]>
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> |
63489f8e | 22 | #include <linux/mmdebug.h> |
174cd4b1 | 23 | #include <linux/sched/signal.h> |
0fe6e20b | 24 | #include <linux/rmap.h> |
c6247f72 | 25 | #include <linux/string_helpers.h> |
fd6a03ed NH |
26 | #include <linux/swap.h> |
27 | #include <linux/swapops.h> | |
8382d914 | 28 | #include <linux/jhash.h> |
98fa15f3 | 29 | #include <linux/numa.h> |
c77c0a8a | 30 | #include <linux/llist.h> |
d6606683 | 31 | |
63551ae0 DG |
32 | #include <asm/page.h> |
33 | #include <asm/pgtable.h> | |
24669e58 | 34 | #include <asm/tlb.h> |
63551ae0 | 35 | |
24669e58 | 36 | #include <linux/io.h> |
63551ae0 | 37 | #include <linux/hugetlb.h> |
9dd540e2 | 38 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 39 | #include <linux/node.h> |
1a1aad8a | 40 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 41 | #include <linux/page_owner.h> |
7835e98b | 42 | #include "internal.h" |
1da177e4 | 43 | |
c3f38a38 | 44 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
45 | unsigned int default_hstate_idx; |
46 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
641844f5 NH |
47 | /* |
48 | * Minimum page order among possible hugepage sizes, set to a proper value | |
49 | * at boot time. | |
50 | */ | |
51 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 52 | |
53ba51d2 JT |
53 | __initdata LIST_HEAD(huge_boot_pages); |
54 | ||
e5ff2159 AK |
55 | /* for command line parsing */ |
56 | static struct hstate * __initdata parsed_hstate; | |
57 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 58 | static unsigned long __initdata default_hstate_size; |
9fee021d | 59 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 60 | |
3935baa9 | 61 | /* |
31caf665 NH |
62 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
63 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 64 | */ |
c3f38a38 | 65 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 66 | |
8382d914 DB |
67 | /* |
68 | * Serializes faults on the same logical page. This is used to | |
69 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
70 | */ | |
71 | static int num_fault_mutexes; | |
c672c7f2 | 72 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 73 | |
7ca02d0a MK |
74 | /* Forward declaration */ |
75 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
76 | ||
90481622 DG |
77 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
78 | { | |
79 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
80 | ||
81 | spin_unlock(&spool->lock); | |
82 | ||
83 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
84 | * remain, give up any reservations mased on minimum size and |
85 | * free the subpool */ | |
86 | if (free) { | |
87 | if (spool->min_hpages != -1) | |
88 | hugetlb_acct_memory(spool->hstate, | |
89 | -spool->min_hpages); | |
90481622 | 90 | kfree(spool); |
7ca02d0a | 91 | } |
90481622 DG |
92 | } |
93 | ||
7ca02d0a MK |
94 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
95 | long min_hpages) | |
90481622 DG |
96 | { |
97 | struct hugepage_subpool *spool; | |
98 | ||
c6a91820 | 99 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
100 | if (!spool) |
101 | return NULL; | |
102 | ||
103 | spin_lock_init(&spool->lock); | |
104 | spool->count = 1; | |
7ca02d0a MK |
105 | spool->max_hpages = max_hpages; |
106 | spool->hstate = h; | |
107 | spool->min_hpages = min_hpages; | |
108 | ||
109 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
110 | kfree(spool); | |
111 | return NULL; | |
112 | } | |
113 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
114 | |
115 | return spool; | |
116 | } | |
117 | ||
118 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
119 | { | |
120 | spin_lock(&spool->lock); | |
121 | BUG_ON(!spool->count); | |
122 | spool->count--; | |
123 | unlock_or_release_subpool(spool); | |
124 | } | |
125 | ||
1c5ecae3 MK |
126 | /* |
127 | * Subpool accounting for allocating and reserving pages. | |
128 | * Return -ENOMEM if there are not enough resources to satisfy the | |
129 | * the request. Otherwise, return the number of pages by which the | |
130 | * global pools must be adjusted (upward). The returned value may | |
131 | * only be different than the passed value (delta) in the case where | |
132 | * a subpool minimum size must be manitained. | |
133 | */ | |
134 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
135 | long delta) |
136 | { | |
1c5ecae3 | 137 | long ret = delta; |
90481622 DG |
138 | |
139 | if (!spool) | |
1c5ecae3 | 140 | return ret; |
90481622 DG |
141 | |
142 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
143 | |
144 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
145 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
146 | spool->used_hpages += delta; | |
147 | else { | |
148 | ret = -ENOMEM; | |
149 | goto unlock_ret; | |
150 | } | |
90481622 | 151 | } |
90481622 | 152 | |
09a95e29 MK |
153 | /* minimum size accounting */ |
154 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
155 | if (delta > spool->rsv_hpages) { |
156 | /* | |
157 | * Asking for more reserves than those already taken on | |
158 | * behalf of subpool. Return difference. | |
159 | */ | |
160 | ret = delta - spool->rsv_hpages; | |
161 | spool->rsv_hpages = 0; | |
162 | } else { | |
163 | ret = 0; /* reserves already accounted for */ | |
164 | spool->rsv_hpages -= delta; | |
165 | } | |
166 | } | |
167 | ||
168 | unlock_ret: | |
169 | spin_unlock(&spool->lock); | |
90481622 DG |
170 | return ret; |
171 | } | |
172 | ||
1c5ecae3 MK |
173 | /* |
174 | * Subpool accounting for freeing and unreserving pages. | |
175 | * Return the number of global page reservations that must be dropped. | |
176 | * The return value may only be different than the passed value (delta) | |
177 | * in the case where a subpool minimum size must be maintained. | |
178 | */ | |
179 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
180 | long delta) |
181 | { | |
1c5ecae3 MK |
182 | long ret = delta; |
183 | ||
90481622 | 184 | if (!spool) |
1c5ecae3 | 185 | return delta; |
90481622 DG |
186 | |
187 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
188 | |
189 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
190 | spool->used_hpages -= delta; | |
191 | ||
09a95e29 MK |
192 | /* minimum size accounting */ |
193 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
194 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
195 | ret = 0; | |
196 | else | |
197 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
198 | ||
199 | spool->rsv_hpages += delta; | |
200 | if (spool->rsv_hpages > spool->min_hpages) | |
201 | spool->rsv_hpages = spool->min_hpages; | |
202 | } | |
203 | ||
204 | /* | |
205 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
206 | * quota reference, free it now. | |
207 | */ | |
90481622 | 208 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
209 | |
210 | return ret; | |
90481622 DG |
211 | } |
212 | ||
213 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
214 | { | |
215 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
216 | } | |
217 | ||
218 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
219 | { | |
496ad9aa | 220 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
221 | } |
222 | ||
0db9d74e MA |
223 | /* Helper that removes a struct file_region from the resv_map cache and returns |
224 | * it for use. | |
225 | */ | |
226 | static struct file_region * | |
227 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
228 | { | |
229 | struct file_region *nrg = NULL; | |
230 | ||
231 | VM_BUG_ON(resv->region_cache_count <= 0); | |
232 | ||
233 | resv->region_cache_count--; | |
234 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
235 | VM_BUG_ON(!nrg); | |
236 | list_del(&nrg->link); | |
237 | ||
238 | nrg->from = from; | |
239 | nrg->to = to; | |
240 | ||
241 | return nrg; | |
242 | } | |
243 | ||
075a61d0 MA |
244 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
245 | struct file_region *rg) | |
246 | { | |
247 | #ifdef CONFIG_CGROUP_HUGETLB | |
248 | nrg->reservation_counter = rg->reservation_counter; | |
249 | nrg->css = rg->css; | |
250 | if (rg->css) | |
251 | css_get(rg->css); | |
252 | #endif | |
253 | } | |
254 | ||
255 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
256 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
257 | struct hstate *h, | |
258 | struct resv_map *resv, | |
259 | struct file_region *nrg) | |
260 | { | |
261 | #ifdef CONFIG_CGROUP_HUGETLB | |
262 | if (h_cg) { | |
263 | nrg->reservation_counter = | |
264 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
265 | nrg->css = &h_cg->css; | |
266 | if (!resv->pages_per_hpage) | |
267 | resv->pages_per_hpage = pages_per_huge_page(h); | |
268 | /* pages_per_hpage should be the same for all entries in | |
269 | * a resv_map. | |
270 | */ | |
271 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
272 | } else { | |
273 | nrg->reservation_counter = NULL; | |
274 | nrg->css = NULL; | |
275 | } | |
276 | #endif | |
277 | } | |
278 | ||
a9b3f867 MA |
279 | static bool has_same_uncharge_info(struct file_region *rg, |
280 | struct file_region *org) | |
281 | { | |
282 | #ifdef CONFIG_CGROUP_HUGETLB | |
283 | return rg && org && | |
284 | rg->reservation_counter == org->reservation_counter && | |
285 | rg->css == org->css; | |
286 | ||
287 | #else | |
288 | return true; | |
289 | #endif | |
290 | } | |
291 | ||
292 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
293 | { | |
294 | struct file_region *nrg = NULL, *prg = NULL; | |
295 | ||
296 | prg = list_prev_entry(rg, link); | |
297 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
298 | has_same_uncharge_info(prg, rg)) { | |
299 | prg->to = rg->to; | |
300 | ||
301 | list_del(&rg->link); | |
302 | kfree(rg); | |
303 | ||
304 | coalesce_file_region(resv, prg); | |
305 | return; | |
306 | } | |
307 | ||
308 | nrg = list_next_entry(rg, link); | |
309 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
310 | has_same_uncharge_info(nrg, rg)) { | |
311 | nrg->from = rg->from; | |
312 | ||
313 | list_del(&rg->link); | |
314 | kfree(rg); | |
315 | ||
316 | coalesce_file_region(resv, nrg); | |
317 | return; | |
318 | } | |
319 | } | |
320 | ||
d75c6af9 MA |
321 | /* Must be called with resv->lock held. Calling this with count_only == true |
322 | * will count the number of pages to be added but will not modify the linked | |
0db9d74e MA |
323 | * list. If regions_needed != NULL and count_only == true, then regions_needed |
324 | * will indicate the number of file_regions needed in the cache to carry out to | |
325 | * add the regions for this range. | |
d75c6af9 MA |
326 | */ |
327 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 MA |
328 | struct hugetlb_cgroup *h_cg, |
329 | struct hstate *h, long *regions_needed, | |
330 | bool count_only) | |
d75c6af9 | 331 | { |
0db9d74e | 332 | long add = 0; |
d75c6af9 | 333 | struct list_head *head = &resv->regions; |
0db9d74e | 334 | long last_accounted_offset = f; |
d75c6af9 MA |
335 | struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; |
336 | ||
0db9d74e MA |
337 | if (regions_needed) |
338 | *regions_needed = 0; | |
d75c6af9 | 339 | |
0db9d74e MA |
340 | /* In this loop, we essentially handle an entry for the range |
341 | * [last_accounted_offset, rg->from), at every iteration, with some | |
342 | * bounds checking. | |
343 | */ | |
344 | list_for_each_entry_safe(rg, trg, head, link) { | |
345 | /* Skip irrelevant regions that start before our range. */ | |
346 | if (rg->from < f) { | |
347 | /* If this region ends after the last accounted offset, | |
348 | * then we need to update last_accounted_offset. | |
349 | */ | |
350 | if (rg->to > last_accounted_offset) | |
351 | last_accounted_offset = rg->to; | |
352 | continue; | |
353 | } | |
d75c6af9 | 354 | |
0db9d74e MA |
355 | /* When we find a region that starts beyond our range, we've |
356 | * finished. | |
357 | */ | |
d75c6af9 MA |
358 | if (rg->from > t) |
359 | break; | |
360 | ||
0db9d74e MA |
361 | /* Add an entry for last_accounted_offset -> rg->from, and |
362 | * update last_accounted_offset. | |
363 | */ | |
364 | if (rg->from > last_accounted_offset) { | |
365 | add += rg->from - last_accounted_offset; | |
366 | if (!count_only) { | |
367 | nrg = get_file_region_entry_from_cache( | |
368 | resv, last_accounted_offset, rg->from); | |
075a61d0 MA |
369 | record_hugetlb_cgroup_uncharge_info(h_cg, h, |
370 | resv, nrg); | |
0db9d74e | 371 | list_add(&nrg->link, rg->link.prev); |
a9b3f867 | 372 | coalesce_file_region(resv, nrg); |
0db9d74e MA |
373 | } else if (regions_needed) |
374 | *regions_needed += 1; | |
375 | } | |
376 | ||
377 | last_accounted_offset = rg->to; | |
378 | } | |
379 | ||
380 | /* Handle the case where our range extends beyond | |
381 | * last_accounted_offset. | |
382 | */ | |
383 | if (last_accounted_offset < t) { | |
384 | add += t - last_accounted_offset; | |
385 | if (!count_only) { | |
386 | nrg = get_file_region_entry_from_cache( | |
387 | resv, last_accounted_offset, t); | |
075a61d0 | 388 | record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); |
0db9d74e | 389 | list_add(&nrg->link, rg->link.prev); |
a9b3f867 | 390 | coalesce_file_region(resv, nrg); |
0db9d74e MA |
391 | } else if (regions_needed) |
392 | *regions_needed += 1; | |
393 | } | |
394 | ||
395 | VM_BUG_ON(add < 0); | |
396 | return add; | |
397 | } | |
398 | ||
399 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
400 | */ | |
401 | static int allocate_file_region_entries(struct resv_map *resv, | |
402 | int regions_needed) | |
403 | __must_hold(&resv->lock) | |
404 | { | |
405 | struct list_head allocated_regions; | |
406 | int to_allocate = 0, i = 0; | |
407 | struct file_region *trg = NULL, *rg = NULL; | |
408 | ||
409 | VM_BUG_ON(regions_needed < 0); | |
410 | ||
411 | INIT_LIST_HEAD(&allocated_regions); | |
412 | ||
413 | /* | |
414 | * Check for sufficient descriptors in the cache to accommodate | |
415 | * the number of in progress add operations plus regions_needed. | |
416 | * | |
417 | * This is a while loop because when we drop the lock, some other call | |
418 | * to region_add or region_del may have consumed some region_entries, | |
419 | * so we keep looping here until we finally have enough entries for | |
420 | * (adds_in_progress + regions_needed). | |
421 | */ | |
422 | while (resv->region_cache_count < | |
423 | (resv->adds_in_progress + regions_needed)) { | |
424 | to_allocate = resv->adds_in_progress + regions_needed - | |
425 | resv->region_cache_count; | |
426 | ||
427 | /* At this point, we should have enough entries in the cache | |
428 | * for all the existings adds_in_progress. We should only be | |
429 | * needing to allocate for regions_needed. | |
d75c6af9 | 430 | */ |
0db9d74e MA |
431 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
432 | ||
433 | spin_unlock(&resv->lock); | |
434 | for (i = 0; i < to_allocate; i++) { | |
435 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
436 | if (!trg) | |
437 | goto out_of_memory; | |
438 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 439 | } |
d75c6af9 | 440 | |
0db9d74e MA |
441 | spin_lock(&resv->lock); |
442 | ||
443 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
d75c6af9 | 444 | list_del(&rg->link); |
0db9d74e MA |
445 | list_add(&rg->link, &resv->region_cache); |
446 | resv->region_cache_count++; | |
d75c6af9 MA |
447 | } |
448 | } | |
449 | ||
0db9d74e | 450 | return 0; |
d75c6af9 | 451 | |
0db9d74e MA |
452 | out_of_memory: |
453 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
454 | list_del(&rg->link); | |
455 | kfree(rg); | |
456 | } | |
457 | return -ENOMEM; | |
d75c6af9 MA |
458 | } |
459 | ||
1dd308a7 MK |
460 | /* |
461 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
462 | * map. Regions will be taken from the cache to fill in this range. |
463 | * Sufficient regions should exist in the cache due to the previous | |
464 | * call to region_chg with the same range, but in some cases the cache will not | |
465 | * have sufficient entries due to races with other code doing region_add or | |
466 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 467 | * |
0db9d74e MA |
468 | * regions_needed is the out value provided by a previous call to region_chg. |
469 | * | |
470 | * Return the number of new huge pages added to the map. This number is greater | |
471 | * than or equal to zero. If file_region entries needed to be allocated for | |
472 | * this operation and we were not able to allocate, it ruturns -ENOMEM. | |
473 | * region_add of regions of length 1 never allocate file_regions and cannot | |
474 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
475 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 476 | */ |
0db9d74e | 477 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
478 | long in_regions_needed, struct hstate *h, |
479 | struct hugetlb_cgroup *h_cg) | |
96822904 | 480 | { |
0db9d74e | 481 | long add = 0, actual_regions_needed = 0; |
96822904 | 482 | |
7b24d861 | 483 | spin_lock(&resv->lock); |
0db9d74e MA |
484 | retry: |
485 | ||
486 | /* Count how many regions are actually needed to execute this add. */ | |
075a61d0 MA |
487 | add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed, |
488 | true); | |
96822904 | 489 | |
5e911373 | 490 | /* |
0db9d74e MA |
491 | * Check for sufficient descriptors in the cache to accommodate |
492 | * this add operation. Note that actual_regions_needed may be greater | |
493 | * than in_regions_needed, as the resv_map may have been modified since | |
494 | * the region_chg call. In this case, we need to make sure that we | |
495 | * allocate extra entries, such that we have enough for all the | |
496 | * existing adds_in_progress, plus the excess needed for this | |
497 | * operation. | |
5e911373 | 498 | */ |
0db9d74e MA |
499 | if (actual_regions_needed > in_regions_needed && |
500 | resv->region_cache_count < | |
501 | resv->adds_in_progress + | |
502 | (actual_regions_needed - in_regions_needed)) { | |
503 | /* region_add operation of range 1 should never need to | |
504 | * allocate file_region entries. | |
505 | */ | |
506 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 507 | |
0db9d74e MA |
508 | if (allocate_file_region_entries( |
509 | resv, actual_regions_needed - in_regions_needed)) { | |
510 | return -ENOMEM; | |
511 | } | |
5e911373 | 512 | |
0db9d74e | 513 | goto retry; |
5e911373 MK |
514 | } |
515 | ||
075a61d0 | 516 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false); |
0db9d74e MA |
517 | |
518 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 519 | |
7b24d861 | 520 | spin_unlock(&resv->lock); |
cf3ad20b MK |
521 | VM_BUG_ON(add < 0); |
522 | return add; | |
96822904 AW |
523 | } |
524 | ||
1dd308a7 MK |
525 | /* |
526 | * Examine the existing reserve map and determine how many | |
527 | * huge pages in the specified range [f, t) are NOT currently | |
528 | * represented. This routine is called before a subsequent | |
529 | * call to region_add that will actually modify the reserve | |
530 | * map to add the specified range [f, t). region_chg does | |
531 | * not change the number of huge pages represented by the | |
0db9d74e MA |
532 | * map. A number of new file_region structures is added to the cache as a |
533 | * placeholder, for the subsequent region_add call to use. At least 1 | |
534 | * file_region structure is added. | |
535 | * | |
536 | * out_regions_needed is the number of regions added to the | |
537 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
538 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
539 | * |
540 | * Returns the number of huge pages that need to be added to the existing | |
541 | * reservation map for the range [f, t). This number is greater or equal to | |
542 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
543 | * is needed and can not be allocated. | |
1dd308a7 | 544 | */ |
0db9d74e MA |
545 | static long region_chg(struct resv_map *resv, long f, long t, |
546 | long *out_regions_needed) | |
96822904 | 547 | { |
96822904 AW |
548 | long chg = 0; |
549 | ||
7b24d861 | 550 | spin_lock(&resv->lock); |
5e911373 | 551 | |
0db9d74e | 552 | /* Count how many hugepages in this range are NOT respresented. */ |
075a61d0 MA |
553 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
554 | out_regions_needed, true); | |
5e911373 | 555 | |
0db9d74e MA |
556 | if (*out_regions_needed == 0) |
557 | *out_regions_needed = 1; | |
5e911373 | 558 | |
0db9d74e MA |
559 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
560 | return -ENOMEM; | |
5e911373 | 561 | |
0db9d74e | 562 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 563 | |
7b24d861 | 564 | spin_unlock(&resv->lock); |
96822904 AW |
565 | return chg; |
566 | } | |
567 | ||
5e911373 MK |
568 | /* |
569 | * Abort the in progress add operation. The adds_in_progress field | |
570 | * of the resv_map keeps track of the operations in progress between | |
571 | * calls to region_chg and region_add. Operations are sometimes | |
572 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
573 | * is called to decrement the adds_in_progress counter. regions_needed |
574 | * is the value returned by the region_chg call, it is used to decrement | |
575 | * the adds_in_progress counter. | |
5e911373 MK |
576 | * |
577 | * NOTE: The range arguments [f, t) are not needed or used in this | |
578 | * routine. They are kept to make reading the calling code easier as | |
579 | * arguments will match the associated region_chg call. | |
580 | */ | |
0db9d74e MA |
581 | static void region_abort(struct resv_map *resv, long f, long t, |
582 | long regions_needed) | |
5e911373 MK |
583 | { |
584 | spin_lock(&resv->lock); | |
585 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 586 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
587 | spin_unlock(&resv->lock); |
588 | } | |
589 | ||
1dd308a7 | 590 | /* |
feba16e2 MK |
591 | * Delete the specified range [f, t) from the reserve map. If the |
592 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
593 | * should be deleted. Locate the regions which intersect [f, t) | |
594 | * and either trim, delete or split the existing regions. | |
595 | * | |
596 | * Returns the number of huge pages deleted from the reserve map. | |
597 | * In the normal case, the return value is zero or more. In the | |
598 | * case where a region must be split, a new region descriptor must | |
599 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
600 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
601 | * a region and possibly return -ENOMEM. Callers specifying | |
602 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 603 | */ |
feba16e2 | 604 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 605 | { |
1406ec9b | 606 | struct list_head *head = &resv->regions; |
96822904 | 607 | struct file_region *rg, *trg; |
feba16e2 MK |
608 | struct file_region *nrg = NULL; |
609 | long del = 0; | |
96822904 | 610 | |
feba16e2 | 611 | retry: |
7b24d861 | 612 | spin_lock(&resv->lock); |
feba16e2 | 613 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
614 | /* |
615 | * Skip regions before the range to be deleted. file_region | |
616 | * ranges are normally of the form [from, to). However, there | |
617 | * may be a "placeholder" entry in the map which is of the form | |
618 | * (from, to) with from == to. Check for placeholder entries | |
619 | * at the beginning of the range to be deleted. | |
620 | */ | |
621 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 622 | continue; |
dbe409e4 | 623 | |
feba16e2 | 624 | if (rg->from >= t) |
96822904 | 625 | break; |
96822904 | 626 | |
feba16e2 MK |
627 | if (f > rg->from && t < rg->to) { /* Must split region */ |
628 | /* | |
629 | * Check for an entry in the cache before dropping | |
630 | * lock and attempting allocation. | |
631 | */ | |
632 | if (!nrg && | |
633 | resv->region_cache_count > resv->adds_in_progress) { | |
634 | nrg = list_first_entry(&resv->region_cache, | |
635 | struct file_region, | |
636 | link); | |
637 | list_del(&nrg->link); | |
638 | resv->region_cache_count--; | |
639 | } | |
96822904 | 640 | |
feba16e2 MK |
641 | if (!nrg) { |
642 | spin_unlock(&resv->lock); | |
643 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
644 | if (!nrg) | |
645 | return -ENOMEM; | |
646 | goto retry; | |
647 | } | |
648 | ||
649 | del += t - f; | |
650 | ||
651 | /* New entry for end of split region */ | |
652 | nrg->from = t; | |
653 | nrg->to = rg->to; | |
075a61d0 MA |
654 | |
655 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
656 | ||
feba16e2 MK |
657 | INIT_LIST_HEAD(&nrg->link); |
658 | ||
659 | /* Original entry is trimmed */ | |
660 | rg->to = f; | |
661 | ||
075a61d0 MA |
662 | hugetlb_cgroup_uncharge_file_region( |
663 | resv, rg, nrg->to - nrg->from); | |
664 | ||
feba16e2 MK |
665 | list_add(&nrg->link, &rg->link); |
666 | nrg = NULL; | |
96822904 | 667 | break; |
feba16e2 MK |
668 | } |
669 | ||
670 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
671 | del += rg->to - rg->from; | |
075a61d0 MA |
672 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
673 | rg->to - rg->from); | |
feba16e2 MK |
674 | list_del(&rg->link); |
675 | kfree(rg); | |
676 | continue; | |
677 | } | |
678 | ||
679 | if (f <= rg->from) { /* Trim beginning of region */ | |
680 | del += t - rg->from; | |
681 | rg->from = t; | |
075a61d0 MA |
682 | |
683 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
684 | t - rg->from); | |
feba16e2 MK |
685 | } else { /* Trim end of region */ |
686 | del += rg->to - f; | |
687 | rg->to = f; | |
075a61d0 MA |
688 | |
689 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
690 | rg->to - f); | |
feba16e2 | 691 | } |
96822904 | 692 | } |
7b24d861 | 693 | |
7b24d861 | 694 | spin_unlock(&resv->lock); |
feba16e2 MK |
695 | kfree(nrg); |
696 | return del; | |
96822904 AW |
697 | } |
698 | ||
b5cec28d MK |
699 | /* |
700 | * A rare out of memory error was encountered which prevented removal of | |
701 | * the reserve map region for a page. The huge page itself was free'ed | |
702 | * and removed from the page cache. This routine will adjust the subpool | |
703 | * usage count, and the global reserve count if needed. By incrementing | |
704 | * these counts, the reserve map entry which could not be deleted will | |
705 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
706 | * counts. | |
