Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * Generic hugetlb support. | |
6d49e352 | 4 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 5 | */ |
1da177e4 LT |
6 | #include <linux/list.h> |
7 | #include <linux/init.h> | |
1da177e4 | 8 | #include <linux/mm.h> |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
97ad1087 | 19 | #include <linux/memblock.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
bbe88753 | 22 | #include <linux/sched/mm.h> |
63489f8e | 23 | #include <linux/mmdebug.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
0fe6e20b | 25 | #include <linux/rmap.h> |
c6247f72 | 26 | #include <linux/string_helpers.h> |
fd6a03ed NH |
27 | #include <linux/swap.h> |
28 | #include <linux/swapops.h> | |
8382d914 | 29 | #include <linux/jhash.h> |
98fa15f3 | 30 | #include <linux/numa.h> |
c77c0a8a | 31 | #include <linux/llist.h> |
cf11e85f | 32 | #include <linux/cma.h> |
d6606683 | 33 | |
63551ae0 | 34 | #include <asm/page.h> |
ca15ca40 | 35 | #include <asm/pgalloc.h> |
24669e58 | 36 | #include <asm/tlb.h> |
63551ae0 | 37 | |
24669e58 | 38 | #include <linux/io.h> |
63551ae0 | 39 | #include <linux/hugetlb.h> |
9dd540e2 | 40 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 41 | #include <linux/node.h> |
1a1aad8a | 42 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 43 | #include <linux/page_owner.h> |
7835e98b | 44 | #include "internal.h" |
1da177e4 | 45 | |
c3f38a38 | 46 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
47 | unsigned int default_hstate_idx; |
48 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 49 | |
dbda8fea | 50 | #ifdef CONFIG_CMA |
cf11e85f | 51 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
dbda8fea BS |
52 | #endif |
53 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 54 | |
641844f5 NH |
55 | /* |
56 | * Minimum page order among possible hugepage sizes, set to a proper value | |
57 | * at boot time. | |
58 | */ | |
59 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 60 | |
53ba51d2 JT |
61 | __initdata LIST_HEAD(huge_boot_pages); |
62 | ||
e5ff2159 AK |
63 | /* for command line parsing */ |
64 | static struct hstate * __initdata parsed_hstate; | |
65 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 66 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 67 | static bool __initdata parsed_default_hugepagesz; |
e5ff2159 | 68 | |
3935baa9 | 69 | /* |
31caf665 NH |
70 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
71 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 72 | */ |
c3f38a38 | 73 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 74 | |
8382d914 DB |
75 | /* |
76 | * Serializes faults on the same logical page. This is used to | |
77 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
78 | */ | |
79 | static int num_fault_mutexes; | |
c672c7f2 | 80 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 81 | |
7ca02d0a MK |
82 | /* Forward declaration */ |
83 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
84 | ||
90481622 DG |
85 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
86 | { | |
87 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
88 | ||
89 | spin_unlock(&spool->lock); | |
90 | ||
91 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 92 | * remain, give up any reservations based on minimum size and |
7ca02d0a MK |
93 | * free the subpool */ |
94 | if (free) { | |
95 | if (spool->min_hpages != -1) | |
96 | hugetlb_acct_memory(spool->hstate, | |
97 | -spool->min_hpages); | |
90481622 | 98 | kfree(spool); |
7ca02d0a | 99 | } |
90481622 DG |
100 | } |
101 | ||
7ca02d0a MK |
102 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
103 | long min_hpages) | |
90481622 DG |
104 | { |
105 | struct hugepage_subpool *spool; | |
106 | ||
c6a91820 | 107 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
108 | if (!spool) |
109 | return NULL; | |
110 | ||
111 | spin_lock_init(&spool->lock); | |
112 | spool->count = 1; | |
7ca02d0a MK |
113 | spool->max_hpages = max_hpages; |
114 | spool->hstate = h; | |
115 | spool->min_hpages = min_hpages; | |
116 | ||
117 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
118 | kfree(spool); | |
119 | return NULL; | |
120 | } | |
121 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
122 | |
123 | return spool; | |
124 | } | |
125 | ||
126 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
127 | { | |
128 | spin_lock(&spool->lock); | |
129 | BUG_ON(!spool->count); | |
130 | spool->count--; | |
131 | unlock_or_release_subpool(spool); | |
132 | } | |
133 | ||
1c5ecae3 MK |
134 | /* |
135 | * Subpool accounting for allocating and reserving pages. | |
136 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 137 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
138 | * global pools must be adjusted (upward). The returned value may |
139 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 140 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
141 | */ |
142 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
143 | long delta) |
144 | { | |
1c5ecae3 | 145 | long ret = delta; |
90481622 DG |
146 | |
147 | if (!spool) | |
1c5ecae3 | 148 | return ret; |
90481622 DG |
149 | |
150 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
151 | |
152 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
153 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
154 | spool->used_hpages += delta; | |
155 | else { | |
156 | ret = -ENOMEM; | |
157 | goto unlock_ret; | |
158 | } | |
90481622 | 159 | } |
90481622 | 160 | |
09a95e29 MK |
161 | /* minimum size accounting */ |
162 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
163 | if (delta > spool->rsv_hpages) { |
164 | /* | |
165 | * Asking for more reserves than those already taken on | |
166 | * behalf of subpool. Return difference. | |
167 | */ | |
168 | ret = delta - spool->rsv_hpages; | |
169 | spool->rsv_hpages = 0; | |
170 | } else { | |
171 | ret = 0; /* reserves already accounted for */ | |
172 | spool->rsv_hpages -= delta; | |
173 | } | |
174 | } | |
175 | ||
176 | unlock_ret: | |
177 | spin_unlock(&spool->lock); | |
90481622 DG |
178 | return ret; |
179 | } | |
180 | ||
1c5ecae3 MK |
181 | /* |
182 | * Subpool accounting for freeing and unreserving pages. | |
183 | * Return the number of global page reservations that must be dropped. | |
184 | * The return value may only be different than the passed value (delta) | |
185 | * in the case where a subpool minimum size must be maintained. | |
186 | */ | |
187 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
188 | long delta) |
189 | { | |
1c5ecae3 MK |
190 | long ret = delta; |
191 | ||
90481622 | 192 | if (!spool) |
1c5ecae3 | 193 | return delta; |
90481622 DG |
194 | |
195 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
196 | |
197 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
198 | spool->used_hpages -= delta; | |
199 | ||
09a95e29 MK |
200 | /* minimum size accounting */ |
201 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
202 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
203 | ret = 0; | |
204 | else | |
205 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
206 | ||
207 | spool->rsv_hpages += delta; | |
208 | if (spool->rsv_hpages > spool->min_hpages) | |
209 | spool->rsv_hpages = spool->min_hpages; | |
210 | } | |
211 | ||
212 | /* | |
213 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
214 | * quota reference, free it now. | |
215 | */ | |
90481622 | 216 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
217 | |
218 | return ret; | |
90481622 DG |
219 | } |
220 | ||
221 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
222 | { | |
223 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
224 | } | |
225 | ||
226 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
227 | { | |
496ad9aa | 228 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
229 | } |
230 | ||
0db9d74e MA |
231 | /* Helper that removes a struct file_region from the resv_map cache and returns |
232 | * it for use. | |
233 | */ | |
234 | static struct file_region * | |
235 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
236 | { | |
237 | struct file_region *nrg = NULL; | |
238 | ||
239 | VM_BUG_ON(resv->region_cache_count <= 0); | |
240 | ||
241 | resv->region_cache_count--; | |
242 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
243 | VM_BUG_ON(!nrg); | |
244 | list_del(&nrg->link); | |
245 | ||
246 | nrg->from = from; | |
247 | nrg->to = to; | |
248 | ||
249 | return nrg; | |
250 | } | |
251 | ||
075a61d0 MA |
252 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
253 | struct file_region *rg) | |
254 | { | |
255 | #ifdef CONFIG_CGROUP_HUGETLB | |
256 | nrg->reservation_counter = rg->reservation_counter; | |
257 | nrg->css = rg->css; | |
258 | if (rg->css) | |
259 | css_get(rg->css); | |
260 | #endif | |
261 | } | |
262 | ||
263 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
264 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
265 | struct hstate *h, | |
266 | struct resv_map *resv, | |
267 | struct file_region *nrg) | |
268 | { | |
269 | #ifdef CONFIG_CGROUP_HUGETLB | |
270 | if (h_cg) { | |
271 | nrg->reservation_counter = | |
272 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
273 | nrg->css = &h_cg->css; | |
274 | if (!resv->pages_per_hpage) | |
275 | resv->pages_per_hpage = pages_per_huge_page(h); | |
276 | /* pages_per_hpage should be the same for all entries in | |
277 | * a resv_map. | |
278 | */ | |
279 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
280 | } else { | |
281 | nrg->reservation_counter = NULL; | |
282 | nrg->css = NULL; | |
283 | } | |
284 | #endif | |
285 | } | |
286 | ||
a9b3f867 MA |
287 | static bool has_same_uncharge_info(struct file_region *rg, |
288 | struct file_region *org) | |
289 | { | |
290 | #ifdef CONFIG_CGROUP_HUGETLB | |
291 | return rg && org && | |
292 | rg->reservation_counter == org->reservation_counter && | |
293 | rg->css == org->css; | |
294 | ||
295 | #else | |
296 | return true; | |
297 | #endif | |
298 | } | |
299 | ||
300 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
301 | { | |
302 | struct file_region *nrg = NULL, *prg = NULL; | |
303 | ||
304 | prg = list_prev_entry(rg, link); | |
305 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
306 | has_same_uncharge_info(prg, rg)) { | |
307 | prg->to = rg->to; | |
308 | ||
309 | list_del(&rg->link); | |
310 | kfree(rg); | |
311 | ||
312 | coalesce_file_region(resv, prg); | |
313 | return; | |
314 | } | |
315 | ||
316 | nrg = list_next_entry(rg, link); | |
317 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
318 | has_same_uncharge_info(nrg, rg)) { | |
319 | nrg->from = rg->from; | |
320 | ||
321 | list_del(&rg->link); | |
322 | kfree(rg); | |
323 | ||
324 | coalesce_file_region(resv, nrg); | |
325 | return; | |
326 | } | |
327 | } | |
328 | ||
d75c6af9 MA |
329 | /* Must be called with resv->lock held. Calling this with count_only == true |
330 | * will count the number of pages to be added but will not modify the linked | |
0db9d74e MA |
331 | * list. If regions_needed != NULL and count_only == true, then regions_needed |
332 | * will indicate the number of file_regions needed in the cache to carry out to | |
333 | * add the regions for this range. | |
d75c6af9 MA |
334 | */ |
335 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 MA |
336 | struct hugetlb_cgroup *h_cg, |
337 | struct hstate *h, long *regions_needed, | |
338 | bool count_only) | |
d75c6af9 | 339 | { |
0db9d74e | 340 | long add = 0; |
d75c6af9 | 341 | struct list_head *head = &resv->regions; |
0db9d74e | 342 | long last_accounted_offset = f; |
d75c6af9 MA |
343 | struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; |
344 | ||
0db9d74e MA |
345 | if (regions_needed) |
346 | *regions_needed = 0; | |
d75c6af9 | 347 | |
0db9d74e MA |
348 | /* In this loop, we essentially handle an entry for the range |
349 | * [last_accounted_offset, rg->from), at every iteration, with some | |
350 | * bounds checking. | |
351 | */ | |
352 | list_for_each_entry_safe(rg, trg, head, link) { | |
353 | /* Skip irrelevant regions that start before our range. */ | |
354 | if (rg->from < f) { | |
355 | /* If this region ends after the last accounted offset, | |
356 | * then we need to update last_accounted_offset. | |
357 | */ | |
358 | if (rg->to > last_accounted_offset) | |
359 | last_accounted_offset = rg->to; | |
360 | continue; | |
361 | } | |
d75c6af9 | 362 | |
0db9d74e MA |
363 | /* When we find a region that starts beyond our range, we've |
364 | * finished. | |
365 | */ | |
d75c6af9 MA |
366 | if (rg->from > t) |
367 | break; | |
368 | ||
0db9d74e MA |
369 | /* Add an entry for last_accounted_offset -> rg->from, and |
370 | * update last_accounted_offset. | |
371 | */ | |
372 | if (rg->from > last_accounted_offset) { | |
373 | add += rg->from - last_accounted_offset; | |
374 | if (!count_only) { | |
375 | nrg = get_file_region_entry_from_cache( | |
376 | resv, last_accounted_offset, rg->from); | |
075a61d0 MA |
377 | record_hugetlb_cgroup_uncharge_info(h_cg, h, |
378 | resv, nrg); | |
0db9d74e | 379 | list_add(&nrg->link, rg->link.prev); |
a9b3f867 | 380 | coalesce_file_region(resv, nrg); |
0db9d74e MA |
381 | } else if (regions_needed) |
382 | *regions_needed += 1; | |
383 | } | |
384 | ||
385 | last_accounted_offset = rg->to; | |
386 | } | |
387 | ||
388 | /* Handle the case where our range extends beyond | |
389 | * last_accounted_offset. | |
390 | */ | |
391 | if (last_accounted_offset < t) { | |
392 | add += t - last_accounted_offset; | |
393 | if (!count_only) { | |
394 | nrg = get_file_region_entry_from_cache( | |
395 | resv, last_accounted_offset, t); | |
075a61d0 | 396 | record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); |
0db9d74e | 397 | list_add(&nrg->link, rg->link.prev); |
a9b3f867 | 398 | coalesce_file_region(resv, nrg); |
0db9d74e MA |
399 | } else if (regions_needed) |
400 | *regions_needed += 1; | |
401 | } | |
402 | ||
403 | VM_BUG_ON(add < 0); | |
404 | return add; | |
405 | } | |
406 | ||
407 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
408 | */ | |
409 | static int allocate_file_region_entries(struct resv_map *resv, | |
410 | int regions_needed) | |
411 | __must_hold(&resv->lock) | |
412 | { | |
413 | struct list_head allocated_regions; | |
414 | int to_allocate = 0, i = 0; | |
415 | struct file_region *trg = NULL, *rg = NULL; | |
416 | ||
417 | VM_BUG_ON(regions_needed < 0); | |
418 | ||
419 | INIT_LIST_HEAD(&allocated_regions); | |
420 | ||
421 | /* | |
422 | * Check for sufficient descriptors in the cache to accommodate | |
423 | * the number of in progress add operations plus regions_needed. | |
424 | * | |
425 | * This is a while loop because when we drop the lock, some other call | |
426 | * to region_add or region_del may have consumed some region_entries, | |
427 | * so we keep looping here until we finally have enough entries for | |
428 | * (adds_in_progress + regions_needed). | |
429 | */ | |
430 | while (resv->region_cache_count < | |
431 | (resv->adds_in_progress + regions_needed)) { | |
432 | to_allocate = resv->adds_in_progress + regions_needed - | |
433 | resv->region_cache_count; | |
434 | ||
435 | /* At this point, we should have enough entries in the cache | |
436 | * for all the existings adds_in_progress. We should only be | |
437 | * needing to allocate for regions_needed. | |
d75c6af9 | 438 | */ |
0db9d74e MA |
439 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
440 | ||
441 | spin_unlock(&resv->lock); | |
442 | for (i = 0; i < to_allocate; i++) { | |
443 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
444 | if (!trg) | |
445 | goto out_of_memory; | |
446 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 447 | } |
d75c6af9 | 448 | |
0db9d74e MA |
449 | spin_lock(&resv->lock); |
450 | ||
451 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
d75c6af9 | 452 | list_del(&rg->link); |
0db9d74e MA |
453 | list_add(&rg->link, &resv->region_cache); |
454 | resv->region_cache_count++; | |
d75c6af9 MA |
455 | } |
456 | } | |
457 | ||
0db9d74e | 458 | return 0; |
d75c6af9 | 459 | |
0db9d74e MA |
460 | out_of_memory: |
461 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
462 | list_del(&rg->link); | |
463 | kfree(rg); | |
464 | } | |
465 | return -ENOMEM; | |
d75c6af9 MA |
466 | } |
467 | ||
1dd308a7 MK |
468 | /* |
469 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
470 | * map. Regions will be taken from the cache to fill in this range. |
471 | * Sufficient regions should exist in the cache due to the previous | |
472 | * call to region_chg with the same range, but in some cases the cache will not | |
473 | * have sufficient entries due to races with other code doing region_add or | |
474 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 475 | * |
0db9d74e MA |
476 | * regions_needed is the out value provided by a previous call to region_chg. |
477 | * | |
478 | * Return the number of new huge pages added to the map. This number is greater | |
479 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 480 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
481 | * region_add of regions of length 1 never allocate file_regions and cannot |
482 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
483 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 484 | */ |
0db9d74e | 485 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
486 | long in_regions_needed, struct hstate *h, |
487 | struct hugetlb_cgroup *h_cg) | |
96822904 | 488 | { |
0db9d74e | 489 | long add = 0, actual_regions_needed = 0; |
96822904 | 490 | |
7b24d861 | 491 | spin_lock(&resv->lock); |
0db9d74e MA |
492 | retry: |
493 | ||
494 | /* Count how many regions are actually needed to execute this add. */ | |
075a61d0 MA |
495 | add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed, |
496 | true); | |
96822904 | 497 | |
5e911373 | 498 | /* |
0db9d74e MA |
499 | * Check for sufficient descriptors in the cache to accommodate |
500 | * this add operation. Note that actual_regions_needed may be greater | |
501 | * than in_regions_needed, as the resv_map may have been modified since | |
502 | * the region_chg call. In this case, we need to make sure that we | |
503 | * allocate extra entries, such that we have enough for all the | |
504 | * existing adds_in_progress, plus the excess needed for this | |
505 | * operation. | |
5e911373 | 506 | */ |
0db9d74e MA |
507 | if (actual_regions_needed > in_regions_needed && |
508 | resv->region_cache_count < | |
509 | resv->adds_in_progress + | |
510 | (actual_regions_needed - in_regions_needed)) { | |
511 | /* region_add operation of range 1 should never need to | |
512 | * allocate file_region entries. | |
513 | */ | |
514 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 515 | |
0db9d74e MA |
516 | if (allocate_file_region_entries( |
517 | resv, actual_regions_needed - in_regions_needed)) { | |
518 | return -ENOMEM; | |
519 | } | |
5e911373 | 520 | |
0db9d74e | 521 | goto retry; |
5e911373 MK |
522 | } |
523 | ||
075a61d0 | 524 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false); |
0db9d74e MA |
525 | |
526 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 527 | |
7b24d861 | 528 | spin_unlock(&resv->lock); |
cf3ad20b MK |
529 | VM_BUG_ON(add < 0); |
530 | return add; | |
96822904 AW |
531 | } |
532 | ||
1dd308a7 MK |
533 | /* |
534 | * Examine the existing reserve map and determine how many | |
535 | * huge pages in the specified range [f, t) are NOT currently | |
536 | * represented. This routine is called before a subsequent | |
537 | * call to region_add that will actually modify the reserve | |
538 | * map to add the specified range [f, t). region_chg does | |
539 | * not change the number of huge pages represented by the | |
0db9d74e MA |
540 | * map. A number of new file_region structures is added to the cache as a |
541 | * placeholder, for the subsequent region_add call to use. At least 1 | |
542 | * file_region structure is added. | |
543 | * | |
544 | * out_regions_needed is the number of regions added to the | |
545 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
546 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
547 | * |
548 | * Returns the number of huge pages that need to be added to the existing | |
549 | * reservation map for the range [f, t). This number is greater or equal to | |
550 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
551 | * is needed and can not be allocated. | |
1dd308a7 | 552 | */ |
0db9d74e MA |
553 | static long region_chg(struct resv_map *resv, long f, long t, |
554 | long *out_regions_needed) | |
96822904 | 555 | { |
96822904 AW |
556 | long chg = 0; |
557 | ||
7b24d861 | 558 | spin_lock(&resv->lock); |
5e911373 | 559 | |
0db9d74e | 560 | /* Count how many hugepages in this range are NOT respresented. */ |
075a61d0 MA |
561 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
562 | out_regions_needed, true); | |
5e911373 | 563 | |
0db9d74e MA |
564 | if (*out_regions_needed == 0) |
565 | *out_regions_needed = 1; | |
5e911373 | 566 | |
0db9d74e MA |
567 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
568 | return -ENOMEM; | |
5e911373 | 569 | |
0db9d74e | 570 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 571 | |
7b24d861 | 572 | spin_unlock(&resv->lock); |
96822904 AW |
573 | return chg; |
574 | } | |
575 | ||
5e911373 MK |
576 | /* |
577 | * Abort the in progress add operation. The adds_in_progress field | |
578 | * of the resv_map keeps track of the operations in progress between | |
579 | * calls to region_chg and region_add. Operations are sometimes | |
580 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
581 | * is called to decrement the adds_in_progress counter. regions_needed |
582 | * is the value returned by the region_chg call, it is used to decrement | |
583 | * the adds_in_progress counter. | |
5e911373 MK |
584 | * |
585 | * NOTE: The range arguments [f, t) are not needed or used in this | |
586 | * routine. They are kept to make reading the calling code easier as | |
587 | * arguments will match the associated region_chg call. | |
588 | */ | |
0db9d74e MA |
589 | static void region_abort(struct resv_map *resv, long f, long t, |
590 | long regions_needed) | |
5e911373 MK |
591 | { |
592 | spin_lock(&resv->lock); | |
593 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 594 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
595 | spin_unlock(&resv->lock); |
596 | } | |
597 | ||
1dd308a7 | 598 | /* |
feba16e2 MK |
599 | * Delete the specified range [f, t) from the reserve map. If the |
600 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
601 | * should be deleted. Locate the regions which intersect [f, t) | |
602 | * and either trim, delete or split the existing regions. | |
603 | * | |
604 | * Returns the number of huge pages deleted from the reserve map. | |
605 | * In the normal case, the return value is zero or more. In the | |
606 | * case where a region must be split, a new region descriptor must | |
607 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
608 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
609 | * a region and possibly return -ENOMEM. Callers specifying | |
610 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 611 | */ |
feba16e2 | 612 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 613 | { |
1406ec9b | 614 | struct list_head *head = &resv->regions; |
96822904 | 615 | struct file_region *rg, *trg; |
feba16e2 MK |
616 | struct file_region *nrg = NULL; |
617 | long del = 0; | |
96822904 | 618 | |
feba16e2 | 619 | retry: |
7b24d861 | 620 | spin_lock(&resv->lock); |
feba16e2 | 621 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
622 | /* |
623 | * Skip regions before the range to be deleted. file_region | |
624 | * ranges are normally of the form [from, to). However, there | |
625 | * may be a "placeholder" entry in the map which is of the form | |
626 | * (from, to) with from == to. Check for placeholder entries | |
627 | * at the beginning of the range to be deleted. | |
628 | */ | |
629 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 630 | continue; |
dbe409e4 | 631 | |
feba16e2 | 632 | if (rg->from >= t) |
96822904 | 633 | break; |
96822904 | 634 | |
feba16e2 MK |
635 | if (f > rg->from && t < rg->to) { /* Must split region */ |
636 | /* | |
637 | * Check for an entry in the cache before dropping | |
638 | * lock and attempting allocation. | |
639 | */ | |
640 | if (!nrg && | |
641 | resv->region_cache_count > resv->adds_in_progress) { | |
642 | nrg = list_first_entry(&resv->region_cache, | |
643 | struct file_region, | |
644 | link); | |
645 | list_del(&nrg->link); | |
646 | resv->region_cache_count--; | |
647 | } | |
96822904 | 648 | |
feba16e2 MK |
649 | if (!nrg) { |
650 | spin_unlock(&resv->lock); | |
651 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
652 | if (!nrg) | |
653 | return -ENOMEM; | |
654 | goto retry; | |
655 | } | |
656 | ||
657 | del += t - f; | |
658 | ||
659 | /* New entry for end of split region */ | |
660 | nrg->from = t; | |
661 | nrg->to = rg->to; | |
075a61d0 MA |
662 | |
663 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
664 | ||
feba16e2 MK |
665 | INIT_LIST_HEAD(&nrg->link); |
666 | ||
667 | /* Original entry is trimmed */ | |
668 | rg->to = f; | |
669 | ||
075a61d0 MA |
670 | hugetlb_cgroup_uncharge_file_region( |
671 | resv, rg, nrg->to - nrg->from); | |
672 | ||
feba16e2 MK |
673 | list_add(&nrg->link, &rg->link); |
674 | nrg = NULL; | |
96822904 | 675 | break; |
feba16e2 MK |
676 | } |
677 | ||
678 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
679 | del += rg->to - rg->from; | |
075a61d0 MA |
680 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
681 | rg->to - rg->from); | |
feba16e2 MK |
682 | list_del(&rg->link); |
683 | kfree(rg); | |
684 | continue; | |
685 | } | |
686 | ||
687 | if (f <= rg->from) { /* Trim beginning of region */ | |
688 | del += t - rg->from; | |
689 | rg->from = t; | |
075a61d0 MA |
690 | |
691 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
692 | t - rg->from); | |
feba16e2 MK |
693 | } else { /* Trim end of region */ |
694 | del += rg->to - f; | |
695 | rg->to = f; | |
075a61d0 MA |
696 | |
697 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
698 | rg->to - f); | |
feba16e2 | 699 | } |
96822904 | 700 | } |
7b24d861 | 701 | |
7b24d861 | 702 | spin_unlock(&resv->lock); |
feba16e2 MK |
703 | kfree(nrg); |
704 | return del; | |
96822904 AW |
705 | } |
706 | ||
b5cec28d MK |
707 | /* |
708 | * A rare out of memory error was encountered which prevented removal of | |
709 | * the reserve map region for a page. The huge page itself was free'ed | |
710 | * and removed from the page cache. This routine will adjust the subpool | |
711 | * usage count, and the global reserve count if needed. By incrementing | |
712 | * these counts, the reserve map entry which could not be deleted will | |
713 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
714 | * counts. | |
715 | */ | |
72e2936c | 716 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
717 | { |
718 | struct hugepage_subpool *spool = subpool_inode(inode); | |
719 | long rsv_adjust; | |
720 | ||
721 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 722 | if (rsv_adjust) { |
b5cec28d MK |
723 | struct hstate *h = hstate_inode(inode); |
724 | ||
725 | hugetlb_acct_memory(h, 1); | |
726 | } | |
727 | } | |
728 | ||
1dd308a7 MK |
729 | /* |
730 | * Count and return the number of huge pages in the reserve map | |
