]>
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 | |
7ffddd49 MS |
82 | static inline bool PageHugeFreed(struct page *head) |
83 | { | |
84 | return page_private(head + 4) == -1UL; | |
85 | } | |
86 | ||
87 | static inline void SetPageHugeFreed(struct page *head) | |
88 | { | |
89 | set_page_private(head + 4, -1UL); | |
90 | } | |
91 | ||
92 | static inline void ClearPageHugeFreed(struct page *head) | |
93 | { | |
94 | set_page_private(head + 4, 0); | |
95 | } | |
96 | ||
7ca02d0a MK |
97 | /* Forward declaration */ |
98 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
99 | ||
90481622 DG |
100 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
101 | { | |
102 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
103 | ||
104 | spin_unlock(&spool->lock); | |
105 | ||
106 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 107 | * remain, give up any reservations based on minimum size and |
7ca02d0a MK |
108 | * free the subpool */ |
109 | if (free) { | |
110 | if (spool->min_hpages != -1) | |
111 | hugetlb_acct_memory(spool->hstate, | |
112 | -spool->min_hpages); | |
90481622 | 113 | kfree(spool); |
7ca02d0a | 114 | } |
90481622 DG |
115 | } |
116 | ||
7ca02d0a MK |
117 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
118 | long min_hpages) | |
90481622 DG |
119 | { |
120 | struct hugepage_subpool *spool; | |
121 | ||
c6a91820 | 122 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
123 | if (!spool) |
124 | return NULL; | |
125 | ||
126 | spin_lock_init(&spool->lock); | |
127 | spool->count = 1; | |
7ca02d0a MK |
128 | spool->max_hpages = max_hpages; |
129 | spool->hstate = h; | |
130 | spool->min_hpages = min_hpages; | |
131 | ||
132 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
133 | kfree(spool); | |
134 | return NULL; | |
135 | } | |
136 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
137 | |
138 | return spool; | |
139 | } | |
140 | ||
141 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
142 | { | |
143 | spin_lock(&spool->lock); | |
144 | BUG_ON(!spool->count); | |
145 | spool->count--; | |
146 | unlock_or_release_subpool(spool); | |
147 | } | |
148 | ||
1c5ecae3 MK |
149 | /* |
150 | * Subpool accounting for allocating and reserving pages. | |
151 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 152 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
153 | * global pools must be adjusted (upward). The returned value may |
154 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 155 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
156 | */ |
157 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
158 | long delta) |
159 | { | |
1c5ecae3 | 160 | long ret = delta; |
90481622 DG |
161 | |
162 | if (!spool) | |
1c5ecae3 | 163 | return ret; |
90481622 DG |
164 | |
165 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
166 | |
167 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
168 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
169 | spool->used_hpages += delta; | |
170 | else { | |
171 | ret = -ENOMEM; | |
172 | goto unlock_ret; | |
173 | } | |
90481622 | 174 | } |
90481622 | 175 | |
09a95e29 MK |
176 | /* minimum size accounting */ |
177 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
178 | if (delta > spool->rsv_hpages) { |
179 | /* | |
180 | * Asking for more reserves than those already taken on | |
181 | * behalf of subpool. Return difference. | |
182 | */ | |
183 | ret = delta - spool->rsv_hpages; | |
184 | spool->rsv_hpages = 0; | |
185 | } else { | |
186 | ret = 0; /* reserves already accounted for */ | |
187 | spool->rsv_hpages -= delta; | |
188 | } | |
189 | } | |
190 | ||
191 | unlock_ret: | |
192 | spin_unlock(&spool->lock); | |
90481622 DG |
193 | return ret; |
194 | } | |
195 | ||
1c5ecae3 MK |
196 | /* |
197 | * Subpool accounting for freeing and unreserving pages. | |
198 | * Return the number of global page reservations that must be dropped. | |
199 | * The return value may only be different than the passed value (delta) | |
200 | * in the case where a subpool minimum size must be maintained. | |
201 | */ | |
202 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
203 | long delta) |
204 | { | |
1c5ecae3 MK |
205 | long ret = delta; |
206 | ||
90481622 | 207 | if (!spool) |
1c5ecae3 | 208 | return delta; |
90481622 DG |
209 | |
210 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
211 | |
212 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
213 | spool->used_hpages -= delta; | |
214 | ||
09a95e29 MK |
215 | /* minimum size accounting */ |
216 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
217 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
218 | ret = 0; | |
219 | else | |
220 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
221 | ||
222 | spool->rsv_hpages += delta; | |
223 | if (spool->rsv_hpages > spool->min_hpages) | |
224 | spool->rsv_hpages = spool->min_hpages; | |
225 | } | |
226 | ||
227 | /* | |
228 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
229 | * quota reference, free it now. | |
230 | */ | |
90481622 | 231 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
232 | |
233 | return ret; | |
90481622 DG |
234 | } |
235 | ||
236 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
237 | { | |
238 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
239 | } | |
240 | ||
241 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
242 | { | |
496ad9aa | 243 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
244 | } |
245 | ||
0db9d74e MA |
246 | /* Helper that removes a struct file_region from the resv_map cache and returns |
247 | * it for use. | |
248 | */ | |
249 | static struct file_region * | |
250 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
251 | { | |
252 | struct file_region *nrg = NULL; | |
253 | ||
254 | VM_BUG_ON(resv->region_cache_count <= 0); | |
255 | ||
256 | resv->region_cache_count--; | |
257 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
258 | list_del(&nrg->link); |
259 | ||
260 | nrg->from = from; | |
261 | nrg->to = to; | |
262 | ||
263 | return nrg; | |
264 | } | |
265 | ||
075a61d0 MA |
266 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
267 | struct file_region *rg) | |
268 | { | |
269 | #ifdef CONFIG_CGROUP_HUGETLB | |
270 | nrg->reservation_counter = rg->reservation_counter; | |
271 | nrg->css = rg->css; | |
272 | if (rg->css) | |
273 | css_get(rg->css); | |
274 | #endif | |
275 | } | |
276 | ||
277 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
278 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
279 | struct hstate *h, | |
280 | struct resv_map *resv, | |
281 | struct file_region *nrg) | |
282 | { | |
283 | #ifdef CONFIG_CGROUP_HUGETLB | |
284 | if (h_cg) { | |
285 | nrg->reservation_counter = | |
286 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
287 | nrg->css = &h_cg->css; | |
0dd70e6b ML |
288 | /* |
289 | * The caller will hold exactly one h_cg->css reference for the | |
290 | * whole contiguous reservation region. But this area might be | |
291 | * scattered when there are already some file_regions reside in | |
292 | * it. As a result, many file_regions may share only one css | |
293 | * reference. In order to ensure that one file_region must hold | |
294 | * exactly one h_cg->css reference, we should do css_get for | |
295 | * each file_region and leave the reference held by caller | |
296 | * untouched. | |
297 | */ | |
298 | css_get(&h_cg->css); | |
075a61d0 MA |
299 | if (!resv->pages_per_hpage) |
300 | resv->pages_per_hpage = pages_per_huge_page(h); | |
301 | /* pages_per_hpage should be the same for all entries in | |
302 | * a resv_map. | |
303 | */ | |
304 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
305 | } else { | |
306 | nrg->reservation_counter = NULL; | |
307 | nrg->css = NULL; | |
308 | } | |
309 | #endif | |
310 | } | |
311 | ||
0dd70e6b ML |
312 | static void put_uncharge_info(struct file_region *rg) |
313 | { | |
314 | #ifdef CONFIG_CGROUP_HUGETLB | |
315 | if (rg->css) | |
316 | css_put(rg->css); | |
317 | #endif | |
318 | } | |
319 | ||
a9b3f867 MA |
320 | static bool has_same_uncharge_info(struct file_region *rg, |
321 | struct file_region *org) | |
322 | { | |
323 | #ifdef CONFIG_CGROUP_HUGETLB | |
324 | return rg && org && | |
325 | rg->reservation_counter == org->reservation_counter && | |
326 | rg->css == org->css; | |
327 | ||
328 | #else | |
329 | return true; | |
330 | #endif | |
331 | } | |
332 | ||
333 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
334 | { | |
335 | struct file_region *nrg = NULL, *prg = NULL; | |
336 | ||
337 | prg = list_prev_entry(rg, link); | |
338 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
339 | has_same_uncharge_info(prg, rg)) { | |
340 | prg->to = rg->to; | |
341 | ||
342 | list_del(&rg->link); | |
0dd70e6b | 343 | put_uncharge_info(rg); |
a9b3f867 MA |
344 | kfree(rg); |
345 | ||
7db5e7b6 | 346 | rg = prg; |
a9b3f867 MA |
347 | } |
348 | ||
349 | nrg = list_next_entry(rg, link); | |
350 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
351 | has_same_uncharge_info(nrg, rg)) { | |
352 | nrg->from = rg->from; | |
353 | ||
354 | list_del(&rg->link); | |
0dd70e6b | 355 | put_uncharge_info(rg); |
a9b3f867 | 356 | kfree(rg); |
a9b3f867 MA |
357 | } |
358 | } | |
359 | ||
972a3da3 WY |
360 | /* |
361 | * Must be called with resv->lock held. | |
362 | * | |
363 | * Calling this with regions_needed != NULL will count the number of pages | |
364 | * to be added but will not modify the linked list. And regions_needed will | |
365 | * indicate the number of file_regions needed in the cache to carry out to add | |
366 | * the regions for this range. | |
d75c6af9 MA |
367 | */ |
368 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 369 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 370 | struct hstate *h, long *regions_needed) |
d75c6af9 | 371 | { |
0db9d74e | 372 | long add = 0; |
d75c6af9 | 373 | struct list_head *head = &resv->regions; |
0db9d74e | 374 | long last_accounted_offset = f; |
d75c6af9 MA |
375 | struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; |
376 | ||
0db9d74e MA |
377 | if (regions_needed) |
378 | *regions_needed = 0; | |
d75c6af9 | 379 | |
0db9d74e MA |
380 | /* In this loop, we essentially handle an entry for the range |
381 | * [last_accounted_offset, rg->from), at every iteration, with some | |
382 | * bounds checking. | |
383 | */ | |
384 | list_for_each_entry_safe(rg, trg, head, link) { | |
385 | /* Skip irrelevant regions that start before our range. */ | |
386 | if (rg->from < f) { | |
387 | /* If this region ends after the last accounted offset, | |
388 | * then we need to update last_accounted_offset. | |
389 | */ | |
390 | if (rg->to > last_accounted_offset) | |
391 | last_accounted_offset = rg->to; | |
392 | continue; | |
393 | } | |
d75c6af9 | 394 | |
0db9d74e MA |
395 | /* When we find a region that starts beyond our range, we've |
396 | * finished. | |
397 | */ | |
d75c6af9 MA |
398 | if (rg->from > t) |
399 | break; | |
400 | ||
0db9d74e MA |
401 | /* Add an entry for last_accounted_offset -> rg->from, and |
402 | * update last_accounted_offset. | |
403 | */ | |
404 | if (rg->from > last_accounted_offset) { | |
405 | add += rg->from - last_accounted_offset; | |
972a3da3 | 406 | if (!regions_needed) { |
0db9d74e MA |
407 | nrg = get_file_region_entry_from_cache( |
408 | resv, last_accounted_offset, rg->from); | |
075a61d0 MA |
409 | record_hugetlb_cgroup_uncharge_info(h_cg, h, |
410 | resv, nrg); | |
0db9d74e | 411 | list_add(&nrg->link, rg->link.prev); |
a9b3f867 | 412 | coalesce_file_region(resv, nrg); |
972a3da3 | 413 | } else |
0db9d74e MA |
414 | *regions_needed += 1; |
415 | } | |
416 | ||
417 | last_accounted_offset = rg->to; | |
418 | } | |
419 | ||
420 | /* Handle the case where our range extends beyond | |
421 | * last_accounted_offset. | |
422 | */ | |
423 | if (last_accounted_offset < t) { | |
424 | add += t - last_accounted_offset; | |
972a3da3 | 425 | if (!regions_needed) { |
0db9d74e MA |
426 | nrg = get_file_region_entry_from_cache( |
427 | resv, last_accounted_offset, t); | |
075a61d0 | 428 | record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); |
0db9d74e | 429 | list_add(&nrg->link, rg->link.prev); |
a9b3f867 | 430 | coalesce_file_region(resv, nrg); |
972a3da3 | 431 | } else |
0db9d74e MA |
432 | *regions_needed += 1; |
433 | } | |
434 | ||
435 | VM_BUG_ON(add < 0); | |
436 | return add; | |
437 | } | |
438 | ||
439 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
440 | */ | |
441 | static int allocate_file_region_entries(struct resv_map *resv, | |
442 | int regions_needed) | |
443 | __must_hold(&resv->lock) | |
444 | { | |
445 | struct list_head allocated_regions; | |
446 | int to_allocate = 0, i = 0; | |
447 | struct file_region *trg = NULL, *rg = NULL; | |
448 | ||
449 | VM_BUG_ON(regions_needed < 0); | |
450 | ||
451 | INIT_LIST_HEAD(&allocated_regions); | |
452 | ||
453 | /* | |
454 | * Check for sufficient descriptors in the cache to accommodate | |
455 | * the number of in progress add operations plus regions_needed. | |
456 | * | |
457 | * This is a while loop because when we drop the lock, some other call | |
458 | * to region_add or region_del may have consumed some region_entries, | |
459 | * so we keep looping here until we finally have enough entries for | |
460 | * (adds_in_progress + regions_needed). | |
461 | */ | |
462 | while (resv->region_cache_count < | |
463 | (resv->adds_in_progress + regions_needed)) { | |
464 | to_allocate = resv->adds_in_progress + regions_needed - | |
465 | resv->region_cache_count; | |
466 | ||
467 | /* At this point, we should have enough entries in the cache | |
468 | * for all the existings adds_in_progress. We should only be | |
469 | * needing to allocate for regions_needed. | |
d75c6af9 | 470 | */ |
0db9d74e MA |
471 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
472 | ||
473 | spin_unlock(&resv->lock); | |
474 | for (i = 0; i < to_allocate; i++) { | |
475 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
476 | if (!trg) | |
477 | goto out_of_memory; | |
478 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 479 | } |
d75c6af9 | 480 | |
0db9d74e MA |
481 | spin_lock(&resv->lock); |
482 | ||
d3ec7b6e WY |
483 | list_splice(&allocated_regions, &resv->region_cache); |
484 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
485 | } |
486 | ||
0db9d74e | 487 | return 0; |
d75c6af9 | 488 | |
0db9d74e MA |
489 | out_of_memory: |
490 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
491 | list_del(&rg->link); | |
492 | kfree(rg); | |
493 | } | |
494 | return -ENOMEM; | |
d75c6af9 MA |
495 | } |
496 | ||
1dd308a7 MK |
497 | /* |
498 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
499 | * map. Regions will be taken from the cache to fill in this range. |
500 | * Sufficient regions should exist in the cache due to the previous | |
501 | * call to region_chg with the same range, but in some cases the cache will not | |
502 | * have sufficient entries due to races with other code doing region_add or | |
503 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 504 | * |
0db9d74e MA |
505 | * regions_needed is the out value provided by a previous call to region_chg. |
506 | * | |
507 | * Return the number of new huge pages added to the map. This number is greater | |
508 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 509 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
510 | * region_add of regions of length 1 never allocate file_regions and cannot |
511 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
512 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 513 | */ |
0db9d74e | 514 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
515 | long in_regions_needed, struct hstate *h, |
516 | struct hugetlb_cgroup *h_cg) | |
96822904 | 517 | { |
0db9d74e | 518 | long add = 0, actual_regions_needed = 0; |
96822904 | 519 | |
7b24d861 | 520 | spin_lock(&resv->lock); |
0db9d74e MA |
521 | retry: |
522 | ||
523 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
524 | add_reservation_in_range(resv, f, t, NULL, NULL, |
525 | &actual_regions_needed); | |
96822904 | 526 | |
5e911373 | 527 | /* |
0db9d74e MA |
528 | * Check for sufficient descriptors in the cache to accommodate |
529 | * this add operation. Note that actual_regions_needed may be greater | |
530 | * than in_regions_needed, as the resv_map may have been modified since | |
531 | * the region_chg call. In this case, we need to make sure that we | |
532 | * allocate extra entries, such that we have enough for all the | |
533 | * existing adds_in_progress, plus the excess needed for this | |
534 | * operation. | |
5e911373 | 535 | */ |
0db9d74e MA |
536 | if (actual_regions_needed > in_regions_needed && |
537 | resv->region_cache_count < | |
538 | resv->adds_in_progress + | |
539 | (actual_regions_needed - in_regions_needed)) { | |
540 | /* region_add operation of range 1 should never need to | |
541 | * allocate file_region entries. | |
542 | */ | |
543 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 544 | |
0db9d74e MA |
545 | if (allocate_file_region_entries( |
546 | resv, actual_regions_needed - in_regions_needed)) { | |
547 | return -ENOMEM; | |
548 | } | |
5e911373 | 549 | |
0db9d74e | 550 | goto retry; |
5e911373 MK |
551 | } |
552 | ||
972a3da3 | 553 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
554 | |
555 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 556 | |
7b24d861 | 557 | spin_unlock(&resv->lock); |
cf3ad20b MK |
558 | VM_BUG_ON(add < 0); |
559 | return add; | |
96822904 AW |
560 | } |
561 | ||
1dd308a7 MK |
562 | /* |
563 | * Examine the existing reserve map and determine how many | |
564 | * huge pages in the specified range [f, t) are NOT currently | |
565 | * represented. This routine is called before a subsequent | |
566 | * call to region_add that will actually modify the reserve | |
567 | * map to add the specified range [f, t). region_chg does | |
568 | * not change the number of huge pages represented by the | |
0db9d74e MA |
569 | * map. A number of new file_region structures is added to the cache as a |
570 | * placeholder, for the subsequent region_add call to use. At least 1 | |
571 | * file_region structure is added. | |
572 | * | |
573 | * out_regions_needed is the number of regions added to the | |
574 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
575 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
576 | * |
577 | * Returns the number of huge pages that need to be added to the existing | |
578 | * reservation map for the range [f, t). This number is greater or equal to | |
579 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
580 | * is needed and can not be allocated. | |
1dd308a7 | 581 | */ |
0db9d74e MA |
582 | static long region_chg(struct resv_map *resv, long f, long t, |
583 | long *out_regions_needed) | |
96822904 | 584 | { |
96822904 AW |
585 | long chg = 0; |
586 | ||
7b24d861 | 587 | spin_lock(&resv->lock); |
5e911373 | 588 | |
972a3da3 | 589 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 590 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 591 | out_regions_needed); |
5e911373 | 592 | |
0db9d74e MA |
593 | if (*out_regions_needed == 0) |
594 | *out_regions_needed = 1; | |
5e911373 | 595 | |
0db9d74e MA |
596 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
597 | return -ENOMEM; | |
5e911373 | 598 | |
0db9d74e | 599 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 600 | |
7b24d861 | 601 | spin_unlock(&resv->lock); |
96822904 AW |
602 | return chg; |
603 | } | |
604 | ||
5e911373 MK |
605 | /* |
606 | * Abort the in progress add operation. The adds_in_progress field | |
607 | * of the resv_map keeps track of the operations in progress between | |
608 | * calls to region_chg and region_add. Operations are sometimes | |
609 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
610 | * is called to decrement the adds_in_progress counter. regions_needed |
611 | * is the value returned by the region_chg call, it is used to decrement | |
612 | * the adds_in_progress counter. | |
5e911373 MK |
613 | * |
614 | * NOTE: The range arguments [f, t) are not needed or used in this | |
615 | * routine. They are kept to make reading the calling code easier as | |
616 | * arguments will match the associated region_chg call. | |
617 | */ | |
0db9d74e MA |
618 | static void region_abort(struct resv_map *resv, long f, long t, |
619 | long regions_needed) | |
5e911373 MK |
620 | { |
621 | spin_lock(&resv->lock); | |
622 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 623 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
624 | spin_unlock(&resv->lock); |
625 | } | |
626 | ||
1dd308a7 | 627 | /* |
feba16e2 MK |
628 | * Delete the specified range [f, t) from the reserve map. If the |
629 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
630 | * should be deleted. Locate the regions which intersect [f, t) | |
631 | * and either trim, delete or split the existing regions. | |
632 | * | |
633 | * Returns the number of huge pages deleted from the reserve map. | |
634 | * In the normal case, the return value is zero or more. In the | |
635 | * case where a region must be split, a new region descriptor must | |
636 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
637 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
638 | * a region and possibly return -ENOMEM. Callers specifying | |
639 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 640 | */ |
feba16e2 | 641 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 642 | { |
1406ec9b | 643 | struct list_head *head = &resv->regions; |
96822904 | 644 | struct file_region *rg, *trg; |
feba16e2 MK |
645 | struct file_region *nrg = NULL; |
646 | long del = 0; | |
96822904 | 647 | |
feba16e2 | 648 | retry: |
7b24d861 | 649 | spin_lock(&resv->lock); |
feba16e2 | 650 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
651 | /* |
652 | * Skip regions before the range to be deleted. file_region | |
653 | * ranges are normally of the form [from, to). However, there | |
654 | * may be a "placeholder" entry in the map which is of the form | |
655 | * (from, to) with from == to. Check for placeholder entries | |
656 | * at the beginning of the range to be deleted. | |
657 | */ | |
658 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 659 | continue; |
dbe409e4 | 660 | |
feba16e2 | 661 | if (rg->from >= t) |
96822904 | 662 | break; |
96822904 | 663 | |
feba16e2 MK |
664 | if (f > rg->from && t < rg->to) { /* Must split region */ |
665 | /* | |
666 | * Check for an entry in the cache before dropping | |
667 | * lock and attempting allocation. | |
668 | */ | |
669 | if (!nrg && | |
670 | resv->region_cache_count > resv->adds_in_progress) { | |
671 | nrg = list_first_entry(&resv->region_cache, | |
672 | struct file_region, | |
673 | link); | |
674 | list_del(&nrg->link); | |
675 | resv->region_cache_count--; | |
676 | } | |
96822904 | 677 | |
feba16e2 MK |
678 | if (!nrg) { |
679 | spin_unlock(&resv->lock); | |
680 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
681 | if (!nrg) | |
682 | return -ENOMEM; | |
683 | goto retry; | |
684 | } | |
685 | ||
686 | del += t - f; | |
79aa925b | 687 | hugetlb_cgroup_uncharge_file_region( |
0dd70e6b | 688 | resv, rg, t - f, false); |
feba16e2 MK |
689 | |
690 | /* New entry for end of split region */ | |
691 | nrg->from = t; | |
692 | nrg->to = rg->to; | |
075a61d0 MA |
693 | |
694 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
695 | ||
feba16e2 MK |
696 | INIT_LIST_HEAD(&nrg->link); |
697 | ||
698 | /* Original entry is trimmed */ | |
699 | rg->to = f; | |
700 | ||
701 | list_add(&nrg->link, &rg->link); | |
702 | nrg = NULL; | |
96822904 | 703 | break; |
feba16e2 MK |
704 | } |
705 | ||
706 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
707 | del += rg->to - rg->from; | |
075a61d0 | 708 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
0dd70e6b | 709 | rg->to - rg->from, true); |
feba16e2 MK |
710 | list_del(&rg->link); |
711 | kfree(rg); | |
712 | continue; | |
713 | } | |
714 | ||
715 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 716 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
0dd70e6b | 717 | t - rg->from, false); |
075a61d0 | 718 | |
79aa925b MK |
719 | del += t - rg->from; |
720 | rg->from = t; | |
721 | } else { /* Trim end of region */ | |
075a61d0 | 722 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
0dd70e6b | 723 | rg->to - f, false); |
79aa925b MK |
724 | |
725 | del += rg->to - f; | |
726 | rg->to = f; | |
feba16e2 | 727 | } |
96822904 | 728 | } |
7b24d861 | 729 | |
7b24d861 | 730 | spin_unlock(&resv->lock); |
feba16e2 MK |
731 | kfree(nrg); |
732 | return del; | |
96822904 AW |
733 | } |
734 | ||
b5cec28d MK |
735 | /* |
736 | * A rare out of memory error was encountered which prevented removal of | |
737 | * the reserve map region for a page. The huge page itself was free'ed | |
738 | * and removed from the page cache. This routine will adjust the subpool | |
739 | * usage count, and the global reserve count if needed. By incrementing | |
740 | * these counts, the reserve map entry which could not be deleted will | |
741 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
742 | * counts. | |
743 | */ | |
72e2936c | 744 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
745 | { |
746 | struct hugepage_subpool *spool = subpool_inode(inode); | |
747 | long rsv_adjust; | |
748 | ||
749 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 750 | if (rsv_adjust) { |
b5cec28d MK |
751 | struct hstate *h = hstate_inode(inode); |
752 | ||
753 | hugetlb_acct_memory(h, 1); | |
754 | } | |
755 | } | |
756 | ||
1dd308a7 MK |
757 | /* |
758 | * Count and return the number of huge pages in the reserve map | |
759 | * that intersect with the range [f, t). | |
760 | */ | |
1406ec9b | 761 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 762 | { |
1406ec9b | 763 | struct list_head *head = &resv->regions; |
84afd99b AW |
764 | struct file_region *rg; |
765 | long chg = 0; | |
766 | ||
7b24d861 | 767 | spin_lock(&resv->lock); |
84afd99b AW |
768 | /* Locate each segment we overlap with, and count that overlap. */ |
769 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
770 | long seg_from; |
771 | long seg_to; | |
84afd99b AW |
772 | |
773 | if (rg->to <= f) | |
774 | continue; | |
775 | if (rg->from >= t) | |
776 | break; | |
777 | ||
778 | seg_from = max(rg->from, f); | |
779 | seg_to = min(rg->to, t); | |
780 | ||
781 | chg += seg_to - seg_from; | |
782 | } | |
7b24d861 | 783 | spin_unlock(&resv->lock); |
84afd99b AW |
784 | |
785 | return chg; | |
786 | } | |
787 | ||
e7c4b0bf AW |
788 | /* |
789 | * Convert the address within this vma to the page offset within | |
790 | * the mapping, in pagecache page units; huge pages here. | |
791 | */ | |
a5516438 AK |
792 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
793 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 794 | { |
a5516438 AK |
795 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
796 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
797 | } |
798 | ||
0fe6e20b NH |
799 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
800 | unsigned long address) | |
801 | { | |
802 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
803 | } | |
dee41079 | 804 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 805 | |
08fba699 MG |
806 | /* |
