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> |
63489f8e | 21 | #include <linux/mmdebug.h> |
174cd4b1 | 22 | #include <linux/sched/signal.h> |
0fe6e20b | 23 | #include <linux/rmap.h> |
c6247f72 | 24 | #include <linux/string_helpers.h> |
fd6a03ed NH |
25 | #include <linux/swap.h> |
26 | #include <linux/swapops.h> | |
8382d914 | 27 | #include <linux/jhash.h> |
d6606683 | 28 | |
63551ae0 DG |
29 | #include <asm/page.h> |
30 | #include <asm/pgtable.h> | |
24669e58 | 31 | #include <asm/tlb.h> |
63551ae0 | 32 | |
24669e58 | 33 | #include <linux/io.h> |
63551ae0 | 34 | #include <linux/hugetlb.h> |
9dd540e2 | 35 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 36 | #include <linux/node.h> |
1a1aad8a | 37 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 38 | #include <linux/page_owner.h> |
7835e98b | 39 | #include "internal.h" |
1da177e4 | 40 | |
c3f38a38 | 41 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
42 | unsigned int default_hstate_idx; |
43 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
44 | /* |
45 | * Minimum page order among possible hugepage sizes, set to a proper value | |
46 | * at boot time. | |
47 | */ | |
48 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 49 | |
53ba51d2 JT |
50 | __initdata LIST_HEAD(huge_boot_pages); |
51 | ||
e5ff2159 AK |
52 | /* for command line parsing */ |
53 | static struct hstate * __initdata parsed_hstate; | |
54 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 55 | static unsigned long __initdata default_hstate_size; |
9fee021d | 56 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 57 | |
3935baa9 | 58 | /* |
31caf665 NH |
59 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
60 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 61 | */ |
c3f38a38 | 62 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 63 | |
8382d914 DB |
64 | /* |
65 | * Serializes faults on the same logical page. This is used to | |
66 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
67 | */ | |
68 | static int num_fault_mutexes; | |
c672c7f2 | 69 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 70 | |
7ca02d0a MK |
71 | /* Forward declaration */ |
72 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
73 | ||
90481622 DG |
74 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
75 | { | |
76 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
77 | ||
78 | spin_unlock(&spool->lock); | |
79 | ||
80 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
81 | * remain, give up any reservations mased on minimum size and |
82 | * free the subpool */ | |
83 | if (free) { | |
84 | if (spool->min_hpages != -1) | |
85 | hugetlb_acct_memory(spool->hstate, | |
86 | -spool->min_hpages); | |
90481622 | 87 | kfree(spool); |
7ca02d0a | 88 | } |
90481622 DG |
89 | } |
90 | ||
7ca02d0a MK |
91 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
92 | long min_hpages) | |
90481622 DG |
93 | { |
94 | struct hugepage_subpool *spool; | |
95 | ||
c6a91820 | 96 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
97 | if (!spool) |
98 | return NULL; | |
99 | ||
100 | spin_lock_init(&spool->lock); | |
101 | spool->count = 1; | |
7ca02d0a MK |
102 | spool->max_hpages = max_hpages; |
103 | spool->hstate = h; | |
104 | spool->min_hpages = min_hpages; | |
105 | ||
106 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
107 | kfree(spool); | |
108 | return NULL; | |
109 | } | |
110 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
111 | |
112 | return spool; | |
113 | } | |
114 | ||
115 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
116 | { | |
117 | spin_lock(&spool->lock); | |
118 | BUG_ON(!spool->count); | |
119 | spool->count--; | |
120 | unlock_or_release_subpool(spool); | |
121 | } | |
122 | ||
1c5ecae3 MK |
123 | /* |
124 | * Subpool accounting for allocating and reserving pages. | |
125 | * Return -ENOMEM if there are not enough resources to satisfy the | |
126 | * the request. Otherwise, return the number of pages by which the | |
127 | * global pools must be adjusted (upward). The returned value may | |
128 | * only be different than the passed value (delta) in the case where | |
129 | * a subpool minimum size must be manitained. | |
130 | */ | |
131 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
132 | long delta) |
133 | { | |
1c5ecae3 | 134 | long ret = delta; |
90481622 DG |
135 | |
136 | if (!spool) | |
1c5ecae3 | 137 | return ret; |
90481622 DG |
138 | |
139 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
140 | |
141 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
142 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
143 | spool->used_hpages += delta; | |
144 | else { | |
145 | ret = -ENOMEM; | |
146 | goto unlock_ret; | |
147 | } | |
90481622 | 148 | } |
90481622 | 149 | |
09a95e29 MK |
150 | /* minimum size accounting */ |
151 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
152 | if (delta > spool->rsv_hpages) { |
153 | /* | |
154 | * Asking for more reserves than those already taken on | |
155 | * behalf of subpool. Return difference. | |
156 | */ | |
157 | ret = delta - spool->rsv_hpages; | |
158 | spool->rsv_hpages = 0; | |
159 | } else { | |
160 | ret = 0; /* reserves already accounted for */ | |
161 | spool->rsv_hpages -= delta; | |
162 | } | |
163 | } | |
164 | ||
165 | unlock_ret: | |
166 | spin_unlock(&spool->lock); | |
90481622 DG |
167 | return ret; |
168 | } | |
169 | ||
1c5ecae3 MK |
170 | /* |
171 | * Subpool accounting for freeing and unreserving pages. | |
172 | * Return the number of global page reservations that must be dropped. | |
173 | * The return value may only be different than the passed value (delta) | |
174 | * in the case where a subpool minimum size must be maintained. | |
175 | */ | |
176 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
177 | long delta) |
178 | { | |
1c5ecae3 MK |
179 | long ret = delta; |
180 | ||
90481622 | 181 | if (!spool) |
1c5ecae3 | 182 | return delta; |
90481622 DG |
183 | |
184 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
185 | |
186 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
187 | spool->used_hpages -= delta; | |
188 | ||
09a95e29 MK |
189 | /* minimum size accounting */ |
190 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
191 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
192 | ret = 0; | |
193 | else | |
194 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
195 | ||
196 | spool->rsv_hpages += delta; | |
197 | if (spool->rsv_hpages > spool->min_hpages) | |
198 | spool->rsv_hpages = spool->min_hpages; | |
199 | } | |
200 | ||
201 | /* | |
202 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
203 | * quota reference, free it now. | |
204 | */ | |
90481622 | 205 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
206 | |
207 | return ret; | |
90481622 DG |
208 | } |
209 | ||
210 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
211 | { | |
212 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
213 | } | |
214 | ||
215 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
216 | { | |
496ad9aa | 217 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
218 | } |
219 | ||
96822904 AW |
220 | /* |
221 | * Region tracking -- allows tracking of reservations and instantiated pages | |
222 | * across the pages in a mapping. | |
84afd99b | 223 | * |
1dd308a7 MK |
224 | * The region data structures are embedded into a resv_map and protected |
225 | * by a resv_map's lock. The set of regions within the resv_map represent | |
226 | * reservations for huge pages, or huge pages that have already been | |
227 | * instantiated within the map. The from and to elements are huge page | |
228 | * indicies into the associated mapping. from indicates the starting index | |
229 | * of the region. to represents the first index past the end of the region. | |
230 | * | |
231 | * For example, a file region structure with from == 0 and to == 4 represents | |
232 | * four huge pages in a mapping. It is important to note that the to element | |
233 | * represents the first element past the end of the region. This is used in | |
234 | * arithmetic as 4(to) - 0(from) = 4 huge pages in the region. | |
235 | * | |
236 | * Interval notation of the form [from, to) will be used to indicate that | |
237 | * the endpoint from is inclusive and to is exclusive. | |
96822904 AW |
238 | */ |
239 | struct file_region { | |
240 | struct list_head link; | |
241 | long from; | |
242 | long to; | |
243 | }; | |
244 | ||
1dd308a7 MK |
245 | /* |
246 | * Add the huge page range represented by [f, t) to the reserve | |
5e911373 MK |
247 | * map. In the normal case, existing regions will be expanded |
248 | * to accommodate the specified range. Sufficient regions should | |
249 | * exist for expansion due to the previous call to region_chg | |
250 | * with the same range. However, it is possible that region_del | |
251 | * could have been called after region_chg and modifed the map | |
252 | * in such a way that no region exists to be expanded. In this | |
253 | * case, pull a region descriptor from the cache associated with | |
254 | * the map and use that for the new range. | |
cf3ad20b MK |
255 | * |
256 | * Return the number of new huge pages added to the map. This | |
257 | * number is greater than or equal to zero. | |
1dd308a7 | 258 | */ |
1406ec9b | 259 | static long region_add(struct resv_map *resv, long f, long t) |
96822904 | 260 | { |
1406ec9b | 261 | struct list_head *head = &resv->regions; |
96822904 | 262 | struct file_region *rg, *nrg, *trg; |
cf3ad20b | 263 | long add = 0; |
96822904 | 264 | |
7b24d861 | 265 | spin_lock(&resv->lock); |
96822904 AW |
266 | /* Locate the region we are either in or before. */ |
267 | list_for_each_entry(rg, head, link) | |
268 | if (f <= rg->to) | |
269 | break; | |
270 | ||
5e911373 MK |
271 | /* |
272 | * If no region exists which can be expanded to include the | |
273 | * specified range, the list must have been modified by an | |
274 | * interleving call to region_del(). Pull a region descriptor | |
275 | * from the cache and use it for this range. | |
276 | */ | |
277 | if (&rg->link == head || t < rg->from) { | |
278 | VM_BUG_ON(resv->region_cache_count <= 0); | |
279 | ||
280 | resv->region_cache_count--; | |
281 | nrg = list_first_entry(&resv->region_cache, struct file_region, | |
282 | link); | |
283 | list_del(&nrg->link); | |
284 | ||
285 | nrg->from = f; | |
286 | nrg->to = t; | |
287 | list_add(&nrg->link, rg->link.prev); | |
288 | ||
289 | add += t - f; | |
290 | goto out_locked; | |
291 | } | |
292 | ||
96822904 AW |
293 | /* Round our left edge to the current segment if it encloses us. */ |
294 | if (f > rg->from) | |
295 | f = rg->from; | |
296 | ||
297 | /* Check for and consume any regions we now overlap with. */ | |
298 | nrg = rg; | |
299 | list_for_each_entry_safe(rg, trg, rg->link.prev, link) { | |
300 | if (&rg->link == head) | |
301 | break; | |
302 | if (rg->from > t) | |
303 | break; | |
304 | ||
305 | /* If this area reaches higher then extend our area to | |
306 | * include it completely. If this is not the first area | |
307 | * which we intend to reuse, free it. */ | |
308 | if (rg->to > t) | |
309 | t = rg->to; | |
310 | if (rg != nrg) { | |
cf3ad20b MK |
311 | /* Decrement return value by the deleted range. |
312 | * Another range will span this area so that by | |
313 | * end of routine add will be >= zero | |
314 | */ | |
315 | add -= (rg->to - rg->from); | |
96822904 AW |
316 | list_del(&rg->link); |
317 | kfree(rg); | |
318 | } | |
319 | } | |
cf3ad20b MK |
320 | |
321 | add += (nrg->from - f); /* Added to beginning of region */ | |
96822904 | 322 | nrg->from = f; |
cf3ad20b | 323 | add += t - nrg->to; /* Added to end of region */ |
96822904 | 324 | nrg->to = t; |
cf3ad20b | 325 | |
5e911373 MK |
326 | out_locked: |
327 | resv->adds_in_progress--; | |
7b24d861 | 328 | spin_unlock(&resv->lock); |
cf3ad20b MK |
329 | VM_BUG_ON(add < 0); |
330 | return add; | |
96822904 AW |
331 | } |
332 | ||
1dd308a7 MK |
333 | /* |
334 | * Examine the existing reserve map and determine how many | |
335 | * huge pages in the specified range [f, t) are NOT currently | |
336 | * represented. This routine is called before a subsequent | |
337 | * call to region_add that will actually modify the reserve | |
338 | * map to add the specified range [f, t). region_chg does | |
339 | * not change the number of huge pages represented by the | |
340 | * map. However, if the existing regions in the map can not | |
341 | * be expanded to represent the new range, a new file_region | |
342 | * structure is added to the map as a placeholder. This is | |
343 | * so that the subsequent region_add call will have all the | |
344 | * regions it needs and will not fail. | |
345 | * | |
5e911373 MK |
346 | * Upon entry, region_chg will also examine the cache of region descriptors |
347 | * associated with the map. If there are not enough descriptors cached, one | |
348 | * will be allocated for the in progress add operation. | |
349 | * | |
350 | * Returns the number of huge pages that need to be added to the existing | |
351 | * reservation map for the range [f, t). This number is greater or equal to | |
352 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
353 | * is needed and can not be allocated. | |
1dd308a7 | 354 | */ |
1406ec9b | 355 | static long region_chg(struct resv_map *resv, long f, long t) |
96822904 | 356 | { |
1406ec9b | 357 | struct list_head *head = &resv->regions; |
7b24d861 | 358 | struct file_region *rg, *nrg = NULL; |
96822904 AW |
359 | long chg = 0; |
360 | ||
7b24d861 DB |
361 | retry: |
362 | spin_lock(&resv->lock); | |
5e911373 MK |
363 | retry_locked: |
364 | resv->adds_in_progress++; | |
365 | ||
366 | /* | |
367 | * Check for sufficient descriptors in the cache to accommodate | |
368 | * the number of in progress add operations. | |
369 | */ | |
370 | if (resv->adds_in_progress > resv->region_cache_count) { | |
371 | struct file_region *trg; | |
372 | ||
373 | VM_BUG_ON(resv->adds_in_progress - resv->region_cache_count > 1); | |
374 | /* Must drop lock to allocate a new descriptor. */ | |
375 | resv->adds_in_progress--; | |
376 | spin_unlock(&resv->lock); | |
377 | ||
378 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
dbe409e4 MK |
379 | if (!trg) { |
380 | kfree(nrg); | |
5e911373 | 381 | return -ENOMEM; |
dbe409e4 | 382 | } |
5e911373 MK |
383 | |
384 | spin_lock(&resv->lock); | |
385 | list_add(&trg->link, &resv->region_cache); | |
386 | resv->region_cache_count++; | |
387 | goto retry_locked; | |
388 | } | |
389 | ||
96822904 AW |
390 | /* Locate the region we are before or in. */ |
391 | list_for_each_entry(rg, head, link) | |
392 | if (f <= rg->to) | |
393 | break; | |
394 | ||
395 | /* If we are below the current region then a new region is required. | |
396 | * Subtle, allocate a new region at the position but make it zero | |
397 | * size such that we can guarantee to record the reservation. */ | |
398 | if (&rg->link == head || t < rg->from) { | |
7b24d861 | 399 | if (!nrg) { |
5e911373 | 400 | resv->adds_in_progress--; |
7b24d861 DB |
401 | spin_unlock(&resv->lock); |
402 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
403 | if (!nrg) | |
404 | return -ENOMEM; | |
405 | ||
406 | nrg->from = f; | |
407 | nrg->to = f; | |
408 | INIT_LIST_HEAD(&nrg->link); | |
409 | goto retry; | |
410 | } | |
96822904 | 411 | |
7b24d861 DB |
412 | list_add(&nrg->link, rg->link.prev); |
413 | chg = t - f; | |
414 | goto out_nrg; | |
96822904 AW |
415 | } |
416 | ||
417 | /* Round our left edge to the current segment if it encloses us. */ | |
418 | if (f > rg->from) | |
419 | f = rg->from; | |
420 | chg = t - f; | |
421 | ||
422 | /* Check for and consume any regions we now overlap with. */ | |
423 | list_for_each_entry(rg, rg->link.prev, link) { | |
424 | if (&rg->link == head) | |
425 | break; | |
426 | if (rg->from > t) | |
7b24d861 | 427 | goto out; |
96822904 | 428 | |
25985edc | 429 | /* We overlap with this area, if it extends further than |
96822904 AW |
430 | * us then we must extend ourselves. Account for its |
431 | * existing reservation. */ | |
432 | if (rg->to > t) { | |
433 | chg += rg->to - t; | |
434 | t = rg->to; | |
435 | } | |
436 | chg -= rg->to - rg->from; | |
437 | } | |
7b24d861 DB |
438 | |
439 | out: | |
440 | spin_unlock(&resv->lock); | |
441 | /* We already know we raced and no longer need the new region */ | |
442 | kfree(nrg); | |
443 | return chg; | |
444 | out_nrg: | |
445 | spin_unlock(&resv->lock); | |
96822904 AW |
446 | return chg; |
447 | } | |
448 | ||
5e911373 MK |
449 | /* |
450 | * Abort the in progress add operation. The adds_in_progress field | |
451 | * of the resv_map keeps track of the operations in progress between | |
452 | * calls to region_chg and region_add. Operations are sometimes | |
453 | * aborted after the call to region_chg. In such cases, region_abort | |
454 | * is called to decrement the adds_in_progress counter. | |
455 | * | |
456 | * NOTE: The range arguments [f, t) are not needed or used in this | |
457 | * routine. They are kept to make reading the calling code easier as | |
458 | * arguments will match the associated region_chg call. | |
459 | */ | |
460 | static void region_abort(struct resv_map *resv, long f, long t) | |
461 | { | |
462 | spin_lock(&resv->lock); | |
463 | VM_BUG_ON(!resv->region_cache_count); | |
464 | resv->adds_in_progress--; | |
465 | spin_unlock(&resv->lock); | |
466 | } | |
467 | ||
1dd308a7 | 468 | /* |
feba16e2 MK |
469 | * Delete the specified range [f, t) from the reserve map. If the |
470 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
471 | * should be deleted. Locate the regions which intersect [f, t) | |
472 | * and either trim, delete or split the existing regions. | |
473 | * | |
474 | * Returns the number of huge pages deleted from the reserve map. | |
475 | * In the normal case, the return value is zero or more. In the | |
476 | * case where a region must be split, a new region descriptor must | |
477 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
478 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
479 | * a region and possibly return -ENOMEM. Callers specifying | |
480 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 481 | */ |
feba16e2 | 482 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 483 | { |
1406ec9b | 484 | struct list_head *head = &resv->regions; |
96822904 | 485 | struct file_region *rg, *trg; |
feba16e2 MK |
486 | struct file_region *nrg = NULL; |
487 | long del = 0; | |
96822904 | 488 | |
feba16e2 | 489 | retry: |
7b24d861 | 490 | spin_lock(&resv->lock); |
feba16e2 | 491 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
492 | /* |
493 | * Skip regions before the range to be deleted. file_region | |
494 | * ranges are normally of the form [from, to). However, there | |
495 | * may be a "placeholder" entry in the map which is of the form | |
496 | * (from, to) with from == to. Check for placeholder entries | |
497 | * at the beginning of the range to be deleted. | |
498 | */ | |
499 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 500 | continue; |
dbe409e4 | 501 | |
feba16e2 | 502 | if (rg->from >= t) |
96822904 | 503 | break; |
96822904 | 504 | |
feba16e2 MK |
505 | if (f > rg->from && t < rg->to) { /* Must split region */ |
506 | /* | |
507 | * Check for an entry in the cache before dropping | |
508 | * lock and attempting allocation. | |
509 | */ | |
510 | if (!nrg && | |
511 | resv->region_cache_count > resv->adds_in_progress) { | |
512 | nrg = list_first_entry(&resv->region_cache, | |
513 | struct file_region, | |
514 | link); | |
515 | list_del(&nrg->link); | |
516 | resv->region_cache_count--; | |
517 | } | |
96822904 | 518 | |
feba16e2 MK |
519 | if (!nrg) { |
520 | spin_unlock(&resv->lock); | |
521 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
522 | if (!nrg) | |
523 | return -ENOMEM; | |
524 | goto retry; | |
525 | } | |
526 | ||
527 | del += t - f; | |
528 | ||
529 | /* New entry for end of split region */ | |
530 | nrg->from = t; | |
531 | nrg->to = rg->to; | |
532 | INIT_LIST_HEAD(&nrg->link); | |
533 | ||
534 | /* Original entry is trimmed */ | |
535 | rg->to = f; | |
536 | ||
537 | list_add(&nrg->link, &rg->link); | |
538 | nrg = NULL; | |
96822904 | 539 | break; |
feba16e2 MK |
540 | } |
541 | ||
542 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
543 | del += rg->to - rg->from; | |
544 | list_del(&rg->link); | |
545 | kfree(rg); | |
546 | continue; | |
547 | } | |
548 | ||
549 | if (f <= rg->from) { /* Trim beginning of region */ | |
550 | del += t - rg->from; | |
551 | rg->from = t; | |
552 | } else { /* Trim end of region */ | |
553 | del += rg->to - f; | |
554 | rg->to = f; | |
555 | } | |
96822904 | 556 | } |
7b24d861 | 557 | |
7b24d861 | 558 | spin_unlock(&resv->lock); |
feba16e2 MK |
559 | kfree(nrg); |
560 | return del; | |
96822904 AW |
561 | } |
562 | ||
b5cec28d MK |
563 | /* |
564 | * A rare out of memory error was encountered which prevented removal of | |
565 | * the reserve map region for a page. The huge page itself was free'ed | |
566 | * and removed from the page cache. This routine will adjust the subpool | |
567 | * usage count, and the global reserve count if needed. By incrementing | |
568 | * these counts, the reserve map entry which could not be deleted will | |
569 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
570 | * counts. | |
571 | */ | |
72e2936c | 572 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
573 | { |
574 | struct hugepage_subpool *spool = subpool_inode(inode); | |
575 | long rsv_adjust; | |
576 | ||
577 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 578 | if (rsv_adjust) { |
b5cec28d MK |
579 | struct hstate *h = hstate_inode(inode); |
580 | ||
581 | hugetlb_acct_memory(h, 1); | |
582 | } | |
583 | } | |
584 | ||
1dd308a7 MK |
585 | /* |
586 | * Count and return the number of huge pages in the reserve map | |
587 | * that intersect with the range [f, t). | |
588 | */ | |
1406ec9b | 589 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 590 | { |
1406ec9b | 591 | struct list_head *head = &resv->regions; |
84afd99b AW |
592 | struct file_region *rg; |
593 | long chg = 0; | |
594 | ||
7b24d861 | 595 | spin_lock(&resv->lock); |
84afd99b AW |
596 | /* Locate each segment we overlap with, and count that overlap. */ |
597 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
598 | long seg_from; |
599 | long seg_to; | |
84afd99b AW |
600 | |
601 | if (rg->to <= f) | |
602 | continue; | |
603 | if (rg->from >= t) | |
604 | break; | |
605 | ||
606 | seg_from = max(rg->from, f); | |
607 | seg_to = min(rg->to, t); | |
608 | ||
609 | chg += seg_to - seg_from; | |
610 | } | |
7b24d861 | 611 | spin_unlock(&resv->lock); |
84afd99b AW |
612 | |
613 | return chg; | |
614 | } | |
615 | ||
e7c4b0bf AW |
616 | /* |
617 | * Convert the address within this vma to the page offset within | |
618 | * the mapping, in pagecache page units; huge pages here. | |
619 | */ | |
a5516438 AK |
620 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
621 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 622 | { |
a5516438 AK |
