
1258 /*
1259 * This address is already reserved by other process(chg == 0),
1260 * so, we should decrement reserved count. Without decrementing,
1261 * reserve count remains after releasing inode, because this
1262 * allocated page will go into page cache and is regarded as
1263 * coming from reserved pool in releasing step. Currently, we
1264 * don't have any other solution to deal with this situation
1265 * properly, so add work-around here.
1266 */
1267 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
1268 return true;
1269 else
1270 return false;
1271 }
1272
1273 /* Shared mappings always use reserves */
1274 if (vma->vm_flags & VM_MAYSHARE) {
1275 /*
1276 * We know VM_NORESERVE is not set. Therefore, there SHOULD
1277 * be a region map for all pages. The only situation where
1278 * there is no region map is if a hole was punched via
1279 * fallocate. In this case, there really are no reserves to
1280 * use. This situation is indicated if chg != 0.
1281 */
1282 if (chg)
1283 return false;
1284 else
1285 return true;
1286 }
1287
1288 /*
1289 * Only the process that called mmap() has reserves for
1290 * private mappings.
1291 */
1292 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
1293 /*
1294 * Like the shared case above, a hole punch or truncate
1295 * could have been performed on the private mapping.
1296 * Examine the value of chg to determine if reserves
1297 * actually exist or were previously consumed.
1298 * Very Subtle - The value of chg comes from a previous
1299 * call to vma_needs_reserves(). The reserve map for
1300 * private mappings has different (opposite) semantics
1301 * than that of shared mappings. vma_needs_reserves()
1302 * has already taken this difference in semantics into
1303 * account. Therefore, the meaning of chg is the same
1304 * as in the shared case above. Code could easily be
1305 * combined, but keeping it separate draws attention to
1306 * subtle differences.
1307 */
1308 if (chg)
1309 return false;
1310 else
1311 return true;
1312 }
1313
1314 return false;
1315 }
1316
enqueue_hugetlb_folio(struct hstate * h,struct folio * folio)1317 static void enqueue_hugetlb_folio(struct hstate *h, struct folio *folio)
1318 {
1319 int nid = folio_nid(folio);
1320
1321 lockdep_assert_held(&hugetlb_lock);
1322 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1323
1324 list_move(&folio->lru, &h->hugepage_freelists[nid]);
1325 h->free_huge_pages++;
1326 h->free_huge_pages_node[nid]++;
1327 folio_set_hugetlb_freed(folio);
1328 }
1329
dequeue_hugetlb_folio_node_exact(struct hstate * h,int nid)1330 static struct folio *dequeue_hugetlb_folio_node_exact(struct hstate *h,
1331 int nid)
1332 {
1333 struct folio *folio;
1334 bool pin = !!(current->flags & PF_MEMALLOC_PIN);
1335
1336 lockdep_assert_held(&hugetlb_lock);
1337 list_for_each_entry(folio, &h->hugepage_freelists[nid], lru) {
1338 if (pin && !folio_is_longterm_pinnable(folio))
1339 continue;
1340
1341 if (folio_test_hwpoison(folio))
1342 continue;
1343
1344 list_move(&folio->lru, &h->hugepage_activelist);
1345 folio_ref_unfreeze(folio, 1);
1346 folio_clear_hugetlb_freed(folio);
1347 h->free_huge_pages--;
1348 h->free_huge_pages_node[nid]--;
1349 return folio;
1350 }
1351
1352 return NULL;
1353 }
1354
dequeue_hugetlb_folio_nodemask(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask)1355 static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_mask,
1356 int nid, nodemask_t *nmask)
1357 {
1358 unsigned int cpuset_mems_cookie;
1359 struct zonelist *zonelist;
1360 struct zone *zone;
1361 struct zoneref *z;
1362 int node = NUMA_NO_NODE;
1363
1364 zonelist = node_zonelist(nid, gfp_mask);
1365
1366 retry_cpuset:
1367 cpuset_mems_cookie = read_mems_allowed_begin();
1368 for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
1369 struct folio *folio;
1370
1371 if (!cpuset_zone_allowed(zone, gfp_mask))
1372 continue;
1373 /*
1374 * no need to ask again on the same node. Pool is node rather than
1375 * zone aware
1376 */
1377 if (zone_to_nid(zone) == node)
1378 continue;
1379 node = zone_to_nid(zone);
1380
1381 folio = dequeue_hugetlb_folio_node_exact(h, node);
1382 if (folio)
1383 return folio;
1384 }
1385 if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
1386 goto retry_cpuset;
1387
1388 return NULL;
1389 }
1390
available_huge_pages(struct hstate * h)1391 static unsigned long available_huge_pages(struct hstate *h)
1392 {
1393 return h->free_huge_pages - h->resv_huge_pages;
1394 }
1395
dequeue_hugetlb_folio_vma(struct hstate * h,struct vm_area_struct * vma,unsigned long address,int avoid_reserve,long chg)1396 static struct folio *dequeue_hugetlb_folio_vma(struct hstate *h,
1397 struct vm_area_struct *vma,
1398 unsigned long address, int avoid_reserve,
1399 long chg)
1400 {
1401 struct folio *folio = NULL;
1402 struct mempolicy *mpol;
1403 gfp_t gfp_mask;
1404 nodemask_t *nodemask;
1405 int nid;
1406
1407 /*
1408 * A child process with MAP_PRIVATE mappings created by their parent
1409 * have no page reserves. This check ensures that reservations are
1410 * not "stolen". The child may still get SIGKILLed
1411 */
1412 if (!vma_has_reserves(vma, chg) && !available_huge_pages(h))
1413 goto err;
1414
1415 /* If reserves cannot be used, ensure enough pages are in the pool */
1416 if (avoid_reserve && !available_huge_pages(h))
1417 goto err;
1418
1419 gfp_mask = htlb_alloc_mask(h);
1420 nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1421
1422 if (mpol_is_preferred_many(mpol)) {
1423 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1424 nid, nodemask);
1425
1426 /* Fallback to all nodes if page==NULL */
1427 nodemask = NULL;
1428 }
1429
1430 if (!folio)
1431 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
1432 nid, nodemask);
1433
1434 if (folio && !avoid_reserve && vma_has_reserves(vma, chg)) {
1435 folio_set_hugetlb_restore_reserve(folio);
1436 h->resv_huge_pages--;
1437 }
1438
1439 mpol_cond_put(mpol);
1440 return folio;
1441
1442 err:
1443 return NULL;
1444 }
1445
1446 /*
1447 * common helper functions for hstate_next_node_to_{alloc|free}.
1448 * We may have allocated or freed a huge page based on a different
1449 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
1450 * be outside of *nodes_allowed. Ensure that we use an allowed
1451 * node for alloc or free.
1452 */
next_node_allowed(int nid,nodemask_t * nodes_allowed)1453 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
1454 {
1455 nid = next_node_in(nid, *nodes_allowed);
1456 VM_BUG_ON(nid >= MAX_NUMNODES);
1457
1458 return nid;
1459 }
1460
get_valid_node_allowed(int nid,nodemask_t * nodes_allowed)1461 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
1462 {
1463 if (!node_isset(nid, *nodes_allowed))
1464 nid = next_node_allowed(nid, nodes_allowed);
1465 return nid;
1466 }
1467
1468 /*
1469 * returns the previously saved node ["this node"] from which to
1470 * allocate a persistent huge page for the pool and advance the
1471 * next node from which to allocate, handling wrap at end of node
1472 * mask.
1473 */
hstate_next_node_to_alloc(struct hstate * h,nodemask_t * nodes_allowed)1474 static int hstate_next_node_to_alloc(struct hstate *h,
1475 nodemask_t *nodes_allowed)
1476 {
1477 int nid;
1478
1479 VM_BUG_ON(!nodes_allowed);
1480
1481 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
1482 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
1483
1484 return nid;
1485 }
1486
1487 /*
1488 * helper for remove_pool_huge_page() - return the previously saved
1489 * node ["this node"] from which to free a huge page. Advance the
1490 * next node id whether or not we find a free huge page to free so
1491 * that the next attempt to free addresses the next node.
1492 */
hstate_next_node_to_free(struct hstate * h,nodemask_t * nodes_allowed)1493 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
1494 {
1495 int nid;
1496
1497 VM_BUG_ON(!nodes_allowed);
1498
1499 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
1500 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
1501
1502 return nid;
1503 }
1504
1505 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1506 for (nr_nodes = nodes_weight(*mask); \
1507 nr_nodes > 0 && \
1508 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1509 nr_nodes--)
1510
1511 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1512 for (nr_nodes = nodes_weight(*mask); \
1513 nr_nodes > 0 && \
1514 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1515 nr_nodes--)
1516
1517 /* used to demote non-gigantic_huge pages as well */
__destroy_compound_gigantic_folio(struct folio * folio,unsigned int order,bool demote)1518 static void __destroy_compound_gigantic_folio(struct folio *folio,
1519 unsigned int order, bool demote)
1520 {
1521 int i;
1522 int nr_pages = 1 << order;
1523 struct page *p;
1524
1525 atomic_set(&folio->_entire_mapcount, 0);
1526 atomic_set(&folio->_nr_pages_mapped, 0);
1527 atomic_set(&folio->_pincount, 0);
1528
1529 for (i = 1; i < nr_pages; i++) {
1530 p = folio_page(folio, i);
1531 p->flags &= ~PAGE_FLAGS_CHECK_AT_FREE;
1532 p->mapping = NULL;
1533 clear_compound_head(p);
1534 if (!demote)
1535 set_page_refcounted(p);
1536 }
1537
1538 __folio_clear_head(folio);
1539 }
1540
destroy_compound_hugetlb_folio_for_demote(struct folio * folio,unsigned int order)1541 static void destroy_compound_hugetlb_folio_for_demote(struct folio *folio,
1542 unsigned int order)
1543 {
1544 __destroy_compound_gigantic_folio(folio, order, true);
1545 }
1546
1547 #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
destroy_compound_gigantic_folio(struct folio * folio,unsigned int order)1548 static void destroy_compound_gigantic_folio(struct folio *folio,
1549 unsigned int order)
1550 {
1551 __destroy_compound_gigantic_folio(folio, order, false);
1552 }
1553
free_gigantic_folio(struct folio * folio,unsigned int order)1554 static void free_gigantic_folio(struct folio *folio, unsigned int order)
1555 {
1556 /*
1557 * If the page isn't allocated using the cma allocator,
1558 * cma_release() returns false.
1559 */
1560 #ifdef CONFIG_CMA
1561 int nid = folio_nid(folio);
1562
1563 if (cma_release(hugetlb_cma[nid], &folio->page, 1 << order))
1564 return;
1565 #endif
1566
1567 free_contig_range(folio_pfn(folio), 1 << order);
1568 }
1569
1570 #ifdef CONFIG_CONTIG_ALLOC
alloc_gigantic_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nodemask)1571 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1572 int nid, nodemask_t *nodemask)
1573 {
1574 struct page *page;
1575 unsigned long nr_pages = pages_per_huge_page(h);
1576 if (nid == NUMA_NO_NODE)
1577 nid = numa_mem_id();
1578
1579 #ifdef CONFIG_CMA
1580 {
1581 int node;
1582
1583 if (hugetlb_cma[nid]) {
1584 page = cma_alloc(hugetlb_cma[nid], nr_pages,
1585 huge_page_order(h), true);
1586 if (page)
1587 return page_folio(page);
1588 }
1589
1590 if (!(gfp_mask & __GFP_THISNODE)) {
1591 for_each_node_mask(node, *nodemask) {
1592 if (node == nid || !hugetlb_cma[node])
1593 continue;
1594
1595 page = cma_alloc(hugetlb_cma[node], nr_pages,
1596 huge_page_order(h), true);
1597 if (page)
1598 return page_folio(page);
1599 }
1600 }
1601 }
1602 #endif
1603
1604 page = alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1605 return page ? page_folio(page) : NULL;
1606 }
1607
1608 #else /* !CONFIG_CONTIG_ALLOC */
alloc_gigantic_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nodemask)1609 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1610 int nid, nodemask_t *nodemask)
1611 {
1612 return NULL;
1613 }
1614 #endif /* CONFIG_CONTIG_ALLOC */
1615
1616 #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
alloc_gigantic_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nodemask)1617 static struct folio *alloc_gigantic_folio(struct hstate *h, gfp_t gfp_mask,
1618 int nid, nodemask_t *nodemask)
1619 {
1620 return NULL;
1621 }
free_gigantic_folio(struct folio * folio,unsigned int order)1622 static inline void free_gigantic_folio(struct folio *folio,
1623 unsigned int order) { }
destroy_compound_gigantic_folio(struct folio * folio,unsigned int order)1624 static inline void destroy_compound_gigantic_folio(struct folio *folio,
1625 unsigned int order) { }
1626 #endif
1627
__clear_hugetlb_destructor(struct hstate * h,struct folio * folio)1628 static inline void __clear_hugetlb_destructor(struct hstate *h,
1629 struct folio *folio)
1630 {
1631 lockdep_assert_held(&hugetlb_lock);
1632
1633 __folio_clear_hugetlb(folio);
1634 }
1635
1636 /*
1637 * Remove hugetlb folio from lists.
1638 * If vmemmap exists for the folio, update dtor so that the folio appears
1639 * as just a compound page. Otherwise, wait until after allocating vmemmap
1640 * to update dtor.
1641 *
1642 * A reference is held on the folio, except in the case of demote.
1643 *
1644 * Must be called with hugetlb lock held.
1645 */
__remove_hugetlb_folio(struct hstate * h,struct folio * folio,bool adjust_surplus,bool demote)1646 static void __remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1647 bool adjust_surplus,
1648 bool demote)
1649 {
1650 int nid = folio_nid(folio);
1651
1652 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio(folio), folio);
1653 VM_BUG_ON_FOLIO(hugetlb_cgroup_from_folio_rsvd(folio), folio);
1654
1655 lockdep_assert_held(&hugetlb_lock);
1656 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1657 return;
1658
1659 list_del(&folio->lru);
1660
1661 if (folio_test_hugetlb_freed(folio)) {
1662 h->free_huge_pages--;
1663 h->free_huge_pages_node[nid]--;
1664 }
1665 if (adjust_surplus) {
1666 h->surplus_huge_pages--;
1667 h->surplus_huge_pages_node[nid]--;
1668 }
1669
1670 /*
1671 * We can only clear the hugetlb destructor after allocating vmemmap
1672 * pages. Otherwise, someone (memory error handling) may try to write
1673 * to tail struct pages.
1674 */
1675 if (!folio_test_hugetlb_vmemmap_optimized(folio))
1676 __clear_hugetlb_destructor(h, folio);
1677
1678 /*
1679 * In the case of demote we do not ref count the page as it will soon
1680 * be turned into a page of smaller size.
1681 */
1682 if (!demote)
1683 folio_ref_unfreeze(folio, 1);
1684
1685 h->nr_huge_pages--;
1686 h->nr_huge_pages_node[nid]--;
1687 }
1688
remove_hugetlb_folio(struct hstate * h,struct folio * folio,bool adjust_surplus)1689 static void remove_hugetlb_folio(struct hstate *h, struct folio *folio,
1690 bool adjust_surplus)
1691 {
1692 __remove_hugetlb_folio(h, folio, adjust_surplus, false);
1693 }
1694
remove_hugetlb_folio_for_demote(struct hstate * h,struct folio * folio,bool adjust_surplus)1695 static void remove_hugetlb_folio_for_demote(struct hstate *h, struct folio *folio,
1696 bool adjust_surplus)
1697 {
1698 __remove_hugetlb_folio(h, folio, adjust_surplus, true);
1699 }
1700
add_hugetlb_folio(struct hstate * h,struct folio * folio,bool adjust_surplus)1701 static void add_hugetlb_folio(struct hstate *h, struct folio *folio,
1702 bool adjust_surplus)
1703 {
1704 int zeroed;
1705 int nid = folio_nid(folio);
1706
1707 VM_BUG_ON_FOLIO(!folio_test_hugetlb_vmemmap_optimized(folio), folio);
1708
1709 lockdep_assert_held(&hugetlb_lock);
1710
1711 INIT_LIST_HEAD(&folio->lru);
1712 h->nr_huge_pages++;
1713 h->nr_huge_pages_node[nid]++;
1714
1715 if (adjust_surplus) {
1716 h->surplus_huge_pages++;
1717 h->surplus_huge_pages_node[nid]++;
1718 }
1719
1720 __folio_set_hugetlb(folio);
1721 folio_change_private(folio, NULL);
1722 /*
1723 * We have to set hugetlb_vmemmap_optimized again as above
1724 * folio_change_private(folio, NULL) cleared it.
1725 */
1726 folio_set_hugetlb_vmemmap_optimized(folio);
1727
1728 /*
1729 * This folio is about to be managed by the hugetlb allocator and
1730 * should have no users. Drop our reference, and check for others
1731 * just in case.
1732 */
1733 zeroed = folio_put_testzero(folio);
1734 if (unlikely(!zeroed))
1735 /*
1736 * It is VERY unlikely soneone else has taken a ref
1737 * on the folio. In this case, we simply return as
1738 * free_huge_folio() will be called when this other ref
1739 * is dropped.
1740 */
1741 return;
1742
1743 arch_clear_hugepage_flags(&folio->page);
1744 enqueue_hugetlb_folio(h, folio);
1745 }
1746
__update_and_free_hugetlb_folio(struct hstate * h,struct folio * folio)1747 static void __update_and_free_hugetlb_folio(struct hstate *h,
1748 struct folio *folio)
1749 {
1750 bool clear_dtor = folio_test_hugetlb_vmemmap_optimized(folio);
1751
1752 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1753 return;
1754
1755 /*
1756 * If we don't know which subpages are hwpoisoned, we can't free
1757 * the hugepage, so it's leaked intentionally.
1758 */
1759 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1760 return;
1761
1762 if (hugetlb_vmemmap_restore(h, &folio->page)) {
1763 spin_lock_irq(&hugetlb_lock);
1764 /*
1765 * If we cannot allocate vmemmap pages, just refuse to free the
1766 * page and put the page back on the hugetlb free list and treat
1767 * as a surplus page.
1768 */
1769 add_hugetlb_folio(h, folio, true);
1770 spin_unlock_irq(&hugetlb_lock);
1771 return;
1772 }
1773
1774 /*
1775 * Move PageHWPoison flag from head page to the raw error pages,
1776 * which makes any healthy subpages reusable.
1777 */
1778 if (unlikely(folio_test_hwpoison(folio)))
1779 folio_clear_hugetlb_hwpoison(folio);
1780
1781 /*
1782 * If vmemmap pages were allocated above, then we need to clear the
1783 * hugetlb destructor under the hugetlb lock.
1784 */
1785 if (clear_dtor) {
1786 spin_lock_irq(&hugetlb_lock);
1787 __clear_hugetlb_destructor(h, folio);
1788 spin_unlock_irq(&hugetlb_lock);
1789 }
1790
1791 /*
1792 * Non-gigantic pages demoted from CMA allocated gigantic pages
1793 * need to be given back to CMA in free_gigantic_folio.
1794 */
1795 if (hstate_is_gigantic(h) ||
1796 hugetlb_cma_folio(folio, huge_page_order(h))) {
1797 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1798 free_gigantic_folio(folio, huge_page_order(h));
1799 } else {
1800 __free_pages(&folio->page, huge_page_order(h));
1801 }
1802 }
1803
1804 /*
1805 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1806 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1807 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1808 * the vmemmap pages.
1809 *
1810 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1811 * freed and frees them one-by-one. As the page->mapping pointer is going
1812 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1813 * structure of a lockless linked list of huge pages to be freed.
1814 */
1815 static LLIST_HEAD(hpage_freelist);
1816
free_hpage_workfn(struct work_struct * work)1817 static void free_hpage_workfn(struct work_struct *work)
1818 {
1819 struct llist_node *node;
1820
1821 node = llist_del_all(&hpage_freelist);
1822
1823 while (node) {
1824 struct page *page;
1825 struct hstate *h;
1826
1827 page = container_of((struct address_space **)node,
1828 struct page, mapping);
1829 node = node->next;
1830 page->mapping = NULL;
1831 /*
1832 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1833 * folio_hstate() is going to trigger because a previous call to
1834 * remove_hugetlb_folio() will clear the hugetlb bit, so do
1835 * not use folio_hstate() directly.
1836 */
1837 h = size_to_hstate(page_size(page));
1838
1839 __update_and_free_hugetlb_folio(h, page_folio(page));
1840
1841 cond_resched();
1842 }
1843 }
1844 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1845
flush_free_hpage_work(struct hstate * h)1846 static inline void flush_free_hpage_work(struct hstate *h)
1847 {
1848 if (hugetlb_vmemmap_optimizable(h))
1849 flush_work(&free_hpage_work);
1850 }
1851
update_and_free_hugetlb_folio(struct hstate * h,struct folio * folio,bool atomic)1852 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1853 bool atomic)
1854 {
1855 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1856 __update_and_free_hugetlb_folio(h, folio);
1857 return;
1858 }
1859
1860 /*
1861 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1862 *
1863 * Only call schedule_work() if hpage_freelist is previously
1864 * empty. Otherwise, schedule_work() had been called but the workfn
1865 * hasn't retrieved the list yet.
1866 */
1867 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1868 schedule_work(&free_hpage_work);
1869 }
1870
update_and_free_pages_bulk(struct hstate * h,struct list_head * list)1871 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1872 {
1873 struct page *page, *t_page;
1874 struct folio *folio;
1875
1876 list_for_each_entry_safe(page, t_page, list, lru) {
1877 folio = page_folio(page);
1878 update_and_free_hugetlb_folio(h, folio, false);
1879 cond_resched();
1880 }
1881 }
1882
size_to_hstate(unsigned long size)1883 struct hstate *size_to_hstate(unsigned long size)
1884 {
1885 struct hstate *h;
1886
1887 for_each_hstate(h) {
1888 if (huge_page_size(h) == size)
1889 return h;
1890 }
1891 return NULL;
1892 }
1893
free_huge_folio(struct folio * folio)1894 void free_huge_folio(struct folio *folio)
1895 {
1896 /*
1897 * Can't pass hstate in here because it is called from the
1898 * compound page destructor.
1899 */
1900 struct hstate *h = folio_hstate(folio);
1901 int nid = folio_nid(folio);
1902 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1903 bool restore_reserve;
1904 unsigned long flags;
1905
1906 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1907 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1908
1909 hugetlb_set_folio_subpool(folio, NULL);
1910 if (folio_test_anon(folio))
1911 __ClearPageAnonExclusive(&folio->page);
1912 folio->mapping = NULL;
1913 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1914 folio_clear_hugetlb_restore_reserve(folio);
1915
1916 /*
1917 * If HPageRestoreReserve was set on page, page allocation consumed a
1918 * reservation. If the page was associated with a subpool, there
1919 * would have been a page reserved in the subpool before allocation
1920 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1921 * reservation, do not call hugepage_subpool_put_pages() as this will
1922 * remove the reserved page from the subpool.
1923 */
1924 if (!restore_reserve) {
1925 /*
1926 * A return code of zero implies that the subpool will be
1927 * under its minimum size if the reservation is not restored
1928 * after page is free. Therefore, force restore_reserve
1929 * operation.
1930 */
1931 if (hugepage_subpool_put_pages(spool, 1) == 0)
1932 restore_reserve = true;
1933 }
1934
1935 spin_lock_irqsave(&hugetlb_lock, flags);
1936 folio_clear_hugetlb_migratable(folio);
1937 hugetlb_cgroup_uncharge_folio(hstate_index(h),
1938 pages_per_huge_page(h), folio);
1939 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1940 pages_per_huge_page(h), folio);
1941 if (restore_reserve)
1942 h->resv_huge_pages++;
1943
1944 if (folio_test_hugetlb_temporary(folio)) {
1945 remove_hugetlb_folio(h, folio, false);
1946 spin_unlock_irqrestore(&hugetlb_lock, flags);
1947 update_and_free_hugetlb_folio(h, folio, true);
1948 } else if (h->surplus_huge_pages_node[nid]) {
1949 /* remove the page from active list */
1950 remove_hugetlb_folio(h, folio, true);
1951 spin_unlock_irqrestore(&hugetlb_lock, flags);
1952 update_and_free_hugetlb_folio(h, folio, true);
1953 } else {
1954 arch_clear_hugepage_flags(&folio->page);
1955 enqueue_hugetlb_folio(h, folio);
1956 spin_unlock_irqrestore(&hugetlb_lock, flags);
1957 }
1958 }
1959
1960 /*
1961 * Must be called with the hugetlb lock held
1962 */
__prep_account_new_huge_page(struct hstate * h,int nid)1963 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1964 {
1965 lockdep_assert_held(&hugetlb_lock);
1966 h->nr_huge_pages++;
1967 h->nr_huge_pages_node[nid]++;
1968 }
1969
__prep_new_hugetlb_folio(struct hstate * h,struct folio * folio)1970 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1971 {
1972 hugetlb_vmemmap_optimize(h, &folio->page);
1973 INIT_LIST_HEAD(&folio->lru);
1974 __folio_set_hugetlb(folio);
1975 hugetlb_set_folio_subpool(folio, NULL);
1976 set_hugetlb_cgroup(folio, NULL);
1977 set_hugetlb_cgroup_rsvd(folio, NULL);
1978 }
1979
prep_new_hugetlb_folio(struct hstate * h,struct folio * folio,int nid)1980 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1981 {
1982 __prep_new_hugetlb_folio(h, folio);
1983 spin_lock_irq(&hugetlb_lock);
1984 __prep_account_new_huge_page(h, nid);
1985 spin_unlock_irq(&hugetlb_lock);
1986 }
1987
__prep_compound_gigantic_folio(struct folio * folio,unsigned int order,bool demote)1988 static bool __prep_compound_gigantic_folio(struct folio *folio,
1989 unsigned int order, bool demote)
1990 {
1991 int i, j;
1992 int nr_pages = 1 << order;
1993 struct page *p;
1994
1995 __folio_clear_reserved(folio);
1996 for (i = 0; i < nr_pages; i++) {
1997 p = folio_page(folio, i);
1998
1999 /*
2000 * For gigantic hugepages allocated through bootmem at
2001 * boot, it's safer to be consistent with the not-gigantic
2002 * hugepages and clear the PG_reserved bit from all tail pages
2003 * too. Otherwise drivers using get_user_pages() to access tail
2004 * pages may get the reference counting wrong if they see
2005 * PG_reserved set on a tail page (despite the head page not
2006 * having PG_reserved set). Enforcing this consistency between
2007 * head and tail pages allows drivers to optimize away a check
2008 * on the head page when they need know if put_page() is needed
2009 * after get_user_pages().
