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> |
d6606683 | 34 | |
63551ae0 | 35 | #include <asm/page.h> |
ca15ca40 | 36 | #include <asm/pgalloc.h> |
24669e58 | 37 | #include <asm/tlb.h> |
63551ae0 | 38 | |
24669e58 | 39 | #include <linux/io.h> |
63551ae0 | 40 | #include <linux/hugetlb.h> |
9dd540e2 | 41 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 42 | #include <linux/node.h> |
ab5ac90a | 43 | #include <linux/page_owner.h> |
7835e98b | 44 | #include "internal.h" |
f41f2ed4 | 45 | #include "hugetlb_vmemmap.h" |
1da177e4 | 46 | |
c3f38a38 | 47 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
48 | unsigned int default_hstate_idx; |
49 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 50 | |
dbda8fea | 51 | #ifdef CONFIG_CMA |
cf11e85f | 52 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
a01f4390 MK |
53 | static bool hugetlb_cma_page(struct page *page, unsigned int order) |
54 | { | |
55 | return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, | |
56 | 1 << order); | |
57 | } | |
58 | #else | |
59 | static bool hugetlb_cma_page(struct page *page, unsigned int order) | |
60 | { | |
61 | return false; | |
62 | } | |
dbda8fea BS |
63 | #endif |
64 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 65 | |
641844f5 NH |
66 | /* |
67 | * Minimum page order among possible hugepage sizes, set to a proper value | |
68 | * at boot time. | |
69 | */ | |
70 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 71 | |
53ba51d2 JT |
72 | __initdata LIST_HEAD(huge_boot_pages); |
73 | ||
e5ff2159 AK |
74 | /* for command line parsing */ |
75 | static struct hstate * __initdata parsed_hstate; | |
76 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 77 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 78 | static bool __initdata parsed_default_hugepagesz; |
e5ff2159 | 79 | |
3935baa9 | 80 | /* |
31caf665 NH |
81 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
82 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 83 | */ |
c3f38a38 | 84 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 85 | |
8382d914 DB |
86 | /* |
87 | * Serializes faults on the same logical page. This is used to | |
88 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
89 | */ | |
90 | static int num_fault_mutexes; | |
c672c7f2 | 91 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 92 | |
7ca02d0a MK |
93 | /* Forward declaration */ |
94 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
95 | ||
1d88433b | 96 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 97 | { |
1d88433b ML |
98 | if (spool->count) |
99 | return false; | |
100 | if (spool->max_hpages != -1) | |
101 | return spool->used_hpages == 0; | |
102 | if (spool->min_hpages != -1) | |
103 | return spool->rsv_hpages == spool->min_hpages; | |
104 | ||
105 | return true; | |
106 | } | |
90481622 | 107 | |
db71ef79 MK |
108 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, |
109 | unsigned long irq_flags) | |
1d88433b | 110 | { |
db71ef79 | 111 | spin_unlock_irqrestore(&spool->lock, irq_flags); |
90481622 DG |
112 | |
113 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 114 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 115 | * free the subpool */ |
1d88433b | 116 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
117 | if (spool->min_hpages != -1) |
118 | hugetlb_acct_memory(spool->hstate, | |
119 | -spool->min_hpages); | |
90481622 | 120 | kfree(spool); |
7ca02d0a | 121 | } |
90481622 DG |
122 | } |
123 | ||
7ca02d0a MK |
124 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
125 | long min_hpages) | |
90481622 DG |
126 | { |
127 | struct hugepage_subpool *spool; | |
128 | ||
c6a91820 | 129 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
130 | if (!spool) |
131 | return NULL; | |
132 | ||
133 | spin_lock_init(&spool->lock); | |
134 | spool->count = 1; | |
7ca02d0a MK |
135 | spool->max_hpages = max_hpages; |
136 | spool->hstate = h; | |
137 | spool->min_hpages = min_hpages; | |
138 | ||
139 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
140 | kfree(spool); | |
141 | return NULL; | |
142 | } | |
143 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
144 | |
145 | return spool; | |
146 | } | |
147 | ||
148 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
149 | { | |
db71ef79 MK |
150 | unsigned long flags; |
151 | ||
152 | spin_lock_irqsave(&spool->lock, flags); | |
90481622 DG |
153 | BUG_ON(!spool->count); |
154 | spool->count--; | |
db71ef79 | 155 | unlock_or_release_subpool(spool, flags); |
90481622 DG |
156 | } |
157 | ||
1c5ecae3 MK |
158 | /* |
159 | * Subpool accounting for allocating and reserving pages. | |
160 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 161 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
162 | * global pools must be adjusted (upward). The returned value may |
163 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 164 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
165 | */ |
166 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
167 | long delta) |
168 | { | |
1c5ecae3 | 169 | long ret = delta; |
90481622 DG |
170 | |
171 | if (!spool) | |
1c5ecae3 | 172 | return ret; |
90481622 | 173 | |
db71ef79 | 174 | spin_lock_irq(&spool->lock); |
1c5ecae3 MK |
175 | |
176 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
177 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
178 | spool->used_hpages += delta; | |
179 | else { | |
180 | ret = -ENOMEM; | |
181 | goto unlock_ret; | |
182 | } | |
90481622 | 183 | } |
90481622 | 184 | |
09a95e29 MK |
185 | /* minimum size accounting */ |
186 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
187 | if (delta > spool->rsv_hpages) { |
188 | /* | |
189 | * Asking for more reserves than those already taken on | |
190 | * behalf of subpool. Return difference. | |
191 | */ | |
192 | ret = delta - spool->rsv_hpages; | |
193 | spool->rsv_hpages = 0; | |
194 | } else { | |
195 | ret = 0; /* reserves already accounted for */ | |
196 | spool->rsv_hpages -= delta; | |
197 | } | |
198 | } | |
199 | ||
200 | unlock_ret: | |
db71ef79 | 201 | spin_unlock_irq(&spool->lock); |
90481622 DG |
202 | return ret; |
203 | } | |
204 | ||
1c5ecae3 MK |
205 | /* |
206 | * Subpool accounting for freeing and unreserving pages. | |
207 | * Return the number of global page reservations that must be dropped. | |
208 | * The return value may only be different than the passed value (delta) | |
209 | * in the case where a subpool minimum size must be maintained. | |
210 | */ | |
211 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
212 | long delta) |
213 | { | |
1c5ecae3 | 214 | long ret = delta; |
db71ef79 | 215 | unsigned long flags; |
1c5ecae3 | 216 | |
90481622 | 217 | if (!spool) |
1c5ecae3 | 218 | return delta; |
90481622 | 219 | |
db71ef79 | 220 | spin_lock_irqsave(&spool->lock, flags); |
1c5ecae3 MK |
221 | |
222 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
223 | spool->used_hpages -= delta; | |
224 | ||
09a95e29 MK |
225 | /* minimum size accounting */ |
226 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
227 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
228 | ret = 0; | |
229 | else | |
230 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
231 | ||
232 | spool->rsv_hpages += delta; | |
233 | if (spool->rsv_hpages > spool->min_hpages) | |
234 | spool->rsv_hpages = spool->min_hpages; | |
235 | } | |
236 | ||
237 | /* | |
238 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
239 | * quota reference, free it now. | |
240 | */ | |
db71ef79 | 241 | unlock_or_release_subpool(spool, flags); |
1c5ecae3 MK |
242 | |
243 | return ret; | |
90481622 DG |
244 | } |
245 | ||
246 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
247 | { | |
248 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
249 | } | |
250 | ||
251 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
252 | { | |
496ad9aa | 253 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
254 | } |
255 | ||
0db9d74e MA |
256 | /* Helper that removes a struct file_region from the resv_map cache and returns |
257 | * it for use. | |
258 | */ | |
259 | static struct file_region * | |
260 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
261 | { | |
262 | struct file_region *nrg = NULL; | |
263 | ||
264 | VM_BUG_ON(resv->region_cache_count <= 0); | |
265 | ||
266 | resv->region_cache_count--; | |
267 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
268 | list_del(&nrg->link); |
269 | ||
270 | nrg->from = from; | |
271 | nrg->to = to; | |
272 | ||
273 | return nrg; | |
274 | } | |
275 | ||
075a61d0 MA |
276 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
277 | struct file_region *rg) | |
278 | { | |
279 | #ifdef CONFIG_CGROUP_HUGETLB | |
280 | nrg->reservation_counter = rg->reservation_counter; | |
281 | nrg->css = rg->css; | |
282 | if (rg->css) | |
283 | css_get(rg->css); | |
284 | #endif | |
285 | } | |
286 | ||
287 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
288 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
289 | struct hstate *h, | |
290 | struct resv_map *resv, | |
291 | struct file_region *nrg) | |
292 | { | |
293 | #ifdef CONFIG_CGROUP_HUGETLB | |
294 | if (h_cg) { | |
295 | nrg->reservation_counter = | |
296 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
297 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
298 | /* |
299 | * The caller will hold exactly one h_cg->css reference for the | |
300 | * whole contiguous reservation region. But this area might be | |
301 | * scattered when there are already some file_regions reside in | |
302 | * it. As a result, many file_regions may share only one css | |
303 | * reference. In order to ensure that one file_region must hold | |
304 | * exactly one h_cg->css reference, we should do css_get for | |
305 | * each file_region and leave the reference held by caller | |
306 | * untouched. | |
307 | */ | |
308 | css_get(&h_cg->css); | |
075a61d0 MA |
309 | if (!resv->pages_per_hpage) |
310 | resv->pages_per_hpage = pages_per_huge_page(h); | |
311 | /* pages_per_hpage should be the same for all entries in | |
312 | * a resv_map. | |
313 | */ | |
314 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
315 | } else { | |
316 | nrg->reservation_counter = NULL; | |
317 | nrg->css = NULL; | |
318 | } | |
319 | #endif | |
320 | } | |
321 | ||
d85aecf2 ML |
322 | static void put_uncharge_info(struct file_region *rg) |
323 | { | |
324 | #ifdef CONFIG_CGROUP_HUGETLB | |
325 | if (rg->css) | |
326 | css_put(rg->css); | |
327 | #endif | |
328 | } | |
329 | ||
a9b3f867 MA |
330 | static bool has_same_uncharge_info(struct file_region *rg, |
331 | struct file_region *org) | |
332 | { | |
333 | #ifdef CONFIG_CGROUP_HUGETLB | |
334 | return rg && org && | |
335 | rg->reservation_counter == org->reservation_counter && | |
336 | rg->css == org->css; | |
337 | ||
338 | #else | |
339 | return true; | |
340 | #endif | |
341 | } | |
342 | ||
343 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
344 | { | |
345 | struct file_region *nrg = NULL, *prg = NULL; | |
346 | ||
347 | prg = list_prev_entry(rg, link); | |
348 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
349 | has_same_uncharge_info(prg, rg)) { | |
350 | prg->to = rg->to; | |
351 | ||
352 | list_del(&rg->link); | |
d85aecf2 | 353 | put_uncharge_info(rg); |
a9b3f867 MA |
354 | kfree(rg); |
355 | ||
7db5e7b6 | 356 | rg = prg; |
a9b3f867 MA |
357 | } |
358 | ||
359 | nrg = list_next_entry(rg, link); | |
360 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
361 | has_same_uncharge_info(nrg, rg)) { | |
362 | nrg->from = rg->from; | |
363 | ||
364 | list_del(&rg->link); | |
d85aecf2 | 365 | put_uncharge_info(rg); |
a9b3f867 | 366 | kfree(rg); |
a9b3f867 MA |
367 | } |
368 | } | |
369 | ||
2103cf9c PX |
370 | static inline long |
371 | hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, | |
372 | long to, struct hstate *h, struct hugetlb_cgroup *cg, | |
373 | long *regions_needed) | |
374 | { | |
375 | struct file_region *nrg; | |
376 | ||
377 | if (!regions_needed) { | |
378 | nrg = get_file_region_entry_from_cache(map, from, to); | |
379 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
380 | list_add(&nrg->link, rg->link.prev); | |
381 | coalesce_file_region(map, nrg); | |
382 | } else | |
383 | *regions_needed += 1; | |
384 | ||
385 | return to - from; | |
386 | } | |
387 | ||
972a3da3 WY |
388 | /* |
389 | * Must be called with resv->lock held. | |
390 | * | |
391 | * Calling this with regions_needed != NULL will count the number of pages | |
392 | * to be added but will not modify the linked list. And regions_needed will | |
393 | * indicate the number of file_regions needed in the cache to carry out to add | |
394 | * the regions for this range. | |
d75c6af9 MA |
395 | */ |
396 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 397 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 398 | struct hstate *h, long *regions_needed) |
d75c6af9 | 399 | { |
0db9d74e | 400 | long add = 0; |
d75c6af9 | 401 | struct list_head *head = &resv->regions; |
0db9d74e | 402 | long last_accounted_offset = f; |
2103cf9c | 403 | struct file_region *rg = NULL, *trg = NULL; |
d75c6af9 | 404 | |
0db9d74e MA |
405 | if (regions_needed) |
406 | *regions_needed = 0; | |
d75c6af9 | 407 | |
0db9d74e MA |
408 | /* In this loop, we essentially handle an entry for the range |
409 | * [last_accounted_offset, rg->from), at every iteration, with some | |
410 | * bounds checking. | |
411 | */ | |
412 | list_for_each_entry_safe(rg, trg, head, link) { | |
413 | /* Skip irrelevant regions that start before our range. */ | |
414 | if (rg->from < f) { | |
415 | /* If this region ends after the last accounted offset, | |
416 | * then we need to update last_accounted_offset. | |
417 | */ | |
418 | if (rg->to > last_accounted_offset) | |
419 | last_accounted_offset = rg->to; | |
420 | continue; | |
421 | } | |
d75c6af9 | 422 | |
0db9d74e MA |
423 | /* When we find a region that starts beyond our range, we've |
424 | * finished. | |
425 | */ | |
ca7e0457 | 426 | if (rg->from >= t) |
d75c6af9 MA |
427 | break; |
428 | ||
0db9d74e MA |
429 | /* Add an entry for last_accounted_offset -> rg->from, and |
430 | * update last_accounted_offset. | |
431 | */ | |
2103cf9c PX |
432 | if (rg->from > last_accounted_offset) |
433 | add += hugetlb_resv_map_add(resv, rg, | |
434 | last_accounted_offset, | |
435 | rg->from, h, h_cg, | |
436 | regions_needed); | |
0db9d74e MA |
437 | |
438 | last_accounted_offset = rg->to; | |
439 | } | |
440 | ||
441 | /* Handle the case where our range extends beyond | |
442 | * last_accounted_offset. | |
443 | */ | |
2103cf9c PX |
444 | if (last_accounted_offset < t) |
445 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
446 | t, h, h_cg, regions_needed); | |
0db9d74e MA |
447 | |
448 | VM_BUG_ON(add < 0); | |
449 | return add; | |
450 | } | |
451 | ||
452 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
453 | */ | |
454 | static int allocate_file_region_entries(struct resv_map *resv, | |
455 | int regions_needed) | |
456 | __must_hold(&resv->lock) | |
457 | { | |
458 | struct list_head allocated_regions; | |
459 | int to_allocate = 0, i = 0; | |
460 | struct file_region *trg = NULL, *rg = NULL; | |
461 | ||
462 | VM_BUG_ON(regions_needed < 0); | |
463 | ||
464 | INIT_LIST_HEAD(&allocated_regions); | |
465 | ||
466 | /* | |
467 | * Check for sufficient descriptors in the cache to accommodate | |
468 | * the number of in progress add operations plus regions_needed. | |
469 | * | |
470 | * This is a while loop because when we drop the lock, some other call | |
471 | * to region_add or region_del may have consumed some region_entries, | |
472 | * so we keep looping here until we finally have enough entries for | |
473 | * (adds_in_progress + regions_needed). | |
474 | */ | |
475 | while (resv->region_cache_count < | |
476 | (resv->adds_in_progress + regions_needed)) { | |
477 | to_allocate = resv->adds_in_progress + regions_needed - | |
478 | resv->region_cache_count; | |
479 | ||
480 | /* At this point, we should have enough entries in the cache | |
f0953a1b | 481 | * for all the existing adds_in_progress. We should only be |
0db9d74e | 482 | * needing to allocate for regions_needed. |
d75c6af9 | 483 | */ |
0db9d74e MA |
484 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
485 | ||
486 | spin_unlock(&resv->lock); | |
487 | for (i = 0; i < to_allocate; i++) { | |
488 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
489 | if (!trg) | |
490 | goto out_of_memory; | |
491 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 492 | } |
d75c6af9 | 493 | |
0db9d74e MA |
494 | spin_lock(&resv->lock); |
495 | ||
d3ec7b6e WY |
496 | list_splice(&allocated_regions, &resv->region_cache); |
497 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
498 | } |
499 | ||
0db9d74e | 500 | return 0; |
d75c6af9 | 501 | |
0db9d74e MA |
502 | out_of_memory: |
503 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
504 | list_del(&rg->link); | |
505 | kfree(rg); | |
506 | } | |
507 | return -ENOMEM; | |
d75c6af9 MA |
508 | } |
509 | ||
1dd308a7 MK |
510 | /* |
511 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
512 | * map. Regions will be taken from the cache to fill in this range. |
513 | * Sufficient regions should exist in the cache due to the previous | |
514 | * call to region_chg with the same range, but in some cases the cache will not | |
515 | * have sufficient entries due to races with other code doing region_add or | |
516 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 517 | * |
0db9d74e MA |
518 | * regions_needed is the out value provided by a previous call to region_chg. |
519 | * | |
520 | * Return the number of new huge pages added to the map. This number is greater | |
521 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 522 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
523 | * region_add of regions of length 1 never allocate file_regions and cannot |
524 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
525 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 526 | */ |
0db9d74e | 527 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
528 | long in_regions_needed, struct hstate *h, |
529 | struct hugetlb_cgroup *h_cg) | |
96822904 | 530 | { |
0db9d74e | 531 | long add = 0, actual_regions_needed = 0; |
96822904 | 532 | |
7b24d861 | 533 | spin_lock(&resv->lock); |
0db9d74e MA |
534 | retry: |
535 | ||
536 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
537 | add_reservation_in_range(resv, f, t, NULL, NULL, |
538 | &actual_regions_needed); | |
96822904 | 539 | |
5e911373 | 540 | /* |
0db9d74e MA |
541 | * Check for sufficient descriptors in the cache to accommodate |
542 | * this add operation. Note that actual_regions_needed may be greater | |
543 | * than in_regions_needed, as the resv_map may have been modified since | |
544 | * the region_chg call. In this case, we need to make sure that we | |
545 | * allocate extra entries, such that we have enough for all the | |
546 | * existing adds_in_progress, plus the excess needed for this | |
547 | * operation. | |
5e911373 | 548 | */ |
0db9d74e MA |
549 | if (actual_regions_needed > in_regions_needed && |
550 | resv->region_cache_count < | |
551 | resv->adds_in_progress + | |
552 | (actual_regions_needed - in_regions_needed)) { | |
553 | /* region_add operation of range 1 should never need to | |
554 | * allocate file_region entries. | |
555 | */ | |
556 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 557 | |
0db9d74e MA |
558 | if (allocate_file_region_entries( |
559 | resv, actual_regions_needed - in_regions_needed)) { | |
560 | return -ENOMEM; | |
561 | } | |
5e911373 | 562 | |
0db9d74e | 563 | goto retry; |
5e911373 MK |
564 | } |
565 | ||
972a3da3 | 566 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
567 | |
568 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 569 | |
7b24d861 | 570 | spin_unlock(&resv->lock); |
cf3ad20b | 571 | return add; |
96822904 AW |
572 | } |
573 | ||
1dd308a7 MK |
574 | /* |
575 | * Examine the existing reserve map and determine how many | |
576 | * huge pages in the specified range [f, t) are NOT currently | |
577 | * represented. This routine is called before a subsequent | |
578 | * call to region_add that will actually modify the reserve | |
579 | * map to add the specified range [f, t). region_chg does | |
580 | * not change the number of huge pages represented by the | |
0db9d74e MA |
581 | * map. A number of new file_region structures is added to the cache as a |
582 | * placeholder, for the subsequent region_add call to use. At least 1 | |
583 | * file_region structure is added. | |
584 | * | |
585 | * out_regions_needed is the number of regions added to the | |
586 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
587 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
588 | * |
589 | * Returns the number of huge pages that need to be added to the existing | |
590 | * reservation map for the range [f, t). This number is greater or equal to | |
591 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
592 | * is needed and can not be allocated. | |
1dd308a7 | 593 | */ |
0db9d74e MA |
594 | static long region_chg(struct resv_map *resv, long f, long t, |
595 | long *out_regions_needed) | |
96822904 | 596 | { |
96822904 AW |
597 | long chg = 0; |
598 | ||
7b24d861 | 599 | spin_lock(&resv->lock); |
5e911373 | 600 | |
972a3da3 | 601 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 602 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 603 | out_regions_needed); |
5e911373 | 604 | |
0db9d74e MA |
605 | if (*out_regions_needed == 0) |
606 | *out_regions_needed = 1; | |
5e911373 | 607 | |
0db9d74e MA |
608 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
609 | return -ENOMEM; | |
5e911373 | 610 | |
0db9d74e | 611 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 612 | |
7b24d861 | 613 | spin_unlock(&resv->lock); |
96822904 AW |
614 | return chg; |
615 | } | |
616 | ||
5e911373 MK |
617 | /* |
618 | * Abort the in progress add operation. The adds_in_progress field | |
619 | * of the resv_map keeps track of the operations in progress between | |
620 | * calls to region_chg and region_add. Operations are sometimes | |
621 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
622 | * is called to decrement the adds_in_progress counter. regions_needed |
623 | * is the value returned by the region_chg call, it is used to decrement | |
624 | * the adds_in_progress counter. | |
5e911373 MK |
625 | * |
626 | * NOTE: The range arguments [f, t) are not needed or used in this | |
627 | * routine. They are kept to make reading the calling code easier as | |
628 | * arguments will match the associated region_chg call. | |
629 | */ | |
0db9d74e MA |
630 | static void region_abort(struct resv_map *resv, long f, long t, |
631 | long regions_needed) | |
5e911373 MK |
632 | { |
633 | spin_lock(&resv->lock); | |
634 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 635 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
636 | spin_unlock(&resv->lock); |
637 | } | |
638 | ||
1dd308a7 | 639 | /* |
feba16e2 MK |
640 | * Delete the specified range [f, t) from the reserve map. If the |
641 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
642 | * should be deleted. Locate the regions which intersect [f, t) | |
643 | * and either trim, delete or split the existing regions. | |
644 | * | |
645 | * Returns the number of huge pages deleted from the reserve map. | |
646 | * In the normal case, the return value is zero or more. In the | |
647 | * case where a region must be split, a new region descriptor must | |
648 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
649 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
650 | * a region and possibly return -ENOMEM. Callers specifying | |
651 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 652 | */ |
feba16e2 | 653 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 654 | { |
1406ec9b | 655 | struct list_head *head = &resv->regions; |
96822904 | 656 | struct file_region *rg, *trg; |
feba16e2 MK |
657 | struct file_region *nrg = NULL; |
658 | long del = 0; | |
96822904 | 659 | |
feba16e2 | 660 | retry: |
7b24d861 | 661 | spin_lock(&resv->lock); |
feba16e2 | 662 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
663 | /* |
664 | * Skip regions before the range to be deleted. file_region | |
665 | * ranges are normally of the form [from, to). However, there | |
666 | * may be a "placeholder" entry in the map which is of the form | |
667 | * (from, to) with from == to. Check for placeholder entries | |
668 | * at the beginning of the range to be deleted. | |
669 | */ | |
670 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 671 | continue; |
dbe409e4 | 672 | |
feba16e2 | 673 | if (rg->from >= t) |
96822904 | 674 | break; |
96822904 | 675 | |
feba16e2 MK |
676 | if (f > rg->from && t < rg->to) { /* Must split region */ |
677 | /* | |
678 | * Check for an entry in the cache before dropping | |
679 | * lock and attempting allocation. | |
680 | */ | |
681 | if (!nrg && | |
682 | resv->region_cache_count > resv->adds_in_progress) { | |
683 | nrg = list_first_entry(&resv->region_cache, | |
684 | struct file_region, | |
685 | link); | |
686 | list_del(&nrg->link); | |
687 | resv->region_cache_count--; | |
688 | } | |
96822904 | 689 | |
feba16e2 MK |
690 | if (!nrg) { |
691 | spin_unlock(&resv->lock); | |
692 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
693 | if (!nrg) | |
694 | return -ENOMEM; | |
695 | goto retry; | |
696 | } | |
697 | ||
698 | del += t - f; | |
79aa925b | 699 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 700 | resv, rg, t - f, false); |
feba16e2 MK |
701 | |
702 | /* New entry for end of split region */ | |
703 | nrg->from = t; | |
704 | nrg->to = rg->to; | |
075a61d0 MA |
705 | |
706 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
707 | ||
feba16e2 MK |
708 | INIT_LIST_HEAD(&nrg->link); |
709 | ||
710 | /* Original entry is trimmed */ | |
711 | rg->to = f; | |
712 | ||
713 | list_add(&nrg->link, &rg->link); | |
714 | nrg = NULL; | |
96822904 | 715 | break; |
feba16e2 MK |
716 | } |
717 | ||
718 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
719 | del += rg->to - rg->from; | |
075a61d0 | 720 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 721 | rg->to - rg->from, true); |
feba16e2 MK |
722 | list_del(&rg->link); |
723 | kfree(rg); | |
724 | continue; | |
725 | } | |
726 | ||
727 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 728 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 729 | t - rg->from, false); |
075a61d0 | 730 | |
79aa925b MK |
731 | del += t - rg->from; |
732 | rg->from = t; | |
733 | } else { /* Trim end of region */ | |
075a61d0 | 734 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 735 | rg->to - f, false); |
79aa925b MK |
736 | |
737 | del += rg->to - f; | |
738 | rg->to = f; | |
feba16e2 | 739 | } |
96822904 | 740 | } |
7b24d861 | 741 | |
7b24d861 | 742 | spin_unlock(&resv->lock); |
feba16e2 MK |
743 | kfree(nrg); |
744 | return del; | |
96822904 AW |
745 | } |
746 | ||
b5cec28d MK |
747 | /* |
748 | * A rare out of memory error was encountered which prevented removal of | |
749 | * the reserve map region for a page. The huge page itself was free'ed | |
750 | * and removed from the page cache. This routine will adjust the subpool | |
751 | * usage count, and the global reserve count if needed. By incrementing | |
752 | * these counts, the reserve map entry which could not be deleted will | |
753 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
754 | * counts. | |
755 | */ | |
72e2936c | 756 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
757 | { |
758 | struct hugepage_subpool *spool = subpool_inode(inode); | |
759 | long rsv_adjust; | |
da56388c | 760 | bool reserved = false; |
b5cec28d MK |
761 | |
762 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 763 | if (rsv_adjust > 0) { |
b5cec28d MK |
764 | struct hstate *h = hstate_inode(inode); |
765 | ||
da56388c ML |
766 | if (!hugetlb_acct_memory(h, 1)) |
767 | reserved = true; | |
768 | } else if (!rsv_adjust) { | |
769 | reserved = true; | |
b5cec28d | 770 | } |
da56388c ML |
771 | |
772 | if (!reserved) | |
773 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
774 | } |
775 | ||
1dd308a7 MK |
776 | /* |
777 | * Count and return the number of huge pages in the reserve map | |
778 | * that intersect with the range [f, t). | |
779 | */ | |
1406ec9b | 780 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 781 | { |
1406ec9b | 782 | struct list_head *head = &resv->regions; |
84afd99b AW |
783 | struct file_region *rg; |
784 | long chg = 0; | |
785 | ||
7b24d861 | 786 | spin_lock(&resv->lock); |
84afd99b AW |
787 | /* Locate each segment we overlap with, and count that overlap. */ |
788 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
789 | long seg_from; |
790 | long seg_to; | |
84afd99b AW |
791 | |
792 | if (rg->to <= f) | |
793 | continue; | |
794 | if (rg->from >= t) | |
795 | break; | |
796 | ||
797 | seg_from = max(rg->from, f); | |
798 | seg_to = min(rg->to, t); | |
799 | ||
800 | chg += seg_to - seg_from; | |
801 | } | |
7b24d861 | 802 | spin_unlock(&resv->lock); |
84afd99b AW |
803 | |
804 | return chg; | |
805 | } | |
806 | ||
e7c4b0bf AW |
807 | /* |
808 | * Convert the address within this vma to the page offset within | |
809 | * the mapping, in pagecache page units; huge pages here. | |
810 | */ | |
a5516438 AK |
811 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
812 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 813 | { |
a5516438 AK |
814 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
815 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
816 | } |
817 | ||
0fe6e20b NH |
818 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
819 | unsigned long address) | |
820 | { | |
821 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
822 | } | |
dee41079 | 823 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 824 | |
08fba699 MG |
825 | /* |
826 | * Return the size of the pages allocated when backing a VMA. In the majority | |
827 | * cases this will be same size as used by the page table entries. | |
828 | */ | |
829 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
830 | { | |
05ea8860 DW |
831 | if (vma->vm_ops && vma->vm_ops->pagesize) |
832 | return vma->vm_ops->pagesize(vma); | |
833 | return PAGE_SIZE; | |
08fba699 | 834 | } |
f340ca0f | 835 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 836 | |
3340289d MG |
837 | /* |
838 | * Return the page size being used by the MMU to back a VMA. In the majority | |
839 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
840 | * architectures where it differs, an architecture-specific 'strong' |
841 | * version of this symbol is required. | |
3340289d | 842 | */ |
09135cc5 | 843 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
844 | { |
845 | return vma_kernel_pagesize(vma); | |
846 | } | |
3340289d | 847 | |
84afd99b AW |
848 | /* |
849 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
850 | * bits of the reservation map pointer, which are always clear due to | |
851 | * alignment. | |
852 | */ | |
853 | #define HPAGE_RESV_OWNER (1UL << 0) | |
854 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 855 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 856 | |
a1e78772 MG |
857 | /* |
858 | * These helpers are used to track how many pages are reserved for | |
859 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
860 | * is guaranteed to have their future faults succeed. | |
861 | * | |
862 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
863 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
864 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
865 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
866 | * |
867 | * The private mapping reservation is represented in a subtly different | |
868 | * manner to a shared mapping. A shared mapping has a region map associated | |
869 | * with the underlying file, this region map represents the backing file | |
870 | * pages which have ever had a reservation assigned which this persists even | |
871 | * after the page is instantiated. A private mapping has a region map | |
872 | * associated with the original mmap which is attached to all VMAs which | |
873 | * reference it, this region map represents those offsets which have consumed | |
874 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 875 | */ |
e7c4b0bf AW |
876 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
877 | { | |
878 | return (unsigned long)vma->vm_private_data; | |
879 | } | |
880 | ||
881 | static void set_vma_private_data(struct vm_area_struct *vma, | |
882 | unsigned long value) | |
883 | { | |
884 | vma->vm_private_data = (void *)value; | |
885 | } | |
886 | ||
e9fe92ae MA |
887 | static void |
888 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
889 | struct hugetlb_cgroup *h_cg, | |
890 | struct hstate *h) | |
891 | { | |
892 | #ifdef CONFIG_CGROUP_HUGETLB | |
893 | if (!h_cg || !h) { | |
894 | resv_map->reservation_counter = NULL; | |
895 | resv_map->pages_per_hpage = 0; | |
896 | resv_map->css = NULL; | |