623 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
624 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
625 | } |
626 | ||
0fe6e20b NH |
627 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
628 | unsigned long address) | |
629 | { | |
630 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
631 | } | |
dee41079 | 632 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 633 | |
08fba699 MG |
634 | /* |
635 | * Return the size of the pages allocated when backing a VMA. In the majority | |
636 | * cases this will be same size as used by the page table entries. | |
637 | */ | |
638 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
639 | { | |
05ea8860 DW |
640 | if (vma->vm_ops && vma->vm_ops->pagesize) |
641 | return vma->vm_ops->pagesize(vma); | |
642 | return PAGE_SIZE; | |
08fba699 | 643 | } |
f340ca0f | 644 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 645 | |
3340289d MG |
646 | /* |
647 | * Return the page size being used by the MMU to back a VMA. In the majority | |
648 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
649 | * architectures where it differs, an architecture-specific 'strong' |
650 | * version of this symbol is required. | |
3340289d | 651 | */ |
09135cc5 | 652 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
653 | { |
654 | return vma_kernel_pagesize(vma); | |
655 | } | |
3340289d | 656 | |
84afd99b AW |
657 | /* |
658 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
659 | * bits of the reservation map pointer, which are always clear due to | |
660 | * alignment. | |
661 | */ | |
662 | #define HPAGE_RESV_OWNER (1UL << 0) | |
663 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 664 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 665 | |
a1e78772 MG |
666 | /* |
667 | * These helpers are used to track how many pages are reserved for | |
668 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
669 | * is guaranteed to have their future faults succeed. | |
670 | * | |
671 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
672 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
673 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
674 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
675 | * |
676 | * The private mapping reservation is represented in a subtly different | |
677 | * manner to a shared mapping. A shared mapping has a region map associated | |
678 | * with the underlying file, this region map represents the backing file | |
679 | * pages which have ever had a reservation assigned which this persists even | |
680 | * after the page is instantiated. A private mapping has a region map | |
681 | * associated with the original mmap which is attached to all VMAs which | |
682 | * reference it, this region map represents those offsets which have consumed | |
683 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 684 | */ |
e7c4b0bf AW |
685 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
686 | { | |
687 | return (unsigned long)vma->vm_private_data; | |
688 | } | |
689 | ||
690 | static void set_vma_private_data(struct vm_area_struct *vma, | |
691 | unsigned long value) | |
692 | { | |
693 | vma->vm_private_data = (void *)value; | |
694 | } | |
695 | ||
9119a41e | 696 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
697 | { |
698 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
699 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
700 | ||
701 | if (!resv_map || !rg) { | |
702 | kfree(resv_map); | |
703 | kfree(rg); | |
84afd99b | 704 | return NULL; |
5e911373 | 705 | } |
84afd99b AW |
706 | |
707 | kref_init(&resv_map->refs); | |
7b24d861 | 708 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
709 | INIT_LIST_HEAD(&resv_map->regions); |
710 | ||
5e911373 MK |
711 | resv_map->adds_in_progress = 0; |
712 | ||
713 | INIT_LIST_HEAD(&resv_map->region_cache); | |
714 | list_add(&rg->link, &resv_map->region_cache); | |
715 | resv_map->region_cache_count = 1; | |
716 | ||
84afd99b AW |
717 | return resv_map; |
718 | } | |
719 | ||
9119a41e | 720 | void resv_map_release(struct kref *ref) |
84afd99b AW |
721 | { |
722 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
723 | struct list_head *head = &resv_map->region_cache; |
724 | struct file_region *rg, *trg; | |
84afd99b AW |
725 | |
726 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 727 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
728 | |
729 | /* ... and any entries left in the cache */ | |
730 | list_for_each_entry_safe(rg, trg, head, link) { | |
731 | list_del(&rg->link); | |
732 | kfree(rg); | |
733 | } | |
734 | ||
735 | VM_BUG_ON(resv_map->adds_in_progress); | |
736 | ||
84afd99b AW |
737 | kfree(resv_map); |
738 | } | |
739 | ||
4e35f483 JK |
740 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
741 | { | |
742 | return inode->i_mapping->private_data; | |
743 | } | |
744 | ||
84afd99b | 745 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 746 | { |
81d1b09c | 747 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
748 | if (vma->vm_flags & VM_MAYSHARE) { |
749 | struct address_space *mapping = vma->vm_file->f_mapping; | |
750 | struct inode *inode = mapping->host; | |
751 | ||
752 | return inode_resv_map(inode); | |
753 | ||
754 | } else { | |
84afd99b AW |
755 | return (struct resv_map *)(get_vma_private_data(vma) & |
756 | ~HPAGE_RESV_MASK); | |
4e35f483 | 757 | } |
a1e78772 MG |
758 | } |
759 | ||
84afd99b | 760 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 761 | { |
81d1b09c SL |
762 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
763 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 764 | |
84afd99b AW |
765 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
766 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
767 | } |
768 | ||
769 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
770 | { | |
81d1b09c SL |
771 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
772 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
773 | |
774 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
775 | } |
776 | ||
777 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
778 | { | |
81d1b09c | 779 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
780 | |
781 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
782 | } |
783 | ||
04f2cbe3 | 784 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
785 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
786 | { | |
81d1b09c | 787 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 788 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
789 | vma->vm_private_data = (void *)0; |
790 | } | |
791 | ||
792 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 793 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 794 | { |
af0ed73e JK |
795 | if (vma->vm_flags & VM_NORESERVE) { |
796 | /* | |
797 | * This address is already reserved by other process(chg == 0), | |
798 | * so, we should decrement reserved count. Without decrementing, | |
799 | * reserve count remains after releasing inode, because this | |
800 | * allocated page will go into page cache and is regarded as | |
801 | * coming from reserved pool in releasing step. Currently, we | |
802 | * don't have any other solution to deal with this situation | |
803 | * properly, so add work-around here. | |
804 | */ | |
805 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 806 | return true; |
af0ed73e | 807 | else |
559ec2f8 | 808 | return false; |
af0ed73e | 809 | } |
a63884e9 JK |
810 | |
811 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
812 | if (vma->vm_flags & VM_MAYSHARE) { |
813 | /* | |
814 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
815 | * be a region map for all pages. The only situation where | |
816 | * there is no region map is if a hole was punched via | |
817 | * fallocate. In this case, there really are no reverves to | |
818 | * use. This situation is indicated if chg != 0. | |
819 | */ | |
820 | if (chg) | |
821 | return false; | |
822 | else | |
823 | return true; | |
824 | } | |
a63884e9 JK |
825 | |
826 | /* | |
827 | * Only the process that called mmap() has reserves for | |
828 | * private mappings. | |
829 | */ | |
67961f9d MK |
830 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
831 | /* | |
832 | * Like the shared case above, a hole punch or truncate | |
833 | * could have been performed on the private mapping. | |
834 | * Examine the value of chg to determine if reserves | |
835 | * actually exist or were previously consumed. | |
836 | * Very Subtle - The value of chg comes from a previous | |
837 | * call to vma_needs_reserves(). The reserve map for | |
838 | * private mappings has different (opposite) semantics | |
839 | * than that of shared mappings. vma_needs_reserves() | |
840 | * has already taken this difference in semantics into | |
841 | * account. Therefore, the meaning of chg is the same | |
842 | * as in the shared case above. Code could easily be | |
843 | * combined, but keeping it separate draws attention to | |
844 | * subtle differences. | |
845 | */ | |
846 | if (chg) | |
847 | return false; | |
848 | else | |
849 | return true; | |
850 | } | |
a63884e9 | 851 | |
559ec2f8 | 852 | return false; |
a1e78772 MG |
853 | } |
854 | ||
a5516438 | 855 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
856 | { |
857 | int nid = page_to_nid(page); | |
0edaecfa | 858 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
859 | h->free_huge_pages++; |
860 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
861 | } |
862 | ||
94310cbc | 863 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
864 | { |
865 | struct page *page; | |
866 | ||
c8721bbb | 867 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 868 | if (!PageHWPoison(page)) |
c8721bbb NH |
869 | break; |
870 | /* | |
871 | * if 'non-isolated free hugepage' not found on the list, | |
872 | * the allocation fails. | |
873 | */ | |
874 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 875 | return NULL; |
0edaecfa | 876 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 877 | set_page_refcounted(page); |
bf50bab2 NH |
878 | h->free_huge_pages--; |
879 | h->free_huge_pages_node[nid]--; | |
880 | return page; | |
881 | } | |
882 | ||
3e59fcb0 MH |
883 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
884 | nodemask_t *nmask) | |
94310cbc | 885 | { |
3e59fcb0 MH |
886 | unsigned int cpuset_mems_cookie; |
887 | struct zonelist *zonelist; | |
888 | struct zone *zone; | |
889 | struct zoneref *z; | |
890 | int node = -1; | |
94310cbc | 891 | |
3e59fcb0 MH |
892 | zonelist = node_zonelist(nid, gfp_mask); |
893 | ||
894 | retry_cpuset: | |
895 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
896 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
897 | struct page *page; | |
898 | ||
899 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
900 | continue; | |
901 | /* | |
902 | * no need to ask again on the same node. Pool is node rather than | |
903 | * zone aware | |
904 | */ | |
905 | if (zone_to_nid(zone) == node) | |
906 | continue; | |
907 | node = zone_to_nid(zone); | |
94310cbc | 908 | |
94310cbc AK |
909 | page = dequeue_huge_page_node_exact(h, node); |
910 | if (page) | |
911 | return page; | |
912 | } | |
3e59fcb0 MH |
913 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
914 | goto retry_cpuset; | |
915 | ||
94310cbc AK |
916 | return NULL; |
917 | } | |
918 | ||
86cdb465 NH |
919 | /* Movability of hugepages depends on migration support. */ |
920 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
921 | { | |
d6cb41cc | 922 | if (hugepage_migration_supported(h)) |
86cdb465 NH |
923 | return GFP_HIGHUSER_MOVABLE; |
924 | else | |
925 | return GFP_HIGHUSER; | |
926 | } | |
927 | ||
a5516438 AK |
928 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
929 | struct vm_area_struct *vma, | |
af0ed73e JK |
930 | unsigned long address, int avoid_reserve, |
931 | long chg) | |
1da177e4 | 932 | { |
3e59fcb0 | 933 | struct page *page; |
480eccf9 | 934 | struct mempolicy *mpol; |
04ec6264 | 935 | gfp_t gfp_mask; |
3e59fcb0 | 936 | nodemask_t *nodemask; |
04ec6264 | 937 | int nid; |
1da177e4 | 938 | |
a1e78772 MG |
939 | /* |
940 | * A child process with MAP_PRIVATE mappings created by their parent | |
941 | * have no page reserves. This check ensures that reservations are | |
942 | * not "stolen". The child may still get SIGKILLed | |
943 | */ | |
af0ed73e | 944 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 945 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 946 | goto err; |
a1e78772 | 947 | |
04f2cbe3 | 948 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 949 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 950 | goto err; |
04f2cbe3 | 951 | |
04ec6264 VB |
952 | gfp_mask = htlb_alloc_mask(h); |
953 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
954 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
955 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
956 | SetPagePrivate(page); | |
957 | h->resv_huge_pages--; | |
1da177e4 | 958 | } |
cc9a6c87 | 959 | |
52cd3b07 | 960 | mpol_cond_put(mpol); |
1da177e4 | 961 | return page; |
cc9a6c87 MG |
962 | |
963 | err: | |
cc9a6c87 | 964 | return NULL; |
1da177e4 LT |
965 | } |
966 | ||
1cac6f2c LC |
967 | /* |
968 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
969 | * We may have allocated or freed a huge page based on a different | |
970 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
971 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
972 | * node for alloc or free. | |
973 | */ | |
974 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
975 | { | |
0edaf86c | 976 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
977 | VM_BUG_ON(nid >= MAX_NUMNODES); |
978 | ||
979 | return nid; | |
980 | } | |
981 | ||
982 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
983 | { | |
984 | if (!node_isset(nid, *nodes_allowed)) | |
985 | nid = next_node_allowed(nid, nodes_allowed); | |
986 | return nid; | |
987 | } | |
988 | ||
989 | /* | |
990 | * returns the previously saved node ["this node"] from which to | |
991 | * allocate a persistent huge page for the pool and advance the | |
992 | * next node from which to allocate, handling wrap at end of node | |
993 | * mask. | |
994 | */ | |
995 | static int hstate_next_node_to_alloc(struct hstate *h, | |
996 | nodemask_t *nodes_allowed) | |
997 | { | |
998 | int nid; | |
999 | ||
1000 | VM_BUG_ON(!nodes_allowed); | |
1001 | ||
1002 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1003 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1004 | ||
1005 | return nid; | |
1006 | } | |
1007 | ||
1008 | /* | |
1009 | * helper for free_pool_huge_page() - return the previously saved | |
1010 | * node ["this node"] from which to free a huge page. Advance the | |
1011 | * next node id whether or not we find a free huge page to free so | |
1012 | * that the next attempt to free addresses the next node. | |
1013 | */ | |
1014 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1015 | { | |
1016 | int nid; | |
1017 | ||
1018 | VM_BUG_ON(!nodes_allowed); | |
1019 | ||
1020 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1021 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1022 | ||
1023 | return nid; | |
1024 | } | |
1025 | ||
1026 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1027 | for (nr_nodes = nodes_weight(*mask); \ | |
1028 | nr_nodes > 0 && \ | |
1029 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1030 | nr_nodes--) | |
1031 | ||
1032 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1033 | for (nr_nodes = nodes_weight(*mask); \ | |
1034 | nr_nodes > 0 && \ | |
1035 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1036 | nr_nodes--) | |
1037 | ||
e1073d1e | 1038 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1039 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1040 | unsigned int order) |
944d9fec LC |
1041 | { |
1042 | int i; | |
1043 | int nr_pages = 1 << order; | |
1044 | struct page *p = page + 1; | |
1045 | ||
c8cc708a | 1046 | atomic_set(compound_mapcount_ptr(page), 0); |
944d9fec | 1047 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1048 | clear_compound_head(p); |
944d9fec | 1049 | set_page_refcounted(p); |
944d9fec LC |
1050 | } |
1051 | ||
1052 | set_compound_order(page, 0); | |
1053 | __ClearPageHead(page); | |
1054 | } | |
1055 | ||
d00181b9 | 1056 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1057 | { |
1058 | free_contig_range(page_to_pfn(page), 1 << order); | |
1059 | } | |
1060 | ||
1061 | static int __alloc_gigantic_page(unsigned long start_pfn, | |
79b63f12 | 1062 | unsigned long nr_pages, gfp_t gfp_mask) |
944d9fec LC |
1063 | { |
1064 | unsigned long end_pfn = start_pfn + nr_pages; | |
ca96b625 | 1065 | return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE, |
79b63f12 | 1066 | gfp_mask); |
944d9fec LC |
1067 | } |
1068 | ||
f44b2dda JK |
1069 | static bool pfn_range_valid_gigantic(struct zone *z, |
1070 | unsigned long start_pfn, unsigned long nr_pages) | |
944d9fec LC |
1071 | { |
1072 | unsigned long i, end_pfn = start_pfn + nr_pages; | |
1073 | struct page *page; | |
1074 | ||
1075 | for (i = start_pfn; i < end_pfn; i++) { | |
1076 | if (!pfn_valid(i)) | |
1077 | return false; | |
1078 | ||
1079 | page = pfn_to_page(i); | |
1080 | ||
f44b2dda JK |
1081 | if (page_zone(page) != z) |
1082 | return false; | |
1083 | ||
944d9fec LC |
1084 | if (PageReserved(page)) |
1085 | return false; | |
1086 | ||
1087 | if (page_count(page) > 0) | |
1088 | return false; | |
1089 | ||
1090 | if (PageHuge(page)) | |
1091 | return false; | |
1092 | } | |
1093 | ||
1094 | return true; | |
1095 | } | |
1096 | ||
1097 | static bool zone_spans_last_pfn(const struct zone *zone, | |
1098 | unsigned long start_pfn, unsigned long nr_pages) | |
1099 | { | |
1100 | unsigned long last_pfn = start_pfn + nr_pages - 1; | |
1101 | return zone_spans_pfn(zone, last_pfn); | |
1102 | } | |
1103 | ||
d9cc948f MH |
1104 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1105 | int nid, nodemask_t *nodemask) | |
944d9fec | 1106 | { |
79b63f12 | 1107 | unsigned int order = huge_page_order(h); |
944d9fec LC |
1108 | unsigned long nr_pages = 1 << order; |
1109 | unsigned long ret, pfn, flags; | |
79b63f12 MH |
1110 | struct zonelist *zonelist; |
1111 | struct zone *zone; | |
1112 | struct zoneref *z; | |
944d9fec | 1113 | |
79b63f12 | 1114 | zonelist = node_zonelist(nid, gfp_mask); |
d9cc948f | 1115 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nodemask) { |
79b63f12 | 1116 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec | 1117 | |
79b63f12 MH |
1118 | pfn = ALIGN(zone->zone_start_pfn, nr_pages); |
1119 | while (zone_spans_last_pfn(zone, pfn, nr_pages)) { | |
1120 | if (pfn_range_valid_gigantic(zone, pfn, nr_pages)) { | |
944d9fec LC |
1121 | /* |
1122 | * We release the zone lock here because | |
1123 | * alloc_contig_range() will also lock the zone | |
1124 | * at some point. If there's an allocation | |
1125 | * spinning on this lock, it may win the race | |
1126 | * and cause alloc_contig_range() to fail... | |
1127 | */ | |
79b63f12 MH |
1128 | spin_unlock_irqrestore(&zone->lock, flags); |
1129 | ret = __alloc_gigantic_page(pfn, nr_pages, gfp_mask); | |
944d9fec LC |
1130 | if (!ret) |
1131 | return pfn_to_page(pfn); | |
79b63f12 | 1132 | spin_lock_irqsave(&zone->lock, flags); |
944d9fec LC |
1133 | } |
1134 | pfn += nr_pages; | |
1135 | } | |
1136 | ||
79b63f12 | 1137 | spin_unlock_irqrestore(&zone->lock, flags); |
944d9fec LC |
1138 | } |
1139 | ||
1140 | return NULL; | |
1141 | } | |
1142 | ||
1143 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1144 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
944d9fec | 1145 | |
e1073d1e | 1146 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
944d9fec | 1147 | static inline bool gigantic_page_supported(void) { return false; } |
d9cc948f MH |
1148 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1149 | int nid, nodemask_t *nodemask) { return NULL; } | |
d00181b9 | 1150 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1151 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1152 | unsigned int order) { } |
944d9fec LC |
1153 | #endif |
1154 | ||
a5516438 | 1155 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1156 | { |
1157 | int i; | |
a5516438 | 1158 | |
944d9fec LC |
1159 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
1160 | return; | |
18229df5 | 1161 | |
a5516438 AK |
1162 | h->nr_huge_pages--; |
1163 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1164 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1165 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1166 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1167 | 1 << PG_active | 1 << PG_private | |
1168 | 1 << PG_writeback); | |
6af2acb6 | 1169 | } |
309381fe | 1170 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
f1e61557 | 1171 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1172 | set_page_refcounted(page); |
944d9fec LC |
1173 | if (hstate_is_gigantic(h)) { |
1174 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1175 | free_gigantic_page(page, huge_page_order(h)); | |
1176 | } else { | |
944d9fec LC |
1177 | __free_pages(page, huge_page_order(h)); |
1178 | } | |
6af2acb6 AL |
1179 | } |
1180 | ||
e5ff2159 AK |
1181 | struct hstate *size_to_hstate(unsigned long size) |
1182 | { | |
1183 | struct hstate *h; | |
1184 | ||
1185 | for_each_hstate(h) { | |
1186 | if (huge_page_size(h) == size) | |
1187 | return h; | |
1188 | } | |
1189 | return NULL; | |
1190 | } | |
1191 | ||
bcc54222 NH |
1192 | /* |
1193 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1194 | * to hstate->hugepage_activelist.) | |
1195 | * | |
1196 | * This function can be called for tail pages, but never returns true for them. | |
1197 | */ | |
1198 | bool page_huge_active(struct page *page) | |
1199 | { | |
1200 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1201 | return PageHead(page) && PagePrivate(&page[1]); | |
1202 | } | |
1203 | ||
1204 | /* never called for tail page */ | |
1205 | static void set_page_huge_active(struct page *page) | |
1206 | { | |
1207 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1208 | SetPagePrivate(&page[1]); | |
1209 | } | |
1210 | ||
1211 | static void clear_page_huge_active(struct page *page) | |
1212 | { | |
1213 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1214 | ClearPagePrivate(&page[1]); | |
1215 | } | |
1216 | ||
ab5ac90a MH |
1217 | /* |
1218 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1219 | * code | |
1220 | */ | |
1221 | static inline bool PageHugeTemporary(struct page *page) | |
1222 | { | |
1223 | if (!PageHuge(page)) | |
1224 | return false; | |
1225 | ||
1226 | return (unsigned long)page[2].mapping == -1U; | |
1227 | } | |
1228 | ||
1229 | static inline void SetPageHugeTemporary(struct page *page) | |
1230 | { | |
1231 | page[2].mapping = (void *)-1U; | |
1232 | } | |
1233 | ||
1234 | static inline void ClearPageHugeTemporary(struct page *page) | |
1235 | { | |
1236 | page[2].mapping = NULL; | |
1237 | } | |
1238 | ||
8f1d26d0 | 1239 | void free_huge_page(struct page *page) |
27a85ef1 | 1240 | { |
a5516438 AK |
