]>
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> |
cf11e85f | 31 | #include <linux/cma.h> |
d6606683 | 32 | |
63551ae0 DG |
33 | #include <asm/page.h> |
34 | #include <asm/pgtable.h> | |
24669e58 | 35 | #include <asm/tlb.h> |
63551ae0 | 36 | |
24669e58 | 37 | #include <linux/io.h> |
63551ae0 | 38 | #include <linux/hugetlb.h> |
9dd540e2 | 39 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 40 | #include <linux/node.h> |
1a1aad8a | 41 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 42 | #include <linux/page_owner.h> |
7835e98b | 43 | #include "internal.h" |
1da177e4 | 44 | |
c3f38a38 | 45 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
46 | unsigned int default_hstate_idx; |
47 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f RG |
48 | |
49 | static struct cma *hugetlb_cma[MAX_NUMNODES]; | |
50 | ||
641844f5 NH |
51 | /* |
52 | * Minimum page order among possible hugepage sizes, set to a proper value | |
53 | * at boot time. | |
54 | */ | |
55 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 56 | |
53ba51d2 JT |
57 | __initdata LIST_HEAD(huge_boot_pages); |
58 | ||
e5ff2159 AK |
59 | /* for command line parsing */ |
60 | static struct hstate * __initdata parsed_hstate; | |
61 | static unsigned long __initdata default_hstate_max_huge_pages; | |
e11bfbfc | 62 | static unsigned long __initdata default_hstate_size; |
9fee021d | 63 | static bool __initdata parsed_valid_hugepagesz = true; |
e5ff2159 | 64 | |
3935baa9 | 65 | /* |
31caf665 NH |
66 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
67 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 68 | */ |
c3f38a38 | 69 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 70 | |
8382d914 DB |
71 | /* |
72 | * Serializes faults on the same logical page. This is used to | |
73 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
74 | */ | |
75 | static int num_fault_mutexes; | |
c672c7f2 | 76 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 77 | |
7ca02d0a MK |
78 | /* Forward declaration */ |
79 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
80 | ||
90481622 DG |
81 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool) |
82 | { | |
83 | bool free = (spool->count == 0) && (spool->used_hpages == 0); | |
84 | ||
85 | spin_unlock(&spool->lock); | |
86 | ||
87 | /* If no pages are used, and no other handles to the subpool | |
7ca02d0a MK |
88 | * remain, give up any reservations mased on minimum size and |
89 | * free the subpool */ | |
90 | if (free) { | |
91 | if (spool->min_hpages != -1) | |
92 | hugetlb_acct_memory(spool->hstate, | |
93 | -spool->min_hpages); | |
90481622 | 94 | kfree(spool); |
7ca02d0a | 95 | } |
90481622 DG |
96 | } |
97 | ||
7ca02d0a MK |
98 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
99 | long min_hpages) | |
90481622 DG |
100 | { |
101 | struct hugepage_subpool *spool; | |
102 | ||
c6a91820 | 103 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
104 | if (!spool) |
105 | return NULL; | |
106 | ||
107 | spin_lock_init(&spool->lock); | |
108 | spool->count = 1; | |
7ca02d0a MK |
109 | spool->max_hpages = max_hpages; |
110 | spool->hstate = h; | |
111 | spool->min_hpages = min_hpages; | |
112 | ||
113 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
114 | kfree(spool); | |
115 | return NULL; | |
116 | } | |
117 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
118 | |
119 | return spool; | |
120 | } | |
121 | ||
122 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
123 | { | |
124 | spin_lock(&spool->lock); | |
125 | BUG_ON(!spool->count); | |
126 | spool->count--; | |
127 | unlock_or_release_subpool(spool); | |
128 | } | |
129 | ||
1c5ecae3 MK |
130 | /* |
131 | * Subpool accounting for allocating and reserving pages. | |
132 | * Return -ENOMEM if there are not enough resources to satisfy the | |
133 | * the request. Otherwise, return the number of pages by which the | |
134 | * global pools must be adjusted (upward). The returned value may | |
135 | * only be different than the passed value (delta) in the case where | |
136 | * a subpool minimum size must be manitained. | |
137 | */ | |
138 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
139 | long delta) |
140 | { | |
1c5ecae3 | 141 | long ret = delta; |
90481622 DG |
142 | |
143 | if (!spool) | |
1c5ecae3 | 144 | return ret; |
90481622 DG |
145 | |
146 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
147 | |
148 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
149 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
150 | spool->used_hpages += delta; | |
151 | else { | |
152 | ret = -ENOMEM; | |
153 | goto unlock_ret; | |
154 | } | |
90481622 | 155 | } |
90481622 | 156 | |
09a95e29 MK |
157 | /* minimum size accounting */ |
158 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
159 | if (delta > spool->rsv_hpages) { |
160 | /* | |
161 | * Asking for more reserves than those already taken on | |
162 | * behalf of subpool. Return difference. | |
163 | */ | |
164 | ret = delta - spool->rsv_hpages; | |
165 | spool->rsv_hpages = 0; | |
166 | } else { | |
167 | ret = 0; /* reserves already accounted for */ | |
168 | spool->rsv_hpages -= delta; | |
169 | } | |
170 | } | |
171 | ||
172 | unlock_ret: | |
173 | spin_unlock(&spool->lock); | |
90481622 DG |
174 | return ret; |
175 | } | |
176 | ||
1c5ecae3 MK |
177 | /* |
178 | * Subpool accounting for freeing and unreserving pages. | |
179 | * Return the number of global page reservations that must be dropped. | |
180 | * The return value may only be different than the passed value (delta) | |
181 | * in the case where a subpool minimum size must be maintained. | |
182 | */ | |
183 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
184 | long delta) |
185 | { | |
1c5ecae3 MK |
186 | long ret = delta; |
187 | ||
90481622 | 188 | if (!spool) |
1c5ecae3 | 189 | return delta; |
90481622 DG |
190 | |
191 | spin_lock(&spool->lock); | |
1c5ecae3 MK |
192 | |
193 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
194 | spool->used_hpages -= delta; | |
195 | ||
09a95e29 MK |
196 | /* minimum size accounting */ |
197 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
198 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
199 | ret = 0; | |
200 | else | |
201 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
202 | ||
203 | spool->rsv_hpages += delta; | |
204 | if (spool->rsv_hpages > spool->min_hpages) | |
205 | spool->rsv_hpages = spool->min_hpages; | |
206 | } | |
207 | ||
208 | /* | |
209 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
210 | * quota reference, free it now. | |
211 | */ | |
90481622 | 212 | unlock_or_release_subpool(spool); |
1c5ecae3 MK |
213 | |
214 | return ret; | |
90481622 DG |
215 | } |
216 | ||
217 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
218 | { | |
219 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
220 | } | |
221 | ||
222 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
223 | { | |
496ad9aa | 224 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
225 | } |
226 | ||
0db9d74e MA |
227 | /* Helper that removes a struct file_region from the resv_map cache and returns |
228 | * it for use. | |
229 | */ | |
230 | static struct file_region * | |
231 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
232 | { | |
233 | struct file_region *nrg = NULL; | |
234 | ||
235 | VM_BUG_ON(resv->region_cache_count <= 0); | |
236 | ||
237 | resv->region_cache_count--; | |
238 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
239 | VM_BUG_ON(!nrg); | |
240 | list_del(&nrg->link); | |
241 | ||
242 | nrg->from = from; | |
243 | nrg->to = to; | |
244 | ||
245 | return nrg; | |
246 | } | |
247 | ||
075a61d0 MA |
248 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
249 | struct file_region *rg) | |
250 | { | |
251 | #ifdef CONFIG_CGROUP_HUGETLB | |
252 | nrg->reservation_counter = rg->reservation_counter; | |
253 | nrg->css = rg->css; | |
254 | if (rg->css) | |
255 | css_get(rg->css); | |
256 | #endif | |
257 | } | |
258 | ||
259 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
260 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
261 | struct hstate *h, | |
262 | struct resv_map *resv, | |
263 | struct file_region *nrg) | |
264 | { | |
265 | #ifdef CONFIG_CGROUP_HUGETLB | |
266 | if (h_cg) { | |
267 | nrg->reservation_counter = | |
268 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
269 | nrg->css = &h_cg->css; | |
270 | if (!resv->pages_per_hpage) | |
271 | resv->pages_per_hpage = pages_per_huge_page(h); | |
272 | /* pages_per_hpage should be the same for all entries in | |
273 | * a resv_map. | |
274 | */ | |
275 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
276 | } else { | |
277 | nrg->reservation_counter = NULL; | |
278 | nrg->css = NULL; | |
279 | } | |
280 | #endif | |
281 | } | |
282 | ||
a9b3f867 MA |
283 | static bool has_same_uncharge_info(struct file_region *rg, |
284 | struct file_region *org) | |
285 | { | |
286 | #ifdef CONFIG_CGROUP_HUGETLB | |
287 | return rg && org && | |
288 | rg->reservation_counter == org->reservation_counter && | |
289 | rg->css == org->css; | |
290 | ||
291 | #else | |
292 | return true; | |
293 | #endif | |
294 | } | |
295 | ||
296 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
297 | { | |
298 | struct file_region *nrg = NULL, *prg = NULL; | |
299 | ||
300 | prg = list_prev_entry(rg, link); | |
301 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
302 | has_same_uncharge_info(prg, rg)) { | |
303 | prg->to = rg->to; | |
304 | ||
305 | list_del(&rg->link); | |
306 | kfree(rg); | |
307 | ||
308 | coalesce_file_region(resv, prg); | |
309 | return; | |
310 | } | |
311 | ||
312 | nrg = list_next_entry(rg, link); | |
313 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
314 | has_same_uncharge_info(nrg, rg)) { | |
315 | nrg->from = rg->from; | |
316 | ||
317 | list_del(&rg->link); | |
318 | kfree(rg); | |
319 | ||
320 | coalesce_file_region(resv, nrg); | |
321 | return; | |
322 | } | |
323 | } | |
324 | ||
d75c6af9 MA |
325 | /* Must be called with resv->lock held. Calling this with count_only == true |
326 | * will count the number of pages to be added but will not modify the linked | |
0db9d74e MA |
327 | * list. If regions_needed != NULL and count_only == true, then regions_needed |
328 | * will indicate the number of file_regions needed in the cache to carry out to | |
329 | * add the regions for this range. | |
d75c6af9 MA |
330 | */ |
331 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 MA |
332 | struct hugetlb_cgroup *h_cg, |
333 | struct hstate *h, long *regions_needed, | |
334 | bool count_only) | |
d75c6af9 | 335 | { |
0db9d74e | 336 | long add = 0; |
d75c6af9 | 337 | struct list_head *head = &resv->regions; |
0db9d74e | 338 | long last_accounted_offset = f; |
d75c6af9 MA |
339 | struct file_region *rg = NULL, *trg = NULL, *nrg = NULL; |
340 | ||
0db9d74e MA |
341 | if (regions_needed) |
342 | *regions_needed = 0; | |
d75c6af9 | 343 | |
0db9d74e MA |
344 | /* In this loop, we essentially handle an entry for the range |
345 | * [last_accounted_offset, rg->from), at every iteration, with some | |
346 | * bounds checking. | |
347 | */ | |
348 | list_for_each_entry_safe(rg, trg, head, link) { | |
349 | /* Skip irrelevant regions that start before our range. */ | |
350 | if (rg->from < f) { | |
351 | /* If this region ends after the last accounted offset, | |
352 | * then we need to update last_accounted_offset. | |
353 | */ | |
354 | if (rg->to > last_accounted_offset) | |
355 | last_accounted_offset = rg->to; | |
356 | continue; | |
357 | } | |
d75c6af9 | 358 | |
0db9d74e MA |
359 | /* When we find a region that starts beyond our range, we've |
360 | * finished. | |
361 | */ | |
d75c6af9 MA |
362 | if (rg->from > t) |
363 | break; | |
364 | ||
0db9d74e MA |
365 | /* Add an entry for last_accounted_offset -> rg->from, and |
366 | * update last_accounted_offset. | |
367 | */ | |
368 | if (rg->from > last_accounted_offset) { | |
369 | add += rg->from - last_accounted_offset; | |
370 | if (!count_only) { | |
371 | nrg = get_file_region_entry_from_cache( | |
372 | resv, last_accounted_offset, rg->from); | |
075a61d0 MA |
373 | record_hugetlb_cgroup_uncharge_info(h_cg, h, |
374 | resv, nrg); | |
0db9d74e | 375 | list_add(&nrg->link, rg->link.prev); |
a9b3f867 | 376 | coalesce_file_region(resv, nrg); |
0db9d74e MA |
377 | } else if (regions_needed) |
378 | *regions_needed += 1; | |
379 | } | |
380 | ||
381 | last_accounted_offset = rg->to; | |
382 | } | |
383 | ||
384 | /* Handle the case where our range extends beyond | |
385 | * last_accounted_offset. | |
386 | */ | |
387 | if (last_accounted_offset < t) { | |
388 | add += t - last_accounted_offset; | |
389 | if (!count_only) { | |
390 | nrg = get_file_region_entry_from_cache( | |
391 | resv, last_accounted_offset, t); | |
075a61d0 | 392 | record_hugetlb_cgroup_uncharge_info(h_cg, h, resv, nrg); |
0db9d74e | 393 | list_add(&nrg->link, rg->link.prev); |
a9b3f867 | 394 | coalesce_file_region(resv, nrg); |
0db9d74e MA |
395 | } else if (regions_needed) |
396 | *regions_needed += 1; | |
397 | } | |
398 | ||
399 | VM_BUG_ON(add < 0); | |
400 | return add; | |
401 | } | |
402 | ||
403 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
404 | */ | |
405 | static int allocate_file_region_entries(struct resv_map *resv, | |
406 | int regions_needed) | |
407 | __must_hold(&resv->lock) | |
408 | { | |
409 | struct list_head allocated_regions; | |
410 | int to_allocate = 0, i = 0; | |
411 | struct file_region *trg = NULL, *rg = NULL; | |
412 | ||
413 | VM_BUG_ON(regions_needed < 0); | |
414 | ||
415 | INIT_LIST_HEAD(&allocated_regions); | |
416 | ||
417 | /* | |
418 | * Check for sufficient descriptors in the cache to accommodate | |
419 | * the number of in progress add operations plus regions_needed. | |
420 | * | |
421 | * This is a while loop because when we drop the lock, some other call | |
422 | * to region_add or region_del may have consumed some region_entries, | |
423 | * so we keep looping here until we finally have enough entries for | |
424 | * (adds_in_progress + regions_needed). | |
425 | */ | |
426 | while (resv->region_cache_count < | |
427 | (resv->adds_in_progress + regions_needed)) { | |
428 | to_allocate = resv->adds_in_progress + regions_needed - | |
429 | resv->region_cache_count; | |
430 | ||
431 | /* At this point, we should have enough entries in the cache | |
432 | * for all the existings adds_in_progress. We should only be | |
433 | * needing to allocate for regions_needed. | |
d75c6af9 | 434 | */ |
0db9d74e MA |
435 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
436 | ||
437 | spin_unlock(&resv->lock); | |
438 | for (i = 0; i < to_allocate; i++) { | |
439 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
440 | if (!trg) | |
441 | goto out_of_memory; | |
442 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 443 | } |
d75c6af9 | 444 | |
0db9d74e MA |
445 | spin_lock(&resv->lock); |
446 | ||
447 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
d75c6af9 | 448 | list_del(&rg->link); |
0db9d74e MA |
449 | list_add(&rg->link, &resv->region_cache); |
450 | resv->region_cache_count++; | |
d75c6af9 MA |
451 | } |
452 | } | |
453 | ||
0db9d74e | 454 | return 0; |
d75c6af9 | 455 | |
0db9d74e MA |
456 | out_of_memory: |
457 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
458 | list_del(&rg->link); | |
459 | kfree(rg); | |
460 | } | |
461 | return -ENOMEM; | |
d75c6af9 MA |
462 | } |
463 | ||
1dd308a7 MK |
464 | /* |
465 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
466 | * map. Regions will be taken from the cache to fill in this range. |
467 | * Sufficient regions should exist in the cache due to the previous | |
468 | * call to region_chg with the same range, but in some cases the cache will not | |
469 | * have sufficient entries due to races with other code doing region_add or | |
470 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 471 | * |
0db9d74e MA |
472 | * regions_needed is the out value provided by a previous call to region_chg. |
473 | * | |
474 | * Return the number of new huge pages added to the map. This number is greater | |
475 | * than or equal to zero. If file_region entries needed to be allocated for | |
476 | * this operation and we were not able to allocate, it ruturns -ENOMEM. | |
477 | * region_add of regions of length 1 never allocate file_regions and cannot | |
478 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
479 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 480 | */ |
0db9d74e | 481 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
482 | long in_regions_needed, struct hstate *h, |
483 | struct hugetlb_cgroup *h_cg) | |
96822904 | 484 | { |
0db9d74e | 485 | long add = 0, actual_regions_needed = 0; |
96822904 | 486 | |
7b24d861 | 487 | spin_lock(&resv->lock); |
0db9d74e MA |
488 | retry: |
489 | ||
490 | /* Count how many regions are actually needed to execute this add. */ | |
075a61d0 MA |
491 | add_reservation_in_range(resv, f, t, NULL, NULL, &actual_regions_needed, |
492 | true); | |
96822904 | 493 | |
5e911373 | 494 | /* |
0db9d74e MA |
495 | * Check for sufficient descriptors in the cache to accommodate |
496 | * this add operation. Note that actual_regions_needed may be greater | |
497 | * than in_regions_needed, as the resv_map may have been modified since | |
498 | * the region_chg call. In this case, we need to make sure that we | |
499 | * allocate extra entries, such that we have enough for all the | |
500 | * existing adds_in_progress, plus the excess needed for this | |
501 | * operation. | |
5e911373 | 502 | */ |
0db9d74e MA |
503 | if (actual_regions_needed > in_regions_needed && |
504 | resv->region_cache_count < | |
505 | resv->adds_in_progress + | |
506 | (actual_regions_needed - in_regions_needed)) { | |
507 | /* region_add operation of range 1 should never need to | |
508 | * allocate file_region entries. | |
509 | */ | |
510 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 511 | |
0db9d74e MA |
512 | if (allocate_file_region_entries( |
513 | resv, actual_regions_needed - in_regions_needed)) { | |
514 | return -ENOMEM; | |
515 | } | |
5e911373 | 516 | |
0db9d74e | 517 | goto retry; |
5e911373 MK |
518 | } |
519 | ||
075a61d0 | 520 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL, false); |
0db9d74e MA |
521 | |
522 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 523 | |
7b24d861 | 524 | spin_unlock(&resv->lock); |
cf3ad20b MK |
525 | VM_BUG_ON(add < 0); |
526 | return add; | |
96822904 AW |
527 | } |
528 | ||
1dd308a7 MK |
529 | /* |
530 | * Examine the existing reserve map and determine how many | |
531 | * huge pages in the specified range [f, t) are NOT currently | |
532 | * represented. This routine is called before a subsequent | |
533 | * call to region_add that will actually modify the reserve | |
534 | * map to add the specified range [f, t). region_chg does | |
535 | * not change the number of huge pages represented by the | |
0db9d74e MA |
536 | * map. A number of new file_region structures is added to the cache as a |
537 | * placeholder, for the subsequent region_add call to use. At least 1 | |
538 | * file_region structure is added. | |
539 | * | |
540 | * out_regions_needed is the number of regions added to the | |
541 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
542 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
543 | * |
544 | * Returns the number of huge pages that need to be added to the existing | |
545 | * reservation map for the range [f, t). This number is greater or equal to | |
546 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
547 | * is needed and can not be allocated. | |
1dd308a7 | 548 | */ |
0db9d74e MA |
549 | static long region_chg(struct resv_map *resv, long f, long t, |
550 | long *out_regions_needed) | |
96822904 | 551 | { |
96822904 AW |
552 | long chg = 0; |
553 | ||
7b24d861 | 554 | spin_lock(&resv->lock); |
5e911373 | 555 | |
0db9d74e | 556 | /* Count how many hugepages in this range are NOT respresented. */ |
075a61d0 MA |
557 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
558 | out_regions_needed, true); | |
5e911373 | 559 | |
0db9d74e MA |
560 | if (*out_regions_needed == 0) |
561 | *out_regions_needed = 1; | |
5e911373 | 562 | |
0db9d74e MA |
563 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
564 | return -ENOMEM; | |
5e911373 | 565 | |
0db9d74e | 566 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 567 | |
7b24d861 | 568 | spin_unlock(&resv->lock); |
96822904 AW |
569 | return chg; |
570 | } | |
571 | ||
5e911373 MK |
572 | /* |
573 | * Abort the in progress add operation. The adds_in_progress field | |
574 | * of the resv_map keeps track of the operations in progress between | |
575 | * calls to region_chg and region_add. Operations are sometimes | |
576 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
577 | * is called to decrement the adds_in_progress counter. regions_needed |
578 | * is the value returned by the region_chg call, it is used to decrement | |
579 | * the adds_in_progress counter. | |
5e911373 MK |
580 | * |
581 | * NOTE: The range arguments [f, t) are not needed or used in this | |
582 | * routine. They are kept to make reading the calling code easier as | |
583 | * arguments will match the associated region_chg call. | |
584 | */ | |
0db9d74e MA |
585 | static void region_abort(struct resv_map *resv, long f, long t, |
586 | long regions_needed) | |
5e911373 MK |
587 | { |
588 | spin_lock(&resv->lock); | |
589 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 590 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
591 | spin_unlock(&resv->lock); |
592 | } | |
593 | ||
1dd308a7 | 594 | /* |
feba16e2 MK |
595 | * Delete the specified range [f, t) from the reserve map. If the |
596 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
597 | * should be deleted. Locate the regions which intersect [f, t) | |
598 | * and either trim, delete or split the existing regions. | |
599 | * | |
600 | * Returns the number of huge pages deleted from the reserve map. | |
601 | * In the normal case, the return value is zero or more. In the | |
602 | * case where a region must be split, a new region descriptor must | |
603 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
604 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
605 | * a region and possibly return -ENOMEM. Callers specifying | |
606 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 607 | */ |
feba16e2 | 608 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 609 | { |
1406ec9b | 610 | struct list_head *head = &resv->regions; |
96822904 | 611 | struct file_region *rg, *trg; |
feba16e2 MK |
612 | struct file_region *nrg = NULL; |
613 | long del = 0; | |
96822904 | 614 | |
feba16e2 | 615 | retry: |
7b24d861 | 616 | spin_lock(&resv->lock); |
feba16e2 | 617 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
618 | /* |
619 | * Skip regions before the range to be deleted. file_region | |
620 | * ranges are normally of the form [from, to). However, there | |
621 | * may be a "placeholder" entry in the map which is of the form | |
622 | * (from, to) with from == to. Check for placeholder entries | |
623 | * at the beginning of the range to be deleted. | |
624 | */ | |
625 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 626 | continue; |
dbe409e4 | 627 | |
feba16e2 | 628 | if (rg->from >= t) |
96822904 | 629 | break; |
96822904 | 630 | |
feba16e2 MK |
631 | if (f > rg->from && t < rg->to) { /* Must split region */ |
632 | /* | |
633 | * Check for an entry in the cache before dropping | |
634 | * lock and attempting allocation. | |
635 | */ | |
636 | if (!nrg && | |
637 | resv->region_cache_count > resv->adds_in_progress) { | |
638 | nrg = list_first_entry(&resv->region_cache, | |
639 | struct file_region, | |
640 | link); | |
641 | list_del(&nrg->link); | |
642 | resv->region_cache_count--; | |
643 | } | |
96822904 | 644 | |
feba16e2 MK |
645 | if (!nrg) { |
646 | spin_unlock(&resv->lock); | |
647 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
648 | if (!nrg) | |
649 | return -ENOMEM; | |
650 | goto retry; | |
651 | } | |
652 | ||
653 | del += t - f; | |
654 | ||
655 | /* New entry for end of split region */ | |
656 | nrg->from = t; | |
657 | nrg->to = rg->to; | |
075a61d0 MA |
658 | |
659 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
660 | ||
feba16e2 MK |
661 | INIT_LIST_HEAD(&nrg->link); |
662 | ||
663 | /* Original entry is trimmed */ | |
664 | rg->to = f; | |
665 | ||
075a61d0 MA |
666 | hugetlb_cgroup_uncharge_file_region( |
667 | resv, rg, nrg->to - nrg->from); | |
668 | ||
feba16e2 MK |
669 | list_add(&nrg->link, &rg->link); |
670 | nrg = NULL; | |
96822904 | 671 | break; |
feba16e2 MK |
672 | } |
673 | ||
674 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
675 | del += rg->to - rg->from; | |
075a61d0 MA |
676 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
677 | rg->to - rg->from); | |
feba16e2 MK |
678 | list_del(&rg->link); |
679 | kfree(rg); | |
680 | continue; | |
681 | } | |
682 | ||
683 | if (f <= rg->from) { /* Trim beginning of region */ | |
684 | del += t - rg->from; | |
685 | rg->from = t; | |
075a61d0 MA |
686 | |
687 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
688 | t - rg->from); | |
feba16e2 MK |
689 | } else { /* Trim end of region */ |
690 | del += rg->to - f; | |
691 | rg->to = f; | |
075a61d0 MA |
692 | |
693 | hugetlb_cgroup_uncharge_file_region(resv, rg, | |
694 | rg->to - f); | |
feba16e2 | 695 | } |
96822904 | 696 | } |
7b24d861 | 697 | |
7b24d861 | 698 | spin_unlock(&resv->lock); |
feba16e2 MK |
699 | kfree(nrg); |
700 | return del; | |
96822904 AW |
701 | } |
702 | ||
b5cec28d MK |
703 | /* |
704 | * A rare out of memory error was encountered which prevented removal of | |
705 | * the reserve map region for a page. The huge page itself was free'ed | |
706 | * and removed from the page cache. This routine will adjust the subpool | |
707 | * usage count, and the global reserve count if needed. By incrementing | |
708 | * these counts, the reserve map entry which could not be deleted will | |
709 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
710 | * counts. | |
711 | */ | |
72e2936c | 712 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
713 | { |
714 | struct hugepage_subpool *spool = subpool_inode(inode); | |
715 | long rsv_adjust; | |
716 | ||
717 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
72e2936c | 718 | if (rsv_adjust) { |
b5cec28d MK |
719 | struct hstate *h = hstate_inode(inode); |
720 | ||
721 | hugetlb_acct_memory(h, 1); | |
722 | } | |
723 | } | |
724 | ||
1dd308a7 MK |
725 | /* |
726 | * Count and return the number of huge pages in the reserve map | |
727 | * that intersect with the range [f, t). | |
728 | */ | |
1406ec9b | 729 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 730 | { |
1406ec9b | 731 | struct list_head *head = &resv->regions; |
84afd99b AW |
732 | struct file_region *rg; |
733 | long chg = 0; | |