807 | * Return the size of the pages allocated when backing a VMA. In the majority | |
808 | * cases this will be same size as used by the page table entries. | |
809 | */ | |
810 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
811 | { | |
05ea8860 DW |
812 | if (vma->vm_ops && vma->vm_ops->pagesize) |
813 | return vma->vm_ops->pagesize(vma); | |
814 | return PAGE_SIZE; | |
08fba699 | 815 | } |
f340ca0f | 816 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 817 | |
3340289d MG |
818 | /* |
819 | * Return the page size being used by the MMU to back a VMA. In the majority | |
820 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
821 | * architectures where it differs, an architecture-specific 'strong' |
822 | * version of this symbol is required. | |
3340289d | 823 | */ |
09135cc5 | 824 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
825 | { |
826 | return vma_kernel_pagesize(vma); | |
827 | } | |
3340289d | 828 | |
84afd99b AW |
829 | /* |
830 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
831 | * bits of the reservation map pointer, which are always clear due to | |
832 | * alignment. | |
833 | */ | |
834 | #define HPAGE_RESV_OWNER (1UL << 0) | |
835 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 836 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 837 | |
a1e78772 MG |
838 | /* |
839 | * These helpers are used to track how many pages are reserved for | |
840 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
841 | * is guaranteed to have their future faults succeed. | |
842 | * | |
843 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
844 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
845 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
846 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
847 | * |
848 | * The private mapping reservation is represented in a subtly different | |
849 | * manner to a shared mapping. A shared mapping has a region map associated | |
850 | * with the underlying file, this region map represents the backing file | |
851 | * pages which have ever had a reservation assigned which this persists even | |
852 | * after the page is instantiated. A private mapping has a region map | |
853 | * associated with the original mmap which is attached to all VMAs which | |
854 | * reference it, this region map represents those offsets which have consumed | |
855 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 856 | */ |
e7c4b0bf AW |
857 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
858 | { | |
859 | return (unsigned long)vma->vm_private_data; | |
860 | } | |
861 | ||
862 | static void set_vma_private_data(struct vm_area_struct *vma, | |
863 | unsigned long value) | |
864 | { | |
865 | vma->vm_private_data = (void *)value; | |
866 | } | |
867 | ||
e9fe92ae MA |
868 | static void |
869 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
870 | struct hugetlb_cgroup *h_cg, | |
871 | struct hstate *h) | |
872 | { | |
873 | #ifdef CONFIG_CGROUP_HUGETLB | |
874 | if (!h_cg || !h) { | |
875 | resv_map->reservation_counter = NULL; | |
876 | resv_map->pages_per_hpage = 0; | |
877 | resv_map->css = NULL; | |
878 | } else { | |
879 | resv_map->reservation_counter = | |
880 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
881 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
882 | resv_map->css = &h_cg->css; | |
883 | } | |
884 | #endif | |
885 | } | |
886 | ||
9119a41e | 887 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
888 | { |
889 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
890 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
891 | ||
892 | if (!resv_map || !rg) { | |
893 | kfree(resv_map); | |
894 | kfree(rg); | |
84afd99b | 895 | return NULL; |
5e911373 | 896 | } |
84afd99b AW |
897 | |
898 | kref_init(&resv_map->refs); | |
7b24d861 | 899 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
900 | INIT_LIST_HEAD(&resv_map->regions); |
901 | ||
5e911373 | 902 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
903 | /* |
904 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
905 | * fields don't do cgroup accounting. On private mappings, these will be | |
906 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
907 | * reservations are to be un-charged from here. | |
908 | */ | |
909 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
910 | |
911 | INIT_LIST_HEAD(&resv_map->region_cache); | |
912 | list_add(&rg->link, &resv_map->region_cache); | |
913 | resv_map->region_cache_count = 1; | |
914 | ||
84afd99b AW |
915 | return resv_map; |
916 | } | |
917 | ||
9119a41e | 918 | void resv_map_release(struct kref *ref) |
84afd99b AW |
919 | { |
920 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
921 | struct list_head *head = &resv_map->region_cache; |
922 | struct file_region *rg, *trg; | |
84afd99b AW |
923 | |
924 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 925 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
926 | |
927 | /* ... and any entries left in the cache */ | |
928 | list_for_each_entry_safe(rg, trg, head, link) { | |
929 | list_del(&rg->link); | |
930 | kfree(rg); | |
931 | } | |
932 | ||
933 | VM_BUG_ON(resv_map->adds_in_progress); | |
934 | ||
84afd99b AW |
935 | kfree(resv_map); |
936 | } | |
937 | ||
4e35f483 JK |
938 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
939 | { | |
f27a5136 MK |
940 | /* |
941 | * At inode evict time, i_mapping may not point to the original | |
942 | * address space within the inode. This original address space | |
943 | * contains the pointer to the resv_map. So, always use the | |
944 | * address space embedded within the inode. | |
945 | * The VERY common case is inode->mapping == &inode->i_data but, | |
946 | * this may not be true for device special inodes. | |
947 | */ | |
948 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
949 | } |
950 | ||
84afd99b | 951 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 952 | { |
81d1b09c | 953 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
954 | if (vma->vm_flags & VM_MAYSHARE) { |
955 | struct address_space *mapping = vma->vm_file->f_mapping; | |
956 | struct inode *inode = mapping->host; | |
957 | ||
958 | return inode_resv_map(inode); | |
959 | ||
960 | } else { | |
84afd99b AW |
961 | return (struct resv_map *)(get_vma_private_data(vma) & |
962 | ~HPAGE_RESV_MASK); | |
4e35f483 | 963 | } |
a1e78772 MG |
964 | } |
965 | ||
84afd99b | 966 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 967 | { |
81d1b09c SL |
968 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
969 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 970 | |
84afd99b AW |
971 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
972 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
973 | } |
974 | ||
975 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
976 | { | |
81d1b09c SL |
977 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
978 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
979 | |
980 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
981 | } |
982 | ||
983 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
984 | { | |
81d1b09c | 985 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
986 | |
987 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
988 | } |
989 | ||
04f2cbe3 | 990 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
991 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
992 | { | |
81d1b09c | 993 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 994 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
995 | vma->vm_private_data = (void *)0; |
996 | } | |
997 | ||
998 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 999 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1000 | { |
af0ed73e JK |
1001 | if (vma->vm_flags & VM_NORESERVE) { |
1002 | /* | |
1003 | * This address is already reserved by other process(chg == 0), | |
1004 | * so, we should decrement reserved count. Without decrementing, | |
1005 | * reserve count remains after releasing inode, because this | |
1006 | * allocated page will go into page cache and is regarded as | |
1007 | * coming from reserved pool in releasing step. Currently, we | |
1008 | * don't have any other solution to deal with this situation | |
1009 | * properly, so add work-around here. | |
1010 | */ | |
1011 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1012 | return true; |
af0ed73e | 1013 | else |
559ec2f8 | 1014 | return false; |
af0ed73e | 1015 | } |
a63884e9 JK |
1016 | |
1017 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1018 | if (vma->vm_flags & VM_MAYSHARE) { |
1019 | /* | |
1020 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1021 | * be a region map for all pages. The only situation where | |
1022 | * there is no region map is if a hole was punched via | |
7c8de358 | 1023 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1024 | * use. This situation is indicated if chg != 0. |
1025 | */ | |
1026 | if (chg) | |
1027 | return false; | |
1028 | else | |
1029 | return true; | |
1030 | } | |
a63884e9 JK |
1031 | |
1032 | /* | |
1033 | * Only the process that called mmap() has reserves for | |
1034 | * private mappings. | |
1035 | */ | |
67961f9d MK |
1036 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1037 | /* | |
1038 | * Like the shared case above, a hole punch or truncate | |
1039 | * could have been performed on the private mapping. | |
1040 | * Examine the value of chg to determine if reserves | |
1041 | * actually exist or were previously consumed. | |
1042 | * Very Subtle - The value of chg comes from a previous | |
1043 | * call to vma_needs_reserves(). The reserve map for | |
1044 | * private mappings has different (opposite) semantics | |
1045 | * than that of shared mappings. vma_needs_reserves() | |
1046 | * has already taken this difference in semantics into | |
1047 | * account. Therefore, the meaning of chg is the same | |
1048 | * as in the shared case above. Code could easily be | |
1049 | * combined, but keeping it separate draws attention to | |
1050 | * subtle differences. | |
1051 | */ | |
1052 | if (chg) | |
1053 | return false; | |
1054 | else | |
1055 | return true; | |
1056 | } | |
a63884e9 | 1057 | |
559ec2f8 | 1058 | return false; |
a1e78772 MG |
1059 | } |
1060 | ||
a5516438 | 1061 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1062 | { |
1063 | int nid = page_to_nid(page); | |
0edaecfa | 1064 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1065 | h->free_huge_pages++; |
1066 | h->free_huge_pages_node[nid]++; | |
7ffddd49 | 1067 | SetPageHugeFreed(page); |
1da177e4 LT |
1068 | } |
1069 | ||
94310cbc | 1070 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1071 | { |
1072 | struct page *page; | |
bbe88753 JK |
1073 | bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); |
1074 | ||
1075 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { | |
1076 | if (nocma && is_migrate_cma_page(page)) | |
1077 | continue; | |
bf50bab2 | 1078 | |
6664bfc8 WY |
1079 | if (PageHWPoison(page)) |
1080 | continue; | |
1081 | ||
1082 | list_move(&page->lru, &h->hugepage_activelist); | |
1083 | set_page_refcounted(page); | |
7ffddd49 | 1084 | ClearPageHugeFreed(page); |
6664bfc8 WY |
1085 | h->free_huge_pages--; |
1086 | h->free_huge_pages_node[nid]--; | |
1087 | return page; | |
bbe88753 JK |
1088 | } |
1089 | ||
6664bfc8 | 1090 | return NULL; |
bf50bab2 NH |
1091 | } |
1092 | ||
3e59fcb0 MH |
1093 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1094 | nodemask_t *nmask) | |
94310cbc | 1095 | { |
3e59fcb0 MH |
1096 | unsigned int cpuset_mems_cookie; |
1097 | struct zonelist *zonelist; | |
1098 | struct zone *zone; | |
1099 | struct zoneref *z; | |
98fa15f3 | 1100 | int node = NUMA_NO_NODE; |
94310cbc | 1101 | |
3e59fcb0 MH |
1102 | zonelist = node_zonelist(nid, gfp_mask); |
1103 | ||
1104 | retry_cpuset: | |
1105 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1106 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1107 | struct page *page; | |
1108 | ||
1109 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1110 | continue; | |
1111 | /* | |
1112 | * no need to ask again on the same node. Pool is node rather than | |
1113 | * zone aware | |
1114 | */ | |
1115 | if (zone_to_nid(zone) == node) | |
1116 | continue; | |
1117 | node = zone_to_nid(zone); | |
94310cbc | 1118 | |
94310cbc AK |
1119 | page = dequeue_huge_page_node_exact(h, node); |
1120 | if (page) | |
1121 | return page; | |
1122 | } | |
3e59fcb0 MH |
1123 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1124 | goto retry_cpuset; | |
1125 | ||
94310cbc AK |
1126 | return NULL; |
1127 | } | |
1128 | ||
a5516438 AK |
1129 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1130 | struct vm_area_struct *vma, | |
af0ed73e JK |
1131 | unsigned long address, int avoid_reserve, |
1132 | long chg) | |
1da177e4 | 1133 | { |
3e59fcb0 | 1134 | struct page *page; |
480eccf9 | 1135 | struct mempolicy *mpol; |
04ec6264 | 1136 | gfp_t gfp_mask; |
3e59fcb0 | 1137 | nodemask_t *nodemask; |
04ec6264 | 1138 | int nid; |
1da177e4 | 1139 | |
a1e78772 MG |
1140 | /* |
1141 | * A child process with MAP_PRIVATE mappings created by their parent | |
1142 | * have no page reserves. This check ensures that reservations are | |
1143 | * not "stolen". The child may still get SIGKILLed | |
1144 | */ | |
af0ed73e | 1145 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1146 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1147 | goto err; |
a1e78772 | 1148 | |
04f2cbe3 | 1149 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1150 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1151 | goto err; |
04f2cbe3 | 1152 | |
04ec6264 VB |
1153 | gfp_mask = htlb_alloc_mask(h); |
1154 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
1155 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
1156 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
1157 | SetPagePrivate(page); | |
1158 | h->resv_huge_pages--; | |
1da177e4 | 1159 | } |
cc9a6c87 | 1160 | |
52cd3b07 | 1161 | mpol_cond_put(mpol); |
1da177e4 | 1162 | return page; |
cc9a6c87 MG |
1163 | |
1164 | err: | |
cc9a6c87 | 1165 | return NULL; |
1da177e4 LT |
1166 | } |
1167 | ||
1cac6f2c LC |
1168 | /* |
1169 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1170 | * We may have allocated or freed a huge page based on a different | |
1171 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1172 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1173 | * node for alloc or free. | |
1174 | */ | |
1175 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1176 | { | |
0edaf86c | 1177 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1178 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1179 | ||
1180 | return nid; | |
1181 | } | |
1182 | ||
1183 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1184 | { | |
1185 | if (!node_isset(nid, *nodes_allowed)) | |
1186 | nid = next_node_allowed(nid, nodes_allowed); | |
1187 | return nid; | |
1188 | } | |
1189 | ||
1190 | /* | |
1191 | * returns the previously saved node ["this node"] from which to | |
1192 | * allocate a persistent huge page for the pool and advance the | |
1193 | * next node from which to allocate, handling wrap at end of node | |
1194 | * mask. | |
1195 | */ | |
1196 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1197 | nodemask_t *nodes_allowed) | |
1198 | { | |
1199 | int nid; | |
1200 | ||
1201 | VM_BUG_ON(!nodes_allowed); | |
1202 | ||
1203 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1204 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1205 | ||
1206 | return nid; | |
1207 | } | |
1208 | ||
1209 | /* | |
1210 | * helper for free_pool_huge_page() - return the previously saved | |
1211 | * node ["this node"] from which to free a huge page. Advance the | |
1212 | * next node id whether or not we find a free huge page to free so | |
1213 | * that the next attempt to free addresses the next node. | |
1214 | */ | |
1215 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1216 | { | |
1217 | int nid; | |
1218 | ||
1219 | VM_BUG_ON(!nodes_allowed); | |
1220 | ||
1221 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1222 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1223 | ||
1224 | return nid; | |
1225 | } | |
1226 | ||
1227 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1228 | for (nr_nodes = nodes_weight(*mask); \ | |
1229 | nr_nodes > 0 && \ | |
1230 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1231 | nr_nodes--) | |
1232 | ||
1233 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1234 | for (nr_nodes = nodes_weight(*mask); \ | |
1235 | nr_nodes > 0 && \ | |
1236 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1237 | nr_nodes--) | |
1238 | ||
e1073d1e | 1239 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1240 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1241 | unsigned int order) |
944d9fec LC |
1242 | { |
1243 | int i; | |
1244 | int nr_pages = 1 << order; | |
1245 | struct page *p = page + 1; | |
1246 | ||
c8cc708a | 1247 | atomic_set(compound_mapcount_ptr(page), 0); |
47e29d32 JH |
1248 | if (hpage_pincount_available(page)) |
1249 | atomic_set(compound_pincount_ptr(page), 0); | |
1250 | ||
944d9fec | 1251 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1252 | clear_compound_head(p); |
944d9fec | 1253 | set_page_refcounted(p); |
944d9fec LC |
1254 | } |
1255 | ||
1256 | set_compound_order(page, 0); | |
ba9c1201 | 1257 | page[1].compound_nr = 0; |
944d9fec LC |
1258 | __ClearPageHead(page); |
1259 | } | |
1260 | ||
d00181b9 | 1261 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1262 | { |
cf11e85f RG |
1263 | /* |
1264 | * If the page isn't allocated using the cma allocator, | |
1265 | * cma_release() returns false. | |
1266 | */ | |
dbda8fea BS |
1267 | #ifdef CONFIG_CMA |
1268 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1269 | return; |
dbda8fea | 1270 | #endif |
cf11e85f | 1271 | |
944d9fec LC |
1272 | free_contig_range(page_to_pfn(page), 1 << order); |
1273 | } | |
1274 | ||
4eb0716e | 1275 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1276 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1277 | int nid, nodemask_t *nodemask) | |
944d9fec | 1278 | { |
5e27a2df | 1279 | unsigned long nr_pages = 1UL << huge_page_order(h); |
953f064a LX |
1280 | if (nid == NUMA_NO_NODE) |
1281 | nid = numa_mem_id(); | |
944d9fec | 1282 | |
dbda8fea BS |
1283 | #ifdef CONFIG_CMA |
1284 | { | |
cf11e85f RG |
1285 | struct page *page; |
1286 | int node; | |
1287 | ||
953f064a LX |
1288 | if (hugetlb_cma[nid]) { |
1289 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1290 | huge_page_order(h), true); | |
cf11e85f RG |
1291 | if (page) |
1292 | return page; | |
1293 | } | |
953f064a LX |
1294 | |
1295 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1296 | for_each_node_mask(node, *nodemask) { | |
1297 | if (node == nid || !hugetlb_cma[node]) | |
1298 | continue; | |
1299 | ||
1300 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1301 | huge_page_order(h), true); | |
1302 | if (page) | |
1303 | return page; | |
1304 | } | |
1305 | } | |
cf11e85f | 1306 | } |
dbda8fea | 1307 | #endif |
cf11e85f | 1308 | |
5e27a2df | 1309 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1310 | } |
1311 | ||
1312 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1313 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1314 | #else /* !CONFIG_CONTIG_ALLOC */ |
1315 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1316 | int nid, nodemask_t *nodemask) | |
1317 | { | |
1318 | return NULL; | |
1319 | } | |
1320 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1321 | |
e1073d1e | 1322 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1323 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1324 | int nid, nodemask_t *nodemask) |
1325 | { | |
1326 | return NULL; | |
1327 | } | |
d00181b9 | 1328 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1329 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1330 | unsigned int order) { } |
944d9fec LC |
1331 | #endif |
1332 | ||
a5516438 | 1333 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1334 | { |
1335 | int i; | |
526c1d39 | 1336 | struct page *subpage = page; |
a5516438 | 1337 | |
4eb0716e | 1338 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1339 | return; |
18229df5 | 1340 | |
a5516438 AK |
1341 | h->nr_huge_pages--; |
1342 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
526c1d39 MK |
1343 | for (i = 0; i < pages_per_huge_page(h); |
1344 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1345 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1346 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1347 | 1 << PG_active | 1 << PG_private | |
1348 | 1 << PG_writeback); | |
6af2acb6 | 1349 | } |
309381fe | 1350 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
1adc4d41 | 1351 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); |
f1e61557 | 1352 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1353 | set_page_refcounted(page); |
944d9fec | 1354 | if (hstate_is_gigantic(h)) { |
cf11e85f RG |
1355 | /* |
1356 | * Temporarily drop the hugetlb_lock, because | |
1357 | * we might block in free_gigantic_page(). | |
1358 | */ | |
1359 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1360 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1361 | free_gigantic_page(page, huge_page_order(h)); | |
cf11e85f | 1362 | spin_lock(&hugetlb_lock); |
944d9fec | 1363 | } else { |
944d9fec LC |
1364 | __free_pages(page, huge_page_order(h)); |
1365 | } | |
6af2acb6 AL |
1366 | } |
1367 | ||
e5ff2159 AK |
1368 | struct hstate *size_to_hstate(unsigned long size) |
1369 | { | |
1370 | struct hstate *h; | |
1371 | ||
1372 | for_each_hstate(h) { | |
1373 | if (huge_page_size(h) == size) | |
1374 | return h; | |
1375 | } | |
1376 | return NULL; | |
1377 | } | |
1378 | ||
bcc54222 NH |
1379 | /* |
1380 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1381 | * to hstate->hugepage_activelist.) | |
1382 | * | |
1383 | * This function can be called for tail pages, but never returns true for them. | |
1384 | */ | |
1385 | bool page_huge_active(struct page *page) | |
1386 | { | |
ecbf4724 | 1387 | return PageHeadHuge(page) && PagePrivate(&page[1]); |
bcc54222 NH |
1388 | } |
1389 | ||
1390 | /* never called for tail page */ | |
585fc0d2 | 1391 | void set_page_huge_active(struct page *page) |
bcc54222 NH |
1392 | { |
1393 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1394 | SetPagePrivate(&page[1]); | |
1395 | } | |
1396 | ||
1397 | static void clear_page_huge_active(struct page *page) | |
1398 | { | |
1399 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1400 | ClearPagePrivate(&page[1]); | |
1401 | } | |
1402 | ||
ab5ac90a MH |
1403 | /* |
1404 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1405 | * code | |
1406 | */ | |
1407 | static inline bool PageHugeTemporary(struct page *page) | |
1408 | { | |
1409 | if (!PageHuge(page)) | |
1410 | return false; | |
1411 | ||
1412 | return (unsigned long)page[2].mapping == -1U; | |
1413 | } | |
1414 | ||
1415 | static inline void SetPageHugeTemporary(struct page *page) | |
1416 | { | |
1417 | page[2].mapping = (void *)-1U; | |
1418 | } | |
1419 | ||
1420 | static inline void ClearPageHugeTemporary(struct page *page) | |
1421 | { | |
1422 | page[2].mapping = NULL; | |
1423 | } | |
1424 | ||
c77c0a8a | 1425 | static void __free_huge_page(struct page *page) |
27a85ef1 | 1426 | { |
a5516438 AK |
1427 | /* |
1428 | * Can't pass hstate in here because it is called from the | |
1429 | * compound page destructor. | |
1430 | */ | |
e5ff2159 | 1431 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1432 | int nid = page_to_nid(page); |
90481622 DG |
1433 | struct hugepage_subpool *spool = |
1434 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1435 | bool restore_reserve; |
27a85ef1 | 1436 | |
b4330afb MK |
1437 | VM_BUG_ON_PAGE(page_count(page), page); |
1438 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc YW |
1439 | |
1440 | set_page_private(page, 0); | |
1441 | page->mapping = NULL; | |
07443a85 | 1442 | restore_reserve = PagePrivate(page); |
16c794b4 | 1443 | ClearPagePrivate(page); |
27a85ef1 | 1444 | |
1c5ecae3 | 1445 | /* |
0919e1b6 MK |
1446 | * If PagePrivate() was set on page, page allocation consumed a |
1447 | * reservation. If the page was associated with a subpool, there | |
1448 | * would have been a page reserved in the subpool before allocation | |
1449 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1450 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1451 | * remove the reserved page from the subpool. | |
1c5ecae3 | 1452 | */ |
0919e1b6 MK |
1453 | if (!restore_reserve) { |
1454 | /* | |
1455 | * A return code of zero implies that the subpool will be | |
1456 | * under its minimum size if the reservation is not restored | |
1457 | * after page is free. Therefore, force restore_reserve | |
1458 | * operation. | |
1459 | */ | |
1460 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1461 | restore_reserve = true; | |
1462 | } | |
1c5ecae3 | 1463 | |
27a85ef1 | 1464 | spin_lock(&hugetlb_lock); |
bcc54222 | 1465 | clear_page_huge_active(page); |
6d76dcf4 AK |
1466 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1467 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1468 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1469 | pages_per_huge_page(h), page); | |
07443a85 JK |
1470 | if (restore_reserve) |
1471 | h->resv_huge_pages++; | |
1472 | ||
ab5ac90a MH |
1473 | if (PageHugeTemporary(page)) { |
1474 | list_del(&page->lru); | |
1475 | ClearPageHugeTemporary(page); | |
1476 | update_and_free_page(h, page); | |
1477 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1478 | /* remove the page from active list */ |
1479 | list_del(&page->lru); | |
a5516438 AK |
1480 | update_and_free_page(h, page); |
1481 | h->surplus_huge_pages--; | |
1482 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1483 | } else { |
5d3a551c | 1484 | arch_clear_hugepage_flags(page); |
a5516438 | 1485 | enqueue_huge_page(h, page); |
7893d1d5 | 1486 | } |
27a85ef1 DG |
1487 | spin_unlock(&hugetlb_lock); |
1488 | } | |
1489 | ||
c77c0a8a WL |
1490 | /* |
1491 | * As free_huge_page() can be called from a non-task context, we have | |
1492 | * to defer the actual freeing in a workqueue to prevent potential | |
1493 | * hugetlb_lock deadlock. | |
1494 | * | |
1495 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1496 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1497 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1498 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1499 | */ | |
1500 | static LLIST_HEAD(hpage_freelist); | |
1501 | ||
1502 | static void free_hpage_workfn(struct work_struct *work) | |
1503 | { | |
1504 | struct llist_node *node; | |
1505 | struct page *page; | |
1506 | ||
1507 | node = llist_del_all(&hpage_freelist); | |
1508 | ||
1509 | while (node) { | |
1510 | page = container_of((struct address_space **)node, | |
1511 | struct page, mapping); | |
1512 | node = node->next; | |
1513 | __free_huge_page(page); | |
1514 | } | |
1515 | } | |
1516 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1517 | ||
1518 | void free_huge_page(struct page *page) | |
1519 | { | |
1520 | /* | |
1521 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1522 | */ | |
1523 | if (!in_task()) { | |
1524 | /* | |
