]>
Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * Generic hugetlb support. | |
6d49e352 | 4 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 5 | */ |
1da177e4 LT |
6 | #include <linux/list.h> |
7 | #include <linux/init.h> | |
1da177e4 | 8 | #include <linux/mm.h> |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
97ad1087 | 19 | #include <linux/memblock.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
63489f8e | 22 | #include <linux/mmdebug.h> |
174cd4b1 | 23 | #include <linux/sched/signal.h> |
0fe6e20b | 24 | #include <linux/rmap.h> |
c6247f72 | 25 | #include <linux/string_helpers.h> |
fd6a03ed NH |
26 | #include <linux/swap.h> |
27 | #include <linux/swapops.h> | |
8382d914 | 28 | #include <linux/jhash.h> |
98fa15f3 | 29 | #include <linux/numa.h> |
c77c0a8a | 30 | #include <linux/llist.h> |
d6606683 | 31 | |
63551ae0 DG |
32 | #include <asm/page.h> |
33 | #include <asm/pgtable.h> | |
24669e58 | 34 | #include <asm/tlb.h> |
63551ae0 | 35 | |
24669e58 | 36 | #include <linux/io.h> |
63551ae0 | 37 | #include <linux/hugetlb.h> |
9dd540e2 | 38 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 39 | #include <linux/node.h> |
1a1aad8a | 40 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 41 | #include <linux/page_owner.h> |
7835e98b | 42 | #include "internal.h" |
1da177e4 | 43 | |
c3f38a38 | 44 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
45 | unsigned int default_hstate_idx; |
46 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
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; | |
e5ff2159 | 52 | |
53ba51d2 JT |
53 | __initdata LIST_HEAD(huge_boot_pages); |
54 | ||
e5ff2159 AK |
55 | /* for command line parsing */ |
56 | static struct hstate * __initdata parsed_hstate; | |
57 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 58 | static unsigned long __initdata default_hstate_size; |
9fee021d | 59 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 60 | |
3935baa9 | 61 | /* |
31caf665 NH |
62 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
63 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 64 | */ |
c3f38a38 | 65 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 66 | |
8382d914 DB |
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; | |
c672c7f2 | 72 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 73 | |
7ca02d0a MK |
74 | /* Forward declaration */ |
75 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
76 | ||
90481622 DG |
77 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
78 | { | |
79 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
80 | ||
81 | spin_unlock(&spool->lock); | |
82 | ||
83 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
84 | * remain, give up any reservations mased on minimum size and |
85 | * free the subpool */ | |
86 | if (free) { | |
87 | if (spool->min_hpages != -1) | |
88 | hugetlb_acct_memory(spool->hstate, | |
89 | -spool->min_hpages); | |
90481622 | 90 | kfree(spool); |
7ca02d0a | 91 | } |
90481622 DG |
92 | } |
93 | ||
7ca02d0a MK |
94 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
95 | long min_hpages) | |
90481622 DG |
96 | { |
97 | struct hugepage_subpool *spool; | |
98 | ||
c6a91820 | 99 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
100 | if (!spool) |
101 | return NULL; | |
102 | ||
103 | spin_lock_init(&spool->lock); | |
104 | spool->count = 1; | |
7ca02d0a MK |
105 | spool->max_hpages = max_hpages; |
106 | spool->hstate = h; | |
107 | spool->min_hpages = min_hpages; | |
108 | ||
109 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
110 | kfree(spool); | |
111 | return NULL; | |
112 | } | |
113 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
114 | |
115 | return spool; | |
116 | } | |
117 | ||
118 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
119 | { | |
120 | spin_lock(&spool->lock); | |
121 | BUG_ON(!spool->count); | |
122 | spool->count--; | |
123 | unlock_or_release_subpool(spool); | |
124 | } | |
125 | ||
1c5ecae3 MK |
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 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
135 | long delta) |
136 | { | |
1c5ecae3 | 137 | long ret = delta; |
90481622 DG |
138 | |
139 | if (!spool) | |
1c5ecae3 | 140 | return ret; |
90481622 DG |
141 | |
142 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
143 | |
144 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
145 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
146 | spool->used_hpages += delta; | |
147 | else { | |
148 | ret = -ENOMEM; | |
149 | goto unlock_ret; | |
150 | } | |
90481622 | 151 | } |
90481622 | 152 | |
09a95e29 MK |
153 | /* minimum size accounting */ |
154 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
155 | if (delta > spool->rsv_hpages) { |
156 | /* | |
157 | * Asking for more reserves than those already taken on | |
158 | * behalf of subpool. Return difference. | |
159 | */ | |
160 | ret = delta - spool->rsv_hpages; | |
161 | spool->rsv_hpages = 0; | |
162 | } else { | |
163 | ret = 0; /* reserves already accounted for */ | |
164 | spool->rsv_hpages -= delta; | |
165 | } | |
166 | } | |
167 | ||
168 | unlock_ret: | |
169 | spin_unlock(&spool->lock); | |
90481622 DG |
170 | return ret; |
171 | } | |
172 | ||
1c5ecae3 MK |
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 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
180 | long delta) |
181 | { | |
1c5ecae3 MK |
182 | long ret = delta; |
183 | ||
90481622 | 184 | if (!spool) |
1c5ecae3 | 185 | return delta; |
90481622 DG |
186 | |
187 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
188 | |
189 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
190 | spool->used_hpages -= delta; | |
191 | ||
09a95e29 MK |
192 | /* minimum size accounting */ |
193 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
194 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
195 | ret = 0; | |
196 | else | |
197 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
198 | ||
199 | spool->rsv_hpages += delta; | |
200 | if (spool->rsv_hpages > spool->min_hpages) | |
201 | 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 | */ | |
90481622 | 208 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
209 | |
210 | return ret; | |
90481622 DG |
211 | } |
212 | ||
213 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
214 | { | |
215 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
216 | } | |
217 | ||
218 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
219 | { | |
496ad9aa | 220 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
221 | } |
222 | ||
96822904 AW |
223 | /* |
224 | * Region tracking -- allows tracking of reservations and instantiated pages | |
225 | * across the pages in a mapping. | |
84afd99b | 226 | * |
1dd308a7 MK |
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. | |
96822904 AW |
241 | */ |
242 | struct file_region { | |
243 | struct list_head link; | |
244 | long from; | |
245 | long to; | |
246 | }; | |
247 | ||
d75c6af9 MA |
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 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
253 | bool count_only) | |
254 | { | |
255 | long chg = 0; | |
256 | struct list_head *head = &resv->regions; | |
257 | struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; | |
258 | ||
259 | /* Locate the region we are before or in. */ | |
260 | list_for_each_entry(rg, head, link) | |
261 | if (f <= rg->to) | |
262 | break; | |
263 | ||
264 | /* Round our left edge to the current segment if it encloses us. */ | |
265 | if (f > rg->from) | |
266 | f = rg->from; | |
267 | ||
268 | chg = t - f; | |
269 | ||
270 | /* Check for and consume any regions we now overlap with. */ | |
271 | nrg = rg; | |
272 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
273 | if (&rg->link == head) | |
274 | break; | |
275 | 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 | if (rg->to > t) { | |
283 | chg += rg->to - t; | |
284 | t = rg->to; | |
285 | } | |
286 | chg -= rg->to - rg->from; | |
287 | ||
288 | if (!count_only && rg != nrg) { | |
289 | list_del(&rg->link); | |
290 | kfree(rg); | |
291 | } | |
292 | } | |
293 | ||
294 | if (!count_only) { | |
295 | nrg->from = f; | |
296 | nrg->to = t; | |
297 | } | |
298 | ||
299 | return chg; | |
300 | } | |
301 | ||
1dd308a7 MK |
302 | /* |
303 | * Add the huge page range represented by [f, t) to the reserve | |
5c911954 MA |
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. | |
cf3ad20b MK |
308 | * |
309 | * Return the number of new huge pages added to the map. This | |
310 | * number is greater than or equal to zero. | |
1dd308a7 | 311 | */ |
1406ec9b | 312 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 313 | { |
1406ec9b | 314 | struct list_head *head = &resv->regions; |
d75c6af9 | 315 | struct file_region *rg, *nrg; |
cf3ad20b | 316 | long add = 0; |
96822904 | 317 | |
7b24d861 | 318 | spin_lock(&resv->lock); |
96822904 AW |
319 | /* Locate the region we are either in or before. */ |
320 | list_for_each_entry(rg, head, link) | |
321 | if (f <= rg->to) | |
322 | break; | |
323 | ||
5e911373 MK |
324 | /* |
325 | * If no region exists which can be expanded to include the | |
5c911954 MA |
326 | * specified range, pull a region descriptor from the cache |
327 | * and use it for this range. | |
5e911373 MK |
328 | */ |
329 | if (&rg->link == head || t < rg->from) { | |
330 | VM_BUG_ON(resv->region_cache_count <= 0); | |
331 | ||
332 | resv->region_cache_count--; | |
333 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
334 | link); | |
335 | list_del(&nrg->link); | |
336 | ||
337 | nrg->from = f; | |
338 | nrg->to = t; | |
339 | list_add(&nrg->link, rg->link.prev); | |
340 | ||
341 | add += t - f; | |
342 | goto out_locked; | |
343 | } | |
344 | ||
d75c6af9 | 345 | add = add_reservation_in_range(resv, f, t, false); |
cf3ad20b | 346 | |
5e911373 MK |
347 | out_locked: |
348 | resv->adds_in_progress--; | |
7b24d861 | 349 | spin_unlock(&resv->lock); |
cf3ad20b MK |
350 | VM_BUG_ON(add < 0); |
351 | return add; | |
96822904 AW |
352 | } |
353 | ||
1dd308a7 MK |
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 | |
5c911954 MA |
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. | |
5e911373 MK |
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. | |
1dd308a7 | 369 | */ |
1406ec9b | 370 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 371 | { |
96822904 AW |
372 | long chg = 0; |
373 | ||
7b24d861 | 374 | spin_lock(&resv->lock); |
5e911373 MK |
375 | retry_locked: |
376 | 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 | if (resv->adds_in_progress > resv->region_cache_count) { | |
383 | struct file_region *trg; | |
384 | ||
385 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
386 | /* Must drop lock to allocate a new descriptor. */ | |
387 | resv->adds_in_progress--; | |
388 | spin_unlock(&resv->lock); | |
389 | ||
390 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
5c911954 | 391 | if (!trg) |
5e911373 MK |
392 | return -ENOMEM; |
393 | ||
394 | spin_lock(&resv->lock); | |
395 | list_add(&trg->link, &resv->region_cache); | |
396 | resv->region_cache_count++; | |
397 | goto retry_locked; | |
398 | } | |
399 | ||
d75c6af9 | 400 | chg = add_reservation_in_range(resv, f, t, true); |
7b24d861 | 401 | |
7b24d861 | 402 | spin_unlock(&resv->lock); |
96822904 AW |
403 | return chg; |
404 | } | |
405 | ||
5e911373 MK |
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 | static void region_abort(struct resv_map *resv, long f, long t) | |
418 | { | |
419 | spin_lock(&resv->lock); | |
420 | VM_BUG_ON(!resv->region_cache_count); | |
421 | resv->adds_in_progress--; | |
422 | spin_unlock(&resv->lock); | |
423 | } | |
424 | ||
1dd308a7 | 425 | /* |
feba16e2 MK |
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. | |
1dd308a7 | 438 | */ |
feba16e2 | 439 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 440 | { |
1406ec9b | 441 | struct list_head *head = &resv->regions; |
96822904 | 442 | struct file_region *rg, *trg; |
feba16e2 MK |
443 | struct file_region *nrg = NULL; |
444 | long del = 0; | |
96822904 | 445 | |
feba16e2 | 446 | retry: |
7b24d861 | 447 | spin_lock(&resv->lock); |
feba16e2 | 448 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
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 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 457 | continue; |
dbe409e4 | 458 | |
feba16e2 | 459 | if (rg->from >= t) |
96822904 | 460 | break; |
96822904 | 461 | |
feba16e2 MK |
462 | 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 | if (!nrg && | |
468 | resv->region_cache_count > resv->adds_in_progress) { | |
469 | nrg = list_first_entry(&resv->region_cache, | |
470 | struct file_region, | |
471 | link); | |
472 | list_del(&nrg->link); | |
473 | resv->region_cache_count--; | |
474 | } | |
96822904 | 475 | |
feba16e2 MK |
476 | if (!nrg) { |
477 | spin_unlock(&resv->lock); | |
478 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
479 | if (!nrg) | |
480 | return -ENOMEM; | |
481 | goto retry; | |
482 | } | |
483 | ||
484 | del += t - f; | |
485 | ||
486 | /* New entry for end of split region */ | |
487 | nrg->from = t; | |
488 | nrg->to = rg->to; | |
489 | INIT_LIST_HEAD(&nrg->link); | |
490 | ||
491 | /* Original entry is trimmed */ | |
492 | rg->to = f; | |
493 | ||
494 | list_add(&nrg->link, &rg->link); | |
495 | nrg = NULL; | |
96822904 | 496 | break; |
feba16e2 MK |
497 | } |
498 | ||
499 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
500 | del += rg->to - rg->from; | |
501 | list_del(&rg->link); | |
502 | kfree(rg); | |
503 | continue; | |
504 | } | |
505 | ||
506 | if (f <= rg->from) { /* Trim beginning of region */ | |
507 | del += t - rg->from; | |
508 | rg->from = t; | |
509 | } else { /* Trim end of region */ | |
510 | del += rg->to - f; | |
511 | rg->to = f; | |
512 | } | |
96822904 | 513 | } |
7b24d861 | 514 | |
7b24d861 | 515 | spin_unlock(&resv->lock); |
feba16e2 MK |
516 | kfree(nrg); |
517 | return del; | |
96822904 AW |
518 | } |
519 | ||
b5cec28d MK |
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 | */ | |
72e2936c | 529 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
530 | { |
531 | struct hugepage_subpool *spool = subpool_inode(inode); | |
532 | long rsv_adjust; | |
533 | ||
534 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 535 | if (rsv_adjust) { |
b5cec28d MK |
536 | struct hstate *h = hstate_inode(inode); |
537 | ||
538 | hugetlb_acct_memory(h, 1); | |
539 | } | |
540 | } | |
541 | ||
1dd308a7 MK |
542 | /* |
543 | * Count and return the number of huge pages in the reserve map | |
544 | * that intersect with the range [f, t). | |
545 | */ | |
1406ec9b | 546 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 547 | { |
1406ec9b | 548 | struct list_head *head = &resv->regions; |
84afd99b AW |
549 | struct file_region *rg; |
550 | long chg = 0; | |
551 | ||
7b24d861 | 552 | spin_lock(&resv->lock); |
84afd99b AW |
553 | /* Locate each segment we overlap with, and count that overlap. */ |
554 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
555 | long seg_from; |
556 | long seg_to; | |
84afd99b AW |
557 | |
558 | if (rg->to <= f) | |
559 | continue; | |
560 | if (rg->from >= t) | |
561 | break; | |
562 | ||
563 | seg_from = max(rg->from, f); | |
564 | seg_to = min(rg->to, t); | |
565 | ||
566 | chg += seg_to - seg_from; | |
567 | } | |
7b24d861 | 568 | spin_unlock(&resv->lock); |
84afd99b AW |
569 | |
570 | return chg; | |
571 | } | |
572 | ||
e7c4b0bf AW |
573 | /* |
574 | * Convert the address within this vma to the page offset within | |
575 | * the mapping, in pagecache page units; huge pages here. | |
576 | */ | |
a5516438 AK |
577 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
578 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 579 | { |
a5516438 AK |
580 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
581 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
582 | } |
583 | ||
0fe6e20b NH |
584 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
585 | unsigned long address) | |
586 | { | |
587 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
588 | } | |
dee41079 | 589 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 590 | |
08fba699 MG |
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 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
596 | { | |
05ea8860 DW |
597 | if (vma->vm_ops && vma->vm_ops->pagesize) |
598 | return vma->vm_ops->pagesize(vma); | |
599 | return PAGE_SIZE; | |
08fba699 | 600 | } |
f340ca0f | 601 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 602 | |
3340289d MG |
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 | |
09135cc5 DW |
606 | * architectures where it differs, an architecture-specific 'strong' |
607 | * version of this symbol is required. | |
3340289d | 608 | */ |
09135cc5 | 609 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
610 | { |
611 | return vma_kernel_pagesize(vma); | |
612 | } | |
3340289d | 613 | |
84afd99b AW |
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) | |
04f2cbe3 | 621 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 622 | |
a1e78772 MG |
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. | |
84afd99b AW |
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. | |
a1e78772 | 641 | */ |
e7c4b0bf AW |
642 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
643 | { | |
644 | return (unsigned long)vma->vm_private_data; | |
645 | } | |
646 | ||
647 | static void set_vma_private_data(struct vm_area_struct *vma, | |
648 | unsigned long value) | |
649 | { | |
650 | vma->vm_private_data = (void *)value; | |
651 | } | |
652 | ||
9119a41e | 653 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
654 | { |
655 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
656 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
657 | ||
658 | if (!resv_map || !rg) { | |
659 | kfree(resv_map); | |
660 | kfree(rg); | |
84afd99b | 661 | return NULL; |
5e911373 | 662 | } |
84afd99b AW |
663 | |
664 | kref_init(&resv_map->refs); | |
7b24d861 | 665 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
666 | INIT_LIST_HEAD(&resv_map->regions); |
667 | ||
5e911373 MK |
668 | resv_map->adds_in_progress = 0; |
669 | ||
670 | INIT_LIST_HEAD(&resv_map->region_cache); | |
671 | list_add(&rg->link, &resv_map->region_cache); | |
672 | resv_map->region_cache_count = 1; | |
673 | ||
84afd99b AW |
674 | return resv_map; |
675 | } | |
676 | ||
9119a41e | 677 | void resv_map_release(struct kref *ref) |
84afd99b AW |
678 | { |
679 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
680 | struct list_head *head = &resv_map->region_cache; |
681 | struct file_region *rg, *trg; | |
84afd99b AW |
682 | |
683 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 684 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
685 | |
686 | /* ... and any entries left in the cache */ | |
687 | list_for_each_entry_safe(rg, trg, head, link) { | |
688 | list_del(&rg->link); | |
689 | kfree(rg); | |
690 | } | |
691 | ||
692 | VM_BUG_ON(resv_map->adds_in_progress); | |
693 | ||
84afd99b AW |
694 | kfree(resv_map); |
695 | } | |
696 | ||
4e35f483 JK |
697 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
698 | { | |
f27a5136 MK |
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 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
708 | } |
709 | ||
84afd99b | 710 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 711 | { |
81d1b09c | 712 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
713 | if (vma->vm_flags & VM_MAYSHARE) { |
714 | struct address_space *mapping = vma->vm_file->f_mapping; | |
715 | struct inode *inode = mapping->host; | |
716 | ||
717 | return inode_resv_map(inode); | |
718 | ||
719 | } else { | |
84afd99b AW |
720 | return (struct resv_map *)(get_vma_private_data(vma) & |
721 | ~HPAGE_RESV_MASK); | |
4e35f483 | 722 | } |
a1e78772 MG |
723 | } |
724 | ||
84afd99b | 725 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 726 | { |
81d1b09c SL |
727 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
728 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 729 | |
84afd99b AW |
730 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
731 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
732 | } |
733 | ||
734 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
735 | { | |
81d1b09c SL |
736 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
737 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
738 | |
739 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
740 | } |
741 | ||
742 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
743 | { | |
81d1b09c | 744 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
745 | |
746 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
747 | } |
748 | ||
04f2cbe3 | 749 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
750 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
751 | { | |
81d1b09c | 752 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 753 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
754 | vma->vm_private_data = (void *)0; |
755 | } | |
756 | ||
757 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 758 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 759 | { |
af0ed73e JK |
760 | 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 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 771 | return true; |
af0ed73e | 772 | else |
559ec2f8 | 773 | return false; |
af0ed73e | 774 | } |
a63884e9 JK |
775 | |
776 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
777 | 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 | if (chg) | |
786 | return false; | |
787 | else | |
788 | return true; | |
789 | } | |
a63884e9 JK |
790 | |
791 | /* | |
792 | * Only the process that called mmap() has reserves for | |
793 | * private mappings. | |
794 | */ | |
67961f9d MK |
795 | 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 | if (chg) | |
812 | return false; | |
813 | else | |
814 | return true; | |
815 | } | |
a63884e9 | 816 | |
559ec2f8 | 817 | return false; |
a1e78772 MG |
818 | } |
819 | ||
a5516438 | 820 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
821 | { |
822 | int nid = page_to_nid(page); | |
0edaecfa | 823 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
824 | h->free_huge_pages++; |
825 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
826 | } |
827 | ||
94310cbc | 828 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
829 | { |
830 | struct page *page; | |
831 | ||
c8721bbb | 832 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 833 | if (!PageHWPoison(page)) |
c8721bbb NH |
834 | break; |
835 | /* | |
836 | * if 'non-isolated free hugepage' not found on the list, | |
837 | * the allocation fails. | |
838 | */ | |
839 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 840 | return NULL; |
0edaecfa | 841 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 842 | set_page_refcounted(page); |
bf50bab2 NH |
843 | h->free_huge_pages--; |
844 | h->free_huge_pages_node[nid]--; | |
845 | return page; | |
846 | } | |
847 | ||
3e59fcb0 MH |
848 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
849 | nodemask_t *nmask) | |
94310cbc | 850 | { |
3e59fcb0 MH |
851 | unsigned int cpuset_mems_cookie; |
852 | struct zonelist *zonelist; | |
853 | struct zone *zone; | |
854 | struct zoneref *z; | |
98fa15f3 | 855 | int node = NUMA_NO_NODE; |
94310cbc | 856 | |
3e59fcb0 MH |
857 | zonelist = node_zonelist(nid, gfp_mask); |
858 | ||
859 | retry_cpuset: | |
860 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
861 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
862 | struct page *page; | |
863 | ||
864 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
