Branch data Line data Source code
1 : : // SPDX-License-Identifier: GPL-2.0-only
2 : : /*
3 : : * Generic hugetlb support.
4 : : * (C) Nadia Yvette Chambers, April 2004
5 : : */
6 : : #include <linux/list.h>
7 : : #include <linux/init.h>
8 : : #include <linux/mm.h>
9 : : #include <linux/seq_file.h>
10 : : #include <linux/sysctl.h>
11 : : #include <linux/highmem.h>
12 : : #include <linux/mmu_notifier.h>
13 : : #include <linux/nodemask.h>
14 : : #include <linux/pagemap.h>
15 : : #include <linux/mempolicy.h>
16 : : #include <linux/compiler.h>
17 : : #include <linux/cpuset.h>
18 : : #include <linux/mutex.h>
19 : : #include <linux/memblock.h>
20 : : #include <linux/sysfs.h>
21 : : #include <linux/slab.h>
22 : : #include <linux/mmdebug.h>
23 : : #include <linux/sched/signal.h>
24 : : #include <linux/rmap.h>
25 : : #include <linux/string_helpers.h>
26 : : #include <linux/swap.h>
27 : : #include <linux/swapops.h>
28 : : #include <linux/jhash.h>
29 : : #include <linux/numa.h>
30 : : #include <linux/llist.h>
31 : :
32 : : #include <asm/page.h>
33 : : #include <asm/pgtable.h>
34 : : #include <asm/tlb.h>
35 : :
36 : : #include <linux/io.h>
37 : : #include <linux/hugetlb.h>
38 : : #include <linux/hugetlb_cgroup.h>
39 : : #include <linux/node.h>
40 : : #include <linux/userfaultfd_k.h>
41 : : #include <linux/page_owner.h>
42 : : #include "internal.h"
43 : :
44 : : int hugetlb_max_hstate __read_mostly;
45 : : unsigned int default_hstate_idx;
46 : : struct hstate hstates[HUGE_MAX_HSTATE];
47 : : /*
48 : : * Minimum page order among possible hugepage sizes, set to a proper value
49 : : * at boot time.
50 : : */
51 : : static unsigned int minimum_order __read_mostly = UINT_MAX;
52 : :
53 : : __initdata LIST_HEAD(huge_boot_pages);
54 : :
55 : : /* for command line parsing */
56 : : static struct hstate * __initdata parsed_hstate;
57 : : static unsigned long __initdata default_hstate_max_huge_pages;
58 : : static unsigned long __initdata default_hstate_size;
59 : : static bool __initdata parsed_valid_hugepagesz = true;
60 : :
61 : : /*
62 : : * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
63 : : * free_huge_pages, and surplus_huge_pages.
64 : : */
65 : : DEFINE_SPINLOCK(hugetlb_lock);
66 : :
67 : : /*
68 : : * Serializes faults on the same logical page. This is used to
69 : : * prevent spurious OOMs when the hugepage pool is fully utilized.
70 : : */
71 : : static int num_fault_mutexes;
72 : : struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp;
73 : :
74 : : /* Forward declaration */
75 : : static int hugetlb_acct_memory(struct hstate *h, long delta);
76 : :
77 : 0 : static inline void unlock_or_release_subpool(struct hugepage_subpool *spool)
78 : : {
79 [ # # # # ]: 0 : bool free = (spool->count == 0) && (spool->used_hpages == 0);
80 : :
81 : 0 : spin_unlock(&spool->lock);
82 : :
83 : : /* If no pages are used, and no other handles to the subpool
84 : : * remain, give up any reservations mased on minimum size and
85 : : * free the subpool */
86 [ # # ]: 0 : if (free) {
87 [ # # ]: 0 : if (spool->min_hpages != -1)
88 : 0 : hugetlb_acct_memory(spool->hstate,
89 : : -spool->min_hpages);
90 : 0 : kfree(spool);
91 : : }
92 : 0 : }
93 : :
94 : 0 : struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages,
95 : : long min_hpages)
96 : : {
97 : 0 : struct hugepage_subpool *spool;
98 : :
99 : 0 : spool = kzalloc(sizeof(*spool), GFP_KERNEL);
100 [ # # ]: 0 : if (!spool)
101 : : return NULL;
102 : :
103 [ # # ]: 0 : spin_lock_init(&spool->lock);
104 : 0 : spool->count = 1;
105 : 0 : spool->max_hpages = max_hpages;
106 : 0 : spool->hstate = h;
107 : 0 : spool->min_hpages = min_hpages;
108 : :
109 [ # # # # ]: 0 : if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) {
110 : 0 : kfree(spool);
111 : 0 : return NULL;
112 : : }
113 : 0 : spool->rsv_hpages = min_hpages;
114 : :
115 : 0 : return spool;
116 : : }
117 : :
118 : 0 : void hugepage_put_subpool(struct hugepage_subpool *spool)
119 : : {
120 : 0 : spin_lock(&spool->lock);
121 [ # # ]: 0 : BUG_ON(!spool->count);
122 : 0 : spool->count--;
123 : 0 : unlock_or_release_subpool(spool);
124 : 0 : }
125 : :
126 : : /*
127 : : * Subpool accounting for allocating and reserving pages.
128 : : * Return -ENOMEM if there are not enough resources to satisfy the
129 : : * the request. Otherwise, return the number of pages by which the
130 : : * global pools must be adjusted (upward). The returned value may
131 : : * only be different than the passed value (delta) in the case where
132 : : * a subpool minimum size must be manitained.
133 : : */
134 : 0 : static long hugepage_subpool_get_pages(struct hugepage_subpool *spool,
135 : : long delta)
136 : : {
137 : 0 : long ret = delta;
138 : :
139 [ # # ]: 0 : if (!spool)
140 : : return ret;
141 : :
142 : 0 : spin_lock(&spool->lock);
143 : :
144 [ # # ]: 0 : if (spool->max_hpages != -1) { /* maximum size accounting */
145 [ # # ]: 0 : if ((spool->used_hpages + delta) <= spool->max_hpages)
146 : 0 : spool->used_hpages += delta;
147 : : else {
148 : 0 : ret = -ENOMEM;
149 : 0 : goto unlock_ret;
150 : : }
151 : : }
152 : :
153 : : /* minimum size accounting */
154 [ # # # # ]: 0 : if (spool->min_hpages != -1 && spool->rsv_hpages) {
155 [ # # ]: 0 : if (delta > spool->rsv_hpages) {
156 : : /*
157 : : * Asking for more reserves than those already taken on
158 : : * behalf of subpool. Return difference.
159 : : */
160 : 0 : ret = delta - spool->rsv_hpages;
161 : 0 : spool->rsv_hpages = 0;
162 : : } else {
163 : 0 : ret = 0; /* reserves already accounted for */
164 : 0 : spool->rsv_hpages -= delta;
165 : : }
166 : : }
167 : :
168 : 0 : unlock_ret:
169 : 0 : spin_unlock(&spool->lock);
170 : 0 : return ret;
171 : : }
172 : :
173 : : /*
174 : : * Subpool accounting for freeing and unreserving pages.
175 : : * Return the number of global page reservations that must be dropped.
176 : : * The return value may only be different than the passed value (delta)
177 : : * in the case where a subpool minimum size must be maintained.
178 : : */
179 : 0 : static long hugepage_subpool_put_pages(struct hugepage_subpool *spool,
180 : : long delta)
181 : : {
182 : 0 : long ret = delta;
183 : :
184 [ # # ]: 0 : if (!spool)
185 : : return delta;
186 : :
187 : 0 : spin_lock(&spool->lock);
188 : :
189 [ # # ]: 0 : if (spool->max_hpages != -1) /* maximum size accounting */
190 : 0 : spool->used_hpages -= delta;
191 : :
192 : : /* minimum size accounting */
193 [ # # # # ]: 0 : if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) {
194 [ # # ]: 0 : if (spool->rsv_hpages + delta <= spool->min_hpages)
195 : : ret = 0;
196 : : else
197 : 0 : ret = spool->rsv_hpages + delta - spool->min_hpages;
198 : :
199 : 0 : spool->rsv_hpages += delta;
200 [ # # ]: 0 : if (spool->rsv_hpages > spool->min_hpages)
201 : 0 : spool->rsv_hpages = spool->min_hpages;
202 : : }
203 : :
204 : : /*
205 : : * If hugetlbfs_put_super couldn't free spool due to an outstanding
206 : : * quota reference, free it now.
207 : : */
208 : 0 : unlock_or_release_subpool(spool);
209 : :
210 : 0 : return ret;
211 : : }
212 : :
213 : 0 : static inline struct hugepage_subpool *subpool_inode(struct inode *inode)
214 : : {
215 [ # # ]: 0 : return HUGETLBFS_SB(inode->i_sb)->spool;
216 : : }
217 : :
218 : 0 : static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma)
219 : : {
220 : 0 : return subpool_inode(file_inode(vma->vm_file));
221 : : }
222 : :
223 : : /*
224 : : * Region tracking -- allows tracking of reservations and instantiated pages
225 : : * across the pages in a mapping.
226 : : *
227 : : * The region data structures are embedded into a resv_map and protected
228 : : * by a resv_map's lock. The set of regions within the resv_map represent
229 : : * reservations for huge pages, or huge pages that have already been
230 : : * instantiated within the map. The from and to elements are huge page
231 : : * indicies into the associated mapping. from indicates the starting index
232 : : * of the region. to represents the first index past the end of the region.
233 : : *
234 : : * For example, a file region structure with from == 0 and to == 4 represents
235 : : * four huge pages in a mapping. It is important to note that the to element
236 : : * represents the first element past the end of the region. This is used in
237 : : * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
238 : : *
239 : : * Interval notation of the form [from, to) will be used to indicate that
240 : : * the endpoint from is inclusive and to is exclusive.
241 : : */
242 : : struct file_region {
243 : : struct list_head link;
244 : : long from;
245 : : long to;
246 : : };
247 : :
248 : : /* Must be called with resv->lock held. Calling this with count_only == true
249 : : * will count the number of pages to be added but will not modify the linked
250 : : * list.
251 : : */
252 : 0 : static long add_reservation_in_range(struct resv_map *resv, long f, long t,
253 : : bool count_only)
254 : : {
255 : 0 : long chg = 0;
256 : 0 : struct list_head *head = &resv->regions;
257 : 0 : struct file_region *rg = NULL, *trg = NULL, *nrg = NULL;
258 : :
259 : : /* Locate the region we are before or in. */
260 [ # # ]: 0 : list_for_each_entry(rg, head, link)
261 [ # # ]: 0 : if (f <= rg->to)
262 : : break;
263 : :
264 : : /* Round our left edge to the current segment if it encloses us. */
265 : 0 : if (f > rg->from)
266 : : f = rg->from;
267 : :
268 : 0 : chg = t - f;
269 : :
270 : : /* Check for and consume any regions we now overlap with. */
271 : 0 : nrg = rg;
272 [ # # ]: 0 : list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
273 [ # # ]: 0 : if (&rg->link == head)
274 : : break;
275 [ # # ]: 0 : if (rg->from > t)
276 : : break;
277 : :
278 : : /* We overlap with this area, if it extends further than
279 : : * us then we must extend ourselves. Account for its
280 : : * existing reservation.
281 : : */
282 [ # # ]: 0 : if (rg->to > t) {
283 : 0 : chg += rg->to - t;
284 : 0 : t = rg->to;
285 : : }
286 : 0 : chg -= rg->to - rg->from;
287 : :
288 [ # # ]: 0 : if (!count_only && rg != nrg) {
289 : 0 : list_del(&rg->link);
290 : 0 : kfree(rg);
291 : : }
292 : : }
293 : :
294 [ # # ]: 0 : if (!count_only) {
295 : 0 : nrg->from = f;
296 : 0 : nrg->to = t;
297 : : }
298 : :
299 : 0 : return chg;
300 : : }
301 : :
302 : : /*
303 : : * Add the huge page range represented by [f, t) to the reserve
304 : : * map. Existing regions will be expanded to accommodate the specified
305 : : * range, or a region will be taken from the cache. Sufficient regions
306 : : * must exist in the cache due to the previous call to region_chg with
307 : : * the same range.
308 : : *
309 : : * Return the number of new huge pages added to the map. This
310 : : * number is greater than or equal to zero.
311 : : */
312 : 0 : static long region_add(struct resv_map *resv, long f, long t)
313 : : {
314 : 0 : struct list_head *head = &resv->regions;
315 : 0 : struct file_region *rg, *nrg;
316 : 0 : long add = 0;
317 : :
318 : 0 : spin_lock(&resv->lock);
319 : : /* Locate the region we are either in or before. */
320 [ # # ]: 0 : list_for_each_entry(rg, head, link)
321 [ # # ]: 0 : if (f <= rg->to)
322 : : break;
323 : :
324 : : /*
325 : : * If no region exists which can be expanded to include the
326 : : * specified range, pull a region descriptor from the cache
327 : : * and use it for this range.
328 : : */
329 [ # # # # ]: 0 : if (&rg->link == head || t < rg->from) {
330 : 0 : VM_BUG_ON(resv->region_cache_count <= 0);
331 : :
332 : 0 : resv->region_cache_count--;
333 : 0 : nrg = list_first_entry(&resv->region_cache, struct file_region,
334 : : link);
335 : 0 : list_del(&nrg->link);
336 : :
337 : 0 : nrg->from = f;
338 : 0 : nrg->to = t;
339 : 0 : list_add(&nrg->link, rg->link.prev);
340 : :
341 : 0 : add += t - f;
342 : 0 : goto out_locked;
343 : : }
344 : :
345 : 0 : add = add_reservation_in_range(resv, f, t, false);
346 : :
347 : 0 : out_locked:
348 : 0 : resv->adds_in_progress--;
349 : 0 : spin_unlock(&resv->lock);
350 : 0 : VM_BUG_ON(add < 0);
351 : 0 : return add;
352 : : }
353 : :
354 : : /*
355 : : * Examine the existing reserve map and determine how many
356 : : * huge pages in the specified range [f, t) are NOT currently
357 : : * represented. This routine is called before a subsequent
358 : : * call to region_add that will actually modify the reserve
359 : : * map to add the specified range [f, t). region_chg does
360 : : * not change the number of huge pages represented by the
361 : : * map. A new file_region structure is added to the cache
362 : : * as a placeholder, so that the subsequent region_add
363 : : * call will have all the regions it needs and will not fail.
364 : : *
365 : : * Returns the number of huge pages that need to be added to the existing
366 : : * reservation map for the range [f, t). This number is greater or equal to
367 : : * zero. -ENOMEM is returned if a new file_region structure or cache entry
368 : : * is needed and can not be allocated.
369 : : */
370 : 0 : static long region_chg(struct resv_map *resv, long f, long t)
371 : : {
372 : 0 : long chg = 0;
373 : :
374 : 0 : spin_lock(&resv->lock);
375 : 0 : retry_locked:
376 : 0 : resv->adds_in_progress++;
377 : :
378 : : /*
379 : : * Check for sufficient descriptors in the cache to accommodate
380 : : * the number of in progress add operations.
381 : : */
382 [ # # ]: 0 : if (resv->adds_in_progress > resv->region_cache_count) {
383 : 0 : struct file_region *trg;
384 : :
385 : 0 : VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1);
386 : : /* Must drop lock to allocate a new descriptor. */
387 : 0 : resv->adds_in_progress--;
388 : 0 : spin_unlock(&resv->lock);
389 : :
390 : 0 : trg = kmalloc(sizeof(*trg), GFP_KERNEL);
391 [ # # ]: 0 : if (!trg)
392 : : return -ENOMEM;
393 : :
394 : 0 : spin_lock(&resv->lock);
395 : 0 : list_add(&trg->link, &resv->region_cache);
396 : 0 : resv->region_cache_count++;
397 : 0 : goto retry_locked;
398 : : }
399 : :
400 : 0 : chg = add_reservation_in_range(resv, f, t, true);
401 : :
402 : 0 : spin_unlock(&resv->lock);
403 : 0 : return chg;
404 : : }
405 : :
406 : : /*
407 : : * Abort the in progress add operation. The adds_in_progress field
408 : : * of the resv_map keeps track of the operations in progress between
409 : : * calls to region_chg and region_add. Operations are sometimes
410 : : * aborted after the call to region_chg. In such cases, region_abort
411 : : * is called to decrement the adds_in_progress counter.
412 : : *
413 : : * NOTE: The range arguments [f, t) are not needed or used in this
414 : : * routine. They are kept to make reading the calling code easier as
415 : : * arguments will match the associated region_chg call.
416 : : */
417 : 0 : static void region_abort(struct resv_map *resv, long f, long t)
418 : : {
419 : 0 : spin_lock(&resv->lock);
420 : 0 : VM_BUG_ON(!resv->region_cache_count);
421 : 0 : resv->adds_in_progress--;
422 : 0 : spin_unlock(&resv->lock);
423 : 0 : }
424 : :
425 : : /*
426 : : * Delete the specified range [f, t) from the reserve map. If the
427 : : * t parameter is LONG_MAX, this indicates that ALL regions after f
428 : : * should be deleted. Locate the regions which intersect [f, t)
429 : : * and either trim, delete or split the existing regions.
430 : : *
431 : : * Returns the number of huge pages deleted from the reserve map.
432 : : * In the normal case, the return value is zero or more. In the
433 : : * case where a region must be split, a new region descriptor must
434 : : * be allocated. If the allocation fails, -ENOMEM will be returned.
435 : : * NOTE: If the parameter t == LONG_MAX, then we will never split
436 : : * a region and possibly return -ENOMEM. Callers specifying
437 : : * t == LONG_MAX do not need to check for -ENOMEM error.
438 : : */
439 : 0 : static long region_del(struct resv_map *resv, long f, long t)
440 : : {
441 : 0 : struct list_head *head = &resv->regions;
442 : 0 : struct file_region *rg, *trg;
443 : 0 : struct file_region *nrg = NULL;
444 : 0 : long del = 0;
445 : :
446 : 0 : retry:
447 : 0 : spin_lock(&resv->lock);
448 [ # # ]: 0 : list_for_each_entry_safe(rg, trg, head, link) {
449 : : /*
450 : : * Skip regions before the range to be deleted. file_region
451 : : * ranges are normally of the form [from, to). However, there
452 : : * may be a "placeholder" entry in the map which is of the form
453 : : * (from, to) with from == to. Check for placeholder entries
454 : : * at the beginning of the range to be deleted.
455 : : */
456 [ # # # # : 0 : if (rg->to <= f && (rg->to != rg->from || rg->to != f))
# # ]
457 : 0 : continue;
458 : :
459 [ # # ]: 0 : if (rg->from >= t)
460 : : break;
461 : :
462 [ # # # # ]: 0 : if (f > rg->from && t < rg->to) { /* Must split region */
463 : : /*
464 : : * Check for an entry in the cache before dropping
465 : : * lock and attempting allocation.
466 : : */
467 [ # # ]: 0 : if (!nrg &&
468 [ # # ]: 0 : resv->region_cache_count > resv->adds_in_progress) {
469 : 0 : nrg = list_first_entry(&resv->region_cache,
470 : : struct file_region,
471 : : link);
472 : 0 : list_del(&nrg->link);
473 : 0 : resv->region_cache_count--;
474 : : }
475 : :
476 [ # # ]: 0 : if (!nrg) {
477 : 0 : spin_unlock(&resv->lock);
478 : 0 : nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
479 [ # # ]: 0 : if (!nrg)
480 : : return -ENOMEM;
481 : 0 : goto retry;
482 : : }
483 : :
484 : 0 : del += t - f;
485 : :
486 : : /* New entry for end of split region */
487 : 0 : nrg->from = t;
488 : 0 : nrg->to = rg->to;
489 : 0 : INIT_LIST_HEAD(&nrg->link);
490 : :
491 : : /* Original entry is trimmed */
492 : 0 : rg->to = f;
493 : :
494 : 0 : list_add(&nrg->link, &rg->link);
495 : 0 : nrg = NULL;
496 : 0 : break;
497 : : }
498 : :
499 [ # # # # ]: 0 : if (f <= rg->from && t >= rg->to) { /* Remove entire region */
500 : 0 : del += rg->to - rg->from;
501 : 0 : list_del(&rg->link);
502 : 0 : kfree(rg);
503 : 0 : continue;
504 : : }
505 : :
506 [ # # ]: 0 : if (f <= rg->from) { /* Trim beginning of region */
507 : 0 : del += t - rg->from;
508 : 0 : rg->from = t;
509 : : } else { /* Trim end of region */
510 : 0 : del += rg->to - f;
511 : 0 : rg->to = f;
512 : : }
513 : : }
514 : :
515 : 0 : spin_unlock(&resv->lock);
516 : 0 : kfree(nrg);
517 : 0 : return del;
518 : : }
519 : :
520 : : /*
521 : : * A rare out of memory error was encountered which prevented removal of
522 : : * the reserve map region for a page. The huge page itself was free'ed
523 : : * and removed from the page cache. This routine will adjust the subpool
524 : : * usage count, and the global reserve count if needed. By incrementing
525 : : * these counts, the reserve map entry which could not be deleted will
526 : : * appear as a "reserved" entry instead of simply dangling with incorrect
527 : : * counts.
528 : : */
529 : 0 : void hugetlb_fix_reserve_counts(struct inode *inode)
530 : : {
531 : 0 : struct hugepage_subpool *spool = subpool_inode(inode);
532 : 0 : long rsv_adjust;
533 : :
534 : 0 : rsv_adjust = hugepage_subpool_get_pages(spool, 1);
535 [ # # ]: 0 : if (rsv_adjust) {
536 : 0 : struct hstate *h = hstate_inode(inode);
537 : :
538 : 0 : hugetlb_acct_memory(h, 1);
539 : : }
540 : 0 : }
541 : :
542 : : /*
543 : : * Count and return the number of huge pages in the reserve map
544 : : * that intersect with the range [f, t).
545 : : */
546 : 0 : static long region_count(struct resv_map *resv, long f, long t)
547 : : {
548 : 0 : struct list_head *head = &resv->regions;
549 : 0 : struct file_region *rg;
550 : 0 : long chg = 0;
551 : :
552 : 0 : spin_lock(&resv->lock);
553 : : /* Locate each segment we overlap with, and count that overlap. */
554 [ # # ]: 0 : list_for_each_entry(rg, head, link) {
555 : 0 : long seg_from;
556 : 0 : long seg_to;
557 : :
558 [ # # ]: 0 : if (rg->to <= f)
559 : 0 : continue;
560 [ # # ]: 0 : if (rg->from >= t)
561 : : break;
562 : :
563 : 0 : seg_from = max(rg->from, f);
564 : 0 : seg_to = min(rg->to, t);
565 : :
566 : 0 : chg += seg_to - seg_from;
567 : : }
568 : 0 : spin_unlock(&resv->lock);
569 : :
570 : 0 : return chg;
571 : : }
572 : :
573 : : /*
574 : : * Convert the address within this vma to the page offset within
575 : : * the mapping, in pagecache page units; huge pages here.
576 : : */
577 : 0 : static pgoff_t vma_hugecache_offset(struct hstate *h,
578 : : struct vm_area_struct *vma, unsigned long address)
579 : : {
580 : 0 : return ((address - vma->vm_start) >> huge_page_shift(h)) +
581 : 0 : (vma->vm_pgoff >> huge_page_order(h));
582 : : }
583 : :
584 : 0 : pgoff_t linear_hugepage_index(struct vm_area_struct *vma,
585 : : unsigned long address)
586 : : {
587 : 0 : return vma_hugecache_offset(hstate_vma(vma), vma, address);
588 : : }
589 : : EXPORT_SYMBOL_GPL(linear_hugepage_index);
590 : :
591 : : /*
592 : : * Return the size of the pages allocated when backing a VMA. In the majority
593 : : * cases this will be same size as used by the page table entries.
594 : : */
595 : 0 : unsigned long vma_kernel_pagesize(struct vm_area_struct *vma)
596 : : {
597 [ # # # # : 0 : if (vma->vm_ops && vma->vm_ops->pagesize)
# # ]
598 : 0 : return vma->vm_ops->pagesize(vma);
599 : : return PAGE_SIZE;
600 : : }
601 : : EXPORT_SYMBOL_GPL(vma_kernel_pagesize);
602 : :
603 : : /*
604 : : * Return the page size being used by the MMU to back a VMA. In the majority
605 : : * of cases, the page size used by the kernel matches the MMU size. On
606 : : * architectures where it differs, an architecture-specific 'strong'
607 : : * version of this symbol is required.
608 : : */
609 : 0 : __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
610 : : {
611 [ # # ]: 0 : return vma_kernel_pagesize(vma);
612 : : }
613 : :
614 : : /*
615 : : * Flags for MAP_PRIVATE reservations. These are stored in the bottom
616 : : * bits of the reservation map pointer, which are always clear due to
617 : : * alignment.
618 : : */
619 : : #define HPAGE_RESV_OWNER (1UL << 0)
620 : : #define HPAGE_RESV_UNMAPPED (1UL << 1)
621 : : #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
622 : :
623 : : /*
624 : : * These helpers are used to track how many pages are reserved for
625 : : * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
626 : : * is guaranteed to have their future faults succeed.
627 : : *
628 : : * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
629 : : * the reserve counters are updated with the hugetlb_lock held. It is safe
630 : : * to reset the VMA at fork() time as it is not in use yet and there is no
631 : : * chance of the global counters getting corrupted as a result of the values.
632 : : *
633 : : * The private mapping reservation is represented in a subtly different
634 : : * manner to a shared mapping. A shared mapping has a region map associated
635 : : * with the underlying file, this region map represents the backing file
636 : : * pages which have ever had a reservation assigned which this persists even
637 : : * after the page is instantiated. A private mapping has a region map
638 : : * associated with the original mmap which is attached to all VMAs which
639 : : * reference it, this region map represents those offsets which have consumed
640 : : * reservation ie. where pages have been instantiated.
