]>
Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * Generic hugetlb support. | |
6d49e352 | 4 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 5 | */ |
1da177e4 LT |
6 | #include <linux/list.h> |
7 | #include <linux/init.h> | |
1da177e4 | 8 | #include <linux/mm.h> |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
97ad1087 | 19 | #include <linux/memblock.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
bbe88753 | 22 | #include <linux/sched/mm.h> |
63489f8e | 23 | #include <linux/mmdebug.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
0fe6e20b | 25 | #include <linux/rmap.h> |
c6247f72 | 26 | #include <linux/string_helpers.h> |
fd6a03ed NH |
27 | #include <linux/swap.h> |
28 | #include <linux/swapops.h> | |
8382d914 | 29 | #include <linux/jhash.h> |
98fa15f3 | 30 | #include <linux/numa.h> |
c77c0a8a | 31 | #include <linux/llist.h> |
cf11e85f | 32 | #include <linux/cma.h> |
8cc5fcbb | 33 | #include <linux/migrate.h> |
d6606683 | 34 | |
63551ae0 | 35 | #include <asm/page.h> |
ca15ca40 | 36 | #include <asm/pgalloc.h> |
24669e58 | 37 | #include <asm/tlb.h> |
63551ae0 | 38 | |
24669e58 | 39 | #include <linux/io.h> |
63551ae0 | 40 | #include <linux/hugetlb.h> |
9dd540e2 | 41 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 42 | #include <linux/node.h> |
ab5ac90a | 43 | #include <linux/page_owner.h> |
7835e98b | 44 | #include "internal.h" |
f41f2ed4 | 45 | #include "hugetlb_vmemmap.h" |
1da177e4 | 46 | |
c3f38a38 | 47 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
48 | unsigned int default_hstate_idx; |
49 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 50 | |
dbda8fea | 51 | #ifdef CONFIG_CMA |
cf11e85f | 52 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
dbda8fea BS |
53 | #endif |
54 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 55 | |
641844f5 NH |
56 | /* |
57 | * Minimum page order among possible hugepage sizes, set to a proper value | |
58 | * at boot time. | |
59 | */ | |
60 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 61 | |
53ba51d2 JT |
62 | __initdata LIST_HEAD(huge_boot_pages); |
63 | ||
e5ff2159 AK |
64 | /* for command line parsing */ |
65 | static struct hstate * __initdata parsed_hstate; | |
66 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 67 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 68 | static bool __initdata parsed_default_hugepagesz; |
e5ff2159 | 69 | |
3935baa9 | 70 | /* |
31caf665 NH |
71 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
72 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 73 | */ |
c3f38a38 | 74 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 75 | |
8382d914 DB |
76 | /* |
77 | * Serializes faults on the same logical page. This is used to | |
78 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
79 | */ | |
80 | static int num_fault_mutexes; | |
c672c7f2 | 81 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 82 | |
7ca02d0a MK |
83 | /* Forward declaration */ |
84 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
85 | ||
1d88433b | 86 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 87 | { |
1d88433b ML |
88 | if (spool->count) |
89 | return false; | |
90 | if (spool->max_hpages != -1) | |
91 | return spool->used_hpages == 0; | |
92 | if (spool->min_hpages != -1) | |
93 | return spool->rsv_hpages == spool->min_hpages; | |
94 | ||
95 | return true; | |
96 | } | |
90481622 | 97 | |
db71ef79 MK |
98 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, |
99 | unsigned long irq_flags) | |
1d88433b | 100 | { |
db71ef79 | 101 | spin_unlock_irqrestore(&spool->lock, irq_flags); |
90481622 DG |
102 | |
103 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 104 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 105 | * free the subpool */ |
1d88433b | 106 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
107 | if (spool->min_hpages != -1) |
108 | hugetlb_acct_memory(spool->hstate, | |
109 | -spool->min_hpages); | |
90481622 | 110 | kfree(spool); |
7ca02d0a | 111 | } |
90481622 DG |
112 | } |
113 | ||
7ca02d0a MK |
114 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
115 | long min_hpages) | |
90481622 DG |
116 | { |
117 | struct hugepage_subpool *spool; | |
118 | ||
c6a91820 | 119 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
120 | if (!spool) |
121 | return NULL; | |
122 | ||
123 | spin_lock_init(&spool->lock); | |
124 | spool->count = 1; | |
7ca02d0a MK |
125 | spool->max_hpages = max_hpages; |
126 | spool->hstate = h; | |
127 | spool->min_hpages = min_hpages; | |
128 | ||
129 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
130 | kfree(spool); | |
131 | return NULL; | |
132 | } | |
133 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
134 | |
135 | return spool; | |
136 | } | |
137 | ||
138 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
139 | { | |
db71ef79 MK |
140 | unsigned long flags; |
141 | ||
142 | spin_lock_irqsave(&spool->lock, flags); | |
90481622 DG |
143 | BUG_ON(!spool->count); |
144 | spool->count--; | |
db71ef79 | 145 | unlock_or_release_subpool(spool, flags); |
90481622 DG |
146 | } |
147 | ||
1c5ecae3 MK |
148 | /* |
149 | * Subpool accounting for allocating and reserving pages. | |
150 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 151 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
152 | * global pools must be adjusted (upward). The returned value may |
153 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 154 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
155 | */ |
156 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
157 | long delta) |
158 | { | |
1c5ecae3 | 159 | long ret = delta; |
90481622 DG |
160 | |
161 | if (!spool) | |
1c5ecae3 | 162 | return ret; |
90481622 | 163 | |
db71ef79 | 164 | spin_lock_irq(&spool->lock); |
1c5ecae3 MK |
165 | |
166 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
167 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
168 | spool->used_hpages += delta; | |
169 | else { | |
170 | ret = -ENOMEM; | |
171 | goto unlock_ret; | |
172 | } | |
90481622 | 173 | } |
90481622 | 174 | |
09a95e29 MK |
175 | /* minimum size accounting */ |
176 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
177 | if (delta > spool->rsv_hpages) { |
178 | /* | |
179 | * Asking for more reserves than those already taken on | |
180 | * behalf of subpool. Return difference. | |
181 | */ | |
182 | ret = delta - spool->rsv_hpages; | |
183 | spool->rsv_hpages = 0; | |
184 | } else { | |
185 | ret = 0; /* reserves already accounted for */ | |
186 | spool->rsv_hpages -= delta; | |
187 | } | |
188 | } | |
189 | ||
190 | unlock_ret: | |
db71ef79 | 191 | spin_unlock_irq(&spool->lock); |
90481622 DG |
192 | return ret; |
193 | } | |
194 | ||
1c5ecae3 MK |
195 | /* |
196 | * Subpool accounting for freeing and unreserving pages. | |
197 | * Return the number of global page reservations that must be dropped. | |
198 | * The return value may only be different than the passed value (delta) | |
199 | * in the case where a subpool minimum size must be maintained. | |
200 | */ | |
201 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
202 | long delta) |
203 | { | |
1c5ecae3 | 204 | long ret = delta; |
db71ef79 | 205 | unsigned long flags; |
1c5ecae3 | 206 | |
90481622 | 207 | if (!spool) |
1c5ecae3 | 208 | return delta; |
90481622 | 209 | |
db71ef79 | 210 | spin_lock_irqsave(&spool->lock, flags); |
1c5ecae3 MK |
211 | |
212 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
213 | spool->used_hpages -= delta; | |
214 | ||
09a95e29 MK |
215 | /* minimum size accounting */ |
216 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
217 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
218 | ret = 0; | |
219 | else | |
220 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
221 | ||
222 | spool->rsv_hpages += delta; | |
223 | if (spool->rsv_hpages > spool->min_hpages) | |
224 | spool->rsv_hpages = spool->min_hpages; | |
225 | } | |
226 | ||
227 | /* | |
228 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
229 | * quota reference, free it now. | |
230 | */ | |
db71ef79 | 231 | unlock_or_release_subpool(spool, flags); |
1c5ecae3 MK |
232 | |
233 | return ret; | |
90481622 DG |
234 | } |
235 | ||
236 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
237 | { | |
238 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
239 | } | |
240 | ||
241 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
242 | { | |
496ad9aa | 243 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
244 | } |
245 | ||
0db9d74e MA |
246 | /* Helper that removes a struct file_region from the resv_map cache and returns |
247 | * it for use. | |
248 | */ | |
249 | static struct file_region * | |
250 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
251 | { | |
252 | struct file_region *nrg = NULL; | |
253 | ||
254 | VM_BUG_ON(resv->region_cache_count <= 0); | |
255 | ||
256 | resv->region_cache_count--; | |
257 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
258 | list_del(&nrg->link); |
259 | ||
260 | nrg->from = from; | |
261 | nrg->to = to; | |
262 | ||
263 | return nrg; | |
264 | } | |
265 | ||
075a61d0 MA |
266 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
267 | struct file_region *rg) | |
268 | { | |
269 | #ifdef CONFIG_CGROUP_HUGETLB | |
270 | nrg->reservation_counter = rg->reservation_counter; | |
271 | nrg->css = rg->css; | |
272 | if (rg->css) | |
273 | css_get(rg->css); | |
274 | #endif | |
275 | } | |
276 | ||
277 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
278 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
279 | struct hstate *h, | |
280 | struct resv_map *resv, | |
281 | struct file_region *nrg) | |
282 | { | |
283 | #ifdef CONFIG_CGROUP_HUGETLB | |
284 | if (h_cg) { | |
285 | nrg->reservation_counter = | |
286 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
287 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
288 | /* |
289 | * The caller will hold exactly one h_cg->css reference for the | |
290 | * whole contiguous reservation region. But this area might be | |
291 | * scattered when there are already some file_regions reside in | |
292 | * it. As a result, many file_regions may share only one css | |
293 | * reference. In order to ensure that one file_region must hold | |
294 | * exactly one h_cg->css reference, we should do css_get for | |
295 | * each file_region and leave the reference held by caller | |
296 | * untouched. | |
297 | */ | |
298 | css_get(&h_cg->css); | |
075a61d0 MA |
299 | if (!resv->pages_per_hpage) |
300 | resv->pages_per_hpage = pages_per_huge_page(h); | |
301 | /* pages_per_hpage should be the same for all entries in | |
302 | * a resv_map. | |
303 | */ | |
304 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
305 | } else { | |
306 | nrg->reservation_counter = NULL; | |
307 | nrg->css = NULL; | |
308 | } | |
309 | #endif | |
310 | } | |
311 | ||
d85aecf2 ML |
312 | static void put_uncharge_info(struct file_region *rg) |
313 | { | |
314 | #ifdef CONFIG_CGROUP_HUGETLB | |
315 | if (rg->css) | |
316 | css_put(rg->css); | |
317 | #endif | |
318 | } | |
319 | ||
a9b3f867 MA |
320 | static bool has_same_uncharge_info(struct file_region *rg, |
321 | struct file_region *org) | |
322 | { | |
323 | #ifdef CONFIG_CGROUP_HUGETLB | |
324 | return rg && org && | |
325 | rg->reservation_counter == org->reservation_counter && | |
326 | rg->css == org->css; | |
327 | ||
328 | #else | |
329 | return true; | |
330 | #endif | |
331 | } | |
332 | ||
333 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
334 | { | |
335 | struct file_region *nrg = NULL, *prg = NULL; | |
336 | ||
337 | prg = list_prev_entry(rg, link); | |
338 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
339 | has_same_uncharge_info(prg, rg)) { | |
340 | prg->to = rg->to; | |
341 | ||
342 | list_del(&rg->link); | |
d85aecf2 | 343 | put_uncharge_info(rg); |
a9b3f867 MA |
344 | kfree(rg); |
345 | ||
7db5e7b6 | 346 | rg = prg; |
a9b3f867 MA |
347 | } |
348 | ||
349 | nrg = list_next_entry(rg, link); | |
350 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
351 | has_same_uncharge_info(nrg, rg)) { | |
352 | nrg->from = rg->from; | |
353 | ||
354 | list_del(&rg->link); | |
d85aecf2 | 355 | put_uncharge_info(rg); |
a9b3f867 | 356 | kfree(rg); |
a9b3f867 MA |
357 | } |
358 | } | |
359 | ||
2103cf9c PX |
360 | static inline long |
361 | hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, | |
362 | long to, struct hstate *h, struct hugetlb_cgroup *cg, | |
363 | long *regions_needed) | |
364 | { | |
365 | struct file_region *nrg; | |
366 | ||
367 | if (!regions_needed) { | |
368 | nrg = get_file_region_entry_from_cache(map, from, to); | |
369 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
370 | list_add(&nrg->link, rg->link.prev); | |
371 | coalesce_file_region(map, nrg); | |
372 | } else | |
373 | *regions_needed += 1; | |
374 | ||
375 | return to - from; | |
376 | } | |
377 | ||
972a3da3 WY |
378 | /* |
379 | * Must be called with resv->lock held. | |
380 | * | |
381 | * Calling this with regions_needed != NULL will count the number of pages | |
382 | * to be added but will not modify the linked list. And regions_needed will | |
383 | * indicate the number of file_regions needed in the cache to carry out to add | |
384 | * the regions for this range. | |
d75c6af9 MA |
385 | */ |
386 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 387 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 388 | struct hstate *h, long *regions_needed) |
d75c6af9 | 389 | { |
0db9d74e | 390 | long add = 0; |
d75c6af9 | 391 | struct list_head *head = &resv->regions; |
0db9d74e | 392 | long last_accounted_offset = f; |
2103cf9c | 393 | struct file_region *rg = NULL, *trg = NULL; |
d75c6af9 | 394 | |
0db9d74e MA |
395 | if (regions_needed) |
396 | *regions_needed = 0; | |
d75c6af9 | 397 | |
0db9d74e MA |
398 | /* In this loop, we essentially handle an entry for the range |
399 | * [last_accounted_offset, rg->from), at every iteration, with some | |
400 | * bounds checking. | |
401 | */ | |
402 | list_for_each_entry_safe(rg, trg, head, link) { | |
403 | /* Skip irrelevant regions that start before our range. */ | |
404 | if (rg->from < f) { | |
405 | /* If this region ends after the last accounted offset, | |
406 | * then we need to update last_accounted_offset. | |
407 | */ | |
408 | if (rg->to > last_accounted_offset) | |
409 | last_accounted_offset = rg->to; | |
410 | continue; | |
411 | } | |
d75c6af9 | 412 | |
0db9d74e MA |
413 | /* When we find a region that starts beyond our range, we've |
414 | * finished. | |
415 | */ | |
ca7e0457 | 416 | if (rg->from >= t) |
d75c6af9 MA |
417 | break; |
418 | ||
0db9d74e MA |
419 | /* Add an entry for last_accounted_offset -> rg->from, and |
420 | * update last_accounted_offset. | |
421 | */ | |
2103cf9c PX |
422 | if (rg->from > last_accounted_offset) |
423 | add += hugetlb_resv_map_add(resv, rg, | |
424 | last_accounted_offset, | |
425 | rg->from, h, h_cg, | |
426 | regions_needed); | |
0db9d74e MA |
427 | |
428 | last_accounted_offset = rg->to; | |
429 | } | |
430 | ||
431 | /* Handle the case where our range extends beyond | |
432 | * last_accounted_offset. | |
433 | */ | |
2103cf9c PX |
434 | if (last_accounted_offset < t) |
435 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
436 | t, h, h_cg, regions_needed); | |
0db9d74e MA |
437 | |
438 | VM_BUG_ON(add < 0); | |
439 | return add; | |
440 | } | |
441 | ||
442 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
443 | */ | |
444 | static int allocate_file_region_entries(struct resv_map *resv, | |
445 | int regions_needed) | |
446 | __must_hold(&resv->lock) | |
447 | { | |
448 | struct list_head allocated_regions; | |
449 | int to_allocate = 0, i = 0; | |
450 | struct file_region *trg = NULL, *rg = NULL; | |
451 | ||
452 | VM_BUG_ON(regions_needed < 0); | |
453 | ||
454 | INIT_LIST_HEAD(&allocated_regions); | |
455 | ||
456 | /* | |
457 | * Check for sufficient descriptors in the cache to accommodate | |
458 | * the number of in progress add operations plus regions_needed. | |
459 | * | |
460 | * This is a while loop because when we drop the lock, some other call | |
461 | * to region_add or region_del may have consumed some region_entries, | |
462 | * so we keep looping here until we finally have enough entries for | |
463 | * (adds_in_progress + regions_needed). | |
464 | */ | |
465 | while (resv->region_cache_count < | |
466 | (resv->adds_in_progress + regions_needed)) { | |
467 | to_allocate = resv->adds_in_progress + regions_needed - | |
468 | resv->region_cache_count; | |
469 | ||
470 | /* At this point, we should have enough entries in the cache | |
f0953a1b | 471 | * for all the existing adds_in_progress. We should only be |
0db9d74e | 472 | * needing to allocate for regions_needed. |
d75c6af9 | 473 | */ |
0db9d74e MA |
474 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
475 | ||
476 | spin_unlock(&resv->lock); | |
477 | for (i = 0; i < to_allocate; i++) { | |
478 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
479 | if (!trg) | |
480 | goto out_of_memory; | |
481 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 482 | } |
d75c6af9 | 483 | |
0db9d74e MA |
484 | spin_lock(&resv->lock); |
485 | ||
d3ec7b6e WY |
486 | list_splice(&allocated_regions, &resv->region_cache); |
487 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
488 | } |
489 | ||
0db9d74e | 490 | return 0; |
d75c6af9 | 491 | |
0db9d74e MA |
492 | out_of_memory: |
493 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
494 | list_del(&rg->link); | |
495 | kfree(rg); | |
496 | } | |
497 | return -ENOMEM; | |
d75c6af9 MA |
498 | } |
499 | ||
1dd308a7 MK |
500 | /* |
501 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
502 | * map. Regions will be taken from the cache to fill in this range. |
503 | * Sufficient regions should exist in the cache due to the previous | |
504 | * call to region_chg with the same range, but in some cases the cache will not | |
505 | * have sufficient entries due to races with other code doing region_add or | |
506 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 507 | * |
0db9d74e MA |
508 | * regions_needed is the out value provided by a previous call to region_chg. |
509 | * | |
510 | * Return the number of new huge pages added to the map. This number is greater | |
511 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 512 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
513 | * region_add of regions of length 1 never allocate file_regions and cannot |
514 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
515 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 516 | */ |
0db9d74e | 517 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
518 | long in_regions_needed, struct hstate *h, |
519 | struct hugetlb_cgroup *h_cg) | |
96822904 | 520 | { |
0db9d74e | 521 | long add = 0, actual_regions_needed = 0; |
96822904 | 522 | |
7b24d861 | 523 | spin_lock(&resv->lock); |
0db9d74e MA |
524 | retry: |
525 | ||
526 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
527 | add_reservation_in_range(resv, f, t, NULL, NULL, |
528 | &actual_regions_needed); | |
96822904 | 529 | |
5e911373 | 530 | /* |
0db9d74e MA |
531 | * Check for sufficient descriptors in the cache to accommodate |
532 | * this add operation. Note that actual_regions_needed may be greater | |
533 | * than in_regions_needed, as the resv_map may have been modified since | |
534 | * the region_chg call. In this case, we need to make sure that we | |
535 | * allocate extra entries, such that we have enough for all the | |
536 | * existing adds_in_progress, plus the excess needed for this | |
537 | * operation. | |
5e911373 | 538 | */ |
0db9d74e MA |
539 | if (actual_regions_needed > in_regions_needed && |
540 | resv->region_cache_count < | |
541 | resv->adds_in_progress + | |
542 | (actual_regions_needed - in_regions_needed)) { | |
543 | /* region_add operation of range 1 should never need to | |
544 | * allocate file_region entries. | |
545 | */ | |
546 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 547 | |
0db9d74e MA |
548 | if (allocate_file_region_entries( |
549 | resv, actual_regions_needed - in_regions_needed)) { | |
550 | return -ENOMEM; | |
551 | } | |
5e911373 | 552 | |
0db9d74e | 553 | goto retry; |
5e911373 MK |
554 | } |
555 | ||
972a3da3 | 556 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
557 | |
558 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 559 | |
7b24d861 | 560 | spin_unlock(&resv->lock); |
cf3ad20b | 561 | return add; |
96822904 AW |
562 | } |
563 | ||
1dd308a7 MK |
564 | /* |
565 | * Examine the existing reserve map and determine how many | |
566 | * huge pages in the specified range [f, t) are NOT currently | |
567 | * represented. This routine is called before a subsequent | |
568 | * call to region_add that will actually modify the reserve | |
569 | * map to add the specified range [f, t). region_chg does | |
570 | * not change the number of huge pages represented by the | |
0db9d74e MA |
571 | * map. A number of new file_region structures is added to the cache as a |
572 | * placeholder, for the subsequent region_add call to use. At least 1 | |
573 | * file_region structure is added. | |
574 | * | |
575 | * out_regions_needed is the number of regions added to the | |
576 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
577 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
578 | * |
579 | * Returns the number of huge pages that need to be added to the existing | |
580 | * reservation map for the range [f, t). This number is greater or equal to | |
581 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
582 | * is needed and can not be allocated. | |
1dd308a7 | 583 | */ |
0db9d74e MA |
584 | static long region_chg(struct resv_map *resv, long f, long t, |
585 | long *out_regions_needed) | |
96822904 | 586 | { |
96822904 AW |
587 | long chg = 0; |
588 | ||
7b24d861 | 589 | spin_lock(&resv->lock); |
5e911373 | 590 | |
972a3da3 | 591 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 592 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 593 | out_regions_needed); |
5e911373 | 594 | |
0db9d74e MA |
595 | if (*out_regions_needed == 0) |
596 | *out_regions_needed = 1; | |
5e911373 | 597 | |
0db9d74e MA |
598 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
599 | return -ENOMEM; | |
5e911373 | 600 | |
0db9d74e | 601 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 602 | |
7b24d861 | 603 | spin_unlock(&resv->lock); |
96822904 AW |
604 | return chg; |
605 | } | |
606 | ||
5e911373 MK |
607 | /* |
608 | * Abort the in progress add operation. The adds_in_progress field | |
609 | * of the resv_map keeps track of the operations in progress between | |
610 | * calls to region_chg and region_add. Operations are sometimes | |
611 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
612 | * is called to decrement the adds_in_progress counter. regions_needed |
613 | * is the value returned by the region_chg call, it is used to decrement | |
614 | * the adds_in_progress counter. | |
5e911373 MK |
615 | * |
616 | * NOTE: The range arguments [f, t) are not needed or used in this | |
617 | * routine. They are kept to make reading the calling code easier as | |
618 | * arguments will match the associated region_chg call. | |
619 | */ | |
0db9d74e MA |
620 | static void region_abort(struct resv_map *resv, long f, long t, |
621 | long regions_needed) | |
5e911373 MK |
622 | { |
623 | spin_lock(&resv->lock); | |
624 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 625 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
626 | spin_unlock(&resv->lock); |
627 | } | |
628 | ||
1dd308a7 | 629 | /* |
feba16e2 MK |
630 | * Delete the specified range [f, t) from the reserve map. If the |
631 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
632 | * should be deleted. Locate the regions which intersect [f, t) | |
633 | * and either trim, delete or split the existing regions. | |
634 | * | |
635 | * Returns the number of huge pages deleted from the reserve map. | |
636 | * In the normal case, the return value is zero or more. In the | |
637 | * case where a region must be split, a new region descriptor must | |
638 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
639 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
640 | * a region and possibly return -ENOMEM. Callers specifying | |
641 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 642 | */ |
feba16e2 | 643 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 644 | { |
1406ec9b | 645 | struct list_head *head = &resv->regions; |
96822904 | 646 | struct file_region *rg, *trg; |
feba16e2 MK |
647 | struct file_region *nrg = NULL; |
648 | long del = 0; | |
96822904 | 649 | |
feba16e2 | 650 | retry: |
7b24d861 | 651 | spin_lock(&resv->lock); |
feba16e2 | 652 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
653 | /* |
654 | * Skip regions before the range to be deleted. file_region | |
655 | * ranges are normally of the form [from, to). However, there | |
656 | * may be a "placeholder" entry in the map which is of the form | |
657 | * (from, to) with from == to. Check for placeholder entries | |
658 | * at the beginning of the range to be deleted. | |
659 | */ | |
660 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 661 | continue; |
dbe409e4 | 662 | |
feba16e2 | 663 | if (rg->from >= t) |
96822904 | 664 | break; |
96822904 | 665 | |
feba16e2 MK |
666 | if (f > rg->from && t < rg->to) { /* Must split region */ |
667 | /* | |
668 | * Check for an entry in the cache before dropping | |
669 | * lock and attempting allocation. | |
670 | */ | |
671 | if (!nrg && | |
672 | resv->region_cache_count > resv->adds_in_progress) { | |
673 | nrg = list_first_entry(&resv->region_cache, | |
674 | struct file_region, | |
675 | link); | |
676 | list_del(&nrg->link); | |
677 | resv->region_cache_count--; | |
678 | } | |
96822904 | 679 | |
feba16e2 MK |
680 | if (!nrg) { |
681 | spin_unlock(&resv->lock); | |
682 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
683 | if (!nrg) | |
684 | return -ENOMEM; | |
685 | goto retry; | |
686 | } | |
687 | ||
688 | del += t - f; | |
79aa925b | 689 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 690 | resv, rg, t - f, false); |
feba16e2 MK |
691 | |
692 | /* New entry for end of split region */ | |
693 | nrg->from = t; | |
694 | nrg->to = rg->to; | |
075a61d0 MA |
695 | |
696 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
697 | ||
feba16e2 MK |
698 | INIT_LIST_HEAD(&nrg->link); |
699 | ||
700 | /* Original entry is trimmed */ | |
701 | rg->to = f; | |
702 | ||
703 | list_add(&nrg->link, &rg->link); | |
704 | nrg = NULL; | |
96822904 | 705 | break; |
feba16e2 MK |
706 | } |
707 | ||
708 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
709 | del += rg->to - rg->from; | |
075a61d0 | 710 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 711 | rg->to - rg->from, true); |
feba16e2 MK |
712 | list_del(&rg->link); |
713 | kfree(rg); | |
714 | continue; | |
715 | } | |
716 | ||
717 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 718 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 719 | t - rg->from, false); |
075a61d0 | 720 | |
79aa925b MK |
721 | del += t - rg->from; |
722 | rg->from = t; | |
723 | } else { /* Trim end of region */ | |
075a61d0 | 724 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 725 | rg->to - f, false); |
79aa925b MK |
726 | |
727 | del += rg->to - f; | |
728 | rg->to = f; | |
feba16e2 | 729 | } |
96822904 | 730 | } |
7b24d861 | 731 | |
7b24d861 | 732 | spin_unlock(&resv->lock); |
feba16e2 MK |
733 | kfree(nrg); |
734 | return del; | |
96822904 AW |
735 | } |
736 | ||
b5cec28d MK |
737 | /* |
738 | * A rare out of memory error was encountered which prevented removal of | |
739 | * the reserve map region for a page. The huge page itself was free'ed | |
740 | * and removed from the page cache. This routine will adjust the subpool | |
741 | * usage count, and the global reserve count if needed. By incrementing | |
742 | * these counts, the reserve map entry which could not be deleted will | |
743 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
744 | * counts. | |
745 | */ | |
72e2936c | 746 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
747 | { |
748 | struct hugepage_subpool *spool = subpool_inode(inode); | |
749 | long rsv_adjust; | |
da56388c | 750 | bool reserved = false; |
b5cec28d MK |
751 | |
752 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 753 | if (rsv_adjust > 0) { |
b5cec28d MK |
754 | struct hstate *h = hstate_inode(inode); |
755 | ||
da56388c ML |
756 | if (!hugetlb_acct_memory(h, 1)) |
757 | reserved = true; | |
758 | } else if (!rsv_adjust) { | |
759 | reserved = true; | |
b5cec28d | 760 | } |
da56388c ML |
761 | |
762 | if (!reserved) | |
763 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
764 | } |
765 | ||
1dd308a7 MK |
766 | /* |
767 | * Count and return the number of huge pages in the reserve map | |
768 | * that intersect with the range [f, t). | |
769 | */ | |
1406ec9b | 770 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 771 | { |
1406ec9b | 772 | struct list_head *head = &resv->regions; |
84afd99b AW |
773 | struct file_region *rg; |
774 | long chg = 0; | |
775 | ||
7b24d861 | 776 | spin_lock(&resv->lock); |
84afd99b AW |
777 | /* Locate each segment we overlap with, and count that overlap. */ |
778 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
779 | long seg_from; |
780 | long seg_to; | |
84afd99b AW |
781 | |
782 | if (rg->to <= f) | |
783 | continue; | |
784 | if (rg->from >= t) | |
785 | break; | |
786 | ||
787 | seg_from = max(rg->from, f); | |
788 | seg_to = min(rg->to, t); | |
789 | ||
790 | chg += seg_to - seg_from; | |
791 | } | |
7b24d861 | 792 | spin_unlock(&resv->lock); |
84afd99b AW |
793 | |
794 | return chg; | |
795 | } | |
796 | ||
e7c4b0bf AW |
797 | /* |
798 | * Convert the address within this vma to the page offset within | |
799 | * the mapping, in pagecache page units; huge pages here. | |
800 | */ | |
a5516438 AK |
801 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
802 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 803 | { |
a5516438 AK |
804 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
805 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
806 | } |
807 | ||
0fe6e20b NH |
808 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
809 | unsigned long address) | |
810 | { | |
811 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
812 | } | |
dee41079 | 813 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 814 | |
08fba699 MG |
815 | /* |
816 | * Return the size of the pages allocated when backing a VMA. In the majority | |
817 | * cases this will be same size as used by the page table entries. | |
818 | */ | |
819 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
820 | { | |
05ea8860 DW |
821 | if (vma->vm_ops && vma->vm_ops->pagesize) |
822 | return vma->vm_ops->pagesize(vma); | |
823 | return PAGE_SIZE; | |
08fba699 | 824 | } |
f340ca0f | 825 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 826 | |
3340289d MG |
827 | /* |
828 | * Return the page size being used by the MMU to back a VMA. In the majority | |
829 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
830 | * architectures where it differs, an architecture-specific 'strong' |