865 | continue; | |
866 | /* | |
867 | * no need to ask again on the same node. Pool is node rather than | |
868 | * zone aware | |
869 | */ | |
870 | if (zone_to_nid(zone) == node) | |
871 | continue; | |
872 | node = zone_to_nid(zone); | |
94310cbc | 873 | |
94310cbc AK |
874 | page = dequeue_huge_page_node_exact(h, node); |
875 | if (page) | |
876 | return page; | |
877 | } | |
3e59fcb0 MH |
878 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
879 | goto retry_cpuset; | |
880 | ||
94310cbc AK |
881 | return NULL; |
882 | } | |
883 | ||
86cdb465 NH |
884 | /* Movability of hugepages depends on migration support. */ |
885 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
886 | { | |
7ed2c31d | 887 | if (hugepage_movable_supported(h)) |
86cdb465 NH |
888 | return GFP_HIGHUSER_MOVABLE; |
889 | else | |
890 | return GFP_HIGHUSER; | |
891 | } | |
892 | ||
a5516438 AK |
893 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
894 | struct vm_area_struct *vma, | |
af0ed73e JK |
895 | unsigned long address, int avoid_reserve, |
896 | long chg) | |
1da177e4 | 897 | { |
3e59fcb0 | 898 | struct page *page; |
480eccf9 | 899 | struct mempolicy *mpol; |
04ec6264 | 900 | gfp_t gfp_mask; |
3e59fcb0 | 901 | nodemask_t *nodemask; |
04ec6264 | 902 | int nid; |
1da177e4 | 903 | |
a1e78772 MG |
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 | */ | |
af0ed73e | 909 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 910 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 911 | goto err; |
a1e78772 | 912 | |
04f2cbe3 | 913 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 914 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 915 | goto err; |
04f2cbe3 | 916 | |
04ec6264 VB |
917 | gfp_mask = htlb_alloc_mask(h); |
918 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
919 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
920 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
921 | SetPagePrivate(page); | |
922 | h->resv_huge_pages--; | |
1da177e4 | 923 | } |
cc9a6c87 | 924 | |
52cd3b07 | 925 | mpol_cond_put(mpol); |
1da177e4 | 926 | return page; |
cc9a6c87 MG |
927 | |
928 | err: | |
cc9a6c87 | 929 | return NULL; |
1da177e4 LT |
930 | } |
931 | ||
1cac6f2c LC |
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 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
940 | { | |
0edaf86c | 941 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
942 | VM_BUG_ON(nid >= MAX_NUMNODES); |
943 | ||
944 | return nid; | |
945 | } | |
946 | ||
947 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
948 | { | |
949 | if (!node_isset(nid, *nodes_allowed)) | |
950 | nid = next_node_allowed(nid, nodes_allowed); | |
951 | 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 | ||
e1073d1e | 1003 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1004 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1005 | unsigned int order) |
944d9fec LC |
1006 | { |
1007 | int i; | |
1008 | int nr_pages = 1 << order; | |
1009 | struct page *p = page + 1; | |
1010 | ||
c8cc708a | 1011 | atomic_set(compound_mapcount_ptr(page), 0); |
47e29d32 JH |
1012 | if (hpage_pincount_available(page)) |
1013 | atomic_set(compound_pincount_ptr(page), 0); | |
1014 | ||
944d9fec | 1015 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1016 | clear_compound_head(p); |
944d9fec | 1017 | set_page_refcounted(p); |
944d9fec LC |
1018 | } |
1019 | ||
1020 | set_compound_order(page, 0); | |
1021 | __ClearPageHead(page); | |
1022 | } | |
1023 | ||
d00181b9 | 1024 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1025 | { |
1026 | free_contig_range(page_to_pfn(page), 1 << order); | |
1027 | } | |
1028 | ||
4eb0716e | 1029 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1030 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1031 | int nid, nodemask_t *nodemask) | |
944d9fec | 1032 | { |
5e27a2df | 1033 | unsigned long nr_pages = 1UL << huge_page_order(h); |
944d9fec | 1034 | |
5e27a2df | 1035 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1036 | } |
1037 | ||
1038 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1039 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1040 | #else /* !CONFIG_CONTIG_ALLOC */ |
1041 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1042 | int nid, nodemask_t *nodemask) | |
1043 | { | |
1044 | return NULL; | |
1045 | } | |
1046 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1047 | |
e1073d1e | 1048 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1049 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1050 | int nid, nodemask_t *nodemask) |
1051 | { | |
1052 | return NULL; | |
1053 | } | |
d00181b9 | 1054 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1055 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1056 | unsigned int order) { } |
944d9fec LC |
1057 | #endif |
1058 | ||
a5516438 | 1059 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1060 | { |
1061 | int i; | |
a5516438 | 1062 | |
4eb0716e | 1063 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1064 | return; |
18229df5 | 1065 | |
a5516438 AK |
1066 | h->nr_huge_pages--; |
1067 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1068 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1069 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1070 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1071 | 1 << PG_active | 1 << PG_private | |
1072 | 1 << PG_writeback); | |
6af2acb6 | 1073 | } |
309381fe | 1074 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1075 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1076 | set_page_refcounted(page); |
944d9fec LC |
1077 | if (hstate_is_gigantic(h)) { |
1078 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1079 | free_gigantic_page(page, huge_page_order(h)); | |
1080 | } else { | |
944d9fec LC |
1081 | __free_pages(page, huge_page_order(h)); |
1082 | } | |
6af2acb6 AL |
1083 | } |
1084 | ||
e5ff2159 AK |
1085 | struct hstate *size_to_hstate(unsigned long size) |
1086 | { | |
1087 | struct hstate *h; | |
1088 | ||
1089 | for_each_hstate(h) { | |
1090 | if (huge_page_size(h) == size) | |
1091 | return h; | |
1092 | } | |
1093 | return NULL; | |
1094 | } | |
1095 | ||
bcc54222 NH |
1096 | /* |
1097 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1098 | * to hstate->hugepage_activelist.) | |
1099 | * | |
1100 | * This function can be called for tail pages, but never returns true for them. | |
1101 | */ | |
1102 | bool page_huge_active(struct page *page) | |
1103 | { | |
1104 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1105 | return PageHead(page) && PagePrivate(&page[1]); | |
1106 | } | |
1107 | ||
1108 | /* never called for tail page */ | |
1109 | static void set_page_huge_active(struct page *page) | |
1110 | { | |
1111 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1112 | SetPagePrivate(&page[1]); | |
1113 | } | |
1114 | ||
1115 | static void clear_page_huge_active(struct page *page) | |
1116 | { | |
1117 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1118 | ClearPagePrivate(&page[1]); | |
1119 | } | |
1120 | ||
ab5ac90a MH |
1121 | /* |
1122 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1123 | * code | |
1124 | */ | |
1125 | static inline bool PageHugeTemporary(struct page *page) | |
1126 | { | |
1127 | if (!PageHuge(page)) | |
1128 | return false; | |
1129 | ||
1130 | return (unsigned long)page[2].mapping == -1U; | |
1131 | } | |
1132 | ||
1133 | static inline void SetPageHugeTemporary(struct page *page) | |
1134 | { | |
1135 | page[2].mapping = (void *)-1U; | |
1136 | } | |
1137 | ||
1138 | static inline void ClearPageHugeTemporary(struct page *page) | |
1139 | { | |
1140 | page[2].mapping = NULL; | |
1141 | } | |
1142 | ||
c77c0a8a | 1143 | static void __free_huge_page(struct page *page) |
27a85ef1 | 1144 | { |
a5516438 AK |
1145 | /* |
1146 | * Can't pass hstate in here because it is called from the | |
1147 | * compound page destructor. | |
1148 | */ | |
e5ff2159 | 1149 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1150 | int nid = page_to_nid(page); |
90481622 DG |
1151 | struct hugepage_subpool *spool = |
1152 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1153 | bool restore_reserve; |
27a85ef1 | 1154 | |
b4330afb MK |
1155 | VM_BUG_ON_PAGE(page_count(page), page); |
1156 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc YW |
1157 | |
1158 | set_page_private(page, 0); | |
1159 | page->mapping = NULL; | |
07443a85 | 1160 | restore_reserve = PagePrivate(page); |
16c794b4 | 1161 | ClearPagePrivate(page); |
27a85ef1 | 1162 | |
1c5ecae3 | 1163 | /* |
0919e1b6 MK |
1164 | * If PagePrivate() was set on page, page allocation consumed a |
1165 | * reservation. If the page was associated with a subpool, there | |
1166 | * would have been a page reserved in the subpool before allocation | |
1167 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1168 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1169 | * remove the reserved page from the subpool. | |
1c5ecae3 | 1170 | */ |
0919e1b6 MK |
1171 | if (!restore_reserve) { |
1172 | /* | |
1173 | * A return code of zero implies that the subpool will be | |
1174 | * under its minimum size if the reservation is not restored | |
1175 | * after page is free. Therefore, force restore_reserve | |
1176 | * operation. | |
1177 | */ | |
1178 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1179 | restore_reserve = true; | |
1180 | } | |
1c5ecae3 | 1181 | |
27a85ef1 | 1182 | spin_lock(&hugetlb_lock); |
bcc54222 | 1183 | clear_page_huge_active(page); |
6d76dcf4 AK |
1184 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1185 | pages_per_huge_page(h), page); | |
07443a85 JK |
1186 | if (restore_reserve) |
1187 | h->resv_huge_pages++; | |
1188 | ||
ab5ac90a MH |
1189 | if (PageHugeTemporary(page)) { |
1190 | list_del(&page->lru); | |
1191 | ClearPageHugeTemporary(page); | |
1192 | update_and_free_page(h, page); | |
1193 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1194 | /* remove the page from active list */ |
1195 | list_del(&page->lru); | |
a5516438 AK |
1196 | update_and_free_page(h, page); |
1197 | h->surplus_huge_pages--; | |
1198 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1199 | } else { |
5d3a551c | 1200 | arch_clear_hugepage_flags(page); |
a5516438 | 1201 | enqueue_huge_page(h, page); |
7893d1d5 | 1202 | } |
27a85ef1 DG |
1203 | spin_unlock(&hugetlb_lock); |
1204 | } | |
1205 | ||
c77c0a8a WL |
1206 | /* |
1207 | * As free_huge_page() can be called from a non-task context, we have | |
1208 | * to defer the actual freeing in a workqueue to prevent potential | |
1209 | * hugetlb_lock deadlock. | |
1210 | * | |
1211 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1212 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1213 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1214 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1215 | */ | |
1216 | static LLIST_HEAD(hpage_freelist); | |
1217 | ||
1218 | static void free_hpage_workfn(struct work_struct *work) | |
1219 | { | |
1220 | struct llist_node *node; | |
1221 | struct page *page; | |
1222 | ||
1223 | node = llist_del_all(&hpage_freelist); | |
1224 | ||
1225 | while (node) { | |
1226 | page = container_of((struct address_space **)node, | |
1227 | struct page, mapping); | |
1228 | node = node->next; | |
1229 | __free_huge_page(page); | |
1230 | } | |
1231 | } | |
1232 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1233 | ||
1234 | void free_huge_page(struct page *page) | |
1235 | { | |
1236 | /* | |
1237 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1238 | */ | |
1239 | if (!in_task()) { | |
1240 | /* | |
1241 | * Only call schedule_work() if hpage_freelist is previously | |
1242 | * empty. Otherwise, schedule_work() had been called but the | |
1243 | * workfn hasn't retrieved the list yet. | |
1244 | */ | |
1245 | if (llist_add((struct llist_node *)&page->mapping, | |
1246 | &hpage_freelist)) | |
1247 | schedule_work(&free_hpage_work); | |
1248 | return; | |
1249 | } | |
1250 | ||
1251 | __free_huge_page(page); | |
1252 | } | |
1253 | ||
a5516438 | 1254 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1255 | { |
0edaecfa | 1256 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1257 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1258 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1259 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1260 | h->nr_huge_pages++; |
1261 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 | 1262 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1263 | } |
1264 | ||
d00181b9 | 1265 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1266 | { |
1267 | int i; | |
1268 | int nr_pages = 1 << order; | |
1269 | struct page *p = page + 1; | |
1270 | ||
1271 | /* we rely on prep_new_huge_page to set the destructor */ | |
1272 | set_compound_order(page, order); | |
ef5a22be | 1273 | __ClearPageReserved(page); |
de09d31d | 1274 | __SetPageHead(page); |
20a0307c | 1275 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1276 | /* |
1277 | * For gigantic hugepages allocated through bootmem at | |
1278 | * boot, it's safer to be consistent with the not-gigantic | |
1279 | * hugepages and clear the PG_reserved bit from all tail pages | |
1280 | * too. Otherwse drivers using get_user_pages() to access tail | |
1281 | * pages may get the reference counting wrong if they see | |
1282 | * PG_reserved set on a tail page (despite the head page not | |
1283 | * having PG_reserved set). Enforcing this consistency between | |
1284 | * head and tail pages allows drivers to optimize away a check | |
1285 | * on the head page when they need know if put_page() is needed | |
1286 | * after get_user_pages(). | |
1287 | */ | |
1288 | __ClearPageReserved(p); | |
58a84aa9 | 1289 | set_page_count(p, 0); |
1d798ca3 | 1290 | set_compound_head(p, page); |
20a0307c | 1291 | } |
b4330afb | 1292 | atomic_set(compound_mapcount_ptr(page), -1); |
47e29d32 JH |
1293 | |
1294 | if (hpage_pincount_available(page)) | |
1295 | atomic_set(compound_pincount_ptr(page), 0); | |
20a0307c WF |
1296 | } |
1297 | ||
7795912c AM |
1298 | /* |
1299 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1300 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1301 | * details. | |
1302 | */ | |
20a0307c WF |
1303 | int PageHuge(struct page *page) |
1304 | { | |
20a0307c WF |
1305 | if (!PageCompound(page)) |
1306 | return 0; | |
1307 | ||
1308 | page = compound_head(page); | |
f1e61557 | 1309 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1310 | } |
43131e14 NH |
1311 | EXPORT_SYMBOL_GPL(PageHuge); |
1312 | ||
27c73ae7 AA |
1313 | /* |
1314 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1315 | * normal or transparent huge pages. | |
1316 | */ | |
1317 | int PageHeadHuge(struct page *page_head) | |
1318 | { | |
27c73ae7 AA |
1319 | if (!PageHead(page_head)) |
1320 | return 0; | |
1321 | ||
758f66a2 | 1322 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1323 | } |
27c73ae7 | 1324 | |
c0d0381a MK |
1325 | /* |
1326 | * Find address_space associated with hugetlbfs page. | |
1327 | * Upon entry page is locked and page 'was' mapped although mapped state | |
1328 | * could change. If necessary, use anon_vma to find vma and associated | |
1329 | * address space. The returned mapping may be stale, but it can not be | |
1330 | * invalid as page lock (which is held) is required to destroy mapping. | |
1331 | */ | |
1332 | static struct address_space *_get_hugetlb_page_mapping(struct page *hpage) | |
1333 | { | |
1334 | struct anon_vma *anon_vma; | |
1335 | pgoff_t pgoff_start, pgoff_end; | |
1336 | struct anon_vma_chain *avc; | |
1337 | struct address_space *mapping = page_mapping(hpage); | |
1338 | ||
1339 | /* Simple file based mapping */ | |
1340 | if (mapping) | |
1341 | return mapping; | |
1342 | ||
1343 | /* | |
1344 | * Even anonymous hugetlbfs mappings are associated with an | |
1345 | * underlying hugetlbfs file (see hugetlb_file_setup in mmap | |
1346 | * code). Find a vma associated with the anonymous vma, and | |
1347 | * use the file pointer to get address_space. | |
1348 | */ | |
1349 | anon_vma = page_lock_anon_vma_read(hpage); | |
1350 | if (!anon_vma) | |
1351 | return mapping; /* NULL */ | |
1352 | ||
1353 | /* Use first found vma */ | |
1354 | pgoff_start = page_to_pgoff(hpage); | |
1355 | pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1; | |
1356 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, | |
1357 | pgoff_start, pgoff_end) { | |
1358 | struct vm_area_struct *vma = avc->vma; | |
1359 | ||
1360 | mapping = vma->vm_file->f_mapping; | |
1361 | break; | |
1362 | } | |
1363 | ||
1364 | anon_vma_unlock_read(anon_vma); | |
1365 | return mapping; | |
1366 | } | |
1367 | ||
1368 | /* | |
1369 | * Find and lock address space (mapping) in write mode. | |
1370 | * | |
1371 | * Upon entry, the page is locked which allows us to find the mapping | |
1372 | * even in the case of an anon page. However, locking order dictates | |
1373 | * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs | |
1374 | * specific. So, we first try to lock the sema while still holding the | |
1375 | * page lock. If this works, great! If not, then we need to drop the | |
1376 | * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of | |
1377 | * course, need to revalidate state along the way. | |
1378 | */ | |
1379 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1380 | { | |
1381 | struct address_space *mapping, *mapping2; | |
1382 | ||
1383 | mapping = _get_hugetlb_page_mapping(hpage); | |
1384 | retry: | |
1385 | if (!mapping) | |
1386 | return mapping; | |
1387 | ||
1388 | /* | |
1389 | * If no contention, take lock and return | |
1390 | */ | |
1391 | if (i_mmap_trylock_write(mapping)) | |
1392 | return mapping; | |
1393 | ||
1394 | /* | |
1395 | * Must drop page lock and wait on mapping sema. | |
1396 | * Note: Once page lock is dropped, mapping could become invalid. | |
1397 | * As a hack, increase map count until we lock page again. | |
1398 | */ | |
1399 | atomic_inc(&hpage->_mapcount); | |
1400 | unlock_page(hpage); | |
1401 | i_mmap_lock_write(mapping); | |
1402 | lock_page(hpage); | |
1403 | atomic_add_negative(-1, &hpage->_mapcount); | |
1404 | ||
1405 | /* verify page is still mapped */ | |
1406 | if (!page_mapped(hpage)) { | |
1407 | i_mmap_unlock_write(mapping); | |
1408 | return NULL; | |
1409 | } | |
1410 | ||
1411 | /* | |
1412 | * Get address space again and verify it is the same one | |
1413 | * we locked. If not, drop lock and retry. | |
1414 | */ | |
1415 | mapping2 = _get_hugetlb_page_mapping(hpage); | |
1416 | if (mapping2 != mapping) { | |
1417 | i_mmap_unlock_write(mapping); | |
1418 | mapping = mapping2; | |
1419 | goto retry; | |
1420 | } | |
1421 | ||
1422 | return mapping; | |
1423 | } | |
1424 | ||
13d60f4b ZY |
1425 | pgoff_t __basepage_index(struct page *page) |
1426 | { | |
1427 | struct page *page_head = compound_head(page); | |
1428 | pgoff_t index = page_index(page_head); | |
1429 | unsigned long compound_idx; | |
1430 | ||
1431 | if (!PageHuge(page_head)) | |
1432 | return page_index(page); | |
1433 | ||
1434 | if (compound_order(page_head) >= MAX_ORDER) | |
1435 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1436 | else | |
1437 | compound_idx = page - page_head; | |
1438 | ||
1439 | return (index << compound_order(page_head)) + compound_idx; | |
1440 | } | |
1441 | ||
0c397dae | 1442 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1443 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1444 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1445 | { |
af0fb9df | 1446 | int order = huge_page_order(h); |
1da177e4 | 1447 | struct page *page; |
f60858f9 | 1448 | bool alloc_try_hard = true; |
f96efd58 | 1449 | |
f60858f9 MK |
1450 | /* |
1451 | * By default we always try hard to allocate the page with | |
1452 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1453 | * a loop (to adjust global huge page counts) and previous allocation | |
1454 | * failed, do not continue to try hard on the same node. Use the | |
1455 | * node_alloc_noretry bitmap to manage this state information. | |
1456 | */ | |
1457 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1458 | alloc_try_hard = false; | |
1459 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1460 | if (alloc_try_hard) | |
1461 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1462 | if (nid == NUMA_NO_NODE) |
1463 | nid = numa_mem_id(); | |
1464 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1465 | if (page) | |
1466 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1467 | else | |
1468 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1469 | |
f60858f9 MK |
1470 | /* |
1471 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1472 | * indicates an overall state change. Clear bit so that we resume | |
1473 | * normal 'try hard' allocations. | |
1474 | */ | |
1475 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1476 | node_clear(nid, *node_alloc_noretry); | |
1477 | ||
1478 | /* | |
1479 | * If we tried hard to get a page but failed, set bit so that | |
1480 | * subsequent attempts will not try as hard until there is an | |
1481 | * overall state change. | |
1482 | */ | |
1483 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1484 | node_set(nid, *node_alloc_noretry); | |
1485 | ||
63b4613c NA |
1486 | return page; |
1487 | } | |
1488 | ||
0c397dae MH |
1489 | /* |
1490 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1491 | * should use this function to get new hugetlb pages | |
1492 | */ | |
1493 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1494 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1495 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1496 | { |
1497 | struct page *page; | |
1498 | ||
1499 | if (hstate_is_gigantic(h)) | |
1500 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1501 | else | |
1502 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1503 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1504 | if (!page) |
1505 | return NULL; | |
1506 | ||
1507 | if (hstate_is_gigantic(h)) | |
1508 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1509 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1510 | ||
1511 | return page; | |
1512 | } | |
1513 | ||
af0fb9df MH |
1514 | /* |
1515 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1516 | * manner. | |
1517 | */ | |
f60858f9 MK |
1518 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1519 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1520 | { |
1521 | struct page *page; | |
1522 | int nr_nodes, node; | |
af0fb9df | 1523 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1524 | |
1525 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1526 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1527 | node_alloc_noretry); | |
af0fb9df | 1528 | if (page) |
b2261026 | 1529 | break; |
b2261026 JK |
1530 | } |
1531 | ||
af0fb9df MH |
1532 | if (!page) |
1533 | return 0; | |
b2261026 | 1534 | |
af0fb9df MH |
1535 | put_page(page); /* free it into the hugepage allocator */ |
1536 | ||
1537 | return 1; | |
b2261026 JK |
1538 | } |
1539 | ||
e8c5c824 LS |
1540 | /* |
1541 | * Free huge page from pool from next node to free. | |
1542 | * Attempt to keep persistent huge pages more or less | |
1543 | * balanced over allowed nodes. | |
1544 | * Called with hugetlb_lock locked. | |
1545 | */ | |
6ae11b27 LS |
1546 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1547 | bool acct_surplus) | |
e8c5c824 | 1548 | { |
b2261026 | 1549 | int nr_nodes, node; |
e8c5c824 LS |
1550 | int ret = 0; |
1551 | ||
