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 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1751 return;
1752
1753 /*
1754 * If we don't know which subpages are hwpoisoned, we can't free
1755 * the hugepage, so it's leaked intentionally.
1756 */
1757 if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1758 return;
1759
1760 if (hugetlb_vmemmap_restore(h, &folio->page)) {
1761 spin_lock_irq(&hugetlb_lock);
1762 /*
1763 * If we cannot allocate vmemmap pages, just refuse to free the
1764 * page and put the page back on the hugetlb free list and treat
1765 * as a surplus page.
1766 */
1767 add_hugetlb_folio(h, folio, true);
1768 spin_unlock_irq(&hugetlb_lock);
1769 return;
1770 }
1771
1772 /*
1773 * Move PageHWPoison flag from head page to the raw error pages,
1774 * which makes any healthy subpages reusable.
1775 */
1776 if (unlikely(folio_test_hwpoison(folio)))
1777 folio_clear_hugetlb_hwpoison(folio);
1778
1779 /*
1780 * If vmemmap pages were allocated above, then we need to clear the
1781 * hugetlb destructor under the hugetlb lock.
1782 */
1783 if (folio_test_hugetlb(folio)) {
1784 spin_lock_irq(&hugetlb_lock);
1785 __clear_hugetlb_destructor(h, folio);
1786 spin_unlock_irq(&hugetlb_lock);
1787 }
1788
1789 /*
1790 * Non-gigantic pages demoted from CMA allocated gigantic pages
1791 * need to be given back to CMA in free_gigantic_folio.
1792 */
1793 if (hstate_is_gigantic(h) ||
1794 hugetlb_cma_folio(folio, huge_page_order(h))) {
1795 destroy_compound_gigantic_folio(folio, huge_page_order(h));
1796 free_gigantic_folio(folio, huge_page_order(h));
1797 } else {
1798 __free_pages(&folio->page, huge_page_order(h));
1799 }
1800 }
1801
1802 /*
1803 * As update_and_free_hugetlb_folio() can be called under any context, so we cannot
1804 * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the
1805 * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate
1806 * the vmemmap pages.
1807 *
1808 * free_hpage_workfn() locklessly retrieves the linked list of pages to be
1809 * freed and frees them one-by-one. As the page->mapping pointer is going
1810 * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node
1811 * structure of a lockless linked list of huge pages to be freed.
1812 */
1813 static LLIST_HEAD(hpage_freelist);
1814
free_hpage_workfn(struct work_struct * work)1815 static void free_hpage_workfn(struct work_struct *work)
1816 {
1817 struct llist_node *node;
1818
1819 node = llist_del_all(&hpage_freelist);
1820
1821 while (node) {
1822 struct page *page;
1823 struct hstate *h;
1824
1825 page = container_of((struct address_space **)node,
1826 struct page, mapping);
1827 node = node->next;
1828 page->mapping = NULL;
1829 /*
1830 * The VM_BUG_ON_FOLIO(!folio_test_hugetlb(folio), folio) in
1831 * folio_hstate() is going to trigger because a previous call to
1832 * remove_hugetlb_folio() will clear the hugetlb bit, so do
1833 * not use folio_hstate() directly.
1834 */
1835 h = size_to_hstate(page_size(page));
1836
1837 __update_and_free_hugetlb_folio(h, page_folio(page));
1838
1839 cond_resched();
1840 }
1841 }
1842 static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1843
flush_free_hpage_work(struct hstate * h)1844 static inline void flush_free_hpage_work(struct hstate *h)
1845 {
1846 if (hugetlb_vmemmap_optimizable(h))
1847 flush_work(&free_hpage_work);
1848 }
1849
update_and_free_hugetlb_folio(struct hstate * h,struct folio * folio,bool atomic)1850 static void update_and_free_hugetlb_folio(struct hstate *h, struct folio *folio,
1851 bool atomic)
1852 {
1853 if (!folio_test_hugetlb_vmemmap_optimized(folio) || !atomic) {
1854 __update_and_free_hugetlb_folio(h, folio);
1855 return;
1856 }
1857
1858 /*
1859 * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages.
1860 *
1861 * Only call schedule_work() if hpage_freelist is previously
1862 * empty. Otherwise, schedule_work() had been called but the workfn
1863 * hasn't retrieved the list yet.
1864 */
1865 if (llist_add((struct llist_node *)&folio->mapping, &hpage_freelist))
1866 schedule_work(&free_hpage_work);
1867 }
1868
update_and_free_pages_bulk(struct hstate * h,struct list_head * list)1869 static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list)
1870 {
1871 struct page *page, *t_page;
1872 struct folio *folio;
1873
1874 list_for_each_entry_safe(page, t_page, list, lru) {
1875 folio = page_folio(page);
1876 update_and_free_hugetlb_folio(h, folio, false);
1877 cond_resched();
1878 }
1879 }
1880
size_to_hstate(unsigned long size)1881 struct hstate *size_to_hstate(unsigned long size)
1882 {
1883 struct hstate *h;
1884
1885 for_each_hstate(h) {
1886 if (huge_page_size(h) == size)
1887 return h;
1888 }
1889 return NULL;
1890 }
1891
free_huge_folio(struct folio * folio)1892 void free_huge_folio(struct folio *folio)
1893 {
1894 /*
1895 * Can't pass hstate in here because it is called from the
1896 * compound page destructor.
1897 */
1898 struct hstate *h = folio_hstate(folio);
1899 int nid = folio_nid(folio);
1900 struct hugepage_subpool *spool = hugetlb_folio_subpool(folio);
1901 bool restore_reserve;
1902 unsigned long flags;
1903
1904 VM_BUG_ON_FOLIO(folio_ref_count(folio), folio);
1905 VM_BUG_ON_FOLIO(folio_mapcount(folio), folio);
1906
1907 hugetlb_set_folio_subpool(folio, NULL);
1908 if (folio_test_anon(folio))
1909 __ClearPageAnonExclusive(&folio->page);
1910 folio->mapping = NULL;
1911 restore_reserve = folio_test_hugetlb_restore_reserve(folio);
1912 folio_clear_hugetlb_restore_reserve(folio);
1913
1914 /*
1915 * If HPageRestoreReserve was set on page, page allocation consumed a
1916 * reservation. If the page was associated with a subpool, there
1917 * would have been a page reserved in the subpool before allocation
1918 * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1919 * reservation, do not call hugepage_subpool_put_pages() as this will
1920 * remove the reserved page from the subpool.
1921 */
1922 if (!restore_reserve) {
1923 /*
1924 * A return code of zero implies that the subpool will be
1925 * under its minimum size if the reservation is not restored
1926 * after page is free. Therefore, force restore_reserve
1927 * operation.
1928 */
1929 if (hugepage_subpool_put_pages(spool, 1) == 0)
1930 restore_reserve = true;
1931 }
1932
1933 spin_lock_irqsave(&hugetlb_lock, flags);
1934 folio_clear_hugetlb_migratable(folio);
1935 hugetlb_cgroup_uncharge_folio(hstate_index(h),
1936 pages_per_huge_page(h), folio);
1937 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
1938 pages_per_huge_page(h), folio);
1939 if (restore_reserve)
1940 h->resv_huge_pages++;
1941
1942 if (folio_test_hugetlb_temporary(folio)) {
1943 remove_hugetlb_folio(h, folio, false);
1944 spin_unlock_irqrestore(&hugetlb_lock, flags);
1945 update_and_free_hugetlb_folio(h, folio, true);
1946 } else if (h->surplus_huge_pages_node[nid]) {
1947 /* remove the page from active list */
1948 remove_hugetlb_folio(h, folio, true);
1949 spin_unlock_irqrestore(&hugetlb_lock, flags);
1950 update_and_free_hugetlb_folio(h, folio, true);
1951 } else {
1952 arch_clear_hugepage_flags(&folio->page);
1953 enqueue_hugetlb_folio(h, folio);
1954 spin_unlock_irqrestore(&hugetlb_lock, flags);
1955 }
1956 }
1957
1958 /*
1959 * Must be called with the hugetlb lock held
1960 */
__prep_account_new_huge_page(struct hstate * h,int nid)1961 static void __prep_account_new_huge_page(struct hstate *h, int nid)
1962 {
1963 lockdep_assert_held(&hugetlb_lock);
1964 h->nr_huge_pages++;
1965 h->nr_huge_pages_node[nid]++;
1966 }
1967
__prep_new_hugetlb_folio(struct hstate * h,struct folio * folio)1968 static void __prep_new_hugetlb_folio(struct hstate *h, struct folio *folio)
1969 {
1970 hugetlb_vmemmap_optimize(h, &folio->page);
1971 INIT_LIST_HEAD(&folio->lru);
1972 __folio_set_hugetlb(folio);
1973 hugetlb_set_folio_subpool(folio, NULL);
1974 set_hugetlb_cgroup(folio, NULL);
1975 set_hugetlb_cgroup_rsvd(folio, NULL);
1976 }
1977
prep_new_hugetlb_folio(struct hstate * h,struct folio * folio,int nid)1978 static void prep_new_hugetlb_folio(struct hstate *h, struct folio *folio, int nid)
1979 {
1980 __prep_new_hugetlb_folio(h, folio);
1981 spin_lock_irq(&hugetlb_lock);
1982 __prep_account_new_huge_page(h, nid);
1983 spin_unlock_irq(&hugetlb_lock);
1984 }
1985
__prep_compound_gigantic_folio(struct folio * folio,unsigned int order,bool demote)1986 static bool __prep_compound_gigantic_folio(struct folio *folio,
1987 unsigned int order, bool demote)
1988 {
1989 int i, j;
1990 int nr_pages = 1 << order;
1991 struct page *p;
1992
1993 __folio_clear_reserved(folio);
1994 for (i = 0; i < nr_pages; i++) {
1995 p = folio_page(folio, i);
1996
1997 /*
1998 * For gigantic hugepages allocated through bootmem at
1999 * boot, it's safer to be consistent with the not-gigantic
2000 * hugepages and clear the PG_reserved bit from all tail pages
2001 * too. Otherwise drivers using get_user_pages() to access tail
2002 * pages may get the reference counting wrong if they see
2003 * PG_reserved set on a tail page (despite the head page not
2004 * having PG_reserved set). Enforcing this consistency between
2005 * head and tail pages allows drivers to optimize away a check
2006 * on the head page when they need know if put_page() is needed
2007 * after get_user_pages().
2008 */
2009 if (i != 0) /* head page cleared above */
2010 __ClearPageReserved(p);
2011 /*
2012 * Subtle and very unlikely
2013 *
2014 * Gigantic 'page allocators' such as memblock or cma will
2015 * return a set of pages with each page ref counted. We need
2016 * to turn this set of pages into a compound page with tail
2017 * page ref counts set to zero. Code such as speculative page
2018 * cache adding could take a ref on a 'to be' tail page.
2019 * We need to respect any increased ref count, and only set
2020 * the ref count to zero if count is currently 1. If count
2021 * is not 1, we return an error. An error return indicates
2022 * the set of pages can not be converted to a gigantic page.
2023 * The caller who allocated the pages should then discard the
2024 * pages using the appropriate free interface.
2025 *
2026 * In the case of demote, the ref count will be zero.
2027 */
2028 if (!demote) {
2029 if (!page_ref_freeze(p, 1)) {
2030 pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n");
2031 goto out_error;
2032 }
2033 } else {
2034 VM_BUG_ON_PAGE(page_count(p), p);
2035 }
2036 if (i != 0)
2037 set_compound_head(p, &folio->page);
2038 }
2039 __folio_set_head(folio);
2040 /* we rely on prep_new_hugetlb_folio to set the destructor */
2041 folio_set_order(folio, order);
2042 atomic_set(&folio->_entire_mapcount, -1);
2043 atomic_set(&folio->_nr_pages_mapped, 0);
2044 atomic_set(&folio->_pincount, 0);
2045 return true;
2046
2047 out_error:
2048 /* undo page modifications made above */
2049 for (j = 0; j < i; j++) {
2050 p = folio_page(folio, j);
2051 if (j != 0)
2052 clear_compound_head(p);
2053 set_page_refcounted(p);
2054 }
2055 /* need to clear PG_reserved on remaining tail pages */
2056 for (; j < nr_pages; j++) {
2057 p = folio_page(folio, j);
2058 __ClearPageReserved(p);
2059 }
2060 return false;
2061 }
2062
prep_compound_gigantic_folio(struct folio * folio,unsigned int order)2063 static bool prep_compound_gigantic_folio(struct folio *folio,
2064 unsigned int order)
2065 {
2066 return __prep_compound_gigantic_folio(folio, order, false);
2067 }
2068
prep_compound_gigantic_folio_for_demote(struct folio * folio,unsigned int order)2069 static bool prep_compound_gigantic_folio_for_demote(struct folio *folio,
2070 unsigned int order)
2071 {
2072 return __prep_compound_gigantic_folio(folio, order, true);
2073 }
2074
2075 /*
2076 * Find and lock address space (mapping) in write mode.
2077 *
2078 * Upon entry, the page is locked which means that page_mapping() is
2079 * stable. Due to locking order, we can only trylock_write. If we can
2080 * not get the lock, simply return NULL to caller.
2081 */
hugetlb_page_mapping_lock_write(struct page * hpage)2082 struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage)
2083 {
2084 struct address_space *mapping = page_mapping(hpage);
2085
2086 if (!mapping)
2087 return mapping;
2088
2089 if (i_mmap_trylock_write(mapping))
2090 return mapping;
2091
2092 return NULL;
2093 }
2094
hugetlb_basepage_index(struct page * page)2095 pgoff_t hugetlb_basepage_index(struct page *page)
2096 {
2097 struct page *page_head = compound_head(page);
2098 pgoff_t index = page_index(page_head);
2099 unsigned long compound_idx;
2100
2101 if (compound_order(page_head) > MAX_ORDER)
2102 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
2103 else
2104 compound_idx = page - page_head;
2105
2106 return (index << compound_order(page_head)) + compound_idx;
2107 }
2108
alloc_buddy_hugetlb_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask,nodemask_t * node_alloc_noretry)2109 static struct folio *alloc_buddy_hugetlb_folio(struct hstate *h,
2110 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2111 nodemask_t *node_alloc_noretry)
2112 {
2113 int order = huge_page_order(h);
2114 struct page *page;
2115 bool alloc_try_hard = true;
2116 bool retry = true;
2117
2118 /*
2119 * By default we always try hard to allocate the page with
2120 * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
2121 * a loop (to adjust global huge page counts) and previous allocation
2122 * failed, do not continue to try hard on the same node. Use the
2123 * node_alloc_noretry bitmap to manage this state information.
2124 */
2125 if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
2126 alloc_try_hard = false;
2127 gfp_mask |= __GFP_COMP|__GFP_NOWARN;
2128 if (alloc_try_hard)
2129 gfp_mask |= __GFP_RETRY_MAYFAIL;
2130 if (nid == NUMA_NO_NODE)
2131 nid = numa_mem_id();
2132 retry:
2133 page = __alloc_pages(gfp_mask, order, nid, nmask);
2134
2135 /* Freeze head page */
2136 if (page && !page_ref_freeze(page, 1)) {
2137 __free_pages(page, order);
2138 if (retry) { /* retry once */
2139 retry = false;
2140 goto retry;
2141 }
2142 /* WOW! twice in a row. */
2143 pr_warn("HugeTLB head page unexpected inflated ref count\n");
2144 page = NULL;
2145 }
2146
2147 /*
2148 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
2149 * indicates an overall state change. Clear bit so that we resume
2150 * normal 'try hard' allocations.
2151 */
2152 if (node_alloc_noretry && page && !alloc_try_hard)
2153 node_clear(nid, *node_alloc_noretry);
2154
2155 /*
2156 * If we tried hard to get a page but failed, set bit so that
2157 * subsequent attempts will not try as hard until there is an
2158 * overall state change.
2159 */
2160 if (node_alloc_noretry && !page && alloc_try_hard)
2161 node_set(nid, *node_alloc_noretry);
2162
2163 if (!page) {
2164 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
2165 return NULL;
2166 }
2167
2168 __count_vm_event(HTLB_BUDDY_PGALLOC);
2169 return page_folio(page);
2170 }
2171
2172 /*
2173 * Common helper to allocate a fresh hugetlb page. All specific allocators
2174 * should use this function to get new hugetlb pages
2175 *
2176 * Note that returned page is 'frozen': ref count of head page and all tail
2177 * pages is zero.
2178 */
alloc_fresh_hugetlb_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask,nodemask_t * node_alloc_noretry)2179 static struct folio *alloc_fresh_hugetlb_folio(struct hstate *h,
2180 gfp_t gfp_mask, int nid, nodemask_t *nmask,
2181 nodemask_t *node_alloc_noretry)
2182 {
2183 struct folio *folio;
2184 bool retry = false;
2185
2186 retry:
2187 if (hstate_is_gigantic(h))
2188 folio = alloc_gigantic_folio(h, gfp_mask, nid, nmask);
2189 else
2190 folio = alloc_buddy_hugetlb_folio(h, gfp_mask,
2191 nid, nmask, node_alloc_noretry);
2192 if (!folio)
2193 return NULL;
2194 if (hstate_is_gigantic(h)) {
2195 if (!prep_compound_gigantic_folio(folio, huge_page_order(h))) {
2196 /*
2197 * Rare failure to convert pages to compound page.
2198 * Free pages and try again - ONCE!
2199 */
2200 free_gigantic_folio(folio, huge_page_order(h));
2201 if (!retry) {
2202 retry = true;
2203 goto retry;
2204 }
2205 return NULL;
2206 }
2207 }
2208 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
2209
2210 return folio;
2211 }
2212
2213 /*
2214 * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
2215 * manner.
2216 */
alloc_pool_huge_page(struct hstate * h,nodemask_t * nodes_allowed,nodemask_t * node_alloc_noretry)2217 static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
2218 nodemask_t *node_alloc_noretry)
2219 {
2220 struct folio *folio;
2221 int nr_nodes, node;
2222 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2223
2224 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2225 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, node,
2226 nodes_allowed, node_alloc_noretry);
2227 if (folio) {
2228 free_huge_folio(folio); /* free it into the hugepage allocator */
2229 return 1;
2230 }
2231 }
2232
2233 return 0;
2234 }
2235
2236 /*
2237 * Remove huge page from pool from next node to free. Attempt to keep
2238 * persistent huge pages more or less balanced over allowed nodes.
2239 * This routine only 'removes' the hugetlb page. The caller must make
2240 * an additional call to free the page to low level allocators.
2241 * Called with hugetlb_lock locked.
2242 */
remove_pool_huge_page(struct hstate * h,nodemask_t * nodes_allowed,bool acct_surplus)2243 static struct page *remove_pool_huge_page(struct hstate *h,
2244 nodemask_t *nodes_allowed,
2245 bool acct_surplus)
2246 {
2247 int nr_nodes, node;
2248 struct page *page = NULL;
2249 struct folio *folio;
2250
2251 lockdep_assert_held(&hugetlb_lock);
2252 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2253 /*
2254 * If we're returning unused surplus pages, only examine
2255 * nodes with surplus pages.
2256 */
2257 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
2258 !list_empty(&h->hugepage_freelists[node])) {
2259 page = list_entry(h->hugepage_freelists[node].next,
2260 struct page, lru);
2261 folio = page_folio(page);
2262 remove_hugetlb_folio(h, folio, acct_surplus);
2263 break;
2264 }
2265 }
2266
2267 return page;
2268 }
2269
2270 /*
2271 * Dissolve a given free hugepage into free buddy pages. This function does
2272 * nothing for in-use hugepages and non-hugepages.
2273 * This function returns values like below:
2274 *
2275 * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages
2276 * when the system is under memory pressure and the feature of
2277 * freeing unused vmemmap pages associated with each hugetlb page
2278 * is enabled.
2279 * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
2280 * (allocated or reserved.)
2281 * 0: successfully dissolved free hugepages or the page is not a
2282 * hugepage (considered as already dissolved)
2283 */
dissolve_free_huge_page(struct page * page)2284 int dissolve_free_huge_page(struct page *page)
2285 {
2286 int rc = -EBUSY;
2287 struct folio *folio = page_folio(page);
2288
2289 retry:
2290 /* Not to disrupt normal path by vainly holding hugetlb_lock */
2291 if (!folio_test_hugetlb(folio))
2292 return 0;
2293
2294 spin_lock_irq(&hugetlb_lock);
2295 if (!folio_test_hugetlb(folio)) {
2296 rc = 0;
2297 goto out;
2298 }
2299
2300 if (!folio_ref_count(folio)) {
2301 struct hstate *h = folio_hstate(folio);
2302 if (!available_huge_pages(h))
2303 goto out;
2304
2305 /*
2306 * We should make sure that the page is already on the free list
2307 * when it is dissolved.
2308 */
2309 if (unlikely(!folio_test_hugetlb_freed(folio))) {
2310 spin_unlock_irq(&hugetlb_lock);
2311 cond_resched();
2312
2313 /*
2314 * Theoretically, we should return -EBUSY when we
2315 * encounter this race. In fact, we have a chance
2316 * to successfully dissolve the page if we do a
2317 * retry. Because the race window is quite small.
2318 * If we seize this opportunity, it is an optimization
2319 * for increasing the success rate of dissolving page.
2320 */
2321 goto retry;
2322 }
2323
2324 remove_hugetlb_folio(h, folio, false);
2325 h->max_huge_pages--;
2326 spin_unlock_irq(&hugetlb_lock);
2327
2328 /*
2329 * Normally update_and_free_hugtlb_folio will allocate required vmemmmap
2330 * before freeing the page. update_and_free_hugtlb_folio will fail to
2331 * free the page if it can not allocate required vmemmap. We
2332 * need to adjust max_huge_pages if the page is not freed.
2333 * Attempt to allocate vmemmmap here so that we can take
2334 * appropriate action on failure.
2335 */
2336 rc = hugetlb_vmemmap_restore(h, &folio->page);
2337 if (!rc) {
2338 update_and_free_hugetlb_folio(h, folio, false);
2339 } else {
2340 spin_lock_irq(&hugetlb_lock);
2341 add_hugetlb_folio(h, folio, false);
2342 h->max_huge_pages++;
2343 spin_unlock_irq(&hugetlb_lock);
2344 }
2345
2346 return rc;
2347 }
2348 out:
2349 spin_unlock_irq(&hugetlb_lock);
2350 return rc;
2351 }
2352
2353 /*
2354 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
2355 * make specified memory blocks removable from the system.
2356 * Note that this will dissolve a free gigantic hugepage completely, if any
2357 * part of it lies within the given range.
2358 * Also note that if dissolve_free_huge_page() returns with an error, all
2359 * free hugepages that were dissolved before that error are lost.
2360 */
dissolve_free_huge_pages(unsigned long start_pfn,unsigned long end_pfn)2361 int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
2362 {
2363 unsigned long pfn;
2364 struct page *page;
2365 int rc = 0;
2366 unsigned int order;
2367 struct hstate *h;
2368
2369 if (!hugepages_supported())
2370 return rc;
2371
2372 order = huge_page_order(&default_hstate);
2373 for_each_hstate(h)
2374 order = min(order, huge_page_order(h));
2375
2376 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) {
2377 page = pfn_to_page(pfn);
2378 rc = dissolve_free_huge_page(page);
2379 if (rc)
2380 break;
2381 }
2382
2383 return rc;
2384 }
2385
2386 /*
2387 * Allocates a fresh surplus page from the page allocator.
2388 */
alloc_surplus_hugetlb_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask)2389 static struct folio *alloc_surplus_hugetlb_folio(struct hstate *h,
2390 gfp_t gfp_mask, int nid, nodemask_t *nmask)
2391 {
2392 struct folio *folio = NULL;
2393
2394 if (hstate_is_gigantic(h))
2395 return NULL;
2396
2397 spin_lock_irq(&hugetlb_lock);
2398 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
2399 goto out_unlock;
2400 spin_unlock_irq(&hugetlb_lock);
2401
2402 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2403 if (!folio)
2404 return NULL;
2405
2406 spin_lock_irq(&hugetlb_lock);
2407 /*
2408 * We could have raced with the pool size change.
2409 * Double check that and simply deallocate the new page
2410 * if we would end up overcommiting the surpluses. Abuse
2411 * temporary page to workaround the nasty free_huge_folio
2412 * codeflow
2413 */
2414 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
2415 folio_set_hugetlb_temporary(folio);
2416 spin_unlock_irq(&hugetlb_lock);
2417 free_huge_folio(folio);
2418 return NULL;
2419 }
2420
2421 h->surplus_huge_pages++;
2422 h->surplus_huge_pages_node[folio_nid(folio)]++;
2423
2424 out_unlock:
2425 spin_unlock_irq(&hugetlb_lock);
2426
2427 return folio;
2428 }
2429
alloc_migrate_hugetlb_folio(struct hstate * h,gfp_t gfp_mask,int nid,nodemask_t * nmask)2430 static struct folio *alloc_migrate_hugetlb_folio(struct hstate *h, gfp_t gfp_mask,
2431 int nid, nodemask_t *nmask)
2432 {
2433 struct folio *folio;
2434
2435 if (hstate_is_gigantic(h))
2436 return NULL;
2437
2438 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid, nmask, NULL);
2439 if (!folio)
2440 return NULL;
2441
2442 /* fresh huge pages are frozen */
2443 folio_ref_unfreeze(folio, 1);
2444 /*
2445 * We do not account these pages as surplus because they are only
2446 * temporary and will be released properly on the last reference
2447 */
2448 folio_set_hugetlb_temporary(folio);
2449
2450 return folio;
2451 }
2452
2453 /*
2454 * Use the VMA's mpolicy to allocate a huge page from the buddy.
