]>
Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * Generic hugetlb support. | |
6d49e352 | 4 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 5 | */ |
1da177e4 LT |
6 | #include <linux/list.h> |
7 | #include <linux/init.h> | |
1da177e4 | 8 | #include <linux/mm.h> |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
97ad1087 | 19 | #include <linux/memblock.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
bbe88753 | 22 | #include <linux/sched/mm.h> |
63489f8e | 23 | #include <linux/mmdebug.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
0fe6e20b | 25 | #include <linux/rmap.h> |
c6247f72 | 26 | #include <linux/string_helpers.h> |
fd6a03ed NH |
27 | #include <linux/swap.h> |
28 | #include <linux/swapops.h> | |
8382d914 | 29 | #include <linux/jhash.h> |
98fa15f3 | 30 | #include <linux/numa.h> |
c77c0a8a | 31 | #include <linux/llist.h> |
cf11e85f | 32 | #include <linux/cma.h> |
d6606683 | 33 | |
63551ae0 | 34 | #include <asm/page.h> |
ca15ca40 | 35 | #include <asm/pgalloc.h> |
24669e58 | 36 | #include <asm/tlb.h> |
63551ae0 | 37 | |
24669e58 | 38 | #include <linux/io.h> |
63551ae0 | 39 | #include <linux/hugetlb.h> |
9dd540e2 | 40 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 41 | #include <linux/node.h> |
ab5ac90a | 42 | #include <linux/page_owner.h> |
7835e98b | 43 | #include "internal.h" |
1da177e4 | 44 | |
c3f38a38 | 45 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
46 | unsigned int default_hstate_idx; |
47 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 48 | |
dbda8fea | 49 | #ifdef CONFIG_CMA |
cf11e85f | 50 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
dbda8fea BS |
51 | #endif |
52 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 53 | |
641844f5 NH |
54 | /* |
55 | * Minimum page order among possible hugepage sizes, set to a proper value | |
56 | * at boot time. | |
57 | */ | |
58 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 59 | |
53ba51d2 JT |
60 | __initdata LIST_HEAD(huge_boot_pages); |
61 | ||
e5ff2159 AK |
62 | /* for command line parsing */ |
63 | static struct hstate * __initdata parsed_hstate; | |
64 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 65 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 66 | static bool __initdata parsed_default_hugepagesz; |
e5ff2159 | 67 | |
3935baa9 | 68 | /* |
31caf665 NH |
69 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
70 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 71 | */ |
c3f38a38 | 72 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 73 | |
8382d914 DB |
74 | /* |
75 | * Serializes faults on the same logical page. This is used to | |
76 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
77 | */ | |
78 | static int num_fault_mutexes; | |
c672c7f2 | 79 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 80 | |
7ca02d0a MK |
81 | /* Forward declaration */ |
82 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
83 | ||
1d88433b | 84 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 85 | { |
1d88433b ML |
86 | if (spool->count) |
87 | return false; | |
88 | if (spool->max_hpages != -1) | |
89 | return spool->used_hpages == 0; | |
90 | if (spool->min_hpages != -1) | |
91 | return spool->rsv_hpages == spool->min_hpages; | |
92 | ||
93 | return true; | |
94 | } | |
90481622 | 95 | |
db71ef79 MK |
96 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, |
97 | unsigned long irq_flags) | |
1d88433b | 98 | { |
db71ef79 | 99 | spin_unlock_irqrestore(&spool->lock, irq_flags); |
90481622 DG |
100 | |
101 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 102 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 103 | * free the subpool */ |
1d88433b | 104 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
105 | if (spool->min_hpages != -1) |
106 | hugetlb_acct_memory(spool->hstate, | |
107 | -spool->min_hpages); | |
90481622 | 108 | kfree(spool); |
7ca02d0a | 109 | } |
90481622 DG |
110 | } |
111 | ||
7ca02d0a MK |
112 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
113 | long min_hpages) | |
90481622 DG |
114 | { |
115 | struct hugepage_subpool *spool; | |
116 | ||
c6a91820 | 117 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
118 | if (!spool) |
119 | return NULL; | |
120 | ||
121 | spin_lock_init(&spool->lock); | |
122 | spool->count = 1; | |
7ca02d0a MK |
123 | spool->max_hpages = max_hpages; |
124 | spool->hstate = h; | |
125 | spool->min_hpages = min_hpages; | |
126 | ||
127 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
128 | kfree(spool); | |
129 | return NULL; | |
130 | } | |
131 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
132 | |
133 | return spool; | |
134 | } | |
135 | ||
136 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
137 | { | |
db71ef79 MK |
138 | unsigned long flags; |
139 | ||
140 | spin_lock_irqsave(&spool->lock, flags); | |
90481622 DG |
141 | BUG_ON(!spool->count); |
142 | spool->count--; | |
db71ef79 | 143 | unlock_or_release_subpool(spool, flags); |
90481622 DG |
144 | } |
145 | ||
1c5ecae3 MK |
146 | /* |
147 | * Subpool accounting for allocating and reserving pages. | |
148 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 149 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
150 | * global pools must be adjusted (upward). The returned value may |
151 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 152 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
153 | */ |
154 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
155 | long delta) |
156 | { | |
1c5ecae3 | 157 | long ret = delta; |
90481622 DG |
158 | |
159 | if (!spool) | |
1c5ecae3 | 160 | return ret; |
90481622 | 161 | |
db71ef79 | 162 | spin_lock_irq(&spool->lock); |
1c5ecae3 MK |
163 | |
164 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
165 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
166 | spool->used_hpages += delta; | |
167 | else { | |
168 | ret = -ENOMEM; | |
169 | goto unlock_ret; | |
170 | } | |
90481622 | 171 | } |
90481622 | 172 | |
09a95e29 MK |
173 | /* minimum size accounting */ |
174 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
175 | if (delta > spool->rsv_hpages) { |
176 | /* | |
177 | * Asking for more reserves than those already taken on | |
178 | * behalf of subpool. Return difference. | |
179 | */ | |
180 | ret = delta - spool->rsv_hpages; | |
181 | spool->rsv_hpages = 0; | |
182 | } else { | |
183 | ret = 0; /* reserves already accounted for */ | |
184 | spool->rsv_hpages -= delta; | |
185 | } | |
186 | } | |
187 | ||
188 | unlock_ret: | |
db71ef79 | 189 | spin_unlock_irq(&spool->lock); |
90481622 DG |
190 | return ret; |
191 | } | |
192 | ||
1c5ecae3 MK |
193 | /* |
194 | * Subpool accounting for freeing and unreserving pages. | |
195 | * Return the number of global page reservations that must be dropped. | |
196 | * The return value may only be different than the passed value (delta) | |
197 | * in the case where a subpool minimum size must be maintained. | |
198 | */ | |
199 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
200 | long delta) |
201 | { | |
1c5ecae3 | 202 | long ret = delta; |
db71ef79 | 203 | unsigned long flags; |
1c5ecae3 | 204 | |
90481622 | 205 | if (!spool) |
1c5ecae3 | 206 | return delta; |
90481622 | 207 | |
db71ef79 | 208 | spin_lock_irqsave(&spool->lock, flags); |
1c5ecae3 MK |
209 | |
210 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
211 | spool->used_hpages -= delta; | |
212 | ||
09a95e29 MK |
213 | /* minimum size accounting */ |
214 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
215 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
216 | ret = 0; | |
217 | else | |
218 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
219 | ||
220 | spool->rsv_hpages += delta; | |
221 | if (spool->rsv_hpages > spool->min_hpages) | |
222 | spool->rsv_hpages = spool->min_hpages; | |
223 | } | |
224 | ||
225 | /* | |
226 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
227 | * quota reference, free it now. | |
228 | */ | |
db71ef79 | 229 | unlock_or_release_subpool(spool, flags); |
1c5ecae3 MK |
230 | |
231 | return ret; | |
90481622 DG |
232 | } |
233 | ||
234 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
235 | { | |
236 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
237 | } | |
238 | ||
239 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
240 | { | |
496ad9aa | 241 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
242 | } |
243 | ||
0db9d74e MA |
244 | /* Helper that removes a struct file_region from the resv_map cache and returns |
245 | * it for use. | |
246 | */ | |
247 | static struct file_region * | |
248 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
249 | { | |
250 | struct file_region *nrg = NULL; | |
251 | ||
252 | VM_BUG_ON(resv->region_cache_count <= 0); | |
253 | ||
254 | resv->region_cache_count--; | |
255 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
256 | list_del(&nrg->link); |
257 | ||
258 | nrg->from = from; | |
259 | nrg->to = to; | |
260 | ||
261 | return nrg; | |
262 | } | |
263 | ||
075a61d0 MA |
264 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
265 | struct file_region *rg) | |
266 | { | |
267 | #ifdef CONFIG_CGROUP_HUGETLB | |
268 | nrg->reservation_counter = rg->reservation_counter; | |
269 | nrg->css = rg->css; | |
270 | if (rg->css) | |
271 | css_get(rg->css); | |
272 | #endif | |
273 | } | |
274 | ||
275 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
276 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
277 | struct hstate *h, | |
278 | struct resv_map *resv, | |
279 | struct file_region *nrg) | |
280 | { | |
281 | #ifdef CONFIG_CGROUP_HUGETLB | |
282 | if (h_cg) { | |
283 | nrg->reservation_counter = | |
284 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
285 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
286 | /* |
287 | * The caller will hold exactly one h_cg->css reference for the | |
288 | * whole contiguous reservation region. But this area might be | |
289 | * scattered when there are already some file_regions reside in | |
290 | * it. As a result, many file_regions may share only one css | |
291 | * reference. In order to ensure that one file_region must hold | |
292 | * exactly one h_cg->css reference, we should do css_get for | |
293 | * each file_region and leave the reference held by caller | |
294 | * untouched. | |
295 | */ | |
296 | css_get(&h_cg->css); | |
075a61d0 MA |
297 | if (!resv->pages_per_hpage) |
298 | resv->pages_per_hpage = pages_per_huge_page(h); | |
299 | /* pages_per_hpage should be the same for all entries in | |
300 | * a resv_map. | |
301 | */ | |
302 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
303 | } else { | |
304 | nrg->reservation_counter = NULL; | |
305 | nrg->css = NULL; | |
306 | } | |
307 | #endif | |
308 | } | |
309 | ||
d85aecf2 ML |
310 | static void put_uncharge_info(struct file_region *rg) |
311 | { | |
312 | #ifdef CONFIG_CGROUP_HUGETLB | |
313 | if (rg->css) | |
314 | css_put(rg->css); | |
315 | #endif | |
316 | } | |
317 | ||
a9b3f867 MA |
318 | static bool has_same_uncharge_info(struct file_region *rg, |
319 | struct file_region *org) | |
320 | { | |
321 | #ifdef CONFIG_CGROUP_HUGETLB | |
322 | return rg && org && | |
323 | rg->reservation_counter == org->reservation_counter && | |
324 | rg->css == org->css; | |
325 | ||
326 | #else | |
327 | return true; | |
328 | #endif | |
329 | } | |
330 | ||
331 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
332 | { | |
333 | struct file_region *nrg = NULL, *prg = NULL; | |
334 | ||
335 | prg = list_prev_entry(rg, link); | |
336 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
337 | has_same_uncharge_info(prg, rg)) { | |
338 | prg->to = rg->to; | |
339 | ||
340 | list_del(&rg->link); | |
d85aecf2 | 341 | put_uncharge_info(rg); |
a9b3f867 MA |
342 | kfree(rg); |
343 | ||
7db5e7b6 | 344 | rg = prg; |
a9b3f867 MA |
345 | } |
346 | ||
347 | nrg = list_next_entry(rg, link); | |
348 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
349 | has_same_uncharge_info(nrg, rg)) { | |
350 | nrg->from = rg->from; | |
351 | ||
352 | list_del(&rg->link); | |
d85aecf2 | 353 | put_uncharge_info(rg); |
a9b3f867 | 354 | kfree(rg); |
a9b3f867 MA |
355 | } |
356 | } | |
357 | ||
2103cf9c PX |
358 | static inline long |
359 | hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, | |
360 | long to, struct hstate *h, struct hugetlb_cgroup *cg, | |
361 | long *regions_needed) | |
362 | { | |
363 | struct file_region *nrg; | |
364 | ||
365 | if (!regions_needed) { | |
366 | nrg = get_file_region_entry_from_cache(map, from, to); | |
367 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
368 | list_add(&nrg->link, rg->link.prev); | |
369 | coalesce_file_region(map, nrg); | |
370 | } else | |
371 | *regions_needed += 1; | |
372 | ||
373 | return to - from; | |
374 | } | |
375 | ||
972a3da3 WY |
376 | /* |
377 | * Must be called with resv->lock held. | |
378 | * | |
379 | * Calling this with regions_needed != NULL will count the number of pages | |
380 | * to be added but will not modify the linked list. And regions_needed will | |
381 | * indicate the number of file_regions needed in the cache to carry out to add | |
382 | * the regions for this range. | |
d75c6af9 MA |
383 | */ |
384 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 385 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 386 | struct hstate *h, long *regions_needed) |
d75c6af9 | 387 | { |
0db9d74e | 388 | long add = 0; |
d75c6af9 | 389 | struct list_head *head = &resv->regions; |
0db9d74e | 390 | long last_accounted_offset = f; |
2103cf9c | 391 | struct file_region *rg = NULL, *trg = NULL; |
d75c6af9 | 392 | |
0db9d74e MA |
393 | if (regions_needed) |
394 | *regions_needed = 0; | |
d75c6af9 | 395 | |
0db9d74e MA |
396 | /* In this loop, we essentially handle an entry for the range |
397 | * [last_accounted_offset, rg->from), at every iteration, with some | |
398 | * bounds checking. | |
399 | */ | |
400 | list_for_each_entry_safe(rg, trg, head, link) { | |
401 | /* Skip irrelevant regions that start before our range. */ | |
402 | if (rg->from < f) { | |
403 | /* If this region ends after the last accounted offset, | |
404 | * then we need to update last_accounted_offset. | |
405 | */ | |
406 | if (rg->to > last_accounted_offset) | |
407 | last_accounted_offset = rg->to; | |
408 | continue; | |
409 | } | |
d75c6af9 | 410 | |
0db9d74e MA |
411 | /* When we find a region that starts beyond our range, we've |
412 | * finished. | |
413 | */ | |
ca7e0457 | 414 | if (rg->from >= t) |
d75c6af9 MA |
415 | break; |
416 | ||
0db9d74e MA |
417 | /* Add an entry for last_accounted_offset -> rg->from, and |
418 | * update last_accounted_offset. | |
419 | */ | |
2103cf9c PX |
420 | if (rg->from > last_accounted_offset) |
421 | add += hugetlb_resv_map_add(resv, rg, | |
422 | last_accounted_offset, | |
423 | rg->from, h, h_cg, | |
424 | regions_needed); | |
0db9d74e MA |
425 | |
426 | last_accounted_offset = rg->to; | |
427 | } | |
428 | ||
429 | /* Handle the case where our range extends beyond | |
430 | * last_accounted_offset. | |
431 | */ | |
2103cf9c PX |
432 | if (last_accounted_offset < t) |
433 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
434 | t, h, h_cg, regions_needed); | |
0db9d74e MA |
435 | |
436 | VM_BUG_ON(add < 0); | |
437 | return add; | |
438 | } | |
439 | ||
440 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
441 | */ | |
442 | static int allocate_file_region_entries(struct resv_map *resv, | |
443 | int regions_needed) | |
444 | __must_hold(&resv->lock) | |
445 | { | |
446 | struct list_head allocated_regions; | |
447 | int to_allocate = 0, i = 0; | |
448 | struct file_region *trg = NULL, *rg = NULL; | |
449 | ||
450 | VM_BUG_ON(regions_needed < 0); | |
451 | ||
452 | INIT_LIST_HEAD(&allocated_regions); | |
453 | ||
454 | /* | |
455 | * Check for sufficient descriptors in the cache to accommodate | |
456 | * the number of in progress add operations plus regions_needed. | |
457 | * | |
458 | * This is a while loop because when we drop the lock, some other call | |
459 | * to region_add or region_del may have consumed some region_entries, | |
460 | * so we keep looping here until we finally have enough entries for | |
461 | * (adds_in_progress + regions_needed). | |
462 | */ | |
463 | while (resv->region_cache_count < | |
464 | (resv->adds_in_progress + regions_needed)) { | |
465 | to_allocate = resv->adds_in_progress + regions_needed - | |
466 | resv->region_cache_count; | |
467 | ||
468 | /* At this point, we should have enough entries in the cache | |
f0953a1b | 469 | * for all the existing adds_in_progress. We should only be |
0db9d74e | 470 | * needing to allocate for regions_needed. |
d75c6af9 | 471 | */ |
0db9d74e MA |
472 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
473 | ||
474 | spin_unlock(&resv->lock); | |
475 | for (i = 0; i < to_allocate; i++) { | |
476 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
477 | if (!trg) | |
478 | goto out_of_memory; | |
479 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 480 | } |
d75c6af9 | 481 | |
0db9d74e MA |
482 | spin_lock(&resv->lock); |
483 | ||
d3ec7b6e WY |
484 | list_splice(&allocated_regions, &resv->region_cache); |
485 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
486 | } |
487 | ||
0db9d74e | 488 | return 0; |
d75c6af9 | 489 | |
0db9d74e MA |
490 | out_of_memory: |
491 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
492 | list_del(&rg->link); | |
493 | kfree(rg); | |
494 | } | |
495 | return -ENOMEM; | |
d75c6af9 MA |
496 | } |
497 | ||
1dd308a7 MK |
498 | /* |
499 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
500 | * map. Regions will be taken from the cache to fill in this range. |
501 | * Sufficient regions should exist in the cache due to the previous | |
502 | * call to region_chg with the same range, but in some cases the cache will not | |
503 | * have sufficient entries due to races with other code doing region_add or | |
504 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 505 | * |
0db9d74e MA |
506 | * regions_needed is the out value provided by a previous call to region_chg. |
507 | * | |
508 | * Return the number of new huge pages added to the map. This number is greater | |
509 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 510 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
511 | * region_add of regions of length 1 never allocate file_regions and cannot |
512 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
513 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 514 | */ |
0db9d74e | 515 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
516 | long in_regions_needed, struct hstate *h, |
517 | struct hugetlb_cgroup *h_cg) | |
96822904 | 518 | { |
0db9d74e | 519 | long add = 0, actual_regions_needed = 0; |
96822904 | 520 | |
7b24d861 | 521 | spin_lock(&resv->lock); |
0db9d74e MA |
522 | retry: |
523 | ||
524 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
525 | add_reservation_in_range(resv, f, t, NULL, NULL, |
526 | &actual_regions_needed); | |
96822904 | 527 | |
5e911373 | 528 | /* |
0db9d74e MA |
529 | * Check for sufficient descriptors in the cache to accommodate |
530 | * this add operation. Note that actual_regions_needed may be greater | |
531 | * than in_regions_needed, as the resv_map may have been modified since | |
532 | * the region_chg call. In this case, we need to make sure that we | |
533 | * allocate extra entries, such that we have enough for all the | |
534 | * existing adds_in_progress, plus the excess needed for this | |
535 | * operation. | |
5e911373 | 536 | */ |
0db9d74e MA |
537 | if (actual_regions_needed > in_regions_needed && |
538 | resv->region_cache_count < | |
539 | resv->adds_in_progress + | |
540 | (actual_regions_needed - in_regions_needed)) { | |
541 | /* region_add operation of range 1 should never need to | |
542 | * allocate file_region entries. | |
543 | */ | |
544 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 545 | |
0db9d74e MA |
546 | if (allocate_file_region_entries( |
547 | resv, actual_regions_needed - in_regions_needed)) { | |
548 | return -ENOMEM; | |
549 | } | |
5e911373 | 550 | |
0db9d74e | 551 | goto retry; |
5e911373 MK |
552 | } |
553 | ||
972a3da3 | 554 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
555 | |
556 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 557 | |
7b24d861 | 558 | spin_unlock(&resv->lock); |
cf3ad20b | 559 | return add; |
96822904 AW |
560 | } |
561 | ||
1dd308a7 MK |
562 | /* |
563 | * Examine the existing reserve map and determine how many | |
564 | * huge pages in the specified range [f, t) are NOT currently | |
565 | * represented. This routine is called before a subsequent | |
566 | * call to region_add that will actually modify the reserve | |
567 | * map to add the specified range [f, t). region_chg does | |
568 | * not change the number of huge pages represented by the | |
0db9d74e MA |
569 | * map. A number of new file_region structures is added to the cache as a |
570 | * placeholder, for the subsequent region_add call to use. At least 1 | |
571 | * file_region structure is added. | |
572 | * | |
573 | * out_regions_needed is the number of regions added to the | |
574 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
575 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
576 | * |
577 | * Returns the number of huge pages that need to be added to the existing | |
578 | * reservation map for the range [f, t). This number is greater or equal to | |
579 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
580 | * is needed and can not be allocated. | |
1dd308a7 | 581 | */ |
0db9d74e MA |
582 | static long region_chg(struct resv_map *resv, long f, long t, |
583 | long *out_regions_needed) | |
96822904 | 584 | { |
96822904 AW |
585 | long chg = 0; |
586 | ||
7b24d861 | 587 | spin_lock(&resv->lock); |
5e911373 | 588 | |
972a3da3 | 589 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 590 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 591 | out_regions_needed); |
5e911373 | 592 | |
0db9d74e MA |
593 | if (*out_regions_needed == 0) |
594 | *out_regions_needed = 1; | |
5e911373 | 595 | |
0db9d74e MA |
596 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
597 | return -ENOMEM; | |
5e911373 | 598 | |
0db9d74e | 599 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 600 | |
7b24d861 | 601 | spin_unlock(&resv->lock); |
96822904 AW |
602 | return chg; |
603 | } | |
604 | ||
5e911373 MK |
605 | /* |
606 | * Abort the in progress add operation. The adds_in_progress field | |
607 | * of the resv_map keeps track of the operations in progress between | |
608 | * calls to region_chg and region_add. Operations are sometimes | |
609 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
610 | * is called to decrement the adds_in_progress counter. regions_needed |
611 | * is the value returned by the region_chg call, it is used to decrement | |
612 | * the adds_in_progress counter. | |
5e911373 MK |
613 | * |
614 | * NOTE: The range arguments [f, t) are not needed or used in this | |
615 | * routine. They are kept to make reading the calling code easier as | |
616 | * arguments will match the associated region_chg call. | |
617 | */ | |
0db9d74e MA |
618 | static void region_abort(struct resv_map *resv, long f, long t, |
619 | long regions_needed) | |
5e911373 MK |
620 | { |
621 | spin_lock(&resv->lock); | |
622 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 623 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
624 | spin_unlock(&resv->lock); |
625 | } | |
626 | ||
1dd308a7 | 627 | /* |
feba16e2 MK |
628 | * Delete the specified range [f, t) from the reserve map. If the |
629 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
630 | * should be deleted. Locate the regions which intersect [f, t) | |
631 | * and either trim, delete or split the existing regions. | |
632 | * | |
633 | * Returns the number of huge pages deleted from the reserve map. | |
634 | * In the normal case, the return value is zero or more. In the | |
635 | * case where a region must be split, a new region descriptor must | |
636 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
637 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
638 | * a region and possibly return -ENOMEM. Callers specifying | |
639 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 640 | */ |
feba16e2 | 641 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 642 | { |
1406ec9b | 643 | struct list_head *head = &resv->regions; |
96822904 | 644 | struct file_region *rg, *trg; |
feba16e2 MK |
645 | struct file_region *nrg = NULL; |
646 | long del = 0; | |
96822904 | 647 | |
feba16e2 | 648 | retry: |
7b24d861 | 649 | spin_lock(&resv->lock); |
feba16e2 | 650 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
651 | /* |
652 | * Skip regions before the range to be deleted. file_region | |
653 | * ranges are normally of the form [from, to). However, there | |
654 | * may be a "placeholder" entry in the map which is of the form | |
655 | * (from, to) with from == to. Check for placeholder entries | |
656 | * at the beginning of the range to be deleted. | |
657 | */ | |
658 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 659 | continue; |
dbe409e4 | 660 | |
feba16e2 | 661 | if (rg->from >= t) |
96822904 | 662 | break; |
96822904 | 663 | |
feba16e2 MK |
664 | if (f > rg->from && t < rg->to) { /* Must split region */ |
665 | /* | |
666 | * Check for an entry in the cache before dropping | |
667 | * lock and attempting allocation. | |
668 | */ | |
669 | if (!nrg && | |
670 | resv->region_cache_count > resv->adds_in_progress) { | |
671 | nrg = list_first_entry(&resv->region_cache, | |
672 | struct file_region, | |
673 | link); | |
674 | list_del(&nrg->link); | |
675 | resv->region_cache_count--; | |
676 | } | |
96822904 | 677 | |
feba16e2 MK |
678 | if (!nrg) { |
679 | spin_unlock(&resv->lock); | |
680 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
681 | if (!nrg) | |
682 | return -ENOMEM; | |
683 | goto retry; | |
684 | } | |
685 | ||
686 | del += t - f; | |
79aa925b | 687 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 688 | resv, rg, t - f, false); |
feba16e2 MK |
689 | |
690 | /* New entry for end of split region */ | |
691 | nrg->from = t; | |
692 | nrg->to = rg->to; | |
075a61d0 MA |
693 | |
694 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
695 | ||
feba16e2 MK |
696 | INIT_LIST_HEAD(&nrg->link); |
697 | ||
698 | /* Original entry is trimmed */ | |
699 | rg->to = f; | |
700 | ||
701 | list_add(&nrg->link, &rg->link); | |
702 | nrg = NULL; | |
96822904 | 703 | break; |
feba16e2 MK |
704 | } |
705 | ||
706 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
707 | del += rg->to - rg->from; | |
075a61d0 | 708 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 709 | rg->to - rg->from, true); |
feba16e2 MK |
710 | list_del(&rg->link); |
711 | kfree(rg); | |
712 | continue; | |
713 | } | |
714 | ||
715 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 716 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 717 | t - rg->from, false); |
075a61d0 | 718 | |
79aa925b MK |
719 | del += t - rg->from; |
720 | rg->from = t; | |
721 | } else { /* Trim end of region */ | |
075a61d0 | 722 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 723 | rg->to - f, false); |
79aa925b MK |
724 | |
725 | del += rg->to - f; | |
726 | rg->to = f; | |
feba16e2 | 727 | } |
96822904 | 728 | } |
7b24d861 | 729 | |
7b24d861 | 730 | spin_unlock(&resv->lock); |
feba16e2 MK |
731 | kfree(nrg); |
732 | return del; | |
96822904 AW |
733 | } |
734 | ||
b5cec28d MK |
735 | /* |
736 | * A rare out of memory error was encountered which prevented removal of | |
737 | * the reserve map region for a page. The huge page itself was free'ed | |
738 | * and removed from the page cache. This routine will adjust the subpool | |
739 | * usage count, and the global reserve count if needed. By incrementing | |
740 | * these counts, the reserve map entry which could not be deleted will | |
741 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
742 | * counts. | |
743 | */ | |
72e2936c | 744 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
745 | { |
746 | struct hugepage_subpool *spool = subpool_inode(inode); | |
747 | long rsv_adjust; | |
da56388c | 748 | bool reserved = false; |
b5cec28d MK |
749 | |
750 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 751 | if (rsv_adjust > 0) { |
b5cec28d MK |
752 | struct hstate *h = hstate_inode(inode); |
753 | ||
da56388c ML |
754 | if (!hugetlb_acct_memory(h, 1)) |
755 | reserved = true; | |
756 | } else if (!rsv_adjust) { | |
757 | reserved = true; | |
b5cec28d | 758 | } |
da56388c ML |
759 | |
760 | if (!reserved) | |
761 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
762 | } |
763 | ||
1dd308a7 MK |
764 | /* |
765 | * Count and return the number of huge pages in the reserve map | |
766 | * that intersect with the range [f, t). | |
767 | */ | |
1406ec9b | 768 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 769 | { |
1406ec9b | 770 | struct list_head *head = &resv->regions; |
84afd99b AW |
771 | struct file_region *rg; |
772 | long chg = 0; | |
773 | ||
7b24d861 | 774 | spin_lock(&resv->lock); |
84afd99b AW |
775 | /* Locate each segment we overlap with, and count that overlap. */ |
776 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
777 | long seg_from; |
778 | long seg_to; | |
84afd99b AW |
779 | |
780 | if (rg->to <= f) | |
781 | continue; | |
782 | if (rg->from >= t) | |
783 | break; | |
784 | ||
785 | seg_from = max(rg->from, f); | |
786 | seg_to = min(rg->to, t); | |
787 | ||
788 | chg += seg_to - seg_from; | |
789 | } | |
7b24d861 | 790 | spin_unlock(&resv->lock); |