b2261026 | 1552 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1553 | /* |
1554 | * If we're returning unused surplus pages, only examine | |
1555 | * nodes with surplus pages. | |
1556 | */ | |
b2261026 JK |
1557 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1558 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1559 | struct page *page = |
b2261026 | 1560 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1561 | struct page, lru); |
1562 | list_del(&page->lru); | |
1563 | h->free_huge_pages--; | |
b2261026 | 1564 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1565 | if (acct_surplus) { |
1566 | h->surplus_huge_pages--; | |
b2261026 | 1567 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1568 | } |
e8c5c824 LS |
1569 | update_and_free_page(h, page); |
1570 | ret = 1; | |
9a76db09 | 1571 | break; |
e8c5c824 | 1572 | } |
b2261026 | 1573 | } |
e8c5c824 LS |
1574 | |
1575 | return ret; | |
1576 | } | |
1577 | ||
c8721bbb NH |
1578 | /* |
1579 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1580 | * nothing for in-use hugepages and non-hugepages. |
1581 | * This function returns values like below: | |
1582 | * | |
1583 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1584 | * (allocated or reserved.) | |
1585 | * 0: successfully dissolved free hugepages or the page is not a | |
1586 | * hugepage (considered as already dissolved) | |
c8721bbb | 1587 | */ |
c3114a84 | 1588 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1589 | { |
6bc9b564 | 1590 | int rc = -EBUSY; |
082d5b6b | 1591 | |
faf53def NH |
1592 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1593 | if (!PageHuge(page)) | |
1594 | return 0; | |
1595 | ||
c8721bbb | 1596 | spin_lock(&hugetlb_lock); |
faf53def NH |
1597 | if (!PageHuge(page)) { |
1598 | rc = 0; | |
1599 | goto out; | |
1600 | } | |
1601 | ||
1602 | if (!page_count(page)) { | |
2247bb33 GS |
1603 | struct page *head = compound_head(page); |
1604 | struct hstate *h = page_hstate(head); | |
1605 | int nid = page_to_nid(head); | |
6bc9b564 | 1606 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1607 | goto out; |
c3114a84 AK |
1608 | /* |
1609 | * Move PageHWPoison flag from head page to the raw error page, | |
1610 | * which makes any subpages rather than the error page reusable. | |
1611 | */ | |
1612 | if (PageHWPoison(head) && page != head) { | |
1613 | SetPageHWPoison(page); | |
1614 | ClearPageHWPoison(head); | |
1615 | } | |
2247bb33 | 1616 | list_del(&head->lru); |
c8721bbb NH |
1617 | h->free_huge_pages--; |
1618 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1619 | h->max_huge_pages--; |
2247bb33 | 1620 | update_and_free_page(h, head); |
6bc9b564 | 1621 | rc = 0; |
c8721bbb | 1622 | } |
082d5b6b | 1623 | out: |
c8721bbb | 1624 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1625 | return rc; |
c8721bbb NH |
1626 | } |
1627 | ||
1628 | /* | |
1629 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1630 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1631 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1632 | * part of it lies within the given range. | |
082d5b6b GS |
1633 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1634 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1635 | */ |
082d5b6b | 1636 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1637 | { |
c8721bbb | 1638 | unsigned long pfn; |
eb03aa00 | 1639 | struct page *page; |
082d5b6b | 1640 | int rc = 0; |
c8721bbb | 1641 | |
d0177639 | 1642 | if (!hugepages_supported()) |
082d5b6b | 1643 | return rc; |
d0177639 | 1644 | |
eb03aa00 GS |
1645 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1646 | page = pfn_to_page(pfn); | |
faf53def NH |
1647 | rc = dissolve_free_huge_page(page); |
1648 | if (rc) | |
1649 | break; | |
eb03aa00 | 1650 | } |
082d5b6b GS |
1651 | |
1652 | return rc; | |
c8721bbb NH |
1653 | } |
1654 | ||
ab5ac90a MH |
1655 | /* |
1656 | * Allocates a fresh surplus page from the page allocator. | |
1657 | */ | |
0c397dae | 1658 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1659 | int nid, nodemask_t *nmask) |
7893d1d5 | 1660 | { |
9980d744 | 1661 | struct page *page = NULL; |
7893d1d5 | 1662 | |
bae7f4ae | 1663 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1664 | return NULL; |
1665 | ||
d1c3fb1f | 1666 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1667 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1668 | goto out_unlock; | |
d1c3fb1f NA |
1669 | spin_unlock(&hugetlb_lock); |
1670 | ||
f60858f9 | 1671 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1672 | if (!page) |
0c397dae | 1673 | return NULL; |
d1c3fb1f NA |
1674 | |
1675 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1676 | /* |
1677 | * We could have raced with the pool size change. | |
1678 | * Double check that and simply deallocate the new page | |
1679 | * if we would end up overcommiting the surpluses. Abuse | |
1680 | * temporary page to workaround the nasty free_huge_page | |
1681 | * codeflow | |
1682 | */ | |
1683 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1684 | SetPageHugeTemporary(page); | |
2bf753e6 | 1685 | spin_unlock(&hugetlb_lock); |
9980d744 | 1686 | put_page(page); |
2bf753e6 | 1687 | return NULL; |
9980d744 | 1688 | } else { |
9980d744 | 1689 | h->surplus_huge_pages++; |
4704dea3 | 1690 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1691 | } |
9980d744 MH |
1692 | |
1693 | out_unlock: | |
d1c3fb1f | 1694 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1695 | |
1696 | return page; | |
1697 | } | |
1698 | ||
9a4e9f3b AK |
1699 | struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
1700 | int nid, nodemask_t *nmask) | |
ab5ac90a MH |
1701 | { |
1702 | struct page *page; | |
1703 | ||
1704 | if (hstate_is_gigantic(h)) | |
1705 | return NULL; | |
1706 | ||
f60858f9 | 1707 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1708 | if (!page) |
1709 | return NULL; | |
1710 | ||
1711 | /* | |
1712 | * We do not account these pages as surplus because they are only | |
1713 | * temporary and will be released properly on the last reference | |
1714 | */ | |
ab5ac90a MH |
1715 | SetPageHugeTemporary(page); |
1716 | ||
1717 | return page; | |
1718 | } | |
1719 | ||
099730d6 DH |
1720 | /* |
1721 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1722 | */ | |
e0ec90ee | 1723 | static |
0c397dae | 1724 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1725 | struct vm_area_struct *vma, unsigned long addr) |
1726 | { | |
aaf14e40 MH |
1727 | struct page *page; |
1728 | struct mempolicy *mpol; | |
1729 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1730 | int nid; | |
1731 | nodemask_t *nodemask; | |
1732 | ||
1733 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1734 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1735 | mpol_cond_put(mpol); |
1736 | ||
1737 | return page; | |
099730d6 DH |
1738 | } |
1739 | ||
ab5ac90a | 1740 | /* page migration callback function */ |
bf50bab2 NH |
1741 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
1742 | { | |
aaf14e40 | 1743 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1744 | struct page *page = NULL; |
bf50bab2 | 1745 | |
aaf14e40 MH |
1746 | if (nid != NUMA_NO_NODE) |
1747 | gfp_mask |= __GFP_THISNODE; | |
1748 | ||
bf50bab2 | 1749 | spin_lock(&hugetlb_lock); |
4ef91848 | 1750 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1751 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1752 | spin_unlock(&hugetlb_lock); |
1753 | ||
94ae8ba7 | 1754 | if (!page) |
0c397dae | 1755 | page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1756 | |
1757 | return page; | |
1758 | } | |
1759 | ||
ab5ac90a | 1760 | /* page migration callback function */ |
3e59fcb0 MH |
1761 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
1762 | nodemask_t *nmask) | |
4db9b2ef | 1763 | { |
aaf14e40 | 1764 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
1765 | |
1766 | spin_lock(&hugetlb_lock); | |
1767 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1768 | struct page *page; |
1769 | ||
1770 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1771 | if (page) { | |
1772 | spin_unlock(&hugetlb_lock); | |
1773 | return page; | |
4db9b2ef MH |
1774 | } |
1775 | } | |
1776 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1777 | |
0c397dae | 1778 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1779 | } |
1780 | ||
ebd63723 | 1781 | /* mempolicy aware migration callback */ |
389c8178 MH |
1782 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1783 | unsigned long address) | |
ebd63723 MH |
1784 | { |
1785 | struct mempolicy *mpol; | |
1786 | nodemask_t *nodemask; | |
1787 | struct page *page; | |
ebd63723 MH |
1788 | gfp_t gfp_mask; |
1789 | int node; | |
1790 | ||
ebd63723 MH |
1791 | gfp_mask = htlb_alloc_mask(h); |
1792 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
1793 | page = alloc_huge_page_nodemask(h, node, nodemask); | |
1794 | mpol_cond_put(mpol); | |
1795 | ||
1796 | return page; | |
1797 | } | |
1798 | ||
e4e574b7 | 1799 | /* |
25985edc | 1800 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1801 | * of size 'delta'. |
1802 | */ | |
a5516438 | 1803 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1804 | { |
1805 | struct list_head surplus_list; | |
1806 | struct page *page, *tmp; | |
1807 | int ret, i; | |
1808 | int needed, allocated; | |
28073b02 | 1809 | bool alloc_ok = true; |
e4e574b7 | 1810 | |
a5516438 | 1811 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1812 | if (needed <= 0) { |
a5516438 | 1813 | h->resv_huge_pages += delta; |
e4e574b7 | 1814 | return 0; |
ac09b3a1 | 1815 | } |
e4e574b7 AL |
1816 | |
1817 | allocated = 0; | |
1818 | INIT_LIST_HEAD(&surplus_list); | |
1819 | ||
1820 | ret = -ENOMEM; | |
1821 | retry: | |
1822 | spin_unlock(&hugetlb_lock); | |
1823 | for (i = 0; i < needed; i++) { | |
0c397dae | 1824 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1825 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1826 | if (!page) { |
1827 | alloc_ok = false; | |
1828 | break; | |
1829 | } | |
e4e574b7 | 1830 | list_add(&page->lru, &surplus_list); |
69ed779a | 1831 | cond_resched(); |
e4e574b7 | 1832 | } |
28073b02 | 1833 | allocated += i; |
e4e574b7 AL |
1834 | |
1835 | /* | |
1836 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1837 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1838 | */ | |
1839 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1840 | needed = (h->resv_huge_pages + delta) - |
1841 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1842 | if (needed > 0) { |
1843 | if (alloc_ok) | |
1844 | goto retry; | |
1845 | /* | |
1846 | * We were not able to allocate enough pages to | |
1847 | * satisfy the entire reservation so we free what | |
1848 | * we've allocated so far. | |
1849 | */ | |
1850 | goto free; | |
1851 | } | |
e4e574b7 AL |
1852 | /* |
1853 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1854 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1855 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1856 | * allocator. Commit the entire reservation here to prevent another |
1857 | * process from stealing the pages as they are added to the pool but | |
1858 | * before they are reserved. | |
e4e574b7 AL |
1859 | */ |
1860 | needed += allocated; | |
a5516438 | 1861 | h->resv_huge_pages += delta; |
e4e574b7 | 1862 | ret = 0; |
a9869b83 | 1863 | |
19fc3f0a | 1864 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1865 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1866 | if ((--needed) < 0) |
1867 | break; | |
a9869b83 NH |
1868 | /* |
1869 | * This page is now managed by the hugetlb allocator and has | |
1870 | * no users -- drop the buddy allocator's reference. | |
1871 | */ | |
1872 | put_page_testzero(page); | |
309381fe | 1873 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1874 | enqueue_huge_page(h, page); |
19fc3f0a | 1875 | } |
28073b02 | 1876 | free: |
b0365c8d | 1877 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1878 | |
1879 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1880 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1881 | put_page(page); | |
a9869b83 | 1882 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1883 | |
1884 | return ret; | |
1885 | } | |
1886 | ||
1887 | /* | |
e5bbc8a6 MK |
1888 | * This routine has two main purposes: |
1889 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1890 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1891 | * to the associated reservation map. | |
1892 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1893 | * the reservation. As many as unused_resv_pages may be freed. | |
1894 | * | |
1895 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1896 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1897 | * we must make sure nobody else can claim pages we are in the process of | |
1898 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1899 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1900 | */ |
a5516438 AK |
1901 | static void return_unused_surplus_pages(struct hstate *h, |
1902 | unsigned long unused_resv_pages) | |
e4e574b7 | 1903 | { |
e4e574b7 AL |
1904 | unsigned long nr_pages; |
1905 | ||
aa888a74 | 1906 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1907 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1908 | goto out; |
aa888a74 | 1909 | |
e5bbc8a6 MK |
1910 | /* |
1911 | * Part (or even all) of the reservation could have been backed | |
1912 | * by pre-allocated pages. Only free surplus pages. | |
1913 | */ | |
a5516438 | 1914 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1915 | |
685f3457 LS |
1916 | /* |
1917 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1918 | * evenly across all nodes with memory. Iterate across these nodes |
1919 | * until we can no longer free unreserved surplus pages. This occurs | |
1920 | * when the nodes with surplus pages have no free pages. | |
1921 | * free_pool_huge_page() will balance the the freed pages across the | |
1922 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1923 | * |
1924 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1925 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1926 | * to cover subsequent pages we may free. | |
685f3457 LS |
1927 | */ |
1928 | while (nr_pages--) { | |
e5bbc8a6 MK |
1929 | h->resv_huge_pages--; |
1930 | unused_resv_pages--; | |
8cebfcd0 | 1931 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1932 | goto out; |
7848a4bf | 1933 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1934 | } |
e5bbc8a6 MK |
1935 | |
1936 | out: | |
1937 | /* Fully uncommit the reservation */ | |
1938 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1939 | } |
1940 | ||
5e911373 | 1941 | |
c37f9fb1 | 1942 | /* |
feba16e2 | 1943 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1944 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1945 | * |
1946 | * vma_needs_reservation is called to determine if the huge page at addr | |
1947 | * within the vma has an associated reservation. If a reservation is | |
1948 | * needed, the value 1 is returned. The caller is then responsible for | |
1949 | * managing the global reservation and subpool usage counts. After | |
1950 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1951 | * to add the page to the reservation map. If the page allocation fails, |
1952 | * the reservation must be ended instead of committed. vma_end_reservation | |
1953 | * is called in such cases. | |
cf3ad20b MK |
1954 | * |
1955 | * In the normal case, vma_commit_reservation returns the same value | |
1956 | * as the preceding vma_needs_reservation call. The only time this | |
1957 | * is not the case is if a reserve map was changed between calls. It | |
1958 | * is the responsibility of the caller to notice the difference and | |
1959 | * take appropriate action. | |
96b96a96 MK |
1960 | * |
1961 | * vma_add_reservation is used in error paths where a reservation must | |
1962 | * be restored when a newly allocated huge page must be freed. It is | |
1963 | * to be called after calling vma_needs_reservation to determine if a | |
1964 | * reservation exists. | |
c37f9fb1 | 1965 | */ |
5e911373 MK |
1966 | enum vma_resv_mode { |
1967 | VMA_NEEDS_RESV, | |
1968 | VMA_COMMIT_RESV, | |
feba16e2 | 1969 | VMA_END_RESV, |
96b96a96 | 1970 | VMA_ADD_RESV, |
5e911373 | 1971 | }; |
cf3ad20b MK |
1972 | static long __vma_reservation_common(struct hstate *h, |
1973 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1974 | enum vma_resv_mode mode) |
c37f9fb1 | 1975 | { |
4e35f483 JK |
1976 | struct resv_map *resv; |
1977 | pgoff_t idx; | |
cf3ad20b | 1978 | long ret; |
c37f9fb1 | 1979 | |
4e35f483 JK |
1980 | resv = vma_resv_map(vma); |
1981 | if (!resv) | |
84afd99b | 1982 | return 1; |
c37f9fb1 | 1983 | |
4e35f483 | 1984 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1985 | switch (mode) { |
1986 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1987 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1988 | break; |
1989 | case VMA_COMMIT_RESV: | |
1990 | ret = region_add(resv, idx, idx + 1); | |
1991 | break; | |
feba16e2 | 1992 | case VMA_END_RESV: |
5e911373 MK |
1993 | region_abort(resv, idx, idx + 1); |
1994 | ret = 0; | |
1995 | break; | |
96b96a96 MK |
1996 | case VMA_ADD_RESV: |
1997 | if (vma->vm_flags & VM_MAYSHARE) | |
1998 | ret = region_add(resv, idx, idx + 1); | |
1999 | else { | |
2000 | region_abort(resv, idx, idx + 1); | |
2001 | ret = region_del(resv, idx, idx + 1); | |
2002 | } | |
2003 | break; | |
5e911373 MK |
2004 | default: |
2005 | BUG(); | |
2006 | } | |
84afd99b | 2007 | |
4e35f483 | 2008 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 2009 | return ret; |
67961f9d MK |
2010 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
2011 | /* | |
2012 | * In most cases, reserves always exist for private mappings. | |
2013 | * However, a file associated with mapping could have been | |
2014 | * hole punched or truncated after reserves were consumed. | |
2015 | * As subsequent fault on such a range will not use reserves. | |
2016 | * Subtle - The reserve map for private mappings has the | |
2017 | * opposite meaning than that of shared mappings. If NO | |
2018 | * entry is in the reserve map, it means a reservation exists. | |
2019 | * If an entry exists in the reserve map, it means the | |
2020 | * reservation has already been consumed. As a result, the | |
2021 | * return value of this routine is the opposite of the | |
2022 | * value returned from reserve map manipulation routines above. | |
2023 | */ | |
2024 | if (ret) | |
2025 | return 0; | |
2026 | else | |
2027 | return 1; | |
2028 | } | |
4e35f483 | 2029 | else |
cf3ad20b | 2030 | return ret < 0 ? ret : 0; |
c37f9fb1 | 2031 | } |
cf3ad20b MK |
2032 | |
2033 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2034 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2035 | { |
5e911373 | 2036 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2037 | } |
84afd99b | 2038 | |
cf3ad20b MK |
2039 | static long vma_commit_reservation(struct hstate *h, |
2040 | struct vm_area_struct *vma, unsigned long addr) | |
2041 | { | |
5e911373 MK |
2042 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2043 | } | |
2044 | ||
feba16e2 | 2045 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2046 | struct vm_area_struct *vma, unsigned long addr) |
2047 | { | |
feba16e2 | 2048 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2049 | } |
2050 | ||
96b96a96 MK |
2051 | static long vma_add_reservation(struct hstate *h, |
2052 | struct vm_area_struct *vma, unsigned long addr) | |
2053 | { | |
2054 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2055 | } | |
2056 | ||
2057 | /* | |
2058 | * This routine is called to restore a reservation on error paths. In the | |
2059 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2060 | * and is about to be freed. If a reservation for the page existed, | |
2061 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2062 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2063 | * the global reservation count will be incremented if PagePrivate is set. | |
2064 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2065 | * reserve map here to be consistent with global reserve count adjustments | |
2066 | * to be made by free_huge_page. | |
2067 | */ | |
2068 | static void restore_reserve_on_error(struct hstate *h, | |
2069 | struct vm_area_struct *vma, unsigned long address, | |
2070 | struct page *page) | |
2071 | { | |
2072 | if (unlikely(PagePrivate(page))) { | |
2073 | long rc = vma_needs_reservation(h, vma, address); | |
2074 | ||
2075 | if (unlikely(rc < 0)) { | |
2076 | /* | |
2077 | * Rare out of memory condition in reserve map | |
2078 | * manipulation. Clear PagePrivate so that | |
2079 | * global reserve count will not be incremented | |
2080 | * by free_huge_page. This will make it appear | |
2081 | * as though the reservation for this page was | |
2082 | * consumed. This may prevent the task from | |
2083 | * faulting in the page at a later time. This | |
2084 | * is better than inconsistent global huge page | |
2085 | * accounting of reserve counts. | |
2086 | */ | |
2087 | ClearPagePrivate(page); | |
2088 | } else if (rc) { | |
2089 | rc = vma_add_reservation(h, vma, address); | |
2090 | if (unlikely(rc < 0)) | |
2091 | /* | |
2092 | * See above comment about rare out of | |
2093 | * memory condition. | |
2094 | */ | |
2095 | ClearPagePrivate(page); | |
2096 | } else | |
2097 | vma_end_reservation(h, vma, address); | |
2098 | } | |
2099 | } | |
2100 | ||
70c3547e | 2101 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2102 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2103 | { |
90481622 | 2104 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2105 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2106 | struct page *page; |
d85f69b0 MK |
2107 | long map_chg, map_commit; |
2108 | long gbl_chg; | |
6d76dcf4 AK |
2109 | int ret, idx; |
2110 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 2111 | |
6d76dcf4 | 2112 | idx = hstate_index(h); |
a1e78772 | 2113 | /* |
d85f69b0 MK |
2114 | * Examine the region/reserve map to determine if the process |
2115 | * has a reservation for the page to be allocated. A return | |
2116 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2117 | */ |
d85f69b0 MK |
2118 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2119 | if (map_chg < 0) | |
76dcee75 | 2120 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2121 | |
2122 | /* | |
2123 | * Processes that did not create the mapping will have no | |
2124 | * reserves as indicated by the region/reserve map. Check | |
2125 | * that the allocation will not exceed the subpool limit. | |
2126 | * Allocations for MAP_NORESERVE mappings also need to be | |
2127 | * checked against any subpool limit. | |
2128 | */ | |
2129 | if (map_chg || avoid_reserve) { | |
2130 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2131 | if (gbl_chg < 0) { | |
feba16e2 | 2132 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2133 | return ERR_PTR(-ENOSPC); |
5e911373 | 2134 | } |
1da177e4 | 2135 | |
d85f69b0 MK |
2136 | /* |
2137 | * Even though there was no reservation in the region/reserve | |
2138 | * map, there could be reservations associated with the | |
2139 | * subpool that can be used. This would be indicated if the | |
2140 | * return value of hugepage_subpool_get_pages() is zero. | |
2141 | * However, if avoid_reserve is specified we still avoid even | |
2142 | * the subpool reservations. | |
2143 | */ | |
2144 | if (avoid_reserve) | |
2145 | gbl_chg = 1; | |
2146 | } | |
2147 | ||
