]>
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> |
ce44b015 | 30 | #include <linux/llist.h> |
d6606683 | 31 | |
63551ae0 DG |
32 | #include <asm/page.h> |
33 | #include <asm/pgtable.h> | |
24669e58 | 34 | #include <asm/tlb.h> |
63551ae0 | 35 | |
24669e58 | 36 | #include <linux/io.h> |
63551ae0 | 37 | #include <linux/hugetlb.h> |
9dd540e2 | 38 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 39 | #include <linux/node.h> |
1a1aad8a | 40 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 41 | #include <linux/page_owner.h> |
7835e98b | 42 | #include "internal.h" |
1da177e4 | 43 | |
c3f38a38 | 44 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
45 | unsigned int default_hstate_idx; |
46 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
47 | /* |
48 | * Minimum page order among possible hugepage sizes, set to a proper value | |
49 | * at boot time. | |
50 | */ | |
51 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 52 | |
53ba51d2 JT |
53 | __initdata LIST_HEAD(huge_boot_pages); |
54 | ||
e5ff2159 AK |
55 | /* for command line parsing */ |
56 | static struct hstate * __initdata parsed_hstate; | |
57 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 58 | static unsigned long __initdata default_hstate_size; |
9fee021d | 59 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 60 | |
3935baa9 | 61 | /* |
31caf665 NH |
62 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
63 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 64 | */ |
c3f38a38 | 65 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 66 | |
8382d914 DB |
67 | /* |
68 | * Serializes faults on the same logical page. This is used to | |
69 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
70 | */ | |
71 | static int num_fault_mutexes; | |
c672c7f2 | 72 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 73 | |
c2f2139b MS |
74 | static inline bool PageHugeFreed(struct page *head) |
75 | { | |
76 | return page_private(head + 4) == -1UL; | |
77 | } | |
78 | ||
79 | static inline void SetPageHugeFreed(struct page *head) | |
80 | { | |
81 | set_page_private(head + 4, -1UL); | |
82 | } | |
83 | ||
84 | static inline void ClearPageHugeFreed(struct page *head) | |
85 | { | |
86 | set_page_private(head + 4, 0); | |
87 | } | |
88 | ||
7ca02d0a MK |
89 | /* Forward declaration */ |
90 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
91 | ||
90481622 DG |
92 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
93 | { | |
94 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
95 | ||
96 | spin_unlock(&spool->lock); | |
97 | ||
98 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
99 | * remain, give up any reservations mased on minimum size and |
100 | * free the subpool */ | |
101 | if (free) { | |
102 | if (spool->min_hpages != -1) | |
103 | hugetlb_acct_memory(spool->hstate, | |
104 | -spool->min_hpages); | |
90481622 | 105 | kfree(spool); |
7ca02d0a | 106 | } |
90481622 DG |
107 | } |
108 | ||
7ca02d0a MK |
109 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
110 | long min_hpages) | |
90481622 DG |
111 | { |
112 | struct hugepage_subpool *spool; | |
113 | ||
c6a91820 | 114 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
115 | if (!spool) |
116 | return NULL; | |
117 | ||
118 | spin_lock_init(&spool->lock); | |
119 | spool->count = 1; | |
7ca02d0a MK |
120 | spool->max_hpages = max_hpages; |
121 | spool->hstate = h; | |
122 | spool->min_hpages = min_hpages; | |
123 | ||
124 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
125 | kfree(spool); | |
126 | return NULL; | |
127 | } | |
128 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
129 | |
130 | return spool; | |
131 | } | |
132 | ||
133 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
134 | { | |
135 | spin_lock(&spool->lock); | |
136 | BUG_ON(!spool->count); | |
137 | spool->count--; | |
138 | unlock_or_release_subpool(spool); | |
139 | } | |
140 | ||
1c5ecae3 MK |
141 | /* |
142 | * Subpool accounting for allocating and reserving pages. | |
143 | * Return -ENOMEM if there are not enough resources to satisfy the | |
144 | * the request. Otherwise, return the number of pages by which the | |
145 | * global pools must be adjusted (upward). The returned value may | |
146 | * only be different than the passed value (delta) in the case where | |
147 | * a subpool minimum size must be manitained. | |
148 | */ | |
149 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
150 | long delta) |
151 | { | |
1c5ecae3 | 152 | long ret = delta; |
90481622 DG |
153 | |
154 | if (!spool) | |
1c5ecae3 | 155 | return ret; |
90481622 DG |
156 | |
157 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
158 | |
159 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
160 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
161 | spool->used_hpages += delta; | |
162 | else { | |
163 | ret = -ENOMEM; | |
164 | goto unlock_ret; | |
165 | } | |
90481622 | 166 | } |
90481622 | 167 | |
09a95e29 MK |
168 | /* minimum size accounting */ |
169 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
170 | if (delta > spool->rsv_hpages) { |
171 | /* | |
172 | * Asking for more reserves than those already taken on | |
173 | * behalf of subpool. Return difference. | |
174 | */ | |
175 | ret = delta - spool->rsv_hpages; | |
176 | spool->rsv_hpages = 0; | |
177 | } else { | |
178 | ret = 0; /* reserves already accounted for */ | |
179 | spool->rsv_hpages -= delta; | |
180 | } | |
181 | } | |
182 | ||
183 | unlock_ret: | |
184 | spin_unlock(&spool->lock); | |
90481622 DG |
185 | return ret; |
186 | } | |
187 | ||
1c5ecae3 MK |
188 | /* |
189 | * Subpool accounting for freeing and unreserving pages. | |
190 | * Return the number of global page reservations that must be dropped. | |
191 | * The return value may only be different than the passed value (delta) | |
192 | * in the case where a subpool minimum size must be maintained. | |
193 | */ | |
194 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
195 | long delta) |
196 | { | |
1c5ecae3 MK |
197 | long ret = delta; |
198 | ||
90481622 | 199 | if (!spool) |
1c5ecae3 | 200 | return delta; |
90481622 DG |
201 | |
202 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
203 | |
204 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
205 | spool->used_hpages -= delta; | |
206 | ||
09a95e29 MK |
207 | /* minimum size accounting */ |
208 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
209 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
210 | ret = 0; | |
211 | else | |
212 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
213 | ||
214 | spool->rsv_hpages += delta; | |
215 | if (spool->rsv_hpages > spool->min_hpages) | |
216 | spool->rsv_hpages = spool->min_hpages; | |
217 | } | |
218 | ||
219 | /* | |
220 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
221 | * quota reference, free it now. | |
222 | */ | |
90481622 | 223 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
224 | |
225 | return ret; | |
90481622 DG |
226 | } |
227 | ||
228 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
229 | { | |
230 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
231 | } | |
232 | ||
233 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
234 | { | |
496ad9aa | 235 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
236 | } |
237 | ||
96822904 AW |
238 | /* |
239 | * Region tracking -- allows tracking of reservations and instantiated pages | |
240 | * across the pages in a mapping. | |
84afd99b | 241 | * |
1dd308a7 MK |
242 | * The region data structures are embedded into a resv_map and protected |
243 | * by a resv_map's lock. The set of regions within the resv_map represent | |
244 | * reservations for huge pages, or huge pages that have already been | |
245 | * instantiated within the map. The from and to elements are huge page | |
246 | * indicies into the associated mapping. from indicates the starting index | |
247 | * of the region. to represents the first index past the end of the region. | |
248 | * | |
249 | * For example, a file region structure with from == 0 and to == 4 represents | |
250 | * four huge pages in a mapping. It is important to note that the to element | |
251 | * represents the first element past the end of the region. This is used in | |
252 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
253 | * | |
254 | * Interval notation of the form [from, to) will be used to indicate that | |
255 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
256 | */ |
257 | struct file_region { | |
258 | struct list_head link; | |
259 | long from; | |
260 | long to; | |
261 | }; | |
262 | ||
1dd308a7 MK |
263 | /* |
264 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
265 | * map. In the normal case, existing regions will be expanded |
266 | * to accommodate the specified range. Sufficient regions should | |
267 | * exist for expansion due to the previous call to region_chg | |
268 | * with the same range. However, it is possible that region_del | |
269 | * could have been called after region_chg and modifed the map | |
270 | * in such a way that no region exists to be expanded. In this | |
271 | * case, pull a region descriptor from the cache associated with | |
272 | * the map and use that for the new range. | |
cf3ad20b MK |
273 | * |
274 | * Return the number of new huge pages added to the map. This | |
275 | * number is greater than or equal to zero. | |
1dd308a7 | 276 | */ |
1406ec9b | 277 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 278 | { |
1406ec9b | 279 | struct list_head *head = &resv->regions; |
96822904 | 280 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 281 | long add = 0; |
96822904 | 282 | |
7b24d861 | 283 | spin_lock(&resv->lock); |
96822904 AW |
284 | /* Locate the region we are either in or before. */ |
285 | list_for_each_entry(rg, head, link) | |
286 | if (f <= rg->to) | |
287 | break; | |
288 | ||
5e911373 MK |
289 | /* |
290 | * If no region exists which can be expanded to include the | |
291 | * specified range, the list must have been modified by an | |
292 | * interleving call to region_del(). Pull a region descriptor | |
293 | * from the cache and use it for this range. | |
294 | */ | |
295 | if (&rg->link == head || t < rg->from) { | |
296 | VM_BUG_ON(resv->region_cache_count <= 0); | |
297 | ||
298 | resv->region_cache_count--; | |
299 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
300 | link); | |
301 | list_del(&nrg->link); | |
302 | ||
303 | nrg->from = f; | |
304 | nrg->to = t; | |
305 | list_add(&nrg->link, rg->link.prev); | |
306 | ||
307 | add += t - f; | |
308 | goto out_locked; | |
309 | } | |
310 | ||
96822904 AW |
311 | /* Round our left edge to the current segment if it encloses us. */ |
312 | if (f > rg->from) | |
313 | f = rg->from; | |
314 | ||
315 | /* Check for and consume any regions we now overlap with. */ | |
316 | nrg = rg; | |
317 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
318 | if (&rg->link == head) | |
319 | break; | |
320 | if (rg->from > t) | |
321 | break; | |
322 | ||
323 | /* If this area reaches higher then extend our area to | |
324 | * include it completely. If this is not the first area | |
325 | * which we intend to reuse, free it. */ | |
326 | if (rg->to > t) | |
327 | t = rg->to; | |
328 | if (rg != nrg) { | |
cf3ad20b MK |
329 | /* Decrement return value by the deleted range. |
330 | * Another range will span this area so that by | |
331 | * end of routine add will be >= zero | |
332 | */ | |
333 | add -= (rg->to - rg->from); | |
96822904 AW |
334 | list_del(&rg->link); |
335 | kfree(rg); | |
336 | } | |
337 | } | |
cf3ad20b MK |
338 | |
339 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 340 | nrg->from = f; |
cf3ad20b | 341 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 342 | nrg->to = t; |
cf3ad20b | 343 | |
5e911373 MK |
344 | out_locked: |
345 | resv->adds_in_progress--; | |
7b24d861 | 346 | spin_unlock(&resv->lock); |
cf3ad20b MK |
347 | VM_BUG_ON(add < 0); |
348 | return add; | |
96822904 AW |
349 | } |
350 | ||
1dd308a7 MK |
351 | /* |
352 | * Examine the existing reserve map and determine how many | |
353 | * huge pages in the specified range [f, t) are NOT currently | |
354 | * represented. This routine is called before a subsequent | |
355 | * call to region_add that will actually modify the reserve | |
356 | * map to add the specified range [f, t). region_chg does | |
357 | * not change the number of huge pages represented by the | |
358 | * map. However, if the existing regions in the map can not | |
359 | * be expanded to represent the new range, a new file_region | |
360 | * structure is added to the map as a placeholder. This is | |
361 | * so that the subsequent region_add call will have all the | |
362 | * regions it needs and will not fail. | |
363 | * | |
5e911373 MK |
364 | * Upon entry, region_chg will also examine the cache of region descriptors |
365 | * associated with the map. If there are not enough descriptors cached, one | |
366 | * will be allocated for the in progress add operation. | |
367 | * | |
368 | * Returns the number of huge pages that need to be added to the existing | |
369 | * reservation map for the range [f, t). This number is greater or equal to | |
370 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
371 | * is needed and can not be allocated. | |
1dd308a7 | 372 | */ |
1406ec9b | 373 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 374 | { |
1406ec9b | 375 | struct list_head *head = &resv->regions; |
7b24d861 | 376 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
377 | long chg = 0; |
378 | ||
7b24d861 DB |
379 | retry: |
380 | spin_lock(&resv->lock); | |
5e911373 MK |
381 | retry_locked: |
382 | resv->adds_in_progress++; | |
383 | ||
384 | /* | |
385 | * Check for sufficient descriptors in the cache to accommodate | |
386 | * the number of in progress add operations. | |
387 | */ | |
388 | if (resv->adds_in_progress > resv->region_cache_count) { | |
389 | struct file_region *trg; | |
390 | ||
391 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
392 | /* Must drop lock to allocate a new descriptor. */ | |
393 | resv->adds_in_progress--; | |
394 | spin_unlock(&resv->lock); | |
395 | ||
396 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
dbe409e4 MK |
397 | if (!trg) { |
398 | kfree(nrg); | |
5e911373 | 399 | return -ENOMEM; |
dbe409e4 | 400 | } |
5e911373 MK |
401 | |
402 | spin_lock(&resv->lock); | |
403 | list_add(&trg->link, &resv->region_cache); | |
404 | resv->region_cache_count++; | |
405 | goto retry_locked; | |
406 | } | |
407 | ||
96822904 AW |
408 | /* Locate the region we are before or in. */ |
409 | list_for_each_entry(rg, head, link) | |
410 | if (f <= rg->to) | |
411 | break; | |
412 | ||
413 | /* If we are below the current region then a new region is required. | |
414 | * Subtle, allocate a new region at the position but make it zero | |
415 | * size such that we can guarantee to record the reservation. */ | |
416 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 417 | if (!nrg) { |
5e911373 | 418 | resv->adds_in_progress--; |
7b24d861 DB |
419 | spin_unlock(&resv->lock); |
420 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
421 | if (!nrg) | |
422 | return -ENOMEM; | |
423 | ||
424 | nrg->from = f; | |
425 | nrg->to = f; | |
426 | INIT_LIST_HEAD(&nrg->link); | |
427 | goto retry; | |
428 | } | |
96822904 | 429 | |
7b24d861 DB |
430 | list_add(&nrg->link, rg->link.prev); |
431 | chg = t - f; | |
432 | goto out_nrg; | |
96822904 AW |
433 | } |
434 | ||
435 | /* Round our left edge to the current segment if it encloses us. */ | |
436 | if (f > rg->from) | |
437 | f = rg->from; | |
438 | chg = t - f; | |
439 | ||
440 | /* Check for and consume any regions we now overlap with. */ | |
441 | list_for_each_entry(rg, rg->link.prev, link) { | |
442 | if (&rg->link == head) | |
443 | break; | |
444 | if (rg->from > t) | |
7b24d861 | 445 | goto out; |
96822904 | 446 | |
25985edc | 447 | /* We overlap with this area, if it extends further than |
96822904 AW |
448 | * us then we must extend ourselves. Account for its |
449 | * existing reservation. */ | |
450 | if (rg->to > t) { | |
451 | chg += rg->to - t; | |
452 | t = rg->to; | |
453 | } | |
454 | chg -= rg->to - rg->from; | |
455 | } | |
7b24d861 DB |
456 | |
457 | out: | |
458 | spin_unlock(&resv->lock); | |
459 | /* We already know we raced and no longer need the new region */ | |
460 | kfree(nrg); | |
461 | return chg; | |
462 | out_nrg: | |
463 | spin_unlock(&resv->lock); | |
96822904 AW |
464 | return chg; |
465 | } | |
466 | ||
5e911373 MK |
467 | /* |
468 | * Abort the in progress add operation. The adds_in_progress field | |
469 | * of the resv_map keeps track of the operations in progress between | |
470 | * calls to region_chg and region_add. Operations are sometimes | |
471 | * aborted after the call to region_chg. In such cases, region_abort | |
472 | * is called to decrement the adds_in_progress counter. | |
473 | * | |
474 | * NOTE: The range arguments [f, t) are not needed or used in this | |
475 | * routine. They are kept to make reading the calling code easier as | |
476 | * arguments will match the associated region_chg call. | |
477 | */ | |
478 | static void region_abort(struct resv_map *resv, long f, long t) | |
479 | { | |
480 | spin_lock(&resv->lock); | |
481 | VM_BUG_ON(!resv->region_cache_count); | |
482 | resv->adds_in_progress--; | |
483 | spin_unlock(&resv->lock); | |
484 | } | |
485 | ||
1dd308a7 | 486 | /* |
feba16e2 MK |
487 | * Delete the specified range [f, t) from the reserve map. If the |
488 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
489 | * should be deleted. Locate the regions which intersect [f, t) | |
490 | * and either trim, delete or split the existing regions. | |
491 | * | |
492 | * Returns the number of huge pages deleted from the reserve map. | |
493 | * In the normal case, the return value is zero or more. In the | |
494 | * case where a region must be split, a new region descriptor must | |
495 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
496 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
497 | * a region and possibly return -ENOMEM. Callers specifying | |
498 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 499 | */ |
feba16e2 | 500 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 501 | { |
1406ec9b | 502 | struct list_head *head = &resv->regions; |
96822904 | 503 | struct file_region *rg, *trg; |
feba16e2 MK |
504 | struct file_region *nrg = NULL; |
505 | long del = 0; | |
96822904 | 506 | |
feba16e2 | 507 | retry: |
7b24d861 | 508 | spin_lock(&resv->lock); |
feba16e2 | 509 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
510 | /* |
511 | * Skip regions before the range to be deleted. file_region | |
512 | * ranges are normally of the form [from, to). However, there | |
513 | * may be a "placeholder" entry in the map which is of the form | |
514 | * (from, to) with from == to. Check for placeholder entries | |
515 | * at the beginning of the range to be deleted. | |
516 | */ | |
517 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 518 | continue; |
dbe409e4 | 519 | |
feba16e2 | 520 | if (rg->from >= t) |
96822904 | 521 | break; |
96822904 | 522 | |
feba16e2 MK |
523 | if (f > rg->from && t < rg->to) { /* Must split region */ |
524 | /* | |
525 | * Check for an entry in the cache before dropping | |
526 | * lock and attempting allocation. | |
527 | */ | |
528 | if (!nrg && | |
529 | resv->region_cache_count > resv->adds_in_progress) { | |
530 | nrg = list_first_entry(&resv->region_cache, | |
531 | struct file_region, | |
532 | link); | |
533 | list_del(&nrg->link); | |
534 | resv->region_cache_count--; | |
535 | } | |
96822904 | 536 | |
feba16e2 MK |
537 | if (!nrg) { |
538 | spin_unlock(&resv->lock); | |
539 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
540 | if (!nrg) | |
541 | return -ENOMEM; | |
542 | goto retry; | |
543 | } | |
544 | ||
545 | del += t - f; | |
546 | ||
547 | /* New entry for end of split region */ | |
548 | nrg->from = t; | |
549 | nrg->to = rg->to; | |
550 | INIT_LIST_HEAD(&nrg->link); | |
551 | ||
552 | /* Original entry is trimmed */ | |
553 | rg->to = f; | |
554 | ||
555 | list_add(&nrg->link, &rg->link); | |
556 | nrg = NULL; | |
96822904 | 557 | break; |
feba16e2 MK |
558 | } |
559 | ||
560 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
561 | del += rg->to - rg->from; | |
562 | list_del(&rg->link); | |
563 | kfree(rg); | |
564 | continue; | |
565 | } | |
566 | ||
567 | if (f <= rg->from) { /* Trim beginning of region */ | |
568 | del += t - rg->from; | |
569 | rg->from = t; | |
570 | } else { /* Trim end of region */ | |
571 | del += rg->to - f; | |
572 | rg->to = f; | |
573 | } | |
96822904 | 574 | } |
7b24d861 | 575 | |
7b24d861 | 576 | spin_unlock(&resv->lock); |
feba16e2 MK |
577 | kfree(nrg); |
578 | return del; | |
96822904 AW |
579 | } |
580 | ||
b5cec28d MK |
581 | /* |
582 | * A rare out of memory error was encountered which prevented removal of | |
583 | * the reserve map region for a page. The huge page itself was free'ed | |
584 | * and removed from the page cache. This routine will adjust the subpool | |
585 | * usage count, and the global reserve count if needed. By incrementing | |
586 | * these counts, the reserve map entry which could not be deleted will | |
587 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
588 | * counts. | |
589 | */ | |
72e2936c | 590 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
591 | { |
592 | struct hugepage_subpool *spool = subpool_inode(inode); | |
593 | long rsv_adjust; | |
a4300479 | 594 | bool reserved = false; |
b5cec28d MK |
595 | |
596 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
a4300479 | 597 | if (rsv_adjust > 0) { |
b5cec28d MK |
598 | struct hstate *h = hstate_inode(inode); |
599 | ||
a4300479 ML |
600 | if (!hugetlb_acct_memory(h, 1)) |
601 | reserved = true; | |
602 | } else if (!rsv_adjust) { | |
603 | reserved = true; | |
b5cec28d | 604 | } |
a4300479 ML |
605 | |
606 | if (!reserved) | |
607 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
608 | } |
609 | ||
1dd308a7 MK |
610 | /* |
611 | * Count and return the number of huge pages in the reserve map | |
612 | * that intersect with the range [f, t). | |
613 | */ | |
1406ec9b | 614 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 615 | { |
1406ec9b | 616 | struct list_head *head = &resv->regions; |
84afd99b AW |
617 | struct file_region *rg; |
618 | long chg = 0; | |
619 | ||
7b24d861 | 620 | spin_lock(&resv->lock); |
84afd99b AW |
621 | /* Locate each segment we overlap with, and count that overlap. */ |
622 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
623 | long seg_from; |
624 | long seg_to; | |
84afd99b AW |
625 | |
626 | if (rg->to <= f) | |
627 | continue; | |
628 | if (rg->from >= t) | |
629 | break; | |
630 | ||
631 | seg_from = max(rg->from, f); | |
632 | seg_to = min(rg->to, t); | |
633 | ||
634 | chg += seg_to - seg_from; | |
635 | } | |
7b24d861 | 636 | spin_unlock(&resv->lock); |
84afd99b AW |
637 | |
638 | return chg; | |
639 | } | |
640 | ||
e7c4b0bf AW |
641 | /* |
642 | * Convert the address within this vma to the page offset within | |
643 | * the mapping, in pagecache page units; huge pages here. | |
644 | */ | |
a5516438 AK |
645 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
646 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 647 | { |
a5516438 AK |
648 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
649 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
650 | } |
651 | ||
0fe6e20b NH |
652 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
653 | unsigned long address) | |
654 | { | |
655 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
656 | } | |
dee41079 | 657 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 658 | |
08fba699 MG |
659 | /* |
660 | * Return the size of the pages allocated when backing a VMA. In the majority | |
661 | * cases this will be same size as used by the page table entries. | |
662 | */ | |
663 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
664 | { | |
05ea8860 DW |
665 | if (vma->vm_ops && vma->vm_ops->pagesize) |
666 | return vma->vm_ops->pagesize(vma); | |
667 | return PAGE_SIZE; | |
08fba699 | 668 | } |
f340ca0f | 669 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 670 | |
3340289d MG |
671 | /* |
672 | * Return the page size being used by the MMU to back a VMA. In the majority | |
673 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
674 | * architectures where it differs, an architecture-specific 'strong' |
675 | * version of this symbol is required. | |
3340289d | 676 | */ |
09135cc5 | 677 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
678 | { |
679 | return vma_kernel_pagesize(vma); | |
680 | } | |
3340289d | 681 | |
84afd99b AW |
682 | /* |
683 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
684 | * bits of the reservation map pointer, which are always clear due to | |
685 | * alignment. | |
686 | */ | |
687 | #define HPAGE_RESV_OWNER (1UL << 0) | |