897 | } else { | |
898 | resv_map->reservation_counter = | |
899 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
900 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
901 | resv_map->css = &h_cg->css; | |
902 | } | |
903 | #endif | |
904 | } | |
905 | ||
9119a41e | 906 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
907 | { |
908 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
909 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
910 | ||
911 | if (!resv_map || !rg) { | |
912 | kfree(resv_map); | |
913 | kfree(rg); | |
84afd99b | 914 | return NULL; |
5e911373 | 915 | } |
84afd99b AW |
916 | |
917 | kref_init(&resv_map->refs); | |
7b24d861 | 918 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
919 | INIT_LIST_HEAD(&resv_map->regions); |
920 | ||
5e911373 | 921 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
922 | /* |
923 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
924 | * fields don't do cgroup accounting. On private mappings, these will be | |
925 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
926 | * reservations are to be un-charged from here. | |
927 | */ | |
928 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
929 | |
930 | INIT_LIST_HEAD(&resv_map->region_cache); | |
931 | list_add(&rg->link, &resv_map->region_cache); | |
932 | resv_map->region_cache_count = 1; | |
933 | ||
84afd99b AW |
934 | return resv_map; |
935 | } | |
936 | ||
9119a41e | 937 | void resv_map_release(struct kref *ref) |
84afd99b AW |
938 | { |
939 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
940 | struct list_head *head = &resv_map->region_cache; |
941 | struct file_region *rg, *trg; | |
84afd99b AW |
942 | |
943 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 944 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
945 | |
946 | /* ... and any entries left in the cache */ | |
947 | list_for_each_entry_safe(rg, trg, head, link) { | |
948 | list_del(&rg->link); | |
949 | kfree(rg); | |
950 | } | |
951 | ||
952 | VM_BUG_ON(resv_map->adds_in_progress); | |
953 | ||
84afd99b AW |
954 | kfree(resv_map); |
955 | } | |
956 | ||
4e35f483 JK |
957 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
958 | { | |
f27a5136 MK |
959 | /* |
960 | * At inode evict time, i_mapping may not point to the original | |
961 | * address space within the inode. This original address space | |
962 | * contains the pointer to the resv_map. So, always use the | |
963 | * address space embedded within the inode. | |
964 | * The VERY common case is inode->mapping == &inode->i_data but, | |
965 | * this may not be true for device special inodes. | |
966 | */ | |
967 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
968 | } |
969 | ||
84afd99b | 970 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 971 | { |
81d1b09c | 972 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
973 | if (vma->vm_flags & VM_MAYSHARE) { |
974 | struct address_space *mapping = vma->vm_file->f_mapping; | |
975 | struct inode *inode = mapping->host; | |
976 | ||
977 | return inode_resv_map(inode); | |
978 | ||
979 | } else { | |
84afd99b AW |
980 | return (struct resv_map *)(get_vma_private_data(vma) & |
981 | ~HPAGE_RESV_MASK); | |
4e35f483 | 982 | } |
a1e78772 MG |
983 | } |
984 | ||
84afd99b | 985 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 986 | { |
81d1b09c SL |
987 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
988 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 989 | |
84afd99b AW |
990 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
991 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
992 | } |
993 | ||
994 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
995 | { | |
81d1b09c SL |
996 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
997 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
998 | |
999 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
1000 | } |
1001 | ||
1002 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
1003 | { | |
81d1b09c | 1004 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
1005 | |
1006 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
1007 | } |
1008 | ||
04f2cbe3 | 1009 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
1010 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
1011 | { | |
81d1b09c | 1012 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 1013 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
1014 | vma->vm_private_data = (void *)0; |
1015 | } | |
1016 | ||
550a7d60 MA |
1017 | /* |
1018 | * Reset and decrement one ref on hugepage private reservation. | |
1019 | * Called with mm->mmap_sem writer semaphore held. | |
1020 | * This function should be only used by move_vma() and operate on | |
1021 | * same sized vma. It should never come here with last ref on the | |
1022 | * reservation. | |
1023 | */ | |
1024 | void clear_vma_resv_huge_pages(struct vm_area_struct *vma) | |
1025 | { | |
1026 | /* | |
1027 | * Clear the old hugetlb private page reservation. | |
1028 | * It has already been transferred to new_vma. | |
1029 | * | |
1030 | * During a mremap() operation of a hugetlb vma we call move_vma() | |
1031 | * which copies vma into new_vma and unmaps vma. After the copy | |
1032 | * operation both new_vma and vma share a reference to the resv_map | |
1033 | * struct, and at that point vma is about to be unmapped. We don't | |
1034 | * want to return the reservation to the pool at unmap of vma because | |
1035 | * the reservation still lives on in new_vma, so simply decrement the | |
1036 | * ref here and remove the resv_map reference from this vma. | |
1037 | */ | |
1038 | struct resv_map *reservations = vma_resv_map(vma); | |
1039 | ||
1040 | if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) | |
1041 | kref_put(&reservations->refs, resv_map_release); | |
1042 | ||
1043 | reset_vma_resv_huge_pages(vma); | |
1044 | } | |
1045 | ||
a1e78772 | 1046 | /* Returns true if the VMA has associated reserve pages */ |
559ec2f8 | 1047 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1048 | { |
af0ed73e JK |
1049 | if (vma->vm_flags & VM_NORESERVE) { |
1050 | /* | |
1051 | * This address is already reserved by other process(chg == 0), | |
1052 | * so, we should decrement reserved count. Without decrementing, | |
1053 | * reserve count remains after releasing inode, because this | |
1054 | * allocated page will go into page cache and is regarded as | |
1055 | * coming from reserved pool in releasing step. Currently, we | |
1056 | * don't have any other solution to deal with this situation | |
1057 | * properly, so add work-around here. | |
1058 | */ | |
1059 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1060 | return true; |
af0ed73e | 1061 | else |
559ec2f8 | 1062 | return false; |
af0ed73e | 1063 | } |
a63884e9 JK |
1064 | |
1065 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1066 | if (vma->vm_flags & VM_MAYSHARE) { |
1067 | /* | |
1068 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1069 | * be a region map for all pages. The only situation where | |
1070 | * there is no region map is if a hole was punched via | |
7c8de358 | 1071 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1072 | * use. This situation is indicated if chg != 0. |
1073 | */ | |
1074 | if (chg) | |
1075 | return false; | |
1076 | else | |
1077 | return true; | |
1078 | } | |
a63884e9 JK |
1079 | |
1080 | /* | |
1081 | * Only the process that called mmap() has reserves for | |
1082 | * private mappings. | |
1083 | */ | |
67961f9d MK |
1084 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1085 | /* | |
1086 | * Like the shared case above, a hole punch or truncate | |
1087 | * could have been performed on the private mapping. | |
1088 | * Examine the value of chg to determine if reserves | |
1089 | * actually exist or were previously consumed. | |
1090 | * Very Subtle - The value of chg comes from a previous | |
1091 | * call to vma_needs_reserves(). The reserve map for | |
1092 | * private mappings has different (opposite) semantics | |
1093 | * than that of shared mappings. vma_needs_reserves() | |
1094 | * has already taken this difference in semantics into | |
1095 | * account. Therefore, the meaning of chg is the same | |
1096 | * as in the shared case above. Code could easily be | |
1097 | * combined, but keeping it separate draws attention to | |
1098 | * subtle differences. | |
1099 | */ | |
1100 | if (chg) | |
1101 | return false; | |
1102 | else | |
1103 | return true; | |
1104 | } | |
a63884e9 | 1105 | |
559ec2f8 | 1106 | return false; |
a1e78772 MG |
1107 | } |
1108 | ||
a5516438 | 1109 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1110 | { |
1111 | int nid = page_to_nid(page); | |
9487ca60 MK |
1112 | |
1113 | lockdep_assert_held(&hugetlb_lock); | |
b65a4eda MK |
1114 | VM_BUG_ON_PAGE(page_count(page), page); |
1115 | ||
0edaecfa | 1116 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1117 | h->free_huge_pages++; |
1118 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1119 | SetHPageFreed(page); |
1da177e4 LT |
1120 | } |
1121 | ||
94310cbc | 1122 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1123 | { |
1124 | struct page *page; | |
1a08ae36 | 1125 | bool pin = !!(current->flags & PF_MEMALLOC_PIN); |
bbe88753 | 1126 | |
9487ca60 | 1127 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 | 1128 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
8e3560d9 | 1129 | if (pin && !is_pinnable_page(page)) |
bbe88753 | 1130 | continue; |
bf50bab2 | 1131 | |
6664bfc8 WY |
1132 | if (PageHWPoison(page)) |
1133 | continue; | |
1134 | ||
1135 | list_move(&page->lru, &h->hugepage_activelist); | |
1136 | set_page_refcounted(page); | |
6c037149 | 1137 | ClearHPageFreed(page); |
6664bfc8 WY |
1138 | h->free_huge_pages--; |
1139 | h->free_huge_pages_node[nid]--; | |
1140 | return page; | |
bbe88753 JK |
1141 | } |
1142 | ||
6664bfc8 | 1143 | return NULL; |
bf50bab2 NH |
1144 | } |
1145 | ||
3e59fcb0 MH |
1146 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1147 | nodemask_t *nmask) | |
94310cbc | 1148 | { |
3e59fcb0 MH |
1149 | unsigned int cpuset_mems_cookie; |
1150 | struct zonelist *zonelist; | |
1151 | struct zone *zone; | |
1152 | struct zoneref *z; | |
98fa15f3 | 1153 | int node = NUMA_NO_NODE; |
94310cbc | 1154 | |
3e59fcb0 MH |
1155 | zonelist = node_zonelist(nid, gfp_mask); |
1156 | ||
1157 | retry_cpuset: | |
1158 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1159 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1160 | struct page *page; | |
1161 | ||
1162 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1163 | continue; | |
1164 | /* | |
1165 | * no need to ask again on the same node. Pool is node rather than | |
1166 | * zone aware | |
1167 | */ | |
1168 | if (zone_to_nid(zone) == node) | |
1169 | continue; | |
1170 | node = zone_to_nid(zone); | |
94310cbc | 1171 | |
94310cbc AK |
1172 | page = dequeue_huge_page_node_exact(h, node); |
1173 | if (page) | |
1174 | return page; | |
1175 | } | |
3e59fcb0 MH |
1176 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1177 | goto retry_cpuset; | |
1178 | ||
94310cbc AK |
1179 | return NULL; |
1180 | } | |
1181 | ||
a5516438 AK |
1182 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1183 | struct vm_area_struct *vma, | |
af0ed73e JK |
1184 | unsigned long address, int avoid_reserve, |
1185 | long chg) | |
1da177e4 | 1186 | { |
cfcaa66f | 1187 | struct page *page = NULL; |
480eccf9 | 1188 | struct mempolicy *mpol; |
04ec6264 | 1189 | gfp_t gfp_mask; |
3e59fcb0 | 1190 | nodemask_t *nodemask; |
04ec6264 | 1191 | int nid; |
1da177e4 | 1192 | |
a1e78772 MG |
1193 | /* |
1194 | * A child process with MAP_PRIVATE mappings created by their parent | |
1195 | * have no page reserves. This check ensures that reservations are | |
1196 | * not "stolen". The child may still get SIGKILLed | |
1197 | */ | |
af0ed73e | 1198 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1199 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1200 | goto err; |
a1e78772 | 1201 | |
04f2cbe3 | 1202 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1203 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1204 | goto err; |
04f2cbe3 | 1205 | |
04ec6264 VB |
1206 | gfp_mask = htlb_alloc_mask(h); |
1207 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
1208 | |
1209 | if (mpol_is_preferred_many(mpol)) { | |
1210 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1211 | ||
1212 | /* Fallback to all nodes if page==NULL */ | |
1213 | nodemask = NULL; | |
1214 | } | |
1215 | ||
1216 | if (!page) | |
1217 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1218 | ||
3e59fcb0 | 1219 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { |
d6995da3 | 1220 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1221 | h->resv_huge_pages--; |
1da177e4 | 1222 | } |
cc9a6c87 | 1223 | |
52cd3b07 | 1224 | mpol_cond_put(mpol); |
1da177e4 | 1225 | return page; |
cc9a6c87 MG |
1226 | |
1227 | err: | |
cc9a6c87 | 1228 | return NULL; |
1da177e4 LT |
1229 | } |
1230 | ||
1cac6f2c LC |
1231 | /* |
1232 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1233 | * We may have allocated or freed a huge page based on a different | |
1234 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1235 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1236 | * node for alloc or free. | |
1237 | */ | |
1238 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1239 | { | |
0edaf86c | 1240 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1241 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1242 | ||
1243 | return nid; | |
1244 | } | |
1245 | ||
1246 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1247 | { | |
1248 | if (!node_isset(nid, *nodes_allowed)) | |
1249 | nid = next_node_allowed(nid, nodes_allowed); | |
1250 | return nid; | |
1251 | } | |
1252 | ||
1253 | /* | |
1254 | * returns the previously saved node ["this node"] from which to | |
1255 | * allocate a persistent huge page for the pool and advance the | |
1256 | * next node from which to allocate, handling wrap at end of node | |
1257 | * mask. | |
1258 | */ | |
1259 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1260 | nodemask_t *nodes_allowed) | |
1261 | { | |
1262 | int nid; | |
1263 | ||
1264 | VM_BUG_ON(!nodes_allowed); | |
1265 | ||
1266 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1267 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1268 | ||
1269 | return nid; | |
1270 | } | |
1271 | ||
1272 | /* | |
10c6ec49 | 1273 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1274 | * node ["this node"] from which to free a huge page. Advance the |
1275 | * next node id whether or not we find a free huge page to free so | |
1276 | * that the next attempt to free addresses the next node. | |
1277 | */ | |
1278 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1279 | { | |
1280 | int nid; | |
1281 | ||
1282 | VM_BUG_ON(!nodes_allowed); | |
1283 | ||
1284 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1285 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1286 | ||
1287 | return nid; | |
1288 | } | |
1289 | ||
1290 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1291 | for (nr_nodes = nodes_weight(*mask); \ | |
1292 | nr_nodes > 0 && \ | |
1293 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1294 | nr_nodes--) | |
1295 | ||
1296 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1297 | for (nr_nodes = nodes_weight(*mask); \ | |
1298 | nr_nodes > 0 && \ | |
1299 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1300 | nr_nodes--) | |
1301 | ||
8531fc6f | 1302 | /* used to demote non-gigantic_huge pages as well */ |
34d9e35b MK |
1303 | static void __destroy_compound_gigantic_page(struct page *page, |
1304 | unsigned int order, bool demote) | |
944d9fec LC |
1305 | { |
1306 | int i; | |
1307 | int nr_pages = 1 << order; | |
1308 | struct page *p = page + 1; | |
1309 | ||
c8cc708a | 1310 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1311 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1312 | |
944d9fec | 1313 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
a01f4390 | 1314 | p->mapping = NULL; |
1d798ca3 | 1315 | clear_compound_head(p); |
34d9e35b MK |
1316 | if (!demote) |
1317 | set_page_refcounted(p); | |
944d9fec LC |
1318 | } |
1319 | ||
1320 | set_compound_order(page, 0); | |
ba9c1201 | 1321 | page[1].compound_nr = 0; |
944d9fec LC |
1322 | __ClearPageHead(page); |
1323 | } | |
1324 | ||
8531fc6f MK |
1325 | static void destroy_compound_hugetlb_page_for_demote(struct page *page, |
1326 | unsigned int order) | |
1327 | { | |
1328 | __destroy_compound_gigantic_page(page, order, true); | |
1329 | } | |
1330 | ||
1331 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE | |
34d9e35b MK |
1332 | static void destroy_compound_gigantic_page(struct page *page, |
1333 | unsigned int order) | |
1334 | { | |
1335 | __destroy_compound_gigantic_page(page, order, false); | |
1336 | } | |
1337 | ||
d00181b9 | 1338 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1339 | { |
cf11e85f RG |
1340 | /* |
1341 | * If the page isn't allocated using the cma allocator, | |
1342 | * cma_release() returns false. | |
1343 | */ | |
dbda8fea BS |
1344 | #ifdef CONFIG_CMA |
1345 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1346 | return; |
dbda8fea | 1347 | #endif |
cf11e85f | 1348 | |
944d9fec LC |
1349 | free_contig_range(page_to_pfn(page), 1 << order); |
1350 | } | |
1351 | ||
4eb0716e | 1352 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1353 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1354 | int nid, nodemask_t *nodemask) | |
944d9fec | 1355 | { |
04adbc3f | 1356 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1357 | if (nid == NUMA_NO_NODE) |
1358 | nid = numa_mem_id(); | |
944d9fec | 1359 | |
dbda8fea BS |
1360 | #ifdef CONFIG_CMA |
1361 | { | |
cf11e85f RG |
1362 | struct page *page; |
1363 | int node; | |
1364 | ||
953f064a LX |
1365 | if (hugetlb_cma[nid]) { |
1366 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1367 | huge_page_order(h), true); | |
cf11e85f RG |
1368 | if (page) |
1369 | return page; | |
1370 | } | |
953f064a LX |
1371 | |
1372 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1373 | for_each_node_mask(node, *nodemask) { | |
1374 | if (node == nid || !hugetlb_cma[node]) | |
1375 | continue; | |
1376 | ||
1377 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1378 | huge_page_order(h), true); | |
1379 | if (page) | |
1380 | return page; | |
1381 | } | |
1382 | } | |
cf11e85f | 1383 | } |
dbda8fea | 1384 | #endif |
cf11e85f | 1385 | |
5e27a2df | 1386 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1387 | } |
1388 | ||
4eb0716e AG |
1389 | #else /* !CONFIG_CONTIG_ALLOC */ |
1390 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1391 | int nid, nodemask_t *nodemask) | |
1392 | { | |
1393 | return NULL; | |
1394 | } | |
1395 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1396 | |
e1073d1e | 1397 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1398 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1399 | int nid, nodemask_t *nodemask) |
1400 | { | |
1401 | return NULL; | |
1402 | } | |
d00181b9 | 1403 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1404 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1405 | unsigned int order) { } |
944d9fec LC |
1406 | #endif |
1407 | ||
6eb4e88a MK |
1408 | /* |
1409 | * Remove hugetlb page from lists, and update dtor so that page appears | |
34d9e35b MK |
1410 | * as just a compound page. |
1411 | * | |
1412 | * A reference is held on the page, except in the case of demote. | |
6eb4e88a MK |
1413 | * |
1414 | * Must be called with hugetlb lock held. | |
1415 | */ | |
34d9e35b MK |
1416 | static void __remove_hugetlb_page(struct hstate *h, struct page *page, |
1417 | bool adjust_surplus, | |
1418 | bool demote) | |
6eb4e88a MK |
1419 | { |
1420 | int nid = page_to_nid(page); | |
1421 | ||
1422 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1423 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1424 | ||
9487ca60 | 1425 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1426 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1427 | return; | |
1428 | ||
1429 | list_del(&page->lru); | |
1430 | ||
1431 | if (HPageFreed(page)) { | |
1432 | h->free_huge_pages--; | |
1433 | h->free_huge_pages_node[nid]--; | |
1434 | } | |
1435 | if (adjust_surplus) { | |
1436 | h->surplus_huge_pages--; | |
1437 | h->surplus_huge_pages_node[nid]--; | |
1438 | } | |
1439 | ||
e32d20c0 MK |
1440 | /* |
1441 | * Very subtle | |
1442 | * | |
1443 | * For non-gigantic pages set the destructor to the normal compound | |
1444 | * page dtor. This is needed in case someone takes an additional | |
1445 | * temporary ref to the page, and freeing is delayed until they drop | |
1446 | * their reference. | |
1447 | * | |
1448 | * For gigantic pages set the destructor to the null dtor. This | |
1449 | * destructor will never be called. Before freeing the gigantic | |
1450 | * page destroy_compound_gigantic_page will turn the compound page | |
1451 | * into a simple group of pages. After this the destructor does not | |
1452 | * apply. | |
1453 | * | |
1454 | * This handles the case where more than one ref is held when and | |
1455 | * after update_and_free_page is called. | |
34d9e35b MK |
1456 | * |
1457 | * In the case of demote we do not ref count the page as it will soon | |
1458 | * be turned into a page of smaller size. | |
e32d20c0 | 1459 | */ |
34d9e35b MK |
1460 | if (!demote) |
1461 | set_page_refcounted(page); | |
e32d20c0 MK |
1462 | if (hstate_is_gigantic(h)) |
1463 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1464 | else | |
1465 | set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); | |
6eb4e88a MK |
1466 | |
1467 | h->nr_huge_pages--; | |
1468 | h->nr_huge_pages_node[nid]--; | |
1469 | } | |
1470 | ||
34d9e35b MK |
1471 | static void remove_hugetlb_page(struct hstate *h, struct page *page, |
1472 | bool adjust_surplus) | |
1473 | { | |
1474 | __remove_hugetlb_page(h, page, adjust_surplus, false); | |
1475 | } | |
1476 | ||
8531fc6f MK |
1477 | static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, |
1478 | bool adjust_surplus) | |
1479 | { | |
1480 | __remove_hugetlb_page(h, page, adjust_surplus, true); | |
1481 | } | |
1482 | ||
ad2fa371 MS |
1483 | static void add_hugetlb_page(struct hstate *h, struct page *page, |
1484 | bool adjust_surplus) | |
1485 | { | |
1486 | int zeroed; | |
1487 | int nid = page_to_nid(page); | |
1488 | ||
1489 | VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); | |
1490 | ||
1491 | lockdep_assert_held(&hugetlb_lock); | |
1492 | ||
1493 | INIT_LIST_HEAD(&page->lru); | |
1494 | h->nr_huge_pages++; | |
1495 | h->nr_huge_pages_node[nid]++; | |
1496 | ||
1497 | if (adjust_surplus) { | |
1498 | h->surplus_huge_pages++; | |
1499 | h->surplus_huge_pages_node[nid]++; | |
1500 | } | |
1501 | ||
1502 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); | |
1503 | set_page_private(page, 0); | |
1504 | SetHPageVmemmapOptimized(page); | |
1505 | ||
1506 | /* | |
b65a4eda MK |
1507 | * This page is about to be managed by the hugetlb allocator and |
1508 | * should have no users. Drop our reference, and check for others | |
1509 | * just in case. | |
ad2fa371 MS |
1510 | */ |
1511 | zeroed = put_page_testzero(page); | |
b65a4eda MK |
1512 | if (!zeroed) |
1513 | /* | |
1514 | * It is VERY unlikely soneone else has taken a ref on | |
1515 | * the page. In this case, we simply return as the | |
1516 | * hugetlb destructor (free_huge_page) will be called | |
1517 | * when this other ref is dropped. | |
1518 | */ | |
1519 | return; | |
1520 | ||
ad2fa371 MS |
1521 | arch_clear_hugepage_flags(page); |
1522 | enqueue_huge_page(h, page); | |
1523 | } | |
1524 | ||
b65d4adb | 1525 | static void __update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1526 | { |
1527 | int i; | |
dbfee5ae | 1528 | struct page *subpage = page; |
a5516438 | 1529 | |
4eb0716e | 1530 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1531 | return; |
18229df5 | 1532 | |
ad2fa371 MS |
1533 | if (alloc_huge_page_vmemmap(h, page)) { |
1534 | spin_lock_irq(&hugetlb_lock); | |
1535 | /* | |
1536 | * If we cannot allocate vmemmap pages, just refuse to free the | |
1537 | * page and put the page back on the hugetlb free list and treat | |
1538 | * as a surplus page. | |
1539 | */ | |
1540 | add_hugetlb_page(h, page, true); | |
1541 | spin_unlock_irq(&hugetlb_lock); | |
1542 | return; | |
1543 | } | |
1544 | ||
dbfee5ae MK |
1545 | for (i = 0; i < pages_per_huge_page(h); |
1546 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1547 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1548 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1549 | 1 << PG_active | 1 << PG_private | |
1550 | 1 << PG_writeback); | |
6af2acb6 | 1551 | } |
a01f4390 MK |
1552 | |
1553 | /* | |
1554 | * Non-gigantic pages demoted from CMA allocated gigantic pages | |
1555 | * need to be given back to CMA in free_gigantic_page. | |
1556 | */ | |
1557 | if (hstate_is_gigantic(h) || | |
1558 | hugetlb_cma_page(page, huge_page_order(h))) { | |
944d9fec LC |
1559 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1560 | free_gigantic_page(page, huge_page_order(h)); | |
1561 | } else { | |
944d9fec LC |
1562 | __free_pages(page, huge_page_order(h)); |
1563 | } | |
6af2acb6 AL |
1564 | } |
1565 | ||
b65d4adb MS |
1566 | /* |
1567 | * As update_and_free_page() can be called under any context, so we cannot | |
1568 | * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the | |
1569 | * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate | |
1570 | * the vmemmap pages. | |
1571 | * | |
1572 | * free_hpage_workfn() locklessly retrieves the linked list of pages to be | |
1573 | * freed and frees them one-by-one. As the page->mapping pointer is going | |
1574 | * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node | |
1575 | * structure of a lockless linked list of huge pages to be freed. | |
1576 | */ | |
1577 | static LLIST_HEAD(hpage_freelist); | |
1578 | ||
1579 | static void free_hpage_workfn(struct work_struct *work) | |
1580 | { | |
1581 | struct llist_node *node; | |
1582 | ||
1583 | node = llist_del_all(&hpage_freelist); | |
1584 | ||
1585 | while (node) { | |
1586 | struct page *page; | |
1587 | struct hstate *h; | |
1588 | ||
1589 | page = container_of((struct address_space **)node, | |
1590 | struct page, mapping); | |
1591 | node = node->next; | |
1592 | page->mapping = NULL; | |
1593 | /* | |
1594 | * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() | |
1595 | * is going to trigger because a previous call to | |
1596 | * remove_hugetlb_page() will set_compound_page_dtor(page, | |
1597 | * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. | |
1598 | */ | |
1599 | h = size_to_hstate(page_size(page)); | |
1600 | ||
1601 | __update_and_free_page(h, page); | |
1602 | ||
1603 | cond_resched(); | |
1604 | } | |
1605 | } | |
1606 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1607 | ||
1608 | static inline void flush_free_hpage_work(struct hstate *h) | |
1609 | { | |
1610 | if (free_vmemmap_pages_per_hpage(h)) | |
1611 | flush_work(&free_hpage_work); | |
1612 | } | |
1613 | ||
1614 | static void update_and_free_page(struct hstate *h, struct page *page, | |
1615 | bool atomic) | |
1616 | { | |
ad2fa371 | 1617 | if (!HPageVmemmapOptimized(page) || !atomic) { |
b65d4adb MS |
1618 | __update_and_free_page(h, page); |
1619 | return; | |
1620 | } | |
1621 | ||
1622 | /* | |
1623 | * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. | |
1624 | * | |
1625 | * Only call schedule_work() if hpage_freelist is previously | |
1626 | * empty. Otherwise, schedule_work() had been called but the workfn | |
1627 | * hasn't retrieved the list yet. | |
1628 | */ | |
1629 | if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) | |
1630 | schedule_work(&free_hpage_work); | |
1631 | } | |
1632 | ||
10c6ec49 MK |
1633 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1634 | { | |
1635 | struct page *page, *t_page; | |
1636 | ||
1637 | list_for_each_entry_safe(page, t_page, list, lru) { | |
b65d4adb | 1638 | update_and_free_page(h, page, false); |
10c6ec49 MK |
1639 | cond_resched(); |
1640 | } | |
1641 | } | |
1642 | ||
e5ff2159 AK |
1643 | struct hstate *size_to_hstate(unsigned long size) |
1644 | { | |
1645 | struct hstate *h; | |
1646 | ||
1647 | for_each_hstate(h) { | |
1648 | if (huge_page_size(h) == size) | |
1649 | return h; | |
1650 | } | |
1651 | return NULL; | |
1652 | } | |
1653 | ||
db71ef79 | 1654 | void free_huge_page(struct page *page) |
27a85ef1 | 1655 | { |
a5516438 AK |
1656 | /* |
1657 | * Can't pass hstate in here because it is called from the | |
1658 | * compound page destructor. | |
1659 | */ | |
e5ff2159 | 1660 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1661 | int nid = page_to_nid(page); |
d6995da3 | 1662 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1663 | bool restore_reserve; |
db71ef79 | 1664 | unsigned long flags; |
27a85ef1 | 1665 | |
b4330afb MK |
1666 | VM_BUG_ON_PAGE(page_count(page), page); |
1667 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1668 | |
d6995da3 | 1669 | hugetlb_set_page_subpool(page, NULL); |
8ace22bc | 1670 | page->mapping = NULL; |
d6995da3 MK |
1671 | restore_reserve = HPageRestoreReserve(page); |
1672 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1673 | |
1c5ecae3 | 1674 | /* |
d6995da3 | 1675 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1676 | * reservation. If the page was associated with a subpool, there |
1677 | * would have been a page reserved in the subpool before allocation | |
1678 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1679 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1680 | * remove the reserved page from the subpool. |
1c5ecae3 | 1681 | */ |
0919e1b6 MK |
1682 | if (!restore_reserve) { |
1683 | /* | |
1684 | * A return code of zero implies that the subpool will be | |
1685 | * under its minimum size if the reservation is not restored | |
1686 | * after page is free. Therefore, force restore_reserve | |
1687 | * operation. | |
1688 | */ | |
1689 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1690 | restore_reserve = true; | |
1691 | } | |
1c5ecae3 | 1692 | |
db71ef79 | 1693 | spin_lock_irqsave(&hugetlb_lock, flags); |
8f251a3d | 1694 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1695 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1696 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1697 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1698 | pages_per_huge_page(h), page); | |
07443a85 JK |
1699 | if (restore_reserve) |
1700 | h->resv_huge_pages++; | |
1701 | ||
9157c311 | 1702 | if (HPageTemporary(page)) { |
6eb4e88a | 1703 | remove_hugetlb_page(h, page, false); |
db71ef79 | 1704 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1705 | update_and_free_page(h, page, true); |
ab5ac90a | 1706 | } else if (h->surplus_huge_pages_node[nid]) { |
0edaecfa | 1707 | /* remove the page from active list */ |
6eb4e88a | 1708 | remove_hugetlb_page(h, page, true); |
db71ef79 | 1709 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1710 | update_and_free_page(h, page, true); |
7893d1d5 | 1711 | } else { |
5d3a551c | 1712 | arch_clear_hugepage_flags(page); |
a5516438 | 1713 | enqueue_huge_page(h, page); |
db71ef79 | 1714 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1715 | } |
c77c0a8a WL |
1716 | } |
1717 | ||
d3d99fcc OS |
1718 | /* |
1719 | * Must be called with the hugetlb lock held | |
1720 | */ | |
1721 | static void __prep_account_new_huge_page(struct hstate *h, int nid) | |
1722 | { | |
1723 | lockdep_assert_held(&hugetlb_lock); | |
1724 | h->nr_huge_pages++; | |
1725 | h->nr_huge_pages_node[nid]++; | |
1726 | } | |
1727 | ||
f41f2ed4 | 1728 | static void __prep_new_huge_page(struct hstate *h, struct page *page) |
b7ba30c6 | 1729 | { |
f41f2ed4 | 1730 | free_huge_page_vmemmap(h, page); |
0edaecfa | 1731 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1732 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1733 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1734 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1735 | set_hugetlb_cgroup_rsvd(page, NULL); |
d3d99fcc OS |
1736 | } |
1737 | ||
1738 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1739 | { | |
f41f2ed4 | 1740 | __prep_new_huge_page(h, page); |
db71ef79 | 1741 | spin_lock_irq(&hugetlb_lock); |
d3d99fcc | 1742 | __prep_account_new_huge_page(h, nid); |
db71ef79 | 1743 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1744 | } |
1745 | ||
34d9e35b MK |
1746 | static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, |
1747 | bool demote) | |
20a0307c | 1748 | { |
7118fc29 | 1749 | int i, j; |
20a0307c WF |
1750 | int nr_pages = 1 << order; |
1751 | struct page *p = page + 1; | |
1752 | ||
1753 | /* we rely on prep_new_huge_page to set the destructor */ | |
1754 | set_compound_order(page, order); | |
ef5a22be | 1755 | __ClearPageReserved(page); |
de09d31d | 1756 | __SetPageHead(page); |
20a0307c | 1757 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1758 | /* |