707 | */ | |
72e2936c | 708 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
709 | { |
710 | struct hugepage_subpool *spool = subpool_inode(inode); | |
711 | long rsv_adjust; | |
712 | ||
713 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 714 | if (rsv_adjust) { |
b5cec28d MK |
715 | struct hstate *h = hstate_inode(inode); |
716 | ||
717 | hugetlb_acct_memory(h, 1); | |
718 | } | |
719 | } | |
720 | ||
1dd308a7 MK |
721 | /* |
722 | * Count and return the number of huge pages in the reserve map | |
723 | * that intersect with the range [f, t). | |
724 | */ | |
1406ec9b | 725 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 726 | { |
1406ec9b | 727 | struct list_head *head = &resv->regions; |
84afd99b AW |
728 | struct file_region *rg; |
729 | long chg = 0; | |
730 | ||
7b24d861 | 731 | spin_lock(&resv->lock); |
84afd99b AW |
732 | /* Locate each segment we overlap with, and count that overlap. */ |
733 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
734 | long seg_from; |
735 | long seg_to; | |
84afd99b AW |
736 | |
737 | if (rg->to <= f) | |
738 | continue; | |
739 | if (rg->from >= t) | |
740 | break; | |
741 | ||
742 | seg_from = max(rg->from, f); | |
743 | seg_to = min(rg->to, t); | |
744 | ||
745 | chg += seg_to - seg_from; | |
746 | } | |
7b24d861 | 747 | spin_unlock(&resv->lock); |
84afd99b AW |
748 | |
749 | return chg; | |
750 | } | |
751 | ||
e7c4b0bf AW |
752 | /* |
753 | * Convert the address within this vma to the page offset within | |
754 | * the mapping, in pagecache page units; huge pages here. | |
755 | */ | |
a5516438 AK |
756 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
757 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 758 | { |
a5516438 AK |
759 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
760 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
761 | } |
762 | ||
0fe6e20b NH |
763 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
764 | unsigned long address) | |
765 | { | |
766 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
767 | } | |
dee41079 | 768 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 769 | |
08fba699 MG |
770 | /* |
771 | * Return the size of the pages allocated when backing a VMA. In the majority | |
772 | * cases this will be same size as used by the page table entries. | |
773 | */ | |
774 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
775 | { | |
05ea8860 DW |
776 | if (vma->vm_ops && vma->vm_ops->pagesize) |
777 | return vma->vm_ops->pagesize(vma); | |
778 | return PAGE_SIZE; | |
08fba699 | 779 | } |
f340ca0f | 780 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 781 | |
3340289d MG |
782 | /* |
783 | * Return the page size being used by the MMU to back a VMA. In the majority | |
784 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
785 | * architectures where it differs, an architecture-specific 'strong' |
786 | * version of this symbol is required. | |
3340289d | 787 | */ |
09135cc5 | 788 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
789 | { |
790 | return vma_kernel_pagesize(vma); | |
791 | } | |
3340289d | 792 | |
84afd99b AW |
793 | /* |
794 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
795 | * bits of the reservation map pointer, which are always clear due to | |
796 | * alignment. | |
797 | */ | |
798 | #define HPAGE_RESV_OWNER (1UL << 0) | |
799 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 800 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 801 | |
a1e78772 MG |
802 | /* |
803 | * These helpers are used to track how many pages are reserved for | |
804 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
805 | * is guaranteed to have their future faults succeed. | |
806 | * | |
807 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
808 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
809 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
810 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
811 | * |
812 | * The private mapping reservation is represented in a subtly different | |
813 | * manner to a shared mapping. A shared mapping has a region map associated | |
814 | * with the underlying file, this region map represents the backing file | |
815 | * pages which have ever had a reservation assigned which this persists even | |
816 | * after the page is instantiated. A private mapping has a region map | |
817 | * associated with the original mmap which is attached to all VMAs which | |
818 | * reference it, this region map represents those offsets which have consumed | |
819 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 820 | */ |
e7c4b0bf AW |
821 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
822 | { | |
823 | return (unsigned long)vma->vm_private_data; | |
824 | } | |
825 | ||
826 | static void set_vma_private_data(struct vm_area_struct *vma, | |
827 | unsigned long value) | |
828 | { | |
829 | vma->vm_private_data = (void *)value; | |
830 | } | |
831 | ||
e9fe92ae MA |
832 | static void |
833 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
834 | struct hugetlb_cgroup *h_cg, | |
835 | struct hstate *h) | |
836 | { | |
837 | #ifdef CONFIG_CGROUP_HUGETLB | |
838 | if (!h_cg || !h) { | |
839 | resv_map->reservation_counter = NULL; | |
840 | resv_map->pages_per_hpage = 0; | |
841 | resv_map->css = NULL; | |
842 | } else { | |
843 | resv_map->reservation_counter = | |
844 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
845 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
846 | resv_map->css = &h_cg->css; | |
847 | } | |
848 | #endif | |
849 | } | |
850 | ||
9119a41e | 851 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
852 | { |
853 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
854 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
855 | ||
856 | if (!resv_map || !rg) { | |
857 | kfree(resv_map); | |
858 | kfree(rg); | |
84afd99b | 859 | return NULL; |
5e911373 | 860 | } |
84afd99b AW |
861 | |
862 | kref_init(&resv_map->refs); | |
7b24d861 | 863 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
864 | INIT_LIST_HEAD(&resv_map->regions); |
865 | ||
5e911373 | 866 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
867 | /* |
868 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
869 | * fields don't do cgroup accounting. On private mappings, these will be | |
870 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
871 | * reservations are to be un-charged from here. | |
872 | */ | |
873 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
874 | |
875 | INIT_LIST_HEAD(&resv_map->region_cache); | |
876 | list_add(&rg->link, &resv_map->region_cache); | |
877 | resv_map->region_cache_count = 1; | |
878 | ||
84afd99b AW |
879 | return resv_map; |
880 | } | |
881 | ||
9119a41e | 882 | void resv_map_release(struct kref *ref) |
84afd99b AW |
883 | { |
884 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
885 | struct list_head *head = &resv_map->region_cache; |
886 | struct file_region *rg, *trg; | |
84afd99b AW |
887 | |
888 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 889 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
890 | |
891 | /* ... and any entries left in the cache */ | |
892 | list_for_each_entry_safe(rg, trg, head, link) { | |
893 | list_del(&rg->link); | |
894 | kfree(rg); | |
895 | } | |
896 | ||
897 | VM_BUG_ON(resv_map->adds_in_progress); | |
898 | ||
84afd99b AW |
899 | kfree(resv_map); |
900 | } | |
901 | ||
4e35f483 JK |
902 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
903 | { | |
f27a5136 MK |
904 | /* |
905 | * At inode evict time, i_mapping may not point to the original | |
906 | * address space within the inode. This original address space | |
907 | * contains the pointer to the resv_map. So, always use the | |
908 | * address space embedded within the inode. | |
909 | * The VERY common case is inode->mapping == &inode->i_data but, | |
910 | * this may not be true for device special inodes. | |
911 | */ | |
912 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
913 | } |
914 | ||
84afd99b | 915 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 916 | { |
81d1b09c | 917 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
918 | if (vma->vm_flags & VM_MAYSHARE) { |
919 | struct address_space *mapping = vma->vm_file->f_mapping; | |
920 | struct inode *inode = mapping->host; | |
921 | ||
922 | return inode_resv_map(inode); | |
923 | ||
924 | } else { | |
84afd99b AW |
925 | return (struct resv_map *)(get_vma_private_data(vma) & |
926 | ~HPAGE_RESV_MASK); | |
4e35f483 | 927 | } |
a1e78772 MG |
928 | } |
929 | ||
84afd99b | 930 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 931 | { |
81d1b09c SL |
932 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
933 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 934 | |
84afd99b AW |
935 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
936 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
937 | } |
938 | ||
939 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
940 | { | |
81d1b09c SL |
941 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
942 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
943 | |
944 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
945 | } |
946 | ||
947 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
948 | { | |
81d1b09c | 949 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
950 | |
951 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
952 | } |
953 | ||
04f2cbe3 | 954 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
955 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
956 | { | |
81d1b09c | 957 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 958 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
959 | vma->vm_private_data = (void *)0; |
960 | } | |
961 | ||
962 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 963 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 964 | { |
af0ed73e JK |
965 | if (vma->vm_flags & VM_NORESERVE) { |
966 | /* | |
967 | * This address is already reserved by other process(chg == 0), | |
968 | * so, we should decrement reserved count. Without decrementing, | |
969 | * reserve count remains after releasing inode, because this | |
970 | * allocated page will go into page cache and is regarded as | |
971 | * coming from reserved pool in releasing step. Currently, we | |
972 | * don't have any other solution to deal with this situation | |
973 | * properly, so add work-around here. | |
974 | */ | |
975 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 976 | return true; |
af0ed73e | 977 | else |
559ec2f8 | 978 | return false; |
af0ed73e | 979 | } |
a63884e9 JK |
980 | |
981 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
982 | if (vma->vm_flags & VM_MAYSHARE) { |
983 | /* | |
984 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
985 | * be a region map for all pages. The only situation where | |
986 | * there is no region map is if a hole was punched via | |
987 | * fallocate. In this case, there really are no reverves to | |
988 | * use. This situation is indicated if chg != 0. | |
989 | */ | |
990 | if (chg) | |
991 | return false; | |
992 | else | |
993 | return true; | |
994 | } | |
a63884e9 JK |
995 | |
996 | /* | |
997 | * Only the process that called mmap() has reserves for | |
998 | * private mappings. | |
999 | */ | |
67961f9d MK |
1000 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1001 | /* | |
1002 | * Like the shared case above, a hole punch or truncate | |
1003 | * could have been performed on the private mapping. | |
1004 | * Examine the value of chg to determine if reserves | |
1005 | * actually exist or were previously consumed. | |
1006 | * Very Subtle - The value of chg comes from a previous | |
1007 | * call to vma_needs_reserves(). The reserve map for | |
1008 | * private mappings has different (opposite) semantics | |
1009 | * than that of shared mappings. vma_needs_reserves() | |
1010 | * has already taken this difference in semantics into | |
1011 | * account. Therefore, the meaning of chg is the same | |
1012 | * as in the shared case above. Code could easily be | |
1013 | * combined, but keeping it separate draws attention to | |
1014 | * subtle differences. | |
1015 | */ | |
1016 | if (chg) | |
1017 | return false; | |
1018 | else | |
1019 | return true; | |
1020 | } | |
a63884e9 | 1021 | |
559ec2f8 | 1022 | return false; |
a1e78772 MG |
1023 | } |
1024 | ||
a5516438 | 1025 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1026 | { |
1027 | int nid = page_to_nid(page); | |
0edaecfa | 1028 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1029 | h->free_huge_pages++; |
1030 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
1031 | } |
1032 | ||
94310cbc | 1033 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1034 | { |
1035 | struct page *page; | |
1036 | ||
c8721bbb | 1037 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 1038 | if (!PageHWPoison(page)) |
c8721bbb NH |
1039 | break; |
1040 | /* | |
1041 | * if 'non-isolated free hugepage' not found on the list, | |
1042 | * the allocation fails. | |
1043 | */ | |
1044 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 1045 | return NULL; |
0edaecfa | 1046 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 1047 | set_page_refcounted(page); |
bf50bab2 NH |
1048 | h->free_huge_pages--; |
1049 | h->free_huge_pages_node[nid]--; | |
1050 | return page; | |
1051 | } | |
1052 | ||
3e59fcb0 MH |
1053 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1054 | nodemask_t *nmask) | |
94310cbc | 1055 | { |
3e59fcb0 MH |
1056 | unsigned int cpuset_mems_cookie; |
1057 | struct zonelist *zonelist; | |
1058 | struct zone *zone; | |
1059 | struct zoneref *z; | |
98fa15f3 | 1060 | int node = NUMA_NO_NODE; |
94310cbc | 1061 | |
3e59fcb0 MH |
1062 | zonelist = node_zonelist(nid, gfp_mask); |
1063 | ||
1064 | retry_cpuset: | |
1065 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1066 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1067 | struct page *page; | |
1068 | ||
1069 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1070 | continue; | |
1071 | /* | |
1072 | * no need to ask again on the same node. Pool is node rather than | |
1073 | * zone aware | |
1074 | */ | |
1075 | if (zone_to_nid(zone) == node) | |
1076 | continue; | |
1077 | node = zone_to_nid(zone); | |
94310cbc | 1078 | |
94310cbc AK |
1079 | page = dequeue_huge_page_node_exact(h, node); |
1080 | if (page) | |
1081 | return page; | |
1082 | } | |
3e59fcb0 MH |
1083 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1084 | goto retry_cpuset; | |
1085 | ||
94310cbc AK |
1086 | return NULL; |
1087 | } | |
1088 | ||
86cdb465 NH |
1089 | /* Movability of hugepages depends on migration support. */ |
1090 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
1091 | { | |
7ed2c31d | 1092 | if (hugepage_movable_supported(h)) |
86cdb465 NH |
1093 | return GFP_HIGHUSER_MOVABLE; |
1094 | else | |
1095 | return GFP_HIGHUSER; | |
1096 | } | |
1097 | ||
a5516438 AK |
1098 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1099 | struct vm_area_struct *vma, | |
af0ed73e JK |
1100 | unsigned long address, int avoid_reserve, |
1101 | long chg) | |
1da177e4 | 1102 | { |
3e59fcb0 | 1103 | struct page *page; |
480eccf9 | 1104 | struct mempolicy *mpol; |
04ec6264 | 1105 | gfp_t gfp_mask; |
3e59fcb0 | 1106 | nodemask_t *nodemask; |
04ec6264 | 1107 | int nid; |
1da177e4 | 1108 | |
a1e78772 MG |
1109 | /* |
1110 | * A child process with MAP_PRIVATE mappings created by their parent | |
1111 | * have no page reserves. This check ensures that reservations are | |
1112 | * not "stolen". The child may still get SIGKILLed | |
1113 | */ | |
af0ed73e | 1114 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1115 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1116 | goto err; |
a1e78772 | 1117 | |
04f2cbe3 | 1118 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1119 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1120 | goto err; |
04f2cbe3 | 1121 | |
04ec6264 VB |
1122 | gfp_mask = htlb_alloc_mask(h); |
1123 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
1124 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
1125 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
1126 | SetPagePrivate(page); | |
1127 | h->resv_huge_pages--; | |
1da177e4 | 1128 | } |
cc9a6c87 | 1129 | |
52cd3b07 | 1130 | mpol_cond_put(mpol); |
1da177e4 | 1131 | return page; |
cc9a6c87 MG |
1132 | |
1133 | err: | |
cc9a6c87 | 1134 | return NULL; |
1da177e4 LT |
1135 | } |
1136 | ||
1cac6f2c LC |
1137 | /* |
1138 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1139 | * We may have allocated or freed a huge page based on a different | |
1140 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1141 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1142 | * node for alloc or free. | |
1143 | */ | |
1144 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1145 | { | |
0edaf86c | 1146 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1147 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1148 | ||
1149 | return nid; | |
1150 | } | |
1151 | ||
1152 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1153 | { | |
1154 | if (!node_isset(nid, *nodes_allowed)) | |
1155 | nid = next_node_allowed(nid, nodes_allowed); | |
1156 | return nid; | |
1157 | } | |
1158 | ||
1159 | /* | |
1160 | * returns the previously saved node ["this node"] from which to | |
1161 | * allocate a persistent huge page for the pool and advance the | |
1162 | * next node from which to allocate, handling wrap at end of node | |
1163 | * mask. | |
1164 | */ | |
1165 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1166 | nodemask_t *nodes_allowed) | |
1167 | { | |
1168 | int nid; | |
1169 | ||
1170 | VM_BUG_ON(!nodes_allowed); | |
1171 | ||
1172 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1173 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1174 | ||
1175 | return nid; | |
1176 | } | |
1177 | ||
1178 | /* | |
1179 | * helper for free_pool_huge_page() - return the previously saved | |
1180 | * node ["this node"] from which to free a huge page. Advance the | |
1181 | * next node id whether or not we find a free huge page to free so | |
1182 | * that the next attempt to free addresses the next node. | |
1183 | */ | |
1184 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1185 | { | |
1186 | int nid; | |
1187 | ||
1188 | VM_BUG_ON(!nodes_allowed); | |
1189 | ||
1190 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1191 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1192 | ||
1193 | return nid; | |
1194 | } | |
1195 | ||
1196 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1197 | for (nr_nodes = nodes_weight(*mask); \ | |
1198 | nr_nodes > 0 && \ | |
1199 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1200 | nr_nodes--) | |
1201 | ||
1202 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1203 | for (nr_nodes = nodes_weight(*mask); \ | |
1204 | nr_nodes > 0 && \ | |
1205 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1206 | nr_nodes--) | |
1207 | ||
e1073d1e | 1208 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1209 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1210 | unsigned int order) |
944d9fec LC |
1211 | { |
1212 | int i; | |
1213 | int nr_pages = 1 << order; | |
1214 | struct page *p = page + 1; | |
1215 | ||
c8cc708a | 1216 | atomic_set(compound_mapcount_ptr(page), 0); |
47e29d32 JH |
1217 | if (hpage_pincount_available(page)) |
1218 | atomic_set(compound_pincount_ptr(page), 0); | |
1219 | ||
944d9fec | 1220 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1221 | clear_compound_head(p); |
944d9fec | 1222 | set_page_refcounted(p); |
944d9fec LC |
1223 | } |
1224 | ||
1225 | set_compound_order(page, 0); | |
1226 | __ClearPageHead(page); | |
1227 | } | |
1228 | ||
d00181b9 | 1229 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec LC |
1230 | { |
1231 | free_contig_range(page_to_pfn(page), 1 << order); | |
1232 | } | |
1233 | ||
4eb0716e | 1234 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1235 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1236 | int nid, nodemask_t *nodemask) | |
944d9fec | 1237 | { |
5e27a2df | 1238 | unsigned long nr_pages = 1UL << huge_page_order(h); |
944d9fec | 1239 | |
5e27a2df | 1240 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1241 | } |
1242 | ||
1243 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1244 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1245 | #else /* !CONFIG_CONTIG_ALLOC */ |
1246 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1247 | int nid, nodemask_t *nodemask) | |
1248 | { | |
1249 | return NULL; | |
1250 | } | |
1251 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1252 | |
e1073d1e | 1253 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1254 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1255 | int nid, nodemask_t *nodemask) |
1256 | { | |
1257 | return NULL; | |
1258 | } | |
d00181b9 | 1259 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1260 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1261 | unsigned int order) { } |
944d9fec LC |
1262 | #endif |
1263 | ||
a5516438 | 1264 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1265 | { |
1266 | int i; | |
a5516438 | 1267 | |
4eb0716e | 1268 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1269 | return; |
18229df5 | 1270 | |
a5516438 AK |
1271 | h->nr_huge_pages--; |
1272 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1273 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1274 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1275 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1276 | 1 << PG_active | 1 << PG_private | |
1277 | 1 << PG_writeback); | |
6af2acb6 | 1278 | } |
309381fe | 1279 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
1adc4d41 | 1280 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); |
f1e61557 | 1281 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1282 | set_page_refcounted(page); |
944d9fec LC |
1283 | if (hstate_is_gigantic(h)) { |
1284 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1285 | free_gigantic_page(page, huge_page_order(h)); | |
1286 | } else { | |
944d9fec LC |
1287 | __free_pages(page, huge_page_order(h)); |
1288 | } | |
6af2acb6 AL |
1289 | } |
1290 | ||
e5ff2159 AK |
1291 | struct hstate *size_to_hstate(unsigned long size) |
1292 | { | |
1293 | struct hstate *h; | |
1294 | ||
1295 | for_each_hstate(h) { | |
1296 | if (huge_page_size(h) == size) | |
1297 | return h; | |
1298 | } | |
1299 | return NULL; | |
1300 | } | |
1301 | ||
bcc54222 NH |
1302 | /* |
1303 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1304 | * to hstate->hugepage_activelist.) | |
1305 | * | |
1306 | * This function can be called for tail pages, but never returns true for them. | |
1307 | */ | |
1308 | bool page_huge_active(struct page *page) | |
1309 | { | |
1310 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1311 | return PageHead(page) && PagePrivate(&page[1]); | |
1312 | } | |
1313 | ||
1314 | /* never called for tail page */ | |
1315 | static void set_page_huge_active(struct page *page) | |
1316 | { | |
1317 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1318 | SetPagePrivate(&page[1]); | |
1319 | } | |
1320 | ||
1321 | static void clear_page_huge_active(struct page *page) | |
1322 | { | |
1323 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1324 | ClearPagePrivate(&page[1]); | |
1325 | } | |
1326 | ||
ab5ac90a MH |
1327 | /* |
1328 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1329 | * code | |
1330 | */ | |
1331 | static inline bool PageHugeTemporary(struct page *page) | |
1332 | { | |
1333 | if (!PageHuge(page)) | |
1334 | return false; | |
1335 | ||
1336 | return (unsigned long)page[2].mapping == -1U; | |
1337 | } | |
1338 | ||
1339 | static inline void SetPageHugeTemporary(struct page *page) | |
1340 | { | |
1341 | page[2].mapping = (void *)-1U; | |
1342 | } | |
1343 | ||
1344 | static inline void ClearPageHugeTemporary(struct page *page) | |
1345 | { | |
1346 | page[2].mapping = NULL; | |
1347 | } | |
1348 | ||
c77c0a8a | 1349 | static void __free_huge_page(struct page *page) |
27a85ef1 | 1350 | { |
a5516438 AK |
1351 | /* |
1352 | * Can't pass hstate in here because it is called from the | |
1353 | * compound page destructor. | |
1354 | */ | |
e5ff2159 | 1355 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1356 | int nid = page_to_nid(page); |
90481622 DG |
1357 | struct hugepage_subpool *spool = |
1358 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1359 | bool restore_reserve; |
27a85ef1 | 1360 | |
b4330afb MK |
1361 | VM_BUG_ON_PAGE(page_count(page), page); |
1362 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc YW |
1363 | |
1364 | set_page_private(page, 0); | |
1365 | page->mapping = NULL; | |
07443a85 | 1366 | restore_reserve = PagePrivate(page); |
16c794b4 | 1367 | ClearPagePrivate(page); |
27a85ef1 | 1368 | |
1c5ecae3 | 1369 | /* |
0919e1b6 MK |
1370 | * If PagePrivate() was set on page, page allocation consumed a |
1371 | * reservation. If the page was associated with a subpool, there | |
1372 | * would have been a page reserved in the subpool before allocation | |
1373 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1374 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1375 | * remove the reserved page from the subpool. | |
1c5ecae3 | 1376 | */ |
0919e1b6 MK |
1377 | if (!restore_reserve) { |
1378 | /* | |
1379 | * A return code of zero implies that the subpool will be | |
1380 | * under its minimum size if the reservation is not restored | |
1381 | * after page is free. Therefore, force restore_reserve | |
1382 | * operation. | |
1383 | */ | |
1384 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1385 | restore_reserve = true; | |
1386 | } | |
1c5ecae3 | 1387 | |
27a85ef1 | 1388 | spin_lock(&hugetlb_lock); |
bcc54222 | 1389 | clear_page_huge_active(page); |
6d76dcf4 AK |
1390 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1391 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1392 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1393 | pages_per_huge_page(h), page); | |
07443a85 JK |
1394 | if (restore_reserve) |
1395 | h->resv_huge_pages++; | |
1396 | ||
ab5ac90a MH |
1397 | if (PageHugeTemporary(page)) { |
1398 | list_del(&page->lru); | |
1399 | ClearPageHugeTemporary(page); | |
1400 | update_and_free_page(h, page); | |
