Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * Generic hugetlb support. | |
6d49e352 | 4 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 5 | */ |
1da177e4 LT |
6 | #include <linux/list.h> |
7 | #include <linux/init.h> | |
1da177e4 | 8 | #include <linux/mm.h> |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
97ad1087 | 19 | #include <linux/memblock.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
bbe88753 | 22 | #include <linux/sched/mm.h> |
63489f8e | 23 | #include <linux/mmdebug.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
0fe6e20b | 25 | #include <linux/rmap.h> |
c6247f72 | 26 | #include <linux/string_helpers.h> |
fd6a03ed NH |
27 | #include <linux/swap.h> |
28 | #include <linux/swapops.h> | |
8382d914 | 29 | #include <linux/jhash.h> |
98fa15f3 | 30 | #include <linux/numa.h> |
c77c0a8a | 31 | #include <linux/llist.h> |
cf11e85f | 32 | #include <linux/cma.h> |
8cc5fcbb | 33 | #include <linux/migrate.h> |
f9317f77 | 34 | #include <linux/nospec.h> |
662ce1dc | 35 | #include <linux/delayacct.h> |
b958d4d0 | 36 | #include <linux/memory.h> |
d6606683 | 37 | |
63551ae0 | 38 | #include <asm/page.h> |
ca15ca40 | 39 | #include <asm/pgalloc.h> |
24669e58 | 40 | #include <asm/tlb.h> |
63551ae0 | 41 | |
24669e58 | 42 | #include <linux/io.h> |
63551ae0 | 43 | #include <linux/hugetlb.h> |
9dd540e2 | 44 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 45 | #include <linux/node.h> |
ab5ac90a | 46 | #include <linux/page_owner.h> |
7835e98b | 47 | #include "internal.h" |
f41f2ed4 | 48 | #include "hugetlb_vmemmap.h" |
1da177e4 | 49 | |
c3f38a38 | 50 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
51 | unsigned int default_hstate_idx; |
52 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 53 | |
dbda8fea | 54 | #ifdef CONFIG_CMA |
cf11e85f | 55 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
38e719ab | 56 | static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; |
a01f4390 MK |
57 | static bool hugetlb_cma_page(struct page *page, unsigned int order) |
58 | { | |
59 | return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, | |
60 | 1 << order); | |
61 | } | |
62 | #else | |
63 | static bool hugetlb_cma_page(struct page *page, unsigned int order) | |
64 | { | |
65 | return false; | |
66 | } | |
dbda8fea BS |
67 | #endif |
68 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 69 | |
53ba51d2 JT |
70 | __initdata LIST_HEAD(huge_boot_pages); |
71 | ||
e5ff2159 AK |
72 | /* for command line parsing */ |
73 | static struct hstate * __initdata parsed_hstate; | |
74 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 75 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 76 | static bool __initdata parsed_default_hugepagesz; |
b5389086 | 77 | static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; |
e5ff2159 | 78 | |
3935baa9 | 79 | /* |
31caf665 NH |
80 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
81 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 82 | */ |
c3f38a38 | 83 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 84 | |
8382d914 DB |
85 | /* |
86 | * Serializes faults on the same logical page. This is used to | |
87 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
88 | */ | |
89 | static int num_fault_mutexes; | |
c672c7f2 | 90 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 91 | |
7ca02d0a MK |
92 | /* Forward declaration */ |
93 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
8d9bfb26 MK |
94 | static void hugetlb_vma_lock_free(struct vm_area_struct *vma); |
95 | static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma); | |
7ca02d0a | 96 | |
1d88433b | 97 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 98 | { |
1d88433b ML |
99 | if (spool->count) |
100 | return false; | |
101 | if (spool->max_hpages != -1) | |
102 | return spool->used_hpages == 0; | |
103 | if (spool->min_hpages != -1) | |
104 | return spool->rsv_hpages == spool->min_hpages; | |
105 | ||
106 | return true; | |
107 | } | |
90481622 | 108 | |
db71ef79 MK |
109 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, |
110 | unsigned long irq_flags) | |
1d88433b | 111 | { |
db71ef79 | 112 | spin_unlock_irqrestore(&spool->lock, irq_flags); |
90481622 DG |
113 | |
114 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 115 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 116 | * free the subpool */ |
1d88433b | 117 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
118 | if (spool->min_hpages != -1) |
119 | hugetlb_acct_memory(spool->hstate, | |
120 | -spool->min_hpages); | |
90481622 | 121 | kfree(spool); |
7ca02d0a | 122 | } |
90481622 DG |
123 | } |
124 | ||
7ca02d0a MK |
125 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
126 | long min_hpages) | |
90481622 DG |
127 | { |
128 | struct hugepage_subpool *spool; | |
129 | ||
c6a91820 | 130 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
131 | if (!spool) |
132 | return NULL; | |
133 | ||
134 | spin_lock_init(&spool->lock); | |
135 | spool->count = 1; | |
7ca02d0a MK |
136 | spool->max_hpages = max_hpages; |
137 | spool->hstate = h; | |
138 | spool->min_hpages = min_hpages; | |
139 | ||
140 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
141 | kfree(spool); | |
142 | return NULL; | |
143 | } | |
144 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
145 | |
146 | return spool; | |
147 | } | |
148 | ||
149 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
150 | { | |
db71ef79 MK |
151 | unsigned long flags; |
152 | ||
153 | spin_lock_irqsave(&spool->lock, flags); | |
90481622 DG |
154 | BUG_ON(!spool->count); |
155 | spool->count--; | |
db71ef79 | 156 | unlock_or_release_subpool(spool, flags); |
90481622 DG |
157 | } |
158 | ||
1c5ecae3 MK |
159 | /* |
160 | * Subpool accounting for allocating and reserving pages. | |
161 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 162 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
163 | * global pools must be adjusted (upward). The returned value may |
164 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 165 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
166 | */ |
167 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
168 | long delta) |
169 | { | |
1c5ecae3 | 170 | long ret = delta; |
90481622 DG |
171 | |
172 | if (!spool) | |
1c5ecae3 | 173 | return ret; |
90481622 | 174 | |
db71ef79 | 175 | spin_lock_irq(&spool->lock); |
1c5ecae3 MK |
176 | |
177 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
178 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
179 | spool->used_hpages += delta; | |
180 | else { | |
181 | ret = -ENOMEM; | |
182 | goto unlock_ret; | |
183 | } | |
90481622 | 184 | } |
90481622 | 185 | |
09a95e29 MK |
186 | /* minimum size accounting */ |
187 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
188 | if (delta > spool->rsv_hpages) { |
189 | /* | |
190 | * Asking for more reserves than those already taken on | |
191 | * behalf of subpool. Return difference. | |
192 | */ | |
193 | ret = delta - spool->rsv_hpages; | |
194 | spool->rsv_hpages = 0; | |
195 | } else { | |
196 | ret = 0; /* reserves already accounted for */ | |
197 | spool->rsv_hpages -= delta; | |
198 | } | |
199 | } | |
200 | ||
201 | unlock_ret: | |
db71ef79 | 202 | spin_unlock_irq(&spool->lock); |
90481622 DG |
203 | return ret; |
204 | } | |
205 | ||
1c5ecae3 MK |
206 | /* |
207 | * Subpool accounting for freeing and unreserving pages. | |
208 | * Return the number of global page reservations that must be dropped. | |
209 | * The return value may only be different than the passed value (delta) | |
210 | * in the case where a subpool minimum size must be maintained. | |
211 | */ | |
212 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
213 | long delta) |
214 | { | |
1c5ecae3 | 215 | long ret = delta; |
db71ef79 | 216 | unsigned long flags; |
1c5ecae3 | 217 | |
90481622 | 218 | if (!spool) |
1c5ecae3 | 219 | return delta; |
90481622 | 220 | |
db71ef79 | 221 | spin_lock_irqsave(&spool->lock, flags); |
1c5ecae3 MK |
222 | |
223 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
224 | spool->used_hpages -= delta; | |
225 | ||
09a95e29 MK |
226 | /* minimum size accounting */ |
227 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
228 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
229 | ret = 0; | |
230 | else | |
231 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
232 | ||
233 | spool->rsv_hpages += delta; | |
234 | if (spool->rsv_hpages > spool->min_hpages) | |
235 | spool->rsv_hpages = spool->min_hpages; | |
236 | } | |
237 | ||
238 | /* | |
239 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
240 | * quota reference, free it now. | |
241 | */ | |
db71ef79 | 242 | unlock_or_release_subpool(spool, flags); |
1c5ecae3 MK |
243 | |
244 | return ret; | |
90481622 DG |
245 | } |
246 | ||
247 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
248 | { | |
249 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
250 | } | |
251 | ||
252 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
253 | { | |
496ad9aa | 254 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
255 | } |
256 | ||
0db9d74e MA |
257 | /* Helper that removes a struct file_region from the resv_map cache and returns |
258 | * it for use. | |
259 | */ | |
260 | static struct file_region * | |
261 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
262 | { | |
3259914f | 263 | struct file_region *nrg; |
0db9d74e MA |
264 | |
265 | VM_BUG_ON(resv->region_cache_count <= 0); | |
266 | ||
267 | resv->region_cache_count--; | |
268 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
269 | list_del(&nrg->link); |
270 | ||
271 | nrg->from = from; | |
272 | nrg->to = to; | |
273 | ||
274 | return nrg; | |
275 | } | |
276 | ||
075a61d0 MA |
277 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
278 | struct file_region *rg) | |
279 | { | |
280 | #ifdef CONFIG_CGROUP_HUGETLB | |
281 | nrg->reservation_counter = rg->reservation_counter; | |
282 | nrg->css = rg->css; | |
283 | if (rg->css) | |
284 | css_get(rg->css); | |
285 | #endif | |
286 | } | |
287 | ||
288 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
289 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
290 | struct hstate *h, | |
291 | struct resv_map *resv, | |
292 | struct file_region *nrg) | |
293 | { | |
294 | #ifdef CONFIG_CGROUP_HUGETLB | |
295 | if (h_cg) { | |
296 | nrg->reservation_counter = | |
297 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
298 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
299 | /* |
300 | * The caller will hold exactly one h_cg->css reference for the | |
301 | * whole contiguous reservation region. But this area might be | |
302 | * scattered when there are already some file_regions reside in | |
303 | * it. As a result, many file_regions may share only one css | |
304 | * reference. In order to ensure that one file_region must hold | |
305 | * exactly one h_cg->css reference, we should do css_get for | |
306 | * each file_region and leave the reference held by caller | |
307 | * untouched. | |
308 | */ | |
309 | css_get(&h_cg->css); | |
075a61d0 MA |
310 | if (!resv->pages_per_hpage) |
311 | resv->pages_per_hpage = pages_per_huge_page(h); | |
312 | /* pages_per_hpage should be the same for all entries in | |
313 | * a resv_map. | |
314 | */ | |
315 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
316 | } else { | |
317 | nrg->reservation_counter = NULL; | |
318 | nrg->css = NULL; | |
319 | } | |
320 | #endif | |
321 | } | |
322 | ||
d85aecf2 ML |
323 | static void put_uncharge_info(struct file_region *rg) |
324 | { | |
325 | #ifdef CONFIG_CGROUP_HUGETLB | |
326 | if (rg->css) | |
327 | css_put(rg->css); | |
328 | #endif | |
329 | } | |
330 | ||
a9b3f867 MA |
331 | static bool has_same_uncharge_info(struct file_region *rg, |
332 | struct file_region *org) | |
333 | { | |
334 | #ifdef CONFIG_CGROUP_HUGETLB | |
0739eb43 | 335 | return rg->reservation_counter == org->reservation_counter && |
a9b3f867 MA |
336 | rg->css == org->css; |
337 | ||
338 | #else | |
339 | return true; | |
340 | #endif | |
341 | } | |
342 | ||
343 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
344 | { | |
3259914f | 345 | struct file_region *nrg, *prg; |
a9b3f867 MA |
346 | |
347 | prg = list_prev_entry(rg, link); | |
348 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
349 | has_same_uncharge_info(prg, rg)) { | |
350 | prg->to = rg->to; | |
351 | ||
352 | list_del(&rg->link); | |
d85aecf2 | 353 | put_uncharge_info(rg); |
a9b3f867 MA |
354 | kfree(rg); |
355 | ||
7db5e7b6 | 356 | rg = prg; |
a9b3f867 MA |
357 | } |
358 | ||
359 | nrg = list_next_entry(rg, link); | |
360 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
361 | has_same_uncharge_info(nrg, rg)) { | |
362 | nrg->from = rg->from; | |
363 | ||
364 | list_del(&rg->link); | |
d85aecf2 | 365 | put_uncharge_info(rg); |
a9b3f867 | 366 | kfree(rg); |
a9b3f867 MA |
367 | } |
368 | } | |
369 | ||
2103cf9c | 370 | static inline long |
84448c8e | 371 | hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, |
2103cf9c PX |
372 | long to, struct hstate *h, struct hugetlb_cgroup *cg, |
373 | long *regions_needed) | |
374 | { | |
375 | struct file_region *nrg; | |
376 | ||
377 | if (!regions_needed) { | |
378 | nrg = get_file_region_entry_from_cache(map, from, to); | |
379 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
84448c8e | 380 | list_add(&nrg->link, rg); |
2103cf9c PX |
381 | coalesce_file_region(map, nrg); |
382 | } else | |
383 | *regions_needed += 1; | |
384 | ||
385 | return to - from; | |
386 | } | |
387 | ||
972a3da3 WY |
388 | /* |
389 | * Must be called with resv->lock held. | |
390 | * | |
391 | * Calling this with regions_needed != NULL will count the number of pages | |
392 | * to be added but will not modify the linked list. And regions_needed will | |
393 | * indicate the number of file_regions needed in the cache to carry out to add | |
394 | * the regions for this range. | |
d75c6af9 MA |
395 | */ |
396 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 397 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 398 | struct hstate *h, long *regions_needed) |
d75c6af9 | 399 | { |
0db9d74e | 400 | long add = 0; |
d75c6af9 | 401 | struct list_head *head = &resv->regions; |
0db9d74e | 402 | long last_accounted_offset = f; |
84448c8e JK |
403 | struct file_region *iter, *trg = NULL; |
404 | struct list_head *rg = NULL; | |
d75c6af9 | 405 | |
0db9d74e MA |
406 | if (regions_needed) |
407 | *regions_needed = 0; | |
d75c6af9 | 408 | |
0db9d74e | 409 | /* In this loop, we essentially handle an entry for the range |
84448c8e | 410 | * [last_accounted_offset, iter->from), at every iteration, with some |
0db9d74e MA |
411 | * bounds checking. |
412 | */ | |
84448c8e | 413 | list_for_each_entry_safe(iter, trg, head, link) { |
0db9d74e | 414 | /* Skip irrelevant regions that start before our range. */ |
84448c8e | 415 | if (iter->from < f) { |
0db9d74e MA |
416 | /* If this region ends after the last accounted offset, |
417 | * then we need to update last_accounted_offset. | |
418 | */ | |
84448c8e JK |
419 | if (iter->to > last_accounted_offset) |
420 | last_accounted_offset = iter->to; | |
0db9d74e MA |
421 | continue; |
422 | } | |
d75c6af9 | 423 | |
0db9d74e MA |
424 | /* When we find a region that starts beyond our range, we've |
425 | * finished. | |
426 | */ | |
84448c8e JK |
427 | if (iter->from >= t) { |
428 | rg = iter->link.prev; | |
d75c6af9 | 429 | break; |
84448c8e | 430 | } |
d75c6af9 | 431 | |
84448c8e | 432 | /* Add an entry for last_accounted_offset -> iter->from, and |
0db9d74e MA |
433 | * update last_accounted_offset. |
434 | */ | |
84448c8e JK |
435 | if (iter->from > last_accounted_offset) |
436 | add += hugetlb_resv_map_add(resv, iter->link.prev, | |
2103cf9c | 437 | last_accounted_offset, |
84448c8e | 438 | iter->from, h, h_cg, |
2103cf9c | 439 | regions_needed); |
0db9d74e | 440 | |
84448c8e | 441 | last_accounted_offset = iter->to; |
0db9d74e MA |
442 | } |
443 | ||
444 | /* Handle the case where our range extends beyond | |
445 | * last_accounted_offset. | |
446 | */ | |
84448c8e JK |
447 | if (!rg) |
448 | rg = head->prev; | |
2103cf9c PX |
449 | if (last_accounted_offset < t) |
450 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
451 | t, h, h_cg, regions_needed); | |
0db9d74e | 452 | |
0db9d74e MA |
453 | return add; |
454 | } | |
455 | ||
456 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
457 | */ | |
458 | static int allocate_file_region_entries(struct resv_map *resv, | |
459 | int regions_needed) | |
460 | __must_hold(&resv->lock) | |
461 | { | |
34665341 | 462 | LIST_HEAD(allocated_regions); |
0db9d74e MA |
463 | int to_allocate = 0, i = 0; |
464 | struct file_region *trg = NULL, *rg = NULL; | |
465 | ||
466 | VM_BUG_ON(regions_needed < 0); | |
467 | ||
0db9d74e MA |
468 | /* |
469 | * Check for sufficient descriptors in the cache to accommodate | |
470 | * the number of in progress add operations plus regions_needed. | |
471 | * | |
472 | * This is a while loop because when we drop the lock, some other call | |
473 | * to region_add or region_del may have consumed some region_entries, | |
474 | * so we keep looping here until we finally have enough entries for | |
475 | * (adds_in_progress + regions_needed). | |
476 | */ | |
477 | while (resv->region_cache_count < | |
478 | (resv->adds_in_progress + regions_needed)) { | |
479 | to_allocate = resv->adds_in_progress + regions_needed - | |
480 | resv->region_cache_count; | |
481 | ||
482 | /* At this point, we should have enough entries in the cache | |
f0953a1b | 483 | * for all the existing adds_in_progress. We should only be |
0db9d74e | 484 | * needing to allocate for regions_needed. |
d75c6af9 | 485 | */ |
0db9d74e MA |
486 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
487 | ||
488 | spin_unlock(&resv->lock); | |
489 | for (i = 0; i < to_allocate; i++) { | |
490 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
491 | if (!trg) | |
492 | goto out_of_memory; | |
493 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 494 | } |
d75c6af9 | 495 | |
0db9d74e MA |
496 | spin_lock(&resv->lock); |
497 | ||
d3ec7b6e WY |
498 | list_splice(&allocated_regions, &resv->region_cache); |
499 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
500 | } |
501 | ||
0db9d74e | 502 | return 0; |
d75c6af9 | 503 | |
0db9d74e MA |
504 | out_of_memory: |
505 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
506 | list_del(&rg->link); | |
507 | kfree(rg); | |
508 | } | |
509 | return -ENOMEM; | |
d75c6af9 MA |
510 | } |
511 | ||
1dd308a7 MK |
512 | /* |
513 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
514 | * map. Regions will be taken from the cache to fill in this range. |
515 | * Sufficient regions should exist in the cache due to the previous | |
516 | * call to region_chg with the same range, but in some cases the cache will not | |
517 | * have sufficient entries due to races with other code doing region_add or | |
518 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 519 | * |
0db9d74e MA |
520 | * regions_needed is the out value provided by a previous call to region_chg. |
521 | * | |
522 | * Return the number of new huge pages added to the map. This number is greater | |
523 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 524 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
525 | * region_add of regions of length 1 never allocate file_regions and cannot |
526 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
527 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 528 | */ |
0db9d74e | 529 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
530 | long in_regions_needed, struct hstate *h, |
531 | struct hugetlb_cgroup *h_cg) | |
96822904 | 532 | { |
0db9d74e | 533 | long add = 0, actual_regions_needed = 0; |
96822904 | 534 | |
7b24d861 | 535 | spin_lock(&resv->lock); |
0db9d74e MA |
536 | retry: |
537 | ||
538 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
539 | add_reservation_in_range(resv, f, t, NULL, NULL, |
540 | &actual_regions_needed); | |
96822904 | 541 | |
5e911373 | 542 | /* |
0db9d74e MA |
543 | * Check for sufficient descriptors in the cache to accommodate |
544 | * this add operation. Note that actual_regions_needed may be greater | |
545 | * than in_regions_needed, as the resv_map may have been modified since | |
546 | * the region_chg call. In this case, we need to make sure that we | |
547 | * allocate extra entries, such that we have enough for all the | |
548 | * existing adds_in_progress, plus the excess needed for this | |
549 | * operation. | |
5e911373 | 550 | */ |
0db9d74e MA |
551 | if (actual_regions_needed > in_regions_needed && |
552 | resv->region_cache_count < | |
553 | resv->adds_in_progress + | |
554 | (actual_regions_needed - in_regions_needed)) { | |
555 | /* region_add operation of range 1 should never need to | |
556 | * allocate file_region entries. | |
557 | */ | |
558 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 559 | |
0db9d74e MA |
560 | if (allocate_file_region_entries( |
561 | resv, actual_regions_needed - in_regions_needed)) { | |
562 | return -ENOMEM; | |
563 | } | |
5e911373 | 564 | |
0db9d74e | 565 | goto retry; |
5e911373 MK |
566 | } |
567 | ||
972a3da3 | 568 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
569 | |
570 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 571 | |
7b24d861 | 572 | spin_unlock(&resv->lock); |
cf3ad20b | 573 | return add; |
96822904 AW |
574 | } |
575 | ||
1dd308a7 MK |
576 | /* |
577 | * Examine the existing reserve map and determine how many | |
578 | * huge pages in the specified range [f, t) are NOT currently | |
579 | * represented. This routine is called before a subsequent | |
580 | * call to region_add that will actually modify the reserve | |
581 | * map to add the specified range [f, t). region_chg does | |
582 | * not change the number of huge pages represented by the | |
0db9d74e MA |
583 | * map. A number of new file_region structures is added to the cache as a |
584 | * placeholder, for the subsequent region_add call to use. At least 1 | |
585 | * file_region structure is added. | |
586 | * | |
587 | * out_regions_needed is the number of regions added to the | |
588 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
589 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
590 | * |
591 | * Returns the number of huge pages that need to be added to the existing | |
592 | * reservation map for the range [f, t). This number is greater or equal to | |
593 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
594 | * is needed and can not be allocated. | |
1dd308a7 | 595 | */ |
0db9d74e MA |
596 | static long region_chg(struct resv_map *resv, long f, long t, |
597 | long *out_regions_needed) | |
96822904 | 598 | { |
96822904 AW |
599 | long chg = 0; |
600 | ||
7b24d861 | 601 | spin_lock(&resv->lock); |
5e911373 | 602 | |
972a3da3 | 603 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 604 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 605 | out_regions_needed); |
5e911373 | 606 | |
0db9d74e MA |
607 | if (*out_regions_needed == 0) |
608 | *out_regions_needed = 1; | |
5e911373 | 609 | |
0db9d74e MA |
610 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
611 | return -ENOMEM; | |
5e911373 | 612 | |
0db9d74e | 613 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 614 | |
7b24d861 | 615 | spin_unlock(&resv->lock); |
96822904 AW |
616 | return chg; |
617 | } | |
618 | ||
5e911373 MK |
619 | /* |
620 | * Abort the in progress add operation. The adds_in_progress field | |
621 | * of the resv_map keeps track of the operations in progress between | |
622 | * calls to region_chg and region_add. Operations are sometimes | |
623 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
624 | * is called to decrement the adds_in_progress counter. regions_needed |
625 | * is the value returned by the region_chg call, it is used to decrement | |
626 | * the adds_in_progress counter. | |
5e911373 MK |
627 | * |
628 | * NOTE: The range arguments [f, t) are not needed or used in this | |
629 | * routine. They are kept to make reading the calling code easier as | |
630 | * arguments will match the associated region_chg call. | |
631 | */ | |
0db9d74e MA |
632 | static void region_abort(struct resv_map *resv, long f, long t, |
633 | long regions_needed) | |
5e911373 MK |
634 | { |
635 | spin_lock(&resv->lock); | |
636 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 637 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
638 | spin_unlock(&resv->lock); |
639 | } | |
640 | ||
1dd308a7 | 641 | /* |
feba16e2 MK |
642 | * Delete the specified range [f, t) from the reserve map. If the |
643 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
644 | * should be deleted. Locate the regions which intersect [f, t) | |
645 | * and either trim, delete or split the existing regions. | |
646 | * | |
647 | * Returns the number of huge pages deleted from the reserve map. | |
648 | * In the normal case, the return value is zero or more. In the | |
649 | * case where a region must be split, a new region descriptor must | |
650 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
651 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
652 | * a region and possibly return -ENOMEM. Callers specifying | |
653 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 654 | */ |
feba16e2 | 655 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 656 | { |
1406ec9b | 657 | struct list_head *head = &resv->regions; |
96822904 | 658 | struct file_region *rg, *trg; |
feba16e2 MK |
659 | struct file_region *nrg = NULL; |
660 | long del = 0; | |
96822904 | 661 | |
feba16e2 | 662 | retry: |
7b24d861 | 663 | spin_lock(&resv->lock); |
feba16e2 | 664 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
665 | /* |
666 | * Skip regions before the range to be deleted. file_region | |
667 | * ranges are normally of the form [from, to). However, there | |
668 | * may be a "placeholder" entry in the map which is of the form | |
669 | * (from, to) with from == to. Check for placeholder entries | |
670 | * at the beginning of the range to be deleted. | |
671 | */ | |
672 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 673 | continue; |
dbe409e4 | 674 | |
feba16e2 | 675 | if (rg->from >= t) |
96822904 | 676 | break; |
96822904 | 677 | |
feba16e2 MK |
678 | if (f > rg->from && t < rg->to) { /* Must split region */ |
679 | /* | |
680 | * Check for an entry in the cache before dropping | |
681 | * lock and attempting allocation. | |
682 | */ | |
683 | if (!nrg && | |
684 | resv->region_cache_count > resv->adds_in_progress) { | |
685 | nrg = list_first_entry(&resv->region_cache, | |
686 | struct file_region, | |
687 | link); | |
688 | list_del(&nrg->link); | |
689 | resv->region_cache_count--; | |
690 | } | |
96822904 | 691 | |
feba16e2 MK |
692 | if (!nrg) { |
693 | spin_unlock(&resv->lock); | |
694 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
695 | if (!nrg) | |
696 | return -ENOMEM; | |
697 | goto retry; | |
698 | } | |
699 | ||
700 | del += t - f; | |
79aa925b | 701 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 702 | resv, rg, t - f, false); |
feba16e2 MK |
703 | |
704 | /* New entry for end of split region */ | |
705 | nrg->from = t; | |
706 | nrg->to = rg->to; | |
075a61d0 MA |
707 | |
708 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
709 | ||
feba16e2 MK |
710 | INIT_LIST_HEAD(&nrg->link); |
711 | ||
712 | /* Original entry is trimmed */ | |
713 | rg->to = f; | |
714 | ||
715 | list_add(&nrg->link, &rg->link); | |
716 | nrg = NULL; | |
96822904 | 717 | break; |
feba16e2 MK |
718 | } |
719 | ||
720 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
721 | del += rg->to - rg->from; | |
075a61d0 | 722 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 723 | rg->to - rg->from, true); |
feba16e2 MK |
724 | list_del(&rg->link); |
725 | kfree(rg); | |
726 | continue; | |
727 | } | |
728 | ||
729 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 730 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 731 | t - rg->from, false); |
075a61d0 | 732 | |
79aa925b MK |
733 | del += t - rg->from; |
734 | rg->from = t; | |
735 | } else { /* Trim end of region */ | |
075a61d0 | 736 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 737 | rg->to - f, false); |
79aa925b MK |
738 | |
739 | del += rg->to - f; | |
740 | rg->to = f; | |
feba16e2 | 741 | } |
96822904 | 742 | } |
7b24d861 | 743 | |
7b24d861 | 744 | spin_unlock(&resv->lock); |
feba16e2 MK |
745 | kfree(nrg); |
746 | return del; | |
96822904 AW |
747 | } |
748 | ||
b5cec28d MK |
749 | /* |
750 | * A rare out of memory error was encountered which prevented removal of | |
751 | * the reserve map region for a page. The huge page itself was free'ed | |
752 | * and removed from the page cache. This routine will adjust the subpool | |
753 | * usage count, and the global reserve count if needed. By incrementing | |
754 | * these counts, the reserve map entry which could not be deleted will | |
755 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
756 | * counts. | |
757 | */ | |
72e2936c | 758 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
759 | { |
760 | struct hugepage_subpool *spool = subpool_inode(inode); | |
761 | long rsv_adjust; | |
da56388c | 762 | bool reserved = false; |
b5cec28d MK |
763 | |
764 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 765 | if (rsv_adjust > 0) { |
b5cec28d MK |
766 | struct hstate *h = hstate_inode(inode); |
767 | ||
da56388c ML |
768 | if (!hugetlb_acct_memory(h, 1)) |
769 | reserved = true; | |
770 | } else if (!rsv_adjust) { | |
771 | reserved = true; | |
b5cec28d | 772 | } |
da56388c ML |
773 | |
774 | if (!reserved) | |
775 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
776 | } |
777 | ||
1dd308a7 MK |
778 | /* |
779 | * Count and return the number of huge pages in the reserve map | |
780 | * that intersect with the range [f, t). | |
781 | */ | |
1406ec9b | 782 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 783 | { |
1406ec9b | 784 | struct list_head *head = &resv->regions; |
84afd99b AW |
785 | struct file_region *rg; |
786 | long chg = 0; | |
787 | ||
7b24d861 | 788 | spin_lock(&resv->lock); |
84afd99b AW |
789 | /* Locate each segment we overlap with, and count that overlap. */ |
790 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
791 | long seg_from; |
792 | long seg_to; | |
84afd99b AW |
793 | |
794 | if (rg->to <= f) | |
795 | continue; | |
796 | if (rg->from >= t) | |
797 | break; | |
798 | ||
799 | seg_from = max(rg->from, f); | |
800 | seg_to = min(rg->to, t); | |
801 | ||
802 | chg += seg_to - seg_from; | |
803 | } | |
7b24d861 | 804 | spin_unlock(&resv->lock); |
84afd99b AW |
805 | |
806 | return chg; | |
807 | } | |
808 | ||
e7c4b0bf AW |
809 | /* |
810 | * Convert the address within this vma to the page offset within | |
811 | * the mapping, in pagecache page units; huge pages here. | |
812 | */ | |
a5516438 AK |
813 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
814 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 815 | { |
a5516438 AK |
816 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
817 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
818 | } |
819 | ||
0fe6e20b NH |
820 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
821 | unsigned long address) | |
822 | { | |
823 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
824 | } | |
dee41079 | 825 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 826 | |
08fba699 MG |
827 | /* |
828 | * Return the size of the pages allocated when backing a VMA. In the majority | |
829 | * cases this will be same size as used by the page table entries. | |
830 | */ | |
831 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
832 | { | |
05ea8860 DW |
833 | if (vma->vm_ops && vma->vm_ops->pagesize) |
834 | return vma->vm_ops->pagesize(vma); | |
835 | return PAGE_SIZE; | |
08fba699 | 836 | } |
f340ca0f | 837 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 838 | |
3340289d MG |
839 | /* |
840 | * Return the page size being used by the MMU to back a VMA. In the majority | |
841 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
842 | * architectures where it differs, an architecture-specific 'strong' |
843 | * version of this symbol is required. | |
3340289d | 844 | */ |
09135cc5 | 845 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
846 | { |
847 | return vma_kernel_pagesize(vma); | |
848 | } | |
3340289d | 849 | |
84afd99b AW |
850 | /* |
851 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
852 | * bits of the reservation map pointer, which are always clear due to | |
853 | * alignment. | |
854 | */ | |
855 | #define HPAGE_RESV_OWNER (1UL << 0) | |
856 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 857 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 858 | |
a1e78772 MG |
859 | /* |
860 | * These helpers are used to track how many pages are reserved for | |
861 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
862 | * is guaranteed to have their future faults succeed. | |
863 | * | |
8d9bfb26 | 864 | * With the exception of hugetlb_dup_vma_private() which is called at fork(), |
a1e78772 MG |
865 | * the reserve counters are updated with the hugetlb_lock held. It is safe |
866 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
867 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
868 | * |
869 | * The private mapping reservation is represented in a subtly different | |
870 | * manner to a shared mapping. A shared mapping has a region map associated | |
871 | * with the underlying file, this region map represents the backing file | |
872 | * pages which have ever had a reservation assigned which this persists even | |
873 | * after the page is instantiated. A private mapping has a region map | |
874 | * associated with the original mmap which is attached to all VMAs which | |
875 | * reference it, this region map represents those offsets which have consumed | |
876 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 877 | */ |
e7c4b0bf AW |
878 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
879 | { | |
880 | return (unsigned long)vma->vm_private_data; | |
881 | } | |
882 | ||
883 | static void set_vma_private_data(struct vm_area_struct *vma, | |
884 | unsigned long value) | |
885 | { | |
886 | vma->vm_private_data = (void *)value; | |
887 | } | |
888 | ||
e9fe92ae MA |
889 | static void |
890 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
891 | struct hugetlb_cgroup *h_cg, | |
892 | struct hstate *h) | |
893 | { | |
894 | #ifdef CONFIG_CGROUP_HUGETLB | |
895 | if (!h_cg || !h) { | |
896 | resv_map->reservation_counter = NULL; | |
897 | resv_map->pages_per_hpage = 0; | |
898 | resv_map->css = NULL; | |
899 | } else { | |
900 | resv_map->reservation_counter = | |
901 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
902 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
903 | resv_map->css = &h_cg->css; | |
904 | } | |
905 | #endif | |
906 | } | |
907 | ||
9119a41e | 908 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
909 | { |
910 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
911 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
912 | ||
913 | if (!resv_map || !rg) { | |
914 | kfree(resv_map); | |
915 | kfree(rg); | |
84afd99b | 916 | return NULL; |
5e911373 | 917 | } |
84afd99b AW |
918 | |
919 | kref_init(&resv_map->refs); | |
7b24d861 | 920 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
921 | INIT_LIST_HEAD(&resv_map->regions); |
922 | ||
5e911373 | 923 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
924 | /* |
925 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
926 | * fields don't do cgroup accounting. On private mappings, these will be | |
927 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
928 | * reservations are to be un-charged from here. | |
929 | */ | |
930 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
931 | |
932 | INIT_LIST_HEAD(&resv_map->region_cache); | |
933 | list_add(&rg->link, &resv_map->region_cache); | |
934 | resv_map->region_cache_count = 1; | |
935 | ||
84afd99b AW |
936 | return resv_map; |
937 | } | |
938 | ||
9119a41e | 939 | void resv_map_release(struct kref *ref) |
84afd99b AW |
940 | { |
941 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
942 | struct list_head *head = &resv_map->region_cache; |
943 | struct file_region *rg, *trg; | |
84afd99b AW |
944 | |
945 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 946 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
947 | |
948 | /* ... and any entries left in the cache */ | |
949 | list_for_each_entry_safe(rg, trg, head, link) { | |
950 | list_del(&rg->link); | |
951 | kfree(rg); | |
952 | } | |
953 | ||
954 | VM_BUG_ON(resv_map->adds_in_progress); | |
955 | ||
84afd99b AW |
956 | kfree(resv_map); |
957 | } | |
958 | ||
4e35f483 JK |
959 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
960 | { | |
f27a5136 MK |
961 | /* |