1759 | * For gigantic hugepages allocated through bootmem at | |
1760 | * boot, it's safer to be consistent with the not-gigantic | |
1761 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1762 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1763 | * pages may get the reference counting wrong if they see |
1764 | * PG_reserved set on a tail page (despite the head page not | |
1765 | * having PG_reserved set). Enforcing this consistency between | |
1766 | * head and tail pages allows drivers to optimize away a check | |
1767 | * on the head page when they need know if put_page() is needed | |
1768 | * after get_user_pages(). | |
1769 | */ | |
1770 | __ClearPageReserved(p); | |
7118fc29 MK |
1771 | /* |
1772 | * Subtle and very unlikely | |
1773 | * | |
1774 | * Gigantic 'page allocators' such as memblock or cma will | |
1775 | * return a set of pages with each page ref counted. We need | |
1776 | * to turn this set of pages into a compound page with tail | |
1777 | * page ref counts set to zero. Code such as speculative page | |
1778 | * cache adding could take a ref on a 'to be' tail page. | |
1779 | * We need to respect any increased ref count, and only set | |
1780 | * the ref count to zero if count is currently 1. If count | |
416d85ed MK |
1781 | * is not 1, we return an error. An error return indicates |
1782 | * the set of pages can not be converted to a gigantic page. | |
1783 | * The caller who allocated the pages should then discard the | |
1784 | * pages using the appropriate free interface. | |
34d9e35b MK |
1785 | * |
1786 | * In the case of demote, the ref count will be zero. | |
7118fc29 | 1787 | */ |
34d9e35b MK |
1788 | if (!demote) { |
1789 | if (!page_ref_freeze(p, 1)) { | |
1790 | pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); | |
1791 | goto out_error; | |
1792 | } | |
1793 | } else { | |
1794 | VM_BUG_ON_PAGE(page_count(p), p); | |
7118fc29 | 1795 | } |
58a84aa9 | 1796 | set_page_count(p, 0); |
1d798ca3 | 1797 | set_compound_head(p, page); |
20a0307c | 1798 | } |
b4330afb | 1799 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1800 | atomic_set(compound_pincount_ptr(page), 0); |
7118fc29 MK |
1801 | return true; |
1802 | ||
1803 | out_error: | |
1804 | /* undo tail page modifications made above */ | |
1805 | p = page + 1; | |
1806 | for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { | |
1807 | clear_compound_head(p); | |
1808 | set_page_refcounted(p); | |
1809 | } | |
1810 | /* need to clear PG_reserved on remaining tail pages */ | |
1811 | for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) | |
1812 | __ClearPageReserved(p); | |
1813 | set_compound_order(page, 0); | |
1814 | page[1].compound_nr = 0; | |
1815 | __ClearPageHead(page); | |
1816 | return false; | |
20a0307c WF |
1817 | } |
1818 | ||
34d9e35b MK |
1819 | static bool prep_compound_gigantic_page(struct page *page, unsigned int order) |
1820 | { | |
1821 | return __prep_compound_gigantic_page(page, order, false); | |
1822 | } | |
1823 | ||
8531fc6f MK |
1824 | static bool prep_compound_gigantic_page_for_demote(struct page *page, |
1825 | unsigned int order) | |
1826 | { | |
1827 | return __prep_compound_gigantic_page(page, order, true); | |
1828 | } | |
1829 | ||
7795912c AM |
1830 | /* |
1831 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1832 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1833 | * details. | |
1834 | */ | |
20a0307c WF |
1835 | int PageHuge(struct page *page) |
1836 | { | |
20a0307c WF |
1837 | if (!PageCompound(page)) |
1838 | return 0; | |
1839 | ||
1840 | page = compound_head(page); | |
f1e61557 | 1841 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1842 | } |
43131e14 NH |
1843 | EXPORT_SYMBOL_GPL(PageHuge); |
1844 | ||
27c73ae7 AA |
1845 | /* |
1846 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1847 | * normal or transparent huge pages. | |
1848 | */ | |
1849 | int PageHeadHuge(struct page *page_head) | |
1850 | { | |
27c73ae7 AA |
1851 | if (!PageHead(page_head)) |
1852 | return 0; | |
1853 | ||
d4af73e3 | 1854 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1855 | } |
27c73ae7 | 1856 | |
c0d0381a MK |
1857 | /* |
1858 | * Find and lock address space (mapping) in write mode. | |
1859 | * | |
336bf30e MK |
1860 | * Upon entry, the page is locked which means that page_mapping() is |
1861 | * stable. Due to locking order, we can only trylock_write. If we can | |
1862 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1863 | */ |
1864 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1865 | { | |
336bf30e | 1866 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1867 | |
c0d0381a MK |
1868 | if (!mapping) |
1869 | return mapping; | |
1870 | ||
c0d0381a MK |
1871 | if (i_mmap_trylock_write(mapping)) |
1872 | return mapping; | |
1873 | ||
336bf30e | 1874 | return NULL; |
c0d0381a MK |
1875 | } |
1876 | ||
fe19bd3d | 1877 | pgoff_t hugetlb_basepage_index(struct page *page) |
13d60f4b ZY |
1878 | { |
1879 | struct page *page_head = compound_head(page); | |
1880 | pgoff_t index = page_index(page_head); | |
1881 | unsigned long compound_idx; | |
1882 | ||
13d60f4b ZY |
1883 | if (compound_order(page_head) >= MAX_ORDER) |
1884 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1885 | else | |
1886 | compound_idx = page - page_head; | |
1887 | ||
1888 | return (index << compound_order(page_head)) + compound_idx; | |
1889 | } | |
1890 | ||
0c397dae | 1891 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1892 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1893 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1894 | { |
af0fb9df | 1895 | int order = huge_page_order(h); |
1da177e4 | 1896 | struct page *page; |
f60858f9 | 1897 | bool alloc_try_hard = true; |
f96efd58 | 1898 | |
f60858f9 MK |
1899 | /* |
1900 | * By default we always try hard to allocate the page with | |
1901 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1902 | * a loop (to adjust global huge page counts) and previous allocation | |
1903 | * failed, do not continue to try hard on the same node. Use the | |
1904 | * node_alloc_noretry bitmap to manage this state information. | |
1905 | */ | |
1906 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1907 | alloc_try_hard = false; | |
1908 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1909 | if (alloc_try_hard) | |
1910 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1911 | if (nid == NUMA_NO_NODE) |
1912 | nid = numa_mem_id(); | |
84172f4b | 1913 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
af0fb9df MH |
1914 | if (page) |
1915 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1916 | else | |
1917 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1918 | |
f60858f9 MK |
1919 | /* |
1920 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1921 | * indicates an overall state change. Clear bit so that we resume | |
1922 | * normal 'try hard' allocations. | |
1923 | */ | |
1924 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1925 | node_clear(nid, *node_alloc_noretry); | |
1926 | ||
1927 | /* | |
1928 | * If we tried hard to get a page but failed, set bit so that | |
1929 | * subsequent attempts will not try as hard until there is an | |
1930 | * overall state change. | |
1931 | */ | |
1932 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1933 | node_set(nid, *node_alloc_noretry); | |
1934 | ||
63b4613c NA |
1935 | return page; |
1936 | } | |
1937 | ||
0c397dae MH |
1938 | /* |
1939 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1940 | * should use this function to get new hugetlb pages | |
1941 | */ | |
1942 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1943 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1944 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1945 | { |
1946 | struct page *page; | |
7118fc29 | 1947 | bool retry = false; |
0c397dae | 1948 | |
7118fc29 | 1949 | retry: |
0c397dae MH |
1950 | if (hstate_is_gigantic(h)) |
1951 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1952 | else | |
1953 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1954 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1955 | if (!page) |
1956 | return NULL; | |
1957 | ||
7118fc29 MK |
1958 | if (hstate_is_gigantic(h)) { |
1959 | if (!prep_compound_gigantic_page(page, huge_page_order(h))) { | |
1960 | /* | |
1961 | * Rare failure to convert pages to compound page. | |
1962 | * Free pages and try again - ONCE! | |
1963 | */ | |
1964 | free_gigantic_page(page, huge_page_order(h)); | |
1965 | if (!retry) { | |
1966 | retry = true; | |
1967 | goto retry; | |
1968 | } | |
7118fc29 MK |
1969 | return NULL; |
1970 | } | |
1971 | } | |
0c397dae MH |
1972 | prep_new_huge_page(h, page, page_to_nid(page)); |
1973 | ||
1974 | return page; | |
1975 | } | |
1976 | ||
af0fb9df MH |
1977 | /* |
1978 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1979 | * manner. | |
1980 | */ | |
f60858f9 MK |
1981 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1982 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1983 | { |
1984 | struct page *page; | |
1985 | int nr_nodes, node; | |
af0fb9df | 1986 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1987 | |
1988 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1989 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1990 | node_alloc_noretry); | |
af0fb9df | 1991 | if (page) |
b2261026 | 1992 | break; |
b2261026 JK |
1993 | } |
1994 | ||
af0fb9df MH |
1995 | if (!page) |
1996 | return 0; | |
b2261026 | 1997 | |
af0fb9df MH |
1998 | put_page(page); /* free it into the hugepage allocator */ |
1999 | ||
2000 | return 1; | |
b2261026 JK |
2001 | } |
2002 | ||
e8c5c824 | 2003 | /* |
10c6ec49 MK |
2004 | * Remove huge page from pool from next node to free. Attempt to keep |
2005 | * persistent huge pages more or less balanced over allowed nodes. | |
2006 | * This routine only 'removes' the hugetlb page. The caller must make | |
2007 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
2008 | * Called with hugetlb_lock locked. |
2009 | */ | |
10c6ec49 MK |
2010 | static struct page *remove_pool_huge_page(struct hstate *h, |
2011 | nodemask_t *nodes_allowed, | |
2012 | bool acct_surplus) | |
e8c5c824 | 2013 | { |
b2261026 | 2014 | int nr_nodes, node; |
10c6ec49 | 2015 | struct page *page = NULL; |
e8c5c824 | 2016 | |
9487ca60 | 2017 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 2018 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
2019 | /* |
2020 | * If we're returning unused surplus pages, only examine | |
2021 | * nodes with surplus pages. | |
2022 | */ | |
b2261026 JK |
2023 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
2024 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 2025 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 2026 | struct page, lru); |
6eb4e88a | 2027 | remove_hugetlb_page(h, page, acct_surplus); |
9a76db09 | 2028 | break; |
e8c5c824 | 2029 | } |
b2261026 | 2030 | } |
e8c5c824 | 2031 | |
10c6ec49 | 2032 | return page; |
e8c5c824 LS |
2033 | } |
2034 | ||
c8721bbb NH |
2035 | /* |
2036 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
2037 | * nothing for in-use hugepages and non-hugepages. |
2038 | * This function returns values like below: | |
2039 | * | |
ad2fa371 MS |
2040 | * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages |
2041 | * when the system is under memory pressure and the feature of | |
2042 | * freeing unused vmemmap pages associated with each hugetlb page | |
2043 | * is enabled. | |
2044 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
2045 | * (allocated or reserved.) | |
2046 | * 0: successfully dissolved free hugepages or the page is not a | |
2047 | * hugepage (considered as already dissolved) | |
c8721bbb | 2048 | */ |
c3114a84 | 2049 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 2050 | { |
6bc9b564 | 2051 | int rc = -EBUSY; |
082d5b6b | 2052 | |
7ffddd49 | 2053 | retry: |
faf53def NH |
2054 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
2055 | if (!PageHuge(page)) | |
2056 | return 0; | |
2057 | ||
db71ef79 | 2058 | spin_lock_irq(&hugetlb_lock); |
faf53def NH |
2059 | if (!PageHuge(page)) { |
2060 | rc = 0; | |
2061 | goto out; | |
2062 | } | |
2063 | ||
2064 | if (!page_count(page)) { | |
2247bb33 GS |
2065 | struct page *head = compound_head(page); |
2066 | struct hstate *h = page_hstate(head); | |
6bc9b564 | 2067 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 2068 | goto out; |
7ffddd49 MS |
2069 | |
2070 | /* | |
2071 | * We should make sure that the page is already on the free list | |
2072 | * when it is dissolved. | |
2073 | */ | |
6c037149 | 2074 | if (unlikely(!HPageFreed(head))) { |
db71ef79 | 2075 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
2076 | cond_resched(); |
2077 | ||
2078 | /* | |
2079 | * Theoretically, we should return -EBUSY when we | |
2080 | * encounter this race. In fact, we have a chance | |
2081 | * to successfully dissolve the page if we do a | |
2082 | * retry. Because the race window is quite small. | |
2083 | * If we seize this opportunity, it is an optimization | |
2084 | * for increasing the success rate of dissolving page. | |
2085 | */ | |
2086 | goto retry; | |
2087 | } | |
2088 | ||
0c5da357 | 2089 | remove_hugetlb_page(h, head, false); |
c1470b33 | 2090 | h->max_huge_pages--; |
db71ef79 | 2091 | spin_unlock_irq(&hugetlb_lock); |
ad2fa371 MS |
2092 | |
2093 | /* | |
2094 | * Normally update_and_free_page will allocate required vmemmmap | |
2095 | * before freeing the page. update_and_free_page will fail to | |
2096 | * free the page if it can not allocate required vmemmap. We | |
2097 | * need to adjust max_huge_pages if the page is not freed. | |
2098 | * Attempt to allocate vmemmmap here so that we can take | |
2099 | * appropriate action on failure. | |
2100 | */ | |
2101 | rc = alloc_huge_page_vmemmap(h, head); | |
2102 | if (!rc) { | |
2103 | /* | |
2104 | * Move PageHWPoison flag from head page to the raw | |
2105 | * error page, which makes any subpages rather than | |
2106 | * the error page reusable. | |
2107 | */ | |
2108 | if (PageHWPoison(head) && page != head) { | |
2109 | SetPageHWPoison(page); | |
2110 | ClearPageHWPoison(head); | |
2111 | } | |
2112 | update_and_free_page(h, head, false); | |
2113 | } else { | |
2114 | spin_lock_irq(&hugetlb_lock); | |
2115 | add_hugetlb_page(h, head, false); | |
2116 | h->max_huge_pages++; | |
2117 | spin_unlock_irq(&hugetlb_lock); | |
2118 | } | |
2119 | ||
2120 | return rc; | |
c8721bbb | 2121 | } |
082d5b6b | 2122 | out: |
db71ef79 | 2123 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 2124 | return rc; |
c8721bbb NH |
2125 | } |
2126 | ||
2127 | /* | |
2128 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
2129 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
2130 | * Note that this will dissolve a free gigantic hugepage completely, if any |
2131 | * part of it lies within the given range. | |
082d5b6b GS |
2132 | * Also note that if dissolve_free_huge_page() returns with an error, all |
2133 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 2134 | */ |
082d5b6b | 2135 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 2136 | { |
c8721bbb | 2137 | unsigned long pfn; |
eb03aa00 | 2138 | struct page *page; |
082d5b6b | 2139 | int rc = 0; |
c8721bbb | 2140 | |
d0177639 | 2141 | if (!hugepages_supported()) |
082d5b6b | 2142 | return rc; |
d0177639 | 2143 | |
eb03aa00 GS |
2144 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
2145 | page = pfn_to_page(pfn); | |
faf53def NH |
2146 | rc = dissolve_free_huge_page(page); |
2147 | if (rc) | |
2148 | break; | |
eb03aa00 | 2149 | } |
082d5b6b GS |
2150 | |
2151 | return rc; | |
c8721bbb NH |
2152 | } |
2153 | ||
ab5ac90a MH |
2154 | /* |
2155 | * Allocates a fresh surplus page from the page allocator. | |
2156 | */ | |
0c397dae | 2157 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
b65a4eda | 2158 | int nid, nodemask_t *nmask, bool zero_ref) |
7893d1d5 | 2159 | { |
9980d744 | 2160 | struct page *page = NULL; |
b65a4eda | 2161 | bool retry = false; |
7893d1d5 | 2162 | |
bae7f4ae | 2163 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2164 | return NULL; |
2165 | ||
db71ef79 | 2166 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2167 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
2168 | goto out_unlock; | |
db71ef79 | 2169 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 2170 | |
b65a4eda | 2171 | retry: |
f60858f9 | 2172 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 2173 | if (!page) |
0c397dae | 2174 | return NULL; |
d1c3fb1f | 2175 | |
db71ef79 | 2176 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2177 | /* |
2178 | * We could have raced with the pool size change. | |
2179 | * Double check that and simply deallocate the new page | |
2180 | * if we would end up overcommiting the surpluses. Abuse | |
2181 | * temporary page to workaround the nasty free_huge_page | |
2182 | * codeflow | |
2183 | */ | |
2184 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 2185 | SetHPageTemporary(page); |
db71ef79 | 2186 | spin_unlock_irq(&hugetlb_lock); |
9980d744 | 2187 | put_page(page); |
2bf753e6 | 2188 | return NULL; |
7893d1d5 | 2189 | } |
9980d744 | 2190 | |
b65a4eda MK |
2191 | if (zero_ref) { |
2192 | /* | |
2193 | * Caller requires a page with zero ref count. | |
2194 | * We will drop ref count here. If someone else is holding | |
2195 | * a ref, the page will be freed when they drop it. Abuse | |
2196 | * temporary page flag to accomplish this. | |
2197 | */ | |
2198 | SetHPageTemporary(page); | |
2199 | if (!put_page_testzero(page)) { | |
2200 | /* | |
2201 | * Unexpected inflated ref count on freshly allocated | |
2202 | * huge. Retry once. | |
2203 | */ | |
2204 | pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); | |
2205 | spin_unlock_irq(&hugetlb_lock); | |
2206 | if (retry) | |
2207 | return NULL; | |
2208 | ||
2209 | retry = true; | |
2210 | goto retry; | |
2211 | } | |
2212 | ClearHPageTemporary(page); | |
2213 | } | |
2214 | ||
2215 | h->surplus_huge_pages++; | |
2216 | h->surplus_huge_pages_node[page_to_nid(page)]++; | |
2217 | ||
9980d744 | 2218 | out_unlock: |
db71ef79 | 2219 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 AL |
2220 | |
2221 | return page; | |
2222 | } | |
2223 | ||
bbe88753 | 2224 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 2225 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
2226 | { |
2227 | struct page *page; | |
2228 | ||
2229 | if (hstate_is_gigantic(h)) | |
2230 | return NULL; | |
2231 | ||
f60858f9 | 2232 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
2233 | if (!page) |
2234 | return NULL; | |
2235 | ||
2236 | /* | |
2237 | * We do not account these pages as surplus because they are only | |
2238 | * temporary and will be released properly on the last reference | |
2239 | */ | |
9157c311 | 2240 | SetHPageTemporary(page); |
ab5ac90a MH |
2241 | |
2242 | return page; | |
2243 | } | |
2244 | ||
099730d6 DH |
2245 | /* |
2246 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
2247 | */ | |
e0ec90ee | 2248 | static |
0c397dae | 2249 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
2250 | struct vm_area_struct *vma, unsigned long addr) |
2251 | { | |
cfcaa66f | 2252 | struct page *page = NULL; |
aaf14e40 MH |
2253 | struct mempolicy *mpol; |
2254 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
2255 | int nid; | |
2256 | nodemask_t *nodemask; | |
2257 | ||
2258 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
2259 | if (mpol_is_preferred_many(mpol)) { |
2260 | gfp_t gfp = gfp_mask | __GFP_NOWARN; | |
2261 | ||
2262 | gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); | |
2263 | page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); | |
aaf14e40 | 2264 | |
cfcaa66f BW |
2265 | /* Fallback to all nodes if page==NULL */ |
2266 | nodemask = NULL; | |
2267 | } | |
2268 | ||
2269 | if (!page) | |
2270 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); | |
2271 | mpol_cond_put(mpol); | |
aaf14e40 | 2272 | return page; |
099730d6 DH |
2273 | } |
2274 | ||
ab5ac90a | 2275 | /* page migration callback function */ |
3e59fcb0 | 2276 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 2277 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 2278 | { |
db71ef79 | 2279 | spin_lock_irq(&hugetlb_lock); |
4db9b2ef | 2280 | if (h->free_huge_pages - h->resv_huge_pages > 0) { |
3e59fcb0 MH |
2281 | struct page *page; |
2282 | ||
2283 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
2284 | if (page) { | |
db71ef79 | 2285 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 2286 | return page; |
4db9b2ef MH |
2287 | } |
2288 | } | |
db71ef79 | 2289 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 2290 | |
0c397dae | 2291 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2292 | } |
2293 | ||
ebd63723 | 2294 | /* mempolicy aware migration callback */ |
389c8178 MH |
2295 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2296 | unsigned long address) | |
ebd63723 MH |
2297 | { |
2298 | struct mempolicy *mpol; | |
2299 | nodemask_t *nodemask; | |
2300 | struct page *page; | |
ebd63723 MH |
2301 | gfp_t gfp_mask; |
2302 | int node; | |
2303 | ||
ebd63723 MH |
2304 | gfp_mask = htlb_alloc_mask(h); |
2305 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 2306 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
2307 | mpol_cond_put(mpol); |
2308 | ||
2309 | return page; | |
2310 | } | |
2311 | ||
e4e574b7 | 2312 | /* |
25985edc | 2313 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2314 | * of size 'delta'. |
2315 | */ | |
0a4f3d1b | 2316 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 2317 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
2318 | { |
2319 | struct list_head surplus_list; | |
2320 | struct page *page, *tmp; | |
0a4f3d1b LX |
2321 | int ret; |
2322 | long i; | |
2323 | long needed, allocated; | |
28073b02 | 2324 | bool alloc_ok = true; |
e4e574b7 | 2325 | |
9487ca60 | 2326 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 2327 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2328 | if (needed <= 0) { |
a5516438 | 2329 | h->resv_huge_pages += delta; |
e4e574b7 | 2330 | return 0; |
ac09b3a1 | 2331 | } |
e4e574b7 AL |
2332 | |
2333 | allocated = 0; | |
2334 | INIT_LIST_HEAD(&surplus_list); | |
2335 | ||
2336 | ret = -ENOMEM; | |
2337 | retry: | |
db71ef79 | 2338 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 2339 | for (i = 0; i < needed; i++) { |
0c397dae | 2340 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
b65a4eda | 2341 | NUMA_NO_NODE, NULL, true); |
28073b02 HD |
2342 | if (!page) { |
2343 | alloc_ok = false; | |
2344 | break; | |
2345 | } | |
e4e574b7 | 2346 | list_add(&page->lru, &surplus_list); |
69ed779a | 2347 | cond_resched(); |
e4e574b7 | 2348 | } |
28073b02 | 2349 | allocated += i; |
e4e574b7 AL |
2350 | |
2351 | /* | |
2352 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2353 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2354 | */ | |
db71ef79 | 2355 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
2356 | needed = (h->resv_huge_pages + delta) - |
2357 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2358 | if (needed > 0) { |
2359 | if (alloc_ok) | |
2360 | goto retry; | |
2361 | /* | |
2362 | * We were not able to allocate enough pages to | |
2363 | * satisfy the entire reservation so we free what | |
2364 | * we've allocated so far. | |
2365 | */ | |
2366 | goto free; | |
2367 | } | |
e4e574b7 AL |
2368 | /* |
2369 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2370 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2371 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2372 | * allocator. Commit the entire reservation here to prevent another |
2373 | * process from stealing the pages as they are added to the pool but | |
2374 | * before they are reserved. | |
e4e574b7 AL |
2375 | */ |
2376 | needed += allocated; | |
a5516438 | 2377 | h->resv_huge_pages += delta; |
e4e574b7 | 2378 | ret = 0; |
a9869b83 | 2379 | |
19fc3f0a | 2380 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2381 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2382 | if ((--needed) < 0) |
2383 | break; | |
b65a4eda | 2384 | /* Add the page to the hugetlb allocator */ |
a5516438 | 2385 | enqueue_huge_page(h, page); |
19fc3f0a | 2386 | } |
28073b02 | 2387 | free: |
db71ef79 | 2388 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a | 2389 | |
b65a4eda MK |
2390 | /* |
2391 | * Free unnecessary surplus pages to the buddy allocator. | |
2392 | * Pages have no ref count, call free_huge_page directly. | |
2393 | */ | |
c0d934ba | 2394 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
b65a4eda | 2395 | free_huge_page(page); |
db71ef79 | 2396 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2397 | |
2398 | return ret; | |
2399 | } | |
2400 | ||
2401 | /* | |
e5bbc8a6 MK |
2402 | * This routine has two main purposes: |
2403 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2404 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2405 | * to the associated reservation map. | |
2406 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2407 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2408 | */ |
a5516438 AK |
2409 | static void return_unused_surplus_pages(struct hstate *h, |
2410 | unsigned long unused_resv_pages) | |
e4e574b7 | 2411 | { |
e4e574b7 | 2412 | unsigned long nr_pages; |
10c6ec49 MK |
2413 | struct page *page; |
2414 | LIST_HEAD(page_list); | |
2415 | ||
9487ca60 | 2416 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2417 | /* Uncommit the reservation */ |
2418 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2419 | |
aa888a74 | 2420 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2421 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2422 | goto out; |
aa888a74 | 2423 | |
e5bbc8a6 MK |
2424 | /* |
2425 | * Part (or even all) of the reservation could have been backed | |
2426 | * by pre-allocated pages. Only free surplus pages. | |
2427 | */ | |
a5516438 | 2428 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2429 | |
685f3457 LS |
2430 | /* |
2431 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2432 | * evenly across all nodes with memory. Iterate across these nodes |
2433 | * until we can no longer free unreserved surplus pages. This occurs | |
2434 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2435 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2436 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2437 | */ |
2438 | while (nr_pages--) { | |
10c6ec49 MK |
2439 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2440 | if (!page) | |
e5bbc8a6 | 2441 | goto out; |
10c6ec49 MK |
2442 | |
2443 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2444 | } |
e5bbc8a6 MK |
2445 | |
2446 | out: | |
db71ef79 | 2447 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2448 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2449 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2450 | } |
2451 | ||
5e911373 | 2452 | |
c37f9fb1 | 2453 | /* |
feba16e2 | 2454 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2455 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2456 | * |
2457 | * vma_needs_reservation is called to determine if the huge page at addr | |
2458 | * within the vma has an associated reservation. If a reservation is | |
2459 | * needed, the value 1 is returned. The caller is then responsible for | |
2460 | * managing the global reservation and subpool usage counts. After | |
2461 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2462 | * to add the page to the reservation map. If the page allocation fails, |
2463 | * the reservation must be ended instead of committed. vma_end_reservation | |
2464 | * is called in such cases. | |
cf3ad20b MK |
2465 | * |
2466 | * In the normal case, vma_commit_reservation returns the same value | |
2467 | * as the preceding vma_needs_reservation call. The only time this | |
2468 | * is not the case is if a reserve map was changed between calls. It | |
2469 | * is the responsibility of the caller to notice the difference and | |
2470 | * take appropriate action. | |
96b96a96 MK |
2471 | * |
2472 | * vma_add_reservation is used in error paths where a reservation must | |
2473 | * be restored when a newly allocated huge page must be freed. It is | |
2474 | * to be called after calling vma_needs_reservation to determine if a | |
2475 | * reservation exists. | |
846be085 MK |
2476 | * |
2477 | * vma_del_reservation is used in error paths where an entry in the reserve | |
2478 | * map was created during huge page allocation and must be removed. It is to | |
2479 | * be called after calling vma_needs_reservation to determine if a reservation | |
2480 | * exists. | |
c37f9fb1 | 2481 | */ |
5e911373 MK |
2482 | enum vma_resv_mode { |
2483 | VMA_NEEDS_RESV, | |
2484 | VMA_COMMIT_RESV, | |
feba16e2 | 2485 | VMA_END_RESV, |
96b96a96 | 2486 | VMA_ADD_RESV, |
846be085 | 2487 | VMA_DEL_RESV, |
5e911373 | 2488 | }; |
cf3ad20b MK |
2489 | static long __vma_reservation_common(struct hstate *h, |
2490 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2491 | enum vma_resv_mode mode) |
c37f9fb1 | 2492 | { |
4e35f483 JK |
2493 | struct resv_map *resv; |
2494 | pgoff_t idx; | |
cf3ad20b | 2495 | long ret; |
0db9d74e | 2496 | long dummy_out_regions_needed; |
c37f9fb1 | 2497 | |
4e35f483 JK |
2498 | resv = vma_resv_map(vma); |
2499 | if (!resv) | |
84afd99b | 2500 | return 1; |
c37f9fb1 | 2501 | |
4e35f483 | 2502 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2503 | switch (mode) { |
2504 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2505 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2506 | /* We assume that vma_reservation_* routines always operate on | |
2507 | * 1 page, and that adding to resv map a 1 page entry can only | |
2508 | * ever require 1 region. | |
2509 | */ | |
2510 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2511 | break; |
2512 | case VMA_COMMIT_RESV: | |
075a61d0 | 2513 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2514 | /* region_add calls of range 1 should never fail. */ |
2515 | VM_BUG_ON(ret < 0); | |
5e911373 | 2516 | break; |
feba16e2 | 2517 | case VMA_END_RESV: |
0db9d74e | 2518 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2519 | ret = 0; |
2520 | break; | |
96b96a96 | 2521 | case VMA_ADD_RESV: |
0db9d74e | 2522 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2523 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2524 | /* region_add calls of range 1 should never fail. */ |
2525 | VM_BUG_ON(ret < 0); | |
2526 | } else { | |
2527 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2528 | ret = region_del(resv, idx, idx + 1); |
2529 | } | |
2530 | break; | |
846be085 MK |
2531 | case VMA_DEL_RESV: |
2532 | if (vma->vm_flags & VM_MAYSHARE) { | |
2533 | region_abort(resv, idx, idx + 1, 1); | |
2534 | ret = region_del(resv, idx, idx + 1); | |
2535 | } else { | |
2536 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2537 | /* region_add calls of range 1 should never fail. */ | |
2538 | VM_BUG_ON(ret < 0); | |
2539 | } | |
2540 | break; | |
5e911373 MK |
2541 | default: |
2542 | BUG(); | |
2543 | } | |
84afd99b | 2544 | |
846be085 | 2545 | if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) |
cf3ad20b | 2546 | return ret; |
bf3d12b9 ML |
2547 | /* |
2548 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2549 | * | |
2550 | * In most cases, reserves always exist for private mappings. | |
2551 | * However, a file associated with mapping could have been | |
2552 | * hole punched or truncated after reserves were consumed. | |
2553 | * As subsequent fault on such a range will not use reserves. | |
2554 | * Subtle - The reserve map for private mappings has the | |
2555 | * opposite meaning than that of shared mappings. If NO | |
2556 | * entry is in the reserve map, it means a reservation exists. | |
2557 | * If an entry exists in the reserve map, it means the | |
2558 | * reservation has already been consumed. As a result, the | |
2559 | * return value of this routine is the opposite of the | |
2560 | * value returned from reserve map manipulation routines above. | |
2561 | */ | |
2562 | if (ret > 0) | |
2563 | return 0; | |
2564 | if (ret == 0) | |
2565 | return 1; | |
2566 | return ret; | |
c37f9fb1 | 2567 | } |
cf3ad20b MK |
2568 | |
2569 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2570 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2571 | { |
5e911373 | 2572 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2573 | } |
84afd99b | 2574 | |
cf3ad20b MK |
2575 | static long vma_commit_reservation(struct hstate *h, |
2576 | struct vm_area_struct *vma, unsigned long addr) | |
2577 | { | |
5e911373 MK |
2578 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2579 | } | |
2580 | ||
feba16e2 | 2581 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2582 | struct vm_area_struct *vma, unsigned long addr) |
2583 | { | |
feba16e2 | 2584 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2585 | } |
2586 | ||
96b96a96 MK |
2587 | static long vma_add_reservation(struct hstate *h, |
2588 | struct vm_area_struct *vma, unsigned long addr) | |
2589 | { | |
2590 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2591 | } | |