1401 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1402 | /* remove the page from active list */ |
1403 | list_del(&page->lru); | |
a5516438 AK |
1404 | update_and_free_page(h, page); |
1405 | h->surplus_huge_pages--; | |
1406 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1407 | } else { |
5d3a551c | 1408 | arch_clear_hugepage_flags(page); |
a5516438 | 1409 | enqueue_huge_page(h, page); |
7893d1d5 | 1410 | } |
27a85ef1 DG |
1411 | spin_unlock(&hugetlb_lock); |
1412 | } | |
1413 | ||
c77c0a8a WL |
1414 | /* |
1415 | * As free_huge_page() can be called from a non-task context, we have | |
1416 | * to defer the actual freeing in a workqueue to prevent potential | |
1417 | * hugetlb_lock deadlock. | |
1418 | * | |
1419 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1420 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1421 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1422 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1423 | */ | |
1424 | static LLIST_HEAD(hpage_freelist); | |
1425 | ||
1426 | static void free_hpage_workfn(struct work_struct *work) | |
1427 | { | |
1428 | struct llist_node *node; | |
1429 | struct page *page; | |
1430 | ||
1431 | node = llist_del_all(&hpage_freelist); | |
1432 | ||
1433 | while (node) { | |
1434 | page = container_of((struct address_space **)node, | |
1435 | struct page, mapping); | |
1436 | node = node->next; | |
1437 | __free_huge_page(page); | |
1438 | } | |
1439 | } | |
1440 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1441 | ||
1442 | void free_huge_page(struct page *page) | |
1443 | { | |
1444 | /* | |
1445 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1446 | */ | |
1447 | if (!in_task()) { | |
1448 | /* | |
1449 | * Only call schedule_work() if hpage_freelist is previously | |
1450 | * empty. Otherwise, schedule_work() had been called but the | |
1451 | * workfn hasn't retrieved the list yet. | |
1452 | */ | |
1453 | if (llist_add((struct llist_node *)&page->mapping, | |
1454 | &hpage_freelist)) | |
1455 | schedule_work(&free_hpage_work); | |
1456 | return; | |
1457 | } | |
1458 | ||
1459 | __free_huge_page(page); | |
1460 | } | |
1461 | ||
a5516438 | 1462 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1463 | { |
0edaecfa | 1464 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1465 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1466 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1467 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1468 | set_hugetlb_cgroup_rsvd(page, NULL); |
a5516438 AK |
1469 | h->nr_huge_pages++; |
1470 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 | 1471 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1472 | } |
1473 | ||
d00181b9 | 1474 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1475 | { |
1476 | int i; | |
1477 | int nr_pages = 1 << order; | |
1478 | struct page *p = page + 1; | |
1479 | ||
1480 | /* we rely on prep_new_huge_page to set the destructor */ | |
1481 | set_compound_order(page, order); | |
ef5a22be | 1482 | __ClearPageReserved(page); |
de09d31d | 1483 | __SetPageHead(page); |
20a0307c | 1484 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1485 | /* |
1486 | * For gigantic hugepages allocated through bootmem at | |
1487 | * boot, it's safer to be consistent with the not-gigantic | |
1488 | * hugepages and clear the PG_reserved bit from all tail pages | |
1489 | * too. Otherwse drivers using get_user_pages() to access tail | |
1490 | * pages may get the reference counting wrong if they see | |
1491 | * PG_reserved set on a tail page (despite the head page not | |
1492 | * having PG_reserved set). Enforcing this consistency between | |
1493 | * head and tail pages allows drivers to optimize away a check | |
1494 | * on the head page when they need know if put_page() is needed | |
1495 | * after get_user_pages(). | |
1496 | */ | |
1497 | __ClearPageReserved(p); | |
58a84aa9 | 1498 | set_page_count(p, 0); |
1d798ca3 | 1499 | set_compound_head(p, page); |
20a0307c | 1500 | } |
b4330afb | 1501 | atomic_set(compound_mapcount_ptr(page), -1); |
47e29d32 JH |
1502 | |
1503 | if (hpage_pincount_available(page)) | |
1504 | atomic_set(compound_pincount_ptr(page), 0); | |
20a0307c WF |
1505 | } |
1506 | ||
7795912c AM |
1507 | /* |
1508 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1509 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1510 | * details. | |
1511 | */ | |
20a0307c WF |
1512 | int PageHuge(struct page *page) |
1513 | { | |
20a0307c WF |
1514 | if (!PageCompound(page)) |
1515 | return 0; | |
1516 | ||
1517 | page = compound_head(page); | |
f1e61557 | 1518 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1519 | } |
43131e14 NH |
1520 | EXPORT_SYMBOL_GPL(PageHuge); |
1521 | ||
27c73ae7 AA |
1522 | /* |
1523 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1524 | * normal or transparent huge pages. | |
1525 | */ | |
1526 | int PageHeadHuge(struct page *page_head) | |
1527 | { | |
27c73ae7 AA |
1528 | if (!PageHead(page_head)) |
1529 | return 0; | |
1530 | ||
d4af73e3 | 1531 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1532 | } |
27c73ae7 | 1533 | |
c0d0381a MK |
1534 | /* |
1535 | * Find address_space associated with hugetlbfs page. | |
1536 | * Upon entry page is locked and page 'was' mapped although mapped state | |
1537 | * could change. If necessary, use anon_vma to find vma and associated | |
1538 | * address space. The returned mapping may be stale, but it can not be | |
1539 | * invalid as page lock (which is held) is required to destroy mapping. | |
1540 | */ | |
1541 | static struct address_space *_get_hugetlb_page_mapping(struct page *hpage) | |
1542 | { | |
1543 | struct anon_vma *anon_vma; | |
1544 | pgoff_t pgoff_start, pgoff_end; | |
1545 | struct anon_vma_chain *avc; | |
1546 | struct address_space *mapping = page_mapping(hpage); | |
1547 | ||
1548 | /* Simple file based mapping */ | |
1549 | if (mapping) | |
1550 | return mapping; | |
1551 | ||
1552 | /* | |
1553 | * Even anonymous hugetlbfs mappings are associated with an | |
1554 | * underlying hugetlbfs file (see hugetlb_file_setup in mmap | |
1555 | * code). Find a vma associated with the anonymous vma, and | |
1556 | * use the file pointer to get address_space. | |
1557 | */ | |
1558 | anon_vma = page_lock_anon_vma_read(hpage); | |
1559 | if (!anon_vma) | |
1560 | return mapping; /* NULL */ | |
1561 | ||
1562 | /* Use first found vma */ | |
1563 | pgoff_start = page_to_pgoff(hpage); | |
1564 | pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1; | |
1565 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, | |
1566 | pgoff_start, pgoff_end) { | |
1567 | struct vm_area_struct *vma = avc->vma; | |
1568 | ||
1569 | mapping = vma->vm_file->f_mapping; | |
1570 | break; | |
1571 | } | |
1572 | ||
1573 | anon_vma_unlock_read(anon_vma); | |
1574 | return mapping; | |
1575 | } | |
1576 | ||
1577 | /* | |
1578 | * Find and lock address space (mapping) in write mode. | |
1579 | * | |
1580 | * Upon entry, the page is locked which allows us to find the mapping | |
1581 | * even in the case of an anon page. However, locking order dictates | |
1582 | * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs | |
1583 | * specific. So, we first try to lock the sema while still holding the | |
1584 | * page lock. If this works, great! If not, then we need to drop the | |
1585 | * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of | |
1586 | * course, need to revalidate state along the way. | |
1587 | */ | |
1588 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1589 | { | |
1590 | struct address_space *mapping, *mapping2; | |
1591 | ||
1592 | mapping = _get_hugetlb_page_mapping(hpage); | |
1593 | retry: | |
1594 | if (!mapping) | |
1595 | return mapping; | |
1596 | ||
1597 | /* | |
1598 | * If no contention, take lock and return | |
1599 | */ | |
1600 | if (i_mmap_trylock_write(mapping)) | |
1601 | return mapping; | |
1602 | ||
1603 | /* | |
1604 | * Must drop page lock and wait on mapping sema. | |
1605 | * Note: Once page lock is dropped, mapping could become invalid. | |
1606 | * As a hack, increase map count until we lock page again. | |
1607 | */ | |
1608 | atomic_inc(&hpage->_mapcount); | |
1609 | unlock_page(hpage); | |
1610 | i_mmap_lock_write(mapping); | |
1611 | lock_page(hpage); | |
1612 | atomic_add_negative(-1, &hpage->_mapcount); | |
1613 | ||
1614 | /* verify page is still mapped */ | |
1615 | if (!page_mapped(hpage)) { | |
1616 | i_mmap_unlock_write(mapping); | |
1617 | return NULL; | |
1618 | } | |
1619 | ||
1620 | /* | |
1621 | * Get address space again and verify it is the same one | |
1622 | * we locked. If not, drop lock and retry. | |
1623 | */ | |
1624 | mapping2 = _get_hugetlb_page_mapping(hpage); | |
1625 | if (mapping2 != mapping) { | |
1626 | i_mmap_unlock_write(mapping); | |
1627 | mapping = mapping2; | |
1628 | goto retry; | |
1629 | } | |
1630 | ||
1631 | return mapping; | |
1632 | } | |
1633 | ||
13d60f4b ZY |
1634 | pgoff_t __basepage_index(struct page *page) |
1635 | { | |
1636 | struct page *page_head = compound_head(page); | |
1637 | pgoff_t index = page_index(page_head); | |
1638 | unsigned long compound_idx; | |
1639 | ||
1640 | if (!PageHuge(page_head)) | |
1641 | return page_index(page); | |
1642 | ||
1643 | if (compound_order(page_head) >= MAX_ORDER) | |
1644 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1645 | else | |
1646 | compound_idx = page - page_head; | |
1647 | ||
1648 | return (index << compound_order(page_head)) + compound_idx; | |
1649 | } | |
1650 | ||
0c397dae | 1651 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1652 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1653 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1654 | { |
af0fb9df | 1655 | int order = huge_page_order(h); |
1da177e4 | 1656 | struct page *page; |
f60858f9 | 1657 | bool alloc_try_hard = true; |
f96efd58 | 1658 | |
f60858f9 MK |
1659 | /* |
1660 | * By default we always try hard to allocate the page with | |
1661 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1662 | * a loop (to adjust global huge page counts) and previous allocation | |
1663 | * failed, do not continue to try hard on the same node. Use the | |
1664 | * node_alloc_noretry bitmap to manage this state information. | |
1665 | */ | |
1666 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1667 | alloc_try_hard = false; | |
1668 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1669 | if (alloc_try_hard) | |
1670 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1671 | if (nid == NUMA_NO_NODE) |
1672 | nid = numa_mem_id(); | |
1673 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1674 | if (page) | |
1675 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1676 | else | |
1677 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1678 | |
f60858f9 MK |
1679 | /* |
1680 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1681 | * indicates an overall state change. Clear bit so that we resume | |
1682 | * normal 'try hard' allocations. | |
1683 | */ | |
1684 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1685 | node_clear(nid, *node_alloc_noretry); | |
1686 | ||
1687 | /* | |
1688 | * If we tried hard to get a page but failed, set bit so that | |
1689 | * subsequent attempts will not try as hard until there is an | |
1690 | * overall state change. | |
1691 | */ | |
1692 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1693 | node_set(nid, *node_alloc_noretry); | |
1694 | ||
63b4613c NA |
1695 | return page; |
1696 | } | |
1697 | ||
0c397dae MH |
1698 | /* |
1699 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1700 | * should use this function to get new hugetlb pages | |
1701 | */ | |
1702 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1703 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1704 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1705 | { |
1706 | struct page *page; | |
1707 | ||
1708 | if (hstate_is_gigantic(h)) | |
1709 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1710 | else | |
1711 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1712 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1713 | if (!page) |
1714 | return NULL; | |
1715 | ||
1716 | if (hstate_is_gigantic(h)) | |
1717 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1718 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1719 | ||
1720 | return page; | |
1721 | } | |
1722 | ||
af0fb9df MH |
1723 | /* |
1724 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1725 | * manner. | |
1726 | */ | |
f60858f9 MK |
1727 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1728 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1729 | { |
1730 | struct page *page; | |
1731 | int nr_nodes, node; | |
af0fb9df | 1732 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1733 | |
1734 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1735 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1736 | node_alloc_noretry); | |
af0fb9df | 1737 | if (page) |
b2261026 | 1738 | break; |
b2261026 JK |
1739 | } |
1740 | ||
af0fb9df MH |
1741 | if (!page) |
1742 | return 0; | |
b2261026 | 1743 | |
af0fb9df MH |
1744 | put_page(page); /* free it into the hugepage allocator */ |
1745 | ||
1746 | return 1; | |
b2261026 JK |
1747 | } |
1748 | ||
e8c5c824 LS |
1749 | /* |
1750 | * Free huge page from pool from next node to free. | |
1751 | * Attempt to keep persistent huge pages more or less | |
1752 | * balanced over allowed nodes. | |
1753 | * Called with hugetlb_lock locked. | |
1754 | */ | |
6ae11b27 LS |
1755 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1756 | bool acct_surplus) | |
e8c5c824 | 1757 | { |
b2261026 | 1758 | int nr_nodes, node; |
e8c5c824 LS |
1759 | int ret = 0; |
1760 | ||
b2261026 | 1761 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1762 | /* |
1763 | * If we're returning unused surplus pages, only examine | |
1764 | * nodes with surplus pages. | |
1765 | */ | |
b2261026 JK |
1766 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1767 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1768 | struct page *page = |
b2261026 | 1769 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1770 | struct page, lru); |
1771 | list_del(&page->lru); | |
1772 | h->free_huge_pages--; | |
b2261026 | 1773 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1774 | if (acct_surplus) { |
1775 | h->surplus_huge_pages--; | |
b2261026 | 1776 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1777 | } |
e8c5c824 LS |
1778 | update_and_free_page(h, page); |
1779 | ret = 1; | |
9a76db09 | 1780 | break; |
e8c5c824 | 1781 | } |
b2261026 | 1782 | } |
e8c5c824 LS |
1783 | |
1784 | return ret; | |
1785 | } | |
1786 | ||
c8721bbb NH |
1787 | /* |
1788 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1789 | * nothing for in-use hugepages and non-hugepages. |
1790 | * This function returns values like below: | |
1791 | * | |
1792 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1793 | * (allocated or reserved.) | |
1794 | * 0: successfully dissolved free hugepages or the page is not a | |
1795 | * hugepage (considered as already dissolved) | |
c8721bbb | 1796 | */ |
c3114a84 | 1797 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1798 | { |
6bc9b564 | 1799 | int rc = -EBUSY; |
082d5b6b | 1800 | |
faf53def NH |
1801 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1802 | if (!PageHuge(page)) | |
1803 | return 0; | |
1804 | ||
c8721bbb | 1805 | spin_lock(&hugetlb_lock); |
faf53def NH |
1806 | if (!PageHuge(page)) { |
1807 | rc = 0; | |
1808 | goto out; | |
1809 | } | |
1810 | ||
1811 | if (!page_count(page)) { | |
2247bb33 GS |
1812 | struct page *head = compound_head(page); |
1813 | struct hstate *h = page_hstate(head); | |
1814 | int nid = page_to_nid(head); | |
6bc9b564 | 1815 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1816 | goto out; |
c3114a84 AK |
1817 | /* |
1818 | * Move PageHWPoison flag from head page to the raw error page, | |
1819 | * which makes any subpages rather than the error page reusable. | |
1820 | */ | |
1821 | if (PageHWPoison(head) && page != head) { | |
1822 | SetPageHWPoison(page); | |
1823 | ClearPageHWPoison(head); | |
1824 | } | |
2247bb33 | 1825 | list_del(&head->lru); |
c8721bbb NH |
1826 | h->free_huge_pages--; |
1827 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1828 | h->max_huge_pages--; |
2247bb33 | 1829 | update_and_free_page(h, head); |
6bc9b564 | 1830 | rc = 0; |
c8721bbb | 1831 | } |
082d5b6b | 1832 | out: |
c8721bbb | 1833 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1834 | return rc; |
c8721bbb NH |
1835 | } |
1836 | ||
1837 | /* | |
1838 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1839 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1840 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1841 | * part of it lies within the given range. | |
082d5b6b GS |
1842 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1843 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1844 | */ |
082d5b6b | 1845 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1846 | { |
c8721bbb | 1847 | unsigned long pfn; |
eb03aa00 | 1848 | struct page *page; |
082d5b6b | 1849 | int rc = 0; |
c8721bbb | 1850 | |
d0177639 | 1851 | if (!hugepages_supported()) |
082d5b6b | 1852 | return rc; |
d0177639 | 1853 | |
eb03aa00 GS |
1854 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1855 | page = pfn_to_page(pfn); | |
faf53def NH |
1856 | rc = dissolve_free_huge_page(page); |
1857 | if (rc) | |
1858 | break; | |
eb03aa00 | 1859 | } |
082d5b6b GS |
1860 | |
1861 | return rc; | |
c8721bbb NH |
1862 | } |
1863 | ||
ab5ac90a MH |
1864 | /* |
1865 | * Allocates a fresh surplus page from the page allocator. | |
1866 | */ | |
0c397dae | 1867 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1868 | int nid, nodemask_t *nmask) |
7893d1d5 | 1869 | { |
9980d744 | 1870 | struct page *page = NULL; |
7893d1d5 | 1871 | |
bae7f4ae | 1872 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1873 | return NULL; |
1874 | ||
d1c3fb1f | 1875 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1876 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1877 | goto out_unlock; | |
d1c3fb1f NA |
1878 | spin_unlock(&hugetlb_lock); |
1879 | ||
f60858f9 | 1880 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1881 | if (!page) |
0c397dae | 1882 | return NULL; |
d1c3fb1f NA |
1883 | |
1884 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1885 | /* |
1886 | * We could have raced with the pool size change. | |
1887 | * Double check that and simply deallocate the new page | |
1888 | * if we would end up overcommiting the surpluses. Abuse | |
1889 | * temporary page to workaround the nasty free_huge_page | |
1890 | * codeflow | |
1891 | */ | |
1892 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1893 | SetPageHugeTemporary(page); | |
2bf753e6 | 1894 | spin_unlock(&hugetlb_lock); |
9980d744 | 1895 | put_page(page); |
2bf753e6 | 1896 | return NULL; |
9980d744 | 1897 | } else { |
9980d744 | 1898 | h->surplus_huge_pages++; |
4704dea3 | 1899 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1900 | } |
9980d744 MH |
1901 | |
1902 | out_unlock: | |
d1c3fb1f | 1903 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1904 | |
1905 | return page; | |
1906 | } | |
1907 | ||
9a4e9f3b AK |
1908 | struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
1909 | int nid, nodemask_t *nmask) | |
ab5ac90a MH |
1910 | { |
1911 | struct page *page; | |
1912 | ||
1913 | if (hstate_is_gigantic(h)) | |
1914 | return NULL; | |
1915 | ||
f60858f9 | 1916 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1917 | if (!page) |
1918 | return NULL; | |
1919 | ||
1920 | /* | |
1921 | * We do not account these pages as surplus because they are only | |
1922 | * temporary and will be released properly on the last reference | |
1923 | */ | |
ab5ac90a MH |
1924 | SetPageHugeTemporary(page); |
1925 | ||
1926 | return page; | |
1927 | } | |
1928 | ||
099730d6 DH |
1929 | /* |
1930 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1931 | */ | |
e0ec90ee | 1932 | static |
0c397dae | 1933 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1934 | struct vm_area_struct *vma, unsigned long addr) |
1935 | { | |
aaf14e40 MH |
1936 | struct page *page; |
1937 | struct mempolicy *mpol; | |
1938 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1939 | int nid; | |
1940 | nodemask_t *nodemask; | |
1941 | ||
1942 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1943 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1944 | mpol_cond_put(mpol); |
1945 | ||
1946 | return page; | |
099730d6 DH |
1947 | } |
1948 | ||
ab5ac90a | 1949 | /* page migration callback function */ |
bf50bab2 NH |
1950 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
1951 | { | |
aaf14e40 | 1952 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1953 | struct page *page = NULL; |
bf50bab2 | 1954 | |
aaf14e40 MH |
1955 | if (nid != NUMA_NO_NODE) |
1956 | gfp_mask |= __GFP_THISNODE; | |
1957 | ||
bf50bab2 | 1958 | spin_lock(&hugetlb_lock); |
4ef91848 | 1959 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1960 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1961 | spin_unlock(&hugetlb_lock); |
1962 | ||
94ae8ba7 | 1963 | if (!page) |
0c397dae | 1964 | page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1965 | |
1966 | return page; | |
1967 | } | |
1968 | ||
ab5ac90a | 1969 | /* page migration callback function */ |
3e59fcb0 MH |
1970 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
1971 | nodemask_t *nmask) | |
4db9b2ef | 1972 | { |
aaf14e40 | 1973 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
1974 | |
1975 | spin_lock(&hugetlb_lock); | |
1976 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
1977 | struct page *page; |
1978 | ||
1979 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1980 | if (page) { | |
1981 | spin_unlock(&hugetlb_lock); | |
1982 | return page; | |
4db9b2ef MH |
1983 | } |
1984 | } | |
1985 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 1986 | |
0c397dae | 1987 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1988 | } |
1989 | ||
ebd63723 | 1990 | /* mempolicy aware migration callback */ |
389c8178 MH |
1991 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1992 | unsigned long address) | |
ebd63723 MH |
1993 | { |
1994 | struct mempolicy *mpol; | |
1995 | nodemask_t *nodemask; | |
1996 | struct page *page; | |
ebd63723 MH |
1997 | gfp_t gfp_mask; |
1998 | int node; | |
1999 | ||
ebd63723 MH |
2000 | gfp_mask = htlb_alloc_mask(h); |
2001 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
2002 | page = alloc_huge_page_nodemask(h, node, nodemask); | |
2003 | mpol_cond_put(mpol); | |
2004 | ||
2005 | return page; | |
2006 | } | |
2007 | ||
e4e574b7 | 2008 | /* |
25985edc | 2009 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2010 | * of size 'delta'. |
2011 | */ | |
a5516438 | 2012 | static int gather_surplus_pages(struct hstate *h, int delta) |
e4e574b7 AL |
2013 | { |
2014 | struct list_head surplus_list; | |
2015 | struct page *page, *tmp; | |
2016 | int ret, i; | |
2017 | int needed, allocated; | |
28073b02 | 2018 | bool alloc_ok = true; |
e4e574b7 | 2019 | |
a5516438 | 2020 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2021 | if (needed <= 0) { |
a5516438 | 2022 | h->resv_huge_pages += delta; |
e4e574b7 | 2023 | return 0; |
ac09b3a1 | 2024 | } |
e4e574b7 AL |
2025 | |
2026 | allocated = 0; | |
2027 | INIT_LIST_HEAD(&surplus_list); | |
2028 | ||
2029 | ret = -ENOMEM; | |
2030 | retry: | |
2031 | spin_unlock(&hugetlb_lock); | |
2032 | for (i = 0; i < needed; i++) { | |
0c397dae | 2033 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 2034 | NUMA_NO_NODE, NULL); |
28073b02 HD |
2035 | if (!page) { |
2036 | alloc_ok = false; | |
2037 | break; | |
2038 | } | |
e4e574b7 | 2039 | list_add(&page->lru, &surplus_list); |
69ed779a | 2040 | cond_resched(); |
e4e574b7 | 2041 | } |
28073b02 | 2042 | allocated += i; |
e4e574b7 AL |
2043 | |
2044 | /* | |
2045 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2046 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2047 | */ | |
2048 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
2049 | needed = (h->resv_huge_pages + delta) - |
2050 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2051 | if (needed > 0) { |
2052 | if (alloc_ok) | |
2053 | goto retry; | |
2054 | /* | |
2055 | * We were not able to allocate enough pages to | |
2056 | * satisfy the entire reservation so we free what | |
2057 | * we've allocated so far. | |
2058 | */ | |
2059 | goto free; | |
2060 | } | |
e4e574b7 AL |
2061 | /* |
2062 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2063 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2064 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2065 | * allocator. Commit the entire reservation here to prevent another |
2066 | * process from stealing the pages as they are added to the pool but | |
2067 | * before they are reserved. | |
e4e574b7 AL |
2068 | */ |
2069 | needed += allocated; | |
a5516438 | 2070 | h->resv_huge_pages += delta; |
e4e574b7 | 2071 | ret = 0; |
a9869b83 | 2072 | |
19fc3f0a | 2073 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2074 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2075 | if ((--needed) < 0) |
2076 | break; | |
a9869b83 NH |
2077 | /* |
2078 | * This page is now managed by the hugetlb allocator and has | |
2079 | * no users -- drop the buddy allocator's reference. | |
2080 | */ | |
2081 | put_page_testzero(page); | |
309381fe | 2082 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 2083 | enqueue_huge_page(h, page); |
19fc3f0a | 2084 | } |
28073b02 | 2085 | free: |
b0365c8d | 2086 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
2087 | |
2088 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
2089 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
2090 | put_page(page); | |
a9869b83 | 2091 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
2092 | |
2093 | return ret; | |
2094 | } | |
2095 | ||
2096 | /* | |
e5bbc8a6 MK |
2097 | * This routine has two main purposes: |
2098 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2099 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2100 | * to the associated reservation map. | |
2101 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2102 | * the reservation. As many as unused_resv_pages may be freed. | |