2455 */
2456 static
alloc_buddy_hugetlb_folio_with_mpol(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2457 struct folio *alloc_buddy_hugetlb_folio_with_mpol(struct hstate *h,
2458 struct vm_area_struct *vma, unsigned long addr)
2459 {
2460 struct folio *folio = NULL;
2461 struct mempolicy *mpol;
2462 gfp_t gfp_mask = htlb_alloc_mask(h);
2463 int nid;
2464 nodemask_t *nodemask;
2465
2466 nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
2467 if (mpol_is_preferred_many(mpol)) {
2468 gfp_t gfp = gfp_mask | __GFP_NOWARN;
2469
2470 gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2471 folio = alloc_surplus_hugetlb_folio(h, gfp, nid, nodemask);
2472
2473 /* Fallback to all nodes if page==NULL */
2474 nodemask = NULL;
2475 }
2476
2477 if (!folio)
2478 folio = alloc_surplus_hugetlb_folio(h, gfp_mask, nid, nodemask);
2479 mpol_cond_put(mpol);
2480 return folio;
2481 }
2482
2483 /* folio migration callback function */
alloc_hugetlb_folio_nodemask(struct hstate * h,int preferred_nid,nodemask_t * nmask,gfp_t gfp_mask)2484 struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid,
2485 nodemask_t *nmask, gfp_t gfp_mask)
2486 {
2487 spin_lock_irq(&hugetlb_lock);
2488 if (available_huge_pages(h)) {
2489 struct folio *folio;
2490
2491 folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask,
2492 preferred_nid, nmask);
2493 if (folio) {
2494 spin_unlock_irq(&hugetlb_lock);
2495 return folio;
2496 }
2497 }
2498 spin_unlock_irq(&hugetlb_lock);
2499
2500 return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
2501 }
2502
2503 /* mempolicy aware migration callback */
alloc_hugetlb_folio_vma(struct hstate * h,struct vm_area_struct * vma,unsigned long address)2504 struct folio *alloc_hugetlb_folio_vma(struct hstate *h, struct vm_area_struct *vma,
2505 unsigned long address)
2506 {
2507 struct mempolicy *mpol;
2508 nodemask_t *nodemask;
2509 struct folio *folio;
2510 gfp_t gfp_mask;
2511 int node;
2512
2513 gfp_mask = htlb_alloc_mask(h);
2514 node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
2515 folio = alloc_hugetlb_folio_nodemask(h, node, nodemask, gfp_mask);
2516 mpol_cond_put(mpol);
2517
2518 return folio;
2519 }
2520
2521 /*
2522 * Increase the hugetlb pool such that it can accommodate a reservation
2523 * of size 'delta'.
2524 */
gather_surplus_pages(struct hstate * h,long delta)2525 static int gather_surplus_pages(struct hstate *h, long delta)
2526 __must_hold(&hugetlb_lock)
2527 {
2528 LIST_HEAD(surplus_list);
2529 struct folio *folio, *tmp;
2530 int ret;
2531 long i;
2532 long needed, allocated;
2533 bool alloc_ok = true;
2534
2535 lockdep_assert_held(&hugetlb_lock);
2536 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
2537 if (needed <= 0) {
2538 h->resv_huge_pages += delta;
2539 return 0;
2540 }
2541
2542 allocated = 0;
2543
2544 ret = -ENOMEM;
2545 retry:
2546 spin_unlock_irq(&hugetlb_lock);
2547 for (i = 0; i < needed; i++) {
2548 folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
2549 NUMA_NO_NODE, NULL);
2550 if (!folio) {
2551 alloc_ok = false;
2552 break;
2553 }
2554 list_add(&folio->lru, &surplus_list);
2555 cond_resched();
2556 }
2557 allocated += i;
2558
2559 /*
2560 * After retaking hugetlb_lock, we need to recalculate 'needed'
2561 * because either resv_huge_pages or free_huge_pages may have changed.
2562 */
2563 spin_lock_irq(&hugetlb_lock);
2564 needed = (h->resv_huge_pages + delta) -
2565 (h->free_huge_pages + allocated);
2566 if (needed > 0) {
2567 if (alloc_ok)
2568 goto retry;
2569 /*
2570 * We were not able to allocate enough pages to
2571 * satisfy the entire reservation so we free what
2572 * we've allocated so far.
2573 */
2574 goto free;
2575 }
2576 /*
2577 * The surplus_list now contains _at_least_ the number of extra pages
2578 * needed to accommodate the reservation. Add the appropriate number
2579 * of pages to the hugetlb pool and free the extras back to the buddy
2580 * allocator. Commit the entire reservation here to prevent another
2581 * process from stealing the pages as they are added to the pool but
2582 * before they are reserved.
2583 */
2584 needed += allocated;
2585 h->resv_huge_pages += delta;
2586 ret = 0;
2587
2588 /* Free the needed pages to the hugetlb pool */
2589 list_for_each_entry_safe(folio, tmp, &surplus_list, lru) {
2590 if ((--needed) < 0)
2591 break;
2592 /* Add the page to the hugetlb allocator */
2593 enqueue_hugetlb_folio(h, folio);
2594 }
2595 free:
2596 spin_unlock_irq(&hugetlb_lock);
2597
2598 /*
2599 * Free unnecessary surplus pages to the buddy allocator.
2600 * Pages have no ref count, call free_huge_folio directly.
2601 */
2602 list_for_each_entry_safe(folio, tmp, &surplus_list, lru)
2603 free_huge_folio(folio);
2604 spin_lock_irq(&hugetlb_lock);
2605
2606 return ret;
2607 }
2608
2609 /*
2610 * This routine has two main purposes:
2611 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
2612 * in unused_resv_pages. This corresponds to the prior adjustments made
2613 * to the associated reservation map.
2614 * 2) Free any unused surplus pages that may have been allocated to satisfy
2615 * the reservation. As many as unused_resv_pages may be freed.
2616 */
return_unused_surplus_pages(struct hstate * h,unsigned long unused_resv_pages)2617 static void return_unused_surplus_pages(struct hstate *h,
2618 unsigned long unused_resv_pages)
2619 {
2620 unsigned long nr_pages;
2621 struct page *page;
2622 LIST_HEAD(page_list);
2623
2624 lockdep_assert_held(&hugetlb_lock);
2625 /* Uncommit the reservation */
2626 h->resv_huge_pages -= unused_resv_pages;
2627
2628 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2629 goto out;
2630
2631 /*
2632 * Part (or even all) of the reservation could have been backed
2633 * by pre-allocated pages. Only free surplus pages.
2634 */
2635 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
2636
2637 /*
2638 * We want to release as many surplus pages as possible, spread
2639 * evenly across all nodes with memory. Iterate across these nodes
2640 * until we can no longer free unreserved surplus pages. This occurs
2641 * when the nodes with surplus pages have no free pages.
2642 * remove_pool_huge_page() will balance the freed pages across the
2643 * on-line nodes with memory and will handle the hstate accounting.
2644 */
2645 while (nr_pages--) {
2646 page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1);
2647 if (!page)
2648 goto out;
2649
2650 list_add(&page->lru, &page_list);
2651 }
2652
2653 out:
2654 spin_unlock_irq(&hugetlb_lock);
2655 update_and_free_pages_bulk(h, &page_list);
2656 spin_lock_irq(&hugetlb_lock);
2657 }
2658
2659
2660 /*
2661 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
2662 * are used by the huge page allocation routines to manage reservations.
2663 *
2664 * vma_needs_reservation is called to determine if the huge page at addr
2665 * within the vma has an associated reservation. If a reservation is
2666 * needed, the value 1 is returned. The caller is then responsible for
2667 * managing the global reservation and subpool usage counts. After
2668 * the huge page has been allocated, vma_commit_reservation is called
2669 * to add the page to the reservation map. If the page allocation fails,
2670 * the reservation must be ended instead of committed. vma_end_reservation
2671 * is called in such cases.
2672 *
2673 * In the normal case, vma_commit_reservation returns the same value
2674 * as the preceding vma_needs_reservation call. The only time this
2675 * is not the case is if a reserve map was changed between calls. It
2676 * is the responsibility of the caller to notice the difference and
2677 * take appropriate action.
2678 *
2679 * vma_add_reservation is used in error paths where a reservation must
2680 * be restored when a newly allocated huge page must be freed. It is
2681 * to be called after calling vma_needs_reservation to determine if a
2682 * reservation exists.
2683 *
2684 * vma_del_reservation is used in error paths where an entry in the reserve
2685 * map was created during huge page allocation and must be removed. It is to
2686 * be called after calling vma_needs_reservation to determine if a reservation
2687 * exists.
2688 */
2689 enum vma_resv_mode {
2690 VMA_NEEDS_RESV,
2691 VMA_COMMIT_RESV,
2692 VMA_END_RESV,
2693 VMA_ADD_RESV,
2694 VMA_DEL_RESV,
2695 };
__vma_reservation_common(struct hstate * h,struct vm_area_struct * vma,unsigned long addr,enum vma_resv_mode mode)2696 static long __vma_reservation_common(struct hstate *h,
2697 struct vm_area_struct *vma, unsigned long addr,
2698 enum vma_resv_mode mode)
2699 {
2700 struct resv_map *resv;
2701 pgoff_t idx;
2702 long ret;
2703 long dummy_out_regions_needed;
2704
2705 resv = vma_resv_map(vma);
2706 if (!resv)
2707 return 1;
2708
2709 idx = vma_hugecache_offset(h, vma, addr);
2710 switch (mode) {
2711 case VMA_NEEDS_RESV:
2712 ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed);
2713 /* We assume that vma_reservation_* routines always operate on
2714 * 1 page, and that adding to resv map a 1 page entry can only
2715 * ever require 1 region.
2716 */
2717 VM_BUG_ON(dummy_out_regions_needed != 1);
2718 break;
2719 case VMA_COMMIT_RESV:
2720 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2721 /* region_add calls of range 1 should never fail. */
2722 VM_BUG_ON(ret < 0);
2723 break;
2724 case VMA_END_RESV:
2725 region_abort(resv, idx, idx + 1, 1);
2726 ret = 0;
2727 break;
2728 case VMA_ADD_RESV:
2729 if (vma->vm_flags & VM_MAYSHARE) {
2730 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2731 /* region_add calls of range 1 should never fail. */
2732 VM_BUG_ON(ret < 0);
2733 } else {
2734 region_abort(resv, idx, idx + 1, 1);
2735 ret = region_del(resv, idx, idx + 1);
2736 }
2737 break;
2738 case VMA_DEL_RESV:
2739 if (vma->vm_flags & VM_MAYSHARE) {
2740 region_abort(resv, idx, idx + 1, 1);
2741 ret = region_del(resv, idx, idx + 1);
2742 } else {
2743 ret = region_add(resv, idx, idx + 1, 1, NULL, NULL);
2744 /* region_add calls of range 1 should never fail. */
2745 VM_BUG_ON(ret < 0);
2746 }
2747 break;
2748 default:
2749 BUG();
2750 }
2751
2752 if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV)
2753 return ret;
2754 /*
2755 * We know private mapping must have HPAGE_RESV_OWNER set.
2756 *
2757 * In most cases, reserves always exist for private mappings.
2758 * However, a file associated with mapping could have been
2759 * hole punched or truncated after reserves were consumed.
2760 * As subsequent fault on such a range will not use reserves.
2761 * Subtle - The reserve map for private mappings has the
2762 * opposite meaning than that of shared mappings. If NO
2763 * entry is in the reserve map, it means a reservation exists.
2764 * If an entry exists in the reserve map, it means the
2765 * reservation has already been consumed. As a result, the
2766 * return value of this routine is the opposite of the
2767 * value returned from reserve map manipulation routines above.
2768 */
2769 if (ret > 0)
2770 return 0;
2771 if (ret == 0)
2772 return 1;
2773 return ret;
2774 }
2775
vma_needs_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2776 static long vma_needs_reservation(struct hstate *h,
2777 struct vm_area_struct *vma, unsigned long addr)
2778 {
2779 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
2780 }
2781
vma_commit_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2782 static long vma_commit_reservation(struct hstate *h,
2783 struct vm_area_struct *vma, unsigned long addr)
2784 {
2785 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
2786 }
2787
vma_end_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2788 static void vma_end_reservation(struct hstate *h,
2789 struct vm_area_struct *vma, unsigned long addr)
2790 {
2791 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
2792 }
2793
vma_add_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2794 static long vma_add_reservation(struct hstate *h,
2795 struct vm_area_struct *vma, unsigned long addr)
2796 {
2797 return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
2798 }
2799
vma_del_reservation(struct hstate * h,struct vm_area_struct * vma,unsigned long addr)2800 static long vma_del_reservation(struct hstate *h,
2801 struct vm_area_struct *vma, unsigned long addr)
2802 {
2803 return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV);
2804 }
2805
2806 /*
2807 * This routine is called to restore reservation information on error paths.
2808 * It should ONLY be called for folios allocated via alloc_hugetlb_folio(),
2809 * and the hugetlb mutex should remain held when calling this routine.
2810 *
2811 * It handles two specific cases:
2812 * 1) A reservation was in place and the folio consumed the reservation.
2813 * hugetlb_restore_reserve is set in the folio.
2814 * 2) No reservation was in place for the page, so hugetlb_restore_reserve is
2815 * not set. However, alloc_hugetlb_folio always updates the reserve map.
2816 *
2817 * In case 1, free_huge_folio later in the error path will increment the
2818 * global reserve count. But, free_huge_folio does not have enough context
2819 * to adjust the reservation map. This case deals primarily with private
2820 * mappings. Adjust the reserve map here to be consistent with global
2821 * reserve count adjustments to be made by free_huge_folio. Make sure the
2822 * reserve map indicates there is a reservation present.
2823 *
2824 * In case 2, simply undo reserve map modifications done by alloc_hugetlb_folio.
2825 */
restore_reserve_on_error(struct hstate * h,struct vm_area_struct * vma,unsigned long address,struct folio * folio)2826 void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma,
2827 unsigned long address, struct folio *folio)
2828 {
2829 long rc = vma_needs_reservation(h, vma, address);
2830
2831 if (folio_test_hugetlb_restore_reserve(folio)) {
2832 if (unlikely(rc < 0))
2833 /*
2834 * Rare out of memory condition in reserve map
2835 * manipulation. Clear hugetlb_restore_reserve so
2836 * that global reserve count will not be incremented
2837 * by free_huge_folio. This will make it appear
2838 * as though the reservation for this folio was
2839 * consumed. This may prevent the task from
2840 * faulting in the folio at a later time. This
2841 * is better than inconsistent global huge page
2842 * accounting of reserve counts.
2843 */
2844 folio_clear_hugetlb_restore_reserve(folio);
2845 else if (rc)
2846 (void)vma_add_reservation(h, vma, address);
2847 else
2848 vma_end_reservation(h, vma, address);
2849 } else {
2850 if (!rc) {
2851 /*
2852 * This indicates there is an entry in the reserve map
2853 * not added by alloc_hugetlb_folio. We know it was added
2854 * before the alloc_hugetlb_folio call, otherwise
2855 * hugetlb_restore_reserve would be set on the folio.
2856 * Remove the entry so that a subsequent allocation
2857 * does not consume a reservation.
2858 */
2859 rc = vma_del_reservation(h, vma, address);
2860 if (rc < 0)
2861 /*
2862 * VERY rare out of memory condition. Since
2863 * we can not delete the entry, set
2864 * hugetlb_restore_reserve so that the reserve
2865 * count will be incremented when the folio
2866 * is freed. This reserve will be consumed
2867 * on a subsequent allocation.
2868 */
2869 folio_set_hugetlb_restore_reserve(folio);
2870 } else if (rc < 0) {
2871 /*
2872 * Rare out of memory condition from
2873 * vma_needs_reservation call. Memory allocation is
2874 * only attempted if a new entry is needed. Therefore,
2875 * this implies there is not an entry in the
2876 * reserve map.
2877 *
2878 * For shared mappings, no entry in the map indicates
2879 * no reservation. We are done.
2880 */
2881 if (!(vma->vm_flags & VM_MAYSHARE))
2882 /*
2883 * For private mappings, no entry indicates
2884 * a reservation is present. Since we can
2885 * not add an entry, set hugetlb_restore_reserve
2886 * on the folio so reserve count will be
2887 * incremented when freed. This reserve will
2888 * be consumed on a subsequent allocation.
2889 */
2890 folio_set_hugetlb_restore_reserve(folio);
2891 } else
2892 /*
2893 * No reservation present, do nothing
2894 */
2895 vma_end_reservation(h, vma, address);
2896 }
2897 }
2898
2899 /*
2900 * alloc_and_dissolve_hugetlb_folio - Allocate a new folio and dissolve
2901 * the old one
2902 * @h: struct hstate old page belongs to
2903 * @old_folio: Old folio to dissolve
2904 * @list: List to isolate the page in case we need to
2905 * Returns 0 on success, otherwise negated error.
2906 */
alloc_and_dissolve_hugetlb_folio(struct hstate * h,struct folio * old_folio,struct list_head * list)2907 static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
2908 struct folio *old_folio, struct list_head *list)
2909 {
2910 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
2911 int nid = folio_nid(old_folio);
2912 struct folio *new_folio;
2913 int ret = 0;
2914
2915 /*
2916 * Before dissolving the folio, we need to allocate a new one for the
2917 * pool to remain stable. Here, we allocate the folio and 'prep' it
2918 * by doing everything but actually updating counters and adding to
2919 * the pool. This simplifies and let us do most of the processing
2920 * under the lock.
2921 */
2922 new_folio = alloc_buddy_hugetlb_folio(h, gfp_mask, nid, NULL, NULL);
2923 if (!new_folio)
2924 return -ENOMEM;
2925 __prep_new_hugetlb_folio(h, new_folio);
2926
2927 retry:
2928 spin_lock_irq(&hugetlb_lock);
2929 if (!folio_test_hugetlb(old_folio)) {
2930 /*
2931 * Freed from under us. Drop new_folio too.
2932 */
2933 goto free_new;
2934 } else if (folio_ref_count(old_folio)) {
2935 bool isolated;
2936
2937 /*
2938 * Someone has grabbed the folio, try to isolate it here.
2939 * Fail with -EBUSY if not possible.
2940 */
2941 spin_unlock_irq(&hugetlb_lock);
2942 isolated = isolate_hugetlb(old_folio, list);
2943 ret = isolated ? 0 : -EBUSY;
2944 spin_lock_irq(&hugetlb_lock);
2945 goto free_new;
2946 } else if (!folio_test_hugetlb_freed(old_folio)) {
2947 /*
2948 * Folio's refcount is 0 but it has not been enqueued in the
2949 * freelist yet. Race window is small, so we can succeed here if
2950 * we retry.
2951 */
2952 spin_unlock_irq(&hugetlb_lock);
2953 cond_resched();
2954 goto retry;
2955 } else {
2956 /*
2957 * Ok, old_folio is still a genuine free hugepage. Remove it from
2958 * the freelist and decrease the counters. These will be
2959 * incremented again when calling __prep_account_new_huge_page()
2960 * and enqueue_hugetlb_folio() for new_folio. The counters will
2961 * remain stable since this happens under the lock.
2962 */
2963 remove_hugetlb_folio(h, old_folio, false);
2964
2965 /*
2966 * Ref count on new_folio is already zero as it was dropped
2967 * earlier. It can be directly added to the pool free list.
2968 */
2969 __prep_account_new_huge_page(h, nid);
2970 enqueue_hugetlb_folio(h, new_folio);
2971
2972 /*
2973 * Folio has been replaced, we can safely free the old one.
2974 */
2975 spin_unlock_irq(&hugetlb_lock);
2976 update_and_free_hugetlb_folio(h, old_folio, false);
2977 }
2978
2979 return ret;
2980
2981 free_new:
2982 spin_unlock_irq(&hugetlb_lock);
2983 /* Folio has a zero ref count, but needs a ref to be freed */
2984 folio_ref_unfreeze(new_folio, 1);
2985 update_and_free_hugetlb_folio(h, new_folio, false);
2986
2987 return ret;
2988 }
2989
isolate_or_dissolve_huge_page(struct page * page,struct list_head * list)2990 int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
2991 {
2992 struct hstate *h;
2993 struct folio *folio = page_folio(page);
2994 int ret = -EBUSY;
2995
2996 /*
2997 * The page might have been dissolved from under our feet, so make sure
2998 * to carefully check the state under the lock.
2999 * Return success when racing as if we dissolved the page ourselves.
3000 */
3001 spin_lock_irq(&hugetlb_lock);
3002 if (folio_test_hugetlb(folio)) {
3003 h = folio_hstate(folio);
3004 } else {
3005 spin_unlock_irq(&hugetlb_lock);
3006 return 0;
3007 }
3008 spin_unlock_irq(&hugetlb_lock);
3009
3010 /*
3011 * Fence off gigantic pages as there is a cyclic dependency between
3012 * alloc_contig_range and them. Return -ENOMEM as this has the effect
3013 * of bailing out right away without further retrying.
3014 */
3015 if (hstate_is_gigantic(h))
3016 return -ENOMEM;
3017
3018 if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
3019 ret = 0;
3020 else if (!folio_ref_count(folio))
3021 ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
3022
3023 return ret;
3024 }
3025
alloc_hugetlb_folio(struct vm_area_struct * vma,unsigned long addr,int avoid_reserve)3026 struct folio *alloc_hugetlb_folio(struct vm_area_struct *vma,
3027 unsigned long addr, int avoid_reserve)
3028 {
3029 struct hugepage_subpool *spool = subpool_vma(vma);
3030 struct hstate *h = hstate_vma(vma);
3031 struct folio *folio;
3032 long map_chg, map_commit;
3033 long gbl_chg;
3034 int ret, idx;
3035 struct hugetlb_cgroup *h_cg = NULL;
3036 bool deferred_reserve;
3037
3038 idx = hstate_index(h);
3039 /*
3040 * Examine the region/reserve map to determine if the process
3041 * has a reservation for the page to be allocated. A return
3042 * code of zero indicates a reservation exists (no change).
3043 */
3044 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
3045 if (map_chg < 0)
3046 return ERR_PTR(-ENOMEM);
3047
3048 /*
3049 * Processes that did not create the mapping will have no
3050 * reserves as indicated by the region/reserve map. Check
3051 * that the allocation will not exceed the subpool limit.
3052 * Allocations for MAP_NORESERVE mappings also need to be
3053 * checked against any subpool limit.
3054 */
3055 if (map_chg || avoid_reserve) {
3056 gbl_chg = hugepage_subpool_get_pages(spool, 1);
3057 if (gbl_chg < 0) {
3058 vma_end_reservation(h, vma, addr);
3059 return ERR_PTR(-ENOSPC);
3060 }
3061
3062 /*
3063 * Even though there was no reservation in the region/reserve
3064 * map, there could be reservations associated with the
3065 * subpool that can be used. This would be indicated if the
3066 * return value of hugepage_subpool_get_pages() is zero.
3067 * However, if avoid_reserve is specified we still avoid even
3068 * the subpool reservations.
3069 */
3070 if (avoid_reserve)
3071 gbl_chg = 1;
3072 }
3073
3074 /* If this allocation is not consuming a reservation, charge it now.
3075 */
3076 deferred_reserve = map_chg || avoid_reserve;
3077 if (deferred_reserve) {
3078 ret = hugetlb_cgroup_charge_cgroup_rsvd(
3079 idx, pages_per_huge_page(h), &h_cg);
3080 if (ret)
3081 goto out_subpool_put;
3082 }
3083
3084 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
3085 if (ret)
3086 goto out_uncharge_cgroup_reservation;
3087
3088 spin_lock_irq(&hugetlb_lock);
3089 /*
3090 * glb_chg is passed to indicate whether or not a page must be taken
3091 * from the global free pool (global change). gbl_chg == 0 indicates
3092 * a reservation exists for the allocation.
3093 */
3094 folio = dequeue_hugetlb_folio_vma(h, vma, addr, avoid_reserve, gbl_chg);
3095 if (!folio) {
3096 spin_unlock_irq(&hugetlb_lock);
3097 folio = alloc_buddy_hugetlb_folio_with_mpol(h, vma, addr);
3098 if (!folio)
3099 goto out_uncharge_cgroup;
3100 spin_lock_irq(&hugetlb_lock);
3101 if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
3102 folio_set_hugetlb_restore_reserve(folio);
3103 h->resv_huge_pages--;
3104 }
3105 list_add(&folio->lru, &h->hugepage_activelist);
3106 folio_ref_unfreeze(folio, 1);
3107 /* Fall through */
3108 }
3109
3110 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, folio);
3111 /* If allocation is not consuming a reservation, also store the
3112 * hugetlb_cgroup pointer on the page.
3113 */
3114 if (deferred_reserve) {
3115 hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h),
3116 h_cg, folio);
3117 }
3118
3119 spin_unlock_irq(&hugetlb_lock);
3120
3121 hugetlb_set_folio_subpool(folio, spool);
3122
3123 map_commit = vma_commit_reservation(h, vma, addr);
3124 if (unlikely(map_chg > map_commit)) {
3125 /*
3126 * The page was added to the reservation map between
3127 * vma_needs_reservation and vma_commit_reservation.
3128 * This indicates a race with hugetlb_reserve_pages.
3129 * Adjust for the subpool count incremented above AND
3130 * in hugetlb_reserve_pages for the same page. Also,
3131 * the reservation count added in hugetlb_reserve_pages
3132 * no longer applies.
