]>
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 | ||
1d88433b | 85 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 86 | { |
1d88433b ML |
87 | if (spool->count) |
88 | return false; | |
89 | if (spool->max_hpages != -1) | |
90 | return spool->used_hpages == 0; | |
91 | if (spool->min_hpages != -1) | |
92 | return spool->rsv_hpages == spool->min_hpages; | |
93 | ||
94 | return true; | |
95 | } | |
90481622 | 96 | |
1d88433b ML |
97 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
98 | { | |
90481622 DG |
99 | spin_unlock(&spool->lock); |
100 | ||
101 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 102 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 103 | * free the subpool */ |
1d88433b | 104 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
105 | if (spool->min_hpages != -1) |
106 | hugetlb_acct_memory(spool->hstate, | |
107 | -spool->min_hpages); | |
90481622 | 108 | kfree(spool); |
7ca02d0a | 109 | } |
90481622 DG |
110 | } |
111 | ||
7ca02d0a MK |
112 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
113 | long min_hpages) | |
90481622 DG |
114 | { |
115 | struct hugepage_subpool *spool; | |
116 | ||
c6a91820 | 117 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
118 | if (!spool) |
119 | return NULL; | |
120 | ||
121 | spin_lock_init(&spool->lock); | |
122 | spool->count = 1; | |
7ca02d0a MK |
123 | spool->max_hpages = max_hpages; |
124 | spool->hstate = h; | |
125 | spool->min_hpages = min_hpages; | |
126 | ||
127 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
128 | kfree(spool); | |
129 | return NULL; | |
130 | } | |
131 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
132 | |
133 | return spool; | |
134 | } | |
135 | ||
136 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
137 | { | |
138 | spin_lock(&spool->lock); | |
139 | BUG_ON(!spool->count); | |
140 | spool->count--; | |
141 | unlock_or_release_subpool(spool); | |
142 | } | |
143 | ||
1c5ecae3 MK |
144 | /* |
145 | * Subpool accounting for allocating and reserving pages. | |
146 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 147 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
148 | * global pools must be adjusted (upward). The returned value may |
149 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 150 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
151 | */ |
152 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
153 | long delta) |
154 | { | |
1c5ecae3 | 155 | long ret = delta; |
90481622 DG |
156 | |
157 | if (!spool) | |
1c5ecae3 | 158 | return ret; |
90481622 DG |
159 | |
160 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
161 | |
162 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
163 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
164 | spool->used_hpages += delta; | |
165 | else { | |
166 | ret = -ENOMEM; | |
167 | goto unlock_ret; | |
168 | } | |
90481622 | 169 | } |
90481622 | 170 | |
09a95e29 MK |
171 | /* minimum size accounting */ |
172 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
173 | if (delta > spool->rsv_hpages) { |
174 | /* | |
175 | * Asking for more reserves than those already taken on | |
176 | * behalf of subpool. Return difference. | |
177 | */ | |
178 | ret = delta - spool->rsv_hpages; | |
179 | spool->rsv_hpages = 0; | |
180 | } else { | |
181 | ret = 0; /* reserves already accounted for */ | |
182 | spool->rsv_hpages -= delta; | |
183 | } | |
184 | } | |
185 | ||
186 | unlock_ret: | |
187 | spin_unlock(&spool->lock); | |
90481622 DG |
188 | return ret; |
189 | } | |
190 | ||
1c5ecae3 MK |
191 | /* |
192 | * Subpool accounting for freeing and unreserving pages. | |
193 | * Return the number of global page reservations that must be dropped. | |
194 | * The return value may only be different than the passed value (delta) | |
195 | * in the case where a subpool minimum size must be maintained. | |
196 | */ | |
197 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
198 | long delta) |
199 | { | |
1c5ecae3 MK |
200 | long ret = delta; |
201 | ||
90481622 | 202 | if (!spool) |
1c5ecae3 | 203 | return delta; |
90481622 DG |
204 | |
205 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
206 | |
207 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
208 | spool->used_hpages -= delta; | |
209 | ||
09a95e29 MK |
210 | /* minimum size accounting */ |
211 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
212 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
213 | ret = 0; | |
214 | else | |
215 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
216 | ||
217 | spool->rsv_hpages += delta; | |
218 | if (spool->rsv_hpages > spool->min_hpages) | |
219 | spool->rsv_hpages = spool->min_hpages; | |
220 | } | |
221 | ||
222 | /* | |
223 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
224 | * quota reference, free it now. | |
225 | */ | |
90481622 | 226 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
227 | |
228 | return ret; | |
90481622 DG |
229 | } |
230 | ||
231 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
232 | { | |
233 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
234 | } | |
235 | ||
236 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
237 | { | |
496ad9aa | 238 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
239 | } |
240 | ||
0db9d74e MA |
241 | /* Helper that removes a struct file_region from the resv_map cache and returns |
242 | * it for use. | |
243 | */ | |
244 | static struct file_region * | |
245 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
246 | { | |
247 | struct file_region *nrg = NULL; | |
248 | ||
249 | VM_BUG_ON(resv->region_cache_count <= 0); | |
250 | ||
251 | resv->region_cache_count--; | |
252 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
253 | list_del(&nrg->link); |
254 | ||
255 | nrg->from = from; | |
256 | nrg->to = to; | |
257 | ||
258 | return nrg; | |
259 | } | |
260 | ||
075a61d0 MA |
261 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
262 | struct file_region *rg) | |
263 | { | |
264 | #ifdef CONFIG_CGROUP_HUGETLB | |
265 | nrg->reservation_counter = rg->reservation_counter; | |
266 | nrg->css = rg->css; | |
267 | if (rg->css) | |
268 | css_get(rg->css); | |
269 | #endif | |
270 | } | |
271 | ||
272 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
273 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
274 | struct hstate *h, | |
275 | struct resv_map *resv, | |
276 | struct file_region *nrg) | |
277 | { | |
278 | #ifdef CONFIG_CGROUP_HUGETLB | |
279 | if (h_cg) { | |
280 | nrg->reservation_counter = | |
281 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
282 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
283 | /* |
284 | * The caller will hold exactly one h_cg->css reference for the | |
285 | * whole contiguous reservation region. But this area might be | |
286 | * scattered when there are already some file_regions reside in | |
287 | * it. As a result, many file_regions may share only one css | |
288 | * reference. In order to ensure that one file_region must hold | |
289 | * exactly one h_cg->css reference, we should do css_get for | |
290 | * each file_region and leave the reference held by caller | |
291 | * untouched. | |
292 | */ | |
293 | css_get(&h_cg->css); | |
075a61d0 MA |
294 | if (!resv->pages_per_hpage) |
295 | resv->pages_per_hpage = pages_per_huge_page(h); | |
296 | /* pages_per_hpage should be the same for all entries in | |
297 | * a resv_map. | |
298 | */ | |
299 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
300 | } else { | |
301 | nrg->reservation_counter = NULL; | |
302 | nrg->css = NULL; | |
303 | } | |
304 | #endif | |
305 | } | |
306 | ||
d85aecf2 ML |
307 | static void put_uncharge_info(struct file_region *rg) |
308 | { | |
309 | #ifdef CONFIG_CGROUP_HUGETLB | |
310 | if (rg->css) | |
311 | css_put(rg->css); | |
312 | #endif | |
313 | } | |
314 | ||
a9b3f867 MA |
315 | static bool has_same_uncharge_info(struct file_region *rg, |
316 | struct file_region *org) | |
317 | { | |
318 | #ifdef CONFIG_CGROUP_HUGETLB | |
319 | return rg && org && | |
320 | rg->reservation_counter == org->reservation_counter && | |
321 | rg->css == org->css; | |
322 | ||
323 | #else | |
324 | return true; | |
325 | #endif | |
326 | } | |
327 | ||
328 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
329 | { | |
330 | struct file_region *nrg = NULL, *prg = NULL; | |
331 | ||
332 | prg = list_prev_entry(rg, link); | |
333 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
334 | has_same_uncharge_info(prg, rg)) { | |
335 | prg->to = rg->to; | |
336 | ||
337 | list_del(&rg->link); | |
d85aecf2 | 338 | put_uncharge_info(rg); |
a9b3f867 MA |
339 | kfree(rg); |
340 | ||
7db5e7b6 | 341 | rg = prg; |
a9b3f867 MA |
342 | } |
343 | ||
344 | nrg = list_next_entry(rg, link); | |
345 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
346 | has_same_uncharge_info(nrg, rg)) { | |
347 | nrg->from = rg->from; | |
348 | ||
349 | list_del(&rg->link); | |
d85aecf2 | 350 | put_uncharge_info(rg); |
a9b3f867 | 351 | kfree(rg); |
a9b3f867 MA |
352 | } |
353 | } | |
354 | ||
2103cf9c PX |
355 | static inline long |
356 | hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, | |
357 | long to, struct hstate *h, struct hugetlb_cgroup *cg, | |
358 | long *regions_needed) | |
359 | { | |
360 | struct file_region *nrg; | |
361 | ||
362 | if (!regions_needed) { | |
363 | nrg = get_file_region_entry_from_cache(map, from, to); | |
364 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
365 | list_add(&nrg->link, rg->link.prev); | |
366 | coalesce_file_region(map, nrg); | |
367 | } else | |
368 | *regions_needed += 1; | |
369 | ||
370 | return to - from; | |
371 | } | |
372 | ||
972a3da3 WY |
373 | /* |
374 | * Must be called with resv->lock held. | |
375 | * | |
376 | * Calling this with regions_needed != NULL will count the number of pages | |
377 | * to be added but will not modify the linked list. And regions_needed will | |
378 | * indicate the number of file_regions needed in the cache to carry out to add | |
379 | * the regions for this range. | |
d75c6af9 MA |
380 | */ |
381 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 382 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 383 | struct hstate *h, long *regions_needed) |
d75c6af9 | 384 | { |
0db9d74e | 385 | long add = 0; |
d75c6af9 | 386 | struct list_head *head = &resv->regions; |
0db9d74e | 387 | long last_accounted_offset = f; |
2103cf9c | 388 | struct file_region *rg = NULL, *trg = NULL; |
d75c6af9 | 389 | |
0db9d74e MA |
390 | if (regions_needed) |
391 | *regions_needed = 0; | |
d75c6af9 | 392 | |
0db9d74e MA |
393 | /* In this loop, we essentially handle an entry for the range |
394 | * [last_accounted_offset, rg->from), at every iteration, with some | |
395 | * bounds checking. | |
396 | */ | |
397 | list_for_each_entry_safe(rg, trg, head, link) { | |
398 | /* Skip irrelevant regions that start before our range. */ | |
399 | if (rg->from < f) { | |
400 | /* If this region ends after the last accounted offset, | |
401 | * then we need to update last_accounted_offset. | |
402 | */ | |
403 | if (rg->to > last_accounted_offset) | |
404 | last_accounted_offset = rg->to; | |
405 | continue; | |
406 | } | |
d75c6af9 | 407 | |
0db9d74e MA |
408 | /* When we find a region that starts beyond our range, we've |
409 | * finished. | |
410 | */ | |
ca7e0457 | 411 | if (rg->from >= t) |
d75c6af9 MA |
412 | break; |
413 | ||
0db9d74e MA |
414 | /* Add an entry for last_accounted_offset -> rg->from, and |
415 | * update last_accounted_offset. | |
416 | */ | |
2103cf9c PX |
417 | if (rg->from > last_accounted_offset) |
418 | add += hugetlb_resv_map_add(resv, rg, | |
419 | last_accounted_offset, | |
420 | rg->from, h, h_cg, | |
421 | regions_needed); | |
0db9d74e MA |
422 | |
423 | last_accounted_offset = rg->to; | |
424 | } | |
425 | ||
426 | /* Handle the case where our range extends beyond | |
427 | * last_accounted_offset. | |
428 | */ | |
2103cf9c PX |
429 | if (last_accounted_offset < t) |
430 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
431 | t, h, h_cg, regions_needed); | |
0db9d74e MA |
432 | |
433 | VM_BUG_ON(add < 0); | |
434 | return add; | |
435 | } | |
436 | ||
437 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
438 | */ | |
439 | static int allocate_file_region_entries(struct resv_map *resv, | |
440 | int regions_needed) | |
441 | __must_hold(&resv->lock) | |
442 | { | |
443 | struct list_head allocated_regions; | |
444 | int to_allocate = 0, i = 0; | |
445 | struct file_region *trg = NULL, *rg = NULL; | |
446 | ||
447 | VM_BUG_ON(regions_needed < 0); | |
448 | ||
449 | INIT_LIST_HEAD(&allocated_regions); | |
450 | ||
451 | /* | |
452 | * Check for sufficient descriptors in the cache to accommodate | |
453 | * the number of in progress add operations plus regions_needed. | |
454 | * | |
455 | * This is a while loop because when we drop the lock, some other call | |
456 | * to region_add or region_del may have consumed some region_entries, | |
457 | * so we keep looping here until we finally have enough entries for | |
458 | * (adds_in_progress + regions_needed). | |
459 | */ | |
460 | while (resv->region_cache_count < | |
461 | (resv->adds_in_progress + regions_needed)) { | |
462 | to_allocate = resv->adds_in_progress + regions_needed - | |
463 | resv->region_cache_count; | |
464 | ||
465 | /* At this point, we should have enough entries in the cache | |
466 | * for all the existings adds_in_progress. We should only be | |
467 | * needing to allocate for regions_needed. | |
d75c6af9 | 468 | */ |
0db9d74e MA |
469 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
470 | ||
471 | spin_unlock(&resv->lock); | |
472 | for (i = 0; i < to_allocate; i++) { | |
473 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
474 | if (!trg) | |
475 | goto out_of_memory; | |
476 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 477 | } |
d75c6af9 | 478 | |
0db9d74e MA |
479 | spin_lock(&resv->lock); |
480 | ||
d3ec7b6e WY |
481 | list_splice(&allocated_regions, &resv->region_cache); |
482 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
483 | } |
484 | ||
0db9d74e | 485 | return 0; |
d75c6af9 | 486 | |
0db9d74e MA |
487 | out_of_memory: |
488 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
489 | list_del(&rg->link); | |
490 | kfree(rg); | |
491 | } | |
492 | return -ENOMEM; | |
d75c6af9 MA |
493 | } |
494 | ||
1dd308a7 MK |
495 | /* |
496 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
497 | * map. Regions will be taken from the cache to fill in this range. |
498 | * Sufficient regions should exist in the cache due to the previous | |
499 | * call to region_chg with the same range, but in some cases the cache will not | |
500 | * have sufficient entries due to races with other code doing region_add or | |
501 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 502 | * |
0db9d74e MA |
503 | * regions_needed is the out value provided by a previous call to region_chg. |
504 | * | |
505 | * Return the number of new huge pages added to the map. This number is greater | |
506 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 507 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
508 | * region_add of regions of length 1 never allocate file_regions and cannot |
509 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
510 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 511 | */ |
0db9d74e | 512 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
513 | long in_regions_needed, struct hstate *h, |
514 | struct hugetlb_cgroup *h_cg) | |
96822904 | 515 | { |
0db9d74e | 516 | long add = 0, actual_regions_needed = 0; |
96822904 | 517 | |
7b24d861 | 518 | spin_lock(&resv->lock); |
0db9d74e MA |
519 | retry: |
520 | ||
521 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
522 | add_reservation_in_range(resv, f, t, NULL, NULL, |
523 | &actual_regions_needed); | |
96822904 | 524 | |
5e911373 | 525 | /* |
0db9d74e MA |
526 | * Check for sufficient descriptors in the cache to accommodate |
527 | * this add operation. Note that actual_regions_needed may be greater | |
528 | * than in_regions_needed, as the resv_map may have been modified since | |
529 | * the region_chg call. In this case, we need to make sure that we | |
530 | * allocate extra entries, such that we have enough for all the | |
531 | * existing adds_in_progress, plus the excess needed for this | |
532 | * operation. | |
5e911373 | 533 | */ |
0db9d74e MA |
534 | if (actual_regions_needed > in_regions_needed && |
535 | resv->region_cache_count < | |
536 | resv->adds_in_progress + | |
537 | (actual_regions_needed - in_regions_needed)) { | |
538 | /* region_add operation of range 1 should never need to | |
539 | * allocate file_region entries. | |
540 | */ | |
541 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 542 | |
0db9d74e MA |
543 | if (allocate_file_region_entries( |
544 | resv, actual_regions_needed - in_regions_needed)) { | |
545 | return -ENOMEM; | |
546 | } | |
5e911373 | 547 | |
0db9d74e | 548 | goto retry; |
5e911373 MK |
549 | } |
550 | ||
972a3da3 | 551 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
552 | |
553 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 554 | |
7b24d861 | 555 | spin_unlock(&resv->lock); |
cf3ad20b | 556 | return add; |
96822904 AW |
557 | } |
558 | ||
1dd308a7 MK |
559 | /* |
560 | * Examine the existing reserve map and determine how many | |
561 | * huge pages in the specified range [f, t) are NOT currently | |
562 | * represented. This routine is called before a subsequent | |
563 | * call to region_add that will actually modify the reserve | |
564 | * map to add the specified range [f, t). region_chg does | |
565 | * not change the number of huge pages represented by the | |
0db9d74e MA |
566 | * map. A number of new file_region structures is added to the cache as a |
567 | * placeholder, for the subsequent region_add call to use. At least 1 | |
568 | * file_region structure is added. | |
569 | * | |
570 | * out_regions_needed is the number of regions added to the | |
571 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
572 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
573 | * |
574 | * Returns the number of huge pages that need to be added to the existing | |
575 | * reservation map for the range [f, t). This number is greater or equal to | |
576 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
577 | * is needed and can not be allocated. | |
1dd308a7 | 578 | */ |
0db9d74e MA |
579 | static long region_chg(struct resv_map *resv, long f, long t, |
580 | long *out_regions_needed) | |
96822904 | 581 | { |
96822904 AW |
582 | long chg = 0; |
583 | ||
7b24d861 | 584 | spin_lock(&resv->lock); |
5e911373 | 585 | |
972a3da3 | 586 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 587 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 588 | out_regions_needed); |
5e911373 | 589 | |
0db9d74e MA |
590 | if (*out_regions_needed == 0) |
591 | *out_regions_needed = 1; | |
5e911373 | 592 | |
0db9d74e MA |
593 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
594 | return -ENOMEM; | |
5e911373 | 595 | |
0db9d74e | 596 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 597 | |
7b24d861 | 598 | spin_unlock(&resv->lock); |
96822904 AW |
599 | return chg; |
600 | } | |
601 | ||
5e911373 MK |
602 | /* |
603 | * Abort the in progress add operation. The adds_in_progress field | |
604 | * of the resv_map keeps track of the operations in progress between | |
605 | * calls to region_chg and region_add. Operations are sometimes | |
606 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
607 | * is called to decrement the adds_in_progress counter. regions_needed |
608 | * is the value returned by the region_chg call, it is used to decrement | |
609 | * the adds_in_progress counter. | |
5e911373 MK |
610 | * |
611 | * NOTE: The range arguments [f, t) are not needed or used in this | |
612 | * routine. They are kept to make reading the calling code easier as | |
613 | * arguments will match the associated region_chg call. | |
614 | */ | |
0db9d74e MA |
615 | static void region_abort(struct resv_map *resv, long f, long t, |
616 | long regions_needed) | |
5e911373 MK |
617 | { |
618 | spin_lock(&resv->lock); | |
619 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 620 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
621 | spin_unlock(&resv->lock); |
622 | } | |
623 | ||
1dd308a7 | 624 | /* |
feba16e2 MK |
625 | * Delete the specified range [f, t) from the reserve map. If the |
626 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
627 | * should be deleted. Locate the regions which intersect [f, t) | |
628 | * and either trim, delete or split the existing regions. | |
629 | * | |
630 | * Returns the number of huge pages deleted from the reserve map. | |
631 | * In the normal case, the return value is zero or more. In the | |
632 | * case where a region must be split, a new region descriptor must | |
633 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
634 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
635 | * a region and possibly return -ENOMEM. Callers specifying | |
636 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 637 | */ |
feba16e2 | 638 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 639 | { |
1406ec9b | 640 | struct list_head *head = &resv->regions; |
96822904 | 641 | struct file_region *rg, *trg; |
feba16e2 MK |
642 | struct file_region *nrg = NULL; |
643 | long del = 0; | |
96822904 | 644 | |
feba16e2 | 645 | retry: |
7b24d861 | 646 | spin_lock(&resv->lock); |
feba16e2 | 647 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
648 | /* |
649 | * Skip regions before the range to be deleted. file_region | |
650 | * ranges are normally of the form [from, to). However, there | |
651 | * may be a "placeholder" entry in the map which is of the form | |
652 | * (from, to) with from == to. Check for placeholder entries | |
653 | * at the beginning of the range to be deleted. | |
654 | */ | |
655 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 656 | continue; |
dbe409e4 | 657 | |
feba16e2 | 658 | if (rg->from >= t) |
96822904 | 659 | break; |
96822904 | 660 | |
feba16e2 MK |
661 | if (f > rg->from && t < rg->to) { /* Must split region */ |
662 | /* | |
663 | * Check for an entry in the cache before dropping | |
664 | * lock and attempting allocation. | |
665 | */ | |
666 | if (!nrg && | |
667 | resv->region_cache_count > resv->adds_in_progress) { | |
668 | nrg = list_first_entry(&resv->region_cache, | |
669 | struct file_region, | |
670 | link); | |
671 | list_del(&nrg->link); | |
672 | resv->region_cache_count--; | |
673 | } | |
96822904 | 674 | |
feba16e2 MK |
675 | if (!nrg) { |
676 | spin_unlock(&resv->lock); | |
677 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
678 | if (!nrg) | |
679 | return -ENOMEM; | |
680 | goto retry; | |
681 | } | |
682 | ||
683 | del += t - f; | |
79aa925b | 684 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 685 | resv, rg, t - f, false); |
feba16e2 MK |
686 | |
687 | /* New entry for end of split region */ | |
688 | nrg->from = t; | |
689 | nrg->to = rg->to; | |
075a61d0 MA |
690 | |
691 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
692 | ||
feba16e2 MK |
693 | INIT_LIST_HEAD(&nrg->link); |
694 | ||
695 | /* Original entry is trimmed */ | |
696 | rg->to = f; | |
697 | ||
698 | list_add(&nrg->link, &rg->link); | |
699 | nrg = NULL; | |
96822904 | 700 | break; |
feba16e2 MK |
701 | } |
702 | ||
703 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
704 | del += rg->to - rg->from; | |
075a61d0 | 705 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 706 | rg->to - rg->from, true); |
feba16e2 MK |
707 | list_del(&rg->link); |
708 | kfree(rg); | |
709 | continue; | |
710 | } | |
711 | ||
712 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 713 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 714 | t - rg->from, false); |
075a61d0 | 715 | |
79aa925b MK |
716 | del += t - rg->from; |
717 | rg->from = t; | |
718 | } else { /* Trim end of region */ | |
075a61d0 | 719 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 720 | rg->to - f, false); |
79aa925b MK |
721 | |
722 | del += rg->to - f; | |
723 | rg->to = f; | |
feba16e2 | 724 | } |
96822904 | 725 | } |
7b24d861 | 726 | |
7b24d861 | 727 | spin_unlock(&resv->lock); |
feba16e2 MK |
728 | kfree(nrg); |
729 | return del; | |
96822904 AW |
730 | } |
731 | ||
b5cec28d MK |
732 | /* |
733 | * A rare out of memory error was encountered which prevented removal of | |
734 | * the reserve map region for a page. The huge page itself was free'ed | |
735 | * and removed from the page cache. This routine will adjust the subpool | |
736 | * usage count, and the global reserve count if needed. By incrementing | |
737 | * these counts, the reserve map entry which could not be deleted will | |
738 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
739 | * counts. | |
740 | */ | |
72e2936c | 741 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
742 | { |
743 | struct hugepage_subpool *spool = subpool_inode(inode); | |
744 | long rsv_adjust; | |
745 | ||
746 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 747 | if (rsv_adjust) { |
b5cec28d MK |
748 | struct hstate *h = hstate_inode(inode); |
749 | ||
750 | hugetlb_acct_memory(h, 1); | |
751 | } | |
752 | } | |
753 | ||
1dd308a7 MK |
754 | /* |
755 | * Count and return the number of huge pages in the reserve map | |
756 | * that intersect with the range [f, t). | |
757 | */ | |
1406ec9b | 758 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 759 | { |
1406ec9b | 760 | struct list_head *head = &resv->regions; |
84afd99b AW |
761 | struct file_region *rg; |
762 | long chg = 0; | |
763 | ||
7b24d861 | 764 | spin_lock(&resv->lock); |
84afd99b AW |
765 | /* Locate each segment we overlap with, and count that overlap. */ |
766 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
767 | long seg_from; |
768 | long seg_to; | |
84afd99b AW |
769 | |
770 | if (rg->to <= f) | |
771 | continue; | |
772 | if (rg->from >= t) | |
773 | break; | |
774 | ||
775 | seg_from = max(rg->from, f); | |
776 | seg_to = min(rg->to, t); | |
777 | ||
778 | chg += seg_to - seg_from; | |
779 | } | |
7b24d861 | 780 | spin_unlock(&resv->lock); |
84afd99b AW |
781 | |
782 | return chg; | |
783 | } | |
784 | ||
e7c4b0bf AW |
785 | /* |
786 | * Convert the address within this vma to the page offset within | |
787 | * the mapping, in pagecache page units; huge pages here. | |
788 | */ | |
a5516438 AK |
789 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
790 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 791 | { |
a5516438 AK |
792 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
793 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
794 | } |
795 | ||
0fe6e20b NH |
796 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
797 | unsigned long address) | |
798 | { | |
799 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
800 | } | |
dee41079 | 801 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 802 | |
08fba699 MG |
803 | /* |
804 | * Return the size of the pages allocated when backing a VMA. In the majority | |
805 | * cases this will be same size as used by the page table entries. | |
806 | */ | |
807 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
808 | { | |
05ea8860 DW |
809 | if (vma->vm_ops && vma->vm_ops->pagesize) |
810 | return vma->vm_ops->pagesize(vma); | |
811 | return PAGE_SIZE; | |
08fba699 | 812 | } |
f340ca0f | 813 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 814 | |
3340289d MG |
815 | /* |
816 | * Return the page size being used by the MMU to back a VMA. In the majority | |
817 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