731 | * that intersect with the range [f, t). | |
732 | */ | |
1406ec9b | 733 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 734 | { |
1406ec9b | 735 | struct list_head *head = &resv->regions; |
84afd99b AW |
736 | struct file_region *rg; |
737 | long chg = 0; | |
738 | ||
7b24d861 | 739 | spin_lock(&resv->lock); |
84afd99b AW |
740 | /* Locate each segment we overlap with, and count that overlap. */ |
741 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
742 | long seg_from; |
743 | long seg_to; | |
84afd99b AW |
744 | |
745 | if (rg->to <= f) | |
746 | continue; | |
747 | if (rg->from >= t) | |
748 | break; | |
749 | ||
750 | seg_from = max(rg->from, f); | |
751 | seg_to = min(rg->to, t); | |
752 | ||
753 | chg += seg_to - seg_from; | |
754 | } | |
7b24d861 | 755 | spin_unlock(&resv->lock); |
84afd99b AW |
756 | |
757 | return chg; | |
758 | } | |
759 | ||
e7c4b0bf AW |
760 | /* |
761 | * Convert the address within this vma to the page offset within | |
762 | * the mapping, in pagecache page units; huge pages here. | |
763 | */ | |
a5516438 AK |
764 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
765 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 766 | { |
a5516438 AK |
767 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
768 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
769 | } |
770 | ||
0fe6e20b NH |
771 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
772 | unsigned long address) | |
773 | { | |
774 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
775 | } | |
dee41079 | 776 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 777 | |
08fba699 MG |
778 | /* |
779 | * Return the size of the pages allocated when backing a VMA. In the majority | |
780 | * cases this will be same size as used by the page table entries. | |
781 | */ | |
782 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
783 | { | |
05ea8860 DW |
784 | if (vma->vm_ops && vma->vm_ops->pagesize) |
785 | return vma->vm_ops->pagesize(vma); | |
786 | return PAGE_SIZE; | |
08fba699 | 787 | } |
f340ca0f | 788 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 789 | |
3340289d MG |
790 | /* |
791 | * Return the page size being used by the MMU to back a VMA. In the majority | |
792 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
793 | * architectures where it differs, an architecture-specific 'strong' |
794 | * version of this symbol is required. | |
3340289d | 795 | */ |
09135cc5 | 796 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
797 | { |
798 | return vma_kernel_pagesize(vma); | |
799 | } | |
3340289d | 800 | |
84afd99b AW |
801 | /* |
802 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
803 | * bits of the reservation map pointer, which are always clear due to | |
804 | * alignment. | |
805 | */ | |
806 | #define HPAGE_RESV_OWNER (1UL << 0) | |
807 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 808 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 809 | |
a1e78772 MG |
810 | /* |
811 | * These helpers are used to track how many pages are reserved for | |
812 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
813 | * is guaranteed to have their future faults succeed. | |
814 | * | |
815 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
816 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
817 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
818 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
819 | * |
820 | * The private mapping reservation is represented in a subtly different | |
821 | * manner to a shared mapping. A shared mapping has a region map associated | |
822 | * with the underlying file, this region map represents the backing file | |
823 | * pages which have ever had a reservation assigned which this persists even | |
824 | * after the page is instantiated. A private mapping has a region map | |
825 | * associated with the original mmap which is attached to all VMAs which | |
826 | * reference it, this region map represents those offsets which have consumed | |
827 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 828 | */ |
e7c4b0bf AW |
829 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
830 | { | |
831 | return (unsigned long)vma->vm_private_data; | |
832 | } | |
833 | ||
834 | static void set_vma_private_data(struct vm_area_struct *vma, | |
835 | unsigned long value) | |
836 | { | |
837 | vma->vm_private_data = (void *)value; | |
838 | } | |
839 | ||
e9fe92ae MA |
840 | static void |
841 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
842 | struct hugetlb_cgroup *h_cg, | |
843 | struct hstate *h) | |
844 | { | |
845 | #ifdef CONFIG_CGROUP_HUGETLB | |
846 | if (!h_cg || !h) { | |
847 | resv_map->reservation_counter = NULL; | |
848 | resv_map->pages_per_hpage = 0; | |
849 | resv_map->css = NULL; | |
850 | } else { | |
851 | resv_map->reservation_counter = | |
852 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
853 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
854 | resv_map->css = &h_cg->css; | |
855 | } | |
856 | #endif | |
857 | } | |
858 | ||
9119a41e | 859 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
860 | { |
861 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
862 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
863 | ||
864 | if (!resv_map || !rg) { | |
865 | kfree(resv_map); | |
866 | kfree(rg); | |
84afd99b | 867 | return NULL; |
5e911373 | 868 | } |
84afd99b AW |
869 | |
870 | kref_init(&resv_map->refs); | |
7b24d861 | 871 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
872 | INIT_LIST_HEAD(&resv_map->regions); |
873 | ||
5e911373 | 874 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
875 | /* |
876 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
877 | * fields don't do cgroup accounting. On private mappings, these will be | |
878 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
879 | * reservations are to be un-charged from here. | |
880 | */ | |
881 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
882 | |
883 | INIT_LIST_HEAD(&resv_map->region_cache); | |
884 | list_add(&rg->link, &resv_map->region_cache); | |
885 | resv_map->region_cache_count = 1; | |
886 | ||
84afd99b AW |
887 | return resv_map; |
888 | } | |
889 | ||
9119a41e | 890 | void resv_map_release(struct kref *ref) |
84afd99b AW |
891 | { |
892 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
893 | struct list_head *head = &resv_map->region_cache; |
894 | struct file_region *rg, *trg; | |
84afd99b AW |
895 | |
896 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 897 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
898 | |
899 | /* ... and any entries left in the cache */ | |
900 | list_for_each_entry_safe(rg, trg, head, link) { | |
901 | list_del(&rg->link); | |
902 | kfree(rg); | |
903 | } | |
904 | ||
905 | VM_BUG_ON(resv_map->adds_in_progress); | |
906 | ||
84afd99b AW |
907 | kfree(resv_map); |
908 | } | |
909 | ||
4e35f483 JK |
910 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
911 | { | |
f27a5136 MK |
912 | /* |
913 | * At inode evict time, i_mapping may not point to the original | |
914 | * address space within the inode. This original address space | |
915 | * contains the pointer to the resv_map. So, always use the | |
916 | * address space embedded within the inode. | |
917 | * The VERY common case is inode->mapping == &inode->i_data but, | |
918 | * this may not be true for device special inodes. | |
919 | */ | |
920 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
921 | } |
922 | ||
84afd99b | 923 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 924 | { |
81d1b09c | 925 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
926 | if (vma->vm_flags & VM_MAYSHARE) { |
927 | struct address_space *mapping = vma->vm_file->f_mapping; | |
928 | struct inode *inode = mapping->host; | |
929 | ||
930 | return inode_resv_map(inode); | |
931 | ||
932 | } else { | |
84afd99b AW |
933 | return (struct resv_map *)(get_vma_private_data(vma) & |
934 | ~HPAGE_RESV_MASK); | |
4e35f483 | 935 | } |
a1e78772 MG |
936 | } |
937 | ||
84afd99b | 938 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 939 | { |
81d1b09c SL |
940 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
941 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 942 | |
84afd99b AW |
943 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
944 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
945 | } |
946 | ||
947 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
948 | { | |
81d1b09c SL |
949 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
950 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
951 | |
952 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
953 | } |
954 | ||
955 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
956 | { | |
81d1b09c | 957 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
958 | |
959 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
960 | } |
961 | ||
04f2cbe3 | 962 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
963 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
964 | { | |
81d1b09c | 965 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 966 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
967 | vma->vm_private_data = (void *)0; |
968 | } | |
969 | ||
970 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 971 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 972 | { |
af0ed73e JK |
973 | if (vma->vm_flags & VM_NORESERVE) { |
974 | /* | |
975 | * This address is already reserved by other process(chg == 0), | |
976 | * so, we should decrement reserved count. Without decrementing, | |
977 | * reserve count remains after releasing inode, because this | |
978 | * allocated page will go into page cache and is regarded as | |
979 | * coming from reserved pool in releasing step. Currently, we | |
980 | * don't have any other solution to deal with this situation | |
981 | * properly, so add work-around here. | |
982 | */ | |
983 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 984 | return true; |
af0ed73e | 985 | else |
559ec2f8 | 986 | return false; |
af0ed73e | 987 | } |
a63884e9 JK |
988 | |
989 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
990 | if (vma->vm_flags & VM_MAYSHARE) { |
991 | /* | |
992 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
993 | * be a region map for all pages. The only situation where | |
994 | * there is no region map is if a hole was punched via | |
7c8de358 | 995 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
996 | * use. This situation is indicated if chg != 0. |
997 | */ | |
998 | if (chg) | |
999 | return false; | |
1000 | else | |
1001 | return true; | |
1002 | } | |
a63884e9 JK |
1003 | |
1004 | /* | |
1005 | * Only the process that called mmap() has reserves for | |
1006 | * private mappings. | |
1007 | */ | |
67961f9d MK |
1008 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1009 | /* | |
1010 | * Like the shared case above, a hole punch or truncate | |
1011 | * could have been performed on the private mapping. | |
1012 | * Examine the value of chg to determine if reserves | |
1013 | * actually exist or were previously consumed. | |
1014 | * Very Subtle - The value of chg comes from a previous | |
1015 | * call to vma_needs_reserves(). The reserve map for | |
1016 | * private mappings has different (opposite) semantics | |
1017 | * than that of shared mappings. vma_needs_reserves() | |
1018 | * has already taken this difference in semantics into | |
1019 | * account. Therefore, the meaning of chg is the same | |
1020 | * as in the shared case above. Code could easily be | |
1021 | * combined, but keeping it separate draws attention to | |
1022 | * subtle differences. | |
1023 | */ | |
1024 | if (chg) | |
1025 | return false; | |
1026 | else | |
1027 | return true; | |
1028 | } | |
a63884e9 | 1029 | |
559ec2f8 | 1030 | return false; |
a1e78772 MG |
1031 | } |
1032 | ||
a5516438 | 1033 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1034 | { |
1035 | int nid = page_to_nid(page); | |
0edaecfa | 1036 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1037 | h->free_huge_pages++; |
1038 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
1039 | } |
1040 | ||
94310cbc | 1041 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1042 | { |
1043 | struct page *page; | |
bbe88753 JK |
1044 | bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); |
1045 | ||
1046 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { | |
1047 | if (nocma && is_migrate_cma_page(page)) | |
1048 | continue; | |
bf50bab2 | 1049 | |
243abd5b | 1050 | if (!PageHWPoison(page)) |
c8721bbb | 1051 | break; |
bbe88753 JK |
1052 | } |
1053 | ||
c8721bbb NH |
1054 | /* |
1055 | * if 'non-isolated free hugepage' not found on the list, | |
1056 | * the allocation fails. | |
1057 | */ | |
1058 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 1059 | return NULL; |
0edaecfa | 1060 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 1061 | set_page_refcounted(page); |
bf50bab2 NH |
1062 | h->free_huge_pages--; |
1063 | h->free_huge_pages_node[nid]--; | |
1064 | return page; | |
1065 | } | |
1066 | ||
3e59fcb0 MH |
1067 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1068 | nodemask_t *nmask) | |
94310cbc | 1069 | { |
3e59fcb0 MH |
1070 | unsigned int cpuset_mems_cookie; |
1071 | struct zonelist *zonelist; | |
1072 | struct zone *zone; | |
1073 | struct zoneref *z; | |
98fa15f3 | 1074 | int node = NUMA_NO_NODE; |
94310cbc | 1075 | |
3e59fcb0 MH |
1076 | zonelist = node_zonelist(nid, gfp_mask); |
1077 | ||
1078 | retry_cpuset: | |
1079 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1080 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1081 | struct page *page; | |
1082 | ||
1083 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1084 | continue; | |
1085 | /* | |
1086 | * no need to ask again on the same node. Pool is node rather than | |
1087 | * zone aware | |
1088 | */ | |
1089 | if (zone_to_nid(zone) == node) | |
1090 | continue; | |
1091 | node = zone_to_nid(zone); | |
94310cbc | 1092 | |
94310cbc AK |
1093 | page = dequeue_huge_page_node_exact(h, node); |
1094 | if (page) | |
1095 | return page; | |
1096 | } | |
3e59fcb0 MH |
1097 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1098 | goto retry_cpuset; | |
1099 | ||
94310cbc AK |
1100 | return NULL; |
1101 | } | |
1102 | ||
a5516438 AK |
1103 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1104 | struct vm_area_struct *vma, | |
af0ed73e JK |
1105 | unsigned long address, int avoid_reserve, |
1106 | long chg) | |
1da177e4 | 1107 | { |
3e59fcb0 | 1108 | struct page *page; |
480eccf9 | 1109 | struct mempolicy *mpol; |
04ec6264 | 1110 | gfp_t gfp_mask; |
3e59fcb0 | 1111 | nodemask_t *nodemask; |
04ec6264 | 1112 | int nid; |
1da177e4 | 1113 | |
a1e78772 MG |
1114 | /* |
1115 | * A child process with MAP_PRIVATE mappings created by their parent | |
1116 | * have no page reserves. This check ensures that reservations are | |
1117 | * not "stolen". The child may still get SIGKILLed | |
1118 | */ | |
af0ed73e | 1119 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1120 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1121 | goto err; |
a1e78772 | 1122 | |
04f2cbe3 | 1123 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1124 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1125 | goto err; |
04f2cbe3 | 1126 | |
04ec6264 VB |
1127 | gfp_mask = htlb_alloc_mask(h); |
1128 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
1129 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
1130 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
1131 | SetPagePrivate(page); | |
1132 | h->resv_huge_pages--; | |
1da177e4 | 1133 | } |
cc9a6c87 | 1134 | |
52cd3b07 | 1135 | mpol_cond_put(mpol); |
1da177e4 | 1136 | return page; |
cc9a6c87 MG |
1137 | |
1138 | err: | |
cc9a6c87 | 1139 | return NULL; |
1da177e4 LT |
1140 | } |
1141 | ||
1cac6f2c LC |
1142 | /* |
1143 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1144 | * We may have allocated or freed a huge page based on a different | |
1145 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1146 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1147 | * node for alloc or free. | |
1148 | */ | |
1149 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1150 | { | |
0edaf86c | 1151 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1152 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1153 | ||
1154 | return nid; | |
1155 | } | |
1156 | ||
1157 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1158 | { | |
1159 | if (!node_isset(nid, *nodes_allowed)) | |
1160 | nid = next_node_allowed(nid, nodes_allowed); | |
1161 | return nid; | |
1162 | } | |
1163 | ||
1164 | /* | |
1165 | * returns the previously saved node ["this node"] from which to | |
1166 | * allocate a persistent huge page for the pool and advance the | |
1167 | * next node from which to allocate, handling wrap at end of node | |
1168 | * mask. | |
1169 | */ | |
1170 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1171 | nodemask_t *nodes_allowed) | |
1172 | { | |
1173 | int nid; | |
1174 | ||
1175 | VM_BUG_ON(!nodes_allowed); | |
1176 | ||
1177 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1178 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1179 | ||
1180 | return nid; | |
1181 | } | |
1182 | ||
1183 | /* | |
1184 | * helper for free_pool_huge_page() - return the previously saved | |
1185 | * node ["this node"] from which to free a huge page. Advance the | |
1186 | * next node id whether or not we find a free huge page to free so | |
1187 | * that the next attempt to free addresses the next node. | |
1188 | */ | |
1189 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1190 | { | |
1191 | int nid; | |
1192 | ||
1193 | VM_BUG_ON(!nodes_allowed); | |
1194 | ||
1195 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1196 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1197 | ||
1198 | return nid; | |
1199 | } | |
1200 | ||
1201 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1202 | for (nr_nodes = nodes_weight(*mask); \ | |
1203 | nr_nodes > 0 && \ | |
1204 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1205 | nr_nodes--) | |
1206 | ||
1207 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1208 | for (nr_nodes = nodes_weight(*mask); \ | |
1209 | nr_nodes > 0 && \ | |
1210 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1211 | nr_nodes--) | |
1212 | ||
e1073d1e | 1213 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1214 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1215 | unsigned int order) |
944d9fec LC |
1216 | { |
1217 | int i; | |
1218 | int nr_pages = 1 << order; | |
1219 | struct page *p = page + 1; | |
1220 | ||
c8cc708a | 1221 | atomic_set(compound_mapcount_ptr(page), 0); |
47e29d32 JH |
1222 | if (hpage_pincount_available(page)) |
1223 | atomic_set(compound_pincount_ptr(page), 0); | |
1224 | ||
944d9fec | 1225 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1226 | clear_compound_head(p); |
944d9fec | 1227 | set_page_refcounted(p); |
944d9fec LC |
1228 | } |
1229 | ||
1230 | set_compound_order(page, 0); | |
1231 | __ClearPageHead(page); | |
1232 | } | |
1233 | ||
d00181b9 | 1234 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1235 | { |
cf11e85f RG |
1236 | /* |
1237 | * If the page isn't allocated using the cma allocator, | |
1238 | * cma_release() returns false. | |
1239 | */ | |
dbda8fea BS |
1240 | #ifdef CONFIG_CMA |
1241 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1242 | return; |
dbda8fea | 1243 | #endif |
cf11e85f | 1244 | |
944d9fec LC |
1245 | free_contig_range(page_to_pfn(page), 1 << order); |
1246 | } | |
1247 | ||
4eb0716e | 1248 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1249 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1250 | int nid, nodemask_t *nodemask) | |
944d9fec | 1251 | { |
5e27a2df | 1252 | unsigned long nr_pages = 1UL << huge_page_order(h); |
944d9fec | 1253 | |
dbda8fea BS |
1254 | #ifdef CONFIG_CMA |
1255 | { | |
cf11e85f RG |
1256 | struct page *page; |
1257 | int node; | |
1258 | ||
1259 | for_each_node_mask(node, *nodemask) { | |
1260 | if (!hugetlb_cma[node]) | |
1261 | continue; | |
1262 | ||
1263 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1264 | huge_page_order(h), true); | |
1265 | if (page) | |
1266 | return page; | |
1267 | } | |
1268 | } | |
dbda8fea | 1269 | #endif |
cf11e85f | 1270 | |
5e27a2df | 1271 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1272 | } |
1273 | ||
1274 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1275 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1276 | #else /* !CONFIG_CONTIG_ALLOC */ |
1277 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1278 | int nid, nodemask_t *nodemask) | |
1279 | { | |
1280 | return NULL; | |
1281 | } | |
1282 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1283 | |
e1073d1e | 1284 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1285 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1286 | int nid, nodemask_t *nodemask) |
1287 | { | |
1288 | return NULL; | |
1289 | } | |
d00181b9 | 1290 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1291 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1292 | unsigned int order) { } |
944d9fec LC |
1293 | #endif |
1294 | ||
a5516438 | 1295 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1296 | { |
1297 | int i; | |
a5516438 | 1298 | |
4eb0716e | 1299 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1300 | return; |
18229df5 | 1301 | |
a5516438 AK |
1302 | h->nr_huge_pages--; |
1303 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1304 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1305 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1306 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1307 | 1 << PG_active | 1 << PG_private | |
1308 | 1 << PG_writeback); | |
6af2acb6 | 1309 | } |
309381fe | 1310 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
1adc4d41 | 1311 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); |
f1e61557 | 1312 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1313 | set_page_refcounted(page); |
944d9fec | 1314 | if (hstate_is_gigantic(h)) { |
cf11e85f RG |
1315 | /* |
1316 | * Temporarily drop the hugetlb_lock, because | |
1317 | * we might block in free_gigantic_page(). | |
1318 | */ | |
1319 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1320 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1321 | free_gigantic_page(page, huge_page_order(h)); | |
cf11e85f | 1322 | spin_lock(&hugetlb_lock); |
944d9fec | 1323 | } else { |
944d9fec LC |
1324 | __free_pages(page, huge_page_order(h)); |
1325 | } | |
6af2acb6 AL |
1326 | } |
1327 | ||
e5ff2159 AK |
1328 | struct hstate *size_to_hstate(unsigned long size) |
1329 | { | |
1330 | struct hstate *h; | |
1331 | ||
1332 | for_each_hstate(h) { | |
1333 | if (huge_page_size(h) == size) | |
1334 | return h; | |
1335 | } | |
1336 | return NULL; | |
1337 | } | |
1338 | ||
bcc54222 NH |
1339 | /* |
1340 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1341 | * to hstate->hugepage_activelist.) | |
1342 | * | |
1343 | * This function can be called for tail pages, but never returns true for them. | |
1344 | */ | |
1345 | bool page_huge_active(struct page *page) | |
1346 | { | |
1347 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1348 | return PageHead(page) && PagePrivate(&page[1]); | |
1349 | } | |
1350 | ||
1351 | /* never called for tail page */ | |
1352 | static void set_page_huge_active(struct page *page) | |
1353 | { | |
1354 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1355 | SetPagePrivate(&page[1]); | |
1356 | } | |
1357 | ||
1358 | static void clear_page_huge_active(struct page *page) | |
1359 | { | |
1360 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1361 | ClearPagePrivate(&page[1]); | |
1362 | } | |
1363 | ||
ab5ac90a MH |
1364 | /* |
1365 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1366 | * code | |
1367 | */ | |
1368 | static inline bool PageHugeTemporary(struct page *page) | |
1369 | { | |
1370 | if (!PageHuge(page)) | |
1371 | return false; | |
1372 | ||
1373 | return (unsigned long)page[2].mapping == -1U; | |
1374 | } | |
1375 | ||
1376 | static inline void SetPageHugeTemporary(struct page *page) | |
1377 | { | |
1378 | page[2].mapping = (void *)-1U; | |
1379 | } | |
1380 | ||
1381 | static inline void ClearPageHugeTemporary(struct page *page) | |
1382 | { | |
1383 | page[2].mapping = NULL; | |
1384 | } | |
1385 | ||
c77c0a8a | 1386 | static void __free_huge_page(struct page *page) |
27a85ef1 | 1387 | { |
a5516438 AK |
1388 | /* |
1389 | * Can't pass hstate in here because it is called from the | |
1390 | * compound page destructor. | |
1391 | */ | |
e5ff2159 | 1392 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1393 | int nid = page_to_nid(page); |
90481622 DG |
1394 | struct hugepage_subpool *spool = |
1395 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1396 | bool restore_reserve; |
27a85ef1 | 1397 | |
b4330afb MK |
1398 | VM_BUG_ON_PAGE(page_count(page), page); |
1399 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc YW |
1400 | |
1401 | set_page_private(page, 0); | |
1402 | page->mapping = NULL; | |
07443a85 | 1403 | restore_reserve = PagePrivate(page); |
16c794b4 | 1404 | ClearPagePrivate(page); |
27a85ef1 | 1405 | |
1c5ecae3 | 1406 | /* |
0919e1b6 MK |
1407 | * If PagePrivate() was set on page, page allocation consumed a |
1408 | * reservation. If the page was associated with a subpool, there | |
1409 | * would have been a page reserved in the subpool before allocation | |
1410 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1411 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1412 | * remove the reserved page from the subpool. | |
1c5ecae3 | 1413 | */ |
0919e1b6 MK |
1414 | if (!restore_reserve) { |
1415 | /* | |
1416 | * A return code of zero implies that the subpool will be | |
1417 | * under its minimum size if the reservation is not restored | |
1418 | * after page is free. Therefore, force restore_reserve | |
1419 | * operation. | |
1420 | */ | |
1421 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1422 | restore_reserve = true; | |
1423 | } | |
1c5ecae3 | 1424 | |
27a85ef1 | 1425 | spin_lock(&hugetlb_lock); |
bcc54222 | 1426 | clear_page_huge_active(page); |
6d76dcf4 AK |
1427 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1428 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1429 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1430 | pages_per_huge_page(h), page); | |
07443a85 JK |
1431 | if (restore_reserve) |
1432 | h->resv_huge_pages++; | |
1433 | ||
ab5ac90a MH |
1434 | if (PageHugeTemporary(page)) { |
1435 | list_del(&page->lru); | |
1436 | ClearPageHugeTemporary(page); | |
1437 | update_and_free_page(h, page); | |
1438 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1439 | /* remove the page from active list */ |
1440 | list_del(&page->lru); | |
a5516438 AK |
1441 | update_and_free_page(h, page); |
1442 | h->surplus_huge_pages--; | |
1443 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1444 | } else { |
5d3a551c | 1445 | arch_clear_hugepage_flags(page); |
a5516438 | 1446 | enqueue_huge_page(h, page); |
7893d1d5 | 1447 | } |
27a85ef1 DG |
1448 | spin_unlock(&hugetlb_lock); |
1449 | } | |
1450 | ||
c77c0a8a WL |
1451 | /* |
1452 | * As free_huge_page() can be called from a non-task context, we have | |
1453 | * to defer the actual freeing in a workqueue to prevent potential | |
1454 | * hugetlb_lock deadlock. | |
1455 | * | |