3133 */
3134 long rsv_adjust;
3135
3136 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
3137 hugetlb_acct_memory(h, -rsv_adjust);
3138 if (deferred_reserve) {
3139 spin_lock_irq(&hugetlb_lock);
3140 hugetlb_cgroup_uncharge_folio_rsvd(hstate_index(h),
3141 pages_per_huge_page(h), folio);
3142 spin_unlock_irq(&hugetlb_lock);
3143 }
3144 }
3145 return folio;
3146
3147 out_uncharge_cgroup:
3148 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
3149 out_uncharge_cgroup_reservation:
3150 if (deferred_reserve)
3151 hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h),
3152 h_cg);
3153 out_subpool_put:
3154 if (map_chg || avoid_reserve)
3155 hugepage_subpool_put_pages(spool, 1);
3156 vma_end_reservation(h, vma, addr);
3157 return ERR_PTR(-ENOSPC);
3158 }
3159
3160 int alloc_bootmem_huge_page(struct hstate *h, int nid)
3161 __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
__alloc_bootmem_huge_page(struct hstate * h,int nid)3162 int __alloc_bootmem_huge_page(struct hstate *h, int nid)
3163 {
3164 struct huge_bootmem_page *m = NULL; /* initialize for clang */
3165 int nr_nodes, node;
3166
3167 /* do node specific alloc */
3168 if (nid != NUMA_NO_NODE) {
3169 m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h),
3170 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
3171 if (!m)
3172 return 0;
3173 goto found;
3174 }
3175 /* allocate from next node when distributing huge pages */
3176 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
3177 m = memblock_alloc_try_nid_raw(
3178 huge_page_size(h), huge_page_size(h),
3179 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
3180 /*
3181 * Use the beginning of the huge page to store the
3182 * huge_bootmem_page struct (until gather_bootmem
3183 * puts them into the mem_map).
3184 */
3185 if (!m)
3186 return 0;
3187 goto found;
3188 }
3189
3190 found:
3191 /* Put them into a private list first because mem_map is not up yet */
3192 INIT_LIST_HEAD(&m->list);
3193 list_add(&m->list, &huge_boot_pages);
3194 m->hstate = h;
3195 return 1;
3196 }
3197
3198 /*
3199 * Put bootmem huge pages into the standard lists after mem_map is up.
3200 * Note: This only applies to gigantic (order > MAX_ORDER) pages.
3201 */
gather_bootmem_prealloc(void)3202 static void __init gather_bootmem_prealloc(void)
3203 {
3204 struct huge_bootmem_page *m;
3205
3206 list_for_each_entry(m, &huge_boot_pages, list) {
3207 struct page *page = virt_to_page(m);
3208 struct folio *folio = page_folio(page);
3209 struct hstate *h = m->hstate;
3210
3211 VM_BUG_ON(!hstate_is_gigantic(h));
3212 WARN_ON(folio_ref_count(folio) != 1);
3213 if (prep_compound_gigantic_folio(folio, huge_page_order(h))) {
3214 WARN_ON(folio_test_reserved(folio));
3215 prep_new_hugetlb_folio(h, folio, folio_nid(folio));
3216 free_huge_folio(folio); /* add to the hugepage allocator */
3217 } else {
3218 /* VERY unlikely inflated ref count on a tail page */
3219 free_gigantic_folio(folio, huge_page_order(h));
3220 }
3221
3222 /*
3223 * We need to restore the 'stolen' pages to totalram_pages
3224 * in order to fix confusing memory reports from free(1) and
3225 * other side-effects, like CommitLimit going negative.
3226 */
3227 adjust_managed_page_count(page, pages_per_huge_page(h));
3228 cond_resched();
3229 }
3230 }
hugetlb_hstate_alloc_pages_onenode(struct hstate * h,int nid)3231 static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid)
3232 {
3233 unsigned long i;
3234 char buf[32];
3235
3236 for (i = 0; i < h->max_huge_pages_node[nid]; ++i) {
3237 if (hstate_is_gigantic(h)) {
3238 if (!alloc_bootmem_huge_page(h, nid))
3239 break;
3240 } else {
3241 struct folio *folio;
3242 gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
3243
3244 folio = alloc_fresh_hugetlb_folio(h, gfp_mask, nid,
3245 &node_states[N_MEMORY], NULL);
3246 if (!folio)
3247 break;
3248 free_huge_folio(folio); /* free it into the hugepage allocator */
3249 }
3250 cond_resched();
3251 }
3252 if (i == h->max_huge_pages_node[nid])
3253 return;
3254
3255 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3256 pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n",
3257 h->max_huge_pages_node[nid], buf, nid, i);
3258 h->max_huge_pages -= (h->max_huge_pages_node[nid] - i);
3259 h->max_huge_pages_node[nid] = i;
3260 }
3261
hugetlb_hstate_alloc_pages(struct hstate * h)3262 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
3263 {
3264 unsigned long i;
3265 nodemask_t *node_alloc_noretry;
3266 bool node_specific_alloc = false;
3267
3268 /* skip gigantic hugepages allocation if hugetlb_cma enabled */
3269 if (hstate_is_gigantic(h) && hugetlb_cma_size) {
3270 pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n");
3271 return;
3272 }
3273
3274 /* do node specific alloc */
3275 for_each_online_node(i) {
3276 if (h->max_huge_pages_node[i] > 0) {
3277 hugetlb_hstate_alloc_pages_onenode(h, i);
3278 node_specific_alloc = true;
3279 }
3280 }
3281
3282 if (node_specific_alloc)
3283 return;
3284
3285 /* below will do all node balanced alloc */
3286 if (!hstate_is_gigantic(h)) {
3287 /*
3288 * Bit mask controlling how hard we retry per-node allocations.
3289 * Ignore errors as lower level routines can deal with
3290 * node_alloc_noretry == NULL. If this kmalloc fails at boot
3291 * time, we are likely in bigger trouble.
3292 */
3293 node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
3294 GFP_KERNEL);
3295 } else {
3296 /* allocations done at boot time */
3297 node_alloc_noretry = NULL;
3298 }
3299
3300 /* bit mask controlling how hard we retry per-node allocations */
3301 if (node_alloc_noretry)
3302 nodes_clear(*node_alloc_noretry);
3303
3304 for (i = 0; i < h->max_huge_pages; ++i) {
3305 if (hstate_is_gigantic(h)) {
3306 if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE))
3307 break;
3308 } else if (!alloc_pool_huge_page(h,
3309 &node_states[N_MEMORY],
3310 node_alloc_noretry))
3311 break;
3312 cond_resched();
3313 }
3314 if (i < h->max_huge_pages) {
3315 char buf[32];
3316
3317 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3318 pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
3319 h->max_huge_pages, buf, i);
3320 h->max_huge_pages = i;
3321 }
3322 kfree(node_alloc_noretry);
3323 }
3324
hugetlb_init_hstates(void)3325 static void __init hugetlb_init_hstates(void)
3326 {
3327 struct hstate *h, *h2;
3328
3329 for_each_hstate(h) {
3330 /* oversize hugepages were init'ed in early boot */
3331 if (!hstate_is_gigantic(h))
3332 hugetlb_hstate_alloc_pages(h);
3333
3334 /*
3335 * Set demote order for each hstate. Note that
3336 * h->demote_order is initially 0.
3337 * - We can not demote gigantic pages if runtime freeing
3338 * is not supported, so skip this.
3339 * - If CMA allocation is possible, we can not demote
3340 * HUGETLB_PAGE_ORDER or smaller size pages.
3341 */
3342 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3343 continue;
3344 if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER)
3345 continue;
3346 for_each_hstate(h2) {
3347 if (h2 == h)
3348 continue;
3349 if (h2->order < h->order &&
3350 h2->order > h->demote_order)
3351 h->demote_order = h2->order;
3352 }
3353 }
3354 }
3355
report_hugepages(void)3356 static void __init report_hugepages(void)
3357 {
3358 struct hstate *h;
3359
3360 for_each_hstate(h) {
3361 char buf[32];
3362
3363 string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
3364 pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n",
3365 buf, h->free_huge_pages);
3366 pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n",
3367 hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf);
3368 }
3369 }
3370
3371 #ifdef CONFIG_HIGHMEM
try_to_free_low(struct hstate * h,unsigned long count,nodemask_t * nodes_allowed)3372 static void try_to_free_low(struct hstate *h, unsigned long count,
3373 nodemask_t *nodes_allowed)
3374 {
3375 int i;
3376 LIST_HEAD(page_list);
3377
3378 lockdep_assert_held(&hugetlb_lock);
3379 if (hstate_is_gigantic(h))
3380 return;
3381
3382 /*
3383 * Collect pages to be freed on a list, and free after dropping lock
3384 */
3385 for_each_node_mask(i, *nodes_allowed) {
3386 struct page *page, *next;
3387 struct list_head *freel = &h->hugepage_freelists[i];
3388 list_for_each_entry_safe(page, next, freel, lru) {
3389 if (count >= h->nr_huge_pages)
3390 goto out;
3391 if (PageHighMem(page))
3392 continue;
3393 remove_hugetlb_folio(h, page_folio(page), false);
3394 list_add(&page->lru, &page_list);
3395 }
3396 }
3397
3398 out:
3399 spin_unlock_irq(&hugetlb_lock);
3400 update_and_free_pages_bulk(h, &page_list);
3401 spin_lock_irq(&hugetlb_lock);
3402 }
3403 #else
try_to_free_low(struct hstate * h,unsigned long count,nodemask_t * nodes_allowed)3404 static inline void try_to_free_low(struct hstate *h, unsigned long count,
3405 nodemask_t *nodes_allowed)
3406 {
3407 }
3408 #endif
3409
3410 /*
3411 * Increment or decrement surplus_huge_pages. Keep node-specific counters
3412 * balanced by operating on them in a round-robin fashion.
3413 * Returns 1 if an adjustment was made.
3414 */
adjust_pool_surplus(struct hstate * h,nodemask_t * nodes_allowed,int delta)3415 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
3416 int delta)
3417 {
3418 int nr_nodes, node;
3419
3420 lockdep_assert_held(&hugetlb_lock);
3421 VM_BUG_ON(delta != -1 && delta != 1);
3422
3423 if (delta < 0) {
3424 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
3425 if (h->surplus_huge_pages_node[node])
3426 goto found;
3427 }
3428 } else {
3429 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3430 if (h->surplus_huge_pages_node[node] <
3431 h->nr_huge_pages_node[node])
3432 goto found;
3433 }
3434 }
3435 return 0;
3436
3437 found:
3438 h->surplus_huge_pages += delta;
3439 h->surplus_huge_pages_node[node] += delta;
3440 return 1;
3441 }
3442
3443 #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)3444 static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
3445 nodemask_t *nodes_allowed)
3446 {
3447 unsigned long min_count, ret;
3448 struct page *page;
3449 LIST_HEAD(page_list);
3450 NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
3451
3452 /*
3453 * Bit mask controlling how hard we retry per-node allocations.
3454 * If we can not allocate the bit mask, do not attempt to allocate
3455 * the requested huge pages.
3456 */
3457 if (node_alloc_noretry)
3458 nodes_clear(*node_alloc_noretry);
3459 else
3460 return -ENOMEM;
3461
3462 /*
3463 * resize_lock mutex prevents concurrent adjustments to number of
3464 * pages in hstate via the proc/sysfs interfaces.
3465 */
3466 mutex_lock(&h->resize_lock);
3467 flush_free_hpage_work(h);
3468 spin_lock_irq(&hugetlb_lock);
3469
3470 /*
3471 * Check for a node specific request.
3472 * Changing node specific huge page count may require a corresponding
3473 * change to the global count. In any case, the passed node mask
3474 * (nodes_allowed) will restrict alloc/free to the specified node.
3475 */
3476 if (nid != NUMA_NO_NODE) {
3477 unsigned long old_count = count;
3478
3479 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
3480 /*
3481 * User may have specified a large count value which caused the
3482 * above calculation to overflow. In this case, they wanted
3483 * to allocate as many huge pages as possible. Set count to
3484 * largest possible value to align with their intention.
3485 */
3486 if (count < old_count)
3487 count = ULONG_MAX;
3488 }
3489
3490 /*
3491 * Gigantic pages runtime allocation depend on the capability for large
3492 * page range allocation.
3493 * If the system does not provide this feature, return an error when
3494 * the user tries to allocate gigantic pages but let the user free the
3495 * boottime allocated gigantic pages.
3496 */
3497 if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
3498 if (count > persistent_huge_pages(h)) {
3499 spin_unlock_irq(&hugetlb_lock);
3500 mutex_unlock(&h->resize_lock);
3501 NODEMASK_FREE(node_alloc_noretry);
3502 return -EINVAL;
3503 }
3504 /* Fall through to decrease pool */
3505 }
3506
3507 /*
3508 * Increase the pool size
3509 * First take pages out of surplus state. Then make up the
3510 * remaining difference by allocating fresh huge pages.
3511 *
3512 * We might race with alloc_surplus_hugetlb_folio() here and be unable
3513 * to convert a surplus huge page to a normal huge page. That is
3514 * not critical, though, it just means the overall size of the
3515 * pool might be one hugepage larger than it needs to be, but
3516 * within all the constraints specified by the sysctls.
3517 */
3518 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
3519 if (!adjust_pool_surplus(h, nodes_allowed, -1))
3520 break;
3521 }
3522
3523 while (count > persistent_huge_pages(h)) {
3524 /*
3525 * If this allocation races such that we no longer need the
3526 * page, free_huge_folio will handle it by freeing the page
3527 * and reducing the surplus.
3528 */
3529 spin_unlock_irq(&hugetlb_lock);
3530
3531 /* yield cpu to avoid soft lockup */
3532 cond_resched();
3533
3534 ret = alloc_pool_huge_page(h, nodes_allowed,
3535 node_alloc_noretry);
3536 spin_lock_irq(&hugetlb_lock);
3537 if (!ret)
3538 goto out;
3539
3540 /* Bail for signals. Probably ctrl-c from user */
3541 if (signal_pending(current))
3542 goto out;
3543 }
3544
3545 /*
3546 * Decrease the pool size
3547 * First return free pages to the buddy allocator (being careful
3548 * to keep enough around to satisfy reservations). Then place
3549 * pages into surplus state as needed so the pool will shrink
3550 * to the desired size as pages become free.
3551 *
3552 * By placing pages into the surplus state independent of the
3553 * overcommit value, we are allowing the surplus pool size to
3554 * exceed overcommit. There are few sane options here. Since
3555 * alloc_surplus_hugetlb_folio() is checking the global counter,
3556 * though, we'll note that we're not allowed to exceed surplus
3557 * and won't grow the pool anywhere else. Not until one of the
3558 * sysctls are changed, or the surplus pages go out of use.
3559 */
3560 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
3561 min_count = max(count, min_count);
3562 try_to_free_low(h, min_count, nodes_allowed);
3563
3564 /*
3565 * Collect pages to be removed on list without dropping lock
3566 */
3567 while (min_count < persistent_huge_pages(h)) {
3568 page = remove_pool_huge_page(h, nodes_allowed, 0);
3569 if (!page)
3570 break;
3571
3572 list_add(&page->lru, &page_list);
3573 }
3574 /* free the pages after dropping lock */
3575 spin_unlock_irq(&hugetlb_lock);
3576 update_and_free_pages_bulk(h, &page_list);
3577 flush_free_hpage_work(h);
3578 spin_lock_irq(&hugetlb_lock);
3579
3580 while (count < persistent_huge_pages(h)) {
3581 if (!adjust_pool_surplus(h, nodes_allowed, 1))
3582 break;
3583 }
3584 out:
3585 h->max_huge_pages = persistent_huge_pages(h);
3586 spin_unlock_irq(&hugetlb_lock);
3587 mutex_unlock(&h->resize_lock);
3588
3589 NODEMASK_FREE(node_alloc_noretry);
3590
3591 return 0;
3592 }
3593
demote_free_hugetlb_folio(struct hstate * h,struct folio * folio)3594 static int demote_free_hugetlb_folio(struct hstate *h, struct folio *folio)
3595 {
3596 int i, nid = folio_nid(folio);
3597 struct hstate *target_hstate;
3598 struct page *subpage;
3599 struct folio *inner_folio;
3600 int rc = 0;
3601
3602 target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order);
3603
3604 remove_hugetlb_folio_for_demote(h, folio, false);
3605 spin_unlock_irq(&hugetlb_lock);
3606
3607 rc = hugetlb_vmemmap_restore(h, &folio->page);
3608 if (rc) {
3609 /* Allocation of vmemmmap failed, we can not demote folio */
3610 spin_lock_irq(&hugetlb_lock);
3611 folio_ref_unfreeze(folio, 1);
3612 add_hugetlb_folio(h, folio, false);
3613 return rc;
3614 }
3615
3616 /*
3617 * Use destroy_compound_hugetlb_folio_for_demote for all huge page
3618 * sizes as it will not ref count folios.
3619 */
3620 destroy_compound_hugetlb_folio_for_demote(folio, huge_page_order(h));
3621
3622 /*
3623 * Taking target hstate mutex synchronizes with set_max_huge_pages.
3624 * Without the mutex, pages added to target hstate could be marked
3625 * as surplus.
3626 *
3627 * Note that we already hold h->resize_lock. To prevent deadlock,
3628 * use the convention of always taking larger size hstate mutex first.
3629 */
3630 mutex_lock(&target_hstate->resize_lock);
3631 for (i = 0; i < pages_per_huge_page(h);
3632 i += pages_per_huge_page(target_hstate)) {
3633 subpage = folio_page(folio, i);
3634 inner_folio = page_folio(subpage);
3635 if (hstate_is_gigantic(target_hstate))
3636 prep_compound_gigantic_folio_for_demote(inner_folio,
3637 target_hstate->order);
3638 else
3639 prep_compound_page(subpage, target_hstate->order);
3640 folio_change_private(inner_folio, NULL);
3641 prep_new_hugetlb_folio(target_hstate, inner_folio, nid);
3642 free_huge_folio(inner_folio);
3643 }
3644 mutex_unlock(&target_hstate->resize_lock);
3645
3646 spin_lock_irq(&hugetlb_lock);
3647
3648 /*
3649 * Not absolutely necessary, but for consistency update max_huge_pages
3650 * based on pool changes for the demoted page.
3651 */
3652 h->max_huge_pages--;
3653 target_hstate->max_huge_pages +=
3654 pages_per_huge_page(h) / pages_per_huge_page(target_hstate);
3655
3656 return rc;
3657 }
3658
demote_pool_huge_page(struct hstate * h,nodemask_t * nodes_allowed)3659 static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
3660 __must_hold(&hugetlb_lock)
3661 {
3662 int nr_nodes, node;
3663 struct folio *folio;
3664
3665 lockdep_assert_held(&hugetlb_lock);
3666
3667 /* We should never get here if no demote order */
3668 if (!h->demote_order) {
3669 pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n");
3670 return -EINVAL; /* internal error */
3671 }
3672
3673 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
3674 list_for_each_entry(folio, &h->hugepage_freelists[node], lru) {
3675 if (folio_test_hwpoison(folio))
3676 continue;
3677 return demote_free_hugetlb_folio(h, folio);
3678 }
3679 }
3680
3681 /*
3682 * Only way to get here is if all pages on free lists are poisoned.
3683 * Return -EBUSY so that caller will not retry.
3684 */
3685 return -EBUSY;
3686 }
3687
3688 #define HSTATE_ATTR_RO(_name) \
3689 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
3690
3691 #define HSTATE_ATTR_WO(_name) \
3692 static struct kobj_attribute _name##_attr = __ATTR_WO(_name)
3693
3694 #define HSTATE_ATTR(_name) \
3695 static struct kobj_attribute _name##_attr = __ATTR_RW(_name)
3696
3697 static struct kobject *hugepages_kobj;
3698 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
3699
3700 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
3701
kobj_to_hstate(struct kobject * kobj,int * nidp)3702 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
3703 {
3704 int i;
3705
3706 for (i = 0; i < HUGE_MAX_HSTATE; i++)
3707 if (hstate_kobjs[i] == kobj) {
3708 if (nidp)
3709 *nidp = NUMA_NO_NODE;
3710 return &hstates[i];
3711 }
3712
3713 return kobj_to_node_hstate(kobj, nidp);
3714 }
3715
nr_hugepages_show_common(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3716 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
3717 struct kobj_attribute *attr, char *buf)
3718 {
3719 struct hstate *h;
3720 unsigned long nr_huge_pages;
3721 int nid;
3722
3723 h = kobj_to_hstate(kobj, &nid);
3724 if (nid == NUMA_NO_NODE)
3725 nr_huge_pages = h->nr_huge_pages;
3726 else
3727 nr_huge_pages = h->nr_huge_pages_node[nid];
3728
3729 return sysfs_emit(buf, "%lu\n", nr_huge_pages);
3730 }
3731
__nr_hugepages_store_common(bool obey_mempolicy,struct hstate * h,int nid,unsigned long count,size_t len)3732 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
3733 struct hstate *h, int nid,
3734 unsigned long count, size_t len)
3735 {
3736 int err;
3737 nodemask_t nodes_allowed, *n_mask;
3738
3739 if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
3740 return -EINVAL;
3741
3742 if (nid == NUMA_NO_NODE) {
3743 /*
3744 * global hstate attribute
3745 */
3746 if (!(obey_mempolicy &&
3747 init_nodemask_of_mempolicy(&nodes_allowed)))
3748 n_mask = &node_states[N_MEMORY];
3749 else
3750 n_mask = &nodes_allowed;
3751 } else {
3752 /*
3753 * Node specific request. count adjustment happens in
3754 * set_max_huge_pages() after acquiring hugetlb_lock.
3755 */
3756 init_nodemask_of_node(&nodes_allowed, nid);
3757 n_mask = &nodes_allowed;
3758 }
3759
3760 err = set_max_huge_pages(h, count, nid, n_mask);
3761
3762 return err ? err : len;
3763 }
3764
nr_hugepages_store_common(bool obey_mempolicy,struct kobject * kobj,const char * buf,size_t len)3765 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
3766 struct kobject *kobj, const char *buf,
3767 size_t len)
3768 {
3769 struct hstate *h;
3770 unsigned long count;
3771 int nid;
3772 int err;
3773
3774 err = kstrtoul(buf, 10, &count);
3775 if (err)
3776 return err;
3777
3778 h = kobj_to_hstate(kobj, &nid);
3779 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
3780 }
3781
nr_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3782 static ssize_t nr_hugepages_show(struct kobject *kobj,
3783 struct kobj_attribute *attr, char *buf)
3784 {
3785 return nr_hugepages_show_common(kobj, attr, buf);
3786 }
3787
nr_hugepages_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)3788 static ssize_t nr_hugepages_store(struct kobject *kobj,
3789 struct kobj_attribute *attr, const char *buf, size_t len)
3790 {
3791 return nr_hugepages_store_common(false, kobj, buf, len);
3792 }
3793 HSTATE_ATTR(nr_hugepages);
3794
3795 #ifdef CONFIG_NUMA
3796
3797 /*
3798 * hstate attribute for optionally mempolicy-based constraint on persistent
3799 * huge page alloc/free.
3800 */
nr_hugepages_mempolicy_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3801 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
3802 struct kobj_attribute *attr,
3803 char *buf)
3804 {
3805 return nr_hugepages_show_common(kobj, attr, buf);
3806 }
3807
nr_hugepages_mempolicy_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)3808 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
3809 struct kobj_attribute *attr, const char *buf, size_t len)
3810 {
3811 return nr_hugepages_store_common(true, kobj, buf, len);
3812 }
3813 HSTATE_ATTR(nr_hugepages_mempolicy);
3814 #endif
3815
3816
nr_overcommit_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3817 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
3818 struct kobj_attribute *attr, char *buf)
3819 {
3820 struct hstate *h = kobj_to_hstate(kobj, NULL);
3821 return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages);
3822 }
3823
nr_overcommit_hugepages_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)3824 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
3825 struct kobj_attribute *attr, const char *buf, size_t count)
3826 {
3827 int err;
3828 unsigned long input;
3829 struct hstate *h = kobj_to_hstate(kobj, NULL);
3830
3831 if (hstate_is_gigantic(h))
3832 return -EINVAL;
3833
3834 err = kstrtoul(buf, 10, &input);
3835 if (err)
3836 return err;
3837
3838 spin_lock_irq(&hugetlb_lock);
3839 h->nr_overcommit_huge_pages = input;
3840 spin_unlock_irq(&hugetlb_lock);
3841
3842 return count;
3843 }
3844 HSTATE_ATTR(nr_overcommit_hugepages);
3845
free_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3846 static ssize_t free_hugepages_show(struct kobject *kobj,
3847 struct kobj_attribute *attr, char *buf)
3848 {
3849 struct hstate *h;
3850 unsigned long free_huge_pages;
3851 int nid;
3852
3853 h = kobj_to_hstate(kobj, &nid);
3854 if (nid == NUMA_NO_NODE)
3855 free_huge_pages = h->free_huge_pages;
3856 else
3857 free_huge_pages = h->free_huge_pages_node[nid];
3858
3859 return sysfs_emit(buf, "%lu\n", free_huge_pages);
3860 }
3861 HSTATE_ATTR_RO(free_hugepages);
3862
resv_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3863 static ssize_t resv_hugepages_show(struct kobject *kobj,
3864 struct kobj_attribute *attr, char *buf)
3865 {
3866 struct hstate *h = kobj_to_hstate(kobj, NULL);
3867 return sysfs_emit(buf, "%lu\n", h->resv_huge_pages);
3868 }
3869 HSTATE_ATTR_RO(resv_hugepages);
3870
surplus_hugepages_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3871 static ssize_t surplus_hugepages_show(struct kobject *kobj,
3872 struct kobj_attribute *attr, char *buf)
3873 {
3874 struct hstate *h;
3875 unsigned long surplus_huge_pages;
3876 int nid;
3877
3878 h = kobj_to_hstate(kobj, &nid);
3879 if (nid == NUMA_NO_NODE)
3880 surplus_huge_pages = h->surplus_huge_pages;
3881 else
3882 surplus_huge_pages = h->surplus_huge_pages_node[nid];
3883
3884 return sysfs_emit(buf, "%lu\n", surplus_huge_pages);
3885 }
3886 HSTATE_ATTR_RO(surplus_hugepages);
3887
demote_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t len)3888 static ssize_t demote_store(struct kobject *kobj,
3889 struct kobj_attribute *attr, const char *buf, size_t len)
3890 {
3891 unsigned long nr_demote;
3892 unsigned long nr_available;
3893 nodemask_t nodes_allowed, *n_mask;
3894 struct hstate *h;
3895 int err;
3896 int nid;
3897
3898 err = kstrtoul(buf, 10, &nr_demote);
3899 if (err)
3900 return err;
3901 h = kobj_to_hstate(kobj, &nid);
3902
3903 if (nid != NUMA_NO_NODE) {
3904 init_nodemask_of_node(&nodes_allowed, nid);
3905 n_mask = &nodes_allowed;
3906 } else {
3907 n_mask = &node_states[N_MEMORY];
3908 }
3909
3910 /* Synchronize with other sysfs operations modifying huge pages */
3911 mutex_lock(&h->resize_lock);
3912 spin_lock_irq(&hugetlb_lock);
3913
3914 while (nr_demote) {
3915 /*
3916 * Check for available pages to demote each time thorough the
3917 * loop as demote_pool_huge_page will drop hugetlb_lock.