1241 | /* |
1242 | * Can't pass hstate in here because it is called from the | |
1243 | * compound page destructor. | |
1244 | */ | |
e5ff2159 | 1245 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1246 | int nid = page_to_nid(page); |
90481622 DG |
1247 | struct hugepage_subpool *spool = |
1248 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1249 | bool restore_reserve; |
27a85ef1 | 1250 | |
e5df70ab | 1251 | set_page_private(page, 0); |
23be7468 | 1252 | page->mapping = NULL; |
b4330afb MK |
1253 | VM_BUG_ON_PAGE(page_count(page), page); |
1254 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
07443a85 | 1255 | restore_reserve = PagePrivate(page); |
16c794b4 | 1256 | ClearPagePrivate(page); |
27a85ef1 | 1257 | |
1c5ecae3 MK |
1258 | /* |
1259 | * A return code of zero implies that the subpool will be under its | |
1260 | * minimum size if the reservation is not restored after page is free. | |
1261 | * Therefore, force restore_reserve operation. | |
1262 | */ | |
1263 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1264 | restore_reserve = true; | |
1265 | ||
27a85ef1 | 1266 | spin_lock(&hugetlb_lock); |
bcc54222 | 1267 | clear_page_huge_active(page); |
6d76dcf4 AK |
1268 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1269 | pages_per_huge_page(h), page); | |
07443a85 JK |
1270 | if (restore_reserve) |
1271 | h->resv_huge_pages++; | |
1272 | ||
ab5ac90a MH |
1273 | if (PageHugeTemporary(page)) { |
1274 | list_del(&page->lru); | |
1275 | ClearPageHugeTemporary(page); | |
1276 | update_and_free_page(h, page); | |
1277 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1278 | /* remove the page from active list */ |
1279 | list_del(&page->lru); | |
a5516438 AK |
1280 | update_and_free_page(h, page); |
1281 | h->surplus_huge_pages--; | |
1282 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1283 | } else { |
5d3a551c | 1284 | arch_clear_hugepage_flags(page); |
a5516438 | 1285 | enqueue_huge_page(h, page); |
7893d1d5 | 1286 | } |
27a85ef1 DG |
1287 | spin_unlock(&hugetlb_lock); |
1288 | } | |
1289 | ||
a5516438 | 1290 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1291 | { |
0edaecfa | 1292 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1293 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1294 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1295 | set_hugetlb_cgroup(page, NULL); |
a5516438 AK |
1296 | h->nr_huge_pages++; |
1297 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 | 1298 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1299 | } |
1300 | ||
d00181b9 | 1301 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1302 | { |
1303 | int i; | |
1304 | int nr_pages = 1 << order; | |
1305 | struct page *p = page + 1; | |
1306 | ||
1307 | /* we rely on prep_new_huge_page to set the destructor */ | |
1308 | set_compound_order(page, order); | |
ef5a22be | 1309 | __ClearPageReserved(page); |
de09d31d | 1310 | __SetPageHead(page); |
20a0307c | 1311 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1312 | /* |
1313 | * For gigantic hugepages allocated through bootmem at | |
1314 | * boot, it's safer to be consistent with the not-gigantic | |
1315 | * hugepages and clear the PG_reserved bit from all tail pages | |
1316 | * too. Otherwse drivers using get_user_pages() to access tail | |
1317 | * pages may get the reference counting wrong if they see | |
1318 | * PG_reserved set on a tail page (despite the head page not | |
1319 | * having PG_reserved set). Enforcing this consistency between | |
1320 | * head and tail pages allows drivers to optimize away a check | |
1321 | * on the head page when they need know if put_page() is needed | |
1322 | * after get_user_pages(). | |
1323 | */ | |
1324 | __ClearPageReserved(p); | |
58a84aa9 | 1325 | set_page_count(p, 0); |
1d798ca3 | 1326 | set_compound_head(p, page); |
20a0307c | 1327 | } |
b4330afb | 1328 | atomic_set(compound_mapcount_ptr(page), -1); |
20a0307c WF |
1329 | } |
1330 | ||
7795912c AM |
1331 | /* |
1332 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1333 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1334 | * details. | |
1335 | */ | |
20a0307c WF |
1336 | int PageHuge(struct page *page) |
1337 | { | |
20a0307c WF |
1338 | if (!PageCompound(page)) |
1339 | return 0; | |
1340 | ||
1341 | page = compound_head(page); | |
f1e61557 | 1342 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1343 | } |
43131e14 NH |
1344 | EXPORT_SYMBOL_GPL(PageHuge); |
1345 | ||
27c73ae7 AA |
1346 | /* |
1347 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1348 | * normal or transparent huge pages. | |
1349 | */ | |
1350 | int PageHeadHuge(struct page *page_head) | |
1351 | { | |
27c73ae7 AA |
1352 | if (!PageHead(page_head)) |
1353 | return 0; | |
1354 | ||
758f66a2 | 1355 | return get_compound_page_dtor(page_head) == free_huge_page; |
27c73ae7 | 1356 | } |
27c73ae7 | 1357 | |
13d60f4b ZY |
1358 | pgoff_t __basepage_index(struct page *page) |
1359 | { | |
1360 | struct page *page_head = compound_head(page); | |
1361 | pgoff_t index = page_index(page_head); | |
1362 | unsigned long compound_idx; | |
1363 | ||
1364 | if (!PageHuge(page_head)) | |
1365 | return page_index(page); | |
1366 | ||
1367 | if (compound_order(page_head) >= MAX_ORDER) | |
1368 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1369 | else | |
1370 | compound_idx = page - page_head; | |
1371 | ||
1372 | return (index << compound_order(page_head)) + compound_idx; | |
1373 | } | |
1374 | ||
0c397dae | 1375 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
af0fb9df | 1376 | gfp_t gfp_mask, int nid, nodemask_t *nmask) |
1da177e4 | 1377 | { |
af0fb9df | 1378 | int order = huge_page_order(h); |
1da177e4 | 1379 | struct page *page; |
f96efd58 | 1380 | |
af0fb9df MH |
1381 | gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN; |
1382 | if (nid == NUMA_NO_NODE) | |
1383 | nid = numa_mem_id(); | |
1384 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1385 | if (page) | |
1386 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1387 | else | |
1388 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c NA |
1389 | |
1390 | return page; | |
1391 | } | |
1392 | ||
0c397dae MH |
1393 | /* |
1394 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1395 | * should use this function to get new hugetlb pages | |
1396 | */ | |
1397 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
1398 | gfp_t gfp_mask, int nid, nodemask_t *nmask) | |
1399 | { | |
1400 | struct page *page; | |
1401 | ||
1402 | if (hstate_is_gigantic(h)) | |
1403 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1404 | else | |
1405 | page = alloc_buddy_huge_page(h, gfp_mask, | |
1406 | nid, nmask); | |
1407 | if (!page) | |
1408 | return NULL; | |
1409 | ||
1410 | if (hstate_is_gigantic(h)) | |
1411 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1412 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1413 | ||
1414 | return page; | |
1415 | } | |
1416 | ||
af0fb9df MH |
1417 | /* |
1418 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1419 | * manner. | |
1420 | */ | |
0c397dae | 1421 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
b2261026 JK |
1422 | { |
1423 | struct page *page; | |
1424 | int nr_nodes, node; | |
af0fb9df | 1425 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1426 | |
1427 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
0c397dae | 1428 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed); |
af0fb9df | 1429 | if (page) |
b2261026 | 1430 | break; |
b2261026 JK |
1431 | } |
1432 | ||
af0fb9df MH |
1433 | if (!page) |
1434 | return 0; | |
b2261026 | 1435 | |
af0fb9df MH |
1436 | put_page(page); /* free it into the hugepage allocator */ |
1437 | ||
1438 | return 1; | |
b2261026 JK |
1439 | } |
1440 | ||
e8c5c824 LS |
1441 | /* |
1442 | * Free huge page from pool from next node to free. | |
1443 | * Attempt to keep persistent huge pages more or less | |
1444 | * balanced over allowed nodes. | |
1445 | * Called with hugetlb_lock locked. | |
1446 | */ | |
6ae11b27 LS |
1447 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1448 | bool acct_surplus) | |
e8c5c824 | 1449 | { |
b2261026 | 1450 | int nr_nodes, node; |
e8c5c824 LS |
1451 | int ret = 0; |
1452 | ||
b2261026 | 1453 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1454 | /* |
1455 | * If we're returning unused surplus pages, only examine | |
1456 | * nodes with surplus pages. | |
1457 | */ | |
b2261026 JK |
1458 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1459 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1460 | struct page *page = |
b2261026 | 1461 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1462 | struct page, lru); |
1463 | list_del(&page->lru); | |
1464 | h->free_huge_pages--; | |
b2261026 | 1465 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1466 | if (acct_surplus) { |
1467 | h->surplus_huge_pages--; | |
b2261026 | 1468 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1469 | } |
e8c5c824 LS |
1470 | update_and_free_page(h, page); |
1471 | ret = 1; | |
9a76db09 | 1472 | break; |
e8c5c824 | 1473 | } |
b2261026 | 1474 | } |
e8c5c824 LS |
1475 | |
1476 | return ret; | |
1477 | } | |
1478 | ||
c8721bbb NH |
1479 | /* |
1480 | * Dissolve a given free hugepage into free buddy pages. This function does | |
082d5b6b GS |
1481 | * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the |
1482 | * number of free hugepages would be reduced below the number of reserved | |
1483 | * hugepages. | |
c8721bbb | 1484 | */ |
c3114a84 | 1485 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1486 | { |
082d5b6b GS |
1487 | int rc = 0; |
1488 | ||
c8721bbb NH |
1489 | spin_lock(&hugetlb_lock); |
1490 | if (PageHuge(page) && !page_count(page)) { | |
2247bb33 GS |
1491 | struct page *head = compound_head(page); |
1492 | struct hstate *h = page_hstate(head); | |
1493 | int nid = page_to_nid(head); | |
082d5b6b GS |
1494 | if (h->free_huge_pages - h->resv_huge_pages == 0) { |
1495 | rc = -EBUSY; | |
1496 | goto out; | |
1497 | } | |
c3114a84 AK |
1498 | /* |
1499 | * Move PageHWPoison flag from head page to the raw error page, | |
1500 | * which makes any subpages rather than the error page reusable. | |
1501 | */ | |
1502 | if (PageHWPoison(head) && page != head) { | |
1503 | SetPageHWPoison(page); | |
1504 | ClearPageHWPoison(head); | |
1505 | } | |
2247bb33 | 1506 | list_del(&head->lru); |
c8721bbb NH |
1507 | h->free_huge_pages--; |
1508 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1509 | h->max_huge_pages--; |
2247bb33 | 1510 | update_and_free_page(h, head); |
c8721bbb | 1511 | } |
082d5b6b | 1512 | out: |
c8721bbb | 1513 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1514 | return rc; |
c8721bbb NH |
1515 | } |
1516 | ||
1517 | /* | |
1518 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1519 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1520 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1521 | * part of it lies within the given range. | |
082d5b6b GS |
1522 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1523 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1524 | */ |
082d5b6b | 1525 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1526 | { |
c8721bbb | 1527 | unsigned long pfn; |
eb03aa00 | 1528 | struct page *page; |
082d5b6b | 1529 | int rc = 0; |
c8721bbb | 1530 | |
d0177639 | 1531 | if (!hugepages_supported()) |
082d5b6b | 1532 | return rc; |
d0177639 | 1533 | |
eb03aa00 GS |
1534 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1535 | page = pfn_to_page(pfn); | |
1536 | if (PageHuge(page) && !page_count(page)) { | |
1537 | rc = dissolve_free_huge_page(page); | |
1538 | if (rc) | |
1539 | break; | |
1540 | } | |
1541 | } | |
082d5b6b GS |
1542 | |
1543 | return rc; | |
c8721bbb NH |
1544 | } |
1545 | ||
ab5ac90a MH |
1546 | /* |
1547 | * Allocates a fresh surplus page from the page allocator. | |
1548 | */ | |
0c397dae | 1549 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1550 | int nid, nodemask_t *nmask) |
7893d1d5 | 1551 | { |
9980d744 | 1552 | struct page *page = NULL; |
7893d1d5 | 1553 | |
bae7f4ae | 1554 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1555 | return NULL; |
1556 | ||
d1c3fb1f | 1557 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1558 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1559 | goto out_unlock; | |
d1c3fb1f NA |
1560 | spin_unlock(&hugetlb_lock); |
1561 | ||
0c397dae | 1562 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); |
9980d744 | 1563 | if (!page) |
0c397dae | 1564 | return NULL; |
d1c3fb1f NA |
1565 | |
1566 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1567 | /* |
1568 | * We could have raced with the pool size change. | |
1569 | * Double check that and simply deallocate the new page | |
1570 | * if we would end up overcommiting the surpluses. Abuse | |
1571 | * temporary page to workaround the nasty free_huge_page | |
1572 | * codeflow | |
1573 | */ | |
1574 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1575 | SetPageHugeTemporary(page); | |
1576 | put_page(page); | |
1577 | page = NULL; | |
1578 | } else { | |
9980d744 | 1579 | h->surplus_huge_pages++; |
4704dea3 | 1580 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1581 | } |
9980d744 MH |
1582 | |
1583 | out_unlock: | |
d1c3fb1f | 1584 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1585 | |
1586 | return page; | |
1587 | } | |
1588 | ||
0c397dae | 1589 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
ab5ac90a MH |
1590 | int nid, nodemask_t *nmask) |
1591 | { | |
1592 | struct page *page; | |
1593 | ||
1594 | if (hstate_is_gigantic(h)) | |
1595 | return NULL; | |
1596 | ||
0c397dae | 1597 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask); |
ab5ac90a MH |
1598 | if (!page) |
1599 | return NULL; | |
1600 | ||
1601 | /* | |
1602 | * We do not account these pages as surplus because they are only | |
1603 | * temporary and will be released properly on the last reference | |
1604 | */ | |
ab5ac90a MH |
1605 | SetPageHugeTemporary(page); |
1606 | ||
1607 | return page; | |
1608 | } | |
1609 | ||
099730d6 DH |
1610 | /* |
1611 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1612 | */ | |
e0ec90ee | 1613 | static |
0c397dae | 1614 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1615 | struct vm_area_struct *vma, unsigned long addr) |
1616 | { | |
aaf14e40 MH |
1617 | struct page *page; |
1618 | struct mempolicy *mpol; | |
1619 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1620 | int nid; | |
1621 | nodemask_t *nodemask; | |
1622 | ||
1623 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1624 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1625 | mpol_cond_put(mpol); |
1626 | ||
1627 | return page; | |
099730d6 DH |
1628 | } |
1629 | ||
ab5ac90a | 1630 | /* page migration callback function */ |
bf50bab2 NH |
1631 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
1632 | { | |
aaf14e40 | 1633 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1634 | struct page *page = NULL; |
bf50bab2 | 1635 | |
aaf14e40 MH |
1636 | if (nid != NUMA_NO_NODE) |
1637 | gfp_mask |= __GFP_THISNODE; | |
1638 | ||
bf50bab2 | 1639 | spin_lock(&hugetlb_lock); |
4ef91848 | 1640 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1641 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1642 | spin_unlock(&hugetlb_lock); |
1643 | ||
94ae8ba7 | 1644 | if (!page) |
0c397dae | 1645 | page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1646 | |
1647 | return page; | |
1648 | } | |
1649 | ||
ab5ac90a | 1650 | /* page migration callback function */ |
3e59fcb0 MH |
1651 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
1652 | nodemask_t *nmask) | |
4db9b2ef | 1653 | { |
aaf14e40 | 1654 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
1655 | |
1656 | spin_lock(&hugetlb_lock); | |
1657 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1658 | struct page *page; |
1659 | ||
1660 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1661 | if (page) { | |
1662 | spin_unlock(&hugetlb_lock); | |
1663 | return page; | |
4db9b2ef MH |
1664 | } |
1665 | } | |
1666 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1667 | |
0c397dae | 1668 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1669 | } |
1670 | ||
ebd63723 | 1671 | /* mempolicy aware migration callback */ |
389c8178 MH |
1672 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1673 | unsigned long address) | |
ebd63723 MH |
1674 | { |
1675 | struct mempolicy *mpol; | |
1676 | nodemask_t *nodemask; | |
1677 | struct page *page; | |
ebd63723 MH |
1678 | gfp_t gfp_mask; |
1679 | int node; | |
1680 | ||
ebd63723 MH |
1681 | gfp_mask = htlb_alloc_mask(h); |
1682 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
1683 | page = alloc_huge_page_nodemask(h, node, nodemask); | |
1684 | mpol_cond_put(mpol); | |
1685 | ||
1686 | return page; | |
1687 | } | |
1688 | ||
e4e574b7 | 1689 | /* |
25985edc | 1690 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1691 | * of size 'delta'. |
1692 | */ | |
a5516438 | 1693 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
1694 | { |
1695 | struct list_head surplus_list; | |
1696 | struct page *page, *tmp; | |
1697 | int ret, i; | |
1698 | int needed, allocated; | |
28073b02 | 1699 | bool alloc_ok = true; |
e4e574b7 | 1700 | |
a5516438 | 1701 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1702 | if (needed <= 0) { |
a5516438 | 1703 | h->resv_huge_pages += delta; |
e4e574b7 | 1704 | return 0; |
ac09b3a1 | 1705 | } |
e4e574b7 AL |
1706 | |
1707 | allocated = 0; | |
1708 | INIT_LIST_HEAD(&surplus_list); | |
1709 | ||
1710 | ret = -ENOMEM; | |
1711 | retry: | |
1712 | spin_unlock(&hugetlb_lock); | |
1713 | for (i = 0; i < needed; i++) { | |
0c397dae | 1714 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1715 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1716 | if (!page) { |
1717 | alloc_ok = false; | |
1718 | break; | |
1719 | } | |
e4e574b7 | 1720 | list_add(&page->lru, &surplus_list); |
69ed779a | 1721 | cond_resched(); |
e4e574b7 | 1722 | } |
28073b02 | 1723 | allocated += i; |
e4e574b7 AL |
1724 | |
1725 | /* | |
1726 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1727 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1728 | */ | |
1729 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1730 | needed = (h->resv_huge_pages + delta) - |
1731 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1732 | if (needed > 0) { |
1733 | if (alloc_ok) | |
1734 | goto retry; | |
1735 | /* | |
1736 | * We were not able to allocate enough pages to | |
1737 | * satisfy the entire reservation so we free what | |
1738 | * we've allocated so far. | |
1739 | */ | |
1740 | goto free; | |
1741 | } | |
e4e574b7 AL |
1742 | /* |
1743 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 1744 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 1745 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
1746 | * allocator. Commit the entire reservation here to prevent another |
1747 | * process from stealing the pages as they are added to the pool but | |
1748 | * before they are reserved. | |
e4e574b7 AL |
1749 | */ |
1750 | needed += allocated; | |
a5516438 | 1751 | h->resv_huge_pages += delta; |
e4e574b7 | 1752 | ret = 0; |
a9869b83 | 1753 | |
19fc3f0a | 1754 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 1755 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
1756 | if ((--needed) < 0) |
1757 | break; | |
a9869b83 NH |
1758 | /* |
1759 | * This page is now managed by the hugetlb allocator and has | |
1760 | * no users -- drop the buddy allocator's reference. | |
1761 | */ | |
1762 | put_page_testzero(page); | |
309381fe | 1763 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 1764 | enqueue_huge_page(h, page); |
19fc3f0a | 1765 | } |
28073b02 | 1766 | free: |
b0365c8d | 1767 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
1768 | |
1769 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
1770 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
1771 | put_page(page); | |
a9869b83 | 1772 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
1773 | |
1774 | return ret; | |
1775 | } | |
1776 | ||
1777 | /* | |
e5bbc8a6 MK |
1778 | * This routine has two main purposes: |
1779 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
1780 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
1781 | * to the associated reservation map. | |
1782 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
1783 | * the reservation. As many as unused_resv_pages may be freed. | |
1784 | * | |
1785 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
1786 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
1787 | * we must make sure nobody else can claim pages we are in the process of | |
1788 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
1789 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 1790 | */ |
a5516438 AK |
1791 | static void return_unused_surplus_pages(struct hstate *h, |
1792 | unsigned long unused_resv_pages) | |
e4e574b7 | 1793 | { |
e4e574b7 AL |
1794 | unsigned long nr_pages; |
1795 | ||
aa888a74 | 1796 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 1797 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 1798 | goto out; |
aa888a74 | 1799 | |
e5bbc8a6 MK |
1800 | /* |
1801 | * Part (or even all) of the reservation could have been backed | |
1802 | * by pre-allocated pages. Only free surplus pages. | |
1803 | */ | |
a5516438 | 1804 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 1805 | |
685f3457 LS |
1806 | /* |
1807 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
1808 | * evenly across all nodes with memory. Iterate across these nodes |
1809 | * until we can no longer free unreserved surplus pages. This occurs | |
1810 | * when the nodes with surplus pages have no free pages. | |
1811 | * free_pool_huge_page() will balance the the freed pages across the | |
1812 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
1813 | * |
1814 | * Note that we decrement resv_huge_pages as we free the pages. If | |
1815 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
1816 | * to cover subsequent pages we may free. | |
685f3457 LS |
1817 | */ |
1818 | while (nr_pages--) { | |
e5bbc8a6 MK |
1819 | h->resv_huge_pages--; |
1820 | unused_resv_pages--; | |
8cebfcd0 | 1821 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 1822 | goto out; |
7848a4bf | 1823 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 1824 | } |
e5bbc8a6 MK |
1825 | |
1826 | out: | |
1827 | /* Fully uncommit the reservation */ | |
1828 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
1829 | } |
1830 | ||
5e911373 | 1831 | |
c37f9fb1 | 1832 | /* |
feba16e2 | 1833 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 1834 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
1835 | * |
1836 | * vma_needs_reservation is called to determine if the huge page at addr | |
1837 | * within the vma has an associated reservation. If a reservation is | |
1838 | * needed, the value 1 is returned. The caller is then responsible for | |
1839 | * managing the global reservation and subpool usage counts. After | |
1840 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
