]>
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); |
944d9fec | 1012 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1013 | clear_compound_head(p); |
944d9fec | 1014 | set_page_refcounted(p); |
944d9fec LC |
1015 | } |
1016 | ||
1017 | set_compound_order(page, 0); | |
1018 | __ClearPageHead(page); | |
1019 | } | |
1020 | ||
d00181b9 | 1021 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1022 | { |
1023 | free_contig_range(page_to_pfn(page), 1 << order); | |
1024 | } | |
1025 | ||
4eb0716e | 1026 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1027 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1028 | int nid, nodemask_t *nodemask) | |
944d9fec | 1029 | { |
5e27a2df | 1030 | unsigned long nr_pages = 1UL << huge_page_order(h); |
944d9fec | 1031 | |
5e27a2df | 1032 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1033 | } |
1034 | ||
1035 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1036 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1037 | #else /* !CONFIG_CONTIG_ALLOC */ |
1038 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1039 | int nid, nodemask_t *nodemask) | |
1040 | { | |
1041 | return NULL; | |
1042 | } | |
1043 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1044 | |
e1073d1e | 1045 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1046 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1047 | int nid, nodemask_t *nodemask) |
1048 | { | |
1049 | return NULL; | |
1050 | } | |
d00181b9 | 1051 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1052 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1053 | unsigned int order) { } |
944d9fec LC |
1054 | #endif |
1055 | ||
a5516438 | 1056 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1057 | { |
1058 | int i; | |
a5516438 | 1059 | |
4eb0716e | 1060 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1061 | return; |
18229df5 | 1062 | |
a5516438 AK |
1063 | h->nr_huge_pages--; |
1064 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1065 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1066 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1067 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1068 | 1 << PG_active | 1 << PG_private | |
1069 | 1 << PG_writeback); | |
6af2acb6 | 1070 | } |
309381fe | 1071 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1072 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1073 | set_page_refcounted(page); |
944d9fec LC |
1074 | if (hstate_is_gigantic(h)) { |
1075 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1076 | free_gigantic_page(page, huge_page_order(h)); | |
1077 | } else { | |
944d9fec LC |
1078 | __free_pages(page, huge_page_order(h)); |
1079 | } | |
6af2acb6 AL |
1080 | } |
1081 | ||
e5ff2159 AK |
1082 | struct hstate *size_to_hstate(unsigned long size) |
1083 | { | |
1084 | struct hstate *h; | |
1085 | ||
1086 | for_each_hstate(h) { | |
1087 | if (huge_page_size(h) == size) | |
1088 | return h; | |
1089 | } | |
1090 | return NULL; | |
1091 | } | |
1092 | ||
bcc54222 NH |
1093 | /* |
1094 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1095 | * to hstate->hugepage_activelist.) | |
1096 | * | |
1097 | * This function can be called for tail pages, but never returns true for them. | |
1098 | */ | |
1099 | bool page_huge_active(struct page *page) | |
1100 | { | |
1101 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1102 | return PageHead(page) && PagePrivate(&page[1]); | |
1103 | } | |
1104 | ||
1105 | /* never called for tail page */ | |
1106 | static void set_page_huge_active(struct page *page) | |
1107 | { | |
1108 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1109 | SetPagePrivate(&page[1]); | |
1110 | } | |
1111 | ||
1112 | static void clear_page_huge_active(struct page *page) | |
1113 | { | |
1114 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1115 | ClearPagePrivate(&page[1]); | |
1116 | } | |
1117 | ||
ab5ac90a MH |
1118 | /* |
1119 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1120 | * code | |
1121 | */ | |
1122 | static inline bool PageHugeTemporary(struct page *page) | |
1123 | { | |
1124 | if (!PageHuge(page)) | |
1125 | return false; | |
1126 | ||
1127 | return (unsigned long)page[2].mapping == -1U; | |
1128 | } | |
1129 | ||
1130 | static inline void SetPageHugeTemporary(struct page *page) | |
1131 | { | |
1132 | page[2].mapping = (void *)-1U; | |
1133 | } | |
1134 | ||
1135 | static inline void ClearPageHugeTemporary(struct page *page) | |
1136 | { | |
1137 | page[2].mapping = NULL; | |
1138 | } | |
1139 | ||
c77c0a8a | 1140 | static void __free_huge_page(struct page *page) |
27a85ef1 | 1141 | { |
a5516438 AK |
1142 | /* |
1143 | * Can't pass hstate in here because it is called from the | |
1144 | * compound page destructor. | |
1145 | */ | |
e5ff2159 | 1146 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1147 | int nid = page_to_nid(page); |
90481622 DG |
1148 | struct hugepage_subpool *spool = |
1149 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1150 | bool restore_reserve; |
27a85ef1 | 1151 | |
b4330afb MK |
1152 | VM_BUG_ON_PAGE(page_count(page), page); |
1153 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc YW |
1154 | |
1155 | set_page_private(page, 0); | |
1156 | page->mapping = NULL; | |
07443a85 | 1157 | restore_reserve = PagePrivate(page); |
16c794b4 | 1158 | ClearPagePrivate(page); |
27a85ef1 | 1159 | |
1c5ecae3 | 1160 | /* |
0919e1b6 MK |
1161 | * If PagePrivate() was set on page, page allocation consumed a |
1162 | * reservation. If the page was associated with a subpool, there | |
1163 | * would have been a page reserved in the subpool before allocation | |
1164 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1165 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1166 | * remove the reserved page from the subpool. | |
1c5ecae3 | 1167 | */ |
0919e1b6 MK |
1168 | if (!restore_reserve) { |
1169 | /* | |
1170 | * A return code of zero implies that the subpool will be | |
1171 | * under its minimum size if the reservation is not restored | |
1172 | * after page is free. Therefore, force restore_reserve | |
1173 | * operation. | |
1174 | */ | |
1175 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1176 | restore_reserve = true; | |
1177 | } | |
1c5ecae3 | 1178 | |
27a85ef1 | 1179 | spin_lock(&hugetlb_lock); |
bcc54222 | 1180 | clear_page_huge_active(page); |
6d76dcf4 AK |
1181 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1182 | pages_per_huge_page(h), page); | |
07443a85 JK |
1183 | if (restore_reserve) |
1184 | h->resv_huge_pages++; | |
1185 | ||
ab5ac90a MH |
1186 | if (PageHugeTemporary(page)) { |
1187 | list_del(&page->lru); | |
1188 | ClearPageHugeTemporary(page); | |
1189 | update_and_free_page(h, page); | |
1190 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1191 | /* remove the page from active list */ |
1192 | list_del(&page->lru); | |
a5516438 AK |
1193 | update_and_free_page(h, page); |
1194 | h->surplus_huge_pages--; | |
1195 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1196 | } else { |
5d3a551c | 1197 | arch_clear_hugepage_flags(page); |
a5516438 | 1198 | enqueue_huge_page(h, page); |
7893d1d5 | 1199 | } |
27a85ef1 DG |
1200 | spin_unlock(&hugetlb_lock); |
1201 | } | |
1202 | ||
c77c0a8a WL |
1203 | /* |
1204 | * As free_huge_page() can be called from a non-task context, we have | |
1205 | * to defer the actual freeing in a workqueue to prevent potential | |
1206 | * hugetlb_lock deadlock. | |
1207 | * | |
1208 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1209 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1210 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1211 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1212 | */ | |
1213 | static LLIST_HEAD(hpage_freelist); | |
1214 | ||
1215 | static void free_hpage_workfn(struct work_struct *work) | |
1216 | { | |
1217 | struct llist_node *node; | |
1218 | struct page *page; | |
1219 | ||
1220 | node = llist_del_all(&hpage_freelist); | |
1221 | ||
1222 | while (node) { | |
1223 | page = container_of((struct address_space **)node, | |
1224 | struct page, mapping); | |
1225 | node = node->next; | |
1226 | __free_huge_page(page); | |
1227 | } | |
1228 | } | |
1229 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1230 | ||
1231 | void free_huge_page(struct page *page) | |
1232 | { | |
1233 | /* | |
1234 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1235 | */ | |
1236 | if (!in_task()) { | |
1237 | /* | |
1238 | * Only call schedule_work() if hpage_freelist is previously | |
1239 | * empty. Otherwise, schedule_work() had been called but the | |
1240 | * workfn hasn't retrieved the list yet. | |
1241 | */ | |
1242 | if (llist_add((struct llist_node *)&page->mapping, | |
1243 | &hpage_freelist)) | |
1244 | schedule_work(&free_hpage_work); | |
1245 | return; | |
1246 | } | |
1247 | ||
1248 | __free_huge_page(page); | |
1249 | } | |
1250 | ||
a5516438 | 1251 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1252 | { |
0edaecfa | 1253 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1254 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1255 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1256 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1257 | h->nr_huge_pages++; |
1258 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 | 1259 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1260 | } |
1261 | ||
d00181b9 | 1262 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1263 | { |
1264 | int i; | |
1265 | int nr_pages = 1 << order; | |
1266 | struct page *p = page + 1; | |
1267 | ||
1268 | /* we rely on prep_new_huge_page to set the destructor */ | |
1269 | set_compound_order(page, order); | |
ef5a22be | 1270 | __ClearPageReserved(page); |
de09d31d | 1271 | __SetPageHead(page); |
20a0307c | 1272 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1273 | /* |
1274 | * For gigantic hugepages allocated through bootmem at | |
1275 | * boot, it's safer to be consistent with the not-gigantic | |
1276 | * hugepages and clear the PG_reserved bit from all tail pages | |
1277 | * too. Otherwse drivers using get_user_pages() to access tail | |
1278 | * pages may get the reference counting wrong if they see | |
1279 | * PG_reserved set on a tail page (despite the head page not | |
1280 | * having PG_reserved set). Enforcing this consistency between | |
1281 | * head and tail pages allows drivers to optimize away a check | |
1282 | * on the head page when they need know if put_page() is needed | |
1283 | * after get_user_pages(). | |
1284 | */ | |
1285 | __ClearPageReserved(p); | |
58a84aa9 | 1286 | set_page_count(p, 0); |
1d798ca3 | 1287 | set_compound_head(p, page); |
20a0307c | 1288 | } |
b4330afb | 1289 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1290 | } |
1291 | ||
7795912c AM |
1292 | /* |
1293 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1294 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1295 | * details. | |
1296 | */ | |
20a0307c WF |
1297 | int PageHuge(struct page *page) |
1298 | { | |
20a0307c WF |
1299 | if (!PageCompound(page)) |
1300 | return 0; | |
1301 | ||
1302 | page = compound_head(page); | |
f1e61557 | 1303 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1304 | } |
43131e14 NH |
1305 | EXPORT_SYMBOL_GPL(PageHuge); |
1306 | ||
27c73ae7 AA |
1307 | /* |
1308 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1309 | * normal or transparent huge pages. | |
1310 | */ | |
1311 | int PageHeadHuge(struct page *page_head) | |
1312 | { | |
27c73ae7 AA |
1313 | if (!PageHead(page_head)) |
1314 | return 0; | |
1315 | ||
758f66a2 | 1316 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1317 | } |
27c73ae7 | 1318 | |
13d60f4b ZY |
1319 | pgoff_t __basepage_index(struct page *page) |
1320 | { | |
1321 | struct page *page_head = compound_head(page); | |
1322 | pgoff_t index = page_index(page_head); | |
1323 | unsigned long compound_idx; | |
1324 | ||
1325 | if (!PageHuge(page_head)) | |
1326 | return page_index(page); | |
1327 | ||
1328 | if (compound_order(page_head) >= MAX_ORDER) | |
1329 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1330 | else | |
1331 | compound_idx = page - page_head; | |
1332 | ||
1333 | return (index << compound_order(page_head)) + compound_idx; | |
1334 | } | |
1335 | ||
0c397dae | 1336 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1337 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1338 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1339 | { |
af0fb9df | 1340 | int order = huge_page_order(h); |
1da177e4 | 1341 | struct page *page; |
f60858f9 | 1342 | bool alloc_try_hard = true; |
f96efd58 | 1343 | |
f60858f9 MK |
1344 | /* |
1345 | * By default we always try hard to allocate the page with | |
1346 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1347 | * a loop (to adjust global huge page counts) and previous allocation | |
1348 | * failed, do not continue to try hard on the same node. Use the | |
1349 | * node_alloc_noretry bitmap to manage this state information. | |
1350 | */ | |
1351 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1352 | alloc_try_hard = false; | |
1353 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1354 | if (alloc_try_hard) | |
1355 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1356 | if (nid == NUMA_NO_NODE) |
1357 | nid = numa_mem_id(); | |
1358 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1359 | if (page) | |
1360 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1361 | else | |
1362 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1363 | |
f60858f9 MK |
1364 | /* |
1365 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1366 | * indicates an overall state change. Clear bit so that we resume | |
1367 | * normal 'try hard' allocations. | |
1368 | */ | |
1369 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1370 | node_clear(nid, *node_alloc_noretry); | |
1371 | ||
1372 | /* | |
1373 | * If we tried hard to get a page but failed, set bit so that | |
1374 | * subsequent attempts will not try as hard until there is an | |
1375 | * overall state change. | |
1376 | */ | |
1377 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1378 | node_set(nid, *node_alloc_noretry); | |
1379 | ||
63b4613c NA |
1380 | return page; |
1381 | } | |
1382 | ||
0c397dae MH |
1383 | /* |
1384 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1385 | * should use this function to get new hugetlb pages | |
1386 | */ | |
1387 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1388 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1389 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1390 | { |
1391 | struct page *page; | |
1392 | ||
1393 | if (hstate_is_gigantic(h)) | |
1394 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1395 | else | |
1396 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1397 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1398 | if (!page) |
1399 | return NULL; | |
1400 | ||
1401 | if (hstate_is_gigantic(h)) | |
1402 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1403 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1404 | ||
1405 | return page; | |
1406 | } | |
1407 | ||
af0fb9df MH |
1408 | /* |
1409 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1410 | * manner. | |
1411 | */ | |
f60858f9 MK |
1412 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1413 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1414 | { |
1415 | struct page *page; | |
1416 | int nr_nodes, node; | |
af0fb9df | 1417 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1418 | |
1419 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1420 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1421 | node_alloc_noretry); | |
af0fb9df | 1422 | if (page) |
b2261026 | 1423 | break; |
b2261026 JK |
1424 | } |
1425 | ||
af0fb9df MH |
1426 | if (!page) |
1427 | return 0; | |
b2261026 | 1428 | |
af0fb9df MH |
1429 | put_page(page); /* free it into the hugepage allocator */ |
1430 | ||
1431 | return 1; | |
b2261026 JK |
1432 | } |
1433 | ||
e8c5c824 LS |
1434 | /* |
1435 | * Free huge page from pool from next node to free. | |
1436 | * Attempt to keep persistent huge pages more or less | |
1437 | * balanced over allowed nodes. | |
1438 | * Called with hugetlb_lock locked. | |
1439 | */ | |
6ae11b27 LS |
1440 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1441 | bool acct_surplus) | |
e8c5c824 | 1442 | { |
b2261026 | 1443 | int nr_nodes, node; |
e8c5c824 LS |
1444 | int ret = 0; |
1445 | ||
b2261026 | 1446 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1447 | /* |
1448 | * If we're returning unused surplus pages, only examine | |
1449 | * nodes with surplus pages. | |
1450 | */ | |
b2261026 JK |
1451 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1452 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1453 | struct page *page = |
b2261026 | 1454 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1455 | struct page, lru); |
1456 | list_del(&page->lru); | |
1457 | h->free_huge_pages--; | |
b2261026 | 1458 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1459 | if (acct_surplus) { |
1460 | h->surplus_huge_pages--; | |
b2261026 | 1461 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1462 | } |
e8c5c824 LS |
1463 | update_and_free_page(h, page); |
1464 | ret = 1; | |
9a76db09 | 1465 | break; |
e8c5c824 | 1466 | } |
b2261026 | 1467 | } |
e8c5c824 LS |
1468 | |
1469 | return ret; | |
1470 | } | |
1471 | ||
c8721bbb NH |
1472 | /* |
1473 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1474 | * nothing for in-use hugepages and non-hugepages. |
1475 | * This function returns values like below: | |
1476 | * | |
1477 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1478 | * (allocated or reserved.) | |
1479 | * 0: successfully dissolved free hugepages or the page is not a | |
1480 | * hugepage (considered as already dissolved) | |
c8721bbb | 1481 | */ |
c3114a84 | 1482 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1483 | { |
6bc9b564 | 1484 | int rc = -EBUSY; |
082d5b6b | 1485 | |
faf53def NH |
1486 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1487 | if (!PageHuge(page)) | |
1488 | return 0; | |
1489 | ||
c8721bbb | 1490 | spin_lock(&hugetlb_lock); |
faf53def NH |
1491 | if (!PageHuge(page)) { |
1492 | rc = 0; | |
1493 | goto out; | |
1494 | } | |
1495 | ||
1496 | if (!page_count(page)) { | |
2247bb33 GS |
1497 | struct page *head = compound_head(page); |
1498 | struct hstate *h = page_hstate(head); | |
1499 | int nid = page_to_nid(head); | |
6bc9b564 | 1500 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1501 | goto out; |
c3114a84 AK |
1502 | /* |
1503 | * Move PageHWPoison flag from head page to the raw error page, | |
1504 | * which makes any subpages rather than the error page reusable. | |
1505 | */ | |
1506 | if (PageHWPoison(head) && page != head) { | |
1507 | SetPageHWPoison(page); | |
1508 | ClearPageHWPoison(head); | |
1509 | } | |
2247bb33 | 1510 | list_del(&head->lru); |
c8721bbb NH |
1511 | h->free_huge_pages--; |
1512 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1513 | h->max_huge_pages--; |
2247bb33 | 1514 | update_and_free_page(h, head); |
6bc9b564 | 1515 | rc = 0; |
c8721bbb | 1516 | } |
082d5b6b | 1517 | out: |
c8721bbb | 1518 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1519 | return rc; |
c8721bbb NH |
1520 | } |
1521 | ||
1522 | /* | |
1523 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1524 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1525 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1526 | * part of it lies within the given range. | |
