]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blob - mm/hugetlb.c
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/hid
[mirror_ubuntu-zesty-kernel.git] / mm / hugetlb.c
1 /*
2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
4 */
5 #include <linux/list.h>
6 #include <linux/init.h>
7 #include <linux/module.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/bootmem.h>
20 #include <linux/sysfs.h>
21 #include <linux/slab.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/page-isolation.h>
26 #include <linux/jhash.h>
27
28 #include <asm/page.h>
29 #include <asm/pgtable.h>
30 #include <asm/tlb.h>
31
32 #include <linux/io.h>
33 #include <linux/hugetlb.h>
34 #include <linux/hugetlb_cgroup.h>
35 #include <linux/node.h>
36 #include "internal.h"
37
38 int hugepages_treat_as_movable;
39
40 int hugetlb_max_hstate __read_mostly;
41 unsigned int default_hstate_idx;
42 struct hstate hstates[HUGE_MAX_HSTATE];
43 /*
44 * Minimum page order among possible hugepage sizes, set to a proper value
45 * at boot time.
46 */
47 static unsigned int minimum_order __read_mostly = UINT_MAX;
48
49 __initdata LIST_HEAD(huge_boot_pages);
50
51 /* for command line parsing */
52 static struct hstate * __initdata parsed_hstate;
53 static unsigned long __initdata default_hstate_max_huge_pages;
54 static unsigned long __initdata default_hstate_size;
55
56 /*
57 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
58 * free_huge_pages, and surplus_huge_pages.
59 */
60 DEFINE_SPINLOCK(hugetlb_lock);
61
62 /*
63 * Serializes faults on the same logical page. This is used to
64 * prevent spurious OOMs when the hugepage pool is fully utilized.
65 */
66 static int num_fault_mutexes;
67 struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
68
69 /* Forward declaration */
70 static int hugetlb_acct_memory(struct hstate *h, long delta);
71
72 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
73 {
74 bool free = (spool->count == 0) && (spool->used_hpages == 0);
75
76 spin_unlock(&spool->lock);
77
78 /* If no pages are used, and no other handles to the subpool
79 * remain, give up any reservations mased on minimum size and
80 * free the subpool */
81 if (free) {
82 if (spool->min_hpages != -1)
83 hugetlb_acct_memory(spool->hstate,
84 -spool->min_hpages);
85 kfree(spool);
86 }
87 }
88
89 struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
90 long min_hpages)
91 {
92 struct hugepage_subpool *spool;
93
94 spool = kzalloc(sizeof(*spool), GFP_KERNEL);
95 if (!spool)
96 return NULL;
97
98 spin_lock_init(&spool->lock);
99 spool->count = 1;
100 spool->max_hpages = max_hpages;
101 spool->hstate = h;
102 spool->min_hpages = min_hpages;
103
104 if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
105 kfree(spool);
106 return NULL;
107 }
108 spool->rsv_hpages = min_hpages;
109
110 return spool;
111 }
112
113 void hugepage_put_subpool(struct hugepage_subpool *spool)
114 {
115 spin_lock(&spool->lock);
116 BUG_ON(!spool->count);
117 spool->count--;
118 unlock_or_release_subpool(spool);
119 }
120
121 /*
122 * Subpool accounting for allocating and reserving pages.
123 * Return -ENOMEM if there are not enough resources to satisfy the
124 * the request. Otherwise, return the number of pages by which the
125 * global pools must be adjusted (upward). The returned value may
126 * only be different than the passed value (delta) in the case where
127 * a subpool minimum size must be manitained.
128 */
129 static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
130 long delta)
131 {
132 long ret = delta;
133
134 if (!spool)
135 return ret;
136
137 spin_lock(&spool->lock);
138
139 if (spool->max_hpages != -1) { /* maximum size accounting */
140 if ((spool->used_hpages + delta) <= spool->max_hpages)
141 spool->used_hpages += delta;
142 else {
143 ret = -ENOMEM;
144 goto unlock_ret;
145 }
146 }
147
148 if (spool->min_hpages != -1) { /* minimum size accounting */
149 if (delta > spool->rsv_hpages) {
150 /*
151 * Asking for more reserves than those already taken on
152 * behalf of subpool. Return difference.
153 */
154 ret = delta - spool->rsv_hpages;
155 spool->rsv_hpages = 0;
156 } else {
157 ret = 0; /* reserves already accounted for */
158 spool->rsv_hpages -= delta;
159 }
160 }
161
162 unlock_ret:
163 spin_unlock(&spool->lock);
164 return ret;
165 }
166
167 /*
168 * Subpool accounting for freeing and unreserving pages.
169 * Return the number of global page reservations that must be dropped.
170 * The return value may only be different than the passed value (delta)
171 * in the case where a subpool minimum size must be maintained.
172 */
173 static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
174 long delta)
175 {
176 long ret = delta;
177
178 if (!spool)
179 return delta;
180
181 spin_lock(&spool->lock);
182
183 if (spool->max_hpages != -1) /* maximum size accounting */
184 spool->used_hpages -= delta;
185
186 if (spool->min_hpages != -1) { /* minimum size accounting */
187 if (spool->rsv_hpages + delta <= spool->min_hpages)
188 ret = 0;
189 else
190 ret = spool->rsv_hpages + delta - spool->min_hpages;
191
192 spool->rsv_hpages += delta;
193 if (spool->rsv_hpages > spool->min_hpages)
194 spool->rsv_hpages = spool->min_hpages;
195 }
196
197 /*
198 * If hugetlbfs_put_super couldn't free spool due to an outstanding
199 * quota reference, free it now.
200 */
201 unlock_or_release_subpool(spool);
202
203 return ret;
204 }
205
206 static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
207 {
208 return HUGETLBFS_SB(inode->i_sb)->spool;
209 }
210
211 static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
212 {
213 return subpool_inode(file_inode(vma->vm_file));
214 }
215
216 /*
217 * Region tracking -- allows tracking of reservations and instantiated pages
218 * across the pages in a mapping.
219 *
220 * The region data structures are embedded into a resv_map and protected
221 * by a resv_map's lock. The set of regions within the resv_map represent
222 * reservations for huge pages, or huge pages that have already been
223 * instantiated within the map. The from and to elements are huge page
224 * indicies into the associated mapping. from indicates the starting index
225 * of the region. to represents the first index past the end of the region.
226 *
227 * For example, a file region structure with from == 0 and to == 4 represents
228 * four huge pages in a mapping. It is important to note that the to element
229 * represents the first element past the end of the region. This is used in
230 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
231 *
232 * Interval notation of the form [from, to) will be used to indicate that
233 * the endpoint from is inclusive and to is exclusive.
234 */
235 struct file_region {
236 struct list_head link;
237 long from;
238 long to;
239 };
240
241 /*
242 * Add the huge page range represented by [f, t) to the reserve
243 * map. In the normal case, existing regions will be expanded
244 * to accommodate the specified range. Sufficient regions should
245 * exist for expansion due to the previous call to region_chg
246 * with the same range. However, it is possible that region_del
247 * could have been called after region_chg and modifed the map
248 * in such a way that no region exists to be expanded. In this
249 * case, pull a region descriptor from the cache associated with
250 * the map and use that for the new range.
251 *
252 * Return the number of new huge pages added to the map. This
253 * number is greater than or equal to zero.
254 */
255 static long region_add(struct resv_map *resv, long f, long t)
256 {
257 struct list_head *head = &resv->regions;
258 struct file_region *rg, *nrg, *trg;
259 long add = 0;
260
261 spin_lock(&resv->lock);
262 /* Locate the region we are either in or before. */
263 list_for_each_entry(rg, head, link)
264 if (f <= rg->to)
265 break;
266
267 /*
268 * If no region exists which can be expanded to include the
269 * specified range, the list must have been modified by an
270 * interleving call to region_del(). Pull a region descriptor
271 * from the cache and use it for this range.
272 */
273 if (&rg->link == head || t < rg->from) {
274 VM_BUG_ON(resv->region_cache_count <= 0);
275
276 resv->region_cache_count--;
277 nrg = list_first_entry(&resv->region_cache, struct file_region,
278 link);
279 list_del(&nrg->link);
280
281 nrg->from = f;
282 nrg->to = t;
283 list_add(&nrg->link, rg->link.prev);
284
285 add += t - f;
286 goto out_locked;
287 }
288
289 /* Round our left edge to the current segment if it encloses us. */
290 if (f > rg->from)
291 f = rg->from;
292
293 /* Check for and consume any regions we now overlap with. */
294 nrg = rg;
295 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
296 if (&rg->link == head)
297 break;
298 if (rg->from > t)
299 break;
300
301 /* If this area reaches higher then extend our area to
302 * include it completely. If this is not the first area
303 * which we intend to reuse, free it. */
304 if (rg->to > t)
305 t = rg->to;
306 if (rg != nrg) {
307 /* Decrement return value by the deleted range.
308 * Another range will span this area so that by
309 * end of routine add will be >= zero
310 */
311 add -= (rg->to - rg->from);
312 list_del(&rg->link);
313 kfree(rg);
314 }
315 }
316
317 add += (nrg->from - f); /* Added to beginning of region */
318 nrg->from = f;
319 add += t - nrg->to; /* Added to end of region */
320 nrg->to = t;
321
322 out_locked:
323 resv->adds_in_progress--;
324 spin_unlock(&resv->lock);
325 VM_BUG_ON(add < 0);
326 return add;
327 }
328
329 /*
330 * Examine the existing reserve map and determine how many
331 * huge pages in the specified range [f, t) are NOT currently
332 * represented. This routine is called before a subsequent
333 * call to region_add that will actually modify the reserve
334 * map to add the specified range [f, t). region_chg does
335 * not change the number of huge pages represented by the
336 * map. However, if the existing regions in the map can not
337 * be expanded to represent the new range, a new file_region
338 * structure is added to the map as a placeholder. This is
339 * so that the subsequent region_add call will have all the
340 * regions it needs and will not fail.
341 *
342 * Upon entry, region_chg will also examine the cache of region descriptors
343 * associated with the map. If there are not enough descriptors cached, one
344 * will be allocated for the in progress add operation.
345 *
346 * Returns the number of huge pages that need to be added to the existing
347 * reservation map for the range [f, t). This number is greater or equal to
348 * zero. -ENOMEM is returned if a new file_region structure or cache entry
349 * is needed and can not be allocated.
350 */
351 static long region_chg(struct resv_map *resv, long f, long t)
352 {
353 struct list_head *head = &resv->regions;
354 struct file_region *rg, *nrg = NULL;
355 long chg = 0;
356
357 retry:
358 spin_lock(&resv->lock);
359 retry_locked:
360 resv->adds_in_progress++;
361
362 /*
363 * Check for sufficient descriptors in the cache to accommodate
364 * the number of in progress add operations.
365 */
366 if (resv->adds_in_progress > resv->region_cache_count) {
367 struct file_region *trg;
368
369 VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
370 /* Must drop lock to allocate a new descriptor. */
371 resv->adds_in_progress--;
372 spin_unlock(&resv->lock);
373
374 trg = kmalloc(sizeof(*trg), GFP_KERNEL);
375 if (!trg)
376 return -ENOMEM;
377
378 spin_lock(&resv->lock);
379 list_add(&trg->link, &resv->region_cache);
380 resv->region_cache_count++;
381 goto retry_locked;
382 }
383
384 /* Locate the region we are before or in. */
385 list_for_each_entry(rg, head, link)
386 if (f <= rg->to)
387 break;
388
389 /* If we are below the current region then a new region is required.
390 * Subtle, allocate a new region at the position but make it zero
391 * size such that we can guarantee to record the reservation. */
392 if (&rg->link == head || t < rg->from) {
393 if (!nrg) {
394 resv->adds_in_progress--;
395 spin_unlock(&resv->lock);
396 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
397 if (!nrg)
398 return -ENOMEM;
399
400 nrg->from = f;
401 nrg->to = f;
402 INIT_LIST_HEAD(&nrg->link);
403 goto retry;
404 }
405
406 list_add(&nrg->link, rg->link.prev);
407 chg = t - f;
408 goto out_nrg;
409 }
410
411 /* Round our left edge to the current segment if it encloses us. */
412 if (f > rg->from)
413 f = rg->from;
414 chg = t - f;
415
416 /* Check for and consume any regions we now overlap with. */
417 list_for_each_entry(rg, rg->link.prev, link) {
418 if (&rg->link == head)
419 break;
420 if (rg->from > t)
421 goto out;
422
423 /* We overlap with this area, if it extends further than
424 * us then we must extend ourselves. Account for its
425 * existing reservation. */
426 if (rg->to > t) {
427 chg += rg->to - t;
428 t = rg->to;
429 }
430 chg -= rg->to - rg->from;
431 }
432
433 out:
434 spin_unlock(&resv->lock);
435 /* We already know we raced and no longer need the new region */
436 kfree(nrg);
437 return chg;
438 out_nrg:
439 spin_unlock(&resv->lock);
440 return chg;
441 }
442
443 /*
444 * Abort the in progress add operation. The adds_in_progress field
445 * of the resv_map keeps track of the operations in progress between
446 * calls to region_chg and region_add. Operations are sometimes
447 * aborted after the call to region_chg. In such cases, region_abort
448 * is called to decrement the adds_in_progress counter.
449 *
450 * NOTE: The range arguments [f, t) are not needed or used in this
451 * routine. They are kept to make reading the calling code easier as
452 * arguments will match the associated region_chg call.
453 */
454 static void region_abort(struct resv_map *resv, long f, long t)
455 {
456 spin_lock(&resv->lock);
457 VM_BUG_ON(!resv->region_cache_count);
458 resv->adds_in_progress--;
459 spin_unlock(&resv->lock);
460 }
461
462 /*
463 * Delete the specified range [f, t) from the reserve map. If the
464 * t parameter is LONG_MAX, this indicates that ALL regions after f
465 * should be deleted. Locate the regions which intersect [f, t)
466 * and either trim, delete or split the existing regions.
467 *
468 * Returns the number of huge pages deleted from the reserve map.
469 * In the normal case, the return value is zero or more. In the
470 * case where a region must be split, a new region descriptor must
471 * be allocated. If the allocation fails, -ENOMEM will be returned.
472 * NOTE: If the parameter t == LONG_MAX, then we will never split
473 * a region and possibly return -ENOMEM. Callers specifying
474 * t == LONG_MAX do not need to check for -ENOMEM error.
475 */
476 static long region_del(struct resv_map *resv, long f, long t)
477 {
478 struct list_head *head = &resv->regions;
479 struct file_region *rg, *trg;
480 struct file_region *nrg = NULL;
481 long del = 0;
482
483 retry:
484 spin_lock(&resv->lock);
485 list_for_each_entry_safe(rg, trg, head, link) {
486 if (rg->to <= f)
487 continue;
488 if (rg->from >= t)
489 break;
490
491 if (f > rg->from && t < rg->to) { /* Must split region */
492 /*
493 * Check for an entry in the cache before dropping
494 * lock and attempting allocation.
495 */
496 if (!nrg &&
497 resv->region_cache_count > resv->adds_in_progress) {
498 nrg = list_first_entry(&resv->region_cache,
499 struct file_region,
500 link);
501 list_del(&nrg->link);
502 resv->region_cache_count--;
503 }
504
505 if (!nrg) {
506 spin_unlock(&resv->lock);
507 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
508 if (!nrg)
509 return -ENOMEM;
510 goto retry;
511 }
512
513 del += t - f;
514
515 /* New entry for end of split region */
516 nrg->from = t;
517 nrg->to = rg->to;
518 INIT_LIST_HEAD(&nrg->link);
519
520 /* Original entry is trimmed */
521 rg->to = f;
522
523 list_add(&nrg->link, &rg->link);
524 nrg = NULL;
525 break;
526 }
527
528 if (f <= rg->from && t >= rg->to) { /* Remove entire region */
529 del += rg->to - rg->from;
530 list_del(&rg->link);
531 kfree(rg);
532 continue;
533 }
534
535 if (f <= rg->from) { /* Trim beginning of region */
536 del += t - rg->from;
537 rg->from = t;
538 } else { /* Trim end of region */
539 del += rg->to - f;
540 rg->to = f;
541 }
542 }
543
544 spin_unlock(&resv->lock);
545 kfree(nrg);
546 return del;
547 }
548
549 /*
550 * A rare out of memory error was encountered which prevented removal of
551 * the reserve map region for a page. The huge page itself was free'ed
552 * and removed from the page cache. This routine will adjust the subpool
553 * usage count, and the global reserve count if needed. By incrementing
554 * these counts, the reserve map entry which could not be deleted will
555 * appear as a "reserved" entry instead of simply dangling with incorrect
556 * counts.
557 */
558 void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve)
559 {
560 struct hugepage_subpool *spool = subpool_inode(inode);
561 long rsv_adjust;
562
563 rsv_adjust = hugepage_subpool_get_pages(spool, 1);
564 if (restore_reserve && rsv_adjust) {
565 struct hstate *h = hstate_inode(inode);
566
567 hugetlb_acct_memory(h, 1);
568 }
569 }
570
571 /*
572 * Count and return the number of huge pages in the reserve map
573 * that intersect with the range [f, t).
