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> |
d6606683 | 30 | |
63551ae0 DG |
31 | #include <asm/page.h> |
32 | #include <asm/pgtable.h> | |
24669e58 | 33 | #include <asm/tlb.h> |
63551ae0 | 34 | |
24669e58 | 35 | #include <linux/io.h> |
63551ae0 | 36 | #include <linux/hugetlb.h> |
9dd540e2 | 37 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 38 | #include <linux/node.h> |
1a1aad8a | 39 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 40 | #include <linux/page_owner.h> |
7835e98b | 41 | #include "internal.h" |
1da177e4 | 42 | |
c3f38a38 | 43 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
44 | unsigned int default_hstate_idx; |
45 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
46 | /* |
47 | * Minimum page order among possible hugepage sizes, set to a proper value | |
48 | * at boot time. | |
49 | */ | |
50 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 51 | |
53ba51d2 JT |
52 | __initdata LIST_HEAD(huge_boot_pages); |
53 | ||
e5ff2159 AK |
54 | /* for command line parsing */ |
55 | static struct hstate * __initdata parsed_hstate; | |
56 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 57 | static unsigned long __initdata default_hstate_size; |
9fee021d | 58 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 59 | |
3935baa9 | 60 | /* |
31caf665 NH |
61 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
62 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 63 | */ |
c3f38a38 | 64 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 65 | |
8382d914 DB |
66 | /* |
67 | * Serializes faults on the same logical page. This is used to | |
68 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
69 | */ | |
70 | static int num_fault_mutexes; | |
c672c7f2 | 71 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 72 | |
7ca02d0a MK |
73 | /* Forward declaration */ |
74 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
75 | ||
90481622 DG |
76 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
77 | { | |
78 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
79 | ||
80 | spin_unlock(&spool->lock); | |
81 | ||
82 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
83 | * remain, give up any reservations mased on minimum size and |
84 | * free the subpool */ | |
85 | if (free) { | |
86 | if (spool->min_hpages != -1) | |
87 | hugetlb_acct_memory(spool->hstate, | |
88 | -spool->min_hpages); | |
90481622 | 89 | kfree(spool); |
7ca02d0a | 90 | } |
90481622 DG |
91 | } |
92 | ||
7ca02d0a MK |
93 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
94 | long min_hpages) | |
90481622 DG |
95 | { |
96 | struct hugepage_subpool *spool; | |
97 | ||
c6a91820 | 98 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
99 | if (!spool) |
100 | return NULL; | |
101 | ||
102 | spin_lock_init(&spool->lock); | |
103 | spool->count = 1; | |
7ca02d0a MK |
104 | spool->max_hpages = max_hpages; |
105 | spool->hstate = h; | |
106 | spool->min_hpages = min_hpages; | |
107 | ||
108 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
109 | kfree(spool); | |
110 | return NULL; | |
111 | } | |
112 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
113 | |
114 | return spool; | |
115 | } | |
116 | ||
117 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
118 | { | |
119 | spin_lock(&spool->lock); | |
120 | BUG_ON(!spool->count); | |
121 | spool->count--; | |
122 | unlock_or_release_subpool(spool); | |
123 | } | |
124 | ||
1c5ecae3 MK |
125 | /* |
126 | * Subpool accounting for allocating and reserving pages. | |
127 | * Return -ENOMEM if there are not enough resources to satisfy the | |
128 | * the request. Otherwise, return the number of pages by which the | |
129 | * global pools must be adjusted (upward). The returned value may | |
130 | * only be different than the passed value (delta) in the case where | |
131 | * a subpool minimum size must be manitained. | |
132 | */ | |
133 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
134 | long delta) |
135 | { | |
1c5ecae3 | 136 | long ret = delta; |
90481622 DG |
137 | |
138 | if (!spool) | |
1c5ecae3 | 139 | return ret; |
90481622 DG |
140 | |
141 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
142 | |
143 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
144 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
145 | spool->used_hpages += delta; | |
146 | else { | |
147 | ret = -ENOMEM; | |
148 | goto unlock_ret; | |
149 | } | |
90481622 | 150 | } |
90481622 | 151 | |
09a95e29 MK |
152 | /* minimum size accounting */ |
153 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
154 | if (delta > spool->rsv_hpages) { |
155 | /* | |
156 | * Asking for more reserves than those already taken on | |
157 | * behalf of subpool. Return difference. | |
158 | */ | |
159 | ret = delta - spool->rsv_hpages; | |
160 | spool->rsv_hpages = 0; | |
161 | } else { | |
162 | ret = 0; /* reserves already accounted for */ | |
163 | spool->rsv_hpages -= delta; | |
164 | } | |
165 | } | |
166 | ||
167 | unlock_ret: | |
168 | spin_unlock(&spool->lock); | |
90481622 DG |
169 | return ret; |
170 | } | |
171 | ||
1c5ecae3 MK |
172 | /* |
173 | * Subpool accounting for freeing and unreserving pages. | |
174 | * Return the number of global page reservations that must be dropped. | |
175 | * The return value may only be different than the passed value (delta) | |
176 | * in the case where a subpool minimum size must be maintained. | |
177 | */ | |
178 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
179 | long delta) |
180 | { | |
1c5ecae3 MK |
181 | long ret = delta; |
182 | ||
90481622 | 183 | if (!spool) |
1c5ecae3 | 184 | return delta; |
90481622 DG |
185 | |
186 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
187 | |
188 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
189 | spool->used_hpages -= delta; | |
190 | ||
09a95e29 MK |
191 | /* minimum size accounting */ |
192 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
193 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
194 | ret = 0; | |
195 | else | |
196 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
197 | ||
198 | spool->rsv_hpages += delta; | |
199 | if (spool->rsv_hpages > spool->min_hpages) | |
200 | spool->rsv_hpages = spool->min_hpages; | |
201 | } | |
202 | ||
203 | /* | |
204 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
205 | * quota reference, free it now. | |
206 | */ | |
90481622 | 207 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
208 | |
209 | return ret; | |
90481622 DG |
210 | } |
211 | ||
212 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
213 | { | |
214 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
215 | } | |
216 | ||
217 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
218 | { | |
496ad9aa | 219 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
220 | } |
221 | ||
96822904 AW |
222 | /* |
223 | * Region tracking -- allows tracking of reservations and instantiated pages | |
224 | * across the pages in a mapping. | |
84afd99b | 225 | * |
1dd308a7 MK |
226 | * The region data structures are embedded into a resv_map and protected |
227 | * by a resv_map's lock. The set of regions within the resv_map represent | |
228 | * reservations for huge pages, or huge pages that have already been | |
229 | * instantiated within the map. The from and to elements are huge page | |
230 | * indicies into the associated mapping. from indicates the starting index | |
231 | * of the region. to represents the first index past the end of the region. | |
232 | * | |
233 | * For example, a file region structure with from == 0 and to == 4 represents | |
234 | * four huge pages in a mapping. It is important to note that the to element | |
235 | * represents the first element past the end of the region. This is used in | |
236 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
237 | * | |
238 | * Interval notation of the form [from, to) will be used to indicate that | |
239 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
240 | */ |
241 | struct file_region { | |
242 | struct list_head link; | |
243 | long from; | |
244 | long to; | |
245 | }; | |
246 | ||
1dd308a7 MK |
247 | /* |
248 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
249 | * map. In the normal case, existing regions will be expanded |
250 | * to accommodate the specified range. Sufficient regions should | |
251 | * exist for expansion due to the previous call to region_chg | |
252 | * with the same range. However, it is possible that region_del | |
253 | * could have been called after region_chg and modifed the map | |
254 | * in such a way that no region exists to be expanded. In this | |
255 | * case, pull a region descriptor from the cache associated with | |
256 | * the map and use that for the new range. | |
cf3ad20b MK |
257 | * |
258 | * Return the number of new huge pages added to the map. This | |
259 | * number is greater than or equal to zero. | |
1dd308a7 | 260 | */ |
1406ec9b | 261 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 262 | { |
1406ec9b | 263 | struct list_head *head = &resv->regions; |
96822904 | 264 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 265 | long add = 0; |
96822904 | 266 | |
7b24d861 | 267 | spin_lock(&resv->lock); |
96822904 AW |
268 | /* Locate the region we are either in or before. */ |
269 | list_for_each_entry(rg, head, link) | |
270 | if (f <= rg->to) | |
271 | break; | |
272 | ||
5e911373 MK |
273 | /* |
274 | * If no region exists which can be expanded to include the | |
275 | * specified range, the list must have been modified by an | |
276 | * interleving call to region_del(). Pull a region descriptor | |
277 | * from the cache and use it for this range. | |
278 | */ | |
279 | if (&rg->link == head || t < rg->from) { | |
280 | VM_BUG_ON(resv->region_cache_count <= 0); | |
281 | ||
282 | resv->region_cache_count--; | |
283 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
284 | link); | |
285 | list_del(&nrg->link); | |
286 | ||
287 | nrg->from = f; | |
288 | nrg->to = t; | |
289 | list_add(&nrg->link, rg->link.prev); | |
290 | ||
291 | add += t - f; | |
292 | goto out_locked; | |
293 | } | |
294 | ||
96822904 AW |
295 | /* Round our left edge to the current segment if it encloses us. */ |
296 | if (f > rg->from) | |
297 | f = rg->from; | |
298 | ||
299 | /* Check for and consume any regions we now overlap with. */ | |
300 | nrg = rg; | |
301 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
302 | if (&rg->link == head) | |
303 | break; | |
304 | if (rg->from > t) | |
305 | break; | |
306 | ||
307 | /* If this area reaches higher then extend our area to | |
308 | * include it completely. If this is not the first area | |
309 | * which we intend to reuse, free it. */ | |
310 | if (rg->to > t) | |
311 | t = rg->to; | |
312 | if (rg != nrg) { | |
cf3ad20b MK |
313 | /* Decrement return value by the deleted range. |
314 | * Another range will span this area so that by | |
315 | * end of routine add will be >= zero | |
316 | */ | |
317 | add -= (rg->to - rg->from); | |
96822904 AW |
318 | list_del(&rg->link); |
319 | kfree(rg); | |
320 | } | |
321 | } | |
cf3ad20b MK |
322 | |
323 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 324 | nrg->from = f; |
cf3ad20b | 325 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 326 | nrg->to = t; |
cf3ad20b | 327 | |
5e911373 MK |
328 | out_locked: |
329 | resv->adds_in_progress--; | |
7b24d861 | 330 | spin_unlock(&resv->lock); |
cf3ad20b MK |
331 | VM_BUG_ON(add < 0); |
332 | return add; | |
96822904 AW |
333 | } |
334 | ||
1dd308a7 MK |
335 | /* |
336 | * Examine the existing reserve map and determine how many | |
337 | * huge pages in the specified range [f, t) are NOT currently | |
338 | * represented. This routine is called before a subsequent | |
339 | * call to region_add that will actually modify the reserve | |
340 | * map to add the specified range [f, t). region_chg does | |
341 | * not change the number of huge pages represented by the | |
342 | * map. However, if the existing regions in the map can not | |
343 | * be expanded to represent the new range, a new file_region | |
344 | * structure is added to the map as a placeholder. This is | |
345 | * so that the subsequent region_add call will have all the | |
346 | * regions it needs and will not fail. | |
347 | * | |
5e911373 MK |
348 | * Upon entry, region_chg will also examine the cache of region descriptors |
349 | * associated with the map. If there are not enough descriptors cached, one | |
350 | * will be allocated for the in progress add operation. | |
351 | * | |
352 | * Returns the number of huge pages that need to be added to the existing | |
353 | * reservation map for the range [f, t). This number is greater or equal to | |
354 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
355 | * is needed and can not be allocated. | |
1dd308a7 | 356 | */ |
1406ec9b | 357 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 358 | { |
1406ec9b | 359 | struct list_head *head = &resv->regions; |
7b24d861 | 360 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
361 | long chg = 0; |
362 | ||
7b24d861 DB |
363 | retry: |
364 | spin_lock(&resv->lock); | |
5e911373 MK |
365 | retry_locked: |
366 | resv->adds_in_progress++; | |
367 | ||
368 | /* | |
369 | * Check for sufficient descriptors in the cache to accommodate | |
370 | * the number of in progress add operations. | |
371 | */ | |
372 | if (resv->adds_in_progress > resv->region_cache_count) { | |
373 | struct file_region *trg; | |
374 | ||
375 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
376 | /* Must drop lock to allocate a new descriptor. */ | |
377 | resv->adds_in_progress--; | |
378 | spin_unlock(&resv->lock); | |
379 | ||
380 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
dbe409e4 MK |
381 | if (!trg) { |
382 | kfree(nrg); | |
5e911373 | 383 | return -ENOMEM; |
dbe409e4 | 384 | } |
5e911373 MK |
385 | |
386 | spin_lock(&resv->lock); | |
387 | list_add(&trg->link, &resv->region_cache); | |
388 | resv->region_cache_count++; | |
389 | goto retry_locked; | |
390 | } | |
391 | ||
96822904 AW |
392 | /* Locate the region we are before or in. */ |
393 | list_for_each_entry(rg, head, link) | |
394 | if (f <= rg->to) | |
395 | break; | |
396 | ||
397 | /* If we are below the current region then a new region is required. | |
398 | * Subtle, allocate a new region at the position but make it zero | |
399 | * size such that we can guarantee to record the reservation. */ | |
400 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 401 | if (!nrg) { |
5e911373 | 402 | resv->adds_in_progress--; |
7b24d861 DB |
403 | spin_unlock(&resv->lock); |
404 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
405 | if (!nrg) | |
406 | return -ENOMEM; | |
407 | ||
408 | nrg->from = f; | |
409 | nrg->to = f; | |
410 | INIT_LIST_HEAD(&nrg->link); | |
411 | goto retry; | |
412 | } | |
96822904 | 413 | |
7b24d861 DB |
414 | list_add(&nrg->link, rg->link.prev); |
415 | chg = t - f; | |
416 | goto out_nrg; | |
96822904 AW |
417 | } |
418 | ||
419 | /* Round our left edge to the current segment if it encloses us. */ | |
420 | if (f > rg->from) | |
421 | f = rg->from; | |
422 | chg = t - f; | |
423 | ||
424 | /* Check for and consume any regions we now overlap with. */ | |
425 | list_for_each_entry(rg, rg->link.prev, link) { | |
426 | if (&rg->link == head) | |
427 | break; | |
428 | if (rg->from > t) | |
7b24d861 | 429 | goto out; |
96822904 | 430 | |
25985edc | 431 | /* We overlap with this area, if it extends further than |
96822904 AW |
432 | * us then we must extend ourselves. Account for its |
433 | * existing reservation. */ | |
434 | if (rg->to > t) { | |
435 | chg += rg->to - t; | |
436 | t = rg->to; | |
437 | } | |
438 | chg -= rg->to - rg->from; | |
439 | } | |
7b24d861 DB |
440 | |
441 | out: | |
442 | spin_unlock(&resv->lock); | |
443 | /* We already know we raced and no longer need the new region */ | |
444 | kfree(nrg); | |
445 | return chg; | |
446 | out_nrg: | |
447 | spin_unlock(&resv->lock); | |
96822904 AW |
448 | return chg; |
449 | } | |
450 | ||
5e911373 MK |
451 | /* |
452 | * Abort the in progress add operation. The adds_in_progress field | |
453 | * of the resv_map keeps track of the operations in progress between | |
454 | * calls to region_chg and region_add. Operations are sometimes | |
455 | * aborted after the call to region_chg. In such cases, region_abort | |
456 | * is called to decrement the adds_in_progress counter. | |
457 | * | |
458 | * NOTE: The range arguments [f, t) are not needed or used in this | |
459 | * routine. They are kept to make reading the calling code easier as | |
460 | * arguments will match the associated region_chg call. | |
461 | */ | |
462 | static void region_abort(struct resv_map *resv, long f, long t) | |
463 | { | |
464 | spin_lock(&resv->lock); | |
465 | VM_BUG_ON(!resv->region_cache_count); | |
466 | resv->adds_in_progress--; | |
467 | spin_unlock(&resv->lock); | |
468 | } | |
469 | ||
1dd308a7 | 470 | /* |
feba16e2 MK |
471 | * Delete the specified range [f, t) from the reserve map. If the |
472 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
473 | * should be deleted. Locate the regions which intersect [f, t) | |
474 | * and either trim, delete or split the existing regions. | |
475 | * | |
476 | * Returns the number of huge pages deleted from the reserve map. | |
477 | * In the normal case, the return value is zero or more. In the | |
478 | * case where a region must be split, a new region descriptor must | |
479 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
480 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
481 | * a region and possibly return -ENOMEM. Callers specifying | |
482 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 483 | */ |
feba16e2 | 484 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 485 | { |
1406ec9b | 486 | struct list_head *head = &resv->regions; |
96822904 | 487 | struct file_region *rg, *trg; |
feba16e2 MK |
488 | struct file_region *nrg = NULL; |
489 | long del = 0; | |
96822904 | 490 | |
feba16e2 | 491 | retry: |
7b24d861 | 492 | spin_lock(&resv->lock); |
feba16e2 | 493 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
494 | /* |
495 | * Skip regions before the range to be deleted. file_region | |
496 | * ranges are normally of the form [from, to). However, there | |
497 | * may be a "placeholder" entry in the map which is of the form | |
498 | * (from, to) with from == to. Check for placeholder entries | |
499 | * at the beginning of the range to be deleted. | |
500 | */ | |
501 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 502 | continue; |
dbe409e4 | 503 | |
feba16e2 | 504 | if (rg->from >= t) |
96822904 | 505 | break; |
96822904 | 506 | |
feba16e2 MK |
507 | if (f > rg->from && t < rg->to) { /* Must split region */ |
508 | /* | |
509 | * Check for an entry in the cache before dropping | |
510 | * lock and attempting allocation. | |
511 | */ | |
512 | if (!nrg && | |
513 | resv->region_cache_count > resv->adds_in_progress) { | |
514 | nrg = list_first_entry(&resv->region_cache, | |
515 | struct file_region, | |
516 | link); | |
517 | list_del(&nrg->link); | |
518 | resv->region_cache_count--; | |
519 | } | |
96822904 | 520 | |
feba16e2 MK |
521 | if (!nrg) { |
522 | spin_unlock(&resv->lock); | |
523 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
524 | if (!nrg) | |
525 | return -ENOMEM; | |
526 | goto retry; | |
527 | } | |
528 | ||
529 | del += t - f; | |
530 | ||
531 | /* New entry for end of split region */ | |
532 | nrg->from = t; | |
533 | nrg->to = rg->to; | |
534 | INIT_LIST_HEAD(&nrg->link); | |
535 | ||
536 | /* Original entry is trimmed */ | |
537 | rg->to = f; | |
538 | ||
539 | list_add(&nrg->link, &rg->link); | |
540 | nrg = NULL; | |
96822904 | 541 | break; |
feba16e2 MK |
542 | } |
543 | ||
544 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
545 | del += rg->to - rg->from; | |
546 | list_del(&rg->link); | |
547 | kfree(rg); | |
548 | continue; | |
549 | } | |
550 | ||
551 | if (f <= rg->from) { /* Trim beginning of region */ | |
552 | del += t - rg->from; | |
553 | rg->from = t; | |
554 | } else { /* Trim end of region */ | |
555 | del += rg->to - f; | |
556 | rg->to = f; | |
557 | } | |
96822904 | 558 | } |
7b24d861 | 559 | |
7b24d861 | 560 | spin_unlock(&resv->lock); |
feba16e2 MK |
561 | kfree(nrg); |
562 | return del; | |
96822904 AW |
563 | } |
564 | ||
b5cec28d MK |
565 | /* |
566 | * A rare out of memory error was encountered which prevented removal of | |
567 | * the reserve map region for a page. The huge page itself was free'ed | |
568 | * and removed from the page cache. This routine will adjust the subpool | |
569 | * usage count, and the global reserve count if needed. By incrementing | |
570 | * these counts, the reserve map entry which could not be deleted will | |
571 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
572 | * counts. | |
573 | */ | |
72e2936c | 574 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
575 | { |
576 | struct hugepage_subpool *spool = subpool_inode(inode); | |
577 | long rsv_adjust; | |
578 | ||
579 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 580 | if (rsv_adjust) { |
b5cec28d MK |
581 | struct hstate *h = hstate_inode(inode); |
582 | ||
583 | hugetlb_acct_memory(h, 1); | |
584 | } | |
585 | } | |
586 | ||
1dd308a7 MK |
587 | /* |
588 | * Count and return the number of huge pages in the reserve map | |
589 | * that intersect with the range [f, t). | |
590 | */ | |
1406ec9b | 591 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 592 | { |
1406ec9b | 593 | struct list_head *head = &resv->regions; |
84afd99b AW |
594 | struct file_region *rg; |
595 | long chg = 0; | |
596 | ||
7b24d861 | 597 | spin_lock(&resv->lock); |
84afd99b AW |
598 | /* Locate each segment we overlap with, and count that overlap. */ |
599 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
600 | long seg_from; |
601 | long seg_to; | |
84afd99b AW |
602 | |
603 | if (rg->to <= f) | |
604 | continue; | |
605 | if (rg->from >= t) | |
606 | break; | |
607 | ||
608 | seg_from = max(rg->from, f); | |
609 | seg_to = min(rg->to, t); | |
610 | ||
611 | chg += seg_to - seg_from; | |
612 | } | |
7b24d861 | 613 | spin_unlock(&resv->lock); |
84afd99b AW |
614 | |
615 | return chg; | |
616 | } | |
617 | ||
e7c4b0bf AW |
618 | /* |
619 | * Convert the address within this vma to the page offset within | |
620 | * the mapping, in pagecache page units; huge pages here. | |
621 | */ | |
a5516438 AK |
622 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
623 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 624 | { |
a5516438 AK |
625 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
626 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
627 | } |
628 | ||
0fe6e20b NH |
629 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
630 | unsigned long address) | |
631 | { | |
632 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
633 | } | |
dee41079 | 634 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 635 | |
08fba699 MG |
636 | /* |
637 | * Return the size of the pages allocated when backing a VMA. In the majority | |
638 | * cases this will be same size as used by the page table entries. | |
639 | */ | |
640 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
641 | { | |
05ea8860 DW |
642 | if (vma->vm_ops && vma->vm_ops->pagesize) |
643 | return vma->vm_ops->pagesize(vma); | |
644 | return PAGE_SIZE; | |
08fba699 | 645 | } |
f340ca0f | 646 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 647 | |
3340289d MG |
648 | /* |
649 | * Return the page size being used by the MMU to back a VMA. In the majority | |
650 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
651 | * architectures where it differs, an architecture-specific 'strong' |
652 | * version of this symbol is required. | |
3340289d | 653 | */ |
09135cc5 | 654 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
655 | { |
656 | return vma_kernel_pagesize(vma); | |
657 | } | |
3340289d | 658 | |
84afd99b AW |
659 | /* |
660 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
661 | * bits of the reservation map pointer, which are always clear due to | |
662 | * alignment. | |
663 | */ | |
664 | #define HPAGE_RESV_OWNER (1UL << 0) | |
665 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 666 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 667 | |
a1e78772 MG |