688 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 689 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 690 | |
a1e78772 MG |
691 | /* |
692 | * These helpers are used to track how many pages are reserved for | |
693 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
694 | * is guaranteed to have their future faults succeed. | |
695 | * | |
696 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
697 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
698 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
699 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
700 | * |
701 | * The private mapping reservation is represented in a subtly different | |
702 | * manner to a shared mapping. A shared mapping has a region map associated | |
703 | * with the underlying file, this region map represents the backing file | |
704 | * pages which have ever had a reservation assigned which this persists even | |
705 | * after the page is instantiated. A private mapping has a region map | |
706 | * associated with the original mmap which is attached to all VMAs which | |
707 | * reference it, this region map represents those offsets which have consumed | |
708 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 709 | */ |
e7c4b0bf AW |
710 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
711 | { | |
712 | return (unsigned long)vma->vm_private_data; | |
713 | } | |
714 | ||
715 | static void set_vma_private_data(struct vm_area_struct *vma, | |
716 | unsigned long value) | |
717 | { | |
718 | vma->vm_private_data = (void *)value; | |
719 | } | |
720 | ||
9119a41e | 721 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
722 | { |
723 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
724 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
725 | ||
726 | if (!resv_map || !rg) { | |
727 | kfree(resv_map); | |
728 | kfree(rg); | |
84afd99b | 729 | return NULL; |
5e911373 | 730 | } |
84afd99b AW |
731 | |
732 | kref_init(&resv_map->refs); | |
7b24d861 | 733 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
734 | INIT_LIST_HEAD(&resv_map->regions); |
735 | ||
5e911373 MK |
736 | resv_map->adds_in_progress = 0; |
737 | ||
738 | INIT_LIST_HEAD(&resv_map->region_cache); | |
739 | list_add(&rg->link, &resv_map->region_cache); | |
740 | resv_map->region_cache_count = 1; | |
741 | ||
84afd99b AW |
742 | return resv_map; |
743 | } | |
744 | ||
9119a41e | 745 | void resv_map_release(struct kref *ref) |
84afd99b AW |
746 | { |
747 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
748 | struct list_head *head = &resv_map->region_cache; |
749 | struct file_region *rg, *trg; | |
84afd99b AW |
750 | |
751 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 752 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
753 | |
754 | /* ... and any entries left in the cache */ | |
755 | list_for_each_entry_safe(rg, trg, head, link) { | |
756 | list_del(&rg->link); | |
757 | kfree(rg); | |
758 | } | |
759 | ||
760 | VM_BUG_ON(resv_map->adds_in_progress); | |
761 | ||
84afd99b AW |
762 | kfree(resv_map); |
763 | } | |
764 | ||
4e35f483 JK |
765 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
766 | { | |
f27a5136 MK |
767 | /* |
768 | * At inode evict time, i_mapping may not point to the original | |
769 | * address space within the inode. This original address space | |
770 | * contains the pointer to the resv_map. So, always use the | |
771 | * address space embedded within the inode. | |
772 | * The VERY common case is inode->mapping == &inode->i_data but, | |
773 | * this may not be true for device special inodes. | |
774 | */ | |
775 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
776 | } |
777 | ||
84afd99b | 778 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 779 | { |
81d1b09c | 780 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
781 | if (vma->vm_flags & VM_MAYSHARE) { |
782 | struct address_space *mapping = vma->vm_file->f_mapping; | |
783 | struct inode *inode = mapping->host; | |
784 | ||
785 | return inode_resv_map(inode); | |
786 | ||
787 | } else { | |
84afd99b AW |
788 | return (struct resv_map *)(get_vma_private_data(vma) & |
789 | ~HPAGE_RESV_MASK); | |
4e35f483 | 790 | } |
a1e78772 MG |
791 | } |
792 | ||
84afd99b | 793 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 794 | { |
81d1b09c SL |
795 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
796 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 797 | |
84afd99b AW |
798 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
799 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
800 | } |
801 | ||
802 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
803 | { | |
81d1b09c SL |
804 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
805 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
806 | |
807 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
808 | } |
809 | ||
810 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
811 | { | |
81d1b09c | 812 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
813 | |
814 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
815 | } |
816 | ||
04f2cbe3 | 817 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
818 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
819 | { | |
81d1b09c | 820 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 821 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
822 | vma->vm_private_data = (void *)0; |
823 | } | |
824 | ||
825 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 826 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 827 | { |
af0ed73e JK |
828 | if (vma->vm_flags & VM_NORESERVE) { |
829 | /* | |
830 | * This address is already reserved by other process(chg == 0), | |
831 | * so, we should decrement reserved count. Without decrementing, | |
832 | * reserve count remains after releasing inode, because this | |
833 | * allocated page will go into page cache and is regarded as | |
834 | * coming from reserved pool in releasing step. Currently, we | |
835 | * don't have any other solution to deal with this situation | |
836 | * properly, so add work-around here. | |
837 | */ | |
838 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 839 | return true; |
af0ed73e | 840 | else |
559ec2f8 | 841 | return false; |
af0ed73e | 842 | } |
a63884e9 JK |
843 | |
844 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
845 | if (vma->vm_flags & VM_MAYSHARE) { |
846 | /* | |
847 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
848 | * be a region map for all pages. The only situation where | |
849 | * there is no region map is if a hole was punched via | |
850 | * fallocate. In this case, there really are no reverves to | |
851 | * use. This situation is indicated if chg != 0. | |
852 | */ | |
853 | if (chg) | |
854 | return false; | |
855 | else | |
856 | return true; | |
857 | } | |
a63884e9 JK |
858 | |
859 | /* | |
860 | * Only the process that called mmap() has reserves for | |
861 | * private mappings. | |
862 | */ | |
67961f9d MK |
863 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
864 | /* | |
865 | * Like the shared case above, a hole punch or truncate | |
866 | * could have been performed on the private mapping. | |
867 | * Examine the value of chg to determine if reserves | |
868 | * actually exist or were previously consumed. | |
869 | * Very Subtle - The value of chg comes from a previous | |
870 | * call to vma_needs_reserves(). The reserve map for | |
871 | * private mappings has different (opposite) semantics | |
872 | * than that of shared mappings. vma_needs_reserves() | |
873 | * has already taken this difference in semantics into | |
874 | * account. Therefore, the meaning of chg is the same | |
875 | * as in the shared case above. Code could easily be | |
876 | * combined, but keeping it separate draws attention to | |
877 | * subtle differences. | |
878 | */ | |
879 | if (chg) | |
880 | return false; | |
881 | else | |
882 | return true; | |
883 | } | |
a63884e9 | 884 | |
559ec2f8 | 885 | return false; |
a1e78772 MG |
886 | } |
887 | ||
a5516438 | 888 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
889 | { |
890 | int nid = page_to_nid(page); | |
0edaecfa | 891 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
892 | h->free_huge_pages++; |
893 | h->free_huge_pages_node[nid]++; | |
c2f2139b | 894 | SetPageHugeFreed(page); |
1da177e4 LT |
895 | } |
896 | ||
94310cbc | 897 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
898 | { |
899 | struct page *page; | |
900 | ||
c8721bbb | 901 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 902 | if (!PageHWPoison(page)) |
c8721bbb NH |
903 | break; |
904 | /* | |
905 | * if 'non-isolated free hugepage' not found on the list, | |
906 | * the allocation fails. | |
907 | */ | |
908 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 909 | return NULL; |
0edaecfa | 910 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 911 | set_page_refcounted(page); |
c2f2139b | 912 | ClearPageHugeFreed(page); |
bf50bab2 NH |
913 | h->free_huge_pages--; |
914 | h->free_huge_pages_node[nid]--; | |
915 | return page; | |
916 | } | |
917 | ||
3e59fcb0 MH |
918 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
919 | nodemask_t *nmask) | |
94310cbc | 920 | { |
3e59fcb0 MH |
921 | unsigned int cpuset_mems_cookie; |
922 | struct zonelist *zonelist; | |
923 | struct zone *zone; | |
924 | struct zoneref *z; | |
98fa15f3 | 925 | int node = NUMA_NO_NODE; |
94310cbc | 926 | |
3e59fcb0 MH |
927 | zonelist = node_zonelist(nid, gfp_mask); |
928 | ||
929 | retry_cpuset: | |
930 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
931 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
932 | struct page *page; | |
933 | ||
934 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
935 | continue; | |
936 | /* | |
937 | * no need to ask again on the same node. Pool is node rather than | |
938 | * zone aware | |
939 | */ | |
940 | if (zone_to_nid(zone) == node) | |
941 | continue; | |
942 | node = zone_to_nid(zone); | |
94310cbc | 943 | |
94310cbc AK |
944 | page = dequeue_huge_page_node_exact(h, node); |
945 | if (page) | |
946 | return page; | |
947 | } | |
3e59fcb0 MH |
948 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
949 | goto retry_cpuset; | |
950 | ||
94310cbc AK |
951 | return NULL; |
952 | } | |
953 | ||
86cdb465 NH |
954 | /* Movability of hugepages depends on migration support. */ |
955 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
956 | { | |
7ed2c31d | 957 | if (hugepage_movable_supported(h)) |
86cdb465 NH |
958 | return GFP_HIGHUSER_MOVABLE; |
959 | else | |
960 | return GFP_HIGHUSER; | |
961 | } | |
962 | ||
a5516438 AK |
963 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
964 | struct vm_area_struct *vma, | |
af0ed73e JK |
965 | unsigned long address, int avoid_reserve, |
966 | long chg) | |
1da177e4 | 967 | { |
3e59fcb0 | 968 | struct page *page; |
480eccf9 | 969 | struct mempolicy *mpol; |
04ec6264 | 970 | gfp_t gfp_mask; |
3e59fcb0 | 971 | nodemask_t *nodemask; |
04ec6264 | 972 | int nid; |
1da177e4 | 973 | |
a1e78772 MG |
974 | /* |
975 | * A child process with MAP_PRIVATE mappings created by their parent | |
976 | * have no page reserves. This check ensures that reservations are | |
977 | * not "stolen". The child may still get SIGKILLed | |
978 | */ | |
af0ed73e | 979 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 980 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 981 | goto err; |
a1e78772 | 982 | |
04f2cbe3 | 983 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 984 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 985 | goto err; |
04f2cbe3 | 986 | |
04ec6264 VB |
987 | gfp_mask = htlb_alloc_mask(h); |
988 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
989 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
990 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
991 | SetPagePrivate(page); | |
992 | h->resv_huge_pages--; | |
1da177e4 | 993 | } |
cc9a6c87 | 994 | |
52cd3b07 | 995 | mpol_cond_put(mpol); |
1da177e4 | 996 | return page; |
cc9a6c87 MG |
997 | |
998 | err: | |
cc9a6c87 | 999 | return NULL; |
1da177e4 LT |
1000 | } |
1001 | ||
1cac6f2c LC |
1002 | /* |
1003 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1004 | * We may have allocated or freed a huge page based on a different | |
1005 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1006 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1007 | * node for alloc or free. | |
1008 | */ | |
1009 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1010 | { | |
0edaf86c | 1011 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1012 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1013 | ||
1014 | return nid; | |
1015 | } | |
1016 | ||
1017 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1018 | { | |
1019 | if (!node_isset(nid, *nodes_allowed)) | |
1020 | nid = next_node_allowed(nid, nodes_allowed); | |
1021 | return nid; | |
1022 | } | |
1023 | ||
1024 | /* | |
1025 | * returns the previously saved node ["this node"] from which to | |
1026 | * allocate a persistent huge page for the pool and advance the | |
1027 | * next node from which to allocate, handling wrap at end of node | |
1028 | * mask. | |
1029 | */ | |
1030 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1031 | nodemask_t *nodes_allowed) | |
1032 | { | |
1033 | int nid; | |
1034 | ||
1035 | VM_BUG_ON(!nodes_allowed); | |
1036 | ||
1037 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1038 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1039 | ||
1040 | return nid; | |
1041 | } | |
1042 | ||
1043 | /* | |
1044 | * helper for free_pool_huge_page() - return the previously saved | |
1045 | * node ["this node"] from which to free a huge page. Advance the | |
1046 | * next node id whether or not we find a free huge page to free so | |
1047 | * that the next attempt to free addresses the next node. | |
1048 | */ | |
1049 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1050 | { | |
1051 | int nid; | |
1052 | ||
1053 | VM_BUG_ON(!nodes_allowed); | |
1054 | ||
1055 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1056 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1057 | ||
1058 | return nid; | |
1059 | } | |
1060 | ||
1061 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1062 | for (nr_nodes = nodes_weight(*mask); \ | |
1063 | nr_nodes > 0 && \ | |
1064 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1065 | nr_nodes--) | |
1066 | ||
1067 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1068 | for (nr_nodes = nodes_weight(*mask); \ | |
1069 | nr_nodes > 0 && \ | |
1070 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1071 | nr_nodes--) | |
1072 | ||
e1073d1e | 1073 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1074 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1075 | unsigned int order) |
944d9fec LC |
1076 | { |
1077 | int i; | |
1078 | int nr_pages = 1 << order; | |
1079 | struct page *p = page + 1; | |
1080 | ||
c8cc708a | 1081 | atomic_set(compound_mapcount_ptr(page), 0); |
944d9fec | 1082 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1083 | clear_compound_head(p); |
944d9fec | 1084 | set_page_refcounted(p); |
944d9fec LC |
1085 | } |
1086 | ||
1087 | set_compound_order(page, 0); | |
1088 | __ClearPageHead(page); | |
1089 | } | |
1090 | ||
d00181b9 | 1091 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1092 | { |
1093 | free_contig_range(page_to_pfn(page), 1 << order); | |
1094 | } | |
1095 | ||
4eb0716e | 1096 | #ifdef CONFIG_CONTIG_ALLOC |
944d9fec | 1097 | static int __alloc_gigantic_page(unsigned long start_pfn, |
79b63f12 | 1098 | unsigned long nr_pages, gfp_t gfp_mask) |
944d9fec LC |
1099 | { |
1100 | unsigned long end_pfn = start_pfn + nr_pages; | |
ca96b625 | 1101 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, |
79b63f12 | 1102 | gfp_mask); |
944d9fec LC |
1103 | } |
1104 | ||
f44b2dda JK |
1105 | static bool pfn_range_valid_gigantic(struct zone *z, |
1106 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1107 | { |
1108 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1109 | struct page *page; | |
1110 | ||
1111 | for (i = start_pfn; i < end_pfn; i++) { | |
f231fe42 DH |
1112 | page = pfn_to_online_page(i); |
1113 | if (!page) | |
944d9fec LC |
1114 | return false; |
1115 | ||
f44b2dda JK |
1116 | if (page_zone(page) != z) |
1117 | return false; | |
1118 | ||
944d9fec LC |
1119 | if (PageReserved(page)) |
1120 | return false; | |
1121 | ||
1122 | if (page_count(page) > 0) | |
1123 | return false; | |
1124 | ||
1125 | if (PageHuge(page)) | |
1126 | return false; | |
1127 | } | |
1128 | ||
1129 | return true; | |
1130 | } | |
1131 | ||
1132 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1133 | unsigned long start_pfn, unsigned long nr_pages) | |
1134 | { | |
1135 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1136 | return zone_spans_pfn(zone, last_pfn); | |
1137 | } | |
1138 | ||
d9cc948f MH |
1139 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1140 | int nid, nodemask_t *nodemask) | |
944d9fec | 1141 | { |
79b63f12 | 1142 | unsigned int order = huge_page_order(h); |
944d9fec LC |
1143 | unsigned long nr_pages = 1 << order; |
1144 | unsigned long ret, pfn, flags; | |
79b63f12 MH |
1145 | struct zonelist *zonelist; |
1146 | struct zone *zone; | |
1147 | struct zoneref *z; | |
944d9fec | 1148 | |
79b63f12 | 1149 | zonelist = node_zonelist(nid, gfp_mask); |
d9cc948f | 1150 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { |
79b63f12 | 1151 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec | 1152 | |
79b63f12 MH |
1153 | pfn = ALIGN(zone->zone_start_pfn, nr_pages); |
1154 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) { | |
1155 | if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { | |
944d9fec LC |
1156 | /* |
1157 | * We release the zone lock here because | |
1158 | * alloc_contig_range() will also lock the zone | |
1159 | * at some point. If there's an allocation | |
1160 | * spinning on this lock, it may win the race | |
1161 | * and cause alloc_contig_range() to fail... | |
1162 | */ | |
79b63f12 MH |
1163 | spin_unlock_irqrestore(&zone->lock, flags); |
1164 | ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); | |
944d9fec LC |
1165 | if (!ret) |
1166 | return pfn_to_page(pfn); | |
79b63f12 | 1167 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec LC |
1168 | } |
1169 | pfn += nr_pages; | |
1170 | } | |
1171 | ||
79b63f12 | 1172 | spin_unlock_irqrestore(&zone->lock, flags); |
944d9fec LC |
1173 | } |
1174 | ||
1175 | return NULL; | |
1176 | } | |
1177 | ||
1178 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1179 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1180 | #else /* !CONFIG_CONTIG_ALLOC */ |
1181 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1182 | int nid, nodemask_t *nodemask) | |
1183 | { | |
1184 | return NULL; | |
1185 | } | |
1186 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1187 | |
e1073d1e | 1188 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1189 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1190 | int nid, nodemask_t *nodemask) |
1191 | { | |
1192 | return NULL; | |
1193 | } | |
d00181b9 | 1194 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1195 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1196 | unsigned int order) { } |
944d9fec LC |
1197 | #endif |
1198 | ||
a5516438 | 1199 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1200 | { |
1201 | int i; | |
fd6126b4 | 1202 | struct page *subpage = page; |
a5516438 | 1203 | |
4eb0716e | 1204 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1205 | return; |
18229df5 | 1206 | |
a5516438 AK |
1207 | h->nr_huge_pages--; |
1208 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
fd6126b4 MK |
1209 | for (i = 0; i < pages_per_huge_page(h); |
1210 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1211 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1212 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1213 | 1 << PG_active | 1 << PG_private | |
1214 | 1 << PG_writeback); | |
6af2acb6 | 1215 | } |
309381fe | 1216 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1217 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1218 | set_page_refcounted(page); |
944d9fec LC |
1219 | if (hstate_is_gigantic(h)) { |
1220 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1221 | free_gigantic_page(page, huge_page_order(h)); | |
1222 | } else { | |
944d9fec LC |
1223 | __free_pages(page, huge_page_order(h)); |
1224 | } | |
6af2acb6 AL |
1225 | } |
1226 | ||
e5ff2159 AK |
1227 | struct hstate *size_to_hstate(unsigned long size) |
1228 | { | |
1229 | struct hstate *h; | |
1230 | ||
1231 | for_each_hstate(h) { | |
1232 | if (huge_page_size(h) == size) | |
1233 | return h; | |
1234 | } | |
1235 | return NULL; | |
1236 | } | |
1237 | ||
bcc54222 NH |
1238 | /* |
1239 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1240 | * to hstate->hugepage_activelist.) | |
1241 | * | |
1242 | * This function can be called for tail pages, but never returns true for them. | |
1243 | */ | |
1244 | bool page_huge_active(struct page *page) | |
1245 | { | |
08f3c9c1 | 1246 | return PageHeadHuge(page) && PagePrivate(&page[1]); |
bcc54222 NH |
1247 | } |
1248 | ||
1249 | /* never called for tail page */ | |
55f5a39f | 1250 | void set_page_huge_active(struct page *page) |
bcc54222 NH |
1251 | { |
1252 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1253 | SetPagePrivate(&page[1]); | |
1254 | } | |
1255 | ||
1256 | static void clear_page_huge_active(struct page *page) | |
1257 | { | |
1258 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1259 | ClearPagePrivate(&page[1]); | |
1260 | } | |
1261 | ||
ab5ac90a MH |
1262 | /* |
1263 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1264 | * code | |
1265 | */ | |
1266 | static inline bool PageHugeTemporary(struct page *page) | |
1267 | { | |
1268 | if (!PageHuge(page)) | |
1269 | return false; | |
1270 | ||
1271 | return (unsigned long)page[2].mapping == -1U; | |
1272 | } | |
1273 | ||
1274 | static inline void SetPageHugeTemporary(struct page *page) | |
1275 | { | |
1276 | page[2].mapping = (void *)-1U; | |
1277 | } | |
1278 | ||
1279 | static inline void ClearPageHugeTemporary(struct page *page) | |
1280 | { | |
1281 | page[2].mapping = NULL; | |
1282 | } | |
1283 | ||
ce44b015 | 1284 | static void __free_huge_page(struct page *page) |
27a85ef1 | 1285 | { |
a5516438 AK |
1286 | /* |
1287 | * Can't pass hstate in here because it is called from the | |
1288 | * compound page destructor. | |
1289 | */ | |
e5ff2159 | 1290 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1291 | int nid = page_to_nid(page); |
90481622 DG |
1292 | struct hugepage_subpool *spool = |
1293 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1294 | bool restore_reserve; |
27a85ef1 | 1295 | |
b4330afb MK |
1296 | VM_BUG_ON_PAGE(page_count(page), page); |
1297 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc YW |
1298 | |
1299 | set_page_private(page, 0); | |
1300 | page->mapping = NULL; | |
07443a85 | 1301 | restore_reserve = PagePrivate(page); |
16c794b4 | 1302 | ClearPagePrivate(page); |
27a85ef1 | 1303 | |
1c5ecae3 | 1304 | /* |
0919e1b6 MK |
1305 | * If PagePrivate() was set on page, page allocation consumed a |
1306 | * reservation. If the page was associated with a subpool, there | |
1307 | * would have been a page reserved in the subpool before allocation | |
1308 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1309 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1310 | * remove the reserved page from the subpool. | |
1c5ecae3 | 1311 | */ |
0919e1b6 MK |
1312 | if (!restore_reserve) { |
1313 | /* | |
1314 | * A return code of zero implies that the subpool will be | |
1315 | * under its minimum size if the reservation is not restored | |
1316 | * after page is free. Therefore, force restore_reserve | |
1317 | * operation. | |
1318 | */ | |
1319 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1320 | restore_reserve = true; | |
1321 | } | |
1c5ecae3 | 1322 | |
27a85ef1 | 1323 | spin_lock(&hugetlb_lock); |
bcc54222 | 1324 | clear_page_huge_active(page); |
6d76dcf4 AK |
1325 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1326 | pages_per_huge_page(h), page); | |
07443a85 JK |
1327 | if (restore_reserve) |
1328 | h->resv_huge_pages++; | |
1329 | ||
ab5ac90a MH |
1330 | if (PageHugeTemporary(page)) { |
1331 | list_del(&page->lru); | |
1332 | ClearPageHugeTemporary(page); | |
1333 | update_and_free_page(h, page); | |
1334 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1335 | /* remove the page from active list */ |
1336 | list_del(&page->lru); | |
a5516438 AK |
1337 | update_and_free_page(h, page); |
1338 | h->surplus_huge_pages--; | |
1339 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1340 | } else { |
5d3a551c | 1341 | arch_clear_hugepage_flags(page); |
a5516438 | 1342 | enqueue_huge_page(h, page); |
7893d1d5 | 1343 | } |
27a85ef1 DG |
1344 | spin_unlock(&hugetlb_lock); |
1345 | } | |
1346 | ||
ce44b015 WL |
1347 | /* |
1348 | * As free_huge_page() can be called from a non-task context, we have | |
