]>
Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * Generic hugetlb support. | |
6d49e352 | 4 | * (C) Nadia Yvette Chambers, April 2004 |
1da177e4 | 5 | */ |
1da177e4 LT |
6 | #include <linux/list.h> |
7 | #include <linux/init.h> | |
1da177e4 | 8 | #include <linux/mm.h> |
e1759c21 | 9 | #include <linux/seq_file.h> |
1da177e4 LT |
10 | #include <linux/sysctl.h> |
11 | #include <linux/highmem.h> | |
cddb8a5c | 12 | #include <linux/mmu_notifier.h> |
1da177e4 | 13 | #include <linux/nodemask.h> |
63551ae0 | 14 | #include <linux/pagemap.h> |
5da7ca86 | 15 | #include <linux/mempolicy.h> |
3b32123d | 16 | #include <linux/compiler.h> |
aea47ff3 | 17 | #include <linux/cpuset.h> |
3935baa9 | 18 | #include <linux/mutex.h> |
97ad1087 | 19 | #include <linux/memblock.h> |
a3437870 | 20 | #include <linux/sysfs.h> |
5a0e3ad6 | 21 | #include <linux/slab.h> |
bbe88753 | 22 | #include <linux/sched/mm.h> |
63489f8e | 23 | #include <linux/mmdebug.h> |
174cd4b1 | 24 | #include <linux/sched/signal.h> |
0fe6e20b | 25 | #include <linux/rmap.h> |
c6247f72 | 26 | #include <linux/string_helpers.h> |
fd6a03ed NH |
27 | #include <linux/swap.h> |
28 | #include <linux/swapops.h> | |
8382d914 | 29 | #include <linux/jhash.h> |
98fa15f3 | 30 | #include <linux/numa.h> |
c77c0a8a | 31 | #include <linux/llist.h> |
cf11e85f | 32 | #include <linux/cma.h> |
d6606683 | 33 | |
63551ae0 | 34 | #include <asm/page.h> |
ca15ca40 | 35 | #include <asm/pgalloc.h> |
24669e58 | 36 | #include <asm/tlb.h> |
63551ae0 | 37 | |
24669e58 | 38 | #include <linux/io.h> |
63551ae0 | 39 | #include <linux/hugetlb.h> |
9dd540e2 | 40 | #include <linux/hugetlb_cgroup.h> |
9a305230 | 41 | #include <linux/node.h> |
1a1aad8a | 42 | #include <linux/userfaultfd_k.h> |
ab5ac90a | 43 | #include <linux/page_owner.h> |
7835e98b | 44 | #include "internal.h" |
1da177e4 | 45 | |
c3f38a38 | 46 | int hugetlb_max_hstate __read_mostly; |
e5ff2159 AK |
47 | unsigned int default_hstate_idx; |
48 | struct hstate hstates[HUGE_MAX_HSTATE]; | |
cf11e85f | 49 | |
dbda8fea | 50 | #ifdef CONFIG_CMA |
cf11e85f | 51 | static struct cma *hugetlb_cma[MAX_NUMNODES]; |
dbda8fea BS |
52 | #endif |
53 | static unsigned long hugetlb_cma_size __initdata; | |
cf11e85f | 54 | |
641844f5 NH |
55 | /* |
56 | * Minimum page order among possible hugepage sizes, set to a proper value | |
57 | * at boot time. | |
58 | */ | |
59 | static unsigned int minimum_order __read_mostly = UINT_MAX; | |
e5ff2159 | 60 | |
53ba51d2 JT |
61 | __initdata LIST_HEAD(huge_boot_pages); |
62 | ||
e5ff2159 AK |
63 | /* for command line parsing */ |
64 | static struct hstate * __initdata parsed_hstate; | |
65 | static unsigned long __initdata default_hstate_max_huge_pages; | |
9fee021d | 66 | static bool __initdata parsed_valid_hugepagesz = true; |
282f4214 | 67 | static bool __initdata parsed_default_hugepagesz; |
e5ff2159 | 68 | |
3935baa9 | 69 | /* |
31caf665 NH |
70 | * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages, |
71 | * free_huge_pages, and surplus_huge_pages. | |
3935baa9 | 72 | */ |
c3f38a38 | 73 | DEFINE_SPINLOCK(hugetlb_lock); |
0bd0f9fb | 74 | |
8382d914 DB |
75 | /* |
76 | * Serializes faults on the same logical page. This is used to | |
77 | * prevent spurious OOMs when the hugepage pool is fully utilized. | |
78 | */ | |
79 | static int num_fault_mutexes; | |
c672c7f2 | 80 | struct mutex *hugetlb_fault_mutex_table ____cacheline_aligned_in_smp; |
8382d914 | 81 | |
7ca02d0a MK |
82 | /* Forward declaration */ |
83 | static int hugetlb_acct_memory(struct hstate *h, long delta); | |
84 | ||
1d88433b | 85 | static inline bool subpool_is_free(struct hugepage_subpool *spool) |
90481622 | 86 | { |
1d88433b ML |
87 | if (spool->count) |
88 | return false; | |
89 | if (spool->max_hpages != -1) | |
90 | return spool->used_hpages == 0; | |
91 | if (spool->min_hpages != -1) | |
92 | return spool->rsv_hpages == spool->min_hpages; | |
93 | ||
94 | return true; | |
95 | } | |
90481622 | 96 | |
db71ef79 MK |
97 | static inline void unlock_or_release_subpool(struct hugepage_subpool *spool, |
98 | unsigned long irq_flags) | |
1d88433b | 99 | { |
db71ef79 | 100 | spin_unlock_irqrestore(&spool->lock, irq_flags); |
90481622 DG |
101 | |
102 | /* If no pages are used, and no other handles to the subpool | |
7c8de358 | 103 | * remain, give up any reservations based on minimum size and |
7ca02d0a | 104 | * free the subpool */ |
1d88433b | 105 | if (subpool_is_free(spool)) { |
7ca02d0a MK |
106 | if (spool->min_hpages != -1) |
107 | hugetlb_acct_memory(spool->hstate, | |
108 | -spool->min_hpages); | |
90481622 | 109 | kfree(spool); |
7ca02d0a | 110 | } |
90481622 DG |
111 | } |
112 | ||
7ca02d0a MK |
113 | struct hugepage_subpool *hugepage_new_subpool(struct hstate *h, long max_hpages, |
114 | long min_hpages) | |
90481622 DG |
115 | { |
116 | struct hugepage_subpool *spool; | |
117 | ||
c6a91820 | 118 | spool = kzalloc(sizeof(*spool), GFP_KERNEL); |
90481622 DG |
119 | if (!spool) |
120 | return NULL; | |
121 | ||
122 | spin_lock_init(&spool->lock); | |
123 | spool->count = 1; | |
7ca02d0a MK |
124 | spool->max_hpages = max_hpages; |
125 | spool->hstate = h; | |
126 | spool->min_hpages = min_hpages; | |
127 | ||
128 | if (min_hpages != -1 && hugetlb_acct_memory(h, min_hpages)) { | |
129 | kfree(spool); | |
130 | return NULL; | |
131 | } | |
132 | spool->rsv_hpages = min_hpages; | |
90481622 DG |
133 | |
134 | return spool; | |
135 | } | |
136 | ||
137 | void hugepage_put_subpool(struct hugepage_subpool *spool) | |
138 | { | |
db71ef79 MK |
139 | unsigned long flags; |
140 | ||
141 | spin_lock_irqsave(&spool->lock, flags); | |
90481622 DG |
142 | BUG_ON(!spool->count); |
143 | spool->count--; | |
db71ef79 | 144 | unlock_or_release_subpool(spool, flags); |
90481622 DG |
145 | } |
146 | ||
1c5ecae3 MK |
147 | /* |
148 | * Subpool accounting for allocating and reserving pages. | |
149 | * Return -ENOMEM if there are not enough resources to satisfy the | |
9e7ee400 | 150 | * request. Otherwise, return the number of pages by which the |
1c5ecae3 MK |
151 | * global pools must be adjusted (upward). The returned value may |
152 | * only be different than the passed value (delta) in the case where | |
7c8de358 | 153 | * a subpool minimum size must be maintained. |
1c5ecae3 MK |
154 | */ |
155 | static long hugepage_subpool_get_pages(struct hugepage_subpool *spool, | |
90481622 DG |
156 | long delta) |
157 | { | |
1c5ecae3 | 158 | long ret = delta; |
90481622 DG |
159 | |
160 | if (!spool) | |
1c5ecae3 | 161 | return ret; |
90481622 | 162 | |
db71ef79 | 163 | spin_lock_irq(&spool->lock); |
1c5ecae3 MK |
164 | |
165 | if (spool->max_hpages != -1) { /* maximum size accounting */ | |
166 | if ((spool->used_hpages + delta) <= spool->max_hpages) | |
167 | spool->used_hpages += delta; | |
168 | else { | |
169 | ret = -ENOMEM; | |
170 | goto unlock_ret; | |
171 | } | |
90481622 | 172 | } |
90481622 | 173 | |
09a95e29 MK |
174 | /* minimum size accounting */ |
175 | if (spool->min_hpages != -1 && spool->rsv_hpages) { | |
1c5ecae3 MK |
176 | if (delta > spool->rsv_hpages) { |
177 | /* | |
178 | * Asking for more reserves than those already taken on | |
179 | * behalf of subpool. Return difference. | |
180 | */ | |
181 | ret = delta - spool->rsv_hpages; | |
182 | spool->rsv_hpages = 0; | |
183 | } else { | |
184 | ret = 0; /* reserves already accounted for */ | |
185 | spool->rsv_hpages -= delta; | |
186 | } | |
187 | } | |
188 | ||
189 | unlock_ret: | |
db71ef79 | 190 | spin_unlock_irq(&spool->lock); |
90481622 DG |
191 | return ret; |
192 | } | |
193 | ||
1c5ecae3 MK |
194 | /* |
195 | * Subpool accounting for freeing and unreserving pages. | |
196 | * Return the number of global page reservations that must be dropped. | |
197 | * The return value may only be different than the passed value (delta) | |
198 | * in the case where a subpool minimum size must be maintained. | |
199 | */ | |
200 | static long hugepage_subpool_put_pages(struct hugepage_subpool *spool, | |
90481622 DG |
201 | long delta) |
202 | { | |
1c5ecae3 | 203 | long ret = delta; |
db71ef79 | 204 | unsigned long flags; |
1c5ecae3 | 205 | |
90481622 | 206 | if (!spool) |
1c5ecae3 | 207 | return delta; |
90481622 | 208 | |
db71ef79 | 209 | spin_lock_irqsave(&spool->lock, flags); |
1c5ecae3 MK |
210 | |
211 | if (spool->max_hpages != -1) /* maximum size accounting */ | |
212 | spool->used_hpages -= delta; | |
213 | ||
09a95e29 MK |
214 | /* minimum size accounting */ |
215 | if (spool->min_hpages != -1 && spool->used_hpages < spool->min_hpages) { | |
1c5ecae3 MK |
216 | if (spool->rsv_hpages + delta <= spool->min_hpages) |
217 | ret = 0; | |
218 | else | |
219 | ret = spool->rsv_hpages + delta - spool->min_hpages; | |
220 | ||
221 | spool->rsv_hpages += delta; | |
222 | if (spool->rsv_hpages > spool->min_hpages) | |
223 | spool->rsv_hpages = spool->min_hpages; | |
224 | } | |
225 | ||
226 | /* | |
227 | * If hugetlbfs_put_super couldn't free spool due to an outstanding | |
228 | * quota reference, free it now. | |
229 | */ | |
db71ef79 | 230 | unlock_or_release_subpool(spool, flags); |
1c5ecae3 MK |
231 | |
232 | return ret; | |
90481622 DG |
233 | } |
234 | ||
235 | static inline struct hugepage_subpool *subpool_inode(struct inode *inode) | |
236 | { | |
237 | return HUGETLBFS_SB(inode->i_sb)->spool; | |
238 | } | |
239 | ||
240 | static inline struct hugepage_subpool *subpool_vma(struct vm_area_struct *vma) | |
241 | { | |
496ad9aa | 242 | return subpool_inode(file_inode(vma->vm_file)); |
90481622 DG |
243 | } |
244 | ||
0db9d74e MA |
245 | /* Helper that removes a struct file_region from the resv_map cache and returns |
246 | * it for use. | |
247 | */ | |
248 | static struct file_region * | |
249 | get_file_region_entry_from_cache(struct resv_map *resv, long from, long to) | |
250 | { | |
251 | struct file_region *nrg = NULL; | |
252 | ||
253 | VM_BUG_ON(resv->region_cache_count <= 0); | |
254 | ||
255 | resv->region_cache_count--; | |
256 | nrg = list_first_entry(&resv->region_cache, struct file_region, link); | |
0db9d74e MA |
257 | list_del(&nrg->link); |
258 | ||
259 | nrg->from = from; | |
260 | nrg->to = to; | |
261 | ||
262 | return nrg; | |
263 | } | |
264 | ||
075a61d0 MA |
265 | static void copy_hugetlb_cgroup_uncharge_info(struct file_region *nrg, |
266 | struct file_region *rg) | |
267 | { | |
268 | #ifdef CONFIG_CGROUP_HUGETLB | |
269 | nrg->reservation_counter = rg->reservation_counter; | |
270 | nrg->css = rg->css; | |
271 | if (rg->css) | |
272 | css_get(rg->css); | |
273 | #endif | |
274 | } | |
275 | ||
276 | /* Helper that records hugetlb_cgroup uncharge info. */ | |
277 | static void record_hugetlb_cgroup_uncharge_info(struct hugetlb_cgroup *h_cg, | |
278 | struct hstate *h, | |
279 | struct resv_map *resv, | |
280 | struct file_region *nrg) | |
281 | { | |
282 | #ifdef CONFIG_CGROUP_HUGETLB | |
283 | if (h_cg) { | |
284 | nrg->reservation_counter = | |
285 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
286 | nrg->css = &h_cg->css; | |
d85aecf2 ML |
287 | /* |
288 | * The caller will hold exactly one h_cg->css reference for the | |
289 | * whole contiguous reservation region. But this area might be | |
290 | * scattered when there are already some file_regions reside in | |
291 | * it. As a result, many file_regions may share only one css | |
292 | * reference. In order to ensure that one file_region must hold | |
293 | * exactly one h_cg->css reference, we should do css_get for | |
294 | * each file_region and leave the reference held by caller | |
295 | * untouched. | |
296 | */ | |
297 | css_get(&h_cg->css); | |
075a61d0 MA |
298 | if (!resv->pages_per_hpage) |
299 | resv->pages_per_hpage = pages_per_huge_page(h); | |
300 | /* pages_per_hpage should be the same for all entries in | |
301 | * a resv_map. | |
302 | */ | |
303 | VM_BUG_ON(resv->pages_per_hpage != pages_per_huge_page(h)); | |
304 | } else { | |
305 | nrg->reservation_counter = NULL; | |
306 | nrg->css = NULL; | |
307 | } | |
308 | #endif | |
309 | } | |
310 | ||
d85aecf2 ML |
311 | static void put_uncharge_info(struct file_region *rg) |
312 | { | |
313 | #ifdef CONFIG_CGROUP_HUGETLB | |
314 | if (rg->css) | |
315 | css_put(rg->css); | |
316 | #endif | |
317 | } | |
318 | ||
a9b3f867 MA |
319 | static bool has_same_uncharge_info(struct file_region *rg, |
320 | struct file_region *org) | |
321 | { | |
322 | #ifdef CONFIG_CGROUP_HUGETLB | |
323 | return rg && org && | |
324 | rg->reservation_counter == org->reservation_counter && | |
325 | rg->css == org->css; | |
326 | ||
327 | #else | |
328 | return true; | |
329 | #endif | |
330 | } | |
331 | ||
332 | static void coalesce_file_region(struct resv_map *resv, struct file_region *rg) | |
333 | { | |
334 | struct file_region *nrg = NULL, *prg = NULL; | |
335 | ||
336 | prg = list_prev_entry(rg, link); | |
337 | if (&prg->link != &resv->regions && prg->to == rg->from && | |
338 | has_same_uncharge_info(prg, rg)) { | |
339 | prg->to = rg->to; | |
340 | ||
341 | list_del(&rg->link); | |
d85aecf2 | 342 | put_uncharge_info(rg); |
a9b3f867 MA |
343 | kfree(rg); |
344 | ||
7db5e7b6 | 345 | rg = prg; |
a9b3f867 MA |
346 | } |
347 | ||
348 | nrg = list_next_entry(rg, link); | |
349 | if (&nrg->link != &resv->regions && nrg->from == rg->to && | |
350 | has_same_uncharge_info(nrg, rg)) { | |
351 | nrg->from = rg->from; | |
352 | ||
353 | list_del(&rg->link); | |
d85aecf2 | 354 | put_uncharge_info(rg); |
a9b3f867 | 355 | kfree(rg); |
a9b3f867 MA |
356 | } |
357 | } | |
358 | ||
2103cf9c PX |
359 | static inline long |
360 | hugetlb_resv_map_add(struct resv_map *map, struct file_region *rg, long from, | |
361 | long to, struct hstate *h, struct hugetlb_cgroup *cg, | |
362 | long *regions_needed) | |
363 | { | |
364 | struct file_region *nrg; | |
365 | ||
366 | if (!regions_needed) { | |
367 | nrg = get_file_region_entry_from_cache(map, from, to); | |
368 | record_hugetlb_cgroup_uncharge_info(cg, h, map, nrg); | |
369 | list_add(&nrg->link, rg->link.prev); | |
370 | coalesce_file_region(map, nrg); | |
371 | } else | |
372 | *regions_needed += 1; | |
373 | ||
374 | return to - from; | |
375 | } | |
376 | ||
972a3da3 WY |
377 | /* |
378 | * Must be called with resv->lock held. | |
379 | * | |
380 | * Calling this with regions_needed != NULL will count the number of pages | |
381 | * to be added but will not modify the linked list. And regions_needed will | |
382 | * indicate the number of file_regions needed in the cache to carry out to add | |
383 | * the regions for this range. | |
d75c6af9 MA |
384 | */ |
385 | static long add_reservation_in_range(struct resv_map *resv, long f, long t, | |
075a61d0 | 386 | struct hugetlb_cgroup *h_cg, |
972a3da3 | 387 | struct hstate *h, long *regions_needed) |
d75c6af9 | 388 | { |
0db9d74e | 389 | long add = 0; |
d75c6af9 | 390 | struct list_head *head = &resv->regions; |
0db9d74e | 391 | long last_accounted_offset = f; |
2103cf9c | 392 | struct file_region *rg = NULL, *trg = NULL; |
d75c6af9 | 393 | |
0db9d74e MA |
394 | if (regions_needed) |
395 | *regions_needed = 0; | |
d75c6af9 | 396 | |
0db9d74e MA |
397 | /* In this loop, we essentially handle an entry for the range |
398 | * [last_accounted_offset, rg->from), at every iteration, with some | |
399 | * bounds checking. | |
400 | */ | |
401 | list_for_each_entry_safe(rg, trg, head, link) { | |
402 | /* Skip irrelevant regions that start before our range. */ | |
403 | if (rg->from < f) { | |
404 | /* If this region ends after the last accounted offset, | |
405 | * then we need to update last_accounted_offset. | |
406 | */ | |
407 | if (rg->to > last_accounted_offset) | |
408 | last_accounted_offset = rg->to; | |
409 | continue; | |
410 | } | |
d75c6af9 | 411 | |
0db9d74e MA |
412 | /* When we find a region that starts beyond our range, we've |
413 | * finished. | |
414 | */ | |
ca7e0457 | 415 | if (rg->from >= t) |
d75c6af9 MA |
416 | break; |
417 | ||
0db9d74e MA |
418 | /* Add an entry for last_accounted_offset -> rg->from, and |
419 | * update last_accounted_offset. | |
420 | */ | |
2103cf9c PX |
421 | if (rg->from > last_accounted_offset) |
422 | add += hugetlb_resv_map_add(resv, rg, | |
423 | last_accounted_offset, | |
424 | rg->from, h, h_cg, | |
425 | regions_needed); | |
0db9d74e MA |
426 | |
427 | last_accounted_offset = rg->to; | |
428 | } | |
429 | ||
430 | /* Handle the case where our range extends beyond | |
431 | * last_accounted_offset. | |
432 | */ | |
2103cf9c PX |
433 | if (last_accounted_offset < t) |
434 | add += hugetlb_resv_map_add(resv, rg, last_accounted_offset, | |
435 | t, h, h_cg, regions_needed); | |
0db9d74e MA |
436 | |
437 | VM_BUG_ON(add < 0); | |
438 | return add; | |
439 | } | |
440 | ||
441 | /* Must be called with resv->lock acquired. Will drop lock to allocate entries. | |
442 | */ | |
443 | static int allocate_file_region_entries(struct resv_map *resv, | |
444 | int regions_needed) | |
445 | __must_hold(&resv->lock) | |
446 | { | |
447 | struct list_head allocated_regions; | |
448 | int to_allocate = 0, i = 0; | |
449 | struct file_region *trg = NULL, *rg = NULL; | |
450 | ||
451 | VM_BUG_ON(regions_needed < 0); | |
452 | ||
453 | INIT_LIST_HEAD(&allocated_regions); | |
454 | ||
455 | /* | |
456 | * Check for sufficient descriptors in the cache to accommodate | |
457 | * the number of in progress add operations plus regions_needed. | |
458 | * | |
459 | * This is a while loop because when we drop the lock, some other call | |
460 | * to region_add or region_del may have consumed some region_entries, | |
461 | * so we keep looping here until we finally have enough entries for | |
462 | * (adds_in_progress + regions_needed). | |
463 | */ | |
464 | while (resv->region_cache_count < | |
465 | (resv->adds_in_progress + regions_needed)) { | |
466 | to_allocate = resv->adds_in_progress + regions_needed - | |
467 | resv->region_cache_count; | |
468 | ||
469 | /* At this point, we should have enough entries in the cache | |
470 | * for all the existings adds_in_progress. We should only be | |
471 | * needing to allocate for regions_needed. | |
d75c6af9 | 472 | */ |
0db9d74e MA |
473 | VM_BUG_ON(resv->region_cache_count < resv->adds_in_progress); |
474 | ||
475 | spin_unlock(&resv->lock); | |
476 | for (i = 0; i < to_allocate; i++) { | |
477 | trg = kmalloc(sizeof(*trg), GFP_KERNEL); | |
478 | if (!trg) | |
479 | goto out_of_memory; | |
480 | list_add(&trg->link, &allocated_regions); | |
d75c6af9 | 481 | } |
d75c6af9 | 482 | |
0db9d74e MA |
483 | spin_lock(&resv->lock); |
484 | ||
d3ec7b6e WY |
485 | list_splice(&allocated_regions, &resv->region_cache); |
486 | resv->region_cache_count += to_allocate; | |
d75c6af9 MA |
487 | } |
488 | ||
0db9d74e | 489 | return 0; |
d75c6af9 | 490 | |
0db9d74e MA |
491 | out_of_memory: |
492 | list_for_each_entry_safe(rg, trg, &allocated_regions, link) { | |
493 | list_del(&rg->link); | |
494 | kfree(rg); | |
495 | } | |
496 | return -ENOMEM; | |
d75c6af9 MA |
497 | } |
498 | ||
1dd308a7 MK |
499 | /* |
500 | * Add the huge page range represented by [f, t) to the reserve | |
0db9d74e MA |
501 | * map. Regions will be taken from the cache to fill in this range. |
502 | * Sufficient regions should exist in the cache due to the previous | |
503 | * call to region_chg with the same range, but in some cases the cache will not | |
504 | * have sufficient entries due to races with other code doing region_add or | |
505 | * region_del. The extra needed entries will be allocated. | |
cf3ad20b | 506 | * |
0db9d74e MA |
507 | * regions_needed is the out value provided by a previous call to region_chg. |
508 | * | |
509 | * Return the number of new huge pages added to the map. This number is greater | |