734 | ||
7b24d861 | 735 | spin_lock(&resv->lock); |
84afd99b AW |
736 | /* Locate each segment we overlap with, and count that overlap. */ |
737 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
738 | long seg_from; |
739 | long seg_to; | |
84afd99b AW |
740 | |
741 | if (rg->to <= f) | |
742 | continue; | |
743 | if (rg->from >= t) | |
744 | break; | |
745 | ||
746 | seg_from = max(rg->from, f); | |
747 | seg_to = min(rg->to, t); | |
748 | ||
749 | chg += seg_to - seg_from; | |
750 | } | |
7b24d861 | 751 | spin_unlock(&resv->lock); |
84afd99b AW |
752 | |
753 | return chg; | |
754 | } | |
755 | ||
e7c4b0bf AW |
756 | /* |
757 | * Convert the address within this vma to the page offset within | |
758 | * the mapping, in pagecache page units; huge pages here. | |
759 | */ | |
a5516438 AK |
760 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
761 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 762 | { |
a5516438 AK |
763 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
764 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
765 | } |
766 | ||
0fe6e20b NH |
767 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
768 | unsigned long address) | |
769 | { | |
770 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
771 | } | |
dee41079 | 772 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 773 | |
08fba699 MG |
774 | /* |
775 | * Return the size of the pages allocated when backing a VMA. In the majority | |
776 | * cases this will be same size as used by the page table entries. | |
777 | */ | |
778 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
779 | { | |
05ea8860 DW |
780 | if (vma->vm_ops && vma->vm_ops->pagesize) |
781 | return vma->vm_ops->pagesize(vma); | |
782 | return PAGE_SIZE; | |
08fba699 | 783 | } |
f340ca0f | 784 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 785 | |
3340289d MG |
786 | /* |
787 | * Return the page size being used by the MMU to back a VMA. In the majority | |
788 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
789 | * architectures where it differs, an architecture-specific 'strong' |
790 | * version of this symbol is required. | |
3340289d | 791 | */ |
09135cc5 | 792 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
793 | { |
794 | return vma_kernel_pagesize(vma); | |
795 | } | |
3340289d | 796 | |
84afd99b AW |
797 | /* |
798 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
799 | * bits of the reservation map pointer, which are always clear due to | |
800 | * alignment. | |
801 | */ | |
802 | #define HPAGE_RESV_OWNER (1UL << 0) | |
803 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 804 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 805 | |
a1e78772 MG |
806 | /* |
807 | * These helpers are used to track how many pages are reserved for | |
808 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
809 | * is guaranteed to have their future faults succeed. | |
810 | * | |
811 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
812 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
813 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
814 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
815 | * |
816 | * The private mapping reservation is represented in a subtly different | |
817 | * manner to a shared mapping. A shared mapping has a region map associated | |
818 | * with the underlying file, this region map represents the backing file | |
819 | * pages which have ever had a reservation assigned which this persists even | |
820 | * after the page is instantiated. A private mapping has a region map | |
821 | * associated with the original mmap which is attached to all VMAs which | |
822 | * reference it, this region map represents those offsets which have consumed | |
823 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 824 | */ |
e7c4b0bf AW |
825 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
826 | { | |
827 | return (unsigned long)vma->vm_private_data; | |
828 | } | |
829 | ||
830 | static void set_vma_private_data(struct vm_area_struct *vma, | |
831 | unsigned long value) | |
832 | { | |
833 | vma->vm_private_data = (void *)value; | |
834 | } | |
835 | ||
e9fe92ae MA |
836 | static void |
837 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
838 | struct hugetlb_cgroup *h_cg, | |
839 | struct hstate *h) | |
840 | { | |
841 | #ifdef CONFIG_CGROUP_HUGETLB | |
842 | if (!h_cg || !h) { | |
843 | resv_map->reservation_counter = NULL; | |
844 | resv_map->pages_per_hpage = 0; | |
845 | resv_map->css = NULL; | |
846 | } else { | |
847 | resv_map->reservation_counter = | |
848 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
849 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
850 | resv_map->css = &h_cg->css; | |
851 | } | |
852 | #endif | |
853 | } | |
854 | ||
9119a41e | 855 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
856 | { |
857 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
858 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
859 | ||
860 | if (!resv_map || !rg) { | |
861 | kfree(resv_map); | |
862 | kfree(rg); | |
84afd99b | 863 | return NULL; |
5e911373 | 864 | } |
84afd99b AW |
865 | |
866 | kref_init(&resv_map->refs); | |
7b24d861 | 867 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
868 | INIT_LIST_HEAD(&resv_map->regions); |
869 | ||
5e911373 | 870 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
871 | /* |
872 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
873 | * fields don't do cgroup accounting. On private mappings, these will be | |
874 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
875 | * reservations are to be un-charged from here. | |
876 | */ | |
877 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
878 | |
879 | INIT_LIST_HEAD(&resv_map->region_cache); | |
880 | list_add(&rg->link, &resv_map->region_cache); | |
881 | resv_map->region_cache_count = 1; | |
882 | ||
84afd99b AW |
883 | return resv_map; |
884 | } | |
885 | ||
9119a41e | 886 | void resv_map_release(struct kref *ref) |
84afd99b AW |
887 | { |
888 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
889 | struct list_head *head = &resv_map->region_cache; |
890 | struct file_region *rg, *trg; | |
84afd99b AW |
891 | |
892 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 893 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
894 | |
895 | /* ... and any entries left in the cache */ | |
896 | list_for_each_entry_safe(rg, trg, head, link) { | |
897 | list_del(&rg->link); | |
898 | kfree(rg); | |
899 | } | |
900 | ||
901 | VM_BUG_ON(resv_map->adds_in_progress); | |
902 | ||
84afd99b AW |
903 | kfree(resv_map); |
904 | } | |
905 | ||
4e35f483 JK |
906 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
907 | { | |
f27a5136 MK |
908 | /* |
909 | * At inode evict time, i_mapping may not point to the original | |
910 | * address space within the inode. This original address space | |
911 | * contains the pointer to the resv_map. So, always use the | |
912 | * address space embedded within the inode. | |
913 | * The VERY common case is inode->mapping == &inode->i_data but, | |
914 | * this may not be true for device special inodes. | |
915 | */ | |
916 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
917 | } |
918 | ||
84afd99b | 919 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 920 | { |
81d1b09c | 921 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
922 | if (vma->vm_flags & VM_MAYSHARE) { |
923 | struct address_space *mapping = vma->vm_file->f_mapping; | |
924 | struct inode *inode = mapping->host; | |
925 | ||
926 | return inode_resv_map(inode); | |
927 | ||
928 | } else { | |
84afd99b AW |
929 | return (struct resv_map *)(get_vma_private_data(vma) & |
930 | ~HPAGE_RESV_MASK); | |
4e35f483 | 931 | } |
a1e78772 MG |
932 | } |
933 | ||
84afd99b | 934 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 935 | { |
81d1b09c SL |
936 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
937 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 938 | |
84afd99b AW |
939 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
940 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
941 | } |
942 | ||
943 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
944 | { | |
81d1b09c SL |
945 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
946 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
947 | |
948 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
949 | } |
950 | ||
951 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
952 | { | |
81d1b09c | 953 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
954 | |
955 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
956 | } |
957 | ||
04f2cbe3 | 958 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
959 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
960 | { | |
81d1b09c | 961 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 962 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
963 | vma->vm_private_data = (void *)0; |
964 | } | |
965 | ||
966 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 967 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 968 | { |
af0ed73e JK |
969 | if (vma->vm_flags & VM_NORESERVE) { |
970 | /* | |
971 | * This address is already reserved by other process(chg == 0), | |
972 | * so, we should decrement reserved count. Without decrementing, | |
973 | * reserve count remains after releasing inode, because this | |
974 | * allocated page will go into page cache and is regarded as | |
975 | * coming from reserved pool in releasing step. Currently, we | |
976 | * don't have any other solution to deal with this situation | |
977 | * properly, so add work-around here. | |
978 | */ | |
979 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 980 | return true; |
af0ed73e | 981 | else |
559ec2f8 | 982 | return false; |
af0ed73e | 983 | } |
a63884e9 JK |
984 | |
985 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
986 | if (vma->vm_flags & VM_MAYSHARE) { |
987 | /* | |
988 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
989 | * be a region map for all pages. The only situation where | |
990 | * there is no region map is if a hole was punched via | |
991 | * fallocate. In this case, there really are no reverves to | |
992 | * use. This situation is indicated if chg != 0. | |
993 | */ | |
994 | if (chg) | |
995 | return false; | |
996 | else | |
997 | return true; | |
998 | } | |
a63884e9 JK |
999 | |
1000 | /* | |
1001 | * Only the process that called mmap() has reserves for | |
1002 | * private mappings. | |
1003 | */ | |
67961f9d MK |
1004 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1005 | /* | |
1006 | * Like the shared case above, a hole punch or truncate | |
1007 | * could have been performed on the private mapping. | |
1008 | * Examine the value of chg to determine if reserves | |
1009 | * actually exist or were previously consumed. | |
1010 | * Very Subtle - The value of chg comes from a previous | |
1011 | * call to vma_needs_reserves(). The reserve map for | |
1012 | * private mappings has different (opposite) semantics | |
1013 | * than that of shared mappings. vma_needs_reserves() | |
1014 | * has already taken this difference in semantics into | |
1015 | * account. Therefore, the meaning of chg is the same | |
1016 | * as in the shared case above. Code could easily be | |
1017 | * combined, but keeping it separate draws attention to | |
1018 | * subtle differences. | |
1019 | */ | |
1020 | if (chg) | |
1021 | return false; | |
1022 | else | |
1023 | return true; | |
1024 | } | |
a63884e9 | 1025 | |
559ec2f8 | 1026 | return false; |
a1e78772 MG |
1027 | } |
1028 | ||
a5516438 | 1029 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1030 | { |
1031 | int nid = page_to_nid(page); | |
0edaecfa | 1032 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1033 | h->free_huge_pages++; |
1034 | h->free_huge_pages_node[nid]++; | |
1da177e4 LT |
1035 | } |
1036 | ||
94310cbc | 1037 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1038 | { |
1039 | struct page *page; | |
1040 | ||
c8721bbb | 1041 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) |
243abd5b | 1042 | if (!PageHWPoison(page)) |
c8721bbb NH |
1043 | break; |
1044 | /* | |
1045 | * if 'non-isolated free hugepage' not found on the list, | |
1046 | * the allocation fails. | |
1047 | */ | |
1048 | if (&h->hugepage_freelists[nid] == &page->lru) | |
bf50bab2 | 1049 | return NULL; |
0edaecfa | 1050 | list_move(&page->lru, &h->hugepage_activelist); |
a9869b83 | 1051 | set_page_refcounted(page); |
bf50bab2 NH |
1052 | h->free_huge_pages--; |
1053 | h->free_huge_pages_node[nid]--; | |
1054 | return page; | |
1055 | } | |
1056 | ||
3e59fcb0 MH |
1057 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1058 | nodemask_t *nmask) | |
94310cbc | 1059 | { |
3e59fcb0 MH |
1060 | unsigned int cpuset_mems_cookie; |
1061 | struct zonelist *zonelist; | |
1062 | struct zone *zone; | |
1063 | struct zoneref *z; | |
98fa15f3 | 1064 | int node = NUMA_NO_NODE; |
94310cbc | 1065 | |
3e59fcb0 MH |
1066 | zonelist = node_zonelist(nid, gfp_mask); |
1067 | ||
1068 | retry_cpuset: | |
1069 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1070 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1071 | struct page *page; | |
1072 | ||
1073 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1074 | continue; | |
1075 | /* | |
1076 | * no need to ask again on the same node. Pool is node rather than | |
1077 | * zone aware | |
1078 | */ | |
1079 | if (zone_to_nid(zone) == node) | |
1080 | continue; | |
1081 | node = zone_to_nid(zone); | |
94310cbc | 1082 | |
94310cbc AK |
1083 | page = dequeue_huge_page_node_exact(h, node); |
1084 | if (page) | |
1085 | return page; | |
1086 | } | |
3e59fcb0 MH |
1087 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1088 | goto retry_cpuset; | |
1089 | ||
94310cbc AK |
1090 | return NULL; |
1091 | } | |
1092 | ||
86cdb465 NH |
1093 | /* Movability of hugepages depends on migration support. */ |
1094 | static inline gfp_t htlb_alloc_mask(struct hstate *h) | |
1095 | { | |
7ed2c31d | 1096 | if (hugepage_movable_supported(h)) |
86cdb465 NH |
1097 | return GFP_HIGHUSER_MOVABLE; |
1098 | else | |
1099 | return GFP_HIGHUSER; | |
1100 | } | |
1101 | ||
a5516438 AK |
1102 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1103 | struct vm_area_struct *vma, | |
af0ed73e JK |
1104 | unsigned long address, int avoid_reserve, |
1105 | long chg) | |
1da177e4 | 1106 | { |
3e59fcb0 | 1107 | struct page *page; |
480eccf9 | 1108 | struct mempolicy *mpol; |
04ec6264 | 1109 | gfp_t gfp_mask; |
3e59fcb0 | 1110 | nodemask_t *nodemask; |
04ec6264 | 1111 | int nid; |
1da177e4 | 1112 | |
a1e78772 MG |
1113 | /* |
1114 | * A child process with MAP_PRIVATE mappings created by their parent | |
1115 | * have no page reserves. This check ensures that reservations are | |
1116 | * not "stolen". The child may still get SIGKILLed | |
1117 | */ | |
af0ed73e | 1118 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1119 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1120 | goto err; |
a1e78772 | 1121 | |
04f2cbe3 | 1122 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1123 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1124 | goto err; |
04f2cbe3 | 1125 | |
04ec6264 VB |
1126 | gfp_mask = htlb_alloc_mask(h); |
1127 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
1128 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
1129 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
1130 | SetPagePrivate(page); | |
1131 | h->resv_huge_pages--; | |
1da177e4 | 1132 | } |
cc9a6c87 | 1133 | |
52cd3b07 | 1134 | mpol_cond_put(mpol); |
1da177e4 | 1135 | return page; |
cc9a6c87 MG |
1136 | |
1137 | err: | |
cc9a6c87 | 1138 | return NULL; |
1da177e4 LT |
1139 | } |
1140 | ||
1cac6f2c LC |
1141 | /* |
1142 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1143 | * We may have allocated or freed a huge page based on a different | |
1144 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1145 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1146 | * node for alloc or free. | |
1147 | */ | |
1148 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1149 | { | |
0edaf86c | 1150 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1151 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1152 | ||
1153 | return nid; | |
1154 | } | |
1155 | ||
1156 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1157 | { | |
1158 | if (!node_isset(nid, *nodes_allowed)) | |
1159 | nid = next_node_allowed(nid, nodes_allowed); | |
1160 | return nid; | |
1161 | } | |
1162 | ||
1163 | /* | |
1164 | * returns the previously saved node ["this node"] from which to | |
1165 | * allocate a persistent huge page for the pool and advance the | |
1166 | * next node from which to allocate, handling wrap at end of node | |
1167 | * mask. | |
1168 | */ | |
1169 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1170 | nodemask_t *nodes_allowed) | |
1171 | { | |
1172 | int nid; | |
1173 | ||
1174 | VM_BUG_ON(!nodes_allowed); | |
1175 | ||
1176 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1177 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1178 | ||
1179 | return nid; | |
1180 | } | |
1181 | ||
1182 | /* | |
1183 | * helper for free_pool_huge_page() - return the previously saved | |
1184 | * node ["this node"] from which to free a huge page. Advance the | |
1185 | * next node id whether or not we find a free huge page to free so | |
1186 | * that the next attempt to free addresses the next node. | |
1187 | */ | |
1188 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1189 | { | |
1190 | int nid; | |
1191 | ||
1192 | VM_BUG_ON(!nodes_allowed); | |
1193 | ||
1194 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1195 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1196 | ||
1197 | return nid; | |
1198 | } | |
1199 | ||
1200 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1201 | for (nr_nodes = nodes_weight(*mask); \ | |
1202 | nr_nodes > 0 && \ | |
1203 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1204 | nr_nodes--) | |
1205 | ||
1206 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1207 | for (nr_nodes = nodes_weight(*mask); \ | |
1208 | nr_nodes > 0 && \ | |
1209 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1210 | nr_nodes--) | |
1211 | ||
e1073d1e | 1212 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1213 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1214 | unsigned int order) |
944d9fec LC |
1215 | { |
1216 | int i; | |
1217 | int nr_pages = 1 << order; | |
1218 | struct page *p = page + 1; | |
1219 | ||
c8cc708a | 1220 | atomic_set(compound_mapcount_ptr(page), 0); |
47e29d32 JH |
1221 | if (hpage_pincount_available(page)) |
1222 | atomic_set(compound_pincount_ptr(page), 0); | |
1223 | ||
944d9fec | 1224 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1225 | clear_compound_head(p); |
944d9fec | 1226 | set_page_refcounted(p); |
944d9fec LC |
1227 | } |
1228 | ||
1229 | set_compound_order(page, 0); | |
1230 | __ClearPageHead(page); | |
1231 | } | |
1232 | ||
d00181b9 | 1233 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1234 | { |
cf11e85f RG |
1235 | /* |
1236 | * If the page isn't allocated using the cma allocator, | |
1237 | * cma_release() returns false. | |
1238 | */ | |
1239 | if (IS_ENABLED(CONFIG_CMA) && | |
1240 | cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
1241 | return; | |
1242 | ||
944d9fec LC |
1243 | free_contig_range(page_to_pfn(page), 1 << order); |
1244 | } | |
1245 | ||
4eb0716e | 1246 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1247 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1248 | int nid, nodemask_t *nodemask) | |
944d9fec | 1249 | { |
5e27a2df | 1250 | unsigned long nr_pages = 1UL << huge_page_order(h); |
944d9fec | 1251 | |
cf11e85f RG |
1252 | if (IS_ENABLED(CONFIG_CMA)) { |
1253 | struct page *page; | |
1254 | int node; | |
1255 | ||
1256 | for_each_node_mask(node, *nodemask) { | |
1257 | if (!hugetlb_cma[node]) | |
1258 | continue; | |
1259 | ||
1260 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1261 | huge_page_order(h), true); | |
1262 | if (page) | |
1263 | return page; | |
1264 | } | |
1265 | } | |
1266 | ||
5e27a2df | 1267 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1268 | } |
1269 | ||
1270 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1271 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1272 | #else /* !CONFIG_CONTIG_ALLOC */ |
1273 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1274 | int nid, nodemask_t *nodemask) | |
1275 | { | |
1276 | return NULL; | |
1277 | } | |
1278 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1279 | |
e1073d1e | 1280 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1281 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1282 | int nid, nodemask_t *nodemask) |
1283 | { | |
1284 | return NULL; | |
1285 | } | |
d00181b9 | 1286 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1287 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1288 | unsigned int order) { } |
944d9fec LC |
1289 | #endif |
1290 | ||
a5516438 | 1291 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1292 | { |
1293 | int i; | |
a5516438 | 1294 | |
4eb0716e | 1295 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1296 | return; |
18229df5 | 1297 | |
a5516438 AK |
1298 | h->nr_huge_pages--; |
1299 | h->nr_huge_pages_node[page_to_nid(page)]--; | |
1300 | for (i = 0; i < pages_per_huge_page(h); i++) { | |
32f84528 CF |
1301 | page[i].flags &= ~(1 << PG_locked | 1 << PG_error | |
1302 | 1 << PG_referenced | 1 << PG_dirty | | |
a7407a27 LC |
1303 | 1 << PG_active | 1 << PG_private | |
1304 | 1 << PG_writeback); | |
6af2acb6 | 1305 | } |
309381fe | 1306 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); |
1adc4d41 | 1307 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); |
f1e61557 | 1308 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); |
6af2acb6 | 1309 | set_page_refcounted(page); |
944d9fec | 1310 | if (hstate_is_gigantic(h)) { |
cf11e85f RG |
1311 | /* |
1312 | * Temporarily drop the hugetlb_lock, because | |
1313 | * we might block in free_gigantic_page(). | |
1314 | */ | |
1315 | spin_unlock(&hugetlb_lock); | |
944d9fec LC |
1316 | destroy_compound_gigantic_page(page, huge_page_order(h)); |
1317 | free_gigantic_page(page, huge_page_order(h)); | |
cf11e85f | 1318 | spin_lock(&hugetlb_lock); |
944d9fec | 1319 | } else { |
944d9fec LC |
1320 | __free_pages(page, huge_page_order(h)); |
1321 | } | |
6af2acb6 AL |
1322 | } |
1323 | ||
e5ff2159 AK |
1324 | struct hstate *size_to_hstate(unsigned long size) |
1325 | { | |
1326 | struct hstate *h; | |
1327 | ||
1328 | for_each_hstate(h) { | |
1329 | if (huge_page_size(h) == size) | |
1330 | return h; | |
1331 | } | |
1332 | return NULL; | |
1333 | } | |
1334 | ||
bcc54222 NH |
1335 | /* |
1336 | * Test to determine whether the hugepage is "active/in-use" (i.e. being linked | |
1337 | * to hstate->hugepage_activelist.) | |
1338 | * | |
1339 | * This function can be called for tail pages, but never returns true for them. | |
1340 | */ | |
1341 | bool page_huge_active(struct page *page) | |
1342 | { | |
1343 | VM_BUG_ON_PAGE(!PageHuge(page), page); | |
1344 | return PageHead(page) && PagePrivate(&page[1]); | |
1345 | } | |
1346 | ||
1347 | /* never called for tail page */ | |
1348 | static void set_page_huge_active(struct page *page) | |
1349 | { | |
1350 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1351 | SetPagePrivate(&page[1]); | |
1352 | } | |
1353 | ||
1354 | static void clear_page_huge_active(struct page *page) | |
1355 | { | |
1356 | VM_BUG_ON_PAGE(!PageHeadHuge(page), page); | |
1357 | ClearPagePrivate(&page[1]); | |
1358 | } | |
1359 | ||
ab5ac90a MH |
1360 | /* |
1361 | * Internal hugetlb specific page flag. Do not use outside of the hugetlb | |
1362 | * code | |
1363 | */ | |
1364 | static inline bool PageHugeTemporary(struct page *page) | |
1365 | { | |
1366 | if (!PageHuge(page)) | |
1367 | return false; | |
1368 | ||
1369 | return (unsigned long)page[2].mapping == -1U; | |
1370 | } | |
1371 | ||
1372 | static inline void SetPageHugeTemporary(struct page *page) | |
1373 | { | |
1374 | page[2].mapping = (void *)-1U; | |
1375 | } | |
1376 | ||
1377 | static inline void ClearPageHugeTemporary(struct page *page) | |
1378 | { | |
1379 | page[2].mapping = NULL; | |
1380 | } | |
1381 | ||
c77c0a8a | 1382 | static void __free_huge_page(struct page *page) |
27a85ef1 | 1383 | { |
a5516438 AK |
1384 | /* |
1385 | * Can't pass hstate in here because it is called from the | |
1386 | * compound page destructor. | |
1387 | */ | |
e5ff2159 | 1388 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1389 | int nid = page_to_nid(page); |
90481622 DG |
1390 | struct hugepage_subpool *spool = |
1391 | (struct hugepage_subpool *)page_private(page); | |
07443a85 | 1392 | bool restore_reserve; |
27a85ef1 | 1393 | |
b4330afb MK |
1394 | VM_BUG_ON_PAGE(page_count(page), page); |
1395 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc YW |
1396 | |
1397 | set_page_private(page, 0); | |
1398 | page->mapping = NULL; | |
07443a85 | 1399 | restore_reserve = PagePrivate(page); |
16c794b4 | 1400 | ClearPagePrivate(page); |
27a85ef1 | 1401 | |
1c5ecae3 | 1402 | /* |
0919e1b6 MK |
1403 | * If PagePrivate() was set on page, page allocation consumed a |
1404 | * reservation. If the page was associated with a subpool, there | |
1405 | * would have been a page reserved in the subpool before allocation | |
1406 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
1407 | * reservtion, do not call hugepage_subpool_put_pages() as this will | |
1408 | * remove the reserved page from the subpool. | |
1c5ecae3 | 1409 | */ |
0919e1b6 MK |
1410 | if (!restore_reserve) { |
1411 | /* | |
1412 | * A return code of zero implies that the subpool will be | |
1413 | * under its minimum size if the reservation is not restored | |
1414 | * after page is free. Therefore, force restore_reserve | |
1415 | * operation. | |
1416 | */ | |
1417 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1418 | restore_reserve = true; | |
1419 | } | |
1c5ecae3 | 1420 | |
27a85ef1 | 1421 | spin_lock(&hugetlb_lock); |
bcc54222 | 1422 | clear_page_huge_active(page); |
6d76dcf4 AK |
1423 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1424 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1425 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1426 | pages_per_huge_page(h), page); | |
07443a85 JK |
1427 | if (restore_reserve) |
1428 | h->resv_huge_pages++; | |
1429 | ||
ab5ac90a MH |
1430 | if (PageHugeTemporary(page)) { |
1431 | list_del(&page->lru); | |
1432 | ClearPageHugeTemporary(page); | |
1433 | update_and_free_page(h, page); | |
1434 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa AK |
1435 | /* remove the page from active list */ |
1436 | list_del(&page->lru); | |
a5516438 AK |
1437 | update_and_free_page(h, page); |
1438 | h->surplus_huge_pages--; | |
1439 | h->surplus_huge_pages_node[nid]--; | |