962 | * At inode evict time, i_mapping may not point to the original | |
963 | * address space within the inode. This original address space | |
964 | * contains the pointer to the resv_map. So, always use the | |
965 | * address space embedded within the inode. | |
966 | * The VERY common case is inode->mapping == &inode->i_data but, | |
967 | * this may not be true for device special inodes. | |
968 | */ | |
969 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
970 | } |
971 | ||
84afd99b | 972 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 973 | { |
81d1b09c | 974 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
975 | if (vma->vm_flags & VM_MAYSHARE) { |
976 | struct address_space *mapping = vma->vm_file->f_mapping; | |
977 | struct inode *inode = mapping->host; | |
978 | ||
979 | return inode_resv_map(inode); | |
980 | ||
981 | } else { | |
84afd99b AW |
982 | return (struct resv_map *)(get_vma_private_data(vma) & |
983 | ~HPAGE_RESV_MASK); | |
4e35f483 | 984 | } |
a1e78772 MG |
985 | } |
986 | ||
84afd99b | 987 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 988 | { |
81d1b09c SL |
989 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
990 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 991 | |
84afd99b AW |
992 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
993 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
994 | } |
995 | ||
996 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
997 | { | |
81d1b09c SL |
998 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
999 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
1000 | |
1001 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
1002 | } |
1003 | ||
1004 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
1005 | { | |
81d1b09c | 1006 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
1007 | |
1008 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
1009 | } |
1010 | ||
8d9bfb26 | 1011 | void hugetlb_dup_vma_private(struct vm_area_struct *vma) |
a1e78772 | 1012 | { |
81d1b09c | 1013 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
8d9bfb26 MK |
1014 | /* |
1015 | * Clear vm_private_data | |
1016 | * - For MAP_PRIVATE mappings, this is the reserve map which does | |
1017 | * not apply to children. Faults generated by the children are | |
1018 | * not guaranteed to succeed, even if read-only. | |
1019 | * - For shared mappings this is a per-vma semaphore that may be | |
1020 | * allocated in a subsequent call to hugetlb_vm_op_open. | |
1021 | */ | |
1022 | vma->vm_private_data = (void *)0; | |
f83a275d | 1023 | if (!(vma->vm_flags & VM_MAYSHARE)) |
8d9bfb26 | 1024 | return; |
a1e78772 MG |
1025 | } |
1026 | ||
550a7d60 MA |
1027 | /* |
1028 | * Reset and decrement one ref on hugepage private reservation. | |
1029 | * Called with mm->mmap_sem writer semaphore held. | |
1030 | * This function should be only used by move_vma() and operate on | |
1031 | * same sized vma. It should never come here with last ref on the | |
1032 | * reservation. | |
1033 | */ | |
1034 | void clear_vma_resv_huge_pages(struct vm_area_struct *vma) | |
1035 | { | |
1036 | /* | |
1037 | * Clear the old hugetlb private page reservation. | |
1038 | * It has already been transferred to new_vma. | |
1039 | * | |
1040 | * During a mremap() operation of a hugetlb vma we call move_vma() | |
1041 | * which copies vma into new_vma and unmaps vma. After the copy | |
1042 | * operation both new_vma and vma share a reference to the resv_map | |
1043 | * struct, and at that point vma is about to be unmapped. We don't | |
1044 | * want to return the reservation to the pool at unmap of vma because | |
1045 | * the reservation still lives on in new_vma, so simply decrement the | |
1046 | * ref here and remove the resv_map reference from this vma. | |
1047 | */ | |
1048 | struct resv_map *reservations = vma_resv_map(vma); | |
1049 | ||
afe041c2 BQM |
1050 | if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1051 | resv_map_put_hugetlb_cgroup_uncharge_info(reservations); | |
550a7d60 | 1052 | kref_put(&reservations->refs, resv_map_release); |
afe041c2 | 1053 | } |
550a7d60 | 1054 | |
8d9bfb26 | 1055 | hugetlb_dup_vma_private(vma); |
550a7d60 MA |
1056 | } |
1057 | ||
a1e78772 | 1058 | /* Returns true if the VMA has associated reserve pages */ |
559ec2f8 | 1059 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1060 | { |
af0ed73e JK |
1061 | if (vma->vm_flags & VM_NORESERVE) { |
1062 | /* | |
1063 | * This address is already reserved by other process(chg == 0), | |
1064 | * so, we should decrement reserved count. Without decrementing, | |
1065 | * reserve count remains after releasing inode, because this | |
1066 | * allocated page will go into page cache and is regarded as | |
1067 | * coming from reserved pool in releasing step. Currently, we | |
1068 | * don't have any other solution to deal with this situation | |
1069 | * properly, so add work-around here. | |
1070 | */ | |
1071 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1072 | return true; |
af0ed73e | 1073 | else |
559ec2f8 | 1074 | return false; |
af0ed73e | 1075 | } |
a63884e9 JK |
1076 | |
1077 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1078 | if (vma->vm_flags & VM_MAYSHARE) { |
1079 | /* | |
1080 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1081 | * be a region map for all pages. The only situation where | |
1082 | * there is no region map is if a hole was punched via | |
7c8de358 | 1083 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1084 | * use. This situation is indicated if chg != 0. |
1085 | */ | |
1086 | if (chg) | |
1087 | return false; | |
1088 | else | |
1089 | return true; | |
1090 | } | |
a63884e9 JK |
1091 | |
1092 | /* | |
1093 | * Only the process that called mmap() has reserves for | |
1094 | * private mappings. | |
1095 | */ | |
67961f9d MK |
1096 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1097 | /* | |
1098 | * Like the shared case above, a hole punch or truncate | |
1099 | * could have been performed on the private mapping. | |
1100 | * Examine the value of chg to determine if reserves | |
1101 | * actually exist or were previously consumed. | |
1102 | * Very Subtle - The value of chg comes from a previous | |
1103 | * call to vma_needs_reserves(). The reserve map for | |
1104 | * private mappings has different (opposite) semantics | |
1105 | * than that of shared mappings. vma_needs_reserves() | |
1106 | * has already taken this difference in semantics into | |
1107 | * account. Therefore, the meaning of chg is the same | |
1108 | * as in the shared case above. Code could easily be | |
1109 | * combined, but keeping it separate draws attention to | |
1110 | * subtle differences. | |
1111 | */ | |
1112 | if (chg) | |
1113 | return false; | |
1114 | else | |
1115 | return true; | |
1116 | } | |
a63884e9 | 1117 | |
559ec2f8 | 1118 | return false; |
a1e78772 MG |
1119 | } |
1120 | ||
a5516438 | 1121 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1122 | { |
1123 | int nid = page_to_nid(page); | |
9487ca60 MK |
1124 | |
1125 | lockdep_assert_held(&hugetlb_lock); | |
b65a4eda MK |
1126 | VM_BUG_ON_PAGE(page_count(page), page); |
1127 | ||
0edaecfa | 1128 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1129 | h->free_huge_pages++; |
1130 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1131 | SetHPageFreed(page); |
1da177e4 LT |
1132 | } |
1133 | ||
94310cbc | 1134 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1135 | { |
1136 | struct page *page; | |
1a08ae36 | 1137 | bool pin = !!(current->flags & PF_MEMALLOC_PIN); |
bbe88753 | 1138 | |
9487ca60 | 1139 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 | 1140 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
6077c943 | 1141 | if (pin && !is_longterm_pinnable_page(page)) |
bbe88753 | 1142 | continue; |
bf50bab2 | 1143 | |
6664bfc8 WY |
1144 | if (PageHWPoison(page)) |
1145 | continue; | |
1146 | ||
1147 | list_move(&page->lru, &h->hugepage_activelist); | |
1148 | set_page_refcounted(page); | |
6c037149 | 1149 | ClearHPageFreed(page); |
6664bfc8 WY |
1150 | h->free_huge_pages--; |
1151 | h->free_huge_pages_node[nid]--; | |
1152 | return page; | |
bbe88753 JK |
1153 | } |
1154 | ||
6664bfc8 | 1155 | return NULL; |
bf50bab2 NH |
1156 | } |
1157 | ||
3e59fcb0 MH |
1158 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1159 | nodemask_t *nmask) | |
94310cbc | 1160 | { |
3e59fcb0 MH |
1161 | unsigned int cpuset_mems_cookie; |
1162 | struct zonelist *zonelist; | |
1163 | struct zone *zone; | |
1164 | struct zoneref *z; | |
98fa15f3 | 1165 | int node = NUMA_NO_NODE; |
94310cbc | 1166 | |
3e59fcb0 MH |
1167 | zonelist = node_zonelist(nid, gfp_mask); |
1168 | ||
1169 | retry_cpuset: | |
1170 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1171 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1172 | struct page *page; | |
1173 | ||
1174 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1175 | continue; | |
1176 | /* | |
1177 | * no need to ask again on the same node. Pool is node rather than | |
1178 | * zone aware | |
1179 | */ | |
1180 | if (zone_to_nid(zone) == node) | |
1181 | continue; | |
1182 | node = zone_to_nid(zone); | |
94310cbc | 1183 | |
94310cbc AK |
1184 | page = dequeue_huge_page_node_exact(h, node); |
1185 | if (page) | |
1186 | return page; | |
1187 | } | |
3e59fcb0 MH |
1188 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1189 | goto retry_cpuset; | |
1190 | ||
94310cbc AK |
1191 | return NULL; |
1192 | } | |
1193 | ||
8346d69d XH |
1194 | static unsigned long available_huge_pages(struct hstate *h) |
1195 | { | |
1196 | return h->free_huge_pages - h->resv_huge_pages; | |
1197 | } | |
1198 | ||
a5516438 AK |
1199 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1200 | struct vm_area_struct *vma, | |
af0ed73e JK |
1201 | unsigned long address, int avoid_reserve, |
1202 | long chg) | |
1da177e4 | 1203 | { |
cfcaa66f | 1204 | struct page *page = NULL; |
480eccf9 | 1205 | struct mempolicy *mpol; |
04ec6264 | 1206 | gfp_t gfp_mask; |
3e59fcb0 | 1207 | nodemask_t *nodemask; |
04ec6264 | 1208 | int nid; |
1da177e4 | 1209 | |
a1e78772 MG |
1210 | /* |
1211 | * A child process with MAP_PRIVATE mappings created by their parent | |
1212 | * have no page reserves. This check ensures that reservations are | |
1213 | * not "stolen". The child may still get SIGKILLed | |
1214 | */ | |
8346d69d | 1215 | if (!vma_has_reserves(vma, chg) && !available_huge_pages(h)) |
c0ff7453 | 1216 | goto err; |
a1e78772 | 1217 | |
04f2cbe3 | 1218 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
8346d69d | 1219 | if (avoid_reserve && !available_huge_pages(h)) |
6eab04a8 | 1220 | goto err; |
04f2cbe3 | 1221 | |
04ec6264 VB |
1222 | gfp_mask = htlb_alloc_mask(h); |
1223 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
1224 | |
1225 | if (mpol_is_preferred_many(mpol)) { | |
1226 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1227 | ||
1228 | /* Fallback to all nodes if page==NULL */ | |
1229 | nodemask = NULL; | |
1230 | } | |
1231 | ||
1232 | if (!page) | |
1233 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1234 | ||
3e59fcb0 | 1235 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { |
d6995da3 | 1236 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1237 | h->resv_huge_pages--; |
1da177e4 | 1238 | } |
cc9a6c87 | 1239 | |
52cd3b07 | 1240 | mpol_cond_put(mpol); |
1da177e4 | 1241 | return page; |
cc9a6c87 MG |
1242 | |
1243 | err: | |
cc9a6c87 | 1244 | return NULL; |
1da177e4 LT |
1245 | } |
1246 | ||
1cac6f2c LC |
1247 | /* |
1248 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1249 | * We may have allocated or freed a huge page based on a different | |
1250 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1251 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1252 | * node for alloc or free. | |
1253 | */ | |
1254 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1255 | { | |
0edaf86c | 1256 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1257 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1258 | ||
1259 | return nid; | |
1260 | } | |
1261 | ||
1262 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1263 | { | |
1264 | if (!node_isset(nid, *nodes_allowed)) | |
1265 | nid = next_node_allowed(nid, nodes_allowed); | |
1266 | return nid; | |
1267 | } | |
1268 | ||
1269 | /* | |
1270 | * returns the previously saved node ["this node"] from which to | |
1271 | * allocate a persistent huge page for the pool and advance the | |
1272 | * next node from which to allocate, handling wrap at end of node | |
1273 | * mask. | |
1274 | */ | |
1275 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1276 | nodemask_t *nodes_allowed) | |
1277 | { | |
1278 | int nid; | |
1279 | ||
1280 | VM_BUG_ON(!nodes_allowed); | |
1281 | ||
1282 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1283 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1284 | ||
1285 | return nid; | |
1286 | } | |
1287 | ||
1288 | /* | |
10c6ec49 | 1289 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1290 | * node ["this node"] from which to free a huge page. Advance the |
1291 | * next node id whether or not we find a free huge page to free so | |
1292 | * that the next attempt to free addresses the next node. | |
1293 | */ | |
1294 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1295 | { | |
1296 | int nid; | |
1297 | ||
1298 | VM_BUG_ON(!nodes_allowed); | |
1299 | ||
1300 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1301 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1302 | ||
1303 | return nid; | |
1304 | } | |
1305 | ||
1306 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1307 | for (nr_nodes = nodes_weight(*mask); \ | |
1308 | nr_nodes > 0 && \ | |
1309 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1310 | nr_nodes--) | |
1311 | ||
1312 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1313 | for (nr_nodes = nodes_weight(*mask); \ | |
1314 | nr_nodes > 0 && \ | |
1315 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1316 | nr_nodes--) | |
1317 | ||
8531fc6f | 1318 | /* used to demote non-gigantic_huge pages as well */ |
34d9e35b MK |
1319 | static void __destroy_compound_gigantic_page(struct page *page, |
1320 | unsigned int order, bool demote) | |
944d9fec LC |
1321 | { |
1322 | int i; | |
1323 | int nr_pages = 1 << order; | |
14455eab | 1324 | struct page *p; |
944d9fec | 1325 | |
c8cc708a | 1326 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1327 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1328 | |
14455eab CL |
1329 | for (i = 1; i < nr_pages; i++) { |
1330 | p = nth_page(page, i); | |
a01f4390 | 1331 | p->mapping = NULL; |
1d798ca3 | 1332 | clear_compound_head(p); |
34d9e35b MK |
1333 | if (!demote) |
1334 | set_page_refcounted(p); | |
944d9fec LC |
1335 | } |
1336 | ||
1337 | set_compound_order(page, 0); | |
5232c63f | 1338 | #ifdef CONFIG_64BIT |
ba9c1201 | 1339 | page[1].compound_nr = 0; |
5232c63f | 1340 | #endif |
944d9fec LC |
1341 | __ClearPageHead(page); |
1342 | } | |
1343 | ||
8531fc6f MK |
1344 | static void destroy_compound_hugetlb_page_for_demote(struct page *page, |
1345 | unsigned int order) | |
1346 | { | |
1347 | __destroy_compound_gigantic_page(page, order, true); | |
1348 | } | |
1349 | ||
1350 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE | |
34d9e35b MK |
1351 | static void destroy_compound_gigantic_page(struct page *page, |
1352 | unsigned int order) | |
1353 | { | |
1354 | __destroy_compound_gigantic_page(page, order, false); | |
1355 | } | |
1356 | ||
d00181b9 | 1357 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1358 | { |
cf11e85f RG |
1359 | /* |
1360 | * If the page isn't allocated using the cma allocator, | |
1361 | * cma_release() returns false. | |
1362 | */ | |
dbda8fea BS |
1363 | #ifdef CONFIG_CMA |
1364 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1365 | return; |
dbda8fea | 1366 | #endif |
cf11e85f | 1367 | |
944d9fec LC |
1368 | free_contig_range(page_to_pfn(page), 1 << order); |
1369 | } | |
1370 | ||
4eb0716e | 1371 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1372 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1373 | int nid, nodemask_t *nodemask) | |
944d9fec | 1374 | { |
04adbc3f | 1375 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1376 | if (nid == NUMA_NO_NODE) |
1377 | nid = numa_mem_id(); | |
944d9fec | 1378 | |
dbda8fea BS |
1379 | #ifdef CONFIG_CMA |
1380 | { | |
cf11e85f RG |
1381 | struct page *page; |
1382 | int node; | |
1383 | ||
953f064a LX |
1384 | if (hugetlb_cma[nid]) { |
1385 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1386 | huge_page_order(h), true); | |
cf11e85f RG |
1387 | if (page) |
1388 | return page; | |
1389 | } | |
953f064a LX |
1390 | |
1391 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1392 | for_each_node_mask(node, *nodemask) { | |
1393 | if (node == nid || !hugetlb_cma[node]) | |
1394 | continue; | |
1395 | ||
1396 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1397 | huge_page_order(h), true); | |
1398 | if (page) | |
1399 | return page; | |
1400 | } | |
1401 | } | |
cf11e85f | 1402 | } |
dbda8fea | 1403 | #endif |
cf11e85f | 1404 | |
5e27a2df | 1405 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1406 | } |
1407 | ||
4eb0716e AG |
1408 | #else /* !CONFIG_CONTIG_ALLOC */ |
1409 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1410 | int nid, nodemask_t *nodemask) | |
1411 | { | |
1412 | return NULL; | |
1413 | } | |
1414 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1415 | |
e1073d1e | 1416 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1417 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1418 | int nid, nodemask_t *nodemask) |
1419 | { | |
1420 | return NULL; | |
1421 | } | |
d00181b9 | 1422 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1423 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1424 | unsigned int order) { } |
944d9fec LC |
1425 | #endif |
1426 | ||
6eb4e88a MK |
1427 | /* |
1428 | * Remove hugetlb page from lists, and update dtor so that page appears | |
34d9e35b MK |
1429 | * as just a compound page. |
1430 | * | |
1431 | * A reference is held on the page, except in the case of demote. | |
6eb4e88a MK |
1432 | * |
1433 | * Must be called with hugetlb lock held. | |
1434 | */ | |
34d9e35b MK |
1435 | static void __remove_hugetlb_page(struct hstate *h, struct page *page, |
1436 | bool adjust_surplus, | |
1437 | bool demote) | |
6eb4e88a MK |
1438 | { |
1439 | int nid = page_to_nid(page); | |
1440 | ||
1441 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1442 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1443 | ||
9487ca60 | 1444 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1445 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1446 | return; | |
1447 | ||
1448 | list_del(&page->lru); | |
1449 | ||
1450 | if (HPageFreed(page)) { | |
1451 | h->free_huge_pages--; | |
1452 | h->free_huge_pages_node[nid]--; | |
1453 | } | |
1454 | if (adjust_surplus) { | |
1455 | h->surplus_huge_pages--; | |
1456 | h->surplus_huge_pages_node[nid]--; | |
1457 | } | |
1458 | ||
e32d20c0 MK |
1459 | /* |
1460 | * Very subtle | |
1461 | * | |
1462 | * For non-gigantic pages set the destructor to the normal compound | |
1463 | * page dtor. This is needed in case someone takes an additional | |
1464 | * temporary ref to the page, and freeing is delayed until they drop | |
1465 | * their reference. | |
1466 | * | |
1467 | * For gigantic pages set the destructor to the null dtor. This | |
1468 | * destructor will never be called. Before freeing the gigantic | |
1469 | * page destroy_compound_gigantic_page will turn the compound page | |
1470 | * into a simple group of pages. After this the destructor does not | |
1471 | * apply. | |
1472 | * | |
1473 | * This handles the case where more than one ref is held when and | |
1474 | * after update_and_free_page is called. | |
34d9e35b MK |
1475 | * |
1476 | * In the case of demote we do not ref count the page as it will soon | |
1477 | * be turned into a page of smaller size. | |
e32d20c0 | 1478 | */ |
34d9e35b MK |
1479 | if (!demote) |
1480 | set_page_refcounted(page); | |
e32d20c0 MK |
1481 | if (hstate_is_gigantic(h)) |
1482 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1483 | else | |
1484 | set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); | |
6eb4e88a MK |
1485 | |
1486 | h->nr_huge_pages--; | |
1487 | h->nr_huge_pages_node[nid]--; | |
1488 | } | |
1489 | ||
34d9e35b MK |
1490 | static void remove_hugetlb_page(struct hstate *h, struct page *page, |
1491 | bool adjust_surplus) | |
1492 | { | |
1493 | __remove_hugetlb_page(h, page, adjust_surplus, false); | |
1494 | } | |
1495 | ||
8531fc6f MK |
1496 | static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, |
1497 | bool adjust_surplus) | |
1498 | { | |
1499 | __remove_hugetlb_page(h, page, adjust_surplus, true); | |
1500 | } | |
1501 | ||
ad2fa371 MS |
1502 | static void add_hugetlb_page(struct hstate *h, struct page *page, |
1503 | bool adjust_surplus) | |
1504 | { | |
1505 | int zeroed; | |
1506 | int nid = page_to_nid(page); | |
1507 | ||
1508 | VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); | |
1509 | ||
1510 | lockdep_assert_held(&hugetlb_lock); | |
1511 | ||
1512 | INIT_LIST_HEAD(&page->lru); | |
1513 | h->nr_huge_pages++; | |
1514 | h->nr_huge_pages_node[nid]++; | |
1515 | ||
1516 | if (adjust_surplus) { | |
1517 | h->surplus_huge_pages++; | |
1518 | h->surplus_huge_pages_node[nid]++; | |
1519 | } | |
1520 | ||
1521 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); | |
1522 | set_page_private(page, 0); | |
a9e1eab2 ML |
1523 | /* |
1524 | * We have to set HPageVmemmapOptimized again as above | |
1525 | * set_page_private(page, 0) cleared it. | |
1526 | */ | |
ad2fa371 MS |
1527 | SetHPageVmemmapOptimized(page); |
1528 | ||
1529 | /* | |
b65a4eda MK |
1530 | * This page is about to be managed by the hugetlb allocator and |
1531 | * should have no users. Drop our reference, and check for others | |
1532 | * just in case. | |
ad2fa371 MS |
1533 | */ |
1534 | zeroed = put_page_testzero(page); | |
b65a4eda MK |
1535 | if (!zeroed) |
1536 | /* | |
1537 | * It is VERY unlikely soneone else has taken a ref on | |
1538 | * the page. In this case, we simply return as the | |
1539 | * hugetlb destructor (free_huge_page) will be called | |
1540 | * when this other ref is dropped. | |
1541 | */ | |
1542 | return; | |
1543 | ||
ad2fa371 MS |
1544 | arch_clear_hugepage_flags(page); |
1545 | enqueue_huge_page(h, page); | |
1546 | } | |
1547 | ||
b65d4adb | 1548 | static void __update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1549 | { |
1550 | int i; | |
14455eab | 1551 | struct page *subpage; |
a5516438 | 1552 | |
4eb0716e | 1553 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1554 | return; |
18229df5 | 1555 | |
161df60e NH |
1556 | /* |
1557 | * If we don't know which subpages are hwpoisoned, we can't free | |
1558 | * the hugepage, so it's leaked intentionally. | |
1559 | */ | |
1560 | if (HPageRawHwpUnreliable(page)) | |
1561 | return; | |
1562 | ||
6213834c | 1563 | if (hugetlb_vmemmap_restore(h, page)) { |
ad2fa371 MS |
1564 | spin_lock_irq(&hugetlb_lock); |
1565 | /* | |
1566 | * If we cannot allocate vmemmap pages, just refuse to free the | |
1567 | * page and put the page back on the hugetlb free list and treat | |
1568 | * as a surplus page. | |
1569 | */ | |
1570 | add_hugetlb_page(h, page, true); | |
1571 | spin_unlock_irq(&hugetlb_lock); | |
1572 | return; | |
1573 | } | |
1574 | ||
161df60e NH |
1575 | /* |
1576 | * Move PageHWPoison flag from head page to the raw error pages, | |
1577 | * which makes any healthy subpages reusable. | |
1578 | */ | |
1579 | if (unlikely(PageHWPoison(page))) | |
1580 | hugetlb_clear_page_hwpoison(page); | |
1581 | ||
14455eab CL |
1582 | for (i = 0; i < pages_per_huge_page(h); i++) { |
1583 | subpage = nth_page(page, i); | |
dbfee5ae | 1584 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | |
32f84528 | 1585 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1586 | 1 << PG_active | 1 << PG_private | |
1587 | 1 << PG_writeback); | |
6af2acb6 | 1588 | } |
a01f4390 MK |
1589 | |
1590 | /* | |
1591 | * Non-gigantic pages demoted from CMA allocated gigantic pages | |
1592 | * need to be given back to CMA in free_gigantic_page. | |
1593 | */ | |
1594 | if (hstate_is_gigantic(h) || | |
1595 | hugetlb_cma_page(page, huge_page_order(h))) { | |
944d9fec LC |
1596 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1597 | free_gigantic_page(page, huge_page_order(h)); | |
1598 | } else { | |
944d9fec LC |
1599 | __free_pages(page, huge_page_order(h)); |
1600 | } | |
6af2acb6 AL |
1601 | } |
1602 | ||
b65d4adb MS |
1603 | /* |
1604 | * As update_and_free_page() can be called under any context, so we cannot | |
1605 | * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the | |
1606 | * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate | |
1607 | * the vmemmap pages. | |
1608 | * | |
1609 | * free_hpage_workfn() locklessly retrieves the linked list of pages to be | |
1610 | * freed and frees them one-by-one. As the page->mapping pointer is going | |
1611 | * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node | |
1612 | * structure of a lockless linked list of huge pages to be freed. | |
1613 | */ | |
1614 | static LLIST_HEAD(hpage_freelist); | |
1615 | ||
1616 | static void free_hpage_workfn(struct work_struct *work) | |
1617 | { | |
1618 | struct llist_node *node; | |
1619 | ||
1620 | node = llist_del_all(&hpage_freelist); | |
1621 | ||
1622 | while (node) { | |
1623 | struct page *page; | |
1624 | struct hstate *h; | |
1625 | ||
1626 | page = container_of((struct address_space **)node, | |
1627 | struct page, mapping); | |
1628 | node = node->next; | |
1629 | page->mapping = NULL; | |
1630 | /* | |
1631 | * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() | |
1632 | * is going to trigger because a previous call to | |
1633 | * remove_hugetlb_page() will set_compound_page_dtor(page, | |
1634 | * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. | |
1635 | */ | |
1636 | h = size_to_hstate(page_size(page)); | |
1637 | ||
1638 | __update_and_free_page(h, page); | |
1639 | ||
1640 | cond_resched(); | |
1641 | } | |
1642 | } | |
1643 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1644 | ||
1645 | static inline void flush_free_hpage_work(struct hstate *h) | |
1646 | { | |
6213834c | 1647 | if (hugetlb_vmemmap_optimizable(h)) |
b65d4adb MS |
1648 | flush_work(&free_hpage_work); |
1649 | } | |
1650 | ||
1651 | static void update_and_free_page(struct hstate *h, struct page *page, | |
1652 | bool atomic) | |
1653 | { | |
ad2fa371 | 1654 | if (!HPageVmemmapOptimized(page) || !atomic) { |
b65d4adb MS |
1655 | __update_and_free_page(h, page); |
1656 | return; | |
1657 | } | |
1658 | ||
1659 | /* | |
1660 | * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. | |
1661 | * | |
1662 | * Only call schedule_work() if hpage_freelist is previously | |
1663 | * empty. Otherwise, schedule_work() had been called but the workfn | |
1664 | * hasn't retrieved the list yet. | |
1665 | */ | |
1666 | if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) | |
1667 | schedule_work(&free_hpage_work); | |
1668 | } | |
1669 | ||
10c6ec49 MK |
1670 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1671 | { | |
1672 | struct page *page, *t_page; | |
1673 | ||
1674 | list_for_each_entry_safe(page, t_page, list, lru) { | |
b65d4adb | 1675 | update_and_free_page(h, page, false); |
10c6ec49 MK |
1676 | cond_resched(); |
1677 | } | |
1678 | } | |
1679 | ||
e5ff2159 AK |
1680 | struct hstate *size_to_hstate(unsigned long size) |
1681 | { | |
1682 | struct hstate *h; | |
1683 | ||
1684 | for_each_hstate(h) { | |
1685 | if (huge_page_size(h) == size) | |
1686 | return h; | |
1687 | } | |
1688 | return NULL; | |
1689 | } | |
1690 | ||
db71ef79 | 1691 | void free_huge_page(struct page *page) |
27a85ef1 | 1692 | { |
a5516438 AK |
1693 | /* |
1694 | * Can't pass hstate in here because it is called from the | |
1695 | * compound page destructor. | |
1696 | */ | |
e5ff2159 | 1697 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1698 | int nid = page_to_nid(page); |
d6995da3 | 1699 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1700 | bool restore_reserve; |
db71ef79 | 1701 | unsigned long flags; |
27a85ef1 | 1702 | |
b4330afb MK |
1703 | VM_BUG_ON_PAGE(page_count(page), page); |
1704 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1705 | |
d6995da3 | 1706 | hugetlb_set_page_subpool(page, NULL); |
78fbe906 DH |
1707 | if (PageAnon(page)) |
1708 | __ClearPageAnonExclusive(page); | |
8ace22bc | 1709 | page->mapping = NULL; |
d6995da3 MK |
1710 | restore_reserve = HPageRestoreReserve(page); |
1711 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1712 | |
1c5ecae3 | 1713 | /* |
d6995da3 | 1714 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1715 | * reservation. If the page was associated with a subpool, there |
1716 | * would have been a page reserved in the subpool before allocation | |
1717 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1718 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1719 | * remove the reserved page from the subpool. |
1c5ecae3 | 1720 | */ |
0919e1b6 MK |
1721 | if (!restore_reserve) { |
1722 | /* | |
1723 | * A return code of zero implies that the subpool will be | |
1724 | * under its minimum size if the reservation is not restored | |
1725 | * after page is free. Therefore, force restore_reserve | |
1726 | * operation. | |
1727 | */ | |
1728 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1729 | restore_reserve = true; | |
1730 | } | |
1c5ecae3 | 1731 | |
db71ef79 | 1732 | spin_lock_irqsave(&hugetlb_lock, flags); |
8f251a3d | 1733 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1734 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1735 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1736 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1737 | pages_per_huge_page(h), page); | |
07443a85 JK |
1738 | if (restore_reserve) |
1739 | h->resv_huge_pages++; | |
1740 | ||
9157c311 | 1741 | if (HPageTemporary(page)) { |
6eb4e88a | 1742 | remove_hugetlb_page(h, page, false); |
db71ef79 | 1743 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1744 | update_and_free_page(h, page, true); |
ab5ac90a | 1745 | } else if (h->surplus_huge_pages_node[nid]) { |
0edaecfa | 1746 | /* remove the page from active list */ |
6eb4e88a | 1747 | remove_hugetlb_page(h, page, true); |
db71ef79 | 1748 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1749 | update_and_free_page(h, page, true); |
7893d1d5 | 1750 | } else { |
5d3a551c | 1751 | arch_clear_hugepage_flags(page); |
a5516438 | 1752 | enqueue_huge_page(h, page); |
db71ef79 | 1753 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1754 | } |
c77c0a8a WL |
1755 | } |
1756 | ||
d3d99fcc OS |
1757 | /* |
1758 | * Must be called with the hugetlb lock held | |
1759 | */ | |
1760 | static void __prep_account_new_huge_page(struct hstate *h, int nid) | |
1761 | { | |
1762 | lockdep_assert_held(&hugetlb_lock); | |
1763 | h->nr_huge_pages++; | |
1764 | h->nr_huge_pages_node[nid]++; | |
1765 | } | |
1766 | ||
f41f2ed4 | 1767 | static void __prep_new_huge_page(struct hstate *h, struct page *page) |
b7ba30c6 | 1768 | { |
6213834c | 1769 | hugetlb_vmemmap_optimize(h, page); |
0edaecfa | 1770 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1771 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1772 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1773 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1774 | set_hugetlb_cgroup_rsvd(page, NULL); |
d3d99fcc OS |
1775 | } |
1776 | ||
1777 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1778 | { | |
f41f2ed4 | 1779 | __prep_new_huge_page(h, page); |
db71ef79 | 1780 | spin_lock_irq(&hugetlb_lock); |
d3d99fcc | 1781 | __prep_account_new_huge_page(h, nid); |
db71ef79 | 1782 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1783 | } |
1784 | ||
34d9e35b MK |
1785 | static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, |
1786 | bool demote) | |
20a0307c | 1787 | { |
7118fc29 | 1788 | int i, j; |
20a0307c | 1789 | int nr_pages = 1 << order; |
14455eab | 1790 | struct page *p; |
20a0307c WF |
1791 | |
1792 | /* we rely on prep_new_huge_page to set the destructor */ | |
1793 | set_compound_order(page, order); | |
de09d31d | 1794 | __SetPageHead(page); |
2b21624f | 1795 | for (i = 0; i < nr_pages; i++) { |
14455eab CL |
1796 | p = nth_page(page, i); |
1797 | ||
ef5a22be AA |
1798 | /* |
1799 | * For gigantic hugepages allocated through bootmem at | |
1800 | * boot, it's safer to be consistent with the not-gigantic | |
1801 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1802 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1803 | * pages may get the reference counting wrong if they see |
1804 | * PG_reserved set on a tail page (despite the head page not | |
1805 | * having PG_reserved set). Enforcing this consistency between | |
1806 | * head and tail pages allows drivers to optimize away a check | |
1807 | * on the head page when they need know if put_page() is needed | |
1808 | * after get_user_pages(). | |
1809 | */ | |
1810 | __ClearPageReserved(p); | |
7118fc29 MK |
1811 | /* |
1812 | * Subtle and very unlikely | |
1813 | * | |
1814 | * Gigantic 'page allocators' such as memblock or cma will | |
1815 | * return a set of pages with each page ref counted. We need | |
1816 | * to turn this set of pages into a compound page with tail | |
1817 | * page ref counts set to zero. Code such as speculative page | |
1818 | * cache adding could take a ref on a 'to be' tail page. | |
1819 | * We need to respect any increased ref count, and only set | |
1820 | * the ref count to zero if count is currently 1. If count | |
416d85ed MK |
1821 | * is not 1, we return an error. An error return indicates |
1822 | * the set of pages can not be converted to a gigantic page. | |
1823 | * The caller who allocated the pages should then discard the | |
1824 | * pages using the appropriate free interface. | |
34d9e35b MK |
1825 | * |
1826 | * In the case of demote, the ref count will be zero. | |
7118fc29 | 1827 | */ |
34d9e35b MK |
1828 | if (!demote) { |
1829 | if (!page_ref_freeze(p, 1)) { | |
1830 | pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); | |
1831 | goto out_error; | |
1832 | } | |
1833 | } else { | |
1834 | VM_BUG_ON_PAGE(page_count(p), p); | |
7118fc29 | 1835 | } |
2b21624f MK |
1836 | if (i != 0) |
1837 | set_compound_head(p, page); | |
20a0307c | 1838 | } |
b4330afb | 1839 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1840 | atomic_set(compound_pincount_ptr(page), 0); |
7118fc29 MK |
1841 | return true; |
1842 | ||
1843 | out_error: | |
2b21624f MK |
1844 | /* undo page modifications made above */ |
1845 | for (j = 0; j < i; j++) { | |
14455eab | 1846 | p = nth_page(page, j); |
2b21624f MK |
1847 | if (j != 0) |
1848 | clear_compound_head(p); | |
7118fc29 MK |
1849 | set_page_refcounted(p); |
1850 | } | |
1851 | /* need to clear PG_reserved on remaining tail pages */ | |
14455eab CL |
1852 | for (; j < nr_pages; j++) { |
1853 | p = nth_page(page, j); | |
7118fc29 | 1854 | __ClearPageReserved(p); |
14455eab | 1855 | } |
7118fc29 | 1856 | set_compound_order(page, 0); |
5232c63f | 1857 | #ifdef CONFIG_64BIT |
7118fc29 | 1858 | page[1].compound_nr = 0; |
5232c63f | 1859 | #endif |
7118fc29 MK |
1860 | __ClearPageHead(page); |
1861 | return false; | |
20a0307c WF |
1862 | } |
1863 | ||
34d9e35b MK |
1864 | static bool prep_compound_gigantic_page(struct page *page, unsigned int order) |
1865 | { | |
1866 | return __prep_compound_gigantic_page(page, order, false); | |
1867 | } | |
1868 | ||
8531fc6f MK |
1869 | static bool prep_compound_gigantic_page_for_demote(struct page *page, |
1870 | unsigned int order) | |
1871 | { | |
1872 | return __prep_compound_gigantic_page(page, order, true); | |
1873 | } | |
1874 | ||
7795912c AM |
1875 | /* |
1876 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1877 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1878 | * details. | |
1879 | */ | |
20a0307c WF |
1880 | int PageHuge(struct page *page) |
1881 | { | |
20a0307c WF |
1882 | if (!PageCompound(page)) |
1883 | return 0; | |
1884 | ||
1885 | page = compound_head(page); | |
f1e61557 | 1886 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1887 | } |
43131e14 NH |
1888 | EXPORT_SYMBOL_GPL(PageHuge); |
1889 | ||
27c73ae7 AA |
1890 | /* |
1891 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1892 | * normal or transparent huge pages. | |
1893 | */ | |
1894 | int PageHeadHuge(struct page *page_head) | |
1895 | { | |
27c73ae7 AA |
1896 | if (!PageHead(page_head)) |
1897 | return 0; | |
1898 | ||
d4af73e3 | 1899 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1900 | } |
4e936ecc | 1901 | EXPORT_SYMBOL_GPL(PageHeadHuge); |
27c73ae7 | 1902 | |
c0d0381a MK |
1903 | /* |