574 */
575 static long region_count(struct resv_map *resv, long f, long t)
576 {
577 struct list_head *head = &resv->regions;
578 struct file_region *rg;
579 long chg = 0;
580
581 spin_lock(&resv->lock);
582 /* Locate each segment we overlap with, and count that overlap. */
583 list_for_each_entry(rg, head, link) {
584 long seg_from;
585 long seg_to;
586
587 if (rg->to <= f)
588 continue;
589 if (rg->from >= t)
590 break;
591
592 seg_from = max(rg->from, f);
593 seg_to = min(rg->to, t);
594
595 chg += seg_to - seg_from;
596 }
597 spin_unlock(&resv->lock);
598
599 return chg;
600 }
601
602 /*
603 * Convert the address within this vma to the page offset within
604 * the mapping, in pagecache page units; huge pages here.
605 */
606 static pgoff_t vma_hugecache_offset(struct hstate *h,
607 struct vm_area_struct *vma, unsigned long address)
608 {
609 return ((address - vma->vm_start) >> huge_page_shift(h)) +
610 (vma->vm_pgoff >> huge_page_order(h));
611 }
612
613 pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
614 unsigned long address)
615 {
616 return vma_hugecache_offset(hstate_vma(vma), vma, address);
617 }
618
619 /*
620 * Return the size of the pages allocated when backing a VMA. In the majority
621 * cases this will be same size as used by the page table entries.
622 */
623 unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
624 {
625 struct hstate *hstate;
626
627 if (!is_vm_hugetlb_page(vma))
628 return PAGE_SIZE;
629
630 hstate = hstate_vma(vma);
631
632 return 1UL << huge_page_shift(hstate);
633 }
634 EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
635
636 /*
637 * Return the page size being used by the MMU to back a VMA. In the majority
638 * of cases, the page size used by the kernel matches the MMU size. On
639 * architectures where it differs, an architecture-specific version of this
640 * function is required.
641 */
642 #ifndef vma_mmu_pagesize
643 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
644 {
645 return vma_kernel_pagesize(vma);
646 }
647 #endif
648
649 /*
650 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
651 * bits of the reservation map pointer, which are always clear due to
652 * alignment.
653 */
654 #define HPAGE_RESV_OWNER (1UL << 0)
655 #define HPAGE_RESV_UNMAPPED (1UL << 1)
656 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
657
658 /*
659 * These helpers are used to track how many pages are reserved for
660 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
661 * is guaranteed to have their future faults succeed.
662 *
663 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
664 * the reserve counters are updated with the hugetlb_lock held. It is safe
665 * to reset the VMA at fork() time as it is not in use yet and there is no
666 * chance of the global counters getting corrupted as a result of the values.
667 *
668 * The private mapping reservation is represented in a subtly different
669 * manner to a shared mapping. A shared mapping has a region map associated
670 * with the underlying file, this region map represents the backing file
671 * pages which have ever had a reservation assigned which this persists even
672 * after the page is instantiated. A private mapping has a region map
673 * associated with the original mmap which is attached to all VMAs which
674 * reference it, this region map represents those offsets which have consumed
675 * reservation ie. where pages have been instantiated.
676 */
677 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
678 {
679 return (unsigned long)vma->vm_private_data;
680 }
681
682 static void set_vma_private_data(struct vm_area_struct *vma,
683 unsigned long value)
684 {
685 vma->vm_private_data = (void *)value;
686 }
687
688 struct resv_map *resv_map_alloc(void)
689 {
690 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
691 struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
692
693 if (!resv_map || !rg) {
694 kfree(resv_map);
695 kfree(rg);
696 return NULL;
697 }
698
699 kref_init(&resv_map->refs);
700 spin_lock_init(&resv_map->lock);
701 INIT_LIST_HEAD(&resv_map->regions);
702
703 resv_map->adds_in_progress = 0;
704
705 INIT_LIST_HEAD(&resv_map->region_cache);
706 list_add(&rg->link, &resv_map->region_cache);
707 resv_map->region_cache_count = 1;
708
709 return resv_map;
710 }
711
712 void resv_map_release(struct kref *ref)
713 {
714 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
715 struct list_head *head = &resv_map->region_cache;
716 struct file_region *rg, *trg;
717
718 /* Clear out any active regions before we release the map. */
719 region_del(resv_map, 0, LONG_MAX);
720
721 /* ... and any entries left in the cache */
722 list_for_each_entry_safe(rg, trg, head, link) {
723 list_del(&rg->link);
724 kfree(rg);
725 }
726
727 VM_BUG_ON(resv_map->adds_in_progress);
728
729 kfree(resv_map);
730 }
731
732 static inline struct resv_map *inode_resv_map(struct inode *inode)
733 {
734 return inode->i_mapping->private_data;
735 }
736
737 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
738 {
739 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
740 if (vma->vm_flags & VM_MAYSHARE) {
741 struct address_space *mapping = vma->vm_file->f_mapping;
742 struct inode *inode = mapping->host;
743
744 return inode_resv_map(inode);
745
746 } else {
747 return (struct resv_map *)(get_vma_private_data(vma) &
748 ~HPAGE_RESV_MASK);
749 }
750 }
751
752 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
753 {
754 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
755 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
756
757 set_vma_private_data(vma, (get_vma_private_data(vma) &
758 HPAGE_RESV_MASK) | (unsigned long)map);
759 }
760
761 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
762 {
763 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
764 VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
765
766 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
767 }
768
769 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
770 {
771 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
772
773 return (get_vma_private_data(vma) & flag) != 0;
774 }
775
776 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
777 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
778 {
779 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
780 if (!(vma->vm_flags & VM_MAYSHARE))
781 vma->vm_private_data = (void *)0;
782 }
783
784 /* Returns true if the VMA has associated reserve pages */
785 static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
786 {
787 if (vma->vm_flags & VM_NORESERVE) {
788 /*
789 * This address is already reserved by other process(chg == 0),
790 * so, we should decrement reserved count. Without decrementing,
791 * reserve count remains after releasing inode, because this
792 * allocated page will go into page cache and is regarded as
793 * coming from reserved pool in releasing step. Currently, we
794 * don't have any other solution to deal with this situation
795 * properly, so add work-around here.
796 */
797 if (vma->vm_flags & VM_MAYSHARE && chg == 0)
798 return true;
799 else
800 return false;
801 }
802
803 /* Shared mappings always use reserves */
804 if (vma->vm_flags & VM_MAYSHARE) {
805 /*
806 * We know VM_NORESERVE is not set. Therefore, there SHOULD
807 * be a region map for all pages. The only situation where
808 * there is no region map is if a hole was punched via
809 * fallocate. In this case, there really are no reverves to
810 * use. This situation is indicated if chg != 0.
811 */
812 if (chg)
813 return false;
814 else
815 return true;
816 }
817
818 /*
819 * Only the process that called mmap() has reserves for
820 * private mappings.
821 */
822 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
823 return true;
824
825 return false;
826 }
827
828 static void enqueue_huge_page(struct hstate *h, struct page *page)
829 {
830 int nid = page_to_nid(page);
831 list_move(&page->lru, &h->hugepage_freelists[nid]);
832 h->free_huge_pages++;
833 h->free_huge_pages_node[nid]++;
834 }
835
836 static struct page *dequeue_huge_page_node(struct hstate *h, int nid)
837 {
838 struct page *page;
839
840 list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
841 if (!is_migrate_isolate_page(page))
842 break;
843 /*
844 * if 'non-isolated free hugepage' not found on the list,
845 * the allocation fails.
846 */
847 if (&h->hugepage_freelists[nid] == &page->lru)
848 return NULL;
849 list_move(&page->lru, &h->hugepage_activelist);
850 set_page_refcounted(page);
851 h->free_huge_pages--;
852 h->free_huge_pages_node[nid]--;
853 return page;
854 }
855
856 /* Movability of hugepages depends on migration support. */
857 static inline gfp_t htlb_alloc_mask(struct hstate *h)
858 {
859 if (hugepages_treat_as_movable || hugepage_migration_supported(h))
860 return GFP_HIGHUSER_MOVABLE;
861 else
862 return GFP_HIGHUSER;
863 }
864
865 static struct page *dequeue_huge_page_vma(struct hstate *h,
866 struct vm_area_struct *vma,
867 unsigned long address, int avoid_reserve,
868 long chg)
869 {
870 struct page *page = NULL;
871 struct mempolicy *mpol;
872 nodemask_t *nodemask;
873 struct zonelist *zonelist;
874 struct zone *zone;
875 struct zoneref *z;
876 unsigned int cpuset_mems_cookie;
877
878 /*
879 * A child process with MAP_PRIVATE mappings created by their parent
880 * have no page reserves. This check ensures that reservations are
881 * not "stolen". The child may still get SIGKILLed
882 */
883 if (!vma_has_reserves(vma, chg) &&
884 h->free_huge_pages - h->resv_huge_pages == 0)
885 goto err;
886
887 /* If reserves cannot be used, ensure enough pages are in the pool */
888 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
889 goto err;
890
891 retry_cpuset:
892 cpuset_mems_cookie = read_mems_allowed_begin();
893 zonelist = huge_zonelist(vma, address,
894 htlb_alloc_mask(h), &mpol, &nodemask);
895
896 for_each_zone_zonelist_nodemask(zone, z, zonelist,
897 MAX_NR_ZONES - 1, nodemask) {
898 if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) {
899 page = dequeue_huge_page_node(h, zone_to_nid(zone));
900 if (page) {
901 if (avoid_reserve)
902 break;
903 if (!vma_has_reserves(vma, chg))
904 break;
905
906 SetPagePrivate(page);
907 h->resv_huge_pages--;
908 break;
909 }
910 }
911 }
912
913 mpol_cond_put(mpol);
914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
915 goto retry_cpuset;
916 return page;
917
918 err:
919 return NULL;
920 }
921
922 /*
923 * common helper functions for hstate_next_node_to_{alloc|free}.
924 * We may have allocated or freed a huge page based on a different
925 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
926 * be outside of *nodes_allowed. Ensure that we use an allowed
927 * node for alloc or free.
928 */
929 static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
930 {
931 nid = next_node(nid, *nodes_allowed);
932 if (nid == MAX_NUMNODES)
933 nid = first_node(*nodes_allowed);
934 VM_BUG_ON(nid >= MAX_NUMNODES);
935
936 return nid;
937 }
938
939 static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
940 {
941 if (!node_isset(nid, *nodes_allowed))
942 nid = next_node_allowed(nid, nodes_allowed);
943 return nid;
944 }
945
946 /*
947 * returns the previously saved node ["this node"] from which to
948 * allocate a persistent huge page for the pool and advance the
949 * next node from which to allocate, handling wrap at end of node
950 * mask.
951 */
952 static int hstate_next_node_to_alloc(struct hstate *h,
953 nodemask_t *nodes_allowed)
954 {
955 int nid;
956
957 VM_BUG_ON(!nodes_allowed);
958
959 nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
960 h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
961
962 return nid;
963 }
964
965 /*
966 * helper for free_pool_huge_page() - return the previously saved
967 * node ["this node"] from which to free a huge page. Advance the
968 * next node id whether or not we find a free huge page to free so
969 * that the next attempt to free addresses the next node.
970 */
971 static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
972 {
973 int nid;
974
975 VM_BUG_ON(!nodes_allowed);
976
977 nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
978 h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
979
980 return nid;
981 }
982
983 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
984 for (nr_nodes = nodes_weight(*mask); \
985 nr_nodes > 0 && \
986 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
987 nr_nodes--)
988
989 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
990 for (nr_nodes = nodes_weight(*mask); \
991 nr_nodes > 0 && \
992 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
993 nr_nodes--)
994
995 #if defined(CONFIG_CMA) && defined(CONFIG_X86_64)
996 static void destroy_compound_gigantic_page(struct page *page,
997 unsigned int order)
998 {
999 int i;
1000 int nr_pages = 1 << order;
1001 struct page *p = page + 1;
1002
1003 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1004 clear_compound_head(p);
1005 set_page_refcounted(p);
1006 }
1007
1008 set_compound_order(page, 0);
1009 __ClearPageHead(page);
1010 }
1011
1012 static void free_gigantic_page(struct page *page, unsigned int order)
1013 {
1014 free_contig_range(page_to_pfn(page), 1 << order);
1015 }
1016
1017 static int __alloc_gigantic_page(unsigned long start_pfn,
1018 unsigned long nr_pages)
1019 {
1020 unsigned long end_pfn = start_pfn + nr_pages;
1021 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE);
1022 }
1023
1024 static bool pfn_range_valid_gigantic(unsigned long start_pfn,
1025 unsigned long nr_pages)
1026 {
1027 unsigned long i, end_pfn = start_pfn + nr_pages;
1028 struct page *page;
1029
1030 for (i = start_pfn; i < end_pfn; i++) {
1031 if (!pfn_valid(i))
1032 return false;
1033
1034 page = pfn_to_page(i);
1035
1036 if (PageReserved(page))
1037 return false;
1038
1039 if (page_count(page) > 0)
1040 return false;
1041
1042 if (PageHuge(page))
1043 return false;
1044 }
1045
1046 return true;
1047 }
1048
1049 static bool zone_spans_last_pfn(const struct zone *zone,
1050 unsigned long start_pfn, unsigned long nr_pages)
1051 {
1052 unsigned long last_pfn = start_pfn + nr_pages - 1;
1053 return zone_spans_pfn(zone, last_pfn);
1054 }
1055
1056 static struct page *alloc_gigantic_page(int nid, unsigned int order)
1057 {
1058 unsigned long nr_pages = 1 << order;
1059 unsigned long ret, pfn, flags;
1060 struct zone *z;
1061
1062 z = NODE_DATA(nid)->node_zones;
1063 for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) {
1064 spin_lock_irqsave(&z->lock, flags);
1065
1066 pfn = ALIGN(z->zone_start_pfn, nr_pages);
1067 while (zone_spans_last_pfn(z, pfn, nr_pages)) {
1068 if (pfn_range_valid_gigantic(pfn, nr_pages)) {
1069 /*
1070 * We release the zone lock here because
1071 * alloc_contig_range() will also lock the zone
1072 * at some point. If there's an allocation
1073 * spinning on this lock, it may win the race
1074 * and cause alloc_contig_range() to fail...
1075 */
1076 spin_unlock_irqrestore(&z->lock, flags);
1077 ret = __alloc_gigantic_page(pfn, nr_pages);
1078 if (!ret)
1079 return pfn_to_page(pfn);
1080 spin_lock_irqsave(&z->lock, flags);
1081 }
1082 pfn += nr_pages;
1083 }
1084
1085 spin_unlock_irqrestore(&z->lock, flags);
1086 }
1087
1088 return NULL;
1089 }
1090
1091 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1092 static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1093
1094 static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
1095 {
1096 struct page *page;
1097
1098 page = alloc_gigantic_page(nid, huge_page_order(h));
1099 if (page) {
1100 prep_compound_gigantic_page(page, huge_page_order(h));
1101 prep_new_huge_page(h, page, nid);
1102 }
1103
1104 return page;
1105 }
1106
1107 static int alloc_fresh_gigantic_page(struct hstate *h,
1108 nodemask_t *nodes_allowed)
1109 {
1110 struct page *page = NULL;
1111 int nr_nodes, node;
1112
1113 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1114 page = alloc_fresh_gigantic_page_node(h, node);
1115 if (page)
1116 return 1;
1117 }
1118
1119 return 0;
1120 }
1121
1122 static inline bool gigantic_page_supported(void) { return true; }
1123 #else
1124 static inline bool gigantic_page_supported(void) { return false; }
1125 static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1126 static inline void destroy_compound_gigantic_page(struct page *page,
1127 unsigned int order) { }
1128 static inline int alloc_fresh_gigantic_page(struct hstate *h,
1129 nodemask_t *nodes_allowed) { return 0; }
1130 #endif
1131
1132 static void update_and_free_page(struct hstate *h, struct page *page)
1133 {
1134 int i;
1135
1136 if (hstate_is_gigantic(h) && !gigantic_page_supported())
1137 return;
1138
1139 h->nr_huge_pages--;
1140 h->nr_huge_pages_node[page_to_nid(page)]--;
1141 for (i = 0; i < pages_per_huge_page(h); i++) {
1142 page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1143 1 << PG_referenced | 1 << PG_dirty |
1144 1 << PG_active | 1 << PG_private |
1145 1 << PG_writeback);
1146 }
1147 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1148 set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1149 set_page_refcounted(page);
1150 if (hstate_is_gigantic(h)) {
1151 destroy_compound_gigantic_page(page, huge_page_order(h));
1152 free_gigantic_page(page, huge_page_order(h));
1153 } else {
1154 __free_pages(page, huge_page_order(h));
1155 }
1156 }
1157
1158 struct hstate *size_to_hstate(unsigned long size)
1159 {
1160 struct hstate *h;
1161
1162 for_each_hstate(h) {
1163 if (huge_page_size(h) == size)
1164 return h;
1165 }
1166 return NULL;
1167 }
1168
1169 /*
1170 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1171 * to hstate->hugepage_activelist.)
1172 *
1173 * This function can be called for tail pages, but never returns true for them.