668 | /* |
669 | * These helpers are used to track how many pages are reserved for | |
670 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
671 | * is guaranteed to have their future faults succeed. | |
672 | * | |
673 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
674 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
675 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
676 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
677 | * |
678 | * The private mapping reservation is represented in a subtly different | |
679 | * manner to a shared mapping. A shared mapping has a region map associated | |
680 | * with the underlying file, this region map represents the backing file | |
681 | * pages which have ever had a reservation assigned which this persists even | |
682 | * after the page is instantiated. A private mapping has a region map | |
683 | * associated with the original mmap which is attached to all VMAs which | |
684 | * reference it, this region map represents those offsets which have consumed | |
685 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 686 | */ |
e7c4b0bf AW |
687 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
688 | { | |
689 | return (unsigned long)vma->vm_private_data; | |
690 | } | |
691 | ||
692 | static void set_vma_private_data(struct vm_area_struct *vma, | |
693 | unsigned long value) | |
694 | { | |
695 | vma->vm_private_data = (void *)value; | |
696 | } | |
697 | ||
9119a41e | 698 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
699 | { |
700 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
701 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
702 | ||
703 | if (!resv_map || !rg) { | |
704 | kfree(resv_map); | |
705 | kfree(rg); | |
84afd99b | 706 | return NULL; |
5e911373 | 707 | } |
84afd99b AW |
708 | |
709 | kref_init(&resv_map->refs); | |
7b24d861 | 710 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
711 | INIT_LIST_HEAD(&resv_map->regions); |
712 | ||
5e911373 MK |
713 | resv_map->adds_in_progress = 0; |
714 | ||
715 | INIT_LIST_HEAD(&resv_map->region_cache); | |
716 | list_add(&rg->link, &resv_map->region_cache); | |
717 | resv_map->region_cache_count = 1; | |
718 | ||
84afd99b AW |
719 | return resv_map; |
720 | } | |
721 | ||
9119a41e | 722 | void resv_map_release(struct kref *ref) |
84afd99b AW |
723 | { |
724 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
725 | struct list_head *head = &resv_map->region_cache; |
726 | struct file_region *rg, *trg; | |
84afd99b AW |
727 | |
728 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 729 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
730 | |
731 | /* ... and any entries left in the cache */ | |
732 | list_for_each_entry_safe(rg, trg, head, link) { | |
733 | list_del(&rg->link); | |
734 | kfree(rg); | |
735 | } | |
736 | ||
737 | VM_BUG_ON(resv_map->adds_in_progress); | |
738 | ||
84afd99b AW |
739 | kfree(resv_map); |
740 | } | |
741 | ||
4e35f483 JK |
742 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
743 | { | |
f27a5136 MK |
744 | /* |
745 | * At inode evict time, i_mapping may not point to the original | |
746 | * address space within the inode. This original address space | |
747 | * contains the pointer to the resv_map. So, always use the | |
748 | * address space embedded within the inode. | |
749 | * The VERY common case is inode->mapping == &inode->i_data but, | |
750 | * this may not be true for device special inodes. | |
751 | */ | |
752 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
753 | } |
754 | ||
84afd99b | 755 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 756 | { |
81d1b09c | 757 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
758 | if (vma->vm_flags & VM_MAYSHARE) { |
759 | struct address_space *mapping = vma->vm_file->f_mapping; | |
760 | struct inode *inode = mapping->host; | |
761 | ||
762 | return inode_resv_map(inode); | |
763 | ||
764 | } else { | |
84afd99b AW |
765 | return (struct resv_map *)(get_vma_private_data(vma) & |
766 | ~HPAGE_RESV_MASK); | |
4e35f483 | 767 | } |
a1e78772 MG |
768 | } |
769 | ||
84afd99b | 770 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 771 | { |
81d1b09c SL |
772 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
773 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 774 | |
84afd99b AW |
775 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
776 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
777 | } |
778 | ||
779 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
780 | { | |
81d1b09c SL |
781 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
782 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
783 | |
784 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
785 | } |
786 | ||
787 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
788 | { | |
81d1b09c | 789 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
790 | |
791 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
792 | } |
793 | ||
04f2cbe3 | 794 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
795 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
796 | { | |
81d1b09c | 797 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 798 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
799 | vma->vm_private_data = (void *)0; |
800 | } | |
801 | ||
802 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 803 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 804 | { |
af0ed73e JK |
805 | if (vma->vm_flags & VM_NORESERVE) { |
806 | /* | |
807 | * This address is already reserved by other process(chg == 0), | |
808 | * so, we should decrement reserved count. Without decrementing, | |
809 | * reserve count remains after releasing inode, because this | |
810 | * allocated page will go into page cache and is regarded as | |
811 | * coming from reserved pool in releasing step. Currently, we | |
812 | * don't have any other solution to deal with this situation | |
813 | * properly, so add work-around here. | |
814 | */ | |
815 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 816 | return true; |
af0ed73e | 817 | else |
559ec2f8 | 818 | return false; |
af0ed73e | 819 | } |
a63884e9 JK |
820 | |
821 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
822 | if (vma->vm_flags & VM_MAYSHARE) { |
823 | /* | |
824 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
825 | * be a region map for all pages. The only situation where | |
826 | * there is no region map is if a hole was punched via | |
827 | * fallocate. In this case, there really are no reverves to | |
828 | * use. This situation is indicated if chg != 0. | |
829 | */ | |
830 | if (chg) | |
831 | return false; | |
832 | else | |
833 | return true; | |
834 | } | |
a63884e9 JK |
835 | |
836 | /* | |
837 | * Only the process that called mmap() has reserves for | |
838 | * private mappings. | |
839 | */ | |
67961f9d MK |
840 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
841 | /* | |
842 | * Like the shared case above, a hole punch or truncate | |
843 | * could have been performed on the private mapping. | |
844 | * Examine the value of chg to determine if reserves | |
845 | * actually exist or were previously consumed. | |
846 | * Very Subtle - The value of chg comes from a previous | |
847 | * call to vma_needs_reserves(). The reserve map for | |
848 | * private mappings has different (opposite) semantics | |
849 | * than that of shared mappings. vma_needs_reserves() | |
850 | * has already taken this difference in semantics into | |
851 | * account. Therefore, the meaning of chg is the same | |
852 | * as in the shared case above. Code could easily be | |
853 | * combined, but keeping it separate draws attention to | |
854 | * subtle differences. | |
855 | */ | |
856 | if (chg) | |
857 | return false; | |
858 | else | |
859 | return true; | |
860 | } | |
a63884e9 | 861 | |
559ec2f8 | 862 | return false; |
a1e78772 MG |
863 | } |
864 | ||
a5516438 | 865 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
866 | { |
867 | int nid = page_to_nid(page); | |
0edaecfa | 868 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
869 | h->free_huge_pages++; |
870 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
871 | } |
872 | ||
94310cbc | 873 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
874 | { |
875 | struct page *page; | |
876 | ||
c8721bbb | 877 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 878 | if (!PageHWPoison(page)) |
c8721bbb NH |
879 | break; |
880 | /* | |
881 | * if 'non-isolated free hugepage' not found on the list, | |
882 | * the allocation fails. | |
883 | */ | |
884 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 885 | return NULL; |
0edaecfa | 886 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 887 | set_page_refcounted(page); |
bf50bab2 NH |
888 | h->free_huge_pages--; |
889 | h->free_huge_pages_node[nid]--; | |
890 | return page; | |
891 | } | |
892 | ||
3e59fcb0 MH |
893 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
894 | nodemask_t *nmask) | |
94310cbc | 895 | { |
3e59fcb0 MH |
896 | unsigned int cpuset_mems_cookie; |
897 | struct zonelist *zonelist; | |
898 | struct zone *zone; | |
899 | struct zoneref *z; | |
98fa15f3 | 900 | int node = NUMA_NO_NODE; |
94310cbc | 901 | |
3e59fcb0 MH |
902 | zonelist = node_zonelist(nid, gfp_mask); |
903 | ||
904 | retry_cpuset: | |
905 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
906 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
907 | struct page *page; | |
908 | ||
909 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
910 | continue; | |
911 | /* | |
912 | * no need to ask again on the same node. Pool is node rather than | |
913 | * zone aware | |
914 | */ | |
915 | if (zone_to_nid(zone) == node) | |
916 | continue; | |
917 | node = zone_to_nid(zone); | |
94310cbc | 918 | |
94310cbc AK |
919 | page = dequeue_huge_page_node_exact(h, node); |
920 | if (page) | |
921 | return page; | |
922 | } | |
3e59fcb0 MH |
923 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
924 | goto retry_cpuset; | |
925 | ||
94310cbc AK |
926 | return NULL; |
927 | } | |
928 | ||
86cdb465 NH |
929 | /* Movability of hugepages depends on migration support. */ |
930 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
931 | { | |
7ed2c31d | 932 | if (hugepage_movable_supported(h)) |
86cdb465 NH |
933 | return GFP_HIGHUSER_MOVABLE; |
934 | else | |
935 | return GFP_HIGHUSER; | |
936 | } | |
937 | ||
a5516438 AK |
938 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
939 | struct vm_area_struct *vma, | |
af0ed73e JK |
940 | unsigned long address, int avoid_reserve, |
941 | long chg) | |
1da177e4 | 942 | { |
3e59fcb0 | 943 | struct page *page; |
480eccf9 | 944 | struct mempolicy *mpol; |
04ec6264 | 945 | gfp_t gfp_mask; |
3e59fcb0 | 946 | nodemask_t *nodemask; |
04ec6264 | 947 | int nid; |
1da177e4 | 948 | |
a1e78772 MG |
949 | /* |
950 | * A child process with MAP_PRIVATE mappings created by their parent | |
951 | * have no page reserves. This check ensures that reservations are | |
952 | * not "stolen". The child may still get SIGKILLed | |
953 | */ | |
af0ed73e | 954 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 955 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 956 | goto err; |
a1e78772 | 957 | |
04f2cbe3 | 958 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 959 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 960 | goto err; |
04f2cbe3 | 961 | |
04ec6264 VB |
962 | gfp_mask = htlb_alloc_mask(h); |
963 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
964 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
965 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
966 | SetPagePrivate(page); | |
967 | h->resv_huge_pages--; | |
1da177e4 | 968 | } |
cc9a6c87 | 969 | |
52cd3b07 | 970 | mpol_cond_put(mpol); |
1da177e4 | 971 | return page; |
cc9a6c87 MG |
972 | |
973 | err: | |
cc9a6c87 | 974 | return NULL; |
1da177e4 LT |
975 | } |
976 | ||
1cac6f2c LC |
977 | /* |
978 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
979 | * We may have allocated or freed a huge page based on a different | |
980 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
981 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
982 | * node for alloc or free. | |
983 | */ | |
984 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
985 | { | |
0edaf86c | 986 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
987 | VM_BUG_ON(nid >= MAX_NUMNODES); |
988 | ||
989 | return nid; | |
990 | } | |
991 | ||
992 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
993 | { | |
994 | if (!node_isset(nid, *nodes_allowed)) | |
995 | nid = next_node_allowed(nid, nodes_allowed); | |
996 | return nid; | |
997 | } | |
998 | ||
999 | /* | |
1000 | * returns the previously saved node ["this node"] from which to | |
1001 | * allocate a persistent huge page for the pool and advance the | |
1002 | * next node from which to allocate, handling wrap at end of node | |
1003 | * mask. | |
1004 | */ | |
1005 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1006 | nodemask_t *nodes_allowed) | |
1007 | { | |
1008 | int nid; | |
1009 | ||
1010 | VM_BUG_ON(!nodes_allowed); | |
1011 | ||
1012 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1013 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1014 | ||
1015 | return nid; | |
1016 | } | |
1017 | ||
1018 | /* | |
1019 | * helper for free_pool_huge_page() - return the previously saved | |
1020 | * node ["this node"] from which to free a huge page. Advance the | |
1021 | * next node id whether or not we find a free huge page to free so | |
1022 | * that the next attempt to free addresses the next node. | |
1023 | */ | |
1024 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1025 | { | |
1026 | int nid; | |
1027 | ||
1028 | VM_BUG_ON(!nodes_allowed); | |
1029 | ||
1030 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1031 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1032 | ||
1033 | return nid; | |
1034 | } | |
1035 | ||
1036 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1037 | for (nr_nodes = nodes_weight(*mask); \ | |
1038 | nr_nodes > 0 && \ | |
1039 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1040 | nr_nodes--) | |
1041 | ||
1042 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1043 | for (nr_nodes = nodes_weight(*mask); \ | |
1044 | nr_nodes > 0 && \ | |
1045 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1046 | nr_nodes--) | |
1047 | ||
e1073d1e | 1048 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1049 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1050 | unsigned int order) |
944d9fec LC |
1051 | { |
1052 | int i; | |
1053 | int nr_pages = 1 << order; | |
1054 | struct page *p = page + 1; | |
1055 | ||
c8cc708a | 1056 | atomic_set(compound_mapcount_ptr(page), 0); |
944d9fec | 1057 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1058 | clear_compound_head(p); |
944d9fec | 1059 | set_page_refcounted(p); |
944d9fec LC |
1060 | } |
1061 | ||
1062 | set_compound_order(page, 0); | |
1063 | __ClearPageHead(page); | |
1064 | } | |
1065 | ||
d00181b9 | 1066 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1067 | { |
1068 | free_contig_range(page_to_pfn(page), 1 << order); | |
1069 | } | |
1070 | ||
4eb0716e | 1071 | #ifdef CONFIG_CONTIG_ALLOC |
944d9fec | 1072 | static int __alloc_gigantic_page(unsigned long start_pfn, |
79b63f12 | 1073 | unsigned long nr_pages, gfp_t gfp_mask) |
944d9fec LC |
1074 | { |
1075 | unsigned long end_pfn = start_pfn + nr_pages; | |
ca96b625 | 1076 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, |
79b63f12 | 1077 | gfp_mask); |
944d9fec LC |
1078 | } |
1079 | ||
f44b2dda JK |
1080 | static bool pfn_range_valid_gigantic(struct zone *z, |
1081 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1082 | { |
1083 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1084 | struct page *page; | |
1085 | ||
1086 | for (i = start_pfn; i < end_pfn; i++) { | |
1087 | if (!pfn_valid(i)) | |
1088 | return false; | |
1089 | ||
1090 | page = pfn_to_page(i); | |
1091 | ||
f44b2dda JK |
1092 | if (page_zone(page) != z) |
1093 | return false; | |
1094 | ||
944d9fec LC |
1095 | if (PageReserved(page)) |
1096 | return false; | |
1097 | ||
1098 | if (page_count(page) > 0) | |
1099 | return false; | |
1100 | ||
1101 | if (PageHuge(page)) | |
1102 | return false; | |
1103 | } | |
1104 | ||
1105 | return true; | |
1106 | } | |
1107 | ||
1108 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1109 | unsigned long start_pfn, unsigned long nr_pages) | |
1110 | { | |
1111 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1112 | return zone_spans_pfn(zone, last_pfn); | |
1113 | } | |
1114 | ||
d9cc948f MH |
1115 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1116 | int nid, nodemask_t *nodemask) | |
944d9fec | 1117 | { |
79b63f12 | 1118 | unsigned int order = huge_page_order(h); |
944d9fec LC |
1119 | unsigned long nr_pages = 1 << order; |
1120 | unsigned long ret, pfn, flags; | |
79b63f12 MH |
1121 | struct zonelist *zonelist; |
1122 | struct zone *zone; | |
1123 | struct zoneref *z; | |
944d9fec | 1124 | |
79b63f12 | 1125 | zonelist = node_zonelist(nid, gfp_mask); |
d9cc948f | 1126 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { |
79b63f12 | 1127 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec | 1128 | |
79b63f12 MH |
1129 | pfn = ALIGN(zone->zone_start_pfn, nr_pages); |
1130 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) { | |
1131 | if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { | |
944d9fec LC |
1132 | /* |
1133 | * We release the zone lock here because | |
1134 | * alloc_contig_range() will also lock the zone | |
1135 | * at some point. If there's an allocation | |
1136 | * spinning on this lock, it may win the race | |
1137 | * and cause alloc_contig_range() to fail... | |
1138 | */ | |
79b63f12 MH |
1139 | spin_unlock_irqrestore(&zone->lock, flags); |
1140 | ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); | |
944d9fec LC |
1141 | if (!ret) |
1142 | return pfn_to_page(pfn); | |
79b63f12 | 1143 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec LC |
1144 | } |
1145 | pfn += nr_pages; | |
1146 | } | |
1147 | ||
79b63f12 | 1148 | spin_unlock_irqrestore(&zone->lock, flags); |
944d9fec LC |
1149 | } |
1150 | ||
1151 | return NULL; | |
1152 | } | |
1153 | ||
1154 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1155 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1156 | #else /* !CONFIG_CONTIG_ALLOC */ |
1157 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1158 | int nid, nodemask_t *nodemask) | |
1159 | { | |
1160 | return NULL; | |
1161 | } | |
1162 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1163 | |
e1073d1e | 1164 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1165 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1166 | int nid, nodemask_t *nodemask) |
1167 | { | |
1168 | return NULL; | |
1169 | } | |
d00181b9 | 1170 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1171 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1172 | unsigned int order) { } |
944d9fec LC |
1173 | #endif |
1174 | ||
a5516438 | 1175 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1176 | { |
1177 | int i; | |
a5516438 | 1178 | |
4eb0716e | 1179 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1180 | return; |
18229df5 | 1181 | |
a5516438 AK |
1182 | h->nr_huge_pages--; |
1183 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1184 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1185 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1186 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1187 | 1 << PG_active | 1 << PG_private | |
1188 | 1 << PG_writeback); | |
6af2acb6 | 1189 | } |
309381fe | 1190 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1191 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1192 | set_page_refcounted(page); |
944d9fec LC |
1193 | if (hstate_is_gigantic(h)) { |
1194 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1195 | free_gigantic_page(page, huge_page_order(h)); | |
1196 | } else { | |
944d9fec LC |
1197 | __free_pages(page, huge_page_order(h)); |
1198 | } | |
6af2acb6 AL |
1199 | } |
1200 | ||
e5ff2159 AK |
1201 | struct hstate *size_to_hstate(unsigned long size) |
1202 | { | |
1203 | struct hstate *h; | |
1204 | ||
1205 | for_each_hstate(h) { | |
1206 | if (huge_page_size(h) == size) | |
1207 | return h; | |
1208 | } | |
1209 | return NULL; | |
1210 | } | |
1211 | ||
bcc54222 NH |
1212 | /* |
1213 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1214 | * to hstate->hugepage_activelist.) | |
1215 | * | |
1216 | * This function can be called for tail pages, but never returns true for them. | |
1217 | */ | |
1218 | bool page_huge_active(struct page *page) | |
1219 | { | |
1220 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1221 | return PageHead(page) && PagePrivate(&page[1]); | |
1222 | } | |
1223 | ||
1224 | /* never called for tail page */ | |
1225 | static void set_page_huge_active(struct page *page) | |
1226 | { | |
1227 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1228 | SetPagePrivate(&page[1]); | |
1229 | } | |
1230 | ||
1231 | static void clear_page_huge_active(struct page *page) | |
1232 | { | |
1233 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1234 | ClearPagePrivate(&page[1]); | |
1235 | } | |
1236 | ||
ab5ac90a MH |
1237 | /* |
1238 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1239 | * code | |
1240 | */ | |
1241 | static inline bool PageHugeTemporary(struct page *page) | |
1242 | { | |
1243 | if (!PageHuge(page)) | |
1244 | return false; | |
1245 | ||
1246 | return (unsigned long)page[2].mapping == -1U; | |
1247 | } | |
1248 | ||
1249 | static inline void SetPageHugeTemporary(struct page *page) | |
1250 | { | |
1251 | page[2].mapping = (void *)-1U; | |
1252 | } | |
1253 | ||
1254 | static inline void ClearPageHugeTemporary(struct page *page) | |
1255 | { | |
1256 | page[2].mapping = NULL; | |
1257 | } | |
1258 | ||
8f1d26d0 | 1259 | void free_huge_page(struct page *page) |
27a85ef1 | 1260 | { |
a5516438 AK |
1261 | /* |
1262 | * Can't pass hstate in here because it is called from the | |
1263 | * compound page destructor. | |
1264 | */ | |
e5ff2159 | 1265 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1266 | int nid = page_to_nid(page); |
90481622 DG |
1267 | struct hugepage_subpool *spool = |
1268 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1269 | bool restore_reserve; |
27a85ef1 | 1270 | |
b4330afb MK |
1271 | VM_BUG_ON_PAGE(page_count(page), page); |
1272 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc YW |
1273 | |
1274 | set_page_private(page, 0); | |
1275 | page->mapping = NULL; | |
07443a85 | 1276 | restore_reserve = PagePrivate(page); |
16c794b4 | 1277 | ClearPagePrivate(page); |
27a85ef1 | 1278 | |
1c5ecae3 | 1279 | /* |
0919e1b6 MK |
1280 | * If PagePrivate() was set on page, page allocation consumed a |
1281 | * reservation. If the page was associated with a subpool, there | |
1282 | * would have been a page reserved in the subpool before allocation | |
1283 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1284 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1285 | * remove the reserved page from the subpool. | |
1c5ecae3 | 1286 | */ |
0919e1b6 MK |
1287 | if (!restore_reserve) { |
1288 | /* | |
1289 | * A return code of zero implies that the subpool will be | |
1290 | * under its minimum size if the reservation is not restored | |
1291 | * after page is free. Therefore, force restore_reserve | |
1292 | * operation. | |
1293 | */ | |
1294 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1295 | restore_reserve = true; | |
1296 | } | |
1c5ecae3 | 1297 | |
27a85ef1 | 1298 | spin_lock(&hugetlb_lock); |
bcc54222 | 1299 | clear_page_huge_active(page); |
6d76dcf4 AK |
1300 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1301 | pages_per_huge_page(h), page); | |
07443a85 JK |
1302 | if (restore_reserve) |
1303 | h->resv_huge_pages++; | |
1304 | ||
ab5ac90a MH |
1305 | if (PageHugeTemporary(page)) { |
1306 | list_del(&page->lru); | |
1307 | ClearPageHugeTemporary(page); | |
1308 | update_and_free_page(h, page); | |
1309 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1310 | /* remove the page from active list */ |
1311 | list_del(&page->lru); | |
a5516438 AK |
1312 | update_and_free_page(h, page); |
1313 | h->surplus_huge_pages--; | |
1314 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1315 | } else { |