082d5b6b GS |
1527 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1528 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1529 | */ |
082d5b6b | 1530 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1531 | { |
c8721bbb | 1532 | unsigned long pfn; |
eb03aa00 | 1533 | struct page *page; |
082d5b6b | 1534 | int rc = 0; |
c8721bbb | 1535 | |
d0177639 | 1536 | if (!hugepages_supported()) |
082d5b6b | 1537 | return rc; |
d0177639 | 1538 | |
eb03aa00 GS |
1539 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1540 | page = pfn_to_page(pfn); | |
faf53def NH |
1541 | rc = dissolve_free_huge_page(page); |
1542 | if (rc) | |
1543 | break; | |
eb03aa00 | 1544 | } |
082d5b6b GS |
1545 | |
1546 | return rc; | |
c8721bbb NH |
1547 | } |
1548 | ||
ab5ac90a MH |
1549 | /* |
1550 | * Allocates a fresh surplus page from the page allocator. | |
1551 | */ | |
0c397dae | 1552 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1553 | int nid, nodemask_t *nmask) |
7893d1d5 | 1554 | { |
9980d744 | 1555 | struct page *page = NULL; |
7893d1d5 | 1556 | |
bae7f4ae | 1557 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1558 | return NULL; |
1559 | ||
d1c3fb1f | 1560 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1561 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1562 | goto out_unlock; | |
d1c3fb1f NA |
1563 | spin_unlock(&hugetlb_lock); |
1564 | ||
f60858f9 | 1565 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1566 | if (!page) |
0c397dae | 1567 | return NULL; |
d1c3fb1f NA |
1568 | |
1569 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1570 | /* |
1571 | * We could have raced with the pool size change. | |
1572 | * Double check that and simply deallocate the new page | |
1573 | * if we would end up overcommiting the surpluses. Abuse | |
1574 | * temporary page to workaround the nasty free_huge_page | |
1575 | * codeflow | |
1576 | */ | |
1577 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1578 | SetPageHugeTemporary(page); | |
2bf753e6 | 1579 | spin_unlock(&hugetlb_lock); |
9980d744 | 1580 | put_page(page); |
2bf753e6 | 1581 | return NULL; |
9980d744 | 1582 | } else { |
9980d744 | 1583 | h->surplus_huge_pages++; |
4704dea3 | 1584 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1585 | } |
9980d744 MH |
1586 | |
1587 | out_unlock: | |
d1c3fb1f | 1588 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1589 | |
1590 | return page; | |
1591 | } | |
1592 | ||
9a4e9f3b AK |
1593 | struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
1594 | int nid, nodemask_t *nmask) | |
ab5ac90a MH |
1595 | { |
1596 | struct page *page; | |
1597 | ||
1598 | if (hstate_is_gigantic(h)) | |
1599 | return NULL; | |
1600 | ||
f60858f9 | 1601 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1602 | if (!page) |
1603 | return NULL; | |
1604 | ||
1605 | /* | |
1606 | * We do not account these pages as surplus because they are only | |
1607 | * temporary and will be released properly on the last reference | |
1608 | */ | |
ab5ac90a MH |
1609 | SetPageHugeTemporary(page); |
1610 | ||
1611 | return page; | |
1612 | } | |
1613 | ||
099730d6 DH |
1614 | /* |
1615 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1616 | */ | |
e0ec90ee | 1617 | static |
0c397dae | 1618 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1619 | struct vm_area_struct *vma, unsigned long addr) |
1620 | { | |
aaf14e40 MH |
1621 | struct page *page; |
1622 | struct mempolicy *mpol; | |
1623 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1624 | int nid; | |
1625 | nodemask_t *nodemask; | |
1626 | ||
1627 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1628 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1629 | mpol_cond_put(mpol); |
1630 | ||
1631 | return page; | |
099730d6 DH |
1632 | } |
1633 | ||
ab5ac90a | 1634 | /* page migration callback function */ |
bf50bab2 NH |
1635 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
1636 | { | |
aaf14e40 | 1637 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1638 | struct page *page = NULL; |
bf50bab2 | 1639 | |
aaf14e40 MH |
1640 | if (nid != NUMA_NO_NODE) |
1641 | gfp_mask |= __GFP_THISNODE; | |
1642 | ||
bf50bab2 | 1643 | spin_lock(&hugetlb_lock); |
4ef91848 | 1644 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1645 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1646 | spin_unlock(&hugetlb_lock); |
1647 | ||
94ae8ba7 | 1648 | if (!page) |
0c397dae | 1649 | page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1650 | |
1651 | return page; | |
1652 | } | |
1653 | ||
ab5ac90a | 1654 | /* page migration callback function */ |
3e59fcb0 MH |
1655 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
1656 | nodemask_t *nmask) | |
4db9b2ef | 1657 | { |
aaf14e40 | 1658 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
1659 | |
1660 | spin_lock(&hugetlb_lock); | |
1661 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1662 | struct page *page; |
1663 | ||
1664 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1665 | if (page) { | |
1666 | spin_unlock(&hugetlb_lock); | |
1667 | return page; | |
4db9b2ef MH |
1668 | } |
1669 | } | |
1670 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1671 | |
0c397dae | 1672 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1673 | } |
1674 | ||
ebd63723 | 1675 | /* mempolicy aware migration callback */ |
389c8178 MH |
1676 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1677 | unsigned long address) | |
ebd63723 MH |
1678 | { |
1679 | struct mempolicy *mpol; | |
1680 | nodemask_t *nodemask; | |
1681 | struct page *page; | |
ebd63723 MH |
1682 | gfp_t gfp_mask; |
1683 | int node; | |
1684 | ||
ebd63723 MH |
1685 | gfp_mask = htlb_alloc_mask(h); |
1686 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
1687 | page = alloc_huge_page_nodemask(h, node, nodemask); | |
1688 | mpol_cond_put(mpol); | |
1689 | ||
1690 | return page; | |
1691 | } | |
1692 | ||
e4e574b7 | 1693 | /* |
25985edc | 1694 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1695 | * of size 'delta'. |
1696 | */ | |
a5516438 | 1697 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1698 | { |
1699 | struct list_head surplus_list; | |
1700 | struct page *page, *tmp; | |
1701 | int ret, i; | |
1702 | int needed, allocated; | |
28073b02 | 1703 | bool alloc_ok = true; |
e4e574b7 | 1704 | |
a5516438 | 1705 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1706 | if (needed <= 0) { |
a5516438 | 1707 | h->resv_huge_pages += delta; |
e4e574b7 | 1708 | return 0; |
ac09b3a1 | 1709 | } |
e4e574b7 AL |
1710 | |
1711 | allocated = 0; | |
1712 | INIT_LIST_HEAD(&surplus_list); | |
1713 | ||
1714 | ret = -ENOMEM; | |
1715 | retry: | |
1716 | spin_unlock(&hugetlb_lock); | |
1717 | for (i = 0; i < needed; i++) { | |
0c397dae | 1718 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1719 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1720 | if (!page) { |
1721 | alloc_ok = false; | |
1722 | break; | |
1723 | } | |
e4e574b7 | 1724 | list_add(&page->lru, &surplus_list); |
69ed779a | 1725 | cond_resched(); |
e4e574b7 | 1726 | } |
28073b02 | 1727 | allocated += i; |
e4e574b7 AL |
1728 | |
1729 | /* | |
1730 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1731 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1732 | */ | |
1733 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1734 | needed = (h->resv_huge_pages + delta) - |
1735 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1736 | if (needed > 0) { |
1737 | if (alloc_ok) | |
1738 | goto retry; | |
1739 | /* | |
1740 | * We were not able to allocate enough pages to | |
1741 | * satisfy the entire reservation so we free what | |
1742 | * we've allocated so far. | |
1743 | */ | |
1744 | goto free; | |
1745 | } | |
e4e574b7 AL |
1746 | /* |
1747 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1748 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1749 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1750 | * allocator. Commit the entire reservation here to prevent another |
1751 | * process from stealing the pages as they are added to the pool but | |
1752 | * before they are reserved. | |
e4e574b7 AL |
1753 | */ |
1754 | needed += allocated; | |
a5516438 | 1755 | h->resv_huge_pages += delta; |
e4e574b7 | 1756 | ret = 0; |
a9869b83 | 1757 | |
19fc3f0a | 1758 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1759 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1760 | if ((--needed) < 0) |
1761 | break; | |
a9869b83 NH |
1762 | /* |
1763 | * This page is now managed by the hugetlb allocator and has | |
1764 | * no users -- drop the buddy allocator's reference. | |
1765 | */ | |
1766 | put_page_testzero(page); | |
309381fe | 1767 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1768 | enqueue_huge_page(h, page); |
19fc3f0a | 1769 | } |
28073b02 | 1770 | free: |
b0365c8d | 1771 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1772 | |
1773 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1774 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1775 | put_page(page); | |
a9869b83 | 1776 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1777 | |
1778 | return ret; | |
1779 | } | |
1780 | ||
1781 | /* | |
e5bbc8a6 MK |
1782 | * This routine has two main purposes: |
1783 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1784 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1785 | * to the associated reservation map. | |
1786 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1787 | * the reservation. As many as unused_resv_pages may be freed. | |
1788 | * | |
1789 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1790 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1791 | * we must make sure nobody else can claim pages we are in the process of | |
1792 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1793 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1794 | */ |
a5516438 AK |
1795 | static void return_unused_surplus_pages(struct hstate *h, |
1796 | unsigned long unused_resv_pages) | |
e4e574b7 | 1797 | { |
e4e574b7 AL |
1798 | unsigned long nr_pages; |
1799 | ||
aa888a74 | 1800 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1801 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1802 | goto out; |
aa888a74 | 1803 | |
e5bbc8a6 MK |
1804 | /* |
1805 | * Part (or even all) of the reservation could have been backed | |
1806 | * by pre-allocated pages. Only free surplus pages. | |
1807 | */ | |
a5516438 | 1808 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1809 | |
685f3457 LS |
1810 | /* |
1811 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1812 | * evenly across all nodes with memory. Iterate across these nodes |
1813 | * until we can no longer free unreserved surplus pages. This occurs | |
1814 | * when the nodes with surplus pages have no free pages. | |
1815 | * free_pool_huge_page() will balance the the freed pages across the | |
1816 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1817 | * |
1818 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1819 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1820 | * to cover subsequent pages we may free. | |
685f3457 LS |
1821 | */ |
1822 | while (nr_pages--) { | |
e5bbc8a6 MK |
1823 | h->resv_huge_pages--; |
1824 | unused_resv_pages--; | |
8cebfcd0 | 1825 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1826 | goto out; |
7848a4bf | 1827 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1828 | } |
e5bbc8a6 MK |
1829 | |
1830 | out: | |
1831 | /* Fully uncommit the reservation */ | |
1832 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1833 | } |
1834 | ||
5e911373 | 1835 | |
c37f9fb1 | 1836 | /* |
feba16e2 | 1837 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1838 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1839 | * |
1840 | * vma_needs_reservation is called to determine if the huge page at addr | |
1841 | * within the vma has an associated reservation. If a reservation is | |
1842 | * needed, the value 1 is returned. The caller is then responsible for | |
1843 | * managing the global reservation and subpool usage counts. After | |
1844 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1845 | * to add the page to the reservation map. If the page allocation fails, |
1846 | * the reservation must be ended instead of committed. vma_end_reservation | |
1847 | * is called in such cases. | |
cf3ad20b MK |
1848 | * |
1849 | * In the normal case, vma_commit_reservation returns the same value | |
1850 | * as the preceding vma_needs_reservation call. The only time this | |
1851 | * is not the case is if a reserve map was changed between calls. It | |
1852 | * is the responsibility of the caller to notice the difference and | |
1853 | * take appropriate action. | |
96b96a96 MK |
1854 | * |
1855 | * vma_add_reservation is used in error paths where a reservation must | |
1856 | * be restored when a newly allocated huge page must be freed. It is | |
1857 | * to be called after calling vma_needs_reservation to determine if a | |
1858 | * reservation exists. | |
c37f9fb1 | 1859 | */ |
5e911373 MK |
1860 | enum vma_resv_mode { |
1861 | VMA_NEEDS_RESV, | |
1862 | VMA_COMMIT_RESV, | |
feba16e2 | 1863 | VMA_END_RESV, |
96b96a96 | 1864 | VMA_ADD_RESV, |
5e911373 | 1865 | }; |
cf3ad20b MK |
1866 | static long __vma_reservation_common(struct hstate *h, |
1867 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1868 | enum vma_resv_mode mode) |
c37f9fb1 | 1869 | { |
4e35f483 JK |
1870 | struct resv_map *resv; |
1871 | pgoff_t idx; | |
cf3ad20b | 1872 | long ret; |
c37f9fb1 | 1873 | |
4e35f483 JK |
1874 | resv = vma_resv_map(vma); |
1875 | if (!resv) | |
84afd99b | 1876 | return 1; |
c37f9fb1 | 1877 | |
4e35f483 | 1878 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1879 | switch (mode) { |
1880 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1881 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1882 | break; |
1883 | case VMA_COMMIT_RESV: | |
1884 | ret = region_add(resv, idx, idx + 1); | |
1885 | break; | |
feba16e2 | 1886 | case VMA_END_RESV: |
5e911373 MK |
1887 | region_abort(resv, idx, idx + 1); |
1888 | ret = 0; | |
1889 | break; | |
96b96a96 MK |
1890 | case VMA_ADD_RESV: |
1891 | if (vma->vm_flags & VM_MAYSHARE) | |
1892 | ret = region_add(resv, idx, idx + 1); | |
1893 | else { | |
1894 | region_abort(resv, idx, idx + 1); | |
1895 | ret = region_del(resv, idx, idx + 1); | |
1896 | } | |
1897 | break; | |
5e911373 MK |
1898 | default: |
1899 | BUG(); | |
1900 | } | |
84afd99b | 1901 | |
4e35f483 | 1902 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1903 | return ret; |
67961f9d MK |
1904 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1905 | /* | |
1906 | * In most cases, reserves always exist for private mappings. | |
1907 | * However, a file associated with mapping could have been | |
1908 | * hole punched or truncated after reserves were consumed. | |
1909 | * As subsequent fault on such a range will not use reserves. | |
1910 | * Subtle - The reserve map for private mappings has the | |
1911 | * opposite meaning than that of shared mappings. If NO | |
1912 | * entry is in the reserve map, it means a reservation exists. | |
1913 | * If an entry exists in the reserve map, it means the | |
1914 | * reservation has already been consumed. As a result, the | |
1915 | * return value of this routine is the opposite of the | |
1916 | * value returned from reserve map manipulation routines above. | |
1917 | */ | |
1918 | if (ret) | |
1919 | return 0; | |
1920 | else | |
1921 | return 1; | |
1922 | } | |
4e35f483 | 1923 | else |
cf3ad20b | 1924 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1925 | } |
cf3ad20b MK |
1926 | |
1927 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1928 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1929 | { |
5e911373 | 1930 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1931 | } |
84afd99b | 1932 | |
cf3ad20b MK |
1933 | static long vma_commit_reservation(struct hstate *h, |
1934 | struct vm_area_struct *vma, unsigned long addr) | |
1935 | { | |
5e911373 MK |
1936 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1937 | } | |
1938 | ||
feba16e2 | 1939 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1940 | struct vm_area_struct *vma, unsigned long addr) |
1941 | { | |
feba16e2 | 1942 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1943 | } |
1944 | ||
96b96a96 MK |
1945 | static long vma_add_reservation(struct hstate *h, |
1946 | struct vm_area_struct *vma, unsigned long addr) | |
1947 | { | |
1948 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
1949 | } | |
1950 | ||
1951 | /* | |
1952 | * This routine is called to restore a reservation on error paths. In the | |
1953 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
1954 | * and is about to be freed. If a reservation for the page existed, | |
1955 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
1956 | * in the newly allocated page. When the page is freed via free_huge_page, | |
1957 | * the global reservation count will be incremented if PagePrivate is set. | |
1958 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
1959 | * reserve map here to be consistent with global reserve count adjustments | |
1960 | * to be made by free_huge_page. | |
1961 | */ | |
1962 | static void restore_reserve_on_error(struct hstate *h, | |
1963 | struct vm_area_struct *vma, unsigned long address, | |
1964 | struct page *page) | |
1965 | { | |
1966 | if (unlikely(PagePrivate(page))) { | |
1967 | long rc = vma_needs_reservation(h, vma, address); | |
1968 | ||
1969 | if (unlikely(rc < 0)) { | |
1970 | /* | |
1971 | * Rare out of memory condition in reserve map | |
1972 | * manipulation. Clear PagePrivate so that | |
1973 | * global reserve count will not be incremented | |
1974 | * by free_huge_page. This will make it appear | |
1975 | * as though the reservation for this page was | |
1976 | * consumed. This may prevent the task from | |
1977 | * faulting in the page at a later time. This | |
1978 | * is better than inconsistent global huge page | |
1979 | * accounting of reserve counts. | |
1980 | */ | |
1981 | ClearPagePrivate(page); | |
1982 | } else if (rc) { | |
1983 | rc = vma_add_reservation(h, vma, address); | |
1984 | if (unlikely(rc < 0)) | |
1985 | /* | |
1986 | * See above comment about rare out of | |
1987 | * memory condition. | |
1988 | */ | |
1989 | ClearPagePrivate(page); | |
1990 | } else | |
1991 | vma_end_reservation(h, vma, address); | |
1992 | } | |
1993 | } | |
1994 | ||
70c3547e | 1995 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1996 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1997 | { |
90481622 | 1998 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1999 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2000 | struct page *page; |
d85f69b0 MK |
2001 | long map_chg, map_commit; |
2002 | long gbl_chg; | |
6d76dcf4 AK |
2003 | int ret, idx; |
2004 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 2005 | |
6d76dcf4 | 2006 | idx = hstate_index(h); |
a1e78772 | 2007 | /* |
d85f69b0 MK |
2008 | * Examine the region/reserve map to determine if the process |
2009 | * has a reservation for the page to be allocated. A return | |
2010 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2011 | */ |
d85f69b0 MK |
2012 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2013 | if (map_chg < 0) | |
76dcee75 | 2014 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2015 | |
2016 | /* | |
2017 | * Processes that did not create the mapping will have no | |
2018 | * reserves as indicated by the region/reserve map. Check | |
