]>
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 | ||
3e59fcb0 MH |
890 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
891 | nodemask_t *nmask) | |
94310cbc | 892 | { |
3e59fcb0 MH |
893 | unsigned int cpuset_mems_cookie; |
894 | struct zonelist *zonelist; | |
895 | struct zone *zone; | |
896 | struct zoneref *z; | |
897 | int node = -1; | |
94310cbc | 898 | |
3e59fcb0 MH |
899 | zonelist = node_zonelist(nid, gfp_mask); |
900 | ||
901 | retry_cpuset: | |
902 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
903 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
904 | struct page *page; | |
905 | ||
906 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
907 | continue; | |
908 | /* | |
909 | * no need to ask again on the same node. Pool is node rather than | |
910 | * zone aware | |
911 | */ | |
912 | if (zone_to_nid(zone) == node) | |
913 | continue; | |
914 | node = zone_to_nid(zone); | |
94310cbc | 915 | |
94310cbc AK |
916 | page = dequeue_huge_page_node_exact(h, node); |
917 | if (page) | |
918 | return page; | |
919 | } | |
3e59fcb0 MH |
920 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
921 | goto retry_cpuset; | |
922 | ||
94310cbc AK |
923 | return NULL; |
924 | } | |
925 | ||
86cdb465 NH |
926 | /* Movability of hugepages depends on migration support. */ |
927 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
928 | { | |
100873d7 | 929 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
930 | return GFP_HIGHUSER_MOVABLE; |
931 | else | |
932 | return GFP_HIGHUSER; | |
933 | } | |
934 | ||
a5516438 AK |
935 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
936 | struct vm_area_struct *vma, | |
af0ed73e JK |
937 | unsigned long address, int avoid_reserve, |
938 | long chg) | |
1da177e4 | 939 | { |
3e59fcb0 | 940 | struct page *page; |
480eccf9 | 941 | struct mempolicy *mpol; |
04ec6264 | 942 | gfp_t gfp_mask; |
3e59fcb0 | 943 | nodemask_t *nodemask; |
04ec6264 | 944 | int nid; |
1da177e4 | 945 | |
a1e78772 MG |
946 | /* |
947 | * A child process with MAP_PRIVATE mappings created by their parent | |
948 | * have no page reserves. This check ensures that reservations are | |
949 | * not "stolen". The child may still get SIGKILLed | |
950 | */ | |
af0ed73e | 951 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 952 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 953 | goto err; |
a1e78772 | 954 | |
04f2cbe3 | 955 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 956 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 957 | goto err; |
04f2cbe3 | 958 | |
04ec6264 VB |
959 | gfp_mask = htlb_alloc_mask(h); |
960 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
961 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
962 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
963 | SetPagePrivate(page); | |
964 | h->resv_huge_pages--; | |
1da177e4 | 965 | } |
cc9a6c87 | 966 | |
52cd3b07 | 967 | mpol_cond_put(mpol); |
1da177e4 | 968 | return page; |
cc9a6c87 MG |
969 | |
970 | err: | |
cc9a6c87 | 971 | return NULL; |
1da177e4 LT |
972 | } |
973 | ||
1cac6f2c LC |
974 | /* |
975 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
976 | * We may have allocated or freed a huge page based on a different | |
977 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
978 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
979 | * node for alloc or free. | |
980 | */ | |
981 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
982 | { | |
0edaf86c | 983 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
984 | VM_BUG_ON(nid >= MAX_NUMNODES); |
985 | ||
986 | return nid; | |
987 | } | |
988 | ||
989 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
990 | { | |
991 | if (!node_isset(nid, *nodes_allowed)) | |
992 | nid = next_node_allowed(nid, nodes_allowed); | |
993 | return nid; | |
994 | } | |
995 | ||
996 | /* | |
997 | * returns the previously saved node ["this node"] from which to | |
998 | * allocate a persistent huge page for the pool and advance the | |
999 | * next node from which to allocate, handling wrap at end of node | |
1000 | * mask. | |
1001 | */ | |
1002 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1003 | nodemask_t *nodes_allowed) | |
1004 | { | |
1005 | int nid; | |
1006 | ||
1007 | VM_BUG_ON(!nodes_allowed); | |
1008 | ||
1009 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1010 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1011 | ||
1012 | return nid; | |
1013 | } | |
1014 | ||
1015 | /* | |
1016 | * helper for free_pool_huge_page() - return the previously saved | |
1017 | * node ["this node"] from which to free a huge page. Advance the | |
1018 | * next node id whether or not we find a free huge page to free so | |
1019 | * that the next attempt to free addresses the next node. | |
1020 | */ | |
1021 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1022 | { | |
1023 | int nid; | |
1024 | ||
1025 | VM_BUG_ON(!nodes_allowed); | |
1026 | ||
1027 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1028 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1029 | ||
1030 | return nid; | |
1031 | } | |
1032 | ||
1033 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1034 | for (nr_nodes = nodes_weight(*mask); \ | |
1035 | nr_nodes > 0 && \ | |
1036 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1037 | nr_nodes--) | |
1038 | ||
1039 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1040 | for (nr_nodes = nodes_weight(*mask); \ | |
1041 | nr_nodes > 0 && \ | |
1042 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1043 | nr_nodes--) | |
1044 | ||
e1073d1e | 1045 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1046 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1047 | unsigned int order) |
944d9fec LC |
1048 | { |
1049 | int i; | |
1050 | int nr_pages = 1 << order; | |
1051 | struct page *p = page + 1; | |
1052 | ||
c8cc708a | 1053 | atomic_set(compound_mapcount_ptr(page), 0); |
944d9fec | 1054 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1055 | clear_compound_head(p); |
944d9fec | 1056 | set_page_refcounted(p); |
944d9fec LC |
1057 | } |
1058 | ||
1059 | set_compound_order(page, 0); | |
1060 | __ClearPageHead(page); | |
1061 | } | |
1062 | ||
d00181b9 | 1063 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1064 | { |
1065 | free_contig_range(page_to_pfn(page), 1 << order); | |
1066 | } | |
1067 | ||
1068 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
79b63f12 | 1069 | unsigned long nr_pages, gfp_t gfp_mask) |
944d9fec LC |
1070 | { |
1071 | unsigned long end_pfn = start_pfn + nr_pages; | |
ca96b625 | 1072 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, |
79b63f12 | 1073 | gfp_mask); |
944d9fec LC |
1074 | } |
1075 | ||
f44b2dda JK |
1076 | static bool pfn_range_valid_gigantic(struct zone *z, |
1077 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1078 | { |
1079 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1080 | struct page *page; | |
1081 | ||
1082 | for (i = start_pfn; i < end_pfn; i++) { | |
1083 | if (!pfn_valid(i)) | |
1084 | return false; | |
1085 | ||
1086 | page = pfn_to_page(i); | |
1087 | ||
f44b2dda JK |
1088 | if (page_zone(page) != z) |
1089 | return false; | |
1090 | ||
944d9fec LC |
1091 | if (PageReserved(page)) |
1092 | return false; | |
1093 | ||
1094 | if (page_count(page) > 0) | |
1095 | return false; | |
1096 | ||
1097 | if (PageHuge(page)) | |
1098 | return false; | |
1099 | } | |
1100 | ||
1101 | return true; | |
1102 | } | |
1103 | ||
1104 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1105 | unsigned long start_pfn, unsigned long nr_pages) | |
1106 | { | |
1107 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1108 | return zone_spans_pfn(zone, last_pfn); | |
1109 | } | |
1110 | ||
79b63f12 | 1111 | static struct page *alloc_gigantic_page(int nid, struct hstate *h) |
944d9fec | 1112 | { |
79b63f12 | 1113 | unsigned int order = huge_page_order(h); |
944d9fec LC |
1114 | unsigned long nr_pages = 1 << order; |
1115 | unsigned long ret, pfn, flags; | |
79b63f12 MH |
1116 | struct zonelist *zonelist; |
1117 | struct zone *zone; | |
1118 | struct zoneref *z; | |
1119 | gfp_t gfp_mask; | |
944d9fec | 1120 | |
79b63f12 MH |
1121 | gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
1122 | zonelist = node_zonelist(nid, gfp_mask); | |
1123 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), NULL) { | |
1124 | spin_lock_irqsave(&zone->lock, flags); | |
944d9fec | 1125 | |
79b63f12 MH |
1126 | pfn = ALIGN(zone->zone_start_pfn, nr_pages); |
1127 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) { | |
1128 | if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { | |
944d9fec LC |
1129 | /* |
1130 | * We release the zone lock here because | |
1131 | * alloc_contig_range() will also lock the zone | |
1132 | * at some point. If there's an allocation | |
1133 | * spinning on this lock, it may win the race | |
1134 | * and cause alloc_contig_range() to fail... | |
1135 | */ | |
79b63f12 MH |
1136 | spin_unlock_irqrestore(&zone->lock, flags); |
1137 | ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); | |
944d9fec LC |
1138 | if (!ret) |
1139 | return pfn_to_page(pfn); | |
79b63f12 | 1140 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec LC |
1141 | } |
1142 | pfn += nr_pages; | |
1143 | } | |
1144 | ||
79b63f12 | 1145 | spin_unlock_irqrestore(&zone->lock, flags); |
944d9fec LC |
1146 | } |
1147 | ||
1148 | return NULL; | |
1149 | } | |
1150 | ||
1151 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1152 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
944d9fec LC |
1153 | |
1154 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
1155 | { | |
1156 | struct page *page; | |
1157 | ||
79b63f12 | 1158 | page = alloc_gigantic_page(nid, h); |
944d9fec LC |
1159 | if (page) { |
1160 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1161 | prep_new_huge_page(h, page, nid); | |
1162 | } | |
1163 | ||
1164 | return page; | |
1165 | } | |
1166 | ||
1167 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
1168 | nodemask_t *nodes_allowed) | |
1169 | { | |
1170 | struct page *page = NULL; | |
1171 | int nr_nodes, node; | |
1172 | ||
1173 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1174 | page = alloc_fresh_gigantic_page_node(h, node); | |
1175 | if (page) | |
1176 | return 1; | |
1177 | } | |
1178 | ||
1179 | return 0; | |
1180 | } | |
1181 | ||
e1073d1e | 1182 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
944d9fec | 1183 | static inline bool gigantic_page_supported(void) { return false; } |
d00181b9 | 1184 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1185 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1186 | unsigned int order) { } |
944d9fec LC |
1187 | static inline int alloc_fresh_gigantic_page(struct hstate *h, |
1188 | nodemask_t *nodes_allowed) { return 0; } | |
1189 | #endif | |
1190 | ||
a5516438 | 1191 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1192 | { |
1193 | int i; | |
a5516438 | 1194 | |
944d9fec LC |
1195 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1196 | return; | |
18229df5 | 1197 | |
a5516438 AK |
1198 | h->nr_huge_pages--; |
1199 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1200 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1201 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1202 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1203 | 1 << PG_active | 1 << PG_private | |
1204 | 1 << PG_writeback); | |
6af2acb6 | 1205 | } |
309381fe | 1206 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1207 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1208 | set_page_refcounted(page); |
944d9fec LC |
1209 | if (hstate_is_gigantic(h)) { |
1210 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1211 | free_gigantic_page(page, huge_page_order(h)); | |
1212 | } else { | |
944d9fec LC |
1213 | __free_pages(page, huge_page_order(h)); |
1214 | } | |
6af2acb6 AL |
1215 | } |
1216 | ||
e5ff2159 AK |
1217 | struct hstate *size_to_hstate(unsigned long size) |
1218 | { | |
1219 | struct hstate *h; | |
1220 | ||
1221 | for_each_hstate(h) { | |
1222 | if (huge_page_size(h) == size) | |
1223 | return h; | |
1224 | } | |
1225 | return NULL; | |
1226 | } | |
1227 | ||
bcc54222 NH |
1228 | /* |
1229 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1230 | * to hstate->hugepage_activelist.) | |
1231 | * | |
1232 | * This function can be called for tail pages, but never returns true for them. | |
1233 | */ | |
1234 | bool page_huge_active(struct page *page) | |
1235 | { | |
1236 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1237 | return PageHead(page) && PagePrivate(&page[1]); | |
1238 | } | |
1239 | ||
1240 | /* never called for tail page */ | |
1241 | static void set_page_huge_active(struct page *page) | |
1242 | { | |
1243 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1244 | SetPagePrivate(&page[1]); | |
1245 | } | |
1246 | ||
1247 | static void clear_page_huge_active(struct page *page) | |
1248 | { | |
1249 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1250 | ClearPagePrivate(&page[1]); | |
1251 | } | |
1252 | ||
8f1d26d0 | 1253 | void free_huge_page(struct page *page) |
27a85ef1 | 1254 | { |
a5516438 AK |
1255 | /* |
1256 | * Can't pass hstate in here because it is called from the | |
1257 | * compound page destructor. | |
1258 | */ | |
e5ff2159 | 1259 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1260 | int nid = page_to_nid(page); |
90481622 DG |
1261 | struct hugepage_subpool *spool = |
1262 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1263 | bool restore_reserve; |
27a85ef1 | 1264 | |
e5df70ab | 1265 | set_page_private(page, 0); |
23be7468 | 1266 | page->mapping = NULL; |
b4330afb MK |
1267 | VM_BUG_ON_PAGE(page_count(page), page); |
1268 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
07443a85 | 1269 | restore_reserve = PagePrivate(page); |
16c794b4 | 1270 | ClearPagePrivate(page); |
27a85ef1 | 1271 | |
1c5ecae3 MK |
1272 | /* |
1273 | * A return code of zero implies that the subpool will be under its | |
1274 | * minimum size if the reservation is not restored after page is free. | |
1275 | * Therefore, force restore_reserve operation. | |
1276 | */ | |
1277 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1278 | restore_reserve = true; | |
1279 | ||
27a85ef1 | 1280 | spin_lock(&hugetlb_lock); |
bcc54222 | 1281 | clear_page_huge_active(page); |
6d76dcf4 AK |
1282 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1283 | pages_per_huge_page(h), page); | |
07443a85 JK |
1284 | if (restore_reserve) |
1285 | h->resv_huge_pages++; | |
1286 | ||
944d9fec | 1287 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1288 | /* remove the page from active list */ |
1289 | list_del(&page->lru); | |
a5516438 AK |
1290 | update_and_free_page(h, page); |
1291 | h->surplus_huge_pages--; | |
1292 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1293 | } else { |
5d3a551c | 1294 | arch_clear_hugepage_flags(page); |
a5516438 | 1295 | enqueue_huge_page(h, page); |
7893d1d5 | 1296 | } |
27a85ef1 DG |
1297 | spin_unlock(&hugetlb_lock); |
1298 | } | |
1299 | ||
a5516438 | 1300 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1301 | { |
0edaecfa | 1302 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1303 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1304 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1305 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1306 | h->nr_huge_pages++; |
1307 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1308 | spin_unlock(&hugetlb_lock); |
1309 | put_page(page); /* free it into the hugepage allocator */ | |
1310 | } | |
1311 | ||
d00181b9 | 1312 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1313 | { |
1314 | int i; | |
1315 | int nr_pages = 1 << order; | |
1316 | struct page *p = page + 1; | |
1317 | ||
1318 | /* we rely on prep_new_huge_page to set the destructor */ | |
1319 | set_compound_order(page, order); | |
ef5a22be | 1320 | __ClearPageReserved(page); |
de09d31d | 1321 | __SetPageHead(page); |
20a0307c | 1322 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1323 | /* |
1324 | * For gigantic hugepages allocated through bootmem at | |
1325 | * boot, it's safer to be consistent with the not-gigantic | |
1326 | * hugepages and clear the PG_reserved bit from all tail pages | |
1327 | * too. Otherwse drivers using get_user_pages() to access tail | |
1328 | * pages may get the reference counting wrong if they see | |
1329 | * PG_reserved set on a tail page (despite the head page not | |
1330 | * having PG_reserved set). Enforcing this consistency between | |
1331 | * head and tail pages allows drivers to optimize away a check | |
1332 | * on the head page when they need know if put_page() is needed | |
1333 | * after get_user_pages(). | |
1334 | */ | |
1335 | __ClearPageReserved(p); | |
58a84aa9 | 1336 | set_page_count(p, 0); |
1d798ca3 | 1337 | set_compound_head(p, page); |
20a0307c | 1338 | } |
b4330afb | 1339 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1340 | } |
1341 | ||
7795912c AM |
1342 | /* |
1343 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1344 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1345 | * details. | |
1346 | */ | |
20a0307c WF |
1347 | int PageHuge(struct page *page) |
1348 | { | |
20a0307c WF |
1349 | if (!PageCompound(page)) |
1350 | return 0; | |
1351 | ||
1352 | page = compound_head(page); | |
f1e61557 | 1353 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1354 | } |
43131e14 NH |
1355 | EXPORT_SYMBOL_GPL(PageHuge); |
1356 | ||
27c73ae7 AA |
1357 | /* |
1358 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1359 | * normal or transparent huge pages. | |
1360 | */ | |
1361 | int PageHeadHuge(struct page *page_head) | |
1362 | { | |
27c73ae7 AA |
1363 | if (!PageHead(page_head)) |
1364 | return 0; | |
1365 | ||
758f66a2 | 1366 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1367 | } |
27c73ae7 | 1368 | |
13d60f4b ZY |
1369 | pgoff_t __basepage_index(struct page *page) |
1370 | { | |
1371 | struct page *page_head = compound_head(page); | |
1372 | pgoff_t index = page_index(page_head); | |
1373 | unsigned long compound_idx; | |
1374 | ||
1375 | if (!PageHuge(page_head)) | |
1376 | return page_index(page); | |
1377 | ||
1378 | if (compound_order(page_head) >= MAX_ORDER) | |
1379 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1380 | else | |
1381 | compound_idx = page - page_head; | |
1382 | ||
1383 | return (index << compound_order(page_head)) + compound_idx; | |
1384 | } | |
1385 | ||
a5516438 | 1386 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1387 | { |
