]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * Generic hugetlb support. | |
6d49e352 | 3 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 4 | */ |
1da177e4 LT |
5 | #include <linux/list.h> |
6 | #include <linux/init.h> | |
1da177e4 | 7 | #include <linux/mm.h> |
e1759c21 | 8 | #include <linux/seq_file.h> |
1da177e4 LT |
9 | #include <linux/sysctl.h> |
10 | #include <linux/highmem.h> | |
cddb8a5c | 11 | #include <linux/mmu_notifier.h> |
1da177e4 | 12 | #include <linux/nodemask.h> |
63551ae0 | 13 | #include <linux/pagemap.h> |
5da7ca86 | 14 | #include <linux/mempolicy.h> |
3b32123d | 15 | #include <linux/compiler.h> |
aea47ff3 | 16 | #include <linux/cpuset.h> |
3935baa9 | 17 | #include <linux/mutex.h> |
aa888a74 | 18 | #include <linux/bootmem.h> |
a3437870 | 19 | #include <linux/sysfs.h> |
5a0e3ad6 | 20 | #include <linux/slab.h> |
174cd4b1 | 21 | #include <linux/sched/signal.h> |
0fe6e20b | 22 | #include <linux/rmap.h> |
c6247f72 | 23 | #include <linux/string_helpers.h> |
fd6a03ed NH |
24 | #include <linux/swap.h> |
25 | #include <linux/swapops.h> | |
8382d914 | 26 | #include <linux/jhash.h> |
d6606683 | 27 | |
63551ae0 DG |
28 | #include <asm/page.h> |
29 | #include <asm/pgtable.h> | |
24669e58 | 30 | #include <asm/tlb.h> |
63551ae0 | 31 | |
24669e58 | 32 | #include <linux/io.h> |
63551ae0 | 33 | #include <linux/hugetlb.h> |
9dd540e2 | 34 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 35 | #include <linux/node.h> |
1a1aad8a | 36 | #include <linux/userfaultfd_k.h> |
7835e98b | 37 | #include "internal.h" |
1da177e4 | 38 | |
753162cd | 39 | int hugepages_treat_as_movable; |
a5516438 | 40 | |
c3f38a38 | 41 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
42 | unsigned int default_hstate_idx; |
43 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
44 | /* |
45 | * Minimum page order among possible hugepage sizes, set to a proper value | |
46 | * at boot time. | |
47 | */ | |
48 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 49 | |
53ba51d2 JT |
50 | __initdata LIST_HEAD(huge_boot_pages); |
51 | ||
e5ff2159 AK |
52 | /* for command line parsing */ |
53 | static struct hstate * __initdata parsed_hstate; | |
54 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 55 | static unsigned long __initdata default_hstate_size; |
9fee021d | 56 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 57 | |
3935baa9 | 58 | /* |
31caf665 NH |
59 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
60 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 61 | */ |
c3f38a38 | 62 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 63 | |
8382d914 DB |
64 | /* |
65 | * Serializes faults on the same logical page. This is used to | |
66 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
67 | */ | |
68 | static int num_fault_mutexes; | |
c672c7f2 | 69 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 70 | |
7ca02d0a MK |
71 | /* Forward declaration */ |
72 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
73 | ||
90481622 DG |
74 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
75 | { | |
76 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
77 | ||
78 | spin_unlock(&spool->lock); | |
79 | ||
80 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
81 | * remain, give up any reservations mased on minimum size and |
82 | * free the subpool */ | |
83 | if (free) { | |
84 | if (spool->min_hpages != -1) | |
85 | hugetlb_acct_memory(spool->hstate, | |
86 | -spool->min_hpages); | |
90481622 | 87 | kfree(spool); |
7ca02d0a | 88 | } |
90481622 DG |
89 | } |
90 | ||
7ca02d0a MK |
91 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
92 | long min_hpages) | |
90481622 DG |
93 | { |
94 | struct hugepage_subpool *spool; | |
95 | ||
c6a91820 | 96 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
97 | if (!spool) |
98 | return NULL; | |
99 | ||
100 | spin_lock_init(&spool->lock); | |
101 | spool->count = 1; | |
7ca02d0a MK |
102 | spool->max_hpages = max_hpages; |
103 | spool->hstate = h; | |
104 | spool->min_hpages = min_hpages; | |
105 | ||
106 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
107 | kfree(spool); | |
108 | return NULL; | |
109 | } | |
110 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
111 | |
112 | return spool; | |
113 | } | |
114 | ||
115 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
116 | { | |
117 | spin_lock(&spool->lock); | |
118 | BUG_ON(!spool->count); | |
119 | spool->count--; | |
120 | unlock_or_release_subpool(spool); | |
121 | } | |
122 | ||
1c5ecae3 MK |
123 | /* |
124 | * Subpool accounting for allocating and reserving pages. | |
125 | * Return -ENOMEM if there are not enough resources to satisfy the | |
126 | * the request. Otherwise, return the number of pages by which the | |
127 | * global pools must be adjusted (upward). The returned value may | |
128 | * only be different than the passed value (delta) in the case where | |
129 | * a subpool minimum size must be manitained. | |
130 | */ | |
131 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
132 | long delta) |
133 | { | |
1c5ecae3 | 134 | long ret = delta; |
90481622 DG |
135 | |
136 | if (!spool) | |
1c5ecae3 | 137 | return ret; |
90481622 DG |
138 | |
139 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
140 | |
141 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
142 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
143 | spool->used_hpages += delta; | |
144 | else { | |
145 | ret = -ENOMEM; | |
146 | goto unlock_ret; | |
147 | } | |
90481622 | 148 | } |
90481622 | 149 | |
09a95e29 MK |
150 | /* minimum size accounting */ |
151 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
152 | if (delta > spool->rsv_hpages) { |
153 | /* | |
154 | * Asking for more reserves than those already taken on | |
155 | * behalf of subpool. Return difference. | |
156 | */ | |
157 | ret = delta - spool->rsv_hpages; | |
158 | spool->rsv_hpages = 0; | |
159 | } else { | |
160 | ret = 0; /* reserves already accounted for */ | |
161 | spool->rsv_hpages -= delta; | |
162 | } | |
163 | } | |
164 | ||
165 | unlock_ret: | |
166 | spin_unlock(&spool->lock); | |
90481622 DG |
167 | return ret; |
168 | } | |
169 | ||
1c5ecae3 MK |
170 | /* |
171 | * Subpool accounting for freeing and unreserving pages. | |
172 | * Return the number of global page reservations that must be dropped. | |
173 | * The return value may only be different than the passed value (delta) | |
174 | * in the case where a subpool minimum size must be maintained. | |
175 | */ | |
176 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
177 | long delta) |
178 | { | |
1c5ecae3 MK |
179 | long ret = delta; |
180 | ||
90481622 | 181 | if (!spool) |
1c5ecae3 | 182 | return delta; |
90481622 DG |
183 | |
184 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
185 | |
186 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
187 | spool->used_hpages -= delta; | |
188 | ||
09a95e29 MK |
189 | /* minimum size accounting */ |
190 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
191 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
192 | ret = 0; | |
193 | else | |
194 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
195 | ||
196 | spool->rsv_hpages += delta; | |
197 | if (spool->rsv_hpages > spool->min_hpages) | |
198 | spool->rsv_hpages = spool->min_hpages; | |
199 | } | |
200 | ||
201 | /* | |
202 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
203 | * quota reference, free it now. | |
204 | */ | |
90481622 | 205 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
206 | |
207 | return ret; | |
90481622 DG |
208 | } |
209 | ||
210 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
211 | { | |
212 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
213 | } | |
214 | ||
215 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
216 | { | |
496ad9aa | 217 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
218 | } |
219 | ||
96822904 AW |
220 | /* |
221 | * Region tracking -- allows tracking of reservations and instantiated pages | |
222 | * across the pages in a mapping. | |
84afd99b | 223 | * |
1dd308a7 MK |
224 | * The region data structures are embedded into a resv_map and protected |
225 | * by a resv_map's lock. The set of regions within the resv_map represent | |
226 | * reservations for huge pages, or huge pages that have already been | |
227 | * instantiated within the map. The from and to elements are huge page | |
228 | * indicies into the associated mapping. from indicates the starting index | |
229 | * of the region. to represents the first index past the end of the region. | |
230 | * | |
231 | * For example, a file region structure with from == 0 and to == 4 represents | |
232 | * four huge pages in a mapping. It is important to note that the to element | |
233 | * represents the first element past the end of the region. This is used in | |
234 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
235 | * | |
236 | * Interval notation of the form [from, to) will be used to indicate that | |
237 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
238 | */ |
239 | struct file_region { | |
240 | struct list_head link; | |
241 | long from; | |
242 | long to; | |
243 | }; | |
244 | ||
1dd308a7 MK |
245 | /* |
246 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
247 | * map. In the normal case, existing regions will be expanded |
248 | * to accommodate the specified range. Sufficient regions should | |
249 | * exist for expansion due to the previous call to region_chg | |
250 | * with the same range. However, it is possible that region_del | |
251 | * could have been called after region_chg and modifed the map | |
252 | * in such a way that no region exists to be expanded. In this | |
253 | * case, pull a region descriptor from the cache associated with | |
254 | * the map and use that for the new range. | |
cf3ad20b MK |
255 | * |
256 | * Return the number of new huge pages added to the map. This | |
257 | * number is greater than or equal to zero. | |
1dd308a7 | 258 | */ |
1406ec9b | 259 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 260 | { |
1406ec9b | 261 | struct list_head *head = &resv->regions; |
96822904 | 262 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 263 | long add = 0; |
96822904 | 264 | |
7b24d861 | 265 | spin_lock(&resv->lock); |
96822904 AW |
266 | /* Locate the region we are either in or before. */ |
267 | list_for_each_entry(rg, head, link) | |
268 | if (f <= rg->to) | |
269 | break; | |
270 | ||
5e911373 MK |
271 | /* |
272 | * If no region exists which can be expanded to include the | |
273 | * specified range, the list must have been modified by an | |
274 | * interleving call to region_del(). Pull a region descriptor | |
275 | * from the cache and use it for this range. | |
276 | */ | |
277 | if (&rg->link == head || t < rg->from) { | |
278 | VM_BUG_ON(resv->region_cache_count <= 0); | |
279 | ||
280 | resv->region_cache_count--; | |
281 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
282 | link); | |
283 | list_del(&nrg->link); | |
284 | ||
285 | nrg->from = f; | |
286 | nrg->to = t; | |
287 | list_add(&nrg->link, rg->link.prev); | |
288 | ||
289 | add += t - f; | |
290 | goto out_locked; | |
291 | } | |
292 | ||
96822904 AW |
293 | /* Round our left edge to the current segment if it encloses us. */ |
294 | if (f > rg->from) | |
295 | f = rg->from; | |
296 | ||
297 | /* Check for and consume any regions we now overlap with. */ | |
298 | nrg = rg; | |
299 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
300 | if (&rg->link == head) | |
301 | break; | |
302 | if (rg->from > t) | |
303 | break; | |
304 | ||
305 | /* If this area reaches higher then extend our area to | |
306 | * include it completely. If this is not the first area | |
307 | * which we intend to reuse, free it. */ | |
308 | if (rg->to > t) | |
309 | t = rg->to; | |
310 | if (rg != nrg) { | |
cf3ad20b MK |
311 | /* Decrement return value by the deleted range. |
312 | * Another range will span this area so that by | |
313 | * end of routine add will be >= zero | |
314 | */ | |
315 | add -= (rg->to - rg->from); | |
96822904 AW |
316 | list_del(&rg->link); |
317 | kfree(rg); | |
318 | } | |
319 | } | |
cf3ad20b MK |
320 | |
321 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 322 | nrg->from = f; |
cf3ad20b | 323 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 324 | nrg->to = t; |
cf3ad20b | 325 | |
5e911373 MK |
326 | out_locked: |
327 | resv->adds_in_progress--; | |
7b24d861 | 328 | spin_unlock(&resv->lock); |
cf3ad20b MK |
329 | VM_BUG_ON(add < 0); |
330 | return add; | |
96822904 AW |
331 | } |
332 | ||
1dd308a7 MK |
333 | /* |
334 | * Examine the existing reserve map and determine how many | |
335 | * huge pages in the specified range [f, t) are NOT currently | |
336 | * represented. This routine is called before a subsequent | |
337 | * call to region_add that will actually modify the reserve | |
338 | * map to add the specified range [f, t). region_chg does | |
339 | * not change the number of huge pages represented by the | |
340 | * map. However, if the existing regions in the map can not | |
341 | * be expanded to represent the new range, a new file_region | |
342 | * structure is added to the map as a placeholder. This is | |
343 | * so that the subsequent region_add call will have all the | |
344 | * regions it needs and will not fail. | |
345 | * | |
5e911373 MK |
346 | * Upon entry, region_chg will also examine the cache of region descriptors |
347 | * associated with the map. If there are not enough descriptors cached, one | |
348 | * will be allocated for the in progress add operation. | |
349 | * | |
350 | * Returns the number of huge pages that need to be added to the existing | |
351 | * reservation map for the range [f, t). This number is greater or equal to | |
352 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
353 | * is needed and can not be allocated. | |
1dd308a7 | 354 | */ |
1406ec9b | 355 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 356 | { |
1406ec9b | 357 | struct list_head *head = &resv->regions; |
7b24d861 | 358 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
359 | long chg = 0; |
360 | ||
7b24d861 DB |
361 | retry: |
362 | spin_lock(&resv->lock); | |
5e911373 MK |
363 | retry_locked: |
364 | resv->adds_in_progress++; | |
365 | ||
366 | /* | |
367 | * Check for sufficient descriptors in the cache to accommodate | |
368 | * the number of in progress add operations. | |
369 | */ | |
370 | if (resv->adds_in_progress > resv->region_cache_count) { | |
371 | struct file_region *trg; | |
372 | ||
373 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
374 | /* Must drop lock to allocate a new descriptor. */ | |
375 | resv->adds_in_progress--; | |
376 | spin_unlock(&resv->lock); | |
377 | ||
378 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
dbe409e4 MK |
379 | if (!trg) { |
380 | kfree(nrg); | |
5e911373 | 381 | return -ENOMEM; |
dbe409e4 | 382 | } |
5e911373 MK |
383 | |
384 | spin_lock(&resv->lock); | |
385 | list_add(&trg->link, &resv->region_cache); | |
386 | resv->region_cache_count++; | |
387 | goto retry_locked; | |
388 | } | |
389 | ||
96822904 AW |
390 | /* Locate the region we are before or in. */ |
391 | list_for_each_entry(rg, head, link) | |
392 | if (f <= rg->to) | |
393 | break; | |
394 | ||
395 | /* If we are below the current region then a new region is required. | |
396 | * Subtle, allocate a new region at the position but make it zero | |
397 | * size such that we can guarantee to record the reservation. */ | |
398 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 399 | if (!nrg) { |
5e911373 | 400 | resv->adds_in_progress--; |
7b24d861 DB |
401 | spin_unlock(&resv->lock); |
402 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
403 | if (!nrg) | |
404 | return -ENOMEM; | |
405 | ||
406 | nrg->from = f; | |
407 | nrg->to = f; | |
408 | INIT_LIST_HEAD(&nrg->link); | |
409 | goto retry; | |
410 | } | |
96822904 | 411 | |
7b24d861 DB |
412 | list_add(&nrg->link, rg->link.prev); |
413 | chg = t - f; | |
414 | goto out_nrg; | |
96822904 AW |
415 | } |
416 | ||
417 | /* Round our left edge to the current segment if it encloses us. */ | |
418 | if (f > rg->from) | |
419 | f = rg->from; | |
420 | chg = t - f; | |
421 | ||
422 | /* Check for and consume any regions we now overlap with. */ | |
423 | list_for_each_entry(rg, rg->link.prev, link) { | |
424 | if (&rg->link == head) | |
425 | break; | |
426 | if (rg->from > t) | |
7b24d861 | 427 | goto out; |
96822904 | 428 | |
25985edc | 429 | /* We overlap with this area, if it extends further than |
96822904 AW |
430 | * us then we must extend ourselves. Account for its |
431 | * existing reservation. */ | |
432 | if (rg->to > t) { | |
433 | chg += rg->to - t; | |
434 | t = rg->to; | |
435 | } | |
436 | chg -= rg->to - rg->from; | |
437 | } | |
7b24d861 DB |
438 | |
439 | out: | |
440 | spin_unlock(&resv->lock); | |
441 | /* We already know we raced and no longer need the new region */ | |
442 | kfree(nrg); | |
443 | return chg; | |
444 | out_nrg: | |
445 | spin_unlock(&resv->lock); | |
96822904 AW |
446 | return chg; |
447 | } | |
448 | ||
5e911373 MK |
449 | /* |
450 | * Abort the in progress add operation. The adds_in_progress field | |
451 | * of the resv_map keeps track of the operations in progress between | |
452 | * calls to region_chg and region_add. Operations are sometimes | |
453 | * aborted after the call to region_chg. In such cases, region_abort | |
454 | * is called to decrement the adds_in_progress counter. | |
455 | * | |
456 | * NOTE: The range arguments [f, t) are not needed or used in this | |
457 | * routine. They are kept to make reading the calling code easier as | |
458 | * arguments will match the associated region_chg call. | |
459 | */ | |
460 | static void region_abort(struct resv_map *resv, long f, long t) | |
461 | { | |
462 | spin_lock(&resv->lock); | |
463 | VM_BUG_ON(!resv->region_cache_count); | |
464 | resv->adds_in_progress--; | |
465 | spin_unlock(&resv->lock); | |
466 | } | |
467 | ||
1dd308a7 | 468 | /* |
feba16e2 MK |
469 | * Delete the specified range [f, t) from the reserve map. If the |
470 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
471 | * should be deleted. Locate the regions which intersect [f, t) | |
472 | * and either trim, delete or split the existing regions. | |
473 | * | |
474 | * Returns the number of huge pages deleted from the reserve map. | |
475 | * In the normal case, the return value is zero or more. In the | |
476 | * case where a region must be split, a new region descriptor must | |
477 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
478 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
479 | * a region and possibly return -ENOMEM. Callers specifying | |
480 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 481 | */ |
feba16e2 | 482 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 483 | { |
1406ec9b | 484 | struct list_head *head = &resv->regions; |
96822904 | 485 | struct file_region *rg, *trg; |
feba16e2 MK |
486 | struct file_region *nrg = NULL; |
487 | long del = 0; | |
96822904 | 488 | |
feba16e2 | 489 | retry: |
7b24d861 | 490 | spin_lock(&resv->lock); |
feba16e2 | 491 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
492 | /* |
493 | * Skip regions before the range to be deleted. file_region | |
494 | * ranges are normally of the form [from, to). However, there | |
495 | * may be a "placeholder" entry in the map which is of the form | |
496 | * (from, to) with from == to. Check for placeholder entries | |
497 | * at the beginning of the range to be deleted. | |
498 | */ | |
499 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 500 | continue; |
dbe409e4 | 501 | |
feba16e2 | 502 | if (rg->from >= t) |
96822904 | 503 | break; |
96822904 | 504 | |
feba16e2 MK |
505 | if (f > rg->from && t < rg->to) { /* Must split region */ |
506 | /* | |
507 | * Check for an entry in the cache before dropping | |
508 | * lock and attempting allocation. | |
509 | */ | |
510 | if (!nrg && | |
511 | resv->region_cache_count > resv->adds_in_progress) { | |
512 | nrg = list_first_entry(&resv->region_cache, | |
513 | struct file_region, | |
514 | link); | |
515 | list_del(&nrg->link); | |
516 | resv->region_cache_count--; | |
517 | } | |
96822904 | 518 | |
feba16e2 MK |
519 | if (!nrg) { |
520 | spin_unlock(&resv->lock); | |
521 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
522 | if (!nrg) | |
523 | return -ENOMEM; | |
524 | goto retry; | |
525 | } | |
526 | ||
527 | del += t - f; | |
528 | ||
529 | /* New entry for end of split region */ | |
530 | nrg->from = t; | |
531 | nrg->to = rg->to; | |
532 | INIT_LIST_HEAD(&nrg->link); | |
533 | ||
534 | /* Original entry is trimmed */ | |
535 | rg->to = f; | |
536 | ||
537 | list_add(&nrg->link, &rg->link); | |
538 | nrg = NULL; | |
96822904 | 539 | break; |
feba16e2 MK |
540 | } |
541 | ||
542 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
543 | del += rg->to - rg->from; | |
544 | list_del(&rg->link); | |
545 | kfree(rg); | |
546 | continue; | |
547 | } | |
548 | ||
549 | if (f <= rg->from) { /* Trim beginning of region */ | |
550 | del += t - rg->from; | |
551 | rg->from = t; | |
552 | } else { /* Trim end of region */ | |
553 | del += rg->to - f; | |
554 | rg->to = f; | |
555 | } | |
96822904 | 556 | } |
7b24d861 | 557 | |
7b24d861 | 558 | spin_unlock(&resv->lock); |
feba16e2 MK |
559 | kfree(nrg); |
560 | return del; | |
96822904 AW |
561 | } |
562 | ||
b5cec28d MK |
563 | /* |
564 | * A rare out of memory error was encountered which prevented removal of | |
565 | * the reserve map region for a page. The huge page itself was free'ed | |
566 | * and removed from the page cache. This routine will adjust the subpool | |
567 | * usage count, and the global reserve count if needed. By incrementing | |
568 | * these counts, the reserve map entry which could not be deleted will | |
569 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
570 | * counts. | |
571 | */ | |
72e2936c | 572 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
573 | { |
574 | struct hugepage_subpool *spool = subpool_inode(inode); | |
575 | long rsv_adjust; | |
576 | ||
577 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 578 | if (rsv_adjust) { |
b5cec28d MK |
579 | struct hstate *h = hstate_inode(inode); |
580 | ||
581 | hugetlb_acct_memory(h, 1); | |
582 | } | |
583 | } | |
584 | ||
1dd308a7 MK |
585 | /* |
586 | * Count and return the number of huge pages in the reserve map | |
587 | * that intersect with the range [f, t). | |
588 | */ | |
1406ec9b | 589 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 590 | { |
1406ec9b | 591 | struct list_head *head = &resv->regions; |
84afd99b AW |
592 | struct file_region *rg; |
593 | long chg = 0; | |
594 | ||
7b24d861 | 595 | spin_lock(&resv->lock); |
84afd99b AW |