2019 | * that the allocation will not exceed the subpool limit. | |
2020 | * Allocations for MAP_NORESERVE mappings also need to be | |
2021 | * checked against any subpool limit. | |
2022 | */ | |
2023 | if (map_chg || avoid_reserve) { | |
2024 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2025 | if (gbl_chg < 0) { | |
feba16e2 | 2026 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2027 | return ERR_PTR(-ENOSPC); |
5e911373 | 2028 | } |
1da177e4 | 2029 | |
d85f69b0 MK |
2030 | /* |
2031 | * Even though there was no reservation in the region/reserve | |
2032 | * map, there could be reservations associated with the | |
2033 | * subpool that can be used. This would be indicated if the | |
2034 | * return value of hugepage_subpool_get_pages() is zero. | |
2035 | * However, if avoid_reserve is specified we still avoid even | |
2036 | * the subpool reservations. | |
2037 | */ | |
2038 | if (avoid_reserve) | |
2039 | gbl_chg = 1; | |
2040 | } | |
2041 | ||
6d76dcf4 | 2042 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2043 | if (ret) |
2044 | goto out_subpool_put; | |
2045 | ||
1da177e4 | 2046 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2047 | /* |
2048 | * glb_chg is passed to indicate whether or not a page must be taken | |
2049 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2050 | * a reservation exists for the allocation. | |
2051 | */ | |
2052 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2053 | if (!page) { |
94ae8ba7 | 2054 | spin_unlock(&hugetlb_lock); |
0c397dae | 2055 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2056 | if (!page) |
2057 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2058 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2059 | SetPagePrivate(page); | |
2060 | h->resv_huge_pages--; | |
2061 | } | |
79dbb236 AK |
2062 | spin_lock(&hugetlb_lock); |
2063 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2064 | /* Fall through */ |
68842c9b | 2065 | } |
81a6fcae JK |
2066 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2067 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2068 | |
90481622 | 2069 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2070 | |
d85f69b0 MK |
2071 | map_commit = vma_commit_reservation(h, vma, addr); |
2072 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2073 | /* |
2074 | * The page was added to the reservation map between | |
2075 | * vma_needs_reservation and vma_commit_reservation. | |
2076 | * This indicates a race with hugetlb_reserve_pages. | |
2077 | * Adjust for the subpool count incremented above AND | |
2078 | * in hugetlb_reserve_pages for the same page. Also, | |
2079 | * the reservation count added in hugetlb_reserve_pages | |
2080 | * no longer applies. | |
2081 | */ | |
2082 | long rsv_adjust; | |
2083 | ||
2084 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2085 | hugetlb_acct_memory(h, -rsv_adjust); | |
2086 | } | |
90d8b7e6 | 2087 | return page; |
8f34af6f JZ |
2088 | |
2089 | out_uncharge_cgroup: | |
2090 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2091 | out_subpool_put: | |
d85f69b0 | 2092 | if (map_chg || avoid_reserve) |
8f34af6f | 2093 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2094 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2095 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2096 | } |
2097 | ||
e24a1307 AK |
2098 | int alloc_bootmem_huge_page(struct hstate *h) |
2099 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2100 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2101 | { |
2102 | struct huge_bootmem_page *m; | |
b2261026 | 2103 | int nr_nodes, node; |
aa888a74 | 2104 | |
b2261026 | 2105 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2106 | void *addr; |
2107 | ||
eb31d559 | 2108 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2109 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2110 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2111 | if (addr) { |
2112 | /* | |
2113 | * Use the beginning of the huge page to store the | |
2114 | * huge_bootmem_page struct (until gather_bootmem | |
2115 | * puts them into the mem_map). | |
2116 | */ | |
2117 | m = addr; | |
91f47662 | 2118 | goto found; |
aa888a74 | 2119 | } |
aa888a74 AK |
2120 | } |
2121 | return 0; | |
2122 | ||
2123 | found: | |
df994ead | 2124 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2125 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2126 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2127 | list_add(&m->list, &huge_boot_pages); |
2128 | m->hstate = h; | |
2129 | return 1; | |
2130 | } | |
2131 | ||
d00181b9 KS |
2132 | static void __init prep_compound_huge_page(struct page *page, |
2133 | unsigned int order) | |
18229df5 AW |
2134 | { |
2135 | if (unlikely(order > (MAX_ORDER - 1))) | |
2136 | prep_compound_gigantic_page(page, order); | |
2137 | else | |
2138 | prep_compound_page(page, order); | |
2139 | } | |
2140 | ||
aa888a74 AK |
2141 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2142 | static void __init gather_bootmem_prealloc(void) | |
2143 | { | |
2144 | struct huge_bootmem_page *m; | |
2145 | ||
2146 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2147 | struct page *page = virt_to_page(m); |
aa888a74 | 2148 | struct hstate *h = m->hstate; |
ee8f248d | 2149 | |
aa888a74 | 2150 | WARN_ON(page_count(page) != 1); |
18229df5 | 2151 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2152 | WARN_ON(PageReserved(page)); |
aa888a74 | 2153 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2154 | put_page(page); /* free it into the hugepage allocator */ |
2155 | ||
b0320c7b RA |
2156 | /* |
2157 | * If we had gigantic hugepages allocated at boot time, we need | |
2158 | * to restore the 'stolen' pages to totalram_pages in order to | |
2159 | * fix confusing memory reports from free(1) and another | |
2160 | * side-effects, like CommitLimit going negative. | |
2161 | */ | |
bae7f4ae | 2162 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2163 | adjust_managed_page_count(page, 1 << h->order); |
520495fe | 2164 | cond_resched(); |
aa888a74 AK |
2165 | } |
2166 | } | |
2167 | ||
8faa8b07 | 2168 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2169 | { |
2170 | unsigned long i; | |
f60858f9 MK |
2171 | nodemask_t *node_alloc_noretry; |
2172 | ||
2173 | if (!hstate_is_gigantic(h)) { | |
2174 | /* | |
2175 | * Bit mask controlling how hard we retry per-node allocations. | |
2176 | * Ignore errors as lower level routines can deal with | |
2177 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2178 | * time, we are likely in bigger trouble. | |
2179 | */ | |
2180 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2181 | GFP_KERNEL); | |
2182 | } else { | |
2183 | /* allocations done at boot time */ | |
2184 | node_alloc_noretry = NULL; | |
2185 | } | |
2186 | ||
2187 | /* bit mask controlling how hard we retry per-node allocations */ | |
2188 | if (node_alloc_noretry) | |
2189 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2190 | |
e5ff2159 | 2191 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2192 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2193 | if (!alloc_bootmem_huge_page(h)) |
2194 | break; | |
0c397dae | 2195 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2196 | &node_states[N_MEMORY], |
2197 | node_alloc_noretry)) | |
1da177e4 | 2198 | break; |
69ed779a | 2199 | cond_resched(); |
1da177e4 | 2200 | } |
d715cf80 LH |
2201 | if (i < h->max_huge_pages) { |
2202 | char buf[32]; | |
2203 | ||
c6247f72 | 2204 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2205 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2206 | h->max_huge_pages, buf, i); | |
2207 | h->max_huge_pages = i; | |
2208 | } | |
f60858f9 MK |
2209 | |
2210 | kfree(node_alloc_noretry); | |
e5ff2159 AK |
2211 | } |
2212 | ||
2213 | static void __init hugetlb_init_hstates(void) | |
2214 | { | |
2215 | struct hstate *h; | |
2216 | ||
2217 | for_each_hstate(h) { | |
641844f5 NH |
2218 | if (minimum_order > huge_page_order(h)) |
2219 | minimum_order = huge_page_order(h); | |
2220 | ||
8faa8b07 | 2221 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2222 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2223 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2224 | } |
641844f5 | 2225 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2226 | } |
2227 | ||
2228 | static void __init report_hugepages(void) | |
2229 | { | |
2230 | struct hstate *h; | |
2231 | ||
2232 | for_each_hstate(h) { | |
4abd32db | 2233 | char buf[32]; |
c6247f72 MW |
2234 | |
2235 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2236 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2237 | buf, h->free_huge_pages); |
e5ff2159 AK |
2238 | } |
2239 | } | |
2240 | ||
1da177e4 | 2241 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2242 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2243 | nodemask_t *nodes_allowed) | |
1da177e4 | 2244 | { |
4415cc8d CL |
2245 | int i; |
2246 | ||
bae7f4ae | 2247 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2248 | return; |
2249 | ||
6ae11b27 | 2250 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2251 | struct page *page, *next; |
a5516438 AK |
2252 | struct list_head *freel = &h->hugepage_freelists[i]; |
2253 | list_for_each_entry_safe(page, next, freel, lru) { | |
2254 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2255 | return; |
1da177e4 LT |
2256 | if (PageHighMem(page)) |
2257 | continue; | |
2258 | list_del(&page->lru); | |
e5ff2159 | 2259 | update_and_free_page(h, page); |
a5516438 AK |
2260 | h->free_huge_pages--; |
2261 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2262 | } |
2263 | } | |
2264 | } | |
2265 | #else | |
6ae11b27 LS |
2266 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2267 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2268 | { |
2269 | } | |
2270 | #endif | |
2271 | ||
20a0307c WF |
2272 | /* |
2273 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2274 | * balanced by operating on them in a round-robin fashion. | |
2275 | * Returns 1 if an adjustment was made. | |
2276 | */ | |
6ae11b27 LS |
2277 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2278 | int delta) | |
20a0307c | 2279 | { |
b2261026 | 2280 | int nr_nodes, node; |
20a0307c WF |
2281 | |
2282 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2283 | |
b2261026 JK |
2284 | if (delta < 0) { |
2285 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2286 | if (h->surplus_huge_pages_node[node]) | |
2287 | goto found; | |
e8c5c824 | 2288 | } |
b2261026 JK |
2289 | } else { |
2290 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2291 | if (h->surplus_huge_pages_node[node] < | |
2292 | h->nr_huge_pages_node[node]) | |
2293 | goto found; | |
e8c5c824 | 2294 | } |
b2261026 JK |
2295 | } |
2296 | return 0; | |
20a0307c | 2297 | |
b2261026 JK |
2298 | found: |
2299 | h->surplus_huge_pages += delta; | |
2300 | h->surplus_huge_pages_node[node] += delta; | |
2301 | return 1; | |
20a0307c WF |
2302 | } |
2303 | ||
a5516438 | 2304 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2305 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2306 | nodemask_t *nodes_allowed) |
1da177e4 | 2307 | { |
7893d1d5 | 2308 | unsigned long min_count, ret; |
f60858f9 MK |
2309 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2310 | ||
2311 | /* | |
2312 | * Bit mask controlling how hard we retry per-node allocations. | |
2313 | * If we can not allocate the bit mask, do not attempt to allocate | |
2314 | * the requested huge pages. | |
2315 | */ | |
2316 | if (node_alloc_noretry) | |
2317 | nodes_clear(*node_alloc_noretry); | |
2318 | else | |
2319 | return -ENOMEM; | |
1da177e4 | 2320 | |
4eb0716e AG |
2321 | spin_lock(&hugetlb_lock); |
2322 | ||
fd875dca MK |
2323 | /* |
2324 | * Check for a node specific request. | |
2325 | * Changing node specific huge page count may require a corresponding | |
2326 | * change to the global count. In any case, the passed node mask | |
2327 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2328 | */ | |
2329 | if (nid != NUMA_NO_NODE) { | |
2330 | unsigned long old_count = count; | |
2331 | ||
2332 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2333 | /* | |
2334 | * User may have specified a large count value which caused the | |
2335 | * above calculation to overflow. In this case, they wanted | |
2336 | * to allocate as many huge pages as possible. Set count to | |
2337 | * largest possible value to align with their intention. | |
2338 | */ | |
2339 | if (count < old_count) | |
2340 | count = ULONG_MAX; | |
2341 | } | |
2342 | ||
4eb0716e AG |
2343 | /* |
2344 | * Gigantic pages runtime allocation depend on the capability for large | |
2345 | * page range allocation. | |
2346 | * If the system does not provide this feature, return an error when | |
2347 | * the user tries to allocate gigantic pages but let the user free the | |
2348 | * boottime allocated gigantic pages. | |
2349 | */ | |
2350 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2351 | if (count > persistent_huge_pages(h)) { | |
2352 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2353 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2354 | return -EINVAL; |
2355 | } | |
2356 | /* Fall through to decrease pool */ | |
2357 | } | |
aa888a74 | 2358 | |
7893d1d5 AL |
2359 | /* |
2360 | * Increase the pool size | |
2361 | * First take pages out of surplus state. Then make up the | |
2362 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2363 | * |
0c397dae | 2364 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2365 | * to convert a surplus huge page to a normal huge page. That is |
2366 | * not critical, though, it just means the overall size of the | |
2367 | * pool might be one hugepage larger than it needs to be, but | |
2368 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2369 | */ |
a5516438 | 2370 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2371 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2372 | break; |
2373 | } | |
2374 | ||
a5516438 | 2375 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2376 | /* |
2377 | * If this allocation races such that we no longer need the | |
2378 | * page, free_huge_page will handle it by freeing the page | |
2379 | * and reducing the surplus. | |
2380 | */ | |
2381 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2382 | |
2383 | /* yield cpu to avoid soft lockup */ | |
2384 | cond_resched(); | |
2385 | ||
f60858f9 MK |
2386 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2387 | node_alloc_noretry); | |
7893d1d5 AL |
2388 | spin_lock(&hugetlb_lock); |
2389 | if (!ret) | |
2390 | goto out; | |
2391 | ||
536240f2 MG |
2392 | /* Bail for signals. Probably ctrl-c from user */ |
2393 | if (signal_pending(current)) | |
2394 | goto out; | |
7893d1d5 | 2395 | } |
7893d1d5 AL |
2396 | |
2397 | /* | |
2398 | * Decrease the pool size | |
2399 | * First return free pages to the buddy allocator (being careful | |
2400 | * to keep enough around to satisfy reservations). Then place | |
2401 | * pages into surplus state as needed so the pool will shrink | |
2402 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2403 | * |
2404 | * By placing pages into the surplus state independent of the | |
2405 | * overcommit value, we are allowing the surplus pool size to | |
2406 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2407 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2408 | * though, we'll note that we're not allowed to exceed surplus |
2409 | * and won't grow the pool anywhere else. Not until one of the | |
2410 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2411 | */ |
a5516438 | 2412 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2413 | min_count = max(count, min_count); |
6ae11b27 | 2414 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2415 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2416 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2417 | break; |
55f67141 | 2418 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2419 | } |
a5516438 | 2420 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2421 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2422 | break; |
2423 | } | |
2424 | out: | |
4eb0716e | 2425 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2426 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2427 | |
f60858f9 MK |
2428 | NODEMASK_FREE(node_alloc_noretry); |
2429 | ||
4eb0716e | 2430 | return 0; |
1da177e4 LT |
2431 | } |
2432 | ||
a3437870 NA |
2433 | #define HSTATE_ATTR_RO(_name) \ |
2434 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2435 | ||
2436 | #define HSTATE_ATTR(_name) \ | |
2437 | static struct kobj_attribute _name##_attr = \ | |
2438 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2439 | ||
2440 | static struct kobject *hugepages_kobj; | |
2441 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2442 | ||
9a305230 LS |
2443 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2444 | ||
2445 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2446 | { |
2447 | int i; | |
9a305230 | 2448 | |
a3437870 | 2449 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2450 | if (hstate_kobjs[i] == kobj) { |
2451 | if (nidp) | |
2452 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2453 | return &hstates[i]; |
9a305230 LS |
2454 | } |
2455 | ||
2456 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2457 | } |
2458 | ||
06808b08 | 2459 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2460 | struct kobj_attribute *attr, char *buf) |
2461 | { | |
9a305230 LS |
2462 | struct hstate *h; |
2463 | unsigned long nr_huge_pages; | |
2464 | int nid; | |
2465 | ||
2466 | h = kobj_to_hstate(kobj, &nid); | |
2467 | if (nid == NUMA_NO_NODE) | |
2468 | nr_huge_pages = h->nr_huge_pages; | |
2469 | else | |
2470 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2471 | ||
2472 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2473 | } |
adbe8726 | 2474 | |
238d3c13 DR |
2475 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2476 | struct hstate *h, int nid, | |
2477 | unsigned long count, size_t len) | |
a3437870 NA |
2478 | { |
2479 | int err; | |
2d0adf7e | 2480 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2481 | |
2d0adf7e OS |
2482 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2483 | return -EINVAL; | |
adbe8726 | 2484 | |
9a305230 LS |
2485 | if (nid == NUMA_NO_NODE) { |
2486 | /* | |
2487 | * global hstate attribute | |
2488 | */ | |
2489 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2490 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2491 | n_mask = &node_states[N_MEMORY]; | |
2492 | else | |
2493 | n_mask = &nodes_allowed; | |
2494 | } else { | |
9a305230 | 2495 | /* |
fd875dca MK |
2496 | * Node specific request. count adjustment happens in |
2497 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2498 | */ |
2d0adf7e OS |
2499 | init_nodemask_of_node(&nodes_allowed, nid); |
2500 | n_mask = &nodes_allowed; | |
fd875dca | 2501 | } |
9a305230 | 2502 | |
2d0adf7e | 2503 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2504 | |
4eb0716e | 2505 | return err ? err : len; |
06808b08 LS |
2506 | } |
2507 | ||
238d3c13 DR |
2508 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2509 | struct kobject *kobj, const char *buf, | |
2510 | size_t len) | |
2511 | { | |
2512 | struct hstate *h; | |
2513 | unsigned long count; | |
2514 | int nid; | |
2515 | int err; | |
2516 | ||
2517 | err = kstrtoul(buf, 10, &count); | |
2518 | if (err) | |
2519 | return err; | |
2520 | ||
2521 | h = kobj_to_hstate(kobj, &nid); | |