1349 | * to defer the actual freeing in a workqueue to prevent potential | |
1350 | * hugetlb_lock deadlock. | |
1351 | * | |
1352 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1353 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1354 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1355 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1356 | */ | |
1357 | static LLIST_HEAD(hpage_freelist); | |
1358 | ||
1359 | static void free_hpage_workfn(struct work_struct *work) | |
1360 | { | |
1361 | struct llist_node *node; | |
1362 | struct page *page; | |
1363 | ||
1364 | node = llist_del_all(&hpage_freelist); | |
1365 | ||
1366 | while (node) { | |
1367 | page = container_of((struct address_space **)node, | |
1368 | struct page, mapping); | |
1369 | node = node->next; | |
1370 | __free_huge_page(page); | |
1371 | } | |
1372 | } | |
1373 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1374 | ||
1375 | void free_huge_page(struct page *page) | |
1376 | { | |
1377 | /* | |
1378 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1379 | */ | |
1380 | if (!in_task()) { | |
1381 | /* | |
1382 | * Only call schedule_work() if hpage_freelist is previously | |
1383 | * empty. Otherwise, schedule_work() had been called but the | |
1384 | * workfn hasn't retrieved the list yet. | |
1385 | */ | |
1386 | if (llist_add((struct llist_node *)&page->mapping, | |
1387 | &hpage_freelist)) | |
1388 | schedule_work(&free_hpage_work); | |
1389 | return; | |
1390 | } | |
1391 | ||
1392 | __free_huge_page(page); | |
1393 | } | |
1394 | ||
a5516438 | 1395 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1396 | { |
0edaecfa | 1397 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1398 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1399 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1400 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1401 | h->nr_huge_pages++; |
1402 | h->nr_huge_pages_node[nid]++; | |
c2f2139b | 1403 | ClearPageHugeFreed(page); |
b7ba30c6 | 1404 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1405 | } |
1406 | ||
d00181b9 | 1407 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1408 | { |
1409 | int i; | |
1410 | int nr_pages = 1 << order; | |
1411 | struct page *p = page + 1; | |
1412 | ||
1413 | /* we rely on prep_new_huge_page to set the destructor */ | |
1414 | set_compound_order(page, order); | |
ef5a22be | 1415 | __ClearPageReserved(page); |
de09d31d | 1416 | __SetPageHead(page); |
20a0307c | 1417 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1418 | /* |
1419 | * For gigantic hugepages allocated through bootmem at | |
1420 | * boot, it's safer to be consistent with the not-gigantic | |
1421 | * hugepages and clear the PG_reserved bit from all tail pages | |
1422 | * too. Otherwse drivers using get_user_pages() to access tail | |
1423 | * pages may get the reference counting wrong if they see | |
1424 | * PG_reserved set on a tail page (despite the head page not | |
1425 | * having PG_reserved set). Enforcing this consistency between | |
1426 | * head and tail pages allows drivers to optimize away a check | |
1427 | * on the head page when they need know if put_page() is needed | |
1428 | * after get_user_pages(). | |
1429 | */ | |
1430 | __ClearPageReserved(p); | |
58a84aa9 | 1431 | set_page_count(p, 0); |
1d798ca3 | 1432 | set_compound_head(p, page); |
20a0307c | 1433 | } |
b4330afb | 1434 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1435 | } |
1436 | ||
7795912c AM |
1437 | /* |
1438 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1439 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1440 | * details. | |
1441 | */ | |
20a0307c WF |
1442 | int PageHuge(struct page *page) |
1443 | { | |
20a0307c WF |
1444 | if (!PageCompound(page)) |
1445 | return 0; | |
1446 | ||
1447 | page = compound_head(page); | |
f1e61557 | 1448 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1449 | } |
43131e14 NH |
1450 | EXPORT_SYMBOL_GPL(PageHuge); |
1451 | ||
27c73ae7 AA |
1452 | /* |
1453 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1454 | * normal or transparent huge pages. | |
1455 | */ | |
1456 | int PageHeadHuge(struct page *page_head) | |
1457 | { | |
27c73ae7 AA |
1458 | if (!PageHead(page_head)) |
1459 | return 0; | |
1460 | ||
758f66a2 | 1461 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1462 | } |
27c73ae7 | 1463 | |
2d57df77 | 1464 | pgoff_t hugetlb_basepage_index(struct page *page) |
13d60f4b ZY |
1465 | { |
1466 | struct page *page_head = compound_head(page); | |
1467 | pgoff_t index = page_index(page_head); | |
1468 | unsigned long compound_idx; | |
1469 | ||
13d60f4b ZY |
1470 | if (compound_order(page_head) >= MAX_ORDER) |
1471 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1472 | else | |
1473 | compound_idx = page - page_head; | |
1474 | ||
1475 | return (index << compound_order(page_head)) + compound_idx; | |
1476 | } | |
1477 | ||
0c397dae | 1478 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1479 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1480 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1481 | { |
af0fb9df | 1482 | int order = huge_page_order(h); |
1da177e4 | 1483 | struct page *page; |
f60858f9 | 1484 | bool alloc_try_hard = true; |
f96efd58 | 1485 | |
f60858f9 MK |
1486 | /* |
1487 | * By default we always try hard to allocate the page with | |
1488 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1489 | * a loop (to adjust global huge page counts) and previous allocation | |
1490 | * failed, do not continue to try hard on the same node. Use the | |
1491 | * node_alloc_noretry bitmap to manage this state information. | |
1492 | */ | |
1493 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1494 | alloc_try_hard = false; | |
1495 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1496 | if (alloc_try_hard) | |
1497 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1498 | if (nid == NUMA_NO_NODE) |
1499 | nid = numa_mem_id(); | |
1500 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1501 | if (page) | |
1502 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1503 | else | |
1504 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1505 | |
f60858f9 MK |
1506 | /* |
1507 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1508 | * indicates an overall state change. Clear bit so that we resume | |
1509 | * normal 'try hard' allocations. | |
1510 | */ | |
1511 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1512 | node_clear(nid, *node_alloc_noretry); | |
1513 | ||
1514 | /* | |
1515 | * If we tried hard to get a page but failed, set bit so that | |
1516 | * subsequent attempts will not try as hard until there is an | |
1517 | * overall state change. | |
1518 | */ | |
1519 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1520 | node_set(nid, *node_alloc_noretry); | |
1521 | ||
63b4613c NA |
1522 | return page; |
1523 | } | |
1524 | ||
0c397dae MH |
1525 | /* |
1526 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1527 | * should use this function to get new hugetlb pages | |
1528 | */ | |
1529 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1530 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1531 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1532 | { |
1533 | struct page *page; | |
1534 | ||
1535 | if (hstate_is_gigantic(h)) | |
1536 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1537 | else | |
1538 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1539 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1540 | if (!page) |
1541 | return NULL; | |
1542 | ||
1543 | if (hstate_is_gigantic(h)) | |
1544 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1545 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1546 | ||
1547 | return page; | |
1548 | } | |
1549 | ||
af0fb9df MH |
1550 | /* |
1551 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1552 | * manner. | |
1553 | */ | |
f60858f9 MK |
1554 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1555 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1556 | { |
1557 | struct page *page; | |
1558 | int nr_nodes, node; | |
af0fb9df | 1559 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1560 | |
1561 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1562 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1563 | node_alloc_noretry); | |
af0fb9df | 1564 | if (page) |
b2261026 | 1565 | break; |
b2261026 JK |
1566 | } |
1567 | ||
af0fb9df MH |
1568 | if (!page) |
1569 | return 0; | |
b2261026 | 1570 | |
af0fb9df MH |
1571 | put_page(page); /* free it into the hugepage allocator */ |
1572 | ||
1573 | return 1; | |
b2261026 JK |
1574 | } |
1575 | ||
e8c5c824 LS |
1576 | /* |
1577 | * Free huge page from pool from next node to free. | |
1578 | * Attempt to keep persistent huge pages more or less | |
1579 | * balanced over allowed nodes. | |
1580 | * Called with hugetlb_lock locked. | |
1581 | */ | |
6ae11b27 LS |
1582 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1583 | bool acct_surplus) | |
e8c5c824 | 1584 | { |
b2261026 | 1585 | int nr_nodes, node; |
e8c5c824 LS |
1586 | int ret = 0; |
1587 | ||
b2261026 | 1588 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1589 | /* |
1590 | * If we're returning unused surplus pages, only examine | |
1591 | * nodes with surplus pages. | |
1592 | */ | |
b2261026 JK |
1593 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1594 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1595 | struct page *page = |
b2261026 | 1596 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1597 | struct page, lru); |
1598 | list_del(&page->lru); | |
1599 | h->free_huge_pages--; | |
b2261026 | 1600 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1601 | if (acct_surplus) { |
1602 | h->surplus_huge_pages--; | |
b2261026 | 1603 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1604 | } |
e8c5c824 LS |
1605 | update_and_free_page(h, page); |
1606 | ret = 1; | |
9a76db09 | 1607 | break; |
e8c5c824 | 1608 | } |
b2261026 | 1609 | } |
e8c5c824 LS |
1610 | |
1611 | return ret; | |
1612 | } | |
1613 | ||
c8721bbb NH |
1614 | /* |
1615 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1616 | * nothing for in-use hugepages and non-hugepages. |
1617 | * This function returns values like below: | |
1618 | * | |
1619 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1620 | * (allocated or reserved.) | |
1621 | * 0: successfully dissolved free hugepages or the page is not a | |
1622 | * hugepage (considered as already dissolved) | |
c8721bbb | 1623 | */ |
c3114a84 | 1624 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1625 | { |
6bc9b564 | 1626 | int rc = -EBUSY; |
082d5b6b | 1627 | |
c2f2139b | 1628 | retry: |
faf53def NH |
1629 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1630 | if (!PageHuge(page)) | |
1631 | return 0; | |
1632 | ||
c8721bbb | 1633 | spin_lock(&hugetlb_lock); |
faf53def NH |
1634 | if (!PageHuge(page)) { |
1635 | rc = 0; | |
1636 | goto out; | |
1637 | } | |
1638 | ||
1639 | if (!page_count(page)) { | |
2247bb33 GS |
1640 | struct page *head = compound_head(page); |
1641 | struct hstate *h = page_hstate(head); | |
1642 | int nid = page_to_nid(head); | |
6bc9b564 | 1643 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1644 | goto out; |
c2f2139b MS |
1645 | |
1646 | /* | |
1647 | * We should make sure that the page is already on the free list | |
1648 | * when it is dissolved. | |
1649 | */ | |
1650 | if (unlikely(!PageHugeFreed(head))) { | |
1651 | spin_unlock(&hugetlb_lock); | |
1652 | cond_resched(); | |
1653 | ||
1654 | /* | |
1655 | * Theoretically, we should return -EBUSY when we | |
1656 | * encounter this race. In fact, we have a chance | |
1657 | * to successfully dissolve the page if we do a | |
1658 | * retry. Because the race window is quite small. | |
1659 | * If we seize this opportunity, it is an optimization | |
1660 | * for increasing the success rate of dissolving page. | |
1661 | */ | |
1662 | goto retry; | |
1663 | } | |
1664 | ||
c3114a84 AK |
1665 | /* |
1666 | * Move PageHWPoison flag from head page to the raw error page, | |
1667 | * which makes any subpages rather than the error page reusable. | |
1668 | */ | |
1669 | if (PageHWPoison(head) && page != head) { | |
1670 | SetPageHWPoison(page); | |
1671 | ClearPageHWPoison(head); | |
1672 | } | |
2247bb33 | 1673 | list_del(&head->lru); |
c8721bbb NH |
1674 | h->free_huge_pages--; |
1675 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1676 | h->max_huge_pages--; |
2247bb33 | 1677 | update_and_free_page(h, head); |
6bc9b564 | 1678 | rc = 0; |
c8721bbb | 1679 | } |
082d5b6b | 1680 | out: |
c8721bbb | 1681 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1682 | return rc; |
c8721bbb NH |
1683 | } |
1684 | ||
1685 | /* | |
1686 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1687 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1688 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1689 | * part of it lies within the given range. | |
082d5b6b GS |
1690 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1691 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1692 | */ |
082d5b6b | 1693 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1694 | { |
c8721bbb | 1695 | unsigned long pfn; |
eb03aa00 | 1696 | struct page *page; |
082d5b6b | 1697 | int rc = 0; |
c8721bbb | 1698 | |
d0177639 | 1699 | if (!hugepages_supported()) |
082d5b6b | 1700 | return rc; |
d0177639 | 1701 | |
eb03aa00 GS |
1702 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1703 | page = pfn_to_page(pfn); | |
faf53def NH |
1704 | rc = dissolve_free_huge_page(page); |
1705 | if (rc) | |
1706 | break; | |
eb03aa00 | 1707 | } |
082d5b6b GS |
1708 | |
1709 | return rc; | |
c8721bbb NH |
1710 | } |
1711 | ||
ab5ac90a MH |
1712 | /* |
1713 | * Allocates a fresh surplus page from the page allocator. | |
1714 | */ | |
0c397dae | 1715 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1716 | int nid, nodemask_t *nmask) |
7893d1d5 | 1717 | { |
9980d744 | 1718 | struct page *page = NULL; |
7893d1d5 | 1719 | |
bae7f4ae | 1720 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1721 | return NULL; |
1722 | ||
d1c3fb1f | 1723 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1724 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1725 | goto out_unlock; | |
d1c3fb1f NA |
1726 | spin_unlock(&hugetlb_lock); |
1727 | ||
f60858f9 | 1728 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1729 | if (!page) |
0c397dae | 1730 | return NULL; |
d1c3fb1f NA |
1731 | |
1732 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1733 | /* |
1734 | * We could have raced with the pool size change. | |
1735 | * Double check that and simply deallocate the new page | |
1736 | * if we would end up overcommiting the surpluses. Abuse | |
1737 | * temporary page to workaround the nasty free_huge_page | |
1738 | * codeflow | |
1739 | */ | |
1740 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1741 | SetPageHugeTemporary(page); | |
2bf753e6 | 1742 | spin_unlock(&hugetlb_lock); |
9980d744 | 1743 | put_page(page); |
2bf753e6 | 1744 | return NULL; |
9980d744 | 1745 | } else { |
9980d744 | 1746 | h->surplus_huge_pages++; |
4704dea3 | 1747 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1748 | } |
9980d744 MH |
1749 | |
1750 | out_unlock: | |
d1c3fb1f | 1751 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1752 | |
1753 | return page; | |
1754 | } | |
1755 | ||
9a4e9f3b AK |
1756 | struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
1757 | int nid, nodemask_t *nmask) | |
ab5ac90a MH |
1758 | { |
1759 | struct page *page; | |
1760 | ||
1761 | if (hstate_is_gigantic(h)) | |
1762 | return NULL; | |
1763 | ||
f60858f9 | 1764 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1765 | if (!page) |
1766 | return NULL; | |
1767 | ||
1768 | /* | |
1769 | * We do not account these pages as surplus because they are only | |
1770 | * temporary and will be released properly on the last reference | |
1771 | */ | |
ab5ac90a MH |
1772 | SetPageHugeTemporary(page); |
1773 | ||
1774 | return page; | |
1775 | } | |
1776 | ||
099730d6 DH |
1777 | /* |
1778 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1779 | */ | |
e0ec90ee | 1780 | static |
0c397dae | 1781 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1782 | struct vm_area_struct *vma, unsigned long addr) |
1783 | { | |
aaf14e40 MH |
1784 | struct page *page; |
1785 | struct mempolicy *mpol; | |
1786 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1787 | int nid; | |
1788 | nodemask_t *nodemask; | |
1789 | ||
1790 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1791 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1792 | mpol_cond_put(mpol); |
1793 | ||
1794 | return page; | |
099730d6 DH |
1795 | } |
1796 | ||
ab5ac90a | 1797 | /* page migration callback function */ |
bf50bab2 NH |
1798 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
1799 | { | |
aaf14e40 | 1800 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1801 | struct page *page = NULL; |
bf50bab2 | 1802 | |
aaf14e40 MH |
1803 | if (nid != NUMA_NO_NODE) |
1804 | gfp_mask |= __GFP_THISNODE; | |
1805 | ||
bf50bab2 | 1806 | spin_lock(&hugetlb_lock); |
4ef91848 | 1807 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1808 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1809 | spin_unlock(&hugetlb_lock); |
1810 | ||
94ae8ba7 | 1811 | if (!page) |
0c397dae | 1812 | page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1813 | |
1814 | return page; | |
1815 | } | |
1816 | ||
ab5ac90a | 1817 | /* page migration callback function */ |
3e59fcb0 MH |
1818 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
1819 | nodemask_t *nmask) | |
4db9b2ef | 1820 | { |
aaf14e40 | 1821 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
1822 | |
1823 | spin_lock(&hugetlb_lock); | |
1824 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1825 | struct page *page; |
1826 | ||
1827 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1828 | if (page) { | |
1829 | spin_unlock(&hugetlb_lock); | |
1830 | return page; | |
4db9b2ef MH |
1831 | } |
1832 | } | |
1833 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1834 | |
0c397dae | 1835 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1836 | } |
1837 | ||
ebd63723 | 1838 | /* mempolicy aware migration callback */ |
389c8178 MH |
1839 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1840 | unsigned long address) | |
ebd63723 MH |
1841 | { |
1842 | struct mempolicy *mpol; | |
1843 | nodemask_t *nodemask; | |
1844 | struct page *page; | |
ebd63723 MH |
1845 | gfp_t gfp_mask; |
1846 | int node; | |
1847 | ||
ebd63723 MH |
1848 | gfp_mask = htlb_alloc_mask(h); |
1849 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
1850 | page = alloc_huge_page_nodemask(h, node, nodemask); | |
1851 | mpol_cond_put(mpol); | |
1852 | ||
1853 | return page; | |
1854 | } | |
1855 | ||
e4e574b7 | 1856 | /* |
25985edc | 1857 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1858 | * of size 'delta'. |
1859 | */ | |
a5516438 | 1860 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1861 | { |
1862 | struct list_head surplus_list; | |
1863 | struct page *page, *tmp; | |
1864 | int ret, i; | |
1865 | int needed, allocated; | |
28073b02 | 1866 | bool alloc_ok = true; |
e4e574b7 | 1867 | |
a5516438 | 1868 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1869 | if (needed <= 0) { |
a5516438 | 1870 | h->resv_huge_pages += delta; |
e4e574b7 | 1871 | return 0; |
ac09b3a1 | 1872 | } |
e4e574b7 AL |
1873 | |
1874 | allocated = 0; | |
1875 | INIT_LIST_HEAD(&surplus_list); | |
1876 | ||
1877 | ret = -ENOMEM; | |
1878 | retry: | |
1879 | spin_unlock(&hugetlb_lock); | |
1880 | for (i = 0; i < needed; i++) { | |
0c397dae | 1881 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1882 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1883 | if (!page) { |
1884 | alloc_ok = false; | |
1885 | break; | |
1886 | } | |
e4e574b7 | 1887 | list_add(&page->lru, &surplus_list); |
69ed779a | 1888 | cond_resched(); |
e4e574b7 | 1889 | } |
28073b02 | 1890 | allocated += i; |
e4e574b7 AL |
1891 | |
1892 | /* | |
1893 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1894 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1895 | */ | |
1896 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1897 | needed = (h->resv_huge_pages + delta) - |
1898 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1899 | if (needed > 0) { |
1900 | if (alloc_ok) | |
1901 | goto retry; | |
1902 | /* | |
1903 | * We were not able to allocate enough pages to | |
1904 | * satisfy the entire reservation so we free what | |
1905 | * we've allocated so far. | |
1906 | */ | |
1907 | goto free; | |
1908 | } | |
e4e574b7 AL |
1909 | /* |
1910 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1911 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1912 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1913 | * allocator. Commit the entire reservation here to prevent another |
1914 | * process from stealing the pages as they are added to the pool but | |
1915 | * before they are reserved. | |
e4e574b7 AL |
1916 | */ |
1917 | needed += allocated; | |
a5516438 | 1918 | h->resv_huge_pages += delta; |
e4e574b7 | 1919 | ret = 0; |
a9869b83 | 1920 | |
19fc3f0a | 1921 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1922 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1923 | if ((--needed) < 0) |
1924 | break; | |
a9869b83 NH |
1925 | /* |
1926 | * This page is now managed by the hugetlb allocator and has | |
1927 | * no users -- drop the buddy allocator's reference. | |
1928 | */ | |
1929 | put_page_testzero(page); | |
309381fe | 1930 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1931 | enqueue_huge_page(h, page); |
19fc3f0a | 1932 | } |
28073b02 | 1933 | free: |
b0365c8d | 1934 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1935 | |
1936 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1937 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1938 | put_page(page); | |
a9869b83 | 1939 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1940 | |
1941 | return ret; | |
1942 | } | |
1943 | ||
1944 | /* | |
e5bbc8a6 MK |
1945 | * This routine has two main purposes: |
1946 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1947 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1948 | * to the associated reservation map. | |
1949 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1950 | * the reservation. As many as unused_resv_pages may be freed. | |
1951 | * | |
1952 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1953 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1954 | * we must make sure nobody else can claim pages we are in the process of | |
1955 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1956 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1957 | */ |
a5516438 AK |
1958 | static void return_unused_surplus_pages(struct hstate *h, |
1959 | unsigned long unused_resv_pages) | |
e4e574b7 | 1960 | { |
e4e574b7 AL |
1961 | unsigned long nr_pages; |
1962 | ||
aa888a74 | 1963 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1964 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1965 | goto out; |
aa888a74 | 1966 | |
e5bbc8a6 MK |
1967 | /* |
1968 | * Part (or even all) of the reservation could have been backed | |
1969 | * by pre-allocated pages. Only free surplus pages. | |
1970 | */ | |
a5516438 | 1971 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1972 | |
685f3457 LS |
1973 | /* |
1974 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1975 | * evenly across all nodes with memory. Iterate across these nodes |
1976 | * until we can no longer free unreserved surplus pages. This occurs | |
1977 | * when the nodes with surplus pages have no free pages. | |
1978 | * free_pool_huge_page() will balance the the freed pages across the | |
1979 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1980 | * |
1981 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1982 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1983 | * to cover subsequent pages we may free. | |
685f3457 LS |
1984 | */ |
1985 | while (nr_pages--) { | |
e5bbc8a6 MK |
1986 | h->resv_huge_pages--; |
1987 | unused_resv_pages--; | |
8cebfcd0 | 1988 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1989 | goto out; |
7848a4bf | 1990 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1991 | } |
e5bbc8a6 MK |
1992 | |
1993 | out: | |
1994 | /* Fully uncommit the reservation */ | |
1995 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1996 | } |
1997 | ||
5e911373 | 1998 | |
c37f9fb1 | 1999 | /* |
feba16e2 | 2000 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2001 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2002 | * |
2003 | * vma_needs_reservation is called to determine if the huge page at addr | |
2004 | * within the vma has an associated reservation. If a reservation is | |
2005 | * needed, the value 1 is returned. The caller is then responsible for | |
2006 | * managing the global reservation and subpool usage counts. After | |