1da177e4 | 1388 | struct page *page; |
f96efd58 | 1389 | |
96db800f | 1390 | page = __alloc_pages_node(nid, |
86cdb465 | 1391 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
dcda9b04 | 1392 | __GFP_RETRY_MAYFAIL|__GFP_NOWARN, |
a5516438 | 1393 | huge_page_order(h)); |
1da177e4 | 1394 | if (page) { |
a5516438 | 1395 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1396 | } |
63b4613c NA |
1397 | |
1398 | return page; | |
1399 | } | |
1400 | ||
b2261026 JK |
1401 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1402 | { | |
1403 | struct page *page; | |
1404 | int nr_nodes, node; | |
1405 | int ret = 0; | |
1406 | ||
1407 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1408 | page = alloc_fresh_huge_page_node(h, node); | |
1409 | if (page) { | |
1410 | ret = 1; | |
1411 | break; | |
1412 | } | |
1413 | } | |
1414 | ||
1415 | if (ret) | |
1416 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1417 | else | |
1418 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1419 | ||
1420 | return ret; | |
1421 | } | |
1422 | ||
e8c5c824 LS |
1423 | /* |
1424 | * Free huge page from pool from next node to free. | |
1425 | * Attempt to keep persistent huge pages more or less | |
1426 | * balanced over allowed nodes. | |
1427 | * Called with hugetlb_lock locked. | |
1428 | */ | |
6ae11b27 LS |
1429 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1430 | bool acct_surplus) | |
e8c5c824 | 1431 | { |
b2261026 | 1432 | int nr_nodes, node; |
e8c5c824 LS |
1433 | int ret = 0; |
1434 | ||
b2261026 | 1435 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1436 | /* |
1437 | * If we're returning unused surplus pages, only examine | |
1438 | * nodes with surplus pages. | |
1439 | */ | |
b2261026 JK |
1440 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1441 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1442 | struct page *page = |
b2261026 | 1443 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1444 | struct page, lru); |
1445 | list_del(&page->lru); | |
1446 | h->free_huge_pages--; | |
b2261026 | 1447 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1448 | if (acct_surplus) { |
1449 | h->surplus_huge_pages--; | |
b2261026 | 1450 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1451 | } |
e8c5c824 LS |
1452 | update_and_free_page(h, page); |
1453 | ret = 1; | |
9a76db09 | 1454 | break; |
e8c5c824 | 1455 | } |
b2261026 | 1456 | } |
e8c5c824 LS |
1457 | |
1458 | return ret; | |
1459 | } | |
1460 | ||
c8721bbb NH |
1461 | /* |
1462 | * Dissolve a given free hugepage into free buddy pages. This function does | |
082d5b6b GS |
1463 | * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the |
1464 | * number of free hugepages would be reduced below the number of reserved | |
1465 | * hugepages. | |
c8721bbb | 1466 | */ |
c3114a84 | 1467 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1468 | { |
082d5b6b GS |
1469 | int rc = 0; |
1470 | ||
c8721bbb NH |
1471 | spin_lock(&hugetlb_lock); |
1472 | if (PageHuge(page) && !page_count(page)) { | |
2247bb33 GS |
1473 | struct page *head = compound_head(page); |
1474 | struct hstate *h = page_hstate(head); | |
1475 | int nid = page_to_nid(head); | |
082d5b6b GS |
1476 | if (h->free_huge_pages - h->resv_huge_pages == 0) { |
1477 | rc = -EBUSY; | |
1478 | goto out; | |
1479 | } | |
c3114a84 AK |
1480 | /* |
1481 | * Move PageHWPoison flag from head page to the raw error page, | |
1482 | * which makes any subpages rather than the error page reusable. | |
1483 | */ | |
1484 | if (PageHWPoison(head) && page != head) { | |
1485 | SetPageHWPoison(page); | |
1486 | ClearPageHWPoison(head); | |
1487 | } | |
2247bb33 | 1488 | list_del(&head->lru); |
c8721bbb NH |
1489 | h->free_huge_pages--; |
1490 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1491 | h->max_huge_pages--; |
2247bb33 | 1492 | update_and_free_page(h, head); |
c8721bbb | 1493 | } |
082d5b6b | 1494 | out: |
c8721bbb | 1495 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1496 | return rc; |
c8721bbb NH |
1497 | } |
1498 | ||
1499 | /* | |
1500 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1501 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1502 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1503 | * part of it lies within the given range. | |
082d5b6b GS |
1504 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1505 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1506 | */ |
082d5b6b | 1507 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1508 | { |
c8721bbb | 1509 | unsigned long pfn; |
eb03aa00 | 1510 | struct page *page; |
082d5b6b | 1511 | int rc = 0; |
c8721bbb | 1512 | |
d0177639 | 1513 | if (!hugepages_supported()) |
082d5b6b | 1514 | return rc; |
d0177639 | 1515 | |
eb03aa00 GS |
1516 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1517 | page = pfn_to_page(pfn); | |
1518 | if (PageHuge(page) && !page_count(page)) { | |
1519 | rc = dissolve_free_huge_page(page); | |
1520 | if (rc) | |
1521 | break; | |
1522 | } | |
1523 | } | |
082d5b6b GS |
1524 | |
1525 | return rc; | |
c8721bbb NH |
1526 | } |
1527 | ||
099730d6 | 1528 | static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, |
aaf14e40 | 1529 | gfp_t gfp_mask, int nid, nodemask_t *nmask) |
099730d6 DH |
1530 | { |
1531 | int order = huge_page_order(h); | |
099730d6 | 1532 | |
dcda9b04 | 1533 | gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; |
aaf14e40 MH |
1534 | if (nid == NUMA_NO_NODE) |
1535 | nid = numa_mem_id(); | |
1536 | return __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
099730d6 DH |
1537 | } |
1538 | ||
aaf14e40 MH |
1539 | static struct page *__alloc_buddy_huge_page(struct hstate *h, gfp_t gfp_mask, |
1540 | int nid, nodemask_t *nmask) | |
7893d1d5 AL |
1541 | { |
1542 | struct page *page; | |
bf50bab2 | 1543 | unsigned int r_nid; |
7893d1d5 | 1544 | |
bae7f4ae | 1545 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1546 | return NULL; |
1547 | ||
d1c3fb1f NA |
1548 | /* |
1549 | * Assume we will successfully allocate the surplus page to | |
1550 | * prevent racing processes from causing the surplus to exceed | |
1551 | * overcommit | |
1552 | * | |
1553 | * This however introduces a different race, where a process B | |
1554 | * tries to grow the static hugepage pool while alloc_pages() is | |
1555 | * called by process A. B will only examine the per-node | |
1556 | * counters in determining if surplus huge pages can be | |
1557 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1558 | * won't be able to increment the per-node counter, until the | |
1559 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1560 | * no more huge pages can be converted from surplus to normal | |
1561 | * state (and doesn't try to convert again). Thus, we have a | |
1562 | * case where a surplus huge page exists, the pool is grown, and | |
1563 | * the surplus huge page still exists after, even though it | |
1564 | * should just have been converted to a normal huge page. This | |
1565 | * does not leak memory, though, as the hugepage will be freed | |
1566 | * once it is out of use. It also does not allow the counters to | |
1567 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1568 | * the node values until we've gotten the hugepage and only the | |
1569 | * per-node value is checked there. | |
1570 | */ | |
1571 | spin_lock(&hugetlb_lock); | |
a5516438 | 1572 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1573 | spin_unlock(&hugetlb_lock); |
1574 | return NULL; | |
1575 | } else { | |
a5516438 AK |
1576 | h->nr_huge_pages++; |
1577 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1578 | } |
1579 | spin_unlock(&hugetlb_lock); | |
1580 | ||
aaf14e40 | 1581 | page = __hugetlb_alloc_buddy_huge_page(h, gfp_mask, nid, nmask); |
d1c3fb1f NA |
1582 | |
1583 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 1584 | if (page) { |
0edaecfa | 1585 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1586 | r_nid = page_to_nid(page); |
f1e61557 | 1587 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
9dd540e2 | 1588 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1589 | /* |
1590 | * We incremented the global counters already | |
1591 | */ | |
bf50bab2 NH |
1592 | h->nr_huge_pages_node[r_nid]++; |
1593 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1594 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1595 | } else { |
a5516438 AK |
1596 | h->nr_huge_pages--; |
1597 | h->surplus_huge_pages--; | |
3b116300 | 1598 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1599 | } |
d1c3fb1f | 1600 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1601 | |
1602 | return page; | |
1603 | } | |
1604 | ||
099730d6 DH |
1605 | /* |
1606 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1607 | */ | |
e0ec90ee | 1608 | static |
099730d6 DH |
1609 | struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, |
1610 | struct vm_area_struct *vma, unsigned long addr) | |
1611 | { | |
aaf14e40 MH |
1612 | struct page *page; |
1613 | struct mempolicy *mpol; | |
1614 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1615 | int nid; | |
1616 | nodemask_t *nodemask; | |
1617 | ||
1618 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
1619 | page = __alloc_buddy_huge_page(h, gfp_mask, nid, nodemask); | |
1620 | mpol_cond_put(mpol); | |
1621 | ||
1622 | return page; | |
099730d6 DH |
1623 | } |
1624 | ||
bf50bab2 NH |
1625 | /* |
1626 | * This allocation function is useful in the context where vma is irrelevant. | |
1627 | * E.g. soft-offlining uses this function because it only cares physical | |
1628 | * address of error page. | |
1629 | */ | |
1630 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1631 | { | |
aaf14e40 | 1632 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1633 | struct page *page = NULL; |
bf50bab2 | 1634 | |
aaf14e40 MH |
1635 | if (nid != NUMA_NO_NODE) |
1636 | gfp_mask |= __GFP_THISNODE; | |
1637 | ||
bf50bab2 | 1638 | spin_lock(&hugetlb_lock); |
4ef91848 | 1639 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1640 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1641 | spin_unlock(&hugetlb_lock); |
1642 | ||
94ae8ba7 | 1643 | if (!page) |
aaf14e40 | 1644 | page = __alloc_buddy_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1645 | |
1646 | return page; | |
1647 | } | |
1648 | ||
3e59fcb0 MH |
1649 | |
1650 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, | |
1651 | nodemask_t *nmask) | |
4db9b2ef | 1652 | { |
aaf14e40 | 1653 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
1654 | |
1655 | spin_lock(&hugetlb_lock); | |
1656 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1657 | struct page *page; |
1658 | ||
1659 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1660 | if (page) { | |
1661 | spin_unlock(&hugetlb_lock); | |
1662 | return page; | |
4db9b2ef MH |
1663 | } |
1664 | } | |
1665 | spin_unlock(&hugetlb_lock); | |
4db9b2ef MH |
1666 | |
1667 | /* No reservations, try to overcommit */ | |
3e59fcb0 MH |
1668 | |
1669 | return __alloc_buddy_huge_page(h, gfp_mask, preferred_nid, nmask); | |
4db9b2ef MH |
1670 | } |
1671 | ||
e4e574b7 | 1672 | /* |
25985edc | 1673 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1674 | * of size 'delta'. |
1675 | */ | |
a5516438 | 1676 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1677 | { |
1678 | struct list_head surplus_list; | |
1679 | struct page *page, *tmp; | |
1680 | int ret, i; | |
1681 | int needed, allocated; | |
28073b02 | 1682 | bool alloc_ok = true; |
e4e574b7 | 1683 | |
a5516438 | 1684 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1685 | if (needed <= 0) { |
a5516438 | 1686 | h->resv_huge_pages += delta; |
e4e574b7 | 1687 | return 0; |
ac09b3a1 | 1688 | } |
e4e574b7 AL |
1689 | |
1690 | allocated = 0; | |
1691 | INIT_LIST_HEAD(&surplus_list); | |
1692 | ||
1693 | ret = -ENOMEM; | |
1694 | retry: | |
1695 | spin_unlock(&hugetlb_lock); | |
1696 | for (i = 0; i < needed; i++) { | |
aaf14e40 MH |
1697 | page = __alloc_buddy_huge_page(h, htlb_alloc_mask(h), |
1698 | NUMA_NO_NODE, NULL); | |
28073b02 HD |
1699 | if (!page) { |
1700 | alloc_ok = false; | |
1701 | break; | |
1702 | } | |
e4e574b7 | 1703 | list_add(&page->lru, &surplus_list); |
69ed779a | 1704 | cond_resched(); |
e4e574b7 | 1705 | } |
28073b02 | 1706 | allocated += i; |
e4e574b7 AL |
1707 | |
1708 | /* | |
1709 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1710 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1711 | */ | |
1712 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1713 | needed = (h->resv_huge_pages + delta) - |
1714 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1715 | if (needed > 0) { |
1716 | if (alloc_ok) | |
1717 | goto retry; | |
1718 | /* | |
1719 | * We were not able to allocate enough pages to | |
1720 | * satisfy the entire reservation so we free what | |
1721 | * we've allocated so far. | |
1722 | */ | |
1723 | goto free; | |
1724 | } | |
e4e574b7 AL |
1725 | /* |
1726 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1727 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1728 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1729 | * allocator. Commit the entire reservation here to prevent another |
1730 | * process from stealing the pages as they are added to the pool but | |
1731 | * before they are reserved. | |
e4e574b7 AL |
1732 | */ |
1733 | needed += allocated; | |
a5516438 | 1734 | h->resv_huge_pages += delta; |
e4e574b7 | 1735 | ret = 0; |
a9869b83 | 1736 | |
19fc3f0a | 1737 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1738 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1739 | if ((--needed) < 0) |
1740 | break; | |
a9869b83 NH |
1741 | /* |
1742 | * This page is now managed by the hugetlb allocator and has | |
1743 | * no users -- drop the buddy allocator's reference. | |
1744 | */ | |
1745 | put_page_testzero(page); | |
309381fe | 1746 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1747 | enqueue_huge_page(h, page); |
19fc3f0a | 1748 | } |
28073b02 | 1749 | free: |
b0365c8d | 1750 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1751 | |
1752 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1753 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1754 | put_page(page); | |
a9869b83 | 1755 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1756 | |
1757 | return ret; | |
1758 | } | |
1759 | ||
1760 | /* | |
e5bbc8a6 MK |
1761 | * This routine has two main purposes: |
1762 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1763 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1764 | * to the associated reservation map. | |
1765 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1766 | * the reservation. As many as unused_resv_pages may be freed. | |
1767 | * | |
1768 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1769 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1770 | * we must make sure nobody else can claim pages we are in the process of | |
1771 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1772 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1773 | */ |
a5516438 AK |
1774 | static void return_unused_surplus_pages(struct hstate *h, |
1775 | unsigned long unused_resv_pages) | |
e4e574b7 | 1776 | { |
e4e574b7 AL |
1777 | unsigned long nr_pages; |
1778 | ||
aa888a74 | 1779 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1780 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1781 | goto out; |
aa888a74 | 1782 | |
e5bbc8a6 MK |
1783 | /* |
1784 | * Part (or even all) of the reservation could have been backed | |
1785 | * by pre-allocated pages. Only free surplus pages. | |
1786 | */ | |
a5516438 | 1787 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1788 | |
685f3457 LS |
1789 | /* |
1790 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1791 | * evenly across all nodes with memory. Iterate across these nodes |
1792 | * until we can no longer free unreserved surplus pages. This occurs | |
1793 | * when the nodes with surplus pages have no free pages. | |
1794 | * free_pool_huge_page() will balance the the freed pages across the | |
1795 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1796 | * |
1797 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1798 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1799 | * to cover subsequent pages we may free. | |
685f3457 LS |
1800 | */ |
1801 | while (nr_pages--) { | |
e5bbc8a6 MK |
1802 | h->resv_huge_pages--; |
1803 | unused_resv_pages--; | |
8cebfcd0 | 1804 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1805 | goto out; |
7848a4bf | 1806 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1807 | } |
e5bbc8a6 MK |
1808 | |
1809 | out: | |
1810 | /* Fully uncommit the reservation */ | |
1811 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1812 | } |
1813 | ||
5e911373 | 1814 | |
c37f9fb1 | 1815 | /* |
feba16e2 | 1816 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1817 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1818 | * |
1819 | * vma_needs_reservation is called to determine if the huge page at addr | |
1820 | * within the vma has an associated reservation. If a reservation is | |
1821 | * needed, the value 1 is returned. The caller is then responsible for | |
1822 | * managing the global reservation and subpool usage counts. After | |
1823 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1824 | * to add the page to the reservation map. If the page allocation fails, |
1825 | * the reservation must be ended instead of committed. vma_end_reservation | |
1826 | * is called in such cases. | |
cf3ad20b MK |
1827 | * |
1828 | * In the normal case, vma_commit_reservation returns the same value | |
1829 | * as the preceding vma_needs_reservation call. The only time this | |
1830 | * is not the case is if a reserve map was changed between calls. It | |
1831 | * is the responsibility of the caller to notice the difference and | |
1832 | * take appropriate action. | |
96b96a96 MK |
1833 | * |
1834 | * vma_add_reservation is used in error paths where a reservation must | |
1835 | * be restored when a newly allocated huge page must be freed. It is | |
1836 | * to be called after calling vma_needs_reservation to determine if a | |
1837 | * reservation exists. | |
c37f9fb1 | 1838 | */ |
5e911373 MK |
1839 | enum vma_resv_mode { |
1840 | VMA_NEEDS_RESV, | |
1841 | VMA_COMMIT_RESV, | |
feba16e2 | 1842 | VMA_END_RESV, |
96b96a96 | 1843 | VMA_ADD_RESV, |
5e911373 | 1844 | }; |
cf3ad20b MK |
1845 | static long __vma_reservation_common(struct hstate *h, |
1846 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1847 | enum vma_resv_mode mode) |
c37f9fb1 | 1848 | { |
