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