510 | * than or equal to zero. If file_region entries needed to be allocated for | |
7c8de358 | 511 | * this operation and we were not able to allocate, it returns -ENOMEM. |
0db9d74e MA |
512 | * region_add of regions of length 1 never allocate file_regions and cannot |
513 | * fail; region_chg will always allocate at least 1 entry and a region_add for | |
514 | * 1 page will only require at most 1 entry. | |
1dd308a7 | 515 | */ |
0db9d74e | 516 | static long region_add(struct resv_map *resv, long f, long t, |
075a61d0 MA |
517 | long in_regions_needed, struct hstate *h, |
518 | struct hugetlb_cgroup *h_cg) | |
96822904 | 519 | { |
0db9d74e | 520 | long add = 0, actual_regions_needed = 0; |
96822904 | 521 | |
7b24d861 | 522 | spin_lock(&resv->lock); |
0db9d74e MA |
523 | retry: |
524 | ||
525 | /* Count how many regions are actually needed to execute this add. */ | |
972a3da3 WY |
526 | add_reservation_in_range(resv, f, t, NULL, NULL, |
527 | &actual_regions_needed); | |
96822904 | 528 | |
5e911373 | 529 | /* |
0db9d74e MA |
530 | * Check for sufficient descriptors in the cache to accommodate |
531 | * this add operation. Note that actual_regions_needed may be greater | |
532 | * than in_regions_needed, as the resv_map may have been modified since | |
533 | * the region_chg call. In this case, we need to make sure that we | |
534 | * allocate extra entries, such that we have enough for all the | |
535 | * existing adds_in_progress, plus the excess needed for this | |
536 | * operation. | |
5e911373 | 537 | */ |
0db9d74e MA |
538 | if (actual_regions_needed > in_regions_needed && |
539 | resv->region_cache_count < | |
540 | resv->adds_in_progress + | |
541 | (actual_regions_needed - in_regions_needed)) { | |
542 | /* region_add operation of range 1 should never need to | |
543 | * allocate file_region entries. | |
544 | */ | |
545 | VM_BUG_ON(t - f <= 1); | |
5e911373 | 546 | |
0db9d74e MA |
547 | if (allocate_file_region_entries( |
548 | resv, actual_regions_needed - in_regions_needed)) { | |
549 | return -ENOMEM; | |
550 | } | |
5e911373 | 551 | |
0db9d74e | 552 | goto retry; |
5e911373 MK |
553 | } |
554 | ||
972a3da3 | 555 | add = add_reservation_in_range(resv, f, t, h_cg, h, NULL); |
0db9d74e MA |
556 | |
557 | resv->adds_in_progress -= in_regions_needed; | |
cf3ad20b | 558 | |
7b24d861 | 559 | spin_unlock(&resv->lock); |
cf3ad20b | 560 | return add; |
96822904 AW |
561 | } |
562 | ||
1dd308a7 MK |
563 | /* |
564 | * Examine the existing reserve map and determine how many | |
565 | * huge pages in the specified range [f, t) are NOT currently | |
566 | * represented. This routine is called before a subsequent | |
567 | * call to region_add that will actually modify the reserve | |
568 | * map to add the specified range [f, t). region_chg does | |
569 | * not change the number of huge pages represented by the | |
0db9d74e MA |
570 | * map. A number of new file_region structures is added to the cache as a |
571 | * placeholder, for the subsequent region_add call to use. At least 1 | |
572 | * file_region structure is added. | |
573 | * | |
574 | * out_regions_needed is the number of regions added to the | |
575 | * resv->adds_in_progress. This value needs to be provided to a follow up call | |
576 | * to region_add or region_abort for proper accounting. | |
5e911373 MK |
577 | * |
578 | * Returns the number of huge pages that need to be added to the existing | |
579 | * reservation map for the range [f, t). This number is greater or equal to | |
580 | * zero. -ENOMEM is returned if a new file_region structure or cache entry | |
581 | * is needed and can not be allocated. | |
1dd308a7 | 582 | */ |
0db9d74e MA |
583 | static long region_chg(struct resv_map *resv, long f, long t, |
584 | long *out_regions_needed) | |
96822904 | 585 | { |
96822904 AW |
586 | long chg = 0; |
587 | ||
7b24d861 | 588 | spin_lock(&resv->lock); |
5e911373 | 589 | |
972a3da3 | 590 | /* Count how many hugepages in this range are NOT represented. */ |
075a61d0 | 591 | chg = add_reservation_in_range(resv, f, t, NULL, NULL, |
972a3da3 | 592 | out_regions_needed); |
5e911373 | 593 | |
0db9d74e MA |
594 | if (*out_regions_needed == 0) |
595 | *out_regions_needed = 1; | |
5e911373 | 596 | |
0db9d74e MA |
597 | if (allocate_file_region_entries(resv, *out_regions_needed)) |
598 | return -ENOMEM; | |
5e911373 | 599 | |
0db9d74e | 600 | resv->adds_in_progress += *out_regions_needed; |
7b24d861 | 601 | |
7b24d861 | 602 | spin_unlock(&resv->lock); |
96822904 AW |
603 | return chg; |
604 | } | |
605 | ||
5e911373 MK |
606 | /* |
607 | * Abort the in progress add operation. The adds_in_progress field | |
608 | * of the resv_map keeps track of the operations in progress between | |
609 | * calls to region_chg and region_add. Operations are sometimes | |
610 | * aborted after the call to region_chg. In such cases, region_abort | |
0db9d74e MA |
611 | * is called to decrement the adds_in_progress counter. regions_needed |
612 | * is the value returned by the region_chg call, it is used to decrement | |
613 | * the adds_in_progress counter. | |
5e911373 MK |
614 | * |
615 | * NOTE: The range arguments [f, t) are not needed or used in this | |
616 | * routine. They are kept to make reading the calling code easier as | |
617 | * arguments will match the associated region_chg call. | |
618 | */ | |
0db9d74e MA |
619 | static void region_abort(struct resv_map *resv, long f, long t, |
620 | long regions_needed) | |
5e911373 MK |
621 | { |
622 | spin_lock(&resv->lock); | |
623 | VM_BUG_ON(!resv->region_cache_count); | |
0db9d74e | 624 | resv->adds_in_progress -= regions_needed; |
5e911373 MK |
625 | spin_unlock(&resv->lock); |
626 | } | |
627 | ||
1dd308a7 | 628 | /* |
feba16e2 MK |
629 | * Delete the specified range [f, t) from the reserve map. If the |
630 | * t parameter is LONG_MAX, this indicates that ALL regions after f | |
631 | * should be deleted. Locate the regions which intersect [f, t) | |
632 | * and either trim, delete or split the existing regions. | |
633 | * | |
634 | * Returns the number of huge pages deleted from the reserve map. | |
635 | * In the normal case, the return value is zero or more. In the | |
636 | * case where a region must be split, a new region descriptor must | |
637 | * be allocated. If the allocation fails, -ENOMEM will be returned. | |
638 | * NOTE: If the parameter t == LONG_MAX, then we will never split | |
639 | * a region and possibly return -ENOMEM. Callers specifying | |
640 | * t == LONG_MAX do not need to check for -ENOMEM error. | |
1dd308a7 | 641 | */ |
feba16e2 | 642 | static long region_del(struct resv_map *resv, long f, long t) |
96822904 | 643 | { |
1406ec9b | 644 | struct list_head *head = &resv->regions; |
96822904 | 645 | struct file_region *rg, *trg; |
feba16e2 MK |
646 | struct file_region *nrg = NULL; |
647 | long del = 0; | |
96822904 | 648 | |
feba16e2 | 649 | retry: |
7b24d861 | 650 | spin_lock(&resv->lock); |
feba16e2 | 651 | list_for_each_entry_safe(rg, trg, head, link) { |
dbe409e4 MK |
652 | /* |
653 | * Skip regions before the range to be deleted. file_region | |
654 | * ranges are normally of the form [from, to). However, there | |
655 | * may be a "placeholder" entry in the map which is of the form | |
656 | * (from, to) with from == to. Check for placeholder entries | |
657 | * at the beginning of the range to be deleted. | |
658 | */ | |
659 | if (rg->to <= f && (rg->to != rg->from || rg->to != f)) | |
feba16e2 | 660 | continue; |
dbe409e4 | 661 | |
feba16e2 | 662 | if (rg->from >= t) |
96822904 | 663 | break; |
96822904 | 664 | |
feba16e2 MK |
665 | if (f > rg->from && t < rg->to) { /* Must split region */ |
666 | /* | |
667 | * Check for an entry in the cache before dropping | |
668 | * lock and attempting allocation. | |
669 | */ | |
670 | if (!nrg && | |
671 | resv->region_cache_count > resv->adds_in_progress) { | |
672 | nrg = list_first_entry(&resv->region_cache, | |
673 | struct file_region, | |
674 | link); | |
675 | list_del(&nrg->link); | |
676 | resv->region_cache_count--; | |
677 | } | |
96822904 | 678 | |
feba16e2 MK |
679 | if (!nrg) { |
680 | spin_unlock(&resv->lock); | |
681 | nrg = kmalloc(sizeof(*nrg), GFP_KERNEL); | |
682 | if (!nrg) | |
683 | return -ENOMEM; | |
684 | goto retry; | |
685 | } | |
686 | ||
687 | del += t - f; | |
79aa925b | 688 | hugetlb_cgroup_uncharge_file_region( |
d85aecf2 | 689 | resv, rg, t - f, false); |
feba16e2 MK |
690 | |
691 | /* New entry for end of split region */ | |
692 | nrg->from = t; | |
693 | nrg->to = rg->to; | |
075a61d0 MA |
694 | |
695 | copy_hugetlb_cgroup_uncharge_info(nrg, rg); | |
696 | ||
feba16e2 MK |
697 | INIT_LIST_HEAD(&nrg->link); |
698 | ||
699 | /* Original entry is trimmed */ | |
700 | rg->to = f; | |
701 | ||
702 | list_add(&nrg->link, &rg->link); | |
703 | nrg = NULL; | |
96822904 | 704 | break; |
feba16e2 MK |
705 | } |
706 | ||
707 | if (f <= rg->from && t >= rg->to) { /* Remove entire region */ | |
708 | del += rg->to - rg->from; | |
075a61d0 | 709 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 710 | rg->to - rg->from, true); |
feba16e2 MK |
711 | list_del(&rg->link); |
712 | kfree(rg); | |
713 | continue; | |
714 | } | |
715 | ||
716 | if (f <= rg->from) { /* Trim beginning of region */ | |
075a61d0 | 717 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 718 | t - rg->from, false); |
075a61d0 | 719 | |
79aa925b MK |
720 | del += t - rg->from; |
721 | rg->from = t; | |
722 | } else { /* Trim end of region */ | |
075a61d0 | 723 | hugetlb_cgroup_uncharge_file_region(resv, rg, |
d85aecf2 | 724 | rg->to - f, false); |
79aa925b MK |
725 | |
726 | del += rg->to - f; | |
727 | rg->to = f; | |
feba16e2 | 728 | } |
96822904 | 729 | } |
7b24d861 | 730 | |
7b24d861 | 731 | spin_unlock(&resv->lock); |
feba16e2 MK |
732 | kfree(nrg); |
733 | return del; | |
96822904 AW |
734 | } |
735 | ||
b5cec28d MK |
736 | /* |
737 | * A rare out of memory error was encountered which prevented removal of | |
738 | * the reserve map region for a page. The huge page itself was free'ed | |
739 | * and removed from the page cache. This routine will adjust the subpool | |
740 | * usage count, and the global reserve count if needed. By incrementing | |
741 | * these counts, the reserve map entry which could not be deleted will | |
742 | * appear as a "reserved" entry instead of simply dangling with incorrect | |
743 | * counts. | |
744 | */ | |
72e2936c | 745 | void hugetlb_fix_reserve_counts(struct inode *inode) |
b5cec28d MK |
746 | { |
747 | struct hugepage_subpool *spool = subpool_inode(inode); | |
748 | long rsv_adjust; | |
da56388c | 749 | bool reserved = false; |
b5cec28d MK |
750 | |
751 | rsv_adjust = hugepage_subpool_get_pages(spool, 1); | |
da56388c | 752 | if (rsv_adjust > 0) { |
b5cec28d MK |
753 | struct hstate *h = hstate_inode(inode); |
754 | ||
da56388c ML |
755 | if (!hugetlb_acct_memory(h, 1)) |
756 | reserved = true; | |
757 | } else if (!rsv_adjust) { | |
758 | reserved = true; | |
b5cec28d | 759 | } |
da56388c ML |
760 | |
761 | if (!reserved) | |
762 | pr_warn("hugetlb: Huge Page Reserved count may go negative.\n"); | |
b5cec28d MK |
763 | } |
764 | ||
1dd308a7 MK |
765 | /* |
766 | * Count and return the number of huge pages in the reserve map | |
767 | * that intersect with the range [f, t). | |
768 | */ | |
1406ec9b | 769 | static long region_count(struct resv_map *resv, long f, long t) |
84afd99b | 770 | { |
1406ec9b | 771 | struct list_head *head = &resv->regions; |
84afd99b AW |
772 | struct file_region *rg; |
773 | long chg = 0; | |
774 | ||
7b24d861 | 775 | spin_lock(&resv->lock); |
84afd99b AW |
776 | /* Locate each segment we overlap with, and count that overlap. */ |
777 | list_for_each_entry(rg, head, link) { | |
f2135a4a WSH |
778 | long seg_from; |
779 | long seg_to; | |
84afd99b AW |
780 | |
781 | if (rg->to <= f) | |
782 | continue; | |
783 | if (rg->from >= t) | |
784 | break; | |
785 | ||
786 | seg_from = max(rg->from, f); | |
787 | seg_to = min(rg->to, t); | |
788 | ||
789 | chg += seg_to - seg_from; | |
790 | } | |
7b24d861 | 791 | spin_unlock(&resv->lock); |
84afd99b AW |
792 | |
793 | return chg; | |
794 | } | |
795 | ||
e7c4b0bf AW |
796 | /* |
797 | * Convert the address within this vma to the page offset within | |
798 | * the mapping, in pagecache page units; huge pages here. | |
799 | */ | |
a5516438 AK |
800 | static pgoff_t vma_hugecache_offset(struct hstate *h, |
801 | struct vm_area_struct *vma, unsigned long address) | |
e7c4b0bf | 802 | { |
a5516438 AK |
803 | return ((address - vma->vm_start) >> huge_page_shift(h)) + |
804 | (vma->vm_pgoff >> huge_page_order(h)); | |
e7c4b0bf AW |
805 | } |
806 | ||
0fe6e20b NH |
807 | pgoff_t linear_hugepage_index(struct vm_area_struct *vma, |
808 | unsigned long address) | |
809 | { | |
810 | return vma_hugecache_offset(hstate_vma(vma), vma, address); | |
811 | } | |
dee41079 | 812 | EXPORT_SYMBOL_GPL(linear_hugepage_index); |
0fe6e20b | 813 | |
08fba699 MG |
814 | /* |
815 | * Return the size of the pages allocated when backing a VMA. In the majority | |
816 | * cases this will be same size as used by the page table entries. | |
817 | */ | |
818 | unsigned long vma_kernel_pagesize(struct vm_area_struct *vma) | |
819 | { | |
05ea8860 DW |
820 | if (vma->vm_ops && vma->vm_ops->pagesize) |
821 | return vma->vm_ops->pagesize(vma); | |
822 | return PAGE_SIZE; | |
08fba699 | 823 | } |
f340ca0f | 824 | EXPORT_SYMBOL_GPL(vma_kernel_pagesize); |
08fba699 | 825 | |
3340289d MG |
826 | /* |
827 | * Return the page size being used by the MMU to back a VMA. In the majority | |
828 | * of cases, the page size used by the kernel matches the MMU size. On | |
09135cc5 DW |
829 | * architectures where it differs, an architecture-specific 'strong' |
830 | * version of this symbol is required. | |
3340289d | 831 | */ |
09135cc5 | 832 | __weak unsigned long vma_mmu_pagesize(struct vm_area_struct *vma) |
3340289d MG |
833 | { |
834 | return vma_kernel_pagesize(vma); | |
835 | } | |
3340289d | 836 | |
84afd99b AW |
837 | /* |
838 | * Flags for MAP_PRIVATE reservations. These are stored in the bottom | |
839 | * bits of the reservation map pointer, which are always clear due to | |
840 | * alignment. | |
841 | */ | |
842 | #define HPAGE_RESV_OWNER (1UL << 0) | |
843 | #define HPAGE_RESV_UNMAPPED (1UL << 1) | |
04f2cbe3 | 844 | #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED) |
84afd99b | 845 | |
a1e78772 MG |
846 | /* |
847 | * These helpers are used to track how many pages are reserved for | |
848 | * faults in a MAP_PRIVATE mapping. Only the process that called mmap() | |
849 | * is guaranteed to have their future faults succeed. | |
850 | * | |
851 | * With the exception of reset_vma_resv_huge_pages() which is called at fork(), | |
852 | * the reserve counters are updated with the hugetlb_lock held. It is safe | |
853 | * to reset the VMA at fork() time as it is not in use yet and there is no | |
854 | * chance of the global counters getting corrupted as a result of the values. | |
84afd99b AW |
855 | * |
856 | * The private mapping reservation is represented in a subtly different | |
857 | * manner to a shared mapping. A shared mapping has a region map associated | |
858 | * with the underlying file, this region map represents the backing file | |
859 | * pages which have ever had a reservation assigned which this persists even | |
860 | * after the page is instantiated. A private mapping has a region map | |
861 | * associated with the original mmap which is attached to all VMAs which | |
862 | * reference it, this region map represents those offsets which have consumed | |
863 | * reservation ie. where pages have been instantiated. | |
a1e78772 | 864 | */ |
e7c4b0bf AW |
865 | static unsigned long get_vma_private_data(struct vm_area_struct *vma) |
866 | { | |
867 | return (unsigned long)vma->vm_private_data; | |
868 | } | |
869 | ||
870 | static void set_vma_private_data(struct vm_area_struct *vma, | |
871 | unsigned long value) | |
872 | { | |
873 | vma->vm_private_data = (void *)value; | |
874 | } | |
875 | ||
e9fe92ae MA |
876 | static void |
877 | resv_map_set_hugetlb_cgroup_uncharge_info(struct resv_map *resv_map, | |
878 | struct hugetlb_cgroup *h_cg, | |
879 | struct hstate *h) | |
880 | { | |
881 | #ifdef CONFIG_CGROUP_HUGETLB | |
882 | if (!h_cg || !h) { | |
883 | resv_map->reservation_counter = NULL; | |
884 | resv_map->pages_per_hpage = 0; | |
885 | resv_map->css = NULL; | |
886 | } else { | |
887 | resv_map->reservation_counter = | |
888 | &h_cg->rsvd_hugepage[hstate_index(h)]; | |
889 | resv_map->pages_per_hpage = pages_per_huge_page(h); | |
890 | resv_map->css = &h_cg->css; | |
891 | } | |
892 | #endif | |
893 | } | |
894 | ||
9119a41e | 895 | struct resv_map *resv_map_alloc(void) |
84afd99b AW |
896 | { |
897 | struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL); | |
5e911373 MK |
898 | struct file_region *rg = kmalloc(sizeof(*rg), GFP_KERNEL); |
899 | ||
900 | if (!resv_map || !rg) { | |
901 | kfree(resv_map); | |
902 | kfree(rg); | |
84afd99b | 903 | return NULL; |
5e911373 | 904 | } |
84afd99b AW |
905 | |
906 | kref_init(&resv_map->refs); | |
7b24d861 | 907 | spin_lock_init(&resv_map->lock); |
84afd99b AW |
908 | INIT_LIST_HEAD(&resv_map->regions); |
909 | ||
5e911373 | 910 | resv_map->adds_in_progress = 0; |
e9fe92ae MA |
911 | /* |
912 | * Initialize these to 0. On shared mappings, 0's here indicate these | |
913 | * fields don't do cgroup accounting. On private mappings, these will be | |
914 | * re-initialized to the proper values, to indicate that hugetlb cgroup | |
915 | * reservations are to be un-charged from here. | |
916 | */ | |
917 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, NULL, NULL); | |
5e911373 MK |
918 | |
919 | INIT_LIST_HEAD(&resv_map->region_cache); | |
920 | list_add(&rg->link, &resv_map->region_cache); | |
921 | resv_map->region_cache_count = 1; | |
922 | ||
84afd99b AW |
923 | return resv_map; |
924 | } | |
925 | ||
9119a41e | 926 | void resv_map_release(struct kref *ref) |
84afd99b AW |
927 | { |
928 | struct resv_map *resv_map = container_of(ref, struct resv_map, refs); | |
5e911373 MK |
929 | struct list_head *head = &resv_map->region_cache; |
930 | struct file_region *rg, *trg; | |
84afd99b AW |
931 | |
932 | /* Clear out any active regions before we release the map. */ | |
feba16e2 | 933 | region_del(resv_map, 0, LONG_MAX); |
5e911373 MK |
934 | |
935 | /* ... and any entries left in the cache */ | |
936 | list_for_each_entry_safe(rg, trg, head, link) { | |
937 | list_del(&rg->link); | |
938 | kfree(rg); | |
939 | } | |
940 | ||
941 | VM_BUG_ON(resv_map->adds_in_progress); | |
942 | ||
84afd99b AW |
943 | kfree(resv_map); |
944 | } | |
945 | ||
4e35f483 JK |
946 | static inline struct resv_map *inode_resv_map(struct inode *inode) |
947 | { | |
f27a5136 MK |
948 | /* |
949 | * At inode evict time, i_mapping may not point to the original | |
950 | * address space within the inode. This original address space | |
951 | * contains the pointer to the resv_map. So, always use the | |
952 | * address space embedded within the inode. | |
953 | * The VERY common case is inode->mapping == &inode->i_data but, | |
954 | * this may not be true for device special inodes. | |
955 | */ | |
956 | return (struct resv_map *)(&inode->i_data)->private_data; | |
4e35f483 JK |
957 | } |
958 | ||
84afd99b | 959 | static struct resv_map *vma_resv_map(struct vm_area_struct *vma) |
a1e78772 | 960 | { |
81d1b09c | 961 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
4e35f483 JK |
962 | if (vma->vm_flags & VM_MAYSHARE) { |
963 | struct address_space *mapping = vma->vm_file->f_mapping; | |