84afd99b AW |
791 | |
792 | return chg; | |
793 | } | |
794 | ||
e7c4b0bf AW |
795 | /* |
796 | * Convert the address within this vma to the page offset within | |
797 | * the mapping, in pagecache page units; huge pages here. | |
798 | */ | |
a5516438 AK |
799 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
800 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 801 | { |
a5516438 AK |
802 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
803 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
804 | } |
805 | ||
0fe6e20b NH |
806 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
807 | unsigned long address) | |
808 | { | |
809 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
810 | } | |
dee41079 | 811 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 812 | |
08fba699 MG |
813 | /* |
814 | * Return the size of the pages allocated when backing a VMA. In the majority | |
815 | * cases this will be same size as used by the page table entries. | |
816 | */ | |
817 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
818 | { | |
05ea8860 DW |
819 | if (vma->vm_ops && vma->vm_ops->pagesize) |
820 | return vma->vm_ops->pagesize(vma); | |
821 | return PAGE_SIZE; | |
08fba699 | 822 | } |
f340ca0f | 823 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 824 | |
3340289d MG |
825 | /* |
826 | * Return the page size being used by the MMU to back a VMA. In the majority | |
827 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
828 | * architectures where it differs, an architecture-specific 'strong' |
829 | * version of this symbol is required. | |
3340289d | 830 | */ |
09135cc5 | 831 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
832 | { |
833 | return vma_kernel_pagesize(vma); | |
834 | } | |
3340289d | 835 | |
84afd99b AW |
836 | /* |
837 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
838 | * bits of the reservation map pointer, which are always clear due to | |
839 | * alignment. | |
840 | */ | |
841 | #define HPAGE_RESV_OWNER (1UL << 0) | |
842 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 843 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 844 | |
a1e78772 MG |
845 | /* |
846 | * These helpers are used to track how many pages are reserved for | |
847 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
848 | * is guaranteed to have their future faults succeed. | |
849 | * | |
850 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
851 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
852 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
853 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
854 | * |
855 | * The private mapping reservation is represented in a subtly different | |
856 | * manner to a shared mapping. A shared mapping has a region map associated | |
857 | * with the underlying file, this region map represents the backing file | |
858 | * pages which have ever had a reservation assigned which this persists even | |
859 | * after the page is instantiated. A private mapping has a region map | |
860 | * associated with the original mmap which is attached to all VMAs which | |
861 | * reference it, this region map represents those offsets which have consumed | |
862 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 863 | */ |
e7c4b0bf AW |
864 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
865 | { | |
866 | return (unsigned long)vma->vm_private_data; | |
867 | } | |
868 | ||
869 | static void set_vma_private_data(struct vm_area_struct *vma, | |
870 | unsigned long value) | |
871 | { | |
872 | vma->vm_private_data = (void *)value; | |
873 | } | |
874 | ||
e9fe92ae MA |
875 | static void |
876 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
877 | struct hugetlb_cgroup *h_cg, | |
878 | struct hstate *h) | |
879 | { | |
880 | #ifdef CONFIG_CGROUP_HUGETLB | |
881 | if (!h_cg || !h) { | |
882 | resv_map->reservation_counter = NULL; | |
883 | resv_map->pages_per_hpage = 0; | |
884 | resv_map->css = NULL; | |
885 | } else { | |
886 | resv_map->reservation_counter = | |
887 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
888 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
889 | resv_map->css = &h_cg->css; | |
890 | } | |
891 | #endif | |
892 | } | |
893 | ||
9119a41e | 894 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
895 | { |
896 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
897 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
898 | ||
899 | if (!resv_map || !rg) { | |
900 | kfree(resv_map); | |
901 | kfree(rg); | |
84afd99b | 902 | return NULL; |
5e911373 | 903 | } |
84afd99b AW |
904 | |
905 | kref_init(&resv_map->refs); | |
7b24d861 | 906 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
907 | INIT_LIST_HEAD(&resv_map->regions); |
908 | ||
5e911373 | 909 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
910 | /* |
911 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
912 | * fields don't do cgroup accounting. On private mappings, these will be | |
913 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
914 | * reservations are to be un-charged from here. | |
915 | */ | |
916 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
917 | |
918 | INIT_LIST_HEAD(&resv_map->region_cache); | |
919 | list_add(&rg->link, &resv_map->region_cache); | |
920 | resv_map->region_cache_count = 1; | |
921 | ||
84afd99b AW |
922 | return resv_map; |
923 | } | |
924 | ||
9119a41e | 925 | void resv_map_release(struct kref *ref) |
84afd99b AW |
926 | { |
927 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
928 | struct list_head *head = &resv_map->region_cache; |
929 | struct file_region *rg, *trg; | |
84afd99b AW |
930 | |
931 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 932 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
933 | |
934 | /* ... and any entries left in the cache */ | |
935 | list_for_each_entry_safe(rg, trg, head, link) { | |
936 | list_del(&rg->link); | |
937 | kfree(rg); | |
938 | } | |
939 | ||
940 | VM_BUG_ON(resv_map->adds_in_progress); | |
941 | ||
84afd99b AW |
942 | kfree(resv_map); |
943 | } | |
944 | ||
4e35f483 JK |
945 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
946 | { | |
f27a5136 MK |
947 | /* |
948 | * At inode evict time, i_mapping may not point to the original | |
949 | * address space within the inode. This original address space | |
950 | * contains the pointer to the resv_map. So, always use the | |
951 | * address space embedded within the inode. | |
952 | * The VERY common case is inode->mapping == &inode->i_data but, | |
953 | * this may not be true for device special inodes. | |
954 | */ | |
955 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
956 | } |
957 | ||
84afd99b | 958 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 959 | { |
81d1b09c | 960 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
961 | if (vma->vm_flags & VM_MAYSHARE) { |
962 | struct address_space *mapping = vma->vm_file->f_mapping; | |
963 | struct inode *inode = mapping->host; | |
964 | ||
965 | return inode_resv_map(inode); | |
966 | ||
967 | } else { | |
84afd99b AW |
968 | return (struct resv_map *)(get_vma_private_data(vma) & |
969 | ~HPAGE_RESV_MASK); | |
4e35f483 | 970 | } |
a1e78772 MG |
971 | } |
972 | ||
84afd99b | 973 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 974 | { |
81d1b09c SL |
975 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
976 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 977 | |
84afd99b AW |
978 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
979 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
980 | } |
981 | ||
982 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
983 | { | |
81d1b09c SL |
984 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
985 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
986 | |
987 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
988 | } |
989 | ||
990 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
991 | { | |
81d1b09c | 992 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
993 | |
994 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
995 | } |
996 | ||
04f2cbe3 | 997 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
998 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
999 | { | |
81d1b09c | 1000 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 1001 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
1002 | vma->vm_private_data = (void *)0; |
1003 | } | |
1004 | ||
1005 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 1006 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1007 | { |
af0ed73e JK |
1008 | if (vma->vm_flags & VM_NORESERVE) { |
1009 | /* | |
1010 | * This address is already reserved by other process(chg == 0), | |
1011 | * so, we should decrement reserved count. Without decrementing, | |
1012 | * reserve count remains after releasing inode, because this | |
1013 | * allocated page will go into page cache and is regarded as | |
1014 | * coming from reserved pool in releasing step. Currently, we | |
1015 | * don't have any other solution to deal with this situation | |
1016 | * properly, so add work-around here. | |
1017 | */ | |
1018 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1019 | return true; |
af0ed73e | 1020 | else |
559ec2f8 | 1021 | return false; |
af0ed73e | 1022 | } |
a63884e9 JK |
1023 | |
1024 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1025 | if (vma->vm_flags & VM_MAYSHARE) { |
1026 | /* | |
1027 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1028 | * be a region map for all pages. The only situation where | |
1029 | * there is no region map is if a hole was punched via | |
7c8de358 | 1030 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1031 | * use. This situation is indicated if chg != 0. |
1032 | */ | |
1033 | if (chg) | |
1034 | return false; | |
1035 | else | |
1036 | return true; | |
1037 | } | |
a63884e9 JK |
1038 | |
1039 | /* | |
1040 | * Only the process that called mmap() has reserves for | |
1041 | * private mappings. | |
1042 | */ | |
67961f9d MK |
1043 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1044 | /* | |
1045 | * Like the shared case above, a hole punch or truncate | |
1046 | * could have been performed on the private mapping. | |
1047 | * Examine the value of chg to determine if reserves | |
1048 | * actually exist or were previously consumed. | |
1049 | * Very Subtle - The value of chg comes from a previous | |
1050 | * call to vma_needs_reserves(). The reserve map for | |
1051 | * private mappings has different (opposite) semantics | |
1052 | * than that of shared mappings. vma_needs_reserves() | |
1053 | * has already taken this difference in semantics into | |
1054 | * account. Therefore, the meaning of chg is the same | |
1055 | * as in the shared case above. Code could easily be | |
1056 | * combined, but keeping it separate draws attention to | |
1057 | * subtle differences. | |
1058 | */ | |
1059 | if (chg) | |
1060 | return false; | |
1061 | else | |
1062 | return true; | |
1063 | } | |
a63884e9 | 1064 | |
559ec2f8 | 1065 | return false; |
a1e78772 MG |
1066 | } |
1067 | ||
a5516438 | 1068 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1069 | { |
1070 | int nid = page_to_nid(page); | |
9487ca60 MK |
1071 | |
1072 | lockdep_assert_held(&hugetlb_lock); | |
0edaecfa | 1073 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1074 | h->free_huge_pages++; |
1075 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1076 | SetHPageFreed(page); |
1da177e4 LT |
1077 | } |
1078 | ||
94310cbc | 1079 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1080 | { |
1081 | struct page *page; | |
1a08ae36 | 1082 | bool pin = !!(current->flags & PF_MEMALLOC_PIN); |
bbe88753 | 1083 | |
9487ca60 | 1084 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 | 1085 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
8e3560d9 | 1086 | if (pin && !is_pinnable_page(page)) |
bbe88753 | 1087 | continue; |
bf50bab2 | 1088 | |
6664bfc8 WY |
1089 | if (PageHWPoison(page)) |
1090 | continue; | |
1091 | ||
1092 | list_move(&page->lru, &h->hugepage_activelist); | |
1093 | set_page_refcounted(page); | |
6c037149 | 1094 | ClearHPageFreed(page); |
6664bfc8 WY |
1095 | h->free_huge_pages--; |
1096 | h->free_huge_pages_node[nid]--; | |
1097 | return page; | |
bbe88753 JK |
1098 | } |
1099 | ||
6664bfc8 | 1100 | return NULL; |
bf50bab2 NH |
1101 | } |
1102 | ||
3e59fcb0 MH |
1103 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1104 | nodemask_t *nmask) | |
94310cbc | 1105 | { |
3e59fcb0 MH |
1106 | unsigned int cpuset_mems_cookie; |
1107 | struct zonelist *zonelist; | |
1108 | struct zone *zone; | |
1109 | struct zoneref *z; | |
98fa15f3 | 1110 | int node = NUMA_NO_NODE; |
94310cbc | 1111 | |
3e59fcb0 MH |
1112 | zonelist = node_zonelist(nid, gfp_mask); |
1113 | ||
1114 | retry_cpuset: | |
1115 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1116 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1117 | struct page *page; | |
1118 | ||
1119 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1120 | continue; | |
1121 | /* | |
1122 | * no need to ask again on the same node. Pool is node rather than | |
1123 | * zone aware | |
1124 | */ | |
1125 | if (zone_to_nid(zone) == node) | |
1126 | continue; | |
1127 | node = zone_to_nid(zone); | |
94310cbc | 1128 | |
94310cbc AK |
1129 | page = dequeue_huge_page_node_exact(h, node); |
1130 | if (page) | |
1131 | return page; | |
1132 | } | |
3e59fcb0 MH |
1133 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1134 | goto retry_cpuset; | |
1135 | ||
94310cbc AK |
1136 | return NULL; |
1137 | } | |
1138 | ||
a5516438 AK |
1139 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1140 | struct vm_area_struct *vma, | |
af0ed73e JK |
1141 | unsigned long address, int avoid_reserve, |
1142 | long chg) | |
1da177e4 | 1143 | { |
3e59fcb0 | 1144 | struct page *page; |
480eccf9 | 1145 | struct mempolicy *mpol; |
04ec6264 | 1146 | gfp_t gfp_mask; |
3e59fcb0 | 1147 | nodemask_t *nodemask; |
04ec6264 | 1148 | int nid; |
1da177e4 | 1149 | |
a1e78772 MG |
1150 | /* |
1151 | * A child process with MAP_PRIVATE mappings created by their parent | |
1152 | * have no page reserves. This check ensures that reservations are | |
1153 | * not "stolen". The child may still get SIGKILLed | |
1154 | */ | |
af0ed73e | 1155 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1156 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1157 | goto err; |
a1e78772 | 1158 | |
04f2cbe3 | 1159 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1160 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1161 | goto err; |
04f2cbe3 | 1162 | |
04ec6264 VB |
1163 | gfp_mask = htlb_alloc_mask(h); |
1164 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
1165 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
1166 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
d6995da3 | 1167 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1168 | h->resv_huge_pages--; |
1da177e4 | 1169 | } |
cc9a6c87 | 1170 | |
52cd3b07 | 1171 | mpol_cond_put(mpol); |
1da177e4 | 1172 | return page; |
cc9a6c87 MG |
1173 | |
1174 | err: | |
cc9a6c87 | 1175 | return NULL; |
1da177e4 LT |
1176 | } |
1177 | ||
1cac6f2c LC |
1178 | /* |
1179 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1180 | * We may have allocated or freed a huge page based on a different | |
1181 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1182 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1183 | * node for alloc or free. | |
1184 | */ | |
1185 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1186 | { | |
0edaf86c | 1187 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1188 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1189 | ||
1190 | return nid; | |
1191 | } | |
1192 | ||
1193 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1194 | { | |
1195 | if (!node_isset(nid, *nodes_allowed)) | |
1196 | nid = next_node_allowed(nid, nodes_allowed); | |
1197 | return nid; | |
1198 | } | |
1199 | ||
1200 | /* | |
1201 | * returns the previously saved node ["this node"] from which to | |
1202 | * allocate a persistent huge page for the pool and advance the | |
1203 | * next node from which to allocate, handling wrap at end of node | |
1204 | * mask. | |
1205 | */ | |
1206 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1207 | nodemask_t *nodes_allowed) | |
1208 | { | |
1209 | int nid; | |
1210 | ||
1211 | VM_BUG_ON(!nodes_allowed); | |
1212 | ||
1213 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1214 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1215 | ||
1216 | return nid; | |
1217 | } | |
1218 | ||
1219 | /* | |
10c6ec49 | 1220 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1221 | * node ["this node"] from which to free a huge page. Advance the |
1222 | * next node id whether or not we find a free huge page to free so | |
1223 | * that the next attempt to free addresses the next node. | |
1224 | */ | |
1225 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1226 | { | |
1227 | int nid; | |
1228 | ||
1229 | VM_BUG_ON(!nodes_allowed); | |
1230 | ||
1231 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1232 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1233 | ||
1234 | return nid; | |
1235 | } | |
1236 | ||
1237 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1238 | for (nr_nodes = nodes_weight(*mask); \ | |
1239 | nr_nodes > 0 && \ | |
1240 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1241 | nr_nodes--) | |
1242 | ||
1243 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1244 | for (nr_nodes = nodes_weight(*mask); \ | |
1245 | nr_nodes > 0 && \ | |
1246 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1247 | nr_nodes--) | |
1248 | ||
e1073d1e | 1249 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1250 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1251 | unsigned int order) |
944d9fec LC |
1252 | { |
1253 | int i; | |
1254 | int nr_pages = 1 << order; | |
1255 | struct page *p = page + 1; | |
1256 | ||
c8cc708a | 1257 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1258 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1259 | |
944d9fec | 1260 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1261 | clear_compound_head(p); |
944d9fec | 1262 | set_page_refcounted(p); |
944d9fec LC |
1263 | } |
1264 | ||
1265 | set_compound_order(page, 0); | |
ba9c1201 | 1266 | page[1].compound_nr = 0; |
944d9fec LC |
1267 | __ClearPageHead(page); |
1268 | } | |
1269 | ||
d00181b9 | 1270 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1271 | { |
cf11e85f RG |
1272 | /* |
1273 | * If the page isn't allocated using the cma allocator, | |
1274 | * cma_release() returns false. | |
1275 | */ | |
dbda8fea BS |
1276 | #ifdef CONFIG_CMA |
1277 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1278 | return; |
dbda8fea | 1279 | #endif |
cf11e85f | 1280 | |
944d9fec LC |
1281 | free_contig_range(page_to_pfn(page), 1 << order); |
1282 | } | |
1283 | ||
4eb0716e | 1284 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1285 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1286 | int nid, nodemask_t *nodemask) | |
944d9fec | 1287 | { |
04adbc3f | 1288 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1289 | if (nid == NUMA_NO_NODE) |
1290 | nid = numa_mem_id(); | |
944d9fec | 1291 | |
dbda8fea BS |
1292 | #ifdef CONFIG_CMA |
1293 | { | |
cf11e85f RG |
1294 | struct page *page; |
1295 | int node; | |
1296 | ||
953f064a LX |
1297 | if (hugetlb_cma[nid]) { |
1298 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1299 | huge_page_order(h), true); | |
cf11e85f RG |
1300 | if (page) |
1301 | return page; | |
1302 | } | |
953f064a LX |
1303 | |
1304 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1305 | for_each_node_mask(node, *nodemask) { | |
1306 | if (node == nid || !hugetlb_cma[node]) | |
1307 | continue; | |
1308 | ||
1309 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1310 | huge_page_order(h), true); | |
1311 | if (page) | |
1312 | return page; | |
1313 | } | |
1314 | } | |
cf11e85f | 1315 | } |
dbda8fea | 1316 | #endif |
cf11e85f | 1317 | |
5e27a2df | 1318 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1319 | } |
1320 | ||
1321 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1322 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1323 | #else /* !CONFIG_CONTIG_ALLOC */ |
1324 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1325 | int nid, nodemask_t *nodemask) | |
1326 | { | |
1327 | return NULL; | |
1328 | } | |
1329 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1330 | |
e1073d1e | 1331 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1332 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1333 | int nid, nodemask_t *nodemask) |
1334 | { | |
1335 | return NULL; | |
1336 | } | |
d00181b9 | 1337 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1338 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1339 | unsigned int order) { } |
944d9fec LC |
1340 | #endif |
1341 | ||
6eb4e88a MK |
1342 | /* |
1343 | * Remove hugetlb page from lists, and update dtor so that page appears | |
1344 | * as just a compound page. A reference is held on the page. | |
1345 | * | |
1346 | * Must be called with hugetlb lock held. | |
1347 | */ | |
1348 | static void remove_hugetlb_page(struct hstate *h, struct page *page, | |
1349 | bool adjust_surplus) | |
1350 | { | |
1351 | int nid = page_to_nid(page); | |
1352 | ||
1353 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1354 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1355 | ||
9487ca60 | 1356 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1357 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1358 | return; | |
1359 | ||
1360 | list_del(&page->lru); | |
1361 | ||
1362 | if (HPageFreed(page)) { | |
1363 | h->free_huge_pages--; | |
1364 | h->free_huge_pages_node[nid]--; | |
1365 | } | |
1366 | if (adjust_surplus) { | |
1367 | h->surplus_huge_pages--; | |
1368 | h->surplus_huge_pages_node[nid]--; | |
1369 | } | |
1370 | ||
1371 | set_page_refcounted(page); | |
1372 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1373 | ||
1374 | h->nr_huge_pages--; | |
1375 | h->nr_huge_pages_node[nid]--; | |
1376 | } | |
1377 | ||
a5516438 | 1378 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1379 | { |
1380 | int i; | |
dbfee5ae | 1381 | struct page *subpage = page; |
a5516438 | 1382 | |
4eb0716e | 1383 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1384 | return; |
18229df5 | 1385 | |
dbfee5ae MK |
1386 | for (i = 0; i < pages_per_huge_page(h); |
1387 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1388 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1389 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1390 | 1 << PG_active | 1 << PG_private | |
1391 | 1 << PG_writeback); | |
6af2acb6 | 1392 | } |
944d9fec LC |
1393 | if (hstate_is_gigantic(h)) { |
1394 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1395 | free_gigantic_page(page, huge_page_order(h)); | |
1396 | } else { | |
944d9fec LC |
1397 | __free_pages(page, huge_page_order(h)); |
1398 | } | |
6af2acb6 AL |
1399 | } |
1400 | ||
10c6ec49 MK |
1401 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1402 | { | |
1403 | struct page *page, *t_page; | |
1404 | ||
1405 | list_for_each_entry_safe(page, t_page, list, lru) { | |
1406 | update_and_free_page(h, page); | |
1407 | cond_resched(); | |
1408 | } | |
1409 | } | |
1410 | ||
e5ff2159 AK |
1411 | struct hstate *size_to_hstate(unsigned long size) |
1412 | { | |
1413 | struct hstate *h; | |
1414 | ||
1415 | for_each_hstate(h) { | |
1416 | if (huge_page_size(h) == size) | |
1417 | return h; | |
1418 | } | |
1419 | return NULL; | |
1420 | } | |
1421 | ||
db71ef79 | 1422 | void free_huge_page(struct page *page) |
27a85ef1 | 1423 | { |
a5516438 AK |
1424 | /* |
1425 | * Can't pass hstate in here because it is called from the | |
1426 | * compound page destructor. | |
1427 | */ | |
e5ff2159 | 1428 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1429 | int nid = page_to_nid(page); |
d6995da3 | 1430 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1431 | bool restore_reserve; |
db71ef79 | 1432 | unsigned long flags; |
27a85ef1 | 1433 | |
b4330afb MK |
1434 | VM_BUG_ON_PAGE(page_count(page), page); |
1435 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1436 | |
d6995da3 | 1437 | hugetlb_set_page_subpool(page, NULL); |
8ace22bc | 1438 | page->mapping = NULL; |
d6995da3 MK |
1439 | restore_reserve = HPageRestoreReserve(page); |
1440 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1441 | |
1c5ecae3 | 1442 | /* |
d6995da3 | 1443 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1444 | * reservation. If the page was associated with a subpool, there |
1445 | * would have been a page reserved in the subpool before allocation | |
1446 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1447 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1448 | * remove the reserved page from the subpool. |
1c5ecae3 | 1449 | */ |
0919e1b6 MK |
1450 | if (!restore_reserve) { |
1451 | /* | |
1452 | * A return code of zero implies that the subpool will be | |
1453 | * under its minimum size if the reservation is not restored | |
1454 | * after page is free. Therefore, force restore_reserve | |
1455 | * operation. | |
1456 | */ | |
1457 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1458 | restore_reserve = true; | |
1459 | } | |
1c5ecae3 | 1460 | |
db71ef79 | 1461 | spin_lock_irqsave(&hugetlb_lock, flags); |
8f251a3d | 1462 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1463 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1464 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1465 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1466 | pages_per_huge_page(h), page); | |
07443a85 JK |
1467 | if (restore_reserve) |
1468 | h->resv_huge_pages++; | |
1469 | ||
9157c311 | 1470 | if (HPageTemporary(page)) { |
6eb4e88a | 1471 | remove_hugetlb_page(h, page, false); |
db71ef79 | 1472 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
ab5ac90a MH |
1473 | update_and_free_page(h, page); |
1474 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa | 1475 | /* remove the page from active list */ |
6eb4e88a | 1476 | remove_hugetlb_page(h, page, true); |
db71ef79 | 1477 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
a5516438 | 1478 | update_and_free_page(h, page); |
7893d1d5 | 1479 | } else { |
5d3a551c | 1480 | arch_clear_hugepage_flags(page); |
a5516438 | 1481 | enqueue_huge_page(h, page); |
db71ef79 | 1482 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1483 | } |
c77c0a8a WL |
1484 | } |
1485 | ||
d3d99fcc OS |
1486 | /* |
1487 | * Must be called with the hugetlb lock held | |
1488 | */ | |
1489 | static void __prep_account_new_huge_page(struct hstate *h, int nid) | |
1490 | { | |
1491 | lockdep_assert_held(&hugetlb_lock); | |
1492 | h->nr_huge_pages++; | |
1493 | h->nr_huge_pages_node[nid]++; | |
1494 | } | |
1495 | ||
1496 | static void __prep_new_huge_page(struct page *page) | |
b7ba30c6 | 1497 | { |
0edaecfa | 1498 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1499 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1500 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1501 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1502 | set_hugetlb_cgroup_rsvd(page, NULL); |
d3d99fcc OS |
1503 | } |
1504 | ||
1505 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1506 | { | |
1507 | __prep_new_huge_page(page); | |
db71ef79 | 1508 | spin_lock_irq(&hugetlb_lock); |
d3d99fcc | 1509 | __prep_account_new_huge_page(h, nid); |
db71ef79 | 1510 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1511 | } |
1512 | ||
d00181b9 | 1513 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1514 | { |
1515 | int i; | |
1516 | int nr_pages = 1 << order; | |
1517 | struct page *p = page + 1; | |
1518 | ||
1519 | /* we rely on prep_new_huge_page to set the destructor */ | |
1520 | set_compound_order(page, order); | |
ef5a22be | 1521 | __ClearPageReserved(page); |
de09d31d | 1522 | __SetPageHead(page); |
20a0307c | 1523 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1524 | /* |
1525 | * For gigantic hugepages allocated through bootmem at | |
1526 | * boot, it's safer to be consistent with the not-gigantic | |
1527 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1528 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1529 | * pages may get the reference counting wrong if they see |
1530 | * PG_reserved set on a tail page (despite the head page not | |
1531 | * having PG_reserved set). Enforcing this consistency between | |
1532 | * head and tail pages allows drivers to optimize away a check | |
1533 | * on the head page when they need know if put_page() is needed | |
1534 | * after get_user_pages(). | |
1535 | */ | |
1536 | __ClearPageReserved(p); | |
58a84aa9 | 1537 | set_page_count(p, 0); |
1d798ca3 | 1538 | set_compound_head(p, page); |
20a0307c | 1539 | } |
b4330afb | 1540 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1541 | atomic_set(compound_pincount_ptr(page), 0); |
20a0307c WF |
1542 | } |
1543 | ||
7795912c AM |
1544 | /* |
1545 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1546 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1547 | * details. | |
1548 | */ | |
20a0307c WF |
1549 | int PageHuge(struct page *page) |
1550 | { | |
20a0307c WF |
1551 | if (!PageCompound(page)) |
1552 | return 0; | |
1553 | ||
1554 | page = compound_head(page); | |
f1e61557 | 1555 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1556 | } |
43131e14 NH |
1557 | EXPORT_SYMBOL_GPL(PageHuge); |
1558 | ||
27c73ae7 AA |
1559 | /* |
1560 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1561 | * normal or transparent huge pages. | |