4e35f483 JK |
1849 | struct resv_map *resv; |
1850 | pgoff_t idx; | |
cf3ad20b | 1851 | long ret; |
c37f9fb1 | 1852 | |
4e35f483 JK |
1853 | resv = vma_resv_map(vma); |
1854 | if (!resv) | |
84afd99b | 1855 | return 1; |
c37f9fb1 | 1856 | |
4e35f483 | 1857 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1858 | switch (mode) { |
1859 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1860 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1861 | break; |
1862 | case VMA_COMMIT_RESV: | |
1863 | ret = region_add(resv, idx, idx + 1); | |
1864 | break; | |
feba16e2 | 1865 | case VMA_END_RESV: |
5e911373 MK |
1866 | region_abort(resv, idx, idx + 1); |
1867 | ret = 0; | |
1868 | break; | |
96b96a96 MK |
1869 | case VMA_ADD_RESV: |
1870 | if (vma->vm_flags & VM_MAYSHARE) | |
1871 | ret = region_add(resv, idx, idx + 1); | |
1872 | else { | |
1873 | region_abort(resv, idx, idx + 1); | |
1874 | ret = region_del(resv, idx, idx + 1); | |
1875 | } | |
1876 | break; | |
5e911373 MK |
1877 | default: |
1878 | BUG(); | |
1879 | } | |
84afd99b | 1880 | |
4e35f483 | 1881 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1882 | return ret; |
67961f9d MK |
1883 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1884 | /* | |
1885 | * In most cases, reserves always exist for private mappings. | |
1886 | * However, a file associated with mapping could have been | |
1887 | * hole punched or truncated after reserves were consumed. | |
1888 | * As subsequent fault on such a range will not use reserves. | |
1889 | * Subtle - The reserve map for private mappings has the | |
1890 | * opposite meaning than that of shared mappings. If NO | |
1891 | * entry is in the reserve map, it means a reservation exists. | |
1892 | * If an entry exists in the reserve map, it means the | |
1893 | * reservation has already been consumed. As a result, the | |
1894 | * return value of this routine is the opposite of the | |
1895 | * value returned from reserve map manipulation routines above. | |
1896 | */ | |
1897 | if (ret) | |
1898 | return 0; | |
1899 | else | |
1900 | return 1; | |
1901 | } | |
4e35f483 | 1902 | else |
cf3ad20b | 1903 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1904 | } |
cf3ad20b MK |
1905 | |
1906 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1907 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1908 | { |
5e911373 | 1909 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1910 | } |
84afd99b | 1911 | |
cf3ad20b MK |
1912 | static long vma_commit_reservation(struct hstate *h, |
1913 | struct vm_area_struct *vma, unsigned long addr) | |
1914 | { | |
5e911373 MK |
1915 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1916 | } | |
1917 | ||
feba16e2 | 1918 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1919 | struct vm_area_struct *vma, unsigned long addr) |
1920 | { | |
feba16e2 | 1921 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1922 | } |
1923 | ||
96b96a96 MK |
1924 | static long vma_add_reservation(struct hstate *h, |
1925 | struct vm_area_struct *vma, unsigned long addr) | |
1926 | { | |
1927 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
1928 | } | |
1929 | ||
1930 | /* | |
1931 | * This routine is called to restore a reservation on error paths. In the | |
1932 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
1933 | * and is about to be freed. If a reservation for the page existed, | |
1934 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
1935 | * in the newly allocated page. When the page is freed via free_huge_page, | |
1936 | * the global reservation count will be incremented if PagePrivate is set. | |
1937 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
1938 | * reserve map here to be consistent with global reserve count adjustments | |
1939 | * to be made by free_huge_page. | |
1940 | */ | |
1941 | static void restore_reserve_on_error(struct hstate *h, | |
1942 | struct vm_area_struct *vma, unsigned long address, | |
1943 | struct page *page) | |
1944 | { | |
1945 | if (unlikely(PagePrivate(page))) { | |
1946 | long rc = vma_needs_reservation(h, vma, address); | |
1947 | ||
1948 | if (unlikely(rc < 0)) { | |
1949 | /* | |
1950 | * Rare out of memory condition in reserve map | |
1951 | * manipulation. Clear PagePrivate so that | |
1952 | * global reserve count will not be incremented | |
1953 | * by free_huge_page. This will make it appear | |
1954 | * as though the reservation for this page was | |
1955 | * consumed. This may prevent the task from | |
1956 | * faulting in the page at a later time. This | |
1957 | * is better than inconsistent global huge page | |
1958 | * accounting of reserve counts. | |
1959 | */ | |
1960 | ClearPagePrivate(page); | |
1961 | } else if (rc) { | |
1962 | rc = vma_add_reservation(h, vma, address); | |
1963 | if (unlikely(rc < 0)) | |
1964 | /* | |
1965 | * See above comment about rare out of | |
1966 | * memory condition. | |
1967 | */ | |
1968 | ClearPagePrivate(page); | |
1969 | } else | |
1970 | vma_end_reservation(h, vma, address); | |
1971 | } | |
1972 | } | |
1973 | ||
70c3547e | 1974 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1975 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1976 | { |
90481622 | 1977 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1978 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1979 | struct page *page; |
d85f69b0 MK |
1980 | long map_chg, map_commit; |
1981 | long gbl_chg; | |
6d76dcf4 AK |
1982 | int ret, idx; |
1983 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1984 | |
6d76dcf4 | 1985 | idx = hstate_index(h); |
a1e78772 | 1986 | /* |
d85f69b0 MK |
1987 | * Examine the region/reserve map to determine if the process |
1988 | * has a reservation for the page to be allocated. A return | |
1989 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 1990 | */ |
d85f69b0 MK |
1991 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
1992 | if (map_chg < 0) | |
76dcee75 | 1993 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
1994 | |
1995 | /* | |
1996 | * Processes that did not create the mapping will have no | |
1997 | * reserves as indicated by the region/reserve map. Check | |
1998 | * that the allocation will not exceed the subpool limit. | |
1999 | * Allocations for MAP_NORESERVE mappings also need to be | |
2000 | * checked against any subpool limit. | |
2001 | */ | |
2002 | if (map_chg || avoid_reserve) { | |
2003 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2004 | if (gbl_chg < 0) { | |
feba16e2 | 2005 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2006 | return ERR_PTR(-ENOSPC); |
5e911373 | 2007 | } |
1da177e4 | 2008 | |
d85f69b0 MK |
2009 | /* |
2010 | * Even though there was no reservation in the region/reserve | |
2011 | * map, there could be reservations associated with the | |
2012 | * subpool that can be used. This would be indicated if the | |
2013 | * return value of hugepage_subpool_get_pages() is zero. | |
2014 | * However, if avoid_reserve is specified we still avoid even | |
2015 | * the subpool reservations. | |
2016 | */ | |
2017 | if (avoid_reserve) | |
2018 | gbl_chg = 1; | |
2019 | } | |
2020 | ||
6d76dcf4 | 2021 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2022 | if (ret) |
2023 | goto out_subpool_put; | |
2024 | ||
1da177e4 | 2025 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2026 | /* |
2027 | * glb_chg is passed to indicate whether or not a page must be taken | |
2028 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2029 | * a reservation exists for the allocation. | |
2030 | */ | |
2031 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2032 | if (!page) { |
94ae8ba7 | 2033 | spin_unlock(&hugetlb_lock); |
099730d6 | 2034 | page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2035 | if (!page) |
2036 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2037 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2038 | SetPagePrivate(page); | |
2039 | h->resv_huge_pages--; | |
2040 | } | |
79dbb236 AK |
2041 | spin_lock(&hugetlb_lock); |
2042 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2043 | /* Fall through */ |
68842c9b | 2044 | } |
81a6fcae JK |
2045 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2046 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2047 | |
90481622 | 2048 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2049 | |
d85f69b0 MK |
2050 | map_commit = vma_commit_reservation(h, vma, addr); |
2051 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2052 | /* |
2053 | * The page was added to the reservation map between | |
2054 | * vma_needs_reservation and vma_commit_reservation. | |
2055 | * This indicates a race with hugetlb_reserve_pages. | |
2056 | * Adjust for the subpool count incremented above AND | |
2057 | * in hugetlb_reserve_pages for the same page. Also, | |
2058 | * the reservation count added in hugetlb_reserve_pages | |
2059 | * no longer applies. | |
2060 | */ | |
2061 | long rsv_adjust; | |
2062 | ||
2063 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2064 | hugetlb_acct_memory(h, -rsv_adjust); | |
2065 | } | |
90d8b7e6 | 2066 | return page; |
8f34af6f JZ |
2067 | |
2068 | out_uncharge_cgroup: | |
2069 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2070 | out_subpool_put: | |
d85f69b0 | 2071 | if (map_chg || avoid_reserve) |
8f34af6f | 2072 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2073 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2074 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2075 | } |
2076 | ||
74060e4d NH |
2077 | /* |
2078 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
2079 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
2080 | * where no ERR_VALUE is expected to be returned. | |
2081 | */ | |
2082 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
2083 | unsigned long addr, int avoid_reserve) | |
2084 | { | |
2085 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
2086 | if (IS_ERR(page)) | |
2087 | page = NULL; | |
2088 | return page; | |
2089 | } | |
2090 | ||
e24a1307 AK |
2091 | int alloc_bootmem_huge_page(struct hstate *h) |
2092 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2093 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2094 | { |
2095 | struct huge_bootmem_page *m; | |
b2261026 | 2096 | int nr_nodes, node; |
aa888a74 | 2097 | |
b2261026 | 2098 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2099 | void *addr; |
2100 | ||
8b89a116 GS |
2101 | addr = memblock_virt_alloc_try_nid_nopanic( |
2102 | huge_page_size(h), huge_page_size(h), | |
2103 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
2104 | if (addr) { |
2105 | /* | |
2106 | * Use the beginning of the huge page to store the | |
2107 | * huge_bootmem_page struct (until gather_bootmem | |
2108 | * puts them into the mem_map). | |
2109 | */ | |
2110 | m = addr; | |
91f47662 | 2111 | goto found; |
aa888a74 | 2112 | } |
aa888a74 AK |
2113 | } |
2114 | return 0; | |
2115 | ||
2116 | found: | |
df994ead | 2117 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
2118 | /* Put them into a private list first because mem_map is not up yet */ |
2119 | list_add(&m->list, &huge_boot_pages); | |
2120 | m->hstate = h; | |
2121 | return 1; | |
2122 | } | |
2123 | ||
d00181b9 KS |
2124 | static void __init prep_compound_huge_page(struct page *page, |
2125 | unsigned int order) | |
18229df5 AW |
2126 | { |
2127 | if (unlikely(order > (MAX_ORDER - 1))) | |
2128 | prep_compound_gigantic_page(page, order); | |
2129 | else | |
2130 | prep_compound_page(page, order); | |
2131 | } | |
2132 | ||
aa888a74 AK |
2133 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2134 | static void __init gather_bootmem_prealloc(void) | |
2135 | { | |
2136 | struct huge_bootmem_page *m; | |
2137 | ||
2138 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 2139 | struct hstate *h = m->hstate; |
ee8f248d BB |
2140 | struct page *page; |
2141 | ||
2142 | #ifdef CONFIG_HIGHMEM | |
2143 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
2144 | memblock_free_late(__pa(m), |
2145 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
2146 | #else |
2147 | page = virt_to_page(m); | |
2148 | #endif | |
aa888a74 | 2149 | WARN_ON(page_count(page) != 1); |
18229df5 | 2150 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2151 | WARN_ON(PageReserved(page)); |
aa888a74 | 2152 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
2153 | /* |
2154 | * If we had gigantic hugepages allocated at boot time, we need | |
2155 | * to restore the 'stolen' pages to totalram_pages in order to | |
2156 | * fix confusing memory reports from free(1) and another | |
2157 | * side-effects, like CommitLimit going negative. | |
2158 | */ | |
bae7f4ae | 2159 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2160 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
2161 | } |
2162 | } | |
2163 | ||
8faa8b07 | 2164 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2165 | { |
2166 | unsigned long i; | |
a5516438 | 2167 | |
e5ff2159 | 2168 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2169 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2170 | if (!alloc_bootmem_huge_page(h)) |
2171 | break; | |
9b5e5d0f | 2172 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 2173 | &node_states[N_MEMORY])) |
1da177e4 | 2174 | break; |
69ed779a | 2175 | cond_resched(); |
1da177e4 | 2176 | } |
d715cf80 LH |
2177 | if (i < h->max_huge_pages) { |
2178 | char buf[32]; | |
2179 | ||
c6247f72 | 2180 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2181 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2182 | h->max_huge_pages, buf, i); | |
2183 | h->max_huge_pages = i; | |
2184 | } | |
e5ff2159 AK |
2185 | } |
2186 | ||
2187 | static void __init hugetlb_init_hstates(void) | |
2188 | { | |
2189 | struct hstate *h; | |
2190 | ||
2191 | for_each_hstate(h) { | |
641844f5 NH |
2192 | if (minimum_order > huge_page_order(h)) |
2193 | minimum_order = huge_page_order(h); | |
2194 | ||
8faa8b07 | 2195 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2196 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2197 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2198 | } |
641844f5 | 2199 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2200 | } |
2201 | ||
2202 | static void __init report_hugepages(void) | |
2203 | { | |
2204 | struct hstate *h; | |
2205 | ||
2206 | for_each_hstate(h) { | |
4abd32db | 2207 | char buf[32]; |
c6247f72 MW |
2208 | |
2209 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2210 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2211 | buf, h->free_huge_pages); |
e5ff2159 AK |
2212 | } |
2213 | } | |
2214 | ||
1da177e4 | 2215 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2216 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2217 | nodemask_t *nodes_allowed) | |
1da177e4 | 2218 | { |
4415cc8d CL |
2219 | int i; |
2220 | ||
bae7f4ae | 2221 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2222 | return; |
2223 | ||
6ae11b27 | 2224 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2225 | struct page *page, *next; |
a5516438 AK |
2226 | struct list_head *freel = &h->hugepage_freelists[i]; |
2227 | list_for_each_entry_safe(page, next, freel, lru) { | |
2228 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2229 | return; |
1da177e4 LT |
2230 | if (PageHighMem(page)) |
2231 | continue; | |
2232 | list_del(&page->lru); | |
e5ff2159 | 2233 | update_and_free_page(h, page); |
a5516438 AK |
2234 | h->free_huge_pages--; |
2235 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2236 | } |
2237 | } | |
2238 | } | |
2239 | #else | |
6ae11b27 LS |
2240 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2241 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2242 | { |
2243 | } | |
2244 | #endif | |
2245 | ||
20a0307c WF |
2246 | /* |
2247 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2248 | * balanced by operating on them in a round-robin fashion. | |
2249 | * Returns 1 if an adjustment was made. | |
2250 | */ | |
6ae11b27 LS |
2251 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2252 | int delta) | |
20a0307c | 2253 | { |
b2261026 | 2254 | int nr_nodes, node; |
20a0307c WF |
2255 | |
2256 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2257 | |
b2261026 JK |
2258 | if (delta < 0) { |
2259 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2260 | if (h->surplus_huge_pages_node[node]) | |
2261 | goto found; | |
e8c5c824 | 2262 | } |
b2261026 JK |
2263 | } else { |
2264 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2265 | if (h->surplus_huge_pages_node[node] < | |
2266 | h->nr_huge_pages_node[node]) | |
2267 | goto found; | |
e8c5c824 | 2268 | } |
b2261026 JK |
2269 | } |
2270 | return 0; | |
20a0307c | 2271 | |
b2261026 JK |
2272 | found: |
2273 | h->surplus_huge_pages += delta; | |
2274 | h->surplus_huge_pages_node[node] += delta; | |
2275 | return 1; | |
20a0307c WF |
2276 | } |
2277 | ||
a5516438 | 2278 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2279 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2280 | nodemask_t *nodes_allowed) | |
1da177e4 | 2281 | { |
7893d1d5 | 2282 | unsigned long min_count, ret; |
1da177e4 | 2283 | |
944d9fec | 2284 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2285 | return h->max_huge_pages; |
2286 | ||
7893d1d5 AL |
2287 | /* |
2288 | * Increase the pool size | |
2289 | * First take pages out of surplus state. Then make up the | |
2290 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2291 | * |
d15c7c09 | 2292 | * We might race with __alloc_buddy_huge_page() here and be unable |
d1c3fb1f NA |
2293 | * to convert a surplus huge page to a normal huge page. That is |
2294 | * not critical, though, it just means the overall size of the | |
2295 | * pool might be one hugepage larger than it needs to be, but | |
2296 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2297 | */ |
1da177e4 | 2298 | spin_lock(&hugetlb_lock); |
a5516438 | 2299 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2300 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2301 | break; |
2302 | } | |
2303 | ||
a5516438 | 2304 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2305 | /* |
2306 | * If this allocation races such that we no longer need the | |
2307 | * page, free_huge_page will handle it by freeing the page | |
2308 | * and reducing the surplus. | |
2309 | */ | |
2310 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2311 | |
2312 | /* yield cpu to avoid soft lockup */ | |
2313 | cond_resched(); | |
2314 | ||
944d9fec LC |
2315 | if (hstate_is_gigantic(h)) |
2316 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
2317 | else | |