964 | struct inode *inode = mapping->host; | |
965 | ||
966 | return inode_resv_map(inode); | |
967 | ||
968 | } else { | |
84afd99b AW |
969 | return (struct resv_map *)(get_vma_private_data(vma) & |
970 | ~HPAGE_RESV_MASK); | |
4e35f483 | 971 | } |
a1e78772 MG |
972 | } |
973 | ||
84afd99b | 974 | static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map) |
a1e78772 | 975 | { |
81d1b09c SL |
976 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
977 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
a1e78772 | 978 | |
84afd99b AW |
979 | set_vma_private_data(vma, (get_vma_private_data(vma) & |
980 | HPAGE_RESV_MASK) | (unsigned long)map); | |
04f2cbe3 MG |
981 | } |
982 | ||
983 | static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags) | |
984 | { | |
81d1b09c SL |
985 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
986 | VM_BUG_ON_VMA(vma->vm_flags & VM_MAYSHARE, vma); | |
e7c4b0bf AW |
987 | |
988 | set_vma_private_data(vma, get_vma_private_data(vma) | flags); | |
04f2cbe3 MG |
989 | } |
990 | ||
991 | static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag) | |
992 | { | |
81d1b09c | 993 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
e7c4b0bf AW |
994 | |
995 | return (get_vma_private_data(vma) & flag) != 0; | |
a1e78772 MG |
996 | } |
997 | ||
04f2cbe3 | 998 | /* Reset counters to 0 and clear all HPAGE_RESV_* flags */ |
a1e78772 MG |
999 | void reset_vma_resv_huge_pages(struct vm_area_struct *vma) |
1000 | { | |
81d1b09c | 1001 | VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma), vma); |
f83a275d | 1002 | if (!(vma->vm_flags & VM_MAYSHARE)) |
a1e78772 MG |
1003 | vma->vm_private_data = (void *)0; |
1004 | } | |
1005 | ||
1006 | /* Returns true if the VMA has associated reserve pages */ | |
559ec2f8 | 1007 | static bool vma_has_reserves(struct vm_area_struct *vma, long chg) |
a1e78772 | 1008 | { |
af0ed73e JK |
1009 | if (vma->vm_flags & VM_NORESERVE) { |
1010 | /* | |
1011 | * This address is already reserved by other process(chg == 0), | |
1012 | * so, we should decrement reserved count. Without decrementing, | |
1013 | * reserve count remains after releasing inode, because this | |
1014 | * allocated page will go into page cache and is regarded as | |
1015 | * coming from reserved pool in releasing step. Currently, we | |
1016 | * don't have any other solution to deal with this situation | |
1017 | * properly, so add work-around here. | |
1018 | */ | |
1019 | if (vma->vm_flags & VM_MAYSHARE && chg == 0) | |
559ec2f8 | 1020 | return true; |
af0ed73e | 1021 | else |
559ec2f8 | 1022 | return false; |
af0ed73e | 1023 | } |
a63884e9 JK |
1024 | |
1025 | /* Shared mappings always use reserves */ | |
1fb1b0e9 MK |
1026 | if (vma->vm_flags & VM_MAYSHARE) { |
1027 | /* | |
1028 | * We know VM_NORESERVE is not set. Therefore, there SHOULD | |
1029 | * be a region map for all pages. The only situation where | |
1030 | * there is no region map is if a hole was punched via | |
7c8de358 | 1031 | * fallocate. In this case, there really are no reserves to |
1fb1b0e9 MK |
1032 | * use. This situation is indicated if chg != 0. |
1033 | */ | |
1034 | if (chg) | |
1035 | return false; | |
1036 | else | |
1037 | return true; | |
1038 | } | |
a63884e9 JK |
1039 | |
1040 | /* | |
1041 | * Only the process that called mmap() has reserves for | |
1042 | * private mappings. | |
1043 | */ | |
67961f9d MK |
1044 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) { |
1045 | /* | |
1046 | * Like the shared case above, a hole punch or truncate | |
1047 | * could have been performed on the private mapping. | |
1048 | * Examine the value of chg to determine if reserves | |
1049 | * actually exist or were previously consumed. | |
1050 | * Very Subtle - The value of chg comes from a previous | |
1051 | * call to vma_needs_reserves(). The reserve map for | |
1052 | * private mappings has different (opposite) semantics | |
1053 | * than that of shared mappings. vma_needs_reserves() | |
1054 | * has already taken this difference in semantics into | |
1055 | * account. Therefore, the meaning of chg is the same | |
1056 | * as in the shared case above. Code could easily be | |
1057 | * combined, but keeping it separate draws attention to | |
1058 | * subtle differences. | |
1059 | */ | |
1060 | if (chg) | |
1061 | return false; | |
1062 | else | |
1063 | return true; | |
1064 | } | |
a63884e9 | 1065 | |
559ec2f8 | 1066 | return false; |
a1e78772 MG |
1067 | } |
1068 | ||
a5516438 | 1069 | static void enqueue_huge_page(struct hstate *h, struct page *page) |
1da177e4 LT |
1070 | { |
1071 | int nid = page_to_nid(page); | |
9487ca60 MK |
1072 | |
1073 | lockdep_assert_held(&hugetlb_lock); | |
0edaecfa | 1074 | list_move(&page->lru, &h->hugepage_freelists[nid]); |
a5516438 AK |
1075 | h->free_huge_pages++; |
1076 | h->free_huge_pages_node[nid]++; | |
6c037149 | 1077 | SetHPageFreed(page); |
1da177e4 LT |
1078 | } |
1079 | ||
94310cbc | 1080 | static struct page *dequeue_huge_page_node_exact(struct hstate *h, int nid) |
bf50bab2 NH |
1081 | { |
1082 | struct page *page; | |
bbe88753 JK |
1083 | bool nocma = !!(current->flags & PF_MEMALLOC_NOCMA); |
1084 | ||
9487ca60 | 1085 | lockdep_assert_held(&hugetlb_lock); |
bbe88753 JK |
1086 | list_for_each_entry(page, &h->hugepage_freelists[nid], lru) { |
1087 | if (nocma && is_migrate_cma_page(page)) | |
1088 | continue; | |
bf50bab2 | 1089 | |
6664bfc8 WY |
1090 | if (PageHWPoison(page)) |
1091 | continue; | |
1092 | ||
1093 | list_move(&page->lru, &h->hugepage_activelist); | |
1094 | set_page_refcounted(page); | |
6c037149 | 1095 | ClearHPageFreed(page); |
6664bfc8 WY |
1096 | h->free_huge_pages--; |
1097 | h->free_huge_pages_node[nid]--; | |
1098 | return page; | |
bbe88753 JK |
1099 | } |
1100 | ||
6664bfc8 | 1101 | return NULL; |
bf50bab2 NH |
1102 | } |
1103 | ||
3e59fcb0 MH |
1104 | static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask, int nid, |
1105 | nodemask_t *nmask) | |
94310cbc | 1106 | { |
3e59fcb0 MH |
1107 | unsigned int cpuset_mems_cookie; |
1108 | struct zonelist *zonelist; | |
1109 | struct zone *zone; | |
1110 | struct zoneref *z; | |
98fa15f3 | 1111 | int node = NUMA_NO_NODE; |
94310cbc | 1112 | |
3e59fcb0 MH |
1113 | zonelist = node_zonelist(nid, gfp_mask); |
1114 | ||
1115 | retry_cpuset: | |
1116 | cpuset_mems_cookie = read_mems_allowed_begin(); | |
1117 | for_each_zone_zonelist_nodemask(zone, z, zonelist, gfp_zone(gfp_mask), nmask) { | |
1118 | struct page *page; | |
1119 | ||
1120 | if (!cpuset_zone_allowed(zone, gfp_mask)) | |
1121 | continue; | |
1122 | /* | |
1123 | * no need to ask again on the same node. Pool is node rather than | |
1124 | * zone aware | |
1125 | */ | |
1126 | if (zone_to_nid(zone) == node) | |
1127 | continue; | |
1128 | node = zone_to_nid(zone); | |
94310cbc | 1129 | |
94310cbc AK |
1130 | page = dequeue_huge_page_node_exact(h, node); |
1131 | if (page) | |
1132 | return page; | |
1133 | } | |
3e59fcb0 MH |
1134 | if (unlikely(read_mems_allowed_retry(cpuset_mems_cookie))) |
1135 | goto retry_cpuset; | |
1136 | ||
94310cbc AK |
1137 | return NULL; |
1138 | } | |
1139 | ||
a5516438 AK |
1140 | static struct page *dequeue_huge_page_vma(struct hstate *h, |
1141 | struct vm_area_struct *vma, | |
af0ed73e JK |
1142 | unsigned long address, int avoid_reserve, |
1143 | long chg) | |
1da177e4 | 1144 | { |
3e59fcb0 | 1145 | struct page *page; |
480eccf9 | 1146 | struct mempolicy *mpol; |
04ec6264 | 1147 | gfp_t gfp_mask; |
3e59fcb0 | 1148 | nodemask_t *nodemask; |
04ec6264 | 1149 | int nid; |
1da177e4 | 1150 | |
a1e78772 MG |
1151 | /* |
1152 | * A child process with MAP_PRIVATE mappings created by their parent | |
1153 | * have no page reserves. This check ensures that reservations are | |
1154 | * not "stolen". The child may still get SIGKILLed | |
1155 | */ | |
af0ed73e | 1156 | if (!vma_has_reserves(vma, chg) && |
a5516438 | 1157 | h->free_huge_pages - h->resv_huge_pages == 0) |
c0ff7453 | 1158 | goto err; |
a1e78772 | 1159 | |
04f2cbe3 | 1160 | /* If reserves cannot be used, ensure enough pages are in the pool */ |
a5516438 | 1161 | if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0) |
6eab04a8 | 1162 | goto err; |
04f2cbe3 | 1163 | |
04ec6264 VB |
1164 | gfp_mask = htlb_alloc_mask(h); |
1165 | nid = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
3e59fcb0 MH |
1166 | page = dequeue_huge_page_nodemask(h, gfp_mask, nid, nodemask); |
1167 | if (page && !avoid_reserve && vma_has_reserves(vma, chg)) { | |
d6995da3 | 1168 | SetHPageRestoreReserve(page); |
3e59fcb0 | 1169 | h->resv_huge_pages--; |
1da177e4 | 1170 | } |
cc9a6c87 | 1171 | |
52cd3b07 | 1172 | mpol_cond_put(mpol); |
1da177e4 | 1173 | return page; |
cc9a6c87 MG |
1174 | |
1175 | err: | |
cc9a6c87 | 1176 | return NULL; |
1da177e4 LT |
1177 | } |
1178 | ||
1cac6f2c LC |
1179 | /* |
1180 | * common helper functions for hstate_next_node_to_{alloc|free}. | |
1181 | * We may have allocated or freed a huge page based on a different | |
1182 | * nodes_allowed previously, so h->next_node_to_{alloc|free} might | |
1183 | * be outside of *nodes_allowed. Ensure that we use an allowed | |
1184 | * node for alloc or free. | |
1185 | */ | |
1186 | static int next_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1187 | { | |
0edaf86c | 1188 | nid = next_node_in(nid, *nodes_allowed); |
1cac6f2c LC |
1189 | VM_BUG_ON(nid >= MAX_NUMNODES); |
1190 | ||
1191 | return nid; | |
1192 | } | |
1193 | ||
1194 | static int get_valid_node_allowed(int nid, nodemask_t *nodes_allowed) | |
1195 | { | |
1196 | if (!node_isset(nid, *nodes_allowed)) | |
1197 | nid = next_node_allowed(nid, nodes_allowed); | |
1198 | return nid; | |
1199 | } | |
1200 | ||
1201 | /* | |
1202 | * returns the previously saved node ["this node"] from which to | |
1203 | * allocate a persistent huge page for the pool and advance the | |
1204 | * next node from which to allocate, handling wrap at end of node | |
1205 | * mask. | |
1206 | */ | |
1207 | static int hstate_next_node_to_alloc(struct hstate *h, | |
1208 | nodemask_t *nodes_allowed) | |
1209 | { | |
1210 | int nid; | |
1211 | ||
1212 | VM_BUG_ON(!nodes_allowed); | |
1213 | ||
1214 | nid = get_valid_node_allowed(h->next_nid_to_alloc, nodes_allowed); | |
1215 | h->next_nid_to_alloc = next_node_allowed(nid, nodes_allowed); | |
1216 | ||
1217 | return nid; | |
1218 | } | |
1219 | ||
1220 | /* | |
10c6ec49 | 1221 | * helper for remove_pool_huge_page() - return the previously saved |
1cac6f2c LC |
1222 | * node ["this node"] from which to free a huge page. Advance the |
1223 | * next node id whether or not we find a free huge page to free so | |
1224 | * that the next attempt to free addresses the next node. | |
1225 | */ | |
1226 | static int hstate_next_node_to_free(struct hstate *h, nodemask_t *nodes_allowed) | |
1227 | { | |
1228 | int nid; | |
1229 | ||
1230 | VM_BUG_ON(!nodes_allowed); | |
1231 | ||
1232 | nid = get_valid_node_allowed(h->next_nid_to_free, nodes_allowed); | |
1233 | h->next_nid_to_free = next_node_allowed(nid, nodes_allowed); | |
1234 | ||
1235 | return nid; | |
1236 | } | |
1237 | ||
1238 | #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \ | |
1239 | for (nr_nodes = nodes_weight(*mask); \ | |
1240 | nr_nodes > 0 && \ | |
1241 | ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \ | |
1242 | nr_nodes--) | |
1243 | ||
1244 | #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \ | |
1245 | for (nr_nodes = nodes_weight(*mask); \ | |
1246 | nr_nodes > 0 && \ | |
1247 | ((node = hstate_next_node_to_free(hs, mask)) || 1); \ | |
1248 | nr_nodes--) | |
1249 | ||
e1073d1e | 1250 | #ifdef CONFIG_ARCH_HAS_GIGANTIC_PAGE |
944d9fec | 1251 | static void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1252 | unsigned int order) |
944d9fec LC |
1253 | { |
1254 | int i; | |
1255 | int nr_pages = 1 << order; | |
1256 | struct page *p = page + 1; | |
1257 | ||
c8cc708a | 1258 | atomic_set(compound_mapcount_ptr(page), 0); |
5291c09b | 1259 | atomic_set(compound_pincount_ptr(page), 0); |
47e29d32 | 1260 | |
944d9fec | 1261 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
1d798ca3 | 1262 | clear_compound_head(p); |
944d9fec | 1263 | set_page_refcounted(p); |
944d9fec LC |
1264 | } |
1265 | ||
1266 | set_compound_order(page, 0); | |
ba9c1201 | 1267 | page[1].compound_nr = 0; |
944d9fec LC |
1268 | __ClearPageHead(page); |
1269 | } | |
1270 | ||
d00181b9 | 1271 | static void free_gigantic_page(struct page *page, unsigned int order) |
944d9fec | 1272 | { |
cf11e85f RG |
1273 | /* |
1274 | * If the page isn't allocated using the cma allocator, | |
1275 | * cma_release() returns false. | |
1276 | */ | |
dbda8fea BS |
1277 | #ifdef CONFIG_CMA |
1278 | if (cma_release(hugetlb_cma[page_to_nid(page)], page, 1 << order)) | |
cf11e85f | 1279 | return; |
dbda8fea | 1280 | #endif |
cf11e85f | 1281 | |
944d9fec LC |
1282 | free_contig_range(page_to_pfn(page), 1 << order); |
1283 | } | |
1284 | ||
4eb0716e | 1285 | #ifdef CONFIG_CONTIG_ALLOC |
d9cc948f MH |
1286 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
1287 | int nid, nodemask_t *nodemask) | |
944d9fec | 1288 | { |
04adbc3f | 1289 | unsigned long nr_pages = pages_per_huge_page(h); |
953f064a LX |
1290 | if (nid == NUMA_NO_NODE) |
1291 | nid = numa_mem_id(); | |
944d9fec | 1292 | |
dbda8fea BS |
1293 | #ifdef CONFIG_CMA |
1294 | { | |
cf11e85f RG |
1295 | struct page *page; |
1296 | int node; | |
1297 | ||
953f064a LX |
1298 | if (hugetlb_cma[nid]) { |
1299 | page = cma_alloc(hugetlb_cma[nid], nr_pages, | |
1300 | huge_page_order(h), true); | |
cf11e85f RG |
1301 | if (page) |
1302 | return page; | |
1303 | } | |
953f064a LX |
1304 | |
1305 | if (!(gfp_mask & __GFP_THISNODE)) { | |
1306 | for_each_node_mask(node, *nodemask) { | |
1307 | if (node == nid || !hugetlb_cma[node]) | |
1308 | continue; | |
1309 | ||
1310 | page = cma_alloc(hugetlb_cma[node], nr_pages, | |
1311 | huge_page_order(h), true); | |
1312 | if (page) | |
1313 | return page; | |
1314 | } | |
1315 | } | |
cf11e85f | 1316 | } |
dbda8fea | 1317 | #endif |
cf11e85f | 1318 | |
5e27a2df | 1319 | return alloc_contig_pages(nr_pages, gfp_mask, nid, nodemask); |
944d9fec LC |
1320 | } |
1321 | ||
1322 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid); | |
d00181b9 | 1323 | static void prep_compound_gigantic_page(struct page *page, unsigned int order); |
4eb0716e AG |
1324 | #else /* !CONFIG_CONTIG_ALLOC */ |
1325 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, | |
1326 | int nid, nodemask_t *nodemask) | |
1327 | { | |
1328 | return NULL; | |
1329 | } | |
1330 | #endif /* CONFIG_CONTIG_ALLOC */ | |
944d9fec | 1331 | |
e1073d1e | 1332 | #else /* !CONFIG_ARCH_HAS_GIGANTIC_PAGE */ |
d9cc948f | 1333 | static struct page *alloc_gigantic_page(struct hstate *h, gfp_t gfp_mask, |
4eb0716e AG |
1334 | int nid, nodemask_t *nodemask) |
1335 | { | |
1336 | return NULL; | |
1337 | } | |
d00181b9 | 1338 | static inline void free_gigantic_page(struct page *page, unsigned int order) { } |
944d9fec | 1339 | static inline void destroy_compound_gigantic_page(struct page *page, |
d00181b9 | 1340 | unsigned int order) { } |
944d9fec LC |
1341 | #endif |
1342 | ||
6eb4e88a MK |
1343 | /* |
1344 | * Remove hugetlb page from lists, and update dtor so that page appears | |
1345 | * as just a compound page. A reference is held on the page. | |
1346 | * | |
1347 | * Must be called with hugetlb lock held. | |
1348 | */ | |
1349 | static void remove_hugetlb_page(struct hstate *h, struct page *page, | |
1350 | bool adjust_surplus) | |
1351 | { | |
1352 | int nid = page_to_nid(page); | |
1353 | ||
1354 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page); | |
1355 | VM_BUG_ON_PAGE(hugetlb_cgroup_from_page_rsvd(page), page); | |
1356 | ||
9487ca60 | 1357 | lockdep_assert_held(&hugetlb_lock); |
6eb4e88a MK |
1358 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
1359 | return; | |
1360 | ||
1361 | list_del(&page->lru); | |
1362 | ||
1363 | if (HPageFreed(page)) { | |
1364 | h->free_huge_pages--; | |
1365 | h->free_huge_pages_node[nid]--; | |
1366 | } | |
1367 | if (adjust_surplus) { | |
1368 | h->surplus_huge_pages--; | |
1369 | h->surplus_huge_pages_node[nid]--; | |
1370 | } | |
1371 | ||
1372 | set_page_refcounted(page); | |
1373 | set_compound_page_dtor(page, NULL_COMPOUND_DTOR); | |
1374 | ||
1375 | h->nr_huge_pages--; | |
1376 | h->nr_huge_pages_node[nid]--; | |
1377 | } | |
1378 | ||
a5516438 | 1379 | static void update_and_free_page(struct hstate *h, struct page *page) |
6af2acb6 AL |
1380 | { |
1381 | int i; | |
dbfee5ae | 1382 | struct page *subpage = page; |
a5516438 | 1383 | |
4eb0716e | 1384 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
944d9fec | 1385 | return; |
18229df5 | 1386 | |
dbfee5ae MK |
1387 | for (i = 0; i < pages_per_huge_page(h); |
1388 | i++, subpage = mem_map_next(subpage, page, i)) { | |
1389 | subpage->flags &= ~(1 << PG_locked | 1 << PG_error | | |
32f84528 | 1390 | 1 << PG_referenced | 1 << PG_dirty | |
a7407a27 LC |
1391 | 1 << PG_active | 1 << PG_private | |
1392 | 1 << PG_writeback); | |
6af2acb6 | 1393 | } |
944d9fec LC |
1394 | if (hstate_is_gigantic(h)) { |
1395 | destroy_compound_gigantic_page(page, huge_page_order(h)); | |
1396 | free_gigantic_page(page, huge_page_order(h)); | |
1397 | } else { | |
944d9fec LC |
1398 | __free_pages(page, huge_page_order(h)); |
1399 | } | |
6af2acb6 AL |
1400 | } |
1401 | ||
10c6ec49 MK |
1402 | static void update_and_free_pages_bulk(struct hstate *h, struct list_head *list) |
1403 | { | |
1404 | struct page *page, *t_page; | |
1405 | ||
1406 | list_for_each_entry_safe(page, t_page, list, lru) { | |
1407 | update_and_free_page(h, page); | |
1408 | cond_resched(); | |
1409 | } | |
1410 | } | |
1411 | ||
e5ff2159 AK |
1412 | struct hstate *size_to_hstate(unsigned long size) |
1413 | { | |
1414 | struct hstate *h; | |
1415 | ||
1416 | for_each_hstate(h) { | |
1417 | if (huge_page_size(h) == size) | |
1418 | return h; | |
1419 | } | |
1420 | return NULL; | |
1421 | } | |
1422 | ||
db71ef79 | 1423 | void free_huge_page(struct page *page) |
27a85ef1 | 1424 | { |
a5516438 AK |
1425 | /* |
1426 | * Can't pass hstate in here because it is called from the | |
1427 | * compound page destructor. | |
1428 | */ | |
e5ff2159 | 1429 | struct hstate *h = page_hstate(page); |
7893d1d5 | 1430 | int nid = page_to_nid(page); |
d6995da3 | 1431 | struct hugepage_subpool *spool = hugetlb_page_subpool(page); |
07443a85 | 1432 | bool restore_reserve; |
db71ef79 | 1433 | unsigned long flags; |
27a85ef1 | 1434 | |
b4330afb MK |
1435 | VM_BUG_ON_PAGE(page_count(page), page); |
1436 | VM_BUG_ON_PAGE(page_mapcount(page), page); | |
8ace22bc | 1437 | |
d6995da3 | 1438 | hugetlb_set_page_subpool(page, NULL); |
8ace22bc | 1439 | page->mapping = NULL; |
d6995da3 MK |
1440 | restore_reserve = HPageRestoreReserve(page); |
1441 | ClearHPageRestoreReserve(page); | |
27a85ef1 | 1442 | |
1c5ecae3 | 1443 | /* |
d6995da3 | 1444 | * If HPageRestoreReserve was set on page, page allocation consumed a |
0919e1b6 MK |
1445 | * reservation. If the page was associated with a subpool, there |
1446 | * would have been a page reserved in the subpool before allocation | |
1447 | * via hugepage_subpool_get_pages(). Since we are 'restoring' the | |
6c26d310 | 1448 | * reservation, do not call hugepage_subpool_put_pages() as this will |
0919e1b6 | 1449 | * remove the reserved page from the subpool. |
1c5ecae3 | 1450 | */ |
0919e1b6 MK |
1451 | if (!restore_reserve) { |
1452 | /* | |
1453 | * A return code of zero implies that the subpool will be | |
1454 | * under its minimum size if the reservation is not restored | |
1455 | * after page is free. Therefore, force restore_reserve | |