1562 | */ | |
1563 | int PageHeadHuge(struct page *page_head) | |
1564 | { | |
27c73ae7 AA |
1565 | if (!PageHead(page_head)) |
1566 | return 0; | |
1567 | ||
d4af73e3 | 1568 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1569 | } |
27c73ae7 | 1570 | |
c0d0381a MK |
1571 | /* |
1572 | * Find and lock address space (mapping) in write mode. | |
1573 | * | |
336bf30e MK |
1574 | * Upon entry, the page is locked which means that page_mapping() is |
1575 | * stable. Due to locking order, we can only trylock_write. If we can | |
1576 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1577 | */ |
1578 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1579 | { | |
336bf30e | 1580 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1581 | |
c0d0381a MK |
1582 | if (!mapping) |
1583 | return mapping; | |
1584 | ||
c0d0381a MK |
1585 | if (i_mmap_trylock_write(mapping)) |
1586 | return mapping; | |
1587 | ||
336bf30e | 1588 | return NULL; |
c0d0381a MK |
1589 | } |
1590 | ||
13d60f4b ZY |
1591 | pgoff_t __basepage_index(struct page *page) |
1592 | { | |
1593 | struct page *page_head = compound_head(page); | |
1594 | pgoff_t index = page_index(page_head); | |
1595 | unsigned long compound_idx; | |
1596 | ||
1597 | if (!PageHuge(page_head)) | |
1598 | return page_index(page); | |
1599 | ||
1600 | if (compound_order(page_head) >= MAX_ORDER) | |
1601 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1602 | else | |
1603 | compound_idx = page - page_head; | |
1604 | ||
1605 | return (index << compound_order(page_head)) + compound_idx; | |
1606 | } | |
1607 | ||
0c397dae | 1608 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1609 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1610 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1611 | { |
af0fb9df | 1612 | int order = huge_page_order(h); |
1da177e4 | 1613 | struct page *page; |
f60858f9 | 1614 | bool alloc_try_hard = true; |
f96efd58 | 1615 | |
f60858f9 MK |
1616 | /* |
1617 | * By default we always try hard to allocate the page with | |
1618 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1619 | * a loop (to adjust global huge page counts) and previous allocation | |
1620 | * failed, do not continue to try hard on the same node. Use the | |
1621 | * node_alloc_noretry bitmap to manage this state information. | |
1622 | */ | |
1623 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1624 | alloc_try_hard = false; | |
1625 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1626 | if (alloc_try_hard) | |
1627 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1628 | if (nid == NUMA_NO_NODE) |
1629 | nid = numa_mem_id(); | |
84172f4b | 1630 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
af0fb9df MH |
1631 | if (page) |
1632 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1633 | else | |
1634 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1635 | |
f60858f9 MK |
1636 | /* |
1637 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1638 | * indicates an overall state change. Clear bit so that we resume | |
1639 | * normal 'try hard' allocations. | |
1640 | */ | |
1641 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1642 | node_clear(nid, *node_alloc_noretry); | |
1643 | ||
1644 | /* | |
1645 | * If we tried hard to get a page but failed, set bit so that | |
1646 | * subsequent attempts will not try as hard until there is an | |
1647 | * overall state change. | |
1648 | */ | |
1649 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1650 | node_set(nid, *node_alloc_noretry); | |
1651 | ||
63b4613c NA |
1652 | return page; |
1653 | } | |
1654 | ||
0c397dae MH |
1655 | /* |
1656 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1657 | * should use this function to get new hugetlb pages | |
1658 | */ | |
1659 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1660 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1661 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1662 | { |
1663 | struct page *page; | |
1664 | ||
1665 | if (hstate_is_gigantic(h)) | |
1666 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1667 | else | |
1668 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1669 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1670 | if (!page) |
1671 | return NULL; | |
1672 | ||
1673 | if (hstate_is_gigantic(h)) | |
1674 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1675 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1676 | ||
1677 | return page; | |
1678 | } | |
1679 | ||
af0fb9df MH |
1680 | /* |
1681 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1682 | * manner. | |
1683 | */ | |
f60858f9 MK |
1684 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1685 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1686 | { |
1687 | struct page *page; | |
1688 | int nr_nodes, node; | |
af0fb9df | 1689 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1690 | |
1691 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1692 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1693 | node_alloc_noretry); | |
af0fb9df | 1694 | if (page) |
b2261026 | 1695 | break; |
b2261026 JK |
1696 | } |
1697 | ||
af0fb9df MH |
1698 | if (!page) |
1699 | return 0; | |
b2261026 | 1700 | |
af0fb9df MH |
1701 | put_page(page); /* free it into the hugepage allocator */ |
1702 | ||
1703 | return 1; | |
b2261026 JK |
1704 | } |
1705 | ||
e8c5c824 | 1706 | /* |
10c6ec49 MK |
1707 | * Remove huge page from pool from next node to free. Attempt to keep |
1708 | * persistent huge pages more or less balanced over allowed nodes. | |
1709 | * This routine only 'removes' the hugetlb page. The caller must make | |
1710 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
1711 | * Called with hugetlb_lock locked. |
1712 | */ | |
10c6ec49 MK |
1713 | static struct page *remove_pool_huge_page(struct hstate *h, |
1714 | nodemask_t *nodes_allowed, | |
1715 | bool acct_surplus) | |
e8c5c824 | 1716 | { |
b2261026 | 1717 | int nr_nodes, node; |
10c6ec49 | 1718 | struct page *page = NULL; |
e8c5c824 | 1719 | |
9487ca60 | 1720 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 1721 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1722 | /* |
1723 | * If we're returning unused surplus pages, only examine | |
1724 | * nodes with surplus pages. | |
1725 | */ | |
b2261026 JK |
1726 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1727 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 1728 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 1729 | struct page, lru); |
6eb4e88a | 1730 | remove_hugetlb_page(h, page, acct_surplus); |
9a76db09 | 1731 | break; |
e8c5c824 | 1732 | } |
b2261026 | 1733 | } |
e8c5c824 | 1734 | |
10c6ec49 | 1735 | return page; |
e8c5c824 LS |
1736 | } |
1737 | ||
c8721bbb NH |
1738 | /* |
1739 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1740 | * nothing for in-use hugepages and non-hugepages. |
1741 | * This function returns values like below: | |
1742 | * | |
1743 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1744 | * (allocated or reserved.) | |
1745 | * 0: successfully dissolved free hugepages or the page is not a | |
1746 | * hugepage (considered as already dissolved) | |
c8721bbb | 1747 | */ |
c3114a84 | 1748 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1749 | { |
6bc9b564 | 1750 | int rc = -EBUSY; |
082d5b6b | 1751 | |
7ffddd49 | 1752 | retry: |
faf53def NH |
1753 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1754 | if (!PageHuge(page)) | |
1755 | return 0; | |
1756 | ||
db71ef79 | 1757 | spin_lock_irq(&hugetlb_lock); |
faf53def NH |
1758 | if (!PageHuge(page)) { |
1759 | rc = 0; | |
1760 | goto out; | |
1761 | } | |
1762 | ||
1763 | if (!page_count(page)) { | |
2247bb33 GS |
1764 | struct page *head = compound_head(page); |
1765 | struct hstate *h = page_hstate(head); | |
6bc9b564 | 1766 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1767 | goto out; |
7ffddd49 MS |
1768 | |
1769 | /* | |
1770 | * We should make sure that the page is already on the free list | |
1771 | * when it is dissolved. | |
1772 | */ | |
6c037149 | 1773 | if (unlikely(!HPageFreed(head))) { |
db71ef79 | 1774 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
1775 | cond_resched(); |
1776 | ||
1777 | /* | |
1778 | * Theoretically, we should return -EBUSY when we | |
1779 | * encounter this race. In fact, we have a chance | |
1780 | * to successfully dissolve the page if we do a | |
1781 | * retry. Because the race window is quite small. | |
1782 | * If we seize this opportunity, it is an optimization | |
1783 | * for increasing the success rate of dissolving page. | |
1784 | */ | |
1785 | goto retry; | |
1786 | } | |
1787 | ||
c3114a84 AK |
1788 | /* |
1789 | * Move PageHWPoison flag from head page to the raw error page, | |
1790 | * which makes any subpages rather than the error page reusable. | |
1791 | */ | |
1792 | if (PageHWPoison(head) && page != head) { | |
1793 | SetPageHWPoison(page); | |
1794 | ClearPageHWPoison(head); | |
1795 | } | |
0c5da357 | 1796 | remove_hugetlb_page(h, head, false); |
c1470b33 | 1797 | h->max_huge_pages--; |
db71ef79 | 1798 | spin_unlock_irq(&hugetlb_lock); |
2247bb33 | 1799 | update_and_free_page(h, head); |
1121828a | 1800 | return 0; |
c8721bbb | 1801 | } |
082d5b6b | 1802 | out: |
db71ef79 | 1803 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 1804 | return rc; |
c8721bbb NH |
1805 | } |
1806 | ||
1807 | /* | |
1808 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1809 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1810 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1811 | * part of it lies within the given range. | |
082d5b6b GS |
1812 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1813 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1814 | */ |
082d5b6b | 1815 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1816 | { |
c8721bbb | 1817 | unsigned long pfn; |
eb03aa00 | 1818 | struct page *page; |
082d5b6b | 1819 | int rc = 0; |
c8721bbb | 1820 | |
d0177639 | 1821 | if (!hugepages_supported()) |
082d5b6b | 1822 | return rc; |
d0177639 | 1823 | |
eb03aa00 GS |
1824 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1825 | page = pfn_to_page(pfn); | |
faf53def NH |
1826 | rc = dissolve_free_huge_page(page); |
1827 | if (rc) | |
1828 | break; | |
eb03aa00 | 1829 | } |
082d5b6b GS |
1830 | |
1831 | return rc; | |
c8721bbb NH |
1832 | } |
1833 | ||
ab5ac90a MH |
1834 | /* |
1835 | * Allocates a fresh surplus page from the page allocator. | |
1836 | */ | |
0c397dae | 1837 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1838 | int nid, nodemask_t *nmask) |
7893d1d5 | 1839 | { |
9980d744 | 1840 | struct page *page = NULL; |
7893d1d5 | 1841 | |
bae7f4ae | 1842 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1843 | return NULL; |
1844 | ||
db71ef79 | 1845 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
1846 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1847 | goto out_unlock; | |
db71ef79 | 1848 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 1849 | |
f60858f9 | 1850 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1851 | if (!page) |
0c397dae | 1852 | return NULL; |
d1c3fb1f | 1853 | |
db71ef79 | 1854 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
1855 | /* |
1856 | * We could have raced with the pool size change. | |
1857 | * Double check that and simply deallocate the new page | |
1858 | * if we would end up overcommiting the surpluses. Abuse | |
1859 | * temporary page to workaround the nasty free_huge_page | |
1860 | * codeflow | |
1861 | */ | |
1862 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 1863 | SetHPageTemporary(page); |
db71ef79 | 1864 | spin_unlock_irq(&hugetlb_lock); |
9980d744 | 1865 | put_page(page); |
2bf753e6 | 1866 | return NULL; |
9980d744 | 1867 | } else { |
9980d744 | 1868 | h->surplus_huge_pages++; |
4704dea3 | 1869 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1870 | } |
9980d744 MH |
1871 | |
1872 | out_unlock: | |
db71ef79 | 1873 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 AL |
1874 | |
1875 | return page; | |
1876 | } | |
1877 | ||
bbe88753 | 1878 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 1879 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
1880 | { |
1881 | struct page *page; | |
1882 | ||
1883 | if (hstate_is_gigantic(h)) | |
1884 | return NULL; | |
1885 | ||
f60858f9 | 1886 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1887 | if (!page) |
1888 | return NULL; | |
1889 | ||
1890 | /* | |
1891 | * We do not account these pages as surplus because they are only | |
1892 | * temporary and will be released properly on the last reference | |
1893 | */ | |
9157c311 | 1894 | SetHPageTemporary(page); |
ab5ac90a MH |
1895 | |
1896 | return page; | |
1897 | } | |
1898 | ||
099730d6 DH |
1899 | /* |
1900 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1901 | */ | |
e0ec90ee | 1902 | static |
0c397dae | 1903 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1904 | struct vm_area_struct *vma, unsigned long addr) |
1905 | { | |
aaf14e40 MH |
1906 | struct page *page; |
1907 | struct mempolicy *mpol; | |
1908 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1909 | int nid; | |
1910 | nodemask_t *nodemask; | |
1911 | ||
1912 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1913 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1914 | mpol_cond_put(mpol); |
1915 | ||
1916 | return page; | |
099730d6 DH |
1917 | } |
1918 | ||
ab5ac90a | 1919 | /* page migration callback function */ |
3e59fcb0 | 1920 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 1921 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 1922 | { |
db71ef79 | 1923 | spin_lock_irq(&hugetlb_lock); |
4db9b2ef | 1924 | if (h->free_huge_pages - h->resv_huge_pages > 0) { |
3e59fcb0 MH |
1925 | struct page *page; |
1926 | ||
1927 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1928 | if (page) { | |
db71ef79 | 1929 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 1930 | return page; |
4db9b2ef MH |
1931 | } |
1932 | } | |
db71ef79 | 1933 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 1934 | |
0c397dae | 1935 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1936 | } |
1937 | ||
ebd63723 | 1938 | /* mempolicy aware migration callback */ |
389c8178 MH |
1939 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1940 | unsigned long address) | |
ebd63723 MH |
1941 | { |
1942 | struct mempolicy *mpol; | |
1943 | nodemask_t *nodemask; | |
1944 | struct page *page; | |
ebd63723 MH |
1945 | gfp_t gfp_mask; |
1946 | int node; | |
1947 | ||
ebd63723 MH |
1948 | gfp_mask = htlb_alloc_mask(h); |
1949 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 1950 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
1951 | mpol_cond_put(mpol); |
1952 | ||
1953 | return page; | |
1954 | } | |
1955 | ||
e4e574b7 | 1956 | /* |
25985edc | 1957 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1958 | * of size 'delta'. |
1959 | */ | |
0a4f3d1b | 1960 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 1961 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
1962 | { |
1963 | struct list_head surplus_list; | |
1964 | struct page *page, *tmp; | |
0a4f3d1b LX |
1965 | int ret; |
1966 | long i; | |
1967 | long needed, allocated; | |
28073b02 | 1968 | bool alloc_ok = true; |
e4e574b7 | 1969 | |
9487ca60 | 1970 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 1971 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1972 | if (needed <= 0) { |
a5516438 | 1973 | h->resv_huge_pages += delta; |
e4e574b7 | 1974 | return 0; |
ac09b3a1 | 1975 | } |
e4e574b7 AL |
1976 | |
1977 | allocated = 0; | |
1978 | INIT_LIST_HEAD(&surplus_list); | |
1979 | ||
1980 | ret = -ENOMEM; | |
1981 | retry: | |
db71ef79 | 1982 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 1983 | for (i = 0; i < needed; i++) { |
0c397dae | 1984 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1985 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1986 | if (!page) { |
1987 | alloc_ok = false; | |
1988 | break; | |
1989 | } | |
e4e574b7 | 1990 | list_add(&page->lru, &surplus_list); |
69ed779a | 1991 | cond_resched(); |
e4e574b7 | 1992 | } |
28073b02 | 1993 | allocated += i; |
e4e574b7 AL |
1994 | |
1995 | /* | |
1996 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1997 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1998 | */ | |
db71ef79 | 1999 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
2000 | needed = (h->resv_huge_pages + delta) - |
2001 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2002 | if (needed > 0) { |
2003 | if (alloc_ok) | |
2004 | goto retry; | |
2005 | /* | |
2006 | * We were not able to allocate enough pages to | |
2007 | * satisfy the entire reservation so we free what | |
2008 | * we've allocated so far. | |
2009 | */ | |
2010 | goto free; | |
2011 | } | |
e4e574b7 AL |
2012 | /* |
2013 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2014 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2015 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2016 | * allocator. Commit the entire reservation here to prevent another |
2017 | * process from stealing the pages as they are added to the pool but | |
2018 | * before they are reserved. | |
e4e574b7 AL |
2019 | */ |
2020 | needed += allocated; | |
a5516438 | 2021 | h->resv_huge_pages += delta; |
e4e574b7 | 2022 | ret = 0; |
a9869b83 | 2023 | |
19fc3f0a | 2024 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2025 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
e558464b MS |
2026 | int zeroed; |
2027 | ||
19fc3f0a AL |
2028 | if ((--needed) < 0) |
2029 | break; | |
a9869b83 NH |
2030 | /* |
2031 | * This page is now managed by the hugetlb allocator and has | |
2032 | * no users -- drop the buddy allocator's reference. | |
2033 | */ | |
e558464b MS |
2034 | zeroed = put_page_testzero(page); |
2035 | VM_BUG_ON_PAGE(!zeroed, page); | |
a5516438 | 2036 | enqueue_huge_page(h, page); |
19fc3f0a | 2037 | } |
28073b02 | 2038 | free: |
db71ef79 | 2039 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a AL |
2040 | |
2041 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
2042 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
2043 | put_page(page); | |
db71ef79 | 2044 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2045 | |
2046 | return ret; | |
2047 | } | |
2048 | ||
2049 | /* | |
e5bbc8a6 MK |
2050 | * This routine has two main purposes: |
2051 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2052 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2053 | * to the associated reservation map. | |
2054 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2055 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2056 | */ |
a5516438 AK |
2057 | static void return_unused_surplus_pages(struct hstate *h, |
2058 | unsigned long unused_resv_pages) | |
e4e574b7 | 2059 | { |
e4e574b7 | 2060 | unsigned long nr_pages; |
10c6ec49 MK |
2061 | struct page *page; |
2062 | LIST_HEAD(page_list); | |
2063 | ||
9487ca60 | 2064 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2065 | /* Uncommit the reservation */ |
2066 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2067 | |
aa888a74 | 2068 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2069 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2070 | goto out; |
aa888a74 | 2071 | |
e5bbc8a6 MK |
2072 | /* |
2073 | * Part (or even all) of the reservation could have been backed | |
2074 | * by pre-allocated pages. Only free surplus pages. | |
2075 | */ | |
a5516438 | 2076 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2077 | |
685f3457 LS |
2078 | /* |
2079 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2080 | * evenly across all nodes with memory. Iterate across these nodes |
2081 | * until we can no longer free unreserved surplus pages. This occurs | |
2082 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2083 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2084 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2085 | */ |
2086 | while (nr_pages--) { | |
10c6ec49 MK |
2087 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2088 | if (!page) | |
e5bbc8a6 | 2089 | goto out; |
10c6ec49 MK |
2090 | |
2091 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2092 | } |
e5bbc8a6 MK |
2093 | |
2094 | out: | |
db71ef79 | 2095 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2096 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2097 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2098 | } |
2099 | ||
5e911373 | 2100 | |
c37f9fb1 | 2101 | /* |
feba16e2 | 2102 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2103 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2104 | * |
2105 | * vma_needs_reservation is called to determine if the huge page at addr | |
2106 | * within the vma has an associated reservation. If a reservation is | |
2107 | * needed, the value 1 is returned. The caller is then responsible for | |
2108 | * managing the global reservation and subpool usage counts. After | |
2109 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2110 | * to add the page to the reservation map. If the page allocation fails, |
2111 | * the reservation must be ended instead of committed. vma_end_reservation | |
2112 | * is called in such cases. | |
cf3ad20b MK |
2113 | * |
2114 | * In the normal case, vma_commit_reservation returns the same value | |
2115 | * as the preceding vma_needs_reservation call. The only time this | |
2116 | * is not the case is if a reserve map was changed between calls. It | |
2117 | * is the responsibility of the caller to notice the difference and | |
2118 | * take appropriate action. | |
96b96a96 MK |
2119 | * |
2120 | * vma_add_reservation is used in error paths where a reservation must | |
2121 | * be restored when a newly allocated huge page must be freed. It is | |
2122 | * to be called after calling vma_needs_reservation to determine if a | |
2123 | * reservation exists. | |
846be085 MK |
2124 | * |
2125 | * vma_del_reservation is used in error paths where an entry in the reserve | |
2126 | * map was created during huge page allocation and must be removed. It is to | |
2127 | * be called after calling vma_needs_reservation to determine if a reservation | |
2128 | * exists. | |
c37f9fb1 | 2129 | */ |
5e911373 MK |
2130 | enum vma_resv_mode { |
2131 | VMA_NEEDS_RESV, | |
2132 | VMA_COMMIT_RESV, | |
feba16e2 | 2133 | VMA_END_RESV, |
96b96a96 | 2134 | VMA_ADD_RESV, |
846be085 | 2135 | VMA_DEL_RESV, |
5e911373 | 2136 | }; |
cf3ad20b MK |
2137 | static long __vma_reservation_common(struct hstate *h, |
2138 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2139 | enum vma_resv_mode mode) |
c37f9fb1 | 2140 | { |
4e35f483 JK |
2141 | struct resv_map *resv; |
2142 | pgoff_t idx; | |
cf3ad20b | 2143 | long ret; |
0db9d74e | 2144 | long dummy_out_regions_needed; |
c37f9fb1 | 2145 | |
4e35f483 JK |
2146 | resv = vma_resv_map(vma); |
2147 | if (!resv) | |
84afd99b | 2148 | return 1; |
c37f9fb1 | 2149 | |
4e35f483 | 2150 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2151 | switch (mode) { |
2152 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2153 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2154 | /* We assume that vma_reservation_* routines always operate on | |
2155 | * 1 page, and that adding to resv map a 1 page entry can only | |
2156 | * ever require 1 region. | |
2157 | */ | |
2158 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2159 | break; |
2160 | case VMA_COMMIT_RESV: | |
075a61d0 | 2161 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2162 | /* region_add calls of range 1 should never fail. */ |
2163 | VM_BUG_ON(ret < 0); | |
5e911373 | 2164 | break; |
feba16e2 | 2165 | case VMA_END_RESV: |
0db9d74e | 2166 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2167 | ret = 0; |
2168 | break; | |
96b96a96 | 2169 | case VMA_ADD_RESV: |
0db9d74e | 2170 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2171 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2172 | /* region_add calls of range 1 should never fail. */ |
2173 | VM_BUG_ON(ret < 0); | |
2174 | } else { | |
2175 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2176 | ret = region_del(resv, idx, idx + 1); |
2177 | } | |
2178 | break; | |
846be085 MK |
2179 | case VMA_DEL_RESV: |
2180 | if (vma->vm_flags & VM_MAYSHARE) { | |
2181 | region_abort(resv, idx, idx + 1, 1); | |
2182 | ret = region_del(resv, idx, idx + 1); | |
2183 | } else { | |
2184 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2185 | /* region_add calls of range 1 should never fail. */ | |
2186 | VM_BUG_ON(ret < 0); | |
2187 | } | |
2188 | break; | |
5e911373 MK |
2189 | default: |
2190 | BUG(); | |
2191 | } | |
84afd99b | 2192 | |
846be085 | 2193 | if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) |
cf3ad20b | 2194 | return ret; |
bf3d12b9 ML |
2195 | /* |
2196 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2197 | * | |
2198 | * In most cases, reserves always exist for private mappings. | |
2199 | * However, a file associated with mapping could have been | |
2200 | * hole punched or truncated after reserves were consumed. | |
2201 | * As subsequent fault on such a range will not use reserves. | |
2202 | * Subtle - The reserve map for private mappings has the | |
2203 | * opposite meaning than that of shared mappings. If NO | |
2204 | * entry is in the reserve map, it means a reservation exists. | |
2205 | * If an entry exists in the reserve map, it means the | |
2206 | * reservation has already been consumed. As a result, the | |
2207 | * return value of this routine is the opposite of the | |
2208 | * value returned from reserve map manipulation routines above. | |
2209 | */ | |
2210 | if (ret > 0) | |
2211 | return 0; | |
2212 | if (ret == 0) | |
2213 | return 1; | |
2214 | return ret; | |
c37f9fb1 | 2215 | } |
cf3ad20b MK |
2216 | |
2217 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2218 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2219 | { |
5e911373 | 2220 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2221 | } |
84afd99b | 2222 | |
cf3ad20b MK |
2223 | static long vma_commit_reservation(struct hstate *h, |
2224 | struct vm_area_struct *vma, unsigned long addr) | |
2225 | { | |
5e911373 MK |
2226 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2227 | } | |
2228 | ||
feba16e2 | 2229 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2230 | struct vm_area_struct *vma, unsigned long addr) |
2231 | { | |
feba16e2 | 2232 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2233 | } |
2234 | ||
96b96a96 MK |
2235 | static long vma_add_reservation(struct hstate *h, |
2236 | struct vm_area_struct *vma, unsigned long addr) | |
2237 | { | |
2238 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2239 | } | |
2240 | ||
846be085 MK |
2241 | static long vma_del_reservation(struct hstate *h, |
2242 | struct vm_area_struct *vma, unsigned long addr) | |
2243 | { | |
2244 | return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); | |
2245 | } | |
2246 | ||
96b96a96 | 2247 | /* |
846be085 MK |
2248 | * This routine is called to restore reservation information on error paths. |
2249 | * It should ONLY be called for pages allocated via alloc_huge_page(), and | |
2250 | * the hugetlb mutex should remain held when calling this routine. | |
2251 | * | |
2252 | * It handles two specific cases: | |
2253 | * 1) A reservation was in place and the page consumed the reservation. | |
2254 | * HPageRestoreReserve is set in the page. | |
2255 | * 2) No reservation was in place for the page, so HPageRestoreReserve is | |
2256 | * not set. However, alloc_huge_page always updates the reserve map. | |
2257 | * | |
2258 | * In case 1, free_huge_page later in the error path will increment the | |
2259 | * global reserve count. But, free_huge_page does not have enough context | |
2260 | * to adjust the reservation map. This case deals primarily with private | |
2261 | * mappings. Adjust the reserve map here to be consistent with global | |
2262 | * reserve count adjustments to be made by free_huge_page. Make sure the | |
2263 | * reserve map indicates there is a reservation present. | |
2264 | * | |
2265 | * In case 2, simply undo reserve map modifications done by alloc_huge_page. | |
96b96a96 | 2266 | */ |
846be085 MK |
2267 | void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, |
2268 | unsigned long address, struct page *page) | |
96b96a96 | 2269 | { |
846be085 | 2270 | long rc = vma_needs_reservation(h, vma, address); |
96b96a96 | 2271 | |
846be085 MK |
2272 | if (HPageRestoreReserve(page)) { |
2273 | if (unlikely(rc < 0)) | |
96b96a96 MK |
2274 | /* |
2275 | * Rare out of memory condition in reserve map | |
d6995da3 | 2276 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2277 | * global reserve count will not be incremented |
2278 | * by free_huge_page. This will make it appear | |
2279 | * as though the reservation for this page was | |
2280 | * consumed. This may prevent the task from | |
2281 | * faulting in the page at a later time. This | |
2282 | * is better than inconsistent global huge page | |
2283 | * accounting of reserve counts. | |
2284 | */ | |
d6995da3 | 2285 | ClearHPageRestoreReserve(page); |
846be085 MK |
2286 | else if (rc) |
2287 | (void)vma_add_reservation(h, vma, address); | |
2288 | else | |
2289 | vma_end_reservation(h, vma, address); | |
2290 | } else { | |
2291 | if (!rc) { | |
2292 | /* | |
2293 | * This indicates there is an entry in the reserve map | |
2294 | * added by alloc_huge_page. We know it was added | |
2295 | * before the alloc_huge_page call, otherwise | |
2296 | * HPageRestoreReserve would be set on the page. | |
2297 | * Remove the entry so that a subsequent allocation | |