1904 | * Find and lock address space (mapping) in write mode. | |
1905 | * | |
336bf30e MK |
1906 | * Upon entry, the page is locked which means that page_mapping() is |
1907 | * stable. Due to locking order, we can only trylock_write. If we can | |
1908 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1909 | */ |
1910 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1911 | { | |
336bf30e | 1912 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1913 | |
c0d0381a MK |
1914 | if (!mapping) |
1915 | return mapping; | |
1916 | ||
c0d0381a MK |
1917 | if (i_mmap_trylock_write(mapping)) |
1918 | return mapping; | |
1919 | ||
336bf30e | 1920 | return NULL; |
c0d0381a MK |
1921 | } |
1922 | ||
fe19bd3d | 1923 | pgoff_t hugetlb_basepage_index(struct page *page) |
13d60f4b ZY |
1924 | { |
1925 | struct page *page_head = compound_head(page); | |
1926 | pgoff_t index = page_index(page_head); | |
1927 | unsigned long compound_idx; | |
1928 | ||
13d60f4b ZY |
1929 | if (compound_order(page_head) >= MAX_ORDER) |
1930 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1931 | else | |
1932 | compound_idx = page - page_head; | |
1933 | ||
1934 | return (index << compound_order(page_head)) + compound_idx; | |
1935 | } | |
1936 | ||
0c397dae | 1937 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1938 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1939 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1940 | { |
af0fb9df | 1941 | int order = huge_page_order(h); |
1da177e4 | 1942 | struct page *page; |
f60858f9 | 1943 | bool alloc_try_hard = true; |
2b21624f | 1944 | bool retry = true; |
f96efd58 | 1945 | |
f60858f9 MK |
1946 | /* |
1947 | * By default we always try hard to allocate the page with | |
1948 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1949 | * a loop (to adjust global huge page counts) and previous allocation | |
1950 | * failed, do not continue to try hard on the same node. Use the | |
1951 | * node_alloc_noretry bitmap to manage this state information. | |
1952 | */ | |
1953 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1954 | alloc_try_hard = false; | |
1955 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1956 | if (alloc_try_hard) | |
1957 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1958 | if (nid == NUMA_NO_NODE) |
1959 | nid = numa_mem_id(); | |
2b21624f | 1960 | retry: |
84172f4b | 1961 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
2b21624f MK |
1962 | |
1963 | /* Freeze head page */ | |
1964 | if (page && !page_ref_freeze(page, 1)) { | |
1965 | __free_pages(page, order); | |
1966 | if (retry) { /* retry once */ | |
1967 | retry = false; | |
1968 | goto retry; | |
1969 | } | |
1970 | /* WOW! twice in a row. */ | |
1971 | pr_warn("HugeTLB head page unexpected inflated ref count\n"); | |
1972 | page = NULL; | |
1973 | } | |
1974 | ||
af0fb9df MH |
1975 | if (page) |
1976 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1977 | else | |
1978 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1979 | |
f60858f9 MK |
1980 | /* |
1981 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1982 | * indicates an overall state change. Clear bit so that we resume | |
1983 | * normal 'try hard' allocations. | |
1984 | */ | |
1985 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1986 | node_clear(nid, *node_alloc_noretry); | |
1987 | ||
1988 | /* | |
1989 | * If we tried hard to get a page but failed, set bit so that | |
1990 | * subsequent attempts will not try as hard until there is an | |
1991 | * overall state change. | |
1992 | */ | |
1993 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1994 | node_set(nid, *node_alloc_noretry); | |
1995 | ||
63b4613c NA |
1996 | return page; |
1997 | } | |
1998 | ||
0c397dae MH |
1999 | /* |
2000 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
2001 | * should use this function to get new hugetlb pages | |
2b21624f MK |
2002 | * |
2003 | * Note that returned page is 'frozen': ref count of head page and all tail | |
2004 | * pages is zero. | |
0c397dae MH |
2005 | */ |
2006 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
2007 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
2008 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
2009 | { |
2010 | struct page *page; | |
7118fc29 | 2011 | bool retry = false; |
0c397dae | 2012 | |
7118fc29 | 2013 | retry: |
0c397dae MH |
2014 | if (hstate_is_gigantic(h)) |
2015 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
2016 | else | |
2017 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 2018 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
2019 | if (!page) |
2020 | return NULL; | |
2021 | ||
7118fc29 MK |
2022 | if (hstate_is_gigantic(h)) { |
2023 | if (!prep_compound_gigantic_page(page, huge_page_order(h))) { | |
2024 | /* | |
2025 | * Rare failure to convert pages to compound page. | |
2026 | * Free pages and try again - ONCE! | |
2027 | */ | |
2028 | free_gigantic_page(page, huge_page_order(h)); | |
2029 | if (!retry) { | |
2030 | retry = true; | |
2031 | goto retry; | |
2032 | } | |
7118fc29 MK |
2033 | return NULL; |
2034 | } | |
2035 | } | |
0c397dae MH |
2036 | prep_new_huge_page(h, page, page_to_nid(page)); |
2037 | ||
2038 | return page; | |
2039 | } | |
2040 | ||
af0fb9df MH |
2041 | /* |
2042 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
2043 | * manner. | |
2044 | */ | |
f60858f9 MK |
2045 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
2046 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
2047 | { |
2048 | struct page *page; | |
2049 | int nr_nodes, node; | |
af0fb9df | 2050 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
2051 | |
2052 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
2053 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
2054 | node_alloc_noretry); | |
af0fb9df | 2055 | if (page) |
b2261026 | 2056 | break; |
b2261026 JK |
2057 | } |
2058 | ||
af0fb9df MH |
2059 | if (!page) |
2060 | return 0; | |
b2261026 | 2061 | |
2b21624f | 2062 | free_huge_page(page); /* free it into the hugepage allocator */ |
af0fb9df MH |
2063 | |
2064 | return 1; | |
b2261026 JK |
2065 | } |
2066 | ||
e8c5c824 | 2067 | /* |
10c6ec49 MK |
2068 | * Remove huge page from pool from next node to free. Attempt to keep |
2069 | * persistent huge pages more or less balanced over allowed nodes. | |
2070 | * This routine only 'removes' the hugetlb page. The caller must make | |
2071 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
2072 | * Called with hugetlb_lock locked. |
2073 | */ | |
10c6ec49 MK |
2074 | static struct page *remove_pool_huge_page(struct hstate *h, |
2075 | nodemask_t *nodes_allowed, | |
2076 | bool acct_surplus) | |
e8c5c824 | 2077 | { |
b2261026 | 2078 | int nr_nodes, node; |
10c6ec49 | 2079 | struct page *page = NULL; |
e8c5c824 | 2080 | |
9487ca60 | 2081 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 2082 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
2083 | /* |
2084 | * If we're returning unused surplus pages, only examine | |
2085 | * nodes with surplus pages. | |
2086 | */ | |
b2261026 JK |
2087 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
2088 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 2089 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 2090 | struct page, lru); |
6eb4e88a | 2091 | remove_hugetlb_page(h, page, acct_surplus); |
9a76db09 | 2092 | break; |
e8c5c824 | 2093 | } |
b2261026 | 2094 | } |
e8c5c824 | 2095 | |
10c6ec49 | 2096 | return page; |
e8c5c824 LS |
2097 | } |
2098 | ||
c8721bbb NH |
2099 | /* |
2100 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
2101 | * nothing for in-use hugepages and non-hugepages. |
2102 | * This function returns values like below: | |
2103 | * | |
ad2fa371 MS |
2104 | * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages |
2105 | * when the system is under memory pressure and the feature of | |
2106 | * freeing unused vmemmap pages associated with each hugetlb page | |
2107 | * is enabled. | |
2108 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
2109 | * (allocated or reserved.) | |
2110 | * 0: successfully dissolved free hugepages or the page is not a | |
2111 | * hugepage (considered as already dissolved) | |
c8721bbb | 2112 | */ |
c3114a84 | 2113 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 2114 | { |
6bc9b564 | 2115 | int rc = -EBUSY; |
082d5b6b | 2116 | |
7ffddd49 | 2117 | retry: |
faf53def NH |
2118 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
2119 | if (!PageHuge(page)) | |
2120 | return 0; | |
2121 | ||
db71ef79 | 2122 | spin_lock_irq(&hugetlb_lock); |
faf53def NH |
2123 | if (!PageHuge(page)) { |
2124 | rc = 0; | |
2125 | goto out; | |
2126 | } | |
2127 | ||
2128 | if (!page_count(page)) { | |
2247bb33 GS |
2129 | struct page *head = compound_head(page); |
2130 | struct hstate *h = page_hstate(head); | |
8346d69d | 2131 | if (!available_huge_pages(h)) |
082d5b6b | 2132 | goto out; |
7ffddd49 MS |
2133 | |
2134 | /* | |
2135 | * We should make sure that the page is already on the free list | |
2136 | * when it is dissolved. | |
2137 | */ | |
6c037149 | 2138 | if (unlikely(!HPageFreed(head))) { |
db71ef79 | 2139 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
2140 | cond_resched(); |
2141 | ||
2142 | /* | |
2143 | * Theoretically, we should return -EBUSY when we | |
2144 | * encounter this race. In fact, we have a chance | |
2145 | * to successfully dissolve the page if we do a | |
2146 | * retry. Because the race window is quite small. | |
2147 | * If we seize this opportunity, it is an optimization | |
2148 | * for increasing the success rate of dissolving page. | |
2149 | */ | |
2150 | goto retry; | |
2151 | } | |
2152 | ||
0c5da357 | 2153 | remove_hugetlb_page(h, head, false); |
c1470b33 | 2154 | h->max_huge_pages--; |
db71ef79 | 2155 | spin_unlock_irq(&hugetlb_lock); |
ad2fa371 MS |
2156 | |
2157 | /* | |
2158 | * Normally update_and_free_page will allocate required vmemmmap | |
2159 | * before freeing the page. update_and_free_page will fail to | |
2160 | * free the page if it can not allocate required vmemmap. We | |
2161 | * need to adjust max_huge_pages if the page is not freed. | |
2162 | * Attempt to allocate vmemmmap here so that we can take | |
2163 | * appropriate action on failure. | |
2164 | */ | |
6213834c | 2165 | rc = hugetlb_vmemmap_restore(h, head); |
ad2fa371 | 2166 | if (!rc) { |
ad2fa371 MS |
2167 | update_and_free_page(h, head, false); |
2168 | } else { | |
2169 | spin_lock_irq(&hugetlb_lock); | |
2170 | add_hugetlb_page(h, head, false); | |
2171 | h->max_huge_pages++; | |
2172 | spin_unlock_irq(&hugetlb_lock); | |
2173 | } | |
2174 | ||
2175 | return rc; | |
c8721bbb | 2176 | } |
082d5b6b | 2177 | out: |
db71ef79 | 2178 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 2179 | return rc; |
c8721bbb NH |
2180 | } |
2181 | ||
2182 | /* | |
2183 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
2184 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
2185 | * Note that this will dissolve a free gigantic hugepage completely, if any |
2186 | * part of it lies within the given range. | |
082d5b6b GS |
2187 | * Also note that if dissolve_free_huge_page() returns with an error, all |
2188 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 2189 | */ |
082d5b6b | 2190 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 2191 | { |
c8721bbb | 2192 | unsigned long pfn; |
eb03aa00 | 2193 | struct page *page; |
082d5b6b | 2194 | int rc = 0; |
dc2628f3 MS |
2195 | unsigned int order; |
2196 | struct hstate *h; | |
c8721bbb | 2197 | |
d0177639 | 2198 | if (!hugepages_supported()) |
082d5b6b | 2199 | return rc; |
d0177639 | 2200 | |
dc2628f3 MS |
2201 | order = huge_page_order(&default_hstate); |
2202 | for_each_hstate(h) | |
2203 | order = min(order, huge_page_order(h)); | |
2204 | ||
2205 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << order) { | |
eb03aa00 | 2206 | page = pfn_to_page(pfn); |
faf53def NH |
2207 | rc = dissolve_free_huge_page(page); |
2208 | if (rc) | |
2209 | break; | |
eb03aa00 | 2210 | } |
082d5b6b GS |
2211 | |
2212 | return rc; | |
c8721bbb NH |
2213 | } |
2214 | ||
ab5ac90a MH |
2215 | /* |
2216 | * Allocates a fresh surplus page from the page allocator. | |
2217 | */ | |
0c397dae | 2218 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
2b21624f | 2219 | int nid, nodemask_t *nmask) |
7893d1d5 | 2220 | { |
9980d744 | 2221 | struct page *page = NULL; |
7893d1d5 | 2222 | |
bae7f4ae | 2223 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2224 | return NULL; |
2225 | ||
db71ef79 | 2226 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2227 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
2228 | goto out_unlock; | |
db71ef79 | 2229 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 2230 | |
f60858f9 | 2231 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 2232 | if (!page) |
0c397dae | 2233 | return NULL; |
d1c3fb1f | 2234 | |
db71ef79 | 2235 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2236 | /* |
2237 | * We could have raced with the pool size change. | |
2238 | * Double check that and simply deallocate the new page | |
2239 | * if we would end up overcommiting the surpluses. Abuse | |
2240 | * temporary page to workaround the nasty free_huge_page | |
2241 | * codeflow | |
2242 | */ | |
2243 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 2244 | SetHPageTemporary(page); |
db71ef79 | 2245 | spin_unlock_irq(&hugetlb_lock); |
2b21624f | 2246 | free_huge_page(page); |
2bf753e6 | 2247 | return NULL; |
7893d1d5 | 2248 | } |
9980d744 | 2249 | |
b65a4eda MK |
2250 | h->surplus_huge_pages++; |
2251 | h->surplus_huge_pages_node[page_to_nid(page)]++; | |
2252 | ||
9980d744 | 2253 | out_unlock: |
db71ef79 | 2254 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 AL |
2255 | |
2256 | return page; | |
2257 | } | |
2258 | ||
bbe88753 | 2259 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 2260 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
2261 | { |
2262 | struct page *page; | |
2263 | ||
2264 | if (hstate_is_gigantic(h)) | |
2265 | return NULL; | |
2266 | ||
f60858f9 | 2267 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
2268 | if (!page) |
2269 | return NULL; | |
2270 | ||
2b21624f MK |
2271 | /* fresh huge pages are frozen */ |
2272 | set_page_refcounted(page); | |
2273 | ||
ab5ac90a MH |
2274 | /* |
2275 | * We do not account these pages as surplus because they are only | |
2276 | * temporary and will be released properly on the last reference | |
2277 | */ | |
9157c311 | 2278 | SetHPageTemporary(page); |
ab5ac90a MH |
2279 | |
2280 | return page; | |
2281 | } | |
2282 | ||
099730d6 DH |
2283 | /* |
2284 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
2285 | */ | |
e0ec90ee | 2286 | static |
0c397dae | 2287 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
2288 | struct vm_area_struct *vma, unsigned long addr) |
2289 | { | |
cfcaa66f | 2290 | struct page *page = NULL; |
aaf14e40 MH |
2291 | struct mempolicy *mpol; |
2292 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
2293 | int nid; | |
2294 | nodemask_t *nodemask; | |
2295 | ||
2296 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
2297 | if (mpol_is_preferred_many(mpol)) { |
2298 | gfp_t gfp = gfp_mask | __GFP_NOWARN; | |
2299 | ||
2300 | gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); | |
2b21624f | 2301 | page = alloc_surplus_huge_page(h, gfp, nid, nodemask); |
aaf14e40 | 2302 | |
cfcaa66f BW |
2303 | /* Fallback to all nodes if page==NULL */ |
2304 | nodemask = NULL; | |
2305 | } | |
2306 | ||
2307 | if (!page) | |
2b21624f | 2308 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
cfcaa66f | 2309 | mpol_cond_put(mpol); |
aaf14e40 | 2310 | return page; |
099730d6 DH |
2311 | } |
2312 | ||
ab5ac90a | 2313 | /* page migration callback function */ |
3e59fcb0 | 2314 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 2315 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 2316 | { |
db71ef79 | 2317 | spin_lock_irq(&hugetlb_lock); |
8346d69d | 2318 | if (available_huge_pages(h)) { |
3e59fcb0 MH |
2319 | struct page *page; |
2320 | ||
2321 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
2322 | if (page) { | |
db71ef79 | 2323 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 2324 | return page; |
4db9b2ef MH |
2325 | } |
2326 | } | |
db71ef79 | 2327 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 2328 | |
0c397dae | 2329 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2330 | } |
2331 | ||
ebd63723 | 2332 | /* mempolicy aware migration callback */ |
389c8178 MH |
2333 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2334 | unsigned long address) | |
ebd63723 MH |
2335 | { |
2336 | struct mempolicy *mpol; | |
2337 | nodemask_t *nodemask; | |
2338 | struct page *page; | |
ebd63723 MH |
2339 | gfp_t gfp_mask; |
2340 | int node; | |
2341 | ||
ebd63723 MH |
2342 | gfp_mask = htlb_alloc_mask(h); |
2343 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 2344 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
2345 | mpol_cond_put(mpol); |
2346 | ||
2347 | return page; | |
2348 | } | |
2349 | ||
e4e574b7 | 2350 | /* |
25985edc | 2351 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2352 | * of size 'delta'. |
2353 | */ | |
0a4f3d1b | 2354 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 2355 | __must_hold(&hugetlb_lock) |
e4e574b7 | 2356 | { |
34665341 | 2357 | LIST_HEAD(surplus_list); |
e4e574b7 | 2358 | struct page *page, *tmp; |
0a4f3d1b LX |
2359 | int ret; |
2360 | long i; | |
2361 | long needed, allocated; | |
28073b02 | 2362 | bool alloc_ok = true; |
e4e574b7 | 2363 | |
9487ca60 | 2364 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 2365 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2366 | if (needed <= 0) { |
a5516438 | 2367 | h->resv_huge_pages += delta; |
e4e574b7 | 2368 | return 0; |
ac09b3a1 | 2369 | } |
e4e574b7 AL |
2370 | |
2371 | allocated = 0; | |
e4e574b7 AL |
2372 | |
2373 | ret = -ENOMEM; | |
2374 | retry: | |
db71ef79 | 2375 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 2376 | for (i = 0; i < needed; i++) { |
0c397dae | 2377 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
2b21624f | 2378 | NUMA_NO_NODE, NULL); |
28073b02 HD |
2379 | if (!page) { |
2380 | alloc_ok = false; | |
2381 | break; | |
2382 | } | |
e4e574b7 | 2383 | list_add(&page->lru, &surplus_list); |
69ed779a | 2384 | cond_resched(); |
e4e574b7 | 2385 | } |
28073b02 | 2386 | allocated += i; |
e4e574b7 AL |
2387 | |
2388 | /* | |
2389 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2390 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2391 | */ | |
db71ef79 | 2392 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
2393 | needed = (h->resv_huge_pages + delta) - |
2394 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2395 | if (needed > 0) { |
2396 | if (alloc_ok) | |
2397 | goto retry; | |
2398 | /* | |
2399 | * We were not able to allocate enough pages to | |
2400 | * satisfy the entire reservation so we free what | |
2401 | * we've allocated so far. | |
2402 | */ | |
2403 | goto free; | |
2404 | } | |
e4e574b7 AL |
2405 | /* |
2406 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2407 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2408 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2409 | * allocator. Commit the entire reservation here to prevent another |
2410 | * process from stealing the pages as they are added to the pool but | |
2411 | * before they are reserved. | |
e4e574b7 AL |
2412 | */ |
2413 | needed += allocated; | |
a5516438 | 2414 | h->resv_huge_pages += delta; |
e4e574b7 | 2415 | ret = 0; |
a9869b83 | 2416 | |
19fc3f0a | 2417 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2418 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2419 | if ((--needed) < 0) |
2420 | break; | |
b65a4eda | 2421 | /* Add the page to the hugetlb allocator */ |
a5516438 | 2422 | enqueue_huge_page(h, page); |
19fc3f0a | 2423 | } |
28073b02 | 2424 | free: |
db71ef79 | 2425 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a | 2426 | |
b65a4eda MK |
2427 | /* |
2428 | * Free unnecessary surplus pages to the buddy allocator. | |
2429 | * Pages have no ref count, call free_huge_page directly. | |
2430 | */ | |
c0d934ba | 2431 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
b65a4eda | 2432 | free_huge_page(page); |
db71ef79 | 2433 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2434 | |
2435 | return ret; | |
2436 | } | |
2437 | ||
2438 | /* | |
e5bbc8a6 MK |
2439 | * This routine has two main purposes: |
2440 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2441 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2442 | * to the associated reservation map. | |
2443 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2444 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2445 | */ |
a5516438 AK |
2446 | static void return_unused_surplus_pages(struct hstate *h, |
2447 | unsigned long unused_resv_pages) | |
e4e574b7 | 2448 | { |
e4e574b7 | 2449 | unsigned long nr_pages; |
10c6ec49 MK |
2450 | struct page *page; |
2451 | LIST_HEAD(page_list); | |
2452 | ||
9487ca60 | 2453 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2454 | /* Uncommit the reservation */ |
2455 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2456 | |
c0531714 | 2457 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
e5bbc8a6 | 2458 | goto out; |
aa888a74 | 2459 | |
e5bbc8a6 MK |
2460 | /* |
2461 | * Part (or even all) of the reservation could have been backed | |
2462 | * by pre-allocated pages. Only free surplus pages. | |
2463 | */ | |
a5516438 | 2464 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2465 | |
685f3457 LS |
2466 | /* |
2467 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2468 | * evenly across all nodes with memory. Iterate across these nodes |
2469 | * until we can no longer free unreserved surplus pages. This occurs | |
2470 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2471 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2472 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2473 | */ |
2474 | while (nr_pages--) { | |
10c6ec49 MK |
2475 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2476 | if (!page) | |
e5bbc8a6 | 2477 | goto out; |
10c6ec49 MK |
2478 | |
2479 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2480 | } |
e5bbc8a6 MK |
2481 | |
2482 | out: | |
db71ef79 | 2483 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2484 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2485 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2486 | } |
2487 | ||
5e911373 | 2488 | |
c37f9fb1 | 2489 | /* |
feba16e2 | 2490 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2491 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2492 | * |
2493 | * vma_needs_reservation is called to determine if the huge page at addr | |
2494 | * within the vma has an associated reservation. If a reservation is | |
2495 | * needed, the value 1 is returned. The caller is then responsible for | |
2496 | * managing the global reservation and subpool usage counts. After | |
2497 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2498 | * to add the page to the reservation map. If the page allocation fails, |
2499 | * the reservation must be ended instead of committed. vma_end_reservation | |
2500 | * is called in such cases. | |
cf3ad20b MK |
2501 | * |
2502 | * In the normal case, vma_commit_reservation returns the same value | |
2503 | * as the preceding vma_needs_reservation call. The only time this | |
2504 | * is not the case is if a reserve map was changed between calls. It | |
2505 | * is the responsibility of the caller to notice the difference and | |
2506 | * take appropriate action. | |
96b96a96 MK |
2507 | * |
2508 | * vma_add_reservation is used in error paths where a reservation must | |
2509 | * be restored when a newly allocated huge page must be freed. It is | |
2510 | * to be called after calling vma_needs_reservation to determine if a | |
2511 | * reservation exists. | |
846be085 MK |
2512 | * |
2513 | * vma_del_reservation is used in error paths where an entry in the reserve | |
2514 | * map was created during huge page allocation and must be removed. It is to | |
2515 | * be called after calling vma_needs_reservation to determine if a reservation | |
2516 | * exists. | |
c37f9fb1 | 2517 | */ |
5e911373 MK |
2518 | enum vma_resv_mode { |
2519 | VMA_NEEDS_RESV, | |
2520 | VMA_COMMIT_RESV, | |
feba16e2 | 2521 | VMA_END_RESV, |
96b96a96 | 2522 | VMA_ADD_RESV, |
846be085 | 2523 | VMA_DEL_RESV, |
5e911373 | 2524 | }; |
cf3ad20b MK |
2525 | static long __vma_reservation_common(struct hstate *h, |
2526 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2527 | enum vma_resv_mode mode) |
c37f9fb1 | 2528 | { |
4e35f483 JK |
2529 | struct resv_map *resv; |
2530 | pgoff_t idx; | |
cf3ad20b | 2531 | long ret; |
0db9d74e | 2532 | long dummy_out_regions_needed; |
c37f9fb1 | 2533 | |
4e35f483 JK |
2534 | resv = vma_resv_map(vma); |
2535 | if (!resv) | |
84afd99b | 2536 | return 1; |
c37f9fb1 | 2537 | |
4e35f483 | 2538 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2539 | switch (mode) { |
2540 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2541 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2542 | /* We assume that vma_reservation_* routines always operate on | |
2543 | * 1 page, and that adding to resv map a 1 page entry can only | |
2544 | * ever require 1 region. | |
2545 | */ | |
2546 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2547 | break; |
2548 | case VMA_COMMIT_RESV: | |
075a61d0 | 2549 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2550 | /* region_add calls of range 1 should never fail. */ |
2551 | VM_BUG_ON(ret < 0); | |
5e911373 | 2552 | break; |
feba16e2 | 2553 | case VMA_END_RESV: |
0db9d74e | 2554 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2555 | ret = 0; |
2556 | break; | |
96b96a96 | 2557 | case VMA_ADD_RESV: |
0db9d74e | 2558 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2559 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2560 | /* region_add calls of range 1 should never fail. */ |
2561 | VM_BUG_ON(ret < 0); | |
2562 | } else { | |
2563 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2564 | ret = region_del(resv, idx, idx + 1); |
2565 | } | |
2566 | break; | |
846be085 MK |
2567 | case VMA_DEL_RESV: |
2568 | if (vma->vm_flags & VM_MAYSHARE) { | |
2569 | region_abort(resv, idx, idx + 1, 1); | |
2570 | ret = region_del(resv, idx, idx + 1); | |
2571 | } else { | |
2572 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2573 | /* region_add calls of range 1 should never fail. */ | |
2574 | VM_BUG_ON(ret < 0); | |
2575 | } | |
2576 | break; | |
5e911373 MK |
2577 | default: |
2578 | BUG(); | |
2579 | } | |
84afd99b | 2580 | |
846be085 | 2581 | if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) |
cf3ad20b | 2582 | return ret; |
bf3d12b9 ML |
2583 | /* |
2584 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2585 | * | |
2586 | * In most cases, reserves always exist for private mappings. | |
2587 | * However, a file associated with mapping could have been | |
2588 | * hole punched or truncated after reserves were consumed. | |
2589 | * As subsequent fault on such a range will not use reserves. | |
2590 | * Subtle - The reserve map for private mappings has the | |
2591 | * opposite meaning than that of shared mappings. If NO | |
2592 | * entry is in the reserve map, it means a reservation exists. | |
2593 | * If an entry exists in the reserve map, it means the | |
2594 | * reservation has already been consumed. As a result, the | |
2595 | * return value of this routine is the opposite of the | |
2596 | * value returned from reserve map manipulation routines above. | |
2597 | */ | |
2598 | if (ret > 0) | |
2599 | return 0; | |
2600 | if (ret == 0) | |
2601 | return 1; | |
2602 | return ret; | |
c37f9fb1 | 2603 | } |
cf3ad20b MK |
2604 | |
2605 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2606 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2607 | { |
5e911373 | 2608 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2609 | } |
84afd99b | 2610 | |
cf3ad20b MK |
2611 | static long vma_commit_reservation(struct hstate *h, |
2612 | struct vm_area_struct *vma, unsigned long addr) | |
2613 | { | |
5e911373 MK |
2614 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2615 | } | |
2616 | ||
feba16e2 | 2617 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2618 | struct vm_area_struct *vma, unsigned long addr) |
2619 | { | |
feba16e2 | 2620 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2621 | } |
2622 | ||
96b96a96 MK |
2623 | static long vma_add_reservation(struct hstate *h, |
2624 | struct vm_area_struct *vma, unsigned long addr) | |
2625 | { | |
2626 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2627 | } | |
2628 | ||
846be085 MK |
2629 | static long vma_del_reservation(struct hstate *h, |
2630 | struct vm_area_struct *vma, unsigned long addr) | |
2631 | { | |
2632 | return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); | |
2633 | } | |
2634 | ||
96b96a96 | 2635 | /* |
846be085 MK |
2636 | * This routine is called to restore reservation information on error paths. |
2637 | * It should ONLY be called for pages allocated via alloc_huge_page(), and | |
2638 | * the hugetlb mutex should remain held when calling this routine. | |
2639 | * | |
2640 | * It handles two specific cases: | |
2641 | * 1) A reservation was in place and the page consumed the reservation. | |
2642 | * HPageRestoreReserve is set in the page. | |
2643 | * 2) No reservation was in place for the page, so HPageRestoreReserve is | |
2644 | * not set. However, alloc_huge_page always updates the reserve map. | |
2645 | * | |
2646 | * In case 1, free_huge_page later in the error path will increment the | |
2647 | * global reserve count. But, free_huge_page does not have enough context | |
2648 | * to adjust the reservation map. This case deals primarily with private | |
2649 | * mappings. Adjust the reserve map here to be consistent with global | |
2650 | * reserve count adjustments to be made by free_huge_page. Make sure the | |
2651 | * reserve map indicates there is a reservation present. | |
2652 | * | |
2653 | * In case 2, simply undo reserve map modifications done by alloc_huge_page. | |
96b96a96 | 2654 | */ |
846be085 MK |
2655 | void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, |
2656 | unsigned long address, struct page *page) | |
96b96a96 | 2657 | { |
846be085 | 2658 | long rc = vma_needs_reservation(h, vma, address); |
96b96a96 | 2659 | |
846be085 MK |
2660 | if (HPageRestoreReserve(page)) { |
2661 | if (unlikely(rc < 0)) | |
96b96a96 MK |
2662 | /* |
2663 | * Rare out of memory condition in reserve map | |
d6995da3 | 2664 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2665 | * global reserve count will not be incremented |
2666 | * by free_huge_page. This will make it appear | |
2667 | * as though the reservation for this page was | |
2668 | * consumed. This may prevent the task from | |
2669 | * faulting in the page at a later time. This | |
2670 | * is better than inconsistent global huge page | |
2671 | * accounting of reserve counts. | |
2672 | */ | |
d6995da3 | 2673 | ClearHPageRestoreReserve(page); |
846be085 MK |
2674 | else if (rc) |
2675 | (void)vma_add_reservation(h, vma, address); | |
2676 | else | |
2677 | vma_end_reservation(h, vma, address); | |
2678 | } else { | |
2679 | if (!rc) { | |
2680 | /* | |
2681 | * This indicates there is an entry in the reserve map | |
c7b1850d | 2682 | * not added by alloc_huge_page. We know it was added |
846be085 MK |
2683 | * before the alloc_huge_page call, otherwise |
2684 | * HPageRestoreReserve would be set on the page. | |
2685 | * Remove the entry so that a subsequent allocation | |
2686 | * does not consume a reservation. | |
2687 | */ | |
2688 | rc = vma_del_reservation(h, vma, address); | |
2689 | if (rc < 0) | |
96b96a96 | 2690 | /* |
846be085 MK |
2691 | * VERY rare out of memory condition. Since |
2692 | * we can not delete the entry, set | |
2693 | * HPageRestoreReserve so that the reserve | |
2694 | * count will be incremented when the page | |
2695 | * is freed. This reserve will be consumed | |
2696 | * on a subsequent allocation. | |
96b96a96 | 2697 | */ |
846be085 MK |
2698 | SetHPageRestoreReserve(page); |
2699 | } else if (rc < 0) { | |
2700 | /* | |
2701 | * Rare out of memory condition from | |
2702 | * vma_needs_reservation call. Memory allocation is | |
2703 | * only attempted if a new entry is needed. Therefore, | |
2704 | * this implies there is not an entry in the | |
2705 | * reserve map. | |
2706 | * | |
2707 | * For shared mappings, no entry in the map indicates | |
2708 | * no reservation. We are done. | |
2709 | */ | |
2710 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
2711 | /* | |
2712 | * For private mappings, no entry indicates | |
2713 | * a reservation is present. Since we can | |
2714 | * not add an entry, set SetHPageRestoreReserve | |
2715 | * on the page so reserve count will be | |
2716 | * incremented when freed. This reserve will | |
2717 | * be consumed on a subsequent allocation. | |
2718 | */ | |
2719 | SetHPageRestoreReserve(page); | |
96b96a96 | 2720 | } else |
846be085 MK |
2721 | /* |
2722 | * No reservation present, do nothing | |
2723 | */ | |
2724 | vma_end_reservation(h, vma, address); | |
96b96a96 MK |
2725 | } |
2726 | } | |
2727 | ||
369fa227 OS |
2728 | /* |
2729 | * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one | |
2730 | * @h: struct hstate old page belongs to | |
2731 | * @old_page: Old page to dissolve | |
ae37c7ff | 2732 | * @list: List to isolate the page in case we need to |
369fa227 OS |
2733 | * Returns 0 on success, otherwise negated error. |
2734 | */ | |
ae37c7ff OS |
2735 | static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, |
2736 | struct list_head *list) | |
369fa227 OS |
2737 | { |
2738 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
2739 | int nid = page_to_nid(old_page); | |
2740 | struct page *new_page; | |
2741 | int ret = 0; | |
2742 | ||
2743 | /* | |
2744 | * Before dissolving the page, we need to allocate a new one for the | |
f41f2ed4 MS |
2745 | * pool to remain stable. Here, we allocate the page and 'prep' it |
2746 | * by doing everything but actually updating counters and adding to | |
2747 | * the pool. This simplifies and let us do most of the processing | |
2748 | * under the lock. | |
369fa227 OS |
2749 | */ |
2750 | new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); | |
2751 | if (!new_page) | |
2752 | return -ENOMEM; | |
f41f2ed4 | 2753 | __prep_new_huge_page(h, new_page); |
369fa227 OS |
2754 | |
2755 | retry: | |
2756 | spin_lock_irq(&hugetlb_lock); | |
2757 | if (!PageHuge(old_page)) { | |
2758 | /* | |
2759 | * Freed from under us. Drop new_page too. | |
2760 | */ | |
2761 | goto free_new; | |
2762 | } else if (page_count(old_page)) { | |
2763 | /* | |
ae37c7ff OS |
2764 | * Someone has grabbed the page, try to isolate it here. |
2765 | * Fail with -EBUSY if not possible. | |
369fa227 | 2766 | */ |
ae37c7ff | 2767 | spin_unlock_irq(&hugetlb_lock); |
7ce82f4c | 2768 | ret = isolate_hugetlb(old_page, list); |
ae37c7ff | 2769 | spin_lock_irq(&hugetlb_lock); |
369fa227 OS |
2770 | goto free_new; |
2771 | } else if (!HPageFreed(old_page)) { | |
2772 | /* | |
2773 | * Page's refcount is 0 but it has not been enqueued in the | |
2774 | * freelist yet. Race window is small, so we can succeed here if | |
2775 | * we retry. | |
2776 | */ | |
2777 | spin_unlock_irq(&hugetlb_lock); | |
2778 | cond_resched(); | |
2779 | goto retry; | |
2780 | } else { | |
2781 | /* | |
2782 | * Ok, old_page is still a genuine free hugepage. Remove it from | |
2783 | * the freelist and decrease the counters. These will be | |
2784 | * incremented again when calling __prep_account_new_huge_page() | |
2785 | * and enqueue_huge_page() for new_page. The counters will remain | |
2786 | * stable since this happens under the lock. | |
2787 | */ | |
2788 | remove_hugetlb_page(h, old_page, false); | |
2789 | ||
2790 | /* | |
b65a4eda MK |
2791 | * Ref count on new page is already zero as it was dropped |
2792 | * earlier. It can be directly added to the pool free list. | |
369fa227 | 2793 | */ |
369fa227 | 2794 | __prep_account_new_huge_page(h, nid); |
369fa227 OS |
2795 | enqueue_huge_page(h, new_page); |
2796 | ||
2797 | /* | |
2798 | * Pages have been replaced, we can safely free the old one. | |
2799 | */ | |
2800 | spin_unlock_irq(&hugetlb_lock); | |
b65d4adb | 2801 | update_and_free_page(h, old_page, false); |
369fa227 OS |
2802 | } |
2803 | ||
2804 | return ret; | |
2805 | ||
2806 | free_new: | |
2807 | spin_unlock_irq(&hugetlb_lock); | |
b65a4eda MK |
2808 | /* Page has a zero ref count, but needs a ref to be freed */ |
2809 | set_page_refcounted(new_page); | |
b65d4adb | 2810 | update_and_free_page(h, new_page, false); |
369fa227 OS |
2811 | |
2812 | return ret; | |
2813 | } | |
2814 | ||
