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> |
8cc5fcbb | 33 | #include <linux/migrate.h> |
f9317f77 | 34 | #include <linux/nospec.h> |
662ce1dc | 35 | #include <linux/delayacct.h> |
d6606683 | 36 | |
63551ae0 | 37 | #include <asm/page.h> |
ca15ca40 | 38 | #include <asm/pgalloc.h> |
24669e58 | 39 | #include <asm/tlb.h> |
63551ae0 | 40 | |
24669e58 | 41 | #include <linux/io.h> |
63551ae0 | 42 | #include <linux/hugetlb.h> |
9dd540e2 | 43 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 44 | #include <linux/node.h> |
ab5ac90a | 45 | #include <linux/page_owner.h> |
7835e98b | 46 | #include "internal.h" |
f41f2ed4 | 47 | #include "hugetlb_vmemmap.h" |
1da177e4 | 48 | |
c3f38a38 | 49 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
50 | unsigned int default_hstate_idx; |
51 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 52 | |
dbda8fea | 53 | #ifdef CONFIG_CMA |
cf11e85f | 54 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
38e719ab | 55 | static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; |
a01f4390 MK |
56 | static bool hugetlb_cma_page(struct page *page, unsigned int order) |
57 | { | |
58 | return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, | |
59 | 1 << order); | |
60 | } | |
61 | #else | |
62 | static bool hugetlb_cma_page(struct page *page, unsigned int order) | |
63 | { | |
64 | return false; | |
65 | } | |
dbda8fea BS |
66 | #endif |
67 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 68 | |
53ba51d2 JT |
69 | __initdata LIST_HEAD(huge_boot_pages); |
70 | ||
e5ff2159 AK |
71 | /* for command line parsing */ |
72 | static struct hstate * __initdata parsed_hstate; | |
73 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 74 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 75 | static bool __initdata parsed_default_hugepagesz; |
b5389086 | 76 | static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; |
e5ff2159 | 77 | |
3935baa9 | 78 | /* |
31caf665 NH |
79 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
80 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 81 | */ |
c3f38a38 | 82 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 83 | |
8382d914 DB |
84 | /* |
85 | * Serializes faults on the same logical page. This is used to | |
86 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
87 | */ | |
88 | static int num_fault_mutexes; | |
c672c7f2 | 89 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 90 | |
7ca02d0a MK |
91 | /* Forward declaration */ |
92 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
93 | ||
1d88433b | 94 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 95 | { |
1d88433b ML |
96 | if (spool->count) |
97 | return false; | |
98 | if (spool->max_hpages != -1) | |
99 | return spool->used_hpages == 0; | |
100 | if (spool->min_hpages != -1) | |
101 | return spool->rsv_hpages == spool->min_hpages; | |
102 | ||
103 | return true; | |
104 | } | |
90481622 | 105 | |
db71ef79 MK |
106 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, |
107 | unsigned long irq_flags) | |
1d88433b | 108 | { |
db71ef79 | 109 | spin_unlock_irqrestore(&spool->lock, irq_flags); |
90481622 DG |
110 | |
111 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 112 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 113 | * free the subpool */ |
1d88433b | 114 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
115 | if (spool->min_hpages != -1) |
116 | hugetlb_acct_memory(spool->hstate, | |
117 | -spool->min_hpages); | |
90481622 | 118 | kfree(spool); |
7ca02d0a | 119 | } |
90481622 DG |
120 | } |
121 | ||
7ca02d0a MK |
122 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
123 | long min_hpages) | |
90481622 DG |
124 | { |
125 | struct hugepage_subpool *spool; | |
126 | ||
c6a91820 | 127 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
128 | if (!spool) |
129 | return NULL; | |
130 | ||
131 | spin_lock_init(&spool->lock); | |
132 | spool->count = 1; | |
7ca02d0a MK |
133 | spool->max_hpages = max_hpages; |
134 | spool->hstate = h; | |
135 | spool->min_hpages = min_hpages; | |
136 | ||
137 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
138 | kfree(spool); | |
139 | return NULL; | |
140 | } | |
141 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
142 | |
143 | return spool; | |
144 | } | |
145 | ||
146 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
147 | { | |
db71ef79 MK |
148 | unsigned long flags; |
149 | ||
150 | spin_lock_irqsave(&spool->lock, flags); | |
90481622 DG |
151 | BUG_ON(!spool->count); |
152 | spool->count--; | |
db71ef79 | 153 | unlock_or_release_subpool(spool, flags); |
90481622 DG |
154 | } |
155 | ||
1c5ecae3 MK |
156 | /* |
157 | * Subpool accounting for allocating and reserving pages. | |
158 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 159 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
160 | * global pools must be adjusted (upward). The returned value may |
161 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 162 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
163 | */ |
164 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
165 | long delta) |
166 | { | |
1c5ecae3 | 167 | long ret = delta; |
90481622 DG |
168 | |
169 | if (!spool) | |
1c5ecae3 | 170 | return ret; |
90481622 | 171 | |
db71ef79 | 172 | spin_lock_irq(&spool->lock); |
1c5ecae3 MK |
173 | |
174 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
175 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
176 | spool->used_hpages += delta; | |
177 | else { | |
178 | ret = -ENOMEM; | |
179 | goto unlock_ret; | |
180 | } | |
90481622 | 181 | } |
90481622 | 182 | |
09a95e29 MK |
183 | /* minimum size accounting */ |
184 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
185 | if (delta > spool->rsv_hpages) { |
186 | /* | |
187 | * Asking for more reserves than those already taken on | |
188 | * behalf of subpool. Return difference. | |
189 | */ | |
190 | ret = delta - spool->rsv_hpages; | |
191 | spool->rsv_hpages = 0; | |
192 | } else { | |
193 | ret = 0; /* reserves already accounted for */ | |
194 | spool->rsv_hpages -= delta; | |
195 | } | |
196 | } | |
197 | ||
198 | unlock_ret: | |
db71ef79 | 199 | spin_unlock_irq(&spool->lock); |
90481622 DG |
200 | return ret; |
201 | } | |
202 | ||
1c5ecae3 MK |
203 | /* |
204 | * Subpool accounting for freeing and unreserving pages. | |
205 | * Return the number of global page reservations that must be dropped. | |
206 | * The return value may only be different than the passed value (delta) | |
207 | * in the case where a subpool minimum size must be maintained. | |
208 | */ | |
209 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
210 | long delta) |
211 | { | |
1c5ecae3 | 212 | long ret = delta; |
db71ef79 | 213 | unsigned long flags; |
1c5ecae3 | 214 | |
90481622 | 215 | if (!spool) |
1c5ecae3 | 216 | return delta; |
90481622 | 217 | |
db71ef79 | 218 | spin_lock_irqsave(&spool->lock, flags); |
1c5ecae3 MK |
219 | |
220 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
221 | spool->used_hpages -= delta; | |
222 | ||
09a95e29 MK |
223 | /* minimum size accounting */ |
224 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
225 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
226 | ret = 0; | |
227 | else | |
228 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
229 | ||
230 | spool->rsv_hpages += delta; | |
231 | if (spool->rsv_hpages > spool->min_hpages) | |
232 | spool->rsv_hpages = spool->min_hpages; | |
233 | } | |
234 | ||
235 | /* | |
236 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
237 | * quota reference, free it now. | |
238 | */ | |
db71ef79 | 239 | unlock_or_release_subpool(spool, flags); |
1c5ecae3 MK |
240 | |
241 | return ret; | |
90481622 DG |
242 | } |
243 | ||
244 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
245 | { | |
246 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
247 | } | |
248 | ||
249 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
250 | { | |
496ad9aa | 251 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
252 | } |
253 | ||
0db9d74e MA |
254 | /* Helper that removes a struct file_region from the resv_map cache and returns |
255 | * it for use. | |
256 | */ | |
257 | static struct file_region * | |
258 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
259 | { | |
260 | struct file_region *nrg = NULL; | |
261 | ||
262 | VM_BUG_ON(resv->region_cache_count <= 0); | |
263 | ||
264 | resv->region_cache_count--; | |
265 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
266 | list_del(&nrg->link); |
267 | ||
268 | nrg->from = from; | |
269 | nrg->to = to; | |
270 | ||
271 | return nrg; | |
272 | } | |
273 | ||
075a61d0 MA |
274 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
275 | struct file_region *rg) | |
276 | { | |
277 | #ifdef CONFIG_CGROUP_HUGETLB | |
278 | nrg->reservation_counter = rg->reservation_counter; | |
279 | nrg->css = rg->css; | |
280 | if (rg->css) | |
281 | css_get(rg->css); | |
282 | #endif | |
283 | } | |
284 | ||
285 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
286 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
287 | struct hstate *h, | |
288 | struct resv_map *resv, | |
289 | struct file_region *nrg) | |
290 | { | |
291 | #ifdef CONFIG_CGROUP_HUGETLB | |
292 | if (h_cg) { | |
293 | nrg->reservation_counter = | |
294 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
295 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
296 | /* |
297 | * The caller will hold exactly one h_cg->css reference for the | |
298 | * whole contiguous reservation region. But this area might be | |
299 | * scattered when there are already some file_regions reside in | |
300 | * it. As a result, many file_regions may share only one css | |
301 | * reference. In order to ensure that one file_region must hold | |
302 | * exactly one h_cg->css reference, we should do css_get for | |
303 | * each file_region and leave the reference held by caller | |
304 | * untouched. | |
305 | */ | |
306 | css_get(&h_cg->css); | |
075a61d0 MA |
307 | if (!resv->pages_per_hpage) |
308 | resv->pages_per_hpage = pages_per_huge_page(h); | |
309 | /* pages_per_hpage should be the same for all entries in | |
310 | * a resv_map. | |
311 | */ | |
312 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
313 | } else { | |
314 | nrg->reservation_counter = NULL; | |
315 | nrg->css = NULL; | |
316 | } | |
317 | #endif | |
318 | } | |
319 | ||
d85aecf2 ML |
320 | static void put_uncharge_info(struct file_region *rg) |
321 | { | |
322 | #ifdef CONFIG_CGROUP_HUGETLB | |
323 | if (rg->css) | |
324 | css_put(rg->css); | |
325 | #endif | |
326 | } | |
327 | ||
a9b3f867 MA |
328 | static bool has_same_uncharge_info(struct file_region *rg, |
329 | struct file_region *org) | |
330 | { | |
331 | #ifdef CONFIG_CGROUP_HUGETLB | |
0739eb43 | 332 | return rg->reservation_counter == org->reservation_counter && |
a9b3f867 MA |
333 | rg->css == org->css; |
334 | ||
335 | #else | |
336 | return true; | |
337 | #endif | |
338 | } | |
339 | ||
340 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
341 | { | |
342 | struct file_region *nrg = NULL, *prg = NULL; | |
343 | ||
344 | prg = list_prev_entry(rg, link); | |
345 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
346 | has_same_uncharge_info(prg, rg)) { | |
347 | prg->to = rg->to; | |
348 | ||
349 | list_del(&rg->link); | |
d85aecf2 | 350 | put_uncharge_info(rg); |
a9b3f867 MA |
351 | kfree(rg); |
352 | ||
7db5e7b6 | 353 | rg = prg; |
a9b3f867 MA |
354 | } |
355 | ||
356 | nrg = list_next_entry(rg, link); | |
357 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
358 | has_same_uncharge_info(nrg, rg)) { | |
359 | nrg->from = rg->from; | |
360 | ||
361 | list_del(&rg->link); | |
d85aecf2 | 362 | put_uncharge_info(rg); |
a9b3f867 | 363 | kfree(rg); |
a9b3f867 MA |
364 | } |
365 | } | |
366 | ||
2103cf9c | 367 | static inline long |
84448c8e | 368 | hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, |
2103cf9c PX |
369 | long to, struct hstate *h, struct hugetlb_cgroup *cg, |
370 | long *regions_needed) | |
371 | { | |
372 | struct file_region *nrg; | |
373 | ||
374 | if (!regions_needed) { | |
375 | nrg = get_file_region_entry_from_cache(map, from, to); | |
376 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
84448c8e | 377 | list_add(&nrg->link, rg); |
2103cf9c PX |
378 | coalesce_file_region(map, nrg); |
379 | } else | |
380 | *regions_needed += 1; | |
381 | ||
382 | return to - from; | |
383 | } | |
384 | ||
972a3da3 WY |
385 | /* |
386 | * Must be called with resv->lock held. | |
387 | * | |
388 | * Calling this with regions_needed != NULL will count the number of pages | |
389 | * to be added but will not modify the linked list. And regions_needed will | |
390 | * indicate the number of file_regions needed in the cache to carry out to add | |
391 | * the regions for this range. | |
d75c6af9 MA |
392 | */ |
393 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 394 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 395 | struct hstate *h, long *regions_needed) |
d75c6af9 | 396 | { |
0db9d74e | 397 | long add = 0; |
d75c6af9 | 398 | struct list_head *head = &resv->regions; |
0db9d74e | 399 | long last_accounted_offset = f; |
84448c8e JK |
400 | struct file_region *iter, *trg = NULL; |
401 | struct list_head *rg = NULL; | |
d75c6af9 | 402 | |
0db9d74e MA |
403 | if (regions_needed) |
404 | *regions_needed = 0; | |
d75c6af9 | 405 | |
0db9d74e | 406 | /* In this loop, we essentially handle an entry for the range |
84448c8e | 407 | * [last_accounted_offset, iter->from), at every iteration, with some |
0db9d74e MA |
408 | * bounds checking. |
409 | */ | |
84448c8e | 410 | list_for_each_entry_safe(iter, trg, head, link) { |
0db9d74e | 411 | /* Skip irrelevant regions that start before our range. */ |
84448c8e | 412 | if (iter->from < f) { |
0db9d74e MA |
413 | /* If this region ends after the last accounted offset, |
414 | * then we need to update last_accounted_offset. | |
415 | */ | |
84448c8e JK |
416 | if (iter->to > last_accounted_offset) |
417 | last_accounted_offset = iter->to; | |
0db9d74e MA |
418 | continue; |
419 | } | |
d75c6af9 | 420 | |
0db9d74e MA |
421 | /* When we find a region that starts beyond our range, we've |
422 | * finished. | |
423 | */ | |
84448c8e JK |
424 | if (iter->from >= t) { |
425 | rg = iter->link.prev; | |
d75c6af9 | 426 | break; |
84448c8e | 427 | } |
d75c6af9 | 428 | |
84448c8e | 429 | /* Add an entry for last_accounted_offset -> iter->from, and |
0db9d74e MA |
430 | * update last_accounted_offset. |
431 | */ | |
84448c8e JK |
432 | if (iter->from > last_accounted_offset) |
433 | add += hugetlb_resv_map_add(resv, iter->link.prev, | |
2103cf9c | 434 | last_accounted_offset, |
84448c8e | 435 | iter->from, h, h_cg, |
2103cf9c | 436 | regions_needed); |
0db9d74e | 437 | |
84448c8e | 438 | last_accounted_offset = iter->to; |
0db9d74e MA |
439 | } |
440 | ||
441 | /* Handle the case where our range extends beyond | |
442 | * last_accounted_offset. | |
443 | */ | |
84448c8e JK |
444 | if (!rg) |
445 | rg = head->prev; | |
2103cf9c PX |
446 | if (last_accounted_offset < t) |
447 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
448 | t, h, h_cg, regions_needed); | |
0db9d74e | 449 | |
0db9d74e MA |
450 | return add; |
451 | } | |
452 | ||
453 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
454 | */ | |
455 | static int allocate_file_region_entries(struct resv_map *resv, | |
456 | int regions_needed) | |
457 | __must_hold(&resv->lock) | |
458 | { | |
459 | struct list_head allocated_regions; | |
460 | int to_allocate = 0, i = 0; | |
461 | struct file_region *trg = NULL, *rg = NULL; | |
462 | ||
463 | VM_BUG_ON(regions_needed < 0); | |
464 | ||
465 | INIT_LIST_HEAD(&allocated_regions); | |
466 | ||
467 | /* | |
468 | * Check for sufficient descriptors in the cache to accommodate | |
469 | * the number of in progress add operations plus regions_needed. | |
470 | * | |
471 | * This is a while loop because when we drop the lock, some other call | |
472 | * to region_add or region_del may have consumed some region_entries, | |
473 | * so we keep looping here until we finally have enough entries for | |
474 | * (adds_in_progress + regions_needed). | |
475 | */ | |
476 | while (resv->region_cache_count < | |
477 | (resv->adds_in_progress + regions_needed)) { | |
478 | to_allocate = resv->adds_in_progress + regions_needed - | |
479 | resv->region_cache_count; | |
480 | ||
481 | /* At this point, we should have enough entries in the cache | |
f0953a1b | 482 | * for all the existing adds_in_progress. We should only be |
0db9d74e | 483 | * needing to allocate for regions_needed. |
d75c6af9 | 484 | */ |
0db9d74e MA |
485 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
486 | ||
487 | spin_unlock(&resv->lock); | |
488 | for (i = 0; i < to_allocate; i++) { | |
489 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
490 | if (!trg) | |
491 | goto out_of_memory; | |
492 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 493 | } |
d75c6af9 | 494 | |
0db9d74e MA |
495 | spin_lock(&resv->lock); |
496 | ||
d3ec7b6e WY |
497 | list_splice(&allocated_regions, &resv->region_cache); |
498 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
499 | } |
500 | ||
0db9d74e | 501 | return 0; |
d75c6af9 | 502 | |
0db9d74e MA |
503 | out_of_memory: |
504 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
505 | list_del(&rg->link); | |
506 | kfree(rg); | |
507 | } | |
508 | return -ENOMEM; | |
d75c6af9 MA |
509 | } |
510 | ||
1dd308a7 MK |
511 | /* |
512 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
513 | * map. Regions will be taken from the cache to fill in this range. |
514 | * Sufficient regions should exist in the cache due to the previous | |
515 | * call to region_chg with the same range, but in some cases the cache will not | |
516 | * have sufficient entries due to races with other code doing region_add or | |
517 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 518 | * |
0db9d74e MA |
519 | * regions_needed is the out value provided by a previous call to region_chg. |
520 | * | |
521 | * Return the number of new huge pages added to the map. This number is greater | |
522 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 523 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
524 | * region_add of regions of length 1 never allocate file_regions and cannot |
525 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
526 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 527 | */ |
0db9d74e | 528 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
529 | long in_regions_needed, struct hstate *h, |
530 | struct hugetlb_cgroup *h_cg) | |
96822904 | 531 | { |
0db9d74e | 532 | long add = 0, actual_regions_needed = 0; |
96822904 | 533 | |
7b24d861 | 534 | spin_lock(&resv->lock); |
0db9d74e MA |
535 | retry: |
536 | ||
537 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
538 | add_reservation_in_range(resv, f, t, NULL, NULL, |
539 | &actual_regions_needed); | |
96822904 | 540 | |
5e911373 | 541 | /* |
0db9d74e MA |
542 | * Check for sufficient descriptors in the cache to accommodate |
543 | * this add operation. Note that actual_regions_needed may be greater | |
544 | * than in_regions_needed, as the resv_map may have been modified since | |
545 | * the region_chg call. In this case, we need to make sure that we | |
546 | * allocate extra entries, such that we have enough for all the | |
547 | * existing adds_in_progress, plus the excess needed for this | |
548 | * operation. | |
5e911373 | 549 | */ |
0db9d74e MA |
550 | if (actual_regions_needed > in_regions_needed && |
551 | resv->region_cache_count < | |
552 | resv->adds_in_progress + | |
553 | (actual_regions_needed - in_regions_needed)) { | |
554 | /* region_add operation of range 1 should never need to | |
555 | * allocate file_region entries. | |
556 | */ | |
557 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 558 | |
0db9d74e MA |
559 | if (allocate_file_region_entries( |
560 | resv, actual_regions_needed - in_regions_needed)) { | |
561 | return -ENOMEM; | |
562 | } | |
5e911373 | 563 | |
0db9d74e | 564 | goto retry; |
5e911373 MK |
565 | } |
566 | ||
972a3da3 | 567 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
568 | |
569 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 570 | |
7b24d861 | 571 | spin_unlock(&resv->lock); |
cf3ad20b | 572 | return add; |
96822904 AW |
573 | } |
574 | ||
1dd308a7 MK |
575 | /* |
576 | * Examine the existing reserve map and determine how many | |
577 | * huge pages in the specified range [f, t) are NOT currently | |
578 | * represented. This routine is called before a subsequent | |
579 | * call to region_add that will actually modify the reserve | |
580 | * map to add the specified range [f, t). region_chg does | |
581 | * not change the number of huge pages represented by the | |
0db9d74e MA |
582 | * map. A number of new file_region structures is added to the cache as a |
583 | * placeholder, for the subsequent region_add call to use. At least 1 | |
584 | * file_region structure is added. | |
585 | * | |
586 | * out_regions_needed is the number of regions added to the | |
587 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
588 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
589 | * |
590 | * Returns the number of huge pages that need to be added to the existing | |
591 | * reservation map for the range [f, t). This number is greater or equal to | |
592 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
593 | * is needed and can not be allocated. | |
1dd308a7 | 594 | */ |
0db9d74e MA |
595 | static long region_chg(struct resv_map *resv, long f, long t, |
596 | long *out_regions_needed) | |
96822904 | 597 | { |
96822904 AW |
598 | long chg = 0; |
599 | ||
7b24d861 | 600 | spin_lock(&resv->lock); |
5e911373 | 601 | |
972a3da3 | 602 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 603 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 604 | out_regions_needed); |
5e911373 | 605 | |
0db9d74e MA |
606 | if (*out_regions_needed == 0) |
607 | *out_regions_needed = 1; | |
5e911373 | 608 | |
0db9d74e MA |
609 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
610 | return -ENOMEM; | |
5e911373 | 611 | |
0db9d74e | 612 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 613 | |
7b24d861 | 614 | spin_unlock(&resv->lock); |
96822904 AW |
615 | return chg; |
616 | } | |
617 | ||
5e911373 MK |
618 | /* |
619 | * Abort the in progress add operation. The adds_in_progress field | |
620 | * of the resv_map keeps track of the operations in progress between | |
621 | * calls to region_chg and region_add. Operations are sometimes | |
622 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
623 | * is called to decrement the adds_in_progress counter. regions_needed |
624 | * is the value returned by the region_chg call, it is used to decrement | |
625 | * the adds_in_progress counter. | |
5e911373 MK |
626 | * |
627 | * NOTE: The range arguments [f, t) are not needed or used in this | |
628 | * routine. They are kept to make reading the calling code easier as | |
629 | * arguments will match the associated region_chg call. | |
630 | */ | |
0db9d74e MA |
631 | static void region_abort(struct resv_map *resv, long f, long t, |
632 | long regions_needed) | |
5e911373 MK |
633 | { |
634 | spin_lock(&resv->lock); | |
635 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 636 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
637 | spin_unlock(&resv->lock); |
638 | } | |
639 | ||
1dd308a7 | 640 | /* |
feba16e2 MK |
641 | * Delete the specified range [f, t) from the reserve map. If the |
642 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
643 | * should be deleted. Locate the regions which intersect [f, t) | |
644 | * and either trim, delete or split the existing regions. | |
645 | * | |
646 | * Returns the number of huge pages deleted from the reserve map. | |
647 | * In the normal case, the return value is zero or more. In the | |
648 | * case where a region must be split, a new region descriptor must | |
649 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
650 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
651 | * a region and possibly return -ENOMEM. Callers specifying | |
652 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 653 | */ |
feba16e2 | 654 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 655 | { |
1406ec9b | 656 | struct list_head *head = &resv->regions; |
96822904 | 657 | struct file_region *rg, *trg; |
feba16e2 MK |
658 | struct file_region *nrg = NULL; |
659 | long del = 0; | |
96822904 | 660 | |
feba16e2 | 661 | retry: |
7b24d861 | 662 | spin_lock(&resv->lock); |
feba16e2 | 663 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
664 | /* |
665 | * Skip regions before the range to be deleted. file_region | |
666 | * ranges are normally of the form [from, to). However, there | |
667 | * may be a "placeholder" entry in the map which is of the form | |
668 | * (from, to) with from == to. Check for placeholder entries | |
669 | * at the beginning of the range to be deleted. | |
670 | */ | |
671 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 672 | continue; |
dbe409e4 | 673 | |
feba16e2 | 674 | if (rg->from >= t) |
96822904 | 675 | break; |
96822904 | 676 | |
feba16e2 MK |
677 | if (f > rg->from && t < rg->to) { /* Must split region */ |
678 | /* | |
679 | * Check for an entry in the cache before dropping | |
680 | * lock and attempting allocation. | |
681 | */ | |
682 | if (!nrg && | |
683 | resv->region_cache_count > resv->adds_in_progress) { | |
684 | nrg = list_first_entry(&resv->region_cache, | |
685 | struct file_region, | |
686 | link); | |
687 | list_del(&nrg->link); | |
688 | resv->region_cache_count--; | |
689 | } | |
96822904 | 690 | |
feba16e2 MK |
691 | if (!nrg) { |
692 | spin_unlock(&resv->lock); | |
693 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
694 | if (!nrg) | |
695 | return -ENOMEM; | |
696 | goto retry; | |
697 | } | |
698 | ||
699 | del += t - f; | |
79aa925b | 700 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 701 | resv, rg, t - f, false); |
feba16e2 MK |
702 | |
703 | /* New entry for end of split region */ | |
704 | nrg->from = t; | |
705 | nrg->to = rg->to; | |
075a61d0 MA |
706 | |
707 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
708 | ||
feba16e2 MK |
709 | INIT_LIST_HEAD(&nrg->link); |
710 | ||
711 | /* Original entry is trimmed */ | |
712 | rg->to = f; | |
713 | ||
714 | list_add(&nrg->link, &rg->link); | |
715 | nrg = NULL; | |
96822904 | 716 | break; |
feba16e2 MK |
717 | } |
718 | ||
719 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
720 | del += rg->to - rg->from; | |
075a61d0 | 721 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 722 | rg->to - rg->from, true); |
feba16e2 MK |
723 | list_del(&rg->link); |
724 | kfree(rg); | |
725 | continue; | |
726 | } | |
727 | ||
728 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 729 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 730 | t - rg->from, false); |
075a61d0 | 731 | |
79aa925b MK |
732 | del += t - rg->from; |
733 | rg->from = t; | |
734 | } else { /* Trim end of region */ | |
075a61d0 | 735 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 736 | rg->to - f, false); |
79aa925b MK |
737 | |
738 | del += rg->to - f; | |
739 | rg->to = f; | |
feba16e2 | 740 | } |
96822904 | 741 | } |
7b24d861 | 742 | |
7b24d861 | 743 | spin_unlock(&resv->lock); |
feba16e2 MK |
744 | kfree(nrg); |
745 | return del; | |
96822904 AW |
746 | } |
747 | ||
b5cec28d MK |
748 | /* |
749 | * A rare out of memory error was encountered which prevented removal of | |
750 | * the reserve map region for a page. The huge page itself was free'ed | |
751 | * and removed from the page cache. This routine will adjust the subpool | |
752 | * usage count, and the global reserve count if needed. By incrementing | |
753 | * these counts, the reserve map entry which could not be deleted will | |
754 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
755 | * counts. | |
756 | */ | |
72e2936c | 757 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
758 | { |
759 | struct hugepage_subpool *spool = subpool_inode(inode); | |
760 | long rsv_adjust; | |
da56388c | 761 | bool reserved = false; |
b5cec28d MK |
762 | |
763 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 764 | if (rsv_adjust > 0) { |
b5cec28d MK |
765 | struct hstate *h = hstate_inode(inode); |
766 | ||
da56388c ML |
767 | if (!hugetlb_acct_memory(h, 1)) |
768 | reserved = true; | |
769 | } else if (!rsv_adjust) { | |
770 | reserved = true; | |
b5cec28d | 771 | } |
da56388c ML |
772 | |
773 | if (!reserved) | |
774 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
775 | } |
776 | ||
1dd308a7 MK |
777 | /* |
778 | * Count and return the number of huge pages in the reserve map | |
779 | * that intersect with the range [f, t). | |
780 | */ | |
1406ec9b | 781 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 782 | { |
1406ec9b | 783 | struct list_head *head = &resv->regions; |
84afd99b AW |
784 | struct file_region *rg; |
785 | long chg = 0; | |
786 | ||
7b24d861 | 787 | spin_lock(&resv->lock); |
84afd99b AW |
788 | /* Locate each segment we overlap with, and count that overlap. */ |
789 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
790 | long seg_from; |
791 | long seg_to; | |
84afd99b AW |
792 | |
793 | if (rg->to <= f) | |
794 | continue; | |
795 | if (rg->from >= t) | |
796 | break; | |
797 | ||
798 | seg_from = max(rg->from, f); | |
799 | seg_to = min(rg->to, t); | |
800 | ||
801 | chg += seg_to - seg_from; | |
802 | } | |
7b24d861 | 803 | spin_unlock(&resv->lock); |
84afd99b AW |
804 | |
805 | return chg; | |
806 | } | |
807 | ||
e7c4b0bf AW |
808 | /* |
809 | * Convert the address within this vma to the page offset within | |
810 | * the mapping, in pagecache page units; huge pages here. | |
811 | */ | |
a5516438 AK |
812 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
813 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 814 | { |
a5516438 AK |
815 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
816 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
817 | } |
818 | ||
0fe6e20b NH |
819 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
820 | unsigned long address) | |
821 | { | |
822 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
823 | } | |
dee41079 | 824 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 825 | |
08fba699 MG |
826 | /* |
827 | * Return the size of the pages allocated when backing a VMA. In the majority | |
828 | * cases this will be same size as used by the page table entries. | |
829 | */ | |
830 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
831 | { | |
05ea8860 DW |
832 | if (vma->vm_ops && vma->vm_ops->pagesize) |
833 | return vma->vm_ops->pagesize(vma); | |
834 | return PAGE_SIZE; | |
08fba699 | 835 | } |
f340ca0f | 836 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 837 | |
3340289d MG |
838 | /* |
839 | * Return the page size being used by the MMU to back a VMA. In the majority | |
840 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
841 | * architectures where it differs, an architecture-specific 'strong' |
842 | * version of this symbol is required. | |
3340289d | 843 | */ |
09135cc5 | 844 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
845 | { |
846 | return vma_kernel_pagesize(vma); | |
847 | } | |
3340289d | 848 | |
84afd99b AW |
849 | /* |
850 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
851 | * bits of the reservation map pointer, which are always clear due to | |
852 | * alignment. | |
853 | */ | |
854 | #define HPAGE_RESV_OWNER (1UL << 0) | |
855 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 856 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 857 | |
a1e78772 MG |
858 | /* |
859 | * These helpers are used to track how many pages are reserved for | |
860 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
861 | * is guaranteed to have their future faults succeed. | |
862 | * | |
863 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
864 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
865 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
866 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
867 | * |
868 | * The private mapping reservation is represented in a subtly different | |
869 | * manner to a shared mapping. A shared mapping has a region map associated | |
870 | * with the underlying file, this region map represents the backing file | |
871 | * pages which have ever had a reservation assigned which this persists even | |
872 | * after the page is instantiated. A private mapping has a region map | |
873 | * associated with the original mmap which is attached to all VMAs which | |
874 | * reference it, this region map represents those offsets which have consumed | |
875 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 876 | */ |
e7c4b0bf AW |
877 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
878 | { | |
879 | return (unsigned long)vma->vm_private_data; | |
880 | } | |
881 | ||
882 | static void set_vma_private_data(struct vm_area_struct *vma, | |
883 | unsigned long value) | |
884 | { | |
885 | vma->vm_private_data = (void *)value; | |
886 | } | |
887 | ||
e9fe92ae MA |
888 | static void |
889 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
890 | struct hugetlb_cgroup *h_cg, | |
891 | struct hstate *h) | |
892 | { | |
893 | #ifdef CONFIG_CGROUP_HUGETLB | |
894 | if (!h_cg || !h) { | |
895 | resv_map->reservation_counter = NULL; | |
896 | resv_map->pages_per_hpage = 0; | |
897 | resv_map->css = NULL; | |
898 | } else { | |
899 | resv_map->reservation_counter = | |
900 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
901 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
902 | resv_map->css = &h_cg->css; | |
903 | } | |
904 | #endif | |
905 | } | |
906 | ||
9119a41e | 907 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
908 | { |
909 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
910 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
911 | ||
912 | if (!resv_map || !rg) { | |
913 | kfree(resv_map); | |
914 | kfree(rg); | |
84afd99b | 915 | return NULL; |
5e911373 | 916 | } |
84afd99b AW |
917 | |
918 | kref_init(&resv_map->refs); | |
7b24d861 | 919 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
920 | INIT_LIST_HEAD(&resv_map->regions); |
921 | ||
5e911373 | 922 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
923 | /* |
924 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
925 | * fields don't do cgroup accounting. On private mappings, these will be | |
926 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
927 | * reservations are to be un-charged from here. | |
928 | */ | |
929 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
930 | |
931 | INIT_LIST_HEAD(&resv_map->region_cache); | |
932 | list_add(&rg->link, &resv_map->region_cache); | |
933 | resv_map->region_cache_count = 1; | |
934 | ||
84afd99b AW |
935 | return resv_map; |
936 | } | |
937 | ||
9119a41e | 938 | void resv_map_release(struct kref *ref) |
84afd99b AW |
939 | { |
940 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
941 | struct list_head *head = &resv_map->region_cache; |
942 | struct file_region *rg, *trg; | |
84afd99b AW |
943 | |
944 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 945 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
946 | |
947 | /* ... and any entries left in the cache */ | |
948 | list_for_each_entry_safe(rg, trg, head, link) { | |
949 | list_del(&rg->link); | |
950 | kfree(rg); | |
951 | } | |
952 | ||
953 | VM_BUG_ON(resv_map->adds_in_progress); | |
954 | ||
84afd99b AW |
955 | kfree(resv_map); |
956 | } | |