1456 | * operation. | |
1457 | */ | |
1458 | if (hugepage_subpool_put_pages(spool, 1) == 0) | |
1459 | restore_reserve = true; | |
1460 | } | |
1c5ecae3 | 1461 | |
db71ef79 | 1462 | spin_lock_irqsave(&hugetlb_lock, flags); |
8f251a3d | 1463 | ClearHPageMigratable(page); |
6d76dcf4 AK |
1464 | hugetlb_cgroup_uncharge_page(hstate_index(h), |
1465 | pages_per_huge_page(h), page); | |
08cf9faf MA |
1466 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), |
1467 | pages_per_huge_page(h), page); | |
07443a85 JK |
1468 | if (restore_reserve) |
1469 | h->resv_huge_pages++; | |
1470 | ||
9157c311 | 1471 | if (HPageTemporary(page)) { |
6eb4e88a | 1472 | remove_hugetlb_page(h, page, false); |
db71ef79 | 1473 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
ab5ac90a MH |
1474 | update_and_free_page(h, page); |
1475 | } else if (h->surplus_huge_pages_node[nid]) { | |
0edaecfa | 1476 | /* remove the page from active list */ |
6eb4e88a | 1477 | remove_hugetlb_page(h, page, true); |
db71ef79 | 1478 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
a5516438 | 1479 | update_and_free_page(h, page); |
7893d1d5 | 1480 | } else { |
5d3a551c | 1481 | arch_clear_hugepage_flags(page); |
a5516438 | 1482 | enqueue_huge_page(h, page); |
db71ef79 | 1483 | spin_unlock_irqrestore(&hugetlb_lock, flags); |
c77c0a8a | 1484 | } |
c77c0a8a WL |
1485 | } |
1486 | ||
a5516438 | 1487 | static void prep_new_huge_page(struct hstate *h, struct page *page, int nid) |
b7ba30c6 | 1488 | { |
0edaecfa | 1489 | INIT_LIST_HEAD(&page->lru); |
f1e61557 | 1490 | set_compound_page_dtor(page, HUGETLB_PAGE_DTOR); |
ff546117 | 1491 | hugetlb_set_page_subpool(page, NULL); |
9dd540e2 | 1492 | set_hugetlb_cgroup(page, NULL); |
1adc4d41 | 1493 | set_hugetlb_cgroup_rsvd(page, NULL); |
db71ef79 | 1494 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
1495 | h->nr_huge_pages++; |
1496 | h->nr_huge_pages_node[nid]++; | |
db71ef79 | 1497 | spin_unlock_irq(&hugetlb_lock); |
b7ba30c6 AK |
1498 | } |
1499 | ||
d00181b9 | 1500 | static void prep_compound_gigantic_page(struct page *page, unsigned int order) |
20a0307c WF |
1501 | { |
1502 | int i; | |
1503 | int nr_pages = 1 << order; | |
1504 | struct page *p = page + 1; | |
1505 | ||
1506 | /* we rely on prep_new_huge_page to set the destructor */ | |
1507 | set_compound_order(page, order); | |
ef5a22be | 1508 | __ClearPageReserved(page); |
de09d31d | 1509 | __SetPageHead(page); |
20a0307c | 1510 | for (i = 1; i < nr_pages; i++, p = mem_map_next(p, page, i)) { |
ef5a22be AA |
1511 | /* |
1512 | * For gigantic hugepages allocated through bootmem at | |
1513 | * boot, it's safer to be consistent with the not-gigantic | |
1514 | * hugepages and clear the PG_reserved bit from all tail pages | |
7c8de358 | 1515 | * too. Otherwise drivers using get_user_pages() to access tail |
ef5a22be AA |
1516 | * pages may get the reference counting wrong if they see |
1517 | * PG_reserved set on a tail page (despite the head page not | |
1518 | * having PG_reserved set). Enforcing this consistency between | |
1519 | * head and tail pages allows drivers to optimize away a check | |
1520 | * on the head page when they need know if put_page() is needed | |
1521 | * after get_user_pages(). | |
1522 | */ | |
1523 | __ClearPageReserved(p); | |
58a84aa9 | 1524 | set_page_count(p, 0); |
1d798ca3 | 1525 | set_compound_head(p, page); |
20a0307c | 1526 | } |
b4330afb | 1527 | atomic_set(compound_mapcount_ptr(page), -1); |
5291c09b | 1528 | atomic_set(compound_pincount_ptr(page), 0); |
20a0307c WF |
1529 | } |
1530 | ||
7795912c AM |
1531 | /* |
1532 | * PageHuge() only returns true for hugetlbfs pages, but not for normal or | |
1533 | * transparent huge pages. See the PageTransHuge() documentation for more | |
1534 | * details. | |
1535 | */ | |
20a0307c WF |
1536 | int PageHuge(struct page *page) |
1537 | { | |
20a0307c WF |
1538 | if (!PageCompound(page)) |
1539 | return 0; | |
1540 | ||
1541 | page = compound_head(page); | |
f1e61557 | 1542 | return page[1].compound_dtor == HUGETLB_PAGE_DTOR; |
20a0307c | 1543 | } |
43131e14 NH |
1544 | EXPORT_SYMBOL_GPL(PageHuge); |
1545 | ||
27c73ae7 AA |
1546 | /* |
1547 | * PageHeadHuge() only returns true for hugetlbfs head page, but not for | |
1548 | * normal or transparent huge pages. | |
1549 | */ | |
1550 | int PageHeadHuge(struct page *page_head) | |
1551 | { | |
27c73ae7 AA |
1552 | if (!PageHead(page_head)) |
1553 | return 0; | |
1554 | ||
d4af73e3 | 1555 | return page_head[1].compound_dtor == HUGETLB_PAGE_DTOR; |
27c73ae7 | 1556 | } |
27c73ae7 | 1557 | |
c0d0381a MK |
1558 | /* |
1559 | * Find and lock address space (mapping) in write mode. | |
1560 | * | |
336bf30e MK |
1561 | * Upon entry, the page is locked which means that page_mapping() is |
1562 | * stable. Due to locking order, we can only trylock_write. If we can | |
1563 | * not get the lock, simply return NULL to caller. | |
c0d0381a MK |
1564 | */ |
1565 | struct address_space *hugetlb_page_mapping_lock_write(struct page *hpage) | |
1566 | { | |
336bf30e | 1567 | struct address_space *mapping = page_mapping(hpage); |
c0d0381a | 1568 | |
c0d0381a MK |
1569 | if (!mapping) |
1570 | return mapping; | |
1571 | ||
c0d0381a MK |
1572 | if (i_mmap_trylock_write(mapping)) |
1573 | return mapping; | |
1574 | ||
336bf30e | 1575 | return NULL; |
c0d0381a MK |
1576 | } |
1577 | ||
13d60f4b ZY |
1578 | pgoff_t __basepage_index(struct page *page) |
1579 | { | |
1580 | struct page *page_head = compound_head(page); | |
1581 | pgoff_t index = page_index(page_head); | |
1582 | unsigned long compound_idx; | |
1583 | ||
1584 | if (!PageHuge(page_head)) | |
1585 | return page_index(page); | |
1586 | ||
1587 | if (compound_order(page_head) >= MAX_ORDER) | |
1588 | compound_idx = page_to_pfn(page) - page_to_pfn(page_head); | |
1589 | else | |
1590 | compound_idx = page - page_head; | |
1591 | ||
1592 | return (index << compound_order(page_head)) + compound_idx; | |
1593 | } | |
1594 | ||
0c397dae | 1595 | static struct page *alloc_buddy_huge_page(struct hstate *h, |
f60858f9 MK |
1596 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1597 | nodemask_t *node_alloc_noretry) | |
1da177e4 | 1598 | { |
af0fb9df | 1599 | int order = huge_page_order(h); |
1da177e4 | 1600 | struct page *page; |
f60858f9 | 1601 | bool alloc_try_hard = true; |
f96efd58 | 1602 | |
f60858f9 MK |
1603 | /* |
1604 | * By default we always try hard to allocate the page with | |
1605 | * __GFP_RETRY_MAYFAIL flag. However, if we are allocating pages in | |
1606 | * a loop (to adjust global huge page counts) and previous allocation | |
1607 | * failed, do not continue to try hard on the same node. Use the | |
1608 | * node_alloc_noretry bitmap to manage this state information. | |
1609 | */ | |
1610 | if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry)) | |
1611 | alloc_try_hard = false; | |
1612 | gfp_mask |= __GFP_COMP|__GFP_NOWARN; | |
1613 | if (alloc_try_hard) | |
1614 | gfp_mask |= __GFP_RETRY_MAYFAIL; | |
af0fb9df MH |
1615 | if (nid == NUMA_NO_NODE) |
1616 | nid = numa_mem_id(); | |
84172f4b | 1617 | page = __alloc_pages(gfp_mask, order, nid, nmask); |
af0fb9df MH |
1618 | if (page) |
1619 | __count_vm_event(HTLB_BUDDY_PGALLOC); | |
1620 | else | |
1621 | __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL); | |
63b4613c | 1622 | |
f60858f9 MK |
1623 | /* |
1624 | * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this | |
1625 | * indicates an overall state change. Clear bit so that we resume | |
1626 | * normal 'try hard' allocations. | |
1627 | */ | |
1628 | if (node_alloc_noretry && page && !alloc_try_hard) | |
1629 | node_clear(nid, *node_alloc_noretry); | |
1630 | ||
1631 | /* | |
1632 | * If we tried hard to get a page but failed, set bit so that | |
1633 | * subsequent attempts will not try as hard until there is an | |
1634 | * overall state change. | |
1635 | */ | |
1636 | if (node_alloc_noretry && !page && alloc_try_hard) | |
1637 | node_set(nid, *node_alloc_noretry); | |
1638 | ||
63b4613c NA |
1639 | return page; |
1640 | } | |
1641 | ||
0c397dae MH |
1642 | /* |
1643 | * Common helper to allocate a fresh hugetlb page. All specific allocators | |
1644 | * should use this function to get new hugetlb pages | |
1645 | */ | |
1646 | static struct page *alloc_fresh_huge_page(struct hstate *h, | |
f60858f9 MK |
1647 | gfp_t gfp_mask, int nid, nodemask_t *nmask, |
1648 | nodemask_t *node_alloc_noretry) | |
0c397dae MH |
1649 | { |
1650 | struct page *page; | |
1651 | ||
1652 | if (hstate_is_gigantic(h)) | |
1653 | page = alloc_gigantic_page(h, gfp_mask, nid, nmask); | |
1654 | else | |
1655 | page = alloc_buddy_huge_page(h, gfp_mask, | |
f60858f9 | 1656 | nid, nmask, node_alloc_noretry); |
0c397dae MH |
1657 | if (!page) |
1658 | return NULL; | |
1659 | ||
1660 | if (hstate_is_gigantic(h)) | |
1661 | prep_compound_gigantic_page(page, huge_page_order(h)); | |
1662 | prep_new_huge_page(h, page, page_to_nid(page)); | |
1663 | ||
1664 | return page; | |
1665 | } | |
1666 | ||
af0fb9df MH |
1667 | /* |
1668 | * Allocates a fresh page to the hugetlb allocator pool in the node interleaved | |
1669 | * manner. | |
1670 | */ | |
f60858f9 MK |
1671 | static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed, |
1672 | nodemask_t *node_alloc_noretry) | |
b2261026 JK |
1673 | { |
1674 | struct page *page; | |
1675 | int nr_nodes, node; | |
af0fb9df | 1676 | gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE; |
b2261026 JK |
1677 | |
1678 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
f60858f9 MK |
1679 | page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed, |
1680 | node_alloc_noretry); | |
af0fb9df | 1681 | if (page) |
b2261026 | 1682 | break; |
b2261026 JK |
1683 | } |
1684 | ||
af0fb9df MH |
1685 | if (!page) |
1686 | return 0; | |
b2261026 | 1687 | |
af0fb9df MH |
1688 | put_page(page); /* free it into the hugepage allocator */ |
1689 | ||
1690 | return 1; | |
b2261026 JK |
1691 | } |
1692 | ||
e8c5c824 | 1693 | /* |
10c6ec49 MK |
1694 | * Remove huge page from pool from next node to free. Attempt to keep |
1695 | * persistent huge pages more or less balanced over allowed nodes. | |
1696 | * This routine only 'removes' the hugetlb page. The caller must make | |
1697 | * an additional call to free the page to low level allocators. | |
e8c5c824 LS |
1698 | * Called with hugetlb_lock locked. |
1699 | */ | |
10c6ec49 MK |
1700 | static struct page *remove_pool_huge_page(struct hstate *h, |
1701 | nodemask_t *nodes_allowed, | |
1702 | bool acct_surplus) | |
e8c5c824 | 1703 | { |
b2261026 | 1704 | int nr_nodes, node; |
10c6ec49 | 1705 | struct page *page = NULL; |
e8c5c824 | 1706 | |
9487ca60 | 1707 | lockdep_assert_held(&hugetlb_lock); |
b2261026 | 1708 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { |
685f3457 LS |
1709 | /* |
1710 | * If we're returning unused surplus pages, only examine | |
1711 | * nodes with surplus pages. | |
1712 | */ | |
b2261026 JK |
1713 | if ((!acct_surplus || h->surplus_huge_pages_node[node]) && |
1714 | !list_empty(&h->hugepage_freelists[node])) { | |
10c6ec49 | 1715 | page = list_entry(h->hugepage_freelists[node].next, |
e8c5c824 | 1716 | struct page, lru); |
6eb4e88a | 1717 | remove_hugetlb_page(h, page, acct_surplus); |
9a76db09 | 1718 | break; |
e8c5c824 | 1719 | } |
b2261026 | 1720 | } |
e8c5c824 | 1721 | |
10c6ec49 | 1722 | return page; |
e8c5c824 LS |
1723 | } |
1724 | ||
c8721bbb NH |
1725 | /* |
1726 | * Dissolve a given free hugepage into free buddy pages. This function does | |
faf53def NH |
1727 | * nothing for in-use hugepages and non-hugepages. |
1728 | * This function returns values like below: | |
1729 | * | |
1730 | * -EBUSY: failed to dissolved free hugepages or the hugepage is in-use | |
1731 | * (allocated or reserved.) | |
1732 | * 0: successfully dissolved free hugepages or the page is not a | |
1733 | * hugepage (considered as already dissolved) | |
c8721bbb | 1734 | */ |
c3114a84 | 1735 | int dissolve_free_huge_page(struct page *page) |
c8721bbb | 1736 | { |
6bc9b564 | 1737 | int rc = -EBUSY; |
082d5b6b | 1738 | |
7ffddd49 | 1739 | retry: |
faf53def NH |
1740 | /* Not to disrupt normal path by vainly holding hugetlb_lock */ |
1741 | if (!PageHuge(page)) | |
1742 | return 0; | |
1743 | ||
db71ef79 | 1744 | spin_lock_irq(&hugetlb_lock); |
faf53def NH |
1745 | if (!PageHuge(page)) { |
1746 | rc = 0; | |
1747 | goto out; | |
1748 | } | |
1749 | ||
1750 | if (!page_count(page)) { | |
2247bb33 GS |
1751 | struct page *head = compound_head(page); |
1752 | struct hstate *h = page_hstate(head); | |
6bc9b564 | 1753 | if (h->free_huge_pages - h->resv_huge_pages == 0) |
082d5b6b | 1754 | goto out; |
7ffddd49 MS |
1755 | |
1756 | /* | |
1757 | * We should make sure that the page is already on the free list | |
1758 | * when it is dissolved. | |
1759 | */ | |
6c037149 | 1760 | if (unlikely(!HPageFreed(head))) { |
db71ef79 | 1761 | spin_unlock_irq(&hugetlb_lock); |
7ffddd49 MS |
1762 | cond_resched(); |
1763 | ||
1764 | /* | |
1765 | * Theoretically, we should return -EBUSY when we | |
1766 | * encounter this race. In fact, we have a chance | |
1767 | * to successfully dissolve the page if we do a | |
1768 | * retry. Because the race window is quite small. | |
1769 | * If we seize this opportunity, it is an optimization | |
1770 | * for increasing the success rate of dissolving page. | |
1771 | */ | |
1772 | goto retry; | |
1773 | } | |
1774 | ||
c3114a84 AK |
1775 | /* |
1776 | * Move PageHWPoison flag from head page to the raw error page, | |
1777 | * which makes any subpages rather than the error page reusable. | |
1778 | */ | |
1779 | if (PageHWPoison(head) && page != head) { | |
1780 | SetPageHWPoison(page); | |
1781 | ClearPageHWPoison(head); | |
1782 | } | |
6eb4e88a | 1783 | remove_hugetlb_page(h, page, false); |
c1470b33 | 1784 | h->max_huge_pages--; |
db71ef79 | 1785 | spin_unlock_irq(&hugetlb_lock); |
2247bb33 | 1786 | update_and_free_page(h, head); |
1121828a | 1787 | return 0; |
c8721bbb | 1788 | } |
082d5b6b | 1789 | out: |
db71ef79 | 1790 | spin_unlock_irq(&hugetlb_lock); |
082d5b6b | 1791 | return rc; |
c8721bbb NH |
1792 | } |
1793 | ||
1794 | /* | |
1795 | * Dissolve free hugepages in a given pfn range. Used by memory hotplug to | |
1796 | * make specified memory blocks removable from the system. | |
2247bb33 GS |
1797 | * Note that this will dissolve a free gigantic hugepage completely, if any |
1798 | * part of it lies within the given range. | |
082d5b6b GS |
1799 | * Also note that if dissolve_free_huge_page() returns with an error, all |
1800 | * free hugepages that were dissolved before that error are lost. | |
c8721bbb | 1801 | */ |
082d5b6b | 1802 | int dissolve_free_huge_pages(unsigned long start_pfn, unsigned long end_pfn) |
c8721bbb | 1803 | { |
c8721bbb | 1804 | unsigned long pfn; |
eb03aa00 | 1805 | struct page *page; |
082d5b6b | 1806 | int rc = 0; |
c8721bbb | 1807 | |
d0177639 | 1808 | if (!hugepages_supported()) |
082d5b6b | 1809 | return rc; |
d0177639 | 1810 | |
eb03aa00 GS |
1811 | for (pfn = start_pfn; pfn < end_pfn; pfn += 1 << minimum_order) { |
1812 | page = pfn_to_page(pfn); | |
faf53def NH |
1813 | rc = dissolve_free_huge_page(page); |
1814 | if (rc) | |
1815 | break; | |
eb03aa00 | 1816 | } |
082d5b6b GS |
1817 | |
1818 | return rc; | |
c8721bbb NH |
1819 | } |
1820 | ||
ab5ac90a MH |
1821 | /* |
1822 | * Allocates a fresh surplus page from the page allocator. | |
1823 | */ | |
0c397dae | 1824 | static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask, |
aaf14e40 | 1825 | int nid, nodemask_t *nmask) |
7893d1d5 | 1826 | { |
9980d744 | 1827 | struct page *page = NULL; |
7893d1d5 | 1828 | |
bae7f4ae | 1829 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
1830 | return NULL; |
1831 | ||
db71ef79 | 1832 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
1833 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) |
1834 | goto out_unlock; | |
db71ef79 | 1835 | spin_unlock_irq(&hugetlb_lock); |
d1c3fb1f | 1836 | |
f60858f9 | 1837 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
9980d744 | 1838 | if (!page) |
0c397dae | 1839 | return NULL; |
d1c3fb1f | 1840 | |
db71ef79 | 1841 | spin_lock_irq(&hugetlb_lock); |
9980d744 MH |
1842 | /* |
1843 | * We could have raced with the pool size change. | |
1844 | * Double check that and simply deallocate the new page | |
1845 | * if we would end up overcommiting the surpluses. Abuse | |
1846 | * temporary page to workaround the nasty free_huge_page | |
1847 | * codeflow | |
1848 | */ | |
1849 | if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) { | |
9157c311 | 1850 | SetHPageTemporary(page); |
db71ef79 | 1851 | spin_unlock_irq(&hugetlb_lock); |
9980d744 | 1852 | put_page(page); |
2bf753e6 | 1853 | return NULL; |
9980d744 | 1854 | } else { |
9980d744 | 1855 | h->surplus_huge_pages++; |
4704dea3 | 1856 | h->surplus_huge_pages_node[page_to_nid(page)]++; |
7893d1d5 | 1857 | } |
9980d744 MH |
1858 | |
1859 | out_unlock: | |
db71ef79 | 1860 | spin_unlock_irq(&hugetlb_lock); |
7893d1d5 AL |
1861 | |
1862 | return page; | |
1863 | } | |
1864 | ||
bbe88753 | 1865 | static struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask, |
9a4e9f3b | 1866 | int nid, nodemask_t *nmask) |
ab5ac90a MH |
1867 | { |
1868 | struct page *page; | |
1869 | ||
1870 | if (hstate_is_gigantic(h)) | |
1871 | return NULL; | |
1872 | ||
f60858f9 | 1873 | page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL); |
ab5ac90a MH |
1874 | if (!page) |
1875 | return NULL; | |
1876 | ||
1877 | /* | |
1878 | * We do not account these pages as surplus because they are only | |
1879 | * temporary and will be released properly on the last reference | |
1880 | */ | |
9157c311 | 1881 | SetHPageTemporary(page); |
ab5ac90a MH |
1882 | |
1883 | return page; | |
1884 | } | |
1885 | ||
099730d6 DH |
1886 | /* |
1887 | * Use the VMA's mpolicy to allocate a huge page from the buddy. | |
1888 | */ | |
e0ec90ee | 1889 | static |
0c397dae | 1890 | struct page *alloc_buddy_huge_page_with_mpol(struct hstate *h, |
099730d6 DH |
1891 | struct vm_area_struct *vma, unsigned long addr) |
1892 | { | |
aaf14e40 MH |
1893 | struct page *page; |
1894 | struct mempolicy *mpol; | |
1895 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
1896 | int nid; | |
1897 | nodemask_t *nodemask; | |
1898 | ||
1899 | nid = huge_node(vma, addr, gfp_mask, &mpol, &nodemask); | |
0c397dae | 1900 | page = alloc_surplus_huge_page(h, gfp_mask, nid, nodemask); |
aaf14e40 MH |
1901 | mpol_cond_put(mpol); |
1902 | ||
1903 | return page; | |
099730d6 DH |
1904 | } |
1905 | ||
ab5ac90a | 1906 | /* page migration callback function */ |
3e59fcb0 | 1907 | struct page *alloc_huge_page_nodemask(struct hstate *h, int preferred_nid, |
d92bbc27 | 1908 | nodemask_t *nmask, gfp_t gfp_mask) |
4db9b2ef | 1909 | { |
db71ef79 | 1910 | spin_lock_irq(&hugetlb_lock); |
4db9b2ef | 1911 | if (h->free_huge_pages - h->resv_huge_pages > 0) { |
3e59fcb0 MH |
1912 | struct page *page; |
1913 | ||
1914 | page = dequeue_huge_page_nodemask(h, gfp_mask, preferred_nid, nmask); | |
1915 | if (page) { | |
db71ef79 | 1916 | spin_unlock_irq(&hugetlb_lock); |
3e59fcb0 | 1917 | return page; |
4db9b2ef MH |
1918 | } |
1919 | } | |
db71ef79 | 1920 | spin_unlock_irq(&hugetlb_lock); |
4db9b2ef | 1921 | |
0c397dae | 1922 | return alloc_migrate_huge_page(h, gfp_mask, preferred_nid, nmask); |
4db9b2ef MH |
1923 | } |
1924 | ||
ebd63723 | 1925 | /* mempolicy aware migration callback */ |
389c8178 MH |
1926 | struct page *alloc_huge_page_vma(struct hstate *h, struct vm_area_struct *vma, |
1927 | unsigned long address) | |
ebd63723 MH |
1928 | { |
1929 | struct mempolicy *mpol; | |