2592 | ||
846be085 MK |
2593 | static long vma_del_reservation(struct hstate *h, |
2594 | struct vm_area_struct *vma, unsigned long addr) | |
2595 | { | |
2596 | return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); | |
2597 | } | |
2598 | ||
96b96a96 | 2599 | /* |
846be085 MK |
2600 | * This routine is called to restore reservation information on error paths. |
2601 | * It should ONLY be called for pages allocated via alloc_huge_page(), and | |
2602 | * the hugetlb mutex should remain held when calling this routine. | |
2603 | * | |
2604 | * It handles two specific cases: | |
2605 | * 1) A reservation was in place and the page consumed the reservation. | |
2606 | * HPageRestoreReserve is set in the page. | |
2607 | * 2) No reservation was in place for the page, so HPageRestoreReserve is | |
2608 | * not set. However, alloc_huge_page always updates the reserve map. | |
2609 | * | |
2610 | * In case 1, free_huge_page later in the error path will increment the | |
2611 | * global reserve count. But, free_huge_page does not have enough context | |
2612 | * to adjust the reservation map. This case deals primarily with private | |
2613 | * mappings. Adjust the reserve map here to be consistent with global | |
2614 | * reserve count adjustments to be made by free_huge_page. Make sure the | |
2615 | * reserve map indicates there is a reservation present. | |
2616 | * | |
2617 | * In case 2, simply undo reserve map modifications done by alloc_huge_page. | |
96b96a96 | 2618 | */ |
846be085 MK |
2619 | void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, |
2620 | unsigned long address, struct page *page) | |
96b96a96 | 2621 | { |
846be085 | 2622 | long rc = vma_needs_reservation(h, vma, address); |
96b96a96 | 2623 | |
846be085 MK |
2624 | if (HPageRestoreReserve(page)) { |
2625 | if (unlikely(rc < 0)) | |
96b96a96 MK |
2626 | /* |
2627 | * Rare out of memory condition in reserve map | |
d6995da3 | 2628 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2629 | * global reserve count will not be incremented |
2630 | * by free_huge_page. This will make it appear | |
2631 | * as though the reservation for this page was | |
2632 | * consumed. This may prevent the task from | |
2633 | * faulting in the page at a later time. This | |
2634 | * is better than inconsistent global huge page | |
2635 | * accounting of reserve counts. | |
2636 | */ | |
d6995da3 | 2637 | ClearHPageRestoreReserve(page); |
846be085 MK |
2638 | else if (rc) |
2639 | (void)vma_add_reservation(h, vma, address); | |
2640 | else | |
2641 | vma_end_reservation(h, vma, address); | |
2642 | } else { | |
2643 | if (!rc) { | |
2644 | /* | |
2645 | * This indicates there is an entry in the reserve map | |
c7b1850d | 2646 | * not added by alloc_huge_page. We know it was added |
846be085 MK |
2647 | * before the alloc_huge_page call, otherwise |
2648 | * HPageRestoreReserve would be set on the page. | |
2649 | * Remove the entry so that a subsequent allocation | |
2650 | * does not consume a reservation. | |
2651 | */ | |
2652 | rc = vma_del_reservation(h, vma, address); | |
2653 | if (rc < 0) | |
96b96a96 | 2654 | /* |
846be085 MK |
2655 | * VERY rare out of memory condition. Since |
2656 | * we can not delete the entry, set | |
2657 | * HPageRestoreReserve so that the reserve | |
2658 | * count will be incremented when the page | |
2659 | * is freed. This reserve will be consumed | |
2660 | * on a subsequent allocation. | |
96b96a96 | 2661 | */ |
846be085 MK |
2662 | SetHPageRestoreReserve(page); |
2663 | } else if (rc < 0) { | |
2664 | /* | |
2665 | * Rare out of memory condition from | |
2666 | * vma_needs_reservation call. Memory allocation is | |
2667 | * only attempted if a new entry is needed. Therefore, | |
2668 | * this implies there is not an entry in the | |
2669 | * reserve map. | |
2670 | * | |
2671 | * For shared mappings, no entry in the map indicates | |
2672 | * no reservation. We are done. | |
2673 | */ | |
2674 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
2675 | /* | |
2676 | * For private mappings, no entry indicates | |
2677 | * a reservation is present. Since we can | |
2678 | * not add an entry, set SetHPageRestoreReserve | |
2679 | * on the page so reserve count will be | |
2680 | * incremented when freed. This reserve will | |
2681 | * be consumed on a subsequent allocation. | |
2682 | */ | |
2683 | SetHPageRestoreReserve(page); | |
96b96a96 | 2684 | } else |
846be085 MK |
2685 | /* |
2686 | * No reservation present, do nothing | |
2687 | */ | |
2688 | vma_end_reservation(h, vma, address); | |
96b96a96 MK |
2689 | } |
2690 | } | |
2691 | ||
369fa227 OS |
2692 | /* |
2693 | * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one | |
2694 | * @h: struct hstate old page belongs to | |
2695 | * @old_page: Old page to dissolve | |
ae37c7ff | 2696 | * @list: List to isolate the page in case we need to |
369fa227 OS |
2697 | * Returns 0 on success, otherwise negated error. |
2698 | */ | |
ae37c7ff OS |
2699 | static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, |
2700 | struct list_head *list) | |
369fa227 OS |
2701 | { |
2702 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
2703 | int nid = page_to_nid(old_page); | |
b65a4eda | 2704 | bool alloc_retry = false; |
369fa227 OS |
2705 | struct page *new_page; |
2706 | int ret = 0; | |
2707 | ||
2708 | /* | |
2709 | * Before dissolving the page, we need to allocate a new one for the | |
f41f2ed4 MS |
2710 | * pool to remain stable. Here, we allocate the page and 'prep' it |
2711 | * by doing everything but actually updating counters and adding to | |
2712 | * the pool. This simplifies and let us do most of the processing | |
2713 | * under the lock. | |
369fa227 | 2714 | */ |
b65a4eda | 2715 | alloc_retry: |
369fa227 OS |
2716 | new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); |
2717 | if (!new_page) | |
2718 | return -ENOMEM; | |
b65a4eda MK |
2719 | /* |
2720 | * If all goes well, this page will be directly added to the free | |
2721 | * list in the pool. For this the ref count needs to be zero. | |
2722 | * Attempt to drop now, and retry once if needed. It is VERY | |
2723 | * unlikely there is another ref on the page. | |
2724 | * | |
2725 | * If someone else has a reference to the page, it will be freed | |
2726 | * when they drop their ref. Abuse temporary page flag to accomplish | |
2727 | * this. Retry once if there is an inflated ref count. | |
2728 | */ | |
2729 | SetHPageTemporary(new_page); | |
2730 | if (!put_page_testzero(new_page)) { | |
2731 | if (alloc_retry) | |
2732 | return -EBUSY; | |
2733 | ||
2734 | alloc_retry = true; | |
2735 | goto alloc_retry; | |
2736 | } | |
2737 | ClearHPageTemporary(new_page); | |
2738 | ||
f41f2ed4 | 2739 | __prep_new_huge_page(h, new_page); |
369fa227 OS |
2740 | |
2741 | retry: | |
2742 | spin_lock_irq(&hugetlb_lock); | |
2743 | if (!PageHuge(old_page)) { | |
2744 | /* | |
2745 | * Freed from under us. Drop new_page too. | |
2746 | */ | |
2747 | goto free_new; | |
2748 | } else if (page_count(old_page)) { | |
2749 | /* | |
ae37c7ff OS |
2750 | * Someone has grabbed the page, try to isolate it here. |
2751 | * Fail with -EBUSY if not possible. | |
369fa227 | 2752 | */ |
ae37c7ff OS |
2753 | spin_unlock_irq(&hugetlb_lock); |
2754 | if (!isolate_huge_page(old_page, list)) | |
2755 | ret = -EBUSY; | |
2756 | spin_lock_irq(&hugetlb_lock); | |
369fa227 OS |
2757 | goto free_new; |
2758 | } else if (!HPageFreed(old_page)) { | |
2759 | /* | |
2760 | * Page's refcount is 0 but it has not been enqueued in the | |
2761 | * freelist yet. Race window is small, so we can succeed here if | |
2762 | * we retry. | |
2763 | */ | |
2764 | spin_unlock_irq(&hugetlb_lock); | |
2765 | cond_resched(); | |
2766 | goto retry; | |
2767 | } else { | |
2768 | /* | |
2769 | * Ok, old_page is still a genuine free hugepage. Remove it from | |
2770 | * the freelist and decrease the counters. These will be | |
2771 | * incremented again when calling __prep_account_new_huge_page() | |
2772 | * and enqueue_huge_page() for new_page. The counters will remain | |
2773 | * stable since this happens under the lock. | |
2774 | */ | |
2775 | remove_hugetlb_page(h, old_page, false); | |
2776 | ||
2777 | /* | |
b65a4eda MK |
2778 | * Ref count on new page is already zero as it was dropped |
2779 | * earlier. It can be directly added to the pool free list. | |
369fa227 | 2780 | */ |
369fa227 | 2781 | __prep_account_new_huge_page(h, nid); |
369fa227 OS |
2782 | enqueue_huge_page(h, new_page); |
2783 | ||
2784 | /* | |
2785 | * Pages have been replaced, we can safely free the old one. | |
2786 | */ | |
2787 | spin_unlock_irq(&hugetlb_lock); | |
b65d4adb | 2788 | update_and_free_page(h, old_page, false); |
369fa227 OS |
2789 | } |
2790 | ||
2791 | return ret; | |
2792 | ||
2793 | free_new: | |
2794 | spin_unlock_irq(&hugetlb_lock); | |
b65a4eda MK |
2795 | /* Page has a zero ref count, but needs a ref to be freed */ |
2796 | set_page_refcounted(new_page); | |
b65d4adb | 2797 | update_and_free_page(h, new_page, false); |
369fa227 OS |
2798 | |
2799 | return ret; | |
2800 | } | |
2801 | ||
ae37c7ff | 2802 | int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) |
369fa227 OS |
2803 | { |
2804 | struct hstate *h; | |
2805 | struct page *head; | |
ae37c7ff | 2806 | int ret = -EBUSY; |
369fa227 OS |
2807 | |
2808 | /* | |
2809 | * The page might have been dissolved from under our feet, so make sure | |
2810 | * to carefully check the state under the lock. | |
2811 | * Return success when racing as if we dissolved the page ourselves. | |
2812 | */ | |
2813 | spin_lock_irq(&hugetlb_lock); | |
2814 | if (PageHuge(page)) { | |
2815 | head = compound_head(page); | |
2816 | h = page_hstate(head); | |
2817 | } else { | |
2818 | spin_unlock_irq(&hugetlb_lock); | |
2819 | return 0; | |
2820 | } | |
2821 | spin_unlock_irq(&hugetlb_lock); | |
2822 | ||
2823 | /* | |
2824 | * Fence off gigantic pages as there is a cyclic dependency between | |
2825 | * alloc_contig_range and them. Return -ENOMEM as this has the effect | |
2826 | * of bailing out right away without further retrying. | |
2827 | */ | |
2828 | if (hstate_is_gigantic(h)) | |
2829 | return -ENOMEM; | |
2830 | ||
ae37c7ff OS |
2831 | if (page_count(head) && isolate_huge_page(head, list)) |
2832 | ret = 0; | |
2833 | else if (!page_count(head)) | |
2834 | ret = alloc_and_dissolve_huge_page(h, head, list); | |
2835 | ||
2836 | return ret; | |
369fa227 OS |
2837 | } |
2838 | ||
70c3547e | 2839 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2840 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2841 | { |
90481622 | 2842 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2843 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2844 | struct page *page; |
d85f69b0 MK |
2845 | long map_chg, map_commit; |
2846 | long gbl_chg; | |
6d76dcf4 AK |
2847 | int ret, idx; |
2848 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2849 | bool deferred_reserve; |
a1e78772 | 2850 | |
6d76dcf4 | 2851 | idx = hstate_index(h); |
a1e78772 | 2852 | /* |
d85f69b0 MK |
2853 | * Examine the region/reserve map to determine if the process |
2854 | * has a reservation for the page to be allocated. A return | |
2855 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2856 | */ |
d85f69b0 MK |
2857 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2858 | if (map_chg < 0) | |
76dcee75 | 2859 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2860 | |
2861 | /* | |
2862 | * Processes that did not create the mapping will have no | |
2863 | * reserves as indicated by the region/reserve map. Check | |
2864 | * that the allocation will not exceed the subpool limit. | |
2865 | * Allocations for MAP_NORESERVE mappings also need to be | |
2866 | * checked against any subpool limit. | |
2867 | */ | |
2868 | if (map_chg || avoid_reserve) { | |
2869 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2870 | if (gbl_chg < 0) { | |
feba16e2 | 2871 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2872 | return ERR_PTR(-ENOSPC); |
5e911373 | 2873 | } |
1da177e4 | 2874 | |
d85f69b0 MK |
2875 | /* |
2876 | * Even though there was no reservation in the region/reserve | |
2877 | * map, there could be reservations associated with the | |
2878 | * subpool that can be used. This would be indicated if the | |
2879 | * return value of hugepage_subpool_get_pages() is zero. | |
2880 | * However, if avoid_reserve is specified we still avoid even | |
2881 | * the subpool reservations. | |
2882 | */ | |
2883 | if (avoid_reserve) | |
2884 | gbl_chg = 1; | |
2885 | } | |
2886 | ||
08cf9faf MA |
2887 | /* If this allocation is not consuming a reservation, charge it now. |
2888 | */ | |
6501fe5f | 2889 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2890 | if (deferred_reserve) { |
2891 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2892 | idx, pages_per_huge_page(h), &h_cg); | |
2893 | if (ret) | |
2894 | goto out_subpool_put; | |
2895 | } | |
2896 | ||
6d76dcf4 | 2897 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2898 | if (ret) |
08cf9faf | 2899 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2900 | |
db71ef79 | 2901 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
2902 | /* |
2903 | * glb_chg is passed to indicate whether or not a page must be taken | |
2904 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2905 | * a reservation exists for the allocation. | |
2906 | */ | |
2907 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2908 | if (!page) { |
db71ef79 | 2909 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 2910 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2911 | if (!page) |
2912 | goto out_uncharge_cgroup; | |
a88c7695 | 2913 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2914 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2915 | h->resv_huge_pages--; |
2916 | } | |
db71ef79 | 2917 | spin_lock_irq(&hugetlb_lock); |
15a8d68e | 2918 | list_add(&page->lru, &h->hugepage_activelist); |
81a6fcae | 2919 | /* Fall through */ |
68842c9b | 2920 | } |
81a6fcae | 2921 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2922 | /* If allocation is not consuming a reservation, also store the |
2923 | * hugetlb_cgroup pointer on the page. | |
2924 | */ | |
2925 | if (deferred_reserve) { | |
2926 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2927 | h_cg, page); | |
2928 | } | |
2929 | ||
db71ef79 | 2930 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 2931 | |
d6995da3 | 2932 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2933 | |
d85f69b0 MK |
2934 | map_commit = vma_commit_reservation(h, vma, addr); |
2935 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2936 | /* |
2937 | * The page was added to the reservation map between | |
2938 | * vma_needs_reservation and vma_commit_reservation. | |
2939 | * This indicates a race with hugetlb_reserve_pages. | |
2940 | * Adjust for the subpool count incremented above AND | |
2941 | * in hugetlb_reserve_pages for the same page. Also, | |
2942 | * the reservation count added in hugetlb_reserve_pages | |
2943 | * no longer applies. | |
2944 | */ | |
2945 | long rsv_adjust; | |
2946 | ||
2947 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2948 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2949 | if (deferred_reserve) |
2950 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2951 | pages_per_huge_page(h), page); | |
33039678 | 2952 | } |
90d8b7e6 | 2953 | return page; |
8f34af6f JZ |
2954 | |
2955 | out_uncharge_cgroup: | |
2956 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2957 | out_uncharge_cgroup_reservation: |
2958 | if (deferred_reserve) | |
2959 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2960 | h_cg); | |
8f34af6f | 2961 | out_subpool_put: |
d85f69b0 | 2962 | if (map_chg || avoid_reserve) |
8f34af6f | 2963 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2964 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2965 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2966 | } |
2967 | ||
e24a1307 AK |
2968 | int alloc_bootmem_huge_page(struct hstate *h) |
2969 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2970 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2971 | { |
2972 | struct huge_bootmem_page *m; | |
b2261026 | 2973 | int nr_nodes, node; |
aa888a74 | 2974 | |
b2261026 | 2975 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2976 | void *addr; |
2977 | ||
eb31d559 | 2978 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2979 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2980 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2981 | if (addr) { |
2982 | /* | |
2983 | * Use the beginning of the huge page to store the | |
2984 | * huge_bootmem_page struct (until gather_bootmem | |
2985 | * puts them into the mem_map). | |
2986 | */ | |
2987 | m = addr; | |
91f47662 | 2988 | goto found; |
aa888a74 | 2989 | } |
aa888a74 AK |
2990 | } |
2991 | return 0; | |
2992 | ||
2993 | found: | |
df994ead | 2994 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2995 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2996 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2997 | list_add(&m->list, &huge_boot_pages); |
2998 | m->hstate = h; | |
2999 | return 1; | |
3000 | } | |
3001 | ||
48b8d744 MK |
3002 | /* |
3003 | * Put bootmem huge pages into the standard lists after mem_map is up. | |
3004 | * Note: This only applies to gigantic (order > MAX_ORDER) pages. | |
3005 | */ | |
aa888a74 AK |
3006 | static void __init gather_bootmem_prealloc(void) |
3007 | { | |
3008 | struct huge_bootmem_page *m; | |
3009 | ||
3010 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 3011 | struct page *page = virt_to_page(m); |
aa888a74 | 3012 | struct hstate *h = m->hstate; |
ee8f248d | 3013 | |
48b8d744 | 3014 | VM_BUG_ON(!hstate_is_gigantic(h)); |
aa888a74 | 3015 | WARN_ON(page_count(page) != 1); |
7118fc29 MK |
3016 | if (prep_compound_gigantic_page(page, huge_page_order(h))) { |
3017 | WARN_ON(PageReserved(page)); | |
3018 | prep_new_huge_page(h, page, page_to_nid(page)); | |
3019 | put_page(page); /* add to the hugepage allocator */ | |
3020 | } else { | |
416d85ed | 3021 | /* VERY unlikely inflated ref count on a tail page */ |
7118fc29 | 3022 | free_gigantic_page(page, huge_page_order(h)); |
7118fc29 | 3023 | } |
af0fb9df | 3024 | |
b0320c7b | 3025 | /* |
48b8d744 MK |
3026 | * We need to restore the 'stolen' pages to totalram_pages |
3027 | * in order to fix confusing memory reports from free(1) and | |
3028 | * other side-effects, like CommitLimit going negative. | |
b0320c7b | 3029 | */ |
48b8d744 | 3030 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 3031 | cond_resched(); |
aa888a74 AK |
3032 | } |
3033 | } | |
3034 | ||
8faa8b07 | 3035 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
3036 | { |
3037 | unsigned long i; | |
f60858f9 MK |
3038 | nodemask_t *node_alloc_noretry; |
3039 | ||
3040 | if (!hstate_is_gigantic(h)) { | |
3041 | /* | |
3042 | * Bit mask controlling how hard we retry per-node allocations. | |
3043 | * Ignore errors as lower level routines can deal with | |
3044 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
3045 | * time, we are likely in bigger trouble. | |
3046 | */ | |
3047 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
3048 | GFP_KERNEL); | |
3049 | } else { | |
3050 | /* allocations done at boot time */ | |
3051 | node_alloc_noretry = NULL; | |
3052 | } | |
3053 | ||
3054 | /* bit mask controlling how hard we retry per-node allocations */ | |
3055 | if (node_alloc_noretry) | |
3056 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 3057 | |
e5ff2159 | 3058 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 3059 | if (hstate_is_gigantic(h)) { |
dbda8fea | 3060 | if (hugetlb_cma_size) { |
cf11e85f | 3061 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); |
7ecc9565 | 3062 | goto free; |
cf11e85f | 3063 | } |
aa888a74 AK |
3064 | if (!alloc_bootmem_huge_page(h)) |
3065 | break; | |
0c397dae | 3066 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
3067 | &node_states[N_MEMORY], |
3068 | node_alloc_noretry)) | |
1da177e4 | 3069 | break; |
69ed779a | 3070 | cond_resched(); |
1da177e4 | 3071 | } |
d715cf80 LH |
3072 | if (i < h->max_huge_pages) { |
3073 | char buf[32]; | |
3074 | ||
c6247f72 | 3075 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
3076 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
3077 | h->max_huge_pages, buf, i); | |
3078 | h->max_huge_pages = i; | |
3079 | } | |
7ecc9565 | 3080 | free: |
f60858f9 | 3081 | kfree(node_alloc_noretry); |
e5ff2159 AK |
3082 | } |
3083 | ||
3084 | static void __init hugetlb_init_hstates(void) | |
3085 | { | |
79dfc695 | 3086 | struct hstate *h, *h2; |
e5ff2159 AK |
3087 | |
3088 | for_each_hstate(h) { | |
641844f5 NH |
3089 | if (minimum_order > huge_page_order(h)) |
3090 | minimum_order = huge_page_order(h); | |
3091 | ||
8faa8b07 | 3092 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 3093 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 3094 | hugetlb_hstate_alloc_pages(h); |
79dfc695 MK |
3095 | |
3096 | /* | |
3097 | * Set demote order for each hstate. Note that | |
3098 | * h->demote_order is initially 0. | |
3099 | * - We can not demote gigantic pages if runtime freeing | |
3100 | * is not supported, so skip this. | |
a01f4390 MK |
3101 | * - If CMA allocation is possible, we can not demote |
3102 | * HUGETLB_PAGE_ORDER or smaller size pages. | |
79dfc695 MK |
3103 | */ |
3104 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
3105 | continue; | |
a01f4390 MK |
3106 | if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) |
3107 | continue; | |
79dfc695 MK |
3108 | for_each_hstate(h2) { |
3109 | if (h2 == h) | |
3110 | continue; | |
3111 | if (h2->order < h->order && | |
3112 | h2->order > h->demote_order) | |
3113 | h->demote_order = h2->order; | |
3114 | } | |
e5ff2159 | 3115 | } |
641844f5 | 3116 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
3117 | } |
3118 | ||
3119 | static void __init report_hugepages(void) | |
3120 | { | |
3121 | struct hstate *h; | |
3122 | ||
3123 | for_each_hstate(h) { | |
4abd32db | 3124 | char buf[32]; |
c6247f72 MW |
3125 | |
3126 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 3127 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 3128 | buf, h->free_huge_pages); |
e5ff2159 AK |
3129 | } |
3130 | } | |
3131 | ||
1da177e4 | 3132 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
3133 | static void try_to_free_low(struct hstate *h, unsigned long count, |
3134 | nodemask_t *nodes_allowed) | |
1da177e4 | 3135 | { |
4415cc8d | 3136 | int i; |
1121828a | 3137 | LIST_HEAD(page_list); |
4415cc8d | 3138 | |
9487ca60 | 3139 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 3140 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
3141 | return; |
3142 | ||
1121828a MK |
3143 | /* |
3144 | * Collect pages to be freed on a list, and free after dropping lock | |
3145 | */ | |
6ae11b27 | 3146 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 3147 | struct page *page, *next; |
a5516438 AK |
3148 | struct list_head *freel = &h->hugepage_freelists[i]; |
3149 | list_for_each_entry_safe(page, next, freel, lru) { | |
3150 | if (count >= h->nr_huge_pages) | |
1121828a | 3151 | goto out; |
1da177e4 LT |
3152 | if (PageHighMem(page)) |
3153 | continue; | |
6eb4e88a | 3154 | remove_hugetlb_page(h, page, false); |
1121828a | 3155 | list_add(&page->lru, &page_list); |
1da177e4 LT |
3156 | } |
3157 | } | |
1121828a MK |
3158 | |
3159 | out: | |
db71ef79 | 3160 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3161 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 3162 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
3163 | } |
3164 | #else | |
6ae11b27 LS |
3165 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
3166 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
3167 | { |
3168 | } | |
3169 | #endif | |
3170 | ||
20a0307c WF |
3171 | /* |
3172 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
3173 | * balanced by operating on them in a round-robin fashion. | |
3174 | * Returns 1 if an adjustment was made. | |
3175 | */ | |
6ae11b27 LS |
3176 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
3177 | int delta) | |
20a0307c | 3178 | { |
b2261026 | 3179 | int nr_nodes, node; |
20a0307c | 3180 | |
9487ca60 | 3181 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 3182 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 3183 | |
b2261026 JK |
3184 | if (delta < 0) { |
3185 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
3186 | if (h->surplus_huge_pages_node[node]) | |
3187 | goto found; | |
e8c5c824 | 3188 | } |
b2261026 JK |
3189 | } else { |
3190 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
3191 | if (h->surplus_huge_pages_node[node] < | |
3192 | h->nr_huge_pages_node[node]) | |
3193 | goto found; | |
e8c5c824 | 3194 | } |
b2261026 JK |
3195 | } |
3196 | return 0; | |
20a0307c | 3197 | |
b2261026 JK |
3198 | found: |
3199 | h->surplus_huge_pages += delta; | |
3200 | h->surplus_huge_pages_node[node] += delta; | |
3201 | return 1; | |
20a0307c WF |
3202 | } |
3203 | ||
a5516438 | 3204 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 3205 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 3206 | nodemask_t *nodes_allowed) |
1da177e4 | 3207 | { |
7893d1d5 | 3208 | unsigned long min_count, ret; |
10c6ec49 MK |
3209 | struct page *page; |
3210 | LIST_HEAD(page_list); | |
f60858f9 MK |
3211 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
3212 | ||
3213 | /* | |
3214 | * Bit mask controlling how hard we retry per-node allocations. | |
3215 | * If we can not allocate the bit mask, do not attempt to allocate | |
3216 | * the requested huge pages. | |
3217 | */ | |
3218 | if (node_alloc_noretry) | |
3219 | nodes_clear(*node_alloc_noretry); | |
3220 | else | |
3221 | return -ENOMEM; | |
1da177e4 | 3222 | |
29383967 MK |
3223 | /* |
3224 | * resize_lock mutex prevents concurrent adjustments to number of | |
3225 | * pages in hstate via the proc/sysfs interfaces. | |
3226 | */ | |
3227 | mutex_lock(&h->resize_lock); | |
b65d4adb | 3228 | flush_free_hpage_work(h); |
db71ef79 | 3229 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 3230 | |
fd875dca MK |
3231 | /* |
3232 | * Check for a node specific request. | |
3233 | * Changing node specific huge page count may require a corresponding | |
3234 | * change to the global count. In any case, the passed node mask | |
3235 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
3236 | */ | |
3237 | if (nid != NUMA_NO_NODE) { | |
3238 | unsigned long old_count = count; | |
3239 | ||
3240 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
3241 | /* | |
3242 | * User may have specified a large count value which caused the | |
3243 | * above calculation to overflow. In this case, they wanted | |
3244 | * to allocate as many huge pages as possible. Set count to | |
3245 | * largest possible value to align with their intention. | |
3246 | */ | |
3247 | if (count < old_count) | |
3248 | count = ULONG_MAX; | |
3249 | } | |
3250 | ||
4eb0716e AG |
3251 | /* |
3252 | * Gigantic pages runtime allocation depend on the capability for large | |
3253 | * page range allocation. | |
3254 | * If the system does not provide this feature, return an error when | |
3255 | * the user tries to allocate gigantic pages but let the user free the | |
3256 | * boottime allocated gigantic pages. | |
3257 | */ | |
3258 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
3259 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 3260 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3261 | mutex_unlock(&h->resize_lock); |
f60858f9 | 3262 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
3263 | return -EINVAL; |
3264 | } | |
3265 | /* Fall through to decrease pool */ | |
3266 | } | |
aa888a74 | 3267 | |
7893d1d5 AL |
3268 | /* |
3269 | * Increase the pool size | |
3270 | * First take pages out of surplus state. Then make up the | |
3271 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 3272 | * |
0c397dae | 3273 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
3274 | * to convert a surplus huge page to a normal huge page. That is |
3275 | * not critical, though, it just means the overall size of the | |
3276 | * pool might be one hugepage larger than it needs to be, but | |
3277 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 3278 | */ |
a5516438 | 3279 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 3280 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
3281 | break; |
3282 | } | |
3283 | ||
a5516438 | 3284 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
3285 | /* |
3286 | * If this allocation races such that we no longer need the | |
3287 | * page, free_huge_page will handle it by freeing the page | |
3288 | * and reducing the surplus. | |
3289 | */ | |
db71ef79 | 3290 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
3291 | |
3292 | /* yield cpu to avoid soft lockup */ | |
3293 | cond_resched(); | |
3294 | ||
f60858f9 MK |
3295 | ret = alloc_pool_huge_page(h, nodes_allowed, |
3296 | node_alloc_noretry); | |
db71ef79 | 3297 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
3298 | if (!ret) |
3299 | goto out; | |
3300 | ||
536240f2 MG |
3301 | /* Bail for signals. Probably ctrl-c from user */ |
3302 | if (signal_pending(current)) | |
3303 | goto out; | |
7893d1d5 | 3304 | } |
7893d1d5 AL |
3305 | |
3306 | /* | |
3307 | * Decrease the pool size | |
3308 | * First return free pages to the buddy allocator (being careful | |
3309 | * to keep enough around to satisfy reservations). Then place | |
3310 | * pages into surplus state as needed so the pool will shrink | |
3311 | * to the desired size as pages become free. | |
d1c3fb1f NA |
3312 | * |
3313 | * By placing pages into the surplus state independent of the | |
3314 | * overcommit value, we are allowing the surplus pool size to | |
3315 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 3316 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
3317 | * though, we'll note that we're not allowed to exceed surplus |
3318 | * and won't grow the pool anywhere else. Not until one of the | |
3319 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 3320 | */ |
a5516438 | 3321 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 3322 | min_count = max(count, min_count); |
6ae11b27 | 3323 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
3324 | |
3325 | /* | |
3326 | * Collect pages to be removed on list without dropping lock | |
3327 | */ | |
a5516438 | 3328 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
3329 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
3330 | if (!page) | |
1da177e4 | 3331 | break; |
10c6ec49 MK |
3332 | |
3333 | list_add(&page->lru, &page_list); | |
1da177e4 | 3334 | } |
10c6ec49 | 3335 | /* free the pages after dropping lock */ |
db71ef79 | 3336 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3337 | update_and_free_pages_bulk(h, &page_list); |
b65d4adb | 3338 | flush_free_hpage_work(h); |
db71ef79 | 3339 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 3340 | |
a5516438 | 3341 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 3342 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
3343 | break; |
3344 | } | |
3345 | out: | |
4eb0716e | 3346 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 3347 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3348 | mutex_unlock(&h->resize_lock); |
4eb0716e | 3349 | |
f60858f9 MK |
3350 | NODEMASK_FREE(node_alloc_noretry); |
3351 | ||
4eb0716e | 3352 | return 0; |
1da177e4 LT |
3353 | } |
3354 | ||
8531fc6f MK |
3355 | static int demote_free_huge_page(struct hstate *h, struct page *page) |
3356 | { | |
3357 | int i, nid = page_to_nid(page); | |
3358 | struct hstate *target_hstate; | |
3359 | int rc = 0; | |
3360 | ||
3361 | target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); | |
3362 | ||
3363 | remove_hugetlb_page_for_demote(h, page, false); | |
3364 | spin_unlock_irq(&hugetlb_lock); | |
3365 | ||
3366 | rc = alloc_huge_page_vmemmap(h, page); | |
3367 | if (rc) { | |
3368 | /* Allocation of vmemmmap failed, we can not demote page */ | |
3369 | spin_lock_irq(&hugetlb_lock); | |
3370 | set_page_refcounted(page); | |
3371 | add_hugetlb_page(h, page, false); | |
3372 | return rc; | |
3373 | } | |
3374 | ||
3375 | /* | |
3376 | * Use destroy_compound_hugetlb_page_for_demote for all huge page | |
3377 | * sizes as it will not ref count pages. | |
3378 | */ | |
3379 | destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); | |
3380 | ||
3381 | /* | |
3382 | * Taking target hstate mutex synchronizes with set_max_huge_pages. | |
3383 | * Without the mutex, pages added to target hstate could be marked | |
3384 | * as surplus. | |
3385 | * | |
3386 | * Note that we already hold h->resize_lock. To prevent deadlock, | |
3387 | * use the convention of always taking larger size hstate mutex first. | |
3388 | */ | |
3389 | mutex_lock(&target_hstate->resize_lock); | |
3390 | for (i = 0; i < pages_per_huge_page(h); | |
3391 | i += pages_per_huge_page(target_hstate)) { | |
3392 | if (hstate_is_gigantic(target_hstate)) | |