818 | * architectures where it differs, an architecture-specific 'strong' |
819 | * version of this symbol is required. | |
3340289d | 820 | */ |
09135cc5 | 821 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
822 | { |
823 | return vma_kernel_pagesize(vma); | |
824 | } | |
3340289d | 825 | |
84afd99b AW |
826 | /* |
827 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
828 | * bits of the reservation map pointer, which are always clear due to | |
829 | * alignment. | |
830 | */ | |
831 | #define HPAGE_RESV_OWNER (1UL << 0) | |
832 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 833 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 834 | |
a1e78772 MG |
835 | /* |
836 | * These helpers are used to track how many pages are reserved for | |
837 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
838 | * is guaranteed to have their future faults succeed. | |
839 | * | |
840 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
841 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
842 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
843 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
844 | * |
845 | * The private mapping reservation is represented in a subtly different | |
846 | * manner to a shared mapping. A shared mapping has a region map associated | |
847 | * with the underlying file, this region map represents the backing file | |
848 | * pages which have ever had a reservation assigned which this persists even | |
849 | * after the page is instantiated. A private mapping has a region map | |
850 | * associated with the original mmap which is attached to all VMAs which | |
851 | * reference it, this region map represents those offsets which have consumed | |
852 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 853 | */ |
e7c4b0bf AW |
854 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
855 | { | |
856 | return (unsigned long)vma->vm_private_data; | |
857 | } | |
858 | ||
859 | static void set_vma_private_data(struct vm_area_struct *vma, | |
860 | unsigned long value) | |
861 | { | |
862 | vma->vm_private_data = (void *)value; | |
863 | } | |
864 | ||
e9fe92ae MA |
865 | static void |
866 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
867 | struct hugetlb_cgroup *h_cg, | |
868 | struct hstate *h) | |
869 | { | |
870 | #ifdef CONFIG_CGROUP_HUGETLB | |
871 | if (!h_cg || !h) { | |
872 | resv_map->reservation_counter = NULL; | |
873 | resv_map->pages_per_hpage = 0; | |
874 | resv_map->css = NULL; | |
875 | } else { | |
876 | resv_map->reservation_counter = | |
877 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
878 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
879 | resv_map->css = &h_cg->css; | |
880 | } | |
881 | #endif | |
882 | } | |
883 | ||
9119a41e | 884 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
885 | { |
886 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
887 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
888 | ||
889 | if (!resv_map || !rg) { | |
890 | kfree(resv_map); | |
891 | kfree(rg); | |
84afd99b | 892 | return NULL; |
5e911373 | 893 | } |
84afd99b AW |
894 | |
895 | kref_init(&resv_map->refs); | |
7b24d861 | 896 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
897 | INIT_LIST_HEAD(&resv_map->regions); |
898 | ||
5e911373 | 899 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
900 | /* |
901 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
902 | * fields don't do cgroup accounting. On private mappings, these will be | |
903 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
904 | * reservations are to be un-charged from here. | |
905 | */ | |
906 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
907 | |
908 | INIT_LIST_HEAD(&resv_map->region_cache); | |
909 | list_add(&rg->link, &resv_map->region_cache); | |
910 | resv_map->region_cache_count = 1; | |
911 | ||
84afd99b AW |
912 | return resv_map; |
913 | } | |
914 | ||
9119a41e | 915 | void resv_map_release(struct kref *ref) |
84afd99b AW |
916 | { |
917 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
918 | struct list_head *head = &resv_map->region_cache; |
919 | struct file_region *rg, *trg; | |
84afd99b AW |
920 | |
921 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 922 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
923 | |
924 | /* ... and any entries left in the cache */ | |
925 | list_for_each_entry_safe(rg, trg, head, link) { | |
926 | list_del(&rg->link); | |
927 | kfree(rg); | |
928 | } | |
929 | ||
930 | VM_BUG_ON(resv_map->adds_in_progress); | |
931 | ||
84afd99b AW |
932 | kfree(resv_map); |
933 | } | |
934 | ||
4e35f483 JK |
935 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
936 | { | |
f27a5136 MK |
937 | /* |
938 | * At inode evict time, i_mapping may not point to the original | |
939 | * address space within the inode. This original address space | |
940 | * contains the pointer to the resv_map. So, always use the | |
941 | * address space embedded within the inode. | |
942 | * The VERY common case is inode->mapping == &inode->i_data but, | |
943 | * this may not be true for device special inodes. | |
944 | */ | |
945 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
946 | } |
947 | ||
84afd99b | 948 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 949 | { |
81d1b09c | 950 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
951 | if (vma->vm_flags & VM_MAYSHARE) { |
952 | struct address_space *mapping = vma->vm_file->f_mapping; | |
953 | struct inode *inode = mapping->host; | |
954 | ||
955 | return inode_resv_map(inode); | |
956 | ||
957 | } else { | |
84afd99b AW |
958 | return (struct resv_map *)(get_vma_private_data(vma) & |
959 | ~HPAGE_RESV_MASK); | |
4e35f483 | 960 | } |
a1e78772 MG |
961 | } |
962 | ||
84afd99b | 963 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 964 | { |
81d1b09c SL |
965 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
966 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 967 | |
84afd99b AW |
968 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
969 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
970 | } |
971 | ||
972 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
973 | { | |
81d1b09c SL |
974 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
975 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
976 | |
977 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
978 | } |
979 | ||
980 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
981 | { | |
81d1b09c | 982 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
983 | |
984 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
985 | } |
986 | ||
04f2cbe3 | 987 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
988 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
989 | { | |
81d1b09c | 990 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 991 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
992 | vma->vm_private_data = (void *)0; |
993 | } | |
994 | ||
995 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 996 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 997 | { |
af0ed73e JK |
998 | if (vma->vm_flags & VM_NORESERVE) { |
999 | /* | |
1000 | * This address is already reserved by other process(chg == 0), | |
1001 | * so, we should decrement reserved count. Without decrementing, | |
1002 | * reserve count remains after releasing inode, because this | |
1003 | * allocated page will go into page cache and is regarded as | |
1004 | * coming from reserved pool in releasing step. Currently, we | |
1005 | * don't have any other solution to deal with this situation | |
1006 | * properly, so add work-around here. | |
1007 | */ | |
1008 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1009 | return true; |
af0ed73e | 1010 | else |
559ec2f8 | 1011 | return false; |
af0ed73e | 1012 | } |
a63884e9 JK |
1013 | |
1014 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1015 | if (vma->vm_flags & VM_MAYSHARE) { |
1016 | /* | |
1017 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1018 | * be a region map for all pages. The only situation where | |
1019 | * there is no region map is if a hole was punched via | |
7c8de358 | 1020 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1021 | * use. This situation is indicated if chg != 0. |
1022 | */ | |
1023 | if (chg) | |
1024 | return false; | |
1025 | else | |
1026 | return true; | |
1027 | } | |
a63884e9 JK |
1028 | |
1029 | /* | |
1030 | * Only the process that called mmap() has reserves for | |
1031 | * private mappings. | |
1032 | */ | |
67961f9d MK |
1033 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1034 | /* | |
1035 | * Like the shared case above, a hole punch or truncate | |
1036 | * could have been performed on the private mapping. | |
1037 | * Examine the value of chg to determine if reserves | |
1038 | * actually exist or were previously consumed. | |
1039 | * Very Subtle - The value of chg comes from a previous | |
1040 | * call to vma_needs_reserves(). The reserve map for | |
1041 | * private mappings has different (opposite) semantics | |
1042 | * than that of shared mappings. vma_needs_reserves() | |
1043 | * has already taken this difference in semantics into | |
1044 | * account. Therefore, the meaning of chg is the same | |
1045 | * as in the shared case above. Code could easily be | |
1046 | * combined, but keeping it separate draws attention to | |
1047 | * subtle differences. | |
1048 | */ | |
1049 | if (chg) | |
1050 | return false; | |
1051 | else | |
1052 | return true; | |
1053 | } | |
a63884e9 | 1054 | |
559ec2f8 | 1055 | return false; |
a1e78772 MG |
1056 | } |
1057 | ||
a5516438 | 1058 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1059 | { |
1060 | int nid = page_to_nid(page); | |
0edaecfa | 1061 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1062 | h->free_huge_pages++; |
1063 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1064 | SetHPageFreed(page); |
1da177e4 LT |
1065 | } |
1066 | ||
94310cbc | 1067 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1068 | { |
1069 | struct page *page; | |
bbe88753 JK |
1070 | bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); |
1071 | ||
1072 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { | |
1073 | if (nocma && is_migrate_cma_page(page)) | |
1074 | continue; | |
bf50bab2 | 1075 | |
6664bfc8 WY |
1076 | if (PageHWPoison(page)) |
1077 | continue; | |
1078 | ||
1079 | list_move(&page->lru, &h->hugepage_activelist); | |
1080 | set_page_refcounted(page); | |
6c037149 | 1081 | ClearHPageFreed(page); |
6664bfc8 WY |
1082 | h->free_huge_pages--; |
1083 | h->free_huge_pages_node[nid]--; | |
1084 | return page; | |
bbe88753 JK |
1085 | } |
1086 | ||
6664bfc8 | 1087 | return NULL; |
bf50bab2 NH |
1088 | } |
1089 | ||
3e59fcb0 MH |
1090 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1091 | nodemask_t *nmask) | |
94310cbc | 1092 | { |
3e59fcb0 MH |
1093 | unsigned int cpuset_mems_cookie; |
1094 | struct zonelist *zonelist; | |
1095 | struct zone *zone; | |
1096 | struct zoneref *z; | |
98fa15f3 | 1097 | int node = NUMA_NO_NODE; |
94310cbc | 1098 | |
3e59fcb0 MH |
1099 | zonelist = node_zonelist(nid, gfp_mask); |
1100 | ||
1101 | retry_cpuset: | |
1102 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1103 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1104 | struct page *page; | |
1105 | ||
1106 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1107 | continue; | |
1108 | /* | |
1109 | * no need to ask again on the same node. Pool is node rather than | |
1110 | * zone aware | |
1111 | */ | |
1112 | if (zone_to_nid(zone) == node) | |
1113 | continue; | |
1114 | node = zone_to_nid(zone); | |
94310cbc | 1115 | |
94310cbc AK |
1116 | page = dequeue_huge_page_node_exact(h, node); |
1117 | if (page) | |
1118 | return page; | |
1119 | } | |
3e59fcb0 MH |
1120 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1121 | goto retry_cpuset; | |
1122 | ||
94310cbc AK |
1123 | return NULL; |
1124 | } | |
1125 | ||
a5516438 AK |
1126 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1127 | struct vm_area_struct *vma, | |
af0ed73e JK |
1128 | unsigned long address, int avoid_reserve, |
1129 | long chg) | |
1da177e4 | 1130 | { |
3e59fcb0 | 1131 | struct page *page; |
480eccf9 | 1132 | struct mempolicy *mpol; |
04ec6264 | 1133 | gfp_t gfp_mask; |
3e59fcb0 | 1134 | nodemask_t *nodemask; |
04ec6264 | 1135 | int nid; |
1da177e4 | 1136 | |
a1e78772 MG |
1137 | /* |
1138 | * A child process with MAP_PRIVATE mappings created by their parent | |
1139 | * have no page reserves. This check ensures that reservations are | |
1140 | * not "stolen". The child may still get SIGKILLed | |
1141 | */ | |
af0ed73e | 1142 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1143 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1144 | goto err; |
a1e78772 | 1145 | |
04f2cbe3 | 1146 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1147 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1148 | goto err; |
04f2cbe3 | 1149 | |
04ec6264 VB |
1150 | gfp_mask = htlb_alloc_mask(h); |
1151 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
1152 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
1153 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
d6995da3 | 1154 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1155 | h->resv_huge_pages--; |
1da177e4 | 1156 | } |
cc9a6c87 | 1157 | |
52cd3b07 | 1158 | mpol_cond_put(mpol); |
1da177e4 | 1159 | return page; |
cc9a6c87 MG |
1160 | |
1161 | err: | |
cc9a6c87 | 1162 | return NULL; |
1da177e4 LT |
1163 | } |
1164 | ||
1cac6f2c LC |
1165 | /* |
1166 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1167 | * We may have allocated or freed a huge page based on a different | |
1168 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1169 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1170 | * node for alloc or free. | |
1171 | */ | |
1172 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1173 | { | |
0edaf86c | 1174 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1175 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1176 | ||
1177 | return nid; | |
1178 | } | |
1179 | ||
1180 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1181 | { | |
1182 | if (!node_isset(nid, *nodes_allowed)) | |
1183 | nid = next_node_allowed(nid, nodes_allowed); | |
1184 | return nid; | |
1185 | } | |
1186 | ||
1187 | /* | |
1188 | * returns the previously saved node ["this node"] from which to | |
1189 | * allocate a persistent huge page for the pool and advance the | |
1190 | * next node from which to allocate, handling wrap at end of node | |
1191 | * mask. | |
1192 | */ | |
1193 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1194 | nodemask_t *nodes_allowed) | |
1195 | { | |
1196 | int nid; | |
1197 | ||
1198 | VM_BUG_ON(!nodes_allowed); | |
1199 | ||
1200 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1201 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1202 | ||
1203 | return nid; | |
1204 | } | |
1205 | ||
1206 | /* | |
1207 | * helper for free_pool_huge_page() - return the previously saved | |
1208 | * node ["this node"] from which to free a huge page. Advance the | |
1209 | * next node id whether or not we find a free huge page to free so | |
1210 | * that the next attempt to free addresses the next node. | |
1211 | */ | |
1212 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1213 | { | |
1214 | int nid; | |
1215 | ||
1216 | VM_BUG_ON(!nodes_allowed); | |
1217 | ||
1218 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1219 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1220 | ||
1221 | return nid; | |
1222 | } | |
1223 | ||
1224 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1225 | for (nr_nodes = nodes_weight(*mask); \ | |
1226 | nr_nodes > 0 && \ | |
1227 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1228 | nr_nodes--) | |
1229 | ||
1230 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1231 | for (nr_nodes = nodes_weight(*mask); \ | |
1232 | nr_nodes > 0 && \ | |
1233 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1234 | nr_nodes--) | |
1235 | ||
e1073d1e | 1236 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1237 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1238 | unsigned int order) |
944d9fec LC |
1239 | { |
1240 | int i; | |
1241 | int nr_pages = 1 << order; | |
1242 | struct page *p = page + 1; | |
1243 | ||
c8cc708a | 1244 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1245 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1246 | |
944d9fec | 1247 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1248 | clear_compound_head(p); |
944d9fec | 1249 | set_page_refcounted(p); |
944d9fec LC |
1250 | } |
1251 | ||
1252 | set_compound_order(page, 0); | |
ba9c1201 | 1253 | page[1].compound_nr = 0; |
944d9fec LC |
1254 | __ClearPageHead(page); |
1255 | } | |
1256 | ||
d00181b9 | 1257 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1258 | { |
cf11e85f RG |
1259 | /* |
1260 | * If the page isn't allocated using the cma allocator, | |
1261 | * cma_release() returns false. | |
1262 | */ | |
dbda8fea BS |
1263 | #ifdef CONFIG_CMA |
1264 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1265 | return; |
dbda8fea | 1266 | #endif |
cf11e85f | 1267 | |
944d9fec LC |
1268 | free_contig_range(page_to_pfn(page), 1 << order); |
1269 | } | |
1270 | ||
4eb0716e | 1271 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1272 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1273 | int nid, nodemask_t *nodemask) | |
944d9fec | 1274 | { |
04adbc3f | 1275 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1276 | if (nid == NUMA_NO_NODE) |
1277 | nid = numa_mem_id(); | |
944d9fec | 1278 | |
dbda8fea BS |
1279 | #ifdef CONFIG_CMA |
1280 | { | |
cf11e85f RG |
1281 | struct page *page; |
1282 | int node; | |
1283 | ||
953f064a LX |
1284 | if (hugetlb_cma[nid]) { |
1285 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1286 | huge_page_order(h), true); | |
cf11e85f RG |
1287 | if (page) |
1288 | return page; | |
1289 | } | |
953f064a LX |
1290 | |
1291 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1292 | for_each_node_mask(node, *nodemask) { | |
1293 | if (node == nid || !hugetlb_cma[node]) | |
1294 | continue; | |
1295 | ||
1296 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1297 | huge_page_order(h), true); | |
1298 | if (page) | |
1299 | return page; | |
1300 | } | |
1301 | } | |
cf11e85f | 1302 | } |
dbda8fea | 1303 | #endif |
cf11e85f | 1304 | |
5e27a2df | 1305 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1306 | } |
1307 | ||
1308 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1309 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1310 | #else /* !CONFIG_CONTIG_ALLOC */ |
1311 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1312 | int nid, nodemask_t *nodemask) | |
1313 | { | |
1314 | return NULL; | |
1315 | } | |
1316 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1317 | |
e1073d1e | 1318 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1319 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1320 | int nid, nodemask_t *nodemask) |
1321 | { | |
1322 | return NULL; | |
1323 | } | |
d00181b9 | 1324 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1325 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1326 | unsigned int order) { } |
944d9fec LC |
1327 | #endif |
1328 | ||
a5516438 | 1329 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1330 | { |
1331 | int i; | |
dbfee5ae | 1332 | struct page *subpage = page; |
a5516438 | 1333 | |
4eb0716e | 1334 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1335 | return; |
18229df5 | 1336 | |
a5516438 AK |
1337 | h->nr_huge_pages--; |
1338 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
dbfee5ae MK |
1339 | for (i = 0; i < pages_per_huge_page(h); |
1340 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1341 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1342 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1343 | 1 << PG_active | 1 << PG_private | |
1344 | 1 << PG_writeback); | |
6af2acb6 | 1345 | } |
309381fe | 1346 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
1adc4d41 | 1347 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); |
f1e61557 | 1348 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1349 | set_page_refcounted(page); |
944d9fec | 1350 | if (hstate_is_gigantic(h)) { |
cf11e85f RG |
1351 | /* |
1352 | * Temporarily drop the hugetlb_lock, because | |
1353 | * we might block in free_gigantic_page(). | |
1354 | */ | |
1355 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1356 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1357 | free_gigantic_page(page, huge_page_order(h)); | |
cf11e85f | 1358 | spin_lock(&hugetlb_lock); |
944d9fec | 1359 | } else { |
944d9fec LC |
1360 | __free_pages(page, huge_page_order(h)); |
1361 | } | |
6af2acb6 AL |
1362 | } |
1363 | ||
e5ff2159 AK |
1364 | struct hstate *size_to_hstate(unsigned long size) |
1365 | { | |
1366 | struct hstate *h; | |
1367 | ||
1368 | for_each_hstate(h) { | |
1369 | if (huge_page_size(h) == size) | |
1370 | return h; | |
1371 | } | |
1372 | return NULL; | |
1373 | } | |
1374 | ||
c77c0a8a | 1375 | static void __free_huge_page(struct page *page) |
27a85ef1 | 1376 | { |
a5516438 AK |
1377 | /* |
1378 | * Can't pass hstate in here because it is called from the | |
1379 | * compound page destructor. | |
1380 | */ | |
e5ff2159 | 1381 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1382 | int nid = page_to_nid(page); |
d6995da3 | 1383 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1384 | bool restore_reserve; |
27a85ef1 | 1385 | |
b4330afb MK |
1386 | VM_BUG_ON_PAGE(page_count(page), page); |
1387 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1388 | |
d6995da3 | 1389 | hugetlb_set_page_subpool(page, NULL); |
8ace22bc | 1390 | page->mapping = NULL; |
d6995da3 MK |
1391 | restore_reserve = HPageRestoreReserve(page); |
1392 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1393 | |
1c5ecae3 | 1394 | /* |
d6995da3 | 1395 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1396 | * reservation. If the page was associated with a subpool, there |
1397 | * would have been a page reserved in the subpool before allocation | |
1398 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1399 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1400 | * remove the reserved page from the subpool. |
1c5ecae3 | 1401 | */ |
0919e1b6 MK |
1402 | if (!restore_reserve) { |
1403 | /* | |
1404 | * A return code of zero implies that the subpool will be | |
1405 | * under its minimum size if the reservation is not restored | |
1406 | * after page is free. Therefore, force restore_reserve | |
1407 | * operation. | |
1408 | */ | |
1409 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1410 | restore_reserve = true; | |
1411 | } | |
1c5ecae3 | 1412 | |
27a85ef1 | 1413 | spin_lock(&hugetlb_lock); |
8f251a3d | 1414 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1415 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1416 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1417 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1418 | pages_per_huge_page(h), page); | |
07443a85 JK |
1419 | if (restore_reserve) |
1420 | h->resv_huge_pages++; | |
1421 | ||
9157c311 | 1422 | if (HPageTemporary(page)) { |
ab5ac90a | 1423 | list_del(&page->lru); |
9157c311 | 1424 | ClearHPageTemporary(page); |
ab5ac90a MH |
1425 | update_and_free_page(h, page); |
1426 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1427 | /* remove the page from active list */ |
1428 | list_del(&page->lru); | |
a5516438 AK |
1429 | update_and_free_page(h, page); |
1430 | h->surplus_huge_pages--; | |
1431 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1432 | } else { |
5d3a551c | 1433 | arch_clear_hugepage_flags(page); |
a5516438 | 1434 | enqueue_huge_page(h, page); |
7893d1d5 | 1435 | } |
27a85ef1 DG |
1436 | spin_unlock(&hugetlb_lock); |
1437 | } | |
1438 | ||
c77c0a8a WL |
1439 | /* |
1440 | * As free_huge_page() can be called from a non-task context, we have | |
1441 | * to defer the actual freeing in a workqueue to prevent potential | |
1442 | * hugetlb_lock deadlock. | |
1443 | * | |
1444 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1445 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1446 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1447 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1448 | */ | |
1449 | static LLIST_HEAD(hpage_freelist); | |
1450 | ||
1451 | static void free_hpage_workfn(struct work_struct *work) | |
1452 | { | |
1453 | struct llist_node *node; | |
1454 | struct page *page; | |
1455 | ||
1456 | node = llist_del_all(&hpage_freelist); | |
1457 | ||
1458 | while (node) { | |
1459 | page = container_of((struct address_space **)node, | |
1460 | struct page, mapping); | |
1461 | node = node->next; | |
1462 | __free_huge_page(page); | |
1463 | } | |
1464 | } | |
1465 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1466 | ||
1467 | void free_huge_page(struct page *page) | |
1468 | { | |
1469 | /* | |
1470 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1471 | */ | |
1472 | if (!in_task()) { | |
1473 | /* | |
1474 | * Only call schedule_work() if hpage_freelist is previously | |
1475 | * empty. Otherwise, schedule_work() had been called but the | |
1476 | * workfn hasn't retrieved the list yet. | |
1477 | */ | |
1478 | if (llist_add((struct llist_node *)&page->mapping, | |
1479 | &hpage_freelist)) | |
1480 | schedule_work(&free_hpage_work); | |
1481 | return; | |
1482 | } | |
1483 | ||
1484 | __free_huge_page(page); | |
1485 | } | |
1486 | ||
a5516438 | 1487 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1488 | { |
0edaecfa | 1489 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1490 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1491 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1492 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1493 | set_hugetlb_cgroup_rsvd(page, NULL); |
2f37511c | 1494 | spin_lock(&hugetlb_lock); |
a5516438 AK |
1495 | h->nr_huge_pages++; |
1496 | h->nr_huge_pages_node[nid]++; | |
6c037149 | 1497 | ClearHPageFreed(page); |
b7ba30c6 | 1498 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1499 | } |
1500 | ||
d00181b9 | 1501 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1502 | { |
1503 | int i; | |
1504 | int nr_pages = 1 << order; | |
1505 | struct page *p = page + 1; | |
1506 | ||
1507 | /* we rely on prep_new_huge_page to set the destructor */ | |
1508 | set_compound_order(page, order); | |
ef5a22be | 1509 | __ClearPageReserved(page); |
de09d31d | 1510 | __SetPageHead(page); |
20a0307c | 1511 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1512 | /* |
1513 | * For gigantic hugepages allocated through bootmem at | |
1514 | * boot, it's safer to be consistent with the not-gigantic | |
1515 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1516 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1517 | * pages may get the reference counting wrong if they see |
1518 | * PG_reserved set on a tail page (despite the head page not | |
1519 | * having PG_reserved set). Enforcing this consistency between | |
1520 | * head and tail pages allows drivers to optimize away a check | |
1521 | * on the head page when they need know if put_page() is needed | |
1522 | * after get_user_pages(). | |
1523 | */ | |
1524 | __ClearPageReserved(p); | |
58a84aa9 | 1525 | set_page_count(p, 0); |
1d798ca3 | 1526 | set_compound_head(p, page); |
20a0307c | 1527 | } |
b4330afb | 1528 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1529 | atomic_set(compound_pincount_ptr(page), 0); |
20a0307c WF |
1530 | } |
1531 | ||
7795912c AM |
1532 | /* |