596 | /* Locate each segment we overlap with, and count that overlap. */ |
597 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
598 | long seg_from; |
599 | long seg_to; | |
84afd99b AW |
600 | |
601 | if (rg->to <= f) | |
602 | continue; | |
603 | if (rg->from >= t) | |
604 | break; | |
605 | ||
606 | seg_from = max(rg->from, f); | |
607 | seg_to = min(rg->to, t); | |
608 | ||
609 | chg += seg_to - seg_from; | |
610 | } | |
7b24d861 | 611 | spin_unlock(&resv->lock); |
84afd99b AW |
612 | |
613 | return chg; | |
614 | } | |
615 | ||
e7c4b0bf AW |
616 | /* |
617 | * Convert the address within this vma to the page offset within | |
618 | * the mapping, in pagecache page units; huge pages here. | |
619 | */ | |
a5516438 AK |
620 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
621 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 622 | { |
a5516438 AK |
623 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
624 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
625 | } |
626 | ||
0fe6e20b NH |
627 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
628 | unsigned long address) | |
629 | { | |
630 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
631 | } | |
dee41079 | 632 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 633 | |
08fba699 MG |
634 | /* |
635 | * Return the size of the pages allocated when backing a VMA. In the majority | |
636 | * cases this will be same size as used by the page table entries. | |
637 | */ | |
638 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
639 | { | |
640 | struct hstate *hstate; | |
641 | ||
642 | if (!is_vm_hugetlb_page(vma)) | |
643 | return PAGE_SIZE; | |
644 | ||
645 | hstate = hstate_vma(vma); | |
646 | ||
2415cf12 | 647 | return 1UL << huge_page_shift(hstate); |
08fba699 | 648 | } |
f340ca0f | 649 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 650 | |
3340289d MG |
651 | /* |
652 | * Return the page size being used by the MMU to back a VMA. In the majority | |
653 | * of cases, the page size used by the kernel matches the MMU size. On | |
654 | * architectures where it differs, an architecture-specific version of this | |
655 | * function is required. | |
656 | */ | |
657 | #ifndef vma_mmu_pagesize | |
658 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
659 | { | |
660 | return vma_kernel_pagesize(vma); | |
661 | } | |
662 | #endif | |
663 | ||
84afd99b AW |
664 | /* |
665 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
666 | * bits of the reservation map pointer, which are always clear due to | |
667 | * alignment. | |
668 | */ | |
669 | #define HPAGE_RESV_OWNER (1UL << 0) | |
670 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 671 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 672 | |
a1e78772 MG |
673 | /* |
674 | * These helpers are used to track how many pages are reserved for | |
675 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
676 | * is guaranteed to have their future faults succeed. | |
677 | * | |
678 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
679 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
680 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
681 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
682 | * |
683 | * The private mapping reservation is represented in a subtly different | |
684 | * manner to a shared mapping. A shared mapping has a region map associated | |
685 | * with the underlying file, this region map represents the backing file | |
686 | * pages which have ever had a reservation assigned which this persists even | |
687 | * after the page is instantiated. A private mapping has a region map | |
688 | * associated with the original mmap which is attached to all VMAs which | |
689 | * reference it, this region map represents those offsets which have consumed | |
690 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 691 | */ |
e7c4b0bf AW |
692 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
693 | { | |
694 | return (unsigned long)vma->vm_private_data; | |
695 | } | |
696 | ||
697 | static void set_vma_private_data(struct vm_area_struct *vma, | |
698 | unsigned long value) | |
699 | { | |
700 | vma->vm_private_data = (void *)value; | |
701 | } | |
702 | ||
9119a41e | 703 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
704 | { |
705 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
706 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
707 | ||
708 | if (!resv_map || !rg) { | |
709 | kfree(resv_map); | |
710 | kfree(rg); | |
84afd99b | 711 | return NULL; |
5e911373 | 712 | } |
84afd99b AW |
713 | |
714 | kref_init(&resv_map->refs); | |
7b24d861 | 715 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
716 | INIT_LIST_HEAD(&resv_map->regions); |
717 | ||
5e911373 MK |
718 | resv_map->adds_in_progress = 0; |
719 | ||
720 | INIT_LIST_HEAD(&resv_map->region_cache); | |
721 | list_add(&rg->link, &resv_map->region_cache); | |
722 | resv_map->region_cache_count = 1; | |
723 | ||
84afd99b AW |
724 | return resv_map; |
725 | } | |
726 | ||
9119a41e | 727 | void resv_map_release(struct kref *ref) |
84afd99b AW |
728 | { |
729 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
730 | struct list_head *head = &resv_map->region_cache; |
731 | struct file_region *rg, *trg; | |
84afd99b AW |
732 | |
733 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 734 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
735 | |
736 | /* ... and any entries left in the cache */ | |
737 | list_for_each_entry_safe(rg, trg, head, link) { | |
738 | list_del(&rg->link); | |
739 | kfree(rg); | |
740 | } | |
741 | ||
742 | VM_BUG_ON(resv_map->adds_in_progress); | |
743 | ||
84afd99b AW |
744 | kfree(resv_map); |
745 | } | |
746 | ||
4e35f483 JK |
747 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
748 | { | |
749 | return inode->i_mapping->private_data; | |
750 | } | |
751 | ||
84afd99b | 752 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 753 | { |
81d1b09c | 754 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
755 | if (vma->vm_flags & VM_MAYSHARE) { |
756 | struct address_space *mapping = vma->vm_file->f_mapping; | |
757 | struct inode *inode = mapping->host; | |
758 | ||
759 | return inode_resv_map(inode); | |
760 | ||
761 | } else { | |
84afd99b AW |
762 | return (struct resv_map *)(get_vma_private_data(vma) & |
763 | ~HPAGE_RESV_MASK); | |
4e35f483 | 764 | } |
a1e78772 MG |
765 | } |
766 | ||
84afd99b | 767 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 768 | { |
81d1b09c SL |
769 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
770 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 771 | |
84afd99b AW |
772 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
773 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
774 | } |
775 | ||
776 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
777 | { | |
81d1b09c SL |
778 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
779 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
780 | |
781 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
782 | } |
783 | ||
784 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
785 | { | |
81d1b09c | 786 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
787 | |
788 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
789 | } |
790 | ||
04f2cbe3 | 791 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
792 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
793 | { | |
81d1b09c | 794 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 795 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
796 | vma->vm_private_data = (void *)0; |
797 | } | |
798 | ||
799 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 800 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 801 | { |
af0ed73e JK |
802 | if (vma->vm_flags & VM_NORESERVE) { |
803 | /* | |
804 | * This address is already reserved by other process(chg == 0), | |
805 | * so, we should decrement reserved count. Without decrementing, | |
806 | * reserve count remains after releasing inode, because this | |
807 | * allocated page will go into page cache and is regarded as | |
808 | * coming from reserved pool in releasing step. Currently, we | |
809 | * don't have any other solution to deal with this situation | |
810 | * properly, so add work-around here. | |
811 | */ | |
812 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 813 | return true; |
af0ed73e | 814 | else |
559ec2f8 | 815 | return false; |
af0ed73e | 816 | } |
a63884e9 JK |
817 | |
818 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
819 | if (vma->vm_flags & VM_MAYSHARE) { |
820 | /* | |
821 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
822 | * be a region map for all pages. The only situation where | |
823 | * there is no region map is if a hole was punched via | |
824 | * fallocate. In this case, there really are no reverves to | |
825 | * use. This situation is indicated if chg != 0. | |
826 | */ | |
827 | if (chg) | |
828 | return false; | |
829 | else | |
830 | return true; | |
831 | } | |
a63884e9 JK |
832 | |
833 | /* | |
834 | * Only the process that called mmap() has reserves for | |
835 | * private mappings. | |
836 | */ | |
67961f9d MK |
837 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
838 | /* | |
839 | * Like the shared case above, a hole punch or truncate | |
840 | * could have been performed on the private mapping. | |
841 | * Examine the value of chg to determine if reserves | |
842 | * actually exist or were previously consumed. | |
843 | * Very Subtle - The value of chg comes from a previous | |
844 | * call to vma_needs_reserves(). The reserve map for | |
845 | * private mappings has different (opposite) semantics | |
846 | * than that of shared mappings. vma_needs_reserves() | |
847 | * has already taken this difference in semantics into | |
848 | * account. Therefore, the meaning of chg is the same | |
849 | * as in the shared case above. Code could easily be | |
850 | * combined, but keeping it separate draws attention to | |
851 | * subtle differences. | |
852 | */ | |
853 | if (chg) | |
854 | return false; | |
855 | else | |
856 | return true; | |
857 | } | |
a63884e9 | 858 | |
559ec2f8 | 859 | return false; |
a1e78772 MG |
860 | } |
861 | ||
a5516438 | 862 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
863 | { |
864 | int nid = page_to_nid(page); | |
0edaecfa | 865 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
866 | h->free_huge_pages++; |
867 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
868 | } |
869 | ||
94310cbc | 870 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
871 | { |
872 | struct page *page; | |
873 | ||
c8721bbb | 874 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 875 | if (!PageHWPoison(page)) |
c8721bbb NH |
876 | break; |
877 | /* | |
878 | * if 'non-isolated free hugepage' not found on the list, | |
879 | * the allocation fails. | |
880 | */ | |
881 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 882 | return NULL; |
0edaecfa | 883 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 884 | set_page_refcounted(page); |
bf50bab2 NH |
885 | h->free_huge_pages--; |
886 | h->free_huge_pages_node[nid]--; | |
887 | return page; | |
888 | } | |
889 | ||
94310cbc AK |
890 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
891 | { | |
892 | struct page *page; | |
893 | int node; | |
894 | ||
895 | if (nid != NUMA_NO_NODE) | |
896 | return dequeue_huge_page_node_exact(h, nid); | |
897 | ||
898 | for_each_online_node(node) { | |
899 | page = dequeue_huge_page_node_exact(h, node); | |
900 | if (page) | |
901 | return page; | |
902 | } | |
903 | return NULL; | |
904 | } | |
905 | ||
86cdb465 NH |
906 | /* Movability of hugepages depends on migration support. */ |
907 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
908 | { | |
100873d7 | 909 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
910 | return GFP_HIGHUSER_MOVABLE; |
911 | else | |
912 | return GFP_HIGHUSER; | |
913 | } | |
914 | ||
a5516438 AK |
915 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
916 | struct vm_area_struct *vma, | |
af0ed73e JK |
917 | unsigned long address, int avoid_reserve, |
918 | long chg) | |
1da177e4 | 919 | { |
b1c12cbc | 920 | struct page *page = NULL; |
480eccf9 | 921 | struct mempolicy *mpol; |
19770b32 | 922 | nodemask_t *nodemask; |
04ec6264 VB |
923 | gfp_t gfp_mask; |
924 | int nid; | |
c0ff7453 | 925 | struct zonelist *zonelist; |
dd1a239f MG |
926 | struct zone *zone; |
927 | struct zoneref *z; | |
cc9a6c87 | 928 | unsigned int cpuset_mems_cookie; |
1da177e4 | 929 | |
a1e78772 MG |
930 | /* |
931 | * A child process with MAP_PRIVATE mappings created by their parent | |
932 | * have no page reserves. This check ensures that reservations are | |
933 | * not "stolen". The child may still get SIGKILLed | |
934 | */ | |
af0ed73e | 935 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 936 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 937 | goto err; |
a1e78772 | 938 | |
04f2cbe3 | 939 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 940 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 941 | goto err; |
04f2cbe3 | 942 | |
9966c4bb | 943 | retry_cpuset: |
d26914d1 | 944 | cpuset_mems_cookie = read_mems_allowed_begin(); |
04ec6264 VB |
945 | gfp_mask = htlb_alloc_mask(h); |
946 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
947 | zonelist = node_zonelist(nid, gfp_mask); | |
9966c4bb | 948 | |
19770b32 MG |
949 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
950 | MAX_NR_ZONES - 1, nodemask) { | |
04ec6264 | 951 | if (cpuset_zone_allowed(zone, gfp_mask)) { |
bf50bab2 NH |
952 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
953 | if (page) { | |
af0ed73e JK |
954 | if (avoid_reserve) |
955 | break; | |
956 | if (!vma_has_reserves(vma, chg)) | |
957 | break; | |
958 | ||
07443a85 | 959 | SetPagePrivate(page); |
af0ed73e | 960 | h->resv_huge_pages--; |
bf50bab2 NH |
961 | break; |
962 | } | |
3abf7afd | 963 | } |
1da177e4 | 964 | } |
cc9a6c87 | 965 | |
52cd3b07 | 966 | mpol_cond_put(mpol); |
d26914d1 | 967 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 968 | goto retry_cpuset; |
1da177e4 | 969 | return page; |
cc9a6c87 MG |
970 | |
971 | err: | |
cc9a6c87 | 972 | return NULL; |
1da177e4 LT |
973 | } |
974 | ||
1cac6f2c LC |
975 | /* |
976 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
977 | * We may have allocated or freed a huge page based on a different | |
978 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
979 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
980 | * node for alloc or free. | |
981 | */ | |
982 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
983 | { | |
0edaf86c | 984 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
985 | VM_BUG_ON(nid >= MAX_NUMNODES); |
986 | ||
987 | return nid; | |
988 | } | |
989 | ||
990 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
991 | { | |
992 | if (!node_isset(nid, *nodes_allowed)) | |
993 | nid = next_node_allowed(nid, nodes_allowed); | |
994 | return nid; | |
995 | } | |
996 | ||
997 | /* | |
998 | * returns the previously saved node ["this node"] from which to | |
999 | * allocate a persistent huge page for the pool and advance the | |
1000 | * next node from which to allocate, handling wrap at end of node | |
1001 | * mask. | |
1002 | */ | |
1003 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1004 | nodemask_t *nodes_allowed) | |
1005 | { | |
1006 | int nid; | |
1007 | ||
1008 | VM_BUG_ON(!nodes_allowed); | |
1009 | ||
1010 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1011 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1012 | ||
1013 | return nid; | |
1014 | } | |
1015 | ||
1016 | /* | |
1017 | * helper for free_pool_huge_page() - return the previously saved | |
1018 | * node ["this node"] from which to free a huge page. Advance the | |
1019 | * next node id whether or not we find a free huge page to free so | |
1020 | * that the next attempt to free addresses the next node. | |
1021 | */ | |
1022 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1023 | { | |
1024 | int nid; | |
1025 | ||
1026 | VM_BUG_ON(!nodes_allowed); | |
1027 | ||
1028 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1029 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1030 | ||
1031 | return nid; | |
1032 | } | |
1033 | ||
1034 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1035 | for (nr_nodes = nodes_weight(*mask); \ | |
1036 | nr_nodes > 0 && \ | |
1037 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1038 | nr_nodes--) | |
1039 | ||
1040 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1041 | for (nr_nodes = nodes_weight(*mask); \ | |
1042 | nr_nodes > 0 && \ | |
1043 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1044 | nr_nodes--) | |
1045 | ||
e1073d1e | 1046 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1047 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1048 | unsigned int order) |
944d9fec LC |
1049 | { |
1050 | int i; | |
1051 | int nr_pages = 1 << order; | |
1052 | struct page *p = page + 1; | |
1053 | ||
c8cc708a | 1054 | atomic_set(compound_mapcount_ptr(page), 0); |
944d9fec | 1055 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1056 | clear_compound_head(p); |
944d9fec | 1057 | set_page_refcounted(p); |
944d9fec LC |
1058 | } |
1059 | ||
1060 | set_compound_order(page, 0); | |
1061 | __ClearPageHead(page); | |
1062 | } | |
1063 | ||
d00181b9 | 1064 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1065 | { |
1066 | free_contig_range(page_to_pfn(page), 1 << order); | |
1067 | } | |
1068 | ||
1069 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
1070 | unsigned long nr_pages) | |
1071 | { | |
1072 | unsigned long end_pfn = start_pfn + nr_pages; | |
ca96b625 LS |
1073 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, |
1074 | GFP_KERNEL); | |
944d9fec LC |
1075 | } |
1076 | ||
f44b2dda JK |
1077 | static bool pfn_range_valid_gigantic(struct zone *z, |
1078 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1079 | { |
1080 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1081 | struct page *page; | |
1082 | ||
1083 | for (i = start_pfn; i < end_pfn; i++) { | |
1084 | if (!pfn_valid(i)) | |
1085 | return false; | |
1086 | ||
1087 | page = pfn_to_page(i); | |
1088 | ||
f44b2dda JK |
1089 | if (page_zone(page) != z) |
1090 | return false; | |
1091 | ||
944d9fec LC |
1092 | if (PageReserved(page)) |
1093 | return false; | |
1094 | ||
1095 | if (page_count(page) > 0) | |
1096 | return false; | |
1097 | ||
1098 | if (PageHuge(page)) | |
1099 | return false; | |
1100 | } | |
1101 | ||
1102 | return true; | |
1103 | } | |
1104 | ||
1105 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1106 | unsigned long start_pfn, unsigned long nr_pages) | |
1107 | { | |
1108 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1109 | return zone_spans_pfn(zone, last_pfn); | |
1110 | } | |
1111 | ||
d00181b9 | 1112 | static struct page *alloc_gigantic_page(int nid, unsigned int order) |
944d9fec LC |
1113 | { |
1114 | unsigned long nr_pages = 1 << order; | |
1115 | unsigned long ret, pfn, flags; | |
1116 | struct zone *z; | |
1117 | ||
1118 | z = NODE_DATA(nid)->node_zones; | |
1119 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
1120 | spin_lock_irqsave(&z->lock, flags); | |
1121 | ||
1122 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
1123 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
f44b2dda | 1124 | if (pfn_range_valid_gigantic(z, pfn, nr_pages)) { |
944d9fec LC |
1125 | /* |
1126 | * We release the zone lock here because | |
1127 | * alloc_contig_range() will also lock the zone | |
1128 | * at some point. If there's an allocation | |
1129 | * spinning on this lock, it may win the race | |
1130 | * and cause alloc_contig_range() to fail... | |
1131 | */ | |
1132 | spin_unlock_irqrestore(&z->lock, flags); | |
1133 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
1134 | if (!ret) | |
1135 | return pfn_to_page(pfn); | |
1136 | spin_lock_irqsave(&z->lock, flags); | |
1137 | } | |
1138 | pfn += nr_pages; | |
1139 | } | |
1140 | ||
1141 | spin_unlock_irqrestore(&z->lock, flags); | |
1142 | } | |
1143 | ||
1144 | return NULL; | |
1145 | } | |
1146 | ||
1147 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1148 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
944d9fec LC |
1149 | |
1150 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
1151 | { | |
1152 | struct page *page; | |
1153 | ||
1154 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
1155 | if (page) { | |
1156 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1157 | prep_new_huge_page(h, page, nid); | |
1158 | } | |
1159 | ||
1160 | return page; | |
1161 | } | |
1162 | ||
1163 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
1164 | nodemask_t *nodes_allowed) | |
1165 | { | |
1166 | struct page *page = NULL; | |
1167 | int nr_nodes, node; | |
1168 | ||
1169 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1170 | page = alloc_fresh_gigantic_page_node(h, node); | |
1171 | if (page) | |
1172 | return 1; | |
1173 | } | |
1174 | ||
1175 | return 0; | |
1176 | } | |
1177 | ||
e1073d1e | 1178 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
944d9fec | 1179 | static inline bool gigantic_page_supported(void) { return false; } |
d00181b9 | 1180 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1181 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1182 | unsigned int order) { } |
944d9fec LC |
1183 | static inline int alloc_fresh_gigantic_page(struct hstate *h, |
1184 | nodemask_t *nodes_allowed) { return 0; } | |
1185 | #endif | |
1186 | ||
a5516438 | 1187 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1188 | { |
1189 | int i; | |
a5516438 | 1190 | |
944d9fec LC |
1191 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1192 | return; | |
18229df5 | 1193 | |
a5516438 AK |
1194 | h->nr_huge_pages--; |
1195 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1196 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1197 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1198 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1199 | 1 << PG_active | 1 << PG_private | |
1200 | 1 << PG_writeback); | |
6af2acb6 | 1201 | } |
309381fe | 1202 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1203 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1204 | set_page_refcounted(page); |
944d9fec LC |
1205 | if (hstate_is_gigantic(h)) { |
1206 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1207 | free_gigantic_page(page, huge_page_order(h)); | |
1208 | } else { | |
944d9fec LC |
1209 | __free_pages(page, huge_page_order(h)); |
1210 | } | |
6af2acb6 AL |
1211 | } |
1212 | ||
e5ff2159 AK |
1213 | struct hstate *size_to_hstate(unsigned long size) |
1214 | { | |
1215 | struct hstate *h; | |
1216 | ||
1217 | for_each_hstate(h) { | |
1218 | if (huge_page_size(h) == size) | |
1219 | return h; | |
1220 | } | |
1221 | return NULL; | |
1222 | } | |