2318 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
2319 | spin_lock(&hugetlb_lock); |
2320 | if (!ret) | |
2321 | goto out; | |
2322 | ||
536240f2 MG |
2323 | /* Bail for signals. Probably ctrl-c from user */ |
2324 | if (signal_pending(current)) | |
2325 | goto out; | |
7893d1d5 | 2326 | } |
7893d1d5 AL |
2327 | |
2328 | /* | |
2329 | * Decrease the pool size | |
2330 | * First return free pages to the buddy allocator (being careful | |
2331 | * to keep enough around to satisfy reservations). Then place | |
2332 | * pages into surplus state as needed so the pool will shrink | |
2333 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2334 | * |
2335 | * By placing pages into the surplus state independent of the | |
2336 | * overcommit value, we are allowing the surplus pool size to | |
2337 | * exceed overcommit. There are few sane options here. Since | |
d15c7c09 | 2338 | * __alloc_buddy_huge_page() is checking the global counter, |
d1c3fb1f NA |
2339 | * though, we'll note that we're not allowed to exceed surplus |
2340 | * and won't grow the pool anywhere else. Not until one of the | |
2341 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2342 | */ |
a5516438 | 2343 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2344 | min_count = max(count, min_count); |
6ae11b27 | 2345 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2346 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2347 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2348 | break; |
55f67141 | 2349 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2350 | } |
a5516438 | 2351 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2352 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2353 | break; |
2354 | } | |
2355 | out: | |
a5516438 | 2356 | ret = persistent_huge_pages(h); |
1da177e4 | 2357 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2358 | return ret; |
1da177e4 LT |
2359 | } |
2360 | ||
a3437870 NA |
2361 | #define HSTATE_ATTR_RO(_name) \ |
2362 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2363 | ||
2364 | #define HSTATE_ATTR(_name) \ | |
2365 | static struct kobj_attribute _name##_attr = \ | |
2366 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2367 | ||
2368 | static struct kobject *hugepages_kobj; | |
2369 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2370 | ||
9a305230 LS |
2371 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2372 | ||
2373 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2374 | { |
2375 | int i; | |
9a305230 | 2376 | |
a3437870 | 2377 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2378 | if (hstate_kobjs[i] == kobj) { |
2379 | if (nidp) | |
2380 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2381 | return &hstates[i]; |
9a305230 LS |
2382 | } |
2383 | ||
2384 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2385 | } |
2386 | ||
06808b08 | 2387 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2388 | struct kobj_attribute *attr, char *buf) |
2389 | { | |
9a305230 LS |
2390 | struct hstate *h; |
2391 | unsigned long nr_huge_pages; | |
2392 | int nid; | |
2393 | ||
2394 | h = kobj_to_hstate(kobj, &nid); | |
2395 | if (nid == NUMA_NO_NODE) | |
2396 | nr_huge_pages = h->nr_huge_pages; | |
2397 | else | |
2398 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2399 | ||
2400 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2401 | } |
adbe8726 | 2402 | |
238d3c13 DR |
2403 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2404 | struct hstate *h, int nid, | |
2405 | unsigned long count, size_t len) | |
a3437870 NA |
2406 | { |
2407 | int err; | |
bad44b5b | 2408 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2409 | |
944d9fec | 2410 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2411 | err = -EINVAL; |
2412 | goto out; | |
2413 | } | |
2414 | ||
9a305230 LS |
2415 | if (nid == NUMA_NO_NODE) { |
2416 | /* | |
2417 | * global hstate attribute | |
2418 | */ | |
2419 | if (!(obey_mempolicy && | |
2420 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2421 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2422 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2423 | } |
2424 | } else if (nodes_allowed) { | |
2425 | /* | |
2426 | * per node hstate attribute: adjust count to global, | |
2427 | * but restrict alloc/free to the specified node. | |
2428 | */ | |
2429 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2430 | init_nodemask_of_node(nodes_allowed, nid); | |
2431 | } else | |
8cebfcd0 | 2432 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2433 | |
06808b08 | 2434 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2435 | |
8cebfcd0 | 2436 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2437 | NODEMASK_FREE(nodes_allowed); |
2438 | ||
2439 | return len; | |
adbe8726 EM |
2440 | out: |
2441 | NODEMASK_FREE(nodes_allowed); | |
2442 | return err; | |
06808b08 LS |
2443 | } |
2444 | ||
238d3c13 DR |
2445 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2446 | struct kobject *kobj, const char *buf, | |
2447 | size_t len) | |
2448 | { | |
2449 | struct hstate *h; | |
2450 | unsigned long count; | |
2451 | int nid; | |
2452 | int err; | |
2453 | ||
2454 | err = kstrtoul(buf, 10, &count); | |
2455 | if (err) | |
2456 | return err; | |
2457 | ||
2458 | h = kobj_to_hstate(kobj, &nid); | |
2459 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2460 | } | |
2461 | ||
06808b08 LS |
2462 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2463 | struct kobj_attribute *attr, char *buf) | |
2464 | { | |
2465 | return nr_hugepages_show_common(kobj, attr, buf); | |
2466 | } | |
2467 | ||
2468 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2469 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2470 | { | |
238d3c13 | 2471 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2472 | } |
2473 | HSTATE_ATTR(nr_hugepages); | |
2474 | ||
06808b08 LS |
2475 | #ifdef CONFIG_NUMA |
2476 | ||
2477 | /* | |
2478 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2479 | * huge page alloc/free. | |
2480 | */ | |
2481 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2482 | struct kobj_attribute *attr, char *buf) | |
2483 | { | |
2484 | return nr_hugepages_show_common(kobj, attr, buf); | |
2485 | } | |
2486 | ||
2487 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2488 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2489 | { | |
238d3c13 | 2490 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2491 | } |
2492 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2493 | #endif | |
2494 | ||
2495 | ||
a3437870 NA |
2496 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2497 | struct kobj_attribute *attr, char *buf) | |
2498 | { | |
9a305230 | 2499 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2500 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2501 | } | |
adbe8726 | 2502 | |
a3437870 NA |
2503 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2504 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2505 | { | |
2506 | int err; | |
2507 | unsigned long input; | |
9a305230 | 2508 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2509 | |
bae7f4ae | 2510 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2511 | return -EINVAL; |
2512 | ||
3dbb95f7 | 2513 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2514 | if (err) |
73ae31e5 | 2515 | return err; |
a3437870 NA |
2516 | |
2517 | spin_lock(&hugetlb_lock); | |
2518 | h->nr_overcommit_huge_pages = input; | |
2519 | spin_unlock(&hugetlb_lock); | |
2520 | ||
2521 | return count; | |
2522 | } | |
2523 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2524 | ||
2525 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2526 | struct kobj_attribute *attr, char *buf) | |
2527 | { | |
9a305230 LS |
2528 | struct hstate *h; |
2529 | unsigned long free_huge_pages; | |
2530 | int nid; | |
2531 | ||
2532 | h = kobj_to_hstate(kobj, &nid); | |
2533 | if (nid == NUMA_NO_NODE) | |
2534 | free_huge_pages = h->free_huge_pages; | |
2535 | else | |
2536 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2537 | ||
2538 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2539 | } |
2540 | HSTATE_ATTR_RO(free_hugepages); | |
2541 | ||
2542 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2543 | struct kobj_attribute *attr, char *buf) | |
2544 | { | |
9a305230 | 2545 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2546 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2547 | } | |
2548 | HSTATE_ATTR_RO(resv_hugepages); | |
2549 | ||
2550 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2551 | struct kobj_attribute *attr, char *buf) | |
2552 | { | |
9a305230 LS |
2553 | struct hstate *h; |
2554 | unsigned long surplus_huge_pages; | |
2555 | int nid; | |
2556 | ||
2557 | h = kobj_to_hstate(kobj, &nid); | |
2558 | if (nid == NUMA_NO_NODE) | |
2559 | surplus_huge_pages = h->surplus_huge_pages; | |
2560 | else | |
2561 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2562 | ||
2563 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2564 | } |
2565 | HSTATE_ATTR_RO(surplus_hugepages); | |
2566 | ||
2567 | static struct attribute *hstate_attrs[] = { | |
2568 | &nr_hugepages_attr.attr, | |
2569 | &nr_overcommit_hugepages_attr.attr, | |
2570 | &free_hugepages_attr.attr, | |
2571 | &resv_hugepages_attr.attr, | |
2572 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2573 | #ifdef CONFIG_NUMA |
2574 | &nr_hugepages_mempolicy_attr.attr, | |
2575 | #endif | |
a3437870 NA |
2576 | NULL, |
2577 | }; | |
2578 | ||
67e5ed96 | 2579 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2580 | .attrs = hstate_attrs, |
2581 | }; | |
2582 | ||
094e9539 JM |
2583 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2584 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2585 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2586 | { |
2587 | int retval; | |
972dc4de | 2588 | int hi = hstate_index(h); |
a3437870 | 2589 | |
9a305230 LS |
2590 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2591 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2592 | return -ENOMEM; |
2593 | ||
9a305230 | 2594 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2595 | if (retval) |
9a305230 | 2596 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2597 | |
2598 | return retval; | |
2599 | } | |
2600 | ||
2601 | static void __init hugetlb_sysfs_init(void) | |
2602 | { | |
2603 | struct hstate *h; | |
2604 | int err; | |
2605 | ||
2606 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2607 | if (!hugepages_kobj) | |
2608 | return; | |
2609 | ||
2610 | for_each_hstate(h) { | |
9a305230 LS |
2611 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2612 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2613 | if (err) |
ffb22af5 | 2614 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2615 | } |
2616 | } | |
2617 | ||
9a305230 LS |
2618 | #ifdef CONFIG_NUMA |
2619 | ||
2620 | /* | |
2621 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2622 | * with node devices in node_devices[] using a parallel array. The array |
2623 | * index of a node device or _hstate == node id. | |
2624 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2625 | * the base kernel, on the hugetlb module. |
2626 | */ | |
2627 | struct node_hstate { | |
2628 | struct kobject *hugepages_kobj; | |
2629 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2630 | }; | |
b4e289a6 | 2631 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2632 | |
2633 | /* | |
10fbcf4c | 2634 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2635 | */ |
2636 | static struct attribute *per_node_hstate_attrs[] = { | |
2637 | &nr_hugepages_attr.attr, | |
2638 | &free_hugepages_attr.attr, | |
2639 | &surplus_hugepages_attr.attr, | |
2640 | NULL, | |
2641 | }; | |
2642 | ||
67e5ed96 | 2643 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
2644 | .attrs = per_node_hstate_attrs, |
2645 | }; | |
2646 | ||
2647 | /* | |
10fbcf4c | 2648 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2649 | * Returns node id via non-NULL nidp. |
2650 | */ | |
2651 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2652 | { | |
2653 | int nid; | |
2654 | ||
2655 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2656 | struct node_hstate *nhs = &node_hstates[nid]; | |
2657 | int i; | |
2658 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2659 | if (nhs->hstate_kobjs[i] == kobj) { | |
2660 | if (nidp) | |
2661 | *nidp = nid; | |
2662 | return &hstates[i]; | |
2663 | } | |
2664 | } | |
2665 | ||
2666 | BUG(); | |
2667 | return NULL; | |
2668 | } | |
2669 | ||
2670 | /* | |
10fbcf4c | 2671 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2672 | * No-op if no hstate attributes attached. |
2673 | */ | |
3cd8b44f | 2674 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2675 | { |
2676 | struct hstate *h; | |
10fbcf4c | 2677 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2678 | |
2679 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2680 | return; /* no hstate attributes */ |
9a305230 | 2681 | |
972dc4de AK |
2682 | for_each_hstate(h) { |
2683 | int idx = hstate_index(h); | |
2684 | if (nhs->hstate_kobjs[idx]) { | |
2685 | kobject_put(nhs->hstate_kobjs[idx]); | |
2686 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2687 | } |
972dc4de | 2688 | } |
9a305230 LS |
2689 | |
2690 | kobject_put(nhs->hugepages_kobj); | |
2691 | nhs->hugepages_kobj = NULL; | |
2692 | } | |
2693 | ||
9a305230 LS |
2694 | |
2695 | /* | |
10fbcf4c | 2696 | * Register hstate attributes for a single node device. |
9a305230 LS |
2697 | * No-op if attributes already registered. |
2698 | */ | |
3cd8b44f | 2699 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2700 | { |
2701 | struct hstate *h; | |
10fbcf4c | 2702 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2703 | int err; |
2704 | ||
2705 | if (nhs->hugepages_kobj) | |
2706 | return; /* already allocated */ | |
2707 | ||
2708 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2709 | &node->dev.kobj); |
9a305230 LS |
2710 | if (!nhs->hugepages_kobj) |
2711 | return; | |
2712 | ||
2713 | for_each_hstate(h) { | |
2714 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2715 | nhs->hstate_kobjs, | |
2716 | &per_node_hstate_attr_group); | |
2717 | if (err) { | |
ffb22af5 AM |
2718 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2719 | h->name, node->dev.id); | |
9a305230 LS |
2720 | hugetlb_unregister_node(node); |
2721 | break; | |
2722 | } | |
2723 | } | |
2724 | } | |
2725 | ||
2726 | /* | |
9b5e5d0f | 2727 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2728 | * devices of nodes that have memory. All on-line nodes should have |
2729 | * registered their associated device by this time. | |
9a305230 | 2730 | */ |
7d9ca000 | 2731 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2732 | { |
2733 | int nid; | |
2734 | ||
8cebfcd0 | 2735 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2736 | struct node *node = node_devices[nid]; |
10fbcf4c | 2737 | if (node->dev.id == nid) |
9a305230 LS |
2738 | hugetlb_register_node(node); |
2739 | } | |
2740 | ||
2741 | /* | |
10fbcf4c | 2742 | * Let the node device driver know we're here so it can |
9a305230 LS |
2743 | * [un]register hstate attributes on node hotplug. |
2744 | */ | |
2745 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2746 | hugetlb_unregister_node); | |
2747 | } | |
2748 | #else /* !CONFIG_NUMA */ | |
2749 | ||
2750 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2751 | { | |
2752 | BUG(); | |
2753 | if (nidp) | |
2754 | *nidp = -1; | |
2755 | return NULL; | |
2756 | } | |
2757 | ||
9a305230 LS |
2758 | static void hugetlb_register_all_nodes(void) { } |
2759 | ||
2760 | #endif | |
2761 | ||
a3437870 NA |
2762 | static int __init hugetlb_init(void) |
2763 | { | |
8382d914 DB |
2764 | int i; |
2765 | ||
457c1b27 | 2766 | if (!hugepages_supported()) |
0ef89d25 | 2767 | return 0; |
a3437870 | 2768 | |
e11bfbfc | 2769 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
2770 | if (default_hstate_size != 0) { |
2771 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
2772 | default_hstate_size, HPAGE_SIZE); | |
2773 | } | |
2774 | ||
e11bfbfc NP |
2775 | default_hstate_size = HPAGE_SIZE; |
2776 | if (!size_to_hstate(default_hstate_size)) | |
2777 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2778 | } |
972dc4de | 2779 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2780 | if (default_hstate_max_huge_pages) { |
2781 | if (!default_hstate.max_huge_pages) | |
2782 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2783 | } | |
a3437870 NA |
2784 | |
2785 | hugetlb_init_hstates(); | |
aa888a74 | 2786 | gather_bootmem_prealloc(); |
a3437870 NA |
2787 | report_hugepages(); |
2788 | ||
2789 | hugetlb_sysfs_init(); | |
9a305230 | 2790 | hugetlb_register_all_nodes(); |
7179e7bf | 2791 | hugetlb_cgroup_file_init(); |
9a305230 | 2792 | |
8382d914 DB |
2793 | #ifdef CONFIG_SMP |
2794 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2795 | #else | |
2796 | num_fault_mutexes = 1; | |
2797 | #endif | |
c672c7f2 | 2798 | hugetlb_fault_mutex_table = |
8382d914 | 2799 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2800 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2801 | |
2802 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2803 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2804 | return 0; |
2805 | } | |
3e89e1c5 | 2806 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2807 | |
2808 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2809 | void __init hugetlb_bad_size(void) |
2810 | { | |
2811 | parsed_valid_hugepagesz = false; | |
2812 | } | |
2813 | ||
d00181b9 | 2814 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2815 | { |
2816 | struct hstate *h; | |
8faa8b07 AK |
2817 | unsigned long i; |
2818 | ||
a3437870 | 2819 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2820 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2821 | return; |
2822 | } | |
47d38344 | 2823 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2824 | BUG_ON(order == 0); |
47d38344 | 2825 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2826 | h->order = order; |
2827 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2828 | h->nr_huge_pages = 0; |
2829 | h->free_huge_pages = 0; | |
2830 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2831 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2832 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2833 | h->next_nid_to_alloc = first_memory_node; |
2834 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2835 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2836 | huge_page_size(h)/1024); | |