957 | ||
4e35f483 JK |
958 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
959 | { | |
f27a5136 MK |
960 | /* |
961 | * At inode evict time, i_mapping may not point to the original | |
962 | * address space within the inode. This original address space | |
963 | * contains the pointer to the resv_map. So, always use the | |
964 | * address space embedded within the inode. | |
965 | * The VERY common case is inode->mapping == &inode->i_data but, | |
966 | * this may not be true for device special inodes. | |
967 | */ | |
968 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
969 | } |
970 | ||
84afd99b | 971 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 972 | { |
81d1b09c | 973 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
974 | if (vma->vm_flags & VM_MAYSHARE) { |
975 | struct address_space *mapping = vma->vm_file->f_mapping; | |
976 | struct inode *inode = mapping->host; | |
977 | ||
978 | return inode_resv_map(inode); | |
979 | ||
980 | } else { | |
84afd99b AW |
981 | return (struct resv_map *)(get_vma_private_data(vma) & |
982 | ~HPAGE_RESV_MASK); | |
4e35f483 | 983 | } |
a1e78772 MG |
984 | } |
985 | ||
84afd99b | 986 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 987 | { |
81d1b09c SL |
988 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
989 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 990 | |
84afd99b AW |
991 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
992 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
993 | } |
994 | ||
995 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
996 | { | |
81d1b09c SL |
997 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
998 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
999 | |
1000 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
1001 | } |
1002 | ||
1003 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
1004 | { | |
81d1b09c | 1005 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
1006 | |
1007 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
1008 | } |
1009 | ||
04f2cbe3 | 1010 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
1011 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
1012 | { | |
81d1b09c | 1013 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 1014 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
1015 | vma->vm_private_data = (void *)0; |
1016 | } | |
1017 | ||
550a7d60 MA |
1018 | /* |
1019 | * Reset and decrement one ref on hugepage private reservation. | |
1020 | * Called with mm->mmap_sem writer semaphore held. | |
1021 | * This function should be only used by move_vma() and operate on | |
1022 | * same sized vma. It should never come here with last ref on the | |
1023 | * reservation. | |
1024 | */ | |
1025 | void clear_vma_resv_huge_pages(struct vm_area_struct *vma) | |
1026 | { | |
1027 | /* | |
1028 | * Clear the old hugetlb private page reservation. | |
1029 | * It has already been transferred to new_vma. | |
1030 | * | |
1031 | * During a mremap() operation of a hugetlb vma we call move_vma() | |
1032 | * which copies vma into new_vma and unmaps vma. After the copy | |
1033 | * operation both new_vma and vma share a reference to the resv_map | |
1034 | * struct, and at that point vma is about to be unmapped. We don't | |
1035 | * want to return the reservation to the pool at unmap of vma because | |
1036 | * the reservation still lives on in new_vma, so simply decrement the | |
1037 | * ref here and remove the resv_map reference from this vma. | |
1038 | */ | |
1039 | struct resv_map *reservations = vma_resv_map(vma); | |
1040 | ||
afe041c2 BQM |
1041 | if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1042 | resv_map_put_hugetlb_cgroup_uncharge_info(reservations); | |
550a7d60 | 1043 | kref_put(&reservations->refs, resv_map_release); |
afe041c2 | 1044 | } |
550a7d60 MA |
1045 | |
1046 | reset_vma_resv_huge_pages(vma); | |
1047 | } | |
1048 | ||
a1e78772 | 1049 | /* Returns true if the VMA has associated reserve pages */ |
559ec2f8 | 1050 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1051 | { |
af0ed73e JK |
1052 | if (vma->vm_flags & VM_NORESERVE) { |
1053 | /* | |
1054 | * This address is already reserved by other process(chg == 0), | |
1055 | * so, we should decrement reserved count. Without decrementing, | |
1056 | * reserve count remains after releasing inode, because this | |
1057 | * allocated page will go into page cache and is regarded as | |
1058 | * coming from reserved pool in releasing step. Currently, we | |
1059 | * don't have any other solution to deal with this situation | |
1060 | * properly, so add work-around here. | |
1061 | */ | |
1062 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1063 | return true; |
af0ed73e | 1064 | else |
559ec2f8 | 1065 | return false; |
af0ed73e | 1066 | } |
a63884e9 JK |
1067 | |
1068 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1069 | if (vma->vm_flags & VM_MAYSHARE) { |
1070 | /* | |
1071 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1072 | * be a region map for all pages. The only situation where | |
1073 | * there is no region map is if a hole was punched via | |
7c8de358 | 1074 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1075 | * use. This situation is indicated if chg != 0. |
1076 | */ | |
1077 | if (chg) | |
1078 | return false; | |
1079 | else | |
1080 | return true; | |
1081 | } | |
a63884e9 JK |
1082 | |
1083 | /* | |
1084 | * Only the process that called mmap() has reserves for | |
1085 | * private mappings. | |
1086 | */ | |
67961f9d MK |
1087 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1088 | /* | |
1089 | * Like the shared case above, a hole punch or truncate | |
1090 | * could have been performed on the private mapping. | |
1091 | * Examine the value of chg to determine if reserves | |
1092 | * actually exist or were previously consumed. | |
1093 | * Very Subtle - The value of chg comes from a previous | |
1094 | * call to vma_needs_reserves(). The reserve map for | |
1095 | * private mappings has different (opposite) semantics | |
1096 | * than that of shared mappings. vma_needs_reserves() | |
1097 | * has already taken this difference in semantics into | |
1098 | * account. Therefore, the meaning of chg is the same | |
1099 | * as in the shared case above. Code could easily be | |
1100 | * combined, but keeping it separate draws attention to | |
1101 | * subtle differences. | |
1102 | */ | |
1103 | if (chg) | |
1104 | return false; | |
1105 | else | |
1106 | return true; | |
1107 | } | |
a63884e9 | 1108 | |
559ec2f8 | 1109 | return false; |
a1e78772 MG |
1110 | } |
1111 | ||
a5516438 | 1112 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1113 | { |
1114 | int nid = page_to_nid(page); | |
9487ca60 MK |
1115 | |
1116 | lockdep_assert_held(&hugetlb_lock); | |
b65a4eda MK |
1117 | VM_BUG_ON_PAGE(page_count(page), page); |
1118 | ||
0edaecfa | 1119 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1120 | h->free_huge_pages++; |
1121 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1122 | SetHPageFreed(page); |
1da177e4 LT |
1123 | } |
1124 | ||
94310cbc | 1125 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1126 | { |
1127 | struct page *page; | |
1a08ae36 | 1128 | bool pin = !!(current->flags & PF_MEMALLOC_PIN); |
bbe88753 | 1129 | |
9487ca60 | 1130 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 | 1131 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
6077c943 | 1132 | if (pin && !is_longterm_pinnable_page(page)) |
bbe88753 | 1133 | continue; |
bf50bab2 | 1134 | |
6664bfc8 WY |
1135 | if (PageHWPoison(page)) |
1136 | continue; | |
1137 | ||
1138 | list_move(&page->lru, &h->hugepage_activelist); | |
1139 | set_page_refcounted(page); | |
6c037149 | 1140 | ClearHPageFreed(page); |
6664bfc8 WY |
1141 | h->free_huge_pages--; |
1142 | h->free_huge_pages_node[nid]--; | |
1143 | return page; | |
bbe88753 JK |
1144 | } |
1145 | ||
6664bfc8 | 1146 | return NULL; |
bf50bab2 NH |
1147 | } |
1148 | ||
3e59fcb0 MH |
1149 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1150 | nodemask_t *nmask) | |
94310cbc | 1151 | { |
3e59fcb0 MH |
1152 | unsigned int cpuset_mems_cookie; |
1153 | struct zonelist *zonelist; | |
1154 | struct zone *zone; | |
1155 | struct zoneref *z; | |
98fa15f3 | 1156 | int node = NUMA_NO_NODE; |
94310cbc | 1157 | |
3e59fcb0 MH |
1158 | zonelist = node_zonelist(nid, gfp_mask); |
1159 | ||
1160 | retry_cpuset: | |
1161 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1162 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1163 | struct page *page; | |
1164 | ||
1165 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1166 | continue; | |
1167 | /* | |
1168 | * no need to ask again on the same node. Pool is node rather than | |
1169 | * zone aware | |
1170 | */ | |
1171 | if (zone_to_nid(zone) == node) | |
1172 | continue; | |
1173 | node = zone_to_nid(zone); | |
94310cbc | 1174 | |
94310cbc AK |
1175 | page = dequeue_huge_page_node_exact(h, node); |
1176 | if (page) | |
1177 | return page; | |
1178 | } | |
3e59fcb0 MH |
1179 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1180 | goto retry_cpuset; | |
1181 | ||
94310cbc AK |
1182 | return NULL; |
1183 | } | |
1184 | ||
a5516438 AK |
1185 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1186 | struct vm_area_struct *vma, | |
af0ed73e JK |
1187 | unsigned long address, int avoid_reserve, |
1188 | long chg) | |
1da177e4 | 1189 | { |
cfcaa66f | 1190 | struct page *page = NULL; |
480eccf9 | 1191 | struct mempolicy *mpol; |
04ec6264 | 1192 | gfp_t gfp_mask; |
3e59fcb0 | 1193 | nodemask_t *nodemask; |
04ec6264 | 1194 | int nid; |
1da177e4 | 1195 | |
a1e78772 MG |
1196 | /* |
1197 | * A child process with MAP_PRIVATE mappings created by their parent | |
1198 | * have no page reserves. This check ensures that reservations are | |
1199 | * not "stolen". The child may still get SIGKILLed | |
1200 | */ | |
af0ed73e | 1201 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1202 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1203 | goto err; |
a1e78772 | 1204 | |
04f2cbe3 | 1205 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1206 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1207 | goto err; |
04f2cbe3 | 1208 | |
04ec6264 VB |
1209 | gfp_mask = htlb_alloc_mask(h); |
1210 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
1211 | |
1212 | if (mpol_is_preferred_many(mpol)) { | |
1213 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1214 | ||
1215 | /* Fallback to all nodes if page==NULL */ | |
1216 | nodemask = NULL; | |
1217 | } | |
1218 | ||
1219 | if (!page) | |
1220 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1221 | ||
3e59fcb0 | 1222 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { |
d6995da3 | 1223 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1224 | h->resv_huge_pages--; |
1da177e4 | 1225 | } |
cc9a6c87 | 1226 | |
52cd3b07 | 1227 | mpol_cond_put(mpol); |
1da177e4 | 1228 | return page; |
cc9a6c87 MG |
1229 | |
1230 | err: | |
cc9a6c87 | 1231 | return NULL; |
1da177e4 LT |
1232 | } |
1233 | ||
1cac6f2c LC |
1234 | /* |
1235 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1236 | * We may have allocated or freed a huge page based on a different | |
1237 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1238 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1239 | * node for alloc or free. | |
1240 | */ | |
1241 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1242 | { | |
0edaf86c | 1243 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1244 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1245 | ||
1246 | return nid; | |
1247 | } | |
1248 | ||
1249 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1250 | { | |
1251 | if (!node_isset(nid, *nodes_allowed)) | |
1252 | nid = next_node_allowed(nid, nodes_allowed); | |
1253 | return nid; | |
1254 | } | |
1255 | ||
1256 | /* | |
1257 | * returns the previously saved node ["this node"] from which to | |
1258 | * allocate a persistent huge page for the pool and advance the | |
1259 | * next node from which to allocate, handling wrap at end of node | |
1260 | * mask. | |
1261 | */ | |
1262 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1263 | nodemask_t *nodes_allowed) | |
1264 | { | |
1265 | int nid; | |
1266 | ||
1267 | VM_BUG_ON(!nodes_allowed); | |
1268 | ||
1269 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1270 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1271 | ||
1272 | return nid; | |
1273 | } | |
1274 | ||
1275 | /* | |
10c6ec49 | 1276 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1277 | * node ["this node"] from which to free a huge page. Advance the |
1278 | * next node id whether or not we find a free huge page to free so | |
1279 | * that the next attempt to free addresses the next node. | |
1280 | */ | |
1281 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1282 | { | |
1283 | int nid; | |
1284 | ||
1285 | VM_BUG_ON(!nodes_allowed); | |
1286 | ||
1287 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1288 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1289 | ||
1290 | return nid; | |
1291 | } | |
1292 | ||
1293 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1294 | for (nr_nodes = nodes_weight(*mask); \ | |
1295 | nr_nodes > 0 && \ | |
1296 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1297 | nr_nodes--) | |
1298 | ||
1299 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1300 | for (nr_nodes = nodes_weight(*mask); \ | |
1301 | nr_nodes > 0 && \ | |
1302 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1303 | nr_nodes--) | |
1304 | ||
8531fc6f | 1305 | /* used to demote non-gigantic_huge pages as well */ |
34d9e35b MK |
1306 | static void __destroy_compound_gigantic_page(struct page *page, |
1307 | unsigned int order, bool demote) | |
944d9fec LC |
1308 | { |
1309 | int i; | |
1310 | int nr_pages = 1 << order; | |
1311 | struct page *p = page + 1; | |
1312 | ||
c8cc708a | 1313 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1314 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1315 | |
944d9fec | 1316 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
a01f4390 | 1317 | p->mapping = NULL; |
1d798ca3 | 1318 | clear_compound_head(p); |
34d9e35b MK |
1319 | if (!demote) |
1320 | set_page_refcounted(p); | |
944d9fec LC |
1321 | } |
1322 | ||
1323 | set_compound_order(page, 0); | |
5232c63f | 1324 | #ifdef CONFIG_64BIT |
ba9c1201 | 1325 | page[1].compound_nr = 0; |
5232c63f | 1326 | #endif |
944d9fec LC |
1327 | __ClearPageHead(page); |
1328 | } | |
1329 | ||
8531fc6f MK |
1330 | static void destroy_compound_hugetlb_page_for_demote(struct page *page, |
1331 | unsigned int order) | |
1332 | { | |
1333 | __destroy_compound_gigantic_page(page, order, true); | |
1334 | } | |
1335 | ||
1336 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE | |
34d9e35b MK |
1337 | static void destroy_compound_gigantic_page(struct page *page, |
1338 | unsigned int order) | |
1339 | { | |
1340 | __destroy_compound_gigantic_page(page, order, false); | |
1341 | } | |
1342 | ||
d00181b9 | 1343 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1344 | { |
cf11e85f RG |
1345 | /* |
1346 | * If the page isn't allocated using the cma allocator, | |
1347 | * cma_release() returns false. | |
1348 | */ | |
dbda8fea BS |
1349 | #ifdef CONFIG_CMA |
1350 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1351 | return; |
dbda8fea | 1352 | #endif |
cf11e85f | 1353 | |
944d9fec LC |
1354 | free_contig_range(page_to_pfn(page), 1 << order); |
1355 | } | |
1356 | ||
4eb0716e | 1357 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1358 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1359 | int nid, nodemask_t *nodemask) | |
944d9fec | 1360 | { |
04adbc3f | 1361 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1362 | if (nid == NUMA_NO_NODE) |
1363 | nid = numa_mem_id(); | |
944d9fec | 1364 | |
dbda8fea BS |
1365 | #ifdef CONFIG_CMA |
1366 | { | |
cf11e85f RG |
1367 | struct page *page; |
1368 | int node; | |
1369 | ||
953f064a LX |
1370 | if (hugetlb_cma[nid]) { |
1371 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1372 | huge_page_order(h), true); | |
cf11e85f RG |
1373 | if (page) |
1374 | return page; | |
1375 | } | |
953f064a LX |
1376 | |
1377 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1378 | for_each_node_mask(node, *nodemask) { | |
1379 | if (node == nid || !hugetlb_cma[node]) | |
1380 | continue; | |
1381 | ||
1382 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1383 | huge_page_order(h), true); | |
1384 | if (page) | |
1385 | return page; | |
1386 | } | |
1387 | } | |
cf11e85f | 1388 | } |
dbda8fea | 1389 | #endif |
cf11e85f | 1390 | |
5e27a2df | 1391 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1392 | } |
1393 | ||
4eb0716e AG |
1394 | #else /* !CONFIG_CONTIG_ALLOC */ |
1395 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1396 | int nid, nodemask_t *nodemask) | |
1397 | { | |
1398 | return NULL; | |
1399 | } | |
1400 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1401 | |
e1073d1e | 1402 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1403 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1404 | int nid, nodemask_t *nodemask) |
1405 | { | |
1406 | return NULL; | |
1407 | } | |
d00181b9 | 1408 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1409 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1410 | unsigned int order) { } |
944d9fec LC |
1411 | #endif |
1412 | ||
6eb4e88a MK |
1413 | /* |
1414 | * Remove hugetlb page from lists, and update dtor so that page appears | |
34d9e35b MK |
1415 | * as just a compound page. |
1416 | * | |
1417 | * A reference is held on the page, except in the case of demote. | |
6eb4e88a MK |
1418 | * |
1419 | * Must be called with hugetlb lock held. | |
1420 | */ | |
34d9e35b MK |
1421 | static void __remove_hugetlb_page(struct hstate *h, struct page *page, |
1422 | bool adjust_surplus, | |
1423 | bool demote) | |
6eb4e88a MK |
1424 | { |
1425 | int nid = page_to_nid(page); | |
1426 | ||
1427 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1428 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1429 | ||
9487ca60 | 1430 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1431 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1432 | return; | |
1433 | ||
1434 | list_del(&page->lru); | |
1435 | ||
1436 | if (HPageFreed(page)) { | |
1437 | h->free_huge_pages--; | |
1438 | h->free_huge_pages_node[nid]--; | |
1439 | } | |
1440 | if (adjust_surplus) { | |
1441 | h->surplus_huge_pages--; | |
1442 | h->surplus_huge_pages_node[nid]--; | |
1443 | } | |
1444 | ||
e32d20c0 MK |
1445 | /* |
1446 | * Very subtle | |
1447 | * | |
1448 | * For non-gigantic pages set the destructor to the normal compound | |
1449 | * page dtor. This is needed in case someone takes an additional | |
1450 | * temporary ref to the page, and freeing is delayed until they drop | |
1451 | * their reference. | |
1452 | * | |
1453 | * For gigantic pages set the destructor to the null dtor. This | |
1454 | * destructor will never be called. Before freeing the gigantic | |
1455 | * page destroy_compound_gigantic_page will turn the compound page | |
1456 | * into a simple group of pages. After this the destructor does not | |
1457 | * apply. | |
1458 | * | |
1459 | * This handles the case where more than one ref is held when and | |
1460 | * after update_and_free_page is called. | |
34d9e35b MK |
1461 | * |
1462 | * In the case of demote we do not ref count the page as it will soon | |
1463 | * be turned into a page of smaller size. | |
e32d20c0 | 1464 | */ |
34d9e35b MK |
1465 | if (!demote) |
1466 | set_page_refcounted(page); | |
e32d20c0 MK |
1467 | if (hstate_is_gigantic(h)) |
1468 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1469 | else | |
1470 | set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); | |
6eb4e88a MK |
1471 | |
1472 | h->nr_huge_pages--; | |
1473 | h->nr_huge_pages_node[nid]--; | |
1474 | } | |
1475 | ||
34d9e35b MK |
1476 | static void remove_hugetlb_page(struct hstate *h, struct page *page, |
1477 | bool adjust_surplus) | |
1478 | { | |
1479 | __remove_hugetlb_page(h, page, adjust_surplus, false); | |
1480 | } | |
1481 | ||
8531fc6f MK |
1482 | static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, |
1483 | bool adjust_surplus) | |
1484 | { | |
1485 | __remove_hugetlb_page(h, page, adjust_surplus, true); | |
1486 | } | |
1487 | ||
ad2fa371 MS |
1488 | static void add_hugetlb_page(struct hstate *h, struct page *page, |
1489 | bool adjust_surplus) | |
1490 | { | |
1491 | int zeroed; | |
1492 | int nid = page_to_nid(page); | |
1493 | ||
1494 | VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); | |
1495 | ||
1496 | lockdep_assert_held(&hugetlb_lock); | |
1497 | ||
1498 | INIT_LIST_HEAD(&page->lru); | |
1499 | h->nr_huge_pages++; | |
1500 | h->nr_huge_pages_node[nid]++; | |
1501 | ||
1502 | if (adjust_surplus) { | |
1503 | h->surplus_huge_pages++; | |
1504 | h->surplus_huge_pages_node[nid]++; | |
1505 | } | |
1506 | ||
1507 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); | |
1508 | set_page_private(page, 0); | |
1509 | SetHPageVmemmapOptimized(page); | |
1510 | ||
1511 | /* | |
b65a4eda MK |
1512 | * This page is about to be managed by the hugetlb allocator and |
1513 | * should have no users. Drop our reference, and check for others | |
1514 | * just in case. | |
ad2fa371 MS |
1515 | */ |
1516 | zeroed = put_page_testzero(page); | |
b65a4eda MK |
1517 | if (!zeroed) |
1518 | /* | |
1519 | * It is VERY unlikely soneone else has taken a ref on | |
1520 | * the page. In this case, we simply return as the | |
1521 | * hugetlb destructor (free_huge_page) will be called | |
1522 | * when this other ref is dropped. | |
1523 | */ | |
1524 | return; | |
1525 | ||
ad2fa371 MS |
1526 | arch_clear_hugepage_flags(page); |
1527 | enqueue_huge_page(h, page); | |
1528 | } | |
1529 | ||
b65d4adb | 1530 | static void __update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1531 | { |
1532 | int i; | |
dbfee5ae | 1533 | struct page *subpage = page; |
a5516438 | 1534 | |
4eb0716e | 1535 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1536 | return; |
18229df5 | 1537 | |
5981611d | 1538 | if (hugetlb_vmemmap_alloc(h, page)) { |
ad2fa371 MS |
1539 | spin_lock_irq(&hugetlb_lock); |
1540 | /* | |
1541 | * If we cannot allocate vmemmap pages, just refuse to free the | |
1542 | * page and put the page back on the hugetlb free list and treat | |
1543 | * as a surplus page. | |
1544 | */ | |
1545 | add_hugetlb_page(h, page, true); | |
1546 | spin_unlock_irq(&hugetlb_lock); | |
1547 | return; | |
1548 | } | |
1549 | ||
dbfee5ae MK |
1550 | for (i = 0; i < pages_per_huge_page(h); |
1551 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1552 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1553 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1554 | 1 << PG_active | 1 << PG_private | |
1555 | 1 << PG_writeback); | |
6af2acb6 | 1556 | } |
a01f4390 MK |
1557 | |
1558 | /* | |
1559 | * Non-gigantic pages demoted from CMA allocated gigantic pages | |
1560 | * need to be given back to CMA in free_gigantic_page. | |
1561 | */ | |
1562 | if (hstate_is_gigantic(h) || | |
1563 | hugetlb_cma_page(page, huge_page_order(h))) { | |
944d9fec LC |
1564 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1565 | free_gigantic_page(page, huge_page_order(h)); | |
1566 | } else { | |
944d9fec LC |
1567 | __free_pages(page, huge_page_order(h)); |
1568 | } | |
6af2acb6 AL |
1569 | } |
1570 | ||
b65d4adb MS |
1571 | /* |
1572 | * As update_and_free_page() can be called under any context, so we cannot | |
1573 | * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the | |
1574 | * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate | |
1575 | * the vmemmap pages. | |
1576 | * | |
1577 | * free_hpage_workfn() locklessly retrieves the linked list of pages to be | |
1578 | * freed and frees them one-by-one. As the page->mapping pointer is going | |
1579 | * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node | |
1580 | * structure of a lockless linked list of huge pages to be freed. | |
1581 | */ | |
1582 | static LLIST_HEAD(hpage_freelist); | |
1583 | ||
1584 | static void free_hpage_workfn(struct work_struct *work) | |
1585 | { | |
1586 | struct llist_node *node; | |
1587 | ||
1588 | node = llist_del_all(&hpage_freelist); | |
1589 | ||
1590 | while (node) { | |
1591 | struct page *page; | |
1592 | struct hstate *h; | |
1593 | ||
1594 | page = container_of((struct address_space **)node, | |
1595 | struct page, mapping); | |
1596 | node = node->next; | |
1597 | page->mapping = NULL; | |
1598 | /* | |
1599 | * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() | |
1600 | * is going to trigger because a previous call to | |
1601 | * remove_hugetlb_page() will set_compound_page_dtor(page, | |
1602 | * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. | |
1603 | */ | |
1604 | h = size_to_hstate(page_size(page)); | |
1605 | ||
1606 | __update_and_free_page(h, page); | |
1607 | ||
1608 | cond_resched(); | |
1609 | } | |
1610 | } | |
1611 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1612 | ||
1613 | static inline void flush_free_hpage_work(struct hstate *h) | |
1614 | { | |
5981611d | 1615 | if (hugetlb_optimize_vmemmap_pages(h)) |
b65d4adb MS |
1616 | flush_work(&free_hpage_work); |
1617 | } | |
1618 | ||
1619 | static void update_and_free_page(struct hstate *h, struct page *page, | |
1620 | bool atomic) | |
1621 | { | |
ad2fa371 | 1622 | if (!HPageVmemmapOptimized(page) || !atomic) { |
b65d4adb MS |
1623 | __update_and_free_page(h, page); |
1624 | return; | |
1625 | } | |
1626 | ||
1627 | /* | |
1628 | * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. | |
1629 | * | |
1630 | * Only call schedule_work() if hpage_freelist is previously | |
1631 | * empty. Otherwise, schedule_work() had been called but the workfn | |
1632 | * hasn't retrieved the list yet. | |
1633 | */ | |
1634 | if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) | |
1635 | schedule_work(&free_hpage_work); | |
1636 | } | |
1637 | ||
10c6ec49 MK |
1638 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1639 | { | |
1640 | struct page *page, *t_page; | |
1641 | ||
1642 | list_for_each_entry_safe(page, t_page, list, lru) { | |
b65d4adb | 1643 | update_and_free_page(h, page, false); |
10c6ec49 MK |
1644 | cond_resched(); |
1645 | } | |
1646 | } | |
1647 | ||
e5ff2159 AK |
1648 | struct hstate *size_to_hstate(unsigned long size) |
1649 | { | |
1650 | struct hstate *h; | |
1651 | ||
1652 | for_each_hstate(h) { | |
1653 | if (huge_page_size(h) == size) | |
1654 | return h; | |
1655 | } | |
1656 | return NULL; | |
1657 | } | |
1658 | ||
db71ef79 | 1659 | void free_huge_page(struct page *page) |
27a85ef1 | 1660 | { |
a5516438 AK |
1661 | /* |
1662 | * Can't pass hstate in here because it is called from the | |
1663 | * compound page destructor. | |
1664 | */ | |
e5ff2159 | 1665 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1666 | int nid = page_to_nid(page); |
d6995da3 | 1667 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1668 | bool restore_reserve; |
db71ef79 | 1669 | unsigned long flags; |
27a85ef1 | 1670 | |
b4330afb MK |
1671 | VM_BUG_ON_PAGE(page_count(page), page); |
1672 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1673 | |
d6995da3 | 1674 | hugetlb_set_page_subpool(page, NULL); |
78fbe906 DH |
1675 | if (PageAnon(page)) |
1676 | __ClearPageAnonExclusive(page); | |
8ace22bc | 1677 | page->mapping = NULL; |
d6995da3 MK |
1678 | restore_reserve = HPageRestoreReserve(page); |
1679 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1680 | |
1c5ecae3 | 1681 | /* |
d6995da3 | 1682 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1683 | * reservation. If the page was associated with a subpool, there |
1684 | * would have been a page reserved in the subpool before allocation | |
1685 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1686 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1687 | * remove the reserved page from the subpool. |
1c5ecae3 | 1688 | */ |
0919e1b6 MK |
1689 | if (!restore_reserve) { |
1690 | /* | |
1691 | * A return code of zero implies that the subpool will be | |
1692 | * under its minimum size if the reservation is not restored | |
1693 | * after page is free. Therefore, force restore_reserve | |
1694 | * operation. | |
1695 | */ | |
1696 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1697 | restore_reserve = true; | |
1698 | } | |
1c5ecae3 | 1699 | |
db71ef79 | 1700 | spin_lock_irqsave(&hugetlb_lock, flags); |
8f251a3d | 1701 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1702 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1703 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1704 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1705 | pages_per_huge_page(h), page); | |
07443a85 JK |
1706 | if (restore_reserve) |
1707 | h->resv_huge_pages++; | |
1708 | ||
9157c311 | 1709 | if (HPageTemporary(page)) { |
6eb4e88a | 1710 | remove_hugetlb_page(h, page, false); |
db71ef79 | 1711 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1712 | update_and_free_page(h, page, true); |
ab5ac90a | 1713 | } else if (h->surplus_huge_pages_node[nid]) { |
0edaecfa | 1714 | /* remove the page from active list */ |
6eb4e88a | 1715 | remove_hugetlb_page(h, page, true); |
db71ef79 | 1716 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1717 | update_and_free_page(h, page, true); |
7893d1d5 | 1718 | } else { |
5d3a551c | 1719 | arch_clear_hugepage_flags(page); |
a5516438 | 1720 | enqueue_huge_page(h, page); |
db71ef79 | 1721 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1722 | } |
c77c0a8a WL |
1723 | } |
1724 | ||
d3d99fcc OS |
1725 | /* |
1726 | * Must be called with the hugetlb lock held | |
1727 | */ | |
1728 | static void __prep_account_new_huge_page(struct hstate *h, int nid) | |
1729 | { | |
1730 | lockdep_assert_held(&hugetlb_lock); | |
1731 | h->nr_huge_pages++; | |
1732 | h->nr_huge_pages_node[nid]++; | |
1733 | } | |
1734 | ||
f41f2ed4 | 1735 | static void __prep_new_huge_page(struct hstate *h, struct page *page) |
b7ba30c6 | 1736 | { |
5981611d | 1737 | hugetlb_vmemmap_free(h, page); |
0edaecfa | 1738 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1739 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1740 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1741 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1742 | set_hugetlb_cgroup_rsvd(page, NULL); |
d3d99fcc OS |
1743 | } |
1744 | ||
1745 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1746 | { | |
f41f2ed4 | 1747 | __prep_new_huge_page(h, page); |
db71ef79 | 1748 | spin_lock_irq(&hugetlb_lock); |
d3d99fcc | 1749 | __prep_account_new_huge_page(h, nid); |
db71ef79 | 1750 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1751 | } |
1752 | ||
34d9e35b MK |
1753 | static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, |
1754 | bool demote) | |
20a0307c | 1755 | { |
7118fc29 | 1756 | int i, j; |
20a0307c WF |
1757 | int nr_pages = 1 << order; |
1758 | struct page *p = page + 1; | |
1759 | ||
1760 | /* we rely on prep_new_huge_page to set the destructor */ | |
1761 | set_compound_order(page, order); | |
ef5a22be | 1762 | __ClearPageReserved(page); |
de09d31d | 1763 | __SetPageHead(page); |
20a0307c | 1764 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1765 | /* |
1766 | * For gigantic hugepages allocated through bootmem at | |
1767 | * boot, it's safer to be consistent with the not-gigantic | |
1768 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1769 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1770 | * pages may get the reference counting wrong if they see |
1771 | * PG_reserved set on a tail page (despite the head page not | |
1772 | * having PG_reserved set). Enforcing this consistency between | |
1773 | * head and tail pages allows drivers to optimize away a check | |
1774 | * on the head page when they need know if put_page() is needed | |
1775 | * after get_user_pages(). | |
1776 | */ | |
1777 | __ClearPageReserved(p); | |
7118fc29 MK |
1778 | /* |
1779 | * Subtle and very unlikely | |
1780 | * | |
1781 | * Gigantic 'page allocators' such as memblock or cma will | |
1782 | * return a set of pages with each page ref counted. We need | |
1783 | * to turn this set of pages into a compound page with tail | |
1784 | * page ref counts set to zero. Code such as speculative page | |
1785 | * cache adding could take a ref on a 'to be' tail page. | |
1786 | * We need to respect any increased ref count, and only set | |
1787 | * the ref count to zero if count is currently 1. If count | |
416d85ed MK |
1788 | * is not 1, we return an error. An error return indicates |
1789 | * the set of pages can not be converted to a gigantic page. | |
1790 | * The caller who allocated the pages should then discard the | |
1791 | * pages using the appropriate free interface. | |
34d9e35b MK |
1792 | * |
1793 | * In the case of demote, the ref count will be zero. | |
7118fc29 | 1794 | */ |
34d9e35b MK |
1795 | if (!demote) { |
1796 | if (!page_ref_freeze(p, 1)) { | |
1797 | pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); | |
1798 | goto out_error; | |
1799 | } | |
1800 | } else { | |
1801 | VM_BUG_ON_PAGE(page_count(p), p); | |
7118fc29 | 1802 | } |
1d798ca3 | 1803 | set_compound_head(p, page); |
20a0307c | 1804 | } |
b4330afb | 1805 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1806 | atomic_set(compound_pincount_ptr(page), 0); |
7118fc29 MK |
1807 | return true; |
1808 | ||
1809 | out_error: | |
1810 | /* undo tail page modifications made above */ | |
1811 | p = page + 1; | |
1812 | for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { | |
1813 | clear_compound_head(p); | |
1814 | set_page_refcounted(p); | |
1815 | } | |
1816 | /* need to clear PG_reserved on remaining tail pages */ | |
1817 | for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) | |
1818 | __ClearPageReserved(p); | |
1819 | set_compound_order(page, 0); | |
5232c63f | 1820 | #ifdef CONFIG_64BIT |
7118fc29 | 1821 | page[1].compound_nr = 0; |
5232c63f | 1822 | #endif |
7118fc29 MK |
1823 | __ClearPageHead(page); |
1824 | return false; | |
20a0307c WF |
1825 | } |
1826 | ||
34d9e35b MK |
1827 | static bool prep_compound_gigantic_page(struct page *page, unsigned int order) |
1828 | { | |
1829 | return __prep_compound_gigantic_page(page, order, false); | |
1830 | } | |
1831 | ||
8531fc6f MK |
1832 | static bool prep_compound_gigantic_page_for_demote(struct page *page, |
1833 | unsigned int order) | |
1834 | { | |
1835 | return __prep_compound_gigantic_page(page, order, true); | |
1836 | } | |
1837 | ||
7795912c AM |
1838 | /* |
1839 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1840 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1841 | * details. | |
1842 | */ | |
20a0307c WF |
1843 | int PageHuge(struct page *page) |
1844 | { | |
20a0307c WF |
1845 | if (!PageCompound(page)) |
1846 | return 0; | |
1847 | ||
1848 | page = compound_head(page); | |
f1e61557 | 1849 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1850 | } |
43131e14 NH |
1851 | EXPORT_SYMBOL_GPL(PageHuge); |
1852 | ||
27c73ae7 AA |
1853 | /* |
1854 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1855 | * normal or transparent huge pages. | |
1856 | */ | |
1857 | int PageHeadHuge(struct page *page_head) | |
1858 | { | |
27c73ae7 AA |
1859 | if (!PageHead(page_head)) |
1860 | return 0; | |
1861 | ||
d4af73e3 | 1862 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1863 | } |
4e936ecc | 1864 | EXPORT_SYMBOL_GPL(PageHeadHuge); |
27c73ae7 | 1865 | |
c0d0381a MK |
1866 | /* |
1867 | * Find and lock address space (mapping) in write mode. | |
1868 | * | |