1533 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1534 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1535 | * details. | |
1536 | */ | |
20a0307c WF |
1537 | int PageHuge(struct page *page) |
1538 | { | |
20a0307c WF |
1539 | if (!PageCompound(page)) |
1540 | return 0; | |
1541 | ||
1542 | page = compound_head(page); | |
f1e61557 | 1543 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1544 | } |
43131e14 NH |
1545 | EXPORT_SYMBOL_GPL(PageHuge); |
1546 | ||
27c73ae7 AA |
1547 | /* |
1548 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1549 | * normal or transparent huge pages. | |
1550 | */ | |
1551 | int PageHeadHuge(struct page *page_head) | |
1552 | { | |
27c73ae7 AA |
1553 | if (!PageHead(page_head)) |
1554 | return 0; | |
1555 | ||
d4af73e3 | 1556 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1557 | } |
27c73ae7 | 1558 | |
c0d0381a MK |
1559 | /* |
1560 | * Find and lock address space (mapping) in write mode. | |
1561 | * | |
336bf30e MK |
1562 | * Upon entry, the page is locked which means that page_mapping() is |
1563 | * stable. Due to locking order, we can only trylock_write. If we can | |
1564 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1565 | */ |
1566 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1567 | { | |
336bf30e | 1568 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1569 | |
c0d0381a MK |
1570 | if (!mapping) |
1571 | return mapping; | |
1572 | ||
c0d0381a MK |
1573 | if (i_mmap_trylock_write(mapping)) |
1574 | return mapping; | |
1575 | ||
336bf30e | 1576 | return NULL; |
c0d0381a MK |
1577 | } |
1578 | ||
13d60f4b ZY |
1579 | pgoff_t __basepage_index(struct page *page) |
1580 | { | |
1581 | struct page *page_head = compound_head(page); | |
1582 | pgoff_t index = page_index(page_head); | |
1583 | unsigned long compound_idx; | |
1584 | ||
1585 | if (!PageHuge(page_head)) | |
1586 | return page_index(page); | |
1587 | ||
1588 | if (compound_order(page_head) >= MAX_ORDER) | |
1589 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1590 | else | |
1591 | compound_idx = page - page_head; | |
1592 | ||
1593 | return (index << compound_order(page_head)) + compound_idx; | |
1594 | } | |
1595 | ||
0c397dae | 1596 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1597 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1598 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1599 | { |
af0fb9df | 1600 | int order = huge_page_order(h); |
1da177e4 | 1601 | struct page *page; |
f60858f9 | 1602 | bool alloc_try_hard = true; |
f96efd58 | 1603 | |
f60858f9 MK |
1604 | /* |
1605 | * By default we always try hard to allocate the page with | |
1606 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1607 | * a loop (to adjust global huge page counts) and previous allocation | |
1608 | * failed, do not continue to try hard on the same node. Use the | |
1609 | * node_alloc_noretry bitmap to manage this state information. | |
1610 | */ | |
1611 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1612 | alloc_try_hard = false; | |
1613 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1614 | if (alloc_try_hard) | |
1615 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1616 | if (nid == NUMA_NO_NODE) |
1617 | nid = numa_mem_id(); | |
84172f4b | 1618 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
af0fb9df MH |
1619 | if (page) |
1620 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1621 | else | |
1622 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1623 | |
f60858f9 MK |
1624 | /* |
1625 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1626 | * indicates an overall state change. Clear bit so that we resume | |
1627 | * normal 'try hard' allocations. | |
1628 | */ | |
1629 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1630 | node_clear(nid, *node_alloc_noretry); | |
1631 | ||
1632 | /* | |
1633 | * If we tried hard to get a page but failed, set bit so that | |
1634 | * subsequent attempts will not try as hard until there is an | |
1635 | * overall state change. | |
1636 | */ | |
1637 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1638 | node_set(nid, *node_alloc_noretry); | |
1639 | ||
63b4613c NA |
1640 | return page; |
1641 | } | |
1642 | ||
0c397dae MH |
1643 | /* |
1644 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1645 | * should use this function to get new hugetlb pages | |
1646 | */ | |
1647 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1648 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1649 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1650 | { |
1651 | struct page *page; | |
1652 | ||
1653 | if (hstate_is_gigantic(h)) | |
1654 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1655 | else | |
1656 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1657 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1658 | if (!page) |
1659 | return NULL; | |
1660 | ||
1661 | if (hstate_is_gigantic(h)) | |
1662 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1663 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1664 | ||
1665 | return page; | |
1666 | } | |
1667 | ||
af0fb9df MH |
1668 | /* |
1669 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1670 | * manner. | |
1671 | */ | |
f60858f9 MK |
1672 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1673 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1674 | { |
1675 | struct page *page; | |
1676 | int nr_nodes, node; | |
af0fb9df | 1677 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1678 | |
1679 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1680 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1681 | node_alloc_noretry); | |
af0fb9df | 1682 | if (page) |
b2261026 | 1683 | break; |
b2261026 JK |
1684 | } |
1685 | ||
af0fb9df MH |
1686 | if (!page) |
1687 | return 0; | |
b2261026 | 1688 | |
af0fb9df MH |
1689 | put_page(page); /* free it into the hugepage allocator */ |
1690 | ||
1691 | return 1; | |
b2261026 JK |
1692 | } |
1693 | ||
e8c5c824 LS |
1694 | /* |
1695 | * Free huge page from pool from next node to free. | |
1696 | * Attempt to keep persistent huge pages more or less | |
1697 | * balanced over allowed nodes. | |
1698 | * Called with hugetlb_lock locked. | |
1699 | */ | |
6ae11b27 LS |
1700 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1701 | bool acct_surplus) | |
e8c5c824 | 1702 | { |
b2261026 | 1703 | int nr_nodes, node; |
e8c5c824 LS |
1704 | int ret = 0; |
1705 | ||
b2261026 | 1706 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1707 | /* |
1708 | * If we're returning unused surplus pages, only examine | |
1709 | * nodes with surplus pages. | |
1710 | */ | |
b2261026 JK |
1711 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1712 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1713 | struct page *page = |
b2261026 | 1714 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1715 | struct page, lru); |
1716 | list_del(&page->lru); | |
1717 | h->free_huge_pages--; | |
b2261026 | 1718 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1719 | if (acct_surplus) { |
1720 | h->surplus_huge_pages--; | |
b2261026 | 1721 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1722 | } |
e8c5c824 LS |
1723 | update_and_free_page(h, page); |
1724 | ret = 1; | |
9a76db09 | 1725 | break; |
e8c5c824 | 1726 | } |
b2261026 | 1727 | } |
e8c5c824 LS |
1728 | |
1729 | return ret; | |
1730 | } | |
1731 | ||
c8721bbb NH |
1732 | /* |
1733 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1734 | * nothing for in-use hugepages and non-hugepages. |
1735 | * This function returns values like below: | |
1736 | * | |
1737 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1738 | * (allocated or reserved.) | |
1739 | * 0: successfully dissolved free hugepages or the page is not a | |
1740 | * hugepage (considered as already dissolved) | |
c8721bbb | 1741 | */ |
c3114a84 | 1742 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1743 | { |
6bc9b564 | 1744 | int rc = -EBUSY; |
082d5b6b | 1745 | |
7ffddd49 | 1746 | retry: |
faf53def NH |
1747 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1748 | if (!PageHuge(page)) | |
1749 | return 0; | |
1750 | ||
c8721bbb | 1751 | spin_lock(&hugetlb_lock); |
faf53def NH |
1752 | if (!PageHuge(page)) { |
1753 | rc = 0; | |
1754 | goto out; | |
1755 | } | |
1756 | ||
1757 | if (!page_count(page)) { | |
2247bb33 GS |
1758 | struct page *head = compound_head(page); |
1759 | struct hstate *h = page_hstate(head); | |
1760 | int nid = page_to_nid(head); | |
6bc9b564 | 1761 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1762 | goto out; |
7ffddd49 MS |
1763 | |
1764 | /* | |
1765 | * We should make sure that the page is already on the free list | |
1766 | * when it is dissolved. | |
1767 | */ | |
6c037149 | 1768 | if (unlikely(!HPageFreed(head))) { |
7ffddd49 MS |
1769 | spin_unlock(&hugetlb_lock); |
1770 | cond_resched(); | |
1771 | ||
1772 | /* | |
1773 | * Theoretically, we should return -EBUSY when we | |
1774 | * encounter this race. In fact, we have a chance | |
1775 | * to successfully dissolve the page if we do a | |
1776 | * retry. Because the race window is quite small. | |
1777 | * If we seize this opportunity, it is an optimization | |
1778 | * for increasing the success rate of dissolving page. | |
1779 | */ | |
1780 | goto retry; | |
1781 | } | |
1782 | ||
c3114a84 AK |
1783 | /* |
1784 | * Move PageHWPoison flag from head page to the raw error page, | |
1785 | * which makes any subpages rather than the error page reusable. | |
1786 | */ | |
1787 | if (PageHWPoison(head) && page != head) { | |
1788 | SetPageHWPoison(page); | |
1789 | ClearPageHWPoison(head); | |
1790 | } | |
2247bb33 | 1791 | list_del(&head->lru); |
c8721bbb NH |
1792 | h->free_huge_pages--; |
1793 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1794 | h->max_huge_pages--; |
2247bb33 | 1795 | update_and_free_page(h, head); |
6bc9b564 | 1796 | rc = 0; |
c8721bbb | 1797 | } |
082d5b6b | 1798 | out: |
c8721bbb | 1799 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1800 | return rc; |
c8721bbb NH |
1801 | } |
1802 | ||
1803 | /* | |
1804 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1805 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1806 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1807 | * part of it lies within the given range. | |
082d5b6b GS |
1808 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1809 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1810 | */ |
082d5b6b | 1811 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1812 | { |
c8721bbb | 1813 | unsigned long pfn; |
eb03aa00 | 1814 | struct page *page; |
082d5b6b | 1815 | int rc = 0; |
c8721bbb | 1816 | |
d0177639 | 1817 | if (!hugepages_supported()) |
082d5b6b | 1818 | return rc; |
d0177639 | 1819 | |
eb03aa00 GS |
1820 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1821 | page = pfn_to_page(pfn); | |
faf53def NH |
1822 | rc = dissolve_free_huge_page(page); |
1823 | if (rc) | |
1824 | break; | |
eb03aa00 | 1825 | } |
082d5b6b GS |
1826 | |
1827 | return rc; | |
c8721bbb NH |
1828 | } |
1829 | ||
ab5ac90a MH |
1830 | /* |
1831 | * Allocates a fresh surplus page from the page allocator. | |
1832 | */ | |
0c397dae | 1833 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1834 | int nid, nodemask_t *nmask) |
7893d1d5 | 1835 | { |
9980d744 | 1836 | struct page *page = NULL; |
7893d1d5 | 1837 | |
bae7f4ae | 1838 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1839 | return NULL; |
1840 | ||
d1c3fb1f | 1841 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1842 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1843 | goto out_unlock; | |
d1c3fb1f NA |
1844 | spin_unlock(&hugetlb_lock); |
1845 | ||
f60858f9 | 1846 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1847 | if (!page) |
0c397dae | 1848 | return NULL; |
d1c3fb1f NA |
1849 | |
1850 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1851 | /* |
1852 | * We could have raced with the pool size change. | |
1853 | * Double check that and simply deallocate the new page | |
1854 | * if we would end up overcommiting the surpluses. Abuse | |
1855 | * temporary page to workaround the nasty free_huge_page | |
1856 | * codeflow | |
1857 | */ | |
1858 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 1859 | SetHPageTemporary(page); |
2bf753e6 | 1860 | spin_unlock(&hugetlb_lock); |
9980d744 | 1861 | put_page(page); |
2bf753e6 | 1862 | return NULL; |
9980d744 | 1863 | } else { |
9980d744 | 1864 | h->surplus_huge_pages++; |
4704dea3 | 1865 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1866 | } |
9980d744 MH |
1867 | |
1868 | out_unlock: | |
d1c3fb1f | 1869 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1870 | |
1871 | return page; | |
1872 | } | |
1873 | ||
bbe88753 | 1874 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 1875 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
1876 | { |
1877 | struct page *page; | |
1878 | ||
1879 | if (hstate_is_gigantic(h)) | |
1880 | return NULL; | |
1881 | ||
f60858f9 | 1882 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1883 | if (!page) |
1884 | return NULL; | |
1885 | ||
1886 | /* | |
1887 | * We do not account these pages as surplus because they are only | |
1888 | * temporary and will be released properly on the last reference | |
1889 | */ | |
9157c311 | 1890 | SetHPageTemporary(page); |
ab5ac90a MH |
1891 | |
1892 | return page; | |
1893 | } | |
1894 | ||
099730d6 DH |
1895 | /* |
1896 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1897 | */ | |
e0ec90ee | 1898 | static |
0c397dae | 1899 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1900 | struct vm_area_struct *vma, unsigned long addr) |
1901 | { | |
aaf14e40 MH |
1902 | struct page *page; |
1903 | struct mempolicy *mpol; | |
1904 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1905 | int nid; | |
1906 | nodemask_t *nodemask; | |
1907 | ||
1908 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1909 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1910 | mpol_cond_put(mpol); |
1911 | ||
1912 | return page; | |
099730d6 DH |
1913 | } |
1914 | ||
ab5ac90a | 1915 | /* page migration callback function */ |
3e59fcb0 | 1916 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 1917 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 1918 | { |
4db9b2ef MH |
1919 | spin_lock(&hugetlb_lock); |
1920 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1921 | struct page *page; |
1922 | ||
1923 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1924 | if (page) { | |
1925 | spin_unlock(&hugetlb_lock); | |
1926 | return page; | |
4db9b2ef MH |
1927 | } |
1928 | } | |
1929 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1930 | |
0c397dae | 1931 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1932 | } |
1933 | ||
ebd63723 | 1934 | /* mempolicy aware migration callback */ |
389c8178 MH |
1935 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1936 | unsigned long address) | |
ebd63723 MH |
1937 | { |
1938 | struct mempolicy *mpol; | |
1939 | nodemask_t *nodemask; | |
1940 | struct page *page; | |
ebd63723 MH |
1941 | gfp_t gfp_mask; |
1942 | int node; | |
1943 | ||
ebd63723 MH |
1944 | gfp_mask = htlb_alloc_mask(h); |
1945 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 1946 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
1947 | mpol_cond_put(mpol); |
1948 | ||
1949 | return page; | |
1950 | } | |
1951 | ||
e4e574b7 | 1952 | /* |
25985edc | 1953 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1954 | * of size 'delta'. |
1955 | */ | |
0a4f3d1b | 1956 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 1957 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
1958 | { |
1959 | struct list_head surplus_list; | |
1960 | struct page *page, *tmp; | |
0a4f3d1b LX |
1961 | int ret; |
1962 | long i; | |
1963 | long needed, allocated; | |
28073b02 | 1964 | bool alloc_ok = true; |
e4e574b7 | 1965 | |
a5516438 | 1966 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1967 | if (needed <= 0) { |
a5516438 | 1968 | h->resv_huge_pages += delta; |
e4e574b7 | 1969 | return 0; |
ac09b3a1 | 1970 | } |
e4e574b7 AL |
1971 | |
1972 | allocated = 0; | |
1973 | INIT_LIST_HEAD(&surplus_list); | |
1974 | ||
1975 | ret = -ENOMEM; | |
1976 | retry: | |
1977 | spin_unlock(&hugetlb_lock); | |
1978 | for (i = 0; i < needed; i++) { | |
0c397dae | 1979 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1980 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1981 | if (!page) { |
1982 | alloc_ok = false; | |
1983 | break; | |
1984 | } | |
e4e574b7 | 1985 | list_add(&page->lru, &surplus_list); |
69ed779a | 1986 | cond_resched(); |
e4e574b7 | 1987 | } |
28073b02 | 1988 | allocated += i; |
e4e574b7 AL |
1989 | |
1990 | /* | |
1991 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1992 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1993 | */ | |
1994 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
1995 | needed = (h->resv_huge_pages + delta) - |
1996 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1997 | if (needed > 0) { |
1998 | if (alloc_ok) | |
1999 | goto retry; | |
2000 | /* | |
2001 | * We were not able to allocate enough pages to | |
2002 | * satisfy the entire reservation so we free what | |
2003 | * we've allocated so far. | |
2004 | */ | |
2005 | goto free; | |
2006 | } | |
e4e574b7 AL |
2007 | /* |
2008 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2009 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2010 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2011 | * allocator. Commit the entire reservation here to prevent another |
2012 | * process from stealing the pages as they are added to the pool but | |
2013 | * before they are reserved. | |
e4e574b7 AL |
2014 | */ |
2015 | needed += allocated; | |
a5516438 | 2016 | h->resv_huge_pages += delta; |
e4e574b7 | 2017 | ret = 0; |
a9869b83 | 2018 | |
19fc3f0a | 2019 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2020 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
e558464b MS |
2021 | int zeroed; |
2022 | ||
19fc3f0a AL |
2023 | if ((--needed) < 0) |
2024 | break; | |
a9869b83 NH |
2025 | /* |
2026 | * This page is now managed by the hugetlb allocator and has | |
2027 | * no users -- drop the buddy allocator's reference. | |
2028 | */ | |
e558464b MS |
2029 | zeroed = put_page_testzero(page); |
2030 | VM_BUG_ON_PAGE(!zeroed, page); | |
a5516438 | 2031 | enqueue_huge_page(h, page); |
19fc3f0a | 2032 | } |
28073b02 | 2033 | free: |
b0365c8d | 2034 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
2035 | |
2036 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
2037 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
2038 | put_page(page); | |
a9869b83 | 2039 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
2040 | |
2041 | return ret; | |
2042 | } | |
2043 | ||
2044 | /* | |
e5bbc8a6 MK |
2045 | * This routine has two main purposes: |
2046 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2047 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2048 | * to the associated reservation map. | |
2049 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2050 | * the reservation. As many as unused_resv_pages may be freed. | |
2051 | * | |
2052 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
2053 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
2054 | * we must make sure nobody else can claim pages we are in the process of | |
2055 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
2056 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 2057 | */ |
a5516438 AK |
2058 | static void return_unused_surplus_pages(struct hstate *h, |
2059 | unsigned long unused_resv_pages) | |
e4e574b7 | 2060 | { |
e4e574b7 AL |
2061 | unsigned long nr_pages; |
2062 | ||
aa888a74 | 2063 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2064 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2065 | goto out; |
aa888a74 | 2066 | |
e5bbc8a6 MK |
2067 | /* |
2068 | * Part (or even all) of the reservation could have been backed | |
2069 | * by pre-allocated pages. Only free surplus pages. | |
2070 | */ | |
a5516438 | 2071 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2072 | |
685f3457 LS |
2073 | /* |
2074 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2075 | * evenly across all nodes with memory. Iterate across these nodes |
2076 | * until we can no longer free unreserved surplus pages. This occurs | |
2077 | * when the nodes with surplus pages have no free pages. | |
9e7ee400 | 2078 | * free_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2079 | * on-line nodes with memory and will handle the hstate accounting. |
e5bbc8a6 MK |
2080 | * |
2081 | * Note that we decrement resv_huge_pages as we free the pages. If | |
2082 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
2083 | * to cover subsequent pages we may free. | |
685f3457 LS |
2084 | */ |
2085 | while (nr_pages--) { | |
e5bbc8a6 MK |
2086 | h->resv_huge_pages--; |
2087 | unused_resv_pages--; | |
8cebfcd0 | 2088 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 2089 | goto out; |
7848a4bf | 2090 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 2091 | } |
e5bbc8a6 MK |
2092 | |
2093 | out: | |
2094 | /* Fully uncommit the reservation */ | |
2095 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
2096 | } |
2097 | ||
5e911373 | 2098 | |
c37f9fb1 | 2099 | /* |
feba16e2 | 2100 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2101 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2102 | * |
2103 | * vma_needs_reservation is called to determine if the huge page at addr | |
2104 | * within the vma has an associated reservation. If a reservation is | |
2105 | * needed, the value 1 is returned. The caller is then responsible for | |
2106 | * managing the global reservation and subpool usage counts. After | |
2107 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2108 | * to add the page to the reservation map. If the page allocation fails, |
2109 | * the reservation must be ended instead of committed. vma_end_reservation | |
2110 | * is called in such cases. | |
cf3ad20b MK |
2111 | * |
2112 | * In the normal case, vma_commit_reservation returns the same value | |
2113 | * as the preceding vma_needs_reservation call. The only time this | |
2114 | * is not the case is if a reserve map was changed between calls. It | |
2115 | * is the responsibility of the caller to notice the difference and | |
2116 | * take appropriate action. | |
96b96a96 MK |
2117 | * |
2118 | * vma_add_reservation is used in error paths where a reservation must | |
2119 | * be restored when a newly allocated huge page must be freed. It is | |
2120 | * to be called after calling vma_needs_reservation to determine if a | |
2121 | * reservation exists. | |
c37f9fb1 | 2122 | */ |
5e911373 MK |
2123 | enum vma_resv_mode { |
2124 | VMA_NEEDS_RESV, | |
2125 | VMA_COMMIT_RESV, | |
feba16e2 | 2126 | VMA_END_RESV, |
96b96a96 | 2127 | VMA_ADD_RESV, |
5e911373 | 2128 | }; |
cf3ad20b MK |
2129 | static long __vma_reservation_common(struct hstate *h, |
2130 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2131 | enum vma_resv_mode mode) |
c37f9fb1 | 2132 | { |
4e35f483 JK |
2133 | struct resv_map *resv; |
2134 | pgoff_t idx; | |
cf3ad20b | 2135 | long ret; |
0db9d74e | 2136 | long dummy_out_regions_needed; |
c37f9fb1 | 2137 | |
4e35f483 JK |
2138 | resv = vma_resv_map(vma); |
2139 | if (!resv) | |
84afd99b | 2140 | return 1; |
c37f9fb1 | 2141 | |
4e35f483 | 2142 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2143 | switch (mode) { |
2144 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2145 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2146 | /* We assume that vma_reservation_* routines always operate on | |
2147 | * 1 page, and that adding to resv map a 1 page entry can only | |
2148 | * ever require 1 region. | |
2149 | */ | |
2150 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2151 | break; |
2152 | case VMA_COMMIT_RESV: | |
075a61d0 | 2153 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2154 | /* region_add calls of range 1 should never fail. */ |
2155 | VM_BUG_ON(ret < 0); | |
5e911373 | 2156 | break; |
feba16e2 | 2157 | case VMA_END_RESV: |
0db9d74e | 2158 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2159 | ret = 0; |
2160 | break; | |
96b96a96 | 2161 | case VMA_ADD_RESV: |
0db9d74e | 2162 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2163 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2164 | /* region_add calls of range 1 should never fail. */ |
2165 | VM_BUG_ON(ret < 0); | |
2166 | } else { | |
2167 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2168 | ret = region_del(resv, idx, idx + 1); |
2169 | } | |
2170 | break; | |
5e911373 MK |
2171 | default: |
2172 | BUG(); | |
2173 | } | |
84afd99b | 2174 | |
4e35f483 | 2175 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 2176 | return ret; |
67961f9d MK |
2177 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
2178 | /* | |
2179 | * In most cases, reserves always exist for private mappings. | |
2180 | * However, a file associated with mapping could have been | |
2181 | * hole punched or truncated after reserves were consumed. | |
2182 | * As subsequent fault on such a range will not use reserves. | |
2183 | * Subtle - The reserve map for private mappings has the | |
2184 | * opposite meaning than that of shared mappings. If NO | |
2185 | * entry is in the reserve map, it means a reservation exists. | |
2186 | * If an entry exists in the reserve map, it means the | |
2187 | * reservation has already been consumed. As a result, the | |
2188 | * return value of this routine is the opposite of the | |
2189 | * value returned from reserve map manipulation routines above. | |
2190 | */ | |
2191 | if (ret) | |
2192 | return 0; | |
2193 | else | |
2194 | return 1; | |
2195 | } | |
4e35f483 | 2196 | else |
cf3ad20b | 2197 | return ret < 0 ? ret : 0; |
c37f9fb1 | 2198 | } |
cf3ad20b MK |
2199 | |
2200 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2201 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2202 | { |
5e911373 | 2203 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2204 | } |
84afd99b | 2205 | |
cf3ad20b MK |
2206 | static long vma_commit_reservation(struct hstate *h, |
2207 | struct vm_area_struct *vma, unsigned long addr) | |
2208 | { | |
5e911373 MK |
2209 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2210 | } | |
2211 | ||
feba16e2 | 2212 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2213 | struct vm_area_struct *vma, unsigned long addr) |
2214 | { | |
feba16e2 | 2215 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2216 | } |
2217 | ||
96b96a96 MK |
2218 | static long vma_add_reservation(struct hstate *h, |
2219 | struct vm_area_struct *vma, unsigned long addr) | |
2220 | { | |
2221 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2222 | } | |
2223 | ||
2224 | /* | |
2225 | * This routine is called to restore a reservation on error paths. In the | |
2226 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2227 | * and is about to be freed. If a reservation for the page existed, | |
d6995da3 MK |
2228 | * alloc_huge_page would have consumed the reservation and set |
2229 | * HPageRestoreReserve in the newly allocated page. When the page is freed | |
2230 | * via free_huge_page, the global reservation count will be incremented if | |
2231 | * HPageRestoreReserve is set. However, free_huge_page can not adjust the | |
2232 | * reserve map. Adjust the reserve map here to be consistent with global | |
2233 | * reserve count adjustments to be made by free_huge_page. | |
96b96a96 MK |
2234 | */ |
2235 | static void restore_reserve_on_error(struct hstate *h, | |
2236 | struct vm_area_struct *vma, unsigned long address, | |
