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> |
d6606683 | 35 | |
63551ae0 | 36 | #include <asm/page.h> |
ca15ca40 | 37 | #include <asm/pgalloc.h> |
24669e58 | 38 | #include <asm/tlb.h> |
63551ae0 | 39 | |
24669e58 | 40 | #include <linux/io.h> |
63551ae0 | 41 | #include <linux/hugetlb.h> |
9dd540e2 | 42 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 43 | #include <linux/node.h> |
ab5ac90a | 44 | #include <linux/page_owner.h> |
7835e98b | 45 | #include "internal.h" |
f41f2ed4 | 46 | #include "hugetlb_vmemmap.h" |
1da177e4 | 47 | |
c3f38a38 | 48 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
49 | unsigned int default_hstate_idx; |
50 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 51 | |
dbda8fea | 52 | #ifdef CONFIG_CMA |
cf11e85f | 53 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
38e719ab | 54 | static unsigned long hugetlb_cma_size_in_node[MAX_NUMNODES] __initdata; |
a01f4390 MK |
55 | static bool hugetlb_cma_page(struct page *page, unsigned int order) |
56 | { | |
57 | return cma_pages_valid(hugetlb_cma[page_to_nid(page)], page, | |
58 | 1 << order); | |
59 | } | |
60 | #else | |
61 | static bool hugetlb_cma_page(struct page *page, unsigned int order) | |
62 | { | |
63 | return false; | |
64 | } | |
dbda8fea BS |
65 | #endif |
66 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 67 | |
641844f5 NH |
68 | /* |
69 | * Minimum page order among possible hugepage sizes, set to a proper value | |
70 | * at boot time. | |
71 | */ | |
72 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 73 | |
53ba51d2 JT |
74 | __initdata LIST_HEAD(huge_boot_pages); |
75 | ||
e5ff2159 AK |
76 | /* for command line parsing */ |
77 | static struct hstate * __initdata parsed_hstate; | |
78 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 79 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 80 | static bool __initdata parsed_default_hugepagesz; |
b5389086 | 81 | static unsigned int default_hugepages_in_node[MAX_NUMNODES] __initdata; |
e5ff2159 | 82 | |
3935baa9 | 83 | /* |
31caf665 NH |
84 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
85 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 86 | */ |
c3f38a38 | 87 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 88 | |
8382d914 DB |
89 | /* |
90 | * Serializes faults on the same logical page. This is used to | |
91 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
92 | */ | |
93 | static int num_fault_mutexes; | |
c672c7f2 | 94 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 95 | |
7ca02d0a MK |
96 | /* Forward declaration */ |
97 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
98 | ||
1d88433b | 99 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 100 | { |
1d88433b ML |
101 | if (spool->count) |
102 | return false; | |
103 | if (spool->max_hpages != -1) | |
104 | return spool->used_hpages == 0; | |
105 | if (spool->min_hpages != -1) | |
106 | return spool->rsv_hpages == spool->min_hpages; | |
107 | ||
108 | return true; | |
109 | } | |
90481622 | 110 | |
db71ef79 MK |
111 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, |
112 | unsigned long irq_flags) | |
1d88433b | 113 | { |
db71ef79 | 114 | spin_unlock_irqrestore(&spool->lock, irq_flags); |
90481622 DG |
115 | |
116 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 117 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 118 | * free the subpool */ |
1d88433b | 119 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
120 | if (spool->min_hpages != -1) |
121 | hugetlb_acct_memory(spool->hstate, | |
122 | -spool->min_hpages); | |
90481622 | 123 | kfree(spool); |
7ca02d0a | 124 | } |
90481622 DG |
125 | } |
126 | ||
7ca02d0a MK |
127 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
128 | long min_hpages) | |
90481622 DG |
129 | { |
130 | struct hugepage_subpool *spool; | |
131 | ||
c6a91820 | 132 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
133 | if (!spool) |
134 | return NULL; | |
135 | ||
136 | spin_lock_init(&spool->lock); | |
137 | spool->count = 1; | |
7ca02d0a MK |
138 | spool->max_hpages = max_hpages; |
139 | spool->hstate = h; | |
140 | spool->min_hpages = min_hpages; | |
141 | ||
142 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
143 | kfree(spool); | |
144 | return NULL; | |
145 | } | |
146 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
147 | |
148 | return spool; | |
149 | } | |
150 | ||
151 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
152 | { | |
db71ef79 MK |
153 | unsigned long flags; |
154 | ||
155 | spin_lock_irqsave(&spool->lock, flags); | |
90481622 DG |
156 | BUG_ON(!spool->count); |
157 | spool->count--; | |
db71ef79 | 158 | unlock_or_release_subpool(spool, flags); |
90481622 DG |
159 | } |
160 | ||
1c5ecae3 MK |
161 | /* |
162 | * Subpool accounting for allocating and reserving pages. | |
163 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 164 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
165 | * global pools must be adjusted (upward). The returned value may |
166 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 167 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
168 | */ |
169 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
170 | long delta) |
171 | { | |
1c5ecae3 | 172 | long ret = delta; |
90481622 DG |
173 | |
174 | if (!spool) | |
1c5ecae3 | 175 | return ret; |
90481622 | 176 | |
db71ef79 | 177 | spin_lock_irq(&spool->lock); |
1c5ecae3 MK |
178 | |
179 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
180 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
181 | spool->used_hpages += delta; | |
182 | else { | |
183 | ret = -ENOMEM; | |
184 | goto unlock_ret; | |
185 | } | |
90481622 | 186 | } |
90481622 | 187 | |
09a95e29 MK |
188 | /* minimum size accounting */ |
189 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
190 | if (delta > spool->rsv_hpages) { |
191 | /* | |
192 | * Asking for more reserves than those already taken on | |
193 | * behalf of subpool. Return difference. | |
194 | */ | |
195 | ret = delta - spool->rsv_hpages; | |
196 | spool->rsv_hpages = 0; | |
197 | } else { | |
198 | ret = 0; /* reserves already accounted for */ | |
199 | spool->rsv_hpages -= delta; | |
200 | } | |
201 | } | |
202 | ||
203 | unlock_ret: | |
db71ef79 | 204 | spin_unlock_irq(&spool->lock); |
90481622 DG |
205 | return ret; |
206 | } | |
207 | ||
1c5ecae3 MK |
208 | /* |
209 | * Subpool accounting for freeing and unreserving pages. | |
210 | * Return the number of global page reservations that must be dropped. | |
211 | * The return value may only be different than the passed value (delta) | |
212 | * in the case where a subpool minimum size must be maintained. | |
213 | */ | |
214 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
215 | long delta) |
216 | { | |
1c5ecae3 | 217 | long ret = delta; |
db71ef79 | 218 | unsigned long flags; |
1c5ecae3 | 219 | |
90481622 | 220 | if (!spool) |
1c5ecae3 | 221 | return delta; |
90481622 | 222 | |
db71ef79 | 223 | spin_lock_irqsave(&spool->lock, flags); |
1c5ecae3 MK |
224 | |
225 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
226 | spool->used_hpages -= delta; | |
227 | ||
09a95e29 MK |
228 | /* minimum size accounting */ |
229 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
230 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
231 | ret = 0; | |
232 | else | |
233 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
234 | ||
235 | spool->rsv_hpages += delta; | |
236 | if (spool->rsv_hpages > spool->min_hpages) | |
237 | spool->rsv_hpages = spool->min_hpages; | |
238 | } | |
239 | ||
240 | /* | |
241 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
242 | * quota reference, free it now. | |
243 | */ | |
db71ef79 | 244 | unlock_or_release_subpool(spool, flags); |
1c5ecae3 MK |
245 | |
246 | return ret; | |
90481622 DG |
247 | } |
248 | ||
249 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
250 | { | |
251 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
252 | } | |
253 | ||
254 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
255 | { | |
496ad9aa | 256 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
257 | } |
258 | ||
0db9d74e MA |
259 | /* Helper that removes a struct file_region from the resv_map cache and returns |
260 | * it for use. | |
261 | */ | |
262 | static struct file_region * | |
263 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
264 | { | |
265 | struct file_region *nrg = NULL; | |
266 | ||
267 | VM_BUG_ON(resv->region_cache_count <= 0); | |
268 | ||
269 | resv->region_cache_count--; | |
270 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
271 | list_del(&nrg->link); |
272 | ||
273 | nrg->from = from; | |
274 | nrg->to = to; | |
275 | ||
276 | return nrg; | |
277 | } | |
278 | ||
075a61d0 MA |
279 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
280 | struct file_region *rg) | |
281 | { | |
282 | #ifdef CONFIG_CGROUP_HUGETLB | |
283 | nrg->reservation_counter = rg->reservation_counter; | |
284 | nrg->css = rg->css; | |
285 | if (rg->css) | |
286 | css_get(rg->css); | |
287 | #endif | |
288 | } | |
289 | ||
290 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
291 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
292 | struct hstate *h, | |
293 | struct resv_map *resv, | |
294 | struct file_region *nrg) | |
295 | { | |
296 | #ifdef CONFIG_CGROUP_HUGETLB | |
297 | if (h_cg) { | |
298 | nrg->reservation_counter = | |
299 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
300 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
301 | /* |
302 | * The caller will hold exactly one h_cg->css reference for the | |
303 | * whole contiguous reservation region. But this area might be | |
304 | * scattered when there are already some file_regions reside in | |
305 | * it. As a result, many file_regions may share only one css | |
306 | * reference. In order to ensure that one file_region must hold | |
307 | * exactly one h_cg->css reference, we should do css_get for | |
308 | * each file_region and leave the reference held by caller | |
309 | * untouched. | |
310 | */ | |
311 | css_get(&h_cg->css); | |
075a61d0 MA |
312 | if (!resv->pages_per_hpage) |
313 | resv->pages_per_hpage = pages_per_huge_page(h); | |
314 | /* pages_per_hpage should be the same for all entries in | |
315 | * a resv_map. | |
316 | */ | |
317 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
318 | } else { | |
319 | nrg->reservation_counter = NULL; | |
320 | nrg->css = NULL; | |
321 | } | |
322 | #endif | |
323 | } | |
324 | ||
d85aecf2 ML |
325 | static void put_uncharge_info(struct file_region *rg) |
326 | { | |
327 | #ifdef CONFIG_CGROUP_HUGETLB | |
328 | if (rg->css) | |
329 | css_put(rg->css); | |
330 | #endif | |
331 | } | |
332 | ||
a9b3f867 MA |
333 | static bool has_same_uncharge_info(struct file_region *rg, |
334 | struct file_region *org) | |
335 | { | |
336 | #ifdef CONFIG_CGROUP_HUGETLB | |
0739eb43 | 337 | return rg->reservation_counter == org->reservation_counter && |
a9b3f867 MA |
338 | rg->css == org->css; |
339 | ||
340 | #else | |
341 | return true; | |
342 | #endif | |
343 | } | |
344 | ||
345 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
346 | { | |
347 | struct file_region *nrg = NULL, *prg = NULL; | |
348 | ||
349 | prg = list_prev_entry(rg, link); | |
350 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
351 | has_same_uncharge_info(prg, rg)) { | |
352 | prg->to = rg->to; | |
353 | ||
354 | list_del(&rg->link); | |
d85aecf2 | 355 | put_uncharge_info(rg); |
a9b3f867 MA |
356 | kfree(rg); |
357 | ||
7db5e7b6 | 358 | rg = prg; |
a9b3f867 MA |
359 | } |
360 | ||
361 | nrg = list_next_entry(rg, link); | |
362 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
363 | has_same_uncharge_info(nrg, rg)) { | |
364 | nrg->from = rg->from; | |
365 | ||
366 | list_del(&rg->link); | |
d85aecf2 | 367 | put_uncharge_info(rg); |
a9b3f867 | 368 | kfree(rg); |
a9b3f867 MA |
369 | } |
370 | } | |
371 | ||
2103cf9c | 372 | static inline long |
84448c8e | 373 | hugetlb_resv_map_add(struct resv_map *map, struct list_head *rg, long from, |
2103cf9c PX |
374 | long to, struct hstate *h, struct hugetlb_cgroup *cg, |
375 | long *regions_needed) | |
376 | { | |
377 | struct file_region *nrg; | |
378 | ||
379 | if (!regions_needed) { | |
380 | nrg = get_file_region_entry_from_cache(map, from, to); | |
381 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
84448c8e | 382 | list_add(&nrg->link, rg); |
2103cf9c PX |
383 | coalesce_file_region(map, nrg); |
384 | } else | |
385 | *regions_needed += 1; | |
386 | ||
387 | return to - from; | |
388 | } | |
389 | ||
972a3da3 WY |
390 | /* |
391 | * Must be called with resv->lock held. | |
392 | * | |
393 | * Calling this with regions_needed != NULL will count the number of pages | |
394 | * to be added but will not modify the linked list. And regions_needed will | |
395 | * indicate the number of file_regions needed in the cache to carry out to add | |
396 | * the regions for this range. | |
d75c6af9 MA |
397 | */ |
398 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 399 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 400 | struct hstate *h, long *regions_needed) |
d75c6af9 | 401 | { |
0db9d74e | 402 | long add = 0; |
d75c6af9 | 403 | struct list_head *head = &resv->regions; |
0db9d74e | 404 | long last_accounted_offset = f; |
84448c8e JK |
405 | struct file_region *iter, *trg = NULL; |
406 | struct list_head *rg = NULL; | |
d75c6af9 | 407 | |
0db9d74e MA |
408 | if (regions_needed) |
409 | *regions_needed = 0; | |
d75c6af9 | 410 | |
0db9d74e | 411 | /* In this loop, we essentially handle an entry for the range |
84448c8e | 412 | * [last_accounted_offset, iter->from), at every iteration, with some |
0db9d74e MA |
413 | * bounds checking. |
414 | */ | |
84448c8e | 415 | list_for_each_entry_safe(iter, trg, head, link) { |
0db9d74e | 416 | /* Skip irrelevant regions that start before our range. */ |
84448c8e | 417 | if (iter->from < f) { |
0db9d74e MA |
418 | /* If this region ends after the last accounted offset, |
419 | * then we need to update last_accounted_offset. | |
420 | */ | |
84448c8e JK |
421 | if (iter->to > last_accounted_offset) |
422 | last_accounted_offset = iter->to; | |
0db9d74e MA |
423 | continue; |
424 | } | |
d75c6af9 | 425 | |
0db9d74e MA |
426 | /* When we find a region that starts beyond our range, we've |
427 | * finished. | |
428 | */ | |
84448c8e JK |
429 | if (iter->from >= t) { |
430 | rg = iter->link.prev; | |
d75c6af9 | 431 | break; |
84448c8e | 432 | } |
d75c6af9 | 433 | |
84448c8e | 434 | /* Add an entry for last_accounted_offset -> iter->from, and |
0db9d74e MA |
435 | * update last_accounted_offset. |
436 | */ | |
84448c8e JK |
437 | if (iter->from > last_accounted_offset) |
438 | add += hugetlb_resv_map_add(resv, iter->link.prev, | |
2103cf9c | 439 | last_accounted_offset, |
84448c8e | 440 | iter->from, h, h_cg, |
2103cf9c | 441 | regions_needed); |
0db9d74e | 442 | |
84448c8e | 443 | last_accounted_offset = iter->to; |
0db9d74e MA |
444 | } |
445 | ||
446 | /* Handle the case where our range extends beyond | |
447 | * last_accounted_offset. | |
448 | */ | |
84448c8e JK |
449 | if (!rg) |
450 | rg = head->prev; | |
2103cf9c PX |
451 | if (last_accounted_offset < t) |
452 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
453 | t, h, h_cg, regions_needed); | |
0db9d74e | 454 | |
0db9d74e MA |
455 | return add; |
456 | } | |
457 | ||
458 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
459 | */ | |
460 | static int allocate_file_region_entries(struct resv_map *resv, | |
461 | int regions_needed) | |
462 | __must_hold(&resv->lock) | |
463 | { | |
464 | struct list_head allocated_regions; | |
465 | int to_allocate = 0, i = 0; | |
466 | struct file_region *trg = NULL, *rg = NULL; | |
467 | ||
468 | VM_BUG_ON(regions_needed < 0); | |
469 | ||
470 | INIT_LIST_HEAD(&allocated_regions); | |
471 | ||
472 | /* | |
473 | * Check for sufficient descriptors in the cache to accommodate | |
474 | * the number of in progress add operations plus regions_needed. | |
475 | * | |
476 | * This is a while loop because when we drop the lock, some other call | |
477 | * to region_add or region_del may have consumed some region_entries, | |
478 | * so we keep looping here until we finally have enough entries for | |
479 | * (adds_in_progress + regions_needed). | |
480 | */ | |
481 | while (resv->region_cache_count < | |
482 | (resv->adds_in_progress + regions_needed)) { | |
483 | to_allocate = resv->adds_in_progress + regions_needed - | |
484 | resv->region_cache_count; | |
485 | ||
486 | /* At this point, we should have enough entries in the cache | |
f0953a1b | 487 | * for all the existing adds_in_progress. We should only be |
0db9d74e | 488 | * needing to allocate for regions_needed. |
d75c6af9 | 489 | */ |
0db9d74e MA |
490 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
491 | ||
492 | spin_unlock(&resv->lock); | |
493 | for (i = 0; i < to_allocate; i++) { | |
494 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
495 | if (!trg) | |
496 | goto out_of_memory; | |
497 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 498 | } |
d75c6af9 | 499 | |
0db9d74e MA |
500 | spin_lock(&resv->lock); |
501 | ||
d3ec7b6e WY |
502 | list_splice(&allocated_regions, &resv->region_cache); |
503 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
504 | } |
505 | ||
0db9d74e | 506 | return 0; |
d75c6af9 | 507 | |
0db9d74e MA |
508 | out_of_memory: |
509 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
510 | list_del(&rg->link); | |
511 | kfree(rg); | |
512 | } | |
513 | return -ENOMEM; | |
d75c6af9 MA |
514 | } |
515 | ||
1dd308a7 MK |
516 | /* |
517 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
518 | * map. Regions will be taken from the cache to fill in this range. |
519 | * Sufficient regions should exist in the cache due to the previous | |
520 | * call to region_chg with the same range, but in some cases the cache will not | |
521 | * have sufficient entries due to races with other code doing region_add or | |
522 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 523 | * |
0db9d74e MA |
524 | * regions_needed is the out value provided by a previous call to region_chg. |
525 | * | |
526 | * Return the number of new huge pages added to the map. This number is greater | |
527 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 528 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
529 | * region_add of regions of length 1 never allocate file_regions and cannot |
530 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
531 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 532 | */ |
0db9d74e | 533 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
534 | long in_regions_needed, struct hstate *h, |
535 | struct hugetlb_cgroup *h_cg) | |
96822904 | 536 | { |
0db9d74e | 537 | long add = 0, actual_regions_needed = 0; |
96822904 | 538 | |
7b24d861 | 539 | spin_lock(&resv->lock); |
0db9d74e MA |
540 | retry: |
541 | ||
542 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
543 | add_reservation_in_range(resv, f, t, NULL, NULL, |
544 | &actual_regions_needed); | |
96822904 | 545 | |
5e911373 | 546 | /* |
0db9d74e MA |
547 | * Check for sufficient descriptors in the cache to accommodate |
548 | * this add operation. Note that actual_regions_needed may be greater | |
549 | * than in_regions_needed, as the resv_map may have been modified since | |
550 | * the region_chg call. In this case, we need to make sure that we | |
551 | * allocate extra entries, such that we have enough for all the | |
552 | * existing adds_in_progress, plus the excess needed for this | |
553 | * operation. | |
5e911373 | 554 | */ |
0db9d74e MA |
555 | if (actual_regions_needed > in_regions_needed && |
556 | resv->region_cache_count < | |
557 | resv->adds_in_progress + | |
558 | (actual_regions_needed - in_regions_needed)) { | |
559 | /* region_add operation of range 1 should never need to | |
560 | * allocate file_region entries. | |
561 | */ | |
562 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 563 | |
0db9d74e MA |
564 | if (allocate_file_region_entries( |
565 | resv, actual_regions_needed - in_regions_needed)) { | |
566 | return -ENOMEM; | |
567 | } | |
5e911373 | 568 | |
0db9d74e | 569 | goto retry; |
5e911373 MK |
570 | } |
571 | ||
972a3da3 | 572 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
573 | |
574 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 575 | |
7b24d861 | 576 | spin_unlock(&resv->lock); |
cf3ad20b | 577 | return add; |
96822904 AW |
578 | } |
579 | ||
1dd308a7 MK |
580 | /* |
581 | * Examine the existing reserve map and determine how many | |
582 | * huge pages in the specified range [f, t) are NOT currently | |
583 | * represented. This routine is called before a subsequent | |
584 | * call to region_add that will actually modify the reserve | |
585 | * map to add the specified range [f, t). region_chg does | |
586 | * not change the number of huge pages represented by the | |
0db9d74e MA |
587 | * map. A number of new file_region structures is added to the cache as a |
588 | * placeholder, for the subsequent region_add call to use. At least 1 | |
589 | * file_region structure is added. | |
590 | * | |
591 | * out_regions_needed is the number of regions added to the | |
592 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
593 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
594 | * |
595 | * Returns the number of huge pages that need to be added to the existing | |
596 | * reservation map for the range [f, t). This number is greater or equal to | |
597 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
598 | * is needed and can not be allocated. | |
1dd308a7 | 599 | */ |
0db9d74e MA |
600 | static long region_chg(struct resv_map *resv, long f, long t, |
601 | long *out_regions_needed) | |
96822904 | 602 | { |
96822904 AW |
603 | long chg = 0; |
604 | ||
7b24d861 | 605 | spin_lock(&resv->lock); |
5e911373 | 606 | |
972a3da3 | 607 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 608 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 609 | out_regions_needed); |
5e911373 | 610 | |
0db9d74e MA |
611 | if (*out_regions_needed == 0) |
612 | *out_regions_needed = 1; | |
5e911373 | 613 | |
0db9d74e MA |
614 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
615 | return -ENOMEM; | |
5e911373 | 616 | |
0db9d74e | 617 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 618 | |
7b24d861 | 619 | spin_unlock(&resv->lock); |
96822904 AW |
620 | return chg; |
621 | } | |
622 | ||
5e911373 MK |
623 | /* |
624 | * Abort the in progress add operation. The adds_in_progress field | |
625 | * of the resv_map keeps track of the operations in progress between | |
626 | * calls to region_chg and region_add. Operations are sometimes | |
627 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
628 | * is called to decrement the adds_in_progress counter. regions_needed |
629 | * is the value returned by the region_chg call, it is used to decrement | |
630 | * the adds_in_progress counter. | |
5e911373 MK |
631 | * |
632 | * NOTE: The range arguments [f, t) are not needed or used in this | |
633 | * routine. They are kept to make reading the calling code easier as | |
634 | * arguments will match the associated region_chg call. | |
635 | */ | |
0db9d74e MA |
636 | static void region_abort(struct resv_map *resv, long f, long t, |
637 | long regions_needed) | |
5e911373 MK |
638 | { |
639 | spin_lock(&resv->lock); | |
640 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 641 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
642 | spin_unlock(&resv->lock); |
643 | } | |
644 | ||
1dd308a7 | 645 | /* |
feba16e2 MK |
646 | * Delete the specified range [f, t) from the reserve map. If the |
647 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
648 | * should be deleted. Locate the regions which intersect [f, t) | |
649 | * and either trim, delete or split the existing regions. | |
650 | * | |
651 | * Returns the number of huge pages deleted from the reserve map. | |
652 | * In the normal case, the return value is zero or more. In the | |
653 | * case where a region must be split, a new region descriptor must | |
654 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
655 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
656 | * a region and possibly return -ENOMEM. Callers specifying | |
657 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 658 | */ |
feba16e2 | 659 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 660 | { |
1406ec9b | 661 | struct list_head *head = &resv->regions; |
96822904 | 662 | struct file_region *rg, *trg; |
feba16e2 MK |
663 | struct file_region *nrg = NULL; |
664 | long del = 0; | |
96822904 | 665 | |
feba16e2 | 666 | retry: |
7b24d861 | 667 | spin_lock(&resv->lock); |
feba16e2 | 668 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
669 | /* |
670 | * Skip regions before the range to be deleted. file_region | |
671 | * ranges are normally of the form [from, to). However, there | |
672 | * may be a "placeholder" entry in the map which is of the form | |
673 | * (from, to) with from == to. Check for placeholder entries | |
674 | * at the beginning of the range to be deleted. | |
675 | */ | |
676 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 677 | continue; |
dbe409e4 | 678 | |
feba16e2 | 679 | if (rg->from >= t) |
96822904 | 680 | break; |
96822904 | 681 | |
feba16e2 MK |
682 | if (f > rg->from && t < rg->to) { /* Must split region */ |
683 | /* | |
684 | * Check for an entry in the cache before dropping | |
685 | * lock and attempting allocation. | |
686 | */ | |
687 | if (!nrg && | |
688 | resv->region_cache_count > resv->adds_in_progress) { | |
689 | nrg = list_first_entry(&resv->region_cache, | |
690 | struct file_region, | |
691 | link); | |
692 | list_del(&nrg->link); | |
693 | resv->region_cache_count--; | |
694 | } | |
96822904 | 695 | |
feba16e2 MK |
696 | if (!nrg) { |
697 | spin_unlock(&resv->lock); | |
698 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
699 | if (!nrg) | |
700 | return -ENOMEM; | |
701 | goto retry; | |
702 | } | |
703 | ||
704 | del += t - f; | |
79aa925b | 705 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 706 | resv, rg, t - f, false); |
feba16e2 MK |
707 | |
708 | /* New entry for end of split region */ | |
709 | nrg->from = t; | |
710 | nrg->to = rg->to; | |
075a61d0 MA |
711 | |
712 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
713 | ||
feba16e2 MK |
714 | INIT_LIST_HEAD(&nrg->link); |
715 | ||
716 | /* Original entry is trimmed */ | |
717 | rg->to = f; | |
718 | ||
719 | list_add(&nrg->link, &rg->link); | |
720 | nrg = NULL; | |
96822904 | 721 | break; |
feba16e2 MK |
722 | } |
723 | ||
724 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
725 | del += rg->to - rg->from; | |
075a61d0 | 726 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 727 | rg->to - rg->from, true); |
feba16e2 MK |
728 | list_del(&rg->link); |
729 | kfree(rg); | |
730 | continue; | |
731 | } | |
732 | ||
733 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 734 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 735 | t - rg->from, false); |
075a61d0 | 736 | |
79aa925b MK |
737 | del += t - rg->from; |
738 | rg->from = t; | |
739 | } else { /* Trim end of region */ | |
075a61d0 | 740 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 741 | rg->to - f, false); |
79aa925b MK |
742 | |
743 | del += rg->to - f; | |
744 | rg->to = f; | |
feba16e2 | 745 | } |
96822904 | 746 | } |
7b24d861 | 747 | |
7b24d861 | 748 | spin_unlock(&resv->lock); |
feba16e2 MK |
749 | kfree(nrg); |
750 | return del; | |
96822904 AW |
751 | } |
752 | ||
b5cec28d MK |
753 | /* |
754 | * A rare out of memory error was encountered which prevented removal of | |
755 | * the reserve map region for a page. The huge page itself was free'ed | |
756 | * and removed from the page cache. This routine will adjust the subpool | |
757 | * usage count, and the global reserve count if needed. By incrementing | |
758 | * these counts, the reserve map entry which could not be deleted will | |
759 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
760 | * counts. | |
761 | */ | |
72e2936c | 762 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
763 | { |
764 | struct hugepage_subpool *spool = subpool_inode(inode); | |
765 | long rsv_adjust; | |
da56388c | 766 | bool reserved = false; |
b5cec28d MK |
767 | |
768 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 769 | if (rsv_adjust > 0) { |
b5cec28d MK |
770 | struct hstate *h = hstate_inode(inode); |
771 | ||
da56388c ML |
772 | if (!hugetlb_acct_memory(h, 1)) |
773 | reserved = true; | |
774 | } else if (!rsv_adjust) { | |
775 | reserved = true; | |
b5cec28d | 776 | } |
da56388c ML |
777 | |
778 | if (!reserved) | |
779 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
780 | } |
781 | ||
1dd308a7 MK |
782 | /* |
783 | * Count and return the number of huge pages in the reserve map | |
784 | * that intersect with the range [f, t). | |
785 | */ | |
1406ec9b | 786 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 787 | { |
1406ec9b | 788 | struct list_head *head = &resv->regions; |
84afd99b AW |
789 | struct file_region *rg; |
790 | long chg = 0; | |
791 | ||
7b24d861 | 792 | spin_lock(&resv->lock); |
84afd99b AW |
793 | /* Locate each segment we overlap with, and count that overlap. */ |
794 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
795 | long seg_from; |
796 | long seg_to; | |
84afd99b AW |
797 | |
798 | if (rg->to <= f) | |
799 | continue; | |
800 | if (rg->from >= t) | |
801 | break; | |
802 | ||
803 | seg_from = max(rg->from, f); | |
804 | seg_to = min(rg->to, t); | |
805 | ||
806 | chg += seg_to - seg_from; | |
807 | } | |
7b24d861 | 808 | spin_unlock(&resv->lock); |
84afd99b AW |
809 | |
810 | return chg; | |
811 | } | |
812 | ||
e7c4b0bf AW |
813 | /* |
814 | * Convert the address within this vma to the page offset within | |
815 | * the mapping, in pagecache page units; huge pages here. | |
816 | */ | |
a5516438 AK |
817 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
818 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 819 | { |
a5516438 AK |
820 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
821 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
822 | } |
823 | ||
0fe6e20b NH |
824 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
825 | unsigned long address) | |
826 | { | |
827 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
828 | } | |
dee41079 | 829 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 830 | |
08fba699 MG |
831 | /* |
832 | * Return the size of the pages allocated when backing a VMA. In the majority | |
833 | * cases this will be same size as used by the page table entries. | |
834 | */ | |
835 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
836 | { | |
05ea8860 DW |
837 | if (vma->vm_ops && vma->vm_ops->pagesize) |
838 | return vma->vm_ops->pagesize(vma); | |
839 | return PAGE_SIZE; | |
08fba699 | 840 | } |
f340ca0f | 841 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 842 | |
3340289d MG |
843 | /* |
844 | * Return the page size being used by the MMU to back a VMA. In the majority | |
845 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
846 | * architectures where it differs, an architecture-specific 'strong' |
847 | * version of this symbol is required. | |
3340289d | 848 | */ |
09135cc5 | 849 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
850 | { |
851 | return vma_kernel_pagesize(vma); | |
852 | } | |
3340289d | 853 | |
84afd99b AW |
854 | /* |
855 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
856 | * bits of the reservation map pointer, which are always clear due to | |
857 | * alignment. | |
858 | */ | |
859 | #define HPAGE_RESV_OWNER (1UL << 0) | |
860 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 861 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 862 | |
a1e78772 MG |
863 | /* |
864 | * These helpers are used to track how many pages are reserved for | |
865 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
866 | * is guaranteed to have their future faults succeed. | |
867 | * | |
868 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
869 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
870 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
871 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
872 | * |
873 | * The private mapping reservation is represented in a subtly different | |
874 | * manner to a shared mapping. A shared mapping has a region map associated | |
875 | * with the underlying file, this region map represents the backing file | |
876 | * pages which have ever had a reservation assigned which this persists even | |
877 | * after the page is instantiated. A private mapping has a region map | |
878 | * associated with the original mmap which is attached to all VMAs which | |
879 | * reference it, this region map represents those offsets which have consumed | |
880 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 881 | */ |
e7c4b0bf AW |
882 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
883 | { | |
884 | return (unsigned long)vma->vm_private_data; | |
885 | } | |
886 | ||
887 | static void set_vma_private_data(struct vm_area_struct *vma, | |
888 | unsigned long value) | |
889 | { | |
890 | vma->vm_private_data = (void *)value; | |
891 | } | |
892 | ||
e9fe92ae MA |
893 | static void |
894 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
895 | struct hugetlb_cgroup *h_cg, | |
896 | struct hstate *h) | |
897 | { | |
898 | #ifdef CONFIG_CGROUP_HUGETLB | |
899 | if (!h_cg || !h) { | |
900 | resv_map->reservation_counter = NULL; | |
901 | resv_map->pages_per_hpage = 0; | |
902 | resv_map->css = NULL; | |
903 | } else { | |
904 | resv_map->reservation_counter = | |
905 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
906 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
907 | resv_map->css = &h_cg->css; | |
908 | } | |
909 | #endif | |
910 | } | |
911 | ||
9119a41e | 912 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
913 | { |
914 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