2522 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2523 | } | |
2524 | ||
06808b08 LS |
2525 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2526 | struct kobj_attribute *attr, char *buf) | |
2527 | { | |
2528 | return nr_hugepages_show_common(kobj, attr, buf); | |
2529 | } | |
2530 | ||
2531 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2532 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2533 | { | |
238d3c13 | 2534 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2535 | } |
2536 | HSTATE_ATTR(nr_hugepages); | |
2537 | ||
06808b08 LS |
2538 | #ifdef CONFIG_NUMA |
2539 | ||
2540 | /* | |
2541 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2542 | * huge page alloc/free. | |
2543 | */ | |
2544 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2545 | struct kobj_attribute *attr, char *buf) | |
2546 | { | |
2547 | return nr_hugepages_show_common(kobj, attr, buf); | |
2548 | } | |
2549 | ||
2550 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2551 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2552 | { | |
238d3c13 | 2553 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2554 | } |
2555 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2556 | #endif | |
2557 | ||
2558 | ||
a3437870 NA |
2559 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2560 | struct kobj_attribute *attr, char *buf) | |
2561 | { | |
9a305230 | 2562 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2563 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2564 | } | |
adbe8726 | 2565 | |
a3437870 NA |
2566 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2567 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2568 | { | |
2569 | int err; | |
2570 | unsigned long input; | |
9a305230 | 2571 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2572 | |
bae7f4ae | 2573 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2574 | return -EINVAL; |
2575 | ||
3dbb95f7 | 2576 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2577 | if (err) |
73ae31e5 | 2578 | return err; |
a3437870 NA |
2579 | |
2580 | spin_lock(&hugetlb_lock); | |
2581 | h->nr_overcommit_huge_pages = input; | |
2582 | spin_unlock(&hugetlb_lock); | |
2583 | ||
2584 | return count; | |
2585 | } | |
2586 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2587 | ||
2588 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2589 | struct kobj_attribute *attr, char *buf) | |
2590 | { | |
9a305230 LS |
2591 | struct hstate *h; |
2592 | unsigned long free_huge_pages; | |
2593 | int nid; | |
2594 | ||
2595 | h = kobj_to_hstate(kobj, &nid); | |
2596 | if (nid == NUMA_NO_NODE) | |
2597 | free_huge_pages = h->free_huge_pages; | |
2598 | else | |
2599 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2600 | ||
2601 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2602 | } |
2603 | HSTATE_ATTR_RO(free_hugepages); | |
2604 | ||
2605 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2606 | struct kobj_attribute *attr, char *buf) | |
2607 | { | |
9a305230 | 2608 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2609 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2610 | } | |
2611 | HSTATE_ATTR_RO(resv_hugepages); | |
2612 | ||
2613 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2614 | struct kobj_attribute *attr, char *buf) | |
2615 | { | |
9a305230 LS |
2616 | struct hstate *h; |
2617 | unsigned long surplus_huge_pages; | |
2618 | int nid; | |
2619 | ||
2620 | h = kobj_to_hstate(kobj, &nid); | |
2621 | if (nid == NUMA_NO_NODE) | |
2622 | surplus_huge_pages = h->surplus_huge_pages; | |
2623 | else | |
2624 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2625 | ||
2626 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2627 | } |
2628 | HSTATE_ATTR_RO(surplus_hugepages); | |
2629 | ||
2630 | static struct attribute *hstate_attrs[] = { | |
2631 | &nr_hugepages_attr.attr, | |
2632 | &nr_overcommit_hugepages_attr.attr, | |
2633 | &free_hugepages_attr.attr, | |
2634 | &resv_hugepages_attr.attr, | |
2635 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2636 | #ifdef CONFIG_NUMA |
2637 | &nr_hugepages_mempolicy_attr.attr, | |
2638 | #endif | |
a3437870 NA |
2639 | NULL, |
2640 | }; | |
2641 | ||
67e5ed96 | 2642 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2643 | .attrs = hstate_attrs, |
2644 | }; | |
2645 | ||
094e9539 JM |
2646 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2647 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2648 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2649 | { |
2650 | int retval; | |
972dc4de | 2651 | int hi = hstate_index(h); |
a3437870 | 2652 | |
9a305230 LS |
2653 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2654 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2655 | return -ENOMEM; |
2656 | ||
9a305230 | 2657 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2658 | if (retval) |
9a305230 | 2659 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2660 | |
2661 | return retval; | |
2662 | } | |
2663 | ||
2664 | static void __init hugetlb_sysfs_init(void) | |
2665 | { | |
2666 | struct hstate *h; | |
2667 | int err; | |
2668 | ||
2669 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2670 | if (!hugepages_kobj) | |
2671 | return; | |
2672 | ||
2673 | for_each_hstate(h) { | |
9a305230 LS |
2674 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2675 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2676 | if (err) |
ffb22af5 | 2677 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2678 | } |
2679 | } | |
2680 | ||
9a305230 LS |
2681 | #ifdef CONFIG_NUMA |
2682 | ||
2683 | /* | |
2684 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2685 | * with node devices in node_devices[] using a parallel array. The array |
2686 | * index of a node device or _hstate == node id. | |
2687 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2688 | * the base kernel, on the hugetlb module. |
2689 | */ | |
2690 | struct node_hstate { | |
2691 | struct kobject *hugepages_kobj; | |
2692 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2693 | }; | |
b4e289a6 | 2694 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2695 | |
2696 | /* | |
10fbcf4c | 2697 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2698 | */ |
2699 | static struct attribute *per_node_hstate_attrs[] = { | |
2700 | &nr_hugepages_attr.attr, | |
2701 | &free_hugepages_attr.attr, | |
2702 | &surplus_hugepages_attr.attr, | |
2703 | NULL, | |
2704 | }; | |
2705 | ||
67e5ed96 | 2706 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
2707 | .attrs = per_node_hstate_attrs, |
2708 | }; | |
2709 | ||
2710 | /* | |
10fbcf4c | 2711 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2712 | * Returns node id via non-NULL nidp. |
2713 | */ | |
2714 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2715 | { | |
2716 | int nid; | |
2717 | ||
2718 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2719 | struct node_hstate *nhs = &node_hstates[nid]; | |
2720 | int i; | |
2721 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2722 | if (nhs->hstate_kobjs[i] == kobj) { | |
2723 | if (nidp) | |
2724 | *nidp = nid; | |
2725 | return &hstates[i]; | |
2726 | } | |
2727 | } | |
2728 | ||
2729 | BUG(); | |
2730 | return NULL; | |
2731 | } | |
2732 | ||
2733 | /* | |
10fbcf4c | 2734 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2735 | * No-op if no hstate attributes attached. |
2736 | */ | |
3cd8b44f | 2737 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2738 | { |
2739 | struct hstate *h; | |
10fbcf4c | 2740 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2741 | |
2742 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2743 | return; /* no hstate attributes */ |
9a305230 | 2744 | |
972dc4de AK |
2745 | for_each_hstate(h) { |
2746 | int idx = hstate_index(h); | |
2747 | if (nhs->hstate_kobjs[idx]) { | |
2748 | kobject_put(nhs->hstate_kobjs[idx]); | |
2749 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2750 | } |
972dc4de | 2751 | } |
9a305230 LS |
2752 | |
2753 | kobject_put(nhs->hugepages_kobj); | |
2754 | nhs->hugepages_kobj = NULL; | |
2755 | } | |
2756 | ||
9a305230 LS |
2757 | |
2758 | /* | |
10fbcf4c | 2759 | * Register hstate attributes for a single node device. |
9a305230 LS |
2760 | * No-op if attributes already registered. |
2761 | */ | |
3cd8b44f | 2762 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2763 | { |
2764 | struct hstate *h; | |
10fbcf4c | 2765 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2766 | int err; |
2767 | ||
2768 | if (nhs->hugepages_kobj) | |
2769 | return; /* already allocated */ | |
2770 | ||
2771 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2772 | &node->dev.kobj); |
9a305230 LS |
2773 | if (!nhs->hugepages_kobj) |
2774 | return; | |
2775 | ||
2776 | for_each_hstate(h) { | |
2777 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2778 | nhs->hstate_kobjs, | |
2779 | &per_node_hstate_attr_group); | |
2780 | if (err) { | |
ffb22af5 AM |
2781 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2782 | h->name, node->dev.id); | |
9a305230 LS |
2783 | hugetlb_unregister_node(node); |
2784 | break; | |
2785 | } | |
2786 | } | |
2787 | } | |
2788 | ||
2789 | /* | |
9b5e5d0f | 2790 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2791 | * devices of nodes that have memory. All on-line nodes should have |
2792 | * registered their associated device by this time. | |
9a305230 | 2793 | */ |
7d9ca000 | 2794 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2795 | { |
2796 | int nid; | |
2797 | ||
8cebfcd0 | 2798 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2799 | struct node *node = node_devices[nid]; |
10fbcf4c | 2800 | if (node->dev.id == nid) |
9a305230 LS |
2801 | hugetlb_register_node(node); |
2802 | } | |
2803 | ||
2804 | /* | |
10fbcf4c | 2805 | * Let the node device driver know we're here so it can |
9a305230 LS |
2806 | * [un]register hstate attributes on node hotplug. |
2807 | */ | |
2808 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2809 | hugetlb_unregister_node); | |
2810 | } | |
2811 | #else /* !CONFIG_NUMA */ | |
2812 | ||
2813 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2814 | { | |
2815 | BUG(); | |
2816 | if (nidp) | |
2817 | *nidp = -1; | |
2818 | return NULL; | |
2819 | } | |
2820 | ||
9a305230 LS |
2821 | static void hugetlb_register_all_nodes(void) { } |
2822 | ||
2823 | #endif | |
2824 | ||
a3437870 NA |
2825 | static int __init hugetlb_init(void) |
2826 | { | |
8382d914 DB |
2827 | int i; |
2828 | ||
457c1b27 | 2829 | if (!hugepages_supported()) |
0ef89d25 | 2830 | return 0; |
a3437870 | 2831 | |
e11bfbfc | 2832 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
2833 | if (default_hstate_size != 0) { |
2834 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
2835 | default_hstate_size, HPAGE_SIZE); | |
2836 | } | |
2837 | ||
e11bfbfc NP |
2838 | default_hstate_size = HPAGE_SIZE; |
2839 | if (!size_to_hstate(default_hstate_size)) | |
2840 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2841 | } |
972dc4de | 2842 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2843 | if (default_hstate_max_huge_pages) { |
2844 | if (!default_hstate.max_huge_pages) | |
2845 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2846 | } | |
a3437870 NA |
2847 | |
2848 | hugetlb_init_hstates(); | |
aa888a74 | 2849 | gather_bootmem_prealloc(); |
a3437870 NA |
2850 | report_hugepages(); |
2851 | ||
2852 | hugetlb_sysfs_init(); | |
9a305230 | 2853 | hugetlb_register_all_nodes(); |
7179e7bf | 2854 | hugetlb_cgroup_file_init(); |
9a305230 | 2855 | |
8382d914 DB |
2856 | #ifdef CONFIG_SMP |
2857 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2858 | #else | |
2859 | num_fault_mutexes = 1; | |
2860 | #endif | |
c672c7f2 | 2861 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
2862 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
2863 | GFP_KERNEL); | |
c672c7f2 | 2864 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2865 | |
2866 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2867 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2868 | return 0; |
2869 | } | |
3e89e1c5 | 2870 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2871 | |
2872 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2873 | void __init hugetlb_bad_size(void) |
2874 | { | |
2875 | parsed_valid_hugepagesz = false; | |
2876 | } | |
2877 | ||
d00181b9 | 2878 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2879 | { |
2880 | struct hstate *h; | |
8faa8b07 AK |
2881 | unsigned long i; |
2882 | ||
a3437870 | 2883 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2884 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2885 | return; |
2886 | } | |
47d38344 | 2887 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2888 | BUG_ON(order == 0); |
47d38344 | 2889 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2890 | h->order = order; |
2891 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2892 | h->nr_huge_pages = 0; |
2893 | h->free_huge_pages = 0; | |
2894 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2895 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2896 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2897 | h->next_nid_to_alloc = first_memory_node; |
2898 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2899 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2900 | huge_page_size(h)/1024); | |
8faa8b07 | 2901 | |
a3437870 NA |
2902 | parsed_hstate = h; |
2903 | } | |
2904 | ||
e11bfbfc | 2905 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2906 | { |
2907 | unsigned long *mhp; | |
8faa8b07 | 2908 | static unsigned long *last_mhp; |
a3437870 | 2909 | |
9fee021d VT |
2910 | if (!parsed_valid_hugepagesz) { |
2911 | pr_warn("hugepages = %s preceded by " | |
2912 | "an unsupported hugepagesz, ignoring\n", s); | |
2913 | parsed_valid_hugepagesz = true; | |
2914 | return 1; | |
2915 | } | |
a3437870 | 2916 | /* |
47d38344 | 2917 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2918 | * so this hugepages= parameter goes to the "default hstate". |
2919 | */ | |
9fee021d | 2920 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2921 | mhp = &default_hstate_max_huge_pages; |
2922 | else | |
2923 | mhp = &parsed_hstate->max_huge_pages; | |
2924 | ||
8faa8b07 | 2925 | if (mhp == last_mhp) { |
598d8091 | 2926 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2927 | return 1; |
2928 | } | |
2929 | ||
a3437870 NA |
2930 | if (sscanf(s, "%lu", mhp) <= 0) |
2931 | *mhp = 0; | |
2932 | ||
8faa8b07 AK |
2933 | /* |
2934 | * Global state is always initialized later in hugetlb_init. | |
2935 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2936 | * use the bootmem allocator. | |
2937 | */ | |
47d38344 | 2938 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2939 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2940 | ||
2941 | last_mhp = mhp; | |
2942 | ||
a3437870 NA |
2943 | return 1; |
2944 | } | |
e11bfbfc NP |
2945 | __setup("hugepages=", hugetlb_nrpages_setup); |
2946 | ||
2947 | static int __init hugetlb_default_setup(char *s) | |
2948 | { | |
2949 | default_hstate_size = memparse(s, &s); | |
2950 | return 1; | |
2951 | } | |
2952 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2953 | |
8a213460 NA |
2954 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2955 | { | |
2956 | int node; | |
2957 | unsigned int nr = 0; | |
2958 | ||
2959 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2960 | nr += array[node]; | |
2961 | ||
2962 | return nr; | |
2963 | } | |
2964 | ||
2965 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2966 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2967 | struct ctl_table *table, int write, | |
2968 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2969 | { |
e5ff2159 | 2970 | struct hstate *h = &default_hstate; |
238d3c13 | 2971 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2972 | int ret; |
e5ff2159 | 2973 | |
457c1b27 | 2974 | if (!hugepages_supported()) |
86613628 | 2975 | return -EOPNOTSUPP; |
457c1b27 | 2976 | |
e5ff2159 AK |
2977 | table->data = &tmp; |
2978 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2979 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2980 | if (ret) | |
2981 | goto out; | |
e5ff2159 | 2982 | |
238d3c13 DR |
2983 | if (write) |
2984 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2985 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2986 | out: |
2987 | return ret; | |
1da177e4 | 2988 | } |
396faf03 | 2989 | |
06808b08 LS |
2990 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2991 | void __user *buffer, size_t *length, loff_t *ppos) | |
2992 | { | |
2993 | ||
2994 | return hugetlb_sysctl_handler_common(false, table, write, | |
2995 | buffer, length, ppos); | |
2996 | } | |
2997 | ||
2998 | #ifdef CONFIG_NUMA | |
2999 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
3000 | void __user *buffer, size_t *length, loff_t *ppos) | |
3001 | { | |
3002 | return hugetlb_sysctl_handler_common(true, table, write, | |
3003 | buffer, length, ppos); | |
3004 | } | |
3005 | #endif /* CONFIG_NUMA */ | |
3006 | ||
a3d0c6aa | 3007 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 3008 | void __user *buffer, |
a3d0c6aa NA |
3009 | size_t *length, loff_t *ppos) |
3010 | { | |
a5516438 | 3011 | struct hstate *h = &default_hstate; |
e5ff2159 | 3012 | unsigned long tmp; |
08d4a246 | 3013 | int ret; |
e5ff2159 | 3014 | |
457c1b27 | 3015 | if (!hugepages_supported()) |
86613628 | 3016 | return -EOPNOTSUPP; |
457c1b27 | 3017 | |
c033a93c | 3018 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3019 | |
bae7f4ae | 3020 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3021 | return -EINVAL; |
3022 | ||
e5ff2159 AK |
3023 | table->data = &tmp; |
3024 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3025 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3026 | if (ret) | |
3027 | goto out; | |
e5ff2159 AK |
3028 | |
3029 | if (write) { | |
3030 | spin_lock(&hugetlb_lock); | |
3031 | h->nr_overcommit_huge_pages = tmp; | |
3032 | spin_unlock(&hugetlb_lock); | |
3033 | } | |
08d4a246 MH |
3034 | out: |
3035 | return ret; | |
a3d0c6aa NA |
3036 | } |
3037 | ||
1da177e4 LT |
3038 | #endif /* CONFIG_SYSCTL */ |
3039 | ||
e1759c21 | 3040 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3041 | { |
fcb2b0c5 RG |
3042 | struct hstate *h; |
3043 | unsigned long total = 0; | |
3044 | ||
457c1b27 NA |
3045 | if (!hugepages_supported()) |
3046 | return; | |
fcb2b0c5 RG |
3047 | |
3048 | for_each_hstate(h) { | |
3049 | unsigned long count = h->nr_huge_pages; | |
3050 | ||
3051 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3052 | ||
3053 | if (h == &default_hstate) | |
3054 | seq_printf(m, | |
3055 | "HugePages_Total: %5lu\n" | |
3056 | "HugePages_Free: %5lu\n" | |
3057 | "HugePages_Rsvd: %5lu\n" | |
3058 | "HugePages_Surp: %5lu\n" | |
3059 | "Hugepagesize: %8lu kB\n", | |
3060 | count, | |
3061 | h->free_huge_pages, | |
3062 | h->resv_huge_pages, | |
3063 | h->surplus_huge_pages, | |
3064 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3065 | } | |
3066 | ||
3067 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3068 | } |
3069 | ||
3070 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3071 | { | |
a5516438 | 3072 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3073 | if (!hugepages_supported()) |
3074 | return 0; | |
1da177e4 LT |
3075 | return sprintf(buf, |
3076 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3077 | "Node %d HugePages_Free: %5u\n" |
3078 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3079 | nid, h->nr_huge_pages_node[nid], |
3080 | nid, h->free_huge_pages_node[nid], | |