5d3a551c | 1316 | arch_clear_hugepage_flags(page); |
a5516438 | 1317 | enqueue_huge_page(h, page); |
7893d1d5 | 1318 | } |
27a85ef1 DG |
1319 | spin_unlock(&hugetlb_lock); |
1320 | } | |
1321 | ||
a5516438 | 1322 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1323 | { |
0edaecfa | 1324 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1325 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1326 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1327 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1328 | h->nr_huge_pages++; |
1329 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 | 1330 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1331 | } |
1332 | ||
d00181b9 | 1333 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1334 | { |
1335 | int i; | |
1336 | int nr_pages = 1 << order; | |
1337 | struct page *p = page + 1; | |
1338 | ||
1339 | /* we rely on prep_new_huge_page to set the destructor */ | |
1340 | set_compound_order(page, order); | |
ef5a22be | 1341 | __ClearPageReserved(page); |
de09d31d | 1342 | __SetPageHead(page); |
20a0307c | 1343 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1344 | /* |
1345 | * For gigantic hugepages allocated through bootmem at | |
1346 | * boot, it's safer to be consistent with the not-gigantic | |
1347 | * hugepages and clear the PG_reserved bit from all tail pages | |
1348 | * too. Otherwse drivers using get_user_pages() to access tail | |
1349 | * pages may get the reference counting wrong if they see | |
1350 | * PG_reserved set on a tail page (despite the head page not | |
1351 | * having PG_reserved set). Enforcing this consistency between | |
1352 | * head and tail pages allows drivers to optimize away a check | |
1353 | * on the head page when they need know if put_page() is needed | |
1354 | * after get_user_pages(). | |
1355 | */ | |
1356 | __ClearPageReserved(p); | |
58a84aa9 | 1357 | set_page_count(p, 0); |
1d798ca3 | 1358 | set_compound_head(p, page); |
20a0307c | 1359 | } |
b4330afb | 1360 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1361 | } |
1362 | ||
7795912c AM |
1363 | /* |
1364 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1365 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1366 | * details. | |
1367 | */ | |
20a0307c WF |
1368 | int PageHuge(struct page *page) |
1369 | { | |
20a0307c WF |
1370 | if (!PageCompound(page)) |
1371 | return 0; | |
1372 | ||
1373 | page = compound_head(page); | |
f1e61557 | 1374 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1375 | } |
43131e14 NH |
1376 | EXPORT_SYMBOL_GPL(PageHuge); |
1377 | ||
27c73ae7 AA |
1378 | /* |
1379 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1380 | * normal or transparent huge pages. | |
1381 | */ | |
1382 | int PageHeadHuge(struct page *page_head) | |
1383 | { | |
27c73ae7 AA |
1384 | if (!PageHead(page_head)) |
1385 | return 0; | |
1386 | ||
758f66a2 | 1387 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1388 | } |
27c73ae7 | 1389 | |
13d60f4b ZY |
1390 | pgoff_t __basepage_index(struct page *page) |
1391 | { | |
1392 | struct page *page_head = compound_head(page); | |
1393 | pgoff_t index = page_index(page_head); | |
1394 | unsigned long compound_idx; | |
1395 | ||
1396 | if (!PageHuge(page_head)) | |
1397 | return page_index(page); | |
1398 | ||
1399 | if (compound_order(page_head) >= MAX_ORDER) | |
1400 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1401 | else | |
1402 | compound_idx = page - page_head; | |
1403 | ||
1404 | return (index << compound_order(page_head)) + compound_idx; | |
1405 | } | |
1406 | ||
0c397dae | 1407 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1408 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1409 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1410 | { |
af0fb9df | 1411 | int order = huge_page_order(h); |
1da177e4 | 1412 | struct page *page; |
f60858f9 | 1413 | bool alloc_try_hard = true; |
f96efd58 | 1414 | |
f60858f9 MK |
1415 | /* |
1416 | * By default we always try hard to allocate the page with | |
1417 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1418 | * a loop (to adjust global huge page counts) and previous allocation | |
1419 | * failed, do not continue to try hard on the same node. Use the | |
1420 | * node_alloc_noretry bitmap to manage this state information. | |
1421 | */ | |
1422 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1423 | alloc_try_hard = false; | |
1424 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1425 | if (alloc_try_hard) | |
1426 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1427 | if (nid == NUMA_NO_NODE) |
1428 | nid = numa_mem_id(); | |
1429 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1430 | if (page) | |
1431 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1432 | else | |
1433 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1434 | |
f60858f9 MK |
1435 | /* |
1436 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1437 | * indicates an overall state change. Clear bit so that we resume | |
1438 | * normal 'try hard' allocations. | |
1439 | */ | |
1440 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1441 | node_clear(nid, *node_alloc_noretry); | |
1442 | ||
1443 | /* | |
1444 | * If we tried hard to get a page but failed, set bit so that | |
1445 | * subsequent attempts will not try as hard until there is an | |
1446 | * overall state change. | |
1447 | */ | |
1448 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1449 | node_set(nid, *node_alloc_noretry); | |
1450 | ||
63b4613c NA |
1451 | return page; |
1452 | } | |
1453 | ||
0c397dae MH |
1454 | /* |
1455 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1456 | * should use this function to get new hugetlb pages | |
1457 | */ | |
1458 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1459 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1460 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1461 | { |
1462 | struct page *page; | |
1463 | ||
1464 | if (hstate_is_gigantic(h)) | |
1465 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1466 | else | |
1467 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1468 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1469 | if (!page) |
1470 | return NULL; | |
1471 | ||
1472 | if (hstate_is_gigantic(h)) | |
1473 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1474 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1475 | ||
1476 | return page; | |
1477 | } | |
1478 | ||
af0fb9df MH |
1479 | /* |
1480 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1481 | * manner. | |
1482 | */ | |
f60858f9 MK |
1483 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1484 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1485 | { |
1486 | struct page *page; | |
1487 | int nr_nodes, node; | |
af0fb9df | 1488 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1489 | |
1490 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1491 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1492 | node_alloc_noretry); | |
af0fb9df | 1493 | if (page) |
b2261026 | 1494 | break; |
b2261026 JK |
1495 | } |
1496 | ||
af0fb9df MH |
1497 | if (!page) |
1498 | return 0; | |
b2261026 | 1499 | |
af0fb9df MH |
1500 | put_page(page); /* free it into the hugepage allocator */ |
1501 | ||
1502 | return 1; | |
b2261026 JK |
1503 | } |
1504 | ||
e8c5c824 LS |
1505 | /* |
1506 | * Free huge page from pool from next node to free. | |
1507 | * Attempt to keep persistent huge pages more or less | |
1508 | * balanced over allowed nodes. | |
1509 | * Called with hugetlb_lock locked. | |
1510 | */ | |
6ae11b27 LS |
1511 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1512 | bool acct_surplus) | |
e8c5c824 | 1513 | { |
b2261026 | 1514 | int nr_nodes, node; |
e8c5c824 LS |
1515 | int ret = 0; |
1516 | ||
b2261026 | 1517 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1518 | /* |
1519 | * If we're returning unused surplus pages, only examine | |
1520 | * nodes with surplus pages. | |
1521 | */ | |
b2261026 JK |
1522 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1523 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1524 | struct page *page = |
b2261026 | 1525 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1526 | struct page, lru); |
1527 | list_del(&page->lru); | |
1528 | h->free_huge_pages--; | |
b2261026 | 1529 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1530 | if (acct_surplus) { |
1531 | h->surplus_huge_pages--; | |
b2261026 | 1532 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1533 | } |
e8c5c824 LS |
1534 | update_and_free_page(h, page); |
1535 | ret = 1; | |
9a76db09 | 1536 | break; |
e8c5c824 | 1537 | } |
b2261026 | 1538 | } |
e8c5c824 LS |
1539 | |
1540 | return ret; | |
1541 | } | |
1542 | ||
c8721bbb NH |
1543 | /* |
1544 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1545 | * nothing for in-use hugepages and non-hugepages. |
1546 | * This function returns values like below: | |
1547 | * | |
1548 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1549 | * (allocated or reserved.) | |
1550 | * 0: successfully dissolved free hugepages or the page is not a | |
1551 | * hugepage (considered as already dissolved) | |
c8721bbb | 1552 | */ |
c3114a84 | 1553 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1554 | { |
6bc9b564 | 1555 | int rc = -EBUSY; |
082d5b6b | 1556 | |
faf53def NH |
1557 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1558 | if (!PageHuge(page)) | |
1559 | return 0; | |
1560 | ||
c8721bbb | 1561 | spin_lock(&hugetlb_lock); |
faf53def NH |
1562 | if (!PageHuge(page)) { |
1563 | rc = 0; | |
1564 | goto out; | |
1565 | } | |
1566 | ||
1567 | if (!page_count(page)) { | |
2247bb33 GS |
1568 | struct page *head = compound_head(page); |
1569 | struct hstate *h = page_hstate(head); | |
1570 | int nid = page_to_nid(head); | |
6bc9b564 | 1571 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1572 | goto out; |
c3114a84 AK |
1573 | /* |
1574 | * Move PageHWPoison flag from head page to the raw error page, | |
1575 | * which makes any subpages rather than the error page reusable. | |
1576 | */ | |
1577 | if (PageHWPoison(head) && page != head) { | |
1578 | SetPageHWPoison(page); | |
1579 | ClearPageHWPoison(head); | |
1580 | } | |
2247bb33 | 1581 | list_del(&head->lru); |
c8721bbb NH |
1582 | h->free_huge_pages--; |
1583 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1584 | h->max_huge_pages--; |
2247bb33 | 1585 | update_and_free_page(h, head); |
6bc9b564 | 1586 | rc = 0; |
c8721bbb | 1587 | } |
082d5b6b | 1588 | out: |
c8721bbb | 1589 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1590 | return rc; |
c8721bbb NH |
1591 | } |
1592 | ||
1593 | /* | |
1594 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1595 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1596 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1597 | * part of it lies within the given range. | |
082d5b6b GS |
1598 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1599 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1600 | */ |
082d5b6b | 1601 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1602 | { |
c8721bbb | 1603 | unsigned long pfn; |
eb03aa00 | 1604 | struct page *page; |
082d5b6b | 1605 | int rc = 0; |
c8721bbb | 1606 | |
d0177639 | 1607 | if (!hugepages_supported()) |
082d5b6b | 1608 | return rc; |
d0177639 | 1609 | |
eb03aa00 GS |
1610 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1611 | page = pfn_to_page(pfn); | |
faf53def NH |
1612 | rc = dissolve_free_huge_page(page); |
1613 | if (rc) | |
1614 | break; | |
eb03aa00 | 1615 | } |
082d5b6b GS |
1616 | |
1617 | return rc; | |
c8721bbb NH |
1618 | } |
1619 | ||
ab5ac90a MH |
1620 | /* |
1621 | * Allocates a fresh surplus page from the page allocator. | |
1622 | */ | |
0c397dae | 1623 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1624 | int nid, nodemask_t *nmask) |
7893d1d5 | 1625 | { |
9980d744 | 1626 | struct page *page = NULL; |
7893d1d5 | 1627 | |
bae7f4ae | 1628 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1629 | return NULL; |
1630 | ||
d1c3fb1f | 1631 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1632 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1633 | goto out_unlock; | |
d1c3fb1f NA |
1634 | spin_unlock(&hugetlb_lock); |
1635 | ||
f60858f9 | 1636 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1637 | if (!page) |
0c397dae | 1638 | return NULL; |
d1c3fb1f NA |
1639 | |
1640 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1641 | /* |
1642 | * We could have raced with the pool size change. | |
1643 | * Double check that and simply deallocate the new page | |
1644 | * if we would end up overcommiting the surpluses. Abuse | |
1645 | * temporary page to workaround the nasty free_huge_page | |
1646 | * codeflow | |
1647 | */ | |
1648 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1649 | SetPageHugeTemporary(page); | |
2bf753e6 | 1650 | spin_unlock(&hugetlb_lock); |
9980d744 | 1651 | put_page(page); |
2bf753e6 | 1652 | return NULL; |
9980d744 | 1653 | } else { |
9980d744 | 1654 | h->surplus_huge_pages++; |
4704dea3 | 1655 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1656 | } |
9980d744 MH |
1657 | |
1658 | out_unlock: | |
d1c3fb1f | 1659 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1660 | |
1661 | return page; | |
1662 | } | |
1663 | ||
9a4e9f3b AK |
1664 | struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
1665 | int nid, nodemask_t *nmask) | |
ab5ac90a MH |
1666 | { |
1667 | struct page *page; | |
1668 | ||
1669 | if (hstate_is_gigantic(h)) | |
1670 | return NULL; | |
1671 | ||
f60858f9 | 1672 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1673 | if (!page) |
1674 | return NULL; | |
1675 | ||
1676 | /* | |
1677 | * We do not account these pages as surplus because they are only | |
1678 | * temporary and will be released properly on the last reference | |
1679 | */ | |
ab5ac90a MH |
1680 | SetPageHugeTemporary(page); |
1681 | ||
1682 | return page; | |
1683 | } | |
1684 | ||
099730d6 DH |
1685 | /* |
1686 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1687 | */ | |
e0ec90ee | 1688 | static |
0c397dae | 1689 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1690 | struct vm_area_struct *vma, unsigned long addr) |
1691 | { | |
aaf14e40 MH |
1692 | struct page *page; |
1693 | struct mempolicy *mpol; | |
1694 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1695 | int nid; | |
1696 | nodemask_t *nodemask; | |
1697 | ||
1698 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1699 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1700 | mpol_cond_put(mpol); |
1701 | ||
1702 | return page; | |
099730d6 DH |
1703 | } |
1704 | ||
ab5ac90a | 1705 | /* page migration callback function */ |
bf50bab2 NH |
1706 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
1707 | { | |
aaf14e40 | 1708 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1709 | struct page *page = NULL; |
bf50bab2 | 1710 | |
aaf14e40 MH |
1711 | if (nid != NUMA_NO_NODE) |
1712 | gfp_mask |= __GFP_THISNODE; | |
1713 | ||
bf50bab2 | 1714 | spin_lock(&hugetlb_lock); |
4ef91848 | 1715 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1716 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1717 | spin_unlock(&hugetlb_lock); |
1718 | ||
94ae8ba7 | 1719 | if (!page) |
0c397dae | 1720 | page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1721 | |
1722 | return page; | |
1723 | } | |
1724 | ||
ab5ac90a | 1725 | /* page migration callback function */ |
3e59fcb0 MH |
1726 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
1727 | nodemask_t *nmask) | |
4db9b2ef | 1728 | { |
aaf14e40 | 1729 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
1730 | |
1731 | spin_lock(&hugetlb_lock); | |
1732 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1733 | struct page *page; |
1734 | ||
1735 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1736 | if (page) { | |
1737 | spin_unlock(&hugetlb_lock); | |
1738 | return page; | |
4db9b2ef MH |
1739 | } |
1740 | } | |
1741 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1742 | |
0c397dae | 1743 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1744 | } |
1745 | ||
ebd63723 | 1746 | /* mempolicy aware migration callback */ |
389c8178 MH |
1747 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1748 | unsigned long address) | |
ebd63723 MH |
1749 | { |
1750 | struct mempolicy *mpol; | |
1751 | nodemask_t *nodemask; | |
1752 | struct page *page; | |
ebd63723 MH |
1753 | gfp_t gfp_mask; |
1754 | int node; | |
1755 | ||
ebd63723 MH |
1756 | gfp_mask = htlb_alloc_mask(h); |
1757 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
1758 | page = alloc_huge_page_nodemask(h, node, nodemask); | |
1759 | mpol_cond_put(mpol); | |
1760 | ||
1761 | return page; | |
1762 | } | |
1763 | ||
e4e574b7 | 1764 | /* |
25985edc | 1765 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1766 | * of size 'delta'. |
1767 | */ | |
a5516438 | 1768 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1769 | { |
1770 | struct list_head surplus_list; | |
1771 | struct page *page, *tmp; | |
1772 | int ret, i; | |
1773 | int needed, allocated; | |
28073b02 | 1774 | bool alloc_ok = true; |
e4e574b7 | 1775 | |
a5516438 | 1776 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1777 | if (needed <= 0) { |
a5516438 | 1778 | h->resv_huge_pages += delta; |
e4e574b7 | 1779 | return 0; |
ac09b3a1 | 1780 | } |
e4e574b7 AL |
1781 | |
1782 | allocated = 0; | |
1783 | INIT_LIST_HEAD(&surplus_list); | |
1784 | ||
1785 | ret = -ENOMEM; | |
1786 | retry: | |
1787 | spin_unlock(&hugetlb_lock); | |
1788 | for (i = 0; i < needed; i++) { | |
0c397dae | 1789 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1790 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1791 | if (!page) { |
1792 | alloc_ok = false; | |
1793 | break; | |
1794 | } | |
e4e574b7 | 1795 | list_add(&page->lru, &surplus_list); |
69ed779a | 1796 | cond_resched(); |
e4e574b7 | 1797 | } |
28073b02 | 1798 | allocated += i; |
e4e574b7 AL |
1799 | |
1800 | /* | |
1801 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1802 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1803 | */ | |
1804 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1805 | needed = (h->resv_huge_pages + delta) - |
1806 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1807 | if (needed > 0) { |
1808 | if (alloc_ok) | |
1809 | goto retry; | |
1810 | /* | |
1811 | * We were not able to allocate enough pages to | |
1812 | * satisfy the entire reservation so we free what | |
1813 | * we've allocated so far. | |
1814 | */ | |
1815 | goto free; | |
1816 | } | |
e4e574b7 AL |
1817 | /* |
1818 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1819 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1820 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1821 | * allocator. Commit the entire reservation here to prevent another |
1822 | * process from stealing the pages as they are added to the pool but | |
1823 | * before they are reserved. | |
e4e574b7 AL |
1824 | */ |
1825 | needed += allocated; | |
a5516438 | 1826 | h->resv_huge_pages += delta; |
e4e574b7 | 1827 | ret = 0; |
a9869b83 | 1828 | |
19fc3f0a | 1829 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1830 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1831 | if ((--needed) < 0) |
1832 | break; | |
a9869b83 NH |
1833 | /* |
1834 | * This page is now managed by the hugetlb allocator and has | |
1835 | * no users -- drop the buddy allocator's reference. | |
1836 | */ | |
1837 | put_page_testzero(page); | |
309381fe | 1838 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1839 | enqueue_huge_page(h, page); |
19fc3f0a | 1840 | } |
28073b02 | 1841 | free: |
b0365c8d | 1842 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1843 | |
1844 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1845 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1846 | put_page(page); | |
a9869b83 | 1847 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1848 | |
1849 | return ret; | |
1850 | } | |
1851 | ||
1852 | /* | |
e5bbc8a6 MK |
1853 | * This routine has two main purposes: |
1854 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1855 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1856 | * to the associated reservation map. | |
1857 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1858 | * the reservation. As many as unused_resv_pages may be freed. | |
1859 | * | |
1860 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1861 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1862 | * we must make sure nobody else can claim pages we are in the process of | |
1863 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1864 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1865 | */ |
a5516438 AK |
1866 | static void return_unused_surplus_pages(struct hstate *h, |
1867 | unsigned long unused_resv_pages) | |
e4e574b7 | 1868 | { |
e4e574b7 AL |
1869 | unsigned long nr_pages; |
1870 | ||
aa888a74 | 1871 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1872 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1873 | goto out; |
aa888a74 | 1874 | |
e5bbc8a6 MK |
1875 | /* |
1876 | * Part (or even all) of the reservation could have been backed | |
1877 | * by pre-allocated pages. Only free surplus pages. | |
1878 | */ | |
a5516438 | 1879 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1880 | |
685f3457 LS |
1881 | /* |
1882 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1883 | * evenly across all nodes with memory. Iterate across these nodes |
1884 | * until we can no longer free unreserved surplus pages. This occurs | |
1885 | * when the nodes with surplus pages have no free pages. | |
1886 | * free_pool_huge_page() will balance the the freed pages across the | |
1887 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1888 | * |
1889 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1890 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1891 | * to cover subsequent pages we may free. | |
685f3457 LS |
1892 | */ |
1893 | while (nr_pages--) { | |
e5bbc8a6 MK |
1894 | h->resv_huge_pages--; |
1895 | unused_resv_pages--; | |
8cebfcd0 | 1896 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1897 | goto out; |
7848a4bf | 1898 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1899 | } |
e5bbc8a6 MK |
1900 | |
1901 | out: | |
1902 | /* Fully uncommit the reservation */ | |
1903 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1904 | } |
1905 | ||
5e911373 | 1906 | |
c37f9fb1 | 1907 | /* |
feba16e2 | 1908 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1909 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1910 | * |
1911 | * vma_needs_reservation is called to determine if the huge page at addr | |
1912 | * within the vma has an associated reservation. If a reservation is | |
1913 | * needed, the value 1 is returned. The caller is then responsible for | |
1914 | * managing the global reservation and subpool usage counts. After | |
1915 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1916 | * to add the page to the reservation map. If the page allocation fails, |
1917 | * the reservation must be ended instead of committed. vma_end_reservation | |
1918 | * is called in such cases. | |
cf3ad20b MK |
1919 | * |
1920 | * In the normal case, vma_commit_reservation returns the same value | |
1921 | * as the preceding vma_needs_reservation call. The only time this | |
1922 | * is not the case is if a reserve map was changed between calls. It | |
1923 | * is the responsibility of the caller to notice the difference and | |
1924 | * take appropriate action. | |
96b96a96 MK |
1925 | * |
1926 | * vma_add_reservation is used in error paths where a reservation must | |
1927 | * be restored when a newly allocated huge page must be freed. It is | |
1928 | * to be called after calling vma_needs_reservation to determine if a | |
1929 | * reservation exists. | |
c37f9fb1 | 1930 | */ |
5e911373 MK |
1931 | enum vma_resv_mode { |
1932 | VMA_NEEDS_RESV, | |
1933 | VMA_COMMIT_RESV, | |
feba16e2 | 1934 | VMA_END_RESV, |
96b96a96 | 1935 | VMA_ADD_RESV, |
5e911373 | 1936 | }; |
cf3ad20b MK |
1937 | static long __vma_reservation_common(struct hstate *h, |
1938 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1939 | enum vma_resv_mode mode) |
c37f9fb1 | 1940 | { |
4e35f483 JK |
1941 | struct resv_map *resv; |
1942 | pgoff_t idx; | |
cf3ad20b | 1943 | long ret; |