7893d1d5 | 1440 | } else { |
5d3a551c | 1441 | arch_clear_hugepage_flags(page); |
a5516438 | 1442 | enqueue_huge_page(h, page); |
7893d1d5 | 1443 | } |
27a85ef1 DG |
1444 | spin_unlock(&hugetlb_lock); |
1445 | } | |
1446 | ||
c77c0a8a WL |
1447 | /* |
1448 | * As free_huge_page() can be called from a non-task context, we have | |
1449 | * to defer the actual freeing in a workqueue to prevent potential | |
1450 | * hugetlb_lock deadlock. | |
1451 | * | |
1452 | * free_hpage_workfn() locklessly retrieves the linked list of pages to | |
1453 | * be freed and frees them one-by-one. As the page->mapping pointer is | |
1454 | * going to be cleared in __free_huge_page() anyway, it is reused as the | |
1455 | * llist_node structure of a lockless linked list of huge pages to be freed. | |
1456 | */ | |
1457 | static LLIST_HEAD(hpage_freelist); | |
1458 | ||
1459 | static void free_hpage_workfn(struct work_struct *work) | |
1460 | { | |
1461 | struct llist_node *node; | |
1462 | struct page *page; | |
1463 | ||
1464 | node = llist_del_all(&hpage_freelist); | |
1465 | ||
1466 | while (node) { | |
1467 | page = container_of((struct address_space **)node, | |
1468 | struct page, mapping); | |
1469 | node = node->next; | |
1470 | __free_huge_page(page); | |
1471 | } | |
1472 | } | |
1473 | static DECLARE_WORK(free_hpage_work, free_hpage_workfn); | |
1474 | ||
1475 | void free_huge_page(struct page *page) | |
1476 | { | |
1477 | /* | |
1478 | * Defer freeing if in non-task context to avoid hugetlb_lock deadlock. | |
1479 | */ | |
1480 | if (!in_task()) { | |
1481 | /* | |
1482 | * Only call schedule_work() if hpage_freelist is previously | |
1483 | * empty. Otherwise, schedule_work() had been called but the | |
1484 | * workfn hasn't retrieved the list yet. | |
1485 | */ | |
1486 | if (llist_add((struct llist_node *)&page->mapping, | |
1487 | &hpage_freelist)) | |
1488 | schedule_work(&free_hpage_work); | |
1489 | return; | |
1490 | } | |
1491 | ||
1492 | __free_huge_page(page); | |
1493 | } | |
1494 | ||
a5516438 | 1495 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1496 | { |
0edaecfa | 1497 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1498 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
b7ba30c6 | 1499 | spin_lock(&hugetlb_lock); |
9dd540e2 | 1500 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1501 | set_hugetlb_cgroup_rsvd(page, NULL); |
a5516438 AK |
1502 | h->nr_huge_pages++; |
1503 | h->nr_huge_pages_node[nid]++; | |
b7ba30c6 | 1504 | spin_unlock(&hugetlb_lock); |
b7ba30c6 AK |
1505 | } |
1506 | ||
d00181b9 | 1507 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1508 | { |
1509 | int i; | |
1510 | int nr_pages = 1 << order; | |
1511 | struct page *p = page + 1; | |
1512 | ||
1513 | /* we rely on prep_new_huge_page to set the destructor */ | |
1514 | set_compound_order(page, order); | |
ef5a22be | 1515 | __ClearPageReserved(page); |
de09d31d | 1516 | __SetPageHead(page); |
20a0307c | 1517 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1518 | /* |
1519 | * For gigantic hugepages allocated through bootmem at | |
1520 | * boot, it's safer to be consistent with the not-gigantic | |
1521 | * hugepages and clear the PG_reserved bit from all tail pages | |
1522 | * too. Otherwse drivers using get_user_pages() to access tail | |
1523 | * pages may get the reference counting wrong if they see | |
1524 | * PG_reserved set on a tail page (despite the head page not | |
1525 | * having PG_reserved set). Enforcing this consistency between | |
1526 | * head and tail pages allows drivers to optimize away a check | |
1527 | * on the head page when they need know if put_page() is needed | |
1528 | * after get_user_pages(). | |
1529 | */ | |
1530 | __ClearPageReserved(p); | |
58a84aa9 | 1531 | set_page_count(p, 0); |
1d798ca3 | 1532 | set_compound_head(p, page); |
20a0307c | 1533 | } |
b4330afb | 1534 | atomic_set(compound_mapcount_ptr(page), -1); |
47e29d32 JH |
1535 | |
1536 | if (hpage_pincount_available(page)) | |
1537 | atomic_set(compound_pincount_ptr(page), 0); | |
20a0307c WF |
1538 | } |
1539 | ||
7795912c AM |
1540 | /* |
1541 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1542 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1543 | * details. | |
1544 | */ | |
20a0307c WF |
1545 | int PageHuge(struct page *page) |
1546 | { | |
20a0307c WF |
1547 | if (!PageCompound(page)) |
1548 | return 0; | |
1549 | ||
1550 | page = compound_head(page); | |
f1e61557 | 1551 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1552 | } |
43131e14 NH |
1553 | EXPORT_SYMBOL_GPL(PageHuge); |
1554 | ||
27c73ae7 AA |
1555 | /* |
1556 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1557 | * normal or transparent huge pages. | |
1558 | */ | |
1559 | int PageHeadHuge(struct page *page_head) | |
1560 | { | |
27c73ae7 AA |
1561 | if (!PageHead(page_head)) |
1562 | return 0; | |
1563 | ||
d4af73e3 | 1564 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1565 | } |
27c73ae7 | 1566 | |
c0d0381a MK |
1567 | /* |
1568 | * Find address_space associated with hugetlbfs page. | |
1569 | * Upon entry page is locked and page 'was' mapped although mapped state | |
1570 | * could change. If necessary, use anon_vma to find vma and associated | |
1571 | * address space. The returned mapping may be stale, but it can not be | |
1572 | * invalid as page lock (which is held) is required to destroy mapping. | |
1573 | */ | |
1574 | static struct address_space *_get_hugetlb_page_mapping(struct page *hpage) | |
1575 | { | |
1576 | struct anon_vma *anon_vma; | |
1577 | pgoff_t pgoff_start, pgoff_end; | |
1578 | struct anon_vma_chain *avc; | |
1579 | struct address_space *mapping = page_mapping(hpage); | |
1580 | ||
1581 | /* Simple file based mapping */ | |
1582 | if (mapping) | |
1583 | return mapping; | |
1584 | ||
1585 | /* | |
1586 | * Even anonymous hugetlbfs mappings are associated with an | |
1587 | * underlying hugetlbfs file (see hugetlb_file_setup in mmap | |
1588 | * code). Find a vma associated with the anonymous vma, and | |
1589 | * use the file pointer to get address_space. | |
1590 | */ | |
1591 | anon_vma = page_lock_anon_vma_read(hpage); | |
1592 | if (!anon_vma) | |
1593 | return mapping; /* NULL */ | |
1594 | ||
1595 | /* Use first found vma */ | |
1596 | pgoff_start = page_to_pgoff(hpage); | |
1597 | pgoff_end = pgoff_start + hpage_nr_pages(hpage) - 1; | |
1598 | anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, | |
1599 | pgoff_start, pgoff_end) { | |
1600 | struct vm_area_struct *vma = avc->vma; | |
1601 | ||
1602 | mapping = vma->vm_file->f_mapping; | |
1603 | break; | |
1604 | } | |
1605 | ||
1606 | anon_vma_unlock_read(anon_vma); | |
1607 | return mapping; | |
1608 | } | |
1609 | ||
1610 | /* | |
1611 | * Find and lock address space (mapping) in write mode. | |
1612 | * | |
1613 | * Upon entry, the page is locked which allows us to find the mapping | |
1614 | * even in the case of an anon page. However, locking order dictates | |
1615 | * the i_mmap_rwsem be acquired BEFORE the page lock. This is hugetlbfs | |
1616 | * specific. So, we first try to lock the sema while still holding the | |
1617 | * page lock. If this works, great! If not, then we need to drop the | |
1618 | * page lock and then acquire i_mmap_rwsem and reacquire page lock. Of | |
1619 | * course, need to revalidate state along the way. | |
1620 | */ | |
1621 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1622 | { | |
1623 | struct address_space *mapping, *mapping2; | |
1624 | ||
1625 | mapping = _get_hugetlb_page_mapping(hpage); | |
1626 | retry: | |
1627 | if (!mapping) | |
1628 | return mapping; | |
1629 | ||
1630 | /* | |
1631 | * If no contention, take lock and return | |
1632 | */ | |
1633 | if (i_mmap_trylock_write(mapping)) | |
1634 | return mapping; | |
1635 | ||
1636 | /* | |
1637 | * Must drop page lock and wait on mapping sema. | |
1638 | * Note: Once page lock is dropped, mapping could become invalid. | |
1639 | * As a hack, increase map count until we lock page again. | |
1640 | */ | |
1641 | atomic_inc(&hpage->_mapcount); | |
1642 | unlock_page(hpage); | |
1643 | i_mmap_lock_write(mapping); | |
1644 | lock_page(hpage); | |
1645 | atomic_add_negative(-1, &hpage->_mapcount); | |
1646 | ||
1647 | /* verify page is still mapped */ | |
1648 | if (!page_mapped(hpage)) { | |
1649 | i_mmap_unlock_write(mapping); | |
1650 | return NULL; | |
1651 | } | |
1652 | ||
1653 | /* | |
1654 | * Get address space again and verify it is the same one | |
1655 | * we locked. If not, drop lock and retry. | |
1656 | */ | |
1657 | mapping2 = _get_hugetlb_page_mapping(hpage); | |
1658 | if (mapping2 != mapping) { | |
1659 | i_mmap_unlock_write(mapping); | |
1660 | mapping = mapping2; | |
1661 | goto retry; | |
1662 | } | |
1663 | ||
1664 | return mapping; | |
1665 | } | |
1666 | ||
13d60f4b ZY |
1667 | pgoff_t __basepage_index(struct page *page) |
1668 | { | |
1669 | struct page *page_head = compound_head(page); | |
1670 | pgoff_t index = page_index(page_head); | |
1671 | unsigned long compound_idx; | |
1672 | ||
1673 | if (!PageHuge(page_head)) | |
1674 | return page_index(page); | |
1675 | ||
1676 | if (compound_order(page_head) >= MAX_ORDER) | |
1677 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1678 | else | |
1679 | compound_idx = page - page_head; | |
1680 | ||
1681 | return (index << compound_order(page_head)) + compound_idx; | |
1682 | } | |
1683 | ||
0c397dae | 1684 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1685 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1686 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1687 | { |
af0fb9df | 1688 | int order = huge_page_order(h); |
1da177e4 | 1689 | struct page *page; |
f60858f9 | 1690 | bool alloc_try_hard = true; |
f96efd58 | 1691 | |
f60858f9 MK |
1692 | /* |
1693 | * By default we always try hard to allocate the page with | |
1694 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1695 | * a loop (to adjust global huge page counts) and previous allocation | |
1696 | * failed, do not continue to try hard on the same node. Use the | |
1697 | * node_alloc_noretry bitmap to manage this state information. | |
1698 | */ | |
1699 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1700 | alloc_try_hard = false; | |
1701 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1702 | if (alloc_try_hard) | |
1703 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1704 | if (nid == NUMA_NO_NODE) |
1705 | nid = numa_mem_id(); | |
1706 | page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask); | |
1707 | if (page) | |
1708 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1709 | else | |
1710 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1711 | |
f60858f9 MK |
1712 | /* |
1713 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1714 | * indicates an overall state change. Clear bit so that we resume | |
1715 | * normal 'try hard' allocations. | |
1716 | */ | |
1717 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1718 | node_clear(nid, *node_alloc_noretry); | |
1719 | ||
1720 | /* | |
1721 | * If we tried hard to get a page but failed, set bit so that | |
1722 | * subsequent attempts will not try as hard until there is an | |
1723 | * overall state change. | |
1724 | */ | |
1725 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1726 | node_set(nid, *node_alloc_noretry); | |
1727 | ||
63b4613c NA |
1728 | return page; |
1729 | } | |
1730 | ||
0c397dae MH |
1731 | /* |
1732 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1733 | * should use this function to get new hugetlb pages | |
1734 | */ | |
1735 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1736 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1737 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1738 | { |
1739 | struct page *page; | |
1740 | ||
1741 | if (hstate_is_gigantic(h)) | |
1742 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1743 | else | |
1744 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1745 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1746 | if (!page) |
1747 | return NULL; | |
1748 | ||
1749 | if (hstate_is_gigantic(h)) | |
1750 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1751 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1752 | ||
1753 | return page; | |
1754 | } | |
1755 | ||
af0fb9df MH |
1756 | /* |
1757 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1758 | * manner. | |
1759 | */ | |
f60858f9 MK |
1760 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1761 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1762 | { |
1763 | struct page *page; | |
1764 | int nr_nodes, node; | |
af0fb9df | 1765 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1766 | |
1767 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1768 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1769 | node_alloc_noretry); | |
af0fb9df | 1770 | if (page) |
b2261026 | 1771 | break; |
b2261026 JK |
1772 | } |
1773 | ||
af0fb9df MH |
1774 | if (!page) |
1775 | return 0; | |
b2261026 | 1776 | |
af0fb9df MH |
1777 | put_page(page); /* free it into the hugepage allocator */ |
1778 | ||
1779 | return 1; | |
b2261026 JK |
1780 | } |
1781 | ||
e8c5c824 LS |
1782 | /* |
1783 | * Free huge page from pool from next node to free. | |
1784 | * Attempt to keep persistent huge pages more or less | |
1785 | * balanced over allowed nodes. | |
1786 | * Called with hugetlb_lock locked. | |
1787 | */ | |
6ae11b27 LS |
1788 | static int free_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1789 | bool acct_surplus) | |
e8c5c824 | 1790 | { |
b2261026 | 1791 | int nr_nodes, node; |
e8c5c824 LS |
1792 | int ret = 0; |
1793 | ||
b2261026 | 1794 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1795 | /* |
1796 | * If we're returning unused surplus pages, only examine | |
1797 | * nodes with surplus pages. | |
1798 | */ | |
b2261026 JK |
1799 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1800 | !list_empty(&h->hugepage_freelists[node])) { | |
e8c5c824 | 1801 | struct page *page = |
b2261026 | 1802 | list_entry(h->hugepage_freelists[node].next, |
e8c5c824 LS |
1803 | struct page, lru); |
1804 | list_del(&page->lru); | |
1805 | h->free_huge_pages--; | |
b2261026 | 1806 | h->free_huge_pages_node[node]--; |
685f3457 LS |
1807 | if (acct_surplus) { |
1808 | h->surplus_huge_pages--; | |
b2261026 | 1809 | h->surplus_huge_pages_node[node]--; |
685f3457 | 1810 | } |
e8c5c824 LS |
1811 | update_and_free_page(h, page); |
1812 | ret = 1; | |
9a76db09 | 1813 | break; |
e8c5c824 | 1814 | } |
b2261026 | 1815 | } |
e8c5c824 LS |
1816 | |
1817 | return ret; | |
1818 | } | |
1819 | ||
c8721bbb NH |
1820 | /* |
1821 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1822 | * nothing for in-use hugepages and non-hugepages. |
1823 | * This function returns values like below: | |
1824 | * | |
1825 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1826 | * (allocated or reserved.) | |
1827 | * 0: successfully dissolved free hugepages or the page is not a | |
1828 | * hugepage (considered as already dissolved) | |
c8721bbb | 1829 | */ |
c3114a84 | 1830 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1831 | { |
6bc9b564 | 1832 | int rc = -EBUSY; |
082d5b6b | 1833 | |
faf53def NH |
1834 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1835 | if (!PageHuge(page)) | |
1836 | return 0; | |
1837 | ||
c8721bbb | 1838 | spin_lock(&hugetlb_lock); |
faf53def NH |
1839 | if (!PageHuge(page)) { |
1840 | rc = 0; | |
1841 | goto out; | |
1842 | } | |
1843 | ||
1844 | if (!page_count(page)) { | |
2247bb33 GS |
1845 | struct page *head = compound_head(page); |
1846 | struct hstate *h = page_hstate(head); | |
1847 | int nid = page_to_nid(head); | |
6bc9b564 | 1848 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1849 | goto out; |
c3114a84 AK |
1850 | /* |
1851 | * Move PageHWPoison flag from head page to the raw error page, | |
1852 | * which makes any subpages rather than the error page reusable. | |
1853 | */ | |
1854 | if (PageHWPoison(head) && page != head) { | |
1855 | SetPageHWPoison(page); | |
1856 | ClearPageHWPoison(head); | |
1857 | } | |
2247bb33 | 1858 | list_del(&head->lru); |
c8721bbb NH |
1859 | h->free_huge_pages--; |
1860 | h->free_huge_pages_node[nid]--; | |
c1470b33 | 1861 | h->max_huge_pages--; |
2247bb33 | 1862 | update_and_free_page(h, head); |
6bc9b564 | 1863 | rc = 0; |
c8721bbb | 1864 | } |
082d5b6b | 1865 | out: |
c8721bbb | 1866 | spin_unlock(&hugetlb_lock); |
082d5b6b | 1867 | return rc; |
c8721bbb NH |
1868 | } |
1869 | ||
1870 | /* | |
1871 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1872 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1873 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1874 | * part of it lies within the given range. | |
082d5b6b GS |
1875 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1876 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1877 | */ |
082d5b6b | 1878 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1879 | { |
c8721bbb | 1880 | unsigned long pfn; |
eb03aa00 | 1881 | struct page *page; |
082d5b6b | 1882 | int rc = 0; |
c8721bbb | 1883 | |
d0177639 | 1884 | if (!hugepages_supported()) |
082d5b6b | 1885 | return rc; |
d0177639 | 1886 | |
eb03aa00 GS |
1887 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1888 | page = pfn_to_page(pfn); | |
faf53def NH |
1889 | rc = dissolve_free_huge_page(page); |
1890 | if (rc) | |
1891 | break; | |
eb03aa00 | 1892 | } |
082d5b6b GS |
1893 | |
1894 | return rc; | |
c8721bbb NH |
1895 | } |
1896 | ||
ab5ac90a MH |
1897 | /* |
1898 | * Allocates a fresh surplus page from the page allocator. | |
1899 | */ | |
0c397dae | 1900 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1901 | int nid, nodemask_t *nmask) |
7893d1d5 | 1902 | { |
9980d744 | 1903 | struct page *page = NULL; |
7893d1d5 | 1904 | |
bae7f4ae | 1905 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1906 | return NULL; |
1907 | ||
d1c3fb1f | 1908 | spin_lock(&hugetlb_lock); |
9980d744 MH |
1909 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1910 | goto out_unlock; | |
d1c3fb1f NA |
1911 | spin_unlock(&hugetlb_lock); |
1912 | ||
f60858f9 | 1913 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1914 | if (!page) |
0c397dae | 1915 | return NULL; |
d1c3fb1f NA |
1916 | |
1917 | spin_lock(&hugetlb_lock); | |
9980d744 MH |
1918 | /* |
1919 | * We could have raced with the pool size change. | |
1920 | * Double check that and simply deallocate the new page | |
1921 | * if we would end up overcommiting the surpluses. Abuse | |
1922 | * temporary page to workaround the nasty free_huge_page | |
1923 | * codeflow | |
1924 | */ | |
1925 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
1926 | SetPageHugeTemporary(page); | |
2bf753e6 | 1927 | spin_unlock(&hugetlb_lock); |
9980d744 | 1928 | put_page(page); |
2bf753e6 | 1929 | return NULL; |
9980d744 | 1930 | } else { |
9980d744 | 1931 | h->surplus_huge_pages++; |
4704dea3 | 1932 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1933 | } |
9980d744 MH |
1934 | |
1935 | out_unlock: | |
d1c3fb1f | 1936 | spin_unlock(&hugetlb_lock); |
7893d1d5 AL |
1937 | |
1938 | return page; | |
1939 | } | |
1940 | ||
9a4e9f3b AK |
1941 | struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
1942 | int nid, nodemask_t *nmask) | |
ab5ac90a MH |
1943 | { |
1944 | struct page *page; | |
1945 | ||
1946 | if (hstate_is_gigantic(h)) | |
1947 | return NULL; | |
1948 | ||
f60858f9 | 1949 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1950 | if (!page) |
1951 | return NULL; | |
1952 | ||
1953 | /* | |
1954 | * We do not account these pages as surplus because they are only | |
1955 | * temporary and will be released properly on the last reference | |
1956 | */ | |
ab5ac90a MH |
1957 | SetPageHugeTemporary(page); |
1958 | ||
1959 | return page; | |
1960 | } | |
1961 | ||
099730d6 DH |
1962 | /* |
1963 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1964 | */ | |
e0ec90ee | 1965 | static |
0c397dae | 1966 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1967 | struct vm_area_struct *vma, unsigned long addr) |
1968 | { | |
aaf14e40 MH |
1969 | struct page *page; |
1970 | struct mempolicy *mpol; | |
1971 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1972 | int nid; | |
1973 | nodemask_t *nodemask; | |
1974 | ||
1975 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1976 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1977 | mpol_cond_put(mpol); |
1978 | ||
1979 | return page; | |
099730d6 DH |
1980 | } |
1981 | ||
ab5ac90a | 1982 | /* page migration callback function */ |
bf50bab2 NH |
1983 | struct page *alloc_huge_page_node(struct hstate *h, int nid) |
1984 | { | |
aaf14e40 | 1985 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4ef91848 | 1986 | struct page *page = NULL; |
bf50bab2 | 1987 | |
aaf14e40 MH |
1988 | if (nid != NUMA_NO_NODE) |
1989 | gfp_mask |= __GFP_THISNODE; | |
1990 | ||
bf50bab2 | 1991 | spin_lock(&hugetlb_lock); |
4ef91848 | 1992 | if (h->free_huge_pages - h->resv_huge_pages > 0) |
3e59fcb0 | 1993 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1994 | spin_unlock(&hugetlb_lock); |
1995 | ||
94ae8ba7 | 1996 | if (!page) |
0c397dae | 1997 | page = alloc_migrate_huge_page(h, gfp_mask, nid, NULL); |
bf50bab2 NH |
1998 | |
1999 | return page; | |
2000 | } | |
2001 | ||
ab5ac90a | 2002 | /* page migration callback function */ |
3e59fcb0 MH |
2003 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
2004 | nodemask_t *nmask) | |
4db9b2ef | 2005 | { |
aaf14e40 | 2006 | gfp_t gfp_mask = htlb_alloc_mask(h); |
4db9b2ef MH |
2007 | |
2008 | spin_lock(&hugetlb_lock); | |
2009 | if (h->free_huge_pages - h->resv_huge_pages > 0) { | |
3e59fcb0 MH |
2010 | struct page *page; |
2011 | ||
2012 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
2013 | if (page) { | |
2014 | spin_unlock(&hugetlb_lock); | |
2015 | return page; | |
4db9b2ef MH |
2016 | } |
2017 | } | |
2018 | spin_unlock(&hugetlb_lock); | |
4db9b2ef | 2019 | |
0c397dae | 2020 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
2021 | } |
2022 | ||
ebd63723 | 2023 | /* mempolicy aware migration callback */ |
389c8178 MH |
2024 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
2025 | unsigned long address) | |
ebd63723 MH |
2026 | { |
2027 | struct mempolicy *mpol; | |
2028 | nodemask_t *nodemask; | |
2029 | struct page *page; | |
ebd63723 MH |
2030 | gfp_t gfp_mask; |
2031 | int node; | |
2032 | ||
ebd63723 MH |
2033 | gfp_mask = htlb_alloc_mask(h); |
2034 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
2035 | page = alloc_huge_page_nodemask(h, node, nodemask); | |
2036 | mpol_cond_put(mpol); | |
2037 | ||
2038 | return page; | |
2039 | } | |
2040 | ||
e4e574b7 | 2041 | /* |
25985edc | 2042 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
2043 | * of size 'delta'. |
2044 | */ | |
a5516438 | 2045 | static int gather_surplus_pages(struct hstate *h, int delta) |
1b2a1e7b | 2046 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
2047 | { |
2048 | struct list_head surplus_list; | |
2049 | struct page *page, *tmp; | |
2050 | int ret, i; | |
2051 | int needed, allocated; | |
28073b02 | 2052 | bool alloc_ok = true; |
e4e574b7 | 2053 | |
a5516438 | 2054 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 2055 | if (needed <= 0) { |
a5516438 | 2056 | h->resv_huge_pages += delta; |
e4e574b7 | 2057 | return 0; |
ac09b3a1 | 2058 | } |
e4e574b7 AL |
2059 | |
2060 | allocated = 0; | |
2061 | INIT_LIST_HEAD(&surplus_list); | |
2062 | ||
2063 | ret = -ENOMEM; | |
2064 | retry: | |
2065 | spin_unlock(&hugetlb_lock); | |
2066 | for (i = 0; i < needed; i++) { | |
0c397dae | 2067 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 2068 | NUMA_NO_NODE, NULL); |
28073b02 HD |
2069 | if (!page) { |
2070 | alloc_ok = false; | |
2071 | break; | |
2072 | } | |
e4e574b7 | 2073 | list_add(&page->lru, &surplus_list); |
69ed779a | 2074 | cond_resched(); |
e4e574b7 | 2075 | } |
28073b02 | 2076 | allocated += i; |
e4e574b7 AL |
2077 | |
2078 | /* | |
2079 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
2080 | * because either resv_huge_pages or free_huge_pages may have changed. | |
2081 | */ | |
2082 | spin_lock(&hugetlb_lock); | |
a5516438 AK |
2083 | needed = (h->resv_huge_pages + delta) - |
2084 | (h->free_huge_pages + allocated); | |
28073b02 HD |
2085 | if (needed > 0) { |
2086 | if (alloc_ok) | |
2087 | goto retry; | |
2088 | /* | |
2089 | * We were not able to allocate enough pages to | |
2090 | * satisfy the entire reservation so we free what | |
2091 | * we've allocated so far. | |
2092 | */ | |
2093 | goto free; | |
2094 | } | |
e4e574b7 AL |
2095 | /* |
2096 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2097 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2098 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2099 | * allocator. Commit the entire reservation here to prevent another |
2100 | * process from stealing the pages as they are added to the pool but | |
2101 | * before they are reserved. | |
e4e574b7 AL |
2102 | */ |
2103 | needed += allocated; | |
a5516438 | 2104 | h->resv_huge_pages += delta; |
e4e574b7 | 2105 | ret = 0; |
a9869b83 | 2106 | |
19fc3f0a | 2107 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2108 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
19fc3f0a AL |
2109 | if ((--needed) < 0) |
2110 | break; | |
a9869b83 NH |
2111 | /* |
2112 | * This page is now managed by the hugetlb allocator and has | |
2113 | * no users -- drop the buddy allocator's reference. | |
2114 | */ | |
2115 | put_page_testzero(page); | |
309381fe | 2116 | VM_BUG_ON_PAGE(page_count(page), page); |
a5516438 | 2117 | enqueue_huge_page(h, page); |
19fc3f0a | 2118 | } |
28073b02 | 2119 | free: |
b0365c8d | 2120 | spin_unlock(&hugetlb_lock); |
19fc3f0a AL |
2121 | |
2122 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
2123 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
2124 | put_page(page); | |
a9869b83 | 2125 | spin_lock(&hugetlb_lock); |
e4e574b7 AL |
2126 | |
2127 | return ret; | |
2128 | } | |
2129 | ||
2130 | /* | |
e5bbc8a6 MK |
2131 | * This routine has two main purposes: |
2132 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2133 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2134 | * to the associated reservation map. | |
2135 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2136 | * the reservation. As many as unused_resv_pages may be freed. | |
2137 | * | |
2138 | * Called with hugetlb_lock held. However, the lock could be dropped (and | |
2139 | * reacquired) during calls to cond_resched_lock. Whenever dropping the lock, | |
2140 | * we must make sure nobody else can claim pages we are in the process of | |
2141 | * freeing. Do this by ensuring resv_huge_page always is greater than the | |
2142 | * number of huge pages we plan to free when dropping the lock. | |
e4e574b7 | 2143 | */ |
a5516438 AK |
2144 | static void return_unused_surplus_pages(struct hstate *h, |
2145 | unsigned long unused_resv_pages) | |