1525 | * Only call schedule_work() if hpage_freelist is previously | |
1526 | * empty. Otherwise, schedule_work() had been called but the | |
1527 | * workfn hasn't retrieved the list yet. | |
1528 | */ | |
1529 | if (llist_add((struct llist_node *)&page->mapping, | |
1530 | &hpage_freelist)) | |
1531 | schedule_work(&free_hpage_work); | |
1532 | return; | |
1533 | } | |
1534 | ||
1535 | __free_huge_page(page); | |
1536 | } | |
1537 | ||
a5516438 | 1538 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1539 | { |
0edaecfa | 1540 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1541 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
9dd540e2 | 1542 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1543 | set_hugetlb_cgroup_rsvd(page, NULL); |
2f37511c | 1544 | spin_lock(&hugetlb_lock); |
a5516438 AK |
1545 | h->nr_huge_pages++; |
1546 | h->nr_huge_pages_node[nid]++; | |
7ffddd49 | 1547 | ClearPageHugeFreed(page); |
b7ba30c6 | 1548 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1549 | } |
1550 | ||
d00181b9 | 1551 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1552 | { |
1553 | int i; | |
1554 | int nr_pages = 1 << order; | |
1555 | struct page *p = page + 1; | |
1556 | ||
1557 | /* we rely on prep_new_huge_page to set the destructor */ | |
1558 | set_compound_order(page, order); | |
ef5a22be | 1559 | __ClearPageReserved(page); |
de09d31d | 1560 | __SetPageHead(page); |
20a0307c | 1561 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1562 | /* |
1563 | * For gigantic hugepages allocated through bootmem at | |
1564 | * boot, it's safer to be consistent with the not-gigantic | |
1565 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1566 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1567 | * pages may get the reference counting wrong if they see |
1568 | * PG_reserved set on a tail page (despite the head page not | |
1569 | * having PG_reserved set). Enforcing this consistency between | |
1570 | * head and tail pages allows drivers to optimize away a check | |
1571 | * on the head page when they need know if put_page() is needed | |
1572 | * after get_user_pages(). | |
1573 | */ | |
1574 | __ClearPageReserved(p); | |
58a84aa9 | 1575 | set_page_count(p, 0); |
1d798ca3 | 1576 | set_compound_head(p, page); |
20a0307c | 1577 | } |
b4330afb | 1578 | atomic_set(compound_mapcount_ptr(page), -1); |
47e29d32 JH |
1579 | |
1580 | if (hpage_pincount_available(page)) | |
1581 | atomic_set(compound_pincount_ptr(page), 0); | |
20a0307c WF |
1582 | } |
1583 | ||
7795912c AM |
1584 | /* |
1585 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1586 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1587 | * details. | |
1588 | */ | |
20a0307c WF |
1589 | int PageHuge(struct page *page) |
1590 | { | |
20a0307c WF |
1591 | if (!PageCompound(page)) |
1592 | return 0; | |
1593 | ||
1594 | page = compound_head(page); | |
f1e61557 | 1595 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1596 | } |
43131e14 NH |
1597 | EXPORT_SYMBOL_GPL(PageHuge); |
1598 | ||
27c73ae7 AA |
1599 | /* |
1600 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1601 | * normal or transparent huge pages. | |
1602 | */ | |
1603 | int PageHeadHuge(struct page *page_head) | |
1604 | { | |
27c73ae7 AA |
1605 | if (!PageHead(page_head)) |
1606 | return 0; | |
1607 | ||
d4af73e3 | 1608 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1609 | } |
27c73ae7 | 1610 | |
c0d0381a MK |
1611 | /* |
1612 | * Find and lock address space (mapping) in write mode. | |
1613 | * | |
336bf30e MK |
1614 | * Upon entry, the page is locked which means that page_mapping() is |
1615 | * stable. Due to locking order, we can only trylock_write. If we can | |
1616 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1617 | */ |
1618 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1619 | { | |
336bf30e | 1620 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1621 | |
c0d0381a MK |
1622 | if (!mapping) |
1623 | return mapping; | |
1624 | ||
c0d0381a MK |
1625 | if (i_mmap_trylock_write(mapping)) |
1626 | return mapping; | |
1627 | ||
336bf30e | 1628 | return NULL; |
c0d0381a MK |
1629 | } |
1630 | ||
13d60f4b ZY |
1631 | pgoff_t __basepage_index(struct page *page) |
1632 | { | |
1633 | struct page *page_head = compound_head(page); | |
1634 | pgoff_t index = page_index(page_head); | |
1635 | unsigned long compound_idx; | |
1636 | ||
1637 | if (!PageHuge(page_head)) | |
1638 | return page_index(page); | |
1639 | ||
1640 | if (compound_order(page_head) >= MAX_ORDER) | |
1641 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1642 | else | |
1643 | compound_idx = page - page_head; | |
1644 | ||
1645 | return (index << compound_order(page_head)) + compound_idx; | |
1646 | } | |
1647 | ||
0c397dae | 1648 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1649 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1650 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1651 | { |
af0fb9df | 1652 | int order = huge_page_order(h); |
1da177e4 | 1653 | struct page *page; |
f60858f9 | 1654 | bool alloc_try_hard = true; |
f96efd58 | 1655 | |
f60858f9 MK |
1656 | /* |
1657 | * By default we always try hard to allocate the page with | |
1658 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1659 | * a loop (to adjust global huge page counts) and previous allocation | |
1660 | * failed, do not continue to try hard on the same node. Use the | |
1661 | * node_alloc_noretry bitmap to manage this state information. | |
1662 | */ | |
1663 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1664 | alloc_try_hard = false; | |
1665 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1666 | if (alloc_try_hard) | |
1667 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1668 | if (nid == NUMA_NO_NODE) |
1669 | nid = numa_mem_id(); | |
1670 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1671 | if (page) | |
1672 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1673 | else | |
1674 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1675 | |
f60858f9 MK |
1676 | /* |
1677 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1678 | * indicates an overall state change. Clear bit so that we resume | |
1679 | * normal 'try hard' allocations. | |
1680 | */ | |
1681 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1682 | node_clear(nid, *node_alloc_noretry); | |
1683 | ||
1684 | /* | |
1685 | * If we tried hard to get a page but failed, set bit so that | |
1686 | * subsequent attempts will not try as hard until there is an | |
1687 | * overall state change. | |
1688 | */ | |
1689 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1690 | node_set(nid, *node_alloc_noretry); | |
1691 | ||
63b4613c NA |
1692 | return page; |
1693 | } | |
1694 | ||
0c397dae MH |
1695 | /* |
1696 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1697 | * should use this function to get new hugetlb pages | |
1698 | */ | |
1699 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1700 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1701 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1702 | { |
1703 | struct page *page; | |
1704 | ||
1705 | if (hstate_is_gigantic(h)) | |
1706 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1707 | else | |
1708 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1709 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1710 | if (!page) |
1711 | return NULL; | |
1712 | ||
1713 | if (hstate_is_gigantic(h)) | |
1714 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1715 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1716 | ||
1717 | return page; | |
1718 | } | |
1719 | ||
af0fb9df MH |
1720 | /* |
1721 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1722 | * manner. | |
1723 | */ | |
f60858f9 MK |
1724 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1725 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1726 | { |
1727 | struct page *page; | |
1728 | int nr_nodes, node; | |
af0fb9df | 1729 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1730 | |
1731 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1732 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1733 | node_alloc_noretry); | |
af0fb9df | 1734 | if (page) |
b2261026 | 1735 | break; |
b2261026 JK |
1736 | } |
1737 | ||
af0fb9df MH |
1738 | if (!page) |
1739 | return 0; | |
b2261026 | 1740 | |
af0fb9df MH |
1741 | put_page(page); /* free it into the hugepage allocator */ |
1742 | ||
1743 | return 1; | |
b2261026 JK |
1744 | } |
1745 | ||
e8c5c824 LS |
1746 | /* |
1747 | * Free huge page from pool from next node to free. | |
1748 | * Attempt to keep persistent huge pages more or less | |
1749 | * balanced over allowed nodes. | |
1750 | * Called with hugetlb_lock locked. | |
1751 | */ | |
6ae11b27 LS |
1752 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1753 | bool acct_surplus) | |
e8c5c824 | 1754 | { |
b2261026 | 1755 | int nr_nodes, node; |
e8c5c824 LS |
1756 | int ret = 0; |
1757 | ||
b2261026 | 1758 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1759 | /* |
1760 | * If we're returning unused surplus pages, only examine | |
1761 | * nodes with surplus pages. | |
1762 | */ | |
b2261026 JK |
1763 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1764 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1765 | struct page *page = |
b2261026 | 1766 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1767 | struct page, lru); |
1768 | list_del(&page->lru); | |
1769 | h->free_huge_pages--; | |
b2261026 | 1770 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1771 | if (acct_surplus) { |
1772 | h->surplus_huge_pages--; | |
b2261026 | 1773 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1774 | } |
e8c5c824 LS |
1775 | update_and_free_page(h, page); |
1776 | ret = 1; | |
9a76db09 | 1777 | break; |
e8c5c824 | 1778 | } |
b2261026 | 1779 | } |
e8c5c824 LS |
1780 | |
1781 | return ret; | |
1782 | } | |
1783 | ||
c8721bbb NH |
1784 | /* |
1785 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1786 | * nothing for in-use hugepages and non-hugepages. |
1787 | * This function returns values like below: | |
1788 | * | |
1789 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1790 | * (allocated or reserved.) | |
1791 | * 0: successfully dissolved free hugepages or the page is not a | |
1792 | * hugepage (considered as already dissolved) | |
c8721bbb | 1793 | */ |
c3114a84 | 1794 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1795 | { |
6bc9b564 | 1796 | int rc = -EBUSY; |
082d5b6b | 1797 | |
7ffddd49 | 1798 | retry: |
faf53def NH |
1799 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1800 | if (!PageHuge(page)) | |
1801 | return 0; | |
1802 | ||
c8721bbb | 1803 | spin_lock(&hugetlb_lock); |
faf53def NH |
1804 | if (!PageHuge(page)) { |
1805 | rc = 0; | |
1806 | goto out; | |
1807 | } | |
1808 | ||
1809 | if (!page_count(page)) { | |
2247bb33 GS |
1810 | struct page *head = compound_head(page); |
1811 | struct hstate *h = page_hstate(head); | |
1812 | int nid = page_to_nid(head); | |
6bc9b564 | 1813 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1814 | goto out; |
7ffddd49 MS |
1815 | |
1816 | /* | |
1817 | * We should make sure that the page is already on the free list | |
1818 | * when it is dissolved. | |
1819 | */ | |
1820 | if (unlikely(!PageHugeFreed(head))) { | |
1821 | spin_unlock(&hugetlb_lock); | |
1822 | cond_resched(); | |
1823 | ||
1824 | /* | |
1825 | * Theoretically, we should return -EBUSY when we | |
1826 | * encounter this race. In fact, we have a chance | |
1827 | * to successfully dissolve the page if we do a | |
1828 | * retry. Because the race window is quite small. | |
1829 | * If we seize this opportunity, it is an optimization | |
1830 | * for increasing the success rate of dissolving page. | |
1831 | */ | |
1832 | goto retry; | |
1833 | } | |
1834 | ||
c3114a84 AK |
1835 | /* |
1836 | * Move PageHWPoison flag from head page to the raw error page, | |
1837 | * which makes any subpages rather than the error page reusable. | |
1838 | */ | |
1839 | if (PageHWPoison(head) && page != head) { | |
1840 | SetPageHWPoison(page); | |
1841 | ClearPageHWPoison(head); | |
1842 | } | |
2247bb33 | 1843 | list_del(&head->lru); |
c8721bbb NH |
1844 | h->free_huge_pages--; |
1845 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1846 | h->max_huge_pages--; |
2247bb33 | 1847 | update_and_free_page(h, head); |
6bc9b564 | 1848 | rc = 0; |
c8721bbb | 1849 | } |
082d5b6b | 1850 | out: |
c8721bbb | 1851 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1852 | return rc; |
c8721bbb NH |
1853 | } |
1854 | ||
1855 | /* | |
1856 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1857 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1858 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1859 | * part of it lies within the given range. | |
082d5b6b GS |
1860 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1861 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1862 | */ |
082d5b6b | 1863 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1864 | { |
c8721bbb | 1865 | unsigned long pfn; |
eb03aa00 | 1866 | struct page *page; |
082d5b6b | 1867 | int rc = 0; |
c8721bbb | 1868 | |
d0177639 | 1869 | if (!hugepages_supported()) |
082d5b6b | 1870 | return rc; |
d0177639 | 1871 | |
eb03aa00 GS |
1872 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1873 | page = pfn_to_page(pfn); | |
faf53def NH |
1874 | rc = dissolve_free_huge_page(page); |
1875 | if (rc) | |
1876 | break; | |
eb03aa00 | 1877 | } |
082d5b6b GS |
1878 | |
1879 | return rc; | |
c8721bbb NH |
1880 | } |
1881 | ||
ab5ac90a MH |
1882 | /* |
1883 | * Allocates a fresh surplus page from the page allocator. | |
1884 | */ | |
0c397dae | 1885 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1886 | int nid, nodemask_t *nmask) |
7893d1d5 | 1887 | { |
9980d744 | 1888 | struct page *page = NULL; |
7893d1d5 | 1889 | |
bae7f4ae | 1890 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1891 | return NULL; |
1892 | ||
d1c3fb1f | 1893 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1894 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1895 | goto out_unlock; | |
d1c3fb1f NA |
1896 | spin_unlock(&hugetlb_lock); |
1897 | ||
f60858f9 | 1898 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1899 | if (!page) |
0c397dae | 1900 | return NULL; |
d1c3fb1f NA |
1901 | |
1902 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1903 | /* |
1904 | * We could have raced with the pool size change. | |
1905 | * Double check that and simply deallocate the new page | |
1906 | * if we would end up overcommiting the surpluses. Abuse | |
1907 | * temporary page to workaround the nasty free_huge_page | |
1908 | * codeflow | |
1909 | */ | |
1910 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1911 | SetPageHugeTemporary(page); | |
2bf753e6 | 1912 | spin_unlock(&hugetlb_lock); |
9980d744 | 1913 | put_page(page); |
2bf753e6 | 1914 | return NULL; |
9980d744 | 1915 | } else { |
9980d744 | 1916 | h->surplus_huge_pages++; |
4704dea3 | 1917 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1918 | } |
9980d744 MH |
1919 | |
1920 | out_unlock: | |
d1c3fb1f | 1921 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1922 | |
1923 | return page; | |
1924 | } | |
1925 | ||
bbe88753 | 1926 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 1927 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
1928 | { |
1929 | struct page *page; | |
1930 | ||
1931 | if (hstate_is_gigantic(h)) | |
1932 | return NULL; | |
1933 | ||
f60858f9 | 1934 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1935 | if (!page) |
1936 | return NULL; | |
1937 | ||
1938 | /* | |
1939 | * We do not account these pages as surplus because they are only | |
1940 | * temporary and will be released properly on the last reference | |
1941 | */ | |
ab5ac90a MH |
1942 | SetPageHugeTemporary(page); |
1943 | ||
1944 | return page; | |
1945 | } | |
1946 | ||
099730d6 DH |
1947 | /* |
1948 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1949 | */ | |
e0ec90ee | 1950 | static |
0c397dae | 1951 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1952 | struct vm_area_struct *vma, unsigned long addr) |
1953 | { | |
aaf14e40 MH |
1954 | struct page *page; |
1955 | struct mempolicy *mpol; | |
1956 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1957 | int nid; | |
1958 | nodemask_t *nodemask; | |
1959 | ||
1960 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1961 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1962 | mpol_cond_put(mpol); |
1963 | ||
1964 | return page; | |
099730d6 DH |
1965 | } |
1966 | ||
ab5ac90a | 1967 | /* page migration callback function */ |
3e59fcb0 | 1968 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 1969 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 1970 | { |
4db9b2ef MH |
1971 | spin_lock(&hugetlb_lock); |
1972 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1973 | struct page *page; |
1974 | ||
1975 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1976 | if (page) { | |
1977 | spin_unlock(&hugetlb_lock); | |
1978 | return page; | |
4db9b2ef MH |
1979 | } |
1980 | } | |
1981 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1982 | |
0c397dae | 1983 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1984 | } |
1985 | ||
ebd63723 | 1986 | /* mempolicy aware migration callback */ |
389c8178 MH |
1987 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1988 | unsigned long address) | |
ebd63723 MH |
1989 | { |
1990 | struct mempolicy *mpol; | |
1991 | nodemask_t *nodemask; | |
1992 | struct page *page; | |
ebd63723 MH |
1993 | gfp_t gfp_mask; |
1994 | int node; | |
1995 | ||
ebd63723 MH |
1996 | gfp_mask = htlb_alloc_mask(h); |
1997 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 1998 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
1999 | mpol_cond_put(mpol); |
2000 | ||
2001 | return page; | |
2002 | } | |
2003 | ||
e4e574b7 | 2004 | /* |
25985edc | 2005 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2006 | * of size 'delta'. |
2007 | */ | |
0a4f3d1b | 2008 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 2009 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
2010 | { |
2011 | struct list_head surplus_list; | |
2012 | struct page *page, *tmp; | |
0a4f3d1b LX |
2013 | int ret; |
2014 | long i; | |
2015 | long needed, allocated; | |
28073b02 | 2016 | bool alloc_ok = true; |
e4e574b7 | 2017 | |
a5516438 | 2018 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2019 | if (needed <= 0) { |
a5516438 | 2020 | h->resv_huge_pages += delta; |
e4e574b7 | 2021 | return 0; |
ac09b3a1 | 2022 | } |
e4e574b7 AL |
2023 | |
2024 | allocated = 0; | |
2025 | INIT_LIST_HEAD(&surplus_list); | |
2026 | ||
2027 | ret = -ENOMEM; | |
2028 | retry: | |
2029 | spin_unlock(&hugetlb_lock); | |
2030 | for (i = 0; i < needed; i++) { | |
0c397dae | 2031 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 2032 | NUMA_NO_NODE, NULL); |
28073b02 HD |
2033 | if (!page) { |
2034 | alloc_ok = false; | |
2035 | break; | |
2036 | } | |
e4e574b7 | 2037 | list_add(&page->lru, &surplus_list); |
69ed779a | 2038 | cond_resched(); |
e4e574b7 | 2039 | } |
28073b02 | 2040 | allocated += i; |
e4e574b7 AL |
2041 | |
2042 | /* | |
2043 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2044 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2045 | */ | |
2046 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
2047 | needed = (h->resv_huge_pages + delta) - |
2048 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2049 | if (needed > 0) { |
2050 | if (alloc_ok) | |
2051 | goto retry; | |
2052 | /* | |
2053 | * We were not able to allocate enough pages to | |
2054 | * satisfy the entire reservation so we free what | |
2055 | * we've allocated so far. | |
2056 | */ | |
2057 | goto free; | |
2058 | } | |
e4e574b7 AL |
2059 | /* |
2060 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2061 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2062 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2063 | * allocator. Commit the entire reservation here to prevent another |
2064 | * process from stealing the pages as they are added to the pool but | |
2065 | * before they are reserved. | |
e4e574b7 AL |
2066 | */ |
2067 | needed += allocated; | |
a5516438 | 2068 | h->resv_huge_pages += delta; |
e4e574b7 | 2069 | ret = 0; |
a9869b83 | 2070 | |
19fc3f0a | 2071 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2072 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
e558464b MS |
2073 | int zeroed; |
2074 | ||
19fc3f0a AL |
2075 | if ((--needed) < 0) |
2076 | break; | |
a9869b83 NH |
2077 | /* |
2078 | * This page is now managed by the hugetlb allocator and has | |
2079 | * no users -- drop the buddy allocator's reference. | |
2080 | */ | |
e558464b MS |
2081 | zeroed = put_page_testzero(page); |
2082 | VM_BUG_ON_PAGE(!zeroed, page); | |
a5516438 | 2083 | enqueue_huge_page(h, page); |
19fc3f0a | 2084 | } |
28073b02 | 2085 | free: |
b0365c8d | 2086 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
2087 | |
2088 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
2089 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
2090 | put_page(page); | |
a9869b83 | 2091 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
2092 | |
2093 | return ret; | |
2094 | } | |
2095 | ||
2096 | /* | |
e5bbc8a6 MK |
2097 | * This routine has two main purposes: |
2098 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2099 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2100 | * to the associated reservation map. | |
2101 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2102 | * the reservation. As many as unused_resv_pages may be freed. | |
2103 | * | |
2104 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
2105 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
2106 | * we must make sure nobody else can claim pages we are in the process of | |
2107 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
2108 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 2109 | */ |
a5516438 AK |
2110 | static void return_unused_surplus_pages(struct hstate *h, |
2111 | unsigned long unused_resv_pages) | |
e4e574b7 | 2112 | { |
e4e574b7 AL |
2113 | unsigned long nr_pages; |
2114 | ||
aa888a74 | 2115 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2116 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2117 | goto out; |
aa888a74 | 2118 | |
e5bbc8a6 MK |
2119 | /* |
2120 | * Part (or even all) of the reservation could have been backed | |
2121 | * by pre-allocated pages. Only free surplus pages. | |
2122 | */ | |
a5516438 | 2123 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2124 | |
685f3457 LS |
2125 | /* |
2126 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2127 | * evenly across all nodes with memory. Iterate across these nodes |
2128 | * until we can no longer free unreserved surplus pages. This occurs | |
2129 | * when the nodes with surplus pages have no free pages. | |
9e7ee400 | 2130 | * free_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2131 | * on-line nodes with memory and will handle the hstate accounting. |
e5bbc8a6 MK |
2132 | * |
2133 | * Note that we decrement resv_huge_pages as we free the pages. If | |
2134 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
2135 | * to cover subsequent pages we may free. | |
685f3457 LS |
2136 | */ |
2137 | while (nr_pages--) { | |
e5bbc8a6 MK |
2138 | h->resv_huge_pages--; |
2139 | unused_resv_pages--; | |
8cebfcd0 | 2140 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 2141 | goto out; |
7848a4bf | 2142 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 2143 | } |
e5bbc8a6 MK |
2144 | |
2145 | out: | |
2146 | /* Fully uncommit the reservation */ | |
2147 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
2148 | } |
2149 | ||
5e911373 | 2150 | |
c37f9fb1 | 2151 | /* |
feba16e2 | 2152 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2153 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2154 | * |
2155 | * vma_needs_reservation is called to determine if the huge page at addr | |
2156 | * within the vma has an associated reservation. If a reservation is | |
2157 | * needed, the value 1 is returned. The caller is then responsible for | |
2158 | * managing the global reservation and subpool usage counts. After | |
2159 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2160 | * to add the page to the reservation map. If the page allocation fails, |
2161 | * the reservation must be ended instead of committed. vma_end_reservation | |
2162 | * is called in such cases. | |
cf3ad20b MK |
2163 | * |
2164 | * In the normal case, vma_commit_reservation returns the same value | |
2165 | * as the preceding vma_needs_reservation call. The only time this | |
2166 | * is not the case is if a reserve map was changed between calls. It | |
2167 | * is the responsibility of the caller to notice the difference and | |
2168 | * take appropriate action. | |
96b96a96 MK |
2169 | * |
2170 | * vma_add_reservation is used in error paths where a reservation must | |
2171 | * be restored when a newly allocated huge page must be freed. It is | |
2172 | * to be called after calling vma_needs_reservation to determine if a | |
2173 | * reservation exists. | |
c37f9fb1 | 2174 | */ |
5e911373 MK |
2175 | enum vma_resv_mode { |
2176 | VMA_NEEDS_RESV, | |
2177 | VMA_COMMIT_RESV, | |
feba16e2 | 2178 | VMA_END_RESV, |
96b96a96 | 2179 | VMA_ADD_RESV, |
5e911373 | 2180 | }; |
cf3ad20b MK |
2181 | static long __vma_reservation_common(struct hstate *h, |
2182 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2183 | enum vma_resv_mode mode) |
c37f9fb1 | 2184 | { |
4e35f483 JK |
2185 | struct resv_map *resv; |
2186 | pgoff_t idx; | |
cf3ad20b | 2187 | long ret; |
0db9d74e | 2188 | long dummy_out_regions_needed; |
c37f9fb1 | 2189 | |
4e35f483 JK |
2190 | resv = vma_resv_map(vma); |
2191 | if (!resv) | |
84afd99b | 2192 | return 1; |
c37f9fb1 | 2193 | |
4e35f483 | 2194 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2195 | switch (mode) { |
2196 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2197 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2198 | /* We assume that vma_reservation_* routines always operate on | |
2199 | * 1 page, and that adding to resv map a 1 page entry can only | |
2200 | * ever require 1 region. | |
2201 | */ | |
2202 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2203 | break; |
2204 | case VMA_COMMIT_RESV: | |
075a61d0 | 2205 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2206 | /* region_add calls of range 1 should never fail. */ |
2207 | VM_BUG_ON(ret < 0); | |
5e911373 | 2208 | break; |
feba16e2 | 2209 | case VMA_END_RESV: |
0db9d74e | 2210 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2211 | ret = 0; |
2212 | break; | |
96b96a96 | 2213 | case VMA_ADD_RESV: |
0db9d74e | 2214 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2215 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2216 | /* region_add calls of range 1 should never fail. */ |
2217 | VM_BUG_ON(ret < 0); | |
2218 | } else { | |
2219 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2220 | ret = region_del(resv, idx, idx + 1); |
2221 | } | |
2222 | break; | |
5e911373 MK |
2223 | default: |
2224 | BUG(); | |
2225 | } | |
84afd99b | 2226 | |
4e35f483 | 2227 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 2228 | return ret; |
67961f9d MK |
2229 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
2230 | /* | |