3393 | prep_compound_gigantic_page_for_demote(page + i, | |
3394 | target_hstate->order); | |
3395 | else | |
3396 | prep_compound_page(page + i, target_hstate->order); | |
3397 | set_page_private(page + i, 0); | |
3398 | set_page_refcounted(page + i); | |
3399 | prep_new_huge_page(target_hstate, page + i, nid); | |
3400 | put_page(page + i); | |
3401 | } | |
3402 | mutex_unlock(&target_hstate->resize_lock); | |
3403 | ||
3404 | spin_lock_irq(&hugetlb_lock); | |
3405 | ||
3406 | /* | |
3407 | * Not absolutely necessary, but for consistency update max_huge_pages | |
3408 | * based on pool changes for the demoted page. | |
3409 | */ | |
3410 | h->max_huge_pages--; | |
3411 | target_hstate->max_huge_pages += pages_per_huge_page(h); | |
3412 | ||
3413 | return rc; | |
3414 | } | |
3415 | ||
79dfc695 MK |
3416 | static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
3417 | __must_hold(&hugetlb_lock) | |
3418 | { | |
8531fc6f MK |
3419 | int nr_nodes, node; |
3420 | struct page *page; | |
79dfc695 MK |
3421 | int rc = 0; |
3422 | ||
3423 | lockdep_assert_held(&hugetlb_lock); | |
3424 | ||
3425 | /* We should never get here if no demote order */ | |
3426 | if (!h->demote_order) { | |
3427 | pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); | |
3428 | return -EINVAL; /* internal error */ | |
3429 | } | |
3430 | ||
8531fc6f MK |
3431 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
3432 | if (!list_empty(&h->hugepage_freelists[node])) { | |
3433 | page = list_entry(h->hugepage_freelists[node].next, | |
3434 | struct page, lru); | |
3435 | rc = demote_free_huge_page(h, page); | |
3436 | break; | |
3437 | } | |
3438 | } | |
3439 | ||
79dfc695 MK |
3440 | return rc; |
3441 | } | |
3442 | ||
a3437870 NA |
3443 | #define HSTATE_ATTR_RO(_name) \ |
3444 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
3445 | ||
79dfc695 MK |
3446 | #define HSTATE_ATTR_WO(_name) \ |
3447 | static struct kobj_attribute _name##_attr = __ATTR_WO(_name) | |
3448 | ||
a3437870 NA |
3449 | #define HSTATE_ATTR(_name) \ |
3450 | static struct kobj_attribute _name##_attr = \ | |
3451 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
3452 | ||
3453 | static struct kobject *hugepages_kobj; | |
3454 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3455 | ||
9a305230 LS |
3456 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
3457 | ||
3458 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
3459 | { |
3460 | int i; | |
9a305230 | 3461 | |
a3437870 | 3462 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
3463 | if (hstate_kobjs[i] == kobj) { |
3464 | if (nidp) | |
3465 | *nidp = NUMA_NO_NODE; | |
a3437870 | 3466 | return &hstates[i]; |
9a305230 LS |
3467 | } |
3468 | ||
3469 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
3470 | } |
3471 | ||
06808b08 | 3472 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
3473 | struct kobj_attribute *attr, char *buf) |
3474 | { | |
9a305230 LS |
3475 | struct hstate *h; |
3476 | unsigned long nr_huge_pages; | |
3477 | int nid; | |
3478 | ||
3479 | h = kobj_to_hstate(kobj, &nid); | |
3480 | if (nid == NUMA_NO_NODE) | |
3481 | nr_huge_pages = h->nr_huge_pages; | |
3482 | else | |
3483 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
3484 | ||
ae7a927d | 3485 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 3486 | } |
adbe8726 | 3487 | |
238d3c13 DR |
3488 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
3489 | struct hstate *h, int nid, | |
3490 | unsigned long count, size_t len) | |
a3437870 NA |
3491 | { |
3492 | int err; | |
2d0adf7e | 3493 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 3494 | |
2d0adf7e OS |
3495 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
3496 | return -EINVAL; | |
adbe8726 | 3497 | |
9a305230 LS |
3498 | if (nid == NUMA_NO_NODE) { |
3499 | /* | |
3500 | * global hstate attribute | |
3501 | */ | |
3502 | if (!(obey_mempolicy && | |
2d0adf7e OS |
3503 | init_nodemask_of_mempolicy(&nodes_allowed))) |
3504 | n_mask = &node_states[N_MEMORY]; | |
3505 | else | |
3506 | n_mask = &nodes_allowed; | |
3507 | } else { | |
9a305230 | 3508 | /* |
fd875dca MK |
3509 | * Node specific request. count adjustment happens in |
3510 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 3511 | */ |
2d0adf7e OS |
3512 | init_nodemask_of_node(&nodes_allowed, nid); |
3513 | n_mask = &nodes_allowed; | |
fd875dca | 3514 | } |
9a305230 | 3515 | |
2d0adf7e | 3516 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 3517 | |
4eb0716e | 3518 | return err ? err : len; |
06808b08 LS |
3519 | } |
3520 | ||
238d3c13 DR |
3521 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
3522 | struct kobject *kobj, const char *buf, | |
3523 | size_t len) | |
3524 | { | |
3525 | struct hstate *h; | |
3526 | unsigned long count; | |
3527 | int nid; | |
3528 | int err; | |
3529 | ||
3530 | err = kstrtoul(buf, 10, &count); | |
3531 | if (err) | |
3532 | return err; | |
3533 | ||
3534 | h = kobj_to_hstate(kobj, &nid); | |
3535 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
3536 | } | |
3537 | ||
06808b08 LS |
3538 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
3539 | struct kobj_attribute *attr, char *buf) | |
3540 | { | |
3541 | return nr_hugepages_show_common(kobj, attr, buf); | |
3542 | } | |
3543 | ||
3544 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
3545 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3546 | { | |
238d3c13 | 3547 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
3548 | } |
3549 | HSTATE_ATTR(nr_hugepages); | |
3550 | ||
06808b08 LS |
3551 | #ifdef CONFIG_NUMA |
3552 | ||
3553 | /* | |
3554 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
3555 | * huge page alloc/free. | |
3556 | */ | |
3557 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
3558 | struct kobj_attribute *attr, |
3559 | char *buf) | |
06808b08 LS |
3560 | { |
3561 | return nr_hugepages_show_common(kobj, attr, buf); | |
3562 | } | |
3563 | ||
3564 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
3565 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3566 | { | |
238d3c13 | 3567 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
3568 | } |
3569 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
3570 | #endif | |
3571 | ||
3572 | ||
a3437870 NA |
3573 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
3574 | struct kobj_attribute *attr, char *buf) | |
3575 | { | |
9a305230 | 3576 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3577 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 3578 | } |
adbe8726 | 3579 | |
a3437870 NA |
3580 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
3581 | struct kobj_attribute *attr, const char *buf, size_t count) | |
3582 | { | |
3583 | int err; | |
3584 | unsigned long input; | |
9a305230 | 3585 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 3586 | |
bae7f4ae | 3587 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
3588 | return -EINVAL; |
3589 | ||
3dbb95f7 | 3590 | err = kstrtoul(buf, 10, &input); |
a3437870 | 3591 | if (err) |
73ae31e5 | 3592 | return err; |
a3437870 | 3593 | |
db71ef79 | 3594 | spin_lock_irq(&hugetlb_lock); |
a3437870 | 3595 | h->nr_overcommit_huge_pages = input; |
db71ef79 | 3596 | spin_unlock_irq(&hugetlb_lock); |
a3437870 NA |
3597 | |
3598 | return count; | |
3599 | } | |
3600 | HSTATE_ATTR(nr_overcommit_hugepages); | |
3601 | ||
3602 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
3603 | struct kobj_attribute *attr, char *buf) | |
3604 | { | |
9a305230 LS |
3605 | struct hstate *h; |
3606 | unsigned long free_huge_pages; | |
3607 | int nid; | |
3608 | ||
3609 | h = kobj_to_hstate(kobj, &nid); | |
3610 | if (nid == NUMA_NO_NODE) | |
3611 | free_huge_pages = h->free_huge_pages; | |
3612 | else | |
3613 | free_huge_pages = h->free_huge_pages_node[nid]; | |
3614 | ||
ae7a927d | 3615 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
3616 | } |
3617 | HSTATE_ATTR_RO(free_hugepages); | |
3618 | ||
3619 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
3620 | struct kobj_attribute *attr, char *buf) | |
3621 | { | |
9a305230 | 3622 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3623 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
3624 | } |
3625 | HSTATE_ATTR_RO(resv_hugepages); | |
3626 | ||
3627 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
3628 | struct kobj_attribute *attr, char *buf) | |
3629 | { | |
9a305230 LS |
3630 | struct hstate *h; |
3631 | unsigned long surplus_huge_pages; | |
3632 | int nid; | |
3633 | ||
3634 | h = kobj_to_hstate(kobj, &nid); | |
3635 | if (nid == NUMA_NO_NODE) | |
3636 | surplus_huge_pages = h->surplus_huge_pages; | |
3637 | else | |
3638 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
3639 | ||
ae7a927d | 3640 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
3641 | } |
3642 | HSTATE_ATTR_RO(surplus_hugepages); | |
3643 | ||
79dfc695 MK |
3644 | static ssize_t demote_store(struct kobject *kobj, |
3645 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3646 | { | |
3647 | unsigned long nr_demote; | |
3648 | unsigned long nr_available; | |
3649 | nodemask_t nodes_allowed, *n_mask; | |
3650 | struct hstate *h; | |
3651 | int err = 0; | |
3652 | int nid; | |
3653 | ||
3654 | err = kstrtoul(buf, 10, &nr_demote); | |
3655 | if (err) | |
3656 | return err; | |
3657 | h = kobj_to_hstate(kobj, &nid); | |
3658 | ||
3659 | if (nid != NUMA_NO_NODE) { | |
3660 | init_nodemask_of_node(&nodes_allowed, nid); | |
3661 | n_mask = &nodes_allowed; | |
3662 | } else { | |
3663 | n_mask = &node_states[N_MEMORY]; | |
3664 | } | |
3665 | ||
3666 | /* Synchronize with other sysfs operations modifying huge pages */ | |
3667 | mutex_lock(&h->resize_lock); | |
3668 | spin_lock_irq(&hugetlb_lock); | |
3669 | ||
3670 | while (nr_demote) { | |
3671 | /* | |
3672 | * Check for available pages to demote each time thorough the | |
3673 | * loop as demote_pool_huge_page will drop hugetlb_lock. | |
79dfc695 MK |
3674 | */ |
3675 | if (nid != NUMA_NO_NODE) | |
3676 | nr_available = h->free_huge_pages_node[nid]; | |
3677 | else | |
3678 | nr_available = h->free_huge_pages; | |
3679 | nr_available -= h->resv_huge_pages; | |
3680 | if (!nr_available) | |
3681 | break; | |
3682 | ||
3683 | err = demote_pool_huge_page(h, n_mask); | |
3684 | if (err) | |
3685 | break; | |
3686 | ||
3687 | nr_demote--; | |
3688 | } | |
3689 | ||
3690 | spin_unlock_irq(&hugetlb_lock); | |
3691 | mutex_unlock(&h->resize_lock); | |
3692 | ||
3693 | if (err) | |
3694 | return err; | |
3695 | return len; | |
3696 | } | |
3697 | HSTATE_ATTR_WO(demote); | |
3698 | ||
3699 | static ssize_t demote_size_show(struct kobject *kobj, | |
3700 | struct kobj_attribute *attr, char *buf) | |
3701 | { | |
3702 | int nid; | |
3703 | struct hstate *h = kobj_to_hstate(kobj, &nid); | |
3704 | unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; | |
3705 | ||
3706 | return sysfs_emit(buf, "%lukB\n", demote_size); | |
3707 | } | |
3708 | ||
3709 | static ssize_t demote_size_store(struct kobject *kobj, | |
3710 | struct kobj_attribute *attr, | |
3711 | const char *buf, size_t count) | |
3712 | { | |
3713 | struct hstate *h, *demote_hstate; | |
3714 | unsigned long demote_size; | |
3715 | unsigned int demote_order; | |
3716 | int nid; | |
3717 | ||
3718 | demote_size = (unsigned long)memparse(buf, NULL); | |
3719 | ||
3720 | demote_hstate = size_to_hstate(demote_size); | |
3721 | if (!demote_hstate) | |
3722 | return -EINVAL; | |
3723 | demote_order = demote_hstate->order; | |
a01f4390 MK |
3724 | if (demote_order < HUGETLB_PAGE_ORDER) |
3725 | return -EINVAL; | |
79dfc695 MK |
3726 | |
3727 | /* demote order must be smaller than hstate order */ | |
3728 | h = kobj_to_hstate(kobj, &nid); | |
3729 | if (demote_order >= h->order) | |
3730 | return -EINVAL; | |
3731 | ||
3732 | /* resize_lock synchronizes access to demote size and writes */ | |
3733 | mutex_lock(&h->resize_lock); | |
3734 | h->demote_order = demote_order; | |
3735 | mutex_unlock(&h->resize_lock); | |
3736 | ||
3737 | return count; | |
3738 | } | |
3739 | HSTATE_ATTR(demote_size); | |
3740 | ||
a3437870 NA |
3741 | static struct attribute *hstate_attrs[] = { |
3742 | &nr_hugepages_attr.attr, | |
3743 | &nr_overcommit_hugepages_attr.attr, | |
3744 | &free_hugepages_attr.attr, | |
3745 | &resv_hugepages_attr.attr, | |
3746 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3747 | #ifdef CONFIG_NUMA |
3748 | &nr_hugepages_mempolicy_attr.attr, | |
3749 | #endif | |
a3437870 NA |
3750 | NULL, |
3751 | }; | |
3752 | ||
67e5ed96 | 3753 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3754 | .attrs = hstate_attrs, |
3755 | }; | |
3756 | ||
79dfc695 MK |
3757 | static struct attribute *hstate_demote_attrs[] = { |
3758 | &demote_size_attr.attr, | |
3759 | &demote_attr.attr, | |
3760 | NULL, | |
3761 | }; | |
3762 | ||
3763 | static const struct attribute_group hstate_demote_attr_group = { | |
3764 | .attrs = hstate_demote_attrs, | |
3765 | }; | |
3766 | ||
094e9539 JM |
3767 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3768 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3769 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3770 | { |
3771 | int retval; | |
972dc4de | 3772 | int hi = hstate_index(h); |
a3437870 | 3773 | |
9a305230 LS |
3774 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3775 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3776 | return -ENOMEM; |
3777 | ||
9a305230 | 3778 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 3779 | if (retval) { |
9a305230 | 3780 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 ML |
3781 | hstate_kobjs[hi] = NULL; |
3782 | } | |
a3437870 | 3783 | |
79dfc695 MK |
3784 | if (h->demote_order) { |
3785 | if (sysfs_create_group(hstate_kobjs[hi], | |
3786 | &hstate_demote_attr_group)) | |
3787 | pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); | |
3788 | } | |
3789 | ||
a3437870 NA |
3790 | return retval; |
3791 | } | |
3792 | ||
3793 | static void __init hugetlb_sysfs_init(void) | |
3794 | { | |
3795 | struct hstate *h; | |
3796 | int err; | |
3797 | ||
3798 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3799 | if (!hugepages_kobj) | |
3800 | return; | |
3801 | ||
3802 | for_each_hstate(h) { | |
9a305230 LS |
3803 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3804 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3805 | if (err) |
282f4214 | 3806 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
3807 | } |
3808 | } | |
3809 | ||
9a305230 LS |
3810 | #ifdef CONFIG_NUMA |
3811 | ||
3812 | /* | |
3813 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3814 | * with node devices in node_devices[] using a parallel array. The array |
3815 | * index of a node device or _hstate == node id. | |
3816 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3817 | * the base kernel, on the hugetlb module. |
3818 | */ | |
3819 | struct node_hstate { | |
3820 | struct kobject *hugepages_kobj; | |
3821 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3822 | }; | |
b4e289a6 | 3823 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3824 | |
3825 | /* | |
10fbcf4c | 3826 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3827 | */ |
3828 | static struct attribute *per_node_hstate_attrs[] = { | |
3829 | &nr_hugepages_attr.attr, | |
3830 | &free_hugepages_attr.attr, | |
3831 | &surplus_hugepages_attr.attr, | |
3832 | NULL, | |
3833 | }; | |
3834 | ||
67e5ed96 | 3835 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3836 | .attrs = per_node_hstate_attrs, |
3837 | }; | |
3838 | ||
3839 | /* | |
10fbcf4c | 3840 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3841 | * Returns node id via non-NULL nidp. |
3842 | */ | |
3843 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3844 | { | |
3845 | int nid; | |
3846 | ||
3847 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3848 | struct node_hstate *nhs = &node_hstates[nid]; | |
3849 | int i; | |
3850 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3851 | if (nhs->hstate_kobjs[i] == kobj) { | |
3852 | if (nidp) | |
3853 | *nidp = nid; | |
3854 | return &hstates[i]; | |
3855 | } | |
3856 | } | |
3857 | ||
3858 | BUG(); | |
3859 | return NULL; | |
3860 | } | |
3861 | ||
3862 | /* | |
10fbcf4c | 3863 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3864 | * No-op if no hstate attributes attached. |
3865 | */ | |
3cd8b44f | 3866 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3867 | { |
3868 | struct hstate *h; | |
10fbcf4c | 3869 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3870 | |
3871 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3872 | return; /* no hstate attributes */ |
9a305230 | 3873 | |
972dc4de AK |
3874 | for_each_hstate(h) { |
3875 | int idx = hstate_index(h); | |
3876 | if (nhs->hstate_kobjs[idx]) { | |
3877 | kobject_put(nhs->hstate_kobjs[idx]); | |
3878 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3879 | } |
972dc4de | 3880 | } |
9a305230 LS |
3881 | |
3882 | kobject_put(nhs->hugepages_kobj); | |
3883 | nhs->hugepages_kobj = NULL; | |
3884 | } | |
3885 | ||
9a305230 LS |
3886 | |
3887 | /* | |
10fbcf4c | 3888 | * Register hstate attributes for a single node device. |
9a305230 LS |
3889 | * No-op if attributes already registered. |
3890 | */ | |
3cd8b44f | 3891 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3892 | { |
3893 | struct hstate *h; | |
10fbcf4c | 3894 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3895 | int err; |
3896 | ||
3897 | if (nhs->hugepages_kobj) | |
3898 | return; /* already allocated */ | |
3899 | ||
3900 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3901 | &node->dev.kobj); |
9a305230 LS |
3902 | if (!nhs->hugepages_kobj) |
3903 | return; | |
3904 | ||
3905 | for_each_hstate(h) { | |
3906 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3907 | nhs->hstate_kobjs, | |
3908 | &per_node_hstate_attr_group); | |
3909 | if (err) { | |
282f4214 | 3910 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3911 | h->name, node->dev.id); |
9a305230 LS |
3912 | hugetlb_unregister_node(node); |
3913 | break; | |
3914 | } | |
3915 | } | |
3916 | } | |
3917 | ||
3918 | /* | |
9b5e5d0f | 3919 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3920 | * devices of nodes that have memory. All on-line nodes should have |
3921 | * registered their associated device by this time. | |
9a305230 | 3922 | */ |
7d9ca000 | 3923 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3924 | { |
3925 | int nid; | |
3926 | ||
8cebfcd0 | 3927 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3928 | struct node *node = node_devices[nid]; |
10fbcf4c | 3929 | if (node->dev.id == nid) |
9a305230 LS |
3930 | hugetlb_register_node(node); |
3931 | } | |
3932 | ||
3933 | /* | |
10fbcf4c | 3934 | * Let the node device driver know we're here so it can |
9a305230 LS |
3935 | * [un]register hstate attributes on node hotplug. |
3936 | */ | |
3937 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3938 | hugetlb_unregister_node); | |
3939 | } | |
3940 | #else /* !CONFIG_NUMA */ | |
3941 | ||
3942 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3943 | { | |
3944 | BUG(); | |
3945 | if (nidp) | |
3946 | *nidp = -1; | |
3947 | return NULL; | |
3948 | } | |
3949 | ||
9a305230 LS |
3950 | static void hugetlb_register_all_nodes(void) { } |
3951 | ||
3952 | #endif | |
3953 | ||
a3437870 NA |
3954 | static int __init hugetlb_init(void) |
3955 | { | |
8382d914 DB |
3956 | int i; |
3957 | ||
d6995da3 MK |
3958 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
3959 | __NR_HPAGEFLAGS); | |
3960 | ||
c2833a5b MK |
3961 | if (!hugepages_supported()) { |
3962 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
3963 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 3964 | return 0; |
c2833a5b | 3965 | } |
a3437870 | 3966 | |
282f4214 MK |
3967 | /* |
3968 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
3969 | * architectures depend on setup being done here. | |
3970 | */ | |
3971 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
3972 | if (!parsed_default_hugepagesz) { | |
3973 | /* | |
3974 | * If we did not parse a default huge page size, set | |
3975 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
3976 | * number of huge pages for this default size was implicitly | |
3977 | * specified, set that here as well. | |
3978 | * Note that the implicit setting will overwrite an explicit | |
3979 | * setting. A warning will be printed in this case. | |
3980 | */ | |
3981 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
3982 | if (default_hstate_max_huge_pages) { | |
3983 | if (default_hstate.max_huge_pages) { | |
3984 | char buf[32]; | |
3985 | ||
3986 | string_get_size(huge_page_size(&default_hstate), | |
3987 | 1, STRING_UNITS_2, buf, 32); | |
3988 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
3989 | default_hstate.max_huge_pages, buf); | |
3990 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
3991 | default_hstate_max_huge_pages); | |
3992 | } | |
3993 | default_hstate.max_huge_pages = | |
3994 | default_hstate_max_huge_pages; | |
d715cf80 | 3995 | } |
f8b74815 | 3996 | } |
a3437870 | 3997 | |
cf11e85f | 3998 | hugetlb_cma_check(); |
a3437870 | 3999 | hugetlb_init_hstates(); |
aa888a74 | 4000 | gather_bootmem_prealloc(); |
a3437870 NA |
4001 | report_hugepages(); |
4002 | ||
4003 | hugetlb_sysfs_init(); | |
9a305230 | 4004 | hugetlb_register_all_nodes(); |
7179e7bf | 4005 | hugetlb_cgroup_file_init(); |
9a305230 | 4006 | |
8382d914 DB |
4007 | #ifdef CONFIG_SMP |
4008 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
4009 | #else | |
4010 | num_fault_mutexes = 1; | |
4011 | #endif | |
c672c7f2 | 4012 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
4013 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
4014 | GFP_KERNEL); | |
c672c7f2 | 4015 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
4016 | |
4017 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 4018 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
4019 | return 0; |
4020 | } | |
3e89e1c5 | 4021 | subsys_initcall(hugetlb_init); |
a3437870 | 4022 | |
ae94da89 MK |
4023 | /* Overwritten by architectures with more huge page sizes */ |
4024 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 4025 | { |
ae94da89 | 4026 | return size == HPAGE_SIZE; |
9fee021d VT |
4027 | } |
4028 | ||
d00181b9 | 4029 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
4030 | { |
4031 | struct hstate *h; | |
8faa8b07 AK |
4032 | unsigned long i; |
4033 | ||
a3437870 | 4034 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
4035 | return; |
4036 | } | |
47d38344 | 4037 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 4038 | BUG_ON(order == 0); |
47d38344 | 4039 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 4040 | mutex_init(&h->resize_lock); |
a3437870 | 4041 | h->order = order; |
aca78307 | 4042 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
4043 | for (i = 0; i < MAX_NUMNODES; ++i) |
4044 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 4045 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
4046 | h->next_nid_to_alloc = first_memory_node; |
4047 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
4048 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
4049 | huge_page_size(h)/1024); | |
77490587 | 4050 | hugetlb_vmemmap_init(h); |
8faa8b07 | 4051 | |
a3437870 NA |
4052 | parsed_hstate = h; |
4053 | } | |
4054 | ||
282f4214 MK |
4055 | /* |
4056 | * hugepages command line processing | |
4057 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
4058 | * specification. If not, ignore the hugepages value. hugepages can also | |
4059 | * be the first huge page command line option in which case it implicitly | |
4060 | * specifies the number of huge pages for the default size. | |
4061 | */ | |
4062 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
4063 | { |
4064 | unsigned long *mhp; | |
8faa8b07 | 4065 | static unsigned long *last_mhp; |
a3437870 | 4066 | |
9fee021d | 4067 | if (!parsed_valid_hugepagesz) { |
282f4214 | 4068 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 4069 | parsed_valid_hugepagesz = true; |
282f4214 | 4070 | return 0; |
9fee021d | 4071 | } |
282f4214 | 4072 | |
a3437870 | 4073 | /* |
282f4214 MK |
4074 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
4075 | * yet, so this hugepages= parameter goes to the "default hstate". | |
4076 | * Otherwise, it goes with the previously parsed hugepagesz or | |
4077 | * default_hugepagesz. | |
a3437870 | 4078 | */ |
9fee021d | 4079 | else if (!hugetlb_max_hstate) |
a3437870 NA |
4080 | mhp = &default_hstate_max_huge_pages; |
4081 | else | |
4082 | mhp = &parsed_hstate->max_huge_pages; | |
4083 | ||
8faa8b07 | 4084 | if (mhp == last_mhp) { |
282f4214 MK |
4085 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
4086 | return 0; | |
8faa8b07 AK |
4087 | } |
4088 | ||
a3437870 NA |
4089 | if (sscanf(s, "%lu", mhp) <= 0) |
4090 | *mhp = 0; | |
4091 | ||
8faa8b07 AK |
4092 | /* |
4093 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 4094 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
4095 | * use the bootmem allocator. |
4096 | */ | |
04adbc3f | 4097 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
4098 | hugetlb_hstate_alloc_pages(parsed_hstate); |
4099 | ||
4100 | last_mhp = mhp; | |
4101 | ||
a3437870 NA |
4102 | return 1; |
4103 | } | |
282f4214 | 4104 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 4105 | |
282f4214 MK |
4106 | /* |
4107 | * hugepagesz command line processing | |
4108 | * A specific huge page size can only be specified once with hugepagesz. | |
4109 | * hugepagesz is followed by hugepages on the command line. The global | |
4110 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
4111 | * hugepagesz argument was valid. | |
4112 | */ | |
359f2544 | 4113 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 4114 | { |
359f2544 | 4115 | unsigned long size; |
282f4214 MK |
4116 | struct hstate *h; |
4117 | ||
4118 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
4119 | size = (unsigned long)memparse(s, NULL); |
4120 | ||
4121 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4122 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
359f2544 MK |
4123 | return 0; |
4124 | } | |
4125 | ||
282f4214 MK |
4126 | h = size_to_hstate(size); |
4127 | if (h) { | |
4128 | /* | |
4129 | * hstate for this size already exists. This is normally | |
4130 | * an error, but is allowed if the existing hstate is the | |
4131 | * default hstate. More specifically, it is only allowed if | |
4132 | * the number of huge pages for the default hstate was not | |
4133 | * previously specified. | |
4134 | */ | |
4135 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
4136 | default_hstate.max_huge_pages) { | |
4137 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
4138 | return 0; | |
4139 | } | |
4140 | ||
4141 | /* | |
4142 | * No need to call hugetlb_add_hstate() as hstate already | |
4143 | * exists. But, do set parsed_hstate so that a following | |
4144 | * hugepages= parameter will be applied to this hstate. | |
4145 | */ | |
4146 | parsed_hstate = h; | |
4147 | parsed_valid_hugepagesz = true; | |
4148 | return 1; | |
38237830 MK |
4149 | } |
4150 | ||
359f2544 | 4151 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 4152 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
4153 | return 1; |
4154 | } | |
359f2544 MK |
4155 | __setup("hugepagesz=", hugepagesz_setup); |
4156 | ||
282f4214 MK |
4157 | /* |
4158 | * default_hugepagesz command line input | |
4159 | * Only one instance of default_hugepagesz allowed on command line. | |
4160 | */ | |
ae94da89 | 4161 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 4162 | { |
ae94da89 MK |
4163 | unsigned long size; |
4164 | ||
282f4214 | 4165 | parsed_valid_hugepagesz = false; |
282f4214 MK |
4166 | if (parsed_default_hugepagesz) { |
4167 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
4168 | return 0; | |
4169 | } | |
4170 | ||
ae94da89 MK |
4171 | size = (unsigned long)memparse(s, NULL); |
4172 | ||
4173 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4174 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
ae94da89 MK |
4175 | return 0; |
4176 | } | |
4177 | ||
282f4214 MK |
4178 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
4179 | parsed_valid_hugepagesz = true; | |
4180 | parsed_default_hugepagesz = true; | |
4181 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
4182 | ||
4183 | /* | |
4184 | * The number of default huge pages (for this size) could have been | |
4185 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
4186 | * then default_hstate_max_huge_pages is set. If the default huge | |
4187 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
4188 | * allocated here from bootmem allocator. | |
4189 | */ | |
4190 | if (default_hstate_max_huge_pages) { | |
4191 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
4192 | if (hstate_is_gigantic(&default_hstate)) | |
4193 | hugetlb_hstate_alloc_pages(&default_hstate); | |
4194 | default_hstate_max_huge_pages = 0; | |
4195 | } | |
4196 | ||
e11bfbfc NP |
4197 | return 1; |
4198 | } | |
ae94da89 | 4199 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 4200 | |
8ca39e68 | 4201 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
4202 | { |
4203 | int node; | |
4204 | unsigned int nr = 0; | |
8ca39e68 MS |
4205 | nodemask_t *mpol_allowed; |
4206 | unsigned int *array = h->free_huge_pages_node; | |
4207 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
4208 | ||
4209 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 4210 | |
8ca39e68 | 4211 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
c93b0a99 | 4212 | if (!mpol_allowed || node_isset(node, *mpol_allowed)) |
8ca39e68 MS |
4213 | nr += array[node]; |
4214 | } | |
8a213460 NA |
4215 | |
4216 | return nr; | |
4217 | } | |
4218 | ||
4219 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
4220 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
4221 | void *buffer, size_t *length, | |
4222 | loff_t *ppos, unsigned long *out) | |
4223 | { | |
4224 | struct ctl_table dup_table; | |
4225 | ||
4226 | /* | |
4227 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
4228 | * can duplicate the @table and alter the duplicate of it. | |
4229 | */ | |
4230 | dup_table = *table; | |
4231 | dup_table.data = out; | |
4232 | ||
4233 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
4234 | } | |
4235 | ||
06808b08 LS |
4236 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
4237 | struct ctl_table *table, int write, | |
32927393 | 4238 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 4239 | { |
e5ff2159 | 4240 | struct hstate *h = &default_hstate; |
238d3c13 | 4241 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 4242 | int ret; |
e5ff2159 | 4243 | |
457c1b27 | 4244 | if (!hugepages_supported()) |
86613628 | 4245 | return -EOPNOTSUPP; |
457c1b27 | 4246 | |
17743798 MS |
4247 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4248 | &tmp); | |
08d4a246 MH |
4249 | if (ret) |
4250 | goto out; | |
e5ff2159 | 4251 | |
238d3c13 DR |
4252 | if (write) |
4253 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
4254 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
4255 | out: |
4256 | return ret; | |
1da177e4 | 4257 | } |
396faf03 | 4258 | |
06808b08 | 4259 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 4260 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4261 | { |
4262 | ||
4263 | return hugetlb_sysctl_handler_common(false, table, write, | |
4264 | buffer, length, ppos); | |
4265 | } | |
4266 | ||
4267 | #ifdef CONFIG_NUMA | |
4268 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 4269 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4270 | { |
4271 | return hugetlb_sysctl_handler_common(true, table, write, | |
4272 | buffer, length, ppos); | |
4273 | } | |
4274 | #endif /* CONFIG_NUMA */ | |
4275 | ||
a3d0c6aa | 4276 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 4277 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 4278 | { |
a5516438 | 4279 | struct hstate *h = &default_hstate; |
e5ff2159 | 4280 | unsigned long tmp; |
08d4a246 | 4281 | int ret; |
e5ff2159 | 4282 | |
457c1b27 | 4283 | if (!hugepages_supported()) |
86613628 | 4284 | return -EOPNOTSUPP; |
457c1b27 | 4285 | |
c033a93c | 4286 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 4287 | |
bae7f4ae | 4288 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
4289 | return -EINVAL; |
4290 | ||
17743798 MS |
4291 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4292 | &tmp); | |
08d4a246 MH |
4293 | if (ret) |
4294 | goto out; | |
e5ff2159 AK |
4295 | |
4296 | if (write) { | |
db71ef79 | 4297 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 4298 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 4299 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 4300 | } |
08d4a246 MH |
4301 | out: |
4302 | return ret; | |
a3d0c6aa NA |
4303 | } |
4304 | ||
1da177e4 LT |
4305 | #endif /* CONFIG_SYSCTL */ |
4306 | ||
e1759c21 | 4307 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 4308 | { |
fcb2b0c5 RG |
4309 | struct hstate *h; |
4310 | unsigned long total = 0; | |
4311 | ||
457c1b27 NA |
4312 | if (!hugepages_supported()) |
4313 | return; | |
fcb2b0c5 RG |
4314 | |
4315 | for_each_hstate(h) { | |