2237 | struct page *page) | |
2238 | { | |
d6995da3 | 2239 | if (unlikely(HPageRestoreReserve(page))) { |
96b96a96 MK |
2240 | long rc = vma_needs_reservation(h, vma, address); |
2241 | ||
2242 | if (unlikely(rc < 0)) { | |
2243 | /* | |
2244 | * Rare out of memory condition in reserve map | |
d6995da3 | 2245 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2246 | * global reserve count will not be incremented |
2247 | * by free_huge_page. This will make it appear | |
2248 | * as though the reservation for this page was | |
2249 | * consumed. This may prevent the task from | |
2250 | * faulting in the page at a later time. This | |
2251 | * is better than inconsistent global huge page | |
2252 | * accounting of reserve counts. | |
2253 | */ | |
d6995da3 | 2254 | ClearHPageRestoreReserve(page); |
96b96a96 MK |
2255 | } else if (rc) { |
2256 | rc = vma_add_reservation(h, vma, address); | |
2257 | if (unlikely(rc < 0)) | |
2258 | /* | |
2259 | * See above comment about rare out of | |
2260 | * memory condition. | |
2261 | */ | |
d6995da3 | 2262 | ClearHPageRestoreReserve(page); |
96b96a96 MK |
2263 | } else |
2264 | vma_end_reservation(h, vma, address); | |
2265 | } | |
2266 | } | |
2267 | ||
70c3547e | 2268 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2269 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2270 | { |
90481622 | 2271 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2272 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2273 | struct page *page; |
d85f69b0 MK |
2274 | long map_chg, map_commit; |
2275 | long gbl_chg; | |
6d76dcf4 AK |
2276 | int ret, idx; |
2277 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2278 | bool deferred_reserve; |
a1e78772 | 2279 | |
6d76dcf4 | 2280 | idx = hstate_index(h); |
a1e78772 | 2281 | /* |
d85f69b0 MK |
2282 | * Examine the region/reserve map to determine if the process |
2283 | * has a reservation for the page to be allocated. A return | |
2284 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2285 | */ |
d85f69b0 MK |
2286 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2287 | if (map_chg < 0) | |
76dcee75 | 2288 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2289 | |
2290 | /* | |
2291 | * Processes that did not create the mapping will have no | |
2292 | * reserves as indicated by the region/reserve map. Check | |
2293 | * that the allocation will not exceed the subpool limit. | |
2294 | * Allocations for MAP_NORESERVE mappings also need to be | |
2295 | * checked against any subpool limit. | |
2296 | */ | |
2297 | if (map_chg || avoid_reserve) { | |
2298 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2299 | if (gbl_chg < 0) { | |
feba16e2 | 2300 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2301 | return ERR_PTR(-ENOSPC); |
5e911373 | 2302 | } |
1da177e4 | 2303 | |
d85f69b0 MK |
2304 | /* |
2305 | * Even though there was no reservation in the region/reserve | |
2306 | * map, there could be reservations associated with the | |
2307 | * subpool that can be used. This would be indicated if the | |
2308 | * return value of hugepage_subpool_get_pages() is zero. | |
2309 | * However, if avoid_reserve is specified we still avoid even | |
2310 | * the subpool reservations. | |
2311 | */ | |
2312 | if (avoid_reserve) | |
2313 | gbl_chg = 1; | |
2314 | } | |
2315 | ||
08cf9faf MA |
2316 | /* If this allocation is not consuming a reservation, charge it now. |
2317 | */ | |
6501fe5f | 2318 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2319 | if (deferred_reserve) { |
2320 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2321 | idx, pages_per_huge_page(h), &h_cg); | |
2322 | if (ret) | |
2323 | goto out_subpool_put; | |
2324 | } | |
2325 | ||
6d76dcf4 | 2326 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2327 | if (ret) |
08cf9faf | 2328 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2329 | |
1da177e4 | 2330 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2331 | /* |
2332 | * glb_chg is passed to indicate whether or not a page must be taken | |
2333 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2334 | * a reservation exists for the allocation. | |
2335 | */ | |
2336 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2337 | if (!page) { |
94ae8ba7 | 2338 | spin_unlock(&hugetlb_lock); |
0c397dae | 2339 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2340 | if (!page) |
2341 | goto out_uncharge_cgroup; | |
a88c7695 | 2342 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2343 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2344 | h->resv_huge_pages--; |
2345 | } | |
79dbb236 | 2346 | spin_lock(&hugetlb_lock); |
15a8d68e | 2347 | list_add(&page->lru, &h->hugepage_activelist); |
81a6fcae | 2348 | /* Fall through */ |
68842c9b | 2349 | } |
81a6fcae | 2350 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2351 | /* If allocation is not consuming a reservation, also store the |
2352 | * hugetlb_cgroup pointer on the page. | |
2353 | */ | |
2354 | if (deferred_reserve) { | |
2355 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2356 | h_cg, page); | |
2357 | } | |
2358 | ||
81a6fcae | 2359 | spin_unlock(&hugetlb_lock); |
348ea204 | 2360 | |
d6995da3 | 2361 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2362 | |
d85f69b0 MK |
2363 | map_commit = vma_commit_reservation(h, vma, addr); |
2364 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2365 | /* |
2366 | * The page was added to the reservation map between | |
2367 | * vma_needs_reservation and vma_commit_reservation. | |
2368 | * This indicates a race with hugetlb_reserve_pages. | |
2369 | * Adjust for the subpool count incremented above AND | |
2370 | * in hugetlb_reserve_pages for the same page. Also, | |
2371 | * the reservation count added in hugetlb_reserve_pages | |
2372 | * no longer applies. | |
2373 | */ | |
2374 | long rsv_adjust; | |
2375 | ||
2376 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2377 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2378 | if (deferred_reserve) |
2379 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2380 | pages_per_huge_page(h), page); | |
33039678 | 2381 | } |
90d8b7e6 | 2382 | return page; |
8f34af6f JZ |
2383 | |
2384 | out_uncharge_cgroup: | |
2385 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2386 | out_uncharge_cgroup_reservation: |
2387 | if (deferred_reserve) | |
2388 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2389 | h_cg); | |
8f34af6f | 2390 | out_subpool_put: |
d85f69b0 | 2391 | if (map_chg || avoid_reserve) |
8f34af6f | 2392 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2393 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2394 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2395 | } |
2396 | ||
e24a1307 AK |
2397 | int alloc_bootmem_huge_page(struct hstate *h) |
2398 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2399 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2400 | { |
2401 | struct huge_bootmem_page *m; | |
b2261026 | 2402 | int nr_nodes, node; |
aa888a74 | 2403 | |
b2261026 | 2404 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2405 | void *addr; |
2406 | ||
eb31d559 | 2407 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2408 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2409 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2410 | if (addr) { |
2411 | /* | |
2412 | * Use the beginning of the huge page to store the | |
2413 | * huge_bootmem_page struct (until gather_bootmem | |
2414 | * puts them into the mem_map). | |
2415 | */ | |
2416 | m = addr; | |
91f47662 | 2417 | goto found; |
aa888a74 | 2418 | } |
aa888a74 AK |
2419 | } |
2420 | return 0; | |
2421 | ||
2422 | found: | |
df994ead | 2423 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2424 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2425 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2426 | list_add(&m->list, &huge_boot_pages); |
2427 | m->hstate = h; | |
2428 | return 1; | |
2429 | } | |
2430 | ||
d00181b9 KS |
2431 | static void __init prep_compound_huge_page(struct page *page, |
2432 | unsigned int order) | |
18229df5 AW |
2433 | { |
2434 | if (unlikely(order > (MAX_ORDER - 1))) | |
2435 | prep_compound_gigantic_page(page, order); | |
2436 | else | |
2437 | prep_compound_page(page, order); | |
2438 | } | |
2439 | ||
aa888a74 AK |
2440 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2441 | static void __init gather_bootmem_prealloc(void) | |
2442 | { | |
2443 | struct huge_bootmem_page *m; | |
2444 | ||
2445 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2446 | struct page *page = virt_to_page(m); |
aa888a74 | 2447 | struct hstate *h = m->hstate; |
ee8f248d | 2448 | |
aa888a74 | 2449 | WARN_ON(page_count(page) != 1); |
c78a7f36 | 2450 | prep_compound_huge_page(page, huge_page_order(h)); |
ef5a22be | 2451 | WARN_ON(PageReserved(page)); |
aa888a74 | 2452 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2453 | put_page(page); /* free it into the hugepage allocator */ |
2454 | ||
b0320c7b RA |
2455 | /* |
2456 | * If we had gigantic hugepages allocated at boot time, we need | |
2457 | * to restore the 'stolen' pages to totalram_pages in order to | |
2458 | * fix confusing memory reports from free(1) and another | |
2459 | * side-effects, like CommitLimit going negative. | |
2460 | */ | |
bae7f4ae | 2461 | if (hstate_is_gigantic(h)) |
c78a7f36 | 2462 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 2463 | cond_resched(); |
aa888a74 AK |
2464 | } |
2465 | } | |
2466 | ||
8faa8b07 | 2467 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2468 | { |
2469 | unsigned long i; | |
f60858f9 MK |
2470 | nodemask_t *node_alloc_noretry; |
2471 | ||
2472 | if (!hstate_is_gigantic(h)) { | |
2473 | /* | |
2474 | * Bit mask controlling how hard we retry per-node allocations. | |
2475 | * Ignore errors as lower level routines can deal with | |
2476 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2477 | * time, we are likely in bigger trouble. | |
2478 | */ | |
2479 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2480 | GFP_KERNEL); | |
2481 | } else { | |
2482 | /* allocations done at boot time */ | |
2483 | node_alloc_noretry = NULL; | |
2484 | } | |
2485 | ||
2486 | /* bit mask controlling how hard we retry per-node allocations */ | |
2487 | if (node_alloc_noretry) | |
2488 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2489 | |
e5ff2159 | 2490 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2491 | if (hstate_is_gigantic(h)) { |
dbda8fea | 2492 | if (hugetlb_cma_size) { |
cf11e85f | 2493 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); |
7ecc9565 | 2494 | goto free; |
cf11e85f | 2495 | } |
aa888a74 AK |
2496 | if (!alloc_bootmem_huge_page(h)) |
2497 | break; | |
0c397dae | 2498 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2499 | &node_states[N_MEMORY], |
2500 | node_alloc_noretry)) | |
1da177e4 | 2501 | break; |
69ed779a | 2502 | cond_resched(); |
1da177e4 | 2503 | } |
d715cf80 LH |
2504 | if (i < h->max_huge_pages) { |
2505 | char buf[32]; | |
2506 | ||
c6247f72 | 2507 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2508 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2509 | h->max_huge_pages, buf, i); | |
2510 | h->max_huge_pages = i; | |
2511 | } | |
7ecc9565 | 2512 | free: |
f60858f9 | 2513 | kfree(node_alloc_noretry); |
e5ff2159 AK |
2514 | } |
2515 | ||
2516 | static void __init hugetlb_init_hstates(void) | |
2517 | { | |
2518 | struct hstate *h; | |
2519 | ||
2520 | for_each_hstate(h) { | |
641844f5 NH |
2521 | if (minimum_order > huge_page_order(h)) |
2522 | minimum_order = huge_page_order(h); | |
2523 | ||
8faa8b07 | 2524 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2525 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2526 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2527 | } |
641844f5 | 2528 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2529 | } |
2530 | ||
2531 | static void __init report_hugepages(void) | |
2532 | { | |
2533 | struct hstate *h; | |
2534 | ||
2535 | for_each_hstate(h) { | |
4abd32db | 2536 | char buf[32]; |
c6247f72 MW |
2537 | |
2538 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2539 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2540 | buf, h->free_huge_pages); |
e5ff2159 AK |
2541 | } |
2542 | } | |
2543 | ||
1da177e4 | 2544 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2545 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2546 | nodemask_t *nodes_allowed) | |
1da177e4 | 2547 | { |
4415cc8d CL |
2548 | int i; |
2549 | ||
bae7f4ae | 2550 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2551 | return; |
2552 | ||
6ae11b27 | 2553 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2554 | struct page *page, *next; |
a5516438 AK |
2555 | struct list_head *freel = &h->hugepage_freelists[i]; |
2556 | list_for_each_entry_safe(page, next, freel, lru) { | |
2557 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2558 | return; |
1da177e4 LT |
2559 | if (PageHighMem(page)) |
2560 | continue; | |
2561 | list_del(&page->lru); | |
e5ff2159 | 2562 | update_and_free_page(h, page); |
a5516438 AK |
2563 | h->free_huge_pages--; |
2564 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2565 | } |
2566 | } | |
2567 | } | |
2568 | #else | |
6ae11b27 LS |
2569 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2570 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2571 | { |
2572 | } | |
2573 | #endif | |
2574 | ||
20a0307c WF |
2575 | /* |
2576 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2577 | * balanced by operating on them in a round-robin fashion. | |
2578 | * Returns 1 if an adjustment was made. | |
2579 | */ | |
6ae11b27 LS |
2580 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2581 | int delta) | |
20a0307c | 2582 | { |
b2261026 | 2583 | int nr_nodes, node; |
20a0307c WF |
2584 | |
2585 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2586 | |
b2261026 JK |
2587 | if (delta < 0) { |
2588 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2589 | if (h->surplus_huge_pages_node[node]) | |
2590 | goto found; | |
e8c5c824 | 2591 | } |
b2261026 JK |
2592 | } else { |
2593 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2594 | if (h->surplus_huge_pages_node[node] < | |
2595 | h->nr_huge_pages_node[node]) | |
2596 | goto found; | |
e8c5c824 | 2597 | } |
b2261026 JK |
2598 | } |
2599 | return 0; | |
20a0307c | 2600 | |
b2261026 JK |
2601 | found: |
2602 | h->surplus_huge_pages += delta; | |
2603 | h->surplus_huge_pages_node[node] += delta; | |
2604 | return 1; | |
20a0307c WF |
2605 | } |
2606 | ||
a5516438 | 2607 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2608 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2609 | nodemask_t *nodes_allowed) |
1da177e4 | 2610 | { |
7893d1d5 | 2611 | unsigned long min_count, ret; |
f60858f9 MK |
2612 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2613 | ||
2614 | /* | |
2615 | * Bit mask controlling how hard we retry per-node allocations. | |
2616 | * If we can not allocate the bit mask, do not attempt to allocate | |
2617 | * the requested huge pages. | |
2618 | */ | |
2619 | if (node_alloc_noretry) | |
2620 | nodes_clear(*node_alloc_noretry); | |
2621 | else | |
2622 | return -ENOMEM; | |
1da177e4 | 2623 | |
4eb0716e AG |
2624 | spin_lock(&hugetlb_lock); |
2625 | ||
fd875dca MK |
2626 | /* |
2627 | * Check for a node specific request. | |
2628 | * Changing node specific huge page count may require a corresponding | |
2629 | * change to the global count. In any case, the passed node mask | |
2630 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2631 | */ | |
2632 | if (nid != NUMA_NO_NODE) { | |
2633 | unsigned long old_count = count; | |
2634 | ||
2635 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2636 | /* | |
2637 | * User may have specified a large count value which caused the | |
2638 | * above calculation to overflow. In this case, they wanted | |
2639 | * to allocate as many huge pages as possible. Set count to | |
2640 | * largest possible value to align with their intention. | |
2641 | */ | |
2642 | if (count < old_count) | |
2643 | count = ULONG_MAX; | |
2644 | } | |
2645 | ||
4eb0716e AG |
2646 | /* |
2647 | * Gigantic pages runtime allocation depend on the capability for large | |
2648 | * page range allocation. | |
2649 | * If the system does not provide this feature, return an error when | |
2650 | * the user tries to allocate gigantic pages but let the user free the | |
2651 | * boottime allocated gigantic pages. | |
2652 | */ | |
2653 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2654 | if (count > persistent_huge_pages(h)) { | |
2655 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2656 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2657 | return -EINVAL; |
2658 | } | |
2659 | /* Fall through to decrease pool */ | |
2660 | } | |
aa888a74 | 2661 | |
7893d1d5 AL |
2662 | /* |
2663 | * Increase the pool size | |
2664 | * First take pages out of surplus state. Then make up the | |
2665 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2666 | * |
0c397dae | 2667 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2668 | * to convert a surplus huge page to a normal huge page. That is |
2669 | * not critical, though, it just means the overall size of the | |
2670 | * pool might be one hugepage larger than it needs to be, but | |
2671 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2672 | */ |
a5516438 | 2673 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2674 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2675 | break; |
2676 | } | |
2677 | ||
a5516438 | 2678 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2679 | /* |
2680 | * If this allocation races such that we no longer need the | |
2681 | * page, free_huge_page will handle it by freeing the page | |
2682 | * and reducing the surplus. | |
2683 | */ | |
2684 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2685 | |
2686 | /* yield cpu to avoid soft lockup */ | |
2687 | cond_resched(); | |
2688 | ||
f60858f9 MK |
2689 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2690 | node_alloc_noretry); | |
7893d1d5 AL |
2691 | spin_lock(&hugetlb_lock); |
2692 | if (!ret) | |
2693 | goto out; | |
2694 | ||
536240f2 MG |
2695 | /* Bail for signals. Probably ctrl-c from user */ |
2696 | if (signal_pending(current)) | |
2697 | goto out; | |
7893d1d5 | 2698 | } |
7893d1d5 AL |
2699 | |
2700 | /* | |
2701 | * Decrease the pool size | |
2702 | * First return free pages to the buddy allocator (being careful | |
2703 | * to keep enough around to satisfy reservations). Then place | |
2704 | * pages into surplus state as needed so the pool will shrink | |
2705 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2706 | * |
2707 | * By placing pages into the surplus state independent of the | |
2708 | * overcommit value, we are allowing the surplus pool size to | |
2709 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2710 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2711 | * though, we'll note that we're not allowed to exceed surplus |
2712 | * and won't grow the pool anywhere else. Not until one of the | |
2713 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2714 | */ |
a5516438 | 2715 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2716 | min_count = max(count, min_count); |
6ae11b27 | 2717 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2718 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2719 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2720 | break; |
55f67141 | 2721 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2722 | } |
a5516438 | 2723 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2724 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2725 | break; |
2726 | } | |
2727 | out: | |
4eb0716e | 2728 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2729 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2730 | |
f60858f9 MK |
2731 | NODEMASK_FREE(node_alloc_noretry); |
2732 | ||
4eb0716e | 2733 | return 0; |
1da177e4 LT |
2734 | } |
2735 | ||
a3437870 NA |
2736 | #define HSTATE_ATTR_RO(_name) \ |
2737 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2738 | ||
2739 | #define HSTATE_ATTR(_name) \ | |
2740 | static struct kobj_attribute _name##_attr = \ | |
2741 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2742 | ||
2743 | static struct kobject *hugepages_kobj; | |
2744 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2745 | ||
9a305230 LS |
2746 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2747 | ||
2748 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2749 | { |
2750 | int i; | |
9a305230 | 2751 | |
a3437870 | 2752 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2753 | if (hstate_kobjs[i] == kobj) { |
2754 | if (nidp) | |
2755 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2756 | return &hstates[i]; |
9a305230 LS |
2757 | } |
2758 | ||
2759 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2760 | } |
2761 | ||
06808b08 | 2762 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2763 | struct kobj_attribute *attr, char *buf) |
2764 | { | |
9a305230 LS |
2765 | struct hstate *h; |
2766 | unsigned long nr_huge_pages; | |
2767 | int nid; | |
2768 | ||
2769 | h = kobj_to_hstate(kobj, &nid); | |
2770 | if (nid == NUMA_NO_NODE) | |
2771 | nr_huge_pages = h->nr_huge_pages; | |
2772 | else | |
2773 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2774 | ||
ae7a927d | 2775 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 2776 | } |
adbe8726 | 2777 | |
238d3c13 DR |
2778 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2779 | struct hstate *h, int nid, | |
2780 | unsigned long count, size_t len) | |
a3437870 NA |
2781 | { |
2782 | int err; | |
2d0adf7e | 2783 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2784 | |
2d0adf7e OS |
2785 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2786 | return -EINVAL; | |
adbe8726 | 2787 | |
9a305230 LS |
2788 | if (nid == NUMA_NO_NODE) { |
2789 | /* | |
2790 | * global hstate attribute | |
2791 | */ | |
2792 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2793 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2794 | n_mask = &node_states[N_MEMORY]; | |
2795 | else | |
2796 | n_mask = &nodes_allowed; | |
2797 | } else { | |
9a305230 | 2798 | /* |
fd875dca MK |
2799 | * Node specific request. count adjustment happens in |
2800 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2801 | */ |
2d0adf7e OS |
2802 | init_nodemask_of_node(&nodes_allowed, nid); |
2803 | n_mask = &nodes_allowed; | |
fd875dca | 2804 | } |
9a305230 | 2805 | |
2d0adf7e | 2806 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2807 | |
4eb0716e | 2808 | return err ? err : len; |
06808b08 LS |
2809 | } |
2810 | ||
238d3c13 DR |
2811 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2812 | struct kobject *kobj, const char *buf, | |
2813 | size_t len) | |
2814 | { | |
2815 | struct hstate *h; | |
2816 | unsigned long count; | |
2817 | int nid; | |
2818 | int err; | |
2819 | ||
2820 | err = kstrtoul(buf, 10, &count); | |
2821 | if (err) | |
2822 | return err; | |
2823 | ||
2824 | h = kobj_to_hstate(kobj, &nid); | |
2825 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2826 | } | |
2827 | ||
06808b08 LS |
2828 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2829 | struct kobj_attribute *attr, char *buf) | |
2830 | { | |
2831 | return nr_hugepages_show_common(kobj, attr, buf); | |
2832 | } | |
2833 | ||
2834 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2835 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2836 | { | |
238d3c13 | 2837 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2838 | } |
2839 | HSTATE_ATTR(nr_hugepages); | |
2840 | ||
06808b08 LS |
2841 | #ifdef CONFIG_NUMA |
2842 | ||
2843 | /* | |
2844 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2845 | * huge page alloc/free. | |
2846 | */ | |
2847 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
2848 | struct kobj_attribute *attr, |
2849 | char *buf) | |
06808b08 LS |
2850 | { |
2851 | return nr_hugepages_show_common(kobj, attr, buf); | |
2852 | } | |
2853 | ||
2854 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2855 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2856 | { | |
238d3c13 | 2857 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2858 | } |
2859 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2860 | #endif | |
2861 | ||
2862 | ||
a3437870 NA |
2863 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2864 | struct kobj_attribute *attr, char *buf) | |
2865 | { | |
9a305230 | 2866 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 2867 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 2868 | } |
adbe8726 | 2869 | |
a3437870 NA |
2870 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2871 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2872 | { | |
2873 | int err; | |
2874 | unsigned long input; | |
9a305230 | 2875 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2876 | |
bae7f4ae | 2877 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2878 | return -EINVAL; |
2879 | ||
3dbb95f7 | 2880 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2881 | if (err) |
73ae31e5 | 2882 | return err; |
a3437870 NA |
2883 | |
2884 | spin_lock(&hugetlb_lock); | |
2885 | h->nr_overcommit_huge_pages = input; | |
2886 | spin_unlock(&hugetlb_lock); | |
2887 | ||
2888 | return count; | |
2889 | } | |
2890 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2891 | ||
2892 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2893 | struct kobj_attribute *attr, char *buf) | |
2894 | { | |
9a305230 LS |
2895 | struct hstate *h; |
2896 | unsigned long free_huge_pages; | |
2897 | int nid; | |
2898 | ||
2899 | h = kobj_to_hstate(kobj, &nid); | |
2900 | if (nid == NUMA_NO_NODE) | |
2901 | free_huge_pages = h->free_huge_pages; | |
2902 | else | |
2903 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2904 | ||
ae7a927d | 2905 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
2906 | } |
2907 | HSTATE_ATTR_RO(free_hugepages); | |
2908 | ||
2909 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2910 | struct kobj_attribute *attr, char *buf) | |
2911 | { | |
9a305230 | 2912 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 2913 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
2914 | } |
2915 | HSTATE_ATTR_RO(resv_hugepages); | |
2916 | ||
2917 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2918 | struct kobj_attribute *attr, char *buf) | |
2919 | { | |
9a305230 LS |
2920 | struct hstate *h; |
2921 | unsigned long surplus_huge_pages; | |
2922 | int nid; | |
2923 | ||
2924 | h = kobj_to_hstate(kobj, &nid); | |
2925 | if (nid == NUMA_NO_NODE) | |
2926 | surplus_huge_pages = h->surplus_huge_pages; | |
2927 | else | |
2928 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2929 | ||
ae7a927d | 2930 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
2931 | } |
2932 | HSTATE_ATTR_RO(surplus_hugepages); | |
2933 | ||
2934 | static struct attribute *hstate_attrs[] = { | |
2935 | &nr_hugepages_attr.attr, | |
2936 | &nr_overcommit_hugepages_attr.attr, | |
2937 | &free_hugepages_attr.attr, | |
2938 | &resv_hugepages_attr.attr, | |
2939 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2940 | #ifdef CONFIG_NUMA |
2941 | &nr_hugepages_mempolicy_attr.attr, | |
2942 | #endif | |
a3437870 NA |
2943 | NULL, |
2944 | }; | |
2945 | ||
67e5ed96 | 2946 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2947 | .attrs = hstate_attrs, |
2948 | }; | |
2949 | ||
094e9539 JM |
2950 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2951 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2952 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2953 | { |
2954 | int retval; | |
972dc4de | 2955 | int hi = hstate_index(h); |