6d76dcf4 | 2148 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2149 | if (ret) |
2150 | goto out_subpool_put; | |
2151 | ||
1da177e4 | 2152 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2153 | /* |
2154 | * glb_chg is passed to indicate whether or not a page must be taken | |
2155 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2156 | * a reservation exists for the allocation. | |
2157 | */ | |
2158 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2159 | if (!page) { |
94ae8ba7 | 2160 | spin_unlock(&hugetlb_lock); |
0c397dae | 2161 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2162 | if (!page) |
2163 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2164 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2165 | SetPagePrivate(page); | |
2166 | h->resv_huge_pages--; | |
2167 | } | |
79dbb236 AK |
2168 | spin_lock(&hugetlb_lock); |
2169 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2170 | /* Fall through */ |
68842c9b | 2171 | } |
81a6fcae JK |
2172 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2173 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2174 | |
90481622 | 2175 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2176 | |
d85f69b0 MK |
2177 | map_commit = vma_commit_reservation(h, vma, addr); |
2178 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2179 | /* |
2180 | * The page was added to the reservation map between | |
2181 | * vma_needs_reservation and vma_commit_reservation. | |
2182 | * This indicates a race with hugetlb_reserve_pages. | |
2183 | * Adjust for the subpool count incremented above AND | |
2184 | * in hugetlb_reserve_pages for the same page. Also, | |
2185 | * the reservation count added in hugetlb_reserve_pages | |
2186 | * no longer applies. | |
2187 | */ | |
2188 | long rsv_adjust; | |
2189 | ||
2190 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2191 | hugetlb_acct_memory(h, -rsv_adjust); | |
2192 | } | |
90d8b7e6 | 2193 | return page; |
8f34af6f JZ |
2194 | |
2195 | out_uncharge_cgroup: | |
2196 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2197 | out_subpool_put: | |
d85f69b0 | 2198 | if (map_chg || avoid_reserve) |
8f34af6f | 2199 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2200 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2201 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2202 | } |
2203 | ||
e24a1307 AK |
2204 | int alloc_bootmem_huge_page(struct hstate *h) |
2205 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2206 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2207 | { |
2208 | struct huge_bootmem_page *m; | |
b2261026 | 2209 | int nr_nodes, node; |
aa888a74 | 2210 | |
b2261026 | 2211 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2212 | void *addr; |
2213 | ||
eb31d559 | 2214 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2215 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2216 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2217 | if (addr) { |
2218 | /* | |
2219 | * Use the beginning of the huge page to store the | |
2220 | * huge_bootmem_page struct (until gather_bootmem | |
2221 | * puts them into the mem_map). | |
2222 | */ | |
2223 | m = addr; | |
91f47662 | 2224 | goto found; |
aa888a74 | 2225 | } |
aa888a74 AK |
2226 | } |
2227 | return 0; | |
2228 | ||
2229 | found: | |
df994ead | 2230 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2231 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2232 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2233 | list_add(&m->list, &huge_boot_pages); |
2234 | m->hstate = h; | |
2235 | return 1; | |
2236 | } | |
2237 | ||
d00181b9 KS |
2238 | static void __init prep_compound_huge_page(struct page *page, |
2239 | unsigned int order) | |
18229df5 AW |
2240 | { |
2241 | if (unlikely(order > (MAX_ORDER - 1))) | |
2242 | prep_compound_gigantic_page(page, order); | |
2243 | else | |
2244 | prep_compound_page(page, order); | |
2245 | } | |
2246 | ||
aa888a74 AK |
2247 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2248 | static void __init gather_bootmem_prealloc(void) | |
2249 | { | |
2250 | struct huge_bootmem_page *m; | |
2251 | ||
2252 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2253 | struct page *page = virt_to_page(m); |
aa888a74 | 2254 | struct hstate *h = m->hstate; |
ee8f248d | 2255 | |
aa888a74 | 2256 | WARN_ON(page_count(page) != 1); |
18229df5 | 2257 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2258 | WARN_ON(PageReserved(page)); |
aa888a74 | 2259 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2260 | put_page(page); /* free it into the hugepage allocator */ |
2261 | ||
b0320c7b RA |
2262 | /* |
2263 | * If we had gigantic hugepages allocated at boot time, we need | |
2264 | * to restore the 'stolen' pages to totalram_pages in order to | |
2265 | * fix confusing memory reports from free(1) and another | |
2266 | * side-effects, like CommitLimit going negative. | |
2267 | */ | |
bae7f4ae | 2268 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2269 | adjust_managed_page_count(page, 1 << h->order); |
520495fe | 2270 | cond_resched(); |
aa888a74 AK |
2271 | } |
2272 | } | |
2273 | ||
8faa8b07 | 2274 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2275 | { |
2276 | unsigned long i; | |
f60858f9 MK |
2277 | nodemask_t *node_alloc_noretry; |
2278 | ||
2279 | if (!hstate_is_gigantic(h)) { | |
2280 | /* | |
2281 | * Bit mask controlling how hard we retry per-node allocations. | |
2282 | * Ignore errors as lower level routines can deal with | |
2283 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2284 | * time, we are likely in bigger trouble. | |
2285 | */ | |
2286 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2287 | GFP_KERNEL); | |
2288 | } else { | |
2289 | /* allocations done at boot time */ | |
2290 | node_alloc_noretry = NULL; | |
2291 | } | |
2292 | ||
2293 | /* bit mask controlling how hard we retry per-node allocations */ | |
2294 | if (node_alloc_noretry) | |
2295 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2296 | |
e5ff2159 | 2297 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2298 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2299 | if (!alloc_bootmem_huge_page(h)) |
2300 | break; | |
0c397dae | 2301 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2302 | &node_states[N_MEMORY], |
2303 | node_alloc_noretry)) | |
1da177e4 | 2304 | break; |
69ed779a | 2305 | cond_resched(); |
1da177e4 | 2306 | } |
d715cf80 LH |
2307 | if (i < h->max_huge_pages) { |
2308 | char buf[32]; | |
2309 | ||
c6247f72 | 2310 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2311 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2312 | h->max_huge_pages, buf, i); | |
2313 | h->max_huge_pages = i; | |
2314 | } | |
f60858f9 MK |
2315 | |
2316 | kfree(node_alloc_noretry); | |
e5ff2159 AK |
2317 | } |
2318 | ||
2319 | static void __init hugetlb_init_hstates(void) | |
2320 | { | |
2321 | struct hstate *h; | |
2322 | ||
2323 | for_each_hstate(h) { | |
641844f5 NH |
2324 | if (minimum_order > huge_page_order(h)) |
2325 | minimum_order = huge_page_order(h); | |
2326 | ||
8faa8b07 | 2327 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2328 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2329 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2330 | } |
641844f5 | 2331 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2332 | } |
2333 | ||
2334 | static void __init report_hugepages(void) | |
2335 | { | |
2336 | struct hstate *h; | |
2337 | ||
2338 | for_each_hstate(h) { | |
4abd32db | 2339 | char buf[32]; |
c6247f72 MW |
2340 | |
2341 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2342 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2343 | buf, h->free_huge_pages); |
e5ff2159 AK |
2344 | } |
2345 | } | |
2346 | ||
1da177e4 | 2347 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2348 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2349 | nodemask_t *nodes_allowed) | |
1da177e4 | 2350 | { |
4415cc8d CL |
2351 | int i; |
2352 | ||
bae7f4ae | 2353 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2354 | return; |
2355 | ||
6ae11b27 | 2356 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2357 | struct page *page, *next; |
a5516438 AK |
2358 | struct list_head *freel = &h->hugepage_freelists[i]; |
2359 | list_for_each_entry_safe(page, next, freel, lru) { | |
2360 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2361 | return; |
1da177e4 LT |
2362 | if (PageHighMem(page)) |
2363 | continue; | |
2364 | list_del(&page->lru); | |
e5ff2159 | 2365 | update_and_free_page(h, page); |
a5516438 AK |
2366 | h->free_huge_pages--; |
2367 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2368 | } |
2369 | } | |
2370 | } | |
2371 | #else | |
6ae11b27 LS |
2372 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2373 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2374 | { |
2375 | } | |
2376 | #endif | |
2377 | ||
20a0307c WF |
2378 | /* |
2379 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2380 | * balanced by operating on them in a round-robin fashion. | |
2381 | * Returns 1 if an adjustment was made. | |
2382 | */ | |
6ae11b27 LS |
2383 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2384 | int delta) | |
20a0307c | 2385 | { |
b2261026 | 2386 | int nr_nodes, node; |
20a0307c WF |
2387 | |
2388 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2389 | |
b2261026 JK |
2390 | if (delta < 0) { |
2391 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2392 | if (h->surplus_huge_pages_node[node]) | |
2393 | goto found; | |
e8c5c824 | 2394 | } |
b2261026 JK |
2395 | } else { |
2396 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2397 | if (h->surplus_huge_pages_node[node] < | |
2398 | h->nr_huge_pages_node[node]) | |
2399 | goto found; | |
e8c5c824 | 2400 | } |
b2261026 JK |
2401 | } |
2402 | return 0; | |
20a0307c | 2403 | |
b2261026 JK |
2404 | found: |
2405 | h->surplus_huge_pages += delta; | |
2406 | h->surplus_huge_pages_node[node] += delta; | |
2407 | return 1; | |
20a0307c WF |
2408 | } |
2409 | ||
a5516438 | 2410 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2411 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2412 | nodemask_t *nodes_allowed) |
1da177e4 | 2413 | { |
7893d1d5 | 2414 | unsigned long min_count, ret; |
f60858f9 MK |
2415 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2416 | ||
2417 | /* | |
2418 | * Bit mask controlling how hard we retry per-node allocations. | |
2419 | * If we can not allocate the bit mask, do not attempt to allocate | |
2420 | * the requested huge pages. | |
2421 | */ | |
2422 | if (node_alloc_noretry) | |
2423 | nodes_clear(*node_alloc_noretry); | |
2424 | else | |
2425 | return -ENOMEM; | |
1da177e4 | 2426 | |
4eb0716e AG |
2427 | spin_lock(&hugetlb_lock); |
2428 | ||
fd875dca MK |
2429 | /* |
2430 | * Check for a node specific request. | |
2431 | * Changing node specific huge page count may require a corresponding | |
2432 | * change to the global count. In any case, the passed node mask | |
2433 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2434 | */ | |
2435 | if (nid != NUMA_NO_NODE) { | |
2436 | unsigned long old_count = count; | |
2437 | ||
2438 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2439 | /* | |
2440 | * User may have specified a large count value which caused the | |
2441 | * above calculation to overflow. In this case, they wanted | |
2442 | * to allocate as many huge pages as possible. Set count to | |
2443 | * largest possible value to align with their intention. | |
2444 | */ | |
2445 | if (count < old_count) | |
2446 | count = ULONG_MAX; | |
2447 | } | |
2448 | ||
4eb0716e AG |
2449 | /* |
2450 | * Gigantic pages runtime allocation depend on the capability for large | |
2451 | * page range allocation. | |
2452 | * If the system does not provide this feature, return an error when | |
2453 | * the user tries to allocate gigantic pages but let the user free the | |
2454 | * boottime allocated gigantic pages. | |
2455 | */ | |
2456 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2457 | if (count > persistent_huge_pages(h)) { | |
2458 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2459 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2460 | return -EINVAL; |
2461 | } | |
2462 | /* Fall through to decrease pool */ | |
2463 | } | |
aa888a74 | 2464 | |
7893d1d5 AL |
2465 | /* |
2466 | * Increase the pool size | |
2467 | * First take pages out of surplus state. Then make up the | |
2468 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2469 | * |
0c397dae | 2470 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2471 | * to convert a surplus huge page to a normal huge page. That is |
2472 | * not critical, though, it just means the overall size of the | |
2473 | * pool might be one hugepage larger than it needs to be, but | |
2474 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2475 | */ |
a5516438 | 2476 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2477 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2478 | break; |
2479 | } | |
2480 | ||
a5516438 | 2481 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2482 | /* |
2483 | * If this allocation races such that we no longer need the | |
2484 | * page, free_huge_page will handle it by freeing the page | |
2485 | * and reducing the surplus. | |
2486 | */ | |
2487 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2488 | |
2489 | /* yield cpu to avoid soft lockup */ | |
2490 | cond_resched(); | |
2491 | ||
f60858f9 MK |
2492 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2493 | node_alloc_noretry); | |
7893d1d5 AL |
2494 | spin_lock(&hugetlb_lock); |
2495 | if (!ret) | |
2496 | goto out; | |
2497 | ||
536240f2 MG |
2498 | /* Bail for signals. Probably ctrl-c from user */ |
2499 | if (signal_pending(current)) | |
2500 | goto out; | |
7893d1d5 | 2501 | } |
7893d1d5 AL |
2502 | |
2503 | /* | |
2504 | * Decrease the pool size | |
2505 | * First return free pages to the buddy allocator (being careful | |
2506 | * to keep enough around to satisfy reservations). Then place | |
2507 | * pages into surplus state as needed so the pool will shrink | |
2508 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2509 | * |
2510 | * By placing pages into the surplus state independent of the | |
2511 | * overcommit value, we are allowing the surplus pool size to | |
2512 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2513 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2514 | * though, we'll note that we're not allowed to exceed surplus |
2515 | * and won't grow the pool anywhere else. Not until one of the | |
2516 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2517 | */ |
a5516438 | 2518 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2519 | min_count = max(count, min_count); |
6ae11b27 | 2520 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2521 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2522 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2523 | break; |
55f67141 | 2524 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2525 | } |
a5516438 | 2526 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2527 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2528 | break; |
2529 | } | |
2530 | out: | |
4eb0716e | 2531 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2532 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2533 | |
f60858f9 MK |
2534 | NODEMASK_FREE(node_alloc_noretry); |
2535 | ||
4eb0716e | 2536 | return 0; |
1da177e4 LT |
2537 | } |
2538 | ||
a3437870 NA |
2539 | #define HSTATE_ATTR_RO(_name) \ |
2540 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2541 | ||
2542 | #define HSTATE_ATTR(_name) \ | |
2543 | static struct kobj_attribute _name##_attr = \ | |
2544 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2545 | ||
2546 | static struct kobject *hugepages_kobj; | |
2547 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2548 | ||
9a305230 LS |
2549 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2550 | ||
2551 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2552 | { |
2553 | int i; | |
9a305230 | 2554 | |
a3437870 | 2555 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2556 | if (hstate_kobjs[i] == kobj) { |
2557 | if (nidp) | |
2558 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2559 | return &hstates[i]; |
9a305230 LS |
2560 | } |
2561 | ||
2562 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2563 | } |
2564 | ||
06808b08 | 2565 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2566 | struct kobj_attribute *attr, char *buf) |
2567 | { | |
9a305230 LS |
2568 | struct hstate *h; |
2569 | unsigned long nr_huge_pages; | |
2570 | int nid; | |
2571 | ||
2572 | h = kobj_to_hstate(kobj, &nid); | |
2573 | if (nid == NUMA_NO_NODE) | |
2574 | nr_huge_pages = h->nr_huge_pages; | |
2575 | else | |
2576 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2577 | ||
2578 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2579 | } |
adbe8726 | 2580 | |
238d3c13 DR |
2581 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2582 | struct hstate *h, int nid, | |
2583 | unsigned long count, size_t len) | |
a3437870 NA |
2584 | { |
2585 | int err; | |
2d0adf7e | 2586 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2587 | |
2d0adf7e OS |
2588 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2589 | return -EINVAL; | |
adbe8726 | 2590 | |
9a305230 LS |
2591 | if (nid == NUMA_NO_NODE) { |
2592 | /* | |
2593 | * global hstate attribute | |
2594 | */ | |
2595 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2596 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2597 | n_mask = &node_states[N_MEMORY]; | |
2598 | else | |
2599 | n_mask = &nodes_allowed; | |
2600 | } else { | |
9a305230 | 2601 | /* |
fd875dca MK |
2602 | * Node specific request. count adjustment happens in |
2603 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2604 | */ |
2d0adf7e OS |
2605 | init_nodemask_of_node(&nodes_allowed, nid); |
2606 | n_mask = &nodes_allowed; | |
fd875dca | 2607 | } |
9a305230 | 2608 | |
2d0adf7e | 2609 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2610 | |
4eb0716e | 2611 | return err ? err : len; |
06808b08 LS |
2612 | } |
2613 | ||
238d3c13 DR |
2614 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2615 | struct kobject *kobj, const char *buf, | |
2616 | size_t len) | |
2617 | { | |
2618 | struct hstate *h; | |
2619 | unsigned long count; | |
2620 | int nid; | |
2621 | int err; | |
2622 | ||
2623 | err = kstrtoul(buf, 10, &count); | |
2624 | if (err) | |
2625 | return err; | |
2626 | ||
2627 | h = kobj_to_hstate(kobj, &nid); | |
2628 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2629 | } | |
2630 | ||
06808b08 LS |
2631 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2632 | struct kobj_attribute *attr, char *buf) | |
2633 | { | |
2634 | return nr_hugepages_show_common(kobj, attr, buf); | |
2635 | } | |
2636 | ||
2637 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2638 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2639 | { | |
238d3c13 | 2640 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2641 | } |
2642 | HSTATE_ATTR(nr_hugepages); | |
2643 | ||
06808b08 LS |
2644 | #ifdef CONFIG_NUMA |
2645 | ||
2646 | /* | |
2647 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2648 | * huge page alloc/free. | |
2649 | */ | |
2650 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2651 | struct kobj_attribute *attr, char *buf) | |
2652 | { | |
2653 | return nr_hugepages_show_common(kobj, attr, buf); | |
2654 | } | |
2655 | ||
2656 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2657 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2658 | { | |
238d3c13 | 2659 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2660 | } |
2661 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2662 | #endif | |
2663 | ||
2664 | ||
a3437870 NA |
2665 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2666 | struct kobj_attribute *attr, char *buf) | |
2667 | { | |
9a305230 | 2668 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2669 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2670 | } | |
adbe8726 | 2671 | |
a3437870 NA |
2672 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2673 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2674 | { | |
2675 | int err; | |
2676 | unsigned long input; | |
9a305230 | 2677 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2678 | |
bae7f4ae | 2679 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2680 | return -EINVAL; |
2681 | ||
3dbb95f7 | 2682 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2683 | if (err) |
73ae31e5 | 2684 | return err; |
a3437870 NA |
2685 | |
2686 | spin_lock(&hugetlb_lock); | |
2687 | h->nr_overcommit_huge_pages = input; | |
2688 | spin_unlock(&hugetlb_lock); | |
2689 | ||
2690 | return count; | |
2691 | } | |
2692 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2693 | ||
2694 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2695 | struct kobj_attribute *attr, char *buf) | |
2696 | { | |
9a305230 LS |
2697 | struct hstate *h; |
2698 | unsigned long free_huge_pages; | |
2699 | int nid; | |
2700 | ||
2701 | h = kobj_to_hstate(kobj, &nid); | |
2702 | if (nid == NUMA_NO_NODE) | |
2703 | free_huge_pages = h->free_huge_pages; | |
2704 | else | |
2705 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2706 | ||
2707 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2708 | } |
2709 | HSTATE_ATTR_RO(free_hugepages); | |
2710 | ||
2711 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2712 | struct kobj_attribute *attr, char *buf) | |
2713 | { | |
9a305230 | 2714 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2715 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2716 | } | |
2717 | HSTATE_ATTR_RO(resv_hugepages); | |
2718 | ||
2719 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2720 | struct kobj_attribute *attr, char *buf) | |
2721 | { | |
9a305230 LS |
2722 | struct hstate *h; |
2723 | unsigned long surplus_huge_pages; | |
2724 | int nid; | |
2725 | ||
2726 | h = kobj_to_hstate(kobj, &nid); | |
2727 | if (nid == NUMA_NO_NODE) | |
2728 | surplus_huge_pages = h->surplus_huge_pages; | |
2729 | else | |
2730 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2731 | ||
2732 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2733 | } |
2734 | HSTATE_ATTR_RO(surplus_hugepages); | |
2735 | ||
2736 | static struct attribute *hstate_attrs[] = { | |
2737 | &nr_hugepages_attr.attr, | |
2738 | &nr_overcommit_hugepages_attr.attr, | |
2739 | &free_hugepages_attr.attr, | |
2740 | &resv_hugepages_attr.attr, | |
2741 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2742 | #ifdef CONFIG_NUMA |
2743 | &nr_hugepages_mempolicy_attr.attr, | |
2744 | #endif | |
a3437870 NA |
2745 | NULL, |
2746 | }; | |
2747 | ||
67e5ed96 | 2748 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2749 | .attrs = hstate_attrs, |
2750 | }; | |
2751 | ||
094e9539 JM |
2752 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2753 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2754 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2755 | { |
2756 | int retval; | |
972dc4de | 2757 | int hi = hstate_index(h); |
a3437870 | 2758 | |
9a305230 LS |
2759 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2760 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2761 | return -ENOMEM; |
2762 | ||
9a305230 | 2763 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2764 | if (retval) |
9a305230 | 2765 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2766 | |
2767 | return retval; | |
2768 | } | |
2769 | ||
2770 | static void __init hugetlb_sysfs_init(void) | |
2771 | { | |
2772 | struct hstate *h; | |