915 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
916 | ||
917 | if (!resv_map || !rg) { | |
918 | kfree(resv_map); | |
919 | kfree(rg); | |
84afd99b | 920 | return NULL; |
5e911373 | 921 | } |
84afd99b AW |
922 | |
923 | kref_init(&resv_map->refs); | |
7b24d861 | 924 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
925 | INIT_LIST_HEAD(&resv_map->regions); |
926 | ||
5e911373 | 927 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
928 | /* |
929 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
930 | * fields don't do cgroup accounting. On private mappings, these will be | |
931 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
932 | * reservations are to be un-charged from here. | |
933 | */ | |
934 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
935 | |
936 | INIT_LIST_HEAD(&resv_map->region_cache); | |
937 | list_add(&rg->link, &resv_map->region_cache); | |
938 | resv_map->region_cache_count = 1; | |
939 | ||
84afd99b AW |
940 | return resv_map; |
941 | } | |
942 | ||
9119a41e | 943 | void resv_map_release(struct kref *ref) |
84afd99b AW |
944 | { |
945 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
946 | struct list_head *head = &resv_map->region_cache; |
947 | struct file_region *rg, *trg; | |
84afd99b AW |
948 | |
949 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 950 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
951 | |
952 | /* ... and any entries left in the cache */ | |
953 | list_for_each_entry_safe(rg, trg, head, link) { | |
954 | list_del(&rg->link); | |
955 | kfree(rg); | |
956 | } | |
957 | ||
958 | VM_BUG_ON(resv_map->adds_in_progress); | |
959 | ||
84afd99b AW |
960 | kfree(resv_map); |
961 | } | |
962 | ||
4e35f483 JK |
963 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
964 | { | |
f27a5136 MK |
965 | /* |
966 | * At inode evict time, i_mapping may not point to the original | |
967 | * address space within the inode. This original address space | |
968 | * contains the pointer to the resv_map. So, always use the | |
969 | * address space embedded within the inode. | |
970 | * The VERY common case is inode->mapping == &inode->i_data but, | |
971 | * this may not be true for device special inodes. | |
972 | */ | |
973 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
974 | } |
975 | ||
84afd99b | 976 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 977 | { |
81d1b09c | 978 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
979 | if (vma->vm_flags & VM_MAYSHARE) { |
980 | struct address_space *mapping = vma->vm_file->f_mapping; | |
981 | struct inode *inode = mapping->host; | |
982 | ||
983 | return inode_resv_map(inode); | |
984 | ||
985 | } else { | |
84afd99b AW |
986 | return (struct resv_map *)(get_vma_private_data(vma) & |
987 | ~HPAGE_RESV_MASK); | |
4e35f483 | 988 | } |
a1e78772 MG |
989 | } |
990 | ||
84afd99b | 991 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 992 | { |
81d1b09c SL |
993 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
994 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 995 | |
84afd99b AW |
996 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
997 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
998 | } |
999 | ||
1000 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
1001 | { | |
81d1b09c SL |
1002 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
1003 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
1004 | |
1005 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
1006 | } |
1007 | ||
1008 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
1009 | { | |
81d1b09c | 1010 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
1011 | |
1012 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
1013 | } |
1014 | ||
04f2cbe3 | 1015 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
1016 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
1017 | { | |
81d1b09c | 1018 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 1019 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
1020 | vma->vm_private_data = (void *)0; |
1021 | } | |
1022 | ||
550a7d60 MA |
1023 | /* |
1024 | * Reset and decrement one ref on hugepage private reservation. | |
1025 | * Called with mm->mmap_sem writer semaphore held. | |
1026 | * This function should be only used by move_vma() and operate on | |
1027 | * same sized vma. It should never come here with last ref on the | |
1028 | * reservation. | |
1029 | */ | |
1030 | void clear_vma_resv_huge_pages(struct vm_area_struct *vma) | |
1031 | { | |
1032 | /* | |
1033 | * Clear the old hugetlb private page reservation. | |
1034 | * It has already been transferred to new_vma. | |
1035 | * | |
1036 | * During a mremap() operation of a hugetlb vma we call move_vma() | |
1037 | * which copies vma into new_vma and unmaps vma. After the copy | |
1038 | * operation both new_vma and vma share a reference to the resv_map | |
1039 | * struct, and at that point vma is about to be unmapped. We don't | |
1040 | * want to return the reservation to the pool at unmap of vma because | |
1041 | * the reservation still lives on in new_vma, so simply decrement the | |
1042 | * ref here and remove the resv_map reference from this vma. | |
1043 | */ | |
1044 | struct resv_map *reservations = vma_resv_map(vma); | |
1045 | ||
afe041c2 BQM |
1046 | if (reservations && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1047 | resv_map_put_hugetlb_cgroup_uncharge_info(reservations); | |
550a7d60 | 1048 | kref_put(&reservations->refs, resv_map_release); |
afe041c2 | 1049 | } |
550a7d60 MA |
1050 | |
1051 | reset_vma_resv_huge_pages(vma); | |
1052 | } | |
1053 | ||
a1e78772 | 1054 | /* Returns true if the VMA has associated reserve pages */ |
559ec2f8 | 1055 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1056 | { |
af0ed73e JK |
1057 | if (vma->vm_flags & VM_NORESERVE) { |
1058 | /* | |
1059 | * This address is already reserved by other process(chg == 0), | |
1060 | * so, we should decrement reserved count. Without decrementing, | |
1061 | * reserve count remains after releasing inode, because this | |
1062 | * allocated page will go into page cache and is regarded as | |
1063 | * coming from reserved pool in releasing step. Currently, we | |
1064 | * don't have any other solution to deal with this situation | |
1065 | * properly, so add work-around here. | |
1066 | */ | |
1067 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1068 | return true; |
af0ed73e | 1069 | else |
559ec2f8 | 1070 | return false; |
af0ed73e | 1071 | } |
a63884e9 JK |
1072 | |
1073 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1074 | if (vma->vm_flags & VM_MAYSHARE) { |
1075 | /* | |
1076 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1077 | * be a region map for all pages. The only situation where | |
1078 | * there is no region map is if a hole was punched via | |
7c8de358 | 1079 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1080 | * use. This situation is indicated if chg != 0. |
1081 | */ | |
1082 | if (chg) | |
1083 | return false; | |
1084 | else | |
1085 | return true; | |
1086 | } | |
a63884e9 JK |
1087 | |
1088 | /* | |
1089 | * Only the process that called mmap() has reserves for | |
1090 | * private mappings. | |
1091 | */ | |
67961f9d MK |
1092 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1093 | /* | |
1094 | * Like the shared case above, a hole punch or truncate | |
1095 | * could have been performed on the private mapping. | |
1096 | * Examine the value of chg to determine if reserves | |
1097 | * actually exist or were previously consumed. | |
1098 | * Very Subtle - The value of chg comes from a previous | |
1099 | * call to vma_needs_reserves(). The reserve map for | |
1100 | * private mappings has different (opposite) semantics | |
1101 | * than that of shared mappings. vma_needs_reserves() | |
1102 | * has already taken this difference in semantics into | |
1103 | * account. Therefore, the meaning of chg is the same | |
1104 | * as in the shared case above. Code could easily be | |
1105 | * combined, but keeping it separate draws attention to | |
1106 | * subtle differences. | |
1107 | */ | |
1108 | if (chg) | |
1109 | return false; | |
1110 | else | |
1111 | return true; | |
1112 | } | |
a63884e9 | 1113 | |
559ec2f8 | 1114 | return false; |
a1e78772 MG |
1115 | } |
1116 | ||
a5516438 | 1117 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1118 | { |
1119 | int nid = page_to_nid(page); | |
9487ca60 MK |
1120 | |
1121 | lockdep_assert_held(&hugetlb_lock); | |
b65a4eda MK |
1122 | VM_BUG_ON_PAGE(page_count(page), page); |
1123 | ||
0edaecfa | 1124 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1125 | h->free_huge_pages++; |
1126 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1127 | SetHPageFreed(page); |
1da177e4 LT |
1128 | } |
1129 | ||
94310cbc | 1130 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1131 | { |
1132 | struct page *page; | |
1a08ae36 | 1133 | bool pin = !!(current->flags & PF_MEMALLOC_PIN); |
bbe88753 | 1134 | |
9487ca60 | 1135 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 | 1136 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
8e3560d9 | 1137 | if (pin && !is_pinnable_page(page)) |
bbe88753 | 1138 | continue; |
bf50bab2 | 1139 | |
6664bfc8 WY |
1140 | if (PageHWPoison(page)) |
1141 | continue; | |
1142 | ||
1143 | list_move(&page->lru, &h->hugepage_activelist); | |
1144 | set_page_refcounted(page); | |
6c037149 | 1145 | ClearHPageFreed(page); |
6664bfc8 WY |
1146 | h->free_huge_pages--; |
1147 | h->free_huge_pages_node[nid]--; | |
1148 | return page; | |
bbe88753 JK |
1149 | } |
1150 | ||
6664bfc8 | 1151 | return NULL; |
bf50bab2 NH |
1152 | } |
1153 | ||
3e59fcb0 MH |
1154 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1155 | nodemask_t *nmask) | |
94310cbc | 1156 | { |
3e59fcb0 MH |
1157 | unsigned int cpuset_mems_cookie; |
1158 | struct zonelist *zonelist; | |
1159 | struct zone *zone; | |
1160 | struct zoneref *z; | |
98fa15f3 | 1161 | int node = NUMA_NO_NODE; |
94310cbc | 1162 | |
3e59fcb0 MH |
1163 | zonelist = node_zonelist(nid, gfp_mask); |
1164 | ||
1165 | retry_cpuset: | |
1166 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1167 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1168 | struct page *page; | |
1169 | ||
1170 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1171 | continue; | |
1172 | /* | |
1173 | * no need to ask again on the same node. Pool is node rather than | |
1174 | * zone aware | |
1175 | */ | |
1176 | if (zone_to_nid(zone) == node) | |
1177 | continue; | |
1178 | node = zone_to_nid(zone); | |
94310cbc | 1179 | |
94310cbc AK |
1180 | page = dequeue_huge_page_node_exact(h, node); |
1181 | if (page) | |
1182 | return page; | |
1183 | } | |
3e59fcb0 MH |
1184 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1185 | goto retry_cpuset; | |
1186 | ||
94310cbc AK |
1187 | return NULL; |
1188 | } | |
1189 | ||
a5516438 AK |
1190 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1191 | struct vm_area_struct *vma, | |
af0ed73e JK |
1192 | unsigned long address, int avoid_reserve, |
1193 | long chg) | |
1da177e4 | 1194 | { |
cfcaa66f | 1195 | struct page *page = NULL; |
480eccf9 | 1196 | struct mempolicy *mpol; |
04ec6264 | 1197 | gfp_t gfp_mask; |
3e59fcb0 | 1198 | nodemask_t *nodemask; |
04ec6264 | 1199 | int nid; |
1da177e4 | 1200 | |
a1e78772 MG |
1201 | /* |
1202 | * A child process with MAP_PRIVATE mappings created by their parent | |
1203 | * have no page reserves. This check ensures that reservations are | |
1204 | * not "stolen". The child may still get SIGKILLed | |
1205 | */ | |
af0ed73e | 1206 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1207 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1208 | goto err; |
a1e78772 | 1209 | |
04f2cbe3 | 1210 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1211 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1212 | goto err; |
04f2cbe3 | 1213 | |
04ec6264 VB |
1214 | gfp_mask = htlb_alloc_mask(h); |
1215 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
1216 | |
1217 | if (mpol_is_preferred_many(mpol)) { | |
1218 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1219 | ||
1220 | /* Fallback to all nodes if page==NULL */ | |
1221 | nodemask = NULL; | |
1222 | } | |
1223 | ||
1224 | if (!page) | |
1225 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); | |
1226 | ||
3e59fcb0 | 1227 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { |
d6995da3 | 1228 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1229 | h->resv_huge_pages--; |
1da177e4 | 1230 | } |
cc9a6c87 | 1231 | |
52cd3b07 | 1232 | mpol_cond_put(mpol); |
1da177e4 | 1233 | return page; |
cc9a6c87 MG |
1234 | |
1235 | err: | |
cc9a6c87 | 1236 | return NULL; |
1da177e4 LT |
1237 | } |
1238 | ||
1cac6f2c LC |
1239 | /* |
1240 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1241 | * We may have allocated or freed a huge page based on a different | |
1242 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1243 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1244 | * node for alloc or free. | |
1245 | */ | |
1246 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1247 | { | |
0edaf86c | 1248 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1249 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1250 | ||
1251 | return nid; | |
1252 | } | |
1253 | ||
1254 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1255 | { | |
1256 | if (!node_isset(nid, *nodes_allowed)) | |
1257 | nid = next_node_allowed(nid, nodes_allowed); | |
1258 | return nid; | |
1259 | } | |
1260 | ||
1261 | /* | |
1262 | * returns the previously saved node ["this node"] from which to | |
1263 | * allocate a persistent huge page for the pool and advance the | |
1264 | * next node from which to allocate, handling wrap at end of node | |
1265 | * mask. | |
1266 | */ | |
1267 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1268 | nodemask_t *nodes_allowed) | |
1269 | { | |
1270 | int nid; | |
1271 | ||
1272 | VM_BUG_ON(!nodes_allowed); | |
1273 | ||
1274 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1275 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1276 | ||
1277 | return nid; | |
1278 | } | |
1279 | ||
1280 | /* | |
10c6ec49 | 1281 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1282 | * node ["this node"] from which to free a huge page. Advance the |
1283 | * next node id whether or not we find a free huge page to free so | |
1284 | * that the next attempt to free addresses the next node. | |
1285 | */ | |
1286 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1287 | { | |
1288 | int nid; | |
1289 | ||
1290 | VM_BUG_ON(!nodes_allowed); | |
1291 | ||
1292 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1293 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1294 | ||
1295 | return nid; | |
1296 | } | |
1297 | ||
1298 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1299 | for (nr_nodes = nodes_weight(*mask); \ | |
1300 | nr_nodes > 0 && \ | |
1301 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1302 | nr_nodes--) | |
1303 | ||
1304 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1305 | for (nr_nodes = nodes_weight(*mask); \ | |
1306 | nr_nodes > 0 && \ | |
1307 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1308 | nr_nodes--) | |
1309 | ||
8531fc6f | 1310 | /* used to demote non-gigantic_huge pages as well */ |
34d9e35b MK |
1311 | static void __destroy_compound_gigantic_page(struct page *page, |
1312 | unsigned int order, bool demote) | |
944d9fec LC |
1313 | { |
1314 | int i; | |
1315 | int nr_pages = 1 << order; | |
1316 | struct page *p = page + 1; | |
1317 | ||
c8cc708a | 1318 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1319 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1320 | |
944d9fec | 1321 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
a01f4390 | 1322 | p->mapping = NULL; |
1d798ca3 | 1323 | clear_compound_head(p); |
34d9e35b MK |
1324 | if (!demote) |
1325 | set_page_refcounted(p); | |
944d9fec LC |
1326 | } |
1327 | ||
1328 | set_compound_order(page, 0); | |
5232c63f | 1329 | #ifdef CONFIG_64BIT |
ba9c1201 | 1330 | page[1].compound_nr = 0; |
5232c63f | 1331 | #endif |
944d9fec LC |
1332 | __ClearPageHead(page); |
1333 | } | |
1334 | ||
8531fc6f MK |
1335 | static void destroy_compound_hugetlb_page_for_demote(struct page *page, |
1336 | unsigned int order) | |
1337 | { | |
1338 | __destroy_compound_gigantic_page(page, order, true); | |
1339 | } | |
1340 | ||
1341 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE | |
34d9e35b MK |
1342 | static void destroy_compound_gigantic_page(struct page *page, |
1343 | unsigned int order) | |
1344 | { | |
1345 | __destroy_compound_gigantic_page(page, order, false); | |
1346 | } | |
1347 | ||
d00181b9 | 1348 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1349 | { |
cf11e85f RG |
1350 | /* |
1351 | * If the page isn't allocated using the cma allocator, | |
1352 | * cma_release() returns false. | |
1353 | */ | |
dbda8fea BS |
1354 | #ifdef CONFIG_CMA |
1355 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1356 | return; |
dbda8fea | 1357 | #endif |
cf11e85f | 1358 | |
944d9fec LC |
1359 | free_contig_range(page_to_pfn(page), 1 << order); |
1360 | } | |
1361 | ||
4eb0716e | 1362 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1363 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1364 | int nid, nodemask_t *nodemask) | |
944d9fec | 1365 | { |
04adbc3f | 1366 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1367 | if (nid == NUMA_NO_NODE) |
1368 | nid = numa_mem_id(); | |
944d9fec | 1369 | |
dbda8fea BS |
1370 | #ifdef CONFIG_CMA |
1371 | { | |
cf11e85f RG |
1372 | struct page *page; |
1373 | int node; | |
1374 | ||
953f064a LX |
1375 | if (hugetlb_cma[nid]) { |
1376 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1377 | huge_page_order(h), true); | |
cf11e85f RG |
1378 | if (page) |
1379 | return page; | |
1380 | } | |
953f064a LX |
1381 | |
1382 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1383 | for_each_node_mask(node, *nodemask) { | |
1384 | if (node == nid || !hugetlb_cma[node]) | |
1385 | continue; | |
1386 | ||
1387 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1388 | huge_page_order(h), true); | |
1389 | if (page) | |
1390 | return page; | |
1391 | } | |
1392 | } | |
cf11e85f | 1393 | } |
dbda8fea | 1394 | #endif |
cf11e85f | 1395 | |
5e27a2df | 1396 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1397 | } |
1398 | ||
4eb0716e AG |
1399 | #else /* !CONFIG_CONTIG_ALLOC */ |
1400 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1401 | int nid, nodemask_t *nodemask) | |
1402 | { | |
1403 | return NULL; | |
1404 | } | |
1405 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1406 | |
e1073d1e | 1407 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1408 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1409 | int nid, nodemask_t *nodemask) |
1410 | { | |
1411 | return NULL; | |
1412 | } | |
d00181b9 | 1413 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1414 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1415 | unsigned int order) { } |
944d9fec LC |
1416 | #endif |
1417 | ||
6eb4e88a MK |
1418 | /* |
1419 | * Remove hugetlb page from lists, and update dtor so that page appears | |
34d9e35b MK |
1420 | * as just a compound page. |
1421 | * | |
1422 | * A reference is held on the page, except in the case of demote. | |
6eb4e88a MK |
1423 | * |
1424 | * Must be called with hugetlb lock held. | |
1425 | */ | |
34d9e35b MK |
1426 | static void __remove_hugetlb_page(struct hstate *h, struct page *page, |
1427 | bool adjust_surplus, | |
1428 | bool demote) | |
6eb4e88a MK |
1429 | { |
1430 | int nid = page_to_nid(page); | |
1431 | ||
1432 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1433 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1434 | ||
9487ca60 | 1435 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1436 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1437 | return; | |
1438 | ||
1439 | list_del(&page->lru); | |
1440 | ||
1441 | if (HPageFreed(page)) { | |
1442 | h->free_huge_pages--; | |
1443 | h->free_huge_pages_node[nid]--; | |
1444 | } | |
1445 | if (adjust_surplus) { | |
1446 | h->surplus_huge_pages--; | |
1447 | h->surplus_huge_pages_node[nid]--; | |
1448 | } | |
1449 | ||
e32d20c0 MK |
1450 | /* |
1451 | * Very subtle | |
1452 | * | |
1453 | * For non-gigantic pages set the destructor to the normal compound | |
1454 | * page dtor. This is needed in case someone takes an additional | |
1455 | * temporary ref to the page, and freeing is delayed until they drop | |
1456 | * their reference. | |
1457 | * | |
1458 | * For gigantic pages set the destructor to the null dtor. This | |
1459 | * destructor will never be called. Before freeing the gigantic | |
1460 | * page destroy_compound_gigantic_page will turn the compound page | |
1461 | * into a simple group of pages. After this the destructor does not | |
1462 | * apply. | |
1463 | * | |
1464 | * This handles the case where more than one ref is held when and | |
1465 | * after update_and_free_page is called. | |
34d9e35b MK |
1466 | * |
1467 | * In the case of demote we do not ref count the page as it will soon | |
1468 | * be turned into a page of smaller size. | |
e32d20c0 | 1469 | */ |
34d9e35b MK |
1470 | if (!demote) |
1471 | set_page_refcounted(page); | |
e32d20c0 MK |
1472 | if (hstate_is_gigantic(h)) |
1473 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1474 | else | |
1475 | set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); | |
6eb4e88a MK |
1476 | |
1477 | h->nr_huge_pages--; | |
1478 | h->nr_huge_pages_node[nid]--; | |
1479 | } | |
1480 | ||
34d9e35b MK |
1481 | static void remove_hugetlb_page(struct hstate *h, struct page *page, |
1482 | bool adjust_surplus) | |
1483 | { | |
1484 | __remove_hugetlb_page(h, page, adjust_surplus, false); | |
1485 | } | |
1486 | ||
8531fc6f MK |
1487 | static void remove_hugetlb_page_for_demote(struct hstate *h, struct page *page, |
1488 | bool adjust_surplus) | |
1489 | { | |
1490 | __remove_hugetlb_page(h, page, adjust_surplus, true); | |
1491 | } | |
1492 | ||
ad2fa371 MS |
1493 | static void add_hugetlb_page(struct hstate *h, struct page *page, |
1494 | bool adjust_surplus) | |
1495 | { | |
1496 | int zeroed; | |
1497 | int nid = page_to_nid(page); | |
1498 | ||
1499 | VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); | |
1500 | ||
1501 | lockdep_assert_held(&hugetlb_lock); | |
1502 | ||
1503 | INIT_LIST_HEAD(&page->lru); | |
1504 | h->nr_huge_pages++; | |
1505 | h->nr_huge_pages_node[nid]++; | |
1506 | ||
1507 | if (adjust_surplus) { | |
1508 | h->surplus_huge_pages++; | |
1509 | h->surplus_huge_pages_node[nid]++; | |
1510 | } | |
1511 | ||
1512 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); | |
1513 | set_page_private(page, 0); | |
1514 | SetHPageVmemmapOptimized(page); | |
1515 | ||
1516 | /* | |
b65a4eda MK |
1517 | * This page is about to be managed by the hugetlb allocator and |
1518 | * should have no users. Drop our reference, and check for others | |
1519 | * just in case. | |
ad2fa371 MS |
1520 | */ |
1521 | zeroed = put_page_testzero(page); | |
b65a4eda MK |
1522 | if (!zeroed) |
1523 | /* | |
1524 | * It is VERY unlikely soneone else has taken a ref on | |
1525 | * the page. In this case, we simply return as the | |
1526 | * hugetlb destructor (free_huge_page) will be called | |
1527 | * when this other ref is dropped. | |
1528 | */ | |
1529 | return; | |
1530 | ||
ad2fa371 MS |
1531 | arch_clear_hugepage_flags(page); |
1532 | enqueue_huge_page(h, page); | |
1533 | } | |
1534 | ||
b65d4adb | 1535 | static void __update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1536 | { |
1537 | int i; | |
dbfee5ae | 1538 | struct page *subpage = page; |
a5516438 | 1539 | |
4eb0716e | 1540 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1541 | return; |
18229df5 | 1542 | |
5981611d | 1543 | if (hugetlb_vmemmap_alloc(h, page)) { |
ad2fa371 MS |
1544 | spin_lock_irq(&hugetlb_lock); |
1545 | /* | |
1546 | * If we cannot allocate vmemmap pages, just refuse to free the | |
1547 | * page and put the page back on the hugetlb free list and treat | |
1548 | * as a surplus page. | |
1549 | */ | |
1550 | add_hugetlb_page(h, page, true); | |
1551 | spin_unlock_irq(&hugetlb_lock); | |
1552 | return; | |
1553 | } | |
1554 | ||
dbfee5ae MK |
1555 | for (i = 0; i < pages_per_huge_page(h); |
1556 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1557 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1558 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1559 | 1 << PG_active | 1 << PG_private | |
1560 | 1 << PG_writeback); | |
6af2acb6 | 1561 | } |
a01f4390 MK |
1562 | |
1563 | /* | |
1564 | * Non-gigantic pages demoted from CMA allocated gigantic pages | |
1565 | * need to be given back to CMA in free_gigantic_page. | |
1566 | */ | |
1567 | if (hstate_is_gigantic(h) || | |
1568 | hugetlb_cma_page(page, huge_page_order(h))) { | |
944d9fec LC |
1569 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1570 | free_gigantic_page(page, huge_page_order(h)); | |
1571 | } else { | |
944d9fec LC |
1572 | __free_pages(page, huge_page_order(h)); |
1573 | } | |
6af2acb6 AL |
1574 | } |
1575 | ||
b65d4adb MS |
1576 | /* |
1577 | * As update_and_free_page() can be called under any context, so we cannot | |
1578 | * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the | |
1579 | * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate | |
1580 | * the vmemmap pages. | |
1581 | * | |
1582 | * free_hpage_workfn() locklessly retrieves the linked list of pages to be | |
1583 | * freed and frees them one-by-one. As the page->mapping pointer is going | |
1584 | * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node | |
1585 | * structure of a lockless linked list of huge pages to be freed. | |
1586 | */ | |
1587 | static LLIST_HEAD(hpage_freelist); | |
1588 | ||
1589 | static void free_hpage_workfn(struct work_struct *work) | |
1590 | { | |
1591 | struct llist_node *node; | |
1592 | ||
1593 | node = llist_del_all(&hpage_freelist); | |
1594 | ||
1595 | while (node) { | |
1596 | struct page *page; | |
1597 | struct hstate *h; | |
1598 | ||
1599 | page = container_of((struct address_space **)node, | |
1600 | struct page, mapping); | |
1601 | node = node->next; | |
1602 | page->mapping = NULL; | |
1603 | /* | |
1604 | * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() | |
1605 | * is going to trigger because a previous call to | |
1606 | * remove_hugetlb_page() will set_compound_page_dtor(page, | |
1607 | * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. | |
1608 | */ | |
1609 | h = size_to_hstate(page_size(page)); | |
1610 | ||
1611 | __update_and_free_page(h, page); | |
1612 | ||
1613 | cond_resched(); | |
1614 | } | |
1615 | } | |
1616 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1617 | ||
1618 | static inline void flush_free_hpage_work(struct hstate *h) | |
1619 | { | |
5981611d | 1620 | if (hugetlb_optimize_vmemmap_pages(h)) |
b65d4adb MS |
1621 | flush_work(&free_hpage_work); |
1622 | } | |
1623 | ||
1624 | static void update_and_free_page(struct hstate *h, struct page *page, | |
1625 | bool atomic) | |
1626 | { | |
ad2fa371 | 1627 | if (!HPageVmemmapOptimized(page) || !atomic) { |
b65d4adb MS |
1628 | __update_and_free_page(h, page); |
1629 | return; | |
1630 | } | |
1631 | ||
1632 | /* | |
1633 | * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. | |
1634 | * | |
1635 | * Only call schedule_work() if hpage_freelist is previously | |
1636 | * empty. Otherwise, schedule_work() had been called but the workfn | |
1637 | * hasn't retrieved the list yet. | |
1638 | */ | |
1639 | if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) | |
1640 | schedule_work(&free_hpage_work); | |
1641 | } | |
1642 | ||
10c6ec49 MK |
1643 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1644 | { | |
1645 | struct page *page, *t_page; | |
1646 | ||
1647 | list_for_each_entry_safe(page, t_page, list, lru) { | |
b65d4adb | 1648 | update_and_free_page(h, page, false); |
10c6ec49 MK |
1649 | cond_resched(); |
1650 | } | |
1651 | } | |
1652 | ||
e5ff2159 AK |
1653 | struct hstate *size_to_hstate(unsigned long size) |
1654 | { | |
1655 | struct hstate *h; | |
1656 | ||
1657 | for_each_hstate(h) { | |
1658 | if (huge_page_size(h) == size) | |
1659 | return h; | |
1660 | } | |
1661 | return NULL; | |
1662 | } | |
1663 | ||
db71ef79 | 1664 | void free_huge_page(struct page *page) |
27a85ef1 | 1665 | { |
a5516438 AK |
1666 | /* |
1667 | * Can't pass hstate in here because it is called from the | |
1668 | * compound page destructor. | |
1669 | */ | |
e5ff2159 | 1670 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1671 | int nid = page_to_nid(page); |
d6995da3 | 1672 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1673 | bool restore_reserve; |
db71ef79 | 1674 | unsigned long flags; |
27a85ef1 | 1675 | |
b4330afb MK |
1676 | VM_BUG_ON_PAGE(page_count(page), page); |
1677 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1678 | |
d6995da3 | 1679 | hugetlb_set_page_subpool(page, NULL); |
78fbe906 DH |
1680 | if (PageAnon(page)) |
1681 | __ClearPageAnonExclusive(page); | |
8ace22bc | 1682 | page->mapping = NULL; |
d6995da3 MK |
1683 | restore_reserve = HPageRestoreReserve(page); |
1684 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1685 | |
1c5ecae3 | 1686 | /* |
d6995da3 | 1687 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1688 | * reservation. If the page was associated with a subpool, there |
1689 | * would have been a page reserved in the subpool before allocation | |
1690 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1691 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1692 | * remove the reserved page from the subpool. |
1c5ecae3 | 1693 | */ |
0919e1b6 MK |
1694 | if (!restore_reserve) { |
1695 | /* | |
1696 | * A return code of zero implies that the subpool will be | |
1697 | * under its minimum size if the reservation is not restored | |
1698 | * after page is free. Therefore, force restore_reserve | |
1699 | * operation. | |
1700 | */ | |
1701 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1702 | restore_reserve = true; | |
1703 | } | |
1c5ecae3 | 1704 | |
db71ef79 | 1705 | spin_lock_irqsave(&hugetlb_lock, flags); |
8f251a3d | 1706 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1707 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1708 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1709 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1710 | pages_per_huge_page(h), page); | |
07443a85 JK |
1711 | if (restore_reserve) |
1712 | h->resv_huge_pages++; | |
1713 | ||
9157c311 | 1714 | if (HPageTemporary(page)) { |
6eb4e88a | 1715 | remove_hugetlb_page(h, page, false); |
db71ef79 | 1716 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1717 | update_and_free_page(h, page, true); |
ab5ac90a | 1718 | } else if (h->surplus_huge_pages_node[nid]) { |
0edaecfa | 1719 | /* remove the page from active list */ |
6eb4e88a | 1720 | remove_hugetlb_page(h, page, true); |
db71ef79 | 1721 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1722 | update_and_free_page(h, page, true); |
7893d1d5 | 1723 | } else { |
5d3a551c | 1724 | arch_clear_hugepage_flags(page); |
a5516438 | 1725 | enqueue_huge_page(h, page); |
db71ef79 | 1726 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1727 | } |
c77c0a8a WL |
1728 | } |
1729 | ||
d3d99fcc OS |
1730 | /* |
1731 | * Must be called with the hugetlb lock held | |
1732 | */ | |
1733 | static void __prep_account_new_huge_page(struct hstate *h, int nid) | |
1734 | { | |
1735 | lockdep_assert_held(&hugetlb_lock); | |
1736 | h->nr_huge_pages++; | |
1737 | h->nr_huge_pages_node[nid]++; | |
1738 | } | |
1739 | ||
f41f2ed4 | 1740 | static void __prep_new_huge_page(struct hstate *h, struct page *page) |
b7ba30c6 | 1741 | { |
5981611d | 1742 | hugetlb_vmemmap_free(h, page); |
0edaecfa | 1743 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1744 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1745 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1746 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1747 | set_hugetlb_cgroup_rsvd(page, NULL); |
d3d99fcc OS |
1748 | } |
1749 | ||
1750 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1751 | { | |
f41f2ed4 | 1752 | __prep_new_huge_page(h, page); |
db71ef79 | 1753 | spin_lock_irq(&hugetlb_lock); |
d3d99fcc | 1754 | __prep_account_new_huge_page(h, nid); |
db71ef79 | 1755 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1756 | } |
1757 | ||
34d9e35b MK |
1758 | static bool __prep_compound_gigantic_page(struct page *page, unsigned int order, |
1759 | bool demote) | |
20a0307c | 1760 | { |
7118fc29 | 1761 | int i, j; |
20a0307c WF |
1762 | int nr_pages = 1 << order; |
1763 | struct page *p = page + 1; | |
1764 | ||
1765 | /* we rely on prep_new_huge_page to set the destructor */ | |
1766 | set_compound_order(page, order); | |
ef5a22be | 1767 | __ClearPageReserved(page); |
de09d31d | 1768 | __SetPageHead(page); |
20a0307c | 1769 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1770 | /* |
1771 | * For gigantic hugepages allocated through bootmem at | |
1772 | * boot, it's safer to be consistent with the not-gigantic | |
1773 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1774 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1775 | * pages may get the reference counting wrong if they see |
1776 | * PG_reserved set on a tail page (despite the head page not | |
1777 | * having PG_reserved set). Enforcing this consistency between | |
1778 | * head and tail pages allows drivers to optimize away a check | |
1779 | * on the head page when they need know if put_page() is needed | |
1780 | * after get_user_pages(). | |
1781 | */ | |
1782 | __ClearPageReserved(p); | |
7118fc29 MK |
1783 | /* |
1784 | * Subtle and very unlikely | |
1785 | * | |
1786 | * Gigantic 'page allocators' such as memblock or cma will | |
1787 | * return a set of pages with each page ref counted. We need | |
1788 | * to turn this set of pages into a compound page with tail | |