c37f9fb1 | 1944 | |
4e35f483 JK |
1945 | resv = vma_resv_map(vma); |
1946 | if (!resv) | |
84afd99b | 1947 | return 1; |
c37f9fb1 | 1948 | |
4e35f483 | 1949 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1950 | switch (mode) { |
1951 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1952 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1953 | break; |
1954 | case VMA_COMMIT_RESV: | |
1955 | ret = region_add(resv, idx, idx + 1); | |
1956 | break; | |
feba16e2 | 1957 | case VMA_END_RESV: |
5e911373 MK |
1958 | region_abort(resv, idx, idx + 1); |
1959 | ret = 0; | |
1960 | break; | |
96b96a96 MK |
1961 | case VMA_ADD_RESV: |
1962 | if (vma->vm_flags & VM_MAYSHARE) | |
1963 | ret = region_add(resv, idx, idx + 1); | |
1964 | else { | |
1965 | region_abort(resv, idx, idx + 1); | |
1966 | ret = region_del(resv, idx, idx + 1); | |
1967 | } | |
1968 | break; | |
5e911373 MK |
1969 | default: |
1970 | BUG(); | |
1971 | } | |
84afd99b | 1972 | |
4e35f483 | 1973 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1974 | return ret; |
67961f9d MK |
1975 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1976 | /* | |
1977 | * In most cases, reserves always exist for private mappings. | |
1978 | * However, a file associated with mapping could have been | |
1979 | * hole punched or truncated after reserves were consumed. | |
1980 | * As subsequent fault on such a range will not use reserves. | |
1981 | * Subtle - The reserve map for private mappings has the | |
1982 | * opposite meaning than that of shared mappings. If NO | |
1983 | * entry is in the reserve map, it means a reservation exists. | |
1984 | * If an entry exists in the reserve map, it means the | |
1985 | * reservation has already been consumed. As a result, the | |
1986 | * return value of this routine is the opposite of the | |
1987 | * value returned from reserve map manipulation routines above. | |
1988 | */ | |
1989 | if (ret) | |
1990 | return 0; | |
1991 | else | |
1992 | return 1; | |
1993 | } | |
4e35f483 | 1994 | else |
cf3ad20b | 1995 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1996 | } |
cf3ad20b MK |
1997 | |
1998 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1999 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2000 | { |
5e911373 | 2001 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2002 | } |
84afd99b | 2003 | |
cf3ad20b MK |
2004 | static long vma_commit_reservation(struct hstate *h, |
2005 | struct vm_area_struct *vma, unsigned long addr) | |
2006 | { | |
5e911373 MK |
2007 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2008 | } | |
2009 | ||
feba16e2 | 2010 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2011 | struct vm_area_struct *vma, unsigned long addr) |
2012 | { | |
feba16e2 | 2013 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2014 | } |
2015 | ||
96b96a96 MK |
2016 | static long vma_add_reservation(struct hstate *h, |
2017 | struct vm_area_struct *vma, unsigned long addr) | |
2018 | { | |
2019 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2020 | } | |
2021 | ||
2022 | /* | |
2023 | * This routine is called to restore a reservation on error paths. In the | |
2024 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2025 | * and is about to be freed. If a reservation for the page existed, | |
2026 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2027 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2028 | * the global reservation count will be incremented if PagePrivate is set. | |
2029 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2030 | * reserve map here to be consistent with global reserve count adjustments | |
2031 | * to be made by free_huge_page. | |
2032 | */ | |
2033 | static void restore_reserve_on_error(struct hstate *h, | |
2034 | struct vm_area_struct *vma, unsigned long address, | |
2035 | struct page *page) | |
2036 | { | |
2037 | if (unlikely(PagePrivate(page))) { | |
2038 | long rc = vma_needs_reservation(h, vma, address); | |
2039 | ||
2040 | if (unlikely(rc < 0)) { | |
2041 | /* | |
2042 | * Rare out of memory condition in reserve map | |
2043 | * manipulation. Clear PagePrivate so that | |
2044 | * global reserve count will not be incremented | |
2045 | * by free_huge_page. This will make it appear | |
2046 | * as though the reservation for this page was | |
2047 | * consumed. This may prevent the task from | |
2048 | * faulting in the page at a later time. This | |
2049 | * is better than inconsistent global huge page | |
2050 | * accounting of reserve counts. | |
2051 | */ | |
2052 | ClearPagePrivate(page); | |
2053 | } else if (rc) { | |
2054 | rc = vma_add_reservation(h, vma, address); | |
2055 | if (unlikely(rc < 0)) | |
2056 | /* | |
2057 | * See above comment about rare out of | |
2058 | * memory condition. | |
2059 | */ | |
2060 | ClearPagePrivate(page); | |
2061 | } else | |
2062 | vma_end_reservation(h, vma, address); | |
2063 | } | |
2064 | } | |
2065 | ||
70c3547e | 2066 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2067 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2068 | { |
90481622 | 2069 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2070 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2071 | struct page *page; |
d85f69b0 MK |
2072 | long map_chg, map_commit; |
2073 | long gbl_chg; | |
6d76dcf4 AK |
2074 | int ret, idx; |
2075 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 2076 | |
6d76dcf4 | 2077 | idx = hstate_index(h); |
a1e78772 | 2078 | /* |
d85f69b0 MK |
2079 | * Examine the region/reserve map to determine if the process |
2080 | * has a reservation for the page to be allocated. A return | |
2081 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2082 | */ |
d85f69b0 MK |
2083 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2084 | if (map_chg < 0) | |
76dcee75 | 2085 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2086 | |
2087 | /* | |
2088 | * Processes that did not create the mapping will have no | |
2089 | * reserves as indicated by the region/reserve map. Check | |
2090 | * that the allocation will not exceed the subpool limit. | |
2091 | * Allocations for MAP_NORESERVE mappings also need to be | |
2092 | * checked against any subpool limit. | |
2093 | */ | |
2094 | if (map_chg || avoid_reserve) { | |
2095 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2096 | if (gbl_chg < 0) { | |
feba16e2 | 2097 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2098 | return ERR_PTR(-ENOSPC); |
5e911373 | 2099 | } |
1da177e4 | 2100 | |
d85f69b0 MK |
2101 | /* |
2102 | * Even though there was no reservation in the region/reserve | |
2103 | * map, there could be reservations associated with the | |
2104 | * subpool that can be used. This would be indicated if the | |
2105 | * return value of hugepage_subpool_get_pages() is zero. | |
2106 | * However, if avoid_reserve is specified we still avoid even | |
2107 | * the subpool reservations. | |
2108 | */ | |
2109 | if (avoid_reserve) | |
2110 | gbl_chg = 1; | |
2111 | } | |
2112 | ||
6d76dcf4 | 2113 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2114 | if (ret) |
2115 | goto out_subpool_put; | |
2116 | ||
1da177e4 | 2117 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2118 | /* |
2119 | * glb_chg is passed to indicate whether or not a page must be taken | |
2120 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2121 | * a reservation exists for the allocation. | |
2122 | */ | |
2123 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2124 | if (!page) { |
94ae8ba7 | 2125 | spin_unlock(&hugetlb_lock); |
0c397dae | 2126 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2127 | if (!page) |
2128 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2129 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2130 | SetPagePrivate(page); | |
2131 | h->resv_huge_pages--; | |
2132 | } | |
79dbb236 AK |
2133 | spin_lock(&hugetlb_lock); |
2134 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2135 | /* Fall through */ |
68842c9b | 2136 | } |
81a6fcae JK |
2137 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2138 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2139 | |
90481622 | 2140 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2141 | |
d85f69b0 MK |
2142 | map_commit = vma_commit_reservation(h, vma, addr); |
2143 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2144 | /* |
2145 | * The page was added to the reservation map between | |
2146 | * vma_needs_reservation and vma_commit_reservation. | |
2147 | * This indicates a race with hugetlb_reserve_pages. | |
2148 | * Adjust for the subpool count incremented above AND | |
2149 | * in hugetlb_reserve_pages for the same page. Also, | |
2150 | * the reservation count added in hugetlb_reserve_pages | |
2151 | * no longer applies. | |
2152 | */ | |
2153 | long rsv_adjust; | |
2154 | ||
2155 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2156 | hugetlb_acct_memory(h, -rsv_adjust); | |
2157 | } | |
90d8b7e6 | 2158 | return page; |
8f34af6f JZ |
2159 | |
2160 | out_uncharge_cgroup: | |
2161 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2162 | out_subpool_put: | |
d85f69b0 | 2163 | if (map_chg || avoid_reserve) |
8f34af6f | 2164 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2165 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2166 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2167 | } |
2168 | ||
e24a1307 AK |
2169 | int alloc_bootmem_huge_page(struct hstate *h) |
2170 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2171 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2172 | { |
2173 | struct huge_bootmem_page *m; | |
b2261026 | 2174 | int nr_nodes, node; |
aa888a74 | 2175 | |
b2261026 | 2176 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2177 | void *addr; |
2178 | ||
eb31d559 | 2179 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2180 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2181 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2182 | if (addr) { |
2183 | /* | |
2184 | * Use the beginning of the huge page to store the | |
2185 | * huge_bootmem_page struct (until gather_bootmem | |
2186 | * puts them into the mem_map). | |
2187 | */ | |
2188 | m = addr; | |
91f47662 | 2189 | goto found; |
aa888a74 | 2190 | } |
aa888a74 AK |
2191 | } |
2192 | return 0; | |
2193 | ||
2194 | found: | |
df994ead | 2195 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2196 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2197 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2198 | list_add(&m->list, &huge_boot_pages); |
2199 | m->hstate = h; | |
2200 | return 1; | |
2201 | } | |
2202 | ||
d00181b9 KS |
2203 | static void __init prep_compound_huge_page(struct page *page, |
2204 | unsigned int order) | |
18229df5 AW |
2205 | { |
2206 | if (unlikely(order > (MAX_ORDER - 1))) | |
2207 | prep_compound_gigantic_page(page, order); | |
2208 | else | |
2209 | prep_compound_page(page, order); | |
2210 | } | |
2211 | ||
aa888a74 AK |
2212 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2213 | static void __init gather_bootmem_prealloc(void) | |
2214 | { | |
2215 | struct huge_bootmem_page *m; | |
2216 | ||
2217 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2218 | struct page *page = virt_to_page(m); |
aa888a74 | 2219 | struct hstate *h = m->hstate; |
ee8f248d | 2220 | |
aa888a74 | 2221 | WARN_ON(page_count(page) != 1); |
18229df5 | 2222 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2223 | WARN_ON(PageReserved(page)); |
aa888a74 | 2224 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2225 | put_page(page); /* free it into the hugepage allocator */ |
2226 | ||
b0320c7b RA |
2227 | /* |
2228 | * If we had gigantic hugepages allocated at boot time, we need | |
2229 | * to restore the 'stolen' pages to totalram_pages in order to | |
2230 | * fix confusing memory reports from free(1) and another | |
2231 | * side-effects, like CommitLimit going negative. | |
2232 | */ | |
bae7f4ae | 2233 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2234 | adjust_managed_page_count(page, 1 << h->order); |
520495fe | 2235 | cond_resched(); |
aa888a74 AK |
2236 | } |
2237 | } | |
2238 | ||
8faa8b07 | 2239 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2240 | { |
2241 | unsigned long i; | |
f60858f9 MK |
2242 | nodemask_t *node_alloc_noretry; |
2243 | ||
2244 | if (!hstate_is_gigantic(h)) { | |
2245 | /* | |
2246 | * Bit mask controlling how hard we retry per-node allocations. | |
2247 | * Ignore errors as lower level routines can deal with | |
2248 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2249 | * time, we are likely in bigger trouble. | |
2250 | */ | |
2251 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2252 | GFP_KERNEL); | |
2253 | } else { | |
2254 | /* allocations done at boot time */ | |
2255 | node_alloc_noretry = NULL; | |
2256 | } | |
2257 | ||
2258 | /* bit mask controlling how hard we retry per-node allocations */ | |
2259 | if (node_alloc_noretry) | |
2260 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2261 | |
e5ff2159 | 2262 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2263 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2264 | if (!alloc_bootmem_huge_page(h)) |
2265 | break; | |
0c397dae | 2266 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2267 | &node_states[N_MEMORY], |
2268 | node_alloc_noretry)) | |
1da177e4 | 2269 | break; |
69ed779a | 2270 | cond_resched(); |
1da177e4 | 2271 | } |
d715cf80 LH |
2272 | if (i < h->max_huge_pages) { |
2273 | char buf[32]; | |
2274 | ||
c6247f72 | 2275 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2276 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2277 | h->max_huge_pages, buf, i); | |
2278 | h->max_huge_pages = i; | |
2279 | } | |
f60858f9 MK |
2280 | |
2281 | kfree(node_alloc_noretry); | |
e5ff2159 AK |
2282 | } |
2283 | ||
2284 | static void __init hugetlb_init_hstates(void) | |
2285 | { | |
2286 | struct hstate *h; | |
2287 | ||
2288 | for_each_hstate(h) { | |
641844f5 NH |
2289 | if (minimum_order > huge_page_order(h)) |
2290 | minimum_order = huge_page_order(h); | |
2291 | ||
8faa8b07 | 2292 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2293 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2294 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2295 | } |
641844f5 | 2296 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2297 | } |
2298 | ||
2299 | static void __init report_hugepages(void) | |
2300 | { | |
2301 | struct hstate *h; | |
2302 | ||
2303 | for_each_hstate(h) { | |
4abd32db | 2304 | char buf[32]; |
c6247f72 MW |
2305 | |
2306 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2307 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2308 | buf, h->free_huge_pages); |
e5ff2159 AK |
2309 | } |
2310 | } | |
2311 | ||
1da177e4 | 2312 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2313 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2314 | nodemask_t *nodes_allowed) | |
1da177e4 | 2315 | { |
4415cc8d CL |
2316 | int i; |
2317 | ||
bae7f4ae | 2318 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2319 | return; |
2320 | ||
6ae11b27 | 2321 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2322 | struct page *page, *next; |
a5516438 AK |
2323 | struct list_head *freel = &h->hugepage_freelists[i]; |
2324 | list_for_each_entry_safe(page, next, freel, lru) { | |
2325 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2326 | return; |
1da177e4 LT |
2327 | if (PageHighMem(page)) |
2328 | continue; | |
2329 | list_del(&page->lru); | |
e5ff2159 | 2330 | update_and_free_page(h, page); |
a5516438 AK |
2331 | h->free_huge_pages--; |
2332 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2333 | } |
2334 | } | |
2335 | } | |
2336 | #else | |
6ae11b27 LS |
2337 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2338 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2339 | { |
2340 | } | |
2341 | #endif | |
2342 | ||
20a0307c WF |
2343 | /* |
2344 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2345 | * balanced by operating on them in a round-robin fashion. | |
2346 | * Returns 1 if an adjustment was made. | |
2347 | */ | |
6ae11b27 LS |
2348 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2349 | int delta) | |
20a0307c | 2350 | { |
b2261026 | 2351 | int nr_nodes, node; |
20a0307c WF |
2352 | |
2353 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2354 | |
b2261026 JK |
2355 | if (delta < 0) { |
2356 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2357 | if (h->surplus_huge_pages_node[node]) | |
2358 | goto found; | |
e8c5c824 | 2359 | } |
b2261026 JK |
2360 | } else { |
2361 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2362 | if (h->surplus_huge_pages_node[node] < | |
2363 | h->nr_huge_pages_node[node]) | |
2364 | goto found; | |
e8c5c824 | 2365 | } |
b2261026 JK |
2366 | } |
2367 | return 0; | |
20a0307c | 2368 | |
b2261026 JK |
2369 | found: |
2370 | h->surplus_huge_pages += delta; | |
2371 | h->surplus_huge_pages_node[node] += delta; | |
2372 | return 1; | |
20a0307c WF |
2373 | } |
2374 | ||
a5516438 | 2375 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2376 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2377 | nodemask_t *nodes_allowed) |
1da177e4 | 2378 | { |
7893d1d5 | 2379 | unsigned long min_count, ret; |
f60858f9 MK |
2380 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2381 | ||
2382 | /* | |
2383 | * Bit mask controlling how hard we retry per-node allocations. | |
2384 | * If we can not allocate the bit mask, do not attempt to allocate | |
2385 | * the requested huge pages. | |
2386 | */ | |
2387 | if (node_alloc_noretry) | |
2388 | nodes_clear(*node_alloc_noretry); | |
2389 | else | |
2390 | return -ENOMEM; | |
1da177e4 | 2391 | |
4eb0716e AG |
2392 | spin_lock(&hugetlb_lock); |
2393 | ||
fd875dca MK |
2394 | /* |
2395 | * Check for a node specific request. | |
2396 | * Changing node specific huge page count may require a corresponding | |
2397 | * change to the global count. In any case, the passed node mask | |
2398 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2399 | */ | |
2400 | if (nid != NUMA_NO_NODE) { | |
2401 | unsigned long old_count = count; | |
2402 | ||
2403 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2404 | /* | |
2405 | * User may have specified a large count value which caused the | |
2406 | * above calculation to overflow. In this case, they wanted | |
2407 | * to allocate as many huge pages as possible. Set count to | |
2408 | * largest possible value to align with their intention. | |
2409 | */ | |
2410 | if (count < old_count) | |
2411 | count = ULONG_MAX; | |
2412 | } | |
2413 | ||
4eb0716e AG |
2414 | /* |
2415 | * Gigantic pages runtime allocation depend on the capability for large | |
2416 | * page range allocation. | |
2417 | * If the system does not provide this feature, return an error when | |
2418 | * the user tries to allocate gigantic pages but let the user free the | |
2419 | * boottime allocated gigantic pages. | |
2420 | */ | |
2421 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2422 | if (count > persistent_huge_pages(h)) { | |
2423 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2424 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2425 | return -EINVAL; |
2426 | } | |
2427 | /* Fall through to decrease pool */ | |
2428 | } | |
aa888a74 | 2429 | |
7893d1d5 AL |
2430 | /* |
2431 | * Increase the pool size | |
2432 | * First take pages out of surplus state. Then make up the | |
2433 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2434 | * |
0c397dae | 2435 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2436 | * to convert a surplus huge page to a normal huge page. That is |
2437 | * not critical, though, it just means the overall size of the | |
2438 | * pool might be one hugepage larger than it needs to be, but | |
2439 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2440 | */ |
a5516438 | 2441 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2442 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2443 | break; |
2444 | } | |
2445 | ||
a5516438 | 2446 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2447 | /* |
2448 | * If this allocation races such that we no longer need the | |
2449 | * page, free_huge_page will handle it by freeing the page | |
2450 | * and reducing the surplus. | |
2451 | */ | |
2452 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2453 | |
2454 | /* yield cpu to avoid soft lockup */ | |
2455 | cond_resched(); | |
2456 | ||
f60858f9 MK |
2457 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2458 | node_alloc_noretry); | |
7893d1d5 AL |
2459 | spin_lock(&hugetlb_lock); |
2460 | if (!ret) | |
2461 | goto out; | |
2462 | ||
536240f2 MG |
2463 | /* Bail for signals. Probably ctrl-c from user */ |
2464 | if (signal_pending(current)) | |
2465 | goto out; | |
7893d1d5 | 2466 | } |
7893d1d5 AL |
2467 | |
2468 | /* | |
2469 | * Decrease the pool size | |
2470 | * First return free pages to the buddy allocator (being careful | |
2471 | * to keep enough around to satisfy reservations). Then place | |
2472 | * pages into surplus state as needed so the pool will shrink | |
2473 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2474 | * |
2475 | * By placing pages into the surplus state independent of the | |
2476 | * overcommit value, we are allowing the surplus pool size to | |
2477 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2478 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2479 | * though, we'll note that we're not allowed to exceed surplus |
2480 | * and won't grow the pool anywhere else. Not until one of the | |
2481 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2482 | */ |
a5516438 | 2483 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2484 | min_count = max(count, min_count); |
6ae11b27 | 2485 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2486 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2487 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2488 | break; |
55f67141 | 2489 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2490 | } |
a5516438 | 2491 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2492 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2493 | break; |
2494 | } | |
2495 | out: | |
4eb0716e | 2496 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2497 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2498 | |
f60858f9 MK |
2499 | NODEMASK_FREE(node_alloc_noretry); |
2500 | ||
4eb0716e | 2501 | return 0; |
1da177e4 LT |
2502 | } |
2503 | ||
a3437870 NA |
2504 | #define HSTATE_ATTR_RO(_name) \ |
2505 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2506 | ||
2507 | #define HSTATE_ATTR(_name) \ | |
2508 | static struct kobj_attribute _name##_attr = \ | |
2509 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2510 | ||
2511 | static struct kobject *hugepages_kobj; | |
2512 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2513 | ||
9a305230 LS |
2514 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2515 | ||
2516 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2517 | { |
2518 | int i; | |
9a305230 | 2519 | |
a3437870 | 2520 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2521 | if (hstate_kobjs[i] == kobj) { |
2522 | if (nidp) | |
2523 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2524 | return &hstates[i]; |
9a305230 LS |
2525 | } |
2526 | ||
2527 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2528 | } |
2529 | ||
06808b08 | 2530 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2531 | struct kobj_attribute *attr, char *buf) |
2532 | { | |
9a305230 LS |
2533 | struct hstate *h; |
2534 | unsigned long nr_huge_pages; | |
2535 | int nid; | |
2536 | ||
2537 | h = kobj_to_hstate(kobj, &nid); | |
2538 | if (nid == NUMA_NO_NODE) | |
2539 | nr_huge_pages = h->nr_huge_pages; | |
2540 | else | |
2541 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2542 | ||
2543 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2544 | } |
adbe8726 | 2545 | |
238d3c13 DR |
2546 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2547 | struct hstate *h, int nid, | |
2548 | unsigned long count, size_t len) | |
a3437870 NA |
2549 | { |
2550 | int err; | |
2d0adf7e | 2551 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2552 | |
2d0adf7e OS |
2553 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2554 | return -EINVAL; | |
adbe8726 | 2555 | |
9a305230 LS |
2556 | if (nid == NUMA_NO_NODE) { |