3918 */
3919 if (nid != NUMA_NO_NODE)
3920 nr_available = h->free_huge_pages_node[nid];
3921 else
3922 nr_available = h->free_huge_pages;
3923 nr_available -= h->resv_huge_pages;
3924 if (!nr_available)
3925 break;
3926
3927 err = demote_pool_huge_page(h, n_mask);
3928 if (err)
3929 break;
3930
3931 nr_demote--;
3932 }
3933
3934 spin_unlock_irq(&hugetlb_lock);
3935 mutex_unlock(&h->resize_lock);
3936
3937 if (err)
3938 return err;
3939 return len;
3940 }
3941 HSTATE_ATTR_WO(demote);
3942
demote_size_show(struct kobject * kobj,struct kobj_attribute * attr,char * buf)3943 static ssize_t demote_size_show(struct kobject *kobj,
3944 struct kobj_attribute *attr, char *buf)
3945 {
3946 struct hstate *h = kobj_to_hstate(kobj, NULL);
3947 unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K;
3948
3949 return sysfs_emit(buf, "%lukB\n", demote_size);
3950 }
3951
demote_size_store(struct kobject * kobj,struct kobj_attribute * attr,const char * buf,size_t count)3952 static ssize_t demote_size_store(struct kobject *kobj,
3953 struct kobj_attribute *attr,
3954 const char *buf, size_t count)
3955 {
3956 struct hstate *h, *demote_hstate;
3957 unsigned long demote_size;
3958 unsigned int demote_order;
3959
3960 demote_size = (unsigned long)memparse(buf, NULL);
3961
3962 demote_hstate = size_to_hstate(demote_size);
3963 if (!demote_hstate)
3964 return -EINVAL;
3965 demote_order = demote_hstate->order;
3966 if (demote_order < HUGETLB_PAGE_ORDER)
3967 return -EINVAL;
3968
3969 /* demote order must be smaller than hstate order */
3970 h = kobj_to_hstate(kobj, NULL);
3971 if (demote_order >= h->order)
3972 return -EINVAL;
3973
3974 /* resize_lock synchronizes access to demote size and writes */
3975 mutex_lock(&h->resize_lock);
3976 h->demote_order = demote_order;
3977 mutex_unlock(&h->resize_lock);
3978
3979 return count;
3980 }
3981 HSTATE_ATTR(demote_size);
3982
3983 static struct attribute *hstate_attrs[] = {
3984 &nr_hugepages_attr.attr,
3985 &nr_overcommit_hugepages_attr.attr,
3986 &free_hugepages_attr.attr,
3987 &resv_hugepages_attr.attr,
3988 &surplus_hugepages_attr.attr,
3989 #ifdef CONFIG_NUMA
3990 &nr_hugepages_mempolicy_attr.attr,
3991 #endif
3992 NULL,
3993 };
3994
3995 static const struct attribute_group hstate_attr_group = {
3996 .attrs = hstate_attrs,
3997 };
3998
3999 static struct attribute *hstate_demote_attrs[] = {
4000 &demote_size_attr.attr,
4001 &demote_attr.attr,
4002 NULL,
4003 };
4004
4005 static const struct attribute_group hstate_demote_attr_group = {
4006 .attrs = hstate_demote_attrs,
4007 };
4008
hugetlb_sysfs_add_hstate(struct hstate * h,struct kobject * parent,struct kobject ** hstate_kobjs,const struct attribute_group * hstate_attr_group)4009 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
4010 struct kobject **hstate_kobjs,
4011 const struct attribute_group *hstate_attr_group)
4012 {
4013 int retval;
4014 int hi = hstate_index(h);
4015
4016 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
4017 if (!hstate_kobjs[hi])
4018 return -ENOMEM;
4019
4020 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
4021 if (retval) {
4022 kobject_put(hstate_kobjs[hi]);
4023 hstate_kobjs[hi] = NULL;
4024 return retval;
4025 }
4026
4027 if (h->demote_order) {
4028 retval = sysfs_create_group(hstate_kobjs[hi],
4029 &hstate_demote_attr_group);
4030 if (retval) {
4031 pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name);
4032 sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group);
4033 kobject_put(hstate_kobjs[hi]);
4034 hstate_kobjs[hi] = NULL;
4035 return retval;
4036 }
4037 }
4038
4039 return 0;
4040 }
4041
4042 #ifdef CONFIG_NUMA
4043 static bool hugetlb_sysfs_initialized __ro_after_init;
4044
4045 /*
4046 * node_hstate/s - associate per node hstate attributes, via their kobjects,
4047 * with node devices in node_devices[] using a parallel array. The array
4048 * index of a node device or _hstate == node id.
4049 * This is here to avoid any static dependency of the node device driver, in
4050 * the base kernel, on the hugetlb module.
4051 */
4052 struct node_hstate {
4053 struct kobject *hugepages_kobj;
4054 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
4055 };
4056 static struct node_hstate node_hstates[MAX_NUMNODES];
4057
4058 /*
4059 * A subset of global hstate attributes for node devices
4060 */
4061 static struct attribute *per_node_hstate_attrs[] = {
4062 &nr_hugepages_attr.attr,
4063 &free_hugepages_attr.attr,
4064 &surplus_hugepages_attr.attr,
4065 NULL,
4066 };
4067
4068 static const struct attribute_group per_node_hstate_attr_group = {
4069 .attrs = per_node_hstate_attrs,
4070 };
4071
4072 /*
4073 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
4074 * Returns node id via non-NULL nidp.
4075 */
kobj_to_node_hstate(struct kobject * kobj,int * nidp)4076 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4077 {
4078 int nid;
4079
4080 for (nid = 0; nid < nr_node_ids; nid++) {
4081 struct node_hstate *nhs = &node_hstates[nid];
4082 int i;
4083 for (i = 0; i < HUGE_MAX_HSTATE; i++)
4084 if (nhs->hstate_kobjs[i] == kobj) {
4085 if (nidp)
4086 *nidp = nid;
4087 return &hstates[i];
4088 }
4089 }
4090
4091 BUG();
4092 return NULL;
4093 }
4094
4095 /*
4096 * Unregister hstate attributes from a single node device.
4097 * No-op if no hstate attributes attached.
4098 */
hugetlb_unregister_node(struct node * node)4099 void hugetlb_unregister_node(struct node *node)
4100 {
4101 struct hstate *h;
4102 struct node_hstate *nhs = &node_hstates[node->dev.id];
4103
4104 if (!nhs->hugepages_kobj)
4105 return; /* no hstate attributes */
4106
4107 for_each_hstate(h) {
4108 int idx = hstate_index(h);
4109 struct kobject *hstate_kobj = nhs->hstate_kobjs[idx];
4110
4111 if (!hstate_kobj)
4112 continue;
4113 if (h->demote_order)
4114 sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group);
4115 sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group);
4116 kobject_put(hstate_kobj);
4117 nhs->hstate_kobjs[idx] = NULL;
4118 }
4119
4120 kobject_put(nhs->hugepages_kobj);
4121 nhs->hugepages_kobj = NULL;
4122 }
4123
4124
4125 /*
4126 * Register hstate attributes for a single node device.
4127 * No-op if attributes already registered.
4128 */
hugetlb_register_node(struct node * node)4129 void hugetlb_register_node(struct node *node)
4130 {
4131 struct hstate *h;
4132 struct node_hstate *nhs = &node_hstates[node->dev.id];
4133 int err;
4134
4135 if (!hugetlb_sysfs_initialized)
4136 return;
4137
4138 if (nhs->hugepages_kobj)
4139 return; /* already allocated */
4140
4141 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
4142 &node->dev.kobj);
4143 if (!nhs->hugepages_kobj)
4144 return;
4145
4146 for_each_hstate(h) {
4147 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
4148 nhs->hstate_kobjs,
4149 &per_node_hstate_attr_group);
4150 if (err) {
4151 pr_err("HugeTLB: Unable to add hstate %s for node %d\n",
4152 h->name, node->dev.id);
4153 hugetlb_unregister_node(node);
4154 break;
4155 }
4156 }
4157 }
4158
4159 /*
4160 * hugetlb init time: register hstate attributes for all registered node
4161 * devices of nodes that have memory. All on-line nodes should have
4162 * registered their associated device by this time.
4163 */
hugetlb_register_all_nodes(void)4164 static void __init hugetlb_register_all_nodes(void)
4165 {
4166 int nid;
4167
4168 for_each_online_node(nid)
4169 hugetlb_register_node(node_devices[nid]);
4170 }
4171 #else /* !CONFIG_NUMA */
4172
kobj_to_node_hstate(struct kobject * kobj,int * nidp)4173 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
4174 {
4175 BUG();
4176 if (nidp)
4177 *nidp = -1;
4178 return NULL;
4179 }
4180
hugetlb_register_all_nodes(void)4181 static void hugetlb_register_all_nodes(void) { }
4182
4183 #endif
4184
4185 #ifdef CONFIG_CMA
4186 static void __init hugetlb_cma_check(void);
4187 #else
hugetlb_cma_check(void)4188 static inline __init void hugetlb_cma_check(void)
4189 {
4190 }
4191 #endif
4192
hugetlb_sysfs_init(void)4193 static void __init hugetlb_sysfs_init(void)
4194 {
4195 struct hstate *h;
4196 int err;
4197
4198 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
4199 if (!hugepages_kobj)
4200 return;
4201
4202 for_each_hstate(h) {
4203 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
4204 hstate_kobjs, &hstate_attr_group);
4205 if (err)
4206 pr_err("HugeTLB: Unable to add hstate %s", h->name);
4207 }
4208
4209 #ifdef CONFIG_NUMA
4210 hugetlb_sysfs_initialized = true;
4211 #endif
4212 hugetlb_register_all_nodes();
4213 }
4214
4215 #ifdef CONFIG_SYSCTL
4216 static void hugetlb_sysctl_init(void);
4217 #else
hugetlb_sysctl_init(void)4218 static inline void hugetlb_sysctl_init(void) { }
4219 #endif
4220
hugetlb_init(void)4221 static int __init hugetlb_init(void)
4222 {
4223 int i;
4224
4225 BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE <
4226 __NR_HPAGEFLAGS);
4227
4228 if (!hugepages_supported()) {
4229 if (hugetlb_max_hstate || default_hstate_max_huge_pages)
4230 pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n");
4231 return 0;
4232 }
4233
4234 /*
4235 * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some
4236 * architectures depend on setup being done here.
4237 */
4238 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
4239 if (!parsed_default_hugepagesz) {
4240 /*
4241 * If we did not parse a default huge page size, set
4242 * default_hstate_idx to HPAGE_SIZE hstate. And, if the
4243 * number of huge pages for this default size was implicitly
4244 * specified, set that here as well.
4245 * Note that the implicit setting will overwrite an explicit
4246 * setting. A warning will be printed in this case.
4247 */
4248 default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE));
4249 if (default_hstate_max_huge_pages) {
4250 if (default_hstate.max_huge_pages) {
4251 char buf[32];
4252
4253 string_get_size(huge_page_size(&default_hstate),
4254 1, STRING_UNITS_2, buf, 32);
4255 pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n",
4256 default_hstate.max_huge_pages, buf);
4257 pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n",
4258 default_hstate_max_huge_pages);
4259 }
4260 default_hstate.max_huge_pages =
4261 default_hstate_max_huge_pages;
4262
4263 for_each_online_node(i)
4264 default_hstate.max_huge_pages_node[i] =
4265 default_hugepages_in_node[i];
4266 }
4267 }
4268
4269 hugetlb_cma_check();
4270 hugetlb_init_hstates();
4271 gather_bootmem_prealloc();
4272 report_hugepages();
4273
4274 hugetlb_sysfs_init();
4275 hugetlb_cgroup_file_init();
4276 hugetlb_sysctl_init();
4277
4278 #ifdef CONFIG_SMP
4279 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
4280 #else
4281 num_fault_mutexes = 1;
4282 #endif
4283 hugetlb_fault_mutex_table =
4284 kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
4285 GFP_KERNEL);
4286 BUG_ON(!hugetlb_fault_mutex_table);
4287
4288 for (i = 0; i < num_fault_mutexes; i++)
4289 mutex_init(&hugetlb_fault_mutex_table[i]);
4290 return 0;
4291 }
4292 subsys_initcall(hugetlb_init);
4293
4294 /* Overwritten by architectures with more huge page sizes */
__init(weak)4295 bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size)
4296 {
4297 return size == HPAGE_SIZE;
4298 }
4299
hugetlb_add_hstate(unsigned int order)4300 void __init hugetlb_add_hstate(unsigned int order)
4301 {
4302 struct hstate *h;
4303 unsigned long i;
4304
4305 if (size_to_hstate(PAGE_SIZE << order)) {
4306 return;
4307 }
4308 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
4309 BUG_ON(order == 0);
4310 h = &hstates[hugetlb_max_hstate++];
4311 mutex_init(&h->resize_lock);
4312 h->order = order;
4313 h->mask = ~(huge_page_size(h) - 1);
4314 for (i = 0; i < MAX_NUMNODES; ++i)
4315 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
4316 INIT_LIST_HEAD(&h->hugepage_activelist);
4317 h->next_nid_to_alloc = first_memory_node;
4318 h->next_nid_to_free = first_memory_node;
4319 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
4320 huge_page_size(h)/SZ_1K);
4321
4322 parsed_hstate = h;
4323 }
4324
hugetlb_node_alloc_supported(void)4325 bool __init __weak hugetlb_node_alloc_supported(void)
4326 {
4327 return true;
4328 }
4329
hugepages_clear_pages_in_node(void)4330 static void __init hugepages_clear_pages_in_node(void)
4331 {
4332 if (!hugetlb_max_hstate) {
4333 default_hstate_max_huge_pages = 0;
4334 memset(default_hugepages_in_node, 0,
4335 sizeof(default_hugepages_in_node));
4336 } else {
4337 parsed_hstate->max_huge_pages = 0;
4338 memset(parsed_hstate->max_huge_pages_node, 0,
4339 sizeof(parsed_hstate->max_huge_pages_node));
4340 }
4341 }
4342
4343 /*
4344 * hugepages command line processing
4345 * hugepages normally follows a valid hugepagsz or default_hugepagsz
4346 * specification. If not, ignore the hugepages value. hugepages can also
4347 * be the first huge page command line option in which case it implicitly
4348 * specifies the number of huge pages for the default size.
4349 */
hugepages_setup(char * s)4350 static int __init hugepages_setup(char *s)
4351 {
4352 unsigned long *mhp;
4353 static unsigned long *last_mhp;
4354 int node = NUMA_NO_NODE;
4355 int count;
4356 unsigned long tmp;
4357 char *p = s;
4358
4359 if (!parsed_valid_hugepagesz) {
4360 pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s);
4361 parsed_valid_hugepagesz = true;
4362 return 1;
4363 }
4364
4365 /*
4366 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter
4367 * yet, so this hugepages= parameter goes to the "default hstate".
4368 * Otherwise, it goes with the previously parsed hugepagesz or
4369 * default_hugepagesz.
4370 */
4371 else if (!hugetlb_max_hstate)
4372 mhp = &default_hstate_max_huge_pages;
4373 else
4374 mhp = &parsed_hstate->max_huge_pages;
4375
4376 if (mhp == last_mhp) {
4377 pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s);
4378 return 1;
4379 }
4380
4381 while (*p) {
4382 count = 0;
4383 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4384 goto invalid;
4385 /* Parameter is node format */
4386 if (p[count] == ':') {
4387 if (!hugetlb_node_alloc_supported()) {
4388 pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n");
4389 return 1;
4390 }
4391 if (tmp >= MAX_NUMNODES || !node_online(tmp))
4392 goto invalid;
4393 node = array_index_nospec(tmp, MAX_NUMNODES);
4394 p += count + 1;
4395 /* Parse hugepages */
4396 if (sscanf(p, "%lu%n", &tmp, &count) != 1)
4397 goto invalid;
4398 if (!hugetlb_max_hstate)
4399 default_hugepages_in_node[node] = tmp;
4400 else
4401 parsed_hstate->max_huge_pages_node[node] = tmp;
4402 *mhp += tmp;
4403 /* Go to parse next node*/
4404 if (p[count] == ',')
4405 p += count + 1;
4406 else
4407 break;
4408 } else {
4409 if (p != s)
4410 goto invalid;
4411 *mhp = tmp;
4412 break;
4413 }
4414 }
4415
4416 /*
4417 * Global state is always initialized later in hugetlb_init.
4418 * But we need to allocate gigantic hstates here early to still
4419 * use the bootmem allocator.
4420 */
4421 if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate))
4422 hugetlb_hstate_alloc_pages(parsed_hstate);
4423
4424 last_mhp = mhp;
4425
4426 return 1;
4427
4428 invalid:
4429 pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p);
4430 hugepages_clear_pages_in_node();
4431 return 1;
4432 }
4433 __setup("hugepages=", hugepages_setup);
4434
4435 /*
4436 * hugepagesz command line processing
4437 * A specific huge page size can only be specified once with hugepagesz.
4438 * hugepagesz is followed by hugepages on the command line. The global
4439 * variable 'parsed_valid_hugepagesz' is used to determine if prior
4440 * hugepagesz argument was valid.
4441 */
hugepagesz_setup(char * s)4442 static int __init hugepagesz_setup(char *s)
4443 {
4444 unsigned long size;
4445 struct hstate *h;
4446
4447 parsed_valid_hugepagesz = false;
4448 size = (unsigned long)memparse(s, NULL);
4449
4450 if (!arch_hugetlb_valid_size(size)) {
4451 pr_err("HugeTLB: unsupported hugepagesz=%s\n", s);
4452 return 1;
4453 }
4454
4455 h = size_to_hstate(size);
4456 if (h) {
4457 /*
4458 * hstate for this size already exists. This is normally
4459 * an error, but is allowed if the existing hstate is the
4460 * default hstate. More specifically, it is only allowed if
4461 * the number of huge pages for the default hstate was not
4462 * previously specified.
4463 */
4464 if (!parsed_default_hugepagesz || h != &default_hstate ||
4465 default_hstate.max_huge_pages) {
4466 pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s);
4467 return 1;
4468 }
4469
4470 /*
4471 * No need to call hugetlb_add_hstate() as hstate already
4472 * exists. But, do set parsed_hstate so that a following
4473 * hugepages= parameter will be applied to this hstate.
4474 */
4475 parsed_hstate = h;
4476 parsed_valid_hugepagesz = true;
4477 return 1;
4478 }
4479
4480 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4481 parsed_valid_hugepagesz = true;
4482 return 1;
4483 }
4484 __setup("hugepagesz=", hugepagesz_setup);
4485
4486 /*
4487 * default_hugepagesz command line input
4488 * Only one instance of default_hugepagesz allowed on command line.
4489 */
default_hugepagesz_setup(char * s)4490 static int __init default_hugepagesz_setup(char *s)
4491 {
4492 unsigned long size;
4493 int i;
4494
4495 parsed_valid_hugepagesz = false;
4496 if (parsed_default_hugepagesz) {
4497 pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s);
4498 return 1;
4499 }
4500
4501 size = (unsigned long)memparse(s, NULL);
4502
4503 if (!arch_hugetlb_valid_size(size)) {
4504 pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s);
4505 return 1;
4506 }
4507
4508 hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT);
4509 parsed_valid_hugepagesz = true;
4510 parsed_default_hugepagesz = true;
4511 default_hstate_idx = hstate_index(size_to_hstate(size));
4512
4513 /*
4514 * The number of default huge pages (for this size) could have been
4515 * specified as the first hugetlb parameter: hugepages=X. If so,
4516 * then default_hstate_max_huge_pages is set. If the default huge
4517 * page size is gigantic (> MAX_ORDER), then the pages must be
4518 * allocated here from bootmem allocator.
4519 */
4520 if (default_hstate_max_huge_pages) {
4521 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
4522 for_each_online_node(i)
4523 default_hstate.max_huge_pages_node[i] =
4524 default_hugepages_in_node[i];
4525 if (hstate_is_gigantic(&default_hstate))
4526 hugetlb_hstate_alloc_pages(&default_hstate);
4527 default_hstate_max_huge_pages = 0;
4528 }
4529
4530 return 1;
4531 }
4532 __setup("default_hugepagesz=", default_hugepagesz_setup);
4533
policy_mbind_nodemask(gfp_t gfp)4534 static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
4535 {
4536 #ifdef CONFIG_NUMA
4537 struct mempolicy *mpol = get_task_policy(current);
4538
4539 /*
4540 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
4541 * (from policy_nodemask) specifically for hugetlb case
4542 */
4543 if (mpol->mode == MPOL_BIND &&
4544 (apply_policy_zone(mpol, gfp_zone(gfp)) &&
4545 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
4546 return &mpol->nodes;
4547 #endif
4548 return NULL;
4549 }
4550
allowed_mems_nr(struct hstate * h)4551 static unsigned int allowed_mems_nr(struct hstate *h)
4552 {
4553 int node;
4554 unsigned int nr = 0;
4555 nodemask_t *mbind_nodemask;
4556 unsigned int *array = h->free_huge_pages_node;
4557 gfp_t gfp_mask = htlb_alloc_mask(h);
4558
4559 mbind_nodemask = policy_mbind_nodemask(gfp_mask);
4560 for_each_node_mask(node, cpuset_current_mems_allowed) {
4561 if (!mbind_nodemask || node_isset(node, *mbind_nodemask))
4562 nr += array[node];
4563 }
4564
4565 return nr;
4566 }
4567
4568 #ifdef CONFIG_SYSCTL
proc_hugetlb_doulongvec_minmax(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos,unsigned long * out)4569 static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write,
4570 void *buffer, size_t *length,
4571 loff_t *ppos, unsigned long *out)
4572 {
4573 struct ctl_table dup_table;
4574
4575 /*
4576 * In order to avoid races with __do_proc_doulongvec_minmax(), we
4577 * can duplicate the @table and alter the duplicate of it.