641 : : */
642 : 0 : static unsigned long get_vma_private_data(struct vm_area_struct *vma)
643 : : {
644 : 0 : return (unsigned long)vma->vm_private_data;
645 : : }
646 : :
647 : 0 : static void set_vma_private_data(struct vm_area_struct *vma,
648 : : unsigned long value)
649 : : {
650 : 0 : vma->vm_private_data = (void *)value;
651 : : }
652 : :
653 : 0 : struct resv_map *resv_map_alloc(void)
654 : : {
655 : 0 : struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
656 : 0 : struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL);
657 : :
658 [ # # ]: 0 : if (!resv_map || !rg) {
659 : 0 : kfree(resv_map);
660 : 0 : kfree(rg);
661 : 0 : return NULL;
662 : : }
663 : :
664 : 0 : kref_init(&resv_map->refs);
665 : 0 : spin_lock_init(&resv_map->lock);
666 : 0 : INIT_LIST_HEAD(&resv_map->regions);
667 : :
668 : 0 : resv_map->adds_in_progress = 0;
669 : :
670 : 0 : INIT_LIST_HEAD(&resv_map->region_cache);
671 : 0 : list_add(&rg->link, &resv_map->region_cache);
672 : 0 : resv_map->region_cache_count = 1;
673 : :
674 : 0 : return resv_map;
675 : : }
676 : :
677 : 0 : void resv_map_release(struct kref *ref)
678 : : {
679 : 0 : struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
680 : 0 : struct list_head *head = &resv_map->region_cache;
681 : 0 : struct file_region *rg, *trg;
682 : :
683 : : /* Clear out any active regions before we release the map. */
684 : 0 : region_del(resv_map, 0, LONG_MAX);
685 : :
686 : : /* ... and any entries left in the cache */
687 [ # # ]: 0 : list_for_each_entry_safe(rg, trg, head, link) {
688 : 0 : list_del(&rg->link);
689 : 0 : kfree(rg);
690 : : }
691 : :
692 : 0 : VM_BUG_ON(resv_map->adds_in_progress);
693 : :
694 : 0 : kfree(resv_map);
695 : 0 : }
696 : :
697 : 0 : static inline struct resv_map *inode_resv_map(struct inode *inode)
698 : : {
699 : : /*
700 : : * At inode evict time, i_mapping may not point to the original
701 : : * address space within the inode. This original address space
702 : : * contains the pointer to the resv_map. So, always use the
703 : : * address space embedded within the inode.
704 : : * The VERY common case is inode->mapping == &inode->i_data but,
705 : : * this may not be true for device special inodes.
706 : : */
707 : 0 : return (struct resv_map *)(&inode->i_data)->private_data;
708 : : }
709 : :
710 : 0 : static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
711 : : {
712 : 0 : VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
713 : 0 : if (vma->vm_flags & VM_MAYSHARE) {
714 : 0 : struct address_space *mapping = vma->vm_file->f_mapping;
715 : 0 : struct inode *inode = mapping->host;
716 : :
717 : 0 : return inode_resv_map(inode);
718 : :
719 : : } else {
720 : 0 : return (struct resv_map *)(get_vma_private_data(vma) &
721 : : ~HPAGE_RESV_MASK);
722 : : }
723 : : }
724 : :
725 : 0 : static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
726 : : {
727 : 0 : VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
728 : 0 : VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
729 : :
730 : 0 : set_vma_private_data(vma, (get_vma_private_data(vma) &
731 : 0 : HPAGE_RESV_MASK) | (unsigned long)map);
732 : : }
733 : :
734 : 0 : static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
735 : : {
736 : 0 : VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
737 : 0 : VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma);
738 : :
739 : 0 : set_vma_private_data(vma, get_vma_private_data(vma) | flags);
740 : 0 : }
741 : :
742 : 0 : static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
743 : : {
744 : 0 : VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
745 : :
746 : 0 : return (get_vma_private_data(vma) & flag) != 0;
747 : : }
748 : :
749 : : /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
750 : 0 : void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
751 : : {
752 : 0 : VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma);
753 [ # # ]: 0 : if (!(vma->vm_flags & VM_MAYSHARE))
754 : 0 : vma->vm_private_data = (void *)0;
755 : 0 : }
756 : :
757 : : /* Returns true if the VMA has associated reserve pages */
758 : 0 : static bool vma_has_reserves(struct vm_area_struct *vma, long chg)
759 : : {
760 [ # # ]: 0 : if (vma->vm_flags & VM_NORESERVE) {
761 : : /*
762 : : * This address is already reserved by other process(chg == 0),
763 : : * so, we should decrement reserved count. Without decrementing,
764 : : * reserve count remains after releasing inode, because this
765 : : * allocated page will go into page cache and is regarded as
766 : : * coming from reserved pool in releasing step. Currently, we
767 : : * don't have any other solution to deal with this situation
768 : : * properly, so add work-around here.
769 : : */
770 [ # # # # ]: 0 : if (vma->vm_flags & VM_MAYSHARE && chg == 0)
771 : : return true;
772 : : else
773 : 0 : return false;
774 : : }
775 : :
776 : : /* Shared mappings always use reserves */
777 [ # # ]: 0 : if (vma->vm_flags & VM_MAYSHARE) {
778 : : /*
779 : : * We know VM_NORESERVE is not set. Therefore, there SHOULD
780 : : * be a region map for all pages. The only situation where
781 : : * there is no region map is if a hole was punched via
782 : : * fallocate. In this case, there really are no reverves to
783 : : * use. This situation is indicated if chg != 0.
784 : : */
785 [ # # ]: 0 : if (chg)
786 : : return false;
787 : : else
788 : 0 : return true;
789 : : }
790 : :
791 : : /*
792 : : * Only the process that called mmap() has reserves for
793 : : * private mappings.
794 : : */
795 [ # # ]: 0 : if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
796 : : /*
797 : : * Like the shared case above, a hole punch or truncate
798 : : * could have been performed on the private mapping.
799 : : * Examine the value of chg to determine if reserves
800 : : * actually exist or were previously consumed.
801 : : * Very Subtle - The value of chg comes from a previous
802 : : * call to vma_needs_reserves(). The reserve map for
803 : : * private mappings has different (opposite) semantics
804 : : * than that of shared mappings. vma_needs_reserves()
805 : : * has already taken this difference in semantics into
806 : : * account. Therefore, the meaning of chg is the same
807 : : * as in the shared case above. Code could easily be
808 : : * combined, but keeping it separate draws attention to
809 : : * subtle differences.
810 : : */
811 [ # # ]: 0 : if (chg)
812 : : return false;
813 : : else
814 : 0 : return true;
815 : : }
816 : :
817 : : return false;
818 : : }
819 : :
820 : 0 : static void enqueue_huge_page(struct hstate *h, struct page *page)
821 : : {
822 : 0 : int nid = page_to_nid(page);
823 : 0 : list_move(&page->lru, &h->hugepage_freelists[nid]);
824 : 0 : h->free_huge_pages++;
825 : 0 : h->free_huge_pages_node[nid]++;
826 : 0 : }
827 : :
828 : 0 : static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid)
829 : : {
830 : 0 : struct page *page;
831 : :
832 : 0 : list_for_each_entry(page, &h->hugepage_freelists[nid], lru)
833 : : if (!PageHWPoison(page))
834 : : break;
835 : : /*
836 : : * if 'non-isolated free hugepage' not found on the list,
837 : : * the allocation fails.
838 : : */
839 [ # # ]: 0 : if (&h->hugepage_freelists[nid] == &page->lru)
840 : : return NULL;
841 : 0 : list_move(&page->lru, &h->hugepage_activelist);
842 : 0 : set_page_refcounted(page);
843 : 0 : h->free_huge_pages--;
844 : 0 : h->free_huge_pages_node[nid]--;
845 : 0 : return page;
846 : : }
847 : :
848 : 0 : static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid,
849 : : nodemask_t *nmask)
850 : : {
851 : 0 : unsigned int cpuset_mems_cookie;
852 : 0 : struct zonelist *zonelist;
853 : 0 : struct zone *zone;
854 : 0 : struct zoneref *z;
855 : 0 : int node = NUMA_NO_NODE;
856 : :
857 [ # # ]: 0 : zonelist = node_zonelist(nid, gfp_mask);
858 : :
859 : 0 : retry_cpuset:
860 : 0 : cpuset_mems_cookie = read_mems_allowed_begin();
861 [ # # # # : 0 : for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) {
# # ]
862 : 0 : struct page *page;
863 : :
864 [ # # ]: 0 : if (!cpuset_zone_allowed(zone, gfp_mask))
865 : 0 : continue;
866 : : /*
867 : : * no need to ask again on the same node. Pool is node rather than
868 : : * zone aware
869 : : */
870 [ # # ]: 0 : if (zone_to_nid(zone) == node)
871 : 0 : continue;
872 : 0 : node = zone_to_nid(zone);
873 : :
874 : 0 : page = dequeue_huge_page_node_exact(h, node);
875 [ # # ]: 0 : if (page)
876 : 0 : return page;
877 : : }
878 [ # # # # ]: 0 : if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie)))
879 : 0 : goto retry_cpuset;
880 : :
881 : : return NULL;
882 : : }
883 : :
884 : : /* Movability of hugepages depends on migration support. */
885 : 0 : static inline gfp_t htlb_alloc_mask(struct hstate *h)
886 : : {
887 : 0 : if (hugepage_movable_supported(h))
888 : : return GFP_HIGHUSER_MOVABLE;
889 : : else
890 : 0 : return GFP_HIGHUSER;
891 : : }
892 : :
893 : 0 : static struct page *dequeue_huge_page_vma(struct hstate *h,
894 : : struct vm_area_struct *vma,
895 : : unsigned long address, int avoid_reserve,
896 : : long chg)
897 : : {
898 : 0 : struct page *page;
899 : 0 : struct mempolicy *mpol;
900 : 0 : gfp_t gfp_mask;
901 : 0 : nodemask_t *nodemask;
902 : 0 : int nid;
903 : :
904 : : /*
905 : : * A child process with MAP_PRIVATE mappings created by their parent
906 : : * have no page reserves. This check ensures that reservations are
907 : : * not "stolen". The child may still get SIGKILLed
908 : : */
909 [ # # ]: 0 : if (!vma_has_reserves(vma, chg) &&
910 [ # # ]: 0 : h->free_huge_pages - h->resv_huge_pages == 0)
911 : 0 : goto err;
912 : :
913 : : /* If reserves cannot be used, ensure enough pages are in the pool */
914 [ # # # # ]: 0 : if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
915 : 0 : goto err;
916 : :
917 [ # # ]: 0 : gfp_mask = htlb_alloc_mask(h);
918 : 0 : nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
919 : 0 : page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask);
920 [ # # # # ]: 0 : if (page && !avoid_reserve && vma_has_reserves(vma, chg)) {
921 : 0 : SetPagePrivate(page);
922 : 0 : h->resv_huge_pages--;
923 : : }
924 : :
925 [ # # ]: 0 : mpol_cond_put(mpol);
926 : : return page;
927 : :
928 : 0 : err:
929 : : return NULL;
930 : : }
931 : :
932 : : /*
933 : : * common helper functions for hstate_next_node_to_{alloc|free}.
934 : : * We may have allocated or freed a huge page based on a different
935 : : * nodes_allowed previously, so h->next_node_to_{alloc|free} might
936 : : * be outside of *nodes_allowed. Ensure that we use an allowed
937 : : * node for alloc or free.
938 : : */
939 : 0 : static int next_node_allowed(int nid, nodemask_t *nodes_allowed)
940 : : {
941 : 0 : nid = next_node_in(nid, *nodes_allowed);
942 : 0 : VM_BUG_ON(nid >= MAX_NUMNODES);
943 : :
944 : 0 : return nid;
945 : : }
946 : :
947 : 0 : static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed)
948 : : {
949 [ # # ]: 0 : if (!node_isset(nid, *nodes_allowed))
950 : 0 : nid = next_node_allowed(nid, nodes_allowed);
951 : 0 : return nid;
952 : : }
953 : :
954 : : /*
955 : : * returns the previously saved node ["this node"] from which to
956 : : * allocate a persistent huge page for the pool and advance the
957 : : * next node from which to allocate, handling wrap at end of node
958 : : * mask.
959 : : */
960 : : static int hstate_next_node_to_alloc(struct hstate *h,
961 : : nodemask_t *nodes_allowed)
962 : : {
963 : : int nid;
964 : :
965 : : VM_BUG_ON(!nodes_allowed);
966 : :
967 : : nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed);
968 : : h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed);
969 : :
970 : : return nid;
971 : : }
972 : :
973 : : /*
974 : : * helper for free_pool_huge_page() - return the previously saved
975 : : * node ["this node"] from which to free a huge page. Advance the
976 : : * next node id whether or not we find a free huge page to free so
977 : : * that the next attempt to free addresses the next node.
978 : : */
979 : : static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed)
980 : : {
981 : : int nid;
982 : :
983 : : VM_BUG_ON(!nodes_allowed);
984 : :
985 : : nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed);
986 : : h->next_nid_to_free = next_node_allowed(nid, nodes_allowed);
987 : :
988 : : return nid;
989 : : }
990 : :
991 : : #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
992 : : for (nr_nodes = nodes_weight(*mask); \
993 : : nr_nodes > 0 && \
994 : : ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
995 : : nr_nodes--)
996 : :
997 : : #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
998 : : for (nr_nodes = nodes_weight(*mask); \
999 : : nr_nodes > 0 && \
1000 : : ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1001 : : nr_nodes--)
1002 : :
1003 : : #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE
1004 : 0 : static void destroy_compound_gigantic_page(struct page *page,
1005 : : unsigned int order)
1006 : : {
1007 : 0 : int i;
1008 : 0 : int nr_pages = 1 << order;
1009 : 0 : struct page *p = page + 1;
1010 : :
1011 : 0 : atomic_set(compound_mapcount_ptr(page), 0);
1012 [ # # ]: 0 : for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1013 : 0 : clear_compound_head(p);
1014 : 0 : set_page_refcounted(p);
1015 : : }
1016 : :
1017 : 0 : set_compound_order(page, 0);
1018 : 0 : __ClearPageHead(page);
1019 : 0 : }
1020 : :
1021 : 0 : static void free_gigantic_page(struct page *page, unsigned int order)
1022 : : {
1023 : 0 : free_contig_range(page_to_pfn(page), 1 << order);
1024 : 0 : }
1025 : :
1026 : : #ifdef CONFIG_CONTIG_ALLOC
1027 : : static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1028 : : int nid, nodemask_t *nodemask)
1029 : : {
1030 : : unsigned long nr_pages = 1UL << huge_page_order(h);
1031 : :
1032 : : return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask);
1033 : : }
1034 : :
1035 : : static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
1036 : : static void prep_compound_gigantic_page(struct page *page, unsigned int order);
1037 : : #else /* !CONFIG_CONTIG_ALLOC */
1038 : : static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1039 : : int nid, nodemask_t *nodemask)
1040 : : {
1041 : : return NULL;
1042 : : }
1043 : : #endif /* CONFIG_CONTIG_ALLOC */
1044 : :
1045 : : #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */
1046 : : static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask,
1047 : : int nid, nodemask_t *nodemask)
1048 : : {
1049 : : return NULL;
1050 : : }
1051 : : static inline void free_gigantic_page(struct page *page, unsigned int order) { }
1052 : : static inline void destroy_compound_gigantic_page(struct page *page,
1053 : : unsigned int order) { }
1054 : : #endif
1055 : :
1056 : 0 : static void update_and_free_page(struct hstate *h, struct page *page)
1057 : : {
1058 : 0 : int i;
1059 : :
1060 : 0 : if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
1061 : : return;
1062 : :
1063 : 0 : h->nr_huge_pages--;
1064 : 0 : h->nr_huge_pages_node[page_to_nid(page)]--;
1065 [ # # ]: 0 : for (i = 0; i < pages_per_huge_page(h); i++) {
1066 : 0 : page[i].flags &= ~(1 << PG_locked | 1 << PG_error |
1067 : : 1 << PG_referenced | 1 << PG_dirty |
1068 : : 1 << PG_active | 1 << PG_private |
1069 : : 1 << PG_writeback);
1070 : : }
1071 : 0 : VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
1072 : 0 : set_compound_page_dtor(page, NULL_COMPOUND_DTOR);
1073 : 0 : set_page_refcounted(page);
1074 [ # # ]: 0 : if (hstate_is_gigantic(h)) {
1075 : 0 : destroy_compound_gigantic_page(page, huge_page_order(h));
1076 : 0 : free_gigantic_page(page, huge_page_order(h));
1077 : : } else {
1078 : 0 : __free_pages(page, huge_page_order(h));
1079 : : }
1080 : : }
1081 : :
1082 : 0 : struct hstate *size_to_hstate(unsigned long size)
1083 : : {
1084 : 52 : struct hstate *h;
1085 : :
1086 [ - + - + : 52 : for_each_hstate(h) {
+ - - - -
- - + ]
1087 [ - - - - : 13 : if (huge_page_size(h) == size)
- + - - -
- - - ]
1088 : 0 : return h;
1089 : : }
1090 : : return NULL;
1091 : : }
1092 : :
1093 : : /*
1094 : : * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1095 : : * to hstate->hugepage_activelist.)
1096 : : *
1097 : : * This function can be called for tail pages, but never returns true for them.
1098 : : */
1099 : 0 : bool page_huge_active(struct page *page)
1100 : : {
1101 : 0 : VM_BUG_ON_PAGE(!PageHuge(page), page);
1102 [ # # # # ]: 0 : return PageHead(page) && PagePrivate(&page[1]);
1103 : : }
1104 : :
1105 : : /* never called for tail page */
1106 : 0 : static void set_page_huge_active(struct page *page)
1107 : : {
1108 : 0 : VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1109 : 0 : SetPagePrivate(&page[1]);
1110 : 0 : }
1111 : :
1112 : 0 : static void clear_page_huge_active(struct page *page)
1113 : : {
1114 : 0 : VM_BUG_ON_PAGE(!PageHeadHuge(page), page);
1115 : 0 : ClearPagePrivate(&page[1]);
1116 : : }
1117 : :
1118 : : /*
1119 : : * Internal hugetlb specific page flag. Do not use outside of the hugetlb
1120 : : * code
1121 : : */
1122 : 0 : static inline bool PageHugeTemporary(struct page *page)
1123 : : {
1124 [ # # # # ]: 0 : if (!PageHuge(page))
1125 : : return false;
1126 : :
1127 : 0 : return (unsigned long)page[2].mapping == -1U;
1128 : : }
1129 : :
1130 : 0 : static inline void SetPageHugeTemporary(struct page *page)
1131 : : {
1132 : 0 : page[2].mapping = (void *)-1U;
1133 : : }
1134 : :
1135 : 0 : static inline void ClearPageHugeTemporary(struct page *page)
1136 : : {
1137 : 0 : page[2].mapping = NULL;
1138 : : }
1139 : :
1140 : 0 : static void __free_huge_page(struct page *page)
1141 : : {
1142 : : /*
1143 : : * Can't pass hstate in here because it is called from the
1144 : : * compound page destructor.
1145 : : */
1146 : 0 : struct hstate *h = page_hstate(page);
1147 : 0 : int nid = page_to_nid(page);
1148 : 0 : struct hugepage_subpool *spool =
1149 : 0 : (struct hugepage_subpool *)page_private(page);
1150 : 0 : bool restore_reserve;
1151 : :
1152 : 0 : VM_BUG_ON_PAGE(page_count(page), page);
1153 : 0 : VM_BUG_ON_PAGE(page_mapcount(page), page);
1154 : :
1155 : 0 : set_page_private(page, 0);
1156 : 0 : page->mapping = NULL;
1157 : 0 : restore_reserve = PagePrivate(page);
1158 : 0 : ClearPagePrivate(page);
1159 : :
1160 : : /*
1161 : : * If PagePrivate() was set on page, page allocation consumed a
1162 : : * reservation. If the page was associated with a subpool, there
1163 : : * would have been a page reserved in the subpool before allocation
1164 : : * via hugepage_subpool_get_pages(). Since we are 'restoring' the
1165 : : * reservtion, do not call hugepage_subpool_put_pages() as this will
1166 : : * remove the reserved page from the subpool.
1167 : : */
1168 [ # # ]: 0 : if (!restore_reserve) {
1169 : : /*
1170 : : * A return code of zero implies that the subpool will be
1171 : : * under its minimum size if the reservation is not restored
1172 : : * after page is free. Therefore, force restore_reserve
1173 : : * operation.
1174 : : */
1175 [ # # ]: 0 : if (hugepage_subpool_put_pages(spool, 1) == 0)
1176 : 0 : restore_reserve = true;
1177 : : }
1178 : :
1179 : 0 : spin_lock(&hugetlb_lock);
1180 : 0 : clear_page_huge_active(page);
1181 [ # # ]: 0 : hugetlb_cgroup_uncharge_page(hstate_index(h),
1182 : : pages_per_huge_page(h), page);
1183 [ # # ]: 0 : if (restore_reserve)
1184 : 0 : h->resv_huge_pages++;
1185 : :
1186 [ # # ]: 0 : if (PageHugeTemporary(page)) {
1187 : 0 : list_del(&page->lru);
1188 : 0 : ClearPageHugeTemporary(page);
1189 : 0 : update_and_free_page(h, page);
1190 [ # # ]: 0 : } else if (h->surplus_huge_pages_node[nid]) {
1191 : : /* remove the page from active list */
1192 : 0 : list_del(&page->lru);
1193 : 0 : update_and_free_page(h, page);
1194 : 0 : h->surplus_huge_pages--;
1195 : 0 : h->surplus_huge_pages_node[nid]--;
1196 : : } else {
1197 : 0 : arch_clear_hugepage_flags(page);
1198 : 0 : enqueue_huge_page(h, page);
1199 : : }
1200 : 0 : spin_unlock(&hugetlb_lock);
1201 : 0 : }
1202 : :
1203 : : /*
1204 : : * As free_huge_page() can be called from a non-task context, we have
1205 : : * to defer the actual freeing in a workqueue to prevent potential
1206 : : * hugetlb_lock deadlock.
1207 : : *
1208 : : * free_hpage_workfn() locklessly retrieves the linked list of pages to
1209 : : * be freed and frees them one-by-one. As the page->mapping pointer is
1210 : : * going to be cleared in __free_huge_page() anyway, it is reused as the
1211 : : * llist_node structure of a lockless linked list of huge pages to be freed.
1212 : : */
1213 : : static LLIST_HEAD(hpage_freelist);
1214 : :
1215 : 0 : static void free_hpage_workfn(struct work_struct *work)
1216 : : {
1217 : 0 : struct llist_node *node;
1218 : 0 : struct page *page;
1219 : :
1220 : 0 : node = llist_del_all(&hpage_freelist);
1221 : :
1222 [ # # ]: 0 : while (node) {
1223 : 0 : page = container_of((struct address_space **)node,
1224 : : struct page, mapping);
1225 : 0 : node = node->next;
1226 : 0 : __free_huge_page(page);
1227 : : }
1228 : 0 : }
1229 : : static DECLARE_WORK(free_hpage_work, free_hpage_workfn);
1230 : :
1231 : 0 : void free_huge_page(struct page *page)
1232 : : {
1233 : : /*
1234 : : * Defer freeing if in non-task context to avoid hugetlb_lock deadlock.
1235 : : */
1236 [ # # ]: 0 : if (!in_task()) {
1237 : : /*
1238 : : * Only call schedule_work() if hpage_freelist is previously
1239 : : * empty. Otherwise, schedule_work() had been called but the
1240 : : * workfn hasn't retrieved the list yet.
1241 : : */
1242 [ # # ]: 0 : if (llist_add((struct llist_node *)&page->mapping,
1243 : : &hpage_freelist))
1244 : 0 : schedule_work(&free_hpage_work);
1245 : 0 : return;
1246 : : }
1247 : :
1248 : 0 : __free_huge_page(page);
1249 : : }
1250 : :
1251 : 0 : static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
1252 : : {
1253 : 0 : INIT_LIST_HEAD(&page->lru);
1254 : 0 : set_compound_page_dtor(page, HUGETLB_PAGE_DTOR);
1255 : 0 : spin_lock(&hugetlb_lock);
1256 : 0 : set_hugetlb_cgroup(page, NULL);
1257 : 0 : h->nr_huge_pages++;
1258 : 0 : h->nr_huge_pages_node[nid]++;
1259 : 0 : spin_unlock(&hugetlb_lock);
1260 : : }
1261 : :
1262 : 0 : static void prep_compound_gigantic_page(struct page *page, unsigned int order)
1263 : : {
1264 : 0 : int i;
1265 : 0 : int nr_pages = 1 << order;
1266 : 0 : struct page *p = page + 1;
1267 : :
1268 : : /* we rely on prep_new_huge_page to set the destructor */
1269 : 0 : set_compound_order(page, order);
1270 : 0 : __ClearPageReserved(page);
1271 : 0 : __SetPageHead(page);
1272 [ # # ]: 0 : for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) {
1273 : : /*
1274 : : * For gigantic hugepages allocated through bootmem at
1275 : : * boot, it's safer to be consistent with the not-gigantic
1276 : : * hugepages and clear the PG_reserved bit from all tail pages
1277 : : * too. Otherwse drivers using get_user_pages() to access tail
1278 : : * pages may get the reference counting wrong if they see
1279 : : * PG_reserved set on a tail page (despite the head page not
1280 : : * having PG_reserved set). Enforcing this consistency between
1281 : : * head and tail pages allows drivers to optimize away a check
1282 : : * on the head page when they need know if put_page() is needed
1283 : : * after get_user_pages().
1284 : : */
1285 : 0 : __ClearPageReserved(p);
1286 : 0 : set_page_count(p, 0);
1287 : 0 : set_compound_head(p, page);
1288 : : }
1289 : 0 : atomic_set(compound_mapcount_ptr(page), -1);
1290 : 0 : }
1291 : :
1292 : : /*
1293 : : * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1294 : : * transparent huge pages. See the PageTransHuge() documentation for more
1295 : : * details.
1296 : : */
1297 : 15714630 : int PageHuge(struct page *page)
1298 : : {
1299 [ - + ]: 31429260 : if (!PageCompound(page))
1300 : : return 0;
1301 : :
1302 [ # # ]: 0 : page = compound_head(page);
1303 : 0 : return page[1].compound_dtor == HUGETLB_PAGE_DTOR;
1304 : : }
1305 : : EXPORT_SYMBOL_GPL(PageHuge);
1306 : :
1307 : : /*
1308 : : * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1309 : : * normal or transparent huge pages.
1310 : : */
1311 : 0 : int PageHeadHuge(struct page *page_head)
1312 : : {
1313 [ # # ]: 0 : if (!PageHead(page_head))
1314 : : return 0;
1315 : :
1316 : 0 : return get_compound_page_dtor(page_head) == free_huge_page;
1317 : : }
1318 : :
1319 : 0 : pgoff_t __basepage_index(struct page *page)
1320 : : {
1321 [ # # ]: 0 : struct page *page_head = compound_head(page);
1322 : 0 : pgoff_t index = page_index(page_head);
1323 : 0 : unsigned long compound_idx;
1324 : :
1325 [ # # ]: 0 : if (!PageHuge(page_head))
1326 : 0 : return page_index(page);
1327 : :
1328 [ # # ]: 0 : if (compound_order(page_head) >= MAX_ORDER)
1329 : 0 : compound_idx = page_to_pfn(page) - page_to_pfn(page_head);
1330 : : else
1331 : 0 : compound_idx = page - page_head;
1332 : :
1333 : 0 : return (index << compound_order(page_head)) + compound_idx;
1334 : : }
1335 : :
1336 : 0 : static struct page *alloc_buddy_huge_page(struct hstate *h,
1337 : : gfp_t gfp_mask, int nid, nodemask_t *nmask,
1338 : : nodemask_t *node_alloc_noretry)
1339 : : {
1340 [ # # ]: 0 : int order = huge_page_order(h);
1341 : 0 : struct page *page;
1342 : 0 : bool alloc_try_hard = true;
1343 : :
1344 : : /*
1345 : : * By default we always try hard to allocate the page with
1346 : : * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in
1347 : : * a loop (to adjust global huge page counts) and previous allocation
1348 : : * failed, do not continue to try hard on the same node. Use the
1349 : : * node_alloc_noretry bitmap to manage this state information.