a3437870 | 2956 | |
9a305230 LS |
2957 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2958 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2959 | return -ENOMEM; |
2960 | ||
9a305230 | 2961 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 2962 | if (retval) { |
9a305230 | 2963 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 ML |
2964 | hstate_kobjs[hi] = NULL; |
2965 | } | |
a3437870 NA |
2966 | |
2967 | return retval; | |
2968 | } | |
2969 | ||
2970 | static void __init hugetlb_sysfs_init(void) | |
2971 | { | |
2972 | struct hstate *h; | |
2973 | int err; | |
2974 | ||
2975 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2976 | if (!hugepages_kobj) | |
2977 | return; | |
2978 | ||
2979 | for_each_hstate(h) { | |
9a305230 LS |
2980 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2981 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2982 | if (err) |
282f4214 | 2983 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
2984 | } |
2985 | } | |
2986 | ||
9a305230 LS |
2987 | #ifdef CONFIG_NUMA |
2988 | ||
2989 | /* | |
2990 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
2991 | * with node devices in node_devices[] using a parallel array. The array |
2992 | * index of a node device or _hstate == node id. | |
2993 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
2994 | * the base kernel, on the hugetlb module. |
2995 | */ | |
2996 | struct node_hstate { | |
2997 | struct kobject *hugepages_kobj; | |
2998 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2999 | }; | |
b4e289a6 | 3000 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3001 | |
3002 | /* | |
10fbcf4c | 3003 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3004 | */ |
3005 | static struct attribute *per_node_hstate_attrs[] = { | |
3006 | &nr_hugepages_attr.attr, | |
3007 | &free_hugepages_attr.attr, | |
3008 | &surplus_hugepages_attr.attr, | |
3009 | NULL, | |
3010 | }; | |
3011 | ||
67e5ed96 | 3012 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3013 | .attrs = per_node_hstate_attrs, |
3014 | }; | |
3015 | ||
3016 | /* | |
10fbcf4c | 3017 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3018 | * Returns node id via non-NULL nidp. |
3019 | */ | |
3020 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3021 | { | |
3022 | int nid; | |
3023 | ||
3024 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3025 | struct node_hstate *nhs = &node_hstates[nid]; | |
3026 | int i; | |
3027 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3028 | if (nhs->hstate_kobjs[i] == kobj) { | |
3029 | if (nidp) | |
3030 | *nidp = nid; | |
3031 | return &hstates[i]; | |
3032 | } | |
3033 | } | |
3034 | ||
3035 | BUG(); | |
3036 | return NULL; | |
3037 | } | |
3038 | ||
3039 | /* | |
10fbcf4c | 3040 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3041 | * No-op if no hstate attributes attached. |
3042 | */ | |
3cd8b44f | 3043 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3044 | { |
3045 | struct hstate *h; | |
10fbcf4c | 3046 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3047 | |
3048 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3049 | return; /* no hstate attributes */ |
9a305230 | 3050 | |
972dc4de AK |
3051 | for_each_hstate(h) { |
3052 | int idx = hstate_index(h); | |
3053 | if (nhs->hstate_kobjs[idx]) { | |
3054 | kobject_put(nhs->hstate_kobjs[idx]); | |
3055 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3056 | } |
972dc4de | 3057 | } |
9a305230 LS |
3058 | |
3059 | kobject_put(nhs->hugepages_kobj); | |
3060 | nhs->hugepages_kobj = NULL; | |
3061 | } | |
3062 | ||
9a305230 LS |
3063 | |
3064 | /* | |
10fbcf4c | 3065 | * Register hstate attributes for a single node device. |
9a305230 LS |
3066 | * No-op if attributes already registered. |
3067 | */ | |
3cd8b44f | 3068 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3069 | { |
3070 | struct hstate *h; | |
10fbcf4c | 3071 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3072 | int err; |
3073 | ||
3074 | if (nhs->hugepages_kobj) | |
3075 | return; /* already allocated */ | |
3076 | ||
3077 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3078 | &node->dev.kobj); |
9a305230 LS |
3079 | if (!nhs->hugepages_kobj) |
3080 | return; | |
3081 | ||
3082 | for_each_hstate(h) { | |
3083 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3084 | nhs->hstate_kobjs, | |
3085 | &per_node_hstate_attr_group); | |
3086 | if (err) { | |
282f4214 | 3087 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3088 | h->name, node->dev.id); |
9a305230 LS |
3089 | hugetlb_unregister_node(node); |
3090 | break; | |
3091 | } | |
3092 | } | |
3093 | } | |
3094 | ||
3095 | /* | |
9b5e5d0f | 3096 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3097 | * devices of nodes that have memory. All on-line nodes should have |
3098 | * registered their associated device by this time. | |
9a305230 | 3099 | */ |
7d9ca000 | 3100 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3101 | { |
3102 | int nid; | |
3103 | ||
8cebfcd0 | 3104 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3105 | struct node *node = node_devices[nid]; |
10fbcf4c | 3106 | if (node->dev.id == nid) |
9a305230 LS |
3107 | hugetlb_register_node(node); |
3108 | } | |
3109 | ||
3110 | /* | |
10fbcf4c | 3111 | * Let the node device driver know we're here so it can |
9a305230 LS |
3112 | * [un]register hstate attributes on node hotplug. |
3113 | */ | |
3114 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3115 | hugetlb_unregister_node); | |
3116 | } | |
3117 | #else /* !CONFIG_NUMA */ | |
3118 | ||
3119 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3120 | { | |
3121 | BUG(); | |
3122 | if (nidp) | |
3123 | *nidp = -1; | |
3124 | return NULL; | |
3125 | } | |
3126 | ||
9a305230 LS |
3127 | static void hugetlb_register_all_nodes(void) { } |
3128 | ||
3129 | #endif | |
3130 | ||
a3437870 NA |
3131 | static int __init hugetlb_init(void) |
3132 | { | |
8382d914 DB |
3133 | int i; |
3134 | ||
d6995da3 MK |
3135 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
3136 | __NR_HPAGEFLAGS); | |
3137 | ||
c2833a5b MK |
3138 | if (!hugepages_supported()) { |
3139 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
3140 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 3141 | return 0; |
c2833a5b | 3142 | } |
a3437870 | 3143 | |
282f4214 MK |
3144 | /* |
3145 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
3146 | * architectures depend on setup being done here. | |
3147 | */ | |
3148 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
3149 | if (!parsed_default_hugepagesz) { | |
3150 | /* | |
3151 | * If we did not parse a default huge page size, set | |
3152 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
3153 | * number of huge pages for this default size was implicitly | |
3154 | * specified, set that here as well. | |
3155 | * Note that the implicit setting will overwrite an explicit | |
3156 | * setting. A warning will be printed in this case. | |
3157 | */ | |
3158 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
3159 | if (default_hstate_max_huge_pages) { | |
3160 | if (default_hstate.max_huge_pages) { | |
3161 | char buf[32]; | |
3162 | ||
3163 | string_get_size(huge_page_size(&default_hstate), | |
3164 | 1, STRING_UNITS_2, buf, 32); | |
3165 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
3166 | default_hstate.max_huge_pages, buf); | |
3167 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
3168 | default_hstate_max_huge_pages); | |
3169 | } | |
3170 | default_hstate.max_huge_pages = | |
3171 | default_hstate_max_huge_pages; | |
d715cf80 | 3172 | } |
f8b74815 | 3173 | } |
a3437870 | 3174 | |
cf11e85f | 3175 | hugetlb_cma_check(); |
a3437870 | 3176 | hugetlb_init_hstates(); |
aa888a74 | 3177 | gather_bootmem_prealloc(); |
a3437870 NA |
3178 | report_hugepages(); |
3179 | ||
3180 | hugetlb_sysfs_init(); | |
9a305230 | 3181 | hugetlb_register_all_nodes(); |
7179e7bf | 3182 | hugetlb_cgroup_file_init(); |
9a305230 | 3183 | |
8382d914 DB |
3184 | #ifdef CONFIG_SMP |
3185 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3186 | #else | |
3187 | num_fault_mutexes = 1; | |
3188 | #endif | |
c672c7f2 | 3189 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3190 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3191 | GFP_KERNEL); | |
c672c7f2 | 3192 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3193 | |
3194 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3195 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3196 | return 0; |
3197 | } | |
3e89e1c5 | 3198 | subsys_initcall(hugetlb_init); |
a3437870 | 3199 | |
ae94da89 MK |
3200 | /* Overwritten by architectures with more huge page sizes */ |
3201 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 3202 | { |
ae94da89 | 3203 | return size == HPAGE_SIZE; |
9fee021d VT |
3204 | } |
3205 | ||
d00181b9 | 3206 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3207 | { |
3208 | struct hstate *h; | |
8faa8b07 AK |
3209 | unsigned long i; |
3210 | ||
a3437870 | 3211 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
3212 | return; |
3213 | } | |
47d38344 | 3214 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3215 | BUG_ON(order == 0); |
47d38344 | 3216 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 | 3217 | h->order = order; |
aca78307 | 3218 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
3219 | for (i = 0; i < MAX_NUMNODES; ++i) |
3220 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3221 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3222 | h->next_nid_to_alloc = first_memory_node; |
3223 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3224 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3225 | huge_page_size(h)/1024); | |
8faa8b07 | 3226 | |
a3437870 NA |
3227 | parsed_hstate = h; |
3228 | } | |
3229 | ||
282f4214 MK |
3230 | /* |
3231 | * hugepages command line processing | |
3232 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
3233 | * specification. If not, ignore the hugepages value. hugepages can also | |
3234 | * be the first huge page command line option in which case it implicitly | |
3235 | * specifies the number of huge pages for the default size. | |
3236 | */ | |
3237 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
3238 | { |
3239 | unsigned long *mhp; | |
8faa8b07 | 3240 | static unsigned long *last_mhp; |
a3437870 | 3241 | |
9fee021d | 3242 | if (!parsed_valid_hugepagesz) { |
282f4214 | 3243 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 3244 | parsed_valid_hugepagesz = true; |
282f4214 | 3245 | return 0; |
9fee021d | 3246 | } |
282f4214 | 3247 | |
a3437870 | 3248 | /* |
282f4214 MK |
3249 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
3250 | * yet, so this hugepages= parameter goes to the "default hstate". | |
3251 | * Otherwise, it goes with the previously parsed hugepagesz or | |
3252 | * default_hugepagesz. | |
a3437870 | 3253 | */ |
9fee021d | 3254 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3255 | mhp = &default_hstate_max_huge_pages; |
3256 | else | |
3257 | mhp = &parsed_hstate->max_huge_pages; | |
3258 | ||
8faa8b07 | 3259 | if (mhp == last_mhp) { |
282f4214 MK |
3260 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
3261 | return 0; | |
8faa8b07 AK |
3262 | } |
3263 | ||
a3437870 NA |
3264 | if (sscanf(s, "%lu", mhp) <= 0) |
3265 | *mhp = 0; | |
3266 | ||
8faa8b07 AK |
3267 | /* |
3268 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 3269 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
3270 | * use the bootmem allocator. |
3271 | */ | |
04adbc3f | 3272 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
3273 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3274 | ||
3275 | last_mhp = mhp; | |
3276 | ||
a3437870 NA |
3277 | return 1; |
3278 | } | |
282f4214 | 3279 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 3280 | |
282f4214 MK |
3281 | /* |
3282 | * hugepagesz command line processing | |
3283 | * A specific huge page size can only be specified once with hugepagesz. | |
3284 | * hugepagesz is followed by hugepages on the command line. The global | |
3285 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
3286 | * hugepagesz argument was valid. | |
3287 | */ | |
359f2544 | 3288 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 3289 | { |
359f2544 | 3290 | unsigned long size; |
282f4214 MK |
3291 | struct hstate *h; |
3292 | ||
3293 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
3294 | size = (unsigned long)memparse(s, NULL); |
3295 | ||
3296 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3297 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
359f2544 MK |
3298 | return 0; |
3299 | } | |
3300 | ||
282f4214 MK |
3301 | h = size_to_hstate(size); |
3302 | if (h) { | |
3303 | /* | |
3304 | * hstate for this size already exists. This is normally | |
3305 | * an error, but is allowed if the existing hstate is the | |
3306 | * default hstate. More specifically, it is only allowed if | |
3307 | * the number of huge pages for the default hstate was not | |
3308 | * previously specified. | |
3309 | */ | |
3310 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
3311 | default_hstate.max_huge_pages) { | |
3312 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
3313 | return 0; | |
3314 | } | |
3315 | ||
3316 | /* | |
3317 | * No need to call hugetlb_add_hstate() as hstate already | |
3318 | * exists. But, do set parsed_hstate so that a following | |
3319 | * hugepages= parameter will be applied to this hstate. | |
3320 | */ | |
3321 | parsed_hstate = h; | |
3322 | parsed_valid_hugepagesz = true; | |
3323 | return 1; | |
38237830 MK |
3324 | } |
3325 | ||
359f2544 | 3326 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 3327 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
3328 | return 1; |
3329 | } | |
359f2544 MK |
3330 | __setup("hugepagesz=", hugepagesz_setup); |
3331 | ||
282f4214 MK |
3332 | /* |
3333 | * default_hugepagesz command line input | |
3334 | * Only one instance of default_hugepagesz allowed on command line. | |
3335 | */ | |
ae94da89 | 3336 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 3337 | { |
ae94da89 MK |
3338 | unsigned long size; |
3339 | ||
282f4214 | 3340 | parsed_valid_hugepagesz = false; |
282f4214 MK |
3341 | if (parsed_default_hugepagesz) { |
3342 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
3343 | return 0; | |
3344 | } | |
3345 | ||
ae94da89 MK |
3346 | size = (unsigned long)memparse(s, NULL); |
3347 | ||
3348 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3349 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
ae94da89 MK |
3350 | return 0; |
3351 | } | |
3352 | ||
282f4214 MK |
3353 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
3354 | parsed_valid_hugepagesz = true; | |
3355 | parsed_default_hugepagesz = true; | |
3356 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
3357 | ||
3358 | /* | |
3359 | * The number of default huge pages (for this size) could have been | |
3360 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
3361 | * then default_hstate_max_huge_pages is set. If the default huge | |
3362 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
3363 | * allocated here from bootmem allocator. | |
3364 | */ | |
3365 | if (default_hstate_max_huge_pages) { | |
3366 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3367 | if (hstate_is_gigantic(&default_hstate)) | |
3368 | hugetlb_hstate_alloc_pages(&default_hstate); | |
3369 | default_hstate_max_huge_pages = 0; | |
3370 | } | |
3371 | ||
e11bfbfc NP |
3372 | return 1; |
3373 | } | |
ae94da89 | 3374 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 3375 | |
8ca39e68 | 3376 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
3377 | { |
3378 | int node; | |
3379 | unsigned int nr = 0; | |
8ca39e68 MS |
3380 | nodemask_t *mpol_allowed; |
3381 | unsigned int *array = h->free_huge_pages_node; | |
3382 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
3383 | ||
3384 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 3385 | |
8ca39e68 | 3386 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
c93b0a99 | 3387 | if (!mpol_allowed || node_isset(node, *mpol_allowed)) |
8ca39e68 MS |
3388 | nr += array[node]; |
3389 | } | |
8a213460 NA |
3390 | |
3391 | return nr; | |
3392 | } | |
3393 | ||
3394 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
3395 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
3396 | void *buffer, size_t *length, | |
3397 | loff_t *ppos, unsigned long *out) | |
3398 | { | |
3399 | struct ctl_table dup_table; | |
3400 | ||
3401 | /* | |
3402 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
3403 | * can duplicate the @table and alter the duplicate of it. | |
3404 | */ | |
3405 | dup_table = *table; | |
3406 | dup_table.data = out; | |
3407 | ||
3408 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
3409 | } | |
3410 | ||
06808b08 LS |
3411 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3412 | struct ctl_table *table, int write, | |
32927393 | 3413 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 3414 | { |
e5ff2159 | 3415 | struct hstate *h = &default_hstate; |
238d3c13 | 3416 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3417 | int ret; |
e5ff2159 | 3418 | |
457c1b27 | 3419 | if (!hugepages_supported()) |
86613628 | 3420 | return -EOPNOTSUPP; |
457c1b27 | 3421 | |
17743798 MS |
3422 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3423 | &tmp); | |
08d4a246 MH |
3424 | if (ret) |
3425 | goto out; | |
e5ff2159 | 3426 | |
238d3c13 DR |
3427 | if (write) |
3428 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3429 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3430 | out: |
3431 | return ret; | |
1da177e4 | 3432 | } |
396faf03 | 3433 | |
06808b08 | 3434 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 3435 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3436 | { |
3437 | ||
3438 | return hugetlb_sysctl_handler_common(false, table, write, | |
3439 | buffer, length, ppos); | |
3440 | } | |
3441 | ||
3442 | #ifdef CONFIG_NUMA | |
3443 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 3444 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3445 | { |
3446 | return hugetlb_sysctl_handler_common(true, table, write, | |
3447 | buffer, length, ppos); | |
3448 | } | |
3449 | #endif /* CONFIG_NUMA */ | |
3450 | ||
a3d0c6aa | 3451 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 3452 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 3453 | { |
a5516438 | 3454 | struct hstate *h = &default_hstate; |
e5ff2159 | 3455 | unsigned long tmp; |
08d4a246 | 3456 | int ret; |
e5ff2159 | 3457 | |
457c1b27 | 3458 | if (!hugepages_supported()) |
86613628 | 3459 | return -EOPNOTSUPP; |
457c1b27 | 3460 | |
c033a93c | 3461 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3462 | |
bae7f4ae | 3463 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3464 | return -EINVAL; |
3465 | ||
17743798 MS |
3466 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3467 | &tmp); | |
08d4a246 MH |
3468 | if (ret) |
3469 | goto out; | |
e5ff2159 AK |
3470 | |
3471 | if (write) { | |
3472 | spin_lock(&hugetlb_lock); | |
3473 | h->nr_overcommit_huge_pages = tmp; | |
3474 | spin_unlock(&hugetlb_lock); | |
3475 | } | |
08d4a246 MH |
3476 | out: |
3477 | return ret; | |
a3d0c6aa NA |
3478 | } |
3479 | ||
1da177e4 LT |
3480 | #endif /* CONFIG_SYSCTL */ |
3481 | ||
e1759c21 | 3482 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3483 | { |
fcb2b0c5 RG |
3484 | struct hstate *h; |
3485 | unsigned long total = 0; | |
3486 | ||
457c1b27 NA |
3487 | if (!hugepages_supported()) |
3488 | return; | |
fcb2b0c5 RG |
3489 | |
3490 | for_each_hstate(h) { | |
3491 | unsigned long count = h->nr_huge_pages; | |
3492 | ||
aca78307 | 3493 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
3494 | |
3495 | if (h == &default_hstate) | |
3496 | seq_printf(m, | |
3497 | "HugePages_Total: %5lu\n" | |
3498 | "HugePages_Free: %5lu\n" | |
3499 | "HugePages_Rsvd: %5lu\n" | |
3500 | "HugePages_Surp: %5lu\n" | |
3501 | "Hugepagesize: %8lu kB\n", | |
3502 | count, | |
3503 | h->free_huge_pages, | |
3504 | h->resv_huge_pages, | |
3505 | h->surplus_huge_pages, | |
aca78307 | 3506 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
3507 | } |
3508 | ||
aca78307 | 3509 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
3510 | } |
3511 | ||
7981593b | 3512 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 3513 | { |
a5516438 | 3514 | struct hstate *h = &default_hstate; |
7981593b | 3515 | |
457c1b27 NA |
3516 | if (!hugepages_supported()) |
3517 | return 0; | |
7981593b JP |
3518 | |
3519 | return sysfs_emit_at(buf, len, | |
3520 | "Node %d HugePages_Total: %5u\n" | |
3521 | "Node %d HugePages_Free: %5u\n" | |
3522 | "Node %d HugePages_Surp: %5u\n", | |
3523 | nid, h->nr_huge_pages_node[nid], | |
3524 | nid, h->free_huge_pages_node[nid], | |
3525 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3526 | } |
3527 | ||
949f7ec5 DR |
3528 | void hugetlb_show_meminfo(void) |
3529 | { | |
3530 | struct hstate *h; | |
3531 | int nid; | |
3532 | ||
457c1b27 NA |
3533 | if (!hugepages_supported()) |
3534 | return; | |
3535 | ||
949f7ec5 DR |
3536 | for_each_node_state(nid, N_MEMORY) |
3537 | for_each_hstate(h) | |
3538 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3539 | nid, | |
3540 | h->nr_huge_pages_node[nid], | |
3541 | h->free_huge_pages_node[nid], | |
3542 | h->surplus_huge_pages_node[nid], | |
aca78307 | 3543 | huge_page_size(h) / SZ_1K); |
949f7ec5 DR |
3544 | } |
3545 | ||
5d317b2b NH |
3546 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3547 | { | |
3548 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3549 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3550 | } | |
3551 | ||
1da177e4 LT |
3552 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3553 | unsigned long hugetlb_total_pages(void) | |
3554 | { | |
d0028588 WL |
3555 | struct hstate *h; |
3556 | unsigned long nr_total_pages = 0; | |
3557 | ||
3558 | for_each_hstate(h) | |
3559 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3560 | return nr_total_pages; | |
1da177e4 | 3561 | } |
1da177e4 | 3562 | |
a5516438 | 3563 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3564 | { |
3565 | int ret = -ENOMEM; | |
3566 | ||
0aa7f354 ML |
3567 | if (!delta) |
3568 | return 0; | |
3569 | ||
fc1b8a73 MG |
3570 | spin_lock(&hugetlb_lock); |
3571 | /* | |
3572 | * When cpuset is configured, it breaks the strict hugetlb page | |
3573 | * reservation as the accounting is done on a global variable. Such | |
3574 | * reservation is completely rubbish in the presence of cpuset because | |
3575 | * the reservation is not checked against page availability for the | |
3576 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3577 | * with lack of free htlb page in cpuset that the task is in. | |
3578 | * Attempt to enforce strict accounting with cpuset is almost | |
3579 | * impossible (or too ugly) because cpuset is too fluid that | |
3580 | * task or memory node can be dynamically moved between cpusets. | |
3581 | * | |
3582 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3583 | * undesirable. However, in order to preserve some of the semantics, | |
3584 | * we fall back to check against current free page availability as | |
3585 | * a best attempt and hopefully to minimize the impact of changing | |
3586 | * semantics that cpuset has. | |
8ca39e68 MS |
3587 | * |
3588 | * Apart from cpuset, we also have memory policy mechanism that | |
3589 | * also determines from which node the kernel will allocate memory | |
3590 | * in a NUMA system. So similar to cpuset, we also should consider | |
3591 | * the memory policy of the current task. Similar to the description | |
3592 | * above. | |
fc1b8a73 MG |
3593 | */ |
3594 | if (delta > 0) { | |
a5516438 | 3595 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3596 | goto out; |
3597 | ||
8ca39e68 | 3598 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 3599 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
3600 | goto out; |
3601 | } | |
3602 | } | |
3603 | ||
3604 | ret = 0; | |
3605 | if (delta < 0) | |
a5516438 | 3606 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3607 | |
3608 | out: | |
3609 | spin_unlock(&hugetlb_lock); | |
3610 | return ret; | |
3611 | } | |
3612 | ||
84afd99b AW |
3613 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3614 | { | |
f522c3ac | 3615 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3616 | |
3617 | /* | |
3618 | * This new VMA should share its siblings reservation map if present. | |
3619 | * The VMA will only ever have a valid reservation map pointer where | |
3620 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3621 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3622 | * after this open call completes. It is therefore safe to take a |
3623 | * new reference here without additional locking. | |
3624 | */ | |
4e35f483 | 3625 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3626 | kref_get(&resv->refs); |
84afd99b AW |
3627 | } |
3628 | ||
a1e78772 MG |
3629 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3630 | { | |
a5516438 | 3631 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3632 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3633 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3634 | unsigned long reserve, start, end; |
1c5ecae3 | 3635 | long gbl_reserve; |
84afd99b | 3636 | |
4e35f483 JK |
3637 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3638 | return; | |
84afd99b | 3639 | |
4e35f483 JK |
3640 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3641 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3642 | |
4e35f483 | 3643 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 3644 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 3645 | if (reserve) { |
1c5ecae3 MK |
3646 | /* |
3647 | * Decrement reserve counts. The global reserve count may be | |
3648 | * adjusted if the subpool has a minimum size. | |
3649 | */ | |
3650 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3651 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3652 | } |
e9fe92ae MA |
3653 | |
3654 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
3655 | } |
3656 | ||
31383c68 DW |
3657 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3658 | { | |
3659 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3660 | return -EINVAL; | |
3661 | return 0; | |
3662 | } | |
3663 | ||
05ea8860 DW |
3664 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3665 | { | |
aca78307 | 3666 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
3667 | } |
3668 | ||
1da177e4 LT |
3669 | /* |
3670 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3671 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 3672 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
3673 | * this far. |
3674 | */ | |
b3ec9f33 | 3675 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3676 | { |
3677 | BUG(); | |
d0217ac0 | 3678 | return 0; |
1da177e4 LT |
3679 | } |
3680 | ||
eec3636a JC |
3681 | /* |
3682 | * When a new function is introduced to vm_operations_struct and added | |
3683 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3684 | * This is because under System V memory model, mappings created via | |
3685 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3686 | * their original vm_ops are overwritten with shm_vm_ops. | |
3687 | */ | |
f0f37e2f | 3688 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3689 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3690 | .open = hugetlb_vm_op_open, |
a1e78772 | 3691 | .close = hugetlb_vm_op_close, |
dd3b614f | 3692 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 3693 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3694 | }; |