8faa8b07 | 2837 | |
a3437870 NA |
2838 | parsed_hstate = h; |
2839 | } | |
2840 | ||
e11bfbfc | 2841 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2842 | { |
2843 | unsigned long *mhp; | |
8faa8b07 | 2844 | static unsigned long *last_mhp; |
a3437870 | 2845 | |
9fee021d VT |
2846 | if (!parsed_valid_hugepagesz) { |
2847 | pr_warn("hugepages = %s preceded by " | |
2848 | "an unsupported hugepagesz, ignoring\n", s); | |
2849 | parsed_valid_hugepagesz = true; | |
2850 | return 1; | |
2851 | } | |
a3437870 | 2852 | /* |
47d38344 | 2853 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2854 | * so this hugepages= parameter goes to the "default hstate". |
2855 | */ | |
9fee021d | 2856 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2857 | mhp = &default_hstate_max_huge_pages; |
2858 | else | |
2859 | mhp = &parsed_hstate->max_huge_pages; | |
2860 | ||
8faa8b07 | 2861 | if (mhp == last_mhp) { |
598d8091 | 2862 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2863 | return 1; |
2864 | } | |
2865 | ||
a3437870 NA |
2866 | if (sscanf(s, "%lu", mhp) <= 0) |
2867 | *mhp = 0; | |
2868 | ||
8faa8b07 AK |
2869 | /* |
2870 | * Global state is always initialized later in hugetlb_init. | |
2871 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2872 | * use the bootmem allocator. | |
2873 | */ | |
47d38344 | 2874 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2875 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2876 | ||
2877 | last_mhp = mhp; | |
2878 | ||
a3437870 NA |
2879 | return 1; |
2880 | } | |
e11bfbfc NP |
2881 | __setup("hugepages=", hugetlb_nrpages_setup); |
2882 | ||
2883 | static int __init hugetlb_default_setup(char *s) | |
2884 | { | |
2885 | default_hstate_size = memparse(s, &s); | |
2886 | return 1; | |
2887 | } | |
2888 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2889 | |
8a213460 NA |
2890 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2891 | { | |
2892 | int node; | |
2893 | unsigned int nr = 0; | |
2894 | ||
2895 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2896 | nr += array[node]; | |
2897 | ||
2898 | return nr; | |
2899 | } | |
2900 | ||
2901 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2902 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2903 | struct ctl_table *table, int write, | |
2904 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2905 | { |
e5ff2159 | 2906 | struct hstate *h = &default_hstate; |
238d3c13 | 2907 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2908 | int ret; |
e5ff2159 | 2909 | |
457c1b27 | 2910 | if (!hugepages_supported()) |
86613628 | 2911 | return -EOPNOTSUPP; |
457c1b27 | 2912 | |
e5ff2159 AK |
2913 | table->data = &tmp; |
2914 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2915 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2916 | if (ret) | |
2917 | goto out; | |
e5ff2159 | 2918 | |
238d3c13 DR |
2919 | if (write) |
2920 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2921 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2922 | out: |
2923 | return ret; | |
1da177e4 | 2924 | } |
396faf03 | 2925 | |
06808b08 LS |
2926 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2927 | void __user *buffer, size_t *length, loff_t *ppos) | |
2928 | { | |
2929 | ||
2930 | return hugetlb_sysctl_handler_common(false, table, write, | |
2931 | buffer, length, ppos); | |
2932 | } | |
2933 | ||
2934 | #ifdef CONFIG_NUMA | |
2935 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2936 | void __user *buffer, size_t *length, loff_t *ppos) | |
2937 | { | |
2938 | return hugetlb_sysctl_handler_common(true, table, write, | |
2939 | buffer, length, ppos); | |
2940 | } | |
2941 | #endif /* CONFIG_NUMA */ | |
2942 | ||
a3d0c6aa | 2943 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2944 | void __user *buffer, |
a3d0c6aa NA |
2945 | size_t *length, loff_t *ppos) |
2946 | { | |
a5516438 | 2947 | struct hstate *h = &default_hstate; |
e5ff2159 | 2948 | unsigned long tmp; |
08d4a246 | 2949 | int ret; |
e5ff2159 | 2950 | |
457c1b27 | 2951 | if (!hugepages_supported()) |
86613628 | 2952 | return -EOPNOTSUPP; |
457c1b27 | 2953 | |
c033a93c | 2954 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2955 | |
bae7f4ae | 2956 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2957 | return -EINVAL; |
2958 | ||
e5ff2159 AK |
2959 | table->data = &tmp; |
2960 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2961 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2962 | if (ret) | |
2963 | goto out; | |
e5ff2159 AK |
2964 | |
2965 | if (write) { | |
2966 | spin_lock(&hugetlb_lock); | |
2967 | h->nr_overcommit_huge_pages = tmp; | |
2968 | spin_unlock(&hugetlb_lock); | |
2969 | } | |
08d4a246 MH |
2970 | out: |
2971 | return ret; | |
a3d0c6aa NA |
2972 | } |
2973 | ||
1da177e4 LT |
2974 | #endif /* CONFIG_SYSCTL */ |
2975 | ||
e1759c21 | 2976 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2977 | { |
a5516438 | 2978 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2979 | if (!hugepages_supported()) |
2980 | return; | |
e1759c21 | 2981 | seq_printf(m, |
4f98a2fe RR |
2982 | "HugePages_Total: %5lu\n" |
2983 | "HugePages_Free: %5lu\n" | |
2984 | "HugePages_Rsvd: %5lu\n" | |
2985 | "HugePages_Surp: %5lu\n" | |
2986 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2987 | h->nr_huge_pages, |
2988 | h->free_huge_pages, | |
2989 | h->resv_huge_pages, | |
2990 | h->surplus_huge_pages, | |
2991 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2992 | } |
2993 | ||
2994 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2995 | { | |
a5516438 | 2996 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2997 | if (!hugepages_supported()) |
2998 | return 0; | |
1da177e4 LT |
2999 | return sprintf(buf, |
3000 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3001 | "Node %d HugePages_Free: %5u\n" |
3002 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3003 | nid, h->nr_huge_pages_node[nid], |
3004 | nid, h->free_huge_pages_node[nid], | |
3005 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3006 | } |
3007 | ||
949f7ec5 DR |
3008 | void hugetlb_show_meminfo(void) |
3009 | { | |
3010 | struct hstate *h; | |
3011 | int nid; | |
3012 | ||
457c1b27 NA |
3013 | if (!hugepages_supported()) |
3014 | return; | |
3015 | ||
949f7ec5 DR |
3016 | for_each_node_state(nid, N_MEMORY) |
3017 | for_each_hstate(h) | |
3018 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3019 | nid, | |
3020 | h->nr_huge_pages_node[nid], | |
3021 | h->free_huge_pages_node[nid], | |
3022 | h->surplus_huge_pages_node[nid], | |
3023 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3024 | } | |
3025 | ||
5d317b2b NH |
3026 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3027 | { | |
3028 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3029 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3030 | } | |
3031 | ||
1da177e4 LT |
3032 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3033 | unsigned long hugetlb_total_pages(void) | |
3034 | { | |
d0028588 WL |
3035 | struct hstate *h; |
3036 | unsigned long nr_total_pages = 0; | |
3037 | ||
3038 | for_each_hstate(h) | |
3039 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3040 | return nr_total_pages; | |
1da177e4 | 3041 | } |
1da177e4 | 3042 | |
a5516438 | 3043 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3044 | { |
3045 | int ret = -ENOMEM; | |
3046 | ||
3047 | spin_lock(&hugetlb_lock); | |
3048 | /* | |
3049 | * When cpuset is configured, it breaks the strict hugetlb page | |
3050 | * reservation as the accounting is done on a global variable. Such | |
3051 | * reservation is completely rubbish in the presence of cpuset because | |
3052 | * the reservation is not checked against page availability for the | |
3053 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3054 | * with lack of free htlb page in cpuset that the task is in. | |
3055 | * Attempt to enforce strict accounting with cpuset is almost | |
3056 | * impossible (or too ugly) because cpuset is too fluid that | |
3057 | * task or memory node can be dynamically moved between cpusets. | |
3058 | * | |
3059 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3060 | * undesirable. However, in order to preserve some of the semantics, | |
3061 | * we fall back to check against current free page availability as | |
3062 | * a best attempt and hopefully to minimize the impact of changing | |
3063 | * semantics that cpuset has. | |
3064 | */ | |
3065 | if (delta > 0) { | |
a5516438 | 3066 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3067 | goto out; |
3068 | ||
a5516438 AK |
3069 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3070 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3071 | goto out; |
3072 | } | |
3073 | } | |
3074 | ||
3075 | ret = 0; | |
3076 | if (delta < 0) | |
a5516438 | 3077 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3078 | |
3079 | out: | |
3080 | spin_unlock(&hugetlb_lock); | |
3081 | return ret; | |
3082 | } | |
3083 | ||
84afd99b AW |
3084 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3085 | { | |
f522c3ac | 3086 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3087 | |
3088 | /* | |
3089 | * This new VMA should share its siblings reservation map if present. | |
3090 | * The VMA will only ever have a valid reservation map pointer where | |
3091 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3092 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3093 | * after this open call completes. It is therefore safe to take a |
3094 | * new reference here without additional locking. | |
3095 | */ | |
4e35f483 | 3096 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3097 | kref_get(&resv->refs); |
84afd99b AW |
3098 | } |
3099 | ||
a1e78772 MG |
3100 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3101 | { | |
a5516438 | 3102 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3103 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3104 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3105 | unsigned long reserve, start, end; |
1c5ecae3 | 3106 | long gbl_reserve; |
84afd99b | 3107 | |
4e35f483 JK |
3108 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3109 | return; | |
84afd99b | 3110 | |
4e35f483 JK |
3111 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3112 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3113 | |
4e35f483 | 3114 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3115 | |
4e35f483 JK |
3116 | kref_put(&resv->refs, resv_map_release); |
3117 | ||
3118 | if (reserve) { | |
1c5ecae3 MK |
3119 | /* |
3120 | * Decrement reserve counts. The global reserve count may be | |
3121 | * adjusted if the subpool has a minimum size. | |
3122 | */ | |
3123 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3124 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3125 | } |
a1e78772 MG |
3126 | } |
3127 | ||
1da177e4 LT |
3128 | /* |
3129 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3130 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3131 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3132 | * this far. | |
3133 | */ | |
11bac800 | 3134 | static int hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3135 | { |
3136 | BUG(); | |
d0217ac0 | 3137 | return 0; |
1da177e4 LT |
3138 | } |
3139 | ||
f0f37e2f | 3140 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3141 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3142 | .open = hugetlb_vm_op_open, |
a1e78772 | 3143 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
3144 | }; |
3145 | ||
1e8f889b DG |
3146 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3147 | int writable) | |
63551ae0 DG |
3148 | { |
3149 | pte_t entry; | |
3150 | ||
1e8f889b | 3151 | if (writable) { |
106c992a GS |
3152 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3153 | vma->vm_page_prot))); | |
63551ae0 | 3154 | } else { |
106c992a GS |
3155 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3156 | vma->vm_page_prot)); | |
63551ae0 DG |
3157 | } |
3158 | entry = pte_mkyoung(entry); | |
3159 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3160 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3161 | |
3162 | return entry; | |
3163 | } | |
3164 | ||
1e8f889b DG |
3165 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3166 | unsigned long address, pte_t *ptep) | |
3167 | { | |
3168 | pte_t entry; | |
3169 | ||
106c992a | 3170 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3171 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3172 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3173 | } |
3174 | ||
d5ed7444 | 3175 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3176 | { |
3177 | swp_entry_t swp; | |
3178 | ||
3179 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3180 | return false; |
4a705fef NH |
3181 | swp = pte_to_swp_entry(pte); |
3182 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3183 | return true; |
4a705fef | 3184 | else |
d5ed7444 | 3185 | return false; |
4a705fef NH |
3186 | } |
3187 | ||
3188 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3189 | { | |
3190 | swp_entry_t swp; | |
3191 | ||
3192 | if (huge_pte_none(pte) || pte_present(pte)) | |
3193 | return 0; | |
3194 | swp = pte_to_swp_entry(pte); | |
3195 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3196 | return 1; | |
3197 | else | |
3198 | return 0; | |
3199 | } | |
1e8f889b | 3200 | |
63551ae0 DG |
3201 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3202 | struct vm_area_struct *vma) | |
3203 | { | |
3204 | pte_t *src_pte, *dst_pte, entry; | |
3205 | struct page *ptepage; | |
1c59827d | 3206 | unsigned long addr; |
1e8f889b | 3207 | int cow; |
a5516438 AK |
3208 | struct hstate *h = hstate_vma(vma); |
3209 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
3210 | unsigned long mmun_start; /* For mmu_notifiers */ |
3211 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3212 | int ret = 0; | |
1e8f889b DG |
3213 | |
3214 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3215 | |
e8569dd2 AS |
3216 | mmun_start = vma->vm_start; |
3217 | mmun_end = vma->vm_end; | |
3218 | if (cow) | |
3219 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
3220 | ||
a5516438 | 3221 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3222 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3223 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3224 | if (!src_pte) |
3225 | continue; | |
a5516438 | 3226 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3227 | if (!dst_pte) { |
3228 | ret = -ENOMEM; | |
3229 | break; | |
3230 | } | |
c5c99429 LW |
3231 | |
3232 | /* If the pagetables are shared don't copy or take references */ | |
3233 | if (dst_pte == src_pte) | |
3234 | continue; | |
3235 | ||
cb900f41 KS |
3236 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3237 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3238 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3239 | entry = huge_ptep_get(src_pte); |
3240 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3241 | ; | |
3242 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3243 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3244 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3245 | ||
3246 | if (is_write_migration_entry(swp_entry) && cow) { | |
3247 | /* | |
3248 | * COW mappings require pages in both | |
3249 | * parent and child to be set to read. | |
3250 | */ | |
3251 | make_migration_entry_read(&swp_entry); | |
3252 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3253 | set_huge_swap_pte_at(src, addr, src_pte, |
3254 | entry, sz); | |
4a705fef | 3255 | } |
e5251fd4 | 3256 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3257 | } else { |
34ee645e | 3258 | if (cow) { |
7f2e9525 | 3259 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
3260 | mmu_notifier_invalidate_range(src, mmun_start, |
3261 | mmun_end); | |
3262 | } | |
0253d634 | 3263 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3264 | ptepage = pte_page(entry); |
3265 | get_page(ptepage); | |
53f9263b | 3266 | page_dup_rmap(ptepage, true); |
1c59827d | 3267 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3268 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3269 | } |
cb900f41 KS |
3270 | spin_unlock(src_ptl); |
3271 | spin_unlock(dst_ptl); | |
63551ae0 | 3272 | } |
63551ae0 | 3273 | |
e8569dd2 AS |
3274 | if (cow) |
3275 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3276 | ||
3277 | return ret; | |
63551ae0 DG |
3278 | } |
3279 | ||
24669e58 AK |
3280 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3281 | unsigned long start, unsigned long end, | |
3282 | struct page *ref_page) | |
63551ae0 DG |
3283 | { |
3284 | struct mm_struct *mm = vma->vm_mm; | |
3285 | unsigned long address; | |
c7546f8f | 3286 | pte_t *ptep; |
63551ae0 | 3287 | pte_t pte; |
cb900f41 | 3288 | spinlock_t *ptl; |
63551ae0 | 3289 | struct page *page; |
a5516438 AK |
3290 | struct hstate *h = hstate_vma(vma); |
3291 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3292 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3293 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3294 | |
63551ae0 | 3295 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3296 | BUG_ON(start & ~huge_page_mask(h)); |
3297 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3298 | |
07e32661 AK |
3299 | /* |
3300 | * This is a hugetlb vma, all the pte entries should point | |
3301 | * to huge page. | |
3302 | */ | |
3303 | tlb_remove_check_page_size_change(tlb, sz); | |
24669e58 | 3304 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3305 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3306 | address = start; |
569f48b8 | 3307 | for (; address < end; address += sz) { |
7868a208 | 3308 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3309 | if (!ptep) |
c7546f8f DG |
3310 | continue; |
3311 | ||
cb900f41 | 3312 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3313 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3314 | spin_unlock(ptl); | |
3315 | continue; | |
3316 | } | |
39dde65c | 3317 | |
6629326b | 3318 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3319 | if (huge_pte_none(pte)) { |
3320 | spin_unlock(ptl); | |
3321 | continue; | |
3322 | } | |
6629326b HD |
3323 | |
3324 | /* | |
9fbc1f63 NH |
3325 | * Migrating hugepage or HWPoisoned hugepage is already |
3326 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3327 | */ |
9fbc1f63 | 3328 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3329 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3330 | spin_unlock(ptl); |
3331 | continue; | |
8c4894c6 | 3332 | } |
6629326b HD |
3333 | |