831 | * version of this symbol is required. | |
3340289d | 832 | */ |
09135cc5 | 833 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
834 | { |
835 | return vma_kernel_pagesize(vma); | |
836 | } | |
3340289d | 837 | |
84afd99b AW |
838 | /* |
839 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
840 | * bits of the reservation map pointer, which are always clear due to | |
841 | * alignment. | |
842 | */ | |
843 | #define HPAGE_RESV_OWNER (1UL << 0) | |
844 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 845 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 846 | |
a1e78772 MG |
847 | /* |
848 | * These helpers are used to track how many pages are reserved for | |
849 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
850 | * is guaranteed to have their future faults succeed. | |
851 | * | |
852 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
853 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
854 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
855 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
856 | * |
857 | * The private mapping reservation is represented in a subtly different | |
858 | * manner to a shared mapping. A shared mapping has a region map associated | |
859 | * with the underlying file, this region map represents the backing file | |
860 | * pages which have ever had a reservation assigned which this persists even | |
861 | * after the page is instantiated. A private mapping has a region map | |
862 | * associated with the original mmap which is attached to all VMAs which | |
863 | * reference it, this region map represents those offsets which have consumed | |
864 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 865 | */ |
e7c4b0bf AW |
866 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
867 | { | |
868 | return (unsigned long)vma->vm_private_data; | |
869 | } | |
870 | ||
871 | static void set_vma_private_data(struct vm_area_struct *vma, | |
872 | unsigned long value) | |
873 | { | |
874 | vma->vm_private_data = (void *)value; | |
875 | } | |
876 | ||
e9fe92ae MA |
877 | static void |
878 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
879 | struct hugetlb_cgroup *h_cg, | |
880 | struct hstate *h) | |
881 | { | |
882 | #ifdef CONFIG_CGROUP_HUGETLB | |
883 | if (!h_cg || !h) { | |
884 | resv_map->reservation_counter = NULL; | |
885 | resv_map->pages_per_hpage = 0; | |
886 | resv_map->css = NULL; | |
887 | } else { | |
888 | resv_map->reservation_counter = | |
889 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
890 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
891 | resv_map->css = &h_cg->css; | |
892 | } | |
893 | #endif | |
894 | } | |
895 | ||
9119a41e | 896 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
897 | { |
898 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
899 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
900 | ||
901 | if (!resv_map || !rg) { | |
902 | kfree(resv_map); | |
903 | kfree(rg); | |
84afd99b | 904 | return NULL; |
5e911373 | 905 | } |
84afd99b AW |
906 | |
907 | kref_init(&resv_map->refs); | |
7b24d861 | 908 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
909 | INIT_LIST_HEAD(&resv_map->regions); |
910 | ||
5e911373 | 911 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
912 | /* |
913 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
914 | * fields don't do cgroup accounting. On private mappings, these will be | |
915 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
916 | * reservations are to be un-charged from here. | |
917 | */ | |
918 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
919 | |
920 | INIT_LIST_HEAD(&resv_map->region_cache); | |
921 | list_add(&rg->link, &resv_map->region_cache); | |
922 | resv_map->region_cache_count = 1; | |
923 | ||
84afd99b AW |
924 | return resv_map; |
925 | } | |
926 | ||
9119a41e | 927 | void resv_map_release(struct kref *ref) |
84afd99b AW |
928 | { |
929 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
930 | struct list_head *head = &resv_map->region_cache; |
931 | struct file_region *rg, *trg; | |
84afd99b AW |
932 | |
933 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 934 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
935 | |
936 | /* ... and any entries left in the cache */ | |
937 | list_for_each_entry_safe(rg, trg, head, link) { | |
938 | list_del(&rg->link); | |
939 | kfree(rg); | |
940 | } | |
941 | ||
942 | VM_BUG_ON(resv_map->adds_in_progress); | |
943 | ||
84afd99b AW |
944 | kfree(resv_map); |
945 | } | |
946 | ||
4e35f483 JK |
947 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
948 | { | |
f27a5136 MK |
949 | /* |
950 | * At inode evict time, i_mapping may not point to the original | |
951 | * address space within the inode. This original address space | |
952 | * contains the pointer to the resv_map. So, always use the | |
953 | * address space embedded within the inode. | |
954 | * The VERY common case is inode->mapping == &inode->i_data but, | |
955 | * this may not be true for device special inodes. | |
956 | */ | |
957 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
958 | } |
959 | ||
84afd99b | 960 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 961 | { |
81d1b09c | 962 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
963 | if (vma->vm_flags & VM_MAYSHARE) { |
964 | struct address_space *mapping = vma->vm_file->f_mapping; | |
965 | struct inode *inode = mapping->host; | |
966 | ||
967 | return inode_resv_map(inode); | |
968 | ||
969 | } else { | |
84afd99b AW |
970 | return (struct resv_map *)(get_vma_private_data(vma) & |
971 | ~HPAGE_RESV_MASK); | |
4e35f483 | 972 | } |
a1e78772 MG |
973 | } |
974 | ||
84afd99b | 975 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 976 | { |
81d1b09c SL |
977 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
978 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 979 | |
84afd99b AW |
980 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
981 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
982 | } |
983 | ||
984 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
985 | { | |
81d1b09c SL |
986 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
987 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
988 | |
989 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
990 | } |
991 | ||
992 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
993 | { | |
81d1b09c | 994 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
995 | |
996 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
997 | } |
998 | ||
04f2cbe3 | 999 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
1000 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
1001 | { | |
81d1b09c | 1002 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 1003 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
1004 | vma->vm_private_data = (void *)0; |
1005 | } | |
1006 | ||
1007 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 1008 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1009 | { |
af0ed73e JK |
1010 | if (vma->vm_flags & VM_NORESERVE) { |
1011 | /* | |
1012 | * This address is already reserved by other process(chg == 0), | |
1013 | * so, we should decrement reserved count. Without decrementing, | |
1014 | * reserve count remains after releasing inode, because this | |
1015 | * allocated page will go into page cache and is regarded as | |
1016 | * coming from reserved pool in releasing step. Currently, we | |
1017 | * don't have any other solution to deal with this situation | |
1018 | * properly, so add work-around here. | |
1019 | */ | |
1020 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1021 | return true; |
af0ed73e | 1022 | else |
559ec2f8 | 1023 | return false; |
af0ed73e | 1024 | } |
a63884e9 JK |
1025 | |
1026 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1027 | if (vma->vm_flags & VM_MAYSHARE) { |
1028 | /* | |
1029 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1030 | * be a region map for all pages. The only situation where | |
1031 | * there is no region map is if a hole was punched via | |
7c8de358 | 1032 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1033 | * use. This situation is indicated if chg != 0. |
1034 | */ | |
1035 | if (chg) | |
1036 | return false; | |
1037 | else | |
1038 | return true; | |
1039 | } | |
a63884e9 JK |
1040 | |
1041 | /* | |
1042 | * Only the process that called mmap() has reserves for | |
1043 | * private mappings. | |
1044 | */ | |
67961f9d MK |
1045 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1046 | /* | |
1047 | * Like the shared case above, a hole punch or truncate | |
1048 | * could have been performed on the private mapping. | |
1049 | * Examine the value of chg to determine if reserves | |
1050 | * actually exist or were previously consumed. | |
1051 | * Very Subtle - The value of chg comes from a previous | |
1052 | * call to vma_needs_reserves(). The reserve map for | |
1053 | * private mappings has different (opposite) semantics | |
1054 | * than that of shared mappings. vma_needs_reserves() | |
1055 | * has already taken this difference in semantics into | |
1056 | * account. Therefore, the meaning of chg is the same | |
1057 | * as in the shared case above. Code could easily be | |
1058 | * combined, but keeping it separate draws attention to | |
1059 | * subtle differences. | |
1060 | */ | |
1061 | if (chg) | |
1062 | return false; | |
1063 | else | |
1064 | return true; | |
1065 | } | |
a63884e9 | 1066 | |
559ec2f8 | 1067 | return false; |
a1e78772 MG |
1068 | } |
1069 | ||
a5516438 | 1070 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1071 | { |
1072 | int nid = page_to_nid(page); | |
9487ca60 MK |
1073 | |
1074 | lockdep_assert_held(&hugetlb_lock); | |
b65a4eda MK |
1075 | VM_BUG_ON_PAGE(page_count(page), page); |
1076 | ||
0edaecfa | 1077 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1078 | h->free_huge_pages++; |
1079 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1080 | SetHPageFreed(page); |
1da177e4 LT |
1081 | } |
1082 | ||
94310cbc | 1083 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1084 | { |
1085 | struct page *page; | |
1a08ae36 | 1086 | bool pin = !!(current->flags & PF_MEMALLOC_PIN); |
bbe88753 | 1087 | |
9487ca60 | 1088 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 | 1089 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
8e3560d9 | 1090 | if (pin && !is_pinnable_page(page)) |
bbe88753 | 1091 | continue; |
bf50bab2 | 1092 | |
6664bfc8 WY |
1093 | if (PageHWPoison(page)) |
1094 | continue; | |
1095 | ||
1096 | list_move(&page->lru, &h->hugepage_activelist); | |
1097 | set_page_refcounted(page); | |
6c037149 | 1098 | ClearHPageFreed(page); |
6664bfc8 WY |
1099 | h->free_huge_pages--; |
1100 | h->free_huge_pages_node[nid]--; | |
1101 | return page; | |
bbe88753 JK |
1102 | } |
1103 | ||
6664bfc8 | 1104 | return NULL; |
bf50bab2 NH |
1105 | } |
1106 | ||
3e59fcb0 MH |
1107 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1108 | nodemask_t *nmask) | |
94310cbc | 1109 | { |
3e59fcb0 MH |
1110 | unsigned int cpuset_mems_cookie; |
1111 | struct zonelist *zonelist; | |
1112 | struct zone *zone; | |
1113 | struct zoneref *z; | |
98fa15f3 | 1114 | int node = NUMA_NO_NODE; |
94310cbc | 1115 | |
3e59fcb0 MH |
1116 | zonelist = node_zonelist(nid, gfp_mask); |
1117 | ||
1118 | retry_cpuset: | |
1119 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1120 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1121 | struct page *page; | |
1122 | ||
1123 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1124 | continue; | |
1125 | /* | |
1126 | * no need to ask again on the same node. Pool is node rather than | |
1127 | * zone aware | |
1128 | */ | |
1129 | if (zone_to_nid(zone) == node) | |
1130 | continue; | |
1131 | node = zone_to_nid(zone); | |
94310cbc | 1132 | |
94310cbc AK |
1133 | page = dequeue_huge_page_node_exact(h, node); |
1134 | if (page) | |
1135 | return page; | |
1136 | } | |
3e59fcb0 MH |
1137 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1138 | goto retry_cpuset; | |
1139 | ||
94310cbc AK |
1140 | return NULL; |
1141 | } | |
1142 | ||
a5516438 AK |
1143 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1144 | struct vm_area_struct *vma, | |
af0ed73e JK |
1145 | unsigned long address, int avoid_reserve, |
1146 | long chg) | |
1da177e4 | 1147 | { |
3e59fcb0 | 1148 | struct page *page; |
480eccf9 | 1149 | struct mempolicy *mpol; |
04ec6264 | 1150 | gfp_t gfp_mask; |
3e59fcb0 | 1151 | nodemask_t *nodemask; |
04ec6264 | 1152 | int nid; |
1da177e4 | 1153 | |
a1e78772 MG |
1154 | /* |
1155 | * A child process with MAP_PRIVATE mappings created by their parent | |
1156 | * have no page reserves. This check ensures that reservations are | |
1157 | * not "stolen". The child may still get SIGKILLed | |
1158 | */ | |
af0ed73e | 1159 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1160 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1161 | goto err; |
a1e78772 | 1162 | |
04f2cbe3 | 1163 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1164 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1165 | goto err; |
04f2cbe3 | 1166 | |
04ec6264 VB |
1167 | gfp_mask = htlb_alloc_mask(h); |
1168 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
1169 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
1170 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
d6995da3 | 1171 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1172 | h->resv_huge_pages--; |
1da177e4 | 1173 | } |
cc9a6c87 | 1174 | |
52cd3b07 | 1175 | mpol_cond_put(mpol); |
1da177e4 | 1176 | return page; |
cc9a6c87 MG |
1177 | |
1178 | err: | |
cc9a6c87 | 1179 | return NULL; |
1da177e4 LT |
1180 | } |
1181 | ||
1cac6f2c LC |
1182 | /* |
1183 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1184 | * We may have allocated or freed a huge page based on a different | |
1185 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1186 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1187 | * node for alloc or free. | |
1188 | */ | |
1189 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1190 | { | |
0edaf86c | 1191 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1192 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1193 | ||
1194 | return nid; | |
1195 | } | |
1196 | ||
1197 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1198 | { | |
1199 | if (!node_isset(nid, *nodes_allowed)) | |
1200 | nid = next_node_allowed(nid, nodes_allowed); | |
1201 | return nid; | |
1202 | } | |
1203 | ||
1204 | /* | |
1205 | * returns the previously saved node ["this node"] from which to | |
1206 | * allocate a persistent huge page for the pool and advance the | |
1207 | * next node from which to allocate, handling wrap at end of node | |
1208 | * mask. | |
1209 | */ | |
1210 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1211 | nodemask_t *nodes_allowed) | |
1212 | { | |
1213 | int nid; | |
1214 | ||
1215 | VM_BUG_ON(!nodes_allowed); | |
1216 | ||
1217 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1218 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1219 | ||
1220 | return nid; | |
1221 | } | |
1222 | ||
1223 | /* | |
10c6ec49 | 1224 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1225 | * node ["this node"] from which to free a huge page. Advance the |
1226 | * next node id whether or not we find a free huge page to free so | |
1227 | * that the next attempt to free addresses the next node. | |
1228 | */ | |
1229 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1230 | { | |
1231 | int nid; | |
1232 | ||
1233 | VM_BUG_ON(!nodes_allowed); | |
1234 | ||
1235 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1236 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1237 | ||
1238 | return nid; | |
1239 | } | |
1240 | ||
1241 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1242 | for (nr_nodes = nodes_weight(*mask); \ | |
1243 | nr_nodes > 0 && \ | |
1244 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1245 | nr_nodes--) | |
1246 | ||
1247 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1248 | for (nr_nodes = nodes_weight(*mask); \ | |
1249 | nr_nodes > 0 && \ | |
1250 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1251 | nr_nodes--) | |
1252 | ||
e1073d1e | 1253 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1254 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1255 | unsigned int order) |
944d9fec LC |
1256 | { |
1257 | int i; | |
1258 | int nr_pages = 1 << order; | |
1259 | struct page *p = page + 1; | |
1260 | ||
c8cc708a | 1261 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1262 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1263 | |
944d9fec | 1264 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1265 | clear_compound_head(p); |
944d9fec | 1266 | set_page_refcounted(p); |
944d9fec LC |
1267 | } |
1268 | ||
1269 | set_compound_order(page, 0); | |
ba9c1201 | 1270 | page[1].compound_nr = 0; |
944d9fec LC |
1271 | __ClearPageHead(page); |
1272 | } | |
1273 | ||
d00181b9 | 1274 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1275 | { |
cf11e85f RG |
1276 | /* |
1277 | * If the page isn't allocated using the cma allocator, | |
1278 | * cma_release() returns false. | |
1279 | */ | |
dbda8fea BS |
1280 | #ifdef CONFIG_CMA |
1281 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1282 | return; |
dbda8fea | 1283 | #endif |
cf11e85f | 1284 | |
944d9fec LC |
1285 | free_contig_range(page_to_pfn(page), 1 << order); |
1286 | } | |
1287 | ||
4eb0716e | 1288 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1289 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1290 | int nid, nodemask_t *nodemask) | |
944d9fec | 1291 | { |
04adbc3f | 1292 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1293 | if (nid == NUMA_NO_NODE) |
1294 | nid = numa_mem_id(); | |
944d9fec | 1295 | |
dbda8fea BS |
1296 | #ifdef CONFIG_CMA |
1297 | { | |
cf11e85f RG |
1298 | struct page *page; |
1299 | int node; | |
1300 | ||
953f064a LX |
1301 | if (hugetlb_cma[nid]) { |
1302 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1303 | huge_page_order(h), true); | |
cf11e85f RG |
1304 | if (page) |
1305 | return page; | |
1306 | } | |
953f064a LX |
1307 | |
1308 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1309 | for_each_node_mask(node, *nodemask) { | |
1310 | if (node == nid || !hugetlb_cma[node]) | |
1311 | continue; | |
1312 | ||
1313 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1314 | huge_page_order(h), true); | |
1315 | if (page) | |
1316 | return page; | |
1317 | } | |
1318 | } | |
cf11e85f | 1319 | } |
dbda8fea | 1320 | #endif |
cf11e85f | 1321 | |
5e27a2df | 1322 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1323 | } |
1324 | ||
4eb0716e AG |
1325 | #else /* !CONFIG_CONTIG_ALLOC */ |
1326 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1327 | int nid, nodemask_t *nodemask) | |
1328 | { | |
1329 | return NULL; | |
1330 | } | |
1331 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1332 | |
e1073d1e | 1333 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1334 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1335 | int nid, nodemask_t *nodemask) |
1336 | { | |
1337 | return NULL; | |
1338 | } | |
d00181b9 | 1339 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1340 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1341 | unsigned int order) { } |
944d9fec LC |
1342 | #endif |
1343 | ||
6eb4e88a MK |
1344 | /* |
1345 | * Remove hugetlb page from lists, and update dtor so that page appears | |
1346 | * as just a compound page. A reference is held on the page. | |
1347 | * | |
1348 | * Must be called with hugetlb lock held. | |
1349 | */ | |
1350 | static void remove_hugetlb_page(struct hstate *h, struct page *page, | |
1351 | bool adjust_surplus) | |
1352 | { | |
1353 | int nid = page_to_nid(page); | |
1354 | ||
1355 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1356 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1357 | ||
9487ca60 | 1358 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1359 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1360 | return; | |
1361 | ||
1362 | list_del(&page->lru); | |
1363 | ||
1364 | if (HPageFreed(page)) { | |
1365 | h->free_huge_pages--; | |
1366 | h->free_huge_pages_node[nid]--; | |
1367 | } | |
1368 | if (adjust_surplus) { | |
1369 | h->surplus_huge_pages--; | |
1370 | h->surplus_huge_pages_node[nid]--; | |
1371 | } | |
1372 | ||
e32d20c0 MK |
1373 | /* |
1374 | * Very subtle | |
1375 | * | |
1376 | * For non-gigantic pages set the destructor to the normal compound | |
1377 | * page dtor. This is needed in case someone takes an additional | |
1378 | * temporary ref to the page, and freeing is delayed until they drop | |
1379 | * their reference. | |
1380 | * | |
1381 | * For gigantic pages set the destructor to the null dtor. This | |
1382 | * destructor will never be called. Before freeing the gigantic | |
1383 | * page destroy_compound_gigantic_page will turn the compound page | |
1384 | * into a simple group of pages. After this the destructor does not | |
1385 | * apply. | |
1386 | * | |
1387 | * This handles the case where more than one ref is held when and | |
1388 | * after update_and_free_page is called. | |
1389 | */ | |
6eb4e88a | 1390 | set_page_refcounted(page); |
e32d20c0 MK |
1391 | if (hstate_is_gigantic(h)) |
1392 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1393 | else | |
1394 | set_compound_page_dtor(page, COMPOUND_PAGE_DTOR); | |
6eb4e88a MK |
1395 | |
1396 | h->nr_huge_pages--; | |
1397 | h->nr_huge_pages_node[nid]--; | |
1398 | } | |
1399 | ||
ad2fa371 MS |
1400 | static void add_hugetlb_page(struct hstate *h, struct page *page, |
1401 | bool adjust_surplus) | |
1402 | { | |
1403 | int zeroed; | |
1404 | int nid = page_to_nid(page); | |
1405 | ||
1406 | VM_BUG_ON_PAGE(!HPageVmemmapOptimized(page), page); | |
1407 | ||
1408 | lockdep_assert_held(&hugetlb_lock); | |
1409 | ||
1410 | INIT_LIST_HEAD(&page->lru); | |
1411 | h->nr_huge_pages++; | |
1412 | h->nr_huge_pages_node[nid]++; | |
1413 | ||
1414 | if (adjust_surplus) { | |
1415 | h->surplus_huge_pages++; | |
1416 | h->surplus_huge_pages_node[nid]++; | |
1417 | } | |
1418 | ||
1419 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); | |
1420 | set_page_private(page, 0); | |
1421 | SetHPageVmemmapOptimized(page); | |
1422 | ||
1423 | /* | |
b65a4eda MK |
1424 | * This page is about to be managed by the hugetlb allocator and |
1425 | * should have no users. Drop our reference, and check for others | |
1426 | * just in case. | |
ad2fa371 MS |
1427 | */ |
1428 | zeroed = put_page_testzero(page); | |
b65a4eda MK |
1429 | if (!zeroed) |
1430 | /* | |
1431 | * It is VERY unlikely soneone else has taken a ref on | |
1432 | * the page. In this case, we simply return as the | |
1433 | * hugetlb destructor (free_huge_page) will be called | |
1434 | * when this other ref is dropped. | |
1435 | */ | |
1436 | return; | |
1437 | ||
ad2fa371 MS |
1438 | arch_clear_hugepage_flags(page); |
1439 | enqueue_huge_page(h, page); | |
1440 | } | |
1441 | ||
b65d4adb | 1442 | static void __update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1443 | { |
1444 | int i; | |
dbfee5ae | 1445 | struct page *subpage = page; |
a5516438 | 1446 | |
4eb0716e | 1447 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1448 | return; |
18229df5 | 1449 | |
ad2fa371 MS |
1450 | if (alloc_huge_page_vmemmap(h, page)) { |
1451 | spin_lock_irq(&hugetlb_lock); | |
1452 | /* | |
1453 | * If we cannot allocate vmemmap pages, just refuse to free the | |
1454 | * page and put the page back on the hugetlb free list and treat | |
1455 | * as a surplus page. | |
1456 | */ | |
1457 | add_hugetlb_page(h, page, true); | |
1458 | spin_unlock_irq(&hugetlb_lock); | |
1459 | return; | |
1460 | } | |
1461 | ||
dbfee5ae MK |
1462 | for (i = 0; i < pages_per_huge_page(h); |
1463 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1464 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1465 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1466 | 1 << PG_active | 1 << PG_private | |
1467 | 1 << PG_writeback); | |
6af2acb6 | 1468 | } |
944d9fec LC |
1469 | if (hstate_is_gigantic(h)) { |
1470 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1471 | free_gigantic_page(page, huge_page_order(h)); | |
1472 | } else { | |
944d9fec LC |
1473 | __free_pages(page, huge_page_order(h)); |
1474 | } | |
6af2acb6 AL |
1475 | } |
1476 | ||
b65d4adb MS |
1477 | /* |
1478 | * As update_and_free_page() can be called under any context, so we cannot | |
1479 | * use GFP_KERNEL to allocate vmemmap pages. However, we can defer the | |
1480 | * actual freeing in a workqueue to prevent from using GFP_ATOMIC to allocate | |
1481 | * the vmemmap pages. | |
1482 | * | |
1483 | * free_hpage_workfn() locklessly retrieves the linked list of pages to be | |
1484 | * freed and frees them one-by-one. As the page->mapping pointer is going | |
1485 | * to be cleared in free_hpage_workfn() anyway, it is reused as the llist_node | |
1486 | * structure of a lockless linked list of huge pages to be freed. | |
1487 | */ | |
1488 | static LLIST_HEAD(hpage_freelist); | |
1489 | ||
1490 | static void free_hpage_workfn(struct work_struct *work) | |
1491 | { | |
1492 | struct llist_node *node; | |
1493 | ||
1494 | node = llist_del_all(&hpage_freelist); | |
1495 | ||
1496 | while (node) { | |
1497 | struct page *page; | |
1498 | struct hstate *h; | |
1499 | ||
1500 | page = container_of((struct address_space **)node, | |
1501 | struct page, mapping); | |
1502 | node = node->next; | |
1503 | page->mapping = NULL; | |
1504 | /* | |
1505 | * The VM_BUG_ON_PAGE(!PageHuge(page), page) in page_hstate() | |
1506 | * is going to trigger because a previous call to | |
1507 | * remove_hugetlb_page() will set_compound_page_dtor(page, | |
1508 | * NULL_COMPOUND_DTOR), so do not use page_hstate() directly. | |
1509 | */ | |
1510 | h = size_to_hstate(page_size(page)); | |
1511 | ||
1512 | __update_and_free_page(h, page); | |
1513 | ||
1514 | cond_resched(); | |
1515 | } | |
1516 | } | |
1517 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1518 | ||
1519 | static inline void flush_free_hpage_work(struct hstate *h) | |
1520 | { | |
1521 | if (free_vmemmap_pages_per_hpage(h)) | |
1522 | flush_work(&free_hpage_work); | |
1523 | } | |
1524 | ||
1525 | static void update_and_free_page(struct hstate *h, struct page *page, | |
1526 | bool atomic) | |
1527 | { | |
ad2fa371 | 1528 | if (!HPageVmemmapOptimized(page) || !atomic) { |
b65d4adb MS |
1529 | __update_and_free_page(h, page); |
1530 | return; | |
1531 | } | |
1532 | ||
1533 | /* | |
1534 | * Defer freeing to avoid using GFP_ATOMIC to allocate vmemmap pages. | |
1535 | * | |
1536 | * Only call schedule_work() if hpage_freelist is previously | |
1537 | * empty. Otherwise, schedule_work() had been called but the workfn | |
1538 | * hasn't retrieved the list yet. | |
1539 | */ | |
1540 | if (llist_add((struct llist_node *)&page->mapping, &hpage_freelist)) | |
1541 | schedule_work(&free_hpage_work); | |
1542 | } | |
1543 | ||
10c6ec49 MK |
1544 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1545 | { | |
1546 | struct page *page, *t_page; | |
1547 | ||
1548 | list_for_each_entry_safe(page, t_page, list, lru) { | |
b65d4adb | 1549 | update_and_free_page(h, page, false); |
10c6ec49 MK |
1550 | cond_resched(); |
1551 | } | |
1552 | } | |
1553 | ||
e5ff2159 AK |
1554 | struct hstate *size_to_hstate(unsigned long size) |
1555 | { | |
1556 | struct hstate *h; | |
1557 | ||
1558 | for_each_hstate(h) { | |
1559 | if (huge_page_size(h) == size) | |
1560 | return h; | |
1561 | } | |
1562 | return NULL; | |
1563 | } | |
1564 | ||
db71ef79 | 1565 | void free_huge_page(struct page *page) |
27a85ef1 | 1566 | { |
a5516438 AK |
1567 | /* |
1568 | * Can't pass hstate in here because it is called from the | |
1569 | * compound page destructor. | |
1570 | */ | |
e5ff2159 | 1571 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1572 | int nid = page_to_nid(page); |
d6995da3 | 1573 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1574 | bool restore_reserve; |
db71ef79 | 1575 | unsigned long flags; |
27a85ef1 | 1576 | |
b4330afb MK |
1577 | VM_BUG_ON_PAGE(page_count(page), page); |
1578 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1579 | |
d6995da3 | 1580 | hugetlb_set_page_subpool(page, NULL); |
8ace22bc | 1581 | page->mapping = NULL; |
d6995da3 MK |
1582 | restore_reserve = HPageRestoreReserve(page); |
1583 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1584 | |
1c5ecae3 | 1585 | /* |
d6995da3 | 1586 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1587 | * reservation. If the page was associated with a subpool, there |
1588 | * would have been a page reserved in the subpool before allocation | |
1589 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1590 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1591 | * remove the reserved page from the subpool. |
1c5ecae3 | 1592 | */ |
0919e1b6 MK |
1593 | if (!restore_reserve) { |
1594 | /* | |
1595 | * A return code of zero implies that the subpool will be | |
1596 | * under its minimum size if the reservation is not restored | |
1597 | * after page is free. Therefore, force restore_reserve | |
1598 | * operation. | |
1599 | */ | |
1600 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1601 | restore_reserve = true; | |
1602 | } | |
1c5ecae3 | 1603 | |
db71ef79 | 1604 | spin_lock_irqsave(&hugetlb_lock, flags); |
8f251a3d | 1605 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1606 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1607 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1608 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1609 | pages_per_huge_page(h), page); | |
07443a85 JK |
1610 | if (restore_reserve) |
1611 | h->resv_huge_pages++; | |
1612 | ||
9157c311 | 1613 | if (HPageTemporary(page)) { |
6eb4e88a | 1614 | remove_hugetlb_page(h, page, false); |
db71ef79 | 1615 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1616 | update_and_free_page(h, page, true); |
ab5ac90a | 1617 | } else if (h->surplus_huge_pages_node[nid]) { |
0edaecfa | 1618 | /* remove the page from active list */ |
6eb4e88a | 1619 | remove_hugetlb_page(h, page, true); |
db71ef79 | 1620 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
b65d4adb | 1621 | update_and_free_page(h, page, true); |
7893d1d5 | 1622 | } else { |
5d3a551c | 1623 | arch_clear_hugepage_flags(page); |
a5516438 | 1624 | enqueue_huge_page(h, page); |
db71ef79 | 1625 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1626 | } |
c77c0a8a WL |
1627 | } |
1628 | ||
d3d99fcc OS |
1629 | /* |
1630 | * Must be called with the hugetlb lock held | |
1631 | */ | |
1632 | static void __prep_account_new_huge_page(struct hstate *h, int nid) | |
1633 | { | |
1634 | lockdep_assert_held(&hugetlb_lock); | |
1635 | h->nr_huge_pages++; | |
1636 | h->nr_huge_pages_node[nid]++; | |
1637 | } | |
1638 | ||
f41f2ed4 | 1639 | static void __prep_new_huge_page(struct hstate *h, struct page *page) |
b7ba30c6 | 1640 | { |
f41f2ed4 | 1641 | free_huge_page_vmemmap(h, page); |
0edaecfa | 1642 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1643 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1644 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1645 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1646 | set_hugetlb_cgroup_rsvd(page, NULL); |