3695 | ||
1e8f889b DG |
3696 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3697 | int writable) | |
63551ae0 DG |
3698 | { |
3699 | pte_t entry; | |
3700 | ||
1e8f889b | 3701 | if (writable) { |
106c992a GS |
3702 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3703 | vma->vm_page_prot))); | |
63551ae0 | 3704 | } else { |
106c992a GS |
3705 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3706 | vma->vm_page_prot)); | |
63551ae0 DG |
3707 | } |
3708 | entry = pte_mkyoung(entry); | |
3709 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3710 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3711 | |
3712 | return entry; | |
3713 | } | |
3714 | ||
1e8f889b DG |
3715 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3716 | unsigned long address, pte_t *ptep) | |
3717 | { | |
3718 | pte_t entry; | |
3719 | ||
106c992a | 3720 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3721 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3722 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3723 | } |
3724 | ||
d5ed7444 | 3725 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3726 | { |
3727 | swp_entry_t swp; | |
3728 | ||
3729 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3730 | return false; |
4a705fef | 3731 | swp = pte_to_swp_entry(pte); |
d79d176a | 3732 | if (is_migration_entry(swp)) |
d5ed7444 | 3733 | return true; |
4a705fef | 3734 | else |
d5ed7444 | 3735 | return false; |
4a705fef NH |
3736 | } |
3737 | ||
3e5c3600 | 3738 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
3739 | { |
3740 | swp_entry_t swp; | |
3741 | ||
3742 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 3743 | return false; |
4a705fef | 3744 | swp = pte_to_swp_entry(pte); |
d79d176a | 3745 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 3746 | return true; |
4a705fef | 3747 | else |
3e5c3600 | 3748 | return false; |
4a705fef | 3749 | } |
1e8f889b | 3750 | |
4eae4efa PX |
3751 | static void |
3752 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
3753 | struct page *new_page) | |
3754 | { | |
3755 | __SetPageUptodate(new_page); | |
3756 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); | |
3757 | hugepage_add_new_anon_rmap(new_page, vma, addr); | |
3758 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); | |
3759 | ClearHPageRestoreReserve(new_page); | |
3760 | SetHPageMigratable(new_page); | |
3761 | } | |
3762 | ||
63551ae0 DG |
3763 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3764 | struct vm_area_struct *vma) | |
3765 | { | |
5e41540c | 3766 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3767 | struct page *ptepage; |
1c59827d | 3768 | unsigned long addr; |
ca6eb14d | 3769 | bool cow = is_cow_mapping(vma->vm_flags); |
a5516438 AK |
3770 | struct hstate *h = hstate_vma(vma); |
3771 | unsigned long sz = huge_page_size(h); | |
4eae4efa | 3772 | unsigned long npages = pages_per_huge_page(h); |
c0d0381a | 3773 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 3774 | struct mmu_notifier_range range; |
e8569dd2 | 3775 | int ret = 0; |
1e8f889b | 3776 | |
ac46d4f3 | 3777 | if (cow) { |
7269f999 | 3778 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3779 | vma->vm_start, |
ac46d4f3 JG |
3780 | vma->vm_end); |
3781 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
3782 | } else { |
3783 | /* | |
3784 | * For shared mappings i_mmap_rwsem must be held to call | |
3785 | * huge_pte_alloc, otherwise the returned ptep could go | |
3786 | * away if part of a shared pmd and another thread calls | |
3787 | * huge_pmd_unshare. | |
3788 | */ | |
3789 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 3790 | } |
e8569dd2 | 3791 | |
a5516438 | 3792 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3793 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3794 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3795 | if (!src_pte) |
3796 | continue; | |
aec44e0f | 3797 | dst_pte = huge_pte_alloc(dst, vma, addr, sz); |
e8569dd2 AS |
3798 | if (!dst_pte) { |
3799 | ret = -ENOMEM; | |
3800 | break; | |
3801 | } | |
c5c99429 | 3802 | |
5e41540c MK |
3803 | /* |
3804 | * If the pagetables are shared don't copy or take references. | |
3805 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3806 | * | |
3807 | * However, src could have 'unshared' and dst shares with | |
3808 | * another vma. If dst_pte !none, this implies sharing. | |
3809 | * Check here before taking page table lock, and once again | |
3810 | * after taking the lock below. | |
3811 | */ | |
3812 | dst_entry = huge_ptep_get(dst_pte); | |
3813 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3814 | continue; |
3815 | ||
cb900f41 KS |
3816 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3817 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3818 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3819 | entry = huge_ptep_get(src_pte); |
5e41540c | 3820 | dst_entry = huge_ptep_get(dst_pte); |
4eae4efa | 3821 | again: |
5e41540c MK |
3822 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { |
3823 | /* | |
3824 | * Skip if src entry none. Also, skip in the | |
3825 | * unlikely case dst entry !none as this implies | |
3826 | * sharing with another vma. | |
3827 | */ | |
4a705fef NH |
3828 | ; |
3829 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3830 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3831 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3832 | ||
3833 | if (is_write_migration_entry(swp_entry) && cow) { | |
3834 | /* | |
3835 | * COW mappings require pages in both | |
3836 | * parent and child to be set to read. | |
3837 | */ | |
3838 | make_migration_entry_read(&swp_entry); | |
3839 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3840 | set_huge_swap_pte_at(src, addr, src_pte, |
3841 | entry, sz); | |
4a705fef | 3842 | } |
e5251fd4 | 3843 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3844 | } else { |
4eae4efa PX |
3845 | entry = huge_ptep_get(src_pte); |
3846 | ptepage = pte_page(entry); | |
3847 | get_page(ptepage); | |
3848 | ||
3849 | /* | |
3850 | * This is a rare case where we see pinned hugetlb | |
3851 | * pages while they're prone to COW. We need to do the | |
3852 | * COW earlier during fork. | |
3853 | * | |
3854 | * When pre-allocating the page or copying data, we | |
3855 | * need to be without the pgtable locks since we could | |
3856 | * sleep during the process. | |
3857 | */ | |
3858 | if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { | |
3859 | pte_t src_pte_old = entry; | |
3860 | struct page *new; | |
3861 | ||
3862 | spin_unlock(src_ptl); | |
3863 | spin_unlock(dst_ptl); | |
3864 | /* Do not use reserve as it's private owned */ | |
3865 | new = alloc_huge_page(vma, addr, 1); | |
3866 | if (IS_ERR(new)) { | |
3867 | put_page(ptepage); | |
3868 | ret = PTR_ERR(new); | |
3869 | break; | |
3870 | } | |
3871 | copy_user_huge_page(new, ptepage, addr, vma, | |
3872 | npages); | |
3873 | put_page(ptepage); | |
3874 | ||
3875 | /* Install the new huge page if src pte stable */ | |
3876 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
3877 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3878 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
3879 | entry = huge_ptep_get(src_pte); | |
3880 | if (!pte_same(src_pte_old, entry)) { | |
3881 | put_page(new); | |
3882 | /* dst_entry won't change as in child */ | |
3883 | goto again; | |
3884 | } | |
3885 | hugetlb_install_page(vma, dst_pte, addr, new); | |
3886 | spin_unlock(src_ptl); | |
3887 | spin_unlock(dst_ptl); | |
3888 | continue; | |
3889 | } | |
3890 | ||
34ee645e | 3891 | if (cow) { |
0f10851e JG |
3892 | /* |
3893 | * No need to notify as we are downgrading page | |
3894 | * table protection not changing it to point | |
3895 | * to a new page. | |
3896 | * | |
ad56b738 | 3897 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3898 | */ |
7f2e9525 | 3899 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3900 | } |
4eae4efa | 3901 | |
53f9263b | 3902 | page_dup_rmap(ptepage, true); |
1c59827d | 3903 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 3904 | hugetlb_count_add(npages, dst); |
1c59827d | 3905 | } |
cb900f41 KS |
3906 | spin_unlock(src_ptl); |
3907 | spin_unlock(dst_ptl); | |
63551ae0 | 3908 | } |
63551ae0 | 3909 | |
e8569dd2 | 3910 | if (cow) |
ac46d4f3 | 3911 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
3912 | else |
3913 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
3914 | |
3915 | return ret; | |
63551ae0 DG |
3916 | } |
3917 | ||
24669e58 AK |
3918 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3919 | unsigned long start, unsigned long end, | |
3920 | struct page *ref_page) | |
63551ae0 DG |
3921 | { |
3922 | struct mm_struct *mm = vma->vm_mm; | |
3923 | unsigned long address; | |
c7546f8f | 3924 | pte_t *ptep; |
63551ae0 | 3925 | pte_t pte; |
cb900f41 | 3926 | spinlock_t *ptl; |
63551ae0 | 3927 | struct page *page; |
a5516438 AK |
3928 | struct hstate *h = hstate_vma(vma); |
3929 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3930 | struct mmu_notifier_range range; |
a5516438 | 3931 | |
63551ae0 | 3932 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3933 | BUG_ON(start & ~huge_page_mask(h)); |
3934 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3935 | |
07e32661 AK |
3936 | /* |
3937 | * This is a hugetlb vma, all the pte entries should point | |
3938 | * to huge page. | |
3939 | */ | |
ed6a7935 | 3940 | tlb_change_page_size(tlb, sz); |
24669e58 | 3941 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3942 | |
3943 | /* | |
3944 | * If sharing possible, alert mmu notifiers of worst case. | |
3945 | */ | |
6f4f13e8 JG |
3946 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3947 | end); | |
ac46d4f3 JG |
3948 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3949 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3950 | address = start; |
569f48b8 | 3951 | for (; address < end; address += sz) { |
7868a208 | 3952 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3953 | if (!ptep) |
c7546f8f DG |
3954 | continue; |
3955 | ||
cb900f41 | 3956 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 3957 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
31d49da5 | 3958 | spin_unlock(ptl); |
dff11abe MK |
3959 | /* |
3960 | * We just unmapped a page of PMDs by clearing a PUD. | |
3961 | * The caller's TLB flush range should cover this area. | |
3962 | */ | |
31d49da5 AK |
3963 | continue; |
3964 | } | |
39dde65c | 3965 | |
6629326b | 3966 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3967 | if (huge_pte_none(pte)) { |
3968 | spin_unlock(ptl); | |
3969 | continue; | |
3970 | } | |
6629326b HD |
3971 | |
3972 | /* | |
9fbc1f63 NH |
3973 | * Migrating hugepage or HWPoisoned hugepage is already |
3974 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3975 | */ |
9fbc1f63 | 3976 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3977 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3978 | spin_unlock(ptl); |
3979 | continue; | |
8c4894c6 | 3980 | } |
6629326b HD |
3981 | |
3982 | page = pte_page(pte); | |
04f2cbe3 MG |
3983 | /* |
3984 | * If a reference page is supplied, it is because a specific | |
3985 | * page is being unmapped, not a range. Ensure the page we | |
3986 | * are about to unmap is the actual page of interest. | |
3987 | */ | |
3988 | if (ref_page) { | |
31d49da5 AK |
3989 | if (page != ref_page) { |
3990 | spin_unlock(ptl); | |
3991 | continue; | |
3992 | } | |
04f2cbe3 MG |
3993 | /* |
3994 | * Mark the VMA as having unmapped its page so that | |
3995 | * future faults in this VMA will fail rather than | |
3996 | * looking like data was lost | |
3997 | */ | |
3998 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3999 | } | |
4000 | ||
c7546f8f | 4001 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 4002 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 4003 | if (huge_pte_dirty(pte)) |
6649a386 | 4004 | set_page_dirty(page); |
9e81130b | 4005 | |
5d317b2b | 4006 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 4007 | page_remove_rmap(page, true); |
31d49da5 | 4008 | |
cb900f41 | 4009 | spin_unlock(ptl); |
e77b0852 | 4010 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
4011 | /* |
4012 | * Bail out after unmapping reference page if supplied | |
4013 | */ | |
4014 | if (ref_page) | |
4015 | break; | |
fe1668ae | 4016 | } |
ac46d4f3 | 4017 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 4018 | tlb_end_vma(tlb, vma); |
1da177e4 | 4019 | } |
63551ae0 | 4020 | |
d833352a MG |
4021 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
4022 | struct vm_area_struct *vma, unsigned long start, | |
4023 | unsigned long end, struct page *ref_page) | |
4024 | { | |
4025 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
4026 | ||
4027 | /* | |
4028 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
4029 | * test will fail on a vma being torn down, and not grab a page table | |
4030 | * on its way out. We're lucky that the flag has such an appropriate | |
4031 | * name, and can in fact be safely cleared here. We could clear it | |
4032 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 4033 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 4034 | * because in the context this is called, the VMA is about to be |
c8c06efa | 4035 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
4036 | */ |
4037 | vma->vm_flags &= ~VM_MAYSHARE; | |
4038 | } | |
4039 | ||
502717f4 | 4040 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 4041 | unsigned long end, struct page *ref_page) |
502717f4 | 4042 | { |
24669e58 | 4043 | struct mmu_gather tlb; |
dff11abe | 4044 | |
a72afd87 | 4045 | tlb_gather_mmu(&tlb, vma->vm_mm); |
24669e58 | 4046 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
ae8eba8b | 4047 | tlb_finish_mmu(&tlb); |
502717f4 KC |
4048 | } |
4049 | ||
04f2cbe3 MG |
4050 | /* |
4051 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 4052 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
4053 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
4054 | * same region. | |
4055 | */ | |
2f4612af DB |
4056 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
4057 | struct page *page, unsigned long address) | |
04f2cbe3 | 4058 | { |
7526674d | 4059 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
4060 | struct vm_area_struct *iter_vma; |
4061 | struct address_space *mapping; | |
04f2cbe3 MG |
4062 | pgoff_t pgoff; |
4063 | ||
4064 | /* | |
4065 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
4066 | * from page cache lookup which is in HPAGE_SIZE units. | |
4067 | */ | |
7526674d | 4068 | address = address & huge_page_mask(h); |
36e4f20a MH |
4069 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
4070 | vma->vm_pgoff; | |
93c76a3d | 4071 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 4072 | |
4eb2b1dc MG |
4073 | /* |
4074 | * Take the mapping lock for the duration of the table walk. As | |
4075 | * this mapping should be shared between all the VMAs, | |
4076 | * __unmap_hugepage_range() is called as the lock is already held | |
4077 | */ | |
83cde9e8 | 4078 | i_mmap_lock_write(mapping); |
6b2dbba8 | 4079 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
4080 | /* Do not unmap the current VMA */ |
4081 | if (iter_vma == vma) | |
4082 | continue; | |
4083 | ||
2f84a899 MG |
4084 | /* |
4085 | * Shared VMAs have their own reserves and do not affect | |
4086 | * MAP_PRIVATE accounting but it is possible that a shared | |
4087 | * VMA is using the same page so check and skip such VMAs. | |
4088 | */ | |
4089 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
4090 | continue; | |
4091 | ||
04f2cbe3 MG |
4092 | /* |
4093 | * Unmap the page from other VMAs without their own reserves. | |
4094 | * They get marked to be SIGKILLed if they fault in these | |
4095 | * areas. This is because a future no-page fault on this VMA | |
4096 | * could insert a zeroed page instead of the data existing | |
4097 | * from the time of fork. This would look like data corruption | |
4098 | */ | |
4099 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
4100 | unmap_hugepage_range(iter_vma, address, |
4101 | address + huge_page_size(h), page); | |
04f2cbe3 | 4102 | } |
83cde9e8 | 4103 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
4104 | } |
4105 | ||
0fe6e20b NH |
4106 | /* |
4107 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
4108 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
4109 | * cannot race with other handlers or page migration. | |
4110 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 4111 | */ |
2b740303 | 4112 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 4113 | unsigned long address, pte_t *ptep, |
3999f52e | 4114 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 4115 | { |
3999f52e | 4116 | pte_t pte; |
a5516438 | 4117 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 4118 | struct page *old_page, *new_page; |
2b740303 SJ |
4119 | int outside_reserve = 0; |
4120 | vm_fault_t ret = 0; | |
974e6d66 | 4121 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 4122 | struct mmu_notifier_range range; |
1e8f889b | 4123 | |
3999f52e | 4124 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
4125 | old_page = pte_page(pte); |
4126 | ||
04f2cbe3 | 4127 | retry_avoidcopy: |
1e8f889b DG |
4128 | /* If no-one else is actually using this page, avoid the copy |
4129 | * and just make the page writable */ | |
37a2140d | 4130 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 4131 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 4132 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 4133 | return 0; |
1e8f889b DG |
4134 | } |
4135 | ||
04f2cbe3 MG |
4136 | /* |
4137 | * If the process that created a MAP_PRIVATE mapping is about to | |
4138 | * perform a COW due to a shared page count, attempt to satisfy | |
4139 | * the allocation without using the existing reserves. The pagecache | |
4140 | * page is used to determine if the reserve at this address was | |
4141 | * consumed or not. If reserves were used, a partial faulted mapping | |
4142 | * at the time of fork() could consume its reserves on COW instead | |
4143 | * of the full address range. | |
4144 | */ | |
5944d011 | 4145 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
4146 | old_page != pagecache_page) |
4147 | outside_reserve = 1; | |
4148 | ||
09cbfeaf | 4149 | get_page(old_page); |
b76c8cfb | 4150 | |
ad4404a2 DB |
4151 | /* |
4152 | * Drop page table lock as buddy allocator may be called. It will | |
4153 | * be acquired again before returning to the caller, as expected. | |
4154 | */ | |
cb900f41 | 4155 | spin_unlock(ptl); |
5b7a1d40 | 4156 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 4157 | |
2fc39cec | 4158 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
4159 | /* |
4160 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4161 | * it is due to references held by a child and an insufficient | |
4162 | * huge page pool. To guarantee the original mappers | |
4163 | * reliability, unmap the page from child processes. The child | |
4164 | * may get SIGKILLed if it later faults. | |
4165 | */ | |
4166 | if (outside_reserve) { | |
e7dd91c4 MK |
4167 | struct address_space *mapping = vma->vm_file->f_mapping; |
4168 | pgoff_t idx; | |
4169 | u32 hash; | |
4170 | ||
09cbfeaf | 4171 | put_page(old_page); |
04f2cbe3 | 4172 | BUG_ON(huge_pte_none(pte)); |
e7dd91c4 MK |
4173 | /* |
4174 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
4175 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
4176 | * in write mode. Dropping i_mmap_rwsem in read mode | |
4177 | * here is OK as COW mappings do not interact with | |
4178 | * PMD sharing. | |
4179 | * | |
4180 | * Reacquire both after unmap operation. | |
4181 | */ | |
4182 | idx = vma_hugecache_offset(h, vma, haddr); | |
4183 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4184 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4185 | i_mmap_unlock_read(mapping); | |
4186 | ||
5b7a1d40 | 4187 | unmap_ref_private(mm, vma, old_page, haddr); |
e7dd91c4 MK |
4188 | |
4189 | i_mmap_lock_read(mapping); | |
4190 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
2f4612af | 4191 | spin_lock(ptl); |
5b7a1d40 | 4192 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
4193 | if (likely(ptep && |
4194 | pte_same(huge_ptep_get(ptep), pte))) | |
4195 | goto retry_avoidcopy; | |
4196 | /* | |
4197 | * race occurs while re-acquiring page table | |
4198 | * lock, and our job is done. | |
4199 | */ | |
4200 | return 0; | |
04f2cbe3 MG |
4201 | } |
4202 | ||
2b740303 | 4203 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 4204 | goto out_release_old; |
1e8f889b DG |
4205 | } |
4206 | ||
0fe6e20b NH |
4207 | /* |
4208 | * When the original hugepage is shared one, it does not have | |
4209 | * anon_vma prepared. | |
4210 | */ | |
44e2aa93 | 4211 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
4212 | ret = VM_FAULT_OOM; |
4213 | goto out_release_all; | |
44e2aa93 | 4214 | } |
0fe6e20b | 4215 | |
974e6d66 | 4216 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 4217 | pages_per_huge_page(h)); |
0ed361de | 4218 | __SetPageUptodate(new_page); |
1e8f889b | 4219 | |
7269f999 | 4220 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 4221 | haddr + huge_page_size(h)); |
ac46d4f3 | 4222 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 4223 | |
b76c8cfb | 4224 | /* |
cb900f41 | 4225 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
4226 | * before the page tables are altered |
4227 | */ | |
cb900f41 | 4228 | spin_lock(ptl); |
5b7a1d40 | 4229 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 4230 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 4231 | ClearHPageRestoreReserve(new_page); |
07443a85 | 4232 | |
1e8f889b | 4233 | /* Break COW */ |
5b7a1d40 | 4234 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 4235 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 4236 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 4237 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 4238 | page_remove_rmap(old_page, true); |
5b7a1d40 | 4239 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
8f251a3d | 4240 | SetHPageMigratable(new_page); |
1e8f889b DG |
4241 | /* Make the old page be freed below */ |
4242 | new_page = old_page; | |
4243 | } | |
cb900f41 | 4244 | spin_unlock(ptl); |
ac46d4f3 | 4245 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 4246 | out_release_all: |
5b7a1d40 | 4247 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 4248 | put_page(new_page); |
ad4404a2 | 4249 | out_release_old: |
09cbfeaf | 4250 | put_page(old_page); |
8312034f | 4251 | |
ad4404a2 DB |
4252 | spin_lock(ptl); /* Caller expects lock to be held */ |
4253 | return ret; | |
1e8f889b DG |
4254 | } |
4255 | ||
04f2cbe3 | 4256 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
4257 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
4258 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
4259 | { |
4260 | struct address_space *mapping; | |
e7c4b0bf | 4261 | pgoff_t idx; |
04f2cbe3 MG |
4262 | |
4263 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 4264 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
4265 | |
4266 | return find_lock_page(mapping, idx); | |
4267 | } | |
4268 | ||
3ae77f43 HD |
4269 | /* |
4270 | * Return whether there is a pagecache page to back given address within VMA. | |
4271 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4272 | */ | |
4273 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
4274 | struct vm_area_struct *vma, unsigned long address) |
4275 | { | |
4276 | struct address_space *mapping; | |
4277 | pgoff_t idx; | |
4278 | struct page *page; | |
4279 | ||
4280 | mapping = vma->vm_file->f_mapping; | |
4281 | idx = vma_hugecache_offset(h, vma, address); | |
4282 | ||
4283 | page = find_get_page(mapping, idx); | |
4284 | if (page) | |
4285 | put_page(page); | |
4286 | return page != NULL; | |
4287 | } | |
4288 | ||
ab76ad54 MK |
4289 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
4290 | pgoff_t idx) | |
4291 | { | |
4292 | struct inode *inode = mapping->host; | |
4293 | struct hstate *h = hstate_inode(inode); | |
4294 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4295 | ||
4296 | if (err) | |
4297 | return err; | |
d6995da3 | 4298 | ClearHPageRestoreReserve(page); |
ab76ad54 | 4299 | |
22146c3c MK |
4300 | /* |
4301 | * set page dirty so that it will not be removed from cache/file | |
4302 | * by non-hugetlbfs specific code paths. | |
4303 | */ | |
4304 | set_page_dirty(page); | |
4305 | ||
ab76ad54 MK |
4306 | spin_lock(&inode->i_lock); |
4307 | inode->i_blocks += blocks_per_huge_page(h); | |
4308 | spin_unlock(&inode->i_lock); | |
4309 | return 0; | |
4310 | } | |
4311 | ||
2b740303 SJ |
4312 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
4313 | struct vm_area_struct *vma, | |
4314 | struct address_space *mapping, pgoff_t idx, | |
4315 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 4316 | { |
a5516438 | 4317 | struct hstate *h = hstate_vma(vma); |
2b740303 | 4318 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 4319 | int anon_rmap = 0; |
4c887265 | 4320 | unsigned long size; |
4c887265 | 4321 | struct page *page; |
1e8f889b | 4322 | pte_t new_pte; |
cb900f41 | 4323 | spinlock_t *ptl; |
285b8dca | 4324 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 4325 | bool new_page = false; |
4c887265 | 4326 | |
04f2cbe3 MG |
4327 | /* |
4328 | * Currently, we are forced to kill the process in the event the | |
4329 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4330 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4331 | */ |
4332 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4333 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4334 | current->pid); |
04f2cbe3 MG |
4335 | return ret; |
4336 | } | |
4337 | ||
4c887265 | 4338 | /* |
87bf91d3 MK |
4339 | * We can not race with truncation due to holding i_mmap_rwsem. |
4340 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4341 | * once for faults beyond end of file. | |
4c887265 | 4342 | */ |
87bf91d3 MK |
4343 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4344 | if (idx >= size) | |
4345 | goto out; | |
4346 | ||
6bda666a CL |
4347 | retry: |
4348 | page = find_lock_page(mapping, idx); | |
4349 | if (!page) { | |
1a1aad8a MK |
4350 | /* |
4351 | * Check for page in userfault range | |
4352 | */ | |
4353 | if (userfaultfd_missing(vma)) { | |
4354 | u32 hash; | |
4355 | struct vm_fault vmf = { | |
4356 | .vma = vma, | |
285b8dca | 4357 | .address = haddr, |
1a1aad8a MK |
4358 | .flags = flags, |
4359 | /* | |
4360 | * Hard to debug if it ends up being | |
4361 | * used by a callee that assumes | |
4362 | * something about the other | |
4363 | * uninitialized fields... same as in | |
4364 | * memory.c | |
4365 | */ | |
4366 | }; | |
4367 | ||
4368 | /* | |
c0d0381a MK |
4369 | * hugetlb_fault_mutex and i_mmap_rwsem must be |
4370 | * dropped before handling userfault. Reacquire | |
4371 | * after handling fault to make calling code simpler. | |
1a1aad8a | 4372 | */ |
188b04a7 | 4373 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
1a1aad8a | 4374 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4375 | i_mmap_unlock_read(mapping); |
1a1aad8a | 4376 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); |
c0d0381a | 4377 | i_mmap_lock_read(mapping); |
1a1aad8a MK |
4378 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
4379 | goto out; | |
4380 | } | |
4381 | ||
285b8dca | 4382 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4383 | if (IS_ERR(page)) { |
4643d67e MK |
4384 | /* |
4385 | * Returning error will result in faulting task being | |
4386 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4387 | * tasks from racing to fault in the same page which | |
