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> |
0fe6e20b | 21 | #include <linux/rmap.h> |
fd6a03ed NH |
22 | #include <linux/swap.h> |
23 | #include <linux/swapops.h> | |
c8721bbb | 24 | #include <linux/page-isolation.h> |
8382d914 | 25 | #include <linux/jhash.h> |
d6606683 | 26 | |
63551ae0 DG |
27 | #include <asm/page.h> |
28 | #include <asm/pgtable.h> | |
24669e58 | 29 | #include <asm/tlb.h> |
63551ae0 | 30 | |
24669e58 | 31 | #include <linux/io.h> |
63551ae0 | 32 | #include <linux/hugetlb.h> |
9dd540e2 | 33 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 34 | #include <linux/node.h> |
7835e98b | 35 | #include "internal.h" |
1da177e4 | 36 | |
753162cd | 37 | int hugepages_treat_as_movable; |
a5516438 | 38 | |
c3f38a38 | 39 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
40 | unsigned int default_hstate_idx; |
41 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
42 | /* |
43 | * Minimum page order among possible hugepage sizes, set to a proper value | |
44 | * at boot time. | |
45 | */ | |
46 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 47 | |
53ba51d2 JT |
48 | __initdata LIST_HEAD(huge_boot_pages); |
49 | ||
e5ff2159 AK |
50 | /* for command line parsing */ |
51 | static struct hstate * __initdata parsed_hstate; | |
52 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 53 | static unsigned long __initdata default_hstate_size; |
9fee021d | 54 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 55 | |
3935baa9 | 56 | /* |
31caf665 NH |
57 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
58 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 59 | */ |
c3f38a38 | 60 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 61 | |
8382d914 DB |
62 | /* |
63 | * Serializes faults on the same logical page. This is used to | |
64 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
65 | */ | |
66 | static int num_fault_mutexes; | |
c672c7f2 | 67 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 68 | |
7ca02d0a MK |
69 | /* Forward declaration */ |
70 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
71 | ||
90481622 DG |
72 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
73 | { | |
74 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
75 | ||
76 | spin_unlock(&spool->lock); | |
77 | ||
78 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
79 | * remain, give up any reservations mased on minimum size and |
80 | * free the subpool */ | |
81 | if (free) { | |
82 | if (spool->min_hpages != -1) | |
83 | hugetlb_acct_memory(spool->hstate, | |
84 | -spool->min_hpages); | |
90481622 | 85 | kfree(spool); |
7ca02d0a | 86 | } |
90481622 DG |
87 | } |
88 | ||
7ca02d0a MK |
89 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
90 | long min_hpages) | |
90481622 DG |
91 | { |
92 | struct hugepage_subpool *spool; | |
93 | ||
c6a91820 | 94 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
95 | if (!spool) |
96 | return NULL; | |
97 | ||
98 | spin_lock_init(&spool->lock); | |
99 | spool->count = 1; | |
7ca02d0a MK |
100 | spool->max_hpages = max_hpages; |
101 | spool->hstate = h; | |
102 | spool->min_hpages = min_hpages; | |
103 | ||
104 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
105 | kfree(spool); | |
106 | return NULL; | |
107 | } | |
108 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
109 | |
110 | return spool; | |
111 | } | |
112 | ||
113 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
114 | { | |
115 | spin_lock(&spool->lock); | |
116 | BUG_ON(!spool->count); | |
117 | spool->count--; | |
118 | unlock_or_release_subpool(spool); | |
119 | } | |
120 | ||
1c5ecae3 MK |
121 | /* |
122 | * Subpool accounting for allocating and reserving pages. | |
123 | * Return -ENOMEM if there are not enough resources to satisfy the | |
124 | * the request. Otherwise, return the number of pages by which the | |
125 | * global pools must be adjusted (upward). The returned value may | |
126 | * only be different than the passed value (delta) in the case where | |
127 | * a subpool minimum size must be manitained. | |
128 | */ | |
129 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
130 | long delta) |
131 | { | |
1c5ecae3 | 132 | long ret = delta; |
90481622 DG |
133 | |
134 | if (!spool) | |
1c5ecae3 | 135 | return ret; |
90481622 DG |
136 | |
137 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
138 | |
139 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
140 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
141 | spool->used_hpages += delta; | |
142 | else { | |
143 | ret = -ENOMEM; | |
144 | goto unlock_ret; | |
145 | } | |
90481622 | 146 | } |
90481622 | 147 | |
09a95e29 MK |
148 | /* minimum size accounting */ |
149 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
150 | if (delta > spool->rsv_hpages) { |
151 | /* | |
152 | * Asking for more reserves than those already taken on | |
153 | * behalf of subpool. Return difference. | |
154 | */ | |
155 | ret = delta - spool->rsv_hpages; | |
156 | spool->rsv_hpages = 0; | |
157 | } else { | |
158 | ret = 0; /* reserves already accounted for */ | |
159 | spool->rsv_hpages -= delta; | |
160 | } | |
161 | } | |
162 | ||
163 | unlock_ret: | |
164 | spin_unlock(&spool->lock); | |
90481622 DG |
165 | return ret; |
166 | } | |
167 | ||
1c5ecae3 MK |
168 | /* |
169 | * Subpool accounting for freeing and unreserving pages. | |
170 | * Return the number of global page reservations that must be dropped. | |
171 | * The return value may only be different than the passed value (delta) | |
172 | * in the case where a subpool minimum size must be maintained. | |
173 | */ | |
174 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
175 | long delta) |
176 | { | |
1c5ecae3 MK |
177 | long ret = delta; |
178 | ||
90481622 | 179 | if (!spool) |
1c5ecae3 | 180 | return delta; |
90481622 DG |
181 | |
182 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
183 | |
184 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
185 | spool->used_hpages -= delta; | |
186 | ||
09a95e29 MK |
187 | /* minimum size accounting */ |
188 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
189 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
190 | ret = 0; | |
191 | else | |
192 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
193 | ||
194 | spool->rsv_hpages += delta; | |
195 | if (spool->rsv_hpages > spool->min_hpages) | |
196 | spool->rsv_hpages = spool->min_hpages; | |
197 | } | |
198 | ||
199 | /* | |
200 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
201 | * quota reference, free it now. | |
202 | */ | |
90481622 | 203 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
204 | |
205 | return ret; | |
90481622 DG |
206 | } |
207 | ||
208 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
209 | { | |
210 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
211 | } | |
212 | ||
213 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
214 | { | |
496ad9aa | 215 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
216 | } |
217 | ||
96822904 AW |
218 | /* |
219 | * Region tracking -- allows tracking of reservations and instantiated pages | |
220 | * across the pages in a mapping. | |
84afd99b | 221 | * |
1dd308a7 MK |
222 | * The region data structures are embedded into a resv_map and protected |
223 | * by a resv_map's lock. The set of regions within the resv_map represent | |
224 | * reservations for huge pages, or huge pages that have already been | |
225 | * instantiated within the map. The from and to elements are huge page | |
226 | * indicies into the associated mapping. from indicates the starting index | |
227 | * of the region. to represents the first index past the end of the region. | |
228 | * | |
229 | * For example, a file region structure with from == 0 and to == 4 represents | |
230 | * four huge pages in a mapping. It is important to note that the to element | |
231 | * represents the first element past the end of the region. This is used in | |
232 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
233 | * | |
234 | * Interval notation of the form [from, to) will be used to indicate that | |
235 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
236 | */ |
237 | struct file_region { | |
238 | struct list_head link; | |
239 | long from; | |
240 | long to; | |
241 | }; | |
242 | ||
1dd308a7 MK |
243 | /* |
244 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
245 | * map. In the normal case, existing regions will be expanded |
246 | * to accommodate the specified range. Sufficient regions should | |
247 | * exist for expansion due to the previous call to region_chg | |
248 | * with the same range. However, it is possible that region_del | |
249 | * could have been called after region_chg and modifed the map | |
250 | * in such a way that no region exists to be expanded. In this | |
251 | * case, pull a region descriptor from the cache associated with | |
252 | * the map and use that for the new range. | |
cf3ad20b MK |
253 | * |
254 | * Return the number of new huge pages added to the map. This | |
255 | * number is greater than or equal to zero. | |
1dd308a7 | 256 | */ |
1406ec9b | 257 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 258 | { |
1406ec9b | 259 | struct list_head *head = &resv->regions; |
96822904 | 260 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 261 | long add = 0; |
96822904 | 262 | |
7b24d861 | 263 | spin_lock(&resv->lock); |
96822904 AW |
264 | /* Locate the region we are either in or before. */ |
265 | list_for_each_entry(rg, head, link) | |
266 | if (f <= rg->to) | |
267 | break; | |
268 | ||
5e911373 MK |
269 | /* |
270 | * If no region exists which can be expanded to include the | |
271 | * specified range, the list must have been modified by an | |
272 | * interleving call to region_del(). Pull a region descriptor | |
273 | * from the cache and use it for this range. | |
274 | */ | |
275 | if (&rg->link == head || t < rg->from) { | |
276 | VM_BUG_ON(resv->region_cache_count <= 0); | |
277 | ||
278 | resv->region_cache_count--; | |
279 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
280 | link); | |
281 | list_del(&nrg->link); | |
282 | ||
283 | nrg->from = f; | |
284 | nrg->to = t; | |
285 | list_add(&nrg->link, rg->link.prev); | |
286 | ||
287 | add += t - f; | |
288 | goto out_locked; | |
289 | } | |
290 | ||
96822904 AW |
291 | /* Round our left edge to the current segment if it encloses us. */ |
292 | if (f > rg->from) | |
293 | f = rg->from; | |
294 | ||
295 | /* Check for and consume any regions we now overlap with. */ | |
296 | nrg = rg; | |
297 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
298 | if (&rg->link == head) | |
299 | break; | |
300 | if (rg->from > t) | |
301 | break; | |
302 | ||
303 | /* If this area reaches higher then extend our area to | |
304 | * include it completely. If this is not the first area | |
305 | * which we intend to reuse, free it. */ | |
306 | if (rg->to > t) | |
307 | t = rg->to; | |
308 | if (rg != nrg) { | |
cf3ad20b MK |
309 | /* Decrement return value by the deleted range. |
310 | * Another range will span this area so that by | |
311 | * end of routine add will be >= zero | |
312 | */ | |
313 | add -= (rg->to - rg->from); | |
96822904 AW |
314 | list_del(&rg->link); |
315 | kfree(rg); | |
316 | } | |
317 | } | |
cf3ad20b MK |
318 | |
319 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 320 | nrg->from = f; |
cf3ad20b | 321 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 322 | nrg->to = t; |
cf3ad20b | 323 | |
5e911373 MK |
324 | out_locked: |
325 | resv->adds_in_progress--; | |
7b24d861 | 326 | spin_unlock(&resv->lock); |
cf3ad20b MK |
327 | VM_BUG_ON(add < 0); |
328 | return add; | |
96822904 AW |
329 | } |
330 | ||
1dd308a7 MK |
331 | /* |
332 | * Examine the existing reserve map and determine how many | |
333 | * huge pages in the specified range [f, t) are NOT currently | |
334 | * represented. This routine is called before a subsequent | |
335 | * call to region_add that will actually modify the reserve | |
336 | * map to add the specified range [f, t). region_chg does | |
337 | * not change the number of huge pages represented by the | |
338 | * map. However, if the existing regions in the map can not | |
339 | * be expanded to represent the new range, a new file_region | |
340 | * structure is added to the map as a placeholder. This is | |
341 | * so that the subsequent region_add call will have all the | |
342 | * regions it needs and will not fail. | |
343 | * | |
5e911373 MK |
344 | * Upon entry, region_chg will also examine the cache of region descriptors |
345 | * associated with the map. If there are not enough descriptors cached, one | |
346 | * will be allocated for the in progress add operation. | |
347 | * | |
348 | * Returns the number of huge pages that need to be added to the existing | |
349 | * reservation map for the range [f, t). This number is greater or equal to | |
350 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
351 | * is needed and can not be allocated. | |
1dd308a7 | 352 | */ |
1406ec9b | 353 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 354 | { |
1406ec9b | 355 | struct list_head *head = &resv->regions; |
7b24d861 | 356 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
357 | long chg = 0; |
358 | ||
7b24d861 DB |
359 | retry: |
360 | spin_lock(&resv->lock); | |
5e911373 MK |
361 | retry_locked: |
362 | resv->adds_in_progress++; | |
363 | ||
364 | /* | |
365 | * Check for sufficient descriptors in the cache to accommodate | |
366 | * the number of in progress add operations. | |
367 | */ | |
368 | if (resv->adds_in_progress > resv->region_cache_count) { | |
369 | struct file_region *trg; | |
370 | ||
371 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
372 | /* Must drop lock to allocate a new descriptor. */ | |
373 | resv->adds_in_progress--; | |
374 | spin_unlock(&resv->lock); | |
375 | ||
376 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
dbe409e4 MK |
377 | if (!trg) { |
378 | kfree(nrg); | |
5e911373 | 379 | return -ENOMEM; |
dbe409e4 | 380 | } |
5e911373 MK |
381 | |
382 | spin_lock(&resv->lock); | |
383 | list_add(&trg->link, &resv->region_cache); | |
384 | resv->region_cache_count++; | |
385 | goto retry_locked; | |
386 | } | |
387 | ||
96822904 AW |
388 | /* Locate the region we are before or in. */ |
389 | list_for_each_entry(rg, head, link) | |
390 | if (f <= rg->to) | |
391 | break; | |
392 | ||
393 | /* If we are below the current region then a new region is required. | |
394 | * Subtle, allocate a new region at the position but make it zero | |
395 | * size such that we can guarantee to record the reservation. */ | |
396 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 397 | if (!nrg) { |
5e911373 | 398 | resv->adds_in_progress--; |
7b24d861 DB |
399 | spin_unlock(&resv->lock); |
400 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
401 | if (!nrg) | |
402 | return -ENOMEM; | |
403 | ||
404 | nrg->from = f; | |
405 | nrg->to = f; | |
406 | INIT_LIST_HEAD(&nrg->link); | |
407 | goto retry; | |
408 | } | |
96822904 | 409 | |
7b24d861 DB |
410 | list_add(&nrg->link, rg->link.prev); |
411 | chg = t - f; | |
412 | goto out_nrg; | |
96822904 AW |
413 | } |
414 | ||
415 | /* Round our left edge to the current segment if it encloses us. */ | |
416 | if (f > rg->from) | |
417 | f = rg->from; | |
418 | chg = t - f; | |
419 | ||
420 | /* Check for and consume any regions we now overlap with. */ | |
421 | list_for_each_entry(rg, rg->link.prev, link) { | |
422 | if (&rg->link == head) | |
423 | break; | |
424 | if (rg->from > t) | |
7b24d861 | 425 | goto out; |
96822904 | 426 | |
25985edc | 427 | /* We overlap with this area, if it extends further than |
96822904 AW |
428 | * us then we must extend ourselves. Account for its |
429 | * existing reservation. */ | |
430 | if (rg->to > t) { | |
431 | chg += rg->to - t; | |
432 | t = rg->to; | |
433 | } | |
434 | chg -= rg->to - rg->from; | |
435 | } | |
7b24d861 DB |
436 | |
437 | out: | |
438 | spin_unlock(&resv->lock); | |
439 | /* We already know we raced and no longer need the new region */ | |
440 | kfree(nrg); | |
441 | return chg; | |
442 | out_nrg: | |
443 | spin_unlock(&resv->lock); | |
96822904 AW |
444 | return chg; |
445 | } | |
446 | ||
5e911373 MK |
447 | /* |
448 | * Abort the in progress add operation. The adds_in_progress field | |
449 | * of the resv_map keeps track of the operations in progress between | |
450 | * calls to region_chg and region_add. Operations are sometimes | |
451 | * aborted after the call to region_chg. In such cases, region_abort | |
452 | * is called to decrement the adds_in_progress counter. | |
453 | * | |
454 | * NOTE: The range arguments [f, t) are not needed or used in this | |
455 | * routine. They are kept to make reading the calling code easier as | |
456 | * arguments will match the associated region_chg call. | |
457 | */ | |
458 | static void region_abort(struct resv_map *resv, long f, long t) | |
459 | { | |
460 | spin_lock(&resv->lock); | |
461 | VM_BUG_ON(!resv->region_cache_count); | |
462 | resv->adds_in_progress--; | |
463 | spin_unlock(&resv->lock); | |
464 | } | |
465 | ||
1dd308a7 | 466 | /* |
feba16e2 MK |
467 | * Delete the specified range [f, t) from the reserve map. If the |
468 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
469 | * should be deleted. Locate the regions which intersect [f, t) | |
470 | * and either trim, delete or split the existing regions. | |
471 | * | |
472 | * Returns the number of huge pages deleted from the reserve map. | |
473 | * In the normal case, the return value is zero or more. In the | |
474 | * case where a region must be split, a new region descriptor must | |
475 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
476 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
477 | * a region and possibly return -ENOMEM. Callers specifying | |
478 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 479 | */ |
feba16e2 | 480 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 481 | { |
1406ec9b | 482 | struct list_head *head = &resv->regions; |
96822904 | 483 | struct file_region *rg, *trg; |
feba16e2 MK |
484 | struct file_region *nrg = NULL; |
485 | long del = 0; | |
96822904 | 486 | |
feba16e2 | 487 | retry: |
7b24d861 | 488 | spin_lock(&resv->lock); |
feba16e2 | 489 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
490 | /* |
491 | * Skip regions before the range to be deleted. file_region | |
492 | * ranges are normally of the form [from, to). However, there | |
493 | * may be a "placeholder" entry in the map which is of the form | |
494 | * (from, to) with from == to. Check for placeholder entries | |
495 | * at the beginning of the range to be deleted. | |
496 | */ | |
497 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 498 | continue; |
dbe409e4 | 499 | |
feba16e2 | 500 | if (rg->from >= t) |
96822904 | 501 | break; |
96822904 | 502 | |
feba16e2 MK |
503 | if (f > rg->from && t < rg->to) { /* Must split region */ |
504 | /* | |
505 | * Check for an entry in the cache before dropping | |
506 | * lock and attempting allocation. | |
507 | */ | |
508 | if (!nrg && | |
509 | resv->region_cache_count > resv->adds_in_progress) { | |
510 | nrg = list_first_entry(&resv->region_cache, | |
511 | struct file_region, | |
512 | link); | |
513 | list_del(&nrg->link); | |
514 | resv->region_cache_count--; | |
515 | } | |
96822904 | 516 | |
feba16e2 MK |
517 | if (!nrg) { |
518 | spin_unlock(&resv->lock); | |
519 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
520 | if (!nrg) | |
521 | return -ENOMEM; | |
522 | goto retry; | |
523 | } | |
524 | ||
525 | del += t - f; | |
526 | ||
527 | /* New entry for end of split region */ | |
528 | nrg->from = t; | |
529 | nrg->to = rg->to; | |
530 | INIT_LIST_HEAD(&nrg->link); | |
531 | ||
532 | /* Original entry is trimmed */ | |
533 | rg->to = f; | |
534 | ||
535 | list_add(&nrg->link, &rg->link); | |
536 | nrg = NULL; | |
96822904 | 537 | break; |
feba16e2 MK |
538 | } |
539 | ||
540 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
541 | del += rg->to - rg->from; | |
542 | list_del(&rg->link); | |
543 | kfree(rg); | |
544 | continue; | |
545 | } | |
546 | ||
547 | if (f <= rg->from) { /* Trim beginning of region */ | |
548 | del += t - rg->from; | |
549 | rg->from = t; | |
550 | } else { /* Trim end of region */ | |
551 | del += rg->to - f; | |
552 | rg->to = f; | |
553 | } | |
96822904 | 554 | } |
7b24d861 | 555 | |
7b24d861 | 556 | spin_unlock(&resv->lock); |
feba16e2 MK |
557 | kfree(nrg); |
558 | return del; | |
96822904 AW |
559 | } |
560 | ||
b5cec28d MK |
561 | /* |
562 | * A rare out of memory error was encountered which prevented removal of | |
563 | * the reserve map region for a page. The huge page itself was free'ed | |
564 | * and removed from the page cache. This routine will adjust the subpool | |
565 | * usage count, and the global reserve count if needed. By incrementing | |
566 | * these counts, the reserve map entry which could not be deleted will | |
567 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
568 | * counts. | |