2010 */
2011 if (i != 0) /* head page cleared above */
2012 __ClearPageReserved(p);
2013 /*
2014 * Subtle and very unlikely
2015 *
2016 * Gigantic 'page allocators' such as memblock or cma will
2017 * return a set of pages with each page ref counted. We need
2018 * to turn this set of pages into a compound page with tail
2019 * page ref counts set to zero. Code such as speculative page
2020 * cache adding could take a ref on a 'to be' tail page.
2021 * We need to respect any increased ref count, and only set
2022 * the ref count to zero if count is currently 1. If count
2023 * is not 1, we return an error. An error return indicates
2024 * the set of pages can not be converted to a gigantic page.
2025 * The caller who allocated the pages should then discard the
2026 * pages using the appropriate free interface.
2027 *
2028 * In the case of demote, the ref count will be zero.
2029 */
2030 if (!demote) {
2031 if (!page_ref_freeze(p, 1)) {
2032 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2033 goto out_error;
2034 }
2035 } else {
2036 VM_BUG_ON_PAGE(page_count(p), p);
2037 }
2038 if (i != 0)
2039 set_compound_head(p, &folio->page);
2040 }
2041 __folio_set_head(folio);
2042 /* we rely on prep_new_hugetlb_folio to set the destructor */
2043 folio_set_order(folio, order);
2044 atomic_set(&folio->_entire_mapcount, -1);
2045 atomic_set(&folio->_nr_pages_mapped, 0);
2046 atomic_set(&folio->_pincount, 0);
2047 return true;
2048
2049 out_error:
2050 /* undo page modifications made above */
2051 for (j = 0; j < i; j++) {
2052 p = folio_page(folio, j);
2053 if (j != 0)
2054 clear_compound_head(p);
2055 set_page_refcounted(p);
2056 }
2057 /* need to clear PG_reserved on remaining tail pages */
2058 for (; j < nr_pages; j++) {
2059 p = folio_page(folio, j);
2060 __ClearPageReserved(p);
2061 }
2062 return false;
2063 }
2064
prep_compound_gigantic_folio(struct folio * folio,unsigned int order)2065 static bool prep_compound_gigantic_folio(struct folio *folio,
2066 unsigned int order)
2067 {
2068 return __prep_compound_gigantic_folio(folio, order, false);
2069 }
2070
prep_compound_gigantic_folio_for_demote(struct folio * folio,unsigned int order)2071 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2072 unsigned int order)
2073 {
2074 return __prep_compound_gigantic_folio(folio, order, true);
2075 }
2076
2077 /*
2078 * Find and lock address space (mapping) in write mode.
2079 *
2080 * Upon entry, the page is locked which means that page_mapping() is
2081 * stable. Due to locking order, we can only trylock_write. If we can
2082 * not get the lock, simply return NULL to caller.
2083 */
hugetlb_page_mapping_lock_write(struct page * hpage)2084 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2085 {
2086 struct address_space *mapping = page_mapping(hpage);
2087
2088 if (!mapping)
2089 return mapping;
2090
2091 if (i_mmap_trylock_write(mapping))
2092 return mapping;
2093
2094 return NULL;
2095 }
2096
hugetlb_basepage_index(struct page * page)2097 pgoff_t hugetlb_basepage_index(struct page *page)
2098 {
2099 struct page *page_head = compound_head(page);
2100 pgoff_t index = page_index(page_head);
2101 unsigned long compound_idx;
2102
2103 if (compound_order(page_head) > MAX_ORDER)
2104 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2105 else
2106 compound_idx = page - page_head;
2107
2108 return (index << compound_order(page_head)) + compound_idx;
2109 }
2110
alloc_buddy_hugetlb_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask,nodemask_t * node_alloc_noretry)2111 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2112 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2113 nodemask_t *node_alloc_noretry)
2114 {
2115 int order = huge_page_order(h);
2116 struct page *page;
2117 bool alloc_try_hard = true;
2118 bool retry = true;
2119
2120 /*
2121 * By default we always try hard to allocate the page with
2122 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2123 * a loop (to adjust global huge page counts) and previous allocation
2124 * failed, do not continue to try hard on the same node. Use the
2125 * node_alloc_noretry bitmap to manage this state information.
2126 */
2127 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2128 alloc_try_hard = false;
2129 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2130 if (alloc_try_hard)
2131 gfp_mask |= __GFP_RETRY_MAYFAIL;
2132 if (nid == NUMA_NO_NODE)
2133 nid = numa_mem_id();
2134 retry:
2135 page = __alloc_pages(gfp_mask, order, nid, nmask);
2136
2137 /* Freeze head page */
2138 if (page && !page_ref_freeze(page, 1)) {
2139 __free_pages(page, order);
2140 if (retry) { /* retry once */
2141 retry = false;
2142 goto retry;
2143 }
2144 /* WOW! twice in a row. */
2145 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2146 page = NULL;
2147 }
2148
2149 /*
2150 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2151 * indicates an overall state change. Clear bit so that we resume
2152 * normal 'try hard' allocations.
2153 */
2154 if (node_alloc_noretry && page && !alloc_try_hard)
2155 node_clear(nid, *node_alloc_noretry);
2156
2157 /*
2158 * If we tried hard to get a page but failed, set bit so that
2159 * subsequent attempts will not try as hard until there is an
2160 * overall state change.
2161 */
2162 if (node_alloc_noretry && !page && alloc_try_hard)
2163 node_set(nid, *node_alloc_noretry);
2164
2165 if (!page) {
2166 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2167 return NULL;
2168 }
2169
2170 __count_vm_event(HTLB_BUDDY_PGALLOC);
2171 return page_folio(page);
2172 }
2173
2174 /*
2175 * Common helper to allocate a fresh hugetlb page. All specific allocators
2176 * should use this function to get new hugetlb pages
2177 *
2178 * Note that returned page is 'frozen': ref count of head page and all tail
2179 * pages is zero.
2180 */
alloc_fresh_hugetlb_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask,nodemask_t * node_alloc_noretry)2181 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2182 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2183 nodemask_t *node_alloc_noretry)
2184 {
2185 struct folio *folio;
2186 bool retry = false;
2187
2188 retry:
2189 if (hstate_is_gigantic(h))
2190 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2191 else
2192 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2193 nid, nmask, node_alloc_noretry);
2194 if (!folio)
2195 return NULL;
2196 if (hstate_is_gigantic(h)) {
2197 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2198 /*
2199 * Rare failure to convert pages to compound page.
2200 * Free pages and try again - ONCE!
2201 */
2202 free_gigantic_folio(folio, huge_page_order(h));
2203 if (!retry) {
2204 retry = true;
2205 goto retry;
2206 }
2207 return NULL;
2208 }
2209 }
2210 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2211
2212 return folio;
2213 }
2214
2215 /*
2216 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2217 * manner.
2218 */
alloc_pool_huge_page(struct hstate * h,nodemask_t * nodes_allowed,nodemask_t * node_alloc_noretry)2219 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2220 nodemask_t *node_alloc_noretry)
2221 {
2222 struct folio *folio;
2223 int nr_nodes, node;
2224 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2225
2226 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2227 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2228 nodes_allowed, node_alloc_noretry);
2229 if (folio) {
2230 free_huge_folio(folio); /* free it into the hugepage allocator */
2231 return 1;
2232 }
2233 }
2234
2235 return 0;
2236 }
2237
2238 /*
2239 * Remove huge page from pool from next node to free. Attempt to keep
2240 * persistent huge pages more or less balanced over allowed nodes.
2241 * This routine only 'removes' the hugetlb page. The caller must make
2242 * an additional call to free the page to low level allocators.
2243 * Called with hugetlb_lock locked.
2244 */
remove_pool_huge_page(struct hstate * h,nodemask_t * nodes_allowed,bool acct_surplus)2245 static struct page *remove_pool_huge_page(struct hstate *h,
2246 nodemask_t *nodes_allowed,
2247 bool acct_surplus)
2248 {
2249 int nr_nodes, node;
2250 struct page *page = NULL;
2251 struct folio *folio;
2252
2253 lockdep_assert_held(&hugetlb_lock);
2254 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2255 /*
2256 * If we're returning unused surplus pages, only examine
2257 * nodes with surplus pages.
2258 */
2259 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2260 !list_empty(&h->hugepage_freelists[node])) {
2261 page = list_entry(h->hugepage_freelists[node].next,
2262 struct page, lru);
2263 folio = page_folio(page);
2264 remove_hugetlb_folio(h, folio, acct_surplus);
2265 break;
2266 }
2267 }
2268
2269 return page;
2270 }
2271
2272 /*
2273 * Dissolve a given free hugepage into free buddy pages. This function does
2274 * nothing for in-use hugepages and non-hugepages.
2275 * This function returns values like below:
2276 *
2277 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2278 * when the system is under memory pressure and the feature of
2279 * freeing unused vmemmap pages associated with each hugetlb page
2280 * is enabled.
2281 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2282 * (allocated or reserved.)
2283 * 0: successfully dissolved free hugepages or the page is not a
2284 * hugepage (considered as already dissolved)
2285 */
dissolve_free_huge_page(struct page * page)2286 int dissolve_free_huge_page(struct page *page)
2287 {
2288 int rc = -EBUSY;
2289 struct folio *folio = page_folio(page);
2290
2291 retry:
2292 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2293 if (!folio_test_hugetlb(folio))
2294 return 0;
2295
2296 spin_lock_irq(&hugetlb_lock);
2297 if (!folio_test_hugetlb(folio)) {
2298 rc = 0;
2299 goto out;
2300 }
2301
2302 if (!folio_ref_count(folio)) {
2303 struct hstate *h = folio_hstate(folio);
2304 if (!available_huge_pages(h))
2305 goto out;
2306
2307 /*
2308 * We should make sure that the page is already on the free list
2309 * when it is dissolved.
2310 */
2311 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2312 spin_unlock_irq(&hugetlb_lock);
2313 cond_resched();
2314
2315 /*
2316 * Theoretically, we should return -EBUSY when we
2317 * encounter this race. In fact, we have a chance
2318 * to successfully dissolve the page if we do a
2319 * retry. Because the race window is quite small.
2320 * If we seize this opportunity, it is an optimization
2321 * for increasing the success rate of dissolving page.
2322 */
2323 goto retry;
2324 }
2325
2326 remove_hugetlb_folio(h, folio, false);
2327 h->max_huge_pages--;
2328 spin_unlock_irq(&hugetlb_lock);
2329
2330 /*
2331 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2332 * before freeing the page. update_and_free_hugtlb_folio will fail to
2333 * free the page if it can not allocate required vmemmap. We
2334 * need to adjust max_huge_pages if the page is not freed.
2335 * Attempt to allocate vmemmmap here so that we can take
2336 * appropriate action on failure.
2337 */
2338 rc = hugetlb_vmemmap_restore(h, &folio->page);
2339 if (!rc) {
2340 update_and_free_hugetlb_folio(h, folio, false);
2341 } else {
2342 spin_lock_irq(&hugetlb_lock);
2343 add_hugetlb_folio(h, folio, false);
2344 h->max_huge_pages++;
2345 spin_unlock_irq(&hugetlb_lock);
2346 }
2347
2348 return rc;
2349 }
2350 out:
2351 spin_unlock_irq(&hugetlb_lock);
2352 return rc;
2353 }
2354
2355 /*
2356 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2357 * make specified memory blocks removable from the system.
2358 * Note that this will dissolve a free gigantic hugepage completely, if any
2359 * part of it lies within the given range.
2360 * Also note that if dissolve_free_huge_page() returns with an error, all
2361 * free hugepages that were dissolved before that error are lost.
2362 */
dissolve_free_huge_pages(unsigned long start_pfn,unsigned long end_pfn)2363 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2364 {
2365 unsigned long pfn;
2366 struct page *page;
2367 int rc = 0;
2368 unsigned int order;
2369 struct hstate *h;
2370
2371 if (!hugepages_supported())
2372 return rc;
2373
2374 order = huge_page_order(&default_hstate);
2375 for_each_hstate(h)
2376 order = min(order, huge_page_order(h));
2377
2378 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2379 page = pfn_to_page(pfn);
2380 rc = dissolve_free_huge_page(page);
2381 if (rc)
2382 break;
2383 }
2384
2385 return rc;
2386 }
2387
2388 /*
2389 * Allocates a fresh surplus page from the page allocator.
2390 */
alloc_surplus_hugetlb_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask)2391 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2392 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2393 {
2394 struct folio *folio = NULL;
2395
2396 if (hstate_is_gigantic(h))
2397 return NULL;
2398
2399 spin_lock_irq(&hugetlb_lock);
2400 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2401 goto out_unlock;
2402 spin_unlock_irq(&hugetlb_lock);
2403
2404 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2405 if (!folio)
2406 return NULL;
2407
2408 spin_lock_irq(&hugetlb_lock);
2409 /*
2410 * We could have raced with the pool size change.
2411 * Double check that and simply deallocate the new page
2412 * if we would end up overcommiting the surpluses. Abuse
2413 * temporary page to workaround the nasty free_huge_folio
2414 * codeflow
2415 */
2416 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2417 folio_set_hugetlb_temporary(folio);
2418 spin_unlock_irq(&hugetlb_lock);
2419 free_huge_folio(folio);
2420 return NULL;
2421 }
2422
2423 h->surplus_huge_pages++;
2424 h->surplus_huge_pages_node[folio_nid(folio)]++;
2425
2426 out_unlock:
2427 spin_unlock_irq(&hugetlb_lock);
2428
2429 return folio;
2430 }
2431
alloc_migrate_hugetlb_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask)2432 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2433 int nid, nodemask_t *nmask)
2434 {
2435 struct folio *folio;
2436
2437 if (hstate_is_gigantic(h))
2438 return NULL;
2439
2440 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2441 if (!folio)
2442 return NULL;
2443
2444 /* fresh huge pages are frozen */
2445 folio_ref_unfreeze(folio, 1);
2446 /*
2447 * We do not account these pages as surplus because they are only
2448 * temporary and will be released properly on the last reference
2449 */
2450 folio_set_hugetlb_temporary(folio);
2451
2452 return folio;
2453 }
2454
2455 /*
2456 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2457 */
2458 static
alloc_buddy_hugetlb_folio_with_mpol(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2459 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2460 struct vm_area_struct *vma, unsigned long addr)
2461 {
2462 struct folio *folio = NULL;
2463 struct mempolicy *mpol;
2464 gfp_t gfp_mask = htlb_alloc_mask(h);
2465 int nid;
2466 nodemask_t *nodemask;
2467
2468 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2469 if (mpol_is_preferred_many(mpol)) {
2470 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2471
2472 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2473 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2474
2475 /* Fallback to all nodes if page==NULL */
2476 nodemask = NULL;
2477 }
2478
2479 if (!folio)
2480 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2481 mpol_cond_put(mpol);
2482 return folio;
2483 }
2484
2485 /* folio migration callback function */
alloc_hugetlb_folio_nodemask(struct hstate * h,int preferred_nid,nodemask_t * nmask,gfp_t gfp_mask)2486 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2487 nodemask_t *nmask, gfp_t gfp_mask)
2488 {
2489 spin_lock_irq(&hugetlb_lock);
2490 if (available_huge_pages(h)) {
2491 struct folio *folio;
2492
2493 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2494 preferred_nid, nmask);
2495 if (folio) {
2496 spin_unlock_irq(&hugetlb_lock);
2497 return folio;
2498 }
2499 }
2500 spin_unlock_irq(&hugetlb_lock);
2501
2502 return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2503 }
2504
2505 /* mempolicy aware migration callback */
alloc_hugetlb_folio_vma(struct hstate * h,struct vm_area_struct * vma,unsigned long address)2506 struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma,
2507 unsigned long address)
2508 {
2509 struct mempolicy *mpol;
2510 nodemask_t *nodemask;
2511 struct folio *folio;
2512 gfp_t gfp_mask;
2513 int node;
2514
2515 gfp_mask = htlb_alloc_mask(h);
2516 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2517 folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
2518 mpol_cond_put(mpol);
2519
2520 return folio;
2521 }
2522
2523 /*
2524 * Increase the hugetlb pool such that it can accommodate a reservation
2525 * of size 'delta'.
2526 */
gather_surplus_pages(struct hstate * h,long delta)2527 static int gather_surplus_pages(struct hstate *h, long delta)
2528 __must_hold(&hugetlb_lock)
2529 {
2530 LIST_HEAD(surplus_list);
2531 struct folio *folio, *tmp;
2532 int ret;
2533 long i;
2534 long needed, allocated;
2535 bool alloc_ok = true;
2536
2537 lockdep_assert_held(&hugetlb_lock);
2538 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2539 if (needed <= 0) {
2540 h->resv_huge_pages += delta;
2541 return 0;
2542 }
2543
2544 allocated = 0;
2545
2546 ret = -ENOMEM;
2547 retry:
2548 spin_unlock_irq(&hugetlb_lock);
2549 for (i = 0; i < needed; i++) {
2550 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2551 NUMA_NO_NODE, NULL);
2552 if (!folio) {
2553 alloc_ok = false;
2554 break;
2555 }
2556 list_add(&folio->lru, &surplus_list);
2557 cond_resched();
2558 }
2559 allocated += i;
2560
2561 /*
2562 * After retaking hugetlb_lock, we need to recalculate 'needed'
2563 * because either resv_huge_pages or free_huge_pages may have changed.
2564 */
2565 spin_lock_irq(&hugetlb_lock);
2566 needed = (h->resv_huge_pages + delta) -
2567 (h->free_huge_pages + allocated);
2568 if (needed > 0) {
2569 if (alloc_ok)
2570 goto retry;
2571 /*
2572 * We were not able to allocate enough pages to
2573 * satisfy the entire reservation so we free what
2574 * we've allocated so far.
2575 */
2576 goto free;
2577 }
2578 /*
2579 * The surplus_list now contains _at_least_ the number of extra pages
2580 * needed to accommodate the reservation. Add the appropriate number
2581 * of pages to the hugetlb pool and free the extras back to the buddy
2582 * allocator. Commit the entire reservation here to prevent another
2583 * process from stealing the pages as they are added to the pool but
2584 * before they are reserved.
2585 */
2586 needed += allocated;
2587 h->resv_huge_pages += delta;
2588 ret = 0;
2589
2590 /* Free the needed pages to the hugetlb pool */
2591 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2592 if ((--needed) < 0)
2593 break;
2594 /* Add the page to the hugetlb allocator */
2595 enqueue_hugetlb_folio(h, folio);
2596 }
2597 free:
2598 spin_unlock_irq(&hugetlb_lock);
2599
2600 /*
2601 * Free unnecessary surplus pages to the buddy allocator.
2602 * Pages have no ref count, call free_huge_folio directly.
2603 */
2604 list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2605 free_huge_folio(folio);
2606 spin_lock_irq(&hugetlb_lock);
2607
2608 return ret;
2609 }
2610
2611 /*
2612 * This routine has two main purposes:
2613 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2614 * in unused_resv_pages. This corresponds to the prior adjustments made
2615 * to the associated reservation map.
2616 * 2) Free any unused surplus pages that may have been allocated to satisfy
2617 * the reservation. As many as unused_resv_pages may be freed.
2618 */
return_unused_surplus_pages(struct hstate * h,unsigned long unused_resv_pages)2619 static void return_unused_surplus_pages(struct hstate *h,
2620 unsigned long unused_resv_pages)
2621 {
2622 unsigned long nr_pages;
2623 struct page *page;
2624 LIST_HEAD(page_list);
2625
2626 lockdep_assert_held(&hugetlb_lock);
2627 /* Uncommit the reservation */
2628 h->resv_huge_pages -= unused_resv_pages;
2629
2630 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2631 goto out;
2632
2633 /*
2634 * Part (or even all) of the reservation could have been backed
2635 * by pre-allocated pages. Only free surplus pages.
2636 */
2637 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2638
2639 /*
2640 * We want to release as many surplus pages as possible, spread
2641 * evenly across all nodes with memory. Iterate across these nodes
2642 * until we can no longer free unreserved surplus pages. This occurs
2643 * when the nodes with surplus pages have no free pages.
2644 * remove_pool_huge_page() will balance the freed pages across the
2645 * on-line nodes with memory and will handle the hstate accounting.
2646 */
2647 while (nr_pages--) {
2648 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2649 if (!page)
2650 goto out;
2651
2652 list_add(&page->lru, &page_list);
2653 }
2654
2655 out:
2656 spin_unlock_irq(&hugetlb_lock);
2657 update_and_free_pages_bulk(h, &page_list);
2658 spin_lock_irq(&hugetlb_lock);
2659 }
2660
2661
2662 /*
2663 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2664 * are used by the huge page allocation routines to manage reservations.
2665 *
2666 * vma_needs_reservation is called to determine if the huge page at addr
2667 * within the vma has an associated reservation. If a reservation is
2668 * needed, the value 1 is returned. The caller is then responsible for
2669 * managing the global reservation and subpool usage counts. After
2670 * the huge page has been allocated, vma_commit_reservation is called
2671 * to add the page to the reservation map. If the page allocation fails,
2672 * the reservation must be ended instead of committed. vma_end_reservation
2673 * is called in such cases.
2674 *
2675 * In the normal case, vma_commit_reservation returns the same value
2676 * as the preceding vma_needs_reservation call. The only time this
2677 * is not the case is if a reserve map was changed between calls. It
2678 * is the responsibility of the caller to notice the difference and
2679 * take appropriate action.
2680 *
2681 * vma_add_reservation is used in error paths where a reservation must
2682 * be restored when a newly allocated huge page must be freed. It is
2683 * to be called after calling vma_needs_reservation to determine if a
2684 * reservation exists.
2685 *
2686 * vma_del_reservation is used in error paths where an entry in the reserve
2687 * map was created during huge page allocation and must be removed. It is to
2688 * be called after calling vma_needs_reservation to determine if a reservation
2689 * exists.
2690 */
2691 enum vma_resv_mode {
2692 VMA_NEEDS_RESV,
2693 VMA_COMMIT_RESV,
2694 VMA_END_RESV,
2695 VMA_ADD_RESV,
2696 VMA_DEL_RESV,
2697 };
__vma_reservation_common(struct hstate * h,struct vm_area_struct * vma,unsigned long addr,enum vma_resv_mode mode)2698 static long __vma_reservation_common(struct hstate *h,
2699 struct vm_area_struct *vma, unsigned long addr,
2700 enum vma_resv_mode mode)
2701 {
2702 struct resv_map *resv;
2703 pgoff_t idx;
2704 long ret;
2705 long dummy_out_regions_needed;
2706
2707 resv = vma_resv_map(vma);
2708 if (!resv)
2709 return 1;
2710
2711 idx = vma_hugecache_offset(h, vma, addr);
2712 switch (mode) {
2713 case VMA_NEEDS_RESV:
2714 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2715 /* We assume that vma_reservation_* routines always operate on
2716 * 1 page, and that adding to resv map a 1 page entry can only
2717 * ever require 1 region.
2718 */
2719 VM_BUG_ON(dummy_out_regions_needed != 1);
2720 break;
2721 case VMA_COMMIT_RESV:
2722 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2723 /* region_add calls of range 1 should never fail. */
2724 VM_BUG_ON(ret < 0);
2725 break;
2726 case VMA_END_RESV:
2727 region_abort(resv, idx, idx + 1, 1);
2728 ret = 0;
2729 break;
2730 case VMA_ADD_RESV:
2731 if (vma->vm_flags & VM_MAYSHARE) {
2732 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2733 /* region_add calls of range 1 should never fail. */
2734 VM_BUG_ON(ret < 0);
2735 } else {
2736 region_abort(resv, idx, idx + 1, 1);
2737 ret = region_del(resv, idx, idx + 1);
2738 }
2739 break;
2740 case VMA_DEL_RESV:
2741 if (vma->vm_flags & VM_MAYSHARE) {
2742 region_abort(resv, idx, idx + 1, 1);
2743 ret = region_del(resv, idx, idx + 1);
2744 } else {
2745 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2746 /* region_add calls of range 1 should never fail. */
2747 VM_BUG_ON(ret < 0);
2748 }
2749 break;
2750 default:
2751 BUG();
2752 }
2753
2754 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2755 return ret;
2756 /*
2757 * We know private mapping must have HPAGE_RESV_OWNER set.
2758 *
2759 * In most cases, reserves always exist for private mappings.
2760 * However, a file associated with mapping could have been
2761 * hole punched or truncated after reserves were consumed.
2762 * As subsequent fault on such a range will not use reserves.
2763 * Subtle - The reserve map for private mappings has the
2764 * opposite meaning than that of shared mappings. If NO
2765 * entry is in the reserve map, it means a reservation exists.
2766 * If an entry exists in the reserve map, it means the
2767 * reservation has already been consumed. As a result, the
2768 * return value of this routine is the opposite of the
2769 * value returned from reserve map manipulation routines above.