1174 */
1175 bool page_huge_active(struct page *page)
1176 {
1177 VM_BUG_ON_PAGE(!PageHuge(page), page);
1178 return PageHead(page) && PagePrivate(&page[1]);
1179 }
1180
1181 /* never called for tail page */
1182 static void set_page_huge_active(struct page *page)
1183 {
1184 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1185 SetPagePrivate(&page[1]);
1186 }
1187
1188 static void clear_page_huge_active(struct page *page)
1189 {
1190 VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1191 ClearPagePrivate(&page[1]);
1192 }
1193
1194 void free_huge_page(struct page *page)
1195 {
1196 /*
1197 * Can't pass hstate in here because it is called from the
1198 * compound page destructor.
1199 */
1200 struct hstate *h = page_hstate(page);
1201 int nid = page_to_nid(page);
1202 struct hugepage_subpool *spool =
1203 (struct hugepage_subpool *)page_private(page);
1204 bool restore_reserve;
1205
1206 set_page_private(page, 0);
1207 page->mapping = NULL;
1208 BUG_ON(page_count(page));
1209 BUG_ON(page_mapcount(page));
1210 restore_reserve = PagePrivate(page);
1211 ClearPagePrivate(page);
1212
1213 /*
1214 * A return code of zero implies that the subpool will be under its
1215 * minimum size if the reservation is not restored after page is free.
1216 * Therefore, force restore_reserve operation.
1217 */
1218 if (hugepage_subpool_put_pages(spool, 1) == 0)
1219 restore_reserve = true;
1220
1221 spin_lock(&hugetlb_lock);
1222 clear_page_huge_active(page);
1223 hugetlb_cgroup_uncharge_page(hstate_index(h),
1224 pages_per_huge_page(h), page);
1225 if (restore_reserve)
1226 h->resv_huge_pages++;
1227
1228 if (h->surplus_huge_pages_node[nid]) {
1229 /* remove the page from active list */
1230 list_del(&page->lru);
1231 update_and_free_page(h, page);
1232 h->surplus_huge_pages--;
1233 h->surplus_huge_pages_node[nid]--;
1234 } else {
1235 arch_clear_hugepage_flags(page);
1236 enqueue_huge_page(h, page);
1237 }
1238 spin_unlock(&hugetlb_lock);
1239 }
1240
1241 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1242 {
1243 INIT_LIST_HEAD(&page->lru);
1244 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1245 spin_lock(&hugetlb_lock);
1246 set_hugetlb_cgroup(page, NULL);
1247 h->nr_huge_pages++;
1248 h->nr_huge_pages_node[nid]++;
1249 spin_unlock(&hugetlb_lock);
1250 put_page(page); /* free it into the hugepage allocator */
1251 }
1252
1253 static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1254 {
1255 int i;
1256 int nr_pages = 1 << order;
1257 struct page *p = page + 1;
1258
1259 /* we rely on prep_new_huge_page to set the destructor */
1260 set_compound_order(page, order);
1261 __SetPageHead(page);
1262 __ClearPageReserved(page);
1263 for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1264 /*
1265 * For gigantic hugepages allocated through bootmem at
1266 * boot, it's safer to be consistent with the not-gigantic
1267 * hugepages and clear the PG_reserved bit from all tail pages
1268 * too. Otherwse drivers using get_user_pages() to access tail
1269 * pages may get the reference counting wrong if they see
1270 * PG_reserved set on a tail page (despite the head page not
1271 * having PG_reserved set). Enforcing this consistency between
1272 * head and tail pages allows drivers to optimize away a check
1273 * on the head page when they need know if put_page() is needed
1274 * after get_user_pages().
1275 */
1276 __ClearPageReserved(p);
1277 set_page_count(p, 0);
1278 set_compound_head(p, page);
1279 }
1280 }
1281
1282 /*
1283 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1284 * transparent huge pages. See the PageTransHuge() documentation for more
1285 * details.
1286 */
1287 int PageHuge(struct page *page)
1288 {
1289 if (!PageCompound(page))
1290 return 0;
1291
1292 page = compound_head(page);
1293 return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1294 }
1295 EXPORT_SYMBOL_GPL(PageHuge);
1296
1297 /*
1298 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1299 * normal or transparent huge pages.
1300 */
1301 int PageHeadHuge(struct page *page_head)
1302 {
1303 if (!PageHead(page_head))
1304 return 0;
1305
1306 return get_compound_page_dtor(page_head) == free_huge_page;
1307 }
1308
1309 pgoff_t __basepage_index(struct page *page)
1310 {
1311 struct page *page_head = compound_head(page);
1312 pgoff_t index = page_index(page_head);
1313 unsigned long compound_idx;
1314
1315 if (!PageHuge(page_head))
1316 return page_index(page);
1317
1318 if (compound_order(page_head) >= MAX_ORDER)
1319 compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1320 else
1321 compound_idx = page - page_head;
1322
1323 return (index << compound_order(page_head)) + compound_idx;
1324 }
1325
1326 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
1327 {
1328 struct page *page;
1329
1330 page = __alloc_pages_node(nid,
1331 htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
1332 __GFP_REPEAT|__GFP_NOWARN,
1333 huge_page_order(h));
1334 if (page) {
1335 prep_new_huge_page(h, page, nid);
1336 }
1337
1338 return page;
1339 }
1340
1341 static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
1342 {
1343 struct page *page;
1344 int nr_nodes, node;
1345 int ret = 0;
1346
1347 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1348 page = alloc_fresh_huge_page_node(h, node);
1349 if (page) {
1350 ret = 1;
1351 break;
1352 }
1353 }
1354
1355 if (ret)
1356 count_vm_event(HTLB_BUDDY_PGALLOC);
1357 else
1358 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1359
1360 return ret;
1361 }
1362
1363 /*
1364 * Free huge page from pool from next node to free.
1365 * Attempt to keep persistent huge pages more or less
1366 * balanced over allowed nodes.
1367 * Called with hugetlb_lock locked.
1368 */
1369 static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1370 bool acct_surplus)
1371 {
1372 int nr_nodes, node;
1373 int ret = 0;
1374
1375 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1376 /*
1377 * If we're returning unused surplus pages, only examine
1378 * nodes with surplus pages.
1379 */
1380 if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
1381 !list_empty(&h->hugepage_freelists[node])) {
1382 struct page *page =
1383 list_entry(h->hugepage_freelists[node].next,
1384 struct page, lru);
1385 list_del(&page->lru);
1386 h->free_huge_pages--;
1387 h->free_huge_pages_node[node]--;
1388 if (acct_surplus) {
1389 h->surplus_huge_pages--;
1390 h->surplus_huge_pages_node[node]--;
1391 }
1392 update_and_free_page(h, page);
1393 ret = 1;
1394 break;
1395 }
1396 }
1397
1398 return ret;
1399 }
1400
1401 /*
1402 * Dissolve a given free hugepage into free buddy pages. This function does
1403 * nothing for in-use (including surplus) hugepages.
1404 */
1405 static void dissolve_free_huge_page(struct page *page)
1406 {
1407 spin_lock(&hugetlb_lock);
1408 if (PageHuge(page) && !page_count(page)) {
1409 struct hstate *h = page_hstate(page);
1410 int nid = page_to_nid(page);
1411 list_del(&page->lru);
1412 h->free_huge_pages--;
1413 h->free_huge_pages_node[nid]--;
1414 update_and_free_page(h, page);
1415 }
1416 spin_unlock(&hugetlb_lock);
1417 }
1418
1419 /*
1420 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1421 * make specified memory blocks removable from the system.
1422 * Note that start_pfn should aligned with (minimum) hugepage size.
1423 */
1424 void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1425 {
1426 unsigned long pfn;
1427
1428 if (!hugepages_supported())
1429 return;
1430
1431 VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order));
1432 for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order)
1433 dissolve_free_huge_page(pfn_to_page(pfn));
1434 }
1435
1436 /*
1437 * There are 3 ways this can get called:
1438 * 1. With vma+addr: we use the VMA's memory policy
1439 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1440 * page from any node, and let the buddy allocator itself figure
1441 * it out.
1442 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1443 * strictly from 'nid'
1444 */
1445 static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h,
1446 struct vm_area_struct *vma, unsigned long addr, int nid)
1447 {
1448 int order = huge_page_order(h);
1449 gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN;
1450 unsigned int cpuset_mems_cookie;
1451
1452 /*
1453 * We need a VMA to get a memory policy. If we do not
1454 * have one, we use the 'nid' argument.
1455 *
1456 * The mempolicy stuff below has some non-inlined bits
1457 * and calls ->vm_ops. That makes it hard to optimize at
1458 * compile-time, even when NUMA is off and it does
1459 * nothing. This helps the compiler optimize it out.
1460 */
1461 if (!IS_ENABLED(CONFIG_NUMA) || !vma) {
1462 /*
1463 * If a specific node is requested, make sure to
1464 * get memory from there, but only when a node
1465 * is explicitly specified.
1466 */
1467 if (nid != NUMA_NO_NODE)
1468 gfp |= __GFP_THISNODE;
1469 /*
1470 * Make sure to call something that can handle
1471 * nid=NUMA_NO_NODE
1472 */
1473 return alloc_pages_node(nid, gfp, order);
1474 }
1475
1476 /*
1477 * OK, so we have a VMA. Fetch the mempolicy and try to
1478 * allocate a huge page with it. We will only reach this
1479 * when CONFIG_NUMA=y.
1480 */
1481 do {
1482 struct page *page;
1483 struct mempolicy *mpol;
1484 struct zonelist *zl;
1485 nodemask_t *nodemask;
1486
1487 cpuset_mems_cookie = read_mems_allowed_begin();
1488 zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask);
1489 mpol_cond_put(mpol);
1490 page = __alloc_pages_nodemask(gfp, order, zl, nodemask);
1491 if (page)
1492 return page;
1493 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1494
1495 return NULL;
1496 }
1497
1498 /*
1499 * There are two ways to allocate a huge page:
1500 * 1. When you have a VMA and an address (like a fault)
1501 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1502 *
1503 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1504 * this case which signifies that the allocation should be done with
1505 * respect for the VMA's memory policy.
1506 *
1507 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1508 * implies that memory policies will not be taken in to account.
1509 */
1510 static struct page *__alloc_buddy_huge_page(struct hstate *h,
1511 struct vm_area_struct *vma, unsigned long addr, int nid)
1512 {
1513 struct page *page;
1514 unsigned int r_nid;
1515
1516 if (hstate_is_gigantic(h))
1517 return NULL;
1518
1519 /*
1520 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1521 * This makes sure the caller is picking _one_ of the modes with which
1522 * we can call this function, not both.
1523 */
1524 if (vma || (addr != -1)) {
1525 VM_WARN_ON_ONCE(addr == -1);
1526 VM_WARN_ON_ONCE(nid != NUMA_NO_NODE);
1527 }
1528 /*
1529 * Assume we will successfully allocate the surplus page to
1530 * prevent racing processes from causing the surplus to exceed
1531 * overcommit
1532 *
1533 * This however introduces a different race, where a process B
1534 * tries to grow the static hugepage pool while alloc_pages() is
1535 * called by process A. B will only examine the per-node
1536 * counters in determining if surplus huge pages can be
1537 * converted to normal huge pages in adjust_pool_surplus(). A
1538 * won't be able to increment the per-node counter, until the
1539 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1540 * no more huge pages can be converted from surplus to normal
1541 * state (and doesn't try to convert again). Thus, we have a
1542 * case where a surplus huge page exists, the pool is grown, and
1543 * the surplus huge page still exists after, even though it
1544 * should just have been converted to a normal huge page. This
1545 * does not leak memory, though, as the hugepage will be freed
1546 * once it is out of use. It also does not allow the counters to
1547 * go out of whack in adjust_pool_surplus() as we don't modify
1548 * the node values until we've gotten the hugepage and only the
1549 * per-node value is checked there.
1550 */
1551 spin_lock(&hugetlb_lock);
1552 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1553 spin_unlock(&hugetlb_lock);
1554 return NULL;
1555 } else {
1556 h->nr_huge_pages++;
1557 h->surplus_huge_pages++;
1558 }
1559 spin_unlock(&hugetlb_lock);
1560
1561 page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid);
1562
1563 spin_lock(&hugetlb_lock);
1564 if (page) {
1565 INIT_LIST_HEAD(&page->lru);
1566 r_nid = page_to_nid(page);
1567 set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1568 set_hugetlb_cgroup(page, NULL);
1569 /*
1570 * We incremented the global counters already
1571 */
1572 h->nr_huge_pages_node[r_nid]++;
1573 h->surplus_huge_pages_node[r_nid]++;
1574 __count_vm_event(HTLB_BUDDY_PGALLOC);
1575 } else {
1576 h->nr_huge_pages--;
1577 h->surplus_huge_pages--;
1578 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1579 }
1580 spin_unlock(&hugetlb_lock);
1581
1582 return page;
1583 }
1584
1585 /*
1586 * Allocate a huge page from 'nid'. Note, 'nid' may be
1587 * NUMA_NO_NODE, which means that it may be allocated
1588 * anywhere.
1589 */
1590 static
1591 struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid)
1592 {
1593 unsigned long addr = -1;
1594
1595 return __alloc_buddy_huge_page(h, NULL, addr, nid);
1596 }
1597
1598 /*
1599 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1600 */
1601 static
1602 struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h,
1603 struct vm_area_struct *vma, unsigned long addr)
1604 {
1605 return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE);
1606 }
1607
1608 /*
1609 * This allocation function is useful in the context where vma is irrelevant.
1610 * E.g. soft-offlining uses this function because it only cares physical
1611 * address of error page.
1612 */
1613 struct page *alloc_huge_page_node(struct hstate *h, int nid)
1614 {
1615 struct page *page = NULL;
1616
1617 spin_lock(&hugetlb_lock);
1618 if (h->free_huge_pages - h->resv_huge_pages > 0)
1619 page = dequeue_huge_page_node(h, nid);
1620 spin_unlock(&hugetlb_lock);
1621
1622 if (!page)
1623 page = __alloc_buddy_huge_page_no_mpol(h, nid);
1624
1625 return page;
1626 }
1627
1628 /*
1629 * Increase the hugetlb pool such that it can accommodate a reservation
1630 * of size 'delta'.
1631 */
1632 static int gather_surplus_pages(struct hstate *h, int delta)
1633 {
1634 struct list_head surplus_list;
1635 struct page *page, *tmp;
1636 int ret, i;
1637 int needed, allocated;
1638 bool alloc_ok = true;
1639
1640 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1641 if (needed <= 0) {
1642 h->resv_huge_pages += delta;
1643 return 0;
1644 }
1645
1646 allocated = 0;
1647 INIT_LIST_HEAD(&surplus_list);
1648
1649 ret = -ENOMEM;
1650 retry:
1651 spin_unlock(&hugetlb_lock);
1652 for (i = 0; i < needed; i++) {
1653 page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE);
1654 if (!page) {
1655 alloc_ok = false;
1656 break;
1657 }
1658 list_add(&page->lru, &surplus_list);
1659 }
1660 allocated += i;
1661
1662 /*
1663 * After retaking hugetlb_lock, we need to recalculate 'needed'
1664 * because either resv_huge_pages or free_huge_pages may have changed.
1665 */
1666 spin_lock(&hugetlb_lock);
1667 needed = (h->resv_huge_pages + delta) -
1668 (h->free_huge_pages + allocated);
1669 if (needed > 0) {
1670 if (alloc_ok)
1671 goto retry;
1672 /*
1673 * We were not able to allocate enough pages to
1674 * satisfy the entire reservation so we free what
1675 * we've allocated so far.
1676 */
1677 goto free;
1678 }
1679 /*
1680 * The surplus_list now contains _at_least_ the number of extra pages
1681 * needed to accommodate the reservation. Add the appropriate number
1682 * of pages to the hugetlb pool and free the extras back to the buddy
1683 * allocator. Commit the entire reservation here to prevent another
1684 * process from stealing the pages as they are added to the pool but
1685 * before they are reserved.
1686 */
1687 needed += allocated;
1688 h->resv_huge_pages += delta;
1689 ret = 0;
1690
1691 /* Free the needed pages to the hugetlb pool */
1692 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1693 if ((--needed) < 0)
1694 break;
1695 /*
1696 * This page is now managed by the hugetlb allocator and has
1697 * no users -- drop the buddy allocator's reference.
1698 */
1699 put_page_testzero(page);
1700 VM_BUG_ON_PAGE(page_count(page), page);
1701 enqueue_huge_page(h, page);
1702 }
1703 free:
1704 spin_unlock(&hugetlb_lock);
1705
1706 /* Free unnecessary surplus pages to the buddy allocator */
1707 list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1708 put_page(page);
1709 spin_lock(&hugetlb_lock);
1710
1711 return ret;
1712 }
1713
1714 /*
1715 * When releasing a hugetlb pool reservation, any surplus pages that were
1716 * allocated to satisfy the reservation must be explicitly freed if they were
1717 * never used.
1718 * Called with hugetlb_lock held.
1719 */
1720 static void return_unused_surplus_pages(struct hstate *h,
1721 unsigned long unused_resv_pages)
1722 {
1723 unsigned long nr_pages;
1724
1725 /* Uncommit the reservation */
1726 h->resv_huge_pages -= unused_resv_pages;
1727
1728 /* Cannot return gigantic pages currently */
1729 if (hstate_is_gigantic(h))
1730 return;
1731
1732 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1733
1734 /*
1735 * We want to release as many surplus pages as possible, spread
1736 * evenly across all nodes with memory. Iterate across these nodes
1737 * until we can no longer free unreserved surplus pages. This occurs
1738 * when the nodes with surplus pages have no free pages.