2103 | * | |
2104 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
2105 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
2106 | * we must make sure nobody else can claim pages we are in the process of | |
2107 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
2108 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 2109 | */ |
a5516438 AK |
2110 | static void return_unused_surplus_pages(struct hstate *h, |
2111 | unsigned long unused_resv_pages) | |
e4e574b7 | 2112 | { |
e4e574b7 AL |
2113 | unsigned long nr_pages; |
2114 | ||
aa888a74 | 2115 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2116 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2117 | goto out; |
aa888a74 | 2118 | |
e5bbc8a6 MK |
2119 | /* |
2120 | * Part (or even all) of the reservation could have been backed | |
2121 | * by pre-allocated pages. Only free surplus pages. | |
2122 | */ | |
a5516438 | 2123 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2124 | |
685f3457 LS |
2125 | /* |
2126 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2127 | * evenly across all nodes with memory. Iterate across these nodes |
2128 | * until we can no longer free unreserved surplus pages. This occurs | |
2129 | * when the nodes with surplus pages have no free pages. | |
2130 | * free_pool_huge_page() will balance the the freed pages across the | |
2131 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
2132 | * |
2133 | * Note that we decrement resv_huge_pages as we free the pages. If | |
2134 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
2135 | * to cover subsequent pages we may free. | |
685f3457 LS |
2136 | */ |
2137 | while (nr_pages--) { | |
e5bbc8a6 MK |
2138 | h->resv_huge_pages--; |
2139 | unused_resv_pages--; | |
8cebfcd0 | 2140 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 2141 | goto out; |
7848a4bf | 2142 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 2143 | } |
e5bbc8a6 MK |
2144 | |
2145 | out: | |
2146 | /* Fully uncommit the reservation */ | |
2147 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
2148 | } |
2149 | ||
5e911373 | 2150 | |
c37f9fb1 | 2151 | /* |
feba16e2 | 2152 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2153 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2154 | * |
2155 | * vma_needs_reservation is called to determine if the huge page at addr | |
2156 | * within the vma has an associated reservation. If a reservation is | |
2157 | * needed, the value 1 is returned. The caller is then responsible for | |
2158 | * managing the global reservation and subpool usage counts. After | |
2159 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2160 | * to add the page to the reservation map. If the page allocation fails, |
2161 | * the reservation must be ended instead of committed. vma_end_reservation | |
2162 | * is called in such cases. | |
cf3ad20b MK |
2163 | * |
2164 | * In the normal case, vma_commit_reservation returns the same value | |
2165 | * as the preceding vma_needs_reservation call. The only time this | |
2166 | * is not the case is if a reserve map was changed between calls. It | |
2167 | * is the responsibility of the caller to notice the difference and | |
2168 | * take appropriate action. | |
96b96a96 MK |
2169 | * |
2170 | * vma_add_reservation is used in error paths where a reservation must | |
2171 | * be restored when a newly allocated huge page must be freed. It is | |
2172 | * to be called after calling vma_needs_reservation to determine if a | |
2173 | * reservation exists. | |
c37f9fb1 | 2174 | */ |
5e911373 MK |
2175 | enum vma_resv_mode { |
2176 | VMA_NEEDS_RESV, | |
2177 | VMA_COMMIT_RESV, | |
feba16e2 | 2178 | VMA_END_RESV, |
96b96a96 | 2179 | VMA_ADD_RESV, |
5e911373 | 2180 | }; |
cf3ad20b MK |
2181 | static long __vma_reservation_common(struct hstate *h, |
2182 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2183 | enum vma_resv_mode mode) |
c37f9fb1 | 2184 | { |
4e35f483 JK |
2185 | struct resv_map *resv; |
2186 | pgoff_t idx; | |
cf3ad20b | 2187 | long ret; |
0db9d74e | 2188 | long dummy_out_regions_needed; |
c37f9fb1 | 2189 | |
4e35f483 JK |
2190 | resv = vma_resv_map(vma); |
2191 | if (!resv) | |
84afd99b | 2192 | return 1; |
c37f9fb1 | 2193 | |
4e35f483 | 2194 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2195 | switch (mode) { |
2196 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2197 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2198 | /* We assume that vma_reservation_* routines always operate on | |
2199 | * 1 page, and that adding to resv map a 1 page entry can only | |
2200 | * ever require 1 region. | |
2201 | */ | |
2202 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2203 | break; |
2204 | case VMA_COMMIT_RESV: | |
075a61d0 | 2205 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2206 | /* region_add calls of range 1 should never fail. */ |
2207 | VM_BUG_ON(ret < 0); | |
5e911373 | 2208 | break; |
feba16e2 | 2209 | case VMA_END_RESV: |
0db9d74e | 2210 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2211 | ret = 0; |
2212 | break; | |
96b96a96 | 2213 | case VMA_ADD_RESV: |
0db9d74e | 2214 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2215 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2216 | /* region_add calls of range 1 should never fail. */ |
2217 | VM_BUG_ON(ret < 0); | |
2218 | } else { | |
2219 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2220 | ret = region_del(resv, idx, idx + 1); |
2221 | } | |
2222 | break; | |
5e911373 MK |
2223 | default: |
2224 | BUG(); | |
2225 | } | |
84afd99b | 2226 | |
4e35f483 | 2227 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 2228 | return ret; |
67961f9d MK |
2229 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
2230 | /* | |
2231 | * In most cases, reserves always exist for private mappings. | |
2232 | * However, a file associated with mapping could have been | |
2233 | * hole punched or truncated after reserves were consumed. | |
2234 | * As subsequent fault on such a range will not use reserves. | |
2235 | * Subtle - The reserve map for private mappings has the | |
2236 | * opposite meaning than that of shared mappings. If NO | |
2237 | * entry is in the reserve map, it means a reservation exists. | |
2238 | * If an entry exists in the reserve map, it means the | |
2239 | * reservation has already been consumed. As a result, the | |
2240 | * return value of this routine is the opposite of the | |
2241 | * value returned from reserve map manipulation routines above. | |
2242 | */ | |
2243 | if (ret) | |
2244 | return 0; | |
2245 | else | |
2246 | return 1; | |
2247 | } | |
4e35f483 | 2248 | else |
cf3ad20b | 2249 | return ret < 0 ? ret : 0; |
c37f9fb1 | 2250 | } |
cf3ad20b MK |
2251 | |
2252 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2253 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2254 | { |
5e911373 | 2255 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2256 | } |
84afd99b | 2257 | |
cf3ad20b MK |
2258 | static long vma_commit_reservation(struct hstate *h, |
2259 | struct vm_area_struct *vma, unsigned long addr) | |
2260 | { | |
5e911373 MK |
2261 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2262 | } | |
2263 | ||
feba16e2 | 2264 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2265 | struct vm_area_struct *vma, unsigned long addr) |
2266 | { | |
feba16e2 | 2267 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2268 | } |
2269 | ||
96b96a96 MK |
2270 | static long vma_add_reservation(struct hstate *h, |
2271 | struct vm_area_struct *vma, unsigned long addr) | |
2272 | { | |
2273 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2274 | } | |
2275 | ||
2276 | /* | |
2277 | * This routine is called to restore a reservation on error paths. In the | |
2278 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2279 | * and is about to be freed. If a reservation for the page existed, | |
2280 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2281 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2282 | * the global reservation count will be incremented if PagePrivate is set. | |
2283 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2284 | * reserve map here to be consistent with global reserve count adjustments | |
2285 | * to be made by free_huge_page. | |
2286 | */ | |
2287 | static void restore_reserve_on_error(struct hstate *h, | |
2288 | struct vm_area_struct *vma, unsigned long address, | |
2289 | struct page *page) | |
2290 | { | |
2291 | if (unlikely(PagePrivate(page))) { | |
2292 | long rc = vma_needs_reservation(h, vma, address); | |
2293 | ||
2294 | if (unlikely(rc < 0)) { | |
2295 | /* | |
2296 | * Rare out of memory condition in reserve map | |
2297 | * manipulation. Clear PagePrivate so that | |
2298 | * global reserve count will not be incremented | |
2299 | * by free_huge_page. This will make it appear | |
2300 | * as though the reservation for this page was | |
2301 | * consumed. This may prevent the task from | |
2302 | * faulting in the page at a later time. This | |
2303 | * is better than inconsistent global huge page | |
2304 | * accounting of reserve counts. | |
2305 | */ | |
2306 | ClearPagePrivate(page); | |
2307 | } else if (rc) { | |
2308 | rc = vma_add_reservation(h, vma, address); | |
2309 | if (unlikely(rc < 0)) | |
2310 | /* | |
2311 | * See above comment about rare out of | |
2312 | * memory condition. | |
2313 | */ | |
2314 | ClearPagePrivate(page); | |
2315 | } else | |
2316 | vma_end_reservation(h, vma, address); | |
2317 | } | |
2318 | } | |
2319 | ||
70c3547e | 2320 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2321 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2322 | { |
90481622 | 2323 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2324 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2325 | struct page *page; |
d85f69b0 MK |
2326 | long map_chg, map_commit; |
2327 | long gbl_chg; | |
6d76dcf4 AK |
2328 | int ret, idx; |
2329 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2330 | bool deferred_reserve; |
a1e78772 | 2331 | |
6d76dcf4 | 2332 | idx = hstate_index(h); |
a1e78772 | 2333 | /* |
d85f69b0 MK |
2334 | * Examine the region/reserve map to determine if the process |
2335 | * has a reservation for the page to be allocated. A return | |
2336 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2337 | */ |
d85f69b0 MK |
2338 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2339 | if (map_chg < 0) | |
76dcee75 | 2340 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2341 | |
2342 | /* | |
2343 | * Processes that did not create the mapping will have no | |
2344 | * reserves as indicated by the region/reserve map. Check | |
2345 | * that the allocation will not exceed the subpool limit. | |
2346 | * Allocations for MAP_NORESERVE mappings also need to be | |
2347 | * checked against any subpool limit. | |
2348 | */ | |
2349 | if (map_chg || avoid_reserve) { | |
2350 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2351 | if (gbl_chg < 0) { | |
feba16e2 | 2352 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2353 | return ERR_PTR(-ENOSPC); |
5e911373 | 2354 | } |
1da177e4 | 2355 | |
d85f69b0 MK |
2356 | /* |
2357 | * Even though there was no reservation in the region/reserve | |
2358 | * map, there could be reservations associated with the | |
2359 | * subpool that can be used. This would be indicated if the | |
2360 | * return value of hugepage_subpool_get_pages() is zero. | |
2361 | * However, if avoid_reserve is specified we still avoid even | |
2362 | * the subpool reservations. | |
2363 | */ | |
2364 | if (avoid_reserve) | |
2365 | gbl_chg = 1; | |
2366 | } | |
2367 | ||
08cf9faf MA |
2368 | /* If this allocation is not consuming a reservation, charge it now. |
2369 | */ | |
2370 | deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); | |
2371 | if (deferred_reserve) { | |
2372 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2373 | idx, pages_per_huge_page(h), &h_cg); | |
2374 | if (ret) | |
2375 | goto out_subpool_put; | |
2376 | } | |
2377 | ||
6d76dcf4 | 2378 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2379 | if (ret) |
08cf9faf | 2380 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2381 | |
1da177e4 | 2382 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2383 | /* |
2384 | * glb_chg is passed to indicate whether or not a page must be taken | |
2385 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2386 | * a reservation exists for the allocation. | |
2387 | */ | |
2388 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2389 | if (!page) { |
94ae8ba7 | 2390 | spin_unlock(&hugetlb_lock); |
0c397dae | 2391 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2392 | if (!page) |
2393 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2394 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2395 | SetPagePrivate(page); | |
2396 | h->resv_huge_pages--; | |
2397 | } | |
79dbb236 AK |
2398 | spin_lock(&hugetlb_lock); |
2399 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2400 | /* Fall through */ |
68842c9b | 2401 | } |
81a6fcae | 2402 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2403 | /* If allocation is not consuming a reservation, also store the |
2404 | * hugetlb_cgroup pointer on the page. | |
2405 | */ | |
2406 | if (deferred_reserve) { | |
2407 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2408 | h_cg, page); | |
2409 | } | |
2410 | ||
81a6fcae | 2411 | spin_unlock(&hugetlb_lock); |
348ea204 | 2412 | |
90481622 | 2413 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2414 | |
d85f69b0 MK |
2415 | map_commit = vma_commit_reservation(h, vma, addr); |
2416 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2417 | /* |
2418 | * The page was added to the reservation map between | |
2419 | * vma_needs_reservation and vma_commit_reservation. | |
2420 | * This indicates a race with hugetlb_reserve_pages. | |
2421 | * Adjust for the subpool count incremented above AND | |
2422 | * in hugetlb_reserve_pages for the same page. Also, | |
2423 | * the reservation count added in hugetlb_reserve_pages | |
2424 | * no longer applies. | |
2425 | */ | |
2426 | long rsv_adjust; | |
2427 | ||
2428 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2429 | hugetlb_acct_memory(h, -rsv_adjust); | |
2430 | } | |
90d8b7e6 | 2431 | return page; |
8f34af6f JZ |
2432 | |
2433 | out_uncharge_cgroup: | |
2434 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2435 | out_uncharge_cgroup_reservation: |
2436 | if (deferred_reserve) | |
2437 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2438 | h_cg); | |
8f34af6f | 2439 | out_subpool_put: |
d85f69b0 | 2440 | if (map_chg || avoid_reserve) |
8f34af6f | 2441 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2442 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2443 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2444 | } |
2445 | ||
e24a1307 AK |
2446 | int alloc_bootmem_huge_page(struct hstate *h) |
2447 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2448 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2449 | { |
2450 | struct huge_bootmem_page *m; | |
b2261026 | 2451 | int nr_nodes, node; |
aa888a74 | 2452 | |
b2261026 | 2453 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2454 | void *addr; |
2455 | ||
eb31d559 | 2456 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2457 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2458 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2459 | if (addr) { |
2460 | /* | |
2461 | * Use the beginning of the huge page to store the | |
2462 | * huge_bootmem_page struct (until gather_bootmem | |
2463 | * puts them into the mem_map). | |
2464 | */ | |
2465 | m = addr; | |
91f47662 | 2466 | goto found; |
aa888a74 | 2467 | } |
aa888a74 AK |
2468 | } |
2469 | return 0; | |
2470 | ||
2471 | found: | |
df994ead | 2472 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2473 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2474 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2475 | list_add(&m->list, &huge_boot_pages); |
2476 | m->hstate = h; | |
2477 | return 1; | |
2478 | } | |
2479 | ||
d00181b9 KS |
2480 | static void __init prep_compound_huge_page(struct page *page, |
2481 | unsigned int order) | |
18229df5 AW |
2482 | { |
2483 | if (unlikely(order > (MAX_ORDER - 1))) | |
2484 | prep_compound_gigantic_page(page, order); | |
2485 | else | |
2486 | prep_compound_page(page, order); | |
2487 | } | |
2488 | ||
aa888a74 AK |
2489 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2490 | static void __init gather_bootmem_prealloc(void) | |
2491 | { | |
2492 | struct huge_bootmem_page *m; | |
2493 | ||
2494 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2495 | struct page *page = virt_to_page(m); |
aa888a74 | 2496 | struct hstate *h = m->hstate; |
ee8f248d | 2497 | |
aa888a74 | 2498 | WARN_ON(page_count(page) != 1); |
18229df5 | 2499 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2500 | WARN_ON(PageReserved(page)); |
aa888a74 | 2501 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2502 | put_page(page); /* free it into the hugepage allocator */ |
2503 | ||
b0320c7b RA |
2504 | /* |
2505 | * If we had gigantic hugepages allocated at boot time, we need | |
2506 | * to restore the 'stolen' pages to totalram_pages in order to | |
2507 | * fix confusing memory reports from free(1) and another | |
2508 | * side-effects, like CommitLimit going negative. | |
2509 | */ | |
bae7f4ae | 2510 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2511 | adjust_managed_page_count(page, 1 << h->order); |
520495fe | 2512 | cond_resched(); |
aa888a74 AK |
2513 | } |
2514 | } | |
2515 | ||
8faa8b07 | 2516 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2517 | { |
2518 | unsigned long i; | |
f60858f9 MK |
2519 | nodemask_t *node_alloc_noretry; |
2520 | ||
2521 | if (!hstate_is_gigantic(h)) { | |
2522 | /* | |
2523 | * Bit mask controlling how hard we retry per-node allocations. | |
2524 | * Ignore errors as lower level routines can deal with | |
2525 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2526 | * time, we are likely in bigger trouble. | |
2527 | */ | |
2528 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2529 | GFP_KERNEL); | |
2530 | } else { | |
2531 | /* allocations done at boot time */ | |
2532 | node_alloc_noretry = NULL; | |
2533 | } | |
2534 | ||
2535 | /* bit mask controlling how hard we retry per-node allocations */ | |
2536 | if (node_alloc_noretry) | |
2537 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2538 | |
e5ff2159 | 2539 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2540 | if (hstate_is_gigantic(h)) { |
aa888a74 AK |
2541 | if (!alloc_bootmem_huge_page(h)) |
2542 | break; | |
0c397dae | 2543 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2544 | &node_states[N_MEMORY], |
2545 | node_alloc_noretry)) | |
1da177e4 | 2546 | break; |
69ed779a | 2547 | cond_resched(); |
1da177e4 | 2548 | } |
d715cf80 LH |
2549 | if (i < h->max_huge_pages) { |
2550 | char buf[32]; | |
2551 | ||
c6247f72 | 2552 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2553 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2554 | h->max_huge_pages, buf, i); | |
2555 | h->max_huge_pages = i; | |
2556 | } | |
f60858f9 MK |
2557 | |
2558 | kfree(node_alloc_noretry); | |
e5ff2159 AK |
2559 | } |
2560 | ||
2561 | static void __init hugetlb_init_hstates(void) | |
2562 | { | |
2563 | struct hstate *h; | |
2564 | ||
2565 | for_each_hstate(h) { | |
641844f5 NH |
2566 | if (minimum_order > huge_page_order(h)) |
2567 | minimum_order = huge_page_order(h); | |
2568 | ||
8faa8b07 | 2569 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2570 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2571 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2572 | } |
641844f5 | 2573 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2574 | } |
2575 | ||
2576 | static void __init report_hugepages(void) | |
2577 | { | |
2578 | struct hstate *h; | |
2579 | ||
2580 | for_each_hstate(h) { | |
4abd32db | 2581 | char buf[32]; |
c6247f72 MW |
2582 | |
2583 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2584 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2585 | buf, h->free_huge_pages); |
e5ff2159 AK |
2586 | } |
2587 | } | |
2588 | ||
1da177e4 | 2589 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2590 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2591 | nodemask_t *nodes_allowed) | |
1da177e4 | 2592 | { |
4415cc8d CL |
2593 | int i; |
2594 | ||
bae7f4ae | 2595 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2596 | return; |
2597 | ||
6ae11b27 | 2598 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2599 | struct page *page, *next; |
a5516438 AK |
2600 | struct list_head *freel = &h->hugepage_freelists[i]; |
2601 | list_for_each_entry_safe(page, next, freel, lru) { | |
2602 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2603 | return; |
1da177e4 LT |
2604 | if (PageHighMem(page)) |
2605 | continue; | |
2606 | list_del(&page->lru); | |
e5ff2159 | 2607 | update_and_free_page(h, page); |
a5516438 AK |
2608 | h->free_huge_pages--; |
2609 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2610 | } |
2611 | } | |
2612 | } | |
2613 | #else | |
6ae11b27 LS |
2614 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2615 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2616 | { |
2617 | } | |
2618 | #endif | |
2619 | ||
20a0307c WF |
2620 | /* |
2621 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2622 | * balanced by operating on them in a round-robin fashion. | |
2623 | * Returns 1 if an adjustment was made. | |
2624 | */ | |
6ae11b27 LS |
2625 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2626 | int delta) | |
20a0307c | 2627 | { |
b2261026 | 2628 | int nr_nodes, node; |
20a0307c WF |
2629 | |
2630 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2631 | |
b2261026 JK |
2632 | if (delta < 0) { |
2633 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2634 | if (h->surplus_huge_pages_node[node]) | |
2635 | goto found; | |
e8c5c824 | 2636 | } |
b2261026 JK |
2637 | } else { |
2638 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2639 | if (h->surplus_huge_pages_node[node] < | |
2640 | h->nr_huge_pages_node[node]) | |
2641 | goto found; | |
e8c5c824 | 2642 | } |
b2261026 JK |
2643 | } |
2644 | return 0; | |
20a0307c | 2645 | |
b2261026 JK |
2646 | found: |
2647 | h->surplus_huge_pages += delta; | |
2648 | h->surplus_huge_pages_node[node] += delta; | |
2649 | return 1; | |
20a0307c WF |
2650 | } |
2651 | ||
a5516438 | 2652 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2653 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2654 | nodemask_t *nodes_allowed) |
1da177e4 | 2655 | { |
7893d1d5 | 2656 | unsigned long min_count, ret; |
f60858f9 MK |
2657 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2658 | ||
2659 | /* | |
2660 | * Bit mask controlling how hard we retry per-node allocations. | |
2661 | * If we can not allocate the bit mask, do not attempt to allocate | |
2662 | * the requested huge pages. | |
2663 | */ | |
2664 | if (node_alloc_noretry) | |
2665 | nodes_clear(*node_alloc_noretry); | |
2666 | else | |
2667 | return -ENOMEM; | |
1da177e4 | 2668 | |
4eb0716e AG |
2669 | spin_lock(&hugetlb_lock); |
2670 | ||
fd875dca MK |
2671 | /* |
2672 | * Check for a node specific request. | |
2673 | * Changing node specific huge page count may require a corresponding | |
2674 | * change to the global count. In any case, the passed node mask | |
2675 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2676 | */ | |
2677 | if (nid != NUMA_NO_NODE) { | |
2678 | unsigned long old_count = count; | |
2679 | ||
2680 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2681 | /* | |
2682 | * User may have specified a large count value which caused the | |
2683 | * above calculation to overflow. In this case, they wanted | |
2684 | * to allocate as many huge pages as possible. Set count to | |
2685 | * largest possible value to align with their intention. | |
2686 | */ | |
2687 | if (count < old_count) | |
2688 | count = ULONG_MAX; | |
2689 | } | |
2690 | ||
4eb0716e AG |
2691 | /* |
2692 | * Gigantic pages runtime allocation depend on the capability for large | |
2693 | * page range allocation. | |
2694 | * If the system does not provide this feature, return an error when | |
2695 | * the user tries to allocate gigantic pages but let the user free the | |
2696 | * boottime allocated gigantic pages. | |
2697 | */ | |
2698 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2699 | if (count > persistent_huge_pages(h)) { | |
2700 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2701 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2702 | return -EINVAL; |
2703 | } | |
2704 | /* Fall through to decrease pool */ | |
2705 | } | |
aa888a74 | 2706 | |
7893d1d5 AL |
2707 | /* |
2708 | * Increase the pool size | |
2709 | * First take pages out of surplus state. Then make up the | |
2710 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2711 | * |
0c397dae | 2712 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2713 | * to convert a surplus huge page to a normal huge page. That is |
2714 | * not critical, though, it just means the overall size of the | |
2715 | * pool might be one hugepage larger than it needs to be, but | |
2716 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2717 | */ |
a5516438 | 2718 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2719 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2720 | break; |
2721 | } | |
2722 | ||
a5516438 | 2723 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2724 | /* |
2725 | * If this allocation races such that we no longer need the | |
2726 | * page, free_huge_page will handle it by freeing the page | |
2727 | * and reducing the surplus. | |
2728 | */ | |
2729 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2730 | |
2731 | /* yield cpu to avoid soft lockup */ | |
2732 | cond_resched(); | |
2733 | ||
f60858f9 MK |
2734 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2735 | node_alloc_noretry); | |
7893d1d5 AL |
2736 | spin_lock(&hugetlb_lock); |
2737 | if (!ret) | |
2738 | goto out; | |
2739 | ||
536240f2 MG |
2740 | /* Bail for signals. Probably ctrl-c from user */ |
2741 | if (signal_pending(current)) | |
2742 | goto out; | |
7893d1d5 | 2743 | } |
7893d1d5 AL |
2744 | |
2745 | /* | |
2746 | * Decrease the pool size | |