2770 */
2771 if (ret > 0)
2772 return 0;
2773 if (ret == 0)
2774 return 1;
2775 return ret;
2776 }
2777
vma_needs_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2778 static long vma_needs_reservation(struct hstate *h,
2779 struct vm_area_struct *vma, unsigned long addr)
2780 {
2781 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2782 }
2783
vma_commit_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2784 static long vma_commit_reservation(struct hstate *h,
2785 struct vm_area_struct *vma, unsigned long addr)
2786 {
2787 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2788 }
2789
vma_end_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2790 static void vma_end_reservation(struct hstate *h,
2791 struct vm_area_struct *vma, unsigned long addr)
2792 {
2793 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2794 }
2795
vma_add_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2796 static long vma_add_reservation(struct hstate *h,
2797 struct vm_area_struct *vma, unsigned long addr)
2798 {
2799 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2800 }
2801
vma_del_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2802 static long vma_del_reservation(struct hstate *h,
2803 struct vm_area_struct *vma, unsigned long addr)
2804 {
2805 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2806 }
2807
2808 /*
2809 * This routine is called to restore reservation information on error paths.
2810 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2811 * and the hugetlb mutex should remain held when calling this routine.
2812 *
2813 * It handles two specific cases:
2814 * 1) A reservation was in place and the folio consumed the reservation.
2815 * hugetlb_restore_reserve is set in the folio.
2816 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2817 * not set. However, alloc_hugetlb_folio always updates the reserve map.
2818 *
2819 * In case 1, free_huge_folio later in the error path will increment the
2820 * global reserve count. But, free_huge_folio does not have enough context
2821 * to adjust the reservation map. This case deals primarily with private
2822 * mappings. Adjust the reserve map here to be consistent with global
2823 * reserve count adjustments to be made by free_huge_folio. Make sure the
2824 * reserve map indicates there is a reservation present.
2825 *
2826 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2827 */
restore_reserve_on_error(struct hstate * h,struct vm_area_struct * vma,unsigned long address,struct folio * folio)2828 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2829 unsigned long address, struct folio *folio)
2830 {
2831 long rc = vma_needs_reservation(h, vma, address);
2832
2833 if (folio_test_hugetlb_restore_reserve(folio)) {
2834 if (unlikely(rc < 0))
2835 /*
2836 * Rare out of memory condition in reserve map
2837 * manipulation. Clear hugetlb_restore_reserve so
2838 * that global reserve count will not be incremented
2839 * by free_huge_folio. This will make it appear
2840 * as though the reservation for this folio was
2841 * consumed. This may prevent the task from
2842 * faulting in the folio at a later time. This
2843 * is better than inconsistent global huge page
2844 * accounting of reserve counts.
2845 */
2846 folio_clear_hugetlb_restore_reserve(folio);
2847 else if (rc)
2848 (void)vma_add_reservation(h, vma, address);
2849 else
2850 vma_end_reservation(h, vma, address);
2851 } else {
2852 if (!rc) {
2853 /*
2854 * This indicates there is an entry in the reserve map
2855 * not added by alloc_hugetlb_folio. We know it was added
2856 * before the alloc_hugetlb_folio call, otherwise
2857 * hugetlb_restore_reserve would be set on the folio.
2858 * Remove the entry so that a subsequent allocation
2859 * does not consume a reservation.
2860 */
2861 rc = vma_del_reservation(h, vma, address);
2862 if (rc < 0)
2863 /*
2864 * VERY rare out of memory condition. Since
2865 * we can not delete the entry, set
2866 * hugetlb_restore_reserve so that the reserve
2867 * count will be incremented when the folio
2868 * is freed. This reserve will be consumed
2869 * on a subsequent allocation.
2870 */
2871 folio_set_hugetlb_restore_reserve(folio);
2872 } else if (rc < 0) {
2873 /*
2874 * Rare out of memory condition from
2875 * vma_needs_reservation call. Memory allocation is
2876 * only attempted if a new entry is needed. Therefore,
2877 * this implies there is not an entry in the
2878 * reserve map.
2879 *
2880 * For shared mappings, no entry in the map indicates
2881 * no reservation. We are done.
2882 */
2883 if (!(vma->vm_flags & VM_MAYSHARE))
2884 /*
2885 * For private mappings, no entry indicates
2886 * a reservation is present. Since we can
2887 * not add an entry, set hugetlb_restore_reserve
2888 * on the folio so reserve count will be
2889 * incremented when freed. This reserve will
2890 * be consumed on a subsequent allocation.
2891 */
2892 folio_set_hugetlb_restore_reserve(folio);
2893 } else
2894 /*
2895 * No reservation present, do nothing
2896 */
2897 vma_end_reservation(h, vma, address);
2898 }
2899 }
2900
2901 /*
2902 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2903 * the old one
2904 * @h: struct hstate old page belongs to
2905 * @old_folio: Old folio to dissolve
2906 * @list: List to isolate the page in case we need to
2907 * Returns 0 on success, otherwise negated error.
2908 */
alloc_and_dissolve_hugetlb_folio(struct hstate * h,struct folio * old_folio,struct list_head * list)2909 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2910 struct folio *old_folio, struct list_head *list)
2911 {
2912 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2913 int nid = folio_nid(old_folio);
2914 struct folio *new_folio;
2915 int ret = 0;
2916
2917 /*
2918 * Before dissolving the folio, we need to allocate a new one for the
2919 * pool to remain stable. Here, we allocate the folio and 'prep' it
2920 * by doing everything but actually updating counters and adding to
2921 * the pool. This simplifies and let us do most of the processing
2922 * under the lock.
2923 */
2924 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2925 if (!new_folio)
2926 return -ENOMEM;
2927 __prep_new_hugetlb_folio(h, new_folio);
2928
2929 retry:
2930 spin_lock_irq(&hugetlb_lock);
2931 if (!folio_test_hugetlb(old_folio)) {
2932 /*
2933 * Freed from under us. Drop new_folio too.
2934 */
2935 goto free_new;
2936 } else if (folio_ref_count(old_folio)) {
2937 bool isolated;
2938
2939 /*
2940 * Someone has grabbed the folio, try to isolate it here.
2941 * Fail with -EBUSY if not possible.
2942 */
2943 spin_unlock_irq(&hugetlb_lock);
2944 isolated = isolate_hugetlb(old_folio, list);
2945 ret = isolated ? 0 : -EBUSY;
2946 spin_lock_irq(&hugetlb_lock);
2947 goto free_new;
2948 } else if (!folio_test_hugetlb_freed(old_folio)) {
2949 /*
2950 * Folio's refcount is 0 but it has not been enqueued in the
2951 * freelist yet. Race window is small, so we can succeed here if
2952 * we retry.
2953 */
2954 spin_unlock_irq(&hugetlb_lock);
2955 cond_resched();
2956 goto retry;
2957 } else {
2958 /*
2959 * Ok, old_folio is still a genuine free hugepage. Remove it from
2960 * the freelist and decrease the counters. These will be
2961 * incremented again when calling __prep_account_new_huge_page()
2962 * and enqueue_hugetlb_folio() for new_folio. The counters will
2963 * remain stable since this happens under the lock.
2964 */
2965 remove_hugetlb_folio(h, old_folio, false);
2966
2967 /*
2968 * Ref count on new_folio is already zero as it was dropped
2969 * earlier. It can be directly added to the pool free list.
2970 */
2971 __prep_account_new_huge_page(h, nid);
2972 enqueue_hugetlb_folio(h, new_folio);
2973
2974 /*
2975 * Folio has been replaced, we can safely free the old one.
2976 */
2977 spin_unlock_irq(&hugetlb_lock);
2978 update_and_free_hugetlb_folio(h, old_folio, false);
2979 }
2980
2981 return ret;
2982
2983 free_new:
2984 spin_unlock_irq(&hugetlb_lock);
2985 /* Folio has a zero ref count, but needs a ref to be freed */
2986 folio_ref_unfreeze(new_folio, 1);
2987 update_and_free_hugetlb_folio(h, new_folio, false);
2988
2989 return ret;
2990 }
2991
isolate_or_dissolve_huge_page(struct page * page,struct list_head * list)2992 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2993 {
2994 struct hstate *h;
2995 struct folio *folio = page_folio(page);
2996 int ret = -EBUSY;
2997
2998 /*
2999 * The page might have been dissolved from under our feet, so make sure
3000 * to carefully check the state under the lock.
3001 * Return success when racing as if we dissolved the page ourselves.
3002 */
3003 spin_lock_irq(&hugetlb_lock);
3004 if (folio_test_hugetlb(folio)) {
3005 h = folio_hstate(folio);
3006 } else {
3007 spin_unlock_irq(&hugetlb_lock);
3008 return 0;
3009 }
3010 spin_unlock_irq(&hugetlb_lock);
3011
3012 /*
3013 * Fence off gigantic pages as there is a cyclic dependency between
3014 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3015 * of bailing out right away without further retrying.
3016 */
3017 if (hstate_is_gigantic(h))
3018 return -ENOMEM;
3019
3020 if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
3021 ret = 0;
3022 else if (!folio_ref_count(folio))
3023 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3024
3025 return ret;
3026 }
3027
alloc_hugetlb_folio(struct vm_area_struct * vma,unsigned long addr,int avoid_reserve)3028 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3029 unsigned long addr, int avoid_reserve)
3030 {
3031 struct hugepage_subpool *spool = subpool_vma(vma);
3032 struct hstate *h = hstate_vma(vma);
3033 struct folio *folio;
3034 long map_chg, map_commit;
3035 long gbl_chg;
3036 int ret, idx;
3037 struct hugetlb_cgroup *h_cg = NULL;
3038 bool deferred_reserve;
3039
3040 idx = hstate_index(h);
3041 /*
3042 * Examine the region/reserve map to determine if the process
3043 * has a reservation for the page to be allocated. A return
3044 * code of zero indicates a reservation exists (no change).
3045 */
3046 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3047 if (map_chg < 0)
3048 return ERR_PTR(-ENOMEM);
3049
3050 /*
3051 * Processes that did not create the mapping will have no
3052 * reserves as indicated by the region/reserve map. Check
3053 * that the allocation will not exceed the subpool limit.
3054 * Allocations for MAP_NORESERVE mappings also need to be
3055 * checked against any subpool limit.
3056 */
3057 if (map_chg || avoid_reserve) {
3058 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3059 if (gbl_chg < 0) {
3060 vma_end_reservation(h, vma, addr);
3061 return ERR_PTR(-ENOSPC);
3062 }
3063
3064 /*
3065 * Even though there was no reservation in the region/reserve
3066 * map, there could be reservations associated with the
3067 * subpool that can be used. This would be indicated if the
3068 * return value of hugepage_subpool_get_pages() is zero.
3069 * However, if avoid_reserve is specified we still avoid even
3070 * the subpool reservations.
3071 */
3072 if (avoid_reserve)
3073 gbl_chg = 1;
3074 }
3075
3076 /* If this allocation is not consuming a reservation, charge it now.
3077 */
3078 deferred_reserve = map_chg || avoid_reserve;
3079 if (deferred_reserve) {
3080 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3081 idx, pages_per_huge_page(h), &h_cg);
3082 if (ret)
3083 goto out_subpool_put;
3084 }
3085
3086 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3087 if (ret)
3088 goto out_uncharge_cgroup_reservation;
3089
3090 spin_lock_irq(&hugetlb_lock);
3091 /*
3092 * glb_chg is passed to indicate whether or not a page must be taken
3093 * from the global free pool (global change). gbl_chg == 0 indicates
3094 * a reservation exists for the allocation.
3095 */
3096 folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3097 if (!folio) {
3098 spin_unlock_irq(&hugetlb_lock);
3099 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3100 if (!folio)
3101 goto out_uncharge_cgroup;
3102 spin_lock_irq(&hugetlb_lock);
3103 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3104 folio_set_hugetlb_restore_reserve(folio);
3105 h->resv_huge_pages--;
3106 }
3107 list_add(&folio->lru, &h->hugepage_activelist);
3108 folio_ref_unfreeze(folio, 1);
3109 /* Fall through */
3110 }
3111
3112 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3113 /* If allocation is not consuming a reservation, also store the
3114 * hugetlb_cgroup pointer on the page.
3115 */
3116 if (deferred_reserve) {
3117 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3118 h_cg, folio);
3119 }
3120
3121 spin_unlock_irq(&hugetlb_lock);
3122
3123 hugetlb_set_folio_subpool(folio, spool);
3124
3125 map_commit = vma_commit_reservation(h, vma, addr);
3126 if (unlikely(map_chg > map_commit)) {
3127 /*
3128 * The page was added to the reservation map between
3129 * vma_needs_reservation and vma_commit_reservation.
3130 * This indicates a race with hugetlb_reserve_pages.
3131 * Adjust for the subpool count incremented above AND
3132 * in hugetlb_reserve_pages for the same page. Also,
3133 * the reservation count added in hugetlb_reserve_pages
3134 * no longer applies.
3135 */
3136 long rsv_adjust;
3137
3138 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3139 hugetlb_acct_memory(h, -rsv_adjust);
3140 if (deferred_reserve) {
3141 spin_lock_irq(&hugetlb_lock);
3142 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3143 pages_per_huge_page(h), folio);
3144 spin_unlock_irq(&hugetlb_lock);
3145 }
3146 }
3147 return folio;
3148
3149 out_uncharge_cgroup:
3150 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3151 out_uncharge_cgroup_reservation:
3152 if (deferred_reserve)
3153 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3154 h_cg);
3155 out_subpool_put:
3156 if (map_chg || avoid_reserve)
3157 hugepage_subpool_put_pages(spool, 1);
3158 vma_end_reservation(h, vma, addr);
3159 return ERR_PTR(-ENOSPC);
3160 }
3161
3162 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3163 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
__alloc_bootmem_huge_page(struct hstate * h,int nid)3164 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3165 {
3166 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3167 int nr_nodes, node;
3168
3169 /* do node specific alloc */
3170 if (nid != NUMA_NO_NODE) {
3171 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3172 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3173 if (!m)
3174 return 0;
3175 goto found;
3176 }
3177 /* allocate from next node when distributing huge pages */
3178 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3179 m = memblock_alloc_try_nid_raw(
3180 huge_page_size(h), huge_page_size(h),
3181 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3182 /*
3183 * Use the beginning of the huge page to store the
3184 * huge_bootmem_page struct (until gather_bootmem
3185 * puts them into the mem_map).
3186 */
3187 if (!m)
3188 return 0;
3189 goto found;
3190 }
3191
3192 found:
3193 /* Put them into a private list first because mem_map is not up yet */
3194 INIT_LIST_HEAD(&m->list);
3195 list_add(&m->list, &huge_boot_pages);
3196 m->hstate = h;
3197 return 1;
3198 }
3199
3200 /*
3201 * Put bootmem huge pages into the standard lists after mem_map is up.
3202 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3203 */
gather_bootmem_prealloc(void)3204 static void __init gather_bootmem_prealloc(void)
3205 {
3206 struct huge_bootmem_page *m;
3207
3208 list_for_each_entry(m, &huge_boot_pages, list) {
3209 struct page *page = virt_to_page(m);
3210 struct folio *folio = page_folio(page);
3211 struct hstate *h = m->hstate;
3212
3213 VM_BUG_ON(!hstate_is_gigantic(h));
3214 WARN_ON(folio_ref_count(folio) != 1);
3215 if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3216 WARN_ON(folio_test_reserved(folio));
3217 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3218 free_huge_folio(folio); /* add to the hugepage allocator */
3219 } else {
3220 /* VERY unlikely inflated ref count on a tail page */
3221 free_gigantic_folio(folio, huge_page_order(h));
3222 }
3223
3224 /*
3225 * We need to restore the 'stolen' pages to totalram_pages
3226 * in order to fix confusing memory reports from free(1) and
3227 * other side-effects, like CommitLimit going negative.
3228 */
3229 adjust_managed_page_count(page, pages_per_huge_page(h));
3230 cond_resched();
3231 }
3232 }
hugetlb_hstate_alloc_pages_onenode(struct hstate * h,int nid)3233 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3234 {
3235 unsigned long i;
3236 char buf[32];
3237
3238 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3239 if (hstate_is_gigantic(h)) {
3240 if (!alloc_bootmem_huge_page(h, nid))
3241 break;
3242 } else {
3243 struct folio *folio;
3244 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3245
3246 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3247 &node_states[N_MEMORY], NULL);
3248 if (!folio)
3249 break;
3250 free_huge_folio(folio); /* free it into the hugepage allocator */
3251 }
3252 cond_resched();
3253 }
3254 if (i == h->max_huge_pages_node[nid])
3255 return;
3256
3257 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3258 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3259 h->max_huge_pages_node[nid], buf, nid, i);
3260 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3261 h->max_huge_pages_node[nid] = i;
3262 }
3263
hugetlb_hstate_alloc_pages(struct hstate * h)3264 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3265 {
3266 unsigned long i;
3267 nodemask_t *node_alloc_noretry;
3268 bool node_specific_alloc = false;
3269
3270 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3271 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3272 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3273 return;
3274 }
3275
3276 /* do node specific alloc */
3277 for_each_online_node(i) {
3278 if (h->max_huge_pages_node[i] > 0) {
3279 hugetlb_hstate_alloc_pages_onenode(h, i);
3280 node_specific_alloc = true;
3281 }
3282 }
3283
3284 if (node_specific_alloc)
3285 return;
3286
3287 /* below will do all node balanced alloc */
3288 if (!hstate_is_gigantic(h)) {
3289 /*
3290 * Bit mask controlling how hard we retry per-node allocations.
3291 * Ignore errors as lower level routines can deal with
3292 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3293 * time, we are likely in bigger trouble.
3294 */
3295 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3296 GFP_KERNEL);
3297 } else {
3298 /* allocations done at boot time */
3299 node_alloc_noretry = NULL;
3300 }
3301
3302 /* bit mask controlling how hard we retry per-node allocations */
3303 if (node_alloc_noretry)
3304 nodes_clear(*node_alloc_noretry);
3305
3306 for (i = 0; i < h->max_huge_pages; ++i) {
3307 if (hstate_is_gigantic(h)) {
3308 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3309 break;
3310 } else if (!alloc_pool_huge_page(h,
3311 &node_states[N_MEMORY],
3312 node_alloc_noretry))
3313 break;
3314 cond_resched();
3315 }
3316 if (i < h->max_huge_pages) {
3317 char buf[32];
3318
3319 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3320 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3321 h->max_huge_pages, buf, i);
3322 h->max_huge_pages = i;
3323 }
3324 kfree(node_alloc_noretry);
3325 }
3326
hugetlb_init_hstates(void)3327 static void __init hugetlb_init_hstates(void)
3328 {
3329 struct hstate *h, *h2;
3330
3331 for_each_hstate(h) {
3332 /* oversize hugepages were init'ed in early boot */
3333 if (!hstate_is_gigantic(h))
3334 hugetlb_hstate_alloc_pages(h);
3335
3336 /*
3337 * Set demote order for each hstate. Note that
3338 * h->demote_order is initially 0.
3339 * - We can not demote gigantic pages if runtime freeing
3340 * is not supported, so skip this.
3341 * - If CMA allocation is possible, we can not demote
3342 * HUGETLB_PAGE_ORDER or smaller size pages.
3343 */
3344 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3345 continue;
3346 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3347 continue;
3348 for_each_hstate(h2) {
3349 if (h2 == h)
3350 continue;
3351 if (h2->order < h->order &&
3352 h2->order > h->demote_order)
3353 h->demote_order = h2->order;
3354 }
3355 }
3356 }
3357
report_hugepages(void)3358 static void __init report_hugepages(void)
3359 {
3360 struct hstate *h;
3361
3362 for_each_hstate(h) {
3363 char buf[32];
3364
3365 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3366 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3367 buf, h->free_huge_pages);
3368 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3369 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3370 }
3371 }
3372
3373 #ifdef CONFIG_HIGHMEM
try_to_free_low(struct hstate * h,unsigned long count,nodemask_t * nodes_allowed)3374 static void try_to_free_low(struct hstate *h, unsigned long count,
3375 nodemask_t *nodes_allowed)
3376 {
3377 int i;
3378 LIST_HEAD(page_list);
3379
3380 lockdep_assert_held(&hugetlb_lock);
3381 if (hstate_is_gigantic(h))
3382 return;
3383
3384 /*
3385 * Collect pages to be freed on a list, and free after dropping lock
3386 */
3387 for_each_node_mask(i, *nodes_allowed) {
3388 struct page *page, *next;
3389 struct list_head *freel = &h->hugepage_freelists[i];
3390 list_for_each_entry_safe(page, next, freel, lru) {
3391 if (count >= h->nr_huge_pages)
3392 goto out;
3393 if (PageHighMem(page))
3394 continue;
3395 remove_hugetlb_folio(h, page_folio(page), false);
3396 list_add(&page->lru, &page_list);
3397 }
3398 }
3399
3400 out:
3401 spin_unlock_irq(&hugetlb_lock);
3402 update_and_free_pages_bulk(h, &page_list);
3403 spin_lock_irq(&hugetlb_lock);
3404 }
3405 #else
try_to_free_low(struct hstate * h,unsigned long count,nodemask_t * nodes_allowed)3406 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3407 nodemask_t *nodes_allowed)
3408 {
3409 }
3410 #endif
3411
3412 /*
3413 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3414 * balanced by operating on them in a round-robin fashion.
3415 * Returns 1 if an adjustment was made.
3416 */
adjust_pool_surplus(struct hstate * h,nodemask_t * nodes_allowed,int delta)3417 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3418 int delta)
3419 {
3420 int nr_nodes, node;
3421
3422 lockdep_assert_held(&hugetlb_lock);
3423 VM_BUG_ON(delta != -1 && delta != 1);
3424
3425 if (delta < 0) {
3426 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3427 if (h->surplus_huge_pages_node[node])
3428 goto found;
3429 }
3430 } else {
3431 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3432 if (h->surplus_huge_pages_node[node] <
3433 h->nr_huge_pages_node[node])
3434 goto found;
3435 }
3436 }
3437 return 0;
3438
3439 found:
3440 h->surplus_huge_pages += delta;
3441 h->surplus_huge_pages_node[node] += delta;
3442 return 1;
3443 }
3444
3445 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
set_max_huge_pages(struct hstate * h,unsigned long count,int nid,nodemask_t * nodes_allowed)3446 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3447 nodemask_t *nodes_allowed)
3448 {
3449 unsigned long min_count, ret;
3450 struct page *page;
3451 LIST_HEAD(page_list);
3452 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3453
3454 /*
3455 * Bit mask controlling how hard we retry per-node allocations.
3456 * If we can not allocate the bit mask, do not attempt to allocate
3457 * the requested huge pages.
3458 */
3459 if (node_alloc_noretry)
3460 nodes_clear(*node_alloc_noretry);
3461 else
3462 return -ENOMEM;
3463
3464 /*
3465 * resize_lock mutex prevents concurrent adjustments to number of
3466 * pages in hstate via the proc/sysfs interfaces.
3467 */
3468 mutex_lock(&h->resize_lock);
3469 flush_free_hpage_work(h);
3470 spin_lock_irq(&hugetlb_lock);
3471
3472 /*
3473 * Check for a node specific request.
3474 * Changing node specific huge page count may require a corresponding
3475 * change to the global count. In any case, the passed node mask
3476 * (nodes_allowed) will restrict alloc/free to the specified node.
3477 */
3478 if (nid != NUMA_NO_NODE) {
3479 unsigned long old_count = count;
3480
3481 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3482 /*
3483 * User may have specified a large count value which caused the
3484 * above calculation to overflow. In this case, they wanted
3485 * to allocate as many huge pages as possible. Set count to
3486 * largest possible value to align with their intention.
3487 */
3488 if (count < old_count)
3489 count = ULONG_MAX;
3490 }
3491
3492 /*
3493 * Gigantic pages runtime allocation depend on the capability for large
3494 * page range allocation.
3495 * If the system does not provide this feature, return an error when
3496 * the user tries to allocate gigantic pages but let the user free the
3497 * boottime allocated gigantic pages.
3498 */
3499 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3500 if (count > persistent_huge_pages(h)) {
3501 spin_unlock_irq(&hugetlb_lock);
3502 mutex_unlock(&h->resize_lock);
3503 NODEMASK_FREE(node_alloc_noretry);
3504 return -EINVAL;
3505 }
3506 /* Fall through to decrease pool */
3507 }
3508
3509 /*
3510 * Increase the pool size
3511 * First take pages out of surplus state. Then make up the
3512 * remaining difference by allocating fresh huge pages.
3513 *
3514 * We might race with alloc_surplus_hugetlb_folio() here and be unable
3515 * to convert a surplus huge page to a normal huge page. That is
3516 * not critical, though, it just means the overall size of the
3517 * pool might be one hugepage larger than it needs to be, but
3518 * within all the constraints specified by the sysctls.
3519 */
3520 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3521 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3522 break;
3523 }
3524
3525 while (count > persistent_huge_pages(h)) {
3526 /*
3527 * If this allocation races such that we no longer need the
3528 * page, free_huge_folio will handle it by freeing the page
3529 * and reducing the surplus.
3530 */
3531 spin_unlock_irq(&hugetlb_lock);
3532
3533 /* yield cpu to avoid soft lockup */
3534 cond_resched();
3535
3536 ret = alloc_pool_huge_page(h, nodes_allowed,
3537 node_alloc_noretry);
3538 spin_lock_irq(&hugetlb_lock);
3539 if (!ret)
3540 goto out;
3541
3542 /* Bail for signals. Probably ctrl-c from user */
3543 if (signal_pending(current))
3544 goto out;
3545 }
3546
3547 /*
3548 * Decrease the pool size
3549 * First return free pages to the buddy allocator (being careful
3550 * to keep enough around to satisfy reservations). Then place
3551 * pages into surplus state as needed so the pool will shrink
3552 * to the desired size as pages become free.