2231 | * In most cases, reserves always exist for private mappings. | |
2232 | * However, a file associated with mapping could have been | |
2233 | * hole punched or truncated after reserves were consumed. | |
2234 | * As subsequent fault on such a range will not use reserves. | |
2235 | * Subtle - The reserve map for private mappings has the | |
2236 | * opposite meaning than that of shared mappings. If NO | |
2237 | * entry is in the reserve map, it means a reservation exists. | |
2238 | * If an entry exists in the reserve map, it means the | |
2239 | * reservation has already been consumed. As a result, the | |
2240 | * return value of this routine is the opposite of the | |
2241 | * value returned from reserve map manipulation routines above. | |
2242 | */ | |
2243 | if (ret) | |
2244 | return 0; | |
2245 | else | |
2246 | return 1; | |
2247 | } | |
4e35f483 | 2248 | else |
cf3ad20b | 2249 | return ret < 0 ? ret : 0; |
c37f9fb1 | 2250 | } |
cf3ad20b MK |
2251 | |
2252 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2253 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2254 | { |
5e911373 | 2255 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2256 | } |
84afd99b | 2257 | |
cf3ad20b MK |
2258 | static long vma_commit_reservation(struct hstate *h, |
2259 | struct vm_area_struct *vma, unsigned long addr) | |
2260 | { | |
5e911373 MK |
2261 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2262 | } | |
2263 | ||
feba16e2 | 2264 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2265 | struct vm_area_struct *vma, unsigned long addr) |
2266 | { | |
feba16e2 | 2267 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2268 | } |
2269 | ||
96b96a96 MK |
2270 | static long vma_add_reservation(struct hstate *h, |
2271 | struct vm_area_struct *vma, unsigned long addr) | |
2272 | { | |
2273 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2274 | } | |
2275 | ||
2276 | /* | |
2277 | * This routine is called to restore a reservation on error paths. In the | |
2278 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2279 | * and is about to be freed. If a reservation for the page existed, | |
2280 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2281 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2282 | * the global reservation count will be incremented if PagePrivate is set. | |
2283 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2284 | * reserve map here to be consistent with global reserve count adjustments | |
2285 | * to be made by free_huge_page. | |
2286 | */ | |
2287 | static void restore_reserve_on_error(struct hstate *h, | |
2288 | struct vm_area_struct *vma, unsigned long address, | |
2289 | struct page *page) | |
2290 | { | |
2291 | if (unlikely(PagePrivate(page))) { | |
2292 | long rc = vma_needs_reservation(h, vma, address); | |
2293 | ||
2294 | if (unlikely(rc < 0)) { | |
2295 | /* | |
2296 | * Rare out of memory condition in reserve map | |
2297 | * manipulation. Clear PagePrivate so that | |
2298 | * global reserve count will not be incremented | |
2299 | * by free_huge_page. This will make it appear | |
2300 | * as though the reservation for this page was | |
2301 | * consumed. This may prevent the task from | |
2302 | * faulting in the page at a later time. This | |
2303 | * is better than inconsistent global huge page | |
2304 | * accounting of reserve counts. | |
2305 | */ | |
2306 | ClearPagePrivate(page); | |
2307 | } else if (rc) { | |
2308 | rc = vma_add_reservation(h, vma, address); | |
2309 | if (unlikely(rc < 0)) | |
2310 | /* | |
2311 | * See above comment about rare out of | |
2312 | * memory condition. | |
2313 | */ | |
2314 | ClearPagePrivate(page); | |
2315 | } else | |
2316 | vma_end_reservation(h, vma, address); | |
2317 | } | |
2318 | } | |
2319 | ||
70c3547e | 2320 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2321 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2322 | { |
90481622 | 2323 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2324 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2325 | struct page *page; |
d85f69b0 MK |
2326 | long map_chg, map_commit; |
2327 | long gbl_chg; | |
6d76dcf4 AK |
2328 | int ret, idx; |
2329 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2330 | bool deferred_reserve; |
a1e78772 | 2331 | |
6d76dcf4 | 2332 | idx = hstate_index(h); |
a1e78772 | 2333 | /* |
d85f69b0 MK |
2334 | * Examine the region/reserve map to determine if the process |
2335 | * has a reservation for the page to be allocated. A return | |
2336 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2337 | */ |
d85f69b0 MK |
2338 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2339 | if (map_chg < 0) | |
76dcee75 | 2340 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2341 | |
2342 | /* | |
2343 | * Processes that did not create the mapping will have no | |
2344 | * reserves as indicated by the region/reserve map. Check | |
2345 | * that the allocation will not exceed the subpool limit. | |
2346 | * Allocations for MAP_NORESERVE mappings also need to be | |
2347 | * checked against any subpool limit. | |
2348 | */ | |
2349 | if (map_chg || avoid_reserve) { | |
2350 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2351 | if (gbl_chg < 0) { | |
feba16e2 | 2352 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2353 | return ERR_PTR(-ENOSPC); |
5e911373 | 2354 | } |
1da177e4 | 2355 | |
d85f69b0 MK |
2356 | /* |
2357 | * Even though there was no reservation in the region/reserve | |
2358 | * map, there could be reservations associated with the | |
2359 | * subpool that can be used. This would be indicated if the | |
2360 | * return value of hugepage_subpool_get_pages() is zero. | |
2361 | * However, if avoid_reserve is specified we still avoid even | |
2362 | * the subpool reservations. | |
2363 | */ | |
2364 | if (avoid_reserve) | |
2365 | gbl_chg = 1; | |
2366 | } | |
2367 | ||
08cf9faf MA |
2368 | /* If this allocation is not consuming a reservation, charge it now. |
2369 | */ | |
2370 | deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); | |
2371 | if (deferred_reserve) { | |
2372 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2373 | idx, pages_per_huge_page(h), &h_cg); | |
2374 | if (ret) | |
2375 | goto out_subpool_put; | |
2376 | } | |
2377 | ||
6d76dcf4 | 2378 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2379 | if (ret) |
08cf9faf | 2380 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2381 | |
1da177e4 | 2382 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2383 | /* |
2384 | * glb_chg is passed to indicate whether or not a page must be taken | |
2385 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2386 | * a reservation exists for the allocation. | |
2387 | */ | |
2388 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2389 | if (!page) { |
94ae8ba7 | 2390 | spin_unlock(&hugetlb_lock); |
0c397dae | 2391 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2392 | if (!page) |
2393 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2394 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2395 | SetPagePrivate(page); | |
2396 | h->resv_huge_pages--; | |
2397 | } | |
79dbb236 | 2398 | spin_lock(&hugetlb_lock); |
15a8d68e | 2399 | list_add(&page->lru, &h->hugepage_activelist); |
81a6fcae | 2400 | /* Fall through */ |
68842c9b | 2401 | } |
81a6fcae | 2402 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2403 | /* If allocation is not consuming a reservation, also store the |
2404 | * hugetlb_cgroup pointer on the page. | |
2405 | */ | |
2406 | if (deferred_reserve) { | |
2407 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2408 | h_cg, page); | |
2409 | } | |
2410 | ||
81a6fcae | 2411 | spin_unlock(&hugetlb_lock); |
348ea204 | 2412 | |
90481622 | 2413 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2414 | |
d85f69b0 MK |
2415 | map_commit = vma_commit_reservation(h, vma, addr); |
2416 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2417 | /* |
2418 | * The page was added to the reservation map between | |
2419 | * vma_needs_reservation and vma_commit_reservation. | |
2420 | * This indicates a race with hugetlb_reserve_pages. | |
2421 | * Adjust for the subpool count incremented above AND | |
2422 | * in hugetlb_reserve_pages for the same page. Also, | |
2423 | * the reservation count added in hugetlb_reserve_pages | |
2424 | * no longer applies. | |
2425 | */ | |
2426 | long rsv_adjust; | |
2427 | ||
2428 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2429 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2430 | if (deferred_reserve) |
2431 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2432 | pages_per_huge_page(h), page); | |
33039678 | 2433 | } |
90d8b7e6 | 2434 | return page; |
8f34af6f JZ |
2435 | |
2436 | out_uncharge_cgroup: | |
2437 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2438 | out_uncharge_cgroup_reservation: |
2439 | if (deferred_reserve) | |
2440 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2441 | h_cg); | |
8f34af6f | 2442 | out_subpool_put: |
d85f69b0 | 2443 | if (map_chg || avoid_reserve) |
8f34af6f | 2444 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2445 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2446 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2447 | } |
2448 | ||
e24a1307 AK |
2449 | int alloc_bootmem_huge_page(struct hstate *h) |
2450 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2451 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2452 | { |
2453 | struct huge_bootmem_page *m; | |
b2261026 | 2454 | int nr_nodes, node; |
aa888a74 | 2455 | |
b2261026 | 2456 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2457 | void *addr; |
2458 | ||
eb31d559 | 2459 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2460 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2461 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2462 | if (addr) { |
2463 | /* | |
2464 | * Use the beginning of the huge page to store the | |
2465 | * huge_bootmem_page struct (until gather_bootmem | |
2466 | * puts them into the mem_map). | |
2467 | */ | |
2468 | m = addr; | |
91f47662 | 2469 | goto found; |
aa888a74 | 2470 | } |
aa888a74 AK |
2471 | } |
2472 | return 0; | |
2473 | ||
2474 | found: | |
df994ead | 2475 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2476 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2477 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2478 | list_add(&m->list, &huge_boot_pages); |
2479 | m->hstate = h; | |
2480 | return 1; | |
2481 | } | |
2482 | ||
d00181b9 KS |
2483 | static void __init prep_compound_huge_page(struct page *page, |
2484 | unsigned int order) | |
18229df5 AW |
2485 | { |
2486 | if (unlikely(order > (MAX_ORDER - 1))) | |
2487 | prep_compound_gigantic_page(page, order); | |
2488 | else | |
2489 | prep_compound_page(page, order); | |
2490 | } | |
2491 | ||
aa888a74 AK |
2492 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2493 | static void __init gather_bootmem_prealloc(void) | |
2494 | { | |
2495 | struct huge_bootmem_page *m; | |
2496 | ||
2497 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2498 | struct page *page = virt_to_page(m); |
aa888a74 | 2499 | struct hstate *h = m->hstate; |
ee8f248d | 2500 | |
aa888a74 | 2501 | WARN_ON(page_count(page) != 1); |
18229df5 | 2502 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2503 | WARN_ON(PageReserved(page)); |
aa888a74 | 2504 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2505 | put_page(page); /* free it into the hugepage allocator */ |
2506 | ||
b0320c7b RA |
2507 | /* |
2508 | * If we had gigantic hugepages allocated at boot time, we need | |
2509 | * to restore the 'stolen' pages to totalram_pages in order to | |
2510 | * fix confusing memory reports from free(1) and another | |
2511 | * side-effects, like CommitLimit going negative. | |
2512 | */ | |
bae7f4ae | 2513 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2514 | adjust_managed_page_count(page, 1 << h->order); |
520495fe | 2515 | cond_resched(); |
aa888a74 AK |
2516 | } |
2517 | } | |
2518 | ||
8faa8b07 | 2519 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2520 | { |
2521 | unsigned long i; | |
f60858f9 MK |
2522 | nodemask_t *node_alloc_noretry; |
2523 | ||
2524 | if (!hstate_is_gigantic(h)) { | |
2525 | /* | |
2526 | * Bit mask controlling how hard we retry per-node allocations. | |
2527 | * Ignore errors as lower level routines can deal with | |
2528 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2529 | * time, we are likely in bigger trouble. | |
2530 | */ | |
2531 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2532 | GFP_KERNEL); | |
2533 | } else { | |
2534 | /* allocations done at boot time */ | |
2535 | node_alloc_noretry = NULL; | |
2536 | } | |
2537 | ||
2538 | /* bit mask controlling how hard we retry per-node allocations */ | |
2539 | if (node_alloc_noretry) | |
2540 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2541 | |
e5ff2159 | 2542 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2543 | if (hstate_is_gigantic(h)) { |
dbda8fea | 2544 | if (hugetlb_cma_size) { |
cf11e85f | 2545 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); |
1ed11ee2 | 2546 | goto free; |
cf11e85f | 2547 | } |
aa888a74 AK |
2548 | if (!alloc_bootmem_huge_page(h)) |
2549 | break; | |
0c397dae | 2550 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2551 | &node_states[N_MEMORY], |
2552 | node_alloc_noretry)) | |
1da177e4 | 2553 | break; |
69ed779a | 2554 | cond_resched(); |
1da177e4 | 2555 | } |
d715cf80 LH |
2556 | if (i < h->max_huge_pages) { |
2557 | char buf[32]; | |
2558 | ||
c6247f72 | 2559 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2560 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2561 | h->max_huge_pages, buf, i); | |
2562 | h->max_huge_pages = i; | |
2563 | } | |
1ed11ee2 | 2564 | free: |
f60858f9 | 2565 | kfree(node_alloc_noretry); |
e5ff2159 AK |
2566 | } |
2567 | ||
2568 | static void __init hugetlb_init_hstates(void) | |
2569 | { | |
2570 | struct hstate *h; | |
2571 | ||
2572 | for_each_hstate(h) { | |
641844f5 NH |
2573 | if (minimum_order > huge_page_order(h)) |
2574 | minimum_order = huge_page_order(h); | |
2575 | ||
8faa8b07 | 2576 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2577 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2578 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2579 | } |
641844f5 | 2580 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2581 | } |
2582 | ||
2583 | static void __init report_hugepages(void) | |
2584 | { | |
2585 | struct hstate *h; | |
2586 | ||
2587 | for_each_hstate(h) { | |
4abd32db | 2588 | char buf[32]; |
c6247f72 MW |
2589 | |
2590 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2591 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2592 | buf, h->free_huge_pages); |
e5ff2159 AK |
2593 | } |
2594 | } | |
2595 | ||
1da177e4 | 2596 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2597 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2598 | nodemask_t *nodes_allowed) | |
1da177e4 | 2599 | { |
4415cc8d CL |
2600 | int i; |
2601 | ||
bae7f4ae | 2602 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2603 | return; |
2604 | ||
6ae11b27 | 2605 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2606 | struct page *page, *next; |
a5516438 AK |
2607 | struct list_head *freel = &h->hugepage_freelists[i]; |
2608 | list_for_each_entry_safe(page, next, freel, lru) { | |
2609 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2610 | return; |
1da177e4 LT |
2611 | if (PageHighMem(page)) |
2612 | continue; | |
2613 | list_del(&page->lru); | |
e5ff2159 | 2614 | update_and_free_page(h, page); |
a5516438 AK |
2615 | h->free_huge_pages--; |
2616 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2617 | } |
2618 | } | |
2619 | } | |
2620 | #else | |
6ae11b27 LS |
2621 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2622 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2623 | { |
2624 | } | |
2625 | #endif | |
2626 | ||
20a0307c WF |
2627 | /* |
2628 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2629 | * balanced by operating on them in a round-robin fashion. | |
2630 | * Returns 1 if an adjustment was made. | |
2631 | */ | |
6ae11b27 LS |
2632 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2633 | int delta) | |
20a0307c | 2634 | { |
b2261026 | 2635 | int nr_nodes, node; |
20a0307c WF |
2636 | |
2637 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2638 | |
b2261026 JK |
2639 | if (delta < 0) { |
2640 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2641 | if (h->surplus_huge_pages_node[node]) | |
2642 | goto found; | |
e8c5c824 | 2643 | } |
b2261026 JK |
2644 | } else { |
2645 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2646 | if (h->surplus_huge_pages_node[node] < | |
2647 | h->nr_huge_pages_node[node]) | |
2648 | goto found; | |
e8c5c824 | 2649 | } |
b2261026 JK |
2650 | } |
2651 | return 0; | |
20a0307c | 2652 | |
b2261026 JK |
2653 | found: |
2654 | h->surplus_huge_pages += delta; | |
2655 | h->surplus_huge_pages_node[node] += delta; | |
2656 | return 1; | |
20a0307c WF |
2657 | } |
2658 | ||
a5516438 | 2659 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2660 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2661 | nodemask_t *nodes_allowed) |
1da177e4 | 2662 | { |
7893d1d5 | 2663 | unsigned long min_count, ret; |
f60858f9 MK |
2664 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2665 | ||
2666 | /* | |
2667 | * Bit mask controlling how hard we retry per-node allocations. | |
2668 | * If we can not allocate the bit mask, do not attempt to allocate | |
2669 | * the requested huge pages. | |
2670 | */ | |
2671 | if (node_alloc_noretry) | |
2672 | nodes_clear(*node_alloc_noretry); | |
2673 | else | |
2674 | return -ENOMEM; | |
1da177e4 | 2675 | |
4eb0716e AG |
2676 | spin_lock(&hugetlb_lock); |
2677 | ||
fd875dca MK |
2678 | /* |
2679 | * Check for a node specific request. | |
2680 | * Changing node specific huge page count may require a corresponding | |
2681 | * change to the global count. In any case, the passed node mask | |
2682 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2683 | */ | |
2684 | if (nid != NUMA_NO_NODE) { | |
2685 | unsigned long old_count = count; | |
2686 | ||
2687 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2688 | /* | |
2689 | * User may have specified a large count value which caused the | |
2690 | * above calculation to overflow. In this case, they wanted | |
2691 | * to allocate as many huge pages as possible. Set count to | |
2692 | * largest possible value to align with their intention. | |
2693 | */ | |
2694 | if (count < old_count) | |
2695 | count = ULONG_MAX; | |
2696 | } | |
2697 | ||
4eb0716e AG |
2698 | /* |
2699 | * Gigantic pages runtime allocation depend on the capability for large | |
2700 | * page range allocation. | |
2701 | * If the system does not provide this feature, return an error when | |
2702 | * the user tries to allocate gigantic pages but let the user free the | |
2703 | * boottime allocated gigantic pages. | |
2704 | */ | |
2705 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2706 | if (count > persistent_huge_pages(h)) { | |
2707 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2708 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2709 | return -EINVAL; |
2710 | } | |
2711 | /* Fall through to decrease pool */ | |
2712 | } | |
aa888a74 | 2713 | |
7893d1d5 AL |
2714 | /* |
2715 | * Increase the pool size | |
2716 | * First take pages out of surplus state. Then make up the | |
2717 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2718 | * |
0c397dae | 2719 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2720 | * to convert a surplus huge page to a normal huge page. That is |
2721 | * not critical, though, it just means the overall size of the | |
2722 | * pool might be one hugepage larger than it needs to be, but | |
2723 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2724 | */ |
a5516438 | 2725 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2726 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2727 | break; |
2728 | } | |
2729 | ||
a5516438 | 2730 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2731 | /* |
2732 | * If this allocation races such that we no longer need the | |
2733 | * page, free_huge_page will handle it by freeing the page | |
2734 | * and reducing the surplus. | |
2735 | */ | |
2736 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2737 | |
2738 | /* yield cpu to avoid soft lockup */ | |
2739 | cond_resched(); | |
2740 | ||
f60858f9 MK |
2741 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2742 | node_alloc_noretry); | |
7893d1d5 AL |
2743 | spin_lock(&hugetlb_lock); |
2744 | if (!ret) | |
2745 | goto out; | |
2746 | ||
536240f2 MG |
2747 | /* Bail for signals. Probably ctrl-c from user */ |
2748 | if (signal_pending(current)) | |
2749 | goto out; | |
7893d1d5 | 2750 | } |
7893d1d5 AL |
2751 | |
2752 | /* | |
2753 | * Decrease the pool size | |
2754 | * First return free pages to the buddy allocator (being careful | |
2755 | * to keep enough around to satisfy reservations). Then place | |
2756 | * pages into surplus state as needed so the pool will shrink | |
2757 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2758 | * |
2759 | * By placing pages into the surplus state independent of the | |
2760 | * overcommit value, we are allowing the surplus pool size to | |
2761 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2762 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2763 | * though, we'll note that we're not allowed to exceed surplus |
2764 | * and won't grow the pool anywhere else. Not until one of the | |
2765 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2766 | */ |
a5516438 | 2767 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2768 | min_count = max(count, min_count); |
6ae11b27 | 2769 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2770 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2771 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2772 | break; |
55f67141 | 2773 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2774 | } |
a5516438 | 2775 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2776 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2777 | break; |
2778 | } | |
2779 | out: | |
4eb0716e | 2780 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2781 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2782 | |
f60858f9 MK |
2783 | NODEMASK_FREE(node_alloc_noretry); |
2784 | ||
4eb0716e | 2785 | return 0; |
1da177e4 LT |
2786 | } |
2787 | ||
a3437870 NA |
2788 | #define HSTATE_ATTR_RO(_name) \ |
2789 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2790 | ||
2791 | #define HSTATE_ATTR(_name) \ | |
2792 | static struct kobj_attribute _name##_attr = \ | |
2793 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2794 | ||
2795 | static struct kobject *hugepages_kobj; | |
2796 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2797 | ||
9a305230 LS |
2798 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2799 | ||
2800 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2801 | { |
2802 | int i; | |
9a305230 | 2803 | |
a3437870 | 2804 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2805 | if (hstate_kobjs[i] == kobj) { |
2806 | if (nidp) | |
2807 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2808 | return &hstates[i]; |
9a305230 LS |
2809 | } |
2810 | ||
2811 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2812 | } |
2813 | ||
06808b08 | 2814 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2815 | struct kobj_attribute *attr, char *buf) |
2816 | { | |
9a305230 LS |
2817 | struct hstate *h; |
2818 | unsigned long nr_huge_pages; | |
2819 | int nid; | |
2820 | ||
2821 | h = kobj_to_hstate(kobj, &nid); | |
2822 | if (nid == NUMA_NO_NODE) | |
2823 | nr_huge_pages = h->nr_huge_pages; | |
2824 | else | |
2825 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2826 | ||
ae7a927d | 2827 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 2828 | } |
adbe8726 | 2829 | |
238d3c13 DR |
2830 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2831 | struct hstate *h, int nid, | |
2832 | unsigned long count, size_t len) | |
a3437870 NA |
2833 | { |
2834 | int err; | |
2d0adf7e | 2835 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2836 | |
2d0adf7e OS |
2837 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2838 | return -EINVAL; | |
adbe8726 | 2839 | |
9a305230 LS |
2840 | if (nid == NUMA_NO_NODE) { |
2841 | /* | |
2842 | * global hstate attribute | |
2843 | */ | |
2844 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2845 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2846 | n_mask = &node_states[N_MEMORY]; | |
2847 | else | |
2848 | n_mask = &nodes_allowed; | |
2849 | } else { | |
9a305230 | 2850 | /* |
fd875dca MK |
2851 | * Node specific request. count adjustment happens in |
2852 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2853 | */ |
2d0adf7e OS |
2854 | init_nodemask_of_node(&nodes_allowed, nid); |
2855 | n_mask = &nodes_allowed; | |
fd875dca | 2856 | } |
9a305230 | 2857 | |
2d0adf7e | 2858 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2859 | |
4eb0716e | 2860 | return err ? err : len; |
06808b08 LS |
2861 | } |
2862 | ||
238d3c13 DR |
2863 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2864 | struct kobject *kobj, const char *buf, | |
2865 | size_t len) | |
2866 | { | |
2867 | struct hstate *h; | |
2868 | unsigned long count; | |
2869 | int nid; | |
2870 | int err; | |
2871 | ||
2872 | err = kstrtoul(buf, 10, &count); | |
2873 | if (err) | |
2874 | return err; | |
2875 | ||
2876 | h = kobj_to_hstate(kobj, &nid); | |
2877 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2878 | } | |
2879 | ||
06808b08 LS |
2880 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2881 | struct kobj_attribute *attr, char *buf) | |
2882 | { | |
2883 | return nr_hugepages_show_common(kobj, attr, buf); | |
2884 | } | |
2885 | ||
2886 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2887 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2888 | { | |
238d3c13 | 2889 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2890 | } |
2891 | HSTATE_ATTR(nr_hugepages); | |
2892 | ||
06808b08 LS |
2893 | #ifdef CONFIG_NUMA |
2894 | ||
2895 | /* | |
2896 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2897 | * huge page alloc/free. | |
2898 | */ | |
2899 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
2900 | struct kobj_attribute *attr, |
2901 | char *buf) | |
06808b08 LS |
2902 | { |
2903 | return nr_hugepages_show_common(kobj, attr, buf); | |
2904 | } | |
2905 | ||
2906 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2907 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2908 | { | |
238d3c13 | 2909 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2910 | } |
2911 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2912 | #endif | |
2913 | ||
2914 | ||
a3437870 NA |
2915 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2916 | struct kobj_attribute *attr, char *buf) | |
2917 | { | |
9a305230 | 2918 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 2919 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 2920 | } |
adbe8726 | 2921 | |
a3437870 NA |
2922 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2923 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2924 | { | |