4316 | unsigned long count = h->nr_huge_pages; | |
4317 | ||
aca78307 | 4318 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
4319 | |
4320 | if (h == &default_hstate) | |
4321 | seq_printf(m, | |
4322 | "HugePages_Total: %5lu\n" | |
4323 | "HugePages_Free: %5lu\n" | |
4324 | "HugePages_Rsvd: %5lu\n" | |
4325 | "HugePages_Surp: %5lu\n" | |
4326 | "Hugepagesize: %8lu kB\n", | |
4327 | count, | |
4328 | h->free_huge_pages, | |
4329 | h->resv_huge_pages, | |
4330 | h->surplus_huge_pages, | |
aca78307 | 4331 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
4332 | } |
4333 | ||
aca78307 | 4334 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
4335 | } |
4336 | ||
7981593b | 4337 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 4338 | { |
a5516438 | 4339 | struct hstate *h = &default_hstate; |
7981593b | 4340 | |
457c1b27 NA |
4341 | if (!hugepages_supported()) |
4342 | return 0; | |
7981593b JP |
4343 | |
4344 | return sysfs_emit_at(buf, len, | |
4345 | "Node %d HugePages_Total: %5u\n" | |
4346 | "Node %d HugePages_Free: %5u\n" | |
4347 | "Node %d HugePages_Surp: %5u\n", | |
4348 | nid, h->nr_huge_pages_node[nid], | |
4349 | nid, h->free_huge_pages_node[nid], | |
4350 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
4351 | } |
4352 | ||
949f7ec5 DR |
4353 | void hugetlb_show_meminfo(void) |
4354 | { | |
4355 | struct hstate *h; | |
4356 | int nid; | |
4357 | ||
457c1b27 NA |
4358 | if (!hugepages_supported()) |
4359 | return; | |
4360 | ||
949f7ec5 DR |
4361 | for_each_node_state(nid, N_MEMORY) |
4362 | for_each_hstate(h) | |
4363 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
4364 | nid, | |
4365 | h->nr_huge_pages_node[nid], | |
4366 | h->free_huge_pages_node[nid], | |
4367 | h->surplus_huge_pages_node[nid], | |
aca78307 | 4368 | huge_page_size(h) / SZ_1K); |
949f7ec5 DR |
4369 | } |
4370 | ||
5d317b2b NH |
4371 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
4372 | { | |
4373 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
4374 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
4375 | } | |
4376 | ||
1da177e4 LT |
4377 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
4378 | unsigned long hugetlb_total_pages(void) | |
4379 | { | |
d0028588 WL |
4380 | struct hstate *h; |
4381 | unsigned long nr_total_pages = 0; | |
4382 | ||
4383 | for_each_hstate(h) | |
4384 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
4385 | return nr_total_pages; | |
1da177e4 | 4386 | } |
1da177e4 | 4387 | |
a5516438 | 4388 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
4389 | { |
4390 | int ret = -ENOMEM; | |
4391 | ||
0aa7f354 ML |
4392 | if (!delta) |
4393 | return 0; | |
4394 | ||
db71ef79 | 4395 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4396 | /* |
4397 | * When cpuset is configured, it breaks the strict hugetlb page | |
4398 | * reservation as the accounting is done on a global variable. Such | |
4399 | * reservation is completely rubbish in the presence of cpuset because | |
4400 | * the reservation is not checked against page availability for the | |
4401 | * current cpuset. Application can still potentially OOM'ed by kernel | |
4402 | * with lack of free htlb page in cpuset that the task is in. | |
4403 | * Attempt to enforce strict accounting with cpuset is almost | |
4404 | * impossible (or too ugly) because cpuset is too fluid that | |
4405 | * task or memory node can be dynamically moved between cpusets. | |
4406 | * | |
4407 | * The change of semantics for shared hugetlb mapping with cpuset is | |
4408 | * undesirable. However, in order to preserve some of the semantics, | |
4409 | * we fall back to check against current free page availability as | |
4410 | * a best attempt and hopefully to minimize the impact of changing | |
4411 | * semantics that cpuset has. | |
8ca39e68 MS |
4412 | * |
4413 | * Apart from cpuset, we also have memory policy mechanism that | |
4414 | * also determines from which node the kernel will allocate memory | |
4415 | * in a NUMA system. So similar to cpuset, we also should consider | |
4416 | * the memory policy of the current task. Similar to the description | |
4417 | * above. | |
fc1b8a73 MG |
4418 | */ |
4419 | if (delta > 0) { | |
a5516438 | 4420 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
4421 | goto out; |
4422 | ||
8ca39e68 | 4423 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 4424 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
4425 | goto out; |
4426 | } | |
4427 | } | |
4428 | ||
4429 | ret = 0; | |
4430 | if (delta < 0) | |
a5516438 | 4431 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
4432 | |
4433 | out: | |
db71ef79 | 4434 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4435 | return ret; |
4436 | } | |
4437 | ||
84afd99b AW |
4438 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
4439 | { | |
f522c3ac | 4440 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
4441 | |
4442 | /* | |
4443 | * This new VMA should share its siblings reservation map if present. | |
4444 | * The VMA will only ever have a valid reservation map pointer where | |
4445 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 4446 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
4447 | * after this open call completes. It is therefore safe to take a |
4448 | * new reference here without additional locking. | |
4449 | */ | |
09a26e83 MK |
4450 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
4451 | resv_map_dup_hugetlb_cgroup_uncharge_info(resv); | |
f522c3ac | 4452 | kref_get(&resv->refs); |
09a26e83 | 4453 | } |
84afd99b AW |
4454 | } |
4455 | ||
a1e78772 MG |
4456 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
4457 | { | |
a5516438 | 4458 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 4459 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 4460 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 4461 | unsigned long reserve, start, end; |
1c5ecae3 | 4462 | long gbl_reserve; |
84afd99b | 4463 | |
4e35f483 JK |
4464 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4465 | return; | |
84afd99b | 4466 | |
4e35f483 JK |
4467 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
4468 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 4469 | |
4e35f483 | 4470 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 4471 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 4472 | if (reserve) { |
1c5ecae3 MK |
4473 | /* |
4474 | * Decrement reserve counts. The global reserve count may be | |
4475 | * adjusted if the subpool has a minimum size. | |
4476 | */ | |
4477 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
4478 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 4479 | } |
e9fe92ae MA |
4480 | |
4481 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
4482 | } |
4483 | ||
31383c68 DW |
4484 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
4485 | { | |
4486 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
4487 | return -EINVAL; | |
4488 | return 0; | |
4489 | } | |
4490 | ||
05ea8860 DW |
4491 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
4492 | { | |
aca78307 | 4493 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
4494 | } |
4495 | ||
1da177e4 LT |
4496 | /* |
4497 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
4498 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 4499 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
4500 | * this far. |
4501 | */ | |
b3ec9f33 | 4502 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
4503 | { |
4504 | BUG(); | |
d0217ac0 | 4505 | return 0; |
1da177e4 LT |
4506 | } |
4507 | ||
eec3636a JC |
4508 | /* |
4509 | * When a new function is introduced to vm_operations_struct and added | |
4510 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
4511 | * This is because under System V memory model, mappings created via | |
4512 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
4513 | * their original vm_ops are overwritten with shm_vm_ops. | |
4514 | */ | |
f0f37e2f | 4515 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 4516 | .fault = hugetlb_vm_op_fault, |
84afd99b | 4517 | .open = hugetlb_vm_op_open, |
a1e78772 | 4518 | .close = hugetlb_vm_op_close, |
dd3b614f | 4519 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 4520 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
4521 | }; |
4522 | ||
1e8f889b DG |
4523 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
4524 | int writable) | |
63551ae0 DG |
4525 | { |
4526 | pte_t entry; | |
79c1c594 | 4527 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
63551ae0 | 4528 | |
1e8f889b | 4529 | if (writable) { |
106c992a GS |
4530 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
4531 | vma->vm_page_prot))); | |
63551ae0 | 4532 | } else { |
106c992a GS |
4533 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
4534 | vma->vm_page_prot)); | |
63551ae0 DG |
4535 | } |
4536 | entry = pte_mkyoung(entry); | |
4537 | entry = pte_mkhuge(entry); | |
79c1c594 | 4538 | entry = arch_make_huge_pte(entry, shift, vma->vm_flags); |
63551ae0 DG |
4539 | |
4540 | return entry; | |
4541 | } | |
4542 | ||
1e8f889b DG |
4543 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
4544 | unsigned long address, pte_t *ptep) | |
4545 | { | |
4546 | pte_t entry; | |
4547 | ||
106c992a | 4548 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 4549 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 4550 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
4551 | } |
4552 | ||
d5ed7444 | 4553 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
4554 | { |
4555 | swp_entry_t swp; | |
4556 | ||
4557 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 4558 | return false; |
4a705fef | 4559 | swp = pte_to_swp_entry(pte); |
d79d176a | 4560 | if (is_migration_entry(swp)) |
d5ed7444 | 4561 | return true; |
4a705fef | 4562 | else |
d5ed7444 | 4563 | return false; |
4a705fef NH |
4564 | } |
4565 | ||
3e5c3600 | 4566 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
4567 | { |
4568 | swp_entry_t swp; | |
4569 | ||
4570 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 4571 | return false; |
4a705fef | 4572 | swp = pte_to_swp_entry(pte); |
d79d176a | 4573 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 4574 | return true; |
4a705fef | 4575 | else |
3e5c3600 | 4576 | return false; |
4a705fef | 4577 | } |
1e8f889b | 4578 | |
4eae4efa PX |
4579 | static void |
4580 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
4581 | struct page *new_page) | |
4582 | { | |
4583 | __SetPageUptodate(new_page); | |
4584 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); | |
4585 | hugepage_add_new_anon_rmap(new_page, vma, addr); | |
4586 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); | |
4587 | ClearHPageRestoreReserve(new_page); | |
4588 | SetHPageMigratable(new_page); | |
4589 | } | |
4590 | ||
63551ae0 DG |
4591 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
4592 | struct vm_area_struct *vma) | |
4593 | { | |
5e41540c | 4594 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 4595 | struct page *ptepage; |
1c59827d | 4596 | unsigned long addr; |
ca6eb14d | 4597 | bool cow = is_cow_mapping(vma->vm_flags); |
a5516438 AK |
4598 | struct hstate *h = hstate_vma(vma); |
4599 | unsigned long sz = huge_page_size(h); | |
4eae4efa | 4600 | unsigned long npages = pages_per_huge_page(h); |
c0d0381a | 4601 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 4602 | struct mmu_notifier_range range; |
e8569dd2 | 4603 | int ret = 0; |
1e8f889b | 4604 | |
ac46d4f3 | 4605 | if (cow) { |
7269f999 | 4606 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 4607 | vma->vm_start, |
ac46d4f3 JG |
4608 | vma->vm_end); |
4609 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
4610 | } else { |
4611 | /* | |
4612 | * For shared mappings i_mmap_rwsem must be held to call | |
4613 | * huge_pte_alloc, otherwise the returned ptep could go | |
4614 | * away if part of a shared pmd and another thread calls | |
4615 | * huge_pmd_unshare. | |
4616 | */ | |
4617 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 4618 | } |
e8569dd2 | 4619 | |
a5516438 | 4620 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 4621 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 4622 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
4623 | if (!src_pte) |
4624 | continue; | |
aec44e0f | 4625 | dst_pte = huge_pte_alloc(dst, vma, addr, sz); |
e8569dd2 AS |
4626 | if (!dst_pte) { |
4627 | ret = -ENOMEM; | |
4628 | break; | |
4629 | } | |
c5c99429 | 4630 | |
5e41540c MK |
4631 | /* |
4632 | * If the pagetables are shared don't copy or take references. | |
4633 | * dst_pte == src_pte is the common case of src/dest sharing. | |
4634 | * | |
4635 | * However, src could have 'unshared' and dst shares with | |
4636 | * another vma. If dst_pte !none, this implies sharing. | |
4637 | * Check here before taking page table lock, and once again | |
4638 | * after taking the lock below. | |
4639 | */ | |
4640 | dst_entry = huge_ptep_get(dst_pte); | |
4641 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
4642 | continue; |
4643 | ||
cb900f41 KS |
4644 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
4645 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4646 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 4647 | entry = huge_ptep_get(src_pte); |
5e41540c | 4648 | dst_entry = huge_ptep_get(dst_pte); |
4eae4efa | 4649 | again: |
5e41540c MK |
4650 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { |
4651 | /* | |
4652 | * Skip if src entry none. Also, skip in the | |
4653 | * unlikely case dst entry !none as this implies | |
4654 | * sharing with another vma. | |
4655 | */ | |
4a705fef NH |
4656 | ; |
4657 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
4658 | is_hugetlb_entry_hwpoisoned(entry))) { | |
4659 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
4660 | ||
4dd845b5 | 4661 | if (is_writable_migration_entry(swp_entry) && cow) { |
4a705fef NH |
4662 | /* |
4663 | * COW mappings require pages in both | |
4664 | * parent and child to be set to read. | |
4665 | */ | |
4dd845b5 AP |
4666 | swp_entry = make_readable_migration_entry( |
4667 | swp_offset(swp_entry)); | |
4a705fef | 4668 | entry = swp_entry_to_pte(swp_entry); |
e5251fd4 PA |
4669 | set_huge_swap_pte_at(src, addr, src_pte, |
4670 | entry, sz); | |
4a705fef | 4671 | } |
e5251fd4 | 4672 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 4673 | } else { |
4eae4efa PX |
4674 | entry = huge_ptep_get(src_pte); |
4675 | ptepage = pte_page(entry); | |
4676 | get_page(ptepage); | |
4677 | ||
4678 | /* | |
4679 | * This is a rare case where we see pinned hugetlb | |
4680 | * pages while they're prone to COW. We need to do the | |
4681 | * COW earlier during fork. | |
4682 | * | |
4683 | * When pre-allocating the page or copying data, we | |
4684 | * need to be without the pgtable locks since we could | |
4685 | * sleep during the process. | |
4686 | */ | |
4687 | if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { | |
4688 | pte_t src_pte_old = entry; | |
4689 | struct page *new; | |
4690 | ||
4691 | spin_unlock(src_ptl); | |
4692 | spin_unlock(dst_ptl); | |
4693 | /* Do not use reserve as it's private owned */ | |
4694 | new = alloc_huge_page(vma, addr, 1); | |
4695 | if (IS_ERR(new)) { | |
4696 | put_page(ptepage); | |
4697 | ret = PTR_ERR(new); | |
4698 | break; | |
4699 | } | |
4700 | copy_user_huge_page(new, ptepage, addr, vma, | |
4701 | npages); | |
4702 | put_page(ptepage); | |
4703 | ||
4704 | /* Install the new huge page if src pte stable */ | |
4705 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
4706 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4707 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4708 | entry = huge_ptep_get(src_pte); | |
4709 | if (!pte_same(src_pte_old, entry)) { | |
846be085 MK |
4710 | restore_reserve_on_error(h, vma, addr, |
4711 | new); | |
4eae4efa PX |
4712 | put_page(new); |
4713 | /* dst_entry won't change as in child */ | |
4714 | goto again; | |
4715 | } | |
4716 | hugetlb_install_page(vma, dst_pte, addr, new); | |
4717 | spin_unlock(src_ptl); | |
4718 | spin_unlock(dst_ptl); | |
4719 | continue; | |
4720 | } | |
4721 | ||
34ee645e | 4722 | if (cow) { |
0f10851e JG |
4723 | /* |
4724 | * No need to notify as we are downgrading page | |
4725 | * table protection not changing it to point | |
4726 | * to a new page. | |
4727 | * | |
ad56b738 | 4728 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4729 | */ |
7f2e9525 | 4730 | huge_ptep_set_wrprotect(src, addr, src_pte); |
84894e1c | 4731 | entry = huge_pte_wrprotect(entry); |
34ee645e | 4732 | } |
4eae4efa | 4733 | |
53f9263b | 4734 | page_dup_rmap(ptepage, true); |
1c59827d | 4735 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 4736 | hugetlb_count_add(npages, dst); |
1c59827d | 4737 | } |
cb900f41 KS |
4738 | spin_unlock(src_ptl); |
4739 | spin_unlock(dst_ptl); | |
63551ae0 | 4740 | } |
63551ae0 | 4741 | |
e8569dd2 | 4742 | if (cow) |
ac46d4f3 | 4743 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
4744 | else |
4745 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
4746 | |
4747 | return ret; | |
63551ae0 DG |
4748 | } |
4749 | ||
550a7d60 MA |
4750 | static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, |
4751 | unsigned long new_addr, pte_t *src_pte) | |
4752 | { | |
4753 | struct hstate *h = hstate_vma(vma); | |
4754 | struct mm_struct *mm = vma->vm_mm; | |
4755 | pte_t *dst_pte, pte; | |
4756 | spinlock_t *src_ptl, *dst_ptl; | |
4757 | ||
4758 | dst_pte = huge_pte_offset(mm, new_addr, huge_page_size(h)); | |
4759 | dst_ptl = huge_pte_lock(h, mm, dst_pte); | |
4760 | src_ptl = huge_pte_lockptr(h, mm, src_pte); | |
4761 | ||
4762 | /* | |
4763 | * We don't have to worry about the ordering of src and dst ptlocks | |
4764 | * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. | |
4765 | */ | |
4766 | if (src_ptl != dst_ptl) | |
4767 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4768 | ||
4769 | pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); | |
4770 | set_huge_pte_at(mm, new_addr, dst_pte, pte); | |
4771 | ||
4772 | if (src_ptl != dst_ptl) | |
4773 | spin_unlock(src_ptl); | |
4774 | spin_unlock(dst_ptl); | |
4775 | } | |
4776 | ||
4777 | int move_hugetlb_page_tables(struct vm_area_struct *vma, | |
4778 | struct vm_area_struct *new_vma, | |
4779 | unsigned long old_addr, unsigned long new_addr, | |
4780 | unsigned long len) | |
4781 | { | |
4782 | struct hstate *h = hstate_vma(vma); | |
4783 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4784 | unsigned long sz = huge_page_size(h); | |
4785 | struct mm_struct *mm = vma->vm_mm; | |
4786 | unsigned long old_end = old_addr + len; | |
4787 | unsigned long old_addr_copy; | |
4788 | pte_t *src_pte, *dst_pte; | |
4789 | struct mmu_notifier_range range; | |
4790 | ||
4791 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, | |
4792 | old_end); | |
4793 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
4794 | mmu_notifier_invalidate_range_start(&range); | |
4795 | /* Prevent race with file truncation */ | |
4796 | i_mmap_lock_write(mapping); | |
4797 | for (; old_addr < old_end; old_addr += sz, new_addr += sz) { | |
4798 | src_pte = huge_pte_offset(mm, old_addr, sz); | |
4799 | if (!src_pte) | |
4800 | continue; | |
4801 | if (huge_pte_none(huge_ptep_get(src_pte))) | |
4802 | continue; | |
4803 | ||
4804 | /* old_addr arg to huge_pmd_unshare() is a pointer and so the | |
4805 | * arg may be modified. Pass a copy instead to preserve the | |
4806 | * value in old_addr. | |
4807 | */ | |
4808 | old_addr_copy = old_addr; | |
4809 | ||
4810 | if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte)) | |
4811 | continue; | |
4812 | ||
4813 | dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); | |
4814 | if (!dst_pte) | |
4815 | break; | |
4816 | ||
4817 | move_huge_pte(vma, old_addr, new_addr, src_pte); | |
4818 | } | |
4819 | i_mmap_unlock_write(mapping); | |
4820 | flush_tlb_range(vma, old_end - len, old_end); | |
4821 | mmu_notifier_invalidate_range_end(&range); | |
4822 | ||
4823 | return len + old_addr - old_end; | |
4824 | } | |
4825 | ||
73c54763 PX |
4826 | static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
4827 | unsigned long start, unsigned long end, | |
4828 | struct page *ref_page) | |
63551ae0 DG |
4829 | { |
4830 | struct mm_struct *mm = vma->vm_mm; | |
4831 | unsigned long address; | |
c7546f8f | 4832 | pte_t *ptep; |
63551ae0 | 4833 | pte_t pte; |
cb900f41 | 4834 | spinlock_t *ptl; |
63551ae0 | 4835 | struct page *page; |
a5516438 AK |
4836 | struct hstate *h = hstate_vma(vma); |
4837 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 4838 | struct mmu_notifier_range range; |
a5516438 | 4839 | |
63551ae0 | 4840 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
4841 | BUG_ON(start & ~huge_page_mask(h)); |
4842 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 4843 | |
07e32661 AK |
4844 | /* |
4845 | * This is a hugetlb vma, all the pte entries should point | |
4846 | * to huge page. | |
4847 | */ | |
ed6a7935 | 4848 | tlb_change_page_size(tlb, sz); |
24669e58 | 4849 | tlb_start_vma(tlb, vma); |
dff11abe MK |
4850 | |
4851 | /* | |
4852 | * If sharing possible, alert mmu notifiers of worst case. | |
4853 | */ | |
6f4f13e8 JG |
4854 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
4855 | end); | |
ac46d4f3 JG |
4856 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
4857 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 4858 | address = start; |
569f48b8 | 4859 | for (; address < end; address += sz) { |
7868a208 | 4860 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 4861 | if (!ptep) |
c7546f8f DG |
4862 | continue; |
4863 | ||
cb900f41 | 4864 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 4865 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
31d49da5 | 4866 | spin_unlock(ptl); |
dff11abe MK |
4867 | /* |
4868 | * We just unmapped a page of PMDs by clearing a PUD. | |
4869 | * The caller's TLB flush range should cover this area. | |
4870 | */ | |
31d49da5 AK |
4871 | continue; |
4872 | } | |
39dde65c | 4873 | |
6629326b | 4874 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
4875 | if (huge_pte_none(pte)) { |
4876 | spin_unlock(ptl); | |
4877 | continue; | |
4878 | } | |
6629326b HD |
4879 | |
4880 | /* | |
9fbc1f63 NH |
4881 | * Migrating hugepage or HWPoisoned hugepage is already |
4882 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 4883 | */ |
9fbc1f63 | 4884 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 4885 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
4886 | spin_unlock(ptl); |
4887 | continue; | |
8c4894c6 | 4888 | } |
6629326b HD |
4889 | |
4890 | page = pte_page(pte); | |
04f2cbe3 MG |
4891 | /* |
4892 | * If a reference page is supplied, it is because a specific | |
4893 | * page is being unmapped, not a range. Ensure the page we | |
4894 | * are about to unmap is the actual page of interest. | |
4895 | */ | |
4896 | if (ref_page) { | |
31d49da5 AK |
4897 | if (page != ref_page) { |
4898 | spin_unlock(ptl); | |
4899 | continue; | |
4900 | } | |
04f2cbe3 MG |
4901 | /* |
4902 | * Mark the VMA as having unmapped its page so that | |
4903 | * future faults in this VMA will fail rather than | |
4904 | * looking like data was lost | |
4905 | */ | |
4906 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
4907 | } | |
4908 | ||
c7546f8f | 4909 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 4910 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 4911 | if (huge_pte_dirty(pte)) |
6649a386 | 4912 | set_page_dirty(page); |
9e81130b | 4913 | |
5d317b2b | 4914 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 4915 | page_remove_rmap(page, true); |
31d49da5 | 4916 | |
cb900f41 | 4917 | spin_unlock(ptl); |
e77b0852 | 4918 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
4919 | /* |
4920 | * Bail out after unmapping reference page if supplied | |
4921 | */ | |
4922 | if (ref_page) | |
4923 | break; | |
fe1668ae | 4924 | } |
ac46d4f3 | 4925 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 4926 | tlb_end_vma(tlb, vma); |
1da177e4 | 4927 | } |
63551ae0 | 4928 | |
d833352a MG |
4929 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
4930 | struct vm_area_struct *vma, unsigned long start, | |
4931 | unsigned long end, struct page *ref_page) | |
4932 | { | |
4933 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
4934 | ||
4935 | /* | |
4936 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
4937 | * test will fail on a vma being torn down, and not grab a page table | |
4938 | * on its way out. We're lucky that the flag has such an appropriate | |
4939 | * name, and can in fact be safely cleared here. We could clear it | |
4940 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 4941 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 4942 | * because in the context this is called, the VMA is about to be |
c8c06efa | 4943 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
4944 | */ |
4945 | vma->vm_flags &= ~VM_MAYSHARE; | |
4946 | } | |
4947 | ||
502717f4 | 4948 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 4949 | unsigned long end, struct page *ref_page) |
502717f4 | 4950 | { |
24669e58 | 4951 | struct mmu_gather tlb; |
dff11abe | 4952 | |
a72afd87 | 4953 | tlb_gather_mmu(&tlb, vma->vm_mm); |
24669e58 | 4954 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
ae8eba8b | 4955 | tlb_finish_mmu(&tlb); |
502717f4 CK |
4956 | } |
4957 | ||
04f2cbe3 MG |
4958 | /* |
4959 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 4960 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
4961 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
4962 | * same region. | |
4963 | */ | |
2f4612af DB |
4964 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
4965 | struct page *page, unsigned long address) | |
04f2cbe3 | 4966 | { |
7526674d | 4967 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
4968 | struct vm_area_struct *iter_vma; |
4969 | struct address_space *mapping; | |
04f2cbe3 MG |
4970 | pgoff_t pgoff; |
4971 | ||
4972 | /* | |
4973 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
4974 | * from page cache lookup which is in HPAGE_SIZE units. | |
4975 | */ | |
7526674d | 4976 | address = address & huge_page_mask(h); |
36e4f20a MH |
4977 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
4978 | vma->vm_pgoff; | |
93c76a3d | 4979 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 4980 | |
4eb2b1dc MG |
4981 | /* |
4982 | * Take the mapping lock for the duration of the table walk. As | |
4983 | * this mapping should be shared between all the VMAs, | |
4984 | * __unmap_hugepage_range() is called as the lock is already held | |
4985 | */ | |
83cde9e8 | 4986 | i_mmap_lock_write(mapping); |
6b2dbba8 | 4987 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
4988 | /* Do not unmap the current VMA */ |
4989 | if (iter_vma == vma) | |
4990 | continue; | |
4991 | ||
2f84a899 MG |
4992 | /* |
4993 | * Shared VMAs have their own reserves and do not affect | |
4994 | * MAP_PRIVATE accounting but it is possible that a shared | |
4995 | * VMA is using the same page so check and skip such VMAs. | |
4996 | */ | |
4997 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
4998 | continue; | |
4999 | ||
04f2cbe3 MG |
5000 | /* |
5001 | * Unmap the page from other VMAs without their own reserves. | |
5002 | * They get marked to be SIGKILLed if they fault in these | |
5003 | * areas. This is because a future no-page fault on this VMA | |
5004 | * could insert a zeroed page instead of the data existing | |
5005 | * from the time of fork. This would look like data corruption | |
5006 | */ | |
5007 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
5008 | unmap_hugepage_range(iter_vma, address, |
5009 | address + huge_page_size(h), page); | |
04f2cbe3 | 5010 | } |
83cde9e8 | 5011 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
5012 | } |
5013 | ||
0fe6e20b NH |
5014 | /* |
5015 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
5016 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
5017 | * cannot race with other handlers or page migration. | |
5018 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 5019 | */ |
2b740303 | 5020 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 5021 | unsigned long address, pte_t *ptep, |
3999f52e | 5022 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 5023 | { |
3999f52e | 5024 | pte_t pte; |
a5516438 | 5025 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 5026 | struct page *old_page, *new_page; |
2b740303 SJ |
5027 | int outside_reserve = 0; |
5028 | vm_fault_t ret = 0; | |
974e6d66 | 5029 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 5030 | struct mmu_notifier_range range; |
1e8f889b | 5031 | |
3999f52e | 5032 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
5033 | old_page = pte_page(pte); |
5034 | ||
04f2cbe3 | 5035 | retry_avoidcopy: |
1e8f889b DG |
5036 | /* If no-one else is actually using this page, avoid the copy |
5037 | * and just make the page writable */ | |
37a2140d | 5038 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 5039 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 5040 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 5041 | return 0; |
1e8f889b DG |
5042 | } |
5043 | ||
04f2cbe3 MG |
5044 | /* |
5045 | * If the process that created a MAP_PRIVATE mapping is about to | |
5046 | * perform a COW due to a shared page count, attempt to satisfy | |
5047 | * the allocation without using the existing reserves. The pagecache | |
5048 | * page is used to determine if the reserve at this address was | |
5049 | * consumed or not. If reserves were used, a partial faulted mapping | |
5050 | * at the time of fork() could consume its reserves on COW instead | |
5051 | * of the full address range. | |
5052 | */ | |
5944d011 | 5053 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
5054 | old_page != pagecache_page) |
5055 | outside_reserve = 1; | |
5056 | ||
09cbfeaf | 5057 | get_page(old_page); |
b76c8cfb | 5058 | |
ad4404a2 DB |
5059 | /* |
5060 | * Drop page table lock as buddy allocator may be called. It will | |
5061 | * be acquired again before returning to the caller, as expected. | |
5062 | */ | |
cb900f41 | 5063 | spin_unlock(ptl); |
5b7a1d40 | 5064 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 5065 | |
2fc39cec | 5066 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
5067 | /* |
5068 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
5069 | * it is due to references held by a child and an insufficient | |
5070 | * huge page pool. To guarantee the original mappers | |
5071 | * reliability, unmap the page from child processes. The child | |
5072 | * may get SIGKILLed if it later faults. | |
5073 | */ | |
5074 | if (outside_reserve) { | |
e7dd91c4 MK |
5075 | struct address_space *mapping = vma->vm_file->f_mapping; |
5076 | pgoff_t idx; | |
5077 | u32 hash; | |
5078 | ||
09cbfeaf | 5079 | put_page(old_page); |
04f2cbe3 | 5080 | BUG_ON(huge_pte_none(pte)); |
e7dd91c4 MK |
5081 | /* |
5082 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
5083 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
5084 | * in write mode. Dropping i_mmap_rwsem in read mode | |
5085 | * here is OK as COW mappings do not interact with | |
5086 | * PMD sharing. | |
5087 | * | |
5088 | * Reacquire both after unmap operation. | |
5089 | */ | |
5090 | idx = vma_hugecache_offset(h, vma, haddr); | |
5091 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5092 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5093 | i_mmap_unlock_read(mapping); | |
5094 | ||
5b7a1d40 | 5095 | unmap_ref_private(mm, vma, old_page, haddr); |
e7dd91c4 MK |
5096 | |
5097 | i_mmap_lock_read(mapping); | |
5098 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
2f4612af | 5099 | spin_lock(ptl); |
5b7a1d40 | 5100 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
5101 | if (likely(ptep && |
5102 | pte_same(huge_ptep_get(ptep), pte))) | |
5103 | goto retry_avoidcopy; | |
5104 | /* | |
5105 | * race occurs while re-acquiring page table | |
5106 | * lock, and our job is done. | |
5107 | */ | |
5108 | return 0; | |
04f2cbe3 MG |
5109 | } |
5110 | ||
2b740303 | 5111 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 5112 | goto out_release_old; |
1e8f889b DG |
5113 | } |
5114 | ||
0fe6e20b NH |
5115 | /* |
5116 | * When the original hugepage is shared one, it does not have | |
5117 | * anon_vma prepared. | |
5118 | */ | |
44e2aa93 | 5119 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
5120 | ret = VM_FAULT_OOM; |
5121 | goto out_release_all; | |
44e2aa93 | 5122 | } |
0fe6e20b | 5123 | |
974e6d66 | 5124 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 5125 | pages_per_huge_page(h)); |
0ed361de | 5126 | __SetPageUptodate(new_page); |
1e8f889b | 5127 | |
7269f999 | 5128 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 5129 | haddr + huge_page_size(h)); |
ac46d4f3 | 5130 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 5131 | |
b76c8cfb | 5132 | /* |
cb900f41 | 5133 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
5134 | * before the page tables are altered |
5135 | */ | |
cb900f41 | 5136 | spin_lock(ptl); |