3553 *
3554 * By placing pages into the surplus state independent of the
3555 * overcommit value, we are allowing the surplus pool size to
3556 * exceed overcommit. There are few sane options here. Since
3557 * alloc_surplus_hugetlb_folio() is checking the global counter,
3558 * though, we'll note that we're not allowed to exceed surplus
3559 * and won't grow the pool anywhere else. Not until one of the
3560 * sysctls are changed, or the surplus pages go out of use.
3561 */
3562 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3563 min_count = max(count, min_count);
3564 try_to_free_low(h, min_count, nodes_allowed);
3565
3566 /*
3567 * Collect pages to be removed on list without dropping lock
3568 */
3569 while (min_count < persistent_huge_pages(h)) {
3570 page = remove_pool_huge_page(h, nodes_allowed, 0);
3571 if (!page)
3572 break;
3573
3574 list_add(&page->lru, &page_list);
3575 }
3576 /* free the pages after dropping lock */
3577 spin_unlock_irq(&hugetlb_lock);
3578 update_and_free_pages_bulk(h, &page_list);
3579 flush_free_hpage_work(h);
3580 spin_lock_irq(&hugetlb_lock);
3581
3582 while (count < persistent_huge_pages(h)) {
3583 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3584 break;
3585 }
3586 out:
3587 h->max_huge_pages = persistent_huge_pages(h);
3588 spin_unlock_irq(&hugetlb_lock);
3589 mutex_unlock(&h->resize_lock);
3590
3591 NODEMASK_FREE(node_alloc_noretry);
3592
3593 return 0;
3594 }
3595
demote_free_hugetlb_folio(struct hstate * h,struct folio * folio)3596 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3597 {
3598 int i, nid = folio_nid(folio);
3599 struct hstate *target_hstate;
3600 struct page *subpage;
3601 struct folio *inner_folio;
3602 int rc = 0;
3603
3604 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3605
3606 remove_hugetlb_folio_for_demote(h, folio, false);
3607 spin_unlock_irq(&hugetlb_lock);
3608
3609 rc = hugetlb_vmemmap_restore(h, &folio->page);
3610 if (rc) {
3611 /* Allocation of vmemmmap failed, we can not demote folio */
3612 spin_lock_irq(&hugetlb_lock);
3613 folio_ref_unfreeze(folio, 1);
3614 add_hugetlb_folio(h, folio, false);
3615 return rc;
3616 }
3617
3618 /*
3619 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3620 * sizes as it will not ref count folios.
3621 */
3622 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3623
3624 /*
3625 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3626 * Without the mutex, pages added to target hstate could be marked
3627 * as surplus.
3628 *
3629 * Note that we already hold h->resize_lock. To prevent deadlock,
3630 * use the convention of always taking larger size hstate mutex first.
3631 */
3632 mutex_lock(&target_hstate->resize_lock);
3633 for (i = 0; i < pages_per_huge_page(h);
3634 i += pages_per_huge_page(target_hstate)) {
3635 subpage = folio_page(folio, i);
3636 inner_folio = page_folio(subpage);
3637 if (hstate_is_gigantic(target_hstate))
3638 prep_compound_gigantic_folio_for_demote(inner_folio,
3639 target_hstate->order);
3640 else
3641 prep_compound_page(subpage, target_hstate->order);
3642 folio_change_private(inner_folio, NULL);
3643 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3644 free_huge_folio(inner_folio);
3645 }
3646 mutex_unlock(&target_hstate->resize_lock);
3647
3648 spin_lock_irq(&hugetlb_lock);
3649
3650 /*
3651 * Not absolutely necessary, but for consistency update max_huge_pages
3652 * based on pool changes for the demoted page.
3653 */
3654 h->max_huge_pages--;
3655 target_hstate->max_huge_pages +=
3656 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3657
3658 return rc;
3659 }
3660
demote_pool_huge_page(struct hstate * h,nodemask_t * nodes_allowed)3661 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3662 __must_hold(&hugetlb_lock)
3663 {
3664 int nr_nodes, node;
3665 struct folio *folio;
3666
3667 lockdep_assert_held(&hugetlb_lock);
3668
3669 /* We should never get here if no demote order */
3670 if (!h->demote_order) {
3671 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3672 return -EINVAL; /* internal error */
3673 }
3674
3675 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3676 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
3677 if (folio_test_hwpoison(folio))
3678 continue;
3679 return demote_free_hugetlb_folio(h, folio);
3680 }
3681 }
3682
3683 /*
3684 * Only way to get here is if all pages on free lists are poisoned.
3685 * Return -EBUSY so that caller will not retry.
3686 */
3687 return -EBUSY;
3688 }
3689
3690 #define HSTATE_ATTR_RO(_name) \
3691 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3692
3693 #define HSTATE_ATTR_WO(_name) \
3694 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3695
3696 #define HSTATE_ATTR(_name) \
3697 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3698
3699 static struct kobject *hugepages_kobj;
3700 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3701
3702 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3703
kobj_to_hstate(struct kobject * kobj,int * nidp)3704 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3705 {
3706 int i;
3707
3708 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3709 if (hstate_kobjs[i] == kobj) {
3710 if (nidp)
3711 *nidp = NUMA_NO_NODE;
3712 return &hstates[i];
3713 }
3714
3715 return kobj_to_node_hstate(kobj, nidp);
3716 }
3717
nr_hugepages_show_common(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3718 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3719 struct kobj_attribute *attr, char *buf)
3720 {
3721 struct hstate *h;
3722 unsigned long nr_huge_pages;
3723 int nid;
3724
3725 h = kobj_to_hstate(kobj, &nid);
3726 if (nid == NUMA_NO_NODE)
3727 nr_huge_pages = h->nr_huge_pages;
3728 else
3729 nr_huge_pages = h->nr_huge_pages_node[nid];
3730
3731 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3732 }
3733
__nr_hugepages_store_common(bool obey_mempolicy,struct hstate * h,int nid,unsigned long count,size_t len)3734 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3735 struct hstate *h, int nid,
3736 unsigned long count, size_t len)
3737 {
3738 int err;
3739 nodemask_t nodes_allowed, *n_mask;
3740
3741 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3742 return -EINVAL;
3743
3744 if (nid == NUMA_NO_NODE) {
3745 /*
3746 * global hstate attribute
3747 */
3748 if (!(obey_mempolicy &&
3749 init_nodemask_of_mempolicy(&nodes_allowed)))
3750 n_mask = &node_states[N_MEMORY];
3751 else
3752 n_mask = &nodes_allowed;
3753 } else {
3754 /*
3755 * Node specific request. count adjustment happens in
3756 * set_max_huge_pages() after acquiring hugetlb_lock.
3757 */
3758 init_nodemask_of_node(&nodes_allowed, nid);
3759 n_mask = &nodes_allowed;
3760 }
3761
3762 err = set_max_huge_pages(h, count, nid, n_mask);
3763
3764 return err ? err : len;
3765 }
3766
nr_hugepages_store_common(bool obey_mempolicy,struct kobject * kobj,const char * buf,size_t len)3767 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3768 struct kobject *kobj, const char *buf,
3769 size_t len)
3770 {
3771 struct hstate *h;
3772 unsigned long count;
3773 int nid;
3774 int err;
3775
3776 err = kstrtoul(buf, 10, &count);
3777 if (err)
3778 return err;
3779
3780 h = kobj_to_hstate(kobj, &nid);
3781 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3782 }
3783
nr_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3784 static ssize_t nr_hugepages_show(struct kobject *kobj,
3785 struct kobj_attribute *attr, char *buf)
3786 {
3787 return nr_hugepages_show_common(kobj, attr, buf);
3788 }
3789
nr_hugepages_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)3790 static ssize_t nr_hugepages_store(struct kobject *kobj,
3791 struct kobj_attribute *attr, const char *buf, size_t len)
3792 {
3793 return nr_hugepages_store_common(false, kobj, buf, len);
3794 }
3795 HSTATE_ATTR(nr_hugepages);
3796
3797 #ifdef CONFIG_NUMA
3798
3799 /*
3800 * hstate attribute for optionally mempolicy-based constraint on persistent
3801 * huge page alloc/free.
3802 */
nr_hugepages_mempolicy_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3803 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3804 struct kobj_attribute *attr,
3805 char *buf)
3806 {
3807 return nr_hugepages_show_common(kobj, attr, buf);
3808 }
3809
nr_hugepages_mempolicy_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)3810 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3811 struct kobj_attribute *attr, const char *buf, size_t len)
3812 {
3813 return nr_hugepages_store_common(true, kobj, buf, len);
3814 }
3815 HSTATE_ATTR(nr_hugepages_mempolicy);
3816 #endif
3817
3818
nr_overcommit_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3819 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3820 struct kobj_attribute *attr, char *buf)
3821 {
3822 struct hstate *h = kobj_to_hstate(kobj, NULL);
3823 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3824 }
3825
nr_overcommit_hugepages_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)3826 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3827 struct kobj_attribute *attr, const char *buf, size_t count)
3828 {
3829 int err;
3830 unsigned long input;
3831 struct hstate *h = kobj_to_hstate(kobj, NULL);
3832
3833 if (hstate_is_gigantic(h))
3834 return -EINVAL;
3835
3836 err = kstrtoul(buf, 10, &input);
3837 if (err)
3838 return err;
3839
3840 spin_lock_irq(&hugetlb_lock);
3841 h->nr_overcommit_huge_pages = input;
3842 spin_unlock_irq(&hugetlb_lock);
3843
3844 return count;
3845 }
3846 HSTATE_ATTR(nr_overcommit_hugepages);
3847
free_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3848 static ssize_t free_hugepages_show(struct kobject *kobj,
3849 struct kobj_attribute *attr, char *buf)
3850 {
3851 struct hstate *h;
3852 unsigned long free_huge_pages;
3853 int nid;
3854
3855 h = kobj_to_hstate(kobj, &nid);
3856 if (nid == NUMA_NO_NODE)
3857 free_huge_pages = h->free_huge_pages;
3858 else
3859 free_huge_pages = h->free_huge_pages_node[nid];
3860
3861 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3862 }
3863 HSTATE_ATTR_RO(free_hugepages);
3864
resv_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3865 static ssize_t resv_hugepages_show(struct kobject *kobj,
3866 struct kobj_attribute *attr, char *buf)
3867 {
3868 struct hstate *h = kobj_to_hstate(kobj, NULL);
3869 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3870 }
3871 HSTATE_ATTR_RO(resv_hugepages);
3872
surplus_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3873 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3874 struct kobj_attribute *attr, char *buf)
3875 {
3876 struct hstate *h;
3877 unsigned long surplus_huge_pages;
3878 int nid;
3879
3880 h = kobj_to_hstate(kobj, &nid);
3881 if (nid == NUMA_NO_NODE)
3882 surplus_huge_pages = h->surplus_huge_pages;
3883 else
3884 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3885
3886 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3887 }
3888 HSTATE_ATTR_RO(surplus_hugepages);
3889
demote_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)3890 static ssize_t demote_store(struct kobject *kobj,
3891 struct kobj_attribute *attr, const char *buf, size_t len)
3892 {
3893 unsigned long nr_demote;
3894 unsigned long nr_available;
3895 nodemask_t nodes_allowed, *n_mask;
3896 struct hstate *h;
3897 int err;
3898 int nid;
3899
3900 err = kstrtoul(buf, 10, &nr_demote);
3901 if (err)
3902 return err;
3903 h = kobj_to_hstate(kobj, &nid);
3904
3905 if (nid != NUMA_NO_NODE) {
3906 init_nodemask_of_node(&nodes_allowed, nid);
3907 n_mask = &nodes_allowed;
3908 } else {
3909 n_mask = &node_states[N_MEMORY];
3910 }
3911
3912 /* Synchronize with other sysfs operations modifying huge pages */
3913 mutex_lock(&h->resize_lock);
3914 spin_lock_irq(&hugetlb_lock);
3915
3916 while (nr_demote) {
3917 /*
3918 * Check for available pages to demote each time thorough the
3919 * loop as demote_pool_huge_page will drop hugetlb_lock.
3920 */
3921 if (nid != NUMA_NO_NODE)
3922 nr_available = h->free_huge_pages_node[nid];
3923 else
3924 nr_available = h->free_huge_pages;
3925 nr_available -= h->resv_huge_pages;
3926 if (!nr_available)
3927 break;
3928
3929 err = demote_pool_huge_page(h, n_mask);
3930 if (err)
3931 break;
3932
3933 nr_demote--;
3934 }
3935
3936 spin_unlock_irq(&hugetlb_lock);
3937 mutex_unlock(&h->resize_lock);
3938
3939 if (err)
3940 return err;
3941 return len;
3942 }
3943 HSTATE_ATTR_WO(demote);
3944
demote_size_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3945 static ssize_t demote_size_show(struct kobject *kobj,
3946 struct kobj_attribute *attr, char *buf)
3947 {
3948 struct hstate *h = kobj_to_hstate(kobj, NULL);
3949 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3950
3951 return sysfs_emit(buf, "%lukB\n", demote_size);
3952 }
3953
demote_size_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)3954 static ssize_t demote_size_store(struct kobject *kobj,
3955 struct kobj_attribute *attr,
3956 const char *buf, size_t count)
3957 {
3958 struct hstate *h, *demote_hstate;
3959 unsigned long demote_size;
3960 unsigned int demote_order;
3961
3962 demote_size = (unsigned long)memparse(buf, NULL);
3963
3964 demote_hstate = size_to_hstate(demote_size);
3965 if (!demote_hstate)
3966 return -EINVAL;
3967 demote_order = demote_hstate->order;
3968 if (demote_order < HUGETLB_PAGE_ORDER)
3969 return -EINVAL;
3970
3971 /* demote order must be smaller than hstate order */
3972 h = kobj_to_hstate(kobj, NULL);
3973 if (demote_order >= h->order)
3974 return -EINVAL;
3975
3976 /* resize_lock synchronizes access to demote size and writes */
3977 mutex_lock(&h->resize_lock);
3978 h->demote_order = demote_order;
3979 mutex_unlock(&h->resize_lock);
3980
3981 return count;
3982 }
3983 HSTATE_ATTR(demote_size);
3984
3985 static struct attribute *hstate_attrs[] = {
3986 &nr_hugepages_attr.attr,
3987 &nr_overcommit_hugepages_attr.attr,
3988 &free_hugepages_attr.attr,
3989 &resv_hugepages_attr.attr,
3990 &surplus_hugepages_attr.attr,
3991 #ifdef CONFIG_NUMA
3992 &nr_hugepages_mempolicy_attr.attr,
3993 #endif
3994 NULL,
3995 };
3996
3997 static const struct attribute_group hstate_attr_group = {
3998 .attrs = hstate_attrs,
3999 };
4000
4001 static struct attribute *hstate_demote_attrs[] = {
4002 &demote_size_attr.attr,
4003 &demote_attr.attr,
4004 NULL,
4005 };
4006
4007 static const struct attribute_group hstate_demote_attr_group = {
4008 .attrs = hstate_demote_attrs,
4009 };
4010
hugetlb_sysfs_add_hstate(struct hstate * h,struct kobject * parent,struct kobject ** hstate_kobjs,const struct attribute_group * hstate_attr_group)4011 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
4012 struct kobject **hstate_kobjs,
4013 const struct attribute_group *hstate_attr_group)
4014 {
4015 int retval;
4016 int hi = hstate_index(h);
4017
4018 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4019 if (!hstate_kobjs[hi])
4020 return -ENOMEM;
4021
4022 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4023 if (retval) {
4024 kobject_put(hstate_kobjs[hi]);
4025 hstate_kobjs[hi] = NULL;
4026 return retval;
4027 }
4028
4029 if (h->demote_order) {
4030 retval = sysfs_create_group(hstate_kobjs[hi],
4031 &hstate_demote_attr_group);
4032 if (retval) {
4033 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4034 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4035 kobject_put(hstate_kobjs[hi]);
4036 hstate_kobjs[hi] = NULL;
4037 return retval;
4038 }
4039 }
4040
4041 return 0;
4042 }
4043
4044 #ifdef CONFIG_NUMA
4045 static bool hugetlb_sysfs_initialized __ro_after_init;
4046
4047 /*
4048 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4049 * with node devices in node_devices[] using a parallel array. The array
4050 * index of a node device or _hstate == node id.
4051 * This is here to avoid any static dependency of the node device driver, in
4052 * the base kernel, on the hugetlb module.
4053 */
4054 struct node_hstate {
4055 struct kobject *hugepages_kobj;
4056 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4057 };
4058 static struct node_hstate node_hstates[MAX_NUMNODES];
4059
4060 /*
4061 * A subset of global hstate attributes for node devices
4062 */
4063 static struct attribute *per_node_hstate_attrs[] = {
4064 &nr_hugepages_attr.attr,
4065 &free_hugepages_attr.attr,
4066 &surplus_hugepages_attr.attr,
4067 NULL,
4068 };
4069
4070 static const struct attribute_group per_node_hstate_attr_group = {
4071 .attrs = per_node_hstate_attrs,
4072 };
4073
4074 /*
4075 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4076 * Returns node id via non-NULL nidp.
4077 */
kobj_to_node_hstate(struct kobject * kobj,int * nidp)4078 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4079 {
4080 int nid;
4081
4082 for (nid = 0; nid < nr_node_ids; nid++) {
4083 struct node_hstate *nhs = &node_hstates[nid];
4084 int i;
4085 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4086 if (nhs->hstate_kobjs[i] == kobj) {
4087 if (nidp)
4088 *nidp = nid;
4089 return &hstates[i];
4090 }
4091 }
4092
4093 BUG();
4094 return NULL;
4095 }
4096
4097 /*
4098 * Unregister hstate attributes from a single node device.
4099 * No-op if no hstate attributes attached.
4100 */
hugetlb_unregister_node(struct node * node)4101 void hugetlb_unregister_node(struct node *node)
4102 {
4103 struct hstate *h;
4104 struct node_hstate *nhs = &node_hstates[node->dev.id];
4105
4106 if (!nhs->hugepages_kobj)
4107 return; /* no hstate attributes */
4108
4109 for_each_hstate(h) {
4110 int idx = hstate_index(h);
4111 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4112
4113 if (!hstate_kobj)
4114 continue;
4115 if (h->demote_order)
4116 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4117 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4118 kobject_put(hstate_kobj);
4119 nhs->hstate_kobjs[idx] = NULL;
4120 }
4121
4122 kobject_put(nhs->hugepages_kobj);
4123 nhs->hugepages_kobj = NULL;
4124 }
4125
4126
4127 /*
4128 * Register hstate attributes for a single node device.
4129 * No-op if attributes already registered.
4130 */
hugetlb_register_node(struct node * node)4131 void hugetlb_register_node(struct node *node)
4132 {
4133 struct hstate *h;
4134 struct node_hstate *nhs = &node_hstates[node->dev.id];
4135 int err;
4136
4137 if (!hugetlb_sysfs_initialized)
4138 return;
4139
4140 if (nhs->hugepages_kobj)
4141 return; /* already allocated */
4142
4143 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4144 &node->dev.kobj);
4145 if (!nhs->hugepages_kobj)
4146 return;
4147
4148 for_each_hstate(h) {
4149 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4150 nhs->hstate_kobjs,
4151 &per_node_hstate_attr_group);
4152 if (err) {
4153 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4154 h->name, node->dev.id);
4155 hugetlb_unregister_node(node);
4156 break;
4157 }
4158 }
4159 }
4160
4161 /*
4162 * hugetlb init time: register hstate attributes for all registered node
4163 * devices of nodes that have memory. All on-line nodes should have
4164 * registered their associated device by this time.
4165 */
hugetlb_register_all_nodes(void)4166 static void __init hugetlb_register_all_nodes(void)
4167 {
4168 int nid;
4169
4170 for_each_online_node(nid)
4171 hugetlb_register_node(node_devices[nid]);
4172 }
4173 #else /* !CONFIG_NUMA */
4174
kobj_to_node_hstate(struct kobject * kobj,int * nidp)4175 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4176 {
4177 BUG();
4178 if (nidp)
4179 *nidp = -1;
4180 return NULL;
4181 }
4182
hugetlb_register_all_nodes(void)4183 static void hugetlb_register_all_nodes(void) { }
4184
4185 #endif
4186
4187 #ifdef CONFIG_CMA
4188 static void __init hugetlb_cma_check(void);
4189 #else
hugetlb_cma_check(void)4190 static inline __init void hugetlb_cma_check(void)
4191 {
4192 }
4193 #endif
4194
hugetlb_sysfs_init(void)4195 static void __init hugetlb_sysfs_init(void)
4196 {
4197 struct hstate *h;
4198 int err;
4199
4200 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4201 if (!hugepages_kobj)
4202 return;
4203
4204 for_each_hstate(h) {
4205 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4206 hstate_kobjs, &hstate_attr_group);
4207 if (err)
4208 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4209 }
4210
4211 #ifdef CONFIG_NUMA
4212 hugetlb_sysfs_initialized = true;
4213 #endif
4214 hugetlb_register_all_nodes();
4215 }
4216
4217 #ifdef CONFIG_SYSCTL
4218 static void hugetlb_sysctl_init(void);
4219 #else
hugetlb_sysctl_init(void)4220 static inline void hugetlb_sysctl_init(void) { }
4221 #endif
4222
hugetlb_init(void)4223 static int __init hugetlb_init(void)
4224 {
4225 int i;
4226
4227 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4228 __NR_HPAGEFLAGS);
4229
4230 if (!hugepages_supported()) {
4231 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4232 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4233 return 0;
4234 }
4235
4236 /*
4237 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4238 * architectures depend on setup being done here.
4239 */
4240 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4241 if (!parsed_default_hugepagesz) {
4242 /*
4243 * If we did not parse a default huge page size, set
4244 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4245 * number of huge pages for this default size was implicitly
4246 * specified, set that here as well.
4247 * Note that the implicit setting will overwrite an explicit
4248 * setting. A warning will be printed in this case.
4249 */
4250 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4251 if (default_hstate_max_huge_pages) {
4252 if (default_hstate.max_huge_pages) {
4253 char buf[32];
4254
4255 string_get_size(huge_page_size(&default_hstate),
4256 1, STRING_UNITS_2, buf, 32);
4257 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4258 default_hstate.max_huge_pages, buf);
4259 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4260 default_hstate_max_huge_pages);
4261 }
4262 default_hstate.max_huge_pages =
4263 default_hstate_max_huge_pages;
4264
4265 for_each_online_node(i)
4266 default_hstate.max_huge_pages_node[i] =
4267 default_hugepages_in_node[i];
4268 }
4269 }
4270
4271 hugetlb_cma_check();
4272 hugetlb_init_hstates();
4273 gather_bootmem_prealloc();
4274 report_hugepages();
4275
4276 hugetlb_sysfs_init();
4277 hugetlb_cgroup_file_init();
4278 hugetlb_sysctl_init();
4279
4280 #ifdef CONFIG_SMP
4281 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4282 #else
4283 num_fault_mutexes = 1;
4284 #endif
4285 hugetlb_fault_mutex_table =
4286 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4287 GFP_KERNEL);
4288 BUG_ON(!hugetlb_fault_mutex_table);
4289
4290 for (i = 0; i < num_fault_mutexes; i++)
4291 mutex_init(&hugetlb_fault_mutex_table[i]);
4292 return 0;
4293 }
4294 subsys_initcall(hugetlb_init);
4295
4296 /* Overwritten by architectures with more huge page sizes */
__init(weak)4297 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4298 {
4299 return size == HPAGE_SIZE;
4300 }
4301
hugetlb_add_hstate(unsigned int order)4302 void __init hugetlb_add_hstate(unsigned int order)
4303 {
4304 struct hstate *h;
4305 unsigned long i;
4306
4307 if (size_to_hstate(PAGE_SIZE << order)) {
4308 return;
4309 }
4310 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4311 BUG_ON(order == 0);
4312 h = &hstates[hugetlb_max_hstate++];
4313 mutex_init(&h->resize_lock);
4314 h->order = order;
4315 h->mask = ~(huge_page_size(h) - 1);
4316 for (i = 0; i < MAX_NUMNODES; ++i)
4317 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4318 INIT_LIST_HEAD(&h->hugepage_activelist);
4319 h->next_nid_to_alloc = first_memory_node;
4320 h->next_nid_to_free = first_memory_node;
4321 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4322 huge_page_size(h)/SZ_1K);
4323
4324 parsed_hstate = h;
4325 }
4326
hugetlb_node_alloc_supported(void)4327 bool __init __weak hugetlb_node_alloc_supported(void)
4328 {
4329 return true;
4330 }
4331
hugepages_clear_pages_in_node(void)4332 static void __init hugepages_clear_pages_in_node(void)
4333 {
4334 if (!hugetlb_max_hstate) {
4335 default_hstate_max_huge_pages = 0;
4336 memset(default_hugepages_in_node, 0,
4337 sizeof(default_hugepages_in_node));
4338 } else {
4339 parsed_hstate->max_huge_pages = 0;
4340 memset(parsed_hstate->max_huge_pages_node, 0,
4341 sizeof(parsed_hstate->max_huge_pages_node));
4342 }
4343 }
4344
4345 /*
4346 * hugepages command line processing
4347 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4348 * specification. If not, ignore the hugepages value. hugepages can also
4349 * be the first huge page command line option in which case it implicitly
4350 * specifies the number of huge pages for the default size.
4351 */
hugepages_setup(char * s)4352 static int __init hugepages_setup(char *s)
4353 {
4354 unsigned long *mhp;
4355 static unsigned long *last_mhp;
4356 int node = NUMA_NO_NODE;
4357 int count;
4358 unsigned long tmp;
4359 char *p = s;
4360
4361 if (!parsed_valid_hugepagesz) {
4362 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4363 parsed_valid_hugepagesz = true;
4364 return 1;
4365 }
4366
4367 /*
4368 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4369 * yet, so this hugepages= parameter goes to the "default hstate".