1223 | ||
bcc54222 NH |
1224 | /* |
1225 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1226 | * to hstate->hugepage_activelist.) | |
1227 | * | |
1228 | * This function can be called for tail pages, but never returns true for them. | |
1229 | */ | |
1230 | bool page_huge_active(struct page *page) | |
1231 | { | |
1232 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1233 | return PageHead(page) && PagePrivate(&page[1]); | |
1234 | } | |
1235 | ||
1236 | /* never called for tail page */ | |
1237 | static void set_page_huge_active(struct page *page) | |
1238 | { | |
1239 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1240 | SetPagePrivate(&page[1]); | |
1241 | } | |
1242 | ||
1243 | static void clear_page_huge_active(struct page *page) | |
1244 | { | |
1245 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1246 | ClearPagePrivate(&page[1]); | |
1247 | } | |
1248 | ||
8f1d26d0 | 1249 | void free_huge_page(struct page *page) |
27a85ef1 | 1250 | { |
a5516438 AK |
1251 | /* |
1252 | * Can't pass hstate in here because it is called from the | |
1253 | * compound page destructor. | |
1254 | */ | |
e5ff2159 | 1255 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1256 | int nid = page_to_nid(page); |
90481622 DG |
1257 | struct hugepage_subpool *spool = |
1258 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1259 | bool restore_reserve; |
27a85ef1 | 1260 | |
e5df70ab | 1261 | set_page_private(page, 0); |
23be7468 | 1262 | page->mapping = NULL; |
b4330afb MK |
1263 | VM_BUG_ON_PAGE(page_count(page), page); |
1264 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
07443a85 | 1265 | restore_reserve = PagePrivate(page); |
16c794b4 | 1266 | ClearPagePrivate(page); |
27a85ef1 | 1267 | |
1c5ecae3 MK |
1268 | /* |
1269 | * A return code of zero implies that the subpool will be under its | |
1270 | * minimum size if the reservation is not restored after page is free. | |
1271 | * Therefore, force restore_reserve operation. | |
1272 | */ | |
1273 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1274 | restore_reserve = true; | |
1275 | ||
27a85ef1 | 1276 | spin_lock(&hugetlb_lock); |
bcc54222 | 1277 | clear_page_huge_active(page); |
6d76dcf4 AK |
1278 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1279 | pages_per_huge_page(h), page); | |
07443a85 JK |
1280 | if (restore_reserve) |
1281 | h->resv_huge_pages++; | |
1282 | ||
944d9fec | 1283 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1284 | /* remove the page from active list */ |
1285 | list_del(&page->lru); | |
a5516438 AK |
1286 | update_and_free_page(h, page); |
1287 | h->surplus_huge_pages--; | |
1288 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1289 | } else { |
5d3a551c | 1290 | arch_clear_hugepage_flags(page); |
a5516438 | 1291 | enqueue_huge_page(h, page); |
7893d1d5 | 1292 | } |
27a85ef1 DG |
1293 | spin_unlock(&hugetlb_lock); |
1294 | } | |
1295 | ||
a5516438 | 1296 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1297 | { |
0edaecfa | 1298 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1299 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1300 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1301 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1302 | h->nr_huge_pages++; |
1303 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1304 | spin_unlock(&hugetlb_lock); |
1305 | put_page(page); /* free it into the hugepage allocator */ | |
1306 | } | |
1307 | ||
d00181b9 | 1308 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1309 | { |
1310 | int i; | |
1311 | int nr_pages = 1 << order; | |
1312 | struct page *p = page + 1; | |
1313 | ||
1314 | /* we rely on prep_new_huge_page to set the destructor */ | |
1315 | set_compound_order(page, order); | |
ef5a22be | 1316 | __ClearPageReserved(page); |
de09d31d | 1317 | __SetPageHead(page); |
20a0307c | 1318 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1319 | /* |
1320 | * For gigantic hugepages allocated through bootmem at | |
1321 | * boot, it's safer to be consistent with the not-gigantic | |
1322 | * hugepages and clear the PG_reserved bit from all tail pages | |
1323 | * too. Otherwse drivers using get_user_pages() to access tail | |
1324 | * pages may get the reference counting wrong if they see | |
1325 | * PG_reserved set on a tail page (despite the head page not | |
1326 | * having PG_reserved set). Enforcing this consistency between | |
1327 | * head and tail pages allows drivers to optimize away a check | |
1328 | * on the head page when they need know if put_page() is needed | |
1329 | * after get_user_pages(). | |
1330 | */ | |
1331 | __ClearPageReserved(p); | |
58a84aa9 | 1332 | set_page_count(p, 0); |
1d798ca3 | 1333 | set_compound_head(p, page); |
20a0307c | 1334 | } |
b4330afb | 1335 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1336 | } |
1337 | ||
7795912c AM |
1338 | /* |
1339 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1340 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1341 | * details. | |
1342 | */ | |
20a0307c WF |
1343 | int PageHuge(struct page *page) |
1344 | { | |
20a0307c WF |
1345 | if (!PageCompound(page)) |
1346 | return 0; | |
1347 | ||
1348 | page = compound_head(page); | |
f1e61557 | 1349 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1350 | } |
43131e14 NH |
1351 | EXPORT_SYMBOL_GPL(PageHuge); |
1352 | ||
27c73ae7 AA |
1353 | /* |
1354 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1355 | * normal or transparent huge pages. | |
1356 | */ | |
1357 | int PageHeadHuge(struct page *page_head) | |
1358 | { | |
27c73ae7 AA |
1359 | if (!PageHead(page_head)) |
1360 | return 0; | |
1361 | ||
758f66a2 | 1362 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1363 | } |
27c73ae7 | 1364 | |
13d60f4b ZY |
1365 | pgoff_t __basepage_index(struct page *page) |
1366 | { | |
1367 | struct page *page_head = compound_head(page); | |
1368 | pgoff_t index = page_index(page_head); | |
1369 | unsigned long compound_idx; | |
1370 | ||
1371 | if (!PageHuge(page_head)) | |
1372 | return page_index(page); | |
1373 | ||
1374 | if (compound_order(page_head) >= MAX_ORDER) | |
1375 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1376 | else | |
1377 | compound_idx = page - page_head; | |
1378 | ||
1379 | return (index << compound_order(page_head)) + compound_idx; | |
1380 | } | |
1381 | ||
a5516438 | 1382 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1383 | { |
1da177e4 | 1384 | struct page *page; |
f96efd58 | 1385 | |
96db800f | 1386 | page = __alloc_pages_node(nid, |
86cdb465 | 1387 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1388 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1389 | huge_page_order(h)); |
1da177e4 | 1390 | if (page) { |
a5516438 | 1391 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1392 | } |
63b4613c NA |
1393 | |
1394 | return page; | |
1395 | } | |
1396 | ||
b2261026 JK |
1397 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1398 | { | |
1399 | struct page *page; | |
1400 | int nr_nodes, node; | |
1401 | int ret = 0; | |
1402 | ||
1403 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1404 | page = alloc_fresh_huge_page_node(h, node); | |
1405 | if (page) { | |
1406 | ret = 1; | |
1407 | break; | |
1408 | } | |
1409 | } | |
1410 | ||
1411 | if (ret) | |
1412 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1413 | else | |
1414 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1415 | ||
1416 | return ret; | |
1417 | } | |
1418 | ||
e8c5c824 LS |
1419 | /* |
1420 | * Free huge page from pool from next node to free. | |
1421 | * Attempt to keep persistent huge pages more or less | |
1422 | * balanced over allowed nodes. | |
1423 | * Called with hugetlb_lock locked. | |
1424 | */ | |
6ae11b27 LS |
1425 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1426 | bool acct_surplus) | |
e8c5c824 | 1427 | { |
b2261026 | 1428 | int nr_nodes, node; |
e8c5c824 LS |
1429 | int ret = 0; |
1430 | ||
b2261026 | 1431 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1432 | /* |
1433 | * If we're returning unused surplus pages, only examine | |
1434 | * nodes with surplus pages. | |
1435 | */ | |
b2261026 JK |
1436 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1437 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1438 | struct page *page = |
b2261026 | 1439 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1440 | struct page, lru); |
1441 | list_del(&page->lru); | |
1442 | h->free_huge_pages--; | |
b2261026 | 1443 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1444 | if (acct_surplus) { |
1445 | h->surplus_huge_pages--; | |
b2261026 | 1446 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1447 | } |
e8c5c824 LS |
1448 | update_and_free_page(h, page); |
1449 | ret = 1; | |
9a76db09 | 1450 | break; |
e8c5c824 | 1451 | } |
b2261026 | 1452 | } |
e8c5c824 LS |
1453 | |
1454 | return ret; | |
1455 | } | |
1456 | ||
c8721bbb NH |
1457 | /* |
1458 | * Dissolve a given free hugepage into free buddy pages. This function does | |
082d5b6b GS |
1459 | * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the |
1460 | * number of free hugepages would be reduced below the number of reserved | |
1461 | * hugepages. | |
c8721bbb | 1462 | */ |
c3114a84 | 1463 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1464 | { |
082d5b6b GS |
1465 | int rc = 0; |
1466 | ||
c8721bbb NH |
1467 | spin_lock(&hugetlb_lock); |
1468 | if (PageHuge(page) && !page_count(page)) { | |
2247bb33 GS |
1469 | struct page *head = compound_head(page); |
1470 | struct hstate *h = page_hstate(head); | |
1471 | int nid = page_to_nid(head); | |
082d5b6b GS |
1472 | if (h->free_huge_pages - h->resv_huge_pages == 0) { |
1473 | rc = -EBUSY; | |
1474 | goto out; | |
1475 | } | |
c3114a84 AK |
1476 | /* |
1477 | * Move PageHWPoison flag from head page to the raw error page, | |
1478 | * which makes any subpages rather than the error page reusable. | |
1479 | */ | |
1480 | if (PageHWPoison(head) && page != head) { | |
1481 | SetPageHWPoison(page); | |
1482 | ClearPageHWPoison(head); | |
1483 | } | |
2247bb33 | 1484 | list_del(&head->lru); |
c8721bbb NH |
1485 | h->free_huge_pages--; |
1486 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1487 | h->max_huge_pages--; |
2247bb33 | 1488 | update_and_free_page(h, head); |
c8721bbb | 1489 | } |
082d5b6b | 1490 | out: |
c8721bbb | 1491 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1492 | return rc; |
c8721bbb NH |
1493 | } |
1494 | ||
1495 | /* | |
1496 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1497 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1498 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1499 | * part of it lies within the given range. | |
082d5b6b GS |
1500 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1501 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1502 | */ |
082d5b6b | 1503 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1504 | { |
c8721bbb | 1505 | unsigned long pfn; |
eb03aa00 | 1506 | struct page *page; |
082d5b6b | 1507 | int rc = 0; |
c8721bbb | 1508 | |
d0177639 | 1509 | if (!hugepages_supported()) |
082d5b6b | 1510 | return rc; |
d0177639 | 1511 | |
eb03aa00 GS |
1512 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1513 | page = pfn_to_page(pfn); | |
1514 | if (PageHuge(page) && !page_count(page)) { | |
1515 | rc = dissolve_free_huge_page(page); | |
1516 | if (rc) | |
1517 | break; | |
1518 | } | |
1519 | } | |
082d5b6b GS |
1520 | |
1521 | return rc; | |
c8721bbb NH |
1522 | } |
1523 | ||
099730d6 DH |
1524 | /* |
1525 | * There are 3 ways this can get called: | |
1526 | * 1. With vma+addr: we use the VMA's memory policy | |
1527 | * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge | |
1528 | * page from any node, and let the buddy allocator itself figure | |
1529 | * it out. | |
1530 | * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page | |
1531 | * strictly from 'nid' | |
1532 | */ | |
1533 | static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, | |
1534 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
1535 | { | |
1536 | int order = huge_page_order(h); | |
1537 | gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; | |
1538 | unsigned int cpuset_mems_cookie; | |
1539 | ||
1540 | /* | |
1541 | * We need a VMA to get a memory policy. If we do not | |
e0ec90ee DH |
1542 | * have one, we use the 'nid' argument. |
1543 | * | |
1544 | * The mempolicy stuff below has some non-inlined bits | |
1545 | * and calls ->vm_ops. That makes it hard to optimize at | |
1546 | * compile-time, even when NUMA is off and it does | |
1547 | * nothing. This helps the compiler optimize it out. | |
099730d6 | 1548 | */ |
e0ec90ee | 1549 | if (!IS_ENABLED(CONFIG_NUMA) || !vma) { |
099730d6 DH |
1550 | /* |
1551 | * If a specific node is requested, make sure to | |
1552 | * get memory from there, but only when a node | |
1553 | * is explicitly specified. | |
1554 | */ | |
1555 | if (nid != NUMA_NO_NODE) | |
1556 | gfp |= __GFP_THISNODE; | |
1557 | /* | |
1558 | * Make sure to call something that can handle | |
1559 | * nid=NUMA_NO_NODE | |
1560 | */ | |
1561 | return alloc_pages_node(nid, gfp, order); | |
1562 | } | |
1563 | ||
1564 | /* | |
1565 | * OK, so we have a VMA. Fetch the mempolicy and try to | |
e0ec90ee DH |
1566 | * allocate a huge page with it. We will only reach this |
1567 | * when CONFIG_NUMA=y. | |
099730d6 DH |
1568 | */ |
1569 | do { | |
1570 | struct page *page; | |
1571 | struct mempolicy *mpol; | |
04ec6264 | 1572 | int nid; |
099730d6 DH |
1573 | nodemask_t *nodemask; |
1574 | ||
1575 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
04ec6264 | 1576 | nid = huge_node(vma, addr, gfp, &mpol, &nodemask); |
099730d6 | 1577 | mpol_cond_put(mpol); |
04ec6264 | 1578 | page = __alloc_pages_nodemask(gfp, order, nid, nodemask); |
099730d6 DH |
1579 | if (page) |
1580 | return page; | |
1581 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); | |
1582 | ||
1583 | return NULL; | |
1584 | } | |
1585 | ||
1586 | /* | |
1587 | * There are two ways to allocate a huge page: | |
1588 | * 1. When you have a VMA and an address (like a fault) | |
1589 | * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) | |
1590 | * | |
1591 | * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in | |
1592 | * this case which signifies that the allocation should be done with | |
1593 | * respect for the VMA's memory policy. | |
1594 | * | |
1595 | * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This | |
1596 | * implies that memory policies will not be taken in to account. | |
1597 | */ | |
1598 | static struct page *__alloc_buddy_huge_page(struct hstate *h, | |
1599 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
7893d1d5 AL |
1600 | { |
1601 | struct page *page; | |
bf50bab2 | 1602 | unsigned int r_nid; |
7893d1d5 | 1603 | |
bae7f4ae | 1604 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1605 | return NULL; |
1606 | ||
099730d6 DH |
1607 | /* |
1608 | * Make sure that anyone specifying 'nid' is not also specifying a VMA. | |
1609 | * This makes sure the caller is picking _one_ of the modes with which | |
1610 | * we can call this function, not both. | |
1611 | */ | |
1612 | if (vma || (addr != -1)) { | |
e0ec90ee DH |
1613 | VM_WARN_ON_ONCE(addr == -1); |
1614 | VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); | |
099730d6 | 1615 | } |
d1c3fb1f NA |
1616 | /* |
1617 | * Assume we will successfully allocate the surplus page to | |
1618 | * prevent racing processes from causing the surplus to exceed | |
1619 | * overcommit | |
1620 | * | |
1621 | * This however introduces a different race, where a process B | |
1622 | * tries to grow the static hugepage pool while alloc_pages() is | |
1623 | * called by process A. B will only examine the per-node | |
1624 | * counters in determining if surplus huge pages can be | |
1625 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1626 | * won't be able to increment the per-node counter, until the | |
1627 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1628 | * no more huge pages can be converted from surplus to normal | |
1629 | * state (and doesn't try to convert again). Thus, we have a | |
1630 | * case where a surplus huge page exists, the pool is grown, and | |
1631 | * the surplus huge page still exists after, even though it | |
1632 | * should just have been converted to a normal huge page. This | |
1633 | * does not leak memory, though, as the hugepage will be freed | |
1634 | * once it is out of use. It also does not allow the counters to | |
1635 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1636 | * the node values until we've gotten the hugepage and only the | |
1637 | * per-node value is checked there. | |
1638 | */ | |
1639 | spin_lock(&hugetlb_lock); | |
a5516438 | 1640 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1641 | spin_unlock(&hugetlb_lock); |
1642 | return NULL; | |
1643 | } else { | |
a5516438 AK |
1644 | h->nr_huge_pages++; |
1645 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1646 | } |
1647 | spin_unlock(&hugetlb_lock); | |
1648 | ||
099730d6 | 1649 | page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); |
d1c3fb1f NA |
1650 | |
1651 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 1652 | if (page) { |
0edaecfa | 1653 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1654 | r_nid = page_to_nid(page); |
f1e61557 | 1655 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
9dd540e2 | 1656 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1657 | /* |
1658 | * We incremented the global counters already | |
1659 | */ | |
bf50bab2 NH |
1660 | h->nr_huge_pages_node[r_nid]++; |
1661 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1662 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1663 | } else { |
a5516438 AK |
1664 | h->nr_huge_pages--; |
1665 | h->surplus_huge_pages--; | |
3b116300 | 1666 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1667 | } |
d1c3fb1f | 1668 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1669 | |
1670 | return page; | |
1671 | } | |
1672 | ||
099730d6 DH |
1673 | /* |
1674 | * Allocate a huge page from 'nid'. Note, 'nid' may be | |
1675 | * NUMA_NO_NODE, which means that it may be allocated | |
1676 | * anywhere. | |
1677 | */ | |
e0ec90ee | 1678 | static |
099730d6 DH |
1679 | struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) |
1680 | { | |
1681 | unsigned long addr = -1; | |
1682 | ||
1683 | return __alloc_buddy_huge_page(h, NULL, addr, nid); | |
1684 | } | |
1685 | ||
1686 | /* | |
1687 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1688 | */ | |
e0ec90ee | 1689 | static |
099730d6 DH |
1690 | struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, |
1691 | struct vm_area_struct *vma, unsigned long addr) | |
1692 | { | |
1693 | return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); | |
1694 | } | |
1695 | ||
bf50bab2 NH |
1696 | /* |
1697 | * This allocation function is useful in the context where vma is irrelevant. | |
1698 | * E.g. soft-offlining uses this function because it only cares physical | |
1699 | * address of error page. | |
1700 | */ | |
1701 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1702 | { | |
4ef91848 | 1703 | struct page *page = NULL; |
bf50bab2 NH |
1704 | |
1705 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1706 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1707 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1708 | spin_unlock(&hugetlb_lock); |
1709 | ||
94ae8ba7 | 1710 | if (!page) |
099730d6 | 1711 | page = __alloc_buddy_huge_page_no_mpol(h, nid); |
bf50bab2 NH |
1712 | |
1713 | return page; | |
1714 | } | |
1715 | ||
4db9b2ef MH |
1716 | struct page *alloc_huge_page_nodemask(struct hstate *h, const nodemask_t *nmask) |
1717 | { | |
1718 | struct page *page = NULL; | |
1719 | int node; | |
1720 | ||
1721 | spin_lock(&hugetlb_lock); | |
1722 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
1723 | for_each_node_mask(node, *nmask) { | |
1724 | page = dequeue_huge_page_node_exact(h, node); | |
1725 | if (page) | |
1726 | break; | |
1727 | } | |
1728 | } | |
1729 | spin_unlock(&hugetlb_lock); | |
1730 | if (page) | |
1731 | return page; | |
1732 | ||
1733 | /* No reservations, try to overcommit */ | |
1734 | for_each_node_mask(node, *nmask) { | |
1735 | page = __alloc_buddy_huge_page_no_mpol(h, node); | |
1736 | if (page) | |
1737 | return page; | |
1738 | } | |
1739 | ||
1740 | return NULL; | |
1741 | } | |
1742 | ||
e4e574b7 | 1743 | /* |
25985edc | 1744 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1745 | * of size 'delta'. |
1746 | */ | |
a5516438 | 1747 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1748 | { |
1749 | struct list_head surplus_list; | |
1750 | struct page *page, *tmp; | |
1751 | int ret, i; | |
1752 | int needed, allocated; | |
28073b02 | 1753 | bool alloc_ok = true; |
e4e574b7 | 1754 | |
a5516438 | 1755 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1756 | if (needed <= 0) { |
a5516438 | 1757 | h->resv_huge_pages += delta; |
e4e574b7 | 1758 | return 0; |
ac09b3a1 | 1759 | } |
e4e574b7 AL |
1760 | |
1761 | allocated = 0; | |
1762 | INIT_LIST_HEAD(&surplus_list); | |
1763 | ||
1764 | ret = -ENOMEM; | |
1765 | retry: | |
1766 | spin_unlock(&hugetlb_lock); | |
1767 | for (i = 0; i < needed; i++) { | |
099730d6 | 1768 | page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); |
28073b02 HD |
1769 | if (!page) { |
1770 | alloc_ok = false; | |
1771 | break; | |
1772 | } | |
e4e574b7 | 1773 | list_add(&page->lru, &surplus_list); |
69ed779a | 1774 | cond_resched(); |
e4e574b7 | 1775 | } |
28073b02 | 1776 | allocated += i; |
e4e574b7 AL |
1777 | |
1778 | /* | |
1779 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1780 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1781 | */ | |
1782 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1783 | needed = (h->resv_huge_pages + delta) - |
1784 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1785 | if (needed > 0) { |
1786 | if (alloc_ok) | |
1787 | goto retry; | |
1788 | /* | |
1789 | * We were not able to allocate enough pages to | |
1790 | * satisfy the entire reservation so we free what | |
1791 | * we've allocated so far. | |
1792 | */ | |
1793 | goto free; | |
1794 | } | |
e4e574b7 AL |
1795 | /* |
1796 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1797 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1798 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1799 | * allocator. Commit the entire reservation here to prevent another |
1800 | * process from stealing the pages as they are added to the pool but | |
1801 | * before they are reserved. | |
e4e574b7 AL |
1802 | */ |
1803 | needed += allocated; | |
a5516438 | 1804 | h->resv_huge_pages += delta; |
e4e574b7 | 1805 | ret = 0; |
a9869b83 | 1806 | |
19fc3f0a | 1807 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1808 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1809 | if ((--needed) < 0) |