e4e574b7 | 2146 | { |
e4e574b7 AL |
2147 | unsigned long nr_pages; |
2148 | ||
aa888a74 | 2149 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2150 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2151 | goto out; |
aa888a74 | 2152 | |
e5bbc8a6 MK |
2153 | /* |
2154 | * Part (or even all) of the reservation could have been backed | |
2155 | * by pre-allocated pages. Only free surplus pages. | |
2156 | */ | |
a5516438 | 2157 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2158 | |
685f3457 LS |
2159 | /* |
2160 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2161 | * evenly across all nodes with memory. Iterate across these nodes |
2162 | * until we can no longer free unreserved surplus pages. This occurs | |
2163 | * when the nodes with surplus pages have no free pages. | |
2164 | * free_pool_huge_page() will balance the the freed pages across the | |
2165 | * on-line nodes with memory and will handle the hstate accounting. | |
e5bbc8a6 MK |
2166 | * |
2167 | * Note that we decrement resv_huge_pages as we free the pages. If | |
2168 | * we drop the lock, resv_huge_pages will still be sufficiently large | |
2169 | * to cover subsequent pages we may free. | |
685f3457 LS |
2170 | */ |
2171 | while (nr_pages--) { | |
e5bbc8a6 MK |
2172 | h->resv_huge_pages--; |
2173 | unused_resv_pages--; | |
8cebfcd0 | 2174 | if (!free_pool_huge_page(h, &node_states[N_MEMORY], 1)) |
e5bbc8a6 | 2175 | goto out; |
7848a4bf | 2176 | cond_resched_lock(&hugetlb_lock); |
e4e574b7 | 2177 | } |
e5bbc8a6 MK |
2178 | |
2179 | out: | |
2180 | /* Fully uncommit the reservation */ | |
2181 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 AL |
2182 | } |
2183 | ||
5e911373 | 2184 | |
c37f9fb1 | 2185 | /* |
feba16e2 | 2186 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2187 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2188 | * |
2189 | * vma_needs_reservation is called to determine if the huge page at addr | |
2190 | * within the vma has an associated reservation. If a reservation is | |
2191 | * needed, the value 1 is returned. The caller is then responsible for | |
2192 | * managing the global reservation and subpool usage counts. After | |
2193 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2194 | * to add the page to the reservation map. If the page allocation fails, |
2195 | * the reservation must be ended instead of committed. vma_end_reservation | |
2196 | * is called in such cases. | |
cf3ad20b MK |
2197 | * |
2198 | * In the normal case, vma_commit_reservation returns the same value | |
2199 | * as the preceding vma_needs_reservation call. The only time this | |
2200 | * is not the case is if a reserve map was changed between calls. It | |
2201 | * is the responsibility of the caller to notice the difference and | |
2202 | * take appropriate action. | |
96b96a96 MK |
2203 | * |
2204 | * vma_add_reservation is used in error paths where a reservation must | |
2205 | * be restored when a newly allocated huge page must be freed. It is | |
2206 | * to be called after calling vma_needs_reservation to determine if a | |
2207 | * reservation exists. | |
c37f9fb1 | 2208 | */ |
5e911373 MK |
2209 | enum vma_resv_mode { |
2210 | VMA_NEEDS_RESV, | |
2211 | VMA_COMMIT_RESV, | |
feba16e2 | 2212 | VMA_END_RESV, |
96b96a96 | 2213 | VMA_ADD_RESV, |
5e911373 | 2214 | }; |
cf3ad20b MK |
2215 | static long __vma_reservation_common(struct hstate *h, |
2216 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2217 | enum vma_resv_mode mode) |
c37f9fb1 | 2218 | { |
4e35f483 JK |
2219 | struct resv_map *resv; |
2220 | pgoff_t idx; | |
cf3ad20b | 2221 | long ret; |
0db9d74e | 2222 | long dummy_out_regions_needed; |
c37f9fb1 | 2223 | |
4e35f483 JK |
2224 | resv = vma_resv_map(vma); |
2225 | if (!resv) | |
84afd99b | 2226 | return 1; |
c37f9fb1 | 2227 | |
4e35f483 | 2228 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2229 | switch (mode) { |
2230 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2231 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2232 | /* We assume that vma_reservation_* routines always operate on | |
2233 | * 1 page, and that adding to resv map a 1 page entry can only | |
2234 | * ever require 1 region. | |
2235 | */ | |
2236 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2237 | break; |
2238 | case VMA_COMMIT_RESV: | |
075a61d0 | 2239 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2240 | /* region_add calls of range 1 should never fail. */ |
2241 | VM_BUG_ON(ret < 0); | |
5e911373 | 2242 | break; |
feba16e2 | 2243 | case VMA_END_RESV: |
0db9d74e | 2244 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2245 | ret = 0; |
2246 | break; | |
96b96a96 | 2247 | case VMA_ADD_RESV: |
0db9d74e | 2248 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2249 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2250 | /* region_add calls of range 1 should never fail. */ |
2251 | VM_BUG_ON(ret < 0); | |
2252 | } else { | |
2253 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2254 | ret = region_del(resv, idx, idx + 1); |
2255 | } | |
2256 | break; | |
5e911373 MK |
2257 | default: |
2258 | BUG(); | |
2259 | } | |
84afd99b | 2260 | |
4e35f483 | 2261 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 2262 | return ret; |
67961f9d MK |
2263 | else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && ret >= 0) { |
2264 | /* | |
2265 | * In most cases, reserves always exist for private mappings. | |
2266 | * However, a file associated with mapping could have been | |
2267 | * hole punched or truncated after reserves were consumed. | |
2268 | * As subsequent fault on such a range will not use reserves. | |
2269 | * Subtle - The reserve map for private mappings has the | |
2270 | * opposite meaning than that of shared mappings. If NO | |
2271 | * entry is in the reserve map, it means a reservation exists. | |
2272 | * If an entry exists in the reserve map, it means the | |
2273 | * reservation has already been consumed. As a result, the | |
2274 | * return value of this routine is the opposite of the | |
2275 | * value returned from reserve map manipulation routines above. | |
2276 | */ | |
2277 | if (ret) | |
2278 | return 0; | |
2279 | else | |
2280 | return 1; | |
2281 | } | |
4e35f483 | 2282 | else |
cf3ad20b | 2283 | return ret < 0 ? ret : 0; |
c37f9fb1 | 2284 | } |
cf3ad20b MK |
2285 | |
2286 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2287 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2288 | { |
5e911373 | 2289 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2290 | } |
84afd99b | 2291 | |
cf3ad20b MK |
2292 | static long vma_commit_reservation(struct hstate *h, |
2293 | struct vm_area_struct *vma, unsigned long addr) | |
2294 | { | |
5e911373 MK |
2295 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2296 | } | |
2297 | ||
feba16e2 | 2298 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2299 | struct vm_area_struct *vma, unsigned long addr) |
2300 | { | |
feba16e2 | 2301 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2302 | } |
2303 | ||
96b96a96 MK |
2304 | static long vma_add_reservation(struct hstate *h, |
2305 | struct vm_area_struct *vma, unsigned long addr) | |
2306 | { | |
2307 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2308 | } | |
2309 | ||
2310 | /* | |
2311 | * This routine is called to restore a reservation on error paths. In the | |
2312 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2313 | * and is about to be freed. If a reservation for the page existed, | |
2314 | * alloc_huge_page would have consumed the reservation and set PagePrivate | |
2315 | * in the newly allocated page. When the page is freed via free_huge_page, | |
2316 | * the global reservation count will be incremented if PagePrivate is set. | |
2317 | * However, free_huge_page can not adjust the reserve map. Adjust the | |
2318 | * reserve map here to be consistent with global reserve count adjustments | |
2319 | * to be made by free_huge_page. | |
2320 | */ | |
2321 | static void restore_reserve_on_error(struct hstate *h, | |
2322 | struct vm_area_struct *vma, unsigned long address, | |
2323 | struct page *page) | |
2324 | { | |
2325 | if (unlikely(PagePrivate(page))) { | |
2326 | long rc = vma_needs_reservation(h, vma, address); | |
2327 | ||
2328 | if (unlikely(rc < 0)) { | |
2329 | /* | |
2330 | * Rare out of memory condition in reserve map | |
2331 | * manipulation. Clear PagePrivate so that | |
2332 | * global reserve count will not be incremented | |
2333 | * by free_huge_page. This will make it appear | |
2334 | * as though the reservation for this page was | |
2335 | * consumed. This may prevent the task from | |
2336 | * faulting in the page at a later time. This | |
2337 | * is better than inconsistent global huge page | |
2338 | * accounting of reserve counts. | |
2339 | */ | |
2340 | ClearPagePrivate(page); | |
2341 | } else if (rc) { | |
2342 | rc = vma_add_reservation(h, vma, address); | |
2343 | if (unlikely(rc < 0)) | |
2344 | /* | |
2345 | * See above comment about rare out of | |
2346 | * memory condition. | |
2347 | */ | |
2348 | ClearPagePrivate(page); | |
2349 | } else | |
2350 | vma_end_reservation(h, vma, address); | |
2351 | } | |
2352 | } | |
2353 | ||
70c3547e | 2354 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2355 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2356 | { |
90481622 | 2357 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2358 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2359 | struct page *page; |
d85f69b0 MK |
2360 | long map_chg, map_commit; |
2361 | long gbl_chg; | |
6d76dcf4 AK |
2362 | int ret, idx; |
2363 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2364 | bool deferred_reserve; |
a1e78772 | 2365 | |
6d76dcf4 | 2366 | idx = hstate_index(h); |
a1e78772 | 2367 | /* |
d85f69b0 MK |
2368 | * Examine the region/reserve map to determine if the process |
2369 | * has a reservation for the page to be allocated. A return | |
2370 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2371 | */ |
d85f69b0 MK |
2372 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2373 | if (map_chg < 0) | |
76dcee75 | 2374 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2375 | |
2376 | /* | |
2377 | * Processes that did not create the mapping will have no | |
2378 | * reserves as indicated by the region/reserve map. Check | |
2379 | * that the allocation will not exceed the subpool limit. | |
2380 | * Allocations for MAP_NORESERVE mappings also need to be | |
2381 | * checked against any subpool limit. | |
2382 | */ | |
2383 | if (map_chg || avoid_reserve) { | |
2384 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2385 | if (gbl_chg < 0) { | |
feba16e2 | 2386 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2387 | return ERR_PTR(-ENOSPC); |
5e911373 | 2388 | } |
1da177e4 | 2389 | |
d85f69b0 MK |
2390 | /* |
2391 | * Even though there was no reservation in the region/reserve | |
2392 | * map, there could be reservations associated with the | |
2393 | * subpool that can be used. This would be indicated if the | |
2394 | * return value of hugepage_subpool_get_pages() is zero. | |
2395 | * However, if avoid_reserve is specified we still avoid even | |
2396 | * the subpool reservations. | |
2397 | */ | |
2398 | if (avoid_reserve) | |
2399 | gbl_chg = 1; | |
2400 | } | |
2401 | ||
08cf9faf MA |
2402 | /* If this allocation is not consuming a reservation, charge it now. |
2403 | */ | |
2404 | deferred_reserve = map_chg || avoid_reserve || !vma_resv_map(vma); | |
2405 | if (deferred_reserve) { | |
2406 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2407 | idx, pages_per_huge_page(h), &h_cg); | |
2408 | if (ret) | |
2409 | goto out_subpool_put; | |
2410 | } | |
2411 | ||
6d76dcf4 | 2412 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2413 | if (ret) |
08cf9faf | 2414 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2415 | |
1da177e4 | 2416 | spin_lock(&hugetlb_lock); |
d85f69b0 MK |
2417 | /* |
2418 | * glb_chg is passed to indicate whether or not a page must be taken | |
2419 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2420 | * a reservation exists for the allocation. | |
2421 | */ | |
2422 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2423 | if (!page) { |
94ae8ba7 | 2424 | spin_unlock(&hugetlb_lock); |
0c397dae | 2425 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2426 | if (!page) |
2427 | goto out_uncharge_cgroup; | |
a88c7695 NH |
2428 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
2429 | SetPagePrivate(page); | |
2430 | h->resv_huge_pages--; | |
2431 | } | |
79dbb236 AK |
2432 | spin_lock(&hugetlb_lock); |
2433 | list_move(&page->lru, &h->hugepage_activelist); | |
81a6fcae | 2434 | /* Fall through */ |
68842c9b | 2435 | } |
81a6fcae | 2436 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2437 | /* If allocation is not consuming a reservation, also store the |
2438 | * hugetlb_cgroup pointer on the page. | |
2439 | */ | |
2440 | if (deferred_reserve) { | |
2441 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2442 | h_cg, page); | |
2443 | } | |
2444 | ||
81a6fcae | 2445 | spin_unlock(&hugetlb_lock); |
348ea204 | 2446 | |
90481622 | 2447 | set_page_private(page, (unsigned long)spool); |
90d8b7e6 | 2448 | |
d85f69b0 MK |
2449 | map_commit = vma_commit_reservation(h, vma, addr); |
2450 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2451 | /* |
2452 | * The page was added to the reservation map between | |
2453 | * vma_needs_reservation and vma_commit_reservation. | |
2454 | * This indicates a race with hugetlb_reserve_pages. | |
2455 | * Adjust for the subpool count incremented above AND | |
2456 | * in hugetlb_reserve_pages for the same page. Also, | |
2457 | * the reservation count added in hugetlb_reserve_pages | |
2458 | * no longer applies. | |
2459 | */ | |
2460 | long rsv_adjust; | |
2461 | ||
2462 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2463 | hugetlb_acct_memory(h, -rsv_adjust); | |
2464 | } | |
90d8b7e6 | 2465 | return page; |
8f34af6f JZ |
2466 | |
2467 | out_uncharge_cgroup: | |
2468 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2469 | out_uncharge_cgroup_reservation: |
2470 | if (deferred_reserve) | |
2471 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2472 | h_cg); | |
8f34af6f | 2473 | out_subpool_put: |
d85f69b0 | 2474 | if (map_chg || avoid_reserve) |
8f34af6f | 2475 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2476 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2477 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2478 | } |
2479 | ||
e24a1307 AK |
2480 | int alloc_bootmem_huge_page(struct hstate *h) |
2481 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2482 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2483 | { |
2484 | struct huge_bootmem_page *m; | |
b2261026 | 2485 | int nr_nodes, node; |
aa888a74 | 2486 | |
b2261026 | 2487 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2488 | void *addr; |
2489 | ||
eb31d559 | 2490 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2491 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2492 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2493 | if (addr) { |
2494 | /* | |
2495 | * Use the beginning of the huge page to store the | |
2496 | * huge_bootmem_page struct (until gather_bootmem | |
2497 | * puts them into the mem_map). | |
2498 | */ | |
2499 | m = addr; | |
91f47662 | 2500 | goto found; |
aa888a74 | 2501 | } |
aa888a74 AK |
2502 | } |
2503 | return 0; | |
2504 | ||
2505 | found: | |
df994ead | 2506 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2507 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2508 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2509 | list_add(&m->list, &huge_boot_pages); |
2510 | m->hstate = h; | |
2511 | return 1; | |
2512 | } | |
2513 | ||
d00181b9 KS |
2514 | static void __init prep_compound_huge_page(struct page *page, |
2515 | unsigned int order) | |
18229df5 AW |
2516 | { |
2517 | if (unlikely(order > (MAX_ORDER - 1))) | |
2518 | prep_compound_gigantic_page(page, order); | |
2519 | else | |
2520 | prep_compound_page(page, order); | |
2521 | } | |
2522 | ||
aa888a74 AK |
2523 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2524 | static void __init gather_bootmem_prealloc(void) | |
2525 | { | |
2526 | struct huge_bootmem_page *m; | |
2527 | ||
2528 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2529 | struct page *page = virt_to_page(m); |
aa888a74 | 2530 | struct hstate *h = m->hstate; |
ee8f248d | 2531 | |
aa888a74 | 2532 | WARN_ON(page_count(page) != 1); |
18229df5 | 2533 | prep_compound_huge_page(page, h->order); |
ef5a22be | 2534 | WARN_ON(PageReserved(page)); |
aa888a74 | 2535 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2536 | put_page(page); /* free it into the hugepage allocator */ |
2537 | ||
b0320c7b RA |
2538 | /* |
2539 | * If we had gigantic hugepages allocated at boot time, we need | |
2540 | * to restore the 'stolen' pages to totalram_pages in order to | |
2541 | * fix confusing memory reports from free(1) and another | |
2542 | * side-effects, like CommitLimit going negative. | |
2543 | */ | |
bae7f4ae | 2544 | if (hstate_is_gigantic(h)) |
3dcc0571 | 2545 | adjust_managed_page_count(page, 1 << h->order); |
520495fe | 2546 | cond_resched(); |
aa888a74 AK |
2547 | } |
2548 | } | |
2549 | ||
8faa8b07 | 2550 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2551 | { |
2552 | unsigned long i; | |
f60858f9 MK |
2553 | nodemask_t *node_alloc_noretry; |
2554 | ||
2555 | if (!hstate_is_gigantic(h)) { | |
2556 | /* | |
2557 | * Bit mask controlling how hard we retry per-node allocations. | |
2558 | * Ignore errors as lower level routines can deal with | |
2559 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2560 | * time, we are likely in bigger trouble. | |
2561 | */ | |
2562 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2563 | GFP_KERNEL); | |
2564 | } else { | |
2565 | /* allocations done at boot time */ | |
2566 | node_alloc_noretry = NULL; | |
2567 | } | |
2568 | ||
2569 | /* bit mask controlling how hard we retry per-node allocations */ | |
2570 | if (node_alloc_noretry) | |
2571 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2572 | |
e5ff2159 | 2573 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2574 | if (hstate_is_gigantic(h)) { |
cf11e85f RG |
2575 | if (IS_ENABLED(CONFIG_CMA) && hugetlb_cma[0]) { |
2576 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); | |
2577 | break; | |
2578 | } | |
aa888a74 AK |
2579 | if (!alloc_bootmem_huge_page(h)) |
2580 | break; | |
0c397dae | 2581 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2582 | &node_states[N_MEMORY], |
2583 | node_alloc_noretry)) | |
1da177e4 | 2584 | break; |
69ed779a | 2585 | cond_resched(); |
1da177e4 | 2586 | } |
d715cf80 LH |
2587 | if (i < h->max_huge_pages) { |
2588 | char buf[32]; | |
2589 | ||
c6247f72 | 2590 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2591 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2592 | h->max_huge_pages, buf, i); | |
2593 | h->max_huge_pages = i; | |
2594 | } | |
f60858f9 MK |
2595 | |
2596 | kfree(node_alloc_noretry); | |
e5ff2159 AK |
2597 | } |
2598 | ||
2599 | static void __init hugetlb_init_hstates(void) | |
2600 | { | |
2601 | struct hstate *h; | |
2602 | ||
2603 | for_each_hstate(h) { | |
641844f5 NH |
2604 | if (minimum_order > huge_page_order(h)) |
2605 | minimum_order = huge_page_order(h); | |
2606 | ||
8faa8b07 | 2607 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2608 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2609 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2610 | } |
641844f5 | 2611 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2612 | } |
2613 | ||
2614 | static void __init report_hugepages(void) | |
2615 | { | |
2616 | struct hstate *h; | |
2617 | ||
2618 | for_each_hstate(h) { | |
4abd32db | 2619 | char buf[32]; |
c6247f72 MW |
2620 | |
2621 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2622 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2623 | buf, h->free_huge_pages); |
e5ff2159 AK |
2624 | } |
2625 | } | |
2626 | ||
1da177e4 | 2627 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2628 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2629 | nodemask_t *nodes_allowed) | |
1da177e4 | 2630 | { |
4415cc8d CL |
2631 | int i; |
2632 | ||
bae7f4ae | 2633 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2634 | return; |
2635 | ||
6ae11b27 | 2636 | for_each_node_mask(i, *nodes_allowed) { |
1da177e4 | 2637 | struct page *page, *next; |
a5516438 AK |
2638 | struct list_head *freel = &h->hugepage_freelists[i]; |
2639 | list_for_each_entry_safe(page, next, freel, lru) { | |
2640 | if (count >= h->nr_huge_pages) | |
6b0c880d | 2641 | return; |
1da177e4 LT |
2642 | if (PageHighMem(page)) |
2643 | continue; | |
2644 | list_del(&page->lru); | |
e5ff2159 | 2645 | update_and_free_page(h, page); |
a5516438 AK |
2646 | h->free_huge_pages--; |
2647 | h->free_huge_pages_node[page_to_nid(page)]--; | |
1da177e4 LT |
2648 | } |
2649 | } | |
2650 | } | |
2651 | #else | |
6ae11b27 LS |
2652 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2653 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2654 | { |
2655 | } | |
2656 | #endif | |
2657 | ||
20a0307c WF |
2658 | /* |
2659 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2660 | * balanced by operating on them in a round-robin fashion. | |
2661 | * Returns 1 if an adjustment was made. | |
2662 | */ | |
6ae11b27 LS |
2663 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2664 | int delta) | |
20a0307c | 2665 | { |
b2261026 | 2666 | int nr_nodes, node; |
20a0307c WF |
2667 | |
2668 | VM_BUG_ON(delta != -1 && delta != 1); | |
20a0307c | 2669 | |
b2261026 JK |
2670 | if (delta < 0) { |
2671 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2672 | if (h->surplus_huge_pages_node[node]) | |
2673 | goto found; | |
e8c5c824 | 2674 | } |
b2261026 JK |
2675 | } else { |
2676 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2677 | if (h->surplus_huge_pages_node[node] < | |
2678 | h->nr_huge_pages_node[node]) | |
2679 | goto found; | |
e8c5c824 | 2680 | } |
b2261026 JK |
2681 | } |
2682 | return 0; | |
20a0307c | 2683 | |
b2261026 JK |
2684 | found: |
2685 | h->surplus_huge_pages += delta; | |
2686 | h->surplus_huge_pages_node[node] += delta; | |
2687 | return 1; | |
20a0307c WF |
2688 | } |
2689 | ||
a5516438 | 2690 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2691 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2692 | nodemask_t *nodes_allowed) |
1da177e4 | 2693 | { |
7893d1d5 | 2694 | unsigned long min_count, ret; |
f60858f9 MK |
2695 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2696 | ||
2697 | /* | |
2698 | * Bit mask controlling how hard we retry per-node allocations. | |
2699 | * If we can not allocate the bit mask, do not attempt to allocate | |
2700 | * the requested huge pages. | |
2701 | */ | |
2702 | if (node_alloc_noretry) | |
2703 | nodes_clear(*node_alloc_noretry); | |
2704 | else | |
2705 | return -ENOMEM; | |
1da177e4 | 2706 | |
4eb0716e AG |
2707 | spin_lock(&hugetlb_lock); |
2708 | ||
fd875dca MK |
2709 | /* |
2710 | * Check for a node specific request. | |
2711 | * Changing node specific huge page count may require a corresponding | |
2712 | * change to the global count. In any case, the passed node mask | |
2713 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2714 | */ | |
2715 | if (nid != NUMA_NO_NODE) { | |
2716 | unsigned long old_count = count; | |
2717 | ||
2718 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2719 | /* | |
2720 | * User may have specified a large count value which caused the | |
2721 | * above calculation to overflow. In this case, they wanted | |
2722 | * to allocate as many huge pages as possible. Set count to | |
2723 | * largest possible value to align with their intention. | |
2724 | */ | |
2725 | if (count < old_count) | |
2726 | count = ULONG_MAX; | |
2727 | } | |
2728 | ||
4eb0716e AG |
2729 | /* |
2730 | * Gigantic pages runtime allocation depend on the capability for large | |
2731 | * page range allocation. | |
2732 | * If the system does not provide this feature, return an error when | |
2733 | * the user tries to allocate gigantic pages but let the user free the | |
2734 | * boottime allocated gigantic pages. | |
2735 | */ | |
2736 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2737 | if (count > persistent_huge_pages(h)) { | |
2738 | spin_unlock(&hugetlb_lock); | |
f60858f9 | 2739 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2740 | return -EINVAL; |
2741 | } | |
2742 | /* Fall through to decrease pool */ | |
2743 | } | |
aa888a74 | 2744 | |
7893d1d5 AL |
2745 | /* |
2746 | * Increase the pool size | |
2747 | * First take pages out of surplus state. Then make up the | |
2748 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2749 | * |
0c397dae | 2750 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2751 | * to convert a surplus huge page to a normal huge page. That is |
2752 | * not critical, though, it just means the overall size of the | |
2753 | * pool might be one hugepage larger than it needs to be, but | |
2754 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2755 | */ |
a5516438 | 2756 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2757 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2758 | break; |
2759 | } | |
2760 | ||
a5516438 | 2761 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2762 | /* |
2763 | * If this allocation races such that we no longer need the | |
2764 | * page, free_huge_page will handle it by freeing the page | |
2765 | * and reducing the surplus. | |
2766 | */ | |
2767 | spin_unlock(&hugetlb_lock); | |
649920c6 JH |
2768 | |
2769 | /* yield cpu to avoid soft lockup */ | |
2770 | cond_resched(); | |
2771 | ||
f60858f9 MK |
2772 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2773 | node_alloc_noretry); | |
7893d1d5 AL |
2774 | spin_lock(&hugetlb_lock); |
2775 | if (!ret) | |
2776 | goto out; | |
2777 | ||
536240f2 MG |
2778 | /* Bail for signals. Probably ctrl-c from user */ |
2779 | if (signal_pending(current)) | |
2780 | goto out; | |
7893d1d5 | 2781 | } |
7893d1d5 AL |
2782 | |
2783 | /* | |
2784 | * Decrease the pool size | |
2785 | * First return free pages to the buddy allocator (being careful | |
2786 | * to keep enough around to satisfy reservations). Then place | |
2787 | * pages into surplus state as needed so the pool will shrink | |
2788 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2789 | * |
2790 | * By placing pages into the surplus state independent of the | |
2791 | * overcommit value, we are allowing the surplus pool size to | |
2792 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2793 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2794 | * though, we'll note that we're not allowed to exceed surplus |
2795 | * and won't grow the pool anywhere else. Not until one of the | |
2796 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2797 | */ |