2007 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2008 | * to add the page to the reservation map. If the page allocation fails, |
2009 | * the reservation must be ended instead of committed. vma_end_reservation | |
2010 | * is called in such cases. | |
cf3ad20b MK |
2011 | * |
2012 | * In the normal case, vma_commit_reservation returns the same value | |
2013 | * as the preceding vma_needs_reservation call. The only time this | |
2014 | * is not the case is if a reserve map was changed between calls. It | |
2015 | * is the responsibility of the caller to notice the difference and | |
2016 | * take appropriate action. | |
96b96a96 MK |
2017 | * |
2018 | * vma_add_reservation is used in error paths where a reservation must | |
2019 | * be restored when a newly allocated huge page must be freed. It is | |
2020 | * to be called after calling vma_needs_reservation to determine if a | |
2021 | * reservation exists. | |
c37f9fb1 | 2022 | */ |
5e911373 MK |
2023 | enum vma_resv_mode { |
2024 | VMA_NEEDS_RESV, | |
2025 | VMA_COMMIT_RESV, | |
feba16e2 | 2026 | VMA_END_RESV, |
96b96a96 | 2027 | VMA_ADD_RESV, |
5e911373 | 2028 | }; |
cf3ad20b MK |
2029 | static long __vma_reservation_common(struct hstate *h, |
2030 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2031 | enum vma_resv_mode mode) |
c37f9fb1 | 2032 | { |
4e35f483 JK |
2033 | struct resv_map *resv; |
2034 | pgoff_t idx; | |
cf3ad20b | 2035 | long ret; |
c37f9fb1 | 2036 | |
4e35f483 JK |
2037 | resv = vma_resv_map(vma); |
2038 | if (!resv) | |
84afd99b | 2039 | return 1; |
c37f9fb1 | 2040 | |
4e35f483 | 2041 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2042 | switch (mode) { |
2043 | case VMA_NEEDS_RESV: | |
cf3ad20b | 2044 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
2045 | break; |
2046 | case VMA_COMMIT_RESV: | |
2047 | ret = region_add(resv, idx, idx + 1); | |
2048 | break; | |
feba16e2 | 2049 | case VMA_END_RESV: |
5e911373 MK |
2050 | region_abort(resv, idx, idx + 1); |
2051 | ret = 0; | |
2052 | break; | |
96b96a96 MK |
2053 | case VMA_ADD_RESV: |
2054 | if (vma->vm_flags & VM_MAYSHARE) | |
2055 | ret = region_add(resv, idx, idx + 1); | |
2056 | else { | |
2057 | region_abort(resv, idx, idx + 1); | |
2058 | ret = region_del(resv, idx, idx + 1); | |
2059 | } | |
2060 | break; | |
5e911373 MK |
2061 | default: |
2062 | BUG(); | |
2063 | } | |
84afd99b | 2064 | |
4e35f483 | 2065 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 2066 | return ret; |
67961f9d MK |
2067 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
2068 | /* | |
2069 | * In most cases, reserves always exist for private mappings. | |
2070 | * However, a file associated with mapping could have been | |
2071 | * hole punched or truncated after reserves were consumed. | |
2072 | * As subsequent fault on such a range will not use reserves. | |
2073 | * Subtle - The reserve map for private mappings has the | |
2074 | * opposite meaning than that of shared mappings. If NO | |
2075 | * entry is in the reserve map, it means a reservation exists. | |
2076 | * If an entry exists in the reserve map, it means the | |
2077 | * reservation has already been consumed. As a result, the | |
2078 | * return value of this routine is the opposite of the | |
2079 | * value returned from reserve map manipulation routines above. | |
2080 | */ | |
2081 | if (ret) | |
2082 | return 0; | |
2083 | else | |
2084 | return 1; | |
2085 | } | |
4e35f483 | 2086 | else |
cf3ad20b | 2087 | return ret < 0 ? ret : 0; |
c37f9fb1 | 2088 | } |
cf3ad20b MK |
2089 | |
2090 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2091 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2092 | { |
5e911373 | 2093 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2094 | } |
84afd99b | 2095 | |
cf3ad20b MK |
2096 | static long vma_commit_reservation(struct hstate *h, |
2097 | struct vm_area_struct *vma, unsigned long addr) | |
2098 | { | |
5e911373 MK |
2099 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2100 | } | |
2101 | ||
feba16e2 | 2102 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2103 | struct vm_area_struct *vma, unsigned long addr) |
2104 | { | |
feba16e2 | 2105 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2106 | } |
2107 | ||
96b96a96 MK |
2108 | static long vma_add_reservation(struct hstate *h, |
2109 | struct vm_area_struct *vma, unsigned long addr) | |
2110 | { | |
2111 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2112 | } | |
2113 | ||
2114 | /* | |
2115 | * This routine is called to restore a reservation on error paths. In the | |
2116 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2117 | * and is about to be freed. If a reservation for the page existed, | |
2118 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2119 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2120 | * the global reservation count will be incremented if PagePrivate is set. | |
2121 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2122 | * reserve map here to be consistent with global reserve count adjustments | |
2123 | * to be made by free_huge_page. | |
2124 | */ | |
2125 | static void restore_reserve_on_error(struct hstate *h, | |
2126 | struct vm_area_struct *vma, unsigned long address, | |
2127 | struct page *page) | |
2128 | { | |
2129 | if (unlikely(PagePrivate(page))) { | |
2130 | long rc = vma_needs_reservation(h, vma, address); | |
2131 | ||
2132 | if (unlikely(rc < 0)) { | |
2133 | /* | |
2134 | * Rare out of memory condition in reserve map | |
2135 | * manipulation. Clear PagePrivate so that | |
2136 | * global reserve count will not be incremented | |
2137 | * by free_huge_page. This will make it appear | |
2138 | * as though the reservation for this page was | |
2139 | * consumed. This may prevent the task from | |
2140 | * faulting in the page at a later time. This | |
2141 | * is better than inconsistent global huge page | |
2142 | * accounting of reserve counts. | |
2143 | */ | |
2144 | ClearPagePrivate(page); | |
2145 | } else if (rc) { | |
2146 | rc = vma_add_reservation(h, vma, address); | |
2147 | if (unlikely(rc < 0)) | |
2148 | /* | |
2149 | * See above comment about rare out of | |
2150 | * memory condition. | |
2151 | */ | |
2152 | ClearPagePrivate(page); | |
2153 | } else | |
2154 | vma_end_reservation(h, vma, address); | |
2155 | } | |
2156 | } | |
2157 | ||
70c3547e | 2158 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2159 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2160 | { |
90481622 | 2161 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2162 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2163 | struct page *page; |
d85f69b0 MK |
2164 | long map_chg, map_commit; |
2165 | long gbl_chg; | |
6d76dcf4 AK |
2166 | int ret, idx; |
2167 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 2168 | |
6d76dcf4 | 2169 | idx = hstate_index(h); |
a1e78772 | 2170 | /* |
d85f69b0 MK |
2171 | * Examine the region/reserve map to determine if the process |
2172 | * has a reservation for the page to be allocated. A return | |
2173 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2174 | */ |
d85f69b0 MK |
2175 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2176 | if (map_chg < 0) | |
76dcee75 | 2177 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2178 | |
2179 | /* | |
2180 | * Processes that did not create the mapping will have no | |
2181 | * reserves as indicated by the region/reserve map. Check | |
2182 | * that the allocation will not exceed the subpool limit. | |
2183 | * Allocations for MAP_NORESERVE mappings also need to be | |
2184 | * checked against any subpool limit. | |
2185 | */ | |
2186 | if (map_chg || avoid_reserve) { | |
2187 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2188 | if (gbl_chg < 0) { | |
feba16e2 | 2189 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2190 | return ERR_PTR(-ENOSPC); |
5e911373 | 2191 | } |
1da177e4 | 2192 | |
d85f69b0 MK |
2193 | /* |
2194 | * Even though there was no reservation in the region/reserve | |
2195 | * map, there could be reservations associated with the | |
2196 | * subpool that can be used. This would be indicated if the | |
2197 | * return value of hugepage_subpool_get_pages() is zero. | |
2198 | * However, if avoid_reserve is specified we still avoid even | |
2199 | * the subpool reservations. | |
2200 | */ | |
2201 | if (avoid_reserve) | |
2202 | gbl_chg = 1; | |
2203 | } | |
2204 | ||
6d76dcf4 | 2205 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2206 | if (ret) |
2207 | goto out_subpool_put; | |
2208 | ||
1da177e4 | 2209 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2210 | /* |
2211 | * glb_chg is passed to indicate whether or not a page must be taken | |
2212 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2213 | * a reservation exists for the allocation. | |
2214 | */ | |
2215 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2216 | if (!page) { |
94ae8ba7 | 2217 | spin_unlock(&hugetlb_lock); |
0c397dae | 2218 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2219 | if (!page) |
2220 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2221 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2222 | SetPagePrivate(page); | |
2223 | h->resv_huge_pages--; | |
2224 | } | |
79dbb236 AK |
2225 | spin_lock(&hugetlb_lock); |
2226 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2227 | /* Fall through */ |
68842c9b | 2228 | } |
81a6fcae JK |
2229 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2230 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2231 | |
90481622 | 2232 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2233 | |
d85f69b0 MK |
2234 | map_commit = vma_commit_reservation(h, vma, addr); |
2235 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2236 | /* |
2237 | * The page was added to the reservation map between | |
2238 | * vma_needs_reservation and vma_commit_reservation. | |
2239 | * This indicates a race with hugetlb_reserve_pages. | |
2240 | * Adjust for the subpool count incremented above AND | |
2241 | * in hugetlb_reserve_pages for the same page. Also, | |
2242 | * the reservation count added in hugetlb_reserve_pages | |
2243 | * no longer applies. | |
2244 | */ | |
2245 | long rsv_adjust; | |
2246 | ||
2247 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2248 | hugetlb_acct_memory(h, -rsv_adjust); | |
2249 | } | |
90d8b7e6 | 2250 | return page; |
8f34af6f JZ |
2251 | |
2252 | out_uncharge_cgroup: | |
2253 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2254 | out_subpool_put: | |
d85f69b0 | 2255 | if (map_chg || avoid_reserve) |
8f34af6f | 2256 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2257 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2258 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2259 | } |
2260 | ||
e24a1307 AK |
2261 | int alloc_bootmem_huge_page(struct hstate *h) |
2262 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2263 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2264 | { |
2265 | struct huge_bootmem_page *m; | |
b2261026 | 2266 | int nr_nodes, node; |
aa888a74 | 2267 | |
b2261026 | 2268 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2269 | void *addr; |
2270 | ||
eb31d559 | 2271 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2272 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2273 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2274 | if (addr) { |
2275 | /* | |
2276 | * Use the beginning of the huge page to store the | |
2277 | * huge_bootmem_page struct (until gather_bootmem | |
2278 | * puts them into the mem_map). | |
2279 | */ | |
2280 | m = addr; | |
91f47662 | 2281 | goto found; |
aa888a74 | 2282 | } |
aa888a74 AK |
2283 | } |
2284 | return 0; | |
2285 | ||
2286 | found: | |
df994ead | 2287 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2288 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2289 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2290 | list_add(&m->list, &huge_boot_pages); |
2291 | m->hstate = h; | |
2292 | return 1; | |
2293 | } | |
2294 | ||
d00181b9 KS |
2295 | static void __init prep_compound_huge_page(struct page *page, |
2296 | unsigned int order) | |
18229df5 AW |
2297 | { |
2298 | if (unlikely(order > (MAX_ORDER - 1))) | |
2299 | prep_compound_gigantic_page(page, order); | |
2300 | else | |
2301 | prep_compound_page(page, order); | |
2302 | } | |
2303 | ||
aa888a74 AK |
2304 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2305 | static void __init gather_bootmem_prealloc(void) | |
2306 | { | |
2307 | struct huge_bootmem_page *m; | |
2308 | ||
2309 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2310 | struct page *page = virt_to_page(m); |
aa888a74 | 2311 | struct hstate *h = m->hstate; |
ee8f248d | 2312 | |
aa888a74 | 2313 | WARN_ON(page_count(page) != 1); |
18229df5 | 2314 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2315 | WARN_ON(PageReserved(page)); |
aa888a74 | 2316 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2317 | put_page(page); /* free it into the hugepage allocator */ |
2318 | ||
b0320c7b RA |
2319 | /* |
2320 | * If we had gigantic hugepages allocated at boot time, we need | |
2321 | * to restore the 'stolen' pages to totalram_pages in order to | |
2322 | * fix confusing memory reports from free(1) and another | |
2323 | * side-effects, like CommitLimit going negative. | |
2324 | */ | |
bae7f4ae | 2325 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2326 | adjust_managed_page_count(page, 1 << h->order); |
520495fe | 2327 | cond_resched(); |
aa888a74 AK |
2328 | } |
2329 | } | |
2330 | ||
8faa8b07 | 2331 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2332 | { |
2333 | unsigned long i; | |
f60858f9 MK |
2334 | nodemask_t *node_alloc_noretry; |
2335 | ||
2336 | if (!hstate_is_gigantic(h)) { | |
2337 | /* | |
2338 | * Bit mask controlling how hard we retry per-node allocations. | |
2339 | * Ignore errors as lower level routines can deal with | |
2340 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2341 | * time, we are likely in bigger trouble. | |
2342 | */ | |
2343 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2344 | GFP_KERNEL); | |
2345 | } else { | |
2346 | /* allocations done at boot time */ | |
2347 | node_alloc_noretry = NULL; | |
2348 | } | |
2349 | ||
2350 | /* bit mask controlling how hard we retry per-node allocations */ | |
2351 | if (node_alloc_noretry) | |
2352 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2353 | |
e5ff2159 | 2354 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2355 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2356 | if (!alloc_bootmem_huge_page(h)) |
2357 | break; | |
0c397dae | 2358 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2359 | &node_states[N_MEMORY], |
2360 | node_alloc_noretry)) | |
1da177e4 | 2361 | break; |
69ed779a | 2362 | cond_resched(); |
1da177e4 | 2363 | } |
d715cf80 LH |
2364 | if (i < h->max_huge_pages) { |
2365 | char buf[32]; | |
2366 | ||
c6247f72 | 2367 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2368 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2369 | h->max_huge_pages, buf, i); | |
2370 | h->max_huge_pages = i; | |
2371 | } | |
f60858f9 MK |
2372 | |
2373 | kfree(node_alloc_noretry); | |
e5ff2159 AK |
2374 | } |
2375 | ||
2376 | static void __init hugetlb_init_hstates(void) | |
2377 | { | |
2378 | struct hstate *h; | |
2379 | ||
2380 | for_each_hstate(h) { | |
641844f5 NH |
2381 | if (minimum_order > huge_page_order(h)) |
2382 | minimum_order = huge_page_order(h); | |
2383 | ||
8faa8b07 | 2384 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2385 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2386 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2387 | } |
641844f5 | 2388 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2389 | } |
2390 | ||
2391 | static void __init report_hugepages(void) | |
2392 | { | |
2393 | struct hstate *h; | |
2394 | ||
2395 | for_each_hstate(h) { | |
4abd32db | 2396 | char buf[32]; |
c6247f72 MW |
2397 | |
2398 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2399 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2400 | buf, h->free_huge_pages); |
e5ff2159 AK |
2401 | } |
2402 | } | |
2403 | ||
1da177e4 | 2404 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2405 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2406 | nodemask_t *nodes_allowed) | |
1da177e4 | 2407 | { |
4415cc8d CL |
2408 | int i; |
2409 | ||
bae7f4ae | 2410 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2411 | return; |
2412 | ||
6ae11b27 | 2413 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2414 | struct page *page, *next; |
a5516438 AK |
2415 | struct list_head *freel = &h->hugepage_freelists[i]; |
2416 | list_for_each_entry_safe(page, next, freel, lru) { | |
2417 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2418 | return; |
1da177e4 LT |
2419 | if (PageHighMem(page)) |
2420 | continue; | |
2421 | list_del(&page->lru); | |
e5ff2159 | 2422 | update_and_free_page(h, page); |
a5516438 AK |
2423 | h->free_huge_pages--; |
2424 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2425 | } |
2426 | } | |
2427 | } | |
2428 | #else | |
6ae11b27 LS |
2429 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2430 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2431 | { |
2432 | } | |
2433 | #endif | |
2434 | ||
20a0307c WF |
2435 | /* |
2436 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2437 | * balanced by operating on them in a round-robin fashion. | |
2438 | * Returns 1 if an adjustment was made. | |
2439 | */ | |
6ae11b27 LS |
2440 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2441 | int delta) | |
20a0307c | 2442 | { |
b2261026 | 2443 | int nr_nodes, node; |
20a0307c WF |
2444 | |
2445 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2446 | |
b2261026 JK |
2447 | if (delta < 0) { |
2448 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2449 | if (h->surplus_huge_pages_node[node]) | |
2450 | goto found; | |
e8c5c824 | 2451 | } |
b2261026 JK |
2452 | } else { |
2453 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2454 | if (h->surplus_huge_pages_node[node] < | |
2455 | h->nr_huge_pages_node[node]) | |
2456 | goto found; | |
e8c5c824 | 2457 | } |
b2261026 JK |
2458 | } |
2459 | return 0; | |
20a0307c | 2460 | |
b2261026 JK |
2461 | found: |
2462 | h->surplus_huge_pages += delta; | |
2463 | h->surplus_huge_pages_node[node] += delta; | |
2464 | return 1; | |
20a0307c WF |
2465 | } |
2466 | ||
a5516438 | 2467 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2468 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2469 | nodemask_t *nodes_allowed) |
1da177e4 | 2470 | { |
7893d1d5 | 2471 | unsigned long min_count, ret; |
f60858f9 MK |
2472 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2473 | ||
2474 | /* | |
2475 | * Bit mask controlling how hard we retry per-node allocations. | |
2476 | * If we can not allocate the bit mask, do not attempt to allocate | |
2477 | * the requested huge pages. | |
2478 | */ | |
2479 | if (node_alloc_noretry) | |
2480 | nodes_clear(*node_alloc_noretry); | |
2481 | else | |
2482 | return -ENOMEM; | |
1da177e4 | 2483 | |
4eb0716e AG |
2484 | spin_lock(&hugetlb_lock); |
2485 | ||
fd875dca MK |
2486 | /* |
2487 | * Check for a node specific request. | |
2488 | * Changing node specific huge page count may require a corresponding | |
2489 | * change to the global count. In any case, the passed node mask | |
2490 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2491 | */ | |
2492 | if (nid != NUMA_NO_NODE) { | |
2493 | unsigned long old_count = count; | |
2494 | ||
2495 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2496 | /* | |
2497 | * User may have specified a large count value which caused the | |
2498 | * above calculation to overflow. In this case, they wanted | |
2499 | * to allocate as many huge pages as possible. Set count to | |
2500 | * largest possible value to align with their intention. | |
2501 | */ | |
2502 | if (count < old_count) | |
2503 | count = ULONG_MAX; | |
2504 | } | |
2505 | ||
4eb0716e AG |
2506 | /* |
2507 | * Gigantic pages runtime allocation depend on the capability for large | |
2508 | * page range allocation. | |
2509 | * If the system does not provide this feature, return an error when | |
2510 | * the user tries to allocate gigantic pages but let the user free the | |
2511 | * boottime allocated gigantic pages. | |
2512 | */ | |
2513 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2514 | if (count > persistent_huge_pages(h)) { | |
2515 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2516 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2517 | return -EINVAL; |
2518 | } | |
2519 | /* Fall through to decrease pool */ | |
2520 | } | |
aa888a74 | 2521 | |
7893d1d5 AL |
2522 | /* |
2523 | * Increase the pool size | |
2524 | * First take pages out of surplus state. Then make up the | |
2525 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2526 | * |
0c397dae | 2527 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2528 | * to convert a surplus huge page to a normal huge page. That is |
2529 | * not critical, though, it just means the overall size of the | |
2530 | * pool might be one hugepage larger than it needs to be, but | |
2531 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2532 | */ |
a5516438 | 2533 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2534 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2535 | break; |
2536 | } | |
2537 | ||
a5516438 | 2538 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2539 | /* |
2540 | * If this allocation races such that we no longer need the | |
2541 | * page, free_huge_page will handle it by freeing the page | |
2542 | * and reducing the surplus. | |
2543 | */ | |
2544 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2545 | |
2546 | /* yield cpu to avoid soft lockup */ | |
2547 | cond_resched(); | |
2548 | ||
f60858f9 MK |
2549 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2550 | node_alloc_noretry); | |
7893d1d5 AL |
2551 | spin_lock(&hugetlb_lock); |
2552 | if (!ret) | |
2553 | goto out; | |
2554 | ||
536240f2 MG |
2555 | /* Bail for signals. Probably ctrl-c from user */ |
2556 | if (signal_pending(current)) | |
2557 | goto out; | |
7893d1d5 | 2558 | } |
7893d1d5 AL |
2559 | |
2560 | /* | |
2561 | * Decrease the pool size | |
2562 | * First return free pages to the buddy allocator (being careful | |
2563 | * to keep enough around to satisfy reservations). Then place | |
2564 | * pages into surplus state as needed so the pool will shrink | |
2565 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2566 | * |
2567 | * By placing pages into the surplus state independent of the | |
2568 | * overcommit value, we are allowing the surplus pool size to | |
2569 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2570 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2571 | * though, we'll note that we're not allowed to exceed surplus |
2572 | * and won't grow the pool anywhere else. Not until one of the | |
2573 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2574 | */ |
a5516438 | 2575 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2576 | min_count = max(count, min_count); |
6ae11b27 | 2577 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2578 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2579 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2580 | break; |
55f67141 | 2581 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2582 | } |
a5516438 | 2583 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2584 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2585 | break; |
2586 | } | |
2587 | out: | |
4eb0716e | 2588 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2589 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2590 | |
f60858f9 MK |
2591 | NODEMASK_FREE(node_alloc_noretry); |
2592 | ||
4eb0716e | 2593 | return 0; |
1da177e4 LT |
2594 | } |
2595 | ||
a3437870 NA |
2596 | #define HSTATE_ATTR_RO(_name) \ |
2597 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2598 | ||
2599 | #define HSTATE_ATTR(_name) \ | |
2600 | static struct kobj_attribute _name##_attr = \ | |
2601 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2602 | ||
2603 | static struct kobject *hugepages_kobj; | |
2604 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2605 | ||
9a305230 LS |
2606 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2607 | ||
2608 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2609 | { |
2610 | int i; | |
9a305230 | 2611 | |
a3437870 | 2612 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2613 | if (hstate_kobjs[i] == kobj) { |
2614 | if (nidp) | |
2615 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2616 | return &hstates[i]; |
9a305230 LS |
2617 | } |
2618 | ||
2619 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2620 | } |
2621 | ||
06808b08 | 2622 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2623 | struct kobj_attribute *attr, char *buf) |
2624 | { | |
9a305230 LS |
2625 | struct hstate *h; |
2626 | unsigned long nr_huge_pages; | |