1810 | break; | |
a9869b83 NH |
1811 | /* |
1812 | * This page is now managed by the hugetlb allocator and has | |
1813 | * no users -- drop the buddy allocator's reference. | |
1814 | */ | |
1815 | put_page_testzero(page); | |
309381fe | 1816 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1817 | enqueue_huge_page(h, page); |
19fc3f0a | 1818 | } |
28073b02 | 1819 | free: |
b0365c8d | 1820 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1821 | |
1822 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1823 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1824 | put_page(page); | |
a9869b83 | 1825 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1826 | |
1827 | return ret; | |
1828 | } | |
1829 | ||
1830 | /* | |
e5bbc8a6 MK |
1831 | * This routine has two main purposes: |
1832 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1833 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1834 | * to the associated reservation map. | |
1835 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1836 | * the reservation. As many as unused_resv_pages may be freed. | |
1837 | * | |
1838 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1839 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1840 | * we must make sure nobody else can claim pages we are in the process of | |
1841 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1842 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1843 | */ |
a5516438 AK |
1844 | static void return_unused_surplus_pages(struct hstate *h, |
1845 | unsigned long unused_resv_pages) | |
e4e574b7 | 1846 | { |
e4e574b7 AL |
1847 | unsigned long nr_pages; |
1848 | ||
aa888a74 | 1849 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1850 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1851 | goto out; |
aa888a74 | 1852 | |
e5bbc8a6 MK |
1853 | /* |
1854 | * Part (or even all) of the reservation could have been backed | |
1855 | * by pre-allocated pages. Only free surplus pages. | |
1856 | */ | |
a5516438 | 1857 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1858 | |
685f3457 LS |
1859 | /* |
1860 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1861 | * evenly across all nodes with memory. Iterate across these nodes |
1862 | * until we can no longer free unreserved surplus pages. This occurs | |
1863 | * when the nodes with surplus pages have no free pages. | |
1864 | * free_pool_huge_page() will balance the the freed pages across the | |
1865 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1866 | * |
1867 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1868 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1869 | * to cover subsequent pages we may free. | |
685f3457 LS |
1870 | */ |
1871 | while (nr_pages--) { | |
e5bbc8a6 MK |
1872 | h->resv_huge_pages--; |
1873 | unused_resv_pages--; | |
8cebfcd0 | 1874 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1875 | goto out; |
7848a4bf | 1876 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1877 | } |
e5bbc8a6 MK |
1878 | |
1879 | out: | |
1880 | /* Fully uncommit the reservation */ | |
1881 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1882 | } |
1883 | ||
5e911373 | 1884 | |
c37f9fb1 | 1885 | /* |
feba16e2 | 1886 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1887 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1888 | * |
1889 | * vma_needs_reservation is called to determine if the huge page at addr | |
1890 | * within the vma has an associated reservation. If a reservation is | |
1891 | * needed, the value 1 is returned. The caller is then responsible for | |
1892 | * managing the global reservation and subpool usage counts. After | |
1893 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1894 | * to add the page to the reservation map. If the page allocation fails, |
1895 | * the reservation must be ended instead of committed. vma_end_reservation | |
1896 | * is called in such cases. | |
cf3ad20b MK |
1897 | * |
1898 | * In the normal case, vma_commit_reservation returns the same value | |
1899 | * as the preceding vma_needs_reservation call. The only time this | |
1900 | * is not the case is if a reserve map was changed between calls. It | |
1901 | * is the responsibility of the caller to notice the difference and | |
1902 | * take appropriate action. | |
96b96a96 MK |
1903 | * |
1904 | * vma_add_reservation is used in error paths where a reservation must | |
1905 | * be restored when a newly allocated huge page must be freed. It is | |
1906 | * to be called after calling vma_needs_reservation to determine if a | |
1907 | * reservation exists. | |
c37f9fb1 | 1908 | */ |
5e911373 MK |
1909 | enum vma_resv_mode { |
1910 | VMA_NEEDS_RESV, | |
1911 | VMA_COMMIT_RESV, | |
feba16e2 | 1912 | VMA_END_RESV, |
96b96a96 | 1913 | VMA_ADD_RESV, |
5e911373 | 1914 | }; |
cf3ad20b MK |
1915 | static long __vma_reservation_common(struct hstate *h, |
1916 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1917 | enum vma_resv_mode mode) |
c37f9fb1 | 1918 | { |
4e35f483 JK |
1919 | struct resv_map *resv; |
1920 | pgoff_t idx; | |
cf3ad20b | 1921 | long ret; |
c37f9fb1 | 1922 | |
4e35f483 JK |
1923 | resv = vma_resv_map(vma); |
1924 | if (!resv) | |
84afd99b | 1925 | return 1; |
c37f9fb1 | 1926 | |
4e35f483 | 1927 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1928 | switch (mode) { |
1929 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1930 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1931 | break; |
1932 | case VMA_COMMIT_RESV: | |
1933 | ret = region_add(resv, idx, idx + 1); | |
1934 | break; | |
feba16e2 | 1935 | case VMA_END_RESV: |
5e911373 MK |
1936 | region_abort(resv, idx, idx + 1); |
1937 | ret = 0; | |
1938 | break; | |
96b96a96 MK |
1939 | case VMA_ADD_RESV: |
1940 | if (vma->vm_flags & VM_MAYSHARE) | |
1941 | ret = region_add(resv, idx, idx + 1); | |
1942 | else { | |
1943 | region_abort(resv, idx, idx + 1); | |
1944 | ret = region_del(resv, idx, idx + 1); | |
1945 | } | |
1946 | break; | |
5e911373 MK |
1947 | default: |
1948 | BUG(); | |
1949 | } | |
84afd99b | 1950 | |
4e35f483 | 1951 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1952 | return ret; |
67961f9d MK |
1953 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1954 | /* | |
1955 | * In most cases, reserves always exist for private mappings. | |
1956 | * However, a file associated with mapping could have been | |
1957 | * hole punched or truncated after reserves were consumed. | |
1958 | * As subsequent fault on such a range will not use reserves. | |
1959 | * Subtle - The reserve map for private mappings has the | |
1960 | * opposite meaning than that of shared mappings. If NO | |
1961 | * entry is in the reserve map, it means a reservation exists. | |
1962 | * If an entry exists in the reserve map, it means the | |
1963 | * reservation has already been consumed. As a result, the | |
1964 | * return value of this routine is the opposite of the | |
1965 | * value returned from reserve map manipulation routines above. | |
1966 | */ | |
1967 | if (ret) | |
1968 | return 0; | |
1969 | else | |
1970 | return 1; | |
1971 | } | |
4e35f483 | 1972 | else |
cf3ad20b | 1973 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1974 | } |
cf3ad20b MK |
1975 | |
1976 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1977 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1978 | { |
5e911373 | 1979 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1980 | } |
84afd99b | 1981 | |
cf3ad20b MK |
1982 | static long vma_commit_reservation(struct hstate *h, |
1983 | struct vm_area_struct *vma, unsigned long addr) | |
1984 | { | |
5e911373 MK |
1985 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1986 | } | |
1987 | ||
feba16e2 | 1988 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1989 | struct vm_area_struct *vma, unsigned long addr) |
1990 | { | |
feba16e2 | 1991 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1992 | } |
1993 | ||
96b96a96 MK |
1994 | static long vma_add_reservation(struct hstate *h, |
1995 | struct vm_area_struct *vma, unsigned long addr) | |
1996 | { | |
1997 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
1998 | } | |
1999 | ||
2000 | /* | |
2001 | * This routine is called to restore a reservation on error paths. In the | |
2002 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2003 | * and is about to be freed. If a reservation for the page existed, | |
2004 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2005 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2006 | * the global reservation count will be incremented if PagePrivate is set. | |
2007 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2008 | * reserve map here to be consistent with global reserve count adjustments | |
2009 | * to be made by free_huge_page. | |
2010 | */ | |
2011 | static void restore_reserve_on_error(struct hstate *h, | |
2012 | struct vm_area_struct *vma, unsigned long address, | |
2013 | struct page *page) | |
2014 | { | |
2015 | if (unlikely(PagePrivate(page))) { | |
2016 | long rc = vma_needs_reservation(h, vma, address); | |
2017 | ||
2018 | if (unlikely(rc < 0)) { | |
2019 | /* | |
2020 | * Rare out of memory condition in reserve map | |
2021 | * manipulation. Clear PagePrivate so that | |
2022 | * global reserve count will not be incremented | |
2023 | * by free_huge_page. This will make it appear | |
2024 | * as though the reservation for this page was | |
2025 | * consumed. This may prevent the task from | |
2026 | * faulting in the page at a later time. This | |
2027 | * is better than inconsistent global huge page | |
2028 | * accounting of reserve counts. | |
2029 | */ | |
2030 | ClearPagePrivate(page); | |
2031 | } else if (rc) { | |
2032 | rc = vma_add_reservation(h, vma, address); | |
2033 | if (unlikely(rc < 0)) | |
2034 | /* | |
2035 | * See above comment about rare out of | |
2036 | * memory condition. | |
2037 | */ | |
2038 | ClearPagePrivate(page); | |
2039 | } else | |
2040 | vma_end_reservation(h, vma, address); | |
2041 | } | |
2042 | } | |
2043 | ||
70c3547e | 2044 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2045 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2046 | { |
90481622 | 2047 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2048 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2049 | struct page *page; |
d85f69b0 MK |
2050 | long map_chg, map_commit; |
2051 | long gbl_chg; | |
6d76dcf4 AK |
2052 | int ret, idx; |
2053 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 2054 | |
6d76dcf4 | 2055 | idx = hstate_index(h); |
a1e78772 | 2056 | /* |
d85f69b0 MK |
2057 | * Examine the region/reserve map to determine if the process |
2058 | * has a reservation for the page to be allocated. A return | |
2059 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2060 | */ |
d85f69b0 MK |
2061 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2062 | if (map_chg < 0) | |
76dcee75 | 2063 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2064 | |
2065 | /* | |
2066 | * Processes that did not create the mapping will have no | |
2067 | * reserves as indicated by the region/reserve map. Check | |
2068 | * that the allocation will not exceed the subpool limit. | |
2069 | * Allocations for MAP_NORESERVE mappings also need to be | |
2070 | * checked against any subpool limit. | |
2071 | */ | |
2072 | if (map_chg || avoid_reserve) { | |
2073 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2074 | if (gbl_chg < 0) { | |
feba16e2 | 2075 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2076 | return ERR_PTR(-ENOSPC); |
5e911373 | 2077 | } |
1da177e4 | 2078 | |
d85f69b0 MK |
2079 | /* |
2080 | * Even though there was no reservation in the region/reserve | |
2081 | * map, there could be reservations associated with the | |
2082 | * subpool that can be used. This would be indicated if the | |
2083 | * return value of hugepage_subpool_get_pages() is zero. | |
2084 | * However, if avoid_reserve is specified we still avoid even | |
2085 | * the subpool reservations. | |
2086 | */ | |
2087 | if (avoid_reserve) | |
2088 | gbl_chg = 1; | |
2089 | } | |
2090 | ||
6d76dcf4 | 2091 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2092 | if (ret) |
2093 | goto out_subpool_put; | |
2094 | ||
1da177e4 | 2095 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2096 | /* |
2097 | * glb_chg is passed to indicate whether or not a page must be taken | |
2098 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2099 | * a reservation exists for the allocation. | |
2100 | */ | |
2101 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2102 | if (!page) { |
94ae8ba7 | 2103 | spin_unlock(&hugetlb_lock); |
099730d6 | 2104 | page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2105 | if (!page) |
2106 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2107 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2108 | SetPagePrivate(page); | |
2109 | h->resv_huge_pages--; | |
2110 | } | |
79dbb236 AK |
2111 | spin_lock(&hugetlb_lock); |
2112 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2113 | /* Fall through */ |
68842c9b | 2114 | } |
81a6fcae JK |
2115 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2116 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2117 | |
90481622 | 2118 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2119 | |
d85f69b0 MK |
2120 | map_commit = vma_commit_reservation(h, vma, addr); |
2121 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2122 | /* |
2123 | * The page was added to the reservation map between | |
2124 | * vma_needs_reservation and vma_commit_reservation. | |
2125 | * This indicates a race with hugetlb_reserve_pages. | |
2126 | * Adjust for the subpool count incremented above AND | |
2127 | * in hugetlb_reserve_pages for the same page. Also, | |
2128 | * the reservation count added in hugetlb_reserve_pages | |
2129 | * no longer applies. | |
2130 | */ | |
2131 | long rsv_adjust; | |
2132 | ||
2133 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2134 | hugetlb_acct_memory(h, -rsv_adjust); | |
2135 | } | |
90d8b7e6 | 2136 | return page; |
8f34af6f JZ |
2137 | |
2138 | out_uncharge_cgroup: | |
2139 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2140 | out_subpool_put: | |
d85f69b0 | 2141 | if (map_chg || avoid_reserve) |
8f34af6f | 2142 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2143 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2144 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2145 | } |
2146 | ||
74060e4d NH |
2147 | /* |
2148 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
2149 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
2150 | * where no ERR_VALUE is expected to be returned. | |
2151 | */ | |
2152 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
2153 | unsigned long addr, int avoid_reserve) | |
2154 | { | |
2155 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
2156 | if (IS_ERR(page)) | |
2157 | page = NULL; | |
2158 | return page; | |
2159 | } | |
2160 | ||
91f47662 | 2161 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
2162 | { |
2163 | struct huge_bootmem_page *m; | |
b2261026 | 2164 | int nr_nodes, node; |
aa888a74 | 2165 | |
b2261026 | 2166 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2167 | void *addr; |
2168 | ||
8b89a116 GS |
2169 | addr = memblock_virt_alloc_try_nid_nopanic( |
2170 | huge_page_size(h), huge_page_size(h), | |
2171 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
2172 | if (addr) { |
2173 | /* | |
2174 | * Use the beginning of the huge page to store the | |
2175 | * huge_bootmem_page struct (until gather_bootmem | |
2176 | * puts them into the mem_map). | |
2177 | */ | |
2178 | m = addr; | |
91f47662 | 2179 | goto found; |
aa888a74 | 2180 | } |
aa888a74 AK |
2181 | } |
2182 | return 0; | |
2183 | ||
2184 | found: | |
df994ead | 2185 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
2186 | /* Put them into a private list first because mem_map is not up yet */ |
2187 | list_add(&m->list, &huge_boot_pages); | |
2188 | m->hstate = h; | |
2189 | return 1; | |
2190 | } | |
2191 | ||
d00181b9 KS |
2192 | static void __init prep_compound_huge_page(struct page *page, |
2193 | unsigned int order) | |
18229df5 AW |
2194 | { |
2195 | if (unlikely(order > (MAX_ORDER - 1))) | |
2196 | prep_compound_gigantic_page(page, order); | |
2197 | else | |
2198 | prep_compound_page(page, order); | |
2199 | } | |
2200 | ||
aa888a74 AK |
2201 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2202 | static void __init gather_bootmem_prealloc(void) | |
2203 | { | |
2204 | struct huge_bootmem_page *m; | |
2205 | ||
2206 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 2207 | struct hstate *h = m->hstate; |
ee8f248d BB |
2208 | struct page *page; |
2209 | ||
2210 | #ifdef CONFIG_HIGHMEM | |
2211 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
2212 | memblock_free_late(__pa(m), |
2213 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
2214 | #else |
2215 | page = virt_to_page(m); | |
2216 | #endif | |
aa888a74 | 2217 | WARN_ON(page_count(page) != 1); |
18229df5 | 2218 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2219 | WARN_ON(PageReserved(page)); |
aa888a74 | 2220 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
2221 | /* |
2222 | * If we had gigantic hugepages allocated at boot time, we need | |
2223 | * to restore the 'stolen' pages to totalram_pages in order to | |
2224 | * fix confusing memory reports from free(1) and another | |
2225 | * side-effects, like CommitLimit going negative. | |
2226 | */ | |
bae7f4ae | 2227 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2228 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
2229 | } |
2230 | } | |
2231 | ||
8faa8b07 | 2232 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2233 | { |
2234 | unsigned long i; | |
a5516438 | 2235 | |
e5ff2159 | 2236 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2237 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2238 | if (!alloc_bootmem_huge_page(h)) |
2239 | break; | |
9b5e5d0f | 2240 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 2241 | &node_states[N_MEMORY])) |
1da177e4 | 2242 | break; |
69ed779a | 2243 | cond_resched(); |
1da177e4 | 2244 | } |
d715cf80 LH |
2245 | if (i < h->max_huge_pages) { |
2246 | char buf[32]; | |
2247 | ||
c6247f72 | 2248 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2249 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2250 | h->max_huge_pages, buf, i); | |
2251 | h->max_huge_pages = i; | |
2252 | } | |
e5ff2159 AK |
2253 | } |
2254 | ||
2255 | static void __init hugetlb_init_hstates(void) | |
2256 | { | |
2257 | struct hstate *h; | |
2258 | ||
2259 | for_each_hstate(h) { | |
641844f5 NH |
2260 | if (minimum_order > huge_page_order(h)) |
2261 | minimum_order = huge_page_order(h); | |
2262 | ||
8faa8b07 | 2263 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2264 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2265 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2266 | } |
641844f5 | 2267 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2268 | } |
2269 | ||
2270 | static void __init report_hugepages(void) | |
2271 | { | |
2272 | struct hstate *h; | |
2273 | ||
2274 | for_each_hstate(h) { | |
4abd32db | 2275 | char buf[32]; |
c6247f72 MW |
2276 | |
2277 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2278 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2279 | buf, h->free_huge_pages); |
e5ff2159 AK |
2280 | } |
2281 | } | |
2282 | ||
1da177e4 | 2283 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2284 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2285 | nodemask_t *nodes_allowed) | |
1da177e4 | 2286 | { |
4415cc8d CL |
2287 | int i; |
2288 | ||
bae7f4ae | 2289 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2290 | return; |
2291 | ||
6ae11b27 | 2292 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2293 | struct page *page, *next; |
a5516438 AK |
2294 | struct list_head *freel = &h->hugepage_freelists[i]; |
2295 | list_for_each_entry_safe(page, next, freel, lru) { | |
2296 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2297 | return; |
1da177e4 LT |
2298 | if (PageHighMem(page)) |
2299 | continue; | |
2300 | list_del(&page->lru); | |
e5ff2159 | 2301 | update_and_free_page(h, page); |
a5516438 AK |
2302 | h->free_huge_pages--; |
2303 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2304 | } |
2305 | } | |
2306 | } | |
2307 | #else | |
6ae11b27 LS |
2308 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2309 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2310 | { |
2311 | } | |
2312 | #endif | |
2313 | ||
20a0307c WF |
2314 | /* |
2315 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2316 | * balanced by operating on them in a round-robin fashion. | |
2317 | * Returns 1 if an adjustment was made. | |
2318 | */ | |
6ae11b27 LS |
2319 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2320 | int delta) | |
20a0307c | 2321 | { |
b2261026 | 2322 | int nr_nodes, node; |
20a0307c WF |
2323 | |
2324 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2325 | |
b2261026 JK |
2326 | if (delta < 0) { |
2327 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2328 | if (h->surplus_huge_pages_node[node]) | |
2329 | goto found; | |
e8c5c824 | 2330 | } |
b2261026 JK |
2331 | } else { |
2332 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2333 | if (h->surplus_huge_pages_node[node] < | |
2334 | h->nr_huge_pages_node[node]) | |
2335 | goto found; | |
e8c5c824 | 2336 | } |
b2261026 JK |
2337 | } |
2338 | return 0; | |
20a0307c | 2339 | |
b2261026 JK |
2340 | found: |
2341 | h->surplus_huge_pages += delta; | |
2342 | h->surplus_huge_pages_node[node] += delta; | |
2343 | return 1; | |
20a0307c WF |
2344 | } |
2345 | ||
a5516438 | 2346 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2347 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2348 | nodemask_t *nodes_allowed) | |
1da177e4 | 2349 | { |
7893d1d5 | 2350 | unsigned long min_count, ret; |
1da177e4 | 2351 | |
944d9fec | 2352 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2353 | return h->max_huge_pages; |
2354 | ||
7893d1d5 AL |
2355 | /* |
2356 | * Increase the pool size | |
2357 | * First take pages out of surplus state. Then make up the | |
2358 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2359 | * |
d15c7c09 | 2360 | * We might race with __alloc_buddy_huge_page() here and be unable |
d1c3fb1f NA |
2361 | * to convert a surplus huge page to a normal huge page. That is |
2362 | * not critical, though, it just means the overall size of the | |
2363 | * pool might be one hugepage larger than it needs to be, but | |
2364 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2365 | */ |
1da177e4 | 2366 | spin_lock(&hugetlb_lock); |
a5516438 | 2367 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2368 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2369 | break; |