4578 */
4579 dup_table = *table;
4580 dup_table.data = out;
4581
4582 return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos);
4583 }
4584
hugetlb_sysctl_handler_common(bool obey_mempolicy,struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)4585 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
4586 struct ctl_table *table, int write,
4587 void *buffer, size_t *length, loff_t *ppos)
4588 {
4589 struct hstate *h = &default_hstate;
4590 unsigned long tmp = h->max_huge_pages;
4591 int ret;
4592
4593 if (!hugepages_supported())
4594 return -EOPNOTSUPP;
4595
4596 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4597 &tmp);
4598 if (ret)
4599 goto out;
4600
4601 if (write)
4602 ret = __nr_hugepages_store_common(obey_mempolicy, h,
4603 NUMA_NO_NODE, tmp, *length);
4604 out:
4605 return ret;
4606 }
4607
hugetlb_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)4608 static int hugetlb_sysctl_handler(struct ctl_table *table, int write,
4609 void *buffer, size_t *length, loff_t *ppos)
4610 {
4611
4612 return hugetlb_sysctl_handler_common(false, table, write,
4613 buffer, length, ppos);
4614 }
4615
4616 #ifdef CONFIG_NUMA
hugetlb_mempolicy_sysctl_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)4617 static int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
4618 void *buffer, size_t *length, loff_t *ppos)
4619 {
4620 return hugetlb_sysctl_handler_common(true, table, write,
4621 buffer, length, ppos);
4622 }
4623 #endif /* CONFIG_NUMA */
4624
hugetlb_overcommit_handler(struct ctl_table * table,int write,void * buffer,size_t * length,loff_t * ppos)4625 static int hugetlb_overcommit_handler(struct ctl_table *table, int write,
4626 void *buffer, size_t *length, loff_t *ppos)
4627 {
4628 struct hstate *h = &default_hstate;
4629 unsigned long tmp;
4630 int ret;
4631
4632 if (!hugepages_supported())
4633 return -EOPNOTSUPP;
4634
4635 tmp = h->nr_overcommit_huge_pages;
4636
4637 if (write && hstate_is_gigantic(h))
4638 return -EINVAL;
4639
4640 ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos,
4641 &tmp);
4642 if (ret)
4643 goto out;
4644
4645 if (write) {
4646 spin_lock_irq(&hugetlb_lock);
4647 h->nr_overcommit_huge_pages = tmp;
4648 spin_unlock_irq(&hugetlb_lock);
4649 }
4650 out:
4651 return ret;
4652 }
4653
4654 static struct ctl_table hugetlb_table[] = {
4655 {
4656 .procname = "nr_hugepages",
4657 .data = NULL,
4658 .maxlen = sizeof(unsigned long),
4659 .mode = 0644,
4660 .proc_handler = hugetlb_sysctl_handler,
4661 },
4662 #ifdef CONFIG_NUMA
4663 {
4664 .procname = "nr_hugepages_mempolicy",
4665 .data = NULL,
4666 .maxlen = sizeof(unsigned long),
4667 .mode = 0644,
4668 .proc_handler = &hugetlb_mempolicy_sysctl_handler,
4669 },
4670 #endif
4671 {
4672 .procname = "hugetlb_shm_group",
4673 .data = &sysctl_hugetlb_shm_group,
4674 .maxlen = sizeof(gid_t),
4675 .mode = 0644,
4676 .proc_handler = proc_dointvec,
4677 },
4678 {
4679 .procname = "nr_overcommit_hugepages",
4680 .data = NULL,
4681 .maxlen = sizeof(unsigned long),
4682 .mode = 0644,
4683 .proc_handler = hugetlb_overcommit_handler,
4684 },
4685 { }
4686 };
4687
hugetlb_sysctl_init(void)4688 static void hugetlb_sysctl_init(void)
4689 {
4690 register_sysctl_init("vm", hugetlb_table);
4691 }
4692 #endif /* CONFIG_SYSCTL */
4693
hugetlb_report_meminfo(struct seq_file * m)4694 void hugetlb_report_meminfo(struct seq_file *m)
4695 {
4696 struct hstate *h;
4697 unsigned long total = 0;
4698
4699 if (!hugepages_supported())
4700 return;
4701
4702 for_each_hstate(h) {
4703 unsigned long count = h->nr_huge_pages;
4704
4705 total += huge_page_size(h) * count;
4706
4707 if (h == &default_hstate)
4708 seq_printf(m,
4709 "HugePages_Total: %5lu\n"
4710 "HugePages_Free: %5lu\n"
4711 "HugePages_Rsvd: %5lu\n"
4712 "HugePages_Surp: %5lu\n"
4713 "Hugepagesize: %8lu kB\n",
4714 count,
4715 h->free_huge_pages,
4716 h->resv_huge_pages,
4717 h->surplus_huge_pages,
4718 huge_page_size(h) / SZ_1K);
4719 }
4720
4721 seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K);
4722 }
4723
hugetlb_report_node_meminfo(char * buf,int len,int nid)4724 int hugetlb_report_node_meminfo(char *buf, int len, int nid)
4725 {
4726 struct hstate *h = &default_hstate;
4727
4728 if (!hugepages_supported())
4729 return 0;
4730
4731 return sysfs_emit_at(buf, len,
4732 "Node %d HugePages_Total: %5u\n"
4733 "Node %d HugePages_Free: %5u\n"
4734 "Node %d HugePages_Surp: %5u\n",
4735 nid, h->nr_huge_pages_node[nid],
4736 nid, h->free_huge_pages_node[nid],
4737 nid, h->surplus_huge_pages_node[nid]);
4738 }
4739
hugetlb_show_meminfo_node(int nid)4740 void hugetlb_show_meminfo_node(int nid)
4741 {
4742 struct hstate *h;
4743
4744 if (!hugepages_supported())
4745 return;
4746
4747 for_each_hstate(h)
4748 printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
4749 nid,
4750 h->nr_huge_pages_node[nid],
4751 h->free_huge_pages_node[nid],
4752 h->surplus_huge_pages_node[nid],
4753 huge_page_size(h) / SZ_1K);
4754 }
4755
hugetlb_report_usage(struct seq_file * m,struct mm_struct * mm)4756 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
4757 {
4758 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
4759 K(atomic_long_read(&mm->hugetlb_usage)));
4760 }
4761
4762 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
hugetlb_total_pages(void)4763 unsigned long hugetlb_total_pages(void)
4764 {
4765 struct hstate *h;
4766 unsigned long nr_total_pages = 0;
4767
4768 for_each_hstate(h)
4769 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
4770 return nr_total_pages;
4771 }
4772
hugetlb_acct_memory(struct hstate * h,long delta)4773 static int hugetlb_acct_memory(struct hstate *h, long delta)
4774 {
4775 int ret = -ENOMEM;
4776
4777 if (!delta)
4778 return 0;
4779
4780 spin_lock_irq(&hugetlb_lock);
4781 /*
4782 * When cpuset is configured, it breaks the strict hugetlb page
4783 * reservation as the accounting is done on a global variable. Such
4784 * reservation is completely rubbish in the presence of cpuset because
4785 * the reservation is not checked against page availability for the
4786 * current cpuset. Application can still potentially OOM'ed by kernel
4787 * with lack of free htlb page in cpuset that the task is in.
4788 * Attempt to enforce strict accounting with cpuset is almost
4789 * impossible (or too ugly) because cpuset is too fluid that
4790 * task or memory node can be dynamically moved between cpusets.
4791 *
4792 * The change of semantics for shared hugetlb mapping with cpuset is
4793 * undesirable. However, in order to preserve some of the semantics,
4794 * we fall back to check against current free page availability as
4795 * a best attempt and hopefully to minimize the impact of changing
4796 * semantics that cpuset has.
4797 *
4798 * Apart from cpuset, we also have memory policy mechanism that
4799 * also determines from which node the kernel will allocate memory
4800 * in a NUMA system. So similar to cpuset, we also should consider
4801 * the memory policy of the current task. Similar to the description
4802 * above.
4803 */
4804 if (delta > 0) {
4805 if (gather_surplus_pages(h, delta) < 0)
4806 goto out;
4807
4808 if (delta > allowed_mems_nr(h)) {
4809 return_unused_surplus_pages(h, delta);
4810 goto out;
4811 }
4812 }
4813
4814 ret = 0;
4815 if (delta < 0)
4816 return_unused_surplus_pages(h, (unsigned long) -delta);
4817
4818 out:
4819 spin_unlock_irq(&hugetlb_lock);
4820 return ret;
4821 }
4822
hugetlb_vm_op_open(struct vm_area_struct * vma)4823 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
4824 {
4825 struct resv_map *resv = vma_resv_map(vma);
4826
4827 /*
4828 * HPAGE_RESV_OWNER indicates a private mapping.
4829 * This new VMA should share its siblings reservation map if present.
4830 * The VMA will only ever have a valid reservation map pointer where
4831 * it is being copied for another still existing VMA. As that VMA
4832 * has a reference to the reservation map it cannot disappear until
4833 * after this open call completes. It is therefore safe to take a
4834 * new reference here without additional locking.
4835 */
4836 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
4837 resv_map_dup_hugetlb_cgroup_uncharge_info(resv);
4838 kref_get(&resv->refs);
4839 }
4840
4841 /*
4842 * vma_lock structure for sharable mappings is vma specific.
4843 * Clear old pointer (if copied via vm_area_dup) and allocate
4844 * new structure. Before clearing, make sure vma_lock is not
4845 * for this vma.
4846 */
4847 if (vma->vm_flags & VM_MAYSHARE) {
4848 struct hugetlb_vma_lock *vma_lock = vma->vm_private_data;
4849
4850 if (vma_lock) {
4851 if (vma_lock->vma != vma) {
4852 vma->vm_private_data = NULL;
4853 hugetlb_vma_lock_alloc(vma);
4854 } else
4855 pr_warn("HugeTLB: vma_lock already exists in %s.\n", __func__);
4856 } else
4857 hugetlb_vma_lock_alloc(vma);
4858 }
4859 }
4860
hugetlb_vm_op_close(struct vm_area_struct * vma)4861 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
4862 {
4863 struct hstate *h = hstate_vma(vma);
4864 struct resv_map *resv;
4865 struct hugepage_subpool *spool = subpool_vma(vma);
4866 unsigned long reserve, start, end;
4867 long gbl_reserve;
4868
4869 hugetlb_vma_lock_free(vma);
4870
4871 resv = vma_resv_map(vma);
4872 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4873 return;
4874
4875 start = vma_hugecache_offset(h, vma, vma->vm_start);
4876 end = vma_hugecache_offset(h, vma, vma->vm_end);
4877
4878 reserve = (end - start) - region_count(resv, start, end);
4879 hugetlb_cgroup_uncharge_counter(resv, start, end);
4880 if (reserve) {
4881 /*
4882 * Decrement reserve counts. The global reserve count may be
4883 * adjusted if the subpool has a minimum size.
4884 */
4885 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
4886 hugetlb_acct_memory(h, -gbl_reserve);
4887 }
4888
4889 kref_put(&resv->refs, resv_map_release);
4890 }
4891
hugetlb_vm_op_split(struct vm_area_struct * vma,unsigned long addr)4892 static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
4893 {
4894 if (addr & ~(huge_page_mask(hstate_vma(vma))))
4895 return -EINVAL;
4896
4897 /*
4898 * PMD sharing is only possible for PUD_SIZE-aligned address ranges
4899 * in HugeTLB VMAs. If we will lose PUD_SIZE alignment due to this
4900 * split, unshare PMDs in the PUD_SIZE interval surrounding addr now.
4901 */
4902 if (addr & ~PUD_MASK) {
4903 /*
4904 * hugetlb_vm_op_split is called right before we attempt to
4905 * split the VMA. We will need to unshare PMDs in the old and
4906 * new VMAs, so let's unshare before we split.
4907 */
4908 unsigned long floor = addr & PUD_MASK;
4909 unsigned long ceil = floor + PUD_SIZE;
4910
4911 if (floor >= vma->vm_start && ceil <= vma->vm_end)
4912 hugetlb_unshare_pmds(vma, floor, ceil);
4913 }
4914
4915 return 0;
4916 }
4917
hugetlb_vm_op_pagesize(struct vm_area_struct * vma)4918 static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
4919 {
4920 return huge_page_size(hstate_vma(vma));
4921 }
4922
4923 /*
4924 * We cannot handle pagefaults against hugetlb pages at all. They cause
4925 * handle_mm_fault() to try to instantiate regular-sized pages in the
4926 * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get
4927 * this far.
4928 */
hugetlb_vm_op_fault(struct vm_fault * vmf)4929 static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
4930 {
4931 BUG();
4932 return 0;
4933 }
4934
4935 /*
4936 * When a new function is introduced to vm_operations_struct and added
4937 * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
4938 * This is because under System V memory model, mappings created via
4939 * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
4940 * their original vm_ops are overwritten with shm_vm_ops.
4941 */
4942 const struct vm_operations_struct hugetlb_vm_ops = {
4943 .fault = hugetlb_vm_op_fault,
4944 .open = hugetlb_vm_op_open,
4945 .close = hugetlb_vm_op_close,
4946 .may_split = hugetlb_vm_op_split,
4947 .pagesize = hugetlb_vm_op_pagesize,
4948 };
4949
make_huge_pte(struct vm_area_struct * vma,struct page * page,int writable)4950 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
4951 int writable)
4952 {
4953 pte_t entry;
4954 unsigned int shift = huge_page_shift(hstate_vma(vma));
4955
4956 if (writable) {
4957 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
4958 vma->vm_page_prot)));
4959 } else {
4960 entry = huge_pte_wrprotect(mk_huge_pte(page,
4961 vma->vm_page_prot));
4962 }
4963 entry = pte_mkyoung(entry);
4964 entry = arch_make_huge_pte(entry, shift, vma->vm_flags);
4965
4966 return entry;
4967 }
4968
set_huge_ptep_writable(struct vm_area_struct * vma,unsigned long address,pte_t * ptep)4969 static void set_huge_ptep_writable(struct vm_area_struct *vma,
4970 unsigned long address, pte_t *ptep)
4971 {
4972 pte_t entry;
4973
4974 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
4975 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
4976 update_mmu_cache(vma, address, ptep);
4977 }
4978
is_hugetlb_entry_migration(pte_t pte)4979 bool is_hugetlb_entry_migration(pte_t pte)
4980 {
4981 swp_entry_t swp;
4982
4983 if (huge_pte_none(pte) || pte_present(pte))
4984 return false;
4985 swp = pte_to_swp_entry(pte);
4986 if (is_migration_entry(swp))
4987 return true;
4988 else
4989 return false;
4990 }
4991
is_hugetlb_entry_hwpoisoned(pte_t pte)4992 static bool is_hugetlb_entry_hwpoisoned(pte_t pte)
4993 {
4994 swp_entry_t swp;
4995
4996 if (huge_pte_none(pte) || pte_present(pte))
4997 return false;
4998 swp = pte_to_swp_entry(pte);
4999 if (is_hwpoison_entry(swp))
5000 return true;
5001 else
5002 return false;
5003 }
5004
5005 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)5006 hugetlb_install_folio(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr,
5007 struct folio *new_folio, pte_t old, unsigned long sz)
5008 {
5009 pte_t newpte = make_huge_pte(vma, &new_folio->page, 1);
5010
5011 __folio_mark_uptodate(new_folio);
5012 hugepage_add_new_anon_rmap(new_folio, vma, addr);
5013 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(old))
5014 newpte = huge_pte_mkuffd_wp(newpte);
5015 set_huge_pte_at(vma->vm_mm, addr, ptep, newpte, sz);
5016 hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm);
5017 folio_set_hugetlb_migratable(new_folio);
5018 }
5019
copy_hugetlb_page_range(struct mm_struct * dst,struct mm_struct * src,struct vm_area_struct * dst_vma,struct vm_area_struct * src_vma)5020 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
5021 struct vm_area_struct *dst_vma,
5022 struct vm_area_struct *src_vma)
5023 {
5024 pte_t *src_pte, *dst_pte, entry;
5025 struct folio *pte_folio;
5026 unsigned long addr;
5027 bool cow = is_cow_mapping(src_vma->vm_flags);
5028 struct hstate *h = hstate_vma(src_vma);
5029 unsigned long sz = huge_page_size(h);
5030 unsigned long npages = pages_per_huge_page(h);
5031 struct mmu_notifier_range range;
5032 unsigned long last_addr_mask;
5033 int ret = 0;
5034
5035 if (cow) {
5036 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src,
5037 src_vma->vm_start,
5038 src_vma->vm_end);
5039 mmu_notifier_invalidate_range_start(&range);
5040 vma_assert_write_locked(src_vma);
5041 raw_write_seqcount_begin(&src->write_protect_seq);
5042 } else {
5043 /*
5044 * For shared mappings the vma lock must be held before
5045 * calling hugetlb_walk() in the src vma. Otherwise, the
5046 * returned ptep could go away if part of a shared pmd and
5047 * another thread calls huge_pmd_unshare.
5048 */
5049 hugetlb_vma_lock_read(src_vma);
5050 }
5051
5052 last_addr_mask = hugetlb_mask_last_page(h);
5053 for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) {
5054 spinlock_t *src_ptl, *dst_ptl;
5055 src_pte = hugetlb_walk(src_vma, addr, sz);
5056 if (!src_pte) {
5057 addr |= last_addr_mask;
5058 continue;
5059 }
5060 dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz);
5061 if (!dst_pte) {
5062 ret = -ENOMEM;
5063 break;
5064 }
5065
5066 /*
5067 * If the pagetables are shared don't copy or take references.
5068 *
5069 * dst_pte == src_pte is the common case of src/dest sharing.
5070 * However, src could have 'unshared' and dst shares with
5071 * another vma. So page_count of ptep page is checked instead
5072 * to reliably determine whether pte is shared.
5073 */
5074 if (page_count(virt_to_page(dst_pte)) > 1) {
5075 addr |= last_addr_mask;
5076 continue;
5077 }
5078
5079 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5080 src_ptl = huge_pte_lockptr(h, src, src_pte);
5081 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5082 entry = huge_ptep_get(src_pte);
5083 again:
5084 if (huge_pte_none(entry)) {
5085 /*
5086 * Skip if src entry none.
5087 */
5088 ;
5089 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) {
5090 if (!userfaultfd_wp(dst_vma))
5091 entry = huge_pte_clear_uffd_wp(entry);
5092 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5093 } else if (unlikely(is_hugetlb_entry_migration(entry))) {
5094 swp_entry_t swp_entry = pte_to_swp_entry(entry);
5095 bool uffd_wp = pte_swp_uffd_wp(entry);
5096
5097 if (!is_readable_migration_entry(swp_entry) && cow) {
5098 /*
5099 * COW mappings require pages in both
5100 * parent and child to be set to read.
5101 */
5102 swp_entry = make_readable_migration_entry(
5103 swp_offset(swp_entry));
5104 entry = swp_entry_to_pte(swp_entry);
5105 if (userfaultfd_wp(src_vma) && uffd_wp)
5106 entry = pte_swp_mkuffd_wp(entry);
5107 set_huge_pte_at(src, addr, src_pte, entry, sz);
5108 }
5109 if (!userfaultfd_wp(dst_vma))
5110 entry = huge_pte_clear_uffd_wp(entry);
5111 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5112 } else if (unlikely(is_pte_marker(entry))) {
5113 pte_marker marker = copy_pte_marker(
5114 pte_to_swp_entry(entry), dst_vma);
5115
5116 if (marker)
5117 set_huge_pte_at(dst, addr, dst_pte,
5118 make_pte_marker(marker), sz);
5119 } else {
5120 entry = huge_ptep_get(src_pte);
5121 pte_folio = page_folio(pte_page(entry));
5122 folio_get(pte_folio);
5123
5124 /*
5125 * Failing to duplicate the anon rmap is a rare case
5126 * where we see pinned hugetlb pages while they're
5127 * prone to COW. We need to do the COW earlier during
5128 * fork.
5129 *
5130 * When pre-allocating the page or copying data, we
5131 * need to be without the pgtable locks since we could
5132 * sleep during the process.
5133 */
5134 if (!folio_test_anon(pte_folio)) {
5135 page_dup_file_rmap(&pte_folio->page, true);
5136 } else if (page_try_dup_anon_rmap(&pte_folio->page,
5137 true, src_vma)) {
5138 pte_t src_pte_old = entry;
5139 struct folio *new_folio;
5140
5141 spin_unlock(src_ptl);
5142 spin_unlock(dst_ptl);
5143 /* Do not use reserve as it's private owned */
5144 new_folio = alloc_hugetlb_folio(dst_vma, addr, 1);
5145 if (IS_ERR(new_folio)) {
5146 folio_put(pte_folio);
5147 ret = PTR_ERR(new_folio);
5148 break;
5149 }
5150 ret = copy_user_large_folio(new_folio,
5151 pte_folio,
5152 addr, dst_vma);
5153 folio_put(pte_folio);
5154 if (ret) {
5155 folio_put(new_folio);
5156 break;
5157 }
5158
5159 /* Install the new hugetlb folio if src pte stable */
5160 dst_ptl = huge_pte_lock(h, dst, dst_pte);
5161 src_ptl = huge_pte_lockptr(h, src, src_pte);
5162 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5163 entry = huge_ptep_get(src_pte);
5164 if (!pte_same(src_pte_old, entry)) {
5165 restore_reserve_on_error(h, dst_vma, addr,
5166 new_folio);
5167 folio_put(new_folio);
5168 /* huge_ptep of dst_pte won't change as in child */
5169 goto again;
5170 }
5171 hugetlb_install_folio(dst_vma, dst_pte, addr,
5172 new_folio, src_pte_old, sz);
5173 spin_unlock(src_ptl);
5174 spin_unlock(dst_ptl);
5175 continue;
5176 }
5177
5178 if (cow) {
5179 /*
5180 * No need to notify as we are downgrading page
5181 * table protection not changing it to point
5182 * to a new page.
5183 *
5184 * See Documentation/mm/mmu_notifier.rst
5185 */
5186 huge_ptep_set_wrprotect(src, addr, src_pte);
5187 entry = huge_pte_wrprotect(entry);
5188 }
5189
5190 if (!userfaultfd_wp(dst_vma))
5191 entry = huge_pte_clear_uffd_wp(entry);
5192
5193 set_huge_pte_at(dst, addr, dst_pte, entry, sz);
5194 hugetlb_count_add(npages, dst);
5195 }
5196 spin_unlock(src_ptl);
5197 spin_unlock(dst_ptl);
5198 }
5199
5200 if (cow) {
5201 raw_write_seqcount_end(&src->write_protect_seq);
5202 mmu_notifier_invalidate_range_end(&range);
5203 } else {
5204 hugetlb_vma_unlock_read(src_vma);
5205 }
5206
5207 return ret;
5208 }
5209
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)5210 static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr,
5211 unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte,
5212 unsigned long sz)
5213 {
5214 struct hstate *h = hstate_vma(vma);
5215 struct mm_struct *mm = vma->vm_mm;
5216 spinlock_t *src_ptl, *dst_ptl;
5217 pte_t pte;
5218
5219 dst_ptl = huge_pte_lock(h, mm, dst_pte);
5220 src_ptl = huge_pte_lockptr(h, mm, src_pte);
5221
5222 /*
5223 * We don't have to worry about the ordering of src and dst ptlocks
5224 * because exclusive mmap_lock (or the i_mmap_lock) prevents deadlock.
5225 */
5226 if (src_ptl != dst_ptl)
5227 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
5228
5229 pte = huge_ptep_get_and_clear(mm, old_addr, src_pte);
5230 set_huge_pte_at(mm, new_addr, dst_pte, pte, sz);
5231
5232 if (src_ptl != dst_ptl)
5233 spin_unlock(src_ptl);
5234 spin_unlock(dst_ptl);
5235 }
5236
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)5237 int move_hugetlb_page_tables(struct vm_area_struct *vma,
5238 struct vm_area_struct *new_vma,
5239 unsigned long old_addr, unsigned long new_addr,
5240 unsigned long len)
5241 {
5242 struct hstate *h = hstate_vma(vma);
5243 struct address_space *mapping = vma->vm_file->f_mapping;
5244 unsigned long sz = huge_page_size(h);
5245 struct mm_struct *mm = vma->vm_mm;
5246 unsigned long old_end = old_addr + len;
5247 unsigned long last_addr_mask;
5248 pte_t *src_pte, *dst_pte;
5249 struct mmu_notifier_range range;
5250 bool shared_pmd = false;
5251
5252 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, old_addr,
5253 old_end);
5254 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5255 /*
5256 * In case of shared PMDs, we should cover the maximum possible
5257 * range.
5258 */
5259 flush_cache_range(vma, range.start, range.end);
5260
5261 mmu_notifier_invalidate_range_start(&range);
5262 last_addr_mask = hugetlb_mask_last_page(h);
5263 /* Prevent race with file truncation */
5264 hugetlb_vma_lock_write(vma);
5265 i_mmap_lock_write(mapping);
5266 for (; old_addr < old_end; old_addr += sz, new_addr += sz) {
5267 src_pte = hugetlb_walk(vma, old_addr, sz);
5268 if (!src_pte) {
5269 old_addr |= last_addr_mask;
5270 new_addr |= last_addr_mask;
5271 continue;
5272 }
5273 if (huge_pte_none(huge_ptep_get(src_pte)))
5274 continue;
5275
5276 if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) {
5277 shared_pmd = true;
5278 old_addr |= last_addr_mask;
5279 new_addr |= last_addr_mask;
5280 continue;
5281 }
5282
5283 dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz);
5284 if (!dst_pte)
5285 break;
5286
5287 move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte, sz);
5288 }
5289
5290 if (shared_pmd)
5291 flush_hugetlb_tlb_range(vma, range.start, range.end);
5292 else
5293 flush_hugetlb_tlb_range(vma, old_end - len, old_end);
5294 mmu_notifier_invalidate_range_end(&range);
5295 i_mmap_unlock_write(mapping);
5296 hugetlb_vma_unlock_write(vma);
5297
5298 return len + old_addr - old_end;
5299 }
5300
__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)5301 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
5302 unsigned long start, unsigned long end,
5303 struct page *ref_page, zap_flags_t zap_flags)
5304 {
5305 struct mm_struct *mm = vma->vm_mm;
5306 unsigned long address;
5307 pte_t *ptep;
5308 pte_t pte;
5309 spinlock_t *ptl;
5310 struct page *page;
5311 struct hstate *h = hstate_vma(vma);
5312 unsigned long sz = huge_page_size(h);
5313 unsigned long last_addr_mask;
5314 bool force_flush = false;
5315
5316 WARN_ON(!is_vm_hugetlb_page(vma));
5317 BUG_ON(start & ~huge_page_mask(h));
5318 BUG_ON(end & ~huge_page_mask(h));
5319
5320 /*
5321 * This is a hugetlb vma, all the pte entries should point
5322 * to huge page.
5323 */
5324 tlb_change_page_size(tlb, sz);
5325 tlb_start_vma(tlb, vma);
5326
5327 last_addr_mask = hugetlb_mask_last_page(h);
5328 address = start;
5329 for (; address < end; address += sz) {
5330 ptep = hugetlb_walk(vma, address, sz);
5331 if (!ptep) {
5332 address |= last_addr_mask;
5333 continue;
5334 }
5335
5336 ptl = huge_pte_lock(h, mm, ptep);
5337 if (huge_pmd_unshare(mm, vma, address, ptep)) {
5338 spin_unlock(ptl);
5339 tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE);
5340 force_flush = true;
5341 address |= last_addr_mask;
5342 continue;
5343 }
5344
5345 pte = huge_ptep_get(ptep);
5346 if (huge_pte_none(pte)) {
5347 spin_unlock(ptl);
5348 continue;
5349 }
5350
5351 /*
5352 * Migrating hugepage or HWPoisoned hugepage is already
5353 * unmapped and its refcount is dropped, so just clear pte here.