569 | */ | |
570 | void hugetlb_fix_reserve_counts(struct inode *inode, bool restore_reserve) | |
571 | { | |
572 | struct hugepage_subpool *spool = subpool_inode(inode); | |
573 | long rsv_adjust; | |
574 | ||
575 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
576 | if (restore_reserve && rsv_adjust) { | |
577 | struct hstate *h = hstate_inode(inode); | |
578 | ||
579 | hugetlb_acct_memory(h, 1); | |
580 | } | |
581 | } | |
582 | ||
1dd308a7 MK |
583 | /* |
584 | * Count and return the number of huge pages in the reserve map | |
585 | * that intersect with the range [f, t). | |
586 | */ | |
1406ec9b | 587 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 588 | { |
1406ec9b | 589 | struct list_head *head = &resv->regions; |
84afd99b AW |
590 | struct file_region *rg; |
591 | long chg = 0; | |
592 | ||
7b24d861 | 593 | spin_lock(&resv->lock); |
84afd99b AW |
594 | /* Locate each segment we overlap with, and count that overlap. */ |
595 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
596 | long seg_from; |
597 | long seg_to; | |
84afd99b AW |
598 | |
599 | if (rg->to <= f) | |
600 | continue; | |
601 | if (rg->from >= t) | |
602 | break; | |
603 | ||
604 | seg_from = max(rg->from, f); | |
605 | seg_to = min(rg->to, t); | |
606 | ||
607 | chg += seg_to - seg_from; | |
608 | } | |
7b24d861 | 609 | spin_unlock(&resv->lock); |
84afd99b AW |
610 | |
611 | return chg; | |
612 | } | |
613 | ||
e7c4b0bf AW |
614 | /* |
615 | * Convert the address within this vma to the page offset within | |
616 | * the mapping, in pagecache page units; huge pages here. | |
617 | */ | |
a5516438 AK |
618 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
619 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 620 | { |
a5516438 AK |
621 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
622 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
623 | } |
624 | ||
0fe6e20b NH |
625 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
626 | unsigned long address) | |
627 | { | |
628 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
629 | } | |
dee41079 | 630 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 631 | |
08fba699 MG |
632 | /* |
633 | * Return the size of the pages allocated when backing a VMA. In the majority | |
634 | * cases this will be same size as used by the page table entries. | |
635 | */ | |
636 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
637 | { | |
638 | struct hstate *hstate; | |
639 | ||
640 | if (!is_vm_hugetlb_page(vma)) | |
641 | return PAGE_SIZE; | |
642 | ||
643 | hstate = hstate_vma(vma); | |
644 | ||
2415cf12 | 645 | return 1UL << huge_page_shift(hstate); |
08fba699 | 646 | } |
f340ca0f | 647 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 648 | |
3340289d MG |
649 | /* |
650 | * Return the page size being used by the MMU to back a VMA. In the majority | |
651 | * of cases, the page size used by the kernel matches the MMU size. On | |
652 | * architectures where it differs, an architecture-specific version of this | |
653 | * function is required. | |
654 | */ | |
655 | #ifndef vma_mmu_pagesize | |
656 | unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) | |
657 | { | |
658 | return vma_kernel_pagesize(vma); | |
659 | } | |
660 | #endif | |
661 | ||
84afd99b AW |
662 | /* |
663 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
664 | * bits of the reservation map pointer, which are always clear due to | |
665 | * alignment. | |
666 | */ | |
667 | #define HPAGE_RESV_OWNER (1UL << 0) | |
668 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 669 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 670 | |
a1e78772 MG |
671 | /* |
672 | * These helpers are used to track how many pages are reserved for | |
673 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
674 | * is guaranteed to have their future faults succeed. | |
675 | * | |
676 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
677 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
678 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
679 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
680 | * |
681 | * The private mapping reservation is represented in a subtly different | |
682 | * manner to a shared mapping. A shared mapping has a region map associated | |
683 | * with the underlying file, this region map represents the backing file | |
684 | * pages which have ever had a reservation assigned which this persists even | |
685 | * after the page is instantiated. A private mapping has a region map | |
686 | * associated with the original mmap which is attached to all VMAs which | |
687 | * reference it, this region map represents those offsets which have consumed | |
688 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 689 | */ |
e7c4b0bf AW |
690 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
691 | { | |
692 | return (unsigned long)vma->vm_private_data; | |
693 | } | |
694 | ||
695 | static void set_vma_private_data(struct vm_area_struct *vma, | |
696 | unsigned long value) | |
697 | { | |
698 | vma->vm_private_data = (void *)value; | |
699 | } | |
700 | ||
9119a41e | 701 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
702 | { |
703 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
704 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
705 | ||
706 | if (!resv_map || !rg) { | |
707 | kfree(resv_map); | |
708 | kfree(rg); | |
84afd99b | 709 | return NULL; |
5e911373 | 710 | } |
84afd99b AW |
711 | |
712 | kref_init(&resv_map->refs); | |
7b24d861 | 713 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
714 | INIT_LIST_HEAD(&resv_map->regions); |
715 | ||
5e911373 MK |
716 | resv_map->adds_in_progress = 0; |
717 | ||
718 | INIT_LIST_HEAD(&resv_map->region_cache); | |
719 | list_add(&rg->link, &resv_map->region_cache); | |
720 | resv_map->region_cache_count = 1; | |
721 | ||
84afd99b AW |
722 | return resv_map; |
723 | } | |
724 | ||
9119a41e | 725 | void resv_map_release(struct kref *ref) |
84afd99b AW |
726 | { |
727 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
728 | struct list_head *head = &resv_map->region_cache; |
729 | struct file_region *rg, *trg; | |
84afd99b AW |
730 | |
731 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 732 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
733 | |
734 | /* ... and any entries left in the cache */ | |
735 | list_for_each_entry_safe(rg, trg, head, link) { | |
736 | list_del(&rg->link); | |
737 | kfree(rg); | |
738 | } | |
739 | ||
740 | VM_BUG_ON(resv_map->adds_in_progress); | |
741 | ||
84afd99b AW |
742 | kfree(resv_map); |
743 | } | |
744 | ||
4e35f483 JK |
745 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
746 | { | |
747 | return inode->i_mapping->private_data; | |
748 | } | |
749 | ||
84afd99b | 750 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 751 | { |
81d1b09c | 752 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
753 | if (vma->vm_flags & VM_MAYSHARE) { |
754 | struct address_space *mapping = vma->vm_file->f_mapping; | |
755 | struct inode *inode = mapping->host; | |
756 | ||
757 | return inode_resv_map(inode); | |
758 | ||
759 | } else { | |
84afd99b AW |
760 | return (struct resv_map *)(get_vma_private_data(vma) & |
761 | ~HPAGE_RESV_MASK); | |
4e35f483 | 762 | } |
a1e78772 MG |
763 | } |
764 | ||
84afd99b | 765 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 766 | { |
81d1b09c SL |
767 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
768 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 769 | |
84afd99b AW |
770 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
771 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
772 | } |
773 | ||
774 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
775 | { | |
81d1b09c SL |
776 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
777 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
778 | |
779 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
780 | } |
781 | ||
782 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
783 | { | |
81d1b09c | 784 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
785 | |
786 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
787 | } |
788 | ||
04f2cbe3 | 789 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
790 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
791 | { | |
81d1b09c | 792 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 793 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
794 | vma->vm_private_data = (void *)0; |
795 | } | |
796 | ||
797 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 798 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 799 | { |
af0ed73e JK |
800 | if (vma->vm_flags & VM_NORESERVE) { |
801 | /* | |
802 | * This address is already reserved by other process(chg == 0), | |
803 | * so, we should decrement reserved count. Without decrementing, | |
804 | * reserve count remains after releasing inode, because this | |
805 | * allocated page will go into page cache and is regarded as | |
806 | * coming from reserved pool in releasing step. Currently, we | |
807 | * don't have any other solution to deal with this situation | |
808 | * properly, so add work-around here. | |
809 | */ | |
810 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 811 | return true; |
af0ed73e | 812 | else |
559ec2f8 | 813 | return false; |
af0ed73e | 814 | } |
a63884e9 JK |
815 | |
816 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
817 | if (vma->vm_flags & VM_MAYSHARE) { |
818 | /* | |
819 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
820 | * be a region map for all pages. The only situation where | |
821 | * there is no region map is if a hole was punched via | |
822 | * fallocate. In this case, there really are no reverves to | |
823 | * use. This situation is indicated if chg != 0. | |
824 | */ | |
825 | if (chg) | |
826 | return false; | |
827 | else | |
828 | return true; | |
829 | } | |
a63884e9 JK |
830 | |
831 | /* | |
832 | * Only the process that called mmap() has reserves for | |
833 | * private mappings. | |
834 | */ | |
7f09ca51 | 835 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
559ec2f8 | 836 | return true; |
a63884e9 | 837 | |
559ec2f8 | 838 | return false; |
a1e78772 MG |
839 | } |
840 | ||
a5516438 | 841 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
842 | { |
843 | int nid = page_to_nid(page); | |
0edaecfa | 844 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
845 | h->free_huge_pages++; |
846 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
847 | } |
848 | ||
bf50bab2 NH |
849 | static struct page *dequeue_huge_page_node(struct hstate *h, int nid) |
850 | { | |
851 | struct page *page; | |
852 | ||
c8721bbb NH |
853 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
854 | if (!is_migrate_isolate_page(page)) | |
855 | break; | |
856 | /* | |
857 | * if 'non-isolated free hugepage' not found on the list, | |
858 | * the allocation fails. | |
859 | */ | |
860 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 861 | return NULL; |
0edaecfa | 862 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 863 | set_page_refcounted(page); |
bf50bab2 NH |
864 | h->free_huge_pages--; |
865 | h->free_huge_pages_node[nid]--; | |
866 | return page; | |
867 | } | |
868 | ||
86cdb465 NH |
869 | /* Movability of hugepages depends on migration support. */ |
870 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
871 | { | |
100873d7 | 872 | if (hugepages_treat_as_movable || hugepage_migration_supported(h)) |
86cdb465 NH |
873 | return GFP_HIGHUSER_MOVABLE; |
874 | else | |
875 | return GFP_HIGHUSER; | |
876 | } | |
877 | ||
a5516438 AK |
878 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
879 | struct vm_area_struct *vma, | |
af0ed73e JK |
880 | unsigned long address, int avoid_reserve, |
881 | long chg) | |
1da177e4 | 882 | { |
b1c12cbc | 883 | struct page *page = NULL; |
480eccf9 | 884 | struct mempolicy *mpol; |
19770b32 | 885 | nodemask_t *nodemask; |
c0ff7453 | 886 | struct zonelist *zonelist; |
dd1a239f MG |
887 | struct zone *zone; |
888 | struct zoneref *z; | |
cc9a6c87 | 889 | unsigned int cpuset_mems_cookie; |
1da177e4 | 890 | |
a1e78772 MG |
891 | /* |
892 | * A child process with MAP_PRIVATE mappings created by their parent | |
893 | * have no page reserves. This check ensures that reservations are | |
894 | * not "stolen". The child may still get SIGKILLed | |
895 | */ | |
af0ed73e | 896 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 897 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 898 | goto err; |
a1e78772 | 899 | |
04f2cbe3 | 900 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 901 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 902 | goto err; |
04f2cbe3 | 903 | |
9966c4bb | 904 | retry_cpuset: |
d26914d1 | 905 | cpuset_mems_cookie = read_mems_allowed_begin(); |
9966c4bb | 906 | zonelist = huge_zonelist(vma, address, |
86cdb465 | 907 | htlb_alloc_mask(h), &mpol, &nodemask); |
9966c4bb | 908 | |
19770b32 MG |
909 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
910 | MAX_NR_ZONES - 1, nodemask) { | |
344736f2 | 911 | if (cpuset_zone_allowed(zone, htlb_alloc_mask(h))) { |
bf50bab2 NH |
912 | page = dequeue_huge_page_node(h, zone_to_nid(zone)); |
913 | if (page) { | |
af0ed73e JK |
914 | if (avoid_reserve) |
915 | break; | |
916 | if (!vma_has_reserves(vma, chg)) | |
917 | break; | |
918 | ||
07443a85 | 919 | SetPagePrivate(page); |
af0ed73e | 920 | h->resv_huge_pages--; |
bf50bab2 NH |
921 | break; |
922 | } | |
3abf7afd | 923 | } |
1da177e4 | 924 | } |
cc9a6c87 | 925 | |
52cd3b07 | 926 | mpol_cond_put(mpol); |
d26914d1 | 927 | if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie))) |
cc9a6c87 | 928 | goto retry_cpuset; |
1da177e4 | 929 | return page; |
cc9a6c87 MG |
930 | |
931 | err: | |
cc9a6c87 | 932 | return NULL; |
1da177e4 LT |
933 | } |
934 | ||
1cac6f2c LC |
935 | /* |
936 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
937 | * We may have allocated or freed a huge page based on a different | |
938 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
939 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
940 | * node for alloc or free. | |
941 | */ | |
942 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
943 | { | |
0edaf86c | 944 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
945 | VM_BUG_ON(nid >= MAX_NUMNODES); |
946 | ||
947 | return nid; | |
948 | } | |
949 | ||
950 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
951 | { | |
952 | if (!node_isset(nid, *nodes_allowed)) | |
953 | nid = next_node_allowed(nid, nodes_allowed); | |
954 | return nid; | |
955 | } | |
956 | ||
957 | /* | |
958 | * returns the previously saved node ["this node"] from which to | |
959 | * allocate a persistent huge page for the pool and advance the | |
960 | * next node from which to allocate, handling wrap at end of node | |
961 | * mask. | |
962 | */ | |
963 | static int hstate_next_node_to_alloc(struct hstate *h, | |
964 | nodemask_t *nodes_allowed) | |
965 | { | |
966 | int nid; | |
967 | ||
968 | VM_BUG_ON(!nodes_allowed); | |
969 | ||
970 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
971 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
972 | ||
973 | return nid; | |
974 | } | |
975 | ||
976 | /* | |
977 | * helper for free_pool_huge_page() - return the previously saved | |
978 | * node ["this node"] from which to free a huge page. Advance the | |
979 | * next node id whether or not we find a free huge page to free so | |
980 | * that the next attempt to free addresses the next node. | |
981 | */ | |
982 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
983 | { | |
984 | int nid; | |
985 | ||
986 | VM_BUG_ON(!nodes_allowed); | |
987 | ||
988 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
989 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
990 | ||
991 | return nid; | |
992 | } | |
993 | ||
994 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
995 | for (nr_nodes = nodes_weight(*mask); \ | |
996 | nr_nodes > 0 && \ | |
997 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
998 | nr_nodes--) | |
999 | ||
1000 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1001 | for (nr_nodes = nodes_weight(*mask); \ | |
1002 | nr_nodes > 0 && \ | |
1003 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1004 | nr_nodes--) | |
1005 | ||
080fe206 | 1006 | #if defined(CONFIG_X86_64) && ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)) |
944d9fec | 1007 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1008 | unsigned int order) |
944d9fec LC |
1009 | { |
1010 | int i; | |
1011 | int nr_pages = 1 << order; | |
1012 | struct page *p = page + 1; | |
1013 | ||
1014 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { | |
1d798ca3 | 1015 | clear_compound_head(p); |
944d9fec | 1016 | set_page_refcounted(p); |
944d9fec LC |
1017 | } |
1018 | ||
1019 | set_compound_order(page, 0); | |
1020 | __ClearPageHead(page); | |
1021 | } | |
1022 | ||
d00181b9 | 1023 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1024 | { |
1025 | free_contig_range(page_to_pfn(page), 1 << order); | |
1026 | } | |
1027 | ||
1028 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
1029 | unsigned long nr_pages) | |
1030 | { | |
1031 | unsigned long end_pfn = start_pfn + nr_pages; | |
1032 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE); | |
1033 | } | |
1034 | ||
f44b2dda JK |
1035 | static bool pfn_range_valid_gigantic(struct zone *z, |
1036 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1037 | { |
1038 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1039 | struct page *page; | |
1040 | ||
1041 | for (i = start_pfn; i < end_pfn; i++) { | |
1042 | if (!pfn_valid(i)) | |
1043 | return false; | |
1044 | ||
1045 | page = pfn_to_page(i); | |
1046 | ||
f44b2dda JK |
1047 | if (page_zone(page) != z) |
1048 | return false; | |
1049 | ||
944d9fec LC |
1050 | if (PageReserved(page)) |
1051 | return false; | |
1052 | ||
1053 | if (page_count(page) > 0) | |
1054 | return false; | |
1055 | ||
1056 | if (PageHuge(page)) | |
1057 | return false; | |
1058 | } | |
1059 | ||
1060 | return true; | |
1061 | } | |
1062 | ||
1063 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1064 | unsigned long start_pfn, unsigned long nr_pages) | |
1065 | { | |
1066 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1067 | return zone_spans_pfn(zone, last_pfn); | |
1068 | } | |
1069 | ||
d00181b9 | 1070 | static struct page *alloc_gigantic_page(int nid, unsigned int order) |
944d9fec LC |
1071 | { |
1072 | unsigned long nr_pages = 1 << order; | |
1073 | unsigned long ret, pfn, flags; | |
1074 | struct zone *z; | |
1075 | ||
1076 | z = NODE_DATA(nid)->node_zones; | |
1077 | for (; z - NODE_DATA(nid)->node_zones < MAX_NR_ZONES; z++) { | |
1078 | spin_lock_irqsave(&z->lock, flags); | |
1079 | ||
1080 | pfn = ALIGN(z->zone_start_pfn, nr_pages); | |
1081 | while (zone_spans_last_pfn(z, pfn, nr_pages)) { | |
f44b2dda | 1082 | if (pfn_range_valid_gigantic(z, pfn, nr_pages)) { |
944d9fec LC |
1083 | /* |
1084 | * We release the zone lock here because | |
1085 | * alloc_contig_range() will also lock the zone | |
1086 | * at some point. If there's an allocation | |
1087 | * spinning on this lock, it may win the race | |
1088 | * and cause alloc_contig_range() to fail... | |
1089 | */ | |
1090 | spin_unlock_irqrestore(&z->lock, flags); | |
1091 | ret = __alloc_gigantic_page(pfn, nr_pages); | |
1092 | if (!ret) | |
1093 | return pfn_to_page(pfn); | |
1094 | spin_lock_irqsave(&z->lock, flags); | |
1095 | } | |
1096 | pfn += nr_pages; | |
1097 | } | |
1098 | ||
1099 | spin_unlock_irqrestore(&z->lock, flags); | |
1100 | } | |
1101 | ||
1102 | return NULL; | |
1103 | } | |
1104 | ||
1105 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1106 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
944d9fec LC |
1107 | |
1108 | static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid) | |
1109 | { | |
1110 | struct page *page; | |
1111 | ||
1112 | page = alloc_gigantic_page(nid, huge_page_order(h)); | |
1113 | if (page) { | |
1114 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1115 | prep_new_huge_page(h, page, nid); | |
1116 | } | |
1117 | ||
1118 | return page; | |
1119 | } | |
1120 | ||
1121 | static int alloc_fresh_gigantic_page(struct hstate *h, | |
1122 | nodemask_t *nodes_allowed) | |
1123 | { | |
1124 | struct page *page = NULL; | |
1125 | int nr_nodes, node; | |
1126 | ||
1127 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1128 | page = alloc_fresh_gigantic_page_node(h, node); | |
1129 | if (page) | |
1130 | return 1; | |
1131 | } | |
1132 | ||
1133 | return 0; | |
1134 | } | |
1135 | ||
1136 | static inline bool gigantic_page_supported(void) { return true; } | |
1137 | #else | |
1138 | static inline bool gigantic_page_supported(void) { return false; } | |
d00181b9 | 1139 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1140 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1141 | unsigned int order) { } |
944d9fec LC |
1142 | static inline int alloc_fresh_gigantic_page(struct hstate *h, |
1143 | nodemask_t *nodes_allowed) { return 0; } | |