d3d99fcc OS |
1647 | } |
1648 | ||
1649 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) | |
1650 | { | |
f41f2ed4 | 1651 | __prep_new_huge_page(h, page); |
db71ef79 | 1652 | spin_lock_irq(&hugetlb_lock); |
d3d99fcc | 1653 | __prep_account_new_huge_page(h, nid); |
db71ef79 | 1654 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1655 | } |
1656 | ||
7118fc29 | 1657 | static bool prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c | 1658 | { |
7118fc29 | 1659 | int i, j; |
20a0307c WF |
1660 | int nr_pages = 1 << order; |
1661 | struct page *p = page + 1; | |
1662 | ||
1663 | /* we rely on prep_new_huge_page to set the destructor */ | |
1664 | set_compound_order(page, order); | |
ef5a22be | 1665 | __ClearPageReserved(page); |
de09d31d | 1666 | __SetPageHead(page); |
20a0307c | 1667 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1668 | /* |
1669 | * For gigantic hugepages allocated through bootmem at | |
1670 | * boot, it's safer to be consistent with the not-gigantic | |
1671 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1672 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1673 | * pages may get the reference counting wrong if they see |
1674 | * PG_reserved set on a tail page (despite the head page not | |
1675 | * having PG_reserved set). Enforcing this consistency between | |
1676 | * head and tail pages allows drivers to optimize away a check | |
1677 | * on the head page when they need know if put_page() is needed | |
1678 | * after get_user_pages(). | |
1679 | */ | |
1680 | __ClearPageReserved(p); | |
7118fc29 MK |
1681 | /* |
1682 | * Subtle and very unlikely | |
1683 | * | |
1684 | * Gigantic 'page allocators' such as memblock or cma will | |
1685 | * return a set of pages with each page ref counted. We need | |
1686 | * to turn this set of pages into a compound page with tail | |
1687 | * page ref counts set to zero. Code such as speculative page | |
1688 | * cache adding could take a ref on a 'to be' tail page. | |
1689 | * We need to respect any increased ref count, and only set | |
1690 | * the ref count to zero if count is currently 1. If count | |
416d85ed MK |
1691 | * is not 1, we return an error. An error return indicates |
1692 | * the set of pages can not be converted to a gigantic page. | |
1693 | * The caller who allocated the pages should then discard the | |
1694 | * pages using the appropriate free interface. | |
7118fc29 MK |
1695 | */ |
1696 | if (!page_ref_freeze(p, 1)) { | |
416d85ed MK |
1697 | pr_warn("HugeTLB page can not be used due to unexpected inflated ref count\n"); |
1698 | goto out_error; | |
7118fc29 | 1699 | } |
58a84aa9 | 1700 | set_page_count(p, 0); |
1d798ca3 | 1701 | set_compound_head(p, page); |
20a0307c | 1702 | } |
b4330afb | 1703 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1704 | atomic_set(compound_pincount_ptr(page), 0); |
7118fc29 MK |
1705 | return true; |
1706 | ||
1707 | out_error: | |
1708 | /* undo tail page modifications made above */ | |
1709 | p = page + 1; | |
1710 | for (j = 1; j < i; j++, p = mem_map_next(p, page, j)) { | |
1711 | clear_compound_head(p); | |
1712 | set_page_refcounted(p); | |
1713 | } | |
1714 | /* need to clear PG_reserved on remaining tail pages */ | |
1715 | for (; j < nr_pages; j++, p = mem_map_next(p, page, j)) | |
1716 | __ClearPageReserved(p); | |
1717 | set_compound_order(page, 0); | |
1718 | page[1].compound_nr = 0; | |
1719 | __ClearPageHead(page); | |
1720 | return false; | |
20a0307c WF |
1721 | } |
1722 | ||
7795912c AM |
1723 | /* |
1724 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1725 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1726 | * details. | |
1727 | */ | |
20a0307c WF |
1728 | int PageHuge(struct page *page) |
1729 | { | |
20a0307c WF |
1730 | if (!PageCompound(page)) |
1731 | return 0; | |
1732 | ||
1733 | page = compound_head(page); | |
f1e61557 | 1734 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1735 | } |
43131e14 NH |
1736 | EXPORT_SYMBOL_GPL(PageHuge); |
1737 | ||
27c73ae7 AA |
1738 | /* |
1739 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1740 | * normal or transparent huge pages. | |
1741 | */ | |
1742 | int PageHeadHuge(struct page *page_head) | |
1743 | { | |
27c73ae7 AA |
1744 | if (!PageHead(page_head)) |
1745 | return 0; | |
1746 | ||
d4af73e3 | 1747 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1748 | } |
27c73ae7 | 1749 | |
c0d0381a MK |
1750 | /* |
1751 | * Find and lock address space (mapping) in write mode. | |
1752 | * | |
336bf30e MK |
1753 | * Upon entry, the page is locked which means that page_mapping() is |
1754 | * stable. Due to locking order, we can only trylock_write. If we can | |
1755 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1756 | */ |
1757 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1758 | { | |
336bf30e | 1759 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1760 | |
c0d0381a MK |
1761 | if (!mapping) |
1762 | return mapping; | |
1763 | ||
c0d0381a MK |
1764 | if (i_mmap_trylock_write(mapping)) |
1765 | return mapping; | |
1766 | ||
336bf30e | 1767 | return NULL; |
c0d0381a MK |
1768 | } |
1769 | ||
fe19bd3d | 1770 | pgoff_t hugetlb_basepage_index(struct page *page) |
13d60f4b ZY |
1771 | { |
1772 | struct page *page_head = compound_head(page); | |
1773 | pgoff_t index = page_index(page_head); | |
1774 | unsigned long compound_idx; | |
1775 | ||
13d60f4b ZY |
1776 | if (compound_order(page_head) >= MAX_ORDER) |
1777 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1778 | else | |
1779 | compound_idx = page - page_head; | |
1780 | ||
1781 | return (index << compound_order(page_head)) + compound_idx; | |
1782 | } | |
1783 | ||
0c397dae | 1784 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1785 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1786 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1787 | { |
af0fb9df | 1788 | int order = huge_page_order(h); |
1da177e4 | 1789 | struct page *page; |
f60858f9 | 1790 | bool alloc_try_hard = true; |
f96efd58 | 1791 | |
f60858f9 MK |
1792 | /* |
1793 | * By default we always try hard to allocate the page with | |
1794 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1795 | * a loop (to adjust global huge page counts) and previous allocation | |
1796 | * failed, do not continue to try hard on the same node. Use the | |
1797 | * node_alloc_noretry bitmap to manage this state information. | |
1798 | */ | |
1799 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1800 | alloc_try_hard = false; | |
1801 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1802 | if (alloc_try_hard) | |
1803 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1804 | if (nid == NUMA_NO_NODE) |
1805 | nid = numa_mem_id(); | |
84172f4b | 1806 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
af0fb9df MH |
1807 | if (page) |
1808 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1809 | else | |
1810 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1811 | |
f60858f9 MK |
1812 | /* |
1813 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1814 | * indicates an overall state change. Clear bit so that we resume | |
1815 | * normal 'try hard' allocations. | |
1816 | */ | |
1817 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1818 | node_clear(nid, *node_alloc_noretry); | |
1819 | ||
1820 | /* | |
1821 | * If we tried hard to get a page but failed, set bit so that | |
1822 | * subsequent attempts will not try as hard until there is an | |
1823 | * overall state change. | |
1824 | */ | |
1825 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1826 | node_set(nid, *node_alloc_noretry); | |
1827 | ||
63b4613c NA |
1828 | return page; |
1829 | } | |
1830 | ||
0c397dae MH |
1831 | /* |
1832 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1833 | * should use this function to get new hugetlb pages | |
1834 | */ | |
1835 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1836 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1837 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1838 | { |
1839 | struct page *page; | |
7118fc29 | 1840 | bool retry = false; |
0c397dae | 1841 | |
7118fc29 | 1842 | retry: |
0c397dae MH |
1843 | if (hstate_is_gigantic(h)) |
1844 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1845 | else | |
1846 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1847 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1848 | if (!page) |
1849 | return NULL; | |
1850 | ||
7118fc29 MK |
1851 | if (hstate_is_gigantic(h)) { |
1852 | if (!prep_compound_gigantic_page(page, huge_page_order(h))) { | |
1853 | /* | |
1854 | * Rare failure to convert pages to compound page. | |
1855 | * Free pages and try again - ONCE! | |
1856 | */ | |
1857 | free_gigantic_page(page, huge_page_order(h)); | |
1858 | if (!retry) { | |
1859 | retry = true; | |
1860 | goto retry; | |
1861 | } | |
7118fc29 MK |
1862 | return NULL; |
1863 | } | |
1864 | } | |
0c397dae MH |
1865 | prep_new_huge_page(h, page, page_to_nid(page)); |
1866 | ||
1867 | return page; | |
1868 | } | |
1869 | ||
af0fb9df MH |
1870 | /* |
1871 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1872 | * manner. | |
1873 | */ | |
f60858f9 MK |
1874 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1875 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1876 | { |
1877 | struct page *page; | |
1878 | int nr_nodes, node; | |
af0fb9df | 1879 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1880 | |
1881 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1882 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1883 | node_alloc_noretry); | |
af0fb9df | 1884 | if (page) |
b2261026 | 1885 | break; |
b2261026 JK |
1886 | } |
1887 | ||
af0fb9df MH |
1888 | if (!page) |
1889 | return 0; | |
b2261026 | 1890 | |
af0fb9df MH |
1891 | put_page(page); /* free it into the hugepage allocator */ |
1892 | ||
1893 | return 1; | |
b2261026 JK |
1894 | } |
1895 | ||
e8c5c824 | 1896 | /* |
10c6ec49 MK |
1897 | * Remove huge page from pool from next node to free. Attempt to keep |
1898 | * persistent huge pages more or less balanced over allowed nodes. | |
1899 | * This routine only 'removes' the hugetlb page. The caller must make | |
1900 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
1901 | * Called with hugetlb_lock locked. |
1902 | */ | |
10c6ec49 MK |
1903 | static struct page *remove_pool_huge_page(struct hstate *h, |
1904 | nodemask_t *nodes_allowed, | |
1905 | bool acct_surplus) | |
e8c5c824 | 1906 | { |
b2261026 | 1907 | int nr_nodes, node; |
10c6ec49 | 1908 | struct page *page = NULL; |
e8c5c824 | 1909 | |
9487ca60 | 1910 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 1911 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1912 | /* |
1913 | * If we're returning unused surplus pages, only examine | |
1914 | * nodes with surplus pages. | |
1915 | */ | |
b2261026 JK |
1916 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1917 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 1918 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 1919 | struct page, lru); |
6eb4e88a | 1920 | remove_hugetlb_page(h, page, acct_surplus); |
9a76db09 | 1921 | break; |
e8c5c824 | 1922 | } |
b2261026 | 1923 | } |
e8c5c824 | 1924 | |
10c6ec49 | 1925 | return page; |
e8c5c824 LS |
1926 | } |
1927 | ||
c8721bbb NH |
1928 | /* |
1929 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1930 | * nothing for in-use hugepages and non-hugepages. |
1931 | * This function returns values like below: | |
1932 | * | |
ad2fa371 MS |
1933 | * -ENOMEM: failed to allocate vmemmap pages to free the freed hugepages |
1934 | * when the system is under memory pressure and the feature of | |
1935 | * freeing unused vmemmap pages associated with each hugetlb page | |
1936 | * is enabled. | |
1937 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1938 | * (allocated or reserved.) | |
1939 | * 0: successfully dissolved free hugepages or the page is not a | |
1940 | * hugepage (considered as already dissolved) | |
c8721bbb | 1941 | */ |
c3114a84 | 1942 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1943 | { |
6bc9b564 | 1944 | int rc = -EBUSY; |
082d5b6b | 1945 | |
7ffddd49 | 1946 | retry: |
faf53def NH |
1947 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1948 | if (!PageHuge(page)) | |
1949 | return 0; | |
1950 | ||
db71ef79 | 1951 | spin_lock_irq(&hugetlb_lock); |
faf53def NH |
1952 | if (!PageHuge(page)) { |
1953 | rc = 0; | |
1954 | goto out; | |
1955 | } | |
1956 | ||
1957 | if (!page_count(page)) { | |
2247bb33 GS |
1958 | struct page *head = compound_head(page); |
1959 | struct hstate *h = page_hstate(head); | |
6bc9b564 | 1960 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1961 | goto out; |
7ffddd49 MS |
1962 | |
1963 | /* | |
1964 | * We should make sure that the page is already on the free list | |
1965 | * when it is dissolved. | |
1966 | */ | |
6c037149 | 1967 | if (unlikely(!HPageFreed(head))) { |
db71ef79 | 1968 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
1969 | cond_resched(); |
1970 | ||
1971 | /* | |
1972 | * Theoretically, we should return -EBUSY when we | |
1973 | * encounter this race. In fact, we have a chance | |
1974 | * to successfully dissolve the page if we do a | |
1975 | * retry. Because the race window is quite small. | |
1976 | * If we seize this opportunity, it is an optimization | |
1977 | * for increasing the success rate of dissolving page. | |
1978 | */ | |
1979 | goto retry; | |
1980 | } | |
1981 | ||
0c5da357 | 1982 | remove_hugetlb_page(h, head, false); |
c1470b33 | 1983 | h->max_huge_pages--; |
db71ef79 | 1984 | spin_unlock_irq(&hugetlb_lock); |
ad2fa371 MS |
1985 | |
1986 | /* | |
1987 | * Normally update_and_free_page will allocate required vmemmmap | |
1988 | * before freeing the page. update_and_free_page will fail to | |
1989 | * free the page if it can not allocate required vmemmap. We | |
1990 | * need to adjust max_huge_pages if the page is not freed. | |
1991 | * Attempt to allocate vmemmmap here so that we can take | |
1992 | * appropriate action on failure. | |
1993 | */ | |
1994 | rc = alloc_huge_page_vmemmap(h, head); | |
1995 | if (!rc) { | |
1996 | /* | |
1997 | * Move PageHWPoison flag from head page to the raw | |
1998 | * error page, which makes any subpages rather than | |
1999 | * the error page reusable. | |
2000 | */ | |
2001 | if (PageHWPoison(head) && page != head) { | |
2002 | SetPageHWPoison(page); | |
2003 | ClearPageHWPoison(head); | |
2004 | } | |
2005 | update_and_free_page(h, head, false); | |
2006 | } else { | |
2007 | spin_lock_irq(&hugetlb_lock); | |
2008 | add_hugetlb_page(h, head, false); | |
2009 | h->max_huge_pages++; | |
2010 | spin_unlock_irq(&hugetlb_lock); | |
2011 | } | |
2012 | ||
2013 | return rc; | |
c8721bbb | 2014 | } |
082d5b6b | 2015 | out: |
db71ef79 | 2016 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 2017 | return rc; |
c8721bbb NH |
2018 | } |
2019 | ||
2020 | /* | |
2021 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
2022 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
2023 | * Note that this will dissolve a free gigantic hugepage completely, if any |
2024 | * part of it lies within the given range. | |
082d5b6b GS |
2025 | * Also note that if dissolve_free_huge_page() returns with an error, all |
2026 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 2027 | */ |
082d5b6b | 2028 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 2029 | { |
c8721bbb | 2030 | unsigned long pfn; |
eb03aa00 | 2031 | struct page *page; |
082d5b6b | 2032 | int rc = 0; |
c8721bbb | 2033 | |
d0177639 | 2034 | if (!hugepages_supported()) |
082d5b6b | 2035 | return rc; |
d0177639 | 2036 | |
eb03aa00 GS |
2037 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
2038 | page = pfn_to_page(pfn); | |
faf53def NH |
2039 | rc = dissolve_free_huge_page(page); |
2040 | if (rc) | |
2041 | break; | |
eb03aa00 | 2042 | } |
082d5b6b GS |
2043 | |
2044 | return rc; | |
c8721bbb NH |
2045 | } |
2046 | ||
ab5ac90a MH |
2047 | /* |
2048 | * Allocates a fresh surplus page from the page allocator. | |
2049 | */ | |
0c397dae | 2050 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
b65a4eda | 2051 | int nid, nodemask_t *nmask, bool zero_ref) |
7893d1d5 | 2052 | { |
9980d744 | 2053 | struct page *page = NULL; |
b65a4eda | 2054 | bool retry = false; |
7893d1d5 | 2055 | |
bae7f4ae | 2056 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2057 | return NULL; |
2058 | ||
db71ef79 | 2059 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2060 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
2061 | goto out_unlock; | |
db71ef79 | 2062 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 2063 | |
b65a4eda | 2064 | retry: |
f60858f9 | 2065 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 2066 | if (!page) |
0c397dae | 2067 | return NULL; |
d1c3fb1f | 2068 | |
db71ef79 | 2069 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
2070 | /* |
2071 | * We could have raced with the pool size change. | |
2072 | * Double check that and simply deallocate the new page | |
2073 | * if we would end up overcommiting the surpluses. Abuse | |
2074 | * temporary page to workaround the nasty free_huge_page | |
2075 | * codeflow | |
2076 | */ | |
2077 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 2078 | SetHPageTemporary(page); |
db71ef79 | 2079 | spin_unlock_irq(&hugetlb_lock); |
9980d744 | 2080 | put_page(page); |
2bf753e6 | 2081 | return NULL; |
7893d1d5 | 2082 | } |
9980d744 | 2083 | |
b65a4eda MK |
2084 | if (zero_ref) { |
2085 | /* | |
2086 | * Caller requires a page with zero ref count. | |
2087 | * We will drop ref count here. If someone else is holding | |
2088 | * a ref, the page will be freed when they drop it. Abuse | |
2089 | * temporary page flag to accomplish this. | |
2090 | */ | |
2091 | SetHPageTemporary(page); | |
2092 | if (!put_page_testzero(page)) { | |
2093 | /* | |
2094 | * Unexpected inflated ref count on freshly allocated | |
2095 | * huge. Retry once. | |
2096 | */ | |
2097 | pr_info("HugeTLB unexpected inflated ref count on freshly allocated page\n"); | |
2098 | spin_unlock_irq(&hugetlb_lock); | |
2099 | if (retry) | |
2100 | return NULL; | |
2101 | ||
2102 | retry = true; | |
2103 | goto retry; | |
2104 | } | |
2105 | ClearHPageTemporary(page); | |
2106 | } | |
2107 | ||
2108 | h->surplus_huge_pages++; | |
2109 | h->surplus_huge_pages_node[page_to_nid(page)]++; | |
2110 | ||
9980d744 | 2111 | out_unlock: |
db71ef79 | 2112 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 AL |
2113 | |
2114 | return page; | |
2115 | } | |
2116 | ||
bbe88753 | 2117 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 2118 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
2119 | { |
2120 | struct page *page; | |
2121 | ||
2122 | if (hstate_is_gigantic(h)) | |
2123 | return NULL; | |
2124 | ||
f60858f9 | 2125 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
2126 | if (!page) |
2127 | return NULL; | |
2128 | ||
2129 | /* | |
2130 | * We do not account these pages as surplus because they are only | |
2131 | * temporary and will be released properly on the last reference | |
2132 | */ | |
9157c311 | 2133 | SetHPageTemporary(page); |
ab5ac90a MH |
2134 | |
2135 | return page; | |
2136 | } | |
2137 | ||
099730d6 DH |
2138 | /* |
2139 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
2140 | */ | |
e0ec90ee | 2141 | static |
0c397dae | 2142 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
2143 | struct vm_area_struct *vma, unsigned long addr) |
2144 | { | |
aaf14e40 MH |
2145 | struct page *page; |
2146 | struct mempolicy *mpol; | |
2147 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
2148 | int nid; | |
2149 | nodemask_t *nodemask; | |
2150 | ||
2151 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
b65a4eda | 2152 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask, false); |
aaf14e40 MH |
2153 | mpol_cond_put(mpol); |
2154 | ||
2155 | return page; | |
099730d6 DH |
2156 | } |
2157 | ||
ab5ac90a | 2158 | /* page migration callback function */ |
3e59fcb0 | 2159 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 2160 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 2161 | { |
db71ef79 | 2162 | spin_lock_irq(&hugetlb_lock); |
4db9b2ef | 2163 | if (h->free_huge_pages - h->resv_huge_pages > 0) { |
3e59fcb0 MH |
2164 | struct page *page; |
2165 | ||
2166 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
2167 | if (page) { | |
db71ef79 | 2168 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 2169 | return page; |
4db9b2ef MH |
2170 | } |
2171 | } | |
db71ef79 | 2172 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 2173 | |
0c397dae | 2174 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2175 | } |
2176 | ||
ebd63723 | 2177 | /* mempolicy aware migration callback */ |
389c8178 MH |
2178 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2179 | unsigned long address) | |
ebd63723 MH |
2180 | { |
2181 | struct mempolicy *mpol; | |
2182 | nodemask_t *nodemask; | |
2183 | struct page *page; | |
ebd63723 MH |
2184 | gfp_t gfp_mask; |
2185 | int node; | |
2186 | ||
ebd63723 MH |
2187 | gfp_mask = htlb_alloc_mask(h); |
2188 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 2189 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
2190 | mpol_cond_put(mpol); |
2191 | ||
2192 | return page; | |
2193 | } | |
2194 | ||
e4e574b7 | 2195 | /* |
25985edc | 2196 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2197 | * of size 'delta'. |
2198 | */ | |
0a4f3d1b | 2199 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 2200 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
2201 | { |
2202 | struct list_head surplus_list; | |
2203 | struct page *page, *tmp; | |
0a4f3d1b LX |
2204 | int ret; |
2205 | long i; | |
2206 | long needed, allocated; | |
28073b02 | 2207 | bool alloc_ok = true; |
e4e574b7 | 2208 | |
9487ca60 | 2209 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 2210 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2211 | if (needed <= 0) { |
a5516438 | 2212 | h->resv_huge_pages += delta; |
e4e574b7 | 2213 | return 0; |
ac09b3a1 | 2214 | } |
e4e574b7 AL |
2215 | |
2216 | allocated = 0; | |
2217 | INIT_LIST_HEAD(&surplus_list); | |
2218 | ||
2219 | ret = -ENOMEM; | |
2220 | retry: | |
db71ef79 | 2221 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 2222 | for (i = 0; i < needed; i++) { |
0c397dae | 2223 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
b65a4eda | 2224 | NUMA_NO_NODE, NULL, true); |
28073b02 HD |
2225 | if (!page) { |
2226 | alloc_ok = false; | |
2227 | break; | |
2228 | } | |
e4e574b7 | 2229 | list_add(&page->lru, &surplus_list); |
69ed779a | 2230 | cond_resched(); |
e4e574b7 | 2231 | } |
28073b02 | 2232 | allocated += i; |
e4e574b7 AL |
2233 | |
2234 | /* | |
2235 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2236 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2237 | */ | |
db71ef79 | 2238 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
2239 | needed = (h->resv_huge_pages + delta) - |
2240 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2241 | if (needed > 0) { |
2242 | if (alloc_ok) | |
2243 | goto retry; | |
2244 | /* | |
2245 | * We were not able to allocate enough pages to | |
2246 | * satisfy the entire reservation so we free what | |
2247 | * we've allocated so far. | |
2248 | */ | |
2249 | goto free; | |
2250 | } | |
e4e574b7 AL |
2251 | /* |
2252 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2253 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2254 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2255 | * allocator. Commit the entire reservation here to prevent another |
2256 | * process from stealing the pages as they are added to the pool but | |
2257 | * before they are reserved. | |
e4e574b7 AL |
2258 | */ |
2259 | needed += allocated; | |
a5516438 | 2260 | h->resv_huge_pages += delta; |
e4e574b7 | 2261 | ret = 0; |
a9869b83 | 2262 | |
19fc3f0a | 2263 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2264 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2265 | if ((--needed) < 0) |
2266 | break; | |
b65a4eda | 2267 | /* Add the page to the hugetlb allocator */ |
a5516438 | 2268 | enqueue_huge_page(h, page); |
19fc3f0a | 2269 | } |
28073b02 | 2270 | free: |
db71ef79 | 2271 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a | 2272 | |
b65a4eda MK |
2273 | /* |
2274 | * Free unnecessary surplus pages to the buddy allocator. | |
2275 | * Pages have no ref count, call free_huge_page directly. | |
2276 | */ | |
c0d934ba | 2277 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
b65a4eda | 2278 | free_huge_page(page); |
db71ef79 | 2279 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2280 | |
2281 | return ret; | |
2282 | } | |
2283 | ||
2284 | /* | |
e5bbc8a6 MK |
2285 | * This routine has two main purposes: |
2286 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2287 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2288 | * to the associated reservation map. | |
2289 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2290 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2291 | */ |
a5516438 AK |
2292 | static void return_unused_surplus_pages(struct hstate *h, |
2293 | unsigned long unused_resv_pages) | |
e4e574b7 | 2294 | { |
e4e574b7 | 2295 | unsigned long nr_pages; |
10c6ec49 MK |
2296 | struct page *page; |
2297 | LIST_HEAD(page_list); | |
2298 | ||
9487ca60 | 2299 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2300 | /* Uncommit the reservation */ |
2301 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2302 | |
aa888a74 | 2303 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2304 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2305 | goto out; |
aa888a74 | 2306 | |
e5bbc8a6 MK |
2307 | /* |
2308 | * Part (or even all) of the reservation could have been backed | |
2309 | * by pre-allocated pages. Only free surplus pages. | |
2310 | */ | |
a5516438 | 2311 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2312 | |
685f3457 LS |
2313 | /* |
2314 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2315 | * evenly across all nodes with memory. Iterate across these nodes |
2316 | * until we can no longer free unreserved surplus pages. This occurs | |
2317 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2318 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2319 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2320 | */ |
2321 | while (nr_pages--) { | |
10c6ec49 MK |
2322 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2323 | if (!page) | |
e5bbc8a6 | 2324 | goto out; |
10c6ec49 MK |
2325 | |
2326 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2327 | } |
e5bbc8a6 MK |
2328 | |
2329 | out: | |
db71ef79 | 2330 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2331 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2332 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2333 | } |
2334 | ||
5e911373 | 2335 | |
c37f9fb1 | 2336 | /* |
feba16e2 | 2337 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2338 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2339 | * |
2340 | * vma_needs_reservation is called to determine if the huge page at addr | |
2341 | * within the vma has an associated reservation. If a reservation is | |
2342 | * needed, the value 1 is returned. The caller is then responsible for | |
2343 | * managing the global reservation and subpool usage counts. After | |
2344 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2345 | * to add the page to the reservation map. If the page allocation fails, |
2346 | * the reservation must be ended instead of committed. vma_end_reservation | |
2347 | * is called in such cases. | |
cf3ad20b MK |
2348 | * |
2349 | * In the normal case, vma_commit_reservation returns the same value | |
2350 | * as the preceding vma_needs_reservation call. The only time this | |
2351 | * is not the case is if a reserve map was changed between calls. It | |
2352 | * is the responsibility of the caller to notice the difference and | |
2353 | * take appropriate action. | |
96b96a96 MK |
2354 | * |
2355 | * vma_add_reservation is used in error paths where a reservation must | |
2356 | * be restored when a newly allocated huge page must be freed. It is | |
2357 | * to be called after calling vma_needs_reservation to determine if a | |
2358 | * reservation exists. | |
846be085 MK |
2359 | * |
2360 | * vma_del_reservation is used in error paths where an entry in the reserve | |
2361 | * map was created during huge page allocation and must be removed. It is to | |
2362 | * be called after calling vma_needs_reservation to determine if a reservation | |
2363 | * exists. | |
c37f9fb1 | 2364 | */ |
5e911373 MK |
2365 | enum vma_resv_mode { |
2366 | VMA_NEEDS_RESV, | |
2367 | VMA_COMMIT_RESV, | |
feba16e2 | 2368 | VMA_END_RESV, |
96b96a96 | 2369 | VMA_ADD_RESV, |
846be085 | 2370 | VMA_DEL_RESV, |
5e911373 | 2371 | }; |
cf3ad20b MK |
2372 | static long __vma_reservation_common(struct hstate *h, |
2373 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2374 | enum vma_resv_mode mode) |
c37f9fb1 | 2375 | { |
4e35f483 JK |
2376 | struct resv_map *resv; |
2377 | pgoff_t idx; | |
cf3ad20b | 2378 | long ret; |
0db9d74e | 2379 | long dummy_out_regions_needed; |
c37f9fb1 | 2380 | |
4e35f483 JK |
2381 | resv = vma_resv_map(vma); |
2382 | if (!resv) | |
84afd99b | 2383 | return 1; |
c37f9fb1 | 2384 | |
4e35f483 | 2385 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2386 | switch (mode) { |
2387 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2388 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2389 | /* We assume that vma_reservation_* routines always operate on | |
2390 | * 1 page, and that adding to resv map a 1 page entry can only | |
2391 | * ever require 1 region. | |
2392 | */ | |
2393 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2394 | break; |
2395 | case VMA_COMMIT_RESV: | |
075a61d0 | 2396 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2397 | /* region_add calls of range 1 should never fail. */ |
2398 | VM_BUG_ON(ret < 0); | |
5e911373 | 2399 | break; |
feba16e2 | 2400 | case VMA_END_RESV: |
0db9d74e | 2401 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2402 | ret = 0; |
2403 | break; | |
96b96a96 | 2404 | case VMA_ADD_RESV: |
0db9d74e | 2405 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2406 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2407 | /* region_add calls of range 1 should never fail. */ |
2408 | VM_BUG_ON(ret < 0); | |
2409 | } else { | |
2410 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2411 | ret = region_del(resv, idx, idx + 1); |
2412 | } | |
2413 | break; | |
846be085 MK |
2414 | case VMA_DEL_RESV: |
2415 | if (vma->vm_flags & VM_MAYSHARE) { | |
2416 | region_abort(resv, idx, idx + 1, 1); | |
2417 | ret = region_del(resv, idx, idx + 1); | |
2418 | } else { | |
2419 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); | |
2420 | /* region_add calls of range 1 should never fail. */ | |
2421 | VM_BUG_ON(ret < 0); | |
2422 | } | |
2423 | break; | |
5e911373 MK |
2424 | default: |
2425 | BUG(); | |
2426 | } | |
84afd99b | 2427 | |
846be085 | 2428 | if (vma->vm_flags & VM_MAYSHARE || mode == VMA_DEL_RESV) |
cf3ad20b | 2429 | return ret; |
bf3d12b9 ML |
2430 | /* |
2431 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2432 | * | |