a5516438 | 2798 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2799 | min_count = max(count, min_count); |
6ae11b27 | 2800 | try_to_free_low(h, min_count, nodes_allowed); |
a5516438 | 2801 | while (min_count < persistent_huge_pages(h)) { |
6ae11b27 | 2802 | if (!free_pool_huge_page(h, nodes_allowed, 0)) |
1da177e4 | 2803 | break; |
55f67141 | 2804 | cond_resched_lock(&hugetlb_lock); |
1da177e4 | 2805 | } |
a5516438 | 2806 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2807 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2808 | break; |
2809 | } | |
2810 | out: | |
4eb0716e | 2811 | h->max_huge_pages = persistent_huge_pages(h); |
1da177e4 | 2812 | spin_unlock(&hugetlb_lock); |
4eb0716e | 2813 | |
f60858f9 MK |
2814 | NODEMASK_FREE(node_alloc_noretry); |
2815 | ||
4eb0716e | 2816 | return 0; |
1da177e4 LT |
2817 | } |
2818 | ||
a3437870 NA |
2819 | #define HSTATE_ATTR_RO(_name) \ |
2820 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2821 | ||
2822 | #define HSTATE_ATTR(_name) \ | |
2823 | static struct kobj_attribute _name##_attr = \ | |
2824 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2825 | ||
2826 | static struct kobject *hugepages_kobj; | |
2827 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2828 | ||
9a305230 LS |
2829 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2830 | ||
2831 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2832 | { |
2833 | int i; | |
9a305230 | 2834 | |
a3437870 | 2835 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2836 | if (hstate_kobjs[i] == kobj) { |
2837 | if (nidp) | |
2838 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2839 | return &hstates[i]; |
9a305230 LS |
2840 | } |
2841 | ||
2842 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2843 | } |
2844 | ||
06808b08 | 2845 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2846 | struct kobj_attribute *attr, char *buf) |
2847 | { | |
9a305230 LS |
2848 | struct hstate *h; |
2849 | unsigned long nr_huge_pages; | |
2850 | int nid; | |
2851 | ||
2852 | h = kobj_to_hstate(kobj, &nid); | |
2853 | if (nid == NUMA_NO_NODE) | |
2854 | nr_huge_pages = h->nr_huge_pages; | |
2855 | else | |
2856 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2857 | ||
2858 | return sprintf(buf, "%lu\n", nr_huge_pages); | |
a3437870 | 2859 | } |
adbe8726 | 2860 | |
238d3c13 DR |
2861 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2862 | struct hstate *h, int nid, | |
2863 | unsigned long count, size_t len) | |
a3437870 NA |
2864 | { |
2865 | int err; | |
2d0adf7e | 2866 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2867 | |
2d0adf7e OS |
2868 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2869 | return -EINVAL; | |
adbe8726 | 2870 | |
9a305230 LS |
2871 | if (nid == NUMA_NO_NODE) { |
2872 | /* | |
2873 | * global hstate attribute | |
2874 | */ | |
2875 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2876 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2877 | n_mask = &node_states[N_MEMORY]; | |
2878 | else | |
2879 | n_mask = &nodes_allowed; | |
2880 | } else { | |
9a305230 | 2881 | /* |
fd875dca MK |
2882 | * Node specific request. count adjustment happens in |
2883 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2884 | */ |
2d0adf7e OS |
2885 | init_nodemask_of_node(&nodes_allowed, nid); |
2886 | n_mask = &nodes_allowed; | |
fd875dca | 2887 | } |
9a305230 | 2888 | |
2d0adf7e | 2889 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2890 | |
4eb0716e | 2891 | return err ? err : len; |
06808b08 LS |
2892 | } |
2893 | ||
238d3c13 DR |
2894 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2895 | struct kobject *kobj, const char *buf, | |
2896 | size_t len) | |
2897 | { | |
2898 | struct hstate *h; | |
2899 | unsigned long count; | |
2900 | int nid; | |
2901 | int err; | |
2902 | ||
2903 | err = kstrtoul(buf, 10, &count); | |
2904 | if (err) | |
2905 | return err; | |
2906 | ||
2907 | h = kobj_to_hstate(kobj, &nid); | |
2908 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2909 | } | |
2910 | ||
06808b08 LS |
2911 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2912 | struct kobj_attribute *attr, char *buf) | |
2913 | { | |
2914 | return nr_hugepages_show_common(kobj, attr, buf); | |
2915 | } | |
2916 | ||
2917 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2918 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2919 | { | |
238d3c13 | 2920 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2921 | } |
2922 | HSTATE_ATTR(nr_hugepages); | |
2923 | ||
06808b08 LS |
2924 | #ifdef CONFIG_NUMA |
2925 | ||
2926 | /* | |
2927 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2928 | * huge page alloc/free. | |
2929 | */ | |
2930 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
2931 | struct kobj_attribute *attr, char *buf) | |
2932 | { | |
2933 | return nr_hugepages_show_common(kobj, attr, buf); | |
2934 | } | |
2935 | ||
2936 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2937 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2938 | { | |
238d3c13 | 2939 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2940 | } |
2941 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2942 | #endif | |
2943 | ||
2944 | ||
a3437870 NA |
2945 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2946 | struct kobj_attribute *attr, char *buf) | |
2947 | { | |
9a305230 | 2948 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2949 | return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages); |
2950 | } | |
adbe8726 | 2951 | |
a3437870 NA |
2952 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2953 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2954 | { | |
2955 | int err; | |
2956 | unsigned long input; | |
9a305230 | 2957 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2958 | |
bae7f4ae | 2959 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2960 | return -EINVAL; |
2961 | ||
3dbb95f7 | 2962 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2963 | if (err) |
73ae31e5 | 2964 | return err; |
a3437870 NA |
2965 | |
2966 | spin_lock(&hugetlb_lock); | |
2967 | h->nr_overcommit_huge_pages = input; | |
2968 | spin_unlock(&hugetlb_lock); | |
2969 | ||
2970 | return count; | |
2971 | } | |
2972 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2973 | ||
2974 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2975 | struct kobj_attribute *attr, char *buf) | |
2976 | { | |
9a305230 LS |
2977 | struct hstate *h; |
2978 | unsigned long free_huge_pages; | |
2979 | int nid; | |
2980 | ||
2981 | h = kobj_to_hstate(kobj, &nid); | |
2982 | if (nid == NUMA_NO_NODE) | |
2983 | free_huge_pages = h->free_huge_pages; | |
2984 | else | |
2985 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2986 | ||
2987 | return sprintf(buf, "%lu\n", free_huge_pages); | |
a3437870 NA |
2988 | } |
2989 | HSTATE_ATTR_RO(free_hugepages); | |
2990 | ||
2991 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2992 | struct kobj_attribute *attr, char *buf) | |
2993 | { | |
9a305230 | 2994 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 NA |
2995 | return sprintf(buf, "%lu\n", h->resv_huge_pages); |
2996 | } | |
2997 | HSTATE_ATTR_RO(resv_hugepages); | |
2998 | ||
2999 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
3000 | struct kobj_attribute *attr, char *buf) | |
3001 | { | |
9a305230 LS |
3002 | struct hstate *h; |
3003 | unsigned long surplus_huge_pages; | |
3004 | int nid; | |
3005 | ||
3006 | h = kobj_to_hstate(kobj, &nid); | |
3007 | if (nid == NUMA_NO_NODE) | |
3008 | surplus_huge_pages = h->surplus_huge_pages; | |
3009 | else | |
3010 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
3011 | ||
3012 | return sprintf(buf, "%lu\n", surplus_huge_pages); | |
a3437870 NA |
3013 | } |
3014 | HSTATE_ATTR_RO(surplus_hugepages); | |
3015 | ||
3016 | static struct attribute *hstate_attrs[] = { | |
3017 | &nr_hugepages_attr.attr, | |
3018 | &nr_overcommit_hugepages_attr.attr, | |
3019 | &free_hugepages_attr.attr, | |
3020 | &resv_hugepages_attr.attr, | |
3021 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
3022 | #ifdef CONFIG_NUMA |
3023 | &nr_hugepages_mempolicy_attr.attr, | |
3024 | #endif | |
a3437870 NA |
3025 | NULL, |
3026 | }; | |
3027 | ||
67e5ed96 | 3028 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
3029 | .attrs = hstate_attrs, |
3030 | }; | |
3031 | ||
094e9539 JM |
3032 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
3033 | struct kobject **hstate_kobjs, | |
67e5ed96 | 3034 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
3035 | { |
3036 | int retval; | |
972dc4de | 3037 | int hi = hstate_index(h); |
a3437870 | 3038 | |
9a305230 LS |
3039 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
3040 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
3041 | return -ENOMEM; |
3042 | ||
9a305230 | 3043 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
a3437870 | 3044 | if (retval) |
9a305230 | 3045 | kobject_put(hstate_kobjs[hi]); |
a3437870 NA |
3046 | |
3047 | return retval; | |
3048 | } | |
3049 | ||
3050 | static void __init hugetlb_sysfs_init(void) | |
3051 | { | |
3052 | struct hstate *h; | |
3053 | int err; | |
3054 | ||
3055 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
3056 | if (!hugepages_kobj) | |
3057 | return; | |
3058 | ||
3059 | for_each_hstate(h) { | |
9a305230 LS |
3060 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
3061 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 3062 | if (err) |
ffb22af5 | 3063 | pr_err("Hugetlb: Unable to add hstate %s", h->name); |
a3437870 NA |
3064 | } |
3065 | } | |
3066 | ||
9a305230 LS |
3067 | #ifdef CONFIG_NUMA |
3068 | ||
3069 | /* | |
3070 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3071 | * with node devices in node_devices[] using a parallel array. The array |
3072 | * index of a node device or _hstate == node id. | |
3073 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3074 | * the base kernel, on the hugetlb module. |
3075 | */ | |
3076 | struct node_hstate { | |
3077 | struct kobject *hugepages_kobj; | |
3078 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3079 | }; | |
b4e289a6 | 3080 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3081 | |
3082 | /* | |
10fbcf4c | 3083 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3084 | */ |
3085 | static struct attribute *per_node_hstate_attrs[] = { | |
3086 | &nr_hugepages_attr.attr, | |
3087 | &free_hugepages_attr.attr, | |
3088 | &surplus_hugepages_attr.attr, | |
3089 | NULL, | |
3090 | }; | |
3091 | ||
67e5ed96 | 3092 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3093 | .attrs = per_node_hstate_attrs, |
3094 | }; | |
3095 | ||
3096 | /* | |
10fbcf4c | 3097 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3098 | * Returns node id via non-NULL nidp. |
3099 | */ | |
3100 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3101 | { | |
3102 | int nid; | |
3103 | ||
3104 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3105 | struct node_hstate *nhs = &node_hstates[nid]; | |
3106 | int i; | |
3107 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3108 | if (nhs->hstate_kobjs[i] == kobj) { | |
3109 | if (nidp) | |
3110 | *nidp = nid; | |
3111 | return &hstates[i]; | |
3112 | } | |
3113 | } | |
3114 | ||
3115 | BUG(); | |
3116 | return NULL; | |
3117 | } | |
3118 | ||
3119 | /* | |
10fbcf4c | 3120 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3121 | * No-op if no hstate attributes attached. |
3122 | */ | |
3cd8b44f | 3123 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3124 | { |
3125 | struct hstate *h; | |
10fbcf4c | 3126 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3127 | |
3128 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3129 | return; /* no hstate attributes */ |
9a305230 | 3130 | |
972dc4de AK |
3131 | for_each_hstate(h) { |
3132 | int idx = hstate_index(h); | |
3133 | if (nhs->hstate_kobjs[idx]) { | |
3134 | kobject_put(nhs->hstate_kobjs[idx]); | |
3135 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3136 | } |
972dc4de | 3137 | } |
9a305230 LS |
3138 | |
3139 | kobject_put(nhs->hugepages_kobj); | |
3140 | nhs->hugepages_kobj = NULL; | |
3141 | } | |
3142 | ||
9a305230 LS |
3143 | |
3144 | /* | |
10fbcf4c | 3145 | * Register hstate attributes for a single node device. |
9a305230 LS |
3146 | * No-op if attributes already registered. |
3147 | */ | |
3cd8b44f | 3148 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3149 | { |
3150 | struct hstate *h; | |
10fbcf4c | 3151 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3152 | int err; |
3153 | ||
3154 | if (nhs->hugepages_kobj) | |
3155 | return; /* already allocated */ | |
3156 | ||
3157 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3158 | &node->dev.kobj); |
9a305230 LS |
3159 | if (!nhs->hugepages_kobj) |
3160 | return; | |
3161 | ||
3162 | for_each_hstate(h) { | |
3163 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3164 | nhs->hstate_kobjs, | |
3165 | &per_node_hstate_attr_group); | |
3166 | if (err) { | |
ffb22af5 AM |
3167 | pr_err("Hugetlb: Unable to add hstate %s for node %d\n", |
3168 | h->name, node->dev.id); | |
9a305230 LS |
3169 | hugetlb_unregister_node(node); |
3170 | break; | |
3171 | } | |
3172 | } | |
3173 | } | |
3174 | ||
3175 | /* | |
9b5e5d0f | 3176 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3177 | * devices of nodes that have memory. All on-line nodes should have |
3178 | * registered their associated device by this time. | |
9a305230 | 3179 | */ |
7d9ca000 | 3180 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3181 | { |
3182 | int nid; | |
3183 | ||
8cebfcd0 | 3184 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3185 | struct node *node = node_devices[nid]; |
10fbcf4c | 3186 | if (node->dev.id == nid) |
9a305230 LS |
3187 | hugetlb_register_node(node); |
3188 | } | |
3189 | ||
3190 | /* | |
10fbcf4c | 3191 | * Let the node device driver know we're here so it can |
9a305230 LS |
3192 | * [un]register hstate attributes on node hotplug. |
3193 | */ | |
3194 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3195 | hugetlb_unregister_node); | |
3196 | } | |
3197 | #else /* !CONFIG_NUMA */ | |
3198 | ||
3199 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3200 | { | |
3201 | BUG(); | |
3202 | if (nidp) | |
3203 | *nidp = -1; | |
3204 | return NULL; | |
3205 | } | |
3206 | ||
9a305230 LS |
3207 | static void hugetlb_register_all_nodes(void) { } |
3208 | ||
3209 | #endif | |
3210 | ||
a3437870 NA |
3211 | static int __init hugetlb_init(void) |
3212 | { | |
8382d914 DB |
3213 | int i; |
3214 | ||
457c1b27 | 3215 | if (!hugepages_supported()) |
0ef89d25 | 3216 | return 0; |
a3437870 | 3217 | |
e11bfbfc | 3218 | if (!size_to_hstate(default_hstate_size)) { |
d715cf80 LH |
3219 | if (default_hstate_size != 0) { |
3220 | pr_err("HugeTLB: unsupported default_hugepagesz %lu. Reverting to %lu\n", | |
3221 | default_hstate_size, HPAGE_SIZE); | |
3222 | } | |
3223 | ||
e11bfbfc NP |
3224 | default_hstate_size = HPAGE_SIZE; |
3225 | if (!size_to_hstate(default_hstate_size)) | |
3226 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
a3437870 | 3227 | } |
972dc4de | 3228 | default_hstate_idx = hstate_index(size_to_hstate(default_hstate_size)); |
f8b74815 VT |
3229 | if (default_hstate_max_huge_pages) { |
3230 | if (!default_hstate.max_huge_pages) | |
3231 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3232 | } | |
a3437870 | 3233 | |
cf11e85f | 3234 | hugetlb_cma_check(); |
a3437870 | 3235 | hugetlb_init_hstates(); |
aa888a74 | 3236 | gather_bootmem_prealloc(); |
a3437870 NA |
3237 | report_hugepages(); |
3238 | ||
3239 | hugetlb_sysfs_init(); | |
9a305230 | 3240 | hugetlb_register_all_nodes(); |
7179e7bf | 3241 | hugetlb_cgroup_file_init(); |
9a305230 | 3242 | |
8382d914 DB |
3243 | #ifdef CONFIG_SMP |
3244 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3245 | #else | |
3246 | num_fault_mutexes = 1; | |
3247 | #endif | |
c672c7f2 | 3248 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3249 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3250 | GFP_KERNEL); | |
c672c7f2 | 3251 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3252 | |
3253 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3254 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3255 | return 0; |
3256 | } | |
3e89e1c5 | 3257 | subsys_initcall(hugetlb_init); |
a3437870 NA |
3258 | |
3259 | /* Should be called on processing a hugepagesz=... option */ | |
9fee021d VT |
3260 | void __init hugetlb_bad_size(void) |
3261 | { | |
3262 | parsed_valid_hugepagesz = false; | |
3263 | } | |
3264 | ||
d00181b9 | 3265 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3266 | { |
3267 | struct hstate *h; | |
8faa8b07 AK |
3268 | unsigned long i; |
3269 | ||
a3437870 | 3270 | if (size_to_hstate(PAGE_SIZE << order)) { |
598d8091 | 3271 | pr_warn("hugepagesz= specified twice, ignoring\n"); |
a3437870 NA |
3272 | return; |
3273 | } | |
47d38344 | 3274 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3275 | BUG_ON(order == 0); |
47d38344 | 3276 | h = &hstates[hugetlb_max_hstate++]; |
a3437870 NA |
3277 | h->order = order; |
3278 | h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1); | |
8faa8b07 AK |
3279 | h->nr_huge_pages = 0; |
3280 | h->free_huge_pages = 0; | |
3281 | for (i = 0; i < MAX_NUMNODES; ++i) | |
3282 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3283 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3284 | h->next_nid_to_alloc = first_memory_node; |
3285 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3286 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3287 | huge_page_size(h)/1024); | |
8faa8b07 | 3288 | |
a3437870 NA |
3289 | parsed_hstate = h; |
3290 | } | |
3291 | ||
e11bfbfc | 3292 | static int __init hugetlb_nrpages_setup(char *s) |
a3437870 NA |
3293 | { |
3294 | unsigned long *mhp; | |
8faa8b07 | 3295 | static unsigned long *last_mhp; |
a3437870 | 3296 | |
9fee021d VT |
3297 | if (!parsed_valid_hugepagesz) { |
3298 | pr_warn("hugepages = %s preceded by " | |
3299 | "an unsupported hugepagesz, ignoring\n", s); | |
3300 | parsed_valid_hugepagesz = true; | |
3301 | return 1; | |
3302 | } | |
a3437870 | 3303 | /* |
47d38344 | 3304 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet, |
a3437870 NA |
3305 | * so this hugepages= parameter goes to the "default hstate". |
3306 | */ | |
9fee021d | 3307 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3308 | mhp = &default_hstate_max_huge_pages; |
3309 | else | |
3310 | mhp = &parsed_hstate->max_huge_pages; | |
3311 | ||
8faa8b07 | 3312 | if (mhp == last_mhp) { |
598d8091 | 3313 | pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n"); |
8faa8b07 AK |
3314 | return 1; |
3315 | } | |
3316 | ||
a3437870 NA |
3317 | if (sscanf(s, "%lu", mhp) <= 0) |
3318 | *mhp = 0; | |
3319 | ||
8faa8b07 AK |
3320 | /* |
3321 | * Global state is always initialized later in hugetlb_init. | |
3322 | * But we need to allocate >= MAX_ORDER hstates here early to still | |
3323 | * use the bootmem allocator. | |
3324 | */ | |
47d38344 | 3325 | if (hugetlb_max_hstate && parsed_hstate->order >= MAX_ORDER) |
8faa8b07 AK |
3326 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3327 | ||
3328 | last_mhp = mhp; | |
3329 | ||
a3437870 NA |
3330 | return 1; |
3331 | } | |
e11bfbfc NP |
3332 | __setup("hugepages=", hugetlb_nrpages_setup); |
3333 | ||
3334 | static int __init hugetlb_default_setup(char *s) | |
3335 | { | |
3336 | default_hstate_size = memparse(s, &s); | |
3337 | return 1; | |
3338 | } | |
3339 | __setup("default_hugepagesz=", hugetlb_default_setup); | |
a3437870 | 3340 | |
8a213460 NA |
3341 | static unsigned int cpuset_mems_nr(unsigned int *array) |
3342 | { | |
3343 | int node; | |
3344 | unsigned int nr = 0; | |
3345 | ||
3346 | for_each_node_mask(node, cpuset_current_mems_allowed) | |
3347 | nr += array[node]; | |
3348 | ||
3349 | return nr; | |
3350 | } | |
3351 | ||
3352 | #ifdef CONFIG_SYSCTL | |
06808b08 LS |
3353 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3354 | struct ctl_table *table, int write, | |
32927393 | 3355 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 3356 | { |
e5ff2159 | 3357 | struct hstate *h = &default_hstate; |
238d3c13 | 3358 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3359 | int ret; |
e5ff2159 | 3360 | |
457c1b27 | 3361 | if (!hugepages_supported()) |
86613628 | 3362 | return -EOPNOTSUPP; |
457c1b27 | 3363 | |
e5ff2159 AK |
3364 | table->data = &tmp; |
3365 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3366 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3367 | if (ret) | |
3368 | goto out; | |
e5ff2159 | 3369 | |
238d3c13 DR |
3370 | if (write) |
3371 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3372 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3373 | out: |
3374 | return ret; | |
1da177e4 | 3375 | } |
396faf03 | 3376 | |
06808b08 | 3377 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 3378 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3379 | { |
3380 | ||
3381 | return hugetlb_sysctl_handler_common(false, table, write, | |
3382 | buffer, length, ppos); | |
3383 | } | |
3384 | ||
3385 | #ifdef CONFIG_NUMA | |
3386 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 3387 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3388 | { |
3389 | return hugetlb_sysctl_handler_common(true, table, write, | |
3390 | buffer, length, ppos); | |
3391 | } | |
3392 | #endif /* CONFIG_NUMA */ | |
3393 | ||
a3d0c6aa | 3394 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 3395 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 3396 | { |
a5516438 | 3397 | struct hstate *h = &default_hstate; |
e5ff2159 | 3398 | unsigned long tmp; |
08d4a246 | 3399 | int ret; |
e5ff2159 | 3400 | |
457c1b27 | 3401 | if (!hugepages_supported()) |
86613628 | 3402 | return -EOPNOTSUPP; |
457c1b27 | 3403 | |
c033a93c | 3404 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3405 | |
bae7f4ae | 3406 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3407 | return -EINVAL; |
3408 | ||
e5ff2159 AK |
3409 | table->data = &tmp; |
3410 | table->maxlen = sizeof(unsigned long); | |
08d4a246 MH |
3411 | ret = proc_doulongvec_minmax(table, write, buffer, length, ppos); |
3412 | if (ret) | |
3413 | goto out; | |
e5ff2159 AK |
3414 | |
3415 | if (write) { | |
3416 | spin_lock(&hugetlb_lock); | |
3417 | h->nr_overcommit_huge_pages = tmp; | |
3418 | spin_unlock(&hugetlb_lock); | |
3419 | } | |
08d4a246 MH |
3420 | out: |
3421 | return ret; | |
a3d0c6aa NA |
3422 | } |
3423 | ||
1da177e4 LT |
3424 | #endif /* CONFIG_SYSCTL */ |
3425 | ||
e1759c21 | 3426 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3427 | { |
fcb2b0c5 RG |
3428 | struct hstate *h; |
3429 | unsigned long total = 0; | |
3430 | ||
457c1b27 NA |
3431 | if (!hugepages_supported()) |
3432 | return; | |
fcb2b0c5 RG |
3433 | |
3434 | for_each_hstate(h) { | |
3435 | unsigned long count = h->nr_huge_pages; | |
3436 | ||
3437 | total += (PAGE_SIZE << huge_page_order(h)) * count; | |
3438 | ||
3439 | if (h == &default_hstate) | |
3440 | seq_printf(m, | |
3441 | "HugePages_Total: %5lu\n" | |
3442 | "HugePages_Free: %5lu\n" | |
3443 | "HugePages_Rsvd: %5lu\n" | |
3444 | "HugePages_Surp: %5lu\n" | |
3445 | "Hugepagesize: %8lu kB\n", | |
3446 | count, | |
3447 | h->free_huge_pages, | |
3448 | h->resv_huge_pages, | |
3449 | h->surplus_huge_pages, | |
3450 | (PAGE_SIZE << huge_page_order(h)) / 1024); | |
3451 | } | |
3452 | ||
3453 | seq_printf(m, "Hugetlb: %8lu kB\n", total / 1024); | |
1da177e4 LT |
3454 | } |
3455 | ||
3456 | int hugetlb_report_node_meminfo(int nid, char *buf) | |
3457 | { | |
a5516438 | 3458 | struct hstate *h = &default_hstate; |
457c1b27 NA |
3459 | if (!hugepages_supported()) |
3460 | return 0; | |
1da177e4 LT |
3461 | return sprintf(buf, |
3462 | "Node %d HugePages_Total: %5u\n" | |
a1de0919 NA |
3463 | "Node %d HugePages_Free: %5u\n" |
3464 | "Node %d HugePages_Surp: %5u\n", | |
a5516438 AK |
3465 | nid, h->nr_huge_pages_node[nid], |
3466 | nid, h->free_huge_pages_node[nid], | |
3467 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3468 | } |
3469 | ||
949f7ec5 DR |
3470 | void hugetlb_show_meminfo(void) |
3471 | { | |
3472 | struct hstate *h; | |
3473 | int nid; | |
3474 | ||
457c1b27 NA |
3475 | if (!hugepages_supported()) |
3476 | return; | |
3477 | ||
949f7ec5 DR |
3478 | for_each_node_state(nid, N_MEMORY) |
3479 | for_each_hstate(h) | |
3480 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3481 | nid, | |
3482 | h->nr_huge_pages_node[nid], | |
3483 | h->free_huge_pages_node[nid], | |
3484 | h->surplus_huge_pages_node[nid], | |
3485 | 1UL << (huge_page_order(h) + PAGE_SHIFT - 10)); | |
3486 | } | |
3487 | ||
5d317b2b NH |
3488 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3489 | { | |
3490 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3491 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3492 | } | |
3493 | ||
1da177e4 LT |
3494 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3495 | unsigned long hugetlb_total_pages(void) | |
3496 | { | |
d0028588 WL |
3497 | struct hstate *h; |
3498 | unsigned long nr_total_pages = 0; | |
3499 | ||
3500 | for_each_hstate(h) | |
3501 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3502 | return nr_total_pages; | |
1da177e4 | 3503 | } |
1da177e4 | 3504 | |
a5516438 | 3505 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3506 | { |
3507 | int ret = -ENOMEM; | |
3508 | ||
3509 | spin_lock(&hugetlb_lock); | |
3510 | /* | |
3511 | * When cpuset is configured, it breaks the strict hugetlb page | |
3512 | * reservation as the accounting is done on a global variable. Such | |
3513 | * reservation is completely rubbish in the presence of cpuset because | |
3514 | * the reservation is not checked against page availability for the | |
3515 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3516 | * with lack of free htlb page in cpuset that the task is in. | |
3517 | * Attempt to enforce strict accounting with cpuset is almost | |
3518 | * impossible (or too ugly) because cpuset is too fluid that | |
3519 | * task or memory node can be dynamically moved between cpusets. | |
3520 | * | |
3521 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3522 | * undesirable. However, in order to preserve some of the semantics, | |
3523 | * we fall back to check against current free page availability as | |
3524 | * a best attempt and hopefully to minimize the impact of changing | |
3525 | * semantics that cpuset has. | |
3526 | */ | |
3527 | if (delta > 0) { | |
a5516438 | 3528 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3529 | goto out; |
3530 | ||
a5516438 AK |
3531 | if (delta > cpuset_mems_nr(h->free_huge_pages_node)) { |
3532 | return_unused_surplus_pages(h, delta); | |
fc1b8a73 MG |
3533 | goto out; |
3534 | } | |
3535 | } | |
3536 | ||
3537 | ret = 0; | |
3538 | if (delta < 0) | |
a5516438 | 3539 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3540 | |
3541 | out: | |
3542 | spin_unlock(&hugetlb_lock); | |
3543 | return ret; | |