2298 | * does not consume a reservation. | |
2299 | */ | |
2300 | rc = vma_del_reservation(h, vma, address); | |
2301 | if (rc < 0) | |
96b96a96 | 2302 | /* |
846be085 MK |
2303 | * VERY rare out of memory condition. Since |
2304 | * we can not delete the entry, set | |
2305 | * HPageRestoreReserve so that the reserve | |
2306 | * count will be incremented when the page | |
2307 | * is freed. This reserve will be consumed | |
2308 | * on a subsequent allocation. | |
96b96a96 | 2309 | */ |
846be085 MK |
2310 | SetHPageRestoreReserve(page); |
2311 | } else if (rc < 0) { | |
2312 | /* | |
2313 | * Rare out of memory condition from | |
2314 | * vma_needs_reservation call. Memory allocation is | |
2315 | * only attempted if a new entry is needed. Therefore, | |
2316 | * this implies there is not an entry in the | |
2317 | * reserve map. | |
2318 | * | |
2319 | * For shared mappings, no entry in the map indicates | |
2320 | * no reservation. We are done. | |
2321 | */ | |
2322 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
2323 | /* | |
2324 | * For private mappings, no entry indicates | |
2325 | * a reservation is present. Since we can | |
2326 | * not add an entry, set SetHPageRestoreReserve | |
2327 | * on the page so reserve count will be | |
2328 | * incremented when freed. This reserve will | |
2329 | * be consumed on a subsequent allocation. | |
2330 | */ | |
2331 | SetHPageRestoreReserve(page); | |
96b96a96 | 2332 | } else |
846be085 MK |
2333 | /* |
2334 | * No reservation present, do nothing | |
2335 | */ | |
2336 | vma_end_reservation(h, vma, address); | |
96b96a96 MK |
2337 | } |
2338 | } | |
2339 | ||
369fa227 OS |
2340 | /* |
2341 | * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one | |
2342 | * @h: struct hstate old page belongs to | |
2343 | * @old_page: Old page to dissolve | |
ae37c7ff | 2344 | * @list: List to isolate the page in case we need to |
369fa227 OS |
2345 | * Returns 0 on success, otherwise negated error. |
2346 | */ | |
ae37c7ff OS |
2347 | static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, |
2348 | struct list_head *list) | |
369fa227 OS |
2349 | { |
2350 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
2351 | int nid = page_to_nid(old_page); | |
2352 | struct page *new_page; | |
2353 | int ret = 0; | |
2354 | ||
2355 | /* | |
2356 | * Before dissolving the page, we need to allocate a new one for the | |
2357 | * pool to remain stable. Using alloc_buddy_huge_page() allows us to | |
2358 | * not having to deal with prep_new_huge_page() and avoids dealing of any | |
2359 | * counters. This simplifies and let us do the whole thing under the | |
2360 | * lock. | |
2361 | */ | |
2362 | new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); | |
2363 | if (!new_page) | |
2364 | return -ENOMEM; | |
2365 | ||
2366 | retry: | |
2367 | spin_lock_irq(&hugetlb_lock); | |
2368 | if (!PageHuge(old_page)) { | |
2369 | /* | |
2370 | * Freed from under us. Drop new_page too. | |
2371 | */ | |
2372 | goto free_new; | |
2373 | } else if (page_count(old_page)) { | |
2374 | /* | |
ae37c7ff OS |
2375 | * Someone has grabbed the page, try to isolate it here. |
2376 | * Fail with -EBUSY if not possible. | |
369fa227 | 2377 | */ |
ae37c7ff OS |
2378 | spin_unlock_irq(&hugetlb_lock); |
2379 | if (!isolate_huge_page(old_page, list)) | |
2380 | ret = -EBUSY; | |
2381 | spin_lock_irq(&hugetlb_lock); | |
369fa227 OS |
2382 | goto free_new; |
2383 | } else if (!HPageFreed(old_page)) { | |
2384 | /* | |
2385 | * Page's refcount is 0 but it has not been enqueued in the | |
2386 | * freelist yet. Race window is small, so we can succeed here if | |
2387 | * we retry. | |
2388 | */ | |
2389 | spin_unlock_irq(&hugetlb_lock); | |
2390 | cond_resched(); | |
2391 | goto retry; | |
2392 | } else { | |
2393 | /* | |
2394 | * Ok, old_page is still a genuine free hugepage. Remove it from | |
2395 | * the freelist and decrease the counters. These will be | |
2396 | * incremented again when calling __prep_account_new_huge_page() | |
2397 | * and enqueue_huge_page() for new_page. The counters will remain | |
2398 | * stable since this happens under the lock. | |
2399 | */ | |
2400 | remove_hugetlb_page(h, old_page, false); | |
2401 | ||
2402 | /* | |
2403 | * new_page needs to be initialized with the standard hugetlb | |
2404 | * state. This is normally done by prep_new_huge_page() but | |
2405 | * that takes hugetlb_lock which is already held so we need to | |
2406 | * open code it here. | |
2407 | * Reference count trick is needed because allocator gives us | |
2408 | * referenced page but the pool requires pages with 0 refcount. | |
2409 | */ | |
2410 | __prep_new_huge_page(new_page); | |
2411 | __prep_account_new_huge_page(h, nid); | |
2412 | page_ref_dec(new_page); | |
2413 | enqueue_huge_page(h, new_page); | |
2414 | ||
2415 | /* | |
2416 | * Pages have been replaced, we can safely free the old one. | |
2417 | */ | |
2418 | spin_unlock_irq(&hugetlb_lock); | |
2419 | update_and_free_page(h, old_page); | |
2420 | } | |
2421 | ||
2422 | return ret; | |
2423 | ||
2424 | free_new: | |
2425 | spin_unlock_irq(&hugetlb_lock); | |
2426 | __free_pages(new_page, huge_page_order(h)); | |
2427 | ||
2428 | return ret; | |
2429 | } | |
2430 | ||
ae37c7ff | 2431 | int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) |
369fa227 OS |
2432 | { |
2433 | struct hstate *h; | |
2434 | struct page *head; | |
ae37c7ff | 2435 | int ret = -EBUSY; |
369fa227 OS |
2436 | |
2437 | /* | |
2438 | * The page might have been dissolved from under our feet, so make sure | |
2439 | * to carefully check the state under the lock. | |
2440 | * Return success when racing as if we dissolved the page ourselves. | |
2441 | */ | |
2442 | spin_lock_irq(&hugetlb_lock); | |
2443 | if (PageHuge(page)) { | |
2444 | head = compound_head(page); | |
2445 | h = page_hstate(head); | |
2446 | } else { | |
2447 | spin_unlock_irq(&hugetlb_lock); | |
2448 | return 0; | |
2449 | } | |
2450 | spin_unlock_irq(&hugetlb_lock); | |
2451 | ||
2452 | /* | |
2453 | * Fence off gigantic pages as there is a cyclic dependency between | |
2454 | * alloc_contig_range and them. Return -ENOMEM as this has the effect | |
2455 | * of bailing out right away without further retrying. | |
2456 | */ | |
2457 | if (hstate_is_gigantic(h)) | |
2458 | return -ENOMEM; | |
2459 | ||
ae37c7ff OS |
2460 | if (page_count(head) && isolate_huge_page(head, list)) |
2461 | ret = 0; | |
2462 | else if (!page_count(head)) | |
2463 | ret = alloc_and_dissolve_huge_page(h, head, list); | |
2464 | ||
2465 | return ret; | |
369fa227 OS |
2466 | } |
2467 | ||
70c3547e | 2468 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2469 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2470 | { |
90481622 | 2471 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2472 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2473 | struct page *page; |
d85f69b0 MK |
2474 | long map_chg, map_commit; |
2475 | long gbl_chg; | |
6d76dcf4 AK |
2476 | int ret, idx; |
2477 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2478 | bool deferred_reserve; |
a1e78772 | 2479 | |
6d76dcf4 | 2480 | idx = hstate_index(h); |
a1e78772 | 2481 | /* |
d85f69b0 MK |
2482 | * Examine the region/reserve map to determine if the process |
2483 | * has a reservation for the page to be allocated. A return | |
2484 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2485 | */ |
d85f69b0 MK |
2486 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2487 | if (map_chg < 0) | |
76dcee75 | 2488 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2489 | |
2490 | /* | |
2491 | * Processes that did not create the mapping will have no | |
2492 | * reserves as indicated by the region/reserve map. Check | |
2493 | * that the allocation will not exceed the subpool limit. | |
2494 | * Allocations for MAP_NORESERVE mappings also need to be | |
2495 | * checked against any subpool limit. | |
2496 | */ | |
2497 | if (map_chg || avoid_reserve) { | |
2498 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2499 | if (gbl_chg < 0) { | |
feba16e2 | 2500 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2501 | return ERR_PTR(-ENOSPC); |
5e911373 | 2502 | } |
1da177e4 | 2503 | |
d85f69b0 MK |
2504 | /* |
2505 | * Even though there was no reservation in the region/reserve | |
2506 | * map, there could be reservations associated with the | |
2507 | * subpool that can be used. This would be indicated if the | |
2508 | * return value of hugepage_subpool_get_pages() is zero. | |
2509 | * However, if avoid_reserve is specified we still avoid even | |
2510 | * the subpool reservations. | |
2511 | */ | |
2512 | if (avoid_reserve) | |
2513 | gbl_chg = 1; | |
2514 | } | |
2515 | ||
08cf9faf MA |
2516 | /* If this allocation is not consuming a reservation, charge it now. |
2517 | */ | |
6501fe5f | 2518 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2519 | if (deferred_reserve) { |
2520 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2521 | idx, pages_per_huge_page(h), &h_cg); | |
2522 | if (ret) | |
2523 | goto out_subpool_put; | |
2524 | } | |
2525 | ||
6d76dcf4 | 2526 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2527 | if (ret) |
08cf9faf | 2528 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2529 | |
db71ef79 | 2530 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
2531 | /* |
2532 | * glb_chg is passed to indicate whether or not a page must be taken | |
2533 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2534 | * a reservation exists for the allocation. | |
2535 | */ | |
2536 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2537 | if (!page) { |
db71ef79 | 2538 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 2539 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2540 | if (!page) |
2541 | goto out_uncharge_cgroup; | |
a88c7695 | 2542 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2543 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2544 | h->resv_huge_pages--; |
2545 | } | |
db71ef79 | 2546 | spin_lock_irq(&hugetlb_lock); |
15a8d68e | 2547 | list_add(&page->lru, &h->hugepage_activelist); |
81a6fcae | 2548 | /* Fall through */ |
68842c9b | 2549 | } |
81a6fcae | 2550 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2551 | /* If allocation is not consuming a reservation, also store the |
2552 | * hugetlb_cgroup pointer on the page. | |
2553 | */ | |
2554 | if (deferred_reserve) { | |
2555 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2556 | h_cg, page); | |
2557 | } | |
2558 | ||
db71ef79 | 2559 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 2560 | |
d6995da3 | 2561 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2562 | |
d85f69b0 MK |
2563 | map_commit = vma_commit_reservation(h, vma, addr); |
2564 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2565 | /* |
2566 | * The page was added to the reservation map between | |
2567 | * vma_needs_reservation and vma_commit_reservation. | |
2568 | * This indicates a race with hugetlb_reserve_pages. | |
2569 | * Adjust for the subpool count incremented above AND | |
2570 | * in hugetlb_reserve_pages for the same page. Also, | |
2571 | * the reservation count added in hugetlb_reserve_pages | |
2572 | * no longer applies. | |
2573 | */ | |
2574 | long rsv_adjust; | |
2575 | ||
2576 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2577 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2578 | if (deferred_reserve) |
2579 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2580 | pages_per_huge_page(h), page); | |
33039678 | 2581 | } |
90d8b7e6 | 2582 | return page; |
8f34af6f JZ |
2583 | |
2584 | out_uncharge_cgroup: | |
2585 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2586 | out_uncharge_cgroup_reservation: |
2587 | if (deferred_reserve) | |
2588 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2589 | h_cg); | |
8f34af6f | 2590 | out_subpool_put: |
d85f69b0 | 2591 | if (map_chg || avoid_reserve) |
8f34af6f | 2592 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2593 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2594 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2595 | } |
2596 | ||
e24a1307 AK |
2597 | int alloc_bootmem_huge_page(struct hstate *h) |
2598 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2599 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2600 | { |
2601 | struct huge_bootmem_page *m; | |
b2261026 | 2602 | int nr_nodes, node; |
aa888a74 | 2603 | |
b2261026 | 2604 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2605 | void *addr; |
2606 | ||
eb31d559 | 2607 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2608 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2609 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2610 | if (addr) { |
2611 | /* | |
2612 | * Use the beginning of the huge page to store the | |
2613 | * huge_bootmem_page struct (until gather_bootmem | |
2614 | * puts them into the mem_map). | |
2615 | */ | |
2616 | m = addr; | |
91f47662 | 2617 | goto found; |
aa888a74 | 2618 | } |
aa888a74 AK |
2619 | } |
2620 | return 0; | |
2621 | ||
2622 | found: | |
df994ead | 2623 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2624 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2625 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2626 | list_add(&m->list, &huge_boot_pages); |
2627 | m->hstate = h; | |
2628 | return 1; | |
2629 | } | |
2630 | ||
d00181b9 KS |
2631 | static void __init prep_compound_huge_page(struct page *page, |
2632 | unsigned int order) | |
18229df5 AW |
2633 | { |
2634 | if (unlikely(order > (MAX_ORDER - 1))) | |
2635 | prep_compound_gigantic_page(page, order); | |
2636 | else | |
2637 | prep_compound_page(page, order); | |
2638 | } | |
2639 | ||
aa888a74 AK |
2640 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2641 | static void __init gather_bootmem_prealloc(void) | |
2642 | { | |
2643 | struct huge_bootmem_page *m; | |
2644 | ||
2645 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2646 | struct page *page = virt_to_page(m); |
aa888a74 | 2647 | struct hstate *h = m->hstate; |
ee8f248d | 2648 | |
aa888a74 | 2649 | WARN_ON(page_count(page) != 1); |
c78a7f36 | 2650 | prep_compound_huge_page(page, huge_page_order(h)); |
ef5a22be | 2651 | WARN_ON(PageReserved(page)); |
aa888a74 | 2652 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2653 | put_page(page); /* free it into the hugepage allocator */ |
2654 | ||
b0320c7b RA |
2655 | /* |
2656 | * If we had gigantic hugepages allocated at boot time, we need | |
2657 | * to restore the 'stolen' pages to totalram_pages in order to | |
2658 | * fix confusing memory reports from free(1) and another | |
2659 | * side-effects, like CommitLimit going negative. | |
2660 | */ | |
bae7f4ae | 2661 | if (hstate_is_gigantic(h)) |
c78a7f36 | 2662 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 2663 | cond_resched(); |
aa888a74 AK |
2664 | } |
2665 | } | |
2666 | ||
8faa8b07 | 2667 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2668 | { |
2669 | unsigned long i; | |
f60858f9 MK |
2670 | nodemask_t *node_alloc_noretry; |
2671 | ||
2672 | if (!hstate_is_gigantic(h)) { | |
2673 | /* | |
2674 | * Bit mask controlling how hard we retry per-node allocations. | |
2675 | * Ignore errors as lower level routines can deal with | |
2676 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2677 | * time, we are likely in bigger trouble. | |
2678 | */ | |
2679 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2680 | GFP_KERNEL); | |
2681 | } else { | |
2682 | /* allocations done at boot time */ | |
2683 | node_alloc_noretry = NULL; | |
2684 | } | |
2685 | ||
2686 | /* bit mask controlling how hard we retry per-node allocations */ | |
2687 | if (node_alloc_noretry) | |
2688 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2689 | |
e5ff2159 | 2690 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2691 | if (hstate_is_gigantic(h)) { |
dbda8fea | 2692 | if (hugetlb_cma_size) { |
cf11e85f | 2693 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); |
7ecc9565 | 2694 | goto free; |
cf11e85f | 2695 | } |
aa888a74 AK |
2696 | if (!alloc_bootmem_huge_page(h)) |
2697 | break; | |
0c397dae | 2698 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2699 | &node_states[N_MEMORY], |
2700 | node_alloc_noretry)) | |
1da177e4 | 2701 | break; |
69ed779a | 2702 | cond_resched(); |
1da177e4 | 2703 | } |
d715cf80 LH |
2704 | if (i < h->max_huge_pages) { |
2705 | char buf[32]; | |
2706 | ||
c6247f72 | 2707 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2708 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2709 | h->max_huge_pages, buf, i); | |
2710 | h->max_huge_pages = i; | |
2711 | } | |
7ecc9565 | 2712 | free: |
f60858f9 | 2713 | kfree(node_alloc_noretry); |
e5ff2159 AK |
2714 | } |
2715 | ||
2716 | static void __init hugetlb_init_hstates(void) | |
2717 | { | |
2718 | struct hstate *h; | |
2719 | ||
2720 | for_each_hstate(h) { | |
641844f5 NH |
2721 | if (minimum_order > huge_page_order(h)) |
2722 | minimum_order = huge_page_order(h); | |
2723 | ||
8faa8b07 | 2724 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2725 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2726 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2727 | } |
641844f5 | 2728 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2729 | } |
2730 | ||
2731 | static void __init report_hugepages(void) | |
2732 | { | |
2733 | struct hstate *h; | |
2734 | ||
2735 | for_each_hstate(h) { | |
4abd32db | 2736 | char buf[32]; |
c6247f72 MW |
2737 | |
2738 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2739 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2740 | buf, h->free_huge_pages); |
e5ff2159 AK |
2741 | } |
2742 | } | |
2743 | ||
1da177e4 | 2744 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2745 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2746 | nodemask_t *nodes_allowed) | |
1da177e4 | 2747 | { |
4415cc8d | 2748 | int i; |
1121828a | 2749 | LIST_HEAD(page_list); |
4415cc8d | 2750 | |
9487ca60 | 2751 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 2752 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2753 | return; |
2754 | ||
1121828a MK |
2755 | /* |
2756 | * Collect pages to be freed on a list, and free after dropping lock | |
2757 | */ | |
6ae11b27 | 2758 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 2759 | struct page *page, *next; |
a5516438 AK |
2760 | struct list_head *freel = &h->hugepage_freelists[i]; |
2761 | list_for_each_entry_safe(page, next, freel, lru) { | |
2762 | if (count >= h->nr_huge_pages) | |
1121828a | 2763 | goto out; |
1da177e4 LT |
2764 | if (PageHighMem(page)) |
2765 | continue; | |
6eb4e88a | 2766 | remove_hugetlb_page(h, page, false); |
1121828a | 2767 | list_add(&page->lru, &page_list); |
1da177e4 LT |
2768 | } |
2769 | } | |
1121828a MK |
2770 | |
2771 | out: | |
db71ef79 | 2772 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2773 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2774 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
2775 | } |
2776 | #else | |
6ae11b27 LS |
2777 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2778 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2779 | { |
2780 | } | |
2781 | #endif | |
2782 | ||
20a0307c WF |
2783 | /* |
2784 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2785 | * balanced by operating on them in a round-robin fashion. | |
2786 | * Returns 1 if an adjustment was made. | |
2787 | */ | |
6ae11b27 LS |
2788 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2789 | int delta) | |
20a0307c | 2790 | { |
b2261026 | 2791 | int nr_nodes, node; |
20a0307c | 2792 | |
9487ca60 | 2793 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 2794 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 2795 | |
b2261026 JK |
2796 | if (delta < 0) { |
2797 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2798 | if (h->surplus_huge_pages_node[node]) | |
2799 | goto found; | |
e8c5c824 | 2800 | } |
b2261026 JK |
2801 | } else { |
2802 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2803 | if (h->surplus_huge_pages_node[node] < | |
2804 | h->nr_huge_pages_node[node]) | |
2805 | goto found; | |
e8c5c824 | 2806 | } |
b2261026 JK |
2807 | } |
2808 | return 0; | |
20a0307c | 2809 | |
b2261026 JK |
2810 | found: |
2811 | h->surplus_huge_pages += delta; | |
2812 | h->surplus_huge_pages_node[node] += delta; | |
2813 | return 1; | |
20a0307c WF |
2814 | } |
2815 | ||
a5516438 | 2816 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2817 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2818 | nodemask_t *nodes_allowed) |
1da177e4 | 2819 | { |
7893d1d5 | 2820 | unsigned long min_count, ret; |
10c6ec49 MK |
2821 | struct page *page; |
2822 | LIST_HEAD(page_list); | |
f60858f9 MK |
2823 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2824 | ||
2825 | /* | |
2826 | * Bit mask controlling how hard we retry per-node allocations. | |
2827 | * If we can not allocate the bit mask, do not attempt to allocate | |
2828 | * the requested huge pages. | |
2829 | */ | |
2830 | if (node_alloc_noretry) | |
2831 | nodes_clear(*node_alloc_noretry); | |
2832 | else | |
2833 | return -ENOMEM; | |
1da177e4 | 2834 | |
29383967 MK |
2835 | /* |
2836 | * resize_lock mutex prevents concurrent adjustments to number of | |
2837 | * pages in hstate via the proc/sysfs interfaces. | |
2838 | */ | |
2839 | mutex_lock(&h->resize_lock); | |
db71ef79 | 2840 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 2841 | |
fd875dca MK |
2842 | /* |
2843 | * Check for a node specific request. | |
2844 | * Changing node specific huge page count may require a corresponding | |
2845 | * change to the global count. In any case, the passed node mask | |
2846 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2847 | */ | |
2848 | if (nid != NUMA_NO_NODE) { | |
2849 | unsigned long old_count = count; | |
2850 | ||
2851 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2852 | /* | |
2853 | * User may have specified a large count value which caused the | |
2854 | * above calculation to overflow. In this case, they wanted | |
2855 | * to allocate as many huge pages as possible. Set count to | |
2856 | * largest possible value to align with their intention. | |
2857 | */ | |
2858 | if (count < old_count) | |
2859 | count = ULONG_MAX; | |
2860 | } | |
2861 | ||
4eb0716e AG |
2862 | /* |
2863 | * Gigantic pages runtime allocation depend on the capability for large | |
2864 | * page range allocation. | |
2865 | * If the system does not provide this feature, return an error when | |
2866 | * the user tries to allocate gigantic pages but let the user free the | |
2867 | * boottime allocated gigantic pages. | |
2868 | */ | |
2869 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2870 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 2871 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 2872 | mutex_unlock(&h->resize_lock); |
f60858f9 | 2873 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2874 | return -EINVAL; |
2875 | } | |
2876 | /* Fall through to decrease pool */ | |
2877 | } | |
aa888a74 | 2878 | |
7893d1d5 AL |
2879 | /* |
2880 | * Increase the pool size | |
2881 | * First take pages out of surplus state. Then make up the | |
2882 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2883 | * |
0c397dae | 2884 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2885 | * to convert a surplus huge page to a normal huge page. That is |
2886 | * not critical, though, it just means the overall size of the | |
2887 | * pool might be one hugepage larger than it needs to be, but | |
2888 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2889 | */ |
a5516438 | 2890 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2891 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2892 | break; |
2893 | } | |
2894 | ||
a5516438 | 2895 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2896 | /* |
2897 | * If this allocation races such that we no longer need the | |
2898 | * page, free_huge_page will handle it by freeing the page | |
2899 | * and reducing the surplus. | |
2900 | */ | |
db71ef79 | 2901 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
2902 | |
2903 | /* yield cpu to avoid soft lockup */ | |
2904 | cond_resched(); | |
2905 | ||
f60858f9 MK |
2906 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2907 | node_alloc_noretry); | |
db71ef79 | 2908 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
2909 | if (!ret) |
2910 | goto out; | |
2911 | ||
536240f2 MG |
2912 | /* Bail for signals. Probably ctrl-c from user */ |
2913 | if (signal_pending(current)) | |
2914 | goto out; | |
7893d1d5 | 2915 | } |
7893d1d5 AL |
2916 | |
2917 | /* | |
2918 | * Decrease the pool size | |
2919 | * First return free pages to the buddy allocator (being careful | |
2920 | * to keep enough around to satisfy reservations). Then place | |
2921 | * pages into surplus state as needed so the pool will shrink | |
2922 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2923 | * |
2924 | * By placing pages into the surplus state independent of the | |
2925 | * overcommit value, we are allowing the surplus pool size to | |
2926 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2927 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2928 | * though, we'll note that we're not allowed to exceed surplus |
2929 | * and won't grow the pool anywhere else. Not until one of the | |
2930 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2931 | */ |
a5516438 | 2932 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2933 | min_count = max(count, min_count); |
6ae11b27 | 2934 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
2935 | |
2936 | /* | |
2937 | * Collect pages to be removed on list without dropping lock | |
2938 | */ | |
a5516438 | 2939 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
2940 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
2941 | if (!page) | |
1da177e4 | 2942 | break; |
10c6ec49 MK |
2943 | |
2944 | list_add(&page->lru, &page_list); | |
1da177e4 | 2945 | } |
10c6ec49 | 2946 | /* free the pages after dropping lock */ |
db71ef79 | 2947 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2948 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2949 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 2950 | |
a5516438 | 2951 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2952 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2953 | break; |
2954 | } | |
2955 | out: | |
4eb0716e | 2956 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 2957 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 2958 | mutex_unlock(&h->resize_lock); |
4eb0716e | 2959 | |
f60858f9 MK |
2960 | NODEMASK_FREE(node_alloc_noretry); |
2961 | ||
4eb0716e | 2962 | return 0; |
1da177e4 LT |
2963 | } |
2964 | ||
a3437870 NA |
2965 | #define HSTATE_ATTR_RO(_name) \ |
2966 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2967 | ||
2968 | #define HSTATE_ATTR(_name) \ | |
2969 | static struct kobj_attribute _name##_attr = \ | |
2970 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2971 | ||
2972 | static struct kobject *hugepages_kobj; | |
2973 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2974 | ||
9a305230 LS |
2975 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2976 | ||
2977 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2978 | { |
2979 | int i; | |
9a305230 | 2980 | |
a3437870 | 2981 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2982 | if (hstate_kobjs[i] == kobj) { |
2983 | if (nidp) | |
2984 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2985 | return &hstates[i]; |
9a305230 LS |
2986 | } |
2987 | ||
2988 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2989 | } |
2990 | ||
06808b08 | 2991 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2992 | struct kobj_attribute *attr, char *buf) |
2993 | { | |
9a305230 LS |
2994 | struct hstate *h; |
2995 | unsigned long nr_huge_pages; | |
2996 | int nid; | |
2997 | ||
2998 | h = kobj_to_hstate(kobj, &nid); | |
2999 | if (nid == NUMA_NO_NODE) | |
3000 | nr_huge_pages = h->nr_huge_pages; | |
3001 | else | |
3002 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
3003 | ||
ae7a927d | 3004 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 3005 | } |
adbe8726 | 3006 | |
238d3c13 DR |
3007 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
3008 | struct hstate *h, int nid, | |
3009 | unsigned long count, size_t len) | |
a3437870 NA |
3010 | { |
3011 | int err; | |
2d0adf7e | 3012 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 3013 | |
2d0adf7e OS |
3014 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
3015 | return -EINVAL; | |
adbe8726 | 3016 | |
9a305230 LS |
3017 | if (nid == NUMA_NO_NODE) { |
3018 | /* | |
3019 | * global hstate attribute | |
3020 | */ | |
3021 | if (!(obey_mempolicy && | |
2d0adf7e OS |
3022 | init_nodemask_of_mempolicy(&nodes_allowed))) |
3023 | n_mask = &node_states[N_MEMORY]; | |
3024 | else | |
3025 | n_mask = &nodes_allowed; | |
3026 | } else { | |
9a305230 | 3027 | /* |
fd875dca MK |
3028 | * Node specific request. count adjustment happens in |
3029 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 3030 | */ |
2d0adf7e OS |
3031 | init_nodemask_of_node(&nodes_allowed, nid); |
3032 | n_mask = &nodes_allowed; | |
fd875dca | 3033 | } |
9a305230 | 3034 | |
2d0adf7e | 3035 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 3036 | |
4eb0716e | 3037 | return err ? err : len; |
06808b08 LS |
3038 | } |
3039 | ||
238d3c13 DR |
3040 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
3041 | struct kobject *kobj, const char *buf, | |
3042 | size_t len) | |
3043 | { | |
3044 | struct hstate *h; | |