2433 | * In most cases, reserves always exist for private mappings. | |
2434 | * However, a file associated with mapping could have been | |
2435 | * hole punched or truncated after reserves were consumed. | |
2436 | * As subsequent fault on such a range will not use reserves. | |
2437 | * Subtle - The reserve map for private mappings has the | |
2438 | * opposite meaning than that of shared mappings. If NO | |
2439 | * entry is in the reserve map, it means a reservation exists. | |
2440 | * If an entry exists in the reserve map, it means the | |
2441 | * reservation has already been consumed. As a result, the | |
2442 | * return value of this routine is the opposite of the | |
2443 | * value returned from reserve map manipulation routines above. | |
2444 | */ | |
2445 | if (ret > 0) | |
2446 | return 0; | |
2447 | if (ret == 0) | |
2448 | return 1; | |
2449 | return ret; | |
c37f9fb1 | 2450 | } |
cf3ad20b MK |
2451 | |
2452 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2453 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2454 | { |
5e911373 | 2455 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2456 | } |
84afd99b | 2457 | |
cf3ad20b MK |
2458 | static long vma_commit_reservation(struct hstate *h, |
2459 | struct vm_area_struct *vma, unsigned long addr) | |
2460 | { | |
5e911373 MK |
2461 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2462 | } | |
2463 | ||
feba16e2 | 2464 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2465 | struct vm_area_struct *vma, unsigned long addr) |
2466 | { | |
feba16e2 | 2467 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2468 | } |
2469 | ||
96b96a96 MK |
2470 | static long vma_add_reservation(struct hstate *h, |
2471 | struct vm_area_struct *vma, unsigned long addr) | |
2472 | { | |
2473 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2474 | } | |
2475 | ||
846be085 MK |
2476 | static long vma_del_reservation(struct hstate *h, |
2477 | struct vm_area_struct *vma, unsigned long addr) | |
2478 | { | |
2479 | return __vma_reservation_common(h, vma, addr, VMA_DEL_RESV); | |
2480 | } | |
2481 | ||
96b96a96 | 2482 | /* |
846be085 MK |
2483 | * This routine is called to restore reservation information on error paths. |
2484 | * It should ONLY be called for pages allocated via alloc_huge_page(), and | |
2485 | * the hugetlb mutex should remain held when calling this routine. | |
2486 | * | |
2487 | * It handles two specific cases: | |
2488 | * 1) A reservation was in place and the page consumed the reservation. | |
2489 | * HPageRestoreReserve is set in the page. | |
2490 | * 2) No reservation was in place for the page, so HPageRestoreReserve is | |
2491 | * not set. However, alloc_huge_page always updates the reserve map. | |
2492 | * | |
2493 | * In case 1, free_huge_page later in the error path will increment the | |
2494 | * global reserve count. But, free_huge_page does not have enough context | |
2495 | * to adjust the reservation map. This case deals primarily with private | |
2496 | * mappings. Adjust the reserve map here to be consistent with global | |
2497 | * reserve count adjustments to be made by free_huge_page. Make sure the | |
2498 | * reserve map indicates there is a reservation present. | |
2499 | * | |
2500 | * In case 2, simply undo reserve map modifications done by alloc_huge_page. | |
96b96a96 | 2501 | */ |
846be085 MK |
2502 | void restore_reserve_on_error(struct hstate *h, struct vm_area_struct *vma, |
2503 | unsigned long address, struct page *page) | |
96b96a96 | 2504 | { |
846be085 | 2505 | long rc = vma_needs_reservation(h, vma, address); |
96b96a96 | 2506 | |
846be085 MK |
2507 | if (HPageRestoreReserve(page)) { |
2508 | if (unlikely(rc < 0)) | |
96b96a96 MK |
2509 | /* |
2510 | * Rare out of memory condition in reserve map | |
d6995da3 | 2511 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2512 | * global reserve count will not be incremented |
2513 | * by free_huge_page. This will make it appear | |
2514 | * as though the reservation for this page was | |
2515 | * consumed. This may prevent the task from | |
2516 | * faulting in the page at a later time. This | |
2517 | * is better than inconsistent global huge page | |
2518 | * accounting of reserve counts. | |
2519 | */ | |
d6995da3 | 2520 | ClearHPageRestoreReserve(page); |
846be085 MK |
2521 | else if (rc) |
2522 | (void)vma_add_reservation(h, vma, address); | |
2523 | else | |
2524 | vma_end_reservation(h, vma, address); | |
2525 | } else { | |
2526 | if (!rc) { | |
2527 | /* | |
2528 | * This indicates there is an entry in the reserve map | |
c7b1850d | 2529 | * not added by alloc_huge_page. We know it was added |
846be085 MK |
2530 | * before the alloc_huge_page call, otherwise |
2531 | * HPageRestoreReserve would be set on the page. | |
2532 | * Remove the entry so that a subsequent allocation | |
2533 | * does not consume a reservation. | |
2534 | */ | |
2535 | rc = vma_del_reservation(h, vma, address); | |
2536 | if (rc < 0) | |
96b96a96 | 2537 | /* |
846be085 MK |
2538 | * VERY rare out of memory condition. Since |
2539 | * we can not delete the entry, set | |
2540 | * HPageRestoreReserve so that the reserve | |
2541 | * count will be incremented when the page | |
2542 | * is freed. This reserve will be consumed | |
2543 | * on a subsequent allocation. | |
96b96a96 | 2544 | */ |
846be085 MK |
2545 | SetHPageRestoreReserve(page); |
2546 | } else if (rc < 0) { | |
2547 | /* | |
2548 | * Rare out of memory condition from | |
2549 | * vma_needs_reservation call. Memory allocation is | |
2550 | * only attempted if a new entry is needed. Therefore, | |
2551 | * this implies there is not an entry in the | |
2552 | * reserve map. | |
2553 | * | |
2554 | * For shared mappings, no entry in the map indicates | |
2555 | * no reservation. We are done. | |
2556 | */ | |
2557 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
2558 | /* | |
2559 | * For private mappings, no entry indicates | |
2560 | * a reservation is present. Since we can | |
2561 | * not add an entry, set SetHPageRestoreReserve | |
2562 | * on the page so reserve count will be | |
2563 | * incremented when freed. This reserve will | |
2564 | * be consumed on a subsequent allocation. | |
2565 | */ | |
2566 | SetHPageRestoreReserve(page); | |
96b96a96 | 2567 | } else |
846be085 MK |
2568 | /* |
2569 | * No reservation present, do nothing | |
2570 | */ | |
2571 | vma_end_reservation(h, vma, address); | |
96b96a96 MK |
2572 | } |
2573 | } | |
2574 | ||
369fa227 OS |
2575 | /* |
2576 | * alloc_and_dissolve_huge_page - Allocate a new page and dissolve the old one | |
2577 | * @h: struct hstate old page belongs to | |
2578 | * @old_page: Old page to dissolve | |
ae37c7ff | 2579 | * @list: List to isolate the page in case we need to |
369fa227 OS |
2580 | * Returns 0 on success, otherwise negated error. |
2581 | */ | |
ae37c7ff OS |
2582 | static int alloc_and_dissolve_huge_page(struct hstate *h, struct page *old_page, |
2583 | struct list_head *list) | |
369fa227 OS |
2584 | { |
2585 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; | |
2586 | int nid = page_to_nid(old_page); | |
b65a4eda | 2587 | bool alloc_retry = false; |
369fa227 OS |
2588 | struct page *new_page; |
2589 | int ret = 0; | |
2590 | ||
2591 | /* | |
2592 | * Before dissolving the page, we need to allocate a new one for the | |
f41f2ed4 MS |
2593 | * pool to remain stable. Here, we allocate the page and 'prep' it |
2594 | * by doing everything but actually updating counters and adding to | |
2595 | * the pool. This simplifies and let us do most of the processing | |
2596 | * under the lock. | |
369fa227 | 2597 | */ |
b65a4eda | 2598 | alloc_retry: |
369fa227 OS |
2599 | new_page = alloc_buddy_huge_page(h, gfp_mask, nid, NULL, NULL); |
2600 | if (!new_page) | |
2601 | return -ENOMEM; | |
b65a4eda MK |
2602 | /* |
2603 | * If all goes well, this page will be directly added to the free | |
2604 | * list in the pool. For this the ref count needs to be zero. | |
2605 | * Attempt to drop now, and retry once if needed. It is VERY | |
2606 | * unlikely there is another ref on the page. | |
2607 | * | |
2608 | * If someone else has a reference to the page, it will be freed | |
2609 | * when they drop their ref. Abuse temporary page flag to accomplish | |
2610 | * this. Retry once if there is an inflated ref count. | |
2611 | */ | |
2612 | SetHPageTemporary(new_page); | |
2613 | if (!put_page_testzero(new_page)) { | |
2614 | if (alloc_retry) | |
2615 | return -EBUSY; | |
2616 | ||
2617 | alloc_retry = true; | |
2618 | goto alloc_retry; | |
2619 | } | |
2620 | ClearHPageTemporary(new_page); | |
2621 | ||
f41f2ed4 | 2622 | __prep_new_huge_page(h, new_page); |
369fa227 OS |
2623 | |
2624 | retry: | |
2625 | spin_lock_irq(&hugetlb_lock); | |
2626 | if (!PageHuge(old_page)) { | |
2627 | /* | |
2628 | * Freed from under us. Drop new_page too. | |
2629 | */ | |
2630 | goto free_new; | |
2631 | } else if (page_count(old_page)) { | |
2632 | /* | |
ae37c7ff OS |
2633 | * Someone has grabbed the page, try to isolate it here. |
2634 | * Fail with -EBUSY if not possible. | |
369fa227 | 2635 | */ |
ae37c7ff OS |
2636 | spin_unlock_irq(&hugetlb_lock); |
2637 | if (!isolate_huge_page(old_page, list)) | |
2638 | ret = -EBUSY; | |
2639 | spin_lock_irq(&hugetlb_lock); | |
369fa227 OS |
2640 | goto free_new; |
2641 | } else if (!HPageFreed(old_page)) { | |
2642 | /* | |
2643 | * Page's refcount is 0 but it has not been enqueued in the | |
2644 | * freelist yet. Race window is small, so we can succeed here if | |
2645 | * we retry. | |
2646 | */ | |
2647 | spin_unlock_irq(&hugetlb_lock); | |
2648 | cond_resched(); | |
2649 | goto retry; | |
2650 | } else { | |
2651 | /* | |
2652 | * Ok, old_page is still a genuine free hugepage. Remove it from | |
2653 | * the freelist and decrease the counters. These will be | |
2654 | * incremented again when calling __prep_account_new_huge_page() | |
2655 | * and enqueue_huge_page() for new_page. The counters will remain | |
2656 | * stable since this happens under the lock. | |
2657 | */ | |
2658 | remove_hugetlb_page(h, old_page, false); | |
2659 | ||
2660 | /* | |
b65a4eda MK |
2661 | * Ref count on new page is already zero as it was dropped |
2662 | * earlier. It can be directly added to the pool free list. | |
369fa227 | 2663 | */ |
369fa227 | 2664 | __prep_account_new_huge_page(h, nid); |
369fa227 OS |
2665 | enqueue_huge_page(h, new_page); |
2666 | ||
2667 | /* | |
2668 | * Pages have been replaced, we can safely free the old one. | |
2669 | */ | |
2670 | spin_unlock_irq(&hugetlb_lock); | |
b65d4adb | 2671 | update_and_free_page(h, old_page, false); |
369fa227 OS |
2672 | } |
2673 | ||
2674 | return ret; | |
2675 | ||
2676 | free_new: | |
2677 | spin_unlock_irq(&hugetlb_lock); | |
b65a4eda MK |
2678 | /* Page has a zero ref count, but needs a ref to be freed */ |
2679 | set_page_refcounted(new_page); | |
b65d4adb | 2680 | update_and_free_page(h, new_page, false); |
369fa227 OS |
2681 | |
2682 | return ret; | |
2683 | } | |
2684 | ||
ae37c7ff | 2685 | int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list) |
369fa227 OS |
2686 | { |
2687 | struct hstate *h; | |
2688 | struct page *head; | |
ae37c7ff | 2689 | int ret = -EBUSY; |
369fa227 OS |
2690 | |
2691 | /* | |
2692 | * The page might have been dissolved from under our feet, so make sure | |
2693 | * to carefully check the state under the lock. | |
2694 | * Return success when racing as if we dissolved the page ourselves. | |
2695 | */ | |
2696 | spin_lock_irq(&hugetlb_lock); | |
2697 | if (PageHuge(page)) { | |
2698 | head = compound_head(page); | |
2699 | h = page_hstate(head); | |
2700 | } else { | |
2701 | spin_unlock_irq(&hugetlb_lock); | |
2702 | return 0; | |
2703 | } | |
2704 | spin_unlock_irq(&hugetlb_lock); | |
2705 | ||
2706 | /* | |
2707 | * Fence off gigantic pages as there is a cyclic dependency between | |
2708 | * alloc_contig_range and them. Return -ENOMEM as this has the effect | |
2709 | * of bailing out right away without further retrying. | |
2710 | */ | |
2711 | if (hstate_is_gigantic(h)) | |
2712 | return -ENOMEM; | |
2713 | ||
ae37c7ff OS |
2714 | if (page_count(head) && isolate_huge_page(head, list)) |
2715 | ret = 0; | |
2716 | else if (!page_count(head)) | |
2717 | ret = alloc_and_dissolve_huge_page(h, head, list); | |
2718 | ||
2719 | return ret; | |
369fa227 OS |
2720 | } |
2721 | ||
70c3547e | 2722 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2723 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2724 | { |
90481622 | 2725 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2726 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2727 | struct page *page; |
d85f69b0 MK |
2728 | long map_chg, map_commit; |
2729 | long gbl_chg; | |
6d76dcf4 AK |
2730 | int ret, idx; |
2731 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2732 | bool deferred_reserve; |
a1e78772 | 2733 | |
6d76dcf4 | 2734 | idx = hstate_index(h); |
a1e78772 | 2735 | /* |
d85f69b0 MK |
2736 | * Examine the region/reserve map to determine if the process |
2737 | * has a reservation for the page to be allocated. A return | |
2738 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2739 | */ |
d85f69b0 MK |
2740 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2741 | if (map_chg < 0) | |
76dcee75 | 2742 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2743 | |
2744 | /* | |
2745 | * Processes that did not create the mapping will have no | |
2746 | * reserves as indicated by the region/reserve map. Check | |
2747 | * that the allocation will not exceed the subpool limit. | |
2748 | * Allocations for MAP_NORESERVE mappings also need to be | |
2749 | * checked against any subpool limit. | |
2750 | */ | |
2751 | if (map_chg || avoid_reserve) { | |
2752 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2753 | if (gbl_chg < 0) { | |
feba16e2 | 2754 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2755 | return ERR_PTR(-ENOSPC); |
5e911373 | 2756 | } |
1da177e4 | 2757 | |
d85f69b0 MK |
2758 | /* |
2759 | * Even though there was no reservation in the region/reserve | |
2760 | * map, there could be reservations associated with the | |
2761 | * subpool that can be used. This would be indicated if the | |
2762 | * return value of hugepage_subpool_get_pages() is zero. | |
2763 | * However, if avoid_reserve is specified we still avoid even | |
2764 | * the subpool reservations. | |
2765 | */ | |
2766 | if (avoid_reserve) | |
2767 | gbl_chg = 1; | |
2768 | } | |
2769 | ||
08cf9faf MA |
2770 | /* If this allocation is not consuming a reservation, charge it now. |
2771 | */ | |
6501fe5f | 2772 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2773 | if (deferred_reserve) { |
2774 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2775 | idx, pages_per_huge_page(h), &h_cg); | |
2776 | if (ret) | |
2777 | goto out_subpool_put; | |
2778 | } | |
2779 | ||
6d76dcf4 | 2780 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2781 | if (ret) |
08cf9faf | 2782 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2783 | |
db71ef79 | 2784 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
2785 | /* |
2786 | * glb_chg is passed to indicate whether or not a page must be taken | |
2787 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2788 | * a reservation exists for the allocation. | |
2789 | */ | |
2790 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2791 | if (!page) { |
db71ef79 | 2792 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 2793 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2794 | if (!page) |
2795 | goto out_uncharge_cgroup; | |
a88c7695 | 2796 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2797 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2798 | h->resv_huge_pages--; |
2799 | } | |
db71ef79 | 2800 | spin_lock_irq(&hugetlb_lock); |
15a8d68e | 2801 | list_add(&page->lru, &h->hugepage_activelist); |
81a6fcae | 2802 | /* Fall through */ |
68842c9b | 2803 | } |
81a6fcae | 2804 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2805 | /* If allocation is not consuming a reservation, also store the |
2806 | * hugetlb_cgroup pointer on the page. | |
2807 | */ | |
2808 | if (deferred_reserve) { | |
2809 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2810 | h_cg, page); | |
2811 | } | |
2812 | ||
db71ef79 | 2813 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 2814 | |
d6995da3 | 2815 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2816 | |
d85f69b0 MK |
2817 | map_commit = vma_commit_reservation(h, vma, addr); |
2818 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2819 | /* |
2820 | * The page was added to the reservation map between | |
2821 | * vma_needs_reservation and vma_commit_reservation. | |
2822 | * This indicates a race with hugetlb_reserve_pages. | |
2823 | * Adjust for the subpool count incremented above AND | |
2824 | * in hugetlb_reserve_pages for the same page. Also, | |
2825 | * the reservation count added in hugetlb_reserve_pages | |
2826 | * no longer applies. | |
2827 | */ | |
2828 | long rsv_adjust; | |
2829 | ||
2830 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2831 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2832 | if (deferred_reserve) |
2833 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2834 | pages_per_huge_page(h), page); | |
33039678 | 2835 | } |
90d8b7e6 | 2836 | return page; |
8f34af6f JZ |
2837 | |
2838 | out_uncharge_cgroup: | |
2839 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2840 | out_uncharge_cgroup_reservation: |
2841 | if (deferred_reserve) | |
2842 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2843 | h_cg); | |
8f34af6f | 2844 | out_subpool_put: |
d85f69b0 | 2845 | if (map_chg || avoid_reserve) |
8f34af6f | 2846 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2847 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2848 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2849 | } |
2850 | ||
e24a1307 AK |
2851 | int alloc_bootmem_huge_page(struct hstate *h) |
2852 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2853 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2854 | { |
2855 | struct huge_bootmem_page *m; | |
b2261026 | 2856 | int nr_nodes, node; |
aa888a74 | 2857 | |
b2261026 | 2858 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2859 | void *addr; |
2860 | ||
eb31d559 | 2861 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2862 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2863 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2864 | if (addr) { |
2865 | /* | |
2866 | * Use the beginning of the huge page to store the | |
2867 | * huge_bootmem_page struct (until gather_bootmem | |
2868 | * puts them into the mem_map). | |
2869 | */ | |
2870 | m = addr; | |
91f47662 | 2871 | goto found; |
aa888a74 | 2872 | } |
aa888a74 AK |
2873 | } |
2874 | return 0; | |
2875 | ||
2876 | found: | |
df994ead | 2877 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2878 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2879 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2880 | list_add(&m->list, &huge_boot_pages); |
2881 | m->hstate = h; | |
2882 | return 1; | |
2883 | } | |
2884 | ||
48b8d744 MK |
2885 | /* |
2886 | * Put bootmem huge pages into the standard lists after mem_map is up. | |
2887 | * Note: This only applies to gigantic (order > MAX_ORDER) pages. | |
2888 | */ | |
aa888a74 AK |
2889 | static void __init gather_bootmem_prealloc(void) |
2890 | { | |
2891 | struct huge_bootmem_page *m; | |
2892 | ||
2893 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2894 | struct page *page = virt_to_page(m); |
aa888a74 | 2895 | struct hstate *h = m->hstate; |
ee8f248d | 2896 | |
48b8d744 | 2897 | VM_BUG_ON(!hstate_is_gigantic(h)); |
aa888a74 | 2898 | WARN_ON(page_count(page) != 1); |
7118fc29 MK |
2899 | if (prep_compound_gigantic_page(page, huge_page_order(h))) { |
2900 | WARN_ON(PageReserved(page)); | |
2901 | prep_new_huge_page(h, page, page_to_nid(page)); | |
2902 | put_page(page); /* add to the hugepage allocator */ | |
2903 | } else { | |
416d85ed | 2904 | /* VERY unlikely inflated ref count on a tail page */ |
7118fc29 | 2905 | free_gigantic_page(page, huge_page_order(h)); |
7118fc29 | 2906 | } |
af0fb9df | 2907 | |
b0320c7b | 2908 | /* |
48b8d744 MK |
2909 | * We need to restore the 'stolen' pages to totalram_pages |
2910 | * in order to fix confusing memory reports from free(1) and | |
2911 | * other side-effects, like CommitLimit going negative. | |
b0320c7b | 2912 | */ |
48b8d744 | 2913 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 2914 | cond_resched(); |
aa888a74 AK |
2915 | } |
2916 | } | |
2917 | ||
8faa8b07 | 2918 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2919 | { |
2920 | unsigned long i; | |
f60858f9 MK |
2921 | nodemask_t *node_alloc_noretry; |
2922 | ||
2923 | if (!hstate_is_gigantic(h)) { | |
2924 | /* | |
2925 | * Bit mask controlling how hard we retry per-node allocations. | |
2926 | * Ignore errors as lower level routines can deal with | |
2927 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2928 | * time, we are likely in bigger trouble. | |
2929 | */ | |
2930 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2931 | GFP_KERNEL); | |
2932 | } else { | |
2933 | /* allocations done at boot time */ | |
2934 | node_alloc_noretry = NULL; | |
2935 | } | |
2936 | ||
2937 | /* bit mask controlling how hard we retry per-node allocations */ | |
2938 | if (node_alloc_noretry) | |
2939 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2940 | |
e5ff2159 | 2941 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2942 | if (hstate_is_gigantic(h)) { |
dbda8fea | 2943 | if (hugetlb_cma_size) { |
cf11e85f | 2944 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); |
7ecc9565 | 2945 | goto free; |
cf11e85f | 2946 | } |
aa888a74 AK |
2947 | if (!alloc_bootmem_huge_page(h)) |
2948 | break; | |
0c397dae | 2949 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2950 | &node_states[N_MEMORY], |
2951 | node_alloc_noretry)) | |
1da177e4 | 2952 | break; |
69ed779a | 2953 | cond_resched(); |
1da177e4 | 2954 | } |
d715cf80 LH |
2955 | if (i < h->max_huge_pages) { |
2956 | char buf[32]; | |
2957 | ||
c6247f72 | 2958 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2959 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2960 | h->max_huge_pages, buf, i); | |
2961 | h->max_huge_pages = i; | |
2962 | } | |
7ecc9565 | 2963 | free: |
f60858f9 | 2964 | kfree(node_alloc_noretry); |
e5ff2159 AK |
2965 | } |
2966 | ||
2967 | static void __init hugetlb_init_hstates(void) | |
2968 | { | |
2969 | struct hstate *h; | |
2970 | ||
2971 | for_each_hstate(h) { | |
641844f5 NH |
2972 | if (minimum_order > huge_page_order(h)) |
2973 | minimum_order = huge_page_order(h); | |
2974 | ||
8faa8b07 | 2975 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2976 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2977 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2978 | } |
641844f5 | 2979 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2980 | } |
2981 | ||
2982 | static void __init report_hugepages(void) | |
2983 | { | |
2984 | struct hstate *h; | |
2985 | ||
2986 | for_each_hstate(h) { | |
4abd32db | 2987 | char buf[32]; |
c6247f72 MW |
2988 | |
2989 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2990 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2991 | buf, h->free_huge_pages); |
e5ff2159 AK |
2992 | } |
2993 | } | |
2994 | ||
1da177e4 | 2995 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2996 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2997 | nodemask_t *nodes_allowed) | |
1da177e4 | 2998 | { |
4415cc8d | 2999 | int i; |
1121828a | 3000 | LIST_HEAD(page_list); |
4415cc8d | 3001 | |
9487ca60 | 3002 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 3003 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
3004 | return; |
3005 | ||
1121828a MK |
3006 | /* |
3007 | * Collect pages to be freed on a list, and free after dropping lock | |
3008 | */ | |
6ae11b27 | 3009 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 3010 | struct page *page, *next; |
a5516438 AK |
3011 | struct list_head *freel = &h->hugepage_freelists[i]; |
3012 | list_for_each_entry_safe(page, next, freel, lru) { | |
3013 | if (count >= h->nr_huge_pages) | |
1121828a | 3014 | goto out; |
1da177e4 LT |
3015 | if (PageHighMem(page)) |
3016 | continue; | |
6eb4e88a | 3017 | remove_hugetlb_page(h, page, false); |
1121828a | 3018 | list_add(&page->lru, &page_list); |
1da177e4 LT |
3019 | } |
3020 | } | |
1121828a MK |
3021 | |
3022 | out: | |
db71ef79 | 3023 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3024 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 3025 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
3026 | } |
3027 | #else | |
6ae11b27 LS |
3028 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
3029 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
3030 | { |
3031 | } | |
3032 | #endif | |
3033 | ||
20a0307c WF |
3034 | /* |
3035 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
3036 | * balanced by operating on them in a round-robin fashion. | |
3037 | * Returns 1 if an adjustment was made. | |
3038 | */ | |
6ae11b27 LS |
3039 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
3040 | int delta) | |
20a0307c | 3041 | { |
b2261026 | 3042 | int nr_nodes, node; |
20a0307c | 3043 | |
9487ca60 | 3044 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 3045 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 3046 | |
b2261026 JK |
3047 | if (delta < 0) { |
3048 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
3049 | if (h->surplus_huge_pages_node[node]) | |
3050 | goto found; | |
e8c5c824 | 3051 | } |
b2261026 JK |
3052 | } else { |
3053 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
3054 | if (h->surplus_huge_pages_node[node] < | |
3055 | h->nr_huge_pages_node[node]) | |
3056 | goto found; | |
e8c5c824 | 3057 | } |
b2261026 JK |
3058 | } |
3059 | return 0; | |
20a0307c | 3060 | |
b2261026 JK |
3061 | found: |
3062 | h->surplus_huge_pages += delta; | |
3063 | h->surplus_huge_pages_node[node] += delta; | |
3064 | return 1; | |
20a0307c WF |
3065 | } |
3066 | ||
a5516438 | 3067 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 3068 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 3069 | nodemask_t *nodes_allowed) |
1da177e4 | 3070 | { |
7893d1d5 | 3071 | unsigned long min_count, ret; |
10c6ec49 MK |
3072 | struct page *page; |
3073 | LIST_HEAD(page_list); | |
f60858f9 MK |
3074 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
3075 | ||
3076 | /* | |
3077 | * Bit mask controlling how hard we retry per-node allocations. | |
3078 | * If we can not allocate the bit mask, do not attempt to allocate | |
3079 | * the requested huge pages. | |
3080 | */ | |
3081 | if (node_alloc_noretry) | |
3082 | nodes_clear(*node_alloc_noretry); | |
3083 | else | |
3084 | return -ENOMEM; | |
1da177e4 | 3085 | |
29383967 MK |
3086 | /* |
3087 | * resize_lock mutex prevents concurrent adjustments to number of | |
3088 | * pages in hstate via the proc/sysfs interfaces. | |
3089 | */ | |
3090 | mutex_lock(&h->resize_lock); | |
b65d4adb | 3091 | flush_free_hpage_work(h); |
db71ef79 | 3092 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 3093 | |
fd875dca MK |
3094 | /* |
3095 | * Check for a node specific request. | |
3096 | * Changing node specific huge page count may require a corresponding | |
3097 | * change to the global count. In any case, the passed node mask | |
3098 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
3099 | */ | |
3100 | if (nid != NUMA_NO_NODE) { | |
3101 | unsigned long old_count = count; | |
3102 | ||
3103 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
3104 | /* | |
3105 | * User may have specified a large count value which caused the | |
3106 | * above calculation to overflow. In this case, they wanted | |
3107 | * to allocate as many huge pages as possible. Set count to | |
3108 | * largest possible value to align with their intention. | |
3109 | */ | |
3110 | if (count < old_count) | |
3111 | count = ULONG_MAX; | |
3112 | } | |
3113 | ||
4eb0716e AG |
3114 | /* |
3115 | * Gigantic pages runtime allocation depend on the capability for large | |
3116 | * page range allocation. | |
3117 | * If the system does not provide this feature, return an error when | |
3118 | * the user tries to allocate gigantic pages but let the user free the | |
3119 | * boottime allocated gigantic pages. | |
3120 | */ | |
3121 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
3122 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 3123 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3124 | mutex_unlock(&h->resize_lock); |
f60858f9 | 3125 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
3126 | return -EINVAL; |
3127 | } | |
3128 | /* Fall through to decrease pool */ | |
3129 | } | |
aa888a74 | 3130 | |
7893d1d5 AL |
3131 | /* |
3132 | * Increase the pool size | |
3133 | * First take pages out of surplus state. Then make up the | |
3134 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 3135 | * |
0c397dae | 3136 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
3137 | * to convert a surplus huge page to a normal huge page. That is |
3138 | * not critical, though, it just means the overall size of the | |
3139 | * pool might be one hugepage larger than it needs to be, but | |
3140 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 3141 | */ |
a5516438 | 3142 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 3143 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
3144 | break; |
3145 | } | |
3146 | ||
a5516438 | 3147 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
3148 | /* |
3149 | * If this allocation races such that we no longer need the | |
3150 | * page, free_huge_page will handle it by freeing the page | |
3151 | * and reducing the surplus. | |
3152 | */ | |
db71ef79 | 3153 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
3154 | |
3155 | /* yield cpu to avoid soft lockup */ | |
3156 | cond_resched(); | |
3157 | ||
f60858f9 MK |
3158 | ret = alloc_pool_huge_page(h, nodes_allowed, |
3159 | node_alloc_noretry); | |
db71ef79 | 3160 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
3161 | if (!ret) |
3162 | goto out; | |
3163 | ||
536240f2 MG |
3164 | /* Bail for signals. Probably ctrl-c from user */ |
3165 | if (signal_pending(current)) | |
3166 | goto out; | |
7893d1d5 | 3167 | } |
7893d1d5 AL |
3168 | |
3169 | /* | |
3170 | * Decrease the pool size | |
3171 | * First return free pages to the buddy allocator (being careful | |
3172 | * to keep enough around to satisfy reservations). Then place | |
3173 | * pages into surplus state as needed so the pool will shrink | |
3174 | * to the desired size as pages become free. | |
d1c3fb1f NA |
3175 | * |
3176 | * By placing pages into the surplus state independent of the | |
3177 | * overcommit value, we are allowing the surplus pool size to | |
3178 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 3179 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
3180 | * though, we'll note that we're not allowed to exceed surplus |
3181 | * and won't grow the pool anywhere else. Not until one of the | |
3182 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 3183 | */ |