1350 : : */
1351 [ # # # # ]: 0 : if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
1352 : 0 : alloc_try_hard = false;
1353 : 0 : gfp_mask |= __GFP_COMP|__GFP_NOWARN;
1354 : 0 : if (alloc_try_hard)
1355 : 0 : gfp_mask |= __GFP_RETRY_MAYFAIL;
1356 [ # # ]: 0 : if (nid == NUMA_NO_NODE)
1357 : 0 : nid = numa_mem_id();
1358 : 0 : page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
1359 [ # # ]: 0 : if (page)
1360 : 0 : __count_vm_event(HTLB_BUDDY_PGALLOC);
1361 : : else
1362 : 0 : __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
1363 : :
1364 : : /*
1365 : : * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
1366 : : * indicates an overall state change. Clear bit so that we resume
1367 : : * normal 'try hard' allocations.
1368 : : */
1369 [ # # # # ]: 0 : if (node_alloc_noretry && page && !alloc_try_hard)
1370 : 0 : node_clear(nid, *node_alloc_noretry);
1371 : :
1372 : : /*
1373 : : * If we tried hard to get a page but failed, set bit so that
1374 : : * subsequent attempts will not try as hard until there is an
1375 : : * overall state change.
1376 : : */
1377 [ # # # # ]: 0 : if (node_alloc_noretry && !page && alloc_try_hard)
1378 : 0 : node_set(nid, *node_alloc_noretry);
1379 : :
1380 : 0 : return page;
1381 : : }
1382 : :
1383 : : /*
1384 : : * Common helper to allocate a fresh hugetlb page. All specific allocators
1385 : : * should use this function to get new hugetlb pages
1386 : : */
1387 : 0 : static struct page *alloc_fresh_huge_page(struct hstate *h,
1388 : : gfp_t gfp_mask, int nid, nodemask_t *nmask,
1389 : : nodemask_t *node_alloc_noretry)
1390 : : {
1391 : 0 : struct page *page;
1392 : :
1393 [ # # ]: 0 : if (hstate_is_gigantic(h))
1394 : : page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
1395 : : else
1396 : 0 : page = alloc_buddy_huge_page(h, gfp_mask,
1397 : : nid, nmask, node_alloc_noretry);
1398 [ # # ]: 0 : if (!page)
1399 : 0 : return NULL;
1400 : :
1401 [ # # ]: 0 : if (hstate_is_gigantic(h))
1402 : 0 : prep_compound_gigantic_page(page, huge_page_order(h));
1403 : 0 : prep_new_huge_page(h, page, page_to_nid(page));
1404 : :
1405 : 0 : return page;
1406 : : }
1407 : :
1408 : : /*
1409 : : * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
1410 : : * manner.
1411 : : */
1412 : 0 : static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1413 : : nodemask_t *node_alloc_noretry)
1414 : : {
1415 : 0 : struct page *page;
1416 : 0 : int nr_nodes, node;
1417 [ # # ]: 0 : gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
1418 : :
1419 [ # # ]: 0 : for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
1420 : 0 : page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
1421 : : node_alloc_noretry);
1422 [ # # ]: 0 : if (page)
1423 : : break;
1424 : : }
1425 : :
1426 [ # # ]: 0 : if (!page)
1427 : : return 0;
1428 : :
1429 : 0 : put_page(page); /* free it into the hugepage allocator */
1430 : :
1431 : 0 : return 1;
1432 : : }
1433 : :
1434 : : /*
1435 : : * Free huge page from pool from next node to free.
1436 : : * Attempt to keep persistent huge pages more or less
1437 : : * balanced over allowed nodes.
1438 : : * Called with hugetlb_lock locked.
1439 : : */
1440 : 0 : static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
1441 : : bool acct_surplus)
1442 : : {
1443 : 0 : int nr_nodes, node;
1444 : 0 : int ret = 0;
1445 : :
1446 [ # # ]: 0 : for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
1447 : : /*
1448 : : * If we're returning unused surplus pages, only examine
1449 : : * nodes with surplus pages.
1450 : : */
1451 [ # # # # : 0 : if ((!acct_surplus || h->surplus_huge_pages_node[node]) &&
# # ]
1452 [ # # ]: 0 : !list_empty(&h->hugepage_freelists[node])) {
1453 : 0 : struct page *page =
1454 : 0 : list_entry(h->hugepage_freelists[node].next,
1455 : : struct page, lru);
1456 [ # # ]: 0 : list_del(&page->lru);
1457 : 0 : h->free_huge_pages--;
1458 : 0 : h->free_huge_pages_node[node]--;
1459 [ # # ]: 0 : if (acct_surplus) {
1460 : 0 : h->surplus_huge_pages--;
1461 : 0 : h->surplus_huge_pages_node[node]--;
1462 : : }
1463 : 0 : update_and_free_page(h, page);
1464 : 0 : ret = 1;
1465 : 0 : break;
1466 : : }
1467 : : }
1468 : :
1469 : 0 : return ret;
1470 : : }
1471 : :
1472 : : /*
1473 : : * Dissolve a given free hugepage into free buddy pages. This function does
1474 : : * nothing for in-use hugepages and non-hugepages.
1475 : : * This function returns values like below:
1476 : : *
1477 : : * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use
1478 : : * (allocated or reserved.)
1479 : : * 0: successfully dissolved free hugepages or the page is not a
1480 : : * hugepage (considered as already dissolved)
1481 : : */
1482 : 0 : int dissolve_free_huge_page(struct page *page)
1483 : : {
1484 : 0 : int rc = -EBUSY;
1485 : :
1486 : : /* Not to disrupt normal path by vainly holding hugetlb_lock */
1487 [ # # ]: 0 : if (!PageHuge(page))
1488 : : return 0;
1489 : :
1490 : 0 : spin_lock(&hugetlb_lock);
1491 [ # # ]: 0 : if (!PageHuge(page)) {
1492 : 0 : rc = 0;
1493 : 0 : goto out;
1494 : : }
1495 : :
1496 [ # # # # ]: 0 : if (!page_count(page)) {
1497 [ # # ]: 0 : struct page *head = compound_head(page);
1498 : 0 : struct hstate *h = page_hstate(head);
1499 [ # # ]: 0 : int nid = page_to_nid(head);
1500 [ # # ]: 0 : if (h->free_huge_pages - h->resv_huge_pages == 0)
1501 : 0 : goto out;
1502 : : /*
1503 : : * Move PageHWPoison flag from head page to the raw error page,
1504 : : * which makes any subpages rather than the error page reusable.
1505 : : */
1506 : 0 : if (PageHWPoison(head) && page != head) {
1507 : : SetPageHWPoison(page);
1508 : : ClearPageHWPoison(head);
1509 : : }
1510 : 0 : list_del(&head->lru);
1511 : 0 : h->free_huge_pages--;
1512 : 0 : h->free_huge_pages_node[nid]--;
1513 : 0 : h->max_huge_pages--;
1514 : 0 : update_and_free_page(h, head);
1515 : 0 : rc = 0;
1516 : : }
1517 : 0 : out:
1518 : 0 : spin_unlock(&hugetlb_lock);
1519 : 0 : return rc;
1520 : : }
1521 : :
1522 : : /*
1523 : : * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1524 : : * make specified memory blocks removable from the system.
1525 : : * Note that this will dissolve a free gigantic hugepage completely, if any
1526 : : * part of it lies within the given range.
1527 : : * Also note that if dissolve_free_huge_page() returns with an error, all
1528 : : * free hugepages that were dissolved before that error are lost.
1529 : : */
1530 : 0 : int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn)
1531 : : {
1532 : 0 : unsigned long pfn;
1533 : 0 : struct page *page;
1534 : 0 : int rc = 0;
1535 : :
1536 [ # # ]: 0 : if (!hugepages_supported())
1537 : : return rc;
1538 : :
1539 [ # # ]: 0 : for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) {
1540 : 0 : page = pfn_to_page(pfn);
1541 : 0 : rc = dissolve_free_huge_page(page);
1542 [ # # ]: 0 : if (rc)
1543 : : break;
1544 : : }
1545 : :
1546 : : return rc;
1547 : : }
1548 : :
1549 : : /*
1550 : : * Allocates a fresh surplus page from the page allocator.
1551 : : */
1552 : 0 : static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
1553 : : int nid, nodemask_t *nmask)
1554 : : {
1555 : 0 : struct page *page = NULL;
1556 : :
1557 [ # # ]: 0 : if (hstate_is_gigantic(h))
1558 : : return NULL;
1559 : :
1560 : 0 : spin_lock(&hugetlb_lock);
1561 [ # # ]: 0 : if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages)
1562 : 0 : goto out_unlock;
1563 : 0 : spin_unlock(&hugetlb_lock);
1564 : :
1565 : 0 : page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1566 [ # # ]: 0 : if (!page)
1567 : : return NULL;
1568 : :
1569 : 0 : spin_lock(&hugetlb_lock);
1570 : : /*
1571 : : * We could have raced with the pool size change.
1572 : : * Double check that and simply deallocate the new page
1573 : : * if we would end up overcommiting the surpluses. Abuse
1574 : : * temporary page to workaround the nasty free_huge_page
1575 : : * codeflow
1576 : : */
1577 [ # # ]: 0 : if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
1578 : 0 : SetPageHugeTemporary(page);
1579 : 0 : spin_unlock(&hugetlb_lock);
1580 : 0 : put_page(page);
1581 : 0 : return NULL;
1582 : : } else {
1583 : 0 : h->surplus_huge_pages++;
1584 : 0 : h->surplus_huge_pages_node[page_to_nid(page)]++;
1585 : : }
1586 : :
1587 : 0 : out_unlock:
1588 : 0 : spin_unlock(&hugetlb_lock);
1589 : :
1590 : 0 : return page;
1591 : : }
1592 : :
1593 : 0 : struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
1594 : : int nid, nodemask_t *nmask)
1595 : : {
1596 : 0 : struct page *page;
1597 : :
1598 [ # # ]: 0 : if (hstate_is_gigantic(h))
1599 : : return NULL;
1600 : :
1601 : 0 : page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
1602 [ # # # # : 0 : if (!page)
# # ]
1603 : : return NULL;
1604 : :
1605 : : /*
1606 : : * We do not account these pages as surplus because they are only
1607 : : * temporary and will be released properly on the last reference
1608 : : */
1609 : 0 : SetPageHugeTemporary(page);
1610 : :
1611 : 0 : return page;
1612 : : }
1613 : :
1614 : : /*
1615 : : * Use the VMA's mpolicy to allocate a huge page from the buddy.
1616 : : */
1617 : : static
1618 : 0 : struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h,
1619 : : struct vm_area_struct *vma, unsigned long addr)
1620 : : {
1621 : 0 : struct page *page;
1622 : 0 : struct mempolicy *mpol;
1623 [ # # ]: 0 : gfp_t gfp_mask = htlb_alloc_mask(h);
1624 : 0 : int nid;
1625 : 0 : nodemask_t *nodemask;
1626 : :
1627 : 0 : nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask);
1628 : 0 : page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask);
1629 [ # # ]: 0 : mpol_cond_put(mpol);
1630 : :
1631 : 0 : return page;
1632 : : }
1633 : :
1634 : : /* page migration callback function */
1635 : 0 : struct page *alloc_huge_page_node(struct hstate *h, int nid)
1636 : : {
1637 [ # # ]: 0 : gfp_t gfp_mask = htlb_alloc_mask(h);
1638 : 0 : struct page *page = NULL;
1639 : :
1640 [ # # ]: 0 : if (nid != NUMA_NO_NODE)
1641 : 0 : gfp_mask |= __GFP_THISNODE;
1642 : :
1643 : 0 : spin_lock(&hugetlb_lock);
1644 [ # # ]: 0 : if (h->free_huge_pages - h->resv_huge_pages > 0)
1645 : 0 : page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL);
1646 : 0 : spin_unlock(&hugetlb_lock);
1647 : :
1648 [ # # ]: 0 : if (!page)
1649 [ # # ]: 0 : page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL);
1650 : :
1651 : 0 : return page;
1652 : : }
1653 : :
1654 : : /* page migration callback function */
1655 : 0 : struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid,
1656 : : nodemask_t *nmask)
1657 : : {
1658 [ # # ]: 0 : gfp_t gfp_mask = htlb_alloc_mask(h);
1659 : :
1660 : 0 : spin_lock(&hugetlb_lock);
1661 [ # # ]: 0 : if (h->free_huge_pages - h->resv_huge_pages > 0) {
1662 : 0 : struct page *page;
1663 : :
1664 : 0 : page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask);
1665 [ # # ]: 0 : if (page) {
1666 : 0 : spin_unlock(&hugetlb_lock);
1667 : 0 : return page;
1668 : : }
1669 : : }
1670 : 0 : spin_unlock(&hugetlb_lock);
1671 : :
1672 [ # # ]: 0 : return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask);
1673 : : }
1674 : :
1675 : : /* mempolicy aware migration callback */
1676 : 0 : struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma,
1677 : : unsigned long address)
1678 : : {
1679 : 0 : struct mempolicy *mpol;
1680 : 0 : nodemask_t *nodemask;
1681 : 0 : struct page *page;
1682 : 0 : gfp_t gfp_mask;
1683 : 0 : int node;
1684 : :
1685 [ # # ]: 0 : gfp_mask = htlb_alloc_mask(h);
1686 : 0 : node = huge_node(vma, address, gfp_mask, &mpol, &nodemask);
1687 : 0 : page = alloc_huge_page_nodemask(h, node, nodemask);
1688 [ # # ]: 0 : mpol_cond_put(mpol);
1689 : :
1690 : 0 : return page;
1691 : : }
1692 : :
1693 : : /*
1694 : : * Increase the hugetlb pool such that it can accommodate a reservation
1695 : : * of size 'delta'.
1696 : : */
1697 : 0 : static int gather_surplus_pages(struct hstate *h, int delta)
1698 : : {
1699 : 0 : struct list_head surplus_list;
1700 : 0 : struct page *page, *tmp;
1701 : 0 : int ret, i;
1702 : 0 : int needed, allocated;
1703 : 0 : bool alloc_ok = true;
1704 : :
1705 : 0 : needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
1706 [ # # ]: 0 : if (needed <= 0) {
1707 : 0 : h->resv_huge_pages += delta;
1708 : 0 : return 0;
1709 : : }
1710 : :
1711 : 0 : allocated = 0;
1712 : 0 : INIT_LIST_HEAD(&surplus_list);
1713 : :
1714 : 0 : ret = -ENOMEM;
1715 : 0 : retry:
1716 : 0 : spin_unlock(&hugetlb_lock);
1717 [ # # ]: 0 : for (i = 0; i < needed; i++) {
1718 [ # # ]: 0 : page = alloc_surplus_huge_page(h, htlb_alloc_mask(h),
1719 : : NUMA_NO_NODE, NULL);
1720 [ # # ]: 0 : if (!page) {
1721 : : alloc_ok = false;
1722 : : break;
1723 : : }
1724 : 0 : list_add(&page->lru, &surplus_list);
1725 : 0 : cond_resched();
1726 : : }
1727 : 0 : allocated += i;
1728 : :
1729 : : /*
1730 : : * After retaking hugetlb_lock, we need to recalculate 'needed'
1731 : : * because either resv_huge_pages or free_huge_pages may have changed.
1732 : : */
1733 : 0 : spin_lock(&hugetlb_lock);
1734 : 0 : needed = (h->resv_huge_pages + delta) -
1735 : 0 : (h->free_huge_pages + allocated);
1736 [ # # ]: 0 : if (needed > 0) {
1737 [ # # ]: 0 : if (alloc_ok)
1738 : 0 : goto retry;
1739 : : /*
1740 : : * We were not able to allocate enough pages to
1741 : : * satisfy the entire reservation so we free what
1742 : : * we've allocated so far.
1743 : : */
1744 : 0 : goto free;
1745 : : }
1746 : : /*
1747 : : * The surplus_list now contains _at_least_ the number of extra pages
1748 : : * needed to accommodate the reservation. Add the appropriate number
1749 : : * of pages to the hugetlb pool and free the extras back to the buddy
1750 : : * allocator. Commit the entire reservation here to prevent another
1751 : : * process from stealing the pages as they are added to the pool but
1752 : : * before they are reserved.
1753 : : */
1754 : 0 : needed += allocated;
1755 : 0 : h->resv_huge_pages += delta;
1756 : 0 : ret = 0;
1757 : :
1758 : : /* Free the needed pages to the hugetlb pool */
1759 [ # # ]: 0 : list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
1760 [ # # ]: 0 : if ((--needed) < 0)
1761 : : break;
1762 : : /*
1763 : : * This page is now managed by the hugetlb allocator and has
1764 : : * no users -- drop the buddy allocator's reference.
1765 : : */
1766 : 0 : put_page_testzero(page);
1767 : 0 : VM_BUG_ON_PAGE(page_count(page), page);
1768 : 0 : enqueue_huge_page(h, page);
1769 : : }
1770 : 0 : free:
1771 : 0 : spin_unlock(&hugetlb_lock);
1772 : :
1773 : : /* Free unnecessary surplus pages to the buddy allocator */
1774 [ # # ]: 0 : list_for_each_entry_safe(page, tmp, &surplus_list, lru)
1775 : 0 : put_page(page);
1776 : 0 : spin_lock(&hugetlb_lock);
1777 : :
1778 : 0 : return ret;
1779 : : }
1780 : :
1781 : : /*
1782 : : * This routine has two main purposes:
1783 : : * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1784 : : * in unused_resv_pages. This corresponds to the prior adjustments made
1785 : : * to the associated reservation map.
1786 : : * 2) Free any unused surplus pages that may have been allocated to satisfy
1787 : : * the reservation. As many as unused_resv_pages may be freed.
1788 : : *
1789 : : * Called with hugetlb_lock held. However, the lock could be dropped (and
1790 : : * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1791 : : * we must make sure nobody else can claim pages we are in the process of
1792 : : * freeing. Do this by ensuring resv_huge_page always is greater than the
1793 : : * number of huge pages we plan to free when dropping the lock.
1794 : : */
1795 : 0 : static void return_unused_surplus_pages(struct hstate *h,
1796 : : unsigned long unused_resv_pages)
1797 : : {
1798 : 0 : unsigned long nr_pages;
1799 : :
1800 : : /* Cannot return gigantic pages currently */
1801 [ # # ]: 0 : if (hstate_is_gigantic(h))
1802 : 0 : goto out;
1803 : :
1804 : : /*
1805 : : * Part (or even all) of the reservation could have been backed
1806 : : * by pre-allocated pages. Only free surplus pages.
1807 : : */
1808 : 0 : nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
1809 : :
1810 : : /*
1811 : : * We want to release as many surplus pages as possible, spread
1812 : : * evenly across all nodes with memory. Iterate across these nodes
1813 : : * until we can no longer free unreserved surplus pages. This occurs
1814 : : * when the nodes with surplus pages have no free pages.
1815 : : * free_pool_huge_page() will balance the the freed pages across the
1816 : : * on-line nodes with memory and will handle the hstate accounting.
1817 : : *
1818 : : * Note that we decrement resv_huge_pages as we free the pages. If
1819 : : * we drop the lock, resv_huge_pages will still be sufficiently large
1820 : : * to cover subsequent pages we may free.
1821 : : */
1822 [ # # ]: 0 : while (nr_pages--) {
1823 : 0 : h->resv_huge_pages--;
1824 : 0 : unused_resv_pages--;
1825 [ # # ]: 0 : if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1))
1826 : 0 : goto out;
1827 : 0 : cond_resched_lock(&hugetlb_lock);
1828 : : }
1829 : :
1830 : 0 : out:
1831 : : /* Fully uncommit the reservation */
1832 : 0 : h->resv_huge_pages -= unused_resv_pages;
1833 : 0 : }
1834 : :
1835 : :
1836 : : /*
1837 : : * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1838 : : * are used by the huge page allocation routines to manage reservations.
1839 : : *
1840 : : * vma_needs_reservation is called to determine if the huge page at addr
1841 : : * within the vma has an associated reservation. If a reservation is
1842 : : * needed, the value 1 is returned. The caller is then responsible for
1843 : : * managing the global reservation and subpool usage counts. After
1844 : : * the huge page has been allocated, vma_commit_reservation is called
1845 : : * to add the page to the reservation map. If the page allocation fails,
1846 : : * the reservation must be ended instead of committed. vma_end_reservation
1847 : : * is called in such cases.
1848 : : *
1849 : : * In the normal case, vma_commit_reservation returns the same value
1850 : : * as the preceding vma_needs_reservation call. The only time this
1851 : : * is not the case is if a reserve map was changed between calls. It
1852 : : * is the responsibility of the caller to notice the difference and
1853 : : * take appropriate action.
1854 : : *
1855 : : * vma_add_reservation is used in error paths where a reservation must
1856 : : * be restored when a newly allocated huge page must be freed. It is
1857 : : * to be called after calling vma_needs_reservation to determine if a
1858 : : * reservation exists.
1859 : : */
1860 : : enum vma_resv_mode {
1861 : : VMA_NEEDS_RESV,
1862 : : VMA_COMMIT_RESV,
1863 : : VMA_END_RESV,
1864 : : VMA_ADD_RESV,
1865 : : };
1866 : 0 : static long __vma_reservation_common(struct hstate *h,
1867 : : struct vm_area_struct *vma, unsigned long addr,
1868 : : enum vma_resv_mode mode)
1869 : : {
1870 : 0 : struct resv_map *resv;
1871 : 0 : pgoff_t idx;
1872 : 0 : long ret;
1873 : :
1874 [ # # ]: 0 : resv = vma_resv_map(vma);
1875 [ # # ]: 0 : if (!resv)
1876 : : return 1;
1877 : :
1878 [ # # # # : 0 : idx = vma_hugecache_offset(h, vma, addr);
# ]
1879 [ # # # # : 0 : switch (mode) {
# ]
1880 : 0 : case VMA_NEEDS_RESV:
1881 : 0 : ret = region_chg(resv, idx, idx + 1);
1882 : 0 : break;
1883 : 0 : case VMA_COMMIT_RESV:
1884 : 0 : ret = region_add(resv, idx, idx + 1);
1885 : 0 : break;
1886 : 0 : case VMA_END_RESV:
1887 : 0 : region_abort(resv, idx, idx + 1);
1888 : 0 : ret = 0;
1889 : 0 : break;
1890 : 0 : case VMA_ADD_RESV:
1891 [ # # ]: 0 : if (vma->vm_flags & VM_MAYSHARE)
1892 : 0 : ret = region_add(resv, idx, idx + 1);
1893 : : else {
1894 : 0 : region_abort(resv, idx, idx + 1);
1895 : 0 : ret = region_del(resv, idx, idx + 1);
1896 : : }
1897 : : break;
1898 : 0 : default:
1899 : 0 : BUG();
1900 : : }
1901 : :
1902 [ # # ]: 0 : if (vma->vm_flags & VM_MAYSHARE)
1903 : : return ret;
1904 [ # # # # ]: 0 : else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) {
1905 : : /*
1906 : : * In most cases, reserves always exist for private mappings.
1907 : : * However, a file associated with mapping could have been
1908 : : * hole punched or truncated after reserves were consumed.
1909 : : * As subsequent fault on such a range will not use reserves.
1910 : : * Subtle - The reserve map for private mappings has the
1911 : : * opposite meaning than that of shared mappings. If NO
1912 : : * entry is in the reserve map, it means a reservation exists.
1913 : : * If an entry exists in the reserve map, it means the
1914 : : * reservation has already been consumed. As a result, the
1915 : : * return value of this routine is the opposite of the
1916 : : * value returned from reserve map manipulation routines above.
1917 : : */
1918 [ # # ]: 0 : if (ret)
1919 : : return 0;
1920 : : else
1921 : 0 : return 1;
1922 : : }
1923 : : else
1924 : 0 : return ret < 0 ? ret : 0;
1925 : : }
1926 : :
1927 : 0 : static long vma_needs_reservation(struct hstate *h,
1928 : : struct vm_area_struct *vma, unsigned long addr)
1929 : : {
1930 : 0 : return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV);
1931 : : }
1932 : :
1933 : 0 : static long vma_commit_reservation(struct hstate *h,
1934 : : struct vm_area_struct *vma, unsigned long addr)
1935 : : {
1936 : 0 : return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV);
1937 : : }
1938 : :
1939 : 0 : static void vma_end_reservation(struct hstate *h,
1940 : : struct vm_area_struct *vma, unsigned long addr)
1941 : : {
1942 : 0 : (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV);
1943 : 0 : }
1944 : :
1945 : 0 : static long vma_add_reservation(struct hstate *h,
1946 : : struct vm_area_struct *vma, unsigned long addr)
1947 : : {
1948 : 0 : return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV);
1949 : : }
1950 : :
1951 : : /*
1952 : : * This routine is called to restore a reservation on error paths. In the
1953 : : * specific error paths, a huge page was allocated (via alloc_huge_page)
1954 : : * and is about to be freed. If a reservation for the page existed,
1955 : : * alloc_huge_page would have consumed the reservation and set PagePrivate
1956 : : * in the newly allocated page. When the page is freed via free_huge_page,
1957 : : * the global reservation count will be incremented if PagePrivate is set.
1958 : : * However, free_huge_page can not adjust the reserve map. Adjust the
1959 : : * reserve map here to be consistent with global reserve count adjustments
1960 : : * to be made by free_huge_page.
1961 : : */
1962 : 0 : static void restore_reserve_on_error(struct hstate *h,
1963 : : struct vm_area_struct *vma, unsigned long address,
1964 : : struct page *page)
1965 : : {
1966 [ # # ]: 0 : if (unlikely(PagePrivate(page))) {
1967 : 0 : long rc = vma_needs_reservation(h, vma, address);
1968 : :
1969 [ # # ]: 0 : if (unlikely(rc < 0)) {
1970 : : /*
1971 : : * Rare out of memory condition in reserve map
1972 : : * manipulation. Clear PagePrivate so that
1973 : : * global reserve count will not be incremented
1974 : : * by free_huge_page. This will make it appear
1975 : : * as though the reservation for this page was
1976 : : * consumed. This may prevent the task from
1977 : : * faulting in the page at a later time. This
1978 : : * is better than inconsistent global huge page
1979 : : * accounting of reserve counts.