3045 | unsigned long count; | |
3046 | int nid; | |
3047 | int err; | |
3048 | ||
3049 | err = kstrtoul(buf, 10, &count); | |
3050 | if (err) | |
3051 | return err; | |
3052 | ||
3053 | h = kobj_to_hstate(kobj, &nid); | |
3054 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
3055 | } | |
3056 | ||
06808b08 LS |
3057 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
3058 | struct kobj_attribute *attr, char *buf) | |
3059 | { | |
3060 | return nr_hugepages_show_common(kobj, attr, buf); | |
3061 | } | |
3062 | ||
3063 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
3064 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3065 | { | |
238d3c13 | 3066 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
3067 | } |
3068 | HSTATE_ATTR(nr_hugepages); | |
3069 | ||
06808b08 LS |
3070 | #ifdef CONFIG_NUMA |
3071 | ||
3072 | /* | |
3073 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
3074 | * huge page alloc/free. | |
3075 | */ | |
3076 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
3077 | struct kobj_attribute *attr, |
3078 | char *buf) | |
06808b08 LS |
3079 | { |
3080 | return nr_hugepages_show_common(kobj, attr, buf); | |
3081 | } | |
3082 | ||
3083 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
3084 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3085 | { | |
238d3c13 | 3086 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
3087 | } |
3088 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
3089 | #endif | |
3090 | ||
3091 | ||
a3437870 NA |
3092 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
3093 | struct kobj_attribute *attr, char *buf) | |
3094 | { | |
9a305230 | 3095 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3096 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 3097 | } |
adbe8726 | 3098 | |
a3437870 NA |
3099 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
3100 | struct kobj_attribute *attr, const char *buf, size_t count) | |
3101 | { | |
3102 | int err; | |
3103 | unsigned long input; | |
9a305230 | 3104 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 3105 | |
bae7f4ae | 3106 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
3107 | return -EINVAL; |
3108 | ||
3dbb95f7 | 3109 | err = kstrtoul(buf, 10, &input); |
a3437870 | 3110 | if (err) |
73ae31e5 | 3111 | return err; |
a3437870 | 3112 | |
db71ef79 | 3113 | spin_lock_irq(&hugetlb_lock); |
a3437870 | 3114 | h->nr_overcommit_huge_pages = input; |
db71ef79 | 3115 | spin_unlock_irq(&hugetlb_lock); |
a3437870 NA |
3116 | |
3117 | return count; | |
3118 | } | |
3119 | HSTATE_ATTR(nr_overcommit_hugepages); | |
3120 | ||
3121 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
3122 | struct kobj_attribute *attr, char *buf) | |
3123 | { | |
9a305230 LS |
3124 | struct hstate *h; |
3125 | unsigned long free_huge_pages; | |
3126 | int nid; | |
3127 | ||
3128 | h = kobj_to_hstate(kobj, &nid); | |
3129 | if (nid == NUMA_NO_NODE) | |
3130 | free_huge_pages = h->free_huge_pages; | |
3131 | else | |
3132 | free_huge_pages = h->free_huge_pages_node[nid]; | |
3133 | ||
ae7a927d | 3134 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
3135 | } |
3136 | HSTATE_ATTR_RO(free_hugepages); | |
3137 | ||
3138 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
3139 | struct kobj_attribute *attr, char *buf) | |
3140 | { | |
9a305230 | 3141 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3142 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
3143 | } |
3144 | HSTATE_ATTR_RO(resv_hugepages); | |
3145 | ||
3146 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
3147 | struct kobj_attribute *attr, char *buf) | |
3148 | { | |
9a305230 LS |
3149 | struct hstate *h; |
3150 | unsigned long surplus_huge_pages; | |
3151 | int nid; | |
3152 | ||
3153 | h = kobj_to_hstate(kobj, &nid); | |
3154 | if (nid == NUMA_NO_NODE) | |
3155 | surplus_huge_pages = h->surplus_huge_pages; | |
3156 | else | |
3157 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
3158 | ||
ae7a927d | 3159 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
3160 | } |
3161 | HSTATE_ATTR_RO(surplus_hugepages); | |
3162 | ||
3163 | static struct attribute *hstate_attrs[] = { | |
3164 | &nr_hugepages_attr.attr, | |
3165 | &nr_overcommit_hugepages_attr.attr, | |
3166 | &free_hugepages_attr.attr, | |
3167 | &resv_hugepages_attr.attr, | |
3168 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3169 | #ifdef CONFIG_NUMA |
3170 | &nr_hugepages_mempolicy_attr.attr, | |
3171 | #endif | |
a3437870 NA |
3172 | NULL, |
3173 | }; | |
3174 | ||
67e5ed96 | 3175 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3176 | .attrs = hstate_attrs, |
3177 | }; | |
3178 | ||
094e9539 JM |
3179 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3180 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3181 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3182 | { |
3183 | int retval; | |
972dc4de | 3184 | int hi = hstate_index(h); |
a3437870 | 3185 | |
9a305230 LS |
3186 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3187 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3188 | return -ENOMEM; |
3189 | ||
9a305230 | 3190 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 3191 | if (retval) { |
9a305230 | 3192 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 ML |
3193 | hstate_kobjs[hi] = NULL; |
3194 | } | |
a3437870 NA |
3195 | |
3196 | return retval; | |
3197 | } | |
3198 | ||
3199 | static void __init hugetlb_sysfs_init(void) | |
3200 | { | |
3201 | struct hstate *h; | |
3202 | int err; | |
3203 | ||
3204 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3205 | if (!hugepages_kobj) | |
3206 | return; | |
3207 | ||
3208 | for_each_hstate(h) { | |
9a305230 LS |
3209 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3210 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3211 | if (err) |
282f4214 | 3212 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
3213 | } |
3214 | } | |
3215 | ||
9a305230 LS |
3216 | #ifdef CONFIG_NUMA |
3217 | ||
3218 | /* | |
3219 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3220 | * with node devices in node_devices[] using a parallel array. The array |
3221 | * index of a node device or _hstate == node id. | |
3222 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3223 | * the base kernel, on the hugetlb module. |
3224 | */ | |
3225 | struct node_hstate { | |
3226 | struct kobject *hugepages_kobj; | |
3227 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3228 | }; | |
b4e289a6 | 3229 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3230 | |
3231 | /* | |
10fbcf4c | 3232 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3233 | */ |
3234 | static struct attribute *per_node_hstate_attrs[] = { | |
3235 | &nr_hugepages_attr.attr, | |
3236 | &free_hugepages_attr.attr, | |
3237 | &surplus_hugepages_attr.attr, | |
3238 | NULL, | |
3239 | }; | |
3240 | ||
67e5ed96 | 3241 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3242 | .attrs = per_node_hstate_attrs, |
3243 | }; | |
3244 | ||
3245 | /* | |
10fbcf4c | 3246 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3247 | * Returns node id via non-NULL nidp. |
3248 | */ | |
3249 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3250 | { | |
3251 | int nid; | |
3252 | ||
3253 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3254 | struct node_hstate *nhs = &node_hstates[nid]; | |
3255 | int i; | |
3256 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3257 | if (nhs->hstate_kobjs[i] == kobj) { | |
3258 | if (nidp) | |
3259 | *nidp = nid; | |
3260 | return &hstates[i]; | |
3261 | } | |
3262 | } | |
3263 | ||
3264 | BUG(); | |
3265 | return NULL; | |
3266 | } | |
3267 | ||
3268 | /* | |
10fbcf4c | 3269 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3270 | * No-op if no hstate attributes attached. |
3271 | */ | |
3cd8b44f | 3272 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3273 | { |
3274 | struct hstate *h; | |
10fbcf4c | 3275 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3276 | |
3277 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3278 | return; /* no hstate attributes */ |
9a305230 | 3279 | |
972dc4de AK |
3280 | for_each_hstate(h) { |
3281 | int idx = hstate_index(h); | |
3282 | if (nhs->hstate_kobjs[idx]) { | |
3283 | kobject_put(nhs->hstate_kobjs[idx]); | |
3284 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3285 | } |
972dc4de | 3286 | } |
9a305230 LS |
3287 | |
3288 | kobject_put(nhs->hugepages_kobj); | |
3289 | nhs->hugepages_kobj = NULL; | |
3290 | } | |
3291 | ||
9a305230 LS |
3292 | |
3293 | /* | |
10fbcf4c | 3294 | * Register hstate attributes for a single node device. |
9a305230 LS |
3295 | * No-op if attributes already registered. |
3296 | */ | |
3cd8b44f | 3297 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3298 | { |
3299 | struct hstate *h; | |
10fbcf4c | 3300 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3301 | int err; |
3302 | ||
3303 | if (nhs->hugepages_kobj) | |
3304 | return; /* already allocated */ | |
3305 | ||
3306 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3307 | &node->dev.kobj); |
9a305230 LS |
3308 | if (!nhs->hugepages_kobj) |
3309 | return; | |
3310 | ||
3311 | for_each_hstate(h) { | |
3312 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3313 | nhs->hstate_kobjs, | |
3314 | &per_node_hstate_attr_group); | |
3315 | if (err) { | |
282f4214 | 3316 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3317 | h->name, node->dev.id); |
9a305230 LS |
3318 | hugetlb_unregister_node(node); |
3319 | break; | |
3320 | } | |
3321 | } | |
3322 | } | |
3323 | ||
3324 | /* | |
9b5e5d0f | 3325 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3326 | * devices of nodes that have memory. All on-line nodes should have |
3327 | * registered their associated device by this time. | |
9a305230 | 3328 | */ |
7d9ca000 | 3329 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3330 | { |
3331 | int nid; | |
3332 | ||
8cebfcd0 | 3333 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3334 | struct node *node = node_devices[nid]; |
10fbcf4c | 3335 | if (node->dev.id == nid) |
9a305230 LS |
3336 | hugetlb_register_node(node); |
3337 | } | |
3338 | ||
3339 | /* | |
10fbcf4c | 3340 | * Let the node device driver know we're here so it can |
9a305230 LS |
3341 | * [un]register hstate attributes on node hotplug. |
3342 | */ | |
3343 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3344 | hugetlb_unregister_node); | |
3345 | } | |
3346 | #else /* !CONFIG_NUMA */ | |
3347 | ||
3348 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3349 | { | |
3350 | BUG(); | |
3351 | if (nidp) | |
3352 | *nidp = -1; | |
3353 | return NULL; | |
3354 | } | |
3355 | ||
9a305230 LS |
3356 | static void hugetlb_register_all_nodes(void) { } |
3357 | ||
3358 | #endif | |
3359 | ||
a3437870 NA |
3360 | static int __init hugetlb_init(void) |
3361 | { | |
8382d914 DB |
3362 | int i; |
3363 | ||
d6995da3 MK |
3364 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
3365 | __NR_HPAGEFLAGS); | |
3366 | ||
c2833a5b MK |
3367 | if (!hugepages_supported()) { |
3368 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
3369 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 3370 | return 0; |
c2833a5b | 3371 | } |
a3437870 | 3372 | |
282f4214 MK |
3373 | /* |
3374 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
3375 | * architectures depend on setup being done here. | |
3376 | */ | |
3377 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
3378 | if (!parsed_default_hugepagesz) { | |
3379 | /* | |
3380 | * If we did not parse a default huge page size, set | |
3381 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
3382 | * number of huge pages for this default size was implicitly | |
3383 | * specified, set that here as well. | |
3384 | * Note that the implicit setting will overwrite an explicit | |
3385 | * setting. A warning will be printed in this case. | |
3386 | */ | |
3387 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
3388 | if (default_hstate_max_huge_pages) { | |
3389 | if (default_hstate.max_huge_pages) { | |
3390 | char buf[32]; | |
3391 | ||
3392 | string_get_size(huge_page_size(&default_hstate), | |
3393 | 1, STRING_UNITS_2, buf, 32); | |
3394 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
3395 | default_hstate.max_huge_pages, buf); | |
3396 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
3397 | default_hstate_max_huge_pages); | |
3398 | } | |
3399 | default_hstate.max_huge_pages = | |
3400 | default_hstate_max_huge_pages; | |
d715cf80 | 3401 | } |
f8b74815 | 3402 | } |
a3437870 | 3403 | |
cf11e85f | 3404 | hugetlb_cma_check(); |
a3437870 | 3405 | hugetlb_init_hstates(); |
aa888a74 | 3406 | gather_bootmem_prealloc(); |
a3437870 NA |
3407 | report_hugepages(); |
3408 | ||
3409 | hugetlb_sysfs_init(); | |
9a305230 | 3410 | hugetlb_register_all_nodes(); |
7179e7bf | 3411 | hugetlb_cgroup_file_init(); |
9a305230 | 3412 | |
8382d914 DB |
3413 | #ifdef CONFIG_SMP |
3414 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3415 | #else | |
3416 | num_fault_mutexes = 1; | |
3417 | #endif | |
c672c7f2 | 3418 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3419 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3420 | GFP_KERNEL); | |
c672c7f2 | 3421 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3422 | |
3423 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3424 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3425 | return 0; |
3426 | } | |
3e89e1c5 | 3427 | subsys_initcall(hugetlb_init); |
a3437870 | 3428 | |
ae94da89 MK |
3429 | /* Overwritten by architectures with more huge page sizes */ |
3430 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 3431 | { |
ae94da89 | 3432 | return size == HPAGE_SIZE; |
9fee021d VT |
3433 | } |
3434 | ||
d00181b9 | 3435 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3436 | { |
3437 | struct hstate *h; | |
8faa8b07 AK |
3438 | unsigned long i; |
3439 | ||
a3437870 | 3440 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
3441 | return; |
3442 | } | |
47d38344 | 3443 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3444 | BUG_ON(order == 0); |
47d38344 | 3445 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 3446 | mutex_init(&h->resize_lock); |
a3437870 | 3447 | h->order = order; |
aca78307 | 3448 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
3449 | for (i = 0; i < MAX_NUMNODES; ++i) |
3450 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3451 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3452 | h->next_nid_to_alloc = first_memory_node; |
3453 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3454 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3455 | huge_page_size(h)/1024); | |
8faa8b07 | 3456 | |
a3437870 NA |
3457 | parsed_hstate = h; |
3458 | } | |
3459 | ||
282f4214 MK |
3460 | /* |
3461 | * hugepages command line processing | |
3462 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
3463 | * specification. If not, ignore the hugepages value. hugepages can also | |
3464 | * be the first huge page command line option in which case it implicitly | |
3465 | * specifies the number of huge pages for the default size. | |
3466 | */ | |
3467 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
3468 | { |
3469 | unsigned long *mhp; | |
8faa8b07 | 3470 | static unsigned long *last_mhp; |
a3437870 | 3471 | |
9fee021d | 3472 | if (!parsed_valid_hugepagesz) { |
282f4214 | 3473 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 3474 | parsed_valid_hugepagesz = true; |
282f4214 | 3475 | return 0; |
9fee021d | 3476 | } |
282f4214 | 3477 | |
a3437870 | 3478 | /* |
282f4214 MK |
3479 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
3480 | * yet, so this hugepages= parameter goes to the "default hstate". | |
3481 | * Otherwise, it goes with the previously parsed hugepagesz or | |
3482 | * default_hugepagesz. | |
a3437870 | 3483 | */ |
9fee021d | 3484 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3485 | mhp = &default_hstate_max_huge_pages; |
3486 | else | |
3487 | mhp = &parsed_hstate->max_huge_pages; | |
3488 | ||
8faa8b07 | 3489 | if (mhp == last_mhp) { |
282f4214 MK |
3490 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
3491 | return 0; | |
8faa8b07 AK |
3492 | } |
3493 | ||
a3437870 NA |
3494 | if (sscanf(s, "%lu", mhp) <= 0) |
3495 | *mhp = 0; | |
3496 | ||
8faa8b07 AK |
3497 | /* |
3498 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 3499 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
3500 | * use the bootmem allocator. |
3501 | */ | |
04adbc3f | 3502 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
3503 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3504 | ||
3505 | last_mhp = mhp; | |
3506 | ||
a3437870 NA |
3507 | return 1; |
3508 | } | |
282f4214 | 3509 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 3510 | |
282f4214 MK |
3511 | /* |
3512 | * hugepagesz command line processing | |
3513 | * A specific huge page size can only be specified once with hugepagesz. | |
3514 | * hugepagesz is followed by hugepages on the command line. The global | |
3515 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
3516 | * hugepagesz argument was valid. | |
3517 | */ | |
359f2544 | 3518 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 3519 | { |
359f2544 | 3520 | unsigned long size; |
282f4214 MK |
3521 | struct hstate *h; |
3522 | ||
3523 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
3524 | size = (unsigned long)memparse(s, NULL); |
3525 | ||
3526 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3527 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
359f2544 MK |
3528 | return 0; |
3529 | } | |
3530 | ||
282f4214 MK |
3531 | h = size_to_hstate(size); |
3532 | if (h) { | |
3533 | /* | |
3534 | * hstate for this size already exists. This is normally | |
3535 | * an error, but is allowed if the existing hstate is the | |
3536 | * default hstate. More specifically, it is only allowed if | |
3537 | * the number of huge pages for the default hstate was not | |
3538 | * previously specified. | |
3539 | */ | |
3540 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
3541 | default_hstate.max_huge_pages) { | |
3542 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
3543 | return 0; | |
3544 | } | |
3545 | ||
3546 | /* | |
3547 | * No need to call hugetlb_add_hstate() as hstate already | |
3548 | * exists. But, do set parsed_hstate so that a following | |
3549 | * hugepages= parameter will be applied to this hstate. | |
3550 | */ | |
3551 | parsed_hstate = h; | |
3552 | parsed_valid_hugepagesz = true; | |
3553 | return 1; | |
38237830 MK |
3554 | } |
3555 | ||
359f2544 | 3556 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 3557 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
3558 | return 1; |
3559 | } | |
359f2544 MK |
3560 | __setup("hugepagesz=", hugepagesz_setup); |
3561 | ||
282f4214 MK |
3562 | /* |
3563 | * default_hugepagesz command line input | |
3564 | * Only one instance of default_hugepagesz allowed on command line. | |
3565 | */ | |
ae94da89 | 3566 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 3567 | { |
ae94da89 MK |
3568 | unsigned long size; |
3569 | ||
282f4214 | 3570 | parsed_valid_hugepagesz = false; |
282f4214 MK |
3571 | if (parsed_default_hugepagesz) { |
3572 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
3573 | return 0; | |
3574 | } | |
3575 | ||
ae94da89 MK |
3576 | size = (unsigned long)memparse(s, NULL); |
3577 | ||
3578 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3579 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
ae94da89 MK |
3580 | return 0; |
3581 | } | |
3582 | ||
282f4214 MK |
3583 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
3584 | parsed_valid_hugepagesz = true; | |
3585 | parsed_default_hugepagesz = true; | |
3586 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
3587 | ||
3588 | /* | |
3589 | * The number of default huge pages (for this size) could have been | |
3590 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
3591 | * then default_hstate_max_huge_pages is set. If the default huge | |
3592 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
3593 | * allocated here from bootmem allocator. | |
3594 | */ | |
3595 | if (default_hstate_max_huge_pages) { | |
3596 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3597 | if (hstate_is_gigantic(&default_hstate)) | |
3598 | hugetlb_hstate_alloc_pages(&default_hstate); | |
3599 | default_hstate_max_huge_pages = 0; | |
3600 | } | |
3601 | ||
e11bfbfc NP |
3602 | return 1; |
3603 | } | |
ae94da89 | 3604 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 3605 | |
8ca39e68 | 3606 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
3607 | { |
3608 | int node; | |
3609 | unsigned int nr = 0; | |
8ca39e68 MS |
3610 | nodemask_t *mpol_allowed; |
3611 | unsigned int *array = h->free_huge_pages_node; | |
3612 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
3613 | ||
3614 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 3615 | |
8ca39e68 | 3616 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
c93b0a99 | 3617 | if (!mpol_allowed || node_isset(node, *mpol_allowed)) |
8ca39e68 MS |
3618 | nr += array[node]; |
3619 | } | |
8a213460 NA |
3620 | |
3621 | return nr; | |
3622 | } | |
3623 | ||
3624 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
3625 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
3626 | void *buffer, size_t *length, | |
3627 | loff_t *ppos, unsigned long *out) | |
3628 | { | |
3629 | struct ctl_table dup_table; | |
3630 | ||
3631 | /* | |
3632 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
3633 | * can duplicate the @table and alter the duplicate of it. | |
3634 | */ | |
3635 | dup_table = *table; | |
3636 | dup_table.data = out; | |
3637 | ||
3638 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
3639 | } | |
3640 | ||
06808b08 LS |
3641 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3642 | struct ctl_table *table, int write, | |
32927393 | 3643 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 3644 | { |
e5ff2159 | 3645 | struct hstate *h = &default_hstate; |
238d3c13 | 3646 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3647 | int ret; |
e5ff2159 | 3648 | |
457c1b27 | 3649 | if (!hugepages_supported()) |
86613628 | 3650 | return -EOPNOTSUPP; |
457c1b27 | 3651 | |
17743798 MS |
3652 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3653 | &tmp); | |
08d4a246 MH |
3654 | if (ret) |
3655 | goto out; | |
e5ff2159 | 3656 | |
238d3c13 DR |
3657 | if (write) |
3658 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3659 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3660 | out: |
3661 | return ret; | |
1da177e4 | 3662 | } |
396faf03 | 3663 | |
06808b08 | 3664 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 3665 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3666 | { |
3667 | ||
3668 | return hugetlb_sysctl_handler_common(false, table, write, | |
3669 | buffer, length, ppos); | |
3670 | } | |
3671 | ||
3672 | #ifdef CONFIG_NUMA | |
3673 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 3674 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3675 | { |
3676 | return hugetlb_sysctl_handler_common(true, table, write, | |
3677 | buffer, length, ppos); | |
3678 | } | |
3679 | #endif /* CONFIG_NUMA */ | |
3680 | ||
a3d0c6aa | 3681 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 3682 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 3683 | { |
a5516438 | 3684 | struct hstate *h = &default_hstate; |
e5ff2159 | 3685 | unsigned long tmp; |
08d4a246 | 3686 | int ret; |
e5ff2159 | 3687 | |
457c1b27 | 3688 | if (!hugepages_supported()) |
86613628 | 3689 | return -EOPNOTSUPP; |
457c1b27 | 3690 | |
c033a93c | 3691 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3692 | |
bae7f4ae | 3693 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3694 | return -EINVAL; |
3695 | ||
17743798 MS |
3696 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3697 | &tmp); | |
08d4a246 MH |
3698 | if (ret) |
3699 | goto out; | |
e5ff2159 AK |
3700 | |
3701 | if (write) { | |
db71ef79 | 3702 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 3703 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 3704 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 3705 | } |
08d4a246 MH |
3706 | out: |
3707 | return ret; | |
a3d0c6aa NA |
3708 | } |
3709 | ||
1da177e4 LT |
3710 | #endif /* CONFIG_SYSCTL */ |
3711 | ||
e1759c21 | 3712 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3713 | { |
fcb2b0c5 RG |
3714 | struct hstate *h; |
3715 | unsigned long total = 0; | |
3716 | ||
457c1b27 NA |
3717 | if (!hugepages_supported()) |
3718 | return; | |
fcb2b0c5 RG |
3719 | |
3720 | for_each_hstate(h) { | |
3721 | unsigned long count = h->nr_huge_pages; | |
3722 | ||
aca78307 | 3723 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
3724 | |
3725 | if (h == &default_hstate) | |
3726 | seq_printf(m, | |
3727 | "HugePages_Total: %5lu\n" | |
3728 | "HugePages_Free: %5lu\n" | |
3729 | "HugePages_Rsvd: %5lu\n" | |
3730 | "HugePages_Surp: %5lu\n" | |
3731 | "Hugepagesize: %8lu kB\n", | |
3732 | count, | |
3733 | h->free_huge_pages, | |
3734 | h->resv_huge_pages, | |
3735 | h->surplus_huge_pages, | |
aca78307 | 3736 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
3737 | } |
3738 | ||
aca78307 | 3739 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
3740 | } |
3741 | ||
7981593b | 3742 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 3743 | { |
a5516438 | 3744 | struct hstate *h = &default_hstate; |
7981593b | 3745 | |
457c1b27 NA |
3746 | if (!hugepages_supported()) |
3747 | return 0; | |
7981593b JP |
3748 | |
3749 | return sysfs_emit_at(buf, len, | |
3750 | "Node %d HugePages_Total: %5u\n" | |
3751 | "Node %d HugePages_Free: %5u\n" | |
3752 | "Node %d HugePages_Surp: %5u\n", | |
3753 | nid, h->nr_huge_pages_node[nid], | |
3754 | nid, h->free_huge_pages_node[nid], | |
3755 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3756 | } |
3757 | ||
949f7ec5 DR |
3758 | void hugetlb_show_meminfo(void) |
3759 | { | |
3760 | struct hstate *h; | |
3761 | int nid; | |
3762 | ||
457c1b27 NA |
3763 | if (!hugepages_supported()) |
3764 | return; | |
3765 | ||
949f7ec5 DR |
3766 | for_each_node_state(nid, N_MEMORY) |
3767 | for_each_hstate(h) | |
3768 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3769 | nid, | |
3770 | h->nr_huge_pages_node[nid], | |
3771 | h->free_huge_pages_node[nid], | |
3772 | h->surplus_huge_pages_node[nid], | |
aca78307 | 3773 | huge_page_size(h) / SZ_1K); |
949f7ec5 DR |
3774 | } |
3775 | ||
5d317b2b NH |
3776 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3777 | { | |
3778 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3779 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3780 | } | |
3781 | ||
1da177e4 LT |
3782 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3783 | unsigned long hugetlb_total_pages(void) | |
3784 | { | |
d0028588 WL |
3785 | struct hstate *h; |
3786 | unsigned long nr_total_pages = 0; | |
3787 | ||
3788 | for_each_hstate(h) | |
3789 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3790 | return nr_total_pages; | |
1da177e4 | 3791 | } |
1da177e4 | 3792 | |
a5516438 | 3793 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3794 | { |
3795 | int ret = -ENOMEM; | |
3796 | ||
0aa7f354 ML |
3797 | if (!delta) |
3798 | return 0; | |
3799 | ||
db71ef79 | 3800 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
3801 | /* |
3802 | * When cpuset is configured, it breaks the strict hugetlb page | |
3803 | * reservation as the accounting is done on a global variable. Such | |
3804 | * reservation is completely rubbish in the presence of cpuset because | |
3805 | * the reservation is not checked against page availability for the | |
3806 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3807 | * with lack of free htlb page in cpuset that the task is in. | |
3808 | * Attempt to enforce strict accounting with cpuset is almost | |
3809 | * impossible (or too ugly) because cpuset is too fluid that | |
3810 | * task or memory node can be dynamically moved between cpusets. | |
3811 | * | |
3812 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3813 | * undesirable. However, in order to preserve some of the semantics, | |
3814 | * we fall back to check against current free page availability as | |
3815 | * a best attempt and hopefully to minimize the impact of changing | |
3816 | * semantics that cpuset has. | |
8ca39e68 MS |
3817 | * |
3818 | * Apart from cpuset, we also have memory policy mechanism that | |
3819 | * also determines from which node the kernel will allocate memory | |
3820 | * in a NUMA system. So similar to cpuset, we also should consider | |
3821 | * the memory policy of the current task. Similar to the description | |
3822 | * above. | |
fc1b8a73 MG |
3823 | */ |
3824 | if (delta > 0) { | |
a5516438 | 3825 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3826 | goto out; |
3827 | ||
8ca39e68 | 3828 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 3829 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
3830 | goto out; |
3831 | } | |
3832 | } | |
3833 | ||
3834 | ret = 0; | |
3835 | if (delta < 0) | |
a5516438 | 3836 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3837 | |
3838 | out: | |
db71ef79 | 3839 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
3840 | return ret; |
3841 | } | |
3842 | ||
84afd99b AW |
3843 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3844 | { | |
f522c3ac | 3845 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3846 | |
3847 | /* | |
3848 | * This new VMA should share its siblings reservation map if present. | |
3849 | * The VMA will only ever have a valid reservation map pointer where | |