2627 | int nid; | |
2628 | ||
2629 | h = kobj_to_hstate(kobj, &nid); | |
2630 | if (nid == NUMA_NO_NODE) | |
2631 | nr_huge_pages = h->nr_huge_pages; | |
2632 | else | |
2633 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2634 | ||
2635 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2636 | } |
adbe8726 | 2637 | |
238d3c13 DR |
2638 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2639 | struct hstate *h, int nid, | |
2640 | unsigned long count, size_t len) | |
a3437870 NA |
2641 | { |
2642 | int err; | |
2d0adf7e | 2643 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2644 | |
2d0adf7e OS |
2645 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2646 | return -EINVAL; | |
adbe8726 | 2647 | |
9a305230 LS |
2648 | if (nid == NUMA_NO_NODE) { |
2649 | /* | |
2650 | * global hstate attribute | |
2651 | */ | |
2652 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2653 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2654 | n_mask = &node_states[N_MEMORY]; | |
2655 | else | |
2656 | n_mask = &nodes_allowed; | |
2657 | } else { | |
9a305230 | 2658 | /* |
fd875dca MK |
2659 | * Node specific request. count adjustment happens in |
2660 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2661 | */ |
2d0adf7e OS |
2662 | init_nodemask_of_node(&nodes_allowed, nid); |
2663 | n_mask = &nodes_allowed; | |
fd875dca | 2664 | } |
9a305230 | 2665 | |
2d0adf7e | 2666 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2667 | |
4eb0716e | 2668 | return err ? err : len; |
06808b08 LS |
2669 | } |
2670 | ||
238d3c13 DR |
2671 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2672 | struct kobject *kobj, const char *buf, | |
2673 | size_t len) | |
2674 | { | |
2675 | struct hstate *h; | |
2676 | unsigned long count; | |
2677 | int nid; | |
2678 | int err; | |
2679 | ||
2680 | err = kstrtoul(buf, 10, &count); | |
2681 | if (err) | |
2682 | return err; | |
2683 | ||
2684 | h = kobj_to_hstate(kobj, &nid); | |
2685 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2686 | } | |
2687 | ||
06808b08 LS |
2688 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2689 | struct kobj_attribute *attr, char *buf) | |
2690 | { | |
2691 | return nr_hugepages_show_common(kobj, attr, buf); | |
2692 | } | |
2693 | ||
2694 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2695 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2696 | { | |
238d3c13 | 2697 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2698 | } |
2699 | HSTATE_ATTR(nr_hugepages); | |
2700 | ||
06808b08 LS |
2701 | #ifdef CONFIG_NUMA |
2702 | ||
2703 | /* | |
2704 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2705 | * huge page alloc/free. | |
2706 | */ | |
2707 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2708 | struct kobj_attribute *attr, char *buf) | |
2709 | { | |
2710 | return nr_hugepages_show_common(kobj, attr, buf); | |
2711 | } | |
2712 | ||
2713 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2714 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2715 | { | |
238d3c13 | 2716 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2717 | } |
2718 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2719 | #endif | |
2720 | ||
2721 | ||
a3437870 NA |
2722 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2723 | struct kobj_attribute *attr, char *buf) | |
2724 | { | |
9a305230 | 2725 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2726 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2727 | } | |
adbe8726 | 2728 | |
a3437870 NA |
2729 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2730 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2731 | { | |
2732 | int err; | |
2733 | unsigned long input; | |
9a305230 | 2734 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2735 | |
bae7f4ae | 2736 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2737 | return -EINVAL; |
2738 | ||
3dbb95f7 | 2739 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2740 | if (err) |
73ae31e5 | 2741 | return err; |
a3437870 NA |
2742 | |
2743 | spin_lock(&hugetlb_lock); | |
2744 | h->nr_overcommit_huge_pages = input; | |
2745 | spin_unlock(&hugetlb_lock); | |
2746 | ||
2747 | return count; | |
2748 | } | |
2749 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2750 | ||
2751 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2752 | struct kobj_attribute *attr, char *buf) | |
2753 | { | |
9a305230 LS |
2754 | struct hstate *h; |
2755 | unsigned long free_huge_pages; | |
2756 | int nid; | |
2757 | ||
2758 | h = kobj_to_hstate(kobj, &nid); | |
2759 | if (nid == NUMA_NO_NODE) | |
2760 | free_huge_pages = h->free_huge_pages; | |
2761 | else | |
2762 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2763 | ||
2764 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2765 | } |
2766 | HSTATE_ATTR_RO(free_hugepages); | |
2767 | ||
2768 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2769 | struct kobj_attribute *attr, char *buf) | |
2770 | { | |
9a305230 | 2771 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2772 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2773 | } | |
2774 | HSTATE_ATTR_RO(resv_hugepages); | |
2775 | ||
2776 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2777 | struct kobj_attribute *attr, char *buf) | |
2778 | { | |
9a305230 LS |
2779 | struct hstate *h; |
2780 | unsigned long surplus_huge_pages; | |
2781 | int nid; | |
2782 | ||
2783 | h = kobj_to_hstate(kobj, &nid); | |
2784 | if (nid == NUMA_NO_NODE) | |
2785 | surplus_huge_pages = h->surplus_huge_pages; | |
2786 | else | |
2787 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2788 | ||
2789 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2790 | } |
2791 | HSTATE_ATTR_RO(surplus_hugepages); | |
2792 | ||
2793 | static struct attribute *hstate_attrs[] = { | |
2794 | &nr_hugepages_attr.attr, | |
2795 | &nr_overcommit_hugepages_attr.attr, | |
2796 | &free_hugepages_attr.attr, | |
2797 | &resv_hugepages_attr.attr, | |
2798 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2799 | #ifdef CONFIG_NUMA |
2800 | &nr_hugepages_mempolicy_attr.attr, | |
2801 | #endif | |
a3437870 NA |
2802 | NULL, |
2803 | }; | |
2804 | ||
67e5ed96 | 2805 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2806 | .attrs = hstate_attrs, |
2807 | }; | |
2808 | ||
094e9539 JM |
2809 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2810 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2811 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2812 | { |
2813 | int retval; | |
972dc4de | 2814 | int hi = hstate_index(h); |
a3437870 | 2815 | |
9a305230 LS |
2816 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2817 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2818 | return -ENOMEM; |
2819 | ||
9a305230 | 2820 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
62fd7270 | 2821 | if (retval) { |
9a305230 | 2822 | kobject_put(hstate_kobjs[hi]); |
62fd7270 ML |
2823 | hstate_kobjs[hi] = NULL; |
2824 | } | |
a3437870 NA |
2825 | |
2826 | return retval; | |
2827 | } | |
2828 | ||
2829 | static void __init hugetlb_sysfs_init(void) | |
2830 | { | |
2831 | struct hstate *h; | |
2832 | int err; | |
2833 | ||
2834 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2835 | if (!hugepages_kobj) | |
2836 | return; | |
2837 | ||
2838 | for_each_hstate(h) { | |
9a305230 LS |
2839 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2840 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2841 | if (err) |
ffb22af5 | 2842 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2843 | } |
2844 | } | |
2845 | ||
9a305230 LS |
2846 | #ifdef CONFIG_NUMA |
2847 | ||
2848 | /* | |
2849 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2850 | * with node devices in node_devices[] using a parallel array. The array |
2851 | * index of a node device or _hstate == node id. | |
2852 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2853 | * the base kernel, on the hugetlb module. |
2854 | */ | |
2855 | struct node_hstate { | |
2856 | struct kobject *hugepages_kobj; | |
2857 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2858 | }; | |
b4e289a6 | 2859 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2860 | |
2861 | /* | |
10fbcf4c | 2862 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2863 | */ |
2864 | static struct attribute *per_node_hstate_attrs[] = { | |
2865 | &nr_hugepages_attr.attr, | |
2866 | &free_hugepages_attr.attr, | |
2867 | &surplus_hugepages_attr.attr, | |
2868 | NULL, | |
2869 | }; | |
2870 | ||
67e5ed96 | 2871 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
2872 | .attrs = per_node_hstate_attrs, |
2873 | }; | |
2874 | ||
2875 | /* | |
10fbcf4c | 2876 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2877 | * Returns node id via non-NULL nidp. |
2878 | */ | |
2879 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2880 | { | |
2881 | int nid; | |
2882 | ||
2883 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2884 | struct node_hstate *nhs = &node_hstates[nid]; | |
2885 | int i; | |
2886 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2887 | if (nhs->hstate_kobjs[i] == kobj) { | |
2888 | if (nidp) | |
2889 | *nidp = nid; | |
2890 | return &hstates[i]; | |
2891 | } | |
2892 | } | |
2893 | ||
2894 | BUG(); | |
2895 | return NULL; | |
2896 | } | |
2897 | ||
2898 | /* | |
10fbcf4c | 2899 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2900 | * No-op if no hstate attributes attached. |
2901 | */ | |
3cd8b44f | 2902 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2903 | { |
2904 | struct hstate *h; | |
10fbcf4c | 2905 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2906 | |
2907 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2908 | return; /* no hstate attributes */ |
9a305230 | 2909 | |
972dc4de AK |
2910 | for_each_hstate(h) { |
2911 | int idx = hstate_index(h); | |
2912 | if (nhs->hstate_kobjs[idx]) { | |
2913 | kobject_put(nhs->hstate_kobjs[idx]); | |
2914 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2915 | } |
972dc4de | 2916 | } |
9a305230 LS |
2917 | |
2918 | kobject_put(nhs->hugepages_kobj); | |
2919 | nhs->hugepages_kobj = NULL; | |
2920 | } | |
2921 | ||
9a305230 LS |
2922 | |
2923 | /* | |
10fbcf4c | 2924 | * Register hstate attributes for a single node device. |
9a305230 LS |
2925 | * No-op if attributes already registered. |
2926 | */ | |
3cd8b44f | 2927 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2928 | { |
2929 | struct hstate *h; | |
10fbcf4c | 2930 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2931 | int err; |
2932 | ||
2933 | if (nhs->hugepages_kobj) | |
2934 | return; /* already allocated */ | |
2935 | ||
2936 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2937 | &node->dev.kobj); |
9a305230 LS |
2938 | if (!nhs->hugepages_kobj) |
2939 | return; | |
2940 | ||
2941 | for_each_hstate(h) { | |
2942 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2943 | nhs->hstate_kobjs, | |
2944 | &per_node_hstate_attr_group); | |
2945 | if (err) { | |
ffb22af5 AM |
2946 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2947 | h->name, node->dev.id); | |
9a305230 LS |
2948 | hugetlb_unregister_node(node); |
2949 | break; | |
2950 | } | |
2951 | } | |
2952 | } | |
2953 | ||
2954 | /* | |
9b5e5d0f | 2955 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2956 | * devices of nodes that have memory. All on-line nodes should have |
2957 | * registered their associated device by this time. | |
9a305230 | 2958 | */ |
7d9ca000 | 2959 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2960 | { |
2961 | int nid; | |
2962 | ||
8cebfcd0 | 2963 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2964 | struct node *node = node_devices[nid]; |
10fbcf4c | 2965 | if (node->dev.id == nid) |
9a305230 LS |
2966 | hugetlb_register_node(node); |
2967 | } | |
2968 | ||
2969 | /* | |
10fbcf4c | 2970 | * Let the node device driver know we're here so it can |
9a305230 LS |
2971 | * [un]register hstate attributes on node hotplug. |
2972 | */ | |
2973 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2974 | hugetlb_unregister_node); | |
2975 | } | |
2976 | #else /* !CONFIG_NUMA */ | |
2977 | ||
2978 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2979 | { | |
2980 | BUG(); | |
2981 | if (nidp) | |
2982 | *nidp = -1; | |
2983 | return NULL; | |
2984 | } | |
2985 | ||
9a305230 LS |
2986 | static void hugetlb_register_all_nodes(void) { } |
2987 | ||
2988 | #endif | |
2989 | ||
a3437870 NA |
2990 | static int __init hugetlb_init(void) |
2991 | { | |
8382d914 DB |
2992 | int i; |
2993 | ||
457c1b27 | 2994 | if (!hugepages_supported()) |
0ef89d25 | 2995 | return 0; |
a3437870 | 2996 | |
e11bfbfc | 2997 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
2998 | if (default_hstate_size != 0) { |
2999 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
3000 | default_hstate_size, HPAGE_SIZE); | |
3001 | } | |
3002 | ||
e11bfbfc NP |
3003 | default_hstate_size = HPAGE_SIZE; |
3004 | if (!size_to_hstate(default_hstate_size)) | |
3005 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 3006 | } |
972dc4de | 3007 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
3008 | if (default_hstate_max_huge_pages) { |
3009 | if (!default_hstate.max_huge_pages) | |
3010 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3011 | } | |
a3437870 NA |
3012 | |
3013 | hugetlb_init_hstates(); | |
aa888a74 | 3014 | gather_bootmem_prealloc(); |
a3437870 NA |
3015 | report_hugepages(); |
3016 | ||
3017 | hugetlb_sysfs_init(); | |
9a305230 | 3018 | hugetlb_register_all_nodes(); |
7179e7bf | 3019 | hugetlb_cgroup_file_init(); |
9a305230 | 3020 | |
8382d914 DB |
3021 | #ifdef CONFIG_SMP |
3022 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3023 | #else | |
3024 | num_fault_mutexes = 1; | |
3025 | #endif | |
c672c7f2 | 3026 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3027 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3028 | GFP_KERNEL); | |
c672c7f2 | 3029 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3030 | |
3031 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3032 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3033 | return 0; |
3034 | } | |
3e89e1c5 | 3035 | subsys_initcall(hugetlb_init); |
a3437870 NA |
3036 | |
3037 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
3038 | void __init hugetlb_bad_size(void) |
3039 | { | |
3040 | parsed_valid_hugepagesz = false; | |
3041 | } | |
3042 | ||
d00181b9 | 3043 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3044 | { |
3045 | struct hstate *h; | |
8faa8b07 AK |
3046 | unsigned long i; |
3047 | ||
a3437870 | 3048 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 3049 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
3050 | return; |
3051 | } | |
47d38344 | 3052 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3053 | BUG_ON(order == 0); |
47d38344 | 3054 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
3055 | h->order = order; |
3056 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
3057 | h->nr_huge_pages = 0; |
3058 | h->free_huge_pages = 0; | |
3059 | for (i = 0; i < MAX_NUMNODES; ++i) | |
3060 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3061 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3062 | h->next_nid_to_alloc = first_memory_node; |
3063 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3064 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3065 | huge_page_size(h)/1024); | |
8faa8b07 | 3066 | |
a3437870 NA |
3067 | parsed_hstate = h; |
3068 | } | |
3069 | ||
e11bfbfc | 3070 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
3071 | { |
3072 | unsigned long *mhp; | |
8faa8b07 | 3073 | static unsigned long *last_mhp; |
a3437870 | 3074 | |
9fee021d VT |
3075 | if (!parsed_valid_hugepagesz) { |
3076 | pr_warn("hugepages = %s preceded by " | |
3077 | "an unsupported hugepagesz, ignoring\n", s); | |
3078 | parsed_valid_hugepagesz = true; | |
3079 | return 1; | |
3080 | } | |
a3437870 | 3081 | /* |
47d38344 | 3082 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
3083 | * so this hugepages= parameter goes to the "default hstate". |
3084 | */ | |
9fee021d | 3085 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3086 | mhp = &default_hstate_max_huge_pages; |
3087 | else | |
3088 | mhp = &parsed_hstate->max_huge_pages; | |
3089 | ||
8faa8b07 | 3090 | if (mhp == last_mhp) { |
598d8091 | 3091 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
3092 | return 1; |
3093 | } | |
3094 | ||
a3437870 NA |
3095 | if (sscanf(s, "%lu", mhp) <= 0) |
3096 | *mhp = 0; | |
3097 | ||
8faa8b07 AK |
3098 | /* |
3099 | * Global state is always initialized later in hugetlb_init. | |
3100 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
3101 | * use the bootmem allocator. | |
3102 | */ | |
47d38344 | 3103 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
3104 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3105 | ||
3106 | last_mhp = mhp; | |
3107 | ||
a3437870 NA |
3108 | return 1; |
3109 | } | |
e11bfbfc NP |
3110 | __setup("hugepages=", hugetlb_nrpages_setup); |
3111 | ||
3112 | static int __init hugetlb_default_setup(char *s) | |
3113 | { | |
3114 | default_hstate_size = memparse(s, &s); | |
3115 | return 1; | |
3116 | } | |
3117 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 3118 | |
8a213460 NA |
3119 | static unsigned int cpuset_mems_nr(unsigned int *array) |
3120 | { | |
3121 | int node; | |
3122 | unsigned int nr = 0; | |
3123 | ||
3124 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
3125 | nr += array[node]; | |
3126 | ||
3127 | return nr; | |
3128 | } | |
3129 | ||
3130 | #ifdef CONFIG_SYSCTL | |
5f84359e MS |
3131 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
3132 | void *buffer, size_t *length, | |
3133 | loff_t *ppos, unsigned long *out) | |
3134 | { | |
3135 | struct ctl_table dup_table; | |
3136 | ||
3137 | /* | |
3138 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
3139 | * can duplicate the @table and alter the duplicate of it. | |
3140 | */ | |
3141 | dup_table = *table; | |
3142 | dup_table.data = out; | |
3143 | ||
3144 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
3145 | } | |
3146 | ||
06808b08 LS |
3147 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3148 | struct ctl_table *table, int write, | |
3149 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 3150 | { |
e5ff2159 | 3151 | struct hstate *h = &default_hstate; |
238d3c13 | 3152 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3153 | int ret; |
e5ff2159 | 3154 | |
457c1b27 | 3155 | if (!hugepages_supported()) |
86613628 | 3156 | return -EOPNOTSUPP; |
457c1b27 | 3157 | |
5f84359e MS |
3158 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3159 | &tmp); | |
08d4a246 MH |
3160 | if (ret) |
3161 | goto out; | |
e5ff2159 | 3162 | |
238d3c13 DR |
3163 | if (write) |
3164 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3165 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3166 | out: |
3167 | return ret; | |
1da177e4 | 3168 | } |
396faf03 | 3169 | |
06808b08 LS |
3170 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
3171 | void __user *buffer, size_t *length, loff_t *ppos) | |
3172 | { | |
3173 | ||
3174 | return hugetlb_sysctl_handler_common(false, table, write, | |
3175 | buffer, length, ppos); | |
3176 | } | |
3177 | ||
3178 | #ifdef CONFIG_NUMA | |
3179 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
3180 | void __user *buffer, size_t *length, loff_t *ppos) | |
3181 | { | |
3182 | return hugetlb_sysctl_handler_common(true, table, write, | |
3183 | buffer, length, ppos); | |
3184 | } | |
3185 | #endif /* CONFIG_NUMA */ | |
3186 | ||
a3d0c6aa | 3187 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 3188 | void __user *buffer, |
a3d0c6aa NA |
3189 | size_t *length, loff_t *ppos) |
3190 | { | |
a5516438 | 3191 | struct hstate *h = &default_hstate; |
e5ff2159 | 3192 | unsigned long tmp; |
08d4a246 | 3193 | int ret; |
e5ff2159 | 3194 | |
457c1b27 | 3195 | if (!hugepages_supported()) |
86613628 | 3196 | return -EOPNOTSUPP; |
457c1b27 | 3197 | |
c033a93c | 3198 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3199 | |
bae7f4ae | 3200 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3201 | return -EINVAL; |
3202 | ||
5f84359e MS |
3203 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3204 | &tmp); | |
08d4a246 MH |
3205 | if (ret) |
3206 | goto out; | |
e5ff2159 AK |
3207 | |
3208 | if (write) { | |
3209 | spin_lock(&hugetlb_lock); | |
3210 | h->nr_overcommit_huge_pages = tmp; | |
3211 | spin_unlock(&hugetlb_lock); | |
3212 | } | |
08d4a246 MH |
3213 | out: |
3214 | return ret; | |
a3d0c6aa NA |
3215 | } |
3216 | ||
1da177e4 LT |
3217 | #endif /* CONFIG_SYSCTL */ |
3218 | ||
e1759c21 | 3219 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3220 | { |
fcb2b0c5 RG |
3221 | struct hstate *h; |
3222 | unsigned long total = 0; | |
3223 | ||
457c1b27 NA |
3224 | if (!hugepages_supported()) |
3225 | return; | |
fcb2b0c5 RG |
3226 | |
3227 | for_each_hstate(h) { | |
3228 | unsigned long count = h->nr_huge_pages; | |
3229 | ||
3230 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3231 | ||
3232 | if (h == &default_hstate) | |
3233 | seq_printf(m, | |
3234 | "HugePages_Total: %5lu\n" | |
3235 | "HugePages_Free: %5lu\n" | |
3236 | "HugePages_Rsvd: %5lu\n" | |
3237 | "HugePages_Surp: %5lu\n" | |
3238 | "Hugepagesize: %8lu kB\n", | |
3239 | count, | |
3240 | h->free_huge_pages, | |
3241 | h->resv_huge_pages, | |
3242 | h->surplus_huge_pages, | |
3243 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3244 | } | |
3245 | ||
3246 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3247 | } |
3248 | ||
3249 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3250 | { | |
a5516438 | 3251 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3252 | if (!hugepages_supported()) |
3253 | return 0; | |
1da177e4 LT |
3254 | return sprintf(buf, |
3255 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3256 | "Node %d HugePages_Free: %5u\n" |
3257 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3258 | nid, h->nr_huge_pages_node[nid], |
3259 | nid, h->free_huge_pages_node[nid], | |
3260 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3261 | } |
3262 | ||
949f7ec5 DR |
3263 | void hugetlb_show_meminfo(void) |
3264 | { | |
3265 | struct hstate *h; | |
3266 | int nid; | |
3267 | ||
457c1b27 NA |
3268 | if (!hugepages_supported()) |
3269 | return; | |
3270 | ||
949f7ec5 DR |
3271 | for_each_node_state(nid, N_MEMORY) |
3272 | for_each_hstate(h) | |
3273 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3274 | nid, | |
3275 | h->nr_huge_pages_node[nid], | |
3276 | h->free_huge_pages_node[nid], | |
3277 | h->surplus_huge_pages_node[nid], | |
3278 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3279 | } | |
3280 | ||
5d317b2b NH |
3281 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3282 | { | |
3283 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3284 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3285 | } | |
3286 | ||
1da177e4 LT |
3287 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3288 | unsigned long hugetlb_total_pages(void) | |
3289 | { | |
d0028588 WL |
3290 | struct hstate *h; |
3291 | unsigned long nr_total_pages = 0; | |
3292 | ||
3293 | for_each_hstate(h) | |
3294 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3295 | return nr_total_pages; | |
1da177e4 | 3296 | } |
1da177e4 | 3297 | |
a5516438 | 3298 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3299 | { |
3300 | int ret = -ENOMEM; | |
3301 | ||
3302 | spin_lock(&hugetlb_lock); | |
3303 | /* | |
3304 | * When cpuset is configured, it breaks the strict hugetlb page | |
3305 | * reservation as the accounting is done on a global variable. Such | |
3306 | * reservation is completely rubbish in the presence of cpuset because | |
3307 | * the reservation is not checked against page availability for the | |
3308 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3309 | * with lack of free htlb page in cpuset that the task is in. | |
3310 | * Attempt to enforce strict accounting with cpuset is almost | |
3311 | * impossible (or too ugly) because cpuset is too fluid that | |
3312 | * task or memory node can be dynamically moved between cpusets. | |
3313 | * | |
3314 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3315 | * undesirable. However, in order to preserve some of the semantics, | |
3316 | * we fall back to check against current free page availability as | |
3317 | * a best attempt and hopefully to minimize the impact of changing | |
3318 | * semantics that cpuset has. | |
3319 | */ | |
3320 | if (delta > 0) { | |
a5516438 | 3321 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3322 | goto out; |
3323 | ||
a5516438 AK |
3324 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3325 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3326 | goto out; |
3327 | } | |
3328 | } | |
3329 | ||
3330 | ret = 0; | |
3331 | if (delta < 0) | |
a5516438 | 3332 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3333 | |
3334 | out: | |
3335 | spin_unlock(&hugetlb_lock); | |
3336 | return ret; | |
3337 | } | |
3338 | ||
84afd99b AW |
3339 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3340 | { | |