1980 : : */
1981 : 0 : ClearPagePrivate(page);
1982 [ # # ]: 0 : } else if (rc) {
1983 : 0 : rc = vma_add_reservation(h, vma, address);
1984 [ # # ]: 0 : if (unlikely(rc < 0))
1985 : : /*
1986 : : * See above comment about rare out of
1987 : : * memory condition.
1988 : : */
1989 : 0 : ClearPagePrivate(page);
1990 : : } else
1991 : 0 : vma_end_reservation(h, vma, address);
1992 : : }
1993 : 0 : }
1994 : :
1995 : 0 : struct page *alloc_huge_page(struct vm_area_struct *vma,
1996 : : unsigned long addr, int avoid_reserve)
1997 : : {
1998 : 0 : struct hugepage_subpool *spool = subpool_vma(vma);
1999 : 0 : struct hstate *h = hstate_vma(vma);
2000 : 0 : struct page *page;
2001 : 0 : long map_chg, map_commit;
2002 : 0 : long gbl_chg;
2003 : 0 : int ret, idx;
2004 : 0 : struct hugetlb_cgroup *h_cg;
2005 : :
2006 : 0 : idx = hstate_index(h);
2007 : : /*
2008 : : * Examine the region/reserve map to determine if the process
2009 : : * has a reservation for the page to be allocated. A return
2010 : : * code of zero indicates a reservation exists (no change).
2011 : : */
2012 : 0 : map_chg = gbl_chg = vma_needs_reservation(h, vma, addr);
2013 [ # # ]: 0 : if (map_chg < 0)
2014 : : return ERR_PTR(-ENOMEM);
2015 : :
2016 : : /*
2017 : : * Processes that did not create the mapping will have no
2018 : : * reserves as indicated by the region/reserve map. Check
2019 : : * that the allocation will not exceed the subpool limit.
2020 : : * Allocations for MAP_NORESERVE mappings also need to be
2021 : : * checked against any subpool limit.
2022 : : */
2023 [ # # ]: 0 : if (map_chg || avoid_reserve) {
2024 : 0 : gbl_chg = hugepage_subpool_get_pages(spool, 1);
2025 [ # # ]: 0 : if (gbl_chg < 0) {
2026 : 0 : vma_end_reservation(h, vma, addr);
2027 : 0 : return ERR_PTR(-ENOSPC);
2028 : : }
2029 : :
2030 : : /*
2031 : : * Even though there was no reservation in the region/reserve
2032 : : * map, there could be reservations associated with the
2033 : : * subpool that can be used. This would be indicated if the
2034 : : * return value of hugepage_subpool_get_pages() is zero.
2035 : : * However, if avoid_reserve is specified we still avoid even
2036 : : * the subpool reservations.
2037 : : */
2038 [ # # ]: 0 : if (avoid_reserve)
2039 : 0 : gbl_chg = 1;
2040 : : }
2041 : :
2042 : 0 : ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg);
2043 : 0 : if (ret)
2044 : : goto out_subpool_put;
2045 : :
2046 : 0 : spin_lock(&hugetlb_lock);
2047 : : /*
2048 : : * glb_chg is passed to indicate whether or not a page must be taken
2049 : : * from the global free pool (global change). gbl_chg == 0 indicates
2050 : : * a reservation exists for the allocation.
2051 : : */
2052 : 0 : page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg);
2053 [ # # ]: 0 : if (!page) {
2054 : 0 : spin_unlock(&hugetlb_lock);
2055 : 0 : page = alloc_buddy_huge_page_with_mpol(h, vma, addr);
2056 [ # # ]: 0 : if (!page)
2057 : 0 : goto out_uncharge_cgroup;
2058 [ # # # # ]: 0 : if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) {
2059 : 0 : SetPagePrivate(page);
2060 : 0 : h->resv_huge_pages--;
2061 : : }
2062 : 0 : spin_lock(&hugetlb_lock);
2063 : 0 : list_move(&page->lru, &h->hugepage_activelist);
2064 : : /* Fall through */
2065 : : }
2066 : 0 : hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page);
2067 : 0 : spin_unlock(&hugetlb_lock);
2068 : :
2069 : 0 : set_page_private(page, (unsigned long)spool);
2070 : :
2071 : 0 : map_commit = vma_commit_reservation(h, vma, addr);
2072 [ # # ]: 0 : if (unlikely(map_chg > map_commit)) {
2073 : : /*
2074 : : * The page was added to the reservation map between
2075 : : * vma_needs_reservation and vma_commit_reservation.
2076 : : * This indicates a race with hugetlb_reserve_pages.
2077 : : * Adjust for the subpool count incremented above AND
2078 : : * in hugetlb_reserve_pages for the same page. Also,
2079 : : * the reservation count added in hugetlb_reserve_pages
2080 : : * no longer applies.
2081 : : */
2082 : 0 : long rsv_adjust;
2083 : :
2084 : 0 : rsv_adjust = hugepage_subpool_put_pages(spool, 1);
2085 : 0 : hugetlb_acct_memory(h, -rsv_adjust);
2086 : : }
2087 : : return page;
2088 : :
2089 : : out_uncharge_cgroup:
2090 [ # # ]: 0 : hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg);
2091 : : out_subpool_put:
2092 [ # # ]: 0 : if (map_chg || avoid_reserve)
2093 : 0 : hugepage_subpool_put_pages(spool, 1);
2094 : 0 : vma_end_reservation(h, vma, addr);
2095 : 0 : return ERR_PTR(-ENOSPC);
2096 : : }
2097 : :
2098 : : int alloc_bootmem_huge_page(struct hstate *h)
2099 : : __attribute__ ((weak, alias("__alloc_bootmem_huge_page")));
2100 : 0 : int __alloc_bootmem_huge_page(struct hstate *h)
2101 : : {
2102 : 0 : struct huge_bootmem_page *m;
2103 : 0 : int nr_nodes, node;
2104 : :
2105 [ # # ]: 0 : for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) {
2106 : 0 : void *addr;
2107 : :
2108 : 0 : addr = memblock_alloc_try_nid_raw(
2109 : 0 : huge_page_size(h), huge_page_size(h),
2110 : : 0, MEMBLOCK_ALLOC_ACCESSIBLE, node);
2111 [ # # ]: 0 : if (addr) {
2112 : : /*
2113 : : * Use the beginning of the huge page to store the
2114 : : * huge_bootmem_page struct (until gather_bootmem
2115 : : * puts them into the mem_map).
2116 : : */
2117 : 0 : m = addr;
2118 : 0 : goto found;
2119 : : }
2120 : : }
2121 : : return 0;
2122 : :
2123 : : found:
2124 [ # # # # ]: 0 : BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h)));
2125 : : /* Put them into a private list first because mem_map is not up yet */
2126 : 0 : INIT_LIST_HEAD(&m->list);
2127 : 0 : list_add(&m->list, &huge_boot_pages);
2128 : 0 : m->hstate = h;
2129 : 0 : return 1;
2130 : : }
2131 : :
2132 : 0 : static void __init prep_compound_huge_page(struct page *page,
2133 : : unsigned int order)
2134 : : {
2135 [ # # ]: 0 : if (unlikely(order > (MAX_ORDER - 1)))
2136 : 0 : prep_compound_gigantic_page(page, order);
2137 : : else
2138 : 0 : prep_compound_page(page, order);
2139 : 0 : }
2140 : :
2141 : : /* Put bootmem huge pages into the standard lists after mem_map is up */
2142 : 13 : static void __init gather_bootmem_prealloc(void)
2143 : : {
2144 : 13 : struct huge_bootmem_page *m;
2145 : :
2146 [ - + ]: 13 : list_for_each_entry(m, &huge_boot_pages, list) {
2147 [ # # ]: 0 : struct page *page = virt_to_page(m);
2148 : 0 : struct hstate *h = m->hstate;
2149 : :
2150 [ # # # # ]: 0 : WARN_ON(page_count(page) != 1);
2151 : 0 : prep_compound_huge_page(page, h->order);
2152 [ # # ]: 0 : WARN_ON(PageReserved(page));
2153 : 0 : prep_new_huge_page(h, page, page_to_nid(page));
2154 : 0 : put_page(page); /* free it into the hugepage allocator */
2155 : :
2156 : : /*
2157 : : * If we had gigantic hugepages allocated at boot time, we need
2158 : : * to restore the 'stolen' pages to totalram_pages in order to
2159 : : * fix confusing memory reports from free(1) and another
2160 : : * side-effects, like CommitLimit going negative.
2161 : : */
2162 [ # # ]: 0 : if (hstate_is_gigantic(h))
2163 : 0 : adjust_managed_page_count(page, 1 << h->order);
2164 : 0 : cond_resched();
2165 : : }
2166 : 13 : }
2167 : :
2168 : 13 : static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
2169 : : {
2170 : 13 : unsigned long i;
2171 : 13 : nodemask_t *node_alloc_noretry;
2172 : :
2173 [ + - ]: 13 : if (!hstate_is_gigantic(h)) {
2174 : : /*
2175 : : * Bit mask controlling how hard we retry per-node allocations.
2176 : : * Ignore errors as lower level routines can deal with
2177 : : * node_alloc_noretry == NULL. If this kmalloc fails at boot
2178 : : * time, we are likely in bigger trouble.
2179 : : */
2180 : 13 : node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
2181 : : GFP_KERNEL);
2182 : : } else {
2183 : : /* allocations done at boot time */
2184 : : node_alloc_noretry = NULL;
2185 : : }
2186 : :
2187 : : /* bit mask controlling how hard we retry per-node allocations */
2188 [ + - ]: 13 : if (node_alloc_noretry)
2189 : 13 : nodes_clear(*node_alloc_noretry);
2190 : :
2191 [ - + ]: 13 : for (i = 0; i < h->max_huge_pages; ++i) {
2192 [ # # ]: 0 : if (hstate_is_gigantic(h)) {
2193 [ # # ]: 0 : if (!alloc_bootmem_huge_page(h))
2194 : : break;
2195 [ # # ]: 0 : } else if (!alloc_pool_huge_page(h,
2196 : : &node_states[N_MEMORY],
2197 : : node_alloc_noretry))
2198 : : break;
2199 : 0 : cond_resched();
2200 : : }
2201 [ - + ]: 13 : if (i < h->max_huge_pages) {
2202 : 0 : char buf[32];
2203 : :
2204 : 0 : string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2205 : 0 : pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n",
2206 : : h->max_huge_pages, buf, i);
2207 : 0 : h->max_huge_pages = i;
2208 : : }
2209 : :
2210 : 13 : kfree(node_alloc_noretry);
2211 : 13 : }
2212 : :
2213 : 13 : static void __init hugetlb_init_hstates(void)
2214 : : {
2215 : 13 : struct hstate *h;
2216 : :
2217 [ + + ]: 26 : for_each_hstate(h) {
2218 [ + - ]: 13 : if (minimum_order > huge_page_order(h))
2219 : 13 : minimum_order = huge_page_order(h);
2220 : :
2221 : : /* oversize hugepages were init'ed in early boot */
2222 [ + - ]: 13 : if (!hstate_is_gigantic(h))
2223 : 13 : hugetlb_hstate_alloc_pages(h);
2224 : : }
2225 : 13 : VM_BUG_ON(minimum_order == UINT_MAX);
2226 : 13 : }
2227 : :
2228 : 13 : static void __init report_hugepages(void)
2229 : : {
2230 : 13 : struct hstate *h;
2231 : :
2232 [ + + ]: 26 : for_each_hstate(h) {
2233 : 13 : char buf[32];
2234 : :
2235 : 13 : string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32);
2236 : 13 : pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2237 : : buf, h->free_huge_pages);
2238 : : }
2239 : 13 : }
2240 : :
2241 : : #ifdef CONFIG_HIGHMEM
2242 : : static void try_to_free_low(struct hstate *h, unsigned long count,
2243 : : nodemask_t *nodes_allowed)
2244 : : {
2245 : : int i;
2246 : :
2247 : : if (hstate_is_gigantic(h))
2248 : : return;
2249 : :
2250 : : for_each_node_mask(i, *nodes_allowed) {
2251 : : struct page *page, *next;
2252 : : struct list_head *freel = &h->hugepage_freelists[i];
2253 : : list_for_each_entry_safe(page, next, freel, lru) {
2254 : : if (count >= h->nr_huge_pages)
2255 : : return;
2256 : : if (PageHighMem(page))
2257 : : continue;
2258 : : list_del(&page->lru);
2259 : : update_and_free_page(h, page);
2260 : : h->free_huge_pages--;
2261 : : h->free_huge_pages_node[page_to_nid(page)]--;
2262 : : }
2263 : : }
2264 : : }
2265 : : #else
2266 : 0 : static inline void try_to_free_low(struct hstate *h, unsigned long count,
2267 : : nodemask_t *nodes_allowed)
2268 : : {
2269 : : }
2270 : : #endif
2271 : :
2272 : : /*
2273 : : * Increment or decrement surplus_huge_pages. Keep node-specific counters
2274 : : * balanced by operating on them in a round-robin fashion.
2275 : : * Returns 1 if an adjustment was made.
2276 : : */
2277 : 0 : static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
2278 : : int delta)
2279 : : {
2280 : 0 : int nr_nodes, node;
2281 : :
2282 : 0 : VM_BUG_ON(delta != -1 && delta != 1);
2283 : :
2284 [ # # ]: 0 : if (delta < 0) {
2285 [ # # ]: 0 : for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
2286 [ # # ]: 0 : if (h->surplus_huge_pages_node[node])
2287 : 0 : goto found;
2288 : : }
2289 : : } else {
2290 [ # # ]: 0 : for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) {
2291 : 0 : if (h->surplus_huge_pages_node[node] <
2292 [ # # ]: 0 : h->nr_huge_pages_node[node])
2293 : 0 : goto found;
2294 : : }
2295 : : }
2296 : : return 0;
2297 : :
2298 : 0 : found:
2299 : 0 : h->surplus_huge_pages += delta;
2300 : 0 : h->surplus_huge_pages_node[node] += delta;
2301 : 0 : return 1;
2302 : : }
2303 : :
2304 : : #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2305 : 0 : static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
2306 : : nodemask_t *nodes_allowed)
2307 : : {
2308 : 0 : unsigned long min_count, ret;
2309 : 0 : NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
2310 : :
2311 : : /*
2312 : : * Bit mask controlling how hard we retry per-node allocations.
2313 : : * If we can not allocate the bit mask, do not attempt to allocate
2314 : : * the requested huge pages.
2315 : : */
2316 : 0 : if (node_alloc_noretry)
2317 : 0 : nodes_clear(*node_alloc_noretry);
2318 : : else
2319 : : return -ENOMEM;
2320 : :
2321 : 0 : spin_lock(&hugetlb_lock);
2322 : :
2323 : : /*
2324 : : * Check for a node specific request.
2325 : : * Changing node specific huge page count may require a corresponding
2326 : : * change to the global count. In any case, the passed node mask
2327 : : * (nodes_allowed) will restrict alloc/free to the specified node.
2328 : : */
2329 [ # # ]: 0 : if (nid != NUMA_NO_NODE) {
2330 : 0 : unsigned long old_count = count;
2331 : :
2332 : 0 : count += h->nr_huge_pages - h->nr_huge_pages_node[nid];
2333 : : /*
2334 : : * User may have specified a large count value which caused the
2335 : : * above calculation to overflow. In this case, they wanted
2336 : : * to allocate as many huge pages as possible. Set count to
2337 : : * largest possible value to align with their intention.
2338 : : */
2339 [ # # ]: 0 : if (count < old_count)
2340 : 0 : count = ULONG_MAX;
2341 : : }
2342 : :
2343 : : /*
2344 : : * Gigantic pages runtime allocation depend on the capability for large
2345 : : * page range allocation.
2346 : : * If the system does not provide this feature, return an error when
2347 : : * the user tries to allocate gigantic pages but let the user free the
2348 : : * boottime allocated gigantic pages.
2349 : : */
2350 [ # # ]: 0 : if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
2351 [ # # ]: 0 : if (count > persistent_huge_pages(h)) {
2352 : 0 : spin_unlock(&hugetlb_lock);
2353 : 0 : NODEMASK_FREE(node_alloc_noretry);
2354 : 0 : return -EINVAL;
2355 : : }
2356 : : /* Fall through to decrease pool */
2357 : : }
2358 : :
2359 : : /*
2360 : : * Increase the pool size
2361 : : * First take pages out of surplus state. Then make up the
2362 : : * remaining difference by allocating fresh huge pages.
2363 : : *
2364 : : * We might race with alloc_surplus_huge_page() here and be unable
2365 : : * to convert a surplus huge page to a normal huge page. That is
2366 : : * not critical, though, it just means the overall size of the
2367 : : * pool might be one hugepage larger than it needs to be, but
2368 : : * within all the constraints specified by the sysctls.
2369 : : */
2370 [ # # # # ]: 0 : while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
2371 [ # # ]: 0 : if (!adjust_pool_surplus(h, nodes_allowed, -1))
2372 : : break;
2373 : : }
2374 : :
2375 [ # # ]: 0 : while (count > persistent_huge_pages(h)) {
2376 : : /*
2377 : : * If this allocation races such that we no longer need the
2378 : : * page, free_huge_page will handle it by freeing the page
2379 : : * and reducing the surplus.
2380 : : */
2381 : 0 : spin_unlock(&hugetlb_lock);
2382 : :
2383 : : /* yield cpu to avoid soft lockup */
2384 : 0 : cond_resched();
2385 : :
2386 : 0 : ret = alloc_pool_huge_page(h, nodes_allowed,
2387 : : node_alloc_noretry);
2388 : 0 : spin_lock(&hugetlb_lock);
2389 [ # # ]: 0 : if (!ret)
2390 : 0 : goto out;
2391 : :
2392 : : /* Bail for signals. Probably ctrl-c from user */
2393 [ # # ]: 0 : if (signal_pending(current))
2394 : 0 : goto out;
2395 : : }
2396 : :
2397 : : /*
2398 : : * Decrease the pool size
2399 : : * First return free pages to the buddy allocator (being careful
2400 : : * to keep enough around to satisfy reservations). Then place
2401 : : * pages into surplus state as needed so the pool will shrink
2402 : : * to the desired size as pages become free.
2403 : : *
2404 : : * By placing pages into the surplus state independent of the
2405 : : * overcommit value, we are allowing the surplus pool size to
2406 : : * exceed overcommit. There are few sane options here. Since
2407 : : * alloc_surplus_huge_page() is checking the global counter,
2408 : : * though, we'll note that we're not allowed to exceed surplus
2409 : : * and won't grow the pool anywhere else. Not until one of the
2410 : : * sysctls are changed, or the surplus pages go out of use.
2411 : : */
2412 : 0 : min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
2413 : 0 : min_count = max(count, min_count);
2414 : 0 : try_to_free_low(h, min_count, nodes_allowed);
2415 [ # # ]: 0 : while (min_count < persistent_huge_pages(h)) {
2416 [ # # ]: 0 : if (!free_pool_huge_page(h, nodes_allowed, 0))
2417 : : break;
2418 : 0 : cond_resched_lock(&hugetlb_lock);
2419 : : }
2420 [ # # ]: 0 : while (count < persistent_huge_pages(h)) {
2421 [ # # ]: 0 : if (!adjust_pool_surplus(h, nodes_allowed, 1))
2422 : : break;
2423 : : }
2424 : 0 : out:
2425 : 0 : h->max_huge_pages = persistent_huge_pages(h);
2426 : 0 : spin_unlock(&hugetlb_lock);
2427 : :
2428 : 0 : NODEMASK_FREE(node_alloc_noretry);
2429 : :
2430 : 0 : return 0;
2431 : : }
2432 : :
2433 : : #define HSTATE_ATTR_RO(_name) \
2434 : : static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2435 : :
2436 : : #define HSTATE_ATTR(_name) \
2437 : : static struct kobj_attribute _name##_attr = \
2438 : : __ATTR(_name, 0644, _name##_show, _name##_store)
2439 : :
2440 : : static struct kobject *hugepages_kobj;
2441 : : static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2442 : :
2443 : : static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp);
2444 : :
2445 : 0 : static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp)
2446 : : {
2447 : 0 : int i;
2448 : :
2449 [ # # # # : 0 : for (i = 0; i < HUGE_MAX_HSTATE; i++)
# # # # #
# # # ]
2450 [ # # # # : 0 : if (hstate_kobjs[i] == kobj) {
# # # # #
# # # ]
2451 : 0 : if (nidp)
2452 : 0 : *nidp = NUMA_NO_NODE;
2453 : 0 : return &hstates[i];
2454 : : }
2455 : :
2456 : 0 : return kobj_to_node_hstate(kobj, nidp);
2457 : : }
2458 : :
2459 : : static ssize_t nr_hugepages_show_common(struct kobject *kobj,
2460 : : struct kobj_attribute *attr, char *buf)
2461 : : {
2462 : : struct hstate *h;
2463 : : unsigned long nr_huge_pages;
2464 : : int nid;
2465 : :
2466 : : h = kobj_to_hstate(kobj, &nid);
2467 : : if (nid == NUMA_NO_NODE)
2468 : : nr_huge_pages = h->nr_huge_pages;
2469 : : else
2470 : : nr_huge_pages = h->nr_huge_pages_node[nid];
2471 : :
2472 : : return sprintf(buf, "%lu\n", nr_huge_pages);
2473 : : }
2474 : :
2475 : 0 : static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
2476 : : struct hstate *h, int nid,
2477 : : unsigned long count, size_t len)
2478 : : {
2479 : 0 : int err;
2480 : 0 : nodemask_t nodes_allowed, *n_mask;
2481 : :
2482 [ # # ]: 0 : if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported())
2483 : : return -EINVAL;
2484 : :
2485 [ # # ]: 0 : if (nid == NUMA_NO_NODE) {
2486 : : /*
2487 : : * global hstate attribute
2488 : : */
2489 [ # # # # ]: 0 : if (!(obey_mempolicy &&
2490 : 0 : init_nodemask_of_mempolicy(&nodes_allowed)))
2491 : : n_mask = &node_states[N_MEMORY];
2492 : : else
2493 : : n_mask = &nodes_allowed;
2494 : : } else {
2495 : : /*
2496 : : * Node specific request. count adjustment happens in
2497 : : * set_max_huge_pages() after acquiring hugetlb_lock.
2498 : : */
2499 : 0 : init_nodemask_of_node(&nodes_allowed, nid);
2500 : 0 : n_mask = &nodes_allowed;
2501 : : }
2502 : :
2503 : 0 : err = set_max_huge_pages(h, count, nid, n_mask);
2504 : :
2505 [ # # ]: 0 : return err ? err : len;
2506 : : }
2507 : :
2508 : 0 : static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
2509 : : struct kobject *kobj, const char *buf,
2510 : : size_t len)
2511 : : {
2512 : 0 : struct hstate *h;
2513 : 0 : unsigned long count;
2514 : 0 : int nid;
2515 : 0 : int err;
2516 : :
2517 : 0 : err = kstrtoul(buf, 10, &count);
2518 [ # # ]: 0 : if (err)
2519 : 0 : return err;
2520 : :
2521 : : h = kobj_to_hstate(kobj, &nid);
2522 : 0 : return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len);
2523 : : }
2524 : :
2525 : 0 : static ssize_t nr_hugepages_show(struct kobject *kobj,
2526 : : struct kobj_attribute *attr, char *buf)
2527 : : {
2528 : 0 : return nr_hugepages_show_common(kobj, attr, buf);
2529 : : }
2530 : :
2531 : 0 : static ssize_t nr_hugepages_store(struct kobject *kobj,
2532 : : struct kobj_attribute *attr, const char *buf, size_t len)
2533 : : {
2534 : 0 : return nr_hugepages_store_common(false, kobj, buf, len);
2535 : : }
2536 : : HSTATE_ATTR(nr_hugepages);
2537 : :
2538 : : #ifdef CONFIG_NUMA
2539 : :
2540 : : /*
2541 : : * hstate attribute for optionally mempolicy-based constraint on persistent
2542 : : * huge page alloc/free.