1789 | * page ref counts set to zero. Code such as speculative page | |
1790 | * cache adding could take a ref on a 'to be' tail page. | |
1791 | * We need to respect any increased ref count, and only set | |
1792 | * the ref count to zero if count is currently 1. If count | |
416d85ed MK |
1793 | * is not 1, we return an error. An error return indicates |
1794 | * the set of pages can not be converted to a gigantic page. | |
1795 | * The caller who allocated the pages should then discard the | |
1796 | * pages using the appropriate free interface. | |
34d9e35b MK |
1797 | * |
1798 | * In the case of demote, the ref count will be zero. | |
7118fc29 | 1799 | */ |
34d9e35b MK |
1800 | if (!demote) { |
1801 | if (!page_ref_freeze(p, 1)) { | |
1802 | pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); | |
1803 | goto out_error; | |
1804 | } | |
1805 | } else { | |
1806 | VM_BUG_ON_PAGE(page_count(p), p); | |
7118fc29 | 1807 | } |
1d798ca3 | 1808 | set_compound_head(p, page); |
20a0307c | 1809 | } |
b4330afb | 1810 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1811 | atomic_set(compound_pincount_ptr(page), 0); |
7118fc29 MK |
1812 | return true; |
1813 | ||
1814 | out_error: | |
1815 | /* undo tail page modifications made above */ | |
1816 | p = page + 1; | |
1817 | for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { | |
1818 | clear_compound_head(p); | |
1819 | set_page_refcounted(p); | |
1820 | } | |
1821 | /* need to clear PG_reserved on remaining tail pages */ | |
1822 | for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) | |
1823 | __ClearPageReserved(p); | |
1824 | set_compound_order(page, 0); | |
5232c63f | 1825 | #ifdef CONFIG_64BIT |
7118fc29 | 1826 | page[1].compound_nr = 0; |
5232c63f | 1827 | #endif |
7118fc29 MK |
1828 | __ClearPageHead(page); |
1829 | return false; | |
20a0307c WF |
1830 | } |
1831 | ||
34d9e35b MK |
1832 | static bool prep_compound_gigantic_page(struct page *page, unsigned int order) |
1833 | { | |
1834 | return __prep_compound_gigantic_page(page, order, false); | |
1835 | } | |
1836 | ||
8531fc6f MK |
1837 | static bool prep_compound_gigantic_page_for_demote(struct page *page, |
1838 | unsigned int order) | |
1839 | { | |
1840 | return __prep_compound_gigantic_page(page, order, true); | |
1841 | } | |
1842 | ||
7795912c AM |
1843 | /* |
1844 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1845 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1846 | * details. | |
1847 | */ | |
20a0307c WF |
1848 | int PageHuge(struct page *page) |
1849 | { | |
20a0307c WF |
1850 | if (!PageCompound(page)) |
1851 | return 0; | |
1852 | ||
1853 | page = compound_head(page); | |
f1e61557 | 1854 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1855 | } |
43131e14 NH |
1856 | EXPORT_SYMBOL_GPL(PageHuge); |
1857 | ||
27c73ae7 AA |
1858 | /* |
1859 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1860 | * normal or transparent huge pages. | |
1861 | */ | |
1862 | int PageHeadHuge(struct page *page_head) | |
1863 | { | |
27c73ae7 AA |
1864 | if (!PageHead(page_head)) |
1865 | return 0; | |
1866 | ||
d4af73e3 | 1867 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1868 | } |
4e936ecc | 1869 | EXPORT_SYMBOL_GPL(PageHeadHuge); |
27c73ae7 | 1870 | |
c0d0381a MK |
1871 | /* |
1872 | * Find and lock address space (mapping) in write mode. | |
1873 | * | |
336bf30e MK |
1874 | * Upon entry, the page is locked which means that page_mapping() is |
1875 | * stable. Due to locking order, we can only trylock_write. If we can | |
1876 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1877 | */ |
1878 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1879 | { | |
336bf30e | 1880 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1881 | |
c0d0381a MK |
1882 | if (!mapping) |
1883 | return mapping; | |
1884 | ||
c0d0381a MK |
1885 | if (i_mmap_trylock_write(mapping)) |
1886 | return mapping; | |
1887 | ||
336bf30e | 1888 | return NULL; |
c0d0381a MK |
1889 | } |
1890 | ||
fe19bd3d | 1891 | pgoff_t hugetlb_basepage_index(struct page *page) |
13d60f4b ZY |
1892 | { |
1893 | struct page *page_head = compound_head(page); | |
1894 | pgoff_t index = page_index(page_head); | |
1895 | unsigned long compound_idx; | |
1896 | ||
13d60f4b ZY |
1897 | if (compound_order(page_head) >= MAX_ORDER) |
1898 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1899 | else | |
1900 | compound_idx = page - page_head; | |
1901 | ||
1902 | return (index << compound_order(page_head)) + compound_idx; | |
1903 | } | |
1904 | ||
0c397dae | 1905 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1906 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1907 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1908 | { |
af0fb9df | 1909 | int order = huge_page_order(h); |
1da177e4 | 1910 | struct page *page; |
f60858f9 | 1911 | bool alloc_try_hard = true; |
f96efd58 | 1912 | |
f60858f9 MK |
1913 | /* |
1914 | * By default we always try hard to allocate the page with | |
1915 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1916 | * a loop (to adjust global huge page counts) and previous allocation | |
1917 | * failed, do not continue to try hard on the same node. Use the | |
1918 | * node_alloc_noretry bitmap to manage this state information. | |
1919 | */ | |
1920 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1921 | alloc_try_hard = false; | |
1922 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1923 | if (alloc_try_hard) | |
1924 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1925 | if (nid == NUMA_NO_NODE) |
1926 | nid = numa_mem_id(); | |
84172f4b | 1927 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
af0fb9df MH |
1928 | if (page) |
1929 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1930 | else | |
1931 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1932 | |
f60858f9 MK |
1933 | /* |
1934 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1935 | * indicates an overall state change. Clear bit so that we resume | |
1936 | * normal 'try hard' allocations. | |
1937 | */ | |
1938 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1939 | node_clear(nid, *node_alloc_noretry); | |
1940 | ||
1941 | /* | |
1942 | * If we tried hard to get a page but failed, set bit so that | |
1943 | * subsequent attempts will not try as hard until there is an | |
1944 | * overall state change. | |
1945 | */ | |
1946 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1947 | node_set(nid, *node_alloc_noretry); | |
1948 | ||
63b4613c NA |
1949 | return page; |
1950 | } | |
1951 | ||
0c397dae MH |
1952 | /* |
1953 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1954 | * should use this function to get new hugetlb pages | |
1955 | */ | |
1956 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1957 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1958 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1959 | { |
1960 | struct page *page; | |
7118fc29 | 1961 | bool retry = false; |
0c397dae | 1962 | |
7118fc29 | 1963 | retry: |
0c397dae MH |
1964 | if (hstate_is_gigantic(h)) |
1965 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1966 | else | |
1967 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1968 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1969 | if (!page) |
1970 | return NULL; | |
1971 | ||
7118fc29 MK |
1972 | if (hstate_is_gigantic(h)) { |
1973 | if (!prep_compound_gigantic_page(page, huge_page_order(h))) { | |
1974 | /* | |
1975 | * Rare failure to convert pages to compound page. | |
1976 | * Free pages and try again - ONCE! | |
1977 | */ | |
1978 | free_gigantic_page(page, huge_page_order(h)); | |
1979 | if (!retry) { | |
1980 | retry = true; | |
1981 | goto retry; | |
1982 | } | |
7118fc29 MK |
1983 | return NULL; |
1984 | } | |
1985 | } | |
0c397dae MH |
1986 | prep_new_huge_page(h, page, page_to_nid(page)); |
1987 | ||
1988 | return page; | |
1989 | } | |
1990 | ||
af0fb9df MH |
1991 | /* |
1992 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1993 | * manner. | |
1994 | */ | |
f60858f9 MK |
1995 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1996 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1997 | { |
1998 | struct page *page; | |
1999 | int nr_nodes, node; | |
af0fb9df | 2000 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
2001 | |
2002 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
2003 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
2004 | node_alloc_noretry); | |
af0fb9df | 2005 | if (page) |
b2261026 | 2006 | break; |
b2261026 JK |
2007 | } |
2008 | ||
af0fb9df MH |
2009 | if (!page) |
2010 | return 0; | |
b2261026 | 2011 | |
af0fb9df MH |
2012 | put_page(page); /* free it into the hugepage allocator */ |
2013 | ||
2014 | return 1; | |
b2261026 JK |
2015 | } |
2016 | ||
e8c5c824 | 2017 | /* |
10c6ec49 MK |
2018 | * Remove huge page from pool from next node to free. Attempt to keep |
2019 | * persistent huge pages more or less balanced over allowed nodes. | |
2020 | * This routine only 'removes' the hugetlb page. The caller must make | |
2021 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
2022 | * Called with hugetlb_lock locked. |
2023 | */ | |
10c6ec49 MK |
2024 | static struct page *remove_pool_huge_page(struct hstate *h, |
2025 | nodemask_t *nodes_allowed, | |
2026 | bool acct_surplus) | |
e8c5c824 | 2027 | { |
b2261026 | 2028 | int nr_nodes, node; |
10c6ec49 | 2029 | struct page *page = NULL; |
e8c5c824 | 2030 | |
9487ca60 | 2031 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 2032 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
2033 | /* |
2034 | * If we're returning unused surplus pages, only examine | |
2035 | * nodes with surplus pages. | |
2036 | */ | |
b2261026 JK |
2037 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
2038 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 2039 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 2040 | struct page, lru); |
6eb4e88a | 2041 | remove_hugetlb_page(h, page, acct_surplus); |
9a76db09 | 2042 | break; |
e8c5c824 | 2043 | } |
b2261026 | 2044 | } |
e8c5c824 | 2045 | |
10c6ec49 | 2046 | return page; |
e8c5c824 LS |
2047 | } |
2048 | ||
c8721bbb NH |
2049 | /* |
2050 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
2051 | * nothing for in-use hugepages and non-hugepages. |
2052 | * This function returns values like below: | |
2053 | * | |
ad2fa371 MS |
2054 | * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages |
2055 | * when the system is under memory pressure and the feature of | |
2056 | * freeing unused vmemmap pages associated with each hugetlb page | |
2057 | * is enabled. | |
2058 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
2059 | * (allocated or reserved.) | |
2060 | * 0: successfully dissolved free hugepages or the page is not a | |
2061 | * hugepage (considered as already dissolved) | |
c8721bbb | 2062 | */ |
c3114a84 | 2063 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 2064 | { |
6bc9b564 | 2065 | int rc = -EBUSY; |
082d5b6b | 2066 | |
7ffddd49 | 2067 | retry: |
faf53def NH |
2068 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
2069 | if (!PageHuge(page)) | |
2070 | return 0; | |
2071 | ||
db71ef79 | 2072 | spin_lock_irq(&hugetlb_lock); |
faf53def NH |
2073 | if (!PageHuge(page)) { |
2074 | rc = 0; | |
2075 | goto out; | |
2076 | } | |
2077 | ||
2078 | if (!page_count(page)) { | |
2247bb33 GS |
2079 | struct page *head = compound_head(page); |
2080 | struct hstate *h = page_hstate(head); | |
6bc9b564 | 2081 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 2082 | goto out; |
7ffddd49 MS |
2083 | |
2084 | /* | |
2085 | * We should make sure that the page is already on the free list | |
2086 | * when it is dissolved. | |
2087 | */ | |
6c037149 | 2088 | if (unlikely(!HPageFreed(head))) { |
db71ef79 | 2089 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
2090 | cond_resched(); |
2091 | ||
2092 | /* | |
2093 | * Theoretically, we should return -EBUSY when we | |
2094 | * encounter this race. In fact, we have a chance | |
2095 | * to successfully dissolve the page if we do a | |
2096 | * retry. Because the race window is quite small. | |
2097 | * If we seize this opportunity, it is an optimization | |
2098 | * for increasing the success rate of dissolving page. | |
2099 | */ | |
2100 | goto retry; | |
2101 | } | |
2102 | ||
0c5da357 | 2103 | remove_hugetlb_page(h, head, false); |
c1470b33 | 2104 | h->max_huge_pages--; |
db71ef79 | 2105 | spin_unlock_irq(&hugetlb_lock); |
ad2fa371 MS |
2106 | |
2107 | /* | |
2108 | * Normally update_and_free_page will allocate required vmemmmap | |
2109 | * before freeing the page. update_and_free_page will fail to | |
2110 | * free the page if it can not allocate required vmemmap. We | |
2111 | * need to adjust max_huge_pages if the page is not freed. | |
2112 | * Attempt to allocate vmemmmap here so that we can take | |
2113 | * appropriate action on failure. | |
2114 | */ | |
5981611d | 2115 | rc = hugetlb_vmemmap_alloc(h, head); |
ad2fa371 MS |
2116 | if (!rc) { |
2117 | /* | |
2118 | * Move PageHWPoison flag from head page to the raw | |
2119 | * error page, which makes any subpages rather than | |
2120 | * the error page reusable. | |
2121 | */ | |
2122 | if (PageHWPoison(head) && page != head) { | |
2123 | SetPageHWPoison(page); | |
2124 | ClearPageHWPoison(head); | |
2125 | } | |
2126 | update_and_free_page(h, head, false); | |
2127 | } else { | |
2128 | spin_lock_irq(&hugetlb_lock); | |
2129 | add_hugetlb_page(h, head, false); | |
2130 | h->max_huge_pages++; | |
2131 | spin_unlock_irq(&hugetlb_lock); | |
2132 | } | |
2133 | ||
2134 | return rc; | |
c8721bbb | 2135 | } |
082d5b6b | 2136 | out: |
db71ef79 | 2137 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 2138 | return rc; |
c8721bbb NH |
2139 | } |
2140 | ||
2141 | /* | |
2142 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
2143 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
2144 | * Note that this will dissolve a free gigantic hugepage completely, if any |
2145 | * part of it lies within the given range. | |
082d5b6b GS |
2146 | * Also note that if dissolve_free_huge_page() returns with an error, all |
2147 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 2148 | */ |
082d5b6b | 2149 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 2150 | { |
c8721bbb | 2151 | unsigned long pfn; |
eb03aa00 | 2152 | struct page *page; |
082d5b6b | 2153 | int rc = 0; |
c8721bbb | 2154 | |
d0177639 | 2155 | if (!hugepages_supported()) |
082d5b6b | 2156 | return rc; |
d0177639 | 2157 | |
eb03aa00 GS |
2158 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
2159 | page = pfn_to_page(pfn); | |
faf53def NH |
2160 | rc = dissolve_free_huge_page(page); |
2161 | if (rc) | |
2162 | break; | |
eb03aa00 | 2163 | } |
082d5b6b GS |
2164 | |
2165 | return rc; | |
c8721bbb NH |
2166 | } |
2167 | ||
ab5ac90a MH |
2168 | /* |
2169 | * Allocates a fresh surplus page from the page allocator. | |
2170 | */ | |
0c397dae | 2171 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
b65a4eda | 2172 | int nid, nodemask_t *nmask, bool zero_ref) |
7893d1d5 | 2173 | { |
9980d744 | 2174 | struct page *page = NULL; |
b65a4eda | 2175 | bool retry = false; |
7893d1d5 | 2176 | |
bae7f4ae | 2177 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2178 | return NULL; |
2179 | ||
db71ef79 | 2180 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2181 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
2182 | goto out_unlock; | |
db71ef79 | 2183 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 2184 | |
b65a4eda | 2185 | retry: |
f60858f9 | 2186 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 2187 | if (!page) |
0c397dae | 2188 | return NULL; |
d1c3fb1f | 2189 | |
db71ef79 | 2190 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2191 | /* |
2192 | * We could have raced with the pool size change. | |
2193 | * Double check that and simply deallocate the new page | |
2194 | * if we would end up overcommiting the surpluses. Abuse | |
2195 | * temporary page to workaround the nasty free_huge_page | |
2196 | * codeflow | |
2197 | */ | |
2198 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 2199 | SetHPageTemporary(page); |
db71ef79 | 2200 | spin_unlock_irq(&hugetlb_lock); |
9980d744 | 2201 | put_page(page); |
2bf753e6 | 2202 | return NULL; |
7893d1d5 | 2203 | } |
9980d744 | 2204 | |
b65a4eda MK |
2205 | if (zero_ref) { |
2206 | /* | |
2207 | * Caller requires a page with zero ref count. | |
2208 | * We will drop ref count here. If someone else is holding | |
2209 | * a ref, the page will be freed when they drop it. Abuse | |
2210 | * temporary page flag to accomplish this. | |
2211 | */ | |
2212 | SetHPageTemporary(page); | |
2213 | if (!put_page_testzero(page)) { | |
2214 | /* | |
2215 | * Unexpected inflated ref count on freshly allocated | |
2216 | * huge. Retry once. | |
2217 | */ | |
2218 | pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); | |
2219 | spin_unlock_irq(&hugetlb_lock); | |
2220 | if (retry) | |
2221 | return NULL; | |
2222 | ||
2223 | retry = true; | |
2224 | goto retry; | |
2225 | } | |
2226 | ClearHPageTemporary(page); | |
2227 | } | |
2228 | ||
2229 | h->surplus_huge_pages++; | |
2230 | h->surplus_huge_pages_node[page_to_nid(page)]++; | |
2231 | ||
9980d744 | 2232 | out_unlock: |
db71ef79 | 2233 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 AL |
2234 | |
2235 | return page; | |
2236 | } | |
2237 | ||
bbe88753 | 2238 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 2239 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
2240 | { |
2241 | struct page *page; | |
2242 | ||
2243 | if (hstate_is_gigantic(h)) | |
2244 | return NULL; | |
2245 | ||
f60858f9 | 2246 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
2247 | if (!page) |
2248 | return NULL; | |
2249 | ||
2250 | /* | |
2251 | * We do not account these pages as surplus because they are only | |
2252 | * temporary and will be released properly on the last reference | |
2253 | */ | |
9157c311 | 2254 | SetHPageTemporary(page); |
ab5ac90a MH |
2255 | |
2256 | return page; | |
2257 | } | |
2258 | ||
099730d6 DH |
2259 | /* |
2260 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
2261 | */ | |
e0ec90ee | 2262 | static |
0c397dae | 2263 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
2264 | struct vm_area_struct *vma, unsigned long addr) |
2265 | { | |
cfcaa66f | 2266 | struct page *page = NULL; |
aaf14e40 MH |
2267 | struct mempolicy *mpol; |
2268 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
2269 | int nid; | |
2270 | nodemask_t *nodemask; | |
2271 | ||
2272 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
cfcaa66f BW |
2273 | if (mpol_is_preferred_many(mpol)) { |
2274 | gfp_t gfp = gfp_mask | __GFP_NOWARN; | |
2275 | ||
2276 | gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL); | |
2277 | page = alloc_surplus_huge_page(h, gfp, nid, nodemask, false); | |
aaf14e40 | 2278 | |
cfcaa66f BW |
2279 | /* Fallback to all nodes if page==NULL */ |
2280 | nodemask = NULL; | |
2281 | } | |
2282 | ||
2283 | if (!page) | |
2284 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); | |
2285 | mpol_cond_put(mpol); | |
aaf14e40 | 2286 | return page; |
099730d6 DH |
2287 | } |
2288 | ||
ab5ac90a | 2289 | /* page migration callback function */ |
3e59fcb0 | 2290 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 2291 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 2292 | { |
db71ef79 | 2293 | spin_lock_irq(&hugetlb_lock); |
4db9b2ef | 2294 | if (h->free_huge_pages - h->resv_huge_pages > 0) { |
3e59fcb0 MH |
2295 | struct page *page; |
2296 | ||
2297 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
2298 | if (page) { | |
db71ef79 | 2299 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 2300 | return page; |
4db9b2ef MH |
2301 | } |
2302 | } | |
db71ef79 | 2303 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 2304 | |
0c397dae | 2305 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2306 | } |
2307 | ||
ebd63723 | 2308 | /* mempolicy aware migration callback */ |
389c8178 MH |
2309 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2310 | unsigned long address) | |
ebd63723 MH |
2311 | { |
2312 | struct mempolicy *mpol; | |
2313 | nodemask_t *nodemask; | |
2314 | struct page *page; | |
ebd63723 MH |
2315 | gfp_t gfp_mask; |
2316 | int node; | |
2317 | ||
ebd63723 MH |
2318 | gfp_mask = htlb_alloc_mask(h); |
2319 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 2320 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
2321 | mpol_cond_put(mpol); |
2322 | ||
2323 | return page; | |
2324 | } | |
2325 | ||
e4e574b7 | 2326 | /* |
25985edc | 2327 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2328 | * of size 'delta'. |
2329 | */ | |
0a4f3d1b | 2330 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 2331 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
2332 | { |
2333 | struct list_head surplus_list; | |
2334 | struct page *page, *tmp; | |
0a4f3d1b LX |
2335 | int ret; |
2336 | long i; | |
2337 | long needed, allocated; | |
28073b02 | 2338 | bool alloc_ok = true; |
e4e574b7 | 2339 | |
9487ca60 | 2340 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 2341 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2342 | if (needed <= 0) { |
a5516438 | 2343 | h->resv_huge_pages += delta; |
e4e574b7 | 2344 | return 0; |
ac09b3a1 | 2345 | } |
e4e574b7 AL |
2346 | |
2347 | allocated = 0; | |
2348 | INIT_LIST_HEAD(&surplus_list); | |
2349 | ||
2350 | ret = -ENOMEM; | |
2351 | retry: | |
db71ef79 | 2352 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 2353 | for (i = 0; i < needed; i++) { |
0c397dae | 2354 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
b65a4eda | 2355 | NUMA_NO_NODE, NULL, true); |
28073b02 HD |
2356 | if (!page) { |
2357 | alloc_ok = false; | |
2358 | break; | |
2359 | } | |
e4e574b7 | 2360 | list_add(&page->lru, &surplus_list); |
69ed779a | 2361 | cond_resched(); |
e4e574b7 | 2362 | } |
28073b02 | 2363 | allocated += i; |
e4e574b7 AL |
2364 | |
2365 | /* | |
2366 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2367 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2368 | */ | |
db71ef79 | 2369 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
2370 | needed = (h->resv_huge_pages + delta) - |
2371 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2372 | if (needed > 0) { |
2373 | if (alloc_ok) | |
2374 | goto retry; | |
2375 | /* | |
2376 | * We were not able to allocate enough pages to | |
2377 | * satisfy the entire reservation so we free what | |
2378 | * we've allocated so far. | |
2379 | */ | |
2380 | goto free; | |
2381 | } | |
e4e574b7 AL |
2382 | /* |
2383 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2384 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2385 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2386 | * allocator. Commit the entire reservation here to prevent another |
2387 | * process from stealing the pages as they are added to the pool but | |
2388 | * before they are reserved. | |
e4e574b7 AL |
2389 | */ |
2390 | needed += allocated; | |
a5516438 | 2391 | h->resv_huge_pages += delta; |
e4e574b7 | 2392 | ret = 0; |
a9869b83 | 2393 | |
19fc3f0a | 2394 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2395 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2396 | if ((--needed) < 0) |
2397 | break; | |
b65a4eda | 2398 | /* Add the page to the hugetlb allocator */ |
a5516438 | 2399 | enqueue_huge_page(h, page); |
19fc3f0a | 2400 | } |
28073b02 | 2401 | free: |
db71ef79 | 2402 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a | 2403 | |
b65a4eda MK |
2404 | /* |
2405 | * Free unnecessary surplus pages to the buddy allocator. | |
2406 | * Pages have no ref count, call free_huge_page directly. | |
2407 | */ | |
c0d934ba | 2408 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
b65a4eda | 2409 | free_huge_page(page); |
db71ef79 | 2410 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2411 | |
2412 | return ret; | |
2413 | } | |
2414 | ||
2415 | /* | |
e5bbc8a6 MK |
2416 | * This routine has two main purposes: |
2417 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2418 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2419 | * to the associated reservation map. | |
2420 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2421 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2422 | */ |
a5516438 AK |
2423 | static void return_unused_surplus_pages(struct hstate *h, |
2424 | unsigned long unused_resv_pages) | |
e4e574b7 | 2425 | { |
e4e574b7 | 2426 | unsigned long nr_pages; |
10c6ec49 MK |
2427 | struct page *page; |
2428 | LIST_HEAD(page_list); | |
2429 | ||
9487ca60 | 2430 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2431 | /* Uncommit the reservation */ |
2432 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2433 | |
aa888a74 | 2434 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2435 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2436 | goto out; |
aa888a74 | 2437 | |
e5bbc8a6 MK |
2438 | /* |
2439 | * Part (or even all) of the reservation could have been backed | |
2440 | * by pre-allocated pages. Only free surplus pages. | |
2441 | */ | |
a5516438 | 2442 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2443 | |
685f3457 LS |
2444 | /* |
2445 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2446 | * evenly across all nodes with memory. Iterate across these nodes |
2447 | * until we can no longer free unreserved surplus pages. This occurs | |
2448 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2449 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2450 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2451 | */ |
2452 | while (nr_pages--) { | |
10c6ec49 MK |
2453 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2454 | if (!page) | |
e5bbc8a6 | 2455 | goto out; |
10c6ec49 MK |
2456 | |
2457 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2458 | } |
e5bbc8a6 MK |
2459 | |
2460 | out: | |
db71ef79 | 2461 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2462 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2463 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2464 | } |
2465 | ||
5e911373 | 2466 | |
c37f9fb1 | 2467 | /* |
feba16e2 | 2468 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2469 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2470 | * |
2471 | * vma_needs_reservation is called to determine if the huge page at addr | |
2472 | * within the vma has an associated reservation. If a reservation is | |
2473 | * needed, the value 1 is returned. The caller is then responsible for | |
2474 | * managing the global reservation and subpool usage counts. After | |
2475 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2476 | * to add the page to the reservation map. If the page allocation fails, |
2477 | * the reservation must be ended instead of committed. vma_end_reservation | |
2478 | * is called in such cases. | |
cf3ad20b MK |
2479 | * |
2480 | * In the normal case, vma_commit_reservation returns the same value | |
2481 | * as the preceding vma_needs_reservation call. The only time this | |
2482 | * is not the case is if a reserve map was changed between calls. It | |
2483 | * is the responsibility of the caller to notice the difference and | |
2484 | * take appropriate action. | |
96b96a96 MK |
2485 | * |
2486 | * vma_add_reservation is used in error paths where a reservation must | |
2487 | * be restored when a newly allocated huge page must be freed. It is | |
2488 | * to be called after calling vma_needs_reservation to determine if a | |
2489 | * reservation exists. | |
846be085 MK |
2490 | * |
2491 | * vma_del_reservation is used in error paths where an entry in the reserve | |
2492 | * map was created during huge page allocation and must be removed. It is to | |
2493 | * be called after calling vma_needs_reservation to determine if a reservation | |
2494 | * exists. | |
c37f9fb1 | 2495 | */ |
5e911373 MK |
2496 | enum vma_resv_mode { |
2497 | VMA_NEEDS_RESV, | |
2498 | VMA_COMMIT_RESV, | |
feba16e2 | 2499 | VMA_END_RESV, |
96b96a96 | 2500 | VMA_ADD_RESV, |
846be085 | 2501 | VMA_DEL_RESV, |
5e911373 | 2502 | }; |
cf3ad20b MK |
2503 | static long __vma_reservation_common(struct hstate *h, |
2504 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2505 | enum vma_resv_mode mode) |
c37f9fb1 | 2506 | { |
4e35f483 JK |
2507 | struct resv_map *resv; |
2508 | pgoff_t idx; | |
cf3ad20b | 2509 | long ret; |
0db9d74e | 2510 | long dummy_out_regions_needed; |
c37f9fb1 | 2511 | |
4e35f483 JK |
2512 | resv = vma_resv_map(vma); |
2513 | if (!resv) | |
84afd99b | 2514 | return 1; |
c37f9fb1 | 2515 | |
4e35f483 | 2516 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2517 | switch (mode) { |
2518 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2519 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2520 | /* We assume that vma_reservation_* routines always operate on | |
2521 | * 1 page, and that adding to resv map a 1 page entry can only | |
2522 | * ever require 1 region. | |
2523 | */ | |
2524 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2525 | break; |
2526 | case VMA_COMMIT_RESV: | |
075a61d0 | 2527 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2528 | /* region_add calls of range 1 should never fail. */ |
2529 | VM_BUG_ON(ret < 0); | |
5e911373 | 2530 | break; |
feba16e2 | 2531 | case VMA_END_RESV: |
0db9d74e | 2532 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2533 | ret = 0; |
2534 | break; | |
96b96a96 | 2535 | case VMA_ADD_RESV: |
0db9d74e | 2536 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2537 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2538 | /* region_add calls of range 1 should never fail. */ |
2539 | VM_BUG_ON(ret < 0); | |
2540 | } else { | |
2541 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2542 | ret = region_del(resv, idx, idx + 1); |
2543 | } | |
2544 | break; | |
846be085 MK |
2545 | case VMA_DEL_RESV: |
2546 | if (vma->vm_flags & VM_MAYSHARE) { | |
2547 | region_abort(resv, idx, idx + 1, 1); | |
2548 | ret = region_del(resv, idx, idx + 1); | |
2549 | } else { | |
2550 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2551 | /* region_add calls of range 1 should never fail. */ | |
2552 | VM_BUG_ON(ret < 0); | |
2553 | } | |
2554 | break; | |
5e911373 MK |
2555 | default: |
2556 | BUG(); | |
2557 | } | |
84afd99b | 2558 | |
846be085 | 2559 | if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) |
cf3ad20b | 2560 | return ret; |
bf3d12b9 ML |
2561 | /* |
2562 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2563 | * | |
2564 | * In most cases, reserves always exist for private mappings. | |
2565 | * However, a file associated with mapping could have been | |
2566 | * hole punched or truncated after reserves were consumed. | |
2567 | * As subsequent fault on such a range will not use reserves. | |
2568 | * Subtle - The reserve map for private mappings has the | |
2569 | * opposite meaning than that of shared mappings. If NO | |
2570 | * entry is in the reserve map, it means a reservation exists. | |
2571 | * If an entry exists in the reserve map, it means the | |
2572 | * reservation has already been consumed. As a result, the | |
2573 | * return value of this routine is the opposite of the | |
2574 | * value returned from reserve map manipulation routines above. | |
2575 | */ | |
2576 | if (ret > 0) | |
2577 | return 0; | |
2578 | if (ret == 0) | |
2579 | return 1; | |
2580 | return ret; | |
c37f9fb1 | 2581 | } |
cf3ad20b MK |
2582 | |
2583 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2584 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2585 | { |
5e911373 | 2586 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2587 | } |
84afd99b | 2588 | |
cf3ad20b MK |
2589 | static long vma_commit_reservation(struct hstate *h, |
2590 | struct vm_area_struct *vma, unsigned long addr) | |
2591 | { | |
5e911373 MK |
2592 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2593 | } | |
2594 | ||
feba16e2 | 2595 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2596 | struct vm_area_struct *vma, unsigned long addr) |
2597 | { | |
feba16e2 | 2598 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2599 | } |
2600 | ||
96b96a96 MK |
2601 | static long vma_add_reservation(struct hstate *h, |
2602 | struct vm_area_struct *vma, unsigned long addr) | |
2603 | { | |
2604 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2605 | } | |
2606 | ||
846be085 MK |
2607 | static long vma_del_reservation(struct hstate *h, |
2608 | struct vm_area_struct *vma, unsigned long addr) | |
2609 | { | |
2610 | return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); | |
2611 | } | |
2612 | ||
96b96a96 | 2613 | /* |
846be085 MK |
2614 | * This routine is called to restore reservation information on error paths. |
2615 | * It should ONLY be called for pages allocated via alloc_huge_page(), and | |
2616 | * the hugetlb mutex should remain held when calling this routine. | |
2617 | * | |
2618 | * It handles two specific cases: | |
2619 | * 1) A reservation was in place and the page consumed the reservation. | |
2620 | * HPageRestoreReserve is set in the page. | |
2621 | * 2) No reservation was in place for the page, so HPageRestoreReserve is | |
2622 | * not set. However, alloc_huge_page always updates the reserve map. | |
2623 | * | |
2624 | * In case 1, free_huge_page later in the error path will increment the | |
2625 | * global reserve count. But, free_huge_page does not have enough context | |
2626 | * to adjust the reservation map. This case deals primarily with private | |
2627 | * mappings. Adjust the reserve map here to be consistent with global | |
2628 | * reserve count adjustments to be made by free_huge_page. Make sure the | |
2629 | * reserve map indicates there is a reservation present. | |
2630 | * | |
2631 | * In case 2, simply undo reserve map modifications done by alloc_huge_page. | |
96b96a96 | 2632 | */ |
846be085 MK |
2633 | void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, |
2634 | unsigned long address, struct page *page) | |
96b96a96 | 2635 | { |
846be085 | 2636 | long rc = vma_needs_reservation(h, vma, address); |
96b96a96 | 2637 | |
846be085 MK |