2370 | } | |
2371 | ||
a5516438 | 2372 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2373 | /* |
2374 | * If this allocation races such that we no longer need the | |
2375 | * page, free_huge_page will handle it by freeing the page | |
2376 | * and reducing the surplus. | |
2377 | */ | |
2378 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2379 | |
2380 | /* yield cpu to avoid soft lockup */ | |
2381 | cond_resched(); | |
2382 | ||
944d9fec LC |
2383 | if (hstate_is_gigantic(h)) |
2384 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
2385 | else | |
2386 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
2387 | spin_lock(&hugetlb_lock); |
2388 | if (!ret) | |
2389 | goto out; | |
2390 | ||
536240f2 MG |
2391 | /* Bail for signals. Probably ctrl-c from user */ |
2392 | if (signal_pending(current)) | |
2393 | goto out; | |
7893d1d5 | 2394 | } |
7893d1d5 AL |
2395 | |
2396 | /* | |
2397 | * Decrease the pool size | |
2398 | * First return free pages to the buddy allocator (being careful | |
2399 | * to keep enough around to satisfy reservations). Then place | |
2400 | * pages into surplus state as needed so the pool will shrink | |
2401 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2402 | * |
2403 | * By placing pages into the surplus state independent of the | |
2404 | * overcommit value, we are allowing the surplus pool size to | |
2405 | * exceed overcommit. There are few sane options here. Since | |
d15c7c09 | 2406 | * __alloc_buddy_huge_page() is checking the global counter, |
d1c3fb1f NA |
2407 | * though, we'll note that we're not allowed to exceed surplus |
2408 | * and won't grow the pool anywhere else. Not until one of the | |
2409 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2410 | */ |
a5516438 | 2411 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2412 | min_count = max(count, min_count); |
6ae11b27 | 2413 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2414 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2415 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2416 | break; |
55f67141 | 2417 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2418 | } |
a5516438 | 2419 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2420 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2421 | break; |
2422 | } | |
2423 | out: | |
a5516438 | 2424 | ret = persistent_huge_pages(h); |
1da177e4 | 2425 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2426 | return ret; |
1da177e4 LT |
2427 | } |
2428 | ||
a3437870 NA |
2429 | #define HSTATE_ATTR_RO(_name) \ |
2430 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2431 | ||
2432 | #define HSTATE_ATTR(_name) \ | |
2433 | static struct kobj_attribute _name##_attr = \ | |
2434 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2435 | ||
2436 | static struct kobject *hugepages_kobj; | |
2437 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2438 | ||
9a305230 LS |
2439 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2440 | ||
2441 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2442 | { |
2443 | int i; | |
9a305230 | 2444 | |
a3437870 | 2445 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2446 | if (hstate_kobjs[i] == kobj) { |
2447 | if (nidp) | |
2448 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2449 | return &hstates[i]; |
9a305230 LS |
2450 | } |
2451 | ||
2452 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2453 | } |
2454 | ||
06808b08 | 2455 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2456 | struct kobj_attribute *attr, char *buf) |
2457 | { | |
9a305230 LS |
2458 | struct hstate *h; |
2459 | unsigned long nr_huge_pages; | |
2460 | int nid; | |
2461 | ||
2462 | h = kobj_to_hstate(kobj, &nid); | |
2463 | if (nid == NUMA_NO_NODE) | |
2464 | nr_huge_pages = h->nr_huge_pages; | |
2465 | else | |
2466 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2467 | ||
2468 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2469 | } |
adbe8726 | 2470 | |
238d3c13 DR |
2471 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2472 | struct hstate *h, int nid, | |
2473 | unsigned long count, size_t len) | |
a3437870 NA |
2474 | { |
2475 | int err; | |
bad44b5b | 2476 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2477 | |
944d9fec | 2478 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2479 | err = -EINVAL; |
2480 | goto out; | |
2481 | } | |
2482 | ||
9a305230 LS |
2483 | if (nid == NUMA_NO_NODE) { |
2484 | /* | |
2485 | * global hstate attribute | |
2486 | */ | |
2487 | if (!(obey_mempolicy && | |
2488 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2489 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2490 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2491 | } |
2492 | } else if (nodes_allowed) { | |
2493 | /* | |
2494 | * per node hstate attribute: adjust count to global, | |
2495 | * but restrict alloc/free to the specified node. | |
2496 | */ | |
2497 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2498 | init_nodemask_of_node(nodes_allowed, nid); | |
2499 | } else | |
8cebfcd0 | 2500 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2501 | |
06808b08 | 2502 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2503 | |
8cebfcd0 | 2504 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2505 | NODEMASK_FREE(nodes_allowed); |
2506 | ||
2507 | return len; | |
adbe8726 EM |
2508 | out: |
2509 | NODEMASK_FREE(nodes_allowed); | |
2510 | return err; | |
06808b08 LS |
2511 | } |
2512 | ||
238d3c13 DR |
2513 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2514 | struct kobject *kobj, const char *buf, | |
2515 | size_t len) | |
2516 | { | |
2517 | struct hstate *h; | |
2518 | unsigned long count; | |
2519 | int nid; | |
2520 | int err; | |
2521 | ||
2522 | err = kstrtoul(buf, 10, &count); | |
2523 | if (err) | |
2524 | return err; | |
2525 | ||
2526 | h = kobj_to_hstate(kobj, &nid); | |
2527 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2528 | } | |
2529 | ||
06808b08 LS |
2530 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2531 | struct kobj_attribute *attr, char *buf) | |
2532 | { | |
2533 | return nr_hugepages_show_common(kobj, attr, buf); | |
2534 | } | |
2535 | ||
2536 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2537 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2538 | { | |
238d3c13 | 2539 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2540 | } |
2541 | HSTATE_ATTR(nr_hugepages); | |
2542 | ||
06808b08 LS |
2543 | #ifdef CONFIG_NUMA |
2544 | ||
2545 | /* | |
2546 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2547 | * huge page alloc/free. | |
2548 | */ | |
2549 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2550 | struct kobj_attribute *attr, char *buf) | |
2551 | { | |
2552 | return nr_hugepages_show_common(kobj, attr, buf); | |
2553 | } | |
2554 | ||
2555 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2556 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2557 | { | |
238d3c13 | 2558 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2559 | } |
2560 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2561 | #endif | |
2562 | ||
2563 | ||
a3437870 NA |
2564 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2565 | struct kobj_attribute *attr, char *buf) | |
2566 | { | |
9a305230 | 2567 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2568 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2569 | } | |
adbe8726 | 2570 | |
a3437870 NA |
2571 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2572 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2573 | { | |
2574 | int err; | |
2575 | unsigned long input; | |
9a305230 | 2576 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2577 | |
bae7f4ae | 2578 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2579 | return -EINVAL; |
2580 | ||
3dbb95f7 | 2581 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2582 | if (err) |
73ae31e5 | 2583 | return err; |
a3437870 NA |
2584 | |
2585 | spin_lock(&hugetlb_lock); | |
2586 | h->nr_overcommit_huge_pages = input; | |
2587 | spin_unlock(&hugetlb_lock); | |
2588 | ||
2589 | return count; | |
2590 | } | |
2591 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2592 | ||
2593 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2594 | struct kobj_attribute *attr, char *buf) | |
2595 | { | |
9a305230 LS |
2596 | struct hstate *h; |
2597 | unsigned long free_huge_pages; | |
2598 | int nid; | |
2599 | ||
2600 | h = kobj_to_hstate(kobj, &nid); | |
2601 | if (nid == NUMA_NO_NODE) | |
2602 | free_huge_pages = h->free_huge_pages; | |
2603 | else | |
2604 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2605 | ||
2606 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2607 | } |
2608 | HSTATE_ATTR_RO(free_hugepages); | |
2609 | ||
2610 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2611 | struct kobj_attribute *attr, char *buf) | |
2612 | { | |
9a305230 | 2613 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2614 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2615 | } | |
2616 | HSTATE_ATTR_RO(resv_hugepages); | |
2617 | ||
2618 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2619 | struct kobj_attribute *attr, char *buf) | |
2620 | { | |
9a305230 LS |
2621 | struct hstate *h; |
2622 | unsigned long surplus_huge_pages; | |
2623 | int nid; | |
2624 | ||
2625 | h = kobj_to_hstate(kobj, &nid); | |
2626 | if (nid == NUMA_NO_NODE) | |
2627 | surplus_huge_pages = h->surplus_huge_pages; | |
2628 | else | |
2629 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2630 | ||
2631 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2632 | } |
2633 | HSTATE_ATTR_RO(surplus_hugepages); | |
2634 | ||
2635 | static struct attribute *hstate_attrs[] = { | |
2636 | &nr_hugepages_attr.attr, | |
2637 | &nr_overcommit_hugepages_attr.attr, | |
2638 | &free_hugepages_attr.attr, | |
2639 | &resv_hugepages_attr.attr, | |
2640 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2641 | #ifdef CONFIG_NUMA |
2642 | &nr_hugepages_mempolicy_attr.attr, | |
2643 | #endif | |
a3437870 NA |
2644 | NULL, |
2645 | }; | |
2646 | ||
2647 | static struct attribute_group hstate_attr_group = { | |
2648 | .attrs = hstate_attrs, | |
2649 | }; | |
2650 | ||
094e9539 JM |
2651 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2652 | struct kobject **hstate_kobjs, | |
2653 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2654 | { |
2655 | int retval; | |
972dc4de | 2656 | int hi = hstate_index(h); |
a3437870 | 2657 | |
9a305230 LS |
2658 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2659 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2660 | return -ENOMEM; |
2661 | ||
9a305230 | 2662 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2663 | if (retval) |
9a305230 | 2664 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2665 | |
2666 | return retval; | |
2667 | } | |
2668 | ||
2669 | static void __init hugetlb_sysfs_init(void) | |
2670 | { | |
2671 | struct hstate *h; | |
2672 | int err; | |
2673 | ||
2674 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2675 | if (!hugepages_kobj) | |
2676 | return; | |
2677 | ||
2678 | for_each_hstate(h) { | |
9a305230 LS |
2679 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2680 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2681 | if (err) |
ffb22af5 | 2682 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2683 | } |
2684 | } | |
2685 | ||
9a305230 LS |
2686 | #ifdef CONFIG_NUMA |
2687 | ||
2688 | /* | |
2689 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2690 | * with node devices in node_devices[] using a parallel array. The array |
2691 | * index of a node device or _hstate == node id. | |
2692 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2693 | * the base kernel, on the hugetlb module. |
2694 | */ | |
2695 | struct node_hstate { | |
2696 | struct kobject *hugepages_kobj; | |
2697 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2698 | }; | |
b4e289a6 | 2699 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2700 | |
2701 | /* | |
10fbcf4c | 2702 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2703 | */ |
2704 | static struct attribute *per_node_hstate_attrs[] = { | |
2705 | &nr_hugepages_attr.attr, | |
2706 | &free_hugepages_attr.attr, | |
2707 | &surplus_hugepages_attr.attr, | |
2708 | NULL, | |
2709 | }; | |
2710 | ||
2711 | static struct attribute_group per_node_hstate_attr_group = { | |
2712 | .attrs = per_node_hstate_attrs, | |
2713 | }; | |
2714 | ||
2715 | /* | |
10fbcf4c | 2716 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2717 | * Returns node id via non-NULL nidp. |
2718 | */ | |
2719 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2720 | { | |
2721 | int nid; | |
2722 | ||
2723 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2724 | struct node_hstate *nhs = &node_hstates[nid]; | |
2725 | int i; | |
2726 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2727 | if (nhs->hstate_kobjs[i] == kobj) { | |
2728 | if (nidp) | |
2729 | *nidp = nid; | |
2730 | return &hstates[i]; | |
2731 | } | |
2732 | } | |
2733 | ||
2734 | BUG(); | |
2735 | return NULL; | |
2736 | } | |
2737 | ||
2738 | /* | |
10fbcf4c | 2739 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2740 | * No-op if no hstate attributes attached. |
2741 | */ | |
3cd8b44f | 2742 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2743 | { |
2744 | struct hstate *h; | |
10fbcf4c | 2745 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2746 | |
2747 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2748 | return; /* no hstate attributes */ |
9a305230 | 2749 | |
972dc4de AK |
2750 | for_each_hstate(h) { |
2751 | int idx = hstate_index(h); | |
2752 | if (nhs->hstate_kobjs[idx]) { | |
2753 | kobject_put(nhs->hstate_kobjs[idx]); | |
2754 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2755 | } |
972dc4de | 2756 | } |
9a305230 LS |
2757 | |
2758 | kobject_put(nhs->hugepages_kobj); | |
2759 | nhs->hugepages_kobj = NULL; | |
2760 | } | |
2761 | ||
9a305230 LS |
2762 | |
2763 | /* | |
10fbcf4c | 2764 | * Register hstate attributes for a single node device. |
9a305230 LS |
2765 | * No-op if attributes already registered. |
2766 | */ | |
3cd8b44f | 2767 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2768 | { |
2769 | struct hstate *h; | |
10fbcf4c | 2770 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2771 | int err; |
2772 | ||
2773 | if (nhs->hugepages_kobj) | |
2774 | return; /* already allocated */ | |
2775 | ||
2776 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2777 | &node->dev.kobj); |
9a305230 LS |
2778 | if (!nhs->hugepages_kobj) |
2779 | return; | |
2780 | ||
2781 | for_each_hstate(h) { | |
2782 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2783 | nhs->hstate_kobjs, | |
2784 | &per_node_hstate_attr_group); | |
2785 | if (err) { | |
ffb22af5 AM |
2786 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2787 | h->name, node->dev.id); | |
9a305230 LS |
2788 | hugetlb_unregister_node(node); |
2789 | break; | |
2790 | } | |
2791 | } | |
2792 | } | |
2793 | ||
2794 | /* | |
9b5e5d0f | 2795 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2796 | * devices of nodes that have memory. All on-line nodes should have |
2797 | * registered their associated device by this time. | |
9a305230 | 2798 | */ |
7d9ca000 | 2799 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2800 | { |
2801 | int nid; | |
2802 | ||
8cebfcd0 | 2803 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2804 | struct node *node = node_devices[nid]; |
10fbcf4c | 2805 | if (node->dev.id == nid) |
9a305230 LS |
2806 | hugetlb_register_node(node); |
2807 | } | |
2808 | ||
2809 | /* | |
10fbcf4c | 2810 | * Let the node device driver know we're here so it can |
9a305230 LS |
2811 | * [un]register hstate attributes on node hotplug. |
2812 | */ | |
2813 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2814 | hugetlb_unregister_node); | |
2815 | } | |
2816 | #else /* !CONFIG_NUMA */ | |
2817 | ||
2818 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2819 | { | |
2820 | BUG(); | |
2821 | if (nidp) | |
2822 | *nidp = -1; | |
2823 | return NULL; | |
2824 | } | |
2825 | ||
9a305230 LS |
2826 | static void hugetlb_register_all_nodes(void) { } |
2827 | ||
2828 | #endif | |
2829 | ||
a3437870 NA |
2830 | static int __init hugetlb_init(void) |
2831 | { | |
8382d914 DB |
2832 | int i; |
2833 | ||
457c1b27 | 2834 | if (!hugepages_supported()) |
0ef89d25 | 2835 | return 0; |
a3437870 | 2836 | |
e11bfbfc | 2837 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
2838 | if (default_hstate_size != 0) { |
2839 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
2840 | default_hstate_size, HPAGE_SIZE); | |
2841 | } | |
2842 | ||
e11bfbfc NP |
2843 | default_hstate_size = HPAGE_SIZE; |
2844 | if (!size_to_hstate(default_hstate_size)) | |
2845 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2846 | } |
972dc4de | 2847 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2848 | if (default_hstate_max_huge_pages) { |
2849 | if (!default_hstate.max_huge_pages) | |
2850 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2851 | } | |
a3437870 NA |
2852 | |
2853 | hugetlb_init_hstates(); | |
aa888a74 | 2854 | gather_bootmem_prealloc(); |
a3437870 NA |
2855 | report_hugepages(); |
2856 | ||
2857 | hugetlb_sysfs_init(); | |
9a305230 | 2858 | hugetlb_register_all_nodes(); |
7179e7bf | 2859 | hugetlb_cgroup_file_init(); |
9a305230 | 2860 | |
8382d914 DB |
2861 | #ifdef CONFIG_SMP |
2862 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2863 | #else | |
2864 | num_fault_mutexes = 1; | |
2865 | #endif | |
c672c7f2 | 2866 | hugetlb_fault_mutex_table = |
8382d914 | 2867 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2868 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2869 | |
2870 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2871 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2872 | return 0; |
2873 | } | |
3e89e1c5 | 2874 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2875 | |
2876 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2877 | void __init hugetlb_bad_size(void) |
2878 | { | |
2879 | parsed_valid_hugepagesz = false; | |
2880 | } | |
2881 | ||
d00181b9 | 2882 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2883 | { |
2884 | struct hstate *h; | |
8faa8b07 AK |
2885 | unsigned long i; |
2886 | ||
a3437870 | 2887 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2888 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2889 | return; |
2890 | } | |
47d38344 | 2891 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2892 | BUG_ON(order == 0); |
47d38344 | 2893 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2894 | h->order = order; |
2895 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2896 | h->nr_huge_pages = 0; |
2897 | h->free_huge_pages = 0; | |
2898 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2899 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2900 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2901 | h->next_nid_to_alloc = first_memory_node; |
2902 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2903 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2904 | huge_page_size(h)/1024); | |
8faa8b07 | 2905 | |
a3437870 NA |
2906 | parsed_hstate = h; |
2907 | } | |
2908 | ||
e11bfbfc | 2909 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2910 | { |
2911 | unsigned long *mhp; | |
8faa8b07 | 2912 | static unsigned long *last_mhp; |
a3437870 | 2913 | |
9fee021d VT |
2914 | if (!parsed_valid_hugepagesz) { |
2915 | pr_warn("hugepages = %s preceded by " | |
2916 | "an unsupported hugepagesz, ignoring\n", s); | |
2917 | parsed_valid_hugepagesz = true; | |
2918 | return 1; | |
2919 | } | |
a3437870 | 2920 | /* |
47d38344 | 2921 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2922 | * so this hugepages= parameter goes to the "default hstate". |
2923 | */ | |
9fee021d | 2924 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2925 | mhp = &default_hstate_max_huge_pages; |
2926 | else | |
2927 | mhp = &parsed_hstate->max_huge_pages; | |
2928 | ||
8faa8b07 | 2929 | if (mhp == last_mhp) { |
598d8091 | 2930 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2931 | return 1; |
2932 | } | |
2933 | ||
a3437870 NA |
2934 | if (sscanf(s, "%lu", mhp) <= 0) |
2935 | *mhp = 0; | |
2936 | ||
8faa8b07 AK |
2937 | /* |
2938 | * Global state is always initialized later in hugetlb_init. | |
2939 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2940 | * use the bootmem allocator. | |
2941 | */ | |
47d38344 | 2942 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2943 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2944 | ||
2945 | last_mhp = mhp; | |
2946 | ||
a3437870 NA |
2947 | return 1; |
2948 | } | |
e11bfbfc NP |
2949 | __setup("hugepages=", hugetlb_nrpages_setup); |
2950 | ||
2951 | static int __init hugetlb_default_setup(char *s) | |
2952 | { | |
2953 | default_hstate_size = memparse(s, &s); | |
2954 | return 1; | |
2955 | } | |
2956 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2957 | |
8a213460 NA |
2958 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2959 | { | |
2960 | int node; | |
2961 | unsigned int nr = 0; | |
2962 | ||
2963 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2964 | nr += array[node]; | |
2965 | ||
2966 | return nr; | |
2967 | } | |
2968 | ||
2969 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2970 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2971 | struct ctl_table *table, int write, | |
2972 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2973 | { |
e5ff2159 | 2974 | struct hstate *h = &default_hstate; |
238d3c13 | 2975 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2976 | int ret; |
e5ff2159 | 2977 | |
457c1b27 | 2978 | if (!hugepages_supported()) |
86613628 | 2979 | return -EOPNOTSUPP; |
457c1b27 | 2980 | |
e5ff2159 AK |
2981 | table->data = &tmp; |
2982 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2983 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2984 | if (ret) | |
2985 | goto out; | |
e5ff2159 | 2986 | |
238d3c13 DR |
2987 | if (write) |
2988 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2989 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2990 | out: |
2991 | return ret; | |
1da177e4 | 2992 | } |
396faf03 | 2993 | |
06808b08 LS |
2994 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2995 | void __user *buffer, size_t *length, loff_t *ppos) | |
2996 | { | |
2997 | ||
2998 | return hugetlb_sysctl_handler_common(false, table, write, | |
2999 | buffer, length, ppos); | |
3000 | } | |
3001 | ||
3002 | #ifdef CONFIG_NUMA | |
3003 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
3004 | void __user *buffer, size_t *length, loff_t *ppos) | |
3005 | { | |
3006 | return hugetlb_sysctl_handler_common(true, table, write, | |
3007 | buffer, length, ppos); | |
3008 | } | |
3009 | #endif /* CONFIG_NUMA */ | |
3010 | ||