2747 | * First return free pages to the buddy allocator (being careful | |
2748 | * to keep enough around to satisfy reservations). Then place | |
2749 | * pages into surplus state as needed so the pool will shrink | |
2750 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2751 | * |
2752 | * By placing pages into the surplus state independent of the | |
2753 | * overcommit value, we are allowing the surplus pool size to | |
2754 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2755 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2756 | * though, we'll note that we're not allowed to exceed surplus |
2757 | * and won't grow the pool anywhere else. Not until one of the | |
2758 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2759 | */ |
a5516438 | 2760 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2761 | min_count = max(count, min_count); |
6ae11b27 | 2762 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2763 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2764 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2765 | break; |
55f67141 | 2766 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2767 | } |
a5516438 | 2768 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2769 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2770 | break; |
2771 | } | |
2772 | out: | |
4eb0716e | 2773 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2774 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2775 | |
f60858f9 MK |
2776 | NODEMASK_FREE(node_alloc_noretry); |
2777 | ||
4eb0716e | 2778 | return 0; |
1da177e4 LT |
2779 | } |
2780 | ||
a3437870 NA |
2781 | #define HSTATE_ATTR_RO(_name) \ |
2782 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2783 | ||
2784 | #define HSTATE_ATTR(_name) \ | |
2785 | static struct kobj_attribute _name##_attr = \ | |
2786 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2787 | ||
2788 | static struct kobject *hugepages_kobj; | |
2789 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2790 | ||
9a305230 LS |
2791 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2792 | ||
2793 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2794 | { |
2795 | int i; | |
9a305230 | 2796 | |
a3437870 | 2797 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2798 | if (hstate_kobjs[i] == kobj) { |
2799 | if (nidp) | |
2800 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2801 | return &hstates[i]; |
9a305230 LS |
2802 | } |
2803 | ||
2804 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2805 | } |
2806 | ||
06808b08 | 2807 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2808 | struct kobj_attribute *attr, char *buf) |
2809 | { | |
9a305230 LS |
2810 | struct hstate *h; |
2811 | unsigned long nr_huge_pages; | |
2812 | int nid; | |
2813 | ||
2814 | h = kobj_to_hstate(kobj, &nid); | |
2815 | if (nid == NUMA_NO_NODE) | |
2816 | nr_huge_pages = h->nr_huge_pages; | |
2817 | else | |
2818 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2819 | ||
2820 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2821 | } |
adbe8726 | 2822 | |
238d3c13 DR |
2823 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2824 | struct hstate *h, int nid, | |
2825 | unsigned long count, size_t len) | |
a3437870 NA |
2826 | { |
2827 | int err; | |
2d0adf7e | 2828 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2829 | |
2d0adf7e OS |
2830 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2831 | return -EINVAL; | |
adbe8726 | 2832 | |
9a305230 LS |
2833 | if (nid == NUMA_NO_NODE) { |
2834 | /* | |
2835 | * global hstate attribute | |
2836 | */ | |
2837 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2838 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2839 | n_mask = &node_states[N_MEMORY]; | |
2840 | else | |
2841 | n_mask = &nodes_allowed; | |
2842 | } else { | |
9a305230 | 2843 | /* |
fd875dca MK |
2844 | * Node specific request. count adjustment happens in |
2845 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2846 | */ |
2d0adf7e OS |
2847 | init_nodemask_of_node(&nodes_allowed, nid); |
2848 | n_mask = &nodes_allowed; | |
fd875dca | 2849 | } |
9a305230 | 2850 | |
2d0adf7e | 2851 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2852 | |
4eb0716e | 2853 | return err ? err : len; |
06808b08 LS |
2854 | } |
2855 | ||
238d3c13 DR |
2856 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2857 | struct kobject *kobj, const char *buf, | |
2858 | size_t len) | |
2859 | { | |
2860 | struct hstate *h; | |
2861 | unsigned long count; | |
2862 | int nid; | |
2863 | int err; | |
2864 | ||
2865 | err = kstrtoul(buf, 10, &count); | |
2866 | if (err) | |
2867 | return err; | |
2868 | ||
2869 | h = kobj_to_hstate(kobj, &nid); | |
2870 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2871 | } | |
2872 | ||
06808b08 LS |
2873 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2874 | struct kobj_attribute *attr, char *buf) | |
2875 | { | |
2876 | return nr_hugepages_show_common(kobj, attr, buf); | |
2877 | } | |
2878 | ||
2879 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2880 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2881 | { | |
238d3c13 | 2882 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2883 | } |
2884 | HSTATE_ATTR(nr_hugepages); | |
2885 | ||
06808b08 LS |
2886 | #ifdef CONFIG_NUMA |
2887 | ||
2888 | /* | |
2889 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2890 | * huge page alloc/free. | |
2891 | */ | |
2892 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2893 | struct kobj_attribute *attr, char *buf) | |
2894 | { | |
2895 | return nr_hugepages_show_common(kobj, attr, buf); | |
2896 | } | |
2897 | ||
2898 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2899 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2900 | { | |
238d3c13 | 2901 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2902 | } |
2903 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2904 | #endif | |
2905 | ||
2906 | ||
a3437870 NA |
2907 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2908 | struct kobj_attribute *attr, char *buf) | |
2909 | { | |
9a305230 | 2910 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2911 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2912 | } | |
adbe8726 | 2913 | |
a3437870 NA |
2914 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2915 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2916 | { | |
2917 | int err; | |
2918 | unsigned long input; | |
9a305230 | 2919 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2920 | |
bae7f4ae | 2921 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2922 | return -EINVAL; |
2923 | ||
3dbb95f7 | 2924 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2925 | if (err) |
73ae31e5 | 2926 | return err; |
a3437870 NA |
2927 | |
2928 | spin_lock(&hugetlb_lock); | |
2929 | h->nr_overcommit_huge_pages = input; | |
2930 | spin_unlock(&hugetlb_lock); | |
2931 | ||
2932 | return count; | |
2933 | } | |
2934 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2935 | ||
2936 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2937 | struct kobj_attribute *attr, char *buf) | |
2938 | { | |
9a305230 LS |
2939 | struct hstate *h; |
2940 | unsigned long free_huge_pages; | |
2941 | int nid; | |
2942 | ||
2943 | h = kobj_to_hstate(kobj, &nid); | |
2944 | if (nid == NUMA_NO_NODE) | |
2945 | free_huge_pages = h->free_huge_pages; | |
2946 | else | |
2947 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2948 | ||
2949 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2950 | } |
2951 | HSTATE_ATTR_RO(free_hugepages); | |
2952 | ||
2953 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2954 | struct kobj_attribute *attr, char *buf) | |
2955 | { | |
9a305230 | 2956 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2957 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2958 | } | |
2959 | HSTATE_ATTR_RO(resv_hugepages); | |
2960 | ||
2961 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2962 | struct kobj_attribute *attr, char *buf) | |
2963 | { | |
9a305230 LS |
2964 | struct hstate *h; |
2965 | unsigned long surplus_huge_pages; | |
2966 | int nid; | |
2967 | ||
2968 | h = kobj_to_hstate(kobj, &nid); | |
2969 | if (nid == NUMA_NO_NODE) | |
2970 | surplus_huge_pages = h->surplus_huge_pages; | |
2971 | else | |
2972 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2973 | ||
2974 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
2975 | } |
2976 | HSTATE_ATTR_RO(surplus_hugepages); | |
2977 | ||
2978 | static struct attribute *hstate_attrs[] = { | |
2979 | &nr_hugepages_attr.attr, | |
2980 | &nr_overcommit_hugepages_attr.attr, | |
2981 | &free_hugepages_attr.attr, | |
2982 | &resv_hugepages_attr.attr, | |
2983 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2984 | #ifdef CONFIG_NUMA |
2985 | &nr_hugepages_mempolicy_attr.attr, | |
2986 | #endif | |
a3437870 NA |
2987 | NULL, |
2988 | }; | |
2989 | ||
67e5ed96 | 2990 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2991 | .attrs = hstate_attrs, |
2992 | }; | |
2993 | ||
094e9539 JM |
2994 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2995 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2996 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2997 | { |
2998 | int retval; | |
972dc4de | 2999 | int hi = hstate_index(h); |
a3437870 | 3000 | |
9a305230 LS |
3001 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3002 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3003 | return -ENOMEM; |
3004 | ||
9a305230 | 3005 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 3006 | if (retval) |
9a305230 | 3007 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
3008 | |
3009 | return retval; | |
3010 | } | |
3011 | ||
3012 | static void __init hugetlb_sysfs_init(void) | |
3013 | { | |
3014 | struct hstate *h; | |
3015 | int err; | |
3016 | ||
3017 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3018 | if (!hugepages_kobj) | |
3019 | return; | |
3020 | ||
3021 | for_each_hstate(h) { | |
9a305230 LS |
3022 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3023 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3024 | if (err) |
ffb22af5 | 3025 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
3026 | } |
3027 | } | |
3028 | ||
9a305230 LS |
3029 | #ifdef CONFIG_NUMA |
3030 | ||
3031 | /* | |
3032 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3033 | * with node devices in node_devices[] using a parallel array. The array |
3034 | * index of a node device or _hstate == node id. | |
3035 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3036 | * the base kernel, on the hugetlb module. |
3037 | */ | |
3038 | struct node_hstate { | |
3039 | struct kobject *hugepages_kobj; | |
3040 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3041 | }; | |
b4e289a6 | 3042 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3043 | |
3044 | /* | |
10fbcf4c | 3045 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3046 | */ |
3047 | static struct attribute *per_node_hstate_attrs[] = { | |
3048 | &nr_hugepages_attr.attr, | |
3049 | &free_hugepages_attr.attr, | |
3050 | &surplus_hugepages_attr.attr, | |
3051 | NULL, | |
3052 | }; | |
3053 | ||
67e5ed96 | 3054 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3055 | .attrs = per_node_hstate_attrs, |
3056 | }; | |
3057 | ||
3058 | /* | |
10fbcf4c | 3059 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3060 | * Returns node id via non-NULL nidp. |
3061 | */ | |
3062 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3063 | { | |
3064 | int nid; | |
3065 | ||
3066 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3067 | struct node_hstate *nhs = &node_hstates[nid]; | |
3068 | int i; | |
3069 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3070 | if (nhs->hstate_kobjs[i] == kobj) { | |
3071 | if (nidp) | |
3072 | *nidp = nid; | |
3073 | return &hstates[i]; | |
3074 | } | |
3075 | } | |
3076 | ||
3077 | BUG(); | |
3078 | return NULL; | |
3079 | } | |
3080 | ||
3081 | /* | |
10fbcf4c | 3082 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3083 | * No-op if no hstate attributes attached. |
3084 | */ | |
3cd8b44f | 3085 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3086 | { |
3087 | struct hstate *h; | |
10fbcf4c | 3088 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3089 | |
3090 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3091 | return; /* no hstate attributes */ |
9a305230 | 3092 | |
972dc4de AK |
3093 | for_each_hstate(h) { |
3094 | int idx = hstate_index(h); | |
3095 | if (nhs->hstate_kobjs[idx]) { | |
3096 | kobject_put(nhs->hstate_kobjs[idx]); | |
3097 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3098 | } |
972dc4de | 3099 | } |
9a305230 LS |
3100 | |
3101 | kobject_put(nhs->hugepages_kobj); | |
3102 | nhs->hugepages_kobj = NULL; | |
3103 | } | |
3104 | ||
9a305230 LS |
3105 | |
3106 | /* | |
10fbcf4c | 3107 | * Register hstate attributes for a single node device. |
9a305230 LS |
3108 | * No-op if attributes already registered. |
3109 | */ | |
3cd8b44f | 3110 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3111 | { |
3112 | struct hstate *h; | |
10fbcf4c | 3113 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3114 | int err; |
3115 | ||
3116 | if (nhs->hugepages_kobj) | |
3117 | return; /* already allocated */ | |
3118 | ||
3119 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3120 | &node->dev.kobj); |
9a305230 LS |
3121 | if (!nhs->hugepages_kobj) |
3122 | return; | |
3123 | ||
3124 | for_each_hstate(h) { | |
3125 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3126 | nhs->hstate_kobjs, | |
3127 | &per_node_hstate_attr_group); | |
3128 | if (err) { | |
ffb22af5 AM |
3129 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
3130 | h->name, node->dev.id); | |
9a305230 LS |
3131 | hugetlb_unregister_node(node); |
3132 | break; | |
3133 | } | |
3134 | } | |
3135 | } | |
3136 | ||
3137 | /* | |
9b5e5d0f | 3138 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3139 | * devices of nodes that have memory. All on-line nodes should have |
3140 | * registered their associated device by this time. | |
9a305230 | 3141 | */ |
7d9ca000 | 3142 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3143 | { |
3144 | int nid; | |
3145 | ||
8cebfcd0 | 3146 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3147 | struct node *node = node_devices[nid]; |
10fbcf4c | 3148 | if (node->dev.id == nid) |
9a305230 LS |
3149 | hugetlb_register_node(node); |
3150 | } | |
3151 | ||
3152 | /* | |
10fbcf4c | 3153 | * Let the node device driver know we're here so it can |
9a305230 LS |
3154 | * [un]register hstate attributes on node hotplug. |
3155 | */ | |
3156 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3157 | hugetlb_unregister_node); | |
3158 | } | |
3159 | #else /* !CONFIG_NUMA */ | |
3160 | ||
3161 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3162 | { | |
3163 | BUG(); | |
3164 | if (nidp) | |
3165 | *nidp = -1; | |
3166 | return NULL; | |
3167 | } | |
3168 | ||
9a305230 LS |
3169 | static void hugetlb_register_all_nodes(void) { } |
3170 | ||
3171 | #endif | |
3172 | ||
a3437870 NA |
3173 | static int __init hugetlb_init(void) |
3174 | { | |
8382d914 DB |
3175 | int i; |
3176 | ||
457c1b27 | 3177 | if (!hugepages_supported()) |
0ef89d25 | 3178 | return 0; |
a3437870 | 3179 | |
e11bfbfc | 3180 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
3181 | if (default_hstate_size != 0) { |
3182 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
3183 | default_hstate_size, HPAGE_SIZE); | |
3184 | } | |
3185 | ||
e11bfbfc NP |
3186 | default_hstate_size = HPAGE_SIZE; |
3187 | if (!size_to_hstate(default_hstate_size)) | |
3188 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 3189 | } |
972dc4de | 3190 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
3191 | if (default_hstate_max_huge_pages) { |
3192 | if (!default_hstate.max_huge_pages) | |
3193 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3194 | } | |
a3437870 NA |
3195 | |
3196 | hugetlb_init_hstates(); | |
aa888a74 | 3197 | gather_bootmem_prealloc(); |
a3437870 NA |
3198 | report_hugepages(); |
3199 | ||
3200 | hugetlb_sysfs_init(); | |
9a305230 | 3201 | hugetlb_register_all_nodes(); |
7179e7bf | 3202 | hugetlb_cgroup_file_init(); |
9a305230 | 3203 | |
8382d914 DB |
3204 | #ifdef CONFIG_SMP |
3205 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3206 | #else | |
3207 | num_fault_mutexes = 1; | |
3208 | #endif | |
c672c7f2 | 3209 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3210 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3211 | GFP_KERNEL); | |
c672c7f2 | 3212 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3213 | |
3214 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3215 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3216 | return 0; |
3217 | } | |
3e89e1c5 | 3218 | subsys_initcall(hugetlb_init); |
a3437870 NA |
3219 | |
3220 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
3221 | void __init hugetlb_bad_size(void) |
3222 | { | |
3223 | parsed_valid_hugepagesz = false; | |
3224 | } | |
3225 | ||
d00181b9 | 3226 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3227 | { |
3228 | struct hstate *h; | |
8faa8b07 AK |
3229 | unsigned long i; |
3230 | ||
a3437870 | 3231 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 3232 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
3233 | return; |
3234 | } | |
47d38344 | 3235 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3236 | BUG_ON(order == 0); |
47d38344 | 3237 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
3238 | h->order = order; |
3239 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
3240 | h->nr_huge_pages = 0; |
3241 | h->free_huge_pages = 0; | |
3242 | for (i = 0; i < MAX_NUMNODES; ++i) | |
3243 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3244 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3245 | h->next_nid_to_alloc = first_memory_node; |
3246 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3247 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3248 | huge_page_size(h)/1024); | |
8faa8b07 | 3249 | |
a3437870 NA |
3250 | parsed_hstate = h; |
3251 | } | |
3252 | ||
e11bfbfc | 3253 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
3254 | { |
3255 | unsigned long *mhp; | |
8faa8b07 | 3256 | static unsigned long *last_mhp; |
a3437870 | 3257 | |
9fee021d VT |
3258 | if (!parsed_valid_hugepagesz) { |
3259 | pr_warn("hugepages = %s preceded by " | |
3260 | "an unsupported hugepagesz, ignoring\n", s); | |
3261 | parsed_valid_hugepagesz = true; | |
3262 | return 1; | |
3263 | } | |
a3437870 | 3264 | /* |
47d38344 | 3265 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
3266 | * so this hugepages= parameter goes to the "default hstate". |
3267 | */ | |
9fee021d | 3268 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3269 | mhp = &default_hstate_max_huge_pages; |
3270 | else | |
3271 | mhp = &parsed_hstate->max_huge_pages; | |
3272 | ||
8faa8b07 | 3273 | if (mhp == last_mhp) { |
598d8091 | 3274 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
3275 | return 1; |
3276 | } | |
3277 | ||
a3437870 NA |
3278 | if (sscanf(s, "%lu", mhp) <= 0) |
3279 | *mhp = 0; | |
3280 | ||
8faa8b07 AK |
3281 | /* |
3282 | * Global state is always initialized later in hugetlb_init. | |
3283 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
3284 | * use the bootmem allocator. | |
3285 | */ | |
47d38344 | 3286 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
3287 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3288 | ||
3289 | last_mhp = mhp; | |
3290 | ||
a3437870 NA |
3291 | return 1; |
3292 | } | |
e11bfbfc NP |
3293 | __setup("hugepages=", hugetlb_nrpages_setup); |
3294 | ||
3295 | static int __init hugetlb_default_setup(char *s) | |
3296 | { | |
3297 | default_hstate_size = memparse(s, &s); | |
3298 | return 1; | |
3299 | } | |
3300 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 3301 | |
8a213460 NA |
3302 | static unsigned int cpuset_mems_nr(unsigned int *array) |
3303 | { | |
3304 | int node; | |
3305 | unsigned int nr = 0; | |
3306 | ||
3307 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
3308 | nr += array[node]; | |
3309 | ||
3310 | return nr; | |
3311 | } | |
3312 | ||
3313 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
3314 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3315 | struct ctl_table *table, int write, | |
3316 | void __user *buffer, size_t *length, loff_t *ppos) | |
1da177e4 | 3317 | { |
e5ff2159 | 3318 | struct hstate *h = &default_hstate; |
238d3c13 | 3319 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3320 | int ret; |
e5ff2159 | 3321 | |
457c1b27 | 3322 | if (!hugepages_supported()) |
86613628 | 3323 | return -EOPNOTSUPP; |
457c1b27 | 3324 | |
e5ff2159 AK |
3325 | table->data = &tmp; |
3326 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3327 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3328 | if (ret) | |
3329 | goto out; | |
e5ff2159 | 3330 | |
238d3c13 DR |
3331 | if (write) |
3332 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3333 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3334 | out: |
3335 | return ret; | |
1da177e4 | 3336 | } |
396faf03 | 3337 | |
06808b08 LS |
3338 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
3339 | void __user *buffer, size_t *length, loff_t *ppos) | |
3340 | { | |
3341 | ||
3342 | return hugetlb_sysctl_handler_common(false, table, write, | |
3343 | buffer, length, ppos); | |
3344 | } | |
3345 | ||
3346 | #ifdef CONFIG_NUMA | |
3347 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
3348 | void __user *buffer, size_t *length, loff_t *ppos) | |
3349 | { | |
3350 | return hugetlb_sysctl_handler_common(true, table, write, | |
3351 | buffer, length, ppos); | |
3352 | } | |
3353 | #endif /* CONFIG_NUMA */ | |
3354 | ||
a3d0c6aa | 3355 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
8d65af78 | 3356 | void __user *buffer, |
a3d0c6aa NA |
3357 | size_t *length, loff_t *ppos) |
3358 | { | |
a5516438 | 3359 | struct hstate *h = &default_hstate; |
e5ff2159 | 3360 | unsigned long tmp; |
08d4a246 | 3361 | int ret; |
e5ff2159 | 3362 | |
457c1b27 | 3363 | if (!hugepages_supported()) |
86613628 | 3364 | return -EOPNOTSUPP; |
457c1b27 | 3365 | |
c033a93c | 3366 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3367 | |
bae7f4ae | 3368 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3369 | return -EINVAL; |
3370 | ||
e5ff2159 AK |
3371 | table->data = &tmp; |
3372 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3373 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3374 | if (ret) | |
3375 | goto out; | |
e5ff2159 AK |
3376 | |
3377 | if (write) { | |
3378 | spin_lock(&hugetlb_lock); | |
3379 | h->nr_overcommit_huge_pages = tmp; | |
3380 | spin_unlock(&hugetlb_lock); | |
3381 | } | |
08d4a246 MH |
3382 | out: |
3383 | return ret; | |
a3d0c6aa NA |
3384 | } |
3385 | ||
1da177e4 LT |
3386 | #endif /* CONFIG_SYSCTL */ |
3387 | ||
e1759c21 | 3388 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3389 | { |
fcb2b0c5 RG |
3390 | struct hstate *h; |
3391 | unsigned long total = 0; | |
3392 | ||
457c1b27 NA |
3393 | if (!hugepages_supported()) |
3394 | return; | |
fcb2b0c5 RG |
3395 | |
3396 | for_each_hstate(h) { | |
3397 | unsigned long count = h->nr_huge_pages; | |
3398 | ||
3399 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3400 | ||
3401 | if (h == &default_hstate) | |
3402 | seq_printf(m, | |
3403 | "HugePages_Total: %5lu\n" | |
3404 | "HugePages_Free: %5lu\n" | |
3405 | "HugePages_Rsvd: %5lu\n" | |
3406 | "HugePages_Surp: %5lu\n" | |
3407 | "Hugepagesize: %8lu kB\n", | |
3408 | count, | |
3409 | h->free_huge_pages, | |
3410 | h->resv_huge_pages, | |
3411 | h->surplus_huge_pages, | |
3412 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3413 | } | |
3414 | ||
3415 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3416 | } |
3417 | ||
3418 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3419 | { | |
a5516438 | 3420 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3421 | if (!hugepages_supported()) |
3422 | return 0; | |
1da177e4 LT |
3423 | return sprintf(buf, |
3424 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3425 | "Node %d HugePages_Free: %5u\n" |
3426 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3427 | nid, h->nr_huge_pages_node[nid], |
3428 | nid, h->free_huge_pages_node[nid], | |
3429 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3430 | } |
3431 | ||
949f7ec5 DR |
3432 | void hugetlb_show_meminfo(void) |
3433 | { | |
3434 | struct hstate *h; | |
3435 | int nid; | |
3436 | ||
457c1b27 NA |
3437 | if (!hugepages_supported()) |
3438 | return; | |
3439 | ||
949f7ec5 DR |
3440 | for_each_node_state(nid, N_MEMORY) |
3441 | for_each_hstate(h) | |
3442 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3443 | nid, | |
3444 | h->nr_huge_pages_node[nid], | |
3445 | h->free_huge_pages_node[nid], | |
3446 | h->surplus_huge_pages_node[nid], | |
3447 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3448 | } | |
3449 | ||
5d317b2b NH |
3450 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3451 | { | |
3452 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3453 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3454 | } | |
3455 | ||
1da177e4 LT |
3456 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3457 | unsigned long hugetlb_total_pages(void) | |
3458 | { | |
d0028588 WL |
3459 | struct hstate *h; |
3460 | unsigned long nr_total_pages = 0; | |
3461 | ||
3462 | for_each_hstate(h) | |
3463 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3464 | return nr_total_pages; | |
1da177e4 | 3465 | } |
1da177e4 | 3466 | |
a5516438 | 3467 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3468 | { |
3469 | int ret = -ENOMEM; | |
3470 | ||
3471 | spin_lock(&hugetlb_lock); | |
3472 | /* | |
3473 | * When cpuset is configured, it breaks the strict hugetlb page | |
3474 | * reservation as the accounting is done on a global variable. Such | |
3475 | * reservation is completely rubbish in the presence of cpuset because | |
3476 | * the reservation is not checked against page availability for the | |
3477 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3478 | * with lack of free htlb page in cpuset that the task is in. | |
3479 | * Attempt to enforce strict accounting with cpuset is almost | |
3480 | * impossible (or too ugly) because cpuset is too fluid that | |
3481 | * task or memory node can be dynamically moved between cpusets. | |