4370 * Otherwise, it goes with the previously parsed hugepagesz or
4371 * default_hugepagesz.
4372 */
4373 else if (!hugetlb_max_hstate)
4374 mhp = &default_hstate_max_huge_pages;
4375 else
4376 mhp = &parsed_hstate->max_huge_pages;
4377
4378 if (mhp == last_mhp) {
4379 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4380 return 1;
4381 }
4382
4383 while (*p) {
4384 count = 0;
4385 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4386 goto invalid;
4387 /* Parameter is node format */
4388 if (p[count] == ':') {
4389 if (!hugetlb_node_alloc_supported()) {
4390 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4391 return 1;
4392 }
4393 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4394 goto invalid;
4395 node = array_index_nospec(tmp, MAX_NUMNODES);
4396 p += count + 1;
4397 /* Parse hugepages */
4398 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4399 goto invalid;
4400 if (!hugetlb_max_hstate)
4401 default_hugepages_in_node[node] = tmp;
4402 else
4403 parsed_hstate->max_huge_pages_node[node] = tmp;
4404 *mhp += tmp;
4405 /* Go to parse next node*/
4406 if (p[count] == ',')
4407 p += count + 1;
4408 else
4409 break;
4410 } else {
4411 if (p != s)
4412 goto invalid;
4413 *mhp = tmp;
4414 break;
4415 }
4416 }
4417
4418 /*
4419 * Global state is always initialized later in hugetlb_init.
4420 * But we need to allocate gigantic hstates here early to still
4421 * use the bootmem allocator.
4422 */
4423 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4424 hugetlb_hstate_alloc_pages(parsed_hstate);
4425
4426 last_mhp = mhp;
4427
4428 return 1;
4429
4430 invalid:
4431 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4432 hugepages_clear_pages_in_node();
4433 return 1;
4434 }
4435 __setup("hugepages=", hugepages_setup);
4436
4437 /*
4438 * hugepagesz command line processing
4439 * A specific huge page size can only be specified once with hugepagesz.
4440 * hugepagesz is followed by hugepages on the command line. The global
4441 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4442 * hugepagesz argument was valid.
4443 */
hugepagesz_setup(char * s)4444 static int __init hugepagesz_setup(char *s)
4445 {
4446 unsigned long size;
4447 struct hstate *h;
4448
4449 parsed_valid_hugepagesz = false;
4450 size = (unsigned long)memparse(s, NULL);
4451
4452 if (!arch_hugetlb_valid_size(size)) {
4453 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4454 return 1;
4455 }
4456
4457 h = size_to_hstate(size);
4458 if (h) {
4459 /*
4460 * hstate for this size already exists. This is normally
4461 * an error, but is allowed if the existing hstate is the
4462 * default hstate. More specifically, it is only allowed if
4463 * the number of huge pages for the default hstate was not
4464 * previously specified.
4465 */
4466 if (!parsed_default_hugepagesz || h != &default_hstate ||
4467 default_hstate.max_huge_pages) {
4468 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4469 return 1;
4470 }
4471
4472 /*
4473 * No need to call hugetlb_add_hstate() as hstate already
4474 * exists. But, do set parsed_hstate so that a following
4475 * hugepages= parameter will be applied to this hstate.
4476 */
4477 parsed_hstate = h;
4478 parsed_valid_hugepagesz = true;
4479 return 1;
4480 }
4481
4482 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4483 parsed_valid_hugepagesz = true;
4484 return 1;
4485 }
4486 __setup("hugepagesz=", hugepagesz_setup);
4487
4488 /*
4489 * default_hugepagesz command line input
4490 * Only one instance of default_hugepagesz allowed on command line.
4491 */
default_hugepagesz_setup(char * s)4492 static int __init default_hugepagesz_setup(char *s)
4493 {
4494 unsigned long size;
4495 int i;
4496
4497 parsed_valid_hugepagesz = false;
4498 if (parsed_default_hugepagesz) {
4499 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4500 return 1;
4501 }
4502
4503 size = (unsigned long)memparse(s, NULL);
4504
4505 if (!arch_hugetlb_valid_size(size)) {
4506 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4507 return 1;
4508 }
4509
4510 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4511 parsed_valid_hugepagesz = true;
4512 parsed_default_hugepagesz = true;
4513 default_hstate_idx = hstate_index(size_to_hstate(size));
4514
4515 /*
4516 * The number of default huge pages (for this size) could have been
4517 * specified as the first hugetlb parameter: hugepages=X. If so,
4518 * then default_hstate_max_huge_pages is set. If the default huge
4519 * page size is gigantic (> MAX_ORDER), then the pages must be
4520 * allocated here from bootmem allocator.
4521 */
4522 if (default_hstate_max_huge_pages) {
4523 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4524 for_each_online_node(i)
4525 default_hstate.max_huge_pages_node[i] =
4526 default_hugepages_in_node[i];
4527 if (hstate_is_gigantic(&default_hstate))
4528 hugetlb_hstate_alloc_pages(&default_hstate);
4529 default_hstate_max_huge_pages = 0;
4530 }
4531
4532 return 1;
4533 }
4534 __setup("default_hugepagesz=", default_hugepagesz_setup);
4535
policy_mbind_nodemask(gfp_t gfp)4536 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4537 {
4538 #ifdef CONFIG_NUMA
4539 struct mempolicy *mpol = get_task_policy(current);
4540
4541 /*
4542 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4543 * (from policy_nodemask) specifically for hugetlb case
4544 */
4545 if (mpol->mode == MPOL_BIND &&
4546 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4547 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4548 return &mpol->nodes;
4549 #endif
4550 return NULL;
4551 }
4552
allowed_mems_nr(struct hstate * h)4553 static unsigned int allowed_mems_nr(struct hstate *h)
4554 {
4555 int node;
4556 unsigned int nr = 0;
4557 nodemask_t *mbind_nodemask;
4558 unsigned int *array = h->free_huge_pages_node;
4559 gfp_t gfp_mask = htlb_alloc_mask(h);
4560
4561 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4562 for_each_node_mask(node, cpuset_current_mems_allowed) {
4563 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4564 nr += array[node];
4565 }
4566
4567 return nr;
4568 }
4569
4570 #ifdef CONFIG_SYSCTL
proc_hugetlb_doulongvec_minmax(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos,unsigned long * out)4571 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4572 void *buffer, size_t *length,
4573 loff_t *ppos, unsigned long *out)
4574 {
4575 struct ctl_table dup_table;
4576
4577 /*
4578 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4579 * can duplicate the @table and alter the duplicate of it.
4580 */
4581 dup_table = *table;
4582 dup_table.data = out;
4583
4584 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4585 }
4586
hugetlb_sysctl_handler_common(bool obey_mempolicy,struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)4587 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4588 struct ctl_table *table, int write,
4589 void *buffer, size_t *length, loff_t *ppos)
4590 {
4591 struct hstate *h = &default_hstate;
4592 unsigned long tmp = h->max_huge_pages;
4593 int ret;
4594
4595 if (!hugepages_supported())
4596 return -EOPNOTSUPP;
4597
4598 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4599 &tmp);
4600 if (ret)
4601 goto out;
4602
4603 if (write)
4604 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4605 NUMA_NO_NODE, tmp, *length);
4606 out:
4607 return ret;
4608 }
4609
hugetlb_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)4610 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4611 void *buffer, size_t *length, loff_t *ppos)
4612 {
4613
4614 return hugetlb_sysctl_handler_common(false, table, write,
4615 buffer, length, ppos);
4616 }
4617
4618 #ifdef CONFIG_NUMA
hugetlb_mempolicy_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)4619 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4620 void *buffer, size_t *length, loff_t *ppos)
4621 {
4622 return hugetlb_sysctl_handler_common(true, table, write,
4623 buffer, length, ppos);
4624 }
4625 #endif /* CONFIG_NUMA */
4626
hugetlb_overcommit_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)4627 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4628 void *buffer, size_t *length, loff_t *ppos)
4629 {
4630 struct hstate *h = &default_hstate;
4631 unsigned long tmp;
4632 int ret;
4633
4634 if (!hugepages_supported())
4635 return -EOPNOTSUPP;
4636
4637 tmp = h->nr_overcommit_huge_pages;
4638
4639 if (write && hstate_is_gigantic(h))
4640 return -EINVAL;
4641
4642 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4643 &tmp);
4644 if (ret)
4645 goto out;
4646
4647 if (write) {
4648 spin_lock_irq(&hugetlb_lock);
4649 h->nr_overcommit_huge_pages = tmp;
4650 spin_unlock_irq(&hugetlb_lock);
4651 }
4652 out:
4653 return ret;
4654 }
4655
4656 static struct ctl_table hugetlb_table[] = {
4657 {
4658 .procname = "nr_hugepages",
4659 .data = NULL,
4660 .maxlen = sizeof(unsigned long),
4661 .mode = 0644,
4662 .proc_handler = hugetlb_sysctl_handler,
4663 },
4664 #ifdef CONFIG_NUMA
4665 {
4666 .procname = "nr_hugepages_mempolicy",
4667 .data = NULL,
4668 .maxlen = sizeof(unsigned long),
4669 .mode = 0644,
4670 .proc_handler = &hugetlb_mempolicy_sysctl_handler,
4671 },
4672 #endif
4673 {
4674 .procname = "hugetlb_shm_group",
4675 .data = &sysctl_hugetlb_shm_group,
4676 .maxlen = sizeof(gid_t),
4677 .mode = 0644,
4678 .proc_handler = proc_dointvec,
4679 },
4680 {
4681 .procname = "nr_overcommit_hugepages",
4682 .data = NULL,
4683 .maxlen = sizeof(unsigned long),
4684 .mode = 0644,
4685 .proc_handler = hugetlb_overcommit_handler,
4686 },
4687 { }
4688 };
4689
hugetlb_sysctl_init(void)4690 static void hugetlb_sysctl_init(void)
4691 {
4692 register_sysctl_init("vm", hugetlb_table);
4693 }
4694 #endif /* CONFIG_SYSCTL */
4695
hugetlb_report_meminfo(struct seq_file * m)4696 void hugetlb_report_meminfo(struct seq_file *m)
4697 {
4698 struct hstate *h;
4699 unsigned long total = 0;
4700
4701 if (!hugepages_supported())
4702 return;
4703
4704 for_each_hstate(h) {
4705 unsigned long count = h->nr_huge_pages;
4706
4707 total += huge_page_size(h) * count;
4708
4709 if (h == &default_hstate)
4710 seq_printf(m,
4711 "HugePages_Total: %5lu\n"
4712 "HugePages_Free: %5lu\n"
4713 "HugePages_Rsvd: %5lu\n"
4714 "HugePages_Surp: %5lu\n"
4715 "Hugepagesize: %8lu kB\n",
4716 count,
4717 h->free_huge_pages,
4718 h->resv_huge_pages,
4719 h->surplus_huge_pages,
4720 huge_page_size(h) / SZ_1K);
4721 }
4722
4723 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4724 }
4725
hugetlb_report_node_meminfo(char * buf,int len,int nid)4726 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4727 {
4728 struct hstate *h = &default_hstate;
4729
4730 if (!hugepages_supported())
4731 return 0;
4732
4733 return sysfs_emit_at(buf, len,
4734 "Node %d HugePages_Total: %5u\n"
4735 "Node %d HugePages_Free: %5u\n"
4736 "Node %d HugePages_Surp: %5u\n",
4737 nid, h->nr_huge_pages_node[nid],
4738 nid, h->free_huge_pages_node[nid],
4739 nid, h->surplus_huge_pages_node[nid]);
4740 }
4741
hugetlb_show_meminfo_node(int nid)4742 void hugetlb_show_meminfo_node(int nid)
4743 {
4744 struct hstate *h;
4745
4746 if (!hugepages_supported())
4747 return;
4748
4749 for_each_hstate(h)
4750 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4751 nid,
4752 h->nr_huge_pages_node[nid],
4753 h->free_huge_pages_node[nid],
4754 h->surplus_huge_pages_node[nid],
4755 huge_page_size(h) / SZ_1K);
4756 }
4757
hugetlb_report_usage(struct seq_file * m,struct mm_struct * mm)4758 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4759 {
4760 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4761 K(atomic_long_read(&mm->hugetlb_usage)));
4762 }
4763
4764 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
hugetlb_total_pages(void)4765 unsigned long hugetlb_total_pages(void)
4766 {
4767 struct hstate *h;
4768 unsigned long nr_total_pages = 0;
4769
4770 for_each_hstate(h)
4771 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4772 return nr_total_pages;
4773 }
4774
hugetlb_acct_memory(struct hstate * h,long delta)4775 static int hugetlb_acct_memory(struct hstate *h, long delta)
4776 {
4777 int ret = -ENOMEM;
4778
4779 if (!delta)
4780 return 0;
4781
4782 spin_lock_irq(&hugetlb_lock);
4783 /*
4784 * When cpuset is configured, it breaks the strict hugetlb page
4785 * reservation as the accounting is done on a global variable. Such
4786 * reservation is completely rubbish in the presence of cpuset because
4787 * the reservation is not checked against page availability for the
4788 * current cpuset. Application can still potentially OOM'ed by kernel
4789 * with lack of free htlb page in cpuset that the task is in.
4790 * Attempt to enforce strict accounting with cpuset is almost
4791 * impossible (or too ugly) because cpuset is too fluid that
4792 * task or memory node can be dynamically moved between cpusets.
4793 *
4794 * The change of semantics for shared hugetlb mapping with cpuset is
4795 * undesirable. However, in order to preserve some of the semantics,
4796 * we fall back to check against current free page availability as
4797 * a best attempt and hopefully to minimize the impact of changing
4798 * semantics that cpuset has.
4799 *
4800 * Apart from cpuset, we also have memory policy mechanism that
4801 * also determines from which node the kernel will allocate memory
4802 * in a NUMA system. So similar to cpuset, we also should consider
4803 * the memory policy of the current task. Similar to the description
4804 * above.
4805 */
4806 if (delta > 0) {
4807 if (gather_surplus_pages(h, delta) < 0)
4808 goto out;
4809
4810 if (delta > allowed_mems_nr(h)) {
4811 return_unused_surplus_pages(h, delta);
4812 goto out;
4813 }
4814 }
4815
4816 ret = 0;
4817 if (delta < 0)
4818 return_unused_surplus_pages(h, (unsigned long) -delta);
4819
4820 out:
4821 spin_unlock_irq(&hugetlb_lock);
4822 return ret;
4823 }
4824
hugetlb_vm_op_open(struct vm_area_struct * vma)4825 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4826 {
4827 struct resv_map *resv = vma_resv_map(vma);
4828
4829 /*
4830 * HPAGE_RESV_OWNER indicates a private mapping.
4831 * This new VMA should share its siblings reservation map if present.
4832 * The VMA will only ever have a valid reservation map pointer where
4833 * it is being copied for another still existing VMA. As that VMA
4834 * has a reference to the reservation map it cannot disappear until
4835 * after this open call completes. It is therefore safe to take a
4836 * new reference here without additional locking.
4837 */
4838 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4839 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4840 kref_get(&resv->refs);
4841 }
4842
4843 /*
4844 * vma_lock structure for sharable mappings is vma specific.
4845 * Clear old pointer (if copied via vm_area_dup) and allocate
4846 * new structure. Before clearing, make sure vma_lock is not
4847 * for this vma.
4848 */
4849 if (vma->vm_flags & VM_MAYSHARE) {
4850 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4851
4852 if (vma_lock) {
4853 if (vma_lock->vma != vma) {
4854 vma->vm_private_data = NULL;
4855 hugetlb_vma_lock_alloc(vma);
4856 } else
4857 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4858 } else
4859 hugetlb_vma_lock_alloc(vma);
4860 }
4861 }
4862
hugetlb_vm_op_close(struct vm_area_struct * vma)4863 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4864 {
4865 struct hstate *h = hstate_vma(vma);
4866 struct resv_map *resv;
4867 struct hugepage_subpool *spool = subpool_vma(vma);
4868 unsigned long reserve, start, end;
4869 long gbl_reserve;
4870
4871 hugetlb_vma_lock_free(vma);
4872
4873 resv = vma_resv_map(vma);
4874 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4875 return;
4876
4877 start = vma_hugecache_offset(h, vma, vma->vm_start);
4878 end = vma_hugecache_offset(h, vma, vma->vm_end);
4879
4880 reserve = (end - start) - region_count(resv, start, end);
4881 hugetlb_cgroup_uncharge_counter(resv, start, end);
4882 if (reserve) {
4883 /*
4884 * Decrement reserve counts. The global reserve count may be
4885 * adjusted if the subpool has a minimum size.
4886 */
4887 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4888 hugetlb_acct_memory(h, -gbl_reserve);
4889 }
4890
4891 kref_put(&resv->refs, resv_map_release);
4892 }
4893
hugetlb_vm_op_split(struct vm_area_struct * vma,unsigned long addr)4894 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4895 {
4896 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4897 return -EINVAL;
4898
4899 /*
4900 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4901 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4902 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4903 */
4904 if (addr & ~PUD_MASK) {
4905 /*
4906 * hugetlb_vm_op_split is called right before we attempt to
4907 * split the VMA. We will need to unshare PMDs in the old and
4908 * new VMAs, so let's unshare before we split.
4909 */
4910 unsigned long floor = addr & PUD_MASK;
4911 unsigned long ceil = floor + PUD_SIZE;
4912
4913 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4914 hugetlb_unshare_pmds(vma, floor, ceil);
4915 }
4916
4917 return 0;
4918 }
4919
hugetlb_vm_op_pagesize(struct vm_area_struct * vma)4920 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4921 {
4922 return huge_page_size(hstate_vma(vma));
4923 }
4924
4925 /*
4926 * We cannot handle pagefaults against hugetlb pages at all. They cause
4927 * handle_mm_fault() to try to instantiate regular-sized pages in the
4928 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4929 * this far.
4930 */
hugetlb_vm_op_fault(struct vm_fault * vmf)4931 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4932 {
4933 BUG();
4934 return 0;
4935 }
4936
4937 /*
4938 * When a new function is introduced to vm_operations_struct and added
4939 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4940 * This is because under System V memory model, mappings created via
4941 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4942 * their original vm_ops are overwritten with shm_vm_ops.
4943 */
4944 const struct vm_operations_struct hugetlb_vm_ops = {
4945 .fault = hugetlb_vm_op_fault,
4946 .open = hugetlb_vm_op_open,
4947 .close = hugetlb_vm_op_close,
4948 .may_split = hugetlb_vm_op_split,
4949 .pagesize = hugetlb_vm_op_pagesize,
4950 };
4951
make_huge_pte(struct vm_area_struct * vma,struct page * page,int writable)4952 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4953 int writable)
4954 {
4955 pte_t entry;
4956 unsigned int shift = huge_page_shift(hstate_vma(vma));
4957
4958 if (writable) {
4959 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4960 vma->vm_page_prot)));
4961 } else {
4962 entry = huge_pte_wrprotect(mk_huge_pte(page,
4963 vma->vm_page_prot));
4964 }
4965 entry = pte_mkyoung(entry);
4966 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4967
4968 return entry;
4969 }
4970
set_huge_ptep_writable(struct vm_area_struct * vma,unsigned long address,pte_t * ptep)4971 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4972 unsigned long address, pte_t *ptep)
4973 {
4974 pte_t entry;
4975
4976 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4977 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4978 update_mmu_cache(vma, address, ptep);
4979 }
4980
is_hugetlb_entry_migration(pte_t pte)4981 bool is_hugetlb_entry_migration(pte_t pte)
4982 {
4983 swp_entry_t swp;
4984
4985 if (huge_pte_none(pte) || pte_present(pte))
4986 return false;
4987 swp = pte_to_swp_entry(pte);
4988 if (is_migration_entry(swp))
4989 return true;
4990 else
4991 return false;
4992 }
4993
is_hugetlb_entry_hwpoisoned(pte_t pte)4994 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4995 {
4996 swp_entry_t swp;
4997
4998 if (huge_pte_none(pte) || pte_present(pte))
4999 return false;
5000 swp = pte_to_swp_entry(pte);
5001 if (is_hwpoison_entry(swp))
5002 return true;
5003 else
5004 return false;
5005 }
5006
5007 static void
hugetlb_install_folio(struct vm_area_struct * vma,pte_t * ptep,unsigned long addr,struct folio * new_folio,pte_t old,unsigned long sz)5008 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
5009 struct folio *new_folio, pte_t old, unsigned long sz)
5010 {
5011 pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
5012
5013 __folio_mark_uptodate(new_folio);
5014 hugepage_add_new_anon_rmap(new_folio, vma, addr);
5015 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
5016 newpte = huge_pte_mkuffd_wp(newpte);
5017 set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
5018 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
5019 folio_set_hugetlb_migratable(new_folio);
5020 }
5021
copy_hugetlb_page_range(struct mm_struct * dst,struct mm_struct * src,struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma)5022 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
5023 struct vm_area_struct *dst_vma,
5024 struct vm_area_struct *src_vma)
5025 {
5026 pte_t *src_pte, *dst_pte, entry;
5027 struct folio *pte_folio;
5028 unsigned long addr;
5029 bool cow = is_cow_mapping(src_vma->vm_flags);
5030 struct hstate *h = hstate_vma(src_vma);
5031 unsigned long sz = huge_page_size(h);
5032 unsigned long npages = pages_per_huge_page(h);
5033 struct mmu_notifier_range range;
5034 unsigned long last_addr_mask;
5035 int ret = 0;
5036
5037 if (cow) {
5038 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5039 src_vma->vm_start,
5040 src_vma->vm_end);
5041 mmu_notifier_invalidate_range_start(&range);
5042 vma_assert_write_locked(src_vma);
5043 raw_write_seqcount_begin(&src->write_protect_seq);
5044 } else {
5045 /*
5046 * For shared mappings the vma lock must be held before
5047 * calling hugetlb_walk() in the src vma. Otherwise, the
5048 * returned ptep could go away if part of a shared pmd and
5049 * another thread calls huge_pmd_unshare.
5050 */
5051 hugetlb_vma_lock_read(src_vma);
5052 }
5053
5054 last_addr_mask = hugetlb_mask_last_page(h);
5055 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5056 spinlock_t *src_ptl, *dst_ptl;
5057 src_pte = hugetlb_walk(src_vma, addr, sz);
5058 if (!src_pte) {
5059 addr |= last_addr_mask;
5060 continue;
5061 }
5062 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5063 if (!dst_pte) {
5064 ret = -ENOMEM;
5065 break;
5066 }
5067
5068 /*
5069 * If the pagetables are shared don't copy or take references.
5070 *
5071 * dst_pte == src_pte is the common case of src/dest sharing.
5072 * However, src could have 'unshared' and dst shares with
5073 * another vma. So page_count of ptep page is checked instead
5074 * to reliably determine whether pte is shared.
5075 */
5076 if (page_count(virt_to_page(dst_pte)) > 1) {
5077 addr |= last_addr_mask;
5078 continue;
5079 }
5080
5081 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5082 src_ptl = huge_pte_lockptr(h, src, src_pte);
5083 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5084 entry = huge_ptep_get(src_pte);
5085 again:
5086 if (huge_pte_none(entry)) {
5087 /*
5088 * Skip if src entry none.
5089 */
5090 ;
5091 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5092 if (!userfaultfd_wp(dst_vma))
5093 entry = huge_pte_clear_uffd_wp(entry);
5094 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5095 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5096 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5097 bool uffd_wp = pte_swp_uffd_wp(entry);
5098
5099 if (!is_readable_migration_entry(swp_entry) && cow) {
5100 /*
5101 * COW mappings require pages in both
5102 * parent and child to be set to read.
5103 */
5104 swp_entry = make_readable_migration_entry(
5105 swp_offset(swp_entry));
5106 entry = swp_entry_to_pte(swp_entry);
5107 if (userfaultfd_wp(src_vma) && uffd_wp)
5108 entry = pte_swp_mkuffd_wp(entry);
5109 set_huge_pte_at(src, addr, src_pte, entry, sz);
5110 }
5111 if (!userfaultfd_wp(dst_vma))
5112 entry = huge_pte_clear_uffd_wp(entry);
5113 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5114 } else if (unlikely(is_pte_marker(entry))) {
5115 pte_marker marker = copy_pte_marker(
5116 pte_to_swp_entry(entry), dst_vma);
5117
5118 if (marker)
5119 set_huge_pte_at(dst, addr, dst_pte,
5120 make_pte_marker(marker), sz);
5121 } else {
5122 entry = huge_ptep_get(src_pte);
5123 pte_folio = page_folio(pte_page(entry));
5124 folio_get(pte_folio);
5125
5126 /*
5127 * Failing to duplicate the anon rmap is a rare case
5128 * where we see pinned hugetlb pages while they're
5129 * prone to COW. We need to do the COW earlier during
5130 * fork.
5131 *
5132 * When pre-allocating the page or copying data, we
5133 * need to be without the pgtable locks since we could
5134 * sleep during the process.
5135 */
5136 if (!folio_test_anon(pte_folio)) {
5137 page_dup_file_rmap(&pte_folio->page, true);
5138 } else if (page_try_dup_anon_rmap(&pte_folio->page,
5139 true, src_vma)) {
5140 pte_t src_pte_old = entry;
5141 struct folio *new_folio;
5142
5143 spin_unlock(src_ptl);
5144 spin_unlock(dst_ptl);
5145 /* Do not use reserve as it's private owned */
5146 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5147 if (IS_ERR(new_folio)) {
5148 folio_put(pte_folio);
5149 ret = PTR_ERR(new_folio);
5150 break;
5151 }
5152 ret = copy_user_large_folio(new_folio,
5153 pte_folio,
5154 addr, dst_vma);
5155 folio_put(pte_folio);
5156 if (ret) {
5157 folio_put(new_folio);
5158 break;
5159 }
5160
5161 /* Install the new hugetlb folio if src pte stable */
5162 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5163 src_ptl = huge_pte_lockptr(h, src, src_pte);
5164 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5165 entry = huge_ptep_get(src_pte);
5166 if (!pte_same(src_pte_old, entry)) {
5167 restore_reserve_on_error(h, dst_vma, addr,
5168 new_folio);
5169 folio_put(new_folio);
5170 /* huge_ptep of dst_pte won't change as in child */
5171 goto again;
5172 }
5173 hugetlb_install_folio(dst_vma, dst_pte, addr,
5174 new_folio, src_pte_old, sz);
5175 spin_unlock(src_ptl);
5176 spin_unlock(dst_ptl);
5177 continue;
5178 }
5179
5180 if (cow) {
5181 /*
5182 * No need to notify as we are downgrading page
5183 * table protection not changing it to point
5184 * to a new page.