1930 | nodemask_t *nodemask; | |
1931 | struct page *page; | |
ebd63723 MH |
1932 | gfp_t gfp_mask; |
1933 | int node; | |
1934 | ||
ebd63723 MH |
1935 | gfp_mask = htlb_alloc_mask(h); |
1936 | node = huge_node(vma, address, gfp_mask, &mpol, &nodemask); | |
d92bbc27 | 1937 | page = alloc_huge_page_nodemask(h, node, nodemask, gfp_mask); |
ebd63723 MH |
1938 | mpol_cond_put(mpol); |
1939 | ||
1940 | return page; | |
1941 | } | |
1942 | ||
e4e574b7 | 1943 | /* |
25985edc | 1944 | * Increase the hugetlb pool such that it can accommodate a reservation |
e4e574b7 AL |
1945 | * of size 'delta'. |
1946 | */ | |
0a4f3d1b | 1947 | static int gather_surplus_pages(struct hstate *h, long delta) |
1b2a1e7b | 1948 | __must_hold(&hugetlb_lock) |
e4e574b7 AL |
1949 | { |
1950 | struct list_head surplus_list; | |
1951 | struct page *page, *tmp; | |
0a4f3d1b LX |
1952 | int ret; |
1953 | long i; | |
1954 | long needed, allocated; | |
28073b02 | 1955 | bool alloc_ok = true; |
e4e574b7 | 1956 | |
9487ca60 | 1957 | lockdep_assert_held(&hugetlb_lock); |
a5516438 | 1958 | needed = (h->resv_huge_pages + delta) - h->free_huge_pages; |
ac09b3a1 | 1959 | if (needed <= 0) { |
a5516438 | 1960 | h->resv_huge_pages += delta; |
e4e574b7 | 1961 | return 0; |
ac09b3a1 | 1962 | } |
e4e574b7 AL |
1963 | |
1964 | allocated = 0; | |
1965 | INIT_LIST_HEAD(&surplus_list); | |
1966 | ||
1967 | ret = -ENOMEM; | |
1968 | retry: | |
db71ef79 | 1969 | spin_unlock_irq(&hugetlb_lock); |
e4e574b7 | 1970 | for (i = 0; i < needed; i++) { |
0c397dae | 1971 | page = alloc_surplus_huge_page(h, htlb_alloc_mask(h), |
aaf14e40 | 1972 | NUMA_NO_NODE, NULL); |
28073b02 HD |
1973 | if (!page) { |
1974 | alloc_ok = false; | |
1975 | break; | |
1976 | } | |
e4e574b7 | 1977 | list_add(&page->lru, &surplus_list); |
69ed779a | 1978 | cond_resched(); |
e4e574b7 | 1979 | } |
28073b02 | 1980 | allocated += i; |
e4e574b7 AL |
1981 | |
1982 | /* | |
1983 | * After retaking hugetlb_lock, we need to recalculate 'needed' | |
1984 | * because either resv_huge_pages or free_huge_pages may have changed. | |
1985 | */ | |
db71ef79 | 1986 | spin_lock_irq(&hugetlb_lock); |
a5516438 AK |
1987 | needed = (h->resv_huge_pages + delta) - |
1988 | (h->free_huge_pages + allocated); | |
28073b02 HD |
1989 | if (needed > 0) { |
1990 | if (alloc_ok) | |
1991 | goto retry; | |
1992 | /* | |
1993 | * We were not able to allocate enough pages to | |
1994 | * satisfy the entire reservation so we free what | |
1995 | * we've allocated so far. | |
1996 | */ | |
1997 | goto free; | |
1998 | } | |
e4e574b7 AL |
1999 | /* |
2000 | * The surplus_list now contains _at_least_ the number of extra pages | |
25985edc | 2001 | * needed to accommodate the reservation. Add the appropriate number |
e4e574b7 | 2002 | * of pages to the hugetlb pool and free the extras back to the buddy |
ac09b3a1 AL |
2003 | * allocator. Commit the entire reservation here to prevent another |
2004 | * process from stealing the pages as they are added to the pool but | |
2005 | * before they are reserved. | |
e4e574b7 AL |
2006 | */ |
2007 | needed += allocated; | |
a5516438 | 2008 | h->resv_huge_pages += delta; |
e4e574b7 | 2009 | ret = 0; |
a9869b83 | 2010 | |
19fc3f0a | 2011 | /* Free the needed pages to the hugetlb pool */ |
e4e574b7 | 2012 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) { |
e558464b MS |
2013 | int zeroed; |
2014 | ||
19fc3f0a AL |
2015 | if ((--needed) < 0) |
2016 | break; | |
a9869b83 NH |
2017 | /* |
2018 | * This page is now managed by the hugetlb allocator and has | |
2019 | * no users -- drop the buddy allocator's reference. | |
2020 | */ | |
e558464b MS |
2021 | zeroed = put_page_testzero(page); |
2022 | VM_BUG_ON_PAGE(!zeroed, page); | |
a5516438 | 2023 | enqueue_huge_page(h, page); |
19fc3f0a | 2024 | } |
28073b02 | 2025 | free: |
db71ef79 | 2026 | spin_unlock_irq(&hugetlb_lock); |
19fc3f0a AL |
2027 | |
2028 | /* Free unnecessary surplus pages to the buddy allocator */ | |
c0d934ba JK |
2029 | list_for_each_entry_safe(page, tmp, &surplus_list, lru) |
2030 | put_page(page); | |
db71ef79 | 2031 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2032 | |
2033 | return ret; | |
2034 | } | |
2035 | ||
2036 | /* | |
e5bbc8a6 MK |
2037 | * This routine has two main purposes: |
2038 | * 1) Decrement the reservation count (resv_huge_pages) by the value passed | |
2039 | * in unused_resv_pages. This corresponds to the prior adjustments made | |
2040 | * to the associated reservation map. | |
2041 | * 2) Free any unused surplus pages that may have been allocated to satisfy | |
2042 | * the reservation. As many as unused_resv_pages may be freed. | |
e4e574b7 | 2043 | */ |
a5516438 AK |
2044 | static void return_unused_surplus_pages(struct hstate *h, |
2045 | unsigned long unused_resv_pages) | |
e4e574b7 | 2046 | { |
e4e574b7 | 2047 | unsigned long nr_pages; |
10c6ec49 MK |
2048 | struct page *page; |
2049 | LIST_HEAD(page_list); | |
2050 | ||
9487ca60 | 2051 | lockdep_assert_held(&hugetlb_lock); |
10c6ec49 MK |
2052 | /* Uncommit the reservation */ |
2053 | h->resv_huge_pages -= unused_resv_pages; | |
e4e574b7 | 2054 | |
aa888a74 | 2055 | /* Cannot return gigantic pages currently */ |
bae7f4ae | 2056 | if (hstate_is_gigantic(h)) |
e5bbc8a6 | 2057 | goto out; |
aa888a74 | 2058 | |
e5bbc8a6 MK |
2059 | /* |
2060 | * Part (or even all) of the reservation could have been backed | |
2061 | * by pre-allocated pages. Only free surplus pages. | |
2062 | */ | |
a5516438 | 2063 | nr_pages = min(unused_resv_pages, h->surplus_huge_pages); |
e4e574b7 | 2064 | |
685f3457 LS |
2065 | /* |
2066 | * We want to release as many surplus pages as possible, spread | |
9b5e5d0f LS |
2067 | * evenly across all nodes with memory. Iterate across these nodes |
2068 | * until we can no longer free unreserved surplus pages. This occurs | |
2069 | * when the nodes with surplus pages have no free pages. | |
10c6ec49 | 2070 | * remove_pool_huge_page() will balance the freed pages across the |
9b5e5d0f | 2071 | * on-line nodes with memory and will handle the hstate accounting. |
685f3457 LS |
2072 | */ |
2073 | while (nr_pages--) { | |
10c6ec49 MK |
2074 | page = remove_pool_huge_page(h, &node_states[N_MEMORY], 1); |
2075 | if (!page) | |
e5bbc8a6 | 2076 | goto out; |
10c6ec49 MK |
2077 | |
2078 | list_add(&page->lru, &page_list); | |
e4e574b7 | 2079 | } |
e5bbc8a6 MK |
2080 | |
2081 | out: | |
db71ef79 | 2082 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2083 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2084 | spin_lock_irq(&hugetlb_lock); |
e4e574b7 AL |
2085 | } |
2086 | ||
5e911373 | 2087 | |
c37f9fb1 | 2088 | /* |
feba16e2 | 2089 | * vma_needs_reservation, vma_commit_reservation and vma_end_reservation |
5e911373 | 2090 | * are used by the huge page allocation routines to manage reservations. |
cf3ad20b MK |
2091 | * |
2092 | * vma_needs_reservation is called to determine if the huge page at addr | |
2093 | * within the vma has an associated reservation. If a reservation is | |
2094 | * needed, the value 1 is returned. The caller is then responsible for | |
2095 | * managing the global reservation and subpool usage counts. After | |
2096 | * the huge page has been allocated, vma_commit_reservation is called | |
feba16e2 MK |
2097 | * to add the page to the reservation map. If the page allocation fails, |
2098 | * the reservation must be ended instead of committed. vma_end_reservation | |
2099 | * is called in such cases. | |
cf3ad20b MK |
2100 | * |
2101 | * In the normal case, vma_commit_reservation returns the same value | |
2102 | * as the preceding vma_needs_reservation call. The only time this | |
2103 | * is not the case is if a reserve map was changed between calls. It | |
2104 | * is the responsibility of the caller to notice the difference and | |
2105 | * take appropriate action. | |
96b96a96 MK |
2106 | * |
2107 | * vma_add_reservation is used in error paths where a reservation must | |
2108 | * be restored when a newly allocated huge page must be freed. It is | |
2109 | * to be called after calling vma_needs_reservation to determine if a | |
2110 | * reservation exists. | |
c37f9fb1 | 2111 | */ |
5e911373 MK |
2112 | enum vma_resv_mode { |
2113 | VMA_NEEDS_RESV, | |
2114 | VMA_COMMIT_RESV, | |
feba16e2 | 2115 | VMA_END_RESV, |
96b96a96 | 2116 | VMA_ADD_RESV, |
5e911373 | 2117 | }; |
cf3ad20b MK |
2118 | static long __vma_reservation_common(struct hstate *h, |
2119 | struct vm_area_struct *vma, unsigned long addr, | |
5e911373 | 2120 | enum vma_resv_mode mode) |
c37f9fb1 | 2121 | { |
4e35f483 JK |
2122 | struct resv_map *resv; |
2123 | pgoff_t idx; | |
cf3ad20b | 2124 | long ret; |
0db9d74e | 2125 | long dummy_out_regions_needed; |
c37f9fb1 | 2126 | |
4e35f483 JK |
2127 | resv = vma_resv_map(vma); |
2128 | if (!resv) | |
84afd99b | 2129 | return 1; |
c37f9fb1 | 2130 | |
4e35f483 | 2131 | idx = vma_hugecache_offset(h, vma, addr); |
5e911373 MK |
2132 | switch (mode) { |
2133 | case VMA_NEEDS_RESV: | |
0db9d74e MA |
2134 | ret = region_chg(resv, idx, idx + 1, &dummy_out_regions_needed); |
2135 | /* We assume that vma_reservation_* routines always operate on | |
2136 | * 1 page, and that adding to resv map a 1 page entry can only | |
2137 | * ever require 1 region. | |
2138 | */ | |
2139 | VM_BUG_ON(dummy_out_regions_needed != 1); | |
5e911373 MK |
2140 | break; |
2141 | case VMA_COMMIT_RESV: | |
075a61d0 | 2142 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2143 | /* region_add calls of range 1 should never fail. */ |
2144 | VM_BUG_ON(ret < 0); | |
5e911373 | 2145 | break; |
feba16e2 | 2146 | case VMA_END_RESV: |
0db9d74e | 2147 | region_abort(resv, idx, idx + 1, 1); |
5e911373 MK |
2148 | ret = 0; |
2149 | break; | |
96b96a96 | 2150 | case VMA_ADD_RESV: |
0db9d74e | 2151 | if (vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 2152 | ret = region_add(resv, idx, idx + 1, 1, NULL, NULL); |
0db9d74e MA |
2153 | /* region_add calls of range 1 should never fail. */ |
2154 | VM_BUG_ON(ret < 0); | |
2155 | } else { | |
2156 | region_abort(resv, idx, idx + 1, 1); | |
96b96a96 MK |
2157 | ret = region_del(resv, idx, idx + 1); |
2158 | } | |
2159 | break; | |
5e911373 MK |
2160 | default: |
2161 | BUG(); | |
2162 | } | |
84afd99b | 2163 | |
4e35f483 | 2164 | if (vma->vm_flags & VM_MAYSHARE) |
cf3ad20b | 2165 | return ret; |
bf3d12b9 ML |
2166 | /* |
2167 | * We know private mapping must have HPAGE_RESV_OWNER set. | |
2168 | * | |
2169 | * In most cases, reserves always exist for private mappings. | |
2170 | * However, a file associated with mapping could have been | |
2171 | * hole punched or truncated after reserves were consumed. | |
2172 | * As subsequent fault on such a range will not use reserves. | |
2173 | * Subtle - The reserve map for private mappings has the | |
2174 | * opposite meaning than that of shared mappings. If NO | |
2175 | * entry is in the reserve map, it means a reservation exists. | |
2176 | * If an entry exists in the reserve map, it means the | |
2177 | * reservation has already been consumed. As a result, the | |
2178 | * return value of this routine is the opposite of the | |
2179 | * value returned from reserve map manipulation routines above. | |
2180 | */ | |
2181 | if (ret > 0) | |
2182 | return 0; | |
2183 | if (ret == 0) | |
2184 | return 1; | |
2185 | return ret; | |
c37f9fb1 | 2186 | } |
cf3ad20b MK |
2187 | |
2188 | static long vma_needs_reservation(struct hstate *h, | |
a5516438 | 2189 | struct vm_area_struct *vma, unsigned long addr) |
c37f9fb1 | 2190 | { |
5e911373 | 2191 | return __vma_reservation_common(h, vma, addr, VMA_NEEDS_RESV); |
cf3ad20b | 2192 | } |
84afd99b | 2193 | |
cf3ad20b MK |
2194 | static long vma_commit_reservation(struct hstate *h, |
2195 | struct vm_area_struct *vma, unsigned long addr) | |
2196 | { | |
5e911373 MK |
2197 | return __vma_reservation_common(h, vma, addr, VMA_COMMIT_RESV); |
2198 | } | |
2199 | ||
feba16e2 | 2200 | static void vma_end_reservation(struct hstate *h, |
5e911373 MK |
2201 | struct vm_area_struct *vma, unsigned long addr) |
2202 | { | |
feba16e2 | 2203 | (void)__vma_reservation_common(h, vma, addr, VMA_END_RESV); |
c37f9fb1 AW |
2204 | } |
2205 | ||
96b96a96 MK |
2206 | static long vma_add_reservation(struct hstate *h, |
2207 | struct vm_area_struct *vma, unsigned long addr) | |
2208 | { | |
2209 | return __vma_reservation_common(h, vma, addr, VMA_ADD_RESV); | |
2210 | } | |
2211 | ||
2212 | /* | |
2213 | * This routine is called to restore a reservation on error paths. In the | |
2214 | * specific error paths, a huge page was allocated (via alloc_huge_page) | |
2215 | * and is about to be freed. If a reservation for the page existed, | |
d6995da3 MK |
2216 | * alloc_huge_page would have consumed the reservation and set |
2217 | * HPageRestoreReserve in the newly allocated page. When the page is freed | |
2218 | * via free_huge_page, the global reservation count will be incremented if | |
2219 | * HPageRestoreReserve is set. However, free_huge_page can not adjust the | |
2220 | * reserve map. Adjust the reserve map here to be consistent with global | |
2221 | * reserve count adjustments to be made by free_huge_page. | |
96b96a96 MK |
2222 | */ |
2223 | static void restore_reserve_on_error(struct hstate *h, | |
2224 | struct vm_area_struct *vma, unsigned long address, | |
2225 | struct page *page) | |
2226 | { | |
d6995da3 | 2227 | if (unlikely(HPageRestoreReserve(page))) { |
96b96a96 MK |
2228 | long rc = vma_needs_reservation(h, vma, address); |
2229 | ||
2230 | if (unlikely(rc < 0)) { | |
2231 | /* | |
2232 | * Rare out of memory condition in reserve map | |
d6995da3 | 2233 | * manipulation. Clear HPageRestoreReserve so that |
96b96a96 MK |
2234 | * global reserve count will not be incremented |
2235 | * by free_huge_page. This will make it appear | |
2236 | * as though the reservation for this page was | |
2237 | * consumed. This may prevent the task from | |
2238 | * faulting in the page at a later time. This | |
2239 | * is better than inconsistent global huge page | |
2240 | * accounting of reserve counts. | |
2241 | */ | |
d6995da3 | 2242 | ClearHPageRestoreReserve(page); |
96b96a96 MK |
2243 | } else if (rc) { |
2244 | rc = vma_add_reservation(h, vma, address); | |
2245 | if (unlikely(rc < 0)) | |
2246 | /* | |
2247 | * See above comment about rare out of | |
2248 | * memory condition. | |
2249 | */ | |
d6995da3 | 2250 | ClearHPageRestoreReserve(page); |
96b96a96 MK |
2251 | } else |
2252 | vma_end_reservation(h, vma, address); | |
2253 | } | |
2254 | } | |
2255 | ||
70c3547e | 2256 | struct page *alloc_huge_page(struct vm_area_struct *vma, |
04f2cbe3 | 2257 | unsigned long addr, int avoid_reserve) |
1da177e4 | 2258 | { |
90481622 | 2259 | struct hugepage_subpool *spool = subpool_vma(vma); |
a5516438 | 2260 | struct hstate *h = hstate_vma(vma); |
348ea204 | 2261 | struct page *page; |
d85f69b0 MK |
2262 | long map_chg, map_commit; |
2263 | long gbl_chg; | |
6d76dcf4 AK |
2264 | int ret, idx; |
2265 | struct hugetlb_cgroup *h_cg; | |
08cf9faf | 2266 | bool deferred_reserve; |
a1e78772 | 2267 | |
6d76dcf4 | 2268 | idx = hstate_index(h); |
a1e78772 | 2269 | /* |
d85f69b0 MK |
2270 | * Examine the region/reserve map to determine if the process |
2271 | * has a reservation for the page to be allocated. A return | |
2272 | * code of zero indicates a reservation exists (no change). | |
a1e78772 | 2273 | */ |
d85f69b0 MK |
2274 | map_chg = gbl_chg = vma_needs_reservation(h, vma, addr); |
2275 | if (map_chg < 0) | |
76dcee75 | 2276 | return ERR_PTR(-ENOMEM); |
d85f69b0 MK |
2277 | |
2278 | /* | |
2279 | * Processes that did not create the mapping will have no | |
2280 | * reserves as indicated by the region/reserve map. Check | |
2281 | * that the allocation will not exceed the subpool limit. | |
2282 | * Allocations for MAP_NORESERVE mappings also need to be | |
2283 | * checked against any subpool limit. | |
2284 | */ | |
2285 | if (map_chg || avoid_reserve) { | |
2286 | gbl_chg = hugepage_subpool_get_pages(spool, 1); | |
2287 | if (gbl_chg < 0) { | |
feba16e2 | 2288 | vma_end_reservation(h, vma, addr); |
76dcee75 | 2289 | return ERR_PTR(-ENOSPC); |
5e911373 | 2290 | } |
1da177e4 | 2291 | |
d85f69b0 MK |
2292 | /* |
2293 | * Even though there was no reservation in the region/reserve | |
2294 | * map, there could be reservations associated with the | |
2295 | * subpool that can be used. This would be indicated if the | |
2296 | * return value of hugepage_subpool_get_pages() is zero. | |
2297 | * However, if avoid_reserve is specified we still avoid even | |
2298 | * the subpool reservations. | |
2299 | */ | |
2300 | if (avoid_reserve) | |
2301 | gbl_chg = 1; | |
2302 | } | |
2303 | ||
08cf9faf MA |
2304 | /* If this allocation is not consuming a reservation, charge it now. |
2305 | */ | |
6501fe5f | 2306 | deferred_reserve = map_chg || avoid_reserve; |
08cf9faf MA |
2307 | if (deferred_reserve) { |
2308 | ret = hugetlb_cgroup_charge_cgroup_rsvd( | |
2309 | idx, pages_per_huge_page(h), &h_cg); | |
2310 | if (ret) | |
2311 | goto out_subpool_put; | |
2312 | } | |
2313 | ||
6d76dcf4 | 2314 | ret = hugetlb_cgroup_charge_cgroup(idx, pages_per_huge_page(h), &h_cg); |
8f34af6f | 2315 | if (ret) |
08cf9faf | 2316 | goto out_uncharge_cgroup_reservation; |
8f34af6f | 2317 | |
db71ef79 | 2318 | spin_lock_irq(&hugetlb_lock); |
d85f69b0 MK |
2319 | /* |
2320 | * glb_chg is passed to indicate whether or not a page must be taken | |
2321 | * from the global free pool (global change). gbl_chg == 0 indicates | |
2322 | * a reservation exists for the allocation. | |
2323 | */ | |
2324 | page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve, gbl_chg); | |
81a6fcae | 2325 | if (!page) { |
db71ef79 | 2326 | spin_unlock_irq(&hugetlb_lock); |
0c397dae | 2327 | page = alloc_buddy_huge_page_with_mpol(h, vma, addr); |
8f34af6f JZ |
2328 | if (!page) |
2329 | goto out_uncharge_cgroup; | |
a88c7695 | 2330 | if (!avoid_reserve && vma_has_reserves(vma, gbl_chg)) { |
d6995da3 | 2331 | SetHPageRestoreReserve(page); |
a88c7695 NH |
2332 | h->resv_huge_pages--; |
2333 | } | |
db71ef79 | 2334 | spin_lock_irq(&hugetlb_lock); |
15a8d68e | 2335 | list_add(&page->lru, &h->hugepage_activelist); |
81a6fcae | 2336 | /* Fall through */ |
68842c9b | 2337 | } |
81a6fcae | 2338 | hugetlb_cgroup_commit_charge(idx, pages_per_huge_page(h), h_cg, page); |
08cf9faf MA |
2339 | /* If allocation is not consuming a reservation, also store the |
2340 | * hugetlb_cgroup pointer on the page. | |
2341 | */ | |
2342 | if (deferred_reserve) { | |
2343 | hugetlb_cgroup_commit_charge_rsvd(idx, pages_per_huge_page(h), | |
2344 | h_cg, page); | |
2345 | } | |
2346 | ||
db71ef79 | 2347 | spin_unlock_irq(&hugetlb_lock); |
348ea204 | 2348 | |
d6995da3 | 2349 | hugetlb_set_page_subpool(page, spool); |
90d8b7e6 | 2350 | |
d85f69b0 MK |
2351 | map_commit = vma_commit_reservation(h, vma, addr); |
2352 | if (unlikely(map_chg > map_commit)) { | |
33039678 MK |
2353 | /* |
2354 | * The page was added to the reservation map between | |
2355 | * vma_needs_reservation and vma_commit_reservation. | |
2356 | * This indicates a race with hugetlb_reserve_pages. | |
2357 | * Adjust for the subpool count incremented above AND | |
2358 | * in hugetlb_reserve_pages for the same page. Also, | |
2359 | * the reservation count added in hugetlb_reserve_pages | |
2360 | * no longer applies. | |
2361 | */ | |
2362 | long rsv_adjust; | |
2363 | ||
2364 | rsv_adjust = hugepage_subpool_put_pages(spool, 1); | |
2365 | hugetlb_acct_memory(h, -rsv_adjust); | |
79aa925b MK |
2366 | if (deferred_reserve) |
2367 | hugetlb_cgroup_uncharge_page_rsvd(hstate_index(h), | |
2368 | pages_per_huge_page(h), page); | |