2773 | int err; | |
2774 | ||
2775 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2776 | if (!hugepages_kobj) | |
2777 | return; | |
2778 | ||
2779 | for_each_hstate(h) { | |
9a305230 LS |
2780 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2781 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2782 | if (err) |
ffb22af5 | 2783 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2784 | } |
2785 | } | |
2786 | ||
9a305230 LS |
2787 | #ifdef CONFIG_NUMA |
2788 | ||
2789 | /* | |
2790 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2791 | * with node devices in node_devices[] using a parallel array. The array |
2792 | * index of a node device or _hstate == node id. | |
2793 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2794 | * the base kernel, on the hugetlb module. |
2795 | */ | |
2796 | struct node_hstate { | |
2797 | struct kobject *hugepages_kobj; | |
2798 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2799 | }; | |
b4e289a6 | 2800 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2801 | |
2802 | /* | |
10fbcf4c | 2803 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2804 | */ |
2805 | static struct attribute *per_node_hstate_attrs[] = { | |
2806 | &nr_hugepages_attr.attr, | |
2807 | &free_hugepages_attr.attr, | |
2808 | &surplus_hugepages_attr.attr, | |
2809 | NULL, | |
2810 | }; | |
2811 | ||
67e5ed96 | 2812 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
2813 | .attrs = per_node_hstate_attrs, |
2814 | }; | |
2815 | ||
2816 | /* | |
10fbcf4c | 2817 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2818 | * Returns node id via non-NULL nidp. |
2819 | */ | |
2820 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2821 | { | |
2822 | int nid; | |
2823 | ||
2824 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2825 | struct node_hstate *nhs = &node_hstates[nid]; | |
2826 | int i; | |
2827 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2828 | if (nhs->hstate_kobjs[i] == kobj) { | |
2829 | if (nidp) | |
2830 | *nidp = nid; | |
2831 | return &hstates[i]; | |
2832 | } | |
2833 | } | |
2834 | ||
2835 | BUG(); | |
2836 | return NULL; | |
2837 | } | |
2838 | ||
2839 | /* | |
10fbcf4c | 2840 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2841 | * No-op if no hstate attributes attached. |
2842 | */ | |
3cd8b44f | 2843 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2844 | { |
2845 | struct hstate *h; | |
10fbcf4c | 2846 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2847 | |
2848 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2849 | return; /* no hstate attributes */ |
9a305230 | 2850 | |
972dc4de AK |
2851 | for_each_hstate(h) { |
2852 | int idx = hstate_index(h); | |
2853 | if (nhs->hstate_kobjs[idx]) { | |
2854 | kobject_put(nhs->hstate_kobjs[idx]); | |
2855 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2856 | } |
972dc4de | 2857 | } |
9a305230 LS |
2858 | |
2859 | kobject_put(nhs->hugepages_kobj); | |
2860 | nhs->hugepages_kobj = NULL; | |
2861 | } | |
2862 | ||
9a305230 LS |
2863 | |
2864 | /* | |
10fbcf4c | 2865 | * Register hstate attributes for a single node device. |
9a305230 LS |
2866 | * No-op if attributes already registered. |
2867 | */ | |
3cd8b44f | 2868 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2869 | { |
2870 | struct hstate *h; | |
10fbcf4c | 2871 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2872 | int err; |
2873 | ||
2874 | if (nhs->hugepages_kobj) | |
2875 | return; /* already allocated */ | |
2876 | ||
2877 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2878 | &node->dev.kobj); |
9a305230 LS |
2879 | if (!nhs->hugepages_kobj) |
2880 | return; | |
2881 | ||
2882 | for_each_hstate(h) { | |
2883 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2884 | nhs->hstate_kobjs, | |
2885 | &per_node_hstate_attr_group); | |
2886 | if (err) { | |
ffb22af5 AM |
2887 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2888 | h->name, node->dev.id); | |
9a305230 LS |
2889 | hugetlb_unregister_node(node); |
2890 | break; | |
2891 | } | |
2892 | } | |
2893 | } | |
2894 | ||
2895 | /* | |
9b5e5d0f | 2896 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2897 | * devices of nodes that have memory. All on-line nodes should have |
2898 | * registered their associated device by this time. | |
9a305230 | 2899 | */ |
7d9ca000 | 2900 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2901 | { |
2902 | int nid; | |
2903 | ||
8cebfcd0 | 2904 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2905 | struct node *node = node_devices[nid]; |
10fbcf4c | 2906 | if (node->dev.id == nid) |
9a305230 LS |
2907 | hugetlb_register_node(node); |
2908 | } | |
2909 | ||
2910 | /* | |
10fbcf4c | 2911 | * Let the node device driver know we're here so it can |
9a305230 LS |
2912 | * [un]register hstate attributes on node hotplug. |
2913 | */ | |
2914 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2915 | hugetlb_unregister_node); | |
2916 | } | |
2917 | #else /* !CONFIG_NUMA */ | |
2918 | ||
2919 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2920 | { | |
2921 | BUG(); | |
2922 | if (nidp) | |
2923 | *nidp = -1; | |
2924 | return NULL; | |
2925 | } | |
2926 | ||
9a305230 LS |
2927 | static void hugetlb_register_all_nodes(void) { } |
2928 | ||
2929 | #endif | |
2930 | ||
a3437870 NA |
2931 | static int __init hugetlb_init(void) |
2932 | { | |
8382d914 DB |
2933 | int i; |
2934 | ||
457c1b27 | 2935 | if (!hugepages_supported()) |
0ef89d25 | 2936 | return 0; |
a3437870 | 2937 | |
e11bfbfc | 2938 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
2939 | if (default_hstate_size != 0) { |
2940 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
2941 | default_hstate_size, HPAGE_SIZE); | |
2942 | } | |
2943 | ||
e11bfbfc NP |
2944 | default_hstate_size = HPAGE_SIZE; |
2945 | if (!size_to_hstate(default_hstate_size)) | |
2946 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2947 | } |
972dc4de | 2948 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2949 | if (default_hstate_max_huge_pages) { |
2950 | if (!default_hstate.max_huge_pages) | |
2951 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2952 | } | |
a3437870 NA |
2953 | |
2954 | hugetlb_init_hstates(); | |
aa888a74 | 2955 | gather_bootmem_prealloc(); |
a3437870 NA |
2956 | report_hugepages(); |
2957 | ||
2958 | hugetlb_sysfs_init(); | |
9a305230 | 2959 | hugetlb_register_all_nodes(); |
7179e7bf | 2960 | hugetlb_cgroup_file_init(); |
9a305230 | 2961 | |
8382d914 DB |
2962 | #ifdef CONFIG_SMP |
2963 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2964 | #else | |
2965 | num_fault_mutexes = 1; | |
2966 | #endif | |
c672c7f2 | 2967 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
2968 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
2969 | GFP_KERNEL); | |
c672c7f2 | 2970 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2971 | |
2972 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2973 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2974 | return 0; |
2975 | } | |
3e89e1c5 | 2976 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2977 | |
2978 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2979 | void __init hugetlb_bad_size(void) |
2980 | { | |
2981 | parsed_valid_hugepagesz = false; | |
2982 | } | |
2983 | ||
d00181b9 | 2984 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2985 | { |
2986 | struct hstate *h; | |
8faa8b07 AK |
2987 | unsigned long i; |
2988 | ||
a3437870 | 2989 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2990 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2991 | return; |
2992 | } | |
47d38344 | 2993 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2994 | BUG_ON(order == 0); |
47d38344 | 2995 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2996 | h->order = order; |
2997 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2998 | h->nr_huge_pages = 0; |
2999 | h->free_huge_pages = 0; | |
3000 | for (i = 0; i < MAX_NUMNODES; ++i) | |
3001 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3002 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3003 | h->next_nid_to_alloc = first_memory_node; |
3004 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3005 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3006 | huge_page_size(h)/1024); | |
8faa8b07 | 3007 | |
a3437870 NA |
3008 | parsed_hstate = h; |
3009 | } | |
3010 | ||
e11bfbfc | 3011 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
3012 | { |
3013 | unsigned long *mhp; | |
8faa8b07 | 3014 | static unsigned long *last_mhp; |
a3437870 | 3015 | |
9fee021d VT |
3016 | if (!parsed_valid_hugepagesz) { |
3017 | pr_warn("hugepages = %s preceded by " | |
3018 | "an unsupported hugepagesz, ignoring\n", s); | |
3019 | parsed_valid_hugepagesz = true; | |
3020 | return 1; | |
3021 | } | |
a3437870 | 3022 | /* |
47d38344 | 3023 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
3024 | * so this hugepages= parameter goes to the "default hstate". |
3025 | */ | |
9fee021d | 3026 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3027 | mhp = &default_hstate_max_huge_pages; |
3028 | else | |
3029 | mhp = &parsed_hstate->max_huge_pages; | |
3030 | ||
8faa8b07 | 3031 | if (mhp == last_mhp) { |
598d8091 | 3032 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
3033 | return 1; |
3034 | } | |
3035 | ||
a3437870 NA |
3036 | if (sscanf(s, "%lu", mhp) <= 0) |
3037 | *mhp = 0; | |
3038 | ||
8faa8b07 AK |
3039 | /* |
3040 | * Global state is always initialized later in hugetlb_init. | |
3041 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
3042 | * use the bootmem allocator. | |
3043 | */ | |
47d38344 | 3044 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
3045 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3046 | ||
3047 | last_mhp = mhp; | |
3048 | ||
a3437870 NA |
3049 | return 1; |
3050 | } | |
e11bfbfc NP |
3051 | __setup("hugepages=", hugetlb_nrpages_setup); |
3052 | ||
3053 | static int __init hugetlb_default_setup(char *s) | |
3054 | { | |
3055 | default_hstate_size = memparse(s, &s); | |
3056 | return 1; | |
3057 | } | |
3058 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 3059 | |
8a213460 NA |
3060 | static unsigned int cpuset_mems_nr(unsigned int *array) |
3061 | { | |
3062 | int node; | |
3063 | unsigned int nr = 0; | |
3064 | ||
3065 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
3066 | nr += array[node]; | |
3067 | ||
3068 | return nr; | |
3069 | } | |
3070 | ||
3071 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
3072 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3073 | struct ctl_table *table, int write, | |
3074 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 3075 | { |
e5ff2159 | 3076 | struct hstate *h = &default_hstate; |
238d3c13 | 3077 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3078 | int ret; |
e5ff2159 | 3079 | |
457c1b27 | 3080 | if (!hugepages_supported()) |
86613628 | 3081 | return -EOPNOTSUPP; |
457c1b27 | 3082 | |
e5ff2159 AK |
3083 | table->data = &tmp; |
3084 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3085 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3086 | if (ret) | |
3087 | goto out; | |
e5ff2159 | 3088 | |
238d3c13 DR |
3089 | if (write) |
3090 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3091 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3092 | out: |
3093 | return ret; | |
1da177e4 | 3094 | } |
396faf03 | 3095 | |
06808b08 LS |
3096 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
3097 | void __user *buffer, size_t *length, loff_t *ppos) | |
3098 | { | |
3099 | ||
3100 | return hugetlb_sysctl_handler_common(false, table, write, | |
3101 | buffer, length, ppos); | |
3102 | } | |
3103 | ||
3104 | #ifdef CONFIG_NUMA | |
3105 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
3106 | void __user *buffer, size_t *length, loff_t *ppos) | |
3107 | { | |
3108 | return hugetlb_sysctl_handler_common(true, table, write, | |
3109 | buffer, length, ppos); | |
3110 | } | |
3111 | #endif /* CONFIG_NUMA */ | |
3112 | ||
a3d0c6aa | 3113 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 3114 | void __user *buffer, |
a3d0c6aa NA |
3115 | size_t *length, loff_t *ppos) |
3116 | { | |
a5516438 | 3117 | struct hstate *h = &default_hstate; |
e5ff2159 | 3118 | unsigned long tmp; |
08d4a246 | 3119 | int ret; |
e5ff2159 | 3120 | |
457c1b27 | 3121 | if (!hugepages_supported()) |
86613628 | 3122 | return -EOPNOTSUPP; |
457c1b27 | 3123 | |
c033a93c | 3124 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3125 | |
bae7f4ae | 3126 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3127 | return -EINVAL; |
3128 | ||
e5ff2159 AK |
3129 | table->data = &tmp; |
3130 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3131 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3132 | if (ret) | |
3133 | goto out; | |
e5ff2159 AK |
3134 | |
3135 | if (write) { | |
3136 | spin_lock(&hugetlb_lock); | |
3137 | h->nr_overcommit_huge_pages = tmp; | |
3138 | spin_unlock(&hugetlb_lock); | |
3139 | } | |
08d4a246 MH |
3140 | out: |
3141 | return ret; | |
a3d0c6aa NA |
3142 | } |
3143 | ||
1da177e4 LT |
3144 | #endif /* CONFIG_SYSCTL */ |
3145 | ||
e1759c21 | 3146 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3147 | { |
fcb2b0c5 RG |
3148 | struct hstate *h; |
3149 | unsigned long total = 0; | |
3150 | ||
457c1b27 NA |
3151 | if (!hugepages_supported()) |
3152 | return; | |
fcb2b0c5 RG |
3153 | |
3154 | for_each_hstate(h) { | |
3155 | unsigned long count = h->nr_huge_pages; | |
3156 | ||
3157 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3158 | ||
3159 | if (h == &default_hstate) | |
3160 | seq_printf(m, | |
3161 | "HugePages_Total: %5lu\n" | |
3162 | "HugePages_Free: %5lu\n" | |
3163 | "HugePages_Rsvd: %5lu\n" | |
3164 | "HugePages_Surp: %5lu\n" | |
3165 | "Hugepagesize: %8lu kB\n", | |
3166 | count, | |
3167 | h->free_huge_pages, | |
3168 | h->resv_huge_pages, | |
3169 | h->surplus_huge_pages, | |
3170 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3171 | } | |
3172 | ||
3173 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3174 | } |
3175 | ||
3176 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3177 | { | |
a5516438 | 3178 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3179 | if (!hugepages_supported()) |
3180 | return 0; | |
1da177e4 LT |
3181 | return sprintf(buf, |
3182 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3183 | "Node %d HugePages_Free: %5u\n" |
3184 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3185 | nid, h->nr_huge_pages_node[nid], |
3186 | nid, h->free_huge_pages_node[nid], | |
3187 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3188 | } |
3189 | ||
949f7ec5 DR |
3190 | void hugetlb_show_meminfo(void) |
3191 | { | |
3192 | struct hstate *h; | |
3193 | int nid; | |
3194 | ||
457c1b27 NA |
3195 | if (!hugepages_supported()) |
3196 | return; | |
3197 | ||
949f7ec5 DR |
3198 | for_each_node_state(nid, N_MEMORY) |
3199 | for_each_hstate(h) | |
3200 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3201 | nid, | |
3202 | h->nr_huge_pages_node[nid], | |
3203 | h->free_huge_pages_node[nid], | |
3204 | h->surplus_huge_pages_node[nid], | |
3205 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3206 | } | |
3207 | ||
5d317b2b NH |
3208 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3209 | { | |
3210 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3211 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3212 | } | |
3213 | ||
1da177e4 LT |
3214 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3215 | unsigned long hugetlb_total_pages(void) | |
3216 | { | |
d0028588 WL |
3217 | struct hstate *h; |
3218 | unsigned long nr_total_pages = 0; | |
3219 | ||
3220 | for_each_hstate(h) | |
3221 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3222 | return nr_total_pages; | |
1da177e4 | 3223 | } |
1da177e4 | 3224 | |
a5516438 | 3225 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3226 | { |
3227 | int ret = -ENOMEM; | |
3228 | ||
3229 | spin_lock(&hugetlb_lock); | |
3230 | /* | |
3231 | * When cpuset is configured, it breaks the strict hugetlb page | |
3232 | * reservation as the accounting is done on a global variable. Such | |
3233 | * reservation is completely rubbish in the presence of cpuset because | |
3234 | * the reservation is not checked against page availability for the | |
3235 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3236 | * with lack of free htlb page in cpuset that the task is in. | |
3237 | * Attempt to enforce strict accounting with cpuset is almost | |
3238 | * impossible (or too ugly) because cpuset is too fluid that | |
3239 | * task or memory node can be dynamically moved between cpusets. | |
3240 | * | |
3241 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3242 | * undesirable. However, in order to preserve some of the semantics, | |
3243 | * we fall back to check against current free page availability as | |
3244 | * a best attempt and hopefully to minimize the impact of changing | |
3245 | * semantics that cpuset has. | |
3246 | */ | |
3247 | if (delta > 0) { | |
a5516438 | 3248 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3249 | goto out; |
3250 | ||
a5516438 AK |
3251 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3252 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3253 | goto out; |
3254 | } | |
3255 | } | |
3256 | ||
3257 | ret = 0; | |
3258 | if (delta < 0) | |
a5516438 | 3259 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3260 | |
3261 | out: | |
3262 | spin_unlock(&hugetlb_lock); | |
3263 | return ret; | |
3264 | } | |
3265 | ||
84afd99b AW |
3266 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3267 | { | |
f522c3ac | 3268 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3269 | |
3270 | /* | |
3271 | * This new VMA should share its siblings reservation map if present. | |
3272 | * The VMA will only ever have a valid reservation map pointer where | |
3273 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3274 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3275 | * after this open call completes. It is therefore safe to take a |
3276 | * new reference here without additional locking. | |
3277 | */ | |
4e35f483 | 3278 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3279 | kref_get(&resv->refs); |
84afd99b AW |
3280 | } |
3281 | ||
a1e78772 MG |
3282 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3283 | { | |
a5516438 | 3284 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3285 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3286 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3287 | unsigned long reserve, start, end; |
1c5ecae3 | 3288 | long gbl_reserve; |
84afd99b | 3289 | |
4e35f483 JK |
3290 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3291 | return; | |
84afd99b | 3292 | |
4e35f483 JK |
3293 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3294 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3295 | |
4e35f483 | 3296 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3297 | |
4e35f483 JK |
3298 | kref_put(&resv->refs, resv_map_release); |
3299 | ||
3300 | if (reserve) { | |
1c5ecae3 MK |
3301 | /* |
3302 | * Decrement reserve counts. The global reserve count may be | |
3303 | * adjusted if the subpool has a minimum size. | |
3304 | */ | |
3305 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3306 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3307 | } |
a1e78772 MG |
3308 | } |
3309 | ||
31383c68 DW |
3310 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3311 | { | |
3312 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3313 | return -EINVAL; | |
3314 | return 0; | |
3315 | } | |
3316 | ||
05ea8860 DW |
3317 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3318 | { | |
3319 | struct hstate *hstate = hstate_vma(vma); | |
3320 | ||
3321 | return 1UL << huge_page_shift(hstate); | |
3322 | } | |
3323 | ||
1da177e4 LT |
3324 | /* |
3325 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3326 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3327 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3328 | * this far. | |
3329 | */ | |
b3ec9f33 | 3330 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3331 | { |
3332 | BUG(); | |
d0217ac0 | 3333 | return 0; |
1da177e4 LT |
3334 | } |
3335 | ||
eec3636a JC |
3336 | /* |
3337 | * When a new function is introduced to vm_operations_struct and added | |
3338 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3339 | * This is because under System V memory model, mappings created via | |
3340 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3341 | * their original vm_ops are overwritten with shm_vm_ops. | |
3342 | */ | |
f0f37e2f | 3343 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3344 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3345 | .open = hugetlb_vm_op_open, |
a1e78772 | 3346 | .close = hugetlb_vm_op_close, |
31383c68 | 3347 | .split = hugetlb_vm_op_split, |
05ea8860 | 3348 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3349 | }; |
3350 | ||
1e8f889b DG |
3351 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3352 | int writable) | |
63551ae0 DG |
3353 | { |
3354 | pte_t entry; | |
3355 | ||
1e8f889b | 3356 | if (writable) { |
106c992a GS |
3357 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3358 | vma->vm_page_prot))); | |
63551ae0 | 3359 | } else { |
106c992a GS |
3360 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3361 | vma->vm_page_prot)); | |
63551ae0 DG |
3362 | } |
3363 | entry = pte_mkyoung(entry); | |
3364 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3365 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3366 | |
3367 | return entry; | |
3368 | } | |
3369 | ||
1e8f889b DG |
3370 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3371 | unsigned long address, pte_t *ptep) | |
3372 | { | |
3373 | pte_t entry; | |
3374 | ||
106c992a | 3375 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3376 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3377 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3378 | } |
3379 | ||
d5ed7444 | 3380 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3381 | { |
3382 | swp_entry_t swp; | |
3383 | ||
3384 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3385 | return false; |
4a705fef NH |
3386 | swp = pte_to_swp_entry(pte); |
3387 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3388 | return true; |
4a705fef | 3389 | else |
d5ed7444 | 3390 | return false; |
4a705fef NH |
3391 | } |
3392 | ||
3393 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3394 | { | |
3395 | swp_entry_t swp; | |
3396 | ||
3397 | if (huge_pte_none(pte) || pte_present(pte)) | |
3398 | return 0; | |
3399 | swp = pte_to_swp_entry(pte); | |
3400 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3401 | return 1; | |
3402 | else | |
3403 | return 0; | |
3404 | } | |
1e8f889b | 3405 | |
63551ae0 DG |
3406 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3407 | struct vm_area_struct *vma) | |
3408 | { | |
5e41540c | 3409 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3410 | struct page *ptepage; |
1c59827d | 3411 | unsigned long addr; |
1e8f889b | 3412 | int cow; |
a5516438 AK |
3413 | struct hstate *h = hstate_vma(vma); |
3414 | unsigned long sz = huge_page_size(h); | |
c0d0381a | 3415 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 3416 | struct mmu_notifier_range range; |
e8569dd2 | 3417 | int ret = 0; |
1e8f889b DG |
3418 | |
3419 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3420 | |
ac46d4f3 | 3421 | if (cow) { |
7269f999 | 3422 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3423 | vma->vm_start, |
ac46d4f3 JG |
3424 | vma->vm_end); |