3334 | page = pte_page(pte); | |
04f2cbe3 MG |
3335 | /* |
3336 | * If a reference page is supplied, it is because a specific | |
3337 | * page is being unmapped, not a range. Ensure the page we | |
3338 | * are about to unmap is the actual page of interest. | |
3339 | */ | |
3340 | if (ref_page) { | |
31d49da5 AK |
3341 | if (page != ref_page) { |
3342 | spin_unlock(ptl); | |
3343 | continue; | |
3344 | } | |
04f2cbe3 MG |
3345 | /* |
3346 | * Mark the VMA as having unmapped its page so that | |
3347 | * future faults in this VMA will fail rather than | |
3348 | * looking like data was lost | |
3349 | */ | |
3350 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3351 | } | |
3352 | ||
c7546f8f | 3353 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3354 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3355 | if (huge_pte_dirty(pte)) |
6649a386 | 3356 | set_page_dirty(page); |
9e81130b | 3357 | |
5d317b2b | 3358 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3359 | page_remove_rmap(page, true); |
31d49da5 | 3360 | |
cb900f41 | 3361 | spin_unlock(ptl); |
e77b0852 | 3362 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3363 | /* |
3364 | * Bail out after unmapping reference page if supplied | |
3365 | */ | |
3366 | if (ref_page) | |
3367 | break; | |
fe1668ae | 3368 | } |
2ec74c3e | 3369 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3370 | tlb_end_vma(tlb, vma); |
1da177e4 | 3371 | } |
63551ae0 | 3372 | |
d833352a MG |
3373 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3374 | struct vm_area_struct *vma, unsigned long start, | |
3375 | unsigned long end, struct page *ref_page) | |
3376 | { | |
3377 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3378 | ||
3379 | /* | |
3380 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3381 | * test will fail on a vma being torn down, and not grab a page table | |
3382 | * on its way out. We're lucky that the flag has such an appropriate | |
3383 | * name, and can in fact be safely cleared here. We could clear it | |
3384 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3385 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3386 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3387 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3388 | */ |
3389 | vma->vm_flags &= ~VM_MAYSHARE; | |
3390 | } | |
3391 | ||
502717f4 | 3392 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3393 | unsigned long end, struct page *ref_page) |
502717f4 | 3394 | { |
24669e58 AK |
3395 | struct mm_struct *mm; |
3396 | struct mmu_gather tlb; | |
3397 | ||
3398 | mm = vma->vm_mm; | |
3399 | ||
2b047252 | 3400 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3401 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3402 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 KC |
3403 | } |
3404 | ||
04f2cbe3 MG |
3405 | /* |
3406 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3407 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3408 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3409 | * same region. | |
3410 | */ | |
2f4612af DB |
3411 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3412 | struct page *page, unsigned long address) | |
04f2cbe3 | 3413 | { |
7526674d | 3414 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3415 | struct vm_area_struct *iter_vma; |
3416 | struct address_space *mapping; | |
04f2cbe3 MG |
3417 | pgoff_t pgoff; |
3418 | ||
3419 | /* | |
3420 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3421 | * from page cache lookup which is in HPAGE_SIZE units. | |
3422 | */ | |
7526674d | 3423 | address = address & huge_page_mask(h); |
36e4f20a MH |
3424 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3425 | vma->vm_pgoff; | |
93c76a3d | 3426 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3427 | |
4eb2b1dc MG |
3428 | /* |
3429 | * Take the mapping lock for the duration of the table walk. As | |
3430 | * this mapping should be shared between all the VMAs, | |
3431 | * __unmap_hugepage_range() is called as the lock is already held | |
3432 | */ | |
83cde9e8 | 3433 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3434 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3435 | /* Do not unmap the current VMA */ |
3436 | if (iter_vma == vma) | |
3437 | continue; | |
3438 | ||
2f84a899 MG |
3439 | /* |
3440 | * Shared VMAs have their own reserves and do not affect | |
3441 | * MAP_PRIVATE accounting but it is possible that a shared | |
3442 | * VMA is using the same page so check and skip such VMAs. | |
3443 | */ | |
3444 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3445 | continue; | |
3446 | ||
04f2cbe3 MG |
3447 | /* |
3448 | * Unmap the page from other VMAs without their own reserves. | |
3449 | * They get marked to be SIGKILLed if they fault in these | |
3450 | * areas. This is because a future no-page fault on this VMA | |
3451 | * could insert a zeroed page instead of the data existing | |
3452 | * from the time of fork. This would look like data corruption | |
3453 | */ | |
3454 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3455 | unmap_hugepage_range(iter_vma, address, |
3456 | address + huge_page_size(h), page); | |
04f2cbe3 | 3457 | } |
83cde9e8 | 3458 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3459 | } |
3460 | ||
0fe6e20b NH |
3461 | /* |
3462 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3463 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3464 | * cannot race with other handlers or page migration. | |
3465 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3466 | */ |
1e8f889b | 3467 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
3999f52e AK |
3468 | unsigned long address, pte_t *ptep, |
3469 | struct page *pagecache_page, spinlock_t *ptl) | |
1e8f889b | 3470 | { |
3999f52e | 3471 | pte_t pte; |
a5516438 | 3472 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3473 | struct page *old_page, *new_page; |
ad4404a2 | 3474 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3475 | unsigned long mmun_start; /* For mmu_notifiers */ |
3476 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b | 3477 | |
3999f52e | 3478 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3479 | old_page = pte_page(pte); |
3480 | ||
04f2cbe3 | 3481 | retry_avoidcopy: |
1e8f889b DG |
3482 | /* If no-one else is actually using this page, avoid the copy |
3483 | * and just make the page writable */ | |
37a2140d | 3484 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3485 | page_move_anon_rmap(old_page, vma); |
1e8f889b | 3486 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3487 | return 0; |
1e8f889b DG |
3488 | } |
3489 | ||
04f2cbe3 MG |
3490 | /* |
3491 | * If the process that created a MAP_PRIVATE mapping is about to | |
3492 | * perform a COW due to a shared page count, attempt to satisfy | |
3493 | * the allocation without using the existing reserves. The pagecache | |
3494 | * page is used to determine if the reserve at this address was | |
3495 | * consumed or not. If reserves were used, a partial faulted mapping | |
3496 | * at the time of fork() could consume its reserves on COW instead | |
3497 | * of the full address range. | |
3498 | */ | |
5944d011 | 3499 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3500 | old_page != pagecache_page) |
3501 | outside_reserve = 1; | |
3502 | ||
09cbfeaf | 3503 | get_page(old_page); |
b76c8cfb | 3504 | |
ad4404a2 DB |
3505 | /* |
3506 | * Drop page table lock as buddy allocator may be called. It will | |
3507 | * be acquired again before returning to the caller, as expected. | |
3508 | */ | |
cb900f41 | 3509 | spin_unlock(ptl); |
04f2cbe3 | 3510 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3511 | |
2fc39cec | 3512 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3513 | /* |
3514 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3515 | * it is due to references held by a child and an insufficient | |
3516 | * huge page pool. To guarantee the original mappers | |
3517 | * reliability, unmap the page from child processes. The child | |
3518 | * may get SIGKILLed if it later faults. | |
3519 | */ | |
3520 | if (outside_reserve) { | |
09cbfeaf | 3521 | put_page(old_page); |
04f2cbe3 | 3522 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3523 | unmap_ref_private(mm, vma, old_page, address); |
3524 | BUG_ON(huge_pte_none(pte)); | |
3525 | spin_lock(ptl); | |
7868a208 PA |
3526 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3527 | huge_page_size(h)); | |
2f4612af DB |
3528 | if (likely(ptep && |
3529 | pte_same(huge_ptep_get(ptep), pte))) | |
3530 | goto retry_avoidcopy; | |
3531 | /* | |
3532 | * race occurs while re-acquiring page table | |
3533 | * lock, and our job is done. | |
3534 | */ | |
3535 | return 0; | |
04f2cbe3 MG |
3536 | } |
3537 | ||
ad4404a2 DB |
3538 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3539 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3540 | goto out_release_old; | |
1e8f889b DG |
3541 | } |
3542 | ||
0fe6e20b NH |
3543 | /* |
3544 | * When the original hugepage is shared one, it does not have | |
3545 | * anon_vma prepared. | |
3546 | */ | |
44e2aa93 | 3547 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3548 | ret = VM_FAULT_OOM; |
3549 | goto out_release_all; | |
44e2aa93 | 3550 | } |
0fe6e20b | 3551 | |
47ad8475 AA |
3552 | copy_user_huge_page(new_page, old_page, address, vma, |
3553 | pages_per_huge_page(h)); | |
0ed361de | 3554 | __SetPageUptodate(new_page); |
bcc54222 | 3555 | set_page_huge_active(new_page); |
1e8f889b | 3556 | |
2ec74c3e SG |
3557 | mmun_start = address & huge_page_mask(h); |
3558 | mmun_end = mmun_start + huge_page_size(h); | |
3559 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3560 | |
b76c8cfb | 3561 | /* |
cb900f41 | 3562 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3563 | * before the page tables are altered |
3564 | */ | |
cb900f41 | 3565 | spin_lock(ptl); |
7868a208 PA |
3566 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3567 | huge_page_size(h)); | |
a9af0c5d | 3568 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3569 | ClearPagePrivate(new_page); |
3570 | ||
1e8f889b | 3571 | /* Break COW */ |
8fe627ec | 3572 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3573 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3574 | set_huge_pte_at(mm, address, ptep, |
3575 | make_huge_pte(vma, new_page, 1)); | |
d281ee61 | 3576 | page_remove_rmap(old_page, true); |
cd67f0d2 | 3577 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3578 | /* Make the old page be freed below */ |
3579 | new_page = old_page; | |
3580 | } | |
cb900f41 | 3581 | spin_unlock(ptl); |
2ec74c3e | 3582 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3583 | out_release_all: |
96b96a96 | 3584 | restore_reserve_on_error(h, vma, address, new_page); |
09cbfeaf | 3585 | put_page(new_page); |
ad4404a2 | 3586 | out_release_old: |
09cbfeaf | 3587 | put_page(old_page); |
8312034f | 3588 | |
ad4404a2 DB |
3589 | spin_lock(ptl); /* Caller expects lock to be held */ |
3590 | return ret; | |
1e8f889b DG |
3591 | } |
3592 | ||
04f2cbe3 | 3593 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3594 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3595 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3596 | { |
3597 | struct address_space *mapping; | |
e7c4b0bf | 3598 | pgoff_t idx; |
04f2cbe3 MG |
3599 | |
3600 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3601 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3602 | |
3603 | return find_lock_page(mapping, idx); | |
3604 | } | |
3605 | ||
3ae77f43 HD |
3606 | /* |
3607 | * Return whether there is a pagecache page to back given address within VMA. | |
3608 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3609 | */ | |
3610 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3611 | struct vm_area_struct *vma, unsigned long address) |
3612 | { | |
3613 | struct address_space *mapping; | |
3614 | pgoff_t idx; | |
3615 | struct page *page; | |
3616 | ||
3617 | mapping = vma->vm_file->f_mapping; | |
3618 | idx = vma_hugecache_offset(h, vma, address); | |
3619 | ||
3620 | page = find_get_page(mapping, idx); | |
3621 | if (page) | |
3622 | put_page(page); | |
3623 | return page != NULL; | |
3624 | } | |
3625 | ||
ab76ad54 MK |
3626 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3627 | pgoff_t idx) | |
3628 | { | |
3629 | struct inode *inode = mapping->host; | |
3630 | struct hstate *h = hstate_inode(inode); | |
3631 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3632 | ||
3633 | if (err) | |
3634 | return err; | |
3635 | ClearPagePrivate(page); | |
3636 | ||
3637 | spin_lock(&inode->i_lock); | |
3638 | inode->i_blocks += blocks_per_huge_page(h); | |
3639 | spin_unlock(&inode->i_lock); | |
3640 | return 0; | |
3641 | } | |
3642 | ||
a1ed3dda | 3643 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3644 | struct address_space *mapping, pgoff_t idx, |
3645 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3646 | { |
a5516438 | 3647 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3648 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3649 | int anon_rmap = 0; |
4c887265 | 3650 | unsigned long size; |
4c887265 | 3651 | struct page *page; |
1e8f889b | 3652 | pte_t new_pte; |
cb900f41 | 3653 | spinlock_t *ptl; |
4c887265 | 3654 | |
04f2cbe3 MG |
3655 | /* |
3656 | * Currently, we are forced to kill the process in the event the | |
3657 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3658 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3659 | */ |
3660 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3661 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3662 | current->pid); |
04f2cbe3 MG |
3663 | return ret; |
3664 | } | |
3665 | ||
4c887265 AL |
3666 | /* |
3667 | * Use page lock to guard against racing truncation | |
3668 | * before we get page_table_lock. | |
3669 | */ | |
6bda666a CL |
3670 | retry: |
3671 | page = find_lock_page(mapping, idx); | |
3672 | if (!page) { | |
a5516438 | 3673 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3674 | if (idx >= size) |
3675 | goto out; | |
1a1aad8a MK |
3676 | |
3677 | /* | |
3678 | * Check for page in userfault range | |
3679 | */ | |
3680 | if (userfaultfd_missing(vma)) { | |
3681 | u32 hash; | |
3682 | struct vm_fault vmf = { | |
3683 | .vma = vma, | |
3684 | .address = address, | |
3685 | .flags = flags, | |
3686 | /* | |
3687 | * Hard to debug if it ends up being | |
3688 | * used by a callee that assumes | |
3689 | * something about the other | |
3690 | * uninitialized fields... same as in | |
3691 | * memory.c | |
3692 | */ | |
3693 | }; | |
3694 | ||
3695 | /* | |
3696 | * hugetlb_fault_mutex must be dropped before | |
3697 | * handling userfault. Reacquire after handling | |
3698 | * fault to make calling code simpler. | |
3699 | */ | |
3700 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, | |
3701 | idx, address); | |
3702 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
3703 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3704 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3705 | goto out; | |
3706 | } | |
3707 | ||
04f2cbe3 | 3708 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3709 | if (IS_ERR(page)) { |
76dcee75 AK |
3710 | ret = PTR_ERR(page); |
3711 | if (ret == -ENOMEM) | |
3712 | ret = VM_FAULT_OOM; | |
3713 | else | |
3714 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3715 | goto out; |
3716 | } | |
47ad8475 | 3717 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3718 | __SetPageUptodate(page); |
bcc54222 | 3719 | set_page_huge_active(page); |
ac9b9c66 | 3720 | |
f83a275d | 3721 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3722 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3723 | if (err) { |
3724 | put_page(page); | |
6bda666a CL |
3725 | if (err == -EEXIST) |
3726 | goto retry; | |
3727 | goto out; | |
3728 | } | |
23be7468 | 3729 | } else { |
6bda666a | 3730 | lock_page(page); |
0fe6e20b NH |
3731 | if (unlikely(anon_vma_prepare(vma))) { |
3732 | ret = VM_FAULT_OOM; | |
3733 | goto backout_unlocked; | |
3734 | } | |
409eb8c2 | 3735 | anon_rmap = 1; |
23be7468 | 3736 | } |
0fe6e20b | 3737 | } else { |
998b4382 NH |
3738 | /* |
3739 | * If memory error occurs between mmap() and fault, some process | |
3740 | * don't have hwpoisoned swap entry for errored virtual address. | |
3741 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3742 | */ | |
3743 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3744 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3745 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3746 | goto backout_unlocked; |
3747 | } | |
6bda666a | 3748 | } |
1e8f889b | 3749 | |
57303d80 AW |
3750 | /* |
3751 | * If we are going to COW a private mapping later, we examine the | |
3752 | * pending reservations for this page now. This will ensure that | |
3753 | * any allocations necessary to record that reservation occur outside | |
3754 | * the spinlock. | |
3755 | */ | |
5e911373 | 3756 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3757 | if (vma_needs_reservation(h, vma, address) < 0) { |
3758 | ret = VM_FAULT_OOM; | |
3759 | goto backout_unlocked; | |
3760 | } | |
5e911373 | 3761 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3762 | vma_end_reservation(h, vma, address); |
5e911373 | 3763 | } |
57303d80 | 3764 | |
8bea8052 | 3765 | ptl = huge_pte_lock(h, mm, ptep); |
a5516438 | 3766 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3767 | if (idx >= size) |
3768 | goto backout; | |
3769 | ||
83c54070 | 3770 | ret = 0; |
7f2e9525 | 3771 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3772 | goto backout; |
3773 | ||
07443a85 JK |
3774 | if (anon_rmap) { |
3775 | ClearPagePrivate(page); | |
409eb8c2 | 3776 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3777 | } else |
53f9263b | 3778 | page_dup_rmap(page, true); |
1e8f889b DG |
3779 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3780 | && (vma->vm_flags & VM_SHARED))); | |
3781 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3782 | ||
5d317b2b | 3783 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3784 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3785 | /* Optimization, do the COW without a second fault */ |
3999f52e | 3786 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
3787 | } |
3788 | ||
cb900f41 | 3789 | spin_unlock(ptl); |
4c887265 AL |
3790 | unlock_page(page); |
3791 | out: | |
ac9b9c66 | 3792 | return ret; |
4c887265 AL |
3793 | |
3794 | backout: | |
cb900f41 | 3795 | spin_unlock(ptl); |
2b26736c | 3796 | backout_unlocked: |
4c887265 | 3797 | unlock_page(page); |