3850 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3851 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3852 | * after this open call completes. It is therefore safe to take a |
3853 | * new reference here without additional locking. | |
3854 | */ | |
4e35f483 | 3855 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3856 | kref_get(&resv->refs); |
84afd99b AW |
3857 | } |
3858 | ||
a1e78772 MG |
3859 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3860 | { | |
a5516438 | 3861 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3862 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3863 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3864 | unsigned long reserve, start, end; |
1c5ecae3 | 3865 | long gbl_reserve; |
84afd99b | 3866 | |
4e35f483 JK |
3867 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3868 | return; | |
84afd99b | 3869 | |
4e35f483 JK |
3870 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3871 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3872 | |
4e35f483 | 3873 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 3874 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 3875 | if (reserve) { |
1c5ecae3 MK |
3876 | /* |
3877 | * Decrement reserve counts. The global reserve count may be | |
3878 | * adjusted if the subpool has a minimum size. | |
3879 | */ | |
3880 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3881 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3882 | } |
e9fe92ae MA |
3883 | |
3884 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
3885 | } |
3886 | ||
31383c68 DW |
3887 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3888 | { | |
3889 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3890 | return -EINVAL; | |
3891 | return 0; | |
3892 | } | |
3893 | ||
05ea8860 DW |
3894 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3895 | { | |
aca78307 | 3896 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
3897 | } |
3898 | ||
1da177e4 LT |
3899 | /* |
3900 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3901 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 3902 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
3903 | * this far. |
3904 | */ | |
b3ec9f33 | 3905 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3906 | { |
3907 | BUG(); | |
d0217ac0 | 3908 | return 0; |
1da177e4 LT |
3909 | } |
3910 | ||
eec3636a JC |
3911 | /* |
3912 | * When a new function is introduced to vm_operations_struct and added | |
3913 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3914 | * This is because under System V memory model, mappings created via | |
3915 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3916 | * their original vm_ops are overwritten with shm_vm_ops. | |
3917 | */ | |
f0f37e2f | 3918 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3919 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3920 | .open = hugetlb_vm_op_open, |
a1e78772 | 3921 | .close = hugetlb_vm_op_close, |
dd3b614f | 3922 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 3923 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3924 | }; |
3925 | ||
1e8f889b DG |
3926 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3927 | int writable) | |
63551ae0 DG |
3928 | { |
3929 | pte_t entry; | |
3930 | ||
1e8f889b | 3931 | if (writable) { |
106c992a GS |
3932 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3933 | vma->vm_page_prot))); | |
63551ae0 | 3934 | } else { |
106c992a GS |
3935 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3936 | vma->vm_page_prot)); | |
63551ae0 DG |
3937 | } |
3938 | entry = pte_mkyoung(entry); | |
3939 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3940 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3941 | |
3942 | return entry; | |
3943 | } | |
3944 | ||
1e8f889b DG |
3945 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3946 | unsigned long address, pte_t *ptep) | |
3947 | { | |
3948 | pte_t entry; | |
3949 | ||
106c992a | 3950 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3951 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3952 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3953 | } |
3954 | ||
d5ed7444 | 3955 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3956 | { |
3957 | swp_entry_t swp; | |
3958 | ||
3959 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3960 | return false; |
4a705fef | 3961 | swp = pte_to_swp_entry(pte); |
d79d176a | 3962 | if (is_migration_entry(swp)) |
d5ed7444 | 3963 | return true; |
4a705fef | 3964 | else |
d5ed7444 | 3965 | return false; |
4a705fef NH |
3966 | } |
3967 | ||
3e5c3600 | 3968 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
3969 | { |
3970 | swp_entry_t swp; | |
3971 | ||
3972 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 3973 | return false; |
4a705fef | 3974 | swp = pte_to_swp_entry(pte); |
d79d176a | 3975 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 3976 | return true; |
4a705fef | 3977 | else |
3e5c3600 | 3978 | return false; |
4a705fef | 3979 | } |
1e8f889b | 3980 | |
4eae4efa PX |
3981 | static void |
3982 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
3983 | struct page *new_page) | |
3984 | { | |
3985 | __SetPageUptodate(new_page); | |
3986 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); | |
3987 | hugepage_add_new_anon_rmap(new_page, vma, addr); | |
3988 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); | |
3989 | ClearHPageRestoreReserve(new_page); | |
3990 | SetHPageMigratable(new_page); | |
3991 | } | |
3992 | ||
63551ae0 DG |
3993 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3994 | struct vm_area_struct *vma) | |
3995 | { | |
5e41540c | 3996 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3997 | struct page *ptepage; |
1c59827d | 3998 | unsigned long addr; |
ca6eb14d | 3999 | bool cow = is_cow_mapping(vma->vm_flags); |
a5516438 AK |
4000 | struct hstate *h = hstate_vma(vma); |
4001 | unsigned long sz = huge_page_size(h); | |
4eae4efa | 4002 | unsigned long npages = pages_per_huge_page(h); |
c0d0381a | 4003 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 4004 | struct mmu_notifier_range range; |
e8569dd2 | 4005 | int ret = 0; |
1e8f889b | 4006 | |
ac46d4f3 | 4007 | if (cow) { |
7269f999 | 4008 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 4009 | vma->vm_start, |
ac46d4f3 JG |
4010 | vma->vm_end); |
4011 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
4012 | } else { |
4013 | /* | |
4014 | * For shared mappings i_mmap_rwsem must be held to call | |
4015 | * huge_pte_alloc, otherwise the returned ptep could go | |
4016 | * away if part of a shared pmd and another thread calls | |
4017 | * huge_pmd_unshare. | |
4018 | */ | |
4019 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 4020 | } |
e8569dd2 | 4021 | |
a5516438 | 4022 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 4023 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 4024 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
4025 | if (!src_pte) |
4026 | continue; | |
aec44e0f | 4027 | dst_pte = huge_pte_alloc(dst, vma, addr, sz); |
e8569dd2 AS |
4028 | if (!dst_pte) { |
4029 | ret = -ENOMEM; | |
4030 | break; | |
4031 | } | |
c5c99429 | 4032 | |
5e41540c MK |
4033 | /* |
4034 | * If the pagetables are shared don't copy or take references. | |
4035 | * dst_pte == src_pte is the common case of src/dest sharing. | |
4036 | * | |
4037 | * However, src could have 'unshared' and dst shares with | |
4038 | * another vma. If dst_pte !none, this implies sharing. | |
4039 | * Check here before taking page table lock, and once again | |
4040 | * after taking the lock below. | |
4041 | */ | |
4042 | dst_entry = huge_ptep_get(dst_pte); | |
4043 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
4044 | continue; |
4045 | ||
cb900f41 KS |
4046 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
4047 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4048 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 4049 | entry = huge_ptep_get(src_pte); |
5e41540c | 4050 | dst_entry = huge_ptep_get(dst_pte); |
4eae4efa | 4051 | again: |
5e41540c MK |
4052 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { |
4053 | /* | |
4054 | * Skip if src entry none. Also, skip in the | |
4055 | * unlikely case dst entry !none as this implies | |
4056 | * sharing with another vma. | |
4057 | */ | |
4a705fef NH |
4058 | ; |
4059 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
4060 | is_hugetlb_entry_hwpoisoned(entry))) { | |
4061 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
4062 | ||
4063 | if (is_write_migration_entry(swp_entry) && cow) { | |
4064 | /* | |
4065 | * COW mappings require pages in both | |
4066 | * parent and child to be set to read. | |
4067 | */ | |
4068 | make_migration_entry_read(&swp_entry); | |
4069 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
4070 | set_huge_swap_pte_at(src, addr, src_pte, |
4071 | entry, sz); | |
4a705fef | 4072 | } |
e5251fd4 | 4073 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 4074 | } else { |
4eae4efa PX |
4075 | entry = huge_ptep_get(src_pte); |
4076 | ptepage = pte_page(entry); | |
4077 | get_page(ptepage); | |
4078 | ||
4079 | /* | |
4080 | * This is a rare case where we see pinned hugetlb | |
4081 | * pages while they're prone to COW. We need to do the | |
4082 | * COW earlier during fork. | |
4083 | * | |
4084 | * When pre-allocating the page or copying data, we | |
4085 | * need to be without the pgtable locks since we could | |
4086 | * sleep during the process. | |
4087 | */ | |
4088 | if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { | |
4089 | pte_t src_pte_old = entry; | |
4090 | struct page *new; | |
4091 | ||
4092 | spin_unlock(src_ptl); | |
4093 | spin_unlock(dst_ptl); | |
4094 | /* Do not use reserve as it's private owned */ | |
4095 | new = alloc_huge_page(vma, addr, 1); | |
4096 | if (IS_ERR(new)) { | |
4097 | put_page(ptepage); | |
4098 | ret = PTR_ERR(new); | |
4099 | break; | |
4100 | } | |
4101 | copy_user_huge_page(new, ptepage, addr, vma, | |
4102 | npages); | |
4103 | put_page(ptepage); | |
4104 | ||
4105 | /* Install the new huge page if src pte stable */ | |
4106 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
4107 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4108 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4109 | entry = huge_ptep_get(src_pte); | |
4110 | if (!pte_same(src_pte_old, entry)) { | |
846be085 MK |
4111 | restore_reserve_on_error(h, vma, addr, |
4112 | new); | |
4eae4efa PX |
4113 | put_page(new); |
4114 | /* dst_entry won't change as in child */ | |
4115 | goto again; | |
4116 | } | |
4117 | hugetlb_install_page(vma, dst_pte, addr, new); | |
4118 | spin_unlock(src_ptl); | |
4119 | spin_unlock(dst_ptl); | |
4120 | continue; | |
4121 | } | |
4122 | ||
34ee645e | 4123 | if (cow) { |
0f10851e JG |
4124 | /* |
4125 | * No need to notify as we are downgrading page | |
4126 | * table protection not changing it to point | |
4127 | * to a new page. | |
4128 | * | |
ad56b738 | 4129 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4130 | */ |
7f2e9525 | 4131 | huge_ptep_set_wrprotect(src, addr, src_pte); |
84894e1c | 4132 | entry = huge_pte_wrprotect(entry); |
34ee645e | 4133 | } |
4eae4efa | 4134 | |
53f9263b | 4135 | page_dup_rmap(ptepage, true); |
1c59827d | 4136 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 4137 | hugetlb_count_add(npages, dst); |
1c59827d | 4138 | } |
cb900f41 KS |
4139 | spin_unlock(src_ptl); |
4140 | spin_unlock(dst_ptl); | |
63551ae0 | 4141 | } |
63551ae0 | 4142 | |
e8569dd2 | 4143 | if (cow) |
ac46d4f3 | 4144 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
4145 | else |
4146 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
4147 | |
4148 | return ret; | |
63551ae0 DG |
4149 | } |
4150 | ||
24669e58 AK |
4151 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
4152 | unsigned long start, unsigned long end, | |
4153 | struct page *ref_page) | |
63551ae0 DG |
4154 | { |
4155 | struct mm_struct *mm = vma->vm_mm; | |
4156 | unsigned long address; | |
c7546f8f | 4157 | pte_t *ptep; |
63551ae0 | 4158 | pte_t pte; |
cb900f41 | 4159 | spinlock_t *ptl; |
63551ae0 | 4160 | struct page *page; |
a5516438 AK |
4161 | struct hstate *h = hstate_vma(vma); |
4162 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 4163 | struct mmu_notifier_range range; |
a5516438 | 4164 | |
63551ae0 | 4165 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
4166 | BUG_ON(start & ~huge_page_mask(h)); |
4167 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 4168 | |
07e32661 AK |
4169 | /* |
4170 | * This is a hugetlb vma, all the pte entries should point | |
4171 | * to huge page. | |
4172 | */ | |
ed6a7935 | 4173 | tlb_change_page_size(tlb, sz); |
24669e58 | 4174 | tlb_start_vma(tlb, vma); |
dff11abe MK |
4175 | |
4176 | /* | |
4177 | * If sharing possible, alert mmu notifiers of worst case. | |
4178 | */ | |
6f4f13e8 JG |
4179 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
4180 | end); | |
ac46d4f3 JG |
4181 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
4182 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 4183 | address = start; |
569f48b8 | 4184 | for (; address < end; address += sz) { |
7868a208 | 4185 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 4186 | if (!ptep) |
c7546f8f DG |
4187 | continue; |
4188 | ||
cb900f41 | 4189 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 4190 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
31d49da5 | 4191 | spin_unlock(ptl); |
dff11abe MK |
4192 | /* |
4193 | * We just unmapped a page of PMDs by clearing a PUD. | |
4194 | * The caller's TLB flush range should cover this area. | |
4195 | */ | |
31d49da5 AK |
4196 | continue; |
4197 | } | |
39dde65c | 4198 | |
6629326b | 4199 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
4200 | if (huge_pte_none(pte)) { |
4201 | spin_unlock(ptl); | |
4202 | continue; | |
4203 | } | |
6629326b HD |
4204 | |
4205 | /* | |
9fbc1f63 NH |
4206 | * Migrating hugepage or HWPoisoned hugepage is already |
4207 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 4208 | */ |
9fbc1f63 | 4209 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 4210 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
4211 | spin_unlock(ptl); |
4212 | continue; | |
8c4894c6 | 4213 | } |
6629326b HD |
4214 | |
4215 | page = pte_page(pte); | |
04f2cbe3 MG |
4216 | /* |
4217 | * If a reference page is supplied, it is because a specific | |
4218 | * page is being unmapped, not a range. Ensure the page we | |
4219 | * are about to unmap is the actual page of interest. | |
4220 | */ | |
4221 | if (ref_page) { | |
31d49da5 AK |
4222 | if (page != ref_page) { |
4223 | spin_unlock(ptl); | |
4224 | continue; | |
4225 | } | |
04f2cbe3 MG |
4226 | /* |
4227 | * Mark the VMA as having unmapped its page so that | |
4228 | * future faults in this VMA will fail rather than | |
4229 | * looking like data was lost | |
4230 | */ | |
4231 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
4232 | } | |
4233 | ||
c7546f8f | 4234 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 4235 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 4236 | if (huge_pte_dirty(pte)) |
6649a386 | 4237 | set_page_dirty(page); |
9e81130b | 4238 | |
5d317b2b | 4239 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 4240 | page_remove_rmap(page, true); |
31d49da5 | 4241 | |
cb900f41 | 4242 | spin_unlock(ptl); |
e77b0852 | 4243 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
4244 | /* |
4245 | * Bail out after unmapping reference page if supplied | |
4246 | */ | |
4247 | if (ref_page) | |
4248 | break; | |
fe1668ae | 4249 | } |
ac46d4f3 | 4250 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 4251 | tlb_end_vma(tlb, vma); |
1da177e4 | 4252 | } |
63551ae0 | 4253 | |
d833352a MG |
4254 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
4255 | struct vm_area_struct *vma, unsigned long start, | |
4256 | unsigned long end, struct page *ref_page) | |
4257 | { | |
4258 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
4259 | ||
4260 | /* | |
4261 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
4262 | * test will fail on a vma being torn down, and not grab a page table | |
4263 | * on its way out. We're lucky that the flag has such an appropriate | |
4264 | * name, and can in fact be safely cleared here. We could clear it | |
4265 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 4266 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 4267 | * because in the context this is called, the VMA is about to be |
c8c06efa | 4268 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
4269 | */ |
4270 | vma->vm_flags &= ~VM_MAYSHARE; | |
4271 | } | |
4272 | ||
502717f4 | 4273 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 4274 | unsigned long end, struct page *ref_page) |
502717f4 | 4275 | { |
24669e58 | 4276 | struct mmu_gather tlb; |
dff11abe | 4277 | |
a72afd87 | 4278 | tlb_gather_mmu(&tlb, vma->vm_mm); |
24669e58 | 4279 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
ae8eba8b | 4280 | tlb_finish_mmu(&tlb); |
502717f4 KC |
4281 | } |
4282 | ||
04f2cbe3 MG |
4283 | /* |
4284 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 4285 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
4286 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
4287 | * same region. | |
4288 | */ | |
2f4612af DB |
4289 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
4290 | struct page *page, unsigned long address) | |
04f2cbe3 | 4291 | { |
7526674d | 4292 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
4293 | struct vm_area_struct *iter_vma; |
4294 | struct address_space *mapping; | |
04f2cbe3 MG |
4295 | pgoff_t pgoff; |
4296 | ||
4297 | /* | |
4298 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
4299 | * from page cache lookup which is in HPAGE_SIZE units. | |
4300 | */ | |
7526674d | 4301 | address = address & huge_page_mask(h); |
36e4f20a MH |
4302 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
4303 | vma->vm_pgoff; | |
93c76a3d | 4304 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 4305 | |
4eb2b1dc MG |
4306 | /* |
4307 | * Take the mapping lock for the duration of the table walk. As | |
4308 | * this mapping should be shared between all the VMAs, | |
4309 | * __unmap_hugepage_range() is called as the lock is already held | |
4310 | */ | |
83cde9e8 | 4311 | i_mmap_lock_write(mapping); |
6b2dbba8 | 4312 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
4313 | /* Do not unmap the current VMA */ |
4314 | if (iter_vma == vma) | |
4315 | continue; | |
4316 | ||
2f84a899 MG |
4317 | /* |
4318 | * Shared VMAs have their own reserves and do not affect | |
4319 | * MAP_PRIVATE accounting but it is possible that a shared | |
4320 | * VMA is using the same page so check and skip such VMAs. | |
4321 | */ | |
4322 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
4323 | continue; | |
4324 | ||
04f2cbe3 MG |
4325 | /* |
4326 | * Unmap the page from other VMAs without their own reserves. | |
4327 | * They get marked to be SIGKILLed if they fault in these | |
4328 | * areas. This is because a future no-page fault on this VMA | |
4329 | * could insert a zeroed page instead of the data existing | |
4330 | * from the time of fork. This would look like data corruption | |
4331 | */ | |
4332 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
4333 | unmap_hugepage_range(iter_vma, address, |
4334 | address + huge_page_size(h), page); | |
04f2cbe3 | 4335 | } |
83cde9e8 | 4336 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
4337 | } |
4338 | ||
0fe6e20b NH |
4339 | /* |
4340 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
4341 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
4342 | * cannot race with other handlers or page migration. | |
4343 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 4344 | */ |
2b740303 | 4345 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 4346 | unsigned long address, pte_t *ptep, |
3999f52e | 4347 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 4348 | { |
3999f52e | 4349 | pte_t pte; |
a5516438 | 4350 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 4351 | struct page *old_page, *new_page; |
2b740303 SJ |
4352 | int outside_reserve = 0; |
4353 | vm_fault_t ret = 0; | |
974e6d66 | 4354 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 4355 | struct mmu_notifier_range range; |
1e8f889b | 4356 | |
3999f52e | 4357 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
4358 | old_page = pte_page(pte); |
4359 | ||
04f2cbe3 | 4360 | retry_avoidcopy: |
1e8f889b DG |
4361 | /* If no-one else is actually using this page, avoid the copy |
4362 | * and just make the page writable */ | |
37a2140d | 4363 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 4364 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 4365 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 4366 | return 0; |
1e8f889b DG |
4367 | } |
4368 | ||
04f2cbe3 MG |
4369 | /* |
4370 | * If the process that created a MAP_PRIVATE mapping is about to | |
4371 | * perform a COW due to a shared page count, attempt to satisfy | |
4372 | * the allocation without using the existing reserves. The pagecache | |
4373 | * page is used to determine if the reserve at this address was | |
4374 | * consumed or not. If reserves were used, a partial faulted mapping | |
4375 | * at the time of fork() could consume its reserves on COW instead | |
4376 | * of the full address range. | |
4377 | */ | |
5944d011 | 4378 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
4379 | old_page != pagecache_page) |
4380 | outside_reserve = 1; | |
4381 | ||
09cbfeaf | 4382 | get_page(old_page); |
b76c8cfb | 4383 | |
ad4404a2 DB |
4384 | /* |
4385 | * Drop page table lock as buddy allocator may be called. It will | |
4386 | * be acquired again before returning to the caller, as expected. | |
4387 | */ | |
cb900f41 | 4388 | spin_unlock(ptl); |
5b7a1d40 | 4389 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 4390 | |
2fc39cec | 4391 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
4392 | /* |
4393 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4394 | * it is due to references held by a child and an insufficient | |
4395 | * huge page pool. To guarantee the original mappers | |
4396 | * reliability, unmap the page from child processes. The child | |
4397 | * may get SIGKILLed if it later faults. | |
4398 | */ | |
4399 | if (outside_reserve) { | |
e7dd91c4 MK |
4400 | struct address_space *mapping = vma->vm_file->f_mapping; |
4401 | pgoff_t idx; | |
4402 | u32 hash; | |
4403 | ||
09cbfeaf | 4404 | put_page(old_page); |
04f2cbe3 | 4405 | BUG_ON(huge_pte_none(pte)); |
e7dd91c4 MK |
4406 | /* |
4407 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
4408 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
4409 | * in write mode. Dropping i_mmap_rwsem in read mode | |
4410 | * here is OK as COW mappings do not interact with | |
4411 | * PMD sharing. | |
4412 | * | |
4413 | * Reacquire both after unmap operation. | |
4414 | */ | |
4415 | idx = vma_hugecache_offset(h, vma, haddr); | |
4416 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4417 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4418 | i_mmap_unlock_read(mapping); | |
4419 | ||
5b7a1d40 | 4420 | unmap_ref_private(mm, vma, old_page, haddr); |
e7dd91c4 MK |
4421 | |
4422 | i_mmap_lock_read(mapping); | |
4423 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
2f4612af | 4424 | spin_lock(ptl); |
5b7a1d40 | 4425 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
4426 | if (likely(ptep && |
4427 | pte_same(huge_ptep_get(ptep), pte))) | |
4428 | goto retry_avoidcopy; | |
4429 | /* | |
4430 | * race occurs while re-acquiring page table | |
4431 | * lock, and our job is done. | |
4432 | */ | |
4433 | return 0; | |
04f2cbe3 MG |
4434 | } |
4435 | ||
2b740303 | 4436 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 4437 | goto out_release_old; |
1e8f889b DG |
4438 | } |
4439 | ||
0fe6e20b NH |
4440 | /* |
4441 | * When the original hugepage is shared one, it does not have | |
4442 | * anon_vma prepared. | |
4443 | */ | |
44e2aa93 | 4444 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
4445 | ret = VM_FAULT_OOM; |
4446 | goto out_release_all; | |
44e2aa93 | 4447 | } |
0fe6e20b | 4448 | |
974e6d66 | 4449 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 4450 | pages_per_huge_page(h)); |
0ed361de | 4451 | __SetPageUptodate(new_page); |
1e8f889b | 4452 | |
7269f999 | 4453 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 4454 | haddr + huge_page_size(h)); |
ac46d4f3 | 4455 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 4456 | |
b76c8cfb | 4457 | /* |
cb900f41 | 4458 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
4459 | * before the page tables are altered |
4460 | */ | |
cb900f41 | 4461 | spin_lock(ptl); |
5b7a1d40 | 4462 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 4463 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 4464 | ClearHPageRestoreReserve(new_page); |
07443a85 | 4465 | |
1e8f889b | 4466 | /* Break COW */ |
5b7a1d40 | 4467 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 4468 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 4469 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 4470 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 4471 | page_remove_rmap(old_page, true); |
5b7a1d40 | 4472 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
8f251a3d | 4473 | SetHPageMigratable(new_page); |
1e8f889b DG |
4474 | /* Make the old page be freed below */ |
4475 | new_page = old_page; | |
4476 | } | |
cb900f41 | 4477 | spin_unlock(ptl); |
ac46d4f3 | 4478 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 4479 | out_release_all: |
5b7a1d40 | 4480 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 4481 | put_page(new_page); |
ad4404a2 | 4482 | out_release_old: |
09cbfeaf | 4483 | put_page(old_page); |
8312034f | 4484 | |
ad4404a2 DB |
4485 | spin_lock(ptl); /* Caller expects lock to be held */ |
4486 | return ret; | |
1e8f889b DG |
4487 | } |
4488 | ||
04f2cbe3 | 4489 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
4490 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
4491 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
4492 | { |
4493 | struct address_space *mapping; | |
e7c4b0bf | 4494 | pgoff_t idx; |
04f2cbe3 MG |
4495 | |
4496 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 4497 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
4498 | |
4499 | return find_lock_page(mapping, idx); | |
4500 | } | |
4501 | ||
3ae77f43 HD |
4502 | /* |
4503 | * Return whether there is a pagecache page to back given address within VMA. | |
4504 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4505 | */ | |
4506 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
4507 | struct vm_area_struct *vma, unsigned long address) |
4508 | { | |
4509 | struct address_space *mapping; | |
4510 | pgoff_t idx; | |
4511 | struct page *page; | |
4512 | ||
4513 | mapping = vma->vm_file->f_mapping; | |
4514 | idx = vma_hugecache_offset(h, vma, address); | |
4515 | ||
4516 | page = find_get_page(mapping, idx); | |
4517 | if (page) | |
4518 | put_page(page); | |
4519 | return page != NULL; | |
4520 | } | |
4521 | ||
ab76ad54 MK |
4522 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
4523 | pgoff_t idx) | |
4524 | { | |
4525 | struct inode *inode = mapping->host; | |
4526 | struct hstate *h = hstate_inode(inode); | |
4527 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4528 | ||
4529 | if (err) | |
4530 | return err; | |
d6995da3 | 4531 | ClearHPageRestoreReserve(page); |
ab76ad54 | 4532 | |
22146c3c MK |
4533 | /* |
4534 | * set page dirty so that it will not be removed from cache/file | |
4535 | * by non-hugetlbfs specific code paths. | |
4536 | */ | |
4537 | set_page_dirty(page); | |
4538 | ||
ab76ad54 MK |
4539 | spin_lock(&inode->i_lock); |
4540 | inode->i_blocks += blocks_per_huge_page(h); | |
4541 | spin_unlock(&inode->i_lock); | |
4542 | return 0; | |
4543 | } | |
4544 | ||
7677f7fd AR |
4545 | static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, |
4546 | struct address_space *mapping, | |
4547 | pgoff_t idx, | |
4548 | unsigned int flags, | |
4549 | unsigned long haddr, | |
4550 | unsigned long reason) | |
4551 | { | |
4552 | vm_fault_t ret; | |
4553 | u32 hash; | |
4554 | struct vm_fault vmf = { | |
4555 | .vma = vma, | |
4556 | .address = haddr, | |
4557 | .flags = flags, | |
4558 | ||
4559 | /* | |
4560 | * Hard to debug if it ends up being | |
4561 | * used by a callee that assumes | |
4562 | * something about the other | |
4563 | * uninitialized fields... same as in | |
4564 | * memory.c | |
4565 | */ | |
4566 | }; | |
4567 | ||
4568 | /* | |
4569 | * hugetlb_fault_mutex and i_mmap_rwsem must be | |
4570 | * dropped before handling userfault. Reacquire | |
4571 | * after handling fault to make calling code simpler. | |
4572 | */ | |
4573 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4574 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4575 | i_mmap_unlock_read(mapping); | |
4576 | ret = handle_userfault(&vmf, reason); | |
4577 | i_mmap_lock_read(mapping); | |
4578 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
4579 | ||
4580 | return ret; | |
4581 | } | |
4582 | ||
2b740303 SJ |
4583 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
4584 | struct vm_area_struct *vma, | |
4585 | struct address_space *mapping, pgoff_t idx, | |
4586 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 4587 | { |
a5516438 | 4588 | struct hstate *h = hstate_vma(vma); |
2b740303 | 4589 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 4590 | int anon_rmap = 0; |
4c887265 | 4591 | unsigned long size; |
4c887265 | 4592 | struct page *page; |