a3d0c6aa | 3011 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 3012 | void __user *buffer, |
a3d0c6aa NA |
3013 | size_t *length, loff_t *ppos) |
3014 | { | |
a5516438 | 3015 | struct hstate *h = &default_hstate; |
e5ff2159 | 3016 | unsigned long tmp; |
08d4a246 | 3017 | int ret; |
e5ff2159 | 3018 | |
457c1b27 | 3019 | if (!hugepages_supported()) |
86613628 | 3020 | return -EOPNOTSUPP; |
457c1b27 | 3021 | |
c033a93c | 3022 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3023 | |
bae7f4ae | 3024 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3025 | return -EINVAL; |
3026 | ||
e5ff2159 AK |
3027 | table->data = &tmp; |
3028 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3029 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3030 | if (ret) | |
3031 | goto out; | |
e5ff2159 AK |
3032 | |
3033 | if (write) { | |
3034 | spin_lock(&hugetlb_lock); | |
3035 | h->nr_overcommit_huge_pages = tmp; | |
3036 | spin_unlock(&hugetlb_lock); | |
3037 | } | |
08d4a246 MH |
3038 | out: |
3039 | return ret; | |
a3d0c6aa NA |
3040 | } |
3041 | ||
1da177e4 LT |
3042 | #endif /* CONFIG_SYSCTL */ |
3043 | ||
e1759c21 | 3044 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3045 | { |
a5516438 | 3046 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3047 | if (!hugepages_supported()) |
3048 | return; | |
e1759c21 | 3049 | seq_printf(m, |
4f98a2fe RR |
3050 | "HugePages_Total: %5lu\n" |
3051 | "HugePages_Free: %5lu\n" | |
3052 | "HugePages_Rsvd: %5lu\n" | |
3053 | "HugePages_Surp: %5lu\n" | |
3054 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
3055 | h->nr_huge_pages, |
3056 | h->free_huge_pages, | |
3057 | h->resv_huge_pages, | |
3058 | h->surplus_huge_pages, | |
3059 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
3060 | } |
3061 | ||
3062 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3063 | { | |
a5516438 | 3064 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3065 | if (!hugepages_supported()) |
3066 | return 0; | |
1da177e4 LT |
3067 | return sprintf(buf, |
3068 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3069 | "Node %d HugePages_Free: %5u\n" |
3070 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3071 | nid, h->nr_huge_pages_node[nid], |
3072 | nid, h->free_huge_pages_node[nid], | |
3073 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3074 | } |
3075 | ||
949f7ec5 DR |
3076 | void hugetlb_show_meminfo(void) |
3077 | { | |
3078 | struct hstate *h; | |
3079 | int nid; | |
3080 | ||
457c1b27 NA |
3081 | if (!hugepages_supported()) |
3082 | return; | |
3083 | ||
949f7ec5 DR |
3084 | for_each_node_state(nid, N_MEMORY) |
3085 | for_each_hstate(h) | |
3086 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3087 | nid, | |
3088 | h->nr_huge_pages_node[nid], | |
3089 | h->free_huge_pages_node[nid], | |
3090 | h->surplus_huge_pages_node[nid], | |
3091 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3092 | } | |
3093 | ||
5d317b2b NH |
3094 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3095 | { | |
3096 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3097 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3098 | } | |
3099 | ||
1da177e4 LT |
3100 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3101 | unsigned long hugetlb_total_pages(void) | |
3102 | { | |
d0028588 WL |
3103 | struct hstate *h; |
3104 | unsigned long nr_total_pages = 0; | |
3105 | ||
3106 | for_each_hstate(h) | |
3107 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3108 | return nr_total_pages; | |
1da177e4 | 3109 | } |
1da177e4 | 3110 | |
a5516438 | 3111 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3112 | { |
3113 | int ret = -ENOMEM; | |
3114 | ||
3115 | spin_lock(&hugetlb_lock); | |
3116 | /* | |
3117 | * When cpuset is configured, it breaks the strict hugetlb page | |
3118 | * reservation as the accounting is done on a global variable. Such | |
3119 | * reservation is completely rubbish in the presence of cpuset because | |
3120 | * the reservation is not checked against page availability for the | |
3121 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3122 | * with lack of free htlb page in cpuset that the task is in. | |
3123 | * Attempt to enforce strict accounting with cpuset is almost | |
3124 | * impossible (or too ugly) because cpuset is too fluid that | |
3125 | * task or memory node can be dynamically moved between cpusets. | |
3126 | * | |
3127 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3128 | * undesirable. However, in order to preserve some of the semantics, | |
3129 | * we fall back to check against current free page availability as | |
3130 | * a best attempt and hopefully to minimize the impact of changing | |
3131 | * semantics that cpuset has. | |
3132 | */ | |
3133 | if (delta > 0) { | |
a5516438 | 3134 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3135 | goto out; |
3136 | ||
a5516438 AK |
3137 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3138 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3139 | goto out; |
3140 | } | |
3141 | } | |
3142 | ||
3143 | ret = 0; | |
3144 | if (delta < 0) | |
a5516438 | 3145 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3146 | |
3147 | out: | |
3148 | spin_unlock(&hugetlb_lock); | |
3149 | return ret; | |
3150 | } | |
3151 | ||
84afd99b AW |
3152 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3153 | { | |
f522c3ac | 3154 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3155 | |
3156 | /* | |
3157 | * This new VMA should share its siblings reservation map if present. | |
3158 | * The VMA will only ever have a valid reservation map pointer where | |
3159 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3160 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3161 | * after this open call completes. It is therefore safe to take a |
3162 | * new reference here without additional locking. | |
3163 | */ | |
4e35f483 | 3164 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3165 | kref_get(&resv->refs); |
84afd99b AW |
3166 | } |
3167 | ||
a1e78772 MG |
3168 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3169 | { | |
a5516438 | 3170 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3171 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3172 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3173 | unsigned long reserve, start, end; |
1c5ecae3 | 3174 | long gbl_reserve; |
84afd99b | 3175 | |
4e35f483 JK |
3176 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3177 | return; | |
84afd99b | 3178 | |
4e35f483 JK |
3179 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3180 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3181 | |
4e35f483 | 3182 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3183 | |
4e35f483 JK |
3184 | kref_put(&resv->refs, resv_map_release); |
3185 | ||
3186 | if (reserve) { | |
1c5ecae3 MK |
3187 | /* |
3188 | * Decrement reserve counts. The global reserve count may be | |
3189 | * adjusted if the subpool has a minimum size. | |
3190 | */ | |
3191 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3192 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3193 | } |
a1e78772 MG |
3194 | } |
3195 | ||
1da177e4 LT |
3196 | /* |
3197 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3198 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3199 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3200 | * this far. | |
3201 | */ | |
11bac800 | 3202 | static int hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3203 | { |
3204 | BUG(); | |
d0217ac0 | 3205 | return 0; |
1da177e4 LT |
3206 | } |
3207 | ||
f0f37e2f | 3208 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3209 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3210 | .open = hugetlb_vm_op_open, |
a1e78772 | 3211 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
3212 | }; |
3213 | ||
1e8f889b DG |
3214 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3215 | int writable) | |
63551ae0 DG |
3216 | { |
3217 | pte_t entry; | |
3218 | ||
1e8f889b | 3219 | if (writable) { |
106c992a GS |
3220 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3221 | vma->vm_page_prot))); | |
63551ae0 | 3222 | } else { |
106c992a GS |
3223 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3224 | vma->vm_page_prot)); | |
63551ae0 DG |
3225 | } |
3226 | entry = pte_mkyoung(entry); | |
3227 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3228 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3229 | |
3230 | return entry; | |
3231 | } | |
3232 | ||
1e8f889b DG |
3233 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3234 | unsigned long address, pte_t *ptep) | |
3235 | { | |
3236 | pte_t entry; | |
3237 | ||
106c992a | 3238 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3239 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3240 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3241 | } |
3242 | ||
d5ed7444 | 3243 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3244 | { |
3245 | swp_entry_t swp; | |
3246 | ||
3247 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3248 | return false; |
4a705fef NH |
3249 | swp = pte_to_swp_entry(pte); |
3250 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3251 | return true; |
4a705fef | 3252 | else |
d5ed7444 | 3253 | return false; |
4a705fef NH |
3254 | } |
3255 | ||
3256 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3257 | { | |
3258 | swp_entry_t swp; | |
3259 | ||
3260 | if (huge_pte_none(pte) || pte_present(pte)) | |
3261 | return 0; | |
3262 | swp = pte_to_swp_entry(pte); | |
3263 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3264 | return 1; | |
3265 | else | |
3266 | return 0; | |
3267 | } | |
1e8f889b | 3268 | |
63551ae0 DG |
3269 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3270 | struct vm_area_struct *vma) | |
3271 | { | |
3272 | pte_t *src_pte, *dst_pte, entry; | |
3273 | struct page *ptepage; | |
1c59827d | 3274 | unsigned long addr; |
1e8f889b | 3275 | int cow; |
a5516438 AK |
3276 | struct hstate *h = hstate_vma(vma); |
3277 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
3278 | unsigned long mmun_start; /* For mmu_notifiers */ |
3279 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3280 | int ret = 0; | |
1e8f889b DG |
3281 | |
3282 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3283 | |
e8569dd2 AS |
3284 | mmun_start = vma->vm_start; |
3285 | mmun_end = vma->vm_end; | |
3286 | if (cow) | |
3287 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
3288 | ||
a5516438 | 3289 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3290 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3291 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3292 | if (!src_pte) |
3293 | continue; | |
a5516438 | 3294 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3295 | if (!dst_pte) { |
3296 | ret = -ENOMEM; | |
3297 | break; | |
3298 | } | |
c5c99429 LW |
3299 | |
3300 | /* If the pagetables are shared don't copy or take references */ | |
3301 | if (dst_pte == src_pte) | |
3302 | continue; | |
3303 | ||
cb900f41 KS |
3304 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3305 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3306 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3307 | entry = huge_ptep_get(src_pte); |
3308 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3309 | ; | |
3310 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3311 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3312 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3313 | ||
3314 | if (is_write_migration_entry(swp_entry) && cow) { | |
3315 | /* | |
3316 | * COW mappings require pages in both | |
3317 | * parent and child to be set to read. | |
3318 | */ | |
3319 | make_migration_entry_read(&swp_entry); | |
3320 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3321 | set_huge_swap_pte_at(src, addr, src_pte, |
3322 | entry, sz); | |
4a705fef | 3323 | } |
e5251fd4 | 3324 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3325 | } else { |
34ee645e | 3326 | if (cow) { |
7f2e9525 | 3327 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
3328 | mmu_notifier_invalidate_range(src, mmun_start, |
3329 | mmun_end); | |
3330 | } | |
0253d634 | 3331 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3332 | ptepage = pte_page(entry); |
3333 | get_page(ptepage); | |
53f9263b | 3334 | page_dup_rmap(ptepage, true); |
1c59827d | 3335 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3336 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3337 | } |
cb900f41 KS |
3338 | spin_unlock(src_ptl); |
3339 | spin_unlock(dst_ptl); | |
63551ae0 | 3340 | } |
63551ae0 | 3341 | |
e8569dd2 AS |
3342 | if (cow) |
3343 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3344 | ||
3345 | return ret; | |
63551ae0 DG |
3346 | } |
3347 | ||
24669e58 AK |
3348 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3349 | unsigned long start, unsigned long end, | |
3350 | struct page *ref_page) | |
63551ae0 DG |
3351 | { |
3352 | struct mm_struct *mm = vma->vm_mm; | |
3353 | unsigned long address; | |
c7546f8f | 3354 | pte_t *ptep; |
63551ae0 | 3355 | pte_t pte; |
cb900f41 | 3356 | spinlock_t *ptl; |
63551ae0 | 3357 | struct page *page; |
a5516438 AK |
3358 | struct hstate *h = hstate_vma(vma); |
3359 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3360 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3361 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3362 | |
63551ae0 | 3363 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3364 | BUG_ON(start & ~huge_page_mask(h)); |
3365 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3366 | |
07e32661 AK |
3367 | /* |
3368 | * This is a hugetlb vma, all the pte entries should point | |
3369 | * to huge page. | |
3370 | */ | |
3371 | tlb_remove_check_page_size_change(tlb, sz); | |
24669e58 | 3372 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3373 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3374 | address = start; |
569f48b8 | 3375 | for (; address < end; address += sz) { |
7868a208 | 3376 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3377 | if (!ptep) |
c7546f8f DG |
3378 | continue; |
3379 | ||
cb900f41 | 3380 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3381 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3382 | spin_unlock(ptl); | |
3383 | continue; | |
3384 | } | |
39dde65c | 3385 | |
6629326b | 3386 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3387 | if (huge_pte_none(pte)) { |
3388 | spin_unlock(ptl); | |
3389 | continue; | |
3390 | } | |
6629326b HD |
3391 | |
3392 | /* | |
9fbc1f63 NH |
3393 | * Migrating hugepage or HWPoisoned hugepage is already |
3394 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3395 | */ |
9fbc1f63 | 3396 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3397 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3398 | spin_unlock(ptl); |
3399 | continue; | |
8c4894c6 | 3400 | } |
6629326b HD |
3401 | |
3402 | page = pte_page(pte); | |
04f2cbe3 MG |
3403 | /* |
3404 | * If a reference page is supplied, it is because a specific | |
3405 | * page is being unmapped, not a range. Ensure the page we | |
3406 | * are about to unmap is the actual page of interest. | |
3407 | */ | |
3408 | if (ref_page) { | |
31d49da5 AK |
3409 | if (page != ref_page) { |
3410 | spin_unlock(ptl); | |
3411 | continue; | |
3412 | } | |
04f2cbe3 MG |
3413 | /* |
3414 | * Mark the VMA as having unmapped its page so that | |
3415 | * future faults in this VMA will fail rather than | |
3416 | * looking like data was lost | |
3417 | */ | |
3418 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3419 | } | |
3420 | ||
c7546f8f | 3421 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3422 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3423 | if (huge_pte_dirty(pte)) |
6649a386 | 3424 | set_page_dirty(page); |
9e81130b | 3425 | |
5d317b2b | 3426 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3427 | page_remove_rmap(page, true); |
31d49da5 | 3428 | |
cb900f41 | 3429 | spin_unlock(ptl); |
e77b0852 | 3430 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3431 | /* |
3432 | * Bail out after unmapping reference page if supplied | |
3433 | */ | |
3434 | if (ref_page) | |
3435 | break; | |
fe1668ae | 3436 | } |
2ec74c3e | 3437 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3438 | tlb_end_vma(tlb, vma); |
1da177e4 | 3439 | } |
63551ae0 | 3440 | |
d833352a MG |
3441 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3442 | struct vm_area_struct *vma, unsigned long start, | |
3443 | unsigned long end, struct page *ref_page) | |
3444 | { | |
3445 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3446 | ||
3447 | /* | |
3448 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3449 | * test will fail on a vma being torn down, and not grab a page table | |
3450 | * on its way out. We're lucky that the flag has such an appropriate | |
3451 | * name, and can in fact be safely cleared here. We could clear it | |
3452 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3453 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3454 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3455 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3456 | */ |
3457 | vma->vm_flags &= ~VM_MAYSHARE; | |
3458 | } | |
3459 | ||
502717f4 | 3460 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3461 | unsigned long end, struct page *ref_page) |
502717f4 | 3462 | { |
24669e58 AK |
3463 | struct mm_struct *mm; |
3464 | struct mmu_gather tlb; | |
3465 | ||
3466 | mm = vma->vm_mm; | |
3467 | ||
2b047252 | 3468 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3469 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3470 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
3471 | } |
3472 | ||
04f2cbe3 MG |
3473 | /* |
3474 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3475 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3476 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3477 | * same region. | |
3478 | */ | |
2f4612af DB |
3479 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3480 | struct page *page, unsigned long address) | |
04f2cbe3 | 3481 | { |
7526674d | 3482 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3483 | struct vm_area_struct *iter_vma; |
3484 | struct address_space *mapping; | |
04f2cbe3 MG |
3485 | pgoff_t pgoff; |
3486 | ||
3487 | /* | |
3488 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3489 | * from page cache lookup which is in HPAGE_SIZE units. | |
3490 | */ | |
7526674d | 3491 | address = address & huge_page_mask(h); |
36e4f20a MH |
3492 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3493 | vma->vm_pgoff; | |
93c76a3d | 3494 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3495 | |
4eb2b1dc MG |
3496 | /* |
3497 | * Take the mapping lock for the duration of the table walk. As | |
3498 | * this mapping should be shared between all the VMAs, | |
3499 | * __unmap_hugepage_range() is called as the lock is already held | |
3500 | */ | |
83cde9e8 | 3501 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3502 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3503 | /* Do not unmap the current VMA */ |
3504 | if (iter_vma == vma) | |
3505 | continue; | |
3506 | ||
2f84a899 MG |
3507 | /* |
3508 | * Shared VMAs have their own reserves and do not affect | |
3509 | * MAP_PRIVATE accounting but it is possible that a shared | |
3510 | * VMA is using the same page so check and skip such VMAs. | |
3511 | */ | |
3512 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3513 | continue; | |
3514 | ||
04f2cbe3 MG |
3515 | /* |
3516 | * Unmap the page from other VMAs without their own reserves. | |
3517 | * They get marked to be SIGKILLed if they fault in these | |
3518 | * areas. This is because a future no-page fault on this VMA | |
3519 | * could insert a zeroed page instead of the data existing | |
3520 | * from the time of fork. This would look like data corruption | |
3521 | */ | |
3522 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3523 | unmap_hugepage_range(iter_vma, address, |
3524 | address + huge_page_size(h), page); | |
04f2cbe3 | 3525 | } |
83cde9e8 | 3526 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3527 | } |
3528 | ||
0fe6e20b NH |
3529 | /* |
3530 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3531 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3532 | * cannot race with other handlers or page migration. | |
3533 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3534 | */ |
1e8f889b | 3535 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
3999f52e AK |
3536 | unsigned long address, pte_t *ptep, |
3537 | struct page *pagecache_page, spinlock_t *ptl) | |
1e8f889b | 3538 | { |
3999f52e | 3539 | pte_t pte; |
a5516438 | 3540 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3541 | struct page *old_page, *new_page; |
ad4404a2 | 3542 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3543 | unsigned long mmun_start; /* For mmu_notifiers */ |
3544 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b | 3545 | |
3999f52e | 3546 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3547 | old_page = pte_page(pte); |
3548 | ||
04f2cbe3 | 3549 | retry_avoidcopy: |
1e8f889b DG |
3550 | /* If no-one else is actually using this page, avoid the copy |
3551 | * and just make the page writable */ | |
37a2140d | 3552 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3553 | page_move_anon_rmap(old_page, vma); |
1e8f889b | 3554 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3555 | return 0; |
1e8f889b DG |
3556 | } |
3557 | ||
04f2cbe3 MG |
3558 | /* |
3559 | * If the process that created a MAP_PRIVATE mapping is about to | |
3560 | * perform a COW due to a shared page count, attempt to satisfy | |
3561 | * the allocation without using the existing reserves. The pagecache | |
3562 | * page is used to determine if the reserve at this address was | |
3563 | * consumed or not. If reserves were used, a partial faulted mapping | |
3564 | * at the time of fork() could consume its reserves on COW instead | |
3565 | * of the full address range. | |
3566 | */ | |
5944d011 | 3567 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3568 | old_page != pagecache_page) |
3569 | outside_reserve = 1; | |
3570 | ||
09cbfeaf | 3571 | get_page(old_page); |
b76c8cfb | 3572 | |
ad4404a2 DB |
3573 | /* |
3574 | * Drop page table lock as buddy allocator may be called. It will | |
3575 | * be acquired again before returning to the caller, as expected. | |
3576 | */ | |
cb900f41 | 3577 | spin_unlock(ptl); |
04f2cbe3 | 3578 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3579 | |
2fc39cec | 3580 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3581 | /* |
3582 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3583 | * it is due to references held by a child and an insufficient | |
3584 | * huge page pool. To guarantee the original mappers | |
3585 | * reliability, unmap the page from child processes. The child | |