2638 | if (HPageRestoreReserve(page)) { |
2639 | if (unlikely(rc < 0)) | |
96b96a96 MK |
2640 | /* |
2641 | * Rare out of memory condition in reserve map | |
d6995da3 | 2642 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2643 | * global reserve count will not be incremented |
2644 | * by free_huge_page. This will make it appear | |
2645 | * as though the reservation for this page was | |
2646 | * consumed. This may prevent the task from | |
2647 | * faulting in the page at a later time. This | |
2648 | * is better than inconsistent global huge page | |
2649 | * accounting of reserve counts. | |
2650 | */ | |
d6995da3 | 2651 | ClearHPageRestoreReserve(page); |
846be085 MK |
2652 | else if (rc) |
2653 | (void)vma_add_reservation(h, vma, address); | |
2654 | else | |
2655 | vma_end_reservation(h, vma, address); | |
2656 | } else { | |
2657 | if (!rc) { | |
2658 | /* | |
2659 | * This indicates there is an entry in the reserve map | |
c7b1850d | 2660 | * not added by alloc_huge_page. We know it was added |
846be085 MK |
2661 | * before the alloc_huge_page call, otherwise |
2662 | * HPageRestoreReserve would be set on the page. | |
2663 | * Remove the entry so that a subsequent allocation | |
2664 | * does not consume a reservation. | |
2665 | */ | |
2666 | rc = vma_del_reservation(h, vma, address); | |
2667 | if (rc < 0) | |
96b96a96 | 2668 | /* |
846be085 MK |
2669 | * VERY rare out of memory condition. Since |
2670 | * we can not delete the entry, set | |
2671 | * HPageRestoreReserve so that the reserve | |
2672 | * count will be incremented when the page | |
2673 | * is freed. This reserve will be consumed | |
2674 | * on a subsequent allocation. | |
96b96a96 | 2675 | */ |
846be085 MK |
2676 | SetHPageRestoreReserve(page); |
2677 | } else if (rc < 0) { | |
2678 | /* | |
2679 | * Rare out of memory condition from | |
2680 | * vma_needs_reservation call. Memory allocation is | |
2681 | * only attempted if a new entry is needed. Therefore, | |
2682 | * this implies there is not an entry in the | |
2683 | * reserve map. | |
2684 | * | |
2685 | * For shared mappings, no entry in the map indicates | |
2686 | * no reservation. We are done. | |
2687 | */ | |
2688 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
2689 | /* | |
2690 | * For private mappings, no entry indicates | |
2691 | * a reservation is present. Since we can | |
2692 | * not add an entry, set SetHPageRestoreReserve | |
2693 | * on the page so reserve count will be | |
2694 | * incremented when freed. This reserve will | |
2695 | * be consumed on a subsequent allocation. | |
2696 | */ | |
2697 | SetHPageRestoreReserve(page); | |
96b96a96 | 2698 | } else |
846be085 MK |
2699 | /* |
2700 | * No reservation present, do nothing | |
2701 | */ | |
2702 | vma_end_reservation(h, vma, address); | |
96b96a96 MK |
2703 | } |
2704 | } | |
2705 | ||
369fa227 OS |
2706 | /* |
2707 | * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one | |
2708 | * @h: struct hstate old page belongs to | |
2709 | * @old_page: Old page to dissolve | |
ae37c7ff | 2710 | * @list: List to isolate the page in case we need to |
369fa227 OS |
2711 | * Returns 0 on success, otherwise negated error. |
2712 | */ | |
ae37c7ff OS |
2713 | static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, |
2714 | struct list_head *list) | |
369fa227 OS |
2715 | { |
2716 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
2717 | int nid = page_to_nid(old_page); | |
b65a4eda | 2718 | bool alloc_retry = false; |
369fa227 OS |
2719 | struct page *new_page; |
2720 | int ret = 0; | |
2721 | ||
2722 | /* | |
2723 | * Before dissolving the page, we need to allocate a new one for the | |
f41f2ed4 MS |
2724 | * pool to remain stable. Here, we allocate the page and 'prep' it |
2725 | * by doing everything but actually updating counters and adding to | |
2726 | * the pool. This simplifies and let us do most of the processing | |
2727 | * under the lock. | |
369fa227 | 2728 | */ |
b65a4eda | 2729 | alloc_retry: |
369fa227 OS |
2730 | new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); |
2731 | if (!new_page) | |
2732 | return -ENOMEM; | |
b65a4eda MK |
2733 | /* |
2734 | * If all goes well, this page will be directly added to the free | |
2735 | * list in the pool. For this the ref count needs to be zero. | |
2736 | * Attempt to drop now, and retry once if needed. It is VERY | |
2737 | * unlikely there is another ref on the page. | |
2738 | * | |
2739 | * If someone else has a reference to the page, it will be freed | |
2740 | * when they drop their ref. Abuse temporary page flag to accomplish | |
2741 | * this. Retry once if there is an inflated ref count. | |
2742 | */ | |
2743 | SetHPageTemporary(new_page); | |
2744 | if (!put_page_testzero(new_page)) { | |
2745 | if (alloc_retry) | |
2746 | return -EBUSY; | |
2747 | ||
2748 | alloc_retry = true; | |
2749 | goto alloc_retry; | |
2750 | } | |
2751 | ClearHPageTemporary(new_page); | |
2752 | ||
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 OS |
2767 | spin_unlock_irq(&hugetlb_lock); |
2768 | if (!isolate_huge_page(old_page, list)) | |
2769 | ret = -EBUSY; | |
2770 | spin_lock_irq(&hugetlb_lock); | |
369fa227 OS |
2771 | goto free_new; |
2772 | } else if (!HPageFreed(old_page)) { | |
2773 | /* | |
2774 | * Page's refcount is 0 but it has not been enqueued in the | |
2775 | * freelist yet. Race window is small, so we can succeed here if | |
2776 | * we retry. | |
2777 | */ | |
2778 | spin_unlock_irq(&hugetlb_lock); | |
2779 | cond_resched(); | |
2780 | goto retry; | |
2781 | } else { | |
2782 | /* | |
2783 | * Ok, old_page is still a genuine free hugepage. Remove it from | |
2784 | * the freelist and decrease the counters. These will be | |
2785 | * incremented again when calling __prep_account_new_huge_page() | |
2786 | * and enqueue_huge_page() for new_page. The counters will remain | |
2787 | * stable since this happens under the lock. | |
2788 | */ | |
2789 | remove_hugetlb_page(h, old_page, false); | |
2790 | ||
2791 | /* | |
b65a4eda MK |
2792 | * Ref count on new page is already zero as it was dropped |
2793 | * earlier. It can be directly added to the pool free list. | |
369fa227 | 2794 | */ |
369fa227 | 2795 | __prep_account_new_huge_page(h, nid); |
369fa227 OS |
2796 | enqueue_huge_page(h, new_page); |
2797 | ||
2798 | /* | |
2799 | * Pages have been replaced, we can safely free the old one. | |
2800 | */ | |
2801 | spin_unlock_irq(&hugetlb_lock); | |
b65d4adb | 2802 | update_and_free_page(h, old_page, false); |
369fa227 OS |
2803 | } |
2804 | ||
2805 | return ret; | |
2806 | ||
2807 | free_new: | |
2808 | spin_unlock_irq(&hugetlb_lock); | |
b65a4eda MK |
2809 | /* Page has a zero ref count, but needs a ref to be freed */ |
2810 | set_page_refcounted(new_page); | |
b65d4adb | 2811 | update_and_free_page(h, new_page, false); |
369fa227 OS |
2812 | |
2813 | return ret; | |
2814 | } | |
2815 | ||
ae37c7ff | 2816 | int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) |
369fa227 OS |
2817 | { |
2818 | struct hstate *h; | |
2819 | struct page *head; | |
ae37c7ff | 2820 | int ret = -EBUSY; |
369fa227 OS |
2821 | |
2822 | /* | |
2823 | * The page might have been dissolved from under our feet, so make sure | |
2824 | * to carefully check the state under the lock. | |
2825 | * Return success when racing as if we dissolved the page ourselves. | |
2826 | */ | |
2827 | spin_lock_irq(&hugetlb_lock); | |
2828 | if (PageHuge(page)) { | |
2829 | head = compound_head(page); | |
2830 | h = page_hstate(head); | |
2831 | } else { | |
2832 | spin_unlock_irq(&hugetlb_lock); | |
2833 | return 0; | |
2834 | } | |
2835 | spin_unlock_irq(&hugetlb_lock); | |
2836 | ||
2837 | /* | |
2838 | * Fence off gigantic pages as there is a cyclic dependency between | |
2839 | * alloc_contig_range and them. Return -ENOMEM as this has the effect | |
2840 | * of bailing out right away without further retrying. | |
2841 | */ | |
2842 | if (hstate_is_gigantic(h)) | |
2843 | return -ENOMEM; | |
2844 | ||
ae37c7ff OS |
2845 | if (page_count(head) && isolate_huge_page(head, list)) |
2846 | ret = 0; | |
2847 | else if (!page_count(head)) | |
2848 | ret = alloc_and_dissolve_huge_page(h, head, list); | |
2849 | ||
2850 | return ret; | |
369fa227 OS |
2851 | } |
2852 | ||
70c3547e | 2853 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2854 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2855 | { |
90481622 | 2856 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2857 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2858 | struct page *page; |
d85f69b0 MK |
2859 | long map_chg, map_commit; |
2860 | long gbl_chg; | |
6d76dcf4 AK |
2861 | int ret, idx; |
2862 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2863 | bool deferred_reserve; |
a1e78772 | 2864 | |
6d76dcf4 | 2865 | idx = hstate_index(h); |
a1e78772 | 2866 | /* |
d85f69b0 MK |
2867 | * Examine the region/reserve map to determine if the process |
2868 | * has a reservation for the page to be allocated. A return | |
2869 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2870 | */ |
d85f69b0 MK |
2871 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2872 | if (map_chg < 0) | |
76dcee75 | 2873 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2874 | |
2875 | /* | |
2876 | * Processes that did not create the mapping will have no | |
2877 | * reserves as indicated by the region/reserve map. Check | |
2878 | * that the allocation will not exceed the subpool limit. | |
2879 | * Allocations for MAP_NORESERVE mappings also need to be | |
2880 | * checked against any subpool limit. | |
2881 | */ | |
2882 | if (map_chg || avoid_reserve) { | |
2883 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2884 | if (gbl_chg < 0) { | |
feba16e2 | 2885 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2886 | return ERR_PTR(-ENOSPC); |
5e911373 | 2887 | } |
1da177e4 | 2888 | |
d85f69b0 MK |
2889 | /* |
2890 | * Even though there was no reservation in the region/reserve | |
2891 | * map, there could be reservations associated with the | |
2892 | * subpool that can be used. This would be indicated if the | |
2893 | * return value of hugepage_subpool_get_pages() is zero. | |
2894 | * However, if avoid_reserve is specified we still avoid even | |
2895 | * the subpool reservations. | |
2896 | */ | |
2897 | if (avoid_reserve) | |
2898 | gbl_chg = 1; | |
2899 | } | |
2900 | ||
08cf9faf MA |
2901 | /* If this allocation is not consuming a reservation, charge it now. |
2902 | */ | |
6501fe5f | 2903 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2904 | if (deferred_reserve) { |
2905 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2906 | idx, pages_per_huge_page(h), &h_cg); | |
2907 | if (ret) | |
2908 | goto out_subpool_put; | |
2909 | } | |
2910 | ||
6d76dcf4 | 2911 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2912 | if (ret) |
08cf9faf | 2913 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2914 | |
db71ef79 | 2915 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
2916 | /* |
2917 | * glb_chg is passed to indicate whether or not a page must be taken | |
2918 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2919 | * a reservation exists for the allocation. | |
2920 | */ | |
2921 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2922 | if (!page) { |
db71ef79 | 2923 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 2924 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2925 | if (!page) |
2926 | goto out_uncharge_cgroup; | |
a88c7695 | 2927 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2928 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2929 | h->resv_huge_pages--; |
2930 | } | |
db71ef79 | 2931 | spin_lock_irq(&hugetlb_lock); |
15a8d68e | 2932 | list_add(&page->lru, &h->hugepage_activelist); |
81a6fcae | 2933 | /* Fall through */ |
68842c9b | 2934 | } |
81a6fcae | 2935 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2936 | /* If allocation is not consuming a reservation, also store the |
2937 | * hugetlb_cgroup pointer on the page. | |
2938 | */ | |
2939 | if (deferred_reserve) { | |
2940 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2941 | h_cg, page); | |
2942 | } | |
2943 | ||
db71ef79 | 2944 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 2945 | |
d6995da3 | 2946 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2947 | |
d85f69b0 MK |
2948 | map_commit = vma_commit_reservation(h, vma, addr); |
2949 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2950 | /* |
2951 | * The page was added to the reservation map between | |
2952 | * vma_needs_reservation and vma_commit_reservation. | |
2953 | * This indicates a race with hugetlb_reserve_pages. | |
2954 | * Adjust for the subpool count incremented above AND | |
2955 | * in hugetlb_reserve_pages for the same page. Also, | |
2956 | * the reservation count added in hugetlb_reserve_pages | |
2957 | * no longer applies. | |
2958 | */ | |
2959 | long rsv_adjust; | |
2960 | ||
2961 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2962 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2963 | if (deferred_reserve) |
2964 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2965 | pages_per_huge_page(h), page); | |
33039678 | 2966 | } |
90d8b7e6 | 2967 | return page; |
8f34af6f JZ |
2968 | |
2969 | out_uncharge_cgroup: | |
2970 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2971 | out_uncharge_cgroup_reservation: |
2972 | if (deferred_reserve) | |
2973 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2974 | h_cg); | |
8f34af6f | 2975 | out_subpool_put: |
d85f69b0 | 2976 | if (map_chg || avoid_reserve) |
8f34af6f | 2977 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2978 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2979 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2980 | } |
2981 | ||
b5389086 | 2982 | int alloc_bootmem_huge_page(struct hstate *h, int nid) |
e24a1307 | 2983 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); |
b5389086 | 2984 | int __alloc_bootmem_huge_page(struct hstate *h, int nid) |
aa888a74 | 2985 | { |
b5389086 | 2986 | struct huge_bootmem_page *m = NULL; /* initialize for clang */ |
b2261026 | 2987 | int nr_nodes, node; |
aa888a74 | 2988 | |
b5389086 ZY |
2989 | /* do node specific alloc */ |
2990 | if (nid != NUMA_NO_NODE) { | |
2991 | m = memblock_alloc_try_nid_raw(huge_page_size(h), huge_page_size(h), | |
2992 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, nid); | |
2993 | if (!m) | |
2994 | return 0; | |
2995 | goto found; | |
2996 | } | |
2997 | /* allocate from next node when distributing huge pages */ | |
b2261026 | 2998 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
b5389086 | 2999 | m = memblock_alloc_try_nid_raw( |
8b89a116 | 3000 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 3001 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
b5389086 ZY |
3002 | /* |
3003 | * Use the beginning of the huge page to store the | |
3004 | * huge_bootmem_page struct (until gather_bootmem | |
3005 | * puts them into the mem_map). | |
3006 | */ | |
3007 | if (!m) | |
3008 | return 0; | |
3009 | goto found; | |
aa888a74 | 3010 | } |
aa888a74 AK |
3011 | |
3012 | found: | |
aa888a74 | 3013 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 3014 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
3015 | list_add(&m->list, &huge_boot_pages); |
3016 | m->hstate = h; | |
3017 | return 1; | |
3018 | } | |
3019 | ||
48b8d744 MK |
3020 | /* |
3021 | * Put bootmem huge pages into the standard lists after mem_map is up. | |
3022 | * Note: This only applies to gigantic (order > MAX_ORDER) pages. | |
3023 | */ | |
aa888a74 AK |
3024 | static void __init gather_bootmem_prealloc(void) |
3025 | { | |
3026 | struct huge_bootmem_page *m; | |
3027 | ||
3028 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 3029 | struct page *page = virt_to_page(m); |
aa888a74 | 3030 | struct hstate *h = m->hstate; |
ee8f248d | 3031 | |
48b8d744 | 3032 | VM_BUG_ON(!hstate_is_gigantic(h)); |
aa888a74 | 3033 | WARN_ON(page_count(page) != 1); |
7118fc29 MK |
3034 | if (prep_compound_gigantic_page(page, huge_page_order(h))) { |
3035 | WARN_ON(PageReserved(page)); | |
3036 | prep_new_huge_page(h, page, page_to_nid(page)); | |
3037 | put_page(page); /* add to the hugepage allocator */ | |
3038 | } else { | |
416d85ed | 3039 | /* VERY unlikely inflated ref count on a tail page */ |
7118fc29 | 3040 | free_gigantic_page(page, huge_page_order(h)); |
7118fc29 | 3041 | } |
af0fb9df | 3042 | |
b0320c7b | 3043 | /* |
48b8d744 MK |
3044 | * We need to restore the 'stolen' pages to totalram_pages |
3045 | * in order to fix confusing memory reports from free(1) and | |
3046 | * other side-effects, like CommitLimit going negative. | |
b0320c7b | 3047 | */ |
48b8d744 | 3048 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 3049 | cond_resched(); |
aa888a74 AK |
3050 | } |
3051 | } | |
b5389086 ZY |
3052 | static void __init hugetlb_hstate_alloc_pages_onenode(struct hstate *h, int nid) |
3053 | { | |
3054 | unsigned long i; | |
3055 | char buf[32]; | |
3056 | ||
3057 | for (i = 0; i < h->max_huge_pages_node[nid]; ++i) { | |
3058 | if (hstate_is_gigantic(h)) { | |
3059 | if (!alloc_bootmem_huge_page(h, nid)) | |
3060 | break; | |
3061 | } else { | |
3062 | struct page *page; | |
3063 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
3064 | ||
3065 | page = alloc_fresh_huge_page(h, gfp_mask, nid, | |
3066 | &node_states[N_MEMORY], NULL); | |
3067 | if (!page) | |
3068 | break; | |
3069 | put_page(page); /* free it into the hugepage allocator */ | |
3070 | } | |
3071 | cond_resched(); | |
3072 | } | |
3073 | if (i == h->max_huge_pages_node[nid]) | |
3074 | return; | |
3075 | ||
3076 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
3077 | pr_warn("HugeTLB: allocating %u of page size %s failed node%d. Only allocated %lu hugepages.\n", | |
3078 | h->max_huge_pages_node[nid], buf, nid, i); | |
3079 | h->max_huge_pages -= (h->max_huge_pages_node[nid] - i); | |
3080 | h->max_huge_pages_node[nid] = i; | |
3081 | } | |
aa888a74 | 3082 | |
8faa8b07 | 3083 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
3084 | { |
3085 | unsigned long i; | |
f60858f9 | 3086 | nodemask_t *node_alloc_noretry; |
b5389086 ZY |
3087 | bool node_specific_alloc = false; |
3088 | ||
3089 | /* skip gigantic hugepages allocation if hugetlb_cma enabled */ | |
3090 | if (hstate_is_gigantic(h) && hugetlb_cma_size) { | |
3091 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); | |
3092 | return; | |
3093 | } | |
3094 | ||
3095 | /* do node specific alloc */ | |
0a7a0f6f | 3096 | for_each_online_node(i) { |
b5389086 ZY |
3097 | if (h->max_huge_pages_node[i] > 0) { |
3098 | hugetlb_hstate_alloc_pages_onenode(h, i); | |
3099 | node_specific_alloc = true; | |
3100 | } | |
3101 | } | |
f60858f9 | 3102 | |
b5389086 ZY |
3103 | if (node_specific_alloc) |
3104 | return; | |
3105 | ||
3106 | /* below will do all node balanced alloc */ | |
f60858f9 MK |
3107 | if (!hstate_is_gigantic(h)) { |
3108 | /* | |
3109 | * Bit mask controlling how hard we retry per-node allocations. | |
3110 | * Ignore errors as lower level routines can deal with | |
3111 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
3112 | * time, we are likely in bigger trouble. | |
3113 | */ | |
3114 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
3115 | GFP_KERNEL); | |
3116 | } else { | |
3117 | /* allocations done at boot time */ | |
3118 | node_alloc_noretry = NULL; | |
3119 | } | |
3120 | ||
3121 | /* bit mask controlling how hard we retry per-node allocations */ | |
3122 | if (node_alloc_noretry) | |
3123 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 3124 | |
e5ff2159 | 3125 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 3126 | if (hstate_is_gigantic(h)) { |
b5389086 | 3127 | if (!alloc_bootmem_huge_page(h, NUMA_NO_NODE)) |
aa888a74 | 3128 | break; |
0c397dae | 3129 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
3130 | &node_states[N_MEMORY], |
3131 | node_alloc_noretry)) | |
1da177e4 | 3132 | break; |
69ed779a | 3133 | cond_resched(); |
1da177e4 | 3134 | } |
d715cf80 LH |
3135 | if (i < h->max_huge_pages) { |
3136 | char buf[32]; | |
3137 | ||
c6247f72 | 3138 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
3139 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
3140 | h->max_huge_pages, buf, i); | |
3141 | h->max_huge_pages = i; | |
3142 | } | |
f60858f9 | 3143 | kfree(node_alloc_noretry); |
e5ff2159 AK |
3144 | } |
3145 | ||
3146 | static void __init hugetlb_init_hstates(void) | |
3147 | { | |
79dfc695 | 3148 | struct hstate *h, *h2; |
e5ff2159 AK |
3149 | |
3150 | for_each_hstate(h) { | |
641844f5 NH |
3151 | if (minimum_order > huge_page_order(h)) |
3152 | minimum_order = huge_page_order(h); | |
3153 | ||
8faa8b07 | 3154 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 3155 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 3156 | hugetlb_hstate_alloc_pages(h); |
79dfc695 MK |
3157 | |
3158 | /* | |
3159 | * Set demote order for each hstate. Note that | |
3160 | * h->demote_order is initially 0. | |
3161 | * - We can not demote gigantic pages if runtime freeing | |
3162 | * is not supported, so skip this. | |
a01f4390 MK |
3163 | * - If CMA allocation is possible, we can not demote |
3164 | * HUGETLB_PAGE_ORDER or smaller size pages. | |
79dfc695 MK |
3165 | */ |
3166 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) | |
3167 | continue; | |
a01f4390 MK |
3168 | if (hugetlb_cma_size && h->order <= HUGETLB_PAGE_ORDER) |
3169 | continue; | |
79dfc695 MK |
3170 | for_each_hstate(h2) { |
3171 | if (h2 == h) | |
3172 | continue; | |
3173 | if (h2->order < h->order && | |
3174 | h2->order > h->demote_order) | |
3175 | h->demote_order = h2->order; | |
3176 | } | |
e5ff2159 | 3177 | } |
641844f5 | 3178 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
3179 | } |
3180 | ||
3181 | static void __init report_hugepages(void) | |
3182 | { | |
3183 | struct hstate *h; | |
3184 | ||
3185 | for_each_hstate(h) { | |
4abd32db | 3186 | char buf[32]; |
c6247f72 MW |
3187 | |
3188 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 3189 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 3190 | buf, h->free_huge_pages); |
e5ff2159 AK |
3191 | } |
3192 | } | |
3193 | ||
1da177e4 | 3194 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
3195 | static void try_to_free_low(struct hstate *h, unsigned long count, |
3196 | nodemask_t *nodes_allowed) | |
1da177e4 | 3197 | { |
4415cc8d | 3198 | int i; |
1121828a | 3199 | LIST_HEAD(page_list); |
4415cc8d | 3200 | |
9487ca60 | 3201 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 3202 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
3203 | return; |
3204 | ||
1121828a MK |
3205 | /* |
3206 | * Collect pages to be freed on a list, and free after dropping lock | |
3207 | */ | |
6ae11b27 | 3208 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 3209 | struct page *page, *next; |
a5516438 AK |
3210 | struct list_head *freel = &h->hugepage_freelists[i]; |
3211 | list_for_each_entry_safe(page, next, freel, lru) { | |
3212 | if (count >= h->nr_huge_pages) | |
1121828a | 3213 | goto out; |
1da177e4 LT |
3214 | if (PageHighMem(page)) |
3215 | continue; | |
6eb4e88a | 3216 | remove_hugetlb_page(h, page, false); |
1121828a | 3217 | list_add(&page->lru, &page_list); |
1da177e4 LT |
3218 | } |
3219 | } | |
1121828a MK |
3220 | |
3221 | out: | |
db71ef79 | 3222 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3223 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 3224 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
3225 | } |
3226 | #else | |
6ae11b27 LS |
3227 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
3228 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
3229 | { |
3230 | } | |
3231 | #endif | |
3232 | ||
20a0307c WF |
3233 | /* |
3234 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
3235 | * balanced by operating on them in a round-robin fashion. | |
3236 | * Returns 1 if an adjustment was made. | |
3237 | */ | |
6ae11b27 LS |
3238 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
3239 | int delta) | |
20a0307c | 3240 | { |
b2261026 | 3241 | int nr_nodes, node; |
20a0307c | 3242 | |
9487ca60 | 3243 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 3244 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 3245 | |
b2261026 JK |
3246 | if (delta < 0) { |
3247 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
3248 | if (h->surplus_huge_pages_node[node]) | |
3249 | goto found; | |
e8c5c824 | 3250 | } |
b2261026 JK |
3251 | } else { |
3252 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
3253 | if (h->surplus_huge_pages_node[node] < | |
3254 | h->nr_huge_pages_node[node]) | |
3255 | goto found; | |
e8c5c824 | 3256 | } |
b2261026 JK |
3257 | } |
3258 | return 0; | |
20a0307c | 3259 | |
b2261026 JK |
3260 | found: |
3261 | h->surplus_huge_pages += delta; | |
3262 | h->surplus_huge_pages_node[node] += delta; | |
3263 | return 1; | |
20a0307c WF |
3264 | } |
3265 | ||
a5516438 | 3266 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 3267 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 3268 | nodemask_t *nodes_allowed) |
1da177e4 | 3269 | { |
7893d1d5 | 3270 | unsigned long min_count, ret; |
10c6ec49 MK |
3271 | struct page *page; |
3272 | LIST_HEAD(page_list); | |
f60858f9 MK |
3273 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
3274 | ||
3275 | /* | |
3276 | * Bit mask controlling how hard we retry per-node allocations. | |
3277 | * If we can not allocate the bit mask, do not attempt to allocate | |
3278 | * the requested huge pages. | |
3279 | */ | |
3280 | if (node_alloc_noretry) | |
3281 | nodes_clear(*node_alloc_noretry); | |
3282 | else | |
3283 | return -ENOMEM; | |
1da177e4 | 3284 | |
29383967 MK |
3285 | /* |
3286 | * resize_lock mutex prevents concurrent adjustments to number of | |
3287 | * pages in hstate via the proc/sysfs interfaces. | |
3288 | */ | |
3289 | mutex_lock(&h->resize_lock); | |
b65d4adb | 3290 | flush_free_hpage_work(h); |
db71ef79 | 3291 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 3292 | |
fd875dca MK |
3293 | /* |
3294 | * Check for a node specific request. | |
3295 | * Changing node specific huge page count may require a corresponding | |
3296 | * change to the global count. In any case, the passed node mask | |
3297 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
3298 | */ | |
3299 | if (nid != NUMA_NO_NODE) { | |
3300 | unsigned long old_count = count; | |
3301 | ||
3302 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
3303 | /* | |
3304 | * User may have specified a large count value which caused the | |
3305 | * above calculation to overflow. In this case, they wanted | |
3306 | * to allocate as many huge pages as possible. Set count to | |
3307 | * largest possible value to align with their intention. | |
3308 | */ | |
3309 | if (count < old_count) | |
3310 | count = ULONG_MAX; | |
3311 | } | |
3312 | ||
4eb0716e AG |
3313 | /* |
3314 | * Gigantic pages runtime allocation depend on the capability for large | |
3315 | * page range allocation. | |
3316 | * If the system does not provide this feature, return an error when | |
3317 | * the user tries to allocate gigantic pages but let the user free the | |
3318 | * boottime allocated gigantic pages. | |
3319 | */ | |
3320 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
3321 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 3322 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3323 | mutex_unlock(&h->resize_lock); |
f60858f9 | 3324 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
3325 | return -EINVAL; |
3326 | } | |
3327 | /* Fall through to decrease pool */ | |
3328 | } | |
aa888a74 | 3329 | |
7893d1d5 AL |
3330 | /* |
3331 | * Increase the pool size | |
3332 | * First take pages out of surplus state. Then make up the | |
3333 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 3334 | * |
0c397dae | 3335 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
3336 | * to convert a surplus huge page to a normal huge page. That is |
3337 | * not critical, though, it just means the overall size of the | |
3338 | * pool might be one hugepage larger than it needs to be, but | |
3339 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 3340 | */ |
a5516438 | 3341 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 3342 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
3343 | break; |
3344 | } | |
3345 | ||
a5516438 | 3346 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
3347 | /* |
3348 | * If this allocation races such that we no longer need the | |
3349 | * page, free_huge_page will handle it by freeing the page | |
3350 | * and reducing the surplus. | |
3351 | */ | |
db71ef79 | 3352 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
3353 | |
3354 | /* yield cpu to avoid soft lockup */ | |
3355 | cond_resched(); | |
3356 | ||
f60858f9 MK |
3357 | ret = alloc_pool_huge_page(h, nodes_allowed, |
3358 | node_alloc_noretry); | |
db71ef79 | 3359 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
3360 | if (!ret) |
3361 | goto out; | |
3362 | ||
536240f2 MG |
3363 | /* Bail for signals. Probably ctrl-c from user */ |
3364 | if (signal_pending(current)) | |
3365 | goto out; | |
7893d1d5 | 3366 | } |
7893d1d5 AL |
3367 | |
3368 | /* | |
3369 | * Decrease the pool size | |
3370 | * First return free pages to the buddy allocator (being careful | |
3371 | * to keep enough around to satisfy reservations). Then place | |
3372 | * pages into surplus state as needed so the pool will shrink | |
3373 | * to the desired size as pages become free. | |
d1c3fb1f NA |
3374 | * |
3375 | * By placing pages into the surplus state independent of the | |
3376 | * overcommit value, we are allowing the surplus pool size to | |
3377 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 3378 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
3379 | * though, we'll note that we're not allowed to exceed surplus |
3380 | * and won't grow the pool anywhere else. Not until one of the | |
3381 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 3382 | */ |
a5516438 | 3383 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 3384 | min_count = max(count, min_count); |
6ae11b27 | 3385 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
3386 | |
3387 | /* | |
3388 | * Collect pages to be removed on list without dropping lock | |
3389 | */ | |
a5516438 | 3390 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
3391 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
3392 | if (!page) | |
1da177e4 | 3393 | break; |
10c6ec49 MK |
3394 | |
3395 | list_add(&page->lru, &page_list); | |
1da177e4 | 3396 | } |
10c6ec49 | 3397 | /* free the pages after dropping lock */ |
db71ef79 | 3398 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3399 | update_and_free_pages_bulk(h, &page_list); |
b65d4adb | 3400 | flush_free_hpage_work(h); |
db71ef79 | 3401 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 3402 | |
a5516438 | 3403 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 3404 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
3405 | break; |
3406 | } | |
3407 | out: | |
4eb0716e | 3408 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 3409 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3410 | mutex_unlock(&h->resize_lock); |
4eb0716e | 3411 | |
f60858f9 MK |
3412 | NODEMASK_FREE(node_alloc_noretry); |
3413 | ||
4eb0716e | 3414 | return 0; |
1da177e4 LT |
3415 | } |
3416 | ||
8531fc6f MK |
3417 | static int demote_free_huge_page(struct hstate *h, struct page *page) |
3418 | { | |
3419 | int i, nid = page_to_nid(page); | |
3420 | struct hstate *target_hstate; | |
3421 | int rc = 0; | |
3422 | ||
3423 | target_hstate = size_to_hstate(PAGE_SIZE << h->demote_order); | |
3424 | ||
3425 | remove_hugetlb_page_for_demote(h, page, false); | |
3426 | spin_unlock_irq(&hugetlb_lock); | |
3427 | ||
5981611d | 3428 | rc = hugetlb_vmemmap_alloc(h, page); |
8531fc6f MK |
3429 | if (rc) { |
3430 | /* Allocation of vmemmmap failed, we can not demote page */ | |
3431 | spin_lock_irq(&hugetlb_lock); | |
3432 | set_page_refcounted(page); | |
3433 | add_hugetlb_page(h, page, false); | |
3434 | return rc; | |
3435 | } | |
3436 | ||
3437 | /* | |
3438 | * Use destroy_compound_hugetlb_page_for_demote for all huge page | |
3439 | * sizes as it will not ref count pages. | |
3440 | */ | |
3441 | destroy_compound_hugetlb_page_for_demote(page, huge_page_order(h)); | |
3442 | ||
3443 | /* | |
3444 | * Taking target hstate mutex synchronizes with set_max_huge_pages. | |
3445 | * Without the mutex, pages added to target hstate could be marked | |
3446 | * as surplus. | |
3447 | * | |
3448 | * Note that we already hold h->resize_lock. To prevent deadlock, | |
3449 | * use the convention of always taking larger size hstate mutex first. | |
3450 | */ | |
3451 | mutex_lock(&target_hstate->resize_lock); | |
3452 | for (i = 0; i < pages_per_huge_page(h); | |
3453 | i += pages_per_huge_page(target_hstate)) { | |
3454 | if (hstate_is_gigantic(target_hstate)) | |
3455 | prep_compound_gigantic_page_for_demote(page + i, | |
3456 | target_hstate->order); | |
3457 | else | |
3458 | prep_compound_page(page + i, target_hstate->order); | |
3459 | set_page_private(page + i, 0); | |
3460 | set_page_refcounted(page + i); | |
3461 | prep_new_huge_page(target_hstate, page + i, nid); | |
3462 | put_page(page + i); | |
3463 | } | |
3464 | mutex_unlock(&target_hstate->resize_lock); | |
3465 | ||
3466 | spin_lock_irq(&hugetlb_lock); | |
3467 | ||
3468 | /* | |
3469 | * Not absolutely necessary, but for consistency update max_huge_pages | |
3470 | * based on pool changes for the demoted page. | |
3471 | */ | |
3472 | h->max_huge_pages--; | |
3473 | target_hstate->max_huge_pages += pages_per_huge_page(h); | |
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; | |
3715 | int err = 0; | |
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 | { | |
3766 | int nid; | |
3767 | struct hstate *h = kobj_to_hstate(kobj, &nid); | |
3768 | unsigned long demote_size = (PAGE_SIZE << h->demote_order) / SZ_1K; | |
3769 | ||
3770 | return sysfs_emit(buf, "%lukB\n", demote_size); | |