1456 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1457 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1458 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1459 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1460 | */ | |
1461 | static LLIST_HEAD(hpage_freelist); | |
1462 | ||
1463 | static void free_hpage_workfn(struct work_struct *work) | |
1464 | { | |
1465 | struct llist_node *node; | |
1466 | struct page *page; | |
1467 | ||
1468 | node = llist_del_all(&hpage_freelist); | |
1469 | ||
1470 | while (node) { | |
1471 | page = container_of((struct address_space **)node, | |
1472 | struct page, mapping); | |
1473 | node = node->next; | |
1474 | __free_huge_page(page); | |
1475 | } | |
1476 | } | |
1477 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1478 | ||
1479 | void free_huge_page(struct page *page) | |
1480 | { | |
1481 | /* | |
1482 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1483 | */ | |
1484 | if (!in_task()) { | |
1485 | /* | |
1486 | * Only call schedule_work() if hpage_freelist is previously | |
1487 | * empty. Otherwise, schedule_work() had been called but the | |
1488 | * workfn hasn't retrieved the list yet. | |
1489 | */ | |
1490 | if (llist_add((struct llist_node *)&page->mapping, | |
1491 | &hpage_freelist)) | |
1492 | schedule_work(&free_hpage_work); | |
1493 | return; | |
1494 | } | |
1495 | ||
1496 | __free_huge_page(page); | |
1497 | } | |
1498 | ||
a5516438 | 1499 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1500 | { |
0edaecfa | 1501 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1502 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1503 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1504 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1505 | set_hugetlb_cgroup_rsvd(page, NULL); |
a5516438 AK |
1506 | h->nr_huge_pages++; |
1507 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 | 1508 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1509 | } |
1510 | ||
d00181b9 | 1511 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1512 | { |
1513 | int i; | |
1514 | int nr_pages = 1 << order; | |
1515 | struct page *p = page + 1; | |
1516 | ||
1517 | /* we rely on prep_new_huge_page to set the destructor */ | |
1518 | set_compound_order(page, order); | |
ef5a22be | 1519 | __ClearPageReserved(page); |
de09d31d | 1520 | __SetPageHead(page); |
20a0307c | 1521 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1522 | /* |
1523 | * For gigantic hugepages allocated through bootmem at | |
1524 | * boot, it's safer to be consistent with the not-gigantic | |
1525 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1526 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1527 | * pages may get the reference counting wrong if they see |
1528 | * PG_reserved set on a tail page (despite the head page not | |
1529 | * having PG_reserved set). Enforcing this consistency between | |
1530 | * head and tail pages allows drivers to optimize away a check | |
1531 | * on the head page when they need know if put_page() is needed | |
1532 | * after get_user_pages(). | |
1533 | */ | |
1534 | __ClearPageReserved(p); | |
58a84aa9 | 1535 | set_page_count(p, 0); |
1d798ca3 | 1536 | set_compound_head(p, page); |
20a0307c | 1537 | } |
b4330afb | 1538 | atomic_set(compound_mapcount_ptr(page), -1); |
47e29d32 JH |
1539 | |
1540 | if (hpage_pincount_available(page)) | |
1541 | atomic_set(compound_pincount_ptr(page), 0); | |
20a0307c WF |
1542 | } |
1543 | ||
7795912c AM |
1544 | /* |
1545 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1546 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1547 | * details. | |
1548 | */ | |
20a0307c WF |
1549 | int PageHuge(struct page *page) |
1550 | { | |
20a0307c WF |
1551 | if (!PageCompound(page)) |
1552 | return 0; | |
1553 | ||
1554 | page = compound_head(page); | |
f1e61557 | 1555 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1556 | } |
43131e14 NH |
1557 | EXPORT_SYMBOL_GPL(PageHuge); |
1558 | ||
27c73ae7 AA |
1559 | /* |
1560 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1561 | * normal or transparent huge pages. | |
1562 | */ | |
1563 | int PageHeadHuge(struct page *page_head) | |
1564 | { | |
27c73ae7 AA |
1565 | if (!PageHead(page_head)) |
1566 | return 0; | |
1567 | ||
d4af73e3 | 1568 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1569 | } |
27c73ae7 | 1570 | |
c0d0381a MK |
1571 | /* |
1572 | * Find address_space associated with hugetlbfs page. | |
1573 | * Upon entry page is locked and page 'was' mapped although mapped state | |
1574 | * could change. If necessary, use anon_vma to find vma and associated | |
1575 | * address space. The returned mapping may be stale, but it can not be | |
1576 | * invalid as page lock (which is held) is required to destroy mapping. | |
1577 | */ | |
1578 | static struct address_space *_get_hugetlb_page_mapping(struct page *hpage) | |
1579 | { | |
1580 | struct anon_vma *anon_vma; | |
1581 | pgoff_t pgoff_start, pgoff_end; | |
1582 | struct anon_vma_chain *avc; | |
1583 | struct address_space *mapping = page_mapping(hpage); | |
1584 | ||
1585 | /* Simple file based mapping */ | |
1586 | if (mapping) | |
1587 | return mapping; | |
1588 | ||
1589 | /* | |
1590 | * Even anonymous hugetlbfs mappings are associated with an | |
1591 | * underlying hugetlbfs file (see hugetlb_file_setup in mmap | |
1592 | * code). Find a vma associated with the anonymous vma, and | |
1593 | * use the file pointer to get address_space. | |
1594 | */ | |
1595 | anon_vma = page_lock_anon_vma_read(hpage); | |
1596 | if (!anon_vma) | |
1597 | return mapping; /* NULL */ | |
1598 | ||
1599 | /* Use first found vma */ | |
1600 | pgoff_start = page_to_pgoff(hpage); | |
1139d336 | 1601 | pgoff_end = pgoff_start + pages_per_huge_page(page_hstate(hpage)) - 1; |
c0d0381a MK |
1602 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, |
1603 | pgoff_start, pgoff_end) { | |
1604 | struct vm_area_struct *vma = avc->vma; | |
1605 | ||
1606 | mapping = vma->vm_file->f_mapping; | |
1607 | break; | |
1608 | } | |
1609 | ||
1610 | anon_vma_unlock_read(anon_vma); | |
1611 | return mapping; | |
1612 | } | |
1613 | ||
1614 | /* | |
1615 | * Find and lock address space (mapping) in write mode. | |
1616 | * | |
1617 | * Upon entry, the page is locked which allows us to find the mapping | |
1618 | * even in the case of an anon page. However, locking order dictates | |
1619 | * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs | |
1620 | * specific. So, we first try to lock the sema while still holding the | |
1621 | * page lock. If this works, great! If not, then we need to drop the | |
1622 | * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of | |
1623 | * course, need to revalidate state along the way. | |
1624 | */ | |
1625 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1626 | { | |
1627 | struct address_space *mapping, *mapping2; | |
1628 | ||
1629 | mapping = _get_hugetlb_page_mapping(hpage); | |
1630 | retry: | |
1631 | if (!mapping) | |
1632 | return mapping; | |
1633 | ||
1634 | /* | |
1635 | * If no contention, take lock and return | |
1636 | */ | |
1637 | if (i_mmap_trylock_write(mapping)) | |
1638 | return mapping; | |
1639 | ||
1640 | /* | |
1641 | * Must drop page lock and wait on mapping sema. | |
1642 | * Note: Once page lock is dropped, mapping could become invalid. | |
1643 | * As a hack, increase map count until we lock page again. | |
1644 | */ | |
1645 | atomic_inc(&hpage->_mapcount); | |
1646 | unlock_page(hpage); | |
1647 | i_mmap_lock_write(mapping); | |
1648 | lock_page(hpage); | |
1649 | atomic_add_negative(-1, &hpage->_mapcount); | |
1650 | ||
1651 | /* verify page is still mapped */ | |
1652 | if (!page_mapped(hpage)) { | |
1653 | i_mmap_unlock_write(mapping); | |
1654 | return NULL; | |
1655 | } | |
1656 | ||
1657 | /* | |
1658 | * Get address space again and verify it is the same one | |
1659 | * we locked. If not, drop lock and retry. | |
1660 | */ | |
1661 | mapping2 = _get_hugetlb_page_mapping(hpage); | |
1662 | if (mapping2 != mapping) { | |
1663 | i_mmap_unlock_write(mapping); | |
1664 | mapping = mapping2; | |
1665 | goto retry; | |
1666 | } | |
1667 | ||
1668 | return mapping; | |
1669 | } | |
1670 | ||
13d60f4b ZY |
1671 | pgoff_t __basepage_index(struct page *page) |
1672 | { | |
1673 | struct page *page_head = compound_head(page); | |
1674 | pgoff_t index = page_index(page_head); | |
1675 | unsigned long compound_idx; | |
1676 | ||
1677 | if (!PageHuge(page_head)) | |
1678 | return page_index(page); | |
1679 | ||
1680 | if (compound_order(page_head) >= MAX_ORDER) | |
1681 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1682 | else | |
1683 | compound_idx = page - page_head; | |
1684 | ||
1685 | return (index << compound_order(page_head)) + compound_idx; | |
1686 | } | |
1687 | ||
0c397dae | 1688 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1689 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1690 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1691 | { |
af0fb9df | 1692 | int order = huge_page_order(h); |
1da177e4 | 1693 | struct page *page; |
f60858f9 | 1694 | bool alloc_try_hard = true; |
f96efd58 | 1695 | |
f60858f9 MK |
1696 | /* |
1697 | * By default we always try hard to allocate the page with | |
1698 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1699 | * a loop (to adjust global huge page counts) and previous allocation | |
1700 | * failed, do not continue to try hard on the same node. Use the | |
1701 | * node_alloc_noretry bitmap to manage this state information. | |
1702 | */ | |
1703 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1704 | alloc_try_hard = false; | |
1705 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1706 | if (alloc_try_hard) | |
1707 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1708 | if (nid == NUMA_NO_NODE) |
1709 | nid = numa_mem_id(); | |
1710 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1711 | if (page) | |
1712 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1713 | else | |
1714 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1715 | |
f60858f9 MK |
1716 | /* |
1717 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1718 | * indicates an overall state change. Clear bit so that we resume | |
1719 | * normal 'try hard' allocations. | |
1720 | */ | |
1721 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1722 | node_clear(nid, *node_alloc_noretry); | |
1723 | ||
1724 | /* | |
1725 | * If we tried hard to get a page but failed, set bit so that | |
1726 | * subsequent attempts will not try as hard until there is an | |
1727 | * overall state change. | |
1728 | */ | |
1729 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1730 | node_set(nid, *node_alloc_noretry); | |
1731 | ||
63b4613c NA |
1732 | return page; |
1733 | } | |
1734 | ||
0c397dae MH |
1735 | /* |
1736 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1737 | * should use this function to get new hugetlb pages | |
1738 | */ | |
1739 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1740 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1741 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1742 | { |
1743 | struct page *page; | |
1744 | ||
1745 | if (hstate_is_gigantic(h)) | |
1746 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1747 | else | |
1748 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1749 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1750 | if (!page) |
1751 | return NULL; | |
1752 | ||
1753 | if (hstate_is_gigantic(h)) | |
1754 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1755 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1756 | ||
1757 | return page; | |
1758 | } | |
1759 | ||
af0fb9df MH |
1760 | /* |
1761 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1762 | * manner. | |
1763 | */ | |
f60858f9 MK |
1764 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1765 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1766 | { |
1767 | struct page *page; | |
1768 | int nr_nodes, node; | |
af0fb9df | 1769 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1770 | |
1771 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1772 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1773 | node_alloc_noretry); | |
af0fb9df | 1774 | if (page) |
b2261026 | 1775 | break; |
b2261026 JK |
1776 | } |
1777 | ||
af0fb9df MH |
1778 | if (!page) |
1779 | return 0; | |
b2261026 | 1780 | |
af0fb9df MH |
1781 | put_page(page); /* free it into the hugepage allocator */ |
1782 | ||
1783 | return 1; | |
b2261026 JK |
1784 | } |
1785 | ||
e8c5c824 LS |
1786 | /* |
1787 | * Free huge page from pool from next node to free. | |
1788 | * Attempt to keep persistent huge pages more or less | |
1789 | * balanced over allowed nodes. | |
1790 | * Called with hugetlb_lock locked. | |
1791 | */ | |
6ae11b27 LS |
1792 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1793 | bool acct_surplus) | |
e8c5c824 | 1794 | { |
b2261026 | 1795 | int nr_nodes, node; |
e8c5c824 LS |
1796 | int ret = 0; |
1797 | ||
b2261026 | 1798 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1799 | /* |
1800 | * If we're returning unused surplus pages, only examine | |
1801 | * nodes with surplus pages. | |
1802 | */ | |
b2261026 JK |
1803 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1804 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1805 | struct page *page = |
b2261026 | 1806 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1807 | struct page, lru); |
1808 | list_del(&page->lru); | |
1809 | h->free_huge_pages--; | |
b2261026 | 1810 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1811 | if (acct_surplus) { |
1812 | h->surplus_huge_pages--; | |
b2261026 | 1813 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1814 | } |
e8c5c824 LS |
1815 | update_and_free_page(h, page); |
1816 | ret = 1; | |
9a76db09 | 1817 | break; |
e8c5c824 | 1818 | } |
b2261026 | 1819 | } |
e8c5c824 LS |
1820 | |
1821 | return ret; | |
1822 | } | |
1823 | ||
c8721bbb NH |
1824 | /* |
1825 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1826 | * nothing for in-use hugepages and non-hugepages. |
1827 | * This function returns values like below: | |
1828 | * | |
1829 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1830 | * (allocated or reserved.) | |
1831 | * 0: successfully dissolved free hugepages or the page is not a | |
1832 | * hugepage (considered as already dissolved) | |
c8721bbb | 1833 | */ |
c3114a84 | 1834 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1835 | { |
6bc9b564 | 1836 | int rc = -EBUSY; |
082d5b6b | 1837 | |
faf53def NH |
1838 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1839 | if (!PageHuge(page)) | |
1840 | return 0; | |
1841 | ||
c8721bbb | 1842 | spin_lock(&hugetlb_lock); |
faf53def NH |
1843 | if (!PageHuge(page)) { |
1844 | rc = 0; | |
1845 | goto out; | |
1846 | } | |
1847 | ||
1848 | if (!page_count(page)) { | |
2247bb33 GS |
1849 | struct page *head = compound_head(page); |
1850 | struct hstate *h = page_hstate(head); | |
1851 | int nid = page_to_nid(head); | |
6bc9b564 | 1852 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1853 | goto out; |
c3114a84 AK |
1854 | /* |
1855 | * Move PageHWPoison flag from head page to the raw error page, | |
1856 | * which makes any subpages rather than the error page reusable. | |
1857 | */ | |
1858 | if (PageHWPoison(head) && page != head) { | |
1859 | SetPageHWPoison(page); | |
1860 | ClearPageHWPoison(head); | |
1861 | } | |
2247bb33 | 1862 | list_del(&head->lru); |
c8721bbb NH |
1863 | h->free_huge_pages--; |
1864 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1865 | h->max_huge_pages--; |
2247bb33 | 1866 | update_and_free_page(h, head); |
6bc9b564 | 1867 | rc = 0; |
c8721bbb | 1868 | } |
082d5b6b | 1869 | out: |
c8721bbb | 1870 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1871 | return rc; |
c8721bbb NH |
1872 | } |
1873 | ||
1874 | /* | |
1875 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1876 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1877 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1878 | * part of it lies within the given range. | |
082d5b6b GS |
1879 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1880 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1881 | */ |
082d5b6b | 1882 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1883 | { |
c8721bbb | 1884 | unsigned long pfn; |
eb03aa00 | 1885 | struct page *page; |
082d5b6b | 1886 | int rc = 0; |
c8721bbb | 1887 | |
d0177639 | 1888 | if (!hugepages_supported()) |
082d5b6b | 1889 | return rc; |
d0177639 | 1890 | |
eb03aa00 GS |
1891 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1892 | page = pfn_to_page(pfn); | |
faf53def NH |
1893 | rc = dissolve_free_huge_page(page); |
1894 | if (rc) | |
1895 | break; | |
eb03aa00 | 1896 | } |
082d5b6b GS |
1897 | |
1898 | return rc; | |
c8721bbb NH |
1899 | } |
1900 | ||
ab5ac90a MH |
1901 | /* |
1902 | * Allocates a fresh surplus page from the page allocator. | |
1903 | */ | |
0c397dae | 1904 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1905 | int nid, nodemask_t *nmask) |
7893d1d5 | 1906 | { |
9980d744 | 1907 | struct page *page = NULL; |
7893d1d5 | 1908 | |
bae7f4ae | 1909 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1910 | return NULL; |
1911 | ||
d1c3fb1f | 1912 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1913 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1914 | goto out_unlock; | |
d1c3fb1f NA |
1915 | spin_unlock(&hugetlb_lock); |
1916 | ||
f60858f9 | 1917 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1918 | if (!page) |
0c397dae | 1919 | return NULL; |
d1c3fb1f NA |
1920 | |
1921 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1922 | /* |
1923 | * We could have raced with the pool size change. | |
1924 | * Double check that and simply deallocate the new page | |
1925 | * if we would end up overcommiting the surpluses. Abuse | |
1926 | * temporary page to workaround the nasty free_huge_page | |
1927 | * codeflow | |
1928 | */ | |
1929 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1930 | SetPageHugeTemporary(page); | |
2bf753e6 | 1931 | spin_unlock(&hugetlb_lock); |
9980d744 | 1932 | put_page(page); |
2bf753e6 | 1933 | return NULL; |
9980d744 | 1934 | } else { |
9980d744 | 1935 | h->surplus_huge_pages++; |
4704dea3 | 1936 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1937 | } |
9980d744 MH |
1938 | |
1939 | out_unlock: | |
d1c3fb1f | 1940 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1941 | |
1942 | return page; | |
1943 | } | |
1944 | ||
bbe88753 | 1945 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 1946 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
1947 | { |
1948 | struct page *page; | |
1949 | ||
1950 | if (hstate_is_gigantic(h)) | |
1951 | return NULL; | |
1952 | ||
f60858f9 | 1953 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1954 | if (!page) |
1955 | return NULL; | |
1956 | ||
1957 | /* | |
1958 | * We do not account these pages as surplus because they are only | |
1959 | * temporary and will be released properly on the last reference | |
1960 | */ | |
ab5ac90a MH |
1961 | SetPageHugeTemporary(page); |
1962 | ||
1963 | return page; | |
1964 | } | |
1965 | ||
099730d6 DH |
1966 | /* |
1967 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1968 | */ | |
e0ec90ee | 1969 | static |
0c397dae | 1970 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1971 | struct vm_area_struct *vma, unsigned long addr) |
1972 | { | |
aaf14e40 MH |
1973 | struct page *page; |
1974 | struct mempolicy *mpol; | |
1975 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1976 | int nid; | |
1977 | nodemask_t *nodemask; | |
1978 | ||
1979 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1980 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1981 | mpol_cond_put(mpol); |
1982 | ||
1983 | return page; | |
099730d6 DH |
1984 | } |
1985 | ||
ab5ac90a | 1986 | /* page migration callback function */ |
3e59fcb0 | 1987 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 1988 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 1989 | { |
4db9b2ef MH |
1990 | spin_lock(&hugetlb_lock); |
1991 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1992 | struct page *page; |
1993 | ||
1994 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1995 | if (page) { | |
1996 | spin_unlock(&hugetlb_lock); | |
1997 | return page; | |
4db9b2ef MH |
1998 | } |
1999 | } | |
2000 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 2001 | |
0c397dae | 2002 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2003 | } |
2004 | ||
ebd63723 | 2005 | /* mempolicy aware migration callback */ |
389c8178 MH |
2006 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2007 | unsigned long address) | |
ebd63723 MH |
2008 | { |
2009 | struct mempolicy *mpol; | |
2010 | nodemask_t *nodemask; | |
2011 | struct page *page; | |
ebd63723 MH |
2012 | gfp_t gfp_mask; |
2013 | int node; | |
2014 | ||
ebd63723 MH |
2015 | gfp_mask = htlb_alloc_mask(h); |
2016 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 2017 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
2018 | mpol_cond_put(mpol); |
2019 | ||
2020 | return page; | |
2021 | } | |
2022 | ||
e4e574b7 | 2023 | /* |
25985edc | 2024 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2025 | * of size 'delta'. |
2026 | */ | |
a5516438 | 2027 | static int gather_surplus_pages(struct hstate *h, int delta) |
1b2a1e7b | 2028 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
2029 | { |
2030 | struct list_head surplus_list; | |
2031 | struct page *page, *tmp; | |
2032 | int ret, i; | |
2033 | int needed, allocated; | |
28073b02 | 2034 | bool alloc_ok = true; |
e4e574b7 | 2035 | |
a5516438 | 2036 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2037 | if (needed <= 0) { |
a5516438 | 2038 | h->resv_huge_pages += delta; |
e4e574b7 | 2039 | return 0; |
ac09b3a1 | 2040 | } |
e4e574b7 AL |
2041 | |
2042 | allocated = 0; | |
2043 | INIT_LIST_HEAD(&surplus_list); | |
2044 | ||
2045 | ret = -ENOMEM; | |
2046 | retry: | |
2047 | spin_unlock(&hugetlb_lock); | |
2048 | for (i = 0; i < needed; i++) { | |
0c397dae | 2049 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 2050 | NUMA_NO_NODE, NULL); |
28073b02 HD |
2051 | if (!page) { |
2052 | alloc_ok = false; | |
2053 | break; | |
2054 | } | |
e4e574b7 | 2055 | list_add(&page->lru, &surplus_list); |
69ed779a | 2056 | cond_resched(); |
e4e574b7 | 2057 | } |
28073b02 | 2058 | allocated += i; |
e4e574b7 AL |
2059 | |
2060 | /* | |
2061 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2062 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2063 | */ | |
2064 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
2065 | needed = (h->resv_huge_pages + delta) - |
2066 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2067 | if (needed > 0) { |
2068 | if (alloc_ok) | |
2069 | goto retry; | |
2070 | /* | |
2071 | * We were not able to allocate enough pages to | |
2072 | * satisfy the entire reservation so we free what | |
2073 | * we've allocated so far. | |
2074 | */ | |
2075 | goto free; | |
2076 | } | |
e4e574b7 AL |
2077 | /* |
2078 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2079 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2080 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2081 | * allocator. Commit the entire reservation here to prevent another |
2082 | * process from stealing the pages as they are added to the pool but | |
2083 | * before they are reserved. | |
e4e574b7 AL |
2084 | */ |
2085 | needed += allocated; | |
a5516438 | 2086 | h->resv_huge_pages += delta; |
e4e574b7 | 2087 | ret = 0; |
a9869b83 | 2088 | |
19fc3f0a | 2089 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2090 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2091 | if ((--needed) < 0) |
2092 | break; | |
a9869b83 NH |
2093 | /* |
2094 | * This page is now managed by the hugetlb allocator and has | |
2095 | * no users -- drop the buddy allocator's reference. | |
2096 | */ | |
2097 | put_page_testzero(page); | |
309381fe | 2098 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 2099 | enqueue_huge_page(h, page); |
19fc3f0a | 2100 | } |
28073b02 | 2101 | free: |
b0365c8d | 2102 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
2103 | |
2104 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
2105 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
2106 | put_page(page); | |
a9869b83 | 2107 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
2108 | |
2109 | return ret; | |
2110 | } | |
2111 | ||
2112 | /* | |
e5bbc8a6 MK |
2113 | * This routine has two main purposes: |
2114 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2115 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2116 | * to the associated reservation map. | |
2117 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2118 | * the reservation. As many as unused_resv_pages may be freed. | |
2119 | * | |
2120 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
2121 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
2122 | * we must make sure nobody else can claim pages we are in the process of | |
2123 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
2124 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 2125 | */ |
a5516438 AK |
2126 | static void return_unused_surplus_pages(struct hstate *h, |
2127 | unsigned long unused_resv_pages) | |
e4e574b7 | 2128 | { |
e4e574b7 AL |
2129 | unsigned long nr_pages; |
2130 | ||
aa888a74 | 2131 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2132 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2133 | goto out; |
aa888a74 | 2134 | |
e5bbc8a6 MK |
2135 | /* |
2136 | * Part (or even all) of the reservation could have been backed | |
2137 | * by pre-allocated pages. Only free surplus pages. | |
2138 | */ | |
a5516438 | 2139 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2140 | |
685f3457 LS |
2141 | /* |
2142 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2143 | * evenly across all nodes with memory. Iterate across these nodes |
2144 | * until we can no longer free unreserved surplus pages. This occurs | |
2145 | * when the nodes with surplus pages have no free pages. | |
9e7ee400 | 2146 | * free_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2147 | * on-line nodes with memory and will handle the hstate accounting. |
e5bbc8a6 MK |
2148 | * |
2149 | * Note that we decrement resv_huge_pages as we free the pages. If | |
2150 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
2151 | * to cover subsequent pages we may free. | |
685f3457 LS |
2152 | */ |
2153 | while (nr_pages--) { | |
e5bbc8a6 MK |
2154 | h->resv_huge_pages--; |
2155 | unused_resv_pages--; | |
8cebfcd0 | 2156 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 2157 | goto out; |
7848a4bf | 2158 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 2159 | } |
e5bbc8a6 MK |
2160 | |
2161 | out: | |
2162 | /* Fully uncommit the reservation */ | |
2163 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
2164 | } |
2165 | ||
5e911373 | 2166 | |
c37f9fb1 | 2167 | /* |
feba16e2 | 2168 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2169 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2170 | * |
2171 | * vma_needs_reservation is called to determine if the huge page at addr | |
2172 | * within the vma has an associated reservation. If a reservation is | |
2173 | * needed, the value 1 is returned. The caller is then responsible for | |
2174 | * managing the global reservation and subpool usage counts. After | |