3586 | * may get SIGKILLed if it later faults. | |
3587 | */ | |
3588 | if (outside_reserve) { | |
09cbfeaf | 3589 | put_page(old_page); |
04f2cbe3 | 3590 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3591 | unmap_ref_private(mm, vma, old_page, address); |
3592 | BUG_ON(huge_pte_none(pte)); | |
3593 | spin_lock(ptl); | |
7868a208 PA |
3594 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3595 | huge_page_size(h)); | |
2f4612af DB |
3596 | if (likely(ptep && |
3597 | pte_same(huge_ptep_get(ptep), pte))) | |
3598 | goto retry_avoidcopy; | |
3599 | /* | |
3600 | * race occurs while re-acquiring page table | |
3601 | * lock, and our job is done. | |
3602 | */ | |
3603 | return 0; | |
04f2cbe3 MG |
3604 | } |
3605 | ||
ad4404a2 DB |
3606 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3607 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3608 | goto out_release_old; | |
1e8f889b DG |
3609 | } |
3610 | ||
0fe6e20b NH |
3611 | /* |
3612 | * When the original hugepage is shared one, it does not have | |
3613 | * anon_vma prepared. | |
3614 | */ | |
44e2aa93 | 3615 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3616 | ret = VM_FAULT_OOM; |
3617 | goto out_release_all; | |
44e2aa93 | 3618 | } |
0fe6e20b | 3619 | |
47ad8475 AA |
3620 | copy_user_huge_page(new_page, old_page, address, vma, |
3621 | pages_per_huge_page(h)); | |
0ed361de | 3622 | __SetPageUptodate(new_page); |
bcc54222 | 3623 | set_page_huge_active(new_page); |
1e8f889b | 3624 | |
2ec74c3e SG |
3625 | mmun_start = address & huge_page_mask(h); |
3626 | mmun_end = mmun_start + huge_page_size(h); | |
3627 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3628 | |
b76c8cfb | 3629 | /* |
cb900f41 | 3630 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3631 | * before the page tables are altered |
3632 | */ | |
cb900f41 | 3633 | spin_lock(ptl); |
7868a208 PA |
3634 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3635 | huge_page_size(h)); | |
a9af0c5d | 3636 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3637 | ClearPagePrivate(new_page); |
3638 | ||
1e8f889b | 3639 | /* Break COW */ |
8fe627ec | 3640 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3641 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3642 | set_huge_pte_at(mm, address, ptep, |
3643 | make_huge_pte(vma, new_page, 1)); | |
d281ee61 | 3644 | page_remove_rmap(old_page, true); |
cd67f0d2 | 3645 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3646 | /* Make the old page be freed below */ |
3647 | new_page = old_page; | |
3648 | } | |
cb900f41 | 3649 | spin_unlock(ptl); |
2ec74c3e | 3650 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3651 | out_release_all: |
96b96a96 | 3652 | restore_reserve_on_error(h, vma, address, new_page); |
09cbfeaf | 3653 | put_page(new_page); |
ad4404a2 | 3654 | out_release_old: |
09cbfeaf | 3655 | put_page(old_page); |
8312034f | 3656 | |
ad4404a2 DB |
3657 | spin_lock(ptl); /* Caller expects lock to be held */ |
3658 | return ret; | |
1e8f889b DG |
3659 | } |
3660 | ||
04f2cbe3 | 3661 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3662 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3663 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3664 | { |
3665 | struct address_space *mapping; | |
e7c4b0bf | 3666 | pgoff_t idx; |
04f2cbe3 MG |
3667 | |
3668 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3669 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3670 | |
3671 | return find_lock_page(mapping, idx); | |
3672 | } | |
3673 | ||
3ae77f43 HD |
3674 | /* |
3675 | * Return whether there is a pagecache page to back given address within VMA. | |
3676 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3677 | */ | |
3678 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3679 | struct vm_area_struct *vma, unsigned long address) |
3680 | { | |
3681 | struct address_space *mapping; | |
3682 | pgoff_t idx; | |
3683 | struct page *page; | |
3684 | ||
3685 | mapping = vma->vm_file->f_mapping; | |
3686 | idx = vma_hugecache_offset(h, vma, address); | |
3687 | ||
3688 | page = find_get_page(mapping, idx); | |
3689 | if (page) | |
3690 | put_page(page); | |
3691 | return page != NULL; | |
3692 | } | |
3693 | ||
ab76ad54 MK |
3694 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3695 | pgoff_t idx) | |
3696 | { | |
3697 | struct inode *inode = mapping->host; | |
3698 | struct hstate *h = hstate_inode(inode); | |
3699 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3700 | ||
3701 | if (err) | |
3702 | return err; | |
3703 | ClearPagePrivate(page); | |
3704 | ||
3705 | spin_lock(&inode->i_lock); | |
3706 | inode->i_blocks += blocks_per_huge_page(h); | |
3707 | spin_unlock(&inode->i_lock); | |
3708 | return 0; | |
3709 | } | |
3710 | ||
a1ed3dda | 3711 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3712 | struct address_space *mapping, pgoff_t idx, |
3713 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3714 | { |
a5516438 | 3715 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3716 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3717 | int anon_rmap = 0; |
4c887265 | 3718 | unsigned long size; |
4c887265 | 3719 | struct page *page; |
1e8f889b | 3720 | pte_t new_pte; |
cb900f41 | 3721 | spinlock_t *ptl; |
4c887265 | 3722 | |
04f2cbe3 MG |
3723 | /* |
3724 | * Currently, we are forced to kill the process in the event the | |
3725 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3726 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3727 | */ |
3728 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3729 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3730 | current->pid); |
04f2cbe3 MG |
3731 | return ret; |
3732 | } | |
3733 | ||
4c887265 AL |
3734 | /* |
3735 | * Use page lock to guard against racing truncation | |
3736 | * before we get page_table_lock. | |
3737 | */ | |
6bda666a CL |
3738 | retry: |
3739 | page = find_lock_page(mapping, idx); | |
3740 | if (!page) { | |
a5516438 | 3741 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3742 | if (idx >= size) |
3743 | goto out; | |
1a1aad8a MK |
3744 | |
3745 | /* | |
3746 | * Check for page in userfault range | |
3747 | */ | |
3748 | if (userfaultfd_missing(vma)) { | |
3749 | u32 hash; | |
3750 | struct vm_fault vmf = { | |
3751 | .vma = vma, | |
3752 | .address = address, | |
3753 | .flags = flags, | |
3754 | /* | |
3755 | * Hard to debug if it ends up being | |
3756 | * used by a callee that assumes | |
3757 | * something about the other | |
3758 | * uninitialized fields... same as in | |
3759 | * memory.c | |
3760 | */ | |
3761 | }; | |
3762 | ||
3763 | /* | |
3764 | * hugetlb_fault_mutex must be dropped before | |
3765 | * handling userfault. Reacquire after handling | |
3766 | * fault to make calling code simpler. | |
3767 | */ | |
3768 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, | |
3769 | idx, address); | |
3770 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
3771 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3772 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3773 | goto out; | |
3774 | } | |
3775 | ||
04f2cbe3 | 3776 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3777 | if (IS_ERR(page)) { |
76dcee75 AK |
3778 | ret = PTR_ERR(page); |
3779 | if (ret == -ENOMEM) | |
3780 | ret = VM_FAULT_OOM; | |
3781 | else | |
3782 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3783 | goto out; |
3784 | } | |
47ad8475 | 3785 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3786 | __SetPageUptodate(page); |
bcc54222 | 3787 | set_page_huge_active(page); |
ac9b9c66 | 3788 | |
f83a275d | 3789 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3790 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3791 | if (err) { |
3792 | put_page(page); | |
6bda666a CL |
3793 | if (err == -EEXIST) |
3794 | goto retry; | |
3795 | goto out; | |
3796 | } | |
23be7468 | 3797 | } else { |
6bda666a | 3798 | lock_page(page); |
0fe6e20b NH |
3799 | if (unlikely(anon_vma_prepare(vma))) { |
3800 | ret = VM_FAULT_OOM; | |
3801 | goto backout_unlocked; | |
3802 | } | |
409eb8c2 | 3803 | anon_rmap = 1; |
23be7468 | 3804 | } |
0fe6e20b | 3805 | } else { |
998b4382 NH |
3806 | /* |
3807 | * If memory error occurs between mmap() and fault, some process | |
3808 | * don't have hwpoisoned swap entry for errored virtual address. | |
3809 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3810 | */ | |
3811 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3812 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3813 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3814 | goto backout_unlocked; |
3815 | } | |
6bda666a | 3816 | } |
1e8f889b | 3817 | |
57303d80 AW |
3818 | /* |
3819 | * If we are going to COW a private mapping later, we examine the | |
3820 | * pending reservations for this page now. This will ensure that | |
3821 | * any allocations necessary to record that reservation occur outside | |
3822 | * the spinlock. | |
3823 | */ | |
5e911373 | 3824 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3825 | if (vma_needs_reservation(h, vma, address) < 0) { |
3826 | ret = VM_FAULT_OOM; | |
3827 | goto backout_unlocked; | |
3828 | } | |
5e911373 | 3829 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3830 | vma_end_reservation(h, vma, address); |
5e911373 | 3831 | } |
57303d80 | 3832 | |
8bea8052 | 3833 | ptl = huge_pte_lock(h, mm, ptep); |
a5516438 | 3834 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3835 | if (idx >= size) |
3836 | goto backout; | |
3837 | ||
83c54070 | 3838 | ret = 0; |
7f2e9525 | 3839 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3840 | goto backout; |
3841 | ||
07443a85 JK |
3842 | if (anon_rmap) { |
3843 | ClearPagePrivate(page); | |
409eb8c2 | 3844 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3845 | } else |
53f9263b | 3846 | page_dup_rmap(page, true); |
1e8f889b DG |
3847 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3848 | && (vma->vm_flags & VM_SHARED))); | |
3849 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3850 | ||
5d317b2b | 3851 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3852 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3853 | /* Optimization, do the COW without a second fault */ |
3999f52e | 3854 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
3855 | } |
3856 | ||
cb900f41 | 3857 | spin_unlock(ptl); |
4c887265 AL |
3858 | unlock_page(page); |
3859 | out: | |
ac9b9c66 | 3860 | return ret; |
4c887265 AL |
3861 | |
3862 | backout: | |
cb900f41 | 3863 | spin_unlock(ptl); |
2b26736c | 3864 | backout_unlocked: |
4c887265 | 3865 | unlock_page(page); |
96b96a96 | 3866 | restore_reserve_on_error(h, vma, address, page); |
4c887265 AL |
3867 | put_page(page); |
3868 | goto out; | |
ac9b9c66 HD |
3869 | } |
3870 | ||
8382d914 | 3871 | #ifdef CONFIG_SMP |
c672c7f2 | 3872 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3873 | struct vm_area_struct *vma, |
3874 | struct address_space *mapping, | |
3875 | pgoff_t idx, unsigned long address) | |
3876 | { | |
3877 | unsigned long key[2]; | |
3878 | u32 hash; | |
3879 | ||
3880 | if (vma->vm_flags & VM_SHARED) { | |
3881 | key[0] = (unsigned long) mapping; | |
3882 | key[1] = idx; | |
3883 | } else { | |
3884 | key[0] = (unsigned long) mm; | |
3885 | key[1] = address >> huge_page_shift(h); | |
3886 | } | |
3887 | ||
3888 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3889 | ||
3890 | return hash & (num_fault_mutexes - 1); | |
3891 | } | |
3892 | #else | |
3893 | /* | |
3894 | * For uniprocesor systems we always use a single mutex, so just | |
3895 | * return 0 and avoid the hashing overhead. | |
3896 | */ | |
c672c7f2 | 3897 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3898 | struct vm_area_struct *vma, |
3899 | struct address_space *mapping, | |
3900 | pgoff_t idx, unsigned long address) | |
3901 | { | |
3902 | return 0; | |
3903 | } | |
3904 | #endif | |
3905 | ||
86e5216f | 3906 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3907 | unsigned long address, unsigned int flags) |
86e5216f | 3908 | { |
8382d914 | 3909 | pte_t *ptep, entry; |
cb900f41 | 3910 | spinlock_t *ptl; |
1e8f889b | 3911 | int ret; |
8382d914 DB |
3912 | u32 hash; |
3913 | pgoff_t idx; | |
0fe6e20b | 3914 | struct page *page = NULL; |
57303d80 | 3915 | struct page *pagecache_page = NULL; |
a5516438 | 3916 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3917 | struct address_space *mapping; |
0f792cf9 | 3918 | int need_wait_lock = 0; |
86e5216f | 3919 | |
1e16a539 KH |
3920 | address &= huge_page_mask(h); |
3921 | ||
7868a208 | 3922 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
fd6a03ed NH |
3923 | if (ptep) { |
3924 | entry = huge_ptep_get(ptep); | |
290408d4 | 3925 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3926 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3927 | return 0; |
3928 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3929 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3930 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
0d777df5 NH |
3931 | } else { |
3932 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); | |
3933 | if (!ptep) | |
3934 | return VM_FAULT_OOM; | |
fd6a03ed NH |
3935 | } |
3936 | ||
8382d914 DB |
3937 | mapping = vma->vm_file->f_mapping; |
3938 | idx = vma_hugecache_offset(h, vma, address); | |
3939 | ||
3935baa9 DG |
3940 | /* |
3941 | * Serialize hugepage allocation and instantiation, so that we don't | |
3942 | * get spurious allocation failures if two CPUs race to instantiate | |
3943 | * the same page in the page cache. | |
3944 | */ | |
c672c7f2 MK |
3945 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3946 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3947 | |
7f2e9525 GS |
3948 | entry = huge_ptep_get(ptep); |
3949 | if (huge_pte_none(entry)) { | |
8382d914 | 3950 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3951 | goto out_mutex; |
3935baa9 | 3952 | } |
86e5216f | 3953 | |
83c54070 | 3954 | ret = 0; |
1e8f889b | 3955 | |
0f792cf9 NH |
3956 | /* |
3957 | * entry could be a migration/hwpoison entry at this point, so this | |
3958 | * check prevents the kernel from going below assuming that we have | |
3959 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3960 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3961 | * handle it. | |
3962 | */ | |
3963 | if (!pte_present(entry)) | |
3964 | goto out_mutex; | |
3965 | ||
57303d80 AW |
3966 | /* |
3967 | * If we are going to COW the mapping later, we examine the pending | |
3968 | * reservations for this page now. This will ensure that any | |
3969 | * allocations necessary to record that reservation occur outside the | |
3970 | * spinlock. For private mappings, we also lookup the pagecache | |
3971 | * page now as it is used to determine if a reservation has been | |
3972 | * consumed. | |
3973 | */ | |
106c992a | 3974 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3975 | if (vma_needs_reservation(h, vma, address) < 0) { |
3976 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3977 | goto out_mutex; |
2b26736c | 3978 | } |
5e911373 | 3979 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3980 | vma_end_reservation(h, vma, address); |
57303d80 | 3981 | |
f83a275d | 3982 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3983 | pagecache_page = hugetlbfs_pagecache_page(h, |
3984 | vma, address); | |
3985 | } | |
3986 | ||
0f792cf9 NH |
3987 | ptl = huge_pte_lock(h, mm, ptep); |
3988 | ||
3989 | /* Check for a racing update before calling hugetlb_cow */ | |
3990 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3991 | goto out_ptl; | |
3992 | ||
56c9cfb1 NH |
3993 | /* |
3994 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3995 | * pagecache_page, so here we need take the former one | |
3996 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3997 | */ |
3998 | page = pte_page(entry); | |
3999 | if (page != pagecache_page) | |
0f792cf9 NH |
4000 | if (!trylock_page(page)) { |
4001 | need_wait_lock = 1; | |
4002 | goto out_ptl; | |
4003 | } | |
b4d1d99f | 4004 | |
0f792cf9 | 4005 | get_page(page); |
b4d1d99f | 4006 | |
788c7df4 | 4007 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4008 | if (!huge_pte_write(entry)) { |
3999f52e AK |
4009 | ret = hugetlb_cow(mm, vma, address, ptep, |
4010 | pagecache_page, ptl); | |
0f792cf9 | 4011 | goto out_put_page; |
b4d1d99f | 4012 | } |
106c992a | 4013 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4014 | } |
4015 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
4016 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
4017 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 4018 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
4019 | out_put_page: |
4020 | if (page != pagecache_page) | |
4021 | unlock_page(page); | |
4022 | put_page(page); | |
cb900f41 KS |
4023 | out_ptl: |
4024 | spin_unlock(ptl); | |
57303d80 AW |
4025 | |
4026 | if (pagecache_page) { | |
4027 | unlock_page(pagecache_page); | |
4028 | put_page(pagecache_page); | |
4029 | } | |
b4d1d99f | 4030 | out_mutex: |
c672c7f2 | 4031 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
4032 | /* |
4033 | * Generally it's safe to hold refcount during waiting page lock. But | |
4034 | * here we just wait to defer the next page fault to avoid busy loop and | |
4035 | * the page is not used after unlocked before returning from the current | |
4036 | * page fault. So we are safe from accessing freed page, even if we wait | |
4037 | * here without taking refcount. | |
4038 | */ | |
4039 | if (need_wait_lock) | |
4040 | wait_on_page_locked(page); | |
1e8f889b | 4041 | return ret; |
86e5216f AL |
4042 | } |
4043 | ||
8fb5debc MK |
4044 | /* |
4045 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4046 | * modifications for huge pages. | |
4047 | */ | |
4048 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4049 | pte_t *dst_pte, | |
4050 | struct vm_area_struct *dst_vma, | |
4051 | unsigned long dst_addr, | |
4052 | unsigned long src_addr, | |
4053 | struct page **pagep) | |
4054 | { | |
1c9e8def | 4055 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4056 | struct hstate *h = hstate_vma(dst_vma); |
4057 | pte_t _dst_pte; | |
4058 | spinlock_t *ptl; | |
4059 | int ret; | |
4060 | struct page *page; | |
4061 | ||
4062 | if (!*pagep) { | |
4063 | ret = -ENOMEM; | |
4064 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4065 | if (IS_ERR(page)) | |
4066 | goto out; | |
4067 | ||
4068 | ret = copy_huge_page_from_user(page, | |
4069 | (const void __user *) src_addr, | |
810a56b9 | 4070 | pages_per_huge_page(h), false); |
8fb5debc MK |
4071 | |
4072 | /* fallback to copy_from_user outside mmap_sem */ | |
4073 | if (unlikely(ret)) { | |
4074 | ret = -EFAULT; | |
4075 | *pagep = page; | |
4076 | /* don't free the page */ | |
4077 | goto out; | |
4078 | } | |
4079 | } else { | |
4080 | page = *pagep; | |
4081 | *pagep = NULL; | |
4082 | } | |
4083 | ||
4084 | /* | |
4085 | * The memory barrier inside __SetPageUptodate makes sure that | |
4086 | * preceding stores to the page contents become visible before | |
4087 | * the set_pte_at() write. | |
4088 | */ | |
4089 | __SetPageUptodate(page); | |
4090 | set_page_huge_active(page); | |
4091 | ||
1c9e8def MK |
4092 | /* |
4093 | * If shared, add to page cache | |
4094 | */ | |
4095 | if (vm_shared) { | |
4096 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
4097 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4098 | ||
4099 | ret = huge_add_to_page_cache(page, mapping, idx); | |
4100 | if (ret) | |
4101 | goto out_release_nounlock; | |
4102 | } | |
4103 | ||
8fb5debc MK |
4104 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4105 | spin_lock(ptl); | |
4106 | ||
4107 | ret = -EEXIST; | |
4108 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4109 | goto out_release_unlock; | |
4110 | ||
1c9e8def MK |
4111 | if (vm_shared) { |
4112 | page_dup_rmap(page, true); | |
4113 | } else { | |
4114 | ClearPagePrivate(page); | |
4115 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4116 | } | |
8fb5debc MK |
4117 | |
4118 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4119 | if (dst_vma->vm_flags & VM_WRITE) | |
4120 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4121 | _dst_pte = pte_mkyoung(_dst_pte); | |
4122 | ||
4123 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4124 | ||
4125 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4126 | dst_vma->vm_flags & VM_WRITE); | |
4127 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4128 | ||
4129 | /* No need to invalidate - it was non-present before */ | |
4130 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4131 | ||
4132 | spin_unlock(ptl); | |
1c9e8def MK |
4133 | if (vm_shared) |
4134 | unlock_page(page); | |
8fb5debc MK |
4135 | ret = 0; |
4136 | out: | |
4137 | return ret; | |
4138 | out_release_unlock: | |
4139 | spin_unlock(ptl); | |
1c9e8def MK |
4140 | out_release_nounlock: |
4141 | if (vm_shared) | |
4142 | unlock_page(page); | |
8fb5debc MK |
4143 | put_page(page); |
4144 | goto out; | |
4145 | } | |
4146 | ||
28a35716 ML |
4147 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4148 | struct page **pages, struct vm_area_struct **vmas, | |
4149 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4150 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4151 | { |
d5d4b0aa KC |
4152 | unsigned long pfn_offset; |
4153 | unsigned long vaddr = *position; | |
28a35716 | 4154 | unsigned long remainder = *nr_pages; |
a5516438 | 4155 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 4156 | |
63551ae0 | 4157 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4158 | pte_t *pte; |
cb900f41 | 4159 | spinlock_t *ptl = NULL; |
2a15efc9 | 4160 | int absent; |
4c887265 | 4161 | struct page *page; |
63551ae0 | 4162 | |
02057967 DR |
4163 | /* |
4164 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4165 | * potentially allocating memory. | |
4166 | */ | |
4167 | if (unlikely(fatal_signal_pending(current))) { | |
4168 | remainder = 0; | |
4169 | break; | |
4170 | } | |
4171 | ||
4c887265 AL |
4172 | /* |
4173 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4174 | * each hugepage. We have to make sure we get the |
4c887265 | 4175 | * first, for the page indexing below to work. |
cb900f41 KS |
4176 | * |
4177 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4178 | */ |
7868a208 PA |
4179 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4180 | huge_page_size(h)); | |
cb900f41 KS |
4181 | if (pte) |
4182 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4183 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4184 | ||
4185 | /* | |
4186 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4187 | * an error where there's an empty slot with no huge pagecache |
4188 | * to back it. This way, we avoid allocating a hugepage, and | |
4189 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4190 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4191 | */ |