2925 | int err; | |
2926 | unsigned long input; | |
9a305230 | 2927 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2928 | |
bae7f4ae | 2929 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2930 | return -EINVAL; |
2931 | ||
3dbb95f7 | 2932 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2933 | if (err) |
73ae31e5 | 2934 | return err; |
a3437870 NA |
2935 | |
2936 | spin_lock(&hugetlb_lock); | |
2937 | h->nr_overcommit_huge_pages = input; | |
2938 | spin_unlock(&hugetlb_lock); | |
2939 | ||
2940 | return count; | |
2941 | } | |
2942 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2943 | ||
2944 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2945 | struct kobj_attribute *attr, char *buf) | |
2946 | { | |
9a305230 LS |
2947 | struct hstate *h; |
2948 | unsigned long free_huge_pages; | |
2949 | int nid; | |
2950 | ||
2951 | h = kobj_to_hstate(kobj, &nid); | |
2952 | if (nid == NUMA_NO_NODE) | |
2953 | free_huge_pages = h->free_huge_pages; | |
2954 | else | |
2955 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2956 | ||
ae7a927d | 2957 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
2958 | } |
2959 | HSTATE_ATTR_RO(free_hugepages); | |
2960 | ||
2961 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2962 | struct kobj_attribute *attr, char *buf) | |
2963 | { | |
9a305230 | 2964 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 2965 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
2966 | } |
2967 | HSTATE_ATTR_RO(resv_hugepages); | |
2968 | ||
2969 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2970 | struct kobj_attribute *attr, char *buf) | |
2971 | { | |
9a305230 LS |
2972 | struct hstate *h; |
2973 | unsigned long surplus_huge_pages; | |
2974 | int nid; | |
2975 | ||
2976 | h = kobj_to_hstate(kobj, &nid); | |
2977 | if (nid == NUMA_NO_NODE) | |
2978 | surplus_huge_pages = h->surplus_huge_pages; | |
2979 | else | |
2980 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2981 | ||
ae7a927d | 2982 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
2983 | } |
2984 | HSTATE_ATTR_RO(surplus_hugepages); | |
2985 | ||
2986 | static struct attribute *hstate_attrs[] = { | |
2987 | &nr_hugepages_attr.attr, | |
2988 | &nr_overcommit_hugepages_attr.attr, | |
2989 | &free_hugepages_attr.attr, | |
2990 | &resv_hugepages_attr.attr, | |
2991 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2992 | #ifdef CONFIG_NUMA |
2993 | &nr_hugepages_mempolicy_attr.attr, | |
2994 | #endif | |
a3437870 NA |
2995 | NULL, |
2996 | }; | |
2997 | ||
67e5ed96 | 2998 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2999 | .attrs = hstate_attrs, |
3000 | }; | |
3001 | ||
094e9539 JM |
3002 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3003 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3004 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3005 | { |
3006 | int retval; | |
972dc4de | 3007 | int hi = hstate_index(h); |
a3437870 | 3008 | |
9a305230 LS |
3009 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3010 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3011 | return -ENOMEM; |
3012 | ||
9a305230 | 3013 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
2710104c | 3014 | if (retval) { |
9a305230 | 3015 | kobject_put(hstate_kobjs[hi]); |
2710104c ML |
3016 | hstate_kobjs[hi] = NULL; |
3017 | } | |
a3437870 NA |
3018 | |
3019 | return retval; | |
3020 | } | |
3021 | ||
3022 | static void __init hugetlb_sysfs_init(void) | |
3023 | { | |
3024 | struct hstate *h; | |
3025 | int err; | |
3026 | ||
3027 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3028 | if (!hugepages_kobj) | |
3029 | return; | |
3030 | ||
3031 | for_each_hstate(h) { | |
9a305230 LS |
3032 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3033 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3034 | if (err) |
282f4214 | 3035 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
3036 | } |
3037 | } | |
3038 | ||
9a305230 LS |
3039 | #ifdef CONFIG_NUMA |
3040 | ||
3041 | /* | |
3042 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3043 | * with node devices in node_devices[] using a parallel array. The array |
3044 | * index of a node device or _hstate == node id. | |
3045 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3046 | * the base kernel, on the hugetlb module. |
3047 | */ | |
3048 | struct node_hstate { | |
3049 | struct kobject *hugepages_kobj; | |
3050 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3051 | }; | |
b4e289a6 | 3052 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3053 | |
3054 | /* | |
10fbcf4c | 3055 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3056 | */ |
3057 | static struct attribute *per_node_hstate_attrs[] = { | |
3058 | &nr_hugepages_attr.attr, | |
3059 | &free_hugepages_attr.attr, | |
3060 | &surplus_hugepages_attr.attr, | |
3061 | NULL, | |
3062 | }; | |
3063 | ||
67e5ed96 | 3064 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3065 | .attrs = per_node_hstate_attrs, |
3066 | }; | |
3067 | ||
3068 | /* | |
10fbcf4c | 3069 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3070 | * Returns node id via non-NULL nidp. |
3071 | */ | |
3072 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3073 | { | |
3074 | int nid; | |
3075 | ||
3076 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3077 | struct node_hstate *nhs = &node_hstates[nid]; | |
3078 | int i; | |
3079 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3080 | if (nhs->hstate_kobjs[i] == kobj) { | |
3081 | if (nidp) | |
3082 | *nidp = nid; | |
3083 | return &hstates[i]; | |
3084 | } | |
3085 | } | |
3086 | ||
3087 | BUG(); | |
3088 | return NULL; | |
3089 | } | |
3090 | ||
3091 | /* | |
10fbcf4c | 3092 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3093 | * No-op if no hstate attributes attached. |
3094 | */ | |
3cd8b44f | 3095 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3096 | { |
3097 | struct hstate *h; | |
10fbcf4c | 3098 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3099 | |
3100 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3101 | return; /* no hstate attributes */ |
9a305230 | 3102 | |
972dc4de AK |
3103 | for_each_hstate(h) { |
3104 | int idx = hstate_index(h); | |
3105 | if (nhs->hstate_kobjs[idx]) { | |
3106 | kobject_put(nhs->hstate_kobjs[idx]); | |
3107 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3108 | } |
972dc4de | 3109 | } |
9a305230 LS |
3110 | |
3111 | kobject_put(nhs->hugepages_kobj); | |
3112 | nhs->hugepages_kobj = NULL; | |
3113 | } | |
3114 | ||
9a305230 LS |
3115 | |
3116 | /* | |
10fbcf4c | 3117 | * Register hstate attributes for a single node device. |
9a305230 LS |
3118 | * No-op if attributes already registered. |
3119 | */ | |
3cd8b44f | 3120 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3121 | { |
3122 | struct hstate *h; | |
10fbcf4c | 3123 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3124 | int err; |
3125 | ||
3126 | if (nhs->hugepages_kobj) | |
3127 | return; /* already allocated */ | |
3128 | ||
3129 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3130 | &node->dev.kobj); |
9a305230 LS |
3131 | if (!nhs->hugepages_kobj) |
3132 | return; | |
3133 | ||
3134 | for_each_hstate(h) { | |
3135 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3136 | nhs->hstate_kobjs, | |
3137 | &per_node_hstate_attr_group); | |
3138 | if (err) { | |
282f4214 | 3139 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3140 | h->name, node->dev.id); |
9a305230 LS |
3141 | hugetlb_unregister_node(node); |
3142 | break; | |
3143 | } | |
3144 | } | |
3145 | } | |
3146 | ||
3147 | /* | |
9b5e5d0f | 3148 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3149 | * devices of nodes that have memory. All on-line nodes should have |
3150 | * registered their associated device by this time. | |
9a305230 | 3151 | */ |
7d9ca000 | 3152 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3153 | { |
3154 | int nid; | |
3155 | ||
8cebfcd0 | 3156 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3157 | struct node *node = node_devices[nid]; |
10fbcf4c | 3158 | if (node->dev.id == nid) |
9a305230 LS |
3159 | hugetlb_register_node(node); |
3160 | } | |
3161 | ||
3162 | /* | |
10fbcf4c | 3163 | * Let the node device driver know we're here so it can |
9a305230 LS |
3164 | * [un]register hstate attributes on node hotplug. |
3165 | */ | |
3166 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3167 | hugetlb_unregister_node); | |
3168 | } | |
3169 | #else /* !CONFIG_NUMA */ | |
3170 | ||
3171 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3172 | { | |
3173 | BUG(); | |
3174 | if (nidp) | |
3175 | *nidp = -1; | |
3176 | return NULL; | |
3177 | } | |
3178 | ||
9a305230 LS |
3179 | static void hugetlb_register_all_nodes(void) { } |
3180 | ||
3181 | #endif | |
3182 | ||
a3437870 NA |
3183 | static int __init hugetlb_init(void) |
3184 | { | |
8382d914 DB |
3185 | int i; |
3186 | ||
c2833a5b MK |
3187 | if (!hugepages_supported()) { |
3188 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
3189 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 3190 | return 0; |
c2833a5b | 3191 | } |
a3437870 | 3192 | |
282f4214 MK |
3193 | /* |
3194 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
3195 | * architectures depend on setup being done here. | |
3196 | */ | |
3197 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
3198 | if (!parsed_default_hugepagesz) { | |
3199 | /* | |
3200 | * If we did not parse a default huge page size, set | |
3201 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
3202 | * number of huge pages for this default size was implicitly | |
3203 | * specified, set that here as well. | |
3204 | * Note that the implicit setting will overwrite an explicit | |
3205 | * setting. A warning will be printed in this case. | |
3206 | */ | |
3207 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
3208 | if (default_hstate_max_huge_pages) { | |
3209 | if (default_hstate.max_huge_pages) { | |
3210 | char buf[32]; | |
3211 | ||
3212 | string_get_size(huge_page_size(&default_hstate), | |
3213 | 1, STRING_UNITS_2, buf, 32); | |
3214 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
3215 | default_hstate.max_huge_pages, buf); | |
3216 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
3217 | default_hstate_max_huge_pages); | |
3218 | } | |
3219 | default_hstate.max_huge_pages = | |
3220 | default_hstate_max_huge_pages; | |
d715cf80 | 3221 | } |
f8b74815 | 3222 | } |
a3437870 | 3223 | |
cf11e85f | 3224 | hugetlb_cma_check(); |
a3437870 | 3225 | hugetlb_init_hstates(); |
aa888a74 | 3226 | gather_bootmem_prealloc(); |
a3437870 NA |
3227 | report_hugepages(); |
3228 | ||
3229 | hugetlb_sysfs_init(); | |
9a305230 | 3230 | hugetlb_register_all_nodes(); |
7179e7bf | 3231 | hugetlb_cgroup_file_init(); |
9a305230 | 3232 | |
8382d914 DB |
3233 | #ifdef CONFIG_SMP |
3234 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3235 | #else | |
3236 | num_fault_mutexes = 1; | |
3237 | #endif | |
c672c7f2 | 3238 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3239 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3240 | GFP_KERNEL); | |
c672c7f2 | 3241 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3242 | |
3243 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3244 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3245 | return 0; |
3246 | } | |
3e89e1c5 | 3247 | subsys_initcall(hugetlb_init); |
a3437870 | 3248 | |
ae94da89 MK |
3249 | /* Overwritten by architectures with more huge page sizes */ |
3250 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 3251 | { |
ae94da89 | 3252 | return size == HPAGE_SIZE; |
9fee021d VT |
3253 | } |
3254 | ||
d00181b9 | 3255 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3256 | { |
3257 | struct hstate *h; | |
8faa8b07 AK |
3258 | unsigned long i; |
3259 | ||
a3437870 | 3260 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
3261 | return; |
3262 | } | |
47d38344 | 3263 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3264 | BUG_ON(order == 0); |
47d38344 | 3265 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
3266 | h->order = order; |
3267 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
3268 | for (i = 0; i < MAX_NUMNODES; ++i) |
3269 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3270 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3271 | h->next_nid_to_alloc = first_memory_node; |
3272 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3273 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3274 | huge_page_size(h)/1024); | |
8faa8b07 | 3275 | |
a3437870 NA |
3276 | parsed_hstate = h; |
3277 | } | |
3278 | ||
282f4214 MK |
3279 | /* |
3280 | * hugepages command line processing | |
3281 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
3282 | * specification. If not, ignore the hugepages value. hugepages can also | |
3283 | * be the first huge page command line option in which case it implicitly | |
3284 | * specifies the number of huge pages for the default size. | |
3285 | */ | |
3286 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
3287 | { |
3288 | unsigned long *mhp; | |
8faa8b07 | 3289 | static unsigned long *last_mhp; |
a3437870 | 3290 | |
9fee021d | 3291 | if (!parsed_valid_hugepagesz) { |
282f4214 | 3292 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 3293 | parsed_valid_hugepagesz = true; |
282f4214 | 3294 | return 0; |
9fee021d | 3295 | } |
282f4214 | 3296 | |
a3437870 | 3297 | /* |
282f4214 MK |
3298 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
3299 | * yet, so this hugepages= parameter goes to the "default hstate". | |
3300 | * Otherwise, it goes with the previously parsed hugepagesz or | |
3301 | * default_hugepagesz. | |
a3437870 | 3302 | */ |
9fee021d | 3303 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3304 | mhp = &default_hstate_max_huge_pages; |
3305 | else | |
3306 | mhp = &parsed_hstate->max_huge_pages; | |
3307 | ||
8faa8b07 | 3308 | if (mhp == last_mhp) { |
282f4214 MK |
3309 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
3310 | return 0; | |
8faa8b07 AK |
3311 | } |
3312 | ||
a3437870 NA |
3313 | if (sscanf(s, "%lu", mhp) <= 0) |
3314 | *mhp = 0; | |
3315 | ||
8faa8b07 AK |
3316 | /* |
3317 | * Global state is always initialized later in hugetlb_init. | |
3318 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
3319 | * use the bootmem allocator. | |
3320 | */ | |
47d38344 | 3321 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
3322 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3323 | ||
3324 | last_mhp = mhp; | |
3325 | ||
a3437870 NA |
3326 | return 1; |
3327 | } | |
282f4214 | 3328 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 3329 | |
282f4214 MK |
3330 | /* |
3331 | * hugepagesz command line processing | |
3332 | * A specific huge page size can only be specified once with hugepagesz. | |
3333 | * hugepagesz is followed by hugepages on the command line. The global | |
3334 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
3335 | * hugepagesz argument was valid. | |
3336 | */ | |
359f2544 | 3337 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 3338 | { |
359f2544 | 3339 | unsigned long size; |
282f4214 MK |
3340 | struct hstate *h; |
3341 | ||
3342 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
3343 | size = (unsigned long)memparse(s, NULL); |
3344 | ||
3345 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3346 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
359f2544 MK |
3347 | return 0; |
3348 | } | |
3349 | ||
282f4214 MK |
3350 | h = size_to_hstate(size); |
3351 | if (h) { | |
3352 | /* | |
3353 | * hstate for this size already exists. This is normally | |
3354 | * an error, but is allowed if the existing hstate is the | |
3355 | * default hstate. More specifically, it is only allowed if | |
3356 | * the number of huge pages for the default hstate was not | |
3357 | * previously specified. | |
3358 | */ | |
3359 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
3360 | default_hstate.max_huge_pages) { | |
3361 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
3362 | return 0; | |
3363 | } | |
3364 | ||
3365 | /* | |
3366 | * No need to call hugetlb_add_hstate() as hstate already | |
3367 | * exists. But, do set parsed_hstate so that a following | |
3368 | * hugepages= parameter will be applied to this hstate. | |
3369 | */ | |
3370 | parsed_hstate = h; | |
3371 | parsed_valid_hugepagesz = true; | |
3372 | return 1; | |
38237830 MK |
3373 | } |
3374 | ||
359f2544 | 3375 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 3376 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
3377 | return 1; |
3378 | } | |
359f2544 MK |
3379 | __setup("hugepagesz=", hugepagesz_setup); |
3380 | ||
282f4214 MK |
3381 | /* |
3382 | * default_hugepagesz command line input | |
3383 | * Only one instance of default_hugepagesz allowed on command line. | |
3384 | */ | |
ae94da89 | 3385 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 3386 | { |
ae94da89 MK |
3387 | unsigned long size; |
3388 | ||
282f4214 | 3389 | parsed_valid_hugepagesz = false; |
282f4214 MK |
3390 | if (parsed_default_hugepagesz) { |
3391 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
3392 | return 0; | |
3393 | } | |
3394 | ||
ae94da89 MK |
3395 | size = (unsigned long)memparse(s, NULL); |
3396 | ||
3397 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3398 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
ae94da89 MK |
3399 | return 0; |
3400 | } | |
3401 | ||
282f4214 MK |
3402 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
3403 | parsed_valid_hugepagesz = true; | |
3404 | parsed_default_hugepagesz = true; | |
3405 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
3406 | ||
3407 | /* | |
3408 | * The number of default huge pages (for this size) could have been | |
3409 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
3410 | * then default_hstate_max_huge_pages is set. If the default huge | |
3411 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
3412 | * allocated here from bootmem allocator. | |
3413 | */ | |
3414 | if (default_hstate_max_huge_pages) { | |
3415 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3416 | if (hstate_is_gigantic(&default_hstate)) | |
3417 | hugetlb_hstate_alloc_pages(&default_hstate); | |
3418 | default_hstate_max_huge_pages = 0; | |
3419 | } | |
3420 | ||
e11bfbfc NP |
3421 | return 1; |
3422 | } | |
ae94da89 | 3423 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 3424 | |
8ca39e68 | 3425 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
3426 | { |
3427 | int node; | |
3428 | unsigned int nr = 0; | |
8ca39e68 MS |
3429 | nodemask_t *mpol_allowed; |
3430 | unsigned int *array = h->free_huge_pages_node; | |
3431 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
3432 | ||
3433 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 3434 | |
8ca39e68 MS |
3435 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
3436 | if (!mpol_allowed || | |
3437 | (mpol_allowed && node_isset(node, *mpol_allowed))) | |
3438 | nr += array[node]; | |
3439 | } | |
8a213460 NA |
3440 | |
3441 | return nr; | |
3442 | } | |
3443 | ||
3444 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
3445 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
3446 | void *buffer, size_t *length, | |
3447 | loff_t *ppos, unsigned long *out) | |
3448 | { | |
3449 | struct ctl_table dup_table; | |
3450 | ||
3451 | /* | |
3452 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
3453 | * can duplicate the @table and alter the duplicate of it. | |
3454 | */ | |
3455 | dup_table = *table; | |
3456 | dup_table.data = out; | |
3457 | ||
3458 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
3459 | } | |
3460 | ||
06808b08 LS |
3461 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3462 | struct ctl_table *table, int write, | |
32927393 | 3463 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 3464 | { |
e5ff2159 | 3465 | struct hstate *h = &default_hstate; |
238d3c13 | 3466 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3467 | int ret; |
e5ff2159 | 3468 | |
457c1b27 | 3469 | if (!hugepages_supported()) |
86613628 | 3470 | return -EOPNOTSUPP; |
457c1b27 | 3471 | |
17743798 MS |
3472 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3473 | &tmp); | |
08d4a246 MH |
3474 | if (ret) |
3475 | goto out; | |
e5ff2159 | 3476 | |
238d3c13 DR |
3477 | if (write) |
3478 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3479 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3480 | out: |
3481 | return ret; | |
1da177e4 | 3482 | } |
396faf03 | 3483 | |
06808b08 | 3484 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 3485 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3486 | { |
3487 | ||
3488 | return hugetlb_sysctl_handler_common(false, table, write, | |
3489 | buffer, length, ppos); | |
3490 | } | |
3491 | ||
3492 | #ifdef CONFIG_NUMA | |
3493 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 3494 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3495 | { |
3496 | return hugetlb_sysctl_handler_common(true, table, write, | |
3497 | buffer, length, ppos); | |
3498 | } | |
3499 | #endif /* CONFIG_NUMA */ | |
3500 | ||
a3d0c6aa | 3501 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 3502 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 3503 | { |
a5516438 | 3504 | struct hstate *h = &default_hstate; |
e5ff2159 | 3505 | unsigned long tmp; |
08d4a246 | 3506 | int ret; |
e5ff2159 | 3507 | |
457c1b27 | 3508 | if (!hugepages_supported()) |
86613628 | 3509 | return -EOPNOTSUPP; |
457c1b27 | 3510 | |
c033a93c | 3511 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3512 | |
bae7f4ae | 3513 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3514 | return -EINVAL; |
3515 | ||
17743798 MS |
3516 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3517 | &tmp); | |
08d4a246 MH |
3518 | if (ret) |
3519 | goto out; | |
e5ff2159 AK |
3520 | |
3521 | if (write) { | |
3522 | spin_lock(&hugetlb_lock); | |
3523 | h->nr_overcommit_huge_pages = tmp; | |
3524 | spin_unlock(&hugetlb_lock); | |
3525 | } | |
08d4a246 MH |
3526 | out: |
3527 | return ret; | |
a3d0c6aa NA |
3528 | } |
3529 | ||
1da177e4 LT |
3530 | #endif /* CONFIG_SYSCTL */ |
3531 | ||
e1759c21 | 3532 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3533 | { |
fcb2b0c5 RG |
3534 | struct hstate *h; |
3535 | unsigned long total = 0; | |
3536 | ||
457c1b27 NA |
3537 | if (!hugepages_supported()) |
3538 | return; | |
fcb2b0c5 RG |
3539 | |
3540 | for_each_hstate(h) { | |
3541 | unsigned long count = h->nr_huge_pages; | |
3542 | ||
3543 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3544 | ||
3545 | if (h == &default_hstate) | |
3546 | seq_printf(m, | |
3547 | "HugePages_Total: %5lu\n" | |
3548 | "HugePages_Free: %5lu\n" | |
3549 | "HugePages_Rsvd: %5lu\n" | |
3550 | "HugePages_Surp: %5lu\n" | |
3551 | "Hugepagesize: %8lu kB\n", | |
3552 | count, | |
3553 | h->free_huge_pages, | |
3554 | h->resv_huge_pages, | |
3555 | h->surplus_huge_pages, | |
3556 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3557 | } | |
3558 | ||
3559 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3560 | } |
3561 | ||
7981593b | 3562 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 3563 | { |
a5516438 | 3564 | struct hstate *h = &default_hstate; |
7981593b | 3565 | |
457c1b27 NA |
3566 | if (!hugepages_supported()) |
3567 | return 0; | |
7981593b JP |
3568 | |
3569 | return sysfs_emit_at(buf, len, | |
3570 | "Node %d HugePages_Total: %5u\n" | |
3571 | "Node %d HugePages_Free: %5u\n" | |
3572 | "Node %d HugePages_Surp: %5u\n", | |
3573 | nid, h->nr_huge_pages_node[nid], | |
3574 | nid, h->free_huge_pages_node[nid], | |
3575 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3576 | } |
3577 | ||
949f7ec5 DR |
3578 | void hugetlb_show_meminfo(void) |
3579 | { | |
3580 | struct hstate *h; | |
3581 | int nid; | |
3582 | ||
457c1b27 NA |
3583 | if (!hugepages_supported()) |
3584 | return; | |
3585 | ||
949f7ec5 DR |
3586 | for_each_node_state(nid, N_MEMORY) |
3587 | for_each_hstate(h) | |
3588 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3589 | nid, | |
3590 | h->nr_huge_pages_node[nid], | |
3591 | h->free_huge_pages_node[nid], | |
3592 | h->surplus_huge_pages_node[nid], | |
3593 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3594 | } | |
3595 | ||
5d317b2b NH |
3596 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3597 | { | |
3598 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3599 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3600 | } | |
3601 | ||
1da177e4 LT |
3602 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3603 | unsigned long hugetlb_total_pages(void) | |
3604 | { | |
d0028588 WL |
3605 | struct hstate *h; |
3606 | unsigned long nr_total_pages = 0; | |
3607 | ||
3608 | for_each_hstate(h) | |
3609 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3610 | return nr_total_pages; | |
1da177e4 | 3611 | } |
1da177e4 | 3612 | |
a5516438 | 3613 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3614 | { |
3615 | int ret = -ENOMEM; | |
3616 | ||
3617 | spin_lock(&hugetlb_lock); | |
3618 | /* | |
3619 | * When cpuset is configured, it breaks the strict hugetlb page | |
3620 | * reservation as the accounting is done on a global variable. Such | |
3621 | * reservation is completely rubbish in the presence of cpuset because | |
3622 | * the reservation is not checked against page availability for the | |
3623 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3624 | * with lack of free htlb page in cpuset that the task is in. | |
3625 | * Attempt to enforce strict accounting with cpuset is almost | |
3626 | * impossible (or too ugly) because cpuset is too fluid that | |
3627 | * task or memory node can be dynamically moved between cpusets. | |
3628 | * | |
3629 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3630 | * undesirable. However, in order to preserve some of the semantics, | |
3631 | * we fall back to check against current free page availability as | |
3632 | * a best attempt and hopefully to minimize the impact of changing | |
3633 | * semantics that cpuset has. | |
8ca39e68 MS |
3634 | * |
3635 | * Apart from cpuset, we also have memory policy mechanism that | |
3636 | * also determines from which node the kernel will allocate memory | |
3637 | * in a NUMA system. So similar to cpuset, we also should consider | |
3638 | * the memory policy of the current task. Similar to the description | |
3639 | * above. | |
fc1b8a73 MG |
3640 | */ |
3641 | if (delta > 0) { | |
a5516438 | 3642 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3643 | goto out; |
3644 | ||
8ca39e68 | 3645 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 3646 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
3647 | goto out; |
3648 | } | |
3649 | } | |
3650 | ||
3651 | ret = 0; | |
3652 | if (delta < 0) | |
a5516438 | 3653 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3654 | |
3655 | out: | |
3656 | spin_unlock(&hugetlb_lock); | |
3657 | return ret; | |
3658 | } | |
3659 | ||
84afd99b AW |
3660 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3661 | { | |
f522c3ac | 3662 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3663 | |
3664 | /* | |
3665 | * This new VMA should share its siblings reservation map if present. | |
3666 | * The VMA will only ever have a valid reservation map pointer where | |