f522c3ac | 3341 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3342 | |
3343 | /* | |
3344 | * This new VMA should share its siblings reservation map if present. | |
3345 | * The VMA will only ever have a valid reservation map pointer where | |
3346 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3347 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3348 | * after this open call completes. It is therefore safe to take a |
3349 | * new reference here without additional locking. | |
3350 | */ | |
4e35f483 | 3351 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3352 | kref_get(&resv->refs); |
84afd99b AW |
3353 | } |
3354 | ||
a1e78772 MG |
3355 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3356 | { | |
a5516438 | 3357 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3358 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3359 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3360 | unsigned long reserve, start, end; |
1c5ecae3 | 3361 | long gbl_reserve; |
84afd99b | 3362 | |
4e35f483 JK |
3363 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3364 | return; | |
84afd99b | 3365 | |
4e35f483 JK |
3366 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3367 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3368 | |
4e35f483 | 3369 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3370 | |
4e35f483 JK |
3371 | kref_put(&resv->refs, resv_map_release); |
3372 | ||
3373 | if (reserve) { | |
1c5ecae3 MK |
3374 | /* |
3375 | * Decrement reserve counts. The global reserve count may be | |
3376 | * adjusted if the subpool has a minimum size. | |
3377 | */ | |
3378 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3379 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3380 | } |
a1e78772 MG |
3381 | } |
3382 | ||
31383c68 DW |
3383 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3384 | { | |
3385 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3386 | return -EINVAL; | |
3387 | return 0; | |
3388 | } | |
3389 | ||
05ea8860 DW |
3390 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3391 | { | |
3392 | struct hstate *hstate = hstate_vma(vma); | |
3393 | ||
3394 | return 1UL << huge_page_shift(hstate); | |
3395 | } | |
3396 | ||
1da177e4 LT |
3397 | /* |
3398 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3399 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3400 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3401 | * this far. | |
3402 | */ | |
b3ec9f33 | 3403 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3404 | { |
3405 | BUG(); | |
d0217ac0 | 3406 | return 0; |
1da177e4 LT |
3407 | } |
3408 | ||
eec3636a JC |
3409 | /* |
3410 | * When a new function is introduced to vm_operations_struct and added | |
3411 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3412 | * This is because under System V memory model, mappings created via | |
3413 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3414 | * their original vm_ops are overwritten with shm_vm_ops. | |
3415 | */ | |
f0f37e2f | 3416 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3417 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3418 | .open = hugetlb_vm_op_open, |
a1e78772 | 3419 | .close = hugetlb_vm_op_close, |
31383c68 | 3420 | .split = hugetlb_vm_op_split, |
05ea8860 | 3421 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3422 | }; |
3423 | ||
1e8f889b DG |
3424 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3425 | int writable) | |
63551ae0 DG |
3426 | { |
3427 | pte_t entry; | |
3428 | ||
1e8f889b | 3429 | if (writable) { |
106c992a GS |
3430 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3431 | vma->vm_page_prot))); | |
63551ae0 | 3432 | } else { |
106c992a GS |
3433 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3434 | vma->vm_page_prot)); | |
63551ae0 DG |
3435 | } |
3436 | entry = pte_mkyoung(entry); | |
3437 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3438 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3439 | |
3440 | return entry; | |
3441 | } | |
3442 | ||
1e8f889b DG |
3443 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3444 | unsigned long address, pte_t *ptep) | |
3445 | { | |
3446 | pte_t entry; | |
3447 | ||
106c992a | 3448 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3449 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3450 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3451 | } |
3452 | ||
d5ed7444 | 3453 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3454 | { |
3455 | swp_entry_t swp; | |
3456 | ||
3457 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3458 | return false; |
4a705fef NH |
3459 | swp = pte_to_swp_entry(pte); |
3460 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3461 | return true; |
4a705fef | 3462 | else |
d5ed7444 | 3463 | return false; |
4a705fef NH |
3464 | } |
3465 | ||
3466 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3467 | { | |
3468 | swp_entry_t swp; | |
3469 | ||
3470 | if (huge_pte_none(pte) || pte_present(pte)) | |
3471 | return 0; | |
3472 | swp = pte_to_swp_entry(pte); | |
3473 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3474 | return 1; | |
3475 | else | |
3476 | return 0; | |
3477 | } | |
1e8f889b | 3478 | |
63551ae0 DG |
3479 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3480 | struct vm_area_struct *vma) | |
3481 | { | |
5e41540c | 3482 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3483 | struct page *ptepage; |
1c59827d | 3484 | unsigned long addr; |
1e8f889b | 3485 | int cow; |
a5516438 AK |
3486 | struct hstate *h = hstate_vma(vma); |
3487 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3488 | struct mmu_notifier_range range; |
e8569dd2 | 3489 | int ret = 0; |
1e8f889b DG |
3490 | |
3491 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3492 | |
ac46d4f3 | 3493 | if (cow) { |
7269f999 | 3494 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3495 | vma->vm_start, |
ac46d4f3 JG |
3496 | vma->vm_end); |
3497 | mmu_notifier_invalidate_range_start(&range); | |
3498 | } | |
e8569dd2 | 3499 | |
a5516438 | 3500 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3501 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3502 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3503 | if (!src_pte) |
3504 | continue; | |
a5516438 | 3505 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3506 | if (!dst_pte) { |
3507 | ret = -ENOMEM; | |
3508 | break; | |
3509 | } | |
c5c99429 | 3510 | |
5e41540c MK |
3511 | /* |
3512 | * If the pagetables are shared don't copy or take references. | |
3513 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3514 | * | |
3515 | * However, src could have 'unshared' and dst shares with | |
3516 | * another vma. If dst_pte !none, this implies sharing. | |
3517 | * Check here before taking page table lock, and once again | |
3518 | * after taking the lock below. | |
3519 | */ | |
3520 | dst_entry = huge_ptep_get(dst_pte); | |
3521 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3522 | continue; |
3523 | ||
cb900f41 KS |
3524 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3525 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3526 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3527 | entry = huge_ptep_get(src_pte); |
5e41540c MK |
3528 | dst_entry = huge_ptep_get(dst_pte); |
3529 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3530 | /* | |
3531 | * Skip if src entry none. Also, skip in the | |
3532 | * unlikely case dst entry !none as this implies | |
3533 | * sharing with another vma. | |
3534 | */ | |
4a705fef NH |
3535 | ; |
3536 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3537 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3538 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3539 | ||
3540 | if (is_write_migration_entry(swp_entry) && cow) { | |
3541 | /* | |
3542 | * COW mappings require pages in both | |
3543 | * parent and child to be set to read. | |
3544 | */ | |
3545 | make_migration_entry_read(&swp_entry); | |
3546 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3547 | set_huge_swap_pte_at(src, addr, src_pte, |
3548 | entry, sz); | |
4a705fef | 3549 | } |
e5251fd4 | 3550 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3551 | } else { |
34ee645e | 3552 | if (cow) { |
0f10851e JG |
3553 | /* |
3554 | * No need to notify as we are downgrading page | |
3555 | * table protection not changing it to point | |
3556 | * to a new page. | |
3557 | * | |
ad56b738 | 3558 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3559 | */ |
7f2e9525 | 3560 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3561 | } |
0253d634 | 3562 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3563 | ptepage = pte_page(entry); |
3564 | get_page(ptepage); | |
53f9263b | 3565 | page_dup_rmap(ptepage, true); |
1c59827d | 3566 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3567 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3568 | } |
cb900f41 KS |
3569 | spin_unlock(src_ptl); |
3570 | spin_unlock(dst_ptl); | |
63551ae0 | 3571 | } |
63551ae0 | 3572 | |
e8569dd2 | 3573 | if (cow) |
ac46d4f3 | 3574 | mmu_notifier_invalidate_range_end(&range); |
e8569dd2 AS |
3575 | |
3576 | return ret; | |
63551ae0 DG |
3577 | } |
3578 | ||
24669e58 AK |
3579 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3580 | unsigned long start, unsigned long end, | |
3581 | struct page *ref_page) | |
63551ae0 DG |
3582 | { |
3583 | struct mm_struct *mm = vma->vm_mm; | |
3584 | unsigned long address; | |
c7546f8f | 3585 | pte_t *ptep; |
63551ae0 | 3586 | pte_t pte; |
cb900f41 | 3587 | spinlock_t *ptl; |
63551ae0 | 3588 | struct page *page; |
a5516438 AK |
3589 | struct hstate *h = hstate_vma(vma); |
3590 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3591 | struct mmu_notifier_range range; |
bbe0decd | 3592 | bool force_flush = false; |
a5516438 | 3593 | |
63551ae0 | 3594 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3595 | BUG_ON(start & ~huge_page_mask(h)); |
3596 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3597 | |
07e32661 AK |
3598 | /* |
3599 | * This is a hugetlb vma, all the pte entries should point | |
3600 | * to huge page. | |
3601 | */ | |
ed6a7935 | 3602 | tlb_change_page_size(tlb, sz); |
24669e58 | 3603 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3604 | |
3605 | /* | |
3606 | * If sharing possible, alert mmu notifiers of worst case. | |
3607 | */ | |
6f4f13e8 JG |
3608 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3609 | end); | |
ac46d4f3 JG |
3610 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3611 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3612 | address = start; |
569f48b8 | 3613 | for (; address < end; address += sz) { |
7868a208 | 3614 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3615 | if (!ptep) |
c7546f8f DG |
3616 | continue; |
3617 | ||
cb900f41 | 3618 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3619 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3620 | spin_unlock(ptl); | |
bbe0decd NA |
3621 | tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); |
3622 | force_flush = true; | |
31d49da5 AK |
3623 | continue; |
3624 | } | |
39dde65c | 3625 | |
6629326b | 3626 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3627 | if (huge_pte_none(pte)) { |
3628 | spin_unlock(ptl); | |
3629 | continue; | |
3630 | } | |
6629326b HD |
3631 | |
3632 | /* | |
9fbc1f63 NH |
3633 | * Migrating hugepage or HWPoisoned hugepage is already |
3634 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3635 | */ |
9fbc1f63 | 3636 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3637 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3638 | spin_unlock(ptl); |
3639 | continue; | |
8c4894c6 | 3640 | } |
6629326b HD |
3641 | |
3642 | page = pte_page(pte); | |
04f2cbe3 MG |
3643 | /* |
3644 | * If a reference page is supplied, it is because a specific | |
3645 | * page is being unmapped, not a range. Ensure the page we | |
3646 | * are about to unmap is the actual page of interest. | |
3647 | */ | |
3648 | if (ref_page) { | |
31d49da5 AK |
3649 | if (page != ref_page) { |
3650 | spin_unlock(ptl); | |
3651 | continue; | |
3652 | } | |
04f2cbe3 MG |
3653 | /* |
3654 | * Mark the VMA as having unmapped its page so that | |
3655 | * future faults in this VMA will fail rather than | |
3656 | * looking like data was lost | |
3657 | */ | |
3658 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3659 | } | |
3660 | ||
c7546f8f | 3661 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3662 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3663 | if (huge_pte_dirty(pte)) |
6649a386 | 3664 | set_page_dirty(page); |
9e81130b | 3665 | |
5d317b2b | 3666 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3667 | page_remove_rmap(page, true); |
31d49da5 | 3668 | |
cb900f41 | 3669 | spin_unlock(ptl); |
e77b0852 | 3670 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3671 | /* |
3672 | * Bail out after unmapping reference page if supplied | |
3673 | */ | |
3674 | if (ref_page) | |
3675 | break; | |
fe1668ae | 3676 | } |
ac46d4f3 | 3677 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 3678 | tlb_end_vma(tlb, vma); |
bbe0decd NA |
3679 | |
3680 | /* | |
3681 | * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We | |
3682 | * could defer the flush until now, since by holding i_mmap_rwsem we | |
3683 | * guaranteed that the last refernece would not be dropped. But we must | |
3684 | * do the flushing before we return, as otherwise i_mmap_rwsem will be | |
3685 | * dropped and the last reference to the shared PMDs page might be | |
3686 | * dropped as well. | |
3687 | * | |
3688 | * In theory we could defer the freeing of the PMD pages as well, but | |
3689 | * huge_pmd_unshare() relies on the exact page_count for the PMD page to | |
3690 | * detect sharing, so we cannot defer the release of the page either. | |
3691 | * Instead, do flush now. | |
3692 | */ | |
3693 | if (force_flush) | |
3694 | tlb_flush_mmu_tlbonly(tlb); | |
1da177e4 | 3695 | } |
63551ae0 | 3696 | |
d833352a MG |
3697 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3698 | struct vm_area_struct *vma, unsigned long start, | |
3699 | unsigned long end, struct page *ref_page) | |
3700 | { | |
3701 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3702 | ||
3703 | /* | |
3704 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3705 | * test will fail on a vma being torn down, and not grab a page table | |
3706 | * on its way out. We're lucky that the flag has such an appropriate | |
3707 | * name, and can in fact be safely cleared here. We could clear it | |
3708 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3709 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3710 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3711 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3712 | */ |
3713 | vma->vm_flags &= ~VM_MAYSHARE; | |
3714 | } | |
3715 | ||
502717f4 | 3716 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3717 | unsigned long end, struct page *ref_page) |
502717f4 | 3718 | { |
24669e58 AK |
3719 | struct mm_struct *mm; |
3720 | struct mmu_gather tlb; | |
dff11abe MK |
3721 | unsigned long tlb_start = start; |
3722 | unsigned long tlb_end = end; | |
3723 | ||
3724 | /* | |
3725 | * If shared PMDs were possibly used within this vma range, adjust | |
3726 | * start/end for worst case tlb flushing. | |
3727 | * Note that we can not be sure if PMDs are shared until we try to | |
3728 | * unmap pages. However, we want to make sure TLB flushing covers | |
3729 | * the largest possible range. | |
3730 | */ | |
3731 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
24669e58 AK |
3732 | |
3733 | mm = vma->vm_mm; | |
3734 | ||
dff11abe | 3735 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); |
24669e58 | 3736 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
dff11abe | 3737 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); |
502717f4 KC |
3738 | } |
3739 | ||
04f2cbe3 MG |
3740 | /* |
3741 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3742 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3743 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3744 | * same region. | |
3745 | */ | |
2f4612af DB |
3746 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3747 | struct page *page, unsigned long address) | |
04f2cbe3 | 3748 | { |
7526674d | 3749 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3750 | struct vm_area_struct *iter_vma; |
3751 | struct address_space *mapping; | |
04f2cbe3 MG |
3752 | pgoff_t pgoff; |
3753 | ||
3754 | /* | |
3755 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3756 | * from page cache lookup which is in HPAGE_SIZE units. | |
3757 | */ | |
7526674d | 3758 | address = address & huge_page_mask(h); |
36e4f20a MH |
3759 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3760 | vma->vm_pgoff; | |
93c76a3d | 3761 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3762 | |
4eb2b1dc MG |
3763 | /* |
3764 | * Take the mapping lock for the duration of the table walk. As | |
3765 | * this mapping should be shared between all the VMAs, | |
3766 | * __unmap_hugepage_range() is called as the lock is already held | |
3767 | */ | |
83cde9e8 | 3768 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3769 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3770 | /* Do not unmap the current VMA */ |
3771 | if (iter_vma == vma) | |
3772 | continue; | |
3773 | ||
2f84a899 MG |
3774 | /* |
3775 | * Shared VMAs have their own reserves and do not affect | |
3776 | * MAP_PRIVATE accounting but it is possible that a shared | |
3777 | * VMA is using the same page so check and skip such VMAs. | |
3778 | */ | |
3779 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3780 | continue; | |
3781 | ||
04f2cbe3 MG |
3782 | /* |
3783 | * Unmap the page from other VMAs without their own reserves. | |
3784 | * They get marked to be SIGKILLed if they fault in these | |
3785 | * areas. This is because a future no-page fault on this VMA | |
3786 | * could insert a zeroed page instead of the data existing | |
3787 | * from the time of fork. This would look like data corruption | |
3788 | */ | |
3789 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3790 | unmap_hugepage_range(iter_vma, address, |
3791 | address + huge_page_size(h), page); | |
04f2cbe3 | 3792 | } |
83cde9e8 | 3793 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3794 | } |
3795 | ||
0fe6e20b NH |
3796 | /* |
3797 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3798 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3799 | * cannot race with other handlers or page migration. | |
3800 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3801 | */ |
2b740303 | 3802 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 3803 | unsigned long address, pte_t *ptep, |
3999f52e | 3804 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3805 | { |
3999f52e | 3806 | pte_t pte; |
a5516438 | 3807 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3808 | struct page *old_page, *new_page; |
2b740303 SJ |
3809 | int outside_reserve = 0; |
3810 | vm_fault_t ret = 0; | |
974e6d66 | 3811 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 3812 | struct mmu_notifier_range range; |
1e8f889b | 3813 | |
3999f52e | 3814 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3815 | old_page = pte_page(pte); |
3816 | ||
04f2cbe3 | 3817 | retry_avoidcopy: |
1e8f889b DG |
3818 | /* If no-one else is actually using this page, avoid the copy |
3819 | * and just make the page writable */ | |
37a2140d | 3820 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3821 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 3822 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 3823 | return 0; |
1e8f889b DG |
3824 | } |
3825 | ||
04f2cbe3 MG |
3826 | /* |
3827 | * If the process that created a MAP_PRIVATE mapping is about to | |
3828 | * perform a COW due to a shared page count, attempt to satisfy | |
3829 | * the allocation without using the existing reserves. The pagecache | |
3830 | * page is used to determine if the reserve at this address was | |
3831 | * consumed or not. If reserves were used, a partial faulted mapping | |
3832 | * at the time of fork() could consume its reserves on COW instead | |
3833 | * of the full address range. | |
3834 | */ | |
5944d011 | 3835 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3836 | old_page != pagecache_page) |
3837 | outside_reserve = 1; | |
3838 | ||
09cbfeaf | 3839 | get_page(old_page); |
b76c8cfb | 3840 | |
ad4404a2 DB |
3841 | /* |
3842 | * Drop page table lock as buddy allocator may be called. It will | |
3843 | * be acquired again before returning to the caller, as expected. | |
3844 | */ | |
cb900f41 | 3845 | spin_unlock(ptl); |
5b7a1d40 | 3846 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 3847 | |
2fc39cec | 3848 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3849 | /* |
3850 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3851 | * it is due to references held by a child and an insufficient | |
3852 | * huge page pool. To guarantee the original mappers | |
3853 | * reliability, unmap the page from child processes. The child | |
3854 | * may get SIGKILLed if it later faults. | |
3855 | */ | |
3856 | if (outside_reserve) { | |
09cbfeaf | 3857 | put_page(old_page); |
04f2cbe3 | 3858 | BUG_ON(huge_pte_none(pte)); |
5b7a1d40 | 3859 | unmap_ref_private(mm, vma, old_page, haddr); |
2f4612af DB |
3860 | BUG_ON(huge_pte_none(pte)); |
3861 | spin_lock(ptl); | |
5b7a1d40 | 3862 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
3863 | if (likely(ptep && |
3864 | pte_same(huge_ptep_get(ptep), pte))) | |
3865 | goto retry_avoidcopy; | |
3866 | /* | |
3867 | * race occurs while re-acquiring page table | |
3868 | * lock, and our job is done. | |
3869 | */ | |
3870 | return 0; | |
04f2cbe3 MG |
3871 | } |
3872 | ||
2b740303 | 3873 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 3874 | goto out_release_old; |
1e8f889b DG |
3875 | } |
3876 | ||
0fe6e20b NH |
3877 | /* |
3878 | * When the original hugepage is shared one, it does not have | |
3879 | * anon_vma prepared. | |
3880 | */ | |
44e2aa93 | 3881 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3882 | ret = VM_FAULT_OOM; |
3883 | goto out_release_all; | |
44e2aa93 | 3884 | } |
0fe6e20b | 3885 | |
974e6d66 | 3886 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 3887 | pages_per_huge_page(h)); |
0ed361de | 3888 | __SetPageUptodate(new_page); |
1e8f889b | 3889 | |
7269f999 | 3890 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 3891 | haddr + huge_page_size(h)); |
ac46d4f3 | 3892 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 3893 | |
b76c8cfb | 3894 | /* |
cb900f41 | 3895 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3896 | * before the page tables are altered |
3897 | */ | |
cb900f41 | 3898 | spin_lock(ptl); |
5b7a1d40 | 3899 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 3900 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3901 | ClearPagePrivate(new_page); |
3902 | ||
1e8f889b | 3903 | /* Break COW */ |
5b7a1d40 | 3904 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 3905 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 3906 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 3907 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 3908 | page_remove_rmap(old_page, true); |
5b7a1d40 | 3909 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
cb6acd01 | 3910 | set_page_huge_active(new_page); |
1e8f889b DG |
3911 | /* Make the old page be freed below */ |
3912 | new_page = old_page; | |
3913 | } | |
cb900f41 | 3914 | spin_unlock(ptl); |
ac46d4f3 | 3915 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 3916 | out_release_all: |
5b7a1d40 | 3917 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 3918 | put_page(new_page); |
ad4404a2 | 3919 | out_release_old: |
09cbfeaf | 3920 | put_page(old_page); |
8312034f | 3921 | |
ad4404a2 DB |
3922 | spin_lock(ptl); /* Caller expects lock to be held */ |
3923 | return ret; | |
1e8f889b DG |
3924 | } |
3925 | ||
04f2cbe3 | 3926 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3927 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3928 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3929 | { |
3930 | struct address_space *mapping; | |
e7c4b0bf | 3931 | pgoff_t idx; |
04f2cbe3 MG |
3932 | |
3933 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3934 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3935 | |
3936 | return find_lock_page(mapping, idx); | |
3937 | } | |
3938 | ||
3ae77f43 HD |
3939 | /* |
3940 | * Return whether there is a pagecache page to back given address within VMA. | |
3941 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3942 | */ | |
3943 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3944 | struct vm_area_struct *vma, unsigned long address) |
3945 | { | |
3946 | struct address_space *mapping; | |
3947 | pgoff_t idx; | |
3948 | struct page *page; | |
3949 | ||
3950 | mapping = vma->vm_file->f_mapping; | |
3951 | idx = vma_hugecache_offset(h, vma, address); | |