ae37c7ff | 2815 | int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) |
369fa227 OS |
2816 | { |
2817 | struct hstate *h; | |
2818 | struct page *head; | |
ae37c7ff | 2819 | int ret = -EBUSY; |
369fa227 OS |
2820 | |
2821 | /* | |
2822 | * The page might have been dissolved from under our feet, so make sure | |
2823 | * to carefully check the state under the lock. | |
2824 | * Return success when racing as if we dissolved the page ourselves. | |
2825 | */ | |
2826 | spin_lock_irq(&hugetlb_lock); | |
2827 | if (PageHuge(page)) { | |
2828 | head = compound_head(page); | |
2829 | h = page_hstate(head); | |
2830 | } else { | |
2831 | spin_unlock_irq(&hugetlb_lock); | |
2832 | return 0; | |
2833 | } | |
2834 | spin_unlock_irq(&hugetlb_lock); | |
2835 | ||
2836 | /* | |
2837 | * Fence off gigantic pages as there is a cyclic dependency between | |
2838 | * alloc_contig_range and them. Return -ENOMEM as this has the effect | |
2839 | * of bailing out right away without further retrying. | |
2840 | */ | |
2841 | if (hstate_is_gigantic(h)) | |
2842 | return -ENOMEM; | |
2843 | ||
7ce82f4c | 2844 | if (page_count(head) && !isolate_hugetlb(head, list)) |
ae37c7ff OS |
2845 | ret = 0; |
2846 | else if (!page_count(head)) | |
2847 | ret = alloc_and_dissolve_huge_page(h, head, list); | |
2848 | ||
2849 | return ret; | |
369fa227 OS |
2850 | } |
2851 | ||
70c3547e | 2852 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2853 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2854 | { |
90481622 | 2855 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2856 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2857 | struct page *page; |
d85f69b0 MK |
2858 | long map_chg, map_commit; |
2859 | long gbl_chg; | |
6d76dcf4 AK |
2860 | int ret, idx; |
2861 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2862 | bool deferred_reserve; |
a1e78772 | 2863 | |
6d76dcf4 | 2864 | idx = hstate_index(h); |
a1e78772 | 2865 | /* |
d85f69b0 MK |
2866 | * Examine the region/reserve map to determine if the process |
2867 | * has a reservation for the page to be allocated. A return | |
2868 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2869 | */ |
d85f69b0 MK |
2870 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2871 | if (map_chg < 0) | |
76dcee75 | 2872 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2873 | |
2874 | /* | |
2875 | * Processes that did not create the mapping will have no | |
2876 | * reserves as indicated by the region/reserve map. Check | |
2877 | * that the allocation will not exceed the subpool limit. | |
2878 | * Allocations for MAP_NORESERVE mappings also need to be | |
2879 | * checked against any subpool limit. | |
2880 | */ | |
2881 | if (map_chg || avoid_reserve) { | |
2882 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2883 | if (gbl_chg < 0) { | |
feba16e2 | 2884 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2885 | return ERR_PTR(-ENOSPC); |
5e911373 | 2886 | } |
1da177e4 | 2887 | |
d85f69b0 MK |
2888 | /* |
2889 | * Even though there was no reservation in the region/reserve | |
2890 | * map, there could be reservations associated with the | |
2891 | * subpool that can be used. This would be indicated if the | |
2892 | * return value of hugepage_subpool_get_pages() is zero. | |
2893 | * However, if avoid_reserve is specified we still avoid even | |
2894 | * the subpool reservations. | |
2895 | */ | |
2896 | if (avoid_reserve) | |
2897 | gbl_chg = 1; | |
2898 | } | |
2899 | ||
08cf9faf MA |
2900 | /* If this allocation is not consuming a reservation, charge it now. |
2901 | */ | |
6501fe5f | 2902 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2903 | if (deferred_reserve) { |
2904 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2905 | idx, pages_per_huge_page(h), &h_cg); | |
2906 | if (ret) | |
2907 | goto out_subpool_put; | |
2908 | } | |
2909 | ||
6d76dcf4 | 2910 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2911 | if (ret) |
08cf9faf | 2912 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2913 | |
db71ef79 | 2914 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
2915 | /* |
2916 | * glb_chg is passed to indicate whether or not a page must be taken | |
2917 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2918 | * a reservation exists for the allocation. | |
2919 | */ | |
2920 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2921 | if (!page) { |
db71ef79 | 2922 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 2923 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2924 | if (!page) |
2925 | goto out_uncharge_cgroup; | |
a88c7695 | 2926 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2927 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2928 | h->resv_huge_pages--; |
2929 | } | |
db71ef79 | 2930 | spin_lock_irq(&hugetlb_lock); |
15a8d68e | 2931 | list_add(&page->lru, &h->hugepage_activelist); |
2b21624f | 2932 | set_page_refcounted(page); |
81a6fcae | 2933 | /* Fall through */ |
68842c9b | 2934 | } |
81a6fcae | 2935 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2936 | /* If allocation is not consuming a reservation, also store the |
2937 | * hugetlb_cgroup pointer on the page. | |
2938 | */ | |
2939 | if (deferred_reserve) { | |
2940 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2941 | h_cg, page); | |
2942 | } | |
2943 | ||
db71ef79 | 2944 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 2945 | |
d6995da3 | 2946 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2947 | |
d85f69b0 MK |
2948 | map_commit = vma_commit_reservation(h, vma, addr); |
2949 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2950 | /* |
2951 | * The page was added to the reservation map between | |
2952 | * vma_needs_reservation and vma_commit_reservation. | |
2953 | * This indicates a race with hugetlb_reserve_pages. | |
2954 | * Adjust for the subpool count incremented above AND | |
2955 | * in hugetlb_reserve_pages for the same page. Also, | |
2956 | * the reservation count added in hugetlb_reserve_pages | |
2957 | * no longer applies. | |
2958 | */ | |
2959 | long rsv_adjust; | |
2960 | ||
2961 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2962 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2963 | if (deferred_reserve) |
2964 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2965 | pages_per_huge_page(h), page); | |
33039678 | 2966 | } |
90d8b7e6 | 2967 | return page; |
8f34af6f JZ |
2968 | |
2969 | out_uncharge_cgroup: | |
2970 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2971 | out_uncharge_cgroup_reservation: |
2972 | if (deferred_reserve) | |
2973 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2974 | h_cg); | |
8f34af6f | 2975 | out_subpool_put: |
d85f69b0 | 2976 | if (map_chg || avoid_reserve) |
8f34af6f | 2977 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2978 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2979 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2980 | } |
2981 | ||
b5389086 | 2982 | int alloc_bootmem_huge_page(struct hstate *h, int nid) |
e24a1307 | 2983 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); |
b5389086 | 2984 | int __alloc_bootmem_huge_page(struct hstate *h, int nid) |
aa888a74 | 2985 | { |
b5389086 | 2986 | struct huge_bootmem_page *m = NULL; /* initialize for clang */ |
b2261026 | 2987 | int nr_nodes, node; |
aa888a74 | 2988 | |
b5389086 ZY |
2989 | /* do node specific alloc */ |
2990 | if (nid != NUMA_NO_NODE) { | |
2991 | m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), | |
2992 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); | |
2993 | if (!m) | |
2994 | return 0; | |
2995 | goto found; | |
2996 | } | |
2997 | /* allocate from next node when distributing huge pages */ | |
b2261026 | 2998 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
b5389086 | 2999 | m = memblock_alloc_try_nid_raw( |
8b89a116 | 3000 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 3001 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
b5389086 ZY |
3002 | /* |
3003 | * Use the beginning of the huge page to store the | |
3004 | * huge_bootmem_page struct (until gather_bootmem | |
3005 | * puts them into the mem_map). | |
3006 | */ | |
3007 | if (!m) | |
3008 | return 0; | |
3009 | goto found; | |
aa888a74 | 3010 | } |
aa888a74 AK |
3011 | |
3012 | found: | |
aa888a74 | 3013 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 3014 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
3015 | list_add(&m->list, &huge_boot_pages); |
3016 | m->hstate = h; | |
3017 | return 1; | |
3018 | } | |
3019 | ||
48b8d744 MK |
3020 | /* |
3021 | * Put bootmem huge pages into the standard lists after mem_map is up. | |
3022 | * Note: This only applies to gigantic (order > MAX_ORDER) pages. | |
3023 | */ | |
aa888a74 AK |
3024 | static void __init gather_bootmem_prealloc(void) |
3025 | { | |
3026 | struct huge_bootmem_page *m; | |
3027 | ||
3028 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 3029 | struct page *page = virt_to_page(m); |
aa888a74 | 3030 | struct hstate *h = m->hstate; |
ee8f248d | 3031 | |
48b8d744 | 3032 | VM_BUG_ON(!hstate_is_gigantic(h)); |
aa888a74 | 3033 | WARN_ON(page_count(page) != 1); |
7118fc29 MK |
3034 | if (prep_compound_gigantic_page(page, huge_page_order(h))) { |
3035 | WARN_ON(PageReserved(page)); | |
3036 | prep_new_huge_page(h, page, page_to_nid(page)); | |
2b21624f | 3037 | free_huge_page(page); /* add to the hugepage allocator */ |
7118fc29 | 3038 | } else { |
416d85ed | 3039 | /* VERY unlikely inflated ref count on a tail page */ |
7118fc29 | 3040 | free_gigantic_page(page, huge_page_order(h)); |
7118fc29 | 3041 | } |
af0fb9df | 3042 | |
b0320c7b | 3043 | /* |
48b8d744 MK |
3044 | * We need to restore the 'stolen' pages to totalram_pages |
3045 | * in order to fix confusing memory reports from free(1) and | |
3046 | * other side-effects, like CommitLimit going negative. | |
b0320c7b | 3047 | */ |
48b8d744 | 3048 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 3049 | cond_resched(); |
aa888a74 AK |
3050 | } |
3051 | } | |
b5389086 ZY |
3052 | static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) |
3053 | { | |
3054 | unsigned long i; | |
3055 | char buf[32]; | |
3056 | ||
3057 | for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { | |
3058 | if (hstate_is_gigantic(h)) { | |
3059 | if (!alloc_bootmem_huge_page(h, nid)) | |
3060 | break; | |
3061 | } else { | |
3062 | struct page *page; | |
3063 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
3064 | ||
3065 | page = alloc_fresh_huge_page(h, gfp_mask, nid, | |
3066 | &node_states[N_MEMORY], NULL); | |
3067 | if (!page) | |
3068 | break; | |
2b21624f | 3069 | free_huge_page(page); /* free it into the hugepage allocator */ |
b5389086 ZY |
3070 | } |
3071 | cond_resched(); | |
3072 | } | |
3073 | if (i == h->max_huge_pages_node[nid]) | |
3074 | return; | |
3075 | ||
3076 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
3077 | pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", | |
3078 | h->max_huge_pages_node[nid], buf, nid, i); | |
3079 | h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); | |
3080 | h->max_huge_pages_node[nid] = i; | |
3081 | } | |
aa888a74 | 3082 | |
8faa8b07 | 3083 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
3084 | { |
3085 | unsigned long i; | |
f60858f9 | 3086 | nodemask_t *node_alloc_noretry; |
b5389086 ZY |
3087 | bool node_specific_alloc = false; |
3088 | ||
3089 | /* skip gigantic hugepages allocation if hugetlb_cma enabled */ | |
3090 | if (hstate_is_gigantic(h) && hugetlb_cma_size) { | |
3091 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); | |
3092 | return; | |
3093 | } | |
3094 | ||
3095 | /* do node specific alloc */ | |
0a7a0f6f | 3096 | for_each_online_node(i) { |
b5389086 ZY |
3097 | if (h->max_huge_pages_node[i] > 0) { |
3098 | hugetlb_hstate_alloc_pages_onenode(h, i); | |
3099 | node_specific_alloc = true; | |
3100 | } | |
3101 | } | |
f60858f9 | 3102 | |
b5389086 ZY |
3103 | if (node_specific_alloc) |
3104 | return; | |
3105 | ||
3106 | /* below will do all node balanced alloc */ | |
f60858f9 MK |
3107 | if (!hstate_is_gigantic(h)) { |
3108 | /* | |
3109 | * Bit mask controlling how hard we retry per-node allocations. | |
3110 | * Ignore errors as lower level routines can deal with | |
3111 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
3112 | * time, we are likely in bigger trouble. | |
3113 | */ | |
3114 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
3115 | GFP_KERNEL); | |
3116 | } else { | |
3117 | /* allocations done at boot time */ | |
3118 | node_alloc_noretry = NULL; | |
3119 | } | |
3120 | ||
3121 | /* bit mask controlling how hard we retry per-node allocations */ | |
3122 | if (node_alloc_noretry) | |
3123 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 3124 | |
e5ff2159 | 3125 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 3126 | if (hstate_is_gigantic(h)) { |
b5389086 | 3127 | if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) |
aa888a74 | 3128 | break; |
0c397dae | 3129 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
3130 | &node_states[N_MEMORY], |
3131 | node_alloc_noretry)) | |
1da177e4 | 3132 | break; |
69ed779a | 3133 | cond_resched(); |
1da177e4 | 3134 | } |
d715cf80 LH |
3135 | if (i < h->max_huge_pages) { |
3136 | char buf[32]; | |
3137 | ||
c6247f72 | 3138 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
3139 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
3140 | h->max_huge_pages, buf, i); | |
3141 | h->max_huge_pages = i; | |
3142 | } | |
f60858f9 | 3143 | kfree(node_alloc_noretry); |
e5ff2159 AK |
3144 | } |
3145 | ||
3146 | static void __init hugetlb_init_hstates(void) | |
3147 | { | |
79dfc695 | 3148 | struct hstate *h, *h2; |
e5ff2159 AK |
3149 | |
3150 | for_each_hstate(h) { | |
8faa8b07 | 3151 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 3152 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 3153 | hugetlb_hstate_alloc_pages(h); |
79dfc695 MK |
3154 | |
3155 | /* | |
3156 | * Set demote order for each hstate. Note that | |
3157 | * h->demote_order is initially 0. | |
3158 | * - We can not demote gigantic pages if runtime freeing | |
3159 | * is not supported, so skip this. | |
a01f4390 MK |
3160 | * - If CMA allocation is possible, we can not demote |
3161 | * HUGETLB_PAGE_ORDER or smaller size pages. | |
79dfc695 MK |
3162 | */ |
3163 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
3164 | continue; | |
a01f4390 MK |
3165 | if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) |
3166 | continue; | |
79dfc695 MK |
3167 | for_each_hstate(h2) { |
3168 | if (h2 == h) | |
3169 | continue; | |
3170 | if (h2->order < h->order && | |
3171 | h2->order > h->demote_order) | |
3172 | h->demote_order = h2->order; | |
3173 | } | |
e5ff2159 AK |
3174 | } |
3175 | } | |
3176 | ||
3177 | static void __init report_hugepages(void) | |
3178 | { | |
3179 | struct hstate *h; | |
3180 | ||
3181 | for_each_hstate(h) { | |
4abd32db | 3182 | char buf[32]; |
c6247f72 MW |
3183 | |
3184 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
6213834c | 3185 | pr_info("HugeTLB: registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 3186 | buf, h->free_huge_pages); |
6213834c MS |
3187 | pr_info("HugeTLB: %d KiB vmemmap can be freed for a %s page\n", |
3188 | hugetlb_vmemmap_optimizable_size(h) / SZ_1K, buf); | |
e5ff2159 AK |
3189 | } |
3190 | } | |
3191 | ||
1da177e4 | 3192 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
3193 | static void try_to_free_low(struct hstate *h, unsigned long count, |
3194 | nodemask_t *nodes_allowed) | |
1da177e4 | 3195 | { |
4415cc8d | 3196 | int i; |
1121828a | 3197 | LIST_HEAD(page_list); |
4415cc8d | 3198 | |
9487ca60 | 3199 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 3200 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
3201 | return; |
3202 | ||
1121828a MK |
3203 | /* |
3204 | * Collect pages to be freed on a list, and free after dropping lock | |
3205 | */ | |
6ae11b27 | 3206 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 3207 | struct page *page, *next; |
a5516438 AK |
3208 | struct list_head *freel = &h->hugepage_freelists[i]; |
3209 | list_for_each_entry_safe(page, next, freel, lru) { | |
3210 | if (count >= h->nr_huge_pages) | |
1121828a | 3211 | goto out; |
1da177e4 LT |
3212 | if (PageHighMem(page)) |
3213 | continue; | |
6eb4e88a | 3214 | remove_hugetlb_page(h, page, false); |
1121828a | 3215 | list_add(&page->lru, &page_list); |
1da177e4 LT |
3216 | } |
3217 | } | |
1121828a MK |
3218 | |
3219 | out: | |
db71ef79 | 3220 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3221 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 3222 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
3223 | } |
3224 | #else | |
6ae11b27 LS |
3225 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
3226 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
3227 | { |
3228 | } | |
3229 | #endif | |
3230 | ||
20a0307c WF |
3231 | /* |
3232 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
3233 | * balanced by operating on them in a round-robin fashion. | |
3234 | * Returns 1 if an adjustment was made. | |
3235 | */ | |
6ae11b27 LS |
3236 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
3237 | int delta) | |
20a0307c | 3238 | { |
b2261026 | 3239 | int nr_nodes, node; |
20a0307c | 3240 | |
9487ca60 | 3241 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 3242 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 3243 | |
b2261026 JK |
3244 | if (delta < 0) { |
3245 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
3246 | if (h->surplus_huge_pages_node[node]) | |
3247 | goto found; | |
e8c5c824 | 3248 | } |
b2261026 JK |
3249 | } else { |
3250 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
3251 | if (h->surplus_huge_pages_node[node] < | |
3252 | h->nr_huge_pages_node[node]) | |
3253 | goto found; | |
e8c5c824 | 3254 | } |
b2261026 JK |
3255 | } |
3256 | return 0; | |
20a0307c | 3257 | |
b2261026 JK |
3258 | found: |
3259 | h->surplus_huge_pages += delta; | |
3260 | h->surplus_huge_pages_node[node] += delta; | |
3261 | return 1; | |
20a0307c WF |
3262 | } |
3263 | ||
a5516438 | 3264 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 3265 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 3266 | nodemask_t *nodes_allowed) |
1da177e4 | 3267 | { |
7893d1d5 | 3268 | unsigned long min_count, ret; |
10c6ec49 MK |
3269 | struct page *page; |
3270 | LIST_HEAD(page_list); | |
f60858f9 MK |
3271 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
3272 | ||
3273 | /* | |
3274 | * Bit mask controlling how hard we retry per-node allocations. | |
3275 | * If we can not allocate the bit mask, do not attempt to allocate | |
3276 | * the requested huge pages. | |
3277 | */ | |
3278 | if (node_alloc_noretry) | |
3279 | nodes_clear(*node_alloc_noretry); | |
3280 | else | |
3281 | return -ENOMEM; | |
1da177e4 | 3282 | |
29383967 MK |
3283 | /* |
3284 | * resize_lock mutex prevents concurrent adjustments to number of | |
3285 | * pages in hstate via the proc/sysfs interfaces. | |
3286 | */ | |
3287 | mutex_lock(&h->resize_lock); | |
b65d4adb | 3288 | flush_free_hpage_work(h); |
db71ef79 | 3289 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 3290 | |
fd875dca MK |
3291 | /* |
3292 | * Check for a node specific request. | |
3293 | * Changing node specific huge page count may require a corresponding | |
3294 | * change to the global count. In any case, the passed node mask | |
3295 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
3296 | */ | |
3297 | if (nid != NUMA_NO_NODE) { | |
3298 | unsigned long old_count = count; | |
3299 | ||
3300 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
3301 | /* | |
3302 | * User may have specified a large count value which caused the | |
3303 | * above calculation to overflow. In this case, they wanted | |
3304 | * to allocate as many huge pages as possible. Set count to | |
3305 | * largest possible value to align with their intention. | |
3306 | */ | |
3307 | if (count < old_count) | |
3308 | count = ULONG_MAX; | |
3309 | } | |
3310 | ||
4eb0716e AG |
3311 | /* |
3312 | * Gigantic pages runtime allocation depend on the capability for large | |
3313 | * page range allocation. | |
3314 | * If the system does not provide this feature, return an error when | |
3315 | * the user tries to allocate gigantic pages but let the user free the | |
3316 | * boottime allocated gigantic pages. | |
3317 | */ | |
3318 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
3319 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 3320 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3321 | mutex_unlock(&h->resize_lock); |
f60858f9 | 3322 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
3323 | return -EINVAL; |
3324 | } | |
3325 | /* Fall through to decrease pool */ | |
3326 | } | |
aa888a74 | 3327 | |
7893d1d5 AL |
3328 | /* |
3329 | * Increase the pool size | |
3330 | * First take pages out of surplus state. Then make up the | |
3331 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 3332 | * |
0c397dae | 3333 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
3334 | * to convert a surplus huge page to a normal huge page. That is |
3335 | * not critical, though, it just means the overall size of the | |
3336 | * pool might be one hugepage larger than it needs to be, but | |
3337 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 3338 | */ |
a5516438 | 3339 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 3340 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
3341 | break; |
3342 | } | |
3343 | ||
a5516438 | 3344 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
3345 | /* |
3346 | * If this allocation races such that we no longer need the | |
3347 | * page, free_huge_page will handle it by freeing the page | |
3348 | * and reducing the surplus. | |
3349 | */ | |
db71ef79 | 3350 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
3351 | |
3352 | /* yield cpu to avoid soft lockup */ | |
3353 | cond_resched(); | |
3354 | ||
f60858f9 MK |
3355 | ret = alloc_pool_huge_page(h, nodes_allowed, |
3356 | node_alloc_noretry); | |
db71ef79 | 3357 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
3358 | if (!ret) |
3359 | goto out; | |
3360 | ||
536240f2 MG |
3361 | /* Bail for signals. Probably ctrl-c from user */ |
3362 | if (signal_pending(current)) | |
3363 | goto out; | |
7893d1d5 | 3364 | } |
7893d1d5 AL |
3365 | |
3366 | /* | |
3367 | * Decrease the pool size | |
3368 | * First return free pages to the buddy allocator (being careful | |
3369 | * to keep enough around to satisfy reservations). Then place | |
3370 | * pages into surplus state as needed so the pool will shrink | |
3371 | * to the desired size as pages become free. | |
d1c3fb1f NA |
3372 | * |
3373 | * By placing pages into the surplus state independent of the | |
3374 | * overcommit value, we are allowing the surplus pool size to | |
3375 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 3376 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
3377 | * though, we'll note that we're not allowed to exceed surplus |
3378 | * and won't grow the pool anywhere else. Not until one of the | |
3379 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 3380 | */ |
a5516438 | 3381 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 3382 | min_count = max(count, min_count); |
6ae11b27 | 3383 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
3384 | |
3385 | /* | |
3386 | * Collect pages to be removed on list without dropping lock | |
3387 | */ | |
a5516438 | 3388 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
3389 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
3390 | if (!page) | |
1da177e4 | 3391 | break; |
10c6ec49 MK |
3392 | |
3393 | list_add(&page->lru, &page_list); | |
1da177e4 | 3394 | } |
10c6ec49 | 3395 | /* free the pages after dropping lock */ |
db71ef79 | 3396 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3397 | update_and_free_pages_bulk(h, &page_list); |
b65d4adb | 3398 | flush_free_hpage_work(h); |
db71ef79 | 3399 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 3400 | |
a5516438 | 3401 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 3402 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
3403 | break; |
3404 | } | |
3405 | out: | |
4eb0716e | 3406 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 3407 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3408 | mutex_unlock(&h->resize_lock); |
4eb0716e | 3409 | |
f60858f9 MK |
3410 | NODEMASK_FREE(node_alloc_noretry); |
3411 | ||
4eb0716e | 3412 | return 0; |
1da177e4 LT |
3413 | } |
3414 | ||
8531fc6f MK |
3415 | static int demote_free_huge_page(struct hstate *h, struct page *page) |
3416 | { | |
3417 | int i, nid = page_to_nid(page); | |
3418 | struct hstate *target_hstate; | |
31731452 | 3419 | struct page *subpage; |
8531fc6f MK |
3420 | int rc = 0; |
3421 | ||
3422 | target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); | |
3423 | ||
3424 | remove_hugetlb_page_for_demote(h, page, false); | |
3425 | spin_unlock_irq(&hugetlb_lock); | |
3426 | ||
6213834c | 3427 | rc = hugetlb_vmemmap_restore(h, page); |
8531fc6f MK |
3428 | if (rc) { |
3429 | /* Allocation of vmemmmap failed, we can not demote page */ | |
3430 | spin_lock_irq(&hugetlb_lock); | |
3431 | set_page_refcounted(page); | |
3432 | add_hugetlb_page(h, page, false); | |
3433 | return rc; | |
3434 | } | |
3435 | ||
3436 | /* | |
3437 | * Use destroy_compound_hugetlb_page_for_demote for all huge page | |
3438 | * sizes as it will not ref count pages. | |
3439 | */ | |
3440 | destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); | |
3441 | ||
3442 | /* | |
3443 | * Taking target hstate mutex synchronizes with set_max_huge_pages. | |
3444 | * Without the mutex, pages added to target hstate could be marked | |
3445 | * as surplus. | |
3446 | * | |
3447 | * Note that we already hold h->resize_lock. To prevent deadlock, | |
3448 | * use the convention of always taking larger size hstate mutex first. | |
3449 | */ | |
3450 | mutex_lock(&target_hstate->resize_lock); | |
3451 | for (i = 0; i < pages_per_huge_page(h); | |
3452 | i += pages_per_huge_page(target_hstate)) { | |
31731452 | 3453 | subpage = nth_page(page, i); |
8531fc6f | 3454 | if (hstate_is_gigantic(target_hstate)) |
31731452 | 3455 | prep_compound_gigantic_page_for_demote(subpage, |
8531fc6f MK |
3456 | target_hstate->order); |
3457 | else | |
31731452 DB |
3458 | prep_compound_page(subpage, target_hstate->order); |
3459 | set_page_private(subpage, 0); | |
31731452 | 3460 | prep_new_huge_page(target_hstate, subpage, nid); |
2b21624f | 3461 | free_huge_page(subpage); |
8531fc6f MK |
3462 | } |
3463 | mutex_unlock(&target_hstate->resize_lock); | |
3464 | ||
3465 | spin_lock_irq(&hugetlb_lock); | |
3466 | ||
3467 | /* | |
3468 | * Not absolutely necessary, but for consistency update max_huge_pages | |
3469 | * based on pool changes for the demoted page. | |
3470 | */ | |
3471 | h->max_huge_pages--; | |
a43a83c7 ML |
3472 | target_hstate->max_huge_pages += |
3473 | pages_per_huge_page(h) / pages_per_huge_page(target_hstate); | |
8531fc6f MK |
3474 | |
3475 | return rc; | |
3476 | } | |
3477 | ||
79dfc695 MK |
3478 | static int demote_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed) |
3479 | __must_hold(&hugetlb_lock) | |
3480 | { | |
8531fc6f MK |
3481 | int nr_nodes, node; |
3482 | struct page *page; | |
79dfc695 MK |
3483 | |
3484 | lockdep_assert_held(&hugetlb_lock); | |
3485 | ||
3486 | /* We should never get here if no demote order */ | |
3487 | if (!h->demote_order) { | |
3488 | pr_warn("HugeTLB: NULL demote order passed to demote_pool_huge_page.\n"); | |
3489 | return -EINVAL; /* internal error */ | |
3490 | } | |
3491 | ||
8531fc6f | 3492 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
5a317412 MK |
3493 | list_for_each_entry(page, &h->hugepage_freelists[node], lru) { |
3494 | if (PageHWPoison(page)) | |
3495 | continue; | |
3496 | ||
3497 | return demote_free_huge_page(h, page); | |
8531fc6f MK |
3498 | } |
3499 | } | |
3500 | ||
5a317412 MK |
3501 | /* |
3502 | * Only way to get here is if all pages on free lists are poisoned. | |
3503 | * Return -EBUSY so that caller will not retry. | |
3504 | */ | |
3505 | return -EBUSY; | |
79dfc695 MK |
3506 | } |
3507 | ||
a3437870 NA |
3508 | #define HSTATE_ATTR_RO(_name) \ |
3509 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
3510 | ||
79dfc695 MK |
3511 | #define HSTATE_ATTR_WO(_name) \ |
3512 | static struct kobj_attribute _name##_attr = __ATTR_WO(_name) | |
3513 | ||
a3437870 | 3514 | #define HSTATE_ATTR(_name) \ |
98bc26ac | 3515 | static struct kobj_attribute _name##_attr = __ATTR_RW(_name) |
a3437870 NA |
3516 | |
3517 | static struct kobject *hugepages_kobj; | |
3518 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3519 | ||
9a305230 LS |
3520 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
3521 | ||
3522 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
3523 | { |
3524 | int i; | |
9a305230 | 3525 | |
a3437870 | 3526 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
3527 | if (hstate_kobjs[i] == kobj) { |
3528 | if (nidp) | |
3529 | *nidp = NUMA_NO_NODE; | |
a3437870 | 3530 | return &hstates[i]; |
9a305230 LS |
3531 | } |
3532 | ||
3533 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
3534 | } |
3535 | ||
06808b08 | 3536 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
3537 | struct kobj_attribute *attr, char *buf) |
3538 | { | |
9a305230 LS |
3539 | struct hstate *h; |
3540 | unsigned long nr_huge_pages; | |
3541 | int nid; | |
3542 | ||
3543 | h = kobj_to_hstate(kobj, &nid); | |
3544 | if (nid == NUMA_NO_NODE) | |
3545 | nr_huge_pages = h->nr_huge_pages; | |
3546 | else | |
3547 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
3548 | ||
ae7a927d | 3549 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 3550 | } |
adbe8726 | 3551 | |
238d3c13 DR |
3552 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
3553 | struct hstate *h, int nid, | |
3554 | unsigned long count, size_t len) | |
a3437870 NA |
3555 | { |
3556 | int err; | |
2d0adf7e | 3557 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 3558 | |
2d0adf7e OS |
3559 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
3560 | return -EINVAL; | |
adbe8726 | 3561 | |
9a305230 LS |
3562 | if (nid == NUMA_NO_NODE) { |
3563 | /* | |
3564 | * global hstate attribute | |
3565 | */ | |
3566 | if (!(obey_mempolicy && | |
2d0adf7e OS |
3567 | init_nodemask_of_mempolicy(&nodes_allowed))) |
3568 | n_mask = &node_states[N_MEMORY]; | |
3569 | else | |
3570 | n_mask = &nodes_allowed; | |
3571 | } else { | |
9a305230 | 3572 | /* |
fd875dca MK |
3573 | * Node specific request. count adjustment happens in |
3574 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 3575 | */ |
2d0adf7e OS |
3576 | init_nodemask_of_node(&nodes_allowed, nid); |
3577 | n_mask = &nodes_allowed; | |
fd875dca | 3578 | } |
9a305230 | 3579 | |
2d0adf7e | 3580 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 3581 | |
4eb0716e | 3582 | return err ? err : len; |
06808b08 LS |
3583 | } |
3584 | ||
238d3c13 DR |
3585 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
3586 | struct kobject *kobj, const char *buf, | |
3587 | size_t len) | |
3588 | { | |
3589 | struct hstate *h; | |
3590 | unsigned long count; | |
3591 | int nid; | |
3592 | int err; | |
3593 | ||
3594 | err = kstrtoul(buf, 10, &count); | |
3595 | if (err) | |
3596 | return err; | |
3597 | ||
3598 | h = kobj_to_hstate(kobj, &nid); | |
3599 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
3600 | } | |
3601 | ||
06808b08 LS |
3602 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
3603 | struct kobj_attribute *attr, char *buf) | |
3604 | { | |
3605 | return nr_hugepages_show_common(kobj, attr, buf); | |
3606 | } | |
3607 | ||
3608 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
3609 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3610 | { | |
238d3c13 | 3611 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
3612 | } |
3613 | HSTATE_ATTR(nr_hugepages); | |
3614 | ||
06808b08 LS |
3615 | #ifdef CONFIG_NUMA |
3616 | ||
3617 | /* | |
3618 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
3619 | * huge page alloc/free. | |
3620 | */ | |
3621 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
3622 | struct kobj_attribute *attr, |
3623 | char *buf) | |
06808b08 LS |
3624 | { |
3625 | return nr_hugepages_show_common(kobj, attr, buf); | |
3626 | } | |
3627 | ||
3628 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
3629 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3630 | { | |
238d3c13 | 3631 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
3632 | } |
3633 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
3634 | #endif | |
3635 | ||
3636 | ||
a3437870 NA |
3637 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
3638 | struct kobj_attribute *attr, char *buf) | |
3639 | { | |
9a305230 | 3640 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3641 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 3642 | } |
adbe8726 | 3643 | |
a3437870 NA |
3644 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
3645 | struct kobj_attribute *attr, const char *buf, size_t count) | |
3646 | { | |
3647 | int err; | |
3648 | unsigned long input; | |
9a305230 | 3649 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 3650 | |
bae7f4ae | 3651 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
3652 | return -EINVAL; |
3653 | ||
3dbb95f7 | 3654 | err = kstrtoul(buf, 10, &input); |
a3437870 | 3655 | if (err) |
73ae31e5 | 3656 | return err; |
a3437870 | 3657 | |
db71ef79 | 3658 | spin_lock_irq(&hugetlb_lock); |
a3437870 | 3659 | h->nr_overcommit_huge_pages = input; |
db71ef79 | 3660 | spin_unlock_irq(&hugetlb_lock); |
a3437870 NA |
3661 | |
3662 | return count; | |
3663 | } | |
3664 | HSTATE_ATTR(nr_overcommit_hugepages); | |
3665 | ||
3666 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
3667 | struct kobj_attribute *attr, char *buf) | |
3668 | { | |
9a305230 LS |
3669 | struct hstate *h; |
3670 | unsigned long free_huge_pages; | |
3671 | int nid; | |
3672 | ||
3673 | h = kobj_to_hstate(kobj, &nid); | |
3674 | if (nid == NUMA_NO_NODE) | |
3675 | free_huge_pages = h->free_huge_pages; | |
3676 | else | |
3677 | free_huge_pages = h->free_huge_pages_node[nid]; | |
3678 | ||
ae7a927d | 3679 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
3680 | } |
3681 | HSTATE_ATTR_RO(free_hugepages); | |
3682 | ||
3683 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
3684 | struct kobj_attribute *attr, char *buf) | |
3685 | { | |
9a305230 | 3686 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3687 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
3688 | } |
3689 | HSTATE_ATTR_RO(resv_hugepages); | |
3690 | ||
3691 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
3692 | struct kobj_attribute *attr, char *buf) | |
3693 | { | |
9a305230 LS |
3694 | struct hstate *h; |
3695 | unsigned long surplus_huge_pages; | |
3696 | int nid; | |
3697 | ||
3698 | h = kobj_to_hstate(kobj, &nid); | |
3699 | if (nid == NUMA_NO_NODE) | |
3700 | surplus_huge_pages = h->surplus_huge_pages; | |
3701 | else | |
3702 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
3703 | ||
ae7a927d | 3704 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
3705 | } |
3706 | HSTATE_ATTR_RO(surplus_hugepages); | |
3707 | ||
79dfc695 MK |
3708 | static ssize_t demote_store(struct kobject *kobj, |
3709 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3710 | { | |
3711 | unsigned long nr_demote; | |
3712 | unsigned long nr_available; | |
3713 | nodemask_t nodes_allowed, *n_mask; | |
3714 | struct hstate *h; | |
8eeda55f | 3715 | int err; |
79dfc695 MK |
3716 | int nid; |
3717 | ||
3718 | err = kstrtoul(buf, 10, &nr_demote); | |
3719 | if (err) | |
3720 | return err; | |
3721 | h = kobj_to_hstate(kobj, &nid); | |
3722 | ||
3723 | if (nid != NUMA_NO_NODE) { | |
3724 | init_nodemask_of_node(&nodes_allowed, nid); | |
3725 | n_mask = &nodes_allowed; | |
3726 | } else { | |
3727 | n_mask = &node_states[N_MEMORY]; | |
3728 | } | |
3729 | ||
3730 | /* Synchronize with other sysfs operations modifying huge pages */ | |
3731 | mutex_lock(&h->resize_lock); | |
3732 | spin_lock_irq(&hugetlb_lock); | |
3733 | ||
3734 | while (nr_demote) { | |
3735 | /* | |
3736 | * Check for available pages to demote each time thorough the | |
3737 | * loop as demote_pool_huge_page will drop hugetlb_lock. | |
79dfc695 MK |
3738 | */ |
3739 | if (nid != NUMA_NO_NODE) | |
3740 | nr_available = h->free_huge_pages_node[nid]; | |
3741 | else | |
3742 | nr_available = h->free_huge_pages; | |
3743 | nr_available -= h->resv_huge_pages; | |
3744 | if (!nr_available) | |
3745 | break; | |
3746 | ||
3747 | err = demote_pool_huge_page(h, n_mask); | |
3748 | if (err) | |
3749 | break; | |
3750 | ||
3751 | nr_demote--; | |
3752 | } | |
3753 | ||
3754 | spin_unlock_irq(&hugetlb_lock); | |
3755 | mutex_unlock(&h->resize_lock); | |
3756 | ||
3757 | if (err) | |