2175 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2176 | * to add the page to the reservation map. If the page allocation fails, |
2177 | * the reservation must be ended instead of committed. vma_end_reservation | |
2178 | * is called in such cases. | |
cf3ad20b MK |
2179 | * |
2180 | * In the normal case, vma_commit_reservation returns the same value | |
2181 | * as the preceding vma_needs_reservation call. The only time this | |
2182 | * is not the case is if a reserve map was changed between calls. It | |
2183 | * is the responsibility of the caller to notice the difference and | |
2184 | * take appropriate action. | |
96b96a96 MK |
2185 | * |
2186 | * vma_add_reservation is used in error paths where a reservation must | |
2187 | * be restored when a newly allocated huge page must be freed. It is | |
2188 | * to be called after calling vma_needs_reservation to determine if a | |
2189 | * reservation exists. | |
c37f9fb1 | 2190 | */ |
5e911373 MK |
2191 | enum vma_resv_mode { |
2192 | VMA_NEEDS_RESV, | |
2193 | VMA_COMMIT_RESV, | |
feba16e2 | 2194 | VMA_END_RESV, |
96b96a96 | 2195 | VMA_ADD_RESV, |
5e911373 | 2196 | }; |
cf3ad20b MK |
2197 | static long __vma_reservation_common(struct hstate *h, |
2198 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2199 | enum vma_resv_mode mode) |
c37f9fb1 | 2200 | { |
4e35f483 JK |
2201 | struct resv_map *resv; |
2202 | pgoff_t idx; | |
cf3ad20b | 2203 | long ret; |
0db9d74e | 2204 | long dummy_out_regions_needed; |
c37f9fb1 | 2205 | |
4e35f483 JK |
2206 | resv = vma_resv_map(vma); |
2207 | if (!resv) | |
84afd99b | 2208 | return 1; |
c37f9fb1 | 2209 | |
4e35f483 | 2210 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2211 | switch (mode) { |
2212 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2213 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2214 | /* We assume that vma_reservation_* routines always operate on | |
2215 | * 1 page, and that adding to resv map a 1 page entry can only | |
2216 | * ever require 1 region. | |
2217 | */ | |
2218 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2219 | break; |
2220 | case VMA_COMMIT_RESV: | |
075a61d0 | 2221 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2222 | /* region_add calls of range 1 should never fail. */ |
2223 | VM_BUG_ON(ret < 0); | |
5e911373 | 2224 | break; |
feba16e2 | 2225 | case VMA_END_RESV: |
0db9d74e | 2226 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2227 | ret = 0; |
2228 | break; | |
96b96a96 | 2229 | case VMA_ADD_RESV: |
0db9d74e | 2230 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2231 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2232 | /* region_add calls of range 1 should never fail. */ |
2233 | VM_BUG_ON(ret < 0); | |
2234 | } else { | |
2235 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2236 | ret = region_del(resv, idx, idx + 1); |
2237 | } | |
2238 | break; | |
5e911373 MK |
2239 | default: |
2240 | BUG(); | |
2241 | } | |
84afd99b | 2242 | |
4e35f483 | 2243 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 2244 | return ret; |
67961f9d MK |
2245 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
2246 | /* | |
2247 | * In most cases, reserves always exist for private mappings. | |
2248 | * However, a file associated with mapping could have been | |
2249 | * hole punched or truncated after reserves were consumed. | |
2250 | * As subsequent fault on such a range will not use reserves. | |
2251 | * Subtle - The reserve map for private mappings has the | |
2252 | * opposite meaning than that of shared mappings. If NO | |
2253 | * entry is in the reserve map, it means a reservation exists. | |
2254 | * If an entry exists in the reserve map, it means the | |
2255 | * reservation has already been consumed. As a result, the | |
2256 | * return value of this routine is the opposite of the | |
2257 | * value returned from reserve map manipulation routines above. | |
2258 | */ | |
2259 | if (ret) | |
2260 | return 0; | |
2261 | else | |
2262 | return 1; | |
2263 | } | |
4e35f483 | 2264 | else |
cf3ad20b | 2265 | return ret < 0 ? ret : 0; |
c37f9fb1 | 2266 | } |
cf3ad20b MK |
2267 | |
2268 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2269 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2270 | { |
5e911373 | 2271 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2272 | } |
84afd99b | 2273 | |
cf3ad20b MK |
2274 | static long vma_commit_reservation(struct hstate *h, |
2275 | struct vm_area_struct *vma, unsigned long addr) | |
2276 | { | |
5e911373 MK |
2277 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2278 | } | |
2279 | ||
feba16e2 | 2280 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2281 | struct vm_area_struct *vma, unsigned long addr) |
2282 | { | |
feba16e2 | 2283 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2284 | } |
2285 | ||
96b96a96 MK |
2286 | static long vma_add_reservation(struct hstate *h, |
2287 | struct vm_area_struct *vma, unsigned long addr) | |
2288 | { | |
2289 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2290 | } | |
2291 | ||
2292 | /* | |
2293 | * This routine is called to restore a reservation on error paths. In the | |
2294 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2295 | * and is about to be freed. If a reservation for the page existed, | |
2296 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2297 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2298 | * the global reservation count will be incremented if PagePrivate is set. | |
2299 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2300 | * reserve map here to be consistent with global reserve count adjustments | |
2301 | * to be made by free_huge_page. | |
2302 | */ | |
2303 | static void restore_reserve_on_error(struct hstate *h, | |
2304 | struct vm_area_struct *vma, unsigned long address, | |
2305 | struct page *page) | |
2306 | { | |
2307 | if (unlikely(PagePrivate(page))) { | |
2308 | long rc = vma_needs_reservation(h, vma, address); | |
2309 | ||
2310 | if (unlikely(rc < 0)) { | |
2311 | /* | |
2312 | * Rare out of memory condition in reserve map | |
2313 | * manipulation. Clear PagePrivate so that | |
2314 | * global reserve count will not be incremented | |
2315 | * by free_huge_page. This will make it appear | |
2316 | * as though the reservation for this page was | |
2317 | * consumed. This may prevent the task from | |
2318 | * faulting in the page at a later time. This | |
2319 | * is better than inconsistent global huge page | |
2320 | * accounting of reserve counts. | |
2321 | */ | |
2322 | ClearPagePrivate(page); | |
2323 | } else if (rc) { | |
2324 | rc = vma_add_reservation(h, vma, address); | |
2325 | if (unlikely(rc < 0)) | |
2326 | /* | |
2327 | * See above comment about rare out of | |
2328 | * memory condition. | |
2329 | */ | |
2330 | ClearPagePrivate(page); | |
2331 | } else | |
2332 | vma_end_reservation(h, vma, address); | |
2333 | } | |
2334 | } | |
2335 | ||
70c3547e | 2336 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2337 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2338 | { |
90481622 | 2339 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2340 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2341 | struct page *page; |
d85f69b0 MK |
2342 | long map_chg, map_commit; |
2343 | long gbl_chg; | |
6d76dcf4 AK |
2344 | int ret, idx; |
2345 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2346 | bool deferred_reserve; |
a1e78772 | 2347 | |
6d76dcf4 | 2348 | idx = hstate_index(h); |
a1e78772 | 2349 | /* |
d85f69b0 MK |
2350 | * Examine the region/reserve map to determine if the process |
2351 | * has a reservation for the page to be allocated. A return | |
2352 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2353 | */ |
d85f69b0 MK |
2354 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2355 | if (map_chg < 0) | |
76dcee75 | 2356 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2357 | |
2358 | /* | |
2359 | * Processes that did not create the mapping will have no | |
2360 | * reserves as indicated by the region/reserve map. Check | |
2361 | * that the allocation will not exceed the subpool limit. | |
2362 | * Allocations for MAP_NORESERVE mappings also need to be | |
2363 | * checked against any subpool limit. | |
2364 | */ | |
2365 | if (map_chg || avoid_reserve) { | |
2366 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2367 | if (gbl_chg < 0) { | |
feba16e2 | 2368 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2369 | return ERR_PTR(-ENOSPC); |
5e911373 | 2370 | } |
1da177e4 | 2371 | |
d85f69b0 MK |
2372 | /* |
2373 | * Even though there was no reservation in the region/reserve | |
2374 | * map, there could be reservations associated with the | |
2375 | * subpool that can be used. This would be indicated if the | |
2376 | * return value of hugepage_subpool_get_pages() is zero. | |
2377 | * However, if avoid_reserve is specified we still avoid even | |
2378 | * the subpool reservations. | |
2379 | */ | |
2380 | if (avoid_reserve) | |
2381 | gbl_chg = 1; | |
2382 | } | |
2383 | ||
08cf9faf MA |
2384 | /* If this allocation is not consuming a reservation, charge it now. |
2385 | */ | |
2386 | deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); | |
2387 | if (deferred_reserve) { | |
2388 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2389 | idx, pages_per_huge_page(h), &h_cg); | |
2390 | if (ret) | |
2391 | goto out_subpool_put; | |
2392 | } | |
2393 | ||
6d76dcf4 | 2394 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2395 | if (ret) |
08cf9faf | 2396 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2397 | |
1da177e4 | 2398 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2399 | /* |
2400 | * glb_chg is passed to indicate whether or not a page must be taken | |
2401 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2402 | * a reservation exists for the allocation. | |
2403 | */ | |
2404 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2405 | if (!page) { |
94ae8ba7 | 2406 | spin_unlock(&hugetlb_lock); |
0c397dae | 2407 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2408 | if (!page) |
2409 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2410 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2411 | SetPagePrivate(page); | |
2412 | h->resv_huge_pages--; | |
2413 | } | |
79dbb236 AK |
2414 | spin_lock(&hugetlb_lock); |
2415 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2416 | /* Fall through */ |
68842c9b | 2417 | } |
81a6fcae | 2418 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2419 | /* If allocation is not consuming a reservation, also store the |
2420 | * hugetlb_cgroup pointer on the page. | |
2421 | */ | |
2422 | if (deferred_reserve) { | |
2423 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2424 | h_cg, page); | |
2425 | } | |
2426 | ||
81a6fcae | 2427 | spin_unlock(&hugetlb_lock); |
348ea204 | 2428 | |
90481622 | 2429 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2430 | |
d85f69b0 MK |
2431 | map_commit = vma_commit_reservation(h, vma, addr); |
2432 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2433 | /* |
2434 | * The page was added to the reservation map between | |
2435 | * vma_needs_reservation and vma_commit_reservation. | |
2436 | * This indicates a race with hugetlb_reserve_pages. | |
2437 | * Adjust for the subpool count incremented above AND | |
2438 | * in hugetlb_reserve_pages for the same page. Also, | |
2439 | * the reservation count added in hugetlb_reserve_pages | |
2440 | * no longer applies. | |
2441 | */ | |
2442 | long rsv_adjust; | |
2443 | ||
2444 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2445 | hugetlb_acct_memory(h, -rsv_adjust); | |
2446 | } | |
90d8b7e6 | 2447 | return page; |
8f34af6f JZ |
2448 | |
2449 | out_uncharge_cgroup: | |
2450 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2451 | out_uncharge_cgroup_reservation: |
2452 | if (deferred_reserve) | |
2453 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2454 | h_cg); | |
8f34af6f | 2455 | out_subpool_put: |
d85f69b0 | 2456 | if (map_chg || avoid_reserve) |
8f34af6f | 2457 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2458 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2459 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2460 | } |
2461 | ||
e24a1307 AK |
2462 | int alloc_bootmem_huge_page(struct hstate *h) |
2463 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2464 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2465 | { |
2466 | struct huge_bootmem_page *m; | |
b2261026 | 2467 | int nr_nodes, node; |
aa888a74 | 2468 | |
b2261026 | 2469 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2470 | void *addr; |
2471 | ||
eb31d559 | 2472 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2473 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2474 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2475 | if (addr) { |
2476 | /* | |
2477 | * Use the beginning of the huge page to store the | |
2478 | * huge_bootmem_page struct (until gather_bootmem | |
2479 | * puts them into the mem_map). | |
2480 | */ | |
2481 | m = addr; | |
91f47662 | 2482 | goto found; |
aa888a74 | 2483 | } |
aa888a74 AK |
2484 | } |
2485 | return 0; | |
2486 | ||
2487 | found: | |
df994ead | 2488 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2489 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2490 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2491 | list_add(&m->list, &huge_boot_pages); |
2492 | m->hstate = h; | |
2493 | return 1; | |
2494 | } | |
2495 | ||
d00181b9 KS |
2496 | static void __init prep_compound_huge_page(struct page *page, |
2497 | unsigned int order) | |
18229df5 AW |
2498 | { |
2499 | if (unlikely(order > (MAX_ORDER - 1))) | |
2500 | prep_compound_gigantic_page(page, order); | |
2501 | else | |
2502 | prep_compound_page(page, order); | |
2503 | } | |
2504 | ||
aa888a74 AK |
2505 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2506 | static void __init gather_bootmem_prealloc(void) | |
2507 | { | |
2508 | struct huge_bootmem_page *m; | |
2509 | ||
2510 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2511 | struct page *page = virt_to_page(m); |
aa888a74 | 2512 | struct hstate *h = m->hstate; |
ee8f248d | 2513 | |
aa888a74 | 2514 | WARN_ON(page_count(page) != 1); |
18229df5 | 2515 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2516 | WARN_ON(PageReserved(page)); |
aa888a74 | 2517 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2518 | put_page(page); /* free it into the hugepage allocator */ |
2519 | ||
b0320c7b RA |
2520 | /* |
2521 | * If we had gigantic hugepages allocated at boot time, we need | |
2522 | * to restore the 'stolen' pages to totalram_pages in order to | |
2523 | * fix confusing memory reports from free(1) and another | |
2524 | * side-effects, like CommitLimit going negative. | |
2525 | */ | |
bae7f4ae | 2526 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2527 | adjust_managed_page_count(page, 1 << h->order); |
520495fe | 2528 | cond_resched(); |
aa888a74 AK |
2529 | } |
2530 | } | |
2531 | ||
8faa8b07 | 2532 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2533 | { |
2534 | unsigned long i; | |
f60858f9 MK |
2535 | nodemask_t *node_alloc_noretry; |
2536 | ||
2537 | if (!hstate_is_gigantic(h)) { | |
2538 | /* | |
2539 | * Bit mask controlling how hard we retry per-node allocations. | |
2540 | * Ignore errors as lower level routines can deal with | |
2541 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2542 | * time, we are likely in bigger trouble. | |
2543 | */ | |
2544 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2545 | GFP_KERNEL); | |
2546 | } else { | |
2547 | /* allocations done at boot time */ | |
2548 | node_alloc_noretry = NULL; | |
2549 | } | |
2550 | ||
2551 | /* bit mask controlling how hard we retry per-node allocations */ | |
2552 | if (node_alloc_noretry) | |
2553 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2554 | |
e5ff2159 | 2555 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2556 | if (hstate_is_gigantic(h)) { |
dbda8fea | 2557 | if (hugetlb_cma_size) { |
cf11e85f RG |
2558 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); |
2559 | break; | |
2560 | } | |
aa888a74 AK |
2561 | if (!alloc_bootmem_huge_page(h)) |
2562 | break; | |
0c397dae | 2563 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2564 | &node_states[N_MEMORY], |
2565 | node_alloc_noretry)) | |
1da177e4 | 2566 | break; |
69ed779a | 2567 | cond_resched(); |
1da177e4 | 2568 | } |
d715cf80 LH |
2569 | if (i < h->max_huge_pages) { |
2570 | char buf[32]; | |
2571 | ||
c6247f72 | 2572 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2573 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2574 | h->max_huge_pages, buf, i); | |
2575 | h->max_huge_pages = i; | |
2576 | } | |
f60858f9 MK |
2577 | |
2578 | kfree(node_alloc_noretry); | |
e5ff2159 AK |
2579 | } |
2580 | ||
2581 | static void __init hugetlb_init_hstates(void) | |
2582 | { | |
2583 | struct hstate *h; | |
2584 | ||
2585 | for_each_hstate(h) { | |
641844f5 NH |
2586 | if (minimum_order > huge_page_order(h)) |
2587 | minimum_order = huge_page_order(h); | |
2588 | ||
8faa8b07 | 2589 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2590 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2591 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2592 | } |
641844f5 | 2593 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2594 | } |
2595 | ||
2596 | static void __init report_hugepages(void) | |
2597 | { | |
2598 | struct hstate *h; | |
2599 | ||
2600 | for_each_hstate(h) { | |
4abd32db | 2601 | char buf[32]; |
c6247f72 MW |
2602 | |
2603 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2604 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2605 | buf, h->free_huge_pages); |
e5ff2159 AK |
2606 | } |
2607 | } | |
2608 | ||
1da177e4 | 2609 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2610 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2611 | nodemask_t *nodes_allowed) | |
1da177e4 | 2612 | { |
4415cc8d CL |
2613 | int i; |
2614 | ||
bae7f4ae | 2615 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2616 | return; |
2617 | ||
6ae11b27 | 2618 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2619 | struct page *page, *next; |
a5516438 AK |
2620 | struct list_head *freel = &h->hugepage_freelists[i]; |
2621 | list_for_each_entry_safe(page, next, freel, lru) { | |
2622 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2623 | return; |
1da177e4 LT |
2624 | if (PageHighMem(page)) |
2625 | continue; | |
2626 | list_del(&page->lru); | |
e5ff2159 | 2627 | update_and_free_page(h, page); |
a5516438 AK |
2628 | h->free_huge_pages--; |
2629 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2630 | } |
2631 | } | |
2632 | } | |
2633 | #else | |
6ae11b27 LS |
2634 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2635 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2636 | { |
2637 | } | |
2638 | #endif | |
2639 | ||
20a0307c WF |
2640 | /* |
2641 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2642 | * balanced by operating on them in a round-robin fashion. | |
2643 | * Returns 1 if an adjustment was made. | |
2644 | */ | |
6ae11b27 LS |
2645 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2646 | int delta) | |
20a0307c | 2647 | { |
b2261026 | 2648 | int nr_nodes, node; |
20a0307c WF |
2649 | |
2650 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2651 | |
b2261026 JK |
2652 | if (delta < 0) { |
2653 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2654 | if (h->surplus_huge_pages_node[node]) | |
2655 | goto found; | |
e8c5c824 | 2656 | } |
b2261026 JK |
2657 | } else { |
2658 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2659 | if (h->surplus_huge_pages_node[node] < | |
2660 | h->nr_huge_pages_node[node]) | |
2661 | goto found; | |
e8c5c824 | 2662 | } |
b2261026 JK |
2663 | } |
2664 | return 0; | |
20a0307c | 2665 | |
b2261026 JK |
2666 | found: |
2667 | h->surplus_huge_pages += delta; | |
2668 | h->surplus_huge_pages_node[node] += delta; | |
2669 | return 1; | |
20a0307c WF |
2670 | } |
2671 | ||
a5516438 | 2672 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2673 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2674 | nodemask_t *nodes_allowed) |
1da177e4 | 2675 | { |
7893d1d5 | 2676 | unsigned long min_count, ret; |
f60858f9 MK |
2677 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2678 | ||
2679 | /* | |
2680 | * Bit mask controlling how hard we retry per-node allocations. | |
2681 | * If we can not allocate the bit mask, do not attempt to allocate | |
2682 | * the requested huge pages. | |
2683 | */ | |
2684 | if (node_alloc_noretry) | |
2685 | nodes_clear(*node_alloc_noretry); | |
2686 | else | |
2687 | return -ENOMEM; | |
1da177e4 | 2688 | |
4eb0716e AG |
2689 | spin_lock(&hugetlb_lock); |
2690 | ||
fd875dca MK |
2691 | /* |
2692 | * Check for a node specific request. | |
2693 | * Changing node specific huge page count may require a corresponding | |
2694 | * change to the global count. In any case, the passed node mask | |
2695 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2696 | */ | |
2697 | if (nid != NUMA_NO_NODE) { | |
2698 | unsigned long old_count = count; | |
2699 | ||
2700 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2701 | /* | |
2702 | * User may have specified a large count value which caused the | |
2703 | * above calculation to overflow. In this case, they wanted | |
2704 | * to allocate as many huge pages as possible. Set count to | |
2705 | * largest possible value to align with their intention. | |
2706 | */ | |
2707 | if (count < old_count) | |
2708 | count = ULONG_MAX; | |
2709 | } | |
2710 | ||
4eb0716e AG |
2711 | /* |
2712 | * Gigantic pages runtime allocation depend on the capability for large | |
2713 | * page range allocation. | |
2714 | * If the system does not provide this feature, return an error when | |
2715 | * the user tries to allocate gigantic pages but let the user free the | |
2716 | * boottime allocated gigantic pages. | |
2717 | */ | |
2718 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2719 | if (count > persistent_huge_pages(h)) { | |
2720 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2721 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2722 | return -EINVAL; |
2723 | } | |
2724 | /* Fall through to decrease pool */ | |
2725 | } | |
aa888a74 | 2726 | |
7893d1d5 AL |
2727 | /* |
2728 | * Increase the pool size | |
2729 | * First take pages out of surplus state. Then make up the | |
2730 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2731 | * |
0c397dae | 2732 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2733 | * to convert a surplus huge page to a normal huge page. That is |
2734 | * not critical, though, it just means the overall size of the | |
2735 | * pool might be one hugepage larger than it needs to be, but | |
2736 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2737 | */ |
a5516438 | 2738 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2739 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2740 | break; |
2741 | } | |
2742 | ||
a5516438 | 2743 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2744 | /* |
2745 | * If this allocation races such that we no longer need the | |
2746 | * page, free_huge_page will handle it by freeing the page | |
2747 | * and reducing the surplus. | |
2748 | */ | |
2749 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2750 | |
2751 | /* yield cpu to avoid soft lockup */ | |
2752 | cond_resched(); | |
2753 | ||
f60858f9 MK |
2754 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2755 | node_alloc_noretry); | |
7893d1d5 AL |
2756 | spin_lock(&hugetlb_lock); |
2757 | if (!ret) | |
2758 | goto out; | |
2759 | ||
536240f2 MG |
2760 | /* Bail for signals. Probably ctrl-c from user */ |
2761 | if (signal_pending(current)) | |
2762 | goto out; | |
7893d1d5 | 2763 | } |
7893d1d5 AL |
2764 | |
2765 | /* | |
2766 | * Decrease the pool size | |
2767 | * First return free pages to the buddy allocator (being careful | |
2768 | * to keep enough around to satisfy reservations). Then place | |
2769 | * pages into surplus state as needed so the pool will shrink | |
2770 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2771 | * |
2772 | * By placing pages into the surplus state independent of the | |
2773 | * overcommit value, we are allowing the surplus pool size to | |
2774 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2775 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2776 | * though, we'll note that we're not allowed to exceed surplus |
2777 | * and won't grow the pool anywhere else. Not until one of the | |
2778 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2779 | */ |
a5516438 | 2780 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2781 | min_count = max(count, min_count); |
6ae11b27 | 2782 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2783 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2784 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2785 | break; |
55f67141 | 2786 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2787 | } |
a5516438 | 2788 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2789 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2790 | break; |
2791 | } | |
2792 | out: | |
4eb0716e | 2793 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2794 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2795 | |
f60858f9 MK |
2796 | NODEMASK_FREE(node_alloc_noretry); |
2797 | ||
4eb0716e | 2798 | return 0; |
1da177e4 LT |
2799 | } |
2800 | ||
a3437870 NA |
2801 | #define HSTATE_ATTR_RO(_name) \ |
2802 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2803 | ||
2804 | #define HSTATE_ATTR(_name) \ | |
2805 | static struct kobj_attribute _name##_attr = \ | |
2806 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2807 | ||
2808 | static struct kobject *hugepages_kobj; | |
2809 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2810 | ||
9a305230 LS |
2811 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2812 | ||
2813 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2814 | { |
2815 | int i; | |
9a305230 | 2816 | |
a3437870 | 2817 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2818 | if (hstate_kobjs[i] == kobj) { |
2819 | if (nidp) | |
2820 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2821 | return &hstates[i]; |
9a305230 LS |
2822 | } |
2823 | ||
2824 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2825 | } |
2826 | ||
06808b08 | 2827 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2828 | struct kobj_attribute *attr, char *buf) |
2829 | { | |
9a305230 LS |
2830 | struct hstate *h; |
2831 | unsigned long nr_huge_pages; | |
2832 | int nid; | |
2833 | ||
2834 | h = kobj_to_hstate(kobj, &nid); | |
2835 | if (nid == NUMA_NO_NODE) | |
2836 | nr_huge_pages = h->nr_huge_pages; | |
2837 | else | |
2838 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2839 | ||
2840 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2841 | } |
adbe8726 | 2842 | |
238d3c13 DR |
2843 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2844 | struct hstate *h, int nid, | |
2845 | unsigned long count, size_t len) | |
a3437870 NA |
2846 | { |
2847 | int err; | |
2d0adf7e | 2848 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2849 | |
2d0adf7e OS |
2850 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2851 | return -EINVAL; | |
adbe8726 | 2852 | |
9a305230 LS |
2853 | if (nid == NUMA_NO_NODE) { |