3952 | ||
3953 | page = find_get_page(mapping, idx); | |
3954 | if (page) | |
3955 | put_page(page); | |
3956 | return page != NULL; | |
3957 | } | |
3958 | ||
ab76ad54 MK |
3959 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3960 | pgoff_t idx) | |
3961 | { | |
3962 | struct inode *inode = mapping->host; | |
3963 | struct hstate *h = hstate_inode(inode); | |
3964 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3965 | ||
3966 | if (err) | |
3967 | return err; | |
3968 | ClearPagePrivate(page); | |
3969 | ||
22146c3c MK |
3970 | /* |
3971 | * set page dirty so that it will not be removed from cache/file | |
3972 | * by non-hugetlbfs specific code paths. | |
3973 | */ | |
3974 | set_page_dirty(page); | |
3975 | ||
ab76ad54 MK |
3976 | spin_lock(&inode->i_lock); |
3977 | inode->i_blocks += blocks_per_huge_page(h); | |
3978 | spin_unlock(&inode->i_lock); | |
3979 | return 0; | |
3980 | } | |
3981 | ||
2b740303 SJ |
3982 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
3983 | struct vm_area_struct *vma, | |
3984 | struct address_space *mapping, pgoff_t idx, | |
3985 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3986 | { |
a5516438 | 3987 | struct hstate *h = hstate_vma(vma); |
2b740303 | 3988 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3989 | int anon_rmap = 0; |
4c887265 | 3990 | unsigned long size; |
4c887265 | 3991 | struct page *page; |
1e8f889b | 3992 | pte_t new_pte; |
cb900f41 | 3993 | spinlock_t *ptl; |
285b8dca | 3994 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 3995 | bool new_page = false; |
4c887265 | 3996 | |
04f2cbe3 MG |
3997 | /* |
3998 | * Currently, we are forced to kill the process in the event the | |
3999 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4000 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4001 | */ |
4002 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4003 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4004 | current->pid); |
04f2cbe3 MG |
4005 | return ret; |
4006 | } | |
4007 | ||
4c887265 | 4008 | /* |
e7c58097 MK |
4009 | * Use page lock to guard against racing truncation |
4010 | * before we get page_table_lock. | |
4c887265 | 4011 | */ |
6bda666a CL |
4012 | retry: |
4013 | page = find_lock_page(mapping, idx); | |
4014 | if (!page) { | |
e7c58097 MK |
4015 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4016 | if (idx >= size) | |
4017 | goto out; | |
4018 | ||
1a1aad8a MK |
4019 | /* |
4020 | * Check for page in userfault range | |
4021 | */ | |
4022 | if (userfaultfd_missing(vma)) { | |
4023 | u32 hash; | |
4024 | struct vm_fault vmf = { | |
4025 | .vma = vma, | |
285b8dca | 4026 | .address = haddr, |
1a1aad8a MK |
4027 | .flags = flags, |
4028 | /* | |
4029 | * Hard to debug if it ends up being | |
4030 | * used by a callee that assumes | |
4031 | * something about the other | |
4032 | * uninitialized fields... same as in | |
4033 | * memory.c | |
4034 | */ | |
4035 | }; | |
4036 | ||
4037 | /* | |
ddeaab32 MK |
4038 | * hugetlb_fault_mutex must be dropped before |
4039 | * handling userfault. Reacquire after handling | |
4040 | * fault to make calling code simpler. | |
1a1aad8a | 4041 | */ |
5e3cb663 | 4042 | hash = hugetlb_fault_mutex_hash(h, mapping, idx); |
1a1aad8a MK |
4043 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
4044 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
4045 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
4046 | goto out; | |
4047 | } | |
4048 | ||
285b8dca | 4049 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4050 | if (IS_ERR(page)) { |
4643d67e MK |
4051 | /* |
4052 | * Returning error will result in faulting task being | |
4053 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4054 | * tasks from racing to fault in the same page which | |
4055 | * could result in false unable to allocate errors. | |
4056 | * Page migration does not take the fault mutex, but | |
4057 | * does a clear then write of pte's under page table | |
4058 | * lock. Page fault code could race with migration, | |
4059 | * notice the clear pte and try to allocate a page | |
4060 | * here. Before returning error, get ptl and make | |
4061 | * sure there really is no pte entry. | |
4062 | */ | |
4063 | ptl = huge_pte_lock(h, mm, ptep); | |
4064 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
4065 | ret = 0; | |
4066 | spin_unlock(ptl); | |
4067 | goto out; | |
4068 | } | |
4069 | spin_unlock(ptl); | |
2b740303 | 4070 | ret = vmf_error(PTR_ERR(page)); |
6bda666a CL |
4071 | goto out; |
4072 | } | |
47ad8475 | 4073 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4074 | __SetPageUptodate(page); |
cb6acd01 | 4075 | new_page = true; |
ac9b9c66 | 4076 | |
f83a275d | 4077 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4078 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4079 | if (err) { |
4080 | put_page(page); | |
6bda666a CL |
4081 | if (err == -EEXIST) |
4082 | goto retry; | |
4083 | goto out; | |
4084 | } | |
23be7468 | 4085 | } else { |
6bda666a | 4086 | lock_page(page); |
0fe6e20b NH |
4087 | if (unlikely(anon_vma_prepare(vma))) { |
4088 | ret = VM_FAULT_OOM; | |
4089 | goto backout_unlocked; | |
4090 | } | |
409eb8c2 | 4091 | anon_rmap = 1; |
23be7468 | 4092 | } |
0fe6e20b | 4093 | } else { |
998b4382 NH |
4094 | /* |
4095 | * If memory error occurs between mmap() and fault, some process | |
4096 | * don't have hwpoisoned swap entry for errored virtual address. | |
4097 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4098 | */ | |
4099 | if (unlikely(PageHWPoison(page))) { | |
61edc516 | 4100 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4101 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4102 | goto backout_unlocked; |
4103 | } | |
6bda666a | 4104 | } |
1e8f889b | 4105 | |
57303d80 AW |
4106 | /* |
4107 | * If we are going to COW a private mapping later, we examine the | |
4108 | * pending reservations for this page now. This will ensure that | |
4109 | * any allocations necessary to record that reservation occur outside | |
4110 | * the spinlock. | |
4111 | */ | |
5e911373 | 4112 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4113 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4114 | ret = VM_FAULT_OOM; |
4115 | goto backout_unlocked; | |
4116 | } | |
5e911373 | 4117 | /* Just decrements count, does not deallocate */ |
285b8dca | 4118 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4119 | } |
57303d80 | 4120 | |
8bea8052 | 4121 | ptl = huge_pte_lock(h, mm, ptep); |
e7c58097 MK |
4122 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4123 | if (idx >= size) | |
4124 | goto backout; | |
4c887265 | 4125 | |
83c54070 | 4126 | ret = 0; |
7f2e9525 | 4127 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4128 | goto backout; |
4129 | ||
07443a85 JK |
4130 | if (anon_rmap) { |
4131 | ClearPagePrivate(page); | |
285b8dca | 4132 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4133 | } else |
53f9263b | 4134 | page_dup_rmap(page, true); |
1e8f889b DG |
4135 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4136 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4137 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4138 | |
5d317b2b | 4139 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4140 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4141 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4142 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4143 | } |
4144 | ||
cb900f41 | 4145 | spin_unlock(ptl); |
cb6acd01 MK |
4146 | |
4147 | /* | |
4148 | * Only make newly allocated pages active. Existing pages found | |
4149 | * in the pagecache could be !page_huge_active() if they have been | |
4150 | * isolated for migration. | |
4151 | */ | |
4152 | if (new_page) | |
4153 | set_page_huge_active(page); | |
4154 | ||
4c887265 AL |
4155 | unlock_page(page); |
4156 | out: | |
ac9b9c66 | 4157 | return ret; |
4c887265 AL |
4158 | |
4159 | backout: | |
cb900f41 | 4160 | spin_unlock(ptl); |
2b26736c | 4161 | backout_unlocked: |
4c887265 | 4162 | unlock_page(page); |
285b8dca | 4163 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4164 | put_page(page); |
4165 | goto out; | |
ac9b9c66 HD |
4166 | } |
4167 | ||
8382d914 | 4168 | #ifdef CONFIG_SMP |
1b426bac | 4169 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping, |
5e3cb663 | 4170 | pgoff_t idx) |
8382d914 DB |
4171 | { |
4172 | unsigned long key[2]; | |
4173 | u32 hash; | |
4174 | ||
1b426bac MK |
4175 | key[0] = (unsigned long) mapping; |
4176 | key[1] = idx; | |
8382d914 | 4177 | |
5e3cb663 | 4178 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4179 | |
4180 | return hash & (num_fault_mutexes - 1); | |
4181 | } | |
4182 | #else | |
4183 | /* | |
4184 | * For uniprocesor systems we always use a single mutex, so just | |
4185 | * return 0 and avoid the hashing overhead. | |
4186 | */ | |
1b426bac | 4187 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct address_space *mapping, |
5e3cb663 | 4188 | pgoff_t idx) |
8382d914 DB |
4189 | { |
4190 | return 0; | |
4191 | } | |
4192 | #endif | |
4193 | ||
2b740303 | 4194 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4195 | unsigned long address, unsigned int flags) |
86e5216f | 4196 | { |
8382d914 | 4197 | pte_t *ptep, entry; |
cb900f41 | 4198 | spinlock_t *ptl; |
2b740303 | 4199 | vm_fault_t ret; |
8382d914 DB |
4200 | u32 hash; |
4201 | pgoff_t idx; | |
0fe6e20b | 4202 | struct page *page = NULL; |
57303d80 | 4203 | struct page *pagecache_page = NULL; |
a5516438 | 4204 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4205 | struct address_space *mapping; |
0f792cf9 | 4206 | int need_wait_lock = 0; |
285b8dca | 4207 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4208 | |
285b8dca | 4209 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed NH |
4210 | if (ptep) { |
4211 | entry = huge_ptep_get(ptep); | |
290408d4 | 4212 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4213 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4214 | return 0; |
4215 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4216 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4217 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
ddeaab32 MK |
4218 | } else { |
4219 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4220 | if (!ptep) | |
4221 | return VM_FAULT_OOM; | |
fd6a03ed NH |
4222 | } |
4223 | ||
8382d914 | 4224 | mapping = vma->vm_file->f_mapping; |
ddeaab32 | 4225 | idx = vma_hugecache_offset(h, vma, haddr); |
8382d914 | 4226 | |
3935baa9 DG |
4227 | /* |
4228 | * Serialize hugepage allocation and instantiation, so that we don't | |
4229 | * get spurious allocation failures if two CPUs race to instantiate | |
4230 | * the same page in the page cache. | |
4231 | */ | |
5e3cb663 | 4232 | hash = hugetlb_fault_mutex_hash(h, mapping, idx); |
c672c7f2 | 4233 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4234 | |
7f2e9525 GS |
4235 | entry = huge_ptep_get(ptep); |
4236 | if (huge_pte_none(entry)) { | |
8382d914 | 4237 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4238 | goto out_mutex; |
3935baa9 | 4239 | } |
86e5216f | 4240 | |
83c54070 | 4241 | ret = 0; |
1e8f889b | 4242 | |
0f792cf9 NH |
4243 | /* |
4244 | * entry could be a migration/hwpoison entry at this point, so this | |
4245 | * check prevents the kernel from going below assuming that we have | |
4246 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
4247 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
4248 | * handle it. | |
4249 | */ | |
4250 | if (!pte_present(entry)) | |
4251 | goto out_mutex; | |
4252 | ||
57303d80 AW |
4253 | /* |
4254 | * If we are going to COW the mapping later, we examine the pending | |
4255 | * reservations for this page now. This will ensure that any | |
4256 | * allocations necessary to record that reservation occur outside the | |
4257 | * spinlock. For private mappings, we also lookup the pagecache | |
4258 | * page now as it is used to determine if a reservation has been | |
4259 | * consumed. | |
4260 | */ | |
106c992a | 4261 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4262 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4263 | ret = VM_FAULT_OOM; |
b4d1d99f | 4264 | goto out_mutex; |
2b26736c | 4265 | } |
5e911373 | 4266 | /* Just decrements count, does not deallocate */ |
285b8dca | 4267 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4268 | |
f83a275d | 4269 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4270 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4271 | vma, haddr); |
57303d80 AW |
4272 | } |
4273 | ||
0f792cf9 NH |
4274 | ptl = huge_pte_lock(h, mm, ptep); |
4275 | ||
4276 | /* Check for a racing update before calling hugetlb_cow */ | |
4277 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4278 | goto out_ptl; | |
4279 | ||
56c9cfb1 NH |
4280 | /* |
4281 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4282 | * pagecache_page, so here we need take the former one | |
4283 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4284 | */ |
4285 | page = pte_page(entry); | |
4286 | if (page != pagecache_page) | |
0f792cf9 NH |
4287 | if (!trylock_page(page)) { |
4288 | need_wait_lock = 1; | |
4289 | goto out_ptl; | |
4290 | } | |
b4d1d99f | 4291 | |
0f792cf9 | 4292 | get_page(page); |
b4d1d99f | 4293 | |
788c7df4 | 4294 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4295 | if (!huge_pte_write(entry)) { |
974e6d66 | 4296 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4297 | pagecache_page, ptl); |
0f792cf9 | 4298 | goto out_put_page; |
b4d1d99f | 4299 | } |
106c992a | 4300 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4301 | } |
4302 | entry = pte_mkyoung(entry); | |
285b8dca | 4303 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4304 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4305 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4306 | out_put_page: |
4307 | if (page != pagecache_page) | |
4308 | unlock_page(page); | |
4309 | put_page(page); | |
cb900f41 KS |
4310 | out_ptl: |
4311 | spin_unlock(ptl); | |
57303d80 AW |
4312 | |
4313 | if (pagecache_page) { | |
4314 | unlock_page(pagecache_page); | |
4315 | put_page(pagecache_page); | |
4316 | } | |
b4d1d99f | 4317 | out_mutex: |
c672c7f2 | 4318 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
4319 | /* |
4320 | * Generally it's safe to hold refcount during waiting page lock. But | |
4321 | * here we just wait to defer the next page fault to avoid busy loop and | |
4322 | * the page is not used after unlocked before returning from the current | |
4323 | * page fault. So we are safe from accessing freed page, even if we wait | |
4324 | * here without taking refcount. | |
4325 | */ | |
4326 | if (need_wait_lock) | |
4327 | wait_on_page_locked(page); | |
1e8f889b | 4328 | return ret; |
86e5216f AL |
4329 | } |
4330 | ||
8fb5debc MK |
4331 | /* |
4332 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4333 | * modifications for huge pages. | |
4334 | */ | |
4335 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4336 | pte_t *dst_pte, | |
4337 | struct vm_area_struct *dst_vma, | |
4338 | unsigned long dst_addr, | |
4339 | unsigned long src_addr, | |
4340 | struct page **pagep) | |
4341 | { | |
1e392147 AA |
4342 | struct address_space *mapping; |
4343 | pgoff_t idx; | |
4344 | unsigned long size; | |
1c9e8def | 4345 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4346 | struct hstate *h = hstate_vma(dst_vma); |
4347 | pte_t _dst_pte; | |
4348 | spinlock_t *ptl; | |
4349 | int ret; | |
4350 | struct page *page; | |
4351 | ||
4352 | if (!*pagep) { | |
2925a98f MA |
4353 | /* If a page already exists, then it's UFFDIO_COPY for |
4354 | * a non-missing case. Return -EEXIST. | |
4355 | */ | |
4356 | if (vm_shared && | |
4357 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
4358 | ret = -EEXIST; | |
4359 | goto out; | |
4360 | } | |
4361 | ||
8fb5debc | 4362 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
2925a98f MA |
4363 | if (IS_ERR(page)) { |
4364 | ret = -ENOMEM; | |
8fb5debc | 4365 | goto out; |
2925a98f | 4366 | } |
8fb5debc MK |
4367 | |
4368 | ret = copy_huge_page_from_user(page, | |
4369 | (const void __user *) src_addr, | |
810a56b9 | 4370 | pages_per_huge_page(h), false); |
8fb5debc MK |
4371 | |
4372 | /* fallback to copy_from_user outside mmap_sem */ | |
4373 | if (unlikely(ret)) { | |
9e368259 | 4374 | ret = -ENOENT; |
8fb5debc MK |
4375 | *pagep = page; |
4376 | /* don't free the page */ | |
4377 | goto out; | |
4378 | } | |
4379 | } else { | |
4380 | page = *pagep; | |
4381 | *pagep = NULL; | |
4382 | } | |
4383 | ||
4384 | /* | |
4385 | * The memory barrier inside __SetPageUptodate makes sure that | |
4386 | * preceding stores to the page contents become visible before | |
4387 | * the set_pte_at() write. | |
4388 | */ | |
4389 | __SetPageUptodate(page); | |
8fb5debc | 4390 | |
1e392147 AA |
4391 | mapping = dst_vma->vm_file->f_mapping; |
4392 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4393 | ||
1c9e8def MK |
4394 | /* |
4395 | * If shared, add to page cache | |
4396 | */ | |
4397 | if (vm_shared) { | |
1e392147 AA |
4398 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4399 | ret = -EFAULT; | |
4400 | if (idx >= size) | |
4401 | goto out_release_nounlock; | |
1c9e8def | 4402 | |
1e392147 AA |
4403 | /* |
4404 | * Serialization between remove_inode_hugepages() and | |
4405 | * huge_add_to_page_cache() below happens through the | |
4406 | * hugetlb_fault_mutex_table that here must be hold by | |
4407 | * the caller. | |
4408 | */ | |
1c9e8def MK |
4409 | ret = huge_add_to_page_cache(page, mapping, idx); |
4410 | if (ret) | |
4411 | goto out_release_nounlock; | |
4412 | } | |
4413 | ||
8fb5debc MK |
4414 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4415 | spin_lock(ptl); | |
4416 | ||
1e392147 AA |
4417 | /* |
4418 | * Recheck the i_size after holding PT lock to make sure not | |
4419 | * to leave any page mapped (as page_mapped()) beyond the end | |
4420 | * of the i_size (remove_inode_hugepages() is strict about | |
4421 | * enforcing that). If we bail out here, we'll also leave a | |
4422 | * page in the radix tree in the vm_shared case beyond the end | |
4423 | * of the i_size, but remove_inode_hugepages() will take care | |
4424 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4425 | */ | |
4426 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4427 | ret = -EFAULT; | |
4428 | if (idx >= size) | |
4429 | goto out_release_unlock; | |
4430 | ||
8fb5debc MK |
4431 | ret = -EEXIST; |
4432 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4433 | goto out_release_unlock; | |
4434 | ||
1c9e8def MK |
4435 | if (vm_shared) { |
4436 | page_dup_rmap(page, true); | |
4437 | } else { | |
4438 | ClearPagePrivate(page); | |
4439 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4440 | } | |
8fb5debc MK |
4441 | |
4442 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4443 | if (dst_vma->vm_flags & VM_WRITE) | |
4444 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4445 | _dst_pte = pte_mkyoung(_dst_pte); | |
4446 | ||
4447 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4448 | ||
4449 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4450 | dst_vma->vm_flags & VM_WRITE); | |
4451 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4452 | ||
4453 | /* No need to invalidate - it was non-present before */ | |
4454 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4455 | ||
4456 | spin_unlock(ptl); | |
cb6acd01 | 4457 | set_page_huge_active(page); |
1c9e8def MK |
4458 | if (vm_shared) |
4459 | unlock_page(page); | |
8fb5debc MK |
4460 | ret = 0; |
4461 | out: | |
4462 | return ret; | |
4463 | out_release_unlock: | |
4464 | spin_unlock(ptl); | |
1c9e8def MK |
4465 | if (vm_shared) |
4466 | unlock_page(page); | |
5af10dfd | 4467 | out_release_nounlock: |
8fb5debc MK |
4468 | put_page(page); |
4469 | goto out; | |
4470 | } | |
4471 | ||
28a35716 ML |
4472 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4473 | struct page **pages, struct vm_area_struct **vmas, | |
4474 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4475 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4476 | { |
d5d4b0aa KC |
4477 | unsigned long pfn_offset; |
4478 | unsigned long vaddr = *position; | |
28a35716 | 4479 | unsigned long remainder = *nr_pages; |
a5516438 | 4480 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4481 | int err = -EFAULT; |
63551ae0 | 4482 | |
63551ae0 | 4483 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4484 | pte_t *pte; |
cb900f41 | 4485 | spinlock_t *ptl = NULL; |
2a15efc9 | 4486 | int absent; |
4c887265 | 4487 | struct page *page; |
63551ae0 | 4488 | |
02057967 DR |
4489 | /* |
4490 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4491 | * potentially allocating memory. | |
4492 | */ | |
fa45f116 | 4493 | if (fatal_signal_pending(current)) { |
02057967 DR |
4494 | remainder = 0; |
4495 | break; | |
4496 | } | |
4497 | ||
4c887265 AL |
4498 | /* |
4499 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4500 | * each hugepage. We have to make sure we get the |
4c887265 | 4501 | * first, for the page indexing below to work. |
cb900f41 KS |
4502 | * |
4503 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4504 | */ |
7868a208 PA |
4505 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4506 | huge_page_size(h)); | |
cb900f41 KS |
4507 | if (pte) |
4508 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4509 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4510 | ||
4511 | /* | |
4512 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4513 | * an error where there's an empty slot with no huge pagecache |
4514 | * to back it. This way, we avoid allocating a hugepage, and | |
4515 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4516 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4517 | */ |
3ae77f43 HD |
4518 | if (absent && (flags & FOLL_DUMP) && |
4519 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4520 | if (pte) |
4521 | spin_unlock(ptl); | |
2a15efc9 HD |
4522 | remainder = 0; |
4523 | break; | |
4524 | } | |
63551ae0 | 4525 | |
9cc3a5bd NH |
4526 | /* |
4527 | * We need call hugetlb_fault for both hugepages under migration | |
4528 | * (in which case hugetlb_fault waits for the migration,) and | |
4529 | * hwpoisoned hugepages (in which case we need to prevent the | |
4530 | * caller from accessing to them.) In order to do this, we use | |
4531 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4532 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4533 | * both cases, and because we can't follow correct pages | |
4534 | * directly from any kind of swap entries. | |
4535 | */ | |
4536 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4537 | ((flags & FOLL_WRITE) && |
4538 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4539 | vm_fault_t ret; |
87ffc118 | 4540 | unsigned int fault_flags = 0; |
63551ae0 | 4541 | |
cb900f41 KS |
4542 | if (pte) |
4543 | spin_unlock(ptl); | |
87ffc118 AA |
4544 | if (flags & FOLL_WRITE) |
4545 | fault_flags |= FAULT_FLAG_WRITE; | |
4546 | if (nonblocking) | |
4547 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4548 | if (flags & FOLL_NOWAIT) | |
4549 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4550 | FAULT_FLAG_RETRY_NOWAIT; | |
4551 | if (flags & FOLL_TRIED) { | |
4552 | VM_WARN_ON_ONCE(fault_flags & | |
4553 | FAULT_FLAG_ALLOW_RETRY); | |
4554 | fault_flags |= FAULT_FLAG_TRIED; | |
4555 | } | |
4556 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4557 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4558 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4559 | remainder = 0; |
4560 | break; | |
4561 | } | |
4562 | if (ret & VM_FAULT_RETRY) { | |
1ac25013 AA |
4563 | if (nonblocking && |
4564 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) | |
87ffc118 AA |
4565 | *nonblocking = 0; |
4566 | *nr_pages = 0; | |
4567 | /* | |
4568 | * VM_FAULT_RETRY must not return an | |
4569 | * error, it will return zero | |
4570 | * instead. | |
4571 | * | |
4572 | * No need to update "position" as the | |
4573 | * caller will not check it after | |
4574 | * *nr_pages is set to 0. | |
4575 | */ | |
4576 | return i; | |
4577 | } | |
4578 | continue; | |
4c887265 AL |
4579 | } |
4580 | ||
a5516438 | 4581 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4582 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca LT |
4583 | |
4584 | /* | |
4585 | * Instead of doing 'try_get_page()' below in the same_page | |
4586 | * loop, just check the count once here. | |
4587 | */ | |
4588 | if (unlikely(page_count(page) <= 0)) { | |
4589 | if (pages) { | |
4590 | spin_unlock(ptl); | |
4591 | remainder = 0; | |
4592 | err = -ENOMEM; | |
4593 | break; | |
4594 | } | |
4595 | } | |
d5d4b0aa | 4596 | same_page: |
d6692183 | 4597 | if (pages) { |
2a15efc9 | 4598 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 4599 | get_page(pages[i]); |
d6692183 | 4600 | } |
63551ae0 DG |
4601 | |
4602 | if (vmas) | |
4603 | vmas[i] = vma; | |
4604 | ||
4605 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4606 | ++pfn_offset; |