3081 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3082 | } |
3083 | ||
949f7ec5 DR |
3084 | void hugetlb_show_meminfo(void) |
3085 | { | |
3086 | struct hstate *h; | |
3087 | int nid; | |
3088 | ||
457c1b27 NA |
3089 | if (!hugepages_supported()) |
3090 | return; | |
3091 | ||
949f7ec5 DR |
3092 | for_each_node_state(nid, N_MEMORY) |
3093 | for_each_hstate(h) | |
3094 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3095 | nid, | |
3096 | h->nr_huge_pages_node[nid], | |
3097 | h->free_huge_pages_node[nid], | |
3098 | h->surplus_huge_pages_node[nid], | |
3099 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3100 | } | |
3101 | ||
5d317b2b NH |
3102 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3103 | { | |
3104 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3105 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3106 | } | |
3107 | ||
1da177e4 LT |
3108 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3109 | unsigned long hugetlb_total_pages(void) | |
3110 | { | |
d0028588 WL |
3111 | struct hstate *h; |
3112 | unsigned long nr_total_pages = 0; | |
3113 | ||
3114 | for_each_hstate(h) | |
3115 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3116 | return nr_total_pages; | |
1da177e4 | 3117 | } |
1da177e4 | 3118 | |
a5516438 | 3119 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3120 | { |
3121 | int ret = -ENOMEM; | |
3122 | ||
3123 | spin_lock(&hugetlb_lock); | |
3124 | /* | |
3125 | * When cpuset is configured, it breaks the strict hugetlb page | |
3126 | * reservation as the accounting is done on a global variable. Such | |
3127 | * reservation is completely rubbish in the presence of cpuset because | |
3128 | * the reservation is not checked against page availability for the | |
3129 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3130 | * with lack of free htlb page in cpuset that the task is in. | |
3131 | * Attempt to enforce strict accounting with cpuset is almost | |
3132 | * impossible (or too ugly) because cpuset is too fluid that | |
3133 | * task or memory node can be dynamically moved between cpusets. | |
3134 | * | |
3135 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3136 | * undesirable. However, in order to preserve some of the semantics, | |
3137 | * we fall back to check against current free page availability as | |
3138 | * a best attempt and hopefully to minimize the impact of changing | |
3139 | * semantics that cpuset has. | |
3140 | */ | |
3141 | if (delta > 0) { | |
a5516438 | 3142 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3143 | goto out; |
3144 | ||
a5516438 AK |
3145 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3146 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3147 | goto out; |
3148 | } | |
3149 | } | |
3150 | ||
3151 | ret = 0; | |
3152 | if (delta < 0) | |
a5516438 | 3153 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3154 | |
3155 | out: | |
3156 | spin_unlock(&hugetlb_lock); | |
3157 | return ret; | |
3158 | } | |
3159 | ||
84afd99b AW |
3160 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3161 | { | |
f522c3ac | 3162 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3163 | |
3164 | /* | |
3165 | * This new VMA should share its siblings reservation map if present. | |
3166 | * The VMA will only ever have a valid reservation map pointer where | |
3167 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3168 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3169 | * after this open call completes. It is therefore safe to take a |
3170 | * new reference here without additional locking. | |
3171 | */ | |
4e35f483 | 3172 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3173 | kref_get(&resv->refs); |
84afd99b AW |
3174 | } |
3175 | ||
a1e78772 MG |
3176 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3177 | { | |
a5516438 | 3178 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3179 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3180 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3181 | unsigned long reserve, start, end; |
1c5ecae3 | 3182 | long gbl_reserve; |
84afd99b | 3183 | |
4e35f483 JK |
3184 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3185 | return; | |
84afd99b | 3186 | |
4e35f483 JK |
3187 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3188 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3189 | |
4e35f483 | 3190 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3191 | |
4e35f483 JK |
3192 | kref_put(&resv->refs, resv_map_release); |
3193 | ||
3194 | if (reserve) { | |
1c5ecae3 MK |
3195 | /* |
3196 | * Decrement reserve counts. The global reserve count may be | |
3197 | * adjusted if the subpool has a minimum size. | |
3198 | */ | |
3199 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3200 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3201 | } |
a1e78772 MG |
3202 | } |
3203 | ||
31383c68 DW |
3204 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3205 | { | |
3206 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3207 | return -EINVAL; | |
3208 | return 0; | |
3209 | } | |
3210 | ||
05ea8860 DW |
3211 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3212 | { | |
3213 | struct hstate *hstate = hstate_vma(vma); | |
3214 | ||
3215 | return 1UL << huge_page_shift(hstate); | |
3216 | } | |
3217 | ||
1da177e4 LT |
3218 | /* |
3219 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3220 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3221 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3222 | * this far. | |
3223 | */ | |
b3ec9f33 | 3224 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3225 | { |
3226 | BUG(); | |
d0217ac0 | 3227 | return 0; |
1da177e4 LT |
3228 | } |
3229 | ||
eec3636a JC |
3230 | /* |
3231 | * When a new function is introduced to vm_operations_struct and added | |
3232 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3233 | * This is because under System V memory model, mappings created via | |
3234 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3235 | * their original vm_ops are overwritten with shm_vm_ops. | |
3236 | */ | |
f0f37e2f | 3237 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3238 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3239 | .open = hugetlb_vm_op_open, |
a1e78772 | 3240 | .close = hugetlb_vm_op_close, |
31383c68 | 3241 | .split = hugetlb_vm_op_split, |
05ea8860 | 3242 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3243 | }; |
3244 | ||
1e8f889b DG |
3245 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3246 | int writable) | |
63551ae0 DG |
3247 | { |
3248 | pte_t entry; | |
3249 | ||
1e8f889b | 3250 | if (writable) { |
106c992a GS |
3251 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3252 | vma->vm_page_prot))); | |
63551ae0 | 3253 | } else { |
106c992a GS |
3254 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3255 | vma->vm_page_prot)); | |
63551ae0 DG |
3256 | } |
3257 | entry = pte_mkyoung(entry); | |
3258 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3259 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3260 | |
3261 | return entry; | |
3262 | } | |
3263 | ||
1e8f889b DG |
3264 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3265 | unsigned long address, pte_t *ptep) | |
3266 | { | |
3267 | pte_t entry; | |
3268 | ||
106c992a | 3269 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3270 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3271 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3272 | } |
3273 | ||
d5ed7444 | 3274 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3275 | { |
3276 | swp_entry_t swp; | |
3277 | ||
3278 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3279 | return false; |
4a705fef NH |
3280 | swp = pte_to_swp_entry(pte); |
3281 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3282 | return true; |
4a705fef | 3283 | else |
d5ed7444 | 3284 | return false; |
4a705fef NH |
3285 | } |
3286 | ||
3287 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3288 | { | |
3289 | swp_entry_t swp; | |
3290 | ||
3291 | if (huge_pte_none(pte) || pte_present(pte)) | |
3292 | return 0; | |
3293 | swp = pte_to_swp_entry(pte); | |
3294 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3295 | return 1; | |
3296 | else | |
3297 | return 0; | |
3298 | } | |
1e8f889b | 3299 | |
63551ae0 DG |
3300 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3301 | struct vm_area_struct *vma) | |
3302 | { | |
5e41540c | 3303 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3304 | struct page *ptepage; |
1c59827d | 3305 | unsigned long addr; |
1e8f889b | 3306 | int cow; |
a5516438 AK |
3307 | struct hstate *h = hstate_vma(vma); |
3308 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3309 | struct mmu_notifier_range range; |
e8569dd2 | 3310 | int ret = 0; |
1e8f889b DG |
3311 | |
3312 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3313 | |
ac46d4f3 | 3314 | if (cow) { |
7269f999 | 3315 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3316 | vma->vm_start, |
ac46d4f3 JG |
3317 | vma->vm_end); |
3318 | mmu_notifier_invalidate_range_start(&range); | |
3319 | } | |
e8569dd2 | 3320 | |
a5516438 | 3321 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3322 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3323 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3324 | if (!src_pte) |
3325 | continue; | |
a5516438 | 3326 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3327 | if (!dst_pte) { |
3328 | ret = -ENOMEM; | |
3329 | break; | |
3330 | } | |
c5c99429 | 3331 | |
5e41540c MK |
3332 | /* |
3333 | * If the pagetables are shared don't copy or take references. | |
3334 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3335 | * | |
3336 | * However, src could have 'unshared' and dst shares with | |
3337 | * another vma. If dst_pte !none, this implies sharing. | |
3338 | * Check here before taking page table lock, and once again | |
3339 | * after taking the lock below. | |
3340 | */ | |
3341 | dst_entry = huge_ptep_get(dst_pte); | |
3342 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3343 | continue; |
3344 | ||
cb900f41 KS |
3345 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3346 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3347 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3348 | entry = huge_ptep_get(src_pte); |
5e41540c MK |
3349 | dst_entry = huge_ptep_get(dst_pte); |
3350 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3351 | /* | |
3352 | * Skip if src entry none. Also, skip in the | |
3353 | * unlikely case dst entry !none as this implies | |
3354 | * sharing with another vma. | |
3355 | */ | |
4a705fef NH |
3356 | ; |
3357 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3358 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3359 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3360 | ||
3361 | if (is_write_migration_entry(swp_entry) && cow) { | |
3362 | /* | |
3363 | * COW mappings require pages in both | |
3364 | * parent and child to be set to read. | |
3365 | */ | |
3366 | make_migration_entry_read(&swp_entry); | |
3367 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3368 | set_huge_swap_pte_at(src, addr, src_pte, |
3369 | entry, sz); | |
4a705fef | 3370 | } |
e5251fd4 | 3371 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3372 | } else { |
34ee645e | 3373 | if (cow) { |
0f10851e JG |
3374 | /* |
3375 | * No need to notify as we are downgrading page | |
3376 | * table protection not changing it to point | |
3377 | * to a new page. | |
3378 | * | |
ad56b738 | 3379 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3380 | */ |
7f2e9525 | 3381 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3382 | } |
0253d634 | 3383 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3384 | ptepage = pte_page(entry); |
3385 | get_page(ptepage); | |
53f9263b | 3386 | page_dup_rmap(ptepage, true); |
1c59827d | 3387 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3388 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3389 | } |
cb900f41 KS |
3390 | spin_unlock(src_ptl); |
3391 | spin_unlock(dst_ptl); | |
63551ae0 | 3392 | } |
63551ae0 | 3393 | |
e8569dd2 | 3394 | if (cow) |
ac46d4f3 | 3395 | mmu_notifier_invalidate_range_end(&range); |
e8569dd2 AS |
3396 | |
3397 | return ret; | |
63551ae0 DG |
3398 | } |
3399 | ||
24669e58 AK |
3400 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3401 | unsigned long start, unsigned long end, | |
3402 | struct page *ref_page) | |
63551ae0 DG |
3403 | { |
3404 | struct mm_struct *mm = vma->vm_mm; | |
3405 | unsigned long address; | |
c7546f8f | 3406 | pte_t *ptep; |
63551ae0 | 3407 | pte_t pte; |
cb900f41 | 3408 | spinlock_t *ptl; |
63551ae0 | 3409 | struct page *page; |
a5516438 AK |
3410 | struct hstate *h = hstate_vma(vma); |
3411 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3412 | struct mmu_notifier_range range; |
a5516438 | 3413 | |
63551ae0 | 3414 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3415 | BUG_ON(start & ~huge_page_mask(h)); |
3416 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3417 | |
07e32661 AK |
3418 | /* |
3419 | * This is a hugetlb vma, all the pte entries should point | |
3420 | * to huge page. | |
3421 | */ | |
ed6a7935 | 3422 | tlb_change_page_size(tlb, sz); |
24669e58 | 3423 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3424 | |
3425 | /* | |
3426 | * If sharing possible, alert mmu notifiers of worst case. | |
3427 | */ | |
6f4f13e8 JG |
3428 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3429 | end); | |
ac46d4f3 JG |
3430 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3431 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3432 | address = start; |
569f48b8 | 3433 | for (; address < end; address += sz) { |
7868a208 | 3434 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3435 | if (!ptep) |
c7546f8f DG |
3436 | continue; |
3437 | ||
cb900f41 | 3438 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3439 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3440 | spin_unlock(ptl); | |
dff11abe MK |
3441 | /* |
3442 | * We just unmapped a page of PMDs by clearing a PUD. | |
3443 | * The caller's TLB flush range should cover this area. | |
3444 | */ | |
31d49da5 AK |
3445 | continue; |
3446 | } | |
39dde65c | 3447 | |
6629326b | 3448 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3449 | if (huge_pte_none(pte)) { |
3450 | spin_unlock(ptl); | |
3451 | continue; | |
3452 | } | |
6629326b HD |
3453 | |
3454 | /* | |
9fbc1f63 NH |
3455 | * Migrating hugepage or HWPoisoned hugepage is already |
3456 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3457 | */ |
9fbc1f63 | 3458 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3459 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3460 | spin_unlock(ptl); |
3461 | continue; | |
8c4894c6 | 3462 | } |
6629326b HD |
3463 | |
3464 | page = pte_page(pte); | |
04f2cbe3 MG |
3465 | /* |
3466 | * If a reference page is supplied, it is because a specific | |
3467 | * page is being unmapped, not a range. Ensure the page we | |
3468 | * are about to unmap is the actual page of interest. | |
3469 | */ | |
3470 | if (ref_page) { | |
31d49da5 AK |
3471 | if (page != ref_page) { |
3472 | spin_unlock(ptl); | |
3473 | continue; | |
3474 | } | |
04f2cbe3 MG |
3475 | /* |
3476 | * Mark the VMA as having unmapped its page so that | |
3477 | * future faults in this VMA will fail rather than | |
3478 | * looking like data was lost | |
3479 | */ | |
3480 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3481 | } | |
3482 | ||
c7546f8f | 3483 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3484 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3485 | if (huge_pte_dirty(pte)) |
6649a386 | 3486 | set_page_dirty(page); |
9e81130b | 3487 | |
5d317b2b | 3488 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3489 | page_remove_rmap(page, true); |
31d49da5 | 3490 | |
cb900f41 | 3491 | spin_unlock(ptl); |
e77b0852 | 3492 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3493 | /* |
3494 | * Bail out after unmapping reference page if supplied | |
3495 | */ | |
3496 | if (ref_page) | |
3497 | break; | |
fe1668ae | 3498 | } |
ac46d4f3 | 3499 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 3500 | tlb_end_vma(tlb, vma); |
1da177e4 | 3501 | } |
63551ae0 | 3502 | |
d833352a MG |
3503 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3504 | struct vm_area_struct *vma, unsigned long start, | |
3505 | unsigned long end, struct page *ref_page) | |
3506 | { | |
3507 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3508 | ||
3509 | /* | |
3510 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3511 | * test will fail on a vma being torn down, and not grab a page table | |
3512 | * on its way out. We're lucky that the flag has such an appropriate | |
3513 | * name, and can in fact be safely cleared here. We could clear it | |
3514 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3515 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3516 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3517 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3518 | */ |
3519 | vma->vm_flags &= ~VM_MAYSHARE; | |
3520 | } | |
3521 | ||
502717f4 | 3522 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3523 | unsigned long end, struct page *ref_page) |
502717f4 | 3524 | { |
24669e58 AK |
3525 | struct mm_struct *mm; |
3526 | struct mmu_gather tlb; | |
dff11abe MK |
3527 | unsigned long tlb_start = start; |
3528 | unsigned long tlb_end = end; | |
3529 | ||
3530 | /* | |
3531 | * If shared PMDs were possibly used within this vma range, adjust | |
3532 | * start/end for worst case tlb flushing. | |
3533 | * Note that we can not be sure if PMDs are shared until we try to | |
3534 | * unmap pages. However, we want to make sure TLB flushing covers | |
3535 | * the largest possible range. | |
3536 | */ | |
3537 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
24669e58 AK |
3538 | |
3539 | mm = vma->vm_mm; | |
3540 | ||
dff11abe | 3541 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); |
24669e58 | 3542 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
dff11abe | 3543 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); |
502717f4 KC |
3544 | } |
3545 | ||
04f2cbe3 MG |
3546 | /* |
3547 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3548 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3549 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3550 | * same region. | |
3551 | */ | |
2f4612af DB |
3552 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3553 | struct page *page, unsigned long address) | |
04f2cbe3 | 3554 | { |
7526674d | 3555 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3556 | struct vm_area_struct *iter_vma; |
3557 | struct address_space *mapping; | |
04f2cbe3 MG |
3558 | pgoff_t pgoff; |
3559 | ||
3560 | /* | |
3561 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3562 | * from page cache lookup which is in HPAGE_SIZE units. | |
3563 | */ | |
7526674d | 3564 | address = address & huge_page_mask(h); |
36e4f20a MH |
3565 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3566 | vma->vm_pgoff; | |
93c76a3d | 3567 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3568 | |
4eb2b1dc MG |
3569 | /* |
3570 | * Take the mapping lock for the duration of the table walk. As | |
3571 | * this mapping should be shared between all the VMAs, | |
3572 | * __unmap_hugepage_range() is called as the lock is already held | |
3573 | */ | |
83cde9e8 | 3574 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3575 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3576 | /* Do not unmap the current VMA */ |