3425 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
3426 | } else { |
3427 | /* | |
3428 | * For shared mappings i_mmap_rwsem must be held to call | |
3429 | * huge_pte_alloc, otherwise the returned ptep could go | |
3430 | * away if part of a shared pmd and another thread calls | |
3431 | * huge_pmd_unshare. | |
3432 | */ | |
3433 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 3434 | } |
e8569dd2 | 3435 | |
a5516438 | 3436 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3437 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3438 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3439 | if (!src_pte) |
3440 | continue; | |
a5516438 | 3441 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3442 | if (!dst_pte) { |
3443 | ret = -ENOMEM; | |
3444 | break; | |
3445 | } | |
c5c99429 | 3446 | |
5e41540c MK |
3447 | /* |
3448 | * If the pagetables are shared don't copy or take references. | |
3449 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3450 | * | |
3451 | * However, src could have 'unshared' and dst shares with | |
3452 | * another vma. If dst_pte !none, this implies sharing. | |
3453 | * Check here before taking page table lock, and once again | |
3454 | * after taking the lock below. | |
3455 | */ | |
3456 | dst_entry = huge_ptep_get(dst_pte); | |
3457 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3458 | continue; |
3459 | ||
cb900f41 KS |
3460 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3461 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3462 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3463 | entry = huge_ptep_get(src_pte); |
5e41540c MK |
3464 | dst_entry = huge_ptep_get(dst_pte); |
3465 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3466 | /* | |
3467 | * Skip if src entry none. Also, skip in the | |
3468 | * unlikely case dst entry !none as this implies | |
3469 | * sharing with another vma. | |
3470 | */ | |
4a705fef NH |
3471 | ; |
3472 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3473 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3474 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3475 | ||
3476 | if (is_write_migration_entry(swp_entry) && cow) { | |
3477 | /* | |
3478 | * COW mappings require pages in both | |
3479 | * parent and child to be set to read. | |
3480 | */ | |
3481 | make_migration_entry_read(&swp_entry); | |
3482 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3483 | set_huge_swap_pte_at(src, addr, src_pte, |
3484 | entry, sz); | |
4a705fef | 3485 | } |
e5251fd4 | 3486 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3487 | } else { |
34ee645e | 3488 | if (cow) { |
0f10851e JG |
3489 | /* |
3490 | * No need to notify as we are downgrading page | |
3491 | * table protection not changing it to point | |
3492 | * to a new page. | |
3493 | * | |
ad56b738 | 3494 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3495 | */ |
7f2e9525 | 3496 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3497 | } |
0253d634 | 3498 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3499 | ptepage = pte_page(entry); |
3500 | get_page(ptepage); | |
53f9263b | 3501 | page_dup_rmap(ptepage, true); |
1c59827d | 3502 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3503 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3504 | } |
cb900f41 KS |
3505 | spin_unlock(src_ptl); |
3506 | spin_unlock(dst_ptl); | |
63551ae0 | 3507 | } |
63551ae0 | 3508 | |
e8569dd2 | 3509 | if (cow) |
ac46d4f3 | 3510 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
3511 | else |
3512 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
3513 | |
3514 | return ret; | |
63551ae0 DG |
3515 | } |
3516 | ||
24669e58 AK |
3517 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3518 | unsigned long start, unsigned long end, | |
3519 | struct page *ref_page) | |
63551ae0 DG |
3520 | { |
3521 | struct mm_struct *mm = vma->vm_mm; | |
3522 | unsigned long address; | |
c7546f8f | 3523 | pte_t *ptep; |
63551ae0 | 3524 | pte_t pte; |
cb900f41 | 3525 | spinlock_t *ptl; |
63551ae0 | 3526 | struct page *page; |
a5516438 AK |
3527 | struct hstate *h = hstate_vma(vma); |
3528 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3529 | struct mmu_notifier_range range; |
a5516438 | 3530 | |
63551ae0 | 3531 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3532 | BUG_ON(start & ~huge_page_mask(h)); |
3533 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3534 | |
07e32661 AK |
3535 | /* |
3536 | * This is a hugetlb vma, all the pte entries should point | |
3537 | * to huge page. | |
3538 | */ | |
ed6a7935 | 3539 | tlb_change_page_size(tlb, sz); |
24669e58 | 3540 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3541 | |
3542 | /* | |
3543 | * If sharing possible, alert mmu notifiers of worst case. | |
3544 | */ | |
6f4f13e8 JG |
3545 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3546 | end); | |
ac46d4f3 JG |
3547 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3548 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3549 | address = start; |
569f48b8 | 3550 | for (; address < end; address += sz) { |
7868a208 | 3551 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3552 | if (!ptep) |
c7546f8f DG |
3553 | continue; |
3554 | ||
cb900f41 | 3555 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3556 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3557 | spin_unlock(ptl); | |
dff11abe MK |
3558 | /* |
3559 | * We just unmapped a page of PMDs by clearing a PUD. | |
3560 | * The caller's TLB flush range should cover this area. | |
3561 | */ | |
31d49da5 AK |
3562 | continue; |
3563 | } | |
39dde65c | 3564 | |
6629326b | 3565 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3566 | if (huge_pte_none(pte)) { |
3567 | spin_unlock(ptl); | |
3568 | continue; | |
3569 | } | |
6629326b HD |
3570 | |
3571 | /* | |
9fbc1f63 NH |
3572 | * Migrating hugepage or HWPoisoned hugepage is already |
3573 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3574 | */ |
9fbc1f63 | 3575 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3576 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3577 | spin_unlock(ptl); |
3578 | continue; | |
8c4894c6 | 3579 | } |
6629326b HD |
3580 | |
3581 | page = pte_page(pte); | |
04f2cbe3 MG |
3582 | /* |
3583 | * If a reference page is supplied, it is because a specific | |
3584 | * page is being unmapped, not a range. Ensure the page we | |
3585 | * are about to unmap is the actual page of interest. | |
3586 | */ | |
3587 | if (ref_page) { | |
31d49da5 AK |
3588 | if (page != ref_page) { |
3589 | spin_unlock(ptl); | |
3590 | continue; | |
3591 | } | |
04f2cbe3 MG |
3592 | /* |
3593 | * Mark the VMA as having unmapped its page so that | |
3594 | * future faults in this VMA will fail rather than | |
3595 | * looking like data was lost | |
3596 | */ | |
3597 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3598 | } | |
3599 | ||
c7546f8f | 3600 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3601 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3602 | if (huge_pte_dirty(pte)) |
6649a386 | 3603 | set_page_dirty(page); |
9e81130b | 3604 | |
5d317b2b | 3605 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3606 | page_remove_rmap(page, true); |
31d49da5 | 3607 | |
cb900f41 | 3608 | spin_unlock(ptl); |
e77b0852 | 3609 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3610 | /* |
3611 | * Bail out after unmapping reference page if supplied | |
3612 | */ | |
3613 | if (ref_page) | |
3614 | break; | |
fe1668ae | 3615 | } |
ac46d4f3 | 3616 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 3617 | tlb_end_vma(tlb, vma); |
1da177e4 | 3618 | } |
63551ae0 | 3619 | |
d833352a MG |
3620 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3621 | struct vm_area_struct *vma, unsigned long start, | |
3622 | unsigned long end, struct page *ref_page) | |
3623 | { | |
3624 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3625 | ||
3626 | /* | |
3627 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3628 | * test will fail on a vma being torn down, and not grab a page table | |
3629 | * on its way out. We're lucky that the flag has such an appropriate | |
3630 | * name, and can in fact be safely cleared here. We could clear it | |
3631 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3632 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3633 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3634 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3635 | */ |
3636 | vma->vm_flags &= ~VM_MAYSHARE; | |
3637 | } | |
3638 | ||
502717f4 | 3639 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3640 | unsigned long end, struct page *ref_page) |
502717f4 | 3641 | { |
24669e58 AK |
3642 | struct mm_struct *mm; |
3643 | struct mmu_gather tlb; | |
dff11abe MK |
3644 | unsigned long tlb_start = start; |
3645 | unsigned long tlb_end = end; | |
3646 | ||
3647 | /* | |
3648 | * If shared PMDs were possibly used within this vma range, adjust | |
3649 | * start/end for worst case tlb flushing. | |
3650 | * Note that we can not be sure if PMDs are shared until we try to | |
3651 | * unmap pages. However, we want to make sure TLB flushing covers | |
3652 | * the largest possible range. | |
3653 | */ | |
3654 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
24669e58 AK |
3655 | |
3656 | mm = vma->vm_mm; | |
3657 | ||
dff11abe | 3658 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); |
24669e58 | 3659 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
dff11abe | 3660 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); |
502717f4 KC |
3661 | } |
3662 | ||
04f2cbe3 MG |
3663 | /* |
3664 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3665 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3666 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3667 | * same region. | |
3668 | */ | |
2f4612af DB |
3669 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3670 | struct page *page, unsigned long address) | |
04f2cbe3 | 3671 | { |
7526674d | 3672 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3673 | struct vm_area_struct *iter_vma; |
3674 | struct address_space *mapping; | |
04f2cbe3 MG |
3675 | pgoff_t pgoff; |
3676 | ||
3677 | /* | |
3678 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3679 | * from page cache lookup which is in HPAGE_SIZE units. | |
3680 | */ | |
7526674d | 3681 | address = address & huge_page_mask(h); |
36e4f20a MH |
3682 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3683 | vma->vm_pgoff; | |
93c76a3d | 3684 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3685 | |
4eb2b1dc MG |
3686 | /* |
3687 | * Take the mapping lock for the duration of the table walk. As | |
3688 | * this mapping should be shared between all the VMAs, | |
3689 | * __unmap_hugepage_range() is called as the lock is already held | |
3690 | */ | |
83cde9e8 | 3691 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3692 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3693 | /* Do not unmap the current VMA */ |
3694 | if (iter_vma == vma) | |
3695 | continue; | |
3696 | ||
2f84a899 MG |
3697 | /* |
3698 | * Shared VMAs have their own reserves and do not affect | |
3699 | * MAP_PRIVATE accounting but it is possible that a shared | |
3700 | * VMA is using the same page so check and skip such VMAs. | |
3701 | */ | |
3702 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3703 | continue; | |
3704 | ||
04f2cbe3 MG |
3705 | /* |
3706 | * Unmap the page from other VMAs without their own reserves. | |
3707 | * They get marked to be SIGKILLed if they fault in these | |
3708 | * areas. This is because a future no-page fault on this VMA | |
3709 | * could insert a zeroed page instead of the data existing | |
3710 | * from the time of fork. This would look like data corruption | |
3711 | */ | |
3712 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3713 | unmap_hugepage_range(iter_vma, address, |
3714 | address + huge_page_size(h), page); | |
04f2cbe3 | 3715 | } |
83cde9e8 | 3716 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3717 | } |
3718 | ||
0fe6e20b NH |
3719 | /* |
3720 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3721 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3722 | * cannot race with other handlers or page migration. | |
3723 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3724 | */ |
2b740303 | 3725 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 3726 | unsigned long address, pte_t *ptep, |
3999f52e | 3727 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3728 | { |
3999f52e | 3729 | pte_t pte; |
a5516438 | 3730 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3731 | struct page *old_page, *new_page; |
2b740303 SJ |
3732 | int outside_reserve = 0; |
3733 | vm_fault_t ret = 0; | |
974e6d66 | 3734 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 3735 | struct mmu_notifier_range range; |
1e8f889b | 3736 | |
3999f52e | 3737 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3738 | old_page = pte_page(pte); |
3739 | ||
04f2cbe3 | 3740 | retry_avoidcopy: |
1e8f889b DG |
3741 | /* If no-one else is actually using this page, avoid the copy |
3742 | * and just make the page writable */ | |
37a2140d | 3743 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3744 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 3745 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 3746 | return 0; |
1e8f889b DG |
3747 | } |
3748 | ||
04f2cbe3 MG |
3749 | /* |
3750 | * If the process that created a MAP_PRIVATE mapping is about to | |
3751 | * perform a COW due to a shared page count, attempt to satisfy | |
3752 | * the allocation without using the existing reserves. The pagecache | |
3753 | * page is used to determine if the reserve at this address was | |
3754 | * consumed or not. If reserves were used, a partial faulted mapping | |
3755 | * at the time of fork() could consume its reserves on COW instead | |
3756 | * of the full address range. | |
3757 | */ | |
5944d011 | 3758 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3759 | old_page != pagecache_page) |
3760 | outside_reserve = 1; | |
3761 | ||
09cbfeaf | 3762 | get_page(old_page); |
b76c8cfb | 3763 | |
ad4404a2 DB |
3764 | /* |
3765 | * Drop page table lock as buddy allocator may be called. It will | |
3766 | * be acquired again before returning to the caller, as expected. | |
3767 | */ | |
cb900f41 | 3768 | spin_unlock(ptl); |
5b7a1d40 | 3769 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 3770 | |
2fc39cec | 3771 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3772 | /* |
3773 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3774 | * it is due to references held by a child and an insufficient | |
3775 | * huge page pool. To guarantee the original mappers | |
3776 | * reliability, unmap the page from child processes. The child | |
3777 | * may get SIGKILLed if it later faults. | |
3778 | */ | |
3779 | if (outside_reserve) { | |
09cbfeaf | 3780 | put_page(old_page); |
04f2cbe3 | 3781 | BUG_ON(huge_pte_none(pte)); |
5b7a1d40 | 3782 | unmap_ref_private(mm, vma, old_page, haddr); |
2f4612af DB |
3783 | BUG_ON(huge_pte_none(pte)); |
3784 | spin_lock(ptl); | |
5b7a1d40 | 3785 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
3786 | if (likely(ptep && |
3787 | pte_same(huge_ptep_get(ptep), pte))) | |
3788 | goto retry_avoidcopy; | |
3789 | /* | |
3790 | * race occurs while re-acquiring page table | |
3791 | * lock, and our job is done. | |
3792 | */ | |
3793 | return 0; | |
04f2cbe3 MG |
3794 | } |
3795 | ||
2b740303 | 3796 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 3797 | goto out_release_old; |
1e8f889b DG |
3798 | } |
3799 | ||
0fe6e20b NH |
3800 | /* |
3801 | * When the original hugepage is shared one, it does not have | |
3802 | * anon_vma prepared. | |
3803 | */ | |
44e2aa93 | 3804 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3805 | ret = VM_FAULT_OOM; |
3806 | goto out_release_all; | |
44e2aa93 | 3807 | } |
0fe6e20b | 3808 | |
974e6d66 | 3809 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 3810 | pages_per_huge_page(h)); |
0ed361de | 3811 | __SetPageUptodate(new_page); |
1e8f889b | 3812 | |
7269f999 | 3813 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 3814 | haddr + huge_page_size(h)); |
ac46d4f3 | 3815 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 3816 | |
b76c8cfb | 3817 | /* |
cb900f41 | 3818 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3819 | * before the page tables are altered |
3820 | */ | |
cb900f41 | 3821 | spin_lock(ptl); |
5b7a1d40 | 3822 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 3823 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3824 | ClearPagePrivate(new_page); |
3825 | ||
1e8f889b | 3826 | /* Break COW */ |
5b7a1d40 | 3827 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 3828 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 3829 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 3830 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 3831 | page_remove_rmap(old_page, true); |
5b7a1d40 | 3832 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
cb6acd01 | 3833 | set_page_huge_active(new_page); |
1e8f889b DG |
3834 | /* Make the old page be freed below */ |
3835 | new_page = old_page; | |
3836 | } | |
cb900f41 | 3837 | spin_unlock(ptl); |
ac46d4f3 | 3838 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 3839 | out_release_all: |
5b7a1d40 | 3840 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 3841 | put_page(new_page); |
ad4404a2 | 3842 | out_release_old: |
09cbfeaf | 3843 | put_page(old_page); |
8312034f | 3844 | |
ad4404a2 DB |
3845 | spin_lock(ptl); /* Caller expects lock to be held */ |
3846 | return ret; | |
1e8f889b DG |
3847 | } |
3848 | ||
04f2cbe3 | 3849 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3850 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3851 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3852 | { |
3853 | struct address_space *mapping; | |
e7c4b0bf | 3854 | pgoff_t idx; |
04f2cbe3 MG |
3855 | |
3856 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3857 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3858 | |
3859 | return find_lock_page(mapping, idx); | |
3860 | } | |
3861 | ||
3ae77f43 HD |
3862 | /* |
3863 | * Return whether there is a pagecache page to back given address within VMA. | |
3864 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3865 | */ | |
3866 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3867 | struct vm_area_struct *vma, unsigned long address) |
3868 | { | |
3869 | struct address_space *mapping; | |
3870 | pgoff_t idx; | |
3871 | struct page *page; | |
3872 | ||
3873 | mapping = vma->vm_file->f_mapping; | |
3874 | idx = vma_hugecache_offset(h, vma, address); | |
3875 | ||
3876 | page = find_get_page(mapping, idx); | |
3877 | if (page) | |
3878 | put_page(page); | |
3879 | return page != NULL; | |
3880 | } | |
3881 | ||
ab76ad54 MK |
3882 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3883 | pgoff_t idx) | |
3884 | { | |
3885 | struct inode *inode = mapping->host; | |
3886 | struct hstate *h = hstate_inode(inode); | |
3887 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3888 | ||
3889 | if (err) | |
3890 | return err; | |
3891 | ClearPagePrivate(page); | |
3892 | ||
22146c3c MK |
3893 | /* |
3894 | * set page dirty so that it will not be removed from cache/file | |
3895 | * by non-hugetlbfs specific code paths. | |
3896 | */ | |
3897 | set_page_dirty(page); | |
3898 | ||
ab76ad54 MK |
3899 | spin_lock(&inode->i_lock); |
3900 | inode->i_blocks += blocks_per_huge_page(h); | |
3901 | spin_unlock(&inode->i_lock); | |
3902 | return 0; | |
3903 | } | |
3904 | ||
2b740303 SJ |
3905 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
3906 | struct vm_area_struct *vma, | |
3907 | struct address_space *mapping, pgoff_t idx, | |
3908 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3909 | { |
a5516438 | 3910 | struct hstate *h = hstate_vma(vma); |
2b740303 | 3911 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3912 | int anon_rmap = 0; |
4c887265 | 3913 | unsigned long size; |
4c887265 | 3914 | struct page *page; |
1e8f889b | 3915 | pte_t new_pte; |
cb900f41 | 3916 | spinlock_t *ptl; |
285b8dca | 3917 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 3918 | bool new_page = false; |
4c887265 | 3919 | |
04f2cbe3 MG |
3920 | /* |
3921 | * Currently, we are forced to kill the process in the event the | |
3922 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3923 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3924 | */ |
3925 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3926 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3927 | current->pid); |
04f2cbe3 MG |
3928 | return ret; |
3929 | } | |
3930 | ||
4c887265 | 3931 | /* |
e7c58097 MK |
3932 | * Use page lock to guard against racing truncation |
3933 | * before we get page_table_lock. | |
4c887265 | 3934 | */ |
6bda666a CL |
3935 | retry: |
3936 | page = find_lock_page(mapping, idx); | |
3937 | if (!page) { | |
e7c58097 MK |
3938 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
3939 | if (idx >= size) | |
3940 | goto out; | |
3941 | ||
1a1aad8a MK |
3942 | /* |
3943 | * Check for page in userfault range | |
3944 | */ | |
3945 | if (userfaultfd_missing(vma)) { | |
3946 | u32 hash; | |
3947 | struct vm_fault vmf = { | |
3948 | .vma = vma, | |
285b8dca | 3949 | .address = haddr, |
1a1aad8a MK |
3950 | .flags = flags, |
3951 | /* | |
3952 | * Hard to debug if it ends up being | |
3953 | * used by a callee that assumes | |
3954 | * something about the other | |
3955 | * uninitialized fields... same as in | |
3956 | * memory.c | |
3957 | */ | |
3958 | }; | |
3959 | ||
3960 | /* | |
c0d0381a MK |
3961 | * hugetlb_fault_mutex and i_mmap_rwsem must be |
3962 | * dropped before handling userfault. Reacquire | |
3963 | * after handling fault to make calling code simpler. | |
1a1aad8a | 3964 | */ |
188b04a7 | 3965 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
1a1aad8a | 3966 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 3967 | i_mmap_unlock_read(mapping); |
1a1aad8a | 3968 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); |
c0d0381a | 3969 | i_mmap_lock_read(mapping); |
1a1aad8a MK |
3970 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
3971 | goto out; | |
3972 | } | |
3973 | ||
285b8dca | 3974 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 3975 | if (IS_ERR(page)) { |
4643d67e MK |
3976 | /* |
3977 | * Returning error will result in faulting task being | |
3978 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
3979 | * tasks from racing to fault in the same page which | |
3980 | * could result in false unable to allocate errors. | |
3981 | * Page migration does not take the fault mutex, but | |
3982 | * does a clear then write of pte's under page table | |
3983 | * lock. Page fault code could race with migration, | |
3984 | * notice the clear pte and try to allocate a page | |
3985 | * here. Before returning error, get ptl and make | |
3986 | * sure there really is no pte entry. | |
3987 | */ | |
3988 | ptl = huge_pte_lock(h, mm, ptep); | |
3989 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
3990 | ret = 0; | |
3991 | spin_unlock(ptl); | |
3992 | goto out; | |
3993 | } | |
3994 | spin_unlock(ptl); | |
2b740303 | 3995 | ret = vmf_error(PTR_ERR(page)); |
6bda666a CL |
3996 | goto out; |
3997 | } | |
47ad8475 | 3998 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3999 | __SetPageUptodate(page); |
cb6acd01 | 4000 | new_page = true; |
ac9b9c66 | 4001 | |
f83a275d | 4002 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4003 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4004 | if (err) { |