1841 | * to add the page to the reservation map. If the page allocation fails, |
1842 | * the reservation must be ended instead of committed. vma_end_reservation | |
1843 | * is called in such cases. | |
cf3ad20b MK |
1844 | * |
1845 | * In the normal case, vma_commit_reservation returns the same value | |
1846 | * as the preceding vma_needs_reservation call. The only time this | |
1847 | * is not the case is if a reserve map was changed between calls. It | |
1848 | * is the responsibility of the caller to notice the difference and | |
1849 | * take appropriate action. | |
96b96a96 MK |
1850 | * |
1851 | * vma_add_reservation is used in error paths where a reservation must | |
1852 | * be restored when a newly allocated huge page must be freed. It is | |
1853 | * to be called after calling vma_needs_reservation to determine if a | |
1854 | * reservation exists. | |
c37f9fb1 | 1855 | */ |
5e911373 MK |
1856 | enum vma_resv_mode { |
1857 | VMA_NEEDS_RESV, | |
1858 | VMA_COMMIT_RESV, | |
feba16e2 | 1859 | VMA_END_RESV, |
96b96a96 | 1860 | VMA_ADD_RESV, |
5e911373 | 1861 | }; |
cf3ad20b MK |
1862 | static long __vma_reservation_common(struct hstate *h, |
1863 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 1864 | enum vma_resv_mode mode) |
c37f9fb1 | 1865 | { |
4e35f483 JK |
1866 | struct resv_map *resv; |
1867 | pgoff_t idx; | |
cf3ad20b | 1868 | long ret; |
c37f9fb1 | 1869 | |
4e35f483 JK |
1870 | resv = vma_resv_map(vma); |
1871 | if (!resv) | |
84afd99b | 1872 | return 1; |
c37f9fb1 | 1873 | |
4e35f483 | 1874 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
1875 | switch (mode) { |
1876 | case VMA_NEEDS_RESV: | |
cf3ad20b | 1877 | ret = region_chg(resv, idx, idx + 1); |
5e911373 MK |
1878 | break; |
1879 | case VMA_COMMIT_RESV: | |
1880 | ret = region_add(resv, idx, idx + 1); | |
1881 | break; | |
feba16e2 | 1882 | case VMA_END_RESV: |
5e911373 MK |
1883 | region_abort(resv, idx, idx + 1); |
1884 | ret = 0; | |
1885 | break; | |
96b96a96 MK |
1886 | case VMA_ADD_RESV: |
1887 | if (vma->vm_flags & VM_MAYSHARE) | |
1888 | ret = region_add(resv, idx, idx + 1); | |
1889 | else { | |
1890 | region_abort(resv, idx, idx + 1); | |
1891 | ret = region_del(resv, idx, idx + 1); | |
1892 | } | |
1893 | break; | |
5e911373 MK |
1894 | default: |
1895 | BUG(); | |
1896 | } | |
84afd99b | 1897 | |
4e35f483 | 1898 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 1899 | return ret; |
67961f9d MK |
1900 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
1901 | /* | |
1902 | * In most cases, reserves always exist for private mappings. | |
1903 | * However, a file associated with mapping could have been | |
1904 | * hole punched or truncated after reserves were consumed. | |
1905 | * As subsequent fault on such a range will not use reserves. | |
1906 | * Subtle - The reserve map for private mappings has the | |
1907 | * opposite meaning than that of shared mappings. If NO | |
1908 | * entry is in the reserve map, it means a reservation exists. | |
1909 | * If an entry exists in the reserve map, it means the | |
1910 | * reservation has already been consumed. As a result, the | |
1911 | * return value of this routine is the opposite of the | |
1912 | * value returned from reserve map manipulation routines above. | |
1913 | */ | |
1914 | if (ret) | |
1915 | return 0; | |
1916 | else | |
1917 | return 1; | |
1918 | } | |
4e35f483 | 1919 | else |
cf3ad20b | 1920 | return ret < 0 ? ret : 0; |
c37f9fb1 | 1921 | } |
cf3ad20b MK |
1922 | |
1923 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 1924 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 1925 | { |
5e911373 | 1926 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 1927 | } |
84afd99b | 1928 | |
cf3ad20b MK |
1929 | static long vma_commit_reservation(struct hstate *h, |
1930 | struct vm_area_struct *vma, unsigned long addr) | |
1931 | { | |
5e911373 MK |
1932 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
1933 | } | |
1934 | ||
feba16e2 | 1935 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
1936 | struct vm_area_struct *vma, unsigned long addr) |
1937 | { | |
feba16e2 | 1938 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
1939 | } |
1940 | ||
96b96a96 MK |
1941 | static long vma_add_reservation(struct hstate *h, |
1942 | struct vm_area_struct *vma, unsigned long addr) | |
1943 | { | |
1944 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
1945 | } | |
1946 | ||
1947 | /* | |
1948 | * This routine is called to restore a reservation on error paths. In the | |
1949 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
1950 | * and is about to be freed. If a reservation for the page existed, | |
1951 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
1952 | * in the newly allocated page. When the page is freed via free_huge_page, | |
1953 | * the global reservation count will be incremented if PagePrivate is set. | |
1954 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
1955 | * reserve map here to be consistent with global reserve count adjustments | |
1956 | * to be made by free_huge_page. | |
1957 | */ | |
1958 | static void restore_reserve_on_error(struct hstate *h, | |
1959 | struct vm_area_struct *vma, unsigned long address, | |
1960 | struct page *page) | |
1961 | { | |
1962 | if (unlikely(PagePrivate(page))) { | |
1963 | long rc = vma_needs_reservation(h, vma, address); | |
1964 | ||
1965 | if (unlikely(rc < 0)) { | |
1966 | /* | |
1967 | * Rare out of memory condition in reserve map | |
1968 | * manipulation. Clear PagePrivate so that | |
1969 | * global reserve count will not be incremented | |
1970 | * by free_huge_page. This will make it appear | |
1971 | * as though the reservation for this page was | |
1972 | * consumed. This may prevent the task from | |
1973 | * faulting in the page at a later time. This | |
1974 | * is better than inconsistent global huge page | |
1975 | * accounting of reserve counts. | |
1976 | */ | |
1977 | ClearPagePrivate(page); | |
1978 | } else if (rc) { | |
1979 | rc = vma_add_reservation(h, vma, address); | |
1980 | if (unlikely(rc < 0)) | |
1981 | /* | |
1982 | * See above comment about rare out of | |
1983 | * memory condition. | |
1984 | */ | |
1985 | ClearPagePrivate(page); | |
1986 | } else | |
1987 | vma_end_reservation(h, vma, address); | |
1988 | } | |
1989 | } | |
1990 | ||
70c3547e | 1991 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 1992 | unsigned long addr, int avoid_reserve) |
1da177e4 | 1993 | { |
90481622 | 1994 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 1995 | struct hstate *h = hstate_vma(vma); |
348ea204 | 1996 | struct page *page; |
d85f69b0 MK |
1997 | long map_chg, map_commit; |
1998 | long gbl_chg; | |
6d76dcf4 AK |
1999 | int ret, idx; |
2000 | struct hugetlb_cgroup *h_cg; | |
a1e78772 | 2001 | |
6d76dcf4 | 2002 | idx = hstate_index(h); |
a1e78772 | 2003 | /* |
d85f69b0 MK |
2004 | * Examine the region/reserve map to determine if the process |
2005 | * has a reservation for the page to be allocated. A return | |
2006 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2007 | */ |
d85f69b0 MK |
2008 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2009 | if (map_chg < 0) | |
76dcee75 | 2010 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2011 | |
2012 | /* | |
2013 | * Processes that did not create the mapping will have no | |
2014 | * reserves as indicated by the region/reserve map. Check | |
2015 | * that the allocation will not exceed the subpool limit. | |
2016 | * Allocations for MAP_NORESERVE mappings also need to be | |
2017 | * checked against any subpool limit. | |
2018 | */ | |
2019 | if (map_chg || avoid_reserve) { | |
2020 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2021 | if (gbl_chg < 0) { | |
feba16e2 | 2022 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2023 | return ERR_PTR(-ENOSPC); |
5e911373 | 2024 | } |
1da177e4 | 2025 | |
d85f69b0 MK |
2026 | /* |
2027 | * Even though there was no reservation in the region/reserve | |
2028 | * map, there could be reservations associated with the | |
2029 | * subpool that can be used. This would be indicated if the | |
2030 | * return value of hugepage_subpool_get_pages() is zero. | |
2031 | * However, if avoid_reserve is specified we still avoid even | |
2032 | * the subpool reservations. | |
2033 | */ | |
2034 | if (avoid_reserve) | |
2035 | gbl_chg = 1; | |
2036 | } | |
2037 | ||
6d76dcf4 | 2038 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f JZ |
2039 | if (ret) |
2040 | goto out_subpool_put; | |
2041 | ||
1da177e4 | 2042 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2043 | /* |
2044 | * glb_chg is passed to indicate whether or not a page must be taken | |
2045 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2046 | * a reservation exists for the allocation. | |
2047 | */ | |
2048 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2049 | if (!page) { |
94ae8ba7 | 2050 | spin_unlock(&hugetlb_lock); |
0c397dae | 2051 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2052 | if (!page) |
2053 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2054 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2055 | SetPagePrivate(page); | |
2056 | h->resv_huge_pages--; | |
2057 | } | |
79dbb236 AK |
2058 | spin_lock(&hugetlb_lock); |
2059 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2060 | /* Fall through */ |
68842c9b | 2061 | } |
81a6fcae JK |
2062 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
2063 | spin_unlock(&hugetlb_lock); | |
348ea204 | 2064 | |
90481622 | 2065 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2066 | |
d85f69b0 MK |
2067 | map_commit = vma_commit_reservation(h, vma, addr); |
2068 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2069 | /* |
2070 | * The page was added to the reservation map between | |
2071 | * vma_needs_reservation and vma_commit_reservation. | |
2072 | * This indicates a race with hugetlb_reserve_pages. | |
2073 | * Adjust for the subpool count incremented above AND | |
2074 | * in hugetlb_reserve_pages for the same page. Also, | |
2075 | * the reservation count added in hugetlb_reserve_pages | |
2076 | * no longer applies. | |
2077 | */ | |
2078 | long rsv_adjust; | |
2079 | ||
2080 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2081 | hugetlb_acct_memory(h, -rsv_adjust); | |
2082 | } | |
90d8b7e6 | 2083 | return page; |
8f34af6f JZ |
2084 | |
2085 | out_uncharge_cgroup: | |
2086 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
2087 | out_subpool_put: | |
d85f69b0 | 2088 | if (map_chg || avoid_reserve) |
8f34af6f | 2089 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2090 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2091 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2092 | } |
2093 | ||
e24a1307 AK |
2094 | int alloc_bootmem_huge_page(struct hstate *h) |
2095 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2096 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2097 | { |
2098 | struct huge_bootmem_page *m; | |
b2261026 | 2099 | int nr_nodes, node; |
aa888a74 | 2100 | |
b2261026 | 2101 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2102 | void *addr; |
2103 | ||
8b89a116 GS |
2104 | addr = memblock_virt_alloc_try_nid_nopanic( |
2105 | huge_page_size(h), huge_page_size(h), | |
2106 | 0, BOOTMEM_ALLOC_ACCESSIBLE, node); | |
aa888a74 AK |
2107 | if (addr) { |
2108 | /* | |
2109 | * Use the beginning of the huge page to store the | |
2110 | * huge_bootmem_page struct (until gather_bootmem | |
2111 | * puts them into the mem_map). | |
2112 | */ | |
2113 | m = addr; | |
91f47662 | 2114 | goto found; |
aa888a74 | 2115 | } |
aa888a74 AK |
2116 | } |
2117 | return 0; | |
2118 | ||
2119 | found: | |
df994ead | 2120 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 AK |
2121 | /* Put them into a private list first because mem_map is not up yet */ |
2122 | list_add(&m->list, &huge_boot_pages); | |
2123 | m->hstate = h; | |
2124 | return 1; | |
2125 | } | |
2126 | ||
d00181b9 KS |
2127 | static void __init prep_compound_huge_page(struct page *page, |
2128 | unsigned int order) | |
18229df5 AW |
2129 | { |
2130 | if (unlikely(order > (MAX_ORDER - 1))) | |
2131 | prep_compound_gigantic_page(page, order); | |
2132 | else | |
2133 | prep_compound_page(page, order); | |
2134 | } | |
2135 | ||
aa888a74 AK |
2136 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2137 | static void __init gather_bootmem_prealloc(void) | |
2138 | { | |
2139 | struct huge_bootmem_page *m; | |
2140 | ||
2141 | list_for_each_entry(m, &huge_boot_pages, list) { | |
aa888a74 | 2142 | struct hstate *h = m->hstate; |
ee8f248d BB |
2143 | struct page *page; |
2144 | ||
2145 | #ifdef CONFIG_HIGHMEM | |
2146 | page = pfn_to_page(m->phys >> PAGE_SHIFT); | |
8b89a116 GS |
2147 | memblock_free_late(__pa(m), |
2148 | sizeof(struct huge_bootmem_page)); | |
ee8f248d BB |
2149 | #else |
2150 | page = virt_to_page(m); | |
2151 | #endif | |
aa888a74 | 2152 | WARN_ON(page_count(page) != 1); |
18229df5 | 2153 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2154 | WARN_ON(PageReserved(page)); |
aa888a74 | 2155 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2156 | put_page(page); /* free it into the hugepage allocator */ |
2157 | ||
b0320c7b RA |
2158 | /* |
2159 | * If we had gigantic hugepages allocated at boot time, we need | |
2160 | * to restore the 'stolen' pages to totalram_pages in order to | |
2161 | * fix confusing memory reports from free(1) and another | |
2162 | * side-effects, like CommitLimit going negative. | |
2163 | */ | |
bae7f4ae | 2164 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2165 | adjust_managed_page_count(page, 1 << h->order); |
aa888a74 AK |
2166 | } |
2167 | } | |
2168 | ||
8faa8b07 | 2169 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2170 | { |
2171 | unsigned long i; | |
a5516438 | 2172 | |
e5ff2159 | 2173 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2174 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2175 | if (!alloc_bootmem_huge_page(h)) |
2176 | break; | |
0c397dae | 2177 | } else if (!alloc_pool_huge_page(h, |
8cebfcd0 | 2178 | &node_states[N_MEMORY])) |
1da177e4 | 2179 | break; |
69ed779a | 2180 | cond_resched(); |
1da177e4 | 2181 | } |
d715cf80 LH |
2182 | if (i < h->max_huge_pages) { |
2183 | char buf[32]; | |
2184 | ||
c6247f72 | 2185 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2186 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2187 | h->max_huge_pages, buf, i); | |
2188 | h->max_huge_pages = i; | |
2189 | } | |
e5ff2159 AK |
2190 | } |
2191 | ||
2192 | static void __init hugetlb_init_hstates(void) | |
2193 | { | |
2194 | struct hstate *h; | |
2195 | ||
2196 | for_each_hstate(h) { | |
641844f5 NH |
2197 | if (minimum_order > huge_page_order(h)) |
2198 | minimum_order = huge_page_order(h); | |
2199 | ||
8faa8b07 | 2200 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2201 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2202 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2203 | } |
641844f5 | 2204 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2205 | } |
2206 | ||
2207 | static void __init report_hugepages(void) | |
2208 | { | |
2209 | struct hstate *h; | |
2210 | ||
2211 | for_each_hstate(h) { | |
4abd32db | 2212 | char buf[32]; |
c6247f72 MW |
2213 | |
2214 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2215 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2216 | buf, h->free_huge_pages); |
e5ff2159 AK |
2217 | } |
2218 | } | |
2219 | ||
1da177e4 | 2220 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2221 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2222 | nodemask_t *nodes_allowed) | |
1da177e4 | 2223 | { |
4415cc8d CL |
2224 | int i; |
2225 | ||
bae7f4ae | 2226 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2227 | return; |
2228 | ||
6ae11b27 | 2229 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2230 | struct page *page, *next; |
a5516438 AK |
2231 | struct list_head *freel = &h->hugepage_freelists[i]; |
2232 | list_for_each_entry_safe(page, next, freel, lru) { | |
2233 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2234 | return; |
1da177e4 LT |
2235 | if (PageHighMem(page)) |
2236 | continue; | |
2237 | list_del(&page->lru); | |
e5ff2159 | 2238 | update_and_free_page(h, page); |
a5516438 AK |
2239 | h->free_huge_pages--; |
2240 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2241 | } |
2242 | } | |
2243 | } | |
2244 | #else | |
6ae11b27 LS |
2245 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2246 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2247 | { |
2248 | } | |
2249 | #endif | |
2250 | ||
20a0307c WF |
2251 | /* |
2252 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2253 | * balanced by operating on them in a round-robin fashion. | |
2254 | * Returns 1 if an adjustment was made. | |
2255 | */ | |
6ae11b27 LS |
2256 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2257 | int delta) | |
20a0307c | 2258 | { |
b2261026 | 2259 | int nr_nodes, node; |
20a0307c WF |
2260 | |
2261 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2262 | |
b2261026 JK |
2263 | if (delta < 0) { |
2264 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2265 | if (h->surplus_huge_pages_node[node]) | |
2266 | goto found; | |
e8c5c824 | 2267 | } |
b2261026 JK |
2268 | } else { |
2269 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2270 | if (h->surplus_huge_pages_node[node] < | |
2271 | h->nr_huge_pages_node[node]) | |
2272 | goto found; | |
e8c5c824 | 2273 | } |
b2261026 JK |
2274 | } |
2275 | return 0; | |
20a0307c | 2276 | |
b2261026 JK |
2277 | found: |
2278 | h->surplus_huge_pages += delta; | |
2279 | h->surplus_huge_pages_node[node] += delta; | |
2280 | return 1; | |
20a0307c WF |
2281 | } |
2282 | ||
a5516438 | 2283 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
6ae11b27 LS |
2284 | static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count, |
2285 | nodemask_t *nodes_allowed) | |
1da177e4 | 2286 | { |
7893d1d5 | 2287 | unsigned long min_count, ret; |
1da177e4 | 2288 | |
944d9fec | 2289 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) |
aa888a74 AK |
2290 | return h->max_huge_pages; |
2291 | ||
7893d1d5 AL |
2292 | /* |
2293 | * Increase the pool size | |
2294 | * First take pages out of surplus state. Then make up the | |
2295 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2296 | * |
0c397dae | 2297 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2298 | * to convert a surplus huge page to a normal huge page. That is |
2299 | * not critical, though, it just means the overall size of the | |
2300 | * pool might be one hugepage larger than it needs to be, but | |
2301 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2302 | */ |
1da177e4 | 2303 | spin_lock(&hugetlb_lock); |
a5516438 | 2304 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2305 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2306 | break; |
2307 | } | |
2308 | ||
a5516438 | 2309 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2310 | /* |
2311 | * If this allocation races such that we no longer need the | |
2312 | * page, free_huge_page will handle it by freeing the page | |
2313 | * and reducing the surplus. | |
2314 | */ | |
2315 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2316 | |
2317 | /* yield cpu to avoid soft lockup */ | |
2318 | cond_resched(); | |
2319 | ||
0c397dae | 2320 | ret = alloc_pool_huge_page(h, nodes_allowed); |
7893d1d5 AL |
2321 | spin_lock(&hugetlb_lock); |
2322 | if (!ret) | |
2323 | goto out; | |
2324 | ||
536240f2 MG |
2325 | /* Bail for signals. Probably ctrl-c from user */ |
2326 | if (signal_pending(current)) | |
2327 | goto out; | |
7893d1d5 | 2328 | } |
7893d1d5 AL |
2329 | |
2330 | /* | |
2331 | * Decrease the pool size | |
2332 | * First return free pages to the buddy allocator (being careful | |
2333 | * to keep enough around to satisfy reservations). Then place | |
2334 | * pages into surplus state as needed so the pool will shrink | |
2335 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2336 | * |
2337 | * By placing pages into the surplus state independent of the | |
2338 | * overcommit value, we are allowing the surplus pool size to | |
2339 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2340 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2341 | * though, we'll note that we're not allowed to exceed surplus |
2342 | * and won't grow the pool anywhere else. Not until one of the | |
2343 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2344 | */ |
a5516438 | 2345 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2346 | min_count = max(count, min_count); |
6ae11b27 | 2347 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2348 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2349 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2350 | break; |
55f67141 | 2351 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2352 | } |
a5516438 | 2353 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2354 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2355 | break; |
2356 | } | |
2357 | out: | |
a5516438 | 2358 | ret = persistent_huge_pages(h); |
1da177e4 | 2359 | spin_unlock(&hugetlb_lock); |
7893d1d5 | 2360 | return ret; |
1da177e4 LT |
2361 | } |
2362 | ||
a3437870 NA |
2363 | #define HSTATE_ATTR_RO(_name) \ |
2364 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2365 | ||
2366 | #define HSTATE_ATTR(_name) \ | |
2367 | static struct kobj_attribute _name##_attr = \ | |
2368 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2369 | ||
2370 | static struct kobject *hugepages_kobj; | |
2371 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2372 | ||
9a305230 LS |
2373 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2374 | ||
2375 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2376 | { |
2377 | int i; | |
9a305230 | 2378 | |
a3437870 | 2379 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2380 | if (hstate_kobjs[i] == kobj) { |
2381 | if (nidp) | |
2382 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2383 | return &hstates[i]; |
9a305230 LS |
2384 | } |
2385 | ||
2386 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2387 | } |
2388 | ||
06808b08 | 2389 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2390 | struct kobj_attribute *attr, char *buf) |
2391 | { | |
9a305230 LS |
2392 | struct hstate *h; |
2393 | unsigned long nr_huge_pages; | |
2394 | int nid; | |
2395 | ||
2396 | h = kobj_to_hstate(kobj, &nid); | |
2397 | if (nid == NUMA_NO_NODE) | |
2398 | nr_huge_pages = h->nr_huge_pages; | |
2399 | else | |
2400 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2401 | ||
2402 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2403 | } |
adbe8726 | 2404 | |
238d3c13 DR |
2405 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2406 | struct hstate *h, int nid, | |
2407 | unsigned long count, size_t len) | |
a3437870 NA |
2408 | { |
2409 | int err; | |
bad44b5b | 2410 | NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY); |
a3437870 | 2411 | |
944d9fec | 2412 | if (hstate_is_gigantic(h) && !gigantic_page_supported()) { |
adbe8726 EM |
2413 | err = -EINVAL; |
2414 | goto out; | |