3544 | } | |
3545 | ||
84afd99b AW |
3546 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3547 | { | |
f522c3ac | 3548 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3549 | |
3550 | /* | |
3551 | * This new VMA should share its siblings reservation map if present. | |
3552 | * The VMA will only ever have a valid reservation map pointer where | |
3553 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3554 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3555 | * after this open call completes. It is therefore safe to take a |
3556 | * new reference here without additional locking. | |
3557 | */ | |
4e35f483 | 3558 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3559 | kref_get(&resv->refs); |
84afd99b AW |
3560 | } |
3561 | ||
a1e78772 MG |
3562 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3563 | { | |
a5516438 | 3564 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3565 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3566 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3567 | unsigned long reserve, start, end; |
1c5ecae3 | 3568 | long gbl_reserve; |
84afd99b | 3569 | |
4e35f483 JK |
3570 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3571 | return; | |
84afd99b | 3572 | |
4e35f483 JK |
3573 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3574 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3575 | |
4e35f483 | 3576 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 3577 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 3578 | if (reserve) { |
1c5ecae3 MK |
3579 | /* |
3580 | * Decrement reserve counts. The global reserve count may be | |
3581 | * adjusted if the subpool has a minimum size. | |
3582 | */ | |
3583 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3584 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3585 | } |
e9fe92ae MA |
3586 | |
3587 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
3588 | } |
3589 | ||
31383c68 DW |
3590 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3591 | { | |
3592 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3593 | return -EINVAL; | |
3594 | return 0; | |
3595 | } | |
3596 | ||
05ea8860 DW |
3597 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3598 | { | |
3599 | struct hstate *hstate = hstate_vma(vma); | |
3600 | ||
3601 | return 1UL << huge_page_shift(hstate); | |
3602 | } | |
3603 | ||
1da177e4 LT |
3604 | /* |
3605 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3606 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
3607 | * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get | |
3608 | * this far. | |
3609 | */ | |
b3ec9f33 | 3610 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3611 | { |
3612 | BUG(); | |
d0217ac0 | 3613 | return 0; |
1da177e4 LT |
3614 | } |
3615 | ||
eec3636a JC |
3616 | /* |
3617 | * When a new function is introduced to vm_operations_struct and added | |
3618 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3619 | * This is because under System V memory model, mappings created via | |
3620 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3621 | * their original vm_ops are overwritten with shm_vm_ops. | |
3622 | */ | |
f0f37e2f | 3623 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3624 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3625 | .open = hugetlb_vm_op_open, |
a1e78772 | 3626 | .close = hugetlb_vm_op_close, |
31383c68 | 3627 | .split = hugetlb_vm_op_split, |
05ea8860 | 3628 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3629 | }; |
3630 | ||
1e8f889b DG |
3631 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3632 | int writable) | |
63551ae0 DG |
3633 | { |
3634 | pte_t entry; | |
3635 | ||
1e8f889b | 3636 | if (writable) { |
106c992a GS |
3637 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3638 | vma->vm_page_prot))); | |
63551ae0 | 3639 | } else { |
106c992a GS |
3640 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3641 | vma->vm_page_prot)); | |
63551ae0 DG |
3642 | } |
3643 | entry = pte_mkyoung(entry); | |
3644 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3645 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3646 | |
3647 | return entry; | |
3648 | } | |
3649 | ||
1e8f889b DG |
3650 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3651 | unsigned long address, pte_t *ptep) | |
3652 | { | |
3653 | pte_t entry; | |
3654 | ||
106c992a | 3655 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3656 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3657 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3658 | } |
3659 | ||
d5ed7444 | 3660 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3661 | { |
3662 | swp_entry_t swp; | |
3663 | ||
3664 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3665 | return false; |
4a705fef NH |
3666 | swp = pte_to_swp_entry(pte); |
3667 | if (non_swap_entry(swp) && is_migration_entry(swp)) | |
d5ed7444 | 3668 | return true; |
4a705fef | 3669 | else |
d5ed7444 | 3670 | return false; |
4a705fef NH |
3671 | } |
3672 | ||
3673 | static int is_hugetlb_entry_hwpoisoned(pte_t pte) | |
3674 | { | |
3675 | swp_entry_t swp; | |
3676 | ||
3677 | if (huge_pte_none(pte) || pte_present(pte)) | |
3678 | return 0; | |
3679 | swp = pte_to_swp_entry(pte); | |
3680 | if (non_swap_entry(swp) && is_hwpoison_entry(swp)) | |
3681 | return 1; | |
3682 | else | |
3683 | return 0; | |
3684 | } | |
1e8f889b | 3685 | |
63551ae0 DG |
3686 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3687 | struct vm_area_struct *vma) | |
3688 | { | |
5e41540c | 3689 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3690 | struct page *ptepage; |
1c59827d | 3691 | unsigned long addr; |
1e8f889b | 3692 | int cow; |
a5516438 AK |
3693 | struct hstate *h = hstate_vma(vma); |
3694 | unsigned long sz = huge_page_size(h); | |
c0d0381a | 3695 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 3696 | struct mmu_notifier_range range; |
e8569dd2 | 3697 | int ret = 0; |
1e8f889b DG |
3698 | |
3699 | cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE; | |
63551ae0 | 3700 | |
ac46d4f3 | 3701 | if (cow) { |
7269f999 | 3702 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3703 | vma->vm_start, |
ac46d4f3 JG |
3704 | vma->vm_end); |
3705 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
3706 | } else { |
3707 | /* | |
3708 | * For shared mappings i_mmap_rwsem must be held to call | |
3709 | * huge_pte_alloc, otherwise the returned ptep could go | |
3710 | * away if part of a shared pmd and another thread calls | |
3711 | * huge_pmd_unshare. | |
3712 | */ | |
3713 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 3714 | } |
e8569dd2 | 3715 | |
a5516438 | 3716 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3717 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3718 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3719 | if (!src_pte) |
3720 | continue; | |
a5516438 | 3721 | dst_pte = huge_pte_alloc(dst, addr, sz); |
e8569dd2 AS |
3722 | if (!dst_pte) { |
3723 | ret = -ENOMEM; | |
3724 | break; | |
3725 | } | |
c5c99429 | 3726 | |
5e41540c MK |
3727 | /* |
3728 | * If the pagetables are shared don't copy or take references. | |
3729 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3730 | * | |
3731 | * However, src could have 'unshared' and dst shares with | |
3732 | * another vma. If dst_pte !none, this implies sharing. | |
3733 | * Check here before taking page table lock, and once again | |
3734 | * after taking the lock below. | |
3735 | */ | |
3736 | dst_entry = huge_ptep_get(dst_pte); | |
3737 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3738 | continue; |
3739 | ||
cb900f41 KS |
3740 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3741 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3742 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3743 | entry = huge_ptep_get(src_pte); |
5e41540c MK |
3744 | dst_entry = huge_ptep_get(dst_pte); |
3745 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { | |
3746 | /* | |
3747 | * Skip if src entry none. Also, skip in the | |
3748 | * unlikely case dst entry !none as this implies | |
3749 | * sharing with another vma. | |
3750 | */ | |
4a705fef NH |
3751 | ; |
3752 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3753 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3754 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3755 | ||
3756 | if (is_write_migration_entry(swp_entry) && cow) { | |
3757 | /* | |
3758 | * COW mappings require pages in both | |
3759 | * parent and child to be set to read. | |
3760 | */ | |
3761 | make_migration_entry_read(&swp_entry); | |
3762 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3763 | set_huge_swap_pte_at(src, addr, src_pte, |
3764 | entry, sz); | |
4a705fef | 3765 | } |
e5251fd4 | 3766 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3767 | } else { |
34ee645e | 3768 | if (cow) { |
0f10851e JG |
3769 | /* |
3770 | * No need to notify as we are downgrading page | |
3771 | * table protection not changing it to point | |
3772 | * to a new page. | |
3773 | * | |
ad56b738 | 3774 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3775 | */ |
7f2e9525 | 3776 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3777 | } |
0253d634 | 3778 | entry = huge_ptep_get(src_pte); |
1c59827d HD |
3779 | ptepage = pte_page(entry); |
3780 | get_page(ptepage); | |
53f9263b | 3781 | page_dup_rmap(ptepage, true); |
1c59827d | 3782 | set_huge_pte_at(dst, addr, dst_pte, entry); |
5d317b2b | 3783 | hugetlb_count_add(pages_per_huge_page(h), dst); |
1c59827d | 3784 | } |
cb900f41 KS |
3785 | spin_unlock(src_ptl); |
3786 | spin_unlock(dst_ptl); | |
63551ae0 | 3787 | } |
63551ae0 | 3788 | |
e8569dd2 | 3789 | if (cow) |
ac46d4f3 | 3790 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
3791 | else |
3792 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
3793 | |
3794 | return ret; | |
63551ae0 DG |
3795 | } |
3796 | ||
24669e58 AK |
3797 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3798 | unsigned long start, unsigned long end, | |
3799 | struct page *ref_page) | |
63551ae0 DG |
3800 | { |
3801 | struct mm_struct *mm = vma->vm_mm; | |
3802 | unsigned long address; | |
c7546f8f | 3803 | pte_t *ptep; |
63551ae0 | 3804 | pte_t pte; |
cb900f41 | 3805 | spinlock_t *ptl; |
63551ae0 | 3806 | struct page *page; |
a5516438 AK |
3807 | struct hstate *h = hstate_vma(vma); |
3808 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3809 | struct mmu_notifier_range range; |
a5516438 | 3810 | |
63551ae0 | 3811 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3812 | BUG_ON(start & ~huge_page_mask(h)); |
3813 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3814 | |
07e32661 AK |
3815 | /* |
3816 | * This is a hugetlb vma, all the pte entries should point | |
3817 | * to huge page. | |
3818 | */ | |
ed6a7935 | 3819 | tlb_change_page_size(tlb, sz); |
24669e58 | 3820 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3821 | |
3822 | /* | |
3823 | * If sharing possible, alert mmu notifiers of worst case. | |
3824 | */ | |
6f4f13e8 JG |
3825 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3826 | end); | |
ac46d4f3 JG |
3827 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3828 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3829 | address = start; |
569f48b8 | 3830 | for (; address < end; address += sz) { |
7868a208 | 3831 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3832 | if (!ptep) |
c7546f8f DG |
3833 | continue; |
3834 | ||
cb900f41 | 3835 | ptl = huge_pte_lock(h, mm, ptep); |
31d49da5 AK |
3836 | if (huge_pmd_unshare(mm, &address, ptep)) { |
3837 | spin_unlock(ptl); | |
dff11abe MK |
3838 | /* |
3839 | * We just unmapped a page of PMDs by clearing a PUD. | |
3840 | * The caller's TLB flush range should cover this area. | |
3841 | */ | |
31d49da5 AK |
3842 | continue; |
3843 | } | |
39dde65c | 3844 | |
6629326b | 3845 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3846 | if (huge_pte_none(pte)) { |
3847 | spin_unlock(ptl); | |
3848 | continue; | |
3849 | } | |
6629326b HD |
3850 | |
3851 | /* | |
9fbc1f63 NH |
3852 | * Migrating hugepage or HWPoisoned hugepage is already |
3853 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3854 | */ |
9fbc1f63 | 3855 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3856 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3857 | spin_unlock(ptl); |
3858 | continue; | |
8c4894c6 | 3859 | } |
6629326b HD |
3860 | |
3861 | page = pte_page(pte); | |
04f2cbe3 MG |
3862 | /* |
3863 | * If a reference page is supplied, it is because a specific | |
3864 | * page is being unmapped, not a range. Ensure the page we | |
3865 | * are about to unmap is the actual page of interest. | |
3866 | */ | |
3867 | if (ref_page) { | |
31d49da5 AK |
3868 | if (page != ref_page) { |
3869 | spin_unlock(ptl); | |
3870 | continue; | |
3871 | } | |
04f2cbe3 MG |
3872 | /* |
3873 | * Mark the VMA as having unmapped its page so that | |
3874 | * future faults in this VMA will fail rather than | |
3875 | * looking like data was lost | |
3876 | */ | |
3877 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
3878 | } | |
3879 | ||
c7546f8f | 3880 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 3881 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 3882 | if (huge_pte_dirty(pte)) |
6649a386 | 3883 | set_page_dirty(page); |
9e81130b | 3884 | |
5d317b2b | 3885 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 3886 | page_remove_rmap(page, true); |
31d49da5 | 3887 | |
cb900f41 | 3888 | spin_unlock(ptl); |
e77b0852 | 3889 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
3890 | /* |
3891 | * Bail out after unmapping reference page if supplied | |
3892 | */ | |
3893 | if (ref_page) | |
3894 | break; | |
fe1668ae | 3895 | } |
ac46d4f3 | 3896 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 3897 | tlb_end_vma(tlb, vma); |
1da177e4 | 3898 | } |
63551ae0 | 3899 | |
d833352a MG |
3900 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
3901 | struct vm_area_struct *vma, unsigned long start, | |
3902 | unsigned long end, struct page *ref_page) | |
3903 | { | |
3904 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
3905 | ||
3906 | /* | |
3907 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
3908 | * test will fail on a vma being torn down, and not grab a page table | |
3909 | * on its way out. We're lucky that the flag has such an appropriate | |
3910 | * name, and can in fact be safely cleared here. We could clear it | |
3911 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 3912 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 3913 | * because in the context this is called, the VMA is about to be |
c8c06efa | 3914 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
3915 | */ |
3916 | vma->vm_flags &= ~VM_MAYSHARE; | |
3917 | } | |
3918 | ||
502717f4 | 3919 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 3920 | unsigned long end, struct page *ref_page) |
502717f4 | 3921 | { |
24669e58 AK |
3922 | struct mm_struct *mm; |
3923 | struct mmu_gather tlb; | |
dff11abe MK |
3924 | unsigned long tlb_start = start; |
3925 | unsigned long tlb_end = end; | |
3926 | ||
3927 | /* | |
3928 | * If shared PMDs were possibly used within this vma range, adjust | |
3929 | * start/end for worst case tlb flushing. | |
3930 | * Note that we can not be sure if PMDs are shared until we try to | |
3931 | * unmap pages. However, we want to make sure TLB flushing covers | |
3932 | * the largest possible range. | |
3933 | */ | |
3934 | adjust_range_if_pmd_sharing_possible(vma, &tlb_start, &tlb_end); | |
24669e58 AK |
3935 | |
3936 | mm = vma->vm_mm; | |
3937 | ||
dff11abe | 3938 | tlb_gather_mmu(&tlb, mm, tlb_start, tlb_end); |
24669e58 | 3939 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
dff11abe | 3940 | tlb_finish_mmu(&tlb, tlb_start, tlb_end); |
502717f4 KC |
3941 | } |
3942 | ||
04f2cbe3 MG |
3943 | /* |
3944 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
3945 | * mappping it owns the reserve page for. The intention is to unmap the page | |
3946 | * from other VMAs and let the children be SIGKILLed if they are faulting the | |
3947 | * same region. | |
3948 | */ | |
2f4612af DB |
3949 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
3950 | struct page *page, unsigned long address) | |
04f2cbe3 | 3951 | { |
7526674d | 3952 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
3953 | struct vm_area_struct *iter_vma; |
3954 | struct address_space *mapping; | |
04f2cbe3 MG |
3955 | pgoff_t pgoff; |
3956 | ||
3957 | /* | |
3958 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
3959 | * from page cache lookup which is in HPAGE_SIZE units. | |
3960 | */ | |
7526674d | 3961 | address = address & huge_page_mask(h); |
36e4f20a MH |
3962 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
3963 | vma->vm_pgoff; | |
93c76a3d | 3964 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 3965 | |
4eb2b1dc MG |
3966 | /* |
3967 | * Take the mapping lock for the duration of the table walk. As | |
3968 | * this mapping should be shared between all the VMAs, | |
3969 | * __unmap_hugepage_range() is called as the lock is already held | |
3970 | */ | |
83cde9e8 | 3971 | i_mmap_lock_write(mapping); |
6b2dbba8 | 3972 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
3973 | /* Do not unmap the current VMA */ |
3974 | if (iter_vma == vma) | |
3975 | continue; | |
3976 | ||
2f84a899 MG |
3977 | /* |
3978 | * Shared VMAs have their own reserves and do not affect | |
3979 | * MAP_PRIVATE accounting but it is possible that a shared | |
3980 | * VMA is using the same page so check and skip such VMAs. | |
3981 | */ | |
3982 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
3983 | continue; | |
3984 | ||
04f2cbe3 MG |
3985 | /* |
3986 | * Unmap the page from other VMAs without their own reserves. | |
3987 | * They get marked to be SIGKILLed if they fault in these | |
3988 | * areas. This is because a future no-page fault on this VMA | |
3989 | * could insert a zeroed page instead of the data existing | |
3990 | * from the time of fork. This would look like data corruption | |
3991 | */ | |
3992 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
3993 | unmap_hugepage_range(iter_vma, address, |
3994 | address + huge_page_size(h), page); | |
04f2cbe3 | 3995 | } |
83cde9e8 | 3996 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
3997 | } |
3998 | ||
0fe6e20b NH |
3999 | /* |
4000 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
4001 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
4002 | * cannot race with other handlers or page migration. | |
4003 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 4004 | */ |
2b740303 | 4005 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 4006 | unsigned long address, pte_t *ptep, |
3999f52e | 4007 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 4008 | { |
3999f52e | 4009 | pte_t pte; |
a5516438 | 4010 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 4011 | struct page *old_page, *new_page; |
2b740303 SJ |
4012 | int outside_reserve = 0; |
4013 | vm_fault_t ret = 0; | |
974e6d66 | 4014 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 4015 | struct mmu_notifier_range range; |
1e8f889b | 4016 | |
3999f52e | 4017 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
4018 | old_page = pte_page(pte); |
4019 | ||
04f2cbe3 | 4020 | retry_avoidcopy: |
1e8f889b DG |
4021 | /* If no-one else is actually using this page, avoid the copy |
4022 | * and just make the page writable */ | |
37a2140d | 4023 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 4024 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 4025 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 4026 | return 0; |
1e8f889b DG |
4027 | } |
4028 | ||
04f2cbe3 MG |
4029 | /* |
4030 | * If the process that created a MAP_PRIVATE mapping is about to | |
4031 | * perform a COW due to a shared page count, attempt to satisfy | |
4032 | * the allocation without using the existing reserves. The pagecache | |
4033 | * page is used to determine if the reserve at this address was | |
4034 | * consumed or not. If reserves were used, a partial faulted mapping | |
4035 | * at the time of fork() could consume its reserves on COW instead | |
4036 | * of the full address range. | |
4037 | */ | |
5944d011 | 4038 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
4039 | old_page != pagecache_page) |
4040 | outside_reserve = 1; | |
4041 | ||
09cbfeaf | 4042 | get_page(old_page); |
b76c8cfb | 4043 | |
ad4404a2 DB |
4044 | /* |
4045 | * Drop page table lock as buddy allocator may be called. It will | |
4046 | * be acquired again before returning to the caller, as expected. | |
4047 | */ | |
cb900f41 | 4048 | spin_unlock(ptl); |
5b7a1d40 | 4049 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 4050 | |
2fc39cec | 4051 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
4052 | /* |
4053 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4054 | * it is due to references held by a child and an insufficient | |
4055 | * huge page pool. To guarantee the original mappers | |
4056 | * reliability, unmap the page from child processes. The child | |
4057 | * may get SIGKILLed if it later faults. | |
4058 | */ | |
4059 | if (outside_reserve) { | |
09cbfeaf | 4060 | put_page(old_page); |
04f2cbe3 | 4061 | BUG_ON(huge_pte_none(pte)); |
5b7a1d40 | 4062 | unmap_ref_private(mm, vma, old_page, haddr); |
2f4612af DB |
4063 | BUG_ON(huge_pte_none(pte)); |
4064 | spin_lock(ptl); | |
5b7a1d40 | 4065 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
4066 | if (likely(ptep && |
4067 | pte_same(huge_ptep_get(ptep), pte))) | |
4068 | goto retry_avoidcopy; | |
4069 | /* | |
4070 | * race occurs while re-acquiring page table | |
4071 | * lock, and our job is done. | |
4072 | */ | |
4073 | return 0; | |
04f2cbe3 MG |
4074 | } |
4075 | ||
2b740303 | 4076 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 4077 | goto out_release_old; |
1e8f889b DG |
4078 | } |
4079 | ||
0fe6e20b NH |
4080 | /* |
4081 | * When the original hugepage is shared one, it does not have | |
4082 | * anon_vma prepared. | |
4083 | */ | |
44e2aa93 | 4084 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
4085 | ret = VM_FAULT_OOM; |
4086 | goto out_release_all; | |
44e2aa93 | 4087 | } |
0fe6e20b | 4088 | |
974e6d66 | 4089 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 4090 | pages_per_huge_page(h)); |
0ed361de | 4091 | __SetPageUptodate(new_page); |
1e8f889b | 4092 | |
7269f999 | 4093 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 4094 | haddr + huge_page_size(h)); |
ac46d4f3 | 4095 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 4096 | |
b76c8cfb | 4097 | /* |
cb900f41 | 4098 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
4099 | * before the page tables are altered |
4100 | */ | |
cb900f41 | 4101 | spin_lock(ptl); |
5b7a1d40 | 4102 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 4103 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
07443a85 JK |
4104 | ClearPagePrivate(new_page); |
4105 | ||
1e8f889b | 4106 | /* Break COW */ |
5b7a1d40 | 4107 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 4108 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 4109 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 4110 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 4111 | page_remove_rmap(old_page, true); |
5b7a1d40 | 4112 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
cb6acd01 | 4113 | set_page_huge_active(new_page); |
1e8f889b DG |
4114 | /* Make the old page be freed below */ |
4115 | new_page = old_page; | |
4116 | } | |
cb900f41 | 4117 | spin_unlock(ptl); |
ac46d4f3 | 4118 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 4119 | out_release_all: |
5b7a1d40 | 4120 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 4121 | put_page(new_page); |
ad4404a2 | 4122 | out_release_old: |
09cbfeaf | 4123 | put_page(old_page); |
8312034f | 4124 | |
ad4404a2 DB |
4125 | spin_lock(ptl); /* Caller expects lock to be held */ |
4126 | return ret; | |
1e8f889b DG |
4127 | } |
4128 | ||
04f2cbe3 | 4129 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
4130 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
4131 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
4132 | { |
4133 | struct address_space *mapping; | |
e7c4b0bf | 4134 | pgoff_t idx; |
04f2cbe3 MG |
4135 | |
4136 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 4137 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
4138 | |
4139 | return find_lock_page(mapping, idx); | |
4140 | } | |
4141 | ||
3ae77f43 HD |
4142 | /* |
4143 | * Return whether there is a pagecache page to back given address within VMA. | |
4144 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4145 | */ | |
4146 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
4147 | struct vm_area_struct *vma, unsigned long address) |
4148 | { | |
4149 | struct address_space *mapping; | |
4150 | pgoff_t idx; | |
4151 | struct page *page; | |
4152 | ||
4153 | mapping = vma->vm_file->f_mapping; | |
4154 | idx = vma_hugecache_offset(h, vma, address); | |
4155 | ||
4156 | page = find_get_page(mapping, idx); | |
4157 | if (page) | |
4158 | put_page(page); | |
4159 | return page != NULL; | |
4160 | } | |
4161 | ||
ab76ad54 MK |
4162 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
4163 | pgoff_t idx) | |
4164 | { | |
4165 | struct inode *inode = mapping->host; | |
4166 | struct hstate *h = hstate_inode(inode); | |
4167 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4168 | ||
4169 | if (err) | |
4170 | return err; | |
4171 | ClearPagePrivate(page); | |
4172 | ||
22146c3c MK |
4173 | /* |
4174 | * set page dirty so that it will not be removed from cache/file | |
4175 | * by non-hugetlbfs specific code paths. | |
4176 | */ | |
4177 | set_page_dirty(page); | |
4178 | ||
ab76ad54 MK |
4179 | spin_lock(&inode->i_lock); |
4180 | inode->i_blocks += blocks_per_huge_page(h); | |
4181 | spin_unlock(&inode->i_lock); | |
4182 | return 0; | |
4183 | } | |
4184 | ||
2b740303 SJ |
4185 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
4186 | struct vm_area_struct *vma, | |
4187 | struct address_space *mapping, pgoff_t idx, | |
4188 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 4189 | { |
a5516438 | 4190 | struct hstate *h = hstate_vma(vma); |
2b740303 | 4191 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 4192 | int anon_rmap = 0; |
4c887265 | 4193 | unsigned long size; |
4c887265 | 4194 | struct page *page; |
1e8f889b | 4195 | pte_t new_pte; |
cb900f41 | 4196 | spinlock_t *ptl; |
285b8dca | 4197 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 4198 | bool new_page = false; |
4c887265 | 4199 | |
04f2cbe3 MG |
4200 | /* |
4201 | * Currently, we are forced to kill the process in the event the | |
4202 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4203 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4204 | */ |
4205 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4206 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4207 | current->pid); |