1739 * free_pool_huge_page() will balance the the freed pages across the
1740 * on-line nodes with memory and will handle the hstate accounting.
1741 */
1742 while (nr_pages--) {
1743 if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1744 break;
1745 cond_resched_lock(&hugetlb_lock);
1746 }
1747 }
1748
1749
1750 /*
1751 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1752 * are used by the huge page allocation routines to manage reservations.
1753 *
1754 * vma_needs_reservation is called to determine if the huge page at addr
1755 * within the vma has an associated reservation. If a reservation is
1756 * needed, the value 1 is returned. The caller is then responsible for
1757 * managing the global reservation and subpool usage counts. After
1758 * the huge page has been allocated, vma_commit_reservation is called
1759 * to add the page to the reservation map. If the page allocation fails,
1760 * the reservation must be ended instead of committed. vma_end_reservation
1761 * is called in such cases.
1762 *
1763 * In the normal case, vma_commit_reservation returns the same value
1764 * as the preceding vma_needs_reservation call. The only time this
1765 * is not the case is if a reserve map was changed between calls. It
1766 * is the responsibility of the caller to notice the difference and
1767 * take appropriate action.
1768 */
1769 enum vma_resv_mode {
1770 VMA_NEEDS_RESV,
1771 VMA_COMMIT_RESV,
1772 VMA_END_RESV,
1773 };
1774 static long __vma_reservation_common(struct hstate *h,
1775 struct vm_area_struct *vma, unsigned long addr,
1776 enum vma_resv_mode mode)
1777 {
1778 struct resv_map *resv;
1779 pgoff_t idx;
1780 long ret;
1781
1782 resv = vma_resv_map(vma);
1783 if (!resv)
1784 return 1;
1785
1786 idx = vma_hugecache_offset(h, vma, addr);
1787 switch (mode) {
1788 case VMA_NEEDS_RESV:
1789 ret = region_chg(resv, idx, idx + 1);
1790 break;
1791 case VMA_COMMIT_RESV:
1792 ret = region_add(resv, idx, idx + 1);
1793 break;
1794 case VMA_END_RESV:
1795 region_abort(resv, idx, idx + 1);
1796 ret = 0;
1797 break;
1798 default:
1799 BUG();
1800 }
1801
1802 if (vma->vm_flags & VM_MAYSHARE)
1803 return ret;
1804 else
1805 return ret < 0 ? ret : 0;
1806 }
1807
1808 static long vma_needs_reservation(struct hstate *h,
1809 struct vm_area_struct *vma, unsigned long addr)
1810 {
1811 return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1812 }
1813
1814 static long vma_commit_reservation(struct hstate *h,
1815 struct vm_area_struct *vma, unsigned long addr)
1816 {
1817 return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1818 }
1819
1820 static void vma_end_reservation(struct hstate *h,
1821 struct vm_area_struct *vma, unsigned long addr)
1822 {
1823 (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1824 }
1825
1826 struct page *alloc_huge_page(struct vm_area_struct *vma,
1827 unsigned long addr, int avoid_reserve)
1828 {
1829 struct hugepage_subpool *spool = subpool_vma(vma);
1830 struct hstate *h = hstate_vma(vma);
1831 struct page *page;
1832 long map_chg, map_commit;
1833 long gbl_chg;
1834 int ret, idx;
1835 struct hugetlb_cgroup *h_cg;
1836
1837 idx = hstate_index(h);
1838 /*
1839 * Examine the region/reserve map to determine if the process
1840 * has a reservation for the page to be allocated. A return
1841 * code of zero indicates a reservation exists (no change).
1842 */
1843 map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
1844 if (map_chg < 0)
1845 return ERR_PTR(-ENOMEM);
1846
1847 /*
1848 * Processes that did not create the mapping will have no
1849 * reserves as indicated by the region/reserve map. Check
1850 * that the allocation will not exceed the subpool limit.
1851 * Allocations for MAP_NORESERVE mappings also need to be
1852 * checked against any subpool limit.
1853 */
1854 if (map_chg || avoid_reserve) {
1855 gbl_chg = hugepage_subpool_get_pages(spool, 1);
1856 if (gbl_chg < 0) {
1857 vma_end_reservation(h, vma, addr);
1858 return ERR_PTR(-ENOSPC);
1859 }
1860
1861 /*
1862 * Even though there was no reservation in the region/reserve
1863 * map, there could be reservations associated with the
1864 * subpool that can be used. This would be indicated if the
1865 * return value of hugepage_subpool_get_pages() is zero.
1866 * However, if avoid_reserve is specified we still avoid even
1867 * the subpool reservations.
1868 */
1869 if (avoid_reserve)
1870 gbl_chg = 1;
1871 }
1872
1873 ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
1874 if (ret)
1875 goto out_subpool_put;
1876
1877 spin_lock(&hugetlb_lock);
1878 /*
1879 * glb_chg is passed to indicate whether or not a page must be taken
1880 * from the global free pool (global change). gbl_chg == 0 indicates
1881 * a reservation exists for the allocation.
1882 */
1883 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
1884 if (!page) {
1885 spin_unlock(&hugetlb_lock);
1886 page = __alloc_buddy_huge_page_with_mpol(h, vma, addr);
1887 if (!page)
1888 goto out_uncharge_cgroup;
1889
1890 spin_lock(&hugetlb_lock);
1891 list_move(&page->lru, &h->hugepage_activelist);
1892 /* Fall through */
1893 }
1894 hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
1895 spin_unlock(&hugetlb_lock);
1896
1897 set_page_private(page, (unsigned long)spool);
1898
1899 map_commit = vma_commit_reservation(h, vma, addr);
1900 if (unlikely(map_chg > map_commit)) {
1901 /*
1902 * The page was added to the reservation map between
1903 * vma_needs_reservation and vma_commit_reservation.
1904 * This indicates a race with hugetlb_reserve_pages.
1905 * Adjust for the subpool count incremented above AND
1906 * in hugetlb_reserve_pages for the same page. Also,
1907 * the reservation count added in hugetlb_reserve_pages
1908 * no longer applies.
1909 */
1910 long rsv_adjust;
1911
1912 rsv_adjust = hugepage_subpool_put_pages(spool, 1);
1913 hugetlb_acct_memory(h, -rsv_adjust);
1914 }
1915 return page;
1916
1917 out_uncharge_cgroup:
1918 hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
1919 out_subpool_put:
1920 if (map_chg || avoid_reserve)
1921 hugepage_subpool_put_pages(spool, 1);
1922 vma_end_reservation(h, vma, addr);
1923 return ERR_PTR(-ENOSPC);
1924 }
1925
1926 /*
1927 * alloc_huge_page()'s wrapper which simply returns the page if allocation
1928 * succeeds, otherwise NULL. This function is called from new_vma_page(),
1929 * where no ERR_VALUE is expected to be returned.
1930 */
1931 struct page *alloc_huge_page_noerr(struct vm_area_struct *vma,
1932 unsigned long addr, int avoid_reserve)
1933 {
1934 struct page *page = alloc_huge_page(vma, addr, avoid_reserve);
1935 if (IS_ERR(page))
1936 page = NULL;
1937 return page;
1938 }
1939
1940 int __weak alloc_bootmem_huge_page(struct hstate *h)
1941 {
1942 struct huge_bootmem_page *m;
1943 int nr_nodes, node;
1944
1945 for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
1946 void *addr;
1947
1948 addr = memblock_virt_alloc_try_nid_nopanic(
1949 huge_page_size(h), huge_page_size(h),
1950 0, BOOTMEM_ALLOC_ACCESSIBLE, node);
1951 if (addr) {
1952 /*
1953 * Use the beginning of the huge page to store the
1954 * huge_bootmem_page struct (until gather_bootmem
1955 * puts them into the mem_map).
1956 */
1957 m = addr;
1958 goto found;
1959 }
1960 }
1961 return 0;
1962
1963 found:
1964 BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
1965 /* Put them into a private list first because mem_map is not up yet */
1966 list_add(&m->list, &huge_boot_pages);
1967 m->hstate = h;
1968 return 1;
1969 }
1970
1971 static void __init prep_compound_huge_page(struct page *page,
1972 unsigned int order)
1973 {
1974 if (unlikely(order > (MAX_ORDER - 1)))
1975 prep_compound_gigantic_page(page, order);
1976 else
1977 prep_compound_page(page, order);
1978 }
1979
1980 /* Put bootmem huge pages into the standard lists after mem_map is up */
1981 static void __init gather_bootmem_prealloc(void)
1982 {
1983 struct huge_bootmem_page *m;
1984
1985 list_for_each_entry(m, &huge_boot_pages, list) {
1986 struct hstate *h = m->hstate;
1987 struct page *page;
1988
1989 #ifdef CONFIG_HIGHMEM
1990 page = pfn_to_page(m->phys >> PAGE_SHIFT);
1991 memblock_free_late(__pa(m),
1992 sizeof(struct huge_bootmem_page));
1993 #else
1994 page = virt_to_page(m);
1995 #endif
1996 WARN_ON(page_count(page) != 1);
1997 prep_compound_huge_page(page, h->order);
1998 WARN_ON(PageReserved(page));
1999 prep_new_huge_page(h, page, page_to_nid(page));
2000 /*
2001 * If we had gigantic hugepages allocated at boot time, we need
2002 * to restore the 'stolen' pages to totalram_pages in order to
2003 * fix confusing memory reports from free(1) and another
2004 * side-effects, like CommitLimit going negative.
2005 */
2006 if (hstate_is_gigantic(h))
2007 adjust_managed_page_count(page, 1 << h->order);
2008 }
2009 }
2010
2011 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2012 {
2013 unsigned long i;
2014
2015 for (i = 0; i < h->max_huge_pages; ++i) {
2016 if (hstate_is_gigantic(h)) {
2017 if (!alloc_bootmem_huge_page(h))
2018 break;
2019 } else if (!alloc_fresh_huge_page(h,
2020 &node_states[N_MEMORY]))
2021 break;
2022 }
2023 h->max_huge_pages = i;
2024 }
2025
2026 static void __init hugetlb_init_hstates(void)
2027 {
2028 struct hstate *h;
2029
2030 for_each_hstate(h) {
2031 if (minimum_order > huge_page_order(h))
2032 minimum_order = huge_page_order(h);
2033
2034 /* oversize hugepages were init'ed in early boot */
2035 if (!hstate_is_gigantic(h))
2036 hugetlb_hstate_alloc_pages(h);
2037 }
2038 VM_BUG_ON(minimum_order == UINT_MAX);
2039 }
2040
2041 static char * __init memfmt(char *buf, unsigned long n)
2042 {
2043 if (n >= (1UL << 30))
2044 sprintf(buf, "%lu GB", n >> 30);
2045 else if (n >= (1UL << 20))
2046 sprintf(buf, "%lu MB", n >> 20);
2047 else
2048 sprintf(buf, "%lu KB", n >> 10);
2049 return buf;
2050 }
2051
2052 static void __init report_hugepages(void)
2053 {
2054 struct hstate *h;
2055
2056 for_each_hstate(h) {
2057 char buf[32];
2058 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2059 memfmt(buf, huge_page_size(h)),
2060 h->free_huge_pages);
2061 }
2062 }
2063
2064 #ifdef CONFIG_HIGHMEM
2065 static void try_to_free_low(struct hstate *h, unsigned long count,
2066 nodemask_t *nodes_allowed)
2067 {
2068 int i;
2069
2070 if (hstate_is_gigantic(h))
2071 return;
2072
2073 for_each_node_mask(i, *nodes_allowed) {
2074 struct page *page, *next;
2075 struct list_head *freel = &h->hugepage_freelists[i];
2076 list_for_each_entry_safe(page, next, freel, lru) {
2077 if (count >= h->nr_huge_pages)
2078 return;
2079 if (PageHighMem(page))
2080 continue;
2081 list_del(&page->lru);
2082 update_and_free_page(h, page);
2083 h->free_huge_pages--;
2084 h->free_huge_pages_node[page_to_nid(page)]--;
2085 }
2086 }
2087 }
2088 #else
2089 static inline void try_to_free_low(struct hstate *h, unsigned long count,
2090 nodemask_t *nodes_allowed)
2091 {
2092 }
2093 #endif
2094
2095 /*
2096 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2097 * balanced by operating on them in a round-robin fashion.
2098 * Returns 1 if an adjustment was made.
2099 */
2100 static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2101 int delta)
2102 {
2103 int nr_nodes, node;
2104
2105 VM_BUG_ON(delta != -1 && delta != 1);
2106
2107 if (delta < 0) {
2108 for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2109 if (h->surplus_huge_pages_node[node])
2110 goto found;
2111 }
2112 } else {
2113 for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2114 if (h->surplus_huge_pages_node[node] <
2115 h->nr_huge_pages_node[node])
2116 goto found;
2117 }
2118 }
2119 return 0;
2120
2121 found:
2122 h->surplus_huge_pages += delta;
2123 h->surplus_huge_pages_node[node] += delta;
2124 return 1;
2125 }
2126
2127 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2128 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
2129 nodemask_t *nodes_allowed)
2130 {
2131 unsigned long min_count, ret;
2132
2133 if (hstate_is_gigantic(h) && !gigantic_page_supported())
2134 return h->max_huge_pages;
2135
2136 /*
2137 * Increase the pool size
2138 * First take pages out of surplus state. Then make up the
2139 * remaining difference by allocating fresh huge pages.
2140 *
2141 * We might race with __alloc_buddy_huge_page() here and be unable
2142 * to convert a surplus huge page to a normal huge page. That is
2143 * not critical, though, it just means the overall size of the
2144 * pool might be one hugepage larger than it needs to be, but
2145 * within all the constraints specified by the sysctls.
2146 */
2147 spin_lock(&hugetlb_lock);
2148 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2149 if (!adjust_pool_surplus(h, nodes_allowed, -1))
2150 break;
2151 }
2152
2153 while (count > persistent_huge_pages(h)) {
2154 /*
2155 * If this allocation races such that we no longer need the
2156 * page, free_huge_page will handle it by freeing the page
2157 * and reducing the surplus.
2158 */
2159 spin_unlock(&hugetlb_lock);
2160 if (hstate_is_gigantic(h))
2161 ret = alloc_fresh_gigantic_page(h, nodes_allowed);
2162 else
2163 ret = alloc_fresh_huge_page(h, nodes_allowed);
2164 spin_lock(&hugetlb_lock);
2165 if (!ret)
2166 goto out;
2167
2168 /* Bail for signals. Probably ctrl-c from user */
2169 if (signal_pending(current))
2170 goto out;
2171 }
2172
2173 /*
2174 * Decrease the pool size
2175 * First return free pages to the buddy allocator (being careful
2176 * to keep enough around to satisfy reservations). Then place
2177 * pages into surplus state as needed so the pool will shrink
2178 * to the desired size as pages become free.
2179 *
2180 * By placing pages into the surplus state independent of the
2181 * overcommit value, we are allowing the surplus pool size to
2182 * exceed overcommit. There are few sane options here. Since
2183 * __alloc_buddy_huge_page() is checking the global counter,
2184 * though, we'll note that we're not allowed to exceed surplus
2185 * and won't grow the pool anywhere else. Not until one of the
2186 * sysctls are changed, or the surplus pages go out of use.
2187 */
2188 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2189 min_count = max(count, min_count);
2190 try_to_free_low(h, min_count, nodes_allowed);
2191 while (min_count < persistent_huge_pages(h)) {
2192 if (!free_pool_huge_page(h, nodes_allowed, 0))
2193 break;
2194 cond_resched_lock(&hugetlb_lock);
2195 }
2196 while (count < persistent_huge_pages(h)) {
2197 if (!adjust_pool_surplus(h, nodes_allowed, 1))
2198 break;
2199 }
2200 out:
2201 ret = persistent_huge_pages(h);
2202 spin_unlock(&hugetlb_lock);
2203 return ret;
2204 }
2205
2206 #define HSTATE_ATTR_RO(_name) \
2207 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2208
2209 #define HSTATE_ATTR(_name) \
2210 static struct kobj_attribute _name##_attr = \
2211 __ATTR(_name, 0644, _name##_show, _name##_store)
2212
2213 static struct kobject *hugepages_kobj;
2214 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2215
2216 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2217
2218 static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2219 {
2220 int i;
2221
2222 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2223 if (hstate_kobjs[i] == kobj) {
2224 if (nidp)
2225 *nidp = NUMA_NO_NODE;
2226 return &hstates[i];
2227 }
2228
2229 return kobj_to_node_hstate(kobj, nidp);
2230 }
2231
2232 static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2233 struct kobj_attribute *attr, char *buf)
2234 {
2235 struct hstate *h;
2236 unsigned long nr_huge_pages;
2237 int nid;
2238
2239 h = kobj_to_hstate(kobj, &nid);
2240 if (nid == NUMA_NO_NODE)
2241 nr_huge_pages = h->nr_huge_pages;
2242 else
2243 nr_huge_pages = h->nr_huge_pages_node[nid];
2244
2245 return sprintf(buf, "%lu\n", nr_huge_pages);
2246 }
2247
2248 static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2249 struct hstate *h, int nid,
2250 unsigned long count, size_t len)
2251 {
2252 int err;
2253 NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
2254
2255 if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
2256 err = -EINVAL;
2257 goto out;
2258 }
2259
2260 if (nid == NUMA_NO_NODE) {
2261 /*
2262 * global hstate attribute
2263 */
2264 if (!(obey_mempolicy &&
2265 init_nodemask_of_mempolicy(nodes_allowed))) {
2266 NODEMASK_FREE(nodes_allowed);
2267 nodes_allowed = &node_states[N_MEMORY];
2268 }
2269 } else if (nodes_allowed) {
2270 /*
2271 * per node hstate attribute: adjust count to global,
2272 * but restrict alloc/free to the specified node.