5354 */
5355 if (unlikely(!pte_present(pte))) {
5356 /*
5357 * If the pte was wr-protected by uffd-wp in any of the
5358 * swap forms, meanwhile the caller does not want to
5359 * drop the uffd-wp bit in this zap, then replace the
5360 * pte with a marker.
5361 */
5362 if (pte_swp_uffd_wp_any(pte) &&
5363 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5364 set_huge_pte_at(mm, address, ptep,
5365 make_pte_marker(PTE_MARKER_UFFD_WP),
5366 sz);
5367 else
5368 huge_pte_clear(mm, address, ptep, sz);
5369 spin_unlock(ptl);
5370 continue;
5371 }
5372
5373 page = pte_page(pte);
5374 /*
5375 * If a reference page is supplied, it is because a specific
5376 * page is being unmapped, not a range. Ensure the page we
5377 * are about to unmap is the actual page of interest.
5378 */
5379 if (ref_page) {
5380 if (page != ref_page) {
5381 spin_unlock(ptl);
5382 continue;
5383 }
5384 /*
5385 * Mark the VMA as having unmapped its page so that
5386 * future faults in this VMA will fail rather than
5387 * looking like data was lost
5388 */
5389 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
5390 }
5391
5392 pte = huge_ptep_get_and_clear(mm, address, ptep);
5393 tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
5394 if (huge_pte_dirty(pte))
5395 set_page_dirty(page);
5396 /* Leave a uffd-wp pte marker if needed */
5397 if (huge_pte_uffd_wp(pte) &&
5398 !(zap_flags & ZAP_FLAG_DROP_MARKER))
5399 set_huge_pte_at(mm, address, ptep,
5400 make_pte_marker(PTE_MARKER_UFFD_WP),
5401 sz);
5402 hugetlb_count_sub(pages_per_huge_page(h), mm);
5403 page_remove_rmap(page, vma, true);
5404
5405 spin_unlock(ptl);
5406 tlb_remove_page_size(tlb, page, huge_page_size(h));
5407 /*
5408 * Bail out after unmapping reference page if supplied
5409 */
5410 if (ref_page)
5411 break;
5412 }
5413 tlb_end_vma(tlb, vma);
5414
5415 /*
5416 * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We
5417 * could defer the flush until now, since by holding i_mmap_rwsem we
5418 * guaranteed that the last refernece would not be dropped. But we must
5419 * do the flushing before we return, as otherwise i_mmap_rwsem will be
5420 * dropped and the last reference to the shared PMDs page might be
5421 * dropped as well.
5422 *
5423 * In theory we could defer the freeing of the PMD pages as well, but
5424 * huge_pmd_unshare() relies on the exact page_count for the PMD page to
5425 * detect sharing, so we cannot defer the release of the page either.
5426 * Instead, do flush now.
5427 */
5428 if (force_flush)
5429 tlb_flush_mmu_tlbonly(tlb);
5430 }
5431
__hugetlb_zap_begin(struct vm_area_struct * vma,unsigned long * start,unsigned long * end)5432 void __hugetlb_zap_begin(struct vm_area_struct *vma,
5433 unsigned long *start, unsigned long *end)
5434 {
5435 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5436 return;
5437
5438 adjust_range_if_pmd_sharing_possible(vma, start, end);
5439 hugetlb_vma_lock_write(vma);
5440 if (vma->vm_file)
5441 i_mmap_lock_write(vma->vm_file->f_mapping);
5442 }
5443
__hugetlb_zap_end(struct vm_area_struct * vma,struct zap_details * details)5444 void __hugetlb_zap_end(struct vm_area_struct *vma,
5445 struct zap_details *details)
5446 {
5447 zap_flags_t zap_flags = details ? details->zap_flags : 0;
5448
5449 if (!vma->vm_file) /* hugetlbfs_file_mmap error */
5450 return;
5451
5452 if (zap_flags & ZAP_FLAG_UNMAP) { /* final unmap */
5453 /*
5454 * Unlock and free the vma lock before releasing i_mmap_rwsem.
5455 * When the vma_lock is freed, this makes the vma ineligible
5456 * for pmd sharing. And, i_mmap_rwsem is required to set up
5457 * pmd sharing. This is important as page tables for this
5458 * unmapped range will be asynchrously deleted. If the page
5459 * tables are shared, there will be issues when accessed by
5460 * someone else.
5461 */
5462 __hugetlb_vma_unlock_write_free(vma);
5463 } else {
5464 hugetlb_vma_unlock_write(vma);
5465 }
5466
5467 if (vma->vm_file)
5468 i_mmap_unlock_write(vma->vm_file->f_mapping);
5469 }
5470
unmap_hugepage_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,struct page * ref_page,zap_flags_t zap_flags)5471 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
5472 unsigned long end, struct page *ref_page,
5473 zap_flags_t zap_flags)
5474 {
5475 struct mmu_notifier_range range;
5476 struct mmu_gather tlb;
5477
5478 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
5479 start, end);
5480 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
5481 mmu_notifier_invalidate_range_start(&range);
5482 tlb_gather_mmu(&tlb, vma->vm_mm);
5483
5484 __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags);
5485
5486 mmu_notifier_invalidate_range_end(&range);
5487 tlb_finish_mmu(&tlb);
5488 }
5489
5490 /*
5491 * This is called when the original mapper is failing to COW a MAP_PRIVATE
5492 * mapping it owns the reserve page for. The intention is to unmap the page
5493 * from other VMAs and let the children be SIGKILLed if they are faulting the
5494 * same region.
5495 */
unmap_ref_private(struct mm_struct * mm,struct vm_area_struct * vma,struct page * page,unsigned long address)5496 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
5497 struct page *page, unsigned long address)
5498 {
5499 struct hstate *h = hstate_vma(vma);
5500 struct vm_area_struct *iter_vma;
5501 struct address_space *mapping;
5502 pgoff_t pgoff;
5503
5504 /*
5505 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
5506 * from page cache lookup which is in HPAGE_SIZE units.
5507 */
5508 address = address & huge_page_mask(h);
5509 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
5510 vma->vm_pgoff;
5511 mapping = vma->vm_file->f_mapping;
5512
5513 /*
5514 * Take the mapping lock for the duration of the table walk. As
5515 * this mapping should be shared between all the VMAs,
5516 * __unmap_hugepage_range() is called as the lock is already held
5517 */
5518 i_mmap_lock_write(mapping);
5519 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
5520 /* Do not unmap the current VMA */
5521 if (iter_vma == vma)
5522 continue;
5523
5524 /*
5525 * Shared VMAs have their own reserves and do not affect
5526 * MAP_PRIVATE accounting but it is possible that a shared
5527 * VMA is using the same page so check and skip such VMAs.
5528 */
5529 if (iter_vma->vm_flags & VM_MAYSHARE)
5530 continue;
5531
5532 /*
5533 * Unmap the page from other VMAs without their own reserves.
5534 * They get marked to be SIGKILLed if they fault in these
5535 * areas. This is because a future no-page fault on this VMA
5536 * could insert a zeroed page instead of the data existing
5537 * from the time of fork. This would look like data corruption
5538 */
5539 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
5540 unmap_hugepage_range(iter_vma, address,
5541 address + huge_page_size(h), page, 0);
5542 }
5543 i_mmap_unlock_write(mapping);
5544 }
5545
5546 /*
5547 * hugetlb_wp() should be called with page lock of the original hugepage held.
5548 * Called with hugetlb_fault_mutex_table held and pte_page locked so we
5549 * cannot race with other handlers or page migration.
5550 * Keep the pte_same checks anyway to make transition from the mutex easier.
5551 */
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)5552 static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma,
5553 unsigned long address, pte_t *ptep, unsigned int flags,
5554 struct folio *pagecache_folio, spinlock_t *ptl)
5555 {
5556 const bool unshare = flags & FAULT_FLAG_UNSHARE;
5557 pte_t pte = huge_ptep_get(ptep);
5558 struct hstate *h = hstate_vma(vma);
5559 struct folio *old_folio;
5560 struct folio *new_folio;
5561 int outside_reserve = 0;
5562 vm_fault_t ret = 0;
5563 unsigned long haddr = address & huge_page_mask(h);
5564 struct mmu_notifier_range range;
5565
5566 /*
5567 * Never handle CoW for uffd-wp protected pages. It should be only
5568 * handled when the uffd-wp protection is removed.
5569 *
5570 * Note that only the CoW optimization path (in hugetlb_no_page())
5571 * can trigger this, because hugetlb_fault() will always resolve
5572 * uffd-wp bit first.
5573 */
5574 if (!unshare && huge_pte_uffd_wp(pte))
5575 return 0;
5576
5577 /*
5578 * hugetlb does not support FOLL_FORCE-style write faults that keep the
5579 * PTE mapped R/O such as maybe_mkwrite() would do.
5580 */
5581 if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE)))
5582 return VM_FAULT_SIGSEGV;
5583
5584 /* Let's take out MAP_SHARED mappings first. */
5585 if (vma->vm_flags & VM_MAYSHARE) {
5586 set_huge_ptep_writable(vma, haddr, ptep);
5587 return 0;
5588 }
5589
5590 old_folio = page_folio(pte_page(pte));
5591
5592 delayacct_wpcopy_start();
5593
5594 retry_avoidcopy:
5595 /*
5596 * If no-one else is actually using this page, we're the exclusive
5597 * owner and can reuse this page.
5598 */
5599 if (folio_mapcount(old_folio) == 1 && folio_test_anon(old_folio)) {
5600 if (!PageAnonExclusive(&old_folio->page))
5601 page_move_anon_rmap(&old_folio->page, vma);
5602 if (likely(!unshare))
5603 set_huge_ptep_writable(vma, haddr, ptep);
5604
5605 delayacct_wpcopy_end();
5606 return 0;
5607 }
5608 VM_BUG_ON_PAGE(folio_test_anon(old_folio) &&
5609 PageAnonExclusive(&old_folio->page), &old_folio->page);
5610
5611 /*
5612 * If the process that created a MAP_PRIVATE mapping is about to
5613 * perform a COW due to a shared page count, attempt to satisfy
5614 * the allocation without using the existing reserves. The pagecache
5615 * page is used to determine if the reserve at this address was
5616 * consumed or not. If reserves were used, a partial faulted mapping
5617 * at the time of fork() could consume its reserves on COW instead
5618 * of the full address range.
5619 */
5620 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
5621 old_folio != pagecache_folio)
5622 outside_reserve = 1;
5623
5624 folio_get(old_folio);
5625
5626 /*
5627 * Drop page table lock as buddy allocator may be called. It will
5628 * be acquired again before returning to the caller, as expected.
5629 */
5630 spin_unlock(ptl);
5631 new_folio = alloc_hugetlb_folio(vma, haddr, outside_reserve);
5632
5633 if (IS_ERR(new_folio)) {
5634 /*
5635 * If a process owning a MAP_PRIVATE mapping fails to COW,
5636 * it is due to references held by a child and an insufficient
5637 * huge page pool. To guarantee the original mappers
5638 * reliability, unmap the page from child processes. The child
5639 * may get SIGKILLed if it later faults.
5640 */
5641 if (outside_reserve) {
5642 struct address_space *mapping = vma->vm_file->f_mapping;
5643 pgoff_t idx;
5644 u32 hash;
5645
5646 folio_put(old_folio);
5647 /*
5648 * Drop hugetlb_fault_mutex and vma_lock before
5649 * unmapping. unmapping needs to hold vma_lock
5650 * in write mode. Dropping vma_lock in read mode
5651 * here is OK as COW mappings do not interact with
5652 * PMD sharing.
5653 *
5654 * Reacquire both after unmap operation.
5655 */
5656 idx = vma_hugecache_offset(h, vma, haddr);
5657 hash = hugetlb_fault_mutex_hash(mapping, idx);
5658 hugetlb_vma_unlock_read(vma);
5659 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5660
5661 unmap_ref_private(mm, vma, &old_folio->page, haddr);
5662
5663 mutex_lock(&hugetlb_fault_mutex_table[hash]);
5664 hugetlb_vma_lock_read(vma);
5665 spin_lock(ptl);
5666 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5667 if (likely(ptep &&
5668 pte_same(huge_ptep_get(ptep), pte)))
5669 goto retry_avoidcopy;
5670 /*
5671 * race occurs while re-acquiring page table
5672 * lock, and our job is done.
5673 */
5674 delayacct_wpcopy_end();
5675 return 0;
5676 }
5677
5678 ret = vmf_error(PTR_ERR(new_folio));
5679 goto out_release_old;
5680 }
5681
5682 /*
5683 * When the original hugepage is shared one, it does not have
5684 * anon_vma prepared.
5685 */
5686 if (unlikely(anon_vma_prepare(vma))) {
5687 ret = VM_FAULT_OOM;
5688 goto out_release_all;
5689 }
5690
5691 if (copy_user_large_folio(new_folio, old_folio, address, vma)) {
5692 ret = VM_FAULT_HWPOISON_LARGE;
5693 goto out_release_all;
5694 }
5695 __folio_mark_uptodate(new_folio);
5696
5697 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, haddr,
5698 haddr + huge_page_size(h));
5699 mmu_notifier_invalidate_range_start(&range);
5700
5701 /*
5702 * Retake the page table lock to check for racing updates
5703 * before the page tables are altered
5704 */
5705 spin_lock(ptl);
5706 ptep = hugetlb_walk(vma, haddr, huge_page_size(h));
5707 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
5708 pte_t newpte = make_huge_pte(vma, &new_folio->page, !unshare);
5709
5710 /* Break COW or unshare */
5711 huge_ptep_clear_flush(vma, haddr, ptep);
5712 page_remove_rmap(&old_folio->page, vma, true);
5713 hugepage_add_new_anon_rmap(new_folio, vma, haddr);
5714 if (huge_pte_uffd_wp(pte))
5715 newpte = huge_pte_mkuffd_wp(newpte);
5716 set_huge_pte_at(mm, haddr, ptep, newpte, huge_page_size(h));
5717 folio_set_hugetlb_migratable(new_folio);
5718 /* Make the old page be freed below */
5719 new_folio = old_folio;
5720 }
5721 spin_unlock(ptl);
5722 mmu_notifier_invalidate_range_end(&range);
5723 out_release_all:
5724 /*
5725 * No restore in case of successful pagetable update (Break COW or
5726 * unshare)
5727 */
5728 if (new_folio != old_folio)
5729 restore_reserve_on_error(h, vma, haddr, new_folio);
5730 folio_put(new_folio);
5731 out_release_old:
5732 folio_put(old_folio);
5733
5734 spin_lock(ptl); /* Caller expects lock to be held */
5735
5736 delayacct_wpcopy_end();
5737 return ret;
5738 }
5739
5740 /*
5741 * Return whether there is a pagecache page to back given address within VMA.
5742 */
hugetlbfs_pagecache_present(struct hstate * h,struct vm_area_struct * vma,unsigned long address)5743 static bool hugetlbfs_pagecache_present(struct hstate *h,
5744 struct vm_area_struct *vma, unsigned long address)
5745 {
5746 struct address_space *mapping = vma->vm_file->f_mapping;
5747 pgoff_t idx = vma_hugecache_offset(h, vma, address);
5748 struct folio *folio;
5749
5750 folio = filemap_get_folio(mapping, idx);
5751 if (IS_ERR(folio))
5752 return false;
5753 folio_put(folio);
5754 return true;
5755 }
5756
hugetlb_add_to_page_cache(struct folio * folio,struct address_space * mapping,pgoff_t idx)5757 int hugetlb_add_to_page_cache(struct folio *folio, struct address_space *mapping,
5758 pgoff_t idx)
5759 {
5760 struct inode *inode = mapping->host;
5761 struct hstate *h = hstate_inode(inode);
5762 int err;
5763
5764 __folio_set_locked(folio);
5765 err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL);
5766
5767 if (unlikely(err)) {
5768 __folio_clear_locked(folio);
5769 return err;
5770 }
5771 folio_clear_hugetlb_restore_reserve(folio);
5772
5773 /*
5774 * mark folio dirty so that it will not be removed from cache/file
5775 * by non-hugetlbfs specific code paths.
5776 */
5777 folio_mark_dirty(folio);
5778
5779 spin_lock(&inode->i_lock);
5780 inode->i_blocks += blocks_per_huge_page(h);
5781 spin_unlock(&inode->i_lock);
5782 return 0;
5783 }
5784
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)5785 static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma,
5786 struct address_space *mapping,
5787 pgoff_t idx,
5788 unsigned int flags,
5789 unsigned long haddr,
5790 unsigned long addr,
5791 unsigned long reason)
5792 {
5793 u32 hash;
5794 struct vm_fault vmf = {
5795 .vma = vma,
5796 .address = haddr,
5797 .real_address = addr,
5798 .flags = flags,
5799
5800 /*
5801 * Hard to debug if it ends up being
5802 * used by a callee that assumes
5803 * something about the other
5804 * uninitialized fields... same as in
5805 * memory.c
5806 */
5807 };
5808
5809 /*
5810 * vma_lock and hugetlb_fault_mutex must be dropped before handling
5811 * userfault. Also mmap_lock could be dropped due to handling
5812 * userfault, any vma operation should be careful from here.
5813 */
5814 hugetlb_vma_unlock_read(vma);
5815 hash = hugetlb_fault_mutex_hash(mapping, idx);
5816 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
5817 return handle_userfault(&vmf, reason);
5818 }
5819
5820 /*
5821 * Recheck pte with pgtable lock. Returns true if pte didn't change, or
5822 * false if pte changed or is changing.
5823 */
hugetlb_pte_stable(struct hstate * h,struct mm_struct * mm,pte_t * ptep,pte_t old_pte)5824 static bool hugetlb_pte_stable(struct hstate *h, struct mm_struct *mm,
5825 pte_t *ptep, pte_t old_pte)
5826 {
5827 spinlock_t *ptl;
5828 bool same;
5829
5830 ptl = huge_pte_lock(h, mm, ptep);
5831 same = pte_same(huge_ptep_get(ptep), old_pte);
5832 spin_unlock(ptl);
5833
5834 return same;
5835 }
5836
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)5837 static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
5838 struct vm_area_struct *vma,
5839 struct address_space *mapping, pgoff_t idx,
5840 unsigned long address, pte_t *ptep,
5841 pte_t old_pte, unsigned int flags)
5842 {
5843 struct hstate *h = hstate_vma(vma);
5844 vm_fault_t ret = VM_FAULT_SIGBUS;
5845 int anon_rmap = 0;
5846 unsigned long size;
5847 struct folio *folio;
5848 pte_t new_pte;
5849 spinlock_t *ptl;
5850 unsigned long haddr = address & huge_page_mask(h);
5851 bool new_folio, new_pagecache_folio = false;
5852 u32 hash = hugetlb_fault_mutex_hash(mapping, idx);
5853
5854 /*
5855 * Currently, we are forced to kill the process in the event the
5856 * original mapper has unmapped pages from the child due to a failed
5857 * COW/unsharing. Warn that such a situation has occurred as it may not
5858 * be obvious.
5859 */
5860 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
5861 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
5862 current->pid);
5863 goto out;
5864 }
5865
5866 /*
5867 * Use page lock to guard against racing truncation
5868 * before we get page_table_lock.
5869 */
5870 new_folio = false;
5871 folio = filemap_lock_folio(mapping, idx);
5872 if (IS_ERR(folio)) {
5873 size = i_size_read(mapping->host) >> huge_page_shift(h);
5874 if (idx >= size)
5875 goto out;
5876 /* Check for page in userfault range */
5877 if (userfaultfd_missing(vma)) {
5878 /*
5879 * Since hugetlb_no_page() was examining pte
5880 * without pgtable lock, we need to re-test under
5881 * lock because the pte may not be stable and could
5882 * have changed from under us. Try to detect
5883 * either changed or during-changing ptes and retry
5884 * properly when needed.
5885 *
5886 * Note that userfaultfd is actually fine with
5887 * false positives (e.g. caused by pte changed),
5888 * but not wrong logical events (e.g. caused by
5889 * reading a pte during changing). The latter can
5890 * confuse the userspace, so the strictness is very
5891 * much preferred. E.g., MISSING event should
5892 * never happen on the page after UFFDIO_COPY has
5893 * correctly installed the page and returned.
5894 */
5895 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5896 ret = 0;
5897 goto out;
5898 }
5899
5900 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5901 haddr, address,
5902 VM_UFFD_MISSING);
5903 }
5904
5905 folio = alloc_hugetlb_folio(vma, haddr, 0);
5906 if (IS_ERR(folio)) {
5907 /*
5908 * Returning error will result in faulting task being
5909 * sent SIGBUS. The hugetlb fault mutex prevents two
5910 * tasks from racing to fault in the same page which
5911 * could result in false unable to allocate errors.
5912 * Page migration does not take the fault mutex, but
5913 * does a clear then write of pte's under page table
5914 * lock. Page fault code could race with migration,
5915 * notice the clear pte and try to allocate a page
5916 * here. Before returning error, get ptl and make
5917 * sure there really is no pte entry.
5918 */
5919 if (hugetlb_pte_stable(h, mm, ptep, old_pte))
5920 ret = vmf_error(PTR_ERR(folio));
5921 else
5922 ret = 0;
5923 goto out;
5924 }
5925 clear_huge_page(&folio->page, address, pages_per_huge_page(h));
5926 __folio_mark_uptodate(folio);
5927 new_folio = true;
5928
5929 if (vma->vm_flags & VM_MAYSHARE) {
5930 int err = hugetlb_add_to_page_cache(folio, mapping, idx);
5931 if (err) {
5932 /*
5933 * err can't be -EEXIST which implies someone
5934 * else consumed the reservation since hugetlb
5935 * fault mutex is held when add a hugetlb page
5936 * to the page cache. So it's safe to call
5937 * restore_reserve_on_error() here.
5938 */
5939 restore_reserve_on_error(h, vma, haddr, folio);
5940 folio_put(folio);
5941 goto out;
5942 }
5943 new_pagecache_folio = true;
5944 } else {
5945 folio_lock(folio);
5946 if (unlikely(anon_vma_prepare(vma))) {
5947 ret = VM_FAULT_OOM;
5948 goto backout_unlocked;
5949 }
5950 anon_rmap = 1;
5951 }
5952 } else {
5953 /*
5954 * If memory error occurs between mmap() and fault, some process
5955 * don't have hwpoisoned swap entry for errored virtual address.
5956 * So we need to block hugepage fault by PG_hwpoison bit check.
5957 */
5958 if (unlikely(folio_test_hwpoison(folio))) {
5959 ret = VM_FAULT_HWPOISON_LARGE |
5960 VM_FAULT_SET_HINDEX(hstate_index(h));
5961 goto backout_unlocked;
5962 }
5963
5964 /* Check for page in userfault range. */
5965 if (userfaultfd_minor(vma)) {
5966 folio_unlock(folio);
5967 folio_put(folio);
5968 /* See comment in userfaultfd_missing() block above */
5969 if (!hugetlb_pte_stable(h, mm, ptep, old_pte)) {
5970 ret = 0;
5971 goto out;
5972 }
5973 return hugetlb_handle_userfault(vma, mapping, idx, flags,
5974 haddr, address,
5975 VM_UFFD_MINOR);
5976 }
5977 }
5978
5979 /*
5980 * If we are going to COW a private mapping later, we examine the
5981 * pending reservations for this page now. This will ensure that
5982 * any allocations necessary to record that reservation occur outside
5983 * the spinlock.
5984 */
5985 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
5986 if (vma_needs_reservation(h, vma, haddr) < 0) {
5987 ret = VM_FAULT_OOM;
5988 goto backout_unlocked;
5989 }
5990 /* Just decrements count, does not deallocate */
5991 vma_end_reservation(h, vma, haddr);
5992 }
5993
5994 ptl = huge_pte_lock(h, mm, ptep);
5995 ret = 0;
5996 /* If pte changed from under us, retry */
5997 if (!pte_same(huge_ptep_get(ptep), old_pte))
5998 goto backout;
5999
6000 if (anon_rmap)
6001 hugepage_add_new_anon_rmap(folio, vma, haddr);
6002 else
6003 page_dup_file_rmap(&folio->page, true);
6004 new_pte = make_huge_pte(vma, &folio->page, ((vma->vm_flags & VM_WRITE)
6005 && (vma->vm_flags & VM_SHARED)));
6006 /*
6007 * If this pte was previously wr-protected, keep it wr-protected even
6008 * if populated.
6009 */
6010 if (unlikely(pte_marker_uffd_wp(old_pte)))
6011 new_pte = huge_pte_mkuffd_wp(new_pte);
6012 set_huge_pte_at(mm, haddr, ptep, new_pte, huge_page_size(h));
6013
6014 hugetlb_count_add(pages_per_huge_page(h), mm);
6015 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
6016 /* Optimization, do the COW without a second fault */
6017 ret = hugetlb_wp(mm, vma, address, ptep, flags, folio, ptl);
6018 }
6019
6020 spin_unlock(ptl);
6021
6022 /*
6023 * Only set hugetlb_migratable in newly allocated pages. Existing pages
6024 * found in the pagecache may not have hugetlb_migratable if they have
6025 * been isolated for migration.
6026 */
6027 if (new_folio)
6028 folio_set_hugetlb_migratable(folio);
6029
6030 folio_unlock(folio);
6031 out:
6032 hugetlb_vma_unlock_read(vma);
6033 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6034 return ret;
6035
6036 backout:
6037 spin_unlock(ptl);
6038 backout_unlocked:
6039 if (new_folio && !new_pagecache_folio)
6040 restore_reserve_on_error(h, vma, haddr, folio);
6041
6042 folio_unlock(folio);
6043 folio_put(folio);
6044 goto out;
6045 }
6046
6047 #ifdef CONFIG_SMP
hugetlb_fault_mutex_hash(struct address_space * mapping,pgoff_t idx)6048 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6049 {
6050 unsigned long key[2];
6051 u32 hash;
6052
6053 key[0] = (unsigned long) mapping;
6054 key[1] = idx;
6055
6056 hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
6057
6058 return hash & (num_fault_mutexes - 1);
6059 }
6060 #else
6061 /*
6062 * For uniprocessor systems we always use a single mutex, so just
6063 * return 0 and avoid the hashing overhead.