336bf30e MK |
1869 | * Upon entry, the page is locked which means that page_mapping() is |
1870 | * stable. Due to locking order, we can only trylock_write. If we can | |
1871 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1872 | */ |
1873 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1874 | { | |
336bf30e | 1875 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1876 | |
c0d0381a MK |
1877 | if (!mapping) |
1878 | return mapping; | |
1879 | ||
c0d0381a MK |
1880 | if (i_mmap_trylock_write(mapping)) |
1881 | return mapping; | |
1882 | ||
336bf30e | 1883 | return NULL; |
c0d0381a MK |
1884 | } |
1885 | ||
fe19bd3d | 1886 | pgoff_t hugetlb_basepage_index(struct page *page) |
13d60f4b ZY |
1887 | { |
1888 | struct page *page_head = compound_head(page); | |
1889 | pgoff_t index = page_index(page_head); | |
1890 | unsigned long compound_idx; | |
1891 | ||
13d60f4b ZY |
1892 | if (compound_order(page_head) >= MAX_ORDER) |
1893 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1894 | else | |
1895 | compound_idx = page - page_head; | |
1896 | ||
1897 | return (index << compound_order(page_head)) + compound_idx; | |
1898 | } | |
1899 | ||
0c397dae | 1900 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1901 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1902 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1903 | { |
af0fb9df | 1904 | int order = huge_page_order(h); |
1da177e4 | 1905 | struct page *page; |
f60858f9 | 1906 | bool alloc_try_hard = true; |
f96efd58 | 1907 | |
f60858f9 MK |
1908 | /* |
1909 | * By default we always try hard to allocate the page with | |
1910 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1911 | * a loop (to adjust global huge page counts) and previous allocation | |
1912 | * failed, do not continue to try hard on the same node. Use the | |
1913 | * node_alloc_noretry bitmap to manage this state information. | |
1914 | */ | |
1915 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1916 | alloc_try_hard = false; | |
1917 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1918 | if (alloc_try_hard) | |
1919 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1920 | if (nid == NUMA_NO_NODE) |
1921 | nid = numa_mem_id(); | |
84172f4b | 1922 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
af0fb9df MH |
1923 | if (page) |
1924 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1925 | else | |
1926 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1927 | |
f60858f9 MK |
1928 | /* |
1929 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1930 | * indicates an overall state change. Clear bit so that we resume | |
1931 | * normal 'try hard' allocations. | |
1932 | */ | |
1933 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1934 | node_clear(nid, *node_alloc_noretry); | |
1935 | ||
1936 | /* | |
1937 | * If we tried hard to get a page but failed, set bit so that | |
1938 | * subsequent attempts will not try as hard until there is an | |
1939 | * overall state change. | |
1940 | */ | |
1941 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1942 | node_set(nid, *node_alloc_noretry); | |
1943 | ||
63b4613c NA |
1944 | return page; |
1945 | } | |
1946 | ||
0c397dae MH |
1947 | /* |
1948 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1949 | * should use this function to get new hugetlb pages | |
1950 | */ | |
1951 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1952 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1953 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1954 | { |
1955 | struct page *page; | |
7118fc29 | 1956 | bool retry = false; |
0c397dae | 1957 | |
7118fc29 | 1958 | retry: |
0c397dae MH |
1959 | if (hstate_is_gigantic(h)) |
1960 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1961 | else | |
1962 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1963 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1964 | if (!page) |
1965 | return NULL; | |
1966 | ||
7118fc29 MK |
1967 | if (hstate_is_gigantic(h)) { |
1968 | if (!prep_compound_gigantic_page(page, huge_page_order(h))) { | |
1969 | /* | |
1970 | * Rare failure to convert pages to compound page. | |
1971 | * Free pages and try again - ONCE! | |
1972 | */ | |
1973 | free_gigantic_page(page, huge_page_order(h)); | |
1974 | if (!retry) { | |
1975 | retry = true; | |
1976 | goto retry; | |
1977 | } | |
7118fc29 MK |
1978 | return NULL; |
1979 | } | |
1980 | } | |
0c397dae MH |
1981 | prep_new_huge_page(h, page, page_to_nid(page)); |
1982 | ||
1983 | return page; | |
1984 | } | |
1985 | ||
af0fb9df MH |
1986 | /* |
1987 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1988 | * manner. | |
1989 | */ | |
f60858f9 MK |
1990 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1991 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1992 | { |
1993 | struct page *page; | |
1994 | int nr_nodes, node; | |
af0fb9df | 1995 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1996 | |
1997 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1998 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1999 | node_alloc_noretry); | |
af0fb9df | 2000 | if (page) |
b2261026 | 2001 | break; |
b2261026 JK |
2002 | } |
2003 | ||
af0fb9df MH |
2004 | if (!page) |
2005 | return 0; | |
b2261026 | 2006 | |
af0fb9df MH |
2007 | put_page(page); /* free it into the hugepage allocator */ |
2008 | ||
2009 | return 1; | |
b2261026 JK |
2010 | } |
2011 | ||
e8c5c824 | 2012 | /* |
10c6ec49 MK |
2013 | * Remove huge page from pool from next node to free. Attempt to keep |
2014 | * persistent huge pages more or less balanced over allowed nodes. | |
2015 | * This routine only 'removes' the hugetlb page. The caller must make | |
2016 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
2017 | * Called with hugetlb_lock locked. |
2018 | */ | |
10c6ec49 MK |
2019 | static struct page *remove_pool_huge_page(struct hstate *h, |
2020 | nodemask_t *nodes_allowed, | |
2021 | bool acct_surplus) | |
e8c5c824 | 2022 | { |
b2261026 | 2023 | int nr_nodes, node; |
10c6ec49 | 2024 | struct page *page = NULL; |
e8c5c824 | 2025 | |
9487ca60 | 2026 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 2027 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
2028 | /* |
2029 | * If we're returning unused surplus pages, only examine | |
2030 | * nodes with surplus pages. | |
2031 | */ | |
b2261026 JK |
2032 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
2033 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 2034 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 2035 | struct page, lru); |
6eb4e88a | 2036 | remove_hugetlb_page(h, page, acct_surplus); |
9a76db09 | 2037 | break; |
e8c5c824 | 2038 | } |
b2261026 | 2039 | } |
e8c5c824 | 2040 | |
10c6ec49 | 2041 | return page; |
e8c5c824 LS |
2042 | } |
2043 | ||
c8721bbb NH |
2044 | /* |
2045 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
2046 | * nothing for in-use hugepages and non-hugepages. |
2047 | * This function returns values like below: | |
2048 | * | |
ad2fa371 MS |
2049 | * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages |
2050 | * when the system is under memory pressure and the feature of | |
2051 | * freeing unused vmemmap pages associated with each hugetlb page | |
2052 | * is enabled. | |
2053 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
2054 | * (allocated or reserved.) | |
2055 | * 0: successfully dissolved free hugepages or the page is not a | |
2056 | * hugepage (considered as already dissolved) | |
c8721bbb | 2057 | */ |
c3114a84 | 2058 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 2059 | { |
6bc9b564 | 2060 | int rc = -EBUSY; |
082d5b6b | 2061 | |
7ffddd49 | 2062 | retry: |
faf53def NH |
2063 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
2064 | if (!PageHuge(page)) | |
2065 | return 0; | |
2066 | ||
db71ef79 | 2067 | spin_lock_irq(&hugetlb_lock); |
faf53def NH |
2068 | if (!PageHuge(page)) { |
2069 | rc = 0; | |
2070 | goto out; | |
2071 | } | |
2072 | ||
2073 | if (!page_count(page)) { | |
2247bb33 GS |
2074 | struct page *head = compound_head(page); |
2075 | struct hstate *h = page_hstate(head); | |
6bc9b564 | 2076 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 2077 | goto out; |
7ffddd49 MS |
2078 | |
2079 | /* | |
2080 | * We should make sure that the page is already on the free list | |
2081 | * when it is dissolved. | |
2082 | */ | |
6c037149 | 2083 | if (unlikely(!HPageFreed(head))) { |
db71ef79 | 2084 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
2085 | cond_resched(); |
2086 | ||
2087 | /* | |
2088 | * Theoretically, we should return -EBUSY when we | |
2089 | * encounter this race. In fact, we have a chance | |
2090 | * to successfully dissolve the page if we do a | |
2091 | * retry. Because the race window is quite small. | |
2092 | * If we seize this opportunity, it is an optimization | |
2093 | * for increasing the success rate of dissolving page. | |
2094 | */ | |
2095 | goto retry; | |
2096 | } | |
2097 | ||
0c5da357 | 2098 | remove_hugetlb_page(h, head, false); |
c1470b33 | 2099 | h->max_huge_pages--; |
db71ef79 | 2100 | spin_unlock_irq(&hugetlb_lock); |
ad2fa371 MS |
2101 | |
2102 | /* | |
2103 | * Normally update_and_free_page will allocate required vmemmmap | |
2104 | * before freeing the page. update_and_free_page will fail to | |
2105 | * free the page if it can not allocate required vmemmap. We | |
2106 | * need to adjust max_huge_pages if the page is not freed. | |
2107 | * Attempt to allocate vmemmmap here so that we can take | |
2108 | * appropriate action on failure. | |
2109 | */ | |
5981611d | 2110 | rc = hugetlb_vmemmap_alloc(h, head); |
ad2fa371 MS |
2111 | if (!rc) { |
2112 | /* | |
2113 | * Move PageHWPoison flag from head page to the raw | |
2114 | * error page, which makes any subpages rather than | |
2115 | * the error page reusable. | |
2116 | */ | |
2117 | if (PageHWPoison(head) && page != head) { | |
2118 | SetPageHWPoison(page); | |
2119 | ClearPageHWPoison(head); | |
2120 | } | |
2121 | update_and_free_page(h, head, false); | |
2122 | } else { | |
2123 | spin_lock_irq(&hugetlb_lock); | |
2124 | add_hugetlb_page(h, head, false); | |
2125 | h->max_huge_pages++; | |
2126 | spin_unlock_irq(&hugetlb_lock); | |
2127 | } | |
2128 | ||
2129 | return rc; | |
c8721bbb | 2130 | } |
082d5b6b | 2131 | out: |
db71ef79 | 2132 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 2133 | return rc; |
c8721bbb NH |
2134 | } |
2135 | ||
2136 | /* | |
2137 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
2138 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
2139 | * Note that this will dissolve a free gigantic hugepage completely, if any |
2140 | * part of it lies within the given range. | |
082d5b6b GS |
2141 | * Also note that if dissolve_free_huge_page() returns with an error, all |
2142 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 2143 | */ |
082d5b6b | 2144 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 2145 | { |
c8721bbb | 2146 | unsigned long pfn; |
eb03aa00 | 2147 | struct page *page; |
082d5b6b | 2148 | int rc = 0; |
dc2628f3 MS |
2149 | unsigned int order; |
2150 | struct hstate *h; | |
c8721bbb | 2151 | |
d0177639 | 2152 | if (!hugepages_supported()) |
082d5b6b | 2153 | return rc; |
d0177639 | 2154 | |
dc2628f3 MS |
2155 | order = huge_page_order(&default_hstate); |
2156 | for_each_hstate(h) | |
2157 | order = min(order, huge_page_order(h)); | |
2158 | ||
2159 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { | |
eb03aa00 | 2160 | page = pfn_to_page(pfn); |
faf53def NH |
2161 | rc = dissolve_free_huge_page(page); |
2162 | if (rc) | |
2163 | break; | |
eb03aa00 | 2164 | } |
082d5b6b GS |
2165 | |
2166 | return rc; | |
c8721bbb NH |
2167 | } |
2168 | ||
ab5ac90a MH |
2169 | /* |
2170 | * Allocates a fresh surplus page from the page allocator. | |
2171 | */ | |
0c397dae | 2172 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
b65a4eda | 2173 | int nid, nodemask_t *nmask, bool zero_ref) |
7893d1d5 | 2174 | { |
9980d744 | 2175 | struct page *page = NULL; |
b65a4eda | 2176 | bool retry = false; |
7893d1d5 | 2177 | |
bae7f4ae | 2178 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2179 | return NULL; |
2180 | ||
db71ef79 | 2181 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2182 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
2183 | goto out_unlock; | |
db71ef79 | 2184 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 2185 | |
b65a4eda | 2186 | retry: |
f60858f9 | 2187 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 2188 | if (!page) |
0c397dae | 2189 | return NULL; |
d1c3fb1f | 2190 | |
db71ef79 | 2191 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2192 | /* |
2193 | * We could have raced with the pool size change. | |
2194 | * Double check that and simply deallocate the new page | |
2195 | * if we would end up overcommiting the surpluses. Abuse | |
2196 | * temporary page to workaround the nasty free_huge_page | |
2197 | * codeflow | |
2198 | */ | |
2199 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 2200 | SetHPageTemporary(page); |
db71ef79 | 2201 | spin_unlock_irq(&hugetlb_lock); |
9980d744 | 2202 | put_page(page); |
2bf753e6 | 2203 | return NULL; |
7893d1d5 | 2204 | } |
9980d744 | 2205 | |
b65a4eda MK |
2206 | if (zero_ref) { |
2207 | /* | |
2208 | * Caller requires a page with zero ref count. | |
2209 | * We will drop ref count here. If someone else is holding | |
2210 | * a ref, the page will be freed when they drop it. Abuse | |
2211 | * temporary page flag to accomplish this. | |
2212 | */ | |
2213 | SetHPageTemporary(page); | |
2214 | if (!put_page_testzero(page)) { | |
2215 | /* | |
2216 | * Unexpected inflated ref count on freshly allocated | |
2217 | * huge. Retry once. | |
2218 | */ | |
2219 | pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); | |
2220 | spin_unlock_irq(&hugetlb_lock); | |
2221 | if (retry) | |
2222 | return NULL; | |
2223 | ||
2224 | retry = true; | |
2225 | goto retry; | |
2226 | } | |
2227 | ClearHPageTemporary(page); | |
2228 | } | |
2229 | ||
2230 | h->surplus_huge_pages++; | |
2231 | h->surplus_huge_pages_node[page_to_nid(page)]++; | |
2232 | ||
9980d744 | 2233 | out_unlock: |
db71ef79 | 2234 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 AL |
2235 | |
2236 | return page; | |
2237 | } | |
2238 | ||
bbe88753 | 2239 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 2240 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
2241 | { |
2242 | struct page *page; | |
2243 | ||
2244 | if (hstate_is_gigantic(h)) | |
2245 | return NULL; | |
2246 | ||
f60858f9 | 2247 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
2248 | if (!page) |
2249 | return NULL; | |
2250 | ||
2251 | /* | |
2252 | * We do not account these pages as surplus because they are only | |
2253 | * temporary and will be released properly on the last reference | |
2254 | */ | |
9157c311 | 2255 | SetHPageTemporary(page); |
ab5ac90a MH |
2256 | |
2257 | return page; | |
2258 | } | |
2259 | ||
099730d6 DH |
2260 | /* |
2261 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
2262 | */ | |
e0ec90ee | 2263 | static |
0c397dae | 2264 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
2265 | struct vm_area_struct *vma, unsigned long addr) |
2266 | { | |
cfcaa66f | 2267 | struct page *page = NULL; |
aaf14e40 MH |
2268 | struct mempolicy *mpol; |
2269 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
2270 | int nid; | |
2271 | nodemask_t *nodemask; | |
2272 | ||
2273 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
2274 | if (mpol_is_preferred_many(mpol)) { |
2275 | gfp_t gfp = gfp_mask | __GFP_NOWARN; | |
2276 | ||
2277 | gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); | |
2278 | page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); | |
aaf14e40 | 2279 | |
cfcaa66f BW |
2280 | /* Fallback to all nodes if page==NULL */ |
2281 | nodemask = NULL; | |
2282 | } | |
2283 | ||
2284 | if (!page) | |
2285 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); | |
2286 | mpol_cond_put(mpol); | |
aaf14e40 | 2287 | return page; |
099730d6 DH |
2288 | } |
2289 | ||
ab5ac90a | 2290 | /* page migration callback function */ |
3e59fcb0 | 2291 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 2292 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 2293 | { |
db71ef79 | 2294 | spin_lock_irq(&hugetlb_lock); |
4db9b2ef | 2295 | if (h->free_huge_pages - h->resv_huge_pages > 0) { |
3e59fcb0 MH |
2296 | struct page *page; |
2297 | ||
2298 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
2299 | if (page) { | |
db71ef79 | 2300 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 2301 | return page; |
4db9b2ef MH |
2302 | } |
2303 | } | |
db71ef79 | 2304 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 2305 | |
0c397dae | 2306 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2307 | } |
2308 | ||
ebd63723 | 2309 | /* mempolicy aware migration callback */ |
389c8178 MH |
2310 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2311 | unsigned long address) | |
ebd63723 MH |
2312 | { |
2313 | struct mempolicy *mpol; | |
2314 | nodemask_t *nodemask; | |
2315 | struct page *page; | |
ebd63723 MH |
2316 | gfp_t gfp_mask; |
2317 | int node; | |
2318 | ||
ebd63723 MH |
2319 | gfp_mask = htlb_alloc_mask(h); |
2320 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 2321 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
2322 | mpol_cond_put(mpol); |
2323 | ||
2324 | return page; | |
2325 | } | |
2326 | ||
e4e574b7 | 2327 | /* |
25985edc | 2328 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2329 | * of size 'delta'. |
2330 | */ | |
0a4f3d1b | 2331 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 2332 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
2333 | { |
2334 | struct list_head surplus_list; | |
2335 | struct page *page, *tmp; | |
0a4f3d1b LX |
2336 | int ret; |
2337 | long i; | |
2338 | long needed, allocated; | |
28073b02 | 2339 | bool alloc_ok = true; |
e4e574b7 | 2340 | |
9487ca60 | 2341 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 2342 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2343 | if (needed <= 0) { |
a5516438 | 2344 | h->resv_huge_pages += delta; |
e4e574b7 | 2345 | return 0; |
ac09b3a1 | 2346 | } |
e4e574b7 AL |
2347 | |
2348 | allocated = 0; | |
2349 | INIT_LIST_HEAD(&surplus_list); | |
2350 | ||
2351 | ret = -ENOMEM; | |
2352 | retry: | |
db71ef79 | 2353 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 2354 | for (i = 0; i < needed; i++) { |
0c397dae | 2355 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
b65a4eda | 2356 | NUMA_NO_NODE, NULL, true); |
28073b02 HD |
2357 | if (!page) { |
2358 | alloc_ok = false; | |
2359 | break; | |
2360 | } | |
e4e574b7 | 2361 | list_add(&page->lru, &surplus_list); |
69ed779a | 2362 | cond_resched(); |
e4e574b7 | 2363 | } |
28073b02 | 2364 | allocated += i; |
e4e574b7 AL |
2365 | |
2366 | /* | |
2367 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2368 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2369 | */ | |
db71ef79 | 2370 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
2371 | needed = (h->resv_huge_pages + delta) - |
2372 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2373 | if (needed > 0) { |
2374 | if (alloc_ok) | |
2375 | goto retry; | |
2376 | /* | |
2377 | * We were not able to allocate enough pages to | |
2378 | * satisfy the entire reservation so we free what | |
2379 | * we've allocated so far. | |
2380 | */ | |
2381 | goto free; | |
2382 | } | |
e4e574b7 AL |
2383 | /* |
2384 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2385 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2386 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2387 | * allocator. Commit the entire reservation here to prevent another |
2388 | * process from stealing the pages as they are added to the pool but | |
2389 | * before they are reserved. | |
e4e574b7 AL |
2390 | */ |
2391 | needed += allocated; | |
a5516438 | 2392 | h->resv_huge_pages += delta; |
e4e574b7 | 2393 | ret = 0; |
a9869b83 | 2394 | |
19fc3f0a | 2395 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2396 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2397 | if ((--needed) < 0) |
2398 | break; | |
b65a4eda | 2399 | /* Add the page to the hugetlb allocator */ |
a5516438 | 2400 | enqueue_huge_page(h, page); |
19fc3f0a | 2401 | } |
28073b02 | 2402 | free: |
db71ef79 | 2403 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a | 2404 | |
b65a4eda MK |
2405 | /* |
2406 | * Free unnecessary surplus pages to the buddy allocator. | |
2407 | * Pages have no ref count, call free_huge_page directly. | |
2408 | */ | |
c0d934ba | 2409 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
b65a4eda | 2410 | free_huge_page(page); |
db71ef79 | 2411 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2412 | |
2413 | return ret; | |
2414 | } | |
2415 | ||
2416 | /* | |
e5bbc8a6 MK |
2417 | * This routine has two main purposes: |
2418 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2419 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2420 | * to the associated reservation map. | |
2421 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2422 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2423 | */ |
a5516438 AK |
2424 | static void return_unused_surplus_pages(struct hstate *h, |
2425 | unsigned long unused_resv_pages) | |
e4e574b7 | 2426 | { |
e4e574b7 | 2427 | unsigned long nr_pages; |
10c6ec49 MK |
2428 | struct page *page; |
2429 | LIST_HEAD(page_list); | |
2430 | ||
9487ca60 | 2431 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2432 | /* Uncommit the reservation */ |
2433 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2434 | |
aa888a74 | 2435 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2436 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2437 | goto out; |
aa888a74 | 2438 | |
e5bbc8a6 MK |
2439 | /* |
2440 | * Part (or even all) of the reservation could have been backed | |
2441 | * by pre-allocated pages. Only free surplus pages. | |
2442 | */ | |
a5516438 | 2443 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2444 | |
685f3457 LS |
2445 | /* |
2446 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2447 | * evenly across all nodes with memory. Iterate across these nodes |
2448 | * until we can no longer free unreserved surplus pages. This occurs | |
2449 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2450 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2451 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2452 | */ |
2453 | while (nr_pages--) { | |
10c6ec49 MK |
2454 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2455 | if (!page) | |
e5bbc8a6 | 2456 | goto out; |
10c6ec49 MK |
2457 | |
2458 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2459 | } |
e5bbc8a6 MK |
2460 | |
2461 | out: | |
db71ef79 | 2462 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2463 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2464 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2465 | } |
2466 | ||
5e911373 | 2467 | |
c37f9fb1 | 2468 | /* |
feba16e2 | 2469 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2470 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2471 | * |
2472 | * vma_needs_reservation is called to determine if the huge page at addr | |
2473 | * within the vma has an associated reservation. If a reservation is | |
2474 | * needed, the value 1 is returned. The caller is then responsible for | |
2475 | * managing the global reservation and subpool usage counts. After | |
2476 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2477 | * to add the page to the reservation map. If the page allocation fails, |
2478 | * the reservation must be ended instead of committed. vma_end_reservation | |
2479 | * is called in such cases. | |
cf3ad20b MK |
2480 | * |
2481 | * In the normal case, vma_commit_reservation returns the same value | |
2482 | * as the preceding vma_needs_reservation call. The only time this | |
2483 | * is not the case is if a reserve map was changed between calls. It | |
2484 | * is the responsibility of the caller to notice the difference and | |
2485 | * take appropriate action. | |
96b96a96 MK |
2486 | * |
2487 | * vma_add_reservation is used in error paths where a reservation must | |
2488 | * be restored when a newly allocated huge page must be freed. It is | |
2489 | * to be called after calling vma_needs_reservation to determine if a | |
2490 | * reservation exists. | |
846be085 MK |
2491 | * |
2492 | * vma_del_reservation is used in error paths where an entry in the reserve | |
2493 | * map was created during huge page allocation and must be removed. It is to | |
2494 | * be called after calling vma_needs_reservation to determine if a reservation | |
2495 | * exists. | |
c37f9fb1 | 2496 | */ |
5e911373 MK |
2497 | enum vma_resv_mode { |
2498 | VMA_NEEDS_RESV, | |
2499 | VMA_COMMIT_RESV, | |
feba16e2 | 2500 | VMA_END_RESV, |
96b96a96 | 2501 | VMA_ADD_RESV, |
846be085 | 2502 | VMA_DEL_RESV, |
5e911373 | 2503 | }; |
cf3ad20b MK |
2504 | static long __vma_reservation_common(struct hstate *h, |
2505 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2506 | enum vma_resv_mode mode) |
c37f9fb1 | 2507 | { |
4e35f483 JK |
2508 | struct resv_map *resv; |
2509 | pgoff_t idx; | |
cf3ad20b | 2510 | long ret; |
0db9d74e | 2511 | long dummy_out_regions_needed; |
c37f9fb1 | 2512 | |
4e35f483 JK |
2513 | resv = vma_resv_map(vma); |
2514 | if (!resv) | |
84afd99b | 2515 | return 1; |
c37f9fb1 | 2516 | |
4e35f483 | 2517 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2518 | switch (mode) { |
2519 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2520 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2521 | /* We assume that vma_reservation_* routines always operate on | |
2522 | * 1 page, and that adding to resv map a 1 page entry can only | |
2523 | * ever require 1 region. | |
2524 | */ | |
2525 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2526 | break; |
2527 | case VMA_COMMIT_RESV: | |
075a61d0 | 2528 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2529 | /* region_add calls of range 1 should never fail. */ |
2530 | VM_BUG_ON(ret < 0); | |
5e911373 | 2531 | break; |
feba16e2 | 2532 | case VMA_END_RESV: |
0db9d74e | 2533 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2534 | ret = 0; |
2535 | break; | |
96b96a96 | 2536 | case VMA_ADD_RESV: |
0db9d74e | 2537 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2538 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2539 | /* region_add calls of range 1 should never fail. */ |
2540 | VM_BUG_ON(ret < 0); | |
2541 | } else { | |
2542 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2543 | ret = region_del(resv, idx, idx + 1); |
2544 | } | |
2545 | break; | |
846be085 MK |
2546 | case VMA_DEL_RESV: |
2547 | if (vma->vm_flags & VM_MAYSHARE) { | |
2548 | region_abort(resv, idx, idx + 1, 1); | |
2549 | ret = region_del(resv, idx, idx + 1); | |
2550 | } else { | |
2551 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2552 | /* region_add calls of range 1 should never fail. */ | |
2553 | VM_BUG_ON(ret < 0); | |
2554 | } | |
2555 | break; | |
5e911373 MK |
2556 | default: |
2557 | BUG(); | |
2558 | } | |
84afd99b | 2559 | |
846be085 | 2560 | if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) |
cf3ad20b | 2561 | return ret; |
bf3d12b9 ML |
2562 | /* |
2563 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2564 | * | |
2565 | * In most cases, reserves always exist for private mappings. | |
2566 | * However, a file associated with mapping could have been | |
2567 | * hole punched or truncated after reserves were consumed. | |
2568 | * As subsequent fault on such a range will not use reserves. | |
2569 | * Subtle - The reserve map for private mappings has the | |
2570 | * opposite meaning than that of shared mappings. If NO | |
2571 | * entry is in the reserve map, it means a reservation exists. | |
2572 | * If an entry exists in the reserve map, it means the | |
2573 | * reservation has already been consumed. As a result, the | |
2574 | * return value of this routine is the opposite of the | |
2575 | * value returned from reserve map manipulation routines above. | |
2576 | */ | |
2577 | if (ret > 0) | |
2578 | return 0; | |
2579 | if (ret == 0) | |
2580 | return 1; | |
2581 | return ret; | |
c37f9fb1 | 2582 | } |
cf3ad20b MK |
2583 | |
2584 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2585 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2586 | { |
5e911373 | 2587 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2588 | } |
84afd99b | 2589 | |
cf3ad20b MK |
2590 | static long vma_commit_reservation(struct hstate *h, |
2591 | struct vm_area_struct *vma, unsigned long addr) | |
2592 | { | |
5e911373 MK |
2593 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2594 | } | |
2595 | ||
feba16e2 | 2596 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2597 | struct vm_area_struct *vma, unsigned long addr) |
2598 | { | |
feba16e2 | 2599 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2600 | } |
2601 | ||
96b96a96 MK |
2602 | static long vma_add_reservation(struct hstate *h, |
2603 | struct vm_area_struct *vma, unsigned long addr) | |
2604 | { | |
2605 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2606 | } | |
2607 | ||
846be085 MK |
2608 | static long vma_del_reservation(struct hstate *h, |
2609 | struct vm_area_struct *vma, unsigned long addr) | |
2610 | { | |
2611 | return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); | |
2612 | } | |
2613 | ||
96b96a96 | 2614 | /* |
846be085 MK |
2615 | * This routine is called to restore reservation information on error paths. |
2616 | * It should ONLY be called for pages allocated via alloc_huge_page(), and | |
2617 | * the hugetlb mutex should remain held when calling this routine. | |
2618 | * | |
2619 | * It handles two specific cases: | |
2620 | * 1) A reservation was in place and the page consumed the reservation. | |
2621 | * HPageRestoreReserve is set in the page. | |
2622 | * 2) No reservation was in place for the page, so HPageRestoreReserve is | |
2623 | * not set. However, alloc_huge_page always updates the reserve map. | |
2624 | * | |
2625 | * In case 1, free_huge_page later in the error path will increment the | |
2626 | * global reserve count. But, free_huge_page does not have enough context | |
2627 | * to adjust the reservation map. This case deals primarily with private | |
2628 | * mappings. Adjust the reserve map here to be consistent with global | |
2629 | * reserve count adjustments to be made by free_huge_page. Make sure the | |
2630 | * reserve map indicates there is a reservation present. | |
2631 | * | |
2632 | * In case 2, simply undo reserve map modifications done by alloc_huge_page. | |
96b96a96 | 2633 | */ |
846be085 MK |
2634 | void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, |
2635 | unsigned long address, struct page *page) | |
96b96a96 | 2636 | { |
846be085 | 2637 | long rc = vma_needs_reservation(h, vma, address); |
96b96a96 | 2638 | |
846be085 MK |
2639 | if (HPageRestoreReserve(page)) { |
2640 | if (unlikely(rc < 0)) | |
96b96a96 MK |
2641 | /* |
2642 | * Rare out of memory condition in reserve map | |
d6995da3 | 2643 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2644 | * global reserve count will not be incremented |
2645 | * by free_huge_page. This will make it appear | |
2646 | * as though the reservation for this page was | |
2647 | * consumed. This may prevent the task from | |
2648 | * faulting in the page at a later time. This | |
2649 | * is better than inconsistent global huge page | |
2650 | * accounting of reserve counts. | |
2651 | */ | |
d6995da3 | 2652 | ClearHPageRestoreReserve(page); |
846be085 MK |
2653 | else if (rc) |
2654 | (void)vma_add_reservation(h, vma, address); | |
2655 | else | |
2656 | vma_end_reservation(h, vma, address); | |
2657 | } else { | |
2658 | if (!rc) { | |
2659 | /* | |
2660 | * This indicates there is an entry in the reserve map | |
c7b1850d | 2661 | * not added by alloc_huge_page. We know it was added |
846be085 MK |
2662 | * before the alloc_huge_page call, otherwise |
2663 | * HPageRestoreReserve would be set on the page. | |
2664 | * Remove the entry so that a subsequent allocation | |
2665 | * does not consume a reservation. | |
2666 | */ | |
2667 | rc = vma_del_reservation(h, vma, address); | |
2668 | if (rc < 0) | |
96b96a96 | 2669 | /* |
846be085 MK |
2670 | * VERY rare out of memory condition. Since |
2671 | * we can not delete the entry, set | |
2672 | * HPageRestoreReserve so that the reserve | |
2673 | * count will be incremented when the page | |
2674 | * is freed. This reserve will be consumed | |
2675 | * on a subsequent allocation. | |
96b96a96 | 2676 | */ |
846be085 MK |
2677 | SetHPageRestoreReserve(page); |
2678 | } else if (rc < 0) { | |
2679 | /* | |
2680 | * Rare out of memory condition from | |
2681 | * vma_needs_reservation call. Memory allocation is | |
2682 | * only attempted if a new entry is needed. Therefore, | |
2683 | * this implies there is not an entry in the | |
2684 | * reserve map. | |
2685 | * | |
2686 | * For shared mappings, no entry in the map indicates | |
2687 | * no reservation. We are done. | |
2688 | */ | |
2689 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
2690 | /* | |
2691 | * For private mappings, no entry indicates | |
2692 | * a reservation is present. Since we can | |
2693 | * not add an entry, set SetHPageRestoreReserve | |
2694 | * on the page so reserve count will be | |
2695 | * incremented when freed. This reserve will | |
2696 | * be consumed on a subsequent allocation. | |
2697 | */ | |
2698 | SetHPageRestoreReserve(page); | |
96b96a96 | 2699 | } else |
846be085 MK |
2700 | /* |
2701 | * No reservation present, do nothing | |
2702 | */ | |
2703 | vma_end_reservation(h, vma, address); | |
96b96a96 MK |
2704 | } |
2705 | } | |
2706 | ||
369fa227 OS |
2707 | /* |
2708 | * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one | |
2709 | * @h: struct hstate old page belongs to | |
2710 | * @old_page: Old page to dissolve | |
ae37c7ff | 2711 | * @list: List to isolate the page in case we need to |
369fa227 OS |
2712 | * Returns 0 on success, otherwise negated error. |
2713 | */ | |
ae37c7ff OS |
2714 | static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, |
2715 | struct list_head *list) | |
369fa227 OS |
2716 | { |
2717 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
2718 | int nid = page_to_nid(old_page); | |
b65a4eda | 2719 | bool alloc_retry = false; |
369fa227 OS |
2720 | struct page *new_page; |
2721 | int ret = 0; | |
2722 | ||
2723 | /* | |
2724 | * Before dissolving the page, we need to allocate a new one for the | |
f41f2ed4 MS |
2725 | * pool to remain stable. Here, we allocate the page and 'prep' it |
2726 | * by doing everything but actually updating counters and adding to | |
2727 | * the pool. This simplifies and let us do most of the processing | |
2728 | * under the lock. | |
369fa227 | 2729 | */ |
b65a4eda | 2730 | alloc_retry: |
369fa227 OS |
2731 | new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); |
2732 | if (!new_page) | |
2733 | return -ENOMEM; | |
b65a4eda MK |
2734 | /* |
2735 | * If all goes well, this page will be directly added to the free | |
2736 | * list in the pool. For this the ref count needs to be zero. | |
2737 | * Attempt to drop now, and retry once if needed. It is VERY | |
2738 | * unlikely there is another ref on the page. | |
2739 | * | |
2740 | * If someone else has a reference to the page, it will be freed | |
2741 | * when they drop their ref. Abuse temporary page flag to accomplish | |
2742 | * this. Retry once if there is an inflated ref count. | |
2743 | */ | |