2415 | } | |
2416 | ||
9a305230 LS |
2417 | if (nid == NUMA_NO_NODE) { |
2418 | /* | |
2419 | * global hstate attribute | |
2420 | */ | |
2421 | if (!(obey_mempolicy && | |
2422 | init_nodemask_of_mempolicy(nodes_allowed))) { | |
2423 | NODEMASK_FREE(nodes_allowed); | |
8cebfcd0 | 2424 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 LS |
2425 | } |
2426 | } else if (nodes_allowed) { | |
2427 | /* | |
2428 | * per node hstate attribute: adjust count to global, | |
2429 | * but restrict alloc/free to the specified node. | |
2430 | */ | |
2431 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2432 | init_nodemask_of_node(nodes_allowed, nid); | |
2433 | } else | |
8cebfcd0 | 2434 | nodes_allowed = &node_states[N_MEMORY]; |
9a305230 | 2435 | |
06808b08 | 2436 | h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed); |
a3437870 | 2437 | |
8cebfcd0 | 2438 | if (nodes_allowed != &node_states[N_MEMORY]) |
06808b08 LS |
2439 | NODEMASK_FREE(nodes_allowed); |
2440 | ||
2441 | return len; | |
adbe8726 EM |
2442 | out: |
2443 | NODEMASK_FREE(nodes_allowed); | |
2444 | return err; | |
06808b08 LS |
2445 | } |
2446 | ||
238d3c13 DR |
2447 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2448 | struct kobject *kobj, const char *buf, | |
2449 | size_t len) | |
2450 | { | |
2451 | struct hstate *h; | |
2452 | unsigned long count; | |
2453 | int nid; | |
2454 | int err; | |
2455 | ||
2456 | err = kstrtoul(buf, 10, &count); | |
2457 | if (err) | |
2458 | return err; | |
2459 | ||
2460 | h = kobj_to_hstate(kobj, &nid); | |
2461 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2462 | } | |
2463 | ||
06808b08 LS |
2464 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2465 | struct kobj_attribute *attr, char *buf) | |
2466 | { | |
2467 | return nr_hugepages_show_common(kobj, attr, buf); | |
2468 | } | |
2469 | ||
2470 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2471 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2472 | { | |
238d3c13 | 2473 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2474 | } |
2475 | HSTATE_ATTR(nr_hugepages); | |
2476 | ||
06808b08 LS |
2477 | #ifdef CONFIG_NUMA |
2478 | ||
2479 | /* | |
2480 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2481 | * huge page alloc/free. | |
2482 | */ | |
2483 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2484 | struct kobj_attribute *attr, char *buf) | |
2485 | { | |
2486 | return nr_hugepages_show_common(kobj, attr, buf); | |
2487 | } | |
2488 | ||
2489 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2490 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2491 | { | |
238d3c13 | 2492 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2493 | } |
2494 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2495 | #endif | |
2496 | ||
2497 | ||
a3437870 NA |
2498 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2499 | struct kobj_attribute *attr, char *buf) | |
2500 | { | |
9a305230 | 2501 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2502 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2503 | } | |
adbe8726 | 2504 | |
a3437870 NA |
2505 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2506 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2507 | { | |
2508 | int err; | |
2509 | unsigned long input; | |
9a305230 | 2510 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2511 | |
bae7f4ae | 2512 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2513 | return -EINVAL; |
2514 | ||
3dbb95f7 | 2515 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2516 | if (err) |
73ae31e5 | 2517 | return err; |
a3437870 NA |
2518 | |
2519 | spin_lock(&hugetlb_lock); | |
2520 | h->nr_overcommit_huge_pages = input; | |
2521 | spin_unlock(&hugetlb_lock); | |
2522 | ||
2523 | return count; | |
2524 | } | |
2525 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2526 | ||
2527 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2528 | struct kobj_attribute *attr, char *buf) | |
2529 | { | |
9a305230 LS |
2530 | struct hstate *h; |
2531 | unsigned long free_huge_pages; | |
2532 | int nid; | |
2533 | ||
2534 | h = kobj_to_hstate(kobj, &nid); | |
2535 | if (nid == NUMA_NO_NODE) | |
2536 | free_huge_pages = h->free_huge_pages; | |
2537 | else | |
2538 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2539 | ||
2540 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2541 | } |
2542 | HSTATE_ATTR_RO(free_hugepages); | |
2543 | ||
2544 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2545 | struct kobj_attribute *attr, char *buf) | |
2546 | { | |
9a305230 | 2547 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2548 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2549 | } | |
2550 | HSTATE_ATTR_RO(resv_hugepages); | |
2551 | ||
2552 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2553 | struct kobj_attribute *attr, char *buf) | |
2554 | { | |
9a305230 LS |
2555 | struct hstate *h; |
2556 | unsigned long surplus_huge_pages; | |
2557 | int nid; | |
2558 | ||
2559 | h = kobj_to_hstate(kobj, &nid); | |
2560 | if (nid == NUMA_NO_NODE) | |
2561 | surplus_huge_pages = h->surplus_huge_pages; | |
2562 | else | |
2563 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2564 | ||
2565 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2566 | } |
2567 | HSTATE_ATTR_RO(surplus_hugepages); | |
2568 | ||
2569 | static struct attribute *hstate_attrs[] = { | |
2570 | &nr_hugepages_attr.attr, | |
2571 | &nr_overcommit_hugepages_attr.attr, | |
2572 | &free_hugepages_attr.attr, | |
2573 | &resv_hugepages_attr.attr, | |
2574 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2575 | #ifdef CONFIG_NUMA |
2576 | &nr_hugepages_mempolicy_attr.attr, | |
2577 | #endif | |
a3437870 NA |
2578 | NULL, |
2579 | }; | |
2580 | ||
67e5ed96 | 2581 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2582 | .attrs = hstate_attrs, |
2583 | }; | |
2584 | ||
094e9539 JM |
2585 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2586 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2587 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2588 | { |
2589 | int retval; | |
972dc4de | 2590 | int hi = hstate_index(h); |
a3437870 | 2591 | |
9a305230 LS |
2592 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2593 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2594 | return -ENOMEM; |
2595 | ||
9a305230 | 2596 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 2597 | if (retval) |
9a305230 | 2598 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
2599 | |
2600 | return retval; | |
2601 | } | |
2602 | ||
2603 | static void __init hugetlb_sysfs_init(void) | |
2604 | { | |
2605 | struct hstate *h; | |
2606 | int err; | |
2607 | ||
2608 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2609 | if (!hugepages_kobj) | |
2610 | return; | |
2611 | ||
2612 | for_each_hstate(h) { | |
9a305230 LS |
2613 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2614 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2615 | if (err) |
ffb22af5 | 2616 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
2617 | } |
2618 | } | |
2619 | ||
9a305230 LS |
2620 | #ifdef CONFIG_NUMA |
2621 | ||
2622 | /* | |
2623 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2624 | * with node devices in node_devices[] using a parallel array. The array |
2625 | * index of a node device or _hstate == node id. | |
2626 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2627 | * the base kernel, on the hugetlb module. |
2628 | */ | |
2629 | struct node_hstate { | |
2630 | struct kobject *hugepages_kobj; | |
2631 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2632 | }; | |
b4e289a6 | 2633 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
2634 | |
2635 | /* | |
10fbcf4c | 2636 | * A subset of global hstate attributes for node devices |
9a305230 LS |
2637 | */ |
2638 | static struct attribute *per_node_hstate_attrs[] = { | |
2639 | &nr_hugepages_attr.attr, | |
2640 | &free_hugepages_attr.attr, | |
2641 | &surplus_hugepages_attr.attr, | |
2642 | NULL, | |
2643 | }; | |
2644 | ||
67e5ed96 | 2645 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
2646 | .attrs = per_node_hstate_attrs, |
2647 | }; | |
2648 | ||
2649 | /* | |
10fbcf4c | 2650 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
2651 | * Returns node id via non-NULL nidp. |
2652 | */ | |
2653 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2654 | { | |
2655 | int nid; | |
2656 | ||
2657 | for (nid = 0; nid < nr_node_ids; nid++) { | |
2658 | struct node_hstate *nhs = &node_hstates[nid]; | |
2659 | int i; | |
2660 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
2661 | if (nhs->hstate_kobjs[i] == kobj) { | |
2662 | if (nidp) | |
2663 | *nidp = nid; | |
2664 | return &hstates[i]; | |
2665 | } | |
2666 | } | |
2667 | ||
2668 | BUG(); | |
2669 | return NULL; | |
2670 | } | |
2671 | ||
2672 | /* | |
10fbcf4c | 2673 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
2674 | * No-op if no hstate attributes attached. |
2675 | */ | |
3cd8b44f | 2676 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
2677 | { |
2678 | struct hstate *h; | |
10fbcf4c | 2679 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2680 | |
2681 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 2682 | return; /* no hstate attributes */ |
9a305230 | 2683 | |
972dc4de AK |
2684 | for_each_hstate(h) { |
2685 | int idx = hstate_index(h); | |
2686 | if (nhs->hstate_kobjs[idx]) { | |
2687 | kobject_put(nhs->hstate_kobjs[idx]); | |
2688 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 2689 | } |
972dc4de | 2690 | } |
9a305230 LS |
2691 | |
2692 | kobject_put(nhs->hugepages_kobj); | |
2693 | nhs->hugepages_kobj = NULL; | |
2694 | } | |
2695 | ||
9a305230 LS |
2696 | |
2697 | /* | |
10fbcf4c | 2698 | * Register hstate attributes for a single node device. |
9a305230 LS |
2699 | * No-op if attributes already registered. |
2700 | */ | |
3cd8b44f | 2701 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
2702 | { |
2703 | struct hstate *h; | |
10fbcf4c | 2704 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
2705 | int err; |
2706 | ||
2707 | if (nhs->hugepages_kobj) | |
2708 | return; /* already allocated */ | |
2709 | ||
2710 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 2711 | &node->dev.kobj); |
9a305230 LS |
2712 | if (!nhs->hugepages_kobj) |
2713 | return; | |
2714 | ||
2715 | for_each_hstate(h) { | |
2716 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
2717 | nhs->hstate_kobjs, | |
2718 | &per_node_hstate_attr_group); | |
2719 | if (err) { | |
ffb22af5 AM |
2720 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
2721 | h->name, node->dev.id); | |
9a305230 LS |
2722 | hugetlb_unregister_node(node); |
2723 | break; | |
2724 | } | |
2725 | } | |
2726 | } | |
2727 | ||
2728 | /* | |
9b5e5d0f | 2729 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
2730 | * devices of nodes that have memory. All on-line nodes should have |
2731 | * registered their associated device by this time. | |
9a305230 | 2732 | */ |
7d9ca000 | 2733 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
2734 | { |
2735 | int nid; | |
2736 | ||
8cebfcd0 | 2737 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 2738 | struct node *node = node_devices[nid]; |
10fbcf4c | 2739 | if (node->dev.id == nid) |
9a305230 LS |
2740 | hugetlb_register_node(node); |
2741 | } | |
2742 | ||
2743 | /* | |
10fbcf4c | 2744 | * Let the node device driver know we're here so it can |
9a305230 LS |
2745 | * [un]register hstate attributes on node hotplug. |
2746 | */ | |
2747 | register_hugetlbfs_with_node(hugetlb_register_node, | |
2748 | hugetlb_unregister_node); | |
2749 | } | |
2750 | #else /* !CONFIG_NUMA */ | |
2751 | ||
2752 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
2753 | { | |
2754 | BUG(); | |
2755 | if (nidp) | |
2756 | *nidp = -1; | |
2757 | return NULL; | |
2758 | } | |
2759 | ||
9a305230 LS |
2760 | static void hugetlb_register_all_nodes(void) { } |
2761 | ||
2762 | #endif | |
2763 | ||
a3437870 NA |
2764 | static int __init hugetlb_init(void) |
2765 | { | |
8382d914 DB |
2766 | int i; |
2767 | ||
457c1b27 | 2768 | if (!hugepages_supported()) |
0ef89d25 | 2769 | return 0; |
a3437870 | 2770 | |
e11bfbfc | 2771 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
2772 | if (default_hstate_size != 0) { |
2773 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
2774 | default_hstate_size, HPAGE_SIZE); | |
2775 | } | |
2776 | ||
e11bfbfc NP |
2777 | default_hstate_size = HPAGE_SIZE; |
2778 | if (!size_to_hstate(default_hstate_size)) | |
2779 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 2780 | } |
972dc4de | 2781 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
2782 | if (default_hstate_max_huge_pages) { |
2783 | if (!default_hstate.max_huge_pages) | |
2784 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
2785 | } | |
a3437870 NA |
2786 | |
2787 | hugetlb_init_hstates(); | |
aa888a74 | 2788 | gather_bootmem_prealloc(); |
a3437870 NA |
2789 | report_hugepages(); |
2790 | ||
2791 | hugetlb_sysfs_init(); | |
9a305230 | 2792 | hugetlb_register_all_nodes(); |
7179e7bf | 2793 | hugetlb_cgroup_file_init(); |
9a305230 | 2794 | |
8382d914 DB |
2795 | #ifdef CONFIG_SMP |
2796 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
2797 | #else | |
2798 | num_fault_mutexes = 1; | |
2799 | #endif | |
c672c7f2 | 2800 | hugetlb_fault_mutex_table = |
8382d914 | 2801 | kmalloc(sizeof(struct mutex) * num_fault_mutexes, GFP_KERNEL); |
c672c7f2 | 2802 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
2803 | |
2804 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 2805 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
2806 | return 0; |
2807 | } | |
3e89e1c5 | 2808 | subsys_initcall(hugetlb_init); |
a3437870 NA |
2809 | |
2810 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
2811 | void __init hugetlb_bad_size(void) |
2812 | { | |
2813 | parsed_valid_hugepagesz = false; | |
2814 | } | |
2815 | ||
d00181b9 | 2816 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
2817 | { |
2818 | struct hstate *h; | |
8faa8b07 AK |
2819 | unsigned long i; |
2820 | ||
a3437870 | 2821 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 2822 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
2823 | return; |
2824 | } | |
47d38344 | 2825 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 2826 | BUG_ON(order == 0); |
47d38344 | 2827 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
2828 | h->order = order; |
2829 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
2830 | h->nr_huge_pages = 0; |
2831 | h->free_huge_pages = 0; | |
2832 | for (i = 0; i < MAX_NUMNODES; ++i) | |
2833 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 2834 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
2835 | h->next_nid_to_alloc = first_memory_node; |
2836 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
2837 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
2838 | huge_page_size(h)/1024); | |
8faa8b07 | 2839 | |
a3437870 NA |
2840 | parsed_hstate = h; |
2841 | } | |
2842 | ||
e11bfbfc | 2843 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
2844 | { |
2845 | unsigned long *mhp; | |
8faa8b07 | 2846 | static unsigned long *last_mhp; |
a3437870 | 2847 | |
9fee021d VT |
2848 | if (!parsed_valid_hugepagesz) { |
2849 | pr_warn("hugepages = %s preceded by " | |
2850 | "an unsupported hugepagesz, ignoring\n", s); | |
2851 | parsed_valid_hugepagesz = true; | |
2852 | return 1; | |
2853 | } | |
a3437870 | 2854 | /* |
47d38344 | 2855 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
2856 | * so this hugepages= parameter goes to the "default hstate". |
2857 | */ | |
9fee021d | 2858 | else if (!hugetlb_max_hstate) |
a3437870 NA |
2859 | mhp = &default_hstate_max_huge_pages; |
2860 | else | |
2861 | mhp = &parsed_hstate->max_huge_pages; | |
2862 | ||
8faa8b07 | 2863 | if (mhp == last_mhp) { |
598d8091 | 2864 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
2865 | return 1; |
2866 | } | |
2867 | ||
a3437870 NA |
2868 | if (sscanf(s, "%lu", mhp) <= 0) |
2869 | *mhp = 0; | |
2870 | ||
8faa8b07 AK |
2871 | /* |
2872 | * Global state is always initialized later in hugetlb_init. | |
2873 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
2874 | * use the bootmem allocator. | |
2875 | */ | |
47d38344 | 2876 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
2877 | hugetlb_hstate_alloc_pages(parsed_hstate); |
2878 | ||
2879 | last_mhp = mhp; | |
2880 | ||
a3437870 NA |
2881 | return 1; |
2882 | } | |
e11bfbfc NP |
2883 | __setup("hugepages=", hugetlb_nrpages_setup); |
2884 | ||
2885 | static int __init hugetlb_default_setup(char *s) | |
2886 | { | |
2887 | default_hstate_size = memparse(s, &s); | |
2888 | return 1; | |
2889 | } | |
2890 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 2891 | |
8a213460 NA |
2892 | static unsigned int cpuset_mems_nr(unsigned int *array) |
2893 | { | |
2894 | int node; | |
2895 | unsigned int nr = 0; | |
2896 | ||
2897 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
2898 | nr += array[node]; | |
2899 | ||
2900 | return nr; | |
2901 | } | |
2902 | ||
2903 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
2904 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
2905 | struct ctl_table *table, int write, | |
2906 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 2907 | { |
e5ff2159 | 2908 | struct hstate *h = &default_hstate; |
238d3c13 | 2909 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 2910 | int ret; |
e5ff2159 | 2911 | |
457c1b27 | 2912 | if (!hugepages_supported()) |
86613628 | 2913 | return -EOPNOTSUPP; |
457c1b27 | 2914 | |
e5ff2159 AK |
2915 | table->data = &tmp; |
2916 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2917 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2918 | if (ret) | |
2919 | goto out; | |
e5ff2159 | 2920 | |
238d3c13 DR |
2921 | if (write) |
2922 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
2923 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
2924 | out: |
2925 | return ret; | |
1da177e4 | 2926 | } |
396faf03 | 2927 | |
06808b08 LS |
2928 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
2929 | void __user *buffer, size_t *length, loff_t *ppos) | |
2930 | { | |
2931 | ||
2932 | return hugetlb_sysctl_handler_common(false, table, write, | |
2933 | buffer, length, ppos); | |
2934 | } | |
2935 | ||
2936 | #ifdef CONFIG_NUMA | |
2937 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
2938 | void __user *buffer, size_t *length, loff_t *ppos) | |
2939 | { | |
2940 | return hugetlb_sysctl_handler_common(true, table, write, | |
2941 | buffer, length, ppos); | |
2942 | } | |
2943 | #endif /* CONFIG_NUMA */ | |
2944 | ||
a3d0c6aa | 2945 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 2946 | void __user *buffer, |
a3d0c6aa NA |
2947 | size_t *length, loff_t *ppos) |
2948 | { | |
a5516438 | 2949 | struct hstate *h = &default_hstate; |
e5ff2159 | 2950 | unsigned long tmp; |
08d4a246 | 2951 | int ret; |
e5ff2159 | 2952 | |
457c1b27 | 2953 | if (!hugepages_supported()) |
86613628 | 2954 | return -EOPNOTSUPP; |
457c1b27 | 2955 | |
c033a93c | 2956 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 2957 | |
bae7f4ae | 2958 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
2959 | return -EINVAL; |
2960 | ||
e5ff2159 AK |
2961 | table->data = &tmp; |
2962 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
2963 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
2964 | if (ret) | |
2965 | goto out; | |
e5ff2159 AK |
2966 | |
2967 | if (write) { | |
2968 | spin_lock(&hugetlb_lock); | |
2969 | h->nr_overcommit_huge_pages = tmp; | |
2970 | spin_unlock(&hugetlb_lock); | |
2971 | } | |
08d4a246 MH |
2972 | out: |
2973 | return ret; | |
a3d0c6aa NA |
2974 | } |
2975 | ||
1da177e4 LT |
2976 | #endif /* CONFIG_SYSCTL */ |
2977 | ||
e1759c21 | 2978 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 2979 | { |
fcb2b0c5 RG |
2980 | struct hstate *h; |
2981 | unsigned long total = 0; | |
2982 | ||
457c1b27 NA |
2983 | if (!hugepages_supported()) |
2984 | return; | |
fcb2b0c5 RG |
2985 | |
2986 | for_each_hstate(h) { | |
2987 | unsigned long count = h->nr_huge_pages; | |
2988 | ||
2989 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
2990 | ||
2991 | if (h == &default_hstate) | |
2992 | seq_printf(m, | |
2993 | "HugePages_Total: %5lu\n" | |
2994 | "HugePages_Free: %5lu\n" | |
2995 | "HugePages_Rsvd: %5lu\n" | |
2996 | "HugePages_Surp: %5lu\n" | |
2997 | "Hugepagesize: %8lu kB\n", | |
2998 | count, | |
2999 | h->free_huge_pages, | |
3000 | h->resv_huge_pages, | |
3001 | h->surplus_huge_pages, | |
3002 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3003 | } | |
3004 | ||
3005 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3006 | } |
3007 | ||
3008 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3009 | { | |
a5516438 | 3010 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3011 | if (!hugepages_supported()) |
3012 | return 0; | |
1da177e4 LT |
3013 | return sprintf(buf, |
3014 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3015 | "Node %d HugePages_Free: %5u\n" |
3016 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3017 | nid, h->nr_huge_pages_node[nid], |
3018 | nid, h->free_huge_pages_node[nid], | |
3019 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3020 | } |
3021 | ||
949f7ec5 DR |
3022 | void hugetlb_show_meminfo(void) |
3023 | { | |
3024 | struct hstate *h; | |
3025 | int nid; | |
3026 | ||
457c1b27 NA |
3027 | if (!hugepages_supported()) |
3028 | return; | |
3029 | ||
949f7ec5 DR |
3030 | for_each_node_state(nid, N_MEMORY) |
3031 | for_each_hstate(h) | |
3032 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3033 | nid, | |
3034 | h->nr_huge_pages_node[nid], | |
3035 | h->free_huge_pages_node[nid], | |
3036 | h->surplus_huge_pages_node[nid], | |
3037 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3038 | } | |
3039 | ||
5d317b2b NH |
3040 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3041 | { | |
3042 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3043 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3044 | } | |
3045 | ||
1da177e4 LT |
3046 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3047 | unsigned long hugetlb_total_pages(void) | |
3048 | { | |
d0028588 WL |
3049 | struct hstate *h; |
3050 | unsigned long nr_total_pages = 0; | |
3051 | ||
3052 | for_each_hstate(h) | |
3053 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3054 | return nr_total_pages; | |
1da177e4 | 3055 | } |
1da177e4 | 3056 | |
a5516438 | 3057 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3058 | { |
3059 | int ret = -ENOMEM; | |
3060 | ||
3061 | spin_lock(&hugetlb_lock); | |
3062 | /* | |
3063 | * When cpuset is configured, it breaks the strict hugetlb page | |
3064 | * reservation as the accounting is done on a global variable. Such | |
3065 | * reservation is completely rubbish in the presence of cpuset because | |
3066 | * the reservation is not checked against page availability for the | |
3067 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3068 | * with lack of free htlb page in cpuset that the task is in. | |
3069 | * Attempt to enforce strict accounting with cpuset is almost | |