2854 | /* | |
2855 | * global hstate attribute | |
2856 | */ | |
2857 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2858 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2859 | n_mask = &node_states[N_MEMORY]; | |
2860 | else | |
2861 | n_mask = &nodes_allowed; | |
2862 | } else { | |
9a305230 | 2863 | /* |
fd875dca MK |
2864 | * Node specific request. count adjustment happens in |
2865 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2866 | */ |
2d0adf7e OS |
2867 | init_nodemask_of_node(&nodes_allowed, nid); |
2868 | n_mask = &nodes_allowed; | |
fd875dca | 2869 | } |
9a305230 | 2870 | |
2d0adf7e | 2871 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2872 | |
4eb0716e | 2873 | return err ? err : len; |
06808b08 LS |
2874 | } |
2875 | ||
238d3c13 DR |
2876 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2877 | struct kobject *kobj, const char *buf, | |
2878 | size_t len) | |
2879 | { | |
2880 | struct hstate *h; | |
2881 | unsigned long count; | |
2882 | int nid; | |
2883 | int err; | |
2884 | ||
2885 | err = kstrtoul(buf, 10, &count); | |
2886 | if (err) | |
2887 | return err; | |
2888 | ||
2889 | h = kobj_to_hstate(kobj, &nid); | |
2890 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2891 | } | |
2892 | ||
06808b08 LS |
2893 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2894 | struct kobj_attribute *attr, char *buf) | |
2895 | { | |
2896 | return nr_hugepages_show_common(kobj, attr, buf); | |
2897 | } | |
2898 | ||
2899 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2900 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2901 | { | |
238d3c13 | 2902 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2903 | } |
2904 | HSTATE_ATTR(nr_hugepages); | |
2905 | ||
06808b08 LS |
2906 | #ifdef CONFIG_NUMA |
2907 | ||
2908 | /* | |
2909 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2910 | * huge page alloc/free. | |
2911 | */ | |
2912 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2913 | struct kobj_attribute *attr, char *buf) | |
2914 | { | |
2915 | return nr_hugepages_show_common(kobj, attr, buf); | |
2916 | } | |
2917 | ||
2918 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2919 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2920 | { | |
238d3c13 | 2921 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2922 | } |
2923 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2924 | #endif | |
2925 | ||
2926 | ||
a3437870 NA |
2927 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2928 | struct kobj_attribute *attr, char *buf) | |
2929 | { | |
9a305230 | 2930 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2931 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2932 | } | |
adbe8726 | 2933 | |
a3437870 NA |
2934 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2935 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2936 | { | |
2937 | int err; | |
2938 | unsigned long input; | |
9a305230 | 2939 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2940 | |
bae7f4ae | 2941 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2942 | return -EINVAL; |
2943 | ||
3dbb95f7 | 2944 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2945 | if (err) |
73ae31e5 | 2946 | return err; |
a3437870 NA |
2947 | |
2948 | spin_lock(&hugetlb_lock); | |
2949 | h->nr_overcommit_huge_pages = input; | |
2950 | spin_unlock(&hugetlb_lock); | |
2951 | ||
2952 | return count; | |
2953 | } | |
2954 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2955 | ||
2956 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2957 | struct kobj_attribute *attr, char *buf) | |
2958 | { | |
9a305230 LS |
2959 | struct hstate *h; |
2960 | unsigned long free_huge_pages; | |
2961 | int nid; | |
2962 | ||
2963 | h = kobj_to_hstate(kobj, &nid); | |
2964 | if (nid == NUMA_NO_NODE) | |
2965 | free_huge_pages = h->free_huge_pages; | |
2966 | else | |
2967 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2968 | ||
2969 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2970 | } |
2971 | HSTATE_ATTR_RO(free_hugepages); | |
2972 | ||
2973 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2974 | struct kobj_attribute *attr, char *buf) | |
2975 | { | |
9a305230 | 2976 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2977 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2978 | } | |
2979 | HSTATE_ATTR_RO(resv_hugepages); | |
2980 | ||
2981 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2982 | struct kobj_attribute *attr, char *buf) | |
2983 | { | |
9a305230 LS |
2984 | struct hstate *h; |
2985 | unsigned long surplus_huge_pages; | |
2986 | int nid; | |
2987 | ||
2988 | h = kobj_to_hstate(kobj, &nid); | |
2989 | if (nid == NUMA_NO_NODE) | |
2990 | surplus_huge_pages = h->surplus_huge_pages; | |
2991 | else | |
2992 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2993 | ||
2994 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2995 | } |
2996 | HSTATE_ATTR_RO(surplus_hugepages); | |
2997 | ||
2998 | static struct attribute *hstate_attrs[] = { | |
2999 | &nr_hugepages_attr.attr, | |
3000 | &nr_overcommit_hugepages_attr.attr, | |
3001 | &free_hugepages_attr.attr, | |
3002 | &resv_hugepages_attr.attr, | |
3003 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3004 | #ifdef CONFIG_NUMA |
3005 | &nr_hugepages_mempolicy_attr.attr, | |
3006 | #endif | |
a3437870 NA |
3007 | NULL, |
3008 | }; | |
3009 | ||
67e5ed96 | 3010 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3011 | .attrs = hstate_attrs, |
3012 | }; | |
3013 | ||
094e9539 JM |
3014 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3015 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3016 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3017 | { |
3018 | int retval; | |
972dc4de | 3019 | int hi = hstate_index(h); |
a3437870 | 3020 | |
9a305230 LS |
3021 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3022 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3023 | return -ENOMEM; |
3024 | ||
9a305230 | 3025 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 3026 | if (retval) |
9a305230 | 3027 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
3028 | |
3029 | return retval; | |
3030 | } | |
3031 | ||
3032 | static void __init hugetlb_sysfs_init(void) | |
3033 | { | |
3034 | struct hstate *h; | |
3035 | int err; | |
3036 | ||
3037 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3038 | if (!hugepages_kobj) | |
3039 | return; | |
3040 | ||
3041 | for_each_hstate(h) { | |
9a305230 LS |
3042 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3043 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3044 | if (err) |
282f4214 | 3045 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
3046 | } |
3047 | } | |
3048 | ||
9a305230 LS |
3049 | #ifdef CONFIG_NUMA |
3050 | ||
3051 | /* | |
3052 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3053 | * with node devices in node_devices[] using a parallel array. The array |
3054 | * index of a node device or _hstate == node id. | |
3055 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3056 | * the base kernel, on the hugetlb module. |
3057 | */ | |
3058 | struct node_hstate { | |
3059 | struct kobject *hugepages_kobj; | |
3060 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3061 | }; | |
b4e289a6 | 3062 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3063 | |
3064 | /* | |
10fbcf4c | 3065 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3066 | */ |
3067 | static struct attribute *per_node_hstate_attrs[] = { | |
3068 | &nr_hugepages_attr.attr, | |
3069 | &free_hugepages_attr.attr, | |
3070 | &surplus_hugepages_attr.attr, | |
3071 | NULL, | |
3072 | }; | |
3073 | ||
67e5ed96 | 3074 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3075 | .attrs = per_node_hstate_attrs, |
3076 | }; | |
3077 | ||
3078 | /* | |
10fbcf4c | 3079 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3080 | * Returns node id via non-NULL nidp. |
3081 | */ | |
3082 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3083 | { | |
3084 | int nid; | |
3085 | ||
3086 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3087 | struct node_hstate *nhs = &node_hstates[nid]; | |
3088 | int i; | |
3089 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3090 | if (nhs->hstate_kobjs[i] == kobj) { | |
3091 | if (nidp) | |
3092 | *nidp = nid; | |
3093 | return &hstates[i]; | |
3094 | } | |
3095 | } | |
3096 | ||
3097 | BUG(); | |
3098 | return NULL; | |
3099 | } | |
3100 | ||
3101 | /* | |
10fbcf4c | 3102 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3103 | * No-op if no hstate attributes attached. |
3104 | */ | |
3cd8b44f | 3105 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3106 | { |
3107 | struct hstate *h; | |
10fbcf4c | 3108 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3109 | |
3110 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3111 | return; /* no hstate attributes */ |
9a305230 | 3112 | |
972dc4de AK |
3113 | for_each_hstate(h) { |
3114 | int idx = hstate_index(h); | |
3115 | if (nhs->hstate_kobjs[idx]) { | |
3116 | kobject_put(nhs->hstate_kobjs[idx]); | |
3117 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3118 | } |
972dc4de | 3119 | } |
9a305230 LS |
3120 | |
3121 | kobject_put(nhs->hugepages_kobj); | |
3122 | nhs->hugepages_kobj = NULL; | |
3123 | } | |
3124 | ||
9a305230 LS |
3125 | |
3126 | /* | |
10fbcf4c | 3127 | * Register hstate attributes for a single node device. |
9a305230 LS |
3128 | * No-op if attributes already registered. |
3129 | */ | |
3cd8b44f | 3130 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3131 | { |
3132 | struct hstate *h; | |
10fbcf4c | 3133 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3134 | int err; |
3135 | ||
3136 | if (nhs->hugepages_kobj) | |
3137 | return; /* already allocated */ | |
3138 | ||
3139 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3140 | &node->dev.kobj); |
9a305230 LS |
3141 | if (!nhs->hugepages_kobj) |
3142 | return; | |
3143 | ||
3144 | for_each_hstate(h) { | |
3145 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3146 | nhs->hstate_kobjs, | |
3147 | &per_node_hstate_attr_group); | |
3148 | if (err) { | |
282f4214 | 3149 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3150 | h->name, node->dev.id); |
9a305230 LS |
3151 | hugetlb_unregister_node(node); |
3152 | break; | |
3153 | } | |
3154 | } | |
3155 | } | |
3156 | ||
3157 | /* | |
9b5e5d0f | 3158 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3159 | * devices of nodes that have memory. All on-line nodes should have |
3160 | * registered their associated device by this time. | |
9a305230 | 3161 | */ |
7d9ca000 | 3162 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3163 | { |
3164 | int nid; | |
3165 | ||
8cebfcd0 | 3166 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3167 | struct node *node = node_devices[nid]; |
10fbcf4c | 3168 | if (node->dev.id == nid) |
9a305230 LS |
3169 | hugetlb_register_node(node); |
3170 | } | |
3171 | ||
3172 | /* | |
10fbcf4c | 3173 | * Let the node device driver know we're here so it can |
9a305230 LS |
3174 | * [un]register hstate attributes on node hotplug. |
3175 | */ | |
3176 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3177 | hugetlb_unregister_node); | |
3178 | } | |
3179 | #else /* !CONFIG_NUMA */ | |
3180 | ||
3181 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3182 | { | |
3183 | BUG(); | |
3184 | if (nidp) | |
3185 | *nidp = -1; | |
3186 | return NULL; | |
3187 | } | |
3188 | ||
9a305230 LS |
3189 | static void hugetlb_register_all_nodes(void) { } |
3190 | ||
3191 | #endif | |
3192 | ||
a3437870 NA |
3193 | static int __init hugetlb_init(void) |
3194 | { | |
8382d914 DB |
3195 | int i; |
3196 | ||
c2833a5b MK |
3197 | if (!hugepages_supported()) { |
3198 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
3199 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 3200 | return 0; |
c2833a5b | 3201 | } |
a3437870 | 3202 | |
282f4214 MK |
3203 | /* |
3204 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
3205 | * architectures depend on setup being done here. | |
3206 | */ | |
3207 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
3208 | if (!parsed_default_hugepagesz) { | |
3209 | /* | |
3210 | * If we did not parse a default huge page size, set | |
3211 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
3212 | * number of huge pages for this default size was implicitly | |
3213 | * specified, set that here as well. | |
3214 | * Note that the implicit setting will overwrite an explicit | |
3215 | * setting. A warning will be printed in this case. | |
3216 | */ | |
3217 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
3218 | if (default_hstate_max_huge_pages) { | |
3219 | if (default_hstate.max_huge_pages) { | |
3220 | char buf[32]; | |
3221 | ||
3222 | string_get_size(huge_page_size(&default_hstate), | |
3223 | 1, STRING_UNITS_2, buf, 32); | |
3224 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
3225 | default_hstate.max_huge_pages, buf); | |
3226 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
3227 | default_hstate_max_huge_pages); | |
3228 | } | |
3229 | default_hstate.max_huge_pages = | |
3230 | default_hstate_max_huge_pages; | |
d715cf80 | 3231 | } |
f8b74815 | 3232 | } |
a3437870 | 3233 | |
cf11e85f | 3234 | hugetlb_cma_check(); |
a3437870 | 3235 | hugetlb_init_hstates(); |
aa888a74 | 3236 | gather_bootmem_prealloc(); |
a3437870 NA |
3237 | report_hugepages(); |
3238 | ||
3239 | hugetlb_sysfs_init(); | |
9a305230 | 3240 | hugetlb_register_all_nodes(); |
7179e7bf | 3241 | hugetlb_cgroup_file_init(); |
9a305230 | 3242 | |
8382d914 DB |
3243 | #ifdef CONFIG_SMP |
3244 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3245 | #else | |
3246 | num_fault_mutexes = 1; | |
3247 | #endif | |
c672c7f2 | 3248 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3249 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3250 | GFP_KERNEL); | |
c672c7f2 | 3251 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3252 | |
3253 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3254 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3255 | return 0; |
3256 | } | |
3e89e1c5 | 3257 | subsys_initcall(hugetlb_init); |
a3437870 | 3258 | |
ae94da89 MK |
3259 | /* Overwritten by architectures with more huge page sizes */ |
3260 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 3261 | { |
ae94da89 | 3262 | return size == HPAGE_SIZE; |
9fee021d VT |
3263 | } |
3264 | ||
d00181b9 | 3265 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3266 | { |
3267 | struct hstate *h; | |
8faa8b07 AK |
3268 | unsigned long i; |
3269 | ||
a3437870 | 3270 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
3271 | return; |
3272 | } | |
47d38344 | 3273 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3274 | BUG_ON(order == 0); |
47d38344 | 3275 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
3276 | h->order = order; |
3277 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
3278 | h->nr_huge_pages = 0; |
3279 | h->free_huge_pages = 0; | |
3280 | for (i = 0; i < MAX_NUMNODES; ++i) | |
3281 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3282 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3283 | h->next_nid_to_alloc = first_memory_node; |
3284 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3285 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3286 | huge_page_size(h)/1024); | |
8faa8b07 | 3287 | |
a3437870 NA |
3288 | parsed_hstate = h; |
3289 | } | |
3290 | ||
282f4214 MK |
3291 | /* |
3292 | * hugepages command line processing | |
3293 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
3294 | * specification. If not, ignore the hugepages value. hugepages can also | |
3295 | * be the first huge page command line option in which case it implicitly | |
3296 | * specifies the number of huge pages for the default size. | |
3297 | */ | |
3298 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
3299 | { |
3300 | unsigned long *mhp; | |
8faa8b07 | 3301 | static unsigned long *last_mhp; |
a3437870 | 3302 | |
9fee021d | 3303 | if (!parsed_valid_hugepagesz) { |
282f4214 | 3304 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 3305 | parsed_valid_hugepagesz = true; |
282f4214 | 3306 | return 0; |
9fee021d | 3307 | } |
282f4214 | 3308 | |
a3437870 | 3309 | /* |
282f4214 MK |
3310 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
3311 | * yet, so this hugepages= parameter goes to the "default hstate". | |
3312 | * Otherwise, it goes with the previously parsed hugepagesz or | |
3313 | * default_hugepagesz. | |
a3437870 | 3314 | */ |
9fee021d | 3315 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3316 | mhp = &default_hstate_max_huge_pages; |
3317 | else | |
3318 | mhp = &parsed_hstate->max_huge_pages; | |
3319 | ||
8faa8b07 | 3320 | if (mhp == last_mhp) { |
282f4214 MK |
3321 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
3322 | return 0; | |
8faa8b07 AK |
3323 | } |
3324 | ||
a3437870 NA |
3325 | if (sscanf(s, "%lu", mhp) <= 0) |
3326 | *mhp = 0; | |
3327 | ||
8faa8b07 AK |
3328 | /* |
3329 | * Global state is always initialized later in hugetlb_init. | |
3330 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
3331 | * use the bootmem allocator. | |
3332 | */ | |
47d38344 | 3333 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
3334 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3335 | ||
3336 | last_mhp = mhp; | |
3337 | ||
a3437870 NA |
3338 | return 1; |
3339 | } | |
282f4214 | 3340 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 3341 | |
282f4214 MK |
3342 | /* |
3343 | * hugepagesz command line processing | |
3344 | * A specific huge page size can only be specified once with hugepagesz. | |
3345 | * hugepagesz is followed by hugepages on the command line. The global | |
3346 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
3347 | * hugepagesz argument was valid. | |
3348 | */ | |
359f2544 | 3349 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 3350 | { |
359f2544 | 3351 | unsigned long size; |
282f4214 MK |
3352 | struct hstate *h; |
3353 | ||
3354 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
3355 | size = (unsigned long)memparse(s, NULL); |
3356 | ||
3357 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3358 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
359f2544 MK |
3359 | return 0; |
3360 | } | |
3361 | ||
282f4214 MK |
3362 | h = size_to_hstate(size); |
3363 | if (h) { | |
3364 | /* | |
3365 | * hstate for this size already exists. This is normally | |
3366 | * an error, but is allowed if the existing hstate is the | |
3367 | * default hstate. More specifically, it is only allowed if | |
3368 | * the number of huge pages for the default hstate was not | |
3369 | * previously specified. | |
3370 | */ | |
3371 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
3372 | default_hstate.max_huge_pages) { | |
3373 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
3374 | return 0; | |
3375 | } | |
3376 | ||
3377 | /* | |
3378 | * No need to call hugetlb_add_hstate() as hstate already | |
3379 | * exists. But, do set parsed_hstate so that a following | |
3380 | * hugepages= parameter will be applied to this hstate. | |
3381 | */ | |
3382 | parsed_hstate = h; | |
3383 | parsed_valid_hugepagesz = true; | |
3384 | return 1; | |
38237830 MK |
3385 | } |
3386 | ||
359f2544 | 3387 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 3388 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
3389 | return 1; |
3390 | } | |
359f2544 MK |
3391 | __setup("hugepagesz=", hugepagesz_setup); |
3392 | ||
282f4214 MK |
3393 | /* |
3394 | * default_hugepagesz command line input | |
3395 | * Only one instance of default_hugepagesz allowed on command line. | |
3396 | */ | |
ae94da89 | 3397 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 3398 | { |
ae94da89 MK |
3399 | unsigned long size; |
3400 | ||
282f4214 | 3401 | parsed_valid_hugepagesz = false; |
282f4214 MK |
3402 | if (parsed_default_hugepagesz) { |
3403 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
3404 | return 0; | |
3405 | } | |
3406 | ||
ae94da89 MK |
3407 | size = (unsigned long)memparse(s, NULL); |
3408 | ||
3409 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3410 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
ae94da89 MK |
3411 | return 0; |
3412 | } | |
3413 | ||
282f4214 MK |
3414 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
3415 | parsed_valid_hugepagesz = true; | |
3416 | parsed_default_hugepagesz = true; | |
3417 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
3418 | ||
3419 | /* | |
3420 | * The number of default huge pages (for this size) could have been | |
3421 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
3422 | * then default_hstate_max_huge_pages is set. If the default huge | |
3423 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
3424 | * allocated here from bootmem allocator. | |
3425 | */ | |
3426 | if (default_hstate_max_huge_pages) { | |
3427 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3428 | if (hstate_is_gigantic(&default_hstate)) | |
3429 | hugetlb_hstate_alloc_pages(&default_hstate); | |
3430 | default_hstate_max_huge_pages = 0; | |
3431 | } | |
3432 | ||
e11bfbfc NP |
3433 | return 1; |
3434 | } | |
ae94da89 | 3435 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 3436 | |
8ca39e68 | 3437 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
3438 | { |
3439 | int node; | |
3440 | unsigned int nr = 0; | |
8ca39e68 MS |
3441 | nodemask_t *mpol_allowed; |
3442 | unsigned int *array = h->free_huge_pages_node; | |
3443 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
3444 | ||
3445 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 3446 | |
8ca39e68 MS |
3447 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
3448 | if (!mpol_allowed || | |
3449 | (mpol_allowed && node_isset(node, *mpol_allowed))) | |
3450 | nr += array[node]; | |
3451 | } | |
8a213460 NA |
3452 | |
3453 | return nr; | |
3454 | } | |
3455 | ||
3456 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
3457 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3458 | struct ctl_table *table, int write, | |
32927393 | 3459 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 3460 | { |
e5ff2159 | 3461 | struct hstate *h = &default_hstate; |
238d3c13 | 3462 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3463 | int ret; |
e5ff2159 | 3464 | |
457c1b27 | 3465 | if (!hugepages_supported()) |
86613628 | 3466 | return -EOPNOTSUPP; |
457c1b27 | 3467 | |
e5ff2159 AK |
3468 | table->data = &tmp; |
3469 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3470 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3471 | if (ret) | |
3472 | goto out; | |
e5ff2159 | 3473 | |
238d3c13 DR |
3474 | if (write) |
3475 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3476 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3477 | out: |
3478 | return ret; | |
1da177e4 | 3479 | } |
396faf03 | 3480 | |
06808b08 | 3481 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 3482 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3483 | { |
3484 | ||
3485 | return hugetlb_sysctl_handler_common(false, table, write, | |
3486 | buffer, length, ppos); | |
3487 | } | |
3488 | ||
3489 | #ifdef CONFIG_NUMA | |
3490 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 3491 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3492 | { |
3493 | return hugetlb_sysctl_handler_common(true, table, write, | |
3494 | buffer, length, ppos); | |
3495 | } | |
3496 | #endif /* CONFIG_NUMA */ | |
3497 | ||
a3d0c6aa | 3498 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 3499 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 3500 | { |
a5516438 | 3501 | struct hstate *h = &default_hstate; |
e5ff2159 | 3502 | unsigned long tmp; |
08d4a246 | 3503 | int ret; |
e5ff2159 | 3504 | |
457c1b27 | 3505 | if (!hugepages_supported()) |
86613628 | 3506 | return -EOPNOTSUPP; |
457c1b27 | 3507 | |
c033a93c | 3508 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3509 | |
bae7f4ae | 3510 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3511 | return -EINVAL; |
3512 | ||
e5ff2159 AK |
3513 | table->data = &tmp; |
3514 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3515 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3516 | if (ret) | |
3517 | goto out; | |
e5ff2159 AK |
3518 | |
3519 | if (write) { | |
3520 | spin_lock(&hugetlb_lock); | |
3521 | h->nr_overcommit_huge_pages = tmp; | |
3522 | spin_unlock(&hugetlb_lock); | |
3523 | } | |
08d4a246 MH |
3524 | out: |
3525 | return ret; | |
a3d0c6aa NA |
3526 | } |
3527 | ||
1da177e4 LT |
3528 | #endif /* CONFIG_SYSCTL */ |
3529 | ||
e1759c21 | 3530 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3531 | { |
fcb2b0c5 RG |
3532 | struct hstate *h; |
3533 | unsigned long total = 0; | |
3534 | ||
457c1b27 NA |
3535 | if (!hugepages_supported()) |
3536 | return; | |
fcb2b0c5 RG |
3537 | |
3538 | for_each_hstate(h) { | |
3539 | unsigned long count = h->nr_huge_pages; | |
3540 | ||
3541 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3542 | ||
3543 | if (h == &default_hstate) | |
3544 | seq_printf(m, | |
3545 | "HugePages_Total: %5lu\n" | |
3546 | "HugePages_Free: %5lu\n" | |
3547 | "HugePages_Rsvd: %5lu\n" | |
3548 | "HugePages_Surp: %5lu\n" | |
3549 | "Hugepagesize: %8lu kB\n", | |
3550 | count, | |
3551 | h->free_huge_pages, | |
3552 | h->resv_huge_pages, | |
3553 | h->surplus_huge_pages, | |
3554 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3555 | } | |
3556 | ||
3557 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3558 | } |
3559 | ||
3560 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3561 | { | |
a5516438 | 3562 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3563 | if (!hugepages_supported()) |
3564 | return 0; | |
1da177e4 LT |
3565 | return sprintf(buf, |
3566 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3567 | "Node %d HugePages_Free: %5u\n" |
3568 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3569 | nid, h->nr_huge_pages_node[nid], |
3570 | nid, h->free_huge_pages_node[nid], | |
3571 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3572 | } |
3573 | ||
949f7ec5 DR |
3574 | void hugetlb_show_meminfo(void) |
3575 | { | |
3576 | struct hstate *h; | |
3577 | int nid; | |
3578 | ||
457c1b27 NA |
3579 | if (!hugepages_supported()) |
3580 | return; | |
3581 | ||
949f7ec5 DR |
3582 | for_each_node_state(nid, N_MEMORY) |
3583 | for_each_hstate(h) | |
3584 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3585 | nid, | |
3586 | h->nr_huge_pages_node[nid], | |
3587 | h->free_huge_pages_node[nid], | |
3588 | h->surplus_huge_pages_node[nid], | |
3589 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3590 | } | |
3591 | ||
5d317b2b NH |
3592 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3593 | { | |
3594 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3595 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3596 | } | |
3597 | ||
1da177e4 LT |
3598 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3599 | unsigned long hugetlb_total_pages(void) | |
3600 | { | |
d0028588 WL |
3601 | struct hstate *h; |
3602 | unsigned long nr_total_pages = 0; | |
3603 | ||
3604 | for_each_hstate(h) | |
3605 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3606 | return nr_total_pages; | |
1da177e4 | 3607 | } |
1da177e4 | 3608 | |
a5516438 | 3609 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3610 | { |
3611 | int ret = -ENOMEM; | |
3612 | ||
3613 | spin_lock(&hugetlb_lock); | |
3614 | /* | |
3615 | * When cpuset is configured, it breaks the strict hugetlb page | |