04f2cbe3 MG |
4208 | return ret; |
4209 | } | |
4210 | ||
4c887265 | 4211 | /* |
87bf91d3 MK |
4212 | * We can not race with truncation due to holding i_mmap_rwsem. |
4213 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4214 | * once for faults beyond end of file. | |
4c887265 | 4215 | */ |
87bf91d3 MK |
4216 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4217 | if (idx >= size) | |
4218 | goto out; | |
4219 | ||
6bda666a CL |
4220 | retry: |
4221 | page = find_lock_page(mapping, idx); | |
4222 | if (!page) { | |
1a1aad8a MK |
4223 | /* |
4224 | * Check for page in userfault range | |
4225 | */ | |
4226 | if (userfaultfd_missing(vma)) { | |
4227 | u32 hash; | |
4228 | struct vm_fault vmf = { | |
4229 | .vma = vma, | |
285b8dca | 4230 | .address = haddr, |
1a1aad8a MK |
4231 | .flags = flags, |
4232 | /* | |
4233 | * Hard to debug if it ends up being | |
4234 | * used by a callee that assumes | |
4235 | * something about the other | |
4236 | * uninitialized fields... same as in | |
4237 | * memory.c | |
4238 | */ | |
4239 | }; | |
4240 | ||
4241 | /* | |
c0d0381a MK |
4242 | * hugetlb_fault_mutex and i_mmap_rwsem must be |
4243 | * dropped before handling userfault. Reacquire | |
4244 | * after handling fault to make calling code simpler. | |
1a1aad8a | 4245 | */ |
188b04a7 | 4246 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
1a1aad8a | 4247 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4248 | i_mmap_unlock_read(mapping); |
1a1aad8a | 4249 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); |
c0d0381a | 4250 | i_mmap_lock_read(mapping); |
1a1aad8a MK |
4251 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
4252 | goto out; | |
4253 | } | |
4254 | ||
285b8dca | 4255 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4256 | if (IS_ERR(page)) { |
4643d67e MK |
4257 | /* |
4258 | * Returning error will result in faulting task being | |
4259 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4260 | * tasks from racing to fault in the same page which | |
4261 | * could result in false unable to allocate errors. | |
4262 | * Page migration does not take the fault mutex, but | |
4263 | * does a clear then write of pte's under page table | |
4264 | * lock. Page fault code could race with migration, | |
4265 | * notice the clear pte and try to allocate a page | |
4266 | * here. Before returning error, get ptl and make | |
4267 | * sure there really is no pte entry. | |
4268 | */ | |
4269 | ptl = huge_pte_lock(h, mm, ptep); | |
4270 | if (!huge_pte_none(huge_ptep_get(ptep))) { | |
4271 | ret = 0; | |
4272 | spin_unlock(ptl); | |
4273 | goto out; | |
4274 | } | |
4275 | spin_unlock(ptl); | |
2b740303 | 4276 | ret = vmf_error(PTR_ERR(page)); |
6bda666a CL |
4277 | goto out; |
4278 | } | |
47ad8475 | 4279 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4280 | __SetPageUptodate(page); |
cb6acd01 | 4281 | new_page = true; |
ac9b9c66 | 4282 | |
f83a275d | 4283 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4284 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4285 | if (err) { |
4286 | put_page(page); | |
6bda666a CL |
4287 | if (err == -EEXIST) |
4288 | goto retry; | |
4289 | goto out; | |
4290 | } | |
23be7468 | 4291 | } else { |
6bda666a | 4292 | lock_page(page); |
0fe6e20b NH |
4293 | if (unlikely(anon_vma_prepare(vma))) { |
4294 | ret = VM_FAULT_OOM; | |
4295 | goto backout_unlocked; | |
4296 | } | |
409eb8c2 | 4297 | anon_rmap = 1; |
23be7468 | 4298 | } |
0fe6e20b | 4299 | } else { |
998b4382 NH |
4300 | /* |
4301 | * If memory error occurs between mmap() and fault, some process | |
4302 | * don't have hwpoisoned swap entry for errored virtual address. | |
4303 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4304 | */ | |
4305 | if (unlikely(PageHWPoison(page))) { | |
32f84528 | 4306 | ret = VM_FAULT_HWPOISON | |
972dc4de | 4307 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4308 | goto backout_unlocked; |
4309 | } | |
6bda666a | 4310 | } |
1e8f889b | 4311 | |
57303d80 AW |
4312 | /* |
4313 | * If we are going to COW a private mapping later, we examine the | |
4314 | * pending reservations for this page now. This will ensure that | |
4315 | * any allocations necessary to record that reservation occur outside | |
4316 | * the spinlock. | |
4317 | */ | |
5e911373 | 4318 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4319 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4320 | ret = VM_FAULT_OOM; |
4321 | goto backout_unlocked; | |
4322 | } | |
5e911373 | 4323 | /* Just decrements count, does not deallocate */ |
285b8dca | 4324 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4325 | } |
57303d80 | 4326 | |
8bea8052 | 4327 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 4328 | ret = 0; |
7f2e9525 | 4329 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4330 | goto backout; |
4331 | ||
07443a85 JK |
4332 | if (anon_rmap) { |
4333 | ClearPagePrivate(page); | |
285b8dca | 4334 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4335 | } else |
53f9263b | 4336 | page_dup_rmap(page, true); |
1e8f889b DG |
4337 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4338 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4339 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4340 | |
5d317b2b | 4341 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4342 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4343 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4344 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4345 | } |
4346 | ||
cb900f41 | 4347 | spin_unlock(ptl); |
cb6acd01 MK |
4348 | |
4349 | /* | |
4350 | * Only make newly allocated pages active. Existing pages found | |
4351 | * in the pagecache could be !page_huge_active() if they have been | |
4352 | * isolated for migration. | |
4353 | */ | |
4354 | if (new_page) | |
4355 | set_page_huge_active(page); | |
4356 | ||
4c887265 AL |
4357 | unlock_page(page); |
4358 | out: | |
ac9b9c66 | 4359 | return ret; |
4c887265 AL |
4360 | |
4361 | backout: | |
cb900f41 | 4362 | spin_unlock(ptl); |
2b26736c | 4363 | backout_unlocked: |
4c887265 | 4364 | unlock_page(page); |
285b8dca | 4365 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4366 | put_page(page); |
4367 | goto out; | |
ac9b9c66 HD |
4368 | } |
4369 | ||
8382d914 | 4370 | #ifdef CONFIG_SMP |
188b04a7 | 4371 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4372 | { |
4373 | unsigned long key[2]; | |
4374 | u32 hash; | |
4375 | ||
1b426bac MK |
4376 | key[0] = (unsigned long) mapping; |
4377 | key[1] = idx; | |
8382d914 | 4378 | |
55254636 | 4379 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4380 | |
4381 | return hash & (num_fault_mutexes - 1); | |
4382 | } | |
4383 | #else | |
4384 | /* | |
4385 | * For uniprocesor systems we always use a single mutex, so just | |
4386 | * return 0 and avoid the hashing overhead. | |
4387 | */ | |
188b04a7 | 4388 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4389 | { |
4390 | return 0; | |
4391 | } | |
4392 | #endif | |
4393 | ||
2b740303 | 4394 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4395 | unsigned long address, unsigned int flags) |
86e5216f | 4396 | { |
8382d914 | 4397 | pte_t *ptep, entry; |
cb900f41 | 4398 | spinlock_t *ptl; |
2b740303 | 4399 | vm_fault_t ret; |
8382d914 DB |
4400 | u32 hash; |
4401 | pgoff_t idx; | |
0fe6e20b | 4402 | struct page *page = NULL; |
57303d80 | 4403 | struct page *pagecache_page = NULL; |
a5516438 | 4404 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4405 | struct address_space *mapping; |
0f792cf9 | 4406 | int need_wait_lock = 0; |
285b8dca | 4407 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4408 | |
285b8dca | 4409 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 4410 | if (ptep) { |
c0d0381a MK |
4411 | /* |
4412 | * Since we hold no locks, ptep could be stale. That is | |
4413 | * OK as we are only making decisions based on content and | |
4414 | * not actually modifying content here. | |
4415 | */ | |
fd6a03ed | 4416 | entry = huge_ptep_get(ptep); |
290408d4 | 4417 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4418 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4419 | return 0; |
4420 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4421 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4422 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
ddeaab32 MK |
4423 | } else { |
4424 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4425 | if (!ptep) | |
4426 | return VM_FAULT_OOM; | |
fd6a03ed NH |
4427 | } |
4428 | ||
c0d0381a MK |
4429 | /* |
4430 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
4431 | * until finished with ptep. This serves two purposes: |
4432 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
4433 | * and making the ptep no longer valid. | |
4434 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
4435 | * |
4436 | * ptep could have already be assigned via huge_pte_offset. That | |
4437 | * is OK, as huge_pte_alloc will return the same value unless | |
4438 | * something has changed. | |
4439 | */ | |
8382d914 | 4440 | mapping = vma->vm_file->f_mapping; |
c0d0381a MK |
4441 | i_mmap_lock_read(mapping); |
4442 | ptep = huge_pte_alloc(mm, haddr, huge_page_size(h)); | |
4443 | if (!ptep) { | |
4444 | i_mmap_unlock_read(mapping); | |
4445 | return VM_FAULT_OOM; | |
4446 | } | |
8382d914 | 4447 | |
3935baa9 DG |
4448 | /* |
4449 | * Serialize hugepage allocation and instantiation, so that we don't | |
4450 | * get spurious allocation failures if two CPUs race to instantiate | |
4451 | * the same page in the page cache. | |
4452 | */ | |
c0d0381a | 4453 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 4454 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4455 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4456 | |
7f2e9525 GS |
4457 | entry = huge_ptep_get(ptep); |
4458 | if (huge_pte_none(entry)) { | |
8382d914 | 4459 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4460 | goto out_mutex; |
3935baa9 | 4461 | } |
86e5216f | 4462 | |
83c54070 | 4463 | ret = 0; |
1e8f889b | 4464 | |
0f792cf9 NH |
4465 | /* |
4466 | * entry could be a migration/hwpoison entry at this point, so this | |
4467 | * check prevents the kernel from going below assuming that we have | |
4468 | * a active hugepage in pagecache. This goto expects the 2nd page fault, | |
4469 | * and is_hugetlb_entry_(migration|hwpoisoned) check will properly | |
4470 | * handle it. | |
4471 | */ | |
4472 | if (!pte_present(entry)) | |
4473 | goto out_mutex; | |
4474 | ||
57303d80 AW |
4475 | /* |
4476 | * If we are going to COW the mapping later, we examine the pending | |
4477 | * reservations for this page now. This will ensure that any | |
4478 | * allocations necessary to record that reservation occur outside the | |
4479 | * spinlock. For private mappings, we also lookup the pagecache | |
4480 | * page now as it is used to determine if a reservation has been | |
4481 | * consumed. | |
4482 | */ | |
106c992a | 4483 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4484 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4485 | ret = VM_FAULT_OOM; |
b4d1d99f | 4486 | goto out_mutex; |
2b26736c | 4487 | } |
5e911373 | 4488 | /* Just decrements count, does not deallocate */ |
285b8dca | 4489 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4490 | |
f83a275d | 4491 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4492 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4493 | vma, haddr); |
57303d80 AW |
4494 | } |
4495 | ||
0f792cf9 NH |
4496 | ptl = huge_pte_lock(h, mm, ptep); |
4497 | ||
4498 | /* Check for a racing update before calling hugetlb_cow */ | |
4499 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4500 | goto out_ptl; | |
4501 | ||
56c9cfb1 NH |
4502 | /* |
4503 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4504 | * pagecache_page, so here we need take the former one | |
4505 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4506 | */ |
4507 | page = pte_page(entry); | |
4508 | if (page != pagecache_page) | |
0f792cf9 NH |
4509 | if (!trylock_page(page)) { |
4510 | need_wait_lock = 1; | |
4511 | goto out_ptl; | |
4512 | } | |
b4d1d99f | 4513 | |
0f792cf9 | 4514 | get_page(page); |
b4d1d99f | 4515 | |
788c7df4 | 4516 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4517 | if (!huge_pte_write(entry)) { |
974e6d66 | 4518 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4519 | pagecache_page, ptl); |
0f792cf9 | 4520 | goto out_put_page; |
b4d1d99f | 4521 | } |
106c992a | 4522 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4523 | } |
4524 | entry = pte_mkyoung(entry); | |
285b8dca | 4525 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4526 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4527 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4528 | out_put_page: |
4529 | if (page != pagecache_page) | |
4530 | unlock_page(page); | |
4531 | put_page(page); | |
cb900f41 KS |
4532 | out_ptl: |
4533 | spin_unlock(ptl); | |
57303d80 AW |
4534 | |
4535 | if (pagecache_page) { | |
4536 | unlock_page(pagecache_page); | |
4537 | put_page(pagecache_page); | |
4538 | } | |
b4d1d99f | 4539 | out_mutex: |
c672c7f2 | 4540 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4541 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
4542 | /* |
4543 | * Generally it's safe to hold refcount during waiting page lock. But | |
4544 | * here we just wait to defer the next page fault to avoid busy loop and | |
4545 | * the page is not used after unlocked before returning from the current | |
4546 | * page fault. So we are safe from accessing freed page, even if we wait | |
4547 | * here without taking refcount. | |
4548 | */ | |
4549 | if (need_wait_lock) | |
4550 | wait_on_page_locked(page); | |
1e8f889b | 4551 | return ret; |
86e5216f AL |
4552 | } |
4553 | ||
8fb5debc MK |
4554 | /* |
4555 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4556 | * modifications for huge pages. | |
4557 | */ | |
4558 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4559 | pte_t *dst_pte, | |
4560 | struct vm_area_struct *dst_vma, | |
4561 | unsigned long dst_addr, | |
4562 | unsigned long src_addr, | |
4563 | struct page **pagep) | |
4564 | { | |
1e392147 AA |
4565 | struct address_space *mapping; |
4566 | pgoff_t idx; | |
4567 | unsigned long size; | |
1c9e8def | 4568 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4569 | struct hstate *h = hstate_vma(dst_vma); |
4570 | pte_t _dst_pte; | |
4571 | spinlock_t *ptl; | |
4572 | int ret; | |
4573 | struct page *page; | |
4574 | ||
4575 | if (!*pagep) { | |
4576 | ret = -ENOMEM; | |
4577 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4578 | if (IS_ERR(page)) | |
4579 | goto out; | |
4580 | ||
4581 | ret = copy_huge_page_from_user(page, | |
4582 | (const void __user *) src_addr, | |
810a56b9 | 4583 | pages_per_huge_page(h), false); |
8fb5debc MK |
4584 | |
4585 | /* fallback to copy_from_user outside mmap_sem */ | |
4586 | if (unlikely(ret)) { | |
9e368259 | 4587 | ret = -ENOENT; |
8fb5debc MK |
4588 | *pagep = page; |
4589 | /* don't free the page */ | |
4590 | goto out; | |
4591 | } | |
4592 | } else { | |
4593 | page = *pagep; | |
4594 | *pagep = NULL; | |
4595 | } | |
4596 | ||
4597 | /* | |
4598 | * The memory barrier inside __SetPageUptodate makes sure that | |
4599 | * preceding stores to the page contents become visible before | |
4600 | * the set_pte_at() write. | |
4601 | */ | |
4602 | __SetPageUptodate(page); | |
8fb5debc | 4603 | |
1e392147 AA |
4604 | mapping = dst_vma->vm_file->f_mapping; |
4605 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4606 | ||
1c9e8def MK |
4607 | /* |
4608 | * If shared, add to page cache | |
4609 | */ | |
4610 | if (vm_shared) { | |
1e392147 AA |
4611 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4612 | ret = -EFAULT; | |
4613 | if (idx >= size) | |
4614 | goto out_release_nounlock; | |
1c9e8def | 4615 | |
1e392147 AA |
4616 | /* |
4617 | * Serialization between remove_inode_hugepages() and | |
4618 | * huge_add_to_page_cache() below happens through the | |
4619 | * hugetlb_fault_mutex_table that here must be hold by | |
4620 | * the caller. | |
4621 | */ | |
1c9e8def MK |
4622 | ret = huge_add_to_page_cache(page, mapping, idx); |
4623 | if (ret) | |
4624 | goto out_release_nounlock; | |
4625 | } | |
4626 | ||
8fb5debc MK |
4627 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4628 | spin_lock(ptl); | |
4629 | ||
1e392147 AA |
4630 | /* |
4631 | * Recheck the i_size after holding PT lock to make sure not | |
4632 | * to leave any page mapped (as page_mapped()) beyond the end | |
4633 | * of the i_size (remove_inode_hugepages() is strict about | |
4634 | * enforcing that). If we bail out here, we'll also leave a | |
4635 | * page in the radix tree in the vm_shared case beyond the end | |
4636 | * of the i_size, but remove_inode_hugepages() will take care | |
4637 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4638 | */ | |
4639 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4640 | ret = -EFAULT; | |
4641 | if (idx >= size) | |
4642 | goto out_release_unlock; | |
4643 | ||
8fb5debc MK |
4644 | ret = -EEXIST; |
4645 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4646 | goto out_release_unlock; | |
4647 | ||
1c9e8def MK |
4648 | if (vm_shared) { |
4649 | page_dup_rmap(page, true); | |
4650 | } else { | |
4651 | ClearPagePrivate(page); | |
4652 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); | |
4653 | } | |
8fb5debc MK |
4654 | |
4655 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4656 | if (dst_vma->vm_flags & VM_WRITE) | |
4657 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4658 | _dst_pte = pte_mkyoung(_dst_pte); | |
4659 | ||
4660 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4661 | ||
4662 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4663 | dst_vma->vm_flags & VM_WRITE); | |
4664 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4665 | ||
4666 | /* No need to invalidate - it was non-present before */ | |
4667 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4668 | ||
4669 | spin_unlock(ptl); | |
cb6acd01 | 4670 | set_page_huge_active(page); |
1c9e8def MK |
4671 | if (vm_shared) |
4672 | unlock_page(page); | |
8fb5debc MK |
4673 | ret = 0; |
4674 | out: | |
4675 | return ret; | |
4676 | out_release_unlock: | |
4677 | spin_unlock(ptl); | |
1c9e8def MK |
4678 | if (vm_shared) |
4679 | unlock_page(page); | |
5af10dfd | 4680 | out_release_nounlock: |
8fb5debc MK |
4681 | put_page(page); |
4682 | goto out; | |
4683 | } | |
4684 | ||
28a35716 ML |
4685 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4686 | struct page **pages, struct vm_area_struct **vmas, | |
4687 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 4688 | long i, unsigned int flags, int *locked) |
63551ae0 | 4689 | { |
d5d4b0aa KC |
4690 | unsigned long pfn_offset; |
4691 | unsigned long vaddr = *position; | |
28a35716 | 4692 | unsigned long remainder = *nr_pages; |
a5516438 | 4693 | struct hstate *h = hstate_vma(vma); |
2be7cfed | 4694 | int err = -EFAULT; |
63551ae0 | 4695 | |
63551ae0 | 4696 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4697 | pte_t *pte; |
cb900f41 | 4698 | spinlock_t *ptl = NULL; |
2a15efc9 | 4699 | int absent; |
4c887265 | 4700 | struct page *page; |
63551ae0 | 4701 | |
02057967 DR |
4702 | /* |
4703 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4704 | * potentially allocating memory. | |
4705 | */ | |
fa45f116 | 4706 | if (fatal_signal_pending(current)) { |
02057967 DR |
4707 | remainder = 0; |
4708 | break; | |
4709 | } | |
4710 | ||
4c887265 AL |
4711 | /* |
4712 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4713 | * each hugepage. We have to make sure we get the |
4c887265 | 4714 | * first, for the page indexing below to work. |
cb900f41 KS |
4715 | * |
4716 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4717 | */ |
7868a208 PA |
4718 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4719 | huge_page_size(h)); | |
cb900f41 KS |
4720 | if (pte) |
4721 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4722 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4723 | ||
4724 | /* | |
4725 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4726 | * an error where there's an empty slot with no huge pagecache |
4727 | * to back it. This way, we avoid allocating a hugepage, and | |
4728 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4729 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4730 | */ |
3ae77f43 HD |
4731 | if (absent && (flags & FOLL_DUMP) && |
4732 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4733 | if (pte) |
4734 | spin_unlock(ptl); | |
2a15efc9 HD |
4735 | remainder = 0; |
4736 | break; | |
4737 | } | |
63551ae0 | 4738 | |
9cc3a5bd NH |
4739 | /* |
4740 | * We need call hugetlb_fault for both hugepages under migration | |
4741 | * (in which case hugetlb_fault waits for the migration,) and | |
4742 | * hwpoisoned hugepages (in which case we need to prevent the | |
4743 | * caller from accessing to them.) In order to do this, we use | |
4744 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4745 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4746 | * both cases, and because we can't follow correct pages | |
4747 | * directly from any kind of swap entries. | |
4748 | */ | |
4749 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4750 | ((flags & FOLL_WRITE) && |
4751 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4752 | vm_fault_t ret; |
87ffc118 | 4753 | unsigned int fault_flags = 0; |
63551ae0 | 4754 | |
cb900f41 KS |
4755 | if (pte) |
4756 | spin_unlock(ptl); | |
87ffc118 AA |
4757 | if (flags & FOLL_WRITE) |
4758 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 4759 | if (locked) |
71335f37 PX |
4760 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
4761 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
4762 | if (flags & FOLL_NOWAIT) |
4763 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4764 | FAULT_FLAG_RETRY_NOWAIT; | |
4765 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
4766 | /* |
4767 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
4768 | * FAULT_FLAG_TRIED can co-exist | |
4769 | */ | |
87ffc118 AA |
4770 | fault_flags |= FAULT_FLAG_TRIED; |
4771 | } | |
4772 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4773 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4774 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4775 | remainder = 0; |
4776 | break; | |
4777 | } | |
4778 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 4779 | if (locked && |
1ac25013 | 4780 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 4781 | *locked = 0; |
87ffc118 AA |
4782 | *nr_pages = 0; |
4783 | /* | |
4784 | * VM_FAULT_RETRY must not return an | |
4785 | * error, it will return zero | |
4786 | * instead. | |
4787 | * | |
4788 | * No need to update "position" as the | |
4789 | * caller will not check it after | |
4790 | * *nr_pages is set to 0. | |
4791 | */ | |
4792 | return i; | |
4793 | } | |
4794 | continue; | |
4c887265 AL |
4795 | } |
4796 | ||
a5516438 | 4797 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4798 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 4799 | |
acbfb087 ZL |
4800 | /* |
4801 | * If subpage information not requested, update counters | |
4802 | * and skip the same_page loop below. | |
4803 | */ | |
4804 | if (!pages && !vmas && !pfn_offset && | |
4805 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4806 | (remainder >= pages_per_huge_page(h))) { | |
4807 | vaddr += huge_page_size(h); | |
4808 | remainder -= pages_per_huge_page(h); | |
4809 | i += pages_per_huge_page(h); | |
4810 | spin_unlock(ptl); | |
4811 | continue; | |
4812 | } | |
4813 | ||
d5d4b0aa | 4814 | same_page: |
d6692183 | 4815 | if (pages) { |
2a15efc9 | 4816 | pages[i] = mem_map_offset(page, pfn_offset); |
3faa52c0 JH |
4817 | /* |
4818 | * try_grab_page() should always succeed here, because: | |
4819 | * a) we hold the ptl lock, and b) we've just checked | |
4820 | * that the huge page is present in the page tables. If | |
4821 | * the huge page is present, then the tail pages must | |
4822 | * also be present. The ptl prevents the head page and | |
4823 | * tail pages from being rearranged in any way. So this | |
4824 | * page must be available at this point, unless the page | |
4825 | * refcount overflowed: | |
4826 | */ | |
4827 | if (WARN_ON_ONCE(!try_grab_page(pages[i], flags))) { | |
4828 | spin_unlock(ptl); | |
4829 | remainder = 0; | |
4830 | err = -ENOMEM; | |
4831 | break; | |
4832 | } | |
d6692183 | 4833 | } |
63551ae0 DG |
4834 | |
4835 | if (vmas) | |
4836 | vmas[i] = vma; | |
4837 | ||
4838 | vaddr += PAGE_SIZE; | |
d5d4b0aa | 4839 | ++pfn_offset; |
63551ae0 DG |
4840 | --remainder; |
4841 | ++i; | |
d5d4b0aa | 4842 | if (vaddr < vma->vm_end && remainder && |
a5516438 | 4843 | pfn_offset < pages_per_huge_page(h)) { |
d5d4b0aa KC |
4844 | /* |
4845 | * We use pfn_offset to avoid touching the pageframes | |
4846 | * of this compound page. | |
4847 | */ | |
4848 | goto same_page; | |
4849 | } | |
cb900f41 | 4850 | spin_unlock(ptl); |
63551ae0 | 4851 | } |
28a35716 | 4852 | *nr_pages = remainder; |
87ffc118 AA |
4853 | /* |
4854 | * setting position is actually required only if remainder is | |
4855 | * not zero but it's faster not to add a "if (remainder)" | |
4856 | * branch. | |
4857 | */ | |
63551ae0 DG |
4858 | *position = vaddr; |
4859 | ||
2be7cfed | 4860 | return i ? i : err; |
63551ae0 | 4861 | } |
8f860591 | 4862 | |
5491ae7b AK |
4863 | #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE |
4864 | /* | |
4865 | * ARCHes with special requirements for evicting HUGETLB backing TLB entries can | |
4866 | * implement this. | |
4867 | */ | |
4868 | #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end) | |
4869 | #endif | |
4870 | ||
7da4d641 | 4871 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
4872 | unsigned long address, unsigned long end, pgprot_t newprot) |
4873 | { | |
4874 | struct mm_struct *mm = vma->vm_mm; | |
4875 | unsigned long start = address; | |
4876 | pte_t *ptep; | |
4877 | pte_t pte; | |
a5516438 | 4878 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 4879 | unsigned long pages = 0; |
dff11abe | 4880 | bool shared_pmd = false; |
ac46d4f3 | 4881 | struct mmu_notifier_range range; |
dff11abe MK |
4882 | |
4883 | /* | |
4884 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 4885 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
4886 | * range if PMD sharing is possible. |
4887 | */ | |
7269f999 JG |
4888 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
4889 | 0, vma, mm, start, end); | |
ac46d4f3 | 4890 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
4891 | |
4892 | BUG_ON(address >= end); | |
ac46d4f3 | 4893 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 4894 | |
ac46d4f3 | 4895 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 4896 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 4897 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 4898 | spinlock_t *ptl; |