2744 | SetHPageTemporary(new_page); | |
2745 | if (!put_page_testzero(new_page)) { | |
2746 | if (alloc_retry) | |
2747 | return -EBUSY; | |
2748 | ||
2749 | alloc_retry = true; | |
2750 | goto alloc_retry; | |
2751 | } | |
2752 | ClearHPageTemporary(new_page); | |
2753 | ||
f41f2ed4 | 2754 | __prep_new_huge_page(h, new_page); |
369fa227 OS |
2755 | |
2756 | retry: | |
2757 | spin_lock_irq(&hugetlb_lock); | |
2758 | if (!PageHuge(old_page)) { | |
2759 | /* | |
2760 | * Freed from under us. Drop new_page too. | |
2761 | */ | |
2762 | goto free_new; | |
2763 | } else if (page_count(old_page)) { | |
2764 | /* | |
ae37c7ff OS |
2765 | * Someone has grabbed the page, try to isolate it here. |
2766 | * Fail with -EBUSY if not possible. | |
369fa227 | 2767 | */ |
ae37c7ff | 2768 | spin_unlock_irq(&hugetlb_lock); |
7ce82f4c | 2769 | ret = isolate_hugetlb(old_page, list); |
ae37c7ff | 2770 | spin_lock_irq(&hugetlb_lock); |
369fa227 OS |
2771 | goto free_new; |
2772 | } else if (!HPageFreed(old_page)) { | |
2773 | /* | |
2774 | * Page's refcount is 0 but it has not been enqueued in the | |
2775 | * freelist yet. Race window is small, so we can succeed here if | |
2776 | * we retry. | |
2777 | */ | |
2778 | spin_unlock_irq(&hugetlb_lock); | |
2779 | cond_resched(); | |
2780 | goto retry; | |
2781 | } else { | |
2782 | /* | |
2783 | * Ok, old_page is still a genuine free hugepage. Remove it from | |
2784 | * the freelist and decrease the counters. These will be | |
2785 | * incremented again when calling __prep_account_new_huge_page() | |
2786 | * and enqueue_huge_page() for new_page. The counters will remain | |
2787 | * stable since this happens under the lock. | |
2788 | */ | |
2789 | remove_hugetlb_page(h, old_page, false); | |
2790 | ||
2791 | /* | |
b65a4eda MK |
2792 | * Ref count on new page is already zero as it was dropped |
2793 | * earlier. It can be directly added to the pool free list. | |
369fa227 | 2794 | */ |
369fa227 | 2795 | __prep_account_new_huge_page(h, nid); |
369fa227 OS |
2796 | enqueue_huge_page(h, new_page); |
2797 | ||
2798 | /* | |
2799 | * Pages have been replaced, we can safely free the old one. | |
2800 | */ | |
2801 | spin_unlock_irq(&hugetlb_lock); | |
b65d4adb | 2802 | update_and_free_page(h, old_page, false); |
369fa227 OS |
2803 | } |
2804 | ||
2805 | return ret; | |
2806 | ||
2807 | free_new: | |
2808 | spin_unlock_irq(&hugetlb_lock); | |
b65a4eda MK |
2809 | /* Page has a zero ref count, but needs a ref to be freed */ |
2810 | set_page_refcounted(new_page); | |
b65d4adb | 2811 | update_and_free_page(h, new_page, false); |
369fa227 OS |
2812 | |
2813 | return ret; | |
2814 | } | |
2815 | ||
ae37c7ff | 2816 | int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) |
369fa227 OS |
2817 | { |
2818 | struct hstate *h; | |
2819 | struct page *head; | |
ae37c7ff | 2820 | int ret = -EBUSY; |
369fa227 OS |
2821 | |
2822 | /* | |
2823 | * The page might have been dissolved from under our feet, so make sure | |
2824 | * to carefully check the state under the lock. | |
2825 | * Return success when racing as if we dissolved the page ourselves. | |
2826 | */ | |
2827 | spin_lock_irq(&hugetlb_lock); | |
2828 | if (PageHuge(page)) { | |
2829 | head = compound_head(page); | |
2830 | h = page_hstate(head); | |
2831 | } else { | |
2832 | spin_unlock_irq(&hugetlb_lock); | |
2833 | return 0; | |
2834 | } | |
2835 | spin_unlock_irq(&hugetlb_lock); | |
2836 | ||
2837 | /* | |
2838 | * Fence off gigantic pages as there is a cyclic dependency between | |
2839 | * alloc_contig_range and them. Return -ENOMEM as this has the effect | |
2840 | * of bailing out right away without further retrying. | |
2841 | */ | |
2842 | if (hstate_is_gigantic(h)) | |
2843 | return -ENOMEM; | |
2844 | ||
7ce82f4c | 2845 | if (page_count(head) && !isolate_hugetlb(head, list)) |
ae37c7ff OS |
2846 | ret = 0; |
2847 | else if (!page_count(head)) | |
2848 | ret = alloc_and_dissolve_huge_page(h, head, list); | |
2849 | ||
2850 | return ret; | |
369fa227 OS |
2851 | } |
2852 | ||
70c3547e | 2853 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2854 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2855 | { |
90481622 | 2856 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2857 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2858 | struct page *page; |
d85f69b0 MK |
2859 | long map_chg, map_commit; |
2860 | long gbl_chg; | |
6d76dcf4 AK |
2861 | int ret, idx; |
2862 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2863 | bool deferred_reserve; |
a1e78772 | 2864 | |
6d76dcf4 | 2865 | idx = hstate_index(h); |
a1e78772 | 2866 | /* |
d85f69b0 MK |
2867 | * Examine the region/reserve map to determine if the process |
2868 | * has a reservation for the page to be allocated. A return | |
2869 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2870 | */ |
d85f69b0 MK |
2871 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2872 | if (map_chg < 0) | |
76dcee75 | 2873 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2874 | |
2875 | /* | |
2876 | * Processes that did not create the mapping will have no | |
2877 | * reserves as indicated by the region/reserve map. Check | |
2878 | * that the allocation will not exceed the subpool limit. | |
2879 | * Allocations for MAP_NORESERVE mappings also need to be | |
2880 | * checked against any subpool limit. | |
2881 | */ | |
2882 | if (map_chg || avoid_reserve) { | |
2883 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2884 | if (gbl_chg < 0) { | |
feba16e2 | 2885 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2886 | return ERR_PTR(-ENOSPC); |
5e911373 | 2887 | } |
1da177e4 | 2888 | |
d85f69b0 MK |
2889 | /* |
2890 | * Even though there was no reservation in the region/reserve | |
2891 | * map, there could be reservations associated with the | |
2892 | * subpool that can be used. This would be indicated if the | |
2893 | * return value of hugepage_subpool_get_pages() is zero. | |
2894 | * However, if avoid_reserve is specified we still avoid even | |
2895 | * the subpool reservations. | |
2896 | */ | |
2897 | if (avoid_reserve) | |
2898 | gbl_chg = 1; | |
2899 | } | |
2900 | ||
08cf9faf MA |
2901 | /* If this allocation is not consuming a reservation, charge it now. |
2902 | */ | |
6501fe5f | 2903 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2904 | if (deferred_reserve) { |
2905 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2906 | idx, pages_per_huge_page(h), &h_cg); | |
2907 | if (ret) | |
2908 | goto out_subpool_put; | |
2909 | } | |
2910 | ||
6d76dcf4 | 2911 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2912 | if (ret) |
08cf9faf | 2913 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2914 | |
db71ef79 | 2915 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
2916 | /* |
2917 | * glb_chg is passed to indicate whether or not a page must be taken | |
2918 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2919 | * a reservation exists for the allocation. | |
2920 | */ | |
2921 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2922 | if (!page) { |
db71ef79 | 2923 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 2924 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2925 | if (!page) |
2926 | goto out_uncharge_cgroup; | |
a88c7695 | 2927 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2928 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2929 | h->resv_huge_pages--; |
2930 | } | |
db71ef79 | 2931 | spin_lock_irq(&hugetlb_lock); |
15a8d68e | 2932 | list_add(&page->lru, &h->hugepage_activelist); |
81a6fcae | 2933 | /* Fall through */ |
68842c9b | 2934 | } |
81a6fcae | 2935 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2936 | /* If allocation is not consuming a reservation, also store the |
2937 | * hugetlb_cgroup pointer on the page. | |
2938 | */ | |
2939 | if (deferred_reserve) { | |
2940 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2941 | h_cg, page); | |
2942 | } | |
2943 | ||
db71ef79 | 2944 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 2945 | |
d6995da3 | 2946 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2947 | |
d85f69b0 MK |
2948 | map_commit = vma_commit_reservation(h, vma, addr); |
2949 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2950 | /* |
2951 | * The page was added to the reservation map between | |
2952 | * vma_needs_reservation and vma_commit_reservation. | |
2953 | * This indicates a race with hugetlb_reserve_pages. | |
2954 | * Adjust for the subpool count incremented above AND | |
2955 | * in hugetlb_reserve_pages for the same page. Also, | |
2956 | * the reservation count added in hugetlb_reserve_pages | |
2957 | * no longer applies. | |
2958 | */ | |
2959 | long rsv_adjust; | |
2960 | ||
2961 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2962 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2963 | if (deferred_reserve) |
2964 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2965 | pages_per_huge_page(h), page); | |
33039678 | 2966 | } |
90d8b7e6 | 2967 | return page; |
8f34af6f JZ |
2968 | |
2969 | out_uncharge_cgroup: | |
2970 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2971 | out_uncharge_cgroup_reservation: |
2972 | if (deferred_reserve) | |
2973 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2974 | h_cg); | |
8f34af6f | 2975 | out_subpool_put: |
d85f69b0 | 2976 | if (map_chg || avoid_reserve) |
8f34af6f | 2977 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2978 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2979 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2980 | } |
2981 | ||
b5389086 | 2982 | int alloc_bootmem_huge_page(struct hstate *h, int nid) |
e24a1307 | 2983 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); |
b5389086 | 2984 | int __alloc_bootmem_huge_page(struct hstate *h, int nid) |
aa888a74 | 2985 | { |
b5389086 | 2986 | struct huge_bootmem_page *m = NULL; /* initialize for clang */ |
b2261026 | 2987 | int nr_nodes, node; |
aa888a74 | 2988 | |
b5389086 ZY |
2989 | /* do node specific alloc */ |
2990 | if (nid != NUMA_NO_NODE) { | |
2991 | m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), | |
2992 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); | |
2993 | if (!m) | |
2994 | return 0; | |
2995 | goto found; | |
2996 | } | |
2997 | /* allocate from next node when distributing huge pages */ | |
b2261026 | 2998 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
b5389086 | 2999 | m = memblock_alloc_try_nid_raw( |
8b89a116 | 3000 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 3001 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
b5389086 ZY |
3002 | /* |
3003 | * Use the beginning of the huge page to store the | |
3004 | * huge_bootmem_page struct (until gather_bootmem | |
3005 | * puts them into the mem_map). | |
3006 | */ | |
3007 | if (!m) | |
3008 | return 0; | |
3009 | goto found; | |
aa888a74 | 3010 | } |
aa888a74 AK |
3011 | |
3012 | found: | |
aa888a74 | 3013 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 3014 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
3015 | list_add(&m->list, &huge_boot_pages); |
3016 | m->hstate = h; | |
3017 | return 1; | |
3018 | } | |
3019 | ||
48b8d744 MK |
3020 | /* |
3021 | * Put bootmem huge pages into the standard lists after mem_map is up. | |
3022 | * Note: This only applies to gigantic (order > MAX_ORDER) pages. | |
3023 | */ | |
aa888a74 AK |
3024 | static void __init gather_bootmem_prealloc(void) |
3025 | { | |
3026 | struct huge_bootmem_page *m; | |
3027 | ||
3028 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 3029 | struct page *page = virt_to_page(m); |
aa888a74 | 3030 | struct hstate *h = m->hstate; |
ee8f248d | 3031 | |
48b8d744 | 3032 | VM_BUG_ON(!hstate_is_gigantic(h)); |
aa888a74 | 3033 | WARN_ON(page_count(page) != 1); |
7118fc29 MK |
3034 | if (prep_compound_gigantic_page(page, huge_page_order(h))) { |
3035 | WARN_ON(PageReserved(page)); | |
3036 | prep_new_huge_page(h, page, page_to_nid(page)); | |
3037 | put_page(page); /* add to the hugepage allocator */ | |
3038 | } else { | |
416d85ed | 3039 | /* VERY unlikely inflated ref count on a tail page */ |
7118fc29 | 3040 | free_gigantic_page(page, huge_page_order(h)); |
7118fc29 | 3041 | } |
af0fb9df | 3042 | |
b0320c7b | 3043 | /* |
48b8d744 MK |
3044 | * We need to restore the 'stolen' pages to totalram_pages |
3045 | * in order to fix confusing memory reports from free(1) and | |
3046 | * other side-effects, like CommitLimit going negative. | |
b0320c7b | 3047 | */ |
48b8d744 | 3048 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 3049 | cond_resched(); |
aa888a74 AK |
3050 | } |
3051 | } | |
b5389086 ZY |
3052 | static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) |
3053 | { | |
3054 | unsigned long i; | |
3055 | char buf[32]; | |
3056 | ||
3057 | for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { | |
3058 | if (hstate_is_gigantic(h)) { | |
3059 | if (!alloc_bootmem_huge_page(h, nid)) | |
3060 | break; | |
3061 | } else { | |
3062 | struct page *page; | |
3063 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
3064 | ||
3065 | page = alloc_fresh_huge_page(h, gfp_mask, nid, | |
3066 | &node_states[N_MEMORY], NULL); | |
3067 | if (!page) | |
3068 | break; | |
3069 | put_page(page); /* free it into the hugepage allocator */ | |
3070 | } | |
3071 | cond_resched(); | |
3072 | } | |
3073 | if (i == h->max_huge_pages_node[nid]) | |
3074 | return; | |
3075 | ||
3076 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
3077 | pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", | |
3078 | h->max_huge_pages_node[nid], buf, nid, i); | |
3079 | h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); | |
3080 | h->max_huge_pages_node[nid] = i; | |
3081 | } | |
aa888a74 | 3082 | |
8faa8b07 | 3083 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
3084 | { |
3085 | unsigned long i; | |
f60858f9 | 3086 | nodemask_t *node_alloc_noretry; |
b5389086 ZY |
3087 | bool node_specific_alloc = false; |
3088 | ||
3089 | /* skip gigantic hugepages allocation if hugetlb_cma enabled */ | |
3090 | if (hstate_is_gigantic(h) && hugetlb_cma_size) { | |
3091 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); | |
3092 | return; | |
3093 | } | |
3094 | ||
3095 | /* do node specific alloc */ | |
0a7a0f6f | 3096 | for_each_online_node(i) { |
b5389086 ZY |
3097 | if (h->max_huge_pages_node[i] > 0) { |
3098 | hugetlb_hstate_alloc_pages_onenode(h, i); | |
3099 | node_specific_alloc = true; | |
3100 | } | |
3101 | } | |
f60858f9 | 3102 | |
b5389086 ZY |
3103 | if (node_specific_alloc) |
3104 | return; | |
3105 | ||
3106 | /* below will do all node balanced alloc */ | |
f60858f9 MK |
3107 | if (!hstate_is_gigantic(h)) { |
3108 | /* | |
3109 | * Bit mask controlling how hard we retry per-node allocations. | |
3110 | * Ignore errors as lower level routines can deal with | |
3111 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
3112 | * time, we are likely in bigger trouble. | |
3113 | */ | |
3114 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
3115 | GFP_KERNEL); | |
3116 | } else { | |
3117 | /* allocations done at boot time */ | |
3118 | node_alloc_noretry = NULL; | |
3119 | } | |
3120 | ||
3121 | /* bit mask controlling how hard we retry per-node allocations */ | |
3122 | if (node_alloc_noretry) | |
3123 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 3124 | |
e5ff2159 | 3125 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 3126 | if (hstate_is_gigantic(h)) { |
b5389086 | 3127 | if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) |
aa888a74 | 3128 | break; |
0c397dae | 3129 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
3130 | &node_states[N_MEMORY], |
3131 | node_alloc_noretry)) | |
1da177e4 | 3132 | break; |
69ed779a | 3133 | cond_resched(); |
1da177e4 | 3134 | } |
d715cf80 LH |
3135 | if (i < h->max_huge_pages) { |
3136 | char buf[32]; | |
3137 | ||
c6247f72 | 3138 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
3139 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
3140 | h->max_huge_pages, buf, i); | |
3141 | h->max_huge_pages = i; | |
3142 | } | |
f60858f9 | 3143 | kfree(node_alloc_noretry); |
e5ff2159 AK |
3144 | } |
3145 | ||
3146 | static void __init hugetlb_init_hstates(void) | |
3147 | { | |
79dfc695 | 3148 | struct hstate *h, *h2; |
e5ff2159 AK |
3149 | |
3150 | for_each_hstate(h) { | |
8faa8b07 | 3151 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 3152 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 3153 | hugetlb_hstate_alloc_pages(h); |
79dfc695 MK |
3154 | |
3155 | /* | |
3156 | * Set demote order for each hstate. Note that | |
3157 | * h->demote_order is initially 0. | |
3158 | * - We can not demote gigantic pages if runtime freeing | |
3159 | * is not supported, so skip this. | |
a01f4390 MK |
3160 | * - If CMA allocation is possible, we can not demote |
3161 | * HUGETLB_PAGE_ORDER or smaller size pages. | |
79dfc695 MK |
3162 | */ |
3163 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
3164 | continue; | |
a01f4390 MK |
3165 | if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) |
3166 | continue; | |
79dfc695 MK |
3167 | for_each_hstate(h2) { |
3168 | if (h2 == h) | |
3169 | continue; | |
3170 | if (h2->order < h->order && | |
3171 | h2->order > h->demote_order) | |
3172 | h->demote_order = h2->order; | |
3173 | } | |
e5ff2159 AK |
3174 | } |
3175 | } | |
3176 | ||
3177 | static void __init report_hugepages(void) | |
3178 | { | |
3179 | struct hstate *h; | |
3180 | ||
3181 | for_each_hstate(h) { | |
4abd32db | 3182 | char buf[32]; |
c6247f72 MW |
3183 | |
3184 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 3185 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 3186 | buf, h->free_huge_pages); |
e5ff2159 AK |
3187 | } |
3188 | } | |
3189 | ||
1da177e4 | 3190 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
3191 | static void try_to_free_low(struct hstate *h, unsigned long count, |
3192 | nodemask_t *nodes_allowed) | |
1da177e4 | 3193 | { |
4415cc8d | 3194 | int i; |
1121828a | 3195 | LIST_HEAD(page_list); |
4415cc8d | 3196 | |
9487ca60 | 3197 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 3198 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
3199 | return; |
3200 | ||
1121828a MK |
3201 | /* |
3202 | * Collect pages to be freed on a list, and free after dropping lock | |
3203 | */ | |
6ae11b27 | 3204 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 3205 | struct page *page, *next; |
a5516438 AK |
3206 | struct list_head *freel = &h->hugepage_freelists[i]; |
3207 | list_for_each_entry_safe(page, next, freel, lru) { | |
3208 | if (count >= h->nr_huge_pages) | |
1121828a | 3209 | goto out; |
1da177e4 LT |
3210 | if (PageHighMem(page)) |
3211 | continue; | |
6eb4e88a | 3212 | remove_hugetlb_page(h, page, false); |
1121828a | 3213 | list_add(&page->lru, &page_list); |
1da177e4 LT |
3214 | } |
3215 | } | |
1121828a MK |
3216 | |
3217 | out: | |
db71ef79 | 3218 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3219 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 3220 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
3221 | } |
3222 | #else | |
6ae11b27 LS |
3223 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
3224 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
3225 | { |
3226 | } | |
3227 | #endif | |
3228 | ||
20a0307c WF |
3229 | /* |
3230 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
3231 | * balanced by operating on them in a round-robin fashion. | |
3232 | * Returns 1 if an adjustment was made. | |
3233 | */ | |
6ae11b27 LS |
3234 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
3235 | int delta) | |
20a0307c | 3236 | { |
b2261026 | 3237 | int nr_nodes, node; |
20a0307c | 3238 | |
9487ca60 | 3239 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 3240 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 3241 | |
b2261026 JK |
3242 | if (delta < 0) { |
3243 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
3244 | if (h->surplus_huge_pages_node[node]) | |
3245 | goto found; | |
e8c5c824 | 3246 | } |
b2261026 JK |
3247 | } else { |
3248 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
3249 | if (h->surplus_huge_pages_node[node] < | |
3250 | h->nr_huge_pages_node[node]) | |
3251 | goto found; | |
e8c5c824 | 3252 | } |
b2261026 JK |
3253 | } |
3254 | return 0; | |
20a0307c | 3255 | |
b2261026 JK |
3256 | found: |
3257 | h->surplus_huge_pages += delta; | |
3258 | h->surplus_huge_pages_node[node] += delta; | |
3259 | return 1; | |
20a0307c WF |
3260 | } |
3261 | ||
a5516438 | 3262 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 3263 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 3264 | nodemask_t *nodes_allowed) |
1da177e4 | 3265 | { |
7893d1d5 | 3266 | unsigned long min_count, ret; |
10c6ec49 MK |
3267 | struct page *page; |
3268 | LIST_HEAD(page_list); | |
f60858f9 MK |
3269 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
3270 | ||
3271 | /* | |
3272 | * Bit mask controlling how hard we retry per-node allocations. | |
3273 | * If we can not allocate the bit mask, do not attempt to allocate | |
3274 | * the requested huge pages. | |
3275 | */ | |
3276 | if (node_alloc_noretry) | |
3277 | nodes_clear(*node_alloc_noretry); | |
3278 | else | |
3279 | return -ENOMEM; | |
1da177e4 | 3280 | |
29383967 MK |
3281 | /* |
3282 | * resize_lock mutex prevents concurrent adjustments to number of | |
3283 | * pages in hstate via the proc/sysfs interfaces. | |
3284 | */ | |
3285 | mutex_lock(&h->resize_lock); | |
b65d4adb | 3286 | flush_free_hpage_work(h); |
db71ef79 | 3287 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 3288 | |
fd875dca MK |
3289 | /* |
3290 | * Check for a node specific request. | |
3291 | * Changing node specific huge page count may require a corresponding | |
3292 | * change to the global count. In any case, the passed node mask | |
3293 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
3294 | */ | |
3295 | if (nid != NUMA_NO_NODE) { | |
3296 | unsigned long old_count = count; | |
3297 | ||
3298 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
3299 | /* | |
3300 | * User may have specified a large count value which caused the | |
3301 | * above calculation to overflow. In this case, they wanted | |
3302 | * to allocate as many huge pages as possible. Set count to | |
3303 | * largest possible value to align with their intention. | |
3304 | */ | |
3305 | if (count < old_count) | |
3306 | count = ULONG_MAX; | |
3307 | } | |
3308 | ||
4eb0716e AG |
3309 | /* |
3310 | * Gigantic pages runtime allocation depend on the capability for large | |
3311 | * page range allocation. | |
3312 | * If the system does not provide this feature, return an error when | |
3313 | * the user tries to allocate gigantic pages but let the user free the | |
3314 | * boottime allocated gigantic pages. | |
3315 | */ | |
3316 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
3317 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 3318 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3319 | mutex_unlock(&h->resize_lock); |
f60858f9 | 3320 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
3321 | return -EINVAL; |
3322 | } | |
3323 | /* Fall through to decrease pool */ | |
3324 | } | |
aa888a74 | 3325 | |
7893d1d5 AL |
3326 | /* |
3327 | * Increase the pool size | |
3328 | * First take pages out of surplus state. Then make up the | |
3329 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 3330 | * |
0c397dae | 3331 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
3332 | * to convert a surplus huge page to a normal huge page. That is |
3333 | * not critical, though, it just means the overall size of the | |
3334 | * pool might be one hugepage larger than it needs to be, but | |
3335 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 3336 | */ |
a5516438 | 3337 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 3338 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
3339 | break; |
3340 | } | |
3341 | ||
a5516438 | 3342 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
3343 | /* |
3344 | * If this allocation races such that we no longer need the | |
3345 | * page, free_huge_page will handle it by freeing the page | |
3346 | * and reducing the surplus. | |
3347 | */ | |
db71ef79 | 3348 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
3349 | |
3350 | /* yield cpu to avoid soft lockup */ | |
3351 | cond_resched(); | |
3352 | ||
f60858f9 MK |
3353 | ret = alloc_pool_huge_page(h, nodes_allowed, |
3354 | node_alloc_noretry); | |
db71ef79 | 3355 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
3356 | if (!ret) |
3357 | goto out; | |
3358 | ||
536240f2 MG |
3359 | /* Bail for signals. Probably ctrl-c from user */ |
3360 | if (signal_pending(current)) | |
3361 | goto out; | |
7893d1d5 | 3362 | } |
7893d1d5 AL |
3363 | |
3364 | /* | |
3365 | * Decrease the pool size | |
3366 | * First return free pages to the buddy allocator (being careful | |
3367 | * to keep enough around to satisfy reservations). Then place | |
3368 | * pages into surplus state as needed so the pool will shrink | |
3369 | * to the desired size as pages become free. | |
d1c3fb1f NA |
3370 | * |
3371 | * By placing pages into the surplus state independent of the | |
3372 | * overcommit value, we are allowing the surplus pool size to | |
3373 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 3374 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
3375 | * though, we'll note that we're not allowed to exceed surplus |
3376 | * and won't grow the pool anywhere else. Not until one of the | |
3377 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 3378 | */ |
a5516438 | 3379 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 3380 | min_count = max(count, min_count); |
6ae11b27 | 3381 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
3382 | |
3383 | /* | |
3384 | * Collect pages to be removed on list without dropping lock | |
3385 | */ | |
a5516438 | 3386 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
3387 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
3388 | if (!page) | |
1da177e4 | 3389 | break; |
10c6ec49 MK |
3390 | |
3391 | list_add(&page->lru, &page_list); | |
1da177e4 | 3392 | } |
10c6ec49 | 3393 | /* free the pages after dropping lock */ |
db71ef79 | 3394 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3395 | update_and_free_pages_bulk(h, &page_list); |
b65d4adb | 3396 | flush_free_hpage_work(h); |
db71ef79 | 3397 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 3398 | |
a5516438 | 3399 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 3400 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
3401 | break; |
3402 | } | |
3403 | out: | |
4eb0716e | 3404 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 3405 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3406 | mutex_unlock(&h->resize_lock); |
4eb0716e | 3407 | |
f60858f9 MK |
3408 | NODEMASK_FREE(node_alloc_noretry); |
3409 | ||
4eb0716e | 3410 | return 0; |
1da177e4 LT |
3411 | } |
3412 | ||
8531fc6f MK |
3413 | static int demote_free_huge_page(struct hstate *h, struct page *page) |
3414 | { | |
3415 | int i, nid = page_to_nid(page); | |
3416 | struct hstate *target_hstate; | |
3417 | int rc = 0; | |
3418 | ||
3419 | target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); | |
3420 | ||
3421 | remove_hugetlb_page_for_demote(h, page, false); | |
3422 | spin_unlock_irq(&hugetlb_lock); | |
3423 | ||
5981611d | 3424 | rc = hugetlb_vmemmap_alloc(h, page); |
8531fc6f MK |
3425 | if (rc) { |
3426 | /* Allocation of vmemmmap failed, we can not demote page */ | |
3427 | spin_lock_irq(&hugetlb_lock); | |
3428 | set_page_refcounted(page); | |
3429 | add_hugetlb_page(h, page, false); | |
3430 | return rc; | |
3431 | } | |
3432 | ||
3433 | /* | |
3434 | * Use destroy_compound_hugetlb_page_for_demote for all huge page | |
3435 | * sizes as it will not ref count pages. | |
3436 | */ | |
3437 | destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); | |
3438 | ||
3439 | /* | |
3440 | * Taking target hstate mutex synchronizes with set_max_huge_pages. | |
3441 | * Without the mutex, pages added to target hstate could be marked | |
3442 | * as surplus. | |
3443 | * | |
3444 | * Note that we already hold h->resize_lock. To prevent deadlock, | |
3445 | * use the convention of always taking larger size hstate mutex first. | |
3446 | */ | |
3447 | mutex_lock(&target_hstate->resize_lock); | |
3448 | for (i = 0; i < pages_per_huge_page(h); | |
3449 | i += pages_per_huge_page(target_hstate)) { | |
3450 | if (hstate_is_gigantic(target_hstate)) | |
3451 | prep_compound_gigantic_page_for_demote(page + i, | |
3452 | target_hstate->order); | |
3453 | else | |
3454 | prep_compound_page(page + i, target_hstate->order); | |
3455 | set_page_private(page + i, 0); | |
3456 | set_page_refcounted(page + i); | |
3457 | prep_new_huge_page(target_hstate, page + i, nid); | |
3458 | put_page(page + i); | |
3459 | } | |
3460 | mutex_unlock(&target_hstate->resize_lock); | |
3461 | ||
3462 | spin_lock_irq(&hugetlb_lock); | |
3463 | ||
3464 | /* | |
3465 | * Not absolutely necessary, but for consistency update max_huge_pages | |
3466 | * based on pool changes for the demoted page. | |
3467 | */ | |
3468 | h->max_huge_pages--; | |
3469 | target_hstate->max_huge_pages += pages_per_huge_page(h); | |
3470 | ||
3471 | return rc; | |
3472 | } | |
3473 | ||
79dfc695 MK |
3474 | static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
3475 | __must_hold(&hugetlb_lock) | |
3476 | { | |
8531fc6f MK |
3477 | int nr_nodes, node; |
3478 | struct page *page; | |
79dfc695 MK |
3479 | |
3480 | lockdep_assert_held(&hugetlb_lock); | |
3481 | ||
3482 | /* We should never get here if no demote order */ | |
3483 | if (!h->demote_order) { | |
3484 | pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); | |
3485 | return -EINVAL; /* internal error */ | |
3486 | } | |
3487 | ||
8531fc6f | 3488 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
5a317412 MK |
3489 | list_for_each_entry(page, &h->hugepage_freelists[node], lru) { |
3490 | if (PageHWPoison(page)) | |
3491 | continue; | |
3492 | ||
3493 | return demote_free_huge_page(h, page); | |
8531fc6f MK |
3494 | } |
3495 | } | |
3496 | ||
5a317412 MK |
3497 | /* |
3498 | * Only way to get here is if all pages on free lists are poisoned. | |
3499 | * Return -EBUSY so that caller will not retry. | |
3500 | */ | |
3501 | return -EBUSY; | |
79dfc695 MK |
3502 | } |
3503 | ||
a3437870 NA |
3504 | #define HSTATE_ATTR_RO(_name) \ |
3505 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
3506 | ||
79dfc695 MK |
3507 | #define HSTATE_ATTR_WO(_name) \ |
3508 | static struct kobj_attribute _name##_attr = __ATTR_WO(_name) | |
3509 | ||
a3437870 | 3510 | #define HSTATE_ATTR(_name) \ |
98bc26ac | 3511 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
a3437870 NA |
3512 | |
3513 | static struct kobject *hugepages_kobj; | |
3514 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3515 | ||
9a305230 LS |
3516 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
3517 | ||
3518 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
3519 | { |
3520 | int i; | |
9a305230 | 3521 | |
a3437870 | 3522 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
3523 | if (hstate_kobjs[i] == kobj) { |
3524 | if (nidp) | |
3525 | *nidp = NUMA_NO_NODE; | |
a3437870 | 3526 | return &hstates[i]; |
9a305230 LS |
3527 | } |
3528 | ||
3529 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
3530 | } |
3531 | ||
06808b08 | 3532 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
3533 | struct kobj_attribute *attr, char *buf) |
3534 | { | |
9a305230 LS |
3535 | struct hstate *h; |
3536 | unsigned long nr_huge_pages; | |
3537 | int nid; | |
3538 | ||
3539 | h = kobj_to_hstate(kobj, &nid); | |
3540 | if (nid == NUMA_NO_NODE) | |
3541 | nr_huge_pages = h->nr_huge_pages; | |
3542 | else | |
3543 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
3544 | ||
ae7a927d | 3545 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 3546 | } |
adbe8726 | 3547 | |
238d3c13 DR |
3548 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
3549 | struct hstate *h, int nid, | |
3550 | unsigned long count, size_t len) | |
a3437870 NA |
3551 | { |
3552 | int err; | |
2d0adf7e | 3553 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 3554 | |
2d0adf7e OS |
3555 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
3556 | return -EINVAL; | |
adbe8726 | 3557 | |
9a305230 LS |
3558 | if (nid == NUMA_NO_NODE) { |
3559 | /* | |
3560 | * global hstate attribute | |
3561 | */ | |
3562 | if (!(obey_mempolicy && | |
2d0adf7e OS |
3563 | init_nodemask_of_mempolicy(&nodes_allowed))) |
3564 | n_mask = &node_states[N_MEMORY]; | |
3565 | else | |
3566 | n_mask = &nodes_allowed; | |
3567 | } else { | |
9a305230 | 3568 | /* |
fd875dca MK |
3569 | * Node specific request. count adjustment happens in |
3570 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 3571 | */ |
2d0adf7e OS |
3572 | init_nodemask_of_node(&nodes_allowed, nid); |
3573 | n_mask = &nodes_allowed; | |
fd875dca | 3574 | } |
9a305230 | 3575 | |
2d0adf7e | 3576 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 3577 | |
4eb0716e | 3578 | return err ? err : len; |
06808b08 LS |
3579 | } |
3580 | ||
238d3c13 DR |
3581 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
3582 | struct kobject *kobj, const char *buf, | |
3583 | size_t len) | |
3584 | { | |
3585 | struct hstate *h; | |
3586 | unsigned long count; | |
3587 | int nid; | |
3588 | int err; | |
3589 | ||
3590 | err = kstrtoul(buf, 10, &count); | |
3591 | if (err) | |
3592 | return err; | |
3593 | ||
3594 | h = kobj_to_hstate(kobj, &nid); | |
3595 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
3596 | } | |
3597 | ||
06808b08 LS |
3598 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
3599 | struct kobj_attribute *attr, char *buf) | |
3600 | { | |
3601 | return nr_hugepages_show_common(kobj, attr, buf); | |
3602 | } | |
3603 | ||
3604 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
3605 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3606 | { | |
238d3c13 | 3607 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
3608 | } |
3609 | HSTATE_ATTR(nr_hugepages); | |
3610 | ||
06808b08 LS |
3611 | #ifdef CONFIG_NUMA |
3612 | ||
3613 | /* | |
3614 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
3615 | * huge page alloc/free. | |
3616 | */ | |
3617 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
3618 | struct kobj_attribute *attr, |
3619 | char *buf) | |
06808b08 LS |
3620 | { |
3621 | return nr_hugepages_show_common(kobj, attr, buf); | |
3622 | } | |
3623 | ||
3624 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
3625 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3626 | { | |
238d3c13 | 3627 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
3628 | } |
3629 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
3630 | #endif | |
3631 | ||
3632 | ||
a3437870 NA |
3633 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
3634 | struct kobj_attribute *attr, char *buf) | |
3635 | { | |
9a305230 | 3636 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3637 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 3638 | } |
adbe8726 | 3639 | |
a3437870 NA |
3640 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
3641 | struct kobj_attribute *attr, const char *buf, size_t count) | |
3642 | { | |
3643 | int err; | |
3644 | unsigned long input; | |
9a305230 | 3645 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 3646 | |
bae7f4ae | 3647 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
3648 | return -EINVAL; |
3649 | ||
3dbb95f7 | 3650 | err = kstrtoul(buf, 10, &input); |
a3437870 | 3651 | if (err) |
73ae31e5 | 3652 | return err; |
a3437870 | 3653 | |
