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