2273 */
2274 count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2275 init_nodemask_of_node(nodes_allowed, nid);
2276 } else
2277 nodes_allowed = &node_states[N_MEMORY];
2278
2279 h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
2280
2281 if (nodes_allowed != &node_states[N_MEMORY])
2282 NODEMASK_FREE(nodes_allowed);
2283
2284 return len;
2285 out:
2286 NODEMASK_FREE(nodes_allowed);
2287 return err;
2288 }
2289
2290 static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2291 struct kobject *kobj, const char *buf,
2292 size_t len)
2293 {
2294 struct hstate *h;
2295 unsigned long count;
2296 int nid;
2297 int err;
2298
2299 err = kstrtoul(buf, 10, &count);
2300 if (err)
2301 return err;
2302
2303 h = kobj_to_hstate(kobj, &nid);
2304 return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2305 }
2306
2307 static ssize_t nr_hugepages_show(struct kobject *kobj,
2308 struct kobj_attribute *attr, char *buf)
2309 {
2310 return nr_hugepages_show_common(kobj, attr, buf);
2311 }
2312
2313 static ssize_t nr_hugepages_store(struct kobject *kobj,
2314 struct kobj_attribute *attr, const char *buf, size_t len)
2315 {
2316 return nr_hugepages_store_common(false, kobj, buf, len);
2317 }
2318 HSTATE_ATTR(nr_hugepages);
2319
2320 #ifdef CONFIG_NUMA
2321
2322 /*
2323 * hstate attribute for optionally mempolicy-based constraint on persistent
2324 * huge page alloc/free.
2325 */
2326 static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2327 struct kobj_attribute *attr, char *buf)
2328 {
2329 return nr_hugepages_show_common(kobj, attr, buf);
2330 }
2331
2332 static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2333 struct kobj_attribute *attr, const char *buf, size_t len)
2334 {
2335 return nr_hugepages_store_common(true, kobj, buf, len);
2336 }
2337 HSTATE_ATTR(nr_hugepages_mempolicy);
2338 #endif
2339
2340
2341 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2342 struct kobj_attribute *attr, char *buf)
2343 {
2344 struct hstate *h = kobj_to_hstate(kobj, NULL);
2345 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2346 }
2347
2348 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2349 struct kobj_attribute *attr, const char *buf, size_t count)
2350 {
2351 int err;
2352 unsigned long input;
2353 struct hstate *h = kobj_to_hstate(kobj, NULL);
2354
2355 if (hstate_is_gigantic(h))
2356 return -EINVAL;
2357
2358 err = kstrtoul(buf, 10, &input);
2359 if (err)
2360 return err;
2361
2362 spin_lock(&hugetlb_lock);
2363 h->nr_overcommit_huge_pages = input;
2364 spin_unlock(&hugetlb_lock);
2365
2366 return count;
2367 }
2368 HSTATE_ATTR(nr_overcommit_hugepages);
2369
2370 static ssize_t free_hugepages_show(struct kobject *kobj,
2371 struct kobj_attribute *attr, char *buf)
2372 {
2373 struct hstate *h;
2374 unsigned long free_huge_pages;
2375 int nid;
2376
2377 h = kobj_to_hstate(kobj, &nid);
2378 if (nid == NUMA_NO_NODE)
2379 free_huge_pages = h->free_huge_pages;
2380 else
2381 free_huge_pages = h->free_huge_pages_node[nid];
2382
2383 return sprintf(buf, "%lu\n", free_huge_pages);
2384 }
2385 HSTATE_ATTR_RO(free_hugepages);
2386
2387 static ssize_t resv_hugepages_show(struct kobject *kobj,
2388 struct kobj_attribute *attr, char *buf)
2389 {
2390 struct hstate *h = kobj_to_hstate(kobj, NULL);
2391 return sprintf(buf, "%lu\n", h->resv_huge_pages);
2392 }
2393 HSTATE_ATTR_RO(resv_hugepages);
2394
2395 static ssize_t surplus_hugepages_show(struct kobject *kobj,
2396 struct kobj_attribute *attr, char *buf)
2397 {
2398 struct hstate *h;
2399 unsigned long surplus_huge_pages;
2400 int nid;
2401
2402 h = kobj_to_hstate(kobj, &nid);
2403 if (nid == NUMA_NO_NODE)
2404 surplus_huge_pages = h->surplus_huge_pages;
2405 else
2406 surplus_huge_pages = h->surplus_huge_pages_node[nid];
2407
2408 return sprintf(buf, "%lu\n", surplus_huge_pages);
2409 }
2410 HSTATE_ATTR_RO(surplus_hugepages);
2411
2412 static struct attribute *hstate_attrs[] = {
2413 &nr_hugepages_attr.attr,
2414 &nr_overcommit_hugepages_attr.attr,
2415 &free_hugepages_attr.attr,
2416 &resv_hugepages_attr.attr,
2417 &surplus_hugepages_attr.attr,
2418 #ifdef CONFIG_NUMA
2419 &nr_hugepages_mempolicy_attr.attr,
2420 #endif
2421 NULL,
2422 };
2423
2424 static struct attribute_group hstate_attr_group = {
2425 .attrs = hstate_attrs,
2426 };
2427
2428 static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2429 struct kobject **hstate_kobjs,
2430 struct attribute_group *hstate_attr_group)
2431 {
2432 int retval;
2433 int hi = hstate_index(h);
2434
2435 hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2436 if (!hstate_kobjs[hi])
2437 return -ENOMEM;
2438
2439 retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2440 if (retval)
2441 kobject_put(hstate_kobjs[hi]);
2442
2443 return retval;
2444 }
2445
2446 static void __init hugetlb_sysfs_init(void)
2447 {
2448 struct hstate *h;
2449 int err;
2450
2451 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2452 if (!hugepages_kobj)
2453 return;
2454
2455 for_each_hstate(h) {
2456 err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2457 hstate_kobjs, &hstate_attr_group);
2458 if (err)
2459 pr_err("Hugetlb: Unable to add hstate %s", h->name);
2460 }
2461 }
2462
2463 #ifdef CONFIG_NUMA
2464
2465 /*
2466 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2467 * with node devices in node_devices[] using a parallel array. The array
2468 * index of a node device or _hstate == node id.
2469 * This is here to avoid any static dependency of the node device driver, in
2470 * the base kernel, on the hugetlb module.
2471 */
2472 struct node_hstate {
2473 struct kobject *hugepages_kobj;
2474 struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2475 };
2476 static struct node_hstate node_hstates[MAX_NUMNODES];
2477
2478 /*
2479 * A subset of global hstate attributes for node devices
2480 */
2481 static struct attribute *per_node_hstate_attrs[] = {
2482 &nr_hugepages_attr.attr,
2483 &free_hugepages_attr.attr,
2484 &surplus_hugepages_attr.attr,
2485 NULL,
2486 };
2487
2488 static struct attribute_group per_node_hstate_attr_group = {
2489 .attrs = per_node_hstate_attrs,
2490 };
2491
2492 /*
2493 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2494 * Returns node id via non-NULL nidp.
2495 */
2496 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2497 {
2498 int nid;
2499
2500 for (nid = 0; nid < nr_node_ids; nid++) {
2501 struct node_hstate *nhs = &node_hstates[nid];
2502 int i;
2503 for (i = 0; i < HUGE_MAX_HSTATE; i++)
2504 if (nhs->hstate_kobjs[i] == kobj) {
2505 if (nidp)
2506 *nidp = nid;
2507 return &hstates[i];
2508 }
2509 }
2510
2511 BUG();
2512 return NULL;
2513 }
2514
2515 /*
2516 * Unregister hstate attributes from a single node device.
2517 * No-op if no hstate attributes attached.
2518 */
2519 static void hugetlb_unregister_node(struct node *node)
2520 {
2521 struct hstate *h;
2522 struct node_hstate *nhs = &node_hstates[node->dev.id];
2523
2524 if (!nhs->hugepages_kobj)
2525 return; /* no hstate attributes */
2526
2527 for_each_hstate(h) {
2528 int idx = hstate_index(h);
2529 if (nhs->hstate_kobjs[idx]) {
2530 kobject_put(nhs->hstate_kobjs[idx]);
2531 nhs->hstate_kobjs[idx] = NULL;
2532 }
2533 }
2534
2535 kobject_put(nhs->hugepages_kobj);
2536 nhs->hugepages_kobj = NULL;
2537 }
2538
2539 /*
2540 * hugetlb module exit: unregister hstate attributes from node devices
2541 * that have them.
2542 */
2543 static void hugetlb_unregister_all_nodes(void)
2544 {
2545 int nid;
2546
2547 /*
2548 * disable node device registrations.
2549 */
2550 register_hugetlbfs_with_node(NULL, NULL);
2551
2552 /*
2553 * remove hstate attributes from any nodes that have them.
2554 */
2555 for (nid = 0; nid < nr_node_ids; nid++)
2556 hugetlb_unregister_node(node_devices[nid]);
2557 }
2558
2559 /*
2560 * Register hstate attributes for a single node device.
2561 * No-op if attributes already registered.
2562 */
2563 static void hugetlb_register_node(struct node *node)
2564 {
2565 struct hstate *h;
2566 struct node_hstate *nhs = &node_hstates[node->dev.id];
2567 int err;
2568
2569 if (nhs->hugepages_kobj)
2570 return; /* already allocated */
2571
2572 nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2573 &node->dev.kobj);
2574 if (!nhs->hugepages_kobj)
2575 return;
2576
2577 for_each_hstate(h) {
2578 err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2579 nhs->hstate_kobjs,
2580 &per_node_hstate_attr_group);
2581 if (err) {
2582 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2583 h->name, node->dev.id);
2584 hugetlb_unregister_node(node);
2585 break;
2586 }
2587 }
2588 }
2589
2590 /*
2591 * hugetlb init time: register hstate attributes for all registered node
2592 * devices of nodes that have memory. All on-line nodes should have
2593 * registered their associated device by this time.
2594 */
2595 static void __init hugetlb_register_all_nodes(void)
2596 {
2597 int nid;
2598
2599 for_each_node_state(nid, N_MEMORY) {
2600 struct node *node = node_devices[nid];
2601 if (node->dev.id == nid)
2602 hugetlb_register_node(node);
2603 }
2604
2605 /*
2606 * Let the node device driver know we're here so it can
2607 * [un]register hstate attributes on node hotplug.
2608 */
2609 register_hugetlbfs_with_node(hugetlb_register_node,
2610 hugetlb_unregister_node);
2611 }
2612 #else /* !CONFIG_NUMA */
2613
2614 static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2615 {
2616 BUG();
2617 if (nidp)
2618 *nidp = -1;
2619 return NULL;
2620 }
2621
2622 static void hugetlb_unregister_all_nodes(void) { }
2623
2624 static void hugetlb_register_all_nodes(void) { }
2625
2626 #endif
2627
2628 static void __exit hugetlb_exit(void)
2629 {
2630 struct hstate *h;
2631
2632 hugetlb_unregister_all_nodes();
2633
2634 for_each_hstate(h) {
2635 kobject_put(hstate_kobjs[hstate_index(h)]);
2636 }
2637
2638 kobject_put(hugepages_kobj);
2639 kfree(hugetlb_fault_mutex_table);
2640 }
2641 module_exit(hugetlb_exit);
2642
2643 static int __init hugetlb_init(void)
2644 {
2645 int i;
2646
2647 if (!hugepages_supported())
2648 return 0;
2649
2650 if (!size_to_hstate(default_hstate_size)) {
2651 default_hstate_size = HPAGE_SIZE;
2652 if (!size_to_hstate(default_hstate_size))
2653 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2654 }
2655 default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2656 if (default_hstate_max_huge_pages)
2657 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2658
2659 hugetlb_init_hstates();
2660 gather_bootmem_prealloc();
2661 report_hugepages();
2662
2663 hugetlb_sysfs_init();
2664 hugetlb_register_all_nodes();
2665 hugetlb_cgroup_file_init();
2666
2667 #ifdef CONFIG_SMP
2668 num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2669 #else
2670 num_fault_mutexes = 1;
2671 #endif
2672 hugetlb_fault_mutex_table =
2673 kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL);
2674 BUG_ON(!hugetlb_fault_mutex_table);
2675
2676 for (i = 0; i < num_fault_mutexes; i++)
2677 mutex_init(&hugetlb_fault_mutex_table[i]);
2678 return 0;
2679 }
2680 module_init(hugetlb_init);
2681
2682 /* Should be called on processing a hugepagesz=... option */
2683 void __init hugetlb_add_hstate(unsigned int order)
2684 {
2685 struct hstate *h;
2686 unsigned long i;
2687
2688 if (size_to_hstate(PAGE_SIZE << order)) {
2689 pr_warning("hugepagesz= specified twice, ignoring\n");
2690 return;
2691 }
2692 BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2693 BUG_ON(order == 0);
2694 h = &hstates[hugetlb_max_hstate++];
2695 h->order = order;
2696 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2697 h->nr_huge_pages = 0;
2698 h->free_huge_pages = 0;
2699 for (i = 0; i < MAX_NUMNODES; ++i)
2700 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2701 INIT_LIST_HEAD(&h->hugepage_activelist);
2702 h->next_nid_to_alloc = first_node(node_states[N_MEMORY]);
2703 h->next_nid_to_free = first_node(node_states[N_MEMORY]);
2704 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2705 huge_page_size(h)/1024);
2706
2707 parsed_hstate = h;
2708 }
2709
2710 static int __init hugetlb_nrpages_setup(char *s)
2711 {
2712 unsigned long *mhp;
2713 static unsigned long *last_mhp;
2714
2715 /*
2716 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2717 * so this hugepages= parameter goes to the "default hstate".
2718 */
2719 if (!hugetlb_max_hstate)
2720 mhp = &default_hstate_max_huge_pages;
2721 else
2722 mhp = &parsed_hstate->max_huge_pages;
2723
2724 if (mhp == last_mhp) {
2725 pr_warning("hugepages= specified twice without "
2726 "interleaving hugepagesz=, ignoring\n");
2727 return 1;
2728 }
2729
2730 if (sscanf(s, "%lu", mhp) <= 0)
2731 *mhp = 0;
2732
2733 /*
2734 * Global state is always initialized later in hugetlb_init.
2735 * But we need to allocate >= MAX_ORDER hstates here early to still
2736 * use the bootmem allocator.