3667 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3668 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3669 | * after this open call completes. It is therefore safe to take a |
3670 | * new reference here without additional locking. | |
3671 | */ | |
4e35f483 | 3672 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3673 | kref_get(&resv->refs); |
84afd99b AW |
3674 | } |
3675 | ||
a1e78772 MG |
3676 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3677 | { | |
a5516438 | 3678 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3679 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3680 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3681 | unsigned long reserve, start, end; |
1c5ecae3 | 3682 | long gbl_reserve; |
84afd99b | 3683 | |
4e35f483 JK |
3684 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3685 | return; | |
84afd99b | 3686 | |
4e35f483 JK |
3687 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3688 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3689 | |
4e35f483 | 3690 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 3691 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 3692 | if (reserve) { |
1c5ecae3 MK |
3693 | /* |
3694 | * Decrement reserve counts. The global reserve count may be | |
3695 | * adjusted if the subpool has a minimum size. | |
3696 | */ | |
3697 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3698 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3699 | } |
e9fe92ae MA |
3700 | |
3701 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
3702 | } |
3703 | ||
31383c68 DW |
3704 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3705 | { | |
3706 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3707 | return -EINVAL; | |
3708 | return 0; | |
3709 | } | |
3710 | ||
05ea8860 DW |
3711 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3712 | { | |
3713 | struct hstate *hstate = hstate_vma(vma); | |
3714 | ||
3715 | return 1UL << huge_page_shift(hstate); | |
3716 | } | |
3717 | ||
1da177e4 LT |
3718 | /* |
3719 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3720 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3721 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3722 | * this far. | |
3723 | */ | |
b3ec9f33 | 3724 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3725 | { |
3726 | BUG(); | |
d0217ac0 | 3727 | return 0; |
1da177e4 LT |
3728 | } |
3729 | ||
eec3636a JC |
3730 | /* |
3731 | * When a new function is introduced to vm_operations_struct and added | |
3732 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3733 | * This is because under System V memory model, mappings created via | |
3734 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3735 | * their original vm_ops are overwritten with shm_vm_ops. | |
3736 | */ | |
f0f37e2f | 3737 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3738 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3739 | .open = hugetlb_vm_op_open, |
a1e78772 | 3740 | .close = hugetlb_vm_op_close, |
dd3b614f | 3741 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 3742 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3743 | }; |
3744 | ||
1e8f889b DG |
3745 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3746 | int writable) | |
63551ae0 DG |
3747 | { |
3748 | pte_t entry; | |
3749 | ||
1e8f889b | 3750 | if (writable) { |
106c992a GS |
3751 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3752 | vma->vm_page_prot))); | |
63551ae0 | 3753 | } else { |
106c992a GS |
3754 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3755 | vma->vm_page_prot)); | |
63551ae0 DG |
3756 | } |
3757 | entry = pte_mkyoung(entry); | |
3758 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3759 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3760 | |
3761 | return entry; | |
3762 | } | |
3763 | ||
1e8f889b DG |
3764 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3765 | unsigned long address, pte_t *ptep) | |
3766 | { | |
3767 | pte_t entry; | |
3768 | ||
106c992a | 3769 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3770 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3771 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3772 | } |
3773 | ||
d5ed7444 | 3774 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3775 | { |
3776 | swp_entry_t swp; | |
3777 | ||
3778 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3779 | return false; |
4a705fef | 3780 | swp = pte_to_swp_entry(pte); |
d79d176a | 3781 | if (is_migration_entry(swp)) |
d5ed7444 | 3782 | return true; |
4a705fef | 3783 | else |
d5ed7444 | 3784 | return false; |
4a705fef NH |
3785 | } |
3786 | ||
3e5c3600 | 3787 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
3788 | { |
3789 | swp_entry_t swp; | |
3790 | ||
3791 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 3792 | return false; |
4a705fef | 3793 | swp = pte_to_swp_entry(pte); |
d79d176a | 3794 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 3795 | return true; |
4a705fef | 3796 | else |
3e5c3600 | 3797 | return false; |
4a705fef | 3798 | } |
1e8f889b | 3799 | |
63551ae0 DG |
3800 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3801 | struct vm_area_struct *vma) | |
3802 | { | |
5e41540c | 3803 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3804 | struct page *ptepage; |
1c59827d | 3805 | unsigned long addr; |
1e8f889b | 3806 | int cow; |
a5516438 AK |
3807 | struct hstate *h = hstate_vma(vma); |
3808 | unsigned long sz = huge_page_size(h); | |
c0d0381a | 3809 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 3810 | struct mmu_notifier_range range; |
e8569dd2 | 3811 | int ret = 0; |
1e8f889b DG |
3812 | |
3813 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3814 | |
ac46d4f3 | 3815 | if (cow) { |
7269f999 | 3816 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3817 | vma->vm_start, |
ac46d4f3 JG |
3818 | vma->vm_end); |
3819 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
3820 | } else { |
3821 | /* | |
3822 | * For shared mappings i_mmap_rwsem must be held to call | |
3823 | * huge_pte_alloc, otherwise the returned ptep could go | |
3824 | * away if part of a shared pmd and another thread calls | |
3825 | * huge_pmd_unshare. | |
3826 | */ | |
3827 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 3828 | } |
e8569dd2 | 3829 | |
a5516438 | 3830 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3831 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3832 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3833 | if (!src_pte) |
3834 | continue; | |
a5516438 | 3835 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3836 | if (!dst_pte) { |
3837 | ret = -ENOMEM; | |
3838 | break; | |
3839 | } | |
c5c99429 | 3840 | |
5e41540c MK |
3841 | /* |
3842 | * If the pagetables are shared don't copy or take references. | |
3843 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3844 | * | |
3845 | * However, src could have 'unshared' and dst shares with | |
3846 | * another vma. If dst_pte !none, this implies sharing. | |
3847 | * Check here before taking page table lock, and once again | |
3848 | * after taking the lock below. | |
3849 | */ | |
3850 | dst_entry = huge_ptep_get(dst_pte); | |
3851 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3852 | continue; |
3853 | ||
cb900f41 KS |
3854 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3855 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3856 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3857 | entry = huge_ptep_get(src_pte); |
5e41540c MK |
3858 | dst_entry = huge_ptep_get(dst_pte); |
3859 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3860 | /* | |
3861 | * Skip if src entry none. Also, skip in the | |
3862 | * unlikely case dst entry !none as this implies | |
3863 | * sharing with another vma. | |
3864 | */ | |
4a705fef NH |
3865 | ; |
3866 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3867 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3868 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3869 | ||
3870 | if (is_write_migration_entry(swp_entry) && cow) { | |
3871 | /* | |
3872 | * COW mappings require pages in both | |
3873 | * parent and child to be set to read. | |
3874 | */ | |
3875 | make_migration_entry_read(&swp_entry); | |
3876 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3877 | set_huge_swap_pte_at(src, addr, src_pte, |
3878 | entry, sz); | |
4a705fef | 3879 | } |
e5251fd4 | 3880 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3881 | } else { |
34ee645e | 3882 | if (cow) { |
0f10851e JG |
3883 | /* |
3884 | * No need to notify as we are downgrading page | |
3885 | * table protection not changing it to point | |
3886 | * to a new page. | |
3887 | * | |
ad56b738 | 3888 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3889 | */ |
7f2e9525 | 3890 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3891 | } |
0253d634 | 3892 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3893 | ptepage = pte_page(entry); |
3894 | get_page(ptepage); | |
53f9263b | 3895 | page_dup_rmap(ptepage, true); |
1c59827d | 3896 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3897 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3898 | } |
cb900f41 KS |
3899 | spin_unlock(src_ptl); |
3900 | spin_unlock(dst_ptl); | |
63551ae0 | 3901 | } |
63551ae0 | 3902 | |
e8569dd2 | 3903 | if (cow) |
ac46d4f3 | 3904 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
3905 | else |
3906 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
3907 | |
3908 | return ret; | |
63551ae0 DG |
3909 | } |
3910 | ||
24669e58 AK |
3911 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3912 | unsigned long start, unsigned long end, | |
3913 | struct page *ref_page) | |
63551ae0 DG |
3914 | { |
3915 | struct mm_struct *mm = vma->vm_mm; | |
3916 | unsigned long address; | |
c7546f8f | 3917 | pte_t *ptep; |
63551ae0 | 3918 | pte_t pte; |
cb900f41 | 3919 | spinlock_t *ptl; |
63551ae0 | 3920 | struct page *page; |
a5516438 AK |
3921 | struct hstate *h = hstate_vma(vma); |
3922 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3923 | struct mmu_notifier_range range; |
a5516438 | 3924 | |
63551ae0 | 3925 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3926 | BUG_ON(start & ~huge_page_mask(h)); |
3927 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3928 | |
07e32661 AK |
3929 | /* |
3930 | * This is a hugetlb vma, all the pte entries should point | |
3931 | * to huge page. | |
3932 | */ | |
ed6a7935 | 3933 | tlb_change_page_size(tlb, sz); |
24669e58 | 3934 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3935 | |
3936 | /* | |
3937 | * If sharing possible, alert mmu notifiers of worst case. | |
3938 | */ | |
6f4f13e8 JG |
3939 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3940 | end); | |
ac46d4f3 JG |
3941 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3942 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3943 | address = start; |
569f48b8 | 3944 | for (; address < end; address += sz) { |
7868a208 | 3945 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3946 | if (!ptep) |
c7546f8f DG |
3947 | continue; |
3948 | ||
cb900f41 | 3949 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 3950 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
31d49da5 | 3951 | spin_unlock(ptl); |
dff11abe MK |
3952 | /* |
3953 | * We just unmapped a page of PMDs by clearing a PUD. | |
3954 | * The caller's TLB flush range should cover this area. | |
3955 | */ | |
31d49da5 AK |
3956 | continue; |
3957 | } | |
39dde65c | 3958 | |
6629326b | 3959 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3960 | if (huge_pte_none(pte)) { |
3961 | spin_unlock(ptl); | |
3962 | continue; | |
3963 | } | |
6629326b HD |
3964 | |
3965 | /* | |
9fbc1f63 NH |
3966 | * Migrating hugepage or HWPoisoned hugepage is already |
3967 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3968 | */ |
9fbc1f63 | 3969 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3970 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3971 | spin_unlock(ptl); |
3972 | continue; | |
8c4894c6 | 3973 | } |
6629326b HD |
3974 | |
3975 | page = pte_page(pte); | |
04f2cbe3 MG |
3976 | /* |
3977 | * If a reference page is supplied, it is because a specific | |
3978 | * page is being unmapped, not a range. Ensure the page we | |
3979 | * are about to unmap is the actual page of interest. | |
3980 | */ | |
3981 | if (ref_page) { | |
31d49da5 AK |
3982 | if (page != ref_page) { |
3983 | spin_unlock(ptl); | |
3984 | continue; | |
3985 | } | |
04f2cbe3 MG |
3986 | /* |
3987 | * Mark the VMA as having unmapped its page so that | |
3988 | * future faults in this VMA will fail rather than | |
3989 | * looking like data was lost | |
3990 | */ | |
3991 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3992 | } | |
3993 | ||
c7546f8f | 3994 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3995 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3996 | if (huge_pte_dirty(pte)) |
6649a386 | 3997 | set_page_dirty(page); |
9e81130b | 3998 | |
5d317b2b | 3999 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 4000 | page_remove_rmap(page, true); |
31d49da5 | 4001 | |
cb900f41 | 4002 | spin_unlock(ptl); |
e77b0852 | 4003 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
4004 | /* |
4005 | * Bail out after unmapping reference page if supplied | |
4006 | */ | |
4007 | if (ref_page) | |
4008 | break; | |
fe1668ae | 4009 | } |
ac46d4f3 | 4010 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 4011 | tlb_end_vma(tlb, vma); |
1da177e4 | 4012 | } |
63551ae0 | 4013 | |
d833352a MG |
4014 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
4015 | struct vm_area_struct *vma, unsigned long start, | |
4016 | unsigned long end, struct page *ref_page) | |
4017 | { | |
4018 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
4019 | ||
4020 | /* | |
4021 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
4022 | * test will fail on a vma being torn down, and not grab a page table | |
4023 | * on its way out. We're lucky that the flag has such an appropriate | |
4024 | * name, and can in fact be safely cleared here. We could clear it | |
4025 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 4026 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 4027 | * because in the context this is called, the VMA is about to be |
c8c06efa | 4028 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
4029 | */ |
4030 | vma->vm_flags &= ~VM_MAYSHARE; | |
4031 | } | |
4032 | ||
502717f4 | 4033 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 4034 | unsigned long end, struct page *ref_page) |
502717f4 | 4035 | { |
24669e58 AK |
4036 | struct mm_struct *mm; |
4037 | struct mmu_gather tlb; | |
dff11abe MK |
4038 | unsigned long tlb_start = start; |
4039 | unsigned long tlb_end = end; | |
4040 | ||
4041 | /* | |
4042 | * If shared PMDs were possibly used within this vma range, adjust | |
4043 | * start/end for worst case tlb flushing. | |
4044 | * Note that we can not be sure if PMDs are shared until we try to | |
4045 | * unmap pages. However, we want to make sure TLB flushing covers | |
4046 | * the largest possible range. | |
4047 | */ | |
4048 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
24669e58 AK |
4049 | |
4050 | mm = vma->vm_mm; | |
4051 | ||
dff11abe | 4052 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); |
24669e58 | 4053 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
dff11abe | 4054 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); |
502717f4 KC |
4055 | } |
4056 | ||
04f2cbe3 MG |
4057 | /* |
4058 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
4059 | * mappping it owns the reserve page for. The intention is to unmap the page | |
4060 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
4061 | * same region. | |
4062 | */ | |
2f4612af DB |
4063 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
4064 | struct page *page, unsigned long address) | |
04f2cbe3 | 4065 | { |
7526674d | 4066 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
4067 | struct vm_area_struct *iter_vma; |
4068 | struct address_space *mapping; | |
04f2cbe3 MG |
4069 | pgoff_t pgoff; |
4070 | ||
4071 | /* | |
4072 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
4073 | * from page cache lookup which is in HPAGE_SIZE units. | |
4074 | */ | |
7526674d | 4075 | address = address & huge_page_mask(h); |
36e4f20a MH |
4076 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
4077 | vma->vm_pgoff; | |
93c76a3d | 4078 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 4079 | |
4eb2b1dc MG |
4080 | /* |
4081 | * Take the mapping lock for the duration of the table walk. As | |
4082 | * this mapping should be shared between all the VMAs, | |
4083 | * __unmap_hugepage_range() is called as the lock is already held | |
4084 | */ | |
83cde9e8 | 4085 | i_mmap_lock_write(mapping); |
6b2dbba8 | 4086 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
4087 | /* Do not unmap the current VMA */ |
4088 | if (iter_vma == vma) | |
4089 | continue; | |
4090 | ||
2f84a899 MG |
4091 | /* |
4092 | * Shared VMAs have their own reserves and do not affect | |
4093 | * MAP_PRIVATE accounting but it is possible that a shared | |
4094 | * VMA is using the same page so check and skip such VMAs. | |
4095 | */ | |
4096 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
4097 | continue; | |
4098 | ||
04f2cbe3 MG |
4099 | /* |
4100 | * Unmap the page from other VMAs without their own reserves. | |
4101 | * They get marked to be SIGKILLed if they fault in these | |
4102 | * areas. This is because a future no-page fault on this VMA | |
4103 | * could insert a zeroed page instead of the data existing | |
4104 | * from the time of fork. This would look like data corruption | |
4105 | */ | |
4106 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
4107 | unmap_hugepage_range(iter_vma, address, |
4108 | address + huge_page_size(h), page); | |
04f2cbe3 | 4109 | } |
83cde9e8 | 4110 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
4111 | } |
4112 | ||
0fe6e20b NH |
4113 | /* |
4114 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
4115 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
4116 | * cannot race with other handlers or page migration. | |
4117 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 4118 | */ |
2b740303 | 4119 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 4120 | unsigned long address, pte_t *ptep, |
3999f52e | 4121 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 4122 | { |
3999f52e | 4123 | pte_t pte; |
a5516438 | 4124 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 4125 | struct page *old_page, *new_page; |
2b740303 SJ |
4126 | int outside_reserve = 0; |
4127 | vm_fault_t ret = 0; | |
974e6d66 | 4128 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 4129 | struct mmu_notifier_range range; |
1e8f889b | 4130 | |
3999f52e | 4131 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
4132 | old_page = pte_page(pte); |
4133 | ||
04f2cbe3 | 4134 | retry_avoidcopy: |
1e8f889b DG |
4135 | /* If no-one else is actually using this page, avoid the copy |
4136 | * and just make the page writable */ | |
37a2140d | 4137 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 4138 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 4139 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 4140 | return 0; |
1e8f889b DG |
4141 | } |
4142 | ||
04f2cbe3 MG |
4143 | /* |
4144 | * If the process that created a MAP_PRIVATE mapping is about to | |
4145 | * perform a COW due to a shared page count, attempt to satisfy | |
4146 | * the allocation without using the existing reserves. The pagecache | |
4147 | * page is used to determine if the reserve at this address was | |
4148 | * consumed or not. If reserves were used, a partial faulted mapping | |
4149 | * at the time of fork() could consume its reserves on COW instead | |
4150 | * of the full address range. | |
4151 | */ | |
5944d011 | 4152 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
4153 | old_page != pagecache_page) |
4154 | outside_reserve = 1; | |
4155 | ||
09cbfeaf | 4156 | get_page(old_page); |
b76c8cfb | 4157 | |
ad4404a2 DB |
4158 | /* |
4159 | * Drop page table lock as buddy allocator may be called. It will | |
4160 | * be acquired again before returning to the caller, as expected. | |
4161 | */ | |
cb900f41 | 4162 | spin_unlock(ptl); |
5b7a1d40 | 4163 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 4164 | |
2fc39cec | 4165 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
4166 | /* |
4167 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4168 | * it is due to references held by a child and an insufficient | |
4169 | * huge page pool. To guarantee the original mappers | |
4170 | * reliability, unmap the page from child processes. The child | |
4171 | * may get SIGKILLed if it later faults. | |
4172 | */ | |
4173 | if (outside_reserve) { | |
e7dd91c4 MK |
4174 | struct address_space *mapping = vma->vm_file->f_mapping; |
4175 | pgoff_t idx; | |
4176 | u32 hash; | |
4177 | ||
09cbfeaf | 4178 | put_page(old_page); |
04f2cbe3 | 4179 | BUG_ON(huge_pte_none(pte)); |
e7dd91c4 MK |
4180 | /* |
4181 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
4182 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
4183 | * in write mode. Dropping i_mmap_rwsem in read mode | |
4184 | * here is OK as COW mappings do not interact with | |
4185 | * PMD sharing. | |
4186 | * | |
4187 | * Reacquire both after unmap operation. | |
4188 | */ | |
4189 | idx = vma_hugecache_offset(h, vma, haddr); | |
4190 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4191 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4192 | i_mmap_unlock_read(mapping); | |
4193 | ||
5b7a1d40 | 4194 | unmap_ref_private(mm, vma, old_page, haddr); |
e7dd91c4 MK |
4195 | |
4196 | i_mmap_lock_read(mapping); | |
4197 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
2f4612af | 4198 | spin_lock(ptl); |
5b7a1d40 | 4199 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
4200 | if (likely(ptep && |
4201 | pte_same(huge_ptep_get(ptep), pte))) | |
4202 | goto retry_avoidcopy; | |
4203 | /* | |
4204 | * race occurs while re-acquiring page table | |
4205 | * lock, and our job is done. | |
4206 | */ | |
4207 | return 0; | |
04f2cbe3 MG |
4208 | } |
4209 | ||
2b740303 | 4210 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 4211 | goto out_release_old; |
1e8f889b DG |
4212 | } |
4213 | ||
0fe6e20b NH |
4214 | /* |
4215 | * When the original hugepage is shared one, it does not have | |
4216 | * anon_vma prepared. | |
4217 | */ | |
44e2aa93 | 4218 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
4219 | ret = VM_FAULT_OOM; |
4220 | goto out_release_all; | |
44e2aa93 | 4221 | } |
0fe6e20b | 4222 | |
974e6d66 | 4223 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 4224 | pages_per_huge_page(h)); |
0ed361de | 4225 | __SetPageUptodate(new_page); |
1e8f889b | 4226 | |
7269f999 | 4227 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 4228 | haddr + huge_page_size(h)); |
ac46d4f3 | 4229 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 4230 | |
b76c8cfb | 4231 | /* |
cb900f41 | 4232 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
4233 | * before the page tables are altered |
4234 | */ | |
cb900f41 | 4235 | spin_lock(ptl); |
5b7a1d40 | 4236 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 4237 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
4238 | ClearPagePrivate(new_page); |
4239 | ||
1e8f889b | 4240 | /* Break COW */ |
5b7a1d40 | 4241 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 4242 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 4243 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 4244 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 4245 | page_remove_rmap(old_page, true); |
5b7a1d40 | 4246 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
cb6acd01 | 4247 | set_page_huge_active(new_page); |
1e8f889b DG |
4248 | /* Make the old page be freed below */ |
4249 | new_page = old_page; | |
4250 | } | |
cb900f41 | 4251 | spin_unlock(ptl); |
ac46d4f3 | 4252 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 4253 | out_release_all: |
5b7a1d40 | 4254 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 4255 | put_page(new_page); |
ad4404a2 | 4256 | out_release_old: |
09cbfeaf | 4257 | put_page(old_page); |
8312034f | 4258 | |
ad4404a2 DB |
4259 | spin_lock(ptl); /* Caller expects lock to be held */ |
4260 | return ret; | |
1e8f889b DG |
4261 | } |
4262 | ||
04f2cbe3 | 4263 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
4264 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
4265 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
4266 | { |
4267 | struct address_space *mapping; | |
e7c4b0bf | 4268 | pgoff_t idx; |
04f2cbe3 MG |
4269 | |
4270 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 4271 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
4272 | |
4273 | return find_lock_page(mapping, idx); | |
4274 | } | |
4275 | ||
3ae77f43 HD |
4276 | /* |
4277 | * Return whether there is a pagecache page to back given address within VMA. | |
4278 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4279 | */ | |
4280 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
4281 | struct vm_area_struct *vma, unsigned long address) |
4282 | { | |
4283 | struct address_space *mapping; | |
4284 | pgoff_t idx; | |
4285 | struct page *page; | |
4286 | ||
4287 | mapping = vma->vm_file->f_mapping; | |
4288 | idx = vma_hugecache_offset(h, vma, address); | |
4289 | ||
4290 | page = find_get_page(mapping, idx); | |
4291 | if (page) | |
4292 | put_page(page); | |
4293 | return page != NULL; | |
4294 | } | |
4295 | ||
ab76ad54 MK |
4296 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
4297 | pgoff_t idx) | |
4298 | { | |
4299 | struct inode *inode = mapping->host; | |
4300 | struct hstate *h = hstate_inode(inode); | |
4301 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4302 | ||
4303 | if (err) | |
4304 | return err; | |
4305 | ClearPagePrivate(page); | |
4306 | ||
22146c3c MK |
4307 | /* |
4308 | * set page dirty so that it will not be removed from cache/file | |
4309 | * by non-hugetlbfs specific code paths. | |
4310 | */ | |
4311 | set_page_dirty(page); | |
4312 | ||
ab76ad54 MK |
4313 | spin_lock(&inode->i_lock); |
4314 | inode->i_blocks += blocks_per_huge_page(h); | |
4315 | spin_unlock(&inode->i_lock); | |
4316 | return 0; | |
4317 | } | |
4318 | ||
2b740303 SJ |
4319 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
4320 | struct vm_area_struct *vma, | |
4321 | struct address_space *mapping, pgoff_t idx, | |
4322 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 4323 | { |
a5516438 | 4324 | struct hstate *h = hstate_vma(vma); |
2b740303 | 4325 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 4326 | int anon_rmap = 0; |
4c887265 | 4327 | unsigned long size; |
4c887265 | 4328 | struct page *page; |
1e8f889b | 4329 | pte_t new_pte; |
cb900f41 | 4330 | spinlock_t *ptl; |
285b8dca | 4331 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 4332 | bool new_page = false; |
4c887265 | 4333 | |
04f2cbe3 MG |
4334 | /* |
4335 | * Currently, we are forced to kill the process in the event the | |
4336 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4337 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4338 | */ |
4339 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4340 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4341 | current->pid); |
04f2cbe3 MG |
4342 | return ret; |
4343 | } | |
4344 | ||
4c887265 | 4345 | /* |