3482 | * | |
3483 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3484 | * undesirable. However, in order to preserve some of the semantics, | |
3485 | * we fall back to check against current free page availability as | |
3486 | * a best attempt and hopefully to minimize the impact of changing | |
3487 | * semantics that cpuset has. | |
3488 | */ | |
3489 | if (delta > 0) { | |
a5516438 | 3490 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3491 | goto out; |
3492 | ||
a5516438 AK |
3493 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3494 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3495 | goto out; |
3496 | } | |
3497 | } | |
3498 | ||
3499 | ret = 0; | |
3500 | if (delta < 0) | |
a5516438 | 3501 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3502 | |
3503 | out: | |
3504 | spin_unlock(&hugetlb_lock); | |
3505 | return ret; | |
3506 | } | |
3507 | ||
84afd99b AW |
3508 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3509 | { | |
f522c3ac | 3510 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3511 | |
3512 | /* | |
3513 | * This new VMA should share its siblings reservation map if present. | |
3514 | * The VMA will only ever have a valid reservation map pointer where | |
3515 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3516 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3517 | * after this open call completes. It is therefore safe to take a |
3518 | * new reference here without additional locking. | |
3519 | */ | |
4e35f483 | 3520 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3521 | kref_get(&resv->refs); |
84afd99b AW |
3522 | } |
3523 | ||
a1e78772 MG |
3524 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3525 | { | |
a5516438 | 3526 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3527 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3528 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3529 | unsigned long reserve, start, end; |
1c5ecae3 | 3530 | long gbl_reserve; |
84afd99b | 3531 | |
4e35f483 JK |
3532 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3533 | return; | |
84afd99b | 3534 | |
4e35f483 JK |
3535 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3536 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3537 | |
4e35f483 | 3538 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 3539 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 3540 | if (reserve) { |
1c5ecae3 MK |
3541 | /* |
3542 | * Decrement reserve counts. The global reserve count may be | |
3543 | * adjusted if the subpool has a minimum size. | |
3544 | */ | |
3545 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3546 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3547 | } |
e9fe92ae MA |
3548 | |
3549 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
3550 | } |
3551 | ||
31383c68 DW |
3552 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3553 | { | |
3554 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3555 | return -EINVAL; | |
3556 | return 0; | |
3557 | } | |
3558 | ||
05ea8860 DW |
3559 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3560 | { | |
3561 | struct hstate *hstate = hstate_vma(vma); | |
3562 | ||
3563 | return 1UL << huge_page_shift(hstate); | |
3564 | } | |
3565 | ||
1da177e4 LT |
3566 | /* |
3567 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3568 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3569 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3570 | * this far. | |
3571 | */ | |
b3ec9f33 | 3572 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3573 | { |
3574 | BUG(); | |
d0217ac0 | 3575 | return 0; |
1da177e4 LT |
3576 | } |
3577 | ||
eec3636a JC |
3578 | /* |
3579 | * When a new function is introduced to vm_operations_struct and added | |
3580 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3581 | * This is because under System V memory model, mappings created via | |
3582 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3583 | * their original vm_ops are overwritten with shm_vm_ops. | |
3584 | */ | |
f0f37e2f | 3585 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3586 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3587 | .open = hugetlb_vm_op_open, |
a1e78772 | 3588 | .close = hugetlb_vm_op_close, |
31383c68 | 3589 | .split = hugetlb_vm_op_split, |
05ea8860 | 3590 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3591 | }; |
3592 | ||
1e8f889b DG |
3593 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3594 | int writable) | |
63551ae0 DG |
3595 | { |
3596 | pte_t entry; | |
3597 | ||
1e8f889b | 3598 | if (writable) { |
106c992a GS |
3599 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3600 | vma->vm_page_prot))); | |
63551ae0 | 3601 | } else { |
106c992a GS |
3602 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3603 | vma->vm_page_prot)); | |
63551ae0 DG |
3604 | } |
3605 | entry = pte_mkyoung(entry); | |
3606 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3607 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3608 | |
3609 | return entry; | |
3610 | } | |
3611 | ||
1e8f889b DG |
3612 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3613 | unsigned long address, pte_t *ptep) | |
3614 | { | |
3615 | pte_t entry; | |
3616 | ||
106c992a | 3617 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3618 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3619 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3620 | } |
3621 | ||
d5ed7444 | 3622 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3623 | { |
3624 | swp_entry_t swp; | |
3625 | ||
3626 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3627 | return false; |
4a705fef NH |
3628 | swp = pte_to_swp_entry(pte); |
3629 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3630 | return true; |
4a705fef | 3631 | else |
d5ed7444 | 3632 | return false; |
4a705fef NH |
3633 | } |
3634 | ||
3635 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3636 | { | |
3637 | swp_entry_t swp; | |
3638 | ||
3639 | if (huge_pte_none(pte) || pte_present(pte)) | |
3640 | return 0; | |
3641 | swp = pte_to_swp_entry(pte); | |
3642 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3643 | return 1; | |
3644 | else | |
3645 | return 0; | |
3646 | } | |
1e8f889b | 3647 | |
63551ae0 DG |
3648 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3649 | struct vm_area_struct *vma) | |
3650 | { | |
5e41540c | 3651 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3652 | struct page *ptepage; |
1c59827d | 3653 | unsigned long addr; |
1e8f889b | 3654 | int cow; |
a5516438 AK |
3655 | struct hstate *h = hstate_vma(vma); |
3656 | unsigned long sz = huge_page_size(h); | |
c0d0381a | 3657 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 3658 | struct mmu_notifier_range range; |
e8569dd2 | 3659 | int ret = 0; |
1e8f889b DG |
3660 | |
3661 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3662 | |
ac46d4f3 | 3663 | if (cow) { |
7269f999 | 3664 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3665 | vma->vm_start, |
ac46d4f3 JG |
3666 | vma->vm_end); |
3667 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
3668 | } else { |
3669 | /* | |
3670 | * For shared mappings i_mmap_rwsem must be held to call | |
3671 | * huge_pte_alloc, otherwise the returned ptep could go | |
3672 | * away if part of a shared pmd and another thread calls | |
3673 | * huge_pmd_unshare. | |
3674 | */ | |
3675 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 3676 | } |
e8569dd2 | 3677 | |
a5516438 | 3678 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3679 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3680 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3681 | if (!src_pte) |
3682 | continue; | |
a5516438 | 3683 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3684 | if (!dst_pte) { |
3685 | ret = -ENOMEM; | |
3686 | break; | |
3687 | } | |
c5c99429 | 3688 | |
5e41540c MK |
3689 | /* |
3690 | * If the pagetables are shared don't copy or take references. | |
3691 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3692 | * | |
3693 | * However, src could have 'unshared' and dst shares with | |
3694 | * another vma. If dst_pte !none, this implies sharing. | |
3695 | * Check here before taking page table lock, and once again | |
3696 | * after taking the lock below. | |
3697 | */ | |
3698 | dst_entry = huge_ptep_get(dst_pte); | |
3699 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3700 | continue; |
3701 | ||
cb900f41 KS |
3702 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3703 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3704 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3705 | entry = huge_ptep_get(src_pte); |
5e41540c MK |
3706 | dst_entry = huge_ptep_get(dst_pte); |
3707 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3708 | /* | |
3709 | * Skip if src entry none. Also, skip in the | |
3710 | * unlikely case dst entry !none as this implies | |
3711 | * sharing with another vma. | |
3712 | */ | |
4a705fef NH |
3713 | ; |
3714 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3715 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3716 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3717 | ||
3718 | if (is_write_migration_entry(swp_entry) && cow) { | |
3719 | /* | |
3720 | * COW mappings require pages in both | |
3721 | * parent and child to be set to read. | |
3722 | */ | |
3723 | make_migration_entry_read(&swp_entry); | |
3724 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3725 | set_huge_swap_pte_at(src, addr, src_pte, |
3726 | entry, sz); | |
4a705fef | 3727 | } |
e5251fd4 | 3728 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3729 | } else { |
34ee645e | 3730 | if (cow) { |
0f10851e JG |
3731 | /* |
3732 | * No need to notify as we are downgrading page | |
3733 | * table protection not changing it to point | |
3734 | * to a new page. | |
3735 | * | |
ad56b738 | 3736 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3737 | */ |
7f2e9525 | 3738 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3739 | } |
0253d634 | 3740 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3741 | ptepage = pte_page(entry); |
3742 | get_page(ptepage); | |
53f9263b | 3743 | page_dup_rmap(ptepage, true); |
1c59827d | 3744 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3745 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3746 | } |
cb900f41 KS |
3747 | spin_unlock(src_ptl); |
3748 | spin_unlock(dst_ptl); | |
63551ae0 | 3749 | } |
63551ae0 | 3750 | |
e8569dd2 | 3751 | if (cow) |
ac46d4f3 | 3752 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
3753 | else |
3754 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
3755 | |
3756 | return ret; | |
63551ae0 DG |
3757 | } |
3758 | ||
24669e58 AK |
3759 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3760 | unsigned long start, unsigned long end, | |
3761 | struct page *ref_page) | |
63551ae0 DG |
3762 | { |
3763 | struct mm_struct *mm = vma->vm_mm; | |
3764 | unsigned long address; | |
c7546f8f | 3765 | pte_t *ptep; |
63551ae0 | 3766 | pte_t pte; |
cb900f41 | 3767 | spinlock_t *ptl; |
63551ae0 | 3768 | struct page *page; |
a5516438 AK |
3769 | struct hstate *h = hstate_vma(vma); |
3770 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3771 | struct mmu_notifier_range range; |
a5516438 | 3772 | |
63551ae0 | 3773 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3774 | BUG_ON(start & ~huge_page_mask(h)); |
3775 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3776 | |
07e32661 AK |
3777 | /* |
3778 | * This is a hugetlb vma, all the pte entries should point | |
3779 | * to huge page. | |
3780 | */ | |
ed6a7935 | 3781 | tlb_change_page_size(tlb, sz); |
24669e58 | 3782 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3783 | |
3784 | /* | |
3785 | * If sharing possible, alert mmu notifiers of worst case. | |
3786 | */ | |
6f4f13e8 JG |
3787 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3788 | end); | |
ac46d4f3 JG |
3789 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3790 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3791 | address = start; |
569f48b8 | 3792 | for (; address < end; address += sz) { |
7868a208 | 3793 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3794 | if (!ptep) |
c7546f8f DG |
3795 | continue; |
3796 | ||
cb900f41 | 3797 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3798 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3799 | spin_unlock(ptl); | |
dff11abe MK |
3800 | /* |
3801 | * We just unmapped a page of PMDs by clearing a PUD. | |
3802 | * The caller's TLB flush range should cover this area. | |
3803 | */ | |
31d49da5 AK |
3804 | continue; |
3805 | } | |
39dde65c | 3806 | |
6629326b | 3807 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3808 | if (huge_pte_none(pte)) { |
3809 | spin_unlock(ptl); | |
3810 | continue; | |
3811 | } | |
6629326b HD |
3812 | |
3813 | /* | |
9fbc1f63 NH |
3814 | * Migrating hugepage or HWPoisoned hugepage is already |
3815 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3816 | */ |
9fbc1f63 | 3817 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3818 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3819 | spin_unlock(ptl); |
3820 | continue; | |
8c4894c6 | 3821 | } |
6629326b HD |
3822 | |
3823 | page = pte_page(pte); | |
04f2cbe3 MG |
3824 | /* |
3825 | * If a reference page is supplied, it is because a specific | |
3826 | * page is being unmapped, not a range. Ensure the page we | |
3827 | * are about to unmap is the actual page of interest. | |
3828 | */ | |
3829 | if (ref_page) { | |
31d49da5 AK |
3830 | if (page != ref_page) { |
3831 | spin_unlock(ptl); | |
3832 | continue; | |
3833 | } | |
04f2cbe3 MG |
3834 | /* |
3835 | * Mark the VMA as having unmapped its page so that | |
3836 | * future faults in this VMA will fail rather than | |
3837 | * looking like data was lost | |
3838 | */ | |
3839 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3840 | } | |
3841 | ||
c7546f8f | 3842 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3843 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3844 | if (huge_pte_dirty(pte)) |
6649a386 | 3845 | set_page_dirty(page); |
9e81130b | 3846 | |
5d317b2b | 3847 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3848 | page_remove_rmap(page, true); |
31d49da5 | 3849 | |
cb900f41 | 3850 | spin_unlock(ptl); |
e77b0852 | 3851 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3852 | /* |
3853 | * Bail out after unmapping reference page if supplied | |
3854 | */ | |
3855 | if (ref_page) | |
3856 | break; | |
fe1668ae | 3857 | } |
ac46d4f3 | 3858 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 3859 | tlb_end_vma(tlb, vma); |
1da177e4 | 3860 | } |
63551ae0 | 3861 | |
d833352a MG |
3862 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3863 | struct vm_area_struct *vma, unsigned long start, | |
3864 | unsigned long end, struct page *ref_page) | |
3865 | { | |
3866 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3867 | ||
3868 | /* | |
3869 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3870 | * test will fail on a vma being torn down, and not grab a page table | |
3871 | * on its way out. We're lucky that the flag has such an appropriate | |
3872 | * name, and can in fact be safely cleared here. We could clear it | |
3873 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3874 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3875 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3876 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3877 | */ |
3878 | vma->vm_flags &= ~VM_MAYSHARE; | |
3879 | } | |
3880 | ||
502717f4 | 3881 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3882 | unsigned long end, struct page *ref_page) |
502717f4 | 3883 | { |
24669e58 AK |
3884 | struct mm_struct *mm; |
3885 | struct mmu_gather tlb; | |
dff11abe MK |
3886 | unsigned long tlb_start = start; |
3887 | unsigned long tlb_end = end; | |
3888 | ||
3889 | /* | |
3890 | * If shared PMDs were possibly used within this vma range, adjust | |
3891 | * start/end for worst case tlb flushing. | |
3892 | * Note that we can not be sure if PMDs are shared until we try to | |
3893 | * unmap pages. However, we want to make sure TLB flushing covers | |
3894 | * the largest possible range. | |
3895 | */ | |
3896 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
24669e58 AK |
3897 | |
3898 | mm = vma->vm_mm; | |
3899 | ||
dff11abe | 3900 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); |
24669e58 | 3901 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
dff11abe | 3902 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); |
502717f4 KC |
3903 | } |
3904 | ||
04f2cbe3 MG |
3905 | /* |
3906 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3907 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3908 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3909 | * same region. | |
3910 | */ | |
2f4612af DB |
3911 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3912 | struct page *page, unsigned long address) | |
04f2cbe3 | 3913 | { |
7526674d | 3914 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3915 | struct vm_area_struct *iter_vma; |
3916 | struct address_space *mapping; | |
04f2cbe3 MG |
3917 | pgoff_t pgoff; |
3918 | ||
3919 | /* | |
3920 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3921 | * from page cache lookup which is in HPAGE_SIZE units. | |
3922 | */ | |
7526674d | 3923 | address = address & huge_page_mask(h); |
36e4f20a MH |
3924 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3925 | vma->vm_pgoff; | |
93c76a3d | 3926 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3927 | |
4eb2b1dc MG |
3928 | /* |
3929 | * Take the mapping lock for the duration of the table walk. As | |
3930 | * this mapping should be shared between all the VMAs, | |
3931 | * __unmap_hugepage_range() is called as the lock is already held | |
3932 | */ | |
83cde9e8 | 3933 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3934 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3935 | /* Do not unmap the current VMA */ |
3936 | if (iter_vma == vma) | |
3937 | continue; | |
3938 | ||
2f84a899 MG |
3939 | /* |
3940 | * Shared VMAs have their own reserves and do not affect | |
3941 | * MAP_PRIVATE accounting but it is possible that a shared | |
3942 | * VMA is using the same page so check and skip such VMAs. | |
3943 | */ | |
3944 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3945 | continue; | |
3946 | ||
04f2cbe3 MG |
3947 | /* |
3948 | * Unmap the page from other VMAs without their own reserves. | |
3949 | * They get marked to be SIGKILLed if they fault in these | |
3950 | * areas. This is because a future no-page fault on this VMA | |
3951 | * could insert a zeroed page instead of the data existing | |
3952 | * from the time of fork. This would look like data corruption | |
3953 | */ | |
3954 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3955 | unmap_hugepage_range(iter_vma, address, |
3956 | address + huge_page_size(h), page); | |
04f2cbe3 | 3957 | } |
83cde9e8 | 3958 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3959 | } |
3960 | ||
0fe6e20b NH |
3961 | /* |
3962 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
3963 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
3964 | * cannot race with other handlers or page migration. | |
3965 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 3966 | */ |
2b740303 | 3967 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 3968 | unsigned long address, pte_t *ptep, |
3999f52e | 3969 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 3970 | { |
3999f52e | 3971 | pte_t pte; |
a5516438 | 3972 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 3973 | struct page *old_page, *new_page; |
2b740303 SJ |
3974 | int outside_reserve = 0; |
3975 | vm_fault_t ret = 0; | |
974e6d66 | 3976 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 3977 | struct mmu_notifier_range range; |
1e8f889b | 3978 | |
3999f52e | 3979 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
3980 | old_page = pte_page(pte); |
3981 | ||
04f2cbe3 | 3982 | retry_avoidcopy: |
1e8f889b DG |
3983 | /* If no-one else is actually using this page, avoid the copy |
3984 | * and just make the page writable */ | |
37a2140d | 3985 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 3986 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 3987 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 3988 | return 0; |
1e8f889b DG |
3989 | } |
3990 | ||
04f2cbe3 MG |
3991 | /* |
3992 | * If the process that created a MAP_PRIVATE mapping is about to | |
3993 | * perform a COW due to a shared page count, attempt to satisfy | |
3994 | * the allocation without using the existing reserves. The pagecache | |
3995 | * page is used to determine if the reserve at this address was | |
3996 | * consumed or not. If reserves were used, a partial faulted mapping | |
3997 | * at the time of fork() could consume its reserves on COW instead | |
3998 | * of the full address range. | |
3999 | */ | |
5944d011 | 4000 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
4001 | old_page != pagecache_page) |
4002 | outside_reserve = 1; | |
4003 | ||
09cbfeaf | 4004 | get_page(old_page); |
b76c8cfb | 4005 | |
ad4404a2 DB |
4006 | /* |
4007 | * Drop page table lock as buddy allocator may be called. It will | |
4008 | * be acquired again before returning to the caller, as expected. | |
4009 | */ | |
cb900f41 | 4010 | spin_unlock(ptl); |
5b7a1d40 | 4011 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 4012 | |
2fc39cec | 4013 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
4014 | /* |
4015 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4016 | * it is due to references held by a child and an insufficient | |
4017 | * huge page pool. To guarantee the original mappers | |
4018 | * reliability, unmap the page from child processes. The child | |
4019 | * may get SIGKILLed if it later faults. | |
4020 | */ | |
4021 | if (outside_reserve) { | |
09cbfeaf | 4022 | put_page(old_page); |
04f2cbe3 | 4023 | BUG_ON(huge_pte_none(pte)); |
5b7a1d40 | 4024 | unmap_ref_private(mm, vma, old_page, haddr); |
2f4612af DB |
4025 | BUG_ON(huge_pte_none(pte)); |
4026 | spin_lock(ptl); | |
5b7a1d40 | 4027 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
4028 | if (likely(ptep && |
4029 | pte_same(huge_ptep_get(ptep), pte))) | |
4030 | goto retry_avoidcopy; | |
4031 | /* | |
4032 | * race occurs while re-acquiring page table | |
4033 | * lock, and our job is done. | |
4034 | */ | |
4035 | return 0; | |
04f2cbe3 MG |
4036 | } |
4037 | ||
2b740303 | 4038 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 4039 | goto out_release_old; |
1e8f889b DG |
4040 | } |
4041 | ||
0fe6e20b NH |
4042 | /* |
4043 | * When the original hugepage is shared one, it does not have | |
4044 | * anon_vma prepared. | |
4045 | */ | |
44e2aa93 | 4046 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
4047 | ret = VM_FAULT_OOM; |
4048 | goto out_release_all; | |
44e2aa93 | 4049 | } |
0fe6e20b | 4050 | |
974e6d66 | 4051 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 4052 | pages_per_huge_page(h)); |
0ed361de | 4053 | __SetPageUptodate(new_page); |
1e8f889b | 4054 | |
7269f999 | 4055 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 4056 | haddr + huge_page_size(h)); |
ac46d4f3 | 4057 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 4058 | |
b76c8cfb | 4059 | /* |
cb900f41 | 4060 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
4061 | * before the page tables are altered |
4062 | */ | |
cb900f41 | 4063 | spin_lock(ptl); |
5b7a1d40 | 4064 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 4065 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
4066 | ClearPagePrivate(new_page); |
4067 | ||
1e8f889b | 4068 | /* Break COW */ |
5b7a1d40 | 4069 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 4070 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 4071 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 4072 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 4073 | page_remove_rmap(old_page, true); |
5b7a1d40 | 4074 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
cb6acd01 | 4075 | set_page_huge_active(new_page); |
1e8f889b DG |
4076 | /* Make the old page be freed below */ |
4077 | new_page = old_page; | |
4078 | } | |
cb900f41 | 4079 | spin_unlock(ptl); |
ac46d4f3 | 4080 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 4081 | out_release_all: |
5b7a1d40 | 4082 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 4083 | put_page(new_page); |
ad4404a2 | 4084 | out_release_old: |
09cbfeaf | 4085 | put_page(old_page); |
8312034f | 4086 | |
ad4404a2 DB |
4087 | spin_lock(ptl); /* Caller expects lock to be held */ |
4088 | return ret; | |
1e8f889b DG |
4089 | } |
4090 | ||
04f2cbe3 | 4091 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
4092 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
4093 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
4094 | { |
4095 | struct address_space *mapping; | |
e7c4b0bf | 4096 | pgoff_t idx; |
04f2cbe3 MG |
4097 | |
4098 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 4099 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
4100 | |
4101 | return find_lock_page(mapping, idx); | |
4102 | } | |
4103 | ||
3ae77f43 HD |
4104 | /* |
4105 | * Return whether there is a pagecache page to back given address within VMA. | |
4106 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4107 | */ | |
4108 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
4109 | struct vm_area_struct *vma, unsigned long address) |
4110 | { | |
4111 | struct address_space *mapping; | |
4112 | pgoff_t idx; | |
4113 | struct page *page; | |
4114 | ||
4115 | mapping = vma->vm_file->f_mapping; | |
4116 | idx = vma_hugecache_offset(h, vma, address); | |
4117 | ||
4118 | page = find_get_page(mapping, idx); | |
4119 | if (page) | |
4120 | put_page(page); | |
4121 | return page != NULL; | |
4122 | } | |
4123 | ||
ab76ad54 MK |
4124 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
4125 | pgoff_t idx) | |
4126 | { | |
4127 | struct inode *inode = mapping->host; | |
4128 | struct hstate *h = hstate_inode(inode); | |
4129 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4130 | ||
4131 | if (err) | |
4132 | return err; | |
4133 | ClearPagePrivate(page); | |
4134 | ||
22146c3c MK |
4135 | /* |
4136 | * set page dirty so that it will not be removed from cache/file | |