2737 */
2738 if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2739 hugetlb_hstate_alloc_pages(parsed_hstate);
2740
2741 last_mhp = mhp;
2742
2743 return 1;
2744 }
2745 __setup("hugepages=", hugetlb_nrpages_setup);
2746
2747 static int __init hugetlb_default_setup(char *s)
2748 {
2749 default_hstate_size = memparse(s, &s);
2750 return 1;
2751 }
2752 __setup("default_hugepagesz=", hugetlb_default_setup);
2753
2754 static unsigned int cpuset_mems_nr(unsigned int *array)
2755 {
2756 int node;
2757 unsigned int nr = 0;
2758
2759 for_each_node_mask(node, cpuset_current_mems_allowed)
2760 nr += array[node];
2761
2762 return nr;
2763 }
2764
2765 #ifdef CONFIG_SYSCTL
2766 static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2767 struct ctl_table *table, int write,
2768 void __user *buffer, size_t *length, loff_t *ppos)
2769 {
2770 struct hstate *h = &default_hstate;
2771 unsigned long tmp = h->max_huge_pages;
2772 int ret;
2773
2774 if (!hugepages_supported())
2775 return -ENOTSUPP;
2776
2777 table->data = &tmp;
2778 table->maxlen = sizeof(unsigned long);
2779 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2780 if (ret)
2781 goto out;
2782
2783 if (write)
2784 ret = __nr_hugepages_store_common(obey_mempolicy, h,
2785 NUMA_NO_NODE, tmp, *length);
2786 out:
2787 return ret;
2788 }
2789
2790 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2791 void __user *buffer, size_t *length, loff_t *ppos)
2792 {
2793
2794 return hugetlb_sysctl_handler_common(false, table, write,
2795 buffer, length, ppos);
2796 }
2797
2798 #ifdef CONFIG_NUMA
2799 int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
2800 void __user *buffer, size_t *length, loff_t *ppos)
2801 {
2802 return hugetlb_sysctl_handler_common(true, table, write,
2803 buffer, length, ppos);
2804 }
2805 #endif /* CONFIG_NUMA */
2806
2807 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
2808 void __user *buffer,
2809 size_t *length, loff_t *ppos)
2810 {
2811 struct hstate *h = &default_hstate;
2812 unsigned long tmp;
2813 int ret;
2814
2815 if (!hugepages_supported())
2816 return -ENOTSUPP;
2817
2818 tmp = h->nr_overcommit_huge_pages;
2819
2820 if (write && hstate_is_gigantic(h))
2821 return -EINVAL;
2822
2823 table->data = &tmp;
2824 table->maxlen = sizeof(unsigned long);
2825 ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2826 if (ret)
2827 goto out;
2828
2829 if (write) {
2830 spin_lock(&hugetlb_lock);
2831 h->nr_overcommit_huge_pages = tmp;
2832 spin_unlock(&hugetlb_lock);
2833 }
2834 out:
2835 return ret;
2836 }
2837
2838 #endif /* CONFIG_SYSCTL */
2839
2840 void hugetlb_report_meminfo(struct seq_file *m)
2841 {
2842 struct hstate *h = &default_hstate;
2843 if (!hugepages_supported())
2844 return;
2845 seq_printf(m,
2846 "HugePages_Total: %5lu\n"
2847 "HugePages_Free: %5lu\n"
2848 "HugePages_Rsvd: %5lu\n"
2849 "HugePages_Surp: %5lu\n"
2850 "Hugepagesize: %8lu kB\n",
2851 h->nr_huge_pages,
2852 h->free_huge_pages,
2853 h->resv_huge_pages,
2854 h->surplus_huge_pages,
2855 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2856 }
2857
2858 int hugetlb_report_node_meminfo(int nid, char *buf)
2859 {
2860 struct hstate *h = &default_hstate;
2861 if (!hugepages_supported())
2862 return 0;
2863 return sprintf(buf,
2864 "Node %d HugePages_Total: %5u\n"
2865 "Node %d HugePages_Free: %5u\n"
2866 "Node %d HugePages_Surp: %5u\n",
2867 nid, h->nr_huge_pages_node[nid],
2868 nid, h->free_huge_pages_node[nid],
2869 nid, h->surplus_huge_pages_node[nid]);
2870 }
2871
2872 void hugetlb_show_meminfo(void)
2873 {
2874 struct hstate *h;
2875 int nid;
2876
2877 if (!hugepages_supported())
2878 return;
2879
2880 for_each_node_state(nid, N_MEMORY)
2881 for_each_hstate(h)
2882 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
2883 nid,
2884 h->nr_huge_pages_node[nid],
2885 h->free_huge_pages_node[nid],
2886 h->surplus_huge_pages_node[nid],
2887 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
2888 }
2889
2890 void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
2891 {
2892 seq_printf(m, "HugetlbPages:\t%8lu kB\n",
2893 atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
2894 }
2895
2896 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
2897 unsigned long hugetlb_total_pages(void)
2898 {
2899 struct hstate *h;
2900 unsigned long nr_total_pages = 0;
2901
2902 for_each_hstate(h)
2903 nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
2904 return nr_total_pages;
2905 }
2906
2907 static int hugetlb_acct_memory(struct hstate *h, long delta)
2908 {
2909 int ret = -ENOMEM;
2910
2911 spin_lock(&hugetlb_lock);
2912 /*
2913 * When cpuset is configured, it breaks the strict hugetlb page
2914 * reservation as the accounting is done on a global variable. Such
2915 * reservation is completely rubbish in the presence of cpuset because
2916 * the reservation is not checked against page availability for the
2917 * current cpuset. Application can still potentially OOM'ed by kernel
2918 * with lack of free htlb page in cpuset that the task is in.
2919 * Attempt to enforce strict accounting with cpuset is almost
2920 * impossible (or too ugly) because cpuset is too fluid that
2921 * task or memory node can be dynamically moved between cpusets.
2922 *
2923 * The change of semantics for shared hugetlb mapping with cpuset is
2924 * undesirable. However, in order to preserve some of the semantics,
2925 * we fall back to check against current free page availability as
2926 * a best attempt and hopefully to minimize the impact of changing
2927 * semantics that cpuset has.
2928 */
2929 if (delta > 0) {
2930 if (gather_surplus_pages(h, delta) < 0)
2931 goto out;
2932
2933 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
2934 return_unused_surplus_pages(h, delta);
2935 goto out;
2936 }
2937 }
2938
2939 ret = 0;
2940 if (delta < 0)
2941 return_unused_surplus_pages(h, (unsigned long) -delta);
2942
2943 out:
2944 spin_unlock(&hugetlb_lock);
2945 return ret;
2946 }
2947
2948 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
2949 {
2950 struct resv_map *resv = vma_resv_map(vma);
2951
2952 /*
2953 * This new VMA should share its siblings reservation map if present.
2954 * The VMA will only ever have a valid reservation map pointer where
2955 * it is being copied for another still existing VMA. As that VMA
2956 * has a reference to the reservation map it cannot disappear until
2957 * after this open call completes. It is therefore safe to take a
2958 * new reference here without additional locking.
2959 */
2960 if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2961 kref_get(&resv->refs);
2962 }
2963
2964 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
2965 {
2966 struct hstate *h = hstate_vma(vma);
2967 struct resv_map *resv = vma_resv_map(vma);
2968 struct hugepage_subpool *spool = subpool_vma(vma);
2969 unsigned long reserve, start, end;
2970 long gbl_reserve;
2971
2972 if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
2973 return;
2974
2975 start = vma_hugecache_offset(h, vma, vma->vm_start);
2976 end = vma_hugecache_offset(h, vma, vma->vm_end);
2977
2978 reserve = (end - start) - region_count(resv, start, end);
2979
2980 kref_put(&resv->refs, resv_map_release);
2981
2982 if (reserve) {
2983 /*
2984 * Decrement reserve counts. The global reserve count may be
2985 * adjusted if the subpool has a minimum size.
2986 */
2987 gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
2988 hugetlb_acct_memory(h, -gbl_reserve);
2989 }
2990 }
2991
2992 /*
2993 * We cannot handle pagefaults against hugetlb pages at all. They cause
2994 * handle_mm_fault() to try to instantiate regular-sized pages in the
2995 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
2996 * this far.
2997 */
2998 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2999 {
3000 BUG();
3001 return 0;
3002 }
3003
3004 const struct vm_operations_struct hugetlb_vm_ops = {
3005 .fault = hugetlb_vm_op_fault,
3006 .open = hugetlb_vm_op_open,
3007 .close = hugetlb_vm_op_close,
3008 };
3009
3010 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3011 int writable)
3012 {
3013 pte_t entry;
3014
3015 if (writable) {
3016 entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3017 vma->vm_page_prot)));
3018 } else {
3019 entry = huge_pte_wrprotect(mk_huge_pte(page,
3020 vma->vm_page_prot));
3021 }
3022 entry = pte_mkyoung(entry);
3023 entry = pte_mkhuge(entry);
3024 entry = arch_make_huge_pte(entry, vma, page, writable);
3025
3026 return entry;
3027 }
3028
3029 static void set_huge_ptep_writable(struct vm_area_struct *vma,
3030 unsigned long address, pte_t *ptep)
3031 {
3032 pte_t entry;
3033
3034 entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3035 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3036 update_mmu_cache(vma, address, ptep);
3037 }
3038
3039 static int is_hugetlb_entry_migration(pte_t pte)
3040 {
3041 swp_entry_t swp;
3042
3043 if (huge_pte_none(pte) || pte_present(pte))
3044 return 0;
3045 swp = pte_to_swp_entry(pte);
3046 if (non_swap_entry(swp) && is_migration_entry(swp))
3047 return 1;
3048 else
3049 return 0;
3050 }
3051
3052 static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3053 {
3054 swp_entry_t swp;
3055
3056 if (huge_pte_none(pte) || pte_present(pte))
3057 return 0;
3058 swp = pte_to_swp_entry(pte);
3059 if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3060 return 1;
3061 else
3062 return 0;
3063 }
3064
3065 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3066 struct vm_area_struct *vma)
3067 {
3068 pte_t *src_pte, *dst_pte, entry;
3069 struct page *ptepage;
3070 unsigned long addr;
3071 int cow;
3072 struct hstate *h = hstate_vma(vma);
3073 unsigned long sz = huge_page_size(h);
3074 unsigned long mmun_start; /* For mmu_notifiers */
3075 unsigned long mmun_end; /* For mmu_notifiers */
3076 int ret = 0;
3077
3078 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3079
3080 mmun_start = vma->vm_start;
3081 mmun_end = vma->vm_end;
3082 if (cow)
3083 mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end);
3084
3085 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3086 spinlock_t *src_ptl, *dst_ptl;
3087 src_pte = huge_pte_offset(src, addr);
3088 if (!src_pte)
3089 continue;
3090 dst_pte = huge_pte_alloc(dst, addr, sz);
3091 if (!dst_pte) {
3092 ret = -ENOMEM;
3093 break;
3094 }
3095
3096 /* If the pagetables are shared don't copy or take references */
3097 if (dst_pte == src_pte)
3098 continue;
3099
3100 dst_ptl = huge_pte_lock(h, dst, dst_pte);
3101 src_ptl = huge_pte_lockptr(h, src, src_pte);
3102 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3103 entry = huge_ptep_get(src_pte);
3104 if (huge_pte_none(entry)) { /* skip none entry */
3105 ;
3106 } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3107 is_hugetlb_entry_hwpoisoned(entry))) {
3108 swp_entry_t swp_entry = pte_to_swp_entry(entry);
3109
3110 if (is_write_migration_entry(swp_entry) && cow) {
3111 /*
3112 * COW mappings require pages in both
3113 * parent and child to be set to read.
3114 */
3115 make_migration_entry_read(&swp_entry);
3116 entry = swp_entry_to_pte(swp_entry);
3117 set_huge_pte_at(src, addr, src_pte, entry);
3118 }
3119 set_huge_pte_at(dst, addr, dst_pte, entry);
3120 } else {
3121 if (cow) {
3122 huge_ptep_set_wrprotect(src, addr, src_pte);
3123 mmu_notifier_invalidate_range(src, mmun_start,
3124 mmun_end);
3125 }
3126 entry = huge_ptep_get(src_pte);
3127 ptepage = pte_page(entry);
3128 get_page(ptepage);
3129 page_dup_rmap(ptepage);
3130 set_huge_pte_at(dst, addr, dst_pte, entry);
3131 hugetlb_count_add(pages_per_huge_page(h), dst);
3132 }
3133 spin_unlock(src_ptl);
3134 spin_unlock(dst_ptl);
3135 }
3136
3137 if (cow)
3138 mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end);
3139
3140 return ret;
3141 }
3142
3143 void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3144 unsigned long start, unsigned long end,
3145 struct page *ref_page)
3146 {
3147 int force_flush = 0;
3148 struct mm_struct *mm = vma->vm_mm;
3149 unsigned long address;
3150 pte_t *ptep;
3151 pte_t pte;
3152 spinlock_t *ptl;
3153 struct page *page;
3154 struct hstate *h = hstate_vma(vma);
3155 unsigned long sz = huge_page_size(h);
3156 const unsigned long mmun_start = start; /* For mmu_notifiers */
3157 const unsigned long mmun_end = end; /* For mmu_notifiers */
3158
3159 WARN_ON(!is_vm_hugetlb_page(vma));
3160 BUG_ON(start & ~huge_page_mask(h));
3161 BUG_ON(end & ~huge_page_mask(h));
3162
3163 tlb_start_vma(tlb, vma);
3164 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3165 address = start;
3166 again:
3167 for (; address < end; address += sz) {
3168 ptep = huge_pte_offset(mm, address);
3169 if (!ptep)
3170 continue;
3171
3172 ptl = huge_pte_lock(h, mm, ptep);
3173 if (huge_pmd_unshare(mm, &address, ptep))
3174 goto unlock;
3175
3176 pte = huge_ptep_get(ptep);
3177 if (huge_pte_none(pte))
3178 goto unlock;
3179
3180 /*
3181 * Migrating hugepage or HWPoisoned hugepage is already
3182 * unmapped and its refcount is dropped, so just clear pte here.
3183 */
3184 if (unlikely(!pte_present(pte))) {
3185 huge_pte_clear(mm, address, ptep);
3186 goto unlock;
3187 }
3188
3189 page = pte_page(pte);
3190 /*
3191 * If a reference page is supplied, it is because a specific
3192 * page is being unmapped, not a range. Ensure the page we
3193 * are about to unmap is the actual page of interest.
3194 */
3195 if (ref_page) {
3196 if (page != ref_page)
3197 goto unlock;
3198
3199 /*
3200 * Mark the VMA as having unmapped its page so that
3201 * future faults in this VMA will fail rather than
3202 * looking like data was lost
3203 */
3204 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3205 }
3206
3207 pte = huge_ptep_get_and_clear(mm, address, ptep);
3208 tlb_remove_tlb_entry(tlb, ptep, address);
3209 if (huge_pte_dirty(pte))
3210 set_page_dirty(page);
3211
3212 hugetlb_count_sub(pages_per_huge_page(h), mm);
3213 page_remove_rmap(page);
3214 force_flush = !__tlb_remove_page(tlb, page);
3215 if (force_flush) {
3216 address += sz;
3217 spin_unlock(ptl);
3218 break;
3219 }
3220 /* Bail out after unmapping reference page if supplied */
3221 if (ref_page) {
3222 spin_unlock(ptl);
3223 break;
3224 }
3225 unlock:
3226 spin_unlock(ptl);
3227 }
3228 /*
3229 * mmu_gather ran out of room to batch pages, we break out of
3230 * the PTE lock to avoid doing the potential expensive TLB invalidate
3231 * and page-free while holding it.
3232 */
3233 if (force_flush) {
3234 force_flush = 0;
3235 tlb_flush_mmu(tlb);
3236 if (address < end && !ref_page)
3237 goto again;
3238 }
3239 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3240 tlb_end_vma(tlb, vma);
3241 }
3242
3243 void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3244 struct vm_area_struct *vma, unsigned long start,
3245 unsigned long end, struct page *ref_page)
3246 {
3247 __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3248
3249 /*
3250 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3251 * test will fail on a vma being torn down, and not grab a page table
3252 * on its way out. We're lucky that the flag has such an appropriate
3253 * name, and can in fact be safely cleared here. We could clear it
3254 * before the __unmap_hugepage_range above, but all that's necessary
3255 * is to clear it before releasing the i_mmap_rwsem. This works
3256 * because in the context this is called, the VMA is about to be
3257 * destroyed and the i_mmap_rwsem is held.
3258 */
3259 vma->vm_flags &= ~VM_MAYSHARE;
3260 }
3261
3262 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3263 unsigned long end, struct page *ref_page)
3264 {
3265 struct mm_struct *mm;
3266 struct mmu_gather tlb;
3267
3268 mm = vma->vm_mm;
3269
3270 tlb_gather_mmu(&tlb, mm, start, end);
3271 __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3272 tlb_finish_mmu(&tlb, start, end);
3273 }
3274
3275 /*
3276 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3277 * mappping it owns the reserve page for. The intention is to unmap the page
3278 * from other VMAs and let the children be SIGKILLed if they are faulting the
3279 * same region.
3280 */
3281 static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3282 struct page *page, unsigned long address)
3283 {
3284 struct hstate *h = hstate_vma(vma);
3285 struct vm_area_struct *iter_vma;
3286 struct address_space *mapping;
3287 pgoff_t pgoff;
3288
3289 /*
3290 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3291 * from page cache lookup which is in HPAGE_SIZE units.
3292 */
3293 address = address & huge_page_mask(h);
3294 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3295 vma->vm_pgoff;
3296 mapping = file_inode(vma->vm_file)->i_mapping;
3297
3298 /*
3299 * Take the mapping lock for the duration of the table walk. As
3300 * this mapping should be shared between all the VMAs,
3301 * __unmap_hugepage_range() is called as the lock is already held
3302 */
3303 i_mmap_lock_write(mapping);
3304 vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3305 /* Do not unmap the current VMA */
3306 if (iter_vma == vma)
3307 continue;
3308
3309 /*
3310 * Shared VMAs have their own reserves and do not affect
3311 * MAP_PRIVATE accounting but it is possible that a shared
3312 * VMA is using the same page so check and skip such VMAs.
3313 */
3314 if (iter_vma->vm_flags & VM_MAYSHARE)
3315 continue;
3316
3317 /*
3318 * Unmap the page from other VMAs without their own reserves.
3319 * They get marked to be SIGKILLed if they fault in these
3320 * areas. This is because a future no-page fault on this VMA
3321 * could insert a zeroed page instead of the data existing
3322 * from the time of fork. This would look like data corruption
3323 */
3324 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3325 unmap_hugepage_range(iter_vma, address,
3326 address + huge_page_size(h), page);
3327 }
3328 i_mmap_unlock_write(mapping);
3329 }
3330
3331 /*
3332 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3333 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3334 * cannot race with other handlers or page migration.
3335 * Keep the pte_same checks anyway to make transition from the mutex easier.
3336 */
3337 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3338 unsigned long address, pte_t *ptep, pte_t pte,
3339 struct page *pagecache_page, spinlock_t *ptl)
3340 {
3341 struct hstate *h = hstate_vma(vma);
3342 struct page *old_page, *new_page;
3343 int ret = 0, outside_reserve = 0;
3344 unsigned long mmun_start; /* For mmu_notifiers */
3345 unsigned long mmun_end; /* For mmu_notifiers */
3346
3347 old_page = pte_page(pte);
3348
3349 retry_avoidcopy:
3350 /* If no-one else is actually using this page, avoid the copy
3351 * and just make the page writable */
3352 if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3353 page_move_anon_rmap(old_page, vma, address);
3354 set_huge_ptep_writable(vma, address, ptep);
3355 return 0;
3356 }
3357
3358 /*
3359 * If the process that created a MAP_PRIVATE mapping is about to
3360 * perform a COW due to a shared page count, attempt to satisfy
3361 * the allocation without using the existing reserves. The pagecache
3362 * page is used to determine if the reserve at this address was
3363 * consumed or not. If reserves were used, a partial faulted mapping
3364 * at the time of fork() could consume its reserves on COW instead
3365 * of the full address range.