a5516438 | 3184 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 3185 | min_count = max(count, min_count); |
6ae11b27 | 3186 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
3187 | |
3188 | /* | |
3189 | * Collect pages to be removed on list without dropping lock | |
3190 | */ | |
a5516438 | 3191 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
3192 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
3193 | if (!page) | |
1da177e4 | 3194 | break; |
10c6ec49 MK |
3195 | |
3196 | list_add(&page->lru, &page_list); | |
1da177e4 | 3197 | } |
10c6ec49 | 3198 | /* free the pages after dropping lock */ |
db71ef79 | 3199 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 3200 | update_and_free_pages_bulk(h, &page_list); |
b65d4adb | 3201 | flush_free_hpage_work(h); |
db71ef79 | 3202 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 3203 | |
a5516438 | 3204 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 3205 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
3206 | break; |
3207 | } | |
3208 | out: | |
4eb0716e | 3209 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 3210 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 3211 | mutex_unlock(&h->resize_lock); |
4eb0716e | 3212 | |
f60858f9 MK |
3213 | NODEMASK_FREE(node_alloc_noretry); |
3214 | ||
4eb0716e | 3215 | return 0; |
1da177e4 LT |
3216 | } |
3217 | ||
a3437870 NA |
3218 | #define HSTATE_ATTR_RO(_name) \ |
3219 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
3220 | ||
3221 | #define HSTATE_ATTR(_name) \ | |
3222 | static struct kobj_attribute _name##_attr = \ | |
3223 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
3224 | ||
3225 | static struct kobject *hugepages_kobj; | |
3226 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3227 | ||
9a305230 LS |
3228 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
3229 | ||
3230 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
3231 | { |
3232 | int i; | |
9a305230 | 3233 | |
a3437870 | 3234 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
3235 | if (hstate_kobjs[i] == kobj) { |
3236 | if (nidp) | |
3237 | *nidp = NUMA_NO_NODE; | |
a3437870 | 3238 | return &hstates[i]; |
9a305230 LS |
3239 | } |
3240 | ||
3241 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
3242 | } |
3243 | ||
06808b08 | 3244 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
3245 | struct kobj_attribute *attr, char *buf) |
3246 | { | |
9a305230 LS |
3247 | struct hstate *h; |
3248 | unsigned long nr_huge_pages; | |
3249 | int nid; | |
3250 | ||
3251 | h = kobj_to_hstate(kobj, &nid); | |
3252 | if (nid == NUMA_NO_NODE) | |
3253 | nr_huge_pages = h->nr_huge_pages; | |
3254 | else | |
3255 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
3256 | ||
ae7a927d | 3257 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 3258 | } |
adbe8726 | 3259 | |
238d3c13 DR |
3260 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
3261 | struct hstate *h, int nid, | |
3262 | unsigned long count, size_t len) | |
a3437870 NA |
3263 | { |
3264 | int err; | |
2d0adf7e | 3265 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 3266 | |
2d0adf7e OS |
3267 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
3268 | return -EINVAL; | |
adbe8726 | 3269 | |
9a305230 LS |
3270 | if (nid == NUMA_NO_NODE) { |
3271 | /* | |
3272 | * global hstate attribute | |
3273 | */ | |
3274 | if (!(obey_mempolicy && | |
2d0adf7e OS |
3275 | init_nodemask_of_mempolicy(&nodes_allowed))) |
3276 | n_mask = &node_states[N_MEMORY]; | |
3277 | else | |
3278 | n_mask = &nodes_allowed; | |
3279 | } else { | |
9a305230 | 3280 | /* |
fd875dca MK |
3281 | * Node specific request. count adjustment happens in |
3282 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 3283 | */ |
2d0adf7e OS |
3284 | init_nodemask_of_node(&nodes_allowed, nid); |
3285 | n_mask = &nodes_allowed; | |
fd875dca | 3286 | } |
9a305230 | 3287 | |
2d0adf7e | 3288 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 3289 | |
4eb0716e | 3290 | return err ? err : len; |
06808b08 LS |
3291 | } |
3292 | ||
238d3c13 DR |
3293 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
3294 | struct kobject *kobj, const char *buf, | |
3295 | size_t len) | |
3296 | { | |
3297 | struct hstate *h; | |
3298 | unsigned long count; | |
3299 | int nid; | |
3300 | int err; | |
3301 | ||
3302 | err = kstrtoul(buf, 10, &count); | |
3303 | if (err) | |
3304 | return err; | |
3305 | ||
3306 | h = kobj_to_hstate(kobj, &nid); | |
3307 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
3308 | } | |
3309 | ||
06808b08 LS |
3310 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
3311 | struct kobj_attribute *attr, char *buf) | |
3312 | { | |
3313 | return nr_hugepages_show_common(kobj, attr, buf); | |
3314 | } | |
3315 | ||
3316 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
3317 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3318 | { | |
238d3c13 | 3319 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
3320 | } |
3321 | HSTATE_ATTR(nr_hugepages); | |
3322 | ||
06808b08 LS |
3323 | #ifdef CONFIG_NUMA |
3324 | ||
3325 | /* | |
3326 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
3327 | * huge page alloc/free. | |
3328 | */ | |
3329 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
3330 | struct kobj_attribute *attr, |
3331 | char *buf) | |
06808b08 LS |
3332 | { |
3333 | return nr_hugepages_show_common(kobj, attr, buf); | |
3334 | } | |
3335 | ||
3336 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
3337 | struct kobj_attribute *attr, const char *buf, size_t len) | |
3338 | { | |
238d3c13 | 3339 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
3340 | } |
3341 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
3342 | #endif | |
3343 | ||
3344 | ||
a3437870 NA |
3345 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
3346 | struct kobj_attribute *attr, char *buf) | |
3347 | { | |
9a305230 | 3348 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3349 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 3350 | } |
adbe8726 | 3351 | |
a3437870 NA |
3352 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
3353 | struct kobj_attribute *attr, const char *buf, size_t count) | |
3354 | { | |
3355 | int err; | |
3356 | unsigned long input; | |
9a305230 | 3357 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 3358 | |
bae7f4ae | 3359 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
3360 | return -EINVAL; |
3361 | ||
3dbb95f7 | 3362 | err = kstrtoul(buf, 10, &input); |
a3437870 | 3363 | if (err) |
73ae31e5 | 3364 | return err; |
a3437870 | 3365 | |
db71ef79 | 3366 | spin_lock_irq(&hugetlb_lock); |
a3437870 | 3367 | h->nr_overcommit_huge_pages = input; |
db71ef79 | 3368 | spin_unlock_irq(&hugetlb_lock); |
a3437870 NA |
3369 | |
3370 | return count; | |
3371 | } | |
3372 | HSTATE_ATTR(nr_overcommit_hugepages); | |
3373 | ||
3374 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
3375 | struct kobj_attribute *attr, char *buf) | |
3376 | { | |
9a305230 LS |
3377 | struct hstate *h; |
3378 | unsigned long free_huge_pages; | |
3379 | int nid; | |
3380 | ||
3381 | h = kobj_to_hstate(kobj, &nid); | |
3382 | if (nid == NUMA_NO_NODE) | |
3383 | free_huge_pages = h->free_huge_pages; | |
3384 | else | |
3385 | free_huge_pages = h->free_huge_pages_node[nid]; | |
3386 | ||
ae7a927d | 3387 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
3388 | } |
3389 | HSTATE_ATTR_RO(free_hugepages); | |
3390 | ||
3391 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
3392 | struct kobj_attribute *attr, char *buf) | |
3393 | { | |
9a305230 | 3394 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 3395 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
3396 | } |
3397 | HSTATE_ATTR_RO(resv_hugepages); | |
3398 | ||
3399 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
3400 | struct kobj_attribute *attr, char *buf) | |
3401 | { | |
9a305230 LS |
3402 | struct hstate *h; |
3403 | unsigned long surplus_huge_pages; | |
3404 | int nid; | |
3405 | ||
3406 | h = kobj_to_hstate(kobj, &nid); | |
3407 | if (nid == NUMA_NO_NODE) | |
3408 | surplus_huge_pages = h->surplus_huge_pages; | |
3409 | else | |
3410 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
3411 | ||
ae7a927d | 3412 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
3413 | } |
3414 | HSTATE_ATTR_RO(surplus_hugepages); | |
3415 | ||
3416 | static struct attribute *hstate_attrs[] = { | |
3417 | &nr_hugepages_attr.attr, | |
3418 | &nr_overcommit_hugepages_attr.attr, | |
3419 | &free_hugepages_attr.attr, | |
3420 | &resv_hugepages_attr.attr, | |
3421 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3422 | #ifdef CONFIG_NUMA |
3423 | &nr_hugepages_mempolicy_attr.attr, | |
3424 | #endif | |
a3437870 NA |
3425 | NULL, |
3426 | }; | |
3427 | ||
67e5ed96 | 3428 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3429 | .attrs = hstate_attrs, |
3430 | }; | |
3431 | ||
094e9539 JM |
3432 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3433 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3434 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3435 | { |
3436 | int retval; | |
972dc4de | 3437 | int hi = hstate_index(h); |
a3437870 | 3438 | |
9a305230 LS |
3439 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3440 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3441 | return -ENOMEM; |
3442 | ||
9a305230 | 3443 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 3444 | if (retval) { |
9a305230 | 3445 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 ML |
3446 | hstate_kobjs[hi] = NULL; |
3447 | } | |
a3437870 NA |
3448 | |
3449 | return retval; | |
3450 | } | |
3451 | ||
3452 | static void __init hugetlb_sysfs_init(void) | |
3453 | { | |
3454 | struct hstate *h; | |
3455 | int err; | |
3456 | ||
3457 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3458 | if (!hugepages_kobj) | |
3459 | return; | |
3460 | ||
3461 | for_each_hstate(h) { | |
9a305230 LS |
3462 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3463 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3464 | if (err) |
282f4214 | 3465 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
3466 | } |
3467 | } | |
3468 | ||
9a305230 LS |
3469 | #ifdef CONFIG_NUMA |
3470 | ||
3471 | /* | |
3472 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3473 | * with node devices in node_devices[] using a parallel array. The array |
3474 | * index of a node device or _hstate == node id. | |
3475 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3476 | * the base kernel, on the hugetlb module. |
3477 | */ | |
3478 | struct node_hstate { | |
3479 | struct kobject *hugepages_kobj; | |
3480 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3481 | }; | |
b4e289a6 | 3482 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3483 | |
3484 | /* | |
10fbcf4c | 3485 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3486 | */ |
3487 | static struct attribute *per_node_hstate_attrs[] = { | |
3488 | &nr_hugepages_attr.attr, | |
3489 | &free_hugepages_attr.attr, | |
3490 | &surplus_hugepages_attr.attr, | |
3491 | NULL, | |
3492 | }; | |
3493 | ||
67e5ed96 | 3494 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3495 | .attrs = per_node_hstate_attrs, |
3496 | }; | |
3497 | ||
3498 | /* | |
10fbcf4c | 3499 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3500 | * Returns node id via non-NULL nidp. |
3501 | */ | |
3502 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3503 | { | |
3504 | int nid; | |
3505 | ||
3506 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3507 | struct node_hstate *nhs = &node_hstates[nid]; | |
3508 | int i; | |
3509 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3510 | if (nhs->hstate_kobjs[i] == kobj) { | |
3511 | if (nidp) | |
3512 | *nidp = nid; | |
3513 | return &hstates[i]; | |
3514 | } | |
3515 | } | |
3516 | ||
3517 | BUG(); | |
3518 | return NULL; | |
3519 | } | |
3520 | ||
3521 | /* | |
10fbcf4c | 3522 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3523 | * No-op if no hstate attributes attached. |
3524 | */ | |
3cd8b44f | 3525 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3526 | { |
3527 | struct hstate *h; | |
10fbcf4c | 3528 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3529 | |
3530 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3531 | return; /* no hstate attributes */ |
9a305230 | 3532 | |
972dc4de AK |
3533 | for_each_hstate(h) { |
3534 | int idx = hstate_index(h); | |
3535 | if (nhs->hstate_kobjs[idx]) { | |
3536 | kobject_put(nhs->hstate_kobjs[idx]); | |
3537 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3538 | } |
972dc4de | 3539 | } |
9a305230 LS |
3540 | |
3541 | kobject_put(nhs->hugepages_kobj); | |
3542 | nhs->hugepages_kobj = NULL; | |
3543 | } | |
3544 | ||
9a305230 LS |
3545 | |
3546 | /* | |
10fbcf4c | 3547 | * Register hstate attributes for a single node device. |
9a305230 LS |
3548 | * No-op if attributes already registered. |
3549 | */ | |
3cd8b44f | 3550 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3551 | { |
3552 | struct hstate *h; | |
10fbcf4c | 3553 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3554 | int err; |
3555 | ||
3556 | if (nhs->hugepages_kobj) | |
3557 | return; /* already allocated */ | |
3558 | ||
3559 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3560 | &node->dev.kobj); |
9a305230 LS |
3561 | if (!nhs->hugepages_kobj) |
3562 | return; | |
3563 | ||
3564 | for_each_hstate(h) { | |
3565 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3566 | nhs->hstate_kobjs, | |
3567 | &per_node_hstate_attr_group); | |
3568 | if (err) { | |
282f4214 | 3569 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3570 | h->name, node->dev.id); |
9a305230 LS |
3571 | hugetlb_unregister_node(node); |
3572 | break; | |
3573 | } | |
3574 | } | |
3575 | } | |
3576 | ||
3577 | /* | |
9b5e5d0f | 3578 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3579 | * devices of nodes that have memory. All on-line nodes should have |
3580 | * registered their associated device by this time. | |
9a305230 | 3581 | */ |
7d9ca000 | 3582 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3583 | { |
3584 | int nid; | |
3585 | ||
8cebfcd0 | 3586 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3587 | struct node *node = node_devices[nid]; |
10fbcf4c | 3588 | if (node->dev.id == nid) |
9a305230 LS |
3589 | hugetlb_register_node(node); |
3590 | } | |
3591 | ||
3592 | /* | |
10fbcf4c | 3593 | * Let the node device driver know we're here so it can |
9a305230 LS |
3594 | * [un]register hstate attributes on node hotplug. |
3595 | */ | |
3596 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3597 | hugetlb_unregister_node); | |
3598 | } | |
3599 | #else /* !CONFIG_NUMA */ | |
3600 | ||
3601 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3602 | { | |
3603 | BUG(); | |
3604 | if (nidp) | |
3605 | *nidp = -1; | |
3606 | return NULL; | |
3607 | } | |
3608 | ||
9a305230 LS |
3609 | static void hugetlb_register_all_nodes(void) { } |
3610 | ||
3611 | #endif | |
3612 | ||
a3437870 NA |
3613 | static int __init hugetlb_init(void) |
3614 | { | |
8382d914 DB |
3615 | int i; |
3616 | ||
d6995da3 MK |
3617 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
3618 | __NR_HPAGEFLAGS); | |
3619 | ||
c2833a5b MK |
3620 | if (!hugepages_supported()) { |
3621 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
3622 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 3623 | return 0; |
c2833a5b | 3624 | } |
a3437870 | 3625 | |
282f4214 MK |
3626 | /* |
3627 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
3628 | * architectures depend on setup being done here. | |
3629 | */ | |
3630 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
3631 | if (!parsed_default_hugepagesz) { | |
3632 | /* | |
3633 | * If we did not parse a default huge page size, set | |
3634 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
3635 | * number of huge pages for this default size was implicitly | |
3636 | * specified, set that here as well. | |
3637 | * Note that the implicit setting will overwrite an explicit | |
3638 | * setting. A warning will be printed in this case. | |
3639 | */ | |
3640 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
3641 | if (default_hstate_max_huge_pages) { | |
3642 | if (default_hstate.max_huge_pages) { | |
3643 | char buf[32]; | |
3644 | ||
3645 | string_get_size(huge_page_size(&default_hstate), | |
3646 | 1, STRING_UNITS_2, buf, 32); | |
3647 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
3648 | default_hstate.max_huge_pages, buf); | |
3649 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
3650 | default_hstate_max_huge_pages); | |
3651 | } | |
3652 | default_hstate.max_huge_pages = | |
3653 | default_hstate_max_huge_pages; | |
d715cf80 | 3654 | } |
f8b74815 | 3655 | } |
a3437870 | 3656 | |
cf11e85f | 3657 | hugetlb_cma_check(); |
a3437870 | 3658 | hugetlb_init_hstates(); |
aa888a74 | 3659 | gather_bootmem_prealloc(); |
a3437870 NA |
3660 | report_hugepages(); |
3661 | ||
3662 | hugetlb_sysfs_init(); | |
9a305230 | 3663 | hugetlb_register_all_nodes(); |
7179e7bf | 3664 | hugetlb_cgroup_file_init(); |
9a305230 | 3665 | |
8382d914 DB |
3666 | #ifdef CONFIG_SMP |
3667 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3668 | #else | |
3669 | num_fault_mutexes = 1; | |
3670 | #endif | |
c672c7f2 | 3671 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3672 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3673 | GFP_KERNEL); | |
c672c7f2 | 3674 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3675 | |
3676 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3677 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3678 | return 0; |
3679 | } | |
3e89e1c5 | 3680 | subsys_initcall(hugetlb_init); |
a3437870 | 3681 | |
ae94da89 MK |
3682 | /* Overwritten by architectures with more huge page sizes */ |
3683 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 3684 | { |
ae94da89 | 3685 | return size == HPAGE_SIZE; |
9fee021d VT |
3686 | } |
3687 | ||
d00181b9 | 3688 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3689 | { |
3690 | struct hstate *h; | |
8faa8b07 AK |
3691 | unsigned long i; |
3692 | ||
a3437870 | 3693 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
3694 | return; |
3695 | } | |
47d38344 | 3696 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3697 | BUG_ON(order == 0); |
47d38344 | 3698 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 3699 | mutex_init(&h->resize_lock); |
a3437870 | 3700 | h->order = order; |
aca78307 | 3701 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
3702 | for (i = 0; i < MAX_NUMNODES; ++i) |
3703 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3704 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3705 | h->next_nid_to_alloc = first_memory_node; |
3706 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3707 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3708 | huge_page_size(h)/1024); | |
77490587 | 3709 | hugetlb_vmemmap_init(h); |
8faa8b07 | 3710 | |
a3437870 NA |
3711 | parsed_hstate = h; |
3712 | } | |
3713 | ||
282f4214 MK |
3714 | /* |
3715 | * hugepages command line processing | |
3716 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
3717 | * specification. If not, ignore the hugepages value. hugepages can also | |
3718 | * be the first huge page command line option in which case it implicitly | |
3719 | * specifies the number of huge pages for the default size. | |
3720 | */ | |
3721 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
3722 | { |
3723 | unsigned long *mhp; | |
8faa8b07 | 3724 | static unsigned long *last_mhp; |
a3437870 | 3725 | |
9fee021d | 3726 | if (!parsed_valid_hugepagesz) { |
282f4214 | 3727 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 3728 | parsed_valid_hugepagesz = true; |
282f4214 | 3729 | return 0; |
9fee021d | 3730 | } |
282f4214 | 3731 | |
a3437870 | 3732 | /* |
282f4214 MK |
3733 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
3734 | * yet, so this hugepages= parameter goes to the "default hstate". | |
3735 | * Otherwise, it goes with the previously parsed hugepagesz or | |
3736 | * default_hugepagesz. | |
a3437870 | 3737 | */ |
9fee021d | 3738 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3739 | mhp = &default_hstate_max_huge_pages; |
3740 | else | |
3741 | mhp = &parsed_hstate->max_huge_pages; | |
3742 | ||
8faa8b07 | 3743 | if (mhp == last_mhp) { |
282f4214 MK |
3744 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
3745 | return 0; | |
8faa8b07 AK |
3746 | } |
3747 | ||
a3437870 NA |
3748 | if (sscanf(s, "%lu", mhp) <= 0) |
3749 | *mhp = 0; | |
3750 | ||
8faa8b07 AK |
3751 | /* |
3752 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 3753 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
3754 | * use the bootmem allocator. |
3755 | */ | |
04adbc3f | 3756 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
3757 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3758 | ||
3759 | last_mhp = mhp; | |
3760 | ||
a3437870 NA |
3761 | return 1; |
3762 | } | |
282f4214 | 3763 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 3764 | |
282f4214 MK |
3765 | /* |
3766 | * hugepagesz command line processing | |
3767 | * A specific huge page size can only be specified once with hugepagesz. | |
3768 | * hugepagesz is followed by hugepages on the command line. The global | |
3769 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
3770 | * hugepagesz argument was valid. | |
3771 | */ | |
359f2544 | 3772 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 3773 | { |
359f2544 | 3774 | unsigned long size; |
282f4214 MK |
3775 | struct hstate *h; |
3776 | ||
3777 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
3778 | size = (unsigned long)memparse(s, NULL); |
3779 | ||
3780 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3781 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
359f2544 MK |
3782 | return 0; |
3783 | } | |
3784 | ||
282f4214 MK |
3785 | h = size_to_hstate(size); |
3786 | if (h) { | |
3787 | /* | |
3788 | * hstate for this size already exists. This is normally | |
3789 | * an error, but is allowed if the existing hstate is the | |
3790 | * default hstate. More specifically, it is only allowed if | |
3791 | * the number of huge pages for the default hstate was not | |
3792 | * previously specified. | |
3793 | */ | |
3794 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
3795 | default_hstate.max_huge_pages) { | |
3796 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
3797 | return 0; | |
3798 | } | |
3799 | ||
3800 | /* | |
3801 | * No need to call hugetlb_add_hstate() as hstate already | |
3802 | * exists. But, do set parsed_hstate so that a following | |
3803 | * hugepages= parameter will be applied to this hstate. | |
3804 | */ | |
3805 | parsed_hstate = h; | |
3806 | parsed_valid_hugepagesz = true; | |
3807 | return 1; | |
38237830 MK |
3808 | } |
3809 | ||
359f2544 | 3810 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 3811 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
3812 | return 1; |
3813 | } | |
359f2544 MK |
3814 | __setup("hugepagesz=", hugepagesz_setup); |
3815 | ||
282f4214 MK |
3816 | /* |
3817 | * default_hugepagesz command line input | |
3818 | * Only one instance of default_hugepagesz allowed on command line. | |
3819 | */ | |
ae94da89 | 3820 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 3821 | { |
ae94da89 MK |
3822 | unsigned long size; |
3823 | ||
282f4214 | 3824 | parsed_valid_hugepagesz = false; |
282f4214 MK |
3825 | if (parsed_default_hugepagesz) { |
3826 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
3827 | return 0; | |
3828 | } | |
3829 | ||
ae94da89 MK |
3830 | size = (unsigned long)memparse(s, NULL); |
3831 | ||
3832 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3833 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
ae94da89 MK |
3834 | return 0; |
3835 | } | |
3836 | ||
282f4214 MK |
3837 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
3838 | parsed_valid_hugepagesz = true; | |
3839 | parsed_default_hugepagesz = true; | |
3840 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
3841 | ||
3842 | /* | |
3843 | * The number of default huge pages (for this size) could have been | |
3844 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
3845 | * then default_hstate_max_huge_pages is set. If the default huge | |
3846 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
3847 | * allocated here from bootmem allocator. | |
3848 | */ | |
3849 | if (default_hstate_max_huge_pages) { | |
3850 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3851 | if (hstate_is_gigantic(&default_hstate)) | |
3852 | hugetlb_hstate_alloc_pages(&default_hstate); | |
3853 | default_hstate_max_huge_pages = 0; | |
3854 | } | |
3855 | ||
e11bfbfc NP |
3856 | return 1; |
3857 | } | |
ae94da89 | 3858 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 3859 | |
8ca39e68 | 3860 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
3861 | { |
3862 | int node; | |
3863 | unsigned int nr = 0; | |
8ca39e68 MS |
3864 | nodemask_t *mpol_allowed; |
3865 | unsigned int *array = h->free_huge_pages_node; | |
3866 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
3867 | ||
3868 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 3869 | |
8ca39e68 | 3870 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
c93b0a99 | 3871 | if (!mpol_allowed || node_isset(node, *mpol_allowed)) |
8ca39e68 MS |
3872 | nr += array[node]; |
3873 | } | |
8a213460 NA |
3874 | |
3875 | return nr; | |
3876 | } | |
3877 | ||
3878 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
3879 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
3880 | void *buffer, size_t *length, | |
3881 | loff_t *ppos, unsigned long *out) | |
3882 | { | |
3883 | struct ctl_table dup_table; | |
3884 | ||
3885 | /* | |
3886 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
3887 | * can duplicate the @table and alter the duplicate of it. | |
3888 | */ | |
3889 | dup_table = *table; | |
3890 | dup_table.data = out; | |
3891 | ||
3892 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
3893 | } | |
3894 | ||
06808b08 LS |
3895 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3896 | struct ctl_table *table, int write, | |
32927393 | 3897 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 3898 | { |
e5ff2159 | 3899 | struct hstate *h = &default_hstate; |
238d3c13 | 3900 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3901 | int ret; |
e5ff2159 | 3902 | |
457c1b27 | 3903 | if (!hugepages_supported()) |
86613628 | 3904 | return -EOPNOTSUPP; |
457c1b27 | 3905 | |
17743798 MS |
3906 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3907 | &tmp); | |
08d4a246 MH |
3908 | if (ret) |
3909 | goto out; | |
e5ff2159 | 3910 | |
238d3c13 DR |
3911 | if (write) |
3912 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3913 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3914 | out: |
3915 | return ret; | |
1da177e4 | 3916 | } |
396faf03 | 3917 | |
06808b08 | 3918 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 3919 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3920 | { |
3921 | ||
3922 | return hugetlb_sysctl_handler_common(false, table, write, | |
3923 | buffer, length, ppos); | |
3924 | } | |
3925 | ||
3926 | #ifdef CONFIG_NUMA | |
3927 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 3928 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3929 | { |
3930 | return hugetlb_sysctl_handler_common(true, table, write, | |
3931 | buffer, length, ppos); | |
3932 | } | |
3933 | #endif /* CONFIG_NUMA */ | |
3934 | ||
a3d0c6aa | 3935 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 3936 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 3937 | { |
a5516438 | 3938 | struct hstate *h = &default_hstate; |
e5ff2159 | 3939 | unsigned long tmp; |
08d4a246 | 3940 | int ret; |
e5ff2159 | 3941 | |
457c1b27 | 3942 | if (!hugepages_supported()) |
86613628 | 3943 | return -EOPNOTSUPP; |
457c1b27 | 3944 | |
c033a93c | 3945 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3946 | |
bae7f4ae | 3947 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3948 | return -EINVAL; |
3949 | ||
17743798 MS |
3950 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3951 | &tmp); | |
08d4a246 MH |
3952 | if (ret) |
3953 | goto out; | |
e5ff2159 AK |
3954 | |
3955 | if (write) { | |
db71ef79 | 3956 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 3957 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 3958 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 3959 | } |
08d4a246 MH |
3960 | out: |
3961 | return ret; | |
a3d0c6aa NA |
3962 | } |
3963 | ||
1da177e4 LT |
3964 | #endif /* CONFIG_SYSCTL */ |
3965 | ||
e1759c21 | 3966 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3967 | { |
fcb2b0c5 RG |
3968 | struct hstate *h; |
3969 | unsigned long total = 0; | |
3970 | ||
457c1b27 NA |
3971 | if (!hugepages_supported()) |
3972 | return; | |
fcb2b0c5 RG |
3973 | |
3974 | for_each_hstate(h) { | |
3975 | unsigned long count = h->nr_huge_pages; | |
3976 | ||
aca78307 | 3977 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
3978 | |
3979 | if (h == &default_hstate) | |
3980 | seq_printf(m, | |
3981 | "HugePages_Total: %5lu\n" | |
3982 | "HugePages_Free: %5lu\n" | |
3983 | "HugePages_Rsvd: %5lu\n" | |
3984 | "HugePages_Surp: %5lu\n" | |
3985 | "Hugepagesize: %8lu kB\n", | |
3986 | count, | |
3987 | h->free_huge_pages, | |
3988 | h->resv_huge_pages, | |
3989 | h->surplus_huge_pages, | |
aca78307 | 3990 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
3991 | } |
3992 | ||
aca78307 | 3993 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
3994 | } |
3995 | ||
7981593b | 3996 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 3997 | { |
a5516438 | 3998 | struct hstate *h = &default_hstate; |
7981593b | 3999 | |
457c1b27 NA |
4000 | if (!hugepages_supported()) |
4001 | return 0; | |
7981593b JP |
4002 | |
4003 | return sysfs_emit_at(buf, len, | |
4004 | "Node %d HugePages_Total: %5u\n" | |
4005 | "Node %d HugePages_Free: %5u\n" | |
4006 | "Node %d HugePages_Surp: %5u\n", | |
4007 | nid, h->nr_huge_pages_node[nid], | |
4008 | nid, h->free_huge_pages_node[nid], | |
4009 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
4010 | } |
4011 | ||
949f7ec5 DR |
4012 | void hugetlb_show_meminfo(void) |
4013 | { | |
4014 | struct hstate *h; | |
4015 | int nid; | |
4016 | ||
457c1b27 NA |
4017 | if (!hugepages_supported()) |
4018 | return; | |
4019 | ||
949f7ec5 DR |
4020 | for_each_node_state(nid, N_MEMORY) |
4021 | for_each_hstate(h) | |
4022 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
4023 | nid, | |
4024 | h->nr_huge_pages_node[nid], | |
4025 | h->free_huge_pages_node[nid], | |
4026 | h->surplus_huge_pages_node[nid], | |
aca78307 | 4027 | huge_page_size(h) / SZ_1K); |
949f7ec5 DR |
4028 | } |
4029 | ||
5d317b2b NH |
4030 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
4031 | { | |
4032 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
4033 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
4034 | } | |
4035 | ||
1da177e4 LT |
4036 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
4037 | unsigned long hugetlb_total_pages(void) | |
4038 | { | |
d0028588 WL |
4039 | struct hstate *h; |
4040 | unsigned long nr_total_pages = 0; | |
4041 | ||
4042 | for_each_hstate(h) | |
4043 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
4044 | return nr_total_pages; | |
1da177e4 | 4045 | } |