1e8f889b | 4593 | pte_t new_pte; |
cb900f41 | 4594 | spinlock_t *ptl; |
285b8dca | 4595 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 4596 | bool new_page = false; |
4c887265 | 4597 | |
04f2cbe3 MG |
4598 | /* |
4599 | * Currently, we are forced to kill the process in the event the | |
4600 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4601 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4602 | */ |
4603 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4604 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4605 | current->pid); |
04f2cbe3 MG |
4606 | return ret; |
4607 | } | |
4608 | ||
4c887265 | 4609 | /* |
87bf91d3 MK |
4610 | * We can not race with truncation due to holding i_mmap_rwsem. |
4611 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4612 | * once for faults beyond end of file. | |
4c887265 | 4613 | */ |
87bf91d3 MK |
4614 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4615 | if (idx >= size) | |
4616 | goto out; | |
4617 | ||
6bda666a CL |
4618 | retry: |
4619 | page = find_lock_page(mapping, idx); | |
4620 | if (!page) { | |
7677f7fd | 4621 | /* Check for page in userfault range */ |
1a1aad8a | 4622 | if (userfaultfd_missing(vma)) { |
7677f7fd AR |
4623 | ret = hugetlb_handle_userfault(vma, mapping, idx, |
4624 | flags, haddr, | |
4625 | VM_UFFD_MISSING); | |
1a1aad8a MK |
4626 | goto out; |
4627 | } | |
4628 | ||
285b8dca | 4629 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4630 | if (IS_ERR(page)) { |
4643d67e MK |
4631 | /* |
4632 | * Returning error will result in faulting task being | |
4633 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4634 | * tasks from racing to fault in the same page which | |
4635 | * could result in false unable to allocate errors. | |
4636 | * Page migration does not take the fault mutex, but | |
4637 | * does a clear then write of pte's under page table | |
4638 | * lock. Page fault code could race with migration, | |
4639 | * notice the clear pte and try to allocate a page | |
4640 | * here. Before returning error, get ptl and make | |
4641 | * sure there really is no pte entry. | |
4642 | */ | |
4643 | ptl = huge_pte_lock(h, mm, ptep); | |
d83e6c8a ML |
4644 | ret = 0; |
4645 | if (huge_pte_none(huge_ptep_get(ptep))) | |
4646 | ret = vmf_error(PTR_ERR(page)); | |
4643d67e | 4647 | spin_unlock(ptl); |
6bda666a CL |
4648 | goto out; |
4649 | } | |
47ad8475 | 4650 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4651 | __SetPageUptodate(page); |
cb6acd01 | 4652 | new_page = true; |
ac9b9c66 | 4653 | |
f83a275d | 4654 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4655 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4656 | if (err) { |
4657 | put_page(page); | |
6bda666a CL |
4658 | if (err == -EEXIST) |
4659 | goto retry; | |
4660 | goto out; | |
4661 | } | |
23be7468 | 4662 | } else { |
6bda666a | 4663 | lock_page(page); |
0fe6e20b NH |
4664 | if (unlikely(anon_vma_prepare(vma))) { |
4665 | ret = VM_FAULT_OOM; | |
4666 | goto backout_unlocked; | |
4667 | } | |
409eb8c2 | 4668 | anon_rmap = 1; |
23be7468 | 4669 | } |
0fe6e20b | 4670 | } else { |
998b4382 NH |
4671 | /* |
4672 | * If memory error occurs between mmap() and fault, some process | |
4673 | * don't have hwpoisoned swap entry for errored virtual address. | |
4674 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4675 | */ | |
4676 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 4677 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4678 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4679 | goto backout_unlocked; |
4680 | } | |
7677f7fd AR |
4681 | |
4682 | /* Check for page in userfault range. */ | |
4683 | if (userfaultfd_minor(vma)) { | |
4684 | unlock_page(page); | |
4685 | put_page(page); | |
4686 | ret = hugetlb_handle_userfault(vma, mapping, idx, | |
4687 | flags, haddr, | |
4688 | VM_UFFD_MINOR); | |
4689 | goto out; | |
4690 | } | |
6bda666a | 4691 | } |
1e8f889b | 4692 | |
57303d80 AW |
4693 | /* |
4694 | * If we are going to COW a private mapping later, we examine the | |
4695 | * pending reservations for this page now. This will ensure that | |
4696 | * any allocations necessary to record that reservation occur outside | |
4697 | * the spinlock. | |
4698 | */ | |
5e911373 | 4699 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4700 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4701 | ret = VM_FAULT_OOM; |
4702 | goto backout_unlocked; | |
4703 | } | |
5e911373 | 4704 | /* Just decrements count, does not deallocate */ |
285b8dca | 4705 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4706 | } |
57303d80 | 4707 | |
8bea8052 | 4708 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 4709 | ret = 0; |
7f2e9525 | 4710 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4711 | goto backout; |
4712 | ||
07443a85 | 4713 | if (anon_rmap) { |
d6995da3 | 4714 | ClearHPageRestoreReserve(page); |
285b8dca | 4715 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4716 | } else |
53f9263b | 4717 | page_dup_rmap(page, true); |
1e8f889b DG |
4718 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4719 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4720 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4721 | |
5d317b2b | 4722 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4723 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4724 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4725 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4726 | } |
4727 | ||
cb900f41 | 4728 | spin_unlock(ptl); |
cb6acd01 MK |
4729 | |
4730 | /* | |
8f251a3d MK |
4731 | * Only set HPageMigratable in newly allocated pages. Existing pages |
4732 | * found in the pagecache may not have HPageMigratableset if they have | |
4733 | * been isolated for migration. | |
cb6acd01 MK |
4734 | */ |
4735 | if (new_page) | |
8f251a3d | 4736 | SetHPageMigratable(page); |
cb6acd01 | 4737 | |
4c887265 AL |
4738 | unlock_page(page); |
4739 | out: | |
ac9b9c66 | 4740 | return ret; |
4c887265 AL |
4741 | |
4742 | backout: | |
cb900f41 | 4743 | spin_unlock(ptl); |
2b26736c | 4744 | backout_unlocked: |
4c887265 | 4745 | unlock_page(page); |
285b8dca | 4746 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4747 | put_page(page); |
4748 | goto out; | |
ac9b9c66 HD |
4749 | } |
4750 | ||
8382d914 | 4751 | #ifdef CONFIG_SMP |
188b04a7 | 4752 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4753 | { |
4754 | unsigned long key[2]; | |
4755 | u32 hash; | |
4756 | ||
1b426bac MK |
4757 | key[0] = (unsigned long) mapping; |
4758 | key[1] = idx; | |
8382d914 | 4759 | |
55254636 | 4760 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4761 | |
4762 | return hash & (num_fault_mutexes - 1); | |
4763 | } | |
4764 | #else | |
4765 | /* | |
6c26d310 | 4766 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
4767 | * return 0 and avoid the hashing overhead. |
4768 | */ | |
188b04a7 | 4769 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4770 | { |
4771 | return 0; | |
4772 | } | |
4773 | #endif | |
4774 | ||
2b740303 | 4775 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4776 | unsigned long address, unsigned int flags) |
86e5216f | 4777 | { |
8382d914 | 4778 | pte_t *ptep, entry; |
cb900f41 | 4779 | spinlock_t *ptl; |
2b740303 | 4780 | vm_fault_t ret; |
8382d914 DB |
4781 | u32 hash; |
4782 | pgoff_t idx; | |
0fe6e20b | 4783 | struct page *page = NULL; |
57303d80 | 4784 | struct page *pagecache_page = NULL; |
a5516438 | 4785 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4786 | struct address_space *mapping; |
0f792cf9 | 4787 | int need_wait_lock = 0; |
285b8dca | 4788 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4789 | |
285b8dca | 4790 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 4791 | if (ptep) { |
c0d0381a MK |
4792 | /* |
4793 | * Since we hold no locks, ptep could be stale. That is | |
4794 | * OK as we are only making decisions based on content and | |
4795 | * not actually modifying content here. | |
4796 | */ | |
fd6a03ed | 4797 | entry = huge_ptep_get(ptep); |
290408d4 | 4798 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4799 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4800 | return 0; |
4801 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4802 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4803 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
4804 | } |
4805 | ||
c0d0381a MK |
4806 | /* |
4807 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
4808 | * until finished with ptep. This serves two purposes: |
4809 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
4810 | * and making the ptep no longer valid. | |
4811 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
4812 | * |
4813 | * ptep could have already be assigned via huge_pte_offset. That | |
4814 | * is OK, as huge_pte_alloc will return the same value unless | |
4815 | * something has changed. | |
4816 | */ | |
8382d914 | 4817 | mapping = vma->vm_file->f_mapping; |
c0d0381a | 4818 | i_mmap_lock_read(mapping); |
aec44e0f | 4819 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); |
c0d0381a MK |
4820 | if (!ptep) { |
4821 | i_mmap_unlock_read(mapping); | |
4822 | return VM_FAULT_OOM; | |
4823 | } | |
8382d914 | 4824 | |
3935baa9 DG |
4825 | /* |
4826 | * Serialize hugepage allocation and instantiation, so that we don't | |
4827 | * get spurious allocation failures if two CPUs race to instantiate | |
4828 | * the same page in the page cache. | |
4829 | */ | |
c0d0381a | 4830 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 4831 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4832 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4833 | |
7f2e9525 GS |
4834 | entry = huge_ptep_get(ptep); |
4835 | if (huge_pte_none(entry)) { | |
8382d914 | 4836 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4837 | goto out_mutex; |
3935baa9 | 4838 | } |
86e5216f | 4839 | |
83c54070 | 4840 | ret = 0; |
1e8f889b | 4841 | |
0f792cf9 NH |
4842 | /* |
4843 | * entry could be a migration/hwpoison entry at this point, so this | |
4844 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
4845 | * an active hugepage in pagecache. This goto expects the 2nd page |
4846 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
4847 | * properly handle it. | |
0f792cf9 NH |
4848 | */ |
4849 | if (!pte_present(entry)) | |
4850 | goto out_mutex; | |
4851 | ||
57303d80 AW |
4852 | /* |
4853 | * If we are going to COW the mapping later, we examine the pending | |
4854 | * reservations for this page now. This will ensure that any | |
4855 | * allocations necessary to record that reservation occur outside the | |
4856 | * spinlock. For private mappings, we also lookup the pagecache | |
4857 | * page now as it is used to determine if a reservation has been | |
4858 | * consumed. | |
4859 | */ | |
106c992a | 4860 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4861 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4862 | ret = VM_FAULT_OOM; |
b4d1d99f | 4863 | goto out_mutex; |
2b26736c | 4864 | } |
5e911373 | 4865 | /* Just decrements count, does not deallocate */ |
285b8dca | 4866 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4867 | |
f83a275d | 4868 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4869 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4870 | vma, haddr); |
57303d80 AW |
4871 | } |
4872 | ||
0f792cf9 NH |
4873 | ptl = huge_pte_lock(h, mm, ptep); |
4874 | ||
4875 | /* Check for a racing update before calling hugetlb_cow */ | |
4876 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4877 | goto out_ptl; | |
4878 | ||
56c9cfb1 NH |
4879 | /* |
4880 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4881 | * pagecache_page, so here we need take the former one | |
4882 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4883 | */ |
4884 | page = pte_page(entry); | |
4885 | if (page != pagecache_page) | |
0f792cf9 NH |
4886 | if (!trylock_page(page)) { |
4887 | need_wait_lock = 1; | |
4888 | goto out_ptl; | |
4889 | } | |
b4d1d99f | 4890 | |
0f792cf9 | 4891 | get_page(page); |
b4d1d99f | 4892 | |
788c7df4 | 4893 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4894 | if (!huge_pte_write(entry)) { |
974e6d66 | 4895 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4896 | pagecache_page, ptl); |
0f792cf9 | 4897 | goto out_put_page; |
b4d1d99f | 4898 | } |
106c992a | 4899 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4900 | } |
4901 | entry = pte_mkyoung(entry); | |
285b8dca | 4902 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4903 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4904 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4905 | out_put_page: |
4906 | if (page != pagecache_page) | |
4907 | unlock_page(page); | |
4908 | put_page(page); | |
cb900f41 KS |
4909 | out_ptl: |
4910 | spin_unlock(ptl); | |
57303d80 AW |
4911 | |
4912 | if (pagecache_page) { | |
4913 | unlock_page(pagecache_page); | |
4914 | put_page(pagecache_page); | |
4915 | } | |
b4d1d99f | 4916 | out_mutex: |
c672c7f2 | 4917 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4918 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
4919 | /* |
4920 | * Generally it's safe to hold refcount during waiting page lock. But | |
4921 | * here we just wait to defer the next page fault to avoid busy loop and | |
4922 | * the page is not used after unlocked before returning from the current | |
4923 | * page fault. So we are safe from accessing freed page, even if we wait | |
4924 | * here without taking refcount. | |
4925 | */ | |
4926 | if (need_wait_lock) | |
4927 | wait_on_page_locked(page); | |
1e8f889b | 4928 | return ret; |
86e5216f AL |
4929 | } |
4930 | ||
714c1891 | 4931 | #ifdef CONFIG_USERFAULTFD |
8fb5debc MK |
4932 | /* |
4933 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4934 | * modifications for huge pages. | |
4935 | */ | |
4936 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4937 | pte_t *dst_pte, | |
4938 | struct vm_area_struct *dst_vma, | |
4939 | unsigned long dst_addr, | |
4940 | unsigned long src_addr, | |
f6191471 | 4941 | enum mcopy_atomic_mode mode, |
8fb5debc MK |
4942 | struct page **pagep) |
4943 | { | |
f6191471 | 4944 | bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); |
1e392147 AA |
4945 | struct address_space *mapping; |
4946 | pgoff_t idx; | |
4947 | unsigned long size; | |
1c9e8def | 4948 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4949 | struct hstate *h = hstate_vma(dst_vma); |
4950 | pte_t _dst_pte; | |
4951 | spinlock_t *ptl; | |
4952 | int ret; | |
4953 | struct page *page; | |
f6191471 | 4954 | int writable; |
8fb5debc | 4955 | |
f6191471 AR |
4956 | mapping = dst_vma->vm_file->f_mapping; |
4957 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4958 | ||
4959 | if (is_continue) { | |
4960 | ret = -EFAULT; | |
4961 | page = find_lock_page(mapping, idx); | |
4962 | if (!page) | |
4963 | goto out; | |
4964 | } else if (!*pagep) { | |
d84cf06e MA |
4965 | /* If a page already exists, then it's UFFDIO_COPY for |
4966 | * a non-missing case. Return -EEXIST. | |
4967 | */ | |
4968 | if (vm_shared && | |
4969 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
4970 | ret = -EEXIST; | |
4971 | goto out; | |
4972 | } | |
4973 | ||
8fb5debc | 4974 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
d84cf06e MA |
4975 | if (IS_ERR(page)) { |
4976 | ret = -ENOMEM; | |
8fb5debc | 4977 | goto out; |
d84cf06e | 4978 | } |
8fb5debc MK |
4979 | |
4980 | ret = copy_huge_page_from_user(page, | |
4981 | (const void __user *) src_addr, | |
810a56b9 | 4982 | pages_per_huge_page(h), false); |
8fb5debc | 4983 | |
c1e8d7c6 | 4984 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 4985 | if (unlikely(ret)) { |
9e368259 | 4986 | ret = -ENOENT; |
8fb5debc MK |
4987 | *pagep = page; |
4988 | /* don't free the page */ | |
4989 | goto out; | |
4990 | } | |
4991 | } else { | |
4992 | page = *pagep; | |
4993 | *pagep = NULL; | |
4994 | } | |
4995 | ||
4996 | /* | |
4997 | * The memory barrier inside __SetPageUptodate makes sure that | |
4998 | * preceding stores to the page contents become visible before | |
4999 | * the set_pte_at() write. | |
5000 | */ | |
5001 | __SetPageUptodate(page); | |
8fb5debc | 5002 | |
f6191471 AR |
5003 | /* Add shared, newly allocated pages to the page cache. */ |
5004 | if (vm_shared && !is_continue) { | |
1e392147 AA |
5005 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5006 | ret = -EFAULT; | |
5007 | if (idx >= size) | |
5008 | goto out_release_nounlock; | |
1c9e8def | 5009 | |
1e392147 AA |
5010 | /* |
5011 | * Serialization between remove_inode_hugepages() and | |
5012 | * huge_add_to_page_cache() below happens through the | |
5013 | * hugetlb_fault_mutex_table that here must be hold by | |
5014 | * the caller. | |
5015 | */ | |
1c9e8def MK |
5016 | ret = huge_add_to_page_cache(page, mapping, idx); |
5017 | if (ret) | |
5018 | goto out_release_nounlock; | |
5019 | } | |
5020 | ||
8fb5debc MK |
5021 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
5022 | spin_lock(ptl); | |
5023 | ||
1e392147 AA |
5024 | /* |
5025 | * Recheck the i_size after holding PT lock to make sure not | |
5026 | * to leave any page mapped (as page_mapped()) beyond the end | |
5027 | * of the i_size (remove_inode_hugepages() is strict about | |
5028 | * enforcing that). If we bail out here, we'll also leave a | |
5029 | * page in the radix tree in the vm_shared case beyond the end | |
5030 | * of the i_size, but remove_inode_hugepages() will take care | |
5031 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
5032 | */ | |
5033 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
5034 | ret = -EFAULT; | |
5035 | if (idx >= size) | |
5036 | goto out_release_unlock; | |
5037 | ||
8fb5debc MK |
5038 | ret = -EEXIST; |
5039 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
5040 | goto out_release_unlock; | |
5041 | ||
1c9e8def MK |
5042 | if (vm_shared) { |
5043 | page_dup_rmap(page, true); | |
5044 | } else { | |
d6995da3 | 5045 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
5046 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
5047 | } | |
8fb5debc | 5048 | |
f6191471 AR |
5049 | /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ |
5050 | if (is_continue && !vm_shared) | |
5051 | writable = 0; | |
5052 | else | |
5053 | writable = dst_vma->vm_flags & VM_WRITE; | |
5054 | ||
5055 | _dst_pte = make_huge_pte(dst_vma, page, writable); | |
5056 | if (writable) | |
8fb5debc MK |
5057 | _dst_pte = huge_pte_mkdirty(_dst_pte); |
5058 | _dst_pte = pte_mkyoung(_dst_pte); | |
5059 | ||
5060 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
5061 | ||
5062 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
5063 | dst_vma->vm_flags & VM_WRITE); | |
5064 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
5065 | ||
5066 | /* No need to invalidate - it was non-present before */ | |
5067 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
5068 | ||
5069 | spin_unlock(ptl); | |
f6191471 AR |
5070 | if (!is_continue) |
5071 | SetHPageMigratable(page); | |
5072 | if (vm_shared || is_continue) | |
1c9e8def | 5073 | unlock_page(page); |
8fb5debc MK |
5074 | ret = 0; |
5075 | out: | |
5076 | return ret; | |
5077 | out_release_unlock: | |
5078 | spin_unlock(ptl); | |
f6191471 | 5079 | if (vm_shared || is_continue) |
1c9e8def | 5080 | unlock_page(page); |
5af10dfd | 5081 | out_release_nounlock: |
846be085 | 5082 | restore_reserve_on_error(h, dst_vma, dst_addr, page); |
8fb5debc MK |
5083 | put_page(page); |
5084 | goto out; | |
5085 | } | |
714c1891 | 5086 | #endif /* CONFIG_USERFAULTFD */ |
8fb5debc | 5087 | |
82e5d378 JM |
5088 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
5089 | int refs, struct page **pages, | |
5090 | struct vm_area_struct **vmas) | |
5091 | { | |
5092 | int nr; | |
5093 | ||
5094 | for (nr = 0; nr < refs; nr++) { | |
5095 | if (likely(pages)) | |
5096 | pages[nr] = mem_map_offset(page, nr); | |
5097 | if (vmas) | |
5098 | vmas[nr] = vma; | |
5099 | } | |
5100 | } | |
5101 | ||
28a35716 ML |
5102 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
5103 | struct page **pages, struct vm_area_struct **vmas, | |
5104 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 5105 | long i, unsigned int flags, int *locked) |
63551ae0 | 5106 | { |
d5d4b0aa KC |
5107 | unsigned long pfn_offset; |
5108 | unsigned long vaddr = *position; | |
28a35716 | 5109 | unsigned long remainder = *nr_pages; |
a5516438 | 5110 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 5111 | int err = -EFAULT, refs; |
63551ae0 | 5112 | |
63551ae0 | 5113 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 5114 | pte_t *pte; |
cb900f41 | 5115 | spinlock_t *ptl = NULL; |
2a15efc9 | 5116 | int absent; |
4c887265 | 5117 | struct page *page; |
63551ae0 | 5118 | |
02057967 DR |
5119 | /* |
5120 | * If we have a pending SIGKILL, don't keep faulting pages and | |
5121 | * potentially allocating memory. | |
5122 | */ | |
fa45f116 | 5123 | if (fatal_signal_pending(current)) { |
02057967 DR |
5124 | remainder = 0; |
5125 | break; | |
5126 | } | |
5127 | ||
4c887265 AL |
5128 | /* |
5129 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 5130 | * each hugepage. We have to make sure we get the |
4c887265 | 5131 | * first, for the page indexing below to work. |
cb900f41 KS |
5132 | * |
5133 | * Note that page table lock is not held when pte is null. | |
4c887265 | 5134 | */ |
7868a208 PA |
5135 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
5136 | huge_page_size(h)); | |
cb900f41 KS |
5137 | if (pte) |
5138 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
5139 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
5140 | ||
5141 | /* | |
5142 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
5143 | * an error where there's an empty slot with no huge pagecache |
5144 | * to back it. This way, we avoid allocating a hugepage, and | |
5145 | * the sparse dumpfile avoids allocating disk blocks, but its | |
5146 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 5147 | */ |
3ae77f43 HD |
5148 | if (absent && (flags & FOLL_DUMP) && |
5149 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
5150 | if (pte) |
5151 | spin_unlock(ptl); | |
2a15efc9 HD |
5152 | remainder = 0; |
5153 | break; | |
5154 | } | |
63551ae0 | 5155 | |
9cc3a5bd NH |
5156 | /* |
5157 | * We need call hugetlb_fault for both hugepages under migration | |
5158 | * (in which case hugetlb_fault waits for the migration,) and | |
5159 | * hwpoisoned hugepages (in which case we need to prevent the | |
5160 | * caller from accessing to them.) In order to do this, we use | |
5161 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
5162 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
5163 | * both cases, and because we can't follow correct pages | |
5164 | * directly from any kind of swap entries. | |
5165 | */ | |
5166 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
5167 | ((flags & FOLL_WRITE) && |
5168 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 5169 | vm_fault_t ret; |
87ffc118 | 5170 | unsigned int fault_flags = 0; |
63551ae0 | 5171 | |
cb900f41 KS |
5172 | if (pte) |
5173 | spin_unlock(ptl); | |
87ffc118 AA |
5174 | if (flags & FOLL_WRITE) |
5175 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 5176 | if (locked) |
71335f37 PX |
5177 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
5178 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
5179 | if (flags & FOLL_NOWAIT) |
5180 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
5181 | FAULT_FLAG_RETRY_NOWAIT; | |
5182 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
5183 | /* |
5184 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
5185 | * FAULT_FLAG_TRIED can co-exist | |
5186 | */ | |
87ffc118 AA |
5187 | fault_flags |= FAULT_FLAG_TRIED; |
5188 | } | |
5189 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
5190 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 5191 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
5192 | remainder = 0; |
5193 | break; | |
5194 | } | |
5195 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 5196 | if (locked && |
1ac25013 | 5197 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 5198 | *locked = 0; |
87ffc118 AA |
5199 | *nr_pages = 0; |
5200 | /* | |
5201 | * VM_FAULT_RETRY must not return an | |
5202 | * error, it will return zero | |
5203 | * instead. | |
5204 | * | |
5205 | * No need to update "position" as the | |
5206 | * caller will not check it after | |
5207 | * *nr_pages is set to 0. | |
5208 | */ | |
5209 | return i; | |
5210 | } | |
5211 | continue; | |
4c887265 AL |
5212 | } |
5213 | ||
a5516438 | 5214 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 5215 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 5216 | |
acbfb087 ZL |
5217 | /* |
5218 | * If subpage information not requested, update counters | |
5219 | * and skip the same_page loop below. | |
5220 | */ | |
5221 | if (!pages && !vmas && !pfn_offset && | |
5222 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
5223 | (remainder >= pages_per_huge_page(h))) { | |
5224 | vaddr += huge_page_size(h); | |
5225 | remainder -= pages_per_huge_page(h); | |
5226 | i += pages_per_huge_page(h); | |
5227 | spin_unlock(ptl); | |
5228 | continue; | |
5229 | } | |
5230 | ||
82e5d378 JM |
5231 | refs = min3(pages_per_huge_page(h) - pfn_offset, |
5232 | (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); | |
0fa5bc40 | 5233 | |
82e5d378 JM |
5234 | if (pages || vmas) |
5235 | record_subpages_vmas(mem_map_offset(page, pfn_offset), | |
5236 | vma, refs, | |
5237 | likely(pages) ? pages + i : NULL, | |
5238 | vmas ? vmas + i : NULL); | |
63551ae0 | 5239 | |
82e5d378 | 5240 | if (pages) { |
0fa5bc40 JM |
5241 | /* |
5242 | * try_grab_compound_head() should always succeed here, | |
5243 | * because: a) we hold the ptl lock, and b) we've just | |
5244 | * checked that the huge page is present in the page | |
5245 | * tables. If the huge page is present, then the tail | |
5246 | * pages must also be present. The ptl prevents the | |
5247 | * head page and tail pages from being rearranged in | |
5248 | * any way. So this page must be available at this | |
5249 | * point, unless the page refcount overflowed: | |
5250 | */ | |
82e5d378 | 5251 | if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], |
0fa5bc40 JM |
5252 | refs, |
5253 | flags))) { | |
5254 | spin_unlock(ptl); | |
5255 | remainder = 0; | |
5256 | err = -ENOMEM; | |
5257 | break; | |
5258 | } | |
d5d4b0aa | 5259 | } |
82e5d378 JM |
5260 | |
5261 | vaddr += (refs << PAGE_SHIFT); | |
5262 | remainder -= refs; | |
5263 | i += refs; | |
5264 | ||
cb900f41 | 5265 | spin_unlock(ptl); |
63551ae0 | 5266 | } |
28a35716 | 5267 | *nr_pages = remainder; |
87ffc118 AA |
5268 | /* |
5269 | * setting position is actually required only if remainder is | |
5270 | * not zero but it's faster not to add a "if (remainder)" | |
5271 | * branch. | |
5272 | */ | |
63551ae0 DG |
5273 | *position = vaddr; |
5274 | ||
2be7cfed | 5275 | return i ? i : err; |
63551ae0 | 5276 | } |
8f860591 | 5277 | |
7da4d641 | 5278 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
5279 | unsigned long address, unsigned long end, pgprot_t newprot) |
5280 | { | |
5281 | struct mm_struct *mm = vma->vm_mm; | |
5282 | unsigned long start = address; | |
5283 | pte_t *ptep; | |
5284 | pte_t pte; | |
a5516438 | 5285 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 5286 | unsigned long pages = 0; |
dff11abe | 5287 | bool shared_pmd = false; |
ac46d4f3 | 5288 | struct mmu_notifier_range range; |
dff11abe MK |
5289 | |
5290 | /* | |
5291 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 5292 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
5293 | * range if PMD sharing is possible. |
5294 | */ | |
7269f999 JG |
5295 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
5296 | 0, vma, mm, start, end); | |
ac46d4f3 | 5297 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
5298 | |
5299 | BUG_ON(address >= end); | |
ac46d4f3 | 5300 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 5301 | |
ac46d4f3 | 5302 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 5303 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 5304 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 5305 | spinlock_t *ptl; |
7868a208 | 5306 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
5307 | if (!ptep) |
5308 | continue; | |
cb900f41 | 5309 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 5310 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
7da4d641 | 5311 | pages++; |
cb900f41 | 5312 | spin_unlock(ptl); |
dff11abe | 5313 | shared_pmd = true; |
39dde65c | 5314 | continue; |
7da4d641 | 5315 | } |
a8bda28d NH |
5316 | pte = huge_ptep_get(ptep); |
5317 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
5318 | spin_unlock(ptl); | |
5319 | continue; | |
5320 | } | |
5321 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
5322 | swp_entry_t entry = pte_to_swp_entry(pte); | |
5323 | ||
5324 | if (is_write_migration_entry(entry)) { | |
5325 | pte_t newpte; | |
5326 | ||
5327 | make_migration_entry_read(&entry); | |
5328 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
5329 | set_huge_swap_pte_at(mm, address, ptep, |
5330 | newpte, huge_page_size(h)); | |
a8bda28d NH |
5331 | pages++; |
5332 | } | |
5333 | spin_unlock(ptl); | |
5334 | continue; | |
5335 | } | |
5336 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
5337 | pte_t old_pte; |
5338 | ||
5339 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
5340 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 5341 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 5342 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 5343 | pages++; |