87bf91d3 MK |
4346 | * We can not race with truncation due to holding i_mmap_rwsem. |
4347 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4348 | * once for faults beyond end of file. | |
4c887265 | 4349 | */ |
87bf91d3 MK |
4350 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4351 | if (idx >= size) | |
4352 | goto out; | |
4353 | ||
6bda666a CL |
4354 | retry: |
4355 | page = find_lock_page(mapping, idx); | |
4356 | if (!page) { | |
1a1aad8a MK |
4357 | /* |
4358 | * Check for page in userfault range | |
4359 | */ | |
4360 | if (userfaultfd_missing(vma)) { | |
4361 | u32 hash; | |
4362 | struct vm_fault vmf = { | |
4363 | .vma = vma, | |
285b8dca | 4364 | .address = haddr, |
1a1aad8a MK |
4365 | .flags = flags, |
4366 | /* | |
4367 | * Hard to debug if it ends up being | |
4368 | * used by a callee that assumes | |
4369 | * something about the other | |
4370 | * uninitialized fields... same as in | |
4371 | * memory.c | |
4372 | */ | |
4373 | }; | |
4374 | ||
4375 | /* | |
c0d0381a MK |
4376 | * hugetlb_fault_mutex and i_mmap_rwsem must be |
4377 | * dropped before handling userfault. Reacquire | |
4378 | * after handling fault to make calling code simpler. | |
1a1aad8a | 4379 | */ |
188b04a7 | 4380 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
1a1aad8a | 4381 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4382 | i_mmap_unlock_read(mapping); |
1a1aad8a | 4383 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); |
c0d0381a | 4384 | i_mmap_lock_read(mapping); |
1a1aad8a MK |
4385 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
4386 | goto out; | |
4387 | } | |
4388 | ||
285b8dca | 4389 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4390 | if (IS_ERR(page)) { |
4643d67e MK |
4391 | /* |
4392 | * Returning error will result in faulting task being | |
4393 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4394 | * tasks from racing to fault in the same page which | |
4395 | * could result in false unable to allocate errors. | |
4396 | * Page migration does not take the fault mutex, but | |
4397 | * does a clear then write of pte's under page table | |
4398 | * lock. Page fault code could race with migration, | |
4399 | * notice the clear pte and try to allocate a page | |
4400 | * here. Before returning error, get ptl and make | |
4401 | * sure there really is no pte entry. | |
4402 | */ | |
4403 | ptl = huge_pte_lock(h, mm, ptep); | |
4404 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
4405 | ret = 0; | |
4406 | spin_unlock(ptl); | |
4407 | goto out; | |
4408 | } | |
4409 | spin_unlock(ptl); | |
2b740303 | 4410 | ret = vmf_error(PTR_ERR(page)); |
6bda666a CL |
4411 | goto out; |
4412 | } | |
47ad8475 | 4413 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4414 | __SetPageUptodate(page); |
cb6acd01 | 4415 | new_page = true; |
ac9b9c66 | 4416 | |
f83a275d | 4417 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4418 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4419 | if (err) { |
4420 | put_page(page); | |
6bda666a CL |
4421 | if (err == -EEXIST) |
4422 | goto retry; | |
4423 | goto out; | |
4424 | } | |
23be7468 | 4425 | } else { |
6bda666a | 4426 | lock_page(page); |
0fe6e20b NH |
4427 | if (unlikely(anon_vma_prepare(vma))) { |
4428 | ret = VM_FAULT_OOM; | |
4429 | goto backout_unlocked; | |
4430 | } | |
409eb8c2 | 4431 | anon_rmap = 1; |
23be7468 | 4432 | } |
0fe6e20b | 4433 | } else { |
998b4382 NH |
4434 | /* |
4435 | * If memory error occurs between mmap() and fault, some process | |
4436 | * don't have hwpoisoned swap entry for errored virtual address. | |
4437 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4438 | */ | |
4439 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 4440 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4441 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4442 | goto backout_unlocked; |
4443 | } | |
6bda666a | 4444 | } |
1e8f889b | 4445 | |
57303d80 AW |
4446 | /* |
4447 | * If we are going to COW a private mapping later, we examine the | |
4448 | * pending reservations for this page now. This will ensure that | |
4449 | * any allocations necessary to record that reservation occur outside | |
4450 | * the spinlock. | |
4451 | */ | |
5e911373 | 4452 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4453 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4454 | ret = VM_FAULT_OOM; |
4455 | goto backout_unlocked; | |
4456 | } | |
5e911373 | 4457 | /* Just decrements count, does not deallocate */ |
285b8dca | 4458 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4459 | } |
57303d80 | 4460 | |
8bea8052 | 4461 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 4462 | ret = 0; |
7f2e9525 | 4463 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4464 | goto backout; |
4465 | ||
07443a85 JK |
4466 | if (anon_rmap) { |
4467 | ClearPagePrivate(page); | |
285b8dca | 4468 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4469 | } else |
53f9263b | 4470 | page_dup_rmap(page, true); |
1e8f889b DG |
4471 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4472 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4473 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4474 | |
5d317b2b | 4475 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4476 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4477 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4478 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4479 | } |
4480 | ||
cb900f41 | 4481 | spin_unlock(ptl); |
cb6acd01 MK |
4482 | |
4483 | /* | |
4484 | * Only make newly allocated pages active. Existing pages found | |
4485 | * in the pagecache could be !page_huge_active() if they have been | |
4486 | * isolated for migration. | |
4487 | */ | |
4488 | if (new_page) | |
4489 | set_page_huge_active(page); | |
4490 | ||
4c887265 AL |
4491 | unlock_page(page); |
4492 | out: | |
ac9b9c66 | 4493 | return ret; |
4c887265 AL |
4494 | |
4495 | backout: | |
cb900f41 | 4496 | spin_unlock(ptl); |
2b26736c | 4497 | backout_unlocked: |
4c887265 | 4498 | unlock_page(page); |
285b8dca | 4499 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4500 | put_page(page); |
4501 | goto out; | |
ac9b9c66 HD |
4502 | } |
4503 | ||
8382d914 | 4504 | #ifdef CONFIG_SMP |
188b04a7 | 4505 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4506 | { |
4507 | unsigned long key[2]; | |
4508 | u32 hash; | |
4509 | ||
1b426bac MK |
4510 | key[0] = (unsigned long) mapping; |
4511 | key[1] = idx; | |
8382d914 | 4512 | |
55254636 | 4513 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4514 | |
4515 | return hash & (num_fault_mutexes - 1); | |
4516 | } | |
4517 | #else | |
4518 | /* | |
4519 | * For uniprocesor systems we always use a single mutex, so just | |
4520 | * return 0 and avoid the hashing overhead. | |
4521 | */ | |
188b04a7 | 4522 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4523 | { |
4524 | return 0; | |
4525 | } | |
4526 | #endif | |
4527 | ||
2b740303 | 4528 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4529 | unsigned long address, unsigned int flags) |
86e5216f | 4530 | { |
8382d914 | 4531 | pte_t *ptep, entry; |
cb900f41 | 4532 | spinlock_t *ptl; |
2b740303 | 4533 | vm_fault_t ret; |
8382d914 DB |
4534 | u32 hash; |
4535 | pgoff_t idx; | |
0fe6e20b | 4536 | struct page *page = NULL; |
57303d80 | 4537 | struct page *pagecache_page = NULL; |
a5516438 | 4538 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4539 | struct address_space *mapping; |
0f792cf9 | 4540 | int need_wait_lock = 0; |
285b8dca | 4541 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4542 | |
285b8dca | 4543 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 4544 | if (ptep) { |
c0d0381a MK |
4545 | /* |
4546 | * Since we hold no locks, ptep could be stale. That is | |
4547 | * OK as we are only making decisions based on content and | |
4548 | * not actually modifying content here. | |
4549 | */ | |
fd6a03ed | 4550 | entry = huge_ptep_get(ptep); |
290408d4 | 4551 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4552 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4553 | return 0; |
4554 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4555 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4556 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
4557 | } |
4558 | ||
c0d0381a MK |
4559 | /* |
4560 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
4561 | * until finished with ptep. This serves two purposes: |
4562 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
4563 | * and making the ptep no longer valid. | |
4564 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
4565 | * |
4566 | * ptep could have already be assigned via huge_pte_offset. That | |
4567 | * is OK, as huge_pte_alloc will return the same value unless | |
4568 | * something has changed. | |
4569 | */ | |
8382d914 | 4570 | mapping = vma->vm_file->f_mapping; |
c0d0381a MK |
4571 | i_mmap_lock_read(mapping); |
4572 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4573 | if (!ptep) { | |
4574 | i_mmap_unlock_read(mapping); | |
4575 | return VM_FAULT_OOM; | |
4576 | } | |
8382d914 | 4577 | |
3935baa9 DG |
4578 | /* |
4579 | * Serialize hugepage allocation and instantiation, so that we don't | |
4580 | * get spurious allocation failures if two CPUs race to instantiate | |
4581 | * the same page in the page cache. | |
4582 | */ | |
c0d0381a | 4583 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 4584 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4585 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4586 | |
7f2e9525 GS |
4587 | entry = huge_ptep_get(ptep); |
4588 | if (huge_pte_none(entry)) { | |
8382d914 | 4589 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4590 | goto out_mutex; |
3935baa9 | 4591 | } |
86e5216f | 4592 | |
83c54070 | 4593 | ret = 0; |
1e8f889b | 4594 | |
0f792cf9 NH |
4595 | /* |
4596 | * entry could be a migration/hwpoison entry at this point, so this | |
4597 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
4598 | * an active hugepage in pagecache. This goto expects the 2nd page |
4599 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
4600 | * properly handle it. | |
0f792cf9 NH |
4601 | */ |
4602 | if (!pte_present(entry)) | |
4603 | goto out_mutex; | |
4604 | ||
57303d80 AW |
4605 | /* |
4606 | * If we are going to COW the mapping later, we examine the pending | |
4607 | * reservations for this page now. This will ensure that any | |
4608 | * allocations necessary to record that reservation occur outside the | |
4609 | * spinlock. For private mappings, we also lookup the pagecache | |
4610 | * page now as it is used to determine if a reservation has been | |
4611 | * consumed. | |
4612 | */ | |
106c992a | 4613 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4614 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4615 | ret = VM_FAULT_OOM; |
b4d1d99f | 4616 | goto out_mutex; |
2b26736c | 4617 | } |
5e911373 | 4618 | /* Just decrements count, does not deallocate */ |
285b8dca | 4619 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4620 | |
f83a275d | 4621 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4622 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4623 | vma, haddr); |
57303d80 AW |
4624 | } |
4625 | ||
0f792cf9 NH |
4626 | ptl = huge_pte_lock(h, mm, ptep); |
4627 | ||
4628 | /* Check for a racing update before calling hugetlb_cow */ | |
4629 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4630 | goto out_ptl; | |
4631 | ||
56c9cfb1 NH |
4632 | /* |
4633 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4634 | * pagecache_page, so here we need take the former one | |
4635 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4636 | */ |
4637 | page = pte_page(entry); | |
4638 | if (page != pagecache_page) | |
0f792cf9 NH |
4639 | if (!trylock_page(page)) { |
4640 | need_wait_lock = 1; | |
4641 | goto out_ptl; | |
4642 | } | |
b4d1d99f | 4643 | |
0f792cf9 | 4644 | get_page(page); |
b4d1d99f | 4645 | |
788c7df4 | 4646 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4647 | if (!huge_pte_write(entry)) { |
974e6d66 | 4648 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4649 | pagecache_page, ptl); |
0f792cf9 | 4650 | goto out_put_page; |
b4d1d99f | 4651 | } |
106c992a | 4652 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4653 | } |
4654 | entry = pte_mkyoung(entry); | |
285b8dca | 4655 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4656 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4657 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4658 | out_put_page: |
4659 | if (page != pagecache_page) | |
4660 | unlock_page(page); | |
4661 | put_page(page); | |
cb900f41 KS |
4662 | out_ptl: |
4663 | spin_unlock(ptl); | |
57303d80 AW |
4664 | |
4665 | if (pagecache_page) { | |
4666 | unlock_page(pagecache_page); | |
4667 | put_page(pagecache_page); | |
4668 | } | |
b4d1d99f | 4669 | out_mutex: |
c672c7f2 | 4670 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4671 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
4672 | /* |
4673 | * Generally it's safe to hold refcount during waiting page lock. But | |
4674 | * here we just wait to defer the next page fault to avoid busy loop and | |
4675 | * the page is not used after unlocked before returning from the current | |
4676 | * page fault. So we are safe from accessing freed page, even if we wait | |
4677 | * here without taking refcount. | |
4678 | */ | |
4679 | if (need_wait_lock) | |
4680 | wait_on_page_locked(page); | |
1e8f889b | 4681 | return ret; |
86e5216f AL |
4682 | } |
4683 | ||
8fb5debc MK |
4684 | /* |
4685 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4686 | * modifications for huge pages. | |
4687 | */ | |
4688 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4689 | pte_t *dst_pte, | |
4690 | struct vm_area_struct *dst_vma, | |
4691 | unsigned long dst_addr, | |
4692 | unsigned long src_addr, | |
4693 | struct page **pagep) | |
4694 | { | |
1e392147 AA |
4695 | struct address_space *mapping; |
4696 | pgoff_t idx; | |
4697 | unsigned long size; | |
1c9e8def | 4698 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4699 | struct hstate *h = hstate_vma(dst_vma); |
4700 | pte_t _dst_pte; | |
4701 | spinlock_t *ptl; | |
4702 | int ret; | |
4703 | struct page *page; | |
4704 | ||
4705 | if (!*pagep) { | |
4706 | ret = -ENOMEM; | |
4707 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4708 | if (IS_ERR(page)) | |
4709 | goto out; | |
4710 | ||
4711 | ret = copy_huge_page_from_user(page, | |
4712 | (const void __user *) src_addr, | |
810a56b9 | 4713 | pages_per_huge_page(h), false); |
8fb5debc | 4714 | |
c1e8d7c6 | 4715 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 4716 | if (unlikely(ret)) { |
9e368259 | 4717 | ret = -ENOENT; |
8fb5debc MK |
4718 | *pagep = page; |
4719 | /* don't free the page */ | |
4720 | goto out; | |
4721 | } | |
4722 | } else { | |
4723 | page = *pagep; | |
4724 | *pagep = NULL; | |
4725 | } | |
4726 | ||
4727 | /* | |
4728 | * The memory barrier inside __SetPageUptodate makes sure that | |
4729 | * preceding stores to the page contents become visible before | |
4730 | * the set_pte_at() write. | |
4731 | */ | |
4732 | __SetPageUptodate(page); | |
8fb5debc | 4733 | |
1e392147 AA |
4734 | mapping = dst_vma->vm_file->f_mapping; |
4735 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4736 | ||
1c9e8def MK |
4737 | /* |
4738 | * If shared, add to page cache | |
4739 | */ | |
4740 | if (vm_shared) { | |
1e392147 AA |
4741 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4742 | ret = -EFAULT; | |
4743 | if (idx >= size) | |
4744 | goto out_release_nounlock; | |
1c9e8def | 4745 | |
1e392147 AA |
4746 | /* |
4747 | * Serialization between remove_inode_hugepages() and | |
4748 | * huge_add_to_page_cache() below happens through the | |
4749 | * hugetlb_fault_mutex_table that here must be hold by | |
4750 | * the caller. | |
4751 | */ | |
1c9e8def MK |
4752 | ret = huge_add_to_page_cache(page, mapping, idx); |
4753 | if (ret) | |
4754 | goto out_release_nounlock; | |
4755 | } | |
4756 | ||
8fb5debc MK |
4757 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4758 | spin_lock(ptl); | |
4759 | ||
1e392147 AA |
4760 | /* |
4761 | * Recheck the i_size after holding PT lock to make sure not | |
4762 | * to leave any page mapped (as page_mapped()) beyond the end | |
4763 | * of the i_size (remove_inode_hugepages() is strict about | |
4764 | * enforcing that). If we bail out here, we'll also leave a | |
4765 | * page in the radix tree in the vm_shared case beyond the end | |
4766 | * of the i_size, but remove_inode_hugepages() will take care | |
4767 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4768 | */ | |
4769 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4770 | ret = -EFAULT; | |
4771 | if (idx >= size) | |
4772 | goto out_release_unlock; | |
4773 | ||
8fb5debc MK |
4774 | ret = -EEXIST; |
4775 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4776 | goto out_release_unlock; | |
4777 | ||
1c9e8def MK |
4778 | if (vm_shared) { |
4779 | page_dup_rmap(page, true); | |
4780 | } else { | |
4781 | ClearPagePrivate(page); | |
4782 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4783 | } | |
8fb5debc MK |
4784 | |
4785 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4786 | if (dst_vma->vm_flags & VM_WRITE) | |
4787 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4788 | _dst_pte = pte_mkyoung(_dst_pte); | |
4789 | ||
4790 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4791 | ||
4792 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4793 | dst_vma->vm_flags & VM_WRITE); | |
4794 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4795 | ||
4796 | /* No need to invalidate - it was non-present before */ | |
4797 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4798 | ||
4799 | spin_unlock(ptl); | |
cb6acd01 | 4800 | set_page_huge_active(page); |
1c9e8def MK |
4801 | if (vm_shared) |
4802 | unlock_page(page); | |
8fb5debc MK |
4803 | ret = 0; |
4804 | out: | |
4805 | return ret; | |
4806 | out_release_unlock: | |
4807 | spin_unlock(ptl); | |
1c9e8def MK |
4808 | if (vm_shared) |
4809 | unlock_page(page); | |
5af10dfd | 4810 | out_release_nounlock: |
8fb5debc MK |
4811 | put_page(page); |
4812 | goto out; | |
4813 | } | |
4814 | ||
28a35716 ML |
4815 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4816 | struct page **pages, struct vm_area_struct **vmas, | |
4817 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 4818 | long i, unsigned int flags, int *locked) |
63551ae0 | 4819 | { |
d5d4b0aa KC |
4820 | unsigned long pfn_offset; |
4821 | unsigned long vaddr = *position; | |
28a35716 | 4822 | unsigned long remainder = *nr_pages; |
a5516438 | 4823 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4824 | int err = -EFAULT; |
63551ae0 | 4825 | |
63551ae0 | 4826 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4827 | pte_t *pte; |
cb900f41 | 4828 | spinlock_t *ptl = NULL; |
2a15efc9 | 4829 | int absent; |
4c887265 | 4830 | struct page *page; |
63551ae0 | 4831 | |
02057967 DR |
4832 | /* |
4833 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4834 | * potentially allocating memory. | |
4835 | */ | |
fa45f116 | 4836 | if (fatal_signal_pending(current)) { |
02057967 DR |
4837 | remainder = 0; |
4838 | break; | |
4839 | } | |
4840 | ||
4c887265 AL |
4841 | /* |
4842 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4843 | * each hugepage. We have to make sure we get the |
4c887265 | 4844 | * first, for the page indexing below to work. |
cb900f41 KS |
4845 | * |
4846 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4847 | */ |
7868a208 PA |
4848 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4849 | huge_page_size(h)); | |
cb900f41 KS |
4850 | if (pte) |
4851 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4852 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4853 | ||
4854 | /* | |
4855 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4856 | * an error where there's an empty slot with no huge pagecache |
4857 | * to back it. This way, we avoid allocating a hugepage, and | |
4858 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4859 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4860 | */ |
3ae77f43 HD |
4861 | if (absent && (flags & FOLL_DUMP) && |
4862 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4863 | if (pte) |
4864 | spin_unlock(ptl); | |
2a15efc9 HD |
4865 | remainder = 0; |
4866 | break; | |
4867 | } | |
63551ae0 | 4868 | |
9cc3a5bd NH |
4869 | /* |
4870 | * We need call hugetlb_fault for both hugepages under migration | |
4871 | * (in which case hugetlb_fault waits for the migration,) and | |
4872 | * hwpoisoned hugepages (in which case we need to prevent the | |
4873 | * caller from accessing to them.) In order to do this, we use | |
4874 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4875 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4876 | * both cases, and because we can't follow correct pages | |
4877 | * directly from any kind of swap entries. | |
4878 | */ | |
4879 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4880 | ((flags & FOLL_WRITE) && |
4881 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4882 | vm_fault_t ret; |
87ffc118 | 4883 | unsigned int fault_flags = 0; |
63551ae0 | 4884 | |
cb900f41 KS |
4885 | if (pte) |
4886 | spin_unlock(ptl); | |
87ffc118 AA |
4887 | if (flags & FOLL_WRITE) |
4888 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 4889 | if (locked) |
71335f37 PX |
4890 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
4891 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
4892 | if (flags & FOLL_NOWAIT) |
4893 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4894 | FAULT_FLAG_RETRY_NOWAIT; | |
4895 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
4896 | /* |
4897 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
4898 | * FAULT_FLAG_TRIED can co-exist | |
4899 | */ | |
87ffc118 AA |
4900 | fault_flags |= FAULT_FLAG_TRIED; |
4901 | } | |
4902 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4903 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4904 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4905 | remainder = 0; |
4906 | break; | |
4907 | } | |
4908 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 4909 | if (locked && |
1ac25013 | 4910 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 4911 | *locked = 0; |
87ffc118 AA |
4912 | *nr_pages = 0; |
4913 | /* | |
4914 | * VM_FAULT_RETRY must not return an | |
4915 | * error, it will return zero | |
4916 | * instead. | |
4917 | * | |
4918 | * No need to update "position" as the | |
4919 | * caller will not check it after | |
4920 | * *nr_pages is set to 0. | |
4921 | */ | |
4922 | return i; | |
4923 | } | |
4924 | continue; | |
4c887265 AL |
4925 | } |
4926 | ||
a5516438 | 4927 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4928 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 4929 | |
acbfb087 ZL |
4930 | /* |
4931 | * If subpage information not requested, update counters | |
4932 | * and skip the same_page loop below. | |
4933 | */ | |
4934 | if (!pages && !vmas && !pfn_offset && | |
4935 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4936 | (remainder >= pages_per_huge_page(h))) { | |
4937 | vaddr += huge_page_size(h); | |
4938 | remainder -= pages_per_huge_page(h); | |
4939 | i += pages_per_huge_page(h); | |
4940 | spin_unlock(ptl); | |
4941 | continue; | |
4942 | } | |
4943 | ||
d5d4b0aa | 4944 | same_page: |
d6692183 | 4945 | if (pages) { |
2a15efc9 | 4946 | pages[i] = mem_map_offset(page, pfn_offset); |
3faa52c0 JH |
4947 | /* |
4948 | * try_grab_page() should always succeed here, because: | |
4949 | * a) we hold the ptl lock, and b) we've just checked | |
4950 | * that the huge page is present in the page tables. If | |
4951 | * the huge page is present, then the tail pages must | |
4952 | * also be present. The ptl prevents the head page and | |
4953 | * tail pages from being rearranged in any way. So this | |
4954 | * page must be available at this point, unless the page | |
4955 | * refcount overflowed: | |
4956 | */ | |
4957 | if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { | |
4958 | spin_unlock(ptl); | |
4959 | remainder = 0; | |
4960 | err = -ENOMEM; | |
4961 | break; | |
4962 | } | |
d6692183 | 4963 | } |
63551ae0 DG |
4964 | |
4965 | if (vmas) | |
4966 | vmas[i] = vma; | |
4967 | ||
4968 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4969 | ++pfn_offset; |
63551ae0 DG |
4970 | --remainder; |
4971 | ++i; | |
d5d4b0aa | 4972 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4973 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4974 | /* |
4975 | * We use pfn_offset to avoid touching the pageframes | |
4976 | * of this compound page. | |
4977 | */ | |
4978 | goto same_page; | |
4979 | } | |
cb900f41 | 4980 | spin_unlock(ptl); |
63551ae0 | 4981 | } |
28a35716 | 4982 | *nr_pages = remainder; |
87ffc118 AA |
4983 | /* |
4984 | * setting position is actually required only if remainder is | |
4985 | * not zero but it's faster not to add a "if (remainder)" | |
4986 | * branch. | |
4987 | */ | |
63551ae0 DG |
4988 | *position = vaddr; |
4989 | ||
2be7cfed | 4990 | return i ? i : err; |
63551ae0 | 4991 | } |
8f860591 | 4992 | |
5491ae7b AK |
4993 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4994 | /* | |
4995 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4996 | * implement this. | |
4997 | */ | |
4998 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4999 | #endif | |
5000 | ||
7da4d641 | 5001 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
5002 | unsigned long address, unsigned long end, pgprot_t newprot) |
5003 | { | |
5004 | struct mm_struct *mm = vma->vm_mm; | |
5005 | unsigned long start = address; | |
5006 | pte_t *ptep; | |
5007 | pte_t pte; | |
a5516438 | 5008 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 5009 | unsigned long pages = 0; |
dff11abe | 5010 | bool shared_pmd = false; |
ac46d4f3 | 5011 | struct mmu_notifier_range range; |
dff11abe MK |
5012 | |
5013 | /* | |
5014 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 5015 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
5016 | * range if PMD sharing is possible. |
5017 | */ | |
7269f999 JG |
5018 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
5019 | 0, vma, mm, start, end); | |
ac46d4f3 | 5020 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
5021 | |
5022 | BUG_ON(address >= end); | |
ac46d4f3 | 5023 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 5024 | |
ac46d4f3 | 5025 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 5026 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 5027 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 5028 | spinlock_t *ptl; |
7868a208 | 5029 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
5030 | if (!ptep) |
5031 | continue; | |
cb900f41 | 5032 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 5033 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
7da4d641 | 5034 | pages++; |
cb900f41 | 5035 | spin_unlock(ptl); |
dff11abe | 5036 | shared_pmd = true; |
39dde65c | 5037 | continue; |
7da4d641 | 5038 | } |
a8bda28d NH |
5039 | pte = huge_ptep_get(ptep); |
5040 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
5041 | spin_unlock(ptl); | |
5042 | continue; | |
5043 | } | |
5044 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