1da177e4 | 4046 | |
a5516438 | 4047 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
4048 | { |
4049 | int ret = -ENOMEM; | |
4050 | ||
0aa7f354 ML |
4051 | if (!delta) |
4052 | return 0; | |
4053 | ||
db71ef79 | 4054 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4055 | /* |
4056 | * When cpuset is configured, it breaks the strict hugetlb page | |
4057 | * reservation as the accounting is done on a global variable. Such | |
4058 | * reservation is completely rubbish in the presence of cpuset because | |
4059 | * the reservation is not checked against page availability for the | |
4060 | * current cpuset. Application can still potentially OOM'ed by kernel | |
4061 | * with lack of free htlb page in cpuset that the task is in. | |
4062 | * Attempt to enforce strict accounting with cpuset is almost | |
4063 | * impossible (or too ugly) because cpuset is too fluid that | |
4064 | * task or memory node can be dynamically moved between cpusets. | |
4065 | * | |
4066 | * The change of semantics for shared hugetlb mapping with cpuset is | |
4067 | * undesirable. However, in order to preserve some of the semantics, | |
4068 | * we fall back to check against current free page availability as | |
4069 | * a best attempt and hopefully to minimize the impact of changing | |
4070 | * semantics that cpuset has. | |
8ca39e68 MS |
4071 | * |
4072 | * Apart from cpuset, we also have memory policy mechanism that | |
4073 | * also determines from which node the kernel will allocate memory | |
4074 | * in a NUMA system. So similar to cpuset, we also should consider | |
4075 | * the memory policy of the current task. Similar to the description | |
4076 | * above. | |
fc1b8a73 MG |
4077 | */ |
4078 | if (delta > 0) { | |
a5516438 | 4079 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
4080 | goto out; |
4081 | ||
8ca39e68 | 4082 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 4083 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
4084 | goto out; |
4085 | } | |
4086 | } | |
4087 | ||
4088 | ret = 0; | |
4089 | if (delta < 0) | |
a5516438 | 4090 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
4091 | |
4092 | out: | |
db71ef79 | 4093 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
4094 | return ret; |
4095 | } | |
4096 | ||
84afd99b AW |
4097 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
4098 | { | |
f522c3ac | 4099 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
4100 | |
4101 | /* | |
4102 | * This new VMA should share its siblings reservation map if present. | |
4103 | * The VMA will only ever have a valid reservation map pointer where | |
4104 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 4105 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
4106 | * after this open call completes. It is therefore safe to take a |
4107 | * new reference here without additional locking. | |
4108 | */ | |
09a26e83 MK |
4109 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
4110 | resv_map_dup_hugetlb_cgroup_uncharge_info(resv); | |
f522c3ac | 4111 | kref_get(&resv->refs); |
09a26e83 | 4112 | } |
84afd99b AW |
4113 | } |
4114 | ||
a1e78772 MG |
4115 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
4116 | { | |
a5516438 | 4117 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 4118 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 4119 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 4120 | unsigned long reserve, start, end; |
1c5ecae3 | 4121 | long gbl_reserve; |
84afd99b | 4122 | |
4e35f483 JK |
4123 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
4124 | return; | |
84afd99b | 4125 | |
4e35f483 JK |
4126 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
4127 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 4128 | |
4e35f483 | 4129 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 4130 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 4131 | if (reserve) { |
1c5ecae3 MK |
4132 | /* |
4133 | * Decrement reserve counts. The global reserve count may be | |
4134 | * adjusted if the subpool has a minimum size. | |
4135 | */ | |
4136 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
4137 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 4138 | } |
e9fe92ae MA |
4139 | |
4140 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
4141 | } |
4142 | ||
31383c68 DW |
4143 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
4144 | { | |
4145 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
4146 | return -EINVAL; | |
4147 | return 0; | |
4148 | } | |
4149 | ||
05ea8860 DW |
4150 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
4151 | { | |
aca78307 | 4152 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
4153 | } |
4154 | ||
1da177e4 LT |
4155 | /* |
4156 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
4157 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 4158 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
4159 | * this far. |
4160 | */ | |
b3ec9f33 | 4161 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
4162 | { |
4163 | BUG(); | |
d0217ac0 | 4164 | return 0; |
1da177e4 LT |
4165 | } |
4166 | ||
eec3636a JC |
4167 | /* |
4168 | * When a new function is introduced to vm_operations_struct and added | |
4169 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
4170 | * This is because under System V memory model, mappings created via | |
4171 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
4172 | * their original vm_ops are overwritten with shm_vm_ops. | |
4173 | */ | |
f0f37e2f | 4174 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 4175 | .fault = hugetlb_vm_op_fault, |
84afd99b | 4176 | .open = hugetlb_vm_op_open, |
a1e78772 | 4177 | .close = hugetlb_vm_op_close, |
dd3b614f | 4178 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 4179 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
4180 | }; |
4181 | ||
1e8f889b DG |
4182 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
4183 | int writable) | |
63551ae0 DG |
4184 | { |
4185 | pte_t entry; | |
79c1c594 | 4186 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
63551ae0 | 4187 | |
1e8f889b | 4188 | if (writable) { |
106c992a GS |
4189 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
4190 | vma->vm_page_prot))); | |
63551ae0 | 4191 | } else { |
106c992a GS |
4192 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
4193 | vma->vm_page_prot)); | |
63551ae0 DG |
4194 | } |
4195 | entry = pte_mkyoung(entry); | |
4196 | entry = pte_mkhuge(entry); | |
79c1c594 | 4197 | entry = arch_make_huge_pte(entry, shift, vma->vm_flags); |
63551ae0 DG |
4198 | |
4199 | return entry; | |
4200 | } | |
4201 | ||
1e8f889b DG |
4202 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
4203 | unsigned long address, pte_t *ptep) | |
4204 | { | |
4205 | pte_t entry; | |
4206 | ||
106c992a | 4207 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 4208 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 4209 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
4210 | } |
4211 | ||
d5ed7444 | 4212 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
4213 | { |
4214 | swp_entry_t swp; | |
4215 | ||
4216 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 4217 | return false; |
4a705fef | 4218 | swp = pte_to_swp_entry(pte); |
d79d176a | 4219 | if (is_migration_entry(swp)) |
d5ed7444 | 4220 | return true; |
4a705fef | 4221 | else |
d5ed7444 | 4222 | return false; |
4a705fef NH |
4223 | } |
4224 | ||
3e5c3600 | 4225 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
4226 | { |
4227 | swp_entry_t swp; | |
4228 | ||
4229 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 4230 | return false; |
4a705fef | 4231 | swp = pte_to_swp_entry(pte); |
d79d176a | 4232 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 4233 | return true; |
4a705fef | 4234 | else |
3e5c3600 | 4235 | return false; |
4a705fef | 4236 | } |
1e8f889b | 4237 | |
4eae4efa PX |
4238 | static void |
4239 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
4240 | struct page *new_page) | |
4241 | { | |
4242 | __SetPageUptodate(new_page); | |
4243 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); | |
4244 | hugepage_add_new_anon_rmap(new_page, vma, addr); | |
4245 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); | |
4246 | ClearHPageRestoreReserve(new_page); | |
4247 | SetHPageMigratable(new_page); | |
4248 | } | |
4249 | ||
63551ae0 DG |
4250 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
4251 | struct vm_area_struct *vma) | |
4252 | { | |
5e41540c | 4253 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 4254 | struct page *ptepage; |
1c59827d | 4255 | unsigned long addr; |
ca6eb14d | 4256 | bool cow = is_cow_mapping(vma->vm_flags); |
a5516438 AK |
4257 | struct hstate *h = hstate_vma(vma); |
4258 | unsigned long sz = huge_page_size(h); | |
4eae4efa | 4259 | unsigned long npages = pages_per_huge_page(h); |
c0d0381a | 4260 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 4261 | struct mmu_notifier_range range; |
e8569dd2 | 4262 | int ret = 0; |
1e8f889b | 4263 | |
ac46d4f3 | 4264 | if (cow) { |
7269f999 | 4265 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 4266 | vma->vm_start, |
ac46d4f3 JG |
4267 | vma->vm_end); |
4268 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
4269 | } else { |
4270 | /* | |
4271 | * For shared mappings i_mmap_rwsem must be held to call | |
4272 | * huge_pte_alloc, otherwise the returned ptep could go | |
4273 | * away if part of a shared pmd and another thread calls | |
4274 | * huge_pmd_unshare. | |
4275 | */ | |
4276 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 4277 | } |
e8569dd2 | 4278 | |
a5516438 | 4279 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 4280 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 4281 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
4282 | if (!src_pte) |
4283 | continue; | |
aec44e0f | 4284 | dst_pte = huge_pte_alloc(dst, vma, addr, sz); |
e8569dd2 AS |
4285 | if (!dst_pte) { |
4286 | ret = -ENOMEM; | |
4287 | break; | |
4288 | } | |
c5c99429 | 4289 | |
5e41540c MK |
4290 | /* |
4291 | * If the pagetables are shared don't copy or take references. | |
4292 | * dst_pte == src_pte is the common case of src/dest sharing. | |
4293 | * | |
4294 | * However, src could have 'unshared' and dst shares with | |
4295 | * another vma. If dst_pte !none, this implies sharing. | |
4296 | * Check here before taking page table lock, and once again | |
4297 | * after taking the lock below. | |
4298 | */ | |
4299 | dst_entry = huge_ptep_get(dst_pte); | |
4300 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
4301 | continue; |
4302 | ||
cb900f41 KS |
4303 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
4304 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4305 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 4306 | entry = huge_ptep_get(src_pte); |
5e41540c | 4307 | dst_entry = huge_ptep_get(dst_pte); |
4eae4efa | 4308 | again: |
5e41540c MK |
4309 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { |
4310 | /* | |
4311 | * Skip if src entry none. Also, skip in the | |
4312 | * unlikely case dst entry !none as this implies | |
4313 | * sharing with another vma. | |
4314 | */ | |
4a705fef NH |
4315 | ; |
4316 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
4317 | is_hugetlb_entry_hwpoisoned(entry))) { | |
4318 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
4319 | ||
4dd845b5 | 4320 | if (is_writable_migration_entry(swp_entry) && cow) { |
4a705fef NH |
4321 | /* |
4322 | * COW mappings require pages in both | |
4323 | * parent and child to be set to read. | |
4324 | */ | |
4dd845b5 AP |
4325 | swp_entry = make_readable_migration_entry( |
4326 | swp_offset(swp_entry)); | |
4a705fef | 4327 | entry = swp_entry_to_pte(swp_entry); |
e5251fd4 PA |
4328 | set_huge_swap_pte_at(src, addr, src_pte, |
4329 | entry, sz); | |
4a705fef | 4330 | } |
e5251fd4 | 4331 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 4332 | } else { |
4eae4efa PX |
4333 | entry = huge_ptep_get(src_pte); |
4334 | ptepage = pte_page(entry); | |
4335 | get_page(ptepage); | |
4336 | ||
4337 | /* | |
4338 | * This is a rare case where we see pinned hugetlb | |
4339 | * pages while they're prone to COW. We need to do the | |
4340 | * COW earlier during fork. | |
4341 | * | |
4342 | * When pre-allocating the page or copying data, we | |
4343 | * need to be without the pgtable locks since we could | |
4344 | * sleep during the process. | |
4345 | */ | |
4346 | if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { | |
4347 | pte_t src_pte_old = entry; | |
4348 | struct page *new; | |
4349 | ||
4350 | spin_unlock(src_ptl); | |
4351 | spin_unlock(dst_ptl); | |
4352 | /* Do not use reserve as it's private owned */ | |
4353 | new = alloc_huge_page(vma, addr, 1); | |
4354 | if (IS_ERR(new)) { | |
4355 | put_page(ptepage); | |
4356 | ret = PTR_ERR(new); | |
4357 | break; | |
4358 | } | |
4359 | copy_user_huge_page(new, ptepage, addr, vma, | |
4360 | npages); | |
4361 | put_page(ptepage); | |
4362 | ||
4363 | /* Install the new huge page if src pte stable */ | |
4364 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
4365 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
4366 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4367 | entry = huge_ptep_get(src_pte); | |
4368 | if (!pte_same(src_pte_old, entry)) { | |
846be085 MK |
4369 | restore_reserve_on_error(h, vma, addr, |
4370 | new); | |
4eae4efa PX |
4371 | put_page(new); |
4372 | /* dst_entry won't change as in child */ | |
4373 | goto again; | |
4374 | } | |
4375 | hugetlb_install_page(vma, dst_pte, addr, new); | |
4376 | spin_unlock(src_ptl); | |
4377 | spin_unlock(dst_ptl); | |
4378 | continue; | |
4379 | } | |
4380 | ||
34ee645e | 4381 | if (cow) { |
0f10851e JG |
4382 | /* |
4383 | * No need to notify as we are downgrading page | |
4384 | * table protection not changing it to point | |
4385 | * to a new page. | |
4386 | * | |
ad56b738 | 4387 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4388 | */ |
7f2e9525 | 4389 | huge_ptep_set_wrprotect(src, addr, src_pte); |
84894e1c | 4390 | entry = huge_pte_wrprotect(entry); |
34ee645e | 4391 | } |
4eae4efa | 4392 | |
53f9263b | 4393 | page_dup_rmap(ptepage, true); |
1c59827d | 4394 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 4395 | hugetlb_count_add(npages, dst); |
1c59827d | 4396 | } |
cb900f41 KS |
4397 | spin_unlock(src_ptl); |
4398 | spin_unlock(dst_ptl); | |
63551ae0 | 4399 | } |
63551ae0 | 4400 | |
e8569dd2 | 4401 | if (cow) |
ac46d4f3 | 4402 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
4403 | else |
4404 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
4405 | |
4406 | return ret; | |
63551ae0 DG |
4407 | } |
4408 | ||
24669e58 AK |
4409 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
4410 | unsigned long start, unsigned long end, | |
4411 | struct page *ref_page) | |
63551ae0 DG |
4412 | { |
4413 | struct mm_struct *mm = vma->vm_mm; | |
4414 | unsigned long address; | |
c7546f8f | 4415 | pte_t *ptep; |
63551ae0 | 4416 | pte_t pte; |
cb900f41 | 4417 | spinlock_t *ptl; |
63551ae0 | 4418 | struct page *page; |
a5516438 AK |
4419 | struct hstate *h = hstate_vma(vma); |
4420 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 4421 | struct mmu_notifier_range range; |
a5516438 | 4422 | |
63551ae0 | 4423 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
4424 | BUG_ON(start & ~huge_page_mask(h)); |
4425 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 4426 | |
07e32661 AK |
4427 | /* |
4428 | * This is a hugetlb vma, all the pte entries should point | |
4429 | * to huge page. | |
4430 | */ | |
ed6a7935 | 4431 | tlb_change_page_size(tlb, sz); |
24669e58 | 4432 | tlb_start_vma(tlb, vma); |
dff11abe MK |
4433 | |
4434 | /* | |
4435 | * If sharing possible, alert mmu notifiers of worst case. | |
4436 | */ | |
6f4f13e8 JG |
4437 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
4438 | end); | |
ac46d4f3 JG |
4439 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
4440 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 4441 | address = start; |
569f48b8 | 4442 | for (; address < end; address += sz) { |
7868a208 | 4443 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 4444 | if (!ptep) |
c7546f8f DG |
4445 | continue; |
4446 | ||
cb900f41 | 4447 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 4448 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
31d49da5 | 4449 | spin_unlock(ptl); |
dff11abe MK |
4450 | /* |
4451 | * We just unmapped a page of PMDs by clearing a PUD. | |
4452 | * The caller's TLB flush range should cover this area. | |
4453 | */ | |
31d49da5 AK |
4454 | continue; |
4455 | } | |
39dde65c | 4456 | |
6629326b | 4457 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
4458 | if (huge_pte_none(pte)) { |
4459 | spin_unlock(ptl); | |
4460 | continue; | |
4461 | } | |
6629326b HD |
4462 | |
4463 | /* | |
9fbc1f63 NH |
4464 | * Migrating hugepage or HWPoisoned hugepage is already |
4465 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 4466 | */ |
9fbc1f63 | 4467 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 4468 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
4469 | spin_unlock(ptl); |
4470 | continue; | |
8c4894c6 | 4471 | } |
6629326b HD |
4472 | |
4473 | page = pte_page(pte); | |
04f2cbe3 MG |
4474 | /* |
4475 | * If a reference page is supplied, it is because a specific | |
4476 | * page is being unmapped, not a range. Ensure the page we | |
4477 | * are about to unmap is the actual page of interest. | |
4478 | */ | |
4479 | if (ref_page) { | |
31d49da5 AK |
4480 | if (page != ref_page) { |
4481 | spin_unlock(ptl); | |
4482 | continue; | |
4483 | } | |
04f2cbe3 MG |
4484 | /* |
4485 | * Mark the VMA as having unmapped its page so that | |
4486 | * future faults in this VMA will fail rather than | |
4487 | * looking like data was lost | |
4488 | */ | |
4489 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
4490 | } | |
4491 | ||
c7546f8f | 4492 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 4493 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 4494 | if (huge_pte_dirty(pte)) |
6649a386 | 4495 | set_page_dirty(page); |
9e81130b | 4496 | |
5d317b2b | 4497 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 4498 | page_remove_rmap(page, true); |
31d49da5 | 4499 | |
cb900f41 | 4500 | spin_unlock(ptl); |
e77b0852 | 4501 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
4502 | /* |
4503 | * Bail out after unmapping reference page if supplied | |
4504 | */ | |
4505 | if (ref_page) | |
4506 | break; | |
fe1668ae | 4507 | } |
ac46d4f3 | 4508 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 4509 | tlb_end_vma(tlb, vma); |
1da177e4 | 4510 | } |
63551ae0 | 4511 | |
d833352a MG |
4512 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
4513 | struct vm_area_struct *vma, unsigned long start, | |
4514 | unsigned long end, struct page *ref_page) | |
4515 | { | |
4516 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
4517 | ||
4518 | /* | |
4519 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
4520 | * test will fail on a vma being torn down, and not grab a page table | |
4521 | * on its way out. We're lucky that the flag has such an appropriate | |
4522 | * name, and can in fact be safely cleared here. We could clear it | |
4523 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 4524 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 4525 | * because in the context this is called, the VMA is about to be |
c8c06efa | 4526 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
4527 | */ |
4528 | vma->vm_flags &= ~VM_MAYSHARE; | |
4529 | } | |
4530 | ||
502717f4 | 4531 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 4532 | unsigned long end, struct page *ref_page) |
502717f4 | 4533 | { |
24669e58 | 4534 | struct mmu_gather tlb; |
dff11abe | 4535 | |
a72afd87 | 4536 | tlb_gather_mmu(&tlb, vma->vm_mm); |
24669e58 | 4537 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
ae8eba8b | 4538 | tlb_finish_mmu(&tlb); |
502717f4 KC |
4539 | } |
4540 | ||
04f2cbe3 MG |
4541 | /* |
4542 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 4543 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
4544 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
4545 | * same region. | |
4546 | */ | |
2f4612af DB |
4547 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
4548 | struct page *page, unsigned long address) | |
04f2cbe3 | 4549 | { |
7526674d | 4550 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
4551 | struct vm_area_struct *iter_vma; |
4552 | struct address_space *mapping; | |
04f2cbe3 MG |
4553 | pgoff_t pgoff; |
4554 | ||
4555 | /* | |
4556 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
4557 | * from page cache lookup which is in HPAGE_SIZE units. | |
4558 | */ | |
7526674d | 4559 | address = address & huge_page_mask(h); |
36e4f20a MH |
4560 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
4561 | vma->vm_pgoff; | |
93c76a3d | 4562 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 4563 | |
4eb2b1dc MG |
4564 | /* |
4565 | * Take the mapping lock for the duration of the table walk. As | |
4566 | * this mapping should be shared between all the VMAs, | |
4567 | * __unmap_hugepage_range() is called as the lock is already held | |
4568 | */ | |
83cde9e8 | 4569 | i_mmap_lock_write(mapping); |
6b2dbba8 | 4570 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
4571 | /* Do not unmap the current VMA */ |
4572 | if (iter_vma == vma) | |
4573 | continue; | |
4574 | ||
2f84a899 MG |
4575 | /* |
4576 | * Shared VMAs have their own reserves and do not affect | |
4577 | * MAP_PRIVATE accounting but it is possible that a shared | |
4578 | * VMA is using the same page so check and skip such VMAs. | |
4579 | */ | |
4580 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
4581 | continue; | |
4582 | ||
04f2cbe3 MG |
4583 | /* |
4584 | * Unmap the page from other VMAs without their own reserves. | |
4585 | * They get marked to be SIGKILLed if they fault in these | |
4586 | * areas. This is because a future no-page fault on this VMA | |
4587 | * could insert a zeroed page instead of the data existing | |
4588 | * from the time of fork. This would look like data corruption | |
4589 | */ | |
4590 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
4591 | unmap_hugepage_range(iter_vma, address, |
4592 | address + huge_page_size(h), page); | |
04f2cbe3 | 4593 | } |
83cde9e8 | 4594 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
4595 | } |
4596 | ||
0fe6e20b NH |
4597 | /* |
4598 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
4599 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
4600 | * cannot race with other handlers or page migration. | |
4601 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 4602 | */ |
2b740303 | 4603 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 4604 | unsigned long address, pte_t *ptep, |
3999f52e | 4605 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 4606 | { |
3999f52e | 4607 | pte_t pte; |
a5516438 | 4608 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 4609 | struct page *old_page, *new_page; |
2b740303 SJ |
4610 | int outside_reserve = 0; |
4611 | vm_fault_t ret = 0; | |
974e6d66 | 4612 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 4613 | struct mmu_notifier_range range; |
1e8f889b | 4614 | |
3999f52e | 4615 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
4616 | old_page = pte_page(pte); |
4617 | ||
04f2cbe3 | 4618 | retry_avoidcopy: |
1e8f889b DG |
4619 | /* If no-one else is actually using this page, avoid the copy |
4620 | * and just make the page writable */ | |
37a2140d | 4621 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 4622 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 4623 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 4624 | return 0; |
1e8f889b DG |
4625 | } |
4626 | ||
04f2cbe3 MG |
4627 | /* |
4628 | * If the process that created a MAP_PRIVATE mapping is about to | |
4629 | * perform a COW due to a shared page count, attempt to satisfy | |
4630 | * the allocation without using the existing reserves. The pagecache | |
4631 | * page is used to determine if the reserve at this address was | |
4632 | * consumed or not. If reserves were used, a partial faulted mapping | |
4633 | * at the time of fork() could consume its reserves on COW instead | |
4634 | * of the full address range. | |
4635 | */ | |
5944d011 | 4636 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
4637 | old_page != pagecache_page) |
4638 | outside_reserve = 1; | |
4639 | ||
09cbfeaf | 4640 | get_page(old_page); |
b76c8cfb | 4641 | |
ad4404a2 DB |
4642 | /* |
4643 | * Drop page table lock as buddy allocator may be called. It will | |
4644 | * be acquired again before returning to the caller, as expected. | |
4645 | */ | |
cb900f41 | 4646 | spin_unlock(ptl); |
5b7a1d40 | 4647 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 4648 | |
2fc39cec | 4649 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
4650 | /* |
4651 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4652 | * it is due to references held by a child and an insufficient | |
4653 | * huge page pool. To guarantee the original mappers | |
4654 | * reliability, unmap the page from child processes. The child | |
4655 | * may get SIGKILLed if it later faults. | |
4656 | */ | |
4657 | if (outside_reserve) { | |
e7dd91c4 MK |
4658 | struct address_space *mapping = vma->vm_file->f_mapping; |
4659 | pgoff_t idx; | |
4660 | u32 hash; | |
4661 | ||
09cbfeaf | 4662 | put_page(old_page); |
04f2cbe3 | 4663 | BUG_ON(huge_pte_none(pte)); |
e7dd91c4 MK |
4664 | /* |
4665 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
4666 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
4667 | * in write mode. Dropping i_mmap_rwsem in read mode | |
4668 | * here is OK as COW mappings do not interact with | |
4669 | * PMD sharing. | |
4670 | * | |
4671 | * Reacquire both after unmap operation. | |
4672 | */ | |
4673 | idx = vma_hugecache_offset(h, vma, haddr); | |
4674 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4675 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4676 | i_mmap_unlock_read(mapping); | |
4677 | ||
5b7a1d40 | 4678 | unmap_ref_private(mm, vma, old_page, haddr); |
e7dd91c4 MK |
4679 | |
4680 | i_mmap_lock_read(mapping); | |
4681 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
2f4612af | 4682 | spin_lock(ptl); |
5b7a1d40 | 4683 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
4684 | if (likely(ptep && |
4685 | pte_same(huge_ptep_get(ptep), pte))) | |
4686 | goto retry_avoidcopy; | |
4687 | /* | |
4688 | * race occurs while re-acquiring page table | |
4689 | * lock, and our job is done. | |
4690 | */ | |
4691 | return 0; | |
04f2cbe3 MG |
4692 | } |
4693 | ||
2b740303 | 4694 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 4695 | goto out_release_old; |
1e8f889b DG |
4696 | } |
4697 | ||
0fe6e20b NH |
4698 | /* |
4699 | * When the original hugepage is shared one, it does not have | |
4700 | * anon_vma prepared. | |
4701 | */ | |
44e2aa93 | 4702 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
4703 | ret = VM_FAULT_OOM; |
4704 | goto out_release_all; | |
44e2aa93 | 4705 | } |
0fe6e20b | 4706 | |
974e6d66 | 4707 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 4708 | pages_per_huge_page(h)); |
0ed361de | 4709 | __SetPageUptodate(new_page); |
1e8f889b | 4710 | |
7269f999 | 4711 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 4712 | haddr + huge_page_size(h)); |
ac46d4f3 | 4713 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 4714 | |
b76c8cfb | 4715 | /* |
cb900f41 | 4716 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
4717 | * before the page tables are altered |
4718 | */ | |
cb900f41 | 4719 | spin_lock(ptl); |
5b7a1d40 | 4720 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 4721 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 4722 | ClearHPageRestoreReserve(new_page); |
07443a85 | 4723 | |
1e8f889b | 4724 | /* Break COW */ |
5b7a1d40 | 4725 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 4726 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 4727 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 4728 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 4729 | page_remove_rmap(old_page, true); |
5b7a1d40 | 4730 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
8f251a3d | 4731 | SetHPageMigratable(new_page); |
1e8f889b DG |
4732 | /* Make the old page be freed below */ |
4733 | new_page = old_page; | |
4734 | } | |
cb900f41 | 4735 | spin_unlock(ptl); |
ac46d4f3 | 4736 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 4737 | out_release_all: |
c7b1850d MK |
4738 | /* No restore in case of successful pagetable update (Break COW) */ |
4739 | if (new_page != old_page) | |
4740 | restore_reserve_on_error(h, vma, haddr, new_page); | |
09cbfeaf | 4741 | put_page(new_page); |
ad4404a2 | 4742 | out_release_old: |
09cbfeaf | 4743 | put_page(old_page); |
8312034f | 4744 | |
ad4404a2 DB |
4745 | spin_lock(ptl); /* Caller expects lock to be held */ |
4746 | return ret; | |
1e8f889b DG |
4747 | } |
4748 | ||
04f2cbe3 | 4749 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
4750 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
4751 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
4752 | { |
4753 | struct address_space *mapping; | |
e7c4b0bf | 4754 | pgoff_t idx; |
04f2cbe3 MG |
4755 | |
4756 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 4757 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
4758 | |
4759 | return find_lock_page(mapping, idx); | |
4760 | } | |
4761 | ||
3ae77f43 HD |
4762 | /* |
4763 | * Return whether there is a pagecache page to back given address within VMA. | |
4764 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4765 | */ | |
4766 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
4767 | struct vm_area_struct *vma, unsigned long address) |
4768 | { | |
4769 | struct address_space *mapping; | |
4770 | pgoff_t idx; | |
4771 | struct page *page; | |
4772 | ||
4773 | mapping = vma->vm_file->f_mapping; | |
4774 | idx = vma_hugecache_offset(h, vma, address); | |
4775 | ||
4776 | page = find_get_page(mapping, idx); | |
4777 | if (page) | |
4778 | put_page(page); | |
4779 | return page != NULL; | |
4780 | } | |
4781 | ||
ab76ad54 MK |
4782 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
4783 | pgoff_t idx) | |
4784 | { | |
4785 | struct inode *inode = mapping->host; | |
4786 | struct hstate *h = hstate_inode(inode); | |
4787 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4788 | ||
4789 | if (err) | |
4790 | return err; | |
d6995da3 | 4791 | ClearHPageRestoreReserve(page); |
ab76ad54 | 4792 | |
22146c3c MK |
4793 | /* |
4794 | * set page dirty so that it will not be removed from cache/file | |
4795 | * by non-hugetlbfs specific code paths. | |
4796 | */ | |
4797 | set_page_dirty(page); | |
4798 | ||
ab76ad54 MK |
4799 | spin_lock(&inode->i_lock); |
4800 | inode->i_blocks += blocks_per_huge_page(h); | |
4801 | spin_unlock(&inode->i_lock); | |
4802 | return 0; | |
4803 | } | |
4804 | ||
7677f7fd AR |
4805 | static inline vm_fault_t hugetlb_handle_userfault(struct vm_area_struct *vma, |
4806 | struct address_space *mapping, | |
4807 | pgoff_t idx, | |
4808 | unsigned int flags, | |
4809 | unsigned long haddr, | |
4810 | unsigned long reason) | |
4811 | { | |
4812 | vm_fault_t ret; | |