1144 | #endif | |
1145 | ||
a5516438 | 1146 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1147 | { |
1148 | int i; | |
a5516438 | 1149 | |
944d9fec LC |
1150 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1151 | return; | |
18229df5 | 1152 | |
a5516438 AK |
1153 | h->nr_huge_pages--; |
1154 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1155 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1156 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1157 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1158 | 1 << PG_active | 1 << PG_private | |
1159 | 1 << PG_writeback); | |
6af2acb6 | 1160 | } |
309381fe | 1161 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1162 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1163 | set_page_refcounted(page); |
944d9fec LC |
1164 | if (hstate_is_gigantic(h)) { |
1165 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1166 | free_gigantic_page(page, huge_page_order(h)); | |
1167 | } else { | |
944d9fec LC |
1168 | __free_pages(page, huge_page_order(h)); |
1169 | } | |
6af2acb6 AL |
1170 | } |
1171 | ||
e5ff2159 AK |
1172 | struct hstate *size_to_hstate(unsigned long size) |
1173 | { | |
1174 | struct hstate *h; | |
1175 | ||
1176 | for_each_hstate(h) { | |
1177 | if (huge_page_size(h) == size) | |
1178 | return h; | |
1179 | } | |
1180 | return NULL; | |
1181 | } | |
1182 | ||
bcc54222 NH |
1183 | /* |
1184 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1185 | * to hstate->hugepage_activelist.) | |
1186 | * | |
1187 | * This function can be called for tail pages, but never returns true for them. | |
1188 | */ | |
1189 | bool page_huge_active(struct page *page) | |
1190 | { | |
1191 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1192 | return PageHead(page) && PagePrivate(&page[1]); | |
1193 | } | |
1194 | ||
1195 | /* never called for tail page */ | |
1196 | static void set_page_huge_active(struct page *page) | |
1197 | { | |
1198 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1199 | SetPagePrivate(&page[1]); | |
1200 | } | |
1201 | ||
1202 | static void clear_page_huge_active(struct page *page) | |
1203 | { | |
1204 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1205 | ClearPagePrivate(&page[1]); | |
1206 | } | |
1207 | ||
8f1d26d0 | 1208 | void free_huge_page(struct page *page) |
27a85ef1 | 1209 | { |
a5516438 AK |
1210 | /* |
1211 | * Can't pass hstate in here because it is called from the | |
1212 | * compound page destructor. | |
1213 | */ | |
e5ff2159 | 1214 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1215 | int nid = page_to_nid(page); |
90481622 DG |
1216 | struct hugepage_subpool *spool = |
1217 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1218 | bool restore_reserve; |
27a85ef1 | 1219 | |
e5df70ab | 1220 | set_page_private(page, 0); |
23be7468 | 1221 | page->mapping = NULL; |
b4330afb MK |
1222 | VM_BUG_ON_PAGE(page_count(page), page); |
1223 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
07443a85 | 1224 | restore_reserve = PagePrivate(page); |
16c794b4 | 1225 | ClearPagePrivate(page); |
27a85ef1 | 1226 | |
1c5ecae3 MK |
1227 | /* |
1228 | * A return code of zero implies that the subpool will be under its | |
1229 | * minimum size if the reservation is not restored after page is free. | |
1230 | * Therefore, force restore_reserve operation. | |
1231 | */ | |
1232 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1233 | restore_reserve = true; | |
1234 | ||
27a85ef1 | 1235 | spin_lock(&hugetlb_lock); |
bcc54222 | 1236 | clear_page_huge_active(page); |
6d76dcf4 AK |
1237 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1238 | pages_per_huge_page(h), page); | |
07443a85 JK |
1239 | if (restore_reserve) |
1240 | h->resv_huge_pages++; | |
1241 | ||
944d9fec | 1242 | if (h->surplus_huge_pages_node[nid]) { |
0edaecfa AK |
1243 | /* remove the page from active list */ |
1244 | list_del(&page->lru); | |
a5516438 AK |
1245 | update_and_free_page(h, page); |
1246 | h->surplus_huge_pages--; | |
1247 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1248 | } else { |
5d3a551c | 1249 | arch_clear_hugepage_flags(page); |
a5516438 | 1250 | enqueue_huge_page(h, page); |
7893d1d5 | 1251 | } |
27a85ef1 DG |
1252 | spin_unlock(&hugetlb_lock); |
1253 | } | |
1254 | ||
a5516438 | 1255 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1256 | { |
0edaecfa | 1257 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1258 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1259 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1260 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1261 | h->nr_huge_pages++; |
1262 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 AK |
1263 | spin_unlock(&hugetlb_lock); |
1264 | put_page(page); /* free it into the hugepage allocator */ | |
1265 | } | |
1266 | ||
d00181b9 | 1267 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1268 | { |
1269 | int i; | |
1270 | int nr_pages = 1 << order; | |
1271 | struct page *p = page + 1; | |
1272 | ||
1273 | /* we rely on prep_new_huge_page to set the destructor */ | |
1274 | set_compound_order(page, order); | |
ef5a22be | 1275 | __ClearPageReserved(page); |
de09d31d | 1276 | __SetPageHead(page); |
20a0307c | 1277 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1278 | /* |
1279 | * For gigantic hugepages allocated through bootmem at | |
1280 | * boot, it's safer to be consistent with the not-gigantic | |
1281 | * hugepages and clear the PG_reserved bit from all tail pages | |
1282 | * too. Otherwse drivers using get_user_pages() to access tail | |
1283 | * pages may get the reference counting wrong if they see | |
1284 | * PG_reserved set on a tail page (despite the head page not | |
1285 | * having PG_reserved set). Enforcing this consistency between | |
1286 | * head and tail pages allows drivers to optimize away a check | |
1287 | * on the head page when they need know if put_page() is needed | |
1288 | * after get_user_pages(). | |
1289 | */ | |
1290 | __ClearPageReserved(p); | |
58a84aa9 | 1291 | set_page_count(p, 0); |
1d798ca3 | 1292 | set_compound_head(p, page); |
20a0307c | 1293 | } |
b4330afb | 1294 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1295 | } |
1296 | ||
7795912c AM |
1297 | /* |
1298 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1299 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1300 | * details. | |
1301 | */ | |
20a0307c WF |
1302 | int PageHuge(struct page *page) |
1303 | { | |
20a0307c WF |
1304 | if (!PageCompound(page)) |
1305 | return 0; | |
1306 | ||
1307 | page = compound_head(page); | |
f1e61557 | 1308 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1309 | } |
43131e14 NH |
1310 | EXPORT_SYMBOL_GPL(PageHuge); |
1311 | ||
27c73ae7 AA |
1312 | /* |
1313 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1314 | * normal or transparent huge pages. | |
1315 | */ | |
1316 | int PageHeadHuge(struct page *page_head) | |
1317 | { | |
27c73ae7 AA |
1318 | if (!PageHead(page_head)) |
1319 | return 0; | |
1320 | ||
758f66a2 | 1321 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1322 | } |
27c73ae7 | 1323 | |
13d60f4b ZY |
1324 | pgoff_t __basepage_index(struct page *page) |
1325 | { | |
1326 | struct page *page_head = compound_head(page); | |
1327 | pgoff_t index = page_index(page_head); | |
1328 | unsigned long compound_idx; | |
1329 | ||
1330 | if (!PageHuge(page_head)) | |
1331 | return page_index(page); | |
1332 | ||
1333 | if (compound_order(page_head) >= MAX_ORDER) | |
1334 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1335 | else | |
1336 | compound_idx = page - page_head; | |
1337 | ||
1338 | return (index << compound_order(page_head)) + compound_idx; | |
1339 | } | |
1340 | ||
a5516438 | 1341 | static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid) |
1da177e4 | 1342 | { |
1da177e4 | 1343 | struct page *page; |
f96efd58 | 1344 | |
96db800f | 1345 | page = __alloc_pages_node(nid, |
86cdb465 | 1346 | htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE| |
551883ae | 1347 | __GFP_REPEAT|__GFP_NOWARN, |
a5516438 | 1348 | huge_page_order(h)); |
1da177e4 | 1349 | if (page) { |
a5516438 | 1350 | prep_new_huge_page(h, page, nid); |
1da177e4 | 1351 | } |
63b4613c NA |
1352 | |
1353 | return page; | |
1354 | } | |
1355 | ||
b2261026 JK |
1356 | static int alloc_fresh_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
1357 | { | |
1358 | struct page *page; | |
1359 | int nr_nodes, node; | |
1360 | int ret = 0; | |
1361 | ||
1362 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
1363 | page = alloc_fresh_huge_page_node(h, node); | |
1364 | if (page) { | |
1365 | ret = 1; | |
1366 | break; | |
1367 | } | |
1368 | } | |
1369 | ||
1370 | if (ret) | |
1371 | count_vm_event(HTLB_BUDDY_PGALLOC); | |
1372 | else | |
1373 | count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
1374 | ||
1375 | return ret; | |
1376 | } | |
1377 | ||
e8c5c824 LS |
1378 | /* |
1379 | * Free huge page from pool from next node to free. | |
1380 | * Attempt to keep persistent huge pages more or less | |
1381 | * balanced over allowed nodes. | |
1382 | * Called with hugetlb_lock locked. | |
1383 | */ | |
6ae11b27 LS |
1384 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1385 | bool acct_surplus) | |
e8c5c824 | 1386 | { |
b2261026 | 1387 | int nr_nodes, node; |
e8c5c824 LS |
1388 | int ret = 0; |
1389 | ||
b2261026 | 1390 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1391 | /* |
1392 | * If we're returning unused surplus pages, only examine | |
1393 | * nodes with surplus pages. | |
1394 | */ | |
b2261026 JK |
1395 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1396 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1397 | struct page *page = |
b2261026 | 1398 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1399 | struct page, lru); |
1400 | list_del(&page->lru); | |
1401 | h->free_huge_pages--; | |
b2261026 | 1402 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1403 | if (acct_surplus) { |
1404 | h->surplus_huge_pages--; | |
b2261026 | 1405 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1406 | } |
e8c5c824 LS |
1407 | update_and_free_page(h, page); |
1408 | ret = 1; | |
9a76db09 | 1409 | break; |
e8c5c824 | 1410 | } |
b2261026 | 1411 | } |
e8c5c824 LS |
1412 | |
1413 | return ret; | |
1414 | } | |
1415 | ||
c8721bbb NH |
1416 | /* |
1417 | * Dissolve a given free hugepage into free buddy pages. This function does | |
1418 | * nothing for in-use (including surplus) hugepages. | |
1419 | */ | |
1420 | static void dissolve_free_huge_page(struct page *page) | |
1421 | { | |
1422 | spin_lock(&hugetlb_lock); | |
1423 | if (PageHuge(page) && !page_count(page)) { | |
1424 | struct hstate *h = page_hstate(page); | |
1425 | int nid = page_to_nid(page); | |
1426 | list_del(&page->lru); | |
1427 | h->free_huge_pages--; | |
1428 | h->free_huge_pages_node[nid]--; | |
1429 | update_and_free_page(h, page); | |
1430 | } | |
1431 | spin_unlock(&hugetlb_lock); | |
1432 | } | |
1433 | ||
1434 | /* | |
1435 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1436 | * make specified memory blocks removable from the system. | |
1437 | * Note that start_pfn should aligned with (minimum) hugepage size. | |
1438 | */ | |
1439 | void dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) | |
1440 | { | |
c8721bbb | 1441 | unsigned long pfn; |
c8721bbb | 1442 | |
d0177639 LZ |
1443 | if (!hugepages_supported()) |
1444 | return; | |
1445 | ||
641844f5 NH |
1446 | VM_BUG_ON(!IS_ALIGNED(start_pfn, 1 << minimum_order)); |
1447 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) | |
c8721bbb NH |
1448 | dissolve_free_huge_page(pfn_to_page(pfn)); |
1449 | } | |
1450 | ||
099730d6 DH |
1451 | /* |
1452 | * There are 3 ways this can get called: | |
1453 | * 1. With vma+addr: we use the VMA's memory policy | |
1454 | * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge | |
1455 | * page from any node, and let the buddy allocator itself figure | |
1456 | * it out. | |
1457 | * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page | |
1458 | * strictly from 'nid' | |
1459 | */ | |
1460 | static struct page *__hugetlb_alloc_buddy_huge_page(struct hstate *h, | |
1461 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
1462 | { | |
1463 | int order = huge_page_order(h); | |
1464 | gfp_t gfp = htlb_alloc_mask(h)|__GFP_COMP|__GFP_REPEAT|__GFP_NOWARN; | |
1465 | unsigned int cpuset_mems_cookie; | |
1466 | ||
1467 | /* | |
1468 | * We need a VMA to get a memory policy. If we do not | |
e0ec90ee DH |
1469 | * have one, we use the 'nid' argument. |
1470 | * | |
1471 | * The mempolicy stuff below has some non-inlined bits | |
1472 | * and calls ->vm_ops. That makes it hard to optimize at | |
1473 | * compile-time, even when NUMA is off and it does | |
1474 | * nothing. This helps the compiler optimize it out. | |
099730d6 | 1475 | */ |
e0ec90ee | 1476 | if (!IS_ENABLED(CONFIG_NUMA) || !vma) { |
099730d6 DH |
1477 | /* |
1478 | * If a specific node is requested, make sure to | |
1479 | * get memory from there, but only when a node | |
1480 | * is explicitly specified. | |
1481 | */ | |
1482 | if (nid != NUMA_NO_NODE) | |
1483 | gfp |= __GFP_THISNODE; | |
1484 | /* | |
1485 | * Make sure to call something that can handle | |
1486 | * nid=NUMA_NO_NODE | |
1487 | */ | |
1488 | return alloc_pages_node(nid, gfp, order); | |
1489 | } | |
1490 | ||
1491 | /* | |
1492 | * OK, so we have a VMA. Fetch the mempolicy and try to | |
e0ec90ee DH |
1493 | * allocate a huge page with it. We will only reach this |
1494 | * when CONFIG_NUMA=y. | |
099730d6 DH |
1495 | */ |
1496 | do { | |
1497 | struct page *page; | |
1498 | struct mempolicy *mpol; | |
1499 | struct zonelist *zl; | |
1500 | nodemask_t *nodemask; | |
1501 | ||
1502 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1503 | zl = huge_zonelist(vma, addr, gfp, &mpol, &nodemask); | |
1504 | mpol_cond_put(mpol); | |
1505 | page = __alloc_pages_nodemask(gfp, order, zl, nodemask); | |
1506 | if (page) | |
1507 | return page; | |
1508 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); | |
1509 | ||
1510 | return NULL; | |
1511 | } | |
1512 | ||
1513 | /* | |
1514 | * There are two ways to allocate a huge page: | |
1515 | * 1. When you have a VMA and an address (like a fault) | |
1516 | * 2. When you have no VMA (like when setting /proc/.../nr_hugepages) | |
1517 | * | |
1518 | * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in | |
1519 | * this case which signifies that the allocation should be done with | |
1520 | * respect for the VMA's memory policy. | |
1521 | * | |
1522 | * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This | |
1523 | * implies that memory policies will not be taken in to account. | |
1524 | */ | |
1525 | static struct page *__alloc_buddy_huge_page(struct hstate *h, | |
1526 | struct vm_area_struct *vma, unsigned long addr, int nid) | |
7893d1d5 AL |
1527 | { |
1528 | struct page *page; | |
bf50bab2 | 1529 | unsigned int r_nid; |
7893d1d5 | 1530 | |
bae7f4ae | 1531 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1532 | return NULL; |
1533 | ||
099730d6 DH |
1534 | /* |
1535 | * Make sure that anyone specifying 'nid' is not also specifying a VMA. | |
1536 | * This makes sure the caller is picking _one_ of the modes with which | |
1537 | * we can call this function, not both. | |
1538 | */ | |
1539 | if (vma || (addr != -1)) { | |
e0ec90ee DH |
1540 | VM_WARN_ON_ONCE(addr == -1); |
1541 | VM_WARN_ON_ONCE(nid != NUMA_NO_NODE); | |
099730d6 | 1542 | } |
d1c3fb1f NA |
1543 | /* |
1544 | * Assume we will successfully allocate the surplus page to | |
1545 | * prevent racing processes from causing the surplus to exceed | |
1546 | * overcommit | |
1547 | * | |
1548 | * This however introduces a different race, where a process B | |
1549 | * tries to grow the static hugepage pool while alloc_pages() is | |
1550 | * called by process A. B will only examine the per-node | |
1551 | * counters in determining if surplus huge pages can be | |
1552 | * converted to normal huge pages in adjust_pool_surplus(). A | |
1553 | * won't be able to increment the per-node counter, until the | |
1554 | * lock is dropped by B, but B doesn't drop hugetlb_lock until | |
1555 | * no more huge pages can be converted from surplus to normal | |
1556 | * state (and doesn't try to convert again). Thus, we have a | |
1557 | * case where a surplus huge page exists, the pool is grown, and | |
1558 | * the surplus huge page still exists after, even though it | |
1559 | * should just have been converted to a normal huge page. This | |
1560 | * does not leak memory, though, as the hugepage will be freed | |
1561 | * once it is out of use. It also does not allow the counters to | |
1562 | * go out of whack in adjust_pool_surplus() as we don't modify | |
1563 | * the node values until we've gotten the hugepage and only the | |
1564 | * per-node value is checked there. | |
1565 | */ | |
1566 | spin_lock(&hugetlb_lock); | |
a5516438 | 1567 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { |
d1c3fb1f NA |
1568 | spin_unlock(&hugetlb_lock); |
1569 | return NULL; | |
1570 | } else { | |
a5516438 AK |
1571 | h->nr_huge_pages++; |
1572 | h->surplus_huge_pages++; | |
d1c3fb1f NA |
1573 | } |
1574 | spin_unlock(&hugetlb_lock); | |
1575 | ||
099730d6 | 1576 | page = __hugetlb_alloc_buddy_huge_page(h, vma, addr, nid); |
d1c3fb1f NA |
1577 | |
1578 | spin_lock(&hugetlb_lock); | |
7893d1d5 | 1579 | if (page) { |
0edaecfa | 1580 | INIT_LIST_HEAD(&page->lru); |
bf50bab2 | 1581 | r_nid = page_to_nid(page); |
f1e61557 | 1582 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
9dd540e2 | 1583 | set_hugetlb_cgroup(page, NULL); |
d1c3fb1f NA |
1584 | /* |
1585 | * We incremented the global counters already | |
1586 | */ | |
bf50bab2 NH |
1587 | h->nr_huge_pages_node[r_nid]++; |
1588 | h->surplus_huge_pages_node[r_nid]++; | |
3b116300 | 1589 | __count_vm_event(HTLB_BUDDY_PGALLOC); |
d1c3fb1f | 1590 | } else { |
a5516438 AK |
1591 | h->nr_huge_pages--; |
1592 | h->surplus_huge_pages--; | |
3b116300 | 1593 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); |
7893d1d5 | 1594 | } |
d1c3fb1f | 1595 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1596 | |
1597 | return page; | |
1598 | } | |
1599 | ||
099730d6 DH |
1600 | /* |
1601 | * Allocate a huge page from 'nid'. Note, 'nid' may be | |
1602 | * NUMA_NO_NODE, which means that it may be allocated | |
1603 | * anywhere. | |
1604 | */ | |
e0ec90ee | 1605 | static |
099730d6 DH |
1606 | struct page *__alloc_buddy_huge_page_no_mpol(struct hstate *h, int nid) |
1607 | { | |
1608 | unsigned long addr = -1; | |
1609 | ||
1610 | return __alloc_buddy_huge_page(h, NULL, addr, nid); | |
1611 | } | |
1612 | ||
1613 | /* | |
1614 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1615 | */ | |
e0ec90ee | 1616 | static |
099730d6 DH |
1617 | struct page *__alloc_buddy_huge_page_with_mpol(struct hstate *h, |
1618 | struct vm_area_struct *vma, unsigned long addr) | |
1619 | { | |
1620 | return __alloc_buddy_huge_page(h, vma, addr, NUMA_NO_NODE); | |
1621 | } | |
1622 | ||
bf50bab2 NH |
1623 | /* |
1624 | * This allocation function is useful in the context where vma is irrelevant. | |
1625 | * E.g. soft-offlining uses this function because it only cares physical | |
1626 | * address of error page. | |
1627 | */ | |
1628 | struct page *alloc_huge_page_node(struct hstate *h, int nid) | |
1629 | { | |
4ef91848 | 1630 | struct page *page = NULL; |
bf50bab2 NH |
1631 | |
1632 | spin_lock(&hugetlb_lock); | |
4ef91848 JK |
1633 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
1634 | page = dequeue_huge_page_node(h, nid); | |
bf50bab2 NH |
1635 | spin_unlock(&hugetlb_lock); |
1636 | ||
94ae8ba7 | 1637 | if (!page) |
099730d6 | 1638 | page = __alloc_buddy_huge_page_no_mpol(h, nid); |
bf50bab2 NH |
1639 | |
1640 | return page; | |
1641 | } | |
1642 | ||
e4e574b7 | 1643 | /* |
25985edc | 1644 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1645 | * of size 'delta'. |
1646 | */ | |
a5516438 | 1647 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1648 | { |
1649 | struct list_head surplus_list; | |
1650 | struct page *page, *tmp; | |
1651 | int ret, i; | |
1652 | int needed, allocated; | |
28073b02 | 1653 | bool alloc_ok = true; |
e4e574b7 | 1654 | |
a5516438 | 1655 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1656 | if (needed <= 0) { |
a5516438 | 1657 | h->resv_huge_pages += delta; |
e4e574b7 | 1658 | return 0; |
ac09b3a1 | 1659 | } |
e4e574b7 AL |
1660 | |
1661 | allocated = 0; | |
1662 | INIT_LIST_HEAD(&surplus_list); | |
1663 | ||
1664 | ret = -ENOMEM; | |
1665 | retry: | |
1666 | spin_unlock(&hugetlb_lock); | |
1667 | for (i = 0; i < needed; i++) { | |
099730d6 | 1668 | page = __alloc_buddy_huge_page_no_mpol(h, NUMA_NO_NODE); |
28073b02 HD |
1669 | if (!page) { |
1670 | alloc_ok = false; | |
1671 | break; | |
1672 | } | |
e4e574b7 AL |
1673 | list_add(&page->lru, &surplus_list); |
1674 | } | |
28073b02 | 1675 | allocated += i; |
e4e574b7 AL |
1676 | |
1677 | /* | |
1678 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1679 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1680 | */ | |