5b7a1d40 | 5137 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 5138 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 5139 | ClearHPageRestoreReserve(new_page); |
07443a85 | 5140 | |
1e8f889b | 5141 | /* Break COW */ |
5b7a1d40 | 5142 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 5143 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 5144 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 5145 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 5146 | page_remove_rmap(old_page, true); |
5b7a1d40 | 5147 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
8f251a3d | 5148 | SetHPageMigratable(new_page); |
1e8f889b DG |
5149 | /* Make the old page be freed below */ |
5150 | new_page = old_page; | |
5151 | } | |
cb900f41 | 5152 | spin_unlock(ptl); |
ac46d4f3 | 5153 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 5154 | out_release_all: |
c7b1850d MK |
5155 | /* No restore in case of successful pagetable update (Break COW) */ |
5156 | if (new_page != old_page) | |
5157 | restore_reserve_on_error(h, vma, haddr, new_page); | |
09cbfeaf | 5158 | put_page(new_page); |
ad4404a2 | 5159 | out_release_old: |
09cbfeaf | 5160 | put_page(old_page); |
8312034f | 5161 | |
ad4404a2 DB |
5162 | spin_lock(ptl); /* Caller expects lock to be held */ |
5163 | return ret; | |
1e8f889b DG |
5164 | } |
5165 | ||
04f2cbe3 | 5166 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
5167 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
5168 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
5169 | { |
5170 | struct address_space *mapping; | |
e7c4b0bf | 5171 | pgoff_t idx; |
04f2cbe3 MG |
5172 | |
5173 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 5174 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
5175 | |
5176 | return find_lock_page(mapping, idx); | |
5177 | } | |
5178 | ||
3ae77f43 HD |
5179 | /* |
5180 | * Return whether there is a pagecache page to back given address within VMA. | |
5181 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
5182 | */ | |
5183 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
5184 | struct vm_area_struct *vma, unsigned long address) |
5185 | { | |
5186 | struct address_space *mapping; | |
5187 | pgoff_t idx; | |
5188 | struct page *page; | |
5189 | ||
5190 | mapping = vma->vm_file->f_mapping; | |
5191 | idx = vma_hugecache_offset(h, vma, address); | |
5192 | ||
5193 | page = find_get_page(mapping, idx); | |
5194 | if (page) | |
5195 | put_page(page); | |
5196 | return page != NULL; | |
5197 | } | |
5198 | ||
ab76ad54 MK |
5199 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
5200 | pgoff_t idx) | |
5201 | { | |
5202 | struct inode *inode = mapping->host; | |
5203 | struct hstate *h = hstate_inode(inode); | |
5204 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
5205 | ||
5206 | if (err) | |
5207 | return err; | |
d6995da3 | 5208 | ClearHPageRestoreReserve(page); |
ab76ad54 | 5209 | |
22146c3c MK |
5210 | /* |
5211 | * set page dirty so that it will not be removed from cache/file | |
5212 | * by non-hugetlbfs specific code paths. | |
5213 | */ | |
5214 | set_page_dirty(page); | |
5215 | ||
ab76ad54 MK |
5216 | spin_lock(&inode->i_lock); |
5217 | inode->i_blocks += blocks_per_huge_page(h); | |
5218 | spin_unlock(&inode->i_lock); | |
5219 | return 0; | |
5220 | } | |
5221 | ||
7677f7fd AR |
5222 | static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, |
5223 | struct address_space *mapping, | |
5224 | pgoff_t idx, | |
5225 | unsigned int flags, | |
5226 | unsigned long haddr, | |
5227 | unsigned long reason) | |
5228 | { | |
5229 | vm_fault_t ret; | |
5230 | u32 hash; | |
5231 | struct vm_fault vmf = { | |
5232 | .vma = vma, | |
5233 | .address = haddr, | |
5234 | .flags = flags, | |
5235 | ||
5236 | /* | |
5237 | * Hard to debug if it ends up being | |
5238 | * used by a callee that assumes | |
5239 | * something about the other | |
5240 | * uninitialized fields... same as in | |
5241 | * memory.c | |
5242 | */ | |
5243 | }; | |
5244 | ||
5245 | /* | |
5246 | * hugetlb_fault_mutex and i_mmap_rwsem must be | |
5247 | * dropped before handling userfault. Reacquire | |
5248 | * after handling fault to make calling code simpler. | |
5249 | */ | |
5250 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5251 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5252 | i_mmap_unlock_read(mapping); | |
5253 | ret = handle_userfault(&vmf, reason); | |
5254 | i_mmap_lock_read(mapping); | |
5255 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
5256 | ||
5257 | return ret; | |
5258 | } | |
5259 | ||
2b740303 SJ |
5260 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
5261 | struct vm_area_struct *vma, | |
5262 | struct address_space *mapping, pgoff_t idx, | |
5263 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 5264 | { |
a5516438 | 5265 | struct hstate *h = hstate_vma(vma); |
2b740303 | 5266 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 5267 | int anon_rmap = 0; |
4c887265 | 5268 | unsigned long size; |
4c887265 | 5269 | struct page *page; |
1e8f889b | 5270 | pte_t new_pte; |
cb900f41 | 5271 | spinlock_t *ptl; |
285b8dca | 5272 | unsigned long haddr = address & huge_page_mask(h); |
c7b1850d | 5273 | bool new_page, new_pagecache_page = false; |
4c887265 | 5274 | |
04f2cbe3 MG |
5275 | /* |
5276 | * Currently, we are forced to kill the process in the event the | |
5277 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 5278 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
5279 | */ |
5280 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 5281 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 5282 | current->pid); |
04f2cbe3 MG |
5283 | return ret; |
5284 | } | |
5285 | ||
4c887265 | 5286 | /* |
87bf91d3 MK |
5287 | * We can not race with truncation due to holding i_mmap_rwsem. |
5288 | * i_size is modified when holding i_mmap_rwsem, so check here | |
5289 | * once for faults beyond end of file. | |
4c887265 | 5290 | */ |
87bf91d3 MK |
5291 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5292 | if (idx >= size) | |
5293 | goto out; | |
5294 | ||
6bda666a | 5295 | retry: |
c7b1850d | 5296 | new_page = false; |
6bda666a CL |
5297 | page = find_lock_page(mapping, idx); |
5298 | if (!page) { | |
7677f7fd | 5299 | /* Check for page in userfault range */ |
1a1aad8a | 5300 | if (userfaultfd_missing(vma)) { |
7677f7fd AR |
5301 | ret = hugetlb_handle_userfault(vma, mapping, idx, |
5302 | flags, haddr, | |
5303 | VM_UFFD_MISSING); | |
1a1aad8a MK |
5304 | goto out; |
5305 | } | |
5306 | ||
285b8dca | 5307 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 5308 | if (IS_ERR(page)) { |
4643d67e MK |
5309 | /* |
5310 | * Returning error will result in faulting task being | |
5311 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
5312 | * tasks from racing to fault in the same page which | |
5313 | * could result in false unable to allocate errors. | |
5314 | * Page migration does not take the fault mutex, but | |
5315 | * does a clear then write of pte's under page table | |
5316 | * lock. Page fault code could race with migration, | |
5317 | * notice the clear pte and try to allocate a page | |
5318 | * here. Before returning error, get ptl and make | |
5319 | * sure there really is no pte entry. | |
5320 | */ | |
5321 | ptl = huge_pte_lock(h, mm, ptep); | |
d83e6c8a ML |
5322 | ret = 0; |
5323 | if (huge_pte_none(huge_ptep_get(ptep))) | |
5324 | ret = vmf_error(PTR_ERR(page)); | |
4643d67e | 5325 | spin_unlock(ptl); |
6bda666a CL |
5326 | goto out; |
5327 | } | |
47ad8475 | 5328 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 5329 | __SetPageUptodate(page); |
cb6acd01 | 5330 | new_page = true; |
ac9b9c66 | 5331 | |
f83a275d | 5332 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 5333 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
5334 | if (err) { |
5335 | put_page(page); | |
6bda666a CL |
5336 | if (err == -EEXIST) |
5337 | goto retry; | |
5338 | goto out; | |
5339 | } | |
c7b1850d | 5340 | new_pagecache_page = true; |
23be7468 | 5341 | } else { |
6bda666a | 5342 | lock_page(page); |
0fe6e20b NH |
5343 | if (unlikely(anon_vma_prepare(vma))) { |
5344 | ret = VM_FAULT_OOM; | |
5345 | goto backout_unlocked; | |
5346 | } | |
409eb8c2 | 5347 | anon_rmap = 1; |
23be7468 | 5348 | } |
0fe6e20b | 5349 | } else { |
998b4382 NH |
5350 | /* |
5351 | * If memory error occurs between mmap() and fault, some process | |
5352 | * don't have hwpoisoned swap entry for errored virtual address. | |
5353 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
5354 | */ | |
5355 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 5356 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5357 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
5358 | goto backout_unlocked; |
5359 | } | |
7677f7fd AR |
5360 | |
5361 | /* Check for page in userfault range. */ | |
5362 | if (userfaultfd_minor(vma)) { | |
5363 | unlock_page(page); | |
5364 | put_page(page); | |
5365 | ret = hugetlb_handle_userfault(vma, mapping, idx, | |
5366 | flags, haddr, | |
5367 | VM_UFFD_MINOR); | |
5368 | goto out; | |
5369 | } | |
6bda666a | 5370 | } |
1e8f889b | 5371 | |
57303d80 AW |
5372 | /* |
5373 | * If we are going to COW a private mapping later, we examine the | |
5374 | * pending reservations for this page now. This will ensure that | |
5375 | * any allocations necessary to record that reservation occur outside | |
5376 | * the spinlock. | |
5377 | */ | |
5e911373 | 5378 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 5379 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
5380 | ret = VM_FAULT_OOM; |
5381 | goto backout_unlocked; | |
5382 | } | |
5e911373 | 5383 | /* Just decrements count, does not deallocate */ |
285b8dca | 5384 | vma_end_reservation(h, vma, haddr); |
5e911373 | 5385 | } |
57303d80 | 5386 | |
8bea8052 | 5387 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 5388 | ret = 0; |
7f2e9525 | 5389 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
5390 | goto backout; |
5391 | ||
07443a85 | 5392 | if (anon_rmap) { |
d6995da3 | 5393 | ClearHPageRestoreReserve(page); |
285b8dca | 5394 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 5395 | } else |
53f9263b | 5396 | page_dup_rmap(page, true); |
1e8f889b DG |
5397 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
5398 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 5399 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 5400 | |
5d317b2b | 5401 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 5402 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 5403 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 5404 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
5405 | } |
5406 | ||
cb900f41 | 5407 | spin_unlock(ptl); |
cb6acd01 MK |
5408 | |
5409 | /* | |
8f251a3d MK |
5410 | * Only set HPageMigratable in newly allocated pages. Existing pages |
5411 | * found in the pagecache may not have HPageMigratableset if they have | |
5412 | * been isolated for migration. | |
cb6acd01 MK |
5413 | */ |
5414 | if (new_page) | |
8f251a3d | 5415 | SetHPageMigratable(page); |
cb6acd01 | 5416 | |
4c887265 AL |
5417 | unlock_page(page); |
5418 | out: | |
ac9b9c66 | 5419 | return ret; |
4c887265 AL |
5420 | |
5421 | backout: | |
cb900f41 | 5422 | spin_unlock(ptl); |
2b26736c | 5423 | backout_unlocked: |
4c887265 | 5424 | unlock_page(page); |
c7b1850d MK |
5425 | /* restore reserve for newly allocated pages not in page cache */ |
5426 | if (new_page && !new_pagecache_page) | |
5427 | restore_reserve_on_error(h, vma, haddr, page); | |
4c887265 AL |
5428 | put_page(page); |
5429 | goto out; | |
ac9b9c66 HD |
5430 | } |
5431 | ||
8382d914 | 5432 | #ifdef CONFIG_SMP |
188b04a7 | 5433 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5434 | { |
5435 | unsigned long key[2]; | |
5436 | u32 hash; | |
5437 | ||
1b426bac MK |
5438 | key[0] = (unsigned long) mapping; |
5439 | key[1] = idx; | |
8382d914 | 5440 | |
55254636 | 5441 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
5442 | |
5443 | return hash & (num_fault_mutexes - 1); | |
5444 | } | |
5445 | #else | |
5446 | /* | |
6c26d310 | 5447 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
5448 | * return 0 and avoid the hashing overhead. |
5449 | */ | |
188b04a7 | 5450 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5451 | { |
5452 | return 0; | |
5453 | } | |
5454 | #endif | |
5455 | ||
2b740303 | 5456 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 5457 | unsigned long address, unsigned int flags) |
86e5216f | 5458 | { |
8382d914 | 5459 | pte_t *ptep, entry; |
cb900f41 | 5460 | spinlock_t *ptl; |
2b740303 | 5461 | vm_fault_t ret; |
8382d914 DB |
5462 | u32 hash; |
5463 | pgoff_t idx; | |
0fe6e20b | 5464 | struct page *page = NULL; |
57303d80 | 5465 | struct page *pagecache_page = NULL; |
a5516438 | 5466 | struct hstate *h = hstate_vma(vma); |
8382d914 | 5467 | struct address_space *mapping; |
0f792cf9 | 5468 | int need_wait_lock = 0; |
285b8dca | 5469 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 5470 | |
285b8dca | 5471 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 5472 | if (ptep) { |
c0d0381a MK |
5473 | /* |
5474 | * Since we hold no locks, ptep could be stale. That is | |
5475 | * OK as we are only making decisions based on content and | |
5476 | * not actually modifying content here. | |
5477 | */ | |
fd6a03ed | 5478 | entry = huge_ptep_get(ptep); |
290408d4 | 5479 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 5480 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
5481 | return 0; |
5482 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 5483 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5484 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
5485 | } |
5486 | ||
c0d0381a MK |
5487 | /* |
5488 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
5489 | * until finished with ptep. This serves two purposes: |
5490 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
5491 | * and making the ptep no longer valid. | |
5492 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
5493 | * |
5494 | * ptep could have already be assigned via huge_pte_offset. That | |
5495 | * is OK, as huge_pte_alloc will return the same value unless | |
5496 | * something has changed. | |
5497 | */ | |
8382d914 | 5498 | mapping = vma->vm_file->f_mapping; |
c0d0381a | 5499 | i_mmap_lock_read(mapping); |
aec44e0f | 5500 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); |
c0d0381a MK |
5501 | if (!ptep) { |
5502 | i_mmap_unlock_read(mapping); | |
5503 | return VM_FAULT_OOM; | |
5504 | } | |
8382d914 | 5505 | |
3935baa9 DG |
5506 | /* |
5507 | * Serialize hugepage allocation and instantiation, so that we don't | |
5508 | * get spurious allocation failures if two CPUs race to instantiate | |
5509 | * the same page in the page cache. | |
5510 | */ | |
c0d0381a | 5511 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 5512 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 5513 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 5514 | |
7f2e9525 GS |
5515 | entry = huge_ptep_get(ptep); |
5516 | if (huge_pte_none(entry)) { | |
8382d914 | 5517 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 5518 | goto out_mutex; |
3935baa9 | 5519 | } |
86e5216f | 5520 | |
83c54070 | 5521 | ret = 0; |
1e8f889b | 5522 | |
0f792cf9 NH |
5523 | /* |
5524 | * entry could be a migration/hwpoison entry at this point, so this | |
5525 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
5526 | * an active hugepage in pagecache. This goto expects the 2nd page |
5527 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
5528 | * properly handle it. | |
0f792cf9 NH |
5529 | */ |
5530 | if (!pte_present(entry)) | |
5531 | goto out_mutex; | |
5532 | ||
57303d80 AW |
5533 | /* |
5534 | * If we are going to COW the mapping later, we examine the pending | |
5535 | * reservations for this page now. This will ensure that any | |
5536 | * allocations necessary to record that reservation occur outside the | |
5537 | * spinlock. For private mappings, we also lookup the pagecache | |
5538 | * page now as it is used to determine if a reservation has been | |
5539 | * consumed. | |
5540 | */ | |
106c992a | 5541 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 5542 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 5543 | ret = VM_FAULT_OOM; |
b4d1d99f | 5544 | goto out_mutex; |
2b26736c | 5545 | } |
5e911373 | 5546 | /* Just decrements count, does not deallocate */ |
285b8dca | 5547 | vma_end_reservation(h, vma, haddr); |
57303d80 | 5548 | |
f83a275d | 5549 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 5550 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 5551 | vma, haddr); |
57303d80 AW |
5552 | } |
5553 | ||
0f792cf9 NH |
5554 | ptl = huge_pte_lock(h, mm, ptep); |
5555 | ||
5556 | /* Check for a racing update before calling hugetlb_cow */ | |
5557 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
5558 | goto out_ptl; | |
5559 | ||
56c9cfb1 NH |
5560 | /* |
5561 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
5562 | * pagecache_page, so here we need take the former one | |
5563 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
5564 | */ |
5565 | page = pte_page(entry); | |
5566 | if (page != pagecache_page) | |
0f792cf9 NH |
5567 | if (!trylock_page(page)) { |
5568 | need_wait_lock = 1; | |
5569 | goto out_ptl; | |
5570 | } | |
b4d1d99f | 5571 | |
0f792cf9 | 5572 | get_page(page); |
b4d1d99f | 5573 | |
788c7df4 | 5574 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 5575 | if (!huge_pte_write(entry)) { |
974e6d66 | 5576 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 5577 | pagecache_page, ptl); |
0f792cf9 | 5578 | goto out_put_page; |
b4d1d99f | 5579 | } |
106c992a | 5580 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
5581 | } |
5582 | entry = pte_mkyoung(entry); | |
285b8dca | 5583 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 5584 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 5585 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
5586 | out_put_page: |
5587 | if (page != pagecache_page) | |
5588 | unlock_page(page); | |
5589 | put_page(page); | |
cb900f41 KS |
5590 | out_ptl: |
5591 | spin_unlock(ptl); | |
57303d80 AW |
5592 | |
5593 | if (pagecache_page) { | |
5594 | unlock_page(pagecache_page); | |
5595 | put_page(pagecache_page); | |
5596 | } | |
b4d1d99f | 5597 | out_mutex: |
c672c7f2 | 5598 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 5599 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
5600 | /* |
5601 | * Generally it's safe to hold refcount during waiting page lock. But | |
5602 | * here we just wait to defer the next page fault to avoid busy loop and | |
5603 | * the page is not used after unlocked before returning from the current | |
5604 | * page fault. So we are safe from accessing freed page, even if we wait | |
5605 | * here without taking refcount. | |
5606 | */ | |
5607 | if (need_wait_lock) | |
5608 | wait_on_page_locked(page); | |
1e8f889b | 5609 | return ret; |
86e5216f AL |
5610 | } |
5611 | ||
714c1891 | 5612 | #ifdef CONFIG_USERFAULTFD |
8fb5debc MK |
5613 | /* |
5614 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
5615 | * modifications for huge pages. | |
5616 | */ | |
5617 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
5618 | pte_t *dst_pte, | |
5619 | struct vm_area_struct *dst_vma, | |
5620 | unsigned long dst_addr, | |
5621 | unsigned long src_addr, | |
f6191471 | 5622 | enum mcopy_atomic_mode mode, |
8fb5debc MK |
5623 | struct page **pagep) |
5624 | { | |
f6191471 | 5625 | bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); |
8cc5fcbb MA |
5626 | struct hstate *h = hstate_vma(dst_vma); |
5627 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
5628 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
1e392147 | 5629 | unsigned long size; |
1c9e8def | 5630 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
5631 | pte_t _dst_pte; |
5632 | spinlock_t *ptl; | |
8cc5fcbb | 5633 | int ret = -ENOMEM; |
8fb5debc | 5634 | struct page *page; |
f6191471 | 5635 | int writable; |
c7b1850d | 5636 | bool new_pagecache_page = false; |
8fb5debc | 5637 | |
f6191471 AR |
5638 | if (is_continue) { |
5639 | ret = -EFAULT; | |
5640 | page = find_lock_page(mapping, idx); | |
5641 | if (!page) | |
5642 | goto out; | |
5643 | } else if (!*pagep) { | |
d84cf06e MA |
5644 | /* If a page already exists, then it's UFFDIO_COPY for |
5645 | * a non-missing case. Return -EEXIST. | |
5646 | */ | |
5647 | if (vm_shared && | |
5648 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5649 | ret = -EEXIST; | |
5650 | goto out; | |
5651 | } | |
5652 | ||
8fb5debc | 5653 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
d84cf06e MA |
5654 | if (IS_ERR(page)) { |
5655 | ret = -ENOMEM; | |
8fb5debc | 5656 | goto out; |
d84cf06e | 5657 | } |
8fb5debc MK |
5658 | |
5659 | ret = copy_huge_page_from_user(page, | |
5660 | (const void __user *) src_addr, | |
810a56b9 | 5661 | pages_per_huge_page(h), false); |
8fb5debc | 5662 | |
c1e8d7c6 | 5663 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 5664 | if (unlikely(ret)) { |
9e368259 | 5665 | ret = -ENOENT; |
8cc5fcbb MA |
5666 | /* Free the allocated page which may have |
5667 | * consumed a reservation. | |
5668 | */ | |
5669 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
5670 | put_page(page); | |
5671 | ||
5672 | /* Allocate a temporary page to hold the copied | |
5673 | * contents. | |
5674 | */ | |
5675 | page = alloc_huge_page_vma(h, dst_vma, dst_addr); | |
5676 | if (!page) { | |
5677 | ret = -ENOMEM; | |
5678 | goto out; | |
5679 | } | |
8fb5debc | 5680 | *pagep = page; |
8cc5fcbb MA |
5681 | /* Set the outparam pagep and return to the caller to |
5682 | * copy the contents outside the lock. Don't free the | |
5683 | * page. | |
5684 | */ | |
8fb5debc MK |
5685 | goto out; |
5686 | } | |
5687 | } else { | |
8cc5fcbb MA |
5688 | if (vm_shared && |
5689 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5690 | put_page(*pagep); | |
5691 | ret = -EEXIST; | |
5692 | *pagep = NULL; | |
5693 | goto out; | |
5694 | } | |
5695 | ||
5696 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
5697 | if (IS_ERR(page)) { | |
5698 | ret = -ENOMEM; | |
5699 | *pagep = NULL; | |
5700 | goto out; | |
5701 | } | |
5702 | copy_huge_page(page, *pagep); | |
5703 | put_page(*pagep); | |
8fb5debc MK |
5704 | *pagep = NULL; |
5705 | } | |
5706 | ||
5707 | /* | |
5708 | * The memory barrier inside __SetPageUptodate makes sure that | |
5709 | * preceding stores to the page contents become visible before | |
5710 | * the set_pte_at() write. | |
5711 | */ | |
5712 | __SetPageUptodate(page); | |
8fb5debc | 5713 | |
f6191471 AR |
5714 | /* Add shared, newly allocated pages to the page cache. */ |
5715 | if (vm_shared && !is_continue) { | |
1e392147 AA |
5716 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5717 | ret = -EFAULT; | |
5718 | if (idx >= size) | |
5719 | goto out_release_nounlock; | |
1c9e8def | 5720 | |
1e392147 AA |
5721 | /* |
5722 | * Serialization between remove_inode_hugepages() and | |
5723 | * huge_add_to_page_cache() below happens through the | |
5724 | * hugetlb_fault_mutex_table that here must be hold by | |
5725 | * the caller. | |
5726 | */ | |
1c9e8def MK |
5727 | ret = huge_add_to_page_cache(page, mapping, idx); |
5728 | if (ret) | |
5729 | goto out_release_nounlock; | |
c7b1850d | 5730 | new_pagecache_page = true; |
1c9e8def MK |
5731 | } |
5732 | ||
8fb5debc MK |
5733 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
5734 | spin_lock(ptl); | |
5735 | ||
1e392147 AA |
5736 | /* |
5737 | * Recheck the i_size after holding PT lock to make sure not | |
5738 | * to leave any page mapped (as page_mapped()) beyond the end | |
5739 | * of the i_size (remove_inode_hugepages() is strict about | |
5740 | * enforcing that). If we bail out here, we'll also leave a | |
5741 | * page in the radix tree in the vm_shared case beyond the end | |
5742 | * of the i_size, but remove_inode_hugepages() will take care | |
5743 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
5744 | */ | |
5745 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
5746 | ret = -EFAULT; | |
5747 | if (idx >= size) | |
5748 | goto out_release_unlock; | |
5749 | ||
8fb5debc MK |
5750 | ret = -EEXIST; |
5751 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
5752 | goto out_release_unlock; | |
5753 | ||
1c9e8def MK |
5754 | if (vm_shared) { |
5755 | page_dup_rmap(page, true); | |
5756 | } else { | |
d6995da3 | 5757 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
5758 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
5759 | } | |
8fb5debc | 5760 | |
f6191471 AR |
5761 | /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ |
5762 | if (is_continue && !vm_shared) | |
5763 | writable = 0; | |
5764 | else | |
5765 | writable = dst_vma->vm_flags & VM_WRITE; | |
5766 | ||
5767 | _dst_pte = make_huge_pte(dst_vma, page, writable); | |
5768 | if (writable) | |
8fb5debc MK |
5769 | _dst_pte = huge_pte_mkdirty(_dst_pte); |
5770 | _dst_pte = pte_mkyoung(_dst_pte); | |
5771 | ||
5772 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
5773 | ||
5774 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
5775 | dst_vma->vm_flags & VM_WRITE); | |
5776 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
5777 | ||
5778 | /* No need to invalidate - it was non-present before */ | |
5779 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
5780 | ||
5781 | spin_unlock(ptl); | |
f6191471 AR |
5782 | if (!is_continue) |
5783 | SetHPageMigratable(page); | |
5784 | if (vm_shared || is_continue) | |
1c9e8def | 5785 | unlock_page(page); |
8fb5debc MK |
5786 | ret = 0; |
5787 | out: | |
5788 | return ret; | |
5789 | out_release_unlock: | |
5790 | spin_unlock(ptl); | |
f6191471 | 5791 | if (vm_shared || is_continue) |
1c9e8def | 5792 | unlock_page(page); |
5af10dfd | 5793 | out_release_nounlock: |
c7b1850d MK |
5794 | if (!new_pagecache_page) |
5795 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
8fb5debc MK |
5796 | put_page(page); |
5797 | goto out; | |
5798 | } | |
714c1891 | 5799 | #endif /* CONFIG_USERFAULTFD */ |
8fb5debc | 5800 | |
82e5d378 JM |
5801 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
5802 | int refs, struct page **pages, | |
5803 | struct vm_area_struct **vmas) | |
5804 | { | |
5805 | int nr; | |
5806 | ||
5807 | for (nr = 0; nr < refs; nr++) { | |
5808 | if (likely(pages)) | |
5809 | pages[nr] = mem_map_offset(page, nr); | |
5810 | if (vmas) | |
5811 | vmas[nr] = vma; | |
5812 | } | |
5813 | } | |
5814 | ||
28a35716 ML |
5815 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
5816 | struct page **pages, struct vm_area_struct **vmas, | |
5817 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 5818 | long i, unsigned int flags, int *locked) |
63551ae0 | 5819 | { |
d5d4b0aa CK |
5820 | unsigned long pfn_offset; |
5821 | unsigned long vaddr = *position; | |
28a35716 | 5822 | unsigned long remainder = *nr_pages; |
a5516438 | 5823 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 5824 | int err = -EFAULT, refs; |
63551ae0 | 5825 | |
63551ae0 | 5826 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 5827 | pte_t *pte; |
cb900f41 | 5828 | spinlock_t *ptl = NULL; |
2a15efc9 | 5829 | int absent; |
4c887265 | 5830 | struct page *page; |
63551ae0 | 5831 | |
02057967 DR |
5832 | /* |
5833 | * If we have a pending SIGKILL, don't keep faulting pages and | |
5834 | * potentially allocating memory. | |
5835 | */ | |
fa45f116 | 5836 | if (fatal_signal_pending(current)) { |
02057967 DR |
5837 | remainder = 0; |
5838 | break; | |
5839 | } | |
5840 | ||
4c887265 AL |
5841 | /* |
5842 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 5843 | * each hugepage. We have to make sure we get the |
4c887265 | 5844 | * first, for the page indexing below to work. |
cb900f41 KS |
5845 | * |
5846 | * Note that page table lock is not held when pte is null. | |
4c887265 | 5847 | */ |
7868a208 PA |
5848 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
5849 | huge_page_size(h)); | |
cb900f41 KS |
5850 | if (pte) |
5851 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
5852 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
5853 | ||
5854 | /* | |
5855 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
5856 | * an error where there's an empty slot with no huge pagecache |
5857 | * to back it. This way, we avoid allocating a hugepage, and | |
5858 | * the sparse dumpfile avoids allocating disk blocks, but its | |
5859 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 5860 | */ |
3ae77f43 HD |
5861 | if (absent && (flags & FOLL_DUMP) && |
5862 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
5863 | if (pte) |
5864 | spin_unlock(ptl); | |
2a15efc9 HD |
5865 | remainder = 0; |
5866 | break; | |
5867 | } | |
63551ae0 | 5868 | |
9cc3a5bd NH |
5869 | /* |
5870 | * We need call hugetlb_fault for both hugepages under migration | |
5871 | * (in which case hugetlb_fault waits for the migration,) and | |
5872 | * hwpoisoned hugepages (in which case we need to prevent the | |
5873 | * caller from accessing to them.) In order to do this, we use | |
5874 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
5875 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
5876 | * both cases, and because we can't follow correct pages | |
5877 | * directly from any kind of swap entries. | |
5878 | */ | |
5879 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
5880 | ((flags & FOLL_WRITE) && |
5881 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 5882 | vm_fault_t ret; |
87ffc118 | 5883 | unsigned int fault_flags = 0; |
63551ae0 | 5884 | |
cb900f41 KS |
5885 | if (pte) |
5886 | spin_unlock(ptl); | |
87ffc118 AA |
5887 | if (flags & FOLL_WRITE) |
5888 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 5889 | if (locked) |
71335f37 PX |
5890 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
5891 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
5892 | if (flags & FOLL_NOWAIT) |
5893 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
5894 | FAULT_FLAG_RETRY_NOWAIT; | |
5895 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
5896 | /* |
5897 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
5898 | * FAULT_FLAG_TRIED can co-exist | |
5899 | */ | |
87ffc118 AA |
5900 | fault_flags |= FAULT_FLAG_TRIED; |
5901 | } | |
5902 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
5903 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 5904 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
5905 | remainder = 0; |
5906 | break; | |
5907 | } | |
5908 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 5909 | if (locked && |
1ac25013 | 5910 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 5911 | *locked = 0; |
87ffc118 AA |
5912 | *nr_pages = 0; |
5913 | /* | |
5914 | * VM_FAULT_RETRY must not return an | |
5915 | * error, it will return zero | |
5916 | * instead. | |
5917 | * | |
5918 | * No need to update "position" as the | |
5919 | * caller will not check it after | |
5920 | * *nr_pages is set to 0. | |
5921 | */ | |
5922 | return i; | |
5923 | } | |
5924 | continue; | |
4c887265 AL |
5925 | } |
5926 | ||
a5516438 | 5927 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 5928 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 5929 | |
acbfb087 ZL |
5930 | /* |
5931 | * If subpage information not requested, update counters | |
5932 | * and skip the same_page loop below. | |
5933 | */ | |
5934 | if (!pages && !vmas && !pfn_offset && | |
5935 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
5936 | (remainder >= pages_per_huge_page(h))) { | |
5937 | vaddr += huge_page_size(h); | |
5938 | remainder -= pages_per_huge_page(h); | |
5939 | i += pages_per_huge_page(h); | |
5940 | spin_unlock(ptl); | |
5941 | continue; | |
5942 | } | |
5943 | ||
d08af0a5 JM |
5944 | /* vaddr may not be aligned to PAGE_SIZE */ |
5945 | refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, | |
5946 | (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); | |
0fa5bc40 | 5947 | |
82e5d378 JM |
5948 | if (pages || vmas) |
5949 | record_subpages_vmas(mem_map_offset(page, pfn_offset), | |
5950 | vma, refs, | |
5951 | likely(pages) ? pages + i : NULL, | |
5952 | vmas ? vmas + i : NULL); | |
63551ae0 | 5953 | |
82e5d378 | 5954 | if (pages) { |
0fa5bc40 JM |
5955 | /* |
5956 | * try_grab_compound_head() should always succeed here, | |
5957 | * because: a) we hold the ptl lock, and b) we've just | |
5958 | * checked that the huge page is present in the page | |
5959 | * tables. If the huge page is present, then the tail | |
5960 | * pages must also be present. The ptl prevents the | |
5961 | * head page and tail pages from being rearranged in | |
5962 | * any way. So this page must be available at this | |
5963 | * point, unless the page refcount overflowed: | |
5964 | */ | |
82e5d378 | 5965 | if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], |
0fa5bc40 JM |
5966 | refs, |
5967 | flags))) { | |
5968 | spin_unlock(ptl); | |
5969 | remainder = 0; | |
5970 | err = -ENOMEM; | |
5971 | break; | |
5972 | } | |
d5d4b0aa | 5973 | } |
82e5d378 JM |
5974 | |
5975 | vaddr += (refs << PAGE_SHIFT); | |
5976 | remainder -= refs; | |
5977 | i += refs; | |
5978 | ||
cb900f41 | 5979 | spin_unlock(ptl); |