4813 | u32 hash; | |
4814 | struct vm_fault vmf = { | |
4815 | .vma = vma, | |
4816 | .address = haddr, | |
4817 | .flags = flags, | |
4818 | ||
4819 | /* | |
4820 | * Hard to debug if it ends up being | |
4821 | * used by a callee that assumes | |
4822 | * something about the other | |
4823 | * uninitialized fields... same as in | |
4824 | * memory.c | |
4825 | */ | |
4826 | }; | |
4827 | ||
4828 | /* | |
4829 | * hugetlb_fault_mutex and i_mmap_rwsem must be | |
4830 | * dropped before handling userfault. Reacquire | |
4831 | * after handling fault to make calling code simpler. | |
4832 | */ | |
4833 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4834 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4835 | i_mmap_unlock_read(mapping); | |
4836 | ret = handle_userfault(&vmf, reason); | |
4837 | i_mmap_lock_read(mapping); | |
4838 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
4839 | ||
4840 | return ret; | |
4841 | } | |
4842 | ||
2b740303 SJ |
4843 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
4844 | struct vm_area_struct *vma, | |
4845 | struct address_space *mapping, pgoff_t idx, | |
4846 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 4847 | { |
a5516438 | 4848 | struct hstate *h = hstate_vma(vma); |
2b740303 | 4849 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 4850 | int anon_rmap = 0; |
4c887265 | 4851 | unsigned long size; |
4c887265 | 4852 | struct page *page; |
1e8f889b | 4853 | pte_t new_pte; |
cb900f41 | 4854 | spinlock_t *ptl; |
285b8dca | 4855 | unsigned long haddr = address & huge_page_mask(h); |
c7b1850d | 4856 | bool new_page, new_pagecache_page = false; |
4c887265 | 4857 | |
04f2cbe3 MG |
4858 | /* |
4859 | * Currently, we are forced to kill the process in the event the | |
4860 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4861 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4862 | */ |
4863 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4864 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4865 | current->pid); |
04f2cbe3 MG |
4866 | return ret; |
4867 | } | |
4868 | ||
4c887265 | 4869 | /* |
87bf91d3 MK |
4870 | * We can not race with truncation due to holding i_mmap_rwsem. |
4871 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4872 | * once for faults beyond end of file. | |
4c887265 | 4873 | */ |
87bf91d3 MK |
4874 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4875 | if (idx >= size) | |
4876 | goto out; | |
4877 | ||
6bda666a | 4878 | retry: |
c7b1850d | 4879 | new_page = false; |
6bda666a CL |
4880 | page = find_lock_page(mapping, idx); |
4881 | if (!page) { | |
7677f7fd | 4882 | /* Check for page in userfault range */ |
1a1aad8a | 4883 | if (userfaultfd_missing(vma)) { |
7677f7fd AR |
4884 | ret = hugetlb_handle_userfault(vma, mapping, idx, |
4885 | flags, haddr, | |
4886 | VM_UFFD_MISSING); | |
1a1aad8a MK |
4887 | goto out; |
4888 | } | |
4889 | ||
285b8dca | 4890 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4891 | if (IS_ERR(page)) { |
4643d67e MK |
4892 | /* |
4893 | * Returning error will result in faulting task being | |
4894 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4895 | * tasks from racing to fault in the same page which | |
4896 | * could result in false unable to allocate errors. | |
4897 | * Page migration does not take the fault mutex, but | |
4898 | * does a clear then write of pte's under page table | |
4899 | * lock. Page fault code could race with migration, | |
4900 | * notice the clear pte and try to allocate a page | |
4901 | * here. Before returning error, get ptl and make | |
4902 | * sure there really is no pte entry. | |
4903 | */ | |
4904 | ptl = huge_pte_lock(h, mm, ptep); | |
d83e6c8a ML |
4905 | ret = 0; |
4906 | if (huge_pte_none(huge_ptep_get(ptep))) | |
4907 | ret = vmf_error(PTR_ERR(page)); | |
4643d67e | 4908 | spin_unlock(ptl); |
6bda666a CL |
4909 | goto out; |
4910 | } | |
47ad8475 | 4911 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4912 | __SetPageUptodate(page); |
cb6acd01 | 4913 | new_page = true; |
ac9b9c66 | 4914 | |
f83a275d | 4915 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4916 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4917 | if (err) { |
4918 | put_page(page); | |
6bda666a CL |
4919 | if (err == -EEXIST) |
4920 | goto retry; | |
4921 | goto out; | |
4922 | } | |
c7b1850d | 4923 | new_pagecache_page = true; |
23be7468 | 4924 | } else { |
6bda666a | 4925 | lock_page(page); |
0fe6e20b NH |
4926 | if (unlikely(anon_vma_prepare(vma))) { |
4927 | ret = VM_FAULT_OOM; | |
4928 | goto backout_unlocked; | |
4929 | } | |
409eb8c2 | 4930 | anon_rmap = 1; |
23be7468 | 4931 | } |
0fe6e20b | 4932 | } else { |
998b4382 NH |
4933 | /* |
4934 | * If memory error occurs between mmap() and fault, some process | |
4935 | * don't have hwpoisoned swap entry for errored virtual address. | |
4936 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4937 | */ | |
4938 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 4939 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4940 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4941 | goto backout_unlocked; |
4942 | } | |
7677f7fd AR |
4943 | |
4944 | /* Check for page in userfault range. */ | |
4945 | if (userfaultfd_minor(vma)) { | |
4946 | unlock_page(page); | |
4947 | put_page(page); | |
4948 | ret = hugetlb_handle_userfault(vma, mapping, idx, | |
4949 | flags, haddr, | |
4950 | VM_UFFD_MINOR); | |
4951 | goto out; | |
4952 | } | |
6bda666a | 4953 | } |
1e8f889b | 4954 | |
57303d80 AW |
4955 | /* |
4956 | * If we are going to COW a private mapping later, we examine the | |
4957 | * pending reservations for this page now. This will ensure that | |
4958 | * any allocations necessary to record that reservation occur outside | |
4959 | * the spinlock. | |
4960 | */ | |
5e911373 | 4961 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4962 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4963 | ret = VM_FAULT_OOM; |
4964 | goto backout_unlocked; | |
4965 | } | |
5e911373 | 4966 | /* Just decrements count, does not deallocate */ |
285b8dca | 4967 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4968 | } |
57303d80 | 4969 | |
8bea8052 | 4970 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 4971 | ret = 0; |
7f2e9525 | 4972 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4973 | goto backout; |
4974 | ||
07443a85 | 4975 | if (anon_rmap) { |
d6995da3 | 4976 | ClearHPageRestoreReserve(page); |
285b8dca | 4977 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4978 | } else |
53f9263b | 4979 | page_dup_rmap(page, true); |
1e8f889b DG |
4980 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4981 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4982 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4983 | |
5d317b2b | 4984 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4985 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4986 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4987 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4988 | } |
4989 | ||
cb900f41 | 4990 | spin_unlock(ptl); |
cb6acd01 MK |
4991 | |
4992 | /* | |
8f251a3d MK |
4993 | * Only set HPageMigratable in newly allocated pages. Existing pages |
4994 | * found in the pagecache may not have HPageMigratableset if they have | |
4995 | * been isolated for migration. | |
cb6acd01 MK |
4996 | */ |
4997 | if (new_page) | |
8f251a3d | 4998 | SetHPageMigratable(page); |
cb6acd01 | 4999 | |
4c887265 AL |
5000 | unlock_page(page); |
5001 | out: | |
ac9b9c66 | 5002 | return ret; |
4c887265 AL |
5003 | |
5004 | backout: | |
cb900f41 | 5005 | spin_unlock(ptl); |
2b26736c | 5006 | backout_unlocked: |
4c887265 | 5007 | unlock_page(page); |
c7b1850d MK |
5008 | /* restore reserve for newly allocated pages not in page cache */ |
5009 | if (new_page && !new_pagecache_page) | |
5010 | restore_reserve_on_error(h, vma, haddr, page); | |
4c887265 AL |
5011 | put_page(page); |
5012 | goto out; | |
ac9b9c66 HD |
5013 | } |
5014 | ||
8382d914 | 5015 | #ifdef CONFIG_SMP |
188b04a7 | 5016 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5017 | { |
5018 | unsigned long key[2]; | |
5019 | u32 hash; | |
5020 | ||
1b426bac MK |
5021 | key[0] = (unsigned long) mapping; |
5022 | key[1] = idx; | |
8382d914 | 5023 | |
55254636 | 5024 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
5025 | |
5026 | return hash & (num_fault_mutexes - 1); | |
5027 | } | |
5028 | #else | |
5029 | /* | |
6c26d310 | 5030 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
5031 | * return 0 and avoid the hashing overhead. |
5032 | */ | |
188b04a7 | 5033 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
5034 | { |
5035 | return 0; | |
5036 | } | |
5037 | #endif | |
5038 | ||
2b740303 | 5039 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 5040 | unsigned long address, unsigned int flags) |
86e5216f | 5041 | { |
8382d914 | 5042 | pte_t *ptep, entry; |
cb900f41 | 5043 | spinlock_t *ptl; |
2b740303 | 5044 | vm_fault_t ret; |
8382d914 DB |
5045 | u32 hash; |
5046 | pgoff_t idx; | |
0fe6e20b | 5047 | struct page *page = NULL; |
57303d80 | 5048 | struct page *pagecache_page = NULL; |
a5516438 | 5049 | struct hstate *h = hstate_vma(vma); |
8382d914 | 5050 | struct address_space *mapping; |
0f792cf9 | 5051 | int need_wait_lock = 0; |
285b8dca | 5052 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 5053 | |
285b8dca | 5054 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 5055 | if (ptep) { |
c0d0381a MK |
5056 | /* |
5057 | * Since we hold no locks, ptep could be stale. That is | |
5058 | * OK as we are only making decisions based on content and | |
5059 | * not actually modifying content here. | |
5060 | */ | |
fd6a03ed | 5061 | entry = huge_ptep_get(ptep); |
290408d4 | 5062 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 5063 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
5064 | return 0; |
5065 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 5066 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 5067 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
5068 | } |
5069 | ||
c0d0381a MK |
5070 | /* |
5071 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
5072 | * until finished with ptep. This serves two purposes: |
5073 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
5074 | * and making the ptep no longer valid. | |
5075 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
5076 | * |
5077 | * ptep could have already be assigned via huge_pte_offset. That | |
5078 | * is OK, as huge_pte_alloc will return the same value unless | |
5079 | * something has changed. | |
5080 | */ | |
8382d914 | 5081 | mapping = vma->vm_file->f_mapping; |
c0d0381a | 5082 | i_mmap_lock_read(mapping); |
aec44e0f | 5083 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); |
c0d0381a MK |
5084 | if (!ptep) { |
5085 | i_mmap_unlock_read(mapping); | |
5086 | return VM_FAULT_OOM; | |
5087 | } | |
8382d914 | 5088 | |
3935baa9 DG |
5089 | /* |
5090 | * Serialize hugepage allocation and instantiation, so that we don't | |
5091 | * get spurious allocation failures if two CPUs race to instantiate | |
5092 | * the same page in the page cache. | |
5093 | */ | |
c0d0381a | 5094 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 5095 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 5096 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 5097 | |
7f2e9525 GS |
5098 | entry = huge_ptep_get(ptep); |
5099 | if (huge_pte_none(entry)) { | |
8382d914 | 5100 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 5101 | goto out_mutex; |
3935baa9 | 5102 | } |
86e5216f | 5103 | |
83c54070 | 5104 | ret = 0; |
1e8f889b | 5105 | |
0f792cf9 NH |
5106 | /* |
5107 | * entry could be a migration/hwpoison entry at this point, so this | |
5108 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
5109 | * an active hugepage in pagecache. This goto expects the 2nd page |
5110 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
5111 | * properly handle it. | |
0f792cf9 NH |
5112 | */ |
5113 | if (!pte_present(entry)) | |
5114 | goto out_mutex; | |
5115 | ||
57303d80 AW |
5116 | /* |
5117 | * If we are going to COW the mapping later, we examine the pending | |
5118 | * reservations for this page now. This will ensure that any | |
5119 | * allocations necessary to record that reservation occur outside the | |
5120 | * spinlock. For private mappings, we also lookup the pagecache | |
5121 | * page now as it is used to determine if a reservation has been | |
5122 | * consumed. | |
5123 | */ | |
106c992a | 5124 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 5125 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 5126 | ret = VM_FAULT_OOM; |
b4d1d99f | 5127 | goto out_mutex; |
2b26736c | 5128 | } |
5e911373 | 5129 | /* Just decrements count, does not deallocate */ |
285b8dca | 5130 | vma_end_reservation(h, vma, haddr); |
57303d80 | 5131 | |
f83a275d | 5132 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 5133 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 5134 | vma, haddr); |
57303d80 AW |
5135 | } |
5136 | ||
0f792cf9 NH |
5137 | ptl = huge_pte_lock(h, mm, ptep); |
5138 | ||
5139 | /* Check for a racing update before calling hugetlb_cow */ | |
5140 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
5141 | goto out_ptl; | |
5142 | ||
56c9cfb1 NH |
5143 | /* |
5144 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
5145 | * pagecache_page, so here we need take the former one | |
5146 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
5147 | */ |
5148 | page = pte_page(entry); | |
5149 | if (page != pagecache_page) | |
0f792cf9 NH |
5150 | if (!trylock_page(page)) { |
5151 | need_wait_lock = 1; | |
5152 | goto out_ptl; | |
5153 | } | |
b4d1d99f | 5154 | |
0f792cf9 | 5155 | get_page(page); |
b4d1d99f | 5156 | |
788c7df4 | 5157 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 5158 | if (!huge_pte_write(entry)) { |
974e6d66 | 5159 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 5160 | pagecache_page, ptl); |
0f792cf9 | 5161 | goto out_put_page; |
b4d1d99f | 5162 | } |
106c992a | 5163 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
5164 | } |
5165 | entry = pte_mkyoung(entry); | |
285b8dca | 5166 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 5167 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 5168 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
5169 | out_put_page: |
5170 | if (page != pagecache_page) | |
5171 | unlock_page(page); | |
5172 | put_page(page); | |
cb900f41 KS |
5173 | out_ptl: |
5174 | spin_unlock(ptl); | |
57303d80 AW |
5175 | |
5176 | if (pagecache_page) { | |
5177 | unlock_page(pagecache_page); | |
5178 | put_page(pagecache_page); | |
5179 | } | |
b4d1d99f | 5180 | out_mutex: |
c672c7f2 | 5181 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 5182 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
5183 | /* |
5184 | * Generally it's safe to hold refcount during waiting page lock. But | |
5185 | * here we just wait to defer the next page fault to avoid busy loop and | |
5186 | * the page is not used after unlocked before returning from the current | |
5187 | * page fault. So we are safe from accessing freed page, even if we wait | |
5188 | * here without taking refcount. | |
5189 | */ | |
5190 | if (need_wait_lock) | |
5191 | wait_on_page_locked(page); | |
1e8f889b | 5192 | return ret; |
86e5216f AL |
5193 | } |
5194 | ||
714c1891 | 5195 | #ifdef CONFIG_USERFAULTFD |
8fb5debc MK |
5196 | /* |
5197 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
5198 | * modifications for huge pages. | |
5199 | */ | |
5200 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
5201 | pte_t *dst_pte, | |
5202 | struct vm_area_struct *dst_vma, | |
5203 | unsigned long dst_addr, | |
5204 | unsigned long src_addr, | |
f6191471 | 5205 | enum mcopy_atomic_mode mode, |
8fb5debc MK |
5206 | struct page **pagep) |
5207 | { | |
f6191471 | 5208 | bool is_continue = (mode == MCOPY_ATOMIC_CONTINUE); |
8cc5fcbb MA |
5209 | struct hstate *h = hstate_vma(dst_vma); |
5210 | struct address_space *mapping = dst_vma->vm_file->f_mapping; | |
5211 | pgoff_t idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
1e392147 | 5212 | unsigned long size; |
1c9e8def | 5213 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
5214 | pte_t _dst_pte; |
5215 | spinlock_t *ptl; | |
8cc5fcbb | 5216 | int ret = -ENOMEM; |
8fb5debc | 5217 | struct page *page; |
f6191471 | 5218 | int writable; |
c7b1850d | 5219 | bool new_pagecache_page = false; |
8fb5debc | 5220 | |
f6191471 AR |
5221 | if (is_continue) { |
5222 | ret = -EFAULT; | |
5223 | page = find_lock_page(mapping, idx); | |
5224 | if (!page) | |
5225 | goto out; | |
5226 | } else if (!*pagep) { | |
d84cf06e MA |
5227 | /* If a page already exists, then it's UFFDIO_COPY for |
5228 | * a non-missing case. Return -EEXIST. | |
5229 | */ | |
5230 | if (vm_shared && | |
5231 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5232 | ret = -EEXIST; | |
5233 | goto out; | |
5234 | } | |
5235 | ||
8fb5debc | 5236 | page = alloc_huge_page(dst_vma, dst_addr, 0); |
d84cf06e MA |
5237 | if (IS_ERR(page)) { |
5238 | ret = -ENOMEM; | |
8fb5debc | 5239 | goto out; |
d84cf06e | 5240 | } |
8fb5debc MK |
5241 | |
5242 | ret = copy_huge_page_from_user(page, | |
5243 | (const void __user *) src_addr, | |
810a56b9 | 5244 | pages_per_huge_page(h), false); |
8fb5debc | 5245 | |
c1e8d7c6 | 5246 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 5247 | if (unlikely(ret)) { |
9e368259 | 5248 | ret = -ENOENT; |
8cc5fcbb MA |
5249 | /* Free the allocated page which may have |
5250 | * consumed a reservation. | |
5251 | */ | |
5252 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
5253 | put_page(page); | |
5254 | ||
5255 | /* Allocate a temporary page to hold the copied | |
5256 | * contents. | |
5257 | */ | |
5258 | page = alloc_huge_page_vma(h, dst_vma, dst_addr); | |
5259 | if (!page) { | |
5260 | ret = -ENOMEM; | |
5261 | goto out; | |
5262 | } | |
8fb5debc | 5263 | *pagep = page; |
8cc5fcbb MA |
5264 | /* Set the outparam pagep and return to the caller to |
5265 | * copy the contents outside the lock. Don't free the | |
5266 | * page. | |
5267 | */ | |
8fb5debc MK |
5268 | goto out; |
5269 | } | |
5270 | } else { | |
8cc5fcbb MA |
5271 | if (vm_shared && |
5272 | hugetlbfs_pagecache_present(h, dst_vma, dst_addr)) { | |
5273 | put_page(*pagep); | |
5274 | ret = -EEXIST; | |
5275 | *pagep = NULL; | |
5276 | goto out; | |
5277 | } | |
5278 | ||
5279 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
5280 | if (IS_ERR(page)) { | |
5281 | ret = -ENOMEM; | |
5282 | *pagep = NULL; | |
5283 | goto out; | |
5284 | } | |
5285 | copy_huge_page(page, *pagep); | |
5286 | put_page(*pagep); | |
8fb5debc MK |
5287 | *pagep = NULL; |
5288 | } | |
5289 | ||
5290 | /* | |
5291 | * The memory barrier inside __SetPageUptodate makes sure that | |
5292 | * preceding stores to the page contents become visible before | |
5293 | * the set_pte_at() write. | |
5294 | */ | |
5295 | __SetPageUptodate(page); | |
8fb5debc | 5296 | |
f6191471 AR |
5297 | /* Add shared, newly allocated pages to the page cache. */ |
5298 | if (vm_shared && !is_continue) { | |
1e392147 AA |
5299 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
5300 | ret = -EFAULT; | |
5301 | if (idx >= size) | |
5302 | goto out_release_nounlock; | |
1c9e8def | 5303 | |
1e392147 AA |
5304 | /* |
5305 | * Serialization between remove_inode_hugepages() and | |
5306 | * huge_add_to_page_cache() below happens through the | |
5307 | * hugetlb_fault_mutex_table that here must be hold by | |
5308 | * the caller. | |
5309 | */ | |
1c9e8def MK |
5310 | ret = huge_add_to_page_cache(page, mapping, idx); |
5311 | if (ret) | |
5312 | goto out_release_nounlock; | |
c7b1850d | 5313 | new_pagecache_page = true; |
1c9e8def MK |
5314 | } |
5315 | ||
8fb5debc MK |
5316 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
5317 | spin_lock(ptl); | |
5318 | ||
1e392147 AA |
5319 | /* |
5320 | * Recheck the i_size after holding PT lock to make sure not | |
5321 | * to leave any page mapped (as page_mapped()) beyond the end | |
5322 | * of the i_size (remove_inode_hugepages() is strict about | |
5323 | * enforcing that). If we bail out here, we'll also leave a | |
5324 | * page in the radix tree in the vm_shared case beyond the end | |
5325 | * of the i_size, but remove_inode_hugepages() will take care | |
5326 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
5327 | */ | |
5328 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
5329 | ret = -EFAULT; | |
5330 | if (idx >= size) | |
5331 | goto out_release_unlock; | |
5332 | ||
8fb5debc MK |
5333 | ret = -EEXIST; |
5334 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
5335 | goto out_release_unlock; | |
5336 | ||
1c9e8def MK |
5337 | if (vm_shared) { |
5338 | page_dup_rmap(page, true); | |
5339 | } else { | |
d6995da3 | 5340 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
5341 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
5342 | } | |
8fb5debc | 5343 | |
f6191471 AR |
5344 | /* For CONTINUE on a non-shared VMA, don't set VM_WRITE for CoW. */ |
5345 | if (is_continue && !vm_shared) | |
5346 | writable = 0; | |
5347 | else | |
5348 | writable = dst_vma->vm_flags & VM_WRITE; | |
5349 | ||
5350 | _dst_pte = make_huge_pte(dst_vma, page, writable); | |
5351 | if (writable) | |
8fb5debc MK |
5352 | _dst_pte = huge_pte_mkdirty(_dst_pte); |
5353 | _dst_pte = pte_mkyoung(_dst_pte); | |
5354 | ||
5355 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
5356 | ||
5357 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
5358 | dst_vma->vm_flags & VM_WRITE); | |
5359 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
5360 | ||
5361 | /* No need to invalidate - it was non-present before */ | |
5362 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
5363 | ||
5364 | spin_unlock(ptl); | |
f6191471 AR |
5365 | if (!is_continue) |
5366 | SetHPageMigratable(page); | |
5367 | if (vm_shared || is_continue) | |
1c9e8def | 5368 | unlock_page(page); |
8fb5debc MK |
5369 | ret = 0; |
5370 | out: | |
5371 | return ret; | |
5372 | out_release_unlock: | |
5373 | spin_unlock(ptl); | |
f6191471 | 5374 | if (vm_shared || is_continue) |
1c9e8def | 5375 | unlock_page(page); |
5af10dfd | 5376 | out_release_nounlock: |
c7b1850d MK |
5377 | if (!new_pagecache_page) |
5378 | restore_reserve_on_error(h, dst_vma, dst_addr, page); | |
8fb5debc MK |
5379 | put_page(page); |
5380 | goto out; | |
5381 | } | |
714c1891 | 5382 | #endif /* CONFIG_USERFAULTFD */ |
8fb5debc | 5383 | |
82e5d378 JM |
5384 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
5385 | int refs, struct page **pages, | |
5386 | struct vm_area_struct **vmas) | |
5387 | { | |
5388 | int nr; | |
5389 | ||
5390 | for (nr = 0; nr < refs; nr++) { | |
5391 | if (likely(pages)) | |
5392 | pages[nr] = mem_map_offset(page, nr); | |
5393 | if (vmas) | |
5394 | vmas[nr] = vma; | |
5395 | } | |
5396 | } | |
5397 | ||
28a35716 ML |
5398 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
5399 | struct page **pages, struct vm_area_struct **vmas, | |
5400 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 5401 | long i, unsigned int flags, int *locked) |
63551ae0 | 5402 | { |
d5d4b0aa KC |
5403 | unsigned long pfn_offset; |
5404 | unsigned long vaddr = *position; | |
28a35716 | 5405 | unsigned long remainder = *nr_pages; |
a5516438 | 5406 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 5407 | int err = -EFAULT, refs; |
63551ae0 | 5408 | |
63551ae0 | 5409 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 5410 | pte_t *pte; |
cb900f41 | 5411 | spinlock_t *ptl = NULL; |
2a15efc9 | 5412 | int absent; |
4c887265 | 5413 | struct page *page; |
63551ae0 | 5414 | |
02057967 DR |
5415 | /* |
5416 | * If we have a pending SIGKILL, don't keep faulting pages and | |
5417 | * potentially allocating memory. | |
5418 | */ | |
fa45f116 | 5419 | if (fatal_signal_pending(current)) { |
02057967 DR |
5420 | remainder = 0; |
5421 | break; | |
5422 | } | |
5423 | ||
4c887265 AL |
5424 | /* |
5425 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 5426 | * each hugepage. We have to make sure we get the |
4c887265 | 5427 | * first, for the page indexing below to work. |
cb900f41 KS |
5428 | * |
5429 | * Note that page table lock is not held when pte is null. | |
4c887265 | 5430 | */ |
7868a208 PA |
5431 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
5432 | huge_page_size(h)); | |
cb900f41 KS |
5433 | if (pte) |
5434 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
5435 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
5436 | ||
5437 | /* | |
5438 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
5439 | * an error where there's an empty slot with no huge pagecache |
5440 | * to back it. This way, we avoid allocating a hugepage, and | |
5441 | * the sparse dumpfile avoids allocating disk blocks, but its | |
5442 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 5443 | */ |
3ae77f43 HD |
5444 | if (absent && (flags & FOLL_DUMP) && |
5445 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
5446 | if (pte) |
5447 | spin_unlock(ptl); | |
2a15efc9 HD |
5448 | remainder = 0; |
5449 | break; | |
5450 | } | |
63551ae0 | 5451 | |
9cc3a5bd NH |
5452 | /* |
5453 | * We need call hugetlb_fault for both hugepages under migration | |
5454 | * (in which case hugetlb_fault waits for the migration,) and | |
5455 | * hwpoisoned hugepages (in which case we need to prevent the | |
5456 | * caller from accessing to them.) In order to do this, we use | |
5457 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
5458 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
5459 | * both cases, and because we can't follow correct pages | |
5460 | * directly from any kind of swap entries. | |
5461 | */ | |
5462 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
5463 | ((flags & FOLL_WRITE) && |
5464 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 5465 | vm_fault_t ret; |
87ffc118 | 5466 | unsigned int fault_flags = 0; |
63551ae0 | 5467 | |
cb900f41 KS |
5468 | if (pte) |
5469 | spin_unlock(ptl); | |
87ffc118 AA |
5470 | if (flags & FOLL_WRITE) |
5471 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 5472 | if (locked) |
71335f37 PX |
5473 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
5474 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
5475 | if (flags & FOLL_NOWAIT) |
5476 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
5477 | FAULT_FLAG_RETRY_NOWAIT; | |
5478 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
5479 | /* |
5480 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
5481 | * FAULT_FLAG_TRIED can co-exist | |
5482 | */ | |
87ffc118 AA |
5483 | fault_flags |= FAULT_FLAG_TRIED; |
5484 | } | |
5485 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
5486 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 5487 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
5488 | remainder = 0; |
5489 | break; | |
5490 | } | |
5491 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 5492 | if (locked && |
1ac25013 | 5493 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 5494 | *locked = 0; |
87ffc118 AA |
5495 | *nr_pages = 0; |
5496 | /* | |
5497 | * VM_FAULT_RETRY must not return an | |
5498 | * error, it will return zero | |
5499 | * instead. | |
5500 | * | |
5501 | * No need to update "position" as the | |
5502 | * caller will not check it after | |
5503 | * *nr_pages is set to 0. | |
5504 | */ | |
5505 | return i; | |
5506 | } | |
5507 | continue; | |
4c887265 AL |
5508 | } |
5509 | ||
a5516438 | 5510 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 5511 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 5512 | |
acbfb087 ZL |
5513 | /* |
5514 | * If subpage information not requested, update counters | |
5515 | * and skip the same_page loop below. | |
5516 | */ | |
5517 | if (!pages && !vmas && !pfn_offset && | |
5518 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
5519 | (remainder >= pages_per_huge_page(h))) { | |
5520 | vaddr += huge_page_size(h); | |
5521 | remainder -= pages_per_huge_page(h); | |
5522 | i += pages_per_huge_page(h); | |
5523 | spin_unlock(ptl); | |
5524 | continue; | |
5525 | } | |
5526 | ||
d08af0a5 JM |
5527 | /* vaddr may not be aligned to PAGE_SIZE */ |
5528 | refs = min3(pages_per_huge_page(h) - pfn_offset, remainder, | |
5529 | (vma->vm_end - ALIGN_DOWN(vaddr, PAGE_SIZE)) >> PAGE_SHIFT); | |
0fa5bc40 | 5530 | |
82e5d378 JM |
5531 | if (pages || vmas) |
5532 | record_subpages_vmas(mem_map_offset(page, pfn_offset), | |
5533 | vma, refs, | |
5534 | likely(pages) ? pages + i : NULL, | |
5535 | vmas ? vmas + i : NULL); | |
63551ae0 | 5536 | |
82e5d378 | 5537 | if (pages) { |
0fa5bc40 JM |
5538 | /* |
5539 | * try_grab_compound_head() should always succeed here, | |
5540 | * because: a) we hold the ptl lock, and b) we've just | |
5541 | * checked that the huge page is present in the page | |
5542 | * tables. If the huge page is present, then the tail | |
5543 | * pages must also be present. The ptl prevents the | |
5544 | * head page and tail pages from being rearranged in | |
5545 | * any way. So this page must be available at this | |
5546 | * point, unless the page refcount overflowed: | |
5547 | */ | |
82e5d378 | 5548 | if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], |
0fa5bc40 JM |
5549 | refs, |
5550 | flags))) { | |
5551 | spin_unlock(ptl); | |
5552 | remainder = 0; | |
5553 | err = -ENOMEM; | |
5554 | break; | |
5555 | } | |
d5d4b0aa | 5556 | } |
82e5d378 JM |
5557 | |
5558 | vaddr += (refs << PAGE_SHIFT); | |
5559 | remainder -= refs; | |
5560 | i += refs; | |
5561 | ||
cb900f41 | 5562 | spin_unlock(ptl); |
63551ae0 | 5563 | } |
28a35716 | 5564 | *nr_pages = remainder; |
87ffc118 AA |
5565 | /* |
5566 | * setting position is actually required only if remainder is | |
5567 | * not zero but it's faster not to add a "if (remainder)" | |
5568 | * branch. | |
5569 | */ | |
63551ae0 DG |
5570 | *position = vaddr; |
5571 | ||
2be7cfed | 5572 | return i ? i : err; |
63551ae0 | 5573 | } |
8f860591 | 5574 | |
7da4d641 | 5575 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
5576 | unsigned long address, unsigned long end, pgprot_t newprot) |
5577 | { | |
5578 | struct mm_struct *mm = vma->vm_mm; | |
5579 | unsigned long start = address; | |
5580 | pte_t *ptep; | |
5581 | pte_t pte; | |
a5516438 | 5582 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 5583 | unsigned long pages = 0; |
dff11abe | 5584 | bool shared_pmd = false; |
ac46d4f3 | 5585 | struct mmu_notifier_range range; |
dff11abe MK |
5586 | |
5587 | /* | |
5588 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 5589 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
5590 | * range if PMD sharing is possible. |
5591 | */ | |
7269f999 JG |
5592 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
5593 | 0, vma, mm, start, end); | |
ac46d4f3 | 5594 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
5595 | |
5596 | BUG_ON(address >= end); | |
ac46d4f3 | 5597 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 5598 | |
ac46d4f3 | 5599 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 5600 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 5601 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 5602 | spinlock_t *ptl; |
7868a208 | 5603 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