2557 | /* | |
2558 | * global hstate attribute | |
2559 | */ | |
2560 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2561 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2562 | n_mask = &node_states[N_MEMORY]; | |
2563 | else | |
2564 | n_mask = &nodes_allowed; | |
2565 | } else { | |
9a305230 | 2566 | /* |
fd875dca MK |
2567 | * Node specific request. count adjustment happens in |
2568 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2569 | */ |
2d0adf7e OS |
2570 | init_nodemask_of_node(&nodes_allowed, nid); |
2571 | n_mask = &nodes_allowed; | |
fd875dca | 2572 | } |
9a305230 | 2573 | |
2d0adf7e | 2574 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2575 | |
4eb0716e | 2576 | return err ? err : len; |
06808b08 LS |
2577 | } |
2578 | ||
238d3c13 DR |
2579 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2580 | struct kobject *kobj, const char *buf, | |
2581 | size_t len) | |
2582 | { | |
2583 | struct hstate *h; | |
2584 | unsigned long count; | |
2585 | int nid; | |
2586 | int err; | |
2587 | ||
2588 | err = kstrtoul(buf, 10, &count); | |
2589 | if (err) | |
2590 | return err; | |
2591 | ||
2592 | h = kobj_to_hstate(kobj, &nid); | |
2593 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2594 | } | |
2595 | ||
06808b08 LS |
2596 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2597 | struct kobj_attribute *attr, char *buf) | |
2598 | { | |
2599 | return nr_hugepages_show_common(kobj, attr, buf); | |
2600 | } | |
2601 | ||
2602 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2603 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2604 | { | |
238d3c13 | 2605 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2606 | } |
2607 | HSTATE_ATTR(nr_hugepages); | |
2608 | ||
06808b08 LS |
2609 | #ifdef CONFIG_NUMA |
2610 | ||
2611 | /* | |
2612 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2613 | * huge page alloc/free. | |
2614 | */ | |
2615 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2616 | struct kobj_attribute *attr, char *buf) | |
2617 | { | |
2618 | return nr_hugepages_show_common(kobj, attr, buf); | |
2619 | } | |
2620 | ||
2621 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2622 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2623 | { | |
238d3c13 | 2624 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2625 | } |
2626 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2627 | #endif | |
2628 | ||
2629 | ||
a3437870 NA |
2630 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2631 | struct kobj_attribute *attr, char *buf) | |
2632 | { | |
9a305230 | 2633 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2634 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2635 | } | |
adbe8726 | 2636 | |
a3437870 NA |
2637 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2638 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2639 | { | |
2640 | int err; | |
2641 | unsigned long input; | |
9a305230 | 2642 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2643 | |
bae7f4ae | 2644 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2645 | return -EINVAL; |
2646 | ||
3dbb95f7 | 2647 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2648 | if (err) |
73ae31e5 | 2649 | return err; |
a3437870 NA |
2650 | |
2651 | spin_lock(&hugetlb_lock); | |
2652 | h->nr_overcommit_huge_pages = input; | |
2653 | spin_unlock(&hugetlb_lock); | |
2654 | ||
2655 | return count; | |
2656 | } | |
2657 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2658 | ||
2659 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2660 | struct kobj_attribute *attr, char *buf) | |
2661 | { | |
9a305230 LS |
2662 | struct hstate *h; |
2663 | unsigned long free_huge_pages; | |
2664 | int nid; | |
2665 | ||
2666 | h = kobj_to_hstate(kobj, &nid); | |
2667 | if (nid == NUMA_NO_NODE) | |
2668 | free_huge_pages = h->free_huge_pages; | |
2669 | else | |
2670 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2671 | ||
2672 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2673 | } |
2674 | HSTATE_ATTR_RO(free_hugepages); | |
2675 | ||
2676 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2677 | struct kobj_attribute *attr, char *buf) | |
2678 | { | |
9a305230 | 2679 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2680 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2681 | } | |
2682 | HSTATE_ATTR_RO(resv_hugepages); | |
2683 | ||
2684 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2685 | struct kobj_attribute *attr, char *buf) | |
2686 | { | |
9a305230 LS |
2687 | struct hstate *h; |
2688 | unsigned long surplus_huge_pages; | |
2689 | int nid; | |
2690 | ||
2691 | h = kobj_to_hstate(kobj, &nid); | |
2692 | if (nid == NUMA_NO_NODE) | |
2693 | surplus_huge_pages = h->surplus_huge_pages; | |
2694 | else | |
2695 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2696 | ||
2697 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2698 | } |
2699 | HSTATE_ATTR_RO(surplus_hugepages); | |
2700 | ||
2701 | static struct attribute *hstate_attrs[] = { | |
2702 | &nr_hugepages_attr.attr, | |
2703 | &nr_overcommit_hugepages_attr.attr, | |
2704 | &free_hugepages_attr.attr, | |
2705 | &resv_hugepages_attr.attr, | |
2706 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2707 | #ifdef CONFIG_NUMA |
2708 | &nr_hugepages_mempolicy_attr.attr, | |
2709 | #endif | |
a3437870 NA |
2710 | NULL, |
2711 | }; | |
2712 | ||
67e5ed96 | 2713 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2714 | .attrs = hstate_attrs, |
2715 | }; | |
2716 | ||
094e9539 JM |
2717 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2718 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2719 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2720 | { |
2721 | int retval; | |
972dc4de | 2722 | int hi = hstate_index(h); |
a3437870 | 2723 | |
9a305230 LS |
2724 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2725 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2726 | return -ENOMEM; |
2727 | ||
9a305230 | 2728 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2729 | if (retval) |
9a305230 | 2730 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2731 | |
2732 | return retval; | |
2733 | } | |
2734 | ||
2735 | static void __init hugetlb_sysfs_init(void) | |
2736 | { | |
2737 | struct hstate *h; | |
2738 | int err; | |
2739 | ||
2740 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2741 | if (!hugepages_kobj) | |
2742 | return; | |
2743 | ||
2744 | for_each_hstate(h) { | |
9a305230 LS |
2745 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2746 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2747 | if (err) |
ffb22af5 | 2748 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2749 | } |
2750 | } | |
2751 | ||
9a305230 LS |
2752 | #ifdef CONFIG_NUMA |
2753 | ||
2754 | /* | |
2755 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2756 | * with node devices in node_devices[] using a parallel array. The array |
2757 | * index of a node device or _hstate == node id. | |
2758 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2759 | * the base kernel, on the hugetlb module. |
2760 | */ | |
2761 | struct node_hstate { | |
2762 | struct kobject *hugepages_kobj; | |
2763 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2764 | }; | |
b4e289a6 | 2765 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2766 | |
2767 | /* | |
10fbcf4c | 2768 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2769 | */ |
2770 | static struct attribute *per_node_hstate_attrs[] = { | |
2771 | &nr_hugepages_attr.attr, | |
2772 | &free_hugepages_attr.attr, | |
2773 | &surplus_hugepages_attr.attr, | |
2774 | NULL, | |
2775 | }; | |
2776 | ||
67e5ed96 | 2777 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
2778 | .attrs = per_node_hstate_attrs, |
2779 | }; | |
2780 | ||
2781 | /* | |
10fbcf4c | 2782 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2783 | * Returns node id via non-NULL nidp. |
2784 | */ | |
2785 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2786 | { | |
2787 | int nid; | |
2788 | ||
2789 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2790 | struct node_hstate *nhs = &node_hstates[nid]; | |
2791 | int i; | |
2792 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2793 | if (nhs->hstate_kobjs[i] == kobj) { | |
2794 | if (nidp) | |
2795 | *nidp = nid; | |
2796 | return &hstates[i]; | |
2797 | } | |
2798 | } | |
2799 | ||
2800 | BUG(); | |
2801 | return NULL; | |
2802 | } | |
2803 | ||
2804 | /* | |
10fbcf4c | 2805 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2806 | * No-op if no hstate attributes attached. |
2807 | */ | |
3cd8b44f | 2808 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2809 | { |
2810 | struct hstate *h; | |
10fbcf4c | 2811 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2812 | |
2813 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2814 | return; /* no hstate attributes */ |
9a305230 | 2815 | |
972dc4de AK |
2816 | for_each_hstate(h) { |
2817 | int idx = hstate_index(h); | |
2818 | if (nhs->hstate_kobjs[idx]) { | |
2819 | kobject_put(nhs->hstate_kobjs[idx]); | |
2820 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2821 | } |
972dc4de | 2822 | } |
9a305230 LS |
2823 | |
2824 | kobject_put(nhs->hugepages_kobj); | |
2825 | nhs->hugepages_kobj = NULL; | |
2826 | } | |
2827 | ||
9a305230 LS |
2828 | |
2829 | /* | |
10fbcf4c | 2830 | * Register hstate attributes for a single node device. |
9a305230 LS |
2831 | * No-op if attributes already registered. |
2832 | */ | |
3cd8b44f | 2833 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2834 | { |
2835 | struct hstate *h; | |
10fbcf4c | 2836 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2837 | int err; |
2838 | ||
2839 | if (nhs->hugepages_kobj) | |
2840 | return; /* already allocated */ | |
2841 | ||
2842 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2843 | &node->dev.kobj); |
9a305230 LS |
2844 | if (!nhs->hugepages_kobj) |
2845 | return; | |
2846 | ||
2847 | for_each_hstate(h) { | |
2848 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2849 | nhs->hstate_kobjs, | |
2850 | &per_node_hstate_attr_group); | |
2851 | if (err) { | |
ffb22af5 AM |
2852 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2853 | h->name, node->dev.id); | |
9a305230 LS |
2854 | hugetlb_unregister_node(node); |
2855 | break; | |
2856 | } | |
2857 | } | |
2858 | } | |
2859 | ||
2860 | /* | |
9b5e5d0f | 2861 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2862 | * devices of nodes that have memory. All on-line nodes should have |
2863 | * registered their associated device by this time. | |
9a305230 | 2864 | */ |
7d9ca000 | 2865 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2866 | { |
2867 | int nid; | |
2868 | ||
8cebfcd0 | 2869 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2870 | struct node *node = node_devices[nid]; |
10fbcf4c | 2871 | if (node->dev.id == nid) |
9a305230 LS |
2872 | hugetlb_register_node(node); |
2873 | } | |
2874 | ||
2875 | /* | |
10fbcf4c | 2876 | * Let the node device driver know we're here so it can |
9a305230 LS |
2877 | * [un]register hstate attributes on node hotplug. |
2878 | */ | |
2879 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2880 | hugetlb_unregister_node); | |
2881 | } | |
2882 | #else /* !CONFIG_NUMA */ | |
2883 | ||
2884 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2885 | { | |
2886 | BUG(); | |
2887 | if (nidp) | |
2888 | *nidp = -1; | |
2889 | return NULL; | |
2890 | } | |
2891 | ||
9a305230 LS |
2892 | static void hugetlb_register_all_nodes(void) { } |
2893 | ||
2894 | #endif | |
2895 | ||
a3437870 NA |
2896 | static int __init hugetlb_init(void) |
2897 | { | |
8382d914 DB |
2898 | int i; |
2899 | ||
457c1b27 | 2900 | if (!hugepages_supported()) |
0ef89d25 | 2901 | return 0; |
a3437870 | 2902 | |
e11bfbfc | 2903 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
2904 | if (default_hstate_size != 0) { |
2905 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
2906 | default_hstate_size, HPAGE_SIZE); | |
2907 | } | |
2908 | ||
e11bfbfc NP |
2909 | default_hstate_size = HPAGE_SIZE; |
2910 | if (!size_to_hstate(default_hstate_size)) | |
2911 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2912 | } |
972dc4de | 2913 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2914 | if (default_hstate_max_huge_pages) { |
2915 | if (!default_hstate.max_huge_pages) | |
2916 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2917 | } | |
a3437870 NA |
2918 | |
2919 | hugetlb_init_hstates(); | |
aa888a74 | 2920 | gather_bootmem_prealloc(); |
a3437870 NA |
2921 | report_hugepages(); |
2922 | ||
2923 | hugetlb_sysfs_init(); | |
9a305230 | 2924 | hugetlb_register_all_nodes(); |
7179e7bf | 2925 | hugetlb_cgroup_file_init(); |
9a305230 | 2926 | |
8382d914 DB |
2927 | #ifdef CONFIG_SMP |
2928 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2929 | #else | |
2930 | num_fault_mutexes = 1; | |
2931 | #endif | |
c672c7f2 | 2932 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
2933 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
2934 | GFP_KERNEL); | |
c672c7f2 | 2935 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2936 | |
2937 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2938 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2939 | return 0; |
2940 | } | |
3e89e1c5 | 2941 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2942 | |
2943 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2944 | void __init hugetlb_bad_size(void) |
2945 | { | |
2946 | parsed_valid_hugepagesz = false; | |
2947 | } | |
2948 | ||
d00181b9 | 2949 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2950 | { |
2951 | struct hstate *h; | |
8faa8b07 AK |
2952 | unsigned long i; |
2953 | ||
a3437870 | 2954 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2955 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2956 | return; |
2957 | } | |
47d38344 | 2958 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2959 | BUG_ON(order == 0); |
47d38344 | 2960 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2961 | h->order = order; |
2962 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2963 | h->nr_huge_pages = 0; |
2964 | h->free_huge_pages = 0; | |
2965 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2966 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2967 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2968 | h->next_nid_to_alloc = first_memory_node; |
2969 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2970 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2971 | huge_page_size(h)/1024); | |
8faa8b07 | 2972 | |
a3437870 NA |
2973 | parsed_hstate = h; |
2974 | } | |
2975 | ||
e11bfbfc | 2976 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2977 | { |
2978 | unsigned long *mhp; | |
8faa8b07 | 2979 | static unsigned long *last_mhp; |
a3437870 | 2980 | |
9fee021d VT |
2981 | if (!parsed_valid_hugepagesz) { |
2982 | pr_warn("hugepages = %s preceded by " | |
2983 | "an unsupported hugepagesz, ignoring\n", s); | |
2984 | parsed_valid_hugepagesz = true; | |
2985 | return 1; | |
2986 | } | |
a3437870 | 2987 | /* |
47d38344 | 2988 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2989 | * so this hugepages= parameter goes to the "default hstate". |
2990 | */ | |
9fee021d | 2991 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2992 | mhp = &default_hstate_max_huge_pages; |
2993 | else | |
2994 | mhp = &parsed_hstate->max_huge_pages; | |
2995 | ||
8faa8b07 | 2996 | if (mhp == last_mhp) { |
598d8091 | 2997 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2998 | return 1; |
2999 | } | |
3000 | ||
a3437870 NA |
3001 | if (sscanf(s, "%lu", mhp) <= 0) |
3002 | *mhp = 0; | |
3003 | ||
8faa8b07 AK |
3004 | /* |
3005 | * Global state is always initialized later in hugetlb_init. | |
3006 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
3007 | * use the bootmem allocator. | |
3008 | */ | |
47d38344 | 3009 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
3010 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3011 | ||
3012 | last_mhp = mhp; | |
3013 | ||
a3437870 NA |
3014 | return 1; |
3015 | } | |
e11bfbfc NP |
3016 | __setup("hugepages=", hugetlb_nrpages_setup); |
3017 | ||
3018 | static int __init hugetlb_default_setup(char *s) | |
3019 | { | |
3020 | default_hstate_size = memparse(s, &s); | |
3021 | return 1; | |
3022 | } | |
3023 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 3024 | |
8a213460 NA |
3025 | static unsigned int cpuset_mems_nr(unsigned int *array) |
3026 | { | |
3027 | int node; | |
3028 | unsigned int nr = 0; | |
3029 | ||
3030 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
3031 | nr += array[node]; | |
3032 | ||
3033 | return nr; | |
3034 | } | |
3035 | ||
3036 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
3037 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3038 | struct ctl_table *table, int write, | |
3039 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 3040 | { |
e5ff2159 | 3041 | struct hstate *h = &default_hstate; |
238d3c13 | 3042 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3043 | int ret; |
e5ff2159 | 3044 | |
457c1b27 | 3045 | if (!hugepages_supported()) |
86613628 | 3046 | return -EOPNOTSUPP; |
457c1b27 | 3047 | |
e5ff2159 AK |
3048 | table->data = &tmp; |
3049 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3050 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3051 | if (ret) | |
3052 | goto out; | |
e5ff2159 | 3053 | |
238d3c13 DR |
3054 | if (write) |
3055 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3056 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3057 | out: |
3058 | return ret; | |
1da177e4 | 3059 | } |
396faf03 | 3060 | |
06808b08 LS |
3061 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
3062 | void __user *buffer, size_t *length, loff_t *ppos) | |
3063 | { | |
3064 | ||
3065 | return hugetlb_sysctl_handler_common(false, table, write, | |
3066 | buffer, length, ppos); | |
3067 | } | |
3068 | ||
3069 | #ifdef CONFIG_NUMA | |
3070 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
3071 | void __user *buffer, size_t *length, loff_t *ppos) | |
3072 | { | |
3073 | return hugetlb_sysctl_handler_common(true, table, write, | |
3074 | buffer, length, ppos); | |
3075 | } | |
3076 | #endif /* CONFIG_NUMA */ | |
3077 | ||
a3d0c6aa | 3078 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 3079 | void __user *buffer, |
a3d0c6aa NA |
3080 | size_t *length, loff_t *ppos) |
3081 | { | |
a5516438 | 3082 | struct hstate *h = &default_hstate; |
e5ff2159 | 3083 | unsigned long tmp; |
08d4a246 | 3084 | int ret; |
e5ff2159 | 3085 | |
457c1b27 | 3086 | if (!hugepages_supported()) |
86613628 | 3087 | return -EOPNOTSUPP; |
457c1b27 | 3088 | |
c033a93c | 3089 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3090 | |
bae7f4ae | 3091 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3092 | return -EINVAL; |
3093 | ||
e5ff2159 AK |
3094 | table->data = &tmp; |
3095 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3096 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3097 | if (ret) | |
3098 | goto out; | |
e5ff2159 AK |
3099 | |
3100 | if (write) { | |
3101 | spin_lock(&hugetlb_lock); | |
3102 | h->nr_overcommit_huge_pages = tmp; | |
3103 | spin_unlock(&hugetlb_lock); | |
3104 | } | |
08d4a246 MH |
3105 | out: |
3106 | return ret; | |
a3d0c6aa NA |
3107 | } |
3108 | ||
1da177e4 LT |
3109 | #endif /* CONFIG_SYSCTL */ |
3110 | ||
e1759c21 | 3111 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3112 | { |
fcb2b0c5 RG |
3113 | struct hstate *h; |
3114 | unsigned long total = 0; | |
3115 | ||
457c1b27 NA |
3116 | if (!hugepages_supported()) |
3117 | return; | |
fcb2b0c5 RG |
3118 | |
3119 | for_each_hstate(h) { | |
3120 | unsigned long count = h->nr_huge_pages; | |
3121 | ||
3122 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3123 | ||
3124 | if (h == &default_hstate) | |
3125 | seq_printf(m, | |
3126 | "HugePages_Total: %5lu\n" | |
3127 | "HugePages_Free: %5lu\n" | |
3128 | "HugePages_Rsvd: %5lu\n" | |
3129 | "HugePages_Surp: %5lu\n" | |
3130 | "Hugepagesize: %8lu kB\n", | |
3131 | count, | |
3132 | h->free_huge_pages, | |
3133 | h->resv_huge_pages, | |
3134 | h->surplus_huge_pages, | |
3135 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3136 | } | |
3137 | ||
3138 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3139 | } |
3140 | ||
3141 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3142 | { | |
a5516438 | 3143 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3144 | if (!hugepages_supported()) |
3145 | return 0; | |
1da177e4 LT |
3146 | return sprintf(buf, |
3147 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3148 | "Node %d HugePages_Free: %5u\n" |
3149 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3150 | nid, h->nr_huge_pages_node[nid], |
3151 | nid, h->free_huge_pages_node[nid], | |
3152 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3153 | } |
3154 | ||
949f7ec5 DR |
3155 | void hugetlb_show_meminfo(void) |
3156 | { | |
3157 | struct hstate *h; | |
3158 | int nid; | |
3159 | ||
457c1b27 NA |
3160 | if (!hugepages_supported()) |
3161 | return; | |
3162 | ||
949f7ec5 DR |
3163 | for_each_node_state(nid, N_MEMORY) |
3164 | for_each_hstate(h) | |
3165 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3166 | nid, | |
3167 | h->nr_huge_pages_node[nid], | |
3168 | h->free_huge_pages_node[nid], | |
3169 | h->surplus_huge_pages_node[nid], | |
3170 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3171 | } | |
3172 | ||
5d317b2b NH |
3173 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3174 | { | |
3175 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3176 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3177 | } | |
3178 | ||
1da177e4 LT |
3179 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3180 | unsigned long hugetlb_total_pages(void) | |
3181 | { | |
d0028588 WL |
3182 | struct hstate *h; |
3183 | unsigned long nr_total_pages = 0; | |
3184 | ||
3185 | for_each_hstate(h) | |
3186 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3187 | return nr_total_pages; | |
1da177e4 | 3188 | } |
1da177e4 | 3189 | |
a5516438 | 3190 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3191 | { |
3192 | int ret = -ENOMEM; | |
3193 | ||
3194 | spin_lock(&hugetlb_lock); | |
3195 | /* | |
3196 | * When cpuset is configured, it breaks the strict hugetlb page | |
3197 | * reservation as the accounting is done on a global variable. Such | |
3198 | * reservation is completely rubbish in the presence of cpuset because | |
3199 | * the reservation is not checked against page availability for the | |
3200 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3201 | * with lack of free htlb page in cpuset that the task is in. | |
3202 | * Attempt to enforce strict accounting with cpuset is almost | |
3203 | * impossible (or too ugly) because cpuset is too fluid that | |
3204 | * task or memory node can be dynamically moved between cpusets. | |
3205 | * | |
3206 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3207 | * undesirable. However, in order to preserve some of the semantics, | |
3208 | * we fall back to check against current free page availability as | |
3209 | * a best attempt and hopefully to minimize the impact of changing | |
3210 | * semantics that cpuset has. | |
3211 | */ | |
3212 | if (delta > 0) { | |
a5516438 | 3213 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3214 | goto out; |
3215 | ||
a5516438 AK |
3216 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3217 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3218 | goto out; |
3219 | } | |
3220 | } | |
3221 | ||
3222 | ret = 0; | |
3223 | if (delta < 0) | |
a5516438 | 3224 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3225 | |
3226 | out: | |
3227 | spin_unlock(&hugetlb_lock); | |
3228 | return ret; | |
3229 | } | |
3230 | ||
84afd99b AW |
3231 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3232 | { | |
f522c3ac | 3233 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3234 | |
3235 | /* | |
3236 | * This new VMA should share its siblings reservation map if present. | |
3237 | * The VMA will only ever have a valid reservation map pointer where | |
3238 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3239 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3240 | * after this open call completes. It is therefore safe to take a |
3241 | * new reference here without additional locking. | |
3242 | */ | |