33039678 | 2369 | } |
90d8b7e6 | 2370 | return page; |
8f34af6f JZ |
2371 | |
2372 | out_uncharge_cgroup: | |
2373 | hugetlb_cgroup_uncharge_cgroup(idx, pages_per_huge_page(h), h_cg); | |
08cf9faf MA |
2374 | out_uncharge_cgroup_reservation: |
2375 | if (deferred_reserve) | |
2376 | hugetlb_cgroup_uncharge_cgroup_rsvd(idx, pages_per_huge_page(h), | |
2377 | h_cg); | |
8f34af6f | 2378 | out_subpool_put: |
d85f69b0 | 2379 | if (map_chg || avoid_reserve) |
8f34af6f | 2380 | hugepage_subpool_put_pages(spool, 1); |
feba16e2 | 2381 | vma_end_reservation(h, vma, addr); |
8f34af6f | 2382 | return ERR_PTR(-ENOSPC); |
b45b5bd6 DG |
2383 | } |
2384 | ||
e24a1307 AK |
2385 | int alloc_bootmem_huge_page(struct hstate *h) |
2386 | __attribute__ ((weak, alias("__alloc_bootmem_huge_page"))); | |
2387 | int __alloc_bootmem_huge_page(struct hstate *h) | |
aa888a74 AK |
2388 | { |
2389 | struct huge_bootmem_page *m; | |
b2261026 | 2390 | int nr_nodes, node; |
aa888a74 | 2391 | |
b2261026 | 2392 | for_each_node_mask_to_alloc(h, nr_nodes, node, &node_states[N_MEMORY]) { |
aa888a74 AK |
2393 | void *addr; |
2394 | ||
eb31d559 | 2395 | addr = memblock_alloc_try_nid_raw( |
8b89a116 | 2396 | huge_page_size(h), huge_page_size(h), |
97ad1087 | 2397 | 0, MEMBLOCK_ALLOC_ACCESSIBLE, node); |
aa888a74 AK |
2398 | if (addr) { |
2399 | /* | |
2400 | * Use the beginning of the huge page to store the | |
2401 | * huge_bootmem_page struct (until gather_bootmem | |
2402 | * puts them into the mem_map). | |
2403 | */ | |
2404 | m = addr; | |
91f47662 | 2405 | goto found; |
aa888a74 | 2406 | } |
aa888a74 AK |
2407 | } |
2408 | return 0; | |
2409 | ||
2410 | found: | |
df994ead | 2411 | BUG_ON(!IS_ALIGNED(virt_to_phys(m), huge_page_size(h))); |
aa888a74 | 2412 | /* Put them into a private list first because mem_map is not up yet */ |
330d6e48 | 2413 | INIT_LIST_HEAD(&m->list); |
aa888a74 AK |
2414 | list_add(&m->list, &huge_boot_pages); |
2415 | m->hstate = h; | |
2416 | return 1; | |
2417 | } | |
2418 | ||
d00181b9 KS |
2419 | static void __init prep_compound_huge_page(struct page *page, |
2420 | unsigned int order) | |
18229df5 AW |
2421 | { |
2422 | if (unlikely(order > (MAX_ORDER - 1))) | |
2423 | prep_compound_gigantic_page(page, order); | |
2424 | else | |
2425 | prep_compound_page(page, order); | |
2426 | } | |
2427 | ||
aa888a74 AK |
2428 | /* Put bootmem huge pages into the standard lists after mem_map is up */ |
2429 | static void __init gather_bootmem_prealloc(void) | |
2430 | { | |
2431 | struct huge_bootmem_page *m; | |
2432 | ||
2433 | list_for_each_entry(m, &huge_boot_pages, list) { | |
40d18ebf | 2434 | struct page *page = virt_to_page(m); |
aa888a74 | 2435 | struct hstate *h = m->hstate; |
ee8f248d | 2436 | |
aa888a74 | 2437 | WARN_ON(page_count(page) != 1); |
c78a7f36 | 2438 | prep_compound_huge_page(page, huge_page_order(h)); |
ef5a22be | 2439 | WARN_ON(PageReserved(page)); |
aa888a74 | 2440 | prep_new_huge_page(h, page, page_to_nid(page)); |
af0fb9df MH |
2441 | put_page(page); /* free it into the hugepage allocator */ |
2442 | ||
b0320c7b RA |
2443 | /* |
2444 | * If we had gigantic hugepages allocated at boot time, we need | |
2445 | * to restore the 'stolen' pages to totalram_pages in order to | |
2446 | * fix confusing memory reports from free(1) and another | |
2447 | * side-effects, like CommitLimit going negative. | |
2448 | */ | |
bae7f4ae | 2449 | if (hstate_is_gigantic(h)) |
c78a7f36 | 2450 | adjust_managed_page_count(page, pages_per_huge_page(h)); |
520495fe | 2451 | cond_resched(); |
aa888a74 AK |
2452 | } |
2453 | } | |
2454 | ||
8faa8b07 | 2455 | static void __init hugetlb_hstate_alloc_pages(struct hstate *h) |
1da177e4 LT |
2456 | { |
2457 | unsigned long i; | |
f60858f9 MK |
2458 | nodemask_t *node_alloc_noretry; |
2459 | ||
2460 | if (!hstate_is_gigantic(h)) { | |
2461 | /* | |
2462 | * Bit mask controlling how hard we retry per-node allocations. | |
2463 | * Ignore errors as lower level routines can deal with | |
2464 | * node_alloc_noretry == NULL. If this kmalloc fails at boot | |
2465 | * time, we are likely in bigger trouble. | |
2466 | */ | |
2467 | node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry), | |
2468 | GFP_KERNEL); | |
2469 | } else { | |
2470 | /* allocations done at boot time */ | |
2471 | node_alloc_noretry = NULL; | |
2472 | } | |
2473 | ||
2474 | /* bit mask controlling how hard we retry per-node allocations */ | |
2475 | if (node_alloc_noretry) | |
2476 | nodes_clear(*node_alloc_noretry); | |
a5516438 | 2477 | |
e5ff2159 | 2478 | for (i = 0; i < h->max_huge_pages; ++i) { |
bae7f4ae | 2479 | if (hstate_is_gigantic(h)) { |
dbda8fea | 2480 | if (hugetlb_cma_size) { |
cf11e85f | 2481 | pr_warn_once("HugeTLB: hugetlb_cma is enabled, skip boot time allocation\n"); |
7ecc9565 | 2482 | goto free; |
cf11e85f | 2483 | } |
aa888a74 AK |
2484 | if (!alloc_bootmem_huge_page(h)) |
2485 | break; | |
0c397dae | 2486 | } else if (!alloc_pool_huge_page(h, |
f60858f9 MK |
2487 | &node_states[N_MEMORY], |
2488 | node_alloc_noretry)) | |
1da177e4 | 2489 | break; |
69ed779a | 2490 | cond_resched(); |
1da177e4 | 2491 | } |
d715cf80 LH |
2492 | if (i < h->max_huge_pages) { |
2493 | char buf[32]; | |
2494 | ||
c6247f72 | 2495 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); |
d715cf80 LH |
2496 | pr_warn("HugeTLB: allocating %lu of page size %s failed. Only allocated %lu hugepages.\n", |
2497 | h->max_huge_pages, buf, i); | |
2498 | h->max_huge_pages = i; | |
2499 | } | |
7ecc9565 | 2500 | free: |
f60858f9 | 2501 | kfree(node_alloc_noretry); |
e5ff2159 AK |
2502 | } |
2503 | ||
2504 | static void __init hugetlb_init_hstates(void) | |
2505 | { | |
2506 | struct hstate *h; | |
2507 | ||
2508 | for_each_hstate(h) { | |
641844f5 NH |
2509 | if (minimum_order > huge_page_order(h)) |
2510 | minimum_order = huge_page_order(h); | |
2511 | ||
8faa8b07 | 2512 | /* oversize hugepages were init'ed in early boot */ |
bae7f4ae | 2513 | if (!hstate_is_gigantic(h)) |
8faa8b07 | 2514 | hugetlb_hstate_alloc_pages(h); |
e5ff2159 | 2515 | } |
641844f5 | 2516 | VM_BUG_ON(minimum_order == UINT_MAX); |
e5ff2159 AK |
2517 | } |
2518 | ||
2519 | static void __init report_hugepages(void) | |
2520 | { | |
2521 | struct hstate *h; | |
2522 | ||
2523 | for_each_hstate(h) { | |
4abd32db | 2524 | char buf[32]; |
c6247f72 MW |
2525 | |
2526 | string_get_size(huge_page_size(h), 1, STRING_UNITS_2, buf, 32); | |
ffb22af5 | 2527 | pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n", |
c6247f72 | 2528 | buf, h->free_huge_pages); |
e5ff2159 AK |
2529 | } |
2530 | } | |
2531 | ||
1da177e4 | 2532 | #ifdef CONFIG_HIGHMEM |
6ae11b27 LS |
2533 | static void try_to_free_low(struct hstate *h, unsigned long count, |
2534 | nodemask_t *nodes_allowed) | |
1da177e4 | 2535 | { |
4415cc8d | 2536 | int i; |
1121828a | 2537 | LIST_HEAD(page_list); |
4415cc8d | 2538 | |
9487ca60 | 2539 | lockdep_assert_held(&hugetlb_lock); |
bae7f4ae | 2540 | if (hstate_is_gigantic(h)) |
aa888a74 AK |
2541 | return; |
2542 | ||
1121828a MK |
2543 | /* |
2544 | * Collect pages to be freed on a list, and free after dropping lock | |
2545 | */ | |
6ae11b27 | 2546 | for_each_node_mask(i, *nodes_allowed) { |
10c6ec49 | 2547 | struct page *page, *next; |
a5516438 AK |
2548 | struct list_head *freel = &h->hugepage_freelists[i]; |
2549 | list_for_each_entry_safe(page, next, freel, lru) { | |
2550 | if (count >= h->nr_huge_pages) | |
1121828a | 2551 | goto out; |
1da177e4 LT |
2552 | if (PageHighMem(page)) |
2553 | continue; | |
6eb4e88a | 2554 | remove_hugetlb_page(h, page, false); |
1121828a | 2555 | list_add(&page->lru, &page_list); |
1da177e4 LT |
2556 | } |
2557 | } | |
1121828a MK |
2558 | |
2559 | out: | |
db71ef79 | 2560 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2561 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2562 | spin_lock_irq(&hugetlb_lock); |
1da177e4 LT |
2563 | } |
2564 | #else | |
6ae11b27 LS |
2565 | static inline void try_to_free_low(struct hstate *h, unsigned long count, |
2566 | nodemask_t *nodes_allowed) | |
1da177e4 LT |
2567 | { |
2568 | } | |
2569 | #endif | |
2570 | ||
20a0307c WF |
2571 | /* |
2572 | * Increment or decrement surplus_huge_pages. Keep node-specific counters | |
2573 | * balanced by operating on them in a round-robin fashion. | |
2574 | * Returns 1 if an adjustment was made. | |
2575 | */ | |
6ae11b27 LS |
2576 | static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed, |
2577 | int delta) | |
20a0307c | 2578 | { |
b2261026 | 2579 | int nr_nodes, node; |
20a0307c | 2580 | |
9487ca60 | 2581 | lockdep_assert_held(&hugetlb_lock); |
20a0307c | 2582 | VM_BUG_ON(delta != -1 && delta != 1); |
20a0307c | 2583 | |
b2261026 JK |
2584 | if (delta < 0) { |
2585 | for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) { | |
2586 | if (h->surplus_huge_pages_node[node]) | |
2587 | goto found; | |
e8c5c824 | 2588 | } |
b2261026 JK |
2589 | } else { |
2590 | for_each_node_mask_to_free(h, nr_nodes, node, nodes_allowed) { | |
2591 | if (h->surplus_huge_pages_node[node] < | |
2592 | h->nr_huge_pages_node[node]) | |
2593 | goto found; | |
e8c5c824 | 2594 | } |
b2261026 JK |
2595 | } |
2596 | return 0; | |
20a0307c | 2597 | |
b2261026 JK |
2598 | found: |
2599 | h->surplus_huge_pages += delta; | |
2600 | h->surplus_huge_pages_node[node] += delta; | |
2601 | return 1; | |
20a0307c WF |
2602 | } |
2603 | ||
a5516438 | 2604 | #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages) |
fd875dca | 2605 | static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid, |
4eb0716e | 2606 | nodemask_t *nodes_allowed) |
1da177e4 | 2607 | { |
7893d1d5 | 2608 | unsigned long min_count, ret; |
10c6ec49 MK |
2609 | struct page *page; |
2610 | LIST_HEAD(page_list); | |
f60858f9 MK |
2611 | NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL); |
2612 | ||
2613 | /* | |
2614 | * Bit mask controlling how hard we retry per-node allocations. | |
2615 | * If we can not allocate the bit mask, do not attempt to allocate | |
2616 | * the requested huge pages. | |
2617 | */ | |
2618 | if (node_alloc_noretry) | |
2619 | nodes_clear(*node_alloc_noretry); | |
2620 | else | |
2621 | return -ENOMEM; | |
1da177e4 | 2622 | |
29383967 MK |
2623 | /* |
2624 | * resize_lock mutex prevents concurrent adjustments to number of | |
2625 | * pages in hstate via the proc/sysfs interfaces. | |
2626 | */ | |
2627 | mutex_lock(&h->resize_lock); | |
db71ef79 | 2628 | spin_lock_irq(&hugetlb_lock); |
4eb0716e | 2629 | |
fd875dca MK |
2630 | /* |
2631 | * Check for a node specific request. | |
2632 | * Changing node specific huge page count may require a corresponding | |
2633 | * change to the global count. In any case, the passed node mask | |
2634 | * (nodes_allowed) will restrict alloc/free to the specified node. | |
2635 | */ | |
2636 | if (nid != NUMA_NO_NODE) { | |
2637 | unsigned long old_count = count; | |
2638 | ||
2639 | count += h->nr_huge_pages - h->nr_huge_pages_node[nid]; | |
2640 | /* | |
2641 | * User may have specified a large count value which caused the | |
2642 | * above calculation to overflow. In this case, they wanted | |
2643 | * to allocate as many huge pages as possible. Set count to | |
2644 | * largest possible value to align with their intention. | |
2645 | */ | |
2646 | if (count < old_count) | |
2647 | count = ULONG_MAX; | |
2648 | } | |
2649 | ||
4eb0716e AG |
2650 | /* |
2651 | * Gigantic pages runtime allocation depend on the capability for large | |
2652 | * page range allocation. | |
2653 | * If the system does not provide this feature, return an error when | |
2654 | * the user tries to allocate gigantic pages but let the user free the | |
2655 | * boottime allocated gigantic pages. | |
2656 | */ | |
2657 | if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) { | |
2658 | if (count > persistent_huge_pages(h)) { | |
db71ef79 | 2659 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 2660 | mutex_unlock(&h->resize_lock); |
f60858f9 | 2661 | NODEMASK_FREE(node_alloc_noretry); |
4eb0716e AG |
2662 | return -EINVAL; |
2663 | } | |
2664 | /* Fall through to decrease pool */ | |
2665 | } | |
aa888a74 | 2666 | |
7893d1d5 AL |
2667 | /* |
2668 | * Increase the pool size | |
2669 | * First take pages out of surplus state. Then make up the | |
2670 | * remaining difference by allocating fresh huge pages. | |
d1c3fb1f | 2671 | * |
0c397dae | 2672 | * We might race with alloc_surplus_huge_page() here and be unable |
d1c3fb1f NA |
2673 | * to convert a surplus huge page to a normal huge page. That is |
2674 | * not critical, though, it just means the overall size of the | |
2675 | * pool might be one hugepage larger than it needs to be, but | |
2676 | * within all the constraints specified by the sysctls. | |
7893d1d5 | 2677 | */ |
a5516438 | 2678 | while (h->surplus_huge_pages && count > persistent_huge_pages(h)) { |
6ae11b27 | 2679 | if (!adjust_pool_surplus(h, nodes_allowed, -1)) |
7893d1d5 AL |
2680 | break; |
2681 | } | |
2682 | ||
a5516438 | 2683 | while (count > persistent_huge_pages(h)) { |
7893d1d5 AL |
2684 | /* |
2685 | * If this allocation races such that we no longer need the | |
2686 | * page, free_huge_page will handle it by freeing the page | |
2687 | * and reducing the surplus. | |
2688 | */ | |
db71ef79 | 2689 | spin_unlock_irq(&hugetlb_lock); |
649920c6 JH |
2690 | |
2691 | /* yield cpu to avoid soft lockup */ | |
2692 | cond_resched(); | |
2693 | ||
f60858f9 MK |
2694 | ret = alloc_pool_huge_page(h, nodes_allowed, |
2695 | node_alloc_noretry); | |
db71ef79 | 2696 | spin_lock_irq(&hugetlb_lock); |
7893d1d5 AL |
2697 | if (!ret) |
2698 | goto out; | |
2699 | ||
536240f2 MG |
2700 | /* Bail for signals. Probably ctrl-c from user */ |
2701 | if (signal_pending(current)) | |
2702 | goto out; | |
7893d1d5 | 2703 | } |
7893d1d5 AL |
2704 | |
2705 | /* | |
2706 | * Decrease the pool size | |
2707 | * First return free pages to the buddy allocator (being careful | |
2708 | * to keep enough around to satisfy reservations). Then place | |
2709 | * pages into surplus state as needed so the pool will shrink | |
2710 | * to the desired size as pages become free. | |
d1c3fb1f NA |
2711 | * |
2712 | * By placing pages into the surplus state independent of the | |
2713 | * overcommit value, we are allowing the surplus pool size to | |
2714 | * exceed overcommit. There are few sane options here. Since | |
0c397dae | 2715 | * alloc_surplus_huge_page() is checking the global counter, |
d1c3fb1f NA |
2716 | * though, we'll note that we're not allowed to exceed surplus |
2717 | * and won't grow the pool anywhere else. Not until one of the | |
2718 | * sysctls are changed, or the surplus pages go out of use. | |
7893d1d5 | 2719 | */ |
a5516438 | 2720 | min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages; |
6b0c880d | 2721 | min_count = max(count, min_count); |
6ae11b27 | 2722 | try_to_free_low(h, min_count, nodes_allowed); |
10c6ec49 MK |
2723 | |
2724 | /* | |
2725 | * Collect pages to be removed on list without dropping lock | |
2726 | */ | |
a5516438 | 2727 | while (min_count < persistent_huge_pages(h)) { |
10c6ec49 MK |
2728 | page = remove_pool_huge_page(h, nodes_allowed, 0); |
2729 | if (!page) | |
1da177e4 | 2730 | break; |
10c6ec49 MK |
2731 | |
2732 | list_add(&page->lru, &page_list); | |
1da177e4 | 2733 | } |
10c6ec49 | 2734 | /* free the pages after dropping lock */ |
db71ef79 | 2735 | spin_unlock_irq(&hugetlb_lock); |
10c6ec49 | 2736 | update_and_free_pages_bulk(h, &page_list); |
db71ef79 | 2737 | spin_lock_irq(&hugetlb_lock); |
10c6ec49 | 2738 | |
a5516438 | 2739 | while (count < persistent_huge_pages(h)) { |
6ae11b27 | 2740 | if (!adjust_pool_surplus(h, nodes_allowed, 1)) |
7893d1d5 AL |
2741 | break; |
2742 | } | |
2743 | out: | |
4eb0716e | 2744 | h->max_huge_pages = persistent_huge_pages(h); |
db71ef79 | 2745 | spin_unlock_irq(&hugetlb_lock); |
29383967 | 2746 | mutex_unlock(&h->resize_lock); |
4eb0716e | 2747 | |
f60858f9 MK |
2748 | NODEMASK_FREE(node_alloc_noretry); |
2749 | ||
4eb0716e | 2750 | return 0; |
1da177e4 LT |
2751 | } |
2752 | ||
a3437870 NA |
2753 | #define HSTATE_ATTR_RO(_name) \ |
2754 | static struct kobj_attribute _name##_attr = __ATTR_RO(_name) | |
2755 | ||
2756 | #define HSTATE_ATTR(_name) \ | |
2757 | static struct kobj_attribute _name##_attr = \ | |
2758 | __ATTR(_name, 0644, _name##_show, _name##_store) | |
2759 | ||
2760 | static struct kobject *hugepages_kobj; | |
2761 | static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
2762 | ||
9a305230 LS |
2763 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp); |
2764 | ||
2765 | static struct hstate *kobj_to_hstate(struct kobject *kobj, int *nidp) | |
a3437870 NA |
2766 | { |
2767 | int i; | |
9a305230 | 2768 | |
a3437870 | 2769 | for (i = 0; i < HUGE_MAX_HSTATE; i++) |
9a305230 LS |
2770 | if (hstate_kobjs[i] == kobj) { |
2771 | if (nidp) | |
2772 | *nidp = NUMA_NO_NODE; | |
a3437870 | 2773 | return &hstates[i]; |
9a305230 LS |
2774 | } |
2775 | ||
2776 | return kobj_to_node_hstate(kobj, nidp); | |
a3437870 NA |
2777 | } |
2778 | ||
06808b08 | 2779 | static ssize_t nr_hugepages_show_common(struct kobject *kobj, |
a3437870 NA |
2780 | struct kobj_attribute *attr, char *buf) |
2781 | { | |
9a305230 LS |
2782 | struct hstate *h; |
2783 | unsigned long nr_huge_pages; | |
2784 | int nid; | |
2785 | ||
2786 | h = kobj_to_hstate(kobj, &nid); | |
2787 | if (nid == NUMA_NO_NODE) | |
2788 | nr_huge_pages = h->nr_huge_pages; | |
2789 | else | |
2790 | nr_huge_pages = h->nr_huge_pages_node[nid]; | |
2791 | ||
ae7a927d | 2792 | return sysfs_emit(buf, "%lu\n", nr_huge_pages); |
a3437870 | 2793 | } |
adbe8726 | 2794 | |
238d3c13 DR |
2795 | static ssize_t __nr_hugepages_store_common(bool obey_mempolicy, |
2796 | struct hstate *h, int nid, | |
2797 | unsigned long count, size_t len) | |
a3437870 NA |
2798 | { |
2799 | int err; | |
2d0adf7e | 2800 | nodemask_t nodes_allowed, *n_mask; |
a3437870 | 2801 | |
2d0adf7e OS |
2802 | if (hstate_is_gigantic(h) && !gigantic_page_runtime_supported()) |
2803 | return -EINVAL; | |
adbe8726 | 2804 | |
9a305230 LS |
2805 | if (nid == NUMA_NO_NODE) { |
2806 | /* | |
2807 | * global hstate attribute | |
2808 | */ | |
2809 | if (!(obey_mempolicy && | |
2d0adf7e OS |
2810 | init_nodemask_of_mempolicy(&nodes_allowed))) |
2811 | n_mask = &node_states[N_MEMORY]; | |
2812 | else | |
2813 | n_mask = &nodes_allowed; | |
2814 | } else { | |
9a305230 | 2815 | /* |
fd875dca MK |
2816 | * Node specific request. count adjustment happens in |
2817 | * set_max_huge_pages() after acquiring hugetlb_lock. | |
9a305230 | 2818 | */ |
2d0adf7e OS |
2819 | init_nodemask_of_node(&nodes_allowed, nid); |
2820 | n_mask = &nodes_allowed; | |
fd875dca | 2821 | } |
9a305230 | 2822 | |
2d0adf7e | 2823 | err = set_max_huge_pages(h, count, nid, n_mask); |
06808b08 | 2824 | |
4eb0716e | 2825 | return err ? err : len; |
06808b08 LS |
2826 | } |
2827 | ||
238d3c13 DR |
2828 | static ssize_t nr_hugepages_store_common(bool obey_mempolicy, |
2829 | struct kobject *kobj, const char *buf, | |
2830 | size_t len) | |
2831 | { | |
2832 | struct hstate *h; | |
2833 | unsigned long count; | |
2834 | int nid; | |
2835 | int err; | |
2836 | ||
2837 | err = kstrtoul(buf, 10, &count); | |
2838 | if (err) | |
2839 | return err; | |
2840 | ||
2841 | h = kobj_to_hstate(kobj, &nid); | |
2842 | return __nr_hugepages_store_common(obey_mempolicy, h, nid, count, len); | |
2843 | } | |
2844 | ||
06808b08 LS |
2845 | static ssize_t nr_hugepages_show(struct kobject *kobj, |
2846 | struct kobj_attribute *attr, char *buf) | |
2847 | { | |
2848 | return nr_hugepages_show_common(kobj, attr, buf); | |
2849 | } | |
2850 | ||