4137 | * by non-hugetlbfs specific code paths. | |
4138 | */ | |
4139 | set_page_dirty(page); | |
4140 | ||
ab76ad54 MK |
4141 | spin_lock(&inode->i_lock); |
4142 | inode->i_blocks += blocks_per_huge_page(h); | |
4143 | spin_unlock(&inode->i_lock); | |
4144 | return 0; | |
4145 | } | |
4146 | ||
2b740303 SJ |
4147 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
4148 | struct vm_area_struct *vma, | |
4149 | struct address_space *mapping, pgoff_t idx, | |
4150 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 4151 | { |
a5516438 | 4152 | struct hstate *h = hstate_vma(vma); |
2b740303 | 4153 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 4154 | int anon_rmap = 0; |
4c887265 | 4155 | unsigned long size; |
4c887265 | 4156 | struct page *page; |
1e8f889b | 4157 | pte_t new_pte; |
cb900f41 | 4158 | spinlock_t *ptl; |
285b8dca | 4159 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 4160 | bool new_page = false; |
4c887265 | 4161 | |
04f2cbe3 MG |
4162 | /* |
4163 | * Currently, we are forced to kill the process in the event the | |
4164 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4165 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4166 | */ |
4167 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4168 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4169 | current->pid); |
04f2cbe3 MG |
4170 | return ret; |
4171 | } | |
4172 | ||
4c887265 | 4173 | /* |
87bf91d3 MK |
4174 | * We can not race with truncation due to holding i_mmap_rwsem. |
4175 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4176 | * once for faults beyond end of file. | |
4c887265 | 4177 | */ |
87bf91d3 MK |
4178 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4179 | if (idx >= size) | |
4180 | goto out; | |
4181 | ||
6bda666a CL |
4182 | retry: |
4183 | page = find_lock_page(mapping, idx); | |
4184 | if (!page) { | |
1a1aad8a MK |
4185 | /* |
4186 | * Check for page in userfault range | |
4187 | */ | |
4188 | if (userfaultfd_missing(vma)) { | |
4189 | u32 hash; | |
4190 | struct vm_fault vmf = { | |
4191 | .vma = vma, | |
285b8dca | 4192 | .address = haddr, |
1a1aad8a MK |
4193 | .flags = flags, |
4194 | /* | |
4195 | * Hard to debug if it ends up being | |
4196 | * used by a callee that assumes | |
4197 | * something about the other | |
4198 | * uninitialized fields... same as in | |
4199 | * memory.c | |
4200 | */ | |
4201 | }; | |
4202 | ||
4203 | /* | |
c0d0381a MK |
4204 | * hugetlb_fault_mutex and i_mmap_rwsem must be |
4205 | * dropped before handling userfault. Reacquire | |
4206 | * after handling fault to make calling code simpler. | |
1a1aad8a | 4207 | */ |
188b04a7 | 4208 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
1a1aad8a | 4209 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4210 | i_mmap_unlock_read(mapping); |
1a1aad8a | 4211 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); |
c0d0381a | 4212 | i_mmap_lock_read(mapping); |
1a1aad8a MK |
4213 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
4214 | goto out; | |
4215 | } | |
4216 | ||
285b8dca | 4217 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4218 | if (IS_ERR(page)) { |
4643d67e MK |
4219 | /* |
4220 | * Returning error will result in faulting task being | |
4221 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4222 | * tasks from racing to fault in the same page which | |
4223 | * could result in false unable to allocate errors. | |
4224 | * Page migration does not take the fault mutex, but | |
4225 | * does a clear then write of pte's under page table | |
4226 | * lock. Page fault code could race with migration, | |
4227 | * notice the clear pte and try to allocate a page | |
4228 | * here. Before returning error, get ptl and make | |
4229 | * sure there really is no pte entry. | |
4230 | */ | |
4231 | ptl = huge_pte_lock(h, mm, ptep); | |
4232 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
4233 | ret = 0; | |
4234 | spin_unlock(ptl); | |
4235 | goto out; | |
4236 | } | |
4237 | spin_unlock(ptl); | |
2b740303 | 4238 | ret = vmf_error(PTR_ERR(page)); |
6bda666a CL |
4239 | goto out; |
4240 | } | |
47ad8475 | 4241 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4242 | __SetPageUptodate(page); |
cb6acd01 | 4243 | new_page = true; |
ac9b9c66 | 4244 | |
f83a275d | 4245 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4246 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4247 | if (err) { |
4248 | put_page(page); | |
6bda666a CL |
4249 | if (err == -EEXIST) |
4250 | goto retry; | |
4251 | goto out; | |
4252 | } | |
23be7468 | 4253 | } else { |
6bda666a | 4254 | lock_page(page); |
0fe6e20b NH |
4255 | if (unlikely(anon_vma_prepare(vma))) { |
4256 | ret = VM_FAULT_OOM; | |
4257 | goto backout_unlocked; | |
4258 | } | |
409eb8c2 | 4259 | anon_rmap = 1; |
23be7468 | 4260 | } |
0fe6e20b | 4261 | } else { |
998b4382 NH |
4262 | /* |
4263 | * If memory error occurs between mmap() and fault, some process | |
4264 | * don't have hwpoisoned swap entry for errored virtual address. | |
4265 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4266 | */ | |
4267 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 4268 | ret = VM_FAULT_HWPOISON | |
972dc4de | 4269 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4270 | goto backout_unlocked; |
4271 | } | |
6bda666a | 4272 | } |
1e8f889b | 4273 | |
57303d80 AW |
4274 | /* |
4275 | * If we are going to COW a private mapping later, we examine the | |
4276 | * pending reservations for this page now. This will ensure that | |
4277 | * any allocations necessary to record that reservation occur outside | |
4278 | * the spinlock. | |
4279 | */ | |
5e911373 | 4280 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4281 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4282 | ret = VM_FAULT_OOM; |
4283 | goto backout_unlocked; | |
4284 | } | |
5e911373 | 4285 | /* Just decrements count, does not deallocate */ |
285b8dca | 4286 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4287 | } |
57303d80 | 4288 | |
8bea8052 | 4289 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 4290 | ret = 0; |
7f2e9525 | 4291 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4292 | goto backout; |
4293 | ||
07443a85 JK |
4294 | if (anon_rmap) { |
4295 | ClearPagePrivate(page); | |
285b8dca | 4296 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4297 | } else |
53f9263b | 4298 | page_dup_rmap(page, true); |
1e8f889b DG |
4299 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4300 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4301 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4302 | |
5d317b2b | 4303 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4304 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4305 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4306 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4307 | } |
4308 | ||
cb900f41 | 4309 | spin_unlock(ptl); |
cb6acd01 MK |
4310 | |
4311 | /* | |
4312 | * Only make newly allocated pages active. Existing pages found | |
4313 | * in the pagecache could be !page_huge_active() if they have been | |
4314 | * isolated for migration. | |
4315 | */ | |
4316 | if (new_page) | |
4317 | set_page_huge_active(page); | |
4318 | ||
4c887265 AL |
4319 | unlock_page(page); |
4320 | out: | |
ac9b9c66 | 4321 | return ret; |
4c887265 AL |
4322 | |
4323 | backout: | |
cb900f41 | 4324 | spin_unlock(ptl); |
2b26736c | 4325 | backout_unlocked: |
4c887265 | 4326 | unlock_page(page); |
285b8dca | 4327 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4328 | put_page(page); |
4329 | goto out; | |
ac9b9c66 HD |
4330 | } |
4331 | ||
8382d914 | 4332 | #ifdef CONFIG_SMP |
188b04a7 | 4333 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4334 | { |
4335 | unsigned long key[2]; | |
4336 | u32 hash; | |
4337 | ||
1b426bac MK |
4338 | key[0] = (unsigned long) mapping; |
4339 | key[1] = idx; | |
8382d914 | 4340 | |
55254636 | 4341 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4342 | |
4343 | return hash & (num_fault_mutexes - 1); | |
4344 | } | |
4345 | #else | |
4346 | /* | |
4347 | * For uniprocesor systems we always use a single mutex, so just | |
4348 | * return 0 and avoid the hashing overhead. | |
4349 | */ | |
188b04a7 | 4350 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4351 | { |
4352 | return 0; | |
4353 | } | |
4354 | #endif | |
4355 | ||
2b740303 | 4356 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4357 | unsigned long address, unsigned int flags) |
86e5216f | 4358 | { |
8382d914 | 4359 | pte_t *ptep, entry; |
cb900f41 | 4360 | spinlock_t *ptl; |
2b740303 | 4361 | vm_fault_t ret; |
8382d914 DB |
4362 | u32 hash; |
4363 | pgoff_t idx; | |
0fe6e20b | 4364 | struct page *page = NULL; |
57303d80 | 4365 | struct page *pagecache_page = NULL; |
a5516438 | 4366 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4367 | struct address_space *mapping; |
0f792cf9 | 4368 | int need_wait_lock = 0; |
285b8dca | 4369 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4370 | |
285b8dca | 4371 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 4372 | if (ptep) { |
c0d0381a MK |
4373 | /* |
4374 | * Since we hold no locks, ptep could be stale. That is | |
4375 | * OK as we are only making decisions based on content and | |
4376 | * not actually modifying content here. | |
4377 | */ | |
fd6a03ed | 4378 | entry = huge_ptep_get(ptep); |
290408d4 | 4379 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4380 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4381 | return 0; |
4382 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4383 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4384 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
ddeaab32 MK |
4385 | } else { |
4386 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4387 | if (!ptep) | |
4388 | return VM_FAULT_OOM; | |
fd6a03ed NH |
4389 | } |
4390 | ||
c0d0381a MK |
4391 | /* |
4392 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
4393 | * until finished with ptep. This serves two purposes: |
4394 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
4395 | * and making the ptep no longer valid. | |
4396 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
4397 | * |
4398 | * ptep could have already be assigned via huge_pte_offset. That | |
4399 | * is OK, as huge_pte_alloc will return the same value unless | |
4400 | * something has changed. | |
4401 | */ | |
8382d914 | 4402 | mapping = vma->vm_file->f_mapping; |
c0d0381a MK |
4403 | i_mmap_lock_read(mapping); |
4404 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4405 | if (!ptep) { | |
4406 | i_mmap_unlock_read(mapping); | |
4407 | return VM_FAULT_OOM; | |
4408 | } | |
8382d914 | 4409 | |
3935baa9 DG |
4410 | /* |
4411 | * Serialize hugepage allocation and instantiation, so that we don't | |
4412 | * get spurious allocation failures if two CPUs race to instantiate | |
4413 | * the same page in the page cache. | |
4414 | */ | |
c0d0381a | 4415 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 4416 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4417 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4418 | |
7f2e9525 GS |
4419 | entry = huge_ptep_get(ptep); |
4420 | if (huge_pte_none(entry)) { | |
8382d914 | 4421 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4422 | goto out_mutex; |
3935baa9 | 4423 | } |
86e5216f | 4424 | |
83c54070 | 4425 | ret = 0; |
1e8f889b | 4426 | |
0f792cf9 NH |
4427 | /* |
4428 | * entry could be a migration/hwpoison entry at this point, so this | |
4429 | * check prevents the kernel from going below assuming that we have | |
4430 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
4431 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
4432 | * handle it. | |
4433 | */ | |
4434 | if (!pte_present(entry)) | |
4435 | goto out_mutex; | |
4436 | ||
57303d80 AW |
4437 | /* |
4438 | * If we are going to COW the mapping later, we examine the pending | |
4439 | * reservations for this page now. This will ensure that any | |
4440 | * allocations necessary to record that reservation occur outside the | |
4441 | * spinlock. For private mappings, we also lookup the pagecache | |
4442 | * page now as it is used to determine if a reservation has been | |
4443 | * consumed. | |
4444 | */ | |
106c992a | 4445 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4446 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4447 | ret = VM_FAULT_OOM; |
b4d1d99f | 4448 | goto out_mutex; |
2b26736c | 4449 | } |
5e911373 | 4450 | /* Just decrements count, does not deallocate */ |
285b8dca | 4451 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4452 | |
f83a275d | 4453 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4454 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4455 | vma, haddr); |
57303d80 AW |
4456 | } |
4457 | ||
0f792cf9 NH |
4458 | ptl = huge_pte_lock(h, mm, ptep); |
4459 | ||
4460 | /* Check for a racing update before calling hugetlb_cow */ | |
4461 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4462 | goto out_ptl; | |
4463 | ||
56c9cfb1 NH |
4464 | /* |
4465 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4466 | * pagecache_page, so here we need take the former one | |
4467 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4468 | */ |
4469 | page = pte_page(entry); | |
4470 | if (page != pagecache_page) | |
0f792cf9 NH |
4471 | if (!trylock_page(page)) { |
4472 | need_wait_lock = 1; | |
4473 | goto out_ptl; | |
4474 | } | |
b4d1d99f | 4475 | |
0f792cf9 | 4476 | get_page(page); |
b4d1d99f | 4477 | |
788c7df4 | 4478 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4479 | if (!huge_pte_write(entry)) { |
974e6d66 | 4480 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4481 | pagecache_page, ptl); |
0f792cf9 | 4482 | goto out_put_page; |
b4d1d99f | 4483 | } |
106c992a | 4484 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4485 | } |
4486 | entry = pte_mkyoung(entry); | |
285b8dca | 4487 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4488 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4489 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4490 | out_put_page: |
4491 | if (page != pagecache_page) | |
4492 | unlock_page(page); | |
4493 | put_page(page); | |
cb900f41 KS |
4494 | out_ptl: |
4495 | spin_unlock(ptl); | |
57303d80 AW |
4496 | |
4497 | if (pagecache_page) { | |
4498 | unlock_page(pagecache_page); | |
4499 | put_page(pagecache_page); | |
4500 | } | |
b4d1d99f | 4501 | out_mutex: |
c672c7f2 | 4502 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4503 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
4504 | /* |
4505 | * Generally it's safe to hold refcount during waiting page lock. But | |
4506 | * here we just wait to defer the next page fault to avoid busy loop and | |
4507 | * the page is not used after unlocked before returning from the current | |
4508 | * page fault. So we are safe from accessing freed page, even if we wait | |
4509 | * here without taking refcount. | |
4510 | */ | |
4511 | if (need_wait_lock) | |
4512 | wait_on_page_locked(page); | |
1e8f889b | 4513 | return ret; |
86e5216f AL |
4514 | } |
4515 | ||
8fb5debc MK |
4516 | /* |
4517 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4518 | * modifications for huge pages. | |
4519 | */ | |
4520 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4521 | pte_t *dst_pte, | |
4522 | struct vm_area_struct *dst_vma, | |
4523 | unsigned long dst_addr, | |
4524 | unsigned long src_addr, | |
4525 | struct page **pagep) | |
4526 | { | |
1e392147 AA |
4527 | struct address_space *mapping; |
4528 | pgoff_t idx; | |
4529 | unsigned long size; | |
1c9e8def | 4530 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4531 | struct hstate *h = hstate_vma(dst_vma); |
4532 | pte_t _dst_pte; | |
4533 | spinlock_t *ptl; | |
4534 | int ret; | |
4535 | struct page *page; | |
4536 | ||
4537 | if (!*pagep) { | |
4538 | ret = -ENOMEM; | |
4539 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4540 | if (IS_ERR(page)) | |
4541 | goto out; | |
4542 | ||
4543 | ret = copy_huge_page_from_user(page, | |
4544 | (const void __user *) src_addr, | |
810a56b9 | 4545 | pages_per_huge_page(h), false); |
8fb5debc MK |
4546 | |
4547 | /* fallback to copy_from_user outside mmap_sem */ | |
4548 | if (unlikely(ret)) { | |
9e368259 | 4549 | ret = -ENOENT; |
8fb5debc MK |
4550 | *pagep = page; |
4551 | /* don't free the page */ | |
4552 | goto out; | |
4553 | } | |
4554 | } else { | |
4555 | page = *pagep; | |
4556 | *pagep = NULL; | |
4557 | } | |
4558 | ||
4559 | /* | |
4560 | * The memory barrier inside __SetPageUptodate makes sure that | |
4561 | * preceding stores to the page contents become visible before | |
4562 | * the set_pte_at() write. | |
4563 | */ | |
4564 | __SetPageUptodate(page); | |
8fb5debc | 4565 | |
1e392147 AA |
4566 | mapping = dst_vma->vm_file->f_mapping; |
4567 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4568 | ||
1c9e8def MK |
4569 | /* |
4570 | * If shared, add to page cache | |
4571 | */ | |
4572 | if (vm_shared) { | |
1e392147 AA |
4573 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4574 | ret = -EFAULT; | |
4575 | if (idx >= size) | |
4576 | goto out_release_nounlock; | |
1c9e8def | 4577 | |
1e392147 AA |
4578 | /* |
4579 | * Serialization between remove_inode_hugepages() and | |
4580 | * huge_add_to_page_cache() below happens through the | |
4581 | * hugetlb_fault_mutex_table that here must be hold by | |
4582 | * the caller. | |
4583 | */ | |
1c9e8def MK |
4584 | ret = huge_add_to_page_cache(page, mapping, idx); |
4585 | if (ret) | |
4586 | goto out_release_nounlock; | |
4587 | } | |
4588 | ||
8fb5debc MK |
4589 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4590 | spin_lock(ptl); | |
4591 | ||
1e392147 AA |
4592 | /* |
4593 | * Recheck the i_size after holding PT lock to make sure not | |
4594 | * to leave any page mapped (as page_mapped()) beyond the end | |
4595 | * of the i_size (remove_inode_hugepages() is strict about | |
4596 | * enforcing that). If we bail out here, we'll also leave a | |
4597 | * page in the radix tree in the vm_shared case beyond the end | |
4598 | * of the i_size, but remove_inode_hugepages() will take care | |
4599 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4600 | */ | |
4601 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4602 | ret = -EFAULT; | |
4603 | if (idx >= size) | |
4604 | goto out_release_unlock; | |
4605 | ||
8fb5debc MK |
4606 | ret = -EEXIST; |
4607 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4608 | goto out_release_unlock; | |
4609 | ||
1c9e8def MK |
4610 | if (vm_shared) { |
4611 | page_dup_rmap(page, true); | |
4612 | } else { | |
4613 | ClearPagePrivate(page); | |
4614 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4615 | } | |
8fb5debc MK |
4616 | |
4617 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4618 | if (dst_vma->vm_flags & VM_WRITE) | |
4619 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4620 | _dst_pte = pte_mkyoung(_dst_pte); | |
4621 | ||
4622 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4623 | ||
4624 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4625 | dst_vma->vm_flags & VM_WRITE); | |
4626 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4627 | ||
4628 | /* No need to invalidate - it was non-present before */ | |
4629 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4630 | ||
4631 | spin_unlock(ptl); | |
cb6acd01 | 4632 | set_page_huge_active(page); |
1c9e8def MK |
4633 | if (vm_shared) |
4634 | unlock_page(page); | |
8fb5debc MK |
4635 | ret = 0; |
4636 | out: | |
4637 | return ret; | |
4638 | out_release_unlock: | |
4639 | spin_unlock(ptl); | |
1c9e8def MK |
4640 | if (vm_shared) |
4641 | unlock_page(page); | |
5af10dfd | 4642 | out_release_nounlock: |
8fb5debc MK |
4643 | put_page(page); |
4644 | goto out; | |
4645 | } | |
4646 | ||
28a35716 ML |
4647 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4648 | struct page **pages, struct vm_area_struct **vmas, | |
4649 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 4650 | long i, unsigned int flags, int *locked) |
63551ae0 | 4651 | { |
d5d4b0aa KC |
4652 | unsigned long pfn_offset; |
4653 | unsigned long vaddr = *position; | |
28a35716 | 4654 | unsigned long remainder = *nr_pages; |
a5516438 | 4655 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4656 | int err = -EFAULT; |
63551ae0 | 4657 | |
63551ae0 | 4658 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4659 | pte_t *pte; |
cb900f41 | 4660 | spinlock_t *ptl = NULL; |
2a15efc9 | 4661 | int absent; |
4c887265 | 4662 | struct page *page; |
63551ae0 | 4663 | |
02057967 DR |
4664 | /* |
4665 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4666 | * potentially allocating memory. | |
4667 | */ | |
fa45f116 | 4668 | if (fatal_signal_pending(current)) { |
02057967 DR |
4669 | remainder = 0; |
4670 | break; | |
4671 | } | |
4672 | ||
4c887265 AL |
4673 | /* |
4674 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4675 | * each hugepage. We have to make sure we get the |
4c887265 | 4676 | * first, for the page indexing below to work. |
cb900f41 KS |
4677 | * |
4678 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4679 | */ |
7868a208 PA |
4680 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4681 | huge_page_size(h)); | |
cb900f41 KS |
4682 | if (pte) |
4683 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4684 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4685 | ||
4686 | /* | |
4687 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4688 | * an error where there's an empty slot with no huge pagecache |
4689 | * to back it. This way, we avoid allocating a hugepage, and | |
4690 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4691 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4692 | */ |
3ae77f43 HD |
4693 | if (absent && (flags & FOLL_DUMP) && |
4694 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4695 | if (pte) |
4696 | spin_unlock(ptl); | |
2a15efc9 HD |
4697 | remainder = 0; |
4698 | break; | |
4699 | } | |
63551ae0 | 4700 | |
9cc3a5bd NH |
4701 | /* |
4702 | * We need call hugetlb_fault for both hugepages under migration | |
4703 | * (in which case hugetlb_fault waits for the migration,) and | |
4704 | * hwpoisoned hugepages (in which case we need to prevent the | |
4705 | * caller from accessing to them.) In order to do this, we use | |
4706 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4707 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4708 | * both cases, and because we can't follow correct pages | |
4709 | * directly from any kind of swap entries. | |
4710 | */ | |
4711 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4712 | ((flags & FOLL_WRITE) && |
4713 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4714 | vm_fault_t ret; |
87ffc118 | 4715 | unsigned int fault_flags = 0; |
63551ae0 | 4716 | |
cb900f41 KS |
4717 | if (pte) |
4718 | spin_unlock(ptl); | |
87ffc118 AA |
4719 | if (flags & FOLL_WRITE) |
4720 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 4721 | if (locked) |
71335f37 PX |
4722 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
4723 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
4724 | if (flags & FOLL_NOWAIT) |
4725 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4726 | FAULT_FLAG_RETRY_NOWAIT; | |
4727 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
4728 | /* |
4729 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
4730 | * FAULT_FLAG_TRIED can co-exist | |
4731 | */ | |
87ffc118 AA |
4732 | fault_flags |= FAULT_FLAG_TRIED; |
4733 | } | |
4734 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4735 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4736 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4737 | remainder = 0; |
4738 | break; | |
4739 | } | |
4740 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 4741 | if (locked && |
1ac25013 | 4742 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 4743 | *locked = 0; |
87ffc118 AA |
4744 | *nr_pages = 0; |
4745 | /* | |
4746 | * VM_FAULT_RETRY must not return an | |
4747 | * error, it will return zero | |
4748 | * instead. | |
4749 | * | |
4750 | * No need to update "position" as the | |
4751 | * caller will not check it after | |
4752 | * *nr_pages is set to 0. | |
4753 | */ | |
4754 | return i; | |
4755 | } | |
4756 | continue; | |
4c887265 AL |
4757 | } |
4758 | ||
a5516438 | 4759 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4760 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 4761 | |
acbfb087 ZL |
4762 | /* |
4763 | * If subpage information not requested, update counters | |
4764 | * and skip the same_page loop below. | |
4765 | */ | |
4766 | if (!pages && !vmas && !pfn_offset && | |
4767 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4768 | (remainder >= pages_per_huge_page(h))) { | |
4769 | vaddr += huge_page_size(h); | |
4770 | remainder -= pages_per_huge_page(h); | |
4771 | i += pages_per_huge_page(h); | |
4772 | spin_unlock(ptl); | |
4773 | continue; | |
4774 | } | |
4775 | ||
d5d4b0aa | 4776 | same_page: |
d6692183 | 4777 | if (pages) { |
2a15efc9 | 4778 | pages[i] = mem_map_offset(page, pfn_offset); |
3faa52c0 JH |
4779 | /* |
4780 | * try_grab_page() should always succeed here, because: | |
4781 | * a) we hold the ptl lock, and b) we've just checked | |
4782 | * that the huge page is present in the page tables. If | |
4783 | * the huge page is present, then the tail pages must | |
4784 | * also be present. The ptl prevents the head page and | |
4785 | * tail pages from being rearranged in any way. So this | |
4786 | * page must be available at this point, unless the page | |
4787 | * refcount overflowed: | |
4788 | */ | |
4789 | if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { | |
4790 | spin_unlock(ptl); | |
4791 | remainder = 0; | |
4792 | err = -ENOMEM; | |
4793 | break; | |
4794 | } | |
d6692183 | 4795 | } |
63551ae0 DG |
4796 | |
4797 | if (vmas) | |
4798 | vmas[i] = vma; | |
4799 | ||
4800 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4801 | ++pfn_offset; |
63551ae0 DG |
4802 | --remainder; |
4803 | ++i; | |
d5d4b0aa | 4804 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4805 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4806 | /* |
4807 | * We use pfn_offset to avoid touching the pageframes | |
4808 | * of this compound page. | |
4809 | */ | |
4810 | goto same_page; | |
4811 | } | |
cb900f41 | 4812 | spin_unlock(ptl); |