3366 */
3367 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3368 old_page != pagecache_page)
3369 outside_reserve = 1;
3370
3371 page_cache_get(old_page);
3372
3373 /*
3374 * Drop page table lock as buddy allocator may be called. It will
3375 * be acquired again before returning to the caller, as expected.
3376 */
3377 spin_unlock(ptl);
3378 new_page = alloc_huge_page(vma, address, outside_reserve);
3379
3380 if (IS_ERR(new_page)) {
3381 /*
3382 * If a process owning a MAP_PRIVATE mapping fails to COW,
3383 * it is due to references held by a child and an insufficient
3384 * huge page pool. To guarantee the original mappers
3385 * reliability, unmap the page from child processes. The child
3386 * may get SIGKILLed if it later faults.
3387 */
3388 if (outside_reserve) {
3389 page_cache_release(old_page);
3390 BUG_ON(huge_pte_none(pte));
3391 unmap_ref_private(mm, vma, old_page, address);
3392 BUG_ON(huge_pte_none(pte));
3393 spin_lock(ptl);
3394 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3395 if (likely(ptep &&
3396 pte_same(huge_ptep_get(ptep), pte)))
3397 goto retry_avoidcopy;
3398 /*
3399 * race occurs while re-acquiring page table
3400 * lock, and our job is done.
3401 */
3402 return 0;
3403 }
3404
3405 ret = (PTR_ERR(new_page) == -ENOMEM) ?
3406 VM_FAULT_OOM : VM_FAULT_SIGBUS;
3407 goto out_release_old;
3408 }
3409
3410 /*
3411 * When the original hugepage is shared one, it does not have
3412 * anon_vma prepared.
3413 */
3414 if (unlikely(anon_vma_prepare(vma))) {
3415 ret = VM_FAULT_OOM;
3416 goto out_release_all;
3417 }
3418
3419 copy_user_huge_page(new_page, old_page, address, vma,
3420 pages_per_huge_page(h));
3421 __SetPageUptodate(new_page);
3422 set_page_huge_active(new_page);
3423
3424 mmun_start = address & huge_page_mask(h);
3425 mmun_end = mmun_start + huge_page_size(h);
3426 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
3427
3428 /*
3429 * Retake the page table lock to check for racing updates
3430 * before the page tables are altered
3431 */
3432 spin_lock(ptl);
3433 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
3434 if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3435 ClearPagePrivate(new_page);
3436
3437 /* Break COW */
3438 huge_ptep_clear_flush(vma, address, ptep);
3439 mmu_notifier_invalidate_range(mm, mmun_start, mmun_end);
3440 set_huge_pte_at(mm, address, ptep,
3441 make_huge_pte(vma, new_page, 1));
3442 page_remove_rmap(old_page);
3443 hugepage_add_new_anon_rmap(new_page, vma, address);
3444 /* Make the old page be freed below */
3445 new_page = old_page;
3446 }
3447 spin_unlock(ptl);
3448 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
3449 out_release_all:
3450 page_cache_release(new_page);
3451 out_release_old:
3452 page_cache_release(old_page);
3453
3454 spin_lock(ptl); /* Caller expects lock to be held */
3455 return ret;
3456 }
3457
3458 /* Return the pagecache page at a given address within a VMA */
3459 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3460 struct vm_area_struct *vma, unsigned long address)
3461 {
3462 struct address_space *mapping;
3463 pgoff_t idx;
3464
3465 mapping = vma->vm_file->f_mapping;
3466 idx = vma_hugecache_offset(h, vma, address);
3467
3468 return find_lock_page(mapping, idx);
3469 }
3470
3471 /*
3472 * Return whether there is a pagecache page to back given address within VMA.
3473 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3474 */
3475 static bool hugetlbfs_pagecache_present(struct hstate *h,
3476 struct vm_area_struct *vma, unsigned long address)
3477 {
3478 struct address_space *mapping;
3479 pgoff_t idx;
3480 struct page *page;
3481
3482 mapping = vma->vm_file->f_mapping;
3483 idx = vma_hugecache_offset(h, vma, address);
3484
3485 page = find_get_page(mapping, idx);
3486 if (page)
3487 put_page(page);
3488 return page != NULL;
3489 }
3490
3491 int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3492 pgoff_t idx)
3493 {
3494 struct inode *inode = mapping->host;
3495 struct hstate *h = hstate_inode(inode);
3496 int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3497
3498 if (err)
3499 return err;
3500 ClearPagePrivate(page);
3501
3502 spin_lock(&inode->i_lock);
3503 inode->i_blocks += blocks_per_huge_page(h);
3504 spin_unlock(&inode->i_lock);
3505 return 0;
3506 }
3507
3508 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
3509 struct address_space *mapping, pgoff_t idx,
3510 unsigned long address, pte_t *ptep, unsigned int flags)
3511 {
3512 struct hstate *h = hstate_vma(vma);
3513 int ret = VM_FAULT_SIGBUS;
3514 int anon_rmap = 0;
3515 unsigned long size;
3516 struct page *page;
3517 pte_t new_pte;
3518 spinlock_t *ptl;
3519
3520 /*
3521 * Currently, we are forced to kill the process in the event the
3522 * original mapper has unmapped pages from the child due to a failed
3523 * COW. Warn that such a situation has occurred as it may not be obvious
3524 */
3525 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3526 pr_warning("PID %d killed due to inadequate hugepage pool\n",
3527 current->pid);
3528 return ret;
3529 }
3530
3531 /*
3532 * Use page lock to guard against racing truncation
3533 * before we get page_table_lock.
3534 */
3535 retry:
3536 page = find_lock_page(mapping, idx);
3537 if (!page) {
3538 size = i_size_read(mapping->host) >> huge_page_shift(h);
3539 if (idx >= size)
3540 goto out;
3541 page = alloc_huge_page(vma, address, 0);
3542 if (IS_ERR(page)) {
3543 ret = PTR_ERR(page);
3544 if (ret == -ENOMEM)
3545 ret = VM_FAULT_OOM;
3546 else
3547 ret = VM_FAULT_SIGBUS;
3548 goto out;
3549 }
3550 clear_huge_page(page, address, pages_per_huge_page(h));
3551 __SetPageUptodate(page);
3552 set_page_huge_active(page);
3553
3554 if (vma->vm_flags & VM_MAYSHARE) {
3555 int err = huge_add_to_page_cache(page, mapping, idx);
3556 if (err) {
3557 put_page(page);
3558 if (err == -EEXIST)
3559 goto retry;
3560 goto out;
3561 }
3562 } else {
3563 lock_page(page);
3564 if (unlikely(anon_vma_prepare(vma))) {
3565 ret = VM_FAULT_OOM;
3566 goto backout_unlocked;
3567 }
3568 anon_rmap = 1;
3569 }
3570 } else {
3571 /*
3572 * If memory error occurs between mmap() and fault, some process
3573 * don't have hwpoisoned swap entry for errored virtual address.
3574 * So we need to block hugepage fault by PG_hwpoison bit check.
3575 */
3576 if (unlikely(PageHWPoison(page))) {
3577 ret = VM_FAULT_HWPOISON |
3578 VM_FAULT_SET_HINDEX(hstate_index(h));
3579 goto backout_unlocked;
3580 }
3581 }
3582
3583 /*
3584 * If we are going to COW a private mapping later, we examine the
3585 * pending reservations for this page now. This will ensure that
3586 * any allocations necessary to record that reservation occur outside
3587 * the spinlock.
3588 */
3589 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3590 if (vma_needs_reservation(h, vma, address) < 0) {
3591 ret = VM_FAULT_OOM;
3592 goto backout_unlocked;
3593 }
3594 /* Just decrements count, does not deallocate */
3595 vma_end_reservation(h, vma, address);
3596 }
3597
3598 ptl = huge_pte_lockptr(h, mm, ptep);
3599 spin_lock(ptl);
3600 size = i_size_read(mapping->host) >> huge_page_shift(h);
3601 if (idx >= size)
3602 goto backout;
3603
3604 ret = 0;
3605 if (!huge_pte_none(huge_ptep_get(ptep)))
3606 goto backout;
3607
3608 if (anon_rmap) {
3609 ClearPagePrivate(page);
3610 hugepage_add_new_anon_rmap(page, vma, address);
3611 } else
3612 page_dup_rmap(page);
3613 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3614 && (vma->vm_flags & VM_SHARED)));
3615 set_huge_pte_at(mm, address, ptep, new_pte);
3616
3617 hugetlb_count_add(pages_per_huge_page(h), mm);
3618 if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3619 /* Optimization, do the COW without a second fault */
3620 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl);
3621 }
3622
3623 spin_unlock(ptl);
3624 unlock_page(page);
3625 out:
3626 return ret;
3627
3628 backout:
3629 spin_unlock(ptl);
3630 backout_unlocked:
3631 unlock_page(page);
3632 put_page(page);
3633 goto out;
3634 }
3635
3636 #ifdef CONFIG_SMP
3637 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3638 struct vm_area_struct *vma,
3639 struct address_space *mapping,
3640 pgoff_t idx, unsigned long address)
3641 {
3642 unsigned long key[2];
3643 u32 hash;
3644
3645 if (vma->vm_flags & VM_SHARED) {
3646 key[0] = (unsigned long) mapping;
3647 key[1] = idx;
3648 } else {
3649 key[0] = (unsigned long) mm;
3650 key[1] = address >> huge_page_shift(h);
3651 }
3652
3653 hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0);
3654
3655 return hash & (num_fault_mutexes - 1);
3656 }
3657 #else
3658 /*
3659 * For uniprocesor systems we always use a single mutex, so just
3660 * return 0 and avoid the hashing overhead.
3661 */
3662 u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm,
3663 struct vm_area_struct *vma,
3664 struct address_space *mapping,
3665 pgoff_t idx, unsigned long address)
3666 {
3667 return 0;
3668 }
3669 #endif
3670
3671 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3672 unsigned long address, unsigned int flags)
3673 {
3674 pte_t *ptep, entry;
3675 spinlock_t *ptl;
3676 int ret;
3677 u32 hash;
3678 pgoff_t idx;
3679 struct page *page = NULL;
3680 struct page *pagecache_page = NULL;
3681 struct hstate *h = hstate_vma(vma);
3682 struct address_space *mapping;
3683 int need_wait_lock = 0;
3684
3685 address &= huge_page_mask(h);
3686
3687 ptep = huge_pte_offset(mm, address);
3688 if (ptep) {
3689 entry = huge_ptep_get(ptep);
3690 if (unlikely(is_hugetlb_entry_migration(entry))) {
3691 migration_entry_wait_huge(vma, mm, ptep);
3692 return 0;
3693 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
3694 return VM_FAULT_HWPOISON_LARGE |
3695 VM_FAULT_SET_HINDEX(hstate_index(h));
3696 }
3697
3698 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
3699 if (!ptep)
3700 return VM_FAULT_OOM;
3701
3702 mapping = vma->vm_file->f_mapping;
3703 idx = vma_hugecache_offset(h, vma, address);
3704
3705 /*
3706 * Serialize hugepage allocation and instantiation, so that we don't
3707 * get spurious allocation failures if two CPUs race to instantiate
3708 * the same page in the page cache.
3709 */
3710 hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address);
3711 mutex_lock(&hugetlb_fault_mutex_table[hash]);
3712
3713 entry = huge_ptep_get(ptep);
3714 if (huge_pte_none(entry)) {
3715 ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
3716 goto out_mutex;
3717 }
3718
3719 ret = 0;
3720
3721 /*
3722 * entry could be a migration/hwpoison entry at this point, so this
3723 * check prevents the kernel from going below assuming that we have
3724 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3725 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3726 * handle it.
3727 */
3728 if (!pte_present(entry))
3729 goto out_mutex;
3730
3731 /*
3732 * If we are going to COW the mapping later, we examine the pending
3733 * reservations for this page now. This will ensure that any
3734 * allocations necessary to record that reservation occur outside the
3735 * spinlock. For private mappings, we also lookup the pagecache
3736 * page now as it is used to determine if a reservation has been
3737 * consumed.
3738 */
3739 if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
3740 if (vma_needs_reservation(h, vma, address) < 0) {
3741 ret = VM_FAULT_OOM;
3742 goto out_mutex;
3743 }
3744 /* Just decrements count, does not deallocate */
3745 vma_end_reservation(h, vma, address);
3746
3747 if (!(vma->vm_flags & VM_MAYSHARE))
3748 pagecache_page = hugetlbfs_pagecache_page(h,
3749 vma, address);
3750 }
3751
3752 ptl = huge_pte_lock(h, mm, ptep);
3753
3754 /* Check for a racing update before calling hugetlb_cow */
3755 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
3756 goto out_ptl;
3757
3758 /*
3759 * hugetlb_cow() requires page locks of pte_page(entry) and
3760 * pagecache_page, so here we need take the former one
3761 * when page != pagecache_page or !pagecache_page.
3762 */
3763 page = pte_page(entry);
3764 if (page != pagecache_page)
3765 if (!trylock_page(page)) {
3766 need_wait_lock = 1;
3767 goto out_ptl;
3768 }
3769
3770 get_page(page);
3771
3772 if (flags & FAULT_FLAG_WRITE) {
3773 if (!huge_pte_write(entry)) {
3774 ret = hugetlb_cow(mm, vma, address, ptep, entry,
3775 pagecache_page, ptl);
3776 goto out_put_page;
3777 }
3778 entry = huge_pte_mkdirty(entry);
3779 }
3780 entry = pte_mkyoung(entry);
3781 if (huge_ptep_set_access_flags(vma, address, ptep, entry,
3782 flags & FAULT_FLAG_WRITE))
3783 update_mmu_cache(vma, address, ptep);
3784 out_put_page:
3785 if (page != pagecache_page)
3786 unlock_page(page);
3787 put_page(page);
3788 out_ptl:
3789 spin_unlock(ptl);
3790
3791 if (pagecache_page) {
3792 unlock_page(pagecache_page);
3793 put_page(pagecache_page);
3794 }
3795 out_mutex:
3796 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3797 /*
3798 * Generally it's safe to hold refcount during waiting page lock. But
3799 * here we just wait to defer the next page fault to avoid busy loop and
3800 * the page is not used after unlocked before returning from the current
3801 * page fault. So we are safe from accessing freed page, even if we wait
3802 * here without taking refcount.
3803 */
3804 if (need_wait_lock)
3805 wait_on_page_locked(page);
3806 return ret;
3807 }
3808
3809 long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
3810 struct page **pages, struct vm_area_struct **vmas,
3811 unsigned long *position, unsigned long *nr_pages,
3812 long i, unsigned int flags)
3813 {
3814 unsigned long pfn_offset;
3815 unsigned long vaddr = *position;
3816 unsigned long remainder = *nr_pages;
3817 struct hstate *h = hstate_vma(vma);
3818
3819 while (vaddr < vma->vm_end && remainder) {
3820 pte_t *pte;
3821 spinlock_t *ptl = NULL;
3822 int absent;
3823 struct page *page;
3824
3825 /*
3826 * If we have a pending SIGKILL, don't keep faulting pages and
3827 * potentially allocating memory.
3828 */
3829 if (unlikely(fatal_signal_pending(current))) {
3830 remainder = 0;
3831 break;
3832 }
3833
3834 /*
3835 * Some archs (sparc64, sh*) have multiple pte_ts to
3836 * each hugepage. We have to make sure we get the
3837 * first, for the page indexing below to work.
3838 *
3839 * Note that page table lock is not held when pte is null.
3840 */
3841 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
3842 if (pte)
3843 ptl = huge_pte_lock(h, mm, pte);
3844 absent = !pte || huge_pte_none(huge_ptep_get(pte));
3845
3846 /*
3847 * When coredumping, it suits get_dump_page if we just return
3848 * an error where there's an empty slot with no huge pagecache
3849 * to back it. This way, we avoid allocating a hugepage, and
3850 * the sparse dumpfile avoids allocating disk blocks, but its
3851 * huge holes still show up with zeroes where they need to be.
3852 */
3853 if (absent && (flags & FOLL_DUMP) &&
3854 !hugetlbfs_pagecache_present(h, vma, vaddr)) {
3855 if (pte)
3856 spin_unlock(ptl);
3857 remainder = 0;
3858 break;
3859 }
3860
3861 /*
3862 * We need call hugetlb_fault for both hugepages under migration
3863 * (in which case hugetlb_fault waits for the migration,) and
3864 * hwpoisoned hugepages (in which case we need to prevent the
3865 * caller from accessing to them.) In order to do this, we use
3866 * here is_swap_pte instead of is_hugetlb_entry_migration and
3867 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
3868 * both cases, and because we can't follow correct pages
3869 * directly from any kind of swap entries.