2851 | static ssize_t nr_hugepages_store(struct kobject *kobj, | |
2852 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2853 | { | |
238d3c13 | 2854 | return nr_hugepages_store_common(false, kobj, buf, len); |
a3437870 NA |
2855 | } |
2856 | HSTATE_ATTR(nr_hugepages); | |
2857 | ||
06808b08 LS |
2858 | #ifdef CONFIG_NUMA |
2859 | ||
2860 | /* | |
2861 | * hstate attribute for optionally mempolicy-based constraint on persistent | |
2862 | * huge page alloc/free. | |
2863 | */ | |
2864 | static ssize_t nr_hugepages_mempolicy_show(struct kobject *kobj, | |
ae7a927d JP |
2865 | struct kobj_attribute *attr, |
2866 | char *buf) | |
06808b08 LS |
2867 | { |
2868 | return nr_hugepages_show_common(kobj, attr, buf); | |
2869 | } | |
2870 | ||
2871 | static ssize_t nr_hugepages_mempolicy_store(struct kobject *kobj, | |
2872 | struct kobj_attribute *attr, const char *buf, size_t len) | |
2873 | { | |
238d3c13 | 2874 | return nr_hugepages_store_common(true, kobj, buf, len); |
06808b08 LS |
2875 | } |
2876 | HSTATE_ATTR(nr_hugepages_mempolicy); | |
2877 | #endif | |
2878 | ||
2879 | ||
a3437870 NA |
2880 | static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj, |
2881 | struct kobj_attribute *attr, char *buf) | |
2882 | { | |
9a305230 | 2883 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 2884 | return sysfs_emit(buf, "%lu\n", h->nr_overcommit_huge_pages); |
a3437870 | 2885 | } |
adbe8726 | 2886 | |
a3437870 NA |
2887 | static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj, |
2888 | struct kobj_attribute *attr, const char *buf, size_t count) | |
2889 | { | |
2890 | int err; | |
2891 | unsigned long input; | |
9a305230 | 2892 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
a3437870 | 2893 | |
bae7f4ae | 2894 | if (hstate_is_gigantic(h)) |
adbe8726 EM |
2895 | return -EINVAL; |
2896 | ||
3dbb95f7 | 2897 | err = kstrtoul(buf, 10, &input); |
a3437870 | 2898 | if (err) |
73ae31e5 | 2899 | return err; |
a3437870 | 2900 | |
db71ef79 | 2901 | spin_lock_irq(&hugetlb_lock); |
a3437870 | 2902 | h->nr_overcommit_huge_pages = input; |
db71ef79 | 2903 | spin_unlock_irq(&hugetlb_lock); |
a3437870 NA |
2904 | |
2905 | return count; | |
2906 | } | |
2907 | HSTATE_ATTR(nr_overcommit_hugepages); | |
2908 | ||
2909 | static ssize_t free_hugepages_show(struct kobject *kobj, | |
2910 | struct kobj_attribute *attr, char *buf) | |
2911 | { | |
9a305230 LS |
2912 | struct hstate *h; |
2913 | unsigned long free_huge_pages; | |
2914 | int nid; | |
2915 | ||
2916 | h = kobj_to_hstate(kobj, &nid); | |
2917 | if (nid == NUMA_NO_NODE) | |
2918 | free_huge_pages = h->free_huge_pages; | |
2919 | else | |
2920 | free_huge_pages = h->free_huge_pages_node[nid]; | |
2921 | ||
ae7a927d | 2922 | return sysfs_emit(buf, "%lu\n", free_huge_pages); |
a3437870 NA |
2923 | } |
2924 | HSTATE_ATTR_RO(free_hugepages); | |
2925 | ||
2926 | static ssize_t resv_hugepages_show(struct kobject *kobj, | |
2927 | struct kobj_attribute *attr, char *buf) | |
2928 | { | |
9a305230 | 2929 | struct hstate *h = kobj_to_hstate(kobj, NULL); |
ae7a927d | 2930 | return sysfs_emit(buf, "%lu\n", h->resv_huge_pages); |
a3437870 NA |
2931 | } |
2932 | HSTATE_ATTR_RO(resv_hugepages); | |
2933 | ||
2934 | static ssize_t surplus_hugepages_show(struct kobject *kobj, | |
2935 | struct kobj_attribute *attr, char *buf) | |
2936 | { | |
9a305230 LS |
2937 | struct hstate *h; |
2938 | unsigned long surplus_huge_pages; | |
2939 | int nid; | |
2940 | ||
2941 | h = kobj_to_hstate(kobj, &nid); | |
2942 | if (nid == NUMA_NO_NODE) | |
2943 | surplus_huge_pages = h->surplus_huge_pages; | |
2944 | else | |
2945 | surplus_huge_pages = h->surplus_huge_pages_node[nid]; | |
2946 | ||
ae7a927d | 2947 | return sysfs_emit(buf, "%lu\n", surplus_huge_pages); |
a3437870 NA |
2948 | } |
2949 | HSTATE_ATTR_RO(surplus_hugepages); | |
2950 | ||
2951 | static struct attribute *hstate_attrs[] = { | |
2952 | &nr_hugepages_attr.attr, | |
2953 | &nr_overcommit_hugepages_attr.attr, | |
2954 | &free_hugepages_attr.attr, | |
2955 | &resv_hugepages_attr.attr, | |
2956 | &surplus_hugepages_attr.attr, | |
06808b08 LS |
2957 | #ifdef CONFIG_NUMA |
2958 | &nr_hugepages_mempolicy_attr.attr, | |
2959 | #endif | |
a3437870 NA |
2960 | NULL, |
2961 | }; | |
2962 | ||
67e5ed96 | 2963 | static const struct attribute_group hstate_attr_group = { |
a3437870 NA |
2964 | .attrs = hstate_attrs, |
2965 | }; | |
2966 | ||
094e9539 JM |
2967 | static int hugetlb_sysfs_add_hstate(struct hstate *h, struct kobject *parent, |
2968 | struct kobject **hstate_kobjs, | |
67e5ed96 | 2969 | const struct attribute_group *hstate_attr_group) |
a3437870 NA |
2970 | { |
2971 | int retval; | |
972dc4de | 2972 | int hi = hstate_index(h); |
a3437870 | 2973 | |
9a305230 LS |
2974 | hstate_kobjs[hi] = kobject_create_and_add(h->name, parent); |
2975 | if (!hstate_kobjs[hi]) | |
a3437870 NA |
2976 | return -ENOMEM; |
2977 | ||
9a305230 | 2978 | retval = sysfs_create_group(hstate_kobjs[hi], hstate_attr_group); |
cc2205a6 | 2979 | if (retval) { |
9a305230 | 2980 | kobject_put(hstate_kobjs[hi]); |
cc2205a6 ML |
2981 | hstate_kobjs[hi] = NULL; |
2982 | } | |
a3437870 NA |
2983 | |
2984 | return retval; | |
2985 | } | |
2986 | ||
2987 | static void __init hugetlb_sysfs_init(void) | |
2988 | { | |
2989 | struct hstate *h; | |
2990 | int err; | |
2991 | ||
2992 | hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj); | |
2993 | if (!hugepages_kobj) | |
2994 | return; | |
2995 | ||
2996 | for_each_hstate(h) { | |
9a305230 LS |
2997 | err = hugetlb_sysfs_add_hstate(h, hugepages_kobj, |
2998 | hstate_kobjs, &hstate_attr_group); | |
a3437870 | 2999 | if (err) |
282f4214 | 3000 | pr_err("HugeTLB: Unable to add hstate %s", h->name); |
a3437870 NA |
3001 | } |
3002 | } | |
3003 | ||
9a305230 LS |
3004 | #ifdef CONFIG_NUMA |
3005 | ||
3006 | /* | |
3007 | * node_hstate/s - associate per node hstate attributes, via their kobjects, | |
10fbcf4c KS |
3008 | * with node devices in node_devices[] using a parallel array. The array |
3009 | * index of a node device or _hstate == node id. | |
3010 | * This is here to avoid any static dependency of the node device driver, in | |
9a305230 LS |
3011 | * the base kernel, on the hugetlb module. |
3012 | */ | |
3013 | struct node_hstate { | |
3014 | struct kobject *hugepages_kobj; | |
3015 | struct kobject *hstate_kobjs[HUGE_MAX_HSTATE]; | |
3016 | }; | |
b4e289a6 | 3017 | static struct node_hstate node_hstates[MAX_NUMNODES]; |
9a305230 LS |
3018 | |
3019 | /* | |
10fbcf4c | 3020 | * A subset of global hstate attributes for node devices |
9a305230 LS |
3021 | */ |
3022 | static struct attribute *per_node_hstate_attrs[] = { | |
3023 | &nr_hugepages_attr.attr, | |
3024 | &free_hugepages_attr.attr, | |
3025 | &surplus_hugepages_attr.attr, | |
3026 | NULL, | |
3027 | }; | |
3028 | ||
67e5ed96 | 3029 | static const struct attribute_group per_node_hstate_attr_group = { |
9a305230 LS |
3030 | .attrs = per_node_hstate_attrs, |
3031 | }; | |
3032 | ||
3033 | /* | |
10fbcf4c | 3034 | * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj. |
9a305230 LS |
3035 | * Returns node id via non-NULL nidp. |
3036 | */ | |
3037 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3038 | { | |
3039 | int nid; | |
3040 | ||
3041 | for (nid = 0; nid < nr_node_ids; nid++) { | |
3042 | struct node_hstate *nhs = &node_hstates[nid]; | |
3043 | int i; | |
3044 | for (i = 0; i < HUGE_MAX_HSTATE; i++) | |
3045 | if (nhs->hstate_kobjs[i] == kobj) { | |
3046 | if (nidp) | |
3047 | *nidp = nid; | |
3048 | return &hstates[i]; | |
3049 | } | |
3050 | } | |
3051 | ||
3052 | BUG(); | |
3053 | return NULL; | |
3054 | } | |
3055 | ||
3056 | /* | |
10fbcf4c | 3057 | * Unregister hstate attributes from a single node device. |
9a305230 LS |
3058 | * No-op if no hstate attributes attached. |
3059 | */ | |
3cd8b44f | 3060 | static void hugetlb_unregister_node(struct node *node) |
9a305230 LS |
3061 | { |
3062 | struct hstate *h; | |
10fbcf4c | 3063 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3064 | |
3065 | if (!nhs->hugepages_kobj) | |
9b5e5d0f | 3066 | return; /* no hstate attributes */ |
9a305230 | 3067 | |
972dc4de AK |
3068 | for_each_hstate(h) { |
3069 | int idx = hstate_index(h); | |
3070 | if (nhs->hstate_kobjs[idx]) { | |
3071 | kobject_put(nhs->hstate_kobjs[idx]); | |
3072 | nhs->hstate_kobjs[idx] = NULL; | |
9a305230 | 3073 | } |
972dc4de | 3074 | } |
9a305230 LS |
3075 | |
3076 | kobject_put(nhs->hugepages_kobj); | |
3077 | nhs->hugepages_kobj = NULL; | |
3078 | } | |
3079 | ||
9a305230 LS |
3080 | |
3081 | /* | |
10fbcf4c | 3082 | * Register hstate attributes for a single node device. |
9a305230 LS |
3083 | * No-op if attributes already registered. |
3084 | */ | |
3cd8b44f | 3085 | static void hugetlb_register_node(struct node *node) |
9a305230 LS |
3086 | { |
3087 | struct hstate *h; | |
10fbcf4c | 3088 | struct node_hstate *nhs = &node_hstates[node->dev.id]; |
9a305230 LS |
3089 | int err; |
3090 | ||
3091 | if (nhs->hugepages_kobj) | |
3092 | return; /* already allocated */ | |
3093 | ||
3094 | nhs->hugepages_kobj = kobject_create_and_add("hugepages", | |
10fbcf4c | 3095 | &node->dev.kobj); |
9a305230 LS |
3096 | if (!nhs->hugepages_kobj) |
3097 | return; | |
3098 | ||
3099 | for_each_hstate(h) { | |
3100 | err = hugetlb_sysfs_add_hstate(h, nhs->hugepages_kobj, | |
3101 | nhs->hstate_kobjs, | |
3102 | &per_node_hstate_attr_group); | |
3103 | if (err) { | |
282f4214 | 3104 | pr_err("HugeTLB: Unable to add hstate %s for node %d\n", |
ffb22af5 | 3105 | h->name, node->dev.id); |
9a305230 LS |
3106 | hugetlb_unregister_node(node); |
3107 | break; | |
3108 | } | |
3109 | } | |
3110 | } | |
3111 | ||
3112 | /* | |
9b5e5d0f | 3113 | * hugetlb init time: register hstate attributes for all registered node |
10fbcf4c KS |
3114 | * devices of nodes that have memory. All on-line nodes should have |
3115 | * registered their associated device by this time. | |
9a305230 | 3116 | */ |
7d9ca000 | 3117 | static void __init hugetlb_register_all_nodes(void) |
9a305230 LS |
3118 | { |
3119 | int nid; | |
3120 | ||
8cebfcd0 | 3121 | for_each_node_state(nid, N_MEMORY) { |
8732794b | 3122 | struct node *node = node_devices[nid]; |
10fbcf4c | 3123 | if (node->dev.id == nid) |
9a305230 LS |
3124 | hugetlb_register_node(node); |
3125 | } | |
3126 | ||
3127 | /* | |
10fbcf4c | 3128 | * Let the node device driver know we're here so it can |
9a305230 LS |
3129 | * [un]register hstate attributes on node hotplug. |
3130 | */ | |
3131 | register_hugetlbfs_with_node(hugetlb_register_node, | |
3132 | hugetlb_unregister_node); | |
3133 | } | |
3134 | #else /* !CONFIG_NUMA */ | |
3135 | ||
3136 | static struct hstate *kobj_to_node_hstate(struct kobject *kobj, int *nidp) | |
3137 | { | |
3138 | BUG(); | |
3139 | if (nidp) | |
3140 | *nidp = -1; | |
3141 | return NULL; | |
3142 | } | |
3143 | ||
9a305230 LS |
3144 | static void hugetlb_register_all_nodes(void) { } |
3145 | ||
3146 | #endif | |
3147 | ||
a3437870 NA |
3148 | static int __init hugetlb_init(void) |
3149 | { | |
8382d914 DB |
3150 | int i; |
3151 | ||
d6995da3 MK |
3152 | BUILD_BUG_ON(sizeof_field(struct page, private) * BITS_PER_BYTE < |
3153 | __NR_HPAGEFLAGS); | |
3154 | ||
c2833a5b MK |
3155 | if (!hugepages_supported()) { |
3156 | if (hugetlb_max_hstate || default_hstate_max_huge_pages) | |
3157 | pr_warn("HugeTLB: huge pages not supported, ignoring associated command-line parameters\n"); | |
0ef89d25 | 3158 | return 0; |
c2833a5b | 3159 | } |
a3437870 | 3160 | |
282f4214 MK |
3161 | /* |
3162 | * Make sure HPAGE_SIZE (HUGETLB_PAGE_ORDER) hstate exists. Some | |
3163 | * architectures depend on setup being done here. | |
3164 | */ | |
3165 | hugetlb_add_hstate(HUGETLB_PAGE_ORDER); | |
3166 | if (!parsed_default_hugepagesz) { | |
3167 | /* | |
3168 | * If we did not parse a default huge page size, set | |
3169 | * default_hstate_idx to HPAGE_SIZE hstate. And, if the | |
3170 | * number of huge pages for this default size was implicitly | |
3171 | * specified, set that here as well. | |
3172 | * Note that the implicit setting will overwrite an explicit | |
3173 | * setting. A warning will be printed in this case. | |
3174 | */ | |
3175 | default_hstate_idx = hstate_index(size_to_hstate(HPAGE_SIZE)); | |
3176 | if (default_hstate_max_huge_pages) { | |
3177 | if (default_hstate.max_huge_pages) { | |
3178 | char buf[32]; | |
3179 | ||
3180 | string_get_size(huge_page_size(&default_hstate), | |
3181 | 1, STRING_UNITS_2, buf, 32); | |
3182 | pr_warn("HugeTLB: Ignoring hugepages=%lu associated with %s page size\n", | |
3183 | default_hstate.max_huge_pages, buf); | |
3184 | pr_warn("HugeTLB: Using hugepages=%lu for number of default huge pages\n", | |
3185 | default_hstate_max_huge_pages); | |
3186 | } | |
3187 | default_hstate.max_huge_pages = | |
3188 | default_hstate_max_huge_pages; | |
d715cf80 | 3189 | } |
f8b74815 | 3190 | } |
a3437870 | 3191 | |
cf11e85f | 3192 | hugetlb_cma_check(); |
a3437870 | 3193 | hugetlb_init_hstates(); |
aa888a74 | 3194 | gather_bootmem_prealloc(); |
a3437870 NA |
3195 | report_hugepages(); |
3196 | ||
3197 | hugetlb_sysfs_init(); | |
9a305230 | 3198 | hugetlb_register_all_nodes(); |
7179e7bf | 3199 | hugetlb_cgroup_file_init(); |
9a305230 | 3200 | |
8382d914 DB |
3201 | #ifdef CONFIG_SMP |
3202 | num_fault_mutexes = roundup_pow_of_two(8 * num_possible_cpus()); | |
3203 | #else | |
3204 | num_fault_mutexes = 1; | |
3205 | #endif | |
c672c7f2 | 3206 | hugetlb_fault_mutex_table = |
6da2ec56 KC |
3207 | kmalloc_array(num_fault_mutexes, sizeof(struct mutex), |
3208 | GFP_KERNEL); | |
c672c7f2 | 3209 | BUG_ON(!hugetlb_fault_mutex_table); |
8382d914 DB |
3210 | |
3211 | for (i = 0; i < num_fault_mutexes; i++) | |
c672c7f2 | 3212 | mutex_init(&hugetlb_fault_mutex_table[i]); |
a3437870 NA |
3213 | return 0; |
3214 | } | |
3e89e1c5 | 3215 | subsys_initcall(hugetlb_init); |
a3437870 | 3216 | |
ae94da89 MK |
3217 | /* Overwritten by architectures with more huge page sizes */ |
3218 | bool __init __attribute((weak)) arch_hugetlb_valid_size(unsigned long size) | |
9fee021d | 3219 | { |
ae94da89 | 3220 | return size == HPAGE_SIZE; |
9fee021d VT |
3221 | } |
3222 | ||
d00181b9 | 3223 | void __init hugetlb_add_hstate(unsigned int order) |
a3437870 NA |
3224 | { |
3225 | struct hstate *h; | |
8faa8b07 AK |
3226 | unsigned long i; |
3227 | ||
a3437870 | 3228 | if (size_to_hstate(PAGE_SIZE << order)) { |
a3437870 NA |
3229 | return; |
3230 | } | |
47d38344 | 3231 | BUG_ON(hugetlb_max_hstate >= HUGE_MAX_HSTATE); |
a3437870 | 3232 | BUG_ON(order == 0); |
47d38344 | 3233 | h = &hstates[hugetlb_max_hstate++]; |
29383967 | 3234 | mutex_init(&h->resize_lock); |
a3437870 | 3235 | h->order = order; |
aca78307 | 3236 | h->mask = ~(huge_page_size(h) - 1); |
8faa8b07 AK |
3237 | for (i = 0; i < MAX_NUMNODES; ++i) |
3238 | INIT_LIST_HEAD(&h->hugepage_freelists[i]); | |
0edaecfa | 3239 | INIT_LIST_HEAD(&h->hugepage_activelist); |
54f18d35 AM |
3240 | h->next_nid_to_alloc = first_memory_node; |
3241 | h->next_nid_to_free = first_memory_node; | |
a3437870 NA |
3242 | snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB", |
3243 | huge_page_size(h)/1024); | |
8faa8b07 | 3244 | |
a3437870 NA |
3245 | parsed_hstate = h; |
3246 | } | |
3247 | ||
282f4214 MK |
3248 | /* |
3249 | * hugepages command line processing | |
3250 | * hugepages normally follows a valid hugepagsz or default_hugepagsz | |
3251 | * specification. If not, ignore the hugepages value. hugepages can also | |
3252 | * be the first huge page command line option in which case it implicitly | |
3253 | * specifies the number of huge pages for the default size. | |
3254 | */ | |
3255 | static int __init hugepages_setup(char *s) | |
a3437870 NA |
3256 | { |
3257 | unsigned long *mhp; | |
8faa8b07 | 3258 | static unsigned long *last_mhp; |
a3437870 | 3259 | |
9fee021d | 3260 | if (!parsed_valid_hugepagesz) { |
282f4214 | 3261 | pr_warn("HugeTLB: hugepages=%s does not follow a valid hugepagesz, ignoring\n", s); |
9fee021d | 3262 | parsed_valid_hugepagesz = true; |
282f4214 | 3263 | return 0; |
9fee021d | 3264 | } |
282f4214 | 3265 | |
a3437870 | 3266 | /* |
282f4214 MK |
3267 | * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter |
3268 | * yet, so this hugepages= parameter goes to the "default hstate". | |
3269 | * Otherwise, it goes with the previously parsed hugepagesz or | |
3270 | * default_hugepagesz. | |
a3437870 | 3271 | */ |
9fee021d | 3272 | else if (!hugetlb_max_hstate) |
a3437870 NA |
3273 | mhp = &default_hstate_max_huge_pages; |
3274 | else | |
3275 | mhp = &parsed_hstate->max_huge_pages; | |
3276 | ||
8faa8b07 | 3277 | if (mhp == last_mhp) { |
282f4214 MK |
3278 | pr_warn("HugeTLB: hugepages= specified twice without interleaving hugepagesz=, ignoring hugepages=%s\n", s); |
3279 | return 0; | |
8faa8b07 AK |
3280 | } |
3281 | ||
a3437870 NA |
3282 | if (sscanf(s, "%lu", mhp) <= 0) |
3283 | *mhp = 0; | |
3284 | ||
8faa8b07 AK |
3285 | /* |
3286 | * Global state is always initialized later in hugetlb_init. | |
04adbc3f | 3287 | * But we need to allocate gigantic hstates here early to still |
8faa8b07 AK |
3288 | * use the bootmem allocator. |
3289 | */ | |
04adbc3f | 3290 | if (hugetlb_max_hstate && hstate_is_gigantic(parsed_hstate)) |
8faa8b07 AK |
3291 | hugetlb_hstate_alloc_pages(parsed_hstate); |
3292 | ||
3293 | last_mhp = mhp; | |
3294 | ||
a3437870 NA |
3295 | return 1; |
3296 | } | |
282f4214 | 3297 | __setup("hugepages=", hugepages_setup); |
e11bfbfc | 3298 | |
282f4214 MK |
3299 | /* |
3300 | * hugepagesz command line processing | |
3301 | * A specific huge page size can only be specified once with hugepagesz. | |
3302 | * hugepagesz is followed by hugepages on the command line. The global | |
3303 | * variable 'parsed_valid_hugepagesz' is used to determine if prior | |
3304 | * hugepagesz argument was valid. | |
3305 | */ | |
359f2544 | 3306 | static int __init hugepagesz_setup(char *s) |
e11bfbfc | 3307 | { |
359f2544 | 3308 | unsigned long size; |
282f4214 MK |
3309 | struct hstate *h; |
3310 | ||
3311 | parsed_valid_hugepagesz = false; | |
359f2544 MK |
3312 | size = (unsigned long)memparse(s, NULL); |
3313 | ||
3314 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3315 | pr_err("HugeTLB: unsupported hugepagesz=%s\n", s); |
359f2544 MK |
3316 | return 0; |
3317 | } | |
3318 | ||
282f4214 MK |
3319 | h = size_to_hstate(size); |
3320 | if (h) { | |
3321 | /* | |
3322 | * hstate for this size already exists. This is normally | |
3323 | * an error, but is allowed if the existing hstate is the | |
3324 | * default hstate. More specifically, it is only allowed if | |
3325 | * the number of huge pages for the default hstate was not | |
3326 | * previously specified. | |
3327 | */ | |
3328 | if (!parsed_default_hugepagesz || h != &default_hstate || | |
3329 | default_hstate.max_huge_pages) { | |
3330 | pr_warn("HugeTLB: hugepagesz=%s specified twice, ignoring\n", s); | |
3331 | return 0; | |
3332 | } | |
3333 | ||
3334 | /* | |
3335 | * No need to call hugetlb_add_hstate() as hstate already | |
3336 | * exists. But, do set parsed_hstate so that a following | |
3337 | * hugepages= parameter will be applied to this hstate. | |
3338 | */ | |
3339 | parsed_hstate = h; | |
3340 | parsed_valid_hugepagesz = true; | |
3341 | return 1; | |
38237830 MK |
3342 | } |
3343 | ||
359f2544 | 3344 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
282f4214 | 3345 | parsed_valid_hugepagesz = true; |
e11bfbfc NP |
3346 | return 1; |
3347 | } | |
359f2544 MK |
3348 | __setup("hugepagesz=", hugepagesz_setup); |
3349 | ||
282f4214 MK |
3350 | /* |
3351 | * default_hugepagesz command line input | |
3352 | * Only one instance of default_hugepagesz allowed on command line. | |