4e35f483 | 3243 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3244 | kref_get(&resv->refs); |
84afd99b AW |
3245 | } |
3246 | ||
a1e78772 MG |
3247 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3248 | { | |
a5516438 | 3249 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3250 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3251 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3252 | unsigned long reserve, start, end; |
1c5ecae3 | 3253 | long gbl_reserve; |
84afd99b | 3254 | |
4e35f483 JK |
3255 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3256 | return; | |
84afd99b | 3257 | |
4e35f483 JK |
3258 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3259 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3260 | |
4e35f483 | 3261 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3262 | |
4e35f483 JK |
3263 | kref_put(&resv->refs, resv_map_release); |
3264 | ||
3265 | if (reserve) { | |
1c5ecae3 MK |
3266 | /* |
3267 | * Decrement reserve counts. The global reserve count may be | |
3268 | * adjusted if the subpool has a minimum size. | |
3269 | */ | |
3270 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3271 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3272 | } |
a1e78772 MG |
3273 | } |
3274 | ||
31383c68 DW |
3275 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3276 | { | |
3277 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3278 | return -EINVAL; | |
3279 | return 0; | |
3280 | } | |
3281 | ||
05ea8860 DW |
3282 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3283 | { | |
3284 | struct hstate *hstate = hstate_vma(vma); | |
3285 | ||
3286 | return 1UL << huge_page_shift(hstate); | |
3287 | } | |
3288 | ||
1da177e4 LT |
3289 | /* |
3290 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3291 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3292 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3293 | * this far. | |
3294 | */ | |
b3ec9f33 | 3295 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3296 | { |
3297 | BUG(); | |
d0217ac0 | 3298 | return 0; |
1da177e4 LT |
3299 | } |
3300 | ||
eec3636a JC |
3301 | /* |
3302 | * When a new function is introduced to vm_operations_struct and added | |
3303 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3304 | * This is because under System V memory model, mappings created via | |
3305 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3306 | * their original vm_ops are overwritten with shm_vm_ops. | |
3307 | */ | |
f0f37e2f | 3308 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3309 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3310 | .open = hugetlb_vm_op_open, |
a1e78772 | 3311 | .close = hugetlb_vm_op_close, |
31383c68 | 3312 | .split = hugetlb_vm_op_split, |
05ea8860 | 3313 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3314 | }; |
3315 | ||
1e8f889b DG |
3316 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3317 | int writable) | |
63551ae0 DG |
3318 | { |
3319 | pte_t entry; | |
3320 | ||
1e8f889b | 3321 | if (writable) { |
106c992a GS |
3322 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3323 | vma->vm_page_prot))); | |
63551ae0 | 3324 | } else { |
106c992a GS |
3325 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3326 | vma->vm_page_prot)); | |
63551ae0 DG |
3327 | } |
3328 | entry = pte_mkyoung(entry); | |
3329 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3330 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3331 | |
3332 | return entry; | |
3333 | } | |
3334 | ||
1e8f889b DG |
3335 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3336 | unsigned long address, pte_t *ptep) | |
3337 | { | |
3338 | pte_t entry; | |
3339 | ||
106c992a | 3340 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3341 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3342 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3343 | } |
3344 | ||
d5ed7444 | 3345 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3346 | { |
3347 | swp_entry_t swp; | |
3348 | ||
3349 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3350 | return false; |
4a705fef NH |
3351 | swp = pte_to_swp_entry(pte); |
3352 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3353 | return true; |
4a705fef | 3354 | else |
d5ed7444 | 3355 | return false; |
4a705fef NH |
3356 | } |
3357 | ||
3358 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3359 | { | |
3360 | swp_entry_t swp; | |
3361 | ||
3362 | if (huge_pte_none(pte) || pte_present(pte)) | |
3363 | return 0; | |
3364 | swp = pte_to_swp_entry(pte); | |
3365 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3366 | return 1; | |
3367 | else | |
3368 | return 0; | |
3369 | } | |
1e8f889b | 3370 | |
63551ae0 DG |
3371 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3372 | struct vm_area_struct *vma) | |
3373 | { | |
5e41540c | 3374 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3375 | struct page *ptepage; |
1c59827d | 3376 | unsigned long addr; |
1e8f889b | 3377 | int cow; |
a5516438 AK |
3378 | struct hstate *h = hstate_vma(vma); |
3379 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3380 | struct mmu_notifier_range range; |
e8569dd2 | 3381 | int ret = 0; |
1e8f889b DG |
3382 | |
3383 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3384 | |
ac46d4f3 | 3385 | if (cow) { |
7269f999 | 3386 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3387 | vma->vm_start, |
ac46d4f3 JG |
3388 | vma->vm_end); |
3389 | mmu_notifier_invalidate_range_start(&range); | |
3390 | } | |
e8569dd2 | 3391 | |
a5516438 | 3392 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3393 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3394 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3395 | if (!src_pte) |
3396 | continue; | |
a5516438 | 3397 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3398 | if (!dst_pte) { |
3399 | ret = -ENOMEM; | |
3400 | break; | |
3401 | } | |
c5c99429 | 3402 | |
5e41540c MK |
3403 | /* |
3404 | * If the pagetables are shared don't copy or take references. | |
3405 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3406 | * | |
3407 | * However, src could have 'unshared' and dst shares with | |
3408 | * another vma. If dst_pte !none, this implies sharing. | |
3409 | * Check here before taking page table lock, and once again | |
3410 | * after taking the lock below. | |
3411 | */ | |
3412 | dst_entry = huge_ptep_get(dst_pte); | |
3413 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3414 | continue; |
3415 | ||
cb900f41 KS |
3416 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3417 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3418 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3419 | entry = huge_ptep_get(src_pte); |
5e41540c MK |
3420 | dst_entry = huge_ptep_get(dst_pte); |
3421 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3422 | /* | |
3423 | * Skip if src entry none. Also, skip in the | |
3424 | * unlikely case dst entry !none as this implies | |
3425 | * sharing with another vma. | |
3426 | */ | |
4a705fef NH |
3427 | ; |
3428 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3429 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3430 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3431 | ||
3432 | if (is_write_migration_entry(swp_entry) && cow) { | |
3433 | /* | |
3434 | * COW mappings require pages in both | |
3435 | * parent and child to be set to read. | |
3436 | */ | |
3437 | make_migration_entry_read(&swp_entry); | |
3438 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3439 | set_huge_swap_pte_at(src, addr, src_pte, |
3440 | entry, sz); | |
4a705fef | 3441 | } |
e5251fd4 | 3442 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3443 | } else { |
34ee645e | 3444 | if (cow) { |
0f10851e JG |
3445 | /* |
3446 | * No need to notify as we are downgrading page | |
3447 | * table protection not changing it to point | |
3448 | * to a new page. | |
3449 | * | |
ad56b738 | 3450 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3451 | */ |
7f2e9525 | 3452 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3453 | } |
0253d634 | 3454 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3455 | ptepage = pte_page(entry); |
3456 | get_page(ptepage); | |
53f9263b | 3457 | page_dup_rmap(ptepage, true); |
1c59827d | 3458 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3459 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3460 | } |
cb900f41 KS |
3461 | spin_unlock(src_ptl); |
3462 | spin_unlock(dst_ptl); | |
63551ae0 | 3463 | } |
63551ae0 | 3464 | |
e8569dd2 | 3465 | if (cow) |
ac46d4f3 | 3466 | mmu_notifier_invalidate_range_end(&range); |
e8569dd2 AS |
3467 | |
3468 | return ret; | |
63551ae0 DG |
3469 | } |
3470 | ||
24669e58 AK |
3471 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3472 | unsigned long start, unsigned long end, | |
3473 | struct page *ref_page) | |
63551ae0 DG |
3474 | { |
3475 | struct mm_struct *mm = vma->vm_mm; | |
3476 | unsigned long address; | |
c7546f8f | 3477 | pte_t *ptep; |
63551ae0 | 3478 | pte_t pte; |
cb900f41 | 3479 | spinlock_t *ptl; |
63551ae0 | 3480 | struct page *page; |
a5516438 AK |
3481 | struct hstate *h = hstate_vma(vma); |
3482 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3483 | struct mmu_notifier_range range; |
a5516438 | 3484 | |
63551ae0 | 3485 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3486 | BUG_ON(start & ~huge_page_mask(h)); |
3487 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3488 | |
07e32661 AK |
3489 | /* |
3490 | * This is a hugetlb vma, all the pte entries should point | |
3491 | * to huge page. | |
3492 | */ | |
ed6a7935 | 3493 | tlb_change_page_size(tlb, sz); |
24669e58 | 3494 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3495 | |
3496 | /* | |
3497 | * If sharing possible, alert mmu notifiers of worst case. | |
3498 | */ | |
6f4f13e8 JG |
3499 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3500 | end); | |
ac46d4f3 JG |
3501 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3502 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3503 | address = start; |
569f48b8 | 3504 | for (; address < end; address += sz) { |
7868a208 | 3505 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3506 | if (!ptep) |
c7546f8f DG |
3507 | continue; |
3508 | ||
cb900f41 | 3509 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3510 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3511 | spin_unlock(ptl); | |
dff11abe MK |
3512 | /* |
3513 | * We just unmapped a page of PMDs by clearing a PUD. | |
3514 | * The caller's TLB flush range should cover this area. | |
3515 | */ | |
31d49da5 AK |
3516 | continue; |
3517 | } | |
39dde65c | 3518 | |
6629326b | 3519 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3520 | if (huge_pte_none(pte)) { |
3521 | spin_unlock(ptl); | |
3522 | continue; | |
3523 | } | |
6629326b HD |
3524 | |
3525 | /* | |
9fbc1f63 NH |
3526 | * Migrating hugepage or HWPoisoned hugepage is already |
3527 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3528 | */ |
9fbc1f63 | 3529 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3530 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3531 | spin_unlock(ptl); |
3532 | continue; | |
8c4894c6 | 3533 | } |
6629326b HD |
3534 | |
3535 | page = pte_page(pte); | |
04f2cbe3 MG |
3536 | /* |
3537 | * If a reference page is supplied, it is because a specific | |
3538 | * page is being unmapped, not a range. Ensure the page we | |
3539 | * are about to unmap is the actual page of interest. | |
3540 | */ | |
3541 | if (ref_page) { | |
31d49da5 AK |
3542 | if (page != ref_page) { |
3543 | spin_unlock(ptl); | |
3544 | continue; | |
3545 | } | |
04f2cbe3 MG |
3546 | /* |
3547 | * Mark the VMA as having unmapped its page so that | |
3548 | * future faults in this VMA will fail rather than | |
3549 | * looking like data was lost | |
3550 | */ | |
3551 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3552 | } | |
3553 | ||
c7546f8f | 3554 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3555 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3556 | if (huge_pte_dirty(pte)) |
6649a386 | 3557 | set_page_dirty(page); |
9e81130b | 3558 | |
5d317b2b | 3559 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3560 | page_remove_rmap(page, true); |
31d49da5 | 3561 | |
cb900f41 | 3562 | spin_unlock(ptl); |
e77b0852 | 3563 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3564 | /* |
3565 | * Bail out after unmapping reference page if supplied | |
3566 | */ | |
3567 | if (ref_page) | |
3568 | break; | |
fe1668ae | 3569 | } |
ac46d4f3 | 3570 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 3571 | tlb_end_vma(tlb, vma); |
1da177e4 | 3572 | } |
63551ae0 | 3573 | |
d833352a MG |
3574 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3575 | struct vm_area_struct *vma, unsigned long start, | |
3576 | unsigned long end, struct page *ref_page) | |
3577 | { | |
3578 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3579 | ||
3580 | /* | |
3581 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3582 | * test will fail on a vma being torn down, and not grab a page table | |
3583 | * on its way out. We're lucky that the flag has such an appropriate | |
3584 | * name, and can in fact be safely cleared here. We could clear it | |
3585 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3586 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3587 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3588 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3589 | */ |
3590 | vma->vm_flags &= ~VM_MAYSHARE; | |
3591 | } | |
3592 | ||
502717f4 | 3593 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3594 | unsigned long end, struct page *ref_page) |
502717f4 | 3595 | { |
24669e58 AK |
3596 | struct mm_struct *mm; |
3597 | struct mmu_gather tlb; | |
dff11abe MK |
3598 | unsigned long tlb_start = start; |
3599 | unsigned long tlb_end = end; | |
3600 | ||
3601 | /* | |
3602 | * If shared PMDs were possibly used within this vma range, adjust | |
3603 | * start/end for worst case tlb flushing. | |
3604 | * Note that we can not be sure if PMDs are shared until we try to | |
3605 | * unmap pages. However, we want to make sure TLB flushing covers | |
3606 | * the largest possible range. | |
3607 | */ | |
3608 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
24669e58 AK |
3609 | |
3610 | mm = vma->vm_mm; | |
3611 | ||
dff11abe | 3612 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); |
24669e58 | 3613 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
dff11abe | 3614 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); |
502717f4 CK |
3615 | } |
3616 | ||
04f2cbe3 MG |
3617 | /* |
3618 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3619 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3620 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3621 | * same region. | |
3622 | */ | |
2f4612af DB |
3623 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3624 | struct page *page, unsigned long address) | |
04f2cbe3 | 3625 | { |
7526674d | 3626 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3627 | struct vm_area_struct *iter_vma; |
3628 | struct address_space *mapping; | |
04f2cbe3 MG |
3629 | pgoff_t pgoff; |
3630 | ||
3631 | /* | |
3632 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3633 | * from page cache lookup which is in HPAGE_SIZE units. | |
3634 | */ | |
7526674d | 3635 | address = address & huge_page_mask(h); |
36e4f20a MH |
3636 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3637 | vma->vm_pgoff; | |
93c76a3d | 3638 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3639 | |
4eb2b1dc MG |
3640 | /* |
3641 | * Take the mapping lock for the duration of the table walk. As | |
3642 | * this mapping should be shared between all the VMAs, | |
3643 | * __unmap_hugepage_range() is called as the lock is already held | |
3644 | */ | |
83cde9e8 | 3645 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3646 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3647 | /* Do not unmap the current VMA */ |
3648 | if (iter_vma == vma) | |
3649 | continue; | |
3650 | ||
2f84a899 MG |
3651 | /* |
3652 | * Shared VMAs have their own reserves and do not affect | |
3653 | * MAP_PRIVATE accounting but it is possible that a shared | |
3654 | * VMA is using the same page so check and skip such VMAs. | |
3655 | */ | |
3656 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3657 | continue; | |
3658 | ||
04f2cbe3 MG |
3659 | /* |
3660 | * Unmap the page from other VMAs without their own reserves. | |
3661 | * They get marked to be SIGKILLed if they fault in these | |
3662 | * areas. This is because a future no-page fault on this VMA | |
3663 | * could insert a zeroed page instead of the data existing | |
3664 | * from the time of fork. This would look like data corruption | |
3665 | */ | |
3666 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3667 | unmap_hugepage_range(iter_vma, address, |
3668 | address + huge_page_size(h), page); | |
04f2cbe3 | 3669 | } |
83cde9e8 | 3670 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3671 | } |
3672 | ||
0fe6e20b NH |
3673 | /* |
3674 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3675 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3676 | * cannot race with other handlers or page migration. | |
3677 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3678 | */ |
2b740303 | 3679 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 3680 | unsigned long address, pte_t *ptep, |
3999f52e | 3681 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3682 | { |
3999f52e | 3683 | pte_t pte; |
a5516438 | 3684 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3685 | struct page *old_page, *new_page; |
2b740303 SJ |
3686 | int outside_reserve = 0; |
3687 | vm_fault_t ret = 0; | |
974e6d66 | 3688 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 3689 | struct mmu_notifier_range range; |
1e8f889b | 3690 | |
3999f52e | 3691 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3692 | old_page = pte_page(pte); |
3693 | ||
04f2cbe3 | 3694 | retry_avoidcopy: |
1e8f889b DG |
3695 | /* If no-one else is actually using this page, avoid the copy |
3696 | * and just make the page writable */ | |
37a2140d | 3697 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3698 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 3699 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 3700 | return 0; |
1e8f889b DG |
3701 | } |
3702 | ||
04f2cbe3 MG |
3703 | /* |
3704 | * If the process that created a MAP_PRIVATE mapping is about to | |
3705 | * perform a COW due to a shared page count, attempt to satisfy | |
3706 | * the allocation without using the existing reserves. The pagecache | |
3707 | * page is used to determine if the reserve at this address was | |
3708 | * consumed or not. If reserves were used, a partial faulted mapping | |
3709 | * at the time of fork() could consume its reserves on COW instead | |
3710 | * of the full address range. | |
3711 | */ | |
5944d011 | 3712 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3713 | old_page != pagecache_page) |
3714 | outside_reserve = 1; | |
3715 | ||
09cbfeaf | 3716 | get_page(old_page); |
b76c8cfb | 3717 | |
ad4404a2 DB |
3718 | /* |
3719 | * Drop page table lock as buddy allocator may be called. It will | |
3720 | * be acquired again before returning to the caller, as expected. | |
3721 | */ | |
cb900f41 | 3722 | spin_unlock(ptl); |
5b7a1d40 | 3723 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 3724 | |
2fc39cec | 3725 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3726 | /* |
3727 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3728 | * it is due to references held by a child and an insufficient | |
3729 | * huge page pool. To guarantee the original mappers | |
3730 | * reliability, unmap the page from child processes. The child | |
3731 | * may get SIGKILLed if it later faults. | |
3732 | */ | |
3733 | if (outside_reserve) { | |
09cbfeaf | 3734 | put_page(old_page); |
04f2cbe3 | 3735 | BUG_ON(huge_pte_none(pte)); |
5b7a1d40 | 3736 | unmap_ref_private(mm, vma, old_page, haddr); |
2f4612af DB |
3737 | BUG_ON(huge_pte_none(pte)); |
3738 | spin_lock(ptl); | |
5b7a1d40 | 3739 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
3740 | if (likely(ptep && |
3741 | pte_same(huge_ptep_get(ptep), pte))) | |
3742 | goto retry_avoidcopy; | |
3743 | /* | |
3744 | * race occurs while re-acquiring page table | |
3745 | * lock, and our job is done. | |
3746 | */ | |
3747 | return 0; | |
04f2cbe3 MG |
3748 | } |
3749 | ||
2b740303 | 3750 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 3751 | goto out_release_old; |
1e8f889b DG |
3752 | } |
3753 | ||
0fe6e20b NH |
3754 | /* |
3755 | * When the original hugepage is shared one, it does not have | |
3756 | * anon_vma prepared. | |
3757 | */ | |
44e2aa93 | 3758 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3759 | ret = VM_FAULT_OOM; |
3760 | goto out_release_all; | |
44e2aa93 | 3761 | } |
0fe6e20b | 3762 | |
974e6d66 | 3763 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 3764 | pages_per_huge_page(h)); |
0ed361de | 3765 | __SetPageUptodate(new_page); |
1e8f889b | 3766 | |
7269f999 | 3767 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 3768 | haddr + huge_page_size(h)); |
ac46d4f3 | 3769 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 3770 | |
b76c8cfb | 3771 | /* |
cb900f41 | 3772 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3773 | * before the page tables are altered |
3774 | */ | |
cb900f41 | 3775 | spin_lock(ptl); |
5b7a1d40 | 3776 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 3777 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3778 | ClearPagePrivate(new_page); |
3779 | ||
1e8f889b | 3780 | /* Break COW */ |
5b7a1d40 | 3781 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 3782 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 3783 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 3784 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 3785 | page_remove_rmap(old_page, true); |
5b7a1d40 | 3786 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
cb6acd01 | 3787 | set_page_huge_active(new_page); |
1e8f889b DG |
3788 | /* Make the old page be freed below */ |
3789 | new_page = old_page; | |
3790 | } | |
cb900f41 | 3791 | spin_unlock(ptl); |
ac46d4f3 | 3792 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 3793 | out_release_all: |
5b7a1d40 | 3794 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 3795 | put_page(new_page); |
ad4404a2 | 3796 | out_release_old: |
09cbfeaf | 3797 | put_page(old_page); |
8312034f | 3798 | |
ad4404a2 DB |
3799 | spin_lock(ptl); /* Caller expects lock to be held */ |
3800 | return ret; | |
1e8f889b DG |
3801 | } |
3802 | ||
04f2cbe3 | 3803 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3804 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3805 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3806 | { |
3807 | struct address_space *mapping; | |
e7c4b0bf | 3808 | pgoff_t idx; |
04f2cbe3 MG |
3809 | |
3810 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3811 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3812 | |
3813 | return find_lock_page(mapping, idx); | |
3814 | } | |
3815 | ||
3ae77f43 HD |
3816 | /* |
3817 | * Return whether there is a pagecache page to back given address within VMA. | |
3818 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3819 | */ | |
3820 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3821 | struct vm_area_struct *vma, unsigned long address) |
3822 | { | |
3823 | struct address_space *mapping; | |
3824 | pgoff_t idx; | |
3825 | struct page *page; | |
3826 | ||
3827 | mapping = vma->vm_file->f_mapping; | |
3828 | idx = vma_hugecache_offset(h, vma, address); | |
3829 | ||
3830 | page = find_get_page(mapping, idx); | |
3831 | if (page) | |
3832 | put_page(page); | |
3833 | return page != NULL; | |
3834 | } | |
3835 | ||
ab76ad54 MK |
3836 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3837 | pgoff_t idx) | |
3838 | { | |
3839 | struct inode *inode = mapping->host; | |
3840 | struct hstate *h = hstate_inode(inode); | |
3841 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3842 | ||
3843 | if (err) | |
3844 | return err; | |
3845 | ClearPagePrivate(page); | |
3846 | ||
22146c3c MK |
3847 | /* |
3848 | * set page dirty so that it will not be removed from cache/file | |
3849 | * by non-hugetlbfs specific code paths. | |
3850 | */ | |
3851 | set_page_dirty(page); | |