db71ef79 | 3654 | spin_lock_irq(&hugetlb_lock); |
a3437870 | 3655 | h->nr_overcommit_huge_pages = input; |
db71ef79 | 3656 | spin_unlock_irq(&hugetlb_lock); |
a3437870 NA |
3657 | |
3658 | return count; | |
3659 | } | |
3660 | HSTATE_ATTR(nr_overcommit_hugepages); | |
3661 | ||
3662 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
3663 | struct kobj_attribute *attr, char *buf) | |
3664 | { | |
9a305230 LS |
3665 | struct hstate *h; |
3666 | unsigned long free_huge_pages; | |
3667 | int nid; | |
3668 | ||
3669 | h = kobj_to_hstate(kobj, &nid); | |
3670 | if (nid == NUMA_NO_NODE) | |
3671 | free_huge_pages = h->free_huge_pages; | |
3672 | else | |
3673 | free_huge_pages = h->free_huge_pages_node[nid]; | |
3674 | ||
ae7a927d | 3675 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
3676 | } |
3677 | HSTATE_ATTR_RO(free_hugepages); | |
3678 | ||
3679 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
3680 | struct kobj_attribute *attr, char *buf) | |
3681 | { | |
9a305230 | 3682 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3683 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
3684 | } |
3685 | HSTATE_ATTR_RO(resv_hugepages); | |
3686 | ||
3687 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
3688 | struct kobj_attribute *attr, char *buf) | |
3689 | { | |
9a305230 LS |
3690 | struct hstate *h; |
3691 | unsigned long surplus_huge_pages; | |
3692 | int nid; | |
3693 | ||
3694 | h = kobj_to_hstate(kobj, &nid); | |
3695 | if (nid == NUMA_NO_NODE) | |
3696 | surplus_huge_pages = h->surplus_huge_pages; | |
3697 | else | |
3698 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
3699 | ||
ae7a927d | 3700 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
3701 | } |
3702 | HSTATE_ATTR_RO(surplus_hugepages); | |
3703 | ||
79dfc695 MK |
3704 | static ssize_t demote_store(struct kobject *kobj, |
3705 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3706 | { | |
3707 | unsigned long nr_demote; | |
3708 | unsigned long nr_available; | |
3709 | nodemask_t nodes_allowed, *n_mask; | |
3710 | struct hstate *h; | |
3711 | int err = 0; | |
3712 | int nid; | |
3713 | ||
3714 | err = kstrtoul(buf, 10, &nr_demote); | |
3715 | if (err) | |
3716 | return err; | |
3717 | h = kobj_to_hstate(kobj, &nid); | |
3718 | ||
3719 | if (nid != NUMA_NO_NODE) { | |
3720 | init_nodemask_of_node(&nodes_allowed, nid); | |
3721 | n_mask = &nodes_allowed; | |
3722 | } else { | |
3723 | n_mask = &node_states[N_MEMORY]; | |
3724 | } | |
3725 | ||
3726 | /* Synchronize with other sysfs operations modifying huge pages */ | |
3727 | mutex_lock(&h->resize_lock); | |
3728 | spin_lock_irq(&hugetlb_lock); | |
3729 | ||
3730 | while (nr_demote) { | |
3731 | /* | |
3732 | * Check for available pages to demote each time thorough the | |
3733 | * loop as demote_pool_huge_page will drop hugetlb_lock. | |
79dfc695 MK |
3734 | */ |
3735 | if (nid != NUMA_NO_NODE) | |
3736 | nr_available = h->free_huge_pages_node[nid]; | |
3737 | else | |
3738 | nr_available = h->free_huge_pages; | |
3739 | nr_available -= h->resv_huge_pages; | |
3740 | if (!nr_available) | |
3741 | break; | |
3742 | ||
3743 | err = demote_pool_huge_page(h, n_mask); | |
3744 | if (err) | |
3745 | break; | |
3746 | ||
3747 | nr_demote--; | |
3748 | } | |
3749 | ||
3750 | spin_unlock_irq(&hugetlb_lock); | |
3751 | mutex_unlock(&h->resize_lock); | |
3752 | ||
3753 | if (err) | |
3754 | return err; | |
3755 | return len; | |
3756 | } | |
3757 | HSTATE_ATTR_WO(demote); | |
3758 | ||
3759 | static ssize_t demote_size_show(struct kobject *kobj, | |
3760 | struct kobj_attribute *attr, char *buf) | |
3761 | { | |
3762 | int nid; | |
3763 | struct hstate *h = kobj_to_hstate(kobj, &nid); | |
3764 | unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; | |
3765 | ||
3766 | return sysfs_emit(buf, "%lukB\n", demote_size); | |
3767 | } | |
3768 | ||
3769 | static ssize_t demote_size_store(struct kobject *kobj, | |
3770 | struct kobj_attribute *attr, | |
3771 | const char *buf, size_t count) | |
3772 | { | |
3773 | struct hstate *h, *demote_hstate; | |
3774 | unsigned long demote_size; | |
3775 | unsigned int demote_order; | |
3776 | int nid; | |
3777 | ||
3778 | demote_size = (unsigned long)memparse(buf, NULL); | |
3779 | ||
3780 | demote_hstate = size_to_hstate(demote_size); | |
3781 | if (!demote_hstate) | |
3782 | return -EINVAL; | |
3783 | demote_order = demote_hstate->order; | |
a01f4390 MK |
3784 | if (demote_order < HUGETLB_PAGE_ORDER) |
3785 | return -EINVAL; | |
79dfc695 MK |
3786 | |
3787 | /* demote order must be smaller than hstate order */ | |
3788 | h = kobj_to_hstate(kobj, &nid); | |
3789 | if (demote_order >= h->order) | |
3790 | return -EINVAL; | |
3791 | ||
3792 | /* resize_lock synchronizes access to demote size and writes */ | |
3793 | mutex_lock(&h->resize_lock); | |
3794 | h->demote_order = demote_order; | |
3795 | mutex_unlock(&h->resize_lock); | |
3796 | ||
3797 | return count; | |
3798 | } | |
3799 | HSTATE_ATTR(demote_size); | |
3800 | ||
a3437870 NA |
3801 | static struct attribute *hstate_attrs[] = { |
3802 | &nr_hugepages_attr.attr, | |
3803 | &nr_overcommit_hugepages_attr.attr, | |
3804 | &free_hugepages_attr.attr, | |
3805 | &resv_hugepages_attr.attr, | |
3806 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3807 | #ifdef CONFIG_NUMA |
3808 | &nr_hugepages_mempolicy_attr.attr, | |
3809 | #endif | |
a3437870 NA |
3810 | NULL, |
3811 | }; | |
3812 | ||
67e5ed96 | 3813 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3814 | .attrs = hstate_attrs, |
3815 | }; | |
3816 | ||
79dfc695 MK |
3817 | static struct attribute *hstate_demote_attrs[] = { |
3818 | &demote_size_attr.attr, | |
3819 | &demote_attr.attr, | |
3820 | NULL, | |
3821 | }; | |
3822 | ||
3823 | static const struct attribute_group hstate_demote_attr_group = { | |
3824 | .attrs = hstate_demote_attrs, | |
3825 | }; | |
3826 | ||
094e9539 JM |
3827 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3828 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3829 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3830 | { |
3831 | int retval; | |
972dc4de | 3832 | int hi = hstate_index(h); |
a3437870 | 3833 | |
9a305230 LS |
3834 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3835 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3836 | return -ENOMEM; |
3837 | ||
9a305230 | 3838 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 3839 | if (retval) { |
9a305230 | 3840 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 ML |
3841 | hstate_kobjs[hi] = NULL; |
3842 | } | |
a3437870 | 3843 | |
79dfc695 MK |
3844 | if (h->demote_order) { |
3845 | if (sysfs_create_group(hstate_kobjs[hi], | |
3846 | &hstate_demote_attr_group)) | |
3847 | pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); | |
3848 | } | |
3849 | ||
a3437870 NA |
3850 | return retval; |
3851 | } | |
3852 | ||
3853 | static void __init hugetlb_sysfs_init(void) | |
3854 | { | |
3855 | struct hstate *h; | |
3856 | int err; | |
3857 | ||
3858 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3859 | if (!hugepages_kobj) | |
3860 | return; | |
3861 | ||
3862 | for_each_hstate(h) { | |
9a305230 LS |
3863 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3864 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3865 | if (err) |
282f4214 | 3866 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
3867 | } |
3868 | } | |
3869 | ||
9a305230 LS |
3870 | #ifdef CONFIG_NUMA |
3871 | ||
3872 | /* | |
3873 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3874 | * with node devices in node_devices[] using a parallel array. The array |
3875 | * index of a node device or _hstate == node id. | |
3876 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3877 | * the base kernel, on the hugetlb module. |
3878 | */ | |
3879 | struct node_hstate { | |
3880 | struct kobject *hugepages_kobj; | |
3881 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3882 | }; | |
b4e289a6 | 3883 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3884 | |
3885 | /* | |
10fbcf4c | 3886 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3887 | */ |
3888 | static struct attribute *per_node_hstate_attrs[] = { | |
3889 | &nr_hugepages_attr.attr, | |
3890 | &free_hugepages_attr.attr, | |
3891 | &surplus_hugepages_attr.attr, | |
3892 | NULL, | |
3893 | }; | |
3894 | ||
67e5ed96 | 3895 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3896 | .attrs = per_node_hstate_attrs, |
3897 | }; | |
3898 | ||
3899 | /* | |
10fbcf4c | 3900 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3901 | * Returns node id via non-NULL nidp. |
3902 | */ | |
3903 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3904 | { | |
3905 | int nid; | |
3906 | ||
3907 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3908 | struct node_hstate *nhs = &node_hstates[nid]; | |
3909 | int i; | |
3910 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3911 | if (nhs->hstate_kobjs[i] == kobj) { | |
3912 | if (nidp) | |
3913 | *nidp = nid; | |
3914 | return &hstates[i]; | |
3915 | } | |
3916 | } | |
3917 | ||
3918 | BUG(); | |
3919 | return NULL; | |
3920 | } | |
3921 | ||
3922 | /* | |
10fbcf4c | 3923 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3924 | * No-op if no hstate attributes attached. |
3925 | */ | |
3cd8b44f | 3926 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3927 | { |
3928 | struct hstate *h; | |
10fbcf4c | 3929 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3930 | |
3931 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3932 | return; /* no hstate attributes */ |
9a305230 | 3933 | |
972dc4de AK |
3934 | for_each_hstate(h) { |
3935 | int idx = hstate_index(h); | |
3936 | if (nhs->hstate_kobjs[idx]) { | |
3937 | kobject_put(nhs->hstate_kobjs[idx]); | |
3938 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3939 | } |
972dc4de | 3940 | } |
9a305230 LS |
3941 | |
3942 | kobject_put(nhs->hugepages_kobj); | |
3943 | nhs->hugepages_kobj = NULL; | |
3944 | } | |
3945 | ||
9a305230 LS |
3946 | |
3947 | /* | |
10fbcf4c | 3948 | * Register hstate attributes for a single node device. |
9a305230 LS |
3949 | * No-op if attributes already registered. |
3950 | */ | |
3cd8b44f | 3951 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3952 | { |
3953 | struct hstate *h; | |
10fbcf4c | 3954 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3955 | int err; |
3956 | ||
3957 | if (nhs->hugepages_kobj) | |
3958 | return; /* already allocated */ | |
3959 | ||
3960 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3961 | &node->dev.kobj); |
9a305230 LS |
3962 | if (!nhs->hugepages_kobj) |
3963 | return; | |
3964 | ||
3965 | for_each_hstate(h) { | |
3966 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3967 | nhs->hstate_kobjs, | |
3968 | &per_node_hstate_attr_group); | |
3969 | if (err) { | |
282f4214 | 3970 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3971 | h->name, node->dev.id); |
9a305230 LS |
3972 | hugetlb_unregister_node(node); |
3973 | break; | |
3974 | } | |
3975 | } | |
3976 | } | |
3977 | ||
3978 | /* | |
9b5e5d0f | 3979 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3980 | * devices of nodes that have memory. All on-line nodes should have |
3981 | * registered their associated device by this time. | |
9a305230 | 3982 | */ |
7d9ca000 | 3983 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3984 | { |
3985 | int nid; | |
3986 | ||
8cebfcd0 | 3987 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3988 | struct node *node = node_devices[nid]; |
10fbcf4c | 3989 | if (node->dev.id == nid) |
9a305230 LS |
3990 | hugetlb_register_node(node); |
3991 | } | |
3992 | ||
3993 | /* | |
10fbcf4c | 3994 | * Let the node device driver know we're here so it can |
9a305230 LS |
3995 | * [un]register hstate attributes on node hotplug. |
3996 | */ | |
3997 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3998 | hugetlb_unregister_node); | |
3999 | } | |
4000 | #else /* !CONFIG_NUMA */ | |
4001 | ||
4002 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
4003 | { | |
4004 | BUG(); | |
4005 | if (nidp) | |
4006 | *nidp = -1; | |
4007 | return NULL; | |
4008 | } | |
4009 | ||
9a305230 LS |
4010 | static void hugetlb_register_all_nodes(void) { } |
4011 | ||
4012 | #endif | |
4013 | ||
a3437870 NA |
4014 | static int __init hugetlb_init(void) |
4015 | { | |
8382d914 DB |
4016 | int i; |
4017 | ||
d6995da3 MK |
4018 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
4019 | __NR_HPAGEFLAGS); | |
4020 | ||
c2833a5b MK |
4021 | if (!hugepages_supported()) { |
4022 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
4023 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 4024 | return 0; |
c2833a5b | 4025 | } |
a3437870 | 4026 | |
282f4214 MK |
4027 | /* |
4028 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
4029 | * architectures depend on setup being done here. | |
4030 | */ | |
4031 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
4032 | if (!parsed_default_hugepagesz) { | |
4033 | /* | |
4034 | * If we did not parse a default huge page size, set | |
4035 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
4036 | * number of huge pages for this default size was implicitly | |
4037 | * specified, set that here as well. | |
4038 | * Note that the implicit setting will overwrite an explicit | |
4039 | * setting. A warning will be printed in this case. | |
4040 | */ | |
4041 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
4042 | if (default_hstate_max_huge_pages) { | |
4043 | if (default_hstate.max_huge_pages) { | |
4044 | char buf[32]; | |
4045 | ||
4046 | string_get_size(huge_page_size(&default_hstate), | |
4047 | 1, STRING_UNITS_2, buf, 32); | |
4048 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
4049 | default_hstate.max_huge_pages, buf); | |
4050 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
4051 | default_hstate_max_huge_pages); | |
4052 | } | |
4053 | default_hstate.max_huge_pages = | |
4054 | default_hstate_max_huge_pages; | |
b5389086 | 4055 | |
0a7a0f6f | 4056 | for_each_online_node(i) |
b5389086 ZY |
4057 | default_hstate.max_huge_pages_node[i] = |
4058 | default_hugepages_in_node[i]; | |
d715cf80 | 4059 | } |
f8b74815 | 4060 | } |
a3437870 | 4061 | |
cf11e85f | 4062 | hugetlb_cma_check(); |
a3437870 | 4063 | hugetlb_init_hstates(); |
aa888a74 | 4064 | gather_bootmem_prealloc(); |
a3437870 NA |
4065 | report_hugepages(); |
4066 | ||
4067 | hugetlb_sysfs_init(); | |
9a305230 | 4068 | hugetlb_register_all_nodes(); |
7179e7bf | 4069 | hugetlb_cgroup_file_init(); |
9a305230 | 4070 | |
8382d914 DB |
4071 | #ifdef CONFIG_SMP |
4072 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
4073 | #else | |
4074 | num_fault_mutexes = 1; | |
4075 | #endif | |
c672c7f2 | 4076 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
4077 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
4078 | GFP_KERNEL); | |
c672c7f2 | 4079 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
4080 | |
4081 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 4082 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
4083 | return 0; |
4084 | } | |
3e89e1c5 | 4085 | subsys_initcall(hugetlb_init); |
a3437870 | 4086 | |
ae94da89 MK |
4087 | /* Overwritten by architectures with more huge page sizes */ |
4088 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 4089 | { |
ae94da89 | 4090 | return size == HPAGE_SIZE; |
9fee021d VT |
4091 | } |
4092 | ||
d00181b9 | 4093 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
4094 | { |
4095 | struct hstate *h; | |
8faa8b07 AK |
4096 | unsigned long i; |
4097 | ||
a3437870 | 4098 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
4099 | return; |
4100 | } | |
47d38344 | 4101 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 4102 | BUG_ON(order == 0); |
47d38344 | 4103 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 4104 | mutex_init(&h->resize_lock); |
a3437870 | 4105 | h->order = order; |
aca78307 | 4106 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
4107 | for (i = 0; i < MAX_NUMNODES; ++i) |
4108 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 4109 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
4110 | h->next_nid_to_alloc = first_memory_node; |
4111 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
4112 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
4113 | huge_page_size(h)/1024); | |
77490587 | 4114 | hugetlb_vmemmap_init(h); |
8faa8b07 | 4115 | |
a3437870 NA |
4116 | parsed_hstate = h; |
4117 | } | |
4118 | ||
b5389086 ZY |
4119 | bool __init __weak hugetlb_node_alloc_supported(void) |
4120 | { | |
4121 | return true; | |
4122 | } | |
f87442f4 PL |
4123 | |
4124 | static void __init hugepages_clear_pages_in_node(void) | |
4125 | { | |
4126 | if (!hugetlb_max_hstate) { | |
4127 | default_hstate_max_huge_pages = 0; | |
4128 | memset(default_hugepages_in_node, 0, | |
4129 | MAX_NUMNODES * sizeof(unsigned int)); | |
4130 | } else { | |
4131 | parsed_hstate->max_huge_pages = 0; | |
4132 | memset(parsed_hstate->max_huge_pages_node, 0, | |
4133 | MAX_NUMNODES * sizeof(unsigned int)); | |
4134 | } | |
4135 | } | |
4136 | ||
282f4214 MK |
4137 | /* |
4138 | * hugepages command line processing | |
4139 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
4140 | * specification. If not, ignore the hugepages value. hugepages can also | |
4141 | * be the first huge page command line option in which case it implicitly | |
4142 | * specifies the number of huge pages for the default size. | |
4143 | */ | |
4144 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
4145 | { |
4146 | unsigned long *mhp; | |
8faa8b07 | 4147 | static unsigned long *last_mhp; |
b5389086 ZY |
4148 | int node = NUMA_NO_NODE; |
4149 | int count; | |
4150 | unsigned long tmp; | |
4151 | char *p = s; | |
a3437870 | 4152 | |
9fee021d | 4153 | if (!parsed_valid_hugepagesz) { |
282f4214 | 4154 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 4155 | parsed_valid_hugepagesz = true; |
f81f6e4b | 4156 | return 1; |
9fee021d | 4157 | } |
282f4214 | 4158 | |
a3437870 | 4159 | /* |
282f4214 MK |
4160 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
4161 | * yet, so this hugepages= parameter goes to the "default hstate". | |
4162 | * Otherwise, it goes with the previously parsed hugepagesz or | |
4163 | * default_hugepagesz. | |
a3437870 | 4164 | */ |
9fee021d | 4165 | else if (!hugetlb_max_hstate) |
a3437870 NA |
4166 | mhp = &default_hstate_max_huge_pages; |
4167 | else | |
4168 | mhp = &parsed_hstate->max_huge_pages; | |
4169 | ||
8faa8b07 | 4170 | if (mhp == last_mhp) { |
282f4214 | 4171 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
f81f6e4b | 4172 | return 1; |
8faa8b07 AK |
4173 | } |
4174 | ||
b5389086 ZY |
4175 | while (*p) { |
4176 | count = 0; | |
4177 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4178 | goto invalid; | |
4179 | /* Parameter is node format */ | |
4180 | if (p[count] == ':') { | |
4181 | if (!hugetlb_node_alloc_supported()) { | |
4182 | pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); | |
f81f6e4b | 4183 | return 1; |
b5389086 | 4184 | } |
0a7a0f6f | 4185 | if (tmp >= MAX_NUMNODES || !node_online(tmp)) |
e79ce983 | 4186 | goto invalid; |
0a7a0f6f | 4187 | node = array_index_nospec(tmp, MAX_NUMNODES); |
b5389086 | 4188 | p += count + 1; |
b5389086 ZY |
4189 | /* Parse hugepages */ |
4190 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4191 | goto invalid; | |
4192 | if (!hugetlb_max_hstate) | |
4193 | default_hugepages_in_node[node] = tmp; | |
4194 | else | |
4195 | parsed_hstate->max_huge_pages_node[node] = tmp; | |
4196 | *mhp += tmp; | |
4197 | /* Go to parse next node*/ | |
4198 | if (p[count] == ',') | |
4199 | p += count + 1; | |
4200 | else | |
4201 | break; | |
4202 | } else { | |
4203 | if (p != s) | |
4204 | goto invalid; | |
4205 | *mhp = tmp; | |
4206 | break; | |
4207 | } | |
4208 | } | |
a3437870 | 4209 | |
8faa8b07 AK |
4210 | /* |
4211 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 4212 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
4213 | * use the bootmem allocator. |
4214 | */ | |
04adbc3f | 4215 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
4216 | hugetlb_hstate_alloc_pages(parsed_hstate); |
4217 | ||
4218 | last_mhp = mhp; | |
4219 | ||
a3437870 | 4220 | return 1; |
b5389086 ZY |
4221 | |
4222 | invalid: | |
4223 | pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); | |
f87442f4 | 4224 | hugepages_clear_pages_in_node(); |
f81f6e4b | 4225 | return 1; |
a3437870 | 4226 | } |
282f4214 | 4227 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 4228 | |
282f4214 MK |
4229 | /* |
4230 | * hugepagesz command line processing | |
4231 | * A specific huge page size can only be specified once with hugepagesz. | |
4232 | * hugepagesz is followed by hugepages on the command line. The global | |
4233 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
4234 | * hugepagesz argument was valid. | |
4235 | */ | |
359f2544 | 4236 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 4237 | { |
359f2544 | 4238 | unsigned long size; |
282f4214 MK |
4239 | struct hstate *h; |
4240 | ||
4241 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
4242 | size = (unsigned long)memparse(s, NULL); |
4243 | ||
4244 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4245 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
f81f6e4b | 4246 | return 1; |
359f2544 MK |
4247 | } |
4248 | ||
282f4214 MK |
4249 | h = size_to_hstate(size); |
4250 | if (h) { | |
4251 | /* | |
4252 | * hstate for this size already exists. This is normally | |
4253 | * an error, but is allowed if the existing hstate is the | |
4254 | * default hstate. More specifically, it is only allowed if | |
4255 | * the number of huge pages for the default hstate was not | |
4256 | * previously specified. | |
4257 | */ | |
4258 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
4259 | default_hstate.max_huge_pages) { | |
4260 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
f81f6e4b | 4261 | return 1; |
282f4214 MK |
4262 | } |
4263 | ||
4264 | /* | |
4265 | * No need to call hugetlb_add_hstate() as hstate already | |
4266 | * exists. But, do set parsed_hstate so that a following | |
4267 | * hugepages= parameter will be applied to this hstate. | |
4268 | */ | |
4269 | parsed_hstate = h; | |
4270 | parsed_valid_hugepagesz = true; | |
4271 | return 1; | |
38237830 MK |
4272 | } |
4273 | ||
359f2544 | 4274 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 4275 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
4276 | return 1; |
4277 | } | |
359f2544 MK |
4278 | __setup("hugepagesz=", hugepagesz_setup); |
4279 | ||
282f4214 MK |
4280 | /* |
4281 | * default_hugepagesz command line input | |
4282 | * Only one instance of default_hugepagesz allowed on command line. | |
4283 | */ | |
ae94da89 | 4284 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 4285 | { |
ae94da89 | 4286 | unsigned long size; |
b5389086 | 4287 | int i; |
ae94da89 | 4288 | |
282f4214 | 4289 | parsed_valid_hugepagesz = false; |
282f4214 MK |
4290 | if (parsed_default_hugepagesz) { |
4291 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
f81f6e4b | 4292 | return 1; |
282f4214 MK |
4293 | } |
4294 | ||
ae94da89 MK |
4295 | size = (unsigned long)memparse(s, NULL); |
4296 | ||
4297 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4298 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
f81f6e4b | 4299 | return 1; |
ae94da89 MK |
4300 | } |
4301 | ||
282f4214 MK |
4302 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
4303 | parsed_valid_hugepagesz = true; | |
4304 | parsed_default_hugepagesz = true; | |
4305 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
4306 | ||
4307 | /* | |
4308 | * The number of default huge pages (for this size) could have been | |
4309 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
4310 | * then default_hstate_max_huge_pages is set. If the default huge | |
4311 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
4312 | * allocated here from bootmem allocator. | |
4313 | */ | |
4314 | if (default_hstate_max_huge_pages) { | |
4315 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
0a7a0f6f | 4316 | for_each_online_node(i) |
b5389086 ZY |
4317 | default_hstate.max_huge_pages_node[i] = |
4318 | default_hugepages_in_node[i]; | |
282f4214 MK |
4319 | if (hstate_is_gigantic(&default_hstate)) |
4320 | hugetlb_hstate_alloc_pages(&default_hstate); | |
4321 | default_hstate_max_huge_pages = 0; | |
4322 | } | |
4323 | ||
e11bfbfc NP |
4324 | return 1; |
4325 | } | |
ae94da89 | 4326 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 4327 | |
8ca39e68 | 4328 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
4329 | { |
4330 | int node; | |
4331 | unsigned int nr = 0; | |
8ca39e68 MS |
4332 | nodemask_t *mpol_allowed; |
4333 | unsigned int *array = h->free_huge_pages_node; | |
4334 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
4335 | ||
4336 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 4337 | |
8ca39e68 | 4338 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
c93b0a99 | 4339 | if (!mpol_allowed || node_isset(node, *mpol_allowed)) |
8ca39e68 MS |
4340 | nr += array[node]; |
4341 | } | |
8a213460 NA |
4342 | |
4343 | return nr; | |
4344 | } | |
4345 | ||
4346 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
4347 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
4348 | void *buffer, size_t *length, | |
4349 | loff_t *ppos, unsigned long *out) | |
4350 | { | |
4351 | struct ctl_table dup_table; | |
4352 | ||
4353 | /* | |
4354 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
4355 | * can duplicate the @table and alter the duplicate of it. | |
4356 | */ | |
4357 | dup_table = *table; | |
4358 | dup_table.data = out; | |
4359 | ||
4360 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
4361 | } | |
4362 | ||
06808b08 LS |
4363 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
4364 | struct ctl_table *table, int write, | |
32927393 | 4365 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 4366 | { |
e5ff2159 | 4367 | struct hstate *h = &default_hstate; |
238d3c13 | 4368 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 4369 | int ret; |
e5ff2159 | 4370 | |
457c1b27 | 4371 | if (!hugepages_supported()) |
86613628 | 4372 | return -EOPNOTSUPP; |
457c1b27 | 4373 | |
17743798 MS |
4374 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4375 | &tmp); | |
08d4a246 MH |
4376 | if (ret) |
4377 | goto out; | |
e5ff2159 | 4378 | |
238d3c13 DR |
4379 | if (write) |
4380 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
4381 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
4382 | out: |
4383 | return ret; | |
1da177e4 | 4384 | } |
396faf03 | 4385 | |
06808b08 | 4386 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 4387 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4388 | { |
4389 | ||
4390 | return hugetlb_sysctl_handler_common(false, table, write, | |
4391 | buffer, length, ppos); | |
4392 | } | |
4393 | ||
4394 | #ifdef CONFIG_NUMA | |
4395 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 4396 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4397 | { |
4398 | return hugetlb_sysctl_handler_common(true, table, write, | |
4399 | buffer, length, ppos); | |
4400 | } | |
4401 | #endif /* CONFIG_NUMA */ | |
4402 | ||
a3d0c6aa | 4403 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 4404 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 4405 | { |
a5516438 | 4406 | struct hstate *h = &default_hstate; |
e5ff2159 | 4407 | unsigned long tmp; |
08d4a246 | 4408 | int ret; |
e5ff2159 | 4409 | |
457c1b27 | 4410 | if (!hugepages_supported()) |
86613628 | 4411 | return -EOPNOTSUPP; |
457c1b27 | 4412 | |
c033a93c | 4413 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 4414 | |
bae7f4ae | 4415 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
4416 | return -EINVAL; |
4417 | ||
17743798 MS |
4418 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4419 | &tmp); | |
08d4a246 MH |
4420 | if (ret) |
4421 | goto out; | |
e5ff2159 AK |
4422 | |
4423 | if (write) { | |
db71ef79 | 4424 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 4425 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 4426 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 4427 | } |
08d4a246 MH |
4428 | out: |
4429 | return ret; | |
a3d0c6aa NA |
4430 | } |
4431 | ||
1da177e4 LT |
4432 | #endif /* CONFIG_SYSCTL */ |
4433 | ||
e1759c21 | 4434 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 4435 | { |
fcb2b0c5 RG |
4436 | struct hstate *h; |
4437 | unsigned long total = 0; | |
4438 | ||
457c1b27 NA |
4439 | if (!hugepages_supported()) |
4440 | return; | |
fcb2b0c5 RG |
4441 | |
4442 | for_each_hstate(h) { | |
4443 | unsigned long count = h->nr_huge_pages; | |
4444 | ||
aca78307 | 4445 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
4446 | |
4447 | if (h == &default_hstate) | |
4448 | seq_printf(m, | |
4449 | "HugePages_Total: %5lu\n" | |
4450 | "HugePages_Free: %5lu\n" | |
4451 | "HugePages_Rsvd: %5lu\n" | |
4452 | "HugePages_Surp: %5lu\n" | |
4453 | "Hugepagesize: %8lu kB\n", | |
4454 | count, | |
4455 | h->free_huge_pages, | |
4456 | h->resv_huge_pages, | |
4457 | h->surplus_huge_pages, | |
aca78307 | 4458 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
4459 | } |
4460 | ||
aca78307 | 4461 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
4462 | } |
4463 | ||
7981593b | 4464 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 4465 | { |
a5516438 | 4466 | struct hstate *h = &default_hstate; |
7981593b | 4467 | |
457c1b27 NA |
4468 | if (!hugepages_supported()) |
4469 | return 0; | |
7981593b JP |
4470 | |
4471 | return sysfs_emit_at(buf, len, | |
4472 | "Node %d HugePages_Total: %5u\n" | |
4473 | "Node %d HugePages_Free: %5u\n" | |
4474 | "Node %d HugePages_Surp: %5u\n", | |
4475 | nid, h->nr_huge_pages_node[nid], | |
4476 | nid, h->free_huge_pages_node[nid], | |
4477 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
4478 | } |
4479 | ||
dcadcf1c | 4480 | void hugetlb_show_meminfo_node(int nid) |
949f7ec5 DR |
4481 | { |
4482 | struct hstate *h; | |
949f7ec5 | 4483 | |
457c1b27 NA |
4484 | if (!hugepages_supported()) |
4485 | return; | |
4486 | ||
dcadcf1c GL |
4487 | for_each_hstate(h) |
4488 | printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
4489 | nid, | |
4490 | h->nr_huge_pages_node[nid], | |
4491 | h->free_huge_pages_node[nid], | |
4492 | h->surplus_huge_pages_node[nid], | |
4493 | huge_page_size(h) / SZ_1K); | |
949f7ec5 DR |
4494 | } |
4495 | ||
5d317b2b NH |
4496 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
4497 | { | |
4498 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
4499 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
4500 | } | |
4501 | ||
1da177e4 LT |
4502 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
4503 | unsigned long hugetlb_total_pages(void) | |
4504 | { | |
d0028588 WL |
4505 | struct hstate *h; |
4506 | unsigned long nr_total_pages = 0; | |
4507 | ||
4508 | for_each_hstate(h) | |
4509 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
4510 | return nr_total_pages; | |
1da177e4 | 4511 | } |
1da177e4 | 4512 | |
a5516438 | 4513 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
4514 | { |
4515 | int ret = -ENOMEM; | |
4516 | ||
0aa7f354 ML |
4517 | if (!delta) |
4518 | return 0; | |
4519 | ||
db71ef79 | 4520 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4521 | /* |
4522 | * When cpuset is configured, it breaks the strict hugetlb page | |
4523 | * reservation as the accounting is done on a global variable. Such | |
4524 | * reservation is completely rubbish in the presence of cpuset because | |
4525 | * the reservation is not checked against page availability for the | |
4526 | * current cpuset. Application can still potentially OOM'ed by kernel | |
4527 | * with lack of free htlb page in cpuset that the task is in. | |
4528 | * Attempt to enforce strict accounting with cpuset is almost | |
4529 | * impossible (or too ugly) because cpuset is too fluid that | |
4530 | * task or memory node can be dynamically moved between cpusets. | |
4531 | * | |
4532 | * The change of semantics for shared hugetlb mapping with cpuset is | |
4533 | * undesirable. However, in order to preserve some of the semantics, | |
4534 | * we fall back to check against current free page availability as | |
4535 | * a best attempt and hopefully to minimize the impact of changing | |
4536 | * semantics that cpuset has. | |
8ca39e68 MS |
4537 | * |
4538 | * Apart from cpuset, we also have memory policy mechanism that | |
4539 | * also determines from which node the kernel will allocate memory | |
4540 | * in a NUMA system. So similar to cpuset, we also should consider | |
4541 | * the memory policy of the current task. Similar to the description | |
4542 | * above. | |
fc1b8a73 MG |
4543 | */ |
4544 | if (delta > 0) { | |
a5516438 | 4545 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
4546 | goto out; |
4547 | ||
8ca39e68 | 4548 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 4549 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
4550 | goto out; |
4551 | } | |
4552 | } | |
4553 | ||
4554 | ret = 0; | |
4555 | if (delta < 0) | |
a5516438 | 4556 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
4557 | |
4558 | out: | |
db71ef79 | 4559 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4560 | return ret; |
4561 | } | |
4562 | ||
84afd99b AW |
4563 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
4564 | { | |
f522c3ac | 4565 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
4566 | |
4567 | /* | |
4568 | * This new VMA should share its siblings reservation map if present. | |
4569 | * The VMA will only ever have a valid reservation map pointer where | |
4570 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 4571 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
4572 | * after this open call completes. It is therefore safe to take a |
4573 | * new reference here without additional locking. | |
4574 | */ | |
09a26e83 MK |
4575 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
4576 | resv_map_dup_hugetlb_cgroup_uncharge_info(resv); | |
f522c3ac | 4577 | kref_get(&resv->refs); |
09a26e83 | 4578 | } |
84afd99b AW |
4579 | } |
4580 | ||
a1e78772 MG |
4581 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
4582 | { | |
a5516438 | 4583 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 4584 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 4585 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 4586 | unsigned long reserve, start, end; |
1c5ecae3 | 4587 | long gbl_reserve; |
84afd99b | 4588 | |
4e35f483 JK |
4589 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4590 | return; | |
84afd99b | 4591 | |
4e35f483 JK |
4592 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
4593 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 4594 | |
4e35f483 | 4595 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 4596 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 4597 | if (reserve) { |
1c5ecae3 MK |
4598 | /* |
4599 | * Decrement reserve counts. The global reserve count may be | |
4600 | * adjusted if the subpool has a minimum size. | |
4601 | */ | |
4602 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
4603 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 4604 | } |
e9fe92ae MA |
4605 | |
4606 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
4607 | } |
4608 | ||
31383c68 DW |
4609 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
4610 | { | |
4611 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
4612 | return -EINVAL; | |
4613 | return 0; | |
4614 | } | |
4615 | ||
05ea8860 DW |
4616 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
4617 | { | |
aca78307 | 4618 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
4619 | } |