3771 | } | |
3772 | ||
3773 | static ssize_t demote_size_store(struct kobject *kobj, | |
3774 | struct kobj_attribute *attr, | |
3775 | const char *buf, size_t count) | |
3776 | { | |
3777 | struct hstate *h, *demote_hstate; | |
3778 | unsigned long demote_size; | |
3779 | unsigned int demote_order; | |
3780 | int nid; | |
3781 | ||
3782 | demote_size = (unsigned long)memparse(buf, NULL); | |
3783 | ||
3784 | demote_hstate = size_to_hstate(demote_size); | |
3785 | if (!demote_hstate) | |
3786 | return -EINVAL; | |
3787 | demote_order = demote_hstate->order; | |
a01f4390 MK |
3788 | if (demote_order < HUGETLB_PAGE_ORDER) |
3789 | return -EINVAL; | |
79dfc695 MK |
3790 | |
3791 | /* demote order must be smaller than hstate order */ | |
3792 | h = kobj_to_hstate(kobj, &nid); | |
3793 | if (demote_order >= h->order) | |
3794 | return -EINVAL; | |
3795 | ||
3796 | /* resize_lock synchronizes access to demote size and writes */ | |
3797 | mutex_lock(&h->resize_lock); | |
3798 | h->demote_order = demote_order; | |
3799 | mutex_unlock(&h->resize_lock); | |
3800 | ||
3801 | return count; | |
3802 | } | |
3803 | HSTATE_ATTR(demote_size); | |
3804 | ||
a3437870 NA |
3805 | static struct attribute *hstate_attrs[] = { |
3806 | &nr_hugepages_attr.attr, | |
3807 | &nr_overcommit_hugepages_attr.attr, | |
3808 | &free_hugepages_attr.attr, | |
3809 | &resv_hugepages_attr.attr, | |
3810 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3811 | #ifdef CONFIG_NUMA |
3812 | &nr_hugepages_mempolicy_attr.attr, | |
3813 | #endif | |
a3437870 NA |
3814 | NULL, |
3815 | }; | |
3816 | ||
67e5ed96 | 3817 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3818 | .attrs = hstate_attrs, |
3819 | }; | |
3820 | ||
79dfc695 MK |
3821 | static struct attribute *hstate_demote_attrs[] = { |
3822 | &demote_size_attr.attr, | |
3823 | &demote_attr.attr, | |
3824 | NULL, | |
3825 | }; | |
3826 | ||
3827 | static const struct attribute_group hstate_demote_attr_group = { | |
3828 | .attrs = hstate_demote_attrs, | |
3829 | }; | |
3830 | ||
094e9539 JM |
3831 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3832 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3833 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3834 | { |
3835 | int retval; | |
972dc4de | 3836 | int hi = hstate_index(h); |
a3437870 | 3837 | |
9a305230 LS |
3838 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3839 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3840 | return -ENOMEM; |
3841 | ||
9a305230 | 3842 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 3843 | if (retval) { |
9a305230 | 3844 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 ML |
3845 | hstate_kobjs[hi] = NULL; |
3846 | } | |
a3437870 | 3847 | |
79dfc695 MK |
3848 | if (h->demote_order) { |
3849 | if (sysfs_create_group(hstate_kobjs[hi], | |
3850 | &hstate_demote_attr_group)) | |
3851 | pr_warn("HugeTLB unable to create demote interfaces for %s\n", h->name); | |
3852 | } | |
3853 | ||
a3437870 NA |
3854 | return retval; |
3855 | } | |
3856 | ||
3857 | static void __init hugetlb_sysfs_init(void) | |
3858 | { | |
3859 | struct hstate *h; | |
3860 | int err; | |
3861 | ||
3862 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3863 | if (!hugepages_kobj) | |
3864 | return; | |
3865 | ||
3866 | for_each_hstate(h) { | |
9a305230 LS |
3867 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3868 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3869 | if (err) |
282f4214 | 3870 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
3871 | } |
3872 | } | |
3873 | ||
9a305230 LS |
3874 | #ifdef CONFIG_NUMA |
3875 | ||
3876 | /* | |
3877 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3878 | * with node devices in node_devices[] using a parallel array. The array |
3879 | * index of a node device or _hstate == node id. | |
3880 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3881 | * the base kernel, on the hugetlb module. |
3882 | */ | |
3883 | struct node_hstate { | |
3884 | struct kobject *hugepages_kobj; | |
3885 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3886 | }; | |
b4e289a6 | 3887 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3888 | |
3889 | /* | |
10fbcf4c | 3890 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3891 | */ |
3892 | static struct attribute *per_node_hstate_attrs[] = { | |
3893 | &nr_hugepages_attr.attr, | |
3894 | &free_hugepages_attr.attr, | |
3895 | &surplus_hugepages_attr.attr, | |
3896 | NULL, | |
3897 | }; | |
3898 | ||
67e5ed96 | 3899 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3900 | .attrs = per_node_hstate_attrs, |
3901 | }; | |
3902 | ||
3903 | /* | |
10fbcf4c | 3904 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3905 | * Returns node id via non-NULL nidp. |
3906 | */ | |
3907 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3908 | { | |
3909 | int nid; | |
3910 | ||
3911 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3912 | struct node_hstate *nhs = &node_hstates[nid]; | |
3913 | int i; | |
3914 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3915 | if (nhs->hstate_kobjs[i] == kobj) { | |
3916 | if (nidp) | |
3917 | *nidp = nid; | |
3918 | return &hstates[i]; | |
3919 | } | |
3920 | } | |
3921 | ||
3922 | BUG(); | |
3923 | return NULL; | |
3924 | } | |
3925 | ||
3926 | /* | |
10fbcf4c | 3927 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3928 | * No-op if no hstate attributes attached. |
3929 | */ | |
3cd8b44f | 3930 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3931 | { |
3932 | struct hstate *h; | |
10fbcf4c | 3933 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3934 | |
3935 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3936 | return; /* no hstate attributes */ |
9a305230 | 3937 | |
972dc4de AK |
3938 | for_each_hstate(h) { |
3939 | int idx = hstate_index(h); | |
3940 | if (nhs->hstate_kobjs[idx]) { | |
3941 | kobject_put(nhs->hstate_kobjs[idx]); | |
3942 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3943 | } |
972dc4de | 3944 | } |
9a305230 LS |
3945 | |
3946 | kobject_put(nhs->hugepages_kobj); | |
3947 | nhs->hugepages_kobj = NULL; | |
3948 | } | |
3949 | ||
9a305230 LS |
3950 | |
3951 | /* | |
10fbcf4c | 3952 | * Register hstate attributes for a single node device. |
9a305230 LS |
3953 | * No-op if attributes already registered. |
3954 | */ | |
3cd8b44f | 3955 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3956 | { |
3957 | struct hstate *h; | |
10fbcf4c | 3958 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3959 | int err; |
3960 | ||
3961 | if (nhs->hugepages_kobj) | |
3962 | return; /* already allocated */ | |
3963 | ||
3964 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3965 | &node->dev.kobj); |
9a305230 LS |
3966 | if (!nhs->hugepages_kobj) |
3967 | return; | |
3968 | ||
3969 | for_each_hstate(h) { | |
3970 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3971 | nhs->hstate_kobjs, | |
3972 | &per_node_hstate_attr_group); | |
3973 | if (err) { | |
282f4214 | 3974 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3975 | h->name, node->dev.id); |
9a305230 LS |
3976 | hugetlb_unregister_node(node); |
3977 | break; | |
3978 | } | |
3979 | } | |
3980 | } | |
3981 | ||
3982 | /* | |
9b5e5d0f | 3983 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3984 | * devices of nodes that have memory. All on-line nodes should have |
3985 | * registered their associated device by this time. | |
9a305230 | 3986 | */ |
7d9ca000 | 3987 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3988 | { |
3989 | int nid; | |
3990 | ||
8cebfcd0 | 3991 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3992 | struct node *node = node_devices[nid]; |
10fbcf4c | 3993 | if (node->dev.id == nid) |
9a305230 LS |
3994 | hugetlb_register_node(node); |
3995 | } | |
3996 | ||
3997 | /* | |
10fbcf4c | 3998 | * Let the node device driver know we're here so it can |
9a305230 LS |
3999 | * [un]register hstate attributes on node hotplug. |
4000 | */ | |
4001 | register_hugetlbfs_with_node(hugetlb_register_node, | |
4002 | hugetlb_unregister_node); | |
4003 | } | |
4004 | #else /* !CONFIG_NUMA */ | |
4005 | ||
4006 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
4007 | { | |
4008 | BUG(); | |
4009 | if (nidp) | |
4010 | *nidp = -1; | |
4011 | return NULL; | |
4012 | } | |
4013 | ||
9a305230 LS |
4014 | static void hugetlb_register_all_nodes(void) { } |
4015 | ||
4016 | #endif | |
4017 | ||
a3437870 NA |
4018 | static int __init hugetlb_init(void) |
4019 | { | |
8382d914 DB |
4020 | int i; |
4021 | ||
d6995da3 MK |
4022 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
4023 | __NR_HPAGEFLAGS); | |
4024 | ||
c2833a5b MK |
4025 | if (!hugepages_supported()) { |
4026 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
4027 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 4028 | return 0; |
c2833a5b | 4029 | } |
a3437870 | 4030 | |
282f4214 MK |
4031 | /* |
4032 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
4033 | * architectures depend on setup being done here. | |
4034 | */ | |
4035 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
4036 | if (!parsed_default_hugepagesz) { | |
4037 | /* | |
4038 | * If we did not parse a default huge page size, set | |
4039 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
4040 | * number of huge pages for this default size was implicitly | |
4041 | * specified, set that here as well. | |
4042 | * Note that the implicit setting will overwrite an explicit | |
4043 | * setting. A warning will be printed in this case. | |
4044 | */ | |
4045 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
4046 | if (default_hstate_max_huge_pages) { | |
4047 | if (default_hstate.max_huge_pages) { | |
4048 | char buf[32]; | |
4049 | ||
4050 | string_get_size(huge_page_size(&default_hstate), | |
4051 | 1, STRING_UNITS_2, buf, 32); | |
4052 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
4053 | default_hstate.max_huge_pages, buf); | |
4054 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
4055 | default_hstate_max_huge_pages); | |
4056 | } | |
4057 | default_hstate.max_huge_pages = | |
4058 | default_hstate_max_huge_pages; | |
b5389086 | 4059 | |
0a7a0f6f | 4060 | for_each_online_node(i) |
b5389086 ZY |
4061 | default_hstate.max_huge_pages_node[i] = |
4062 | default_hugepages_in_node[i]; | |
d715cf80 | 4063 | } |
f8b74815 | 4064 | } |
a3437870 | 4065 | |
cf11e85f | 4066 | hugetlb_cma_check(); |
a3437870 | 4067 | hugetlb_init_hstates(); |
aa888a74 | 4068 | gather_bootmem_prealloc(); |
a3437870 NA |
4069 | report_hugepages(); |
4070 | ||
4071 | hugetlb_sysfs_init(); | |
9a305230 | 4072 | hugetlb_register_all_nodes(); |
7179e7bf | 4073 | hugetlb_cgroup_file_init(); |
9a305230 | 4074 | |
8382d914 DB |
4075 | #ifdef CONFIG_SMP |
4076 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
4077 | #else | |
4078 | num_fault_mutexes = 1; | |
4079 | #endif | |
c672c7f2 | 4080 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
4081 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
4082 | GFP_KERNEL); | |
c672c7f2 | 4083 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
4084 | |
4085 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 4086 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
4087 | return 0; |
4088 | } | |
3e89e1c5 | 4089 | subsys_initcall(hugetlb_init); |
a3437870 | 4090 | |
ae94da89 MK |
4091 | /* Overwritten by architectures with more huge page sizes */ |
4092 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 4093 | { |
ae94da89 | 4094 | return size == HPAGE_SIZE; |
9fee021d VT |
4095 | } |
4096 | ||
d00181b9 | 4097 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
4098 | { |
4099 | struct hstate *h; | |
8faa8b07 AK |
4100 | unsigned long i; |
4101 | ||
a3437870 | 4102 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
4103 | return; |
4104 | } | |
47d38344 | 4105 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 4106 | BUG_ON(order == 0); |
47d38344 | 4107 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 4108 | mutex_init(&h->resize_lock); |
a3437870 | 4109 | h->order = order; |
aca78307 | 4110 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
4111 | for (i = 0; i < MAX_NUMNODES; ++i) |
4112 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 4113 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
4114 | h->next_nid_to_alloc = first_memory_node; |
4115 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
4116 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
4117 | huge_page_size(h)/1024); | |
77490587 | 4118 | hugetlb_vmemmap_init(h); |
8faa8b07 | 4119 | |
a3437870 NA |
4120 | parsed_hstate = h; |
4121 | } | |
4122 | ||
b5389086 ZY |
4123 | bool __init __weak hugetlb_node_alloc_supported(void) |
4124 | { | |
4125 | return true; | |
4126 | } | |
f87442f4 PL |
4127 | |
4128 | static void __init hugepages_clear_pages_in_node(void) | |
4129 | { | |
4130 | if (!hugetlb_max_hstate) { | |
4131 | default_hstate_max_huge_pages = 0; | |
4132 | memset(default_hugepages_in_node, 0, | |
4133 | MAX_NUMNODES * sizeof(unsigned int)); | |
4134 | } else { | |
4135 | parsed_hstate->max_huge_pages = 0; | |
4136 | memset(parsed_hstate->max_huge_pages_node, 0, | |
4137 | MAX_NUMNODES * sizeof(unsigned int)); | |
4138 | } | |
4139 | } | |
4140 | ||
282f4214 MK |
4141 | /* |
4142 | * hugepages command line processing | |
4143 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
4144 | * specification. If not, ignore the hugepages value. hugepages can also | |
4145 | * be the first huge page command line option in which case it implicitly | |
4146 | * specifies the number of huge pages for the default size. | |
4147 | */ | |
4148 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
4149 | { |
4150 | unsigned long *mhp; | |
8faa8b07 | 4151 | static unsigned long *last_mhp; |
b5389086 ZY |
4152 | int node = NUMA_NO_NODE; |
4153 | int count; | |
4154 | unsigned long tmp; | |
4155 | char *p = s; | |
a3437870 | 4156 | |
9fee021d | 4157 | if (!parsed_valid_hugepagesz) { |
282f4214 | 4158 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 4159 | parsed_valid_hugepagesz = true; |
f81f6e4b | 4160 | return 1; |
9fee021d | 4161 | } |
282f4214 | 4162 | |
a3437870 | 4163 | /* |
282f4214 MK |
4164 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
4165 | * yet, so this hugepages= parameter goes to the "default hstate". | |
4166 | * Otherwise, it goes with the previously parsed hugepagesz or | |
4167 | * default_hugepagesz. | |
a3437870 | 4168 | */ |
9fee021d | 4169 | else if (!hugetlb_max_hstate) |
a3437870 NA |
4170 | mhp = &default_hstate_max_huge_pages; |
4171 | else | |
4172 | mhp = &parsed_hstate->max_huge_pages; | |
4173 | ||
8faa8b07 | 4174 | if (mhp == last_mhp) { |
282f4214 | 4175 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
f81f6e4b | 4176 | return 1; |
8faa8b07 AK |
4177 | } |
4178 | ||
b5389086 ZY |
4179 | while (*p) { |
4180 | count = 0; | |
4181 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4182 | goto invalid; | |
4183 | /* Parameter is node format */ | |
4184 | if (p[count] == ':') { | |
4185 | if (!hugetlb_node_alloc_supported()) { | |
4186 | pr_warn("HugeTLB: architecture can't support node specific alloc, ignoring!\n"); | |
f81f6e4b | 4187 | return 1; |
b5389086 | 4188 | } |
0a7a0f6f | 4189 | if (tmp >= MAX_NUMNODES || !node_online(tmp)) |
e79ce983 | 4190 | goto invalid; |
0a7a0f6f | 4191 | node = array_index_nospec(tmp, MAX_NUMNODES); |
b5389086 | 4192 | p += count + 1; |
b5389086 ZY |
4193 | /* Parse hugepages */ |
4194 | if (sscanf(p, "%lu%n", &tmp, &count) != 1) | |
4195 | goto invalid; | |
4196 | if (!hugetlb_max_hstate) | |
4197 | default_hugepages_in_node[node] = tmp; | |
4198 | else | |
4199 | parsed_hstate->max_huge_pages_node[node] = tmp; | |
4200 | *mhp += tmp; | |
4201 | /* Go to parse next node*/ | |
4202 | if (p[count] == ',') | |
4203 | p += count + 1; | |
4204 | else | |
4205 | break; | |
4206 | } else { | |
4207 | if (p != s) | |
4208 | goto invalid; | |
4209 | *mhp = tmp; | |
4210 | break; | |
4211 | } | |
4212 | } | |
a3437870 | 4213 | |
8faa8b07 AK |
4214 | /* |
4215 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 4216 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
4217 | * use the bootmem allocator. |
4218 | */ | |
04adbc3f | 4219 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
4220 | hugetlb_hstate_alloc_pages(parsed_hstate); |
4221 | ||
4222 | last_mhp = mhp; | |
4223 | ||
a3437870 | 4224 | return 1; |
b5389086 ZY |
4225 | |
4226 | invalid: | |
4227 | pr_warn("HugeTLB: Invalid hugepages parameter %s\n", p); | |
f87442f4 | 4228 | hugepages_clear_pages_in_node(); |
f81f6e4b | 4229 | return 1; |
a3437870 | 4230 | } |
282f4214 | 4231 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 4232 | |
282f4214 MK |
4233 | /* |
4234 | * hugepagesz command line processing | |
4235 | * A specific huge page size can only be specified once with hugepagesz. | |
4236 | * hugepagesz is followed by hugepages on the command line. The global | |
4237 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
4238 | * hugepagesz argument was valid. | |
4239 | */ | |
359f2544 | 4240 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 4241 | { |
359f2544 | 4242 | unsigned long size; |
282f4214 MK |
4243 | struct hstate *h; |
4244 | ||
4245 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
4246 | size = (unsigned long)memparse(s, NULL); |
4247 | ||
4248 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4249 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
f81f6e4b | 4250 | return 1; |
359f2544 MK |
4251 | } |
4252 | ||
282f4214 MK |
4253 | h = size_to_hstate(size); |
4254 | if (h) { | |
4255 | /* | |
4256 | * hstate for this size already exists. This is normally | |
4257 | * an error, but is allowed if the existing hstate is the | |
4258 | * default hstate. More specifically, it is only allowed if | |
4259 | * the number of huge pages for the default hstate was not | |
4260 | * previously specified. | |
4261 | */ | |
4262 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
4263 | default_hstate.max_huge_pages) { | |
4264 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
f81f6e4b | 4265 | return 1; |
282f4214 MK |
4266 | } |
4267 | ||
4268 | /* | |
4269 | * No need to call hugetlb_add_hstate() as hstate already | |
4270 | * exists. But, do set parsed_hstate so that a following | |
4271 | * hugepages= parameter will be applied to this hstate. | |
4272 | */ | |
4273 | parsed_hstate = h; | |
4274 | parsed_valid_hugepagesz = true; | |
4275 | return 1; | |
38237830 MK |
4276 | } |
4277 | ||
359f2544 | 4278 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 4279 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
4280 | return 1; |
4281 | } | |
359f2544 MK |
4282 | __setup("hugepagesz=", hugepagesz_setup); |
4283 | ||
282f4214 MK |
4284 | /* |
4285 | * default_hugepagesz command line input | |
4286 | * Only one instance of default_hugepagesz allowed on command line. | |
4287 | */ | |
ae94da89 | 4288 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 4289 | { |
ae94da89 | 4290 | unsigned long size; |
b5389086 | 4291 | int i; |
ae94da89 | 4292 | |
282f4214 | 4293 | parsed_valid_hugepagesz = false; |
282f4214 MK |
4294 | if (parsed_default_hugepagesz) { |
4295 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
f81f6e4b | 4296 | return 1; |
282f4214 MK |
4297 | } |
4298 | ||
ae94da89 MK |
4299 | size = (unsigned long)memparse(s, NULL); |
4300 | ||
4301 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 4302 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
f81f6e4b | 4303 | return 1; |
ae94da89 MK |
4304 | } |
4305 | ||
282f4214 MK |
4306 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
4307 | parsed_valid_hugepagesz = true; | |
4308 | parsed_default_hugepagesz = true; | |
4309 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
4310 | ||
4311 | /* | |
4312 | * The number of default huge pages (for this size) could have been | |
4313 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
4314 | * then default_hstate_max_huge_pages is set. If the default huge | |
4315 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
4316 | * allocated here from bootmem allocator. | |
4317 | */ | |
4318 | if (default_hstate_max_huge_pages) { | |
4319 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
0a7a0f6f | 4320 | for_each_online_node(i) |
b5389086 ZY |
4321 | default_hstate.max_huge_pages_node[i] = |
4322 | default_hugepages_in_node[i]; | |
282f4214 MK |
4323 | if (hstate_is_gigantic(&default_hstate)) |
4324 | hugetlb_hstate_alloc_pages(&default_hstate); | |
4325 | default_hstate_max_huge_pages = 0; | |
4326 | } | |
4327 | ||
e11bfbfc NP |
4328 | return 1; |
4329 | } | |
ae94da89 | 4330 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 4331 | |
8ca39e68 | 4332 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
4333 | { |
4334 | int node; | |
4335 | unsigned int nr = 0; | |
8ca39e68 MS |
4336 | nodemask_t *mpol_allowed; |
4337 | unsigned int *array = h->free_huge_pages_node; | |
4338 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
4339 | ||
4340 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 4341 | |
8ca39e68 | 4342 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
c93b0a99 | 4343 | if (!mpol_allowed || node_isset(node, *mpol_allowed)) |
8ca39e68 MS |
4344 | nr += array[node]; |
4345 | } | |
8a213460 NA |
4346 | |
4347 | return nr; | |
4348 | } | |
4349 | ||
4350 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
4351 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
4352 | void *buffer, size_t *length, | |
4353 | loff_t *ppos, unsigned long *out) | |
4354 | { | |
4355 | struct ctl_table dup_table; | |
4356 | ||
4357 | /* | |
4358 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
4359 | * can duplicate the @table and alter the duplicate of it. | |
4360 | */ | |
4361 | dup_table = *table; | |
4362 | dup_table.data = out; | |
4363 | ||
4364 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
4365 | } | |
4366 | ||
06808b08 LS |
4367 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
4368 | struct ctl_table *table, int write, | |
32927393 | 4369 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 4370 | { |
e5ff2159 | 4371 | struct hstate *h = &default_hstate; |
238d3c13 | 4372 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 4373 | int ret; |
e5ff2159 | 4374 | |
457c1b27 | 4375 | if (!hugepages_supported()) |
86613628 | 4376 | return -EOPNOTSUPP; |
457c1b27 | 4377 | |
17743798 MS |
4378 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4379 | &tmp); | |
08d4a246 MH |
4380 | if (ret) |
4381 | goto out; | |
e5ff2159 | 4382 | |
238d3c13 DR |
4383 | if (write) |
4384 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
4385 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
4386 | out: |
4387 | return ret; | |
1da177e4 | 4388 | } |
396faf03 | 4389 | |
06808b08 | 4390 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 4391 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4392 | { |
4393 | ||
4394 | return hugetlb_sysctl_handler_common(false, table, write, | |
4395 | buffer, length, ppos); | |
4396 | } | |
4397 | ||
4398 | #ifdef CONFIG_NUMA | |
4399 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 4400 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
4401 | { |
4402 | return hugetlb_sysctl_handler_common(true, table, write, | |
4403 | buffer, length, ppos); | |
4404 | } | |
4405 | #endif /* CONFIG_NUMA */ | |
4406 | ||
a3d0c6aa | 4407 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 4408 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 4409 | { |
a5516438 | 4410 | struct hstate *h = &default_hstate; |
e5ff2159 | 4411 | unsigned long tmp; |
08d4a246 | 4412 | int ret; |
e5ff2159 | 4413 | |
457c1b27 | 4414 | if (!hugepages_supported()) |
86613628 | 4415 | return -EOPNOTSUPP; |
457c1b27 | 4416 | |
c033a93c | 4417 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 4418 | |
bae7f4ae | 4419 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
4420 | return -EINVAL; |
4421 | ||
17743798 MS |
4422 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
4423 | &tmp); | |
08d4a246 MH |
4424 | if (ret) |
4425 | goto out; | |
e5ff2159 AK |
4426 | |
4427 | if (write) { | |
db71ef79 | 4428 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 4429 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 4430 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 4431 | } |
08d4a246 MH |
4432 | out: |
4433 | return ret; | |
a3d0c6aa NA |
4434 | } |
4435 | ||
1da177e4 LT |
4436 | #endif /* CONFIG_SYSCTL */ |
4437 | ||
e1759c21 | 4438 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 4439 | { |
fcb2b0c5 RG |
4440 | struct hstate *h; |
4441 | unsigned long total = 0; | |
4442 | ||
457c1b27 NA |
4443 | if (!hugepages_supported()) |
4444 | return; | |
fcb2b0c5 RG |
4445 | |
4446 | for_each_hstate(h) { | |
4447 | unsigned long count = h->nr_huge_pages; | |
4448 | ||
aca78307 | 4449 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
4450 | |
4451 | if (h == &default_hstate) | |
4452 | seq_printf(m, | |
4453 | "HugePages_Total: %5lu\n" | |
4454 | "HugePages_Free: %5lu\n" | |
4455 | "HugePages_Rsvd: %5lu\n" | |
4456 | "HugePages_Surp: %5lu\n" | |
4457 | "Hugepagesize: %8lu kB\n", | |
4458 | count, | |
4459 | h->free_huge_pages, | |
4460 | h->resv_huge_pages, | |
4461 | h->surplus_huge_pages, | |
aca78307 | 4462 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
4463 | } |
4464 | ||
aca78307 | 4465 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
4466 | } |
4467 | ||
7981593b | 4468 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 4469 | { |
a5516438 | 4470 | struct hstate *h = &default_hstate; |
7981593b | 4471 | |
457c1b27 NA |
4472 | if (!hugepages_supported()) |
4473 | return 0; | |
7981593b JP |
4474 | |
4475 | return sysfs_emit_at(buf, len, | |
4476 | "Node %d HugePages_Total: %5u\n" | |
4477 | "Node %d HugePages_Free: %5u\n" | |
4478 | "Node %d HugePages_Surp: %5u\n", | |
4479 | nid, h->nr_huge_pages_node[nid], | |
4480 | nid, h->free_huge_pages_node[nid], | |
4481 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
4482 | } |
4483 | ||
949f7ec5 DR |
4484 | void hugetlb_show_meminfo(void) |
4485 | { | |
4486 | struct hstate *h; | |
4487 | int nid; | |
4488 | ||
457c1b27 NA |
4489 | if (!hugepages_supported()) |
4490 | return; | |
4491 | ||
949f7ec5 DR |
4492 | for_each_node_state(nid, N_MEMORY) |
4493 | for_each_hstate(h) | |
4494 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
4495 | nid, | |
4496 | h->nr_huge_pages_node[nid], | |
4497 | h->free_huge_pages_node[nid], | |
4498 | h->surplus_huge_pages_node[nid], | |
aca78307 | 4499 | huge_page_size(h) / SZ_1K); |
949f7ec5 DR |
4500 | } |
4501 | ||
5d317b2b NH |
4502 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
4503 | { | |
4504 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
4505 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
4506 | } | |
4507 | ||
1da177e4 LT |
4508 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
4509 | unsigned long hugetlb_total_pages(void) | |
4510 | { | |
d0028588 WL |
4511 | struct hstate *h; |
4512 | unsigned long nr_total_pages = 0; | |
4513 | ||
4514 | for_each_hstate(h) | |
4515 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
4516 | return nr_total_pages; | |
1da177e4 | 4517 | } |
1da177e4 | 4518 | |
a5516438 | 4519 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
4520 | { |
4521 | int ret = -ENOMEM; | |
4522 | ||
0aa7f354 ML |
4523 | if (!delta) |
4524 | return 0; | |
4525 | ||
db71ef79 | 4526 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4527 | /* |
4528 | * When cpuset is configured, it breaks the strict hugetlb page | |
4529 | * reservation as the accounting is done on a global variable. Such | |
4530 | * reservation is completely rubbish in the presence of cpuset because | |
4531 | * the reservation is not checked against page availability for the | |
4532 | * current cpuset. Application can still potentially OOM'ed by kernel | |
4533 | * with lack of free htlb page in cpuset that the task is in. | |
4534 | * Attempt to enforce strict accounting with cpuset is almost | |
4535 | * impossible (or too ugly) because cpuset is too fluid that | |
4536 | * task or memory node can be dynamically moved between cpusets. | |
4537 | * | |
4538 | * The change of semantics for shared hugetlb mapping with cpuset is | |
4539 | * undesirable. However, in order to preserve some of the semantics, | |
4540 | * we fall back to check against current free page availability as | |
4541 | * a best attempt and hopefully to minimize the impact of changing | |
4542 | * semantics that cpuset has. | |
8ca39e68 MS |
4543 | * |
4544 | * Apart from cpuset, we also have memory policy mechanism that | |
4545 | * also determines from which node the kernel will allocate memory | |
4546 | * in a NUMA system. So similar to cpuset, we also should consider | |
4547 | * the memory policy of the current task. Similar to the description | |
4548 | * above. | |
fc1b8a73 MG |
4549 | */ |
4550 | if (delta > 0) { | |
a5516438 | 4551 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
4552 | goto out; |
4553 | ||
8ca39e68 | 4554 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 4555 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
4556 | goto out; |
4557 | } | |
4558 | } | |
4559 | ||
4560 | ret = 0; | |
4561 | if (delta < 0) | |
a5516438 | 4562 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
4563 | |
4564 | out: | |
db71ef79 | 4565 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4566 | return ret; |
4567 | } | |
4568 | ||
84afd99b AW |
4569 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
4570 | { | |
f522c3ac | 4571 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
4572 | |
4573 | /* | |
4574 | * This new VMA should share its siblings reservation map if present. | |
4575 | * The VMA will only ever have a valid reservation map pointer where | |
4576 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 4577 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
4578 | * after this open call completes. It is therefore safe to take a |
4579 | * new reference here without additional locking. | |
4580 | */ | |
09a26e83 MK |
4581 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
4582 | resv_map_dup_hugetlb_cgroup_uncharge_info(resv); | |
f522c3ac | 4583 | kref_get(&resv->refs); |
09a26e83 | 4584 | } |
84afd99b AW |
4585 | } |
4586 | ||
a1e78772 MG |
4587 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
4588 | { | |
a5516438 | 4589 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 4590 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 4591 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 4592 | unsigned long reserve, start, end; |
1c5ecae3 | 4593 | long gbl_reserve; |
84afd99b | 4594 | |
4e35f483 JK |
4595 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4596 | return; | |
84afd99b | 4597 | |
4e35f483 JK |
4598 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
4599 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 4600 | |
4e35f483 | 4601 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 4602 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 4603 | if (reserve) { |
1c5ecae3 MK |
4604 | /* |
4605 | * Decrement reserve counts. The global reserve count may be | |
4606 | * adjusted if the subpool has a minimum size. | |
4607 | */ | |
4608 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
4609 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 4610 | } |
e9fe92ae MA |
4611 | |
4612 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
4613 | } |
4614 | ||
31383c68 DW |
4615 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
4616 | { | |
4617 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
4618 | return -EINVAL; | |
4619 | return 0; | |
4620 | } | |
4621 | ||
05ea8860 DW |
4622 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
4623 | { | |
aca78307 | 4624 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
4625 | } |
4626 | ||
1da177e4 LT |
4627 | /* |
4628 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
4629 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 4630 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
4631 | * this far. |
4632 | */ | |
b3ec9f33 | 4633 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
4634 | { |
4635 | BUG(); | |
d0217ac0 | 4636 | return 0; |
1da177e4 LT |
4637 | } |
4638 | ||
eec3636a JC |
4639 | /* |
4640 | * When a new function is introduced to vm_operations_struct and added | |
4641 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
4642 | * This is because under System V memory model, mappings created via | |
4643 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
4644 | * their original vm_ops are overwritten with shm_vm_ops. | |
4645 | */ | |
f0f37e2f | 4646 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 4647 | .fault = hugetlb_vm_op_fault, |
84afd99b | 4648 | .open = hugetlb_vm_op_open, |
a1e78772 | 4649 | .close = hugetlb_vm_op_close, |
dd3b614f | 4650 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 4651 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
4652 | }; |
4653 | ||
1e8f889b DG |
4654 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
4655 | int writable) | |
63551ae0 DG |
4656 | { |
4657 | pte_t entry; | |
79c1c594 | 4658 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
63551ae0 | 4659 | |
1e8f889b | 4660 | if (writable) { |
106c992a GS |
4661 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
4662 | vma->vm_page_prot))); | |
63551ae0 | 4663 | } else { |
106c992a GS |
4664 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
4665 | vma->vm_page_prot)); | |
63551ae0 DG |
4666 | } |
4667 | entry = pte_mkyoung(entry); | |
79c1c594 | 4668 | entry = arch_make_huge_pte(entry, shift, vma->vm_flags); |
63551ae0 DG |
4669 | |
4670 | return entry; | |
4671 | } | |
4672 | ||
1e8f889b DG |
4673 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
4674 | unsigned long address, pte_t *ptep) | |
4675 | { | |
4676 | pte_t entry; | |
4677 | ||
106c992a | 4678 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 4679 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 4680 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