3852 | ||
ab76ad54 MK |
3853 | spin_lock(&inode->i_lock); |
3854 | inode->i_blocks += blocks_per_huge_page(h); | |
3855 | spin_unlock(&inode->i_lock); | |
3856 | return 0; | |
3857 | } | |
3858 | ||
2b740303 SJ |
3859 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
3860 | struct vm_area_struct *vma, | |
3861 | struct address_space *mapping, pgoff_t idx, | |
3862 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3863 | { |
a5516438 | 3864 | struct hstate *h = hstate_vma(vma); |
2b740303 | 3865 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3866 | int anon_rmap = 0; |
4c887265 | 3867 | unsigned long size; |
4c887265 | 3868 | struct page *page; |
1e8f889b | 3869 | pte_t new_pte; |
cb900f41 | 3870 | spinlock_t *ptl; |
285b8dca | 3871 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 3872 | bool new_page = false; |
4c887265 | 3873 | |
04f2cbe3 MG |
3874 | /* |
3875 | * Currently, we are forced to kill the process in the event the | |
3876 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3877 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3878 | */ |
3879 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3880 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3881 | current->pid); |
04f2cbe3 MG |
3882 | return ret; |
3883 | } | |
3884 | ||
4c887265 | 3885 | /* |
e7c58097 MK |
3886 | * Use page lock to guard against racing truncation |
3887 | * before we get page_table_lock. | |
4c887265 | 3888 | */ |
6bda666a CL |
3889 | retry: |
3890 | page = find_lock_page(mapping, idx); | |
3891 | if (!page) { | |
e7c58097 MK |
3892 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
3893 | if (idx >= size) | |
3894 | goto out; | |
3895 | ||
1a1aad8a MK |
3896 | /* |
3897 | * Check for page in userfault range | |
3898 | */ | |
3899 | if (userfaultfd_missing(vma)) { | |
3900 | u32 hash; | |
3901 | struct vm_fault vmf = { | |
3902 | .vma = vma, | |
285b8dca | 3903 | .address = haddr, |
1a1aad8a MK |
3904 | .flags = flags, |
3905 | /* | |
3906 | * Hard to debug if it ends up being | |
3907 | * used by a callee that assumes | |
3908 | * something about the other | |
3909 | * uninitialized fields... same as in | |
3910 | * memory.c | |
3911 | */ | |
3912 | }; | |
3913 | ||
3914 | /* | |
ddeaab32 MK |
3915 | * hugetlb_fault_mutex must be dropped before |
3916 | * handling userfault. Reacquire after handling | |
3917 | * fault to make calling code simpler. | |
1a1aad8a | 3918 | */ |
1b426bac | 3919 | hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr); |
1a1aad8a MK |
3920 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
3921 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3922 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3923 | goto out; | |
3924 | } | |
3925 | ||
285b8dca | 3926 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 3927 | if (IS_ERR(page)) { |
4643d67e MK |
3928 | /* |
3929 | * Returning error will result in faulting task being | |
3930 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
3931 | * tasks from racing to fault in the same page which | |
3932 | * could result in false unable to allocate errors. | |
3933 | * Page migration does not take the fault mutex, but | |
3934 | * does a clear then write of pte's under page table | |
3935 | * lock. Page fault code could race with migration, | |
3936 | * notice the clear pte and try to allocate a page | |
3937 | * here. Before returning error, get ptl and make | |
3938 | * sure there really is no pte entry. | |
3939 | */ | |
3940 | ptl = huge_pte_lock(h, mm, ptep); | |
3941 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
3942 | ret = 0; | |
3943 | spin_unlock(ptl); | |
3944 | goto out; | |
3945 | } | |
3946 | spin_unlock(ptl); | |
2b740303 | 3947 | ret = vmf_error(PTR_ERR(page)); |
6bda666a CL |
3948 | goto out; |
3949 | } | |
47ad8475 | 3950 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3951 | __SetPageUptodate(page); |
cb6acd01 | 3952 | new_page = true; |
ac9b9c66 | 3953 | |
f83a275d | 3954 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3955 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3956 | if (err) { |
3957 | put_page(page); | |
6bda666a CL |
3958 | if (err == -EEXIST) |
3959 | goto retry; | |
3960 | goto out; | |
3961 | } | |
23be7468 | 3962 | } else { |
6bda666a | 3963 | lock_page(page); |
0fe6e20b NH |
3964 | if (unlikely(anon_vma_prepare(vma))) { |
3965 | ret = VM_FAULT_OOM; | |
3966 | goto backout_unlocked; | |
3967 | } | |
409eb8c2 | 3968 | anon_rmap = 1; |
23be7468 | 3969 | } |
0fe6e20b | 3970 | } else { |
998b4382 NH |
3971 | /* |
3972 | * If memory error occurs between mmap() and fault, some process | |
3973 | * don't have hwpoisoned swap entry for errored virtual address. | |
3974 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3975 | */ | |
3976 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3977 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3978 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3979 | goto backout_unlocked; |
3980 | } | |
6bda666a | 3981 | } |
1e8f889b | 3982 | |
57303d80 AW |
3983 | /* |
3984 | * If we are going to COW a private mapping later, we examine the | |
3985 | * pending reservations for this page now. This will ensure that | |
3986 | * any allocations necessary to record that reservation occur outside | |
3987 | * the spinlock. | |
3988 | */ | |
5e911373 | 3989 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 3990 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
3991 | ret = VM_FAULT_OOM; |
3992 | goto backout_unlocked; | |
3993 | } | |
5e911373 | 3994 | /* Just decrements count, does not deallocate */ |
285b8dca | 3995 | vma_end_reservation(h, vma, haddr); |
5e911373 | 3996 | } |
57303d80 | 3997 | |
8bea8052 | 3998 | ptl = huge_pte_lock(h, mm, ptep); |
e7c58097 MK |
3999 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4000 | if (idx >= size) | |
4001 | goto backout; | |
4c887265 | 4002 | |
83c54070 | 4003 | ret = 0; |
7f2e9525 | 4004 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4005 | goto backout; |
4006 | ||
07443a85 JK |
4007 | if (anon_rmap) { |
4008 | ClearPagePrivate(page); | |
285b8dca | 4009 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4010 | } else |
53f9263b | 4011 | page_dup_rmap(page, true); |
1e8f889b DG |
4012 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4013 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4014 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4015 | |
5d317b2b | 4016 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4017 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4018 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4019 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4020 | } |
4021 | ||
cb900f41 | 4022 | spin_unlock(ptl); |
cb6acd01 MK |
4023 | |
4024 | /* | |
4025 | * Only make newly allocated pages active. Existing pages found | |
4026 | * in the pagecache could be !page_huge_active() if they have been | |
4027 | * isolated for migration. | |
4028 | */ | |
4029 | if (new_page) | |
4030 | set_page_huge_active(page); | |
4031 | ||
4c887265 AL |
4032 | unlock_page(page); |
4033 | out: | |
ac9b9c66 | 4034 | return ret; |
4c887265 AL |
4035 | |
4036 | backout: | |
cb900f41 | 4037 | spin_unlock(ptl); |
2b26736c | 4038 | backout_unlocked: |
4c887265 | 4039 | unlock_page(page); |
285b8dca | 4040 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4041 | put_page(page); |
4042 | goto out; | |
ac9b9c66 HD |
4043 | } |
4044 | ||
8382d914 | 4045 | #ifdef CONFIG_SMP |
1b426bac | 4046 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping, |
8382d914 DB |
4047 | pgoff_t idx, unsigned long address) |
4048 | { | |
4049 | unsigned long key[2]; | |
4050 | u32 hash; | |
4051 | ||
1b426bac MK |
4052 | key[0] = (unsigned long) mapping; |
4053 | key[1] = idx; | |
8382d914 DB |
4054 | |
4055 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
4056 | ||
4057 | return hash & (num_fault_mutexes - 1); | |
4058 | } | |
4059 | #else | |
4060 | /* | |
4061 | * For uniprocesor systems we always use a single mutex, so just | |
4062 | * return 0 and avoid the hashing overhead. | |
4063 | */ | |
1b426bac | 4064 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping, |
8382d914 DB |
4065 | pgoff_t idx, unsigned long address) |
4066 | { | |
4067 | return 0; | |
4068 | } | |
4069 | #endif | |
4070 | ||
2b740303 | 4071 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4072 | unsigned long address, unsigned int flags) |
86e5216f | 4073 | { |
8382d914 | 4074 | pte_t *ptep, entry; |
cb900f41 | 4075 | spinlock_t *ptl; |
2b740303 | 4076 | vm_fault_t ret; |
8382d914 DB |
4077 | u32 hash; |
4078 | pgoff_t idx; | |
0fe6e20b | 4079 | struct page *page = NULL; |
57303d80 | 4080 | struct page *pagecache_page = NULL; |
a5516438 | 4081 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4082 | struct address_space *mapping; |
0f792cf9 | 4083 | int need_wait_lock = 0; |
285b8dca | 4084 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4085 | |
285b8dca | 4086 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed NH |
4087 | if (ptep) { |
4088 | entry = huge_ptep_get(ptep); | |
290408d4 | 4089 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4090 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4091 | return 0; |
4092 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4093 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4094 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
ddeaab32 MK |
4095 | } else { |
4096 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4097 | if (!ptep) | |
4098 | return VM_FAULT_OOM; | |
fd6a03ed NH |
4099 | } |
4100 | ||
8382d914 | 4101 | mapping = vma->vm_file->f_mapping; |
ddeaab32 | 4102 | idx = vma_hugecache_offset(h, vma, haddr); |
8382d914 | 4103 | |
3935baa9 DG |
4104 | /* |
4105 | * Serialize hugepage allocation and instantiation, so that we don't | |
4106 | * get spurious allocation failures if two CPUs race to instantiate | |
4107 | * the same page in the page cache. | |
4108 | */ | |
1b426bac | 4109 | hash = hugetlb_fault_mutex_hash(h, mapping, idx, haddr); |
c672c7f2 | 4110 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4111 | |
7f2e9525 GS |
4112 | entry = huge_ptep_get(ptep); |
4113 | if (huge_pte_none(entry)) { | |
8382d914 | 4114 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4115 | goto out_mutex; |
3935baa9 | 4116 | } |
86e5216f | 4117 | |
83c54070 | 4118 | ret = 0; |
1e8f889b | 4119 | |
0f792cf9 NH |
4120 | /* |
4121 | * entry could be a migration/hwpoison entry at this point, so this | |
4122 | * check prevents the kernel from going below assuming that we have | |
4123 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
4124 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
4125 | * handle it. | |
4126 | */ | |
4127 | if (!pte_present(entry)) | |
4128 | goto out_mutex; | |
4129 | ||
57303d80 AW |
4130 | /* |
4131 | * If we are going to COW the mapping later, we examine the pending | |
4132 | * reservations for this page now. This will ensure that any | |
4133 | * allocations necessary to record that reservation occur outside the | |
4134 | * spinlock. For private mappings, we also lookup the pagecache | |
4135 | * page now as it is used to determine if a reservation has been | |
4136 | * consumed. | |
4137 | */ | |
106c992a | 4138 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4139 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4140 | ret = VM_FAULT_OOM; |
b4d1d99f | 4141 | goto out_mutex; |
2b26736c | 4142 | } |
5e911373 | 4143 | /* Just decrements count, does not deallocate */ |
285b8dca | 4144 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4145 | |
f83a275d | 4146 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4147 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4148 | vma, haddr); |
57303d80 AW |
4149 | } |
4150 | ||
0f792cf9 NH |
4151 | ptl = huge_pte_lock(h, mm, ptep); |
4152 | ||
4153 | /* Check for a racing update before calling hugetlb_cow */ | |
4154 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4155 | goto out_ptl; | |
4156 | ||
56c9cfb1 NH |
4157 | /* |
4158 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4159 | * pagecache_page, so here we need take the former one | |
4160 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4161 | */ |
4162 | page = pte_page(entry); | |
4163 | if (page != pagecache_page) | |
0f792cf9 NH |
4164 | if (!trylock_page(page)) { |
4165 | need_wait_lock = 1; | |
4166 | goto out_ptl; | |
4167 | } | |
b4d1d99f | 4168 | |
0f792cf9 | 4169 | get_page(page); |
b4d1d99f | 4170 | |
788c7df4 | 4171 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4172 | if (!huge_pte_write(entry)) { |
974e6d66 | 4173 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4174 | pagecache_page, ptl); |
0f792cf9 | 4175 | goto out_put_page; |
b4d1d99f | 4176 | } |
106c992a | 4177 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4178 | } |
4179 | entry = pte_mkyoung(entry); | |
285b8dca | 4180 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4181 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4182 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4183 | out_put_page: |
4184 | if (page != pagecache_page) | |
4185 | unlock_page(page); | |
4186 | put_page(page); | |
cb900f41 KS |
4187 | out_ptl: |
4188 | spin_unlock(ptl); | |
57303d80 AW |
4189 | |
4190 | if (pagecache_page) { | |
4191 | unlock_page(pagecache_page); | |
4192 | put_page(pagecache_page); | |
4193 | } | |
b4d1d99f | 4194 | out_mutex: |
c672c7f2 | 4195 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
4196 | /* |
4197 | * Generally it's safe to hold refcount during waiting page lock. But | |
4198 | * here we just wait to defer the next page fault to avoid busy loop and | |
4199 | * the page is not used after unlocked before returning from the current | |
4200 | * page fault. So we are safe from accessing freed page, even if we wait | |
4201 | * here without taking refcount. | |
4202 | */ | |
4203 | if (need_wait_lock) | |
4204 | wait_on_page_locked(page); | |
1e8f889b | 4205 | return ret; |
86e5216f AL |
4206 | } |
4207 | ||
8fb5debc MK |
4208 | /* |
4209 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4210 | * modifications for huge pages. | |
4211 | */ | |
4212 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4213 | pte_t *dst_pte, | |
4214 | struct vm_area_struct *dst_vma, | |
4215 | unsigned long dst_addr, | |
4216 | unsigned long src_addr, | |
4217 | struct page **pagep) | |
4218 | { | |
1e392147 AA |
4219 | struct address_space *mapping; |
4220 | pgoff_t idx; | |
4221 | unsigned long size; | |
1c9e8def | 4222 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4223 | struct hstate *h = hstate_vma(dst_vma); |
4224 | pte_t _dst_pte; | |
4225 | spinlock_t *ptl; | |
4226 | int ret; | |
4227 | struct page *page; | |
4228 | ||
4229 | if (!*pagep) { | |
4230 | ret = -ENOMEM; | |
4231 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4232 | if (IS_ERR(page)) | |
4233 | goto out; | |
4234 | ||
4235 | ret = copy_huge_page_from_user(page, | |
4236 | (const void __user *) src_addr, | |
810a56b9 | 4237 | pages_per_huge_page(h), false); |
8fb5debc MK |
4238 | |
4239 | /* fallback to copy_from_user outside mmap_sem */ | |
4240 | if (unlikely(ret)) { | |
9e368259 | 4241 | ret = -ENOENT; |
8fb5debc MK |
4242 | *pagep = page; |
4243 | /* don't free the page */ | |
4244 | goto out; | |
4245 | } | |
4246 | } else { | |
4247 | page = *pagep; | |
4248 | *pagep = NULL; | |
4249 | } | |
4250 | ||
4251 | /* | |
4252 | * The memory barrier inside __SetPageUptodate makes sure that | |
4253 | * preceding stores to the page contents become visible before | |
4254 | * the set_pte_at() write. | |
4255 | */ | |
4256 | __SetPageUptodate(page); | |
8fb5debc | 4257 | |
1e392147 AA |
4258 | mapping = dst_vma->vm_file->f_mapping; |
4259 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4260 | ||
1c9e8def MK |
4261 | /* |
4262 | * If shared, add to page cache | |
4263 | */ | |
4264 | if (vm_shared) { | |
1e392147 AA |
4265 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4266 | ret = -EFAULT; | |
4267 | if (idx >= size) | |
4268 | goto out_release_nounlock; | |
1c9e8def | 4269 | |
1e392147 AA |
4270 | /* |
4271 | * Serialization between remove_inode_hugepages() and | |
4272 | * huge_add_to_page_cache() below happens through the | |
4273 | * hugetlb_fault_mutex_table that here must be hold by | |
4274 | * the caller. | |
4275 | */ | |
1c9e8def MK |
4276 | ret = huge_add_to_page_cache(page, mapping, idx); |
4277 | if (ret) | |
4278 | goto out_release_nounlock; | |
4279 | } | |
4280 | ||
8fb5debc MK |
4281 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4282 | spin_lock(ptl); | |
4283 | ||
1e392147 AA |
4284 | /* |
4285 | * Recheck the i_size after holding PT lock to make sure not | |
4286 | * to leave any page mapped (as page_mapped()) beyond the end | |
4287 | * of the i_size (remove_inode_hugepages() is strict about | |
4288 | * enforcing that). If we bail out here, we'll also leave a | |
4289 | * page in the radix tree in the vm_shared case beyond the end | |
4290 | * of the i_size, but remove_inode_hugepages() will take care | |
4291 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4292 | */ | |
4293 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4294 | ret = -EFAULT; | |
4295 | if (idx >= size) | |
4296 | goto out_release_unlock; | |
4297 | ||
8fb5debc MK |
4298 | ret = -EEXIST; |
4299 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4300 | goto out_release_unlock; | |
4301 | ||
1c9e8def MK |
4302 | if (vm_shared) { |
4303 | page_dup_rmap(page, true); | |
4304 | } else { | |
4305 | ClearPagePrivate(page); | |
4306 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4307 | } | |
8fb5debc MK |
4308 | |
4309 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4310 | if (dst_vma->vm_flags & VM_WRITE) | |
4311 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4312 | _dst_pte = pte_mkyoung(_dst_pte); | |
4313 | ||
4314 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4315 | ||
4316 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4317 | dst_vma->vm_flags & VM_WRITE); | |
4318 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4319 | ||
4320 | /* No need to invalidate - it was non-present before */ | |
4321 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4322 | ||
4323 | spin_unlock(ptl); | |
cb6acd01 | 4324 | set_page_huge_active(page); |
1c9e8def MK |
4325 | if (vm_shared) |
4326 | unlock_page(page); | |
8fb5debc MK |
4327 | ret = 0; |
4328 | out: | |
4329 | return ret; | |
4330 | out_release_unlock: | |
4331 | spin_unlock(ptl); | |
1c9e8def MK |
4332 | if (vm_shared) |
4333 | unlock_page(page); | |
5af10dfd | 4334 | out_release_nounlock: |
8fb5debc MK |
4335 | put_page(page); |
4336 | goto out; | |
4337 | } | |
4338 | ||
28a35716 ML |
4339 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4340 | struct page **pages, struct vm_area_struct **vmas, | |
4341 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4342 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4343 | { |
d5d4b0aa CK |
4344 | unsigned long pfn_offset; |
4345 | unsigned long vaddr = *position; | |
28a35716 | 4346 | unsigned long remainder = *nr_pages; |
a5516438 | 4347 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4348 | int err = -EFAULT; |
63551ae0 | 4349 | |
63551ae0 | 4350 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4351 | pte_t *pte; |
cb900f41 | 4352 | spinlock_t *ptl = NULL; |
2a15efc9 | 4353 | int absent; |
4c887265 | 4354 | struct page *page; |
63551ae0 | 4355 | |
02057967 DR |
4356 | /* |
4357 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4358 | * potentially allocating memory. | |
4359 | */ | |
fa45f116 | 4360 | if (fatal_signal_pending(current)) { |
02057967 DR |
4361 | remainder = 0; |
4362 | break; | |
4363 | } | |
4364 | ||
4c887265 AL |
4365 | /* |
4366 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4367 | * each hugepage. We have to make sure we get the |
4c887265 | 4368 | * first, for the page indexing below to work. |
cb900f41 KS |
4369 | * |
4370 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4371 | */ |
7868a208 PA |
4372 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4373 | huge_page_size(h)); | |
cb900f41 KS |
4374 | if (pte) |
4375 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4376 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4377 | ||
4378 | /* | |
4379 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4380 | * an error where there's an empty slot with no huge pagecache |
4381 | * to back it. This way, we avoid allocating a hugepage, and | |
4382 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4383 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4384 | */ |
3ae77f43 HD |
4385 | if (absent && (flags & FOLL_DUMP) && |
4386 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4387 | if (pte) |
4388 | spin_unlock(ptl); | |
2a15efc9 HD |
4389 | remainder = 0; |
4390 | break; | |
4391 | } | |
63551ae0 | 4392 | |
9cc3a5bd NH |
4393 | /* |
4394 | * We need call hugetlb_fault for both hugepages under migration | |
4395 | * (in which case hugetlb_fault waits for the migration,) and | |
4396 | * hwpoisoned hugepages (in which case we need to prevent the | |
4397 | * caller from accessing to them.) In order to do this, we use | |
4398 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4399 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4400 | * both cases, and because we can't follow correct pages | |
4401 | * directly from any kind of swap entries. | |
4402 | */ | |
4403 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4404 | ((flags & FOLL_WRITE) && |
4405 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4406 | vm_fault_t ret; |
87ffc118 | 4407 | unsigned int fault_flags = 0; |
63551ae0 | 4408 | |
cb900f41 KS |
4409 | if (pte) |
4410 | spin_unlock(ptl); | |
87ffc118 AA |
4411 | if (flags & FOLL_WRITE) |
4412 | fault_flags |= FAULT_FLAG_WRITE; | |
4413 | if (nonblocking) | |
4414 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4415 | if (flags & FOLL_NOWAIT) | |
4416 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4417 | FAULT_FLAG_RETRY_NOWAIT; | |
4418 | if (flags & FOLL_TRIED) { | |
4419 | VM_WARN_ON_ONCE(fault_flags & | |
4420 | FAULT_FLAG_ALLOW_RETRY); | |
4421 | fault_flags |= FAULT_FLAG_TRIED; | |
4422 | } | |
4423 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4424 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4425 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4426 | remainder = 0; |
4427 | break; | |
4428 | } | |
4429 | if (ret & VM_FAULT_RETRY) { | |
1ac25013 AA |
4430 | if (nonblocking && |
4431 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) | |
87ffc118 AA |
4432 | *nonblocking = 0; |
4433 | *nr_pages = 0; | |
4434 | /* | |
4435 | * VM_FAULT_RETRY must not return an | |
4436 | * error, it will return zero | |
4437 | * instead. | |
4438 | * | |
4439 | * No need to update "position" as the | |
4440 | * caller will not check it after | |
4441 | * *nr_pages is set to 0. | |
4442 | */ | |
4443 | return i; | |
4444 | } | |
4445 | continue; | |
4c887265 AL |
4446 | } |
4447 | ||
a5516438 | 4448 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4449 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca LT |
4450 | |
4451 | /* | |
4452 | * Instead of doing 'try_get_page()' below in the same_page | |
4453 | * loop, just check the count once here. | |
4454 | */ | |
4455 | if (unlikely(page_count(page) <= 0)) { | |
4456 | if (pages) { | |
4457 | spin_unlock(ptl); | |
4458 | remainder = 0; | |
4459 | err = -ENOMEM; | |
4460 | break; | |
4461 | } | |
4462 | } | |
d5d4b0aa | 4463 | same_page: |
d6692183 | 4464 | if (pages) { |
2a15efc9 | 4465 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 4466 | get_page(pages[i]); |
d6692183 | 4467 | } |
63551ae0 DG |
4468 | |
4469 | if (vmas) | |
4470 | vmas[i] = vma; | |
4471 | ||
4472 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4473 | ++pfn_offset; |
63551ae0 DG |
4474 | --remainder; |
4475 | ++i; | |
d5d4b0aa | 4476 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4477 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
4478 | /* |
4479 | * We use pfn_offset to avoid touching the pageframes | |
4480 | * of this compound page. | |
4481 | */ | |
4482 | goto same_page; | |
4483 | } | |
cb900f41 | 4484 | spin_unlock(ptl); |
63551ae0 | 4485 | } |