3758 | return err; | |
3759 | return len; | |
3760 | } | |
3761 | HSTATE_ATTR_WO(demote); | |
3762 | ||
3763 | static ssize_t demote_size_show(struct kobject *kobj, | |
3764 | struct kobj_attribute *attr, char *buf) | |
3765 | { | |
12658abf | 3766 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
79dfc695 MK |
3767 | unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; |
3768 | ||
3769 | return sysfs_emit(buf, "%lukB\n", demote_size); | |
3770 | } | |
3771 | ||
3772 | static ssize_t demote_size_store(struct kobject *kobj, | |
3773 | struct kobj_attribute *attr, | |
3774 | const char *buf, size_t count) | |
3775 | { | |
3776 | struct hstate *h, *demote_hstate; | |
3777 | unsigned long demote_size; | |
3778 | unsigned int demote_order; | |
79dfc695 MK |
3779 | |
3780 | demote_size = (unsigned long)memparse(buf, NULL); | |
3781 | ||
3782 | demote_hstate = size_to_hstate(demote_size); | |
3783 | if (!demote_hstate) | |
3784 | return -EINVAL; | |
3785 | demote_order = demote_hstate->order; | |
a01f4390 MK |
3786 | if (demote_order < HUGETLB_PAGE_ORDER) |
3787 | return -EINVAL; | |
79dfc695 MK |
3788 | |
3789 | /* demote order must be smaller than hstate order */ | |
12658abf | 3790 | h = kobj_to_hstate(kobj, NULL); |
79dfc695 MK |
3791 | if (demote_order >= h->order) |
3792 | return -EINVAL; | |
3793 | ||
3794 | /* resize_lock synchronizes access to demote size and writes */ | |
3795 | mutex_lock(&h->resize_lock); | |
3796 | h->demote_order = demote_order; | |
3797 | mutex_unlock(&h->resize_lock); | |
3798 | ||
3799 | return count; | |
3800 | } | |
3801 | HSTATE_ATTR(demote_size); | |
3802 | ||
a3437870 NA |
3803 | static struct attribute *hstate_attrs[] = { |
3804 | &nr_hugepages_attr.attr, | |
3805 | &nr_overcommit_hugepages_attr.attr, | |
3806 | &free_hugepages_attr.attr, | |
3807 | &resv_hugepages_attr.attr, | |
3808 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3809 | #ifdef CONFIG_NUMA |
3810 | &nr_hugepages_mempolicy_attr.attr, | |
3811 | #endif | |
a3437870 NA |
3812 | NULL, |
3813 | }; | |
3814 | ||
67e5ed96 | 3815 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3816 | .attrs = hstate_attrs, |
3817 | }; | |
3818 | ||
79dfc695 MK |
3819 | static struct attribute *hstate_demote_attrs[] = { |
3820 | &demote_size_attr.attr, | |
3821 | &demote_attr.attr, | |
3822 | NULL, | |
3823 | }; | |
3824 | ||
3825 | static const struct attribute_group hstate_demote_attr_group = { | |
3826 | .attrs = hstate_demote_attrs, | |
3827 | }; | |
3828 | ||
094e9539 JM |
3829 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3830 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3831 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3832 | { |
3833 | int retval; | |
972dc4de | 3834 | int hi = hstate_index(h); |
a3437870 | 3835 | |
9a305230 LS |
3836 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3837 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3838 | return -ENOMEM; |
3839 | ||
9a305230 | 3840 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 3841 | if (retval) { |
9a305230 | 3842 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 | 3843 | hstate_kobjs[hi] = NULL; |
3a6bdda0 | 3844 | return retval; |
cc2205a6 | 3845 | } |
a3437870 | 3846 | |
79dfc695 | 3847 | if (h->demote_order) { |
01088a60 ML |
3848 | retval = sysfs_create_group(hstate_kobjs[hi], |
3849 | &hstate_demote_attr_group); | |
3850 | if (retval) { | |
79dfc695 | 3851 | pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); |
01088a60 ML |
3852 | sysfs_remove_group(hstate_kobjs[hi], hstate_attr_group); |
3853 | kobject_put(hstate_kobjs[hi]); | |
3854 | hstate_kobjs[hi] = NULL; | |
3855 | return retval; | |
3856 | } | |
79dfc695 MK |
3857 | } |
3858 | ||
01088a60 | 3859 | return 0; |
a3437870 NA |
3860 | } |
3861 | ||
9a305230 | 3862 | #ifdef CONFIG_NUMA |
a4a00b45 | 3863 | static bool hugetlb_sysfs_initialized __ro_after_init; |
9a305230 LS |
3864 | |
3865 | /* | |
3866 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3867 | * with node devices in node_devices[] using a parallel array. The array |
3868 | * index of a node device or _hstate == node id. | |
3869 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3870 | * the base kernel, on the hugetlb module. |
3871 | */ | |
3872 | struct node_hstate { | |
3873 | struct kobject *hugepages_kobj; | |
3874 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3875 | }; | |
b4e289a6 | 3876 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3877 | |
3878 | /* | |
10fbcf4c | 3879 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3880 | */ |
3881 | static struct attribute *per_node_hstate_attrs[] = { | |
3882 | &nr_hugepages_attr.attr, | |
3883 | &free_hugepages_attr.attr, | |
3884 | &surplus_hugepages_attr.attr, | |
3885 | NULL, | |
3886 | }; | |
3887 | ||
67e5ed96 | 3888 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3889 | .attrs = per_node_hstate_attrs, |
3890 | }; | |
3891 | ||
3892 | /* | |
10fbcf4c | 3893 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3894 | * Returns node id via non-NULL nidp. |
3895 | */ | |
3896 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3897 | { | |
3898 | int nid; | |
3899 | ||
3900 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3901 | struct node_hstate *nhs = &node_hstates[nid]; | |
3902 | int i; | |
3903 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3904 | if (nhs->hstate_kobjs[i] == kobj) { | |
3905 | if (nidp) | |
3906 | *nidp = nid; | |
3907 | return &hstates[i]; | |
3908 | } | |
3909 | } | |
3910 | ||
3911 | BUG(); | |
3912 | return NULL; | |
3913 | } | |
3914 | ||
3915 | /* | |
10fbcf4c | 3916 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3917 | * No-op if no hstate attributes attached. |
3918 | */ | |
a4a00b45 | 3919 | void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3920 | { |
3921 | struct hstate *h; | |
10fbcf4c | 3922 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3923 | |
3924 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3925 | return; /* no hstate attributes */ |
9a305230 | 3926 | |
972dc4de AK |
3927 | for_each_hstate(h) { |
3928 | int idx = hstate_index(h); | |
01088a60 ML |
3929 | struct kobject *hstate_kobj = nhs->hstate_kobjs[idx]; |
3930 | ||
3931 | if (!hstate_kobj) | |
3932 | continue; | |
3933 | if (h->demote_order) | |
3934 | sysfs_remove_group(hstate_kobj, &hstate_demote_attr_group); | |
3935 | sysfs_remove_group(hstate_kobj, &per_node_hstate_attr_group); | |
3936 | kobject_put(hstate_kobj); | |
3937 | nhs->hstate_kobjs[idx] = NULL; | |
972dc4de | 3938 | } |
9a305230 LS |
3939 | |
3940 | kobject_put(nhs->hugepages_kobj); | |
3941 | nhs->hugepages_kobj = NULL; | |
3942 | } | |
3943 | ||
9a305230 LS |
3944 | |
3945 | /* | |
10fbcf4c | 3946 | * Register hstate attributes for a single node device. |
9a305230 LS |
3947 | * No-op if attributes already registered. |
3948 | */ | |
a4a00b45 | 3949 | void hugetlb_register_node(struct node *node) |
9a305230 LS |
3950 | { |
3951 | struct hstate *h; | |
10fbcf4c | 3952 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3953 | int err; |
3954 | ||
a4a00b45 MS |
3955 | if (!hugetlb_sysfs_initialized) |
3956 | return; | |
3957 | ||
9a305230 LS |
3958 | if (nhs->hugepages_kobj) |
3959 | return; /* already allocated */ | |
3960 | ||
3961 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3962 | &node->dev.kobj); |
9a305230 LS |
3963 | if (!nhs->hugepages_kobj) |
3964 | return; | |
3965 | ||
3966 | for_each_hstate(h) { | |
3967 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3968 | nhs->hstate_kobjs, | |
3969 | &per_node_hstate_attr_group); | |
3970 | if (err) { | |
282f4214 | 3971 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3972 | h->name, node->dev.id); |
9a305230 LS |
3973 | hugetlb_unregister_node(node); |
3974 | break; | |
3975 | } | |
3976 | } | |
3977 | } | |
3978 | ||
3979 | /* | |
9b5e5d0f | 3980 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3981 | * devices of nodes that have memory. All on-line nodes should have |
3982 | * registered their associated device by this time. | |
9a305230 | 3983 | */ |
7d9ca000 | 3984 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3985 | { |
3986 | int nid; | |
3987 | ||
a4a00b45 | 3988 | for_each_online_node(nid) |
b958d4d0 | 3989 | hugetlb_register_node(node_devices[nid]); |
9a305230 LS |
3990 | } |
3991 | #else /* !CONFIG_NUMA */ | |
3992 | ||
3993 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3994 | { | |
3995 | BUG(); | |
3996 | if (nidp) | |
3997 | *nidp = -1; | |
3998 | return NULL; | |
3999 | } | |
4000 | ||
9a305230 LS |
4001 | static void hugetlb_register_all_nodes(void) { } |
4002 | ||
4003 | #endif | |
4004 | ||
263b8998 ML |
4005 | #ifdef CONFIG_CMA |
4006 | static void __init hugetlb_cma_check(void); | |
4007 | #else | |
4008 | static inline __init void hugetlb_cma_check(void) | |
4009 | { | |
4010 | } | |
4011 | #endif | |
4012 | ||
a4a00b45 MS |
4013 | static void __init hugetlb_sysfs_init(void) |
4014 | { | |
4015 | struct hstate *h; | |
4016 | int err; | |
4017 | ||
4018 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
4019 | if (!hugepages_kobj) | |
4020 | return; | |
4021 | ||
4022 | for_each_hstate(h) { | |
4023 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, | |
4024 | hstate_kobjs, &hstate_attr_group); | |
4025 | if (err) | |
4026 | pr_err("HugeTLB: Unable to add hstate %s", h->name); | |
4027 | } | |
4028 | ||
4029 | #ifdef CONFIG_NUMA | |
4030 | hugetlb_sysfs_initialized = true; | |
4031 | #endif | |
4032 | hugetlb_register_all_nodes(); | |
4033 | } | |
4034 | ||
a3437870 NA |
4035 | static int __init hugetlb_init(void) |
4036 | { | |
8382d914 DB |
4037 | int i; |
4038 | ||
d6995da3 MK |
4039 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
4040 | __NR_HPAGEFLAGS); | |
4041 | ||
c2833a5b MK |
4042 | if (!hugepages_supported()) { |
4043 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
4044 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 4045 | return 0; |
c2833a5b | 4046 | } |
a3437870 | 4047 | |
282f4214 MK |
4048 | /* |
4049 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
4050 | * architectures depend on setup being done here. | |
4051 | */ | |
4052 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
4053 | if (!parsed_default_hugepagesz) { | |
4054 | /* | |
4055 | * If we did not parse a default huge page size, set | |
4056 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
4057 | * number of huge pages for this default size was implicitly | |
4058 | * specified, set that here as well. | |
4059 | * Note that the implicit setting will overwrite an explicit | |
4060 | * setting. A warning will be printed in this case. | |
4061 | */ | |
4062 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
4063 | if (default_hstate_max_huge_pages) { | |
4064 | if (default_hstate.max_huge_pages) { | |
4065 | char buf[32]; | |
4066 | ||
4067 | string_get_size(huge_page_size(&default_hstate), | |
4068 | 1, STRING_UNITS_2, buf, 32); | |
4069 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
4070 | default_hstate.max_huge_pages, buf); | |
4071 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
4072 | default_hstate_max_huge_pages); | |
4073 | } | |
4074 | default_hstate.max_huge_pages = | |
4075 | default_hstate_max_huge_pages; | |
b5389086 | 4076 | |
0a7a0f6f | 4077 | for_each_online_node(i) |
b5389086 ZY |
4078 | default_hstate.max_huge_pages_node[i] = |
4079 | default_hugepages_in_node[i]; | |
d715cf80 | 4080 | } |
f8b74815 | 4081 | } |
a3437870 | 4082 | |
cf11e85f | 4083 | hugetlb_cma_check(); |
a3437870 | 4084 | hugetlb_init_hstates(); |
aa888a74 | 4085 | gather_bootmem_prealloc(); |
a3437870 NA |
4086 | report_hugepages(); |
4087 | ||
4088 | hugetlb_sysfs_init(); | |
7179e7bf | 4089 | hugetlb_cgroup_file_init(); |
9a305230 | 4090 | |
8382d914 DB |
4091 | #ifdef CONFIG_SMP |
4092 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
4093 | #else | |
4094 | num_fault_mutexes = 1; | |
4095 | #endif | |
c672c7f2 | 4096 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
4097 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
4098 | GFP_KERNEL); | |
c672c7f2 | 4099 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
4100 | |
4101 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 4102 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
4103 | return 0; |
4104 | } | |
3e89e1c5 | 4105 | subsys_initcall(hugetlb_init); |
a3437870 | 4106 | |
ae94da89 MK |
4107 | /* Overwritten by architectures with more huge page sizes */ |
4108 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 4109 | { |
ae94da89 | 4110 | return size == HPAGE_SIZE; |
9fee021d VT |
4111 | } |
4112 | ||
d00181b9 | 4113 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
4114 | { |
4115 | struct hstate *h; | |
8faa8b07 AK |
4116 | unsigned long i; |
4117 | ||
a3437870 | 4118 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
4119 | return; |
4120 | } | |
47d38344 | 4121 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 4122 | BUG_ON(order == 0); |
47d38344 | 4123 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 4124 | mutex_init(&h->resize_lock); |
a3437870 | 4125 | h->order = order; |
aca78307 | 4126 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
4127 | for (i = 0; i < MAX_NUMNODES; ++i) |
4128 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 4129 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
4130 | h->next_nid_to_alloc = first_memory_node; |
4131 | h->next_nid_to_free = first_memory_node; | |
a3437870 | 4132 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
c2c3a60a | 4133 | huge_page_size(h)/SZ_1K); |
8faa8b07 | 4134 | |
a3437870 NA |
4135 | parsed_hstate = h; |
4136 | } | |
4137 | ||
b5389086 ZY |
4138 | bool __init __weak hugetlb_node_alloc_supported(void) |
4139 | { | |
4140 | return true; | |
4141 | } | |
f87442f4 PL |
4142 | |
4143 | static void __init hugepages_clear_pages_in_node(void) | |
4144 | { | |
4145 | if (!hugetlb_max_hstate) { | |
4146 | default_hstate_max_huge_pages = 0; | |
4147 | memset(default_hugepages_in_node, 0, | |
10395680 | 4148 | sizeof(default_hugepages_in_node)); |
f87442f4 PL |
4149 | } else { |
4150 | parsed_hstate->max_huge_pages = 0; | |
4151 | memset(parsed_hstate->max_huge_pages_node, 0, | |
10395680 | 4152 | sizeof(parsed_hstate->max_huge_pages_node)); |
f87442f4 PL |
4153 | } |
4154 | } | |
4155 | ||
282f4214 MK |
4156 | /* |
4157 | * hugepages command line processing | |
4158 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
4159 | * specification. If not, ignore the hugepages value. hugepages can also | |
4160 | * be the first huge page command line option in which case it implicitly | |
4161 | * specifies the number of huge pages for the default size. | |
4162 | */ | |
4163 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
4164 | { |
4165 | unsigned long *mhp; | |
8faa8b07 | 4166 | static unsigned long *last_mhp; |
b5389086 ZY |
4167 | int node = NUMA_NO_NODE; |
4168 | int count; | |
4169 | unsigned long tmp; | |
4170 | char *p = s; | |
a3437870 | 4171 | |
9fee021d | 4172 | if (!parsed_valid_hugepagesz) { |
282f4214 | 4173 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 4174 | parsed_valid_hugepagesz = true; |
f81f6e4b | 4175 | return 1; |
9fee021d | 4176 | } |
282f4214 | 4177 | |
a3437870 | 4178 | /* |
282f4214 MK |
4179 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
4180 | * yet, so this hugepages= parameter goes to the "default hstate". | |
4181 | * Otherwise, it goes with the previously parsed hugepagesz or | |
4182 | * default_hugepagesz. | |
a3437870 | 4183 | */ |
9fee021d | 4184 | else if (!hugetlb_max_hstate) |
a3437870 NA |
4185 | mhp = &default_hstate_max_huge_pages; |
4186 | else | |
4187 | mhp = &parsed_hstate->max_huge_pages; | |
4188 | ||
8faa8b07 | 4189 | if (mhp == last_mhp) { |
282f4214 | 4190 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
f81f6e4b | 4191 | return 1; |
8faa8b07 AK |
4192 | } |
4193 | ||
b5389086 ZY |
4194 | while (*p) { |
4195 | count = 0; | |
4196 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4197 | goto invalid; | |
4198 | /* Parameter is node format */ | |
4199 | if (p[count] == ':') { | |
4200 | if (!hugetlb_node_alloc_supported()) { | |
4201 | pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); | |
f81f6e4b | 4202 | return 1; |
b5389086 | 4203 | } |
0a7a0f6f | 4204 | if (tmp >= MAX_NUMNODES || !node_online(tmp)) |
e79ce983 | 4205 | goto invalid; |
0a7a0f6f | 4206 | node = array_index_nospec(tmp, MAX_NUMNODES); |
b5389086 | 4207 | p += count + 1; |
b5389086 ZY |
4208 | /* Parse hugepages */ |
4209 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4210 | goto invalid; | |
4211 | if (!hugetlb_max_hstate) | |
4212 | default_hugepages_in_node[node] = tmp; | |
4213 | else | |
4214 | parsed_hstate->max_huge_pages_node[node] = tmp; | |
4215 | *mhp += tmp; | |
4216 | /* Go to parse next node*/ | |
4217 | if (p[count] == ',') | |
4218 | p += count + 1; | |
4219 | else | |
4220 | break; | |
4221 | } else { | |
4222 | if (p != s) | |
4223 | goto invalid; | |
4224 | *mhp = tmp; | |
4225 | break; | |
4226 | } | |
4227 | } | |
a3437870 | 4228 | |
8faa8b07 AK |
4229 | /* |
4230 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 4231 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
4232 | * use the bootmem allocator. |
4233 | */ | |
04adbc3f | 4234 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
4235 | hugetlb_hstate_alloc_pages(parsed_hstate); |
4236 | ||
4237 | last_mhp = mhp; | |
4238 | ||
a3437870 | 4239 | return 1; |
b5389086 ZY |
4240 | |
4241 | invalid: | |
4242 | pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); | |
f87442f4 | 4243 | hugepages_clear_pages_in_node(); |
f81f6e4b | 4244 | return 1; |
a3437870 | 4245 | } |
282f4214 | 4246 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 4247 | |
282f4214 MK |
4248 | /* |
4249 | * hugepagesz command line processing | |
4250 | * A specific huge page size can only be specified once with hugepagesz. | |
4251 | * hugepagesz is followed by hugepages on the command line. The global | |
4252 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
4253 | * hugepagesz argument was valid. | |
4254 | */ | |
359f2544 | 4255 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 4256 | { |
359f2544 | 4257 | unsigned long size; |
282f4214 MK |
4258 | struct hstate *h; |
4259 | ||
4260 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
4261 | size = (unsigned long)memparse(s, NULL); |
4262 | ||
4263 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4264 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
f81f6e4b | 4265 | return 1; |
359f2544 MK |
4266 | } |
4267 | ||
282f4214 MK |
4268 | h = size_to_hstate(size); |
4269 | if (h) { | |
4270 | /* | |
4271 | * hstate for this size already exists. This is normally | |
4272 | * an error, but is allowed if the existing hstate is the | |
4273 | * default hstate. More specifically, it is only allowed if | |
4274 | * the number of huge pages for the default hstate was not | |
4275 | * previously specified. | |
4276 | */ | |
4277 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
4278 | default_hstate.max_huge_pages) { | |
4279 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
f81f6e4b | 4280 | return 1; |
282f4214 MK |
4281 | } |
4282 | ||
4283 | /* | |
4284 | * No need to call hugetlb_add_hstate() as hstate already | |
4285 | * exists. But, do set parsed_hstate so that a following | |
4286 | * hugepages= parameter will be applied to this hstate. | |
4287 | */ | |
4288 | parsed_hstate = h; | |
4289 | parsed_valid_hugepagesz = true; | |
4290 | return 1; | |
38237830 MK |
4291 | } |
4292 | ||
359f2544 | 4293 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 4294 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
4295 | return 1; |
4296 | } | |
359f2544 MK |
4297 | __setup("hugepagesz=", hugepagesz_setup); |
4298 | ||
282f4214 MK |
4299 | /* |
4300 | * default_hugepagesz command line input | |
4301 | * Only one instance of default_hugepagesz allowed on command line. | |
4302 | */ | |
ae94da89 | 4303 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 4304 | { |
ae94da89 | 4305 | unsigned long size; |
b5389086 | 4306 | int i; |
ae94da89 | 4307 | |
282f4214 | 4308 | parsed_valid_hugepagesz = false; |
282f4214 MK |
4309 | if (parsed_default_hugepagesz) { |
4310 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
f81f6e4b | 4311 | return 1; |
282f4214 MK |
4312 | } |
4313 | ||
ae94da89 MK |
4314 | size = (unsigned long)memparse(s, NULL); |
4315 | ||
4316 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4317 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
f81f6e4b | 4318 | return 1; |
ae94da89 MK |
4319 | } |
4320 | ||
282f4214 MK |
4321 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
4322 | parsed_valid_hugepagesz = true; | |
4323 | parsed_default_hugepagesz = true; | |
4324 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
4325 | ||
4326 | /* | |
4327 | * The number of default huge pages (for this size) could have been | |
4328 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
4329 | * then default_hstate_max_huge_pages is set. If the default huge | |
4330 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
4331 | * allocated here from bootmem allocator. | |
4332 | */ | |
4333 | if (default_hstate_max_huge_pages) { | |
4334 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
0a7a0f6f | 4335 | for_each_online_node(i) |
b5389086 ZY |
4336 | default_hstate.max_huge_pages_node[i] = |
4337 | default_hugepages_in_node[i]; | |
282f4214 MK |
4338 | if (hstate_is_gigantic(&default_hstate)) |
4339 | hugetlb_hstate_alloc_pages(&default_hstate); | |
4340 | default_hstate_max_huge_pages = 0; | |
4341 | } | |
4342 | ||
e11bfbfc NP |
4343 | return 1; |
4344 | } | |
ae94da89 | 4345 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 4346 | |
d2226ebd FT |
4347 | static nodemask_t *policy_mbind_nodemask(gfp_t gfp) |
4348 | { | |
4349 | #ifdef CONFIG_NUMA | |
4350 | struct mempolicy *mpol = get_task_policy(current); | |
4351 | ||
4352 | /* | |
4353 | * Only enforce MPOL_BIND policy which overlaps with cpuset policy | |
4354 | * (from policy_nodemask) specifically for hugetlb case | |
4355 | */ | |
4356 | if (mpol->mode == MPOL_BIND && | |
4357 | (apply_policy_zone(mpol, gfp_zone(gfp)) && | |
4358 | cpuset_nodemask_valid_mems_allowed(&mpol->nodes))) | |
4359 | return &mpol->nodes; | |
4360 | #endif | |
4361 | return NULL; | |
4362 | } | |
4363 | ||
8ca39e68 | 4364 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
4365 | { |
4366 | int node; | |
4367 | unsigned int nr = 0; | |
d2226ebd | 4368 | nodemask_t *mbind_nodemask; |
8ca39e68 MS |
4369 | unsigned int *array = h->free_huge_pages_node; |
4370 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
4371 | ||
d2226ebd | 4372 | mbind_nodemask = policy_mbind_nodemask(gfp_mask); |
8ca39e68 | 4373 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
d2226ebd | 4374 | if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) |
8ca39e68 MS |
4375 | nr += array[node]; |
4376 | } | |
8a213460 NA |
4377 | |
4378 | return nr; | |
4379 | } | |
4380 | ||
4381 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
4382 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
4383 | void *buffer, size_t *length, | |
4384 | loff_t *ppos, unsigned long *out) | |
4385 | { | |
4386 | struct ctl_table dup_table; | |
4387 | ||
4388 | /* | |
4389 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
4390 | * can duplicate the @table and alter the duplicate of it. | |
4391 | */ | |
4392 | dup_table = *table; | |
4393 | dup_table.data = out; | |
4394 | ||
4395 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
4396 | } | |
4397 | ||
06808b08 LS |
4398 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
4399 | struct ctl_table *table, int write, | |
32927393 | 4400 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 4401 | { |
e5ff2159 | 4402 | struct hstate *h = &default_hstate; |
238d3c13 | 4403 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 4404 | int ret; |
e5ff2159 | 4405 | |
457c1b27 | 4406 | if (!hugepages_supported()) |
86613628 | 4407 | return -EOPNOTSUPP; |
457c1b27 | 4408 | |
17743798 MS |
4409 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4410 | &tmp); | |
08d4a246 MH |
4411 | if (ret) |
4412 | goto out; | |
e5ff2159 | 4413 | |
238d3c13 DR |
4414 | if (write) |
4415 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
4416 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
4417 | out: |
4418 | return ret; | |
1da177e4 | 4419 | } |
396faf03 | 4420 | |
06808b08 | 4421 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 4422 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4423 | { |
4424 | ||
4425 | return hugetlb_sysctl_handler_common(false, table, write, | |
4426 | buffer, length, ppos); | |
4427 | } | |
4428 | ||
4429 | #ifdef CONFIG_NUMA | |
4430 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 4431 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4432 | { |
4433 | return hugetlb_sysctl_handler_common(true, table, write, | |
4434 | buffer, length, ppos); | |
4435 | } | |
4436 | #endif /* CONFIG_NUMA */ | |
4437 | ||
a3d0c6aa | 4438 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 4439 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 4440 | { |
a5516438 | 4441 | struct hstate *h = &default_hstate; |
e5ff2159 | 4442 | unsigned long tmp; |
08d4a246 | 4443 | int ret; |
e5ff2159 | 4444 | |
457c1b27 | 4445 | if (!hugepages_supported()) |
86613628 | 4446 | return -EOPNOTSUPP; |
457c1b27 | 4447 | |
c033a93c | 4448 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 4449 | |
bae7f4ae | 4450 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
4451 | return -EINVAL; |
4452 | ||
17743798 MS |
4453 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4454 | &tmp); | |
08d4a246 MH |
4455 | if (ret) |
4456 | goto out; | |
e5ff2159 AK |
4457 | |
4458 | if (write) { | |
db71ef79 | 4459 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 4460 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 4461 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 4462 | } |
08d4a246 MH |
4463 | out: |
4464 | return ret; | |
a3d0c6aa NA |
4465 | } |
4466 | ||
1da177e4 LT |
4467 | #endif /* CONFIG_SYSCTL */ |
4468 | ||
e1759c21 | 4469 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 4470 | { |
fcb2b0c5 RG |
4471 | struct hstate *h; |
4472 | unsigned long total = 0; | |
4473 | ||
457c1b27 NA |
4474 | if (!hugepages_supported()) |
4475 | return; | |
fcb2b0c5 RG |
4476 | |
4477 | for_each_hstate(h) { | |
4478 | unsigned long count = h->nr_huge_pages; | |
4479 | ||
aca78307 | 4480 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
4481 | |
4482 | if (h == &default_hstate) | |
4483 | seq_printf(m, | |
4484 | "HugePages_Total: %5lu\n" | |
4485 | "HugePages_Free: %5lu\n" | |
4486 | "HugePages_Rsvd: %5lu\n" | |
4487 | "HugePages_Surp: %5lu\n" | |
4488 | "Hugepagesize: %8lu kB\n", | |
4489 | count, | |
4490 | h->free_huge_pages, | |
4491 | h->resv_huge_pages, | |
4492 | h->surplus_huge_pages, | |
aca78307 | 4493 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
4494 | } |
4495 | ||
aca78307 | 4496 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
4497 | } |
4498 | ||
7981593b | 4499 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 4500 | { |
a5516438 | 4501 | struct hstate *h = &default_hstate; |
7981593b | 4502 | |
457c1b27 NA |
4503 | if (!hugepages_supported()) |
4504 | return 0; | |
7981593b JP |
4505 | |
4506 | return sysfs_emit_at(buf, len, | |
4507 | "Node %d HugePages_Total: %5u\n" | |
4508 | "Node %d HugePages_Free: %5u\n" | |
4509 | "Node %d HugePages_Surp: %5u\n", | |
4510 | nid, h->nr_huge_pages_node[nid], | |
4511 | nid, h->free_huge_pages_node[nid], | |
4512 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
4513 | } |
4514 | ||
dcadcf1c | 4515 | void hugetlb_show_meminfo_node(int nid) |
949f7ec5 DR |
4516 | { |
4517 | struct hstate *h; | |
949f7ec5 | 4518 | |
457c1b27 NA |
4519 | if (!hugepages_supported()) |
4520 | return; | |
4521 | ||
dcadcf1c GL |
4522 | for_each_hstate(h) |
4523 | printk("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
4524 | nid, | |
4525 | h->nr_huge_pages_node[nid], | |
4526 | h->free_huge_pages_node[nid], | |
4527 | h->surplus_huge_pages_node[nid], | |
4528 | huge_page_size(h) / SZ_1K); | |
949f7ec5 DR |
4529 | } |
4530 | ||
5d317b2b NH |
4531 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
4532 | { | |
4533 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
4534 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
4535 | } | |
4536 | ||
1da177e4 LT |
4537 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
4538 | unsigned long hugetlb_total_pages(void) | |
4539 | { | |
d0028588 WL |
4540 | struct hstate *h; |
4541 | unsigned long nr_total_pages = 0; | |
4542 | ||
4543 | for_each_hstate(h) | |
4544 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
4545 | return nr_total_pages; | |
1da177e4 | 4546 | } |
1da177e4 | 4547 | |
a5516438 | 4548 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
4549 | { |
4550 | int ret = -ENOMEM; | |
4551 | ||
0aa7f354 ML |
4552 | if (!delta) |
4553 | return 0; | |
4554 | ||
db71ef79 | 4555 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4556 | /* |
4557 | * When cpuset is configured, it breaks the strict hugetlb page | |
4558 | * reservation as the accounting is done on a global variable. Such | |
4559 | * reservation is completely rubbish in the presence of cpuset because | |
4560 | * the reservation is not checked against page availability for the | |
4561 | * current cpuset. Application can still potentially OOM'ed by kernel | |
4562 | * with lack of free htlb page in cpuset that the task is in. | |
4563 | * Attempt to enforce strict accounting with cpuset is almost | |
4564 | * impossible (or too ugly) because cpuset is too fluid that | |
4565 | * task or memory node can be dynamically moved between cpusets. | |
4566 | * | |
4567 | * The change of semantics for shared hugetlb mapping with cpuset is | |
4568 | * undesirable. However, in order to preserve some of the semantics, | |
4569 | * we fall back to check against current free page availability as | |
4570 | * a best attempt and hopefully to minimize the impact of changing | |
4571 | * semantics that cpuset has. | |
8ca39e68 MS |
4572 | * |
4573 | * Apart from cpuset, we also have memory policy mechanism that | |
4574 | * also determines from which node the kernel will allocate memory | |
4575 | * in a NUMA system. So similar to cpuset, we also should consider | |
4576 | * the memory policy of the current task. Similar to the description | |
4577 | * above. | |
fc1b8a73 MG |
4578 | */ |
4579 | if (delta > 0) { | |
a5516438 | 4580 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
4581 | goto out; |
4582 | ||
8ca39e68 | 4583 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 4584 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
4585 | goto out; |
4586 | } | |
4587 | } | |
4588 | ||
4589 | ret = 0; | |
4590 | if (delta < 0) | |
a5516438 | 4591 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
4592 | |
4593 | out: | |
db71ef79 | 4594 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4595 | return ret; |
4596 | } | |
4597 | ||
84afd99b AW |
4598 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
4599 | { | |
f522c3ac | 4600 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
4601 | |
4602 | /* | |
4603 | * This new VMA should share its siblings reservation map if present. | |
4604 | * The VMA will only ever have a valid reservation map pointer where | |
4605 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 4606 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
4607 | * after this open call completes. It is therefore safe to take a |
4608 | * new reference here without additional locking. | |
4609 | */ | |
09a26e83 MK |
4610 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
4611 | resv_map_dup_hugetlb_cgroup_uncharge_info(resv); | |
f522c3ac | 4612 | kref_get(&resv->refs); |
09a26e83 | 4613 | } |
8d9bfb26 MK |
4614 | |
4615 | hugetlb_vma_lock_alloc(vma); | |
84afd99b AW |
4616 | } |
4617 | ||
a1e78772 MG |
4618 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
4619 | { | |
a5516438 | 4620 | struct hstate *h = hstate_vma(vma); |
8d9bfb26 | 4621 | struct resv_map *resv; |
90481622 | 4622 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 4623 | unsigned long reserve, start, end; |
1c5ecae3 | 4624 | long gbl_reserve; |
84afd99b | 4625 | |
8d9bfb26 MK |
4626 | hugetlb_vma_lock_free(vma); |
4627 | ||
4628 | resv = vma_resv_map(vma); | |
4e35f483 JK |
4629 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4630 | return; | |
84afd99b | 4631 | |
4e35f483 JK |
4632 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
4633 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 4634 | |
4e35f483 | 4635 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 4636 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 4637 | if (reserve) { |
1c5ecae3 MK |
4638 | /* |
4639 | * Decrement reserve counts. The global reserve count may be | |
4640 | * adjusted if the subpool has a minimum size. | |
4641 | */ | |
4642 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
4643 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 4644 | } |
e9fe92ae MA |
4645 | |
4646 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
4647 | } |
4648 | ||
31383c68 DW |
4649 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
4650 | { | |
4651 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
4652 | return -EINVAL; | |
4653 | return 0; | |
4654 | } | |
4655 | ||
05ea8860 DW |
4656 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
4657 | { | |
aca78307 | 4658 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
4659 | } |
4660 | ||
1da177e4 LT |
4661 | /* |
4662 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
4663 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 4664 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
4665 | * this far. |
4666 | */ | |
b3ec9f33 | 4667 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
4668 | { |
4669 | BUG(); | |
d0217ac0 | 4670 | return 0; |
1da177e4 LT |
4671 | } |
4672 | ||
eec3636a JC |
4673 | /* |
4674 | * When a new function is introduced to vm_operations_struct and added | |
4675 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
4676 | * This is because under System V memory model, mappings created via | |
4677 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
4678 | * their original vm_ops are overwritten with shm_vm_ops. | |
4679 | */ | |
f0f37e2f | 4680 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 4681 | .fault = hugetlb_vm_op_fault, |
84afd99b | 4682 | .open = hugetlb_vm_op_open, |
a1e78772 | 4683 | .close = hugetlb_vm_op_close, |
dd3b614f | 4684 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 4685 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
4686 | }; |
4687 | ||
1e8f889b DG |
4688 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
4689 | int writable) | |
63551ae0 DG |
4690 | { |
4691 | pte_t entry; | |
79c1c594 | 4692 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
63551ae0 | 4693 | |
1e8f889b | 4694 | if (writable) { |
106c992a GS |
4695 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
4696 | vma->vm_page_prot))); | |
63551ae0 | 4697 | } else { |
106c992a GS |
4698 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
4699 | vma->vm_page_prot)); | |
63551ae0 DG |
4700 | } |
4701 | entry = pte_mkyoung(entry); | |
79c1c594 | 4702 | entry = arch_make_huge_pte(entry, shift, vma->vm_flags); |
63551ae0 DG |
4703 | |
4704 | return entry; | |
4705 | } | |
4706 | ||
1e8f889b DG |
4707 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
4708 | unsigned long address, pte_t *ptep) | |
4709 | { | |
4710 | pte_t entry; | |
4711 | ||
106c992a | 4712 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 4713 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 4714 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
4715 | } |
4716 | ||
d5ed7444 | 4717 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
4718 | { |
4719 | swp_entry_t swp; | |
4720 | ||
4721 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 4722 | return false; |
4a705fef | 4723 | swp = pte_to_swp_entry(pte); |
d79d176a | 4724 | if (is_migration_entry(swp)) |
d5ed7444 | 4725 | return true; |
4a705fef | 4726 | else |
d5ed7444 | 4727 | return false; |
4a705fef NH |
4728 | } |
4729 | ||
3e5c3600 | 4730 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
4731 | { |
4732 | swp_entry_t swp; | |
4733 | ||
4734 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 4735 | return false; |
4a705fef | 4736 | swp = pte_to_swp_entry(pte); |
d79d176a | 4737 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 4738 | return true; |
4a705fef | 4739 | else |
3e5c3600 | 4740 | return false; |
4a705fef | 4741 | } |
1e8f889b | 4742 | |
4eae4efa PX |
4743 | static void |