4681 | } |
4682 | ||
d5ed7444 | 4683 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
4684 | { |
4685 | swp_entry_t swp; | |
4686 | ||
4687 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 4688 | return false; |
4a705fef | 4689 | swp = pte_to_swp_entry(pte); |
d79d176a | 4690 | if (is_migration_entry(swp)) |
d5ed7444 | 4691 | return true; |
4a705fef | 4692 | else |
d5ed7444 | 4693 | return false; |
4a705fef NH |
4694 | } |
4695 | ||
3e5c3600 | 4696 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
4697 | { |
4698 | swp_entry_t swp; | |
4699 | ||
4700 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 4701 | return false; |
4a705fef | 4702 | swp = pte_to_swp_entry(pte); |
d79d176a | 4703 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 4704 | return true; |
4a705fef | 4705 | else |
3e5c3600 | 4706 | return false; |
4a705fef | 4707 | } |
1e8f889b | 4708 | |
4eae4efa PX |
4709 | static void |
4710 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
4711 | struct page *new_page) | |
4712 | { | |
4713 | __SetPageUptodate(new_page); | |
4eae4efa | 4714 | hugepage_add_new_anon_rmap(new_page, vma, addr); |
1eba86c0 | 4715 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); |
4eae4efa PX |
4716 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); |
4717 | ClearHPageRestoreReserve(new_page); | |
4718 | SetHPageMigratable(new_page); | |
4719 | } | |
4720 | ||
63551ae0 | 4721 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
bc70fbf2 PX |
4722 | struct vm_area_struct *dst_vma, |
4723 | struct vm_area_struct *src_vma) | |
63551ae0 | 4724 | { |
5e41540c | 4725 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 4726 | struct page *ptepage; |
1c59827d | 4727 | unsigned long addr; |
bc70fbf2 PX |
4728 | bool cow = is_cow_mapping(src_vma->vm_flags); |
4729 | struct hstate *h = hstate_vma(src_vma); | |
a5516438 | 4730 | unsigned long sz = huge_page_size(h); |
4eae4efa | 4731 | unsigned long npages = pages_per_huge_page(h); |
bc70fbf2 | 4732 | struct address_space *mapping = src_vma->vm_file->f_mapping; |
ac46d4f3 | 4733 | struct mmu_notifier_range range; |
e8569dd2 | 4734 | int ret = 0; |
1e8f889b | 4735 | |
ac46d4f3 | 4736 | if (cow) { |
bc70fbf2 PX |
4737 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, src_vma, src, |
4738 | src_vma->vm_start, | |
4739 | src_vma->vm_end); | |
ac46d4f3 | 4740 | mmu_notifier_invalidate_range_start(&range); |
623a1ddf DH |
4741 | mmap_assert_write_locked(src); |
4742 | raw_write_seqcount_begin(&src->write_protect_seq); | |
c0d0381a MK |
4743 | } else { |
4744 | /* | |
4745 | * For shared mappings i_mmap_rwsem must be held to call | |
4746 | * huge_pte_alloc, otherwise the returned ptep could go | |
4747 | * away if part of a shared pmd and another thread calls | |
4748 | * huge_pmd_unshare. | |
4749 | */ | |
4750 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 4751 | } |
e8569dd2 | 4752 | |
bc70fbf2 | 4753 | for (addr = src_vma->vm_start; addr < src_vma->vm_end; addr += sz) { |
cb900f41 | 4754 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 4755 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
4756 | if (!src_pte) |
4757 | continue; | |
bc70fbf2 | 4758 | dst_pte = huge_pte_alloc(dst, dst_vma, addr, sz); |
e8569dd2 AS |
4759 | if (!dst_pte) { |
4760 | ret = -ENOMEM; | |
4761 | break; | |
4762 | } | |
c5c99429 | 4763 | |
5e41540c MK |
4764 | /* |
4765 | * If the pagetables are shared don't copy or take references. | |
4766 | * dst_pte == src_pte is the common case of src/dest sharing. | |
4767 | * | |
4768 | * However, src could have 'unshared' and dst shares with | |
4769 | * another vma. If dst_pte !none, this implies sharing. | |
4770 | * Check here before taking page table lock, and once again | |
4771 | * after taking the lock below. | |
4772 | */ | |
4773 | dst_entry = huge_ptep_get(dst_pte); | |
4774 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
4775 | continue; |
4776 | ||
cb900f41 KS |
4777 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
4778 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4779 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 4780 | entry = huge_ptep_get(src_pte); |
5e41540c | 4781 | dst_entry = huge_ptep_get(dst_pte); |
4eae4efa | 4782 | again: |
5e41540c MK |
4783 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { |
4784 | /* | |
4785 | * Skip if src entry none. Also, skip in the | |
4786 | * unlikely case dst entry !none as this implies | |
4787 | * sharing with another vma. | |
4788 | */ | |
4a705fef NH |
4789 | ; |
4790 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
4791 | is_hugetlb_entry_hwpoisoned(entry))) { | |
4792 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
bc70fbf2 | 4793 | bool uffd_wp = huge_pte_uffd_wp(entry); |
4a705fef | 4794 | |
6c287605 | 4795 | if (!is_readable_migration_entry(swp_entry) && cow) { |
4a705fef NH |
4796 | /* |
4797 | * COW mappings require pages in both | |
4798 | * parent and child to be set to read. | |
4799 | */ | |
4dd845b5 AP |
4800 | swp_entry = make_readable_migration_entry( |
4801 | swp_offset(swp_entry)); | |
4a705fef | 4802 | entry = swp_entry_to_pte(swp_entry); |
bc70fbf2 PX |
4803 | if (userfaultfd_wp(src_vma) && uffd_wp) |
4804 | entry = huge_pte_mkuffd_wp(entry); | |
e5251fd4 PA |
4805 | set_huge_swap_pte_at(src, addr, src_pte, |
4806 | entry, sz); | |
4a705fef | 4807 | } |
bc70fbf2 PX |
4808 | if (!userfaultfd_wp(dst_vma) && uffd_wp) |
4809 | entry = huge_pte_clear_uffd_wp(entry); | |
e5251fd4 | 4810 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
bc70fbf2 PX |
4811 | } else if (unlikely(is_pte_marker(entry))) { |
4812 | /* | |
4813 | * We copy the pte marker only if the dst vma has | |
4814 | * uffd-wp enabled. | |
4815 | */ | |
4816 | if (userfaultfd_wp(dst_vma)) | |
4817 | set_huge_pte_at(dst, addr, dst_pte, entry); | |
4a705fef | 4818 | } else { |
4eae4efa PX |
4819 | entry = huge_ptep_get(src_pte); |
4820 | ptepage = pte_page(entry); | |
4821 | get_page(ptepage); | |
4822 | ||
4823 | /* | |
fb3d824d DH |
4824 | * Failing to duplicate the anon rmap is a rare case |
4825 | * where we see pinned hugetlb pages while they're | |
4826 | * prone to COW. We need to do the COW earlier during | |
4827 | * fork. | |
4eae4efa PX |
4828 | * |
4829 | * When pre-allocating the page or copying data, we | |
4830 | * need to be without the pgtable locks since we could | |
4831 | * sleep during the process. | |
4832 | */ | |
fb3d824d DH |
4833 | if (!PageAnon(ptepage)) { |
4834 | page_dup_file_rmap(ptepage, true); | |
bc70fbf2 PX |
4835 | } else if (page_try_dup_anon_rmap(ptepage, true, |
4836 | src_vma)) { | |
4eae4efa PX |
4837 | pte_t src_pte_old = entry; |
4838 | struct page *new; | |
4839 | ||
4840 | spin_unlock(src_ptl); | |
4841 | spin_unlock(dst_ptl); | |
4842 | /* Do not use reserve as it's private owned */ | |
bc70fbf2 | 4843 | new = alloc_huge_page(dst_vma, addr, 1); |
4eae4efa PX |
4844 | if (IS_ERR(new)) { |
4845 | put_page(ptepage); | |
4846 | ret = PTR_ERR(new); | |
4847 | break; | |
4848 | } | |
bc70fbf2 | 4849 | copy_user_huge_page(new, ptepage, addr, dst_vma, |
4eae4efa PX |
4850 | npages); |
4851 | put_page(ptepage); | |
4852 | ||
4853 | /* Install the new huge page if src pte stable */ | |
4854 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
4855 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4856 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4857 | entry = huge_ptep_get(src_pte); | |
4858 | if (!pte_same(src_pte_old, entry)) { | |
bc70fbf2 | 4859 | restore_reserve_on_error(h, dst_vma, addr, |
846be085 | 4860 | new); |
4eae4efa PX |
4861 | put_page(new); |
4862 | /* dst_entry won't change as in child */ | |
4863 | goto again; | |
4864 | } | |
bc70fbf2 | 4865 | hugetlb_install_page(dst_vma, dst_pte, addr, new); |
4eae4efa PX |
4866 | spin_unlock(src_ptl); |
4867 | spin_unlock(dst_ptl); | |
4868 | continue; | |
4869 | } | |
4870 | ||
34ee645e | 4871 | if (cow) { |
0f10851e JG |
4872 | /* |
4873 | * No need to notify as we are downgrading page | |
4874 | * table protection not changing it to point | |
4875 | * to a new page. | |
4876 | * | |
ad56b738 | 4877 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4878 | */ |
7f2e9525 | 4879 | huge_ptep_set_wrprotect(src, addr, src_pte); |
84894e1c | 4880 | entry = huge_pte_wrprotect(entry); |
34ee645e | 4881 | } |
4eae4efa | 4882 | |
1c59827d | 4883 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 4884 | hugetlb_count_add(npages, dst); |
1c59827d | 4885 | } |
cb900f41 KS |
4886 | spin_unlock(src_ptl); |
4887 | spin_unlock(dst_ptl); | |
63551ae0 | 4888 | } |
63551ae0 | 4889 | |
623a1ddf DH |
4890 | if (cow) { |
4891 | raw_write_seqcount_end(&src->write_protect_seq); | |
ac46d4f3 | 4892 | mmu_notifier_invalidate_range_end(&range); |
623a1ddf | 4893 | } else { |
c0d0381a | 4894 | i_mmap_unlock_read(mapping); |
623a1ddf | 4895 | } |
e8569dd2 AS |
4896 | |
4897 | return ret; | |
63551ae0 DG |
4898 | } |
4899 | ||
550a7d60 | 4900 | static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, |
db110a99 | 4901 | unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte) |
550a7d60 MA |
4902 | { |
4903 | struct hstate *h = hstate_vma(vma); | |
4904 | struct mm_struct *mm = vma->vm_mm; | |
550a7d60 | 4905 | spinlock_t *src_ptl, *dst_ptl; |
db110a99 | 4906 | pte_t pte; |
550a7d60 | 4907 | |
550a7d60 MA |
4908 | dst_ptl = huge_pte_lock(h, mm, dst_pte); |
4909 | src_ptl = huge_pte_lockptr(h, mm, src_pte); | |
4910 | ||
4911 | /* | |
4912 | * We don't have to worry about the ordering of src and dst ptlocks | |
4913 | * because exclusive mmap_sem (or the i_mmap_lock) prevents deadlock. | |
4914 | */ | |
4915 | if (src_ptl != dst_ptl) | |
4916 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4917 | ||
4918 | pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); | |
4919 | set_huge_pte_at(mm, new_addr, dst_pte, pte); | |
4920 | ||
4921 | if (src_ptl != dst_ptl) | |
4922 | spin_unlock(src_ptl); | |
4923 | spin_unlock(dst_ptl); | |
4924 | } | |
4925 | ||
4926 | int move_hugetlb_page_tables(struct vm_area_struct *vma, | |
4927 | struct vm_area_struct *new_vma, | |
4928 | unsigned long old_addr, unsigned long new_addr, | |
4929 | unsigned long len) | |
4930 | { | |
4931 | struct hstate *h = hstate_vma(vma); | |
4932 | struct address_space *mapping = vma->vm_file->f_mapping; | |
4933 | unsigned long sz = huge_page_size(h); | |
4934 | struct mm_struct *mm = vma->vm_mm; | |
4935 | unsigned long old_end = old_addr + len; | |
4936 | unsigned long old_addr_copy; | |
4937 | pte_t *src_pte, *dst_pte; | |
4938 | struct mmu_notifier_range range; | |
3d0b95cd | 4939 | bool shared_pmd = false; |
550a7d60 MA |
4940 | |
4941 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, old_addr, | |
4942 | old_end); | |
4943 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); | |
3d0b95cd BW |
4944 | /* |
4945 | * In case of shared PMDs, we should cover the maximum possible | |
4946 | * range. | |
4947 | */ | |
4948 | flush_cache_range(vma, range.start, range.end); | |
4949 | ||
550a7d60 MA |
4950 | mmu_notifier_invalidate_range_start(&range); |
4951 | /* Prevent race with file truncation */ | |
4952 | i_mmap_lock_write(mapping); | |
4953 | for (; old_addr < old_end; old_addr += sz, new_addr += sz) { | |
4954 | src_pte = huge_pte_offset(mm, old_addr, sz); | |
4955 | if (!src_pte) | |
4956 | continue; | |
4957 | if (huge_pte_none(huge_ptep_get(src_pte))) | |
4958 | continue; | |
4959 | ||
4960 | /* old_addr arg to huge_pmd_unshare() is a pointer and so the | |
4961 | * arg may be modified. Pass a copy instead to preserve the | |
4962 | * value in old_addr. | |
4963 | */ | |
4964 | old_addr_copy = old_addr; | |
4965 | ||
3d0b95cd BW |
4966 | if (huge_pmd_unshare(mm, vma, &old_addr_copy, src_pte)) { |
4967 | shared_pmd = true; | |
550a7d60 | 4968 | continue; |
3d0b95cd | 4969 | } |
550a7d60 MA |
4970 | |
4971 | dst_pte = huge_pte_alloc(mm, new_vma, new_addr, sz); | |
4972 | if (!dst_pte) | |
4973 | break; | |
4974 | ||
db110a99 | 4975 | move_huge_pte(vma, old_addr, new_addr, src_pte, dst_pte); |
550a7d60 | 4976 | } |
3d0b95cd BW |
4977 | |
4978 | if (shared_pmd) | |
4979 | flush_tlb_range(vma, range.start, range.end); | |
4980 | else | |
4981 | flush_tlb_range(vma, old_end - len, old_end); | |
550a7d60 | 4982 | mmu_notifier_invalidate_range_end(&range); |
13e4ad2c | 4983 | i_mmap_unlock_write(mapping); |
550a7d60 MA |
4984 | |
4985 | return len + old_addr - old_end; | |
4986 | } | |
4987 | ||
73c54763 PX |
4988 | static void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
4989 | unsigned long start, unsigned long end, | |
05e90bd0 | 4990 | struct page *ref_page, zap_flags_t zap_flags) |
63551ae0 DG |
4991 | { |
4992 | struct mm_struct *mm = vma->vm_mm; | |
4993 | unsigned long address; | |
c7546f8f | 4994 | pte_t *ptep; |
63551ae0 | 4995 | pte_t pte; |
cb900f41 | 4996 | spinlock_t *ptl; |
63551ae0 | 4997 | struct page *page; |
a5516438 AK |
4998 | struct hstate *h = hstate_vma(vma); |
4999 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 5000 | struct mmu_notifier_range range; |
a4a118f2 | 5001 | bool force_flush = false; |
a5516438 | 5002 | |
63551ae0 | 5003 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
5004 | BUG_ON(start & ~huge_page_mask(h)); |
5005 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 5006 | |
07e32661 AK |
5007 | /* |
5008 | * This is a hugetlb vma, all the pte entries should point | |
5009 | * to huge page. | |
5010 | */ | |
ed6a7935 | 5011 | tlb_change_page_size(tlb, sz); |
24669e58 | 5012 | tlb_start_vma(tlb, vma); |
dff11abe MK |
5013 | |
5014 | /* | |
5015 | * If sharing possible, alert mmu notifiers of worst case. | |
5016 | */ | |
6f4f13e8 JG |
5017 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
5018 | end); | |
ac46d4f3 JG |
5019 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
5020 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 5021 | address = start; |
569f48b8 | 5022 | for (; address < end; address += sz) { |
7868a208 | 5023 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 5024 | if (!ptep) |
c7546f8f DG |
5025 | continue; |
5026 | ||
cb900f41 | 5027 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 5028 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
31d49da5 | 5029 | spin_unlock(ptl); |
a4a118f2 NA |
5030 | tlb_flush_pmd_range(tlb, address & PUD_MASK, PUD_SIZE); |
5031 | force_flush = true; | |
31d49da5 AK |
5032 | continue; |
5033 | } | |
39dde65c | 5034 | |
6629326b | 5035 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
5036 | if (huge_pte_none(pte)) { |
5037 | spin_unlock(ptl); | |
5038 | continue; | |
5039 | } | |
6629326b HD |
5040 | |
5041 | /* | |
9fbc1f63 NH |
5042 | * Migrating hugepage or HWPoisoned hugepage is already |
5043 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 5044 | */ |
9fbc1f63 | 5045 | if (unlikely(!pte_present(pte))) { |
05e90bd0 PX |
5046 | /* |
5047 | * If the pte was wr-protected by uffd-wp in any of the | |
5048 | * swap forms, meanwhile the caller does not want to | |
5049 | * drop the uffd-wp bit in this zap, then replace the | |
5050 | * pte with a marker. | |
5051 | */ | |
5052 | if (pte_swp_uffd_wp_any(pte) && | |
5053 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5054 | set_huge_pte_at(mm, address, ptep, | |
5055 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5056 | else | |
5057 | huge_pte_clear(mm, address, ptep, sz); | |
31d49da5 AK |
5058 | spin_unlock(ptl); |
5059 | continue; | |
8c4894c6 | 5060 | } |
6629326b HD |
5061 | |
5062 | page = pte_page(pte); | |
04f2cbe3 MG |
5063 | /* |
5064 | * If a reference page is supplied, it is because a specific | |
5065 | * page is being unmapped, not a range. Ensure the page we | |
5066 | * are about to unmap is the actual page of interest. | |
5067 | */ | |
5068 | if (ref_page) { | |
31d49da5 AK |
5069 | if (page != ref_page) { |
5070 | spin_unlock(ptl); | |
5071 | continue; | |
5072 | } | |
04f2cbe3 MG |
5073 | /* |
5074 | * Mark the VMA as having unmapped its page so that | |
5075 | * future faults in this VMA will fail rather than | |
5076 | * looking like data was lost | |
5077 | */ | |
5078 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
5079 | } | |
5080 | ||
c7546f8f | 5081 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 5082 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 5083 | if (huge_pte_dirty(pte)) |
6649a386 | 5084 | set_page_dirty(page); |
05e90bd0 PX |
5085 | /* Leave a uffd-wp pte marker if needed */ |
5086 | if (huge_pte_uffd_wp(pte) && | |
5087 | !(zap_flags & ZAP_FLAG_DROP_MARKER)) | |
5088 | set_huge_pte_at(mm, address, ptep, | |
5089 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
5d317b2b | 5090 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
cea86fe2 | 5091 | page_remove_rmap(page, vma, true); |
31d49da5 | 5092 | |
cb900f41 | 5093 | spin_unlock(ptl); |
e77b0852 | 5094 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
5095 | /* |
5096 | * Bail out after unmapping reference page if supplied | |
5097 | */ | |
5098 | if (ref_page) | |
5099 | break; | |
fe1668ae | 5100 | } |
ac46d4f3 | 5101 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 5102 | tlb_end_vma(tlb, vma); |
a4a118f2 NA |
5103 | |
5104 | /* | |
5105 | * If we unshared PMDs, the TLB flush was not recorded in mmu_gather. We | |
5106 | * could defer the flush until now, since by holding i_mmap_rwsem we | |
5107 | * guaranteed that the last refernece would not be dropped. But we must | |
5108 | * do the flushing before we return, as otherwise i_mmap_rwsem will be | |
5109 | * dropped and the last reference to the shared PMDs page might be | |
5110 | * dropped as well. | |
5111 | * | |
5112 | * In theory we could defer the freeing of the PMD pages as well, but | |
5113 | * huge_pmd_unshare() relies on the exact page_count for the PMD page to | |
5114 | * detect sharing, so we cannot defer the release of the page either. | |
5115 | * Instead, do flush now. | |
5116 | */ | |
5117 | if (force_flush) | |
5118 | tlb_flush_mmu_tlbonly(tlb); | |
1da177e4 | 5119 | } |
63551ae0 | 5120 | |
d833352a MG |
5121 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
5122 | struct vm_area_struct *vma, unsigned long start, | |
05e90bd0 PX |
5123 | unsigned long end, struct page *ref_page, |
5124 | zap_flags_t zap_flags) | |
d833352a | 5125 | { |
05e90bd0 | 5126 | __unmap_hugepage_range(tlb, vma, start, end, ref_page, zap_flags); |
d833352a MG |
5127 | |
5128 | /* | |
5129 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
5130 | * test will fail on a vma being torn down, and not grab a page table | |
5131 | * on its way out. We're lucky that the flag has such an appropriate | |
5132 | * name, and can in fact be safely cleared here. We could clear it | |
5133 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 5134 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 5135 | * because in the context this is called, the VMA is about to be |
c8c06efa | 5136 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
5137 | */ |
5138 | vma->vm_flags &= ~VM_MAYSHARE; | |
5139 | } | |
5140 | ||
502717f4 | 5141 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
05e90bd0 PX |
5142 | unsigned long end, struct page *ref_page, |
5143 | zap_flags_t zap_flags) | |
502717f4 | 5144 | { |
24669e58 | 5145 | struct mmu_gather tlb; |
dff11abe | 5146 | |
a72afd87 | 5147 | tlb_gather_mmu(&tlb, vma->vm_mm); |
05e90bd0 | 5148 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page, zap_flags); |
ae8eba8b | 5149 | tlb_finish_mmu(&tlb); |
502717f4 CK |
5150 | } |
5151 | ||
04f2cbe3 MG |
5152 | /* |
5153 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 5154 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
5155 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
5156 | * same region. | |
5157 | */ | |
2f4612af DB |
5158 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
5159 | struct page *page, unsigned long address) | |
04f2cbe3 | 5160 | { |
7526674d | 5161 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
5162 | struct vm_area_struct *iter_vma; |
5163 | struct address_space *mapping; | |
04f2cbe3 MG |
5164 | pgoff_t pgoff; |
5165 | ||
5166 | /* | |
5167 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
5168 | * from page cache lookup which is in HPAGE_SIZE units. | |
5169 | */ | |
7526674d | 5170 | address = address & huge_page_mask(h); |
36e4f20a MH |
5171 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
5172 | vma->vm_pgoff; | |
93c76a3d | 5173 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 5174 | |
4eb2b1dc MG |
5175 | /* |
5176 | * Take the mapping lock for the duration of the table walk. As | |
5177 | * this mapping should be shared between all the VMAs, | |
5178 | * __unmap_hugepage_range() is called as the lock is already held | |
5179 | */ | |
83cde9e8 | 5180 | i_mmap_lock_write(mapping); |
6b2dbba8 | 5181 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
5182 | /* Do not unmap the current VMA */ |
5183 | if (iter_vma == vma) | |
5184 | continue; | |
5185 | ||
2f84a899 MG |
5186 | /* |
5187 | * Shared VMAs have their own reserves and do not affect | |
5188 | * MAP_PRIVATE accounting but it is possible that a shared | |
5189 | * VMA is using the same page so check and skip such VMAs. | |
5190 | */ | |
5191 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
5192 | continue; | |
5193 | ||
04f2cbe3 MG |
5194 | /* |
5195 | * Unmap the page from other VMAs without their own reserves. | |
5196 | * They get marked to be SIGKILLed if they fault in these | |
5197 | * areas. This is because a future no-page fault on this VMA | |
5198 | * could insert a zeroed page instead of the data existing | |
5199 | * from the time of fork. This would look like data corruption | |
5200 | */ | |
5201 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 | 5202 | unmap_hugepage_range(iter_vma, address, |
05e90bd0 | 5203 | address + huge_page_size(h), page, 0); |
04f2cbe3 | 5204 | } |
83cde9e8 | 5205 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
5206 | } |
5207 | ||
0fe6e20b | 5208 | /* |
c89357e2 | 5209 | * hugetlb_wp() should be called with page lock of the original hugepage held. |
aa6d2e8c | 5210 | * Called with hugetlb_fault_mutex_table held and pte_page locked so we |
ef009b25 MH |
5211 | * cannot race with other handlers or page migration. |
5212 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 5213 | */ |
c89357e2 DH |
5214 | static vm_fault_t hugetlb_wp(struct mm_struct *mm, struct vm_area_struct *vma, |
5215 | unsigned long address, pte_t *ptep, unsigned int flags, | |
3999f52e | 5216 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 5217 | { |
c89357e2 | 5218 | const bool unshare = flags & FAULT_FLAG_UNSHARE; |
3999f52e | 5219 | pte_t pte; |
a5516438 | 5220 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 5221 | struct page *old_page, *new_page; |
2b740303 SJ |
5222 | int outside_reserve = 0; |
5223 | vm_fault_t ret = 0; | |
974e6d66 | 5224 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 5225 | struct mmu_notifier_range range; |
1e8f889b | 5226 | |
c89357e2 DH |
5227 | VM_BUG_ON(unshare && (flags & FOLL_WRITE)); |
5228 | VM_BUG_ON(!unshare && !(flags & FOLL_WRITE)); | |
5229 | ||
3999f52e | 5230 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
5231 | old_page = pte_page(pte); |
5232 | ||
04f2cbe3 | 5233 | retry_avoidcopy: |
c89357e2 DH |
5234 | /* |
5235 | * If no-one else is actually using this page, we're the exclusive | |
5236 | * owner and can reuse this page. | |
5237 | */ | |
37a2140d | 5238 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
c89357e2 DH |
5239 | if (!PageAnonExclusive(old_page)) |
5240 | page_move_anon_rmap(old_page, vma); | |
5241 | if (likely(!unshare)) | |
5242 | set_huge_ptep_writable(vma, haddr, ptep); | |
83c54070 | 5243 | return 0; |
1e8f889b | 5244 | } |
6c287605 DH |
5245 | VM_BUG_ON_PAGE(PageAnon(old_page) && PageAnonExclusive(old_page), |
5246 | old_page); | |
1e8f889b | 5247 | |
04f2cbe3 MG |
5248 | /* |
5249 | * If the process that created a MAP_PRIVATE mapping is about to | |
5250 | * perform a COW due to a shared page count, attempt to satisfy | |
5251 | * the allocation without using the existing reserves. The pagecache | |
5252 | * page is used to determine if the reserve at this address was | |
5253 | * consumed or not. If reserves were used, a partial faulted mapping | |
5254 | * at the time of fork() could consume its reserves on COW instead | |
5255 | * of the full address range. | |
5256 | */ | |
5944d011 | 5257 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
5258 | old_page != pagecache_page) |
5259 | outside_reserve = 1; | |
5260 | ||
09cbfeaf | 5261 | get_page(old_page); |
b76c8cfb | 5262 | |
ad4404a2 DB |
5263 | /* |
5264 | * Drop page table lock as buddy allocator may be called. It will | |
5265 | * be acquired again before returning to the caller, as expected. | |
5266 | */ | |
cb900f41 | 5267 | spin_unlock(ptl); |
5b7a1d40 | 5268 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 5269 | |
2fc39cec | 5270 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
5271 | /* |
5272 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
5273 | * it is due to references held by a child and an insufficient | |
5274 | * huge page pool. To guarantee the original mappers | |
5275 | * reliability, unmap the page from child processes. The child | |
5276 | * may get SIGKILLed if it later faults. | |
5277 | */ | |
5278 | if (outside_reserve) { | |
e7dd91c4 MK |
5279 | struct address_space *mapping = vma->vm_file->f_mapping; |
5280 | pgoff_t idx; | |
5281 | u32 hash; | |
5282 | ||
09cbfeaf | 5283 | put_page(old_page); |
04f2cbe3 | 5284 | BUG_ON(huge_pte_none(pte)); |
e7dd91c4 MK |
5285 | /* |
5286 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
5287 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
5288 | * in write mode. Dropping i_mmap_rwsem in read mode | |
5289 | * here is OK as COW mappings do not interact with | |
5290 | * PMD sharing. | |
5291 | * | |
5292 | * Reacquire both after unmap operation. | |
5293 | */ | |
5294 | idx = vma_hugecache_offset(h, vma, haddr); | |
5295 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5296 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5297 | i_mmap_unlock_read(mapping); | |
5298 | ||
5b7a1d40 | 5299 | unmap_ref_private(mm, vma, old_page, haddr); |
e7dd91c4 MK |
5300 | |
5301 | i_mmap_lock_read(mapping); | |
5302 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
2f4612af | 5303 | spin_lock(ptl); |
5b7a1d40 | 5304 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
5305 | if (likely(ptep && |
5306 | pte_same(huge_ptep_get(ptep), pte))) | |
5307 | goto retry_avoidcopy; | |
5308 | /* | |
5309 | * race occurs while re-acquiring page table | |
5310 | * lock, and our job is done. | |
5311 | */ | |
5312 | return 0; | |
04f2cbe3 MG |
5313 | } |
5314 | ||
2b740303 | 5315 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 5316 | goto out_release_old; |
1e8f889b DG |
5317 | } |
5318 | ||
0fe6e20b NH |
5319 | /* |
5320 | * When the original hugepage is shared one, it does not have | |
5321 | * anon_vma prepared. | |
5322 | */ | |
44e2aa93 | 5323 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
5324 | ret = VM_FAULT_OOM; |
5325 | goto out_release_all; | |
44e2aa93 | 5326 | } |
0fe6e20b | 5327 | |
974e6d66 | 5328 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 5329 | pages_per_huge_page(h)); |
0ed361de | 5330 | __SetPageUptodate(new_page); |
1e8f889b | 5331 | |
7269f999 | 5332 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 5333 | haddr + huge_page_size(h)); |
ac46d4f3 | 5334 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 5335 | |
b76c8cfb | 5336 | /* |
cb900f41 | 5337 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
5338 | * before the page tables are altered |
5339 | */ | |
cb900f41 | 5340 | spin_lock(ptl); |
5b7a1d40 | 5341 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 5342 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 5343 | ClearHPageRestoreReserve(new_page); |
07443a85 | 5344 | |
c89357e2 | 5345 | /* Break COW or unshare */ |
5b7a1d40 | 5346 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 5347 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
cea86fe2 | 5348 | page_remove_rmap(old_page, vma, true); |
5b7a1d40 | 5349 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
1eba86c0 | 5350 | set_huge_pte_at(mm, haddr, ptep, |
c89357e2 | 5351 | make_huge_pte(vma, new_page, !unshare)); |
8f251a3d | 5352 | SetHPageMigratable(new_page); |
1e8f889b DG |
5353 | /* Make the old page be freed below */ |
5354 | new_page = old_page; | |
5355 | } | |
cb900f41 | 5356 | spin_unlock(ptl); |
ac46d4f3 | 5357 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 5358 | out_release_all: |
c89357e2 DH |
5359 | /* |
5360 | * No restore in case of successful pagetable update (Break COW or | |
5361 | * unshare) | |
5362 | */ | |
c7b1850d MK |
5363 | if (new_page != old_page) |
5364 | restore_reserve_on_error(h, vma, haddr, new_page); | |
09cbfeaf | 5365 | put_page(new_page); |
ad4404a2 | 5366 | out_release_old: |
09cbfeaf | 5367 | put_page(old_page); |
8312034f | 5368 | |
ad4404a2 DB |
5369 | spin_lock(ptl); /* Caller expects lock to be held */ |
5370 | return ret; | |
1e8f889b DG |
5371 | } |
5372 | ||
04f2cbe3 | 5373 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
5374 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
5375 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
5376 | { |
5377 | struct address_space *mapping; | |
e7c4b0bf | 5378 | pgoff_t idx; |
04f2cbe3 MG |
5379 | |
5380 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 5381 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
5382 | |
5383 | return find_lock_page(mapping, idx); | |
5384 | } | |
5385 | ||
3ae77f43 HD |
5386 | /* |
5387 | * Return whether there is a pagecache page to back given address within VMA. | |
5388 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
5389 | */ | |
5390 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
5391 | struct vm_area_struct *vma, unsigned long address) |
5392 | { | |
5393 | struct address_space *mapping; | |
5394 | pgoff_t idx; | |
5395 | struct page *page; | |
5396 | ||
5397 | mapping = vma->vm_file->f_mapping; | |
5398 | idx = vma_hugecache_offset(h, vma, address); | |
5399 | ||
5400 | page = find_get_page(mapping, idx); | |
5401 | if (page) | |
5402 | put_page(page); | |
5403 | return page != NULL; | |
5404 | } | |
5405 | ||
ab76ad54 MK |
5406 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
5407 | pgoff_t idx) | |
5408 | { | |
5409 | struct inode *inode = mapping->host; | |
5410 | struct hstate *h = hstate_inode(inode); | |
5411 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
5412 | ||
5413 | if (err) | |
5414 | return err; | |
d6995da3 | 5415 | ClearHPageRestoreReserve(page); |
ab76ad54 | 5416 | |
22146c3c MK |
5417 | /* |
5418 | * set page dirty so that it will not be removed from cache/file | |
5419 | * by non-hugetlbfs specific code paths. | |
5420 | */ | |
5421 | set_page_dirty(page); | |
5422 | ||
ab76ad54 MK |
5423 | spin_lock(&inode->i_lock); |
5424 | inode->i_blocks += blocks_per_huge_page(h); | |
5425 | spin_unlock(&inode->i_lock); | |
5426 | return 0; | |
5427 | } | |
5428 | ||
7677f7fd AR |
5429 | static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, |
5430 | struct address_space *mapping, | |
5431 | pgoff_t idx, | |
5432 | unsigned int flags, | |
5433 | unsigned long haddr, | |
824ddc60 | 5434 | unsigned long addr, |
7677f7fd AR |
5435 | unsigned long reason) |
5436 | { | |
5437 | vm_fault_t ret; | |
5438 | u32 hash; | |
5439 | struct vm_fault vmf = { | |
5440 | .vma = vma, | |
5441 | .address = haddr, | |
824ddc60 | 5442 | .real_address = addr, |
7677f7fd AR |
5443 | .flags = flags, |
5444 | ||
5445 | /* | |
5446 | * Hard to debug if it ends up being | |
5447 | * used by a callee that assumes | |
5448 | * something about the other | |
5449 | * uninitialized fields... same as in | |
5450 | * memory.c | |
5451 | */ | |
5452 | }; | |
5453 | ||
5454 | /* | |
5455 | * hugetlb_fault_mutex and i_mmap_rwsem must be | |
5456 | * dropped before handling userfault. Reacquire | |
5457 | * after handling fault to make calling code simpler. | |
5458 | */ | |
5459 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
5460 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5461 | i_mmap_unlock_read(mapping); | |
5462 | ret = handle_userfault(&vmf, reason); | |
5463 | i_mmap_lock_read(mapping); | |
5464 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
5465 | ||
5466 | return ret; | |
5467 | } | |
5468 | ||
2b740303 SJ |
5469 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
5470 | struct vm_area_struct *vma, | |
5471 | struct address_space *mapping, pgoff_t idx, | |
c64e912c PX |
5472 | unsigned long address, pte_t *ptep, |
5473 | pte_t old_pte, unsigned int flags) | |
ac9b9c66 | 5474 | { |
a5516438 | 5475 | struct hstate *h = hstate_vma(vma); |
2b740303 | 5476 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 5477 | int anon_rmap = 0; |
4c887265 | 5478 | unsigned long size; |
4c887265 | 5479 | struct page *page; |
1e8f889b | 5480 | pte_t new_pte; |
cb900f41 | 5481 | spinlock_t *ptl; |
285b8dca | 5482 | unsigned long haddr = address & huge_page_mask(h); |
c7b1850d | 5483 | bool new_page, new_pagecache_page = false; |
4c887265 | 5484 | |
04f2cbe3 MG |
5485 | /* |
5486 | * Currently, we are forced to kill the process in the event the | |
5487 | * original mapper has unmapped pages from the child due to a failed | |
c89357e2 DH |
5488 | * COW/unsharing. Warn that such a situation has occurred as it may not |
5489 | * be obvious. | |
04f2cbe3 MG |
5490 | */ |
5491 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 5492 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 5493 | current->pid); |
04f2cbe3 MG |
5494 | return ret; |
5495 | } | |
5496 | ||
4c887265 | 5497 | /* |
87bf91d3 MK |
5498 | * We can not race with truncation due to holding i_mmap_rwsem. |
5499 | * i_size is modified when holding i_mmap_rwsem, so check here | |