4620 | ||
1da177e4 LT |
4621 | /* |
4622 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
4623 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 4624 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
4625 | * this far. |
4626 | */ | |
b3ec9f33 | 4627 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
4628 | { |
4629 | BUG(); | |
d0217ac0 | 4630 | return 0; |
1da177e4 LT |
4631 | } |
4632 | ||
eec3636a JC |
4633 | /* |
4634 | * When a new function is introduced to vm_operations_struct and added | |
4635 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
4636 | * This is because under System V memory model, mappings created via | |
4637 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
4638 | * their original vm_ops are overwritten with shm_vm_ops. | |
4639 | */ | |
f0f37e2f | 4640 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 4641 | .fault = hugetlb_vm_op_fault, |
84afd99b | 4642 | .open = hugetlb_vm_op_open, |
a1e78772 | 4643 | .close = hugetlb_vm_op_close, |
dd3b614f | 4644 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 4645 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
4646 | }; |
4647 | ||
1e8f889b DG |
4648 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
4649 | int writable) | |
63551ae0 DG |
4650 | { |
4651 | pte_t entry; | |
79c1c594 | 4652 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
63551ae0 | 4653 | |
1e8f889b | 4654 | if (writable) { |
106c992a GS |
4655 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
4656 | vma->vm_page_prot))); | |
63551ae0 | 4657 | } else { |
106c992a GS |
4658 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
4659 | vma->vm_page_prot)); | |
63551ae0 DG |
4660 | } |
4661 | entry = pte_mkyoung(entry); | |
79c1c594 | 4662 | entry = arch_make_huge_pte(entry, shift, vma->vm_flags); |
63551ae0 DG |
4663 | |
4664 | return entry; | |
4665 | } | |
4666 | ||
1e8f889b DG |
4667 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
4668 | unsigned long address, pte_t *ptep) | |
4669 | { | |
4670 | pte_t entry; | |
4671 | ||
106c992a | 4672 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 4673 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 4674 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
4675 | } |
4676 | ||
d5ed7444 | 4677 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
4678 | { |
4679 | swp_entry_t swp; | |
4680 | ||
4681 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 4682 | return false; |
4a705fef | 4683 | swp = pte_to_swp_entry(pte); |
d79d176a | 4684 | if (is_migration_entry(swp)) |
d5ed7444 | 4685 | return true; |
4a705fef | 4686 | else |
d5ed7444 | 4687 | return false; |
4a705fef NH |
4688 | } |
4689 | ||
3e5c3600 | 4690 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
4691 | { |
4692 | swp_entry_t swp; | |
4693 | ||
4694 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 4695 | return false; |
4a705fef | 4696 | swp = pte_to_swp_entry(pte); |
d79d176a | 4697 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 4698 | return true; |
4a705fef | 4699 | else |
3e5c3600 | 4700 | return false; |
4a705fef | 4701 | } |
1e8f889b | 4702 | |
4eae4efa PX |
4703 | static void |
4704 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
4705 | struct page *new_page) | |
4706 | { | |
4707 | __SetPageUptodate(new_page); | |
4eae4efa | 4708 | hugepage_add_new_anon_rmap(new_page, vma, addr); |
1eba86c0 | 4709 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); |
4eae4efa PX |
4710 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); |
4711 | ClearHPageRestoreReserve(new_page); | |
4712 | SetHPageMigratable(new_page); | |
4713 | } | |
4714 | ||
63551ae0 | 4715 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
bc70fbf2 PX |
4716 | struct vm_area_struct *dst_vma, |
4717 | struct vm_area_struct *src_vma) | |
63551ae0 | 4718 | { |
5e41540c | 4719 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 4720 | struct page *ptepage; |
1c59827d | 4721 | unsigned long addr; |
bc70fbf2 PX |
4722 | bool cow = is_cow_mapping(src_vma->vm_flags); |
4723 | struct hstate *h = hstate_vma(src_vma); | |
a5516438 | 4724 | unsigned long sz = huge_page_size(h); |
4eae4efa | 4725 | unsigned long npages = pages_per_huge_page(h); |
bc70fbf2 | 4726 | struct address_space *mapping = src_vma->vm_file->f_mapping; |
ac46d4f3 | 4727 | struct mmu_notifier_range range; |
e95a9851 | 4728 | unsigned long last_addr_mask; |
e8569dd2 | 4729 | int ret = 0; |
1e8f889b | 4730 | |
ac46d4f3 | 4731 | if (cow) { |
bc70fbf2 PX |
4732 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, |
4733 | src_vma->vm_start, | |
4734 | src_vma->vm_end); | |
ac46d4f3 | 4735 | mmu_notifier_invalidate_range_start(&range); |
623a1ddf DH |
4736 | mmap_assert_write_locked(src); |
4737 | raw_write_seqcount_begin(&src->write_protect_seq); | |
c0d0381a MK |
4738 | } else { |
4739 | /* | |
4740 | * For shared mappings i_mmap_rwsem must be held to call | |
4741 | * huge_pte_alloc, otherwise the returned ptep could go | |
4742 | * away if part of a shared pmd and another thread calls | |
4743 | * huge_pmd_unshare. | |
4744 | */ | |
4745 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 4746 | } |
e8569dd2 | 4747 | |
e95a9851 | 4748 | last_addr_mask = hugetlb_mask_last_page(h); |
bc70fbf2 | 4749 | for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { |
cb900f41 | 4750 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 4751 | src_pte = huge_pte_offset(src, addr, sz); |
e95a9851 MK |
4752 | if (!src_pte) { |
4753 | addr |= last_addr_mask; | |
c74df32c | 4754 | continue; |
e95a9851 | 4755 | } |
bc70fbf2 | 4756 | dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); |
e8569dd2 AS |
4757 | if (!dst_pte) { |
4758 | ret = -ENOMEM; | |
4759 | break; | |
4760 | } | |
c5c99429 | 4761 | |
5e41540c MK |
4762 | /* |
4763 | * If the pagetables are shared don't copy or take references. | |
4764 | * dst_pte == src_pte is the common case of src/dest sharing. | |
4765 | * | |
4766 | * However, src could have 'unshared' and dst shares with | |
4767 | * another vma. If dst_pte !none, this implies sharing. | |
4768 | * Check here before taking page table lock, and once again | |
4769 | * after taking the lock below. | |
4770 | */ | |
4771 | dst_entry = huge_ptep_get(dst_pte); | |
e95a9851 MK |
4772 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) { |
4773 | addr |= last_addr_mask; | |
c5c99429 | 4774 | continue; |
e95a9851 | 4775 | } |
c5c99429 | 4776 | |
cb900f41 KS |
4777 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
4778 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4779 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 4780 | entry = huge_ptep_get(src_pte); |
5e41540c | 4781 | dst_entry = huge_ptep_get(dst_pte); |
4eae4efa | 4782 | again: |
5e41540c MK |
4783 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { |
4784 | /* | |
4785 | * Skip if src entry none. Also, skip in the | |
4786 | * unlikely case dst entry !none as this implies | |
4787 | * sharing with another vma. | |
4788 | */ | |
4a705fef | 4789 | ; |
c2cb0dcc NH |
4790 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { |
4791 | bool uffd_wp = huge_pte_uffd_wp(entry); | |
4792 | ||
4793 | if (!userfaultfd_wp(dst_vma) && uffd_wp) | |
4794 | entry = huge_pte_clear_uffd_wp(entry); | |
4795 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
4796 | } else if (unlikely(is_hugetlb_entry_migration(entry))) { | |
4a705fef | 4797 | swp_entry_t swp_entry = pte_to_swp_entry(entry); |
bc70fbf2 | 4798 | bool uffd_wp = huge_pte_uffd_wp(entry); |
4a705fef | 4799 | |
6c287605 | 4800 | if (!is_readable_migration_entry(swp_entry) && cow) { |
4a705fef NH |
4801 | /* |
4802 | * COW mappings require pages in both | |
4803 | * parent and child to be set to read. | |
4804 | */ | |
4dd845b5 AP |
4805 | swp_entry = make_readable_migration_entry( |
4806 | swp_offset(swp_entry)); | |
4a705fef | 4807 | entry = swp_entry_to_pte(swp_entry); |
bc70fbf2 PX |
4808 | if (userfaultfd_wp(src_vma) && uffd_wp) |
4809 | entry = huge_pte_mkuffd_wp(entry); | |
18f39629 | 4810 | set_huge_pte_at(src, addr, src_pte, entry); |
4a705fef | 4811 | } |
bc70fbf2 PX |
4812 | if (!userfaultfd_wp(dst_vma) && uffd_wp) |
4813 | entry = huge_pte_clear_uffd_wp(entry); | |
18f39629 | 4814 | set_huge_pte_at(dst, addr, dst_pte, entry); |
bc70fbf2 PX |
4815 | } else if (unlikely(is_pte_marker(entry))) { |
4816 | /* | |
4817 | * We copy the pte marker only if the dst vma has | |
4818 | * uffd-wp enabled. | |
4819 | */ | |
4820 | if (userfaultfd_wp(dst_vma)) | |
4821 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
4a705fef | 4822 | } else { |
4eae4efa PX |
4823 | entry = huge_ptep_get(src_pte); |
4824 | ptepage = pte_page(entry); | |
4825 | get_page(ptepage); | |
4826 | ||
4827 | /* | |
fb3d824d DH |
4828 | * Failing to duplicate the anon rmap is a rare case |
4829 | * where we see pinned hugetlb pages while they're | |
4830 | * prone to COW. We need to do the COW earlier during | |
4831 | * fork. | |
4eae4efa PX |
4832 | * |
4833 | * When pre-allocating the page or copying data, we | |
4834 | * need to be without the pgtable locks since we could | |
4835 | * sleep during the process. | |
4836 | */ | |
fb3d824d DH |
4837 | if (!PageAnon(ptepage)) { |
4838 | page_dup_file_rmap(ptepage, true); | |
bc70fbf2 PX |
4839 | } else if (page_try_dup_anon_rmap(ptepage, true, |
4840 | src_vma)) { | |
4eae4efa PX |
4841 | pte_t src_pte_old = entry; |
4842 | struct page *new; | |
4843 | ||
4844 | spin_unlock(src_ptl); | |
4845 | spin_unlock(dst_ptl); | |
4846 | /* Do not use reserve as it's private owned */ | |
bc70fbf2 | 4847 | new = alloc_huge_page(dst_vma, addr, 1); |
4eae4efa PX |
4848 | if (IS_ERR(new)) { |
4849 | put_page(ptepage); | |
4850 | ret = PTR_ERR(new); | |
4851 | break; | |
4852 | } | |
bc70fbf2 | 4853 | copy_user_huge_page(new, ptepage, addr, dst_vma, |
4eae4efa PX |
4854 | npages); |
4855 | put_page(ptepage); | |
4856 | ||
4857 | /* Install the new huge page if src pte stable */ | |
4858 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
4859 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4860 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4861 | entry = huge_ptep_get(src_pte); | |
4862 | if (!pte_same(src_pte_old, entry)) { | |
bc70fbf2 | 4863 | restore_reserve_on_error(h, dst_vma, addr, |
846be085 | 4864 | new); |
4eae4efa PX |
4865 | put_page(new); |
4866 | /* dst_entry won't change as in child */ | |
4867 | goto again; | |
4868 | } | |
bc70fbf2 | 4869 | hugetlb_install_page(dst_vma, dst_pte, addr, new); |
4eae4efa PX |
4870 | spin_unlock(src_ptl); |
4871 | spin_unlock(dst_ptl); | |
4872 | continue; | |
4873 | } | |
4874 | ||
34ee645e | 4875 | if (cow) { |
0f10851e JG |
4876 | /* |
4877 | * No need to notify as we are downgrading page | |
4878 | * table protection not changing it to point | |
4879 | * to a new page. | |
4880 | * | |
ee65728e | 4881 | * See Documentation/mm/mmu_notifier.rst |
0f10851e | 4882 | */ |
7f2e9525 | 4883 | huge_ptep_set_wrprotect(src, addr, src_pte); |
84894e1c | 4884 | entry = huge_pte_wrprotect(entry); |
34ee645e | 4885 | } |
4eae4efa | 4886 | |
1c59827d | 4887 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 4888 | hugetlb_count_add(npages, dst); |
1c59827d | 4889 | } |
cb900f41 KS |
4890 | spin_unlock(src_ptl); |
4891 | spin_unlock(dst_ptl); | |
63551ae0 | 4892 | } |
63551ae0 | 4893 | |
623a1ddf DH |
4894 | if (cow) { |
4895 | raw_write_seqcount_end(&src->write_protect_seq); | |
ac46d4f3 | 4896 | mmu_notifier_invalidate_range_end(&range); |
623a1ddf | 4897 | } else { |
c0d0381a | 4898 | i_mmap_unlock_read(mapping); |
623a1ddf | 4899 | } |
e8569dd2 AS |
4900 | |
4901 | return ret; | |
63551ae0 DG |
4902 | } |
4903 | ||
550a7d60 | 4904 | static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, |
db110a99 | 4905 | unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) |
550a7d60 MA |
4906 | { |
4907 | struct hstate *h = hstate_vma(vma); | |
4908 | struct mm_struct *mm = vma->vm_mm; | |
550a7d60 | 4909 | spinlock_t *src_ptl, *dst_ptl; |
db110a99 | 4910 | pte_t pte; |
550a7d60 | 4911 | |
550a7d60 MA |
4912 | dst_ptl = huge_pte_lock(h, mm, dst_pte); |
4913 | src_ptl = huge_pte_lockptr(h, mm, src_pte); | |
4914 | ||
4915 | /* | |
4916 | * We don't have to worry about the ordering of src and dst ptlocks | |
4917 | * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. | |
4918 | */ | |
4919 | if (src_ptl != dst_ptl) | |
4920 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4921 | ||
4922 | pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); | |
4923 | set_huge_pte_at(mm, new_addr, dst_pte, pte); | |
4924 | ||
4925 | if (src_ptl != dst_ptl) | |
4926 | spin_unlock(src_ptl); | |
4927 | spin_unlock(dst_ptl); | |
4928 | } | |
4929 | ||
4930 | int move_hugetlb_page_tables(struct vm_area_struct *vma, | |
4931 | struct vm_area_struct *new_vma, | |
4932 | unsigned long old_addr, unsigned long new_addr, | |
4933 | unsigned long len) | |
4934 | { | |
4935 | struct hstate *h = hstate_vma(vma); | |
4936 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4937 | unsigned long sz = huge_page_size(h); | |
4938 | struct mm_struct *mm = vma->vm_mm; | |
4939 | unsigned long old_end = old_addr + len; | |
e95a9851 | 4940 | unsigned long last_addr_mask; |
550a7d60 MA |
4941 | pte_t *src_pte, *dst_pte; |
4942 | struct mmu_notifier_range range; | |
3d0b95cd | 4943 | bool shared_pmd = false; |
550a7d60 MA |
4944 | |
4945 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, | |
4946 | old_end); | |
4947 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
3d0b95cd BW |
4948 | /* |
4949 | * In case of shared PMDs, we should cover the maximum possible | |
4950 | * range. | |
4951 | */ | |
4952 | flush_cache_range(vma, range.start, range.end); | |
4953 | ||
550a7d60 | 4954 | mmu_notifier_invalidate_range_start(&range); |
e95a9851 | 4955 | last_addr_mask = hugetlb_mask_last_page(h); |
550a7d60 MA |
4956 | /* Prevent race with file truncation */ |
4957 | i_mmap_lock_write(mapping); | |
4958 | for (; old_addr < old_end; old_addr += sz, new_addr += sz) { | |
4959 | src_pte = huge_pte_offset(mm, old_addr, sz); | |
e95a9851 MK |
4960 | if (!src_pte) { |
4961 | old_addr |= last_addr_mask; | |
4962 | new_addr |= last_addr_mask; | |
550a7d60 | 4963 | continue; |
e95a9851 | 4964 | } |
550a7d60 MA |
4965 | if (huge_pte_none(huge_ptep_get(src_pte))) |
4966 | continue; | |
4967 | ||
4ddb4d91 | 4968 | if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { |
3d0b95cd | 4969 | shared_pmd = true; |
4ddb4d91 MK |
4970 | old_addr |= last_addr_mask; |
4971 | new_addr |= last_addr_mask; | |
550a7d60 | 4972 | continue; |
3d0b95cd | 4973 | } |
550a7d60 MA |
4974 | |
4975 | dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); | |
4976 | if (!dst_pte) | |
4977 | break; | |
4978 | ||
db110a99 | 4979 | move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); |
550a7d60 | 4980 | } |
3d0b95cd BW |
4981 | |
4982 | if (shared_pmd) | |
4983 | flush_tlb_range(vma, range.start, range.end); | |
4984 | else | |
4985 | flush_tlb_range(vma, old_end - len, old_end); | |
550a7d60 | 4986 | mmu_notifier_invalidate_range_end(&range); |
13e4ad2c | 4987 | i_mmap_unlock_write(mapping); |
550a7d60 MA |
4988 | |
4989 | return len + old_addr - old_end; | |
4990 | } | |
4991 | ||
73c54763 PX |
4992 | static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
4993 | unsigned long start, unsigned long end, | |
05e90bd0 | 4994 | struct page *ref_page, zap_flags_t zap_flags) |
63551ae0 DG |
4995 | { |
4996 | struct mm_struct *mm = vma->vm_mm; | |
4997 | unsigned long address; | |
c7546f8f | 4998 | pte_t *ptep; |
63551ae0 | 4999 | pte_t pte; |
cb900f41 | 5000 | spinlock_t *ptl; |
63551ae0 | 5001 | struct page *page; |
a5516438 AK |
5002 | struct hstate *h = hstate_vma(vma); |
5003 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 5004 | struct mmu_notifier_range range; |
e95a9851 | 5005 | unsigned long last_addr_mask; |
a4a118f2 | 5006 | bool force_flush = false; |
a5516438 | 5007 | |
63551ae0 | 5008 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
5009 | BUG_ON(start & ~huge_page_mask(h)); |
5010 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 5011 | |
07e32661 AK |
5012 | /* |
5013 | * This is a hugetlb vma, all the pte entries should point | |
5014 | * to huge page. | |
5015 | */ | |
ed6a7935 | 5016 | tlb_change_page_size(tlb, sz); |
24669e58 | 5017 | tlb_start_vma(tlb, vma); |
dff11abe MK |
5018 | |
5019 | /* | |
5020 | * If sharing possible, alert mmu notifiers of worst case. | |
5021 | */ | |
6f4f13e8 JG |
5022 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
5023 | end); | |
ac46d4f3 JG |
5024 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
5025 | mmu_notifier_invalidate_range_start(&range); | |
e95a9851 | 5026 | last_addr_mask = hugetlb_mask_last_page(h); |
569f48b8 | 5027 | address = start; |
569f48b8 | 5028 | for (; address < end; address += sz) { |
7868a208 | 5029 | ptep = huge_pte_offset(mm, address, sz); |
e95a9851 MK |
5030 | if (!ptep) { |
5031 | address |= last_addr_mask; | |
c7546f8f | 5032 | continue; |
e95a9851 | 5033 | } |
c7546f8f | 5034 | |
cb900f41 | 5035 | ptl = huge_pte_lock(h, mm, ptep); |
4ddb4d91 | 5036 | if (huge_pmd_unshare(mm, vma, address, ptep)) { |
31d49da5 | 5037 | spin_unlock(ptl); |
a4a118f2 NA |
5038 | tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); |
5039 | force_flush = true; | |
4ddb4d91 | 5040 | address |= last_addr_mask; |
31d49da5 AK |
5041 | continue; |
5042 | } | |
39dde65c | 5043 | |
6629326b | 5044 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
5045 | if (huge_pte_none(pte)) { |
5046 | spin_unlock(ptl); | |
5047 | continue; | |
5048 | } | |
6629326b HD |
5049 | |
5050 | /* | |
9fbc1f63 NH |
5051 | * Migrating hugepage or HWPoisoned hugepage is already |
5052 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 5053 | */ |
9fbc1f63 | 5054 | if (unlikely(!pte_present(pte))) { |
05e90bd0 PX |
5055 | /* |
5056 | * If the pte was wr-protected by uffd-wp in any of the | |
5057 | * swap forms, meanwhile the caller does not want to | |
5058 | * drop the uffd-wp bit in this zap, then replace the | |
5059 | * pte with a marker. | |
5060 | */ | |
5061 | if (pte_swp_uffd_wp_any(pte) && | |
5062 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5063 | set_huge_pte_at(mm, address, ptep, | |
5064 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5065 | else | |
5066 | huge_pte_clear(mm, address, ptep, sz); | |
31d49da5 AK |
5067 | spin_unlock(ptl); |
5068 | continue; | |
8c4894c6 | 5069 | } |
6629326b HD |
5070 | |
5071 | page = pte_page(pte); | |
04f2cbe3 MG |
5072 | /* |
5073 | * If a reference page is supplied, it is because a specific | |
5074 | * page is being unmapped, not a range. Ensure the page we | |
5075 | * are about to unmap is the actual page of interest. | |
5076 | */ | |
5077 | if (ref_page) { | |
31d49da5 AK |
5078 | if (page != ref_page) { |
5079 | spin_unlock(ptl); | |
5080 | continue; | |
5081 | } | |
04f2cbe3 MG |
5082 | /* |
5083 | * Mark the VMA as having unmapped its page so that | |
5084 | * future faults in this VMA will fail rather than | |
5085 | * looking like data was lost | |
5086 | */ | |
5087 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
5088 | } | |
5089 | ||
c7546f8f | 5090 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 5091 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 5092 | if (huge_pte_dirty(pte)) |
6649a386 | 5093 | set_page_dirty(page); |
05e90bd0 PX |
5094 | /* Leave a uffd-wp pte marker if needed */ |
5095 | if (huge_pte_uffd_wp(pte) && | |
5096 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5097 | set_huge_pte_at(mm, address, ptep, | |
5098 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5d317b2b | 5099 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
cea86fe2 | 5100 | page_remove_rmap(page, vma, true); |
31d49da5 | 5101 | |
cb900f41 | 5102 | spin_unlock(ptl); |
e77b0852 | 5103 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
5104 | /* |
5105 | * Bail out after unmapping reference page if supplied | |
5106 | */ | |
5107 | if (ref_page) | |
5108 | break; | |
fe1668ae | 5109 | } |
ac46d4f3 | 5110 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 5111 | tlb_end_vma(tlb, vma); |
a4a118f2 NA |
5112 | |
5113 | /* | |
5114 | * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We | |
5115 | * could defer the flush until now, since by holding i_mmap_rwsem we | |
5116 | * guaranteed that the last refernece would not be dropped. But we must | |
5117 | * do the flushing before we return, as otherwise i_mmap_rwsem will be | |
5118 | * dropped and the last reference to the shared PMDs page might be | |
5119 | * dropped as well. | |
5120 | * | |
5121 | * In theory we could defer the freeing of the PMD pages as well, but | |
5122 | * huge_pmd_unshare() relies on the exact page_count for the PMD page to | |
5123 | * detect sharing, so we cannot defer the release of the page either. | |
5124 | * Instead, do flush now. | |
5125 | */ | |
5126 | if (force_flush) | |
5127 | tlb_flush_mmu_tlbonly(tlb); | |
1da177e4 | 5128 | } |
63551ae0 | 5129 | |
d833352a MG |
5130 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
5131 | struct vm_area_struct *vma, unsigned long start, | |
05e90bd0 PX |
5132 | unsigned long end, struct page *ref_page, |
5133 | zap_flags_t zap_flags) | |
d833352a | 5134 | { |
05e90bd0 | 5135 | __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); |
d833352a MG |
5136 | |
5137 | /* | |
5138 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
5139 | * test will fail on a vma being torn down, and not grab a page table | |
5140 | * on its way out. We're lucky that the flag has such an appropriate | |
5141 | * name, and can in fact be safely cleared here. We could clear it | |
5142 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 5143 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 5144 | * because in the context this is called, the VMA is about to be |
c8c06efa | 5145 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
5146 | */ |
5147 | vma->vm_flags &= ~VM_MAYSHARE; | |
5148 | } | |
5149 | ||
502717f4 | 5150 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
05e90bd0 PX |
5151 | unsigned long end, struct page *ref_page, |
5152 | zap_flags_t zap_flags) | |
502717f4 | 5153 | { |
24669e58 | 5154 | struct mmu_gather tlb; |
dff11abe | 5155 | |
a72afd87 | 5156 | tlb_gather_mmu(&tlb, vma->vm_mm); |
05e90bd0 | 5157 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); |
ae8eba8b | 5158 | tlb_finish_mmu(&tlb); |
502717f4 CK |
5159 | } |
5160 | ||
04f2cbe3 MG |
5161 | /* |
5162 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 5163 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
5164 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
5165 | * same region. | |
5166 | */ | |
2f4612af DB |
5167 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
5168 | struct page *page, unsigned long address) | |
04f2cbe3 | 5169 | { |
7526674d | 5170 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
5171 | struct vm_area_struct *iter_vma; |
5172 | struct address_space *mapping; | |
04f2cbe3 MG |
5173 | pgoff_t pgoff; |
5174 | ||
5175 | /* | |
5176 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
5177 | * from page cache lookup which is in HPAGE_SIZE units. | |
5178 | */ | |
7526674d | 5179 | address = address & huge_page_mask(h); |
36e4f20a MH |
5180 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
5181 | vma->vm_pgoff; | |
93c76a3d | 5182 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 5183 | |
4eb2b1dc MG |
5184 | /* |
5185 | * Take the mapping lock for the duration of the table walk. As | |
5186 | * this mapping should be shared between all the VMAs, | |
5187 | * __unmap_hugepage_range() is called as the lock is already held | |
5188 | */ | |
83cde9e8 | 5189 | i_mmap_lock_write(mapping); |
6b2dbba8 | 5190 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
5191 | /* Do not unmap the current VMA */ |
5192 | if (iter_vma == vma) | |
5193 | continue; | |
5194 | ||
2f84a899 MG |
5195 | /* |
5196 | * Shared VMAs have their own reserves and do not affect | |
5197 | * MAP_PRIVATE accounting but it is possible that a shared | |
5198 | * VMA is using the same page so check and skip such VMAs. | |
5199 | */ | |
5200 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
5201 | continue; | |
5202 | ||
04f2cbe3 MG |
5203 | /* |
5204 | * Unmap the page from other VMAs without their own reserves. | |
5205 | * They get marked to be SIGKILLed if they fault in these | |
5206 | * areas. This is because a future no-page fault on this VMA | |
5207 | * could insert a zeroed page instead of the data existing | |
5208 | * from the time of fork. This would look like data corruption | |
5209 | */ | |
5210 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 | 5211 | unmap_hugepage_range(iter_vma, address, |
05e90bd0 | 5212 | address + huge_page_size(h), page, 0); |
04f2cbe3 | 5213 | } |
83cde9e8 | 5214 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
5215 | } |
5216 | ||
0fe6e20b | 5217 | /* |
c89357e2 | 5218 | * hugetlb_wp() should be called with page lock of the original hugepage held. |
aa6d2e8c | 5219 | * Called with hugetlb_fault_mutex_table held and pte_page locked so we |
ef009b25 MH |
5220 | * cannot race with other handlers or page migration. |
5221 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 5222 | */ |
c89357e2 DH |
5223 | static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, |
5224 | unsigned long address, pte_t *ptep, unsigned int flags, | |
3999f52e | 5225 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 5226 | { |
c89357e2 | 5227 | const bool unshare = flags & FAULT_FLAG_UNSHARE; |
3999f52e | 5228 | pte_t pte; |
a5516438 | 5229 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 5230 | struct page *old_page, *new_page; |
2b740303 SJ |
5231 | int outside_reserve = 0; |
5232 | vm_fault_t ret = 0; | |
974e6d66 | 5233 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 5234 | struct mmu_notifier_range range; |
1e8f889b | 5235 | |
c89357e2 DH |
5236 | VM_BUG_ON(unshare && (flags & FOLL_WRITE)); |
5237 | VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); | |
5238 | ||
3999f52e | 5239 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
5240 | old_page = pte_page(pte); |
5241 | ||
662ce1dc YY |
5242 | delayacct_wpcopy_start(); |
5243 | ||
04f2cbe3 | 5244 | retry_avoidcopy: |
c89357e2 DH |
5245 | /* |
5246 | * If no-one else is actually using this page, we're the exclusive | |
5247 | * owner and can reuse this page. | |
5248 | */ | |
37a2140d | 5249 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
c89357e2 DH |
5250 | if (!PageAnonExclusive(old_page)) |
5251 | page_move_anon_rmap(old_page, vma); | |
5252 | if (likely(!unshare)) | |
5253 | set_huge_ptep_writable(vma, haddr, ptep); | |
662ce1dc YY |
5254 | |
5255 | delayacct_wpcopy_end(); | |
83c54070 | 5256 | return 0; |
1e8f889b | 5257 | } |
6c287605 DH |
5258 | VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), |
5259 | old_page); | |
1e8f889b | 5260 | |
04f2cbe3 MG |
5261 | /* |
5262 | * If the process that created a MAP_PRIVATE mapping is about to | |
5263 | * perform a COW due to a shared page count, attempt to satisfy | |
5264 | * the allocation without using the existing reserves. The pagecache | |
5265 | * page is used to determine if the reserve at this address was | |
5266 | * consumed or not. If reserves were used, a partial faulted mapping | |
5267 | * at the time of fork() could consume its reserves on COW instead | |
5268 | * of the full address range. | |
5269 | */ | |
5944d011 | 5270 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
5271 | old_page != pagecache_page) |
5272 | outside_reserve = 1; | |
5273 | ||
09cbfeaf | 5274 | get_page(old_page); |
b76c8cfb | 5275 | |
ad4404a2 DB |
5276 | /* |
5277 | * Drop page table lock as buddy allocator may be called. It will | |
5278 | * be acquired again before returning to the caller, as expected. | |
5279 | */ | |
cb900f41 | 5280 | spin_unlock(ptl); |
5b7a1d40 | 5281 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 5282 | |
2fc39cec | 5283 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
5284 | /* |
5285 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
5286 | * it is due to references held by a child and an insufficient | |
5287 | * huge page pool. To guarantee the original mappers | |
5288 | * reliability, unmap the page from child processes. The child | |
5289 | * may get SIGKILLed if it later faults. | |
5290 | */ | |
5291 | if (outside_reserve) { | |
e7dd91c4 MK |
5292 | struct address_space *mapping = vma->vm_file->f_mapping; |
5293 | pgoff_t idx; | |
5294 | u32 hash; | |
5295 | ||
09cbfeaf | 5296 | put_page(old_page); |
04f2cbe3 | 5297 | BUG_ON(huge_pte_none(pte)); |
e7dd91c4 MK |
5298 | /* |
5299 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
5300 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
5301 | * in write mode. Dropping i_mmap_rwsem in read mode | |
5302 | * here is OK as COW mappings do not interact with | |
5303 | * PMD sharing. | |
5304 | * | |
5305 | * Reacquire both after unmap operation. | |
5306 | */ | |
5307 | idx = vma_hugecache_offset(h, vma, haddr); | |
5308 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5309 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5310 | i_mmap_unlock_read(mapping); | |
5311 | ||
5b7a1d40 | 5312 | unmap_ref_private(mm, vma, old_page, haddr); |
e7dd91c4 MK |
5313 | |
5314 | i_mmap_lock_read(mapping); | |
5315 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
2f4612af | 5316 | spin_lock(ptl); |
5b7a1d40 | 5317 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
5318 | if (likely(ptep && |
5319 | pte_same(huge_ptep_get(ptep), pte))) | |
5320 | goto retry_avoidcopy; | |
5321 | /* | |
5322 | * race occurs while re-acquiring page table | |
5323 | * lock, and our job is done. | |
5324 | */ | |
662ce1dc | 5325 | delayacct_wpcopy_end(); |
2f4612af | 5326 | return 0; |
04f2cbe3 MG |
5327 | } |
5328 | ||
2b740303 | 5329 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 5330 | goto out_release_old; |
1e8f889b DG |
5331 | } |
5332 | ||
0fe6e20b NH |
5333 | /* |
5334 | * When the original hugepage is shared one, it does not have | |
5335 | * anon_vma prepared. | |
5336 | */ | |
44e2aa93 | 5337 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
5338 | ret = VM_FAULT_OOM; |
5339 | goto out_release_all; | |
44e2aa93 | 5340 | } |
0fe6e20b | 5341 | |
974e6d66 | 5342 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 5343 | pages_per_huge_page(h)); |
0ed361de | 5344 | __SetPageUptodate(new_page); |
1e8f889b | 5345 | |
7269f999 | 5346 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 5347 | haddr + huge_page_size(h)); |
ac46d4f3 | 5348 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 5349 | |
b76c8cfb | 5350 | /* |
cb900f41 | 5351 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
5352 | * before the page tables are altered |
5353 | */ | |
cb900f41 | 5354 | spin_lock(ptl); |
5b7a1d40 | 5355 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 5356 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 5357 | ClearHPageRestoreReserve(new_page); |
07443a85 | 5358 | |
c89357e2 | 5359 | /* Break COW or unshare */ |
5b7a1d40 | 5360 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 5361 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
cea86fe2 | 5362 | page_remove_rmap(old_page, vma, true); |
5b7a1d40 | 5363 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
1eba86c0 | 5364 | set_huge_pte_at(mm, haddr, ptep, |
c89357e2 | 5365 | make_huge_pte(vma, new_page, !unshare)); |
8f251a3d | 5366 | SetHPageMigratable(new_page); |
1e8f889b DG |
5367 | /* Make the old page be freed below */ |
5368 | new_page = old_page; | |
5369 | } | |
cb900f41 | 5370 | spin_unlock(ptl); |
ac46d4f3 | 5371 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 5372 | out_release_all: |
c89357e2 DH |
5373 | /* |
5374 | * No restore in case of successful pagetable update (Break COW or | |
5375 | * unshare) | |
5376 | */ | |
c7b1850d MK |
5377 | if (new_page != old_page) |
5378 | restore_reserve_on_error(h, vma, haddr, new_page); | |
09cbfeaf | 5379 | put_page(new_page); |
ad4404a2 | 5380 | out_release_old: |
09cbfeaf | 5381 | put_page(old_page); |
8312034f | 5382 | |
ad4404a2 | 5383 | spin_lock(ptl); /* Caller expects lock to be held */ |
662ce1dc YY |
5384 | |
5385 | delayacct_wpcopy_end(); | |
ad4404a2 | 5386 | return ret; |
1e8f889b DG |
5387 | } |
5388 | ||
04f2cbe3 | 5389 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
5390 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
5391 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
5392 | { |
5393 | struct address_space *mapping; | |
e7c4b0bf | 5394 | pgoff_t idx; |
04f2cbe3 MG |
5395 | |
5396 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 5397 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
5398 | |
5399 | return find_lock_page(mapping, idx); | |
5400 | } | |
5401 | ||
3ae77f43 HD |
5402 | /* |
5403 | * Return whether there is a pagecache page to back given address within VMA. | |
5404 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
5405 | */ | |
5406 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
5407 | struct vm_area_struct *vma, unsigned long address) |
5408 | { | |
5409 | struct address_space *mapping; | |
5410 | pgoff_t idx; | |
5411 | struct page *page; | |
5412 | ||
5413 | mapping = vma->vm_file->f_mapping; | |
5414 | idx = vma_hugecache_offset(h, vma, address); | |
5415 | ||
5416 | page = find_get_page(mapping, idx); | |
5417 | if (page) | |
5418 | put_page(page); | |
5419 | return page != NULL; | |
5420 | } | |
5421 | ||
ab76ad54 MK |
5422 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
5423 | pgoff_t idx) | |
5424 | { | |
d9ef44de | 5425 | struct folio *folio = page_folio(page); |
ab76ad54 MK |
5426 | struct inode *inode = mapping->host; |
5427 | struct hstate *h = hstate_inode(inode); | |
d9ef44de | 5428 | int err; |
ab76ad54 | 5429 | |
d9ef44de MWO |
5430 | __folio_set_locked(folio); |
5431 | err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); | |
5432 | ||
5433 | if (unlikely(err)) { | |
5434 | __folio_clear_locked(folio); | |
ab76ad54 | 5435 | return err; |
d9ef44de | 5436 | } |
d6995da3 | 5437 | ClearHPageRestoreReserve(page); |
ab76ad54 | 5438 | |
22146c3c | 5439 | /* |
d9ef44de | 5440 | * mark folio dirty so that it will not be removed from cache/file |
22146c3c MK |
5441 | * by non-hugetlbfs specific code paths. |
5442 | */ | |
d9ef44de | 5443 | folio_mark_dirty(folio); |
22146c3c | 5444 | |
ab76ad54 MK |
5445 | spin_lock(&inode->i_lock); |
5446 | inode->i_blocks += blocks_per_huge_page(h); | |
5447 | spin_unlock(&inode->i_lock); | |
5448 | return 0; | |
5449 | } | |
5450 | ||
7677f7fd AR |
5451 | static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, |
5452 | struct address_space *mapping, | |
5453 | pgoff_t idx, | |
5454 | unsigned int flags, | |
5455 | unsigned long haddr, | |
824ddc60 | 5456 | unsigned long addr, |
7677f7fd AR |
5457 | unsigned long reason) |
5458 | { | |
5459 | vm_fault_t ret; | |
5460 | u32 hash; | |
5461 | struct vm_fault vmf = { | |
5462 | .vma = vma, | |
5463 | .address = haddr, | |
824ddc60 | 5464 | .real_address = addr, |
7677f7fd AR |
5465 | .flags = flags, |
5466 | ||
5467 | /* | |
5468 | * Hard to debug if it ends up being | |
5469 | * used by a callee that assumes | |
5470 | * something about the other | |
5471 | * uninitialized fields... same as in | |
5472 | * memory.c | |
5473 | */ | |
5474 | }; | |
5475 | ||
5476 | /* | |
5477 | * hugetlb_fault_mutex and i_mmap_rwsem must be | |
5478 | * dropped before handling userfault. Reacquire | |