3070 | * impossible (or too ugly) because cpuset is too fluid that | |
3071 | * task or memory node can be dynamically moved between cpusets. | |
3072 | * | |
3073 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3074 | * undesirable. However, in order to preserve some of the semantics, | |
3075 | * we fall back to check against current free page availability as | |
3076 | * a best attempt and hopefully to minimize the impact of changing | |
3077 | * semantics that cpuset has. | |
3078 | */ | |
3079 | if (delta > 0) { | |
a5516438 | 3080 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3081 | goto out; |
3082 | ||
a5516438 AK |
3083 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3084 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3085 | goto out; |
3086 | } | |
3087 | } | |
3088 | ||
3089 | ret = 0; | |
3090 | if (delta < 0) | |
a5516438 | 3091 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3092 | |
3093 | out: | |
3094 | spin_unlock(&hugetlb_lock); | |
3095 | return ret; | |
3096 | } | |
3097 | ||
84afd99b AW |
3098 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3099 | { | |
f522c3ac | 3100 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3101 | |
3102 | /* | |
3103 | * This new VMA should share its siblings reservation map if present. | |
3104 | * The VMA will only ever have a valid reservation map pointer where | |
3105 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3106 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3107 | * after this open call completes. It is therefore safe to take a |
3108 | * new reference here without additional locking. | |
3109 | */ | |
4e35f483 | 3110 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3111 | kref_get(&resv->refs); |
84afd99b AW |
3112 | } |
3113 | ||
a1e78772 MG |
3114 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3115 | { | |
a5516438 | 3116 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3117 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3118 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3119 | unsigned long reserve, start, end; |
1c5ecae3 | 3120 | long gbl_reserve; |
84afd99b | 3121 | |
4e35f483 JK |
3122 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3123 | return; | |
84afd99b | 3124 | |
4e35f483 JK |
3125 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3126 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3127 | |
4e35f483 | 3128 | reserve = (end - start) - region_count(resv, start, end); |
84afd99b | 3129 | |
4e35f483 JK |
3130 | kref_put(&resv->refs, resv_map_release); |
3131 | ||
3132 | if (reserve) { | |
1c5ecae3 MK |
3133 | /* |
3134 | * Decrement reserve counts. The global reserve count may be | |
3135 | * adjusted if the subpool has a minimum size. | |
3136 | */ | |
3137 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3138 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3139 | } |
a1e78772 MG |
3140 | } |
3141 | ||
31383c68 DW |
3142 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3143 | { | |
3144 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3145 | return -EINVAL; | |
3146 | return 0; | |
3147 | } | |
3148 | ||
05ea8860 DW |
3149 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3150 | { | |
3151 | struct hstate *hstate = hstate_vma(vma); | |
3152 | ||
3153 | return 1UL << huge_page_shift(hstate); | |
3154 | } | |
3155 | ||
1da177e4 LT |
3156 | /* |
3157 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3158 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3159 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3160 | * this far. | |
3161 | */ | |
11bac800 | 3162 | static int hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3163 | { |
3164 | BUG(); | |
d0217ac0 | 3165 | return 0; |
1da177e4 LT |
3166 | } |
3167 | ||
f0f37e2f | 3168 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3169 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3170 | .open = hugetlb_vm_op_open, |
a1e78772 | 3171 | .close = hugetlb_vm_op_close, |
31383c68 | 3172 | .split = hugetlb_vm_op_split, |
05ea8860 | 3173 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3174 | }; |
3175 | ||
1e8f889b DG |
3176 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3177 | int writable) | |
63551ae0 DG |
3178 | { |
3179 | pte_t entry; | |
3180 | ||
1e8f889b | 3181 | if (writable) { |
106c992a GS |
3182 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3183 | vma->vm_page_prot))); | |
63551ae0 | 3184 | } else { |
106c992a GS |
3185 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3186 | vma->vm_page_prot)); | |
63551ae0 DG |
3187 | } |
3188 | entry = pte_mkyoung(entry); | |
3189 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3190 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3191 | |
3192 | return entry; | |
3193 | } | |
3194 | ||
1e8f889b DG |
3195 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3196 | unsigned long address, pte_t *ptep) | |
3197 | { | |
3198 | pte_t entry; | |
3199 | ||
106c992a | 3200 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3201 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3202 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3203 | } |
3204 | ||
d5ed7444 | 3205 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3206 | { |
3207 | swp_entry_t swp; | |
3208 | ||
3209 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3210 | return false; |
4a705fef NH |
3211 | swp = pte_to_swp_entry(pte); |
3212 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3213 | return true; |
4a705fef | 3214 | else |
d5ed7444 | 3215 | return false; |
4a705fef NH |
3216 | } |
3217 | ||
3218 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3219 | { | |
3220 | swp_entry_t swp; | |
3221 | ||
3222 | if (huge_pte_none(pte) || pte_present(pte)) | |
3223 | return 0; | |
3224 | swp = pte_to_swp_entry(pte); | |
3225 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3226 | return 1; | |
3227 | else | |
3228 | return 0; | |
3229 | } | |
1e8f889b | 3230 | |
63551ae0 DG |
3231 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3232 | struct vm_area_struct *vma) | |
3233 | { | |
3234 | pte_t *src_pte, *dst_pte, entry; | |
3235 | struct page *ptepage; | |
1c59827d | 3236 | unsigned long addr; |
1e8f889b | 3237 | int cow; |
a5516438 AK |
3238 | struct hstate *h = hstate_vma(vma); |
3239 | unsigned long sz = huge_page_size(h); | |
e8569dd2 AS |
3240 | unsigned long mmun_start; /* For mmu_notifiers */ |
3241 | unsigned long mmun_end; /* For mmu_notifiers */ | |
3242 | int ret = 0; | |
1e8f889b DG |
3243 | |
3244 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3245 | |
e8569dd2 AS |
3246 | mmun_start = vma->vm_start; |
3247 | mmun_end = vma->vm_end; | |
3248 | if (cow) | |
3249 | mmu_notifier_invalidate_range_start(src, mmun_start, mmun_end); | |
3250 | ||
a5516438 | 3251 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3252 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3253 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3254 | if (!src_pte) |
3255 | continue; | |
a5516438 | 3256 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3257 | if (!dst_pte) { |
3258 | ret = -ENOMEM; | |
3259 | break; | |
3260 | } | |
c5c99429 LW |
3261 | |
3262 | /* If the pagetables are shared don't copy or take references */ | |
3263 | if (dst_pte == src_pte) | |
3264 | continue; | |
3265 | ||
cb900f41 KS |
3266 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3267 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3268 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef NH |
3269 | entry = huge_ptep_get(src_pte); |
3270 | if (huge_pte_none(entry)) { /* skip none entry */ | |
3271 | ; | |
3272 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3273 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3274 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3275 | ||
3276 | if (is_write_migration_entry(swp_entry) && cow) { | |
3277 | /* | |
3278 | * COW mappings require pages in both | |
3279 | * parent and child to be set to read. | |
3280 | */ | |
3281 | make_migration_entry_read(&swp_entry); | |
3282 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3283 | set_huge_swap_pte_at(src, addr, src_pte, |
3284 | entry, sz); | |
4a705fef | 3285 | } |
e5251fd4 | 3286 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3287 | } else { |
34ee645e | 3288 | if (cow) { |
0f10851e JG |
3289 | /* |
3290 | * No need to notify as we are downgrading page | |
3291 | * table protection not changing it to point | |
3292 | * to a new page. | |
3293 | * | |
3294 | * See Documentation/vm/mmu_notifier.txt | |
3295 | */ | |
7f2e9525 | 3296 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3297 | } |
0253d634 | 3298 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3299 | ptepage = pte_page(entry); |
3300 | get_page(ptepage); | |
53f9263b | 3301 | page_dup_rmap(ptepage, true); |
1c59827d | 3302 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3303 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3304 | } |
cb900f41 KS |
3305 | spin_unlock(src_ptl); |
3306 | spin_unlock(dst_ptl); | |
63551ae0 | 3307 | } |
63551ae0 | 3308 | |
e8569dd2 AS |
3309 | if (cow) |
3310 | mmu_notifier_invalidate_range_end(src, mmun_start, mmun_end); | |
3311 | ||
3312 | return ret; | |
63551ae0 DG |
3313 | } |
3314 | ||
24669e58 AK |
3315 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3316 | unsigned long start, unsigned long end, | |
3317 | struct page *ref_page) | |
63551ae0 DG |
3318 | { |
3319 | struct mm_struct *mm = vma->vm_mm; | |
3320 | unsigned long address; | |
c7546f8f | 3321 | pte_t *ptep; |
63551ae0 | 3322 | pte_t pte; |
cb900f41 | 3323 | spinlock_t *ptl; |
63551ae0 | 3324 | struct page *page; |
a5516438 AK |
3325 | struct hstate *h = hstate_vma(vma); |
3326 | unsigned long sz = huge_page_size(h); | |
2ec74c3e SG |
3327 | const unsigned long mmun_start = start; /* For mmu_notifiers */ |
3328 | const unsigned long mmun_end = end; /* For mmu_notifiers */ | |
a5516438 | 3329 | |
63551ae0 | 3330 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3331 | BUG_ON(start & ~huge_page_mask(h)); |
3332 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3333 | |
07e32661 AK |
3334 | /* |
3335 | * This is a hugetlb vma, all the pte entries should point | |
3336 | * to huge page. | |
3337 | */ | |
3338 | tlb_remove_check_page_size_change(tlb, sz); | |
24669e58 | 3339 | tlb_start_vma(tlb, vma); |
2ec74c3e | 3340 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
569f48b8 | 3341 | address = start; |
569f48b8 | 3342 | for (; address < end; address += sz) { |
7868a208 | 3343 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3344 | if (!ptep) |
c7546f8f DG |
3345 | continue; |
3346 | ||
cb900f41 | 3347 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3348 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3349 | spin_unlock(ptl); | |
3350 | continue; | |
3351 | } | |
39dde65c | 3352 | |
6629326b | 3353 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3354 | if (huge_pte_none(pte)) { |
3355 | spin_unlock(ptl); | |
3356 | continue; | |
3357 | } | |
6629326b HD |
3358 | |
3359 | /* | |
9fbc1f63 NH |
3360 | * Migrating hugepage or HWPoisoned hugepage is already |
3361 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3362 | */ |
9fbc1f63 | 3363 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3364 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3365 | spin_unlock(ptl); |
3366 | continue; | |
8c4894c6 | 3367 | } |
6629326b HD |
3368 | |
3369 | page = pte_page(pte); | |
04f2cbe3 MG |
3370 | /* |
3371 | * If a reference page is supplied, it is because a specific | |
3372 | * page is being unmapped, not a range. Ensure the page we | |
3373 | * are about to unmap is the actual page of interest. | |
3374 | */ | |
3375 | if (ref_page) { | |
31d49da5 AK |
3376 | if (page != ref_page) { |
3377 | spin_unlock(ptl); | |
3378 | continue; | |
3379 | } | |
04f2cbe3 MG |
3380 | /* |
3381 | * Mark the VMA as having unmapped its page so that | |
3382 | * future faults in this VMA will fail rather than | |
3383 | * looking like data was lost | |
3384 | */ | |
3385 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3386 | } | |
3387 | ||
c7546f8f | 3388 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3389 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3390 | if (huge_pte_dirty(pte)) |
6649a386 | 3391 | set_page_dirty(page); |
9e81130b | 3392 | |
5d317b2b | 3393 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3394 | page_remove_rmap(page, true); |
31d49da5 | 3395 | |
cb900f41 | 3396 | spin_unlock(ptl); |
e77b0852 | 3397 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3398 | /* |
3399 | * Bail out after unmapping reference page if supplied | |
3400 | */ | |
3401 | if (ref_page) | |
3402 | break; | |
fe1668ae | 3403 | } |
2ec74c3e | 3404 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
24669e58 | 3405 | tlb_end_vma(tlb, vma); |
1da177e4 | 3406 | } |
63551ae0 | 3407 | |
d833352a MG |
3408 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3409 | struct vm_area_struct *vma, unsigned long start, | |
3410 | unsigned long end, struct page *ref_page) | |
3411 | { | |
3412 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3413 | ||
3414 | /* | |
3415 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3416 | * test will fail on a vma being torn down, and not grab a page table | |
3417 | * on its way out. We're lucky that the flag has such an appropriate | |
3418 | * name, and can in fact be safely cleared here. We could clear it | |
3419 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3420 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3421 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3422 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3423 | */ |
3424 | vma->vm_flags &= ~VM_MAYSHARE; | |
3425 | } | |
3426 | ||
502717f4 | 3427 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3428 | unsigned long end, struct page *ref_page) |
502717f4 | 3429 | { |
24669e58 AK |
3430 | struct mm_struct *mm; |
3431 | struct mmu_gather tlb; | |
3432 | ||
3433 | mm = vma->vm_mm; | |
3434 | ||
2b047252 | 3435 | tlb_gather_mmu(&tlb, mm, start, end); |
24669e58 AK |
3436 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
3437 | tlb_finish_mmu(&tlb, start, end); | |
502717f4 CK |
3438 | } |
3439 | ||
04f2cbe3 MG |
3440 | /* |
3441 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3442 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3443 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3444 | * same region. | |
3445 | */ | |
2f4612af DB |
3446 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3447 | struct page *page, unsigned long address) | |
04f2cbe3 | 3448 | { |
7526674d | 3449 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3450 | struct vm_area_struct *iter_vma; |
3451 | struct address_space *mapping; | |
04f2cbe3 MG |
3452 | pgoff_t pgoff; |
3453 | ||
3454 | /* | |
3455 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3456 | * from page cache lookup which is in HPAGE_SIZE units. | |
3457 | */ | |
7526674d | 3458 | address = address & huge_page_mask(h); |
36e4f20a MH |
3459 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3460 | vma->vm_pgoff; | |
93c76a3d | 3461 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3462 | |
4eb2b1dc MG |
3463 | /* |
3464 | * Take the mapping lock for the duration of the table walk. As | |
3465 | * this mapping should be shared between all the VMAs, | |
3466 | * __unmap_hugepage_range() is called as the lock is already held | |
3467 | */ | |
83cde9e8 | 3468 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3469 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3470 | /* Do not unmap the current VMA */ |
3471 | if (iter_vma == vma) | |
3472 | continue; | |
3473 | ||
2f84a899 MG |
3474 | /* |
3475 | * Shared VMAs have their own reserves and do not affect | |
3476 | * MAP_PRIVATE accounting but it is possible that a shared | |
3477 | * VMA is using the same page so check and skip such VMAs. | |
3478 | */ | |
3479 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3480 | continue; | |
3481 | ||
04f2cbe3 MG |
3482 | /* |
3483 | * Unmap the page from other VMAs without their own reserves. | |
3484 | * They get marked to be SIGKILLed if they fault in these | |
3485 | * areas. This is because a future no-page fault on this VMA | |
3486 | * could insert a zeroed page instead of the data existing | |
3487 | * from the time of fork. This would look like data corruption | |
3488 | */ | |
3489 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3490 | unmap_hugepage_range(iter_vma, address, |
3491 | address + huge_page_size(h), page); | |
04f2cbe3 | 3492 | } |
83cde9e8 | 3493 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3494 | } |
3495 | ||
0fe6e20b NH |
3496 | /* |
3497 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3498 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3499 | * cannot race with other handlers or page migration. | |
3500 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3501 | */ |
1e8f889b | 3502 | static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
3999f52e AK |
3503 | unsigned long address, pte_t *ptep, |
3504 | struct page *pagecache_page, spinlock_t *ptl) | |
1e8f889b | 3505 | { |
3999f52e | 3506 | pte_t pte; |
a5516438 | 3507 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3508 | struct page *old_page, *new_page; |
ad4404a2 | 3509 | int ret = 0, outside_reserve = 0; |
2ec74c3e SG |
3510 | unsigned long mmun_start; /* For mmu_notifiers */ |
3511 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1e8f889b | 3512 | |
3999f52e | 3513 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3514 | old_page = pte_page(pte); |
3515 | ||
04f2cbe3 | 3516 | retry_avoidcopy: |
1e8f889b DG |
3517 | /* If no-one else is actually using this page, avoid the copy |
3518 | * and just make the page writable */ | |
37a2140d | 3519 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3520 | page_move_anon_rmap(old_page, vma); |
1e8f889b | 3521 | set_huge_ptep_writable(vma, address, ptep); |
83c54070 | 3522 | return 0; |
1e8f889b DG |
3523 | } |
3524 | ||
04f2cbe3 MG |
3525 | /* |
3526 | * If the process that created a MAP_PRIVATE mapping is about to | |
3527 | * perform a COW due to a shared page count, attempt to satisfy | |
3528 | * the allocation without using the existing reserves. The pagecache | |
3529 | * page is used to determine if the reserve at this address was | |
3530 | * consumed or not. If reserves were used, a partial faulted mapping | |
3531 | * at the time of fork() could consume its reserves on COW instead | |
3532 | * of the full address range. | |
3533 | */ | |
5944d011 | 3534 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
3535 | old_page != pagecache_page) |
3536 | outside_reserve = 1; | |
3537 | ||
09cbfeaf | 3538 | get_page(old_page); |
b76c8cfb | 3539 | |
ad4404a2 DB |
3540 | /* |
3541 | * Drop page table lock as buddy allocator may be called. It will | |
3542 | * be acquired again before returning to the caller, as expected. | |
3543 | */ | |
cb900f41 | 3544 | spin_unlock(ptl); |
04f2cbe3 | 3545 | new_page = alloc_huge_page(vma, address, outside_reserve); |
1e8f889b | 3546 | |
2fc39cec | 3547 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
3548 | /* |
3549 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
3550 | * it is due to references held by a child and an insufficient | |
3551 | * huge page pool. To guarantee the original mappers | |
3552 | * reliability, unmap the page from child processes. The child | |
3553 | * may get SIGKILLed if it later faults. | |
3554 | */ | |
3555 | if (outside_reserve) { | |
09cbfeaf | 3556 | put_page(old_page); |
04f2cbe3 | 3557 | BUG_ON(huge_pte_none(pte)); |
2f4612af DB |
3558 | unmap_ref_private(mm, vma, old_page, address); |
3559 | BUG_ON(huge_pte_none(pte)); | |
3560 | spin_lock(ptl); | |
7868a208 PA |
3561 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3562 | huge_page_size(h)); | |
2f4612af DB |
3563 | if (likely(ptep && |
3564 | pte_same(huge_ptep_get(ptep), pte))) | |
3565 | goto retry_avoidcopy; | |
3566 | /* | |
3567 | * race occurs while re-acquiring page table | |
3568 | * lock, and our job is done. | |
3569 | */ | |
3570 | return 0; | |
04f2cbe3 MG |
3571 | } |
3572 | ||
ad4404a2 DB |
3573 | ret = (PTR_ERR(new_page) == -ENOMEM) ? |
3574 | VM_FAULT_OOM : VM_FAULT_SIGBUS; | |
3575 | goto out_release_old; | |
1e8f889b DG |
3576 | } |
3577 | ||
0fe6e20b NH |
3578 | /* |
3579 | * When the original hugepage is shared one, it does not have | |
3580 | * anon_vma prepared. | |
3581 | */ | |
44e2aa93 | 3582 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
3583 | ret = VM_FAULT_OOM; |
3584 | goto out_release_all; | |
44e2aa93 | 3585 | } |
0fe6e20b | 3586 | |
47ad8475 AA |
3587 | copy_user_huge_page(new_page, old_page, address, vma, |
3588 | pages_per_huge_page(h)); | |
0ed361de | 3589 | __SetPageUptodate(new_page); |
bcc54222 | 3590 | set_page_huge_active(new_page); |
1e8f889b | 3591 | |
2ec74c3e SG |
3592 | mmun_start = address & huge_page_mask(h); |
3593 | mmun_end = mmun_start + huge_page_size(h); | |
3594 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
ad4404a2 | 3595 | |
b76c8cfb | 3596 | /* |
cb900f41 | 3597 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
3598 | * before the page tables are altered |
3599 | */ | |
cb900f41 | 3600 | spin_lock(ptl); |
7868a208 PA |
3601 | ptep = huge_pte_offset(mm, address & huge_page_mask(h), |
3602 | huge_page_size(h)); | |
a9af0c5d | 3603 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
3604 | ClearPagePrivate(new_page); |
3605 | ||
1e8f889b | 3606 | /* Break COW */ |
8fe627ec | 3607 | huge_ptep_clear_flush(vma, address, ptep); |
34ee645e | 3608 | mmu_notifier_invalidate_range(mm, mmun_start, mmun_end); |
1e8f889b DG |
3609 | set_huge_pte_at(mm, address, ptep, |
3610 | make_huge_pte(vma, new_page, 1)); | |
d281ee61 | 3611 | page_remove_rmap(old_page, true); |
cd67f0d2 | 3612 | hugepage_add_new_anon_rmap(new_page, vma, address); |
1e8f889b DG |
3613 | /* Make the old page be freed below */ |
3614 | new_page = old_page; | |
3615 | } | |
cb900f41 | 3616 | spin_unlock(ptl); |
2ec74c3e | 3617 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
ad4404a2 | 3618 | out_release_all: |
96b96a96 | 3619 | restore_reserve_on_error(h, vma, address, new_page); |
09cbfeaf | 3620 | put_page(new_page); |
ad4404a2 | 3621 | out_release_old: |
09cbfeaf | 3622 | put_page(old_page); |
8312034f | 3623 | |
ad4404a2 DB |
3624 | spin_lock(ptl); /* Caller expects lock to be held */ |
3625 | return ret; | |
1e8f889b DG |
3626 | } |
3627 | ||
04f2cbe3 | 3628 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
3629 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
3630 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
3631 | { |
3632 | struct address_space *mapping; | |
e7c4b0bf | 3633 | pgoff_t idx; |
04f2cbe3 MG |
3634 | |
3635 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 3636 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
3637 | |
3638 | return find_lock_page(mapping, idx); | |
3639 | } | |
3640 | ||
3ae77f43 HD |
3641 | /* |