96b96a96 | 3798 | restore_reserve_on_error(h, vma, address, page); |
4c887265 AL |
3799 | put_page(page); |
3800 | goto out; | |
ac9b9c66 HD |
3801 | } |
3802 | ||
8382d914 | 3803 | #ifdef CONFIG_SMP |
c672c7f2 | 3804 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3805 | struct vm_area_struct *vma, |
3806 | struct address_space *mapping, | |
3807 | pgoff_t idx, unsigned long address) | |
3808 | { | |
3809 | unsigned long key[2]; | |
3810 | u32 hash; | |
3811 | ||
3812 | if (vma->vm_flags & VM_SHARED) { | |
3813 | key[0] = (unsigned long) mapping; | |
3814 | key[1] = idx; | |
3815 | } else { | |
3816 | key[0] = (unsigned long) mm; | |
3817 | key[1] = address >> huge_page_shift(h); | |
3818 | } | |
3819 | ||
3820 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3821 | ||
3822 | return hash & (num_fault_mutexes - 1); | |
3823 | } | |
3824 | #else | |
3825 | /* | |
3826 | * For uniprocesor systems we always use a single mutex, so just | |
3827 | * return 0 and avoid the hashing overhead. | |
3828 | */ | |
c672c7f2 | 3829 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3830 | struct vm_area_struct *vma, |
3831 | struct address_space *mapping, | |
3832 | pgoff_t idx, unsigned long address) | |
3833 | { | |
3834 | return 0; | |
3835 | } | |
3836 | #endif | |
3837 | ||
86e5216f | 3838 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3839 | unsigned long address, unsigned int flags) |
86e5216f | 3840 | { |
8382d914 | 3841 | pte_t *ptep, entry; |
cb900f41 | 3842 | spinlock_t *ptl; |
1e8f889b | 3843 | int ret; |
8382d914 DB |
3844 | u32 hash; |
3845 | pgoff_t idx; | |
0fe6e20b | 3846 | struct page *page = NULL; |
57303d80 | 3847 | struct page *pagecache_page = NULL; |
a5516438 | 3848 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3849 | struct address_space *mapping; |
0f792cf9 | 3850 | int need_wait_lock = 0; |
86e5216f | 3851 | |
1e16a539 KH |
3852 | address &= huge_page_mask(h); |
3853 | ||
7868a208 | 3854 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
fd6a03ed NH |
3855 | if (ptep) { |
3856 | entry = huge_ptep_get(ptep); | |
290408d4 | 3857 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3858 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3859 | return 0; |
3860 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3861 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3862 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
0d777df5 NH |
3863 | } else { |
3864 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); | |
3865 | if (!ptep) | |
3866 | return VM_FAULT_OOM; | |
fd6a03ed NH |
3867 | } |
3868 | ||
8382d914 DB |
3869 | mapping = vma->vm_file->f_mapping; |
3870 | idx = vma_hugecache_offset(h, vma, address); | |
3871 | ||
3935baa9 DG |
3872 | /* |
3873 | * Serialize hugepage allocation and instantiation, so that we don't | |
3874 | * get spurious allocation failures if two CPUs race to instantiate | |
3875 | * the same page in the page cache. | |
3876 | */ | |
c672c7f2 MK |
3877 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3878 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3879 | |
7f2e9525 GS |
3880 | entry = huge_ptep_get(ptep); |
3881 | if (huge_pte_none(entry)) { | |
8382d914 | 3882 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3883 | goto out_mutex; |
3935baa9 | 3884 | } |
86e5216f | 3885 | |
83c54070 | 3886 | ret = 0; |
1e8f889b | 3887 | |
0f792cf9 NH |
3888 | /* |
3889 | * entry could be a migration/hwpoison entry at this point, so this | |
3890 | * check prevents the kernel from going below assuming that we have | |
3891 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3892 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3893 | * handle it. | |
3894 | */ | |
3895 | if (!pte_present(entry)) | |
3896 | goto out_mutex; | |
3897 | ||
57303d80 AW |
3898 | /* |
3899 | * If we are going to COW the mapping later, we examine the pending | |
3900 | * reservations for this page now. This will ensure that any | |
3901 | * allocations necessary to record that reservation occur outside the | |
3902 | * spinlock. For private mappings, we also lookup the pagecache | |
3903 | * page now as it is used to determine if a reservation has been | |
3904 | * consumed. | |
3905 | */ | |
106c992a | 3906 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3907 | if (vma_needs_reservation(h, vma, address) < 0) { |
3908 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3909 | goto out_mutex; |
2b26736c | 3910 | } |
5e911373 | 3911 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3912 | vma_end_reservation(h, vma, address); |
57303d80 | 3913 | |
f83a275d | 3914 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3915 | pagecache_page = hugetlbfs_pagecache_page(h, |
3916 | vma, address); | |
3917 | } | |
3918 | ||
0f792cf9 NH |
3919 | ptl = huge_pte_lock(h, mm, ptep); |
3920 | ||
3921 | /* Check for a racing update before calling hugetlb_cow */ | |
3922 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3923 | goto out_ptl; | |
3924 | ||
56c9cfb1 NH |
3925 | /* |
3926 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3927 | * pagecache_page, so here we need take the former one | |
3928 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3929 | */ |
3930 | page = pte_page(entry); | |
3931 | if (page != pagecache_page) | |
0f792cf9 NH |
3932 | if (!trylock_page(page)) { |
3933 | need_wait_lock = 1; | |
3934 | goto out_ptl; | |
3935 | } | |
b4d1d99f | 3936 | |
0f792cf9 | 3937 | get_page(page); |
b4d1d99f | 3938 | |
788c7df4 | 3939 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3940 | if (!huge_pte_write(entry)) { |
3999f52e AK |
3941 | ret = hugetlb_cow(mm, vma, address, ptep, |
3942 | pagecache_page, ptl); | |
0f792cf9 | 3943 | goto out_put_page; |
b4d1d99f | 3944 | } |
106c992a | 3945 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3946 | } |
3947 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3948 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3949 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3950 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3951 | out_put_page: |
3952 | if (page != pagecache_page) | |
3953 | unlock_page(page); | |
3954 | put_page(page); | |
cb900f41 KS |
3955 | out_ptl: |
3956 | spin_unlock(ptl); | |
57303d80 AW |
3957 | |
3958 | if (pagecache_page) { | |
3959 | unlock_page(pagecache_page); | |
3960 | put_page(pagecache_page); | |
3961 | } | |
b4d1d99f | 3962 | out_mutex: |
c672c7f2 | 3963 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3964 | /* |
3965 | * Generally it's safe to hold refcount during waiting page lock. But | |
3966 | * here we just wait to defer the next page fault to avoid busy loop and | |
3967 | * the page is not used after unlocked before returning from the current | |
3968 | * page fault. So we are safe from accessing freed page, even if we wait | |
3969 | * here without taking refcount. | |
3970 | */ | |
3971 | if (need_wait_lock) | |
3972 | wait_on_page_locked(page); | |
1e8f889b | 3973 | return ret; |
86e5216f AL |
3974 | } |
3975 | ||
8fb5debc MK |
3976 | /* |
3977 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
3978 | * modifications for huge pages. | |
3979 | */ | |
3980 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
3981 | pte_t *dst_pte, | |
3982 | struct vm_area_struct *dst_vma, | |
3983 | unsigned long dst_addr, | |
3984 | unsigned long src_addr, | |
3985 | struct page **pagep) | |
3986 | { | |
1e392147 AA |
3987 | struct address_space *mapping; |
3988 | pgoff_t idx; | |
3989 | unsigned long size; | |
1c9e8def | 3990 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
3991 | struct hstate *h = hstate_vma(dst_vma); |
3992 | pte_t _dst_pte; | |
3993 | spinlock_t *ptl; | |
3994 | int ret; | |
3995 | struct page *page; | |
3996 | ||
3997 | if (!*pagep) { | |
3998 | ret = -ENOMEM; | |
3999 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4000 | if (IS_ERR(page)) | |
4001 | goto out; | |
4002 | ||
4003 | ret = copy_huge_page_from_user(page, | |
4004 | (const void __user *) src_addr, | |
810a56b9 | 4005 | pages_per_huge_page(h), false); |
8fb5debc MK |
4006 | |
4007 | /* fallback to copy_from_user outside mmap_sem */ | |
4008 | if (unlikely(ret)) { | |
4009 | ret = -EFAULT; | |
4010 | *pagep = page; | |
4011 | /* don't free the page */ | |
4012 | goto out; | |
4013 | } | |
4014 | } else { | |
4015 | page = *pagep; | |
4016 | *pagep = NULL; | |
4017 | } | |
4018 | ||
4019 | /* | |
4020 | * The memory barrier inside __SetPageUptodate makes sure that | |
4021 | * preceding stores to the page contents become visible before | |
4022 | * the set_pte_at() write. | |
4023 | */ | |
4024 | __SetPageUptodate(page); | |
4025 | set_page_huge_active(page); | |
4026 | ||
1e392147 AA |
4027 | mapping = dst_vma->vm_file->f_mapping; |
4028 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4029 | ||
1c9e8def MK |
4030 | /* |
4031 | * If shared, add to page cache | |
4032 | */ | |
4033 | if (vm_shared) { | |
1e392147 AA |
4034 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4035 | ret = -EFAULT; | |
4036 | if (idx >= size) | |
4037 | goto out_release_nounlock; | |
1c9e8def | 4038 | |
1e392147 AA |
4039 | /* |
4040 | * Serialization between remove_inode_hugepages() and | |
4041 | * huge_add_to_page_cache() below happens through the | |
4042 | * hugetlb_fault_mutex_table that here must be hold by | |
4043 | * the caller. | |
4044 | */ | |
1c9e8def MK |
4045 | ret = huge_add_to_page_cache(page, mapping, idx); |
4046 | if (ret) | |
4047 | goto out_release_nounlock; | |
4048 | } | |
4049 | ||
8fb5debc MK |
4050 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4051 | spin_lock(ptl); | |
4052 | ||
1e392147 AA |
4053 | /* |
4054 | * Recheck the i_size after holding PT lock to make sure not | |
4055 | * to leave any page mapped (as page_mapped()) beyond the end | |
4056 | * of the i_size (remove_inode_hugepages() is strict about | |
4057 | * enforcing that). If we bail out here, we'll also leave a | |
4058 | * page in the radix tree in the vm_shared case beyond the end | |
4059 | * of the i_size, but remove_inode_hugepages() will take care | |
4060 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4061 | */ | |
4062 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4063 | ret = -EFAULT; | |
4064 | if (idx >= size) | |
4065 | goto out_release_unlock; | |
4066 | ||
8fb5debc MK |
4067 | ret = -EEXIST; |
4068 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4069 | goto out_release_unlock; | |
4070 | ||
1c9e8def MK |
4071 | if (vm_shared) { |
4072 | page_dup_rmap(page, true); | |
4073 | } else { | |
4074 | ClearPagePrivate(page); | |
4075 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4076 | } | |
8fb5debc MK |
4077 | |
4078 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4079 | if (dst_vma->vm_flags & VM_WRITE) | |
4080 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4081 | _dst_pte = pte_mkyoung(_dst_pte); | |
4082 | ||
4083 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4084 | ||
4085 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4086 | dst_vma->vm_flags & VM_WRITE); | |
4087 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4088 | ||
4089 | /* No need to invalidate - it was non-present before */ | |
4090 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4091 | ||
4092 | spin_unlock(ptl); | |
1c9e8def MK |
4093 | if (vm_shared) |
4094 | unlock_page(page); | |
8fb5debc MK |
4095 | ret = 0; |
4096 | out: | |
4097 | return ret; | |
4098 | out_release_unlock: | |
4099 | spin_unlock(ptl); | |
1c9e8def MK |
4100 | if (vm_shared) |
4101 | unlock_page(page); | |
5af10dfd | 4102 | out_release_nounlock: |
8fb5debc MK |
4103 | put_page(page); |
4104 | goto out; | |
4105 | } | |
4106 | ||
28a35716 ML |
4107 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4108 | struct page **pages, struct vm_area_struct **vmas, | |
4109 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4110 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4111 | { |
d5d4b0aa KC |
4112 | unsigned long pfn_offset; |
4113 | unsigned long vaddr = *position; | |
28a35716 | 4114 | unsigned long remainder = *nr_pages; |
a5516438 | 4115 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4116 | int err = -EFAULT; |
63551ae0 | 4117 | |
63551ae0 | 4118 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4119 | pte_t *pte; |
cb900f41 | 4120 | spinlock_t *ptl = NULL; |
2a15efc9 | 4121 | int absent; |
4c887265 | 4122 | struct page *page; |
63551ae0 | 4123 | |
02057967 DR |
4124 | /* |
4125 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4126 | * potentially allocating memory. | |
4127 | */ | |
4128 | if (unlikely(fatal_signal_pending(current))) { | |
4129 | remainder = 0; | |
4130 | break; | |
4131 | } | |
4132 | ||
4c887265 AL |
4133 | /* |
4134 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4135 | * each hugepage. We have to make sure we get the |
4c887265 | 4136 | * first, for the page indexing below to work. |
cb900f41 KS |
4137 | * |
4138 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4139 | */ |
7868a208 PA |
4140 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4141 | huge_page_size(h)); | |
cb900f41 KS |
4142 | if (pte) |
4143 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4144 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4145 | ||
4146 | /* | |
4147 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4148 | * an error where there's an empty slot with no huge pagecache |
4149 | * to back it. This way, we avoid allocating a hugepage, and | |
4150 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4151 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4152 | */ |
3ae77f43 HD |
4153 | if (absent && (flags & FOLL_DUMP) && |
4154 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4155 | if (pte) |
4156 | spin_unlock(ptl); | |
2a15efc9 HD |
4157 | remainder = 0; |
4158 | break; | |
4159 | } | |
63551ae0 | 4160 | |
9cc3a5bd NH |
4161 | /* |
4162 | * We need call hugetlb_fault for both hugepages under migration | |
4163 | * (in which case hugetlb_fault waits for the migration,) and | |
4164 | * hwpoisoned hugepages (in which case we need to prevent the | |
4165 | * caller from accessing to them.) In order to do this, we use | |
4166 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4167 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4168 | * both cases, and because we can't follow correct pages | |
4169 | * directly from any kind of swap entries. | |
4170 | */ | |
4171 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4172 | ((flags & FOLL_WRITE) && |
4173 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 4174 | int ret; |
87ffc118 | 4175 | unsigned int fault_flags = 0; |
63551ae0 | 4176 | |
cb900f41 KS |
4177 | if (pte) |
4178 | spin_unlock(ptl); | |
87ffc118 AA |
4179 | if (flags & FOLL_WRITE) |
4180 | fault_flags |= FAULT_FLAG_WRITE; | |
4181 | if (nonblocking) | |
4182 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4183 | if (flags & FOLL_NOWAIT) | |
4184 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4185 | FAULT_FLAG_RETRY_NOWAIT; | |
4186 | if (flags & FOLL_TRIED) { | |
4187 | VM_WARN_ON_ONCE(fault_flags & | |
4188 | FAULT_FLAG_ALLOW_RETRY); | |
4189 | fault_flags |= FAULT_FLAG_TRIED; | |
4190 | } | |
4191 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4192 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4193 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4194 | remainder = 0; |
4195 | break; | |
4196 | } | |
4197 | if (ret & VM_FAULT_RETRY) { | |
4198 | if (nonblocking) | |
4199 | *nonblocking = 0; | |
4200 | *nr_pages = 0; | |
4201 | /* | |
4202 | * VM_FAULT_RETRY must not return an | |
4203 | * error, it will return zero | |
4204 | * instead. | |
4205 | * | |
4206 | * No need to update "position" as the | |
4207 | * caller will not check it after | |
4208 | * *nr_pages is set to 0. | |
4209 | */ | |
4210 | return i; | |
4211 | } | |
4212 | continue; | |
4c887265 AL |
4213 | } |
4214 | ||
a5516438 | 4215 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4216 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 4217 | same_page: |
d6692183 | 4218 | if (pages) { |
2a15efc9 | 4219 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 4220 | get_page(pages[i]); |
d6692183 | 4221 | } |
63551ae0 DG |
4222 | |
4223 | if (vmas) | |
4224 | vmas[i] = vma; | |
4225 | ||
4226 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4227 | ++pfn_offset; |
63551ae0 DG |
4228 | --remainder; |
4229 | ++i; | |
d5d4b0aa | 4230 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4231 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4232 | /* |
4233 | * We use pfn_offset to avoid touching the pageframes | |
4234 | * of this compound page. | |
4235 | */ | |
4236 | goto same_page; | |
4237 | } | |
cb900f41 | 4238 | spin_unlock(ptl); |
63551ae0 | 4239 | } |
28a35716 | 4240 | *nr_pages = remainder; |
87ffc118 AA |
4241 | /* |
4242 | * setting position is actually required only if remainder is | |
4243 | * not zero but it's faster not to add a "if (remainder)" | |
4244 | * branch. | |
4245 | */ | |
63551ae0 DG |
4246 | *position = vaddr; |
4247 | ||
2be7cfed | 4248 | return i ? i : err; |
63551ae0 | 4249 | } |
8f860591 | 4250 | |
5491ae7b AK |
4251 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4252 | /* | |
4253 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4254 | * implement this. | |
4255 | */ | |
4256 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4257 | #endif | |
4258 | ||
7da4d641 | 4259 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4260 | unsigned long address, unsigned long end, pgprot_t newprot) |
4261 | { | |
4262 | struct mm_struct *mm = vma->vm_mm; | |
4263 | unsigned long start = address; | |
4264 | pte_t *ptep; | |
4265 | pte_t pte; | |
a5516438 | 4266 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4267 | unsigned long pages = 0; |
8f860591 ZY |
4268 | |
4269 | BUG_ON(address >= end); | |
4270 | flush_cache_range(vma, address, end); | |
4271 | ||
a5338093 | 4272 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 4273 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4274 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4275 | spinlock_t *ptl; |
7868a208 | 4276 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4277 | if (!ptep) |
4278 | continue; | |
cb900f41 | 4279 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4280 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4281 | pages++; | |
cb900f41 | 4282 | spin_unlock(ptl); |
39dde65c | 4283 | continue; |
7da4d641 | 4284 | } |
a8bda28d NH |
4285 | pte = huge_ptep_get(ptep); |