3577 | if (iter_vma == vma) | |
3578 | continue; | |
3579 | ||
2f84a899 MG |
3580 | /* |
3581 | * Shared VMAs have their own reserves and do not affect | |
3582 | * MAP_PRIVATE accounting but it is possible that a shared | |
3583 | * VMA is using the same page so check and skip such VMAs. | |
3584 | */ | |
3585 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3586 | continue; | |
3587 | ||
04f2cbe3 MG |
3588 | /* |
3589 | * Unmap the page from other VMAs without their own reserves. | |
3590 | * They get marked to be SIGKILLed if they fault in these | |
3591 | * areas. This is because a future no-page fault on this VMA | |
3592 | * could insert a zeroed page instead of the data existing | |
3593 | * from the time of fork. This would look like data corruption | |
3594 | */ | |
3595 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3596 | unmap_hugepage_range(iter_vma, address, |
3597 | address + huge_page_size(h), page); | |
04f2cbe3 | 3598 | } |
83cde9e8 | 3599 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3600 | } |
3601 | ||
0fe6e20b NH |
3602 | /* |
3603 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3604 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3605 | * cannot race with other handlers or page migration. | |
3606 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3607 | */ |
2b740303 | 3608 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 3609 | unsigned long address, pte_t *ptep, |
3999f52e | 3610 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3611 | { |
3999f52e | 3612 | pte_t pte; |
a5516438 | 3613 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3614 | struct page *old_page, *new_page; |
2b740303 SJ |
3615 | int outside_reserve = 0; |
3616 | vm_fault_t ret = 0; | |
974e6d66 | 3617 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 3618 | struct mmu_notifier_range range; |
1e8f889b | 3619 | |
3999f52e | 3620 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3621 | old_page = pte_page(pte); |
3622 | ||
04f2cbe3 | 3623 | retry_avoidcopy: |
1e8f889b DG |
3624 | /* If no-one else is actually using this page, avoid the copy |
3625 | * and just make the page writable */ | |
37a2140d | 3626 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3627 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 3628 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 3629 | return 0; |
1e8f889b DG |
3630 | } |
3631 | ||
04f2cbe3 MG |
3632 | /* |
3633 | * If the process that created a MAP_PRIVATE mapping is about to | |
3634 | * perform a COW due to a shared page count, attempt to satisfy | |
3635 | * the allocation without using the existing reserves. The pagecache | |
3636 | * page is used to determine if the reserve at this address was | |
3637 | * consumed or not. If reserves were used, a partial faulted mapping | |
3638 | * at the time of fork() could consume its reserves on COW instead | |
3639 | * of the full address range. | |
3640 | */ | |
5944d011 | 3641 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3642 | old_page != pagecache_page) |
3643 | outside_reserve = 1; | |
3644 | ||
09cbfeaf | 3645 | get_page(old_page); |
b76c8cfb | 3646 | |
ad4404a2 DB |
3647 | /* |
3648 | * Drop page table lock as buddy allocator may be called. It will | |
3649 | * be acquired again before returning to the caller, as expected. | |
3650 | */ | |
cb900f41 | 3651 | spin_unlock(ptl); |
5b7a1d40 | 3652 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 3653 | |
2fc39cec | 3654 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3655 | /* |
3656 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3657 | * it is due to references held by a child and an insufficient | |
3658 | * huge page pool. To guarantee the original mappers | |
3659 | * reliability, unmap the page from child processes. The child | |
3660 | * may get SIGKILLed if it later faults. | |
3661 | */ | |
3662 | if (outside_reserve) { | |
09cbfeaf | 3663 | put_page(old_page); |
04f2cbe3 | 3664 | BUG_ON(huge_pte_none(pte)); |
5b7a1d40 | 3665 | unmap_ref_private(mm, vma, old_page, haddr); |
2f4612af DB |
3666 | BUG_ON(huge_pte_none(pte)); |
3667 | spin_lock(ptl); | |
5b7a1d40 | 3668 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
3669 | if (likely(ptep && |
3670 | pte_same(huge_ptep_get(ptep), pte))) | |
3671 | goto retry_avoidcopy; | |
3672 | /* | |
3673 | * race occurs while re-acquiring page table | |
3674 | * lock, and our job is done. | |
3675 | */ | |
3676 | return 0; | |
04f2cbe3 MG |
3677 | } |
3678 | ||
2b740303 | 3679 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 3680 | goto out_release_old; |
1e8f889b DG |
3681 | } |
3682 | ||
0fe6e20b NH |
3683 | /* |
3684 | * When the original hugepage is shared one, it does not have | |
3685 | * anon_vma prepared. | |
3686 | */ | |
44e2aa93 | 3687 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3688 | ret = VM_FAULT_OOM; |
3689 | goto out_release_all; | |
44e2aa93 | 3690 | } |
0fe6e20b | 3691 | |
974e6d66 | 3692 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 3693 | pages_per_huge_page(h)); |
0ed361de | 3694 | __SetPageUptodate(new_page); |
1e8f889b | 3695 | |
7269f999 | 3696 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 3697 | haddr + huge_page_size(h)); |
ac46d4f3 | 3698 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 3699 | |
b76c8cfb | 3700 | /* |
cb900f41 | 3701 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3702 | * before the page tables are altered |
3703 | */ | |
cb900f41 | 3704 | spin_lock(ptl); |
5b7a1d40 | 3705 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 3706 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3707 | ClearPagePrivate(new_page); |
3708 | ||
1e8f889b | 3709 | /* Break COW */ |
5b7a1d40 | 3710 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 3711 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 3712 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 3713 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 3714 | page_remove_rmap(old_page, true); |
5b7a1d40 | 3715 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
cb6acd01 | 3716 | set_page_huge_active(new_page); |
1e8f889b DG |
3717 | /* Make the old page be freed below */ |
3718 | new_page = old_page; | |
3719 | } | |
cb900f41 | 3720 | spin_unlock(ptl); |
ac46d4f3 | 3721 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 3722 | out_release_all: |
5b7a1d40 | 3723 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 3724 | put_page(new_page); |
ad4404a2 | 3725 | out_release_old: |
09cbfeaf | 3726 | put_page(old_page); |
8312034f | 3727 | |
ad4404a2 DB |
3728 | spin_lock(ptl); /* Caller expects lock to be held */ |
3729 | return ret; | |
1e8f889b DG |
3730 | } |
3731 | ||
04f2cbe3 | 3732 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3733 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3734 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3735 | { |
3736 | struct address_space *mapping; | |
e7c4b0bf | 3737 | pgoff_t idx; |
04f2cbe3 MG |
3738 | |
3739 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3740 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3741 | |
3742 | return find_lock_page(mapping, idx); | |
3743 | } | |
3744 | ||
3ae77f43 HD |
3745 | /* |
3746 | * Return whether there is a pagecache page to back given address within VMA. | |
3747 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3748 | */ | |
3749 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3750 | struct vm_area_struct *vma, unsigned long address) |
3751 | { | |
3752 | struct address_space *mapping; | |
3753 | pgoff_t idx; | |
3754 | struct page *page; | |
3755 | ||
3756 | mapping = vma->vm_file->f_mapping; | |
3757 | idx = vma_hugecache_offset(h, vma, address); | |
3758 | ||
3759 | page = find_get_page(mapping, idx); | |
3760 | if (page) | |
3761 | put_page(page); | |
3762 | return page != NULL; | |
3763 | } | |
3764 | ||
ab76ad54 MK |
3765 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3766 | pgoff_t idx) | |
3767 | { | |
3768 | struct inode *inode = mapping->host; | |
3769 | struct hstate *h = hstate_inode(inode); | |
3770 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3771 | ||
3772 | if (err) | |
3773 | return err; | |
3774 | ClearPagePrivate(page); | |
3775 | ||
22146c3c MK |
3776 | /* |
3777 | * set page dirty so that it will not be removed from cache/file | |
3778 | * by non-hugetlbfs specific code paths. | |
3779 | */ | |
3780 | set_page_dirty(page); | |
3781 | ||
ab76ad54 MK |
3782 | spin_lock(&inode->i_lock); |
3783 | inode->i_blocks += blocks_per_huge_page(h); | |
3784 | spin_unlock(&inode->i_lock); | |
3785 | return 0; | |
3786 | } | |
3787 | ||
2b740303 SJ |
3788 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
3789 | struct vm_area_struct *vma, | |
3790 | struct address_space *mapping, pgoff_t idx, | |
3791 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3792 | { |
a5516438 | 3793 | struct hstate *h = hstate_vma(vma); |
2b740303 | 3794 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3795 | int anon_rmap = 0; |
4c887265 | 3796 | unsigned long size; |
4c887265 | 3797 | struct page *page; |
1e8f889b | 3798 | pte_t new_pte; |
cb900f41 | 3799 | spinlock_t *ptl; |
285b8dca | 3800 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 3801 | bool new_page = false; |
4c887265 | 3802 | |
04f2cbe3 MG |
3803 | /* |
3804 | * Currently, we are forced to kill the process in the event the | |
3805 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3806 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3807 | */ |
3808 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3809 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3810 | current->pid); |
04f2cbe3 MG |
3811 | return ret; |
3812 | } | |
3813 | ||
4c887265 | 3814 | /* |
e7c58097 MK |
3815 | * Use page lock to guard against racing truncation |
3816 | * before we get page_table_lock. | |
4c887265 | 3817 | */ |
6bda666a CL |
3818 | retry: |
3819 | page = find_lock_page(mapping, idx); | |
3820 | if (!page) { | |
e7c58097 MK |
3821 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
3822 | if (idx >= size) | |
3823 | goto out; | |
3824 | ||
1a1aad8a MK |
3825 | /* |
3826 | * Check for page in userfault range | |
3827 | */ | |
3828 | if (userfaultfd_missing(vma)) { | |
3829 | u32 hash; | |
3830 | struct vm_fault vmf = { | |
3831 | .vma = vma, | |
285b8dca | 3832 | .address = haddr, |
1a1aad8a MK |
3833 | .flags = flags, |
3834 | /* | |
3835 | * Hard to debug if it ends up being | |
3836 | * used by a callee that assumes | |
3837 | * something about the other | |
3838 | * uninitialized fields... same as in | |
3839 | * memory.c | |
3840 | */ | |
3841 | }; | |
3842 | ||
3843 | /* | |
ddeaab32 MK |
3844 | * hugetlb_fault_mutex must be dropped before |
3845 | * handling userfault. Reacquire after handling | |
3846 | * fault to make calling code simpler. | |
1a1aad8a | 3847 | */ |
188b04a7 | 3848 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
1a1aad8a MK |
3849 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
3850 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3851 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3852 | goto out; | |
3853 | } | |
3854 | ||
285b8dca | 3855 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 3856 | if (IS_ERR(page)) { |
4643d67e MK |
3857 | /* |
3858 | * Returning error will result in faulting task being | |
3859 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
3860 | * tasks from racing to fault in the same page which | |
3861 | * could result in false unable to allocate errors. | |
3862 | * Page migration does not take the fault mutex, but | |
3863 | * does a clear then write of pte's under page table | |
3864 | * lock. Page fault code could race with migration, | |
3865 | * notice the clear pte and try to allocate a page | |
3866 | * here. Before returning error, get ptl and make | |
3867 | * sure there really is no pte entry. | |
3868 | */ | |
3869 | ptl = huge_pte_lock(h, mm, ptep); | |
3870 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
3871 | ret = 0; | |
3872 | spin_unlock(ptl); | |
3873 | goto out; | |
3874 | } | |
3875 | spin_unlock(ptl); | |
2b740303 | 3876 | ret = vmf_error(PTR_ERR(page)); |
6bda666a CL |
3877 | goto out; |
3878 | } | |
47ad8475 | 3879 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3880 | __SetPageUptodate(page); |
cb6acd01 | 3881 | new_page = true; |
ac9b9c66 | 3882 | |
f83a275d | 3883 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3884 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3885 | if (err) { |
3886 | put_page(page); | |
6bda666a CL |
3887 | if (err == -EEXIST) |
3888 | goto retry; | |
3889 | goto out; | |
3890 | } | |
23be7468 | 3891 | } else { |
6bda666a | 3892 | lock_page(page); |
0fe6e20b NH |
3893 | if (unlikely(anon_vma_prepare(vma))) { |
3894 | ret = VM_FAULT_OOM; | |
3895 | goto backout_unlocked; | |
3896 | } | |
409eb8c2 | 3897 | anon_rmap = 1; |
23be7468 | 3898 | } |
0fe6e20b | 3899 | } else { |
998b4382 NH |
3900 | /* |
3901 | * If memory error occurs between mmap() and fault, some process | |
3902 | * don't have hwpoisoned swap entry for errored virtual address. | |
3903 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3904 | */ | |
3905 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3906 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3907 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3908 | goto backout_unlocked; |
3909 | } | |
6bda666a | 3910 | } |
1e8f889b | 3911 | |
57303d80 AW |
3912 | /* |
3913 | * If we are going to COW a private mapping later, we examine the | |
3914 | * pending reservations for this page now. This will ensure that | |
3915 | * any allocations necessary to record that reservation occur outside | |
3916 | * the spinlock. | |
3917 | */ | |
5e911373 | 3918 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 3919 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
3920 | ret = VM_FAULT_OOM; |
3921 | goto backout_unlocked; | |
3922 | } | |
5e911373 | 3923 | /* Just decrements count, does not deallocate */ |
285b8dca | 3924 | vma_end_reservation(h, vma, haddr); |
5e911373 | 3925 | } |
57303d80 | 3926 | |
8bea8052 | 3927 | ptl = huge_pte_lock(h, mm, ptep); |
e7c58097 MK |
3928 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
3929 | if (idx >= size) | |
3930 | goto backout; | |
4c887265 | 3931 | |
83c54070 | 3932 | ret = 0; |
7f2e9525 | 3933 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3934 | goto backout; |
3935 | ||
07443a85 JK |
3936 | if (anon_rmap) { |
3937 | ClearPagePrivate(page); | |
285b8dca | 3938 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 3939 | } else |
53f9263b | 3940 | page_dup_rmap(page, true); |
1e8f889b DG |
3941 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3942 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 3943 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 3944 | |
5d317b2b | 3945 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3946 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3947 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 3948 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
3949 | } |
3950 | ||
cb900f41 | 3951 | spin_unlock(ptl); |
cb6acd01 MK |
3952 | |
3953 | /* | |
3954 | * Only make newly allocated pages active. Existing pages found | |
3955 | * in the pagecache could be !page_huge_active() if they have been | |
3956 | * isolated for migration. | |
3957 | */ | |
3958 | if (new_page) | |
3959 | set_page_huge_active(page); | |
3960 | ||
4c887265 AL |
3961 | unlock_page(page); |
3962 | out: | |
ac9b9c66 | 3963 | return ret; |
4c887265 AL |
3964 | |
3965 | backout: | |
cb900f41 | 3966 | spin_unlock(ptl); |
2b26736c | 3967 | backout_unlocked: |
4c887265 | 3968 | unlock_page(page); |
285b8dca | 3969 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
3970 | put_page(page); |
3971 | goto out; | |
ac9b9c66 HD |
3972 | } |
3973 | ||
8382d914 | 3974 | #ifdef CONFIG_SMP |
188b04a7 | 3975 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
3976 | { |
3977 | unsigned long key[2]; | |
3978 | u32 hash; | |
3979 | ||
1b426bac MK |
3980 | key[0] = (unsigned long) mapping; |
3981 | key[1] = idx; | |
8382d914 | 3982 | |
55254636 | 3983 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
3984 | |
3985 | return hash & (num_fault_mutexes - 1); | |
3986 | } | |
3987 | #else | |
3988 | /* | |
3989 | * For uniprocesor systems we always use a single mutex, so just | |
3990 | * return 0 and avoid the hashing overhead. | |
3991 | */ | |
188b04a7 | 3992 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
3993 | { |
3994 | return 0; | |
3995 | } | |
3996 | #endif | |
3997 | ||
2b740303 | 3998 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3999 | unsigned long address, unsigned int flags) |
86e5216f | 4000 | { |
8382d914 | 4001 | pte_t *ptep, entry; |
cb900f41 | 4002 | spinlock_t *ptl; |
2b740303 | 4003 | vm_fault_t ret; |
8382d914 DB |
4004 | u32 hash; |
4005 | pgoff_t idx; | |
0fe6e20b | 4006 | struct page *page = NULL; |
57303d80 | 4007 | struct page *pagecache_page = NULL; |
a5516438 | 4008 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4009 | struct address_space *mapping; |
0f792cf9 | 4010 | int need_wait_lock = 0; |
285b8dca | 4011 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4012 | |
285b8dca | 4013 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed NH |
4014 | if (ptep) { |
4015 | entry = huge_ptep_get(ptep); | |
290408d4 | 4016 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4017 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4018 | return 0; |
4019 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4020 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4021 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
ddeaab32 MK |
4022 | } else { |
4023 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4024 | if (!ptep) | |
4025 | return VM_FAULT_OOM; | |
fd6a03ed NH |
4026 | } |
4027 | ||
8382d914 | 4028 | mapping = vma->vm_file->f_mapping; |
ddeaab32 | 4029 | idx = vma_hugecache_offset(h, vma, haddr); |
8382d914 | 4030 | |
3935baa9 DG |
4031 | /* |
4032 | * Serialize hugepage allocation and instantiation, so that we don't | |
4033 | * get spurious allocation failures if two CPUs race to instantiate | |
4034 | * the same page in the page cache. | |
4035 | */ | |
188b04a7 | 4036 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4037 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4038 | |
7f2e9525 GS |
4039 | entry = huge_ptep_get(ptep); |
4040 | if (huge_pte_none(entry)) { | |
8382d914 | 4041 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4042 | goto out_mutex; |
3935baa9 | 4043 | } |
86e5216f | 4044 | |
83c54070 | 4045 | ret = 0; |
1e8f889b | 4046 | |
0f792cf9 NH |
4047 | /* |
4048 | * entry could be a migration/hwpoison entry at this point, so this | |
4049 | * check prevents the kernel from going below assuming that we have | |
4050 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
4051 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
4052 | * handle it. | |
4053 | */ | |
4054 | if (!pte_present(entry)) | |
4055 | goto out_mutex; | |
4056 | ||
57303d80 AW |
4057 | /* |
4058 | * If we are going to COW the mapping later, we examine the pending | |
4059 | * reservations for this page now. This will ensure that any | |
4060 | * allocations necessary to record that reservation occur outside the | |
4061 | * spinlock. For private mappings, we also lookup the pagecache | |
4062 | * page now as it is used to determine if a reservation has been | |