3ae77f43 HD |
4192 | if (absent && (flags & FOLL_DUMP) && |
4193 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4194 | if (pte) |
4195 | spin_unlock(ptl); | |
2a15efc9 HD |
4196 | remainder = 0; |
4197 | break; | |
4198 | } | |
63551ae0 | 4199 | |
9cc3a5bd NH |
4200 | /* |
4201 | * We need call hugetlb_fault for both hugepages under migration | |
4202 | * (in which case hugetlb_fault waits for the migration,) and | |
4203 | * hwpoisoned hugepages (in which case we need to prevent the | |
4204 | * caller from accessing to them.) In order to do this, we use | |
4205 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4206 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4207 | * both cases, and because we can't follow correct pages | |
4208 | * directly from any kind of swap entries. | |
4209 | */ | |
4210 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4211 | ((flags & FOLL_WRITE) && |
4212 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 4213 | int ret; |
87ffc118 | 4214 | unsigned int fault_flags = 0; |
63551ae0 | 4215 | |
cb900f41 KS |
4216 | if (pte) |
4217 | spin_unlock(ptl); | |
87ffc118 AA |
4218 | if (flags & FOLL_WRITE) |
4219 | fault_flags |= FAULT_FLAG_WRITE; | |
4220 | if (nonblocking) | |
4221 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4222 | if (flags & FOLL_NOWAIT) | |
4223 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4224 | FAULT_FLAG_RETRY_NOWAIT; | |
4225 | if (flags & FOLL_TRIED) { | |
4226 | VM_WARN_ON_ONCE(fault_flags & | |
4227 | FAULT_FLAG_ALLOW_RETRY); | |
4228 | fault_flags |= FAULT_FLAG_TRIED; | |
4229 | } | |
4230 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4231 | if (ret & VM_FAULT_ERROR) { | |
9a291a7c JM |
4232 | int err = vm_fault_to_errno(ret, flags); |
4233 | ||
4234 | if (err) | |
4235 | return err; | |
4236 | ||
87ffc118 AA |
4237 | remainder = 0; |
4238 | break; | |
4239 | } | |
4240 | if (ret & VM_FAULT_RETRY) { | |
4241 | if (nonblocking) | |
4242 | *nonblocking = 0; | |
4243 | *nr_pages = 0; | |
4244 | /* | |
4245 | * VM_FAULT_RETRY must not return an | |
4246 | * error, it will return zero | |
4247 | * instead. | |
4248 | * | |
4249 | * No need to update "position" as the | |
4250 | * caller will not check it after | |
4251 | * *nr_pages is set to 0. | |
4252 | */ | |
4253 | return i; | |
4254 | } | |
4255 | continue; | |
4c887265 AL |
4256 | } |
4257 | ||
a5516438 | 4258 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4259 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 4260 | same_page: |
d6692183 | 4261 | if (pages) { |
2a15efc9 | 4262 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 4263 | get_page(pages[i]); |
d6692183 | 4264 | } |
63551ae0 DG |
4265 | |
4266 | if (vmas) | |
4267 | vmas[i] = vma; | |
4268 | ||
4269 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4270 | ++pfn_offset; |
63551ae0 DG |
4271 | --remainder; |
4272 | ++i; | |
d5d4b0aa | 4273 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4274 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4275 | /* |
4276 | * We use pfn_offset to avoid touching the pageframes | |
4277 | * of this compound page. | |
4278 | */ | |
4279 | goto same_page; | |
4280 | } | |
cb900f41 | 4281 | spin_unlock(ptl); |
63551ae0 | 4282 | } |
28a35716 | 4283 | *nr_pages = remainder; |
87ffc118 AA |
4284 | /* |
4285 | * setting position is actually required only if remainder is | |
4286 | * not zero but it's faster not to add a "if (remainder)" | |
4287 | * branch. | |
4288 | */ | |
63551ae0 DG |
4289 | *position = vaddr; |
4290 | ||
2a15efc9 | 4291 | return i ? i : -EFAULT; |
63551ae0 | 4292 | } |
8f860591 | 4293 | |
5491ae7b AK |
4294 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4295 | /* | |
4296 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4297 | * implement this. | |
4298 | */ | |
4299 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4300 | #endif | |
4301 | ||
7da4d641 | 4302 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4303 | unsigned long address, unsigned long end, pgprot_t newprot) |
4304 | { | |
4305 | struct mm_struct *mm = vma->vm_mm; | |
4306 | unsigned long start = address; | |
4307 | pte_t *ptep; | |
4308 | pte_t pte; | |
a5516438 | 4309 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4310 | unsigned long pages = 0; |
8f860591 ZY |
4311 | |
4312 | BUG_ON(address >= end); | |
4313 | flush_cache_range(vma, address, end); | |
4314 | ||
a5338093 | 4315 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 4316 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4317 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4318 | spinlock_t *ptl; |
7868a208 | 4319 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4320 | if (!ptep) |
4321 | continue; | |
cb900f41 | 4322 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4323 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4324 | pages++; | |
cb900f41 | 4325 | spin_unlock(ptl); |
39dde65c | 4326 | continue; |
7da4d641 | 4327 | } |
a8bda28d NH |
4328 | pte = huge_ptep_get(ptep); |
4329 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4330 | spin_unlock(ptl); | |
4331 | continue; | |
4332 | } | |
4333 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4334 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4335 | ||
4336 | if (is_write_migration_entry(entry)) { | |
4337 | pte_t newpte; | |
4338 | ||
4339 | make_migration_entry_read(&entry); | |
4340 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4341 | set_huge_swap_pte_at(mm, address, ptep, |
4342 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4343 | pages++; |
4344 | } | |
4345 | spin_unlock(ptl); | |
4346 | continue; | |
4347 | } | |
4348 | if (!huge_pte_none(pte)) { | |
8f860591 | 4349 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 4350 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 4351 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 4352 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 4353 | pages++; |
8f860591 | 4354 | } |
cb900f41 | 4355 | spin_unlock(ptl); |
8f860591 | 4356 | } |
d833352a | 4357 | /* |
c8c06efa | 4358 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4359 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4360 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
4361 | * and that page table be reused and filled with junk. |
4362 | */ | |
5491ae7b | 4363 | flush_hugetlb_tlb_range(vma, start, end); |
34ee645e | 4364 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 4365 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 4366 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
4367 | |
4368 | return pages << h->order; | |
8f860591 ZY |
4369 | } |
4370 | ||
a1e78772 MG |
4371 | int hugetlb_reserve_pages(struct inode *inode, |
4372 | long from, long to, | |
5a6fe125 | 4373 | struct vm_area_struct *vma, |
ca16d140 | 4374 | vm_flags_t vm_flags) |
e4e574b7 | 4375 | { |
17c9d12e | 4376 | long ret, chg; |
a5516438 | 4377 | struct hstate *h = hstate_inode(inode); |
90481622 | 4378 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4379 | struct resv_map *resv_map; |
1c5ecae3 | 4380 | long gbl_reserve; |
e4e574b7 | 4381 | |
17c9d12e MG |
4382 | /* |
4383 | * Only apply hugepage reservation if asked. At fault time, an | |
4384 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4385 | * without using reserves |
17c9d12e | 4386 | */ |
ca16d140 | 4387 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4388 | return 0; |
4389 | ||
a1e78772 MG |
4390 | /* |
4391 | * Shared mappings base their reservation on the number of pages that | |
4392 | * are already allocated on behalf of the file. Private mappings need | |
4393 | * to reserve the full area even if read-only as mprotect() may be | |
4394 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4395 | */ | |
9119a41e | 4396 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 4397 | resv_map = inode_resv_map(inode); |
9119a41e | 4398 | |
1406ec9b | 4399 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4400 | |
4401 | } else { | |
4402 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4403 | if (!resv_map) |
4404 | return -ENOMEM; | |
4405 | ||
a1e78772 | 4406 | chg = to - from; |
84afd99b | 4407 | |
17c9d12e MG |
4408 | set_vma_resv_map(vma, resv_map); |
4409 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4410 | } | |
4411 | ||
c50ac050 DH |
4412 | if (chg < 0) { |
4413 | ret = chg; | |
4414 | goto out_err; | |
4415 | } | |
8a630112 | 4416 | |
1c5ecae3 MK |
4417 | /* |
4418 | * There must be enough pages in the subpool for the mapping. If | |
4419 | * the subpool has a minimum size, there may be some global | |
4420 | * reservations already in place (gbl_reserve). | |
4421 | */ | |
4422 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4423 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4424 | ret = -ENOSPC; |
4425 | goto out_err; | |
4426 | } | |
5a6fe125 MG |
4427 | |
4428 | /* | |
17c9d12e | 4429 | * Check enough hugepages are available for the reservation. |
90481622 | 4430 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4431 | */ |
1c5ecae3 | 4432 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4433 | if (ret < 0) { |
1c5ecae3 MK |
4434 | /* put back original number of pages, chg */ |
4435 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4436 | goto out_err; |
68842c9b | 4437 | } |
17c9d12e MG |
4438 | |
4439 | /* | |
4440 | * Account for the reservations made. Shared mappings record regions | |
4441 | * that have reservations as they are shared by multiple VMAs. | |
4442 | * When the last VMA disappears, the region map says how much | |
4443 | * the reservation was and the page cache tells how much of | |
4444 | * the reservation was consumed. Private mappings are per-VMA and | |
4445 | * only the consumed reservations are tracked. When the VMA | |
4446 | * disappears, the original reservation is the VMA size and the | |
4447 | * consumed reservations are stored in the map. Hence, nothing | |
4448 | * else has to be done for private mappings here | |
4449 | */ | |
33039678 MK |
4450 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4451 | long add = region_add(resv_map, from, to); | |
4452 | ||
4453 | if (unlikely(chg > add)) { | |
4454 | /* | |
4455 | * pages in this range were added to the reserve | |
4456 | * map between region_chg and region_add. This | |
4457 | * indicates a race with alloc_huge_page. Adjust | |
4458 | * the subpool and reserve counts modified above | |
4459 | * based on the difference. | |
4460 | */ | |
4461 | long rsv_adjust; | |
4462 | ||
4463 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4464 | chg - add); | |
4465 | hugetlb_acct_memory(h, -rsv_adjust); | |
4466 | } | |
4467 | } | |
a43a8c39 | 4468 | return 0; |
c50ac050 | 4469 | out_err: |
5e911373 | 4470 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4471 | /* Don't call region_abort if region_chg failed */ |
4472 | if (chg >= 0) | |
4473 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4474 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4475 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4476 | return ret; |
a43a8c39 KC |
4477 | } |
4478 | ||
b5cec28d MK |
4479 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4480 | long freed) | |
a43a8c39 | 4481 | { |
a5516438 | 4482 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4483 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4484 | long chg = 0; |
90481622 | 4485 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4486 | long gbl_reserve; |
45c682a6 | 4487 | |
b5cec28d MK |
4488 | if (resv_map) { |
4489 | chg = region_del(resv_map, start, end); | |
4490 | /* | |
4491 | * region_del() can fail in the rare case where a region | |
4492 | * must be split and another region descriptor can not be | |
4493 | * allocated. If end == LONG_MAX, it will not fail. | |
4494 | */ | |
4495 | if (chg < 0) | |
4496 | return chg; | |
4497 | } | |
4498 | ||
45c682a6 | 4499 | spin_lock(&inode->i_lock); |
e4c6f8be | 4500 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4501 | spin_unlock(&inode->i_lock); |
4502 | ||
1c5ecae3 MK |
4503 | /* |
4504 | * If the subpool has a minimum size, the number of global | |
4505 | * reservations to be released may be adjusted. | |
4506 | */ | |
4507 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4508 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4509 | |
4510 | return 0; | |
a43a8c39 | 4511 | } |
93f70f90 | 4512 | |
3212b535 SC |
4513 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4514 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4515 | struct vm_area_struct *vma, | |
4516 | unsigned long addr, pgoff_t idx) | |
4517 | { | |
4518 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4519 | svma->vm_start; | |
4520 | unsigned long sbase = saddr & PUD_MASK; | |
4521 | unsigned long s_end = sbase + PUD_SIZE; | |
4522 | ||
4523 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4524 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4525 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4526 | |
4527 | /* | |
4528 | * match the virtual addresses, permission and the alignment of the | |
4529 | * page table page. | |
4530 | */ | |
4531 | if (pmd_index(addr) != pmd_index(saddr) || | |
4532 | vm_flags != svm_flags || | |
4533 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4534 | return 0; | |
4535 | ||
4536 | return saddr; | |
4537 | } | |
4538 | ||
31aafb45 | 4539 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4540 | { |
4541 | unsigned long base = addr & PUD_MASK; | |
4542 | unsigned long end = base + PUD_SIZE; | |
4543 | ||
4544 | /* | |
4545 | * check on proper vm_flags and page table alignment | |
4546 | */ | |
4547 | if (vma->vm_flags & VM_MAYSHARE && | |
4548 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4549 | return true; |
4550 | return false; | |
3212b535 SC |
4551 | } |
4552 | ||
4553 | /* | |
4554 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4555 | * and returns the corresponding pte. While this is not necessary for the | |
4556 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4557 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4558 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4559 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4560 | * bad pmd for sharing. | |
4561 | */ | |
4562 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4563 | { | |
4564 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4565 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4566 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4567 | vma->vm_pgoff; | |
4568 | struct vm_area_struct *svma; | |
4569 | unsigned long saddr; | |
4570 | pte_t *spte = NULL; | |
4571 | pte_t *pte; | |
cb900f41 | 4572 | spinlock_t *ptl; |
3212b535 SC |
4573 | |
4574 | if (!vma_shareable(vma, addr)) | |
4575 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4576 | ||
83cde9e8 | 4577 | i_mmap_lock_write(mapping); |
3212b535 SC |
4578 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4579 | if (svma == vma) | |
4580 | continue; | |
4581 | ||
4582 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4583 | if (saddr) { | |
7868a208 PA |
4584 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4585 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4586 | if (spte) { |
4587 | get_page(virt_to_page(spte)); | |
4588 | break; | |
4589 | } | |
4590 | } | |
4591 | } | |
4592 | ||
4593 | if (!spte) | |
4594 | goto out; | |
4595 | ||
8bea8052 | 4596 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4597 | if (pud_none(*pud)) { |
3212b535 SC |
4598 | pud_populate(mm, pud, |
4599 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4600 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4601 | } else { |
3212b535 | 4602 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4603 | } |
cb900f41 | 4604 | spin_unlock(ptl); |
3212b535 SC |
4605 | out: |
4606 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4607 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4608 | return pte; |
4609 | } | |
4610 | ||
4611 | /* | |
4612 | * unmap huge page backed by shared pte. | |
4613 | * | |
4614 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4615 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4616 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4617 | * | |
cb900f41 | 4618 | * called with page table lock held. |
3212b535 SC |
4619 | * |
4620 | * returns: 1 successfully unmapped a shared pte page | |
4621 | * 0 the underlying pte page is not shared, or it is the last user | |
4622 | */ | |
4623 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4624 | { | |
4625 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4626 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4627 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4628 | |
4629 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4630 | if (page_count(virt_to_page(ptep)) == 1) | |
4631 | return 0; | |
4632 | ||
4633 | pud_clear(pud); | |
4634 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4635 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4636 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4637 | return 1; | |
4638 | } | |
9e5fc74c SC |
4639 | #define want_pmd_share() (1) |
4640 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4641 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4642 | { | |
4643 | return NULL; | |
4644 | } | |
e81f2d22 ZZ |
4645 | |
4646 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4647 | { | |
4648 | return 0; | |
4649 | } | |
9e5fc74c | 4650 | #define want_pmd_share() (0) |
3212b535 SC |
4651 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4652 | ||
9e5fc74c SC |
4653 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4654 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4655 | unsigned long addr, unsigned long sz) | |
4656 | { | |
4657 | pgd_t *pgd; | |
c2febafc | 4658 | p4d_t *p4d; |
9e5fc74c SC |
4659 | pud_t *pud; |
4660 | pte_t *pte = NULL; | |
4661 | ||
4662 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4663 | p4d = p4d_offset(pgd, addr); |
4664 | pud = pud_alloc(mm, p4d, addr); | |
9e5fc74c SC |
4665 | if (pud) { |
4666 | if (sz == PUD_SIZE) { | |
4667 | pte = (pte_t *)pud; | |
4668 | } else { | |
4669 | BUG_ON(sz != PMD_SIZE); | |
4670 | if (want_pmd_share() && pud_none(*pud)) | |
4671 | pte = huge_pmd_share(mm, addr, pud); | |
4672 | else | |
4673 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4674 | } | |
4675 | } | |
4e666314 | 4676 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
4677 | |
4678 | return pte; | |
4679 | } | |
4680 | ||
7868a208 PA |
4681 | pte_t *huge_pte_offset(struct mm_struct *mm, |
4682 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
4683 | { |
4684 | pgd_t *pgd; | |
c2febafc | 4685 | p4d_t *p4d; |
9e5fc74c | 4686 | pud_t *pud; |
c2febafc | 4687 | pmd_t *pmd; |
9e5fc74c SC |
4688 | |
4689 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4690 | if (!pgd_present(*pgd)) |
4691 | return NULL; | |
4692 | p4d = p4d_offset(pgd, addr); | |
4693 | if (!p4d_present(*p4d)) | |
4694 | return NULL; | |
4695 | pud = pud_offset(p4d, addr); | |
4696 | if (!pud_present(*pud)) | |
4697 | return NULL; | |
4698 | if (pud_huge(*pud)) | |
4699 | return (pte_t *)pud; | |
4700 | pmd = pmd_offset(pud, addr); | |
9e5fc74c SC |
4701 | return (pte_t *) pmd; |
4702 | } | |
4703 | ||
61f77eda NH |
4704 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4705 | ||
4706 | /* | |
4707 | * These functions are overwritable if your architecture needs its own | |
4708 | * behavior. | |
4709 | */ | |
4710 | struct page * __weak | |
4711 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4712 | int write) | |
4713 | { | |
4714 | return ERR_PTR(-EINVAL); | |
4715 | } | |
4716 | ||
4dc71451 AK |
4717 | struct page * __weak |
4718 | follow_huge_pd(struct vm_area_struct *vma, | |
4719 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
4720 | { | |
4721 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
4722 | return NULL; | |
4723 | } | |
4724 | ||
61f77eda | 4725 | struct page * __weak |
9e5fc74c | 4726 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4727 | pmd_t *pmd, int flags) |
9e5fc74c | 4728 | { |
e66f17ff NH |
4729 | struct page *page = NULL; |
4730 | spinlock_t *ptl; | |
c9d398fa | 4731 | pte_t pte; |
e66f17ff NH |
4732 | retry: |
4733 | ptl = pmd_lockptr(mm, pmd); | |
4734 | spin_lock(ptl); | |
4735 | /* | |
4736 | * make sure that the address range covered by this pmd is not | |
4737 | * unmapped from other threads. | |
4738 | */ | |
4739 | if (!pmd_huge(*pmd)) | |
4740 | goto out; | |
c9d398fa NH |
4741 | pte = huge_ptep_get((pte_t *)pmd); |
4742 | if (pte_present(pte)) { | |
97534127 | 4743 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4744 | if (flags & FOLL_GET) |
4745 | get_page(page); | |
4746 | } else { | |
c9d398fa | 4747 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
4748 | spin_unlock(ptl); |
4749 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4750 | goto retry; | |
4751 | } | |
4752 | /* | |
4753 | * hwpoisoned entry is treated as no_page_table in | |
4754 | * follow_page_mask(). | |
4755 | */ | |
4756 | } | |
4757 | out: | |
4758 | spin_unlock(ptl); | |
9e5fc74c SC |
4759 | return page; |
4760 | } | |
4761 | ||
61f77eda | 4762 | struct page * __weak |
9e5fc74c | 4763 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4764 | pud_t *pud, int flags) |
9e5fc74c | 4765 | { |
e66f17ff NH |
4766 | if (flags & FOLL_GET) |
4767 | return NULL; | |
9e5fc74c | 4768 | |
e66f17ff | 4769 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4770 | } |
4771 | ||
faaa5b62 AK |
4772 | struct page * __weak |
4773 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
4774 | { | |
4775 | if (flags & FOLL_GET) | |
4776 | return NULL; | |
4777 | ||
4778 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
4779 | } | |
4780 | ||
31caf665 NH |
4781 | bool isolate_huge_page(struct page *page, struct list_head *list) |
4782 | { | |
bcc54222 NH |
4783 | bool ret = true; |
4784 | ||
309381fe | 4785 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4786 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4787 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4788 | ret = false; | |
4789 | goto unlock; | |
4790 | } | |
4791 | clear_page_huge_active(page); | |
31caf665 | 4792 | list_move_tail(&page->lru, list); |
bcc54222 | 4793 | unlock: |
31caf665 | 4794 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4795 | return ret; |
31caf665 NH |
4796 | } |
4797 | ||
4798 | void putback_active_hugepage(struct page *page) | |
4799 | { | |
309381fe | 4800 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4801 | spin_lock(&hugetlb_lock); |
bcc54222 | 4802 | set_page_huge_active(page); |
31caf665 NH |
4803 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4804 | spin_unlock(&hugetlb_lock); | |
4805 | put_page(page); | |
4806 | } |