3616 | * reservation as the accounting is done on a global variable. Such | |
3617 | * reservation is completely rubbish in the presence of cpuset because | |
3618 | * the reservation is not checked against page availability for the | |
3619 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3620 | * with lack of free htlb page in cpuset that the task is in. | |
3621 | * Attempt to enforce strict accounting with cpuset is almost | |
3622 | * impossible (or too ugly) because cpuset is too fluid that | |
3623 | * task or memory node can be dynamically moved between cpusets. | |
3624 | * | |
3625 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3626 | * undesirable. However, in order to preserve some of the semantics, | |
3627 | * we fall back to check against current free page availability as | |
3628 | * a best attempt and hopefully to minimize the impact of changing | |
3629 | * semantics that cpuset has. | |
8ca39e68 MS |
3630 | * |
3631 | * Apart from cpuset, we also have memory policy mechanism that | |
3632 | * also determines from which node the kernel will allocate memory | |
3633 | * in a NUMA system. So similar to cpuset, we also should consider | |
3634 | * the memory policy of the current task. Similar to the description | |
3635 | * above. | |
fc1b8a73 MG |
3636 | */ |
3637 | if (delta > 0) { | |
a5516438 | 3638 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3639 | goto out; |
3640 | ||
8ca39e68 | 3641 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 3642 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
3643 | goto out; |
3644 | } | |
3645 | } | |
3646 | ||
3647 | ret = 0; | |
3648 | if (delta < 0) | |
a5516438 | 3649 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3650 | |
3651 | out: | |
3652 | spin_unlock(&hugetlb_lock); | |
3653 | return ret; | |
3654 | } | |
3655 | ||
84afd99b AW |
3656 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3657 | { | |
f522c3ac | 3658 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3659 | |
3660 | /* | |
3661 | * This new VMA should share its siblings reservation map if present. | |
3662 | * The VMA will only ever have a valid reservation map pointer where | |
3663 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3664 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3665 | * after this open call completes. It is therefore safe to take a |
3666 | * new reference here without additional locking. | |
3667 | */ | |
4e35f483 | 3668 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3669 | kref_get(&resv->refs); |
84afd99b AW |
3670 | } |
3671 | ||
a1e78772 MG |
3672 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3673 | { | |
a5516438 | 3674 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3675 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3676 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3677 | unsigned long reserve, start, end; |
1c5ecae3 | 3678 | long gbl_reserve; |
84afd99b | 3679 | |
4e35f483 JK |
3680 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3681 | return; | |
84afd99b | 3682 | |
4e35f483 JK |
3683 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3684 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3685 | |
4e35f483 | 3686 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 3687 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 3688 | if (reserve) { |
1c5ecae3 MK |
3689 | /* |
3690 | * Decrement reserve counts. The global reserve count may be | |
3691 | * adjusted if the subpool has a minimum size. | |
3692 | */ | |
3693 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3694 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3695 | } |
e9fe92ae MA |
3696 | |
3697 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
3698 | } |
3699 | ||
31383c68 DW |
3700 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3701 | { | |
3702 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3703 | return -EINVAL; | |
3704 | return 0; | |
3705 | } | |
3706 | ||
05ea8860 DW |
3707 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3708 | { | |
3709 | struct hstate *hstate = hstate_vma(vma); | |
3710 | ||
3711 | return 1UL << huge_page_shift(hstate); | |
3712 | } | |
3713 | ||
1da177e4 LT |
3714 | /* |
3715 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3716 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3717 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3718 | * this far. | |
3719 | */ | |
b3ec9f33 | 3720 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3721 | { |
3722 | BUG(); | |
d0217ac0 | 3723 | return 0; |
1da177e4 LT |
3724 | } |
3725 | ||
eec3636a JC |
3726 | /* |
3727 | * When a new function is introduced to vm_operations_struct and added | |
3728 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3729 | * This is because under System V memory model, mappings created via | |
3730 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3731 | * their original vm_ops are overwritten with shm_vm_ops. | |
3732 | */ | |
f0f37e2f | 3733 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3734 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3735 | .open = hugetlb_vm_op_open, |
a1e78772 | 3736 | .close = hugetlb_vm_op_close, |
31383c68 | 3737 | .split = hugetlb_vm_op_split, |
05ea8860 | 3738 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3739 | }; |
3740 | ||
1e8f889b DG |
3741 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3742 | int writable) | |
63551ae0 DG |
3743 | { |
3744 | pte_t entry; | |
3745 | ||
1e8f889b | 3746 | if (writable) { |
106c992a GS |
3747 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3748 | vma->vm_page_prot))); | |
63551ae0 | 3749 | } else { |
106c992a GS |
3750 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3751 | vma->vm_page_prot)); | |
63551ae0 DG |
3752 | } |
3753 | entry = pte_mkyoung(entry); | |
3754 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3755 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3756 | |
3757 | return entry; | |
3758 | } | |
3759 | ||
1e8f889b DG |
3760 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3761 | unsigned long address, pte_t *ptep) | |
3762 | { | |
3763 | pte_t entry; | |
3764 | ||
106c992a | 3765 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3766 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3767 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3768 | } |
3769 | ||
d5ed7444 | 3770 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3771 | { |
3772 | swp_entry_t swp; | |
3773 | ||
3774 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3775 | return false; |
4a705fef NH |
3776 | swp = pte_to_swp_entry(pte); |
3777 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3778 | return true; |
4a705fef | 3779 | else |
d5ed7444 | 3780 | return false; |
4a705fef NH |
3781 | } |
3782 | ||
3783 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3784 | { | |
3785 | swp_entry_t swp; | |
3786 | ||
3787 | if (huge_pte_none(pte) || pte_present(pte)) | |
3788 | return 0; | |
3789 | swp = pte_to_swp_entry(pte); | |
3790 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3791 | return 1; | |
3792 | else | |
3793 | return 0; | |
3794 | } | |
1e8f889b | 3795 | |
63551ae0 DG |
3796 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3797 | struct vm_area_struct *vma) | |
3798 | { | |
5e41540c | 3799 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3800 | struct page *ptepage; |
1c59827d | 3801 | unsigned long addr; |
1e8f889b | 3802 | int cow; |
a5516438 AK |
3803 | struct hstate *h = hstate_vma(vma); |
3804 | unsigned long sz = huge_page_size(h); | |
c0d0381a | 3805 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 3806 | struct mmu_notifier_range range; |
e8569dd2 | 3807 | int ret = 0; |
1e8f889b DG |
3808 | |
3809 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3810 | |
ac46d4f3 | 3811 | if (cow) { |
7269f999 | 3812 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3813 | vma->vm_start, |
ac46d4f3 JG |
3814 | vma->vm_end); |
3815 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
3816 | } else { |
3817 | /* | |
3818 | * For shared mappings i_mmap_rwsem must be held to call | |
3819 | * huge_pte_alloc, otherwise the returned ptep could go | |
3820 | * away if part of a shared pmd and another thread calls | |
3821 | * huge_pmd_unshare. | |
3822 | */ | |
3823 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 3824 | } |
e8569dd2 | 3825 | |
a5516438 | 3826 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3827 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3828 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3829 | if (!src_pte) |
3830 | continue; | |
a5516438 | 3831 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3832 | if (!dst_pte) { |
3833 | ret = -ENOMEM; | |
3834 | break; | |
3835 | } | |
c5c99429 | 3836 | |
5e41540c MK |
3837 | /* |
3838 | * If the pagetables are shared don't copy or take references. | |
3839 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3840 | * | |
3841 | * However, src could have 'unshared' and dst shares with | |
3842 | * another vma. If dst_pte !none, this implies sharing. | |
3843 | * Check here before taking page table lock, and once again | |
3844 | * after taking the lock below. | |
3845 | */ | |
3846 | dst_entry = huge_ptep_get(dst_pte); | |
3847 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3848 | continue; |
3849 | ||
cb900f41 KS |
3850 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3851 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3852 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3853 | entry = huge_ptep_get(src_pte); |
5e41540c MK |
3854 | dst_entry = huge_ptep_get(dst_pte); |
3855 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3856 | /* | |
3857 | * Skip if src entry none. Also, skip in the | |
3858 | * unlikely case dst entry !none as this implies | |
3859 | * sharing with another vma. | |
3860 | */ | |
4a705fef NH |
3861 | ; |
3862 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3863 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3864 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3865 | ||
3866 | if (is_write_migration_entry(swp_entry) && cow) { | |
3867 | /* | |
3868 | * COW mappings require pages in both | |
3869 | * parent and child to be set to read. | |
3870 | */ | |
3871 | make_migration_entry_read(&swp_entry); | |
3872 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3873 | set_huge_swap_pte_at(src, addr, src_pte, |
3874 | entry, sz); | |
4a705fef | 3875 | } |
e5251fd4 | 3876 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3877 | } else { |
34ee645e | 3878 | if (cow) { |
0f10851e JG |
3879 | /* |
3880 | * No need to notify as we are downgrading page | |
3881 | * table protection not changing it to point | |
3882 | * to a new page. | |
3883 | * | |
ad56b738 | 3884 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3885 | */ |
7f2e9525 | 3886 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3887 | } |
0253d634 | 3888 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3889 | ptepage = pte_page(entry); |
3890 | get_page(ptepage); | |
53f9263b | 3891 | page_dup_rmap(ptepage, true); |
1c59827d | 3892 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3893 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3894 | } |
cb900f41 KS |
3895 | spin_unlock(src_ptl); |
3896 | spin_unlock(dst_ptl); | |
63551ae0 | 3897 | } |
63551ae0 | 3898 | |
e8569dd2 | 3899 | if (cow) |
ac46d4f3 | 3900 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
3901 | else |
3902 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
3903 | |
3904 | return ret; | |
63551ae0 DG |
3905 | } |
3906 | ||
24669e58 AK |
3907 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3908 | unsigned long start, unsigned long end, | |
3909 | struct page *ref_page) | |
63551ae0 DG |
3910 | { |
3911 | struct mm_struct *mm = vma->vm_mm; | |
3912 | unsigned long address; | |
c7546f8f | 3913 | pte_t *ptep; |
63551ae0 | 3914 | pte_t pte; |
cb900f41 | 3915 | spinlock_t *ptl; |
63551ae0 | 3916 | struct page *page; |
a5516438 AK |
3917 | struct hstate *h = hstate_vma(vma); |
3918 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3919 | struct mmu_notifier_range range; |
a5516438 | 3920 | |
63551ae0 | 3921 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3922 | BUG_ON(start & ~huge_page_mask(h)); |
3923 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3924 | |
07e32661 AK |
3925 | /* |
3926 | * This is a hugetlb vma, all the pte entries should point | |
3927 | * to huge page. | |
3928 | */ | |
ed6a7935 | 3929 | tlb_change_page_size(tlb, sz); |
24669e58 | 3930 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3931 | |
3932 | /* | |
3933 | * If sharing possible, alert mmu notifiers of worst case. | |
3934 | */ | |
6f4f13e8 JG |
3935 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3936 | end); | |
ac46d4f3 JG |
3937 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3938 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3939 | address = start; |
569f48b8 | 3940 | for (; address < end; address += sz) { |
7868a208 | 3941 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3942 | if (!ptep) |
c7546f8f DG |
3943 | continue; |
3944 | ||
cb900f41 | 3945 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 3946 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
31d49da5 | 3947 | spin_unlock(ptl); |
dff11abe MK |
3948 | /* |
3949 | * We just unmapped a page of PMDs by clearing a PUD. | |
3950 | * The caller's TLB flush range should cover this area. | |
3951 | */ | |
31d49da5 AK |
3952 | continue; |
3953 | } | |
39dde65c | 3954 | |
6629326b | 3955 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3956 | if (huge_pte_none(pte)) { |
3957 | spin_unlock(ptl); | |
3958 | continue; | |
3959 | } | |
6629326b HD |
3960 | |
3961 | /* | |
9fbc1f63 NH |
3962 | * Migrating hugepage or HWPoisoned hugepage is already |
3963 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3964 | */ |
9fbc1f63 | 3965 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3966 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3967 | spin_unlock(ptl); |
3968 | continue; | |
8c4894c6 | 3969 | } |
6629326b HD |
3970 | |
3971 | page = pte_page(pte); | |
04f2cbe3 MG |
3972 | /* |
3973 | * If a reference page is supplied, it is because a specific | |
3974 | * page is being unmapped, not a range. Ensure the page we | |
3975 | * are about to unmap is the actual page of interest. | |
3976 | */ | |
3977 | if (ref_page) { | |
31d49da5 AK |
3978 | if (page != ref_page) { |
3979 | spin_unlock(ptl); | |
3980 | continue; | |
3981 | } | |
04f2cbe3 MG |
3982 | /* |
3983 | * Mark the VMA as having unmapped its page so that | |
3984 | * future faults in this VMA will fail rather than | |
3985 | * looking like data was lost | |
3986 | */ | |
3987 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3988 | } | |
3989 | ||
c7546f8f | 3990 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3991 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3992 | if (huge_pte_dirty(pte)) |
6649a386 | 3993 | set_page_dirty(page); |
9e81130b | 3994 | |
5d317b2b | 3995 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3996 | page_remove_rmap(page, true); |
31d49da5 | 3997 | |
cb900f41 | 3998 | spin_unlock(ptl); |
e77b0852 | 3999 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
4000 | /* |
4001 | * Bail out after unmapping reference page if supplied | |
4002 | */ | |
4003 | if (ref_page) | |
4004 | break; | |
fe1668ae | 4005 | } |
ac46d4f3 | 4006 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 4007 | tlb_end_vma(tlb, vma); |
1da177e4 | 4008 | } |
63551ae0 | 4009 | |
d833352a MG |
4010 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
4011 | struct vm_area_struct *vma, unsigned long start, | |
4012 | unsigned long end, struct page *ref_page) | |
4013 | { | |
4014 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
4015 | ||
4016 | /* | |
4017 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
4018 | * test will fail on a vma being torn down, and not grab a page table | |
4019 | * on its way out. We're lucky that the flag has such an appropriate | |
4020 | * name, and can in fact be safely cleared here. We could clear it | |
4021 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 4022 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 4023 | * because in the context this is called, the VMA is about to be |
c8c06efa | 4024 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
4025 | */ |
4026 | vma->vm_flags &= ~VM_MAYSHARE; | |
4027 | } | |
4028 | ||
502717f4 | 4029 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 4030 | unsigned long end, struct page *ref_page) |
502717f4 | 4031 | { |
24669e58 AK |
4032 | struct mm_struct *mm; |
4033 | struct mmu_gather tlb; | |
dff11abe MK |
4034 | unsigned long tlb_start = start; |
4035 | unsigned long tlb_end = end; | |
4036 | ||
4037 | /* | |
4038 | * If shared PMDs were possibly used within this vma range, adjust | |
4039 | * start/end for worst case tlb flushing. | |
4040 | * Note that we can not be sure if PMDs are shared until we try to | |
4041 | * unmap pages. However, we want to make sure TLB flushing covers | |
4042 | * the largest possible range. | |
4043 | */ | |
4044 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
24669e58 AK |
4045 | |
4046 | mm = vma->vm_mm; | |
4047 | ||
dff11abe | 4048 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); |
24669e58 | 4049 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
dff11abe | 4050 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); |
502717f4 CK |
4051 | } |
4052 | ||
04f2cbe3 MG |
4053 | /* |
4054 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
4055 | * mappping it owns the reserve page for. The intention is to unmap the page | |
4056 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
4057 | * same region. | |
4058 | */ | |
2f4612af DB |
4059 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
4060 | struct page *page, unsigned long address) | |
04f2cbe3 | 4061 | { |
7526674d | 4062 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
4063 | struct vm_area_struct *iter_vma; |
4064 | struct address_space *mapping; | |
04f2cbe3 MG |
4065 | pgoff_t pgoff; |
4066 | ||
4067 | /* | |
4068 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
4069 | * from page cache lookup which is in HPAGE_SIZE units. | |
4070 | */ | |
7526674d | 4071 | address = address & huge_page_mask(h); |
36e4f20a MH |
4072 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
4073 | vma->vm_pgoff; | |
93c76a3d | 4074 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 4075 | |
4eb2b1dc MG |
4076 | /* |
4077 | * Take the mapping lock for the duration of the table walk. As | |
4078 | * this mapping should be shared between all the VMAs, | |
4079 | * __unmap_hugepage_range() is called as the lock is already held | |
4080 | */ | |
83cde9e8 | 4081 | i_mmap_lock_write(mapping); |
6b2dbba8 | 4082 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
4083 | /* Do not unmap the current VMA */ |
4084 | if (iter_vma == vma) | |
4085 | continue; | |
4086 | ||
2f84a899 MG |
4087 | /* |
4088 | * Shared VMAs have their own reserves and do not affect | |
4089 | * MAP_PRIVATE accounting but it is possible that a shared | |
4090 | * VMA is using the same page so check and skip such VMAs. | |
4091 | */ | |
4092 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
4093 | continue; | |
4094 | ||
04f2cbe3 MG |
4095 | /* |
4096 | * Unmap the page from other VMAs without their own reserves. | |
4097 | * They get marked to be SIGKILLed if they fault in these | |
4098 | * areas. This is because a future no-page fault on this VMA | |
4099 | * could insert a zeroed page instead of the data existing | |
4100 | * from the time of fork. This would look like data corruption | |
4101 | */ | |
4102 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
4103 | unmap_hugepage_range(iter_vma, address, |
4104 | address + huge_page_size(h), page); | |
04f2cbe3 | 4105 | } |
83cde9e8 | 4106 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
4107 | } |
4108 | ||
0fe6e20b NH |
4109 | /* |
4110 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
4111 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
4112 | * cannot race with other handlers or page migration. | |
4113 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 4114 | */ |
2b740303 | 4115 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 4116 | unsigned long address, pte_t *ptep, |
3999f52e | 4117 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 4118 | { |
3999f52e | 4119 | pte_t pte; |
a5516438 | 4120 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 4121 | struct page *old_page, *new_page; |
2b740303 SJ |
4122 | int outside_reserve = 0; |
4123 | vm_fault_t ret = 0; | |
974e6d66 | 4124 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 4125 | struct mmu_notifier_range range; |
1e8f889b | 4126 | |
3999f52e | 4127 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
4128 | old_page = pte_page(pte); |
4129 | ||
04f2cbe3 | 4130 | retry_avoidcopy: |
1e8f889b DG |
4131 | /* If no-one else is actually using this page, avoid the copy |
4132 | * and just make the page writable */ | |
37a2140d | 4133 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 4134 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 4135 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 4136 | return 0; |
1e8f889b DG |
4137 | } |
4138 | ||
04f2cbe3 MG |
4139 | /* |
4140 | * If the process that created a MAP_PRIVATE mapping is about to | |
4141 | * perform a COW due to a shared page count, attempt to satisfy | |
4142 | * the allocation without using the existing reserves. The pagecache | |
4143 | * page is used to determine if the reserve at this address was | |
4144 | * consumed or not. If reserves were used, a partial faulted mapping | |
4145 | * at the time of fork() could consume its reserves on COW instead | |
4146 | * of the full address range. | |
4147 | */ | |
5944d011 | 4148 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
4149 | old_page != pagecache_page) |
4150 | outside_reserve = 1; | |
4151 | ||
09cbfeaf | 4152 | get_page(old_page); |
b76c8cfb | 4153 | |
ad4404a2 DB |
4154 | /* |
4155 | * Drop page table lock as buddy allocator may be called. It will | |
4156 | * be acquired again before returning to the caller, as expected. | |
4157 | */ | |
cb900f41 | 4158 | spin_unlock(ptl); |
5b7a1d40 | 4159 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 4160 | |
2fc39cec | 4161 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
4162 | /* |
4163 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4164 | * it is due to references held by a child and an insufficient | |
4165 | * huge page pool. To guarantee the original mappers | |
4166 | * reliability, unmap the page from child processes. The child | |
4167 | * may get SIGKILLed if it later faults. | |
4168 | */ | |
4169 | if (outside_reserve) { | |
09cbfeaf | 4170 | put_page(old_page); |
04f2cbe3 | 4171 | BUG_ON(huge_pte_none(pte)); |
5b7a1d40 | 4172 | unmap_ref_private(mm, vma, old_page, haddr); |
2f4612af DB |
4173 | BUG_ON(huge_pte_none(pte)); |
4174 | spin_lock(ptl); | |
5b7a1d40 | 4175 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
4176 | if (likely(ptep && |
4177 | pte_same(huge_ptep_get(ptep), pte))) | |
4178 | goto retry_avoidcopy; | |
4179 | /* | |
4180 | * race occurs while re-acquiring page table | |
4181 | * lock, and our job is done. | |
4182 | */ | |
4183 | return 0; | |
04f2cbe3 MG |
4184 | } |
4185 | ||
2b740303 | 4186 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 4187 | goto out_release_old; |
1e8f889b DG |
4188 | } |
4189 | ||
0fe6e20b NH |
4190 | /* |
4191 | * When the original hugepage is shared one, it does not have | |
4192 | * anon_vma prepared. | |
4193 | */ | |
44e2aa93 | 4194 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
4195 | ret = VM_FAULT_OOM; |
4196 | goto out_release_all; | |
44e2aa93 | 4197 | } |
0fe6e20b | 4198 | |
974e6d66 | 4199 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 4200 | pages_per_huge_page(h)); |
0ed361de | 4201 | __SetPageUptodate(new_page); |
1e8f889b | 4202 | |
7269f999 | 4203 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 4204 | haddr + huge_page_size(h)); |
ac46d4f3 | 4205 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 4206 | |
b76c8cfb | 4207 | /* |
cb900f41 | 4208 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
4209 | * before the page tables are altered |
4210 | */ | |
cb900f41 | 4211 | spin_lock(ptl); |
5b7a1d40 | 4212 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 4213 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
4214 | ClearPagePrivate(new_page); |
4215 | ||
1e8f889b | 4216 | /* Break COW */ |
5b7a1d40 | 4217 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 4218 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 4219 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 4220 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 4221 | page_remove_rmap(old_page, true); |
5b7a1d40 | 4222 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
cb6acd01 | 4223 | set_page_huge_active(new_page); |
1e8f889b DG |
4224 | /* Make the old page be freed below */ |
4225 | new_page = old_page; | |
4226 | } | |
cb900f41 | 4227 | spin_unlock(ptl); |
ac46d4f3 | 4228 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 4229 | out_release_all: |
5b7a1d40 | 4230 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 4231 | put_page(new_page); |
ad4404a2 | 4232 | out_release_old: |
09cbfeaf | 4233 | put_page(old_page); |
8312034f | 4234 | |
ad4404a2 DB |
4235 | spin_lock(ptl); /* Caller expects lock to be held */ |
4236 | return ret; | |
1e8f889b DG |
4237 | } |
4238 | ||
04f2cbe3 | 4239 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
4240 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
4241 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
4242 | { |
4243 | struct address_space *mapping; | |
e7c4b0bf | 4244 | pgoff_t idx; |
04f2cbe3 MG |
4245 | |
4246 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 4247 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
4248 | |
4249 | return find_lock_page(mapping, idx); | |
4250 | } | |
4251 | ||
3ae77f43 HD |
4252 | /* |
4253 | * Return whether there is a pagecache page to back given address within VMA. | |
4254 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4255 | */ | |
4256 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
4257 | struct vm_area_struct *vma, unsigned long address) |
4258 | { | |
4259 | struct address_space *mapping; | |
4260 | pgoff_t idx; | |
4261 | struct page *page; | |
4262 | ||
4263 | mapping = vma->vm_file->f_mapping; | |
4264 | idx = vma_hugecache_offset(h, vma, address); | |
4265 | ||
4266 | page = find_get_page(mapping, idx); | |
4267 | if (page) | |
4268 | put_page(page); | |
4269 | return page != NULL; | |
4270 | } | |
4271 | ||
ab76ad54 MK |
4272 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
4273 | pgoff_t idx) | |
4274 | { | |
4275 | struct inode *inode = mapping->host; | |
4276 | struct hstate *h = hstate_inode(inode); | |
4277 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4278 | ||
4279 | if (err) | |
4280 | return err; | |
4281 | ClearPagePrivate(page); | |
4282 | ||
22146c3c MK |
4283 | /* |
4284 | * set page dirty so that it will not be removed from cache/file | |
4285 | * by non-hugetlbfs specific code paths. | |
4286 | */ | |
4287 | set_page_dirty(page); | |