63551ae0 | 5980 | } |
28a35716 | 5981 | *nr_pages = remainder; |
87ffc118 AA |
5982 | /* |
5983 | * setting position is actually required only if remainder is | |
5984 | * not zero but it's faster not to add a "if (remainder)" | |
5985 | * branch. | |
5986 | */ | |
63551ae0 DG |
5987 | *position = vaddr; |
5988 | ||
2be7cfed | 5989 | return i ? i : err; |
63551ae0 | 5990 | } |
8f860591 | 5991 | |
7da4d641 | 5992 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
5993 | unsigned long address, unsigned long end, pgprot_t newprot) |
5994 | { | |
5995 | struct mm_struct *mm = vma->vm_mm; | |
5996 | unsigned long start = address; | |
5997 | pte_t *ptep; | |
5998 | pte_t pte; | |
a5516438 | 5999 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 6000 | unsigned long pages = 0; |
dff11abe | 6001 | bool shared_pmd = false; |
ac46d4f3 | 6002 | struct mmu_notifier_range range; |
dff11abe MK |
6003 | |
6004 | /* | |
6005 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 6006 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
6007 | * range if PMD sharing is possible. |
6008 | */ | |
7269f999 JG |
6009 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
6010 | 0, vma, mm, start, end); | |
ac46d4f3 | 6011 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
6012 | |
6013 | BUG_ON(address >= end); | |
ac46d4f3 | 6014 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 6015 | |
ac46d4f3 | 6016 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 6017 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 6018 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 6019 | spinlock_t *ptl; |
7868a208 | 6020 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
6021 | if (!ptep) |
6022 | continue; | |
cb900f41 | 6023 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 6024 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
7da4d641 | 6025 | pages++; |
cb900f41 | 6026 | spin_unlock(ptl); |
dff11abe | 6027 | shared_pmd = true; |
39dde65c | 6028 | continue; |
7da4d641 | 6029 | } |
a8bda28d NH |
6030 | pte = huge_ptep_get(ptep); |
6031 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
6032 | spin_unlock(ptl); | |
6033 | continue; | |
6034 | } | |
6035 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
6036 | swp_entry_t entry = pte_to_swp_entry(pte); | |
6037 | ||
4dd845b5 | 6038 | if (is_writable_migration_entry(entry)) { |
a8bda28d NH |
6039 | pte_t newpte; |
6040 | ||
4dd845b5 AP |
6041 | entry = make_readable_migration_entry( |
6042 | swp_offset(entry)); | |
a8bda28d | 6043 | newpte = swp_entry_to_pte(entry); |
e5251fd4 PA |
6044 | set_huge_swap_pte_at(mm, address, ptep, |
6045 | newpte, huge_page_size(h)); | |
a8bda28d NH |
6046 | pages++; |
6047 | } | |
6048 | spin_unlock(ptl); | |
6049 | continue; | |
6050 | } | |
6051 | if (!huge_pte_none(pte)) { | |
023bdd00 | 6052 | pte_t old_pte; |
79c1c594 | 6053 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
023bdd00 AK |
6054 | |
6055 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
6056 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
79c1c594 | 6057 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
023bdd00 | 6058 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 6059 | pages++; |
8f860591 | 6060 | } |
cb900f41 | 6061 | spin_unlock(ptl); |
8f860591 | 6062 | } |
d833352a | 6063 | /* |
c8c06efa | 6064 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 6065 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 6066 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
6067 | * and that page table be reused and filled with junk. If we actually |
6068 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 6069 | */ |
dff11abe | 6070 | if (shared_pmd) |
ac46d4f3 | 6071 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
6072 | else |
6073 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
6074 | /* |
6075 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
6076 | * page table protection not changing it to point to a new page. | |
6077 | * | |
ad56b738 | 6078 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 6079 | */ |
83cde9e8 | 6080 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 6081 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
6082 | |
6083 | return pages << h->order; | |
8f860591 ZY |
6084 | } |
6085 | ||
33b8f84a MK |
6086 | /* Return true if reservation was successful, false otherwise. */ |
6087 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 6088 | long from, long to, |
5a6fe125 | 6089 | struct vm_area_struct *vma, |
ca16d140 | 6090 | vm_flags_t vm_flags) |
e4e574b7 | 6091 | { |
33b8f84a | 6092 | long chg, add = -1; |
a5516438 | 6093 | struct hstate *h = hstate_inode(inode); |
90481622 | 6094 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 6095 | struct resv_map *resv_map; |
075a61d0 | 6096 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 6097 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 6098 | |
63489f8e MK |
6099 | /* This should never happen */ |
6100 | if (from > to) { | |
6101 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 6102 | return false; |
63489f8e MK |
6103 | } |
6104 | ||
17c9d12e MG |
6105 | /* |
6106 | * Only apply hugepage reservation if asked. At fault time, an | |
6107 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 6108 | * without using reserves |
17c9d12e | 6109 | */ |
ca16d140 | 6110 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 6111 | return true; |
17c9d12e | 6112 | |
a1e78772 MG |
6113 | /* |
6114 | * Shared mappings base their reservation on the number of pages that | |
6115 | * are already allocated on behalf of the file. Private mappings need | |
6116 | * to reserve the full area even if read-only as mprotect() may be | |
6117 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
6118 | */ | |
9119a41e | 6119 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
6120 | /* |
6121 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
6122 | * called for inodes for which resv_maps were created (see | |
6123 | * hugetlbfs_get_inode). | |
6124 | */ | |
4e35f483 | 6125 | resv_map = inode_resv_map(inode); |
9119a41e | 6126 | |
0db9d74e | 6127 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
6128 | |
6129 | } else { | |
e9fe92ae | 6130 | /* Private mapping. */ |
9119a41e | 6131 | resv_map = resv_map_alloc(); |
17c9d12e | 6132 | if (!resv_map) |
33b8f84a | 6133 | return false; |
17c9d12e | 6134 | |
a1e78772 | 6135 | chg = to - from; |
84afd99b | 6136 | |
17c9d12e MG |
6137 | set_vma_resv_map(vma, resv_map); |
6138 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
6139 | } | |
6140 | ||
33b8f84a | 6141 | if (chg < 0) |
c50ac050 | 6142 | goto out_err; |
8a630112 | 6143 | |
33b8f84a MK |
6144 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
6145 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 6146 | goto out_err; |
075a61d0 MA |
6147 | |
6148 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
6149 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
6150 | * of the resv_map. | |
6151 | */ | |
6152 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
6153 | } | |
6154 | ||
1c5ecae3 MK |
6155 | /* |
6156 | * There must be enough pages in the subpool for the mapping. If | |
6157 | * the subpool has a minimum size, there may be some global | |
6158 | * reservations already in place (gbl_reserve). | |
6159 | */ | |
6160 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 6161 | if (gbl_reserve < 0) |
075a61d0 | 6162 | goto out_uncharge_cgroup; |
5a6fe125 MG |
6163 | |
6164 | /* | |
17c9d12e | 6165 | * Check enough hugepages are available for the reservation. |
90481622 | 6166 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 6167 | */ |
33b8f84a | 6168 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 6169 | goto out_put_pages; |
17c9d12e MG |
6170 | |
6171 | /* | |
6172 | * Account for the reservations made. Shared mappings record regions | |
6173 | * that have reservations as they are shared by multiple VMAs. | |
6174 | * When the last VMA disappears, the region map says how much | |
6175 | * the reservation was and the page cache tells how much of | |
6176 | * the reservation was consumed. Private mappings are per-VMA and | |
6177 | * only the consumed reservations are tracked. When the VMA | |
6178 | * disappears, the original reservation is the VMA size and the | |
6179 | * consumed reservations are stored in the map. Hence, nothing | |
6180 | * else has to be done for private mappings here | |
6181 | */ | |
33039678 | 6182 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 6183 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
6184 | |
6185 | if (unlikely(add < 0)) { | |
6186 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 6187 | goto out_put_pages; |
0db9d74e | 6188 | } else if (unlikely(chg > add)) { |
33039678 MK |
6189 | /* |
6190 | * pages in this range were added to the reserve | |
6191 | * map between region_chg and region_add. This | |
6192 | * indicates a race with alloc_huge_page. Adjust | |
6193 | * the subpool and reserve counts modified above | |
6194 | * based on the difference. | |
6195 | */ | |
6196 | long rsv_adjust; | |
6197 | ||
d85aecf2 ML |
6198 | /* |
6199 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
6200 | * reference to h_cg->css. See comment below for detail. | |
6201 | */ | |
075a61d0 MA |
6202 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
6203 | hstate_index(h), | |
6204 | (chg - add) * pages_per_huge_page(h), h_cg); | |
6205 | ||
33039678 MK |
6206 | rsv_adjust = hugepage_subpool_put_pages(spool, |
6207 | chg - add); | |
6208 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
6209 | } else if (h_cg) { |
6210 | /* | |
6211 | * The file_regions will hold their own reference to | |
6212 | * h_cg->css. So we should release the reference held | |
6213 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
6214 | * done. | |
6215 | */ | |
6216 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
6217 | } |
6218 | } | |
33b8f84a MK |
6219 | return true; |
6220 | ||
075a61d0 MA |
6221 | out_put_pages: |
6222 | /* put back original number of pages, chg */ | |
6223 | (void)hugepage_subpool_put_pages(spool, chg); | |
6224 | out_uncharge_cgroup: | |
6225 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
6226 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 6227 | out_err: |
5e911373 | 6228 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
6229 | /* Only call region_abort if the region_chg succeeded but the |
6230 | * region_add failed or didn't run. | |
6231 | */ | |
6232 | if (chg >= 0 && add < 0) | |
6233 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
6234 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
6235 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 6236 | return false; |
a43a8c39 CK |
6237 | } |
6238 | ||
b5cec28d MK |
6239 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
6240 | long freed) | |
a43a8c39 | 6241 | { |
a5516438 | 6242 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 6243 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 6244 | long chg = 0; |
90481622 | 6245 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 6246 | long gbl_reserve; |
45c682a6 | 6247 | |
f27a5136 MK |
6248 | /* |
6249 | * Since this routine can be called in the evict inode path for all | |
6250 | * hugetlbfs inodes, resv_map could be NULL. | |
6251 | */ | |
b5cec28d MK |
6252 | if (resv_map) { |
6253 | chg = region_del(resv_map, start, end); | |
6254 | /* | |
6255 | * region_del() can fail in the rare case where a region | |
6256 | * must be split and another region descriptor can not be | |
6257 | * allocated. If end == LONG_MAX, it will not fail. | |
6258 | */ | |
6259 | if (chg < 0) | |
6260 | return chg; | |
6261 | } | |
6262 | ||
45c682a6 | 6263 | spin_lock(&inode->i_lock); |
e4c6f8be | 6264 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
6265 | spin_unlock(&inode->i_lock); |
6266 | ||
1c5ecae3 MK |
6267 | /* |
6268 | * If the subpool has a minimum size, the number of global | |
6269 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
6270 | * |
6271 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
6272 | * won't go negative. | |
1c5ecae3 MK |
6273 | */ |
6274 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
6275 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
6276 | |
6277 | return 0; | |
a43a8c39 | 6278 | } |
93f70f90 | 6279 | |
3212b535 SC |
6280 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
6281 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
6282 | struct vm_area_struct *vma, | |
6283 | unsigned long addr, pgoff_t idx) | |
6284 | { | |
6285 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
6286 | svma->vm_start; | |
6287 | unsigned long sbase = saddr & PUD_MASK; | |
6288 | unsigned long s_end = sbase + PUD_SIZE; | |
6289 | ||
6290 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
6291 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
6292 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
6293 | |
6294 | /* | |
6295 | * match the virtual addresses, permission and the alignment of the | |
6296 | * page table page. | |
6297 | */ | |
6298 | if (pmd_index(addr) != pmd_index(saddr) || | |
6299 | vm_flags != svm_flags || | |
07e51edf | 6300 | !range_in_vma(svma, sbase, s_end)) |
3212b535 SC |
6301 | return 0; |
6302 | ||
6303 | return saddr; | |
6304 | } | |
6305 | ||
31aafb45 | 6306 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
6307 | { |
6308 | unsigned long base = addr & PUD_MASK; | |
6309 | unsigned long end = base + PUD_SIZE; | |
6310 | ||
6311 | /* | |
6312 | * check on proper vm_flags and page table alignment | |
6313 | */ | |
017b1660 | 6314 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
6315 | return true; |
6316 | return false; | |
3212b535 SC |
6317 | } |
6318 | ||
c1991e07 PX |
6319 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
6320 | { | |
6321 | #ifdef CONFIG_USERFAULTFD | |
6322 | if (uffd_disable_huge_pmd_share(vma)) | |
6323 | return false; | |
6324 | #endif | |
6325 | return vma_shareable(vma, addr); | |
6326 | } | |
6327 | ||
017b1660 MK |
6328 | /* |
6329 | * Determine if start,end range within vma could be mapped by shared pmd. | |
6330 | * If yes, adjust start and end to cover range associated with possible | |
6331 | * shared pmd mappings. | |
6332 | */ | |
6333 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6334 | unsigned long *start, unsigned long *end) | |
6335 | { | |
a1ba9da8 LX |
6336 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
6337 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 6338 | |
a1ba9da8 | 6339 | /* |
f0953a1b IM |
6340 | * vma needs to span at least one aligned PUD size, and the range |
6341 | * must be at least partially within in. | |
a1ba9da8 LX |
6342 | */ |
6343 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
6344 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
6345 | return; |
6346 | ||
75802ca6 | 6347 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
6348 | if (*start > v_start) |
6349 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 6350 | |
a1ba9da8 LX |
6351 | if (*end < v_end) |
6352 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
6353 | } |
6354 | ||
3212b535 SC |
6355 | /* |
6356 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
6357 | * and returns the corresponding pte. While this is not necessary for the | |
6358 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
6359 | * code much cleaner. |
6360 | * | |
0bf7b64e MK |
6361 | * This routine must be called with i_mmap_rwsem held in at least read mode if |
6362 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
6363 | * table entries associated with the address space. This is important as we | |
6364 | * are setting up sharing based on existing page table entries (mappings). | |
3212b535 | 6365 | */ |
aec44e0f PX |
6366 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6367 | unsigned long addr, pud_t *pud) | |
3212b535 | 6368 | { |
3212b535 SC |
6369 | struct address_space *mapping = vma->vm_file->f_mapping; |
6370 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
6371 | vma->vm_pgoff; | |
6372 | struct vm_area_struct *svma; | |
6373 | unsigned long saddr; | |
6374 | pte_t *spte = NULL; | |
6375 | pte_t *pte; | |
cb900f41 | 6376 | spinlock_t *ptl; |
3212b535 | 6377 | |
0bf7b64e | 6378 | i_mmap_assert_locked(mapping); |
3212b535 SC |
6379 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
6380 | if (svma == vma) | |
6381 | continue; | |
6382 | ||
6383 | saddr = page_table_shareable(svma, vma, addr, idx); | |
6384 | if (saddr) { | |
7868a208 PA |
6385 | spte = huge_pte_offset(svma->vm_mm, saddr, |
6386 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
6387 | if (spte) { |
6388 | get_page(virt_to_page(spte)); | |
6389 | break; | |
6390 | } | |
6391 | } | |
6392 | } | |
6393 | ||
6394 | if (!spte) | |
6395 | goto out; | |
6396 | ||
8bea8052 | 6397 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 6398 | if (pud_none(*pud)) { |
3212b535 SC |
6399 | pud_populate(mm, pud, |
6400 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 6401 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 6402 | } else { |
3212b535 | 6403 | put_page(virt_to_page(spte)); |
dc6c9a35 | 6404 | } |
cb900f41 | 6405 | spin_unlock(ptl); |
3212b535 SC |
6406 | out: |
6407 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
6408 | return pte; |
6409 | } | |
6410 | ||
6411 | /* | |
6412 | * unmap huge page backed by shared pte. | |
6413 | * | |
6414 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
6415 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
6416 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
6417 | * | |
c0d0381a | 6418 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
6419 | * |
6420 | * returns: 1 successfully unmapped a shared pte page | |
6421 | * 0 the underlying pte page is not shared, or it is the last user | |
6422 | */ | |
34ae204f MK |
6423 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
6424 | unsigned long *addr, pte_t *ptep) | |
3212b535 SC |
6425 | { |
6426 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
6427 | p4d_t *p4d = p4d_offset(pgd, *addr); |
6428 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 | 6429 | |
34ae204f | 6430 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
6431 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
6432 | if (page_count(virt_to_page(ptep)) == 1) | |
6433 | return 0; | |
6434 | ||
6435 | pud_clear(pud); | |
6436 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 6437 | mm_dec_nr_pmds(mm); |
3212b535 SC |
6438 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
6439 | return 1; | |
6440 | } | |
c1991e07 | 6441 | |
9e5fc74c | 6442 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
aec44e0f PX |
6443 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6444 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
6445 | { |
6446 | return NULL; | |
6447 | } | |
e81f2d22 | 6448 | |
34ae204f MK |
6449 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
6450 | unsigned long *addr, pte_t *ptep) | |
e81f2d22 ZZ |
6451 | { |
6452 | return 0; | |
6453 | } | |
017b1660 MK |
6454 | |
6455 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6456 | unsigned long *start, unsigned long *end) | |
6457 | { | |
6458 | } | |
c1991e07 PX |
6459 | |
6460 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
6461 | { | |
6462 | return false; | |
6463 | } | |
3212b535 SC |
6464 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
6465 | ||
9e5fc74c | 6466 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 6467 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
6468 | unsigned long addr, unsigned long sz) |
6469 | { | |
6470 | pgd_t *pgd; | |
c2febafc | 6471 | p4d_t *p4d; |
9e5fc74c SC |
6472 | pud_t *pud; |
6473 | pte_t *pte = NULL; | |
6474 | ||
6475 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
6476 | p4d = p4d_alloc(mm, pgd, addr); |
6477 | if (!p4d) | |
6478 | return NULL; | |
c2febafc | 6479 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
6480 | if (pud) { |
6481 | if (sz == PUD_SIZE) { | |
6482 | pte = (pte_t *)pud; | |
6483 | } else { | |
6484 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 6485 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 6486 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
6487 | else |
6488 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
6489 | } | |
6490 | } | |
4e666314 | 6491 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
6492 | |
6493 | return pte; | |
6494 | } | |
6495 | ||
9b19df29 PA |
6496 | /* |
6497 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
6498 | * entry at address @addr | |
6499 | * | |
8ac0b81a LX |
6500 | * Return: Pointer to page table entry (PUD or PMD) for |
6501 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
6502 | * size @sz doesn't match the hugepage size at this level of the page |
6503 | * table. | |
6504 | */ | |
7868a208 PA |
6505 | pte_t *huge_pte_offset(struct mm_struct *mm, |
6506 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
6507 | { |
6508 | pgd_t *pgd; | |
c2febafc | 6509 | p4d_t *p4d; |
8ac0b81a LX |
6510 | pud_t *pud; |
6511 | pmd_t *pmd; | |
9e5fc74c SC |
6512 | |
6513 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
6514 | if (!pgd_present(*pgd)) |
6515 | return NULL; | |
6516 | p4d = p4d_offset(pgd, addr); | |
6517 | if (!p4d_present(*p4d)) | |
6518 | return NULL; | |
9b19df29 | 6519 | |
c2febafc | 6520 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
6521 | if (sz == PUD_SIZE) |
6522 | /* must be pud huge, non-present or none */ | |
c2febafc | 6523 | return (pte_t *)pud; |
8ac0b81a | 6524 | if (!pud_present(*pud)) |
9b19df29 | 6525 | return NULL; |
8ac0b81a | 6526 | /* must have a valid entry and size to go further */ |
9b19df29 | 6527 | |
8ac0b81a LX |
6528 | pmd = pmd_offset(pud, addr); |
6529 | /* must be pmd huge, non-present or none */ | |
6530 | return (pte_t *)pmd; | |
9e5fc74c SC |
6531 | } |
6532 | ||
61f77eda NH |
6533 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
6534 | ||
6535 | /* | |
6536 | * These functions are overwritable if your architecture needs its own | |
6537 | * behavior. | |
6538 | */ | |
6539 | struct page * __weak | |
6540 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
6541 | int write) | |
6542 | { | |
6543 | return ERR_PTR(-EINVAL); | |
6544 | } | |
6545 | ||
4dc71451 AK |
6546 | struct page * __weak |
6547 | follow_huge_pd(struct vm_area_struct *vma, | |
6548 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
6549 | { | |
6550 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
6551 | return NULL; | |
6552 | } | |
6553 | ||
61f77eda | 6554 | struct page * __weak |
9e5fc74c | 6555 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 6556 | pmd_t *pmd, int flags) |
9e5fc74c | 6557 | { |
e66f17ff NH |
6558 | struct page *page = NULL; |
6559 | spinlock_t *ptl; | |
c9d398fa | 6560 | pte_t pte; |
3faa52c0 JH |
6561 | |
6562 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
6563 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
6564 | (FOLL_PIN | FOLL_GET))) | |
6565 | return NULL; | |
6566 | ||
e66f17ff NH |
6567 | retry: |
6568 | ptl = pmd_lockptr(mm, pmd); | |
6569 | spin_lock(ptl); | |
6570 | /* | |
6571 | * make sure that the address range covered by this pmd is not | |
6572 | * unmapped from other threads. | |
6573 | */ | |
6574 | if (!pmd_huge(*pmd)) | |
6575 | goto out; | |
c9d398fa NH |
6576 | pte = huge_ptep_get((pte_t *)pmd); |
6577 | if (pte_present(pte)) { | |
97534127 | 6578 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
6579 | /* |
6580 | * try_grab_page() should always succeed here, because: a) we | |
6581 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
6582 | * huge pmd (head) page is present in the page tables. The ptl | |
6583 | * prevents the head page and tail pages from being rearranged | |
6584 | * in any way. So this page must be available at this point, | |
6585 | * unless the page refcount overflowed: | |
6586 | */ | |
6587 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
6588 | page = NULL; | |
6589 | goto out; | |
6590 | } | |
e66f17ff | 6591 | } else { |
c9d398fa | 6592 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
6593 | spin_unlock(ptl); |
6594 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
6595 | goto retry; | |
6596 | } | |
6597 | /* | |
6598 | * hwpoisoned entry is treated as no_page_table in | |
6599 | * follow_page_mask(). | |
6600 | */ | |
6601 | } | |
6602 | out: | |
6603 | spin_unlock(ptl); | |
9e5fc74c SC |
6604 | return page; |
6605 | } | |
6606 | ||
61f77eda | 6607 | struct page * __weak |
9e5fc74c | 6608 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 6609 | pud_t *pud, int flags) |
9e5fc74c | 6610 | { |
3faa52c0 | 6611 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 6612 | return NULL; |
9e5fc74c | 6613 | |
e66f17ff | 6614 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
6615 | } |
6616 | ||
faaa5b62 AK |
6617 | struct page * __weak |
6618 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
6619 | { | |
3faa52c0 | 6620 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
6621 | return NULL; |
6622 | ||
6623 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
6624 | } | |
6625 | ||
31caf665 NH |
6626 | bool isolate_huge_page(struct page *page, struct list_head *list) |
6627 | { | |
bcc54222 NH |
6628 | bool ret = true; |
6629 | ||
db71ef79 | 6630 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
6631 | if (!PageHeadHuge(page) || |
6632 | !HPageMigratable(page) || | |
0eb2df2b | 6633 | !get_page_unless_zero(page)) { |
bcc54222 NH |
6634 | ret = false; |
6635 | goto unlock; | |
6636 | } | |
8f251a3d | 6637 | ClearHPageMigratable(page); |
31caf665 | 6638 | list_move_tail(&page->lru, list); |
bcc54222 | 6639 | unlock: |
db71ef79 | 6640 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 6641 | return ret; |
31caf665 NH |
6642 | } |
6643 | ||
25182f05 NH |
6644 | int get_hwpoison_huge_page(struct page *page, bool *hugetlb) |
6645 | { | |
6646 | int ret = 0; | |
6647 | ||
6648 | *hugetlb = false; | |
6649 | spin_lock_irq(&hugetlb_lock); | |
6650 | if (PageHeadHuge(page)) { | |
6651 | *hugetlb = true; | |
6652 | if (HPageFreed(page) || HPageMigratable(page)) | |
6653 | ret = get_page_unless_zero(page); | |
0ed950d1 NH |
6654 | else |
6655 | ret = -EBUSY; | |
25182f05 NH |
6656 | } |
6657 | spin_unlock_irq(&hugetlb_lock); | |
6658 | return ret; | |
6659 | } | |
6660 | ||
31caf665 NH |
6661 | void putback_active_hugepage(struct page *page) |
6662 | { | |
db71ef79 | 6663 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 6664 | SetHPageMigratable(page); |
31caf665 | 6665 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 6666 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
6667 | put_page(page); |
6668 | } | |
ab5ac90a MH |
6669 | |
6670 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
6671 | { | |
6672 | struct hstate *h = page_hstate(oldpage); | |
6673 | ||
6674 | hugetlb_cgroup_migrate(oldpage, newpage); | |
6675 | set_page_owner_migrate_reason(newpage, reason); | |
6676 | ||
6677 | /* | |
6678 | * transfer temporary state of the new huge page. This is | |
6679 | * reverse to other transitions because the newpage is going to | |
6680 | * be final while the old one will be freed so it takes over | |
6681 | * the temporary status. | |
6682 | * | |
6683 | * Also note that we have to transfer the per-node surplus state | |
6684 | * here as well otherwise the global surplus count will not match | |
6685 | * the per-node's. | |
6686 | */ | |
9157c311 | 6687 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
6688 | int old_nid = page_to_nid(oldpage); |
6689 | int new_nid = page_to_nid(newpage); | |
6690 | ||
9157c311 MK |
6691 | SetHPageTemporary(oldpage); |
6692 | ClearHPageTemporary(newpage); | |
ab5ac90a | 6693 | |
5af1ab1d ML |
6694 | /* |
6695 | * There is no need to transfer the per-node surplus state | |
6696 | * when we do not cross the node. | |
6697 | */ | |
6698 | if (new_nid == old_nid) | |
6699 | return; | |
db71ef79 | 6700 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
6701 | if (h->surplus_huge_pages_node[old_nid]) { |
6702 | h->surplus_huge_pages_node[old_nid]--; | |
6703 | h->surplus_huge_pages_node[new_nid]++; | |
6704 | } | |
db71ef79 | 6705 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
6706 | } |
6707 | } | |
cf11e85f | 6708 | |
6dfeaff9 PX |
6709 | /* |
6710 | * This function will unconditionally remove all the shared pmd pgtable entries | |
6711 | * within the specific vma for a hugetlbfs memory range. | |
6712 | */ | |
6713 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
6714 | { | |
6715 | struct hstate *h = hstate_vma(vma); | |
6716 | unsigned long sz = huge_page_size(h); | |
6717 | struct mm_struct *mm = vma->vm_mm; | |
6718 | struct mmu_notifier_range range; | |
6719 | unsigned long address, start, end; | |
6720 | spinlock_t *ptl; | |
6721 | pte_t *ptep; | |
6722 | ||
6723 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
6724 | return; | |
6725 | ||
6726 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
6727 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
6728 | ||
6729 | if (start >= end) | |
6730 | return; | |
6731 | ||
6732 | /* | |
6733 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
6734 | * we have already done the PUD_SIZE alignment. | |
6735 | */ | |
6736 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
6737 | start, end); | |
6738 | mmu_notifier_invalidate_range_start(&range); | |
6739 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
6740 | for (address = start; address < end; address += PUD_SIZE) { | |
6741 | unsigned long tmp = address; | |
6742 | ||
6743 | ptep = huge_pte_offset(mm, address, sz); | |
6744 | if (!ptep) | |
6745 | continue; | |
6746 | ptl = huge_pte_lock(h, mm, ptep); | |
6747 | /* We don't want 'address' to be changed */ | |
6748 | huge_pmd_unshare(mm, vma, &tmp, ptep); | |
6749 | spin_unlock(ptl); | |
6750 | } | |
6751 | flush_hugetlb_tlb_range(vma, start, end); | |
6752 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
6753 | /* | |
6754 | * No need to call mmu_notifier_invalidate_range(), see | |
6755 | * Documentation/vm/mmu_notifier.rst. | |
6756 | */ | |
6757 | mmu_notifier_invalidate_range_end(&range); | |
6758 | } | |
6759 | ||
cf11e85f | 6760 | #ifdef CONFIG_CMA |
cf11e85f RG |
6761 | static bool cma_reserve_called __initdata; |
6762 | ||
6763 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
6764 | { | |
6765 | hugetlb_cma_size = memparse(p, &p); | |
6766 | return 0; | |
6767 | } | |
6768 | ||
6769 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
6770 | ||
6771 | void __init hugetlb_cma_reserve(int order) | |
6772 | { | |
6773 | unsigned long size, reserved, per_node; | |
6774 | int nid; | |
6775 | ||
6776 | cma_reserve_called = true; | |
6777 | ||
6778 | if (!hugetlb_cma_size) | |
6779 | return; | |
6780 | ||
6781 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
6782 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
6783 | (PAGE_SIZE << order) / SZ_1M); | |
a01f4390 | 6784 | hugetlb_cma_size = 0; |
cf11e85f RG |
6785 | return; |
6786 | } | |
6787 | ||
6788 | /* | |
6789 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
6790 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
6791 | */ | |
6792 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
6793 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
6794 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
6795 | ||
6796 | reserved = 0; | |
6797 | for_each_node_state(nid, N_ONLINE) { | |
6798 | int res; | |
2281f797 | 6799 | char name[CMA_MAX_NAME]; |
cf11e85f RG |
6800 | |
6801 | size = min(per_node, hugetlb_cma_size - reserved); | |
6802 | size = round_up(size, PAGE_SIZE << order); | |
6803 | ||
2281f797 | 6804 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
a01f4390 MK |
6805 | /* |
6806 | * Note that 'order per bit' is based on smallest size that | |
6807 | * may be returned to CMA allocator in the case of | |
6808 | * huge page demotion. | |
6809 | */ | |
6810 | res = cma_declare_contiguous_nid(0, size, 0, | |
6811 | PAGE_SIZE << HUGETLB_PAGE_ORDER, | |
29d0f41d | 6812 | 0, false, name, |
cf11e85f RG |
6813 | &hugetlb_cma[nid], nid); |
6814 | if (res) { | |
6815 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
6816 | res, nid); | |
6817 | continue; | |
6818 | } | |
6819 | ||
6820 | reserved += size; | |
6821 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
6822 | size / SZ_1M, nid); | |
6823 | ||
6824 | if (reserved >= hugetlb_cma_size) | |
6825 | break; | |
6826 | } | |
a01f4390 MK |
6827 | |
6828 | if (!reserved) | |
6829 | /* | |
6830 | * hugetlb_cma_size is used to determine if allocations from | |
6831 | * cma are possible. Set to zero if no cma regions are set up. | |
6832 | */ | |
6833 | hugetlb_cma_size = 0; | |
cf11e85f RG |
6834 | } |
6835 | ||
6836 | void __init hugetlb_cma_check(void) | |
6837 | { | |
6838 | if (!hugetlb_cma_size || cma_reserve_called) | |
6839 | return; | |
6840 | ||
6841 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
6842 | } | |
6843 | ||
6844 | #endif /* CONFIG_CMA */ |