2543 : : */
2544 : 0 : static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj,
2545 : : struct kobj_attribute *attr, char *buf)
2546 : : {
2547 : 0 : return nr_hugepages_show_common(kobj, attr, buf);
2548 : : }
2549 : :
2550 : 0 : static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj,
2551 : : struct kobj_attribute *attr, const char *buf, size_t len)
2552 : : {
2553 : 0 : return nr_hugepages_store_common(true, kobj, buf, len);
2554 : : }
2555 : : HSTATE_ATTR(nr_hugepages_mempolicy);
2556 : : #endif
2557 : :
2558 : :
2559 : 0 : static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
2560 : : struct kobj_attribute *attr, char *buf)
2561 : : {
2562 : 0 : struct hstate *h = kobj_to_hstate(kobj, NULL);
2563 : 0 : return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
2564 : : }
2565 : :
2566 : 0 : static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
2567 : : struct kobj_attribute *attr, const char *buf, size_t count)
2568 : : {
2569 : 0 : int err;
2570 : 0 : unsigned long input;
2571 : 0 : struct hstate *h = kobj_to_hstate(kobj, NULL);
2572 : :
2573 [ # # ]: 0 : if (hstate_is_gigantic(h))
2574 : : return -EINVAL;
2575 : :
2576 : 0 : err = kstrtoul(buf, 10, &input);
2577 [ # # ]: 0 : if (err)
2578 : 0 : return err;
2579 : :
2580 : 0 : spin_lock(&hugetlb_lock);
2581 : 0 : h->nr_overcommit_huge_pages = input;
2582 : 0 : spin_unlock(&hugetlb_lock);
2583 : :
2584 : 0 : return count;
2585 : : }
2586 : : HSTATE_ATTR(nr_overcommit_hugepages);
2587 : :
2588 : 0 : static ssize_t free_hugepages_show(struct kobject *kobj,
2589 : : struct kobj_attribute *attr, char *buf)
2590 : : {
2591 : 0 : struct hstate *h;
2592 : 0 : unsigned long free_huge_pages;
2593 : 0 : int nid;
2594 : :
2595 : 0 : h = kobj_to_hstate(kobj, &nid);
2596 [ # # ]: 0 : if (nid == NUMA_NO_NODE)
2597 : 0 : free_huge_pages = h->free_huge_pages;
2598 : : else
2599 : 0 : free_huge_pages = h->free_huge_pages_node[nid];
2600 : :
2601 : 0 : return sprintf(buf, "%lu\n", free_huge_pages);
2602 : : }
2603 : : HSTATE_ATTR_RO(free_hugepages);
2604 : :
2605 : 0 : static ssize_t resv_hugepages_show(struct kobject *kobj,
2606 : : struct kobj_attribute *attr, char *buf)
2607 : : {
2608 : 0 : struct hstate *h = kobj_to_hstate(kobj, NULL);
2609 : 0 : return sprintf(buf, "%lu\n", h->resv_huge_pages);
2610 : : }
2611 : : HSTATE_ATTR_RO(resv_hugepages);
2612 : :
2613 : 0 : static ssize_t surplus_hugepages_show(struct kobject *kobj,
2614 : : struct kobj_attribute *attr, char *buf)
2615 : : {
2616 : 0 : struct hstate *h;
2617 : 0 : unsigned long surplus_huge_pages;
2618 : 0 : int nid;
2619 : :
2620 : 0 : h = kobj_to_hstate(kobj, &nid);
2621 [ # # ]: 0 : if (nid == NUMA_NO_NODE)
2622 : 0 : surplus_huge_pages = h->surplus_huge_pages;
2623 : : else
2624 : 0 : surplus_huge_pages = h->surplus_huge_pages_node[nid];
2625 : :
2626 : 0 : return sprintf(buf, "%lu\n", surplus_huge_pages);
2627 : : }
2628 : : HSTATE_ATTR_RO(surplus_hugepages);
2629 : :
2630 : : static struct attribute *hstate_attrs[] = {
2631 : : &nr_hugepages_attr.attr,
2632 : : &nr_overcommit_hugepages_attr.attr,
2633 : : &free_hugepages_attr.attr,
2634 : : &resv_hugepages_attr.attr,
2635 : : &surplus_hugepages_attr.attr,
2636 : : #ifdef CONFIG_NUMA
2637 : : &nr_hugepages_mempolicy_attr.attr,
2638 : : #endif
2639 : : NULL,
2640 : : };
2641 : :
2642 : : static const struct attribute_group hstate_attr_group = {
2643 : : .attrs = hstate_attrs,
2644 : : };
2645 : :
2646 : 26 : static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent,
2647 : : struct kobject **hstate_kobjs,
2648 : : const struct attribute_group *hstate_attr_group)
2649 : : {
2650 : 26 : int retval;
2651 : 26 : int hi = hstate_index(h);
2652 : :
2653 : 26 : hstate_kobjs[hi] = kobject_create_and_add(h->name, parent);
2654 [ + - ]: 26 : if (!hstate_kobjs[hi])
2655 : : return -ENOMEM;
2656 : :
2657 : 26 : retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group);
2658 [ - + ]: 26 : if (retval)
2659 : 0 : kobject_put(hstate_kobjs[hi]);
2660 : :
2661 : : return retval;
2662 : : }
2663 : :
2664 : 13 : static void __init hugetlb_sysfs_init(void)
2665 : : {
2666 : 13 : struct hstate *h;
2667 : 13 : int err;
2668 : :
2669 : 13 : hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
2670 [ + - ]: 13 : if (!hugepages_kobj)
2671 : : return;
2672 : :
2673 [ + + ]: 26 : for_each_hstate(h) {
2674 : 13 : err = hugetlb_sysfs_add_hstate(h, hugepages_kobj,
2675 : : hstate_kobjs, &hstate_attr_group);
2676 [ - + ]: 13 : if (err)
2677 : 0 : pr_err("Hugetlb: Unable to add hstate %s", h->name);
2678 : : }
2679 : : }
2680 : :
2681 : : #ifdef CONFIG_NUMA
2682 : :
2683 : : /*
2684 : : * node_hstate/s - associate per node hstate attributes, via their kobjects,
2685 : : * with node devices in node_devices[] using a parallel array. The array
2686 : : * index of a node device or _hstate == node id.
2687 : : * This is here to avoid any static dependency of the node device driver, in
2688 : : * the base kernel, on the hugetlb module.
2689 : : */
2690 : : struct node_hstate {
2691 : : struct kobject *hugepages_kobj;
2692 : : struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
2693 : : };
2694 : : static struct node_hstate node_hstates[MAX_NUMNODES];
2695 : :
2696 : : /*
2697 : : * A subset of global hstate attributes for node devices
2698 : : */
2699 : : static struct attribute *per_node_hstate_attrs[] = {
2700 : : &nr_hugepages_attr.attr,
2701 : : &free_hugepages_attr.attr,
2702 : : &surplus_hugepages_attr.attr,
2703 : : NULL,
2704 : : };
2705 : :
2706 : : static const struct attribute_group per_node_hstate_attr_group = {
2707 : : .attrs = per_node_hstate_attrs,
2708 : : };
2709 : :
2710 : : /*
2711 : : * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2712 : : * Returns node id via non-NULL nidp.
2713 : : */
2714 : 0 : static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2715 : : {
2716 : 0 : int nid;
2717 : :
2718 [ # # ]: 0 : for (nid = 0; nid < nr_node_ids; nid++) {
2719 : : struct node_hstate *nhs = &node_hstates[nid];
2720 : : int i;
2721 [ # # ]: 0 : for (i = 0; i < HUGE_MAX_HSTATE; i++)
2722 [ # # ]: 0 : if (nhs->hstate_kobjs[i] == kobj) {
2723 [ # # ]: 0 : if (nidp)
2724 : 0 : *nidp = nid;
2725 : 0 : return &hstates[i];
2726 : : }
2727 : : }
2728 : :
2729 : 0 : BUG();
2730 : : return NULL;
2731 : : }
2732 : :
2733 : : /*
2734 : : * Unregister hstate attributes from a single node device.
2735 : : * No-op if no hstate attributes attached.
2736 : : */
2737 : 0 : static void hugetlb_unregister_node(struct node *node)
2738 : : {
2739 : 0 : struct hstate *h;
2740 : 0 : struct node_hstate *nhs = &node_hstates[node->dev.id];
2741 : :
2742 [ # # ]: 0 : if (!nhs->hugepages_kobj)
2743 : : return; /* no hstate attributes */
2744 : :
2745 [ # # ]: 0 : for_each_hstate(h) {
2746 [ # # ]: 0 : int idx = hstate_index(h);
2747 [ # # ]: 0 : if (nhs->hstate_kobjs[idx]) {
2748 : 0 : kobject_put(nhs->hstate_kobjs[idx]);
2749 : 0 : nhs->hstate_kobjs[idx] = NULL;
2750 : : }
2751 : : }
2752 : :
2753 : 0 : kobject_put(nhs->hugepages_kobj);
2754 : 0 : nhs->hugepages_kobj = NULL;
2755 : : }
2756 : :
2757 : :
2758 : : /*
2759 : : * Register hstate attributes for a single node device.
2760 : : * No-op if attributes already registered.
2761 : : */
2762 : 13 : static void hugetlb_register_node(struct node *node)
2763 : : {
2764 : 13 : struct hstate *h;
2765 : 13 : struct node_hstate *nhs = &node_hstates[node->dev.id];
2766 : 13 : int err;
2767 : :
2768 [ + - ]: 13 : if (nhs->hugepages_kobj)
2769 : : return; /* already allocated */
2770 : :
2771 : 13 : nhs->hugepages_kobj = kobject_create_and_add("hugepages",
2772 : : &node->dev.kobj);
2773 [ + - ]: 13 : if (!nhs->hugepages_kobj)
2774 : : return;
2775 : :
2776 [ + + ]: 26 : for_each_hstate(h) {
2777 : 13 : err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj,
2778 : 13 : nhs->hstate_kobjs,
2779 : : &per_node_hstate_attr_group);
2780 [ - + ]: 13 : if (err) {
2781 : 0 : pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2782 : : h->name, node->dev.id);
2783 : 0 : hugetlb_unregister_node(node);
2784 : 0 : break;
2785 : : }
2786 : : }
2787 : : }
2788 : :
2789 : : /*
2790 : : * hugetlb init time: register hstate attributes for all registered node
2791 : : * devices of nodes that have memory. All on-line nodes should have
2792 : : * registered their associated device by this time.
2793 : : */
2794 : 13 : static void __init hugetlb_register_all_nodes(void)
2795 : : {
2796 : 13 : int nid;
2797 : :
2798 [ + + ]: 52 : for_each_node_state(nid, N_MEMORY) {
2799 : 13 : struct node *node = node_devices[nid];
2800 [ + - ]: 13 : if (node->dev.id == nid)
2801 : 13 : hugetlb_register_node(node);
2802 : : }
2803 : :
2804 : : /*
2805 : : * Let the node device driver know we're here so it can
2806 : : * [un]register hstate attributes on node hotplug.
2807 : : */
2808 : 13 : register_hugetlbfs_with_node(hugetlb_register_node,
2809 : : hugetlb_unregister_node);
2810 : 13 : }
2811 : : #else /* !CONFIG_NUMA */
2812 : :
2813 : : static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp)
2814 : : {
2815 : : BUG();
2816 : : if (nidp)
2817 : : *nidp = -1;
2818 : : return NULL;
2819 : : }
2820 : :
2821 : : static void hugetlb_register_all_nodes(void) { }
2822 : :
2823 : : #endif
2824 : :
2825 : 13 : static int __init hugetlb_init(void)
2826 : : {
2827 : 13 : int i;
2828 : :
2829 [ + - ]: 13 : if (!hugepages_supported())
2830 : : return 0;
2831 : :
2832 [ + - ]: 26 : if (!size_to_hstate(default_hstate_size)) {
2833 [ - + ]: 13 : if (default_hstate_size != 0) {
2834 : 0 : pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n",
2835 : : default_hstate_size, HPAGE_SIZE);
2836 : : }
2837 : :
2838 : 13 : default_hstate_size = HPAGE_SIZE;
2839 [ + - ]: 26 : if (!size_to_hstate(default_hstate_size))
2840 : 13 : hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
2841 : : }
2842 : 13 : default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size));
2843 [ - + ]: 13 : if (default_hstate_max_huge_pages) {
2844 [ # # ]: 0 : if (!default_hstate.max_huge_pages)
2845 : 0 : default_hstate.max_huge_pages = default_hstate_max_huge_pages;
2846 : : }
2847 : :
2848 : 13 : hugetlb_init_hstates();
2849 : 13 : gather_bootmem_prealloc();
2850 : 13 : report_hugepages();
2851 : :
2852 : 13 : hugetlb_sysfs_init();
2853 : 13 : hugetlb_register_all_nodes();
2854 : 13 : hugetlb_cgroup_file_init();
2855 : :
2856 : : #ifdef CONFIG_SMP
2857 : 13 : num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus());
2858 : : #else
2859 : : num_fault_mutexes = 1;
2860 : : #endif
2861 : 26 : hugetlb_fault_mutex_table =
2862 : 13 : kmalloc_array(num_fault_mutexes, sizeof(struct mutex),
2863 : : GFP_KERNEL);
2864 [ - + ]: 13 : BUG_ON(!hugetlb_fault_mutex_table);
2865 : :
2866 [ + + ]: 117 : for (i = 0; i < num_fault_mutexes; i++)
2867 : 104 : mutex_init(&hugetlb_fault_mutex_table[i]);
2868 : : return 0;
2869 : : }
2870 : : subsys_initcall(hugetlb_init);
2871 : :
2872 : : /* Should be called on processing a hugepagesz=... option */
2873 : 0 : void __init hugetlb_bad_size(void)
2874 : : {
2875 : 0 : parsed_valid_hugepagesz = false;
2876 : 0 : }
2877 : :
2878 : 13 : void __init hugetlb_add_hstate(unsigned int order)
2879 : : {
2880 : 13 : struct hstate *h;
2881 : 13 : unsigned long i;
2882 : :
2883 [ - + ]: 26 : if (size_to_hstate(PAGE_SIZE << order)) {
2884 : 0 : pr_warn("hugepagesz= specified twice, ignoring\n");
2885 : 0 : return;
2886 : : }
2887 [ - + ]: 13 : BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE);
2888 [ - + ]: 13 : BUG_ON(order == 0);
2889 : 13 : h = &hstates[hugetlb_max_hstate++];
2890 : 13 : h->order = order;
2891 : 13 : h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
2892 : 13 : h->nr_huge_pages = 0;
2893 : 13 : h->free_huge_pages = 0;
2894 [ + + ]: 845 : for (i = 0; i < MAX_NUMNODES; ++i)
2895 : 832 : INIT_LIST_HEAD(&h->hugepage_freelists[i]);
2896 : 13 : INIT_LIST_HEAD(&h->hugepage_activelist);
2897 : 13 : h->next_nid_to_alloc = first_memory_node;
2898 : 13 : h->next_nid_to_free = first_memory_node;
2899 : 13 : snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
2900 : 13 : huge_page_size(h)/1024);
2901 : :
2902 : 13 : parsed_hstate = h;
2903 : : }
2904 : :
2905 : 0 : static int __init hugetlb_nrpages_setup(char *s)
2906 : : {
2907 : 0 : unsigned long *mhp;
2908 : 0 : static unsigned long *last_mhp;
2909 : :
2910 [ # # ]: 0 : if (!parsed_valid_hugepagesz) {
2911 : 0 : pr_warn("hugepages = %s preceded by "
2912 : : "an unsupported hugepagesz, ignoring\n", s);
2913 : 0 : parsed_valid_hugepagesz = true;
2914 : 0 : return 1;
2915 : : }
2916 : : /*
2917 : : * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2918 : : * so this hugepages= parameter goes to the "default hstate".
2919 : : */
2920 [ # # ]: 0 : else if (!hugetlb_max_hstate)
2921 : : mhp = &default_hstate_max_huge_pages;
2922 : : else
2923 : 0 : mhp = &parsed_hstate->max_huge_pages;
2924 : :
2925 [ # # ]: 0 : if (mhp == last_mhp) {
2926 : 0 : pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2927 : 0 : return 1;
2928 : : }
2929 : :
2930 [ # # ]: 0 : if (sscanf(s, "%lu", mhp) <= 0)
2931 : 0 : *mhp = 0;
2932 : :
2933 : : /*
2934 : : * Global state is always initialized later in hugetlb_init.
2935 : : * But we need to allocate >= MAX_ORDER hstates here early to still
2936 : : * use the bootmem allocator.
2937 : : */
2938 [ # # # # ]: 0 : if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER)
2939 : 0 : hugetlb_hstate_alloc_pages(parsed_hstate);
2940 : :
2941 : 0 : last_mhp = mhp;
2942 : :
2943 : 0 : return 1;
2944 : : }
2945 : : __setup("hugepages=", hugetlb_nrpages_setup);
2946 : :
2947 : 0 : static int __init hugetlb_default_setup(char *s)
2948 : : {
2949 : 0 : default_hstate_size = memparse(s, &s);
2950 : 0 : return 1;
2951 : : }
2952 : : __setup("default_hugepagesz=", hugetlb_default_setup);
2953 : :
2954 : 0 : static unsigned int cpuset_mems_nr(unsigned int *array)
2955 : : {
2956 : 0 : int node;
2957 : 0 : unsigned int nr = 0;
2958 : :
2959 [ # # ]: 0 : for_each_node_mask(node, cpuset_current_mems_allowed)
2960 : 0 : nr += array[node];
2961 : :
2962 : 0 : return nr;
2963 : : }
2964 : :
2965 : : #ifdef CONFIG_SYSCTL
2966 : 0 : static int hugetlb_sysctl_handler_common(bool obey_mempolicy,
2967 : : struct ctl_table *table, int write,
2968 : : void __user *buffer, size_t *length, loff_t *ppos)
2969 : : {
2970 : 0 : struct hstate *h = &default_hstate;
2971 : 0 : unsigned long tmp = h->max_huge_pages;
2972 : 0 : int ret;
2973 : :
2974 [ # # ]: 0 : if (!hugepages_supported())
2975 : : return -EOPNOTSUPP;
2976 : :
2977 : 0 : table->data = &tmp;
2978 : 0 : table->maxlen = sizeof(unsigned long);
2979 : 0 : ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
2980 [ # # ]: 0 : if (ret)
2981 : 0 : goto out;
2982 : :
2983 [ # # ]: 0 : if (write)
2984 : 0 : ret = __nr_hugepages_store_common(obey_mempolicy, h,
2985 : : NUMA_NO_NODE, tmp, *length);
2986 : 0 : out:
2987 : : return ret;
2988 : : }
2989 : :
2990 : 0 : int hugetlb_sysctl_handler(struct ctl_table *table, int write,
2991 : : void __user *buffer, size_t *length, loff_t *ppos)
2992 : : {
2993 : :
2994 : 0 : return hugetlb_sysctl_handler_common(false, table, write,
2995 : : buffer, length, ppos);
2996 : : }
2997 : :
2998 : : #ifdef CONFIG_NUMA
2999 : 0 : int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write,
3000 : : void __user *buffer, size_t *length, loff_t *ppos)
3001 : : {
3002 : 0 : return hugetlb_sysctl_handler_common(true, table, write,
3003 : : buffer, length, ppos);
3004 : : }
3005 : : #endif /* CONFIG_NUMA */
3006 : :
3007 : 0 : int hugetlb_overcommit_handler(struct ctl_table *table, int write,
3008 : : void __user *buffer,
3009 : : size_t *length, loff_t *ppos)
3010 : : {
3011 : 0 : struct hstate *h = &default_hstate;
3012 : 0 : unsigned long tmp;
3013 : 0 : int ret;
3014 : :
3015 [ # # ]: 0 : if (!hugepages_supported())
3016 : : return -EOPNOTSUPP;
3017 : :
3018 : 0 : tmp = h->nr_overcommit_huge_pages;
3019 : :
3020 [ # # # # ]: 0 : if (write && hstate_is_gigantic(h))
3021 : : return -EINVAL;
3022 : :
3023 : 0 : table->data = &tmp;
3024 : 0 : table->maxlen = sizeof(unsigned long);
3025 : 0 : ret = proc_doulongvec_minmax(table, write, buffer, length, ppos);
3026 [ # # ]: 0 : if (ret)
3027 : 0 : goto out;
3028 : :
3029 [ # # ]: 0 : if (write) {
3030 : 0 : spin_lock(&hugetlb_lock);
3031 : 0 : h->nr_overcommit_huge_pages = tmp;
3032 : 0 : spin_unlock(&hugetlb_lock);
3033 : : }
3034 : 0 : out:
3035 : : return ret;
3036 : : }
3037 : :
3038 : : #endif /* CONFIG_SYSCTL */
3039 : :
3040 : 13 : void hugetlb_report_meminfo(struct seq_file *m)
3041 : : {
3042 : 13 : struct hstate *h;
3043 : 13 : unsigned long total = 0;
3044 : :
3045 [ + - ]: 13 : if (!hugepages_supported())
3046 : : return;
3047 : :
3048 [ + + ]: 26 : for_each_hstate(h) {
3049 : 13 : unsigned long count = h->nr_huge_pages;
3050 : :
3051 [ + - ]: 13 : total += (PAGE_SIZE << huge_page_order(h)) * count;
3052 : :
3053 [ + - ]: 13 : if (h == &default_hstate)
3054 : 13 : seq_printf(m,
3055 : : "HugePages_Total: %5lu\n"
3056 : : "HugePages_Free: %5lu\n"
3057 : : "HugePages_Rsvd: %5lu\n"
3058 : : "HugePages_Surp: %5lu\n"
3059 : : "Hugepagesize: %8lu kB\n",
3060 : : count,
3061 : : h->free_huge_pages,
3062 : : h->resv_huge_pages,
3063 : : h->surplus_huge_pages,
3064 : : (PAGE_SIZE << huge_page_order(h)) / 1024);
3065 : : }
3066 : :
3067 : 13 : seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024);
3068 : : }
3069 : :
3070 : 0 : int hugetlb_report_node_meminfo(int nid, char *buf)
3071 : : {
3072 : 0 : struct hstate *h = &default_hstate;
3073 [ # # ]: 0 : if (!hugepages_supported())
3074 : : return 0;
3075 : 0 : return sprintf(buf,
3076 : : "Node %d HugePages_Total: %5u\n"
3077 : : "Node %d HugePages_Free: %5u\n"
3078 : : "Node %d HugePages_Surp: %5u\n",
3079 : : nid, h->nr_huge_pages_node[nid],
3080 : : nid, h->free_huge_pages_node[nid],
3081 : : nid, h->surplus_huge_pages_node[nid]);
3082 : : }
3083 : :
3084 : 0 : void hugetlb_show_meminfo(void)
3085 : : {
3086 : 0 : struct hstate *h;
3087 : 0 : int nid;
3088 : :
3089 [ # # ]: 0 : if (!hugepages_supported())
3090 : : return;
3091 : :
3092 [ # # ]: 0 : for_each_node_state(nid, N_MEMORY)
3093 [ # # ]: 0 : for_each_hstate(h)
3094 : 0 : pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3095 : : nid,
3096 : : h->nr_huge_pages_node[nid],
3097 : : h->free_huge_pages_node[nid],
3098 : : h->surplus_huge_pages_node[nid],
3099 : : 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
3100 : : }
3101 : :
3102 : 287 : void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm)
3103 : : {
3104 : 287 : seq_printf(m, "HugetlbPages:\t%8lu kB\n",
3105 : 287 : atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10));
3106 : 287 : }
3107 : :
3108 : : /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3109 : 13 : unsigned long hugetlb_total_pages(void)
3110 : : {
3111 : 13 : struct hstate *h;
3112 : 13 : unsigned long nr_total_pages = 0;
3113 : :
3114 [ + + ]: 26 : for_each_hstate(h)
3115 : 13 : nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h);
3116 : 13 : return nr_total_pages;
3117 : : }
3118 : :
3119 : 0 : static int hugetlb_acct_memory(struct hstate *h, long delta)
3120 : : {
3121 : 0 : int ret = -ENOMEM;
3122 : :
3123 : 0 : spin_lock(&hugetlb_lock);
3124 : : /*
3125 : : * When cpuset is configured, it breaks the strict hugetlb page
3126 : : * reservation as the accounting is done on a global variable. Such
3127 : : * reservation is completely rubbish in the presence of cpuset because
3128 : : * the reservation is not checked against page availability for the
3129 : : * current cpuset. Application can still potentially OOM'ed by kernel
3130 : : * with lack of free htlb page in cpuset that the task is in.
3131 : : * Attempt to enforce strict accounting with cpuset is almost
3132 : : * impossible (or too ugly) because cpuset is too fluid that
3133 : : * task or memory node can be dynamically moved between cpusets.
3134 : : *
3135 : : * The change of semantics for shared hugetlb mapping with cpuset is
3136 : : * undesirable. However, in order to preserve some of the semantics,
3137 : : * we fall back to check against current free page availability as
3138 : : * a best attempt and hopefully to minimize the impact of changing
3139 : : * semantics that cpuset has.
3140 : : */
3141 [ # # ]: 0 : if (delta > 0) {
3142 [ # # ]: 0 : if (gather_surplus_pages(h, delta) < 0)
3143 : 0 : goto out;
3144 : :
3145 [ # # ]: 0 : if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
3146 : 0 : return_unused_surplus_pages(h, delta);
3147 : 0 : goto out;
3148 : : }
3149 : : }
3150 : :
3151 : 0 : ret = 0;
3152 [ # # ]: 0 : if (delta < 0)
3153 : 0 : return_unused_surplus_pages(h, (unsigned long) -delta);
3154 : :
3155 : 0 : out:
3156 : 0 : spin_unlock(&hugetlb_lock);
3157 : 0 : return ret;
3158 : : }
3159 : :
3160 : 0 : static void hugetlb_vm_op_open(struct vm_area_struct *vma)
3161 : : {
3162 [ # # ]: 0 : struct resv_map *resv = vma_resv_map(vma);
3163 : :
3164 : : /*
3165 : : * This new VMA should share its siblings reservation map if present.
3166 : : * The VMA will only ever have a valid reservation map pointer where
3167 : : * it is being copied for another still existing VMA. As that VMA
3168 : : * has a reference to the reservation map it cannot disappear until
3169 : : * after this open call completes. It is therefore safe to take a
3170 : : * new reference here without additional locking.
3171 : : */
3172 [ # # # # ]: 0 : if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3173 : 0 : kref_get(&resv->refs);
3174 : 0 : }
3175 : :
3176 : 0 : static void hugetlb_vm_op_close(struct vm_area_struct *vma)
3177 : : {
3178 [ # # ]: 0 : struct hstate *h = hstate_vma(vma);
3179 [ # # ]: 0 : struct resv_map *resv = vma_resv_map(vma);
3180 [ # # ]: 0 : struct hugepage_subpool *spool = subpool_vma(vma);
3181 : 0 : unsigned long reserve, start, end;
3182 : 0 : long gbl_reserve;
3183 : :
3184 [ # # # # ]: 0 : if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER))
3185 : : return;
3186 : :
3187 : 0 : start = vma_hugecache_offset(h, vma, vma->vm_start);
3188 : 0 : end = vma_hugecache_offset(h, vma, vma->vm_end);
3189 : :
3190 : 0 : reserve = (end - start) - region_count(resv, start, end);
3191 : :
3192 : 0 : kref_put(&resv->refs, resv_map_release);
3193 : :
3194 [ # # ]: 0 : if (reserve) {
3195 : : /*
3196 : : * Decrement reserve counts. The global reserve count may be
3197 : : * adjusted if the subpool has a minimum size.
3198 : : */
3199 : 0 : gbl_reserve = hugepage_subpool_put_pages(spool, reserve);
3200 : 0 : hugetlb_acct_memory(h, -gbl_reserve);
3201 : : }
3202 : : }
3203 : :
3204 : 0 : static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr)
3205 : : {
3206 [ # # ]: 0 : if (addr & ~(huge_page_mask(hstate_vma(vma))))
3207 : 0 : return -EINVAL;
3208 : : return 0;
3209 : : }
3210 : :
3211 : 0 : static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma)
3212 : : {
3213 : 0 : struct hstate *hstate = hstate_vma(vma);
3214 : :
3215 : 0 : return 1UL << huge_page_shift(hstate);
3216 : : }
3217 : :
3218 : : /*
3219 : : * We cannot handle pagefaults against hugetlb pages at all. They cause
3220 : : * handle_mm_fault() to try to instantiate regular-sized pages in the
3221 : : * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3222 : : * this far.
3223 : : */
3224 : 0 : static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf)
3225 : : {
3226 : 0 : BUG();
3227 : : return 0;
3228 : : }
3229 : :
3230 : : /*
3231 : : * When a new function is introduced to vm_operations_struct and added
3232 : : * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops.
3233 : : * This is because under System V memory model, mappings created via
3234 : : * shmget/shmat with "huge page" specified are backed by hugetlbfs files,
3235 : : * their original vm_ops are overwritten with shm_vm_ops.