4744 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
4745 | struct page *new_page) | |
4746 | { | |
4747 | __SetPageUptodate(new_page); | |
4eae4efa | 4748 | hugepage_add_new_anon_rmap(new_page, vma, addr); |
1eba86c0 | 4749 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); |
4eae4efa PX |
4750 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); |
4751 | ClearHPageRestoreReserve(new_page); | |
4752 | SetHPageMigratable(new_page); | |
4753 | } | |
4754 | ||
63551ae0 | 4755 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
bc70fbf2 PX |
4756 | struct vm_area_struct *dst_vma, |
4757 | struct vm_area_struct *src_vma) | |
63551ae0 | 4758 | { |
3aa4ed80 | 4759 | pte_t *src_pte, *dst_pte, entry; |
63551ae0 | 4760 | struct page *ptepage; |
1c59827d | 4761 | unsigned long addr; |
bc70fbf2 PX |
4762 | bool cow = is_cow_mapping(src_vma->vm_flags); |
4763 | struct hstate *h = hstate_vma(src_vma); | |
a5516438 | 4764 | unsigned long sz = huge_page_size(h); |
4eae4efa | 4765 | unsigned long npages = pages_per_huge_page(h); |
ac46d4f3 | 4766 | struct mmu_notifier_range range; |
e95a9851 | 4767 | unsigned long last_addr_mask; |
e8569dd2 | 4768 | int ret = 0; |
1e8f889b | 4769 | |
ac46d4f3 | 4770 | if (cow) { |
bc70fbf2 PX |
4771 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, |
4772 | src_vma->vm_start, | |
4773 | src_vma->vm_end); | |
ac46d4f3 | 4774 | mmu_notifier_invalidate_range_start(&range); |
623a1ddf DH |
4775 | mmap_assert_write_locked(src); |
4776 | raw_write_seqcount_begin(&src->write_protect_seq); | |
40549ba8 MK |
4777 | } else { |
4778 | /* | |
4779 | * For shared mappings the vma lock must be held before | |
4780 | * calling huge_pte_offset in the src vma. Otherwise, the | |
4781 | * returned ptep could go away if part of a shared pmd and | |
4782 | * another thread calls huge_pmd_unshare. | |
4783 | */ | |
4784 | hugetlb_vma_lock_read(src_vma); | |
ac46d4f3 | 4785 | } |
e8569dd2 | 4786 | |
e95a9851 | 4787 | last_addr_mask = hugetlb_mask_last_page(h); |
bc70fbf2 | 4788 | for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { |
cb900f41 | 4789 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 4790 | src_pte = huge_pte_offset(src, addr, sz); |
e95a9851 MK |
4791 | if (!src_pte) { |
4792 | addr |= last_addr_mask; | |
c74df32c | 4793 | continue; |
e95a9851 | 4794 | } |
bc70fbf2 | 4795 | dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); |
e8569dd2 AS |
4796 | if (!dst_pte) { |
4797 | ret = -ENOMEM; | |
4798 | break; | |
4799 | } | |
c5c99429 | 4800 | |
5e41540c MK |
4801 | /* |
4802 | * If the pagetables are shared don't copy or take references. | |
5e41540c | 4803 | * |
3aa4ed80 | 4804 | * dst_pte == src_pte is the common case of src/dest sharing. |
5e41540c | 4805 | * However, src could have 'unshared' and dst shares with |
3aa4ed80 ML |
4806 | * another vma. So page_count of ptep page is checked instead |
4807 | * to reliably determine whether pte is shared. | |
5e41540c | 4808 | */ |
3aa4ed80 | 4809 | if (page_count(virt_to_page(dst_pte)) > 1) { |
e95a9851 | 4810 | addr |= last_addr_mask; |
c5c99429 | 4811 | continue; |
e95a9851 | 4812 | } |
c5c99429 | 4813 | |
cb900f41 KS |
4814 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
4815 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4816 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 4817 | entry = huge_ptep_get(src_pte); |
4eae4efa | 4818 | again: |
3aa4ed80 | 4819 | if (huge_pte_none(entry)) { |
5e41540c | 4820 | /* |
3aa4ed80 | 4821 | * Skip if src entry none. |
5e41540c | 4822 | */ |
4a705fef | 4823 | ; |
c2cb0dcc NH |
4824 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) { |
4825 | bool uffd_wp = huge_pte_uffd_wp(entry); | |
4826 | ||
4827 | if (!userfaultfd_wp(dst_vma) && uffd_wp) | |
4828 | entry = huge_pte_clear_uffd_wp(entry); | |
4829 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
4830 | } else if (unlikely(is_hugetlb_entry_migration(entry))) { | |
4a705fef | 4831 | swp_entry_t swp_entry = pte_to_swp_entry(entry); |
bc70fbf2 | 4832 | bool uffd_wp = huge_pte_uffd_wp(entry); |
4a705fef | 4833 | |
6c287605 | 4834 | if (!is_readable_migration_entry(swp_entry) && cow) { |
4a705fef NH |
4835 | /* |
4836 | * COW mappings require pages in both | |
4837 | * parent and child to be set to read. | |
4838 | */ | |
4dd845b5 AP |
4839 | swp_entry = make_readable_migration_entry( |
4840 | swp_offset(swp_entry)); | |
4a705fef | 4841 | entry = swp_entry_to_pte(swp_entry); |
bc70fbf2 PX |
4842 | if (userfaultfd_wp(src_vma) && uffd_wp) |
4843 | entry = huge_pte_mkuffd_wp(entry); | |
18f39629 | 4844 | set_huge_pte_at(src, addr, src_pte, entry); |
4a705fef | 4845 | } |
bc70fbf2 PX |
4846 | if (!userfaultfd_wp(dst_vma) && uffd_wp) |
4847 | entry = huge_pte_clear_uffd_wp(entry); | |
18f39629 | 4848 | set_huge_pte_at(dst, addr, dst_pte, entry); |
bc70fbf2 PX |
4849 | } else if (unlikely(is_pte_marker(entry))) { |
4850 | /* | |
4851 | * We copy the pte marker only if the dst vma has | |
4852 | * uffd-wp enabled. | |
4853 | */ | |
4854 | if (userfaultfd_wp(dst_vma)) | |
4855 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
4a705fef | 4856 | } else { |
4eae4efa PX |
4857 | entry = huge_ptep_get(src_pte); |
4858 | ptepage = pte_page(entry); | |
4859 | get_page(ptepage); | |
4860 | ||
4861 | /* | |
fb3d824d DH |
4862 | * Failing to duplicate the anon rmap is a rare case |
4863 | * where we see pinned hugetlb pages while they're | |
4864 | * prone to COW. We need to do the COW earlier during | |
4865 | * fork. | |
4eae4efa PX |
4866 | * |
4867 | * When pre-allocating the page or copying data, we | |
4868 | * need to be without the pgtable locks since we could | |
4869 | * sleep during the process. | |
4870 | */ | |
fb3d824d DH |
4871 | if (!PageAnon(ptepage)) { |
4872 | page_dup_file_rmap(ptepage, true); | |
bc70fbf2 PX |
4873 | } else if (page_try_dup_anon_rmap(ptepage, true, |
4874 | src_vma)) { | |
4eae4efa PX |
4875 | pte_t src_pte_old = entry; |
4876 | struct page *new; | |
4877 | ||
4878 | spin_unlock(src_ptl); | |
4879 | spin_unlock(dst_ptl); | |
4880 | /* Do not use reserve as it's private owned */ | |
bc70fbf2 | 4881 | new = alloc_huge_page(dst_vma, addr, 1); |
4eae4efa PX |
4882 | if (IS_ERR(new)) { |
4883 | put_page(ptepage); | |
4884 | ret = PTR_ERR(new); | |
4885 | break; | |
4886 | } | |
bc70fbf2 | 4887 | copy_user_huge_page(new, ptepage, addr, dst_vma, |
4eae4efa PX |
4888 | npages); |
4889 | put_page(ptepage); | |
4890 | ||
4891 | /* Install the new huge page if src pte stable */ | |
4892 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
4893 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4894 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4895 | entry = huge_ptep_get(src_pte); | |
4896 | if (!pte_same(src_pte_old, entry)) { | |
bc70fbf2 | 4897 | restore_reserve_on_error(h, dst_vma, addr, |
846be085 | 4898 | new); |
4eae4efa | 4899 | put_page(new); |
3aa4ed80 | 4900 | /* huge_ptep of dst_pte won't change as in child */ |
4eae4efa PX |
4901 | goto again; |
4902 | } | |
bc70fbf2 | 4903 | hugetlb_install_page(dst_vma, dst_pte, addr, new); |
4eae4efa PX |
4904 | spin_unlock(src_ptl); |
4905 | spin_unlock(dst_ptl); | |
4906 | continue; | |
4907 | } | |
4908 | ||
34ee645e | 4909 | if (cow) { |
0f10851e JG |
4910 | /* |
4911 | * No need to notify as we are downgrading page | |
4912 | * table protection not changing it to point | |
4913 | * to a new page. | |
4914 | * | |
ee65728e | 4915 | * See Documentation/mm/mmu_notifier.rst |
0f10851e | 4916 | */ |
7f2e9525 | 4917 | huge_ptep_set_wrprotect(src, addr, src_pte); |
84894e1c | 4918 | entry = huge_pte_wrprotect(entry); |
34ee645e | 4919 | } |
4eae4efa | 4920 | |
1c59827d | 4921 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 4922 | hugetlb_count_add(npages, dst); |
1c59827d | 4923 | } |
cb900f41 KS |
4924 | spin_unlock(src_ptl); |
4925 | spin_unlock(dst_ptl); | |
63551ae0 | 4926 | } |
63551ae0 | 4927 | |
623a1ddf DH |
4928 | if (cow) { |
4929 | raw_write_seqcount_end(&src->write_protect_seq); | |
ac46d4f3 | 4930 | mmu_notifier_invalidate_range_end(&range); |
40549ba8 MK |
4931 | } else { |
4932 | hugetlb_vma_unlock_read(src_vma); | |
623a1ddf | 4933 | } |
e8569dd2 AS |
4934 | |
4935 | return ret; | |
63551ae0 DG |
4936 | } |
4937 | ||
550a7d60 | 4938 | static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, |
db110a99 | 4939 | unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) |
550a7d60 MA |
4940 | { |
4941 | struct hstate *h = hstate_vma(vma); | |
4942 | struct mm_struct *mm = vma->vm_mm; | |
550a7d60 | 4943 | spinlock_t *src_ptl, *dst_ptl; |
db110a99 | 4944 | pte_t pte; |
550a7d60 | 4945 | |
550a7d60 MA |
4946 | dst_ptl = huge_pte_lock(h, mm, dst_pte); |
4947 | src_ptl = huge_pte_lockptr(h, mm, src_pte); | |
4948 | ||
4949 | /* | |
4950 | * We don't have to worry about the ordering of src and dst ptlocks | |
4951 | * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. | |
4952 | */ | |
4953 | if (src_ptl != dst_ptl) | |
4954 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4955 | ||
4956 | pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); | |
4957 | set_huge_pte_at(mm, new_addr, dst_pte, pte); | |
4958 | ||
4959 | if (src_ptl != dst_ptl) | |
4960 | spin_unlock(src_ptl); | |
4961 | spin_unlock(dst_ptl); | |
4962 | } | |
4963 | ||
4964 | int move_hugetlb_page_tables(struct vm_area_struct *vma, | |
4965 | struct vm_area_struct *new_vma, | |
4966 | unsigned long old_addr, unsigned long new_addr, | |
4967 | unsigned long len) | |
4968 | { | |
4969 | struct hstate *h = hstate_vma(vma); | |
4970 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4971 | unsigned long sz = huge_page_size(h); | |
4972 | struct mm_struct *mm = vma->vm_mm; | |
4973 | unsigned long old_end = old_addr + len; | |
e95a9851 | 4974 | unsigned long last_addr_mask; |
550a7d60 MA |
4975 | pte_t *src_pte, *dst_pte; |
4976 | struct mmu_notifier_range range; | |
3d0b95cd | 4977 | bool shared_pmd = false; |
550a7d60 MA |
4978 | |
4979 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, | |
4980 | old_end); | |
4981 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
3d0b95cd BW |
4982 | /* |
4983 | * In case of shared PMDs, we should cover the maximum possible | |
4984 | * range. | |
4985 | */ | |
4986 | flush_cache_range(vma, range.start, range.end); | |
4987 | ||
550a7d60 | 4988 | mmu_notifier_invalidate_range_start(&range); |
e95a9851 | 4989 | last_addr_mask = hugetlb_mask_last_page(h); |
550a7d60 | 4990 | /* Prevent race with file truncation */ |
40549ba8 | 4991 | hugetlb_vma_lock_write(vma); |
550a7d60 MA |
4992 | i_mmap_lock_write(mapping); |
4993 | for (; old_addr < old_end; old_addr += sz, new_addr += sz) { | |
4994 | src_pte = huge_pte_offset(mm, old_addr, sz); | |
e95a9851 MK |
4995 | if (!src_pte) { |
4996 | old_addr |= last_addr_mask; | |
4997 | new_addr |= last_addr_mask; | |
550a7d60 | 4998 | continue; |
e95a9851 | 4999 | } |
550a7d60 MA |
5000 | if (huge_pte_none(huge_ptep_get(src_pte))) |
5001 | continue; | |
5002 | ||
4ddb4d91 | 5003 | if (huge_pmd_unshare(mm, vma, old_addr, src_pte)) { |
3d0b95cd | 5004 | shared_pmd = true; |
4ddb4d91 MK |
5005 | old_addr |= last_addr_mask; |
5006 | new_addr |= last_addr_mask; | |
550a7d60 | 5007 | continue; |
3d0b95cd | 5008 | } |
550a7d60 MA |
5009 | |
5010 | dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); | |
5011 | if (!dst_pte) | |
5012 | break; | |
5013 | ||
db110a99 | 5014 | move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); |
550a7d60 | 5015 | } |
3d0b95cd BW |
5016 | |
5017 | if (shared_pmd) | |
5018 | flush_tlb_range(vma, range.start, range.end); | |
5019 | else | |
5020 | flush_tlb_range(vma, old_end - len, old_end); | |
550a7d60 | 5021 | mmu_notifier_invalidate_range_end(&range); |
13e4ad2c | 5022 | i_mmap_unlock_write(mapping); |
40549ba8 | 5023 | hugetlb_vma_unlock_write(vma); |
550a7d60 MA |
5024 | |
5025 | return len + old_addr - old_end; | |
5026 | } | |
5027 | ||
73c54763 PX |
5028 | static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
5029 | unsigned long start, unsigned long end, | |
05e90bd0 | 5030 | struct page *ref_page, zap_flags_t zap_flags) |
63551ae0 DG |
5031 | { |
5032 | struct mm_struct *mm = vma->vm_mm; | |
5033 | unsigned long address; | |
c7546f8f | 5034 | pte_t *ptep; |
63551ae0 | 5035 | pte_t pte; |
cb900f41 | 5036 | spinlock_t *ptl; |
63551ae0 | 5037 | struct page *page; |
a5516438 AK |
5038 | struct hstate *h = hstate_vma(vma); |
5039 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 5040 | struct mmu_notifier_range range; |
e95a9851 | 5041 | unsigned long last_addr_mask; |
a4a118f2 | 5042 | bool force_flush = false; |
a5516438 | 5043 | |
63551ae0 | 5044 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
5045 | BUG_ON(start & ~huge_page_mask(h)); |
5046 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 5047 | |
07e32661 AK |
5048 | /* |
5049 | * This is a hugetlb vma, all the pte entries should point | |
5050 | * to huge page. | |
5051 | */ | |
ed6a7935 | 5052 | tlb_change_page_size(tlb, sz); |
24669e58 | 5053 | tlb_start_vma(tlb, vma); |
dff11abe MK |
5054 | |
5055 | /* | |
5056 | * If sharing possible, alert mmu notifiers of worst case. | |
5057 | */ | |
6f4f13e8 JG |
5058 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
5059 | end); | |
ac46d4f3 JG |
5060 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
5061 | mmu_notifier_invalidate_range_start(&range); | |
e95a9851 | 5062 | last_addr_mask = hugetlb_mask_last_page(h); |
569f48b8 | 5063 | address = start; |
569f48b8 | 5064 | for (; address < end; address += sz) { |
7868a208 | 5065 | ptep = huge_pte_offset(mm, address, sz); |
e95a9851 MK |
5066 | if (!ptep) { |
5067 | address |= last_addr_mask; | |
c7546f8f | 5068 | continue; |
e95a9851 | 5069 | } |
c7546f8f | 5070 | |
cb900f41 | 5071 | ptl = huge_pte_lock(h, mm, ptep); |
4ddb4d91 | 5072 | if (huge_pmd_unshare(mm, vma, address, ptep)) { |
31d49da5 | 5073 | spin_unlock(ptl); |
a4a118f2 NA |
5074 | tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); |
5075 | force_flush = true; | |
4ddb4d91 | 5076 | address |= last_addr_mask; |
31d49da5 AK |
5077 | continue; |
5078 | } | |
39dde65c | 5079 | |
6629326b | 5080 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
5081 | if (huge_pte_none(pte)) { |
5082 | spin_unlock(ptl); | |
5083 | continue; | |
5084 | } | |
6629326b HD |
5085 | |
5086 | /* | |
9fbc1f63 NH |
5087 | * Migrating hugepage or HWPoisoned hugepage is already |
5088 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 5089 | */ |
9fbc1f63 | 5090 | if (unlikely(!pte_present(pte))) { |
05e90bd0 PX |
5091 | /* |
5092 | * If the pte was wr-protected by uffd-wp in any of the | |
5093 | * swap forms, meanwhile the caller does not want to | |
5094 | * drop the uffd-wp bit in this zap, then replace the | |
5095 | * pte with a marker. | |
5096 | */ | |
5097 | if (pte_swp_uffd_wp_any(pte) && | |
5098 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5099 | set_huge_pte_at(mm, address, ptep, | |
5100 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5101 | else | |
5102 | huge_pte_clear(mm, address, ptep, sz); | |
31d49da5 AK |
5103 | spin_unlock(ptl); |
5104 | continue; | |
8c4894c6 | 5105 | } |
6629326b HD |
5106 | |
5107 | page = pte_page(pte); | |
04f2cbe3 MG |
5108 | /* |
5109 | * If a reference page is supplied, it is because a specific | |
5110 | * page is being unmapped, not a range. Ensure the page we | |
5111 | * are about to unmap is the actual page of interest. | |
5112 | */ | |
5113 | if (ref_page) { | |
31d49da5 AK |
5114 | if (page != ref_page) { |
5115 | spin_unlock(ptl); | |
5116 | continue; | |
5117 | } | |
04f2cbe3 MG |
5118 | /* |
5119 | * Mark the VMA as having unmapped its page so that | |
5120 | * future faults in this VMA will fail rather than | |
5121 | * looking like data was lost | |
5122 | */ | |
5123 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
5124 | } | |
5125 | ||
c7546f8f | 5126 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 5127 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 5128 | if (huge_pte_dirty(pte)) |
6649a386 | 5129 | set_page_dirty(page); |
05e90bd0 PX |
5130 | /* Leave a uffd-wp pte marker if needed */ |
5131 | if (huge_pte_uffd_wp(pte) && | |
5132 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5133 | set_huge_pte_at(mm, address, ptep, | |
5134 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5d317b2b | 5135 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
cea86fe2 | 5136 | page_remove_rmap(page, vma, true); |
31d49da5 | 5137 | |
cb900f41 | 5138 | spin_unlock(ptl); |
e77b0852 | 5139 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
5140 | /* |
5141 | * Bail out after unmapping reference page if supplied | |
5142 | */ | |
5143 | if (ref_page) | |
5144 | break; | |
fe1668ae | 5145 | } |
ac46d4f3 | 5146 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 5147 | tlb_end_vma(tlb, vma); |
a4a118f2 NA |
5148 | |
5149 | /* | |
5150 | * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We | |
5151 | * could defer the flush until now, since by holding i_mmap_rwsem we | |
5152 | * guaranteed that the last refernece would not be dropped. But we must | |
5153 | * do the flushing before we return, as otherwise i_mmap_rwsem will be | |
5154 | * dropped and the last reference to the shared PMDs page might be | |
5155 | * dropped as well. | |
5156 | * | |
5157 | * In theory we could defer the freeing of the PMD pages as well, but | |
5158 | * huge_pmd_unshare() relies on the exact page_count for the PMD page to | |
5159 | * detect sharing, so we cannot defer the release of the page either. | |
5160 | * Instead, do flush now. | |
5161 | */ | |
5162 | if (force_flush) | |
5163 | tlb_flush_mmu_tlbonly(tlb); | |
1da177e4 | 5164 | } |
63551ae0 | 5165 | |
d833352a MG |
5166 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
5167 | struct vm_area_struct *vma, unsigned long start, | |
05e90bd0 PX |
5168 | unsigned long end, struct page *ref_page, |
5169 | zap_flags_t zap_flags) | |
d833352a | 5170 | { |
05e90bd0 | 5171 | __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); |
d833352a MG |
5172 | |
5173 | /* | |
5174 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
5175 | * test will fail on a vma being torn down, and not grab a page table | |
5176 | * on its way out. We're lucky that the flag has such an appropriate | |
5177 | * name, and can in fact be safely cleared here. We could clear it | |
5178 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 5179 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 5180 | * because in the context this is called, the VMA is about to be |
c8c06efa | 5181 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
5182 | */ |
5183 | vma->vm_flags &= ~VM_MAYSHARE; | |
5184 | } | |
5185 | ||
502717f4 | 5186 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
05e90bd0 PX |
5187 | unsigned long end, struct page *ref_page, |
5188 | zap_flags_t zap_flags) | |
502717f4 | 5189 | { |
24669e58 | 5190 | struct mmu_gather tlb; |
dff11abe | 5191 | |
a72afd87 | 5192 | tlb_gather_mmu(&tlb, vma->vm_mm); |
05e90bd0 | 5193 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); |
ae8eba8b | 5194 | tlb_finish_mmu(&tlb); |
502717f4 CK |
5195 | } |
5196 | ||
04f2cbe3 MG |
5197 | /* |
5198 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 5199 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
5200 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
5201 | * same region. | |
5202 | */ | |
2f4612af DB |
5203 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
5204 | struct page *page, unsigned long address) | |
04f2cbe3 | 5205 | { |
7526674d | 5206 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
5207 | struct vm_area_struct *iter_vma; |
5208 | struct address_space *mapping; | |
04f2cbe3 MG |
5209 | pgoff_t pgoff; |
5210 | ||
5211 | /* | |
5212 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
5213 | * from page cache lookup which is in HPAGE_SIZE units. | |
5214 | */ | |
7526674d | 5215 | address = address & huge_page_mask(h); |
36e4f20a MH |
5216 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
5217 | vma->vm_pgoff; | |
93c76a3d | 5218 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 5219 | |
4eb2b1dc MG |
5220 | /* |
5221 | * Take the mapping lock for the duration of the table walk. As | |
5222 | * this mapping should be shared between all the VMAs, | |
5223 | * __unmap_hugepage_range() is called as the lock is already held | |
5224 | */ | |
83cde9e8 | 5225 | i_mmap_lock_write(mapping); |
6b2dbba8 | 5226 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
5227 | /* Do not unmap the current VMA */ |
5228 | if (iter_vma == vma) | |
5229 | continue; | |
5230 | ||
2f84a899 MG |
5231 | /* |
5232 | * Shared VMAs have their own reserves and do not affect | |
5233 | * MAP_PRIVATE accounting but it is possible that a shared | |
5234 | * VMA is using the same page so check and skip such VMAs. | |
5235 | */ | |
5236 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
5237 | continue; | |
5238 | ||
04f2cbe3 MG |
5239 | /* |
5240 | * Unmap the page from other VMAs without their own reserves. | |
5241 | * They get marked to be SIGKILLed if they fault in these | |
5242 | * areas. This is because a future no-page fault on this VMA | |
5243 | * could insert a zeroed page instead of the data existing | |
5244 | * from the time of fork. This would look like data corruption | |
5245 | */ | |
5246 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 | 5247 | unmap_hugepage_range(iter_vma, address, |
05e90bd0 | 5248 | address + huge_page_size(h), page, 0); |
04f2cbe3 | 5249 | } |
83cde9e8 | 5250 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
5251 | } |
5252 | ||
0fe6e20b | 5253 | /* |
c89357e2 | 5254 | * hugetlb_wp() should be called with page lock of the original hugepage held. |
aa6d2e8c | 5255 | * Called with hugetlb_fault_mutex_table held and pte_page locked so we |
ef009b25 MH |
5256 | * cannot race with other handlers or page migration. |
5257 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 5258 | */ |
c89357e2 DH |
5259 | static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, |
5260 | unsigned long address, pte_t *ptep, unsigned int flags, | |
3999f52e | 5261 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 5262 | { |
c89357e2 | 5263 | const bool unshare = flags & FAULT_FLAG_UNSHARE; |
3999f52e | 5264 | pte_t pte; |
a5516438 | 5265 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 5266 | struct page *old_page, *new_page; |
2b740303 SJ |
5267 | int outside_reserve = 0; |
5268 | vm_fault_t ret = 0; | |
974e6d66 | 5269 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 5270 | struct mmu_notifier_range range; |
1e8f889b | 5271 | |
c89357e2 DH |
5272 | VM_BUG_ON(unshare && (flags & FOLL_WRITE)); |
5273 | VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); | |
5274 | ||
1d8d1464 DH |
5275 | /* |
5276 | * hugetlb does not support FOLL_FORCE-style write faults that keep the | |
5277 | * PTE mapped R/O such as maybe_mkwrite() would do. | |
5278 | */ | |
5279 | if (WARN_ON_ONCE(!unshare && !(vma->vm_flags & VM_WRITE))) | |
5280 | return VM_FAULT_SIGSEGV; | |
5281 | ||
5282 | /* Let's take out MAP_SHARED mappings first. */ | |
5283 | if (vma->vm_flags & VM_MAYSHARE) { | |
5284 | if (unlikely(unshare)) | |
5285 | return 0; | |
5286 | set_huge_ptep_writable(vma, haddr, ptep); | |
5287 | return 0; | |
5288 | } | |
5289 | ||
3999f52e | 5290 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
5291 | old_page = pte_page(pte); |
5292 | ||
662ce1dc YY |
5293 | delayacct_wpcopy_start(); |
5294 | ||
04f2cbe3 | 5295 | retry_avoidcopy: |
c89357e2 DH |
5296 | /* |
5297 | * If no-one else is actually using this page, we're the exclusive | |
5298 | * owner and can reuse this page. | |
5299 | */ | |
37a2140d | 5300 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
c89357e2 DH |
5301 | if (!PageAnonExclusive(old_page)) |
5302 | page_move_anon_rmap(old_page, vma); | |
5303 | if (likely(!unshare)) | |
5304 | set_huge_ptep_writable(vma, haddr, ptep); | |
662ce1dc YY |
5305 | |
5306 | delayacct_wpcopy_end(); | |
83c54070 | 5307 | return 0; |
1e8f889b | 5308 | } |
6c287605 DH |
5309 | VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), |
5310 | old_page); | |
1e8f889b | 5311 | |
04f2cbe3 MG |
5312 | /* |
5313 | * If the process that created a MAP_PRIVATE mapping is about to | |
5314 | * perform a COW due to a shared page count, attempt to satisfy | |
5315 | * the allocation without using the existing reserves. The pagecache | |
5316 | * page is used to determine if the reserve at this address was | |
5317 | * consumed or not. If reserves were used, a partial faulted mapping | |
5318 | * at the time of fork() could consume its reserves on COW instead | |
5319 | * of the full address range. | |
5320 | */ | |
5944d011 | 5321 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
5322 | old_page != pagecache_page) |
5323 | outside_reserve = 1; | |
5324 | ||
09cbfeaf | 5325 | get_page(old_page); |
b76c8cfb | 5326 | |
ad4404a2 DB |
5327 | /* |
5328 | * Drop page table lock as buddy allocator may be called. It will | |
5329 | * be acquired again before returning to the caller, as expected. | |
5330 | */ | |
cb900f41 | 5331 | spin_unlock(ptl); |
5b7a1d40 | 5332 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 5333 | |
2fc39cec | 5334 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
5335 | /* |
5336 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
5337 | * it is due to references held by a child and an insufficient | |
5338 | * huge page pool. To guarantee the original mappers | |
5339 | * reliability, unmap the page from child processes. The child | |
5340 | * may get SIGKILLed if it later faults. | |
5341 | */ | |
5342 | if (outside_reserve) { | |
40549ba8 MK |
5343 | struct address_space *mapping = vma->vm_file->f_mapping; |
5344 | pgoff_t idx; | |
5345 | u32 hash; | |
5346 | ||
09cbfeaf | 5347 | put_page(old_page); |
40549ba8 MK |
5348 | /* |
5349 | * Drop hugetlb_fault_mutex and vma_lock before | |
5350 | * unmapping. unmapping needs to hold vma_lock | |
5351 | * in write mode. Dropping vma_lock in read mode | |
5352 | * here is OK as COW mappings do not interact with | |
5353 | * PMD sharing. | |
5354 | * | |
5355 | * Reacquire both after unmap operation. | |
5356 | */ | |
5357 | idx = vma_hugecache_offset(h, vma, haddr); | |
5358 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5359 | hugetlb_vma_unlock_read(vma); | |
5360 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5361 | ||
5b7a1d40 | 5362 | unmap_ref_private(mm, vma, old_page, haddr); |
40549ba8 MK |
5363 | |
5364 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
5365 | hugetlb_vma_lock_read(vma); | |
2f4612af | 5366 | spin_lock(ptl); |
5b7a1d40 | 5367 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
5368 | if (likely(ptep && |
5369 | pte_same(huge_ptep_get(ptep), pte))) | |
5370 | goto retry_avoidcopy; | |
5371 | /* | |
5372 | * race occurs while re-acquiring page table | |
5373 | * lock, and our job is done. | |
5374 | */ | |
662ce1dc | 5375 | delayacct_wpcopy_end(); |
2f4612af | 5376 | return 0; |
04f2cbe3 MG |
5377 | } |
5378 | ||
2b740303 | 5379 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 5380 | goto out_release_old; |
1e8f889b DG |
5381 | } |
5382 | ||
0fe6e20b NH |
5383 | /* |
5384 | * When the original hugepage is shared one, it does not have | |
5385 | * anon_vma prepared. | |
5386 | */ | |
44e2aa93 | 5387 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
5388 | ret = VM_FAULT_OOM; |
5389 | goto out_release_all; | |
44e2aa93 | 5390 | } |
0fe6e20b | 5391 | |
974e6d66 | 5392 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 5393 | pages_per_huge_page(h)); |
0ed361de | 5394 | __SetPageUptodate(new_page); |
1e8f889b | 5395 | |
7269f999 | 5396 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 5397 | haddr + huge_page_size(h)); |
ac46d4f3 | 5398 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 5399 | |
b76c8cfb | 5400 | /* |
cb900f41 | 5401 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
5402 | * before the page tables are altered |
5403 | */ | |
cb900f41 | 5404 | spin_lock(ptl); |
5b7a1d40 | 5405 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 5406 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 5407 | ClearHPageRestoreReserve(new_page); |
07443a85 | 5408 | |
c89357e2 | 5409 | /* Break COW or unshare */ |
5b7a1d40 | 5410 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 5411 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
cea86fe2 | 5412 | page_remove_rmap(old_page, vma, true); |
5b7a1d40 | 5413 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
1eba86c0 | 5414 | set_huge_pte_at(mm, haddr, ptep, |
c89357e2 | 5415 | make_huge_pte(vma, new_page, !unshare)); |
8f251a3d | 5416 | SetHPageMigratable(new_page); |
1e8f889b DG |
5417 | /* Make the old page be freed below */ |
5418 | new_page = old_page; | |
5419 | } | |
cb900f41 | 5420 | spin_unlock(ptl); |
ac46d4f3 | 5421 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 5422 | out_release_all: |
c89357e2 DH |
5423 | /* |
5424 | * No restore in case of successful pagetable update (Break COW or | |
5425 | * unshare) | |
5426 | */ | |
c7b1850d MK |
5427 | if (new_page != old_page) |
5428 | restore_reserve_on_error(h, vma, haddr, new_page); | |
09cbfeaf | 5429 | put_page(new_page); |
ad4404a2 | 5430 | out_release_old: |
09cbfeaf | 5431 | put_page(old_page); |
8312034f | 5432 | |
ad4404a2 | 5433 | spin_lock(ptl); /* Caller expects lock to be held */ |
662ce1dc YY |
5434 | |
5435 | delayacct_wpcopy_end(); | |
ad4404a2 | 5436 | return ret; |
1e8f889b DG |
5437 | } |
5438 | ||
3ae77f43 HD |
5439 | /* |
5440 | * Return whether there is a pagecache page to back given address within VMA. | |
5441 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
5442 | */ | |
5443 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
5444 | struct vm_area_struct *vma, unsigned long address) |
5445 | { | |
5446 | struct address_space *mapping; | |
5447 | pgoff_t idx; | |
5448 | struct page *page; | |
5449 | ||
5450 | mapping = vma->vm_file->f_mapping; | |
5451 | idx = vma_hugecache_offset(h, vma, address); | |
5452 | ||
5453 | page = find_get_page(mapping, idx); | |
5454 | if (page) | |
5455 | put_page(page); | |
5456 | return page != NULL; | |
5457 | } | |
5458 | ||
7e1813d4 | 5459 | int hugetlb_add_to_page_cache(struct page *page, struct address_space *mapping, |
ab76ad54 MK |
5460 | pgoff_t idx) |
5461 | { | |
d9ef44de | 5462 | struct folio *folio = page_folio(page); |
ab76ad54 MK |
5463 | struct inode *inode = mapping->host; |
5464 | struct hstate *h = hstate_inode(inode); | |
d9ef44de | 5465 | int err; |
ab76ad54 | 5466 | |
d9ef44de MWO |
5467 | __folio_set_locked(folio); |
5468 | err = __filemap_add_folio(mapping, folio, idx, GFP_KERNEL, NULL); | |
5469 | ||
5470 | if (unlikely(err)) { | |
5471 | __folio_clear_locked(folio); | |
ab76ad54 | 5472 | return err; |
d9ef44de | 5473 | } |
d6995da3 | 5474 | ClearHPageRestoreReserve(page); |
ab76ad54 | 5475 | |
22146c3c | 5476 | /* |
d9ef44de | 5477 | * mark folio dirty so that it will not be removed from cache/file |
22146c3c MK |
5478 | * by non-hugetlbfs specific code paths. |
5479 | */ | |
d9ef44de | 5480 | folio_mark_dirty(folio); |
22146c3c | 5481 | |
ab76ad54 MK |
5482 | spin_lock(&inode->i_lock); |
5483 | inode->i_blocks += blocks_per_huge_page(h); | |
5484 | spin_unlock(&inode->i_lock); | |
5485 | return 0; | |
5486 | } | |
5487 | ||
7677f7fd AR |
5488 | static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, |
5489 | struct address_space *mapping, | |
5490 | pgoff_t idx, | |
5491 | unsigned int flags, | |
5492 | unsigned long haddr, | |
824ddc60 | 5493 | unsigned long addr, |
7677f7fd AR |
5494 | unsigned long reason) |
5495 | { | |
7677f7fd AR |
5496 | u32 hash; |
5497 | struct vm_fault vmf = { | |
5498 | .vma = vma, | |
5499 | .address = haddr, | |
824ddc60 | 5500 | .real_address = addr, |
7677f7fd AR |
5501 | .flags = flags, |
5502 | ||
5503 | /* | |
5504 | * Hard to debug if it ends up being | |
5505 | * used by a callee that assumes | |
5506 | * something about the other | |
5507 | * uninitialized fields... same as in | |
5508 | * memory.c | |
5509 | */ | |
5510 | }; | |
5511 | ||
5512 | /* | |
958f32ce LS |
5513 | * vma_lock and hugetlb_fault_mutex must be dropped before handling |
5514 | * userfault. Also mmap_lock could be dropped due to handling | |
5515 | * userfault, any vma operation should be careful from here. | |
7677f7fd | 5516 | */ |
40549ba8 | 5517 | hugetlb_vma_unlock_read(vma); |
7677f7fd AR |
5518 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
5519 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
958f32ce | 5520 | return handle_userfault(&vmf, reason); |
7677f7fd AR |
5521 | } |
5522 | ||
2b740303 SJ |
5523 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
5524 | struct vm_area_struct *vma, | |
5525 | struct address_space *mapping, pgoff_t idx, | |
c64e912c PX |
5526 | unsigned long address, pte_t *ptep, |
5527 | pte_t old_pte, unsigned int flags) | |
ac9b9c66 | 5528 | { |
a5516438 | 5529 | struct hstate *h = hstate_vma(vma); |
2b740303 | 5530 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 5531 | int anon_rmap = 0; |
4c887265 | 5532 | unsigned long size; |
4c887265 | 5533 | struct page *page; |
1e8f889b | 5534 | pte_t new_pte; |
cb900f41 | 5535 | spinlock_t *ptl; |
285b8dca | 5536 | unsigned long haddr = address & huge_page_mask(h); |
c7b1850d | 5537 | bool new_page, new_pagecache_page = false; |
958f32ce | 5538 | u32 hash = hugetlb_fault_mutex_hash(mapping, idx); |
4c887265 | 5539 | |
04f2cbe3 MG |
5540 | /* |
5541 | * Currently, we are forced to kill the process in the event the | |
5542 | * original mapper has unmapped pages from the child due to a failed | |
c89357e2 DH |
5543 | * COW/unsharing. Warn that such a situation has occurred as it may not |
5544 | * be obvious. | |
04f2cbe3 MG |
5545 | */ |
5546 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 5547 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 5548 | current->pid); |
958f32ce | 5549 | goto out; |
04f2cbe3 MG |
5550 | } |
5551 | ||
4c887265 | 5552 | /* |
188a3972 MK |
5553 | * Use page lock to guard against racing truncation |
5554 | * before we get page_table_lock. | |
4c887265 | 5555 | */ |
c7b1850d | 5556 | new_page = false; |
6bda666a CL |
5557 | page = find_lock_page(mapping, idx); |
5558 | if (!page) { | |
188a3972 MK |
5559 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5560 | if (idx >= size) | |
5561 | goto out; | |
7677f7fd | 5562 | /* Check for page in userfault range */ |
958f32ce LS |
5563 | if (userfaultfd_missing(vma)) |
5564 | return hugetlb_handle_userfault(vma, mapping, idx, | |
824ddc60 | 5565 | flags, haddr, address, |
7677f7fd | 5566 | VM_UFFD_MISSING); |
1a1aad8a | 5567 | |
285b8dca | 5568 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 5569 | if (IS_ERR(page)) { |
4643d67e MK |
5570 | /* |
5571 | * Returning error will result in faulting task being | |
5572 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
5573 | * tasks from racing to fault in the same page which | |
5574 | * could result in false unable to allocate errors. | |
5575 | * Page migration does not take the fault mutex, but | |
5576 | * does a clear then write of pte's under page table | |
5577 | * lock. Page fault code could race with migration, | |
5578 | * notice the clear pte and try to allocate a page | |
5579 | * here. Before returning error, get ptl and make | |
5580 | * sure there really is no pte entry. | |
5581 | */ | |
5582 | ptl = huge_pte_lock(h, mm, ptep); | |
d83e6c8a ML |
5583 | ret = 0; |
5584 | if (huge_pte_none(huge_ptep_get(ptep))) | |
5585 | ret = vmf_error(PTR_ERR(page)); | |
4643d67e | 5586 | spin_unlock(ptl); |
6bda666a CL |
5587 | goto out; |
5588 | } | |
47ad8475 | 5589 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 5590 | __SetPageUptodate(page); |
cb6acd01 | 5591 | new_page = true; |
ac9b9c66 | 5592 | |
f83a275d | 5593 | if (vma->vm_flags & VM_MAYSHARE) { |
7e1813d4 | 5594 | int err = hugetlb_add_to_page_cache(page, mapping, idx); |
6bda666a | 5595 | if (err) { |
3a5497a2 ML |
5596 | /* |
5597 | * err can't be -EEXIST which implies someone | |
5598 | * else consumed the reservation since hugetlb | |
5599 | * fault mutex is held when add a hugetlb page | |
5600 | * to the page cache. So it's safe to call | |
5601 | * restore_reserve_on_error() here. | |
5602 | */ | |
5603 | restore_reserve_on_error(h, vma, haddr, page); | |
6bda666a | 5604 | put_page(page); |
6bda666a CL |
5605 | goto out; |
5606 | } | |
c7b1850d | 5607 | new_pagecache_page = true; |
23be7468 | 5608 | } else { |
6bda666a | 5609 | lock_page(page); |
0fe6e20b NH |
5610 | if (unlikely(anon_vma_prepare(vma))) { |
5611 | ret = VM_FAULT_OOM; | |
5612 | goto backout_unlocked; | |
5613 | } | |
409eb8c2 | 5614 | anon_rmap = 1; |
23be7468 | 5615 | } |
0fe6e20b | 5616 | } else { |