1681 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1682 | needed = (h->resv_huge_pages + delta) - |
1683 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1684 | if (needed > 0) { |
1685 | if (alloc_ok) | |
1686 | goto retry; | |
1687 | /* | |
1688 | * We were not able to allocate enough pages to | |
1689 | * satisfy the entire reservation so we free what | |
1690 | * we've allocated so far. | |
1691 | */ | |
1692 | goto free; | |
1693 | } | |
e4e574b7 AL |
1694 | /* |
1695 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1696 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1697 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1698 | * allocator. Commit the entire reservation here to prevent another |
1699 | * process from stealing the pages as they are added to the pool but | |
1700 | * before they are reserved. | |
e4e574b7 AL |
1701 | */ |
1702 | needed += allocated; | |
a5516438 | 1703 | h->resv_huge_pages += delta; |
e4e574b7 | 1704 | ret = 0; |
a9869b83 | 1705 | |
19fc3f0a | 1706 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1707 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1708 | if ((--needed) < 0) |
1709 | break; | |
a9869b83 NH |
1710 | /* |
1711 | * This page is now managed by the hugetlb allocator and has | |
1712 | * no users -- drop the buddy allocator's reference. | |
1713 | */ | |
1714 | put_page_testzero(page); | |
309381fe | 1715 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1716 | enqueue_huge_page(h, page); |
19fc3f0a | 1717 | } |
28073b02 | 1718 | free: |
b0365c8d | 1719 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1720 | |
1721 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1722 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1723 | put_page(page); | |
a9869b83 | 1724 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1725 | |
1726 | return ret; | |
1727 | } | |
1728 | ||
1729 | /* | |
1730 | * When releasing a hugetlb pool reservation, any surplus pages that were | |
1731 | * allocated to satisfy the reservation must be explicitly freed if they were | |
1732 | * never used. | |
685f3457 | 1733 | * Called with hugetlb_lock held. |
e4e574b7 | 1734 | */ |
a5516438 AK |
1735 | static void return_unused_surplus_pages(struct hstate *h, |
1736 | unsigned long unused_resv_pages) | |
e4e574b7 | 1737 | { |
e4e574b7 AL |
1738 | unsigned long nr_pages; |
1739 | ||
ac09b3a1 | 1740 | /* Uncommit the reservation */ |
a5516438 | 1741 | h->resv_huge_pages -= unused_resv_pages; |
ac09b3a1 | 1742 | |
aa888a74 | 1743 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1744 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1745 | return; |
1746 | ||
a5516438 | 1747 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1748 | |
685f3457 LS |
1749 | /* |
1750 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1751 | * evenly across all nodes with memory. Iterate across these nodes |
1752 | * until we can no longer free unreserved surplus pages. This occurs | |
1753 | * when the nodes with surplus pages have no free pages. | |
1754 | * free_pool_huge_page() will balance the the freed pages across the | |
1755 | * on-line nodes with memory and will handle the hstate accounting. | |
685f3457 LS |
1756 | */ |
1757 | while (nr_pages--) { | |
8cebfcd0 | 1758 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
685f3457 | 1759 | break; |
7848a4bf | 1760 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 AL |
1761 | } |
1762 | } | |
1763 | ||
5e911373 | 1764 | |
c37f9fb1 | 1765 | /* |
feba16e2 | 1766 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1767 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1768 | * |
1769 | * vma_needs_reservation is called to determine if the huge page at addr | |
1770 | * within the vma has an associated reservation. If a reservation is | |
1771 | * needed, the value 1 is returned. The caller is then responsible for | |
1772 | * managing the global reservation and subpool usage counts. After | |
1773 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1774 | * to add the page to the reservation map. If the page allocation fails, |
1775 | * the reservation must be ended instead of committed. vma_end_reservation | |
1776 | * is called in such cases. | |
cf3ad20b MK |
1777 | * |
1778 | * In the normal case, vma_commit_reservation returns the same value | |
1779 | * as the preceding vma_needs_reservation call. The only time this | |
1780 | * is not the case is if a reserve map was changed between calls. It | |
1781 | * is the responsibility of the caller to notice the difference and | |
1782 | * take appropriate action. | |
c37f9fb1 | 1783 | */ |
5e911373 MK |
1784 | enum vma_resv_mode { |
1785 | VMA_NEEDS_RESV, | |
1786 | VMA_COMMIT_RESV, | |
feba16e2 | 1787 | VMA_END_RESV, |
5e911373 | 1788 | }; |
cf3ad20b MK |
1789 | static long __vma_reservation_common(struct hstate *h, |
1790 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1791 | enum vma_resv_mode mode) |
c37f9fb1 | 1792 | { |
4e35f483 JK |
1793 | struct resv_map *resv; |
1794 | pgoff_t idx; | |
cf3ad20b | 1795 | long ret; |
c37f9fb1 | 1796 | |
4e35f483 JK |
1797 | resv = vma_resv_map(vma); |
1798 | if (!resv) | |
84afd99b | 1799 | return 1; |
c37f9fb1 | 1800 | |
4e35f483 | 1801 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1802 | switch (mode) { |
1803 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1804 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1805 | break; |
1806 | case VMA_COMMIT_RESV: | |
1807 | ret = region_add(resv, idx, idx + 1); | |
1808 | break; | |
feba16e2 | 1809 | case VMA_END_RESV: |
5e911373 MK |
1810 | region_abort(resv, idx, idx + 1); |
1811 | ret = 0; | |
1812 | break; | |
1813 | default: | |
1814 | BUG(); | |
1815 | } | |
84afd99b | 1816 | |
4e35f483 | 1817 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1818 | return ret; |
4e35f483 | 1819 | else |
cf3ad20b | 1820 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1821 | } |
cf3ad20b MK |
1822 | |
1823 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1824 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1825 | { |
5e911373 | 1826 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1827 | } |
84afd99b | 1828 | |
cf3ad20b MK |
1829 | static long vma_commit_reservation(struct hstate *h, |
1830 | struct vm_area_struct *vma, unsigned long addr) | |
1831 | { | |
5e911373 MK |
1832 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1833 | } | |
1834 | ||
feba16e2 | 1835 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1836 | struct vm_area_struct *vma, unsigned long addr) |
1837 | { | |
feba16e2 | 1838 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1839 | } |
1840 | ||
70c3547e | 1841 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1842 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1843 | { |
90481622 | 1844 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1845 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1846 | struct page *page; |
d85f69b0 MK |
1847 | long map_chg, map_commit; |
1848 | long gbl_chg; | |
6d76dcf4 AK |
1849 | int ret, idx; |
1850 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 1851 | |
6d76dcf4 | 1852 | idx = hstate_index(h); |
a1e78772 | 1853 | /* |
d85f69b0 MK |
1854 | * Examine the region/reserve map to determine if the process |
1855 | * has a reservation for the page to be allocated. A return | |
1856 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 1857 | */ |
d85f69b0 MK |
1858 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
1859 | if (map_chg < 0) | |
76dcee75 | 1860 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
1861 | |
1862 | /* | |
1863 | * Processes that did not create the mapping will have no | |
1864 | * reserves as indicated by the region/reserve map. Check | |
1865 | * that the allocation will not exceed the subpool limit. | |
1866 | * Allocations for MAP_NORESERVE mappings also need to be | |
1867 | * checked against any subpool limit. | |
1868 | */ | |
1869 | if (map_chg || avoid_reserve) { | |
1870 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
1871 | if (gbl_chg < 0) { | |
feba16e2 | 1872 | vma_end_reservation(h, vma, addr); |
76dcee75 | 1873 | return ERR_PTR(-ENOSPC); |
5e911373 | 1874 | } |
1da177e4 | 1875 | |
d85f69b0 MK |
1876 | /* |
1877 | * Even though there was no reservation in the region/reserve | |
1878 | * map, there could be reservations associated with the | |
1879 | * subpool that can be used. This would be indicated if the | |
1880 | * return value of hugepage_subpool_get_pages() is zero. | |
1881 | * However, if avoid_reserve is specified we still avoid even | |
1882 | * the subpool reservations. | |
1883 | */ | |
1884 | if (avoid_reserve) | |
1885 | gbl_chg = 1; | |
1886 | } | |
1887 | ||
6d76dcf4 | 1888 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
1889 | if (ret) |
1890 | goto out_subpool_put; | |
1891 | ||
1da177e4 | 1892 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
1893 | /* |
1894 | * glb_chg is passed to indicate whether or not a page must be taken | |
1895 | * from the global free pool (global change). gbl_chg == 0 indicates | |
1896 | * a reservation exists for the allocation. | |
1897 | */ | |
1898 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 1899 | if (!page) { |
94ae8ba7 | 1900 | spin_unlock(&hugetlb_lock); |
099730d6 | 1901 | page = __alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
1902 | if (!page) |
1903 | goto out_uncharge_cgroup; | |
a88c7695 NH |
1904 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
1905 | SetPagePrivate(page); | |
1906 | h->resv_huge_pages--; | |
1907 | } | |
79dbb236 AK |
1908 | spin_lock(&hugetlb_lock); |
1909 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 1910 | /* Fall through */ |
68842c9b | 1911 | } |
81a6fcae JK |
1912 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
1913 | spin_unlock(&hugetlb_lock); | |
348ea204 | 1914 | |
90481622 | 1915 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 1916 | |
d85f69b0 MK |
1917 | map_commit = vma_commit_reservation(h, vma, addr); |
1918 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
1919 | /* |
1920 | * The page was added to the reservation map between | |
1921 | * vma_needs_reservation and vma_commit_reservation. | |
1922 | * This indicates a race with hugetlb_reserve_pages. | |
1923 | * Adjust for the subpool count incremented above AND | |
1924 | * in hugetlb_reserve_pages for the same page. Also, | |
1925 | * the reservation count added in hugetlb_reserve_pages | |
1926 | * no longer applies. | |
1927 | */ | |
1928 | long rsv_adjust; | |
1929 | ||
1930 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
1931 | hugetlb_acct_memory(h, -rsv_adjust); | |
1932 | } | |
90d8b7e6 | 1933 | return page; |
8f34af6f JZ |
1934 | |
1935 | out_uncharge_cgroup: | |
1936 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
1937 | out_subpool_put: | |
d85f69b0 | 1938 | if (map_chg || avoid_reserve) |
8f34af6f | 1939 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 1940 | vma_end_reservation(h, vma, addr); |
8f34af6f | 1941 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
1942 | } |
1943 | ||
74060e4d NH |
1944 | /* |
1945 | * alloc_huge_page()'s wrapper which simply returns the page if allocation | |
1946 | * succeeds, otherwise NULL. This function is called from new_vma_page(), | |
1947 | * where no ERR_VALUE is expected to be returned. | |
1948 | */ | |
1949 | struct page *alloc_huge_page_noerr(struct vm_area_struct *vma, | |
1950 | unsigned long addr, int avoid_reserve) | |
1951 | { | |
1952 | struct page *page = alloc_huge_page(vma, addr, avoid_reserve); | |
1953 | if (IS_ERR(page)) | |
1954 | page = NULL; | |
1955 | return page; | |
1956 | } | |
1957 | ||
91f47662 | 1958 | int __weak alloc_bootmem_huge_page(struct hstate *h) |
aa888a74 AK |
1959 | { |
1960 | struct huge_bootmem_page *m; | |
b2261026 | 1961 | int nr_nodes, node; |
aa888a74 | 1962 | |
b2261026 | 1963 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
1964 | void *addr; |
1965 | ||
8b89a116 GS |
1966 | addr = memblock_virt_alloc_try_nid_nopanic( |
1967 | huge_page_size(h), huge_page_size(h), | |
1968 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
1969 | if (addr) { |
1970 | /* | |
1971 | * Use the beginning of the huge page to store the | |
1972 | * huge_bootmem_page struct (until gather_bootmem | |
1973 | * puts them into the mem_map). | |
1974 | */ | |
1975 | m = addr; | |
91f47662 | 1976 | goto found; |
aa888a74 | 1977 | } |
aa888a74 AK |
1978 | } |
1979 | return 0; | |
1980 | ||
1981 | found: | |
df994ead | 1982 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
1983 | /* Put them into a private list first because mem_map is not up yet */ |
1984 | list_add(&m->list, &huge_boot_pages); | |
1985 | m->hstate = h; | |
1986 | return 1; | |
1987 | } | |
1988 | ||
d00181b9 KS |
1989 | static void __init prep_compound_huge_page(struct page *page, |
1990 | unsigned int order) | |
18229df5 AW |
1991 | { |
1992 | if (unlikely(order > (MAX_ORDER - 1))) | |
1993 | prep_compound_gigantic_page(page, order); | |
1994 | else | |
1995 | prep_compound_page(page, order); | |
1996 | } | |
1997 | ||
aa888a74 AK |
1998 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
1999 | static void __init gather_bootmem_prealloc(void) | |
2000 | { | |
2001 | struct huge_bootmem_page *m; | |
2002 | ||
2003 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 2004 | struct hstate *h = m->hstate; |
ee8f248d BB |
2005 | struct page *page; |
2006 | ||
2007 | #ifdef CONFIG_HIGHMEM | |
2008 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
2009 | memblock_free_late(__pa(m), |
2010 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
2011 | #else |
2012 | page = virt_to_page(m); | |
2013 | #endif | |
aa888a74 | 2014 | WARN_ON(page_count(page) != 1); |
18229df5 | 2015 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2016 | WARN_ON(PageReserved(page)); |
aa888a74 | 2017 | prep_new_huge_page(h, page, page_to_nid(page)); |
b0320c7b RA |
2018 | /* |
2019 | * If we had gigantic hugepages allocated at boot time, we need | |
2020 | * to restore the 'stolen' pages to totalram_pages in order to | |
2021 | * fix confusing memory reports from free(1) and another | |
2022 | * side-effects, like CommitLimit going negative. | |
2023 | */ | |
bae7f4ae | 2024 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2025 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
2026 | } |
2027 | } | |
2028 | ||
8faa8b07 | 2029 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2030 | { |
2031 | unsigned long i; | |
a5516438 | 2032 | |
e5ff2159 | 2033 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2034 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2035 | if (!alloc_bootmem_huge_page(h)) |
2036 | break; | |
9b5e5d0f | 2037 | } else if (!alloc_fresh_huge_page(h, |
8cebfcd0 | 2038 | &node_states[N_MEMORY])) |
1da177e4 | 2039 | break; |
1da177e4 | 2040 | } |
8faa8b07 | 2041 | h->max_huge_pages = i; |
e5ff2159 AK |
2042 | } |
2043 | ||
2044 | static void __init hugetlb_init_hstates(void) | |
2045 | { | |
2046 | struct hstate *h; | |
2047 | ||
2048 | for_each_hstate(h) { | |
641844f5 NH |
2049 | if (minimum_order > huge_page_order(h)) |
2050 | minimum_order = huge_page_order(h); | |
2051 | ||
8faa8b07 | 2052 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2053 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2054 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2055 | } |
641844f5 | 2056 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2057 | } |
2058 | ||
4abd32db AK |
2059 | static char * __init memfmt(char *buf, unsigned long n) |
2060 | { | |
2061 | if (n >= (1UL << 30)) | |
2062 | sprintf(buf, "%lu GB", n >> 30); | |
2063 | else if (n >= (1UL << 20)) | |
2064 | sprintf(buf, "%lu MB", n >> 20); | |
2065 | else | |
2066 | sprintf(buf, "%lu KB", n >> 10); | |
2067 | return buf; | |
2068 | } | |
2069 | ||
e5ff2159 AK |
2070 | static void __init report_hugepages(void) |
2071 | { | |
2072 | struct hstate *h; | |
2073 | ||
2074 | for_each_hstate(h) { | |
4abd32db | 2075 | char buf[32]; |
ffb22af5 | 2076 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
4abd32db AK |
2077 | memfmt(buf, huge_page_size(h)), |
2078 | h->free_huge_pages); | |
e5ff2159 AK |
2079 | } |
2080 | } | |
2081 | ||
1da177e4 | 2082 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2083 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2084 | nodemask_t *nodes_allowed) | |
1da177e4 | 2085 | { |
4415cc8d CL |
2086 | int i; |
2087 | ||
bae7f4ae | 2088 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2089 | return; |
2090 | ||
6ae11b27 | 2091 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2092 | struct page *page, *next; |
a5516438 AK |
2093 | struct list_head *freel = &h->hugepage_freelists[i]; |
2094 | list_for_each_entry_safe(page, next, freel, lru) { | |
2095 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2096 | return; |
1da177e4 LT |
2097 | if (PageHighMem(page)) |
2098 | continue; | |
2099 | list_del(&page->lru); | |
e5ff2159 | 2100 | update_and_free_page(h, page); |
a5516438 AK |
2101 | h->free_huge_pages--; |
2102 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2103 | } |
2104 | } | |
2105 | } | |
2106 | #else | |
6ae11b27 LS |
2107 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2108 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2109 | { |
2110 | } | |
2111 | #endif | |
2112 | ||
20a0307c WF |
2113 | /* |
2114 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2115 | * balanced by operating on them in a round-robin fashion. | |
2116 | * Returns 1 if an adjustment was made. | |
2117 | */ | |
6ae11b27 LS |
2118 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2119 | int delta) | |
20a0307c | 2120 | { |
b2261026 | 2121 | int nr_nodes, node; |
20a0307c WF |
2122 | |
2123 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2124 | |
b2261026 JK |
2125 | if (delta < 0) { |
2126 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2127 | if (h->surplus_huge_pages_node[node]) | |
2128 | goto found; | |
e8c5c824 | 2129 | } |
b2261026 JK |
2130 | } else { |
2131 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2132 | if (h->surplus_huge_pages_node[node] < | |
2133 | h->nr_huge_pages_node[node]) | |
2134 | goto found; | |
e8c5c824 | 2135 | } |
b2261026 JK |
2136 | } |
2137 | return 0; | |
20a0307c | 2138 | |
b2261026 JK |
2139 | found: |
2140 | h->surplus_huge_pages += delta; | |
2141 | h->surplus_huge_pages_node[node] += delta; | |
2142 | return 1; | |
20a0307c WF |
2143 | } |
2144 | ||
a5516438 | 2145 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2146 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2147 | nodemask_t *nodes_allowed) | |
1da177e4 | 2148 | { |
7893d1d5 | 2149 | unsigned long min_count, ret; |
1da177e4 | 2150 | |
944d9fec | 2151 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2152 | return h->max_huge_pages; |
2153 | ||
7893d1d5 AL |
2154 | /* |
2155 | * Increase the pool size | |
2156 | * First take pages out of surplus state. Then make up the | |
2157 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2158 | * |
d15c7c09 | 2159 | * We might race with __alloc_buddy_huge_page() here and be unable |
d1c3fb1f NA |
2160 | * to convert a surplus huge page to a normal huge page. That is |
2161 | * not critical, though, it just means the overall size of the | |
2162 | * pool might be one hugepage larger than it needs to be, but | |
2163 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2164 | */ |
1da177e4 | 2165 | spin_lock(&hugetlb_lock); |
a5516438 | 2166 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2167 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2168 | break; |
2169 | } | |
2170 | ||
a5516438 | 2171 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2172 | /* |
2173 | * If this allocation races such that we no longer need the | |
2174 | * page, free_huge_page will handle it by freeing the page | |
2175 | * and reducing the surplus. | |
2176 | */ | |
2177 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
2178 | if (hstate_is_gigantic(h)) |
2179 | ret = alloc_fresh_gigantic_page(h, nodes_allowed); | |
2180 | else | |
2181 | ret = alloc_fresh_huge_page(h, nodes_allowed); | |
7893d1d5 AL |
2182 | spin_lock(&hugetlb_lock); |
2183 | if (!ret) | |
2184 | goto out; | |
2185 | ||
536240f2 MG |
2186 | /* Bail for signals. Probably ctrl-c from user */ |
2187 | if (signal_pending(current)) | |
2188 | goto out; | |
7893d1d5 | 2189 | } |
7893d1d5 AL |
2190 | |
2191 | /* | |
2192 | * Decrease the pool size | |
2193 | * First return free pages to the buddy allocator (being careful | |
2194 | * to keep enough around to satisfy reservations). Then place | |
2195 | * pages into surplus state as needed so the pool will shrink | |
2196 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2197 | * |
2198 | * By placing pages into the surplus state independent of the | |
2199 | * overcommit value, we are allowing the surplus pool size to | |
2200 | * exceed overcommit. There are few sane options here. Since | |