5604 | if (!ptep) |
5605 | continue; | |
cb900f41 | 5606 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 5607 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
7da4d641 | 5608 | pages++; |
cb900f41 | 5609 | spin_unlock(ptl); |
dff11abe | 5610 | shared_pmd = true; |
39dde65c | 5611 | continue; |
7da4d641 | 5612 | } |
a8bda28d NH |
5613 | pte = huge_ptep_get(ptep); |
5614 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
5615 | spin_unlock(ptl); | |
5616 | continue; | |
5617 | } | |
5618 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
5619 | swp_entry_t entry = pte_to_swp_entry(pte); | |
5620 | ||
4dd845b5 | 5621 | if (is_writable_migration_entry(entry)) { |
a8bda28d NH |
5622 | pte_t newpte; |
5623 | ||
4dd845b5 AP |
5624 | entry = make_readable_migration_entry( |
5625 | swp_offset(entry)); | |
a8bda28d | 5626 | newpte = swp_entry_to_pte(entry); |
e5251fd4 PA |
5627 | set_huge_swap_pte_at(mm, address, ptep, |
5628 | newpte, huge_page_size(h)); | |
a8bda28d NH |
5629 | pages++; |
5630 | } | |
5631 | spin_unlock(ptl); | |
5632 | continue; | |
5633 | } | |
5634 | if (!huge_pte_none(pte)) { | |
023bdd00 | 5635 | pte_t old_pte; |
79c1c594 | 5636 | unsigned int shift = huge_page_shift(hstate_vma(vma)); |
023bdd00 AK |
5637 | |
5638 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
5639 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
79c1c594 | 5640 | pte = arch_make_huge_pte(pte, shift, vma->vm_flags); |
023bdd00 | 5641 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 5642 | pages++; |
8f860591 | 5643 | } |
cb900f41 | 5644 | spin_unlock(ptl); |
8f860591 | 5645 | } |
d833352a | 5646 | /* |
c8c06efa | 5647 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 5648 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 5649 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
5650 | * and that page table be reused and filled with junk. If we actually |
5651 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 5652 | */ |
dff11abe | 5653 | if (shared_pmd) |
ac46d4f3 | 5654 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
5655 | else |
5656 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
5657 | /* |
5658 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
5659 | * page table protection not changing it to point to a new page. | |
5660 | * | |
ad56b738 | 5661 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 5662 | */ |
83cde9e8 | 5663 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 5664 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
5665 | |
5666 | return pages << h->order; | |
8f860591 ZY |
5667 | } |
5668 | ||
33b8f84a MK |
5669 | /* Return true if reservation was successful, false otherwise. */ |
5670 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 5671 | long from, long to, |
5a6fe125 | 5672 | struct vm_area_struct *vma, |
ca16d140 | 5673 | vm_flags_t vm_flags) |
e4e574b7 | 5674 | { |
33b8f84a | 5675 | long chg, add = -1; |
a5516438 | 5676 | struct hstate *h = hstate_inode(inode); |
90481622 | 5677 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 5678 | struct resv_map *resv_map; |
075a61d0 | 5679 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 5680 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 5681 | |
63489f8e MK |
5682 | /* This should never happen */ |
5683 | if (from > to) { | |
5684 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 5685 | return false; |
63489f8e MK |
5686 | } |
5687 | ||
17c9d12e MG |
5688 | /* |
5689 | * Only apply hugepage reservation if asked. At fault time, an | |
5690 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 5691 | * without using reserves |
17c9d12e | 5692 | */ |
ca16d140 | 5693 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 5694 | return true; |
17c9d12e | 5695 | |
a1e78772 MG |
5696 | /* |
5697 | * Shared mappings base their reservation on the number of pages that | |
5698 | * are already allocated on behalf of the file. Private mappings need | |
5699 | * to reserve the full area even if read-only as mprotect() may be | |
5700 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
5701 | */ | |
9119a41e | 5702 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
5703 | /* |
5704 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
5705 | * called for inodes for which resv_maps were created (see | |
5706 | * hugetlbfs_get_inode). | |
5707 | */ | |
4e35f483 | 5708 | resv_map = inode_resv_map(inode); |
9119a41e | 5709 | |
0db9d74e | 5710 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
5711 | |
5712 | } else { | |
e9fe92ae | 5713 | /* Private mapping. */ |
9119a41e | 5714 | resv_map = resv_map_alloc(); |
17c9d12e | 5715 | if (!resv_map) |
33b8f84a | 5716 | return false; |
17c9d12e | 5717 | |
a1e78772 | 5718 | chg = to - from; |
84afd99b | 5719 | |
17c9d12e MG |
5720 | set_vma_resv_map(vma, resv_map); |
5721 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
5722 | } | |
5723 | ||
33b8f84a | 5724 | if (chg < 0) |
c50ac050 | 5725 | goto out_err; |
8a630112 | 5726 | |
33b8f84a MK |
5727 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
5728 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 5729 | goto out_err; |
075a61d0 MA |
5730 | |
5731 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
5732 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
5733 | * of the resv_map. | |
5734 | */ | |
5735 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
5736 | } | |
5737 | ||
1c5ecae3 MK |
5738 | /* |
5739 | * There must be enough pages in the subpool for the mapping. If | |
5740 | * the subpool has a minimum size, there may be some global | |
5741 | * reservations already in place (gbl_reserve). | |
5742 | */ | |
5743 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 5744 | if (gbl_reserve < 0) |
075a61d0 | 5745 | goto out_uncharge_cgroup; |
5a6fe125 MG |
5746 | |
5747 | /* | |
17c9d12e | 5748 | * Check enough hugepages are available for the reservation. |
90481622 | 5749 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 5750 | */ |
33b8f84a | 5751 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 5752 | goto out_put_pages; |
17c9d12e MG |
5753 | |
5754 | /* | |
5755 | * Account for the reservations made. Shared mappings record regions | |
5756 | * that have reservations as they are shared by multiple VMAs. | |
5757 | * When the last VMA disappears, the region map says how much | |
5758 | * the reservation was and the page cache tells how much of | |
5759 | * the reservation was consumed. Private mappings are per-VMA and | |
5760 | * only the consumed reservations are tracked. When the VMA | |
5761 | * disappears, the original reservation is the VMA size and the | |
5762 | * consumed reservations are stored in the map. Hence, nothing | |
5763 | * else has to be done for private mappings here | |
5764 | */ | |
33039678 | 5765 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 5766 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
5767 | |
5768 | if (unlikely(add < 0)) { | |
5769 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 5770 | goto out_put_pages; |
0db9d74e | 5771 | } else if (unlikely(chg > add)) { |
33039678 MK |
5772 | /* |
5773 | * pages in this range were added to the reserve | |
5774 | * map between region_chg and region_add. This | |
5775 | * indicates a race with alloc_huge_page. Adjust | |
5776 | * the subpool and reserve counts modified above | |
5777 | * based on the difference. | |
5778 | */ | |
5779 | long rsv_adjust; | |
5780 | ||
d85aecf2 ML |
5781 | /* |
5782 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
5783 | * reference to h_cg->css. See comment below for detail. | |
5784 | */ | |
075a61d0 MA |
5785 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
5786 | hstate_index(h), | |
5787 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5788 | ||
33039678 MK |
5789 | rsv_adjust = hugepage_subpool_put_pages(spool, |
5790 | chg - add); | |
5791 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
5792 | } else if (h_cg) { |
5793 | /* | |
5794 | * The file_regions will hold their own reference to | |
5795 | * h_cg->css. So we should release the reference held | |
5796 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
5797 | * done. | |
5798 | */ | |
5799 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
5800 | } |
5801 | } | |
33b8f84a MK |
5802 | return true; |
5803 | ||
075a61d0 MA |
5804 | out_put_pages: |
5805 | /* put back original number of pages, chg */ | |
5806 | (void)hugepage_subpool_put_pages(spool, chg); | |
5807 | out_uncharge_cgroup: | |
5808 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5809 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 5810 | out_err: |
5e911373 | 5811 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
5812 | /* Only call region_abort if the region_chg succeeded but the |
5813 | * region_add failed or didn't run. | |
5814 | */ | |
5815 | if (chg >= 0 && add < 0) | |
5816 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
5817 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
5818 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 5819 | return false; |
a43a8c39 KC |
5820 | } |
5821 | ||
b5cec28d MK |
5822 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
5823 | long freed) | |
a43a8c39 | 5824 | { |
a5516438 | 5825 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 5826 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 5827 | long chg = 0; |
90481622 | 5828 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 5829 | long gbl_reserve; |
45c682a6 | 5830 | |
f27a5136 MK |
5831 | /* |
5832 | * Since this routine can be called in the evict inode path for all | |
5833 | * hugetlbfs inodes, resv_map could be NULL. | |
5834 | */ | |
b5cec28d MK |
5835 | if (resv_map) { |
5836 | chg = region_del(resv_map, start, end); | |
5837 | /* | |
5838 | * region_del() can fail in the rare case where a region | |
5839 | * must be split and another region descriptor can not be | |
5840 | * allocated. If end == LONG_MAX, it will not fail. | |
5841 | */ | |
5842 | if (chg < 0) | |
5843 | return chg; | |
5844 | } | |
5845 | ||
45c682a6 | 5846 | spin_lock(&inode->i_lock); |
e4c6f8be | 5847 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
5848 | spin_unlock(&inode->i_lock); |
5849 | ||
1c5ecae3 MK |
5850 | /* |
5851 | * If the subpool has a minimum size, the number of global | |
5852 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
5853 | * |
5854 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
5855 | * won't go negative. | |
1c5ecae3 MK |
5856 | */ |
5857 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5858 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
5859 | |
5860 | return 0; | |
a43a8c39 | 5861 | } |
93f70f90 | 5862 | |
3212b535 SC |
5863 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
5864 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5865 | struct vm_area_struct *vma, | |
5866 | unsigned long addr, pgoff_t idx) | |
5867 | { | |
5868 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5869 | svma->vm_start; | |
5870 | unsigned long sbase = saddr & PUD_MASK; | |
5871 | unsigned long s_end = sbase + PUD_SIZE; | |
5872 | ||
5873 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
5874 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
5875 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
5876 | |
5877 | /* | |
5878 | * match the virtual addresses, permission and the alignment of the | |
5879 | * page table page. | |
5880 | */ | |
5881 | if (pmd_index(addr) != pmd_index(saddr) || | |
5882 | vm_flags != svm_flags || | |
07e51edf | 5883 | !range_in_vma(svma, sbase, s_end)) |
3212b535 SC |
5884 | return 0; |
5885 | ||
5886 | return saddr; | |
5887 | } | |
5888 | ||
31aafb45 | 5889 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
5890 | { |
5891 | unsigned long base = addr & PUD_MASK; | |
5892 | unsigned long end = base + PUD_SIZE; | |
5893 | ||
5894 | /* | |
5895 | * check on proper vm_flags and page table alignment | |
5896 | */ | |
017b1660 | 5897 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
5898 | return true; |
5899 | return false; | |
3212b535 SC |
5900 | } |
5901 | ||
c1991e07 PX |
5902 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
5903 | { | |
5904 | #ifdef CONFIG_USERFAULTFD | |
5905 | if (uffd_disable_huge_pmd_share(vma)) | |
5906 | return false; | |
5907 | #endif | |
5908 | return vma_shareable(vma, addr); | |
5909 | } | |
5910 | ||
017b1660 MK |
5911 | /* |
5912 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5913 | * If yes, adjust start and end to cover range associated with possible | |
5914 | * shared pmd mappings. | |
5915 | */ | |
5916 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5917 | unsigned long *start, unsigned long *end) | |
5918 | { | |
a1ba9da8 LX |
5919 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
5920 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 5921 | |
a1ba9da8 | 5922 | /* |
f0953a1b IM |
5923 | * vma needs to span at least one aligned PUD size, and the range |
5924 | * must be at least partially within in. | |
a1ba9da8 LX |
5925 | */ |
5926 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
5927 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
5928 | return; |
5929 | ||
75802ca6 | 5930 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
5931 | if (*start > v_start) |
5932 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 5933 | |
a1ba9da8 LX |
5934 | if (*end < v_end) |
5935 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
5936 | } |
5937 | ||
3212b535 SC |
5938 | /* |
5939 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5940 | * and returns the corresponding pte. While this is not necessary for the | |
5941 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
5942 | * code much cleaner. |
5943 | * | |
0bf7b64e MK |
5944 | * This routine must be called with i_mmap_rwsem held in at least read mode if |
5945 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
5946 | * table entries associated with the address space. This is important as we | |
5947 | * are setting up sharing based on existing page table entries (mappings). | |
5948 | * | |
5949 | * NOTE: This routine is only called from huge_pte_alloc. Some callers of | |
5950 | * huge_pte_alloc know that sharing is not possible and do not take | |
5951 | * i_mmap_rwsem as a performance optimization. This is handled by the | |
5952 | * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is | |
5953 | * only required for subsequent processing. | |
3212b535 | 5954 | */ |
aec44e0f PX |
5955 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
5956 | unsigned long addr, pud_t *pud) | |
3212b535 | 5957 | { |
3212b535 SC |
5958 | struct address_space *mapping = vma->vm_file->f_mapping; |
5959 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5960 | vma->vm_pgoff; | |
5961 | struct vm_area_struct *svma; | |
5962 | unsigned long saddr; | |
5963 | pte_t *spte = NULL; | |
5964 | pte_t *pte; | |
cb900f41 | 5965 | spinlock_t *ptl; |
3212b535 | 5966 | |
0bf7b64e | 5967 | i_mmap_assert_locked(mapping); |
3212b535 SC |
5968 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
5969 | if (svma == vma) | |
5970 | continue; | |
5971 | ||
5972 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5973 | if (saddr) { | |
7868a208 PA |
5974 | spte = huge_pte_offset(svma->vm_mm, saddr, |
5975 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
5976 | if (spte) { |
5977 | get_page(virt_to_page(spte)); | |
5978 | break; | |
5979 | } | |
5980 | } | |
5981 | } | |
5982 | ||
5983 | if (!spte) | |
5984 | goto out; | |
5985 | ||
8bea8052 | 5986 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 5987 | if (pud_none(*pud)) { |
3212b535 SC |
5988 | pud_populate(mm, pud, |
5989 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5990 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5991 | } else { |
3212b535 | 5992 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5993 | } |
cb900f41 | 5994 | spin_unlock(ptl); |
3212b535 SC |
5995 | out: |
5996 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
5997 | return pte; |
5998 | } | |
5999 | ||
6000 | /* | |
6001 | * unmap huge page backed by shared pte. | |
6002 | * | |
6003 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
6004 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
6005 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
6006 | * | |
c0d0381a | 6007 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
6008 | * |
6009 | * returns: 1 successfully unmapped a shared pte page | |
6010 | * 0 the underlying pte page is not shared, or it is the last user | |
6011 | */ | |
34ae204f MK |
6012 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
6013 | unsigned long *addr, pte_t *ptep) | |
3212b535 SC |
6014 | { |
6015 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
6016 | p4d_t *p4d = p4d_offset(pgd, *addr); |
6017 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 | 6018 | |
34ae204f | 6019 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
6020 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
6021 | if (page_count(virt_to_page(ptep)) == 1) | |
6022 | return 0; | |
6023 | ||
6024 | pud_clear(pud); | |
6025 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 6026 | mm_dec_nr_pmds(mm); |
3212b535 SC |
6027 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
6028 | return 1; | |
6029 | } | |
c1991e07 | 6030 | |
9e5fc74c | 6031 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
aec44e0f PX |
6032 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
6033 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
6034 | { |
6035 | return NULL; | |
6036 | } | |
e81f2d22 | 6037 | |
34ae204f MK |
6038 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
6039 | unsigned long *addr, pte_t *ptep) | |
e81f2d22 ZZ |
6040 | { |
6041 | return 0; | |
6042 | } | |
017b1660 MK |
6043 | |
6044 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
6045 | unsigned long *start, unsigned long *end) | |
6046 | { | |
6047 | } | |
c1991e07 PX |
6048 | |
6049 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
6050 | { | |
6051 | return false; | |
6052 | } | |
3212b535 SC |
6053 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
6054 | ||
9e5fc74c | 6055 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 6056 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
6057 | unsigned long addr, unsigned long sz) |
6058 | { | |
6059 | pgd_t *pgd; | |
c2febafc | 6060 | p4d_t *p4d; |
9e5fc74c SC |
6061 | pud_t *pud; |
6062 | pte_t *pte = NULL; | |
6063 | ||
6064 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
6065 | p4d = p4d_alloc(mm, pgd, addr); |
6066 | if (!p4d) | |
6067 | return NULL; | |
c2febafc | 6068 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
6069 | if (pud) { |
6070 | if (sz == PUD_SIZE) { | |
6071 | pte = (pte_t *)pud; | |
6072 | } else { | |
6073 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 6074 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 6075 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
6076 | else |
6077 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
6078 | } | |
6079 | } | |
4e666314 | 6080 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
6081 | |
6082 | return pte; | |
6083 | } | |
6084 | ||
9b19df29 PA |
6085 | /* |
6086 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
6087 | * entry at address @addr | |
6088 | * | |
8ac0b81a LX |
6089 | * Return: Pointer to page table entry (PUD or PMD) for |
6090 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
6091 | * size @sz doesn't match the hugepage size at this level of the page |
6092 | * table. | |
6093 | */ | |
7868a208 PA |
6094 | pte_t *huge_pte_offset(struct mm_struct *mm, |
6095 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
6096 | { |
6097 | pgd_t *pgd; | |
c2febafc | 6098 | p4d_t *p4d; |
8ac0b81a LX |
6099 | pud_t *pud; |
6100 | pmd_t *pmd; | |
9e5fc74c SC |
6101 | |
6102 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
6103 | if (!pgd_present(*pgd)) |
6104 | return NULL; | |
6105 | p4d = p4d_offset(pgd, addr); | |
6106 | if (!p4d_present(*p4d)) | |
6107 | return NULL; | |
9b19df29 | 6108 | |
c2febafc | 6109 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
6110 | if (sz == PUD_SIZE) |
6111 | /* must be pud huge, non-present or none */ | |
c2febafc | 6112 | return (pte_t *)pud; |
8ac0b81a | 6113 | if (!pud_present(*pud)) |
9b19df29 | 6114 | return NULL; |
8ac0b81a | 6115 | /* must have a valid entry and size to go further */ |
9b19df29 | 6116 | |
8ac0b81a LX |
6117 | pmd = pmd_offset(pud, addr); |
6118 | /* must be pmd huge, non-present or none */ | |
6119 | return (pte_t *)pmd; | |
9e5fc74c SC |
6120 | } |
6121 | ||
61f77eda NH |
6122 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
6123 | ||
6124 | /* | |
6125 | * These functions are overwritable if your architecture needs its own | |
6126 | * behavior. | |
6127 | */ | |
6128 | struct page * __weak | |
6129 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
6130 | int write) | |
6131 | { | |
6132 | return ERR_PTR(-EINVAL); | |
6133 | } | |
6134 | ||
4dc71451 AK |
6135 | struct page * __weak |
6136 | follow_huge_pd(struct vm_area_struct *vma, | |
6137 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
6138 | { | |
6139 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
6140 | return NULL; | |
6141 | } | |
6142 | ||
61f77eda | 6143 | struct page * __weak |
9e5fc74c | 6144 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 6145 | pmd_t *pmd, int flags) |
9e5fc74c | 6146 | { |
e66f17ff NH |
6147 | struct page *page = NULL; |
6148 | spinlock_t *ptl; | |
c9d398fa | 6149 | pte_t pte; |
3faa52c0 JH |
6150 | |
6151 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
6152 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
6153 | (FOLL_PIN | FOLL_GET))) | |
6154 | return NULL; | |
6155 | ||
e66f17ff NH |
6156 | retry: |
6157 | ptl = pmd_lockptr(mm, pmd); | |
6158 | spin_lock(ptl); | |
6159 | /* | |
6160 | * make sure that the address range covered by this pmd is not | |
6161 | * unmapped from other threads. | |
6162 | */ | |
6163 | if (!pmd_huge(*pmd)) | |
6164 | goto out; | |
c9d398fa NH |
6165 | pte = huge_ptep_get((pte_t *)pmd); |
6166 | if (pte_present(pte)) { | |
97534127 | 6167 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
6168 | /* |
6169 | * try_grab_page() should always succeed here, because: a) we | |
6170 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
6171 | * huge pmd (head) page is present in the page tables. The ptl | |
6172 | * prevents the head page and tail pages from being rearranged | |
6173 | * in any way. So this page must be available at this point, | |
6174 | * unless the page refcount overflowed: | |
6175 | */ | |
6176 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
6177 | page = NULL; | |
6178 | goto out; | |
6179 | } | |
e66f17ff | 6180 | } else { |
c9d398fa | 6181 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
6182 | spin_unlock(ptl); |
6183 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
6184 | goto retry; | |
6185 | } | |
6186 | /* | |
6187 | * hwpoisoned entry is treated as no_page_table in | |
6188 | * follow_page_mask(). | |
6189 | */ | |
6190 | } | |
6191 | out: | |
6192 | spin_unlock(ptl); | |
9e5fc74c SC |
6193 | return page; |
6194 | } | |
6195 | ||
61f77eda | 6196 | struct page * __weak |
9e5fc74c | 6197 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 6198 | pud_t *pud, int flags) |
9e5fc74c | 6199 | { |
3faa52c0 | 6200 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 6201 | return NULL; |
9e5fc74c | 6202 | |
e66f17ff | 6203 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
6204 | } |
6205 | ||
faaa5b62 AK |
6206 | struct page * __weak |
6207 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
6208 | { | |
3faa52c0 | 6209 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
6210 | return NULL; |
6211 | ||
6212 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
6213 | } | |
6214 | ||
31caf665 NH |
6215 | bool isolate_huge_page(struct page *page, struct list_head *list) |
6216 | { | |
bcc54222 NH |
6217 | bool ret = true; |
6218 | ||
db71ef79 | 6219 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
6220 | if (!PageHeadHuge(page) || |
6221 | !HPageMigratable(page) || | |
0eb2df2b | 6222 | !get_page_unless_zero(page)) { |
bcc54222 NH |
6223 | ret = false; |
6224 | goto unlock; | |
6225 | } | |
8f251a3d | 6226 | ClearHPageMigratable(page); |
31caf665 | 6227 | list_move_tail(&page->lru, list); |
bcc54222 | 6228 | unlock: |
db71ef79 | 6229 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 6230 | return ret; |
31caf665 NH |
6231 | } |
6232 | ||
25182f05 NH |
6233 | int get_hwpoison_huge_page(struct page *page, bool *hugetlb) |
6234 | { | |
6235 | int ret = 0; | |
6236 | ||
6237 | *hugetlb = false; | |
6238 | spin_lock_irq(&hugetlb_lock); | |
6239 | if (PageHeadHuge(page)) { | |
6240 | *hugetlb = true; | |
6241 | if (HPageFreed(page) || HPageMigratable(page)) | |
6242 | ret = get_page_unless_zero(page); | |
0ed950d1 NH |
6243 | else |
6244 | ret = -EBUSY; | |
25182f05 NH |
6245 | } |
6246 | spin_unlock_irq(&hugetlb_lock); | |
6247 | return ret; | |
6248 | } | |
6249 | ||
31caf665 NH |
6250 | void putback_active_hugepage(struct page *page) |
6251 | { | |
db71ef79 | 6252 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 6253 | SetHPageMigratable(page); |
31caf665 | 6254 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 6255 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
6256 | put_page(page); |
6257 | } | |
ab5ac90a MH |
6258 | |
6259 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
6260 | { | |
6261 | struct hstate *h = page_hstate(oldpage); | |
6262 | ||
6263 | hugetlb_cgroup_migrate(oldpage, newpage); | |
6264 | set_page_owner_migrate_reason(newpage, reason); | |
6265 | ||
6266 | /* | |
6267 | * transfer temporary state of the new huge page. This is | |
6268 | * reverse to other transitions because the newpage is going to | |
6269 | * be final while the old one will be freed so it takes over | |
6270 | * the temporary status. | |
6271 | * | |
6272 | * Also note that we have to transfer the per-node surplus state | |
6273 | * here as well otherwise the global surplus count will not match | |
6274 | * the per-node's. | |
6275 | */ | |
9157c311 | 6276 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
6277 | int old_nid = page_to_nid(oldpage); |
6278 | int new_nid = page_to_nid(newpage); | |
6279 | ||
9157c311 MK |
6280 | SetHPageTemporary(oldpage); |
6281 | ClearHPageTemporary(newpage); | |
ab5ac90a | 6282 | |
5af1ab1d ML |
6283 | /* |
6284 | * There is no need to transfer the per-node surplus state | |
6285 | * when we do not cross the node. | |
6286 | */ | |
6287 | if (new_nid == old_nid) | |
6288 | return; | |
db71ef79 | 6289 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
6290 | if (h->surplus_huge_pages_node[old_nid]) { |
6291 | h->surplus_huge_pages_node[old_nid]--; | |
6292 | h->surplus_huge_pages_node[new_nid]++; | |
6293 | } | |
db71ef79 | 6294 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
6295 | } |
6296 | } | |
cf11e85f | 6297 | |
6dfeaff9 PX |
6298 | /* |
6299 | * This function will unconditionally remove all the shared pmd pgtable entries | |
6300 | * within the specific vma for a hugetlbfs memory range. | |
6301 | */ | |
6302 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
6303 | { | |
6304 | struct hstate *h = hstate_vma(vma); | |
6305 | unsigned long sz = huge_page_size(h); | |
6306 | struct mm_struct *mm = vma->vm_mm; | |
6307 | struct mmu_notifier_range range; | |
6308 | unsigned long address, start, end; | |
6309 | spinlock_t *ptl; | |
6310 | pte_t *ptep; | |
6311 | ||
6312 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
6313 | return; | |
6314 | ||
6315 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
6316 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
6317 | ||
6318 | if (start >= end) | |
6319 | return; | |
6320 | ||
6321 | /* | |
6322 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
6323 | * we have already done the PUD_SIZE alignment. | |
6324 | */ | |
6325 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
6326 | start, end); | |
6327 | mmu_notifier_invalidate_range_start(&range); | |
6328 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
6329 | for (address = start; address < end; address += PUD_SIZE) { | |
6330 | unsigned long tmp = address; | |
6331 | ||
6332 | ptep = huge_pte_offset(mm, address, sz); | |
6333 | if (!ptep) | |
6334 | continue; | |
6335 | ptl = huge_pte_lock(h, mm, ptep); | |
6336 | /* We don't want 'address' to be changed */ | |
6337 | huge_pmd_unshare(mm, vma, &tmp, ptep); | |
6338 | spin_unlock(ptl); | |
6339 | } | |
6340 | flush_hugetlb_tlb_range(vma, start, end); | |
6341 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
6342 | /* | |
6343 | * No need to call mmu_notifier_invalidate_range(), see | |
6344 | * Documentation/vm/mmu_notifier.rst. | |
6345 | */ | |
6346 | mmu_notifier_invalidate_range_end(&range); | |
6347 | } | |
6348 | ||
cf11e85f | 6349 | #ifdef CONFIG_CMA |
cf11e85f RG |
6350 | static bool cma_reserve_called __initdata; |
6351 | ||
6352 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
6353 | { | |
6354 | hugetlb_cma_size = memparse(p, &p); | |
6355 | return 0; | |
6356 | } | |
6357 | ||
6358 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
6359 | ||
6360 | void __init hugetlb_cma_reserve(int order) | |
6361 | { | |
6362 | unsigned long size, reserved, per_node; | |
6363 | int nid; | |
6364 | ||
6365 | cma_reserve_called = true; | |
6366 | ||
6367 | if (!hugetlb_cma_size) | |
6368 | return; | |
6369 | ||
6370 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
6371 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
6372 | (PAGE_SIZE << order) / SZ_1M); | |
6373 | return; | |
6374 | } | |
6375 | ||
6376 | /* | |
6377 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
6378 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
6379 | */ | |
6380 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
6381 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
6382 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
6383 | ||
6384 | reserved = 0; | |
6385 | for_each_node_state(nid, N_ONLINE) { | |
6386 | int res; | |
2281f797 | 6387 | char name[CMA_MAX_NAME]; |
cf11e85f RG |
6388 | |
6389 | size = min(per_node, hugetlb_cma_size - reserved); | |
6390 | size = round_up(size, PAGE_SIZE << order); | |
6391 | ||
2281f797 | 6392 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
cf11e85f | 6393 | res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, |
29d0f41d | 6394 | 0, false, name, |
cf11e85f RG |
6395 | &hugetlb_cma[nid], nid); |
6396 | if (res) { | |
6397 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
6398 | res, nid); | |
6399 | continue; | |
6400 | } | |
6401 | ||
6402 | reserved += size; | |
6403 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
6404 | size / SZ_1M, nid); | |
6405 | ||
6406 | if (reserved >= hugetlb_cma_size) | |
6407 | break; | |
6408 | } | |
6409 | } | |
6410 | ||
6411 | void __init hugetlb_cma_check(void) | |
6412 | { | |
6413 | if (!hugetlb_cma_size || cma_reserve_called) | |
6414 | return; | |
6415 | ||
6416 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
6417 | } | |
6418 | ||
6419 | #endif /* CONFIG_CMA */ |