5045 | swp_entry_t entry = pte_to_swp_entry(pte); | |
5046 | ||
5047 | if (is_write_migration_entry(entry)) { | |
5048 | pte_t newpte; | |
5049 | ||
5050 | make_migration_entry_read(&entry); | |
5051 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
5052 | set_huge_swap_pte_at(mm, address, ptep, |
5053 | newpte, huge_page_size(h)); | |
a8bda28d NH |
5054 | pages++; |
5055 | } | |
5056 | spin_unlock(ptl); | |
5057 | continue; | |
5058 | } | |
5059 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
5060 | pte_t old_pte; |
5061 | ||
5062 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
5063 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 5064 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 5065 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 5066 | pages++; |
8f860591 | 5067 | } |
cb900f41 | 5068 | spin_unlock(ptl); |
8f860591 | 5069 | } |
d833352a | 5070 | /* |
c8c06efa | 5071 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 5072 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 5073 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
5074 | * and that page table be reused and filled with junk. If we actually |
5075 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 5076 | */ |
dff11abe | 5077 | if (shared_pmd) |
ac46d4f3 | 5078 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
5079 | else |
5080 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
5081 | /* |
5082 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
5083 | * page table protection not changing it to point to a new page. | |
5084 | * | |
ad56b738 | 5085 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 5086 | */ |
83cde9e8 | 5087 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 5088 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
5089 | |
5090 | return pages << h->order; | |
8f860591 ZY |
5091 | } |
5092 | ||
a1e78772 MG |
5093 | int hugetlb_reserve_pages(struct inode *inode, |
5094 | long from, long to, | |
5a6fe125 | 5095 | struct vm_area_struct *vma, |
ca16d140 | 5096 | vm_flags_t vm_flags) |
e4e574b7 | 5097 | { |
0db9d74e | 5098 | long ret, chg, add = -1; |
a5516438 | 5099 | struct hstate *h = hstate_inode(inode); |
90481622 | 5100 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 5101 | struct resv_map *resv_map; |
075a61d0 | 5102 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 5103 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 5104 | |
63489f8e MK |
5105 | /* This should never happen */ |
5106 | if (from > to) { | |
5107 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
5108 | return -EINVAL; | |
5109 | } | |
5110 | ||
17c9d12e MG |
5111 | /* |
5112 | * Only apply hugepage reservation if asked. At fault time, an | |
5113 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 5114 | * without using reserves |
17c9d12e | 5115 | */ |
ca16d140 | 5116 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
5117 | return 0; |
5118 | ||
a1e78772 MG |
5119 | /* |
5120 | * Shared mappings base their reservation on the number of pages that | |
5121 | * are already allocated on behalf of the file. Private mappings need | |
5122 | * to reserve the full area even if read-only as mprotect() may be | |
5123 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
5124 | */ | |
9119a41e | 5125 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
5126 | /* |
5127 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
5128 | * called for inodes for which resv_maps were created (see | |
5129 | * hugetlbfs_get_inode). | |
5130 | */ | |
4e35f483 | 5131 | resv_map = inode_resv_map(inode); |
9119a41e | 5132 | |
0db9d74e | 5133 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
5134 | |
5135 | } else { | |
e9fe92ae | 5136 | /* Private mapping. */ |
9119a41e | 5137 | resv_map = resv_map_alloc(); |
17c9d12e MG |
5138 | if (!resv_map) |
5139 | return -ENOMEM; | |
5140 | ||
a1e78772 | 5141 | chg = to - from; |
84afd99b | 5142 | |
17c9d12e MG |
5143 | set_vma_resv_map(vma, resv_map); |
5144 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
5145 | } | |
5146 | ||
c50ac050 DH |
5147 | if (chg < 0) { |
5148 | ret = chg; | |
5149 | goto out_err; | |
5150 | } | |
8a630112 | 5151 | |
075a61d0 MA |
5152 | ret = hugetlb_cgroup_charge_cgroup_rsvd( |
5153 | hstate_index(h), chg * pages_per_huge_page(h), &h_cg); | |
5154 | ||
5155 | if (ret < 0) { | |
5156 | ret = -ENOMEM; | |
5157 | goto out_err; | |
5158 | } | |
5159 | ||
5160 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
5161 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
5162 | * of the resv_map. | |
5163 | */ | |
5164 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
5165 | } | |
5166 | ||
1c5ecae3 MK |
5167 | /* |
5168 | * There must be enough pages in the subpool for the mapping. If | |
5169 | * the subpool has a minimum size, there may be some global | |
5170 | * reservations already in place (gbl_reserve). | |
5171 | */ | |
5172 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
5173 | if (gbl_reserve < 0) { | |
c50ac050 | 5174 | ret = -ENOSPC; |
075a61d0 | 5175 | goto out_uncharge_cgroup; |
c50ac050 | 5176 | } |
5a6fe125 MG |
5177 | |
5178 | /* | |
17c9d12e | 5179 | * Check enough hugepages are available for the reservation. |
90481622 | 5180 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 5181 | */ |
1c5ecae3 | 5182 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 5183 | if (ret < 0) { |
075a61d0 | 5184 | goto out_put_pages; |
68842c9b | 5185 | } |
17c9d12e MG |
5186 | |
5187 | /* | |
5188 | * Account for the reservations made. Shared mappings record regions | |
5189 | * that have reservations as they are shared by multiple VMAs. | |
5190 | * When the last VMA disappears, the region map says how much | |
5191 | * the reservation was and the page cache tells how much of | |
5192 | * the reservation was consumed. Private mappings are per-VMA and | |
5193 | * only the consumed reservations are tracked. When the VMA | |
5194 | * disappears, the original reservation is the VMA size and the | |
5195 | * consumed reservations are stored in the map. Hence, nothing | |
5196 | * else has to be done for private mappings here | |
5197 | */ | |
33039678 | 5198 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 5199 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
5200 | |
5201 | if (unlikely(add < 0)) { | |
5202 | hugetlb_acct_memory(h, -gbl_reserve); | |
7fc2513a | 5203 | ret = add; |
075a61d0 | 5204 | goto out_put_pages; |
0db9d74e | 5205 | } else if (unlikely(chg > add)) { |
33039678 MK |
5206 | /* |
5207 | * pages in this range were added to the reserve | |
5208 | * map between region_chg and region_add. This | |
5209 | * indicates a race with alloc_huge_page. Adjust | |
5210 | * the subpool and reserve counts modified above | |
5211 | * based on the difference. | |
5212 | */ | |
5213 | long rsv_adjust; | |
5214 | ||
0dd70e6b ML |
5215 | /* |
5216 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
5217 | * reference to h_cg->css. See comment below for detail. | |
5218 | */ | |
075a61d0 MA |
5219 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
5220 | hstate_index(h), | |
5221 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5222 | ||
33039678 MK |
5223 | rsv_adjust = hugepage_subpool_put_pages(spool, |
5224 | chg - add); | |
5225 | hugetlb_acct_memory(h, -rsv_adjust); | |
0dd70e6b ML |
5226 | } else if (h_cg) { |
5227 | /* | |
5228 | * The file_regions will hold their own reference to | |
5229 | * h_cg->css. So we should release the reference held | |
5230 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
5231 | * done. | |
5232 | */ | |
5233 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
5234 | } |
5235 | } | |
a43a8c39 | 5236 | return 0; |
075a61d0 MA |
5237 | out_put_pages: |
5238 | /* put back original number of pages, chg */ | |
5239 | (void)hugepage_subpool_put_pages(spool, chg); | |
5240 | out_uncharge_cgroup: | |
5241 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5242 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 5243 | out_err: |
5e911373 | 5244 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
5245 | /* Only call region_abort if the region_chg succeeded but the |
5246 | * region_add failed or didn't run. | |
5247 | */ | |
5248 | if (chg >= 0 && add < 0) | |
5249 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
5250 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
5251 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 5252 | return ret; |
a43a8c39 KC |
5253 | } |
5254 | ||
b5cec28d MK |
5255 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
5256 | long freed) | |
a43a8c39 | 5257 | { |
a5516438 | 5258 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 5259 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 5260 | long chg = 0; |
90481622 | 5261 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 5262 | long gbl_reserve; |
45c682a6 | 5263 | |
f27a5136 MK |
5264 | /* |
5265 | * Since this routine can be called in the evict inode path for all | |
5266 | * hugetlbfs inodes, resv_map could be NULL. | |
5267 | */ | |
b5cec28d MK |
5268 | if (resv_map) { |
5269 | chg = region_del(resv_map, start, end); | |
5270 | /* | |
5271 | * region_del() can fail in the rare case where a region | |
5272 | * must be split and another region descriptor can not be | |
5273 | * allocated. If end == LONG_MAX, it will not fail. | |
5274 | */ | |
5275 | if (chg < 0) | |
5276 | return chg; | |
5277 | } | |
5278 | ||
45c682a6 | 5279 | spin_lock(&inode->i_lock); |
e4c6f8be | 5280 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
5281 | spin_unlock(&inode->i_lock); |
5282 | ||
1c5ecae3 MK |
5283 | /* |
5284 | * If the subpool has a minimum size, the number of global | |
5285 | * reservations to be released may be adjusted. | |
5286 | */ | |
5287 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5288 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
5289 | |
5290 | return 0; | |
a43a8c39 | 5291 | } |
93f70f90 | 5292 | |
3212b535 SC |
5293 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
5294 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5295 | struct vm_area_struct *vma, | |
5296 | unsigned long addr, pgoff_t idx) | |
5297 | { | |
5298 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5299 | svma->vm_start; | |
5300 | unsigned long sbase = saddr & PUD_MASK; | |
5301 | unsigned long s_end = sbase + PUD_SIZE; | |
5302 | ||
5303 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
5304 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
5305 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
5306 | |
5307 | /* | |
5308 | * match the virtual addresses, permission and the alignment of the | |
5309 | * page table page. | |
5310 | */ | |
5311 | if (pmd_index(addr) != pmd_index(saddr) || | |
5312 | vm_flags != svm_flags || | |
5313 | sbase < svma->vm_start || svma->vm_end < s_end) | |
5314 | return 0; | |
5315 | ||
5316 | return saddr; | |
5317 | } | |
5318 | ||
31aafb45 | 5319 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
5320 | { |
5321 | unsigned long base = addr & PUD_MASK; | |
5322 | unsigned long end = base + PUD_SIZE; | |
5323 | ||
5324 | /* | |
5325 | * check on proper vm_flags and page table alignment | |
5326 | */ | |
017b1660 | 5327 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
5328 | return true; |
5329 | return false; | |
3212b535 SC |
5330 | } |
5331 | ||
017b1660 MK |
5332 | /* |
5333 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5334 | * If yes, adjust start and end to cover range associated with possible | |
5335 | * shared pmd mappings. | |
5336 | */ | |
5337 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5338 | unsigned long *start, unsigned long *end) | |
5339 | { | |
6d6c2964 LX |
5340 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
5341 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 5342 | |
6d6c2964 LX |
5343 | /* |
5344 | * vma need span at least one aligned PUD size and the start,end range | |
5345 | * must at least partialy within it. | |
5346 | */ | |
5347 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
5348 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
5349 | return; |
5350 | ||
75802ca6 | 5351 | /* Extend the range to be PUD aligned for a worst case scenario */ |
6d6c2964 LX |
5352 | if (*start > v_start) |
5353 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 5354 | |
6d6c2964 LX |
5355 | if (*end < v_end) |
5356 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
5357 | } |
5358 | ||
3212b535 SC |
5359 | /* |
5360 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5361 | * and returns the corresponding pte. While this is not necessary for the | |
5362 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
5363 | * code much cleaner. |
5364 | * | |
0bf7b64e MK |
5365 | * This routine must be called with i_mmap_rwsem held in at least read mode if |
5366 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
5367 | * table entries associated with the address space. This is important as we | |
5368 | * are setting up sharing based on existing page table entries (mappings). | |
5369 | * | |
5370 | * NOTE: This routine is only called from huge_pte_alloc. Some callers of | |
5371 | * huge_pte_alloc know that sharing is not possible and do not take | |
5372 | * i_mmap_rwsem as a performance optimization. This is handled by the | |
5373 | * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is | |
5374 | * only required for subsequent processing. | |
3212b535 SC |
5375 | */ |
5376 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5377 | { | |
5378 | struct vm_area_struct *vma = find_vma(mm, addr); | |
5379 | struct address_space *mapping = vma->vm_file->f_mapping; | |
5380 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5381 | vma->vm_pgoff; | |
5382 | struct vm_area_struct *svma; | |
5383 | unsigned long saddr; | |
5384 | pte_t *spte = NULL; | |
5385 | pte_t *pte; | |
cb900f41 | 5386 | spinlock_t *ptl; |
3212b535 SC |
5387 | |
5388 | if (!vma_shareable(vma, addr)) | |
5389 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
5390 | ||
0bf7b64e | 5391 | i_mmap_assert_locked(mapping); |
3212b535 SC |
5392 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
5393 | if (svma == vma) | |
5394 | continue; | |
5395 | ||
5396 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5397 | if (saddr) { | |
7868a208 PA |
5398 | spte = huge_pte_offset(svma->vm_mm, saddr, |
5399 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
5400 | if (spte) { |
5401 | get_page(virt_to_page(spte)); | |
5402 | break; | |
5403 | } | |
5404 | } | |
5405 | } | |
5406 | ||
5407 | if (!spte) | |
5408 | goto out; | |
5409 | ||
8bea8052 | 5410 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 5411 | if (pud_none(*pud)) { |
3212b535 SC |
5412 | pud_populate(mm, pud, |
5413 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5414 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5415 | } else { |
3212b535 | 5416 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5417 | } |
cb900f41 | 5418 | spin_unlock(ptl); |
3212b535 SC |
5419 | out: |
5420 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
5421 | return pte; |
5422 | } | |
5423 | ||
5424 | /* | |
5425 | * unmap huge page backed by shared pte. | |
5426 | * | |
5427 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5428 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5429 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5430 | * | |
c0d0381a | 5431 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
5432 | * |
5433 | * returns: 1 successfully unmapped a shared pte page | |
5434 | * 0 the underlying pte page is not shared, or it is the last user | |
5435 | */ | |
34ae204f MK |
5436 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5437 | unsigned long *addr, pte_t *ptep) | |
3212b535 SC |
5438 | { |
5439 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
5440 | p4d_t *p4d = p4d_offset(pgd, *addr); |
5441 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 | 5442 | |
34ae204f | 5443 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
5444 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
5445 | if (page_count(virt_to_page(ptep)) == 1) | |
5446 | return 0; | |
5447 | ||
5448 | pud_clear(pud); | |
5449 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 5450 | mm_dec_nr_pmds(mm); |
3212b535 SC |
5451 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
5452 | return 1; | |
5453 | } | |
9e5fc74c SC |
5454 | #define want_pmd_share() (1) |
5455 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
5456 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5457 | { | |
5458 | return NULL; | |
5459 | } | |
e81f2d22 | 5460 | |
34ae204f MK |
5461 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5462 | unsigned long *addr, pte_t *ptep) | |
e81f2d22 ZZ |
5463 | { |
5464 | return 0; | |
5465 | } | |
017b1660 MK |
5466 | |
5467 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5468 | unsigned long *start, unsigned long *end) | |
5469 | { | |
5470 | } | |
9e5fc74c | 5471 | #define want_pmd_share() (0) |
3212b535 SC |
5472 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5473 | ||
9e5fc74c SC |
5474 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
5475 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
5476 | unsigned long addr, unsigned long sz) | |
5477 | { | |
5478 | pgd_t *pgd; | |
c2febafc | 5479 | p4d_t *p4d; |
9e5fc74c SC |
5480 | pud_t *pud; |
5481 | pte_t *pte = NULL; | |
5482 | ||
5483 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5484 | p4d = p4d_alloc(mm, pgd, addr); |
5485 | if (!p4d) | |
5486 | return NULL; | |
c2febafc | 5487 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5488 | if (pud) { |
5489 | if (sz == PUD_SIZE) { | |
5490 | pte = (pte_t *)pud; | |
5491 | } else { | |
5492 | BUG_ON(sz != PMD_SIZE); | |
5493 | if (want_pmd_share() && pud_none(*pud)) | |
5494 | pte = huge_pmd_share(mm, addr, pud); | |
5495 | else | |
5496 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5497 | } | |
5498 | } | |
4e666314 | 5499 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5500 | |
5501 | return pte; | |
5502 | } | |
5503 | ||
9b19df29 PA |
5504 | /* |
5505 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5506 | * entry at address @addr | |
5507 | * | |
8ac0b81a LX |
5508 | * Return: Pointer to page table entry (PUD or PMD) for |
5509 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
5510 | * size @sz doesn't match the hugepage size at this level of the page |
5511 | * table. | |
5512 | */ | |
7868a208 PA |
5513 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5514 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5515 | { |
5516 | pgd_t *pgd; | |
c2febafc | 5517 | p4d_t *p4d; |
8ac0b81a LX |
5518 | pud_t *pud; |
5519 | pmd_t *pmd; | |
9e5fc74c SC |
5520 | |
5521 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5522 | if (!pgd_present(*pgd)) |
5523 | return NULL; | |
5524 | p4d = p4d_offset(pgd, addr); | |
5525 | if (!p4d_present(*p4d)) | |
5526 | return NULL; | |
9b19df29 | 5527 | |
c2febafc | 5528 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
5529 | if (sz == PUD_SIZE) |
5530 | /* must be pud huge, non-present or none */ | |
c2febafc | 5531 | return (pte_t *)pud; |
8ac0b81a | 5532 | if (!pud_present(*pud)) |
9b19df29 | 5533 | return NULL; |
8ac0b81a | 5534 | /* must have a valid entry and size to go further */ |
9b19df29 | 5535 | |
8ac0b81a LX |
5536 | pmd = pmd_offset(pud, addr); |
5537 | /* must be pmd huge, non-present or none */ | |
5538 | return (pte_t *)pmd; | |
9e5fc74c SC |
5539 | } |
5540 | ||
61f77eda NH |
5541 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5542 | ||
5543 | /* | |
5544 | * These functions are overwritable if your architecture needs its own | |
5545 | * behavior. | |
5546 | */ | |
5547 | struct page * __weak | |
5548 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5549 | int write) | |
5550 | { | |
5551 | return ERR_PTR(-EINVAL); | |
5552 | } | |
5553 | ||
4dc71451 AK |
5554 | struct page * __weak |
5555 | follow_huge_pd(struct vm_area_struct *vma, | |
5556 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5557 | { | |
5558 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5559 | return NULL; | |
5560 | } | |
5561 | ||
61f77eda | 5562 | struct page * __weak |
9e5fc74c | 5563 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5564 | pmd_t *pmd, int flags) |
9e5fc74c | 5565 | { |
e66f17ff NH |
5566 | struct page *page = NULL; |
5567 | spinlock_t *ptl; | |
c9d398fa | 5568 | pte_t pte; |
3faa52c0 JH |
5569 | |
5570 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5571 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5572 | (FOLL_PIN | FOLL_GET))) | |
5573 | return NULL; | |
5574 | ||
e66f17ff NH |
5575 | retry: |
5576 | ptl = pmd_lockptr(mm, pmd); | |
5577 | spin_lock(ptl); | |
5578 | /* | |
5579 | * make sure that the address range covered by this pmd is not | |
5580 | * unmapped from other threads. | |
5581 | */ | |
5582 | if (!pmd_huge(*pmd)) | |
5583 | goto out; | |
c9d398fa NH |
5584 | pte = huge_ptep_get((pte_t *)pmd); |
5585 | if (pte_present(pte)) { | |
97534127 | 5586 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
5587 | /* |
5588 | * try_grab_page() should always succeed here, because: a) we | |
5589 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5590 | * huge pmd (head) page is present in the page tables. The ptl | |
5591 | * prevents the head page and tail pages from being rearranged | |
5592 | * in any way. So this page must be available at this point, | |
5593 | * unless the page refcount overflowed: | |
5594 | */ | |
5595 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5596 | page = NULL; | |
5597 | goto out; | |
5598 | } | |
e66f17ff | 5599 | } else { |
c9d398fa | 5600 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5601 | spin_unlock(ptl); |
5602 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5603 | goto retry; | |
5604 | } | |
5605 | /* | |
5606 | * hwpoisoned entry is treated as no_page_table in | |
5607 | * follow_page_mask(). | |
5608 | */ | |
5609 | } | |
5610 | out: | |
5611 | spin_unlock(ptl); | |
9e5fc74c SC |
5612 | return page; |
5613 | } | |
5614 | ||
61f77eda | 5615 | struct page * __weak |
9e5fc74c | 5616 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5617 | pud_t *pud, int flags) |
9e5fc74c | 5618 | { |
3faa52c0 | 5619 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5620 | return NULL; |
9e5fc74c | 5621 | |
e66f17ff | 5622 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5623 | } |
5624 | ||
faaa5b62 AK |
5625 | struct page * __weak |
5626 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5627 | { | |
3faa52c0 | 5628 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5629 | return NULL; |
5630 | ||
5631 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5632 | } | |
5633 | ||
31caf665 NH |
5634 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5635 | { | |
bcc54222 NH |
5636 | bool ret = true; |
5637 | ||
31caf665 | 5638 | spin_lock(&hugetlb_lock); |
0eb2df2b MS |
5639 | if (!PageHeadHuge(page) || !page_huge_active(page) || |
5640 | !get_page_unless_zero(page)) { | |
bcc54222 NH |
5641 | ret = false; |
5642 | goto unlock; | |
5643 | } | |
5644 | clear_page_huge_active(page); | |
31caf665 | 5645 | list_move_tail(&page->lru, list); |
bcc54222 | 5646 | unlock: |
31caf665 | 5647 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5648 | return ret; |
31caf665 NH |
5649 | } |
5650 | ||
5651 | void putback_active_hugepage(struct page *page) | |
5652 | { | |
309381fe | 5653 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5654 | spin_lock(&hugetlb_lock); |
bcc54222 | 5655 | set_page_huge_active(page); |
31caf665 NH |
5656 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5657 | spin_unlock(&hugetlb_lock); | |
5658 | put_page(page); | |
5659 | } | |
ab5ac90a MH |
5660 | |
5661 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5662 | { | |
5663 | struct hstate *h = page_hstate(oldpage); | |
5664 | ||
5665 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5666 | set_page_owner_migrate_reason(newpage, reason); | |
5667 | ||
5668 | /* | |
5669 | * transfer temporary state of the new huge page. This is | |
5670 | * reverse to other transitions because the newpage is going to | |
5671 | * be final while the old one will be freed so it takes over | |
5672 | * the temporary status. | |
5673 | * | |
5674 | * Also note that we have to transfer the per-node surplus state | |
5675 | * here as well otherwise the global surplus count will not match | |
5676 | * the per-node's. | |
5677 | */ | |
5678 | if (PageHugeTemporary(newpage)) { | |
5679 | int old_nid = page_to_nid(oldpage); | |
5680 | int new_nid = page_to_nid(newpage); | |
5681 | ||
5682 | SetPageHugeTemporary(oldpage); | |
5683 | ClearPageHugeTemporary(newpage); | |
5684 | ||
5685 | spin_lock(&hugetlb_lock); | |
5686 | if (h->surplus_huge_pages_node[old_nid]) { | |
5687 | h->surplus_huge_pages_node[old_nid]--; | |
5688 | h->surplus_huge_pages_node[new_nid]++; | |
5689 | } | |
5690 | spin_unlock(&hugetlb_lock); | |
5691 | } | |
5692 | } | |
cf11e85f RG |
5693 | |
5694 | #ifdef CONFIG_CMA | |
cf11e85f RG |
5695 | static bool cma_reserve_called __initdata; |
5696 | ||
5697 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
5698 | { | |
5699 | hugetlb_cma_size = memparse(p, &p); | |
5700 | return 0; | |
5701 | } | |
5702 | ||
5703 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
5704 | ||
5705 | void __init hugetlb_cma_reserve(int order) | |
5706 | { | |
5707 | unsigned long size, reserved, per_node; | |
5708 | int nid; | |
5709 | ||
5710 | cma_reserve_called = true; | |
5711 | ||
5712 | if (!hugetlb_cma_size) | |
5713 | return; | |
5714 | ||
5715 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
5716 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
5717 | (PAGE_SIZE << order) / SZ_1M); | |
5718 | return; | |
5719 | } | |
5720 | ||
5721 | /* | |
5722 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
5723 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
5724 | */ | |
5725 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
5726 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
5727 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
5728 | ||
5729 | reserved = 0; | |
5730 | for_each_node_state(nid, N_ONLINE) { | |
5731 | int res; | |
2281f797 | 5732 | char name[CMA_MAX_NAME]; |
cf11e85f RG |
5733 | |
5734 | size = min(per_node, hugetlb_cma_size - reserved); | |
5735 | size = round_up(size, PAGE_SIZE << order); | |
5736 | ||
2281f797 | 5737 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
cf11e85f | 5738 | res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, |
29d0f41d | 5739 | 0, false, name, |
cf11e85f RG |
5740 | &hugetlb_cma[nid], nid); |
5741 | if (res) { | |
5742 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
5743 | res, nid); | |
5744 | continue; | |
5745 | } | |
5746 | ||
5747 | reserved += size; | |
5748 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
5749 | size / SZ_1M, nid); | |
5750 | ||
5751 | if (reserved >= hugetlb_cma_size) | |
5752 | break; | |
5753 | } | |
5754 | } | |
5755 | ||
5756 | void __init hugetlb_cma_check(void) | |
5757 | { | |
5758 | if (!hugetlb_cma_size || cma_reserve_called) | |
5759 | return; | |
5760 | ||
5761 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
5762 | } | |
5763 | ||
5764 | #endif /* CONFIG_CMA */ |