d15c7c09 | 2201 | * __alloc_buddy_huge_page() is checking the global counter, |
d1c3fb1f NA |
2202 | * though, we'll note that we're not allowed to exceed surplus |
2203 | * and won't grow the pool anywhere else. Not until one of the | |
2204 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2205 | */ |
a5516438 | 2206 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2207 | min_count = max(count, min_count); |
6ae11b27 | 2208 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2209 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2210 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2211 | break; |
55f67141 | 2212 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2213 | } |
a5516438 | 2214 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2215 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2216 | break; |
2217 | } | |
2218 | out: | |
a5516438 | 2219 | ret = persistent_huge_pages(h); |
1da177e4 | 2220 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2221 | return ret; |
1da177e4 LT |
2222 | } |
2223 | ||
a3437870 NA |
2224 | #define HSTATE_ATTR_RO(_name) \ |
2225 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2226 | ||
2227 | #define HSTATE_ATTR(_name) \ | |
2228 | static struct kobj_attribute _name##_attr = \ | |
2229 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2230 | ||
2231 | static struct kobject *hugepages_kobj; | |
2232 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2233 | ||
9a305230 LS |
2234 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2235 | ||
2236 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2237 | { |
2238 | int i; | |
9a305230 | 2239 | |
a3437870 | 2240 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2241 | if (hstate_kobjs[i] == kobj) { |
2242 | if (nidp) | |
2243 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2244 | return &hstates[i]; |
9a305230 LS |
2245 | } |
2246 | ||
2247 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2248 | } |
2249 | ||
06808b08 | 2250 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2251 | struct kobj_attribute *attr, char *buf) |
2252 | { | |
9a305230 LS |
2253 | struct hstate *h; |
2254 | unsigned long nr_huge_pages; | |
2255 | int nid; | |
2256 | ||
2257 | h = kobj_to_hstate(kobj, &nid); | |
2258 | if (nid == NUMA_NO_NODE) | |
2259 | nr_huge_pages = h->nr_huge_pages; | |
2260 | else | |
2261 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2262 | ||
2263 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2264 | } |
adbe8726 | 2265 | |
238d3c13 DR |
2266 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2267 | struct hstate *h, int nid, | |
2268 | unsigned long count, size_t len) | |
a3437870 NA |
2269 | { |
2270 | int err; | |
bad44b5b | 2271 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2272 | |
944d9fec | 2273 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2274 | err = -EINVAL; |
2275 | goto out; | |
2276 | } | |
2277 | ||
9a305230 LS |
2278 | if (nid == NUMA_NO_NODE) { |
2279 | /* | |
2280 | * global hstate attribute | |
2281 | */ | |
2282 | if (!(obey_mempolicy && | |
2283 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2284 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2285 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2286 | } |
2287 | } else if (nodes_allowed) { | |
2288 | /* | |
2289 | * per node hstate attribute: adjust count to global, | |
2290 | * but restrict alloc/free to the specified node. | |
2291 | */ | |
2292 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2293 | init_nodemask_of_node(nodes_allowed, nid); | |
2294 | } else | |
8cebfcd0 | 2295 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2296 | |
06808b08 | 2297 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2298 | |
8cebfcd0 | 2299 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2300 | NODEMASK_FREE(nodes_allowed); |
2301 | ||
2302 | return len; | |
adbe8726 EM |
2303 | out: |
2304 | NODEMASK_FREE(nodes_allowed); | |
2305 | return err; | |
06808b08 LS |
2306 | } |
2307 | ||
238d3c13 DR |
2308 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2309 | struct kobject *kobj, const char *buf, | |
2310 | size_t len) | |
2311 | { | |
2312 | struct hstate *h; | |
2313 | unsigned long count; | |
2314 | int nid; | |
2315 | int err; | |
2316 | ||
2317 | err = kstrtoul(buf, 10, &count); | |
2318 | if (err) | |
2319 | return err; | |
2320 | ||
2321 | h = kobj_to_hstate(kobj, &nid); | |
2322 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2323 | } | |
2324 | ||
06808b08 LS |
2325 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2326 | struct kobj_attribute *attr, char *buf) | |
2327 | { | |
2328 | return nr_hugepages_show_common(kobj, attr, buf); | |
2329 | } | |
2330 | ||
2331 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2332 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2333 | { | |
238d3c13 | 2334 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2335 | } |
2336 | HSTATE_ATTR(nr_hugepages); | |
2337 | ||
06808b08 LS |
2338 | #ifdef CONFIG_NUMA |
2339 | ||
2340 | /* | |
2341 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2342 | * huge page alloc/free. | |
2343 | */ | |
2344 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2345 | struct kobj_attribute *attr, char *buf) | |
2346 | { | |
2347 | return nr_hugepages_show_common(kobj, attr, buf); | |
2348 | } | |
2349 | ||
2350 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2351 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2352 | { | |
238d3c13 | 2353 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2354 | } |
2355 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2356 | #endif | |
2357 | ||
2358 | ||
a3437870 NA |
2359 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2360 | struct kobj_attribute *attr, char *buf) | |
2361 | { | |
9a305230 | 2362 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2363 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2364 | } | |
adbe8726 | 2365 | |
a3437870 NA |
2366 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2367 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2368 | { | |
2369 | int err; | |
2370 | unsigned long input; | |
9a305230 | 2371 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2372 | |
bae7f4ae | 2373 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2374 | return -EINVAL; |
2375 | ||
3dbb95f7 | 2376 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2377 | if (err) |
73ae31e5 | 2378 | return err; |
a3437870 NA |
2379 | |
2380 | spin_lock(&hugetlb_lock); | |
2381 | h->nr_overcommit_huge_pages = input; | |
2382 | spin_unlock(&hugetlb_lock); | |
2383 | ||
2384 | return count; | |
2385 | } | |
2386 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2387 | ||
2388 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2389 | struct kobj_attribute *attr, char *buf) | |
2390 | { | |
9a305230 LS |
2391 | struct hstate *h; |
2392 | unsigned long free_huge_pages; | |
2393 | int nid; | |
2394 | ||
2395 | h = kobj_to_hstate(kobj, &nid); | |
2396 | if (nid == NUMA_NO_NODE) | |
2397 | free_huge_pages = h->free_huge_pages; | |
2398 | else | |
2399 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2400 | ||
2401 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2402 | } |
2403 | HSTATE_ATTR_RO(free_hugepages); | |
2404 | ||
2405 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2406 | struct kobj_attribute *attr, char *buf) | |
2407 | { | |
9a305230 | 2408 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2409 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2410 | } | |
2411 | HSTATE_ATTR_RO(resv_hugepages); | |
2412 | ||
2413 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2414 | struct kobj_attribute *attr, char *buf) | |
2415 | { | |
9a305230 LS |
2416 | struct hstate *h; |
2417 | unsigned long surplus_huge_pages; | |
2418 | int nid; | |
2419 | ||
2420 | h = kobj_to_hstate(kobj, &nid); | |
2421 | if (nid == NUMA_NO_NODE) | |
2422 | surplus_huge_pages = h->surplus_huge_pages; | |
2423 | else | |
2424 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2425 | ||
2426 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2427 | } |
2428 | HSTATE_ATTR_RO(surplus_hugepages); | |
2429 | ||
2430 | static struct attribute *hstate_attrs[] = { | |
2431 | &nr_hugepages_attr.attr, | |
2432 | &nr_overcommit_hugepages_attr.attr, | |
2433 | &free_hugepages_attr.attr, | |
2434 | &resv_hugepages_attr.attr, | |
2435 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2436 | #ifdef CONFIG_NUMA |
2437 | &nr_hugepages_mempolicy_attr.attr, | |
2438 | #endif | |
a3437870 NA |
2439 | NULL, |
2440 | }; | |
2441 | ||
2442 | static struct attribute_group hstate_attr_group = { | |
2443 | .attrs = hstate_attrs, | |
2444 | }; | |
2445 | ||
094e9539 JM |
2446 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2447 | struct kobject **hstate_kobjs, | |
2448 | struct attribute_group *hstate_attr_group) | |
a3437870 NA |
2449 | { |
2450 | int retval; | |
972dc4de | 2451 | int hi = hstate_index(h); |
a3437870 | 2452 | |
9a305230 LS |
2453 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2454 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2455 | return -ENOMEM; |
2456 | ||
9a305230 | 2457 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2458 | if (retval) |
9a305230 | 2459 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2460 | |
2461 | return retval; | |
2462 | } | |
2463 | ||
2464 | static void __init hugetlb_sysfs_init(void) | |
2465 | { | |
2466 | struct hstate *h; | |
2467 | int err; | |
2468 | ||
2469 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2470 | if (!hugepages_kobj) | |
2471 | return; | |
2472 | ||
2473 | for_each_hstate(h) { | |
9a305230 LS |
2474 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2475 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2476 | if (err) |
ffb22af5 | 2477 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2478 | } |
2479 | } | |
2480 | ||
9a305230 LS |
2481 | #ifdef CONFIG_NUMA |
2482 | ||
2483 | /* | |
2484 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2485 | * with node devices in node_devices[] using a parallel array. The array |
2486 | * index of a node device or _hstate == node id. | |
2487 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2488 | * the base kernel, on the hugetlb module. |
2489 | */ | |
2490 | struct node_hstate { | |
2491 | struct kobject *hugepages_kobj; | |
2492 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2493 | }; | |
b4e289a6 | 2494 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2495 | |
2496 | /* | |
10fbcf4c | 2497 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2498 | */ |
2499 | static struct attribute *per_node_hstate_attrs[] = { | |
2500 | &nr_hugepages_attr.attr, | |
2501 | &free_hugepages_attr.attr, | |
2502 | &surplus_hugepages_attr.attr, | |
2503 | NULL, | |
2504 | }; | |
2505 | ||
2506 | static struct attribute_group per_node_hstate_attr_group = { | |
2507 | .attrs = per_node_hstate_attrs, | |
2508 | }; | |
2509 | ||
2510 | /* | |
10fbcf4c | 2511 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2512 | * Returns node id via non-NULL nidp. |
2513 | */ | |
2514 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2515 | { | |
2516 | int nid; | |
2517 | ||
2518 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2519 | struct node_hstate *nhs = &node_hstates[nid]; | |
2520 | int i; | |
2521 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2522 | if (nhs->hstate_kobjs[i] == kobj) { | |
2523 | if (nidp) | |
2524 | *nidp = nid; | |
2525 | return &hstates[i]; | |
2526 | } | |
2527 | } | |
2528 | ||
2529 | BUG(); | |
2530 | return NULL; | |
2531 | } | |
2532 | ||
2533 | /* | |
10fbcf4c | 2534 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2535 | * No-op if no hstate attributes attached. |
2536 | */ | |
3cd8b44f | 2537 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2538 | { |
2539 | struct hstate *h; | |
10fbcf4c | 2540 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2541 | |
2542 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2543 | return; /* no hstate attributes */ |
9a305230 | 2544 | |
972dc4de AK |
2545 | for_each_hstate(h) { |
2546 | int idx = hstate_index(h); | |
2547 | if (nhs->hstate_kobjs[idx]) { | |
2548 | kobject_put(nhs->hstate_kobjs[idx]); | |
2549 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2550 | } |
972dc4de | 2551 | } |
9a305230 LS |
2552 | |
2553 | kobject_put(nhs->hugepages_kobj); | |
2554 | nhs->hugepages_kobj = NULL; | |
2555 | } | |
2556 | ||
9a305230 LS |
2557 | |
2558 | /* | |
10fbcf4c | 2559 | * Register hstate attributes for a single node device. |
9a305230 LS |
2560 | * No-op if attributes already registered. |
2561 | */ | |
3cd8b44f | 2562 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2563 | { |
2564 | struct hstate *h; | |
10fbcf4c | 2565 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2566 | int err; |
2567 | ||
2568 | if (nhs->hugepages_kobj) | |
2569 | return; /* already allocated */ | |
2570 | ||
2571 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2572 | &node->dev.kobj); |
9a305230 LS |
2573 | if (!nhs->hugepages_kobj) |
2574 | return; | |
2575 | ||
2576 | for_each_hstate(h) { | |
2577 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2578 | nhs->hstate_kobjs, | |
2579 | &per_node_hstate_attr_group); | |
2580 | if (err) { | |
ffb22af5 AM |
2581 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2582 | h->name, node->dev.id); | |
9a305230 LS |
2583 | hugetlb_unregister_node(node); |
2584 | break; | |
2585 | } | |
2586 | } | |
2587 | } | |
2588 | ||
2589 | /* | |
9b5e5d0f | 2590 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2591 | * devices of nodes that have memory. All on-line nodes should have |
2592 | * registered their associated device by this time. | |
9a305230 | 2593 | */ |
7d9ca000 | 2594 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2595 | { |
2596 | int nid; | |
2597 | ||
8cebfcd0 | 2598 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2599 | struct node *node = node_devices[nid]; |
10fbcf4c | 2600 | if (node->dev.id == nid) |
9a305230 LS |
2601 | hugetlb_register_node(node); |
2602 | } | |
2603 | ||
2604 | /* | |
10fbcf4c | 2605 | * Let the node device driver know we're here so it can |
9a305230 LS |
2606 | * [un]register hstate attributes on node hotplug. |
2607 | */ | |
2608 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2609 | hugetlb_unregister_node); | |
2610 | } | |
2611 | #else /* !CONFIG_NUMA */ | |
2612 | ||
2613 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2614 | { | |
2615 | BUG(); | |
2616 | if (nidp) | |
2617 | *nidp = -1; | |
2618 | return NULL; | |
2619 | } | |
2620 | ||
9a305230 LS |
2621 | static void hugetlb_register_all_nodes(void) { } |
2622 | ||
2623 | #endif | |
2624 | ||
a3437870 NA |
2625 | static int __init hugetlb_init(void) |
2626 | { | |
8382d914 DB |
2627 | int i; |
2628 | ||
457c1b27 | 2629 | if (!hugepages_supported()) |
0ef89d25 | 2630 | return 0; |
a3437870 | 2631 | |
e11bfbfc NP |
2632 | if (!size_to_hstate(default_hstate_size)) { |
2633 | default_hstate_size = HPAGE_SIZE; | |
2634 | if (!size_to_hstate(default_hstate_size)) | |
2635 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2636 | } |
972dc4de | 2637 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2638 | if (default_hstate_max_huge_pages) { |
2639 | if (!default_hstate.max_huge_pages) | |
2640 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2641 | } | |
a3437870 NA |
2642 | |
2643 | hugetlb_init_hstates(); | |
aa888a74 | 2644 | gather_bootmem_prealloc(); |
a3437870 NA |
2645 | report_hugepages(); |
2646 | ||
2647 | hugetlb_sysfs_init(); | |
9a305230 | 2648 | hugetlb_register_all_nodes(); |
7179e7bf | 2649 | hugetlb_cgroup_file_init(); |
9a305230 | 2650 | |
8382d914 DB |
2651 | #ifdef CONFIG_SMP |
2652 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2653 | #else | |
2654 | num_fault_mutexes = 1; | |
2655 | #endif | |
c672c7f2 | 2656 | hugetlb_fault_mutex_table = |
8382d914 | 2657 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2658 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2659 | |
2660 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2661 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2662 | return 0; |
2663 | } | |
3e89e1c5 | 2664 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2665 | |
2666 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2667 | void __init hugetlb_bad_size(void) |
2668 | { | |
2669 | parsed_valid_hugepagesz = false; | |
2670 | } | |
2671 | ||
d00181b9 | 2672 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2673 | { |
2674 | struct hstate *h; | |
8faa8b07 AK |
2675 | unsigned long i; |
2676 | ||
a3437870 | 2677 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2678 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2679 | return; |
2680 | } | |
47d38344 | 2681 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2682 | BUG_ON(order == 0); |
47d38344 | 2683 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2684 | h->order = order; |
2685 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2686 | h->nr_huge_pages = 0; |
2687 | h->free_huge_pages = 0; | |
2688 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2689 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2690 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2691 | h->next_nid_to_alloc = first_memory_node; |
2692 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2693 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2694 | huge_page_size(h)/1024); | |
8faa8b07 | 2695 | |
a3437870 NA |
2696 | parsed_hstate = h; |
2697 | } | |
2698 | ||
e11bfbfc | 2699 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2700 | { |
2701 | unsigned long *mhp; | |
8faa8b07 | 2702 | static unsigned long *last_mhp; |
a3437870 | 2703 | |
9fee021d VT |
2704 | if (!parsed_valid_hugepagesz) { |
2705 | pr_warn("hugepages = %s preceded by " | |
2706 | "an unsupported hugepagesz, ignoring\n", s); | |
2707 | parsed_valid_hugepagesz = true; | |
2708 | return 1; | |
2709 | } | |
a3437870 | 2710 | /* |
47d38344 | 2711 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2712 | * so this hugepages= parameter goes to the "default hstate". |
2713 | */ | |
9fee021d | 2714 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2715 | mhp = &default_hstate_max_huge_pages; |
2716 | else | |
2717 | mhp = &parsed_hstate->max_huge_pages; | |
2718 | ||
8faa8b07 | 2719 | if (mhp == last_mhp) { |
598d8091 | 2720 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2721 | return 1; |
2722 | } | |
2723 | ||
a3437870 NA |
2724 | if (sscanf(s, "%lu", mhp) <= 0) |
2725 | *mhp = 0; | |
2726 | ||
8faa8b07 AK |
2727 | /* |
2728 | * Global state is always initialized later in hugetlb_init. | |
2729 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2730 | * use the bootmem allocator. | |
2731 | */ | |
47d38344 | 2732 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2733 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2734 | ||
2735 | last_mhp = mhp; | |
2736 | ||
a3437870 NA |
2737 | return 1; |
2738 | } | |
e11bfbfc NP |
2739 | __setup("hugepages=", hugetlb_nrpages_setup); |
2740 | ||
2741 | static int __init hugetlb_default_setup(char *s) | |
2742 | { | |
2743 | default_hstate_size = memparse(s, &s); | |
2744 | return 1; | |
2745 | } | |
2746 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2747 | |
8a213460 NA |
2748 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2749 | { | |
2750 | int node; | |
2751 | unsigned int nr = 0; | |
2752 | ||
2753 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2754 | nr += array[node]; | |
2755 | ||
2756 | return nr; | |
2757 | } | |
2758 | ||
2759 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2760 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2761 | struct ctl_table *table, int write, | |
2762 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2763 | { |
e5ff2159 | 2764 | struct hstate *h = &default_hstate; |
238d3c13 | 2765 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2766 | int ret; |
e5ff2159 | 2767 | |
457c1b27 | 2768 | if (!hugepages_supported()) |
86613628 | 2769 | return -EOPNOTSUPP; |
457c1b27 | 2770 | |
e5ff2159 AK |
2771 | table->data = &tmp; |
2772 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2773 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2774 | if (ret) | |
2775 | goto out; | |
e5ff2159 | 2776 | |
238d3c13 DR |
2777 | if (write) |
2778 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2779 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2780 | out: |
2781 | return ret; | |
1da177e4 | 2782 | } |
396faf03 | 2783 | |
06808b08 LS |
2784 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2785 | void __user *buffer, size_t *length, loff_t *ppos) | |
2786 | { | |
2787 | ||
2788 | return hugetlb_sysctl_handler_common(false, table, write, | |
2789 | buffer, length, ppos); | |
2790 | } | |
2791 | ||
2792 | #ifdef CONFIG_NUMA | |
2793 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2794 | void __user *buffer, size_t *length, loff_t *ppos) | |
2795 | { | |