7868a208 | 4899 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
4900 | if (!ptep) |
4901 | continue; | |
cb900f41 | 4902 | ptl = huge_pte_lock(h, mm, ptep); |
7da4d641 PZ |
4903 | if (huge_pmd_unshare(mm, &address, ptep)) { |
4904 | pages++; | |
cb900f41 | 4905 | spin_unlock(ptl); |
dff11abe | 4906 | shared_pmd = true; |
39dde65c | 4907 | continue; |
7da4d641 | 4908 | } |
a8bda28d NH |
4909 | pte = huge_ptep_get(ptep); |
4910 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
4911 | spin_unlock(ptl); | |
4912 | continue; | |
4913 | } | |
4914 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
4915 | swp_entry_t entry = pte_to_swp_entry(pte); | |
4916 | ||
4917 | if (is_write_migration_entry(entry)) { | |
4918 | pte_t newpte; | |
4919 | ||
4920 | make_migration_entry_read(&entry); | |
4921 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
4922 | set_huge_swap_pte_at(mm, address, ptep, |
4923 | newpte, huge_page_size(h)); | |
a8bda28d NH |
4924 | pages++; |
4925 | } | |
4926 | spin_unlock(ptl); | |
4927 | continue; | |
4928 | } | |
4929 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
4930 | pte_t old_pte; |
4931 | ||
4932 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
4933 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 4934 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 4935 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 4936 | pages++; |
8f860591 | 4937 | } |
cb900f41 | 4938 | spin_unlock(ptl); |
8f860591 | 4939 | } |
d833352a | 4940 | /* |
c8c06efa | 4941 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 4942 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 4943 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
4944 | * and that page table be reused and filled with junk. If we actually |
4945 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 4946 | */ |
dff11abe | 4947 | if (shared_pmd) |
ac46d4f3 | 4948 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
4949 | else |
4950 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
4951 | /* |
4952 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
4953 | * page table protection not changing it to point to a new page. | |
4954 | * | |
ad56b738 | 4955 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 4956 | */ |
83cde9e8 | 4957 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 4958 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
4959 | |
4960 | return pages << h->order; | |
8f860591 ZY |
4961 | } |
4962 | ||
a1e78772 MG |
4963 | int hugetlb_reserve_pages(struct inode *inode, |
4964 | long from, long to, | |
5a6fe125 | 4965 | struct vm_area_struct *vma, |
ca16d140 | 4966 | vm_flags_t vm_flags) |
e4e574b7 | 4967 | { |
0db9d74e | 4968 | long ret, chg, add = -1; |
a5516438 | 4969 | struct hstate *h = hstate_inode(inode); |
90481622 | 4970 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 4971 | struct resv_map *resv_map; |
075a61d0 | 4972 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 4973 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 4974 | |
63489f8e MK |
4975 | /* This should never happen */ |
4976 | if (from > to) { | |
4977 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
4978 | return -EINVAL; | |
4979 | } | |
4980 | ||
17c9d12e MG |
4981 | /* |
4982 | * Only apply hugepage reservation if asked. At fault time, an | |
4983 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 4984 | * without using reserves |
17c9d12e | 4985 | */ |
ca16d140 | 4986 | if (vm_flags & VM_NORESERVE) |
17c9d12e MG |
4987 | return 0; |
4988 | ||
a1e78772 MG |
4989 | /* |
4990 | * Shared mappings base their reservation on the number of pages that | |
4991 | * are already allocated on behalf of the file. Private mappings need | |
4992 | * to reserve the full area even if read-only as mprotect() may be | |
4993 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
4994 | */ | |
9119a41e | 4995 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
4996 | /* |
4997 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
4998 | * called for inodes for which resv_maps were created (see | |
4999 | * hugetlbfs_get_inode). | |
5000 | */ | |
4e35f483 | 5001 | resv_map = inode_resv_map(inode); |
9119a41e | 5002 | |
0db9d74e | 5003 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
5004 | |
5005 | } else { | |
e9fe92ae | 5006 | /* Private mapping. */ |
9119a41e | 5007 | resv_map = resv_map_alloc(); |
17c9d12e MG |
5008 | if (!resv_map) |
5009 | return -ENOMEM; | |
5010 | ||
a1e78772 | 5011 | chg = to - from; |
84afd99b | 5012 | |
17c9d12e MG |
5013 | set_vma_resv_map(vma, resv_map); |
5014 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
5015 | } | |
5016 | ||
c50ac050 DH |
5017 | if (chg < 0) { |
5018 | ret = chg; | |
5019 | goto out_err; | |
5020 | } | |
8a630112 | 5021 | |
075a61d0 MA |
5022 | ret = hugetlb_cgroup_charge_cgroup_rsvd( |
5023 | hstate_index(h), chg * pages_per_huge_page(h), &h_cg); | |
5024 | ||
5025 | if (ret < 0) { | |
5026 | ret = -ENOMEM; | |
5027 | goto out_err; | |
5028 | } | |
5029 | ||
5030 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
5031 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
5032 | * of the resv_map. | |
5033 | */ | |
5034 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
5035 | } | |
5036 | ||
1c5ecae3 MK |
5037 | /* |
5038 | * There must be enough pages in the subpool for the mapping. If | |
5039 | * the subpool has a minimum size, there may be some global | |
5040 | * reservations already in place (gbl_reserve). | |
5041 | */ | |
5042 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
5043 | if (gbl_reserve < 0) { | |
c50ac050 | 5044 | ret = -ENOSPC; |
075a61d0 | 5045 | goto out_uncharge_cgroup; |
c50ac050 | 5046 | } |
5a6fe125 MG |
5047 | |
5048 | /* | |
17c9d12e | 5049 | * Check enough hugepages are available for the reservation. |
90481622 | 5050 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 5051 | */ |
1c5ecae3 | 5052 | ret = hugetlb_acct_memory(h, gbl_reserve); |
68842c9b | 5053 | if (ret < 0) { |
075a61d0 | 5054 | goto out_put_pages; |
68842c9b | 5055 | } |
17c9d12e MG |
5056 | |
5057 | /* | |
5058 | * Account for the reservations made. Shared mappings record regions | |
5059 | * that have reservations as they are shared by multiple VMAs. | |
5060 | * When the last VMA disappears, the region map says how much | |
5061 | * the reservation was and the page cache tells how much of | |
5062 | * the reservation was consumed. Private mappings are per-VMA and | |
5063 | * only the consumed reservations are tracked. When the VMA | |
5064 | * disappears, the original reservation is the VMA size and the | |
5065 | * consumed reservations are stored in the map. Hence, nothing | |
5066 | * else has to be done for private mappings here | |
5067 | */ | |
33039678 | 5068 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 5069 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
5070 | |
5071 | if (unlikely(add < 0)) { | |
5072 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 5073 | goto out_put_pages; |
0db9d74e | 5074 | } else if (unlikely(chg > add)) { |
33039678 MK |
5075 | /* |
5076 | * pages in this range were added to the reserve | |
5077 | * map between region_chg and region_add. This | |
5078 | * indicates a race with alloc_huge_page. Adjust | |
5079 | * the subpool and reserve counts modified above | |
5080 | * based on the difference. | |
5081 | */ | |
5082 | long rsv_adjust; | |
5083 | ||
075a61d0 MA |
5084 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
5085 | hstate_index(h), | |
5086 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5087 | ||
33039678 MK |
5088 | rsv_adjust = hugepage_subpool_put_pages(spool, |
5089 | chg - add); | |
5090 | hugetlb_acct_memory(h, -rsv_adjust); | |
5091 | } | |
5092 | } | |
a43a8c39 | 5093 | return 0; |
075a61d0 MA |
5094 | out_put_pages: |
5095 | /* put back original number of pages, chg */ | |
5096 | (void)hugepage_subpool_put_pages(spool, chg); | |
5097 | out_uncharge_cgroup: | |
5098 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5099 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 5100 | out_err: |
5e911373 | 5101 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
5102 | /* Only call region_abort if the region_chg succeeded but the |
5103 | * region_add failed or didn't run. | |
5104 | */ | |
5105 | if (chg >= 0 && add < 0) | |
5106 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
5107 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
5108 | kref_put(&resv_map->refs, resv_map_release); | |
c50ac050 | 5109 | return ret; |
a43a8c39 KC |
5110 | } |
5111 | ||
b5cec28d MK |
5112 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
5113 | long freed) | |
a43a8c39 | 5114 | { |
a5516438 | 5115 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 5116 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 5117 | long chg = 0; |
90481622 | 5118 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 5119 | long gbl_reserve; |
45c682a6 | 5120 | |
f27a5136 MK |
5121 | /* |
5122 | * Since this routine can be called in the evict inode path for all | |
5123 | * hugetlbfs inodes, resv_map could be NULL. | |
5124 | */ | |
b5cec28d MK |
5125 | if (resv_map) { |
5126 | chg = region_del(resv_map, start, end); | |
5127 | /* | |
5128 | * region_del() can fail in the rare case where a region | |
5129 | * must be split and another region descriptor can not be | |
5130 | * allocated. If end == LONG_MAX, it will not fail. | |
5131 | */ | |
5132 | if (chg < 0) | |
5133 | return chg; | |
5134 | } | |
5135 | ||
45c682a6 | 5136 | spin_lock(&inode->i_lock); |
e4c6f8be | 5137 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
5138 | spin_unlock(&inode->i_lock); |
5139 | ||
1c5ecae3 MK |
5140 | /* |
5141 | * If the subpool has a minimum size, the number of global | |
5142 | * reservations to be released may be adjusted. | |
5143 | */ | |
5144 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5145 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
5146 | |
5147 | return 0; | |
a43a8c39 | 5148 | } |
93f70f90 | 5149 | |
3212b535 SC |
5150 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
5151 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5152 | struct vm_area_struct *vma, | |
5153 | unsigned long addr, pgoff_t idx) | |
5154 | { | |
5155 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5156 | svma->vm_start; | |
5157 | unsigned long sbase = saddr & PUD_MASK; | |
5158 | unsigned long s_end = sbase + PUD_SIZE; | |
5159 | ||
5160 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
5161 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
5162 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
5163 | |
5164 | /* | |
5165 | * match the virtual addresses, permission and the alignment of the | |
5166 | * page table page. | |
5167 | */ | |
5168 | if (pmd_index(addr) != pmd_index(saddr) || | |
5169 | vm_flags != svm_flags || | |
5170 | sbase < svma->vm_start || svma->vm_end < s_end) | |
5171 | return 0; | |
5172 | ||
5173 | return saddr; | |
5174 | } | |
5175 | ||
31aafb45 | 5176 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
5177 | { |
5178 | unsigned long base = addr & PUD_MASK; | |
5179 | unsigned long end = base + PUD_SIZE; | |
5180 | ||
5181 | /* | |
5182 | * check on proper vm_flags and page table alignment | |
5183 | */ | |
017b1660 | 5184 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
5185 | return true; |
5186 | return false; | |
3212b535 SC |
5187 | } |
5188 | ||
017b1660 MK |
5189 | /* |
5190 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5191 | * If yes, adjust start and end to cover range associated with possible | |
5192 | * shared pmd mappings. | |
5193 | */ | |
5194 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5195 | unsigned long *start, unsigned long *end) | |
5196 | { | |
353b2de4 | 5197 | unsigned long check_addr; |
017b1660 MK |
5198 | |
5199 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
5200 | return; | |
5201 | ||
5202 | for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) { | |
5203 | unsigned long a_start = check_addr & PUD_MASK; | |
5204 | unsigned long a_end = a_start + PUD_SIZE; | |
5205 | ||
5206 | /* | |
5207 | * If sharing is possible, adjust start/end if necessary. | |
5208 | */ | |
5209 | if (range_in_vma(vma, a_start, a_end)) { | |
5210 | if (a_start < *start) | |
5211 | *start = a_start; | |
5212 | if (a_end > *end) | |
5213 | *end = a_end; | |
5214 | } | |
5215 | } | |
5216 | } | |
5217 | ||
3212b535 SC |
5218 | /* |
5219 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5220 | * and returns the corresponding pte. While this is not necessary for the | |
5221 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
5222 | * code much cleaner. |
5223 | * | |
5224 | * This routine must be called with i_mmap_rwsem held in at least read mode. | |
5225 | * For hugetlbfs, this prevents removal of any page table entries associated | |
5226 | * with the address space. This is important as we are setting up sharing | |
5227 | * based on existing page table entries (mappings). | |
3212b535 SC |
5228 | */ |
5229 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5230 | { | |
5231 | struct vm_area_struct *vma = find_vma(mm, addr); | |
5232 | struct address_space *mapping = vma->vm_file->f_mapping; | |
5233 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5234 | vma->vm_pgoff; | |
5235 | struct vm_area_struct *svma; | |
5236 | unsigned long saddr; | |
5237 | pte_t *spte = NULL; | |
5238 | pte_t *pte; | |
cb900f41 | 5239 | spinlock_t *ptl; |
3212b535 SC |
5240 | |
5241 | if (!vma_shareable(vma, addr)) | |
5242 | return (pte_t *)pmd_alloc(mm, pud, addr); | |
5243 | ||
3212b535 SC |
5244 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
5245 | if (svma == vma) | |
5246 | continue; | |
5247 | ||
5248 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5249 | if (saddr) { | |
7868a208 PA |
5250 | spte = huge_pte_offset(svma->vm_mm, saddr, |
5251 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
5252 | if (spte) { |
5253 | get_page(virt_to_page(spte)); | |
5254 | break; | |
5255 | } | |
5256 | } | |
5257 | } | |
5258 | ||
5259 | if (!spte) | |
5260 | goto out; | |
5261 | ||
8bea8052 | 5262 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 5263 | if (pud_none(*pud)) { |
3212b535 SC |
5264 | pud_populate(mm, pud, |
5265 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5266 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5267 | } else { |
3212b535 | 5268 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5269 | } |
cb900f41 | 5270 | spin_unlock(ptl); |
3212b535 SC |
5271 | out: |
5272 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
5273 | return pte; |
5274 | } | |
5275 | ||
5276 | /* | |
5277 | * unmap huge page backed by shared pte. | |
5278 | * | |
5279 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5280 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5281 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5282 | * | |
c0d0381a | 5283 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
5284 | * |
5285 | * returns: 1 successfully unmapped a shared pte page | |
5286 | * 0 the underlying pte page is not shared, or it is the last user | |
5287 | */ | |
5288 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
5289 | { | |
5290 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
5291 | p4d_t *p4d = p4d_offset(pgd, *addr); |
5292 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 SC |
5293 | |
5294 | BUG_ON(page_count(virt_to_page(ptep)) == 0); | |
5295 | if (page_count(virt_to_page(ptep)) == 1) | |
5296 | return 0; | |
5297 | ||
5298 | pud_clear(pud); | |
5299 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 5300 | mm_dec_nr_pmds(mm); |
3212b535 SC |
5301 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
5302 | return 1; | |
5303 | } | |
9e5fc74c SC |
5304 | #define want_pmd_share() (1) |
5305 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ | |
5306 | pte_t *huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud) | |
5307 | { | |
5308 | return NULL; | |
5309 | } | |
e81f2d22 ZZ |
5310 | |
5311 | int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep) | |
5312 | { | |
5313 | return 0; | |
5314 | } | |
017b1660 MK |
5315 | |
5316 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5317 | unsigned long *start, unsigned long *end) | |
5318 | { | |
5319 | } | |
9e5fc74c | 5320 | #define want_pmd_share() (0) |
3212b535 SC |
5321 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5322 | ||
9e5fc74c SC |
5323 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
5324 | pte_t *huge_pte_alloc(struct mm_struct *mm, | |
5325 | unsigned long addr, unsigned long sz) | |
5326 | { | |
5327 | pgd_t *pgd; | |
c2febafc | 5328 | p4d_t *p4d; |
9e5fc74c SC |
5329 | pud_t *pud; |
5330 | pte_t *pte = NULL; | |
5331 | ||
5332 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5333 | p4d = p4d_alloc(mm, pgd, addr); |
5334 | if (!p4d) | |
5335 | return NULL; | |
c2febafc | 5336 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5337 | if (pud) { |
5338 | if (sz == PUD_SIZE) { | |
5339 | pte = (pte_t *)pud; | |
5340 | } else { | |
5341 | BUG_ON(sz != PMD_SIZE); | |
5342 | if (want_pmd_share() && pud_none(*pud)) | |
5343 | pte = huge_pmd_share(mm, addr, pud); | |
5344 | else | |
5345 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5346 | } | |
5347 | } | |
4e666314 | 5348 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5349 | |
5350 | return pte; | |
5351 | } | |
5352 | ||
9b19df29 PA |
5353 | /* |
5354 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5355 | * entry at address @addr | |
5356 | * | |
5357 | * Return: Pointer to page table or swap entry (PUD or PMD) for | |
5358 | * address @addr, or NULL if a p*d_none() entry is encountered and the | |
5359 | * size @sz doesn't match the hugepage size at this level of the page | |
5360 | * table. | |
5361 | */ | |
7868a208 PA |
5362 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5363 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5364 | { |
5365 | pgd_t *pgd; | |
c2febafc | 5366 | p4d_t *p4d; |
3c1d7e6c L |
5367 | pud_t *pud, pud_entry; |
5368 | pmd_t *pmd, pmd_entry; | |
9e5fc74c SC |
5369 | |
5370 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5371 | if (!pgd_present(*pgd)) |
5372 | return NULL; | |
5373 | p4d = p4d_offset(pgd, addr); | |
5374 | if (!p4d_present(*p4d)) | |
5375 | return NULL; | |
9b19df29 | 5376 | |
c2febafc | 5377 | pud = pud_offset(p4d, addr); |
3c1d7e6c L |
5378 | pud_entry = READ_ONCE(*pud); |
5379 | if (sz != PUD_SIZE && pud_none(pud_entry)) | |
c2febafc | 5380 | return NULL; |
9b19df29 | 5381 | /* hugepage or swap? */ |
3c1d7e6c | 5382 | if (pud_huge(pud_entry) || !pud_present(pud_entry)) |
c2febafc | 5383 | return (pte_t *)pud; |
9b19df29 | 5384 | |
c2febafc | 5385 | pmd = pmd_offset(pud, addr); |
3c1d7e6c L |
5386 | pmd_entry = READ_ONCE(*pmd); |
5387 | if (sz != PMD_SIZE && pmd_none(pmd_entry)) | |
9b19df29 PA |
5388 | return NULL; |
5389 | /* hugepage or swap? */ | |
3c1d7e6c | 5390 | if (pmd_huge(pmd_entry) || !pmd_present(pmd_entry)) |
9b19df29 PA |
5391 | return (pte_t *)pmd; |
5392 | ||
5393 | return NULL; | |
9e5fc74c SC |
5394 | } |
5395 | ||
61f77eda NH |
5396 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5397 | ||
5398 | /* | |
5399 | * These functions are overwritable if your architecture needs its own | |
5400 | * behavior. | |
5401 | */ | |
5402 | struct page * __weak | |
5403 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5404 | int write) | |
5405 | { | |
5406 | return ERR_PTR(-EINVAL); | |
5407 | } | |
5408 | ||
4dc71451 AK |
5409 | struct page * __weak |
5410 | follow_huge_pd(struct vm_area_struct *vma, | |
5411 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5412 | { | |
5413 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5414 | return NULL; | |
5415 | } | |
5416 | ||
61f77eda | 5417 | struct page * __weak |
9e5fc74c | 5418 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5419 | pmd_t *pmd, int flags) |
9e5fc74c | 5420 | { |
e66f17ff NH |
5421 | struct page *page = NULL; |
5422 | spinlock_t *ptl; | |
c9d398fa | 5423 | pte_t pte; |
3faa52c0 JH |
5424 | |
5425 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5426 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5427 | (FOLL_PIN | FOLL_GET))) | |
5428 | return NULL; | |
5429 | ||
e66f17ff NH |
5430 | retry: |
5431 | ptl = pmd_lockptr(mm, pmd); | |
5432 | spin_lock(ptl); | |
5433 | /* | |
5434 | * make sure that the address range covered by this pmd is not | |
5435 | * unmapped from other threads. | |
5436 | */ | |
5437 | if (!pmd_huge(*pmd)) | |
5438 | goto out; | |
c9d398fa NH |
5439 | pte = huge_ptep_get((pte_t *)pmd); |
5440 | if (pte_present(pte)) { | |
97534127 | 5441 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
5442 | /* |
5443 | * try_grab_page() should always succeed here, because: a) we | |
5444 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5445 | * huge pmd (head) page is present in the page tables. The ptl | |
5446 | * prevents the head page and tail pages from being rearranged | |
5447 | * in any way. So this page must be available at this point, | |
5448 | * unless the page refcount overflowed: | |
5449 | */ | |
5450 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5451 | page = NULL; | |
5452 | goto out; | |
5453 | } | |
e66f17ff | 5454 | } else { |
c9d398fa | 5455 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5456 | spin_unlock(ptl); |
5457 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5458 | goto retry; | |
5459 | } | |
5460 | /* | |
5461 | * hwpoisoned entry is treated as no_page_table in | |
5462 | * follow_page_mask(). | |
5463 | */ | |
5464 | } | |
5465 | out: | |
5466 | spin_unlock(ptl); | |
9e5fc74c SC |
5467 | return page; |
5468 | } | |
5469 | ||
61f77eda | 5470 | struct page * __weak |
9e5fc74c | 5471 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5472 | pud_t *pud, int flags) |
9e5fc74c | 5473 | { |
3faa52c0 | 5474 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5475 | return NULL; |
9e5fc74c | 5476 | |
e66f17ff | 5477 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5478 | } |
5479 | ||
faaa5b62 AK |
5480 | struct page * __weak |
5481 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5482 | { | |
3faa52c0 | 5483 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5484 | return NULL; |
5485 | ||
5486 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5487 | } | |
5488 | ||
31caf665 NH |
5489 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5490 | { | |
bcc54222 NH |
5491 | bool ret = true; |
5492 | ||
309381fe | 5493 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5494 | spin_lock(&hugetlb_lock); |
bcc54222 NH |
5495 | if (!page_huge_active(page) || !get_page_unless_zero(page)) { |
5496 | ret = false; | |
5497 | goto unlock; | |
5498 | } | |
5499 | clear_page_huge_active(page); | |
31caf665 | 5500 | list_move_tail(&page->lru, list); |
bcc54222 | 5501 | unlock: |
31caf665 | 5502 | spin_unlock(&hugetlb_lock); |
bcc54222 | 5503 | return ret; |
31caf665 NH |
5504 | } |
5505 | ||
5506 | void putback_active_hugepage(struct page *page) | |
5507 | { | |
309381fe | 5508 | VM_BUG_ON_PAGE(!PageHead(page), page); |
31caf665 | 5509 | spin_lock(&hugetlb_lock); |
bcc54222 | 5510 | set_page_huge_active(page); |
31caf665 NH |
5511 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
5512 | spin_unlock(&hugetlb_lock); | |
5513 | put_page(page); | |
5514 | } | |
ab5ac90a MH |
5515 | |
5516 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5517 | { | |
5518 | struct hstate *h = page_hstate(oldpage); | |
5519 | ||
5520 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5521 | set_page_owner_migrate_reason(newpage, reason); | |
5522 | ||
5523 | /* | |
5524 | * transfer temporary state of the new huge page. This is | |
5525 | * reverse to other transitions because the newpage is going to | |
5526 | * be final while the old one will be freed so it takes over | |
5527 | * the temporary status. | |
5528 | * | |
5529 | * Also note that we have to transfer the per-node surplus state | |
5530 | * here as well otherwise the global surplus count will not match | |
5531 | * the per-node's. | |
5532 | */ | |
5533 | if (PageHugeTemporary(newpage)) { | |
5534 | int old_nid = page_to_nid(oldpage); | |
5535 | int new_nid = page_to_nid(newpage); | |
5536 | ||
5537 | SetPageHugeTemporary(oldpage); | |
5538 | ClearPageHugeTemporary(newpage); | |
5539 | ||
5540 | spin_lock(&hugetlb_lock); | |
5541 | if (h->surplus_huge_pages_node[old_nid]) { | |
5542 | h->surplus_huge_pages_node[old_nid]--; | |
5543 | h->surplus_huge_pages_node[new_nid]++; | |
5544 | } | |
5545 | spin_unlock(&hugetlb_lock); | |
5546 | } | |
5547 | } | |
cf11e85f RG |
5548 | |
5549 | #ifdef CONFIG_CMA | |
5550 | static unsigned long hugetlb_cma_size __initdata; | |
5551 | static bool cma_reserve_called __initdata; | |
5552 | ||
5553 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
5554 | { | |
5555 | hugetlb_cma_size = memparse(p, &p); | |
5556 | return 0; | |
5557 | } | |
5558 | ||
5559 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
5560 | ||
5561 | void __init hugetlb_cma_reserve(int order) | |
5562 | { | |
5563 | unsigned long size, reserved, per_node; | |
5564 | int nid; | |
5565 | ||
5566 | cma_reserve_called = true; | |
5567 | ||
5568 | if (!hugetlb_cma_size) | |
5569 | return; | |
5570 | ||
5571 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
5572 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
5573 | (PAGE_SIZE << order) / SZ_1M); | |
5574 | return; | |
5575 | } | |
5576 | ||
5577 | /* | |
5578 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
5579 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
5580 | */ | |
5581 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
5582 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
5583 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
5584 | ||
5585 | reserved = 0; | |
5586 | for_each_node_state(nid, N_ONLINE) { | |
5587 | int res; | |
5588 | ||
5589 | size = min(per_node, hugetlb_cma_size - reserved); | |
5590 | size = round_up(size, PAGE_SIZE << order); | |
5591 | ||
5592 | res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, | |
5593 | 0, false, "hugetlb", | |
5594 | &hugetlb_cma[nid], nid); | |
5595 | if (res) { | |
5596 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
5597 | res, nid); | |
5598 | continue; | |
5599 | } | |
5600 | ||
5601 | reserved += size; | |
5602 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
5603 | size / SZ_1M, nid); | |
5604 | ||
5605 | if (reserved >= hugetlb_cma_size) | |
5606 | break; | |
5607 | } | |
5608 | } | |
5609 | ||
5610 | void __init hugetlb_cma_check(void) | |
5611 | { | |
5612 | if (!hugetlb_cma_size || cma_reserve_called) | |
5613 | return; | |
5614 | ||
5615 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
5616 | } | |
5617 | ||
5618 | #endif /* CONFIG_CMA */ |