8f860591 | 5344 | } |
cb900f41 | 5345 | spin_unlock(ptl); |
8f860591 | 5346 | } |
d833352a | 5347 | /* |
c8c06efa | 5348 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 5349 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 5350 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
5351 | * and that page table be reused and filled with junk. If we actually |
5352 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 5353 | */ |
dff11abe | 5354 | if (shared_pmd) |
ac46d4f3 | 5355 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
5356 | else |
5357 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
5358 | /* |
5359 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
5360 | * page table protection not changing it to point to a new page. | |
5361 | * | |
ad56b738 | 5362 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 5363 | */ |
83cde9e8 | 5364 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 5365 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
5366 | |
5367 | return pages << h->order; | |
8f860591 ZY |
5368 | } |
5369 | ||
33b8f84a MK |
5370 | /* Return true if reservation was successful, false otherwise. */ |
5371 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 5372 | long from, long to, |
5a6fe125 | 5373 | struct vm_area_struct *vma, |
ca16d140 | 5374 | vm_flags_t vm_flags) |
e4e574b7 | 5375 | { |
33b8f84a | 5376 | long chg, add = -1; |
a5516438 | 5377 | struct hstate *h = hstate_inode(inode); |
90481622 | 5378 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 5379 | struct resv_map *resv_map; |
075a61d0 | 5380 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 5381 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 5382 | |
63489f8e MK |
5383 | /* This should never happen */ |
5384 | if (from > to) { | |
5385 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 5386 | return false; |
63489f8e MK |
5387 | } |
5388 | ||
17c9d12e MG |
5389 | /* |
5390 | * Only apply hugepage reservation if asked. At fault time, an | |
5391 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 5392 | * without using reserves |
17c9d12e | 5393 | */ |
ca16d140 | 5394 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 5395 | return true; |
17c9d12e | 5396 | |
a1e78772 MG |
5397 | /* |
5398 | * Shared mappings base their reservation on the number of pages that | |
5399 | * are already allocated on behalf of the file. Private mappings need | |
5400 | * to reserve the full area even if read-only as mprotect() may be | |
5401 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
5402 | */ | |
9119a41e | 5403 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
5404 | /* |
5405 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
5406 | * called for inodes for which resv_maps were created (see | |
5407 | * hugetlbfs_get_inode). | |
5408 | */ | |
4e35f483 | 5409 | resv_map = inode_resv_map(inode); |
9119a41e | 5410 | |
0db9d74e | 5411 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
5412 | |
5413 | } else { | |
e9fe92ae | 5414 | /* Private mapping. */ |
9119a41e | 5415 | resv_map = resv_map_alloc(); |
17c9d12e | 5416 | if (!resv_map) |
33b8f84a | 5417 | return false; |
17c9d12e | 5418 | |
a1e78772 | 5419 | chg = to - from; |
84afd99b | 5420 | |
17c9d12e MG |
5421 | set_vma_resv_map(vma, resv_map); |
5422 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
5423 | } | |
5424 | ||
33b8f84a | 5425 | if (chg < 0) |
c50ac050 | 5426 | goto out_err; |
8a630112 | 5427 | |
33b8f84a MK |
5428 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
5429 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 5430 | goto out_err; |
075a61d0 MA |
5431 | |
5432 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
5433 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
5434 | * of the resv_map. | |
5435 | */ | |
5436 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
5437 | } | |
5438 | ||
1c5ecae3 MK |
5439 | /* |
5440 | * There must be enough pages in the subpool for the mapping. If | |
5441 | * the subpool has a minimum size, there may be some global | |
5442 | * reservations already in place (gbl_reserve). | |
5443 | */ | |
5444 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 5445 | if (gbl_reserve < 0) |
075a61d0 | 5446 | goto out_uncharge_cgroup; |
5a6fe125 MG |
5447 | |
5448 | /* | |
17c9d12e | 5449 | * Check enough hugepages are available for the reservation. |
90481622 | 5450 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 5451 | */ |
33b8f84a | 5452 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 5453 | goto out_put_pages; |
17c9d12e MG |
5454 | |
5455 | /* | |
5456 | * Account for the reservations made. Shared mappings record regions | |
5457 | * that have reservations as they are shared by multiple VMAs. | |
5458 | * When the last VMA disappears, the region map says how much | |
5459 | * the reservation was and the page cache tells how much of | |
5460 | * the reservation was consumed. Private mappings are per-VMA and | |
5461 | * only the consumed reservations are tracked. When the VMA | |
5462 | * disappears, the original reservation is the VMA size and the | |
5463 | * consumed reservations are stored in the map. Hence, nothing | |
5464 | * else has to be done for private mappings here | |
5465 | */ | |
33039678 | 5466 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 5467 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
5468 | |
5469 | if (unlikely(add < 0)) { | |
5470 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 5471 | goto out_put_pages; |
0db9d74e | 5472 | } else if (unlikely(chg > add)) { |
33039678 MK |
5473 | /* |
5474 | * pages in this range were added to the reserve | |
5475 | * map between region_chg and region_add. This | |
5476 | * indicates a race with alloc_huge_page. Adjust | |
5477 | * the subpool and reserve counts modified above | |
5478 | * based on the difference. | |
5479 | */ | |
5480 | long rsv_adjust; | |
5481 | ||
d85aecf2 ML |
5482 | /* |
5483 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
5484 | * reference to h_cg->css. See comment below for detail. | |
5485 | */ | |
075a61d0 MA |
5486 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
5487 | hstate_index(h), | |
5488 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5489 | ||
33039678 MK |
5490 | rsv_adjust = hugepage_subpool_put_pages(spool, |
5491 | chg - add); | |
5492 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
5493 | } else if (h_cg) { |
5494 | /* | |
5495 | * The file_regions will hold their own reference to | |
5496 | * h_cg->css. So we should release the reference held | |
5497 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
5498 | * done. | |
5499 | */ | |
5500 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
5501 | } |
5502 | } | |
33b8f84a MK |
5503 | return true; |
5504 | ||
075a61d0 MA |
5505 | out_put_pages: |
5506 | /* put back original number of pages, chg */ | |
5507 | (void)hugepage_subpool_put_pages(spool, chg); | |
5508 | out_uncharge_cgroup: | |
5509 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5510 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 5511 | out_err: |
5e911373 | 5512 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
5513 | /* Only call region_abort if the region_chg succeeded but the |
5514 | * region_add failed or didn't run. | |
5515 | */ | |
5516 | if (chg >= 0 && add < 0) | |
5517 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
5518 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
5519 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 5520 | return false; |
a43a8c39 KC |
5521 | } |
5522 | ||
b5cec28d MK |
5523 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
5524 | long freed) | |
a43a8c39 | 5525 | { |
a5516438 | 5526 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 5527 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 5528 | long chg = 0; |
90481622 | 5529 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 5530 | long gbl_reserve; |
45c682a6 | 5531 | |
f27a5136 MK |
5532 | /* |
5533 | * Since this routine can be called in the evict inode path for all | |
5534 | * hugetlbfs inodes, resv_map could be NULL. | |
5535 | */ | |
b5cec28d MK |
5536 | if (resv_map) { |
5537 | chg = region_del(resv_map, start, end); | |
5538 | /* | |
5539 | * region_del() can fail in the rare case where a region | |
5540 | * must be split and another region descriptor can not be | |
5541 | * allocated. If end == LONG_MAX, it will not fail. | |
5542 | */ | |
5543 | if (chg < 0) | |
5544 | return chg; | |
5545 | } | |
5546 | ||
45c682a6 | 5547 | spin_lock(&inode->i_lock); |
e4c6f8be | 5548 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
5549 | spin_unlock(&inode->i_lock); |
5550 | ||
1c5ecae3 MK |
5551 | /* |
5552 | * If the subpool has a minimum size, the number of global | |
5553 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
5554 | * |
5555 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
5556 | * won't go negative. | |
1c5ecae3 MK |
5557 | */ |
5558 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5559 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
5560 | |
5561 | return 0; | |
a43a8c39 | 5562 | } |
93f70f90 | 5563 | |
3212b535 SC |
5564 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
5565 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5566 | struct vm_area_struct *vma, | |
5567 | unsigned long addr, pgoff_t idx) | |
5568 | { | |
5569 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5570 | svma->vm_start; | |
5571 | unsigned long sbase = saddr & PUD_MASK; | |
5572 | unsigned long s_end = sbase + PUD_SIZE; | |
5573 | ||
5574 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
5575 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
5576 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
5577 | |
5578 | /* | |
5579 | * match the virtual addresses, permission and the alignment of the | |
5580 | * page table page. | |
5581 | */ | |
5582 | if (pmd_index(addr) != pmd_index(saddr) || | |
5583 | vm_flags != svm_flags || | |
07e51edf | 5584 | !range_in_vma(svma, sbase, s_end)) |
3212b535 SC |
5585 | return 0; |
5586 | ||
5587 | return saddr; | |
5588 | } | |
5589 | ||
31aafb45 | 5590 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
5591 | { |
5592 | unsigned long base = addr & PUD_MASK; | |
5593 | unsigned long end = base + PUD_SIZE; | |
5594 | ||
5595 | /* | |
5596 | * check on proper vm_flags and page table alignment | |
5597 | */ | |
017b1660 | 5598 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
5599 | return true; |
5600 | return false; | |
3212b535 SC |
5601 | } |
5602 | ||
c1991e07 PX |
5603 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
5604 | { | |
5605 | #ifdef CONFIG_USERFAULTFD | |
5606 | if (uffd_disable_huge_pmd_share(vma)) | |
5607 | return false; | |
5608 | #endif | |
5609 | return vma_shareable(vma, addr); | |
5610 | } | |
5611 | ||
017b1660 MK |
5612 | /* |
5613 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5614 | * If yes, adjust start and end to cover range associated with possible | |
5615 | * shared pmd mappings. | |
5616 | */ | |
5617 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5618 | unsigned long *start, unsigned long *end) | |
5619 | { | |
a1ba9da8 LX |
5620 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
5621 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 5622 | |
a1ba9da8 | 5623 | /* |
f0953a1b IM |
5624 | * vma needs to span at least one aligned PUD size, and the range |
5625 | * must be at least partially within in. | |
a1ba9da8 LX |
5626 | */ |
5627 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
5628 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
5629 | return; |
5630 | ||
75802ca6 | 5631 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
5632 | if (*start > v_start) |
5633 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 5634 | |
a1ba9da8 LX |
5635 | if (*end < v_end) |
5636 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
5637 | } |
5638 | ||
3212b535 SC |
5639 | /* |
5640 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5641 | * and returns the corresponding pte. While this is not necessary for the | |
5642 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
5643 | * code much cleaner. |
5644 | * | |
0bf7b64e MK |
5645 | * This routine must be called with i_mmap_rwsem held in at least read mode if |
5646 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
5647 | * table entries associated with the address space. This is important as we | |
5648 | * are setting up sharing based on existing page table entries (mappings). | |
5649 | * | |
5650 | * NOTE: This routine is only called from huge_pte_alloc. Some callers of | |
5651 | * huge_pte_alloc know that sharing is not possible and do not take | |
5652 | * i_mmap_rwsem as a performance optimization. This is handled by the | |
5653 | * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is | |
5654 | * only required for subsequent processing. | |
3212b535 | 5655 | */ |
aec44e0f PX |
5656 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
5657 | unsigned long addr, pud_t *pud) | |
3212b535 | 5658 | { |
3212b535 SC |
5659 | struct address_space *mapping = vma->vm_file->f_mapping; |
5660 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5661 | vma->vm_pgoff; | |
5662 | struct vm_area_struct *svma; | |
5663 | unsigned long saddr; | |
5664 | pte_t *spte = NULL; | |
5665 | pte_t *pte; | |
cb900f41 | 5666 | spinlock_t *ptl; |
3212b535 | 5667 | |
0bf7b64e | 5668 | i_mmap_assert_locked(mapping); |
3212b535 SC |
5669 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
5670 | if (svma == vma) | |
5671 | continue; | |
5672 | ||
5673 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5674 | if (saddr) { | |
7868a208 PA |
5675 | spte = huge_pte_offset(svma->vm_mm, saddr, |
5676 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
5677 | if (spte) { |
5678 | get_page(virt_to_page(spte)); | |
5679 | break; | |
5680 | } | |
5681 | } | |
5682 | } | |
5683 | ||
5684 | if (!spte) | |
5685 | goto out; | |
5686 | ||
8bea8052 | 5687 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 5688 | if (pud_none(*pud)) { |
3212b535 SC |
5689 | pud_populate(mm, pud, |
5690 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5691 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5692 | } else { |
3212b535 | 5693 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5694 | } |
cb900f41 | 5695 | spin_unlock(ptl); |
3212b535 SC |
5696 | out: |
5697 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
5698 | return pte; |
5699 | } | |
5700 | ||
5701 | /* | |
5702 | * unmap huge page backed by shared pte. | |
5703 | * | |
5704 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5705 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5706 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5707 | * | |
c0d0381a | 5708 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
5709 | * |
5710 | * returns: 1 successfully unmapped a shared pte page | |
5711 | * 0 the underlying pte page is not shared, or it is the last user | |
5712 | */ | |
34ae204f MK |
5713 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5714 | unsigned long *addr, pte_t *ptep) | |
3212b535 SC |
5715 | { |
5716 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
5717 | p4d_t *p4d = p4d_offset(pgd, *addr); |
5718 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 | 5719 | |
34ae204f | 5720 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
5721 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
5722 | if (page_count(virt_to_page(ptep)) == 1) | |
5723 | return 0; | |
5724 | ||
5725 | pud_clear(pud); | |
5726 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 5727 | mm_dec_nr_pmds(mm); |
3212b535 SC |
5728 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
5729 | return 1; | |
5730 | } | |
c1991e07 | 5731 | |
9e5fc74c | 5732 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
aec44e0f PX |
5733 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
5734 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
5735 | { |
5736 | return NULL; | |
5737 | } | |
e81f2d22 | 5738 | |
34ae204f MK |
5739 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5740 | unsigned long *addr, pte_t *ptep) | |
e81f2d22 ZZ |
5741 | { |
5742 | return 0; | |
5743 | } | |
017b1660 MK |
5744 | |
5745 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5746 | unsigned long *start, unsigned long *end) | |
5747 | { | |
5748 | } | |
c1991e07 PX |
5749 | |
5750 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
5751 | { | |
5752 | return false; | |
5753 | } | |
3212b535 SC |
5754 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5755 | ||
9e5fc74c | 5756 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 5757 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
5758 | unsigned long addr, unsigned long sz) |
5759 | { | |
5760 | pgd_t *pgd; | |
c2febafc | 5761 | p4d_t *p4d; |
9e5fc74c SC |
5762 | pud_t *pud; |
5763 | pte_t *pte = NULL; | |
5764 | ||
5765 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5766 | p4d = p4d_alloc(mm, pgd, addr); |
5767 | if (!p4d) | |
5768 | return NULL; | |
c2febafc | 5769 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5770 | if (pud) { |
5771 | if (sz == PUD_SIZE) { | |
5772 | pte = (pte_t *)pud; | |
5773 | } else { | |
5774 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 5775 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 5776 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
5777 | else |
5778 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5779 | } | |
5780 | } | |
4e666314 | 5781 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5782 | |
5783 | return pte; | |
5784 | } | |
5785 | ||
9b19df29 PA |
5786 | /* |
5787 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5788 | * entry at address @addr | |
5789 | * | |
8ac0b81a LX |
5790 | * Return: Pointer to page table entry (PUD or PMD) for |
5791 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
5792 | * size @sz doesn't match the hugepage size at this level of the page |
5793 | * table. | |
5794 | */ | |
7868a208 PA |
5795 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5796 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5797 | { |
5798 | pgd_t *pgd; | |
c2febafc | 5799 | p4d_t *p4d; |
8ac0b81a LX |
5800 | pud_t *pud; |
5801 | pmd_t *pmd; | |
9e5fc74c SC |
5802 | |
5803 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5804 | if (!pgd_present(*pgd)) |
5805 | return NULL; | |
5806 | p4d = p4d_offset(pgd, addr); | |
5807 | if (!p4d_present(*p4d)) | |
5808 | return NULL; | |
9b19df29 | 5809 | |
c2febafc | 5810 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
5811 | if (sz == PUD_SIZE) |
5812 | /* must be pud huge, non-present or none */ | |
c2febafc | 5813 | return (pte_t *)pud; |
8ac0b81a | 5814 | if (!pud_present(*pud)) |
9b19df29 | 5815 | return NULL; |
8ac0b81a | 5816 | /* must have a valid entry and size to go further */ |
9b19df29 | 5817 | |
8ac0b81a LX |
5818 | pmd = pmd_offset(pud, addr); |
5819 | /* must be pmd huge, non-present or none */ | |
5820 | return (pte_t *)pmd; | |
9e5fc74c SC |
5821 | } |
5822 | ||
61f77eda NH |
5823 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5824 | ||
5825 | /* | |
5826 | * These functions are overwritable if your architecture needs its own | |
5827 | * behavior. | |
5828 | */ | |
5829 | struct page * __weak | |
5830 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5831 | int write) | |
5832 | { | |
5833 | return ERR_PTR(-EINVAL); | |
5834 | } | |
5835 | ||
4dc71451 AK |
5836 | struct page * __weak |
5837 | follow_huge_pd(struct vm_area_struct *vma, | |
5838 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5839 | { | |
5840 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5841 | return NULL; | |
5842 | } | |
5843 | ||
61f77eda | 5844 | struct page * __weak |
9e5fc74c | 5845 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5846 | pmd_t *pmd, int flags) |
9e5fc74c | 5847 | { |
e66f17ff NH |
5848 | struct page *page = NULL; |
5849 | spinlock_t *ptl; | |
c9d398fa | 5850 | pte_t pte; |
3faa52c0 JH |
5851 | |
5852 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5853 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5854 | (FOLL_PIN | FOLL_GET))) | |
5855 | return NULL; | |
5856 | ||
e66f17ff NH |
5857 | retry: |
5858 | ptl = pmd_lockptr(mm, pmd); | |
5859 | spin_lock(ptl); | |
5860 | /* | |
5861 | * make sure that the address range covered by this pmd is not | |
5862 | * unmapped from other threads. | |
5863 | */ | |
5864 | if (!pmd_huge(*pmd)) | |
5865 | goto out; | |
c9d398fa NH |
5866 | pte = huge_ptep_get((pte_t *)pmd); |
5867 | if (pte_present(pte)) { | |
97534127 | 5868 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
5869 | /* |
5870 | * try_grab_page() should always succeed here, because: a) we | |
5871 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5872 | * huge pmd (head) page is present in the page tables. The ptl | |
5873 | * prevents the head page and tail pages from being rearranged | |
5874 | * in any way. So this page must be available at this point, | |
5875 | * unless the page refcount overflowed: | |
5876 | */ | |
5877 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5878 | page = NULL; | |
5879 | goto out; | |
5880 | } | |
e66f17ff | 5881 | } else { |
c9d398fa | 5882 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5883 | spin_unlock(ptl); |
5884 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5885 | goto retry; | |
5886 | } | |
5887 | /* | |
5888 | * hwpoisoned entry is treated as no_page_table in | |
5889 | * follow_page_mask(). | |
5890 | */ | |
5891 | } | |
5892 | out: | |
5893 | spin_unlock(ptl); | |
9e5fc74c SC |
5894 | return page; |
5895 | } | |
5896 | ||
61f77eda | 5897 | struct page * __weak |
9e5fc74c | 5898 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5899 | pud_t *pud, int flags) |
9e5fc74c | 5900 | { |
3faa52c0 | 5901 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5902 | return NULL; |
9e5fc74c | 5903 | |
e66f17ff | 5904 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5905 | } |
5906 | ||
faaa5b62 AK |
5907 | struct page * __weak |
5908 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5909 | { | |
3faa52c0 | 5910 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5911 | return NULL; |
5912 | ||
5913 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5914 | } | |
5915 | ||
31caf665 NH |
5916 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5917 | { | |
bcc54222 NH |
5918 | bool ret = true; |
5919 | ||
db71ef79 | 5920 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
5921 | if (!PageHeadHuge(page) || |
5922 | !HPageMigratable(page) || | |
0eb2df2b | 5923 | !get_page_unless_zero(page)) { |
bcc54222 NH |
5924 | ret = false; |
5925 | goto unlock; | |
5926 | } | |
8f251a3d | 5927 | ClearHPageMigratable(page); |
31caf665 | 5928 | list_move_tail(&page->lru, list); |
bcc54222 | 5929 | unlock: |
db71ef79 | 5930 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 5931 | return ret; |
31caf665 NH |
5932 | } |
5933 | ||
25182f05 NH |
5934 | int get_hwpoison_huge_page(struct page *page, bool *hugetlb) |
5935 | { | |
5936 | int ret = 0; | |
5937 | ||
5938 | *hugetlb = false; | |
5939 | spin_lock_irq(&hugetlb_lock); | |
5940 | if (PageHeadHuge(page)) { | |
5941 | *hugetlb = true; | |
5942 | if (HPageFreed(page) || HPageMigratable(page)) | |
5943 | ret = get_page_unless_zero(page); | |
5944 | } | |
5945 | spin_unlock_irq(&hugetlb_lock); | |
5946 | return ret; | |
5947 | } | |
5948 | ||
31caf665 NH |
5949 | void putback_active_hugepage(struct page *page) |
5950 | { | |
db71ef79 | 5951 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 5952 | SetHPageMigratable(page); |
31caf665 | 5953 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 5954 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
5955 | put_page(page); |
5956 | } | |
ab5ac90a MH |
5957 | |
5958 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5959 | { | |
5960 | struct hstate *h = page_hstate(oldpage); | |
5961 | ||
5962 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5963 | set_page_owner_migrate_reason(newpage, reason); | |
5964 | ||
5965 | /* | |
5966 | * transfer temporary state of the new huge page. This is | |
5967 | * reverse to other transitions because the newpage is going to | |
5968 | * be final while the old one will be freed so it takes over | |
5969 | * the temporary status. | |
5970 | * | |
5971 | * Also note that we have to transfer the per-node surplus state | |
5972 | * here as well otherwise the global surplus count will not match | |
5973 | * the per-node's. | |
5974 | */ | |
9157c311 | 5975 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
5976 | int old_nid = page_to_nid(oldpage); |
5977 | int new_nid = page_to_nid(newpage); | |
5978 | ||
9157c311 MK |
5979 | SetHPageTemporary(oldpage); |
5980 | ClearHPageTemporary(newpage); | |
ab5ac90a | 5981 | |
5af1ab1d ML |
5982 | /* |
5983 | * There is no need to transfer the per-node surplus state | |
5984 | * when we do not cross the node. | |
5985 | */ | |
5986 | if (new_nid == old_nid) | |
5987 | return; | |
db71ef79 | 5988 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
5989 | if (h->surplus_huge_pages_node[old_nid]) { |
5990 | h->surplus_huge_pages_node[old_nid]--; | |
5991 | h->surplus_huge_pages_node[new_nid]++; | |
5992 | } | |
db71ef79 | 5993 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
5994 | } |
5995 | } | |
cf11e85f | 5996 | |
6dfeaff9 PX |
5997 | /* |
5998 | * This function will unconditionally remove all the shared pmd pgtable entries | |
5999 | * within the specific vma for a hugetlbfs memory range. | |
6000 | */ | |
6001 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
6002 | { | |
6003 | struct hstate *h = hstate_vma(vma); | |
6004 | unsigned long sz = huge_page_size(h); | |
6005 | struct mm_struct *mm = vma->vm_mm; | |
6006 | struct mmu_notifier_range range; | |
6007 | unsigned long address, start, end; | |
6008 | spinlock_t *ptl; | |
6009 | pte_t *ptep; | |
6010 | ||
6011 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
6012 | return; | |
6013 | ||
6014 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
6015 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
6016 | ||
6017 | if (start >= end) | |
6018 | return; | |
6019 | ||
6020 | /* | |
6021 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
6022 | * we have already done the PUD_SIZE alignment. | |
6023 | */ | |
6024 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
6025 | start, end); | |
6026 | mmu_notifier_invalidate_range_start(&range); | |
6027 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
6028 | for (address = start; address < end; address += PUD_SIZE) { | |
6029 | unsigned long tmp = address; | |
6030 | ||
6031 | ptep = huge_pte_offset(mm, address, sz); | |
6032 | if (!ptep) | |
6033 | continue; | |
6034 | ptl = huge_pte_lock(h, mm, ptep); | |
6035 | /* We don't want 'address' to be changed */ | |
6036 | huge_pmd_unshare(mm, vma, &tmp, ptep); | |
6037 | spin_unlock(ptl); | |
6038 | } | |
6039 | flush_hugetlb_tlb_range(vma, start, end); | |
6040 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
6041 | /* | |
6042 | * No need to call mmu_notifier_invalidate_range(), see | |
6043 | * Documentation/vm/mmu_notifier.rst. | |
6044 | */ | |
6045 | mmu_notifier_invalidate_range_end(&range); | |
6046 | } | |
6047 | ||
cf11e85f | 6048 | #ifdef CONFIG_CMA |
cf11e85f RG |
6049 | static bool cma_reserve_called __initdata; |
6050 | ||
6051 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
6052 | { | |
6053 | hugetlb_cma_size = memparse(p, &p); | |
6054 | return 0; | |
6055 | } | |
6056 | ||
6057 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
6058 | ||
6059 | void __init hugetlb_cma_reserve(int order) | |
6060 | { | |
6061 | unsigned long size, reserved, per_node; | |
6062 | int nid; | |
6063 | ||
6064 | cma_reserve_called = true; | |
6065 | ||
6066 | if (!hugetlb_cma_size) | |
6067 | return; | |
6068 | ||
6069 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
6070 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
6071 | (PAGE_SIZE << order) / SZ_1M); | |
6072 | return; | |
6073 | } | |
6074 | ||
6075 | /* | |
6076 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
6077 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
6078 | */ | |
6079 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
6080 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
6081 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
6082 | ||
6083 | reserved = 0; | |
6084 | for_each_node_state(nid, N_ONLINE) { | |
6085 | int res; | |
2281f797 | 6086 | char name[CMA_MAX_NAME]; |
cf11e85f RG |
6087 | |
6088 | size = min(per_node, hugetlb_cma_size - reserved); | |
6089 | size = round_up(size, PAGE_SIZE << order); | |
6090 | ||
2281f797 | 6091 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
cf11e85f | 6092 | res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, |
29d0f41d | 6093 | 0, false, name, |
cf11e85f RG |
6094 | &hugetlb_cma[nid], nid); |
6095 | if (res) { | |
6096 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
6097 | res, nid); | |
6098 | continue; | |
6099 | } | |
6100 | ||
6101 | reserved += size; | |
6102 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
6103 | size / SZ_1M, nid); | |
6104 | ||
6105 | if (reserved >= hugetlb_cma_size) | |
6106 | break; | |
6107 | } | |
6108 | } | |
6109 | ||
6110 | void __init hugetlb_cma_check(void) | |
6111 | { | |
6112 | if (!hugetlb_cma_size || cma_reserve_called) | |
6113 | return; | |
6114 | ||
6115 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
6116 | } | |
6117 | ||
6118 | #endif /* CONFIG_CMA */ |