4288 | ||
ab76ad54 MK |
4289 | spin_lock(&inode->i_lock); |
4290 | inode->i_blocks += blocks_per_huge_page(h); | |
4291 | spin_unlock(&inode->i_lock); | |
4292 | return 0; | |
4293 | } | |
4294 | ||
2b740303 SJ |
4295 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
4296 | struct vm_area_struct *vma, | |
4297 | struct address_space *mapping, pgoff_t idx, | |
4298 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 4299 | { |
a5516438 | 4300 | struct hstate *h = hstate_vma(vma); |
2b740303 | 4301 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 4302 | int anon_rmap = 0; |
4c887265 | 4303 | unsigned long size; |
4c887265 | 4304 | struct page *page; |
1e8f889b | 4305 | pte_t new_pte; |
cb900f41 | 4306 | spinlock_t *ptl; |
285b8dca | 4307 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 4308 | bool new_page = false; |
4c887265 | 4309 | |
04f2cbe3 MG |
4310 | /* |
4311 | * Currently, we are forced to kill the process in the event the | |
4312 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4313 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4314 | */ |
4315 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4316 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4317 | current->pid); |
04f2cbe3 MG |
4318 | return ret; |
4319 | } | |
4320 | ||
4c887265 | 4321 | /* |
87bf91d3 MK |
4322 | * We can not race with truncation due to holding i_mmap_rwsem. |
4323 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4324 | * once for faults beyond end of file. | |
4c887265 | 4325 | */ |
87bf91d3 MK |
4326 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4327 | if (idx >= size) | |
4328 | goto out; | |
4329 | ||
6bda666a CL |
4330 | retry: |
4331 | page = find_lock_page(mapping, idx); | |
4332 | if (!page) { | |
1a1aad8a MK |
4333 | /* |
4334 | * Check for page in userfault range | |
4335 | */ | |
4336 | if (userfaultfd_missing(vma)) { | |
4337 | u32 hash; | |
4338 | struct vm_fault vmf = { | |
4339 | .vma = vma, | |
285b8dca | 4340 | .address = haddr, |
1a1aad8a MK |
4341 | .flags = flags, |
4342 | /* | |
4343 | * Hard to debug if it ends up being | |
4344 | * used by a callee that assumes | |
4345 | * something about the other | |
4346 | * uninitialized fields... same as in | |
4347 | * memory.c | |
4348 | */ | |
4349 | }; | |
4350 | ||
4351 | /* | |
c0d0381a MK |
4352 | * hugetlb_fault_mutex and i_mmap_rwsem must be |
4353 | * dropped before handling userfault. Reacquire | |
4354 | * after handling fault to make calling code simpler. | |
1a1aad8a | 4355 | */ |
188b04a7 | 4356 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
1a1aad8a | 4357 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4358 | i_mmap_unlock_read(mapping); |
1a1aad8a | 4359 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); |
c0d0381a | 4360 | i_mmap_lock_read(mapping); |
1a1aad8a MK |
4361 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
4362 | goto out; | |
4363 | } | |
4364 | ||
285b8dca | 4365 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4366 | if (IS_ERR(page)) { |
4643d67e MK |
4367 | /* |
4368 | * Returning error will result in faulting task being | |
4369 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4370 | * tasks from racing to fault in the same page which | |
4371 | * could result in false unable to allocate errors. | |
4372 | * Page migration does not take the fault mutex, but | |
4373 | * does a clear then write of pte's under page table | |
4374 | * lock. Page fault code could race with migration, | |
4375 | * notice the clear pte and try to allocate a page | |
4376 | * here. Before returning error, get ptl and make | |
4377 | * sure there really is no pte entry. | |
4378 | */ | |
4379 | ptl = huge_pte_lock(h, mm, ptep); | |
4380 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
4381 | ret = 0; | |
4382 | spin_unlock(ptl); | |
4383 | goto out; | |
4384 | } | |
4385 | spin_unlock(ptl); | |
2b740303 | 4386 | ret = vmf_error(PTR_ERR(page)); |
6bda666a CL |
4387 | goto out; |
4388 | } | |
47ad8475 | 4389 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4390 | __SetPageUptodate(page); |
cb6acd01 | 4391 | new_page = true; |
ac9b9c66 | 4392 | |
f83a275d | 4393 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4394 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4395 | if (err) { |
4396 | put_page(page); | |
6bda666a CL |
4397 | if (err == -EEXIST) |
4398 | goto retry; | |
4399 | goto out; | |
4400 | } | |
23be7468 | 4401 | } else { |
6bda666a | 4402 | lock_page(page); |
0fe6e20b NH |
4403 | if (unlikely(anon_vma_prepare(vma))) { |
4404 | ret = VM_FAULT_OOM; | |
4405 | goto backout_unlocked; | |
4406 | } | |
409eb8c2 | 4407 | anon_rmap = 1; |
23be7468 | 4408 | } |
0fe6e20b | 4409 | } else { |
998b4382 NH |
4410 | /* |
4411 | * If memory error occurs between mmap() and fault, some process | |
4412 | * don't have hwpoisoned swap entry for errored virtual address. | |
4413 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4414 | */ | |
4415 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 4416 | ret = VM_FAULT_HWPOISON | |
972dc4de | 4417 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4418 | goto backout_unlocked; |
4419 | } | |
6bda666a | 4420 | } |
1e8f889b | 4421 | |
57303d80 AW |
4422 | /* |
4423 | * If we are going to COW a private mapping later, we examine the | |
4424 | * pending reservations for this page now. This will ensure that | |
4425 | * any allocations necessary to record that reservation occur outside | |
4426 | * the spinlock. | |
4427 | */ | |
5e911373 | 4428 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4429 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4430 | ret = VM_FAULT_OOM; |
4431 | goto backout_unlocked; | |
4432 | } | |
5e911373 | 4433 | /* Just decrements count, does not deallocate */ |
285b8dca | 4434 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4435 | } |
57303d80 | 4436 | |
8bea8052 | 4437 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 4438 | ret = 0; |
7f2e9525 | 4439 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4440 | goto backout; |
4441 | ||
07443a85 JK |
4442 | if (anon_rmap) { |
4443 | ClearPagePrivate(page); | |
285b8dca | 4444 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4445 | } else |
53f9263b | 4446 | page_dup_rmap(page, true); |
1e8f889b DG |
4447 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4448 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4449 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4450 | |
5d317b2b | 4451 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4452 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4453 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4454 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4455 | } |
4456 | ||
cb900f41 | 4457 | spin_unlock(ptl); |
cb6acd01 MK |
4458 | |
4459 | /* | |
4460 | * Only make newly allocated pages active. Existing pages found | |
4461 | * in the pagecache could be !page_huge_active() if they have been | |
4462 | * isolated for migration. | |
4463 | */ | |
4464 | if (new_page) | |
4465 | set_page_huge_active(page); | |
4466 | ||
4c887265 AL |
4467 | unlock_page(page); |
4468 | out: | |
ac9b9c66 | 4469 | return ret; |
4c887265 AL |
4470 | |
4471 | backout: | |
cb900f41 | 4472 | spin_unlock(ptl); |
2b26736c | 4473 | backout_unlocked: |
4c887265 | 4474 | unlock_page(page); |
285b8dca | 4475 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4476 | put_page(page); |
4477 | goto out; | |
ac9b9c66 HD |
4478 | } |
4479 | ||
8382d914 | 4480 | #ifdef CONFIG_SMP |
188b04a7 | 4481 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4482 | { |
4483 | unsigned long key[2]; | |
4484 | u32 hash; | |
4485 | ||
1b426bac MK |
4486 | key[0] = (unsigned long) mapping; |
4487 | key[1] = idx; | |
8382d914 | 4488 | |
55254636 | 4489 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4490 | |
4491 | return hash & (num_fault_mutexes - 1); | |
4492 | } | |
4493 | #else | |
4494 | /* | |
4495 | * For uniprocesor systems we always use a single mutex, so just | |
4496 | * return 0 and avoid the hashing overhead. | |
4497 | */ | |
188b04a7 | 4498 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4499 | { |
4500 | return 0; | |
4501 | } | |
4502 | #endif | |
4503 | ||
2b740303 | 4504 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4505 | unsigned long address, unsigned int flags) |
86e5216f | 4506 | { |
8382d914 | 4507 | pte_t *ptep, entry; |
cb900f41 | 4508 | spinlock_t *ptl; |
2b740303 | 4509 | vm_fault_t ret; |
8382d914 DB |
4510 | u32 hash; |
4511 | pgoff_t idx; | |
0fe6e20b | 4512 | struct page *page = NULL; |
57303d80 | 4513 | struct page *pagecache_page = NULL; |
a5516438 | 4514 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4515 | struct address_space *mapping; |
0f792cf9 | 4516 | int need_wait_lock = 0; |
285b8dca | 4517 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4518 | |
285b8dca | 4519 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 4520 | if (ptep) { |
c0d0381a MK |
4521 | /* |
4522 | * Since we hold no locks, ptep could be stale. That is | |
4523 | * OK as we are only making decisions based on content and | |
4524 | * not actually modifying content here. | |
4525 | */ | |
fd6a03ed | 4526 | entry = huge_ptep_get(ptep); |
290408d4 | 4527 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4528 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4529 | return 0; |
4530 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4531 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4532 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
4533 | } |
4534 | ||
c0d0381a MK |
4535 | /* |
4536 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
4537 | * until finished with ptep. This serves two purposes: |
4538 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
4539 | * and making the ptep no longer valid. | |
4540 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
4541 | * |
4542 | * ptep could have already be assigned via huge_pte_offset. That | |
4543 | * is OK, as huge_pte_alloc will return the same value unless | |
4544 | * something has changed. | |
4545 | */ | |
8382d914 | 4546 | mapping = vma->vm_file->f_mapping; |
c0d0381a MK |
4547 | i_mmap_lock_read(mapping); |
4548 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4549 | if (!ptep) { | |
4550 | i_mmap_unlock_read(mapping); | |
4551 | return VM_FAULT_OOM; | |
4552 | } | |
8382d914 | 4553 | |
3935baa9 DG |
4554 | /* |
4555 | * Serialize hugepage allocation and instantiation, so that we don't | |
4556 | * get spurious allocation failures if two CPUs race to instantiate | |
4557 | * the same page in the page cache. | |
4558 | */ | |
c0d0381a | 4559 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 4560 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4561 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4562 | |
7f2e9525 GS |
4563 | entry = huge_ptep_get(ptep); |
4564 | if (huge_pte_none(entry)) { | |
8382d914 | 4565 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4566 | goto out_mutex; |
3935baa9 | 4567 | } |
86e5216f | 4568 | |
83c54070 | 4569 | ret = 0; |
1e8f889b | 4570 | |
0f792cf9 NH |
4571 | /* |
4572 | * entry could be a migration/hwpoison entry at this point, so this | |
4573 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
4574 | * an active hugepage in pagecache. This goto expects the 2nd page |
4575 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
4576 | * properly handle it. | |
0f792cf9 NH |
4577 | */ |
4578 | if (!pte_present(entry)) | |
4579 | goto out_mutex; | |
4580 | ||
57303d80 AW |
4581 | /* |
4582 | * If we are going to COW the mapping later, we examine the pending | |
4583 | * reservations for this page now. This will ensure that any | |
4584 | * allocations necessary to record that reservation occur outside the | |
4585 | * spinlock. For private mappings, we also lookup the pagecache | |
4586 | * page now as it is used to determine if a reservation has been | |
4587 | * consumed. | |
4588 | */ | |
106c992a | 4589 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4590 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4591 | ret = VM_FAULT_OOM; |
b4d1d99f | 4592 | goto out_mutex; |
2b26736c | 4593 | } |
5e911373 | 4594 | /* Just decrements count, does not deallocate */ |
285b8dca | 4595 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4596 | |
f83a275d | 4597 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4598 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4599 | vma, haddr); |
57303d80 AW |
4600 | } |
4601 | ||
0f792cf9 NH |
4602 | ptl = huge_pte_lock(h, mm, ptep); |
4603 | ||
4604 | /* Check for a racing update before calling hugetlb_cow */ | |
4605 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4606 | goto out_ptl; | |
4607 | ||
56c9cfb1 NH |
4608 | /* |
4609 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4610 | * pagecache_page, so here we need take the former one | |
4611 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4612 | */ |
4613 | page = pte_page(entry); | |
4614 | if (page != pagecache_page) | |
0f792cf9 NH |
4615 | if (!trylock_page(page)) { |
4616 | need_wait_lock = 1; | |
4617 | goto out_ptl; | |
4618 | } | |
b4d1d99f | 4619 | |
0f792cf9 | 4620 | get_page(page); |
b4d1d99f | 4621 | |
788c7df4 | 4622 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4623 | if (!huge_pte_write(entry)) { |
974e6d66 | 4624 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4625 | pagecache_page, ptl); |
0f792cf9 | 4626 | goto out_put_page; |
b4d1d99f | 4627 | } |
106c992a | 4628 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4629 | } |
4630 | entry = pte_mkyoung(entry); | |
285b8dca | 4631 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4632 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4633 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4634 | out_put_page: |
4635 | if (page != pagecache_page) | |
4636 | unlock_page(page); | |
4637 | put_page(page); | |
cb900f41 KS |
4638 | out_ptl: |
4639 | spin_unlock(ptl); | |
57303d80 AW |
4640 | |
4641 | if (pagecache_page) { | |
4642 | unlock_page(pagecache_page); | |
4643 | put_page(pagecache_page); | |
4644 | } | |
b4d1d99f | 4645 | out_mutex: |
c672c7f2 | 4646 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4647 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
4648 | /* |
4649 | * Generally it's safe to hold refcount during waiting page lock. But | |
4650 | * here we just wait to defer the next page fault to avoid busy loop and | |
4651 | * the page is not used after unlocked before returning from the current | |
4652 | * page fault. So we are safe from accessing freed page, even if we wait | |
4653 | * here without taking refcount. | |
4654 | */ | |
4655 | if (need_wait_lock) | |
4656 | wait_on_page_locked(page); | |
1e8f889b | 4657 | return ret; |
86e5216f AL |
4658 | } |
4659 | ||
8fb5debc MK |
4660 | /* |
4661 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4662 | * modifications for huge pages. | |
4663 | */ | |
4664 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4665 | pte_t *dst_pte, | |
4666 | struct vm_area_struct *dst_vma, | |
4667 | unsigned long dst_addr, | |
4668 | unsigned long src_addr, | |
4669 | struct page **pagep) | |
4670 | { | |
1e392147 AA |
4671 | struct address_space *mapping; |
4672 | pgoff_t idx; | |
4673 | unsigned long size; | |
1c9e8def | 4674 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4675 | struct hstate *h = hstate_vma(dst_vma); |
4676 | pte_t _dst_pte; | |
4677 | spinlock_t *ptl; | |
4678 | int ret; | |
4679 | struct page *page; | |
4680 | ||
4681 | if (!*pagep) { | |
4682 | ret = -ENOMEM; | |
4683 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4684 | if (IS_ERR(page)) | |
4685 | goto out; | |
4686 | ||
4687 | ret = copy_huge_page_from_user(page, | |
4688 | (const void __user *) src_addr, | |
810a56b9 | 4689 | pages_per_huge_page(h), false); |
8fb5debc | 4690 | |
c1e8d7c6 | 4691 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 4692 | if (unlikely(ret)) { |
9e368259 | 4693 | ret = -ENOENT; |
8fb5debc MK |
4694 | *pagep = page; |
4695 | /* don't free the page */ | |
4696 | goto out; | |
4697 | } | |
4698 | } else { | |
4699 | page = *pagep; | |
4700 | *pagep = NULL; | |
4701 | } | |
4702 | ||
4703 | /* | |
4704 | * The memory barrier inside __SetPageUptodate makes sure that | |
4705 | * preceding stores to the page contents become visible before | |
4706 | * the set_pte_at() write. | |
4707 | */ | |
4708 | __SetPageUptodate(page); | |
8fb5debc | 4709 | |
1e392147 AA |
4710 | mapping = dst_vma->vm_file->f_mapping; |
4711 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4712 | ||
1c9e8def MK |
4713 | /* |
4714 | * If shared, add to page cache | |
4715 | */ | |
4716 | if (vm_shared) { | |
1e392147 AA |
4717 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4718 | ret = -EFAULT; | |
4719 | if (idx >= size) | |
4720 | goto out_release_nounlock; | |
1c9e8def | 4721 | |
1e392147 AA |
4722 | /* |
4723 | * Serialization between remove_inode_hugepages() and | |
4724 | * huge_add_to_page_cache() below happens through the | |
4725 | * hugetlb_fault_mutex_table that here must be hold by | |
4726 | * the caller. | |
4727 | */ | |
1c9e8def MK |
4728 | ret = huge_add_to_page_cache(page, mapping, idx); |
4729 | if (ret) | |
4730 | goto out_release_nounlock; | |
4731 | } | |
4732 | ||
8fb5debc MK |
4733 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4734 | spin_lock(ptl); | |
4735 | ||
1e392147 AA |
4736 | /* |
4737 | * Recheck the i_size after holding PT lock to make sure not | |
4738 | * to leave any page mapped (as page_mapped()) beyond the end | |
4739 | * of the i_size (remove_inode_hugepages() is strict about | |
4740 | * enforcing that). If we bail out here, we'll also leave a | |
4741 | * page in the radix tree in the vm_shared case beyond the end | |
4742 | * of the i_size, but remove_inode_hugepages() will take care | |
4743 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4744 | */ | |
4745 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4746 | ret = -EFAULT; | |
4747 | if (idx >= size) | |
4748 | goto out_release_unlock; | |
4749 | ||
8fb5debc MK |
4750 | ret = -EEXIST; |
4751 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4752 | goto out_release_unlock; | |
4753 | ||
1c9e8def MK |
4754 | if (vm_shared) { |
4755 | page_dup_rmap(page, true); | |
4756 | } else { | |
4757 | ClearPagePrivate(page); | |
4758 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4759 | } | |
8fb5debc MK |
4760 | |
4761 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4762 | if (dst_vma->vm_flags & VM_WRITE) | |
4763 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4764 | _dst_pte = pte_mkyoung(_dst_pte); | |
4765 | ||
4766 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4767 | ||
4768 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4769 | dst_vma->vm_flags & VM_WRITE); | |
4770 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4771 | ||
4772 | /* No need to invalidate - it was non-present before */ | |
4773 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4774 | ||
4775 | spin_unlock(ptl); | |
cb6acd01 | 4776 | set_page_huge_active(page); |
1c9e8def MK |
4777 | if (vm_shared) |
4778 | unlock_page(page); | |
8fb5debc MK |
4779 | ret = 0; |
4780 | out: | |
4781 | return ret; | |
4782 | out_release_unlock: | |
4783 | spin_unlock(ptl); | |
1c9e8def MK |
4784 | if (vm_shared) |
4785 | unlock_page(page); | |
5af10dfd | 4786 | out_release_nounlock: |
8fb5debc MK |
4787 | put_page(page); |
4788 | goto out; | |
4789 | } | |
4790 | ||
28a35716 ML |
4791 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4792 | struct page **pages, struct vm_area_struct **vmas, | |
4793 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 4794 | long i, unsigned int flags, int *locked) |
63551ae0 | 4795 | { |
d5d4b0aa CK |
4796 | unsigned long pfn_offset; |
4797 | unsigned long vaddr = *position; | |
28a35716 | 4798 | unsigned long remainder = *nr_pages; |
a5516438 | 4799 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4800 | int err = -EFAULT; |
63551ae0 | 4801 | |
63551ae0 | 4802 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4803 | pte_t *pte; |
cb900f41 | 4804 | spinlock_t *ptl = NULL; |
2a15efc9 | 4805 | int absent; |
4c887265 | 4806 | struct page *page; |
63551ae0 | 4807 | |
02057967 DR |
4808 | /* |
4809 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4810 | * potentially allocating memory. | |
4811 | */ | |
fa45f116 | 4812 | if (fatal_signal_pending(current)) { |
02057967 DR |
4813 | remainder = 0; |
4814 | break; | |
4815 | } | |
4816 | ||
4c887265 AL |
4817 | /* |
4818 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4819 | * each hugepage. We have to make sure we get the |
4c887265 | 4820 | * first, for the page indexing below to work. |
cb900f41 KS |
4821 | * |
4822 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4823 | */ |
7868a208 PA |
4824 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4825 | huge_page_size(h)); | |
cb900f41 KS |
4826 | if (pte) |
4827 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4828 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4829 | ||
4830 | /* | |
4831 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4832 | * an error where there's an empty slot with no huge pagecache |
4833 | * to back it. This way, we avoid allocating a hugepage, and | |
4834 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4835 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4836 | */ |
3ae77f43 HD |
4837 | if (absent && (flags & FOLL_DUMP) && |
4838 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4839 | if (pte) |
4840 | spin_unlock(ptl); | |
2a15efc9 HD |
4841 | remainder = 0; |
4842 | break; | |
4843 | } | |
63551ae0 | 4844 | |
9cc3a5bd NH |
4845 | /* |
4846 | * We need call hugetlb_fault for both hugepages under migration | |
4847 | * (in which case hugetlb_fault waits for the migration,) and | |
4848 | * hwpoisoned hugepages (in which case we need to prevent the | |
4849 | * caller from accessing to them.) In order to do this, we use | |
4850 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4851 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4852 | * both cases, and because we can't follow correct pages | |
4853 | * directly from any kind of swap entries. | |
4854 | */ | |
4855 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4856 | ((flags & FOLL_WRITE) && |
4857 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4858 | vm_fault_t ret; |
87ffc118 | 4859 | unsigned int fault_flags = 0; |
63551ae0 | 4860 | |
cb900f41 KS |
4861 | if (pte) |
4862 | spin_unlock(ptl); | |
87ffc118 AA |
4863 | if (flags & FOLL_WRITE) |
4864 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 4865 | if (locked) |
71335f37 PX |
4866 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
4867 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
4868 | if (flags & FOLL_NOWAIT) |
4869 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4870 | FAULT_FLAG_RETRY_NOWAIT; | |
4871 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
4872 | /* |
4873 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
4874 | * FAULT_FLAG_TRIED can co-exist | |
4875 | */ | |
87ffc118 AA |
4876 | fault_flags |= FAULT_FLAG_TRIED; |
4877 | } | |
4878 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4879 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4880 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4881 | remainder = 0; |
4882 | break; | |
4883 | } | |
4884 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 4885 | if (locked && |
1ac25013 | 4886 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 4887 | *locked = 0; |
87ffc118 AA |
4888 | *nr_pages = 0; |
4889 | /* | |
4890 | * VM_FAULT_RETRY must not return an | |
4891 | * error, it will return zero | |
4892 | * instead. | |
4893 | * | |
4894 | * No need to update "position" as the | |
4895 | * caller will not check it after | |
4896 | * *nr_pages is set to 0. | |
4897 | */ | |
4898 | return i; | |
4899 | } | |
4900 | continue; | |
4c887265 AL |
4901 | } |
4902 | ||
a5516438 | 4903 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4904 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 4905 | |
acbfb087 ZL |
4906 | /* |
4907 | * If subpage information not requested, update counters | |
4908 | * and skip the same_page loop below. | |
4909 | */ | |
4910 | if (!pages && !vmas && !pfn_offset && | |
4911 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4912 | (remainder >= pages_per_huge_page(h))) { | |
4913 | vaddr += huge_page_size(h); | |
4914 | remainder -= pages_per_huge_page(h); | |
4915 | i += pages_per_huge_page(h); | |
4916 | spin_unlock(ptl); | |
4917 | continue; | |
4918 | } | |
4919 | ||
d5d4b0aa | 4920 | same_page: |
d6692183 | 4921 | if (pages) { |
2a15efc9 | 4922 | pages[i] = mem_map_offset(page, pfn_offset); |
3faa52c0 JH |
4923 | /* |
4924 | * try_grab_page() should always succeed here, because: | |
4925 | * a) we hold the ptl lock, and b) we've just checked | |
4926 | * that the huge page is present in the page tables. If | |
4927 | * the huge page is present, then the tail pages must | |
4928 | * also be present. The ptl prevents the head page and | |
4929 | * tail pages from being rearranged in any way. So this | |
4930 | * page must be available at this point, unless the page | |
4931 | * refcount overflowed: | |
4932 | */ | |
4933 | if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { | |
4934 | spin_unlock(ptl); | |
4935 | remainder = 0; | |
4936 | err = -ENOMEM; | |
4937 | break; | |
4938 | } | |
d6692183 | 4939 | } |
63551ae0 DG |
4940 | |
4941 | if (vmas) | |
4942 | vmas[i] = vma; | |
4943 | ||
4944 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4945 | ++pfn_offset; |
63551ae0 DG |
4946 | --remainder; |
4947 | ++i; | |
d5d4b0aa | 4948 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4949 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa CK |
4950 | /* |
4951 | * We use pfn_offset to avoid touching the pageframes | |
4952 | * of this compound page. | |
4953 | */ | |
4954 | goto same_page; | |
4955 | } | |
cb900f41 | 4956 | spin_unlock(ptl); |
63551ae0 | 4957 | } |
28a35716 | 4958 | *nr_pages = remainder; |
87ffc118 AA |
4959 | /* |
4960 | * setting position is actually required only if remainder is | |
4961 | * not zero but it's faster not to add a "if (remainder)" | |
4962 | * branch. | |
4963 | */ | |
63551ae0 DG |
4964 | *position = vaddr; |
4965 | ||
2be7cfed | 4966 | return i ? i : err; |
63551ae0 | 4967 | } |
8f860591 | 4968 | |
5491ae7b AK |
4969 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4970 | /* | |
4971 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4972 | * implement this. | |
4973 | */ | |
4974 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4975 | #endif | |
4976 | ||
7da4d641 | 4977 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4978 | unsigned long address, unsigned long end, pgprot_t newprot) |
4979 | { | |
4980 | struct mm_struct *mm = vma->vm_mm; | |
4981 | unsigned long start = address; | |
4982 | pte_t *ptep; | |
4983 | pte_t pte; | |
a5516438 | 4984 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4985 | unsigned long pages = 0; |
dff11abe | 4986 | bool shared_pmd = false; |
ac46d4f3 | 4987 | struct mmu_notifier_range range; |
dff11abe MK |
4988 | |
4989 | /* | |
4990 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 4991 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