3642 | * Return whether there is a pagecache page to back given address within VMA. | |
3643 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
3644 | */ | |
3645 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
3646 | struct vm_area_struct *vma, unsigned long address) |
3647 | { | |
3648 | struct address_space *mapping; | |
3649 | pgoff_t idx; | |
3650 | struct page *page; | |
3651 | ||
3652 | mapping = vma->vm_file->f_mapping; | |
3653 | idx = vma_hugecache_offset(h, vma, address); | |
3654 | ||
3655 | page = find_get_page(mapping, idx); | |
3656 | if (page) | |
3657 | put_page(page); | |
3658 | return page != NULL; | |
3659 | } | |
3660 | ||
ab76ad54 MK |
3661 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
3662 | pgoff_t idx) | |
3663 | { | |
3664 | struct inode *inode = mapping->host; | |
3665 | struct hstate *h = hstate_inode(inode); | |
3666 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
3667 | ||
3668 | if (err) | |
3669 | return err; | |
3670 | ClearPagePrivate(page); | |
3671 | ||
3672 | spin_lock(&inode->i_lock); | |
3673 | inode->i_blocks += blocks_per_huge_page(h); | |
3674 | spin_unlock(&inode->i_lock); | |
3675 | return 0; | |
3676 | } | |
3677 | ||
a1ed3dda | 3678 | static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma, |
8382d914 DB |
3679 | struct address_space *mapping, pgoff_t idx, |
3680 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 3681 | { |
a5516438 | 3682 | struct hstate *h = hstate_vma(vma); |
ac9b9c66 | 3683 | int ret = VM_FAULT_SIGBUS; |
409eb8c2 | 3684 | int anon_rmap = 0; |
4c887265 | 3685 | unsigned long size; |
4c887265 | 3686 | struct page *page; |
1e8f889b | 3687 | pte_t new_pte; |
cb900f41 | 3688 | spinlock_t *ptl; |
4c887265 | 3689 | |
04f2cbe3 MG |
3690 | /* |
3691 | * Currently, we are forced to kill the process in the event the | |
3692 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 3693 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
3694 | */ |
3695 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 3696 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 3697 | current->pid); |
04f2cbe3 MG |
3698 | return ret; |
3699 | } | |
3700 | ||
4c887265 AL |
3701 | /* |
3702 | * Use page lock to guard against racing truncation | |
3703 | * before we get page_table_lock. | |
3704 | */ | |
6bda666a CL |
3705 | retry: |
3706 | page = find_lock_page(mapping, idx); | |
3707 | if (!page) { | |
a5516438 | 3708 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
ebed4bfc HD |
3709 | if (idx >= size) |
3710 | goto out; | |
1a1aad8a MK |
3711 | |
3712 | /* | |
3713 | * Check for page in userfault range | |
3714 | */ | |
3715 | if (userfaultfd_missing(vma)) { | |
3716 | u32 hash; | |
3717 | struct vm_fault vmf = { | |
3718 | .vma = vma, | |
3719 | .address = address, | |
3720 | .flags = flags, | |
3721 | /* | |
3722 | * Hard to debug if it ends up being | |
3723 | * used by a callee that assumes | |
3724 | * something about the other | |
3725 | * uninitialized fields... same as in | |
3726 | * memory.c | |
3727 | */ | |
3728 | }; | |
3729 | ||
3730 | /* | |
3731 | * hugetlb_fault_mutex must be dropped before | |
3732 | * handling userfault. Reacquire after handling | |
3733 | * fault to make calling code simpler. | |
3734 | */ | |
3735 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, | |
3736 | idx, address); | |
3737 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
3738 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); | |
3739 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
3740 | goto out; | |
3741 | } | |
3742 | ||
04f2cbe3 | 3743 | page = alloc_huge_page(vma, address, 0); |
2fc39cec | 3744 | if (IS_ERR(page)) { |
76dcee75 AK |
3745 | ret = PTR_ERR(page); |
3746 | if (ret == -ENOMEM) | |
3747 | ret = VM_FAULT_OOM; | |
3748 | else | |
3749 | ret = VM_FAULT_SIGBUS; | |
6bda666a CL |
3750 | goto out; |
3751 | } | |
47ad8475 | 3752 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 3753 | __SetPageUptodate(page); |
bcc54222 | 3754 | set_page_huge_active(page); |
ac9b9c66 | 3755 | |
f83a275d | 3756 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 3757 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
3758 | if (err) { |
3759 | put_page(page); | |
6bda666a CL |
3760 | if (err == -EEXIST) |
3761 | goto retry; | |
3762 | goto out; | |
3763 | } | |
23be7468 | 3764 | } else { |
6bda666a | 3765 | lock_page(page); |
0fe6e20b NH |
3766 | if (unlikely(anon_vma_prepare(vma))) { |
3767 | ret = VM_FAULT_OOM; | |
3768 | goto backout_unlocked; | |
3769 | } | |
409eb8c2 | 3770 | anon_rmap = 1; |
23be7468 | 3771 | } |
0fe6e20b | 3772 | } else { |
998b4382 NH |
3773 | /* |
3774 | * If memory error occurs between mmap() and fault, some process | |
3775 | * don't have hwpoisoned swap entry for errored virtual address. | |
3776 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
3777 | */ | |
3778 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 3779 | ret = VM_FAULT_HWPOISON | |
972dc4de | 3780 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
3781 | goto backout_unlocked; |
3782 | } | |
6bda666a | 3783 | } |
1e8f889b | 3784 | |
57303d80 AW |
3785 | /* |
3786 | * If we are going to COW a private mapping later, we examine the | |
3787 | * pending reservations for this page now. This will ensure that | |
3788 | * any allocations necessary to record that reservation occur outside | |
3789 | * the spinlock. | |
3790 | */ | |
5e911373 | 3791 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
2b26736c AW |
3792 | if (vma_needs_reservation(h, vma, address) < 0) { |
3793 | ret = VM_FAULT_OOM; | |
3794 | goto backout_unlocked; | |
3795 | } | |
5e911373 | 3796 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3797 | vma_end_reservation(h, vma, address); |
5e911373 | 3798 | } |
57303d80 | 3799 | |
8bea8052 | 3800 | ptl = huge_pte_lock(h, mm, ptep); |
a5516438 | 3801 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4c887265 AL |
3802 | if (idx >= size) |
3803 | goto backout; | |
3804 | ||
83c54070 | 3805 | ret = 0; |
7f2e9525 | 3806 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
3807 | goto backout; |
3808 | ||
07443a85 JK |
3809 | if (anon_rmap) { |
3810 | ClearPagePrivate(page); | |
409eb8c2 | 3811 | hugepage_add_new_anon_rmap(page, vma, address); |
ac714904 | 3812 | } else |
53f9263b | 3813 | page_dup_rmap(page, true); |
1e8f889b DG |
3814 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
3815 | && (vma->vm_flags & VM_SHARED))); | |
3816 | set_huge_pte_at(mm, address, ptep, new_pte); | |
3817 | ||
5d317b2b | 3818 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 3819 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 3820 | /* Optimization, do the COW without a second fault */ |
3999f52e | 3821 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
3822 | } |
3823 | ||
cb900f41 | 3824 | spin_unlock(ptl); |
4c887265 AL |
3825 | unlock_page(page); |
3826 | out: | |
ac9b9c66 | 3827 | return ret; |
4c887265 AL |
3828 | |
3829 | backout: | |
cb900f41 | 3830 | spin_unlock(ptl); |
2b26736c | 3831 | backout_unlocked: |
4c887265 | 3832 | unlock_page(page); |
96b96a96 | 3833 | restore_reserve_on_error(h, vma, address, page); |
4c887265 AL |
3834 | put_page(page); |
3835 | goto out; | |
ac9b9c66 HD |
3836 | } |
3837 | ||
8382d914 | 3838 | #ifdef CONFIG_SMP |
c672c7f2 | 3839 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3840 | struct vm_area_struct *vma, |
3841 | struct address_space *mapping, | |
3842 | pgoff_t idx, unsigned long address) | |
3843 | { | |
3844 | unsigned long key[2]; | |
3845 | u32 hash; | |
3846 | ||
3847 | if (vma->vm_flags & VM_SHARED) { | |
3848 | key[0] = (unsigned long) mapping; | |
3849 | key[1] = idx; | |
3850 | } else { | |
3851 | key[0] = (unsigned long) mm; | |
3852 | key[1] = address >> huge_page_shift(h); | |
3853 | } | |
3854 | ||
3855 | hash = jhash2((u32 *)&key, sizeof(key)/sizeof(u32), 0); | |
3856 | ||
3857 | return hash & (num_fault_mutexes - 1); | |
3858 | } | |
3859 | #else | |
3860 | /* | |
3861 | * For uniprocesor systems we always use a single mutex, so just | |
3862 | * return 0 and avoid the hashing overhead. | |
3863 | */ | |
c672c7f2 | 3864 | u32 hugetlb_fault_mutex_hash(struct hstate *h, struct mm_struct *mm, |
8382d914 DB |
3865 | struct vm_area_struct *vma, |
3866 | struct address_space *mapping, | |
3867 | pgoff_t idx, unsigned long address) | |
3868 | { | |
3869 | return 0; | |
3870 | } | |
3871 | #endif | |
3872 | ||
86e5216f | 3873 | int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 3874 | unsigned long address, unsigned int flags) |
86e5216f | 3875 | { |
8382d914 | 3876 | pte_t *ptep, entry; |
cb900f41 | 3877 | spinlock_t *ptl; |
1e8f889b | 3878 | int ret; |
8382d914 DB |
3879 | u32 hash; |
3880 | pgoff_t idx; | |
0fe6e20b | 3881 | struct page *page = NULL; |
57303d80 | 3882 | struct page *pagecache_page = NULL; |
a5516438 | 3883 | struct hstate *h = hstate_vma(vma); |
8382d914 | 3884 | struct address_space *mapping; |
0f792cf9 | 3885 | int need_wait_lock = 0; |
86e5216f | 3886 | |
1e16a539 KH |
3887 | address &= huge_page_mask(h); |
3888 | ||
7868a208 | 3889 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
fd6a03ed NH |
3890 | if (ptep) { |
3891 | entry = huge_ptep_get(ptep); | |
290408d4 | 3892 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 3893 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
3894 | return 0; |
3895 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 3896 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 3897 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
0d777df5 NH |
3898 | } else { |
3899 | ptep = huge_pte_alloc(mm, address, huge_page_size(h)); | |
3900 | if (!ptep) | |
3901 | return VM_FAULT_OOM; | |
fd6a03ed NH |
3902 | } |
3903 | ||
8382d914 DB |
3904 | mapping = vma->vm_file->f_mapping; |
3905 | idx = vma_hugecache_offset(h, vma, address); | |
3906 | ||
3935baa9 DG |
3907 | /* |
3908 | * Serialize hugepage allocation and instantiation, so that we don't | |
3909 | * get spurious allocation failures if two CPUs race to instantiate | |
3910 | * the same page in the page cache. | |
3911 | */ | |
c672c7f2 MK |
3912 | hash = hugetlb_fault_mutex_hash(h, mm, vma, mapping, idx, address); |
3913 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
8382d914 | 3914 | |
7f2e9525 GS |
3915 | entry = huge_ptep_get(ptep); |
3916 | if (huge_pte_none(entry)) { | |
8382d914 | 3917 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 3918 | goto out_mutex; |
3935baa9 | 3919 | } |
86e5216f | 3920 | |
83c54070 | 3921 | ret = 0; |
1e8f889b | 3922 | |
0f792cf9 NH |
3923 | /* |
3924 | * entry could be a migration/hwpoison entry at this point, so this | |
3925 | * check prevents the kernel from going below assuming that we have | |
3926 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
3927 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
3928 | * handle it. | |
3929 | */ | |
3930 | if (!pte_present(entry)) | |
3931 | goto out_mutex; | |
3932 | ||
57303d80 AW |
3933 | /* |
3934 | * If we are going to COW the mapping later, we examine the pending | |
3935 | * reservations for this page now. This will ensure that any | |
3936 | * allocations necessary to record that reservation occur outside the | |
3937 | * spinlock. For private mappings, we also lookup the pagecache | |
3938 | * page now as it is used to determine if a reservation has been | |
3939 | * consumed. | |
3940 | */ | |
106c992a | 3941 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
2b26736c AW |
3942 | if (vma_needs_reservation(h, vma, address) < 0) { |
3943 | ret = VM_FAULT_OOM; | |
b4d1d99f | 3944 | goto out_mutex; |
2b26736c | 3945 | } |
5e911373 | 3946 | /* Just decrements count, does not deallocate */ |
feba16e2 | 3947 | vma_end_reservation(h, vma, address); |
57303d80 | 3948 | |
f83a275d | 3949 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 AW |
3950 | pagecache_page = hugetlbfs_pagecache_page(h, |
3951 | vma, address); | |
3952 | } | |
3953 | ||
0f792cf9 NH |
3954 | ptl = huge_pte_lock(h, mm, ptep); |
3955 | ||
3956 | /* Check for a racing update before calling hugetlb_cow */ | |
3957 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
3958 | goto out_ptl; | |
3959 | ||
56c9cfb1 NH |
3960 | /* |
3961 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
3962 | * pagecache_page, so here we need take the former one | |
3963 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
3964 | */ |
3965 | page = pte_page(entry); | |
3966 | if (page != pagecache_page) | |
0f792cf9 NH |
3967 | if (!trylock_page(page)) { |
3968 | need_wait_lock = 1; | |
3969 | goto out_ptl; | |
3970 | } | |
b4d1d99f | 3971 | |
0f792cf9 | 3972 | get_page(page); |
b4d1d99f | 3973 | |
788c7df4 | 3974 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 3975 | if (!huge_pte_write(entry)) { |
3999f52e AK |
3976 | ret = hugetlb_cow(mm, vma, address, ptep, |
3977 | pagecache_page, ptl); | |
0f792cf9 | 3978 | goto out_put_page; |
b4d1d99f | 3979 | } |
106c992a | 3980 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
3981 | } |
3982 | entry = pte_mkyoung(entry); | |
788c7df4 HD |
3983 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, |
3984 | flags & FAULT_FLAG_WRITE)) | |
4b3073e1 | 3985 | update_mmu_cache(vma, address, ptep); |
0f792cf9 NH |
3986 | out_put_page: |
3987 | if (page != pagecache_page) | |
3988 | unlock_page(page); | |
3989 | put_page(page); | |
cb900f41 KS |
3990 | out_ptl: |
3991 | spin_unlock(ptl); | |
57303d80 AW |
3992 | |
3993 | if (pagecache_page) { | |
3994 | unlock_page(pagecache_page); | |
3995 | put_page(pagecache_page); | |
3996 | } | |
b4d1d99f | 3997 | out_mutex: |
c672c7f2 | 3998 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
3999 | /* |
4000 | * Generally it's safe to hold refcount during waiting page lock. But | |
4001 | * here we just wait to defer the next page fault to avoid busy loop and | |
4002 | * the page is not used after unlocked before returning from the current | |
4003 | * page fault. So we are safe from accessing freed page, even if we wait | |
4004 | * here without taking refcount. | |
4005 | */ | |
4006 | if (need_wait_lock) | |
4007 | wait_on_page_locked(page); | |
1e8f889b | 4008 | return ret; |
86e5216f AL |
4009 | } |
4010 | ||
8fb5debc MK |
4011 | /* |
4012 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4013 | * modifications for huge pages. | |
4014 | */ | |
4015 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4016 | pte_t *dst_pte, | |
4017 | struct vm_area_struct *dst_vma, | |
4018 | unsigned long dst_addr, | |
4019 | unsigned long src_addr, | |
4020 | struct page **pagep) | |
4021 | { | |
1e392147 AA |
4022 | struct address_space *mapping; |
4023 | pgoff_t idx; | |
4024 | unsigned long size; | |
1c9e8def | 4025 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4026 | struct hstate *h = hstate_vma(dst_vma); |
4027 | pte_t _dst_pte; | |
4028 | spinlock_t *ptl; | |
4029 | int ret; | |
4030 | struct page *page; | |
4031 | ||
4032 | if (!*pagep) { | |
4033 | ret = -ENOMEM; | |
4034 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4035 | if (IS_ERR(page)) | |
4036 | goto out; | |
4037 | ||
4038 | ret = copy_huge_page_from_user(page, | |
4039 | (const void __user *) src_addr, | |
810a56b9 | 4040 | pages_per_huge_page(h), false); |
8fb5debc MK |
4041 | |
4042 | /* fallback to copy_from_user outside mmap_sem */ | |
4043 | if (unlikely(ret)) { | |
4044 | ret = -EFAULT; | |
4045 | *pagep = page; | |
4046 | /* don't free the page */ | |
4047 | goto out; | |
4048 | } | |
4049 | } else { | |
4050 | page = *pagep; | |
4051 | *pagep = NULL; | |
4052 | } | |
4053 | ||
4054 | /* | |
4055 | * The memory barrier inside __SetPageUptodate makes sure that | |
4056 | * preceding stores to the page contents become visible before | |
4057 | * the set_pte_at() write. | |
4058 | */ | |
4059 | __SetPageUptodate(page); | |
4060 | set_page_huge_active(page); | |
4061 | ||
1e392147 AA |
4062 | mapping = dst_vma->vm_file->f_mapping; |
4063 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4064 | ||
1c9e8def MK |
4065 | /* |
4066 | * If shared, add to page cache | |
4067 | */ | |
4068 | if (vm_shared) { | |
1e392147 AA |
4069 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4070 | ret = -EFAULT; | |
4071 | if (idx >= size) | |
4072 | goto out_release_nounlock; | |
1c9e8def | 4073 | |
1e392147 AA |
4074 | /* |
4075 | * Serialization between remove_inode_hugepages() and | |
4076 | * huge_add_to_page_cache() below happens through the | |
4077 | * hugetlb_fault_mutex_table that here must be hold by | |
4078 | * the caller. | |
4079 | */ | |
1c9e8def MK |
4080 | ret = huge_add_to_page_cache(page, mapping, idx); |
4081 | if (ret) | |
4082 | goto out_release_nounlock; | |
4083 | } | |
4084 | ||
8fb5debc MK |
4085 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4086 | spin_lock(ptl); | |
4087 | ||
1e392147 AA |
4088 | /* |
4089 | * Recheck the i_size after holding PT lock to make sure not | |
4090 | * to leave any page mapped (as page_mapped()) beyond the end | |
4091 | * of the i_size (remove_inode_hugepages() is strict about | |
4092 | * enforcing that). If we bail out here, we'll also leave a | |
4093 | * page in the radix tree in the vm_shared case beyond the end | |
4094 | * of the i_size, but remove_inode_hugepages() will take care | |
4095 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4096 | */ | |
4097 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4098 | ret = -EFAULT; | |
4099 | if (idx >= size) | |
4100 | goto out_release_unlock; | |
4101 | ||
8fb5debc MK |
4102 | ret = -EEXIST; |
4103 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4104 | goto out_release_unlock; | |
4105 | ||
1c9e8def MK |
4106 | if (vm_shared) { |
4107 | page_dup_rmap(page, true); | |
4108 | } else { | |
4109 | ClearPagePrivate(page); | |
4110 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4111 | } | |
8fb5debc MK |
4112 | |
4113 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4114 | if (dst_vma->vm_flags & VM_WRITE) | |
4115 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4116 | _dst_pte = pte_mkyoung(_dst_pte); | |
4117 | ||
4118 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4119 | ||
4120 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4121 | dst_vma->vm_flags & VM_WRITE); | |
4122 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4123 | ||
4124 | /* No need to invalidate - it was non-present before */ | |
4125 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4126 | ||
4127 | spin_unlock(ptl); | |
1c9e8def MK |
4128 | if (vm_shared) |
4129 | unlock_page(page); | |
8fb5debc MK |
4130 | ret = 0; |
4131 | out: | |
4132 | return ret; | |
4133 | out_release_unlock: | |
4134 | spin_unlock(ptl); | |
1c9e8def MK |
4135 | if (vm_shared) |
4136 | unlock_page(page); | |
5af10dfd | 4137 | out_release_nounlock: |
8fb5debc MK |
4138 | put_page(page); |
4139 | goto out; | |
4140 | } | |
4141 | ||
28a35716 ML |
4142 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4143 | struct page **pages, struct vm_area_struct **vmas, | |
4144 | unsigned long *position, unsigned long *nr_pages, | |
87ffc118 | 4145 | long i, unsigned int flags, int *nonblocking) |
63551ae0 | 4146 | { |
d5d4b0aa CK |
4147 | unsigned long pfn_offset; |
4148 | unsigned long vaddr = *position; | |
28a35716 | 4149 | unsigned long remainder = *nr_pages; |
a5516438 | 4150 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4151 | int err = -EFAULT; |
63551ae0 | 4152 | |
63551ae0 | 4153 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4154 | pte_t *pte; |
cb900f41 | 4155 | spinlock_t *ptl = NULL; |
2a15efc9 | 4156 | int absent; |
4c887265 | 4157 | struct page *page; |
63551ae0 | 4158 | |
02057967 DR |
4159 | /* |
4160 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4161 | * potentially allocating memory. | |
4162 | */ | |
4163 | if (unlikely(fatal_signal_pending(current))) { | |
4164 | remainder = 0; | |
4165 | break; | |
4166 | } | |
4167 | ||
4c887265 AL |
4168 | /* |
4169 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4170 | * each hugepage. We have to make sure we get the |
4c887265 | 4171 | * first, for the page indexing below to work. |
cb900f41 KS |
4172 | * |
4173 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4174 | */ |
7868a208 PA |
4175 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4176 | huge_page_size(h)); | |
cb900f41 KS |
4177 | if (pte) |
4178 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4179 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4180 | ||
4181 | /* | |
4182 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4183 | * an error where there's an empty slot with no huge pagecache |
4184 | * to back it. This way, we avoid allocating a hugepage, and | |
4185 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4186 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4187 | */ |
3ae77f43 HD |
4188 | if (absent && (flags & FOLL_DUMP) && |
4189 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4190 | if (pte) |
4191 | spin_unlock(ptl); | |
2a15efc9 HD |
4192 | remainder = 0; |
4193 | break; | |
4194 | } | |
63551ae0 | 4195 | |
9cc3a5bd NH |
4196 | /* |
4197 | * We need call hugetlb_fault for both hugepages under migration | |
4198 | * (in which case hugetlb_fault waits for the migration,) and | |
4199 | * hwpoisoned hugepages (in which case we need to prevent the | |
4200 | * caller from accessing to them.) In order to do this, we use | |
4201 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4202 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4203 | * both cases, and because we can't follow correct pages | |
4204 | * directly from any kind of swap entries. | |
4205 | */ | |
4206 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4207 | ((flags & FOLL_WRITE) && |
4208 | !huge_pte_write(huge_ptep_get(pte)))) { | |
4c887265 | 4209 | int ret; |
87ffc118 | 4210 | unsigned int fault_flags = 0; |
63551ae0 | 4211 | |
cb900f41 KS |
4212 | if (pte) |
4213 | spin_unlock(ptl); | |
87ffc118 AA |
4214 | if (flags & FOLL_WRITE) |
4215 | fault_flags |= FAULT_FLAG_WRITE; | |
4216 | if (nonblocking) | |
4217 | fault_flags |= FAULT_FLAG_ALLOW_RETRY; | |
4218 | if (flags & FOLL_NOWAIT) | |
4219 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4220 | FAULT_FLAG_RETRY_NOWAIT; | |
4221 | if (flags & FOLL_TRIED) { | |
4222 | VM_WARN_ON_ONCE(fault_flags & | |
4223 | FAULT_FLAG_ALLOW_RETRY); | |
4224 | fault_flags |= FAULT_FLAG_TRIED; | |
4225 | } | |
4226 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4227 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4228 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4229 | remainder = 0; |
4230 | break; | |
4231 | } | |
4232 | if (ret & VM_FAULT_RETRY) { | |
4233 | if (nonblocking) | |
4234 | *nonblocking = 0; | |
4235 | *nr_pages = 0; | |
4236 | /* | |
4237 | * VM_FAULT_RETRY must not return an | |
4238 | * error, it will return zero | |
4239 | * instead. | |
4240 | * | |
4241 | * No need to update "position" as the | |
4242 | * caller will not check it after | |
4243 | * *nr_pages is set to 0. | |
4244 | */ | |
4245 | return i; | |
4246 | } | |
4247 | continue; | |
4c887265 AL |
4248 | } |
4249 | ||