4063 | * consumed. | |
4064 | */ | |
106c992a | 4065 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4066 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4067 | ret = VM_FAULT_OOM; |
b4d1d99f | 4068 | goto out_mutex; |
2b26736c | 4069 | } |
5e911373 | 4070 | /* Just decrements count, does not deallocate */ |
285b8dca | 4071 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4072 | |
f83a275d | 4073 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4074 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4075 | vma, haddr); |
57303d80 AW |
4076 | } |
4077 | ||
0f792cf9 NH |
4078 | ptl = huge_pte_lock(h, mm, ptep); |
4079 | ||
4080 | /* Check for a racing update before calling hugetlb_cow */ | |
4081 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4082 | goto out_ptl; | |
4083 | ||
56c9cfb1 NH |
4084 | /* |
4085 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4086 | * pagecache_page, so here we need take the former one | |
4087 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4088 | */ |
4089 | page = pte_page(entry); | |
4090 | if (page != pagecache_page) | |
0f792cf9 NH |
4091 | if (!trylock_page(page)) { |
4092 | need_wait_lock = 1; | |
4093 | goto out_ptl; | |
4094 | } | |
b4d1d99f | 4095 | |
0f792cf9 | 4096 | get_page(page); |
b4d1d99f | 4097 | |
788c7df4 | 4098 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4099 | if (!huge_pte_write(entry)) { |
974e6d66 | 4100 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4101 | pagecache_page, ptl); |
0f792cf9 | 4102 | goto out_put_page; |
b4d1d99f | 4103 | } |
106c992a | 4104 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4105 | } |
4106 | entry = pte_mkyoung(entry); | |
285b8dca | 4107 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4108 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4109 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4110 | out_put_page: |
4111 | if (page != pagecache_page) | |
4112 | unlock_page(page); | |
4113 | put_page(page); | |
cb900f41 KS |
4114 | out_ptl: |
4115 | spin_unlock(ptl); | |
57303d80 AW |
4116 | |
4117 | if (pagecache_page) { | |
4118 | unlock_page(pagecache_page); | |
4119 | put_page(pagecache_page); | |
4120 | } | |
b4d1d99f | 4121 | out_mutex: |
c672c7f2 | 4122 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
4123 | /* |
4124 | * Generally it's safe to hold refcount during waiting page lock. But | |
4125 | * here we just wait to defer the next page fault to avoid busy loop and | |
4126 | * the page is not used after unlocked before returning from the current | |
4127 | * page fault. So we are safe from accessing freed page, even if we wait | |
4128 | * here without taking refcount. | |
4129 | */ | |
4130 | if (need_wait_lock) | |
4131 | wait_on_page_locked(page); | |
1e8f889b | 4132 | return ret; |
86e5216f AL |
4133 | } |
4134 | ||
8fb5debc MK |
4135 | /* |
4136 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4137 | * modifications for huge pages. | |
4138 | */ | |
4139 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4140 | pte_t *dst_pte, | |
4141 | struct vm_area_struct *dst_vma, | |
4142 | unsigned long dst_addr, | |
4143 | unsigned long src_addr, | |
4144 | struct page **pagep) | |
4145 | { | |
1e392147 AA |
4146 | struct address_space *mapping; |
4147 | pgoff_t idx; | |
4148 | unsigned long size; | |
1c9e8def | 4149 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4150 | struct hstate *h = hstate_vma(dst_vma); |
4151 | pte_t _dst_pte; | |
4152 | spinlock_t *ptl; | |
4153 | int ret; | |
4154 | struct page *page; | |
4155 | ||
4156 | if (!*pagep) { | |
4157 | ret = -ENOMEM; | |
4158 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4159 | if (IS_ERR(page)) | |
4160 | goto out; | |
4161 | ||
4162 | ret = copy_huge_page_from_user(page, | |
4163 | (const void __user *) src_addr, | |
810a56b9 | 4164 | pages_per_huge_page(h), false); |
8fb5debc MK |
4165 | |
4166 | /* fallback to copy_from_user outside mmap_sem */ | |
4167 | if (unlikely(ret)) { | |
9e368259 | 4168 | ret = -ENOENT; |
8fb5debc MK |
4169 | *pagep = page; |
4170 | /* don't free the page */ | |
4171 | goto out; | |
4172 | } | |
4173 | } else { | |
4174 | page = *pagep; | |
4175 | *pagep = NULL; | |
4176 | } | |
4177 | ||
4178 | /* | |
4179 | * The memory barrier inside __SetPageUptodate makes sure that | |
4180 | * preceding stores to the page contents become visible before | |
4181 | * the set_pte_at() write. | |
4182 | */ | |
4183 | __SetPageUptodate(page); | |
8fb5debc | 4184 | |
1e392147 AA |
4185 | mapping = dst_vma->vm_file->f_mapping; |
4186 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4187 | ||
1c9e8def MK |
4188 | /* |
4189 | * If shared, add to page cache | |
4190 | */ | |
4191 | if (vm_shared) { | |
1e392147 AA |
4192 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4193 | ret = -EFAULT; | |
4194 | if (idx >= size) | |
4195 | goto out_release_nounlock; | |
1c9e8def | 4196 | |
1e392147 AA |
4197 | /* |
4198 | * Serialization between remove_inode_hugepages() and | |
4199 | * huge_add_to_page_cache() below happens through the | |
4200 | * hugetlb_fault_mutex_table that here must be hold by | |
4201 | * the caller. | |
4202 | */ | |
1c9e8def MK |
4203 | ret = huge_add_to_page_cache(page, mapping, idx); |
4204 | if (ret) | |
4205 | goto out_release_nounlock; | |
4206 | } | |
4207 | ||
8fb5debc MK |
4208 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4209 | spin_lock(ptl); | |
4210 | ||
1e392147 AA |
4211 | /* |
4212 | * Recheck the i_size after holding PT lock to make sure not | |
4213 | * to leave any page mapped (as page_mapped()) beyond the end | |
4214 | * of the i_size (remove_inode_hugepages() is strict about | |
4215 | * enforcing that). If we bail out here, we'll also leave a | |
4216 | * page in the radix tree in the vm_shared case beyond the end | |
4217 | * of the i_size, but remove_inode_hugepages() will take care | |
4218 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4219 | */ | |
4220 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4221 | ret = -EFAULT; | |
4222 | if (idx >= size) | |
4223 | goto out_release_unlock; | |
4224 | ||
8fb5debc MK |
4225 | ret = -EEXIST; |
4226 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4227 | goto out_release_unlock; | |
4228 | ||
1c9e8def MK |
4229 | if (vm_shared) { |
4230 | page_dup_rmap(page, true); | |
4231 | } else { | |
4232 | ClearPagePrivate(page); | |
4233 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4234 | } | |
8fb5debc MK |
4235 | |
4236 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4237 | if (dst_vma->vm_flags & VM_WRITE) | |
4238 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4239 | _dst_pte = pte_mkyoung(_dst_pte); | |
4240 | ||
4241 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4242 | ||
4243 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4244 | dst_vma->vm_flags & VM_WRITE); | |
4245 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4246 | ||
4247 | /* No need to invalidate - it was non-present before */ | |
4248 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4249 | ||
4250 | spin_unlock(ptl); | |
cb6acd01 | 4251 | set_page_huge_active(page); |
1c9e8def MK |
4252 | if (vm_shared) |
4253 | unlock_page(page); | |
8fb5debc MK |
4254 | ret = 0; |
4255 | out: | |
4256 | return ret; | |
4257 | out_release_unlock: | |
4258 | spin_unlock(ptl); | |
1c9e8def MK |
4259 | if (vm_shared) |
4260 | unlock_page(page); | |
5af10dfd | 4261 | out_release_nounlock: |
8fb5debc MK |
4262 | put_page(page); |
4263 | goto out; | |
4264 | } | |
4265 | ||
28a35716 ML |
4266 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4267 | struct page **pages, struct vm_area_struct **vmas, | |
4268 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4269 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4270 | { |
d5d4b0aa KC |
4271 | unsigned long pfn_offset; |
4272 | unsigned long vaddr = *position; | |
28a35716 | 4273 | unsigned long remainder = *nr_pages; |
a5516438 | 4274 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4275 | int err = -EFAULT; |
63551ae0 | 4276 | |
63551ae0 | 4277 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4278 | pte_t *pte; |
cb900f41 | 4279 | spinlock_t *ptl = NULL; |
2a15efc9 | 4280 | int absent; |
4c887265 | 4281 | struct page *page; |
63551ae0 | 4282 | |
02057967 DR |
4283 | /* |
4284 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4285 | * potentially allocating memory. | |
4286 | */ | |
fa45f116 | 4287 | if (fatal_signal_pending(current)) { |
02057967 DR |
4288 | remainder = 0; |
4289 | break; | |
4290 | } | |
4291 | ||
4c887265 AL |
4292 | /* |
4293 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4294 | * each hugepage. We have to make sure we get the |
4c887265 | 4295 | * first, for the page indexing below to work. |
cb900f41 KS |
4296 | * |
4297 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4298 | */ |
7868a208 PA |
4299 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4300 | huge_page_size(h)); | |
cb900f41 KS |
4301 | if (pte) |
4302 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4303 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4304 | ||
4305 | /* | |
4306 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4307 | * an error where there's an empty slot with no huge pagecache |
4308 | * to back it. This way, we avoid allocating a hugepage, and | |
4309 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4310 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4311 | */ |
3ae77f43 HD |
4312 | if (absent && (flags & FOLL_DUMP) && |
4313 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4314 | if (pte) |
4315 | spin_unlock(ptl); | |
2a15efc9 HD |
4316 | remainder = 0; |
4317 | break; | |
4318 | } | |
63551ae0 | 4319 | |
9cc3a5bd NH |
4320 | /* |
4321 | * We need call hugetlb_fault for both hugepages under migration | |
4322 | * (in which case hugetlb_fault waits for the migration,) and | |
4323 | * hwpoisoned hugepages (in which case we need to prevent the | |
4324 | * caller from accessing to them.) In order to do this, we use | |
4325 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4326 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4327 | * both cases, and because we can't follow correct pages | |
4328 | * directly from any kind of swap entries. | |
4329 | */ | |
4330 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4331 | ((flags & FOLL_WRITE) && |
4332 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4333 | vm_fault_t ret; |
87ffc118 | 4334 | unsigned int fault_flags = 0; |
63551ae0 | 4335 | |
cb900f41 KS |
4336 | if (pte) |
4337 | spin_unlock(ptl); | |
87ffc118 AA |
4338 | if (flags & FOLL_WRITE) |
4339 | fault_flags |= FAULT_FLAG_WRITE; | |
4340 | if (nonblocking) | |
4341 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4342 | if (flags & FOLL_NOWAIT) | |
4343 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4344 | FAULT_FLAG_RETRY_NOWAIT; | |
4345 | if (flags & FOLL_TRIED) { | |
4346 | VM_WARN_ON_ONCE(fault_flags & | |
4347 | FAULT_FLAG_ALLOW_RETRY); | |
4348 | fault_flags |= FAULT_FLAG_TRIED; | |
4349 | } | |
4350 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4351 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4352 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4353 | remainder = 0; |
4354 | break; | |
4355 | } | |
4356 | if (ret & VM_FAULT_RETRY) { | |
1ac25013 AA |
4357 | if (nonblocking && |
4358 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) | |
87ffc118 AA |
4359 | *nonblocking = 0; |
4360 | *nr_pages = 0; | |
4361 | /* | |
4362 | * VM_FAULT_RETRY must not return an | |
4363 | * error, it will return zero | |
4364 | * instead. | |
4365 | * | |
4366 | * No need to update "position" as the | |
4367 | * caller will not check it after | |
4368 | * *nr_pages is set to 0. | |
4369 | */ | |
4370 | return i; | |
4371 | } | |
4372 | continue; | |
4c887265 AL |
4373 | } |
4374 | ||
a5516438 | 4375 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4376 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 4377 | |
acbfb087 ZL |
4378 | /* |
4379 | * If subpage information not requested, update counters | |
4380 | * and skip the same_page loop below. | |
4381 | */ | |
4382 | if (!pages && !vmas && !pfn_offset && | |
4383 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4384 | (remainder >= pages_per_huge_page(h))) { | |
4385 | vaddr += huge_page_size(h); | |
4386 | remainder -= pages_per_huge_page(h); | |
4387 | i += pages_per_huge_page(h); | |
4388 | spin_unlock(ptl); | |
4389 | continue; | |
4390 | } | |
4391 | ||
d5d4b0aa | 4392 | same_page: |
d6692183 | 4393 | if (pages) { |
2a15efc9 | 4394 | pages[i] = mem_map_offset(page, pfn_offset); |
3faa52c0 JH |
4395 | /* |
4396 | * try_grab_page() should always succeed here, because: | |
4397 | * a) we hold the ptl lock, and b) we've just checked | |
4398 | * that the huge page is present in the page tables. If | |
4399 | * the huge page is present, then the tail pages must | |
4400 | * also be present. The ptl prevents the head page and | |
4401 | * tail pages from being rearranged in any way. So this | |
4402 | * page must be available at this point, unless the page | |
4403 | * refcount overflowed: | |
4404 | */ | |
4405 | if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { | |
4406 | spin_unlock(ptl); | |
4407 | remainder = 0; | |
4408 | err = -ENOMEM; | |
4409 | break; | |
4410 | } | |
d6692183 | 4411 | } |
63551ae0 DG |
4412 | |
4413 | if (vmas) | |
4414 | vmas[i] = vma; | |
4415 | ||
4416 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4417 | ++pfn_offset; |
63551ae0 DG |
4418 | --remainder; |
4419 | ++i; | |
d5d4b0aa | 4420 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4421 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4422 | /* |
4423 | * We use pfn_offset to avoid touching the pageframes | |
4424 | * of this compound page. | |
4425 | */ | |
4426 | goto same_page; | |
4427 | } | |
cb900f41 | 4428 | spin_unlock(ptl); |
63551ae0 | 4429 | } |
28a35716 | 4430 | *nr_pages = remainder; |
87ffc118 AA |
4431 | /* |
4432 | * setting position is actually required only if remainder is | |
4433 | * not zero but it's faster not to add a "if (remainder)" | |
4434 | * branch. | |
4435 | */ | |
63551ae0 DG |
4436 | *position = vaddr; |
4437 | ||
2be7cfed | 4438 | return i ? i : err; |
63551ae0 | 4439 | } |
8f860591 | 4440 | |
5491ae7b AK |
4441 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4442 | /* | |
4443 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4444 | * implement this. | |
4445 | */ | |
4446 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4447 | #endif | |
4448 | ||
7da4d641 | 4449 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4450 | unsigned long address, unsigned long end, pgprot_t newprot) |
4451 | { | |
4452 | struct mm_struct *mm = vma->vm_mm; | |
4453 | unsigned long start = address; | |
4454 | pte_t *ptep; | |
4455 | pte_t pte; | |
a5516438 | 4456 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4457 | unsigned long pages = 0; |
dff11abe | 4458 | bool shared_pmd = false; |
ac46d4f3 | 4459 | struct mmu_notifier_range range; |
dff11abe MK |
4460 | |
4461 | /* | |
4462 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 4463 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
4464 | * range if PMD sharing is possible. |
4465 | */ | |
7269f999 JG |
4466 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
4467 | 0, vma, mm, start, end); | |
ac46d4f3 | 4468 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
4469 | |
4470 | BUG_ON(address >= end); | |
ac46d4f3 | 4471 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 4472 | |
ac46d4f3 | 4473 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 4474 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4475 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4476 | spinlock_t *ptl; |
7868a208 | 4477 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4478 | if (!ptep) |
4479 | continue; | |
cb900f41 | 4480 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4481 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4482 | pages++; | |
cb900f41 | 4483 | spin_unlock(ptl); |
dff11abe | 4484 | shared_pmd = true; |
39dde65c | 4485 | continue; |
7da4d641 | 4486 | } |
a8bda28d NH |
4487 | pte = huge_ptep_get(ptep); |
4488 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4489 | spin_unlock(ptl); | |
4490 | continue; | |
4491 | } | |
4492 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4493 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4494 | ||
4495 | if (is_write_migration_entry(entry)) { | |
4496 | pte_t newpte; | |
4497 | ||
4498 | make_migration_entry_read(&entry); | |
4499 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4500 | set_huge_swap_pte_at(mm, address, ptep, |
4501 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4502 | pages++; |
4503 | } | |
4504 | spin_unlock(ptl); | |
4505 | continue; | |
4506 | } | |
4507 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
4508 | pte_t old_pte; |
4509 | ||
4510 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
4511 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 4512 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 4513 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 4514 | pages++; |
8f860591 | 4515 | } |
cb900f41 | 4516 | spin_unlock(ptl); |
8f860591 | 4517 | } |
d833352a | 4518 | /* |
c8c06efa | 4519 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4520 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4521 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
4522 | * and that page table be reused and filled with junk. If we actually |
4523 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 4524 | */ |
dff11abe | 4525 | if (shared_pmd) |
ac46d4f3 | 4526 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
4527 | else |
4528 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
4529 | /* |
4530 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4531 | * page table protection not changing it to point to a new page. | |
4532 | * | |
ad56b738 | 4533 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4534 | */ |
83cde9e8 | 4535 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 4536 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
4537 | |
4538 | return pages << h->order; | |
8f860591 ZY |
4539 | } |
4540 | ||
a1e78772 MG |
4541 | int hugetlb_reserve_pages(struct inode *inode, |
4542 | long from, long to, | |
5a6fe125 | 4543 | struct vm_area_struct *vma, |
ca16d140 | 4544 | vm_flags_t vm_flags) |
e4e574b7 | 4545 | { |
17c9d12e | 4546 | long ret, chg; |
a5516438 | 4547 | struct hstate *h = hstate_inode(inode); |
90481622 | 4548 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4549 | struct resv_map *resv_map; |
1c5ecae3 | 4550 | long gbl_reserve; |
e4e574b7 | 4551 | |
63489f8e MK |
4552 | /* This should never happen */ |
4553 | if (from > to) { | |
4554 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
4555 | return -EINVAL; | |
4556 | } | |
4557 | ||
17c9d12e MG |
4558 | /* |
4559 | * Only apply hugepage reservation if asked. At fault time, an | |
4560 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4561 | * without using reserves |
17c9d12e | 4562 | */ |
ca16d140 | 4563 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4564 | return 0; |
4565 | ||
a1e78772 MG |
4566 | /* |
4567 | * Shared mappings base their reservation on the number of pages that | |
4568 | * are already allocated on behalf of the file. Private mappings need | |
4569 | * to reserve the full area even if read-only as mprotect() may be | |