998b4382 NH |
5617 | /* |
5618 | * If memory error occurs between mmap() and fault, some process | |
5619 | * don't have hwpoisoned swap entry for errored virtual address. | |
5620 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
5621 | */ | |
5622 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 5623 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5624 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
5625 | goto backout_unlocked; |
5626 | } | |
7677f7fd AR |
5627 | |
5628 | /* Check for page in userfault range. */ | |
5629 | if (userfaultfd_minor(vma)) { | |
5630 | unlock_page(page); | |
5631 | put_page(page); | |
958f32ce | 5632 | return hugetlb_handle_userfault(vma, mapping, idx, |
824ddc60 | 5633 | flags, haddr, address, |
7677f7fd | 5634 | VM_UFFD_MINOR); |
7677f7fd | 5635 | } |
6bda666a | 5636 | } |
1e8f889b | 5637 | |
57303d80 AW |
5638 | /* |
5639 | * If we are going to COW a private mapping later, we examine the | |
5640 | * pending reservations for this page now. This will ensure that | |
5641 | * any allocations necessary to record that reservation occur outside | |
5642 | * the spinlock. | |
5643 | */ | |
5e911373 | 5644 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 5645 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
5646 | ret = VM_FAULT_OOM; |
5647 | goto backout_unlocked; | |
5648 | } | |
5e911373 | 5649 | /* Just decrements count, does not deallocate */ |
285b8dca | 5650 | vma_end_reservation(h, vma, haddr); |
5e911373 | 5651 | } |
57303d80 | 5652 | |
8bea8052 | 5653 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 5654 | ret = 0; |
c64e912c PX |
5655 | /* If pte changed from under us, retry */ |
5656 | if (!pte_same(huge_ptep_get(ptep), old_pte)) | |
4c887265 AL |
5657 | goto backout; |
5658 | ||
07443a85 | 5659 | if (anon_rmap) { |
d6995da3 | 5660 | ClearHPageRestoreReserve(page); |
285b8dca | 5661 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 5662 | } else |
fb3d824d | 5663 | page_dup_file_rmap(page, true); |
1e8f889b DG |
5664 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
5665 | && (vma->vm_flags & VM_SHARED))); | |
c64e912c PX |
5666 | /* |
5667 | * If this pte was previously wr-protected, keep it wr-protected even | |
5668 | * if populated. | |
5669 | */ | |
5670 | if (unlikely(pte_marker_uffd_wp(old_pte))) | |
5671 | new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); | |
285b8dca | 5672 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 5673 | |
5d317b2b | 5674 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 5675 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 5676 | /* Optimization, do the COW without a second fault */ |
c89357e2 | 5677 | ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); |
1e8f889b DG |
5678 | } |
5679 | ||
cb900f41 | 5680 | spin_unlock(ptl); |
cb6acd01 MK |
5681 | |
5682 | /* | |
8f251a3d MK |
5683 | * Only set HPageMigratable in newly allocated pages. Existing pages |
5684 | * found in the pagecache may not have HPageMigratableset if they have | |
5685 | * been isolated for migration. | |
cb6acd01 MK |
5686 | */ |
5687 | if (new_page) | |
8f251a3d | 5688 | SetHPageMigratable(page); |
cb6acd01 | 5689 | |
4c887265 AL |
5690 | unlock_page(page); |
5691 | out: | |
958f32ce LS |
5692 | hugetlb_vma_unlock_read(vma); |
5693 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
ac9b9c66 | 5694 | return ret; |
4c887265 AL |
5695 | |
5696 | backout: | |
cb900f41 | 5697 | spin_unlock(ptl); |
2b26736c | 5698 | backout_unlocked: |
c7b1850d MK |
5699 | if (new_page && !new_pagecache_page) |
5700 | restore_reserve_on_error(h, vma, haddr, page); | |
fa27759a MK |
5701 | |
5702 | unlock_page(page); | |
4c887265 AL |
5703 | put_page(page); |
5704 | goto out; | |
ac9b9c66 HD |
5705 | } |
5706 | ||
8382d914 | 5707 | #ifdef CONFIG_SMP |
188b04a7 | 5708 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5709 | { |
5710 | unsigned long key[2]; | |
5711 | u32 hash; | |
5712 | ||
1b426bac MK |
5713 | key[0] = (unsigned long) mapping; |
5714 | key[1] = idx; | |
8382d914 | 5715 | |
55254636 | 5716 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
5717 | |
5718 | return hash & (num_fault_mutexes - 1); | |
5719 | } | |
5720 | #else | |
5721 | /* | |
6c26d310 | 5722 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
5723 | * return 0 and avoid the hashing overhead. |
5724 | */ | |
188b04a7 | 5725 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5726 | { |
5727 | return 0; | |
5728 | } | |
5729 | #endif | |
5730 | ||
2b740303 | 5731 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 5732 | unsigned long address, unsigned int flags) |
86e5216f | 5733 | { |
8382d914 | 5734 | pte_t *ptep, entry; |
cb900f41 | 5735 | spinlock_t *ptl; |
2b740303 | 5736 | vm_fault_t ret; |
8382d914 DB |
5737 | u32 hash; |
5738 | pgoff_t idx; | |
0fe6e20b | 5739 | struct page *page = NULL; |
57303d80 | 5740 | struct page *pagecache_page = NULL; |
a5516438 | 5741 | struct hstate *h = hstate_vma(vma); |
8382d914 | 5742 | struct address_space *mapping; |
0f792cf9 | 5743 | int need_wait_lock = 0; |
285b8dca | 5744 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 5745 | |
285b8dca | 5746 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 5747 | if (ptep) { |
40549ba8 MK |
5748 | /* |
5749 | * Since we hold no locks, ptep could be stale. That is | |
5750 | * OK as we are only making decisions based on content and | |
5751 | * not actually modifying content here. | |
5752 | */ | |
fd6a03ed | 5753 | entry = huge_ptep_get(ptep); |
290408d4 | 5754 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
ad1ac596 | 5755 | migration_entry_wait_huge(vma, ptep); |
290408d4 NH |
5756 | return 0; |
5757 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 5758 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5759 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
5760 | } |
5761 | ||
3935baa9 DG |
5762 | /* |
5763 | * Serialize hugepage allocation and instantiation, so that we don't | |
5764 | * get spurious allocation failures if two CPUs race to instantiate | |
5765 | * the same page in the page cache. | |
5766 | */ | |
40549ba8 MK |
5767 | mapping = vma->vm_file->f_mapping; |
5768 | idx = vma_hugecache_offset(h, vma, haddr); | |
188b04a7 | 5769 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 5770 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 5771 | |
40549ba8 MK |
5772 | /* |
5773 | * Acquire vma lock before calling huge_pte_alloc and hold | |
5774 | * until finished with ptep. This prevents huge_pmd_unshare from | |
5775 | * being called elsewhere and making the ptep no longer valid. | |
5776 | * | |
5777 | * ptep could have already be assigned via huge_pte_offset. That | |
5778 | * is OK, as huge_pte_alloc will return the same value unless | |
5779 | * something has changed. | |
5780 | */ | |
5781 | hugetlb_vma_lock_read(vma); | |
5782 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); | |
5783 | if (!ptep) { | |
5784 | hugetlb_vma_unlock_read(vma); | |
5785 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5786 | return VM_FAULT_OOM; | |
5787 | } | |
5788 | ||
7f2e9525 | 5789 | entry = huge_ptep_get(ptep); |
c64e912c | 5790 | /* PTE markers should be handled the same way as none pte */ |
958f32ce LS |
5791 | if (huge_pte_none_mostly(entry)) |
5792 | /* | |
5793 | * hugetlb_no_page will drop vma lock and hugetlb fault | |
5794 | * mutex internally, which make us return immediately. | |
5795 | */ | |
5796 | return hugetlb_no_page(mm, vma, mapping, idx, address, ptep, | |
c64e912c | 5797 | entry, flags); |
86e5216f | 5798 | |
83c54070 | 5799 | ret = 0; |
1e8f889b | 5800 | |
0f792cf9 NH |
5801 | /* |
5802 | * entry could be a migration/hwpoison entry at this point, so this | |
5803 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
5804 | * an active hugepage in pagecache. This goto expects the 2nd page |
5805 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
5806 | * properly handle it. | |
0f792cf9 NH |
5807 | */ |
5808 | if (!pte_present(entry)) | |
5809 | goto out_mutex; | |
5810 | ||
57303d80 | 5811 | /* |
c89357e2 DH |
5812 | * If we are going to COW/unshare the mapping later, we examine the |
5813 | * pending reservations for this page now. This will ensure that any | |
57303d80 | 5814 | * allocations necessary to record that reservation occur outside the |
1d8d1464 DH |
5815 | * spinlock. Also lookup the pagecache page now as it is used to |
5816 | * determine if a reservation has been consumed. | |
57303d80 | 5817 | */ |
c89357e2 | 5818 | if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && |
1d8d1464 | 5819 | !(vma->vm_flags & VM_MAYSHARE) && !huge_pte_write(entry)) { |
285b8dca | 5820 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 5821 | ret = VM_FAULT_OOM; |
b4d1d99f | 5822 | goto out_mutex; |
2b26736c | 5823 | } |
5e911373 | 5824 | /* Just decrements count, does not deallocate */ |
285b8dca | 5825 | vma_end_reservation(h, vma, haddr); |
57303d80 | 5826 | |
29be8426 | 5827 | pagecache_page = find_lock_page(mapping, idx); |
57303d80 AW |
5828 | } |
5829 | ||
0f792cf9 NH |
5830 | ptl = huge_pte_lock(h, mm, ptep); |
5831 | ||
c89357e2 | 5832 | /* Check for a racing update before calling hugetlb_wp() */ |
0f792cf9 NH |
5833 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
5834 | goto out_ptl; | |
5835 | ||
166f3ecc PX |
5836 | /* Handle userfault-wp first, before trying to lock more pages */ |
5837 | if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && | |
5838 | (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { | |
5839 | struct vm_fault vmf = { | |
5840 | .vma = vma, | |
5841 | .address = haddr, | |
5842 | .real_address = address, | |
5843 | .flags = flags, | |
5844 | }; | |
5845 | ||
5846 | spin_unlock(ptl); | |
5847 | if (pagecache_page) { | |
5848 | unlock_page(pagecache_page); | |
5849 | put_page(pagecache_page); | |
5850 | } | |
40549ba8 | 5851 | hugetlb_vma_unlock_read(vma); |
166f3ecc | 5852 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
166f3ecc PX |
5853 | return handle_userfault(&vmf, VM_UFFD_WP); |
5854 | } | |
5855 | ||
56c9cfb1 | 5856 | /* |
c89357e2 | 5857 | * hugetlb_wp() requires page locks of pte_page(entry) and |
56c9cfb1 NH |
5858 | * pagecache_page, so here we need take the former one |
5859 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
5860 | */ |
5861 | page = pte_page(entry); | |
5862 | if (page != pagecache_page) | |
0f792cf9 NH |
5863 | if (!trylock_page(page)) { |
5864 | need_wait_lock = 1; | |
5865 | goto out_ptl; | |
5866 | } | |
b4d1d99f | 5867 | |
0f792cf9 | 5868 | get_page(page); |
b4d1d99f | 5869 | |
c89357e2 | 5870 | if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { |
106c992a | 5871 | if (!huge_pte_write(entry)) { |
c89357e2 DH |
5872 | ret = hugetlb_wp(mm, vma, address, ptep, flags, |
5873 | pagecache_page, ptl); | |
0f792cf9 | 5874 | goto out_put_page; |
c89357e2 DH |
5875 | } else if (likely(flags & FAULT_FLAG_WRITE)) { |
5876 | entry = huge_pte_mkdirty(entry); | |
b4d1d99f | 5877 | } |
b4d1d99f DG |
5878 | } |
5879 | entry = pte_mkyoung(entry); | |
285b8dca | 5880 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 5881 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 5882 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
5883 | out_put_page: |
5884 | if (page != pagecache_page) | |
5885 | unlock_page(page); | |
5886 | put_page(page); | |
cb900f41 KS |
5887 | out_ptl: |
5888 | spin_unlock(ptl); | |
57303d80 AW |
5889 | |
5890 | if (pagecache_page) { | |
5891 | unlock_page(pagecache_page); | |
5892 | put_page(pagecache_page); | |
5893 | } | |
b4d1d99f | 5894 | out_mutex: |
40549ba8 | 5895 | hugetlb_vma_unlock_read(vma); |
c672c7f2 | 5896 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
0f792cf9 NH |
5897 | /* |
5898 | * Generally it's safe to hold refcount during waiting page lock. But | |
5899 | * here we just wait to defer the next page fault to avoid busy loop and | |
5900 | * the page is not used after unlocked before returning from the current | |
5901 | * page fault. So we are safe from accessing freed page, even if we wait | |
5902 | * here without taking refcount. | |
5903 | */ | |
5904 | if (need_wait_lock) | |
5905 | wait_on_page_locked(page); | |
1e8f889b | 5906 | return ret; |
86e5216f AL |
5907 | } |
5908 | ||
714c1891 | 5909 | #ifdef CONFIG_USERFAULTFD |
8fb5debc MK |
5910 | /* |
5911 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
5912 | * modifications for huge pages. | |
5913 | */ | |
5914 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
5915 | pte_t *dst_pte, | |
5916 | struct vm_area_struct *dst_vma, | |
5917 | unsigned long dst_addr, | |
5918 | unsigned long src_addr, | |
f6191471 | 5919 | enum mcopy_atomic_mode mode, |
6041c691 PX |
5920 | struct page **pagep, |
5921 | bool wp_copy) | |
8fb5debc | 5922 | { |
f6191471 | 5923 | bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); |
8cc5fcbb MA |
5924 | struct hstate *h = hstate_vma(dst_vma); |
5925 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
5926 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
1e392147 | 5927 | unsigned long size; |
1c9e8def | 5928 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
5929 | pte_t _dst_pte; |
5930 | spinlock_t *ptl; | |
8cc5fcbb | 5931 | int ret = -ENOMEM; |
8fb5debc | 5932 | struct page *page; |
f6191471 | 5933 | int writable; |
cc30042d | 5934 | bool page_in_pagecache = false; |
8fb5debc | 5935 | |
f6191471 AR |
5936 | if (is_continue) { |
5937 | ret = -EFAULT; | |
5938 | page = find_lock_page(mapping, idx); | |
5939 | if (!page) | |
5940 | goto out; | |
cc30042d | 5941 | page_in_pagecache = true; |
f6191471 | 5942 | } else if (!*pagep) { |
d84cf06e MA |
5943 | /* If a page already exists, then it's UFFDIO_COPY for |
5944 | * a non-missing case. Return -EEXIST. | |
5945 | */ | |
5946 | if (vm_shared && | |
5947 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5948 | ret = -EEXIST; | |
5949 | goto out; | |
5950 | } | |
5951 | ||
8fb5debc | 5952 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
d84cf06e MA |
5953 | if (IS_ERR(page)) { |
5954 | ret = -ENOMEM; | |
8fb5debc | 5955 | goto out; |
d84cf06e | 5956 | } |
8fb5debc MK |
5957 | |
5958 | ret = copy_huge_page_from_user(page, | |
5959 | (const void __user *) src_addr, | |
810a56b9 | 5960 | pages_per_huge_page(h), false); |
8fb5debc | 5961 | |
c1e8d7c6 | 5962 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 5963 | if (unlikely(ret)) { |
9e368259 | 5964 | ret = -ENOENT; |
8cc5fcbb MA |
5965 | /* Free the allocated page which may have |
5966 | * consumed a reservation. | |
5967 | */ | |
5968 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
5969 | put_page(page); | |
5970 | ||
5971 | /* Allocate a temporary page to hold the copied | |
5972 | * contents. | |
5973 | */ | |
5974 | page = alloc_huge_page_vma(h, dst_vma, dst_addr); | |
5975 | if (!page) { | |
5976 | ret = -ENOMEM; | |
5977 | goto out; | |
5978 | } | |
8fb5debc | 5979 | *pagep = page; |
8cc5fcbb MA |
5980 | /* Set the outparam pagep and return to the caller to |
5981 | * copy the contents outside the lock. Don't free the | |
5982 | * page. | |
5983 | */ | |
8fb5debc MK |
5984 | goto out; |
5985 | } | |
5986 | } else { | |
8cc5fcbb MA |
5987 | if (vm_shared && |
5988 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5989 | put_page(*pagep); | |
5990 | ret = -EEXIST; | |
5991 | *pagep = NULL; | |
5992 | goto out; | |
5993 | } | |
5994 | ||
5995 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
5996 | if (IS_ERR(page)) { | |
da9a298f | 5997 | put_page(*pagep); |
8cc5fcbb MA |
5998 | ret = -ENOMEM; |
5999 | *pagep = NULL; | |
6000 | goto out; | |
6001 | } | |
34892366 MS |
6002 | copy_user_huge_page(page, *pagep, dst_addr, dst_vma, |
6003 | pages_per_huge_page(h)); | |
8cc5fcbb | 6004 | put_page(*pagep); |
8fb5debc MK |
6005 | *pagep = NULL; |
6006 | } | |
6007 | ||
6008 | /* | |
6009 | * The memory barrier inside __SetPageUptodate makes sure that | |
6010 | * preceding stores to the page contents become visible before | |
6011 | * the set_pte_at() write. | |
6012 | */ | |
6013 | __SetPageUptodate(page); | |
8fb5debc | 6014 | |
f6191471 AR |
6015 | /* Add shared, newly allocated pages to the page cache. */ |
6016 | if (vm_shared && !is_continue) { | |
1e392147 AA |
6017 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
6018 | ret = -EFAULT; | |
6019 | if (idx >= size) | |
6020 | goto out_release_nounlock; | |
1c9e8def | 6021 | |
1e392147 AA |
6022 | /* |
6023 | * Serialization between remove_inode_hugepages() and | |
7e1813d4 | 6024 | * hugetlb_add_to_page_cache() below happens through the |
1e392147 AA |
6025 | * hugetlb_fault_mutex_table that here must be hold by |
6026 | * the caller. | |
6027 | */ | |
7e1813d4 | 6028 | ret = hugetlb_add_to_page_cache(page, mapping, idx); |
1c9e8def MK |
6029 | if (ret) |
6030 | goto out_release_nounlock; | |
cc30042d | 6031 | page_in_pagecache = true; |
1c9e8def MK |
6032 | } |
6033 | ||
bcc66543 | 6034 | ptl = huge_pte_lock(h, dst_mm, dst_pte); |
8fb5debc | 6035 | |
6041c691 PX |
6036 | /* |
6037 | * We allow to overwrite a pte marker: consider when both MISSING|WP | |
6038 | * registered, we firstly wr-protect a none pte which has no page cache | |
6039 | * page backing it, then access the page. | |
6040 | */ | |
fa27759a | 6041 | ret = -EEXIST; |
6041c691 | 6042 | if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) |
8fb5debc MK |
6043 | goto out_release_unlock; |
6044 | ||
ab74ef70 | 6045 | if (page_in_pagecache) { |
fb3d824d | 6046 | page_dup_file_rmap(page, true); |
1c9e8def | 6047 | } else { |
d6995da3 | 6048 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
6049 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
6050 | } | |
8fb5debc | 6051 | |
6041c691 PX |
6052 | /* |
6053 | * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY | |
6054 | * with wp flag set, don't set pte write bit. | |
6055 | */ | |
6056 | if (wp_copy || (is_continue && !vm_shared)) | |
f6191471 AR |
6057 | writable = 0; |
6058 | else | |
6059 | writable = dst_vma->vm_flags & VM_WRITE; | |
6060 | ||
6061 | _dst_pte = make_huge_pte(dst_vma, page, writable); | |
6041c691 PX |
6062 | /* |
6063 | * Always mark UFFDIO_COPY page dirty; note that this may not be | |
6064 | * extremely important for hugetlbfs for now since swapping is not | |
6065 | * supported, but we should still be clear in that this page cannot be | |
6066 | * thrown away at will, even if write bit not set. | |
6067 | */ | |
6068 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
8fb5debc MK |
6069 | _dst_pte = pte_mkyoung(_dst_pte); |
6070 | ||
6041c691 PX |
6071 | if (wp_copy) |
6072 | _dst_pte = huge_pte_mkuffd_wp(_dst_pte); | |
6073 | ||
8fb5debc MK |
6074 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); |
6075 | ||
8fb5debc MK |
6076 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); |
6077 | ||
6078 | /* No need to invalidate - it was non-present before */ | |
6079 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
6080 | ||
6081 | spin_unlock(ptl); | |
f6191471 AR |
6082 | if (!is_continue) |
6083 | SetHPageMigratable(page); | |
6084 | if (vm_shared || is_continue) | |
1c9e8def | 6085 | unlock_page(page); |
8fb5debc MK |
6086 | ret = 0; |
6087 | out: | |
6088 | return ret; | |
6089 | out_release_unlock: | |
6090 | spin_unlock(ptl); | |
f6191471 | 6091 | if (vm_shared || is_continue) |
1c9e8def | 6092 | unlock_page(page); |
5af10dfd | 6093 | out_release_nounlock: |
cc30042d | 6094 | if (!page_in_pagecache) |
c7b1850d | 6095 | restore_reserve_on_error(h, dst_vma, dst_addr, page); |
8fb5debc MK |
6096 | put_page(page); |
6097 | goto out; | |
6098 | } | |
714c1891 | 6099 | #endif /* CONFIG_USERFAULTFD */ |
8fb5debc | 6100 | |
82e5d378 JM |
6101 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
6102 | int refs, struct page **pages, | |
6103 | struct vm_area_struct **vmas) | |
6104 | { | |
6105 | int nr; | |
6106 | ||
6107 | for (nr = 0; nr < refs; nr++) { | |
6108 | if (likely(pages)) | |
14455eab | 6109 | pages[nr] = nth_page(page, nr); |
82e5d378 JM |
6110 | if (vmas) |
6111 | vmas[nr] = vma; | |
6112 | } | |
6113 | } | |
6114 | ||
a7f22660 DH |
6115 | static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, |
6116 | bool *unshare) | |
6117 | { | |
6118 | pte_t pteval = huge_ptep_get(pte); | |
6119 | ||
6120 | *unshare = false; | |
6121 | if (is_swap_pte(pteval)) | |
6122 | return true; | |
6123 | if (huge_pte_write(pteval)) | |
6124 | return false; | |
6125 | if (flags & FOLL_WRITE) | |
6126 | return true; | |
6127 | if (gup_must_unshare(flags, pte_page(pteval))) { | |
6128 | *unshare = true; | |
6129 | return true; | |
6130 | } | |
6131 | return false; | |
6132 | } | |
6133 | ||
28a35716 ML |
6134 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
6135 | struct page **pages, struct vm_area_struct **vmas, | |
6136 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 6137 | long i, unsigned int flags, int *locked) |
63551ae0 | 6138 | { |
d5d4b0aa CK |
6139 | unsigned long pfn_offset; |
6140 | unsigned long vaddr = *position; | |
28a35716 | 6141 | unsigned long remainder = *nr_pages; |
a5516438 | 6142 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 6143 | int err = -EFAULT, refs; |
63551ae0 | 6144 | |
63551ae0 | 6145 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 6146 | pte_t *pte; |
cb900f41 | 6147 | spinlock_t *ptl = NULL; |
a7f22660 | 6148 | bool unshare = false; |
2a15efc9 | 6149 | int absent; |
4c887265 | 6150 | struct page *page; |
63551ae0 | 6151 | |
02057967 DR |
6152 | /* |
6153 | * If we have a pending SIGKILL, don't keep faulting pages and | |
6154 | * potentially allocating memory. | |
6155 | */ | |
fa45f116 | 6156 | if (fatal_signal_pending(current)) { |
02057967 DR |
6157 | remainder = 0; |
6158 | break; | |
6159 | } | |
6160 | ||
4c887265 AL |
6161 | /* |
6162 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 6163 | * each hugepage. We have to make sure we get the |
4c887265 | 6164 | * first, for the page indexing below to work. |
cb900f41 KS |
6165 | * |
6166 | * Note that page table lock is not held when pte is null. | |
4c887265 | 6167 | */ |
7868a208 PA |
6168 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
6169 | huge_page_size(h)); | |
cb900f41 KS |
6170 | if (pte) |
6171 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
6172 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
6173 | ||
6174 | /* | |
6175 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
6176 | * an error where there's an empty slot with no huge pagecache |
6177 | * to back it. This way, we avoid allocating a hugepage, and | |
6178 | * the sparse dumpfile avoids allocating disk blocks, but its | |
6179 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 6180 | */ |
3ae77f43 HD |
6181 | if (absent && (flags & FOLL_DUMP) && |
6182 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
6183 | if (pte) |
6184 | spin_unlock(ptl); | |
2a15efc9 HD |
6185 | remainder = 0; |
6186 | break; | |
6187 | } | |
63551ae0 | 6188 | |
9cc3a5bd NH |
6189 | /* |
6190 | * We need call hugetlb_fault for both hugepages under migration | |
6191 | * (in which case hugetlb_fault waits for the migration,) and | |
6192 | * hwpoisoned hugepages (in which case we need to prevent the | |
6193 | * caller from accessing to them.) In order to do this, we use | |
6194 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
6195 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
6196 | * both cases, and because we can't follow correct pages | |
6197 | * directly from any kind of swap entries. | |
6198 | */ | |
a7f22660 DH |
6199 | if (absent || |
6200 | __follow_hugetlb_must_fault(flags, pte, &unshare)) { | |
2b740303 | 6201 | vm_fault_t ret; |
87ffc118 | 6202 | unsigned int fault_flags = 0; |
63551ae0 | 6203 | |
cb900f41 KS |
6204 | if (pte) |
6205 | spin_unlock(ptl); | |
87ffc118 AA |
6206 | if (flags & FOLL_WRITE) |
6207 | fault_flags |= FAULT_FLAG_WRITE; | |
a7f22660 DH |
6208 | else if (unshare) |
6209 | fault_flags |= FAULT_FLAG_UNSHARE; | |
4f6da934 | 6210 | if (locked) |
71335f37 PX |
6211 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
6212 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
6213 | if (flags & FOLL_NOWAIT) |
6214 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
6215 | FAULT_FLAG_RETRY_NOWAIT; | |
6216 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
6217 | /* |
6218 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
6219 | * FAULT_FLAG_TRIED can co-exist | |
6220 | */ | |
87ffc118 AA |
6221 | fault_flags |= FAULT_FLAG_TRIED; |
6222 | } | |
6223 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
6224 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 6225 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
6226 | remainder = 0; |
6227 | break; | |
6228 | } | |
6229 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 6230 | if (locked && |
1ac25013 | 6231 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 6232 | *locked = 0; |
87ffc118 AA |
6233 | *nr_pages = 0; |
6234 | /* | |
6235 | * VM_FAULT_RETRY must not return an | |
6236 | * error, it will return zero | |
6237 | * instead. | |
6238 | * | |
6239 | * No need to update "position" as the | |
6240 | * caller will not check it after | |
6241 | * *nr_pages is set to 0. | |
6242 | */ | |
6243 | return i; | |
6244 | } | |
6245 | continue; | |
4c887265 AL |
6246 | } |
6247 | ||
a5516438 | 6248 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 6249 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 6250 | |
b6a2619c DH |
6251 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && |
6252 | !PageAnonExclusive(page), page); | |
6253 | ||
acbfb087 ZL |
6254 | /* |
6255 | * If subpage information not requested, update counters | |
6256 | * and skip the same_page loop below. | |
6257 | */ | |
6258 | if (!pages && !vmas && !pfn_offset && | |
6259 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
6260 | (remainder >= pages_per_huge_page(h))) { | |
6261 | vaddr += huge_page_size(h); | |
6262 | remainder -= pages_per_huge_page(h); | |
6263 | i += pages_per_huge_page(h); | |
6264 | spin_unlock(ptl); | |
6265 | continue; | |
6266 | } | |
6267 | ||
d08af0a5 JM |
6268 | /* vaddr may not be aligned to PAGE_SIZE */ |
6269 | refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, | |
6270 | (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); | |
0fa5bc40 | 6271 | |
82e5d378 | 6272 | if (pages || vmas) |
14455eab | 6273 | record_subpages_vmas(nth_page(page, pfn_offset), |
82e5d378 JM |
6274 | vma, refs, |
6275 | likely(pages) ? pages + i : NULL, | |
6276 | vmas ? vmas + i : NULL); | |
63551ae0 | 6277 | |
82e5d378 | 6278 | if (pages) { |
0fa5bc40 | 6279 | /* |
822951d8 | 6280 | * try_grab_folio() should always succeed here, |
0fa5bc40 JM |
6281 | * because: a) we hold the ptl lock, and b) we've just |
6282 | * checked that the huge page is present in the page | |
6283 | * tables. If the huge page is present, then the tail | |
6284 | * pages must also be present. The ptl prevents the | |
6285 | * head page and tail pages from being rearranged in | |
6286 | * any way. So this page must be available at this | |
6287 | * point, unless the page refcount overflowed: | |
6288 | */ | |
822951d8 MWO |
6289 | if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, |
6290 | flags))) { | |
0fa5bc40 JM |
6291 | spin_unlock(ptl); |
6292 | remainder = 0; | |
6293 | err = -ENOMEM; | |
6294 | break; | |
6295 | } | |
d5d4b0aa | 6296 | } |
82e5d378 JM |
6297 | |
6298 | vaddr += (refs << PAGE_SHIFT); | |
6299 | remainder -= refs; | |
6300 | i += refs; | |
6301 | ||
cb900f41 | 6302 | spin_unlock(ptl); |
63551ae0 | 6303 | } |
28a35716 | 6304 | *nr_pages = remainder; |
87ffc118 AA |
6305 | /* |
6306 | * setting position is actually required only if remainder is | |
6307 | * not zero but it's faster not to add a "if (remainder)" | |
6308 | * branch. | |
6309 | */ | |
63551ae0 DG |
6310 | *position = vaddr; |
6311 | ||
2be7cfed | 6312 | return i ? i : err; |
63551ae0 | 6313 | } |
8f860591 | 6314 | |
7da4d641 | 6315 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
5a90d5a1 PX |
6316 | unsigned long address, unsigned long end, |
6317 | pgprot_t newprot, unsigned long cp_flags) | |
8f860591 ZY |
6318 | { |
6319 | struct mm_struct *mm = vma->vm_mm; | |
6320 | unsigned long start = address; | |
6321 | pte_t *ptep; | |
6322 | pte_t pte; | |
a5516438 | 6323 | struct hstate *h = hstate_vma(vma); |
60dfaad6 | 6324 | unsigned long pages = 0, psize = huge_page_size(h); |
dff11abe | 6325 | bool shared_pmd = false; |
ac46d4f3 | 6326 | struct mmu_notifier_range range; |
e95a9851 | 6327 | unsigned long last_addr_mask; |
5a90d5a1 PX |
6328 | bool uffd_wp = cp_flags & MM_CP_UFFD_WP; |
6329 | bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; | |
dff11abe MK |
6330 | |
6331 | /* | |
6332 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 6333 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
6334 | * range if PMD sharing is possible. |
6335 | */ | |
7269f999 JG |
6336 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
6337 | 0, vma, mm, start, end); | |
ac46d4f3 | 6338 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
6339 | |
6340 | BUG_ON(address >= end); | |
ac46d4f3 | 6341 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 6342 | |
ac46d4f3 | 6343 | mmu_notifier_invalidate_range_start(&range); |
40549ba8 | 6344 | hugetlb_vma_lock_write(vma); |
83cde9e8 | 6345 | i_mmap_lock_write(vma->vm_file->f_mapping); |
40549ba8 | 6346 | last_addr_mask = hugetlb_mask_last_page(h); |
60dfaad6 | 6347 | for (; address < end; address += psize) { |
cb900f41 | 6348 | spinlock_t *ptl; |
60dfaad6 | 6349 | ptep = huge_pte_offset(mm, address, psize); |
e95a9851 MK |
6350 | if (!ptep) { |
6351 | address |= last_addr_mask; | |
8f860591 | 6352 | continue; |
e95a9851 | 6353 | } |
cb900f41 | 6354 | ptl = huge_pte_lock(h, mm, ptep); |
4ddb4d91 | 6355 | if (huge_pmd_unshare(mm, vma, address, ptep)) { |
60dfaad6 PX |
6356 | /* |
6357 | * When uffd-wp is enabled on the vma, unshare | |
6358 | * shouldn't happen at all. Warn about it if it | |
6359 | * happened due to some reason. | |
6360 | */ | |
6361 | WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); | |
7da4d641 | 6362 | pages++; |
cb900f41 | 6363 | spin_unlock(ptl); |
dff11abe | 6364 | shared_pmd = true; |
4ddb4d91 | 6365 | address |= last_addr_mask; |
39dde65c | 6366 | continue; |
7da4d641 | 6367 | } |
a8bda28d NH |
6368 | pte = huge_ptep_get(ptep); |
6369 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
6370 | spin_unlock(ptl); | |
6371 | continue; | |
6372 | } | |
6373 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
6374 | swp_entry_t entry = pte_to_swp_entry(pte); | |
6c287605 | 6375 | struct page *page = pfn_swap_entry_to_page(entry); |
a8bda28d | 6376 | |
6c287605 | 6377 | if (!is_readable_migration_entry(entry)) { |
a8bda28d NH |
6378 | pte_t newpte; |
6379 | ||
6c287605 DH |
6380 | if (PageAnon(page)) |
6381 | entry = make_readable_exclusive_migration_entry( | |
6382 | swp_offset(entry)); | |
6383 | else | |
6384 | entry = make_readable_migration_entry( | |
6385 | swp_offset(entry)); | |
a8bda28d | 6386 | newpte = swp_entry_to_pte(entry); |
5a90d5a1 PX |
6387 | if (uffd_wp) |
6388 | newpte = pte_swp_mkuffd_wp(newpte); | |
6389 | else if (uffd_wp_resolve) | |
6390 | newpte = pte_swp_clear_uffd_wp(newpte); | |
18f39629 | 6391 | set_huge_pte_at(mm, address, ptep, newpte); |
a8bda28d NH |
6392 | pages++; |
6393 | } | |
6394 | spin_unlock(ptl); | |
6395 | continue; | |
6396 | } | |
60dfaad6 PX |
6397 | if (unlikely(pte_marker_uffd_wp(pte))) { |
6398 | /* | |
6399 | * This is changing a non-present pte into a none pte, | |
6400 | * no need for huge_ptep_modify_prot_start/commit(). | |
6401 | */ | |
6402 | if (uffd_wp_resolve) | |
6403 | huge_pte_clear(mm, address, ptep, psize); | |
6404 | } | |
a8bda28d | 6405 | if (!huge_pte_none(pte)) { |
023bdd00 | 6406 | pte_t old_pte; |
79c1c594 | 6407 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
023bdd00 AK |
6408 | |
6409 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
16785bd7 | 6410 | pte = huge_pte_modify(old_pte, newprot); |
79c1c594 | 6411 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
5a90d5a1 PX |
6412 | if (uffd_wp) |
6413 | pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); | |
6414 | else if (uffd_wp_resolve) | |
6415 | pte = huge_pte_clear_uffd_wp(pte); | |
023bdd00 | 6416 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 6417 | pages++; |
60dfaad6 PX |
6418 | } else { |
6419 | /* None pte */ | |
6420 | if (unlikely(uffd_wp)) | |
6421 | /* Safe to modify directly (none->non-present). */ | |
6422 | set_huge_pte_at(mm, address, ptep, | |
6423 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
8f860591 | 6424 | } |
cb900f41 | 6425 | spin_unlock(ptl); |
8f860591 | 6426 | } |
d833352a | 6427 | /* |
c8c06efa | 6428 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 6429 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 6430 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
6431 | * and that page table be reused and filled with junk. If we actually |
6432 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 6433 | */ |
dff11abe | 6434 | if (shared_pmd) |
ac46d4f3 | 6435 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
6436 | else |
6437 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
6438 | /* |
6439 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
6440 | * page table protection not changing it to point to a new page. | |
6441 | * | |
ee65728e | 6442 | * See Documentation/mm/mmu_notifier.rst |
0f10851e | 6443 | */ |
83cde9e8 | 6444 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
40549ba8 | 6445 | hugetlb_vma_unlock_write(vma); |
ac46d4f3 | 6446 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
6447 | |
6448 | return pages << h->order; | |
8f860591 ZY |
6449 | } |
6450 | ||
33b8f84a MK |
6451 | /* Return true if reservation was successful, false otherwise. */ |
6452 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 6453 | long from, long to, |
5a6fe125 | 6454 | struct vm_area_struct *vma, |
ca16d140 | 6455 | vm_flags_t vm_flags) |
e4e574b7 | 6456 | { |
33b8f84a | 6457 | long chg, add = -1; |
a5516438 | 6458 | struct hstate *h = hstate_inode(inode); |
90481622 | 6459 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 6460 | struct resv_map *resv_map; |
075a61d0 | 6461 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 6462 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 6463 | |
63489f8e MK |
6464 | /* This should never happen */ |
6465 | if (from > to) { | |
6466 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 6467 | return false; |
63489f8e MK |
6468 | } |
6469 | ||
8d9bfb26 MK |
6470 | /* |
6471 | * vma specific semaphore used for pmd sharing synchronization | |
6472 | */ | |
6473 | hugetlb_vma_lock_alloc(vma); | |
6474 | ||
17c9d12e MG |
6475 | /* |
6476 | * Only apply hugepage reservation if asked. At fault time, an | |
6477 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 6478 | * without using reserves |
17c9d12e | 6479 | */ |
ca16d140 | 6480 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 6481 | return true; |
17c9d12e | 6482 | |
a1e78772 MG |
6483 | /* |
6484 | * Shared mappings base their reservation on the number of pages that | |
6485 | * are already allocated on behalf of the file. Private mappings need | |
6486 | * to reserve the full area even if read-only as mprotect() may be | |
6487 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
6488 | */ | |
9119a41e | 6489 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
6490 | /* |
6491 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
6492 | * called for inodes for which resv_maps were created (see | |
6493 | * hugetlbfs_get_inode). | |
6494 | */ | |
4e35f483 | 6495 | resv_map = inode_resv_map(inode); |
9119a41e | 6496 | |
0db9d74e | 6497 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e | 6498 | } else { |
e9fe92ae | 6499 | /* Private mapping. */ |
9119a41e | 6500 | resv_map = resv_map_alloc(); |
17c9d12e | 6501 | if (!resv_map) |
8d9bfb26 | 6502 | goto out_err; |
17c9d12e | 6503 | |
a1e78772 | 6504 | chg = to - from; |
84afd99b | 6505 | |
17c9d12e MG |
6506 | set_vma_resv_map(vma, resv_map); |
6507 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
6508 | } | |
6509 | ||
33b8f84a | 6510 | if (chg < 0) |
c50ac050 | 6511 | goto out_err; |
8a630112 | 6512 | |
33b8f84a MK |
6513 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
6514 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 6515 | goto out_err; |
075a61d0 MA |
6516 | |
6517 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
6518 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
6519 | * of the resv_map. | |
6520 | */ | |
6521 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
6522 | } | |
6523 | ||
1c5ecae3 MK |
6524 | /* |
6525 | * There must be enough pages in the subpool for the mapping. If | |