3236 : : */
3237 : : const struct vm_operations_struct hugetlb_vm_ops = {
3238 : : .fault = hugetlb_vm_op_fault,
3239 : : .open = hugetlb_vm_op_open,
3240 : : .close = hugetlb_vm_op_close,
3241 : : .split = hugetlb_vm_op_split,
3242 : : .pagesize = hugetlb_vm_op_pagesize,
3243 : : };
3244 : :
3245 : 0 : static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
3246 : : int writable)
3247 : : {
3248 : 0 : pte_t entry;
3249 : :
3250 : 0 : if (writable) {
3251 : 0 : entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page,
3252 : : vma->vm_page_prot)));
3253 : : } else {
3254 : : entry = huge_pte_wrprotect(mk_huge_pte(page,
3255 : : vma->vm_page_prot));
3256 : : }
3257 : 0 : entry = pte_mkyoung(entry);
3258 : 0 : entry = pte_mkhuge(entry);
3259 : 0 : entry = arch_make_huge_pte(entry, vma, page, writable);
3260 : :
3261 : 0 : return entry;
3262 : : }
3263 : :
3264 : 0 : static void set_huge_ptep_writable(struct vm_area_struct *vma,
3265 : : unsigned long address, pte_t *ptep)
3266 : : {
3267 : 0 : pte_t entry;
3268 : :
3269 : 0 : entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep)));
3270 : 0 : if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1))
3271 : : update_mmu_cache(vma, address, ptep);
3272 : : }
3273 : :
3274 : 0 : bool is_hugetlb_entry_migration(pte_t pte)
3275 : : {
3276 : 0 : swp_entry_t swp;
3277 : :
3278 [ # # # # ]: 0 : if (huge_pte_none(pte) || pte_present(pte))
3279 : : return false;
3280 [ # # ]: 0 : swp = pte_to_swp_entry(pte);
3281 [ # # ]: 0 : if (non_swap_entry(swp) && is_migration_entry(swp))
3282 : : return true;
3283 : : else
3284 : 0 : return false;
3285 : : }
3286 : :
3287 : 0 : static int is_hugetlb_entry_hwpoisoned(pte_t pte)
3288 : : {
3289 : 0 : swp_entry_t swp;
3290 : :
3291 [ # # ]: 0 : if (huge_pte_none(pte) || pte_present(pte))
3292 : 0 : return 0;
3293 : : swp = pte_to_swp_entry(pte);
3294 : : if (non_swap_entry(swp) && is_hwpoison_entry(swp))
3295 : : return 1;
3296 : : else
3297 : : return 0;
3298 : : }
3299 : :
3300 : 0 : int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
3301 : : struct vm_area_struct *vma)
3302 : : {
3303 : 0 : pte_t *src_pte, *dst_pte, entry, dst_entry;
3304 : 0 : struct page *ptepage;
3305 : 0 : unsigned long addr;
3306 : 0 : int cow;
3307 [ # # ]: 0 : struct hstate *h = hstate_vma(vma);
3308 [ # # ]: 0 : unsigned long sz = huge_page_size(h);
3309 : 0 : struct mmu_notifier_range range;
3310 : 0 : int ret = 0;
3311 : :
3312 : 0 : cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
3313 : :
3314 [ # # ]: 0 : if (cow) {
3315 : 0 : mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src,
3316 : : vma->vm_start,
3317 : : vma->vm_end);
3318 : 0 : mmu_notifier_invalidate_range_start(&range);
3319 : : }
3320 : :
3321 [ # # ]: 0 : for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
3322 : 0 : spinlock_t *src_ptl, *dst_ptl;
3323 : 0 : src_pte = huge_pte_offset(src, addr, sz);
3324 [ # # ]: 0 : if (!src_pte)
3325 : 0 : continue;
3326 : 0 : dst_pte = huge_pte_alloc(dst, addr, sz);
3327 [ # # ]: 0 : if (!dst_pte) {
3328 : : ret = -ENOMEM;
3329 : : break;
3330 : : }
3331 : :
3332 : : /*
3333 : : * If the pagetables are shared don't copy or take references.
3334 : : * dst_pte == src_pte is the common case of src/dest sharing.
3335 : : *
3336 : : * However, src could have 'unshared' and dst shares with
3337 : : * another vma. If dst_pte !none, this implies sharing.
3338 : : * Check here before taking page table lock, and once again
3339 : : * after taking the lock below.
3340 : : */
3341 [ # # ]: 0 : dst_entry = huge_ptep_get(dst_pte);
3342 [ # # # # ]: 0 : if ((dst_pte == src_pte) || !huge_pte_none(dst_entry))
3343 : 0 : continue;
3344 : :
3345 : 0 : dst_ptl = huge_pte_lock(h, dst, dst_pte);
3346 [ # # ]: 0 : src_ptl = huge_pte_lockptr(h, src, src_pte);
3347 : 0 : spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
3348 [ # # ]: 0 : entry = huge_ptep_get(src_pte);
3349 : 0 : dst_entry = huge_ptep_get(dst_pte);
3350 [ # # # # ]: 0 : if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) {
3351 : : /*
3352 : : * Skip if src entry none. Also, skip in the
3353 : : * unlikely case dst entry !none as this implies
3354 : : * sharing with another vma.
3355 : : */
3356 : : ;
3357 [ # # ]: 0 : } else if (unlikely(is_hugetlb_entry_migration(entry) ||
3358 : 0 : is_hugetlb_entry_hwpoisoned(entry))) {
3359 [ # # ]: 0 : swp_entry_t swp_entry = pte_to_swp_entry(entry);
3360 : :
3361 [ # # # # ]: 0 : if (is_write_migration_entry(swp_entry) && cow) {
3362 : : /*
3363 : : * COW mappings require pages in both
3364 : : * parent and child to be set to read.
3365 : : */
3366 : 0 : make_migration_entry_read(&swp_entry);
3367 : 0 : entry = swp_entry_to_pte(swp_entry);
3368 : 0 : set_huge_swap_pte_at(src, addr, src_pte,
3369 : : entry, sz);
3370 : : }
3371 : 0 : set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz);
3372 : : } else {
3373 [ # # ]: 0 : if (cow) {
3374 : : /*
3375 : : * No need to notify as we are downgrading page
3376 : : * table protection not changing it to point
3377 : : * to a new page.
3378 : : *
3379 : : * See Documentation/vm/mmu_notifier.rst
3380 : : */
3381 : 0 : huge_ptep_set_wrprotect(src, addr, src_pte);
3382 : : }
3383 [ # # ]: 0 : entry = huge_ptep_get(src_pte);
3384 [ # # ]: 0 : ptepage = pte_page(entry);
3385 [ # # ]: 0 : get_page(ptepage);
3386 : 0 : page_dup_rmap(ptepage, true);
3387 : 0 : set_huge_pte_at(dst, addr, dst_pte, entry);
3388 : 0 : hugetlb_count_add(pages_per_huge_page(h), dst);
3389 : : }
3390 : 0 : spin_unlock(src_ptl);
3391 : 0 : spin_unlock(dst_ptl);
3392 : : }
3393 : :
3394 [ # # ]: 0 : if (cow)
3395 : 0 : mmu_notifier_invalidate_range_end(&range);
3396 : :
3397 : 0 : return ret;
3398 : : }
3399 : :
3400 : 0 : void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma,
3401 : : unsigned long start, unsigned long end,
3402 : : struct page *ref_page)
3403 : : {
3404 : 0 : struct mm_struct *mm = vma->vm_mm;
3405 : 0 : unsigned long address;
3406 : 0 : pte_t *ptep;
3407 : 0 : pte_t pte;
3408 : 0 : spinlock_t *ptl;
3409 : 0 : struct page *page;
3410 [ # # ]: 0 : struct hstate *h = hstate_vma(vma);
3411 [ # # ]: 0 : unsigned long sz = huge_page_size(h);
3412 : 0 : struct mmu_notifier_range range;
3413 : :
3414 [ # # ]: 0 : WARN_ON(!is_vm_hugetlb_page(vma));
3415 [ # # ]: 0 : BUG_ON(start & ~huge_page_mask(h));
3416 [ # # ]: 0 : BUG_ON(end & ~huge_page_mask(h));
3417 : :
3418 : : /*
3419 : : * This is a hugetlb vma, all the pte entries should point
3420 : : * to huge page.
3421 : : */
3422 : 0 : tlb_change_page_size(tlb, sz);
3423 : 0 : tlb_start_vma(tlb, vma);
3424 : :
3425 : : /*
3426 : : * If sharing possible, alert mmu notifiers of worst case.
3427 : : */
3428 : 0 : mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start,
3429 : : end);
3430 : 0 : adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
3431 : 0 : mmu_notifier_invalidate_range_start(&range);
3432 : 0 : address = start;
3433 [ # # ]: 0 : for (; address < end; address += sz) {
3434 : 0 : ptep = huge_pte_offset(mm, address, sz);
3435 [ # # ]: 0 : if (!ptep)
3436 : 0 : continue;
3437 : :
3438 : 0 : ptl = huge_pte_lock(h, mm, ptep);
3439 [ # # ]: 0 : if (huge_pmd_unshare(mm, &address, ptep)) {
3440 : 0 : spin_unlock(ptl);
3441 : : /*
3442 : : * We just unmapped a page of PMDs by clearing a PUD.
3443 : : * The caller's TLB flush range should cover this area.
3444 : : */
3445 : 0 : continue;
3446 : : }
3447 : :
3448 [ # # ]: 0 : pte = huge_ptep_get(ptep);
3449 [ # # ]: 0 : if (huge_pte_none(pte)) {
3450 : 0 : spin_unlock(ptl);
3451 : 0 : continue;
3452 : : }
3453 : :
3454 : : /*
3455 : : * Migrating hugepage or HWPoisoned hugepage is already
3456 : : * unmapped and its refcount is dropped, so just clear pte here.
3457 : : */
3458 [ # # ]: 0 : if (unlikely(!pte_present(pte))) {
3459 : 0 : huge_pte_clear(mm, address, ptep, sz);
3460 : 0 : spin_unlock(ptl);
3461 : 0 : continue;
3462 : : }
3463 : :
3464 [ # # ]: 0 : page = pte_page(pte);
3465 : : /*
3466 : : * If a reference page is supplied, it is because a specific
3467 : : * page is being unmapped, not a range. Ensure the page we
3468 : : * are about to unmap is the actual page of interest.
3469 : : */
3470 [ # # ]: 0 : if (ref_page) {
3471 [ # # ]: 0 : if (page != ref_page) {
3472 : 0 : spin_unlock(ptl);
3473 : 0 : continue;
3474 : : }
3475 : : /*
3476 : : * Mark the VMA as having unmapped its page so that
3477 : : * future faults in this VMA will fail rather than
3478 : : * looking like data was lost
3479 : : */
3480 : 0 : set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
3481 : : }
3482 : :
3483 : 0 : pte = huge_ptep_get_and_clear(mm, address, ptep);
3484 [ # # # # ]: 0 : tlb_remove_huge_tlb_entry(h, tlb, ptep, address);
3485 [ # # ]: 0 : if (huge_pte_dirty(pte))
3486 : 0 : set_page_dirty(page);
3487 : :
3488 : 0 : hugetlb_count_sub(pages_per_huge_page(h), mm);
3489 : 0 : page_remove_rmap(page, true);
3490 : :
3491 : 0 : spin_unlock(ptl);
3492 : 0 : tlb_remove_page_size(tlb, page, huge_page_size(h));
3493 : : /*
3494 : : * Bail out after unmapping reference page if supplied
3495 : : */
3496 [ # # ]: 0 : if (ref_page)
3497 : : break;
3498 : : }
3499 : 0 : mmu_notifier_invalidate_range_end(&range);
3500 : 0 : tlb_end_vma(tlb, vma);
3501 : 0 : }
3502 : :
3503 : 0 : void __unmap_hugepage_range_final(struct mmu_gather *tlb,
3504 : : struct vm_area_struct *vma, unsigned long start,
3505 : : unsigned long end, struct page *ref_page)
3506 : : {
3507 : 0 : __unmap_hugepage_range(tlb, vma, start, end, ref_page);
3508 : :
3509 : : /*
3510 : : * Clear this flag so that x86's huge_pmd_share page_table_shareable
3511 : : * test will fail on a vma being torn down, and not grab a page table
3512 : : * on its way out. We're lucky that the flag has such an appropriate
3513 : : * name, and can in fact be safely cleared here. We could clear it
3514 : : * before the __unmap_hugepage_range above, but all that's necessary
3515 : : * is to clear it before releasing the i_mmap_rwsem. This works
3516 : : * because in the context this is called, the VMA is about to be
3517 : : * destroyed and the i_mmap_rwsem is held.
3518 : : */
3519 : 0 : vma->vm_flags &= ~VM_MAYSHARE;
3520 : 0 : }
3521 : :
3522 : 0 : void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
3523 : : unsigned long end, struct page *ref_page)
3524 : : {
3525 : 0 : struct mm_struct *mm;
3526 : 0 : struct mmu_gather tlb;
3527 : 0 : unsigned long tlb_start = start;
3528 : 0 : unsigned long tlb_end = end;
3529 : :
3530 : : /*
3531 : : * If shared PMDs were possibly used within this vma range, adjust
3532 : : * start/end for worst case tlb flushing.
3533 : : * Note that we can not be sure if PMDs are shared until we try to
3534 : : * unmap pages. However, we want to make sure TLB flushing covers
3535 : : * the largest possible range.
3536 : : */
3537 : 0 : adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end);
3538 : :
3539 : 0 : mm = vma->vm_mm;
3540 : :
3541 : 0 : tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end);
3542 : 0 : __unmap_hugepage_range(&tlb, vma, start, end, ref_page);
3543 : 0 : tlb_finish_mmu(&tlb, tlb_start, tlb_end);
3544 : 0 : }
3545 : :
3546 : : /*
3547 : : * This is called when the original mapper is failing to COW a MAP_PRIVATE
3548 : : * mappping it owns the reserve page for. The intention is to unmap the page
3549 : : * from other VMAs and let the children be SIGKILLed if they are faulting the
3550 : : * same region.
3551 : : */
3552 : : static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
3553 : : struct page *page, unsigned long address)
3554 : : {
3555 : : struct hstate *h = hstate_vma(vma);
3556 : : struct vm_area_struct *iter_vma;
3557 : : struct address_space *mapping;
3558 : : pgoff_t pgoff;
3559 : :
3560 : : /*
3561 : : * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3562 : : * from page cache lookup which is in HPAGE_SIZE units.
3563 : : */
3564 : : address = address & huge_page_mask(h);
3565 : : pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) +
3566 : : vma->vm_pgoff;
3567 : : mapping = vma->vm_file->f_mapping;
3568 : :
3569 : : /*
3570 : : * Take the mapping lock for the duration of the table walk. As
3571 : : * this mapping should be shared between all the VMAs,
3572 : : * __unmap_hugepage_range() is called as the lock is already held
3573 : : */
3574 : : i_mmap_lock_write(mapping);
3575 : : vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) {
3576 : : /* Do not unmap the current VMA */
3577 : : if (iter_vma == vma)
3578 : : continue;
3579 : :
3580 : : /*
3581 : : * Shared VMAs have their own reserves and do not affect
3582 : : * MAP_PRIVATE accounting but it is possible that a shared
3583 : : * VMA is using the same page so check and skip such VMAs.
3584 : : */
3585 : : if (iter_vma->vm_flags & VM_MAYSHARE)
3586 : : continue;
3587 : :
3588 : : /*
3589 : : * Unmap the page from other VMAs without their own reserves.
3590 : : * They get marked to be SIGKILLed if they fault in these
3591 : : * areas. This is because a future no-page fault on this VMA
3592 : : * could insert a zeroed page instead of the data existing
3593 : : * from the time of fork. This would look like data corruption
3594 : : */
3595 : : if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
3596 : : unmap_hugepage_range(iter_vma, address,
3597 : : address + huge_page_size(h), page);
3598 : : }
3599 : : i_mmap_unlock_write(mapping);
3600 : : }
3601 : :
3602 : : /*
3603 : : * Hugetlb_cow() should be called with page lock of the original hugepage held.
3604 : : * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3605 : : * cannot race with other handlers or page migration.
3606 : : * Keep the pte_same checks anyway to make transition from the mutex easier.
3607 : : */
3608 : 0 : static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
3609 : : unsigned long address, pte_t *ptep,
3610 : : struct page *pagecache_page, spinlock_t *ptl)
3611 : : {
3612 : 0 : pte_t pte;
3613 [ # # ]: 0 : struct hstate *h = hstate_vma(vma);
3614 : 0 : struct page *old_page, *new_page;
3615 : 0 : int outside_reserve = 0;
3616 : 0 : vm_fault_t ret = 0;
3617 [ # # ]: 0 : unsigned long haddr = address & huge_page_mask(h);
3618 : 0 : struct mmu_notifier_range range;
3619 : :
3620 [ # # ]: 0 : pte = huge_ptep_get(ptep);
3621 [ # # ]: 0 : old_page = pte_page(pte);
3622 : :
3623 : 0 : retry_avoidcopy:
3624 : : /* If no-one else is actually using this page, avoid the copy
3625 : : * and just make the page writable */
3626 [ # # # # ]: 0 : if (page_mapcount(old_page) == 1 && PageAnon(old_page)) {
3627 : 0 : page_move_anon_rmap(old_page, vma);
3628 : 0 : set_huge_ptep_writable(vma, haddr, ptep);
3629 : 0 : return 0;
3630 : : }
3631 : :
3632 : : /*
3633 : : * If the process that created a MAP_PRIVATE mapping is about to
3634 : : * perform a COW due to a shared page count, attempt to satisfy
3635 : : * the allocation without using the existing reserves. The pagecache
3636 : : * page is used to determine if the reserve at this address was
3637 : : * consumed or not. If reserves were used, a partial faulted mapping
3638 : : * at the time of fork() could consume its reserves on COW instead
3639 : : * of the full address range.
3640 : : */
3641 [ # # # # ]: 0 : if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
3642 : : old_page != pagecache_page)
3643 : 0 : outside_reserve = 1;
3644 : :
3645 [ # # ]: 0 : get_page(old_page);
3646 : :
3647 : : /*
3648 : : * Drop page table lock as buddy allocator may be called. It will
3649 : : * be acquired again before returning to the caller, as expected.
3650 : : */
3651 : 0 : spin_unlock(ptl);
3652 : 0 : new_page = alloc_huge_page(vma, haddr, outside_reserve);
3653 : :
3654 [ # # ]: 0 : if (IS_ERR(new_page)) {
3655 : : /*
3656 : : * If a process owning a MAP_PRIVATE mapping fails to COW,
3657 : : * it is due to references held by a child and an insufficient
3658 : : * huge page pool. To guarantee the original mappers
3659 : : * reliability, unmap the page from child processes. The child
3660 : : * may get SIGKILLed if it later faults.
3661 : : */
3662 [ # # ]: 0 : if (outside_reserve) {
3663 : 0 : put_page(old_page);
3664 [ # # ]: 0 : BUG_ON(huge_pte_none(pte));
3665 : 0 : unmap_ref_private(mm, vma, old_page, haddr);
3666 : 0 : BUG_ON(huge_pte_none(pte));
3667 : 0 : spin_lock(ptl);
3668 : 0 : ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3669 [ # # # # ]: 0 : if (likely(ptep &&
3670 : : pte_same(huge_ptep_get(ptep), pte)))
3671 : 0 : goto retry_avoidcopy;
3672 : : /*
3673 : : * race occurs while re-acquiring page table
3674 : : * lock, and our job is done.
3675 : : */
3676 : : return 0;
3677 : : }
3678 : :
3679 [ # # ]: 0 : ret = vmf_error(PTR_ERR(new_page));
3680 : 0 : goto out_release_old;
3681 : : }
3682 : :
3683 : : /*
3684 : : * When the original hugepage is shared one, it does not have
3685 : : * anon_vma prepared.
3686 : : */
3687 [ # # # # ]: 0 : if (unlikely(anon_vma_prepare(vma))) {
3688 : 0 : ret = VM_FAULT_OOM;
3689 : 0 : goto out_release_all;
3690 : : }
3691 : :
3692 : 0 : copy_user_huge_page(new_page, old_page, address, vma,
3693 : : pages_per_huge_page(h));
3694 : 0 : __SetPageUptodate(new_page);
3695 : :
3696 : 0 : mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr,
3697 : 0 : haddr + huge_page_size(h));
3698 : 0 : mmu_notifier_invalidate_range_start(&range);
3699 : :
3700 : : /*
3701 : : * Retake the page table lock to check for racing updates
3702 : : * before the page tables are altered
3703 : : */
3704 : 0 : spin_lock(ptl);
3705 : 0 : ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
3706 [ # # # # ]: 0 : if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) {
3707 : 0 : ClearPagePrivate(new_page);
3708 : :
3709 : : /* Break COW */
3710 : 0 : huge_ptep_clear_flush(vma, haddr, ptep);
3711 [ # # ]: 0 : mmu_notifier_invalidate_range(mm, range.start, range.end);
3712 : 0 : set_huge_pte_at(mm, haddr, ptep,
3713 : : make_huge_pte(vma, new_page, 1));
3714 : 0 : page_remove_rmap(old_page, true);
3715 : 0 : hugepage_add_new_anon_rmap(new_page, vma, haddr);
3716 : 0 : set_page_huge_active(new_page);
3717 : : /* Make the old page be freed below */
3718 : 0 : new_page = old_page;
3719 : : }
3720 : 0 : spin_unlock(ptl);
3721 : 0 : mmu_notifier_invalidate_range_end(&range);
3722 : 0 : out_release_all:
3723 : 0 : restore_reserve_on_error(h, vma, haddr, new_page);
3724 : 0 : put_page(new_page);
3725 : 0 : out_release_old:
3726 : 0 : put_page(old_page);
3727 : :
3728 : 0 : spin_lock(ptl); /* Caller expects lock to be held */
3729 : 0 : return ret;
3730 : : }
3731 : :
3732 : : /* Return the pagecache page at a given address within a VMA */
3733 : 0 : static struct page *hugetlbfs_pagecache_page(struct hstate *h,
3734 : : struct vm_area_struct *vma, unsigned long address)
3735 : : {
3736 : 0 : struct address_space *mapping;
3737 : 0 : pgoff_t idx;
3738 : :
3739 : 0 : mapping = vma->vm_file->f_mapping;
3740 : 0 : idx = vma_hugecache_offset(h, vma, address);
3741 : :
3742 : 0 : return find_lock_page(mapping, idx);
3743 : : }
3744 : :
3745 : : /*
3746 : : * Return whether there is a pagecache page to back given address within VMA.
3747 : : * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3748 : : */
3749 : 0 : static bool hugetlbfs_pagecache_present(struct hstate *h,
3750 : : struct vm_area_struct *vma, unsigned long address)
3751 : : {
3752 : 0 : struct address_space *mapping;
3753 : 0 : pgoff_t idx;
3754 : 0 : struct page *page;
3755 : :
3756 : 0 : mapping = vma->vm_file->f_mapping;
3757 : 0 : idx = vma_hugecache_offset(h, vma, address);
3758 : :
3759 : 0 : page = find_get_page(mapping, idx);
3760 [ # # ]: 0 : if (page)
3761 : 0 : put_page(page);
3762 : 0 : return page != NULL;
3763 : : }
3764 : :
3765 : 0 : int huge_add_to_page_cache(struct page *page, struct address_space *mapping,
3766 : : pgoff_t idx)
3767 : : {
3768 : 0 : struct inode *inode = mapping->host;
3769 : 0 : struct hstate *h = hstate_inode(inode);
3770 : 0 : int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
3771 : :
3772 [ # # ]: 0 : if (err)
3773 : : return err;
3774 : 0 : ClearPagePrivate(page);
3775 : :
3776 : : /*
3777 : : * set page dirty so that it will not be removed from cache/file
3778 : : * by non-hugetlbfs specific code paths.
3779 : : */
3780 : 0 : set_page_dirty(page);
3781 : :
3782 : 0 : spin_lock(&inode->i_lock);
3783 : 0 : inode->i_blocks += blocks_per_huge_page(h);
3784 : 0 : spin_unlock(&inode->i_lock);
3785 : 0 : return 0;
3786 : : }
3787 : :
3788 : 0 : static vm_fault_t hugetlb_no_page(struct mm_struct *mm,
3789 : : struct vm_area_struct *vma,
3790 : : struct address_space *mapping, pgoff_t idx,
3791 : : unsigned long address, pte_t *ptep, unsigned int flags)
3792 : : {
3793 [ # # ]: 0 : struct hstate *h = hstate_vma(vma);
3794 : 0 : vm_fault_t ret = VM_FAULT_SIGBUS;
3795 : 0 : int anon_rmap = 0;
3796 : 0 : unsigned long size;
3797 : 0 : struct page *page;
3798 : 0 : pte_t new_pte;
3799 : 0 : spinlock_t *ptl;
3800 [ # # ]: 0 : unsigned long haddr = address & huge_page_mask(h);
3801 : 0 : bool new_page = false;
3802 : :
3803 : : /*
3804 : : * Currently, we are forced to kill the process in the event the
3805 : : * original mapper has unmapped pages from the child due to a failed
3806 : : * COW. Warn that such a situation has occurred as it may not be obvious
3807 : : */
3808 [ # # ]: 0 : if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
3809 [ # # ]: 0 : pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3810 : : current->pid);
3811 : 0 : return ret;
3812 : : }
3813 : :
3814 : : /*
3815 : : * Use page lock to guard against racing truncation
3816 : : * before we get page_table_lock.
3817 : : */
3818 : 0 : retry:
3819 : 0 : page = find_lock_page(mapping, idx);
3820 [ # # ]: 0 : if (!page) {
3821 [ # # ]: 0 : size = i_size_read(mapping->host) >> huge_page_shift(h);
3822 [ # # ]: 0 : if (idx >= size)
3823 : 0 : goto out;
3824 : :
3825 : : /*
3826 : : * Check for page in userfault range
3827 : : */
3828 : 0 : if (userfaultfd_missing(vma)) {
3829 : : u32 hash;
3830 : : struct vm_fault vmf = {
3831 : : .vma = vma,
3832 : : .address = haddr,
3833 : : .flags = flags,
3834 : : /*
3835 : : * Hard to debug if it ends up being
3836 : : * used by a callee that assumes
3837 : : * something about the other
3838 : : * uninitialized fields... same as in
3839 : : * memory.c
3840 : : */
3841 : : };
3842 : :
3843 : : /*
3844 : : * hugetlb_fault_mutex must be dropped before
3845 : : * handling userfault. Reacquire after handling
3846 : : * fault to make calling code simpler.
3847 : : */
3848 : : hash = hugetlb_fault_mutex_hash(mapping, idx);
3849 : : mutex_unlock(&hugetlb_fault_mutex_table[hash]);
3850 : : ret = handle_userfault(&vmf, VM_UFFD_MISSING);
3851 : : mutex_lock(&hugetlb_fault_mutex_table[hash]);
3852 : : goto out;
3853 : : }
3854 : :
3855 : 0 : page = alloc_huge_page(vma, haddr, 0);
3856 [ # # ]: 0 : if (IS_ERR(page)) {
3857 : : /*
3858 : : * Returning error will result in faulting task being
3859 : : * sent SIGBUS. The hugetlb fault mutex prevents two
3860 : : * tasks from racing to fault in the same page which
3861 : : * could result in false unable to allocate errors.