5500 | * once for faults beyond end of file. | |
4c887265 | 5501 | */ |
87bf91d3 MK |
5502 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5503 | if (idx >= size) | |
5504 | goto out; | |
5505 | ||
6bda666a | 5506 | retry: |
c7b1850d | 5507 | new_page = false; |
6bda666a CL |
5508 | page = find_lock_page(mapping, idx); |
5509 | if (!page) { | |
7677f7fd | 5510 | /* Check for page in userfault range */ |
1a1aad8a | 5511 | if (userfaultfd_missing(vma)) { |
7677f7fd | 5512 | ret = hugetlb_handle_userfault(vma, mapping, idx, |
824ddc60 | 5513 | flags, haddr, address, |
7677f7fd | 5514 | VM_UFFD_MISSING); |
1a1aad8a MK |
5515 | goto out; |
5516 | } | |
5517 | ||
285b8dca | 5518 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 5519 | if (IS_ERR(page)) { |
4643d67e MK |
5520 | /* |
5521 | * Returning error will result in faulting task being | |
5522 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
5523 | * tasks from racing to fault in the same page which | |
5524 | * could result in false unable to allocate errors. | |
5525 | * Page migration does not take the fault mutex, but | |
5526 | * does a clear then write of pte's under page table | |
5527 | * lock. Page fault code could race with migration, | |
5528 | * notice the clear pte and try to allocate a page | |
5529 | * here. Before returning error, get ptl and make | |
5530 | * sure there really is no pte entry. | |
5531 | */ | |
5532 | ptl = huge_pte_lock(h, mm, ptep); | |
d83e6c8a ML |
5533 | ret = 0; |
5534 | if (huge_pte_none(huge_ptep_get(ptep))) | |
5535 | ret = vmf_error(PTR_ERR(page)); | |
4643d67e | 5536 | spin_unlock(ptl); |
6bda666a CL |
5537 | goto out; |
5538 | } | |
47ad8475 | 5539 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 5540 | __SetPageUptodate(page); |
cb6acd01 | 5541 | new_page = true; |
ac9b9c66 | 5542 | |
f83a275d | 5543 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 5544 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
5545 | if (err) { |
5546 | put_page(page); | |
6bda666a CL |
5547 | if (err == -EEXIST) |
5548 | goto retry; | |
5549 | goto out; | |
5550 | } | |
c7b1850d | 5551 | new_pagecache_page = true; |
23be7468 | 5552 | } else { |
6bda666a | 5553 | lock_page(page); |
0fe6e20b NH |
5554 | if (unlikely(anon_vma_prepare(vma))) { |
5555 | ret = VM_FAULT_OOM; | |
5556 | goto backout_unlocked; | |
5557 | } | |
409eb8c2 | 5558 | anon_rmap = 1; |
23be7468 | 5559 | } |
0fe6e20b | 5560 | } else { |
998b4382 NH |
5561 | /* |
5562 | * If memory error occurs between mmap() and fault, some process | |
5563 | * don't have hwpoisoned swap entry for errored virtual address. | |
5564 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
5565 | */ | |
5566 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 5567 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5568 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
5569 | goto backout_unlocked; |
5570 | } | |
7677f7fd AR |
5571 | |
5572 | /* Check for page in userfault range. */ | |
5573 | if (userfaultfd_minor(vma)) { | |
5574 | unlock_page(page); | |
5575 | put_page(page); | |
5576 | ret = hugetlb_handle_userfault(vma, mapping, idx, | |
824ddc60 | 5577 | flags, haddr, address, |
7677f7fd AR |
5578 | VM_UFFD_MINOR); |
5579 | goto out; | |
5580 | } | |
6bda666a | 5581 | } |
1e8f889b | 5582 | |
57303d80 AW |
5583 | /* |
5584 | * If we are going to COW a private mapping later, we examine the | |
5585 | * pending reservations for this page now. This will ensure that | |
5586 | * any allocations necessary to record that reservation occur outside | |
5587 | * the spinlock. | |
5588 | */ | |
5e911373 | 5589 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 5590 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
5591 | ret = VM_FAULT_OOM; |
5592 | goto backout_unlocked; | |
5593 | } | |
5e911373 | 5594 | /* Just decrements count, does not deallocate */ |
285b8dca | 5595 | vma_end_reservation(h, vma, haddr); |
5e911373 | 5596 | } |
57303d80 | 5597 | |
8bea8052 | 5598 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 5599 | ret = 0; |
c64e912c PX |
5600 | /* If pte changed from under us, retry */ |
5601 | if (!pte_same(huge_ptep_get(ptep), old_pte)) | |
4c887265 AL |
5602 | goto backout; |
5603 | ||
07443a85 | 5604 | if (anon_rmap) { |
d6995da3 | 5605 | ClearHPageRestoreReserve(page); |
285b8dca | 5606 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 5607 | } else |
fb3d824d | 5608 | page_dup_file_rmap(page, true); |
1e8f889b DG |
5609 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
5610 | && (vma->vm_flags & VM_SHARED))); | |
c64e912c PX |
5611 | /* |
5612 | * If this pte was previously wr-protected, keep it wr-protected even | |
5613 | * if populated. | |
5614 | */ | |
5615 | if (unlikely(pte_marker_uffd_wp(old_pte))) | |
5616 | new_pte = huge_pte_wrprotect(huge_pte_mkuffd_wp(new_pte)); | |
285b8dca | 5617 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 5618 | |
5d317b2b | 5619 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 5620 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 5621 | /* Optimization, do the COW without a second fault */ |
c89357e2 | 5622 | ret = hugetlb_wp(mm, vma, address, ptep, flags, page, ptl); |
1e8f889b DG |
5623 | } |
5624 | ||
cb900f41 | 5625 | spin_unlock(ptl); |
cb6acd01 MK |
5626 | |
5627 | /* | |
8f251a3d MK |
5628 | * Only set HPageMigratable in newly allocated pages. Existing pages |
5629 | * found in the pagecache may not have HPageMigratableset if they have | |
5630 | * been isolated for migration. | |
cb6acd01 MK |
5631 | */ |
5632 | if (new_page) | |
8f251a3d | 5633 | SetHPageMigratable(page); |
cb6acd01 | 5634 | |
4c887265 AL |
5635 | unlock_page(page); |
5636 | out: | |
ac9b9c66 | 5637 | return ret; |
4c887265 AL |
5638 | |
5639 | backout: | |
cb900f41 | 5640 | spin_unlock(ptl); |
2b26736c | 5641 | backout_unlocked: |
4c887265 | 5642 | unlock_page(page); |
c7b1850d MK |
5643 | /* restore reserve for newly allocated pages not in page cache */ |
5644 | if (new_page && !new_pagecache_page) | |
5645 | restore_reserve_on_error(h, vma, haddr, page); | |
4c887265 AL |
5646 | put_page(page); |
5647 | goto out; | |
ac9b9c66 HD |
5648 | } |
5649 | ||
8382d914 | 5650 | #ifdef CONFIG_SMP |
188b04a7 | 5651 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5652 | { |
5653 | unsigned long key[2]; | |
5654 | u32 hash; | |
5655 | ||
1b426bac MK |
5656 | key[0] = (unsigned long) mapping; |
5657 | key[1] = idx; | |
8382d914 | 5658 | |
55254636 | 5659 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
5660 | |
5661 | return hash & (num_fault_mutexes - 1); | |
5662 | } | |
5663 | #else | |
5664 | /* | |
6c26d310 | 5665 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
5666 | * return 0 and avoid the hashing overhead. |
5667 | */ | |
188b04a7 | 5668 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5669 | { |
5670 | return 0; | |
5671 | } | |
5672 | #endif | |
5673 | ||
2b740303 | 5674 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 5675 | unsigned long address, unsigned int flags) |
86e5216f | 5676 | { |
8382d914 | 5677 | pte_t *ptep, entry; |
cb900f41 | 5678 | spinlock_t *ptl; |
2b740303 | 5679 | vm_fault_t ret; |
8382d914 DB |
5680 | u32 hash; |
5681 | pgoff_t idx; | |
0fe6e20b | 5682 | struct page *page = NULL; |
57303d80 | 5683 | struct page *pagecache_page = NULL; |
a5516438 | 5684 | struct hstate *h = hstate_vma(vma); |
8382d914 | 5685 | struct address_space *mapping; |
0f792cf9 | 5686 | int need_wait_lock = 0; |
285b8dca | 5687 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 5688 | |
285b8dca | 5689 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 5690 | if (ptep) { |
c0d0381a MK |
5691 | /* |
5692 | * Since we hold no locks, ptep could be stale. That is | |
5693 | * OK as we are only making decisions based on content and | |
5694 | * not actually modifying content here. | |
5695 | */ | |
fd6a03ed | 5696 | entry = huge_ptep_get(ptep); |
290408d4 | 5697 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 5698 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
5699 | return 0; |
5700 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 5701 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5702 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
5703 | } |
5704 | ||
c0d0381a MK |
5705 | /* |
5706 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
5707 | * until finished with ptep. This serves two purposes: |
5708 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
5709 | * and making the ptep no longer valid. | |
5710 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
5711 | * |
5712 | * ptep could have already be assigned via huge_pte_offset. That | |
5713 | * is OK, as huge_pte_alloc will return the same value unless | |
5714 | * something has changed. | |
5715 | */ | |
8382d914 | 5716 | mapping = vma->vm_file->f_mapping; |
c0d0381a | 5717 | i_mmap_lock_read(mapping); |
aec44e0f | 5718 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); |
c0d0381a MK |
5719 | if (!ptep) { |
5720 | i_mmap_unlock_read(mapping); | |
5721 | return VM_FAULT_OOM; | |
5722 | } | |
8382d914 | 5723 | |
3935baa9 DG |
5724 | /* |
5725 | * Serialize hugepage allocation and instantiation, so that we don't | |
5726 | * get spurious allocation failures if two CPUs race to instantiate | |
5727 | * the same page in the page cache. | |
5728 | */ | |
c0d0381a | 5729 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 5730 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 5731 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 5732 | |
7f2e9525 | 5733 | entry = huge_ptep_get(ptep); |
c64e912c PX |
5734 | /* PTE markers should be handled the same way as none pte */ |
5735 | if (huge_pte_none_mostly(entry)) { | |
5736 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, | |
5737 | entry, flags); | |
b4d1d99f | 5738 | goto out_mutex; |
3935baa9 | 5739 | } |
86e5216f | 5740 | |
83c54070 | 5741 | ret = 0; |
1e8f889b | 5742 | |
0f792cf9 NH |
5743 | /* |
5744 | * entry could be a migration/hwpoison entry at this point, so this | |
5745 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
5746 | * an active hugepage in pagecache. This goto expects the 2nd page |
5747 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
5748 | * properly handle it. | |
0f792cf9 NH |
5749 | */ |
5750 | if (!pte_present(entry)) | |
5751 | goto out_mutex; | |
5752 | ||
57303d80 | 5753 | /* |
c89357e2 DH |
5754 | * If we are going to COW/unshare the mapping later, we examine the |
5755 | * pending reservations for this page now. This will ensure that any | |
57303d80 AW |
5756 | * allocations necessary to record that reservation occur outside the |
5757 | * spinlock. For private mappings, we also lookup the pagecache | |
5758 | * page now as it is used to determine if a reservation has been | |
5759 | * consumed. | |
5760 | */ | |
c89357e2 DH |
5761 | if ((flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) && |
5762 | !huge_pte_write(entry)) { | |
285b8dca | 5763 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 5764 | ret = VM_FAULT_OOM; |
b4d1d99f | 5765 | goto out_mutex; |
2b26736c | 5766 | } |
5e911373 | 5767 | /* Just decrements count, does not deallocate */ |
285b8dca | 5768 | vma_end_reservation(h, vma, haddr); |
57303d80 | 5769 | |
f83a275d | 5770 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 5771 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 5772 | vma, haddr); |
57303d80 AW |
5773 | } |
5774 | ||
0f792cf9 NH |
5775 | ptl = huge_pte_lock(h, mm, ptep); |
5776 | ||
c89357e2 | 5777 | /* Check for a racing update before calling hugetlb_wp() */ |
0f792cf9 NH |
5778 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) |
5779 | goto out_ptl; | |
5780 | ||
166f3ecc PX |
5781 | /* Handle userfault-wp first, before trying to lock more pages */ |
5782 | if (userfaultfd_wp(vma) && huge_pte_uffd_wp(huge_ptep_get(ptep)) && | |
5783 | (flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { | |
5784 | struct vm_fault vmf = { | |
5785 | .vma = vma, | |
5786 | .address = haddr, | |
5787 | .real_address = address, | |
5788 | .flags = flags, | |
5789 | }; | |
5790 | ||
5791 | spin_unlock(ptl); | |
5792 | if (pagecache_page) { | |
5793 | unlock_page(pagecache_page); | |
5794 | put_page(pagecache_page); | |
5795 | } | |
5796 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
5797 | i_mmap_unlock_read(mapping); | |
5798 | return handle_userfault(&vmf, VM_UFFD_WP); | |
5799 | } | |
5800 | ||
56c9cfb1 | 5801 | /* |
c89357e2 | 5802 | * hugetlb_wp() requires page locks of pte_page(entry) and |
56c9cfb1 NH |
5803 | * pagecache_page, so here we need take the former one |
5804 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
5805 | */ |
5806 | page = pte_page(entry); | |
5807 | if (page != pagecache_page) | |
0f792cf9 NH |
5808 | if (!trylock_page(page)) { |
5809 | need_wait_lock = 1; | |
5810 | goto out_ptl; | |
5811 | } | |
b4d1d99f | 5812 | |
0f792cf9 | 5813 | get_page(page); |
b4d1d99f | 5814 | |
c89357e2 | 5815 | if (flags & (FAULT_FLAG_WRITE|FAULT_FLAG_UNSHARE)) { |
106c992a | 5816 | if (!huge_pte_write(entry)) { |
c89357e2 DH |
5817 | ret = hugetlb_wp(mm, vma, address, ptep, flags, |
5818 | pagecache_page, ptl); | |
0f792cf9 | 5819 | goto out_put_page; |
c89357e2 DH |
5820 | } else if (likely(flags & FAULT_FLAG_WRITE)) { |
5821 | entry = huge_pte_mkdirty(entry); | |
b4d1d99f | 5822 | } |
b4d1d99f DG |
5823 | } |
5824 | entry = pte_mkyoung(entry); | |
285b8dca | 5825 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 5826 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 5827 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
5828 | out_put_page: |
5829 | if (page != pagecache_page) | |
5830 | unlock_page(page); | |
5831 | put_page(page); | |
cb900f41 KS |
5832 | out_ptl: |
5833 | spin_unlock(ptl); | |
57303d80 AW |
5834 | |
5835 | if (pagecache_page) { | |
5836 | unlock_page(pagecache_page); | |
5837 | put_page(pagecache_page); | |
5838 | } | |
b4d1d99f | 5839 | out_mutex: |
c672c7f2 | 5840 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 5841 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
5842 | /* |
5843 | * Generally it's safe to hold refcount during waiting page lock. But | |
5844 | * here we just wait to defer the next page fault to avoid busy loop and | |
5845 | * the page is not used after unlocked before returning from the current | |
5846 | * page fault. So we are safe from accessing freed page, even if we wait | |
5847 | * here without taking refcount. | |
5848 | */ | |
5849 | if (need_wait_lock) | |
5850 | wait_on_page_locked(page); | |
1e8f889b | 5851 | return ret; |
86e5216f AL |
5852 | } |
5853 | ||
714c1891 | 5854 | #ifdef CONFIG_USERFAULTFD |
8fb5debc MK |
5855 | /* |
5856 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
5857 | * modifications for huge pages. | |
5858 | */ | |
5859 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
5860 | pte_t *dst_pte, | |
5861 | struct vm_area_struct *dst_vma, | |
5862 | unsigned long dst_addr, | |
5863 | unsigned long src_addr, | |
f6191471 | 5864 | enum mcopy_atomic_mode mode, |
6041c691 PX |
5865 | struct page **pagep, |
5866 | bool wp_copy) | |
8fb5debc | 5867 | { |
f6191471 | 5868 | bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); |
8cc5fcbb MA |
5869 | struct hstate *h = hstate_vma(dst_vma); |
5870 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
5871 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
1e392147 | 5872 | unsigned long size; |
1c9e8def | 5873 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
5874 | pte_t _dst_pte; |
5875 | spinlock_t *ptl; | |
8cc5fcbb | 5876 | int ret = -ENOMEM; |
8fb5debc | 5877 | struct page *page; |
f6191471 | 5878 | int writable; |
cc30042d | 5879 | bool page_in_pagecache = false; |
8fb5debc | 5880 | |
f6191471 AR |
5881 | if (is_continue) { |
5882 | ret = -EFAULT; | |
5883 | page = find_lock_page(mapping, idx); | |
5884 | if (!page) | |
5885 | goto out; | |
cc30042d | 5886 | page_in_pagecache = true; |
f6191471 | 5887 | } else if (!*pagep) { |
d84cf06e MA |
5888 | /* If a page already exists, then it's UFFDIO_COPY for |
5889 | * a non-missing case. Return -EEXIST. | |
5890 | */ | |
5891 | if (vm_shared && | |
5892 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5893 | ret = -EEXIST; | |
5894 | goto out; | |
5895 | } | |
5896 | ||
8fb5debc | 5897 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
d84cf06e MA |
5898 | if (IS_ERR(page)) { |
5899 | ret = -ENOMEM; | |
8fb5debc | 5900 | goto out; |
d84cf06e | 5901 | } |
8fb5debc MK |
5902 | |
5903 | ret = copy_huge_page_from_user(page, | |
5904 | (const void __user *) src_addr, | |
810a56b9 | 5905 | pages_per_huge_page(h), false); |
8fb5debc | 5906 | |
c1e8d7c6 | 5907 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 5908 | if (unlikely(ret)) { |
9e368259 | 5909 | ret = -ENOENT; |
8cc5fcbb MA |
5910 | /* Free the allocated page which may have |
5911 | * consumed a reservation. | |
5912 | */ | |
5913 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
5914 | put_page(page); | |
5915 | ||
5916 | /* Allocate a temporary page to hold the copied | |
5917 | * contents. | |
5918 | */ | |
5919 | page = alloc_huge_page_vma(h, dst_vma, dst_addr); | |
5920 | if (!page) { | |
5921 | ret = -ENOMEM; | |
5922 | goto out; | |
5923 | } | |
8fb5debc | 5924 | *pagep = page; |
8cc5fcbb MA |
5925 | /* Set the outparam pagep and return to the caller to |
5926 | * copy the contents outside the lock. Don't free the | |
5927 | * page. | |
5928 | */ | |
8fb5debc MK |
5929 | goto out; |
5930 | } | |
5931 | } else { | |
8cc5fcbb MA |
5932 | if (vm_shared && |
5933 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5934 | put_page(*pagep); | |
5935 | ret = -EEXIST; | |
5936 | *pagep = NULL; | |
5937 | goto out; | |
5938 | } | |
5939 | ||
5940 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
5941 | if (IS_ERR(page)) { | |
5942 | ret = -ENOMEM; | |
5943 | *pagep = NULL; | |
5944 | goto out; | |
5945 | } | |
34892366 MS |
5946 | copy_user_huge_page(page, *pagep, dst_addr, dst_vma, |
5947 | pages_per_huge_page(h)); | |
8cc5fcbb | 5948 | put_page(*pagep); |
8fb5debc MK |
5949 | *pagep = NULL; |
5950 | } | |
5951 | ||
5952 | /* | |
5953 | * The memory barrier inside __SetPageUptodate makes sure that | |
5954 | * preceding stores to the page contents become visible before | |
5955 | * the set_pte_at() write. | |
5956 | */ | |
5957 | __SetPageUptodate(page); | |
8fb5debc | 5958 | |
f6191471 AR |
5959 | /* Add shared, newly allocated pages to the page cache. */ |
5960 | if (vm_shared && !is_continue) { | |
1e392147 AA |
5961 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5962 | ret = -EFAULT; | |
5963 | if (idx >= size) | |
5964 | goto out_release_nounlock; | |
1c9e8def | 5965 | |
1e392147 AA |
5966 | /* |
5967 | * Serialization between remove_inode_hugepages() and | |
5968 | * huge_add_to_page_cache() below happens through the | |
5969 | * hugetlb_fault_mutex_table that here must be hold by | |
5970 | * the caller. | |
5971 | */ | |
1c9e8def MK |
5972 | ret = huge_add_to_page_cache(page, mapping, idx); |
5973 | if (ret) | |
5974 | goto out_release_nounlock; | |
cc30042d | 5975 | page_in_pagecache = true; |
1c9e8def MK |
5976 | } |
5977 | ||
8fb5debc MK |
5978 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
5979 | spin_lock(ptl); | |
5980 | ||
1e392147 AA |
5981 | /* |
5982 | * Recheck the i_size after holding PT lock to make sure not | |
5983 | * to leave any page mapped (as page_mapped()) beyond the end | |
5984 | * of the i_size (remove_inode_hugepages() is strict about | |
5985 | * enforcing that). If we bail out here, we'll also leave a | |
5986 | * page in the radix tree in the vm_shared case beyond the end | |
5987 | * of the i_size, but remove_inode_hugepages() will take care | |
5988 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
5989 | */ | |
5990 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
5991 | ret = -EFAULT; | |
5992 | if (idx >= size) | |
5993 | goto out_release_unlock; | |
5994 | ||
8fb5debc | 5995 | ret = -EEXIST; |
6041c691 PX |
5996 | /* |
5997 | * We allow to overwrite a pte marker: consider when both MISSING|WP | |
5998 | * registered, we firstly wr-protect a none pte which has no page cache | |
5999 | * page backing it, then access the page. | |
6000 | */ | |
6001 | if (!huge_pte_none_mostly(huge_ptep_get(dst_pte))) | |
8fb5debc MK |
6002 | goto out_release_unlock; |
6003 | ||
1c9e8def | 6004 | if (vm_shared) { |
fb3d824d | 6005 | page_dup_file_rmap(page, true); |
1c9e8def | 6006 | } else { |
d6995da3 | 6007 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
6008 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
6009 | } | |
8fb5debc | 6010 | |
6041c691 PX |
6011 | /* |
6012 | * For either: (1) CONTINUE on a non-shared VMA, or (2) UFFDIO_COPY | |
6013 | * with wp flag set, don't set pte write bit. | |
6014 | */ | |
6015 | if (wp_copy || (is_continue && !vm_shared)) | |
f6191471 AR |
6016 | writable = 0; |
6017 | else | |
6018 | writable = dst_vma->vm_flags & VM_WRITE; | |
6019 | ||
6020 | _dst_pte = make_huge_pte(dst_vma, page, writable); | |
6041c691 PX |
6021 | /* |
6022 | * Always mark UFFDIO_COPY page dirty; note that this may not be | |
6023 | * extremely important for hugetlbfs for now since swapping is not | |
6024 | * supported, but we should still be clear in that this page cannot be | |
6025 | * thrown away at will, even if write bit not set. | |
6026 | */ | |
6027 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
8fb5debc MK |
6028 | _dst_pte = pte_mkyoung(_dst_pte); |
6029 | ||
6041c691 PX |
6030 | if (wp_copy) |
6031 | _dst_pte = huge_pte_mkuffd_wp(_dst_pte); | |
6032 | ||
8fb5debc MK |
6033 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); |
6034 | ||
6035 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
6036 | dst_vma->vm_flags & VM_WRITE); | |
6037 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
6038 | ||
6039 | /* No need to invalidate - it was non-present before */ | |
6040 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
6041 | ||
6042 | spin_unlock(ptl); | |
f6191471 AR |
6043 | if (!is_continue) |
6044 | SetHPageMigratable(page); | |
6045 | if (vm_shared || is_continue) | |
1c9e8def | 6046 | unlock_page(page); |
8fb5debc MK |
6047 | ret = 0; |
6048 | out: | |
6049 | return ret; | |
6050 | out_release_unlock: | |
6051 | spin_unlock(ptl); | |
f6191471 | 6052 | if (vm_shared || is_continue) |
1c9e8def | 6053 | unlock_page(page); |
5af10dfd | 6054 | out_release_nounlock: |
cc30042d | 6055 | if (!page_in_pagecache) |
c7b1850d | 6056 | restore_reserve_on_error(h, dst_vma, dst_addr, page); |
8fb5debc MK |
6057 | put_page(page); |
6058 | goto out; | |
6059 | } | |
714c1891 | 6060 | #endif /* CONFIG_USERFAULTFD */ |
8fb5debc | 6061 | |
82e5d378 JM |
6062 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
6063 | int refs, struct page **pages, | |
6064 | struct vm_area_struct **vmas) | |
6065 | { | |
6066 | int nr; | |
6067 | ||
6068 | for (nr = 0; nr < refs; nr++) { | |
6069 | if (likely(pages)) | |
6070 | pages[nr] = mem_map_offset(page, nr); | |
6071 | if (vmas) | |
6072 | vmas[nr] = vma; | |
6073 | } | |
6074 | } | |
6075 | ||
a7f22660 DH |
6076 | static inline bool __follow_hugetlb_must_fault(unsigned int flags, pte_t *pte, |
6077 | bool *unshare) | |
6078 | { | |
6079 | pte_t pteval = huge_ptep_get(pte); | |
6080 | ||
6081 | *unshare = false; | |
6082 | if (is_swap_pte(pteval)) | |
6083 | return true; | |
6084 | if (huge_pte_write(pteval)) | |
6085 | return false; | |
6086 | if (flags & FOLL_WRITE) | |
6087 | return true; | |
6088 | if (gup_must_unshare(flags, pte_page(pteval))) { | |
6089 | *unshare = true; | |
6090 | return true; | |
6091 | } | |
6092 | return false; | |
6093 | } | |
6094 | ||
28a35716 ML |
6095 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
6096 | struct page **pages, struct vm_area_struct **vmas, | |
6097 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 6098 | long i, unsigned int flags, int *locked) |
63551ae0 | 6099 | { |
d5d4b0aa CK |
6100 | unsigned long pfn_offset; |
6101 | unsigned long vaddr = *position; | |
28a35716 | 6102 | unsigned long remainder = *nr_pages; |
a5516438 | 6103 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 6104 | int err = -EFAULT, refs; |
63551ae0 | 6105 | |
63551ae0 | 6106 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 6107 | pte_t *pte; |
cb900f41 | 6108 | spinlock_t *ptl = NULL; |
a7f22660 | 6109 | bool unshare = false; |
2a15efc9 | 6110 | int absent; |
4c887265 | 6111 | struct page *page; |
63551ae0 | 6112 | |
02057967 DR |
6113 | /* |
6114 | * If we have a pending SIGKILL, don't keep faulting pages and | |
6115 | * potentially allocating memory. | |
6116 | */ | |
fa45f116 | 6117 | if (fatal_signal_pending(current)) { |
02057967 DR |
6118 | remainder = 0; |
6119 | break; | |
6120 | } | |
6121 | ||
4c887265 AL |
6122 | /* |
6123 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 6124 | * each hugepage. We have to make sure we get the |
4c887265 | 6125 | * first, for the page indexing below to work. |
cb900f41 KS |
6126 | * |
6127 | * Note that page table lock is not held when pte is null. | |
4c887265 | 6128 | */ |
7868a208 PA |
6129 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
6130 | huge_page_size(h)); | |
cb900f41 KS |
6131 | if (pte) |
6132 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
6133 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
6134 | ||
6135 | /* | |
6136 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
6137 | * an error where there's an empty slot with no huge pagecache |
6138 | * to back it. This way, we avoid allocating a hugepage, and | |
6139 | * the sparse dumpfile avoids allocating disk blocks, but its | |
6140 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 6141 | */ |
3ae77f43 HD |
6142 | if (absent && (flags & FOLL_DUMP) && |
6143 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
6144 | if (pte) |
6145 | spin_unlock(ptl); | |
2a15efc9 HD |
6146 | remainder = 0; |
6147 | break; | |
6148 | } | |
63551ae0 | 6149 | |
9cc3a5bd NH |
6150 | /* |
6151 | * We need call hugetlb_fault for both hugepages under migration | |
6152 | * (in which case hugetlb_fault waits for the migration,) and | |
6153 | * hwpoisoned hugepages (in which case we need to prevent the | |
6154 | * caller from accessing to them.) In order to do this, we use | |
6155 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
6156 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
6157 | * both cases, and because we can't follow correct pages | |
6158 | * directly from any kind of swap entries. | |
6159 | */ | |
a7f22660 DH |
6160 | if (absent || |
6161 | __follow_hugetlb_must_fault(flags, pte, &unshare)) { | |
2b740303 | 6162 | vm_fault_t ret; |
87ffc118 | 6163 | unsigned int fault_flags = 0; |
63551ae0 | 6164 | |
cb900f41 KS |
6165 | if (pte) |
6166 | spin_unlock(ptl); | |
87ffc118 AA |
6167 | if (flags & FOLL_WRITE) |
6168 | fault_flags |= FAULT_FLAG_WRITE; | |
a7f22660 DH |
6169 | else if (unshare) |
6170 | fault_flags |= FAULT_FLAG_UNSHARE; | |
4f6da934 | 6171 | if (locked) |
71335f37 PX |
6172 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
6173 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
6174 | if (flags & FOLL_NOWAIT) |
6175 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
6176 | FAULT_FLAG_RETRY_NOWAIT; | |
6177 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
6178 | /* |
6179 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
6180 | * FAULT_FLAG_TRIED can co-exist | |
6181 | */ | |
87ffc118 AA |
6182 | fault_flags |= FAULT_FLAG_TRIED; |
6183 | } | |
6184 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
6185 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 6186 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
6187 | remainder = 0; |
6188 | break; | |
6189 | } | |
6190 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 6191 | if (locked && |
1ac25013 | 6192 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 6193 | *locked = 0; |
87ffc118 AA |
6194 | *nr_pages = 0; |
6195 | /* | |
6196 | * VM_FAULT_RETRY must not return an | |
6197 | * error, it will return zero | |
6198 | * instead. | |
6199 | * | |
6200 | * No need to update "position" as the | |
6201 | * caller will not check it after | |
6202 | * *nr_pages is set to 0. | |
6203 | */ | |
6204 | return i; | |
6205 | } | |
6206 | continue; | |
4c887265 AL |
6207 | } |
6208 | ||
a5516438 | 6209 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 6210 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 6211 | |
b6a2619c DH |
6212 | VM_BUG_ON_PAGE((flags & FOLL_PIN) && PageAnon(page) && |
6213 | !PageAnonExclusive(page), page); | |
6214 | ||
acbfb087 ZL |
6215 | /* |
6216 | * If subpage information not requested, update counters | |
6217 | * and skip the same_page loop below. | |
6218 | */ | |
6219 | if (!pages && !vmas && !pfn_offset && | |
6220 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
6221 | (remainder >= pages_per_huge_page(h))) { | |
6222 | vaddr += huge_page_size(h); | |
6223 | remainder -= pages_per_huge_page(h); | |
6224 | i += pages_per_huge_page(h); | |
6225 | spin_unlock(ptl); | |
6226 | continue; | |
6227 | } | |
6228 | ||
d08af0a5 JM |
6229 | /* vaddr may not be aligned to PAGE_SIZE */ |
6230 | refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, | |
6231 | (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); | |
0fa5bc40 | 6232 | |
82e5d378 JM |
6233 | if (pages || vmas) |
6234 | record_subpages_vmas(mem_map_offset(page, pfn_offset), | |
6235 | vma, refs, | |
6236 | likely(pages) ? pages + i : NULL, | |
6237 | vmas ? vmas + i : NULL); | |
63551ae0 | 6238 | |
82e5d378 | 6239 | if (pages) { |
0fa5bc40 | 6240 | /* |
822951d8 | 6241 | * try_grab_folio() should always succeed here, |
0fa5bc40 JM |
6242 | * because: a) we hold the ptl lock, and b) we've just |
6243 | * checked that the huge page is present in the page | |
6244 | * tables. If the huge page is present, then the tail | |
6245 | * pages must also be present. The ptl prevents the | |
6246 | * head page and tail pages from being rearranged in | |
6247 | * any way. So this page must be available at this | |
6248 | * point, unless the page refcount overflowed: | |
6249 | */ | |
822951d8 MWO |
6250 | if (WARN_ON_ONCE(!try_grab_folio(pages[i], refs, |
6251 | flags))) { | |
0fa5bc40 JM |
6252 | spin_unlock(ptl); |
6253 | remainder = 0; | |
6254 | err = -ENOMEM; | |
6255 | break; | |
6256 | } | |
d5d4b0aa | 6257 | } |
82e5d378 JM |
6258 | |
6259 | vaddr += (refs << PAGE_SHIFT); | |
6260 | remainder -= refs; | |
6261 | i += refs; | |
6262 | ||
cb900f41 | 6263 | spin_unlock(ptl); |
63551ae0 | 6264 | } |
28a35716 | 6265 | *nr_pages = remainder; |
87ffc118 AA |
6266 | /* |
6267 | * setting position is actually required only if remainder is | |
6268 | * not zero but it's faster not to add a "if (remainder)" | |
6269 | * branch. | |
6270 | */ | |
63551ae0 DG |
6271 | *position = vaddr; |
6272 | ||
2be7cfed | 6273 | return i ? i : err; |
63551ae0 | 6274 | } |
8f860591 | 6275 | |
7da4d641 | 6276 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
5a90d5a1 PX |
6277 | unsigned long address, unsigned long end, |
6278 | pgprot_t newprot, unsigned long cp_flags) | |
8f860591 ZY |
6279 | { |
6280 | struct mm_struct *mm = vma->vm_mm; | |
6281 | unsigned long start = address; | |
6282 | pte_t *ptep; | |
6283 | pte_t pte; | |
a5516438 | 6284 | struct hstate *h = hstate_vma(vma); |
60dfaad6 | 6285 | unsigned long pages = 0, psize = huge_page_size(h); |
dff11abe | 6286 | bool shared_pmd = false; |
ac46d4f3 | 6287 | struct mmu_notifier_range range; |
5a90d5a1 PX |
6288 | bool uffd_wp = cp_flags & MM_CP_UFFD_WP; |
6289 | bool uffd_wp_resolve = cp_flags & MM_CP_UFFD_WP_RESOLVE; | |
dff11abe MK |
6290 | |
6291 | /* | |
6292 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 6293 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
6294 | * range if PMD sharing is possible. |
6295 | */ | |
7269f999 JG |
6296 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
6297 | 0, vma, mm, start, end); | |
ac46d4f3 | 6298 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
6299 | |
6300 | BUG_ON(address >= end); | |
ac46d4f3 | 6301 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 6302 | |
ac46d4f3 | 6303 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 6304 | i_mmap_lock_write(vma->vm_file->f_mapping); |
60dfaad6 | 6305 | for (; address < end; address += psize) { |
cb900f41 | 6306 | spinlock_t *ptl; |
60dfaad6 | 6307 | ptep = huge_pte_offset(mm, address, psize); |
8f860591 ZY |
6308 | if (!ptep) |
6309 | continue; | |
cb900f41 | 6310 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 6311 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
60dfaad6 PX |
6312 | /* |
6313 | * When uffd-wp is enabled on the vma, unshare | |
6314 | * shouldn't happen at all. Warn about it if it | |
6315 | * happened due to some reason. | |
6316 | */ | |
6317 | WARN_ON_ONCE(uffd_wp || uffd_wp_resolve); | |
7da4d641 | 6318 | pages++; |
cb900f41 | 6319 | spin_unlock(ptl); |
dff11abe | 6320 | shared_pmd = true; |
39dde65c | 6321 | continue; |
7da4d641 | 6322 | } |
a8bda28d NH |
6323 | pte = huge_ptep_get(ptep); |
6324 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
6325 | spin_unlock(ptl); | |
6326 | continue; | |
6327 | } | |
6328 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
6329 | swp_entry_t entry = pte_to_swp_entry(pte); | |
6c287605 | 6330 | struct page *page = pfn_swap_entry_to_page(entry); |
a8bda28d | 6331 | |
6c287605 | 6332 | if (!is_readable_migration_entry(entry)) { |
a8bda28d NH |
6333 | pte_t newpte; |
6334 | ||
6c287605 DH |
6335 | if (PageAnon(page)) |
6336 | entry = make_readable_exclusive_migration_entry( | |
6337 | swp_offset(entry)); | |
6338 | else | |
6339 | entry = make_readable_migration_entry( | |
6340 | swp_offset(entry)); | |
a8bda28d | 6341 | newpte = swp_entry_to_pte(entry); |
5a90d5a1 PX |
6342 | if (uffd_wp) |
6343 | newpte = pte_swp_mkuffd_wp(newpte); | |
6344 | else if (uffd_wp_resolve) | |
6345 | newpte = pte_swp_clear_uffd_wp(newpte); | |