4005 | put_page(page); | |
6bda666a CL |
4006 | if (err == -EEXIST) |
4007 | goto retry; | |
4008 | goto out; | |
4009 | } | |
23be7468 | 4010 | } else { |
6bda666a | 4011 | lock_page(page); |
0fe6e20b NH |
4012 | if (unlikely(anon_vma_prepare(vma))) { |
4013 | ret = VM_FAULT_OOM; | |
4014 | goto backout_unlocked; | |
4015 | } | |
409eb8c2 | 4016 | anon_rmap = 1; |
23be7468 | 4017 | } |
0fe6e20b | 4018 | } else { |
998b4382 NH |
4019 | /* |
4020 | * If memory error occurs between mmap() and fault, some process | |
4021 | * don't have hwpoisoned swap entry for errored virtual address. | |
4022 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4023 | */ | |
4024 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 4025 | ret = VM_FAULT_HWPOISON | |
972dc4de | 4026 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4027 | goto backout_unlocked; |
4028 | } | |
6bda666a | 4029 | } |
1e8f889b | 4030 | |
57303d80 AW |
4031 | /* |
4032 | * If we are going to COW a private mapping later, we examine the | |
4033 | * pending reservations for this page now. This will ensure that | |
4034 | * any allocations necessary to record that reservation occur outside | |
4035 | * the spinlock. | |
4036 | */ | |
5e911373 | 4037 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4038 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4039 | ret = VM_FAULT_OOM; |
4040 | goto backout_unlocked; | |
4041 | } | |
5e911373 | 4042 | /* Just decrements count, does not deallocate */ |
285b8dca | 4043 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4044 | } |
57303d80 | 4045 | |
8bea8052 | 4046 | ptl = huge_pte_lock(h, mm, ptep); |
e7c58097 MK |
4047 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4048 | if (idx >= size) | |
4049 | goto backout; | |
4c887265 | 4050 | |
83c54070 | 4051 | ret = 0; |
7f2e9525 | 4052 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4053 | goto backout; |
4054 | ||
07443a85 JK |
4055 | if (anon_rmap) { |
4056 | ClearPagePrivate(page); | |
285b8dca | 4057 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4058 | } else |
53f9263b | 4059 | page_dup_rmap(page, true); |
1e8f889b DG |
4060 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4061 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4062 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4063 | |
5d317b2b | 4064 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4065 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4066 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4067 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4068 | } |
4069 | ||
cb900f41 | 4070 | spin_unlock(ptl); |
cb6acd01 MK |
4071 | |
4072 | /* | |
4073 | * Only make newly allocated pages active. Existing pages found | |
4074 | * in the pagecache could be !page_huge_active() if they have been | |
4075 | * isolated for migration. | |
4076 | */ | |
4077 | if (new_page) | |
4078 | set_page_huge_active(page); | |
4079 | ||
4c887265 AL |
4080 | unlock_page(page); |
4081 | out: | |
ac9b9c66 | 4082 | return ret; |
4c887265 AL |
4083 | |
4084 | backout: | |
cb900f41 | 4085 | spin_unlock(ptl); |
2b26736c | 4086 | backout_unlocked: |
4c887265 | 4087 | unlock_page(page); |
285b8dca | 4088 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4089 | put_page(page); |
4090 | goto out; | |
ac9b9c66 HD |
4091 | } |
4092 | ||
8382d914 | 4093 | #ifdef CONFIG_SMP |
188b04a7 | 4094 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4095 | { |
4096 | unsigned long key[2]; | |
4097 | u32 hash; | |
4098 | ||
1b426bac MK |
4099 | key[0] = (unsigned long) mapping; |
4100 | key[1] = idx; | |
8382d914 | 4101 | |
55254636 | 4102 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4103 | |
4104 | return hash & (num_fault_mutexes - 1); | |
4105 | } | |
4106 | #else | |
4107 | /* | |
4108 | * For uniprocesor systems we always use a single mutex, so just | |
4109 | * return 0 and avoid the hashing overhead. | |
4110 | */ | |
188b04a7 | 4111 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4112 | { |
4113 | return 0; | |
4114 | } | |
4115 | #endif | |
4116 | ||
2b740303 | 4117 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4118 | unsigned long address, unsigned int flags) |
86e5216f | 4119 | { |
8382d914 | 4120 | pte_t *ptep, entry; |
cb900f41 | 4121 | spinlock_t *ptl; |
2b740303 | 4122 | vm_fault_t ret; |
8382d914 DB |
4123 | u32 hash; |
4124 | pgoff_t idx; | |
0fe6e20b | 4125 | struct page *page = NULL; |
57303d80 | 4126 | struct page *pagecache_page = NULL; |
a5516438 | 4127 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4128 | struct address_space *mapping; |
0f792cf9 | 4129 | int need_wait_lock = 0; |
285b8dca | 4130 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4131 | |
285b8dca | 4132 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 4133 | if (ptep) { |
c0d0381a MK |
4134 | /* |
4135 | * Since we hold no locks, ptep could be stale. That is | |
4136 | * OK as we are only making decisions based on content and | |
4137 | * not actually modifying content here. | |
4138 | */ | |
fd6a03ed | 4139 | entry = huge_ptep_get(ptep); |
290408d4 | 4140 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4141 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4142 | return 0; |
4143 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4144 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4145 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
ddeaab32 MK |
4146 | } else { |
4147 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4148 | if (!ptep) | |
4149 | return VM_FAULT_OOM; | |
fd6a03ed NH |
4150 | } |
4151 | ||
c0d0381a MK |
4152 | /* |
4153 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
4154 | * until finished with ptep. This prevents huge_pmd_unshare from | |
4155 | * being called elsewhere and making the ptep no longer valid. | |
4156 | * | |
4157 | * ptep could have already be assigned via huge_pte_offset. That | |
4158 | * is OK, as huge_pte_alloc will return the same value unless | |
4159 | * something has changed. | |
4160 | */ | |
8382d914 | 4161 | mapping = vma->vm_file->f_mapping; |
c0d0381a MK |
4162 | i_mmap_lock_read(mapping); |
4163 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4164 | if (!ptep) { | |
4165 | i_mmap_unlock_read(mapping); | |
4166 | return VM_FAULT_OOM; | |
4167 | } | |
8382d914 | 4168 | |
3935baa9 DG |
4169 | /* |
4170 | * Serialize hugepage allocation and instantiation, so that we don't | |
4171 | * get spurious allocation failures if two CPUs race to instantiate | |
4172 | * the same page in the page cache. | |
4173 | */ | |
c0d0381a | 4174 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 4175 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4176 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4177 | |
7f2e9525 GS |
4178 | entry = huge_ptep_get(ptep); |
4179 | if (huge_pte_none(entry)) { | |
8382d914 | 4180 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4181 | goto out_mutex; |
3935baa9 | 4182 | } |
86e5216f | 4183 | |
83c54070 | 4184 | ret = 0; |
1e8f889b | 4185 | |
0f792cf9 NH |
4186 | /* |
4187 | * entry could be a migration/hwpoison entry at this point, so this | |
4188 | * check prevents the kernel from going below assuming that we have | |
4189 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
4190 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
4191 | * handle it. | |
4192 | */ | |
4193 | if (!pte_present(entry)) | |
4194 | goto out_mutex; | |
4195 | ||
57303d80 AW |
4196 | /* |
4197 | * If we are going to COW the mapping later, we examine the pending | |
4198 | * reservations for this page now. This will ensure that any | |
4199 | * allocations necessary to record that reservation occur outside the | |
4200 | * spinlock. For private mappings, we also lookup the pagecache | |
4201 | * page now as it is used to determine if a reservation has been | |
4202 | * consumed. | |
4203 | */ | |
106c992a | 4204 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4205 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4206 | ret = VM_FAULT_OOM; |
b4d1d99f | 4207 | goto out_mutex; |
2b26736c | 4208 | } |
5e911373 | 4209 | /* Just decrements count, does not deallocate */ |
285b8dca | 4210 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4211 | |
f83a275d | 4212 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4213 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4214 | vma, haddr); |
57303d80 AW |
4215 | } |
4216 | ||
0f792cf9 NH |
4217 | ptl = huge_pte_lock(h, mm, ptep); |
4218 | ||
4219 | /* Check for a racing update before calling hugetlb_cow */ | |
4220 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4221 | goto out_ptl; | |
4222 | ||
56c9cfb1 NH |
4223 | /* |
4224 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4225 | * pagecache_page, so here we need take the former one | |
4226 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4227 | */ |
4228 | page = pte_page(entry); | |
4229 | if (page != pagecache_page) | |
0f792cf9 NH |
4230 | if (!trylock_page(page)) { |
4231 | need_wait_lock = 1; | |
4232 | goto out_ptl; | |
4233 | } | |
b4d1d99f | 4234 | |
0f792cf9 | 4235 | get_page(page); |
b4d1d99f | 4236 | |
788c7df4 | 4237 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4238 | if (!huge_pte_write(entry)) { |
974e6d66 | 4239 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4240 | pagecache_page, ptl); |
0f792cf9 | 4241 | goto out_put_page; |
b4d1d99f | 4242 | } |
106c992a | 4243 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4244 | } |
4245 | entry = pte_mkyoung(entry); | |
285b8dca | 4246 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4247 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4248 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4249 | out_put_page: |
4250 | if (page != pagecache_page) | |
4251 | unlock_page(page); | |
4252 | put_page(page); | |
cb900f41 KS |
4253 | out_ptl: |
4254 | spin_unlock(ptl); | |
57303d80 AW |
4255 | |
4256 | if (pagecache_page) { | |
4257 | unlock_page(pagecache_page); | |
4258 | put_page(pagecache_page); | |
4259 | } | |
b4d1d99f | 4260 | out_mutex: |
c672c7f2 | 4261 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4262 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
4263 | /* |
4264 | * Generally it's safe to hold refcount during waiting page lock. But | |
4265 | * here we just wait to defer the next page fault to avoid busy loop and | |
4266 | * the page is not used after unlocked before returning from the current | |
4267 | * page fault. So we are safe from accessing freed page, even if we wait | |
4268 | * here without taking refcount. | |
4269 | */ | |
4270 | if (need_wait_lock) | |
4271 | wait_on_page_locked(page); | |
1e8f889b | 4272 | return ret; |
86e5216f AL |
4273 | } |
4274 | ||
8fb5debc MK |
4275 | /* |
4276 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4277 | * modifications for huge pages. | |
4278 | */ | |
4279 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4280 | pte_t *dst_pte, | |
4281 | struct vm_area_struct *dst_vma, | |
4282 | unsigned long dst_addr, | |
4283 | unsigned long src_addr, | |
4284 | struct page **pagep) | |
4285 | { | |
1e392147 AA |
4286 | struct address_space *mapping; |
4287 | pgoff_t idx; | |
4288 | unsigned long size; | |
1c9e8def | 4289 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4290 | struct hstate *h = hstate_vma(dst_vma); |
4291 | pte_t _dst_pte; | |
4292 | spinlock_t *ptl; | |
4293 | int ret; | |
4294 | struct page *page; | |
4295 | ||
4296 | if (!*pagep) { | |
4297 | ret = -ENOMEM; | |
4298 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4299 | if (IS_ERR(page)) | |
4300 | goto out; | |
4301 | ||
4302 | ret = copy_huge_page_from_user(page, | |
4303 | (const void __user *) src_addr, | |
810a56b9 | 4304 | pages_per_huge_page(h), false); |
8fb5debc MK |
4305 | |
4306 | /* fallback to copy_from_user outside mmap_sem */ | |
4307 | if (unlikely(ret)) { | |
9e368259 | 4308 | ret = -ENOENT; |
8fb5debc MK |
4309 | *pagep = page; |
4310 | /* don't free the page */ | |
4311 | goto out; | |
4312 | } | |
4313 | } else { | |
4314 | page = *pagep; | |
4315 | *pagep = NULL; | |
4316 | } | |
4317 | ||
4318 | /* | |
4319 | * The memory barrier inside __SetPageUptodate makes sure that | |
4320 | * preceding stores to the page contents become visible before | |
4321 | * the set_pte_at() write. | |
4322 | */ | |
4323 | __SetPageUptodate(page); | |
8fb5debc | 4324 | |
1e392147 AA |
4325 | mapping = dst_vma->vm_file->f_mapping; |
4326 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4327 | ||
1c9e8def MK |
4328 | /* |
4329 | * If shared, add to page cache | |
4330 | */ | |
4331 | if (vm_shared) { | |
1e392147 AA |
4332 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4333 | ret = -EFAULT; | |
4334 | if (idx >= size) | |
4335 | goto out_release_nounlock; | |
1c9e8def | 4336 | |
1e392147 AA |
4337 | /* |
4338 | * Serialization between remove_inode_hugepages() and | |
4339 | * huge_add_to_page_cache() below happens through the | |
4340 | * hugetlb_fault_mutex_table that here must be hold by | |
4341 | * the caller. | |
4342 | */ | |
1c9e8def MK |
4343 | ret = huge_add_to_page_cache(page, mapping, idx); |
4344 | if (ret) | |
4345 | goto out_release_nounlock; | |
4346 | } | |
4347 | ||
8fb5debc MK |
4348 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4349 | spin_lock(ptl); | |
4350 | ||
1e392147 AA |
4351 | /* |
4352 | * Recheck the i_size after holding PT lock to make sure not | |
4353 | * to leave any page mapped (as page_mapped()) beyond the end | |
4354 | * of the i_size (remove_inode_hugepages() is strict about | |
4355 | * enforcing that). If we bail out here, we'll also leave a | |
4356 | * page in the radix tree in the vm_shared case beyond the end | |
4357 | * of the i_size, but remove_inode_hugepages() will take care | |
4358 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4359 | */ | |
4360 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4361 | ret = -EFAULT; | |
4362 | if (idx >= size) | |
4363 | goto out_release_unlock; | |
4364 | ||
8fb5debc MK |
4365 | ret = -EEXIST; |
4366 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4367 | goto out_release_unlock; | |
4368 | ||
1c9e8def MK |
4369 | if (vm_shared) { |
4370 | page_dup_rmap(page, true); | |
4371 | } else { | |
4372 | ClearPagePrivate(page); | |
4373 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4374 | } | |
8fb5debc MK |
4375 | |
4376 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4377 | if (dst_vma->vm_flags & VM_WRITE) | |
4378 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4379 | _dst_pte = pte_mkyoung(_dst_pte); | |
4380 | ||
4381 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4382 | ||
4383 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4384 | dst_vma->vm_flags & VM_WRITE); | |
4385 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4386 | ||
4387 | /* No need to invalidate - it was non-present before */ | |
4388 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4389 | ||
4390 | spin_unlock(ptl); | |
cb6acd01 | 4391 | set_page_huge_active(page); |
1c9e8def MK |
4392 | if (vm_shared) |
4393 | unlock_page(page); | |
8fb5debc MK |
4394 | ret = 0; |
4395 | out: | |
4396 | return ret; | |
4397 | out_release_unlock: | |
4398 | spin_unlock(ptl); | |
1c9e8def MK |
4399 | if (vm_shared) |
4400 | unlock_page(page); | |
5af10dfd | 4401 | out_release_nounlock: |
8fb5debc MK |
4402 | put_page(page); |
4403 | goto out; | |
4404 | } | |
4405 | ||
28a35716 ML |
4406 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4407 | struct page **pages, struct vm_area_struct **vmas, | |
4408 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 4409 | long i, unsigned int flags, int *locked) |
63551ae0 | 4410 | { |
d5d4b0aa KC |
4411 | unsigned long pfn_offset; |
4412 | unsigned long vaddr = *position; | |
28a35716 | 4413 | unsigned long remainder = *nr_pages; |
a5516438 | 4414 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4415 | int err = -EFAULT; |
63551ae0 | 4416 | |
63551ae0 | 4417 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4418 | pte_t *pte; |
cb900f41 | 4419 | spinlock_t *ptl = NULL; |
2a15efc9 | 4420 | int absent; |
4c887265 | 4421 | struct page *page; |
63551ae0 | 4422 | |
02057967 DR |
4423 | /* |
4424 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4425 | * potentially allocating memory. | |
4426 | */ | |
fa45f116 | 4427 | if (fatal_signal_pending(current)) { |
02057967 DR |
4428 | remainder = 0; |
4429 | break; | |
4430 | } | |
4431 | ||
4c887265 AL |
4432 | /* |
4433 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4434 | * each hugepage. We have to make sure we get the |
4c887265 | 4435 | * first, for the page indexing below to work. |
cb900f41 KS |
4436 | * |
4437 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4438 | */ |
7868a208 PA |
4439 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4440 | huge_page_size(h)); | |
cb900f41 KS |
4441 | if (pte) |
4442 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4443 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4444 | ||
4445 | /* | |
4446 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4447 | * an error where there's an empty slot with no huge pagecache |
4448 | * to back it. This way, we avoid allocating a hugepage, and | |
4449 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4450 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4451 | */ |
3ae77f43 HD |
4452 | if (absent && (flags & FOLL_DUMP) && |
4453 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4454 | if (pte) |
4455 | spin_unlock(ptl); | |
2a15efc9 HD |
4456 | remainder = 0; |
4457 | break; | |
4458 | } | |
63551ae0 | 4459 | |
9cc3a5bd NH |
4460 | /* |
4461 | * We need call hugetlb_fault for both hugepages under migration | |
4462 | * (in which case hugetlb_fault waits for the migration,) and | |
4463 | * hwpoisoned hugepages (in which case we need to prevent the | |
4464 | * caller from accessing to them.) In order to do this, we use | |
4465 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4466 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4467 | * both cases, and because we can't follow correct pages | |
4468 | * directly from any kind of swap entries. | |
4469 | */ | |
4470 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4471 | ((flags & FOLL_WRITE) && |
4472 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4473 | vm_fault_t ret; |
87ffc118 | 4474 | unsigned int fault_flags = 0; |
63551ae0 | 4475 | |
cb900f41 KS |
4476 | if (pte) |
4477 | spin_unlock(ptl); | |
87ffc118 AA |
4478 | if (flags & FOLL_WRITE) |
4479 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 4480 | if (locked) |
71335f37 PX |
4481 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
4482 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
4483 | if (flags & FOLL_NOWAIT) |
4484 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4485 | FAULT_FLAG_RETRY_NOWAIT; | |
4486 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
4487 | /* |
4488 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
4489 | * FAULT_FLAG_TRIED can co-exist | |
4490 | */ | |
87ffc118 AA |
4491 | fault_flags |= FAULT_FLAG_TRIED; |
4492 | } | |
4493 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4494 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4495 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4496 | remainder = 0; |
4497 | break; | |
4498 | } | |
4499 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 4500 | if (locked && |
1ac25013 | 4501 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 4502 | *locked = 0; |
87ffc118 AA |
4503 | *nr_pages = 0; |
4504 | /* | |
4505 | * VM_FAULT_RETRY must not return an | |
4506 | * error, it will return zero | |
4507 | * instead. | |
4508 | * | |
4509 | * No need to update "position" as the | |
4510 | * caller will not check it after | |
4511 | * *nr_pages is set to 0. | |
4512 | */ | |
4513 | return i; | |
4514 | } | |
4515 | continue; | |
4c887265 AL |
4516 | } |
4517 | ||
a5516438 | 4518 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4519 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 4520 | |
acbfb087 ZL |
4521 | /* |
4522 | * If subpage information not requested, update counters | |
4523 | * and skip the same_page loop below. | |
4524 | */ | |
4525 | if (!pages && !vmas && !pfn_offset && | |
4526 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4527 | (remainder >= pages_per_huge_page(h))) { | |
4528 | vaddr += huge_page_size(h); | |
4529 | remainder -= pages_per_huge_page(h); | |
4530 | i += pages_per_huge_page(h); | |
4531 | spin_unlock(ptl); | |
4532 | continue; | |
4533 | } | |
4534 | ||
d5d4b0aa | 4535 | same_page: |
d6692183 | 4536 | if (pages) { |
2a15efc9 | 4537 | pages[i] = mem_map_offset(page, pfn_offset); |
3faa52c0 JH |
4538 | /* |
4539 | * try_grab_page() should always succeed here, because: | |
4540 | * a) we hold the ptl lock, and b) we've just checked | |
4541 | * that the huge page is present in the page tables. If | |
4542 | * the huge page is present, then the tail pages must | |
4543 | * also be present. The ptl prevents the head page and | |
4544 | * tail pages from being rearranged in any way. So this | |
4545 | * page must be available at this point, unless the page | |
4546 | * refcount overflowed: | |
4547 | */ | |
4548 | if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { | |
4549 | spin_unlock(ptl); | |
4550 | remainder = 0; | |
4551 | err = -ENOMEM; | |
4552 | break; | |
4553 | } | |
d6692183 | 4554 | } |
63551ae0 DG |
4555 | |
4556 | if (vmas) | |
4557 | vmas[i] = vma; | |
4558 | ||
4559 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4560 | ++pfn_offset; |
63551ae0 DG |
4561 | --remainder; |
4562 | ++i; | |
d5d4b0aa | 4563 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4564 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4565 | /* |
4566 | * We use pfn_offset to avoid touching the pageframes | |
4567 | * of this compound page. | |
4568 | */ | |
4569 | goto same_page; | |
4570 | } | |
cb900f41 | 4571 | spin_unlock(ptl); |
63551ae0 | 4572 | } |
28a35716 | 4573 | *nr_pages = remainder; |
87ffc118 AA |
4574 | /* |
4575 | * setting position is actually required only if remainder is | |
4576 | * not zero but it's faster not to add a "if (remainder)" | |
4577 | * branch. | |
4578 | */ | |
63551ae0 DG |
4579 | *position = vaddr; |
4580 | ||
2be7cfed | 4581 | return i ? i : err; |
63551ae0 | 4582 | } |
8f860591 | 4583 | |
5491ae7b AK |
4584 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4585 | /* | |
4586 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4587 | * implement this. | |
4588 | */ | |
4589 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4590 | #endif | |
4591 | ||
7da4d641 | 4592 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4593 | unsigned long address, unsigned long end, pgprot_t newprot) |
4594 | { | |
4595 | struct mm_struct *mm = vma->vm_mm; | |
4596 | unsigned long start = address; | |
4597 | pte_t *ptep; | |
4598 | pte_t pte; | |
a5516438 | 4599 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4600 | unsigned long pages = 0; |
dff11abe | 4601 | bool shared_pmd = false; |
ac46d4f3 | 4602 | struct mmu_notifier_range range; |
dff11abe MK |
4603 | |
4604 | /* | |
4605 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 4606 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
4607 | * range if PMD sharing is possible. |
4608 | */ | |
7269f999 JG |
4609 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
4610 | 0, vma, mm, start, end); | |
ac46d4f3 | 4611 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