5185 *
5186 * See Documentation/mm/mmu_notifier.rst
5187 */
5188 huge_ptep_set_wrprotect(src, addr, src_pte);
5189 entry = huge_pte_wrprotect(entry);
5190 }
5191
5192 if (!userfaultfd_wp(dst_vma))
5193 entry = huge_pte_clear_uffd_wp(entry);
5194
5195 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5196 hugetlb_count_add(npages, dst);
5197 }
5198 spin_unlock(src_ptl);
5199 spin_unlock(dst_ptl);
5200 }
5201
5202 if (cow) {
5203 raw_write_seqcount_end(&src->write_protect_seq);
5204 mmu_notifier_invalidate_range_end(&range);
5205 } else {
5206 hugetlb_vma_unlock_read(src_vma);
5207 }
5208
5209 return ret;
5210 }
5211
move_huge_pte(struct vm_area_struct * vma,unsigned long old_addr,unsigned long new_addr,pte_t * src_pte,pte_t * dst_pte,unsigned long sz)5212 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5213 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5214 unsigned long sz)
5215 {
5216 struct hstate *h = hstate_vma(vma);
5217 struct mm_struct *mm = vma->vm_mm;
5218 spinlock_t *src_ptl, *dst_ptl;
5219 pte_t pte;
5220
5221 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5222 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5223
5224 /*
5225 * We don't have to worry about the ordering of src and dst ptlocks
5226 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5227 */
5228 if (src_ptl != dst_ptl)
5229 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5230
5231 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5232 set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5233
5234 if (src_ptl != dst_ptl)
5235 spin_unlock(src_ptl);
5236 spin_unlock(dst_ptl);
5237 }
5238
move_hugetlb_page_tables(struct vm_area_struct * vma,struct vm_area_struct * new_vma,unsigned long old_addr,unsigned long new_addr,unsigned long len)5239 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5240 struct vm_area_struct *new_vma,
5241 unsigned long old_addr, unsigned long new_addr,
5242 unsigned long len)
5243 {
5244 struct hstate *h = hstate_vma(vma);
5245 struct address_space *mapping = vma->vm_file->f_mapping;
5246 unsigned long sz = huge_page_size(h);
5247 struct mm_struct *mm = vma->vm_mm;
5248 unsigned long old_end = old_addr + len;
5249 unsigned long last_addr_mask;
5250 pte_t *src_pte, *dst_pte;
5251 struct mmu_notifier_range range;
5252 bool shared_pmd = false;
5253
5254 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5255 old_end);
5256 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5257 /*
5258 * In case of shared PMDs, we should cover the maximum possible
5259 * range.
5260 */
5261 flush_cache_range(vma, range.start, range.end);
5262
5263 mmu_notifier_invalidate_range_start(&range);
5264 last_addr_mask = hugetlb_mask_last_page(h);
5265 /* Prevent race with file truncation */
5266 hugetlb_vma_lock_write(vma);
5267 i_mmap_lock_write(mapping);
5268 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5269 src_pte = hugetlb_walk(vma, old_addr, sz);
5270 if (!src_pte) {
5271 old_addr |= last_addr_mask;
5272 new_addr |= last_addr_mask;
5273 continue;
5274 }
5275 if (huge_pte_none(huge_ptep_get(src_pte)))
5276 continue;
5277
5278 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5279 shared_pmd = true;
5280 old_addr |= last_addr_mask;
5281 new_addr |= last_addr_mask;
5282 continue;
5283 }
5284
5285 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5286 if (!dst_pte)
5287 break;
5288
5289 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5290 }
5291
5292 if (shared_pmd)
5293 flush_hugetlb_tlb_range(vma, range.start, range.end);
5294 else
5295 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5296 mmu_notifier_invalidate_range_end(&range);
5297 i_mmap_unlock_write(mapping);
5298 hugetlb_vma_unlock_write(vma);
5299
5300 return len + old_addr - old_end;
5301 }
5302
__unmap_hugepage_range(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long start,unsigned long end,struct page * ref_page,zap_flags_t zap_flags)5303 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5304 unsigned long start, unsigned long end,
5305 struct page *ref_page, zap_flags_t zap_flags)
5306 {
5307 struct mm_struct *mm = vma->vm_mm;
5308 unsigned long address;
5309 pte_t *ptep;
5310 pte_t pte;
5311 spinlock_t *ptl;
5312 struct page *page;
5313 struct hstate *h = hstate_vma(vma);
5314 unsigned long sz = huge_page_size(h);
5315 unsigned long last_addr_mask;
5316 bool force_flush = false;
5317
5318 WARN_ON(!is_vm_hugetlb_page(vma));
5319 BUG_ON(start & ~huge_page_mask(h));
5320 BUG_ON(end & ~huge_page_mask(h));
5321
5322 /*
5323 * This is a hugetlb vma, all the pte entries should point
5324 * to huge page.
5325 */
5326 tlb_change_page_size(tlb, sz);
5327 tlb_start_vma(tlb, vma);
5328
5329 last_addr_mask = hugetlb_mask_last_page(h);
5330 address = start;
5331 for (; address < end; address += sz) {
5332 ptep = hugetlb_walk(vma, address, sz);
5333 if (!ptep) {
5334 address |= last_addr_mask;
5335 continue;
5336 }
5337
5338 ptl = huge_pte_lock(h, mm, ptep);
5339 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5340 spin_unlock(ptl);
5341 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5342 force_flush = true;
5343 address |= last_addr_mask;
5344 continue;
5345 }
5346
5347 pte = huge_ptep_get(ptep);
5348 if (huge_pte_none(pte)) {
5349 spin_unlock(ptl);
5350 continue;
5351 }
5352
5353 /*
5354 * Migrating hugepage or HWPoisoned hugepage is already
5355 * unmapped and its refcount is dropped, so just clear pte here.
5356 */
5357 if (unlikely(!pte_present(pte))) {
5358 /*
5359 * If the pte was wr-protected by uffd-wp in any of the
5360 * swap forms, meanwhile the caller does not want to
5361 * drop the uffd-wp bit in this zap, then replace the
5362 * pte with a marker.
5363 */
5364 if (pte_swp_uffd_wp_any(pte) &&
5365 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5366 set_huge_pte_at(mm, address, ptep,
5367 make_pte_marker(PTE_MARKER_UFFD_WP),
5368 sz);
5369 else
5370 huge_pte_clear(mm, address, ptep, sz);
5371 spin_unlock(ptl);
5372 continue;
5373 }
5374
5375 page = pte_page(pte);
5376 /*
5377 * If a reference page is supplied, it is because a specific
5378 * page is being unmapped, not a range. Ensure the page we
5379 * are about to unmap is the actual page of interest.
5380 */
5381 if (ref_page) {
5382 if (page != ref_page) {
5383 spin_unlock(ptl);
5384 continue;
5385 }
5386 /*
5387 * Mark the VMA as having unmapped its page so that
5388 * future faults in this VMA will fail rather than
5389 * looking like data was lost
5390 */
5391 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5392 }
5393
5394 pte = huge_ptep_get_and_clear(mm, address, ptep);
5395 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5396 if (huge_pte_dirty(pte))
5397 set_page_dirty(page);
5398 /* Leave a uffd-wp pte marker if needed */
5399 if (huge_pte_uffd_wp(pte) &&
5400 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5401 set_huge_pte_at(mm, address, ptep,
5402 make_pte_marker(PTE_MARKER_UFFD_WP),
5403 sz);
5404 hugetlb_count_sub(pages_per_huge_page(h), mm);
5405 page_remove_rmap(page, vma, true);
5406
5407 spin_unlock(ptl);
5408 tlb_remove_page_size(tlb, page, huge_page_size(h));
5409 /*
5410 * Bail out after unmapping reference page if supplied
5411 */
5412 if (ref_page)
5413 break;
5414 }
5415 tlb_end_vma(tlb, vma);
5416
5417 /*
5418 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5419 * could defer the flush until now, since by holding i_mmap_rwsem we
5420 * guaranteed that the last refernece would not be dropped. But we must
5421 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5422 * dropped and the last reference to the shared PMDs page might be
5423 * dropped as well.
5424 *
5425 * In theory we could defer the freeing of the PMD pages as well, but
5426 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5427 * detect sharing, so we cannot defer the release of the page either.
5428 * Instead, do flush now.
5429 */
5430 if (force_flush)
5431 tlb_flush_mmu_tlbonly(tlb);
5432 }
5433
__hugetlb_zap_begin(struct vm_area_struct * vma,unsigned long * start,unsigned long * end)5434 void __hugetlb_zap_begin(struct vm_area_struct *vma,
5435 unsigned long *start, unsigned long *end)
5436 {
5437 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5438 return;
5439
5440 adjust_range_if_pmd_sharing_possible(vma, start, end);
5441 hugetlb_vma_lock_write(vma);
5442 if (vma->vm_file)
5443 i_mmap_lock_write(vma->vm_file->f_mapping);
5444 }
5445
__hugetlb_zap_end(struct vm_area_struct * vma,struct zap_details * details)5446 void __hugetlb_zap_end(struct vm_area_struct *vma,
5447 struct zap_details *details)
5448 {
5449 zap_flags_t zap_flags = details ? details->zap_flags : 0;
5450
5451 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5452 return;
5453
5454 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5455 /*
5456 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5457 * When the vma_lock is freed, this makes the vma ineligible
5458 * for pmd sharing. And, i_mmap_rwsem is required to set up
5459 * pmd sharing. This is important as page tables for this
5460 * unmapped range will be asynchrously deleted. If the page
5461 * tables are shared, there will be issues when accessed by
5462 * someone else.
5463 */
5464 __hugetlb_vma_unlock_write_free(vma);
5465 } else {
5466 hugetlb_vma_unlock_write(vma);
5467 }
5468
5469 if (vma->vm_file)
5470 i_mmap_unlock_write(vma->vm_file->f_mapping);
5471 }
5472
unmap_hugepage_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,struct page * ref_page,zap_flags_t zap_flags)5473 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5474 unsigned long end, struct page *ref_page,
5475 zap_flags_t zap_flags)
5476 {
5477 struct mmu_notifier_range range;
5478 struct mmu_gather tlb;
5479
5480 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5481 start, end);
5482 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5483 mmu_notifier_invalidate_range_start(&range);
5484 tlb_gather_mmu(&tlb, vma->vm_mm);
5485
5486 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5487
5488 mmu_notifier_invalidate_range_end(&range);
5489 tlb_finish_mmu(&tlb);
5490 }
5491
5492 /*
5493 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5494 * mapping it owns the reserve page for. The intention is to unmap the page
5495 * from other VMAs and let the children be SIGKILLed if they are faulting the
5496 * same region.
5497 */
unmap_ref_private(struct mm_struct * mm,struct vm_area_struct * vma,struct page * page,unsigned long address)5498 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5499 struct page *page, unsigned long address)
5500 {
5501 struct hstate *h = hstate_vma(vma);
5502 struct vm_area_struct *iter_vma;
5503 struct address_space *mapping;
5504 pgoff_t pgoff;
5505
5506 /*
5507 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5508 * from page cache lookup which is in HPAGE_SIZE units.
5509 */
5510 address = address & huge_page_mask(h);
5511 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5512 vma->vm_pgoff;
5513 mapping = vma->vm_file->f_mapping;
5514
5515 /*
5516 * Take the mapping lock for the duration of the table walk. As
5517 * this mapping should be shared between all the VMAs,
5518 * __unmap_hugepage_range() is called as the lock is already held
5519 */
5520 i_mmap_lock_write(mapping);
5521 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5522 /* Do not unmap the current VMA */
5523 if (iter_vma == vma)
5524 continue;
5525
5526 /*
5527 * Shared VMAs have their own reserves and do not affect
5528 * MAP_PRIVATE accounting but it is possible that a shared
5529 * VMA is using the same page so check and skip such VMAs.
5530 */
5531 if (iter_vma->vm_flags & VM_MAYSHARE)
5532 continue;
5533
5534 /*
5535 * Unmap the page from other VMAs without their own reserves.
5536 * They get marked to be SIGKILLed if they fault in these
5537 * areas. This is because a future no-page fault on this VMA
5538 * could insert a zeroed page instead of the data existing
5539 * from the time of fork. This would look like data corruption
5540 */
5541 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5542 unmap_hugepage_range(iter_vma, address,
5543 address + huge_page_size(h), page, 0);
5544 }
5545 i_mmap_unlock_write(mapping);
5546 }
5547
5548 /*
5549 * hugetlb_wp() should be called with page lock of the original hugepage held.
5550 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5551 * cannot race with other handlers or page migration.
5552 * Keep the pte_same checks anyway to make transition from the mutex easier.
5553 */
hugetlb_wp(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,pte_t * ptep,unsigned int flags,struct folio * pagecache_folio,spinlock_t * ptl)5554 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5555 unsigned long address, pte_t *ptep, unsigned int flags,
5556 struct folio *pagecache_folio, spinlock_t *ptl)
5557 {
5558 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5559 pte_t pte = huge_ptep_get(ptep);
5560 struct hstate *h = hstate_vma(vma);
5561 struct folio *old_folio;
5562 struct folio *new_folio;
5563 int outside_reserve = 0;
5564 vm_fault_t ret = 0;
5565 unsigned long haddr = address & huge_page_mask(h);
5566 struct mmu_notifier_range range;
5567
5568 /*
5569 * Never handle CoW for uffd-wp protected pages. It should be only
5570 * handled when the uffd-wp protection is removed.
5571 *
5572 * Note that only the CoW optimization path (in hugetlb_no_page())
5573 * can trigger this, because hugetlb_fault() will always resolve
5574 * uffd-wp bit first.
5575 */
5576 if (!unshare && huge_pte_uffd_wp(pte))
5577 return 0;
5578
5579 /*
5580 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5581 * PTE mapped R/O such as maybe_mkwrite() would do.
5582 */
5583 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5584 return VM_FAULT_SIGSEGV;
5585
5586 /* Let's take out MAP_SHARED mappings first. */
5587 if (vma->vm_flags & VM_MAYSHARE) {
5588 set_huge_ptep_writable(vma, haddr, ptep);
5589 return 0;
5590 }
5591
5592 old_folio = page_folio(pte_page(pte));
5593
5594 delayacct_wpcopy_start();
5595
5596 retry_avoidcopy:
5597 /*
5598 * If no-one else is actually using this page, we're the exclusive
5599 * owner and can reuse this page.
5600 */
5601 if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5602 if (!PageAnonExclusive(&old_folio->page))
5603 page_move_anon_rmap(&old_folio->page, vma);
5604 if (likely(!unshare))
5605 set_huge_ptep_writable(vma, haddr, ptep);
5606
5607 delayacct_wpcopy_end();
5608 return 0;
5609 }
5610 VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5611 PageAnonExclusive(&old_folio->page), &old_folio->page);
5612
5613 /*
5614 * If the process that created a MAP_PRIVATE mapping is about to
5615 * perform a COW due to a shared page count, attempt to satisfy
5616 * the allocation without using the existing reserves. The pagecache
5617 * page is used to determine if the reserve at this address was
5618 * consumed or not. If reserves were used, a partial faulted mapping
5619 * at the time of fork() could consume its reserves on COW instead
5620 * of the full address range.
5621 */
5622 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5623 old_folio != pagecache_folio)
5624 outside_reserve = 1;
5625
5626 folio_get(old_folio);
5627
5628 /*
5629 * Drop page table lock as buddy allocator may be called. It will
5630 * be acquired again before returning to the caller, as expected.
5631 */
5632 spin_unlock(ptl);
5633 new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
5634
5635 if (IS_ERR(new_folio)) {
5636 /*
5637 * If a process owning a MAP_PRIVATE mapping fails to COW,
5638 * it is due to references held by a child and an insufficient
5639 * huge page pool. To guarantee the original mappers
5640 * reliability, unmap the page from child processes. The child
5641 * may get SIGKILLed if it later faults.
5642 */
5643 if (outside_reserve) {
5644 struct address_space *mapping = vma->vm_file->f_mapping;
5645 pgoff_t idx;
5646 u32 hash;
5647
5648 folio_put(old_folio);
5649 /*
5650 * Drop hugetlb_fault_mutex and vma_lock before
5651 * unmapping. unmapping needs to hold vma_lock
5652 * in write mode. Dropping vma_lock in read mode
5653 * here is OK as COW mappings do not interact with
5654 * PMD sharing.
5655 *
5656 * Reacquire both after unmap operation.
5657 */
5658 idx = vma_hugecache_offset(h, vma, haddr);
5659 hash = hugetlb_fault_mutex_hash(mapping, idx);
5660 hugetlb_vma_unlock_read(vma);
5661 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5662
5663 unmap_ref_private(mm, vma, &old_folio->page, haddr);
5664
5665 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5666 hugetlb_vma_lock_read(vma);
5667 spin_lock(ptl);
5668 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5669 if (likely(ptep &&
5670 pte_same(huge_ptep_get(ptep), pte)))
5671 goto retry_avoidcopy;
5672 /*
5673 * race occurs while re-acquiring page table
5674 * lock, and our job is done.
5675 */
5676 delayacct_wpcopy_end();
5677 return 0;
5678 }
5679
5680 ret = vmf_error(PTR_ERR(new_folio));
5681 goto out_release_old;
5682 }
5683
5684 /*
5685 * When the original hugepage is shared one, it does not have
5686 * anon_vma prepared.
5687 */
5688 if (unlikely(anon_vma_prepare(vma))) {
5689 ret = VM_FAULT_OOM;
5690 goto out_release_all;
5691 }
5692
5693 if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
5694 ret = VM_FAULT_HWPOISON_LARGE;
5695 goto out_release_all;
5696 }
5697 __folio_mark_uptodate(new_folio);
5698
5699 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
5700 haddr + huge_page_size(h));
5701 mmu_notifier_invalidate_range_start(&range);
5702
5703 /*
5704 * Retake the page table lock to check for racing updates
5705 * before the page tables are altered
5706 */
5707 spin_lock(ptl);
5708 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5709 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5710 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
5711
5712 /* Break COW or unshare */
5713 huge_ptep_clear_flush(vma, haddr, ptep);
5714 page_remove_rmap(&old_folio->page, vma, true);
5715 hugepage_add_new_anon_rmap(new_folio, vma, haddr);
5716 if (huge_pte_uffd_wp(pte))
5717 newpte = huge_pte_mkuffd_wp(newpte);
5718 set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
5719 folio_set_hugetlb_migratable(new_folio);
5720 /* Make the old page be freed below */
5721 new_folio = old_folio;
5722 }
5723 spin_unlock(ptl);
5724 mmu_notifier_invalidate_range_end(&range);
5725 out_release_all:
5726 /*
5727 * No restore in case of successful pagetable update (Break COW or
5728 * unshare)
5729 */
5730 if (new_folio != old_folio)
5731 restore_reserve_on_error(h, vma, haddr, new_folio);
5732 folio_put(new_folio);
5733 out_release_old:
5734 folio_put(old_folio);
5735
5736 spin_lock(ptl); /* Caller expects lock to be held */
5737
5738 delayacct_wpcopy_end();
5739 return ret;
5740 }
5741
5742 /*
5743 * Return whether there is a pagecache page to back given address within VMA.
5744 */
hugetlbfs_pagecache_present(struct hstate * h,struct vm_area_struct * vma,unsigned long address)5745 static bool hugetlbfs_pagecache_present(struct hstate *h,
5746 struct vm_area_struct *vma, unsigned long address)
5747 {
5748 struct address_space *mapping = vma->vm_file->f_mapping;
5749 pgoff_t idx = vma_hugecache_offset(h, vma, address);
5750 struct folio *folio;
5751
5752 folio = filemap_get_folio(mapping, idx);
5753 if (IS_ERR(folio))
5754 return false;
5755 folio_put(folio);
5756 return true;
5757 }
5758
hugetlb_add_to_page_cache(struct folio * folio,struct address_space * mapping,pgoff_t idx)5759 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
5760 pgoff_t idx)
5761 {
5762 struct inode *inode = mapping->host;
5763 struct hstate *h = hstate_inode(inode);
5764 int err;
5765
5766 __folio_set_locked(folio);
5767 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5768
5769 if (unlikely(err)) {
5770 __folio_clear_locked(folio);
5771 return err;
5772 }
5773 folio_clear_hugetlb_restore_reserve(folio);
5774
5775 /*
5776 * mark folio dirty so that it will not be removed from cache/file
5777 * by non-hugetlbfs specific code paths.
5778 */
5779 folio_mark_dirty(folio);
5780
5781 spin_lock(&inode->i_lock);
5782 inode->i_blocks += blocks_per_huge_page(h);
5783 spin_unlock(&inode->i_lock);
5784 return 0;
5785 }
5786
hugetlb_handle_userfault(struct vm_area_struct * vma,struct address_space * mapping,pgoff_t idx,unsigned int flags,unsigned long haddr,unsigned long addr,unsigned long reason)5787 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5788 struct address_space *mapping,
5789 pgoff_t idx,
5790 unsigned int flags,
5791 unsigned long haddr,
5792 unsigned long addr,
5793 unsigned long reason)
5794 {
5795 u32 hash;
5796 struct vm_fault vmf = {
5797 .vma = vma,
5798 .address = haddr,
5799 .real_address = addr,
5800 .flags = flags,
5801
5802 /*
5803 * Hard to debug if it ends up being
5804 * used by a callee that assumes
5805 * something about the other
5806 * uninitialized fields... same as in
5807 * memory.c
5808 */
5809 };
5810
5811 /*
5812 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5813 * userfault. Also mmap_lock could be dropped due to handling
5814 * userfault, any vma operation should be careful from here.
5815 */
5816 hugetlb_vma_unlock_read(vma);
5817 hash = hugetlb_fault_mutex_hash(mapping, idx);
5818 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5819 return handle_userfault(&vmf, reason);
5820 }
5821
5822 /*
5823 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5824 * false if pte changed or is changing.
5825 */
hugetlb_pte_stable(struct hstate * h,struct mm_struct * mm,pte_t * ptep,pte_t old_pte)5826 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5827 pte_t *ptep, pte_t old_pte)
5828 {
5829 spinlock_t *ptl;
5830 bool same;
5831
5832 ptl = huge_pte_lock(h, mm, ptep);
5833 same = pte_same(huge_ptep_get(ptep), old_pte);
5834 spin_unlock(ptl);
5835
5836 return same;
5837 }
5838
hugetlb_no_page(struct mm_struct * mm,struct vm_area_struct * vma,struct address_space * mapping,pgoff_t idx,unsigned long address,pte_t * ptep,pte_t old_pte,unsigned int flags)5839 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5840 struct vm_area_struct *vma,
5841 struct address_space *mapping, pgoff_t idx,
5842 unsigned long address, pte_t *ptep,
5843 pte_t old_pte, unsigned int flags)
5844 {
5845 struct hstate *h = hstate_vma(vma);
5846 vm_fault_t ret = VM_FAULT_SIGBUS;
5847 int anon_rmap = 0;
5848 unsigned long size;
5849 struct folio *folio;
5850 pte_t new_pte;
5851 spinlock_t *ptl;
5852 unsigned long haddr = address & huge_page_mask(h);
5853 bool new_folio, new_pagecache_folio = false;
5854 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5855
5856 /*
5857 * Currently, we are forced to kill the process in the event the
5858 * original mapper has unmapped pages from the child due to a failed
5859 * COW/unsharing. Warn that such a situation has occurred as it may not
5860 * be obvious.
5861 */
5862 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5863 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5864 current->pid);
5865 goto out;
5866 }
5867
5868 /*
5869 * Use page lock to guard against racing truncation
5870 * before we get page_table_lock.
5871 */
5872 new_folio = false;
5873 folio = filemap_lock_folio(mapping, idx);
5874 if (IS_ERR(folio)) {
5875 size = i_size_read(mapping->host) >> huge_page_shift(h);
5876 if (idx >= size)
5877 goto out;
5878 /* Check for page in userfault range */
5879 if (userfaultfd_missing(vma)) {
5880 /*
5881 * Since hugetlb_no_page() was examining pte
5882 * without pgtable lock, we need to re-test under
5883 * lock because the pte may not be stable and could
5884 * have changed from under us. Try to detect
5885 * either changed or during-changing ptes and retry
5886 * properly when needed.
5887 *
5888 * Note that userfaultfd is actually fine with
5889 * false positives (e.g. caused by pte changed),
5890 * but not wrong logical events (e.g. caused by
5891 * reading a pte during changing). The latter can
5892 * confuse the userspace, so the strictness is very
5893 * much preferred. E.g., MISSING event should
5894 * never happen on the page after UFFDIO_COPY has
5895 * correctly installed the page and returned.
5896 */
5897 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5898 ret = 0;
5899 goto out;
5900 }
5901
5902 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5903 haddr, address,
5904 VM_UFFD_MISSING);
5905 }
5906
5907 folio = alloc_hugetlb_folio(vma, haddr, 0);
5908 if (IS_ERR(folio)) {
5909 /*
5910 * Returning error will result in faulting task being
5911 * sent SIGBUS. The hugetlb fault mutex prevents two
5912 * tasks from racing to fault in the same page which
5913 * could result in false unable to allocate errors.