3862 : : * Page migration does not take the fault mutex, but
3863 : : * does a clear then write of pte's under page table
3864 : : * lock. Page fault code could race with migration,
3865 : : * notice the clear pte and try to allocate a page
3866 : : * here. Before returning error, get ptl and make
3867 : : * sure there really is no pte entry.
3868 : : */
3869 : 0 : ptl = huge_pte_lock(h, mm, ptep);
3870 [ # # ]: 0 : if (!huge_pte_none(huge_ptep_get(ptep))) {
3871 : 0 : ret = 0;
3872 : 0 : spin_unlock(ptl);
3873 : 0 : goto out;
3874 : : }
3875 : 0 : spin_unlock(ptl);
3876 [ # # ]: 0 : ret = vmf_error(PTR_ERR(page));
3877 : 0 : goto out;
3878 : : }
3879 : 0 : clear_huge_page(page, address, pages_per_huge_page(h));
3880 : 0 : __SetPageUptodate(page);
3881 : 0 : new_page = true;
3882 : :
3883 [ # # ]: 0 : if (vma->vm_flags & VM_MAYSHARE) {
3884 : 0 : int err = huge_add_to_page_cache(page, mapping, idx);
3885 [ # # ]: 0 : if (err) {
3886 : 0 : put_page(page);
3887 [ # # ]: 0 : if (err == -EEXIST)
3888 : 0 : goto retry;
3889 : 0 : goto out;
3890 : : }
3891 : : } else {
3892 : 0 : lock_page(page);
3893 [ # # # # ]: 0 : if (unlikely(anon_vma_prepare(vma))) {
3894 : 0 : ret = VM_FAULT_OOM;
3895 : 0 : goto backout_unlocked;
3896 : : }
3897 : : anon_rmap = 1;
3898 : : }
3899 : : } else {
3900 : : /*
3901 : : * If memory error occurs between mmap() and fault, some process
3902 : : * don't have hwpoisoned swap entry for errored virtual address.
3903 : : * So we need to block hugepage fault by PG_hwpoison bit check.
3904 : : */
3905 : : if (unlikely(PageHWPoison(page))) {
3906 : : ret = VM_FAULT_HWPOISON |
3907 : : VM_FAULT_SET_HINDEX(hstate_index(h));
3908 : : goto backout_unlocked;
3909 : : }
3910 : : }
3911 : :
3912 : : /*
3913 : : * If we are going to COW a private mapping later, we examine the
3914 : : * pending reservations for this page now. This will ensure that
3915 : : * any allocations necessary to record that reservation occur outside
3916 : : * the spinlock.
3917 : : */
3918 [ # # # # ]: 0 : if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3919 [ # # ]: 0 : if (vma_needs_reservation(h, vma, haddr) < 0) {
3920 : 0 : ret = VM_FAULT_OOM;
3921 : 0 : goto backout_unlocked;
3922 : : }
3923 : : /* Just decrements count, does not deallocate */
3924 : 0 : vma_end_reservation(h, vma, haddr);
3925 : : }
3926 : :
3927 : 0 : ptl = huge_pte_lock(h, mm, ptep);
3928 [ # # ]: 0 : size = i_size_read(mapping->host) >> huge_page_shift(h);
3929 [ # # ]: 0 : if (idx >= size)
3930 : 0 : goto backout;
3931 : :
3932 : 0 : ret = 0;
3933 [ # # ]: 0 : if (!huge_pte_none(huge_ptep_get(ptep)))
3934 : 0 : goto backout;
3935 : :
3936 [ # # ]: 0 : if (anon_rmap) {
3937 : 0 : ClearPagePrivate(page);
3938 : 0 : hugepage_add_new_anon_rmap(page, vma, haddr);
3939 : : } else
3940 : 0 : page_dup_rmap(page, true);
3941 : 0 : new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
3942 : 0 : && (vma->vm_flags & VM_SHARED)));
3943 : 0 : set_huge_pte_at(mm, haddr, ptep, new_pte);
3944 : :
3945 : 0 : hugetlb_count_add(pages_per_huge_page(h), mm);
3946 [ # # # # ]: 0 : if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) {
3947 : : /* Optimization, do the COW without a second fault */
3948 : 0 : ret = hugetlb_cow(mm, vma, address, ptep, page, ptl);
3949 : : }
3950 : :
3951 : 0 : spin_unlock(ptl);
3952 : :
3953 : : /*
3954 : : * Only make newly allocated pages active. Existing pages found
3955 : : * in the pagecache could be !page_huge_active() if they have been
3956 : : * isolated for migration.
3957 : : */
3958 [ # # ]: 0 : if (new_page)
3959 : 0 : set_page_huge_active(page);
3960 : :
3961 : 0 : unlock_page(page);
3962 : : out:
3963 : : return ret;
3964 : :
3965 : 0 : backout:
3966 : 0 : spin_unlock(ptl);
3967 : 0 : backout_unlocked:
3968 : 0 : unlock_page(page);
3969 : 0 : restore_reserve_on_error(h, vma, haddr, page);
3970 : 0 : put_page(page);
3971 : 0 : goto out;
3972 : : }
3973 : :
3974 : : #ifdef CONFIG_SMP
3975 : 0 : u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
3976 : : {
3977 : 0 : unsigned long key[2];
3978 : 0 : u32 hash;
3979 : :
3980 : 0 : key[0] = (unsigned long) mapping;
3981 : 0 : key[1] = idx;
3982 : :
3983 : 0 : hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0);
3984 : :
3985 : 0 : return hash & (num_fault_mutexes - 1);
3986 : : }
3987 : : #else
3988 : : /*
3989 : : * For uniprocesor systems we always use a single mutex, so just
3990 : : * return 0 and avoid the hashing overhead.
3991 : : */
3992 : : u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx)
3993 : : {
3994 : : return 0;
3995 : : }
3996 : : #endif
3997 : :
3998 : 0 : vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3999 : : unsigned long address, unsigned int flags)
4000 : : {
4001 : 0 : pte_t *ptep, entry;
4002 : 0 : spinlock_t *ptl;
4003 : 0 : vm_fault_t ret;
4004 : 0 : u32 hash;
4005 : 0 : pgoff_t idx;
4006 : 0 : struct page *page = NULL;
4007 : 0 : struct page *pagecache_page = NULL;
4008 : 0 : struct hstate *h = hstate_vma(vma);
4009 : 0 : struct address_space *mapping;
4010 : 0 : int need_wait_lock = 0;
4011 : 0 : unsigned long haddr = address & huge_page_mask(h);
4012 : :
4013 : 0 : ptep = huge_pte_offset(mm, haddr, huge_page_size(h));
4014 [ # # ]: 0 : if (ptep) {
4015 [ # # ]: 0 : entry = huge_ptep_get(ptep);
4016 [ # # ]: 0 : if (unlikely(is_hugetlb_entry_migration(entry))) {
4017 : 0 : migration_entry_wait_huge(vma, mm, ptep);
4018 : 0 : return 0;
4019 : : } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry)))
4020 : : return VM_FAULT_HWPOISON_LARGE |
4021 : : VM_FAULT_SET_HINDEX(hstate_index(h));
4022 : : } else {
4023 : 0 : ptep = huge_pte_alloc(mm, haddr, huge_page_size(h));
4024 [ # # ]: 0 : if (!ptep)
4025 : : return VM_FAULT_OOM;
4026 : : }
4027 : :
4028 : 0 : mapping = vma->vm_file->f_mapping;
4029 : 0 : idx = vma_hugecache_offset(h, vma, haddr);
4030 : :
4031 : : /*
4032 : : * Serialize hugepage allocation and instantiation, so that we don't
4033 : : * get spurious allocation failures if two CPUs race to instantiate
4034 : : * the same page in the page cache.
4035 : : */
4036 : 0 : hash = hugetlb_fault_mutex_hash(mapping, idx);
4037 : 0 : mutex_lock(&hugetlb_fault_mutex_table[hash]);
4038 : :
4039 [ # # ]: 0 : entry = huge_ptep_get(ptep);
4040 [ # # ]: 0 : if (huge_pte_none(entry)) {
4041 : 0 : ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags);
4042 : 0 : goto out_mutex;
4043 : : }
4044 : :
4045 : 0 : ret = 0;
4046 : :
4047 : : /*
4048 : : * entry could be a migration/hwpoison entry at this point, so this
4049 : : * check prevents the kernel from going below assuming that we have
4050 : : * a active hugepage in pagecache. This goto expects the 2nd page fault,
4051 : : * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
4052 : : * handle it.
4053 : : */
4054 [ # # ]: 0 : if (!pte_present(entry))
4055 : 0 : goto out_mutex;
4056 : :
4057 : : /*
4058 : : * If we are going to COW the mapping later, we examine the pending
4059 : : * reservations for this page now. This will ensure that any
4060 : : * allocations necessary to record that reservation occur outside the
4061 : : * spinlock. For private mappings, we also lookup the pagecache
4062 : : * page now as it is used to determine if a reservation has been
4063 : : * consumed.
4064 : : */
4065 [ # # # # ]: 0 : if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
4066 [ # # ]: 0 : if (vma_needs_reservation(h, vma, haddr) < 0) {
4067 : 0 : ret = VM_FAULT_OOM;
4068 : 0 : goto out_mutex;
4069 : : }
4070 : : /* Just decrements count, does not deallocate */
4071 : 0 : vma_end_reservation(h, vma, haddr);
4072 : :
4073 [ # # ]: 0 : if (!(vma->vm_flags & VM_MAYSHARE))
4074 : 0 : pagecache_page = hugetlbfs_pagecache_page(h,
4075 : : vma, haddr);
4076 : : }
4077 : :
4078 : 0 : ptl = huge_pte_lock(h, mm, ptep);
4079 : :
4080 : : /* Check for a racing update before calling hugetlb_cow */
4081 [ # # ]: 0 : if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
4082 : 0 : goto out_ptl;
4083 : :
4084 : : /*
4085 : : * hugetlb_cow() requires page locks of pte_page(entry) and
4086 : : * pagecache_page, so here we need take the former one
4087 : : * when page != pagecache_page or !pagecache_page.
4088 : : */
4089 [ # # ]: 0 : page = pte_page(entry);
4090 [ # # ]: 0 : if (page != pagecache_page)
4091 [ # # # # ]: 0 : if (!trylock_page(page)) {
4092 : 0 : need_wait_lock = 1;
4093 : 0 : goto out_ptl;
4094 : : }
4095 : :
4096 [ # # ]: 0 : get_page(page);
4097 : :
4098 [ # # ]: 0 : if (flags & FAULT_FLAG_WRITE) {
4099 [ # # ]: 0 : if (!huge_pte_write(entry)) {
4100 : 0 : ret = hugetlb_cow(mm, vma, address, ptep,
4101 : : pagecache_page, ptl);
4102 : 0 : goto out_put_page;
4103 : : }
4104 : 0 : entry = huge_pte_mkdirty(entry);
4105 : : }
4106 : 0 : entry = pte_mkyoung(entry);
4107 : 0 : if (huge_ptep_set_access_flags(vma, haddr, ptep, entry,
4108 : 0 : flags & FAULT_FLAG_WRITE))
4109 : : update_mmu_cache(vma, haddr, ptep);
4110 : 0 : out_put_page:
4111 [ # # ]: 0 : if (page != pagecache_page)
4112 : 0 : unlock_page(page);
4113 : 0 : put_page(page);
4114 : 0 : out_ptl:
4115 : 0 : spin_unlock(ptl);
4116 : :
4117 [ # # ]: 0 : if (pagecache_page) {
4118 : 0 : unlock_page(pagecache_page);
4119 : 0 : put_page(pagecache_page);
4120 : : }
4121 : 0 : out_mutex:
4122 : 0 : mutex_unlock(&hugetlb_fault_mutex_table[hash]);
4123 : : /*
4124 : : * Generally it's safe to hold refcount during waiting page lock. But
4125 : : * here we just wait to defer the next page fault to avoid busy loop and
4126 : : * the page is not used after unlocked before returning from the current
4127 : : * page fault. So we are safe from accessing freed page, even if we wait
4128 : : * here without taking refcount.
4129 : : */
4130 [ # # ]: 0 : if (need_wait_lock)
4131 : 0 : wait_on_page_locked(page);
4132 : : return ret;
4133 : : }
4134 : :
4135 : : /*
4136 : : * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
4137 : : * modifications for huge pages.
4138 : : */
4139 : 0 : int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
4140 : : pte_t *dst_pte,
4141 : : struct vm_area_struct *dst_vma,
4142 : : unsigned long dst_addr,
4143 : : unsigned long src_addr,
4144 : : struct page **pagep)
4145 : : {
4146 : 0 : struct address_space *mapping;
4147 : 0 : pgoff_t idx;
4148 : 0 : unsigned long size;
4149 : 0 : int vm_shared = dst_vma->vm_flags & VM_SHARED;
4150 [ # # ]: 0 : struct hstate *h = hstate_vma(dst_vma);
4151 : 0 : pte_t _dst_pte;
4152 : 0 : spinlock_t *ptl;
4153 : 0 : int ret;
4154 : 0 : struct page *page;
4155 : :
4156 [ # # ]: 0 : if (!*pagep) {
4157 : 0 : ret = -ENOMEM;
4158 : 0 : page = alloc_huge_page(dst_vma, dst_addr, 0);
4159 [ # # ]: 0 : if (IS_ERR(page))
4160 : 0 : goto out;
4161 : :
4162 : 0 : ret = copy_huge_page_from_user(page,
4163 : : (const void __user *) src_addr,
4164 : : pages_per_huge_page(h), false);
4165 : :
4166 : : /* fallback to copy_from_user outside mmap_sem */
4167 [ # # ]: 0 : if (unlikely(ret)) {
4168 : 0 : ret = -ENOENT;
4169 : 0 : *pagep = page;
4170 : : /* don't free the page */
4171 : 0 : goto out;
4172 : : }
4173 : : } else {
4174 : 0 : page = *pagep;
4175 : 0 : *pagep = NULL;
4176 : : }
4177 : :
4178 : : /*
4179 : : * The memory barrier inside __SetPageUptodate makes sure that
4180 : : * preceding stores to the page contents become visible before
4181 : : * the set_pte_at() write.
4182 : : */
4183 : 0 : __SetPageUptodate(page);
4184 : :
4185 : 0 : mapping = dst_vma->vm_file->f_mapping;
4186 [ # # ]: 0 : idx = vma_hugecache_offset(h, dst_vma, dst_addr);
4187 : :
4188 : : /*
4189 : : * If shared, add to page cache
4190 : : */
4191 [ # # ]: 0 : if (vm_shared) {
4192 [ # # ]: 0 : size = i_size_read(mapping->host) >> huge_page_shift(h);
4193 : 0 : ret = -EFAULT;
4194 [ # # ]: 0 : if (idx >= size)
4195 : 0 : goto out_release_nounlock;
4196 : :
4197 : : /*
4198 : : * Serialization between remove_inode_hugepages() and
4199 : : * huge_add_to_page_cache() below happens through the
4200 : : * hugetlb_fault_mutex_table that here must be hold by
4201 : : * the caller.
4202 : : */
4203 : 0 : ret = huge_add_to_page_cache(page, mapping, idx);
4204 [ # # ]: 0 : if (ret)
4205 : 0 : goto out_release_nounlock;
4206 : : }
4207 : :
4208 [ # # ]: 0 : ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
4209 : 0 : spin_lock(ptl);
4210 : :
4211 : : /*
4212 : : * Recheck the i_size after holding PT lock to make sure not
4213 : : * to leave any page mapped (as page_mapped()) beyond the end
4214 : : * of the i_size (remove_inode_hugepages() is strict about
4215 : : * enforcing that). If we bail out here, we'll also leave a
4216 : : * page in the radix tree in the vm_shared case beyond the end
4217 : : * of the i_size, but remove_inode_hugepages() will take care
4218 : : * of it as soon as we drop the hugetlb_fault_mutex_table.
4219 : : */
4220 [ # # ]: 0 : size = i_size_read(mapping->host) >> huge_page_shift(h);
4221 : 0 : ret = -EFAULT;
4222 [ # # ]: 0 : if (idx >= size)
4223 : 0 : goto out_release_unlock;
4224 : :
4225 : 0 : ret = -EEXIST;
4226 [ # # ]: 0 : if (!huge_pte_none(huge_ptep_get(dst_pte)))
4227 : 0 : goto out_release_unlock;
4228 : :
4229 [ # # ]: 0 : if (vm_shared) {
4230 : 0 : page_dup_rmap(page, true);
4231 : : } else {
4232 : 0 : ClearPagePrivate(page);
4233 : 0 : hugepage_add_new_anon_rmap(page, dst_vma, dst_addr);
4234 : : }
4235 : :
4236 : 0 : _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE);
4237 [ # # ]: 0 : if (dst_vma->vm_flags & VM_WRITE)
4238 : 0 : _dst_pte = huge_pte_mkdirty(_dst_pte);
4239 : 0 : _dst_pte = pte_mkyoung(_dst_pte);
4240 : :
4241 : 0 : set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
4242 : :
4243 : 0 : (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte,
4244 : 0 : dst_vma->vm_flags & VM_WRITE);
4245 : 0 : hugetlb_count_add(pages_per_huge_page(h), dst_mm);
4246 : :
4247 : : /* No need to invalidate - it was non-present before */
4248 : 0 : update_mmu_cache(dst_vma, dst_addr, dst_pte);
4249 : :
4250 : 0 : spin_unlock(ptl);
4251 : 0 : set_page_huge_active(page);
4252 [ # # ]: 0 : if (vm_shared)
4253 : 0 : unlock_page(page);
4254 : : ret = 0;
4255 : 0 : out:
4256 : 0 : return ret;
4257 : 0 : out_release_unlock:
4258 : 0 : spin_unlock(ptl);
4259 [ # # ]: 0 : if (vm_shared)
4260 : 0 : unlock_page(page);
4261 : 0 : out_release_nounlock:
4262 : 0 : put_page(page);
4263 : 0 : goto out;
4264 : : }
4265 : :
4266 : 0 : long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
4267 : : struct page **pages, struct vm_area_struct **vmas,
4268 : : unsigned long *position, unsigned long *nr_pages,
4269 : : long i, unsigned int flags, int *nonblocking)
4270 : : {
4271 : 0 : unsigned long pfn_offset;
4272 : 0 : unsigned long vaddr = *position;
4273 : 0 : unsigned long remainder = *nr_pages;
4274 : 0 : struct hstate *h = hstate_vma(vma);
4275 : 0 : int err = -EFAULT;
4276 : :
4277 [ # # # # ]: 0 : while (vaddr < vma->vm_end && remainder) {
4278 : 0 : pte_t *pte;
4279 : 0 : spinlock_t *ptl = NULL;
4280 : 0 : int absent;
4281 : 0 : struct page *page;
4282 : :
4283 : : /*
4284 : : * If we have a pending SIGKILL, don't keep faulting pages and
4285 : : * potentially allocating memory.
4286 : : */
4287 [ # # ]: 0 : if (fatal_signal_pending(current)) {
4288 : : remainder = 0;
4289 : : break;
4290 : : }
4291 : :
4292 : : /*
4293 : : * Some archs (sparc64, sh*) have multiple pte_ts to
4294 : : * each hugepage. We have to make sure we get the
4295 : : * first, for the page indexing below to work.
4296 : : *
4297 : : * Note that page table lock is not held when pte is null.
4298 : : */
4299 : 0 : pte = huge_pte_offset(mm, vaddr & huge_page_mask(h),
4300 : : huge_page_size(h));
4301 [ # # ]: 0 : if (pte)
4302 : 0 : ptl = huge_pte_lock(h, mm, pte);
4303 [ # # # # ]: 0 : absent = !pte || huge_pte_none(huge_ptep_get(pte));
4304 : :
4305 : : /*
4306 : : * When coredumping, it suits get_dump_page if we just return
4307 : : * an error where there's an empty slot with no huge pagecache
4308 : : * to back it. This way, we avoid allocating a hugepage, and
4309 : : * the sparse dumpfile avoids allocating disk blocks, but its
4310 : : * huge holes still show up with zeroes where they need to be.
4311 : : */
4312 [ # # # # ]: 0 : if (absent && (flags & FOLL_DUMP) &&
4313 : 0 : !hugetlbfs_pagecache_present(h, vma, vaddr)) {
4314 [ # # ]: 0 : if (pte)
4315 : 0 : spin_unlock(ptl);
4316 : : remainder = 0;
4317 : : break;
4318 : : }
4319 : :
4320 : : /*
4321 : : * We need call hugetlb_fault for both hugepages under migration
4322 : : * (in which case hugetlb_fault waits for the migration,) and
4323 : : * hwpoisoned hugepages (in which case we need to prevent the
4324 : : * caller from accessing to them.) In order to do this, we use
4325 : : * here is_swap_pte instead of is_hugetlb_entry_migration and
4326 : : * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4327 : : * both cases, and because we can't follow correct pages
4328 : : * directly from any kind of swap entries.
4329 : : */
4330 [ # # ]: 0 : if (absent || is_swap_pte(huge_ptep_get(pte)) ||
4331 [ # # # # ]: 0 : ((flags & FOLL_WRITE) &&
4332 : : !huge_pte_write(huge_ptep_get(pte)))) {
4333 : 0 : vm_fault_t ret;
4334 : 0 : unsigned int fault_flags = 0;
4335 : :
4336 [ # # ]: 0 : if (pte)
4337 : 0 : spin_unlock(ptl);
4338 : 0 : if (flags & FOLL_WRITE)
4339 : : fault_flags |= FAULT_FLAG_WRITE;
4340 [ # # ]: 0 : if (nonblocking)
4341 : 0 : fault_flags |= FAULT_FLAG_ALLOW_RETRY;
4342 [ # # ]: 0 : if (flags & FOLL_NOWAIT)
4343 : 0 : fault_flags |= FAULT_FLAG_ALLOW_RETRY |
4344 : : FAULT_FLAG_RETRY_NOWAIT;
4345 [ # # ]: 0 : if (flags & FOLL_TRIED) {
4346 : 0 : VM_WARN_ON_ONCE(fault_flags &
4347 : : FAULT_FLAG_ALLOW_RETRY);
4348 : 0 : fault_flags |= FAULT_FLAG_TRIED;
4349 : : }
4350 : 0 : ret = hugetlb_fault(mm, vma, vaddr, fault_flags);
4351 [ # # ]: 0 : if (ret & VM_FAULT_ERROR) {
4352 [ # # ]: 0 : err = vm_fault_to_errno(ret, flags);
4353 : : remainder = 0;
4354 : : break;
4355 : : }
4356 [ # # ]: 0 : if (ret & VM_FAULT_RETRY) {
4357 [ # # ]: 0 : if (nonblocking &&
4358 [ # # ]: 0 : !(fault_flags & FAULT_FLAG_RETRY_NOWAIT))
4359 : 0 : *nonblocking = 0;
4360 : 0 : *nr_pages = 0;
4361 : : /*
4362 : : * VM_FAULT_RETRY must not return an
4363 : : * error, it will return zero
4364 : : * instead.
4365 : : *
4366 : : * No need to update "position" as the
4367 : : * caller will not check it after
4368 : : * *nr_pages is set to 0.
4369 : : */
4370 : 0 : return i;
4371 : : }
4372 : 0 : continue;
4373 : : }
4374 : :
4375 [ # # ]: 0 : pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
4376 [ # # ]: 0 : page = pte_page(huge_ptep_get(pte));
4377 : :
4378 : : /*
4379 : : * Instead of doing 'try_get_page()' below in the same_page
4380 : : * loop, just check the count once here.
4381 : : */
4382 [ # # # # ]: 0 : if (unlikely(page_count(page) <= 0)) {
4383 [ # # ]: 0 : if (pages) {
4384 : 0 : spin_unlock(ptl);
4385 : 0 : remainder = 0;
4386 : 0 : err = -ENOMEM;
4387 : 0 : break;
4388 : : }
4389 : : }
4390 : :
4391 : : /*
4392 : : * If subpage information not requested, update counters
4393 : : * and skip the same_page loop below.
4394 : : */
4395 [ # # # # : 0 : if (!pages && !vmas && !pfn_offset &&
# # ]
4396 [ # # ]: 0 : (vaddr + huge_page_size(h) < vma->vm_end) &&
4397 [ # # ]: 0 : (remainder >= pages_per_huge_page(h))) {
4398 : 0 : vaddr += huge_page_size(h);
4399 : 0 : remainder -= pages_per_huge_page(h);
4400 : 0 : i += pages_per_huge_page(h);
4401 : 0 : spin_unlock(ptl);
4402 : 0 : continue;
4403 : : }
4404 : :
4405 : 0 : same_page:
4406 [ # # ]: 0 : if (pages) {
4407 [ # # ]: 0 : pages[i] = mem_map_offset(page, pfn_offset);
4408 [ # # ]: 0 : get_page(pages[i]);
4409 : : }
4410 : :
4411 [ # # ]: 0 : if (vmas)
4412 : 0 : vmas[i] = vma;
4413 : :
4414 : 0 : vaddr += PAGE_SIZE;
4415 : 0 : ++pfn_offset;
4416 : 0 : --remainder;
4417 : 0 : ++i;
4418 [ # # # # : 0 : if (vaddr < vma->vm_end && remainder &&
# # ]
4419 [ # # ]: 0 : pfn_offset < pages_per_huge_page(h)) {
4420 : : /*
4421 : : * We use pfn_offset to avoid touching the pageframes
4422 : : * of this compound page.
4423 : : */
4424 : 0 : goto same_page;
4425 : : }
4426 : 0 : spin_unlock(ptl);
4427 : : }
4428 : 0 : *nr_pages = remainder;
4429 : : /*
4430 : : * setting position is actually required only if remainder is
4431 : : * not zero but it's faster not to add a "if (remainder)"
4432 : : * branch.
4433 : : */
4434 : 0 : *position = vaddr;
4435 : :
4436 [ # # ]: 0 : return i ? i : err;
4437 : : }
4438 : :
4439 : : #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4440 : : /*
4441 : : * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4442 : : * implement this.
4443 : : */
4444 : : #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4445 : : #endif
4446 : :
4447 : 0 : unsigned long hugetlb_change_protection(struct vm_area_struct *vma,
4448 : : unsigned long address, unsigned long end, pgprot_t newprot)
4449 : : {
4450 : 0 : struct mm_struct *mm = vma->vm_mm;
4451 : 0 : unsigned long start = address;
4452 : 0 : pte_t *ptep;
4453 : 0 : pte_t pte;
4454 : 0 : struct hstate *h = hstate_vma(vma);
4455 : 0 : unsigned long pages = 0;
4456 : 0 : bool shared_pmd = false;
4457 : 0 : struct mmu_notifier_range range;
4458 : :
4459 : : /*
4460 : : * In the case of shared PMDs, the area to flush could be beyond
4461 : : * start/end. Set range.start/range.end to cover the maximum possible
4462 : : * range if PMD sharing is possible.