a5516438 | 4250 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4251 | page = pte_page(huge_ptep_get(pte)); |
d5d4b0aa | 4252 | same_page: |
d6692183 | 4253 | if (pages) { |
2a15efc9 | 4254 | pages[i] = mem_map_offset(page, pfn_offset); |
ddc58f27 | 4255 | get_page(pages[i]); |
d6692183 | 4256 | } |
63551ae0 DG |
4257 | |
4258 | if (vmas) | |
4259 | vmas[i] = vma; | |
4260 | ||
4261 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4262 | ++pfn_offset; |
63551ae0 DG |
4263 | --remainder; |
4264 | ++i; | |
d5d4b0aa | 4265 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4266 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
4267 | /* |
4268 | * We use pfn_offset to avoid touching the pageframes | |
4269 | * of this compound page. | |
4270 | */ | |
4271 | goto same_page; | |
4272 | } | |
cb900f41 | 4273 | spin_unlock(ptl); |
63551ae0 | 4274 | } |
28a35716 | 4275 | *nr_pages = remainder; |
87ffc118 AA |
4276 | /* |
4277 | * setting position is actually required only if remainder is | |
4278 | * not zero but it's faster not to add a "if (remainder)" | |
4279 | * branch. | |
4280 | */ | |
63551ae0 DG |
4281 | *position = vaddr; |
4282 | ||
2be7cfed | 4283 | return i ? i : err; |
63551ae0 | 4284 | } |
8f860591 | 4285 | |
5491ae7b AK |
4286 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4287 | /* | |
4288 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4289 | * implement this. | |
4290 | */ | |
4291 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4292 | #endif | |
4293 | ||
7da4d641 | 4294 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4295 | unsigned long address, unsigned long end, pgprot_t newprot) |
4296 | { | |
4297 | struct mm_struct *mm = vma->vm_mm; | |
4298 | unsigned long start = address; | |
4299 | pte_t *ptep; | |
4300 | pte_t pte; | |
a5516438 | 4301 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4302 | unsigned long pages = 0; |
8f860591 ZY |
4303 | |
4304 | BUG_ON(address >= end); | |
4305 | flush_cache_range(vma, address, end); | |
4306 | ||
a5338093 | 4307 | mmu_notifier_invalidate_range_start(mm, start, end); |
83cde9e8 | 4308 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4309 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4310 | spinlock_t *ptl; |
7868a208 | 4311 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4312 | if (!ptep) |
4313 | continue; | |
cb900f41 | 4314 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4315 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4316 | pages++; | |
cb900f41 | 4317 | spin_unlock(ptl); |
39dde65c | 4318 | continue; |
7da4d641 | 4319 | } |
a8bda28d NH |
4320 | pte = huge_ptep_get(ptep); |
4321 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4322 | spin_unlock(ptl); | |
4323 | continue; | |
4324 | } | |
4325 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4326 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4327 | ||
4328 | if (is_write_migration_entry(entry)) { | |
4329 | pte_t newpte; | |
4330 | ||
4331 | make_migration_entry_read(&entry); | |
4332 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4333 | set_huge_swap_pte_at(mm, address, ptep, |
4334 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4335 | pages++; |
4336 | } | |
4337 | spin_unlock(ptl); | |
4338 | continue; | |
4339 | } | |
4340 | if (!huge_pte_none(pte)) { | |
8f860591 | 4341 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
106c992a | 4342 | pte = pte_mkhuge(huge_pte_modify(pte, newprot)); |
be7517d6 | 4343 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
8f860591 | 4344 | set_huge_pte_at(mm, address, ptep, pte); |
7da4d641 | 4345 | pages++; |
8f860591 | 4346 | } |
cb900f41 | 4347 | spin_unlock(ptl); |
8f860591 | 4348 | } |
d833352a | 4349 | /* |
c8c06efa | 4350 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4351 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4352 | * once we release i_mmap_rwsem, another task can do the final put_page |
d833352a MG |
4353 | * and that page table be reused and filled with junk. |
4354 | */ | |
5491ae7b | 4355 | flush_hugetlb_tlb_range(vma, start, end); |
0f10851e JG |
4356 | /* |
4357 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4358 | * page table protection not changing it to point to a new page. | |
4359 | * | |
4360 | * See Documentation/vm/mmu_notifier.txt | |
4361 | */ | |
83cde9e8 | 4362 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
a5338093 | 4363 | mmu_notifier_invalidate_range_end(mm, start, end); |
7da4d641 PZ |
4364 | |
4365 | return pages << h->order; | |
8f860591 ZY |
4366 | } |
4367 | ||
a1e78772 MG |
4368 | int hugetlb_reserve_pages(struct inode *inode, |
4369 | long from, long to, | |
5a6fe125 | 4370 | struct vm_area_struct *vma, |
ca16d140 | 4371 | vm_flags_t vm_flags) |
e4e574b7 | 4372 | { |
17c9d12e | 4373 | long ret, chg; |
a5516438 | 4374 | struct hstate *h = hstate_inode(inode); |
90481622 | 4375 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4376 | struct resv_map *resv_map; |
1c5ecae3 | 4377 | long gbl_reserve; |
e4e574b7 | 4378 | |
63489f8e MK |
4379 | /* This should never happen */ |
4380 | if (from > to) { | |
4381 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
4382 | return -EINVAL; | |
4383 | } | |
4384 | ||
17c9d12e MG |
4385 | /* |
4386 | * Only apply hugepage reservation if asked. At fault time, an | |
4387 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4388 | * without using reserves |
17c9d12e | 4389 | */ |
ca16d140 | 4390 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4391 | return 0; |
4392 | ||
a1e78772 MG |
4393 | /* |
4394 | * Shared mappings base their reservation on the number of pages that | |
4395 | * are already allocated on behalf of the file. Private mappings need | |
4396 | * to reserve the full area even if read-only as mprotect() may be | |
4397 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4398 | */ | |
9119a41e | 4399 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4e35f483 | 4400 | resv_map = inode_resv_map(inode); |
9119a41e | 4401 | |
1406ec9b | 4402 | chg = region_chg(resv_map, from, to); |
9119a41e JK |
4403 | |
4404 | } else { | |
4405 | resv_map = resv_map_alloc(); | |
17c9d12e MG |
4406 | if (!resv_map) |
4407 | return -ENOMEM; | |
4408 | ||
a1e78772 | 4409 | chg = to - from; |
84afd99b | 4410 | |
17c9d12e MG |
4411 | set_vma_resv_map(vma, resv_map); |
4412 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4413 | } | |
4414 | ||
c50ac050 DH |
4415 | if (chg < 0) { |
4416 | ret = chg; | |
4417 | goto out_err; | |
4418 | } | |
8a630112 | 4419 | |
1c5ecae3 MK |
4420 | /* |
4421 | * There must be enough pages in the subpool for the mapping. If | |
4422 | * the subpool has a minimum size, there may be some global | |
4423 | * reservations already in place (gbl_reserve). | |
4424 | */ | |
4425 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
4426 | if (gbl_reserve < 0) { | |
c50ac050 DH |
4427 | ret = -ENOSPC; |
4428 | goto out_err; | |
4429 | } | |
5a6fe125 MG |
4430 | |
4431 | /* | |
17c9d12e | 4432 | * Check enough hugepages are available for the reservation. |
90481622 | 4433 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 4434 | */ |
1c5ecae3 | 4435 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 4436 | if (ret < 0) { |
1c5ecae3 MK |
4437 | /* put back original number of pages, chg */ |
4438 | (void)hugepage_subpool_put_pages(spool, chg); | |
c50ac050 | 4439 | goto out_err; |
68842c9b | 4440 | } |
17c9d12e MG |
4441 | |
4442 | /* | |
4443 | * Account for the reservations made. Shared mappings record regions | |
4444 | * that have reservations as they are shared by multiple VMAs. | |
4445 | * When the last VMA disappears, the region map says how much | |
4446 | * the reservation was and the page cache tells how much of | |
4447 | * the reservation was consumed. Private mappings are per-VMA and | |
4448 | * only the consumed reservations are tracked. When the VMA | |
4449 | * disappears, the original reservation is the VMA size and the | |
4450 | * consumed reservations are stored in the map. Hence, nothing | |
4451 | * else has to be done for private mappings here | |
4452 | */ | |
33039678 MK |
4453 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
4454 | long add = region_add(resv_map, from, to); | |
4455 | ||
4456 | if (unlikely(chg > add)) { | |
4457 | /* | |
4458 | * pages in this range were added to the reserve | |
4459 | * map between region_chg and region_add. This | |
4460 | * indicates a race with alloc_huge_page. Adjust | |
4461 | * the subpool and reserve counts modified above | |
4462 | * based on the difference. | |
4463 | */ | |
4464 | long rsv_adjust; | |
4465 | ||
4466 | rsv_adjust = hugepage_subpool_put_pages(spool, | |
4467 | chg - add); | |
4468 | hugetlb_acct_memory(h, -rsv_adjust); | |
4469 | } | |
4470 | } | |
a43a8c39 | 4471 | return 0; |
c50ac050 | 4472 | out_err: |
5e911373 | 4473 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
ff8c0c53 MK |
4474 | /* Don't call region_abort if region_chg failed */ |
4475 | if (chg >= 0) | |
4476 | region_abort(resv_map, from, to); | |
f031dd27 JK |
4477 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4478 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 4479 | return ret; |
a43a8c39 CK |
4480 | } |
4481 | ||
b5cec28d MK |
4482 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
4483 | long freed) | |
a43a8c39 | 4484 | { |
a5516438 | 4485 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 4486 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 4487 | long chg = 0; |
90481622 | 4488 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 4489 | long gbl_reserve; |
45c682a6 | 4490 | |
b5cec28d MK |
4491 | if (resv_map) { |
4492 | chg = region_del(resv_map, start, end); | |
4493 | /* | |
4494 | * region_del() can fail in the rare case where a region | |
4495 | * must be split and another region descriptor can not be | |
4496 | * allocated. If end == LONG_MAX, it will not fail. | |
4497 | */ | |
4498 | if (chg < 0) | |
4499 | return chg; | |
4500 | } | |
4501 | ||
45c682a6 | 4502 | spin_lock(&inode->i_lock); |
e4c6f8be | 4503 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
4504 | spin_unlock(&inode->i_lock); |
4505 | ||
1c5ecae3 MK |
4506 | /* |
4507 | * If the subpool has a minimum size, the number of global | |
4508 | * reservations to be released may be adjusted. | |
4509 | */ | |
4510 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
4511 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
4512 | |
4513 | return 0; | |
a43a8c39 | 4514 | } |
93f70f90 | 4515 | |
3212b535 SC |
4516 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
4517 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
4518 | struct vm_area_struct *vma, | |
4519 | unsigned long addr, pgoff_t idx) | |
4520 | { | |
4521 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
4522 | svma->vm_start; | |
4523 | unsigned long sbase = saddr & PUD_MASK; | |
4524 | unsigned long s_end = sbase + PUD_SIZE; | |
4525 | ||
4526 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
4527 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
4528 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
4529 | |
4530 | /* | |
4531 | * match the virtual addresses, permission and the alignment of the | |
4532 | * page table page. | |
4533 | */ | |
4534 | if (pmd_index(addr) != pmd_index(saddr) || | |
4535 | vm_flags != svm_flags || | |
4536 | sbase < svma->vm_start || svma->vm_end < s_end) | |
4537 | return 0; | |
4538 | ||
4539 | return saddr; | |
4540 | } | |
4541 | ||
31aafb45 | 4542 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
4543 | { |
4544 | unsigned long base = addr & PUD_MASK; | |
4545 | unsigned long end = base + PUD_SIZE; | |
4546 | ||
4547 | /* | |
4548 | * check on proper vm_flags and page table alignment | |
4549 | */ | |
4550 | if (vma->vm_flags & VM_MAYSHARE && | |
4551 | vma->vm_start <= base && end <= vma->vm_end) | |
31aafb45 NK |
4552 | return true; |
4553 | return false; | |
3212b535 SC |
4554 | } |
4555 | ||
4556 | /* | |
4557 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
4558 | * and returns the corresponding pte. While this is not necessary for the | |
4559 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
4560 | * code much cleaner. pmd allocation is essential for the shared case because | |
c8c06efa | 4561 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise |
3212b535 SC |
4562 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a |
4563 | * bad pmd for sharing. | |
4564 | */ | |
4565 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4566 | { | |
4567 | struct vm_area_struct *vma = find_vma(mm, addr); | |
4568 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4569 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
4570 | vma->vm_pgoff; | |
4571 | struct vm_area_struct *svma; | |
4572 | unsigned long saddr; | |
4573 | pte_t *spte = NULL; | |
4574 | pte_t *pte; | |
cb900f41 | 4575 | spinlock_t *ptl; |
3212b535 SC |
4576 | |
4577 | if (!vma_shareable(vma, addr)) | |
4578 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
4579 | ||
83cde9e8 | 4580 | i_mmap_lock_write(mapping); |
3212b535 SC |
4581 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
4582 | if (svma == vma) | |
4583 | continue; | |
4584 | ||
4585 | saddr = page_table_shareable(svma, vma, addr, idx); | |
4586 | if (saddr) { | |
7868a208 PA |
4587 | spte = huge_pte_offset(svma->vm_mm, saddr, |
4588 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
4589 | if (spte) { |
4590 | get_page(virt_to_page(spte)); | |
4591 | break; | |
4592 | } | |
4593 | } | |
4594 | } | |
4595 | ||
4596 | if (!spte) | |
4597 | goto out; | |
4598 | ||
8bea8052 | 4599 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 4600 | if (pud_none(*pud)) { |
3212b535 SC |
4601 | pud_populate(mm, pud, |
4602 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 4603 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 4604 | } else { |
3212b535 | 4605 | put_page(virt_to_page(spte)); |
dc6c9a35 | 4606 | } |
cb900f41 | 4607 | spin_unlock(ptl); |
3212b535 SC |
4608 | out: |
4609 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
83cde9e8 | 4610 | i_mmap_unlock_write(mapping); |
3212b535 SC |
4611 | return pte; |
4612 | } | |
4613 | ||
4614 | /* | |
4615 | * unmap huge page backed by shared pte. | |
4616 | * | |
4617 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
4618 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
4619 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
4620 | * | |
cb900f41 | 4621 | * called with page table lock held. |
3212b535 SC |
4622 | * |
4623 | * returns: 1 successfully unmapped a shared pte page | |
4624 | * 0 the underlying pte page is not shared, or it is the last user | |
4625 | */ | |
4626 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4627 | { | |
4628 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
4629 | p4d_t *p4d = p4d_offset(pgd, *addr); |
4630 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
4631 | |
4632 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
4633 | if (page_count(virt_to_page(ptep)) == 1) | |
4634 | return 0; | |
4635 | ||
4636 | pud_clear(pud); | |
4637 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 4638 | mm_dec_nr_pmds(mm); |
3212b535 SC |
4639 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
4640 | return 1; | |
4641 | } | |
9e5fc74c SC |
4642 | #define want_pmd_share() (1) |
4643 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
4644 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
4645 | { | |
4646 | return NULL; | |
4647 | } | |
e81f2d22 ZZ |
4648 | |
4649 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
4650 | { | |
4651 | return 0; | |
4652 | } | |
9e5fc74c | 4653 | #define want_pmd_share() (0) |
3212b535 SC |
4654 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
4655 | ||
9e5fc74c SC |
4656 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
4657 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
4658 | unsigned long addr, unsigned long sz) | |
4659 | { | |
4660 | pgd_t *pgd; | |
c2febafc | 4661 | p4d_t *p4d; |
9e5fc74c SC |
4662 | pud_t *pud; |
4663 | pte_t *pte = NULL; | |
4664 | ||
4665 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
4666 | p4d = p4d_alloc(mm, pgd, addr); |
4667 | if (!p4d) | |
4668 | return NULL; | |
c2febafc | 4669 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
4670 | if (pud) { |
4671 | if (sz == PUD_SIZE) { | |
4672 | pte = (pte_t *)pud; | |
4673 | } else { | |
4674 | BUG_ON(sz != PMD_SIZE); | |
4675 | if (want_pmd_share() && pud_none(*pud)) | |
4676 | pte = huge_pmd_share(mm, addr, pud); | |
4677 | else | |
4678 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
4679 | } | |
4680 | } | |
4e666314 | 4681 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
4682 | |
4683 | return pte; | |
4684 | } | |
4685 | ||
9b19df29 PA |
4686 | /* |
4687 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
4688 | * entry at address @addr | |
4689 | * | |
4690 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
4691 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
4692 | * size @sz doesn't match the hugepage size at this level of the page | |
4693 | * table. | |
4694 | */ | |
7868a208 PA |
4695 | pte_t *huge_pte_offset(struct mm_struct *mm, |
4696 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
4697 | { |
4698 | pgd_t *pgd; | |
c2febafc | 4699 | p4d_t *p4d; |
9e5fc74c | 4700 | pud_t *pud; |
c2febafc | 4701 | pmd_t *pmd; |
9e5fc74c SC |
4702 | |
4703 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
4704 | if (!pgd_present(*pgd)) |
4705 | return NULL; | |
4706 | p4d = p4d_offset(pgd, addr); | |
4707 | if (!p4d_present(*p4d)) | |
4708 | return NULL; | |
9b19df29 | 4709 | |
c2febafc | 4710 | pud = pud_offset(p4d, addr); |
9b19df29 | 4711 | if (sz != PUD_SIZE && pud_none(*pud)) |
c2febafc | 4712 | return NULL; |
9b19df29 PA |
4713 | /* hugepage or swap? */ |
4714 | if (pud_huge(*pud) || !pud_present(*pud)) | |
c2febafc | 4715 | return (pte_t *)pud; |
9b19df29 | 4716 | |
c2febafc | 4717 | pmd = pmd_offset(pud, addr); |
9b19df29 PA |
4718 | if (sz != PMD_SIZE && pmd_none(*pmd)) |
4719 | return NULL; | |
4720 | /* hugepage or swap? */ | |
4721 | if (pmd_huge(*pmd) || !pmd_present(*pmd)) | |
4722 | return (pte_t *)pmd; | |
4723 | ||
4724 | return NULL; | |
9e5fc74c SC |
4725 | } |
4726 | ||
61f77eda NH |
4727 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
4728 | ||
4729 | /* | |
4730 | * These functions are overwritable if your architecture needs its own | |
4731 | * behavior. | |
4732 | */ | |
4733 | struct page * __weak | |
4734 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
4735 | int write) | |
4736 | { | |
4737 | return ERR_PTR(-EINVAL); | |
4738 | } | |
4739 | ||
4dc71451 AK |
4740 | struct page * __weak |
4741 | follow_huge_pd(struct vm_area_struct *vma, | |
4742 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
4743 | { | |
4744 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
4745 | return NULL; | |
4746 | } | |
4747 | ||
61f77eda | 4748 | struct page * __weak |
9e5fc74c | 4749 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4750 | pmd_t *pmd, int flags) |
9e5fc74c | 4751 | { |
e66f17ff NH |
4752 | struct page *page = NULL; |
4753 | spinlock_t *ptl; | |
c9d398fa | 4754 | pte_t pte; |
e66f17ff NH |
4755 | retry: |
4756 | ptl = pmd_lockptr(mm, pmd); | |
4757 | spin_lock(ptl); | |
4758 | /* | |
4759 | * make sure that the address range covered by this pmd is not | |
4760 | * unmapped from other threads. | |
4761 | */ | |
4762 | if (!pmd_huge(*pmd)) | |
4763 | goto out; | |
c9d398fa NH |
4764 | pte = huge_ptep_get((pte_t *)pmd); |
4765 | if (pte_present(pte)) { | |
97534127 | 4766 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
e66f17ff NH |
4767 | if (flags & FOLL_GET) |
4768 | get_page(page); | |
4769 | } else { | |
c9d398fa | 4770 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
4771 | spin_unlock(ptl); |
4772 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
4773 | goto retry; | |
4774 | } | |
4775 | /* | |
4776 | * hwpoisoned entry is treated as no_page_table in | |
4777 | * follow_page_mask(). | |
4778 | */ | |
4779 | } | |
4780 | out: | |
4781 | spin_unlock(ptl); | |
9e5fc74c SC |
4782 | return page; |
4783 | } | |
4784 | ||
61f77eda | 4785 | struct page * __weak |
9e5fc74c | 4786 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 4787 | pud_t *pud, int flags) |
9e5fc74c | 4788 | { |
e66f17ff NH |
4789 | if (flags & FOLL_GET) |
4790 | return NULL; | |
9e5fc74c | 4791 | |
e66f17ff | 4792 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
4793 | } |
4794 | ||
faaa5b62 AK |
4795 | struct page * __weak |
4796 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
4797 | { | |
4798 | if (flags & FOLL_GET) | |
4799 | return NULL; | |
4800 | ||
4801 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
4802 | } | |
4803 | ||
31caf665 NH |
4804 | bool isolate_huge_page(struct page *page, struct list_head *list) |
4805 | { | |
bcc54222 NH |
4806 | bool ret = true; |
4807 | ||
309381fe | 4808 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4809 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
4810 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
4811 | ret = false; | |
4812 | goto unlock; | |
4813 | } | |
4814 | clear_page_huge_active(page); | |
31caf665 | 4815 | list_move_tail(&page->lru, list); |
bcc54222 | 4816 | unlock: |
31caf665 | 4817 | spin_unlock(&hugetlb_lock); |
bcc54222 | 4818 | return ret; |
31caf665 NH |
4819 | } |
4820 | ||
4821 | void putback_active_hugepage(struct page *page) | |
4822 | { | |
309381fe | 4823 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 4824 | spin_lock(&hugetlb_lock); |
bcc54222 | 4825 | set_page_huge_active(page); |
31caf665 NH |
4826 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
4827 | spin_unlock(&hugetlb_lock); | |
4828 | put_page(page); | |
4829 | } | |
ab5ac90a MH |
4830 | |
4831 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
4832 | { | |
4833 | struct hstate *h = page_hstate(oldpage); | |
4834 | ||
4835 | hugetlb_cgroup_migrate(oldpage, newpage); | |
4836 | set_page_owner_migrate_reason(newpage, reason); | |
4837 | ||
4838 | /* | |
4839 | * transfer temporary state of the new huge page. This is | |
4840 | * reverse to other transitions because the newpage is going to | |
4841 | * be final while the old one will be freed so it takes over | |
4842 | * the temporary status. | |
4843 | * | |
4844 | * Also note that we have to transfer the per-node surplus state | |
4845 | * here as well otherwise the global surplus count will not match | |
4846 | * the per-node's. | |
4847 | */ | |
4848 | if (PageHugeTemporary(newpage)) { | |
4849 | int old_nid = page_to_nid(oldpage); | |
4850 | int new_nid = page_to_nid(newpage); | |
4851 | ||
4852 | SetPageHugeTemporary(oldpage); | |
4853 | ClearPageHugeTemporary(newpage); | |
4854 | ||
4855 | spin_lock(&hugetlb_lock); | |
4856 | if (h->surplus_huge_pages_node[old_nid]) { | |
4857 | h->surplus_huge_pages_node[old_nid]--; | |
4858 | h->surplus_huge_pages_node[new_nid]++; | |
4859 | } | |
4860 | spin_unlock(&hugetlb_lock); | |
4861 | } | |
4862 | } |