2796 | return hugetlb_sysctl_handler_common(true, table, write, | |
2797 | buffer, length, ppos); | |
2798 | } | |
2799 | #endif /* CONFIG_NUMA */ | |
2800 | ||
a3d0c6aa | 2801 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2802 | void __user *buffer, |
a3d0c6aa NA |
2803 | size_t *length, loff_t *ppos) |
2804 | { | |
a5516438 | 2805 | struct hstate *h = &default_hstate; |
e5ff2159 | 2806 | unsigned long tmp; |
08d4a246 | 2807 | int ret; |
e5ff2159 | 2808 | |
457c1b27 | 2809 | if (!hugepages_supported()) |
86613628 | 2810 | return -EOPNOTSUPP; |
457c1b27 | 2811 | |
c033a93c | 2812 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2813 | |
bae7f4ae | 2814 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2815 | return -EINVAL; |
2816 | ||
e5ff2159 AK |
2817 | table->data = &tmp; |
2818 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2819 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2820 | if (ret) | |
2821 | goto out; | |
e5ff2159 AK |
2822 | |
2823 | if (write) { | |
2824 | spin_lock(&hugetlb_lock); | |
2825 | h->nr_overcommit_huge_pages = tmp; | |
2826 | spin_unlock(&hugetlb_lock); | |
2827 | } | |
08d4a246 MH |
2828 | out: |
2829 | return ret; | |
a3d0c6aa NA |
2830 | } |
2831 | ||
1da177e4 LT |
2832 | #endif /* CONFIG_SYSCTL */ |
2833 | ||
e1759c21 | 2834 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2835 | { |
a5516438 | 2836 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2837 | if (!hugepages_supported()) |
2838 | return; | |
e1759c21 | 2839 | seq_printf(m, |
4f98a2fe RR |
2840 | "HugePages_Total: %5lu\n" |
2841 | "HugePages_Free: %5lu\n" | |
2842 | "HugePages_Rsvd: %5lu\n" | |
2843 | "HugePages_Surp: %5lu\n" | |
2844 | "Hugepagesize: %8lu kB\n", | |
a5516438 AK |
2845 | h->nr_huge_pages, |
2846 | h->free_huge_pages, | |
2847 | h->resv_huge_pages, | |
2848 | h->surplus_huge_pages, | |
2849 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
1da177e4 LT |
2850 | } |
2851 | ||
2852 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
2853 | { | |
a5516438 | 2854 | struct hstate *h = &default_hstate; |
457c1b27 NA |
2855 | if (!hugepages_supported()) |
2856 | return 0; | |
1da177e4 LT |
2857 | return sprintf(buf, |
2858 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
2859 | "Node %d HugePages_Free: %5u\n" |
2860 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
2861 | nid, h->nr_huge_pages_node[nid], |
2862 | nid, h->free_huge_pages_node[nid], | |
2863 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
2864 | } |
2865 | ||
949f7ec5 DR |
2866 | void hugetlb_show_meminfo(void) |
2867 | { | |
2868 | struct hstate *h; | |
2869 | int nid; | |
2870 | ||
457c1b27 NA |
2871 | if (!hugepages_supported()) |
2872 | return; | |
2873 | ||
949f7ec5 DR |
2874 | for_each_node_state(nid, N_MEMORY) |
2875 | for_each_hstate(h) | |
2876 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
2877 | nid, | |
2878 | h->nr_huge_pages_node[nid], | |
2879 | h->free_huge_pages_node[nid], | |
2880 | h->surplus_huge_pages_node[nid], | |
2881 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
2882 | } | |
2883 | ||
5d317b2b NH |
2884 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
2885 | { | |
2886 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
2887 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
2888 | } | |
2889 | ||
1da177e4 LT |
2890 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
2891 | unsigned long hugetlb_total_pages(void) | |
2892 | { | |
d0028588 WL |
2893 | struct hstate *h; |
2894 | unsigned long nr_total_pages = 0; | |
2895 | ||
2896 | for_each_hstate(h) | |
2897 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
2898 | return nr_total_pages; | |
1da177e4 | 2899 | } |
1da177e4 | 2900 | |
a5516438 | 2901 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
2902 | { |
2903 | int ret = -ENOMEM; | |
2904 | ||
2905 | spin_lock(&hugetlb_lock); | |
2906 | /* | |
2907 | * When cpuset is configured, it breaks the strict hugetlb page | |
2908 | * reservation as the accounting is done on a global variable. Such | |
2909 | * reservation is completely rubbish in the presence of cpuset because | |
2910 | * the reservation is not checked against page availability for the | |
2911 | * current cpuset. Application can still potentially OOM'ed by kernel | |
2912 | * with lack of free htlb page in cpuset that the task is in. | |
2913 | * Attempt to enforce strict accounting with cpuset is almost | |
2914 | * impossible (or too ugly) because cpuset is too fluid that | |
2915 | * task or memory node can be dynamically moved between cpusets. | |
2916 | * | |
2917 | * The change of semantics for shared hugetlb mapping with cpuset is | |
2918 | * undesirable. However, in order to preserve some of the semantics, | |
2919 | * we fall back to check against current free page availability as | |
2920 | * a best attempt and hopefully to minimize the impact of changing | |
2921 | * semantics that cpuset has. | |
2922 | */ | |
2923 | if (delta > 0) { | |
a5516438 | 2924 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
2925 | goto out; |
2926 | ||
a5516438 AK |
2927 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
2928 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
2929 | goto out; |
2930 | } | |
2931 | } | |
2932 | ||
2933 | ret = 0; | |
2934 | if (delta < 0) | |
a5516438 | 2935 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
2936 | |
2937 | out: | |
2938 | spin_unlock(&hugetlb_lock); | |
2939 | return ret; | |
2940 | } | |
2941 | ||
84afd99b AW |
2942 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
2943 | { | |
f522c3ac | 2944 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
2945 | |
2946 | /* | |
2947 | * This new VMA should share its siblings reservation map if present. | |
2948 | * The VMA will only ever have a valid reservation map pointer where | |
2949 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 2950 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
2951 | * after this open call completes. It is therefore safe to take a |
2952 | * new reference here without additional locking. | |
2953 | */ | |
4e35f483 | 2954 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 2955 | kref_get(&resv->refs); |
84afd99b AW |
2956 | } |
2957 | ||
a1e78772 MG |
2958 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
2959 | { | |
a5516438 | 2960 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 2961 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 2962 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 2963 | unsigned long reserve, start, end; |
1c5ecae3 | 2964 | long gbl_reserve; |
84afd99b | 2965 | |
4e35f483 JK |
2966 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
2967 | return; | |
84afd99b | 2968 | |
4e35f483 JK |
2969 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
2970 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 2971 | |
4e35f483 | 2972 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 2973 | |
4e35f483 JK |
2974 | kref_put(&resv->refs, resv_map_release); |
2975 | ||
2976 | if (reserve) { | |
1c5ecae3 MK |
2977 | /* |
2978 | * Decrement reserve counts. The global reserve count may be | |
2979 | * adjusted if the subpool has a minimum size. | |
2980 | */ | |
2981 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
2982 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 2983 | } |
a1e78772 MG |
2984 | } |
2985 | ||
1da177e4 LT |
2986 | /* |
2987 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
2988 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
2989 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
2990 | * this far. | |
2991 | */ | |
d0217ac0 | 2992 | static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf) |
1da177e4 LT |
2993 | { |
2994 | BUG(); | |
d0217ac0 | 2995 | return 0; |
1da177e4 LT |
2996 | } |
2997 | ||
f0f37e2f | 2998 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 2999 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3000 | .open = hugetlb_vm_op_open, |
a1e78772 | 3001 | .close = hugetlb_vm_op_close, |
1da177e4 LT |
3002 | }; |
3003 | ||
1e8f889b DG |
3004 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3005 | int writable) | |
63551ae0 DG |
3006 | { |
3007 | pte_t entry; | |
3008 | ||
1e8f889b | 3009 | if (writable) { |
106c992a GS |
3010 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3011 | vma->vm_page_prot))); | |
63551ae0 | 3012 | } else { |
106c992a GS |
3013 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3014 | vma->vm_page_prot)); | |
63551ae0 DG |
3015 | } |
3016 | entry = pte_mkyoung(entry); | |
3017 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3018 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3019 | |
3020 | return entry; | |
3021 | } | |
3022 | ||
1e8f889b DG |
3023 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3024 | unsigned long address, pte_t *ptep) | |
3025 | { | |
3026 | pte_t entry; | |
3027 | ||
106c992a | 3028 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3029 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3030 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3031 | } |
3032 | ||
4a705fef NH |
3033 | static int is_hugetlb_entry_migration(pte_t pte) |
3034 | { | |
3035 | swp_entry_t swp; | |
3036 | ||
3037 | if (huge_pte_none(pte) || pte_present(pte)) | |
3038 | return 0; | |
3039 | swp = pte_to_swp_entry(pte); | |
3040 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
3041 | return 1; | |
3042 | else | |
3043 | return 0; | |
3044 | } | |
3045 | ||
3046 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3047 | { | |
3048 | swp_entry_t swp; | |
3049 | ||
3050 | if (huge_pte_none(pte) || pte_present(pte)) | |
3051 | return 0; | |
3052 | swp = pte_to_swp_entry(pte); | |
3053 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3054 | return 1; | |
3055 | else | |
3056 | return 0; | |
3057 | } | |
1e8f889b | 3058 | |
63551ae0 DG |
3059 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3060 | struct vm_area_struct *vma) | |
3061 | { | |
3062 | pte_t *src_pte, *dst_pte, entry; | |
3063 | struct page *ptepage; | |
1c59827d | 3064 | unsigned long addr; |
1e8f889b | 3065 | int cow; |
a5516438 AK |
3066 | struct hstate *h = hstate_vma(vma); |
3067 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
3068 | unsigned long mmun_start; /* For mmu_notifiers */ |
3069 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3070 | int ret = 0; | |
1e8f889b DG |
3071 | |
3072 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3073 | |
e8569dd2 AS |
3074 | mmun_start = vma->vm_start; |
3075 | mmun_end = vma->vm_end; | |
3076 | if (cow) | |
3077 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
3078 | ||
a5516438 | 3079 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3080 | spinlock_t *src_ptl, *dst_ptl; |
c74df32c HD |
3081 | src_pte = huge_pte_offset(src, addr); |
3082 | if (!src_pte) | |
3083 | continue; | |
a5516438 | 3084 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3085 | if (!dst_pte) { |
3086 | ret = -ENOMEM; | |
3087 | break; | |
3088 | } | |
c5c99429 LW |
3089 | |
3090 | /* If the pagetables are shared don't copy or take references */ | |
3091 | if (dst_pte == src_pte) | |
3092 | continue; | |
3093 | ||
cb900f41 KS |
3094 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3095 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3096 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3097 | entry = huge_ptep_get(src_pte); |
3098 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3099 | ; | |
3100 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3101 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3102 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3103 | ||
3104 | if (is_write_migration_entry(swp_entry) && cow) { | |
3105 | /* | |
3106 | * COW mappings require pages in both | |
3107 | * parent and child to be set to read. | |
3108 | */ | |
3109 | make_migration_entry_read(&swp_entry); | |
3110 | entry = swp_entry_to_pte(swp_entry); | |
3111 | set_huge_pte_at(src, addr, src_pte, entry); | |
3112 | } | |
3113 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
3114 | } else { | |
34ee645e | 3115 | if (cow) { |
7f2e9525 | 3116 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e JR |
3117 | mmu_notifier_invalidate_range(src, mmun_start, |
3118 | mmun_end); | |
3119 | } | |
0253d634 | 3120 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3121 | ptepage = pte_page(entry); |
3122 | get_page(ptepage); | |
53f9263b | 3123 | page_dup_rmap(ptepage, true); |
1c59827d | 3124 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3125 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3126 | } |
cb900f41 KS |
3127 | spin_unlock(src_ptl); |
3128 | spin_unlock(dst_ptl); | |
63551ae0 | 3129 | } |
63551ae0 | 3130 | |
e8569dd2 AS |
3131 | if (cow) |
3132 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3133 | ||
3134 | return ret; | |
63551ae0 DG |
3135 | } |
3136 | ||
24669e58 AK |
3137 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3138 | unsigned long start, unsigned long end, | |
3139 | struct page *ref_page) | |
63551ae0 | 3140 | { |
24669e58 | 3141 | int force_flush = 0; |
63551ae0 DG |
3142 | struct mm_struct *mm = vma->vm_mm; |
3143 | unsigned long address; | |
c7546f8f | 3144 | pte_t *ptep; |
63551ae0 | 3145 | pte_t pte; |
cb900f41 | 3146 | spinlock_t *ptl; |
63551ae0 | 3147 | struct page *page; |
a5516438 AK |
3148 | struct hstate *h = hstate_vma(vma); |
3149 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3150 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3151 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3152 | |
63551ae0 | 3153 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3154 | BUG_ON(start & ~huge_page_mask(h)); |
3155 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3156 | |
24669e58 | 3157 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3158 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3159 | address = start; |
24669e58 | 3160 | again: |
569f48b8 | 3161 | for (; address < end; address += sz) { |
c7546f8f | 3162 | ptep = huge_pte_offset(mm, address); |
4c887265 | 3163 | if (!ptep) |
c7546f8f DG |
3164 | continue; |
3165 | ||
cb900f41 | 3166 | ptl = huge_pte_lock(h, mm, ptep); |
39dde65c | 3167 | if (huge_pmd_unshare(mm, &address, ptep)) |
cb900f41 | 3168 | goto unlock; |
39dde65c | 3169 | |
6629326b HD |
3170 | pte = huge_ptep_get(ptep); |
3171 | if (huge_pte_none(pte)) | |
cb900f41 | 3172 | goto unlock; |
6629326b HD |
3173 | |
3174 | /* | |
9fbc1f63 NH |
3175 | * Migrating hugepage or HWPoisoned hugepage is already |
3176 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3177 | */ |
9fbc1f63 | 3178 | if (unlikely(!pte_present(pte))) { |
106c992a | 3179 | huge_pte_clear(mm, address, ptep); |
cb900f41 | 3180 | goto unlock; |
8c4894c6 | 3181 | } |
6629326b HD |
3182 | |
3183 | page = pte_page(pte); | |
04f2cbe3 MG |
3184 | /* |
3185 | * If a reference page is supplied, it is because a specific | |
3186 | * page is being unmapped, not a range. Ensure the page we | |
3187 | * are about to unmap is the actual page of interest. | |
3188 | */ | |
3189 | if (ref_page) { | |
04f2cbe3 | 3190 | if (page != ref_page) |
cb900f41 | 3191 | goto unlock; |
04f2cbe3 MG |
3192 | |
3193 | /* | |
3194 | * Mark the VMA as having unmapped its page so that | |
3195 | * future faults in this VMA will fail rather than | |
3196 | * looking like data was lost | |
3197 | */ | |
3198 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3199 | } | |
3200 | ||
c7546f8f | 3201 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
24669e58 | 3202 | tlb_remove_tlb_entry(tlb, ptep, address); |
106c992a | 3203 | if (huge_pte_dirty(pte)) |
6649a386 | 3204 | set_page_dirty(page); |
9e81130b | 3205 | |
5d317b2b | 3206 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3207 | page_remove_rmap(page, true); |
24669e58 | 3208 | force_flush = !__tlb_remove_page(tlb, page); |
cb900f41 | 3209 | if (force_flush) { |
569f48b8 | 3210 | address += sz; |
cb900f41 | 3211 | spin_unlock(ptl); |
24669e58 | 3212 | break; |
cb900f41 | 3213 | } |
9e81130b | 3214 | /* Bail out after unmapping reference page if supplied */ |
cb900f41 KS |
3215 | if (ref_page) { |
3216 | spin_unlock(ptl); | |
9e81130b | 3217 | break; |
cb900f41 KS |
3218 | } |
3219 | unlock: | |
3220 | spin_unlock(ptl); | |
63551ae0 | 3221 | } |
24669e58 AK |
3222 | /* |
3223 | * mmu_gather ran out of room to batch pages, we break out of | |
3224 | * the PTE lock to avoid doing the potential expensive TLB invalidate | |
3225 | * and page-free while holding it. | |
3226 | */ | |
3227 | if (force_flush) { | |
3228 | force_flush = 0; | |
3229 | tlb_flush_mmu(tlb); | |
3230 | if (address < end && !ref_page) | |
3231 | goto again; | |
fe1668ae | 3232 | } |
2ec74c3e | 3233 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3234 | tlb_end_vma(tlb, vma); |
1da177e4 | 3235 | } |
63551ae0 | 3236 | |
d833352a MG |
3237 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3238 | struct vm_area_struct *vma, unsigned long start, | |
3239 | unsigned long end, struct page *ref_page) | |
3240 | { | |
3241 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3242 | ||
3243 | /* | |
3244 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3245 | * test will fail on a vma being torn down, and not grab a page table | |
3246 | * on its way out. We're lucky that the flag has such an appropriate | |
3247 | * name, and can in fact be safely cleared here. We could clear it | |
3248 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3249 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3250 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3251 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3252 | */ |
3253 | vma->vm_flags &= ~VM_MAYSHARE; | |
3254 | } | |
3255 | ||
502717f4 | 3256 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3257 | unsigned long end, struct page *ref_page) |
502717f4 | 3258 | { |
24669e58 AK |
3259 | struct mm_struct *mm; |
3260 | struct mmu_gather tlb; | |
3261 | ||
3262 | mm = vma->vm_mm; | |
3263 | ||
2b047252 | 3264 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3265 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3266 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
3267 | } |
3268 | ||
04f2cbe3 MG |
3269 | /* |
3270 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3271 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3272 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3273 | * same region. | |
3274 | */ | |
2f4612af DB |
3275 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3276 | struct page *page, unsigned long address) | |
04f2cbe3 | 3277 | { |
7526674d | 3278 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3279 | struct vm_area_struct *iter_vma; |
3280 | struct address_space *mapping; | |
04f2cbe3 MG |
3281 | pgoff_t pgoff; |
3282 | ||
3283 | /* | |
3284 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3285 | * from page cache lookup which is in HPAGE_SIZE units. | |
3286 | */ | |
7526674d | 3287 | address = address & huge_page_mask(h); |
36e4f20a MH |
3288 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3289 | vma->vm_pgoff; | |
496ad9aa | 3290 | mapping = file_inode(vma->vm_file)->i_mapping; |
04f2cbe3 | 3291 | |
4eb2b1dc MG |
3292 | /* |
3293 | * Take the mapping lock for the duration of the table walk. As | |
3294 | * this mapping should be shared between all the VMAs, | |
3295 | * __unmap_hugepage_range() is called as the lock is already held | |
3296 | */ | |
83cde9e8 | 3297 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3298 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3299 | /* Do not unmap the current VMA */ |
3300 | if (iter_vma == vma) | |
3301 | continue; | |
3302 | ||
2f84a899 MG |
3303 | /* |
3304 | * Shared VMAs have their own reserves and do not affect | |
3305 | * MAP_PRIVATE accounting but it is possible that a shared | |
3306 | * VMA is using the same page so check and skip such VMAs. | |
3307 | */ | |
3308 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3309 | continue; | |
3310 | ||
04f2cbe3 MG |
3311 | /* |
3312 | * Unmap the page from other VMAs without their own reserves. | |
3313 | * They get marked to be SIGKILLed if they fault in these | |
3314 | * areas. This is because a future no-page fault on this VMA | |
3315 | * could insert a zeroed page instead of the data existing | |
3316 | * from the time of fork. This would look like data corruption | |
3317 | */ | |
3318 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3319 | unmap_hugepage_range(iter_vma, address, |
3320 | address + huge_page_size(h), page); | |
04f2cbe3 | 3321 | } |
83cde9e8 | 3322 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3323 | } |
3324 | ||
0fe6e20b NH |
3325 | /* |
3326 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3327 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3328 | * cannot race with other handlers or page migration. | |
3329 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3330 | */ |
1e8f889b | 3331 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
04f2cbe3 | 3332 | unsigned long address, pte_t *ptep, pte_t pte, |