4388 | * could result in false unable to allocate errors. | |
4389 | * Page migration does not take the fault mutex, but | |
4390 | * does a clear then write of pte's under page table | |
4391 | * lock. Page fault code could race with migration, | |
4392 | * notice the clear pte and try to allocate a page | |
4393 | * here. Before returning error, get ptl and make | |
4394 | * sure there really is no pte entry. | |
4395 | */ | |
4396 | ptl = huge_pte_lock(h, mm, ptep); | |
d83e6c8a ML |
4397 | ret = 0; |
4398 | if (huge_pte_none(huge_ptep_get(ptep))) | |
4399 | ret = vmf_error(PTR_ERR(page)); | |
4643d67e | 4400 | spin_unlock(ptl); |
6bda666a CL |
4401 | goto out; |
4402 | } | |
47ad8475 | 4403 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4404 | __SetPageUptodate(page); |
cb6acd01 | 4405 | new_page = true; |
ac9b9c66 | 4406 | |
f83a275d | 4407 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4408 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4409 | if (err) { |
4410 | put_page(page); | |
6bda666a CL |
4411 | if (err == -EEXIST) |
4412 | goto retry; | |
4413 | goto out; | |
4414 | } | |
23be7468 | 4415 | } else { |
6bda666a | 4416 | lock_page(page); |
0fe6e20b NH |
4417 | if (unlikely(anon_vma_prepare(vma))) { |
4418 | ret = VM_FAULT_OOM; | |
4419 | goto backout_unlocked; | |
4420 | } | |
409eb8c2 | 4421 | anon_rmap = 1; |
23be7468 | 4422 | } |
0fe6e20b | 4423 | } else { |
998b4382 NH |
4424 | /* |
4425 | * If memory error occurs between mmap() and fault, some process | |
4426 | * don't have hwpoisoned swap entry for errored virtual address. | |
4427 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4428 | */ | |
4429 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 4430 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4431 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4432 | goto backout_unlocked; |
4433 | } | |
6bda666a | 4434 | } |
1e8f889b | 4435 | |
57303d80 AW |
4436 | /* |
4437 | * If we are going to COW a private mapping later, we examine the | |
4438 | * pending reservations for this page now. This will ensure that | |
4439 | * any allocations necessary to record that reservation occur outside | |
4440 | * the spinlock. | |
4441 | */ | |
5e911373 | 4442 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4443 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4444 | ret = VM_FAULT_OOM; |
4445 | goto backout_unlocked; | |
4446 | } | |
5e911373 | 4447 | /* Just decrements count, does not deallocate */ |
285b8dca | 4448 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4449 | } |
57303d80 | 4450 | |
8bea8052 | 4451 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 4452 | ret = 0; |
7f2e9525 | 4453 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4454 | goto backout; |
4455 | ||
07443a85 | 4456 | if (anon_rmap) { |
d6995da3 | 4457 | ClearHPageRestoreReserve(page); |
285b8dca | 4458 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4459 | } else |
53f9263b | 4460 | page_dup_rmap(page, true); |
1e8f889b DG |
4461 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4462 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4463 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4464 | |
5d317b2b | 4465 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4466 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4467 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4468 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4469 | } |
4470 | ||
cb900f41 | 4471 | spin_unlock(ptl); |
cb6acd01 MK |
4472 | |
4473 | /* | |
8f251a3d MK |
4474 | * Only set HPageMigratable in newly allocated pages. Existing pages |
4475 | * found in the pagecache may not have HPageMigratableset if they have | |
4476 | * been isolated for migration. | |
cb6acd01 MK |
4477 | */ |
4478 | if (new_page) | |
8f251a3d | 4479 | SetHPageMigratable(page); |
cb6acd01 | 4480 | |
4c887265 AL |
4481 | unlock_page(page); |
4482 | out: | |
ac9b9c66 | 4483 | return ret; |
4c887265 AL |
4484 | |
4485 | backout: | |
cb900f41 | 4486 | spin_unlock(ptl); |
2b26736c | 4487 | backout_unlocked: |
4c887265 | 4488 | unlock_page(page); |
285b8dca | 4489 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4490 | put_page(page); |
4491 | goto out; | |
ac9b9c66 HD |
4492 | } |
4493 | ||
8382d914 | 4494 | #ifdef CONFIG_SMP |
188b04a7 | 4495 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4496 | { |
4497 | unsigned long key[2]; | |
4498 | u32 hash; | |
4499 | ||
1b426bac MK |
4500 | key[0] = (unsigned long) mapping; |
4501 | key[1] = idx; | |
8382d914 | 4502 | |
55254636 | 4503 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4504 | |
4505 | return hash & (num_fault_mutexes - 1); | |
4506 | } | |
4507 | #else | |
4508 | /* | |
6c26d310 | 4509 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
4510 | * return 0 and avoid the hashing overhead. |
4511 | */ | |
188b04a7 | 4512 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4513 | { |
4514 | return 0; | |
4515 | } | |
4516 | #endif | |
4517 | ||
2b740303 | 4518 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4519 | unsigned long address, unsigned int flags) |
86e5216f | 4520 | { |
8382d914 | 4521 | pte_t *ptep, entry; |
cb900f41 | 4522 | spinlock_t *ptl; |
2b740303 | 4523 | vm_fault_t ret; |
8382d914 DB |
4524 | u32 hash; |
4525 | pgoff_t idx; | |
0fe6e20b | 4526 | struct page *page = NULL; |
57303d80 | 4527 | struct page *pagecache_page = NULL; |
a5516438 | 4528 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4529 | struct address_space *mapping; |
0f792cf9 | 4530 | int need_wait_lock = 0; |
285b8dca | 4531 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4532 | |
285b8dca | 4533 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 4534 | if (ptep) { |
c0d0381a MK |
4535 | /* |
4536 | * Since we hold no locks, ptep could be stale. That is | |
4537 | * OK as we are only making decisions based on content and | |
4538 | * not actually modifying content here. | |
4539 | */ | |
fd6a03ed | 4540 | entry = huge_ptep_get(ptep); |
290408d4 | 4541 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4542 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4543 | return 0; |
4544 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4545 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4546 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
4547 | } |
4548 | ||
c0d0381a MK |
4549 | /* |
4550 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
4551 | * until finished with ptep. This serves two purposes: |
4552 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
4553 | * and making the ptep no longer valid. | |
4554 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
4555 | * |
4556 | * ptep could have already be assigned via huge_pte_offset. That | |
4557 | * is OK, as huge_pte_alloc will return the same value unless | |
4558 | * something has changed. | |
4559 | */ | |
8382d914 | 4560 | mapping = vma->vm_file->f_mapping; |
c0d0381a | 4561 | i_mmap_lock_read(mapping); |
aec44e0f | 4562 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); |
c0d0381a MK |
4563 | if (!ptep) { |
4564 | i_mmap_unlock_read(mapping); | |
4565 | return VM_FAULT_OOM; | |
4566 | } | |
8382d914 | 4567 | |
3935baa9 DG |
4568 | /* |
4569 | * Serialize hugepage allocation and instantiation, so that we don't | |
4570 | * get spurious allocation failures if two CPUs race to instantiate | |
4571 | * the same page in the page cache. | |
4572 | */ | |
c0d0381a | 4573 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 4574 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4575 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4576 | |
7f2e9525 GS |
4577 | entry = huge_ptep_get(ptep); |
4578 | if (huge_pte_none(entry)) { | |
8382d914 | 4579 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4580 | goto out_mutex; |
3935baa9 | 4581 | } |
86e5216f | 4582 | |
83c54070 | 4583 | ret = 0; |
1e8f889b | 4584 | |
0f792cf9 NH |
4585 | /* |
4586 | * entry could be a migration/hwpoison entry at this point, so this | |
4587 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
4588 | * an active hugepage in pagecache. This goto expects the 2nd page |
4589 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
4590 | * properly handle it. | |
0f792cf9 NH |
4591 | */ |
4592 | if (!pte_present(entry)) | |
4593 | goto out_mutex; | |
4594 | ||
57303d80 AW |
4595 | /* |
4596 | * If we are going to COW the mapping later, we examine the pending | |
4597 | * reservations for this page now. This will ensure that any | |
4598 | * allocations necessary to record that reservation occur outside the | |
4599 | * spinlock. For private mappings, we also lookup the pagecache | |
4600 | * page now as it is used to determine if a reservation has been | |
4601 | * consumed. | |
4602 | */ | |
106c992a | 4603 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4604 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4605 | ret = VM_FAULT_OOM; |
b4d1d99f | 4606 | goto out_mutex; |
2b26736c | 4607 | } |
5e911373 | 4608 | /* Just decrements count, does not deallocate */ |
285b8dca | 4609 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4610 | |
f83a275d | 4611 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4612 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4613 | vma, haddr); |
57303d80 AW |
4614 | } |
4615 | ||
0f792cf9 NH |
4616 | ptl = huge_pte_lock(h, mm, ptep); |
4617 | ||
4618 | /* Check for a racing update before calling hugetlb_cow */ | |
4619 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4620 | goto out_ptl; | |
4621 | ||
56c9cfb1 NH |
4622 | /* |
4623 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4624 | * pagecache_page, so here we need take the former one | |
4625 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4626 | */ |
4627 | page = pte_page(entry); | |
4628 | if (page != pagecache_page) | |
0f792cf9 NH |
4629 | if (!trylock_page(page)) { |
4630 | need_wait_lock = 1; | |
4631 | goto out_ptl; | |
4632 | } | |
b4d1d99f | 4633 | |
0f792cf9 | 4634 | get_page(page); |
b4d1d99f | 4635 | |
788c7df4 | 4636 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4637 | if (!huge_pte_write(entry)) { |
974e6d66 | 4638 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4639 | pagecache_page, ptl); |
0f792cf9 | 4640 | goto out_put_page; |
b4d1d99f | 4641 | } |
106c992a | 4642 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4643 | } |
4644 | entry = pte_mkyoung(entry); | |
285b8dca | 4645 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4646 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4647 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4648 | out_put_page: |
4649 | if (page != pagecache_page) | |
4650 | unlock_page(page); | |
4651 | put_page(page); | |
cb900f41 KS |
4652 | out_ptl: |
4653 | spin_unlock(ptl); | |
57303d80 AW |
4654 | |
4655 | if (pagecache_page) { | |
4656 | unlock_page(pagecache_page); | |
4657 | put_page(pagecache_page); | |
4658 | } | |
b4d1d99f | 4659 | out_mutex: |
c672c7f2 | 4660 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4661 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
4662 | /* |
4663 | * Generally it's safe to hold refcount during waiting page lock. But | |
4664 | * here we just wait to defer the next page fault to avoid busy loop and | |
4665 | * the page is not used after unlocked before returning from the current | |
4666 | * page fault. So we are safe from accessing freed page, even if we wait | |
4667 | * here without taking refcount. | |
4668 | */ | |
4669 | if (need_wait_lock) | |
4670 | wait_on_page_locked(page); | |
1e8f889b | 4671 | return ret; |
86e5216f AL |
4672 | } |
4673 | ||
8fb5debc MK |
4674 | /* |
4675 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4676 | * modifications for huge pages. | |
4677 | */ | |
4678 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4679 | pte_t *dst_pte, | |
4680 | struct vm_area_struct *dst_vma, | |
4681 | unsigned long dst_addr, | |
4682 | unsigned long src_addr, | |
4683 | struct page **pagep) | |
4684 | { | |
1e392147 AA |
4685 | struct address_space *mapping; |
4686 | pgoff_t idx; | |
4687 | unsigned long size; | |
1c9e8def | 4688 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4689 | struct hstate *h = hstate_vma(dst_vma); |
4690 | pte_t _dst_pte; | |
4691 | spinlock_t *ptl; | |
4692 | int ret; | |
4693 | struct page *page; | |
4694 | ||
4695 | if (!*pagep) { | |
4696 | ret = -ENOMEM; | |
4697 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4698 | if (IS_ERR(page)) | |
4699 | goto out; | |
4700 | ||
4701 | ret = copy_huge_page_from_user(page, | |
4702 | (const void __user *) src_addr, | |
810a56b9 | 4703 | pages_per_huge_page(h), false); |
8fb5debc | 4704 | |
c1e8d7c6 | 4705 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 4706 | if (unlikely(ret)) { |
9e368259 | 4707 | ret = -ENOENT; |
8fb5debc MK |
4708 | *pagep = page; |
4709 | /* don't free the page */ | |
4710 | goto out; | |
4711 | } | |
4712 | } else { | |
4713 | page = *pagep; | |
4714 | *pagep = NULL; | |
4715 | } | |
4716 | ||
4717 | /* | |
4718 | * The memory barrier inside __SetPageUptodate makes sure that | |
4719 | * preceding stores to the page contents become visible before | |
4720 | * the set_pte_at() write. | |
4721 | */ | |
4722 | __SetPageUptodate(page); | |
8fb5debc | 4723 | |
1e392147 AA |
4724 | mapping = dst_vma->vm_file->f_mapping; |
4725 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4726 | ||
1c9e8def MK |
4727 | /* |
4728 | * If shared, add to page cache | |
4729 | */ | |
4730 | if (vm_shared) { | |
1e392147 AA |
4731 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4732 | ret = -EFAULT; | |
4733 | if (idx >= size) | |
4734 | goto out_release_nounlock; | |
1c9e8def | 4735 | |
1e392147 AA |
4736 | /* |
4737 | * Serialization between remove_inode_hugepages() and | |
4738 | * huge_add_to_page_cache() below happens through the | |
4739 | * hugetlb_fault_mutex_table that here must be hold by | |
4740 | * the caller. | |
4741 | */ | |
1c9e8def MK |
4742 | ret = huge_add_to_page_cache(page, mapping, idx); |
4743 | if (ret) | |
4744 | goto out_release_nounlock; | |
4745 | } | |
4746 | ||
8fb5debc MK |
4747 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4748 | spin_lock(ptl); | |
4749 | ||
1e392147 AA |
4750 | /* |
4751 | * Recheck the i_size after holding PT lock to make sure not | |
4752 | * to leave any page mapped (as page_mapped()) beyond the end | |
4753 | * of the i_size (remove_inode_hugepages() is strict about | |
4754 | * enforcing that). If we bail out here, we'll also leave a | |
4755 | * page in the radix tree in the vm_shared case beyond the end | |
4756 | * of the i_size, but remove_inode_hugepages() will take care | |
4757 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4758 | */ | |
4759 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4760 | ret = -EFAULT; | |
4761 | if (idx >= size) | |
4762 | goto out_release_unlock; | |
4763 | ||
8fb5debc MK |
4764 | ret = -EEXIST; |
4765 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4766 | goto out_release_unlock; | |
4767 | ||
1c9e8def MK |
4768 | if (vm_shared) { |
4769 | page_dup_rmap(page, true); | |
4770 | } else { | |
d6995da3 | 4771 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
4772 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
4773 | } | |
8fb5debc MK |
4774 | |
4775 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4776 | if (dst_vma->vm_flags & VM_WRITE) | |
4777 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4778 | _dst_pte = pte_mkyoung(_dst_pte); | |
4779 | ||
4780 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4781 | ||
4782 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4783 | dst_vma->vm_flags & VM_WRITE); | |
4784 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4785 | ||
4786 | /* No need to invalidate - it was non-present before */ | |
4787 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4788 | ||
4789 | spin_unlock(ptl); | |
8f251a3d | 4790 | SetHPageMigratable(page); |
1c9e8def MK |
4791 | if (vm_shared) |
4792 | unlock_page(page); | |
8fb5debc MK |
4793 | ret = 0; |
4794 | out: | |
4795 | return ret; | |
4796 | out_release_unlock: | |
4797 | spin_unlock(ptl); | |
1c9e8def MK |
4798 | if (vm_shared) |
4799 | unlock_page(page); | |
5af10dfd | 4800 | out_release_nounlock: |
8fb5debc MK |
4801 | put_page(page); |
4802 | goto out; | |
4803 | } | |
4804 | ||
82e5d378 JM |
4805 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
4806 | int refs, struct page **pages, | |
4807 | struct vm_area_struct **vmas) | |
4808 | { | |
4809 | int nr; | |
4810 | ||
4811 | for (nr = 0; nr < refs; nr++) { | |
4812 | if (likely(pages)) | |
4813 | pages[nr] = mem_map_offset(page, nr); | |
4814 | if (vmas) | |
4815 | vmas[nr] = vma; | |
4816 | } | |
4817 | } | |
4818 | ||
28a35716 ML |
4819 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4820 | struct page **pages, struct vm_area_struct **vmas, | |
4821 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 4822 | long i, unsigned int flags, int *locked) |
63551ae0 | 4823 | { |
d5d4b0aa KC |
4824 | unsigned long pfn_offset; |
4825 | unsigned long vaddr = *position; | |
28a35716 | 4826 | unsigned long remainder = *nr_pages; |
a5516438 | 4827 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 4828 | int err = -EFAULT, refs; |
63551ae0 | 4829 | |
63551ae0 | 4830 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4831 | pte_t *pte; |
cb900f41 | 4832 | spinlock_t *ptl = NULL; |
2a15efc9 | 4833 | int absent; |
4c887265 | 4834 | struct page *page; |
63551ae0 | 4835 | |
02057967 DR |
4836 | /* |
4837 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4838 | * potentially allocating memory. | |
4839 | */ | |
fa45f116 | 4840 | if (fatal_signal_pending(current)) { |
02057967 DR |
4841 | remainder = 0; |
4842 | break; | |
4843 | } | |
4844 | ||
4c887265 AL |
4845 | /* |
4846 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4847 | * each hugepage. We have to make sure we get the |
4c887265 | 4848 | * first, for the page indexing below to work. |
cb900f41 KS |
4849 | * |
4850 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4851 | */ |
7868a208 PA |
4852 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4853 | huge_page_size(h)); | |
cb900f41 KS |
4854 | if (pte) |
4855 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4856 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4857 | ||
4858 | /* | |
4859 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4860 | * an error where there's an empty slot with no huge pagecache |
4861 | * to back it. This way, we avoid allocating a hugepage, and | |
4862 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4863 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4864 | */ |
3ae77f43 HD |
4865 | if (absent && (flags & FOLL_DUMP) && |
4866 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4867 | if (pte) |
4868 | spin_unlock(ptl); | |
2a15efc9 HD |
4869 | remainder = 0; |
4870 | break; | |
4871 | } | |
63551ae0 | 4872 | |
9cc3a5bd NH |
4873 | /* |
4874 | * We need call hugetlb_fault for both hugepages under migration | |
4875 | * (in which case hugetlb_fault waits for the migration,) and | |
4876 | * hwpoisoned hugepages (in which case we need to prevent the | |
4877 | * caller from accessing to them.) In order to do this, we use | |
4878 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4879 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4880 | * both cases, and because we can't follow correct pages | |
4881 | * directly from any kind of swap entries. | |
4882 | */ | |
4883 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4884 | ((flags & FOLL_WRITE) && |
4885 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4886 | vm_fault_t ret; |
87ffc118 | 4887 | unsigned int fault_flags = 0; |
63551ae0 | 4888 | |
cb900f41 KS |
4889 | if (pte) |
4890 | spin_unlock(ptl); | |
87ffc118 AA |
4891 | if (flags & FOLL_WRITE) |
4892 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 4893 | if (locked) |
71335f37 PX |
4894 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
4895 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
4896 | if (flags & FOLL_NOWAIT) |
4897 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4898 | FAULT_FLAG_RETRY_NOWAIT; | |
4899 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
4900 | /* |
4901 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
4902 | * FAULT_FLAG_TRIED can co-exist | |
4903 | */ | |
87ffc118 AA |
4904 | fault_flags |= FAULT_FLAG_TRIED; |
4905 | } | |
4906 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4907 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4908 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4909 | remainder = 0; |
4910 | break; | |
4911 | } | |
4912 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 4913 | if (locked && |
1ac25013 | 4914 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 4915 | *locked = 0; |
87ffc118 AA |
4916 | *nr_pages = 0; |
4917 | /* | |
4918 | * VM_FAULT_RETRY must not return an | |
4919 | * error, it will return zero | |
4920 | * instead. | |
4921 | * | |
4922 | * No need to update "position" as the | |
4923 | * caller will not check it after | |
4924 | * *nr_pages is set to 0. | |
4925 | */ | |
4926 | return i; | |
4927 | } | |
4928 | continue; | |
4c887265 AL |
4929 | } |
4930 | ||
a5516438 | 4931 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4932 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 4933 | |
acbfb087 ZL |
4934 | /* |
4935 | * If subpage information not requested, update counters | |
4936 | * and skip the same_page loop below. | |
4937 | */ | |
4938 | if (!pages && !vmas && !pfn_offset && | |
4939 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4940 | (remainder >= pages_per_huge_page(h))) { | |
4941 | vaddr += huge_page_size(h); | |
4942 | remainder -= pages_per_huge_page(h); | |
4943 | i += pages_per_huge_page(h); | |
4944 | spin_unlock(ptl); | |
4945 | continue; | |
4946 | } | |
4947 | ||
82e5d378 JM |
4948 | refs = min3(pages_per_huge_page(h) - pfn_offset, |
4949 | (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); | |
0fa5bc40 | 4950 | |
82e5d378 JM |
4951 | if (pages || vmas) |
4952 | record_subpages_vmas(mem_map_offset(page, pfn_offset), | |
4953 | vma, refs, | |
4954 | likely(pages) ? pages + i : NULL, | |
4955 | vmas ? vmas + i : NULL); | |
63551ae0 | 4956 | |
82e5d378 | 4957 | if (pages) { |
0fa5bc40 JM |
4958 | /* |
4959 | * try_grab_compound_head() should always succeed here, | |
4960 | * because: a) we hold the ptl lock, and b) we've just | |
4961 | * checked that the huge page is present in the page | |
4962 | * tables. If the huge page is present, then the tail | |
4963 | * pages must also be present. The ptl prevents the | |
4964 | * head page and tail pages from being rearranged in | |
4965 | * any way. So this page must be available at this | |
4966 | * point, unless the page refcount overflowed: | |
4967 | */ | |
82e5d378 | 4968 | if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], |
0fa5bc40 JM |
4969 | refs, |
4970 | flags))) { | |
4971 | spin_unlock(ptl); | |
4972 | remainder = 0; | |
4973 | err = -ENOMEM; | |
4974 | break; | |
4975 | } | |
d5d4b0aa | 4976 | } |
82e5d378 JM |
4977 | |
4978 | vaddr += (refs << PAGE_SHIFT); | |
4979 | remainder -= refs; | |
4980 | i += refs; | |
4981 | ||
cb900f41 | 4982 | spin_unlock(ptl); |
63551ae0 | 4983 | } |
28a35716 | 4984 | *nr_pages = remainder; |
87ffc118 AA |
4985 | /* |
4986 | * setting position is actually required only if remainder is | |
4987 | * not zero but it's faster not to add a "if (remainder)" | |
4988 | * branch. | |
4989 | */ | |
63551ae0 DG |
4990 | *position = vaddr; |
4991 | ||
2be7cfed | 4992 | return i ? i : err; |
63551ae0 | 4993 | } |
8f860591 | 4994 | |
7da4d641 | 4995 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4996 | unsigned long address, unsigned long end, pgprot_t newprot) |
4997 | { | |
4998 | struct mm_struct *mm = vma->vm_mm; | |
4999 | unsigned long start = address; | |
5000 | pte_t *ptep; | |
5001 | pte_t pte; | |
a5516438 | 5002 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 5003 | unsigned long pages = 0; |
dff11abe | 5004 | bool shared_pmd = false; |
ac46d4f3 | 5005 | struct mmu_notifier_range range; |
dff11abe MK |
5006 | |
5007 | /* | |
5008 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 5009 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
5010 | * range if PMD sharing is possible. |
5011 | */ | |
7269f999 JG |
5012 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
5013 | 0, vma, mm, start, end); | |
ac46d4f3 | 5014 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
5015 | |
5016 | BUG_ON(address >= end); | |
ac46d4f3 | 5017 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 5018 | |
ac46d4f3 | 5019 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 5020 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 5021 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 5022 | spinlock_t *ptl; |
7868a208 | 5023 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
5024 | if (!ptep) |
5025 | continue; | |
cb900f41 | 5026 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 5027 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
7da4d641 | 5028 | pages++; |
cb900f41 | 5029 | spin_unlock(ptl); |
dff11abe | 5030 | shared_pmd = true; |
39dde65c | 5031 | continue; |
7da4d641 | 5032 | } |
a8bda28d NH |
5033 | pte = huge_ptep_get(ptep); |
5034 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
5035 | spin_unlock(ptl); | |
5036 | continue; | |
5037 | } | |
5038 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
5039 | swp_entry_t entry = pte_to_swp_entry(pte); | |
5040 | ||
5041 | if (is_write_migration_entry(entry)) { | |
5042 | pte_t newpte; | |
5043 | ||
5044 | make_migration_entry_read(&entry); | |
5045 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
5046 | set_huge_swap_pte_at(mm, address, ptep, |
5047 | newpte, huge_page_size(h)); | |
a8bda28d NH |
5048 | pages++; |
5049 | } | |
5050 | spin_unlock(ptl); | |
5051 | continue; | |
5052 | } | |
5053 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
5054 | pte_t old_pte; |
5055 | ||
5056 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
5057 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 5058 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 5059 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 5060 | pages++; |
8f860591 | 5061 | } |
cb900f41 | 5062 | spin_unlock(ptl); |
8f860591 | 5063 | } |
d833352a | 5064 | /* |
c8c06efa | 5065 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 5066 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 5067 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
5068 | * and that page table be reused and filled with junk. If we actually |
5069 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 5070 | */ |
dff11abe | 5071 | if (shared_pmd) |
ac46d4f3 | 5072 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
5073 | else |
5074 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
5075 | /* |
5076 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
5077 | * page table protection not changing it to point to a new page. | |
5078 | * | |
ad56b738 | 5079 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 5080 | */ |
83cde9e8 | 5081 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 5082 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
5083 | |
5084 | return pages << h->order; | |
8f860591 ZY |
5085 | } |
5086 | ||
33b8f84a MK |
5087 | /* Return true if reservation was successful, false otherwise. */ |
5088 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 5089 | long from, long to, |