6064 */
hugetlb_fault_mutex_hash(struct address_space * mapping,pgoff_t idx)6065 u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
6066 {
6067 return 0;
6068 }
6069 #endif
6070
hugetlb_fault(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long address,unsigned int flags)6071 vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
6072 unsigned long address, unsigned int flags)
6073 {
6074 pte_t *ptep, entry;
6075 spinlock_t *ptl;
6076 vm_fault_t ret;
6077 u32 hash;
6078 pgoff_t idx;
6079 struct folio *folio = NULL;
6080 struct folio *pagecache_folio = NULL;
6081 struct hstate *h = hstate_vma(vma);
6082 struct address_space *mapping;
6083 int need_wait_lock = 0;
6084 unsigned long haddr = address & huge_page_mask(h);
6085
6086 /* TODO: Handle faults under the VMA lock */
6087 if (flags & FAULT_FLAG_VMA_LOCK) {
6088 vma_end_read(vma);
6089 return VM_FAULT_RETRY;
6090 }
6091
6092 /*
6093 * Serialize hugepage allocation and instantiation, so that we don't
6094 * get spurious allocation failures if two CPUs race to instantiate
6095 * the same page in the page cache.
6096 */
6097 mapping = vma->vm_file->f_mapping;
6098 idx = vma_hugecache_offset(h, vma, haddr);
6099 hash = hugetlb_fault_mutex_hash(mapping, idx);
6100 mutex_lock(&hugetlb_fault_mutex_table[hash]);
6101
6102 /*
6103 * Acquire vma lock before calling huge_pte_alloc and hold
6104 * until finished with ptep. This prevents huge_pmd_unshare from
6105 * being called elsewhere and making the ptep no longer valid.
6106 */
6107 hugetlb_vma_lock_read(vma);
6108 ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h));
6109 if (!ptep) {
6110 hugetlb_vma_unlock_read(vma);
6111 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6112 return VM_FAULT_OOM;
6113 }
6114
6115 entry = huge_ptep_get(ptep);
6116 if (huge_pte_none_mostly(entry)) {
6117 if (is_pte_marker(entry)) {
6118 pte_marker marker =
6119 pte_marker_get(pte_to_swp_entry(entry));
6120
6121 if (marker & PTE_MARKER_POISONED) {
6122 ret = VM_FAULT_HWPOISON_LARGE;
6123 goto out_mutex;
6124 }
6125 }
6126
6127 /*
6128 * Other PTE markers should be handled the same way as none PTE.
6129 *
6130 * hugetlb_no_page will drop vma lock and hugetlb fault
6131 * mutex internally, which make us return immediately.
6132 */
6133 return hugetlb_no_page(mm, vma, mapping, idx, address, ptep,
6134 entry, flags);
6135 }
6136
6137 ret = 0;
6138
6139 /*
6140 * entry could be a migration/hwpoison entry at this point, so this
6141 * check prevents the kernel from going below assuming that we have
6142 * an active hugepage in pagecache. This goto expects the 2nd page
6143 * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will
6144 * properly handle it.
6145 */
6146 if (!pte_present(entry)) {
6147 if (unlikely(is_hugetlb_entry_migration(entry))) {
6148 /*
6149 * Release the hugetlb fault lock now, but retain
6150 * the vma lock, because it is needed to guard the
6151 * huge_pte_lockptr() later in
6152 * migration_entry_wait_huge(). The vma lock will
6153 * be released there.
6154 */
6155 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6156 migration_entry_wait_huge(vma, ptep);
6157 return 0;
6158 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
6159 ret = VM_FAULT_HWPOISON_LARGE |
6160 VM_FAULT_SET_HINDEX(hstate_index(h));
6161 goto out_mutex;
6162 }
6163
6164 /*
6165 * If we are going to COW/unshare the mapping later, we examine the
6166 * pending reservations for this page now. This will ensure that any
6167 * allocations necessary to record that reservation occur outside the
6168 * spinlock. Also lookup the pagecache page now as it is used to
6169 * determine if a reservation has been consumed.
6170 */
6171 if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) &&
6172 !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) {
6173 if (vma_needs_reservation(h, vma, haddr) < 0) {
6174 ret = VM_FAULT_OOM;
6175 goto out_mutex;
6176 }
6177 /* Just decrements count, does not deallocate */
6178 vma_end_reservation(h, vma, haddr);
6179
6180 pagecache_folio = filemap_lock_folio(mapping, idx);
6181 if (IS_ERR(pagecache_folio))
6182 pagecache_folio = NULL;
6183 }
6184
6185 ptl = huge_pte_lock(h, mm, ptep);
6186
6187 /* Check for a racing update before calling hugetlb_wp() */
6188 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
6189 goto out_ptl;
6190
6191 /* Handle userfault-wp first, before trying to lock more pages */
6192 if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) &&
6193 (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
6194 struct vm_fault vmf = {
6195 .vma = vma,
6196 .address = haddr,
6197 .real_address = address,
6198 .flags = flags,
6199 };
6200
6201 spin_unlock(ptl);
6202 if (pagecache_folio) {
6203 folio_unlock(pagecache_folio);
6204 folio_put(pagecache_folio);
6205 }
6206 hugetlb_vma_unlock_read(vma);
6207 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6208 return handle_userfault(&vmf, VM_UFFD_WP);
6209 }
6210
6211 /*
6212 * hugetlb_wp() requires page locks of pte_page(entry) and
6213 * pagecache_folio, so here we need take the former one
6214 * when folio != pagecache_folio or !pagecache_folio.
6215 */
6216 folio = page_folio(pte_page(entry));
6217 if (folio != pagecache_folio)
6218 if (!folio_trylock(folio)) {
6219 need_wait_lock = 1;
6220 goto out_ptl;
6221 }
6222
6223 folio_get(folio);
6224
6225 if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) {
6226 if (!huge_pte_write(entry)) {
6227 ret = hugetlb_wp(mm, vma, address, ptep, flags,
6228 pagecache_folio, ptl);
6229 goto out_put_page;
6230 } else if (likely(flags & FAULT_FLAG_WRITE)) {
6231 entry = huge_pte_mkdirty(entry);
6232 }
6233 }
6234 entry = pte_mkyoung(entry);
6235 if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
6236 flags & FAULT_FLAG_WRITE))
6237 update_mmu_cache(vma, haddr, ptep);
6238 out_put_page:
6239 if (folio != pagecache_folio)
6240 folio_unlock(folio);
6241 folio_put(folio);
6242 out_ptl:
6243 spin_unlock(ptl);
6244
6245 if (pagecache_folio) {
6246 folio_unlock(pagecache_folio);
6247 folio_put(pagecache_folio);
6248 }
6249 out_mutex:
6250 hugetlb_vma_unlock_read(vma);
6251 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
6252 /*
6253 * Generally it's safe to hold refcount during waiting page lock. But
6254 * here we just wait to defer the next page fault to avoid busy loop and
6255 * the page is not used after unlocked before returning from the current
6256 * page fault. So we are safe from accessing freed page, even if we wait
6257 * here without taking refcount.
6258 */
6259 if (need_wait_lock)
6260 folio_wait_locked(folio);
6261 return ret;
6262 }
6263
6264 #ifdef CONFIG_USERFAULTFD
6265 /*
6266 * Used by userfaultfd UFFDIO_* ioctls. Based on userfaultfd's mfill_atomic_pte
6267 * with modifications for hugetlb pages.
6268 */
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)6269 int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
6270 struct vm_area_struct *dst_vma,
6271 unsigned long dst_addr,
6272 unsigned long src_addr,
6273 uffd_flags_t flags,
6274 struct folio **foliop)
6275 {
6276 struct mm_struct *dst_mm = dst_vma->vm_mm;
6277 bool is_continue = uffd_flags_mode_is(flags, MFILL_ATOMIC_CONTINUE);
6278 bool wp_enabled = (flags & MFILL_ATOMIC_WP);
6279 struct hstate *h = hstate_vma(dst_vma);
6280 struct address_space *mapping = dst_vma->vm_file->f_mapping;
6281 pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr);
6282 unsigned long size;
6283 int vm_shared = dst_vma->vm_flags & VM_SHARED;
6284 pte_t _dst_pte;
6285 spinlock_t *ptl;
6286 int ret = -ENOMEM;
6287 struct folio *folio;
6288 int writable;
6289 bool folio_in_pagecache = false;
6290
6291 if (uffd_flags_mode_is(flags, MFILL_ATOMIC_POISON)) {
6292 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6293
6294 /* Don't overwrite any existing PTEs (even markers) */
6295 if (!huge_pte_none(huge_ptep_get(dst_pte))) {
6296 spin_unlock(ptl);
6297 return -EEXIST;
6298 }
6299
6300 _dst_pte = make_pte_marker(PTE_MARKER_POISONED);
6301 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte,
6302 huge_page_size(h));
6303
6304 /* No need to invalidate - it was non-present before */
6305 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6306
6307 spin_unlock(ptl);
6308 return 0;
6309 }
6310
6311 if (is_continue) {
6312 ret = -EFAULT;
6313 folio = filemap_lock_folio(mapping, idx);
6314 if (IS_ERR(folio))
6315 goto out;
6316 folio_in_pagecache = true;
6317 } else if (!*foliop) {
6318 /* If a folio already exists, then it's UFFDIO_COPY for
6319 * a non-missing case. Return -EEXIST.
6320 */
6321 if (vm_shared &&
6322 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6323 ret = -EEXIST;
6324 goto out;
6325 }
6326
6327 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6328 if (IS_ERR(folio)) {
6329 ret = -ENOMEM;
6330 goto out;
6331 }
6332
6333 ret = copy_folio_from_user(folio, (const void __user *) src_addr,
6334 false);
6335
6336 /* fallback to copy_from_user outside mmap_lock */
6337 if (unlikely(ret)) {
6338 ret = -ENOENT;
6339 /* Free the allocated folio which may have
6340 * consumed a reservation.
6341 */
6342 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6343 folio_put(folio);
6344
6345 /* Allocate a temporary folio to hold the copied
6346 * contents.
6347 */
6348 folio = alloc_hugetlb_folio_vma(h, dst_vma, dst_addr);
6349 if (!folio) {
6350 ret = -ENOMEM;
6351 goto out;
6352 }
6353 *foliop = folio;
6354 /* Set the outparam foliop and return to the caller to
6355 * copy the contents outside the lock. Don't free the
6356 * folio.
6357 */
6358 goto out;
6359 }
6360 } else {
6361 if (vm_shared &&
6362 hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) {
6363 folio_put(*foliop);
6364 ret = -EEXIST;
6365 *foliop = NULL;
6366 goto out;
6367 }
6368
6369 folio = alloc_hugetlb_folio(dst_vma, dst_addr, 0);
6370 if (IS_ERR(folio)) {
6371 folio_put(*foliop);
6372 ret = -ENOMEM;
6373 *foliop = NULL;
6374 goto out;
6375 }
6376 ret = copy_user_large_folio(folio, *foliop, dst_addr, dst_vma);
6377 folio_put(*foliop);
6378 *foliop = NULL;
6379 if (ret) {
6380 folio_put(folio);
6381 goto out;
6382 }
6383 }
6384
6385 /*
6386 * The memory barrier inside __folio_mark_uptodate makes sure that
6387 * preceding stores to the page contents become visible before
6388 * the set_pte_at() write.
6389 */
6390 __folio_mark_uptodate(folio);
6391
6392 /* Add shared, newly allocated pages to the page cache. */
6393 if (vm_shared && !is_continue) {
6394 size = i_size_read(mapping->host) >> huge_page_shift(h);
6395 ret = -EFAULT;
6396 if (idx >= size)
6397 goto out_release_nounlock;
6398
6399 /*
6400 * Serialization between remove_inode_hugepages() and
6401 * hugetlb_add_to_page_cache() below happens through the
6402 * hugetlb_fault_mutex_table that here must be hold by
6403 * the caller.
6404 */
6405 ret = hugetlb_add_to_page_cache(folio, mapping, idx);
6406 if (ret)
6407 goto out_release_nounlock;
6408 folio_in_pagecache = true;
6409 }
6410
6411 ptl = huge_pte_lock(h, dst_mm, dst_pte);
6412
6413 ret = -EIO;
6414 if (folio_test_hwpoison(folio))
6415 goto out_release_unlock;
6416
6417 /*
6418 * We allow to overwrite a pte marker: consider when both MISSING|WP
6419 * registered, we firstly wr-protect a none pte which has no page cache
6420 * page backing it, then access the page.
6421 */
6422 ret = -EEXIST;
6423 if (!huge_pte_none_mostly(huge_ptep_get(dst_pte)))
6424 goto out_release_unlock;
6425
6426 if (folio_in_pagecache)
6427 page_dup_file_rmap(&folio->page, true);
6428 else
6429 hugepage_add_new_anon_rmap(folio, dst_vma, dst_addr);
6430
6431 /*
6432 * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY
6433 * with wp flag set, don't set pte write bit.
6434 */
6435 if (wp_enabled || (is_continue && !vm_shared))
6436 writable = 0;
6437 else
6438 writable = dst_vma->vm_flags & VM_WRITE;
6439
6440 _dst_pte = make_huge_pte(dst_vma, &folio->page, writable);
6441 /*
6442 * Always mark UFFDIO_COPY page dirty; note that this may not be
6443 * extremely important for hugetlbfs for now since swapping is not
6444 * supported, but we should still be clear in that this page cannot be
6445 * thrown away at will, even if write bit not set.
6446 */
6447 _dst_pte = huge_pte_mkdirty(_dst_pte);
6448 _dst_pte = pte_mkyoung(_dst_pte);
6449
6450 if (wp_enabled)
6451 _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
6452
6453 set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte, huge_page_size(h));
6454
6455 hugetlb_count_add(pages_per_huge_page(h), dst_mm);
6456
6457 /* No need to invalidate - it was non-present before */
6458 update_mmu_cache(dst_vma, dst_addr, dst_pte);
6459
6460 spin_unlock(ptl);
6461 if (!is_continue)
6462 folio_set_hugetlb_migratable(folio);
6463 if (vm_shared || is_continue)
6464 folio_unlock(folio);
6465 ret = 0;
6466 out:
6467 return ret;
6468 out_release_unlock:
6469 spin_unlock(ptl);
6470 if (vm_shared || is_continue)
6471 folio_unlock(folio);
6472 out_release_nounlock:
6473 if (!folio_in_pagecache)
6474 restore_reserve_on_error(h, dst_vma, dst_addr, folio);
6475 folio_put(folio);
6476 goto out;
6477 }
6478 #endif /* CONFIG_USERFAULTFD */
6479
hugetlb_follow_page_mask(struct vm_area_struct * vma,unsigned long address,unsigned int flags,unsigned int * page_mask)6480 struct page *hugetlb_follow_page_mask(struct vm_area_struct *vma,
6481 unsigned long address, unsigned int flags,
6482 unsigned int *page_mask)
6483 {
6484 struct hstate *h = hstate_vma(vma);
6485 struct mm_struct *mm = vma->vm_mm;
6486 unsigned long haddr = address & huge_page_mask(h);
6487 struct page *page = NULL;
6488 spinlock_t *ptl;
6489 pte_t *pte, entry;
6490 int ret;
6491
6492 hugetlb_vma_lock_read(vma);
6493 pte = hugetlb_walk(vma, haddr, huge_page_size(h));
6494 if (!pte)
6495 goto out_unlock;
6496
6497 ptl = huge_pte_lock(h, mm, pte);
6498 entry = huge_ptep_get(pte);
6499 if (pte_present(entry)) {
6500 page = pte_page(entry);
6501
6502 if (!huge_pte_write(entry)) {
6503 if (flags & FOLL_WRITE) {
6504 page = NULL;
6505 goto out;
6506 }
6507
6508 if (gup_must_unshare(vma, flags, page)) {
6509 /* Tell the caller to do unsharing */
6510 page = ERR_PTR(-EMLINK);
6511 goto out;
6512 }
6513 }
6514
6515 page = nth_page(page, ((address & ~huge_page_mask(h)) >> PAGE_SHIFT));
6516
6517 /*
6518 * Note that page may be a sub-page, and with vmemmap
6519 * optimizations the page struct may be read only.
6520 * try_grab_page() will increase the ref count on the
6521 * head page, so this will be OK.
6522 *
6523 * try_grab_page() should always be able to get the page here,
6524 * because we hold the ptl lock and have verified pte_present().
6525 */
6526 ret = try_grab_page(page, flags);
6527
6528 if (WARN_ON_ONCE(ret)) {
6529 page = ERR_PTR(ret);
6530 goto out;
6531 }
6532
6533 *page_mask = (1U << huge_page_order(h)) - 1;
6534 }
6535 out:
6536 spin_unlock(ptl);
6537 out_unlock:
6538 hugetlb_vma_unlock_read(vma);
6539
6540 /*
6541 * Fixup retval for dump requests: if pagecache doesn't exist,
6542 * don't try to allocate a new page but just skip it.
6543 */
6544 if (!page && (flags & FOLL_DUMP) &&
6545 !hugetlbfs_pagecache_present(h, vma, address))
6546 page = ERR_PTR(-EFAULT);
6547
6548 return page;
6549 }
6550
hugetlb_change_protection(struct vm_area_struct * vma,unsigned long address,unsigned long end,pgprot_t newprot,unsigned long cp_flags)6551 long hugetlb_change_protection(struct vm_area_struct *vma,
6552 unsigned long address, unsigned long end,
6553 pgprot_t newprot, unsigned long cp_flags)
6554 {
6555 struct mm_struct *mm = vma->vm_mm;
6556 unsigned long start = address;
6557 pte_t *ptep;
6558 pte_t pte;
6559 struct hstate *h = hstate_vma(vma);
6560 long pages = 0, psize = huge_page_size(h);
6561 bool shared_pmd = false;
6562 struct mmu_notifier_range range;
6563 unsigned long last_addr_mask;
6564 bool uffd_wp = cp_flags & MM_CP_UFFD_WP;
6565 bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE;
6566
6567 /*
6568 * In the case of shared PMDs, the area to flush could be beyond
6569 * start/end. Set range.start/range.end to cover the maximum possible
6570 * range if PMD sharing is possible.
6571 */
6572 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
6573 0, mm, start, end);
6574 adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
6575
6576 BUG_ON(address >= end);
6577 flush_cache_range(vma, range.start, range.end);
6578
6579 mmu_notifier_invalidate_range_start(&range);
6580 hugetlb_vma_lock_write(vma);
6581 i_mmap_lock_write(vma->vm_file->f_mapping);
6582 last_addr_mask = hugetlb_mask_last_page(h);
6583 for (; address < end; address += psize) {
6584 spinlock_t *ptl;
6585 ptep = hugetlb_walk(vma, address, psize);
6586 if (!ptep) {
6587 if (!uffd_wp) {
6588 address |= last_addr_mask;
6589 continue;
6590 }
6591 /*
6592 * Userfaultfd wr-protect requires pgtable
6593 * pre-allocations to install pte markers.
6594 */
6595 ptep = huge_pte_alloc(mm, vma, address, psize);
6596 if (!ptep) {
6597 pages = -ENOMEM;
6598 break;
6599 }
6600 }
6601 ptl = huge_pte_lock(h, mm, ptep);
6602 if (huge_pmd_unshare(mm, vma, address, ptep)) {
6603 /*
6604 * When uffd-wp is enabled on the vma, unshare
6605 * shouldn't happen at all. Warn about it if it
6606 * happened due to some reason.
6607 */
6608 WARN_ON_ONCE(uffd_wp || uffd_wp_resolve);
6609 pages++;
6610 spin_unlock(ptl);
6611 shared_pmd = true;
6612 address |= last_addr_mask;
6613 continue;
6614 }
6615 pte = huge_ptep_get(ptep);
6616 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
6617 /* Nothing to do. */
6618 } else if (unlikely(is_hugetlb_entry_migration(pte))) {
6619 swp_entry_t entry = pte_to_swp_entry(pte);
6620 struct page *page = pfn_swap_entry_to_page(entry);
6621 pte_t newpte = pte;
6622
6623 if (is_writable_migration_entry(entry)) {
6624 if (PageAnon(page))
6625 entry = make_readable_exclusive_migration_entry(
6626 swp_offset(entry));
6627 else
6628 entry = make_readable_migration_entry(
6629 swp_offset(entry));
6630 newpte = swp_entry_to_pte(entry);
6631 pages++;
6632 }
6633
6634 if (uffd_wp)
6635 newpte = pte_swp_mkuffd_wp(newpte);
6636 else if (uffd_wp_resolve)
6637 newpte = pte_swp_clear_uffd_wp(newpte);
6638 if (!pte_same(pte, newpte))
6639 set_huge_pte_at(mm, address, ptep, newpte, psize);
6640 } else if (unlikely(is_pte_marker(pte))) {
6641 /*
6642 * Do nothing on a poison marker; page is
6643 * corrupted, permissons do not apply. Here
6644 * pte_marker_uffd_wp()==true implies !poison
6645 * because they're mutual exclusive.
6646 */
6647 if (pte_marker_uffd_wp(pte) && uffd_wp_resolve)
6648 /* Safe to modify directly (non-present->none). */
6649 huge_pte_clear(mm, address, ptep, psize);
6650 } else if (!huge_pte_none(pte)) {
6651 pte_t old_pte;
6652 unsigned int shift = huge_page_shift(hstate_vma(vma));
6653
6654 old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
6655 pte = huge_pte_modify(old_pte, newprot);
6656 pte = arch_make_huge_pte(pte, shift, vma->vm_flags);
6657 if (uffd_wp)
6658 pte = huge_pte_mkuffd_wp(pte);
6659 else if (uffd_wp_resolve)
6660 pte = huge_pte_clear_uffd_wp(pte);
6661 huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
6662 pages++;
6663 } else {
6664 /* None pte */
6665 if (unlikely(uffd_wp))
6666 /* Safe to modify directly (none->non-present). */
6667 set_huge_pte_at(mm, address, ptep,
6668 make_pte_marker(PTE_MARKER_UFFD_WP),
6669 psize);
6670 }
6671 spin_unlock(ptl);
6672 }
6673 /*
6674 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
6675 * may have cleared our pud entry and done put_page on the page table:
6676 * once we release i_mmap_rwsem, another task can do the final put_page
6677 * and that page table be reused and filled with junk. If we actually
6678 * did unshare a page of pmds, flush the range corresponding to the pud.
6679 */
6680 if (shared_pmd)
6681 flush_hugetlb_tlb_range(vma, range.start, range.end);
6682 else
6683 flush_hugetlb_tlb_range(vma, start, end);
6684 /*
6685 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs() we are
6686 * downgrading page table protection not changing it to point to a new
6687 * page.
6688 *
6689 * See Documentation/mm/mmu_notifier.rst
6690 */
6691 i_mmap_unlock_write(vma->vm_file->f_mapping);
6692 hugetlb_vma_unlock_write(vma);
6693 mmu_notifier_invalidate_range_end(&range);
6694
6695 return pages > 0 ? (pages << h->order) : pages;
6696 }
6697
6698 /* 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)6699 bool hugetlb_reserve_pages(struct inode *inode,
6700 long from, long to,
6701 struct vm_area_struct *vma,
6702 vm_flags_t vm_flags)
6703 {
6704 long chg = -1, add = -1;
6705 struct hstate *h = hstate_inode(inode);
6706 struct hugepage_subpool *spool = subpool_inode(inode);
6707 struct resv_map *resv_map;
6708 struct hugetlb_cgroup *h_cg = NULL;
6709 long gbl_reserve, regions_needed = 0;
6710
6711 /* This should never happen */
6712 if (from > to) {
6713 VM_WARN(1, "%s called with a negative range\n", __func__);
6714 return false;
6715 }
6716
6717 /*
6718 * vma specific semaphore used for pmd sharing and fault/truncation
6719 * synchronization
6720 */
6721 hugetlb_vma_lock_alloc(vma);
6722
6723 /*
6724 * Only apply hugepage reservation if asked. At fault time, an
6725 * attempt will be made for VM_NORESERVE to allocate a page
6726 * without using reserves
6727 */
6728 if (vm_flags & VM_NORESERVE)
6729 return true;
6730
6731 /*
6732 * Shared mappings base their reservation on the number of pages that
6733 * are already allocated on behalf of the file. Private mappings need
6734 * to reserve the full area even if read-only as mprotect() may be
6735 * called to make the mapping read-write. Assume !vma is a shm mapping
6736 */
6737 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6738 /*
6739 * resv_map can not be NULL as hugetlb_reserve_pages is only
6740 * called for inodes for which resv_maps were created (see
6741 * hugetlbfs_get_inode).
6742 */
6743 resv_map = inode_resv_map(inode);
6744
6745 chg = region_chg(resv_map, from, to, ®ions_needed);
6746 } else {
6747 /* Private mapping. */
6748 resv_map = resv_map_alloc();
6749 if (!resv_map)
6750 goto out_err;
6751
6752 chg = to - from;
6753
6754 set_vma_resv_map(vma, resv_map);
6755 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
6756 }
6757
6758 if (chg < 0)
6759 goto out_err;
6760
6761 if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h),
6762 chg * pages_per_huge_page(h), &h_cg) < 0)
6763 goto out_err;
6764
6765 if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) {
6766 /* For private mappings, the hugetlb_cgroup uncharge info hangs
6767 * of the resv_map.