4463 : : */
4464 : 0 : mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA,
4465 : : 0, vma, mm, start, end);
4466 : 0 : adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end);
4467 : :
4468 [ # # ]: 0 : BUG_ON(address >= end);
4469 : 0 : flush_cache_range(vma, range.start, range.end);
4470 : :
4471 : 0 : mmu_notifier_invalidate_range_start(&range);
4472 : 0 : i_mmap_lock_write(vma->vm_file->f_mapping);
4473 [ # # ]: 0 : for (; address < end; address += huge_page_size(h)) {
4474 : 0 : spinlock_t *ptl;
4475 : 0 : ptep = huge_pte_offset(mm, address, huge_page_size(h));
4476 [ # # ]: 0 : if (!ptep)
4477 : 0 : continue;
4478 : 0 : ptl = huge_pte_lock(h, mm, ptep);
4479 [ # # ]: 0 : if (huge_pmd_unshare(mm, &address, ptep)) {
4480 : 0 : pages++;
4481 : 0 : spin_unlock(ptl);
4482 : 0 : shared_pmd = true;
4483 : 0 : continue;
4484 : : }
4485 [ # # ]: 0 : pte = huge_ptep_get(ptep);
4486 [ # # ]: 0 : if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) {
4487 : : spin_unlock(ptl);
4488 : : continue;
4489 : : }
4490 [ # # ]: 0 : if (unlikely(is_hugetlb_entry_migration(pte))) {
4491 [ # # ]: 0 : swp_entry_t entry = pte_to_swp_entry(pte);
4492 : :
4493 [ # # ]: 0 : if (is_write_migration_entry(entry)) {
4494 : 0 : pte_t newpte;
4495 : :
4496 : 0 : make_migration_entry_read(&entry);
4497 : 0 : newpte = swp_entry_to_pte(entry);
4498 : 0 : set_huge_swap_pte_at(mm, address, ptep,
4499 : : newpte, huge_page_size(h));
4500 : 0 : pages++;
4501 : : }
4502 : 0 : spin_unlock(ptl);
4503 : 0 : continue;
4504 : : }
4505 [ # # ]: 0 : if (!huge_pte_none(pte)) {
4506 : 0 : pte_t old_pte;
4507 : :
4508 : 0 : old_pte = huge_ptep_modify_prot_start(vma, address, ptep);
4509 : 0 : pte = pte_mkhuge(huge_pte_modify(old_pte, newprot));
4510 : 0 : pte = arch_make_huge_pte(pte, vma, NULL, 0);
4511 : 0 : huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte);
4512 : 0 : pages++;
4513 : : }
4514 : 0 : spin_unlock(ptl);
4515 : : }
4516 : : /*
4517 : : * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4518 : : * may have cleared our pud entry and done put_page on the page table:
4519 : : * once we release i_mmap_rwsem, another task can do the final put_page
4520 : : * and that page table be reused and filled with junk. If we actually
4521 : : * did unshare a page of pmds, flush the range corresponding to the pud.
4522 : : */
4523 [ # # ]: 0 : if (shared_pmd)
4524 [ # # ]: 0 : flush_hugetlb_tlb_range(vma, range.start, range.end);
4525 : : else
4526 [ # # ]: 0 : flush_hugetlb_tlb_range(vma, start, end);
4527 : : /*
4528 : : * No need to call mmu_notifier_invalidate_range() we are downgrading
4529 : : * page table protection not changing it to point to a new page.
4530 : : *
4531 : : * See Documentation/vm/mmu_notifier.rst
4532 : : */
4533 : 0 : i_mmap_unlock_write(vma->vm_file->f_mapping);
4534 : 0 : mmu_notifier_invalidate_range_end(&range);
4535 : :
4536 : 0 : return pages << h->order;
4537 : : }
4538 : :
4539 : 0 : int hugetlb_reserve_pages(struct inode *inode,
4540 : : long from, long to,
4541 : : struct vm_area_struct *vma,
4542 : : vm_flags_t vm_flags)
4543 : : {
4544 : 0 : long ret, chg;
4545 [ # # ]: 0 : struct hstate *h = hstate_inode(inode);
4546 : 0 : struct hugepage_subpool *spool = subpool_inode(inode);
4547 : 0 : struct resv_map *resv_map;
4548 : 0 : long gbl_reserve;
4549 : :
4550 : : /* This should never happen */
4551 [ # # ]: 0 : if (from > to) {
4552 : : VM_WARN(1, "%s called with a negative range\n", __func__);
4553 : : return -EINVAL;
4554 : : }
4555 : :
4556 : : /*
4557 : : * Only apply hugepage reservation if asked. At fault time, an
4558 : : * attempt will be made for VM_NORESERVE to allocate a page
4559 : : * without using reserves
4560 : : */
4561 [ # # ]: 0 : if (vm_flags & VM_NORESERVE)
4562 : : return 0;
4563 : :
4564 : : /*
4565 : : * Shared mappings base their reservation on the number of pages that
4566 : : * are already allocated on behalf of the file. Private mappings need
4567 : : * to reserve the full area even if read-only as mprotect() may be
4568 : : * called to make the mapping read-write. Assume !vma is a shm mapping
4569 : : */
4570 [ # # # # ]: 0 : if (!vma || vma->vm_flags & VM_MAYSHARE) {
4571 : : /*
4572 : : * resv_map can not be NULL as hugetlb_reserve_pages is only
4573 : : * called for inodes for which resv_maps were created (see
4574 : : * hugetlbfs_get_inode).
4575 : : */
4576 : 0 : resv_map = inode_resv_map(inode);
4577 : :
4578 : 0 : chg = region_chg(resv_map, from, to);
4579 : :
4580 : : } else {
4581 : 0 : resv_map = resv_map_alloc();
4582 [ # # ]: 0 : if (!resv_map)
4583 : : return -ENOMEM;
4584 : :
4585 : 0 : chg = to - from;
4586 : :
4587 : 0 : set_vma_resv_map(vma, resv_map);
4588 : 0 : set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
4589 : : }
4590 : :
4591 [ # # ]: 0 : if (chg < 0) {
4592 : 0 : ret = chg;
4593 : 0 : goto out_err;
4594 : : }
4595 : :
4596 : : /*
4597 : : * There must be enough pages in the subpool for the mapping. If
4598 : : * the subpool has a minimum size, there may be some global
4599 : : * reservations already in place (gbl_reserve).
4600 : : */
4601 : 0 : gbl_reserve = hugepage_subpool_get_pages(spool, chg);
4602 [ # # ]: 0 : if (gbl_reserve < 0) {
4603 : 0 : ret = -ENOSPC;
4604 : 0 : goto out_err;
4605 : : }
4606 : :
4607 : : /*
4608 : : * Check enough hugepages are available for the reservation.
4609 : : * Hand the pages back to the subpool if there are not
4610 : : */
4611 : 0 : ret = hugetlb_acct_memory(h, gbl_reserve);
4612 [ # # ]: 0 : if (ret < 0) {
4613 : : /* put back original number of pages, chg */
4614 : 0 : (void)hugepage_subpool_put_pages(spool, chg);
4615 : 0 : goto out_err;
4616 : : }
4617 : :
4618 : : /*
4619 : : * Account for the reservations made. Shared mappings record regions
4620 : : * that have reservations as they are shared by multiple VMAs.
4621 : : * When the last VMA disappears, the region map says how much
4622 : : * the reservation was and the page cache tells how much of
4623 : : * the reservation was consumed. Private mappings are per-VMA and
4624 : : * only the consumed reservations are tracked. When the VMA
4625 : : * disappears, the original reservation is the VMA size and the
4626 : : * consumed reservations are stored in the map. Hence, nothing
4627 : : * else has to be done for private mappings here
4628 : : */
4629 [ # # # # ]: 0 : if (!vma || vma->vm_flags & VM_MAYSHARE) {
4630 : 0 : long add = region_add(resv_map, from, to);
4631 : :
4632 [ # # ]: 0 : if (unlikely(chg > add)) {
4633 : : /*
4634 : : * pages in this range were added to the reserve
4635 : : * map between region_chg and region_add. This
4636 : : * indicates a race with alloc_huge_page. Adjust
4637 : : * the subpool and reserve counts modified above
4638 : : * based on the difference.
4639 : : */
4640 : 0 : long rsv_adjust;
4641 : :
4642 : 0 : rsv_adjust = hugepage_subpool_put_pages(spool,
4643 : : chg - add);
4644 : 0 : hugetlb_acct_memory(h, -rsv_adjust);
4645 : : }
4646 : : }
4647 : : return 0;
4648 : 0 : out_err:
4649 [ # # # # ]: 0 : if (!vma || vma->vm_flags & VM_MAYSHARE)
4650 : : /* Don't call region_abort if region_chg failed */
4651 [ # # ]: 0 : if (chg >= 0)
4652 : 0 : region_abort(resv_map, from, to);
4653 [ # # # # ]: 0 : if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
4654 : 0 : kref_put(&resv_map->refs, resv_map_release);
4655 : 0 : return ret;
4656 : : }
4657 : :
4658 : 0 : long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
4659 : : long freed)
4660 : : {
4661 [ # # ]: 0 : struct hstate *h = hstate_inode(inode);
4662 : 0 : struct resv_map *resv_map = inode_resv_map(inode);
4663 : 0 : long chg = 0;
4664 : 0 : struct hugepage_subpool *spool = subpool_inode(inode);
4665 : 0 : long gbl_reserve;
4666 : :
4667 : : /*
4668 : : * Since this routine can be called in the evict inode path for all
4669 : : * hugetlbfs inodes, resv_map could be NULL.
4670 : : */
4671 [ # # ]: 0 : if (resv_map) {
4672 : 0 : chg = region_del(resv_map, start, end);
4673 : : /*
4674 : : * region_del() can fail in the rare case where a region
4675 : : * must be split and another region descriptor can not be
4676 : : * allocated. If end == LONG_MAX, it will not fail.
4677 : : */
4678 [ # # ]: 0 : if (chg < 0)
4679 : : return chg;
4680 : : }
4681 : :
4682 : 0 : spin_lock(&inode->i_lock);
4683 : 0 : inode->i_blocks -= (blocks_per_huge_page(h) * freed);
4684 : 0 : spin_unlock(&inode->i_lock);
4685 : :
4686 : : /*
4687 : : * If the subpool has a minimum size, the number of global
4688 : : * reservations to be released may be adjusted.
4689 : : */
4690 : 0 : gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed));
4691 : 0 : hugetlb_acct_memory(h, -gbl_reserve);
4692 : :
4693 : 0 : return 0;
4694 : : }
4695 : :
4696 : : #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4697 : : static unsigned long page_table_shareable(struct vm_area_struct *svma,
4698 : : struct vm_area_struct *vma,
4699 : : unsigned long addr, pgoff_t idx)
4700 : : {
4701 : : unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
4702 : : svma->vm_start;
4703 : : unsigned long sbase = saddr & PUD_MASK;
4704 : : unsigned long s_end = sbase + PUD_SIZE;
4705 : :
4706 : : /* Allow segments to share if only one is marked locked */
4707 : : unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK;
4708 : : unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK;
4709 : :
4710 : : /*
4711 : : * match the virtual addresses, permission and the alignment of the
4712 : : * page table page.
4713 : : */
4714 : : if (pmd_index(addr) != pmd_index(saddr) ||
4715 : : vm_flags != svm_flags ||
4716 : : sbase < svma->vm_start || svma->vm_end < s_end)
4717 : : return 0;
4718 : :
4719 : : return saddr;
4720 : : }
4721 : :
4722 : 0 : static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
4723 : : {
4724 : 0 : unsigned long base = addr & PUD_MASK;
4725 : 0 : unsigned long end = base + PUD_SIZE;
4726 : :
4727 : : /*
4728 : : * check on proper vm_flags and page table alignment
4729 : : */
4730 [ # # ]: 0 : if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end))
4731 : 0 : return true;
4732 : : return false;
4733 : : }
4734 : :
4735 : : /*
4736 : : * Determine if start,end range within vma could be mapped by shared pmd.
4737 : : * If yes, adjust start and end to cover range associated with possible
4738 : : * shared pmd mappings.
4739 : : */
4740 : 0 : void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4741 : : unsigned long *start, unsigned long *end)
4742 : : {
4743 : 0 : unsigned long check_addr = *start;
4744 : :
4745 [ # # ]: 0 : if (!(vma->vm_flags & VM_MAYSHARE))
4746 : : return;
4747 : :
4748 [ # # ]: 0 : for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
4749 : 0 : unsigned long a_start = check_addr & PUD_MASK;
4750 : 0 : unsigned long a_end = a_start + PUD_SIZE;
4751 : :
4752 : : /*
4753 : : * If sharing is possible, adjust start/end if necessary.
4754 : : */
4755 [ # # # # ]: 0 : if (range_in_vma(vma, a_start, a_end)) {
4756 [ # # ]: 0 : if (a_start < *start)
4757 : 0 : *start = a_start;
4758 [ # # ]: 0 : if (a_end > *end)
4759 : 0 : *end = a_end;
4760 : : }
4761 : : }
4762 : : }
4763 : :
4764 : : /*
4765 : : * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4766 : : * and returns the corresponding pte. While this is not necessary for the
4767 : : * !shared pmd case because we can allocate the pmd later as well, it makes the
4768 : : * code much cleaner. pmd allocation is essential for the shared case because
4769 : : * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4770 : : * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4771 : : * bad pmd for sharing.
4772 : : */
4773 : 0 : pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4774 : : {
4775 : 0 : struct vm_area_struct *vma = find_vma(mm, addr);
4776 : 0 : struct address_space *mapping = vma->vm_file->f_mapping;
4777 : 0 : pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
4778 : 0 : vma->vm_pgoff;
4779 : 0 : struct vm_area_struct *svma;
4780 : 0 : unsigned long saddr;
4781 : 0 : pte_t *spte = NULL;
4782 : 0 : pte_t *pte;
4783 : 0 : spinlock_t *ptl;
4784 : :
4785 [ # # ]: 0 : if (!vma_shareable(vma, addr))
4786 : 0 : return (pte_t *)pmd_alloc(mm, pud, addr);
4787 : :
4788 : 0 : i_mmap_lock_read(mapping);
4789 [ # # ]: 0 : vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) {
4790 [ # # ]: 0 : if (svma == vma)
4791 : 0 : continue;
4792 : :
4793 : 0 : saddr = page_table_shareable(svma, vma, addr, idx);
4794 [ # # ]: 0 : if (saddr) {
4795 : 0 : spte = huge_pte_offset(svma->vm_mm, saddr,
4796 : : vma_mmu_pagesize(svma));
4797 [ # # ]: 0 : if (spte) {
4798 [ # # # # ]: 0 : get_page(virt_to_page(spte));
4799 : : break;
4800 : : }
4801 : : }
4802 : : }
4803 : :
4804 [ # # ]: 0 : if (!spte)
4805 : 0 : goto out;
4806 : :
4807 : 0 : ptl = huge_pte_lock(hstate_vma(vma), mm, spte);
4808 [ # # ]: 0 : if (pud_none(*pud)) {
4809 : 0 : pud_populate(mm, pud,
4810 [ # # ]: 0 : (pmd_t *)((unsigned long)spte & PAGE_MASK));
4811 : 0 : mm_inc_nr_pmds(mm);
4812 : : } else {
4813 [ # # ]: 0 : put_page(virt_to_page(spte));
4814 : : }
4815 : 0 : spin_unlock(ptl);
4816 : 0 : out:
4817 : 0 : pte = (pte_t *)pmd_alloc(mm, pud, addr);
4818 : 0 : i_mmap_unlock_read(mapping);
4819 : 0 : return pte;
4820 : : }
4821 : :
4822 : : /*
4823 : : * unmap huge page backed by shared pte.
4824 : : *
4825 : : * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4826 : : * indicated by page_count > 1, unmap is achieved by clearing pud and
4827 : : * decrementing the ref count. If count == 1, the pte page is not shared.
4828 : : *
4829 : : * called with page table lock held.
4830 : : *
4831 : : * returns: 1 successfully unmapped a shared pte page
4832 : : * 0 the underlying pte page is not shared, or it is the last user
4833 : : */
4834 : 0 : int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4835 : : {
4836 : 0 : pgd_t *pgd = pgd_offset(mm, *addr);
4837 : 0 : p4d_t *p4d = p4d_offset(pgd, *addr);
4838 [ # # ]: 0 : pud_t *pud = pud_offset(p4d, *addr);
4839 : :
4840 [ # # # # : 0 : BUG_ON(page_count(virt_to_page(ptep)) == 0);
# # ]
4841 [ # # # # : 0 : if (page_count(virt_to_page(ptep)) == 1)
# # ]
4842 : : return 0;
4843 : :
4844 [ # # ]: 0 : pud_clear(pud);
4845 [ # # ]: 0 : put_page(virt_to_page(ptep));
4846 : 0 : mm_dec_nr_pmds(mm);
4847 : 0 : *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
4848 : 0 : return 1;
4849 : : }
4850 : : #define want_pmd_share() (1)
4851 : : #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4852 : : pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
4853 : : {
4854 : : return NULL;
4855 : : }
4856 : :
4857 : : int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
4858 : : {
4859 : : return 0;
4860 : : }
4861 : :
4862 : : void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
4863 : : unsigned long *start, unsigned long *end)
4864 : : {
4865 : : }
4866 : : #define want_pmd_share() (0)
4867 : : #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4868 : :
4869 : : #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4870 : 0 : pte_t *huge_pte_alloc(struct mm_struct *mm,
4871 : : unsigned long addr, unsigned long sz)
4872 : : {
4873 : 0 : pgd_t *pgd;
4874 : 0 : p4d_t *p4d;
4875 : 0 : pud_t *pud;
4876 : 0 : pte_t *pte = NULL;
4877 : :
4878 : 0 : pgd = pgd_offset(mm, addr);
4879 : 0 : p4d = p4d_alloc(mm, pgd, addr);
4880 [ # # ]: 0 : if (!p4d)
4881 : : return NULL;
4882 : 0 : pud = pud_alloc(mm, p4d, addr);
4883 [ # # ]: 0 : if (pud) {
4884 [ # # ]: 0 : if (sz == PUD_SIZE) {
4885 : : pte = (pte_t *)pud;
4886 : : } else {
4887 [ # # ]: 0 : BUG_ON(sz != PMD_SIZE);
4888 [ # # ]: 0 : if (want_pmd_share() && pud_none(*pud))
4889 : 0 : pte = huge_pmd_share(mm, addr, pud);
4890 : : else
4891 : 0 : pte = (pte_t *)pmd_alloc(mm, pud, addr);
4892 : : }
4893 : : }
4894 [ # # # # : 0 : BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte));
# # ]
4895 : :
4896 : : return pte;
4897 : : }
4898 : :
4899 : : /*
4900 : : * huge_pte_offset() - Walk the page table to resolve the hugepage
4901 : : * entry at address @addr
4902 : : *
4903 : : * Return: Pointer to page table or swap entry (PUD or PMD) for
4904 : : * address @addr, or NULL if a p*d_none() entry is encountered and the
4905 : : * size @sz doesn't match the hugepage size at this level of the page
4906 : : * table.
4907 : : */
4908 : 0 : pte_t *huge_pte_offset(struct mm_struct *mm,
4909 : : unsigned long addr, unsigned long sz)
4910 : : {
4911 : 0 : pgd_t *pgd;
4912 : 0 : p4d_t *p4d;
4913 : 0 : pud_t *pud;
4914 : 0 : pmd_t *pmd;
4915 : :
4916 : 0 : pgd = pgd_offset(mm, addr);
4917 [ # # ]: 0 : if (!pgd_present(*pgd))
4918 : : return NULL;
4919 : 0 : p4d = p4d_offset(pgd, addr);
4920 [ # # ]: 0 : if (!p4d_present(*p4d))
4921 : : return NULL;
4922 : :
4923 [ # # ]: 0 : pud = pud_offset(p4d, addr);
4924 [ # # # # ]: 0 : if (sz != PUD_SIZE && pud_none(*pud))
4925 : : return NULL;
4926 : : /* hugepage or swap? */
4927 [ # # # # ]: 0 : if (pud_huge(*pud) || !pud_present(*pud))
4928 : : return (pte_t *)pud;
4929 : :
4930 [ # # ]: 0 : pmd = pmd_offset(pud, addr);
4931 [ # # # # ]: 0 : if (sz != PMD_SIZE && pmd_none(*pmd))
4932 : : return NULL;
4933 : : /* hugepage or swap? */
4934 [ # # # # ]: 0 : if (pmd_huge(*pmd) || !pmd_present(*pmd))
4935 : 0 : return (pte_t *)pmd;
4936 : :
4937 : : return NULL;
4938 : : }
4939 : :
4940 : : #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4941 : :
4942 : : /*
4943 : : * These functions are overwritable if your architecture needs its own
4944 : : * behavior.
4945 : : */
4946 : : struct page * __weak
4947 : 45997 : follow_huge_addr(struct mm_struct *mm, unsigned long address,
4948 : : int write)
4949 : : {
4950 : 45997 : return ERR_PTR(-EINVAL);
4951 : : }
4952 : :
4953 : : struct page * __weak
4954 : 0 : follow_huge_pd(struct vm_area_struct *vma,
4955 : : unsigned long address, hugepd_t hpd, int flags, int pdshift)
4956 : : {
4957 : 0 : WARN(1, "hugepd follow called with no support for hugepage directory format\n");
4958 : 0 : return NULL;
4959 : : }
4960 : :
4961 : : struct page * __weak
4962 : 0 : follow_huge_pmd(struct mm_struct *mm, unsigned long address,
4963 : : pmd_t *pmd, int flags)
4964 : : {
4965 : 0 : struct page *page = NULL;
4966 : 0 : spinlock_t *ptl;
4967 : 0 : pte_t pte;
4968 : 0 : retry:
4969 [ # # ]: 0 : ptl = pmd_lockptr(mm, pmd);
4970 : 0 : spin_lock(ptl);
4971 : : /*
4972 : : * make sure that the address range covered by this pmd is not
4973 : : * unmapped from other threads.
4974 : : */
4975 [ # # ]: 0 : if (!pmd_huge(*pmd))
4976 : 0 : goto out;
4977 [ # # ]: 0 : pte = huge_ptep_get((pte_t *)pmd);
4978 [ # # ]: 0 : if (pte_present(pte)) {
4979 [ # # ]: 0 : page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT);
4980 [ # # ]: 0 : if (flags & FOLL_GET)
4981 [ # # ]: 0 : get_page(page);
4982 : : } else {
4983 [ # # ]: 0 : if (is_hugetlb_entry_migration(pte)) {
4984 : 0 : spin_unlock(ptl);
4985 : 0 : __migration_entry_wait(mm, (pte_t *)pmd, ptl);
4986 : 0 : goto retry;
4987 : : }
4988 : : /*
4989 : : * hwpoisoned entry is treated as no_page_table in
4990 : : * follow_page_mask().
4991 : : */
4992 : : }
4993 : 0 : out:
4994 : 0 : spin_unlock(ptl);
4995 : 0 : return page;
4996 : : }
4997 : :
4998 : : struct page * __weak
4999 : 0 : follow_huge_pud(struct mm_struct *mm, unsigned long address,
5000 : : pud_t *pud, int flags)
5001 : : {
5002 [ # # ]: 0 : if (flags & FOLL_GET)
5003 : : return NULL;
5004 : :
5005 [ # # ]: 0 : return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT);
5006 : : }
5007 : :
5008 : : struct page * __weak
5009 : 0 : follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags)
5010 : : {
5011 [ # # ]: 0 : if (flags & FOLL_GET)
5012 : : return NULL;
5013 : :
5014 [ # # ]: 0 : return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT);
5015 : : }
5016 : :
5017 : 0 : bool isolate_huge_page(struct page *page, struct list_head *list)
5018 : : {
5019 : 0 : bool ret = true;
5020 : :
5021 : 0 : VM_BUG_ON_PAGE(!PageHead(page), page);
5022 : 0 : spin_lock(&hugetlb_lock);
5023 [ # # # # ]: 0 : if (!page_huge_active(page) || !get_page_unless_zero(page)) {
5024 : 0 : ret = false;
5025 : 0 : goto unlock;
5026 : : }
5027 : 0 : clear_page_huge_active(page);
5028 : 0 : list_move_tail(&page->lru, list);
5029 : 0 : unlock:
5030 : 0 : spin_unlock(&hugetlb_lock);
5031 : 0 : return ret;
5032 : : }
5033 : :
5034 : 0 : void putback_active_hugepage(struct page *page)
5035 : : {
5036 : 0 : VM_BUG_ON_PAGE(!PageHead(page), page);
5037 : 0 : spin_lock(&hugetlb_lock);
5038 : 0 : set_page_huge_active(page);
5039 : 0 : list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist);
5040 : 0 : spin_unlock(&hugetlb_lock);
5041 : 0 : put_page(page);
5042 : 0 : }
5043 : :
5044 : 0 : void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason)
5045 : : {
5046 : 0 : struct hstate *h = page_hstate(oldpage);
5047 : :
5048 : 0 : hugetlb_cgroup_migrate(oldpage, newpage);
5049 : 0 : set_page_owner_migrate_reason(newpage, reason);
5050 : :
5051 : : /*
5052 : : * transfer temporary state of the new huge page. This is
5053 : : * reverse to other transitions because the newpage is going to
5054 : : * be final while the old one will be freed so it takes over
5055 : : * the temporary status.
5056 : : *
5057 : : * Also note that we have to transfer the per-node surplus state
5058 : : * here as well otherwise the global surplus count will not match
5059 : : * the per-node's.
5060 : : */
5061 [ # # ]: 0 : if (PageHugeTemporary(newpage)) {
5062 : 0 : int old_nid = page_to_nid(oldpage);
5063 : 0 : int new_nid = page_to_nid(newpage);
5064 : :
5065 : 0 : SetPageHugeTemporary(oldpage);
5066 : 0 : ClearPageHugeTemporary(newpage);
5067 : :
5068 : 0 : spin_lock(&hugetlb_lock);
5069 [ # # ]: 0 : if (h->surplus_huge_pages_node[old_nid]) {
5070 : 0 : h->surplus_huge_pages_node[old_nid]--;
5071 : 0 : h->surplus_huge_pages_node[new_nid]++;
5072 : : }
5073 : 0 : spin_unlock(&hugetlb_lock);
5074 : : }
5075 : 0 : }
|