5914 * Page migration does not take the fault mutex, but
5915 * does a clear then write of pte's under page table
5916 * lock. Page fault code could race with migration,
5917 * notice the clear pte and try to allocate a page
5918 * here. Before returning error, get ptl and make
5919 * sure there really is no pte entry.
5920 */
5921 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5922 ret = vmf_error(PTR_ERR(folio));
5923 else
5924 ret = 0;
5925 goto out;
5926 }
5927 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
5928 __folio_mark_uptodate(folio);
5929 new_folio = true;
5930
5931 if (vma->vm_flags & VM_MAYSHARE) {
5932 int err = hugetlb_add_to_page_cache(folio, mapping, idx);
5933 if (err) {
5934 /*
5935 * err can't be -EEXIST which implies someone
5936 * else consumed the reservation since hugetlb
5937 * fault mutex is held when add a hugetlb page
5938 * to the page cache. So it's safe to call
5939 * restore_reserve_on_error() here.
5940 */
5941 restore_reserve_on_error(h, vma, haddr, folio);
5942 folio_put(folio);
5943 goto out;
5944 }
5945 new_pagecache_folio = true;
5946 } else {
5947 folio_lock(folio);
5948 if (unlikely(anon_vma_prepare(vma))) {
5949 ret = VM_FAULT_OOM;
5950 goto backout_unlocked;
5951 }
5952 anon_rmap = 1;
5953 }
5954 } else {
5955 /*
5956 * If memory error occurs between mmap() and fault, some process
5957 * don't have hwpoisoned swap entry for errored virtual address.
5958 * So we need to block hugepage fault by PG_hwpoison bit check.
5959 */
5960 if (unlikely(folio_test_hwpoison(folio))) {
5961 ret = VM_FAULT_HWPOISON_LARGE |
5962 VM_FAULT_SET_HINDEX(hstate_index(h));
5963 goto backout_unlocked;
5964 }
5965
5966 /* Check for page in userfault range. */
5967 if (userfaultfd_minor(vma)) {
5968 folio_unlock(folio);
5969 folio_put(folio);
5970 /* See comment in userfaultfd_missing() block above */
5971 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5972 ret = 0;
5973 goto out;
5974 }
5975 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5976 haddr, address,
5977 VM_UFFD_MINOR);
5978 }
5979 }
5980
5981 /*
5982 * If we are going to COW a private mapping later, we examine the
5983 * pending reservations for this page now. This will ensure that
5984 * any allocations necessary to record that reservation occur outside
5985 * the spinlock.
5986 */
5987 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5988 if (vma_needs_reservation(h, vma, haddr) < 0) {
5989 ret = VM_FAULT_OOM;
5990 goto backout_unlocked;
5991 }
5992 /* Just decrements count, does not deallocate */
5993 vma_end_reservation(h, vma, haddr);
5994 }
5995
5996 ptl = huge_pte_lock(h, mm, ptep);
5997 ret = 0;
5998 /* If pte changed from under us, retry */
5999 if (!pte_same(huge_ptep_get(ptep), old_pte))
6000 goto backout;
6001
6002 if (anon_rmap)
6003 hugepage_add_new_anon_rmap(folio, vma, haddr);
6004 else
6005 page_dup_file_rmap(&folio->page, true);
6006 new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
6007 && (vma->vm_flags & VM_SHARED)));
6008 /*
6009 * If this pte was previously wr-protected, keep it wr-protected even
6010 * if populated.
6011 */
6012 if (unlikely(pte_marker_uffd_wp(old_pte)))
6013 new_pte = huge_pte_mkuffd_wp(new_pte);
6014 set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
6015
6016 hugetlb_count_add(pages_per_huge_page(h), mm);
6017 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6018 /* Optimization, do the COW without a second fault */
6019 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl);
6020 }
6021
6022 spin_unlock(ptl);
6023
6024 /*
6025 * Only set hugetlb_migratable in newly allocated pages. Existing pages
6026 * found in the pagecache may not have hugetlb_migratable if they have
6027 * been isolated for migration.
6028 */
6029 if (new_folio)
6030 folio_set_hugetlb_migratable(folio);
6031
6032 folio_unlock(folio);
6033 out:
6034 hugetlb_vma_unlock_read(vma);
6035 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6036 return ret;
6037
6038 backout:
6039 spin_unlock(ptl);
6040 backout_unlocked:
6041 if (new_folio && !new_pagecache_folio)
6042 restore_reserve_on_error(h, vma, haddr, folio);
6043
6044 folio_unlock(folio);
6045 folio_put(folio);
6046 goto out;
6047 }
6048
6049 #ifdef CONFIG_SMP
hugetlb_fault_mutex_hash(struct address_space * mapping,pgoff_t idx)6050 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6051 {
6052 unsigned long key[2];
6053 u32 hash;
6054
6055 key[0] = (unsigned long) mapping;
6056 key[1] = idx;
6057
6058 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6059
6060 return hash & (num_fault_mutexes - 1);
6061 }
6062 #else
6063 /*
6064 * For uniprocessor systems we always use a single mutex, so just
6065 * return 0 and avoid the hashing overhead.
6066 */
hugetlb_fault_mutex_hash(struct address_space * mapping,pgoff_t idx)6067 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6068 {
6069 return 0;
6070 }
6071 #endif
6072
hugetlb_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,unsigned int flags)6073 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6074 unsigned long address, unsigned int flags)
6075 {
6076 pte_t *ptep, entry;
6077 spinlock_t *ptl;
6078 vm_fault_t ret;
6079 u32 hash;
6080 pgoff_t idx;
6081 struct folio *folio = NULL;
6082 struct folio *pagecache_folio = NULL;
6083 struct hstate *h = hstate_vma(vma);
6084 struct address_space *mapping;
6085 int need_wait_lock = 0;
6086 unsigned long haddr = address & huge_page_mask(h);
6087
6088 /* TODO: Handle faults under the VMA lock */
6089 if (flags & FAULT_FLAG_VMA_LOCK) {
6090 vma_end_read(vma);
6091 return VM_FAULT_RETRY;
6092 }
6093
6094 /*
6095 * Serialize hugepage allocation and instantiation, so that we don't
6096 * get spurious allocation failures if two CPUs race to instantiate
6097 * the same page in the page cache.
6098 */
6099 mapping = vma->vm_file->f_mapping;
6100 idx = vma_hugecache_offset(h, vma, haddr);
6101 hash = hugetlb_fault_mutex_hash(mapping, idx);
6102 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6103
6104 /*
6105 * Acquire vma lock before calling huge_pte_alloc and hold
6106 * until finished with ptep. This prevents huge_pmd_unshare from
6107 * being called elsewhere and making the ptep no longer valid.
6108 */
6109 hugetlb_vma_lock_read(vma);
6110 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6111 if (!ptep) {
6112 hugetlb_vma_unlock_read(vma);
6113 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6114 return VM_FAULT_OOM;
6115 }
6116
6117 entry = huge_ptep_get(ptep);
6118 if (huge_pte_none_mostly(entry)) {
6119 if (is_pte_marker(entry)) {
6120 pte_marker marker =
6121 pte_marker_get(pte_to_swp_entry(entry));
6122
6123 if (marker & PTE_MARKER_POISONED) {
6124 ret = VM_FAULT_HWPOISON_LARGE;
6125 goto out_mutex;
6126 }
6127 }
6128
6129 /*
6130 * Other PTE markers should be handled the same way as none PTE.
6131 *
6132 * hugetlb_no_page will drop vma lock and hugetlb fault
6133 * mutex internally, which make us return immediately.
6134 */
6135 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6136 entry, flags);
6137 }
6138
6139 ret = 0;
6140
6141 /*
6142 * entry could be a migration/hwpoison entry at this point, so this
6143 * check prevents the kernel from going below assuming that we have
6144 * an active hugepage in pagecache. This goto expects the 2nd page
6145 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6146 * properly handle it.
6147 */
6148 if (!pte_present(entry)) {
6149 if (unlikely(is_hugetlb_entry_migration(entry))) {
6150 /*
6151 * Release the hugetlb fault lock now, but retain
6152 * the vma lock, because it is needed to guard the
6153 * huge_pte_lockptr() later in
6154 * migration_entry_wait_huge(). The vma lock will
6155 * be released there.
6156 */
6157 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6158 migration_entry_wait_huge(vma, ptep);
6159 return 0;
6160 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6161 ret = VM_FAULT_HWPOISON_LARGE |
6162 VM_FAULT_SET_HINDEX(hstate_index(h));
6163 goto out_mutex;
6164 }
6165
6166 /*
6167 * If we are going to COW/unshare the mapping later, we examine the
6168 * pending reservations for this page now. This will ensure that any
6169 * allocations necessary to record that reservation occur outside the
6170 * spinlock. Also lookup the pagecache page now as it is used to
6171 * determine if a reservation has been consumed.
6172 */
6173 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6174 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6175 if (vma_needs_reservation(h, vma, haddr) < 0) {
6176 ret = VM_FAULT_OOM;
6177 goto out_mutex;
6178 }
6179 /* Just decrements count, does not deallocate */
6180 vma_end_reservation(h, vma, haddr);
6181
6182 pagecache_folio = filemap_lock_folio(mapping, idx);
6183 if (IS_ERR(pagecache_folio))
6184 pagecache_folio = NULL;
6185 }
6186
6187 ptl = huge_pte_lock(h, mm, ptep);
6188
6189 /* Check for a racing update before calling hugetlb_wp() */
6190 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6191 goto out_ptl;
6192
6193 /* Handle userfault-wp first, before trying to lock more pages */
6194 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6195 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6196 struct vm_fault vmf = {
6197 .vma = vma,
6198 .address = haddr,
6199 .real_address = address,
6200 .flags = flags,
6201 };
6202
6203 spin_unlock(ptl);
6204 if (pagecache_folio) {
6205 folio_unlock(pagecache_folio);
6206 folio_put(pagecache_folio);
6207 }
6208 hugetlb_vma_unlock_read(vma);
6209 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6210 return handle_userfault(&vmf, VM_UFFD_WP);
6211 }
6212
6213 /*
6214 * hugetlb_wp() requires page locks of pte_page(entry) and
6215 * pagecache_folio, so here we need take the former one
6216 * when folio != pagecache_folio or !pagecache_folio.
6217 */
6218 folio = page_folio(pte_page(entry));
6219 if (folio != pagecache_folio)
6220 if (!folio_trylock(folio)) {
6221 need_wait_lock = 1;
6222 goto out_ptl;
6223 }
6224
6225 folio_get(folio);
6226
6227 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6228 if (!huge_pte_write(entry)) {
6229 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6230 pagecache_folio, ptl);
6231 goto out_put_page;
6232 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6233 entry = huge_pte_mkdirty(entry);
6234 }
6235 }
6236 entry = pte_mkyoung(entry);
6237 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6238 flags & FAULT_FLAG_WRITE))
6239 update_mmu_cache(vma, haddr, ptep);
6240 out_put_page:
6241 if (folio != pagecache_folio)
6242 folio_unlock(folio);
6243 folio_put(folio);
6244 out_ptl:
6245 spin_unlock(ptl);
6246
6247 if (pagecache_folio) {
6248 folio_unlock(pagecache_folio);
6249 folio_put(pagecache_folio);
6250 }
6251 out_mutex:
6252 hugetlb_vma_unlock_read(vma);
6253 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6254 /*
6255 * Generally it's safe to hold refcount during waiting page lock. But
6256 * here we just wait to defer the next page fault to avoid busy loop and
6257 * the page is not used after unlocked before returning from the current
6258 * page fault. So we are safe from accessing freed page, even if we wait
6259 * here without taking refcount.
6260 */
6261 if (need_wait_lock)
6262 folio_wait_locked(folio);
6263 return ret;
6264 }
6265
6266 #ifdef CONFIG_USERFAULTFD
6267 /*
6268 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6269 * with modifications for hugetlb pages.
6270 */
hugetlb_mfill_atomic_pte(pte_t * dst_pte,struct vm_area_struct * dst_vma,unsigned long dst_addr,unsigned long src_addr,uffd_flags_t flags,struct folio ** foliop)6271 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6272 struct vm_area_struct *dst_vma,
6273 unsigned long dst_addr,
6274 unsigned long src_addr,
6275 uffd_flags_t flags,
6276 struct folio **foliop)
6277 {
6278 struct mm_struct *dst_mm = dst_vma->vm_mm;
6279 bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6280 bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6281 struct hstate *h = hstate_vma(dst_vma);
6282 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6283 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6284 unsigned long size;
6285 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6286 pte_t _dst_pte;
6287 spinlock_t *ptl;
6288 int ret = -ENOMEM;
6289 struct folio *folio;
6290 int writable;
6291 bool folio_in_pagecache = false;
6292
6293 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6294 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6295
6296 /* Don't overwrite any existing PTEs (even markers) */
6297 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6298 spin_unlock(ptl);
6299 return -EEXIST;
6300 }
6301
6302 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6303 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6304 huge_page_size(h));
6305
6306 /* No need to invalidate - it was non-present before */
6307 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6308
6309 spin_unlock(ptl);
6310 return 0;
6311 }
6312
6313 if (is_continue) {
6314 ret = -EFAULT;
6315 folio = filemap_lock_folio(mapping, idx);
6316 if (IS_ERR(folio))
6317 goto out;
6318 folio_in_pagecache = true;
6319 } else if (!*foliop) {
6320 /* If a folio already exists, then it's UFFDIO_COPY for
6321 * a non-missing case. Return -EEXIST.
6322 */
6323 if (vm_shared &&
6324 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6325 ret = -EEXIST;
6326 goto out;
6327 }
6328
6329 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6330 if (IS_ERR(folio)) {
6331 ret = -ENOMEM;
6332 goto out;
6333 }
6334
6335 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6336 false);
6337
6338 /* fallback to copy_from_user outside mmap_lock */
6339 if (unlikely(ret)) {
6340 ret = -ENOENT;
6341 /* Free the allocated folio which may have
6342 * consumed a reservation.
6343 */
6344 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6345 folio_put(folio);
6346
6347 /* Allocate a temporary folio to hold the copied
6348 * contents.
6349 */
6350 folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6351 if (!folio) {
6352 ret = -ENOMEM;
6353 goto out;
6354 }
6355 *foliop = folio;
6356 /* Set the outparam foliop and return to the caller to
6357 * copy the contents outside the lock. Don't free the
6358 * folio.
6359 */
6360 goto out;
6361 }
6362 } else {
6363 if (vm_shared &&
6364 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6365 folio_put(*foliop);
6366 ret = -EEXIST;
6367 *foliop = NULL;
6368 goto out;
6369 }
6370
6371 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6372 if (IS_ERR(folio)) {
6373 folio_put(*foliop);
6374 ret = -ENOMEM;
6375 *foliop = NULL;
6376 goto out;
6377 }
6378 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6379 folio_put(*foliop);
6380 *foliop = NULL;
6381 if (ret) {
6382 folio_put(folio);
6383 goto out;
6384 }
6385 }
6386
6387 /*
6388 * The memory barrier inside __folio_mark_uptodate makes sure that
6389 * preceding stores to the page contents become visible before
6390 * the set_pte_at() write.
6391 */
6392 __folio_mark_uptodate(folio);
6393
6394 /* Add shared, newly allocated pages to the page cache. */
6395 if (vm_shared && !is_continue) {
6396 size = i_size_read(mapping->host) >> huge_page_shift(h);
6397 ret = -EFAULT;
6398 if (idx >= size)
6399 goto out_release_nounlock;
6400
6401 /*
6402 * Serialization between remove_inode_hugepages() and
6403 * hugetlb_add_to_page_cache() below happens through the
6404 * hugetlb_fault_mutex_table that here must be hold by
6405 * the caller.
6406 */
6407 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6408 if (ret)
6409 goto out_release_nounlock;
6410 folio_in_pagecache = true;
6411 }
6412
6413 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6414
6415 ret = -EIO;
6416 if (folio_test_hwpoison(folio))
6417 goto out_release_unlock;
6418
6419 /*
6420 * We allow to overwrite a pte marker: consider when both MISSING|WP
6421 * registered, we firstly wr-protect a none pte which has no page cache
6422 * page backing it, then access the page.
6423 */
6424 ret = -EEXIST;
6425 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6426 goto out_release_unlock;
6427
6428 if (folio_in_pagecache)
6429 page_dup_file_rmap(&folio->page, true);
6430 else
6431 hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr);
6432
6433 /*
6434 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6435 * with wp flag set, don't set pte write bit.
6436 */
6437 if (wp_enabled || (is_continue && !vm_shared))
6438 writable = 0;
6439 else
6440 writable = dst_vma->vm_flags & VM_WRITE;
6441
6442 _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6443 /*
6444 * Always mark UFFDIO_COPY page dirty; note that this may not be
6445 * extremely important for hugetlbfs for now since swapping is not
6446 * supported, but we should still be clear in that this page cannot be
6447 * thrown away at will, even if write bit not set.
6448 */
6449 _dst_pte = huge_pte_mkdirty(_dst_pte);
6450 _dst_pte = pte_mkyoung(_dst_pte);
6451
6452 if (wp_enabled)
6453 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6454
6455 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6456
6457 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6458
6459 /* No need to invalidate - it was non-present before */
6460 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6461
6462 spin_unlock(ptl);
6463 if (!is_continue)
6464 folio_set_hugetlb_migratable(folio);
6465 if (vm_shared || is_continue)
6466 folio_unlock(folio);
6467 ret = 0;
6468 out:
6469 return ret;
6470 out_release_unlock:
6471 spin_unlock(ptl);
6472 if (vm_shared || is_continue)
6473 folio_unlock(folio);
6474 out_release_nounlock:
6475 if (!folio_in_pagecache)
6476 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6477 folio_put(folio);
6478 goto out;
6479 }
6480 #endif /* CONFIG_USERFAULTFD */
6481
hugetlb_follow_page_mask(struct vm_area_struct * vma,unsigned long address,unsigned int flags,unsigned int * page_mask)6482 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6483 unsigned long address, unsigned int flags,
6484 unsigned int *page_mask)
6485 {
6486 struct hstate *h = hstate_vma(vma);
6487 struct mm_struct *mm = vma->vm_mm;
6488 unsigned long haddr = address & huge_page_mask(h);
6489 struct page *page = NULL;
6490 spinlock_t *ptl;
6491 pte_t *pte, entry;
6492 int ret;
6493
6494 hugetlb_vma_lock_read(vma);
6495 pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6496 if (!pte)
6497 goto out_unlock;
6498
6499 ptl = huge_pte_lock(h, mm, pte);
6500 entry = huge_ptep_get(pte);
6501 if (pte_present(entry)) {
6502 page = pte_page(entry);
6503
6504 if (!huge_pte_write(entry)) {
6505 if (flags & FOLL_WRITE) {
6506 page = NULL;
6507 goto out;
6508 }
6509
6510 if (gup_must_unshare(vma, flags, page)) {
6511 /* Tell the caller to do unsharing */
6512 page = ERR_PTR(-EMLINK);
6513 goto out;
6514 }
6515 }
6516
6517 page = nth_page(page, ((address & ~huge_page_mask(h)) >> PAGE_SHIFT));
6518
6519 /*
6520 * Note that page may be a sub-page, and with vmemmap
6521 * optimizations the page struct may be read only.
6522 * try_grab_page() will increase the ref count on the
6523 * head page, so this will be OK.
6524 *
6525 * try_grab_page() should always be able to get the page here,
6526 * because we hold the ptl lock and have verified pte_present().
6527 */
6528 ret = try_grab_page(page, flags);
6529
6530 if (WARN_ON_ONCE(ret)) {
6531 page = ERR_PTR(ret);
6532 goto out;
6533 }
6534
6535 *page_mask = (1U << huge_page_order(h)) - 1;
6536 }
6537 out:
6538 spin_unlock(ptl);
6539 out_unlock:
6540 hugetlb_vma_unlock_read(vma);
6541
6542 /*
6543 * Fixup retval for dump requests: if pagecache doesn't exist,
6544 * don't try to allocate a new page but just skip it.
6545 */
6546 if (!page && (flags & FOLL_DUMP) &&
6547 !hugetlbfs_pagecache_present(h, vma, address))
6548 page = ERR_PTR(-EFAULT);
6549
6550 return page;
6551 }
6552
hugetlb_change_protection(struct vm_area_struct * vma,unsigned long address,unsigned long end,pgprot_t newprot,unsigned long cp_flags)6553 long hugetlb_change_protection(struct vm_area_struct *vma,
6554 unsigned long address, unsigned long end,
6555 pgprot_t newprot, unsigned long cp_flags)
6556 {
6557 struct mm_struct *mm = vma->vm_mm;
6558 unsigned long start = address;
6559 pte_t *ptep;
6560 pte_t pte;
6561 struct hstate *h = hstate_vma(vma);
6562 long pages = 0, psize = huge_page_size(h);
6563 bool shared_pmd = false;
6564 struct mmu_notifier_range range;
6565 unsigned long last_addr_mask;
6566 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6567 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6568
6569 /*
6570 * In the case of shared PMDs, the area to flush could be beyond
6571 * start/end. Set range.start/range.end to cover the maximum possible
6572 * range if PMD sharing is possible.
6573 */
6574 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6575 0, mm, start, end);
6576 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6577
6578 BUG_ON(address >= end);
6579 flush_cache_range(vma, range.start, range.end);
6580
6581 mmu_notifier_invalidate_range_start(&range);
6582 hugetlb_vma_lock_write(vma);
6583 i_mmap_lock_write(vma->vm_file->f_mapping);
6584 last_addr_mask = hugetlb_mask_last_page(h);
6585 for (; address < end; address += psize) {
6586 spinlock_t *ptl;
6587 ptep = hugetlb_walk(vma, address, psize);
6588 if (!ptep) {
6589 if (!uffd_wp) {
6590 address |= last_addr_mask;
6591 continue;
6592 }
6593 /*
6594 * Userfaultfd wr-protect requires pgtable
6595 * pre-allocations to install pte markers.
6596 */
6597 ptep = huge_pte_alloc(mm, vma, address, psize);
6598 if (!ptep) {
6599 pages = -ENOMEM;
6600 break;
6601 }
6602 }
6603 ptl = huge_pte_lock(h, mm, ptep);
6604 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6605 /*
6606 * When uffd-wp is enabled on the vma, unshare
6607 * shouldn't happen at all. Warn about it if it
6608 * happened due to some reason.
6609 */
6610 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6611 pages++;
6612 spin_unlock(ptl);
6613 shared_pmd = true;
6614 address |= last_addr_mask;
6615 continue;
6616 }
6617 pte = huge_ptep_get(ptep);
6618 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6619 /* Nothing to do. */
6620 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6621 swp_entry_t entry = pte_to_swp_entry(pte);
6622 struct page *page = pfn_swap_entry_to_page(entry);
6623 pte_t newpte = pte;
6624
6625 if (is_writable_migration_entry(entry)) {
6626 if (PageAnon(page))
6627 entry = make_readable_exclusive_migration_entry(
6628 swp_offset(entry));
6629 else
6630 entry = make_readable_migration_entry(
6631 swp_offset(entry));
6632 newpte = swp_entry_to_pte(entry);
6633 pages++;
6634 }
6635
6636 if (uffd_wp)
6637 newpte = pte_swp_mkuffd_wp(newpte);
6638 else if (uffd_wp_resolve)
6639 newpte = pte_swp_clear_uffd_wp(newpte);
6640 if (!pte_same(pte, newpte))
6641 set_huge_pte_at(mm, address, ptep, newpte, psize);
6642 } else if (unlikely(is_pte_marker(pte))) {
6643 /*
6644 * Do nothing on a poison marker; page is
6645 * corrupted, permissons do not apply. Here
6646 * pte_marker_uffd_wp()==true implies !poison
6647 * because they're mutual exclusive.
6648 */
6649 if (pte_marker_uffd_wp(pte) && uffd_wp_resolve)
6650 /* Safe to modify directly (non-present->none). */
6651 huge_pte_clear(mm, address, ptep, psize);
6652 } else if (!huge_pte_none(pte)) {
6653 pte_t old_pte;
6654 unsigned int shift = huge_page_shift(hstate_vma(vma));
6655
6656 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6657 pte = huge_pte_modify(old_pte, newprot);
6658 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6659 if (uffd_wp)
6660 pte = huge_pte_mkuffd_wp(pte);
6661 else if (uffd_wp_resolve)
6662 pte = huge_pte_clear_uffd_wp(pte);
6663 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6664 pages++;
6665 } else {
6666 /* None pte */
6667 if (unlikely(uffd_wp))
6668 /* Safe to modify directly (none->non-present). */
6669 set_huge_pte_at(mm, address, ptep,
6670 make_pte_marker(PTE_MARKER_UFFD_WP),
6671 psize);
6672 }
6673 spin_unlock(ptl);
6674 }
6675 /*
6676 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6677 * may have cleared our pud entry and done put_page on the page table:
6678 * once we release i_mmap_rwsem, another task can do the final put_page
6679 * and that page table be reused and filled with junk. If we actually
6680 * did unshare a page of pmds, flush the range corresponding to the pud.
6681 */
6682 if (shared_pmd)
6683 flush_hugetlb_tlb_range(vma, range.start, range.end);
6684 else
6685 flush_hugetlb_tlb_range(vma, start, end);
6686 /*
6687 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
6688 * downgrading page table protection not changing it to point to a new
6689 * page.