5479 | * after handling fault to make calling code simpler. | |
5480 | */ | |
5481 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5482 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5483 | i_mmap_unlock_read(mapping); | |
5484 | ret = handle_userfault(&vmf, reason); | |
5485 | i_mmap_lock_read(mapping); | |
5486 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
5487 | ||
5488 | return ret; | |
5489 | } | |
5490 | ||
2b740303 SJ |
5491 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
5492 | struct vm_area_struct *vma, | |
5493 | struct address_space *mapping, pgoff_t idx, | |
c64e912c PX |
5494 | unsigned long address, pte_t *ptep, |
5495 | pte_t old_pte, unsigned int flags) | |
ac9b9c66 | 5496 | { |
a5516438 | 5497 | struct hstate *h = hstate_vma(vma); |
2b740303 | 5498 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 5499 | int anon_rmap = 0; |
4c887265 | 5500 | unsigned long size; |
4c887265 | 5501 | struct page *page; |
1e8f889b | 5502 | pte_t new_pte; |
cb900f41 | 5503 | spinlock_t *ptl; |
285b8dca | 5504 | unsigned long haddr = address & huge_page_mask(h); |
c7b1850d | 5505 | bool new_page, new_pagecache_page = false; |
4c887265 | 5506 | |
04f2cbe3 MG |
5507 | /* |
5508 | * Currently, we are forced to kill the process in the event the | |
5509 | * original mapper has unmapped pages from the child due to a failed | |
c89357e2 DH |
5510 | * COW/unsharing. Warn that such a situation has occurred as it may not |
5511 | * be obvious. | |
04f2cbe3 MG |
5512 | */ |
5513 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 5514 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 5515 | current->pid); |
04f2cbe3 MG |
5516 | return ret; |
5517 | } | |
5518 | ||
4c887265 | 5519 | /* |
87bf91d3 MK |
5520 | * We can not race with truncation due to holding i_mmap_rwsem. |
5521 | * i_size is modified when holding i_mmap_rwsem, so check here | |
5522 | * once for faults beyond end of file. | |
4c887265 | 5523 | */ |
87bf91d3 MK |
5524 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5525 | if (idx >= size) | |
5526 | goto out; | |
5527 | ||
6bda666a | 5528 | retry: |
c7b1850d | 5529 | new_page = false; |
6bda666a CL |
5530 | page = find_lock_page(mapping, idx); |
5531 | if (!page) { | |
7677f7fd | 5532 | /* Check for page in userfault range */ |
1a1aad8a | 5533 | if (userfaultfd_missing(vma)) { |
7677f7fd | 5534 | ret = hugetlb_handle_userfault(vma, mapping, idx, |
824ddc60 | 5535 | flags, haddr, address, |
7677f7fd | 5536 | VM_UFFD_MISSING); |
1a1aad8a MK |
5537 | goto out; |
5538 | } | |
5539 | ||
285b8dca | 5540 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 5541 | if (IS_ERR(page)) { |
4643d67e MK |
5542 | /* |
5543 | * Returning error will result in faulting task being | |
5544 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
5545 | * tasks from racing to fault in the same page which | |
5546 | * could result in false unable to allocate errors. | |
5547 | * Page migration does not take the fault mutex, but | |
5548 | * does a clear then write of pte's under page table | |
5549 | * lock. Page fault code could race with migration, | |
5550 | * notice the clear pte and try to allocate a page | |
5551 | * here. Before returning error, get ptl and make | |
5552 | * sure there really is no pte entry. | |
5553 | */ | |
5554 | ptl = huge_pte_lock(h, mm, ptep); | |
d83e6c8a ML |
5555 | ret = 0; |
5556 | if (huge_pte_none(huge_ptep_get(ptep))) | |
5557 | ret = vmf_error(PTR_ERR(page)); | |
4643d67e | 5558 | spin_unlock(ptl); |
6bda666a CL |
5559 | goto out; |
5560 | } | |
47ad8475 | 5561 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 5562 | __SetPageUptodate(page); |
cb6acd01 | 5563 | new_page = true; |
ac9b9c66 | 5564 | |
f83a275d | 5565 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 5566 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
5567 | if (err) { |
5568 | put_page(page); | |
6bda666a CL |
5569 | if (err == -EEXIST) |
5570 | goto retry; | |
5571 | goto out; | |
5572 | } | |
c7b1850d | 5573 | new_pagecache_page = true; |
23be7468 | 5574 | } else { |
6bda666a | 5575 | lock_page(page); |
0fe6e20b NH |
5576 | if (unlikely(anon_vma_prepare(vma))) { |
5577 | ret = VM_FAULT_OOM; | |
5578 | goto backout_unlocked; | |
5579 | } | |
409eb8c2 | 5580 | anon_rmap = 1; |
23be7468 | 5581 | } |
0fe6e20b | 5582 | } else { |
998b4382 NH |
5583 | /* |
5584 | * If memory error occurs between mmap() and fault, some process | |
5585 | * don't have hwpoisoned swap entry for errored virtual address. | |
5586 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
5587 | */ | |
5588 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 5589 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5590 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
5591 | goto backout_unlocked; |
5592 | } | |
7677f7fd AR |
5593 | |
5594 | /* Check for page in userfault range. */ | |
5595 | if (userfaultfd_minor(vma)) { | |
5596 | unlock_page(page); | |
5597 | put_page(page); | |
5598 | ret = hugetlb_handle_userfault(vma, mapping, idx, | |
824ddc60 | 5599 | flags, haddr, address, |
7677f7fd AR |
5600 | VM_UFFD_MINOR); |
5601 | goto out; | |
5602 | } | |
6bda666a | 5603 | } |
1e8f889b | 5604 | |
57303d80 AW |
5605 | /* |
5606 | * If we are going to COW a private mapping later, we examine the | |
5607 | * pending reservations for this page now. This will ensure that | |
5608 | * any allocations necessary to record that reservation occur outside | |
5609 | * the spinlock. | |
5610 | */ | |
5e911373 | 5611 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 5612 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
5613 | ret = VM_FAULT_OOM; |
5614 | goto backout_unlocked; | |
5615 | } | |
5e911373 | 5616 | /* Just decrements count, does not deallocate */ |
285b8dca | 5617 | vma_end_reservation(h, vma, haddr); |
5e911373 | 5618 | } |
57303d80 | 5619 | |
8bea8052 | 5620 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 5621 | ret = 0; |
c64e912c PX |
5622 | /* If pte changed from under us, retry */ |
5623 | if (!pte_same(huge_ptep_get(ptep), old_pte)) | |
4c887265 AL |
5624 | goto backout; |
5625 | ||
07443a85 | 5626 | if (anon_rmap) { |
d6995da3 | 5627 | ClearHPageRestoreReserve(page); |
285b8dca | 5628 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 5629 | } else |
fb3d824d | 5630 | page_dup_file_rmap(page, true); |
1e8f889b DG |
5631 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
5632 | && (vma->vm_flags & VM_SHARED))); | |
c64e912c PX |
5633 | /* |
5634 | * If this pte was previously wr-protected, keep it wr-protected even | |
5635 | * if populated. | |
5636 | */ | |
5637 | if (unlikely(pte_marker_uffd_wp(old_pte))) | |
5638 | new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); | |
285b8dca | 5639 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 5640 | |
5d317b2b | 5641 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 5642 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 5643 | /* Optimization, do the COW without a second fault */ |
c89357e2 | 5644 | ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); |
1e8f889b DG |
5645 | } |
5646 | ||
cb900f41 | 5647 | spin_unlock(ptl); |
cb6acd01 MK |
5648 | |
5649 | /* | |
8f251a3d MK |
5650 | * Only set HPageMigratable in newly allocated pages. Existing pages |
5651 | * found in the pagecache may not have HPageMigratableset if they have | |
5652 | * been isolated for migration. | |
cb6acd01 MK |
5653 | */ |
5654 | if (new_page) | |
8f251a3d | 5655 | SetHPageMigratable(page); |
cb6acd01 | 5656 | |
4c887265 AL |
5657 | unlock_page(page); |
5658 | out: | |
ac9b9c66 | 5659 | return ret; |
4c887265 AL |
5660 | |
5661 | backout: | |
cb900f41 | 5662 | spin_unlock(ptl); |
2b26736c | 5663 | backout_unlocked: |
4c887265 | 5664 | unlock_page(page); |
c7b1850d MK |
5665 | /* restore reserve for newly allocated pages not in page cache */ |
5666 | if (new_page && !new_pagecache_page) | |
5667 | restore_reserve_on_error(h, vma, haddr, page); | |
4c887265 AL |
5668 | put_page(page); |
5669 | goto out; | |
ac9b9c66 HD |
5670 | } |
5671 | ||
8382d914 | 5672 | #ifdef CONFIG_SMP |
188b04a7 | 5673 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5674 | { |
5675 | unsigned long key[2]; | |
5676 | u32 hash; | |
5677 | ||
1b426bac MK |
5678 | key[0] = (unsigned long) mapping; |
5679 | key[1] = idx; | |
8382d914 | 5680 | |
55254636 | 5681 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
5682 | |
5683 | return hash & (num_fault_mutexes - 1); | |
5684 | } | |
5685 | #else | |
5686 | /* | |
6c26d310 | 5687 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
5688 | * return 0 and avoid the hashing overhead. |
5689 | */ | |
188b04a7 | 5690 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5691 | { |
5692 | return 0; | |
5693 | } | |
5694 | #endif | |
5695 | ||
2b740303 | 5696 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 5697 | unsigned long address, unsigned int flags) |
86e5216f | 5698 | { |
8382d914 | 5699 | pte_t *ptep, entry; |
cb900f41 | 5700 | spinlock_t *ptl; |
2b740303 | 5701 | vm_fault_t ret; |
8382d914 DB |
5702 | u32 hash; |
5703 | pgoff_t idx; | |
0fe6e20b | 5704 | struct page *page = NULL; |
57303d80 | 5705 | struct page *pagecache_page = NULL; |
a5516438 | 5706 | struct hstate *h = hstate_vma(vma); |
8382d914 | 5707 | struct address_space *mapping; |
0f792cf9 | 5708 | int need_wait_lock = 0; |
285b8dca | 5709 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 5710 | |
285b8dca | 5711 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 5712 | if (ptep) { |
c0d0381a MK |
5713 | /* |
5714 | * Since we hold no locks, ptep could be stale. That is | |
5715 | * OK as we are only making decisions based on content and | |
5716 | * not actually modifying content here. | |
5717 | */ | |
fd6a03ed | 5718 | entry = huge_ptep_get(ptep); |
290408d4 | 5719 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
ad1ac596 | 5720 | migration_entry_wait_huge(vma, ptep); |
290408d4 NH |
5721 | return 0; |
5722 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 5723 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5724 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
5725 | } |
5726 | ||
c0d0381a MK |
5727 | /* |
5728 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
5729 | * until finished with ptep. This serves two purposes: |
5730 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
5731 | * and making the ptep no longer valid. | |
5732 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
5733 | * |
5734 | * ptep could have already be assigned via huge_pte_offset. That | |
5735 | * is OK, as huge_pte_alloc will return the same value unless | |
5736 | * something has changed. | |
5737 | */ | |
8382d914 | 5738 | mapping = vma->vm_file->f_mapping; |
c0d0381a | 5739 | i_mmap_lock_read(mapping); |
aec44e0f | 5740 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); |
c0d0381a MK |
5741 | if (!ptep) { |
5742 | i_mmap_unlock_read(mapping); | |
5743 | return VM_FAULT_OOM; | |
5744 | } | |
8382d914 | 5745 | |
3935baa9 DG |
5746 | /* |
5747 | * Serialize hugepage allocation and instantiation, so that we don't | |
5748 | * get spurious allocation failures if two CPUs race to instantiate | |
5749 | * the same page in the page cache. | |
5750 | */ | |
c0d0381a | 5751 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 5752 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 5753 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 5754 | |
7f2e9525 | 5755 | entry = huge_ptep_get(ptep); |
c64e912c PX |
5756 | /* PTE markers should be handled the same way as none pte */ |
5757 | if (huge_pte_none_mostly(entry)) { | |
5758 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, | |
5759 | entry, flags); | |
b4d1d99f | 5760 | goto out_mutex; |
3935baa9 | 5761 | } |
86e5216f | 5762 | |
83c54070 | 5763 | ret = 0; |
1e8f889b | 5764 | |
0f792cf9 NH |
5765 | /* |
5766 | * entry could be a migration/hwpoison entry at this point, so this | |
5767 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
5768 | * an active hugepage in pagecache. This goto expects the 2nd page |
5769 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
5770 | * properly handle it. | |
0f792cf9 NH |
5771 | */ |
5772 | if (!pte_present(entry)) | |
5773 | goto out_mutex; | |
5774 | ||
57303d80 | 5775 | /* |
c89357e2 DH |
5776 | * If we are going to COW/unshare the mapping later, we examine the |
5777 | * pending reservations for this page now. This will ensure that any | |
57303d80 AW |
5778 | * allocations necessary to record that reservation occur outside the |
5779 | * spinlock. For private mappings, we also lookup the pagecache | |
5780 | * page now as it is used to determine if a reservation has been | |
5781 | * consumed. | |
5782 | */ | |
c89357e2 DH |
5783 | if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && |
5784 | !huge_pte_write(entry)) { | |
285b8dca | 5785 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 5786 | ret = VM_FAULT_OOM; |
b4d1d99f | 5787 | goto out_mutex; |
2b26736c | 5788 | } |
5e911373 | 5789 | /* Just decrements count, does not deallocate */ |
285b8dca | 5790 | vma_end_reservation(h, vma, haddr); |
57303d80 | 5791 | |
f83a275d | 5792 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 5793 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 5794 | vma, haddr); |
57303d80 AW |
5795 | } |
5796 | ||
0f792cf9 NH |
5797 | ptl = huge_pte_lock(h, mm, ptep); |
5798 | ||
c89357e2 | 5799 | /* Check for a racing update before calling hugetlb_wp() */ |
0f792cf9 NH |
5800 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
5801 | goto out_ptl; | |
5802 | ||
166f3ecc PX |
5803 | /* Handle userfault-wp first, before trying to lock more pages */ |
5804 | if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && | |
5805 | (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { | |
5806 | struct vm_fault vmf = { | |
5807 | .vma = vma, | |
5808 | .address = haddr, | |
5809 | .real_address = address, | |
5810 | .flags = flags, | |
5811 | }; | |
5812 | ||
5813 | spin_unlock(ptl); | |
5814 | if (pagecache_page) { | |
5815 | unlock_page(pagecache_page); | |
5816 | put_page(pagecache_page); | |
5817 | } | |
5818 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5819 | i_mmap_unlock_read(mapping); | |
5820 | return handle_userfault(&vmf, VM_UFFD_WP); | |
5821 | } | |
5822 | ||
56c9cfb1 | 5823 | /* |
c89357e2 | 5824 | * hugetlb_wp() requires page locks of pte_page(entry) and |
56c9cfb1 NH |
5825 | * pagecache_page, so here we need take the former one |
5826 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
5827 | */ |
5828 | page = pte_page(entry); | |
5829 | if (page != pagecache_page) | |
0f792cf9 NH |
5830 | if (!trylock_page(page)) { |
5831 | need_wait_lock = 1; | |
5832 | goto out_ptl; | |
5833 | } | |
b4d1d99f | 5834 | |
0f792cf9 | 5835 | get_page(page); |
b4d1d99f | 5836 | |
c89357e2 | 5837 | if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { |
106c992a | 5838 | if (!huge_pte_write(entry)) { |
c89357e2 DH |
5839 | ret = hugetlb_wp(mm, vma, address, ptep, flags, |
5840 | pagecache_page, ptl); | |
0f792cf9 | 5841 | goto out_put_page; |
c89357e2 DH |
5842 | } else if (likely(flags & FAULT_FLAG_WRITE)) { |
5843 | entry = huge_pte_mkdirty(entry); | |
b4d1d99f | 5844 | } |
b4d1d99f DG |
5845 | } |
5846 | entry = pte_mkyoung(entry); | |
285b8dca | 5847 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 5848 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 5849 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
5850 | out_put_page: |
5851 | if (page != pagecache_page) | |
5852 | unlock_page(page); | |
5853 | put_page(page); | |
cb900f41 KS |
5854 | out_ptl: |
5855 | spin_unlock(ptl); | |
57303d80 AW |
5856 | |
5857 | if (pagecache_page) { | |
5858 | unlock_page(pagecache_page); | |
5859 | put_page(pagecache_page); | |
5860 | } | |
b4d1d99f | 5861 | out_mutex: |
c672c7f2 | 5862 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 5863 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
5864 | /* |
5865 | * Generally it's safe to hold refcount during waiting page lock. But | |
5866 | * here we just wait to defer the next page fault to avoid busy loop and | |
5867 | * the page is not used after unlocked before returning from the current | |
5868 | * page fault. So we are safe from accessing freed page, even if we wait | |
5869 | * here without taking refcount. | |
5870 | */ | |
5871 | if (need_wait_lock) | |
5872 | wait_on_page_locked(page); | |
1e8f889b | 5873 | return ret; |
86e5216f AL |
5874 | } |
5875 | ||
714c1891 | 5876 | #ifdef CONFIG_USERFAULTFD |
8fb5debc MK |
5877 | /* |
5878 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
5879 | * modifications for huge pages. | |
5880 | */ | |
5881 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
5882 | pte_t *dst_pte, | |
5883 | struct vm_area_struct *dst_vma, | |
5884 | unsigned long dst_addr, | |
5885 | unsigned long src_addr, | |
f6191471 | 5886 | enum mcopy_atomic_mode mode, |
6041c691 PX |
5887 | struct page **pagep, |
5888 | bool wp_copy) | |
8fb5debc | 5889 | { |
f6191471 | 5890 | bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); |
8cc5fcbb MA |
5891 | struct hstate *h = hstate_vma(dst_vma); |
5892 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
5893 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
1e392147 | 5894 | unsigned long size; |
1c9e8def | 5895 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
5896 | pte_t _dst_pte; |
5897 | spinlock_t *ptl; | |
8cc5fcbb | 5898 | int ret = -ENOMEM; |
8fb5debc | 5899 | struct page *page; |
f6191471 | 5900 | int writable; |
cc30042d | 5901 | bool page_in_pagecache = false; |
8fb5debc | 5902 | |
f6191471 AR |
5903 | if (is_continue) { |
5904 | ret = -EFAULT; | |
5905 | page = find_lock_page(mapping, idx); | |
5906 | if (!page) | |
5907 | goto out; | |
cc30042d | 5908 | page_in_pagecache = true; |
f6191471 | 5909 | } else if (!*pagep) { |
d84cf06e MA |
5910 | /* If a page already exists, then it's UFFDIO_COPY for |
5911 | * a non-missing case. Return -EEXIST. | |
5912 | */ | |
5913 | if (vm_shared && | |
5914 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5915 | ret = -EEXIST; | |
5916 | goto out; | |
5917 | } | |
5918 | ||
8fb5debc | 5919 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
d84cf06e MA |
5920 | if (IS_ERR(page)) { |
5921 | ret = -ENOMEM; | |
8fb5debc | 5922 | goto out; |
d84cf06e | 5923 | } |
8fb5debc MK |
5924 | |
5925 | ret = copy_huge_page_from_user(page, | |
5926 | (const void __user *) src_addr, | |
810a56b9 | 5927 | pages_per_huge_page(h), false); |
8fb5debc | 5928 | |
c1e8d7c6 | 5929 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 5930 | if (unlikely(ret)) { |
9e368259 | 5931 | ret = -ENOENT; |
8cc5fcbb MA |
5932 | /* Free the allocated page which may have |
5933 | * consumed a reservation. | |
5934 | */ | |
5935 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
5936 | put_page(page); | |
5937 | ||
5938 | /* Allocate a temporary page to hold the copied | |
5939 | * contents. | |
5940 | */ | |
5941 | page = alloc_huge_page_vma(h, dst_vma, dst_addr); | |
5942 | if (!page) { | |
5943 | ret = -ENOMEM; | |
5944 | goto out; | |
5945 | } | |
8fb5debc | 5946 | *pagep = page; |
8cc5fcbb MA |
5947 | /* Set the outparam pagep and return to the caller to |
5948 | * copy the contents outside the lock. Don't free the | |
5949 | * page. | |
5950 | */ | |
8fb5debc MK |
5951 | goto out; |
5952 | } | |
5953 | } else { | |
8cc5fcbb MA |
5954 | if (vm_shared && |
5955 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5956 | put_page(*pagep); | |
5957 | ret = -EEXIST; | |
5958 | *pagep = NULL; | |
5959 | goto out; | |
5960 | } | |
5961 | ||
5962 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
5963 | if (IS_ERR(page)) { | |
da9a298f | 5964 | put_page(*pagep); |
8cc5fcbb MA |
5965 | ret = -ENOMEM; |
5966 | *pagep = NULL; | |
5967 | goto out; | |
5968 | } | |
34892366 MS |
5969 | copy_user_huge_page(page, *pagep, dst_addr, dst_vma, |
5970 | pages_per_huge_page(h)); | |
8cc5fcbb | 5971 | put_page(*pagep); |
8fb5debc MK |
5972 | *pagep = NULL; |
5973 | } | |
5974 | ||
5975 | /* | |
5976 | * The memory barrier inside __SetPageUptodate makes sure that | |
5977 | * preceding stores to the page contents become visible before | |
5978 | * the set_pte_at() write. | |
5979 | */ | |
5980 | __SetPageUptodate(page); | |
8fb5debc | 5981 | |
f6191471 AR |
5982 | /* Add shared, newly allocated pages to the page cache. */ |
5983 | if (vm_shared && !is_continue) { | |
1e392147 AA |
5984 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5985 | ret = -EFAULT; | |
5986 | if (idx >= size) | |
5987 | goto out_release_nounlock; | |
1c9e8def | 5988 | |
1e392147 AA |
5989 | /* |
5990 | * Serialization between remove_inode_hugepages() and | |
5991 | * huge_add_to_page_cache() below happens through the | |
5992 | * hugetlb_fault_mutex_table that here must be hold by | |
5993 | * the caller. | |
5994 | */ | |
1c9e8def MK |
5995 | ret = huge_add_to_page_cache(page, mapping, idx); |
5996 | if (ret) | |
5997 | goto out_release_nounlock; | |
cc30042d | 5998 | page_in_pagecache = true; |
1c9e8def MK |
5999 | } |
6000 | ||
8fb5debc MK |
6001 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
6002 | spin_lock(ptl); | |
6003 | ||
1e392147 AA |
6004 | /* |
6005 | * Recheck the i_size after holding PT lock to make sure not | |
6006 | * to leave any page mapped (as page_mapped()) beyond the end | |
6007 | * of the i_size (remove_inode_hugepages() is strict about | |
6008 | * enforcing that). If we bail out here, we'll also leave a | |
6009 | * page in the radix tree in the vm_shared case beyond the end | |
6010 | * of the i_size, but remove_inode_hugepages() will take care | |
6011 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
6012 | */ | |
6013 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
6014 | ret = -EFAULT; | |
6015 | if (idx >= size) | |
6016 | goto out_release_unlock; | |
6017 | ||
8fb5debc | 6018 | ret = -EEXIST; |
6041c691 PX |
6019 | /* |
6020 | * We allow to overwrite a pte marker: consider when both MISSING|WP | |
6021 | * registered, we firstly wr-protect a none pte which has no page cache | |
6022 | * page backing it, then access the page. | |
6023 | */ | |
6024 | if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) | |
8fb5debc MK |
6025 | goto out_release_unlock; |
6026 | ||
1c9e8def | 6027 | if (vm_shared) { |
fb3d824d | 6028 | page_dup_file_rmap(page, true); |
1c9e8def | 6029 | } else { |
d6995da3 | 6030 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
6031 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
6032 | } | |
8fb5debc | 6033 | |
6041c691 PX |
6034 | /* |
6035 | * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY | |
6036 | * with wp flag set, don't set pte write bit. | |
6037 | */ | |
6038 | if (wp_copy || (is_continue && !vm_shared)) | |
f6191471 AR |
6039 | writable = 0; |
6040 | else | |
6041 | writable = dst_vma->vm_flags & VM_WRITE; | |
6042 | ||
6043 | _dst_pte = make_huge_pte(dst_vma, page, writable); | |
6041c691 PX |
6044 | /* |
6045 | * Always mark UFFDIO_COPY page dirty; note that this may not be | |
6046 | * extremely important for hugetlbfs for now since swapping is not | |
6047 | * supported, but we should still be clear in that this page cannot be | |
6048 | * thrown away at will, even if write bit not set. | |
6049 | */ | |
6050 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
8fb5debc MK |
6051 | _dst_pte = pte_mkyoung(_dst_pte); |
6052 | ||
6041c691 PX |
6053 | if (wp_copy) |
6054 | _dst_pte = huge_pte_mkuffd_wp(_dst_pte); | |
6055 | ||
8fb5debc MK |
6056 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); |
6057 | ||
8fb5debc MK |
6058 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); |
6059 | ||
6060 | /* No need to invalidate - it was non-present before */ | |
6061 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
6062 | ||
6063 | spin_unlock(ptl); | |
f6191471 AR |
6064 | if (!is_continue) |
6065 | SetHPageMigratable(page); | |
6066 | if (vm_shared || is_continue) | |
1c9e8def | 6067 | unlock_page(page); |
8fb5debc MK |
6068 | ret = 0; |
6069 | out: | |
6070 | return ret; | |
6071 | out_release_unlock: | |
6072 | spin_unlock(ptl); | |
f6191471 | 6073 | if (vm_shared || is_continue) |
1c9e8def | 6074 | unlock_page(page); |
5af10dfd | 6075 | out_release_nounlock: |
cc30042d | 6076 | if (!page_in_pagecache) |
c7b1850d | 6077 | restore_reserve_on_error(h, dst_vma, dst_addr, page); |
8fb5debc MK |
6078 | put_page(page); |
6079 | goto out; | |
6080 | } | |
714c1891 | 6081 | #endif /* CONFIG_USERFAULTFD */ |
8fb5debc | 6082 | |
82e5d378 JM |
6083 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
6084 | int refs, struct page **pages, | |
6085 | struct vm_area_struct **vmas) | |
6086 | { | |
6087 | int nr; | |
6088 | ||
6089 | for (nr = 0; nr < refs; nr++) { | |
6090 | if (likely(pages)) | |
6091 | pages[nr] = mem_map_offset(page, nr); | |
6092 | if (vmas) | |
6093 | vmas[nr] = vma; | |
6094 | } | |
6095 | } | |
6096 | ||
a7f22660 DH |
6097 | static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, |
6098 | bool *unshare) | |
6099 | { | |
6100 | pte_t pteval = huge_ptep_get(pte); | |
6101 | ||
6102 | *unshare = false; | |
6103 | if (is_swap_pte(pteval)) | |
6104 | return true; | |
6105 | if (huge_pte_write(pteval)) | |
6106 | return false; | |
6107 | if (flags & FOLL_WRITE) | |
6108 | return true; | |
6109 | if (gup_must_unshare(flags, pte_page(pteval))) { | |
6110 | *unshare = true; | |
6111 | return true; | |
6112 | } | |
6113 | return false; | |
6114 | } | |
6115 | ||
28a35716 ML |
6116 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
6117 | struct page **pages, struct vm_area_struct **vmas, | |
6118 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 6119 | long i, unsigned int flags, int *locked) |
63551ae0 | 6120 | { |
d5d4b0aa CK |
6121 | unsigned long pfn_offset; |
6122 | unsigned long vaddr = *position; | |
28a35716 | 6123 | unsigned long remainder = *nr_pages; |
a5516438 | 6124 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 6125 | int err = -EFAULT, refs; |
63551ae0 | 6126 | |
63551ae0 | 6127 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 6128 | pte_t *pte; |
cb900f41 | 6129 | spinlock_t *ptl = NULL; |
a7f22660 | 6130 | bool unshare = false; |
2a15efc9 | 6131 | int absent; |
4c887265 | 6132 | struct page *page; |
63551ae0 | 6133 | |
02057967 DR |
6134 | /* |
6135 | * If we have a pending SIGKILL, don't keep faulting pages and | |
6136 | * potentially allocating memory. | |
6137 | */ | |
fa45f116 | 6138 | if (fatal_signal_pending(current)) { |
02057967 DR |
6139 | remainder = 0; |
6140 | break; | |
6141 | } | |
6142 | ||
4c887265 AL |
6143 | /* |
6144 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 6145 | * each hugepage. We have to make sure we get the |
4c887265 | 6146 | * first, for the page indexing below to work. |
cb900f41 KS |
6147 | * |
6148 | * Note that page table lock is not held when pte is null. | |
4c887265 | 6149 | */ |
7868a208 PA |
6150 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
6151 | huge_page_size(h)); | |
cb900f41 KS |
6152 | if (pte) |
6153 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
6154 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
6155 | ||
6156 | /* | |
6157 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
6158 | * an error where there's an empty slot with no huge pagecache |
6159 | * to back it. This way, we avoid allocating a hugepage, and | |
6160 | * the sparse dumpfile avoids allocating disk blocks, but its | |
6161 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 6162 | */ |
3ae77f43 HD |
6163 | if (absent && (flags & FOLL_DUMP) && |
6164 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
6165 | if (pte) |
6166 | spin_unlock(ptl); | |
2a15efc9 HD |
6167 | remainder = 0; |
6168 | break; | |
6169 | } | |
63551ae0 | 6170 | |
9cc3a5bd NH |
6171 | /* |
6172 | * We need call hugetlb_fault for both hugepages under migration | |
6173 | * (in which case hugetlb_fault waits for the migration,) and | |
6174 | * hwpoisoned hugepages (in which case we need to prevent the | |
6175 | * caller from accessing to them.) In order to do this, we use | |
6176 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
6177 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
6178 | * both cases, and because we can't follow correct pages | |
6179 | * directly from any kind of swap entries. | |
6180 | */ | |
a7f22660 DH |
6181 | if (absent || |
6182 | __follow_hugetlb_must_fault(flags, pte, &unshare)) { | |
2b740303 | 6183 | vm_fault_t ret; |
87ffc118 | 6184 | unsigned int fault_flags = 0; |
63551ae0 | 6185 | |
cb900f41 KS |
6186 | if (pte) |
6187 | spin_unlock(ptl); | |
87ffc118 AA |
6188 | if (flags & FOLL_WRITE) |
6189 | fault_flags |= FAULT_FLAG_WRITE; | |
a7f22660 DH |
6190 | else if (unshare) |
6191 | fault_flags |= FAULT_FLAG_UNSHARE; | |
4f6da934 | 6192 | if (locked) |
71335f37 PX |
6193 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
6194 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
6195 | if (flags & FOLL_NOWAIT) |
6196 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
6197 | FAULT_FLAG_RETRY_NOWAIT; | |
6198 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
6199 | /* |
6200 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
6201 | * FAULT_FLAG_TRIED can co-exist | |
6202 | */ | |
87ffc118 AA |
6203 | fault_flags |= FAULT_FLAG_TRIED; |
6204 | } | |
6205 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
6206 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 6207 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
6208 | remainder = 0; |
6209 | break; | |
6210 | } | |
6211 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 6212 | if (locked && |
1ac25013 | 6213 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 6214 | *locked = 0; |
87ffc118 AA |
6215 | *nr_pages = 0; |
6216 | /* | |
6217 | * VM_FAULT_RETRY must not return an | |
6218 | * error, it will return zero | |
6219 | * instead. | |
6220 | * | |
6221 | * No need to update "position" as the | |
6222 | * caller will not check it after | |
6223 | * *nr_pages is set to 0. | |
6224 | */ | |
6225 | return i; | |
6226 | } | |
6227 | continue; | |
4c887265 AL |
6228 | } |
6229 | ||
a5516438 | 6230 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 6231 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 6232 | |
b6a2619c DH |
6233 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && |
6234 | !PageAnonExclusive(page), page); | |
6235 | ||
acbfb087 ZL |
6236 | /* |
6237 | * If subpage information not requested, update counters | |
6238 | * and skip the same_page loop below. | |
6239 | */ | |
6240 | if (!pages && !vmas && !pfn_offset && | |
6241 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
6242 | (remainder >= pages_per_huge_page(h))) { | |
6243 | vaddr += huge_page_size(h); | |
6244 | remainder -= pages_per_huge_page(h); | |
6245 | i += pages_per_huge_page(h); | |
6246 | spin_unlock(ptl); | |
6247 | continue; | |
6248 | } | |
6249 | ||
d08af0a5 JM |
6250 | /* vaddr may not be aligned to PAGE_SIZE */ |
6251 | refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, | |
6252 | (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); | |
0fa5bc40 | 6253 | |
82e5d378 JM |
6254 | if (pages || vmas) |
6255 | record_subpages_vmas(mem_map_offset(page, pfn_offset), | |
6256 | vma, refs, | |
6257 | likely(pages) ? pages + i : NULL, | |
6258 | vmas ? vmas + i : NULL); | |
63551ae0 | 6259 | |
82e5d378 | 6260 | if (pages) { |
0fa5bc40 | 6261 | /* |
822951d8 | 6262 | * try_grab_folio() should always succeed here, |
0fa5bc40 JM |
6263 | * because: a) we hold the ptl lock, and b) we've just |
6264 | * checked that the huge page is present in the page | |
6265 | * tables. If the huge page is present, then the tail | |
6266 | * pages must also be present. The ptl prevents the | |
6267 | * head page and tail pages from being rearranged in | |
6268 | * any way. So this page must be available at this | |
6269 | * point, unless the page refcount overflowed: | |
6270 | */ | |
822951d8 MWO |
6271 | if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, |
6272 | flags))) { | |
0fa5bc40 JM |
6273 | spin_unlock(ptl); |
6274 | remainder = 0; | |
6275 | err = -ENOMEM; | |
6276 | break; | |
6277 | } | |
d5d4b0aa | 6278 | } |
82e5d378 JM |
6279 | |
6280 | vaddr += (refs << PAGE_SHIFT); | |
6281 | remainder -= refs; | |
6282 | i += refs; | |
6283 | ||
cb900f41 | 6284 | spin_unlock(ptl); |
63551ae0 | 6285 | } |
28a35716 | 6286 | *nr_pages = remainder; |
87ffc118 AA |
6287 | /* |
6288 | * setting position is actually required only if remainder is | |
6289 | * not zero but it's faster not to add a "if (remainder)" | |
6290 | * branch. | |
6291 | */ | |
63551ae0 DG |
6292 | *position = vaddr; |
6293 | ||
2be7cfed | 6294 | return i ? i : err; |
63551ae0 | 6295 | } |
8f860591 | 6296 | |
7da4d641 | 6297 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
5a90d5a1 PX |
6298 | unsigned long address, unsigned long end, |
6299 | pgprot_t newprot, unsigned long cp_flags) | |
8f860591 ZY |
6300 | { |
6301 | struct mm_struct *mm = vma->vm_mm; | |
6302 | unsigned long start = address; | |
6303 | pte_t *ptep; | |
6304 | pte_t pte; | |
a5516438 | 6305 | struct hstate *h = hstate_vma(vma); |
60dfaad6 | 6306 | unsigned long pages = 0, psize = huge_page_size(h); |
dff11abe | 6307 | bool shared_pmd = false; |
ac46d4f3 | 6308 | struct mmu_notifier_range range; |
e95a9851 | 6309 | unsigned long last_addr_mask; |
5a90d5a1 PX |
6310 | bool uffd_wp = cp_flags & MM_CP_UFFD_WP; |
6311 | bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; | |
dff11abe MK |
6312 | |
6313 | /* | |
6314 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 6315 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
6316 | * range if PMD sharing is possible. |
6317 | */ | |
7269f999 JG |
6318 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
6319 | 0, vma, mm, start, end); | |
ac46d4f3 | 6320 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
6321 | |
6322 | BUG_ON(address >= end); | |
ac46d4f3 | 6323 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 6324 | |
ac46d4f3 | 6325 | mmu_notifier_invalidate_range_start(&range); |
e95a9851 | 6326 | last_addr_mask = hugetlb_mask_last_page(h); |
83cde9e8 | 6327 | i_mmap_lock_write(vma->vm_file->f_mapping); |
60dfaad6 | 6328 | for (; address < end; address += psize) { |
cb900f41 | 6329 | spinlock_t *ptl; |
60dfaad6 | 6330 | ptep = huge_pte_offset(mm, address, psize); |
e95a9851 MK |
6331 | if (!ptep) { |
6332 | address |= last_addr_mask; | |
8f860591 | 6333 | continue; |
e95a9851 | 6334 | } |
cb900f41 | 6335 | ptl = huge_pte_lock(h, mm, ptep); |
4ddb4d91 | 6336 | if (huge_pmd_unshare(mm, vma, address, ptep)) { |
60dfaad6 PX |
6337 | /* |
6338 | * When uffd-wp is enabled on the vma, unshare | |
6339 | * shouldn't happen at all. Warn about it if it | |
6340 | * happened due to some reason. | |
6341 | */ | |
6342 | WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); | |
7da4d641 | 6343 | pages++; |
cb900f41 | 6344 | spin_unlock(ptl); |
dff11abe | 6345 | shared_pmd = true; |
4ddb4d91 | 6346 | address |= last_addr_mask; |
39dde65c | 6347 | continue; |
7da4d641 | 6348 | } |
a8bda28d NH |
6349 | pte = huge_ptep_get(ptep); |
6350 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
6351 | spin_unlock(ptl); | |
6352 | continue; | |
6353 | } | |
6354 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
6355 | swp_entry_t entry = pte_to_swp_entry(pte); | |
6c287605 | 6356 | struct page *page = pfn_swap_entry_to_page(entry); |
a8bda28d | 6357 | |
6c287605 | 6358 | if (!is_readable_migration_entry(entry)) { |
a8bda28d NH |