28a35716 | 4486 | *nr_pages = remainder; |
87ffc118 AA |
4487 | /* |
4488 | * setting position is actually required only if remainder is | |
4489 | * not zero but it's faster not to add a "if (remainder)" | |
4490 | * branch. | |
4491 | */ | |
63551ae0 DG |
4492 | *position = vaddr; |
4493 | ||
2be7cfed | 4494 | return i ? i : err; |
63551ae0 | 4495 | } |
8f860591 | 4496 | |
5491ae7b AK |
4497 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4498 | /* | |
4499 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4500 | * implement this. | |
4501 | */ | |
4502 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4503 | #endif | |
4504 | ||
7da4d641 | 4505 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4506 | unsigned long address, unsigned long end, pgprot_t newprot) |
4507 | { | |
4508 | struct mm_struct *mm = vma->vm_mm; | |
4509 | unsigned long start = address; | |
4510 | pte_t *ptep; | |
4511 | pte_t pte; | |
a5516438 | 4512 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4513 | unsigned long pages = 0; |
dff11abe | 4514 | bool shared_pmd = false; |
ac46d4f3 | 4515 | struct mmu_notifier_range range; |
dff11abe MK |
4516 | |
4517 | /* | |
4518 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 4519 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
4520 | * range if PMD sharing is possible. |
4521 | */ | |
7269f999 JG |
4522 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
4523 | 0, vma, mm, start, end); | |
ac46d4f3 | 4524 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
4525 | |
4526 | BUG_ON(address >= end); | |
ac46d4f3 | 4527 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 4528 | |
ac46d4f3 | 4529 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 4530 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4531 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4532 | spinlock_t *ptl; |
7868a208 | 4533 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4534 | if (!ptep) |
4535 | continue; | |
cb900f41 | 4536 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4537 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4538 | pages++; | |
cb900f41 | 4539 | spin_unlock(ptl); |
dff11abe | 4540 | shared_pmd = true; |
39dde65c | 4541 | continue; |
7da4d641 | 4542 | } |
a8bda28d NH |
4543 | pte = huge_ptep_get(ptep); |
4544 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4545 | spin_unlock(ptl); | |
4546 | continue; | |
4547 | } | |
4548 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4549 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4550 | ||
4551 | if (is_write_migration_entry(entry)) { | |
4552 | pte_t newpte; | |
4553 | ||
4554 | make_migration_entry_read(&entry); | |
4555 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4556 | set_huge_swap_pte_at(mm, address, ptep, |
4557 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4558 | pages++; |
4559 | } | |
4560 | spin_unlock(ptl); | |
4561 | continue; | |
4562 | } | |
4563 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
4564 | pte_t old_pte; |
4565 | ||
4566 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
4567 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 4568 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 4569 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 4570 | pages++; |
8f860591 | 4571 | } |
cb900f41 | 4572 | spin_unlock(ptl); |
8f860591 | 4573 | } |
d833352a | 4574 | /* |
c8c06efa | 4575 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4576 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4577 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
4578 | * and that page table be reused and filled with junk. If we actually |
4579 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 4580 | */ |
dff11abe | 4581 | if (shared_pmd) |
ac46d4f3 | 4582 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
4583 | else |
4584 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
4585 | /* |
4586 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4587 | * page table protection not changing it to point to a new page. | |
4588 | * | |
ad56b738 | 4589 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4590 | */ |
83cde9e8 | 4591 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 4592 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
4593 | |
4594 | return pages << h->order; | |
8f860591 ZY |
4595 | } |
4596 | ||
a1e78772 MG |
4597 | int hugetlb_reserve_pages(struct inode *inode, |
4598 | long from, long to, | |
5a6fe125 | 4599 | struct vm_area_struct *vma, |
ca16d140 | 4600 | vm_flags_t vm_flags) |
e4e574b7 | 4601 | { |
17c9d12e | 4602 | long ret, chg; |
a5516438 | 4603 | struct hstate *h = hstate_inode(inode); |
90481622 | 4604 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4605 | struct resv_map *resv_map; |
1c5ecae3 | 4606 | long gbl_reserve; |
e4e574b7 | 4607 | |
63489f8e MK |
4608 | /* This should never happen */ |
4609 | if (from > to) { | |
4610 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
4611 | return -EINVAL; | |
4612 | } | |
4613 | ||
17c9d12e MG |
4614 | /* |
4615 | * Only apply hugepage reservation if asked. At fault time, an | |
4616 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4617 | * without using reserves |
17c9d12e | 4618 | */ |
ca16d140 | 4619 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4620 | return 0; |
4621 | ||
a1e78772 MG |
4622 | /* |
4623 | * Shared mappings base their reservation on the number of pages that | |
4624 | * are already allocated on behalf of the file. Private mappings need | |
4625 | * to reserve the full area even if read-only as mprotect() may be | |
4626 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4627 | */ | |
9119a41e | 4628 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
4629 | /* |
4630 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
4631 | * called for inodes for which resv_maps were created (see | |
4632 | * hugetlbfs_get_inode). | |
4633 | */ | |
4e35f483 | 4634 | resv_map = inode_resv_map(inode); |
9119a41e | 4635 | |
1406ec9b | 4636 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4637 | |
4638 | } else { | |
4639 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4640 | if (!resv_map) |
4641 | return -ENOMEM; | |
4642 | ||
a1e78772 | 4643 | chg = to - from; |
84afd99b | 4644 | |
17c9d12e MG |
4645 | set_vma_resv_map(vma, resv_map); |
4646 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4647 | } | |
4648 | ||
c50ac050 DH |
4649 | if (chg < 0) { |
4650 | ret = chg; | |
4651 | goto out_err; | |
4652 | } | |
8a630112 | 4653 | |
1c5ecae3 MK |
4654 | /* |
4655 | * There must be enough pages in the subpool for the mapping. If | |
4656 | * the subpool has a minimum size, there may be some global | |
4657 | * reservations already in place (gbl_reserve). | |
4658 | */ | |
4659 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4660 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4661 | ret = -ENOSPC; |
4662 | goto out_err; | |
4663 | } | |
5a6fe125 MG |
4664 | |
4665 | /* | |
17c9d12e | 4666 | * Check enough hugepages are available for the reservation. |
90481622 | 4667 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4668 | */ |
1c5ecae3 | 4669 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4670 | if (ret < 0) { |
1c5ecae3 MK |
4671 | /* put back original number of pages, chg */ |
4672 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4673 | goto out_err; |
68842c9b | 4674 | } |
17c9d12e MG |
4675 | |
4676 | /* | |
4677 | * Account for the reservations made. Shared mappings record regions | |
4678 | * that have reservations as they are shared by multiple VMAs. | |
4679 | * When the last VMA disappears, the region map says how much | |
4680 | * the reservation was and the page cache tells how much of | |
4681 | * the reservation was consumed. Private mappings are per-VMA and | |
4682 | * only the consumed reservations are tracked. When the VMA | |
4683 | * disappears, the original reservation is the VMA size and the | |
4684 | * consumed reservations are stored in the map. Hence, nothing | |
4685 | * else has to be done for private mappings here | |
4686 | */ | |
33039678 MK |
4687 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4688 | long add = region_add(resv_map, from, to); | |
4689 | ||
4690 | if (unlikely(chg > add)) { | |
4691 | /* | |
4692 | * pages in this range were added to the reserve | |
4693 | * map between region_chg and region_add. This | |
4694 | * indicates a race with alloc_huge_page. Adjust | |
4695 | * the subpool and reserve counts modified above | |
4696 | * based on the difference. | |
4697 | */ | |
4698 | long rsv_adjust; | |
4699 | ||
4700 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4701 | chg - add); | |
4702 | hugetlb_acct_memory(h, -rsv_adjust); | |
4703 | } | |
4704 | } | |
a43a8c39 | 4705 | return 0; |
c50ac050 | 4706 | out_err: |
5e911373 | 4707 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4708 | /* Don't call region_abort if region_chg failed */ |
4709 | if (chg >= 0) | |
4710 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4711 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4712 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4713 | return ret; |
a43a8c39 CK |
4714 | } |
4715 | ||
b5cec28d MK |
4716 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4717 | long freed) | |
a43a8c39 | 4718 | { |
a5516438 | 4719 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4720 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4721 | long chg = 0; |
90481622 | 4722 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4723 | long gbl_reserve; |
45c682a6 | 4724 | |
f27a5136 MK |
4725 | /* |
4726 | * Since this routine can be called in the evict inode path for all | |
4727 | * hugetlbfs inodes, resv_map could be NULL. | |
4728 | */ | |
b5cec28d MK |
4729 | if (resv_map) { |
4730 | chg = region_del(resv_map, start, end); | |
4731 | /* | |
4732 | * region_del() can fail in the rare case where a region | |
4733 | * must be split and another region descriptor can not be | |
4734 | * allocated. If end == LONG_MAX, it will not fail. | |
4735 | */ | |
4736 | if (chg < 0) | |
4737 | return chg; | |
4738 | } | |
4739 | ||
45c682a6 | 4740 | spin_lock(&inode->i_lock); |
e4c6f8be | 4741 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4742 | spin_unlock(&inode->i_lock); |
4743 | ||
1c5ecae3 MK |
4744 | /* |
4745 | * If the subpool has a minimum size, the number of global | |
4746 | * reservations to be released may be adjusted. | |
4747 | */ | |
4748 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4749 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4750 | |
4751 | return 0; | |
a43a8c39 | 4752 | } |
93f70f90 | 4753 | |
3212b535 SC |
4754 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4755 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4756 | struct vm_area_struct *vma, | |
4757 | unsigned long addr, pgoff_t idx) | |
4758 | { | |
4759 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4760 | svma->vm_start; | |
4761 | unsigned long sbase = saddr & PUD_MASK; | |
4762 | unsigned long s_end = sbase + PUD_SIZE; | |
4763 | ||
4764 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4765 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4766 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4767 | |
4768 | /* | |
4769 | * match the virtual addresses, permission and the alignment of the | |
4770 | * page table page. | |
4771 | */ | |
4772 | if (pmd_index(addr) != pmd_index(saddr) || | |
4773 | vm_flags != svm_flags || | |
4774 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4775 | return 0; | |
4776 | ||
4777 | return saddr; | |
4778 | } | |
4779 | ||
31aafb45 | 4780 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4781 | { |
4782 | unsigned long base = addr & PUD_MASK; | |
4783 | unsigned long end = base + PUD_SIZE; | |
4784 | ||
4785 | /* | |
4786 | * check on proper vm_flags and page table alignment | |
4787 | */ | |
017b1660 | 4788 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
4789 | return true; |
4790 | return false; | |
3212b535 SC |
4791 | } |
4792 | ||
017b1660 MK |
4793 | /* |
4794 | * Determine if start,end range within vma could be mapped by shared pmd. | |
4795 | * If yes, adjust start and end to cover range associated with possible | |
4796 | * shared pmd mappings. | |
4797 | */ | |
4798 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
4799 | unsigned long *start, unsigned long *end) | |
4800 | { | |
4801 | unsigned long check_addr = *start; | |
4802 | ||
4803 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
4804 | return; | |
4805 | ||
4806 | for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { | |
4807 | unsigned long a_start = check_addr & PUD_MASK; | |
4808 | unsigned long a_end = a_start + PUD_SIZE; | |
4809 | ||
4810 | /* | |
4811 | * If sharing is possible, adjust start/end if necessary. | |
4812 | */ | |
4813 | if (range_in_vma(vma, a_start, a_end)) { | |
4814 | if (a_start < *start) | |
4815 | *start = a_start; | |
4816 | if (a_end > *end) | |
4817 | *end = a_end; | |
4818 | } | |
4819 | } | |
4820 | } | |
4821 | ||
3212b535 SC |
4822 | /* |
4823 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4824 | * and returns the corresponding pte. While this is not necessary for the | |
4825 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
ddeaab32 MK |
4826 | * code much cleaner. pmd allocation is essential for the shared case because |
4827 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise | |
4828 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
4829 | * bad pmd for sharing. | |
3212b535 SC |
4830 | */ |
4831 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4832 | { | |
4833 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4834 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4835 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4836 | vma->vm_pgoff; | |
4837 | struct vm_area_struct *svma; | |
4838 | unsigned long saddr; | |
4839 | pte_t *spte = NULL; | |
4840 | pte_t *pte; | |
cb900f41 | 4841 | spinlock_t *ptl; |
3212b535 SC |
4842 | |
4843 | if (!vma_shareable(vma, addr)) | |
4844 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4845 | ||
ddeaab32 | 4846 | i_mmap_lock_write(mapping); |
3212b535 SC |
4847 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4848 | if (svma == vma) | |
4849 | continue; | |
4850 | ||
4851 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4852 | if (saddr) { | |
7868a208 PA |
4853 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4854 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4855 | if (spte) { |
4856 | get_page(virt_to_page(spte)); | |
4857 | break; | |
4858 | } | |
4859 | } | |
4860 | } | |
4861 | ||
4862 | if (!spte) | |
4863 | goto out; | |
4864 | ||
8bea8052 | 4865 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4866 | if (pud_none(*pud)) { |
3212b535 SC |
4867 | pud_populate(mm, pud, |
4868 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4869 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4870 | } else { |
3212b535 | 4871 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4872 | } |
cb900f41 | 4873 | spin_unlock(ptl); |
3212b535 SC |
4874 | out: |
4875 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
ddeaab32 | 4876 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4877 | return pte; |
4878 | } | |
4879 | ||
4880 | /* | |
4881 | * unmap huge page backed by shared pte. | |
4882 | * | |
4883 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4884 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4885 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4886 | * | |
ddeaab32 | 4887 | * called with page table lock held. |
3212b535 SC |
4888 | * |
4889 | * returns: 1 successfully unmapped a shared pte page | |
4890 | * 0 the underlying pte page is not shared, or it is the last user | |
4891 | */ | |
4892 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4893 | { | |
4894 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4895 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4896 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4897 | |
4898 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4899 | if (page_count(virt_to_page(ptep)) == 1) | |
4900 | return 0; | |
4901 | ||
4902 | pud_clear(pud); | |
4903 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4904 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4905 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4906 | return 1; | |
4907 | } | |
9e5fc74c SC |
4908 | #define want_pmd_share() (1) |
4909 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4910 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4911 | { | |
4912 | return NULL; | |
4913 | } | |
e81f2d22 ZZ |
4914 | |
4915 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4916 | { | |
4917 | return 0; | |
4918 | } | |
017b1660 MK |
4919 | |
4920 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
4921 | unsigned long *start, unsigned long *end) | |
4922 | { | |
4923 | } | |
9e5fc74c | 4924 | #define want_pmd_share() (0) |
3212b535 SC |
4925 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4926 | ||
9e5fc74c SC |
4927 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4928 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4929 | unsigned long addr, unsigned long sz) | |
4930 | { | |
4931 | pgd_t *pgd; | |
c2febafc | 4932 | p4d_t *p4d; |
9e5fc74c SC |
4933 | pud_t *pud; |
4934 | pte_t *pte = NULL; | |
4935 | ||
4936 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
4937 | p4d = p4d_alloc(mm, pgd, addr); |
4938 | if (!p4d) | |
4939 | return NULL; | |
c2febafc | 4940 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
4941 | if (pud) { |
4942 | if (sz == PUD_SIZE) { | |
4943 | pte = (pte_t *)pud; | |
4944 | } else { | |
4945 | BUG_ON(sz != PMD_SIZE); | |
4946 | if (want_pmd_share() && pud_none(*pud)) | |
4947 | pte = huge_pmd_share(mm, addr, pud); | |
4948 | else | |
4949 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4950 | } | |
4951 | } | |
4e666314 | 4952 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
4953 | |
4954 | return pte; | |
4955 | } | |
4956 | ||
9b19df29 PA |
4957 | /* |
4958 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
4959 | * entry at address @addr | |
4960 | * | |
4961 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
4962 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
4963 | * size @sz doesn't match the hugepage size at this level of the page | |
4964 | * table. | |
4965 | */ | |
7868a208 PA |
4966 | pte_t *huge_pte_offset(struct mm_struct *mm, |
4967 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
4968 | { |
4969 | pgd_t *pgd; | |
c2febafc | 4970 | p4d_t *p4d; |
9e5fc74c | 4971 | pud_t *pud; |
c2febafc | 4972 | pmd_t *pmd; |
9e5fc74c SC |
4973 | |
4974 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4975 | if (!pgd_present(*pgd)) |
4976 | return NULL; | |
4977 | p4d = p4d_offset(pgd, addr); | |
4978 | if (!p4d_present(*p4d)) | |
4979 | return NULL; | |
9b19df29 | 4980 | |
c2febafc | 4981 | pud = pud_offset(p4d, addr); |
9b19df29 | 4982 | if (sz != PUD_SIZE && pud_none(*pud)) |
c2febafc | 4983 | return NULL; |
9b19df29 PA |
4984 | /* hugepage or swap? */ |
4985 | if (pud_huge(*pud) || !pud_present(*pud)) | |
c2febafc | 4986 | return (pte_t *)pud; |
9b19df29 | 4987 | |
c2febafc | 4988 | pmd = pmd_offset(pud, addr); |
9b19df29 PA |
4989 | if (sz != PMD_SIZE && pmd_none(*pmd)) |
4990 | return NULL; | |
4991 | /* hugepage or swap? */ | |
4992 | if (pmd_huge(*pmd) || !pmd_present(*pmd)) | |
4993 | return (pte_t *)pmd; | |
4994 | ||
4995 | return NULL; | |
9e5fc74c SC |
4996 | } |
4997 | ||
61f77eda NH |
4998 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4999 | ||
5000 | /* | |
5001 | * These functions are overwritable if your architecture needs its own | |
5002 | * behavior. | |
5003 | */ | |
5004 | struct page * __weak | |
5005 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5006 | int write) | |
5007 | { | |
5008 | return ERR_PTR(-EINVAL); | |
5009 | } | |
5010 | ||
4dc71451 AK |
5011 | struct page * __weak |
5012 | follow_huge_pd(struct vm_area_struct *vma, | |
5013 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5014 | { | |
5015 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5016 | return NULL; | |
5017 | } | |
5018 | ||
61f77eda | 5019 | struct page * __weak |
9e5fc74c | 5020 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5021 | pmd_t *pmd, int flags) |
9e5fc74c | 5022 | { |
e66f17ff NH |
5023 | struct page *page = NULL; |
5024 | spinlock_t *ptl; | |
c9d398fa | 5025 | pte_t pte; |
e66f17ff NH |
5026 | retry: |
5027 | ptl = pmd_lockptr(mm, pmd); | |
5028 | spin_lock(ptl); | |
5029 | /* | |
5030 | * make sure that the address range covered by this pmd is not | |
5031 | * unmapped from other threads. | |
5032 | */ | |
5033 | if (!pmd_huge(*pmd)) | |
5034 | goto out; | |
c9d398fa NH |
5035 | pte = huge_ptep_get((pte_t *)pmd); |
5036 | if (pte_present(pte)) { | |
97534127 | 5037 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
5038 | if (flags & FOLL_GET) |
5039 | get_page(page); | |
5040 | } else { | |
c9d398fa | 5041 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5042 | spin_unlock(ptl); |
5043 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5044 | goto retry; | |
5045 | } | |
5046 | /* | |
5047 | * hwpoisoned entry is treated as no_page_table in | |
5048 | * follow_page_mask(). | |
5049 | */ | |
5050 | } | |
5051 | out: | |
5052 | spin_unlock(ptl); | |
9e5fc74c SC |
5053 | return page; |
5054 | } | |
5055 | ||
61f77eda | 5056 | struct page * __weak |
9e5fc74c | 5057 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5058 | pud_t *pud, int flags) |
9e5fc74c | 5059 | { |
e66f17ff NH |
5060 | if (flags & FOLL_GET) |
5061 | return NULL; | |
9e5fc74c | 5062 | |
e66f17ff | 5063 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5064 | } |
5065 | ||
faaa5b62 AK |
5066 | struct page * __weak |
5067 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5068 | { | |
5069 | if (flags & FOLL_GET) | |
5070 | return NULL; | |
5071 | ||
5072 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5073 | } | |
5074 | ||
31caf665 NH |
5075 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5076 | { | |
bcc54222 NH |
5077 | bool ret = true; |
5078 | ||
309381fe | 5079 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5080 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
5081 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
5082 | ret = false; | |
5083 | goto unlock; | |
5084 | } | |
5085 | clear_page_huge_active(page); | |
31caf665 | 5086 | list_move_tail(&page->lru, list); |
bcc54222 | 5087 | unlock: |
31caf665 | 5088 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5089 | return ret; |
31caf665 NH |
5090 | } |
5091 | ||
5092 | void putback_active_hugepage(struct page *page) | |
5093 | { | |
309381fe | 5094 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5095 | spin_lock(&hugetlb_lock); |
bcc54222 | 5096 | set_page_huge_active(page); |
31caf665 NH |
5097 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5098 | spin_unlock(&hugetlb_lock); | |
5099 | put_page(page); | |
5100 | } | |
ab5ac90a MH |
5101 | |
5102 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5103 | { | |
5104 | struct hstate *h = page_hstate(oldpage); | |
5105 | ||
5106 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5107 | set_page_owner_migrate_reason(newpage, reason); | |
5108 | ||
5109 | /* | |
5110 | * transfer temporary state of the new huge page. This is | |
5111 | * reverse to other transitions because the newpage is going to | |
5112 | * be final while the old one will be freed so it takes over | |
5113 | * the temporary status. | |
5114 | * | |
5115 | * Also note that we have to transfer the per-node surplus state | |
5116 | * here as well otherwise the global surplus count will not match | |
5117 | * the per-node's. | |
5118 | */ | |
5119 | if (PageHugeTemporary(newpage)) { | |
5120 | int old_nid = page_to_nid(oldpage); | |
5121 | int new_nid = page_to_nid(newpage); | |
5122 | ||
5123 | SetPageHugeTemporary(oldpage); | |
5124 | ClearPageHugeTemporary(newpage); | |
5125 | ||
5126 | spin_lock(&hugetlb_lock); | |
5127 | if (h->surplus_huge_pages_node[old_nid]) { | |
5128 | h->surplus_huge_pages_node[old_nid]--; | |
5129 | h->surplus_huge_pages_node[new_nid]++; | |
5130 | } | |
5131 | spin_unlock(&hugetlb_lock); | |
5132 | } | |
5133 | } |