cb900f41 | 3333 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3334 | { |
a5516438 | 3335 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3336 | struct page *old_page, *new_page; |
ad4404a2 | 3337 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3338 | unsigned long mmun_start; /* For mmu_notifiers */ |
3339 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b DG |
3340 | |
3341 | old_page = pte_page(pte); | |
3342 | ||
04f2cbe3 | 3343 | retry_avoidcopy: |
1e8f889b DG |
3344 | /* If no-one else is actually using this page, avoid the copy |
3345 | * and just make the page writable */ | |
37a2140d JK |
3346 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
3347 | page_move_anon_rmap(old_page, vma, address); | |
1e8f889b | 3348 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3349 | return 0; |
1e8f889b DG |
3350 | } |
3351 | ||
04f2cbe3 MG |
3352 | /* |
3353 | * If the process that created a MAP_PRIVATE mapping is about to | |
3354 | * perform a COW due to a shared page count, attempt to satisfy | |
3355 | * the allocation without using the existing reserves. The pagecache | |
3356 | * page is used to determine if the reserve at this address was | |
3357 | * consumed or not. If reserves were used, a partial faulted mapping | |
3358 | * at the time of fork() could consume its reserves on COW instead | |
3359 | * of the full address range. | |
3360 | */ | |
5944d011 | 3361 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3362 | old_page != pagecache_page) |
3363 | outside_reserve = 1; | |
3364 | ||
09cbfeaf | 3365 | get_page(old_page); |
b76c8cfb | 3366 | |
ad4404a2 DB |
3367 | /* |
3368 | * Drop page table lock as buddy allocator may be called. It will | |
3369 | * be acquired again before returning to the caller, as expected. | |
3370 | */ | |
cb900f41 | 3371 | spin_unlock(ptl); |
04f2cbe3 | 3372 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3373 | |
2fc39cec | 3374 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3375 | /* |
3376 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3377 | * it is due to references held by a child and an insufficient | |
3378 | * huge page pool. To guarantee the original mappers | |
3379 | * reliability, unmap the page from child processes. The child | |
3380 | * may get SIGKILLed if it later faults. | |
3381 | */ | |
3382 | if (outside_reserve) { | |
09cbfeaf | 3383 | put_page(old_page); |
04f2cbe3 | 3384 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3385 | unmap_ref_private(mm, vma, old_page, address); |
3386 | BUG_ON(huge_pte_none(pte)); | |
3387 | spin_lock(ptl); | |
3388 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); | |
3389 | if (likely(ptep && | |
3390 | pte_same(huge_ptep_get(ptep), pte))) | |
3391 | goto retry_avoidcopy; | |
3392 | /* | |
3393 | * race occurs while re-acquiring page table | |
3394 | * lock, and our job is done. | |
3395 | */ | |
3396 | return 0; | |
04f2cbe3 MG |
3397 | } |
3398 | ||
ad4404a2 DB |
3399 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3400 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3401 | goto out_release_old; | |
1e8f889b DG |
3402 | } |
3403 | ||
0fe6e20b NH |
3404 | /* |
3405 | * When the original hugepage is shared one, it does not have | |
3406 | * anon_vma prepared. | |
3407 | */ | |
44e2aa93 | 3408 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3409 | ret = VM_FAULT_OOM; |
3410 | goto out_release_all; | |
44e2aa93 | 3411 | } |
0fe6e20b | 3412 | |
47ad8475 AA |
3413 | copy_user_huge_page(new_page, old_page, address, vma, |
3414 | pages_per_huge_page(h)); | |
0ed361de | 3415 | __SetPageUptodate(new_page); |
bcc54222 | 3416 | set_page_huge_active(new_page); |
1e8f889b | 3417 | |
2ec74c3e SG |
3418 | mmun_start = address & huge_page_mask(h); |
3419 | mmun_end = mmun_start + huge_page_size(h); | |
3420 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3421 | |
b76c8cfb | 3422 | /* |
cb900f41 | 3423 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3424 | * before the page tables are altered |
3425 | */ | |
cb900f41 | 3426 | spin_lock(ptl); |
a5516438 | 3427 | ptep = huge_pte_offset(mm, address & huge_page_mask(h)); |
a9af0c5d | 3428 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3429 | ClearPagePrivate(new_page); |
3430 | ||
1e8f889b | 3431 | /* Break COW */ |
8fe627ec | 3432 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3433 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3434 | set_huge_pte_at(mm, address, ptep, |
3435 | make_huge_pte(vma, new_page, 1)); | |
d281ee61 | 3436 | page_remove_rmap(old_page, true); |
cd67f0d2 | 3437 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3438 | /* Make the old page be freed below */ |
3439 | new_page = old_page; | |
3440 | } | |
cb900f41 | 3441 | spin_unlock(ptl); |
2ec74c3e | 3442 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3443 | out_release_all: |
09cbfeaf | 3444 | put_page(new_page); |
ad4404a2 | 3445 | out_release_old: |
09cbfeaf | 3446 | put_page(old_page); |
8312034f | 3447 | |
ad4404a2 DB |
3448 | spin_lock(ptl); /* Caller expects lock to be held */ |
3449 | return ret; | |
1e8f889b DG |
3450 | } |
3451 | ||
04f2cbe3 | 3452 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3453 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3454 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3455 | { |
3456 | struct address_space *mapping; | |
e7c4b0bf | 3457 | pgoff_t idx; |
04f2cbe3 MG |
3458 | |
3459 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3460 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3461 | |
3462 | return find_lock_page(mapping, idx); | |
3463 | } | |
3464 | ||
3ae77f43 HD |
3465 | /* |
3466 | * Return whether there is a pagecache page to back given address within VMA. | |
3467 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3468 | */ | |
3469 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3470 | struct vm_area_struct *vma, unsigned long address) |
3471 | { | |
3472 | struct address_space *mapping; | |
3473 | pgoff_t idx; | |
3474 | struct page *page; | |
3475 | ||
3476 | mapping = vma->vm_file->f_mapping; | |
3477 | idx = vma_hugecache_offset(h, vma, address); | |
3478 | ||
3479 | page = find_get_page(mapping, idx); | |
3480 | if (page) | |
3481 | put_page(page); | |
3482 | return page != NULL; | |
3483 | } | |
3484 | ||
ab76ad54 MK |
3485 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3486 | pgoff_t idx) | |
3487 | { | |
3488 | struct inode *inode = mapping->host; | |
3489 | struct hstate *h = hstate_inode(inode); | |
3490 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3491 | ||
3492 | if (err) | |
3493 | return err; | |
3494 | ClearPagePrivate(page); | |
3495 | ||
3496 | spin_lock(&inode->i_lock); | |
3497 | inode->i_blocks += blocks_per_huge_page(h); | |
3498 | spin_unlock(&inode->i_lock); | |
3499 | return 0; | |
3500 | } | |
3501 | ||
a1ed3dda | 3502 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3503 | struct address_space *mapping, pgoff_t idx, |
3504 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3505 | { |
a5516438 | 3506 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3507 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3508 | int anon_rmap = 0; |
4c887265 | 3509 | unsigned long size; |
4c887265 | 3510 | struct page *page; |
1e8f889b | 3511 | pte_t new_pte; |
cb900f41 | 3512 | spinlock_t *ptl; |
4c887265 | 3513 | |
04f2cbe3 MG |
3514 | /* |
3515 | * Currently, we are forced to kill the process in the event the | |
3516 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3517 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3518 | */ |
3519 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3520 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3521 | current->pid); |
04f2cbe3 MG |
3522 | return ret; |
3523 | } | |
3524 | ||
4c887265 AL |
3525 | /* |
3526 | * Use page lock to guard against racing truncation | |
3527 | * before we get page_table_lock. | |
3528 | */ | |
6bda666a CL |
3529 | retry: |
3530 | page = find_lock_page(mapping, idx); | |
3531 | if (!page) { | |
a5516438 | 3532 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3533 | if (idx >= size) |
3534 | goto out; | |
04f2cbe3 | 3535 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3536 | if (IS_ERR(page)) { |
76dcee75 AK |
3537 | ret = PTR_ERR(page); |
3538 | if (ret == -ENOMEM) | |
3539 | ret = VM_FAULT_OOM; | |
3540 | else | |
3541 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3542 | goto out; |
3543 | } | |
47ad8475 | 3544 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3545 | __SetPageUptodate(page); |
bcc54222 | 3546 | set_page_huge_active(page); |
ac9b9c66 | 3547 | |
f83a275d | 3548 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3549 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3550 | if (err) { |
3551 | put_page(page); | |
6bda666a CL |
3552 | if (err == -EEXIST) |
3553 | goto retry; | |
3554 | goto out; | |
3555 | } | |
23be7468 | 3556 | } else { |
6bda666a | 3557 | lock_page(page); |
0fe6e20b NH |
3558 | if (unlikely(anon_vma_prepare(vma))) { |
3559 | ret = VM_FAULT_OOM; | |
3560 | goto backout_unlocked; | |
3561 | } | |
409eb8c2 | 3562 | anon_rmap = 1; |
23be7468 | 3563 | } |
0fe6e20b | 3564 | } else { |
998b4382 NH |
3565 | /* |
3566 | * If memory error occurs between mmap() and fault, some process | |
3567 | * don't have hwpoisoned swap entry for errored virtual address. | |
3568 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3569 | */ | |
3570 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3571 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3572 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3573 | goto backout_unlocked; |
3574 | } | |
6bda666a | 3575 | } |
1e8f889b | 3576 | |
57303d80 AW |
3577 | /* |
3578 | * If we are going to COW a private mapping later, we examine the | |
3579 | * pending reservations for this page now. This will ensure that | |
3580 | * any allocations necessary to record that reservation occur outside | |
3581 | * the spinlock. | |
3582 | */ | |
5e911373 | 3583 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3584 | if (vma_needs_reservation(h, vma, address) < 0) { |
3585 | ret = VM_FAULT_OOM; | |
3586 | goto backout_unlocked; | |
3587 | } | |
5e911373 | 3588 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3589 | vma_end_reservation(h, vma, address); |
5e911373 | 3590 | } |
57303d80 | 3591 | |
cb900f41 KS |
3592 | ptl = huge_pte_lockptr(h, mm, ptep); |
3593 | spin_lock(ptl); | |
a5516438 | 3594 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3595 | if (idx >= size) |
3596 | goto backout; | |
3597 | ||
83c54070 | 3598 | ret = 0; |
7f2e9525 | 3599 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3600 | goto backout; |
3601 | ||
07443a85 JK |
3602 | if (anon_rmap) { |
3603 | ClearPagePrivate(page); | |
409eb8c2 | 3604 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3605 | } else |
53f9263b | 3606 | page_dup_rmap(page, true); |
1e8f889b DG |
3607 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3608 | && (vma->vm_flags & VM_SHARED))); | |
3609 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3610 | ||
5d317b2b | 3611 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3612 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3613 | /* Optimization, do the COW without a second fault */ |
cb900f41 | 3614 | ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page, ptl); |
1e8f889b DG |
3615 | } |
3616 | ||
cb900f41 | 3617 | spin_unlock(ptl); |
4c887265 AL |
3618 | unlock_page(page); |
3619 | out: | |
ac9b9c66 | 3620 | return ret; |
4c887265 AL |
3621 | |
3622 | backout: | |
cb900f41 | 3623 | spin_unlock(ptl); |
2b26736c | 3624 | backout_unlocked: |
4c887265 AL |
3625 | unlock_page(page); |
3626 | put_page(page); | |
3627 | goto out; | |
ac9b9c66 HD |
3628 | } |
3629 | ||
8382d914 | 3630 | #ifdef CONFIG_SMP |
c672c7f2 | 3631 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3632 | struct vm_area_struct *vma, |
3633 | struct address_space *mapping, | |
3634 | pgoff_t idx, unsigned long address) | |
3635 | { | |
3636 | unsigned long key[2]; | |
3637 | u32 hash; | |
3638 | ||
3639 | if (vma->vm_flags & VM_SHARED) { | |
3640 | key[0] = (unsigned long) mapping; | |
3641 | key[1] = idx; | |
3642 | } else { | |
3643 | key[0] = (unsigned long) mm; | |
3644 | key[1] = address >> huge_page_shift(h); | |
3645 | } | |
3646 | ||
3647 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3648 | ||
3649 | return hash & (num_fault_mutexes - 1); | |
3650 | } | |
3651 | #else | |
3652 | /* | |
3653 | * For uniprocesor systems we always use a single mutex, so just | |
3654 | * return 0 and avoid the hashing overhead. | |
3655 | */ | |
c672c7f2 | 3656 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3657 | struct vm_area_struct *vma, |
3658 | struct address_space *mapping, | |
3659 | pgoff_t idx, unsigned long address) | |
3660 | { | |
3661 | return 0; | |
3662 | } | |
3663 | #endif | |
3664 | ||
86e5216f | 3665 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3666 | unsigned long address, unsigned int flags) |
86e5216f | 3667 | { |
8382d914 | 3668 | pte_t *ptep, entry; |
cb900f41 | 3669 | spinlock_t *ptl; |
1e8f889b | 3670 | int ret; |
8382d914 DB |
3671 | u32 hash; |
3672 | pgoff_t idx; | |
0fe6e20b | 3673 | struct page *page = NULL; |
57303d80 | 3674 | struct page *pagecache_page = NULL; |
a5516438 | 3675 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3676 | struct address_space *mapping; |
0f792cf9 | 3677 | int need_wait_lock = 0; |
86e5216f | 3678 | |
1e16a539 KH |
3679 | address &= huge_page_mask(h); |
3680 | ||
fd6a03ed NH |
3681 | ptep = huge_pte_offset(mm, address); |
3682 | if (ptep) { | |
3683 | entry = huge_ptep_get(ptep); | |
290408d4 | 3684 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3685 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3686 | return 0; |
3687 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3688 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3689 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
0d777df5 NH |
3690 | } else { |
3691 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); | |
3692 | if (!ptep) | |
3693 | return VM_FAULT_OOM; | |
fd6a03ed NH |
3694 | } |
3695 | ||
8382d914 DB |
3696 | mapping = vma->vm_file->f_mapping; |
3697 | idx = vma_hugecache_offset(h, vma, address); | |
3698 | ||
3935baa9 DG |
3699 | /* |
3700 | * Serialize hugepage allocation and instantiation, so that we don't | |
3701 | * get spurious allocation failures if two CPUs race to instantiate | |
3702 | * the same page in the page cache. | |
3703 | */ | |
c672c7f2 MK |
3704 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3705 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3706 | |
7f2e9525 GS |
3707 | entry = huge_ptep_get(ptep); |
3708 | if (huge_pte_none(entry)) { | |
8382d914 | 3709 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3710 | goto out_mutex; |
3935baa9 | 3711 | } |
86e5216f | 3712 | |
83c54070 | 3713 | ret = 0; |
1e8f889b | 3714 | |
0f792cf9 NH |
3715 | /* |
3716 | * entry could be a migration/hwpoison entry at this point, so this | |
3717 | * check prevents the kernel from going below assuming that we have | |
3718 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3719 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3720 | * handle it. | |
3721 | */ | |
3722 | if (!pte_present(entry)) | |
3723 | goto out_mutex; | |
3724 | ||
57303d80 AW |
3725 | /* |
3726 | * If we are going to COW the mapping later, we examine the pending | |
3727 | * reservations for this page now. This will ensure that any | |
3728 | * allocations necessary to record that reservation occur outside the | |
3729 | * spinlock. For private mappings, we also lookup the pagecache | |
3730 | * page now as it is used to determine if a reservation has been | |
3731 | * consumed. | |
3732 | */ | |
106c992a | 3733 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3734 | if (vma_needs_reservation(h, vma, address) < 0) { |
3735 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3736 | goto out_mutex; |
2b26736c | 3737 | } |
5e911373 | 3738 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3739 | vma_end_reservation(h, vma, address); |
57303d80 | 3740 | |
f83a275d | 3741 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3742 | pagecache_page = hugetlbfs_pagecache_page(h, |
3743 | vma, address); | |
3744 | } | |
3745 | ||
0f792cf9 NH |
3746 | ptl = huge_pte_lock(h, mm, ptep); |
3747 | ||
3748 | /* Check for a racing update before calling hugetlb_cow */ | |
3749 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3750 | goto out_ptl; | |
3751 | ||
56c9cfb1 NH |
3752 | /* |
3753 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3754 | * pagecache_page, so here we need take the former one | |
3755 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3756 | */ |
3757 | page = pte_page(entry); | |
3758 | if (page != pagecache_page) | |
0f792cf9 NH |
3759 | if (!trylock_page(page)) { |
3760 | need_wait_lock = 1; | |
3761 | goto out_ptl; | |
3762 | } | |
b4d1d99f | 3763 | |
0f792cf9 | 3764 | get_page(page); |
b4d1d99f | 3765 | |
788c7df4 | 3766 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3767 | if (!huge_pte_write(entry)) { |
57303d80 | 3768 | ret = hugetlb_cow(mm, vma, address, ptep, entry, |
cb900f41 | 3769 | pagecache_page, ptl); |
0f792cf9 | 3770 | goto out_put_page; |
b4d1d99f | 3771 | } |
106c992a | 3772 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3773 | } |
3774 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3775 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3776 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3777 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3778 | out_put_page: |
3779 | if (page != pagecache_page) | |
3780 | unlock_page(page); | |
3781 | put_page(page); | |
cb900f41 KS |
3782 | out_ptl: |
3783 | spin_unlock(ptl); | |
57303d80 AW |
3784 | |
3785 | if (pagecache_page) { | |
3786 | unlock_page(pagecache_page); | |
3787 | put_page(pagecache_page); | |
3788 | } | |
b4d1d99f | 3789 | out_mutex: |
c672c7f2 | 3790 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3791 | /* |
3792 | * Generally it's safe to hold refcount during waiting page lock. But | |
3793 | * here we just wait to defer the next page fault to avoid busy loop and | |
3794 | * the page is not used after unlocked before returning from the current | |
3795 | * page fault. So we are safe from accessing freed page, even if we wait | |
3796 | * here without taking refcount. | |
3797 | */ | |
3798 | if (need_wait_lock) | |
3799 | wait_on_page_locked(page); | |
1e8f889b | 3800 | return ret; |
86e5216f AL |
3801 | } |
3802 | ||
28a35716 ML |
3803 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3804 | struct page **pages, struct vm_area_struct **vmas, | |
3805 | unsigned long *position, unsigned long *nr_pages, | |
3806 | long i, unsigned int flags) | |
63551ae0 | 3807 | { |
d5d4b0aa CK |
3808 | unsigned long pfn_offset; |
3809 | unsigned long vaddr = *position; | |
28a35716 | 3810 | unsigned long remainder = *nr_pages; |
a5516438 | 3811 | struct hstate *h = hstate_vma(vma); |
63551ae0 | 3812 | |
63551ae0 | 3813 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 3814 | pte_t *pte; |
cb900f41 | 3815 | spinlock_t *ptl = NULL; |
2a15efc9 | 3816 | int absent; |
4c887265 | 3817 | struct page *page; |
63551ae0 | 3818 | |
02057967 DR |
3819 | /* |
3820 | * If we have a pending SIGKILL, don't keep faulting pages and | |
3821 | * potentially allocating memory. | |
3822 | */ | |
3823 | if (unlikely(fatal_signal_pending(current))) { | |
3824 | remainder = 0; | |
3825 | break; | |
3826 | } | |
3827 | ||
4c887265 AL |
3828 | /* |
3829 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 3830 | * each hugepage. We have to make sure we get the |
4c887265 | 3831 | * first, for the page indexing below to work. |
cb900f41 KS |
3832 | * |
3833 | * Note that page table lock is not held when pte is null. | |
4c887265 | 3834 | */ |
a5516438 | 3835 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h)); |
cb900f41 KS |
3836 | if (pte) |
3837 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
3838 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
3839 | ||
3840 | /* | |
3841 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
3842 | * an error where there's an empty slot with no huge pagecache |
3843 | * to back it. This way, we avoid allocating a hugepage, and | |
3844 | * the sparse dumpfile avoids allocating disk blocks, but its | |
3845 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 3846 | */ |
3ae77f43 HD |
3847 | if (absent && (flags & FOLL_DUMP) && |
3848 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
3849 | if (pte) |
3850 | spin_unlock(ptl); | |
2a15efc9 HD |
3851 | remainder = 0; |
3852 | break; | |
3853 | } | |
63551ae0 | 3854 | |
9cc3a5bd NH |
3855 | /* |
3856 | * We need call hugetlb_fault for both hugepages under migration | |
3857 | * (in which case hugetlb_fault waits for the migration,) and | |
3858 | * hwpoisoned hugepages (in which case we need to prevent the | |