e5251fd4 | 6346 | set_huge_swap_pte_at(mm, address, ptep, |
60dfaad6 | 6347 | newpte, psize); |
a8bda28d NH |
6348 | pages++; |
6349 | } | |
6350 | spin_unlock(ptl); | |
6351 | continue; | |
6352 | } | |
60dfaad6 PX |
6353 | if (unlikely(pte_marker_uffd_wp(pte))) { |
6354 | /* | |
6355 | * This is changing a non-present pte into a none pte, | |
6356 | * no need for huge_ptep_modify_prot_start/commit(). | |
6357 | */ | |
6358 | if (uffd_wp_resolve) | |
6359 | huge_pte_clear(mm, address, ptep, psize); | |
6360 | } | |
a8bda28d | 6361 | if (!huge_pte_none(pte)) { |
023bdd00 | 6362 | pte_t old_pte; |
79c1c594 | 6363 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
023bdd00 AK |
6364 | |
6365 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
16785bd7 | 6366 | pte = huge_pte_modify(old_pte, newprot); |
79c1c594 | 6367 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
5a90d5a1 PX |
6368 | if (uffd_wp) |
6369 | pte = huge_pte_mkuffd_wp(huge_pte_wrprotect(pte)); | |
6370 | else if (uffd_wp_resolve) | |
6371 | pte = huge_pte_clear_uffd_wp(pte); | |
023bdd00 | 6372 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 6373 | pages++; |
60dfaad6 PX |
6374 | } else { |
6375 | /* None pte */ | |
6376 | if (unlikely(uffd_wp)) | |
6377 | /* Safe to modify directly (none->non-present). */ | |
6378 | set_huge_pte_at(mm, address, ptep, | |
6379 | make_pte_marker(PTE_MARKER_UFFD_WP)); | |
8f860591 | 6380 | } |
cb900f41 | 6381 | spin_unlock(ptl); |
8f860591 | 6382 | } |
d833352a | 6383 | /* |
c8c06efa | 6384 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 6385 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 6386 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
6387 | * and that page table be reused and filled with junk. If we actually |
6388 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 6389 | */ |
dff11abe | 6390 | if (shared_pmd) |
ac46d4f3 | 6391 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
6392 | else |
6393 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
6394 | /* |
6395 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
6396 | * page table protection not changing it to point to a new page. | |
6397 | * | |
ad56b738 | 6398 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 6399 | */ |
83cde9e8 | 6400 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 6401 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
6402 | |
6403 | return pages << h->order; | |
8f860591 ZY |
6404 | } |
6405 | ||
33b8f84a MK |
6406 | /* Return true if reservation was successful, false otherwise. */ |
6407 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 6408 | long from, long to, |
5a6fe125 | 6409 | struct vm_area_struct *vma, |
ca16d140 | 6410 | vm_flags_t vm_flags) |
e4e574b7 | 6411 | { |
33b8f84a | 6412 | long chg, add = -1; |
a5516438 | 6413 | struct hstate *h = hstate_inode(inode); |
90481622 | 6414 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 6415 | struct resv_map *resv_map; |
075a61d0 | 6416 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 6417 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 6418 | |
63489f8e MK |
6419 | /* This should never happen */ |
6420 | if (from > to) { | |
6421 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 6422 | return false; |
63489f8e MK |
6423 | } |
6424 | ||
17c9d12e MG |
6425 | /* |
6426 | * Only apply hugepage reservation if asked. At fault time, an | |
6427 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 6428 | * without using reserves |
17c9d12e | 6429 | */ |
ca16d140 | 6430 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 6431 | return true; |
17c9d12e | 6432 | |
a1e78772 MG |
6433 | /* |
6434 | * Shared mappings base their reservation on the number of pages that | |
6435 | * are already allocated on behalf of the file. Private mappings need | |
6436 | * to reserve the full area even if read-only as mprotect() may be | |
6437 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
6438 | */ | |
9119a41e | 6439 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
6440 | /* |
6441 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
6442 | * called for inodes for which resv_maps were created (see | |
6443 | * hugetlbfs_get_inode). | |
6444 | */ | |
4e35f483 | 6445 | resv_map = inode_resv_map(inode); |
9119a41e | 6446 | |
0db9d74e | 6447 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
6448 | |
6449 | } else { | |
e9fe92ae | 6450 | /* Private mapping. */ |
9119a41e | 6451 | resv_map = resv_map_alloc(); |
17c9d12e | 6452 | if (!resv_map) |
33b8f84a | 6453 | return false; |
17c9d12e | 6454 | |
a1e78772 | 6455 | chg = to - from; |
84afd99b | 6456 | |
17c9d12e MG |
6457 | set_vma_resv_map(vma, resv_map); |
6458 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
6459 | } | |
6460 | ||
33b8f84a | 6461 | if (chg < 0) |
c50ac050 | 6462 | goto out_err; |
8a630112 | 6463 | |
33b8f84a MK |
6464 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
6465 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 6466 | goto out_err; |
075a61d0 MA |
6467 | |
6468 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
6469 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
6470 | * of the resv_map. | |
6471 | */ | |
6472 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
6473 | } | |
6474 | ||
1c5ecae3 MK |
6475 | /* |
6476 | * There must be enough pages in the subpool for the mapping. If | |
6477 | * the subpool has a minimum size, there may be some global | |
6478 | * reservations already in place (gbl_reserve). | |
6479 | */ | |
6480 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 6481 | if (gbl_reserve < 0) |
075a61d0 | 6482 | goto out_uncharge_cgroup; |
5a6fe125 MG |
6483 | |
6484 | /* | |
17c9d12e | 6485 | * Check enough hugepages are available for the reservation. |
90481622 | 6486 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 6487 | */ |
33b8f84a | 6488 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 6489 | goto out_put_pages; |
17c9d12e MG |
6490 | |
6491 | /* | |
6492 | * Account for the reservations made. Shared mappings record regions | |
6493 | * that have reservations as they are shared by multiple VMAs. | |
6494 | * When the last VMA disappears, the region map says how much | |
6495 | * the reservation was and the page cache tells how much of | |
6496 | * the reservation was consumed. Private mappings are per-VMA and | |
6497 | * only the consumed reservations are tracked. When the VMA | |
6498 | * disappears, the original reservation is the VMA size and the | |
6499 | * consumed reservations are stored in the map. Hence, nothing | |
6500 | * else has to be done for private mappings here | |
6501 | */ | |
33039678 | 6502 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 6503 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
6504 | |
6505 | if (unlikely(add < 0)) { | |
6506 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 6507 | goto out_put_pages; |
0db9d74e | 6508 | } else if (unlikely(chg > add)) { |
33039678 MK |
6509 | /* |
6510 | * pages in this range were added to the reserve | |
6511 | * map between region_chg and region_add. This | |
6512 | * indicates a race with alloc_huge_page. Adjust | |
6513 | * the subpool and reserve counts modified above | |
6514 | * based on the difference. | |
6515 | */ | |
6516 | long rsv_adjust; | |
6517 | ||
d85aecf2 ML |
6518 | /* |
6519 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
6520 | * reference to h_cg->css. See comment below for detail. | |
6521 | */ | |
075a61d0 MA |
6522 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
6523 | hstate_index(h), | |
6524 | (chg - add) * pages_per_huge_page(h), h_cg); | |
6525 | ||
33039678 MK |
6526 | rsv_adjust = hugepage_subpool_put_pages(spool, |
6527 | chg - add); | |
6528 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
6529 | } else if (h_cg) { |
6530 | /* | |
6531 | * The file_regions will hold their own reference to | |
6532 | * h_cg->css. So we should release the reference held | |
6533 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
6534 | * done. | |
6535 | */ | |
6536 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
6537 | } |
6538 | } | |
33b8f84a MK |
6539 | return true; |
6540 | ||
075a61d0 MA |
6541 | out_put_pages: |
6542 | /* put back original number of pages, chg */ | |
6543 | (void)hugepage_subpool_put_pages(spool, chg); | |
6544 | out_uncharge_cgroup: | |
6545 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
6546 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 6547 | out_err: |
5e911373 | 6548 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
6549 | /* Only call region_abort if the region_chg succeeded but the |
6550 | * region_add failed or didn't run. | |
6551 | */ | |
6552 | if (chg >= 0 && add < 0) | |
6553 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
6554 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
6555 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 6556 | return false; |
a43a8c39 CK |
6557 | } |
6558 | ||
b5cec28d MK |
6559 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
6560 | long freed) | |
a43a8c39 | 6561 | { |
a5516438 | 6562 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 6563 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 6564 | long chg = 0; |
90481622 | 6565 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 6566 | long gbl_reserve; |
45c682a6 | 6567 | |
f27a5136 MK |
6568 | /* |
6569 | * Since this routine can be called in the evict inode path for all | |
6570 | * hugetlbfs inodes, resv_map could be NULL. | |
6571 | */ | |
b5cec28d MK |
6572 | if (resv_map) { |
6573 | chg = region_del(resv_map, start, end); | |
6574 | /* | |
6575 | * region_del() can fail in the rare case where a region | |
6576 | * must be split and another region descriptor can not be | |
6577 | * allocated. If end == LONG_MAX, it will not fail. | |
6578 | */ | |
6579 | if (chg < 0) | |
6580 | return chg; | |
6581 | } | |
6582 | ||
45c682a6 | 6583 | spin_lock(&inode->i_lock); |
e4c6f8be | 6584 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
6585 | spin_unlock(&inode->i_lock); |
6586 | ||
1c5ecae3 MK |
6587 | /* |
6588 | * If the subpool has a minimum size, the number of global | |
6589 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
6590 | * |
6591 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
6592 | * won't go negative. | |
1c5ecae3 MK |
6593 | */ |
6594 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
6595 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
6596 | |
6597 | return 0; | |
a43a8c39 | 6598 | } |
93f70f90 | 6599 | |
3212b535 SC |
6600 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
6601 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
6602 | struct vm_area_struct *vma, | |
6603 | unsigned long addr, pgoff_t idx) | |
6604 | { | |
6605 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
6606 | svma->vm_start; | |
6607 | unsigned long sbase = saddr & PUD_MASK; | |
6608 | unsigned long s_end = sbase + PUD_SIZE; | |
6609 | ||
6610 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
6611 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
6612 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
6613 | |
6614 | /* | |
6615 | * match the virtual addresses, permission and the alignment of the | |
6616 | * page table page. | |
6617 | */ | |
6618 | if (pmd_index(addr) != pmd_index(saddr) || | |
6619 | vm_flags != svm_flags || | |
07e51edf | 6620 | !range_in_vma(svma, sbase, s_end)) |
3212b535 SC |
6621 | return 0; |
6622 | ||
6623 | return saddr; | |
6624 | } | |
6625 | ||
31aafb45 | 6626 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
6627 | { |
6628 | unsigned long base = addr & PUD_MASK; | |
6629 | unsigned long end = base + PUD_SIZE; | |
6630 | ||
6631 | /* | |
6632 | * check on proper vm_flags and page table alignment | |
6633 | */ | |
017b1660 | 6634 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
6635 | return true; |
6636 | return false; | |
3212b535 SC |
6637 | } |
6638 | ||
c1991e07 PX |
6639 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
6640 | { | |
6641 | #ifdef CONFIG_USERFAULTFD | |
6642 | if (uffd_disable_huge_pmd_share(vma)) | |
6643 | return false; | |
6644 | #endif | |
6645 | return vma_shareable(vma, addr); | |
6646 | } | |
6647 | ||
017b1660 MK |
6648 | /* |
6649 | * Determine if start,end range within vma could be mapped by shared pmd. | |
6650 | * If yes, adjust start and end to cover range associated with possible | |
6651 | * shared pmd mappings. | |
6652 | */ | |
6653 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6654 | unsigned long *start, unsigned long *end) | |
6655 | { | |
a1ba9da8 LX |
6656 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
6657 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 6658 | |
a1ba9da8 | 6659 | /* |
f0953a1b IM |
6660 | * vma needs to span at least one aligned PUD size, and the range |
6661 | * must be at least partially within in. | |
a1ba9da8 LX |
6662 | */ |
6663 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
6664 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
6665 | return; |
6666 | ||
75802ca6 | 6667 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
6668 | if (*start > v_start) |
6669 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 6670 | |
a1ba9da8 LX |
6671 | if (*end < v_end) |
6672 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
6673 | } |
6674 | ||
3212b535 SC |
6675 | /* |
6676 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
6677 | * and returns the corresponding pte. While this is not necessary for the | |
6678 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
6679 | * code much cleaner. |
6680 | * | |
0bf7b64e MK |
6681 | * This routine must be called with i_mmap_rwsem held in at least read mode if |
6682 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
6683 | * table entries associated with the address space. This is important as we | |
6684 | * are setting up sharing based on existing page table entries (mappings). | |
3212b535 | 6685 | */ |
aec44e0f PX |
6686 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6687 | unsigned long addr, pud_t *pud) | |
3212b535 | 6688 | { |
3212b535 SC |
6689 | struct address_space *mapping = vma->vm_file->f_mapping; |
6690 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
6691 | vma->vm_pgoff; | |
6692 | struct vm_area_struct *svma; | |
6693 | unsigned long saddr; | |
6694 | pte_t *spte = NULL; | |
6695 | pte_t *pte; | |
cb900f41 | 6696 | spinlock_t *ptl; |
3212b535 | 6697 | |
0bf7b64e | 6698 | i_mmap_assert_locked(mapping); |
3212b535 SC |
6699 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
6700 | if (svma == vma) | |
6701 | continue; | |
6702 | ||
6703 | saddr = page_table_shareable(svma, vma, addr, idx); | |
6704 | if (saddr) { | |
7868a208 PA |
6705 | spte = huge_pte_offset(svma->vm_mm, saddr, |
6706 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
6707 | if (spte) { |
6708 | get_page(virt_to_page(spte)); | |
6709 | break; | |
6710 | } | |
6711 | } | |
6712 | } | |
6713 | ||
6714 | if (!spte) | |
6715 | goto out; | |
6716 | ||
8bea8052 | 6717 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 6718 | if (pud_none(*pud)) { |
3212b535 SC |
6719 | pud_populate(mm, pud, |
6720 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 6721 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 6722 | } else { |
3212b535 | 6723 | put_page(virt_to_page(spte)); |
dc6c9a35 | 6724 | } |
cb900f41 | 6725 | spin_unlock(ptl); |
3212b535 SC |
6726 | out: |
6727 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
6728 | return pte; |
6729 | } | |
6730 | ||
6731 | /* | |
6732 | * unmap huge page backed by shared pte. | |
6733 | * | |
6734 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
6735 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
6736 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
6737 | * | |
c0d0381a | 6738 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
6739 | * |
6740 | * returns: 1 successfully unmapped a shared pte page | |
6741 | * 0 the underlying pte page is not shared, or it is the last user | |
6742 | */ | |
34ae204f MK |
6743 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
6744 | unsigned long *addr, pte_t *ptep) | |
3212b535 SC |
6745 | { |
6746 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
6747 | p4d_t *p4d = p4d_offset(pgd, *addr); |
6748 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 | 6749 | |
34ae204f | 6750 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
6751 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
6752 | if (page_count(virt_to_page(ptep)) == 1) | |
6753 | return 0; | |
6754 | ||
6755 | pud_clear(pud); | |
6756 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 6757 | mm_dec_nr_pmds(mm); |
3212b535 SC |
6758 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
6759 | return 1; | |
6760 | } | |
c1991e07 | 6761 | |
9e5fc74c | 6762 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
aec44e0f PX |
6763 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6764 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
6765 | { |
6766 | return NULL; | |
6767 | } | |
e81f2d22 | 6768 | |
34ae204f MK |
6769 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
6770 | unsigned long *addr, pte_t *ptep) | |
e81f2d22 ZZ |
6771 | { |
6772 | return 0; | |
6773 | } | |
017b1660 MK |
6774 | |
6775 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6776 | unsigned long *start, unsigned long *end) | |
6777 | { | |
6778 | } | |
c1991e07 PX |
6779 | |
6780 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
6781 | { | |
6782 | return false; | |
6783 | } | |
3212b535 SC |
6784 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
6785 | ||
9e5fc74c | 6786 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 6787 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
6788 | unsigned long addr, unsigned long sz) |
6789 | { | |
6790 | pgd_t *pgd; | |
c2febafc | 6791 | p4d_t *p4d; |
9e5fc74c SC |
6792 | pud_t *pud; |
6793 | pte_t *pte = NULL; | |
6794 | ||
6795 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
6796 | p4d = p4d_alloc(mm, pgd, addr); |
6797 | if (!p4d) | |
6798 | return NULL; | |
c2febafc | 6799 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
6800 | if (pud) { |
6801 | if (sz == PUD_SIZE) { | |
6802 | pte = (pte_t *)pud; | |
6803 | } else { | |
6804 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 6805 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 6806 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
6807 | else |
6808 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
6809 | } | |
6810 | } | |
4e666314 | 6811 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
6812 | |
6813 | return pte; | |
6814 | } | |
6815 | ||
9b19df29 PA |
6816 | /* |
6817 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
6818 | * entry at address @addr | |
6819 | * | |
8ac0b81a LX |
6820 | * Return: Pointer to page table entry (PUD or PMD) for |
6821 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
6822 | * size @sz doesn't match the hugepage size at this level of the page |
6823 | * table. | |
6824 | */ | |
7868a208 PA |
6825 | pte_t *huge_pte_offset(struct mm_struct *mm, |
6826 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
6827 | { |
6828 | pgd_t *pgd; | |
c2febafc | 6829 | p4d_t *p4d; |
8ac0b81a LX |
6830 | pud_t *pud; |
6831 | pmd_t *pmd; | |
9e5fc74c SC |
6832 | |
6833 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
6834 | if (!pgd_present(*pgd)) |
6835 | return NULL; | |
6836 | p4d = p4d_offset(pgd, addr); | |
6837 | if (!p4d_present(*p4d)) | |
6838 | return NULL; | |
9b19df29 | 6839 | |
c2febafc | 6840 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
6841 | if (sz == PUD_SIZE) |
6842 | /* must be pud huge, non-present or none */ | |
c2febafc | 6843 | return (pte_t *)pud; |
8ac0b81a | 6844 | if (!pud_present(*pud)) |
9b19df29 | 6845 | return NULL; |
8ac0b81a | 6846 | /* must have a valid entry and size to go further */ |
9b19df29 | 6847 | |
8ac0b81a LX |
6848 | pmd = pmd_offset(pud, addr); |
6849 | /* must be pmd huge, non-present or none */ | |
6850 | return (pte_t *)pmd; | |
9e5fc74c SC |
6851 | } |
6852 | ||
61f77eda NH |
6853 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
6854 | ||
6855 | /* | |
6856 | * These functions are overwritable if your architecture needs its own | |
6857 | * behavior. | |
6858 | */ | |
6859 | struct page * __weak | |
6860 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
6861 | int write) | |
6862 | { | |
6863 | return ERR_PTR(-EINVAL); | |
6864 | } | |
6865 | ||
4dc71451 AK |
6866 | struct page * __weak |
6867 | follow_huge_pd(struct vm_area_struct *vma, | |
6868 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
6869 | { | |
6870 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
6871 | return NULL; | |
6872 | } | |
6873 | ||
61f77eda | 6874 | struct page * __weak |
9e5fc74c | 6875 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 6876 | pmd_t *pmd, int flags) |
9e5fc74c | 6877 | { |
e66f17ff NH |
6878 | struct page *page = NULL; |
6879 | spinlock_t *ptl; | |
c9d398fa | 6880 | pte_t pte; |
3faa52c0 | 6881 | |
8909691b DH |
6882 | /* |
6883 | * FOLL_PIN is not supported for follow_page(). Ordinary GUP goes via | |
6884 | * follow_hugetlb_page(). | |
6885 | */ | |
6886 | if (WARN_ON_ONCE(flags & FOLL_PIN)) | |
3faa52c0 JH |
6887 | return NULL; |
6888 | ||
e66f17ff NH |
6889 | retry: |
6890 | ptl = pmd_lockptr(mm, pmd); | |
6891 | spin_lock(ptl); | |
6892 | /* | |
6893 | * make sure that the address range covered by this pmd is not | |
6894 | * unmapped from other threads. | |
6895 | */ | |
6896 | if (!pmd_huge(*pmd)) | |
6897 | goto out; | |
c9d398fa NH |
6898 | pte = huge_ptep_get((pte_t *)pmd); |
6899 | if (pte_present(pte)) { | |
97534127 | 6900 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
6901 | /* |
6902 | * try_grab_page() should always succeed here, because: a) we | |
6903 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
6904 | * huge pmd (head) page is present in the page tables. The ptl | |
6905 | * prevents the head page and tail pages from being rearranged | |
6906 | * in any way. So this page must be available at this point, | |
6907 | * unless the page refcount overflowed: | |
6908 | */ | |
6909 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
6910 | page = NULL; | |
6911 | goto out; | |
6912 | } | |
e66f17ff | 6913 | } else { |
c9d398fa | 6914 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
6915 | spin_unlock(ptl); |
6916 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
6917 | goto retry; | |
6918 | } | |
6919 | /* | |
6920 | * hwpoisoned entry is treated as no_page_table in | |
6921 | * follow_page_mask(). | |
6922 | */ | |
6923 | } | |
6924 | out: | |
6925 | spin_unlock(ptl); | |
9e5fc74c SC |
6926 | return page; |
6927 | } | |
6928 | ||
61f77eda | 6929 | struct page * __weak |
9e5fc74c | 6930 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 6931 | pud_t *pud, int flags) |
9e5fc74c | 6932 | { |
3faa52c0 | 6933 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 6934 | return NULL; |
9e5fc74c | 6935 | |
e66f17ff | 6936 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
6937 | } |
6938 | ||
faaa5b62 AK |
6939 | struct page * __weak |
6940 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
6941 | { | |
3faa52c0 | 6942 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
6943 | return NULL; |
6944 | ||
6945 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
6946 | } | |
6947 | ||
31caf665 NH |
6948 | bool isolate_huge_page(struct page *page, struct list_head *list) |
6949 | { | |
bcc54222 NH |
6950 | bool ret = true; |
6951 | ||
db71ef79 | 6952 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
6953 | if (!PageHeadHuge(page) || |
6954 | !HPageMigratable(page) || | |
0eb2df2b | 6955 | !get_page_unless_zero(page)) { |
bcc54222 NH |
6956 | ret = false; |
6957 | goto unlock; | |
6958 | } | |
8f251a3d | 6959 | ClearHPageMigratable(page); |
31caf665 | 6960 | list_move_tail(&page->lru, list); |
bcc54222 | 6961 | unlock: |
db71ef79 | 6962 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 6963 | return ret; |
31caf665 NH |
6964 | } |
6965 | ||
25182f05 NH |
6966 | int get_hwpoison_huge_page(struct page *page, bool *hugetlb) |
6967 | { | |
6968 | int ret = 0; | |
6969 | ||
6970 | *hugetlb = false; | |
6971 | spin_lock_irq(&hugetlb_lock); | |
6972 | if (PageHeadHuge(page)) { | |
6973 | *hugetlb = true; | |
b283d983 NH |
6974 | if (HPageFreed(page)) |
6975 | ret = 0; | |
6976 | else if (HPageMigratable(page)) | |
25182f05 | 6977 | ret = get_page_unless_zero(page); |
0ed950d1 NH |
6978 | else |
6979 | ret = -EBUSY; | |
25182f05 NH |
6980 | } |
6981 | spin_unlock_irq(&hugetlb_lock); | |
6982 | return ret; | |
6983 | } | |
6984 | ||
405ce051 NH |
6985 | int get_huge_page_for_hwpoison(unsigned long pfn, int flags) |
6986 | { | |
6987 | int ret; | |
6988 | ||
6989 | spin_lock_irq(&hugetlb_lock); | |
6990 | ret = __get_huge_page_for_hwpoison(pfn, flags); | |
6991 | spin_unlock_irq(&hugetlb_lock); | |
6992 | return ret; | |
6993 | } | |
6994 | ||
31caf665 NH |
6995 | void putback_active_hugepage(struct page *page) |
6996 | { | |
db71ef79 | 6997 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 6998 | SetHPageMigratable(page); |
31caf665 | 6999 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 7000 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
7001 | put_page(page); |
7002 | } | |
ab5ac90a MH |
7003 | |
7004 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
7005 | { | |
7006 | struct hstate *h = page_hstate(oldpage); | |
7007 | ||
7008 | hugetlb_cgroup_migrate(oldpage, newpage); | |
7009 | set_page_owner_migrate_reason(newpage, reason); | |
7010 | ||
7011 | /* | |
7012 | * transfer temporary state of the new huge page. This is | |
7013 | * reverse to other transitions because the newpage is going to | |
7014 | * be final while the old one will be freed so it takes over | |
7015 | * the temporary status. | |
7016 | * | |
7017 | * Also note that we have to transfer the per-node surplus state | |
7018 | * here as well otherwise the global surplus count will not match | |
7019 | * the per-node's. | |
7020 | */ | |
9157c311 | 7021 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
7022 | int old_nid = page_to_nid(oldpage); |
7023 | int new_nid = page_to_nid(newpage); | |
7024 | ||
9157c311 MK |
7025 | SetHPageTemporary(oldpage); |
7026 | ClearHPageTemporary(newpage); | |
ab5ac90a | 7027 | |
5af1ab1d ML |
7028 | /* |
7029 | * There is no need to transfer the per-node surplus state | |
7030 | * when we do not cross the node. | |
7031 | */ | |
7032 | if (new_nid == old_nid) | |
7033 | return; | |
db71ef79 | 7034 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
7035 | if (h->surplus_huge_pages_node[old_nid]) { |
7036 | h->surplus_huge_pages_node[old_nid]--; | |
7037 | h->surplus_huge_pages_node[new_nid]++; | |
7038 | } | |
db71ef79 | 7039 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
7040 | } |
7041 | } | |
cf11e85f | 7042 | |
6dfeaff9 PX |
7043 | /* |
7044 | * This function will unconditionally remove all the shared pmd pgtable entries | |
7045 | * within the specific vma for a hugetlbfs memory range. | |
7046 | */ | |
7047 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
7048 | { | |
7049 | struct hstate *h = hstate_vma(vma); | |
7050 | unsigned long sz = huge_page_size(h); | |
7051 | struct mm_struct *mm = vma->vm_mm; | |
7052 | struct mmu_notifier_range range; | |
7053 | unsigned long address, start, end; | |
7054 | spinlock_t *ptl; | |
7055 | pte_t *ptep; | |
7056 | ||
7057 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
7058 | return; | |
7059 | ||
7060 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
7061 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
7062 | ||
7063 | if (start >= end) | |
7064 | return; | |
7065 | ||
9c8bbfac | 7066 | flush_cache_range(vma, start, end); |
6dfeaff9 PX |
7067 | /* |
7068 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
7069 | * we have already done the PUD_SIZE alignment. | |
7070 | */ | |
7071 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
7072 | start, end); | |
7073 | mmu_notifier_invalidate_range_start(&range); | |
7074 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
7075 | for (address = start; address < end; address += PUD_SIZE) { | |
7076 | unsigned long tmp = address; | |
7077 | ||
7078 | ptep = huge_pte_offset(mm, address, sz); | |
7079 | if (!ptep) | |
7080 | continue; | |
7081 | ptl = huge_pte_lock(h, mm, ptep); | |
7082 | /* We don't want 'address' to be changed */ | |
7083 | huge_pmd_unshare(mm, vma, &tmp, ptep); | |
7084 | spin_unlock(ptl); | |
7085 | } | |
7086 | flush_hugetlb_tlb_range(vma, start, end); | |
7087 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
7088 | /* | |
7089 | * No need to call mmu_notifier_invalidate_range(), see | |
7090 | * Documentation/vm/mmu_notifier.rst. | |
7091 | */ | |
7092 | mmu_notifier_invalidate_range_end(&range); | |
7093 | } | |
7094 | ||
cf11e85f | 7095 | #ifdef CONFIG_CMA |
cf11e85f RG |
7096 | static bool cma_reserve_called __initdata; |
7097 | ||
7098 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
7099 | { | |
38e719ab BW |
7100 | int nid, count = 0; |
7101 | unsigned long tmp; | |
7102 | char *s = p; | |
7103 | ||
7104 | while (*s) { | |
7105 | if (sscanf(s, "%lu%n", &tmp, &count) != 1) | |
7106 | break; | |
7107 | ||
7108 | if (s[count] == ':') { | |
f9317f77 | 7109 | if (tmp >= MAX_NUMNODES) |
38e719ab | 7110 | break; |
f9317f77 | 7111 | nid = array_index_nospec(tmp, MAX_NUMNODES); |
38e719ab BW |
7112 | |
7113 | s += count + 1; | |
7114 | tmp = memparse(s, &s); | |
7115 | hugetlb_cma_size_in_node[nid] = tmp; | |
7116 | hugetlb_cma_size += tmp; | |
7117 | ||
7118 | /* | |
7119 | * Skip the separator if have one, otherwise | |
7120 | * break the parsing. | |
7121 | */ | |
7122 | if (*s == ',') | |
7123 | s++; | |
7124 | else | |
7125 | break; | |
7126 | } else { | |
7127 | hugetlb_cma_size = memparse(p, &p); | |
7128 | break; | |
7129 | } | |
7130 | } | |
7131 | ||
cf11e85f RG |
7132 | return 0; |
7133 | } | |
7134 | ||
7135 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
7136 | ||
7137 | void __init hugetlb_cma_reserve(int order) | |
7138 | { | |
7139 | unsigned long size, reserved, per_node; | |
38e719ab | 7140 | bool node_specific_cma_alloc = false; |
cf11e85f RG |
7141 | int nid; |
7142 | ||
7143 | cma_reserve_called = true; | |
7144 | ||
38e719ab BW |
7145 | if (!hugetlb_cma_size) |
7146 | return; | |
7147 | ||
7148 | for (nid = 0; nid < MAX_NUMNODES; nid++) { | |
7149 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7150 | continue; | |
7151 | ||
30a51400 | 7152 | if (!node_online(nid)) { |
38e719ab BW |
7153 | pr_warn("hugetlb_cma: invalid node %d specified\n", nid); |
7154 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7155 | hugetlb_cma_size_in_node[nid] = 0; | |
7156 | continue; | |
7157 | } | |
7158 | ||
7159 | if (hugetlb_cma_size_in_node[nid] < (PAGE_SIZE << order)) { | |
7160 | pr_warn("hugetlb_cma: cma area of node %d should be at least %lu MiB\n", | |
7161 | nid, (PAGE_SIZE << order) / SZ_1M); | |
7162 | hugetlb_cma_size -= hugetlb_cma_size_in_node[nid]; | |
7163 | hugetlb_cma_size_in_node[nid] = 0; | |
7164 | } else { | |
7165 | node_specific_cma_alloc = true; | |
7166 | } | |
7167 | } | |
7168 | ||
7169 | /* Validate the CMA size again in case some invalid nodes specified. */ | |
cf11e85f RG |
7170 | if (!hugetlb_cma_size) |
7171 | return; | |
7172 | ||
7173 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
7174 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
7175 | (PAGE_SIZE << order) / SZ_1M); | |
a01f4390 | 7176 | hugetlb_cma_size = 0; |
cf11e85f RG |
7177 | return; |
7178 | } | |
7179 | ||
38e719ab BW |
7180 | if (!node_specific_cma_alloc) { |
7181 | /* | |
7182 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
7183 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
7184 | */ | |
7185 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
7186 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
7187 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
7188 | } | |
cf11e85f RG |
7189 | |
7190 | reserved = 0; | |
30a51400 | 7191 | for_each_online_node(nid) { |
cf11e85f | 7192 | int res; |
2281f797 | 7193 | char name[CMA_MAX_NAME]; |
cf11e85f | 7194 | |
38e719ab BW |
7195 | if (node_specific_cma_alloc) { |
7196 | if (hugetlb_cma_size_in_node[nid] == 0) | |
7197 | continue; | |
7198 | ||
7199 | size = hugetlb_cma_size_in_node[nid]; | |
7200 | } else { | |
7201 | size = min(per_node, hugetlb_cma_size - reserved); | |
7202 | } | |
7203 | ||
cf11e85f RG |
7204 | size = round_up(size, PAGE_SIZE << order); |
7205 | ||
2281f797 | 7206 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
a01f4390 MK |
7207 | /* |
7208 | * Note that 'order per bit' is based on smallest size that | |
7209 | * may be returned to CMA allocator in the case of | |
7210 | * huge page demotion. | |
7211 | */ | |
7212 | res = cma_declare_contiguous_nid(0, size, 0, | |
7213 | PAGE_SIZE << HUGETLB_PAGE_ORDER, | |
29d0f41d | 7214 | 0, false, name, |
cf11e85f RG |
7215 | &hugetlb_cma[nid], nid); |
7216 | if (res) { | |
7217 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
7218 | res, nid); | |
7219 | continue; | |
7220 | } | |
7221 | ||
7222 | reserved += size; | |
7223 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
7224 | size / SZ_1M, nid); | |
7225 | ||
7226 | if (reserved >= hugetlb_cma_size) | |
7227 | break; | |
7228 | } | |
a01f4390 MK |
7229 | |
7230 | if (!reserved) | |
7231 | /* | |
7232 | * hugetlb_cma_size is used to determine if allocations from | |
7233 | * cma are possible. Set to zero if no cma regions are set up. | |
7234 | */ | |
7235 | hugetlb_cma_size = 0; | |
cf11e85f RG |
7236 | } |
7237 | ||
7238 | void __init hugetlb_cma_check(void) | |
7239 | { | |
7240 | if (!hugetlb_cma_size || cma_reserve_called) | |
7241 | return; | |
7242 | ||
7243 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
7244 | } | |
7245 | ||
7246 | #endif /* CONFIG_CMA */ |