4612 | |
4613 | BUG_ON(address >= end); | |
ac46d4f3 | 4614 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 4615 | |
ac46d4f3 | 4616 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 4617 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4618 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4619 | spinlock_t *ptl; |
7868a208 | 4620 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4621 | if (!ptep) |
4622 | continue; | |
cb900f41 | 4623 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4624 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4625 | pages++; | |
cb900f41 | 4626 | spin_unlock(ptl); |
dff11abe | 4627 | shared_pmd = true; |
39dde65c | 4628 | continue; |
7da4d641 | 4629 | } |
a8bda28d NH |
4630 | pte = huge_ptep_get(ptep); |
4631 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4632 | spin_unlock(ptl); | |
4633 | continue; | |
4634 | } | |
4635 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4636 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4637 | ||
4638 | if (is_write_migration_entry(entry)) { | |
4639 | pte_t newpte; | |
4640 | ||
4641 | make_migration_entry_read(&entry); | |
4642 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4643 | set_huge_swap_pte_at(mm, address, ptep, |
4644 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4645 | pages++; |
4646 | } | |
4647 | spin_unlock(ptl); | |
4648 | continue; | |
4649 | } | |
4650 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
4651 | pte_t old_pte; |
4652 | ||
4653 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
4654 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 4655 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 4656 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 4657 | pages++; |
8f860591 | 4658 | } |
cb900f41 | 4659 | spin_unlock(ptl); |
8f860591 | 4660 | } |
d833352a | 4661 | /* |
c8c06efa | 4662 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4663 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4664 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
4665 | * and that page table be reused and filled with junk. If we actually |
4666 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 4667 | */ |
dff11abe | 4668 | if (shared_pmd) |
ac46d4f3 | 4669 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
4670 | else |
4671 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
4672 | /* |
4673 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4674 | * page table protection not changing it to point to a new page. | |
4675 | * | |
ad56b738 | 4676 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4677 | */ |
83cde9e8 | 4678 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 4679 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
4680 | |
4681 | return pages << h->order; | |
8f860591 ZY |
4682 | } |
4683 | ||
a1e78772 MG |
4684 | int hugetlb_reserve_pages(struct inode *inode, |
4685 | long from, long to, | |
5a6fe125 | 4686 | struct vm_area_struct *vma, |
ca16d140 | 4687 | vm_flags_t vm_flags) |
e4e574b7 | 4688 | { |
17c9d12e | 4689 | long ret, chg; |
a5516438 | 4690 | struct hstate *h = hstate_inode(inode); |
90481622 | 4691 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4692 | struct resv_map *resv_map; |
1c5ecae3 | 4693 | long gbl_reserve; |
e4e574b7 | 4694 | |
63489f8e MK |
4695 | /* This should never happen */ |
4696 | if (from > to) { | |
4697 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
4698 | return -EINVAL; | |
4699 | } | |
4700 | ||
17c9d12e MG |
4701 | /* |
4702 | * Only apply hugepage reservation if asked. At fault time, an | |
4703 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4704 | * without using reserves |
17c9d12e | 4705 | */ |
ca16d140 | 4706 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4707 | return 0; |
4708 | ||
a1e78772 MG |
4709 | /* |
4710 | * Shared mappings base their reservation on the number of pages that | |
4711 | * are already allocated on behalf of the file. Private mappings need | |
4712 | * to reserve the full area even if read-only as mprotect() may be | |
4713 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4714 | */ | |
9119a41e | 4715 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
4716 | /* |
4717 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
4718 | * called for inodes for which resv_maps were created (see | |
4719 | * hugetlbfs_get_inode). | |
4720 | */ | |
4e35f483 | 4721 | resv_map = inode_resv_map(inode); |
9119a41e | 4722 | |
1406ec9b | 4723 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4724 | |
4725 | } else { | |
4726 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4727 | if (!resv_map) |
4728 | return -ENOMEM; | |
4729 | ||
a1e78772 | 4730 | chg = to - from; |
84afd99b | 4731 | |
17c9d12e MG |
4732 | set_vma_resv_map(vma, resv_map); |
4733 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4734 | } | |
4735 | ||
c50ac050 DH |
4736 | if (chg < 0) { |
4737 | ret = chg; | |
4738 | goto out_err; | |
4739 | } | |
8a630112 | 4740 | |
1c5ecae3 MK |
4741 | /* |
4742 | * There must be enough pages in the subpool for the mapping. If | |
4743 | * the subpool has a minimum size, there may be some global | |
4744 | * reservations already in place (gbl_reserve). | |
4745 | */ | |
4746 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4747 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4748 | ret = -ENOSPC; |
4749 | goto out_err; | |
4750 | } | |
5a6fe125 MG |
4751 | |
4752 | /* | |
17c9d12e | 4753 | * Check enough hugepages are available for the reservation. |
90481622 | 4754 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4755 | */ |
1c5ecae3 | 4756 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4757 | if (ret < 0) { |
1c5ecae3 MK |
4758 | /* put back original number of pages, chg */ |
4759 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4760 | goto out_err; |
68842c9b | 4761 | } |
17c9d12e MG |
4762 | |
4763 | /* | |
4764 | * Account for the reservations made. Shared mappings record regions | |
4765 | * that have reservations as they are shared by multiple VMAs. | |
4766 | * When the last VMA disappears, the region map says how much | |
4767 | * the reservation was and the page cache tells how much of | |
4768 | * the reservation was consumed. Private mappings are per-VMA and | |
4769 | * only the consumed reservations are tracked. When the VMA | |
4770 | * disappears, the original reservation is the VMA size and the | |
4771 | * consumed reservations are stored in the map. Hence, nothing | |
4772 | * else has to be done for private mappings here | |
4773 | */ | |
33039678 MK |
4774 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4775 | long add = region_add(resv_map, from, to); | |
4776 | ||
4777 | if (unlikely(chg > add)) { | |
4778 | /* | |
4779 | * pages in this range were added to the reserve | |
4780 | * map between region_chg and region_add. This | |
4781 | * indicates a race with alloc_huge_page. Adjust | |
4782 | * the subpool and reserve counts modified above | |
4783 | * based on the difference. | |
4784 | */ | |
4785 | long rsv_adjust; | |
4786 | ||
4787 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4788 | chg - add); | |
4789 | hugetlb_acct_memory(h, -rsv_adjust); | |
4790 | } | |
4791 | } | |
a43a8c39 | 4792 | return 0; |
c50ac050 | 4793 | out_err: |
5e911373 | 4794 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4795 | /* Don't call region_abort if region_chg failed */ |
4796 | if (chg >= 0) | |
4797 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4798 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4799 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4800 | return ret; |
a43a8c39 KC |
4801 | } |
4802 | ||
b5cec28d MK |
4803 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4804 | long freed) | |
a43a8c39 | 4805 | { |
a5516438 | 4806 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4807 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4808 | long chg = 0; |
90481622 | 4809 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4810 | long gbl_reserve; |
45c682a6 | 4811 | |
f27a5136 MK |
4812 | /* |
4813 | * Since this routine can be called in the evict inode path for all | |
4814 | * hugetlbfs inodes, resv_map could be NULL. | |
4815 | */ | |
b5cec28d MK |
4816 | if (resv_map) { |
4817 | chg = region_del(resv_map, start, end); | |
4818 | /* | |
4819 | * region_del() can fail in the rare case where a region | |
4820 | * must be split and another region descriptor can not be | |
4821 | * allocated. If end == LONG_MAX, it will not fail. | |
4822 | */ | |
4823 | if (chg < 0) | |
4824 | return chg; | |
4825 | } | |
4826 | ||
45c682a6 | 4827 | spin_lock(&inode->i_lock); |
e4c6f8be | 4828 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4829 | spin_unlock(&inode->i_lock); |
4830 | ||
1c5ecae3 MK |
4831 | /* |
4832 | * If the subpool has a minimum size, the number of global | |
4833 | * reservations to be released may be adjusted. | |
4834 | */ | |
4835 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4836 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4837 | |
4838 | return 0; | |
a43a8c39 | 4839 | } |
93f70f90 | 4840 | |
3212b535 SC |
4841 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4842 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4843 | struct vm_area_struct *vma, | |
4844 | unsigned long addr, pgoff_t idx) | |
4845 | { | |
4846 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4847 | svma->vm_start; | |
4848 | unsigned long sbase = saddr & PUD_MASK; | |
4849 | unsigned long s_end = sbase + PUD_SIZE; | |
4850 | ||
4851 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4852 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4853 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4854 | |
4855 | /* | |
4856 | * match the virtual addresses, permission and the alignment of the | |
4857 | * page table page. | |
4858 | */ | |
4859 | if (pmd_index(addr) != pmd_index(saddr) || | |
4860 | vm_flags != svm_flags || | |
4861 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4862 | return 0; | |
4863 | ||
4864 | return saddr; | |
4865 | } | |
4866 | ||
31aafb45 | 4867 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4868 | { |
4869 | unsigned long base = addr & PUD_MASK; | |
4870 | unsigned long end = base + PUD_SIZE; | |
4871 | ||
4872 | /* | |
4873 | * check on proper vm_flags and page table alignment | |
4874 | */ | |
017b1660 | 4875 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
4876 | return true; |
4877 | return false; | |
3212b535 SC |
4878 | } |
4879 | ||
017b1660 MK |
4880 | /* |
4881 | * Determine if start,end range within vma could be mapped by shared pmd. | |
4882 | * If yes, adjust start and end to cover range associated with possible | |
4883 | * shared pmd mappings. | |
4884 | */ | |
4885 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
4886 | unsigned long *start, unsigned long *end) | |
4887 | { | |
4888 | unsigned long check_addr = *start; | |
4889 | ||
4890 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
4891 | return; | |
4892 | ||
4893 | for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { | |
4894 | unsigned long a_start = check_addr & PUD_MASK; | |
4895 | unsigned long a_end = a_start + PUD_SIZE; | |
4896 | ||
4897 | /* | |
4898 | * If sharing is possible, adjust start/end if necessary. | |
4899 | */ | |
4900 | if (range_in_vma(vma, a_start, a_end)) { | |
4901 | if (a_start < *start) | |
4902 | *start = a_start; | |
4903 | if (a_end > *end) | |
4904 | *end = a_end; | |
4905 | } | |
4906 | } | |
4907 | } | |
4908 | ||
3212b535 SC |
4909 | /* |
4910 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4911 | * and returns the corresponding pte. While this is not necessary for the | |
4912 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
4913 | * code much cleaner. |
4914 | * | |
4915 | * This routine must be called with i_mmap_rwsem held in at least read mode. | |
4916 | * For hugetlbfs, this prevents removal of any page table entries associated | |
4917 | * with the address space. This is important as we are setting up sharing | |
4918 | * based on existing page table entries (mappings). | |
3212b535 SC |
4919 | */ |
4920 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4921 | { | |
4922 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4923 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4924 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4925 | vma->vm_pgoff; | |
4926 | struct vm_area_struct *svma; | |
4927 | unsigned long saddr; | |
4928 | pte_t *spte = NULL; | |
4929 | pte_t *pte; | |
cb900f41 | 4930 | spinlock_t *ptl; |
3212b535 SC |
4931 | |
4932 | if (!vma_shareable(vma, addr)) | |
4933 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4934 | ||
3212b535 SC |
4935 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4936 | if (svma == vma) | |
4937 | continue; | |
4938 | ||
4939 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4940 | if (saddr) { | |
7868a208 PA |
4941 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4942 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4943 | if (spte) { |
4944 | get_page(virt_to_page(spte)); | |
4945 | break; | |
4946 | } | |
4947 | } | |
4948 | } | |
4949 | ||
4950 | if (!spte) | |
4951 | goto out; | |
4952 | ||
8bea8052 | 4953 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4954 | if (pud_none(*pud)) { |
3212b535 SC |
4955 | pud_populate(mm, pud, |
4956 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4957 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4958 | } else { |
3212b535 | 4959 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4960 | } |
cb900f41 | 4961 | spin_unlock(ptl); |
3212b535 SC |
4962 | out: |
4963 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
4964 | return pte; |
4965 | } | |
4966 | ||
4967 | /* | |
4968 | * unmap huge page backed by shared pte. | |
4969 | * | |
4970 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4971 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4972 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4973 | * | |
c0d0381a | 4974 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
4975 | * |
4976 | * returns: 1 successfully unmapped a shared pte page | |
4977 | * 0 the underlying pte page is not shared, or it is the last user | |
4978 | */ | |
4979 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4980 | { | |
4981 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4982 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4983 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4984 | |
4985 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4986 | if (page_count(virt_to_page(ptep)) == 1) | |
4987 | return 0; | |
4988 | ||
4989 | pud_clear(pud); | |
4990 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4991 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4992 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4993 | return 1; | |
4994 | } | |
9e5fc74c SC |
4995 | #define want_pmd_share() (1) |
4996 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4997 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4998 | { | |
4999 | return NULL; | |
5000 | } | |
e81f2d22 ZZ |
5001 | |
5002 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
5003 | { | |
5004 | return 0; | |
5005 | } | |
017b1660 MK |
5006 | |
5007 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5008 | unsigned long *start, unsigned long *end) | |
5009 | { | |
5010 | } | |
9e5fc74c | 5011 | #define want_pmd_share() (0) |
3212b535 SC |
5012 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5013 | ||
9e5fc74c SC |
5014 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
5015 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
5016 | unsigned long addr, unsigned long sz) | |
5017 | { | |
5018 | pgd_t *pgd; | |
c2febafc | 5019 | p4d_t *p4d; |
9e5fc74c SC |
5020 | pud_t *pud; |
5021 | pte_t *pte = NULL; | |
5022 | ||
5023 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5024 | p4d = p4d_alloc(mm, pgd, addr); |
5025 | if (!p4d) | |
5026 | return NULL; | |
c2febafc | 5027 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5028 | if (pud) { |
5029 | if (sz == PUD_SIZE) { | |
5030 | pte = (pte_t *)pud; | |
5031 | } else { | |
5032 | BUG_ON(sz != PMD_SIZE); | |
5033 | if (want_pmd_share() && pud_none(*pud)) | |
5034 | pte = huge_pmd_share(mm, addr, pud); | |
5035 | else | |
5036 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5037 | } | |
5038 | } | |
4e666314 | 5039 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5040 | |
5041 | return pte; | |
5042 | } | |
5043 | ||
9b19df29 PA |
5044 | /* |
5045 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5046 | * entry at address @addr | |
5047 | * | |
5048 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
5049 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
5050 | * size @sz doesn't match the hugepage size at this level of the page | |
5051 | * table. | |
5052 | */ | |
7868a208 PA |
5053 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5054 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5055 | { |
5056 | pgd_t *pgd; | |
c2febafc | 5057 | p4d_t *p4d; |
9e5fc74c | 5058 | pud_t *pud; |
c2febafc | 5059 | pmd_t *pmd; |
9e5fc74c SC |
5060 | |
5061 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5062 | if (!pgd_present(*pgd)) |
5063 | return NULL; | |
5064 | p4d = p4d_offset(pgd, addr); | |
5065 | if (!p4d_present(*p4d)) | |
5066 | return NULL; | |
9b19df29 | 5067 | |
c2febafc | 5068 | pud = pud_offset(p4d, addr); |
9b19df29 | 5069 | if (sz != PUD_SIZE && pud_none(*pud)) |
c2febafc | 5070 | return NULL; |
9b19df29 PA |
5071 | /* hugepage or swap? */ |
5072 | if (pud_huge(*pud) || !pud_present(*pud)) | |
c2febafc | 5073 | return (pte_t *)pud; |
9b19df29 | 5074 | |
c2febafc | 5075 | pmd = pmd_offset(pud, addr); |
9b19df29 PA |
5076 | if (sz != PMD_SIZE && pmd_none(*pmd)) |
5077 | return NULL; | |
5078 | /* hugepage or swap? */ | |
5079 | if (pmd_huge(*pmd) || !pmd_present(*pmd)) | |
5080 | return (pte_t *)pmd; | |
5081 | ||
5082 | return NULL; | |
9e5fc74c SC |
5083 | } |
5084 | ||
61f77eda NH |
5085 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5086 | ||
5087 | /* | |
5088 | * These functions are overwritable if your architecture needs its own | |
5089 | * behavior. | |
5090 | */ | |
5091 | struct page * __weak | |
5092 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5093 | int write) | |
5094 | { | |
5095 | return ERR_PTR(-EINVAL); | |
5096 | } | |
5097 | ||
4dc71451 AK |
5098 | struct page * __weak |
5099 | follow_huge_pd(struct vm_area_struct *vma, | |
5100 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5101 | { | |
5102 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5103 | return NULL; | |
5104 | } | |
5105 | ||
61f77eda | 5106 | struct page * __weak |
9e5fc74c | 5107 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5108 | pmd_t *pmd, int flags) |
9e5fc74c | 5109 | { |
e66f17ff NH |
5110 | struct page *page = NULL; |
5111 | spinlock_t *ptl; | |
c9d398fa | 5112 | pte_t pte; |
3faa52c0 JH |
5113 | |
5114 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5115 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5116 | (FOLL_PIN | FOLL_GET))) | |
5117 | return NULL; | |
5118 | ||
e66f17ff NH |
5119 | retry: |
5120 | ptl = pmd_lockptr(mm, pmd); | |
5121 | spin_lock(ptl); | |
5122 | /* | |
5123 | * make sure that the address range covered by this pmd is not | |
5124 | * unmapped from other threads. | |
5125 | */ | |
5126 | if (!pmd_huge(*pmd)) | |
5127 | goto out; | |
c9d398fa NH |
5128 | pte = huge_ptep_get((pte_t *)pmd); |
5129 | if (pte_present(pte)) { | |
97534127 | 5130 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
5131 | /* |
5132 | * try_grab_page() should always succeed here, because: a) we | |
5133 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5134 | * huge pmd (head) page is present in the page tables. The ptl | |
5135 | * prevents the head page and tail pages from being rearranged | |
5136 | * in any way. So this page must be available at this point, | |
5137 | * unless the page refcount overflowed: | |
5138 | */ | |
5139 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5140 | page = NULL; | |
5141 | goto out; | |
5142 | } | |
e66f17ff | 5143 | } else { |
c9d398fa | 5144 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5145 | spin_unlock(ptl); |
5146 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5147 | goto retry; | |
5148 | } | |
5149 | /* | |
5150 | * hwpoisoned entry is treated as no_page_table in | |
5151 | * follow_page_mask(). | |
5152 | */ | |
5153 | } | |
5154 | out: | |
5155 | spin_unlock(ptl); | |
9e5fc74c SC |
5156 | return page; |
5157 | } | |
5158 | ||
61f77eda | 5159 | struct page * __weak |
9e5fc74c | 5160 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5161 | pud_t *pud, int flags) |
9e5fc74c | 5162 | { |
3faa52c0 | 5163 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5164 | return NULL; |
9e5fc74c | 5165 | |
e66f17ff | 5166 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5167 | } |
5168 | ||
faaa5b62 AK |
5169 | struct page * __weak |
5170 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5171 | { | |
3faa52c0 | 5172 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5173 | return NULL; |
5174 | ||
5175 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5176 | } | |
5177 | ||
31caf665 NH |
5178 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5179 | { | |
bcc54222 NH |
5180 | bool ret = true; |
5181 | ||
309381fe | 5182 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5183 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
5184 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
5185 | ret = false; | |
5186 | goto unlock; | |
5187 | } | |
5188 | clear_page_huge_active(page); | |
31caf665 | 5189 | list_move_tail(&page->lru, list); |
bcc54222 | 5190 | unlock: |
31caf665 | 5191 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5192 | return ret; |
31caf665 NH |
5193 | } |
5194 | ||
5195 | void putback_active_hugepage(struct page *page) | |
5196 | { | |
309381fe | 5197 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5198 | spin_lock(&hugetlb_lock); |
bcc54222 | 5199 | set_page_huge_active(page); |
31caf665 NH |
5200 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5201 | spin_unlock(&hugetlb_lock); | |
5202 | put_page(page); | |
5203 | } | |
ab5ac90a MH |
5204 | |
5205 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5206 | { | |
5207 | struct hstate *h = page_hstate(oldpage); | |
5208 | ||
5209 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5210 | set_page_owner_migrate_reason(newpage, reason); | |
5211 | ||
5212 | /* | |
5213 | * transfer temporary state of the new huge page. This is | |
5214 | * reverse to other transitions because the newpage is going to | |
5215 | * be final while the old one will be freed so it takes over | |
5216 | * the temporary status. | |
5217 | * | |
5218 | * Also note that we have to transfer the per-node surplus state | |
5219 | * here as well otherwise the global surplus count will not match | |
5220 | * the per-node's. | |
5221 | */ | |
5222 | if (PageHugeTemporary(newpage)) { | |
5223 | int old_nid = page_to_nid(oldpage); | |
5224 | int new_nid = page_to_nid(newpage); | |
5225 | ||
5226 | SetPageHugeTemporary(oldpage); | |
5227 | ClearPageHugeTemporary(newpage); | |
5228 | ||
5229 | spin_lock(&hugetlb_lock); | |
5230 | if (h->surplus_huge_pages_node[old_nid]) { | |
5231 | h->surplus_huge_pages_node[old_nid]--; | |
5232 | h->surplus_huge_pages_node[new_nid]++; | |
5233 | } | |
5234 | spin_unlock(&hugetlb_lock); | |
5235 | } | |
5236 | } |