6526 | * the subpool has a minimum size, there may be some global | |
6527 | * reservations already in place (gbl_reserve). | |
6528 | */ | |
6529 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 6530 | if (gbl_reserve < 0) |
075a61d0 | 6531 | goto out_uncharge_cgroup; |
5a6fe125 MG |
6532 | |
6533 | /* | |
17c9d12e | 6534 | * Check enough hugepages are available for the reservation. |
90481622 | 6535 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 6536 | */ |
33b8f84a | 6537 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 6538 | goto out_put_pages; |
17c9d12e MG |
6539 | |
6540 | /* | |
6541 | * Account for the reservations made. Shared mappings record regions | |
6542 | * that have reservations as they are shared by multiple VMAs. | |
6543 | * When the last VMA disappears, the region map says how much | |
6544 | * the reservation was and the page cache tells how much of | |
6545 | * the reservation was consumed. Private mappings are per-VMA and | |
6546 | * only the consumed reservations are tracked. When the VMA | |
6547 | * disappears, the original reservation is the VMA size and the | |
6548 | * consumed reservations are stored in the map. Hence, nothing | |
6549 | * else has to be done for private mappings here | |
6550 | */ | |
33039678 | 6551 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 6552 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
6553 | |
6554 | if (unlikely(add < 0)) { | |
6555 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 6556 | goto out_put_pages; |
0db9d74e | 6557 | } else if (unlikely(chg > add)) { |
33039678 MK |
6558 | /* |
6559 | * pages in this range were added to the reserve | |
6560 | * map between region_chg and region_add. This | |
6561 | * indicates a race with alloc_huge_page. Adjust | |
6562 | * the subpool and reserve counts modified above | |
6563 | * based on the difference. | |
6564 | */ | |
6565 | long rsv_adjust; | |
6566 | ||
d85aecf2 ML |
6567 | /* |
6568 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
6569 | * reference to h_cg->css. See comment below for detail. | |
6570 | */ | |
075a61d0 MA |
6571 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
6572 | hstate_index(h), | |
6573 | (chg - add) * pages_per_huge_page(h), h_cg); | |
6574 | ||
33039678 MK |
6575 | rsv_adjust = hugepage_subpool_put_pages(spool, |
6576 | chg - add); | |
6577 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
6578 | } else if (h_cg) { |
6579 | /* | |
6580 | * The file_regions will hold their own reference to | |
6581 | * h_cg->css. So we should release the reference held | |
6582 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
6583 | * done. | |
6584 | */ | |
6585 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
6586 | } |
6587 | } | |
33b8f84a MK |
6588 | return true; |
6589 | ||
075a61d0 MA |
6590 | out_put_pages: |
6591 | /* put back original number of pages, chg */ | |
6592 | (void)hugepage_subpool_put_pages(spool, chg); | |
6593 | out_uncharge_cgroup: | |
6594 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
6595 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 6596 | out_err: |
8d9bfb26 | 6597 | hugetlb_vma_lock_free(vma); |
5e911373 | 6598 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
6599 | /* Only call region_abort if the region_chg succeeded but the |
6600 | * region_add failed or didn't run. | |
6601 | */ | |
6602 | if (chg >= 0 && add < 0) | |
6603 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
6604 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
6605 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 6606 | return false; |
a43a8c39 CK |
6607 | } |
6608 | ||
b5cec28d MK |
6609 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
6610 | long freed) | |
a43a8c39 | 6611 | { |
a5516438 | 6612 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 6613 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 6614 | long chg = 0; |
90481622 | 6615 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 6616 | long gbl_reserve; |
45c682a6 | 6617 | |
f27a5136 MK |
6618 | /* |
6619 | * Since this routine can be called in the evict inode path for all | |
6620 | * hugetlbfs inodes, resv_map could be NULL. | |
6621 | */ | |
b5cec28d MK |
6622 | if (resv_map) { |
6623 | chg = region_del(resv_map, start, end); | |
6624 | /* | |
6625 | * region_del() can fail in the rare case where a region | |
6626 | * must be split and another region descriptor can not be | |
6627 | * allocated. If end == LONG_MAX, it will not fail. | |
6628 | */ | |
6629 | if (chg < 0) | |
6630 | return chg; | |
6631 | } | |
6632 | ||
45c682a6 | 6633 | spin_lock(&inode->i_lock); |
e4c6f8be | 6634 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
6635 | spin_unlock(&inode->i_lock); |
6636 | ||
1c5ecae3 MK |
6637 | /* |
6638 | * If the subpool has a minimum size, the number of global | |
6639 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
6640 | * |
6641 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
6642 | * won't go negative. | |
1c5ecae3 MK |
6643 | */ |
6644 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
6645 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
6646 | |
6647 | return 0; | |
a43a8c39 | 6648 | } |
93f70f90 | 6649 | |
3212b535 SC |
6650 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
6651 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
6652 | struct vm_area_struct *vma, | |
6653 | unsigned long addr, pgoff_t idx) | |
6654 | { | |
6655 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
6656 | svma->vm_start; | |
6657 | unsigned long sbase = saddr & PUD_MASK; | |
6658 | unsigned long s_end = sbase + PUD_SIZE; | |
6659 | ||
6660 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
6661 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
6662 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
6663 | |
6664 | /* | |
6665 | * match the virtual addresses, permission and the alignment of the | |
6666 | * page table page. | |
6667 | */ | |
6668 | if (pmd_index(addr) != pmd_index(saddr) || | |
6669 | vm_flags != svm_flags || | |
07e51edf | 6670 | !range_in_vma(svma, sbase, s_end)) |
3212b535 SC |
6671 | return 0; |
6672 | ||
6673 | return saddr; | |
6674 | } | |
6675 | ||
12710fd6 | 6676 | static bool __vma_aligned_range_pmd_shareable(struct vm_area_struct *vma, |
8d9bfb26 MK |
6677 | unsigned long start, unsigned long end, |
6678 | bool check_vma_lock) | |
3212b535 | 6679 | { |
8d9bfb26 MK |
6680 | #ifdef CONFIG_USERFAULTFD |
6681 | if (uffd_disable_huge_pmd_share(vma)) | |
6682 | return false; | |
6683 | #endif | |
3212b535 SC |
6684 | /* |
6685 | * check on proper vm_flags and page table alignment | |
6686 | */ | |
8d9bfb26 MK |
6687 | if (!(vma->vm_flags & VM_MAYSHARE)) |
6688 | return false; | |
6689 | if (check_vma_lock && !vma->vm_private_data) | |
6690 | return false; | |
6691 | if (!range_in_vma(vma, start, end)) | |
6692 | return false; | |
6693 | return true; | |
6694 | } | |
6695 | ||
6696 | static bool vma_pmd_shareable(struct vm_area_struct *vma) | |
6697 | { | |
6698 | unsigned long start = ALIGN(vma->vm_start, PUD_SIZE), | |
6699 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
6700 | ||
6701 | if (start >= end) | |
6702 | return false; | |
6703 | ||
6704 | return __vma_aligned_range_pmd_shareable(vma, start, end, false); | |
3212b535 SC |
6705 | } |
6706 | ||
12710fd6 MK |
6707 | static bool vma_addr_pmd_shareable(struct vm_area_struct *vma, |
6708 | unsigned long addr) | |
6709 | { | |
6710 | unsigned long start = addr & PUD_MASK; | |
6711 | unsigned long end = start + PUD_SIZE; | |
6712 | ||
8d9bfb26 | 6713 | return __vma_aligned_range_pmd_shareable(vma, start, end, true); |
12710fd6 MK |
6714 | } |
6715 | ||
c1991e07 PX |
6716 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
6717 | { | |
12710fd6 | 6718 | return vma_addr_pmd_shareable(vma, addr); |
c1991e07 PX |
6719 | } |
6720 | ||
017b1660 MK |
6721 | /* |
6722 | * Determine if start,end range within vma could be mapped by shared pmd. | |
6723 | * If yes, adjust start and end to cover range associated with possible | |
6724 | * shared pmd mappings. | |
6725 | */ | |
6726 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6727 | unsigned long *start, unsigned long *end) | |
6728 | { | |
a1ba9da8 LX |
6729 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
6730 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 6731 | |
a1ba9da8 | 6732 | /* |
f0953a1b IM |
6733 | * vma needs to span at least one aligned PUD size, and the range |
6734 | * must be at least partially within in. | |
a1ba9da8 LX |
6735 | */ |
6736 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
6737 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
6738 | return; |
6739 | ||
75802ca6 | 6740 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
6741 | if (*start > v_start) |
6742 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 6743 | |
a1ba9da8 LX |
6744 | if (*end < v_end) |
6745 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
6746 | } |
6747 | ||
8d9bfb26 MK |
6748 | static bool __vma_shareable_flags_pmd(struct vm_area_struct *vma) |
6749 | { | |
6750 | return vma->vm_flags & (VM_MAYSHARE | VM_SHARED) && | |
6751 | vma->vm_private_data; | |
6752 | } | |
6753 | ||
6754 | void hugetlb_vma_lock_read(struct vm_area_struct *vma) | |
6755 | { | |
6756 | if (__vma_shareable_flags_pmd(vma)) { | |
6757 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6758 | ||
6759 | down_read(&vma_lock->rw_sema); | |
6760 | } | |
6761 | } | |
6762 | ||
6763 | void hugetlb_vma_unlock_read(struct vm_area_struct *vma) | |
6764 | { | |
6765 | if (__vma_shareable_flags_pmd(vma)) { | |
6766 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6767 | ||
6768 | up_read(&vma_lock->rw_sema); | |
6769 | } | |
6770 | } | |
6771 | ||
6772 | void hugetlb_vma_lock_write(struct vm_area_struct *vma) | |
6773 | { | |
6774 | if (__vma_shareable_flags_pmd(vma)) { | |
6775 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6776 | ||
6777 | down_write(&vma_lock->rw_sema); | |
6778 | } | |
6779 | } | |
6780 | ||
6781 | void hugetlb_vma_unlock_write(struct vm_area_struct *vma) | |
6782 | { | |
6783 | if (__vma_shareable_flags_pmd(vma)) { | |
6784 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6785 | ||
6786 | up_write(&vma_lock->rw_sema); | |
6787 | } | |
6788 | } | |
6789 | ||
6790 | int hugetlb_vma_trylock_write(struct vm_area_struct *vma) | |
6791 | { | |
6792 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6793 | ||
6794 | if (!__vma_shareable_flags_pmd(vma)) | |
6795 | return 1; | |
6796 | ||
6797 | return down_write_trylock(&vma_lock->rw_sema); | |
6798 | } | |
6799 | ||
6800 | void hugetlb_vma_assert_locked(struct vm_area_struct *vma) | |
6801 | { | |
6802 | if (__vma_shareable_flags_pmd(vma)) { | |
6803 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6804 | ||
6805 | lockdep_assert_held(&vma_lock->rw_sema); | |
6806 | } | |
6807 | } | |
6808 | ||
6809 | void hugetlb_vma_lock_release(struct kref *kref) | |
6810 | { | |
6811 | struct hugetlb_vma_lock *vma_lock = container_of(kref, | |
6812 | struct hugetlb_vma_lock, refs); | |
6813 | ||
6814 | kfree(vma_lock); | |
6815 | } | |
6816 | ||
6817 | static void hugetlb_vma_lock_free(struct vm_area_struct *vma) | |
6818 | { | |
6819 | /* | |
6820 | * Only present in sharable vmas. See comment in | |
6821 | * __unmap_hugepage_range_final about how VM_SHARED could | |
6822 | * be set without VM_MAYSHARE. As a result, we need to | |
6823 | * check if either is set in the free path. | |
6824 | */ | |
6825 | if (!vma || !(vma->vm_flags & (VM_MAYSHARE | VM_SHARED))) | |
6826 | return; | |
6827 | ||
6828 | if (vma->vm_private_data) { | |
6829 | struct hugetlb_vma_lock *vma_lock = vma->vm_private_data; | |
6830 | ||
6831 | /* | |
6832 | * vma_lock structure may or not be released, but it | |
6833 | * certainly will no longer be attached to vma so clear | |
6834 | * pointer. | |
6835 | */ | |
6836 | vma_lock->vma = NULL; | |
6837 | kref_put(&vma_lock->refs, hugetlb_vma_lock_release); | |
6838 | vma->vm_private_data = NULL; | |
6839 | } | |
6840 | } | |
6841 | ||
6842 | static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) | |
6843 | { | |
6844 | struct hugetlb_vma_lock *vma_lock; | |
6845 | ||
6846 | /* Only establish in (flags) sharable vmas */ | |
6847 | if (!vma || !(vma->vm_flags & VM_MAYSHARE)) | |
6848 | return; | |
6849 | ||
6850 | /* Should never get here with non-NULL vm_private_data */ | |
6851 | if (vma->vm_private_data) | |
6852 | return; | |
6853 | ||
6854 | /* Check size/alignment for pmd sharing possible */ | |
6855 | if (!vma_pmd_shareable(vma)) | |
6856 | return; | |
6857 | ||
6858 | vma_lock = kmalloc(sizeof(*vma_lock), GFP_KERNEL); | |
6859 | if (!vma_lock) | |
6860 | /* | |
6861 | * If we can not allocate structure, then vma can not | |
6862 | * participate in pmd sharing. | |
6863 | */ | |
6864 | return; | |
6865 | ||
6866 | kref_init(&vma_lock->refs); | |
6867 | init_rwsem(&vma_lock->rw_sema); | |
6868 | vma_lock->vma = vma; | |
6869 | vma->vm_private_data = vma_lock; | |
6870 | } | |
6871 | ||
3212b535 SC |
6872 | /* |
6873 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
6874 | * and returns the corresponding pte. While this is not necessary for the | |
6875 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
3a47c54f MK |
6876 | * code much cleaner. pmd allocation is essential for the shared case because |
6877 | * pud has to be populated inside the same i_mmap_rwsem section - otherwise | |
6878 | * racing tasks could either miss the sharing (see huge_pte_offset) or select a | |
6879 | * bad pmd for sharing. | |
3212b535 | 6880 | */ |
aec44e0f PX |
6881 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6882 | unsigned long addr, pud_t *pud) | |
3212b535 | 6883 | { |
3212b535 SC |
6884 | struct address_space *mapping = vma->vm_file->f_mapping; |
6885 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
6886 | vma->vm_pgoff; | |
6887 | struct vm_area_struct *svma; | |
6888 | unsigned long saddr; | |
6889 | pte_t *spte = NULL; | |
6890 | pte_t *pte; | |
cb900f41 | 6891 | spinlock_t *ptl; |
3212b535 | 6892 | |
3a47c54f | 6893 | i_mmap_lock_read(mapping); |
3212b535 SC |
6894 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
6895 | if (svma == vma) | |
6896 | continue; | |
6897 | ||
6898 | saddr = page_table_shareable(svma, vma, addr, idx); | |
6899 | if (saddr) { | |
7868a208 PA |
6900 | spte = huge_pte_offset(svma->vm_mm, saddr, |
6901 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
6902 | if (spte) { |
6903 | get_page(virt_to_page(spte)); | |
6904 | break; | |
6905 | } | |
6906 | } | |
6907 | } | |
6908 | ||
6909 | if (!spte) | |
6910 | goto out; | |
6911 | ||
8bea8052 | 6912 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 6913 | if (pud_none(*pud)) { |
3212b535 SC |
6914 | pud_populate(mm, pud, |
6915 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 6916 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 6917 | } else { |
3212b535 | 6918 | put_page(virt_to_page(spte)); |
dc6c9a35 | 6919 | } |
cb900f41 | 6920 | spin_unlock(ptl); |
3212b535 SC |
6921 | out: |
6922 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3a47c54f | 6923 | i_mmap_unlock_read(mapping); |
3212b535 SC |
6924 | return pte; |
6925 | } | |
6926 | ||
6927 | /* | |
6928 | * unmap huge page backed by shared pte. | |
6929 | * | |
6930 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
6931 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
6932 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
6933 | * | |
3a47c54f | 6934 | * Called with page table lock held. |
3212b535 SC |
6935 | * |
6936 | * returns: 1 successfully unmapped a shared pte page | |
6937 | * 0 the underlying pte page is not shared, or it is the last user | |
6938 | */ | |
34ae204f | 6939 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
4ddb4d91 | 6940 | unsigned long addr, pte_t *ptep) |
3212b535 | 6941 | { |
4ddb4d91 MK |
6942 | pgd_t *pgd = pgd_offset(mm, addr); |
6943 | p4d_t *p4d = p4d_offset(pgd, addr); | |
6944 | pud_t *pud = pud_offset(p4d, addr); | |
3212b535 | 6945 | |
34ae204f | 6946 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
40549ba8 | 6947 | hugetlb_vma_assert_locked(vma); |
3212b535 SC |
6948 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
6949 | if (page_count(virt_to_page(ptep)) == 1) | |
6950 | return 0; | |
6951 | ||
6952 | pud_clear(pud); | |
6953 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 6954 | mm_dec_nr_pmds(mm); |
3212b535 SC |
6955 | return 1; |
6956 | } | |
c1991e07 | 6957 | |
9e5fc74c | 6958 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
8d9bfb26 | 6959 | |
40549ba8 MK |
6960 | void hugetlb_vma_lock_read(struct vm_area_struct *vma) |
6961 | { | |
6962 | } | |
6963 | ||
6964 | void hugetlb_vma_unlock_read(struct vm_area_struct *vma) | |
6965 | { | |
6966 | } | |
6967 | ||
6968 | void hugetlb_vma_lock_write(struct vm_area_struct *vma) | |
6969 | { | |
6970 | } | |
6971 | ||
6972 | void hugetlb_vma_unlock_write(struct vm_area_struct *vma) | |
6973 | { | |
6974 | } | |
6975 | ||
6976 | int hugetlb_vma_trylock_write(struct vm_area_struct *vma) | |
6977 | { | |
6978 | return 1; | |
6979 | } | |
6980 | ||
6981 | void hugetlb_vma_assert_locked(struct vm_area_struct *vma) | |
6982 | { | |
6983 | } | |
6984 | ||
8d9bfb26 MK |
6985 | void hugetlb_vma_lock_release(struct kref *kref) |
6986 | { | |
6987 | } | |
6988 | ||
6989 | static void hugetlb_vma_lock_free(struct vm_area_struct *vma) | |
6990 | { | |
6991 | } | |
6992 | ||
6993 | static void hugetlb_vma_lock_alloc(struct vm_area_struct *vma) | |
6994 | { | |
6995 | } | |
6996 | ||
aec44e0f PX |
6997 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6998 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
6999 | { |
7000 | return NULL; | |
7001 | } | |
e81f2d22 | 7002 | |
34ae204f | 7003 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
4ddb4d91 | 7004 | unsigned long addr, pte_t *ptep) |
e81f2d22 ZZ |
7005 | { |
7006 | return 0; | |
7007 | } | |
017b1660 MK |
7008 | |
7009 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
7010 | unsigned long *start, unsigned long *end) | |
7011 | { | |
7012 | } | |
c1991e07 PX |
7013 | |
7014 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
7015 | { | |
7016 | return false; | |
7017 | } | |
3212b535 SC |
7018 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
7019 | ||
9e5fc74c | 7020 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 7021 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
7022 | unsigned long addr, unsigned long sz) |
7023 | { | |
7024 | pgd_t *pgd; | |
c2febafc | 7025 | p4d_t *p4d; |
9e5fc74c SC |
7026 | pud_t *pud; |
7027 | pte_t *pte = NULL; | |
7028 | ||
7029 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
7030 | p4d = p4d_alloc(mm, pgd, addr); |
7031 | if (!p4d) | |
7032 | return NULL; | |
c2febafc | 7033 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
7034 | if (pud) { |
7035 | if (sz == PUD_SIZE) { | |
7036 | pte = (pte_t *)pud; | |
7037 | } else { | |
7038 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 7039 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 7040 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
7041 | else |
7042 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
7043 | } | |
7044 | } | |
4e666314 | 7045 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
7046 | |
7047 | return pte; | |
7048 | } | |
7049 | ||
9b19df29 PA |
7050 | /* |
7051 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
7052 | * entry at address @addr | |
7053 | * | |
8ac0b81a LX |
7054 | * Return: Pointer to page table entry (PUD or PMD) for |
7055 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
7056 | * size @sz doesn't match the hugepage size at this level of the page |
7057 | * table. | |
7058 | */ | |
7868a208 PA |
7059 | pte_t *huge_pte_offset(struct mm_struct *mm, |
7060 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
7061 | { |
7062 | pgd_t *pgd; | |
c2febafc | 7063 | p4d_t *p4d; |
8ac0b81a LX |
7064 | pud_t *pud; |
7065 | pmd_t *pmd; | |
9e5fc74c SC |
7066 | |
7067 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
7068 | if (!pgd_present(*pgd)) |
7069 | return NULL; | |
7070 | p4d = p4d_offset(pgd, addr); | |
7071 | if (!p4d_present(*p4d)) | |
7072 | return NULL; | |
9b19df29 | 7073 | |
c2febafc | 7074 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
7075 | if (sz == PUD_SIZE) |
7076 | /* must be pud huge, non-present or none */ | |
c2febafc | 7077 | return (pte_t *)pud; |
8ac0b81a | 7078 | if (!pud_present(*pud)) |
9b19df29 | 7079 | return NULL; |
8ac0b81a | 7080 | /* must have a valid entry and size to go further */ |
9b19df29 | 7081 | |
8ac0b81a LX |
7082 | pmd = pmd_offset(pud, addr); |
7083 | /* must be pmd huge, non-present or none */ | |
7084 | return (pte_t *)pmd; | |
9e5fc74c SC |
7085 | } |
7086 | ||
e95a9851 MK |
7087 | /* |
7088 | * Return a mask that can be used to update an address to the last huge | |
7089 | * page in a page table page mapping size. Used to skip non-present | |
7090 | * page table entries when linearly scanning address ranges. Architectures | |
7091 | * with unique huge page to page table relationships can define their own | |
7092 | * version of this routine. | |
7093 | */ | |
7094 | unsigned long hugetlb_mask_last_page(struct hstate *h) | |
7095 | { | |
7096 | unsigned long hp_size = huge_page_size(h); | |
7097 | ||
7098 | if (hp_size == PUD_SIZE) | |
7099 | return P4D_SIZE - PUD_SIZE; | |
7100 | else if (hp_size == PMD_SIZE) | |
7101 | return PUD_SIZE - PMD_SIZE; | |
7102 | else | |
7103 | return 0UL; | |
7104 | } | |
7105 | ||
7106 | #else | |
7107 | ||
7108 | /* See description above. Architectures can provide their own version. */ | |
7109 | __weak unsigned long hugetlb_mask_last_page(struct hstate *h) | |
7110 | { | |
4ddb4d91 MK |
7111 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
7112 | if (huge_page_size(h) == PMD_SIZE) | |
7113 | return PUD_SIZE - PMD_SIZE; | |
7114 | #endif | |
e95a9851 MK |
7115 | return 0UL; |
7116 | } | |
7117 | ||
61f77eda NH |
7118 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
7119 | ||
7120 | /* | |
7121 | * These functions are overwritable if your architecture needs its own | |
7122 | * behavior. | |
7123 | */ | |
7124 | struct page * __weak | |
7125 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
7126 | int write) | |
7127 | { | |
7128 | return ERR_PTR(-EINVAL); | |
7129 | } | |
7130 | ||
4dc71451 AK |
7131 | struct page * __weak |
7132 | follow_huge_pd(struct vm_area_struct *vma, | |
7133 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
7134 | { | |
7135 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
7136 | return NULL; | |
7137 | } | |
7138 | ||
61f77eda | 7139 | struct page * __weak |
9e5fc74c | 7140 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 7141 | pmd_t *pmd, int flags) |
9e5fc74c | 7142 | { |
e66f17ff NH |
7143 | struct page *page = NULL; |
7144 | spinlock_t *ptl; | |
c9d398fa | 7145 | pte_t pte; |
3faa52c0 | 7146 | |
8909691b DH |
7147 | /* |
7148 | * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via | |
7149 | * follow_hugetlb_page(). | |
7150 | */ | |
7151 | if (WARN_ON_ONCE(flags & FOLL_PIN)) | |
3faa52c0 JH |
7152 | return NULL; |
7153 | ||
e66f17ff NH |
7154 | retry: |
7155 | ptl = pmd_lockptr(mm, pmd); | |
7156 | spin_lock(ptl); | |
7157 | /* | |
7158 | * make sure that the address range covered by this pmd is not | |
7159 | * unmapped from other threads. | |
7160 | */ | |
7161 | if (!pmd_huge(*pmd)) | |
7162 | goto out; | |
c9d398fa NH |
7163 | pte = huge_ptep_get((pte_t *)pmd); |
7164 | if (pte_present(pte)) { | |
97534127 | 7165 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
7166 | /* |
7167 | * try_grab_page() should always succeed here, because: a) we | |
7168 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
7169 | * huge pmd (head) page is present in the page tables. The ptl | |
7170 | * prevents the head page and tail pages from being rearranged | |
7171 | * in any way. So this page must be available at this point, | |
7172 | * unless the page refcount overflowed: | |
7173 | */ | |
7174 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
7175 | page = NULL; | |
7176 | goto out; | |
7177 | } | |
e66f17ff | 7178 | } else { |
c9d398fa | 7179 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff | 7180 | spin_unlock(ptl); |
ad1ac596 | 7181 | __migration_entry_wait_huge((pte_t *)pmd, ptl); |
e66f17ff NH |
7182 | goto retry; |
7183 | } | |
7184 | /* | |
7185 | * hwpoisoned entry is treated as no_page_table in | |
7186 | * follow_page_mask(). | |
7187 | */ | |
7188 | } | |
7189 | out: | |
7190 | spin_unlock(ptl); | |
9e5fc74c SC |
7191 | return page; |
7192 | } | |
7193 | ||
61f77eda | 7194 | struct page * __weak |
9e5fc74c | 7195 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 7196 | pud_t *pud, int flags) |
9e5fc74c | 7197 | { |
3a194f3f NH |
7198 | struct page *page = NULL; |
7199 | spinlock_t *ptl; | |
7200 | pte_t pte; | |
7201 | ||
7202 | if (WARN_ON_ONCE(flags & FOLL_PIN)) | |
e66f17ff | 7203 | return NULL; |
9e5fc74c | 7204 | |
3a194f3f NH |
7205 | retry: |
7206 | ptl = huge_pte_lock(hstate_sizelog(PUD_SHIFT), mm, (pte_t *)pud); | |
7207 | if (!pud_huge(*pud)) | |
7208 | goto out; | |
7209 | pte = huge_ptep_get((pte_t *)pud); | |
7210 | if (pte_present(pte)) { | |
7211 | page = pud_page(*pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); | |
7212 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
7213 | page = NULL; | |
7214 | goto out; | |
7215 | } | |
7216 | } else { | |
7217 | if (is_hugetlb_entry_migration(pte)) { | |
7218 | spin_unlock(ptl); | |
7219 | __migration_entry_wait(mm, (pte_t *)pud, ptl); | |
7220 | goto retry; | |
7221 | } | |
7222 | /* | |
7223 | * hwpoisoned entry is treated as no_page_table in | |
7224 | * follow_page_mask(). | |
7225 | */ | |
7226 | } | |
7227 | out: | |
7228 | spin_unlock(ptl); | |
7229 | return page; | |
9e5fc74c SC |
7230 | } |
7231 | ||
faaa5b62 AK |
7232 | struct page * __weak |
7233 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
7234 | { | |
3faa52c0 | 7235 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
7236 | return NULL; |
7237 | ||
7238 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
7239 | } | |
7240 | ||
7ce82f4c | 7241 | int isolate_hugetlb(struct page *page, struct list_head *list) |
31caf665 | 7242 | { |
7ce82f4c | 7243 | int ret = 0; |
bcc54222 | 7244 | |
db71ef79 | 7245 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
7246 | if (!PageHeadHuge(page) || |
7247 | !HPageMigratable(page) || | |
0eb2df2b | 7248 | !get_page_unless_zero(page)) { |
7ce82f4c | 7249 | ret = -EBUSY; |
bcc54222 NH |
7250 | goto unlock; |
7251 | } | |
8f251a3d | 7252 | ClearHPageMigratable(page); |
31caf665 | 7253 | list_move_tail(&page->lru, list); |
bcc54222 | 7254 | unlock: |
db71ef79 | 7255 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 7256 | return ret; |
31caf665 NH |
7257 | } |
7258 | ||
25182f05 NH |
7259 | int get_hwpoison_huge_page(struct page *page, bool *hugetlb) |
7260 | { | |
7261 | int ret = 0; | |
7262 | ||
7263 | *hugetlb = false; | |
7264 | spin_lock_irq(&hugetlb_lock); | |
7265 | if (PageHeadHuge(page)) { | |
7266 | *hugetlb = true; | |
b283d983 NH |
7267 | if (HPageFreed(page)) |
7268 | ret = 0; | |
7269 | else if (HPageMigratable(page)) | |
25182f05 | 7270 | ret = get_page_unless_zero(page); |
0ed950d1 NH |
7271 | else |
7272 | ret = -EBUSY; | |
25182f05 NH |
7273 | } |
7274 | spin_unlock_irq(&hugetlb_lock); | |
7275 | return ret; | |
7276 | } | |
7277 | ||
405ce051 NH |
7278 | int get_huge_page_for_hwpoison(unsigned long pfn, int flags) |
7279 | { | |
7280 | int ret; | |
7281 | ||
7282 | spin_lock_irq(&hugetlb_lock); | |
7283 | ret = __get_huge_page_for_hwpoison(pfn, flags); | |
7284 | spin_unlock_irq(&hugetlb_lock); | |
7285 | return ret; | |
7286 | } | |
7287 | ||
31caf665 NH |
7288 | void putback_active_hugepage(struct page *page) |
7289 | { | |
db71ef79 | 7290 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 7291 | SetHPageMigratable(page); |
31caf665 | 7292 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 7293 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
7294 | put_page(page); |
7295 | } | |
ab5ac90a MH |
7296 | |
7297 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
7298 | { | |
7299 | struct hstate *h = page_hstate(oldpage); | |
7300 | ||
7301 | hugetlb_cgroup_migrate(oldpage, newpage); | |
7302 | set_page_owner_migrate_reason(newpage, reason); | |
7303 | ||
7304 | /* | |
7305 | * transfer temporary state of the new huge page. This is | |
7306 | * reverse to other transitions because the newpage is going to | |
7307 | * be final while the old one will be freed so it takes over | |
7308 | * the temporary status. | |
7309 | * | |
7310 | * Also note that we have to transfer the per-node surplus state | |
7311 | * here as well otherwise the global surplus count will not match | |
7312 | * the per-node's. | |
7313 | */ | |
9157c311 | 7314 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
7315 | int old_nid = page_to_nid(oldpage); |
7316 | int new_nid = page_to_nid(newpage); | |
7317 | ||
9157c311 MK |
7318 | SetHPageTemporary(oldpage); |
7319 | ClearHPageTemporary(newpage); | |
ab5ac90a | 7320 | |
5af1ab1d ML |
7321 | /* |
7322 | * There is no need to transfer the per-node surplus state | |
7323 | * when we do not cross the node. | |
7324 | */ | |
7325 | if (new_nid == old_nid) | |
7326 | return; | |
db71ef79 | 7327 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
7328 | if (h->surplus_huge_pages_node[old_nid]) { |
7329 | h->surplus_huge_pages_node[old_nid]--; | |
7330 | h->surplus_huge_pages_node[new_nid]++; | |
7331 | } | |
db71ef79 | 7332 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
7333 | } |
7334 | } | |
cf11e85f | 7335 | |
6dfeaff9 PX |
7336 | /* |
7337 | * This function will unconditionally remove all the shared pmd pgtable entries | |
7338 | * within the specific vma for a hugetlbfs memory range. | |
7339 | */ | |
7340 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
7341 | { | |
7342 | struct hstate *h = hstate_vma(vma); | |
7343 | unsigned long sz = huge_page_size(h); | |
7344 | struct mm_struct *mm = vma->vm_mm; | |
7345 | struct mmu_notifier_range range; | |
7346 | unsigned long address, start, end; | |
7347 | spinlock_t *ptl; | |
7348 | pte_t *ptep; | |
7349 | ||
7350 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
7351 | return; | |
7352 | ||
7353 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
7354 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
7355 | ||
7356 | if (start >= end) | |
7357 | return; | |
7358 | ||
9c8bbfac | 7359 | flush_cache_range(vma, start, end); |
6dfeaff9 PX |
7360 | /* |
7361 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
7362 | * we have already done the PUD_SIZE alignment. | |
7363 | */ | |
7364 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
7365 | start, end); | |
7366 | mmu_notifier_invalidate_range_start(&range); | |
40549ba8 | 7367 | hugetlb_vma_lock_write(vma); |
6dfeaff9 PX |
7368 | i_mmap_lock_write(vma->vm_file->f_mapping); |
7369 | for (address = start; address < end; address += PUD_SIZE) { | |
6dfeaff9 PX |
7370 | ptep = huge_pte_offset(mm, address, sz); |
7371 | if (!ptep) | |
7372 | continue; | |
7373 | ptl = huge_pte_lock(h, mm, ptep); | |
4ddb4d91 | 7374 | huge_pmd_unshare(mm, vma, address, ptep); |
6dfeaff9 PX |
7375 | spin_unlock(ptl); |
7376 | } | |
7377 | flush_hugetlb_tlb_range(vma, start, end); | |
7378 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
40549ba8 | 7379 | hugetlb_vma_unlock_write(vma); |
6dfeaff9 PX |
7380 | /* |
7381 | * No need to call mmu_notifier_invalidate_range(), see | |
ee65728e | 7382 | * Documentation/mm/mmu_notifier.rst. |
6dfeaff9 PX |
7383 | */ |
7384 | mmu_notifier_invalidate_range_end(&range); | |
7385 | } | |
7386 | ||
cf11e85f | 7387 | #ifdef CONFIG_CMA |
cf11e85f RG |
7388 | static bool cma_reserve_called __initdata; |
7389 | ||
7390 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
7391 | { | |
38e719ab BW |
7392 | int nid, count = 0; |
7393 | unsigned long tmp; | |
7394 | char *s = p; | |
7395 | ||
7396 | while (*s) { | |
7397 | if (sscanf(s, "%lu%n", &tmp, &count) != 1) | |
7398 | break; | |
7399 | ||
7400 | if (s[count] == ':') { | |
f9317f77 | 7401 | if (tmp >= MAX_NUMNODES) |
38e719ab | 7402 | break; |
f9317f77 | 7403 | nid = array_index_nospec(tmp, MAX_NUMNODES); |
38e719ab BW |
7404 | |
7405 | s += count + 1; | |
7406 | tmp = memparse(s, &s); | |
7407 | hugetlb_cma_size_in_node[nid] = tmp; | |
7408 | hugetlb_cma_size += tmp; | |
7409 | ||
7410 | /* | |
7411 | * Skip the separator if have one, otherwise | |
7412 | * break the parsing. | |
7413 | */ | |
7414 | if (*s == ',') | |
7415 | s++; | |
7416 | else | |
7417 | break; | |
7418 | } else { | |
7419 | hugetlb_cma_size = memparse(p, &p); | |
7420 | break; | |
7421 | } | |
7422 | } | |
7423 | ||
cf11e85f RG |
7424 | return 0; |
7425 | } | |
7426 | ||
7427 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
7428 | ||
7429 | void __init hugetlb_cma_reserve(int order) | |
7430 | { | |
7431 | unsigned long size, reserved, per_node; | |
38e719ab | 7432 | bool node_specific_cma_alloc = false; |
cf11e85f RG |
7433 | int nid; |
7434 | ||
7435 | cma_reserve_called = true; | |
7436 | ||
38e719ab BW |
7437 | if (!hugetlb_cma_size) |
7438 | return; | |
7439 | ||
7440 | for (nid = 0; nid < MAX_NUMNODES; nid++) { | |
7441 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7442 | continue; | |
7443 | ||
30a51400 | 7444 | if (!node_online(nid)) { |
38e719ab BW |
7445 | pr_warn("hugetlb_cma: invalid node %d specified\n", nid); |
7446 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7447 | hugetlb_cma_size_in_node[nid] = 0; | |
7448 | continue; | |
7449 | } | |
7450 | ||
7451 | if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { | |
7452 | pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", | |
7453 | nid, (PAGE_SIZE << order) / SZ_1M); | |
7454 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7455 | hugetlb_cma_size_in_node[nid] = 0; | |
7456 | } else { | |
7457 | node_specific_cma_alloc = true; | |
7458 | } | |
7459 | } | |
7460 | ||
7461 | /* Validate the CMA size again in case some invalid nodes specified. */ | |
cf11e85f RG |
7462 | if (!hugetlb_cma_size) |
7463 | return; | |
7464 | ||
7465 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
7466 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
7467 | (PAGE_SIZE << order) / SZ_1M); | |
a01f4390 | 7468 | hugetlb_cma_size = 0; |
cf11e85f RG |
7469 | return; |
7470 | } | |
7471 | ||
38e719ab BW |
7472 | if (!node_specific_cma_alloc) { |
7473 | /* | |
7474 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
7475 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
7476 | */ | |
7477 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
7478 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
7479 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
7480 | } | |
cf11e85f RG |
7481 | |
7482 | reserved = 0; | |
30a51400 | 7483 | for_each_online_node(nid) { |
cf11e85f | 7484 | int res; |
2281f797 | 7485 | char name[CMA_MAX_NAME]; |
cf11e85f | 7486 | |
38e719ab BW |
7487 | if (node_specific_cma_alloc) { |
7488 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7489 | continue; | |
7490 | ||
7491 | size = hugetlb_cma_size_in_node[nid]; | |
7492 | } else { | |
7493 | size = min(per_node, hugetlb_cma_size - reserved); | |
7494 | } | |
7495 | ||
cf11e85f RG |
7496 | size = round_up(size, PAGE_SIZE << order); |
7497 | ||
2281f797 | 7498 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
a01f4390 MK |
7499 | /* |
7500 | * Note that 'order per bit' is based on smallest size that | |
7501 | * may be returned to CMA allocator in the case of | |
7502 | * huge page demotion. | |
7503 | */ | |
7504 | res = cma_declare_contiguous_nid(0, size, 0, | |
7505 | PAGE_SIZE << HUGETLB_PAGE_ORDER, | |
29d0f41d | 7506 | 0, false, name, |
cf11e85f RG |
7507 | &hugetlb_cma[nid], nid); |
7508 | if (res) { | |
7509 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
7510 | res, nid); | |
7511 | continue; | |
7512 | } | |
7513 | ||
7514 | reserved += size; | |
7515 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
7516 | size / SZ_1M, nid); | |
7517 | ||
7518 | if (reserved >= hugetlb_cma_size) | |
7519 | break; | |
7520 | } | |
a01f4390 MK |
7521 | |
7522 | if (!reserved) | |
7523 | /* | |
7524 | * hugetlb_cma_size is used to determine if allocations from | |
7525 | * cma are possible. Set to zero if no cma regions are set up. | |
7526 | */ | |
7527 | hugetlb_cma_size = 0; | |
cf11e85f RG |
7528 | } |
7529 | ||
263b8998 | 7530 | static void __init hugetlb_cma_check(void) |
cf11e85f RG |
7531 | { |
7532 | if (!hugetlb_cma_size || cma_reserve_called) | |
7533 | return; | |
7534 | ||
7535 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
7536 | } | |
7537 | ||
7538 | #endif /* CONFIG_CMA */ |