6690 *
6691 * See Documentation/mm/mmu_notifier.rst
6692 */
6693 i_mmap_unlock_write(vma->vm_file->f_mapping);
6694 hugetlb_vma_unlock_write(vma);
6695 mmu_notifier_invalidate_range_end(&range);
6696
6697 return pages > 0 ? (pages << h->order) : pages;
6698 }
6699
6700 /* Return true if reservation was successful, false otherwise. */
hugetlb_reserve_pages(struct inode * inode,long from,long to,struct vm_area_struct * vma,vm_flags_t vm_flags)6701 bool hugetlb_reserve_pages(struct inode *inode,
6702 long from, long to,
6703 struct vm_area_struct *vma,
6704 vm_flags_t vm_flags)
6705 {
6706 long chg = -1, add = -1;
6707 struct hstate *h = hstate_inode(inode);
6708 struct hugepage_subpool *spool = subpool_inode(inode);
6709 struct resv_map *resv_map;
6710 struct hugetlb_cgroup *h_cg = NULL;
6711 long gbl_reserve, regions_needed = 0;
6712
6713 /* This should never happen */
6714 if (from > to) {
6715 VM_WARN(1, "%s called with a negative range\n", __func__);
6716 return false;
6717 }
6718
6719 /*
6720 * vma specific semaphore used for pmd sharing and fault/truncation
6721 * synchronization
6722 */
6723 hugetlb_vma_lock_alloc(vma);
6724
6725 /*
6726 * Only apply hugepage reservation if asked. At fault time, an
6727 * attempt will be made for VM_NORESERVE to allocate a page
6728 * without using reserves
6729 */
6730 if (vm_flags & VM_NORESERVE)
6731 return true;
6732
6733 /*
6734 * Shared mappings base their reservation on the number of pages that
6735 * are already allocated on behalf of the file. Private mappings need
6736 * to reserve the full area even if read-only as mprotect() may be
6737 * called to make the mapping read-write. Assume !vma is a shm mapping
6738 */
6739 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6740 /*
6741 * resv_map can not be NULL as hugetlb_reserve_pages is only
6742 * called for inodes for which resv_maps were created (see
6743 * hugetlbfs_get_inode).
6744 */
6745 resv_map = inode_resv_map(inode);
6746
6747 chg = region_chg(resv_map, from, to, ®ions_needed);
6748 } else {
6749 /* Private mapping. */
6750 resv_map = resv_map_alloc();
6751 if (!resv_map)
6752 goto out_err;
6753
6754 chg = to - from;
6755
6756 set_vma_resv_map(vma, resv_map);
6757 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6758 }
6759
6760 if (chg < 0)
6761 goto out_err;
6762
6763 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6764 chg * pages_per_huge_page(h), &h_cg) < 0)
6765 goto out_err;
6766
6767 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6768 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6769 * of the resv_map.
6770 */
6771 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6772 }
6773
6774 /*
6775 * There must be enough pages in the subpool for the mapping. If
6776 * the subpool has a minimum size, there may be some global
6777 * reservations already in place (gbl_reserve).
6778 */
6779 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6780 if (gbl_reserve < 0)
6781 goto out_uncharge_cgroup;
6782
6783 /*
6784 * Check enough hugepages are available for the reservation.
6785 * Hand the pages back to the subpool if there are not
6786 */
6787 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6788 goto out_put_pages;
6789
6790 /*
6791 * Account for the reservations made. Shared mappings record regions
6792 * that have reservations as they are shared by multiple VMAs.
6793 * When the last VMA disappears, the region map says how much
6794 * the reservation was and the page cache tells how much of
6795 * the reservation was consumed. Private mappings are per-VMA and
6796 * only the consumed reservations are tracked. When the VMA
6797 * disappears, the original reservation is the VMA size and the
6798 * consumed reservations are stored in the map. Hence, nothing
6799 * else has to be done for private mappings here
6800 */
6801 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6802 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6803
6804 if (unlikely(add < 0)) {
6805 hugetlb_acct_memory(h, -gbl_reserve);
6806 goto out_put_pages;
6807 } else if (unlikely(chg > add)) {
6808 /*
6809 * pages in this range were added to the reserve
6810 * map between region_chg and region_add. This
6811 * indicates a race with alloc_hugetlb_folio. Adjust
6812 * the subpool and reserve counts modified above
6813 * based on the difference.
6814 */
6815 long rsv_adjust;
6816
6817 /*
6818 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6819 * reference to h_cg->css. See comment below for detail.
6820 */
6821 hugetlb_cgroup_uncharge_cgroup_rsvd(
6822 hstate_index(h),
6823 (chg - add) * pages_per_huge_page(h), h_cg);
6824
6825 rsv_adjust = hugepage_subpool_put_pages(spool,
6826 chg - add);
6827 hugetlb_acct_memory(h, -rsv_adjust);
6828 } else if (h_cg) {
6829 /*
6830 * The file_regions will hold their own reference to
6831 * h_cg->css. So we should release the reference held
6832 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6833 * done.
6834 */
6835 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6836 }
6837 }
6838 return true;
6839
6840 out_put_pages:
6841 /* put back original number of pages, chg */
6842 (void)hugepage_subpool_put_pages(spool, chg);
6843 out_uncharge_cgroup:
6844 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6845 chg * pages_per_huge_page(h), h_cg);
6846 out_err:
6847 hugetlb_vma_lock_free(vma);
6848 if (!vma || vma->vm_flags & VM_MAYSHARE)
6849 /* Only call region_abort if the region_chg succeeded but the
6850 * region_add failed or didn't run.
6851 */
6852 if (chg >= 0 && add < 0)
6853 region_abort(resv_map, from, to, regions_needed);
6854 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
6855 kref_put(&resv_map->refs, resv_map_release);
6856 set_vma_resv_map(vma, NULL);
6857 }
6858 return false;
6859 }
6860
hugetlb_unreserve_pages(struct inode * inode,long start,long end,long freed)6861 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6862 long freed)
6863 {
6864 struct hstate *h = hstate_inode(inode);
6865 struct resv_map *resv_map = inode_resv_map(inode);
6866 long chg = 0;
6867 struct hugepage_subpool *spool = subpool_inode(inode);
6868 long gbl_reserve;
6869
6870 /*
6871 * Since this routine can be called in the evict inode path for all
6872 * hugetlbfs inodes, resv_map could be NULL.
6873 */
6874 if (resv_map) {
6875 chg = region_del(resv_map, start, end);
6876 /*
6877 * region_del() can fail in the rare case where a region
6878 * must be split and another region descriptor can not be
6879 * allocated. If end == LONG_MAX, it will not fail.
6880 */
6881 if (chg < 0)
6882 return chg;
6883 }
6884
6885 spin_lock(&inode->i_lock);
6886 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6887 spin_unlock(&inode->i_lock);
6888
6889 /*
6890 * If the subpool has a minimum size, the number of global
6891 * reservations to be released may be adjusted.
6892 *
6893 * Note that !resv_map implies freed == 0. So (chg - freed)
6894 * won't go negative.
6895 */
6896 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6897 hugetlb_acct_memory(h, -gbl_reserve);
6898
6899 return 0;
6900 }
6901
6902 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
page_table_shareable(struct vm_area_struct * svma,struct vm_area_struct * vma,unsigned long addr,pgoff_t idx)6903 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6904 struct vm_area_struct *vma,
6905 unsigned long addr, pgoff_t idx)
6906 {
6907 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6908 svma->vm_start;
6909 unsigned long sbase = saddr & PUD_MASK;
6910 unsigned long s_end = sbase + PUD_SIZE;
6911
6912 /* Allow segments to share if only one is marked locked */
6913 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
6914 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
6915
6916 /*
6917 * match the virtual addresses, permission and the alignment of the
6918 * page table page.
6919 *
6920 * Also, vma_lock (vm_private_data) is required for sharing.
6921 */
6922 if (pmd_index(addr) != pmd_index(saddr) ||
6923 vm_flags != svm_flags ||
6924 !range_in_vma(svma, sbase, s_end) ||
6925 !svma->vm_private_data)
6926 return 0;
6927
6928 return saddr;
6929 }
6930
want_pmd_share(struct vm_area_struct * vma,unsigned long addr)6931 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6932 {
6933 unsigned long start = addr & PUD_MASK;
6934 unsigned long end = start + PUD_SIZE;
6935
6936 #ifdef CONFIG_USERFAULTFD
6937 if (uffd_disable_huge_pmd_share(vma))
6938 return false;
6939 #endif
6940 /*
6941 * check on proper vm_flags and page table alignment
6942 */
6943 if (!(vma->vm_flags & VM_MAYSHARE))
6944 return false;
6945 if (!vma->vm_private_data) /* vma lock required for sharing */
6946 return false;
6947 if (!range_in_vma(vma, start, end))
6948 return false;
6949 return true;
6950 }
6951
6952 /*
6953 * Determine if start,end range within vma could be mapped by shared pmd.
6954 * If yes, adjust start and end to cover range associated with possible
6955 * shared pmd mappings.
6956 */
adjust_range_if_pmd_sharing_possible(struct vm_area_struct * vma,unsigned long * start,unsigned long * end)6957 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6958 unsigned long *start, unsigned long *end)
6959 {
6960 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6961 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6962
6963 /*
6964 * vma needs to span at least one aligned PUD size, and the range
6965 * must be at least partially within in.
6966 */
6967 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6968 (*end <= v_start) || (*start >= v_end))
6969 return;
6970
6971 /* Extend the range to be PUD aligned for a worst case scenario */
6972 if (*start > v_start)
6973 *start = ALIGN_DOWN(*start, PUD_SIZE);
6974
6975 if (*end < v_end)
6976 *end = ALIGN(*end, PUD_SIZE);
6977 }
6978
6979 /*
6980 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6981 * and returns the corresponding pte. While this is not necessary for the
6982 * !shared pmd case because we can allocate the pmd later as well, it makes the
6983 * code much cleaner. pmd allocation is essential for the shared case because
6984 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
6985 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
6986 * bad pmd for sharing.
6987 */
huge_pmd_share(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pud_t * pud)6988 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6989 unsigned long addr, pud_t *pud)
6990 {
6991 struct address_space *mapping = vma->vm_file->f_mapping;
6992 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6993 vma->vm_pgoff;
6994 struct vm_area_struct *svma;
6995 unsigned long saddr;
6996 pte_t *spte = NULL;
6997 pte_t *pte;
6998
6999 i_mmap_lock_read(mapping);
7000 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
7001 if (svma == vma)
7002 continue;
7003
7004 saddr = page_table_shareable(svma, vma, addr, idx);
7005 if (saddr) {
7006 spte = hugetlb_walk(svma, saddr,
7007 vma_mmu_pagesize(svma));
7008 if (spte) {
7009 get_page(virt_to_page(spte));
7010 break;
7011 }
7012 }
7013 }
7014
7015 if (!spte)
7016 goto out;
7017
7018 spin_lock(&mm->page_table_lock);
7019 if (pud_none(*pud)) {
7020 pud_populate(mm, pud,
7021 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7022 mm_inc_nr_pmds(mm);
7023 } else {
7024 put_page(virt_to_page(spte));
7025 }
7026 spin_unlock(&mm->page_table_lock);
7027 out:
7028 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7029 i_mmap_unlock_read(mapping);
7030 return pte;
7031 }
7032
7033 /*
7034 * unmap huge page backed by shared pte.
7035 *
7036 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7037 * indicated by page_count > 1, unmap is achieved by clearing pud and
7038 * decrementing the ref count. If count == 1, the pte page is not shared.
7039 *
7040 * Called with page table lock held.
7041 *
7042 * returns: 1 successfully unmapped a shared pte page
7043 * 0 the underlying pte page is not shared, or it is the last user
7044 */
huge_pmd_unshare(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pte_t * ptep)7045 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7046 unsigned long addr, pte_t *ptep)
7047 {
7048 pgd_t *pgd = pgd_offset(mm, addr);
7049 p4d_t *p4d = p4d_offset(pgd, addr);
7050 pud_t *pud = pud_offset(p4d, addr);
7051
7052 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7053 hugetlb_vma_assert_locked(vma);
7054 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7055 if (page_count(virt_to_page(ptep)) == 1)
7056 return 0;
7057
7058 pud_clear(pud);
7059 put_page(virt_to_page(ptep));
7060 mm_dec_nr_pmds(mm);
7061 return 1;
7062 }
7063
7064 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7065
huge_pmd_share(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pud_t * pud)7066 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7067 unsigned long addr, pud_t *pud)
7068 {
7069 return NULL;
7070 }
7071
huge_pmd_unshare(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pte_t * ptep)7072 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7073 unsigned long addr, pte_t *ptep)
7074 {
7075 return 0;
7076 }
7077
adjust_range_if_pmd_sharing_possible(struct vm_area_struct * vma,unsigned long * start,unsigned long * end)7078 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7079 unsigned long *start, unsigned long *end)
7080 {
7081 }
7082
want_pmd_share(struct vm_area_struct * vma,unsigned long addr)7083 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7084 {
7085 return false;
7086 }
7087 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7088
7089 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
huge_pte_alloc(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,unsigned long sz)7090 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7091 unsigned long addr, unsigned long sz)
7092 {
7093 pgd_t *pgd;
7094 p4d_t *p4d;
7095 pud_t *pud;
7096 pte_t *pte = NULL;
7097
7098 pgd = pgd_offset(mm, addr);
7099 p4d = p4d_alloc(mm, pgd, addr);
7100 if (!p4d)
7101 return NULL;
7102 pud = pud_alloc(mm, p4d, addr);
7103 if (pud) {
7104 if (sz == PUD_SIZE) {
7105 pte = (pte_t *)pud;
7106 } else {
7107 BUG_ON(sz != PMD_SIZE);
7108 if (want_pmd_share(vma, addr) && pud_none(*pud))
7109 pte = huge_pmd_share(mm, vma, addr, pud);
7110 else
7111 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7112 }
7113 }
7114
7115 if (pte) {
7116 pte_t pteval = ptep_get_lockless(pte);
7117
7118 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7119 }
7120
7121 return pte;
7122 }
7123
7124 /*
7125 * huge_pte_offset() - Walk the page table to resolve the hugepage
7126 * entry at address @addr
7127 *
7128 * Return: Pointer to page table entry (PUD or PMD) for
7129 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7130 * size @sz doesn't match the hugepage size at this level of the page
7131 * table.
7132 */
huge_pte_offset(struct mm_struct * mm,unsigned long addr,unsigned long sz)7133 pte_t *huge_pte_offset(struct mm_struct *mm,
7134 unsigned long addr, unsigned long sz)
7135 {
7136 pgd_t *pgd;
7137 p4d_t *p4d;
7138 pud_t *pud;
7139 pmd_t *pmd;
7140
7141 pgd = pgd_offset(mm, addr);
7142 if (!pgd_present(*pgd))
7143 return NULL;
7144 p4d = p4d_offset(pgd, addr);
7145 if (!p4d_present(*p4d))
7146 return NULL;
7147
7148 pud = pud_offset(p4d, addr);
7149 if (sz == PUD_SIZE)
7150 /* must be pud huge, non-present or none */
7151 return (pte_t *)pud;
7152 if (!pud_present(*pud))
7153 return NULL;
7154 /* must have a valid entry and size to go further */
7155
7156 pmd = pmd_offset(pud, addr);
7157 /* must be pmd huge, non-present or none */
7158 return (pte_t *)pmd;
7159 }
7160
7161 /*
7162 * Return a mask that can be used to update an address to the last huge
7163 * page in a page table page mapping size. Used to skip non-present
7164 * page table entries when linearly scanning address ranges. Architectures
7165 * with unique huge page to page table relationships can define their own
7166 * version of this routine.
7167 */
hugetlb_mask_last_page(struct hstate * h)7168 unsigned long hugetlb_mask_last_page(struct hstate *h)
7169 {
7170 unsigned long hp_size = huge_page_size(h);
7171
7172 if (hp_size == PUD_SIZE)
7173 return P4D_SIZE - PUD_SIZE;
7174 else if (hp_size == PMD_SIZE)
7175 return PUD_SIZE - PMD_SIZE;
7176 else
7177 return 0UL;
7178 }
7179
7180 #else
7181
7182 /* See description above. Architectures can provide their own version. */
hugetlb_mask_last_page(struct hstate * h)7183 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7184 {
7185 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7186 if (huge_page_size(h) == PMD_SIZE)
7187 return PUD_SIZE - PMD_SIZE;
7188 #endif
7189 return 0UL;
7190 }
7191
7192 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7193
7194 /*
7195 * These functions are overwritable if your architecture needs its own
7196 * behavior.
7197 */
isolate_hugetlb(struct folio * folio,struct list_head * list)7198 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7199 {
7200 bool ret = true;
7201
7202 spin_lock_irq(&hugetlb_lock);
7203 if (!folio_test_hugetlb(folio) ||
7204 !folio_test_hugetlb_migratable(folio) ||
7205 !folio_try_get(folio)) {
7206 ret = false;
7207 goto unlock;
7208 }
7209 folio_clear_hugetlb_migratable(folio);
7210 list_move_tail(&folio->lru, list);
7211 unlock:
7212 spin_unlock_irq(&hugetlb_lock);
7213 return ret;
7214 }
7215
get_hwpoison_hugetlb_folio(struct folio * folio,bool * hugetlb,bool unpoison)7216 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7217 {
7218 int ret = 0;
7219
7220 *hugetlb = false;
7221 spin_lock_irq(&hugetlb_lock);
7222 if (folio_test_hugetlb(folio)) {
7223 *hugetlb = true;
7224 if (folio_test_hugetlb_freed(folio))
7225 ret = 0;
7226 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7227 ret = folio_try_get(folio);
7228 else
7229 ret = -EBUSY;
7230 }
7231 spin_unlock_irq(&hugetlb_lock);
7232 return ret;
7233 }
7234
get_huge_page_for_hwpoison(unsigned long pfn,int flags,bool * migratable_cleared)7235 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7236 bool *migratable_cleared)
7237 {
7238 int ret;
7239
7240 spin_lock_irq(&hugetlb_lock);
7241 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7242 spin_unlock_irq(&hugetlb_lock);
7243 return ret;
7244 }
7245
folio_putback_active_hugetlb(struct folio * folio)7246 void folio_putback_active_hugetlb(struct folio *folio)
7247 {
7248 spin_lock_irq(&hugetlb_lock);
7249 folio_set_hugetlb_migratable(folio);
7250 list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7251 spin_unlock_irq(&hugetlb_lock);
7252 folio_put(folio);
7253 }
7254
move_hugetlb_state(struct folio * old_folio,struct folio * new_folio,int reason)7255 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7256 {
7257 struct hstate *h = folio_hstate(old_folio);
7258
7259 hugetlb_cgroup_migrate(old_folio, new_folio);
7260 set_page_owner_migrate_reason(&new_folio->page, reason);
7261
7262 /*
7263 * transfer temporary state of the new hugetlb folio. This is
7264 * reverse to other transitions because the newpage is going to
7265 * be final while the old one will be freed so it takes over
7266 * the temporary status.
7267 *
7268 * Also note that we have to transfer the per-node surplus state
7269 * here as well otherwise the global surplus count will not match
7270 * the per-node's.
7271 */
7272 if (folio_test_hugetlb_temporary(new_folio)) {
7273 int old_nid = folio_nid(old_folio);
7274 int new_nid = folio_nid(new_folio);
7275
7276 folio_set_hugetlb_temporary(old_folio);
7277 folio_clear_hugetlb_temporary(new_folio);
7278
7279
7280 /*
7281 * There is no need to transfer the per-node surplus state
7282 * when we do not cross the node.
7283 */
7284 if (new_nid == old_nid)
7285 return;
7286 spin_lock_irq(&hugetlb_lock);
7287 if (h->surplus_huge_pages_node[old_nid]) {
7288 h->surplus_huge_pages_node[old_nid]--;
7289 h->surplus_huge_pages_node[new_nid]++;
7290 }
7291 spin_unlock_irq(&hugetlb_lock);
7292 }
7293 }
7294
hugetlb_unshare_pmds(struct vm_area_struct * vma,unsigned long start,unsigned long end)7295 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7296 unsigned long start,
7297 unsigned long end)
7298 {
7299 struct hstate *h = hstate_vma(vma);
7300 unsigned long sz = huge_page_size(h);
7301 struct mm_struct *mm = vma->vm_mm;
7302 struct mmu_notifier_range range;
7303 unsigned long address;
7304 spinlock_t *ptl;
7305 pte_t *ptep;
7306
7307 if (!(vma->vm_flags & VM_MAYSHARE))
7308 return;
7309
7310 if (start >= end)
7311 return;
7312
7313 flush_cache_range(vma, start, end);
7314 /*
7315 * No need to call adjust_range_if_pmd_sharing_possible(), because
7316 * we have already done the PUD_SIZE alignment.
7317 */
7318 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7319 start, end);
7320 mmu_notifier_invalidate_range_start(&range);
7321 hugetlb_vma_lock_write(vma);
7322 i_mmap_lock_write(vma->vm_file->f_mapping);
7323 for (address = start; address < end; address += PUD_SIZE) {
7324 ptep = hugetlb_walk(vma, address, sz);
7325 if (!ptep)
7326 continue;
7327 ptl = huge_pte_lock(h, mm, ptep);
7328 huge_pmd_unshare(mm, vma, address, ptep);
7329 spin_unlock(ptl);
7330 }
7331 flush_hugetlb_tlb_range(vma, start, end);
7332 i_mmap_unlock_write(vma->vm_file->f_mapping);
7333 hugetlb_vma_unlock_write(vma);
7334 /*
7335 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7336 * Documentation/mm/mmu_notifier.rst.
7337 */
7338 mmu_notifier_invalidate_range_end(&range);
7339 }
7340
7341 /*
7342 * This function will unconditionally remove all the shared pmd pgtable entries
7343 * within the specific vma for a hugetlbfs memory range.
7344 */
hugetlb_unshare_all_pmds(struct vm_area_struct * vma)7345 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7346 {
7347 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7348 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7349 }
7350
7351 #ifdef CONFIG_CMA
7352 static bool cma_reserve_called __initdata;
7353
cmdline_parse_hugetlb_cma(char * p)7354 static int __init cmdline_parse_hugetlb_cma(char *p)
7355 {
7356 int nid, count = 0;
7357 unsigned long tmp;
7358 char *s = p;
7359
7360 while (*s) {
7361 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7362 break;
7363
7364 if (s[count] == ':') {
7365 if (tmp >= MAX_NUMNODES)
7366 break;
7367 nid = array_index_nospec(tmp, MAX_NUMNODES);
7368
7369 s += count + 1;
7370 tmp = memparse(s, &s);
7371 hugetlb_cma_size_in_node[nid] = tmp;
7372 hugetlb_cma_size += tmp;
7373
7374 /*
7375 * Skip the separator if have one, otherwise
7376 * break the parsing.
7377 */
7378 if (*s == ',')
7379 s++;
7380 else
7381 break;
7382 } else {
7383 hugetlb_cma_size = memparse(p, &p);
7384 break;
7385 }
7386 }
7387
7388 return 0;
7389 }
7390
7391 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7392
hugetlb_cma_reserve(int order)7393 void __init hugetlb_cma_reserve(int order)
7394 {
7395 unsigned long size, reserved, per_node;
7396 bool node_specific_cma_alloc = false;
7397 int nid;
7398
7399 cma_reserve_called = true;
7400
7401 if (!hugetlb_cma_size)
7402 return;
7403
7404 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7405 if (hugetlb_cma_size_in_node[nid] == 0)
7406 continue;
7407
7408 if (!node_online(nid)) {
7409 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7410 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7411 hugetlb_cma_size_in_node[nid] = 0;
7412 continue;
7413 }
7414
7415 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7416 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7417 nid, (PAGE_SIZE << order) / SZ_1M);
7418 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7419 hugetlb_cma_size_in_node[nid] = 0;
7420 } else {
7421 node_specific_cma_alloc = true;
7422 }
7423 }
7424
7425 /* Validate the CMA size again in case some invalid nodes specified. */
7426 if (!hugetlb_cma_size)
7427 return;
7428
7429 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7430 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7431 (PAGE_SIZE << order) / SZ_1M);
7432 hugetlb_cma_size = 0;
7433 return;
7434 }
7435
7436 if (!node_specific_cma_alloc) {
7437 /*
7438 * If 3 GB area is requested on a machine with 4 numa nodes,
7439 * let's allocate 1 GB on first three nodes and ignore the last one.
7440 */
7441 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7442 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7443 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7444 }
7445
7446 reserved = 0;
7447 for_each_online_node(nid) {
7448 int res;
7449 char name[CMA_MAX_NAME];
7450
7451 if (node_specific_cma_alloc) {
7452 if (hugetlb_cma_size_in_node[nid] == 0)
7453 continue;
7454
7455 size = hugetlb_cma_size_in_node[nid];
7456 } else {
7457 size = min(per_node, hugetlb_cma_size - reserved);
7458 }
7459
7460 size = round_up(size, PAGE_SIZE << order);
7461
7462 snprintf(name, sizeof(name), "hugetlb%d", nid);
7463 /*
7464 * Note that 'order per bit' is based on smallest size that
7465 * may be returned to CMA allocator in the case of
7466 * huge page demotion.
7467 */
7468 res = cma_declare_contiguous_nid(0, size, 0,
7469 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7470 0, false, name,
7471 &hugetlb_cma[nid], nid);
7472 if (res) {
7473 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7474 res, nid);
7475 continue;
7476 }
7477
7478 reserved += size;
7479 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7480 size / SZ_1M, nid);
7481
7482 if (reserved >= hugetlb_cma_size)
7483 break;
7484 }
7485
7486 if (!reserved)
7487 /*
7488 * hugetlb_cma_size is used to determine if allocations from
7489 * cma are possible. Set to zero if no cma regions are set up.
7490 */
7491 hugetlb_cma_size = 0;
7492 }
7493
hugetlb_cma_check(void)7494 static void __init hugetlb_cma_check(void)
7495 {
7496 if (!hugetlb_cma_size || cma_reserve_called)
7497 return;
7498
7499 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7500 }
7501
7502 #endif /* CONFIG_CMA */
7503