3859 | * caller from accessing to them.) In order to do this, we use | |
3860 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
3861 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
3862 | * both cases, and because we can't follow correct pages | |
3863 | * directly from any kind of swap entries. | |
3864 | */ | |
3865 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
3866 | ((flags & FOLL_WRITE) && |
3867 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 3868 | int ret; |
63551ae0 | 3869 | |
cb900f41 KS |
3870 | if (pte) |
3871 | spin_unlock(ptl); | |
2a15efc9 HD |
3872 | ret = hugetlb_fault(mm, vma, vaddr, |
3873 | (flags & FOLL_WRITE) ? FAULT_FLAG_WRITE : 0); | |
a89182c7 | 3874 | if (!(ret & VM_FAULT_ERROR)) |
4c887265 | 3875 | continue; |
63551ae0 | 3876 | |
4c887265 | 3877 | remainder = 0; |
4c887265 AL |
3878 | break; |
3879 | } | |
3880 | ||
a5516438 | 3881 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 3882 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 3883 | same_page: |
d6692183 | 3884 | if (pages) { |
2a15efc9 | 3885 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 3886 | get_page(pages[i]); |
d6692183 | 3887 | } |
63551ae0 DG |
3888 | |
3889 | if (vmas) | |
3890 | vmas[i] = vma; | |
3891 | ||
3892 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 3893 | ++pfn_offset; |
63551ae0 DG |
3894 | --remainder; |
3895 | ++i; | |
d5d4b0aa | 3896 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 3897 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
3898 | /* |
3899 | * We use pfn_offset to avoid touching the pageframes | |
3900 | * of this compound page. | |
3901 | */ | |
3902 | goto same_page; | |
3903 | } | |
cb900f41 | 3904 | spin_unlock(ptl); |
63551ae0 | 3905 | } |
28a35716 | 3906 | *nr_pages = remainder; |
63551ae0 DG |
3907 | *position = vaddr; |
3908 | ||
2a15efc9 | 3909 | return i ? i : -EFAULT; |
63551ae0 | 3910 | } |
8f860591 | 3911 | |
7da4d641 | 3912 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
3913 | unsigned long address, unsigned long end, pgprot_t newprot) |
3914 | { | |
3915 | struct mm_struct *mm = vma->vm_mm; | |
3916 | unsigned long start = address; | |
3917 | pte_t *ptep; | |
3918 | pte_t pte; | |
a5516438 | 3919 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 3920 | unsigned long pages = 0; |
8f860591 ZY |
3921 | |
3922 | BUG_ON(address >= end); | |
3923 | flush_cache_range(vma, address, end); | |
3924 | ||
a5338093 | 3925 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 3926 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 3927 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 3928 | spinlock_t *ptl; |
8f860591 ZY |
3929 | ptep = huge_pte_offset(mm, address); |
3930 | if (!ptep) | |
3931 | continue; | |
cb900f41 | 3932 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
3933 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3934 | pages++; | |
cb900f41 | 3935 | spin_unlock(ptl); |
39dde65c | 3936 | continue; |
7da4d641 | 3937 | } |
a8bda28d NH |
3938 | pte = huge_ptep_get(ptep); |
3939 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
3940 | spin_unlock(ptl); | |
3941 | continue; | |
3942 | } | |
3943 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
3944 | swp_entry_t entry = pte_to_swp_entry(pte); | |
3945 | ||
3946 | if (is_write_migration_entry(entry)) { | |
3947 | pte_t newpte; | |
3948 | ||
3949 | make_migration_entry_read(&entry); | |
3950 | newpte = swp_entry_to_pte(entry); | |
3951 | set_huge_pte_at(mm, address, ptep, newpte); | |
3952 | pages++; | |
3953 | } | |
3954 | spin_unlock(ptl); | |
3955 | continue; | |
3956 | } | |
3957 | if (!huge_pte_none(pte)) { | |
8f860591 | 3958 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 3959 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 3960 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 3961 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 3962 | pages++; |
8f860591 | 3963 | } |
cb900f41 | 3964 | spin_unlock(ptl); |
8f860591 | 3965 | } |
d833352a | 3966 | /* |
c8c06efa | 3967 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 3968 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 3969 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
3970 | * and that page table be reused and filled with junk. |
3971 | */ | |
8f860591 | 3972 | flush_tlb_range(vma, start, end); |
34ee645e | 3973 | mmu_notifier_invalidate_range(mm, start, end); |
83cde9e8 | 3974 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 3975 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
3976 | |
3977 | return pages << h->order; | |
8f860591 ZY |
3978 | } |
3979 | ||
a1e78772 MG |
3980 | int hugetlb_reserve_pages(struct inode *inode, |
3981 | long from, long to, | |
5a6fe125 | 3982 | struct vm_area_struct *vma, |
ca16d140 | 3983 | vm_flags_t vm_flags) |
e4e574b7 | 3984 | { |
17c9d12e | 3985 | long ret, chg; |
a5516438 | 3986 | struct hstate *h = hstate_inode(inode); |
90481622 | 3987 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 3988 | struct resv_map *resv_map; |
1c5ecae3 | 3989 | long gbl_reserve; |
e4e574b7 | 3990 | |
17c9d12e MG |
3991 | /* |
3992 | * Only apply hugepage reservation if asked. At fault time, an | |
3993 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 3994 | * without using reserves |
17c9d12e | 3995 | */ |
ca16d140 | 3996 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
3997 | return 0; |
3998 | ||
a1e78772 MG |
3999 | /* |
4000 | * Shared mappings base their reservation on the number of pages that | |
4001 | * are already allocated on behalf of the file. Private mappings need | |
4002 | * to reserve the full area even if read-only as mprotect() may be | |
4003 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4004 | */ | |
9119a41e | 4005 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 4006 | resv_map = inode_resv_map(inode); |
9119a41e | 4007 | |
1406ec9b | 4008 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4009 | |
4010 | } else { | |
4011 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4012 | if (!resv_map) |
4013 | return -ENOMEM; | |
4014 | ||
a1e78772 | 4015 | chg = to - from; |
84afd99b | 4016 | |
17c9d12e MG |
4017 | set_vma_resv_map(vma, resv_map); |
4018 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4019 | } | |
4020 | ||
c50ac050 DH |
4021 | if (chg < 0) { |
4022 | ret = chg; | |
4023 | goto out_err; | |
4024 | } | |
8a630112 | 4025 | |
1c5ecae3 MK |
4026 | /* |
4027 | * There must be enough pages in the subpool for the mapping. If | |
4028 | * the subpool has a minimum size, there may be some global | |
4029 | * reservations already in place (gbl_reserve). | |
4030 | */ | |
4031 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4032 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4033 | ret = -ENOSPC; |
4034 | goto out_err; | |
4035 | } | |
5a6fe125 MG |
4036 | |
4037 | /* | |
17c9d12e | 4038 | * Check enough hugepages are available for the reservation. |
90481622 | 4039 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4040 | */ |
1c5ecae3 | 4041 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4042 | if (ret < 0) { |
1c5ecae3 MK |
4043 | /* put back original number of pages, chg */ |
4044 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4045 | goto out_err; |
68842c9b | 4046 | } |
17c9d12e MG |
4047 | |
4048 | /* | |
4049 | * Account for the reservations made. Shared mappings record regions | |
4050 | * that have reservations as they are shared by multiple VMAs. | |
4051 | * When the last VMA disappears, the region map says how much | |
4052 | * the reservation was and the page cache tells how much of | |
4053 | * the reservation was consumed. Private mappings are per-VMA and | |
4054 | * only the consumed reservations are tracked. When the VMA | |
4055 | * disappears, the original reservation is the VMA size and the | |
4056 | * consumed reservations are stored in the map. Hence, nothing | |
4057 | * else has to be done for private mappings here | |
4058 | */ | |
33039678 MK |
4059 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4060 | long add = region_add(resv_map, from, to); | |
4061 | ||
4062 | if (unlikely(chg > add)) { | |
4063 | /* | |
4064 | * pages in this range were added to the reserve | |
4065 | * map between region_chg and region_add. This | |
4066 | * indicates a race with alloc_huge_page. Adjust | |
4067 | * the subpool and reserve counts modified above | |
4068 | * based on the difference. | |
4069 | */ | |
4070 | long rsv_adjust; | |
4071 | ||
4072 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4073 | chg - add); | |
4074 | hugetlb_acct_memory(h, -rsv_adjust); | |
4075 | } | |
4076 | } | |
a43a8c39 | 4077 | return 0; |
c50ac050 | 4078 | out_err: |
5e911373 MK |
4079 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
4080 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4081 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4082 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4083 | return ret; |
a43a8c39 CK |
4084 | } |
4085 | ||
b5cec28d MK |
4086 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4087 | long freed) | |
a43a8c39 | 4088 | { |
a5516438 | 4089 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4090 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4091 | long chg = 0; |
90481622 | 4092 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4093 | long gbl_reserve; |
45c682a6 | 4094 | |
b5cec28d MK |
4095 | if (resv_map) { |
4096 | chg = region_del(resv_map, start, end); | |
4097 | /* | |
4098 | * region_del() can fail in the rare case where a region | |
4099 | * must be split and another region descriptor can not be | |
4100 | * allocated. If end == LONG_MAX, it will not fail. | |
4101 | */ | |
4102 | if (chg < 0) | |
4103 | return chg; | |
4104 | } | |
4105 | ||
45c682a6 | 4106 | spin_lock(&inode->i_lock); |
e4c6f8be | 4107 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4108 | spin_unlock(&inode->i_lock); |
4109 | ||
1c5ecae3 MK |
4110 | /* |
4111 | * If the subpool has a minimum size, the number of global | |
4112 | * reservations to be released may be adjusted. | |
4113 | */ | |
4114 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4115 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4116 | |
4117 | return 0; | |
a43a8c39 | 4118 | } |
93f70f90 | 4119 | |
3212b535 SC |
4120 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4121 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4122 | struct vm_area_struct *vma, | |
4123 | unsigned long addr, pgoff_t idx) | |
4124 | { | |
4125 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4126 | svma->vm_start; | |
4127 | unsigned long sbase = saddr & PUD_MASK; | |
4128 | unsigned long s_end = sbase + PUD_SIZE; | |
4129 | ||
4130 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4131 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4132 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4133 | |
4134 | /* | |
4135 | * match the virtual addresses, permission and the alignment of the | |
4136 | * page table page. | |
4137 | */ | |
4138 | if (pmd_index(addr) != pmd_index(saddr) || | |
4139 | vm_flags != svm_flags || | |
4140 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4141 | return 0; | |
4142 | ||
4143 | return saddr; | |
4144 | } | |
4145 | ||
31aafb45 | 4146 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4147 | { |
4148 | unsigned long base = addr & PUD_MASK; | |
4149 | unsigned long end = base + PUD_SIZE; | |
4150 | ||
4151 | /* | |
4152 | * check on proper vm_flags and page table alignment | |
4153 | */ | |
4154 | if (vma->vm_flags & VM_MAYSHARE && | |
4155 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4156 | return true; |
4157 | return false; | |
3212b535 SC |
4158 | } |
4159 | ||
4160 | /* | |
4161 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4162 | * and returns the corresponding pte. While this is not necessary for the | |
4163 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4164 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4165 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4166 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4167 | * bad pmd for sharing. | |
4168 | */ | |
4169 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4170 | { | |
4171 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4172 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4173 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4174 | vma->vm_pgoff; | |
4175 | struct vm_area_struct *svma; | |
4176 | unsigned long saddr; | |
4177 | pte_t *spte = NULL; | |
4178 | pte_t *pte; | |
cb900f41 | 4179 | spinlock_t *ptl; |
3212b535 SC |
4180 | |
4181 | if (!vma_shareable(vma, addr)) | |
4182 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4183 | ||
83cde9e8 | 4184 | i_mmap_lock_write(mapping); |
3212b535 SC |
4185 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4186 | if (svma == vma) | |
4187 | continue; | |
4188 | ||
4189 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4190 | if (saddr) { | |
4191 | spte = huge_pte_offset(svma->vm_mm, saddr); | |
4192 | if (spte) { | |
dc6c9a35 | 4193 | mm_inc_nr_pmds(mm); |
3212b535 SC |
4194 | get_page(virt_to_page(spte)); |
4195 | break; | |
4196 | } | |
4197 | } | |
4198 | } | |
4199 | ||
4200 | if (!spte) | |
4201 | goto out; | |
4202 | ||
cb900f41 KS |
4203 | ptl = huge_pte_lockptr(hstate_vma(vma), mm, spte); |
4204 | spin_lock(ptl); | |
dc6c9a35 | 4205 | if (pud_none(*pud)) { |
3212b535 SC |
4206 | pud_populate(mm, pud, |
4207 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
dc6c9a35 | 4208 | } else { |
3212b535 | 4209 | put_page(virt_to_page(spte)); |
dc6c9a35 KS |
4210 | mm_inc_nr_pmds(mm); |
4211 | } | |
cb900f41 | 4212 | spin_unlock(ptl); |
3212b535 SC |
4213 | out: |
4214 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4215 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4216 | return pte; |
4217 | } | |
4218 | ||
4219 | /* | |
4220 | * unmap huge page backed by shared pte. | |
4221 | * | |
4222 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4223 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4224 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4225 | * | |
cb900f41 | 4226 | * called with page table lock held. |
3212b535 SC |
4227 | * |
4228 | * returns: 1 successfully unmapped a shared pte page | |
4229 | * 0 the underlying pte page is not shared, or it is the last user | |
4230 | */ | |
4231 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4232 | { | |
4233 | pgd_t *pgd = pgd_offset(mm, *addr); | |
4234 | pud_t *pud = pud_offset(pgd, *addr); | |
4235 | ||
4236 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4237 | if (page_count(virt_to_page(ptep)) == 1) | |
4238 | return 0; | |
4239 | ||
4240 | pud_clear(pud); | |
4241 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4242 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4243 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4244 | return 1; | |
4245 | } | |
9e5fc74c SC |
4246 | #define want_pmd_share() (1) |
4247 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4248 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4249 | { | |
4250 | return NULL; | |
4251 | } | |
e81f2d22 ZZ |
4252 | |
4253 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4254 | { | |
4255 | return 0; | |
4256 | } | |
9e5fc74c | 4257 | #define want_pmd_share() (0) |
3212b535 SC |
4258 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4259 | ||
9e5fc74c SC |
4260 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4261 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4262 | unsigned long addr, unsigned long sz) | |
4263 | { | |
4264 | pgd_t *pgd; | |
4265 | pud_t *pud; | |
4266 | pte_t *pte = NULL; | |
4267 | ||
4268 | pgd = pgd_offset(mm, addr); | |
4269 | pud = pud_alloc(mm, pgd, addr); | |
4270 | if (pud) { | |
4271 | if (sz == PUD_SIZE) { | |
4272 | pte = (pte_t *)pud; | |
4273 | } else { | |
4274 | BUG_ON(sz != PMD_SIZE); | |
4275 | if (want_pmd_share() && pud_none(*pud)) | |
4276 | pte = huge_pmd_share(mm, addr, pud); | |
4277 | else | |
4278 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4279 | } | |
4280 | } | |
4281 | BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte)); | |
4282 | ||
4283 | return pte; | |
4284 | } | |
4285 | ||
4286 | pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) | |
4287 | { | |
4288 | pgd_t *pgd; | |
4289 | pud_t *pud; | |
4290 | pmd_t *pmd = NULL; | |
4291 | ||
4292 | pgd = pgd_offset(mm, addr); | |
4293 | if (pgd_present(*pgd)) { | |
4294 | pud = pud_offset(pgd, addr); | |
4295 | if (pud_present(*pud)) { | |
4296 | if (pud_huge(*pud)) | |
4297 | return (pte_t *)pud; | |
4298 | pmd = pmd_offset(pud, addr); | |
4299 | } | |
4300 | } | |
4301 | return (pte_t *) pmd; | |
4302 | } | |
4303 | ||
61f77eda NH |
4304 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4305 | ||
4306 | /* | |
4307 | * These functions are overwritable if your architecture needs its own | |
4308 | * behavior. | |
4309 | */ | |
4310 | struct page * __weak | |
4311 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4312 | int write) | |
4313 | { | |
4314 | return ERR_PTR(-EINVAL); | |
4315 | } | |
4316 | ||
4317 | struct page * __weak | |
9e5fc74c | 4318 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4319 | pmd_t *pmd, int flags) |
9e5fc74c | 4320 | { |
e66f17ff NH |
4321 | struct page *page = NULL; |
4322 | spinlock_t *ptl; | |
4323 | retry: | |
4324 | ptl = pmd_lockptr(mm, pmd); | |
4325 | spin_lock(ptl); | |
4326 | /* | |
4327 | * make sure that the address range covered by this pmd is not | |
4328 | * unmapped from other threads. | |
4329 | */ | |
4330 | if (!pmd_huge(*pmd)) | |
4331 | goto out; | |
4332 | if (pmd_present(*pmd)) { | |
97534127 | 4333 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4334 | if (flags & FOLL_GET) |
4335 | get_page(page); | |
4336 | } else { | |
4337 | if (is_hugetlb_entry_migration(huge_ptep_get((pte_t *)pmd))) { | |
4338 | spin_unlock(ptl); | |
4339 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4340 | goto retry; | |
4341 | } | |
4342 | /* | |
4343 | * hwpoisoned entry is treated as no_page_table in | |
4344 | * follow_page_mask(). | |
4345 | */ | |
4346 | } | |
4347 | out: | |
4348 | spin_unlock(ptl); | |
9e5fc74c SC |
4349 | return page; |
4350 | } | |
4351 | ||
61f77eda | 4352 | struct page * __weak |
9e5fc74c | 4353 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4354 | pud_t *pud, int flags) |
9e5fc74c | 4355 | { |
e66f17ff NH |
4356 | if (flags & FOLL_GET) |
4357 | return NULL; | |
9e5fc74c | 4358 | |
e66f17ff | 4359 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4360 | } |
4361 | ||
d5bd9106 AK |
4362 | #ifdef CONFIG_MEMORY_FAILURE |
4363 | ||
93f70f90 NH |
4364 | /* |
4365 | * This function is called from memory failure code. | |
4366 | * Assume the caller holds page lock of the head page. | |
4367 | */ | |
6de2b1aa | 4368 | int dequeue_hwpoisoned_huge_page(struct page *hpage) |
93f70f90 NH |
4369 | { |
4370 | struct hstate *h = page_hstate(hpage); | |
4371 | int nid = page_to_nid(hpage); | |
6de2b1aa | 4372 | int ret = -EBUSY; |
93f70f90 NH |
4373 | |
4374 | spin_lock(&hugetlb_lock); | |
7e1f049e NH |
4375 | /* |
4376 | * Just checking !page_huge_active is not enough, because that could be | |
4377 | * an isolated/hwpoisoned hugepage (which have >0 refcount). | |
4378 | */ | |
4379 | if (!page_huge_active(hpage) && !page_count(hpage)) { | |
56f2fb14 NH |
4380 | /* |
4381 | * Hwpoisoned hugepage isn't linked to activelist or freelist, | |
4382 | * but dangling hpage->lru can trigger list-debug warnings | |
4383 | * (this happens when we call unpoison_memory() on it), | |
4384 | * so let it point to itself with list_del_init(). | |
4385 | */ | |
4386 | list_del_init(&hpage->lru); | |
8c6c2ecb | 4387 | set_page_refcounted(hpage); |
6de2b1aa NH |
4388 | h->free_huge_pages--; |
4389 | h->free_huge_pages_node[nid]--; | |
4390 | ret = 0; | |
4391 | } | |
93f70f90 | 4392 | spin_unlock(&hugetlb_lock); |
6de2b1aa | 4393 | return ret; |
93f70f90 | 4394 | } |
6de2b1aa | 4395 | #endif |
31caf665 NH |
4396 | |
4397 | bool isolate_huge_page(struct page *page, struct list_head *list) | |
4398 | { | |
bcc54222 NH |
4399 | bool ret = true; |
4400 | ||
309381fe | 4401 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4402 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4403 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4404 | ret = false; | |
4405 | goto unlock; | |
4406 | } | |
4407 | clear_page_huge_active(page); | |
31caf665 | 4408 | list_move_tail(&page->lru, list); |
bcc54222 | 4409 | unlock: |
31caf665 | 4410 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4411 | return ret; |
31caf665 NH |
4412 | } |
4413 | ||
4414 | void putback_active_hugepage(struct page *page) | |
4415 | { | |
309381fe | 4416 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4417 | spin_lock(&hugetlb_lock); |
bcc54222 | 4418 | set_page_huge_active(page); |
31caf665 NH |
4419 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4420 | spin_unlock(&hugetlb_lock); | |
4421 | put_page(page); | |
4422 | } |