4286 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4287 | spin_unlock(ptl); | |
4288 | continue; | |
4289 | } | |
4290 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4291 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4292 | ||
4293 | if (is_write_migration_entry(entry)) { | |
4294 | pte_t newpte; | |
4295 | ||
4296 | make_migration_entry_read(&entry); | |
4297 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4298 | set_huge_swap_pte_at(mm, address, ptep, |
4299 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4300 | pages++; |
4301 | } | |
4302 | spin_unlock(ptl); | |
4303 | continue; | |
4304 | } | |
4305 | if (!huge_pte_none(pte)) { | |
8f860591 | 4306 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 4307 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 4308 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 4309 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 4310 | pages++; |
8f860591 | 4311 | } |
cb900f41 | 4312 | spin_unlock(ptl); |
8f860591 | 4313 | } |
d833352a | 4314 | /* |
c8c06efa | 4315 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4316 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4317 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
4318 | * and that page table be reused and filled with junk. |
4319 | */ | |
5491ae7b | 4320 | flush_hugetlb_tlb_range(vma, start, end); |
34ee645e | 4321 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 4322 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 4323 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
4324 | |
4325 | return pages << h->order; | |
8f860591 ZY |
4326 | } |
4327 | ||
a1e78772 MG |
4328 | int hugetlb_reserve_pages(struct inode *inode, |
4329 | long from, long to, | |
5a6fe125 | 4330 | struct vm_area_struct *vma, |
ca16d140 | 4331 | vm_flags_t vm_flags) |
e4e574b7 | 4332 | { |
17c9d12e | 4333 | long ret, chg; |
a5516438 | 4334 | struct hstate *h = hstate_inode(inode); |
90481622 | 4335 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4336 | struct resv_map *resv_map; |
1c5ecae3 | 4337 | long gbl_reserve; |
e4e574b7 | 4338 | |
17c9d12e MG |
4339 | /* |
4340 | * Only apply hugepage reservation if asked. At fault time, an | |
4341 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4342 | * without using reserves |
17c9d12e | 4343 | */ |
ca16d140 | 4344 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4345 | return 0; |
4346 | ||
a1e78772 MG |
4347 | /* |
4348 | * Shared mappings base their reservation on the number of pages that | |
4349 | * are already allocated on behalf of the file. Private mappings need | |
4350 | * to reserve the full area even if read-only as mprotect() may be | |
4351 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4352 | */ | |
9119a41e | 4353 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 4354 | resv_map = inode_resv_map(inode); |
9119a41e | 4355 | |
1406ec9b | 4356 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4357 | |
4358 | } else { | |
4359 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4360 | if (!resv_map) |
4361 | return -ENOMEM; | |
4362 | ||
a1e78772 | 4363 | chg = to - from; |
84afd99b | 4364 | |
17c9d12e MG |
4365 | set_vma_resv_map(vma, resv_map); |
4366 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4367 | } | |
4368 | ||
c50ac050 DH |
4369 | if (chg < 0) { |
4370 | ret = chg; | |
4371 | goto out_err; | |
4372 | } | |
8a630112 | 4373 | |
1c5ecae3 MK |
4374 | /* |
4375 | * There must be enough pages in the subpool for the mapping. If | |
4376 | * the subpool has a minimum size, there may be some global | |
4377 | * reservations already in place (gbl_reserve). | |
4378 | */ | |
4379 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4380 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4381 | ret = -ENOSPC; |
4382 | goto out_err; | |
4383 | } | |
5a6fe125 MG |
4384 | |
4385 | /* | |
17c9d12e | 4386 | * Check enough hugepages are available for the reservation. |
90481622 | 4387 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4388 | */ |
1c5ecae3 | 4389 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4390 | if (ret < 0) { |
1c5ecae3 MK |
4391 | /* put back original number of pages, chg */ |
4392 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4393 | goto out_err; |
68842c9b | 4394 | } |
17c9d12e MG |
4395 | |
4396 | /* | |
4397 | * Account for the reservations made. Shared mappings record regions | |
4398 | * that have reservations as they are shared by multiple VMAs. | |
4399 | * When the last VMA disappears, the region map says how much | |
4400 | * the reservation was and the page cache tells how much of | |
4401 | * the reservation was consumed. Private mappings are per-VMA and | |
4402 | * only the consumed reservations are tracked. When the VMA | |
4403 | * disappears, the original reservation is the VMA size and the | |
4404 | * consumed reservations are stored in the map. Hence, nothing | |
4405 | * else has to be done for private mappings here | |
4406 | */ | |
33039678 MK |
4407 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4408 | long add = region_add(resv_map, from, to); | |
4409 | ||
4410 | if (unlikely(chg > add)) { | |
4411 | /* | |
4412 | * pages in this range were added to the reserve | |
4413 | * map between region_chg and region_add. This | |
4414 | * indicates a race with alloc_huge_page. Adjust | |
4415 | * the subpool and reserve counts modified above | |
4416 | * based on the difference. | |
4417 | */ | |
4418 | long rsv_adjust; | |
4419 | ||
4420 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4421 | chg - add); | |
4422 | hugetlb_acct_memory(h, -rsv_adjust); | |
4423 | } | |
4424 | } | |
a43a8c39 | 4425 | return 0; |
c50ac050 | 4426 | out_err: |
5e911373 | 4427 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4428 | /* Don't call region_abort if region_chg failed */ |
4429 | if (chg >= 0) | |
4430 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4431 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4432 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4433 | return ret; |
a43a8c39 KC |
4434 | } |
4435 | ||
b5cec28d MK |
4436 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4437 | long freed) | |
a43a8c39 | 4438 | { |
a5516438 | 4439 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4440 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4441 | long chg = 0; |
90481622 | 4442 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4443 | long gbl_reserve; |
45c682a6 | 4444 | |
b5cec28d MK |
4445 | if (resv_map) { |
4446 | chg = region_del(resv_map, start, end); | |
4447 | /* | |
4448 | * region_del() can fail in the rare case where a region | |
4449 | * must be split and another region descriptor can not be | |
4450 | * allocated. If end == LONG_MAX, it will not fail. | |
4451 | */ | |
4452 | if (chg < 0) | |
4453 | return chg; | |
4454 | } | |
4455 | ||
45c682a6 | 4456 | spin_lock(&inode->i_lock); |
e4c6f8be | 4457 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4458 | spin_unlock(&inode->i_lock); |
4459 | ||
1c5ecae3 MK |
4460 | /* |
4461 | * If the subpool has a minimum size, the number of global | |
4462 | * reservations to be released may be adjusted. | |
4463 | */ | |
4464 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4465 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4466 | |
4467 | return 0; | |
a43a8c39 | 4468 | } |
93f70f90 | 4469 | |
3212b535 SC |
4470 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4471 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4472 | struct vm_area_struct *vma, | |
4473 | unsigned long addr, pgoff_t idx) | |
4474 | { | |
4475 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4476 | svma->vm_start; | |
4477 | unsigned long sbase = saddr & PUD_MASK; | |
4478 | unsigned long s_end = sbase + PUD_SIZE; | |
4479 | ||
4480 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4481 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4482 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4483 | |
4484 | /* | |
4485 | * match the virtual addresses, permission and the alignment of the | |
4486 | * page table page. | |
4487 | */ | |
4488 | if (pmd_index(addr) != pmd_index(saddr) || | |
4489 | vm_flags != svm_flags || | |
4490 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4491 | return 0; | |
4492 | ||
4493 | return saddr; | |
4494 | } | |
4495 | ||
31aafb45 | 4496 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4497 | { |
4498 | unsigned long base = addr & PUD_MASK; | |
4499 | unsigned long end = base + PUD_SIZE; | |
4500 | ||
4501 | /* | |
4502 | * check on proper vm_flags and page table alignment | |
4503 | */ | |
4504 | if (vma->vm_flags & VM_MAYSHARE && | |
4505 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4506 | return true; |
4507 | return false; | |
3212b535 SC |
4508 | } |
4509 | ||
4510 | /* | |
4511 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4512 | * and returns the corresponding pte. While this is not necessary for the | |
4513 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4514 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4515 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4516 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4517 | * bad pmd for sharing. | |
4518 | */ | |
4519 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4520 | { | |
4521 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4522 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4523 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4524 | vma->vm_pgoff; | |
4525 | struct vm_area_struct *svma; | |
4526 | unsigned long saddr; | |
4527 | pte_t *spte = NULL; | |
4528 | pte_t *pte; | |
cb900f41 | 4529 | spinlock_t *ptl; |
3212b535 SC |
4530 | |
4531 | if (!vma_shareable(vma, addr)) | |
4532 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4533 | ||
83cde9e8 | 4534 | i_mmap_lock_write(mapping); |
3212b535 SC |
4535 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4536 | if (svma == vma) | |
4537 | continue; | |
4538 | ||
4539 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4540 | if (saddr) { | |
7868a208 PA |
4541 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4542 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4543 | if (spte) { |
4544 | get_page(virt_to_page(spte)); | |
4545 | break; | |
4546 | } | |
4547 | } | |
4548 | } | |
4549 | ||
4550 | if (!spte) | |
4551 | goto out; | |
4552 | ||
8bea8052 | 4553 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4554 | if (pud_none(*pud)) { |
3212b535 SC |
4555 | pud_populate(mm, pud, |
4556 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4557 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4558 | } else { |
3212b535 | 4559 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4560 | } |
cb900f41 | 4561 | spin_unlock(ptl); |
3212b535 SC |
4562 | out: |
4563 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4564 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4565 | return pte; |
4566 | } | |
4567 | ||
4568 | /* | |
4569 | * unmap huge page backed by shared pte. | |
4570 | * | |
4571 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4572 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4573 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4574 | * | |
cb900f41 | 4575 | * called with page table lock held. |
3212b535 SC |
4576 | * |
4577 | * returns: 1 successfully unmapped a shared pte page | |
4578 | * 0 the underlying pte page is not shared, or it is the last user | |
4579 | */ | |
4580 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4581 | { | |
4582 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4583 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4584 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4585 | |
4586 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4587 | if (page_count(virt_to_page(ptep)) == 1) | |
4588 | return 0; | |
4589 | ||
4590 | pud_clear(pud); | |
4591 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4592 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4593 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4594 | return 1; | |
4595 | } | |
9e5fc74c SC |
4596 | #define want_pmd_share() (1) |
4597 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4598 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4599 | { | |
4600 | return NULL; | |
4601 | } | |
e81f2d22 ZZ |
4602 | |
4603 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4604 | { | |
4605 | return 0; | |
4606 | } | |
9e5fc74c | 4607 | #define want_pmd_share() (0) |
3212b535 SC |
4608 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4609 | ||
9e5fc74c SC |
4610 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4611 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4612 | unsigned long addr, unsigned long sz) | |
4613 | { | |
4614 | pgd_t *pgd; | |
c2febafc | 4615 | p4d_t *p4d; |
9e5fc74c SC |
4616 | pud_t *pud; |
4617 | pte_t *pte = NULL; | |
4618 | ||
4619 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4620 | p4d = p4d_offset(pgd, addr); |
4621 | pud = pud_alloc(mm, p4d, addr); | |
9e5fc74c SC |
4622 | if (pud) { |
4623 | if (sz == PUD_SIZE) { | |
4624 | pte = (pte_t *)pud; | |
4625 | } else { | |
4626 | BUG_ON(sz != PMD_SIZE); | |
4627 | if (want_pmd_share() && pud_none(*pud)) | |
4628 | pte = huge_pmd_share(mm, addr, pud); | |
4629 | else | |
4630 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4631 | } | |
4632 | } | |
4e666314 | 4633 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
4634 | |
4635 | return pte; | |
4636 | } | |
4637 | ||
9b19df29 PA |
4638 | /* |
4639 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
4640 | * entry at address @addr | |
4641 | * | |
4642 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
4643 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
4644 | * size @sz doesn't match the hugepage size at this level of the page | |
4645 | * table. | |
4646 | */ | |
7868a208 PA |
4647 | pte_t *huge_pte_offset(struct mm_struct *mm, |
4648 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
4649 | { |
4650 | pgd_t *pgd; | |
c2febafc | 4651 | p4d_t *p4d; |
9e5fc74c | 4652 | pud_t *pud; |
c2febafc | 4653 | pmd_t *pmd; |
9e5fc74c SC |
4654 | |
4655 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4656 | if (!pgd_present(*pgd)) |
4657 | return NULL; | |
4658 | p4d = p4d_offset(pgd, addr); | |
4659 | if (!p4d_present(*p4d)) | |
4660 | return NULL; | |
9b19df29 | 4661 | |
c2febafc | 4662 | pud = pud_offset(p4d, addr); |
9b19df29 | 4663 | if (sz != PUD_SIZE && pud_none(*pud)) |
c2febafc | 4664 | return NULL; |
9b19df29 PA |
4665 | /* hugepage or swap? */ |
4666 | if (pud_huge(*pud) || !pud_present(*pud)) | |
c2febafc | 4667 | return (pte_t *)pud; |
9b19df29 | 4668 | |
c2febafc | 4669 | pmd = pmd_offset(pud, addr); |
9b19df29 PA |
4670 | if (sz != PMD_SIZE && pmd_none(*pmd)) |
4671 | return NULL; | |
4672 | /* hugepage or swap? */ | |
4673 | if (pmd_huge(*pmd) || !pmd_present(*pmd)) | |
4674 | return (pte_t *)pmd; | |
4675 | ||
4676 | return NULL; | |
9e5fc74c SC |
4677 | } |
4678 | ||
61f77eda NH |
4679 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4680 | ||
4681 | /* | |
4682 | * These functions are overwritable if your architecture needs its own | |
4683 | * behavior. | |
4684 | */ | |
4685 | struct page * __weak | |
4686 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4687 | int write) | |
4688 | { | |
4689 | return ERR_PTR(-EINVAL); | |
4690 | } | |
4691 | ||
4dc71451 AK |
4692 | struct page * __weak |
4693 | follow_huge_pd(struct vm_area_struct *vma, | |
4694 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
4695 | { | |
4696 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
4697 | return NULL; | |
4698 | } | |
4699 | ||
61f77eda | 4700 | struct page * __weak |
9e5fc74c | 4701 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4702 | pmd_t *pmd, int flags) |
9e5fc74c | 4703 | { |
e66f17ff NH |
4704 | struct page *page = NULL; |
4705 | spinlock_t *ptl; | |
c9d398fa | 4706 | pte_t pte; |
e66f17ff NH |
4707 | retry: |
4708 | ptl = pmd_lockptr(mm, pmd); | |
4709 | spin_lock(ptl); | |
4710 | /* | |
4711 | * make sure that the address range covered by this pmd is not | |
4712 | * unmapped from other threads. | |
4713 | */ | |
4714 | if (!pmd_huge(*pmd)) | |
4715 | goto out; | |
c9d398fa NH |
4716 | pte = huge_ptep_get((pte_t *)pmd); |
4717 | if (pte_present(pte)) { | |
97534127 | 4718 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4719 | if (flags & FOLL_GET) |
4720 | get_page(page); | |
4721 | } else { | |
c9d398fa | 4722 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
4723 | spin_unlock(ptl); |
4724 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4725 | goto retry; | |
4726 | } | |
4727 | /* | |
4728 | * hwpoisoned entry is treated as no_page_table in | |
4729 | * follow_page_mask(). | |
4730 | */ | |
4731 | } | |
4732 | out: | |
4733 | spin_unlock(ptl); | |
9e5fc74c SC |
4734 | return page; |
4735 | } | |
4736 | ||
61f77eda | 4737 | struct page * __weak |
9e5fc74c | 4738 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4739 | pud_t *pud, int flags) |
9e5fc74c | 4740 | { |
e66f17ff NH |
4741 | if (flags & FOLL_GET) |
4742 | return NULL; | |
9e5fc74c | 4743 | |
e66f17ff | 4744 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4745 | } |
4746 | ||
faaa5b62 AK |
4747 | struct page * __weak |
4748 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
4749 | { | |
4750 | if (flags & FOLL_GET) | |
4751 | return NULL; | |
4752 | ||
4753 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
4754 | } | |
4755 | ||
31caf665 NH |
4756 | bool isolate_huge_page(struct page *page, struct list_head *list) |
4757 | { | |
bcc54222 NH |
4758 | bool ret = true; |
4759 | ||
309381fe | 4760 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4761 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4762 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4763 | ret = false; | |
4764 | goto unlock; | |
4765 | } | |
4766 | clear_page_huge_active(page); | |
31caf665 | 4767 | list_move_tail(&page->lru, list); |
bcc54222 | 4768 | unlock: |
31caf665 | 4769 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4770 | return ret; |
31caf665 NH |
4771 | } |
4772 | ||
4773 | void putback_active_hugepage(struct page *page) | |
4774 | { | |
309381fe | 4775 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4776 | spin_lock(&hugetlb_lock); |
bcc54222 | 4777 | set_page_huge_active(page); |
31caf665 NH |
4778 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4779 | spin_unlock(&hugetlb_lock); | |
4780 | put_page(page); | |
4781 | } |