4570 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4571 | */ | |
9119a41e | 4572 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
4573 | /* |
4574 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
4575 | * called for inodes for which resv_maps were created (see | |
4576 | * hugetlbfs_get_inode). | |
4577 | */ | |
4e35f483 | 4578 | resv_map = inode_resv_map(inode); |
9119a41e | 4579 | |
1406ec9b | 4580 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4581 | |
4582 | } else { | |
4583 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4584 | if (!resv_map) |
4585 | return -ENOMEM; | |
4586 | ||
a1e78772 | 4587 | chg = to - from; |
84afd99b | 4588 | |
17c9d12e MG |
4589 | set_vma_resv_map(vma, resv_map); |
4590 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4591 | } | |
4592 | ||
c50ac050 DH |
4593 | if (chg < 0) { |
4594 | ret = chg; | |
4595 | goto out_err; | |
4596 | } | |
8a630112 | 4597 | |
1c5ecae3 MK |
4598 | /* |
4599 | * There must be enough pages in the subpool for the mapping. If | |
4600 | * the subpool has a minimum size, there may be some global | |
4601 | * reservations already in place (gbl_reserve). | |
4602 | */ | |
4603 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4604 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4605 | ret = -ENOSPC; |
4606 | goto out_err; | |
4607 | } | |
5a6fe125 MG |
4608 | |
4609 | /* | |
17c9d12e | 4610 | * Check enough hugepages are available for the reservation. |
90481622 | 4611 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4612 | */ |
1c5ecae3 | 4613 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4614 | if (ret < 0) { |
1c5ecae3 MK |
4615 | /* put back original number of pages, chg */ |
4616 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4617 | goto out_err; |
68842c9b | 4618 | } |
17c9d12e MG |
4619 | |
4620 | /* | |
4621 | * Account for the reservations made. Shared mappings record regions | |
4622 | * that have reservations as they are shared by multiple VMAs. | |
4623 | * When the last VMA disappears, the region map says how much | |
4624 | * the reservation was and the page cache tells how much of | |
4625 | * the reservation was consumed. Private mappings are per-VMA and | |
4626 | * only the consumed reservations are tracked. When the VMA | |
4627 | * disappears, the original reservation is the VMA size and the | |
4628 | * consumed reservations are stored in the map. Hence, nothing | |
4629 | * else has to be done for private mappings here | |
4630 | */ | |
33039678 MK |
4631 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4632 | long add = region_add(resv_map, from, to); | |
4633 | ||
4634 | if (unlikely(chg > add)) { | |
4635 | /* | |
4636 | * pages in this range were added to the reserve | |
4637 | * map between region_chg and region_add. This | |
4638 | * indicates a race with alloc_huge_page. Adjust | |
4639 | * the subpool and reserve counts modified above | |
4640 | * based on the difference. | |
4641 | */ | |
4642 | long rsv_adjust; | |
4643 | ||
4644 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4645 | chg - add); | |
4646 | hugetlb_acct_memory(h, -rsv_adjust); | |
4647 | } | |
4648 | } | |
a43a8c39 | 4649 | return 0; |
c50ac050 | 4650 | out_err: |
5e911373 | 4651 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4652 | /* Don't call region_abort if region_chg failed */ |
4653 | if (chg >= 0) | |
4654 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4655 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4656 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4657 | return ret; |
a43a8c39 KC |
4658 | } |
4659 | ||
b5cec28d MK |
4660 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4661 | long freed) | |
a43a8c39 | 4662 | { |
a5516438 | 4663 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4664 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4665 | long chg = 0; |
90481622 | 4666 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4667 | long gbl_reserve; |
45c682a6 | 4668 | |
f27a5136 MK |
4669 | /* |
4670 | * Since this routine can be called in the evict inode path for all | |
4671 | * hugetlbfs inodes, resv_map could be NULL. | |
4672 | */ | |
b5cec28d MK |
4673 | if (resv_map) { |
4674 | chg = region_del(resv_map, start, end); | |
4675 | /* | |
4676 | * region_del() can fail in the rare case where a region | |
4677 | * must be split and another region descriptor can not be | |
4678 | * allocated. If end == LONG_MAX, it will not fail. | |
4679 | */ | |
4680 | if (chg < 0) | |
4681 | return chg; | |
4682 | } | |
4683 | ||
45c682a6 | 4684 | spin_lock(&inode->i_lock); |
e4c6f8be | 4685 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4686 | spin_unlock(&inode->i_lock); |
4687 | ||
1c5ecae3 MK |
4688 | /* |
4689 | * If the subpool has a minimum size, the number of global | |
4690 | * reservations to be released may be adjusted. | |
4691 | */ | |
4692 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4693 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4694 | |
4695 | return 0; | |
a43a8c39 | 4696 | } |
93f70f90 | 4697 | |
3212b535 SC |
4698 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4699 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4700 | struct vm_area_struct *vma, | |
4701 | unsigned long addr, pgoff_t idx) | |
4702 | { | |
4703 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4704 | svma->vm_start; | |
4705 | unsigned long sbase = saddr & PUD_MASK; | |
4706 | unsigned long s_end = sbase + PUD_SIZE; | |
4707 | ||
4708 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4709 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4710 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4711 | |
4712 | /* | |
4713 | * match the virtual addresses, permission and the alignment of the | |
4714 | * page table page. | |
4715 | */ | |
4716 | if (pmd_index(addr) != pmd_index(saddr) || | |
4717 | vm_flags != svm_flags || | |
4718 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4719 | return 0; | |
4720 | ||
4721 | return saddr; | |
4722 | } | |
4723 | ||
31aafb45 | 4724 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4725 | { |
4726 | unsigned long base = addr & PUD_MASK; | |
4727 | unsigned long end = base + PUD_SIZE; | |
4728 | ||
4729 | /* | |
4730 | * check on proper vm_flags and page table alignment | |
4731 | */ | |
017b1660 | 4732 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
4733 | return true; |
4734 | return false; | |
3212b535 SC |
4735 | } |
4736 | ||
017b1660 MK |
4737 | /* |
4738 | * Determine if start,end range within vma could be mapped by shared pmd. | |
4739 | * If yes, adjust start and end to cover range associated with possible | |
4740 | * shared pmd mappings. | |
4741 | */ | |
4742 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
4743 | unsigned long *start, unsigned long *end) | |
4744 | { | |
4745 | unsigned long check_addr = *start; | |
4746 | ||
4747 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
4748 | return; | |
4749 | ||
4750 | for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { | |
4751 | unsigned long a_start = check_addr & PUD_MASK; | |
4752 | unsigned long a_end = a_start + PUD_SIZE; | |
4753 | ||
4754 | /* | |
4755 | * If sharing is possible, adjust start/end if necessary. | |
4756 | */ | |
4757 | if (range_in_vma(vma, a_start, a_end)) { | |
4758 | if (a_start < *start) | |
4759 | *start = a_start; | |
4760 | if (a_end > *end) | |
4761 | *end = a_end; | |
4762 | } | |
4763 | } | |
4764 | } | |
4765 | ||
3212b535 SC |
4766 | /* |
4767 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4768 | * and returns the corresponding pte. While this is not necessary for the | |
4769 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
ddeaab32 MK |
4770 | * code much cleaner. pmd allocation is essential for the shared case because |
4771 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise | |
4772 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
4773 | * bad pmd for sharing. | |
3212b535 SC |
4774 | */ |
4775 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4776 | { | |
4777 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4778 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4779 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4780 | vma->vm_pgoff; | |
4781 | struct vm_area_struct *svma; | |
4782 | unsigned long saddr; | |
4783 | pte_t *spte = NULL; | |
4784 | pte_t *pte; | |
cb900f41 | 4785 | spinlock_t *ptl; |
3212b535 SC |
4786 | |
4787 | if (!vma_shareable(vma, addr)) | |
4788 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4789 | ||
930668c3 | 4790 | i_mmap_lock_read(mapping); |
3212b535 SC |
4791 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4792 | if (svma == vma) | |
4793 | continue; | |
4794 | ||
4795 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4796 | if (saddr) { | |
7868a208 PA |
4797 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4798 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4799 | if (spte) { |
4800 | get_page(virt_to_page(spte)); | |
4801 | break; | |
4802 | } | |
4803 | } | |
4804 | } | |
4805 | ||
4806 | if (!spte) | |
4807 | goto out; | |
4808 | ||
8bea8052 | 4809 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4810 | if (pud_none(*pud)) { |
3212b535 SC |
4811 | pud_populate(mm, pud, |
4812 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4813 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4814 | } else { |
3212b535 | 4815 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4816 | } |
cb900f41 | 4817 | spin_unlock(ptl); |
3212b535 SC |
4818 | out: |
4819 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
930668c3 | 4820 | i_mmap_unlock_read(mapping); |
3212b535 SC |
4821 | return pte; |
4822 | } | |
4823 | ||
4824 | /* | |
4825 | * unmap huge page backed by shared pte. | |
4826 | * | |
4827 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4828 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4829 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4830 | * | |
ddeaab32 | 4831 | * called with page table lock held. |
3212b535 SC |
4832 | * |
4833 | * returns: 1 successfully unmapped a shared pte page | |
4834 | * 0 the underlying pte page is not shared, or it is the last user | |
4835 | */ | |
4836 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4837 | { | |
4838 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4839 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4840 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4841 | |
4842 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4843 | if (page_count(virt_to_page(ptep)) == 1) | |
4844 | return 0; | |
4845 | ||
4846 | pud_clear(pud); | |
4847 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4848 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4849 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4850 | return 1; | |
4851 | } | |
9e5fc74c SC |
4852 | #define want_pmd_share() (1) |
4853 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4854 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4855 | { | |
4856 | return NULL; | |
4857 | } | |
e81f2d22 ZZ |
4858 | |
4859 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4860 | { | |
4861 | return 0; | |
4862 | } | |
017b1660 MK |
4863 | |
4864 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
4865 | unsigned long *start, unsigned long *end) | |
4866 | { | |
4867 | } | |
9e5fc74c | 4868 | #define want_pmd_share() (0) |
3212b535 SC |
4869 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4870 | ||
9e5fc74c SC |
4871 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4872 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4873 | unsigned long addr, unsigned long sz) | |
4874 | { | |
4875 | pgd_t *pgd; | |
c2febafc | 4876 | p4d_t *p4d; |
9e5fc74c SC |
4877 | pud_t *pud; |
4878 | pte_t *pte = NULL; | |
4879 | ||
4880 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
4881 | p4d = p4d_alloc(mm, pgd, addr); |
4882 | if (!p4d) | |
4883 | return NULL; | |
c2febafc | 4884 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
4885 | if (pud) { |
4886 | if (sz == PUD_SIZE) { | |
4887 | pte = (pte_t *)pud; | |
4888 | } else { | |
4889 | BUG_ON(sz != PMD_SIZE); | |
4890 | if (want_pmd_share() && pud_none(*pud)) | |
4891 | pte = huge_pmd_share(mm, addr, pud); | |
4892 | else | |
4893 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4894 | } | |
4895 | } | |
4e666314 | 4896 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
4897 | |
4898 | return pte; | |
4899 | } | |
4900 | ||
9b19df29 PA |
4901 | /* |
4902 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
4903 | * entry at address @addr | |
4904 | * | |
4905 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
4906 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
4907 | * size @sz doesn't match the hugepage size at this level of the page | |
4908 | * table. | |
4909 | */ | |
7868a208 PA |
4910 | pte_t *huge_pte_offset(struct mm_struct *mm, |
4911 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
4912 | { |
4913 | pgd_t *pgd; | |
c2febafc | 4914 | p4d_t *p4d; |
9e5fc74c | 4915 | pud_t *pud; |
c2febafc | 4916 | pmd_t *pmd; |
9e5fc74c SC |
4917 | |
4918 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4919 | if (!pgd_present(*pgd)) |
4920 | return NULL; | |
4921 | p4d = p4d_offset(pgd, addr); | |
4922 | if (!p4d_present(*p4d)) | |
4923 | return NULL; | |
9b19df29 | 4924 | |
c2febafc | 4925 | pud = pud_offset(p4d, addr); |
9b19df29 | 4926 | if (sz != PUD_SIZE && pud_none(*pud)) |
c2febafc | 4927 | return NULL; |
9b19df29 PA |
4928 | /* hugepage or swap? */ |
4929 | if (pud_huge(*pud) || !pud_present(*pud)) | |
c2febafc | 4930 | return (pte_t *)pud; |
9b19df29 | 4931 | |
c2febafc | 4932 | pmd = pmd_offset(pud, addr); |
9b19df29 PA |
4933 | if (sz != PMD_SIZE && pmd_none(*pmd)) |
4934 | return NULL; | |
4935 | /* hugepage or swap? */ | |
4936 | if (pmd_huge(*pmd) || !pmd_present(*pmd)) | |
4937 | return (pte_t *)pmd; | |
4938 | ||
4939 | return NULL; | |
9e5fc74c SC |
4940 | } |
4941 | ||
61f77eda NH |
4942 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4943 | ||
4944 | /* | |
4945 | * These functions are overwritable if your architecture needs its own | |
4946 | * behavior. | |
4947 | */ | |
4948 | struct page * __weak | |
4949 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4950 | int write) | |
4951 | { | |
4952 | return ERR_PTR(-EINVAL); | |
4953 | } | |
4954 | ||
4dc71451 AK |
4955 | struct page * __weak |
4956 | follow_huge_pd(struct vm_area_struct *vma, | |
4957 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
4958 | { | |
4959 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
4960 | return NULL; | |
4961 | } | |
4962 | ||
61f77eda | 4963 | struct page * __weak |
9e5fc74c | 4964 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4965 | pmd_t *pmd, int flags) |
9e5fc74c | 4966 | { |
e66f17ff NH |
4967 | struct page *page = NULL; |
4968 | spinlock_t *ptl; | |
c9d398fa | 4969 | pte_t pte; |
3faa52c0 JH |
4970 | |
4971 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
4972 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
4973 | (FOLL_PIN | FOLL_GET))) | |
4974 | return NULL; | |
4975 | ||
e66f17ff NH |
4976 | retry: |
4977 | ptl = pmd_lockptr(mm, pmd); | |
4978 | spin_lock(ptl); | |
4979 | /* | |
4980 | * make sure that the address range covered by this pmd is not | |
4981 | * unmapped from other threads. | |
4982 | */ | |
4983 | if (!pmd_huge(*pmd)) | |
4984 | goto out; | |
c9d398fa NH |
4985 | pte = huge_ptep_get((pte_t *)pmd); |
4986 | if (pte_present(pte)) { | |
97534127 | 4987 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
4988 | /* |
4989 | * try_grab_page() should always succeed here, because: a) we | |
4990 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
4991 | * huge pmd (head) page is present in the page tables. The ptl | |
4992 | * prevents the head page and tail pages from being rearranged | |
4993 | * in any way. So this page must be available at this point, | |
4994 | * unless the page refcount overflowed: | |
4995 | */ | |
4996 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
4997 | page = NULL; | |
4998 | goto out; | |
4999 | } | |
e66f17ff | 5000 | } else { |
c9d398fa | 5001 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5002 | spin_unlock(ptl); |
5003 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5004 | goto retry; | |
5005 | } | |
5006 | /* | |
5007 | * hwpoisoned entry is treated as no_page_table in | |
5008 | * follow_page_mask(). | |
5009 | */ | |
5010 | } | |
5011 | out: | |
5012 | spin_unlock(ptl); | |
9e5fc74c SC |
5013 | return page; |
5014 | } | |
5015 | ||
61f77eda | 5016 | struct page * __weak |
9e5fc74c | 5017 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5018 | pud_t *pud, int flags) |
9e5fc74c | 5019 | { |
3faa52c0 | 5020 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5021 | return NULL; |
9e5fc74c | 5022 | |
e66f17ff | 5023 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5024 | } |
5025 | ||
faaa5b62 AK |
5026 | struct page * __weak |
5027 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5028 | { | |
3faa52c0 | 5029 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5030 | return NULL; |
5031 | ||
5032 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5033 | } | |
5034 | ||
31caf665 NH |
5035 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5036 | { | |
bcc54222 NH |
5037 | bool ret = true; |
5038 | ||
309381fe | 5039 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5040 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
5041 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
5042 | ret = false; | |
5043 | goto unlock; | |
5044 | } | |
5045 | clear_page_huge_active(page); | |
31caf665 | 5046 | list_move_tail(&page->lru, list); |
bcc54222 | 5047 | unlock: |
31caf665 | 5048 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5049 | return ret; |
31caf665 NH |
5050 | } |
5051 | ||
5052 | void putback_active_hugepage(struct page *page) | |
5053 | { | |
309381fe | 5054 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5055 | spin_lock(&hugetlb_lock); |
bcc54222 | 5056 | set_page_huge_active(page); |
31caf665 NH |
5057 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5058 | spin_unlock(&hugetlb_lock); | |
5059 | put_page(page); | |
5060 | } | |
ab5ac90a MH |
5061 | |
5062 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5063 | { | |
5064 | struct hstate *h = page_hstate(oldpage); | |
5065 | ||
5066 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5067 | set_page_owner_migrate_reason(newpage, reason); | |
5068 | ||
5069 | /* | |
5070 | * transfer temporary state of the new huge page. This is | |
5071 | * reverse to other transitions because the newpage is going to | |
5072 | * be final while the old one will be freed so it takes over | |
5073 | * the temporary status. | |
5074 | * | |
5075 | * Also note that we have to transfer the per-node surplus state | |
5076 | * here as well otherwise the global surplus count will not match | |
5077 | * the per-node's. | |
5078 | */ | |
5079 | if (PageHugeTemporary(newpage)) { | |
5080 | int old_nid = page_to_nid(oldpage); | |
5081 | int new_nid = page_to_nid(newpage); | |
5082 | ||
5083 | SetPageHugeTemporary(oldpage); | |
5084 | ClearPageHugeTemporary(newpage); | |
5085 | ||
5086 | spin_lock(&hugetlb_lock); | |
5087 | if (h->surplus_huge_pages_node[old_nid]) { | |
5088 | h->surplus_huge_pages_node[old_nid]--; | |
5089 | h->surplus_huge_pages_node[new_nid]++; | |
5090 | } | |
5091 | spin_unlock(&hugetlb_lock); | |
5092 | } | |
5093 | } |