5a6fe125 | 5090 | struct vm_area_struct *vma, |
ca16d140 | 5091 | vm_flags_t vm_flags) |
e4e574b7 | 5092 | { |
33b8f84a | 5093 | long chg, add = -1; |
a5516438 | 5094 | struct hstate *h = hstate_inode(inode); |
90481622 | 5095 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 5096 | struct resv_map *resv_map; |
075a61d0 | 5097 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 5098 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 5099 | |
63489f8e MK |
5100 | /* This should never happen */ |
5101 | if (from > to) { | |
5102 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 5103 | return false; |
63489f8e MK |
5104 | } |
5105 | ||
17c9d12e MG |
5106 | /* |
5107 | * Only apply hugepage reservation if asked. At fault time, an | |
5108 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 5109 | * without using reserves |
17c9d12e | 5110 | */ |
ca16d140 | 5111 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 5112 | return true; |
17c9d12e | 5113 | |
a1e78772 MG |
5114 | /* |
5115 | * Shared mappings base their reservation on the number of pages that | |
5116 | * are already allocated on behalf of the file. Private mappings need | |
5117 | * to reserve the full area even if read-only as mprotect() may be | |
5118 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
5119 | */ | |
9119a41e | 5120 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
5121 | /* |
5122 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
5123 | * called for inodes for which resv_maps were created (see | |
5124 | * hugetlbfs_get_inode). | |
5125 | */ | |
4e35f483 | 5126 | resv_map = inode_resv_map(inode); |
9119a41e | 5127 | |
0db9d74e | 5128 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
5129 | |
5130 | } else { | |
e9fe92ae | 5131 | /* Private mapping. */ |
9119a41e | 5132 | resv_map = resv_map_alloc(); |
17c9d12e | 5133 | if (!resv_map) |
33b8f84a | 5134 | return false; |
17c9d12e | 5135 | |
a1e78772 | 5136 | chg = to - from; |
84afd99b | 5137 | |
17c9d12e MG |
5138 | set_vma_resv_map(vma, resv_map); |
5139 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
5140 | } | |
5141 | ||
33b8f84a | 5142 | if (chg < 0) |
c50ac050 | 5143 | goto out_err; |
8a630112 | 5144 | |
33b8f84a MK |
5145 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
5146 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 5147 | goto out_err; |
075a61d0 MA |
5148 | |
5149 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
5150 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
5151 | * of the resv_map. | |
5152 | */ | |
5153 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
5154 | } | |
5155 | ||
1c5ecae3 MK |
5156 | /* |
5157 | * There must be enough pages in the subpool for the mapping. If | |
5158 | * the subpool has a minimum size, there may be some global | |
5159 | * reservations already in place (gbl_reserve). | |
5160 | */ | |
5161 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 5162 | if (gbl_reserve < 0) |
075a61d0 | 5163 | goto out_uncharge_cgroup; |
5a6fe125 MG |
5164 | |
5165 | /* | |
17c9d12e | 5166 | * Check enough hugepages are available for the reservation. |
90481622 | 5167 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 5168 | */ |
33b8f84a | 5169 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 5170 | goto out_put_pages; |
17c9d12e MG |
5171 | |
5172 | /* | |
5173 | * Account for the reservations made. Shared mappings record regions | |
5174 | * that have reservations as they are shared by multiple VMAs. | |
5175 | * When the last VMA disappears, the region map says how much | |
5176 | * the reservation was and the page cache tells how much of | |
5177 | * the reservation was consumed. Private mappings are per-VMA and | |
5178 | * only the consumed reservations are tracked. When the VMA | |
5179 | * disappears, the original reservation is the VMA size and the | |
5180 | * consumed reservations are stored in the map. Hence, nothing | |
5181 | * else has to be done for private mappings here | |
5182 | */ | |
33039678 | 5183 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 5184 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
5185 | |
5186 | if (unlikely(add < 0)) { | |
5187 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 5188 | goto out_put_pages; |
0db9d74e | 5189 | } else if (unlikely(chg > add)) { |
33039678 MK |
5190 | /* |
5191 | * pages in this range were added to the reserve | |
5192 | * map between region_chg and region_add. This | |
5193 | * indicates a race with alloc_huge_page. Adjust | |
5194 | * the subpool and reserve counts modified above | |
5195 | * based on the difference. | |
5196 | */ | |
5197 | long rsv_adjust; | |
5198 | ||
d85aecf2 ML |
5199 | /* |
5200 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
5201 | * reference to h_cg->css. See comment below for detail. | |
5202 | */ | |
075a61d0 MA |
5203 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
5204 | hstate_index(h), | |
5205 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5206 | ||
33039678 MK |
5207 | rsv_adjust = hugepage_subpool_put_pages(spool, |
5208 | chg - add); | |
5209 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
5210 | } else if (h_cg) { |
5211 | /* | |
5212 | * The file_regions will hold their own reference to | |
5213 | * h_cg->css. So we should release the reference held | |
5214 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
5215 | * done. | |
5216 | */ | |
5217 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
5218 | } |
5219 | } | |
33b8f84a MK |
5220 | return true; |
5221 | ||
075a61d0 MA |
5222 | out_put_pages: |
5223 | /* put back original number of pages, chg */ | |
5224 | (void)hugepage_subpool_put_pages(spool, chg); | |
5225 | out_uncharge_cgroup: | |
5226 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5227 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 5228 | out_err: |
5e911373 | 5229 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
5230 | /* Only call region_abort if the region_chg succeeded but the |
5231 | * region_add failed or didn't run. | |
5232 | */ | |
5233 | if (chg >= 0 && add < 0) | |
5234 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
5235 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
5236 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 5237 | return false; |
a43a8c39 KC |
5238 | } |
5239 | ||
b5cec28d MK |
5240 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
5241 | long freed) | |
a43a8c39 | 5242 | { |
a5516438 | 5243 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 5244 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 5245 | long chg = 0; |
90481622 | 5246 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 5247 | long gbl_reserve; |
45c682a6 | 5248 | |
f27a5136 MK |
5249 | /* |
5250 | * Since this routine can be called in the evict inode path for all | |
5251 | * hugetlbfs inodes, resv_map could be NULL. | |
5252 | */ | |
b5cec28d MK |
5253 | if (resv_map) { |
5254 | chg = region_del(resv_map, start, end); | |
5255 | /* | |
5256 | * region_del() can fail in the rare case where a region | |
5257 | * must be split and another region descriptor can not be | |
5258 | * allocated. If end == LONG_MAX, it will not fail. | |
5259 | */ | |
5260 | if (chg < 0) | |
5261 | return chg; | |
5262 | } | |
5263 | ||
45c682a6 | 5264 | spin_lock(&inode->i_lock); |
e4c6f8be | 5265 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
5266 | spin_unlock(&inode->i_lock); |
5267 | ||
1c5ecae3 MK |
5268 | /* |
5269 | * If the subpool has a minimum size, the number of global | |
5270 | * reservations to be released may be adjusted. | |
5271 | */ | |
5272 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5273 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
5274 | |
5275 | return 0; | |
a43a8c39 | 5276 | } |
93f70f90 | 5277 | |
3212b535 SC |
5278 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
5279 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5280 | struct vm_area_struct *vma, | |
5281 | unsigned long addr, pgoff_t idx) | |
5282 | { | |
5283 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5284 | svma->vm_start; | |
5285 | unsigned long sbase = saddr & PUD_MASK; | |
5286 | unsigned long s_end = sbase + PUD_SIZE; | |
5287 | ||
5288 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
5289 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
5290 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
5291 | |
5292 | /* | |
5293 | * match the virtual addresses, permission and the alignment of the | |
5294 | * page table page. | |
5295 | */ | |
5296 | if (pmd_index(addr) != pmd_index(saddr) || | |
5297 | vm_flags != svm_flags || | |
07e51edf | 5298 | !range_in_vma(svma, sbase, s_end)) |
3212b535 SC |
5299 | return 0; |
5300 | ||
5301 | return saddr; | |
5302 | } | |
5303 | ||
31aafb45 | 5304 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
5305 | { |
5306 | unsigned long base = addr & PUD_MASK; | |
5307 | unsigned long end = base + PUD_SIZE; | |
5308 | ||
5309 | /* | |
5310 | * check on proper vm_flags and page table alignment | |
5311 | */ | |
017b1660 | 5312 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
5313 | return true; |
5314 | return false; | |
3212b535 SC |
5315 | } |
5316 | ||
c1991e07 PX |
5317 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
5318 | { | |
5319 | #ifdef CONFIG_USERFAULTFD | |
5320 | if (uffd_disable_huge_pmd_share(vma)) | |
5321 | return false; | |
5322 | #endif | |
5323 | return vma_shareable(vma, addr); | |
5324 | } | |
5325 | ||
017b1660 MK |
5326 | /* |
5327 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5328 | * If yes, adjust start and end to cover range associated with possible | |
5329 | * shared pmd mappings. | |
5330 | */ | |
5331 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5332 | unsigned long *start, unsigned long *end) | |
5333 | { | |
a1ba9da8 LX |
5334 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
5335 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 5336 | |
a1ba9da8 LX |
5337 | /* |
5338 | * vma need span at least one aligned PUD size and the start,end range | |
5339 | * must at least partialy within it. | |
5340 | */ | |
5341 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
5342 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
5343 | return; |
5344 | ||
75802ca6 | 5345 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
5346 | if (*start > v_start) |
5347 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 5348 | |
a1ba9da8 LX |
5349 | if (*end < v_end) |
5350 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
5351 | } |
5352 | ||
3212b535 SC |
5353 | /* |
5354 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5355 | * and returns the corresponding pte. While this is not necessary for the | |
5356 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
5357 | * code much cleaner. |
5358 | * | |
0bf7b64e MK |
5359 | * This routine must be called with i_mmap_rwsem held in at least read mode if |
5360 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
5361 | * table entries associated with the address space. This is important as we | |
5362 | * are setting up sharing based on existing page table entries (mappings). | |
5363 | * | |
5364 | * NOTE: This routine is only called from huge_pte_alloc. Some callers of | |
5365 | * huge_pte_alloc know that sharing is not possible and do not take | |
5366 | * i_mmap_rwsem as a performance optimization. This is handled by the | |
5367 | * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is | |
5368 | * only required for subsequent processing. | |
3212b535 | 5369 | */ |
aec44e0f PX |
5370 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
5371 | unsigned long addr, pud_t *pud) | |
3212b535 | 5372 | { |
3212b535 SC |
5373 | struct address_space *mapping = vma->vm_file->f_mapping; |
5374 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5375 | vma->vm_pgoff; | |
5376 | struct vm_area_struct *svma; | |
5377 | unsigned long saddr; | |
5378 | pte_t *spte = NULL; | |
5379 | pte_t *pte; | |
cb900f41 | 5380 | spinlock_t *ptl; |
3212b535 | 5381 | |
0bf7b64e | 5382 | i_mmap_assert_locked(mapping); |
3212b535 SC |
5383 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
5384 | if (svma == vma) | |
5385 | continue; | |
5386 | ||
5387 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5388 | if (saddr) { | |
7868a208 PA |
5389 | spte = huge_pte_offset(svma->vm_mm, saddr, |
5390 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
5391 | if (spte) { |
5392 | get_page(virt_to_page(spte)); | |
5393 | break; | |
5394 | } | |
5395 | } | |
5396 | } | |
5397 | ||
5398 | if (!spte) | |
5399 | goto out; | |
5400 | ||
8bea8052 | 5401 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 5402 | if (pud_none(*pud)) { |
3212b535 SC |
5403 | pud_populate(mm, pud, |
5404 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5405 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5406 | } else { |
3212b535 | 5407 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5408 | } |
cb900f41 | 5409 | spin_unlock(ptl); |
3212b535 SC |
5410 | out: |
5411 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
5412 | return pte; |
5413 | } | |
5414 | ||
5415 | /* | |
5416 | * unmap huge page backed by shared pte. | |
5417 | * | |
5418 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5419 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5420 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5421 | * | |
c0d0381a | 5422 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
5423 | * |
5424 | * returns: 1 successfully unmapped a shared pte page | |
5425 | * 0 the underlying pte page is not shared, or it is the last user | |
5426 | */ | |
34ae204f MK |
5427 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5428 | unsigned long *addr, pte_t *ptep) | |
3212b535 SC |
5429 | { |
5430 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
5431 | p4d_t *p4d = p4d_offset(pgd, *addr); |
5432 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 | 5433 | |
34ae204f | 5434 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
5435 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
5436 | if (page_count(virt_to_page(ptep)) == 1) | |
5437 | return 0; | |
5438 | ||
5439 | pud_clear(pud); | |
5440 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 5441 | mm_dec_nr_pmds(mm); |
3212b535 SC |
5442 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
5443 | return 1; | |
5444 | } | |
c1991e07 | 5445 | |
9e5fc74c | 5446 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
aec44e0f PX |
5447 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
5448 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
5449 | { |
5450 | return NULL; | |
5451 | } | |
e81f2d22 | 5452 | |
34ae204f MK |
5453 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5454 | unsigned long *addr, pte_t *ptep) | |
e81f2d22 ZZ |
5455 | { |
5456 | return 0; | |
5457 | } | |
017b1660 MK |
5458 | |
5459 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5460 | unsigned long *start, unsigned long *end) | |
5461 | { | |
5462 | } | |
c1991e07 PX |
5463 | |
5464 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
5465 | { | |
5466 | return false; | |
5467 | } | |
3212b535 SC |
5468 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5469 | ||
9e5fc74c | 5470 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 5471 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
5472 | unsigned long addr, unsigned long sz) |
5473 | { | |
5474 | pgd_t *pgd; | |
c2febafc | 5475 | p4d_t *p4d; |
9e5fc74c SC |
5476 | pud_t *pud; |
5477 | pte_t *pte = NULL; | |
5478 | ||
5479 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5480 | p4d = p4d_alloc(mm, pgd, addr); |
5481 | if (!p4d) | |
5482 | return NULL; | |
c2febafc | 5483 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5484 | if (pud) { |
5485 | if (sz == PUD_SIZE) { | |
5486 | pte = (pte_t *)pud; | |
5487 | } else { | |
5488 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 5489 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 5490 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
5491 | else |
5492 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5493 | } | |
5494 | } | |
4e666314 | 5495 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5496 | |
5497 | return pte; | |
5498 | } | |
5499 | ||
9b19df29 PA |
5500 | /* |
5501 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5502 | * entry at address @addr | |
5503 | * | |
8ac0b81a LX |
5504 | * Return: Pointer to page table entry (PUD or PMD) for |
5505 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
5506 | * size @sz doesn't match the hugepage size at this level of the page |
5507 | * table. | |
5508 | */ | |
7868a208 PA |
5509 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5510 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5511 | { |
5512 | pgd_t *pgd; | |
c2febafc | 5513 | p4d_t *p4d; |
8ac0b81a LX |
5514 | pud_t *pud; |
5515 | pmd_t *pmd; | |
9e5fc74c SC |
5516 | |
5517 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5518 | if (!pgd_present(*pgd)) |
5519 | return NULL; | |
5520 | p4d = p4d_offset(pgd, addr); | |
5521 | if (!p4d_present(*p4d)) | |
5522 | return NULL; | |
9b19df29 | 5523 | |
c2febafc | 5524 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
5525 | if (sz == PUD_SIZE) |
5526 | /* must be pud huge, non-present or none */ | |
c2febafc | 5527 | return (pte_t *)pud; |
8ac0b81a | 5528 | if (!pud_present(*pud)) |
9b19df29 | 5529 | return NULL; |
8ac0b81a | 5530 | /* must have a valid entry and size to go further */ |
9b19df29 | 5531 | |
8ac0b81a LX |
5532 | pmd = pmd_offset(pud, addr); |
5533 | /* must be pmd huge, non-present or none */ | |
5534 | return (pte_t *)pmd; | |
9e5fc74c SC |
5535 | } |
5536 | ||
61f77eda NH |
5537 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5538 | ||
5539 | /* | |
5540 | * These functions are overwritable if your architecture needs its own | |
5541 | * behavior. | |
5542 | */ | |
5543 | struct page * __weak | |
5544 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5545 | int write) | |
5546 | { | |
5547 | return ERR_PTR(-EINVAL); | |
5548 | } | |
5549 | ||
4dc71451 AK |
5550 | struct page * __weak |
5551 | follow_huge_pd(struct vm_area_struct *vma, | |
5552 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5553 | { | |
5554 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5555 | return NULL; | |
5556 | } | |
5557 | ||
61f77eda | 5558 | struct page * __weak |
9e5fc74c | 5559 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5560 | pmd_t *pmd, int flags) |
9e5fc74c | 5561 | { |
e66f17ff NH |
5562 | struct page *page = NULL; |
5563 | spinlock_t *ptl; | |
c9d398fa | 5564 | pte_t pte; |
3faa52c0 JH |
5565 | |
5566 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5567 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5568 | (FOLL_PIN | FOLL_GET))) | |
5569 | return NULL; | |
5570 | ||
e66f17ff NH |
5571 | retry: |
5572 | ptl = pmd_lockptr(mm, pmd); | |
5573 | spin_lock(ptl); | |
5574 | /* | |
5575 | * make sure that the address range covered by this pmd is not | |
5576 | * unmapped from other threads. | |
5577 | */ | |
5578 | if (!pmd_huge(*pmd)) | |
5579 | goto out; | |
c9d398fa NH |
5580 | pte = huge_ptep_get((pte_t *)pmd); |
5581 | if (pte_present(pte)) { | |
97534127 | 5582 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
5583 | /* |
5584 | * try_grab_page() should always succeed here, because: a) we | |
5585 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5586 | * huge pmd (head) page is present in the page tables. The ptl | |
5587 | * prevents the head page and tail pages from being rearranged | |
5588 | * in any way. So this page must be available at this point, | |
5589 | * unless the page refcount overflowed: | |
5590 | */ | |
5591 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5592 | page = NULL; | |
5593 | goto out; | |
5594 | } | |
e66f17ff | 5595 | } else { |
c9d398fa | 5596 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5597 | spin_unlock(ptl); |
5598 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5599 | goto retry; | |
5600 | } | |
5601 | /* | |
5602 | * hwpoisoned entry is treated as no_page_table in | |
5603 | * follow_page_mask(). | |
5604 | */ | |
5605 | } | |
5606 | out: | |
5607 | spin_unlock(ptl); | |
9e5fc74c SC |
5608 | return page; |
5609 | } | |
5610 | ||
61f77eda | 5611 | struct page * __weak |
9e5fc74c | 5612 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5613 | pud_t *pud, int flags) |
9e5fc74c | 5614 | { |
3faa52c0 | 5615 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5616 | return NULL; |
9e5fc74c | 5617 | |
e66f17ff | 5618 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5619 | } |
5620 | ||
faaa5b62 AK |
5621 | struct page * __weak |
5622 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5623 | { | |
3faa52c0 | 5624 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5625 | return NULL; |
5626 | ||
5627 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5628 | } | |
5629 | ||
31caf665 NH |
5630 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5631 | { | |
bcc54222 NH |
5632 | bool ret = true; |
5633 | ||
31caf665 | 5634 | spin_lock(&hugetlb_lock); |
8f251a3d MK |
5635 | if (!PageHeadHuge(page) || |
5636 | !HPageMigratable(page) || | |
0eb2df2b | 5637 | !get_page_unless_zero(page)) { |
bcc54222 NH |
5638 | ret = false; |
5639 | goto unlock; | |
5640 | } | |
8f251a3d | 5641 | ClearHPageMigratable(page); |
31caf665 | 5642 | list_move_tail(&page->lru, list); |
bcc54222 | 5643 | unlock: |
31caf665 | 5644 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5645 | return ret; |
31caf665 NH |
5646 | } |
5647 | ||
5648 | void putback_active_hugepage(struct page *page) | |
5649 | { | |
31caf665 | 5650 | spin_lock(&hugetlb_lock); |
8f251a3d | 5651 | SetHPageMigratable(page); |
31caf665 NH |
5652 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5653 | spin_unlock(&hugetlb_lock); | |
5654 | put_page(page); | |
5655 | } | |
ab5ac90a MH |
5656 | |
5657 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5658 | { | |
5659 | struct hstate *h = page_hstate(oldpage); | |
5660 | ||
5661 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5662 | set_page_owner_migrate_reason(newpage, reason); | |
5663 | ||
5664 | /* | |
5665 | * transfer temporary state of the new huge page. This is | |
5666 | * reverse to other transitions because the newpage is going to | |
5667 | * be final while the old one will be freed so it takes over | |
5668 | * the temporary status. | |
5669 | * | |
5670 | * Also note that we have to transfer the per-node surplus state | |
5671 | * here as well otherwise the global surplus count will not match | |
5672 | * the per-node's. | |
5673 | */ | |
9157c311 | 5674 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
5675 | int old_nid = page_to_nid(oldpage); |
5676 | int new_nid = page_to_nid(newpage); | |
5677 | ||
9157c311 MK |
5678 | SetHPageTemporary(oldpage); |
5679 | ClearHPageTemporary(newpage); | |
ab5ac90a | 5680 | |
5af1ab1d ML |
5681 | /* |
5682 | * There is no need to transfer the per-node surplus state | |
5683 | * when we do not cross the node. | |
5684 | */ | |
5685 | if (new_nid == old_nid) | |
5686 | return; | |
ab5ac90a MH |
5687 | spin_lock(&hugetlb_lock); |
5688 | if (h->surplus_huge_pages_node[old_nid]) { | |
5689 | h->surplus_huge_pages_node[old_nid]--; | |
5690 | h->surplus_huge_pages_node[new_nid]++; | |
5691 | } | |
5692 | spin_unlock(&hugetlb_lock); | |
5693 | } | |
5694 | } | |
cf11e85f | 5695 | |
6dfeaff9 PX |
5696 | /* |
5697 | * This function will unconditionally remove all the shared pmd pgtable entries | |
5698 | * within the specific vma for a hugetlbfs memory range. | |
5699 | */ | |
5700 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
5701 | { | |
5702 | struct hstate *h = hstate_vma(vma); | |
5703 | unsigned long sz = huge_page_size(h); | |
5704 | struct mm_struct *mm = vma->vm_mm; | |
5705 | struct mmu_notifier_range range; | |
5706 | unsigned long address, start, end; | |
5707 | spinlock_t *ptl; | |
5708 | pte_t *ptep; | |
5709 | ||
5710 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
5711 | return; | |
5712 | ||
5713 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
5714 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
5715 | ||
5716 | if (start >= end) | |
5717 | return; | |
5718 | ||
5719 | /* | |
5720 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
5721 | * we have already done the PUD_SIZE alignment. | |
5722 | */ | |
5723 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
5724 | start, end); | |
5725 | mmu_notifier_invalidate_range_start(&range); | |
5726 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
5727 | for (address = start; address < end; address += PUD_SIZE) { | |
5728 | unsigned long tmp = address; | |
5729 | ||
5730 | ptep = huge_pte_offset(mm, address, sz); | |
5731 | if (!ptep) | |
5732 | continue; | |
5733 | ptl = huge_pte_lock(h, mm, ptep); | |
5734 | /* We don't want 'address' to be changed */ | |
5735 | huge_pmd_unshare(mm, vma, &tmp, ptep); | |
5736 | spin_unlock(ptl); | |
5737 | } | |
5738 | flush_hugetlb_tlb_range(vma, start, end); | |
5739 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
5740 | /* | |
5741 | * No need to call mmu_notifier_invalidate_range(), see | |
5742 | * Documentation/vm/mmu_notifier.rst. | |
5743 | */ | |
5744 | mmu_notifier_invalidate_range_end(&range); | |
5745 | } | |
5746 | ||
cf11e85f | 5747 | #ifdef CONFIG_CMA |
cf11e85f RG |
5748 | static bool cma_reserve_called __initdata; |
5749 | ||
5750 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
5751 | { | |
5752 | hugetlb_cma_size = memparse(p, &p); | |
5753 | return 0; | |
5754 | } | |
5755 | ||
5756 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
5757 | ||
5758 | void __init hugetlb_cma_reserve(int order) | |
5759 | { | |
5760 | unsigned long size, reserved, per_node; | |
5761 | int nid; | |
5762 | ||
5763 | cma_reserve_called = true; | |
5764 | ||
5765 | if (!hugetlb_cma_size) | |
5766 | return; | |
5767 | ||
5768 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
5769 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
5770 | (PAGE_SIZE << order) / SZ_1M); | |
5771 | return; | |
5772 | } | |
5773 | ||
5774 | /* | |
5775 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
5776 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
5777 | */ | |
5778 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
5779 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
5780 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
5781 | ||
5782 | reserved = 0; | |
5783 | for_each_node_state(nid, N_ONLINE) { | |
5784 | int res; | |
2281f797 | 5785 | char name[CMA_MAX_NAME]; |
cf11e85f RG |
5786 | |
5787 | size = min(per_node, hugetlb_cma_size - reserved); | |
5788 | size = round_up(size, PAGE_SIZE << order); | |
5789 | ||
2281f797 | 5790 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
cf11e85f | 5791 | res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, |
29d0f41d | 5792 | 0, false, name, |
cf11e85f RG |
5793 | &hugetlb_cma[nid], nid); |
5794 | if (res) { | |
5795 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
5796 | res, nid); | |
5797 | continue; | |
5798 | } | |
5799 | ||
5800 | reserved += size; | |
5801 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
5802 | size / SZ_1M, nid); | |
5803 | ||
5804 | if (reserved >= hugetlb_cma_size) | |
5805 | break; | |
5806 | } | |
5807 | } | |
5808 | ||
5809 | void __init hugetlb_cma_check(void) | |
5810 | { | |
5811 | if (!hugetlb_cma_size || cma_reserve_called) | |
5812 | return; | |
5813 | ||
5814 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
5815 | } | |
5816 | ||
5817 | #endif /* CONFIG_CMA */ |