6359 | pte_t newpte; |
6360 | ||
6c287605 DH |
6361 | if (PageAnon(page)) |
6362 | entry = make_readable_exclusive_migration_entry( | |
6363 | swp_offset(entry)); | |
6364 | else | |
6365 | entry = make_readable_migration_entry( | |
6366 | swp_offset(entry)); | |
a8bda28d | 6367 | newpte = swp_entry_to_pte(entry); |
5a90d5a1 PX |
6368 | if (uffd_wp) |
6369 | newpte = pte_swp_mkuffd_wp(newpte); | |
6370 | else if (uffd_wp_resolve) | |
6371 | newpte = pte_swp_clear_uffd_wp(newpte); | |
18f39629 | 6372 | set_huge_pte_at(mm, address, ptep, newpte); |
a8bda28d NH |
6373 | pages++; |
6374 | } | |
6375 | spin_unlock(ptl); | |
6376 | continue; | |
6377 | } | |
60dfaad6 PX |
6378 | if (unlikely(pte_marker_uffd_wp(pte))) { |
6379 | /* | |
6380 | * This is changing a non-present pte into a none pte, | |
6381 | * no need for huge_ptep_modify_prot_start/commit(). | |
6382 | */ | |
6383 | if (uffd_wp_resolve) | |
6384 | huge_pte_clear(mm, address, ptep, psize); | |
6385 | } | |
a8bda28d | 6386 | if (!huge_pte_none(pte)) { |
023bdd00 | 6387 | pte_t old_pte; |
79c1c594 | 6388 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
023bdd00 AK |
6389 | |
6390 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
16785bd7 | 6391 | pte = huge_pte_modify(old_pte, newprot); |
79c1c594 | 6392 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
5a90d5a1 PX |
6393 | if (uffd_wp) |
6394 | pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); | |
6395 | else if (uffd_wp_resolve) | |
6396 | pte = huge_pte_clear_uffd_wp(pte); | |
023bdd00 | 6397 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 6398 | pages++; |
60dfaad6 PX |
6399 | } else { |
6400 | /* None pte */ | |
6401 | if (unlikely(uffd_wp)) | |
6402 | /* Safe to modify directly (none->non-present). */ | |
6403 | set_huge_pte_at(mm, address, ptep, | |
6404 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
8f860591 | 6405 | } |
cb900f41 | 6406 | spin_unlock(ptl); |
8f860591 | 6407 | } |
d833352a | 6408 | /* |
c8c06efa | 6409 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 6410 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 6411 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
6412 | * and that page table be reused and filled with junk. If we actually |
6413 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 6414 | */ |
dff11abe | 6415 | if (shared_pmd) |
ac46d4f3 | 6416 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
6417 | else |
6418 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
6419 | /* |
6420 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
6421 | * page table protection not changing it to point to a new page. | |
6422 | * | |
ee65728e | 6423 | * See Documentation/mm/mmu_notifier.rst |
0f10851e | 6424 | */ |
83cde9e8 | 6425 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 6426 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
6427 | |
6428 | return pages << h->order; | |
8f860591 ZY |
6429 | } |
6430 | ||
33b8f84a MK |
6431 | /* Return true if reservation was successful, false otherwise. */ |
6432 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 6433 | long from, long to, |
5a6fe125 | 6434 | struct vm_area_struct *vma, |
ca16d140 | 6435 | vm_flags_t vm_flags) |
e4e574b7 | 6436 | { |
33b8f84a | 6437 | long chg, add = -1; |
a5516438 | 6438 | struct hstate *h = hstate_inode(inode); |
90481622 | 6439 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 6440 | struct resv_map *resv_map; |
075a61d0 | 6441 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 6442 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 6443 | |
63489f8e MK |
6444 | /* This should never happen */ |
6445 | if (from > to) { | |
6446 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 6447 | return false; |
63489f8e MK |
6448 | } |
6449 | ||
17c9d12e MG |
6450 | /* |
6451 | * Only apply hugepage reservation if asked. At fault time, an | |
6452 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 6453 | * without using reserves |
17c9d12e | 6454 | */ |
ca16d140 | 6455 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 6456 | return true; |
17c9d12e | 6457 | |
a1e78772 MG |
6458 | /* |
6459 | * Shared mappings base their reservation on the number of pages that | |
6460 | * are already allocated on behalf of the file. Private mappings need | |
6461 | * to reserve the full area even if read-only as mprotect() may be | |
6462 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
6463 | */ | |
9119a41e | 6464 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
6465 | /* |
6466 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
6467 | * called for inodes for which resv_maps were created (see | |
6468 | * hugetlbfs_get_inode). | |
6469 | */ | |
4e35f483 | 6470 | resv_map = inode_resv_map(inode); |
9119a41e | 6471 | |
0db9d74e | 6472 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
6473 | |
6474 | } else { | |
e9fe92ae | 6475 | /* Private mapping. */ |
9119a41e | 6476 | resv_map = resv_map_alloc(); |
17c9d12e | 6477 | if (!resv_map) |
33b8f84a | 6478 | return false; |
17c9d12e | 6479 | |
a1e78772 | 6480 | chg = to - from; |
84afd99b | 6481 | |
17c9d12e MG |
6482 | set_vma_resv_map(vma, resv_map); |
6483 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
6484 | } | |
6485 | ||
33b8f84a | 6486 | if (chg < 0) |
c50ac050 | 6487 | goto out_err; |
8a630112 | 6488 | |
33b8f84a MK |
6489 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
6490 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 6491 | goto out_err; |
075a61d0 MA |
6492 | |
6493 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
6494 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
6495 | * of the resv_map. | |
6496 | */ | |
6497 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
6498 | } | |
6499 | ||
1c5ecae3 MK |
6500 | /* |
6501 | * There must be enough pages in the subpool for the mapping. If | |
6502 | * the subpool has a minimum size, there may be some global | |
6503 | * reservations already in place (gbl_reserve). | |
6504 | */ | |
6505 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 6506 | if (gbl_reserve < 0) |
075a61d0 | 6507 | goto out_uncharge_cgroup; |
5a6fe125 MG |
6508 | |
6509 | /* | |
17c9d12e | 6510 | * Check enough hugepages are available for the reservation. |
90481622 | 6511 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 6512 | */ |
33b8f84a | 6513 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 6514 | goto out_put_pages; |
17c9d12e MG |
6515 | |
6516 | /* | |
6517 | * Account for the reservations made. Shared mappings record regions | |
6518 | * that have reservations as they are shared by multiple VMAs. | |
6519 | * When the last VMA disappears, the region map says how much | |
6520 | * the reservation was and the page cache tells how much of | |
6521 | * the reservation was consumed. Private mappings are per-VMA and | |
6522 | * only the consumed reservations are tracked. When the VMA | |
6523 | * disappears, the original reservation is the VMA size and the | |
6524 | * consumed reservations are stored in the map. Hence, nothing | |
6525 | * else has to be done for private mappings here | |
6526 | */ | |
33039678 | 6527 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 6528 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
6529 | |
6530 | if (unlikely(add < 0)) { | |
6531 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 6532 | goto out_put_pages; |
0db9d74e | 6533 | } else if (unlikely(chg > add)) { |
33039678 MK |
6534 | /* |
6535 | * pages in this range were added to the reserve | |
6536 | * map between region_chg and region_add. This | |
6537 | * indicates a race with alloc_huge_page. Adjust | |
6538 | * the subpool and reserve counts modified above | |
6539 | * based on the difference. | |
6540 | */ | |
6541 | long rsv_adjust; | |
6542 | ||
d85aecf2 ML |
6543 | /* |
6544 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
6545 | * reference to h_cg->css. See comment below for detail. | |
6546 | */ | |
075a61d0 MA |
6547 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
6548 | hstate_index(h), | |
6549 | (chg - add) * pages_per_huge_page(h), h_cg); | |
6550 | ||
33039678 MK |
6551 | rsv_adjust = hugepage_subpool_put_pages(spool, |
6552 | chg - add); | |
6553 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
6554 | } else if (h_cg) { |
6555 | /* | |
6556 | * The file_regions will hold their own reference to | |
6557 | * h_cg->css. So we should release the reference held | |
6558 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
6559 | * done. | |
6560 | */ | |
6561 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
6562 | } |
6563 | } | |
33b8f84a MK |
6564 | return true; |
6565 | ||
075a61d0 MA |
6566 | out_put_pages: |
6567 | /* put back original number of pages, chg */ | |
6568 | (void)hugepage_subpool_put_pages(spool, chg); | |
6569 | out_uncharge_cgroup: | |
6570 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
6571 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 6572 | out_err: |
5e911373 | 6573 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
6574 | /* Only call region_abort if the region_chg succeeded but the |
6575 | * region_add failed or didn't run. | |
6576 | */ | |
6577 | if (chg >= 0 && add < 0) | |
6578 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
6579 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
6580 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 6581 | return false; |
a43a8c39 CK |
6582 | } |
6583 | ||
b5cec28d MK |
6584 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
6585 | long freed) | |
a43a8c39 | 6586 | { |
a5516438 | 6587 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 6588 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 6589 | long chg = 0; |
90481622 | 6590 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 6591 | long gbl_reserve; |
45c682a6 | 6592 | |
f27a5136 MK |
6593 | /* |
6594 | * Since this routine can be called in the evict inode path for all | |
6595 | * hugetlbfs inodes, resv_map could be NULL. | |
6596 | */ | |
b5cec28d MK |
6597 | if (resv_map) { |
6598 | chg = region_del(resv_map, start, end); | |
6599 | /* | |
6600 | * region_del() can fail in the rare case where a region | |
6601 | * must be split and another region descriptor can not be | |
6602 | * allocated. If end == LONG_MAX, it will not fail. | |
6603 | */ | |
6604 | if (chg < 0) | |
6605 | return chg; | |
6606 | } | |
6607 | ||
45c682a6 | 6608 | spin_lock(&inode->i_lock); |
e4c6f8be | 6609 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
6610 | spin_unlock(&inode->i_lock); |
6611 | ||
1c5ecae3 MK |
6612 | /* |
6613 | * If the subpool has a minimum size, the number of global | |
6614 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
6615 | * |
6616 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
6617 | * won't go negative. | |
1c5ecae3 MK |
6618 | */ |
6619 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
6620 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
6621 | |
6622 | return 0; | |
a43a8c39 | 6623 | } |
93f70f90 | 6624 | |
3212b535 SC |
6625 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
6626 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
6627 | struct vm_area_struct *vma, | |
6628 | unsigned long addr, pgoff_t idx) | |
6629 | { | |
6630 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
6631 | svma->vm_start; | |
6632 | unsigned long sbase = saddr & PUD_MASK; | |
6633 | unsigned long s_end = sbase + PUD_SIZE; | |
6634 | ||
6635 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
6636 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
6637 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
6638 | |
6639 | /* | |
6640 | * match the virtual addresses, permission and the alignment of the | |
6641 | * page table page. | |
6642 | */ | |
6643 | if (pmd_index(addr) != pmd_index(saddr) || | |
6644 | vm_flags != svm_flags || | |
07e51edf | 6645 | !range_in_vma(svma, sbase, s_end)) |
3212b535 SC |
6646 | return 0; |
6647 | ||
6648 | return saddr; | |
6649 | } | |
6650 | ||
31aafb45 | 6651 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
6652 | { |
6653 | unsigned long base = addr & PUD_MASK; | |
6654 | unsigned long end = base + PUD_SIZE; | |
6655 | ||
6656 | /* | |
6657 | * check on proper vm_flags and page table alignment | |
6658 | */ | |
017b1660 | 6659 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
6660 | return true; |
6661 | return false; | |
3212b535 SC |
6662 | } |
6663 | ||
c1991e07 PX |
6664 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
6665 | { | |
6666 | #ifdef CONFIG_USERFAULTFD | |
6667 | if (uffd_disable_huge_pmd_share(vma)) | |
6668 | return false; | |
6669 | #endif | |
6670 | return vma_shareable(vma, addr); | |
6671 | } | |
6672 | ||
017b1660 MK |
6673 | /* |
6674 | * Determine if start,end range within vma could be mapped by shared pmd. | |
6675 | * If yes, adjust start and end to cover range associated with possible | |
6676 | * shared pmd mappings. | |
6677 | */ | |
6678 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6679 | unsigned long *start, unsigned long *end) | |
6680 | { | |
a1ba9da8 LX |
6681 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
6682 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 6683 | |
a1ba9da8 | 6684 | /* |
f0953a1b IM |
6685 | * vma needs to span at least one aligned PUD size, and the range |
6686 | * must be at least partially within in. | |
a1ba9da8 LX |
6687 | */ |
6688 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
6689 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
6690 | return; |
6691 | ||
75802ca6 | 6692 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
6693 | if (*start > v_start) |
6694 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 6695 | |
a1ba9da8 LX |
6696 | if (*end < v_end) |
6697 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
6698 | } |
6699 | ||
3212b535 SC |
6700 | /* |
6701 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
6702 | * and returns the corresponding pte. While this is not necessary for the | |
6703 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
6704 | * code much cleaner. |
6705 | * | |
0bf7b64e MK |
6706 | * This routine must be called with i_mmap_rwsem held in at least read mode if |
6707 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
6708 | * table entries associated with the address space. This is important as we | |
6709 | * are setting up sharing based on existing page table entries (mappings). | |
3212b535 | 6710 | */ |
aec44e0f PX |
6711 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6712 | unsigned long addr, pud_t *pud) | |
3212b535 | 6713 | { |
3212b535 SC |
6714 | struct address_space *mapping = vma->vm_file->f_mapping; |
6715 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
6716 | vma->vm_pgoff; | |
6717 | struct vm_area_struct *svma; | |
6718 | unsigned long saddr; | |
6719 | pte_t *spte = NULL; | |
6720 | pte_t *pte; | |
cb900f41 | 6721 | spinlock_t *ptl; |
3212b535 | 6722 | |
0bf7b64e | 6723 | i_mmap_assert_locked(mapping); |
3212b535 SC |
6724 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
6725 | if (svma == vma) | |
6726 | continue; | |
6727 | ||
6728 | saddr = page_table_shareable(svma, vma, addr, idx); | |
6729 | if (saddr) { | |
7868a208 PA |
6730 | spte = huge_pte_offset(svma->vm_mm, saddr, |
6731 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
6732 | if (spte) { |
6733 | get_page(virt_to_page(spte)); | |
6734 | break; | |
6735 | } | |
6736 | } | |
6737 | } | |
6738 | ||
6739 | if (!spte) | |
6740 | goto out; | |
6741 | ||
8bea8052 | 6742 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 6743 | if (pud_none(*pud)) { |
3212b535 SC |
6744 | pud_populate(mm, pud, |
6745 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 6746 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 6747 | } else { |
3212b535 | 6748 | put_page(virt_to_page(spte)); |
dc6c9a35 | 6749 | } |
cb900f41 | 6750 | spin_unlock(ptl); |
3212b535 SC |
6751 | out: |
6752 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
6753 | return pte; |
6754 | } | |
6755 | ||
6756 | /* | |
6757 | * unmap huge page backed by shared pte. | |
6758 | * | |
6759 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
6760 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
6761 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
6762 | * | |
c0d0381a | 6763 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
6764 | * |
6765 | * returns: 1 successfully unmapped a shared pte page | |
6766 | * 0 the underlying pte page is not shared, or it is the last user | |
6767 | */ | |
34ae204f | 6768 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
4ddb4d91 | 6769 | unsigned long addr, pte_t *ptep) |
3212b535 | 6770 | { |
4ddb4d91 MK |
6771 | pgd_t *pgd = pgd_offset(mm, addr); |
6772 | p4d_t *p4d = p4d_offset(pgd, addr); | |
6773 | pud_t *pud = pud_offset(p4d, addr); | |
3212b535 | 6774 | |
34ae204f | 6775 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
6776 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
6777 | if (page_count(virt_to_page(ptep)) == 1) | |
6778 | return 0; | |
6779 | ||
6780 | pud_clear(pud); | |
6781 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 6782 | mm_dec_nr_pmds(mm); |
3212b535 SC |
6783 | return 1; |
6784 | } | |
c1991e07 | 6785 | |
9e5fc74c | 6786 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
aec44e0f PX |
6787 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6788 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
6789 | { |
6790 | return NULL; | |
6791 | } | |
e81f2d22 | 6792 | |
34ae204f | 6793 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
4ddb4d91 | 6794 | unsigned long addr, pte_t *ptep) |
e81f2d22 ZZ |
6795 | { |
6796 | return 0; | |
6797 | } | |
017b1660 MK |
6798 | |
6799 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6800 | unsigned long *start, unsigned long *end) | |
6801 | { | |
6802 | } | |
c1991e07 PX |
6803 | |
6804 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
6805 | { | |
6806 | return false; | |
6807 | } | |
3212b535 SC |
6808 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
6809 | ||
9e5fc74c | 6810 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 6811 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
6812 | unsigned long addr, unsigned long sz) |
6813 | { | |
6814 | pgd_t *pgd; | |
c2febafc | 6815 | p4d_t *p4d; |
9e5fc74c SC |
6816 | pud_t *pud; |
6817 | pte_t *pte = NULL; | |
6818 | ||
6819 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
6820 | p4d = p4d_alloc(mm, pgd, addr); |
6821 | if (!p4d) | |
6822 | return NULL; | |
c2febafc | 6823 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
6824 | if (pud) { |
6825 | if (sz == PUD_SIZE) { | |
6826 | pte = (pte_t *)pud; | |
6827 | } else { | |
6828 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 6829 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 6830 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
6831 | else |
6832 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
6833 | } | |
6834 | } | |
4e666314 | 6835 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
6836 | |
6837 | return pte; | |
6838 | } | |
6839 | ||
9b19df29 PA |
6840 | /* |
6841 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
6842 | * entry at address @addr | |
6843 | * | |
8ac0b81a LX |
6844 | * Return: Pointer to page table entry (PUD or PMD) for |
6845 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
6846 | * size @sz doesn't match the hugepage size at this level of the page |
6847 | * table. | |
6848 | */ | |
7868a208 PA |
6849 | pte_t *huge_pte_offset(struct mm_struct *mm, |
6850 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
6851 | { |
6852 | pgd_t *pgd; | |
c2febafc | 6853 | p4d_t *p4d; |
8ac0b81a LX |
6854 | pud_t *pud; |
6855 | pmd_t *pmd; | |
9e5fc74c SC |
6856 | |
6857 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
6858 | if (!pgd_present(*pgd)) |
6859 | return NULL; | |
6860 | p4d = p4d_offset(pgd, addr); | |
6861 | if (!p4d_present(*p4d)) | |
6862 | return NULL; | |
9b19df29 | 6863 | |
c2febafc | 6864 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
6865 | if (sz == PUD_SIZE) |
6866 | /* must be pud huge, non-present or none */ | |
c2febafc | 6867 | return (pte_t *)pud; |
8ac0b81a | 6868 | if (!pud_present(*pud)) |
9b19df29 | 6869 | return NULL; |
8ac0b81a | 6870 | /* must have a valid entry and size to go further */ |
9b19df29 | 6871 | |
8ac0b81a LX |
6872 | pmd = pmd_offset(pud, addr); |
6873 | /* must be pmd huge, non-present or none */ | |
6874 | return (pte_t *)pmd; | |
9e5fc74c SC |
6875 | } |
6876 | ||
e95a9851 MK |
6877 | /* |
6878 | * Return a mask that can be used to update an address to the last huge | |
6879 | * page in a page table page mapping size. Used to skip non-present | |
6880 | * page table entries when linearly scanning address ranges. Architectures | |
6881 | * with unique huge page to page table relationships can define their own | |
6882 | * version of this routine. | |
6883 | */ | |
6884 | unsigned long hugetlb_mask_last_page(struct hstate *h) | |
6885 | { | |
6886 | unsigned long hp_size = huge_page_size(h); | |
6887 | ||
6888 | if (hp_size == PUD_SIZE) | |
6889 | return P4D_SIZE - PUD_SIZE; | |
6890 | else if (hp_size == PMD_SIZE) | |
6891 | return PUD_SIZE - PMD_SIZE; | |
6892 | else | |
6893 | return 0UL; | |
6894 | } | |
6895 | ||
6896 | #else | |
6897 | ||
6898 | /* See description above. Architectures can provide their own version. */ | |
6899 | __weak unsigned long hugetlb_mask_last_page(struct hstate *h) | |
6900 | { | |
4ddb4d91 MK |
6901 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
6902 | if (huge_page_size(h) == PMD_SIZE) | |
6903 | return PUD_SIZE - PMD_SIZE; | |
6904 | #endif | |
e95a9851 MK |
6905 | return 0UL; |
6906 | } | |
6907 | ||
61f77eda NH |
6908 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
6909 | ||
6910 | /* | |
6911 | * These functions are overwritable if your architecture needs its own | |
6912 | * behavior. | |
6913 | */ | |
6914 | struct page * __weak | |
6915 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
6916 | int write) | |
6917 | { | |
6918 | return ERR_PTR(-EINVAL); | |
6919 | } | |
6920 | ||
4dc71451 AK |
6921 | struct page * __weak |
6922 | follow_huge_pd(struct vm_area_struct *vma, | |
6923 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
6924 | { | |
6925 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
6926 | return NULL; | |
6927 | } | |
6928 | ||
61f77eda | 6929 | struct page * __weak |
9e5fc74c | 6930 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 6931 | pmd_t *pmd, int flags) |
9e5fc74c | 6932 | { |
e66f17ff NH |
6933 | struct page *page = NULL; |
6934 | spinlock_t *ptl; | |
c9d398fa | 6935 | pte_t pte; |
3faa52c0 | 6936 | |
8909691b DH |
6937 | /* |
6938 | * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via | |
6939 | * follow_hugetlb_page(). | |
6940 | */ | |
6941 | if (WARN_ON_ONCE(flags & FOLL_PIN)) | |
3faa52c0 JH |
6942 | return NULL; |
6943 | ||
e66f17ff NH |
6944 | retry: |
6945 | ptl = pmd_lockptr(mm, pmd); | |
6946 | spin_lock(ptl); | |
6947 | /* | |
6948 | * make sure that the address range covered by this pmd is not | |
6949 | * unmapped from other threads. | |
6950 | */ | |
6951 | if (!pmd_huge(*pmd)) | |
6952 | goto out; | |
c9d398fa NH |
6953 | pte = huge_ptep_get((pte_t *)pmd); |
6954 | if (pte_present(pte)) { | |
97534127 | 6955 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
6956 | /* |
6957 | * try_grab_page() should always succeed here, because: a) we | |
6958 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
6959 | * huge pmd (head) page is present in the page tables. The ptl | |
6960 | * prevents the head page and tail pages from being rearranged | |
6961 | * in any way. So this page must be available at this point, | |
6962 | * unless the page refcount overflowed: | |
6963 | */ | |
6964 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
6965 | page = NULL; | |
6966 | goto out; | |
6967 | } | |
e66f17ff | 6968 | } else { |
c9d398fa | 6969 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff | 6970 | spin_unlock(ptl); |
ad1ac596 | 6971 | __migration_entry_wait_huge((pte_t *)pmd, ptl); |
e66f17ff NH |
6972 | goto retry; |
6973 | } | |
6974 | /* | |
6975 | * hwpoisoned entry is treated as no_page_table in | |
6976 | * follow_page_mask(). | |
6977 | */ | |
6978 | } | |
6979 | out: | |
6980 | spin_unlock(ptl); | |
9e5fc74c SC |
6981 | return page; |
6982 | } | |
6983 | ||
61f77eda | 6984 | struct page * __weak |
9e5fc74c | 6985 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 6986 | pud_t *pud, int flags) |
9e5fc74c | 6987 | { |
3faa52c0 | 6988 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 6989 | return NULL; |
9e5fc74c | 6990 | |
e66f17ff | 6991 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
6992 | } |
6993 | ||
faaa5b62 AK |
6994 | struct page * __weak |
6995 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
6996 | { | |
3faa52c0 | 6997 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
6998 | return NULL; |
6999 | ||
7000 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
7001 | } | |
7002 | ||
7ce82f4c | 7003 | int isolate_hugetlb(struct page *page, struct list_head *list) |
31caf665 | 7004 | { |
7ce82f4c | 7005 | int ret = 0; |
bcc54222 | 7006 | |
db71ef79 | 7007 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
7008 | if (!PageHeadHuge(page) || |
7009 | !HPageMigratable(page) || | |
0eb2df2b | 7010 | !get_page_unless_zero(page)) { |
7ce82f4c | 7011 | ret = -EBUSY; |
bcc54222 NH |
7012 | goto unlock; |
7013 | } | |
8f251a3d | 7014 | ClearHPageMigratable(page); |
31caf665 | 7015 | list_move_tail(&page->lru, list); |
bcc54222 | 7016 | unlock: |
db71ef79 | 7017 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 7018 | return ret; |
31caf665 NH |
7019 | } |
7020 | ||
25182f05 NH |
7021 | int get_hwpoison_huge_page(struct page *page, bool *hugetlb) |
7022 | { | |
7023 | int ret = 0; | |
7024 | ||
7025 | *hugetlb = false; | |
7026 | spin_lock_irq(&hugetlb_lock); | |
7027 | if (PageHeadHuge(page)) { | |
7028 | *hugetlb = true; | |
b283d983 NH |
7029 | if (HPageFreed(page)) |
7030 | ret = 0; | |
7031 | else if (HPageMigratable(page)) | |
25182f05 | 7032 | ret = get_page_unless_zero(page); |
0ed950d1 NH |
7033 | else |
7034 | ret = -EBUSY; | |
25182f05 NH |
7035 | } |
7036 | spin_unlock_irq(&hugetlb_lock); | |
7037 | return ret; | |
7038 | } | |
7039 | ||
405ce051 NH |
7040 | int get_huge_page_for_hwpoison(unsigned long pfn, int flags) |
7041 | { | |
7042 | int ret; | |
7043 | ||
7044 | spin_lock_irq(&hugetlb_lock); | |
7045 | ret = __get_huge_page_for_hwpoison(pfn, flags); | |
7046 | spin_unlock_irq(&hugetlb_lock); | |
7047 | return ret; | |
7048 | } | |
7049 | ||
31caf665 NH |
7050 | void putback_active_hugepage(struct page *page) |
7051 | { | |
db71ef79 | 7052 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 7053 | SetHPageMigratable(page); |
31caf665 | 7054 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 7055 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
7056 | put_page(page); |
7057 | } | |
ab5ac90a MH |
7058 | |
7059 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
7060 | { | |
7061 | struct hstate *h = page_hstate(oldpage); | |
7062 | ||
7063 | hugetlb_cgroup_migrate(oldpage, newpage); | |
7064 | set_page_owner_migrate_reason(newpage, reason); | |
7065 | ||
7066 | /* | |
7067 | * transfer temporary state of the new huge page. This is | |
7068 | * reverse to other transitions because the newpage is going to | |
7069 | * be final while the old one will be freed so it takes over | |
7070 | * the temporary status. | |
7071 | * | |
7072 | * Also note that we have to transfer the per-node surplus state | |
7073 | * here as well otherwise the global surplus count will not match | |
7074 | * the per-node's. | |
7075 | */ | |
9157c311 | 7076 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
7077 | int old_nid = page_to_nid(oldpage); |
7078 | int new_nid = page_to_nid(newpage); | |
7079 | ||
9157c311 MK |
7080 | SetHPageTemporary(oldpage); |
7081 | ClearHPageTemporary(newpage); | |
ab5ac90a | 7082 | |
5af1ab1d ML |
7083 | /* |
7084 | * There is no need to transfer the per-node surplus state | |
7085 | * when we do not cross the node. | |
7086 | */ | |
7087 | if (new_nid == old_nid) | |
7088 | return; | |
db71ef79 | 7089 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
7090 | if (h->surplus_huge_pages_node[old_nid]) { |
7091 | h->surplus_huge_pages_node[old_nid]--; | |
7092 | h->surplus_huge_pages_node[new_nid]++; | |
7093 | } | |
db71ef79 | 7094 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
7095 | } |
7096 | } | |
cf11e85f | 7097 | |
6dfeaff9 PX |
7098 | /* |
7099 | * This function will unconditionally remove all the shared pmd pgtable entries | |
7100 | * within the specific vma for a hugetlbfs memory range. | |
7101 | */ | |
7102 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
7103 | { | |
7104 | struct hstate *h = hstate_vma(vma); | |
7105 | unsigned long sz = huge_page_size(h); | |
7106 | struct mm_struct *mm = vma->vm_mm; | |
7107 | struct mmu_notifier_range range; | |
7108 | unsigned long address, start, end; | |
7109 | spinlock_t *ptl; | |
7110 | pte_t *ptep; | |
7111 | ||
7112 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
7113 | return; | |
7114 | ||
7115 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
7116 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
7117 | ||
7118 | if (start >= end) | |
7119 | return; | |
7120 | ||
9c8bbfac | 7121 | flush_cache_range(vma, start, end); |
6dfeaff9 PX |
7122 | /* |
7123 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
7124 | * we have already done the PUD_SIZE alignment. | |
7125 | */ | |
7126 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
7127 | start, end); | |
7128 | mmu_notifier_invalidate_range_start(&range); | |
7129 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
7130 | for (address = start; address < end; address += PUD_SIZE) { | |
6dfeaff9 PX |
7131 | ptep = huge_pte_offset(mm, address, sz); |
7132 | if (!ptep) | |
7133 | continue; | |
7134 | ptl = huge_pte_lock(h, mm, ptep); | |
4ddb4d91 | 7135 | huge_pmd_unshare(mm, vma, address, ptep); |
6dfeaff9 PX |
7136 | spin_unlock(ptl); |
7137 | } | |
7138 | flush_hugetlb_tlb_range(vma, start, end); | |
7139 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
7140 | /* | |
7141 | * No need to call mmu_notifier_invalidate_range(), see | |
ee65728e | 7142 | * Documentation/mm/mmu_notifier.rst. |
6dfeaff9 PX |
7143 | */ |
7144 | mmu_notifier_invalidate_range_end(&range); | |
7145 | } | |
7146 | ||
cf11e85f | 7147 | #ifdef CONFIG_CMA |
cf11e85f RG |
7148 | static bool cma_reserve_called __initdata; |
7149 | ||
7150 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
7151 | { | |
38e719ab BW |
7152 | int nid, count = 0; |
7153 | unsigned long tmp; | |
7154 | char *s = p; | |
7155 | ||
7156 | while (*s) { | |
7157 | if (sscanf(s, "%lu%n", &tmp, &count) != 1) | |
7158 | break; | |
7159 | ||
7160 | if (s[count] == ':') { | |
f9317f77 | 7161 | if (tmp >= MAX_NUMNODES) |
38e719ab | 7162 | break; |
f9317f77 | 7163 | nid = array_index_nospec(tmp, MAX_NUMNODES); |
38e719ab BW |
7164 | |
7165 | s += count + 1; | |
7166 | tmp = memparse(s, &s); | |
7167 | hugetlb_cma_size_in_node[nid] = tmp; | |
7168 | hugetlb_cma_size += tmp; | |
7169 | ||
7170 | /* | |
7171 | * Skip the separator if have one, otherwise | |
7172 | * break the parsing. | |
7173 | */ | |
7174 | if (*s == ',') | |
7175 | s++; | |
7176 | else | |
7177 | break; | |
7178 | } else { | |
7179 | hugetlb_cma_size = memparse(p, &p); | |
7180 | break; | |
7181 | } | |
7182 | } | |
7183 | ||
cf11e85f RG |
7184 | return 0; |
7185 | } | |
7186 | ||
7187 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
7188 | ||
7189 | void __init hugetlb_cma_reserve(int order) | |
7190 | { | |
7191 | unsigned long size, reserved, per_node; | |
38e719ab | 7192 | bool node_specific_cma_alloc = false; |
cf11e85f RG |
7193 | int nid; |
7194 | ||
7195 | cma_reserve_called = true; | |
7196 | ||
38e719ab BW |
7197 | if (!hugetlb_cma_size) |
7198 | return; | |
7199 | ||
7200 | for (nid = 0; nid < MAX_NUMNODES; nid++) { | |
7201 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7202 | continue; | |
7203 | ||
30a51400 | 7204 | if (!node_online(nid)) { |
38e719ab BW |
7205 | pr_warn("hugetlb_cma: invalid node %d specified\n", nid); |
7206 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7207 | hugetlb_cma_size_in_node[nid] = 0; | |
7208 | continue; | |
7209 | } | |
7210 | ||
7211 | if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { | |
7212 | pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", | |
7213 | nid, (PAGE_SIZE << order) / SZ_1M); | |
7214 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7215 | hugetlb_cma_size_in_node[nid] = 0; | |
7216 | } else { | |
7217 | node_specific_cma_alloc = true; | |
7218 | } | |
7219 | } | |
7220 | ||
7221 | /* Validate the CMA size again in case some invalid nodes specified. */ | |
cf11e85f RG |
7222 | if (!hugetlb_cma_size) |
7223 | return; | |
7224 | ||
7225 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
7226 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
7227 | (PAGE_SIZE << order) / SZ_1M); | |
a01f4390 | 7228 | hugetlb_cma_size = 0; |
cf11e85f RG |
7229 | return; |
7230 | } | |
7231 | ||
38e719ab BW |
7232 | if (!node_specific_cma_alloc) { |
7233 | /* | |
7234 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
7235 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
7236 | */ | |
7237 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
7238 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
7239 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
7240 | } | |
cf11e85f RG |
7241 | |
7242 | reserved = 0; | |
30a51400 | 7243 | for_each_online_node(nid) { |
cf11e85f | 7244 | int res; |
2281f797 | 7245 | char name[CMA_MAX_NAME]; |
cf11e85f | 7246 | |
38e719ab BW |
7247 | if (node_specific_cma_alloc) { |
7248 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7249 | continue; | |
7250 | ||
7251 | size = hugetlb_cma_size_in_node[nid]; | |
7252 | } else { | |
7253 | size = min(per_node, hugetlb_cma_size - reserved); | |
7254 | } | |
7255 | ||
cf11e85f RG |
7256 | size = round_up(size, PAGE_SIZE << order); |
7257 | ||
2281f797 | 7258 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
a01f4390 MK |
7259 | /* |
7260 | * Note that 'order per bit' is based on smallest size that | |
7261 | * may be returned to CMA allocator in the case of | |
7262 | * huge page demotion. | |
7263 | */ | |
7264 | res = cma_declare_contiguous_nid(0, size, 0, | |
7265 | PAGE_SIZE << HUGETLB_PAGE_ORDER, | |
29d0f41d | 7266 | 0, false, name, |
cf11e85f RG |
7267 | &hugetlb_cma[nid], nid); |
7268 | if (res) { | |
7269 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
7270 | res, nid); | |
7271 | continue; | |
7272 | } | |
7273 | ||
7274 | reserved += size; | |
7275 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
7276 | size / SZ_1M, nid); | |
7277 | ||
7278 | if (reserved >= hugetlb_cma_size) | |
7279 | break; | |
7280 | } | |
a01f4390 MK |
7281 | |
7282 | if (!reserved) | |
7283 | /* | |
7284 | * hugetlb_cma_size is used to determine if allocations from | |
7285 | * cma are possible. Set to zero if no cma regions are set up. | |
7286 | */ | |
7287 | hugetlb_cma_size = 0; | |
cf11e85f RG |
7288 | } |
7289 | ||
7290 | void __init hugetlb_cma_check(void) | |
7291 | { | |
7292 | if (!hugetlb_cma_size || cma_reserve_called) | |
7293 | return; | |
7294 | ||
7295 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
7296 | } | |
7297 | ||
7298 | #endif /* CONFIG_CMA */ |