6768 */
6769 resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h);
6770 }
6771
6772 /*
6773 * There must be enough pages in the subpool for the mapping. If
6774 * the subpool has a minimum size, there may be some global
6775 * reservations already in place (gbl_reserve).
6776 */
6777 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
6778 if (gbl_reserve < 0)
6779 goto out_uncharge_cgroup;
6780
6781 /*
6782 * Check enough hugepages are available for the reservation.
6783 * Hand the pages back to the subpool if there are not
6784 */
6785 if (hugetlb_acct_memory(h, gbl_reserve) < 0)
6786 goto out_put_pages;
6787
6788 /*
6789 * Account for the reservations made. Shared mappings record regions
6790 * that have reservations as they are shared by multiple VMAs.
6791 * When the last VMA disappears, the region map says how much
6792 * the reservation was and the page cache tells how much of
6793 * the reservation was consumed. Private mappings are per-VMA and
6794 * only the consumed reservations are tracked. When the VMA
6795 * disappears, the original reservation is the VMA size and the
6796 * consumed reservations are stored in the map. Hence, nothing
6797 * else has to be done for private mappings here
6798 */
6799 if (!vma || vma->vm_flags & VM_MAYSHARE) {
6800 add = region_add(resv_map, from, to, regions_needed, h, h_cg);
6801
6802 if (unlikely(add < 0)) {
6803 hugetlb_acct_memory(h, -gbl_reserve);
6804 goto out_put_pages;
6805 } else if (unlikely(chg > add)) {
6806 /*
6807 * pages in this range were added to the reserve
6808 * map between region_chg and region_add. This
6809 * indicates a race with alloc_hugetlb_folio. Adjust
6810 * the subpool and reserve counts modified above
6811 * based on the difference.
6812 */
6813 long rsv_adjust;
6814
6815 /*
6816 * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the
6817 * reference to h_cg->css. See comment below for detail.
6818 */
6819 hugetlb_cgroup_uncharge_cgroup_rsvd(
6820 hstate_index(h),
6821 (chg - add) * pages_per_huge_page(h), h_cg);
6822
6823 rsv_adjust = hugepage_subpool_put_pages(spool,
6824 chg - add);
6825 hugetlb_acct_memory(h, -rsv_adjust);
6826 } else if (h_cg) {
6827 /*
6828 * The file_regions will hold their own reference to
6829 * h_cg->css. So we should release the reference held
6830 * via hugetlb_cgroup_charge_cgroup_rsvd() when we are
6831 * done.
6832 */
6833 hugetlb_cgroup_put_rsvd_cgroup(h_cg);
6834 }
6835 }
6836 return true;
6837
6838 out_put_pages:
6839 /* put back original number of pages, chg */
6840 (void)hugepage_subpool_put_pages(spool, chg);
6841 out_uncharge_cgroup:
6842 hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h),
6843 chg * pages_per_huge_page(h), h_cg);
6844 out_err:
6845 hugetlb_vma_lock_free(vma);
6846 if (!vma || vma->vm_flags & VM_MAYSHARE)
6847 /* Only call region_abort if the region_chg succeeded but the
6848 * region_add failed or didn't run.
6849 */
6850 if (chg >= 0 && add < 0)
6851 region_abort(resv_map, from, to, regions_needed);
6852 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
6853 kref_put(&resv_map->refs, resv_map_release);
6854 set_vma_resv_map(vma, NULL);
6855 }
6856 return false;
6857 }
6858
hugetlb_unreserve_pages(struct inode * inode,long start,long end,long freed)6859 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
6860 long freed)
6861 {
6862 struct hstate *h = hstate_inode(inode);
6863 struct resv_map *resv_map = inode_resv_map(inode);
6864 long chg = 0;
6865 struct hugepage_subpool *spool = subpool_inode(inode);
6866 long gbl_reserve;
6867
6868 /*
6869 * Since this routine can be called in the evict inode path for all
6870 * hugetlbfs inodes, resv_map could be NULL.
6871 */
6872 if (resv_map) {
6873 chg = region_del(resv_map, start, end);
6874 /*
6875 * region_del() can fail in the rare case where a region
6876 * must be split and another region descriptor can not be
6877 * allocated. If end == LONG_MAX, it will not fail.
6878 */
6879 if (chg < 0)
6880 return chg;
6881 }
6882
6883 spin_lock(&inode->i_lock);
6884 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
6885 spin_unlock(&inode->i_lock);
6886
6887 /*
6888 * If the subpool has a minimum size, the number of global
6889 * reservations to be released may be adjusted.
6890 *
6891 * Note that !resv_map implies freed == 0. So (chg - freed)
6892 * won't go negative.
6893 */
6894 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
6895 hugetlb_acct_memory(h, -gbl_reserve);
6896
6897 return 0;
6898 }
6899
6900 #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)6901 static unsigned long page_table_shareable(struct vm_area_struct *svma,
6902 struct vm_area_struct *vma,
6903 unsigned long addr, pgoff_t idx)
6904 {
6905 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
6906 svma->vm_start;
6907 unsigned long sbase = saddr & PUD_MASK;
6908 unsigned long s_end = sbase + PUD_SIZE;
6909
6910 /* Allow segments to share if only one is marked locked */
6911 unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED_MASK;
6912 unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED_MASK;
6913
6914 /*
6915 * match the virtual addresses, permission and the alignment of the
6916 * page table page.
6917 *
6918 * Also, vma_lock (vm_private_data) is required for sharing.
6919 */
6920 if (pmd_index(addr) != pmd_index(saddr) ||
6921 vm_flags != svm_flags ||
6922 !range_in_vma(svma, sbase, s_end) ||
6923 !svma->vm_private_data)
6924 return 0;
6925
6926 return saddr;
6927 }
6928
want_pmd_share(struct vm_area_struct * vma,unsigned long addr)6929 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
6930 {
6931 unsigned long start = addr & PUD_MASK;
6932 unsigned long end = start + PUD_SIZE;
6933
6934 #ifdef CONFIG_USERFAULTFD
6935 if (uffd_disable_huge_pmd_share(vma))
6936 return false;
6937 #endif
6938 /*
6939 * check on proper vm_flags and page table alignment
6940 */
6941 if (!(vma->vm_flags & VM_MAYSHARE))
6942 return false;
6943 if (!vma->vm_private_data) /* vma lock required for sharing */
6944 return false;
6945 if (!range_in_vma(vma, start, end))
6946 return false;
6947 return true;
6948 }
6949
6950 /*
6951 * Determine if start,end range within vma could be mapped by shared pmd.
6952 * If yes, adjust start and end to cover range associated with possible
6953 * shared pmd mappings.
6954 */
adjust_range_if_pmd_sharing_possible(struct vm_area_struct * vma,unsigned long * start,unsigned long * end)6955 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
6956 unsigned long *start, unsigned long *end)
6957 {
6958 unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE),
6959 v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE);
6960
6961 /*
6962 * vma needs to span at least one aligned PUD size, and the range
6963 * must be at least partially within in.
6964 */
6965 if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) ||
6966 (*end <= v_start) || (*start >= v_end))
6967 return;
6968
6969 /* Extend the range to be PUD aligned for a worst case scenario */
6970 if (*start > v_start)
6971 *start = ALIGN_DOWN(*start, PUD_SIZE);
6972
6973 if (*end < v_end)
6974 *end = ALIGN(*end, PUD_SIZE);
6975 }
6976
6977 /*
6978 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
6979 * and returns the corresponding pte. While this is not necessary for the
6980 * !shared pmd case because we can allocate the pmd later as well, it makes the
6981 * code much cleaner. pmd allocation is essential for the shared case because
6982 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
6983 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
6984 * bad pmd for sharing.
6985 */
huge_pmd_share(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pud_t * pud)6986 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
6987 unsigned long addr, pud_t *pud)
6988 {
6989 struct address_space *mapping = vma->vm_file->f_mapping;
6990 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
6991 vma->vm_pgoff;
6992 struct vm_area_struct *svma;
6993 unsigned long saddr;
6994 pte_t *spte = NULL;
6995 pte_t *pte;
6996
6997 i_mmap_lock_read(mapping);
6998 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
6999 if (svma == vma)
7000 continue;
7001
7002 saddr = page_table_shareable(svma, vma, addr, idx);
7003 if (saddr) {
7004 spte = hugetlb_walk(svma, saddr,
7005 vma_mmu_pagesize(svma));
7006 if (spte) {
7007 get_page(virt_to_page(spte));
7008 break;
7009 }
7010 }
7011 }
7012
7013 if (!spte)
7014 goto out;
7015
7016 spin_lock(&mm->page_table_lock);
7017 if (pud_none(*pud)) {
7018 pud_populate(mm, pud,
7019 (pmd_t *)((unsigned long)spte & PAGE_MASK));
7020 mm_inc_nr_pmds(mm);
7021 } else {
7022 put_page(virt_to_page(spte));
7023 }
7024 spin_unlock(&mm->page_table_lock);
7025 out:
7026 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7027 i_mmap_unlock_read(mapping);
7028 return pte;
7029 }
7030
7031 /*
7032 * unmap huge page backed by shared pte.
7033 *
7034 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
7035 * indicated by page_count > 1, unmap is achieved by clearing pud and
7036 * decrementing the ref count. If count == 1, the pte page is not shared.
7037 *
7038 * Called with page table lock held.
7039 *
7040 * returns: 1 successfully unmapped a shared pte page
7041 * 0 the underlying pte page is not shared, or it is the last user
7042 */
huge_pmd_unshare(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pte_t * ptep)7043 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7044 unsigned long addr, pte_t *ptep)
7045 {
7046 pgd_t *pgd = pgd_offset(mm, addr);
7047 p4d_t *p4d = p4d_offset(pgd, addr);
7048 pud_t *pud = pud_offset(p4d, addr);
7049
7050 i_mmap_assert_write_locked(vma->vm_file->f_mapping);
7051 hugetlb_vma_assert_locked(vma);
7052 BUG_ON(page_count(virt_to_page(ptep)) == 0);
7053 if (page_count(virt_to_page(ptep)) == 1)
7054 return 0;
7055
7056 pud_clear(pud);
7057 put_page(virt_to_page(ptep));
7058 mm_dec_nr_pmds(mm);
7059 return 1;
7060 }
7061
7062 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7063
huge_pmd_share(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pud_t * pud)7064 pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma,
7065 unsigned long addr, pud_t *pud)
7066 {
7067 return NULL;
7068 }
7069
huge_pmd_unshare(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,pte_t * ptep)7070 int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
7071 unsigned long addr, pte_t *ptep)
7072 {
7073 return 0;
7074 }
7075
adjust_range_if_pmd_sharing_possible(struct vm_area_struct * vma,unsigned long * start,unsigned long * end)7076 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
7077 unsigned long *start, unsigned long *end)
7078 {
7079 }
7080
want_pmd_share(struct vm_area_struct * vma,unsigned long addr)7081 bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr)
7082 {
7083 return false;
7084 }
7085 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
7086
7087 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
huge_pte_alloc(struct mm_struct * mm,struct vm_area_struct * vma,unsigned long addr,unsigned long sz)7088 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
7089 unsigned long addr, unsigned long sz)
7090 {
7091 pgd_t *pgd;
7092 p4d_t *p4d;
7093 pud_t *pud;
7094 pte_t *pte = NULL;
7095
7096 pgd = pgd_offset(mm, addr);
7097 p4d = p4d_alloc(mm, pgd, addr);
7098 if (!p4d)
7099 return NULL;
7100 pud = pud_alloc(mm, p4d, addr);
7101 if (pud) {
7102 if (sz == PUD_SIZE) {
7103 pte = (pte_t *)pud;
7104 } else {
7105 BUG_ON(sz != PMD_SIZE);
7106 if (want_pmd_share(vma, addr) && pud_none(*pud))
7107 pte = huge_pmd_share(mm, vma, addr, pud);
7108 else
7109 pte = (pte_t *)pmd_alloc(mm, pud, addr);
7110 }
7111 }
7112
7113 if (pte) {
7114 pte_t pteval = ptep_get_lockless(pte);
7115
7116 BUG_ON(pte_present(pteval) && !pte_huge(pteval));
7117 }
7118
7119 return pte;
7120 }
7121
7122 /*
7123 * huge_pte_offset() - Walk the page table to resolve the hugepage
7124 * entry at address @addr
7125 *
7126 * Return: Pointer to page table entry (PUD or PMD) for
7127 * address @addr, or NULL if a !p*d_present() entry is encountered and the
7128 * size @sz doesn't match the hugepage size at this level of the page
7129 * table.
7130 */
huge_pte_offset(struct mm_struct * mm,unsigned long addr,unsigned long sz)7131 pte_t *huge_pte_offset(struct mm_struct *mm,
7132 unsigned long addr, unsigned long sz)
7133 {
7134 pgd_t *pgd;
7135 p4d_t *p4d;
7136 pud_t *pud;
7137 pmd_t *pmd;
7138
7139 pgd = pgd_offset(mm, addr);
7140 if (!pgd_present(*pgd))
7141 return NULL;
7142 p4d = p4d_offset(pgd, addr);
7143 if (!p4d_present(*p4d))
7144 return NULL;
7145
7146 pud = pud_offset(p4d, addr);
7147 if (sz == PUD_SIZE)
7148 /* must be pud huge, non-present or none */
7149 return (pte_t *)pud;
7150 if (!pud_present(*pud))
7151 return NULL;
7152 /* must have a valid entry and size to go further */
7153
7154 pmd = pmd_offset(pud, addr);
7155 /* must be pmd huge, non-present or none */
7156 return (pte_t *)pmd;
7157 }
7158
7159 /*
7160 * Return a mask that can be used to update an address to the last huge
7161 * page in a page table page mapping size. Used to skip non-present
7162 * page table entries when linearly scanning address ranges. Architectures
7163 * with unique huge page to page table relationships can define their own
7164 * version of this routine.
7165 */
hugetlb_mask_last_page(struct hstate * h)7166 unsigned long hugetlb_mask_last_page(struct hstate *h)
7167 {
7168 unsigned long hp_size = huge_page_size(h);
7169
7170 if (hp_size == PUD_SIZE)
7171 return P4D_SIZE - PUD_SIZE;
7172 else if (hp_size == PMD_SIZE)
7173 return PUD_SIZE - PMD_SIZE;
7174 else
7175 return 0UL;
7176 }
7177
7178 #else
7179
7180 /* See description above. Architectures can provide their own version. */
hugetlb_mask_last_page(struct hstate * h)7181 __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
7182 {
7183 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
7184 if (huge_page_size(h) == PMD_SIZE)
7185 return PUD_SIZE - PMD_SIZE;
7186 #endif
7187 return 0UL;
7188 }
7189
7190 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
7191
7192 /*
7193 * These functions are overwritable if your architecture needs its own
7194 * behavior.
7195 */
isolate_hugetlb(struct folio * folio,struct list_head * list)7196 bool isolate_hugetlb(struct folio *folio, struct list_head *list)
7197 {
7198 bool ret = true;
7199
7200 spin_lock_irq(&hugetlb_lock);
7201 if (!folio_test_hugetlb(folio) ||
7202 !folio_test_hugetlb_migratable(folio) ||
7203 !folio_try_get(folio)) {
7204 ret = false;
7205 goto unlock;
7206 }
7207 folio_clear_hugetlb_migratable(folio);
7208 list_move_tail(&folio->lru, list);
7209 unlock:
7210 spin_unlock_irq(&hugetlb_lock);
7211 return ret;
7212 }
7213
get_hwpoison_hugetlb_folio(struct folio * folio,bool * hugetlb,bool unpoison)7214 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison)
7215 {
7216 int ret = 0;
7217
7218 *hugetlb = false;
7219 spin_lock_irq(&hugetlb_lock);
7220 if (folio_test_hugetlb(folio)) {
7221 *hugetlb = true;
7222 if (folio_test_hugetlb_freed(folio))
7223 ret = 0;
7224 else if (folio_test_hugetlb_migratable(folio) || unpoison)
7225 ret = folio_try_get(folio);
7226 else
7227 ret = -EBUSY;
7228 }
7229 spin_unlock_irq(&hugetlb_lock);
7230 return ret;
7231 }
7232
get_huge_page_for_hwpoison(unsigned long pfn,int flags,bool * migratable_cleared)7233 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
7234 bool *migratable_cleared)
7235 {
7236 int ret;
7237
7238 spin_lock_irq(&hugetlb_lock);
7239 ret = __get_huge_page_for_hwpoison(pfn, flags, migratable_cleared);
7240 spin_unlock_irq(&hugetlb_lock);
7241 return ret;
7242 }
7243
folio_putback_active_hugetlb(struct folio * folio)7244 void folio_putback_active_hugetlb(struct folio *folio)
7245 {
7246 spin_lock_irq(&hugetlb_lock);
7247 folio_set_hugetlb_migratable(folio);
7248 list_move_tail(&folio->lru, &(folio_hstate(folio))->hugepage_activelist);
7249 spin_unlock_irq(&hugetlb_lock);
7250 folio_put(folio);
7251 }
7252
move_hugetlb_state(struct folio * old_folio,struct folio * new_folio,int reason)7253 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason)
7254 {
7255 struct hstate *h = folio_hstate(old_folio);
7256
7257 hugetlb_cgroup_migrate(old_folio, new_folio);
7258 set_page_owner_migrate_reason(&new_folio->page, reason);
7259
7260 /*
7261 * transfer temporary state of the new hugetlb folio. This is
7262 * reverse to other transitions because the newpage is going to
7263 * be final while the old one will be freed so it takes over
7264 * the temporary status.
7265 *
7266 * Also note that we have to transfer the per-node surplus state
7267 * here as well otherwise the global surplus count will not match
7268 * the per-node's.
7269 */
7270 if (folio_test_hugetlb_temporary(new_folio)) {
7271 int old_nid = folio_nid(old_folio);
7272 int new_nid = folio_nid(new_folio);
7273
7274 folio_set_hugetlb_temporary(old_folio);
7275 folio_clear_hugetlb_temporary(new_folio);
7276
7277
7278 /*
7279 * There is no need to transfer the per-node surplus state
7280 * when we do not cross the node.
7281 */
7282 if (new_nid == old_nid)
7283 return;
7284 spin_lock_irq(&hugetlb_lock);
7285 if (h->surplus_huge_pages_node[old_nid]) {
7286 h->surplus_huge_pages_node[old_nid]--;
7287 h->surplus_huge_pages_node[new_nid]++;
7288 }
7289 spin_unlock_irq(&hugetlb_lock);
7290 }
7291 }
7292
hugetlb_unshare_pmds(struct vm_area_struct * vma,unsigned long start,unsigned long end)7293 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
7294 unsigned long start,
7295 unsigned long end)
7296 {
7297 struct hstate *h = hstate_vma(vma);
7298 unsigned long sz = huge_page_size(h);
7299 struct mm_struct *mm = vma->vm_mm;
7300 struct mmu_notifier_range range;
7301 unsigned long address;
7302 spinlock_t *ptl;
7303 pte_t *ptep;
7304
7305 if (!(vma->vm_flags & VM_MAYSHARE))
7306 return;
7307
7308 if (start >= end)
7309 return;
7310
7311 flush_cache_range(vma, start, end);
7312 /*
7313 * No need to call adjust_range_if_pmd_sharing_possible(), because
7314 * we have already done the PUD_SIZE alignment.
7315 */
7316 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
7317 start, end);
7318 mmu_notifier_invalidate_range_start(&range);
7319 hugetlb_vma_lock_write(vma);
7320 i_mmap_lock_write(vma->vm_file->f_mapping);
7321 for (address = start; address < end; address += PUD_SIZE) {
7322 ptep = hugetlb_walk(vma, address, sz);
7323 if (!ptep)
7324 continue;
7325 ptl = huge_pte_lock(h, mm, ptep);
7326 huge_pmd_unshare(mm, vma, address, ptep);
7327 spin_unlock(ptl);
7328 }
7329 flush_hugetlb_tlb_range(vma, start, end);
7330 i_mmap_unlock_write(vma->vm_file->f_mapping);
7331 hugetlb_vma_unlock_write(vma);
7332 /*
7333 * No need to call mmu_notifier_arch_invalidate_secondary_tlbs(), see
7334 * Documentation/mm/mmu_notifier.rst.
7335 */
7336 mmu_notifier_invalidate_range_end(&range);
7337 }
7338
7339 /*
7340 * This function will unconditionally remove all the shared pmd pgtable entries
7341 * within the specific vma for a hugetlbfs memory range.
7342 */
hugetlb_unshare_all_pmds(struct vm_area_struct * vma)7343 void hugetlb_unshare_all_pmds(struct vm_area_struct *vma)
7344 {
7345 hugetlb_unshare_pmds(vma, ALIGN(vma->vm_start, PUD_SIZE),
7346 ALIGN_DOWN(vma->vm_end, PUD_SIZE));
7347 }
7348
7349 #ifdef CONFIG_CMA
7350 static bool cma_reserve_called __initdata;
7351
cmdline_parse_hugetlb_cma(char * p)7352 static int __init cmdline_parse_hugetlb_cma(char *p)
7353 {
7354 int nid, count = 0;
7355 unsigned long tmp;
7356 char *s = p;
7357
7358 while (*s) {
7359 if (sscanf(s, "%lu%n", &tmp, &count) != 1)
7360 break;
7361
7362 if (s[count] == ':') {
7363 if (tmp >= MAX_NUMNODES)
7364 break;
7365 nid = array_index_nospec(tmp, MAX_NUMNODES);
7366
7367 s += count + 1;
7368 tmp = memparse(s, &s);
7369 hugetlb_cma_size_in_node[nid] = tmp;
7370 hugetlb_cma_size += tmp;
7371
7372 /*
7373 * Skip the separator if have one, otherwise
7374 * break the parsing.
7375 */
7376 if (*s == ',')
7377 s++;
7378 else
7379 break;
7380 } else {
7381 hugetlb_cma_size = memparse(p, &p);
7382 break;
7383 }
7384 }
7385
7386 return 0;
7387 }
7388
7389 early_param("hugetlb_cma", cmdline_parse_hugetlb_cma);
7390
hugetlb_cma_reserve(int order)7391 void __init hugetlb_cma_reserve(int order)
7392 {
7393 unsigned long size, reserved, per_node;
7394 bool node_specific_cma_alloc = false;
7395 int nid;
7396
7397 cma_reserve_called = true;
7398
7399 if (!hugetlb_cma_size)
7400 return;
7401
7402 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7403 if (hugetlb_cma_size_in_node[nid] == 0)
7404 continue;
7405
7406 if (!node_online(nid)) {
7407 pr_warn("hugetlb_cma: invalid node %d specified\n", nid);
7408 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7409 hugetlb_cma_size_in_node[nid] = 0;
7410 continue;
7411 }
7412
7413 if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) {
7414 pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n",
7415 nid, (PAGE_SIZE << order) / SZ_1M);
7416 hugetlb_cma_size -= hugetlb_cma_size_in_node[nid];
7417 hugetlb_cma_size_in_node[nid] = 0;
7418 } else {
7419 node_specific_cma_alloc = true;
7420 }
7421 }
7422
7423 /* Validate the CMA size again in case some invalid nodes specified. */
7424 if (!hugetlb_cma_size)
7425 return;
7426
7427 if (hugetlb_cma_size < (PAGE_SIZE << order)) {
7428 pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n",
7429 (PAGE_SIZE << order) / SZ_1M);
7430 hugetlb_cma_size = 0;
7431 return;
7432 }
7433
7434 if (!node_specific_cma_alloc) {
7435 /*
7436 * If 3 GB area is requested on a machine with 4 numa nodes,
7437 * let's allocate 1 GB on first three nodes and ignore the last one.
7438 */
7439 per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
7440 pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
7441 hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
7442 }
7443
7444 reserved = 0;
7445 for_each_online_node(nid) {
7446 int res;
7447 char name[CMA_MAX_NAME];
7448
7449 if (node_specific_cma_alloc) {
7450 if (hugetlb_cma_size_in_node[nid] == 0)
7451 continue;
7452
7453 size = hugetlb_cma_size_in_node[nid];
7454 } else {
7455 size = min(per_node, hugetlb_cma_size - reserved);
7456 }
7457
7458 size = round_up(size, PAGE_SIZE << order);
7459
7460 snprintf(name, sizeof(name), "hugetlb%d", nid);
7461 /*
7462 * Note that 'order per bit' is based on smallest size that
7463 * may be returned to CMA allocator in the case of
7464 * huge page demotion.
7465 */
7466 res = cma_declare_contiguous_nid(0, size, 0,
7467 PAGE_SIZE << HUGETLB_PAGE_ORDER,
7468 HUGETLB_PAGE_ORDER, false, name,
7469 &hugetlb_cma[nid], nid);
7470 if (res) {
7471 pr_warn("hugetlb_cma: reservation failed: err %d, node %d",
7472 res, nid);
7473 continue;
7474 }
7475
7476 reserved += size;
7477 pr_info("hugetlb_cma: reserved %lu MiB on node %d\n",
7478 size / SZ_1M, nid);
7479
7480 if (reserved >= hugetlb_cma_size)
7481 break;
7482 }
7483
7484 if (!reserved)
7485 /*
7486 * hugetlb_cma_size is used to determine if allocations from
7487 * cma are possible. Set to zero if no cma regions are set up.
7488 */
7489 hugetlb_cma_size = 0;
7490 }
7491
hugetlb_cma_check(void)7492 static void __init hugetlb_cma_check(void)
7493 {
7494 if (!hugetlb_cma_size || cma_reserve_called)
7495 return;
7496
7497 pr_warn("hugetlb_cma: the option isn't supported by current arch\n");
7498 }
7499
7500 #endif /* CONFIG_CMA */
7501