63551ae0 | 4813 | } |
28a35716 | 4814 | *nr_pages = remainder; |
87ffc118 AA |
4815 | /* |
4816 | * setting position is actually required only if remainder is | |
4817 | * not zero but it's faster not to add a "if (remainder)" | |
4818 | * branch. | |
4819 | */ | |
63551ae0 DG |
4820 | *position = vaddr; |
4821 | ||
2be7cfed | 4822 | return i ? i : err; |
63551ae0 | 4823 | } |
8f860591 | 4824 | |
5491ae7b AK |
4825 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4826 | /* | |
4827 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4828 | * implement this. | |
4829 | */ | |
4830 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4831 | #endif | |
4832 | ||
7da4d641 | 4833 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4834 | unsigned long address, unsigned long end, pgprot_t newprot) |
4835 | { | |
4836 | struct mm_struct *mm = vma->vm_mm; | |
4837 | unsigned long start = address; | |
4838 | pte_t *ptep; | |
4839 | pte_t pte; | |
a5516438 | 4840 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4841 | unsigned long pages = 0; |
dff11abe | 4842 | bool shared_pmd = false; |
ac46d4f3 | 4843 | struct mmu_notifier_range range; |
dff11abe MK |
4844 | |
4845 | /* | |
4846 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 4847 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
4848 | * range if PMD sharing is possible. |
4849 | */ | |
7269f999 JG |
4850 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
4851 | 0, vma, mm, start, end); | |
ac46d4f3 | 4852 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
4853 | |
4854 | BUG_ON(address >= end); | |
ac46d4f3 | 4855 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 4856 | |
ac46d4f3 | 4857 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 4858 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4859 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4860 | spinlock_t *ptl; |
7868a208 | 4861 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4862 | if (!ptep) |
4863 | continue; | |
cb900f41 | 4864 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4865 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4866 | pages++; | |
cb900f41 | 4867 | spin_unlock(ptl); |
dff11abe | 4868 | shared_pmd = true; |
39dde65c | 4869 | continue; |
7da4d641 | 4870 | } |
a8bda28d NH |
4871 | pte = huge_ptep_get(ptep); |
4872 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4873 | spin_unlock(ptl); | |
4874 | continue; | |
4875 | } | |
4876 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4877 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4878 | ||
4879 | if (is_write_migration_entry(entry)) { | |
4880 | pte_t newpte; | |
4881 | ||
4882 | make_migration_entry_read(&entry); | |
4883 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4884 | set_huge_swap_pte_at(mm, address, ptep, |
4885 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4886 | pages++; |
4887 | } | |
4888 | spin_unlock(ptl); | |
4889 | continue; | |
4890 | } | |
4891 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
4892 | pte_t old_pte; |
4893 | ||
4894 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
4895 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 4896 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 4897 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 4898 | pages++; |
8f860591 | 4899 | } |
cb900f41 | 4900 | spin_unlock(ptl); |
8f860591 | 4901 | } |
d833352a | 4902 | /* |
c8c06efa | 4903 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4904 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4905 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
4906 | * and that page table be reused and filled with junk. If we actually |
4907 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 4908 | */ |
dff11abe | 4909 | if (shared_pmd) |
ac46d4f3 | 4910 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
4911 | else |
4912 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
4913 | /* |
4914 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4915 | * page table protection not changing it to point to a new page. | |
4916 | * | |
ad56b738 | 4917 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4918 | */ |
83cde9e8 | 4919 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 4920 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
4921 | |
4922 | return pages << h->order; | |
8f860591 ZY |
4923 | } |
4924 | ||
a1e78772 MG |
4925 | int hugetlb_reserve_pages(struct inode *inode, |
4926 | long from, long to, | |
5a6fe125 | 4927 | struct vm_area_struct *vma, |
ca16d140 | 4928 | vm_flags_t vm_flags) |
e4e574b7 | 4929 | { |
0db9d74e | 4930 | long ret, chg, add = -1; |
a5516438 | 4931 | struct hstate *h = hstate_inode(inode); |
90481622 | 4932 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4933 | struct resv_map *resv_map; |
075a61d0 | 4934 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 4935 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 4936 | |
63489f8e MK |
4937 | /* This should never happen */ |
4938 | if (from > to) { | |
4939 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
4940 | return -EINVAL; | |
4941 | } | |
4942 | ||
17c9d12e MG |
4943 | /* |
4944 | * Only apply hugepage reservation if asked. At fault time, an | |
4945 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4946 | * without using reserves |
17c9d12e | 4947 | */ |
ca16d140 | 4948 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4949 | return 0; |
4950 | ||
a1e78772 MG |
4951 | /* |
4952 | * Shared mappings base their reservation on the number of pages that | |
4953 | * are already allocated on behalf of the file. Private mappings need | |
4954 | * to reserve the full area even if read-only as mprotect() may be | |
4955 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4956 | */ | |
9119a41e | 4957 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
4958 | /* |
4959 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
4960 | * called for inodes for which resv_maps were created (see | |
4961 | * hugetlbfs_get_inode). | |
4962 | */ | |
4e35f483 | 4963 | resv_map = inode_resv_map(inode); |
9119a41e | 4964 | |
0db9d74e | 4965 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
4966 | |
4967 | } else { | |
e9fe92ae | 4968 | /* Private mapping. */ |
9119a41e | 4969 | resv_map = resv_map_alloc(); |
17c9d12e MG |
4970 | if (!resv_map) |
4971 | return -ENOMEM; | |
4972 | ||
a1e78772 | 4973 | chg = to - from; |
84afd99b | 4974 | |
17c9d12e MG |
4975 | set_vma_resv_map(vma, resv_map); |
4976 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
4977 | } | |
4978 | ||
c50ac050 DH |
4979 | if (chg < 0) { |
4980 | ret = chg; | |
4981 | goto out_err; | |
4982 | } | |
8a630112 | 4983 | |
075a61d0 MA |
4984 | ret = hugetlb_cgroup_charge_cgroup_rsvd( |
4985 | hstate_index(h), chg * pages_per_huge_page(h), &h_cg); | |
4986 | ||
4987 | if (ret < 0) { | |
4988 | ret = -ENOMEM; | |
4989 | goto out_err; | |
4990 | } | |
4991 | ||
4992 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
4993 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
4994 | * of the resv_map. | |
4995 | */ | |
4996 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
4997 | } | |
4998 | ||
1c5ecae3 MK |
4999 | /* |
5000 | * There must be enough pages in the subpool for the mapping. If | |
5001 | * the subpool has a minimum size, there may be some global | |
5002 | * reservations already in place (gbl_reserve). | |
5003 | */ | |
5004 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
5005 | if (gbl_reserve < 0) { | |
c50ac050 | 5006 | ret = -ENOSPC; |
075a61d0 | 5007 | goto out_uncharge_cgroup; |
c50ac050 | 5008 | } |
5a6fe125 MG |
5009 | |
5010 | /* | |
17c9d12e | 5011 | * Check enough hugepages are available for the reservation. |
90481622 | 5012 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 5013 | */ |
1c5ecae3 | 5014 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 5015 | if (ret < 0) { |
075a61d0 | 5016 | goto out_put_pages; |
68842c9b | 5017 | } |
17c9d12e MG |
5018 | |
5019 | /* | |
5020 | * Account for the reservations made. Shared mappings record regions | |
5021 | * that have reservations as they are shared by multiple VMAs. | |
5022 | * When the last VMA disappears, the region map says how much | |
5023 | * the reservation was and the page cache tells how much of | |
5024 | * the reservation was consumed. Private mappings are per-VMA and | |
5025 | * only the consumed reservations are tracked. When the VMA | |
5026 | * disappears, the original reservation is the VMA size and the | |
5027 | * consumed reservations are stored in the map. Hence, nothing | |
5028 | * else has to be done for private mappings here | |
5029 | */ | |
33039678 | 5030 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 5031 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
5032 | |
5033 | if (unlikely(add < 0)) { | |
5034 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 5035 | goto out_put_pages; |
0db9d74e | 5036 | } else if (unlikely(chg > add)) { |
33039678 MK |
5037 | /* |
5038 | * pages in this range were added to the reserve | |
5039 | * map between region_chg and region_add. This | |
5040 | * indicates a race with alloc_huge_page. Adjust | |
5041 | * the subpool and reserve counts modified above | |
5042 | * based on the difference. | |
5043 | */ | |
5044 | long rsv_adjust; | |
5045 | ||
075a61d0 MA |
5046 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
5047 | hstate_index(h), | |
5048 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5049 | ||
33039678 MK |
5050 | rsv_adjust = hugepage_subpool_put_pages(spool, |
5051 | chg - add); | |
5052 | hugetlb_acct_memory(h, -rsv_adjust); | |
5053 | } | |
5054 | } | |
a43a8c39 | 5055 | return 0; |
075a61d0 MA |
5056 | out_put_pages: |
5057 | /* put back original number of pages, chg */ | |
5058 | (void)hugepage_subpool_put_pages(spool, chg); | |
5059 | out_uncharge_cgroup: | |
5060 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5061 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 5062 | out_err: |
5e911373 | 5063 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
5064 | /* Only call region_abort if the region_chg succeeded but the |
5065 | * region_add failed or didn't run. | |
5066 | */ | |
5067 | if (chg >= 0 && add < 0) | |
5068 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
5069 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
5070 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 5071 | return ret; |
a43a8c39 KC |
5072 | } |
5073 | ||
b5cec28d MK |
5074 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
5075 | long freed) | |
a43a8c39 | 5076 | { |
a5516438 | 5077 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 5078 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 5079 | long chg = 0; |
90481622 | 5080 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 5081 | long gbl_reserve; |
45c682a6 | 5082 | |
f27a5136 MK |
5083 | /* |
5084 | * Since this routine can be called in the evict inode path for all | |
5085 | * hugetlbfs inodes, resv_map could be NULL. | |
5086 | */ | |
b5cec28d MK |
5087 | if (resv_map) { |
5088 | chg = region_del(resv_map, start, end); | |
5089 | /* | |
5090 | * region_del() can fail in the rare case where a region | |
5091 | * must be split and another region descriptor can not be | |
5092 | * allocated. If end == LONG_MAX, it will not fail. | |
5093 | */ | |
5094 | if (chg < 0) | |
5095 | return chg; | |
5096 | } | |
5097 | ||
45c682a6 | 5098 | spin_lock(&inode->i_lock); |
e4c6f8be | 5099 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
5100 | spin_unlock(&inode->i_lock); |
5101 | ||
1c5ecae3 MK |
5102 | /* |
5103 | * If the subpool has a minimum size, the number of global | |
5104 | * reservations to be released may be adjusted. | |
5105 | */ | |
5106 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5107 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
5108 | |
5109 | return 0; | |
a43a8c39 | 5110 | } |
93f70f90 | 5111 | |
3212b535 SC |
5112 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
5113 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5114 | struct vm_area_struct *vma, | |
5115 | unsigned long addr, pgoff_t idx) | |
5116 | { | |
5117 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5118 | svma->vm_start; | |
5119 | unsigned long sbase = saddr & PUD_MASK; | |
5120 | unsigned long s_end = sbase + PUD_SIZE; | |
5121 | ||
5122 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
5123 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
5124 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
5125 | |
5126 | /* | |
5127 | * match the virtual addresses, permission and the alignment of the | |
5128 | * page table page. | |
5129 | */ | |
5130 | if (pmd_index(addr) != pmd_index(saddr) || | |
5131 | vm_flags != svm_flags || | |
5132 | sbase < svma->vm_start || svma->vm_end < s_end) | |
5133 | return 0; | |
5134 | ||
5135 | return saddr; | |
5136 | } | |
5137 | ||
31aafb45 | 5138 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
5139 | { |
5140 | unsigned long base = addr & PUD_MASK; | |
5141 | unsigned long end = base + PUD_SIZE; | |
5142 | ||
5143 | /* | |
5144 | * check on proper vm_flags and page table alignment | |
5145 | */ | |
017b1660 | 5146 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
5147 | return true; |
5148 | return false; | |
3212b535 SC |
5149 | } |
5150 | ||
017b1660 MK |
5151 | /* |
5152 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5153 | * If yes, adjust start and end to cover range associated with possible | |
5154 | * shared pmd mappings. | |
5155 | */ | |
5156 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5157 | unsigned long *start, unsigned long *end) | |
5158 | { | |
353b2de4 | 5159 | unsigned long check_addr; |
017b1660 MK |
5160 | |
5161 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
5162 | return; | |
5163 | ||
5164 | for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { | |
5165 | unsigned long a_start = check_addr & PUD_MASK; | |
5166 | unsigned long a_end = a_start + PUD_SIZE; | |
5167 | ||
5168 | /* | |
5169 | * If sharing is possible, adjust start/end if necessary. | |
5170 | */ | |
5171 | if (range_in_vma(vma, a_start, a_end)) { | |
5172 | if (a_start < *start) | |
5173 | *start = a_start; | |
5174 | if (a_end > *end) | |
5175 | *end = a_end; | |
5176 | } | |
5177 | } | |
5178 | } | |
5179 | ||
3212b535 SC |
5180 | /* |
5181 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5182 | * and returns the corresponding pte. While this is not necessary for the | |
5183 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
5184 | * code much cleaner. |
5185 | * | |
5186 | * This routine must be called with i_mmap_rwsem held in at least read mode. | |
5187 | * For hugetlbfs, this prevents removal of any page table entries associated | |
5188 | * with the address space. This is important as we are setting up sharing | |
5189 | * based on existing page table entries (mappings). | |
3212b535 SC |
5190 | */ |
5191 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5192 | { | |
5193 | struct vm_area_struct *vma = find_vma(mm, addr); | |
5194 | struct address_space *mapping = vma->vm_file->f_mapping; | |
5195 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5196 | vma->vm_pgoff; | |
5197 | struct vm_area_struct *svma; | |
5198 | unsigned long saddr; | |
5199 | pte_t *spte = NULL; | |
5200 | pte_t *pte; | |
cb900f41 | 5201 | spinlock_t *ptl; |
3212b535 SC |
5202 | |
5203 | if (!vma_shareable(vma, addr)) | |
5204 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
5205 | ||
3212b535 SC |
5206 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
5207 | if (svma == vma) | |
5208 | continue; | |
5209 | ||
5210 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5211 | if (saddr) { | |
7868a208 PA |
5212 | spte = huge_pte_offset(svma->vm_mm, saddr, |
5213 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
5214 | if (spte) { |
5215 | get_page(virt_to_page(spte)); | |
5216 | break; | |
5217 | } | |
5218 | } | |
5219 | } | |
5220 | ||
5221 | if (!spte) | |
5222 | goto out; | |
5223 | ||
8bea8052 | 5224 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 5225 | if (pud_none(*pud)) { |
3212b535 SC |
5226 | pud_populate(mm, pud, |
5227 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5228 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5229 | } else { |
3212b535 | 5230 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5231 | } |
cb900f41 | 5232 | spin_unlock(ptl); |
3212b535 SC |
5233 | out: |
5234 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
5235 | return pte; |
5236 | } | |
5237 | ||
5238 | /* | |
5239 | * unmap huge page backed by shared pte. | |
5240 | * | |
5241 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5242 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5243 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5244 | * | |
c0d0381a | 5245 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
5246 | * |
5247 | * returns: 1 successfully unmapped a shared pte page | |
5248 | * 0 the underlying pte page is not shared, or it is the last user | |
5249 | */ | |
5250 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
5251 | { | |
5252 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
5253 | p4d_t *p4d = p4d_offset(pgd, *addr); |
5254 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
5255 | |
5256 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
5257 | if (page_count(virt_to_page(ptep)) == 1) | |
5258 | return 0; | |
5259 | ||
5260 | pud_clear(pud); | |
5261 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 5262 | mm_dec_nr_pmds(mm); |
3212b535 SC |
5263 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
5264 | return 1; | |
5265 | } | |
9e5fc74c SC |
5266 | #define want_pmd_share() (1) |
5267 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
5268 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5269 | { | |
5270 | return NULL; | |
5271 | } | |
e81f2d22 ZZ |
5272 | |
5273 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
5274 | { | |
5275 | return 0; | |
5276 | } | |
017b1660 MK |
5277 | |
5278 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5279 | unsigned long *start, unsigned long *end) | |
5280 | { | |
5281 | } | |
9e5fc74c | 5282 | #define want_pmd_share() (0) |
3212b535 SC |
5283 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5284 | ||
9e5fc74c SC |
5285 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
5286 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
5287 | unsigned long addr, unsigned long sz) | |
5288 | { | |
5289 | pgd_t *pgd; | |
c2febafc | 5290 | p4d_t *p4d; |
9e5fc74c SC |
5291 | pud_t *pud; |
5292 | pte_t *pte = NULL; | |
5293 | ||
5294 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5295 | p4d = p4d_alloc(mm, pgd, addr); |
5296 | if (!p4d) | |
5297 | return NULL; | |
c2febafc | 5298 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5299 | if (pud) { |
5300 | if (sz == PUD_SIZE) { | |
5301 | pte = (pte_t *)pud; | |
5302 | } else { | |
5303 | BUG_ON(sz != PMD_SIZE); | |
5304 | if (want_pmd_share() && pud_none(*pud)) | |
5305 | pte = huge_pmd_share(mm, addr, pud); | |
5306 | else | |
5307 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5308 | } | |
5309 | } | |
4e666314 | 5310 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5311 | |
5312 | return pte; | |
5313 | } | |
5314 | ||
9b19df29 PA |
5315 | /* |
5316 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5317 | * entry at address @addr | |
5318 | * | |
5319 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
5320 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
5321 | * size @sz doesn't match the hugepage size at this level of the page | |
5322 | * table. | |
5323 | */ | |
7868a208 PA |
5324 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5325 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5326 | { |
5327 | pgd_t *pgd; | |
c2febafc | 5328 | p4d_t *p4d; |
9e5fc74c | 5329 | pud_t *pud; |
c2febafc | 5330 | pmd_t *pmd; |
9e5fc74c SC |
5331 | |
5332 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5333 | if (!pgd_present(*pgd)) |
5334 | return NULL; | |
5335 | p4d = p4d_offset(pgd, addr); | |
5336 | if (!p4d_present(*p4d)) | |
5337 | return NULL; | |
9b19df29 | 5338 | |
c2febafc | 5339 | pud = pud_offset(p4d, addr); |
9b19df29 | 5340 | if (sz != PUD_SIZE && pud_none(*pud)) |
c2febafc | 5341 | return NULL; |
9b19df29 PA |
5342 | /* hugepage or swap? */ |
5343 | if (pud_huge(*pud) || !pud_present(*pud)) | |
c2febafc | 5344 | return (pte_t *)pud; |
9b19df29 | 5345 | |
c2febafc | 5346 | pmd = pmd_offset(pud, addr); |
9b19df29 PA |
5347 | if (sz != PMD_SIZE && pmd_none(*pmd)) |
5348 | return NULL; | |
5349 | /* hugepage or swap? */ | |
5350 | if (pmd_huge(*pmd) || !pmd_present(*pmd)) | |
5351 | return (pte_t *)pmd; | |
5352 | ||
5353 | return NULL; | |
9e5fc74c SC |
5354 | } |
5355 | ||
61f77eda NH |
5356 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5357 | ||
5358 | /* | |
5359 | * These functions are overwritable if your architecture needs its own | |
5360 | * behavior. | |
5361 | */ | |
5362 | struct page * __weak | |
5363 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5364 | int write) | |
5365 | { | |
5366 | return ERR_PTR(-EINVAL); | |
5367 | } | |
5368 | ||
4dc71451 AK |
5369 | struct page * __weak |
5370 | follow_huge_pd(struct vm_area_struct *vma, | |
5371 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5372 | { | |
5373 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5374 | return NULL; | |
5375 | } | |
5376 | ||
61f77eda | 5377 | struct page * __weak |
9e5fc74c | 5378 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5379 | pmd_t *pmd, int flags) |
9e5fc74c | 5380 | { |
e66f17ff NH |
5381 | struct page *page = NULL; |
5382 | spinlock_t *ptl; | |
c9d398fa | 5383 | pte_t pte; |
3faa52c0 JH |
5384 | |
5385 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5386 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5387 | (FOLL_PIN | FOLL_GET))) | |
5388 | return NULL; | |
5389 | ||
e66f17ff NH |
5390 | retry: |
5391 | ptl = pmd_lockptr(mm, pmd); | |
5392 | spin_lock(ptl); | |
5393 | /* | |
5394 | * make sure that the address range covered by this pmd is not | |
5395 | * unmapped from other threads. | |
5396 | */ | |
5397 | if (!pmd_huge(*pmd)) | |
5398 | goto out; | |
c9d398fa NH |
5399 | pte = huge_ptep_get((pte_t *)pmd); |
5400 | if (pte_present(pte)) { | |
97534127 | 5401 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
5402 | /* |
5403 | * try_grab_page() should always succeed here, because: a) we | |
5404 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5405 | * huge pmd (head) page is present in the page tables. The ptl | |
5406 | * prevents the head page and tail pages from being rearranged | |
5407 | * in any way. So this page must be available at this point, | |
5408 | * unless the page refcount overflowed: | |
5409 | */ | |
5410 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5411 | page = NULL; | |
5412 | goto out; | |
5413 | } | |
e66f17ff | 5414 | } else { |
c9d398fa | 5415 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5416 | spin_unlock(ptl); |
5417 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5418 | goto retry; | |
5419 | } | |
5420 | /* | |
5421 | * hwpoisoned entry is treated as no_page_table in | |
5422 | * follow_page_mask(). | |
5423 | */ | |
5424 | } | |
5425 | out: | |
5426 | spin_unlock(ptl); | |
9e5fc74c SC |
5427 | return page; |
5428 | } | |
5429 | ||
61f77eda | 5430 | struct page * __weak |
9e5fc74c | 5431 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5432 | pud_t *pud, int flags) |
9e5fc74c | 5433 | { |
3faa52c0 | 5434 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5435 | return NULL; |
9e5fc74c | 5436 | |
e66f17ff | 5437 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5438 | } |
5439 | ||
faaa5b62 AK |
5440 | struct page * __weak |
5441 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5442 | { | |
3faa52c0 | 5443 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5444 | return NULL; |
5445 | ||
5446 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5447 | } | |
5448 | ||
31caf665 NH |
5449 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5450 | { | |
bcc54222 NH |
5451 | bool ret = true; |
5452 | ||
309381fe | 5453 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5454 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
5455 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
5456 | ret = false; | |
5457 | goto unlock; | |
5458 | } | |
5459 | clear_page_huge_active(page); | |
31caf665 | 5460 | list_move_tail(&page->lru, list); |
bcc54222 | 5461 | unlock: |
31caf665 | 5462 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5463 | return ret; |
31caf665 NH |
5464 | } |
5465 | ||
5466 | void putback_active_hugepage(struct page *page) | |
5467 | { | |
309381fe | 5468 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5469 | spin_lock(&hugetlb_lock); |
bcc54222 | 5470 | set_page_huge_active(page); |
31caf665 NH |
5471 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5472 | spin_unlock(&hugetlb_lock); | |
5473 | put_page(page); | |
5474 | } | |
ab5ac90a MH |
5475 | |
5476 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5477 | { | |
5478 | struct hstate *h = page_hstate(oldpage); | |
5479 | ||
5480 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5481 | set_page_owner_migrate_reason(newpage, reason); | |
5482 | ||
5483 | /* | |
5484 | * transfer temporary state of the new huge page. This is | |
5485 | * reverse to other transitions because the newpage is going to | |
5486 | * be final while the old one will be freed so it takes over | |
5487 | * the temporary status. | |
5488 | * | |
5489 | * Also note that we have to transfer the per-node surplus state | |
5490 | * here as well otherwise the global surplus count will not match | |
5491 | * the per-node's. | |
5492 | */ | |
5493 | if (PageHugeTemporary(newpage)) { | |
5494 | int old_nid = page_to_nid(oldpage); | |
5495 | int new_nid = page_to_nid(newpage); | |
5496 | ||
5497 | SetPageHugeTemporary(oldpage); | |
5498 | ClearPageHugeTemporary(newpage); | |
5499 | ||
5500 | spin_lock(&hugetlb_lock); | |
5501 | if (h->surplus_huge_pages_node[old_nid]) { | |
5502 | h->surplus_huge_pages_node[old_nid]--; | |
5503 | h->surplus_huge_pages_node[new_nid]++; | |
5504 | } | |
5505 | spin_unlock(&hugetlb_lock); | |
5506 | } | |
5507 | } |