3870 */
3871 if (absent || is_swap_pte(huge_ptep_get(pte)) ||
3872 ((flags & FOLL_WRITE) &&
3873 !huge_pte_write(huge_ptep_get(pte)))) {
3874 int ret;
3875
3876 if (pte)
3877 spin_unlock(ptl);
3878 ret = hugetlb_fault(mm, vma, vaddr,
3879 (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0);
3880 if (!(ret & VM_FAULT_ERROR))
3881 continue;
3882
3883 remainder = 0;
3884 break;
3885 }
3886
3887 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
3888 page = pte_page(huge_ptep_get(pte));
3889 same_page:
3890 if (pages) {
3891 pages[i] = mem_map_offset(page, pfn_offset);
3892 get_page_foll(pages[i]);
3893 }
3894
3895 if (vmas)
3896 vmas[i] = vma;
3897
3898 vaddr += PAGE_SIZE;
3899 ++pfn_offset;
3900 --remainder;
3901 ++i;
3902 if (vaddr < vma->vm_end && remainder &&
3903 pfn_offset < pages_per_huge_page(h)) {
3904 /*
3905 * We use pfn_offset to avoid touching the pageframes
3906 * of this compound page.
3907 */
3908 goto same_page;
3909 }
3910 spin_unlock(ptl);
3911 }
3912 *nr_pages = remainder;
3913 *position = vaddr;
3914
3915 return i ? i : -EFAULT;
3916 }
3917
3918 unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
3919 unsigned long address, unsigned long end, pgprot_t newprot)
3920 {
3921 struct mm_struct *mm = vma->vm_mm;
3922 unsigned long start = address;
3923 pte_t *ptep;
3924 pte_t pte;
3925 struct hstate *h = hstate_vma(vma);
3926 unsigned long pages = 0;
3927
3928 BUG_ON(address >= end);
3929 flush_cache_range(vma, address, end);
3930
3931 mmu_notifier_invalidate_range_start(mm, start, end);
3932 i_mmap_lock_write(vma->vm_file->f_mapping);
3933 for (; address < end; address += huge_page_size(h)) {
3934 spinlock_t *ptl;
3935 ptep = huge_pte_offset(mm, address);
3936 if (!ptep)
3937 continue;
3938 ptl = huge_pte_lock(h, mm, ptep);
3939 if (huge_pmd_unshare(mm, &address, ptep)) {
3940 pages++;
3941 spin_unlock(ptl);
3942 continue;
3943 }
3944 pte = huge_ptep_get(ptep);
3945 if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
3946 spin_unlock(ptl);
3947 continue;
3948 }
3949 if (unlikely(is_hugetlb_entry_migration(pte))) {
3950 swp_entry_t entry = pte_to_swp_entry(pte);
3951
3952 if (is_write_migration_entry(entry)) {
3953 pte_t newpte;
3954
3955 make_migration_entry_read(&entry);
3956 newpte = swp_entry_to_pte(entry);
3957 set_huge_pte_at(mm, address, ptep, newpte);
3958 pages++;
3959 }
3960 spin_unlock(ptl);
3961 continue;
3962 }
3963 if (!huge_pte_none(pte)) {
3964 pte = huge_ptep_get_and_clear(mm, address, ptep);
3965 pte = pte_mkhuge(huge_pte_modify(pte, newprot));
3966 pte = arch_make_huge_pte(pte, vma, NULL, 0);
3967 set_huge_pte_at(mm, address, ptep, pte);
3968 pages++;
3969 }
3970 spin_unlock(ptl);
3971 }
3972 /*
3973 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
3974 * may have cleared our pud entry and done put_page on the page table:
3975 * once we release i_mmap_rwsem, another task can do the final put_page
3976 * and that page table be reused and filled with junk.
3977 */
3978 flush_tlb_range(vma, start, end);
3979 mmu_notifier_invalidate_range(mm, start, end);
3980 i_mmap_unlock_write(vma->vm_file->f_mapping);
3981 mmu_notifier_invalidate_range_end(mm, start, end);
3982
3983 return pages << h->order;
3984 }
3985
3986 int hugetlb_reserve_pages(struct inode *inode,
3987 long from, long to,
3988 struct vm_area_struct *vma,
3989 vm_flags_t vm_flags)
3990 {
3991 long ret, chg;
3992 struct hstate *h = hstate_inode(inode);
3993 struct hugepage_subpool *spool = subpool_inode(inode);
3994 struct resv_map *resv_map;
3995 long gbl_reserve;
3996
3997 /*
3998 * Only apply hugepage reservation if asked. At fault time, an
3999 * attempt will be made for VM_NORESERVE to allocate a page
4000 * without using reserves
4001 */
4002 if (vm_flags & VM_NORESERVE)
4003 return 0;
4004
4005 /*
4006 * Shared mappings base their reservation on the number of pages that
4007 * are already allocated on behalf of the file. Private mappings need
4008 * to reserve the full area even if read-only as mprotect() may be
4009 * called to make the mapping read-write. Assume !vma is a shm mapping
4010 */
4011 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4012 resv_map = inode_resv_map(inode);
4013
4014 chg = region_chg(resv_map, from, to);
4015
4016 } else {
4017 resv_map = resv_map_alloc();
4018 if (!resv_map)
4019 return -ENOMEM;
4020
4021 chg = to - from;
4022
4023 set_vma_resv_map(vma, resv_map);
4024 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4025 }
4026
4027 if (chg < 0) {
4028 ret = chg;
4029 goto out_err;
4030 }
4031
4032 /*
4033 * There must be enough pages in the subpool for the mapping. If
4034 * the subpool has a minimum size, there may be some global
4035 * reservations already in place (gbl_reserve).
4036 */
4037 gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4038 if (gbl_reserve < 0) {
4039 ret = -ENOSPC;
4040 goto out_err;
4041 }
4042
4043 /*
4044 * Check enough hugepages are available for the reservation.
4045 * Hand the pages back to the subpool if there are not
4046 */
4047 ret = hugetlb_acct_memory(h, gbl_reserve);
4048 if (ret < 0) {
4049 /* put back original number of pages, chg */
4050 (void)hugepage_subpool_put_pages(spool, chg);
4051 goto out_err;
4052 }
4053
4054 /*
4055 * Account for the reservations made. Shared mappings record regions
4056 * that have reservations as they are shared by multiple VMAs.
4057 * When the last VMA disappears, the region map says how much
4058 * the reservation was and the page cache tells how much of
4059 * the reservation was consumed. Private mappings are per-VMA and
4060 * only the consumed reservations are tracked. When the VMA
4061 * disappears, the original reservation is the VMA size and the
4062 * consumed reservations are stored in the map. Hence, nothing
4063 * else has to be done for private mappings here
4064 */
4065 if (!vma || vma->vm_flags & VM_MAYSHARE) {
4066 long add = region_add(resv_map, from, to);
4067
4068 if (unlikely(chg > add)) {
4069 /*
4070 * pages in this range were added to the reserve
4071 * map between region_chg and region_add. This
4072 * indicates a race with alloc_huge_page. Adjust
4073 * the subpool and reserve counts modified above
4074 * based on the difference.
4075 */
4076 long rsv_adjust;
4077
4078 rsv_adjust = hugepage_subpool_put_pages(spool,
4079 chg - add);
4080 hugetlb_acct_memory(h, -rsv_adjust);
4081 }
4082 }
4083 return 0;
4084 out_err:
4085 if (!vma || vma->vm_flags & VM_MAYSHARE)
4086 region_abort(resv_map, from, to);
4087 if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4088 kref_put(&resv_map->refs, resv_map_release);
4089 return ret;
4090 }
4091
4092 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4093 long freed)
4094 {
4095 struct hstate *h = hstate_inode(inode);
4096 struct resv_map *resv_map = inode_resv_map(inode);
4097 long chg = 0;
4098 struct hugepage_subpool *spool = subpool_inode(inode);
4099 long gbl_reserve;
4100
4101 if (resv_map) {
4102 chg = region_del(resv_map, start, end);
4103 /*
4104 * region_del() can fail in the rare case where a region
4105 * must be split and another region descriptor can not be
4106 * allocated. If end == LONG_MAX, it will not fail.
4107 */
4108 if (chg < 0)
4109 return chg;
4110 }
4111
4112 spin_lock(&inode->i_lock);
4113 inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4114 spin_unlock(&inode->i_lock);
4115
4116 /*
4117 * If the subpool has a minimum size, the number of global
4118 * reservations to be released may be adjusted.
4119 */
4120 gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4121 hugetlb_acct_memory(h, -gbl_reserve);
4122
4123 return 0;
4124 }
4125
4126 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4127 static unsigned long page_table_shareable(struct vm_area_struct *svma,
4128 struct vm_area_struct *vma,
4129 unsigned long addr, pgoff_t idx)
4130 {
4131 unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4132 svma->vm_start;
4133 unsigned long sbase = saddr & PUD_MASK;
4134 unsigned long s_end = sbase + PUD_SIZE;
4135
4136 /* Allow segments to share if only one is marked locked */
4137 unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4138 unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4139
4140 /*
4141 * match the virtual addresses, permission and the alignment of the
4142 * page table page.
4143 */
4144 if (pmd_index(addr) != pmd_index(saddr) ||
4145 vm_flags != svm_flags ||
4146 sbase < svma->vm_start || svma->vm_end < s_end)
4147 return 0;
4148
4149 return saddr;
4150 }
4151
4152 static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4153 {
4154 unsigned long base = addr & PUD_MASK;
4155 unsigned long end = base + PUD_SIZE;
4156
4157 /*
4158 * check on proper vm_flags and page table alignment
4159 */
4160 if (vma->vm_flags & VM_MAYSHARE &&
4161 vma->vm_start <= base && end <= vma->vm_end)
4162 return true;
4163 return false;
4164 }
4165
4166 /*
4167 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4168 * and returns the corresponding pte. While this is not necessary for the
4169 * !shared pmd case because we can allocate the pmd later as well, it makes the
4170 * code much cleaner. pmd allocation is essential for the shared case because
4171 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4172 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4173 * bad pmd for sharing.
4174 */
4175 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4176 {
4177 struct vm_area_struct *vma = find_vma(mm, addr);
4178 struct address_space *mapping = vma->vm_file->f_mapping;
4179 pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4180 vma->vm_pgoff;
4181 struct vm_area_struct *svma;
4182 unsigned long saddr;
4183 pte_t *spte = NULL;
4184 pte_t *pte;
4185 spinlock_t *ptl;
4186
4187 if (!vma_shareable(vma, addr))
4188 return (pte_t *)pmd_alloc(mm, pud, addr);
4189
4190 i_mmap_lock_write(mapping);
4191 vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4192 if (svma == vma)
4193 continue;
4194
4195 saddr = page_table_shareable(svma, vma, addr, idx);
4196 if (saddr) {
4197 spte = huge_pte_offset(svma->vm_mm, saddr);
4198 if (spte) {
4199 mm_inc_nr_pmds(mm);
4200 get_page(virt_to_page(spte));
4201 break;
4202 }
4203 }
4204 }
4205
4206 if (!spte)
4207 goto out;
4208
4209 ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte);
4210 spin_lock(ptl);
4211 if (pud_none(*pud)) {
4212 pud_populate(mm, pud,
4213 (pmd_t *)((unsigned long)spte & PAGE_MASK));
4214 } else {
4215 put_page(virt_to_page(spte));
4216 mm_inc_nr_pmds(mm);
4217 }
4218 spin_unlock(ptl);
4219 out:
4220 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4221 i_mmap_unlock_write(mapping);
4222 return pte;
4223 }
4224
4225 /*
4226 * unmap huge page backed by shared pte.
4227 *
4228 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4229 * indicated by page_count > 1, unmap is achieved by clearing pud and
4230 * decrementing the ref count. If count == 1, the pte page is not shared.
4231 *
4232 * called with page table lock held.
4233 *
4234 * returns: 1 successfully unmapped a shared pte page
4235 * 0 the underlying pte page is not shared, or it is the last user
4236 */
4237 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4238 {
4239 pgd_t *pgd = pgd_offset(mm, *addr);
4240 pud_t *pud = pud_offset(pgd, *addr);
4241
4242 BUG_ON(page_count(virt_to_page(ptep)) == 0);
4243 if (page_count(virt_to_page(ptep)) == 1)
4244 return 0;
4245
4246 pud_clear(pud);
4247 put_page(virt_to_page(ptep));
4248 mm_dec_nr_pmds(mm);
4249 *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4250 return 1;
4251 }
4252 #define want_pmd_share() (1)
4253 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4254 pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4255 {
4256 return NULL;
4257 }
4258
4259 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4260 {
4261 return 0;
4262 }
4263 #define want_pmd_share() (0)
4264 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4265
4266 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4267 pte_t *huge_pte_alloc(struct mm_struct *mm,
4268 unsigned long addr, unsigned long sz)
4269 {
4270 pgd_t *pgd;
4271 pud_t *pud;
4272 pte_t *pte = NULL;
4273
4274 pgd = pgd_offset(mm, addr);
4275 pud = pud_alloc(mm, pgd, addr);
4276 if (pud) {
4277 if (sz == PUD_SIZE) {
4278 pte = (pte_t *)pud;
4279 } else {
4280 BUG_ON(sz != PMD_SIZE);
4281 if (want_pmd_share() && pud_none(*pud))
4282 pte = huge_pmd_share(mm, addr, pud);
4283 else
4284 pte = (pte_t *)pmd_alloc(mm, pud, addr);
4285 }
4286 }
4287 BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
4288
4289 return pte;
4290 }
4291
4292 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
4293 {
4294 pgd_t *pgd;
4295 pud_t *pud;
4296 pmd_t *pmd = NULL;
4297
4298 pgd = pgd_offset(mm, addr);
4299 if (pgd_present(*pgd)) {
4300 pud = pud_offset(pgd, addr);
4301 if (pud_present(*pud)) {
4302 if (pud_huge(*pud))
4303 return (pte_t *)pud;
4304 pmd = pmd_offset(pud, addr);
4305 }
4306 }
4307 return (pte_t *) pmd;
4308 }
4309
4310 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4311
4312 /*
4313 * These functions are overwritable if your architecture needs its own
4314 * behavior.
4315 */
4316 struct page * __weak
4317 follow_huge_addr(struct mm_struct *mm, unsigned long address,
4318 int write)
4319 {
4320 return ERR_PTR(-EINVAL);
4321 }
4322
4323 struct page * __weak
4324 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4325 pmd_t *pmd, int flags)
4326 {
4327 struct page *page = NULL;
4328 spinlock_t *ptl;
4329 retry:
4330 ptl = pmd_lockptr(mm, pmd);
4331 spin_lock(ptl);
4332 /*
4333 * make sure that the address range covered by this pmd is not
4334 * unmapped from other threads.
4335 */
4336 if (!pmd_huge(*pmd))
4337 goto out;
4338 if (pmd_present(*pmd)) {
4339 page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4340 if (flags & FOLL_GET)
4341 get_page(page);
4342 } else {
4343 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) {
4344 spin_unlock(ptl);
4345 __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4346 goto retry;
4347 }
4348 /*
4349 * hwpoisoned entry is treated as no_page_table in
4350 * follow_page_mask().
4351 */
4352 }
4353 out:
4354 spin_unlock(ptl);
4355 return page;
4356 }
4357
4358 struct page * __weak
4359 follow_huge_pud(struct mm_struct *mm, unsigned long address,
4360 pud_t *pud, int flags)
4361 {
4362 if (flags & FOLL_GET)
4363 return NULL;
4364
4365 return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
4366 }
4367
4368 #ifdef CONFIG_MEMORY_FAILURE
4369
4370 /*
4371 * This function is called from memory failure code.
4372 * Assume the caller holds page lock of the head page.
4373 */
4374 int dequeue_hwpoisoned_huge_page(struct page *hpage)
4375 {
4376 struct hstate *h = page_hstate(hpage);
4377 int nid = page_to_nid(hpage);
4378 int ret = -EBUSY;
4379
4380 spin_lock(&hugetlb_lock);
4381 /*
4382 * Just checking !page_huge_active is not enough, because that could be
4383 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4384 */
4385 if (!page_huge_active(hpage) && !page_count(hpage)) {
4386 /*
4387 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4388 * but dangling hpage->lru can trigger list-debug warnings
4389 * (this happens when we call unpoison_memory() on it),
4390 * so let it point to itself with list_del_init().
4391 */
4392 list_del_init(&hpage->lru);
4393 set_page_refcounted(hpage);
4394 h->free_huge_pages--;
4395 h->free_huge_pages_node[nid]--;
4396 ret = 0;
4397 }
4398 spin_unlock(&hugetlb_lock);
4399 return ret;
4400 }
4401 #endif
4402
4403 bool isolate_huge_page(struct page *page, struct list_head *list)
4404 {
4405 bool ret = true;
4406
4407 VM_BUG_ON_PAGE(!PageHead(page), page);
4408 spin_lock(&hugetlb_lock);
4409 if (!page_huge_active(page) || !get_page_unless_zero(page)) {
4410 ret = false;
4411 goto unlock;
4412 }
4413 clear_page_huge_active(page);
4414 list_move_tail(&page->lru, list);
4415 unlock:
4416 spin_unlock(&hugetlb_lock);
4417 return ret;
4418 }
4419
4420 void putback_active_hugepage(struct page *page)
4421 {
4422 VM_BUG_ON_PAGE(!PageHead(page), page);
4423 spin_lock(&hugetlb_lock);
4424 set_page_huge_active(page);
4425 list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
4426 spin_unlock(&hugetlb_lock);
4427 put_page(page);
4428 }