4992 | * range if PMD sharing is possible. |
4993 | */ | |
7269f999 JG |
4994 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
4995 | 0, vma, mm, start, end); | |
ac46d4f3 | 4996 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
4997 | |
4998 | BUG_ON(address >= end); | |
ac46d4f3 | 4999 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 5000 | |
ac46d4f3 | 5001 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 5002 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 5003 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 5004 | spinlock_t *ptl; |
7868a208 | 5005 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
5006 | if (!ptep) |
5007 | continue; | |
cb900f41 | 5008 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 5009 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
7da4d641 | 5010 | pages++; |
cb900f41 | 5011 | spin_unlock(ptl); |
dff11abe | 5012 | shared_pmd = true; |
39dde65c | 5013 | continue; |
7da4d641 | 5014 | } |
a8bda28d NH |
5015 | pte = huge_ptep_get(ptep); |
5016 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
5017 | spin_unlock(ptl); | |
5018 | continue; | |
5019 | } | |
5020 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
5021 | swp_entry_t entry = pte_to_swp_entry(pte); | |
5022 | ||
5023 | if (is_write_migration_entry(entry)) { | |
5024 | pte_t newpte; | |
5025 | ||
5026 | make_migration_entry_read(&entry); | |
5027 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
5028 | set_huge_swap_pte_at(mm, address, ptep, |
5029 | newpte, huge_page_size(h)); | |
a8bda28d NH |
5030 | pages++; |
5031 | } | |
5032 | spin_unlock(ptl); | |
5033 | continue; | |
5034 | } | |
5035 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
5036 | pte_t old_pte; |
5037 | ||
5038 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
5039 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 5040 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 5041 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 5042 | pages++; |
8f860591 | 5043 | } |
cb900f41 | 5044 | spin_unlock(ptl); |
8f860591 | 5045 | } |
d833352a | 5046 | /* |
c8c06efa | 5047 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 5048 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 5049 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
5050 | * and that page table be reused and filled with junk. If we actually |
5051 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 5052 | */ |
dff11abe | 5053 | if (shared_pmd) |
ac46d4f3 | 5054 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
5055 | else |
5056 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
5057 | /* |
5058 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
5059 | * page table protection not changing it to point to a new page. | |
5060 | * | |
ad56b738 | 5061 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 5062 | */ |
83cde9e8 | 5063 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 5064 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
5065 | |
5066 | return pages << h->order; | |
8f860591 ZY |
5067 | } |
5068 | ||
a1e78772 MG |
5069 | int hugetlb_reserve_pages(struct inode *inode, |
5070 | long from, long to, | |
5a6fe125 | 5071 | struct vm_area_struct *vma, |
ca16d140 | 5072 | vm_flags_t vm_flags) |
e4e574b7 | 5073 | { |
0db9d74e | 5074 | long ret, chg, add = -1; |
a5516438 | 5075 | struct hstate *h = hstate_inode(inode); |
90481622 | 5076 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 5077 | struct resv_map *resv_map; |
075a61d0 | 5078 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 5079 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 5080 | |
63489f8e MK |
5081 | /* This should never happen */ |
5082 | if (from > to) { | |
5083 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
5084 | return -EINVAL; | |
5085 | } | |
5086 | ||
17c9d12e MG |
5087 | /* |
5088 | * Only apply hugepage reservation if asked. At fault time, an | |
5089 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 5090 | * without using reserves |
17c9d12e | 5091 | */ |
ca16d140 | 5092 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
5093 | return 0; |
5094 | ||
a1e78772 MG |
5095 | /* |
5096 | * Shared mappings base their reservation on the number of pages that | |
5097 | * are already allocated on behalf of the file. Private mappings need | |
5098 | * to reserve the full area even if read-only as mprotect() may be | |
5099 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
5100 | */ | |
9119a41e | 5101 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
5102 | /* |
5103 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
5104 | * called for inodes for which resv_maps were created (see | |
5105 | * hugetlbfs_get_inode). | |
5106 | */ | |
4e35f483 | 5107 | resv_map = inode_resv_map(inode); |
9119a41e | 5108 | |
0db9d74e | 5109 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
5110 | |
5111 | } else { | |
e9fe92ae | 5112 | /* Private mapping. */ |
9119a41e | 5113 | resv_map = resv_map_alloc(); |
17c9d12e MG |
5114 | if (!resv_map) |
5115 | return -ENOMEM; | |
5116 | ||
a1e78772 | 5117 | chg = to - from; |
84afd99b | 5118 | |
17c9d12e MG |
5119 | set_vma_resv_map(vma, resv_map); |
5120 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
5121 | } | |
5122 | ||
c50ac050 DH |
5123 | if (chg < 0) { |
5124 | ret = chg; | |
5125 | goto out_err; | |
5126 | } | |
8a630112 | 5127 | |
075a61d0 MA |
5128 | ret = hugetlb_cgroup_charge_cgroup_rsvd( |
5129 | hstate_index(h), chg * pages_per_huge_page(h), &h_cg); | |
5130 | ||
5131 | if (ret < 0) { | |
5132 | ret = -ENOMEM; | |
5133 | goto out_err; | |
5134 | } | |
5135 | ||
5136 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
5137 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
5138 | * of the resv_map. | |
5139 | */ | |
5140 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
5141 | } | |
5142 | ||
1c5ecae3 MK |
5143 | /* |
5144 | * There must be enough pages in the subpool for the mapping. If | |
5145 | * the subpool has a minimum size, there may be some global | |
5146 | * reservations already in place (gbl_reserve). | |
5147 | */ | |
5148 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
5149 | if (gbl_reserve < 0) { | |
c50ac050 | 5150 | ret = -ENOSPC; |
075a61d0 | 5151 | goto out_uncharge_cgroup; |
c50ac050 | 5152 | } |
5a6fe125 MG |
5153 | |
5154 | /* | |
17c9d12e | 5155 | * Check enough hugepages are available for the reservation. |
90481622 | 5156 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 5157 | */ |
1c5ecae3 | 5158 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 5159 | if (ret < 0) { |
075a61d0 | 5160 | goto out_put_pages; |
68842c9b | 5161 | } |
17c9d12e MG |
5162 | |
5163 | /* | |
5164 | * Account for the reservations made. Shared mappings record regions | |
5165 | * that have reservations as they are shared by multiple VMAs. | |
5166 | * When the last VMA disappears, the region map says how much | |
5167 | * the reservation was and the page cache tells how much of | |
5168 | * the reservation was consumed. Private mappings are per-VMA and | |
5169 | * only the consumed reservations are tracked. When the VMA | |
5170 | * disappears, the original reservation is the VMA size and the | |
5171 | * consumed reservations are stored in the map. Hence, nothing | |
5172 | * else has to be done for private mappings here | |
5173 | */ | |
33039678 | 5174 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 5175 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
5176 | |
5177 | if (unlikely(add < 0)) { | |
5178 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 5179 | goto out_put_pages; |
0db9d74e | 5180 | } else if (unlikely(chg > add)) { |
33039678 MK |
5181 | /* |
5182 | * pages in this range were added to the reserve | |
5183 | * map between region_chg and region_add. This | |
5184 | * indicates a race with alloc_huge_page. Adjust | |
5185 | * the subpool and reserve counts modified above | |
5186 | * based on the difference. | |
5187 | */ | |
5188 | long rsv_adjust; | |
5189 | ||
075a61d0 MA |
5190 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
5191 | hstate_index(h), | |
5192 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5193 | ||
33039678 MK |
5194 | rsv_adjust = hugepage_subpool_put_pages(spool, |
5195 | chg - add); | |
5196 | hugetlb_acct_memory(h, -rsv_adjust); | |
5197 | } | |
5198 | } | |
a43a8c39 | 5199 | return 0; |
075a61d0 MA |
5200 | out_put_pages: |
5201 | /* put back original number of pages, chg */ | |
5202 | (void)hugepage_subpool_put_pages(spool, chg); | |
5203 | out_uncharge_cgroup: | |
5204 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5205 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 5206 | out_err: |
5e911373 | 5207 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
5208 | /* Only call region_abort if the region_chg succeeded but the |
5209 | * region_add failed or didn't run. | |
5210 | */ | |
5211 | if (chg >= 0 && add < 0) | |
5212 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
5213 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
5214 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 5215 | return ret; |
a43a8c39 CK |
5216 | } |
5217 | ||
b5cec28d MK |
5218 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
5219 | long freed) | |
a43a8c39 | 5220 | { |
a5516438 | 5221 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 5222 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 5223 | long chg = 0; |
90481622 | 5224 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 5225 | long gbl_reserve; |
45c682a6 | 5226 | |
f27a5136 MK |
5227 | /* |
5228 | * Since this routine can be called in the evict inode path for all | |
5229 | * hugetlbfs inodes, resv_map could be NULL. | |
5230 | */ | |
b5cec28d MK |
5231 | if (resv_map) { |
5232 | chg = region_del(resv_map, start, end); | |
5233 | /* | |
5234 | * region_del() can fail in the rare case where a region | |
5235 | * must be split and another region descriptor can not be | |
5236 | * allocated. If end == LONG_MAX, it will not fail. | |
5237 | */ | |
5238 | if (chg < 0) | |
5239 | return chg; | |
5240 | } | |
5241 | ||
45c682a6 | 5242 | spin_lock(&inode->i_lock); |
e4c6f8be | 5243 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
5244 | spin_unlock(&inode->i_lock); |
5245 | ||
1c5ecae3 MK |
5246 | /* |
5247 | * If the subpool has a minimum size, the number of global | |
5248 | * reservations to be released may be adjusted. | |
5249 | */ | |
5250 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5251 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
5252 | |
5253 | return 0; | |
a43a8c39 | 5254 | } |
93f70f90 | 5255 | |
3212b535 SC |
5256 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
5257 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5258 | struct vm_area_struct *vma, | |
5259 | unsigned long addr, pgoff_t idx) | |
5260 | { | |
5261 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5262 | svma->vm_start; | |
5263 | unsigned long sbase = saddr & PUD_MASK; | |
5264 | unsigned long s_end = sbase + PUD_SIZE; | |
5265 | ||
5266 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
5267 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
5268 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
5269 | |
5270 | /* | |
5271 | * match the virtual addresses, permission and the alignment of the | |
5272 | * page table page. | |
5273 | */ | |
5274 | if (pmd_index(addr) != pmd_index(saddr) || | |
5275 | vm_flags != svm_flags || | |
5276 | sbase < svma->vm_start || svma->vm_end < s_end) | |
5277 | return 0; | |
5278 | ||
5279 | return saddr; | |
5280 | } | |
5281 | ||
31aafb45 | 5282 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
5283 | { |
5284 | unsigned long base = addr & PUD_MASK; | |
5285 | unsigned long end = base + PUD_SIZE; | |
5286 | ||
5287 | /* | |
5288 | * check on proper vm_flags and page table alignment | |
5289 | */ | |
017b1660 | 5290 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
5291 | return true; |
5292 | return false; | |
3212b535 SC |
5293 | } |
5294 | ||
017b1660 MK |
5295 | /* |
5296 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5297 | * If yes, adjust start and end to cover range associated with possible | |
5298 | * shared pmd mappings. | |
5299 | */ | |
5300 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5301 | unsigned long *start, unsigned long *end) | |
5302 | { | |
75802ca6 | 5303 | unsigned long a_start, a_end; |
017b1660 MK |
5304 | |
5305 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
5306 | return; | |
5307 | ||
75802ca6 PX |
5308 | /* Extend the range to be PUD aligned for a worst case scenario */ |
5309 | a_start = ALIGN_DOWN(*start, PUD_SIZE); | |
5310 | a_end = ALIGN(*end, PUD_SIZE); | |
017b1660 | 5311 | |
75802ca6 PX |
5312 | /* |
5313 | * Intersect the range with the vma range, since pmd sharing won't be | |
5314 | * across vma after all | |
5315 | */ | |
5316 | *start = max(vma->vm_start, a_start); | |
5317 | *end = min(vma->vm_end, a_end); | |
017b1660 MK |
5318 | } |
5319 | ||
3212b535 SC |
5320 | /* |
5321 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5322 | * and returns the corresponding pte. While this is not necessary for the | |
5323 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
5324 | * code much cleaner. |
5325 | * | |
5326 | * This routine must be called with i_mmap_rwsem held in at least read mode. | |
5327 | * For hugetlbfs, this prevents removal of any page table entries associated | |
5328 | * with the address space. This is important as we are setting up sharing | |
5329 | * based on existing page table entries (mappings). | |
3212b535 SC |
5330 | */ |
5331 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5332 | { | |
5333 | struct vm_area_struct *vma = find_vma(mm, addr); | |
5334 | struct address_space *mapping = vma->vm_file->f_mapping; | |
5335 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5336 | vma->vm_pgoff; | |
5337 | struct vm_area_struct *svma; | |
5338 | unsigned long saddr; | |
5339 | pte_t *spte = NULL; | |
5340 | pte_t *pte; | |
cb900f41 | 5341 | spinlock_t *ptl; |
3212b535 SC |
5342 | |
5343 | if (!vma_shareable(vma, addr)) | |
5344 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
5345 | ||
3212b535 SC |
5346 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
5347 | if (svma == vma) | |
5348 | continue; | |
5349 | ||
5350 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5351 | if (saddr) { | |
7868a208 PA |
5352 | spte = huge_pte_offset(svma->vm_mm, saddr, |
5353 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
5354 | if (spte) { |
5355 | get_page(virt_to_page(spte)); | |
5356 | break; | |
5357 | } | |
5358 | } | |
5359 | } | |
5360 | ||
5361 | if (!spte) | |
5362 | goto out; | |
5363 | ||
8bea8052 | 5364 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 5365 | if (pud_none(*pud)) { |
3212b535 SC |
5366 | pud_populate(mm, pud, |
5367 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5368 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5369 | } else { |
3212b535 | 5370 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5371 | } |
cb900f41 | 5372 | spin_unlock(ptl); |
3212b535 SC |
5373 | out: |
5374 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
5375 | return pte; |
5376 | } | |
5377 | ||
5378 | /* | |
5379 | * unmap huge page backed by shared pte. | |
5380 | * | |
5381 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5382 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5383 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5384 | * | |
c0d0381a | 5385 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
5386 | * |
5387 | * returns: 1 successfully unmapped a shared pte page | |
5388 | * 0 the underlying pte page is not shared, or it is the last user | |
5389 | */ | |
34ae204f MK |
5390 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5391 | unsigned long *addr, pte_t *ptep) | |
3212b535 SC |
5392 | { |
5393 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
5394 | p4d_t *p4d = p4d_offset(pgd, *addr); |
5395 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 | 5396 | |
34ae204f | 5397 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
5398 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
5399 | if (page_count(virt_to_page(ptep)) == 1) | |
5400 | return 0; | |
5401 | ||
5402 | pud_clear(pud); | |
5403 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 5404 | mm_dec_nr_pmds(mm); |
3212b535 SC |
5405 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
5406 | return 1; | |
5407 | } | |
9e5fc74c SC |
5408 | #define want_pmd_share() (1) |
5409 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
5410 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5411 | { | |
5412 | return NULL; | |
5413 | } | |
e81f2d22 | 5414 | |
34ae204f MK |
5415 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5416 | unsigned long *addr, pte_t *ptep) | |
e81f2d22 ZZ |
5417 | { |
5418 | return 0; | |
5419 | } | |
017b1660 MK |
5420 | |
5421 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5422 | unsigned long *start, unsigned long *end) | |
5423 | { | |
5424 | } | |
9e5fc74c | 5425 | #define want_pmd_share() (0) |
3212b535 SC |
5426 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5427 | ||
9e5fc74c SC |
5428 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
5429 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
5430 | unsigned long addr, unsigned long sz) | |
5431 | { | |
5432 | pgd_t *pgd; | |
c2febafc | 5433 | p4d_t *p4d; |
9e5fc74c SC |
5434 | pud_t *pud; |
5435 | pte_t *pte = NULL; | |
5436 | ||
5437 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5438 | p4d = p4d_alloc(mm, pgd, addr); |
5439 | if (!p4d) | |
5440 | return NULL; | |
c2febafc | 5441 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5442 | if (pud) { |
5443 | if (sz == PUD_SIZE) { | |
5444 | pte = (pte_t *)pud; | |
5445 | } else { | |
5446 | BUG_ON(sz != PMD_SIZE); | |
5447 | if (want_pmd_share() && pud_none(*pud)) | |
5448 | pte = huge_pmd_share(mm, addr, pud); | |
5449 | else | |
5450 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5451 | } | |
5452 | } | |
4e666314 | 5453 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5454 | |
5455 | return pte; | |
5456 | } | |
5457 | ||
9b19df29 PA |
5458 | /* |
5459 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5460 | * entry at address @addr | |
5461 | * | |
8ac0b81a LX |
5462 | * Return: Pointer to page table entry (PUD or PMD) for |
5463 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
5464 | * size @sz doesn't match the hugepage size at this level of the page |
5465 | * table. | |
5466 | */ | |
7868a208 PA |
5467 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5468 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5469 | { |
5470 | pgd_t *pgd; | |
c2febafc | 5471 | p4d_t *p4d; |
8ac0b81a LX |
5472 | pud_t *pud; |
5473 | pmd_t *pmd; | |
9e5fc74c SC |
5474 | |
5475 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5476 | if (!pgd_present(*pgd)) |
5477 | return NULL; | |
5478 | p4d = p4d_offset(pgd, addr); | |
5479 | if (!p4d_present(*p4d)) | |
5480 | return NULL; | |
9b19df29 | 5481 | |
c2febafc | 5482 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
5483 | if (sz == PUD_SIZE) |
5484 | /* must be pud huge, non-present or none */ | |
c2febafc | 5485 | return (pte_t *)pud; |
8ac0b81a | 5486 | if (!pud_present(*pud)) |
9b19df29 | 5487 | return NULL; |
8ac0b81a | 5488 | /* must have a valid entry and size to go further */ |
9b19df29 | 5489 | |
8ac0b81a LX |
5490 | pmd = pmd_offset(pud, addr); |
5491 | /* must be pmd huge, non-present or none */ | |
5492 | return (pte_t *)pmd; | |
9e5fc74c SC |
5493 | } |
5494 | ||
61f77eda NH |
5495 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5496 | ||
5497 | /* | |
5498 | * These functions are overwritable if your architecture needs its own | |
5499 | * behavior. | |
5500 | */ | |
5501 | struct page * __weak | |
5502 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5503 | int write) | |
5504 | { | |
5505 | return ERR_PTR(-EINVAL); | |
5506 | } | |
5507 | ||
4dc71451 AK |
5508 | struct page * __weak |
5509 | follow_huge_pd(struct vm_area_struct *vma, | |
5510 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5511 | { | |
5512 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5513 | return NULL; | |
5514 | } | |
5515 | ||
61f77eda | 5516 | struct page * __weak |
9e5fc74c | 5517 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5518 | pmd_t *pmd, int flags) |
9e5fc74c | 5519 | { |
e66f17ff NH |
5520 | struct page *page = NULL; |
5521 | spinlock_t *ptl; | |
c9d398fa | 5522 | pte_t pte; |
3faa52c0 JH |
5523 | |
5524 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5525 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5526 | (FOLL_PIN | FOLL_GET))) | |
5527 | return NULL; | |
5528 | ||
e66f17ff NH |
5529 | retry: |
5530 | ptl = pmd_lockptr(mm, pmd); | |
5531 | spin_lock(ptl); | |
5532 | /* | |
5533 | * make sure that the address range covered by this pmd is not | |
5534 | * unmapped from other threads. | |
5535 | */ | |
5536 | if (!pmd_huge(*pmd)) | |
5537 | goto out; | |
c9d398fa NH |
5538 | pte = huge_ptep_get((pte_t *)pmd); |
5539 | if (pte_present(pte)) { | |
97534127 | 5540 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
5541 | /* |
5542 | * try_grab_page() should always succeed here, because: a) we | |
5543 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5544 | * huge pmd (head) page is present in the page tables. The ptl | |
5545 | * prevents the head page and tail pages from being rearranged | |
5546 | * in any way. So this page must be available at this point, | |
5547 | * unless the page refcount overflowed: | |
5548 | */ | |
5549 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5550 | page = NULL; | |
5551 | goto out; | |
5552 | } | |
e66f17ff | 5553 | } else { |
c9d398fa | 5554 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5555 | spin_unlock(ptl); |
5556 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5557 | goto retry; | |
5558 | } | |
5559 | /* | |
5560 | * hwpoisoned entry is treated as no_page_table in | |
5561 | * follow_page_mask(). | |
5562 | */ | |
5563 | } | |
5564 | out: | |
5565 | spin_unlock(ptl); | |
9e5fc74c SC |
5566 | return page; |
5567 | } | |
5568 | ||
61f77eda | 5569 | struct page * __weak |
9e5fc74c | 5570 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5571 | pud_t *pud, int flags) |
9e5fc74c | 5572 | { |
3faa52c0 | 5573 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5574 | return NULL; |
9e5fc74c | 5575 | |
e66f17ff | 5576 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5577 | } |
5578 | ||
faaa5b62 AK |
5579 | struct page * __weak |
5580 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5581 | { | |
3faa52c0 | 5582 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5583 | return NULL; |
5584 | ||
5585 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5586 | } | |
5587 | ||
31caf665 NH |
5588 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5589 | { | |
bcc54222 NH |
5590 | bool ret = true; |
5591 | ||
309381fe | 5592 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5593 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
5594 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
5595 | ret = false; | |
5596 | goto unlock; | |
5597 | } | |
5598 | clear_page_huge_active(page); | |
31caf665 | 5599 | list_move_tail(&page->lru, list); |
bcc54222 | 5600 | unlock: |
31caf665 | 5601 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5602 | return ret; |
31caf665 NH |
5603 | } |
5604 | ||
5605 | void putback_active_hugepage(struct page *page) | |
5606 | { | |
309381fe | 5607 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5608 | spin_lock(&hugetlb_lock); |
bcc54222 | 5609 | set_page_huge_active(page); |
31caf665 NH |
5610 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5611 | spin_unlock(&hugetlb_lock); | |
5612 | put_page(page); | |
5613 | } | |
ab5ac90a MH |
5614 | |
5615 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5616 | { | |
5617 | struct hstate *h = page_hstate(oldpage); | |
5618 | ||
5619 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5620 | set_page_owner_migrate_reason(newpage, reason); | |
5621 | ||
5622 | /* | |
5623 | * transfer temporary state of the new huge page. This is | |
5624 | * reverse to other transitions because the newpage is going to | |
5625 | * be final while the old one will be freed so it takes over | |
5626 | * the temporary status. | |
5627 | * | |
5628 | * Also note that we have to transfer the per-node surplus state | |
5629 | * here as well otherwise the global surplus count will not match | |
5630 | * the per-node's. | |
5631 | */ | |
5632 | if (PageHugeTemporary(newpage)) { | |
5633 | int old_nid = page_to_nid(oldpage); | |
5634 | int new_nid = page_to_nid(newpage); | |
5635 | ||
5636 | SetPageHugeTemporary(oldpage); | |
5637 | ClearPageHugeTemporary(newpage); | |
5638 | ||
5639 | spin_lock(&hugetlb_lock); | |
5640 | if (h->surplus_huge_pages_node[old_nid]) { | |
5641 | h->surplus_huge_pages_node[old_nid]--; | |
5642 | h->surplus_huge_pages_node[new_nid]++; | |
5643 | } | |
5644 | spin_unlock(&hugetlb_lock); | |
5645 | } | |
5646 | } | |
cf11e85f RG |
5647 | |
5648 | #ifdef CONFIG_CMA | |
cf11e85f RG |
5649 | static bool cma_reserve_called __initdata; |
5650 | ||
5651 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
5652 | { | |
5653 | hugetlb_cma_size = memparse(p, &p); | |
5654 | return 0; | |
5655 | } | |
5656 | ||
5657 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
5658 | ||
5659 | void __init hugetlb_cma_reserve(int order) | |
5660 | { | |
5661 | unsigned long size, reserved, per_node; | |
5662 | int nid; | |
5663 | ||
5664 | cma_reserve_called = true; | |
5665 | ||
5666 | if (!hugetlb_cma_size) | |
5667 | return; | |
5668 | ||
5669 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
5670 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
5671 | (PAGE_SIZE << order) / SZ_1M); | |
5672 | return; | |
5673 | } | |
5674 | ||
5675 | /* | |
5676 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
5677 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
5678 | */ | |
5679 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
5680 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
5681 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
5682 | ||
5683 | reserved = 0; | |
5684 | for_each_node_state(nid, N_ONLINE) { | |
5685 | int res; | |
29d0f41d | 5686 | char name[20]; |
cf11e85f RG |
5687 | |
5688 | size = min(per_node, hugetlb_cma_size - reserved); | |
5689 | size = round_up(size, PAGE_SIZE << order); | |
5690 | ||
29d0f41d | 5691 | snprintf(name, 20, "hugetlb%d", nid); |
cf11e85f | 5692 | res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, |
29d0f41d | 5693 | 0, false, name, |
cf11e85f RG |
5694 | &hugetlb_cma[nid], nid); |
5695 | if (res) { | |
5696 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
5697 | res, nid); | |
5698 | continue; | |
5699 | } | |
5700 | ||
5701 | reserved += size; | |
5702 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
5703 | size / SZ_1M, nid); | |
5704 | ||
5705 | if (reserved >= hugetlb_cma_size) | |
5706 | break; | |
5707 | } | |
5708 | } | |
5709 | ||
5710 | void __init hugetlb_cma_check(void) | |
5711 | { | |
5712 | if (!hugetlb_cma_size || cma_reserve_called) | |
5713 | return; | |
5714 | ||
5715 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
5716 | } | |
5717 | ||
5718 | #endif /* CONFIG_CMA */ |