63551ae0 DG |
4607 | --remainder; |
4608 | ++i; | |
d5d4b0aa | 4609 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4610 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4611 | /* |
4612 | * We use pfn_offset to avoid touching the pageframes | |
4613 | * of this compound page. | |
4614 | */ | |
4615 | goto same_page; | |
4616 | } | |
cb900f41 | 4617 | spin_unlock(ptl); |
63551ae0 | 4618 | } |
28a35716 | 4619 | *nr_pages = remainder; |
87ffc118 AA |
4620 | /* |
4621 | * setting position is actually required only if remainder is | |
4622 | * not zero but it's faster not to add a "if (remainder)" | |
4623 | * branch. | |
4624 | */ | |
63551ae0 DG |
4625 | *position = vaddr; |
4626 | ||
2be7cfed | 4627 | return i ? i : err; |
63551ae0 | 4628 | } |
8f860591 | 4629 | |
5491ae7b AK |
4630 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4631 | /* | |
4632 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4633 | * implement this. | |
4634 | */ | |
4635 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4636 | #endif | |
4637 | ||
7da4d641 | 4638 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4639 | unsigned long address, unsigned long end, pgprot_t newprot) |
4640 | { | |
4641 | struct mm_struct *mm = vma->vm_mm; | |
4642 | unsigned long start = address; | |
4643 | pte_t *ptep; | |
4644 | pte_t pte; | |
a5516438 | 4645 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4646 | unsigned long pages = 0; |
dff11abe | 4647 | bool shared_pmd = false; |
ac46d4f3 | 4648 | struct mmu_notifier_range range; |
dff11abe MK |
4649 | |
4650 | /* | |
4651 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 4652 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
4653 | * range if PMD sharing is possible. |
4654 | */ | |
7269f999 JG |
4655 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
4656 | 0, vma, mm, start, end); | |
ac46d4f3 | 4657 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
4658 | |
4659 | BUG_ON(address >= end); | |
ac46d4f3 | 4660 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 4661 | |
ac46d4f3 | 4662 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 4663 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4664 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4665 | spinlock_t *ptl; |
7868a208 | 4666 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4667 | if (!ptep) |
4668 | continue; | |
cb900f41 | 4669 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4670 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4671 | pages++; | |
cb900f41 | 4672 | spin_unlock(ptl); |
dff11abe | 4673 | shared_pmd = true; |
39dde65c | 4674 | continue; |
7da4d641 | 4675 | } |
a8bda28d NH |
4676 | pte = huge_ptep_get(ptep); |
4677 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4678 | spin_unlock(ptl); | |
4679 | continue; | |
4680 | } | |
4681 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4682 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4683 | ||
4684 | if (is_write_migration_entry(entry)) { | |
4685 | pte_t newpte; | |
4686 | ||
4687 | make_migration_entry_read(&entry); | |
4688 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4689 | set_huge_swap_pte_at(mm, address, ptep, |
4690 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4691 | pages++; |
4692 | } | |
4693 | spin_unlock(ptl); | |
4694 | continue; | |
4695 | } | |
4696 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
4697 | pte_t old_pte; |
4698 | ||
4699 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
4700 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 4701 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 4702 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 4703 | pages++; |
8f860591 | 4704 | } |
cb900f41 | 4705 | spin_unlock(ptl); |
8f860591 | 4706 | } |
d833352a | 4707 | /* |
c8c06efa | 4708 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4709 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4710 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
4711 | * and that page table be reused and filled with junk. If we actually |
4712 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 4713 | */ |
dff11abe | 4714 | if (shared_pmd) |
ac46d4f3 | 4715 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
4716 | else |
4717 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
4718 | /* |
4719 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4720 | * page table protection not changing it to point to a new page. | |
4721 | * | |
ad56b738 | 4722 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4723 | */ |
83cde9e8 | 4724 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 4725 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
4726 | |
4727 | return pages << h->order; | |
8f860591 ZY |
4728 | } |
4729 | ||
a1e78772 MG |
4730 | int hugetlb_reserve_pages(struct inode *inode, |
4731 | long from, long to, | |
5a6fe125 | 4732 | struct vm_area_struct *vma, |
ca16d140 | 4733 | vm_flags_t vm_flags) |
e4e574b7 | 4734 | { |
17c9d12e | 4735 | long ret, chg; |
a5516438 | 4736 | struct hstate *h = hstate_inode(inode); |
90481622 | 4737 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4738 | struct resv_map *resv_map; |
1c5ecae3 | 4739 | long gbl_reserve; |
e4e574b7 | 4740 | |
63489f8e MK |
4741 | /* This should never happen */ |
4742 | if (from > to) { | |
4743 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
4744 | return -EINVAL; | |
4745 | } | |
4746 | ||
17c9d12e MG |
4747 | /* |
4748 | * Only apply hugepage reservation if asked. At fault time, an | |
4749 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4750 | * without using reserves |
17c9d12e | 4751 | */ |
ca16d140 | 4752 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4753 | return 0; |
4754 | ||
a1e78772 MG |
4755 | /* |
4756 | * Shared mappings base their reservation on the number of pages that | |
4757 | * are already allocated on behalf of the file. Private mappings need | |
4758 | * to reserve the full area even if read-only as mprotect() may be | |
4759 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4760 | */ | |
9119a41e | 4761 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
4762 | /* |
4763 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
4764 | * called for inodes for which resv_maps were created (see | |
4765 | * hugetlbfs_get_inode). | |
4766 | */ | |
4e35f483 | 4767 | resv_map = inode_resv_map(inode); |
9119a41e | 4768 | |
1406ec9b | 4769 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4770 | |
4771 | } else { | |
4772 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4773 | if (!resv_map) |
4774 | return -ENOMEM; | |
4775 | ||
a1e78772 | 4776 | chg = to - from; |
84afd99b | 4777 | |
17c9d12e MG |
4778 | set_vma_resv_map(vma, resv_map); |
4779 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4780 | } | |
4781 | ||
c50ac050 DH |
4782 | if (chg < 0) { |
4783 | ret = chg; | |
4784 | goto out_err; | |
4785 | } | |
8a630112 | 4786 | |
1c5ecae3 MK |
4787 | /* |
4788 | * There must be enough pages in the subpool for the mapping. If | |
4789 | * the subpool has a minimum size, there may be some global | |
4790 | * reservations already in place (gbl_reserve). | |
4791 | */ | |
4792 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4793 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4794 | ret = -ENOSPC; |
4795 | goto out_err; | |
4796 | } | |
5a6fe125 MG |
4797 | |
4798 | /* | |
17c9d12e | 4799 | * Check enough hugepages are available for the reservation. |
90481622 | 4800 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4801 | */ |
1c5ecae3 | 4802 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4803 | if (ret < 0) { |
1c5ecae3 MK |
4804 | /* put back original number of pages, chg */ |
4805 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4806 | goto out_err; |
68842c9b | 4807 | } |
17c9d12e MG |
4808 | |
4809 | /* | |
4810 | * Account for the reservations made. Shared mappings record regions | |
4811 | * that have reservations as they are shared by multiple VMAs. | |
4812 | * When the last VMA disappears, the region map says how much | |
4813 | * the reservation was and the page cache tells how much of | |
4814 | * the reservation was consumed. Private mappings are per-VMA and | |
4815 | * only the consumed reservations are tracked. When the VMA | |
4816 | * disappears, the original reservation is the VMA size and the | |
4817 | * consumed reservations are stored in the map. Hence, nothing | |
4818 | * else has to be done for private mappings here | |
4819 | */ | |
33039678 MK |
4820 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4821 | long add = region_add(resv_map, from, to); | |
4822 | ||
4823 | if (unlikely(chg > add)) { | |
4824 | /* | |
4825 | * pages in this range were added to the reserve | |
4826 | * map between region_chg and region_add. This | |
4827 | * indicates a race with alloc_huge_page. Adjust | |
4828 | * the subpool and reserve counts modified above | |
4829 | * based on the difference. | |
4830 | */ | |
4831 | long rsv_adjust; | |
4832 | ||
4833 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4834 | chg - add); | |
4835 | hugetlb_acct_memory(h, -rsv_adjust); | |
4836 | } | |
4837 | } | |
a43a8c39 | 4838 | return 0; |
c50ac050 | 4839 | out_err: |
5e911373 | 4840 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4841 | /* Don't call region_abort if region_chg failed */ |
4842 | if (chg >= 0) | |
4843 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4844 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4845 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4846 | return ret; |
a43a8c39 KC |
4847 | } |
4848 | ||
b5cec28d MK |
4849 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4850 | long freed) | |
a43a8c39 | 4851 | { |
a5516438 | 4852 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4853 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4854 | long chg = 0; |
90481622 | 4855 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4856 | long gbl_reserve; |
45c682a6 | 4857 | |
f27a5136 MK |
4858 | /* |
4859 | * Since this routine can be called in the evict inode path for all | |
4860 | * hugetlbfs inodes, resv_map could be NULL. | |
4861 | */ | |
b5cec28d MK |
4862 | if (resv_map) { |
4863 | chg = region_del(resv_map, start, end); | |
4864 | /* | |
4865 | * region_del() can fail in the rare case where a region | |
4866 | * must be split and another region descriptor can not be | |
4867 | * allocated. If end == LONG_MAX, it will not fail. | |
4868 | */ | |
4869 | if (chg < 0) | |
4870 | return chg; | |
4871 | } | |
4872 | ||
45c682a6 | 4873 | spin_lock(&inode->i_lock); |
e4c6f8be | 4874 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4875 | spin_unlock(&inode->i_lock); |
4876 | ||
1c5ecae3 MK |
4877 | /* |
4878 | * If the subpool has a minimum size, the number of global | |
4879 | * reservations to be released may be adjusted. | |
4880 | */ | |
4881 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4882 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4883 | |
4884 | return 0; | |
a43a8c39 | 4885 | } |
93f70f90 | 4886 | |
3212b535 SC |
4887 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4888 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4889 | struct vm_area_struct *vma, | |
4890 | unsigned long addr, pgoff_t idx) | |
4891 | { | |
4892 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4893 | svma->vm_start; | |
4894 | unsigned long sbase = saddr & PUD_MASK; | |
4895 | unsigned long s_end = sbase + PUD_SIZE; | |
4896 | ||
4897 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4898 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4899 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4900 | |
4901 | /* | |
4902 | * match the virtual addresses, permission and the alignment of the | |
4903 | * page table page. | |
4904 | */ | |
4905 | if (pmd_index(addr) != pmd_index(saddr) || | |
4906 | vm_flags != svm_flags || | |
4907 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4908 | return 0; | |
4909 | ||
4910 | return saddr; | |
4911 | } | |
4912 | ||
31aafb45 | 4913 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4914 | { |
4915 | unsigned long base = addr & PUD_MASK; | |
4916 | unsigned long end = base + PUD_SIZE; | |
4917 | ||
4918 | /* | |
4919 | * check on proper vm_flags and page table alignment | |
4920 | */ | |
017b1660 | 4921 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
4922 | return true; |
4923 | return false; | |
3212b535 SC |
4924 | } |
4925 | ||
017b1660 MK |
4926 | /* |
4927 | * Determine if start,end range within vma could be mapped by shared pmd. | |
4928 | * If yes, adjust start and end to cover range associated with possible | |
4929 | * shared pmd mappings. | |
4930 | */ | |
4931 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
4932 | unsigned long *start, unsigned long *end) | |
4933 | { | |
ce202e4b LX |
4934 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
4935 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 4936 | |
ce202e4b LX |
4937 | /* |
4938 | * vma need span at least one aligned PUD size and the start,end range | |
4939 | * must at least partialy within it. | |
4940 | */ | |
4941 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
4942 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
4943 | return; |
4944 | ||
6727dc93 | 4945 | /* Extend the range to be PUD aligned for a worst case scenario */ |
ce202e4b LX |
4946 | if (*start > v_start) |
4947 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 4948 | |
ce202e4b LX |
4949 | if (*end < v_end) |
4950 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
4951 | } |
4952 | ||
3212b535 SC |
4953 | /* |
4954 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4955 | * and returns the corresponding pte. While this is not necessary for the | |
4956 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
ddeaab32 MK |
4957 | * code much cleaner. pmd allocation is essential for the shared case because |
4958 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise | |
4959 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
4960 | * bad pmd for sharing. | |
3212b535 SC |
4961 | */ |
4962 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4963 | { | |
4964 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4965 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4966 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4967 | vma->vm_pgoff; | |
4968 | struct vm_area_struct *svma; | |
4969 | unsigned long saddr; | |
4970 | pte_t *spte = NULL; | |
4971 | pte_t *pte; | |
cb900f41 | 4972 | spinlock_t *ptl; |
3212b535 SC |
4973 | |
4974 | if (!vma_shareable(vma, addr)) | |
4975 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4976 | ||
6f21bb93 | 4977 | i_mmap_lock_read(mapping); |
3212b535 SC |
4978 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4979 | if (svma == vma) | |
4980 | continue; | |
4981 | ||
4982 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4983 | if (saddr) { | |
7868a208 PA |
4984 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4985 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4986 | if (spte) { |
4987 | get_page(virt_to_page(spte)); | |
4988 | break; | |
4989 | } | |
4990 | } | |
4991 | } | |
4992 | ||
4993 | if (!spte) | |
4994 | goto out; | |
4995 | ||
8bea8052 | 4996 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4997 | if (pud_none(*pud)) { |
3212b535 SC |
4998 | pud_populate(mm, pud, |
4999 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5000 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5001 | } else { |
3212b535 | 5002 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5003 | } |
cb900f41 | 5004 | spin_unlock(ptl); |
3212b535 SC |
5005 | out: |
5006 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
6f21bb93 | 5007 | i_mmap_unlock_read(mapping); |
3212b535 SC |
5008 | return pte; |
5009 | } | |
5010 | ||
5011 | /* | |
5012 | * unmap huge page backed by shared pte. | |
5013 | * | |
5014 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5015 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5016 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5017 | * | |
ddeaab32 | 5018 | * called with page table lock held. |
3212b535 SC |
5019 | * |
5020 | * returns: 1 successfully unmapped a shared pte page | |
5021 | * 0 the underlying pte page is not shared, or it is the last user | |
5022 | */ | |
5023 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
5024 | { | |
5025 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
5026 | p4d_t *p4d = p4d_offset(pgd, *addr); |
5027 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
5028 | |
5029 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
5030 | if (page_count(virt_to_page(ptep)) == 1) | |
5031 | return 0; | |
5032 | ||
5033 | pud_clear(pud); | |
5034 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 5035 | mm_dec_nr_pmds(mm); |
3212b535 SC |
5036 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
5037 | return 1; | |
5038 | } | |
9e5fc74c SC |
5039 | #define want_pmd_share() (1) |
5040 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
5041 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5042 | { | |
5043 | return NULL; | |
5044 | } | |
e81f2d22 ZZ |
5045 | |
5046 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
5047 | { | |
5048 | return 0; | |
5049 | } | |
017b1660 MK |
5050 | |
5051 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5052 | unsigned long *start, unsigned long *end) | |
5053 | { | |
5054 | } | |
9e5fc74c | 5055 | #define want_pmd_share() (0) |
3212b535 SC |
5056 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5057 | ||
9e5fc74c SC |
5058 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
5059 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
5060 | unsigned long addr, unsigned long sz) | |
5061 | { | |
5062 | pgd_t *pgd; | |
c2febafc | 5063 | p4d_t *p4d; |
9e5fc74c SC |
5064 | pud_t *pud; |
5065 | pte_t *pte = NULL; | |
5066 | ||
5067 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5068 | p4d = p4d_alloc(mm, pgd, addr); |
5069 | if (!p4d) | |
5070 | return NULL; | |
c2febafc | 5071 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5072 | if (pud) { |
5073 | if (sz == PUD_SIZE) { | |
5074 | pte = (pte_t *)pud; | |
5075 | } else { | |
5076 | BUG_ON(sz != PMD_SIZE); | |
5077 | if (want_pmd_share() && pud_none(*pud)) | |
5078 | pte = huge_pmd_share(mm, addr, pud); | |
5079 | else | |
5080 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5081 | } | |
5082 | } | |
4e666314 | 5083 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5084 | |
5085 | return pte; | |
5086 | } | |
5087 | ||
9b19df29 PA |
5088 | /* |
5089 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5090 | * entry at address @addr | |
5091 | * | |
5092 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
5093 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
5094 | * size @sz doesn't match the hugepage size at this level of the page | |
5095 | * table. | |
5096 | */ | |
7868a208 PA |
5097 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5098 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5099 | { |
5100 | pgd_t *pgd; | |
c2febafc | 5101 | p4d_t *p4d; |
37a6e55e L |
5102 | pud_t *pud, pud_entry; |
5103 | pmd_t *pmd, pmd_entry; | |
9e5fc74c SC |
5104 | |
5105 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5106 | if (!pgd_present(*pgd)) |
5107 | return NULL; | |
5108 | p4d = p4d_offset(pgd, addr); | |
5109 | if (!p4d_present(*p4d)) | |
5110 | return NULL; | |
9b19df29 | 5111 | |
c2febafc | 5112 | pud = pud_offset(p4d, addr); |
37a6e55e L |
5113 | pud_entry = READ_ONCE(*pud); |
5114 | if (sz != PUD_SIZE && pud_none(pud_entry)) | |
c2febafc | 5115 | return NULL; |
9b19df29 | 5116 | /* hugepage or swap? */ |
37a6e55e | 5117 | if (pud_huge(pud_entry) || !pud_present(pud_entry)) |
c2febafc | 5118 | return (pte_t *)pud; |
9b19df29 | 5119 | |
c2febafc | 5120 | pmd = pmd_offset(pud, addr); |
37a6e55e L |
5121 | pmd_entry = READ_ONCE(*pmd); |
5122 | if (sz != PMD_SIZE && pmd_none(pmd_entry)) | |
9b19df29 PA |
5123 | return NULL; |
5124 | /* hugepage or swap? */ | |
37a6e55e | 5125 | if (pmd_huge(pmd_entry) || !pmd_present(pmd_entry)) |
9b19df29 PA |
5126 | return (pte_t *)pmd; |
5127 | ||
5128 | return NULL; | |
9e5fc74c SC |
5129 | } |
5130 | ||
61f77eda NH |
5131 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5132 | ||
5133 | /* | |
5134 | * These functions are overwritable if your architecture needs its own | |
5135 | * behavior. | |
5136 | */ | |
5137 | struct page * __weak | |
5138 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5139 | int write) | |
5140 | { | |
5141 | return ERR_PTR(-EINVAL); | |
5142 | } | |
5143 | ||
4dc71451 AK |
5144 | struct page * __weak |
5145 | follow_huge_pd(struct vm_area_struct *vma, | |
5146 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5147 | { | |
5148 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5149 | return NULL; | |
5150 | } | |
5151 | ||
61f77eda | 5152 | struct page * __weak |
9e5fc74c | 5153 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5154 | pmd_t *pmd, int flags) |
9e5fc74c | 5155 | { |
e66f17ff NH |
5156 | struct page *page = NULL; |
5157 | spinlock_t *ptl; | |
c9d398fa | 5158 | pte_t pte; |
e66f17ff NH |
5159 | retry: |
5160 | ptl = pmd_lockptr(mm, pmd); | |
5161 | spin_lock(ptl); | |
5162 | /* | |
5163 | * make sure that the address range covered by this pmd is not | |
5164 | * unmapped from other threads. | |
5165 | */ | |
5166 | if (!pmd_huge(*pmd)) | |
5167 | goto out; | |
c9d398fa NH |
5168 | pte = huge_ptep_get((pte_t *)pmd); |
5169 | if (pte_present(pte)) { | |
97534127 | 5170 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
5171 | if (flags & FOLL_GET) |
5172 | get_page(page); | |
5173 | } else { | |
c9d398fa | 5174 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5175 | spin_unlock(ptl); |
5176 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5177 | goto retry; | |
5178 | } | |
5179 | /* | |
5180 | * hwpoisoned entry is treated as no_page_table in | |
5181 | * follow_page_mask(). | |
5182 | */ | |
5183 | } | |
5184 | out: | |
5185 | spin_unlock(ptl); | |
9e5fc74c SC |
5186 | return page; |
5187 | } | |
5188 | ||
61f77eda | 5189 | struct page * __weak |
9e5fc74c | 5190 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5191 | pud_t *pud, int flags) |
9e5fc74c | 5192 | { |
e66f17ff NH |
5193 | if (flags & FOLL_GET) |
5194 | return NULL; | |
9e5fc74c | 5195 | |
e66f17ff | 5196 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5197 | } |
5198 | ||
faaa5b62 AK |
5199 | struct page * __weak |
5200 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5201 | { | |
5202 | if (flags & FOLL_GET) | |
5203 | return NULL; | |
5204 | ||
5205 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5206 | } | |
5207 | ||
31caf665 NH |
5208 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5209 | { | |
bcc54222 NH |
5210 | bool ret = true; |
5211 | ||
31caf665 | 5212 | spin_lock(&hugetlb_lock); |
9b7a670c MS |
5213 | if (!PageHeadHuge(page) || !page_huge_active(page) || |
5214 | !get_page_unless_zero(page)) { | |
bcc54222 NH |
5215 | ret = false; |
5216 | goto unlock; | |
5217 | } | |
5218 | clear_page_huge_active(page); | |
31caf665 | 5219 | list_move_tail(&page->lru, list); |
bcc54222 | 5220 | unlock: |
31caf665 | 5221 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5222 | return ret; |
31caf665 NH |
5223 | } |
5224 | ||
5225 | void putback_active_hugepage(struct page *page) | |
5226 | { | |
309381fe | 5227 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5228 | spin_lock(&hugetlb_lock); |
bcc54222 | 5229 | set_page_huge_active(page); |
31caf665 NH |
5230 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5231 | spin_unlock(&hugetlb_lock); | |
5232 | put_page(page); | |
5233 | } | |
ab5ac90a MH |
5234 | |
5235 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5236 | { | |
5237 | struct hstate *h = page_hstate(oldpage); | |
5238 | ||
5239 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5240 | set_page_owner_migrate_reason(newpage, reason); | |
5241 | ||
5242 | /* | |
5243 | * transfer temporary state of the new huge page. This is | |
5244 | * reverse to other transitions because the newpage is going to | |
5245 | * be final while the old one will be freed so it takes over | |
5246 | * the temporary status. | |
5247 | * | |
5248 | * Also note that we have to transfer the per-node surplus state | |
5249 | * here as well otherwise the global surplus count will not match | |
5250 | * the per-node's. | |
5251 | */ | |
5252 | if (PageHugeTemporary(newpage)) { | |
5253 | int old_nid = page_to_nid(oldpage); | |
5254 | int new_nid = page_to_nid(newpage); | |
5255 | ||
5256 | SetPageHugeTemporary(oldpage); | |
5257 | ClearPageHugeTemporary(newpage); | |
5258 | ||
5259 | spin_lock(&hugetlb_lock); | |
5260 | if (h->surplus_huge_pages_node[old_nid]) { | |
5261 | h->surplus_huge_pages_node[old_nid]--; | |
5262 | h->surplus_huge_pages_node[new_nid]++; | |
5263 | } | |
5264 | spin_unlock(&hugetlb_lock); | |
5265 | } | |
5266 | } |