3353 | */ | |
ae94da89 | 3354 | static int __init default_hugepagesz_setup(char *s) |
e11bfbfc | 3355 | { |
ae94da89 MK |
3356 | unsigned long size; |
3357 | ||
282f4214 | 3358 | parsed_valid_hugepagesz = false; |
282f4214 MK |
3359 | if (parsed_default_hugepagesz) { |
3360 | pr_err("HugeTLB: default_hugepagesz previously specified, ignoring %s\n", s); | |
3361 | return 0; | |
3362 | } | |
3363 | ||
ae94da89 MK |
3364 | size = (unsigned long)memparse(s, NULL); |
3365 | ||
3366 | if (!arch_hugetlb_valid_size(size)) { | |
282f4214 | 3367 | pr_err("HugeTLB: unsupported default_hugepagesz=%s\n", s); |
ae94da89 MK |
3368 | return 0; |
3369 | } | |
3370 | ||
282f4214 MK |
3371 | hugetlb_add_hstate(ilog2(size) - PAGE_SHIFT); |
3372 | parsed_valid_hugepagesz = true; | |
3373 | parsed_default_hugepagesz = true; | |
3374 | default_hstate_idx = hstate_index(size_to_hstate(size)); | |
3375 | ||
3376 | /* | |
3377 | * The number of default huge pages (for this size) could have been | |
3378 | * specified as the first hugetlb parameter: hugepages=X. If so, | |
3379 | * then default_hstate_max_huge_pages is set. If the default huge | |
3380 | * page size is gigantic (>= MAX_ORDER), then the pages must be | |
3381 | * allocated here from bootmem allocator. | |
3382 | */ | |
3383 | if (default_hstate_max_huge_pages) { | |
3384 | default_hstate.max_huge_pages = default_hstate_max_huge_pages; | |
3385 | if (hstate_is_gigantic(&default_hstate)) | |
3386 | hugetlb_hstate_alloc_pages(&default_hstate); | |
3387 | default_hstate_max_huge_pages = 0; | |
3388 | } | |
3389 | ||
e11bfbfc NP |
3390 | return 1; |
3391 | } | |
ae94da89 | 3392 | __setup("default_hugepagesz=", default_hugepagesz_setup); |
a3437870 | 3393 | |
8ca39e68 | 3394 | static unsigned int allowed_mems_nr(struct hstate *h) |
8a213460 NA |
3395 | { |
3396 | int node; | |
3397 | unsigned int nr = 0; | |
8ca39e68 MS |
3398 | nodemask_t *mpol_allowed; |
3399 | unsigned int *array = h->free_huge_pages_node; | |
3400 | gfp_t gfp_mask = htlb_alloc_mask(h); | |
3401 | ||
3402 | mpol_allowed = policy_nodemask_current(gfp_mask); | |
8a213460 | 3403 | |
8ca39e68 | 3404 | for_each_node_mask(node, cpuset_current_mems_allowed) { |
c93b0a99 | 3405 | if (!mpol_allowed || node_isset(node, *mpol_allowed)) |
8ca39e68 MS |
3406 | nr += array[node]; |
3407 | } | |
8a213460 NA |
3408 | |
3409 | return nr; | |
3410 | } | |
3411 | ||
3412 | #ifdef CONFIG_SYSCTL | |
17743798 MS |
3413 | static int proc_hugetlb_doulongvec_minmax(struct ctl_table *table, int write, |
3414 | void *buffer, size_t *length, | |
3415 | loff_t *ppos, unsigned long *out) | |
3416 | { | |
3417 | struct ctl_table dup_table; | |
3418 | ||
3419 | /* | |
3420 | * In order to avoid races with __do_proc_doulongvec_minmax(), we | |
3421 | * can duplicate the @table and alter the duplicate of it. | |
3422 | */ | |
3423 | dup_table = *table; | |
3424 | dup_table.data = out; | |
3425 | ||
3426 | return proc_doulongvec_minmax(&dup_table, write, buffer, length, ppos); | |
3427 | } | |
3428 | ||
06808b08 LS |
3429 | static int hugetlb_sysctl_handler_common(bool obey_mempolicy, |
3430 | struct ctl_table *table, int write, | |
32927393 | 3431 | void *buffer, size_t *length, loff_t *ppos) |
1da177e4 | 3432 | { |
e5ff2159 | 3433 | struct hstate *h = &default_hstate; |
238d3c13 | 3434 | unsigned long tmp = h->max_huge_pages; |
08d4a246 | 3435 | int ret; |
e5ff2159 | 3436 | |
457c1b27 | 3437 | if (!hugepages_supported()) |
86613628 | 3438 | return -EOPNOTSUPP; |
457c1b27 | 3439 | |
17743798 MS |
3440 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3441 | &tmp); | |
08d4a246 MH |
3442 | if (ret) |
3443 | goto out; | |
e5ff2159 | 3444 | |
238d3c13 DR |
3445 | if (write) |
3446 | ret = __nr_hugepages_store_common(obey_mempolicy, h, | |
3447 | NUMA_NO_NODE, tmp, *length); | |
08d4a246 MH |
3448 | out: |
3449 | return ret; | |
1da177e4 | 3450 | } |
396faf03 | 3451 | |
06808b08 | 3452 | int hugetlb_sysctl_handler(struct ctl_table *table, int write, |
32927393 | 3453 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3454 | { |
3455 | ||
3456 | return hugetlb_sysctl_handler_common(false, table, write, | |
3457 | buffer, length, ppos); | |
3458 | } | |
3459 | ||
3460 | #ifdef CONFIG_NUMA | |
3461 | int hugetlb_mempolicy_sysctl_handler(struct ctl_table *table, int write, | |
32927393 | 3462 | void *buffer, size_t *length, loff_t *ppos) |
06808b08 LS |
3463 | { |
3464 | return hugetlb_sysctl_handler_common(true, table, write, | |
3465 | buffer, length, ppos); | |
3466 | } | |
3467 | #endif /* CONFIG_NUMA */ | |
3468 | ||
a3d0c6aa | 3469 | int hugetlb_overcommit_handler(struct ctl_table *table, int write, |
32927393 | 3470 | void *buffer, size_t *length, loff_t *ppos) |
a3d0c6aa | 3471 | { |
a5516438 | 3472 | struct hstate *h = &default_hstate; |
e5ff2159 | 3473 | unsigned long tmp; |
08d4a246 | 3474 | int ret; |
e5ff2159 | 3475 | |
457c1b27 | 3476 | if (!hugepages_supported()) |
86613628 | 3477 | return -EOPNOTSUPP; |
457c1b27 | 3478 | |
c033a93c | 3479 | tmp = h->nr_overcommit_huge_pages; |
e5ff2159 | 3480 | |
bae7f4ae | 3481 | if (write && hstate_is_gigantic(h)) |
adbe8726 EM |
3482 | return -EINVAL; |
3483 | ||
17743798 MS |
3484 | ret = proc_hugetlb_doulongvec_minmax(table, write, buffer, length, ppos, |
3485 | &tmp); | |
08d4a246 MH |
3486 | if (ret) |
3487 | goto out; | |
e5ff2159 AK |
3488 | |
3489 | if (write) { | |
db71ef79 | 3490 | spin_lock_irq(&hugetlb_lock); |
e5ff2159 | 3491 | h->nr_overcommit_huge_pages = tmp; |
db71ef79 | 3492 | spin_unlock_irq(&hugetlb_lock); |
e5ff2159 | 3493 | } |
08d4a246 MH |
3494 | out: |
3495 | return ret; | |
a3d0c6aa NA |
3496 | } |
3497 | ||
1da177e4 LT |
3498 | #endif /* CONFIG_SYSCTL */ |
3499 | ||
e1759c21 | 3500 | void hugetlb_report_meminfo(struct seq_file *m) |
1da177e4 | 3501 | { |
fcb2b0c5 RG |
3502 | struct hstate *h; |
3503 | unsigned long total = 0; | |
3504 | ||
457c1b27 NA |
3505 | if (!hugepages_supported()) |
3506 | return; | |
fcb2b0c5 RG |
3507 | |
3508 | for_each_hstate(h) { | |
3509 | unsigned long count = h->nr_huge_pages; | |
3510 | ||
aca78307 | 3511 | total += huge_page_size(h) * count; |
fcb2b0c5 RG |
3512 | |
3513 | if (h == &default_hstate) | |
3514 | seq_printf(m, | |
3515 | "HugePages_Total: %5lu\n" | |
3516 | "HugePages_Free: %5lu\n" | |
3517 | "HugePages_Rsvd: %5lu\n" | |
3518 | "HugePages_Surp: %5lu\n" | |
3519 | "Hugepagesize: %8lu kB\n", | |
3520 | count, | |
3521 | h->free_huge_pages, | |
3522 | h->resv_huge_pages, | |
3523 | h->surplus_huge_pages, | |
aca78307 | 3524 | huge_page_size(h) / SZ_1K); |
fcb2b0c5 RG |
3525 | } |
3526 | ||
aca78307 | 3527 | seq_printf(m, "Hugetlb: %8lu kB\n", total / SZ_1K); |
1da177e4 LT |
3528 | } |
3529 | ||
7981593b | 3530 | int hugetlb_report_node_meminfo(char *buf, int len, int nid) |
1da177e4 | 3531 | { |
a5516438 | 3532 | struct hstate *h = &default_hstate; |
7981593b | 3533 | |
457c1b27 NA |
3534 | if (!hugepages_supported()) |
3535 | return 0; | |
7981593b JP |
3536 | |
3537 | return sysfs_emit_at(buf, len, | |
3538 | "Node %d HugePages_Total: %5u\n" | |
3539 | "Node %d HugePages_Free: %5u\n" | |
3540 | "Node %d HugePages_Surp: %5u\n", | |
3541 | nid, h->nr_huge_pages_node[nid], | |
3542 | nid, h->free_huge_pages_node[nid], | |
3543 | nid, h->surplus_huge_pages_node[nid]); | |
1da177e4 LT |
3544 | } |
3545 | ||
949f7ec5 DR |
3546 | void hugetlb_show_meminfo(void) |
3547 | { | |
3548 | struct hstate *h; | |
3549 | int nid; | |
3550 | ||
457c1b27 NA |
3551 | if (!hugepages_supported()) |
3552 | return; | |
3553 | ||
949f7ec5 DR |
3554 | for_each_node_state(nid, N_MEMORY) |
3555 | for_each_hstate(h) | |
3556 | pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n", | |
3557 | nid, | |
3558 | h->nr_huge_pages_node[nid], | |
3559 | h->free_huge_pages_node[nid], | |
3560 | h->surplus_huge_pages_node[nid], | |
aca78307 | 3561 | huge_page_size(h) / SZ_1K); |
949f7ec5 DR |
3562 | } |
3563 | ||
5d317b2b NH |
3564 | void hugetlb_report_usage(struct seq_file *m, struct mm_struct *mm) |
3565 | { | |
3566 | seq_printf(m, "HugetlbPages:\t%8lu kB\n", | |
3567 | atomic_long_read(&mm->hugetlb_usage) << (PAGE_SHIFT - 10)); | |
3568 | } | |
3569 | ||
1da177e4 LT |
3570 | /* Return the number pages of memory we physically have, in PAGE_SIZE units. */ |
3571 | unsigned long hugetlb_total_pages(void) | |
3572 | { | |
d0028588 WL |
3573 | struct hstate *h; |
3574 | unsigned long nr_total_pages = 0; | |
3575 | ||
3576 | for_each_hstate(h) | |
3577 | nr_total_pages += h->nr_huge_pages * pages_per_huge_page(h); | |
3578 | return nr_total_pages; | |
1da177e4 | 3579 | } |
1da177e4 | 3580 | |
a5516438 | 3581 | static int hugetlb_acct_memory(struct hstate *h, long delta) |
fc1b8a73 MG |
3582 | { |
3583 | int ret = -ENOMEM; | |
3584 | ||
0aa7f354 ML |
3585 | if (!delta) |
3586 | return 0; | |
3587 | ||
db71ef79 | 3588 | spin_lock_irq(&hugetlb_lock); |
fc1b8a73 MG |
3589 | /* |
3590 | * When cpuset is configured, it breaks the strict hugetlb page | |
3591 | * reservation as the accounting is done on a global variable. Such | |
3592 | * reservation is completely rubbish in the presence of cpuset because | |
3593 | * the reservation is not checked against page availability for the | |
3594 | * current cpuset. Application can still potentially OOM'ed by kernel | |
3595 | * with lack of free htlb page in cpuset that the task is in. | |
3596 | * Attempt to enforce strict accounting with cpuset is almost | |
3597 | * impossible (or too ugly) because cpuset is too fluid that | |
3598 | * task or memory node can be dynamically moved between cpusets. | |
3599 | * | |
3600 | * The change of semantics for shared hugetlb mapping with cpuset is | |
3601 | * undesirable. However, in order to preserve some of the semantics, | |
3602 | * we fall back to check against current free page availability as | |
3603 | * a best attempt and hopefully to minimize the impact of changing | |
3604 | * semantics that cpuset has. | |
8ca39e68 MS |
3605 | * |
3606 | * Apart from cpuset, we also have memory policy mechanism that | |
3607 | * also determines from which node the kernel will allocate memory | |
3608 | * in a NUMA system. So similar to cpuset, we also should consider | |
3609 | * the memory policy of the current task. Similar to the description | |
3610 | * above. | |
fc1b8a73 MG |
3611 | */ |
3612 | if (delta > 0) { | |
a5516438 | 3613 | if (gather_surplus_pages(h, delta) < 0) |
fc1b8a73 MG |
3614 | goto out; |
3615 | ||
8ca39e68 | 3616 | if (delta > allowed_mems_nr(h)) { |
a5516438 | 3617 | return_unused_surplus_pages(h, delta); |
fc1b8a73 MG |
3618 | goto out; |
3619 | } | |
3620 | } | |
3621 | ||
3622 | ret = 0; | |
3623 | if (delta < 0) | |
a5516438 | 3624 | return_unused_surplus_pages(h, (unsigned long) -delta); |
fc1b8a73 MG |
3625 | |
3626 | out: | |
db71ef79 | 3627 | spin_unlock_irq(&hugetlb_lock); |
fc1b8a73 MG |
3628 | return ret; |
3629 | } | |
3630 | ||
84afd99b AW |
3631 | static void hugetlb_vm_op_open(struct vm_area_struct *vma) |
3632 | { | |
f522c3ac | 3633 | struct resv_map *resv = vma_resv_map(vma); |
84afd99b AW |
3634 | |
3635 | /* | |
3636 | * This new VMA should share its siblings reservation map if present. | |
3637 | * The VMA will only ever have a valid reservation map pointer where | |
3638 | * it is being copied for another still existing VMA. As that VMA | |
25985edc | 3639 | * has a reference to the reservation map it cannot disappear until |
84afd99b AW |
3640 | * after this open call completes. It is therefore safe to take a |
3641 | * new reference here without additional locking. | |
3642 | */ | |
4e35f483 | 3643 | if (resv && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
f522c3ac | 3644 | kref_get(&resv->refs); |
84afd99b AW |
3645 | } |
3646 | ||
a1e78772 MG |
3647 | static void hugetlb_vm_op_close(struct vm_area_struct *vma) |
3648 | { | |
a5516438 | 3649 | struct hstate *h = hstate_vma(vma); |
f522c3ac | 3650 | struct resv_map *resv = vma_resv_map(vma); |
90481622 | 3651 | struct hugepage_subpool *spool = subpool_vma(vma); |
4e35f483 | 3652 | unsigned long reserve, start, end; |
1c5ecae3 | 3653 | long gbl_reserve; |
84afd99b | 3654 | |
4e35f483 JK |
3655 | if (!resv || !is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
3656 | return; | |
84afd99b | 3657 | |
4e35f483 JK |
3658 | start = vma_hugecache_offset(h, vma, vma->vm_start); |
3659 | end = vma_hugecache_offset(h, vma, vma->vm_end); | |
84afd99b | 3660 | |
4e35f483 | 3661 | reserve = (end - start) - region_count(resv, start, end); |
e9fe92ae | 3662 | hugetlb_cgroup_uncharge_counter(resv, start, end); |
4e35f483 | 3663 | if (reserve) { |
1c5ecae3 MK |
3664 | /* |
3665 | * Decrement reserve counts. The global reserve count may be | |
3666 | * adjusted if the subpool has a minimum size. | |
3667 | */ | |
3668 | gbl_reserve = hugepage_subpool_put_pages(spool, reserve); | |
3669 | hugetlb_acct_memory(h, -gbl_reserve); | |
84afd99b | 3670 | } |
e9fe92ae MA |
3671 | |
3672 | kref_put(&resv->refs, resv_map_release); | |
a1e78772 MG |
3673 | } |
3674 | ||
31383c68 DW |
3675 | static int hugetlb_vm_op_split(struct vm_area_struct *vma, unsigned long addr) |
3676 | { | |
3677 | if (addr & ~(huge_page_mask(hstate_vma(vma)))) | |
3678 | return -EINVAL; | |
3679 | return 0; | |
3680 | } | |
3681 | ||
05ea8860 DW |
3682 | static unsigned long hugetlb_vm_op_pagesize(struct vm_area_struct *vma) |
3683 | { | |
aca78307 | 3684 | return huge_page_size(hstate_vma(vma)); |
05ea8860 DW |
3685 | } |
3686 | ||
1da177e4 LT |
3687 | /* |
3688 | * We cannot handle pagefaults against hugetlb pages at all. They cause | |
3689 | * handle_mm_fault() to try to instantiate regular-sized pages in the | |
6c26d310 | 3690 | * hugepage VMA. do_page_fault() is supposed to trap this, so BUG is we get |
1da177e4 LT |
3691 | * this far. |
3692 | */ | |
b3ec9f33 | 3693 | static vm_fault_t hugetlb_vm_op_fault(struct vm_fault *vmf) |
1da177e4 LT |
3694 | { |
3695 | BUG(); | |
d0217ac0 | 3696 | return 0; |
1da177e4 LT |
3697 | } |
3698 | ||
eec3636a JC |
3699 | /* |
3700 | * When a new function is introduced to vm_operations_struct and added | |
3701 | * to hugetlb_vm_ops, please consider adding the function to shm_vm_ops. | |
3702 | * This is because under System V memory model, mappings created via | |
3703 | * shmget/shmat with "huge page" specified are backed by hugetlbfs files, | |
3704 | * their original vm_ops are overwritten with shm_vm_ops. | |
3705 | */ | |
f0f37e2f | 3706 | const struct vm_operations_struct hugetlb_vm_ops = { |
d0217ac0 | 3707 | .fault = hugetlb_vm_op_fault, |
84afd99b | 3708 | .open = hugetlb_vm_op_open, |
a1e78772 | 3709 | .close = hugetlb_vm_op_close, |
dd3b614f | 3710 | .may_split = hugetlb_vm_op_split, |
05ea8860 | 3711 | .pagesize = hugetlb_vm_op_pagesize, |
1da177e4 LT |
3712 | }; |
3713 | ||
1e8f889b DG |
3714 | static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page, |
3715 | int writable) | |
63551ae0 DG |
3716 | { |
3717 | pte_t entry; | |
3718 | ||
1e8f889b | 3719 | if (writable) { |
106c992a GS |
3720 | entry = huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page, |
3721 | vma->vm_page_prot))); | |
63551ae0 | 3722 | } else { |
106c992a GS |
3723 | entry = huge_pte_wrprotect(mk_huge_pte(page, |
3724 | vma->vm_page_prot)); | |
63551ae0 DG |
3725 | } |
3726 | entry = pte_mkyoung(entry); | |
3727 | entry = pte_mkhuge(entry); | |
d9ed9faa | 3728 | entry = arch_make_huge_pte(entry, vma, page, writable); |
63551ae0 DG |
3729 | |
3730 | return entry; | |
3731 | } | |
3732 | ||
1e8f889b DG |
3733 | static void set_huge_ptep_writable(struct vm_area_struct *vma, |
3734 | unsigned long address, pte_t *ptep) | |
3735 | { | |
3736 | pte_t entry; | |
3737 | ||
106c992a | 3738 | entry = huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep))); |
32f84528 | 3739 | if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) |
4b3073e1 | 3740 | update_mmu_cache(vma, address, ptep); |
1e8f889b DG |
3741 | } |
3742 | ||
d5ed7444 | 3743 | bool is_hugetlb_entry_migration(pte_t pte) |
4a705fef NH |
3744 | { |
3745 | swp_entry_t swp; | |
3746 | ||
3747 | if (huge_pte_none(pte) || pte_present(pte)) | |
d5ed7444 | 3748 | return false; |
4a705fef | 3749 | swp = pte_to_swp_entry(pte); |
d79d176a | 3750 | if (is_migration_entry(swp)) |
d5ed7444 | 3751 | return true; |
4a705fef | 3752 | else |
d5ed7444 | 3753 | return false; |
4a705fef NH |
3754 | } |
3755 | ||
3e5c3600 | 3756 | static bool is_hugetlb_entry_hwpoisoned(pte_t pte) |
4a705fef NH |
3757 | { |
3758 | swp_entry_t swp; | |
3759 | ||
3760 | if (huge_pte_none(pte) || pte_present(pte)) | |
3e5c3600 | 3761 | return false; |
4a705fef | 3762 | swp = pte_to_swp_entry(pte); |
d79d176a | 3763 | if (is_hwpoison_entry(swp)) |
3e5c3600 | 3764 | return true; |
4a705fef | 3765 | else |
3e5c3600 | 3766 | return false; |
4a705fef | 3767 | } |
1e8f889b | 3768 | |
4eae4efa PX |
3769 | static void |
3770 | hugetlb_install_page(struct vm_area_struct *vma, pte_t *ptep, unsigned long addr, | |
3771 | struct page *new_page) | |
3772 | { | |
3773 | __SetPageUptodate(new_page); | |
3774 | set_huge_pte_at(vma->vm_mm, addr, ptep, make_huge_pte(vma, new_page, 1)); | |
3775 | hugepage_add_new_anon_rmap(new_page, vma, addr); | |
3776 | hugetlb_count_add(pages_per_huge_page(hstate_vma(vma)), vma->vm_mm); | |
3777 | ClearHPageRestoreReserve(new_page); | |
3778 | SetHPageMigratable(new_page); | |
3779 | } | |
3780 | ||
63551ae0 DG |
3781 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
3782 | struct vm_area_struct *vma) | |
3783 | { | |
5e41540c | 3784 | pte_t *src_pte, *dst_pte, entry, dst_entry; |
63551ae0 | 3785 | struct page *ptepage; |
1c59827d | 3786 | unsigned long addr; |
ca6eb14d | 3787 | bool cow = is_cow_mapping(vma->vm_flags); |
a5516438 AK |
3788 | struct hstate *h = hstate_vma(vma); |
3789 | unsigned long sz = huge_page_size(h); | |
4eae4efa | 3790 | unsigned long npages = pages_per_huge_page(h); |
c0d0381a | 3791 | struct address_space *mapping = vma->vm_file->f_mapping; |
ac46d4f3 | 3792 | struct mmu_notifier_range range; |
e8569dd2 | 3793 | int ret = 0; |
1e8f889b | 3794 | |
ac46d4f3 | 3795 | if (cow) { |
7269f999 | 3796 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, src, |
6f4f13e8 | 3797 | vma->vm_start, |
ac46d4f3 JG |
3798 | vma->vm_end); |
3799 | mmu_notifier_invalidate_range_start(&range); | |
c0d0381a MK |
3800 | } else { |
3801 | /* | |
3802 | * For shared mappings i_mmap_rwsem must be held to call | |
3803 | * huge_pte_alloc, otherwise the returned ptep could go | |
3804 | * away if part of a shared pmd and another thread calls | |
3805 | * huge_pmd_unshare. | |
3806 | */ | |
3807 | i_mmap_lock_read(mapping); | |
ac46d4f3 | 3808 | } |
e8569dd2 | 3809 | |
a5516438 | 3810 | for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) { |
cb900f41 | 3811 | spinlock_t *src_ptl, *dst_ptl; |
7868a208 | 3812 | src_pte = huge_pte_offset(src, addr, sz); |
c74df32c HD |
3813 | if (!src_pte) |
3814 | continue; | |
aec44e0f | 3815 | dst_pte = huge_pte_alloc(dst, vma, addr, sz); |
e8569dd2 AS |
3816 | if (!dst_pte) { |
3817 | ret = -ENOMEM; | |
3818 | break; | |
3819 | } | |
c5c99429 | 3820 | |
5e41540c MK |
3821 | /* |
3822 | * If the pagetables are shared don't copy or take references. | |
3823 | * dst_pte == src_pte is the common case of src/dest sharing. | |
3824 | * | |
3825 | * However, src could have 'unshared' and dst shares with | |
3826 | * another vma. If dst_pte !none, this implies sharing. | |
3827 | * Check here before taking page table lock, and once again | |
3828 | * after taking the lock below. | |
3829 | */ | |
3830 | dst_entry = huge_ptep_get(dst_pte); | |
3831 | if ((dst_pte == src_pte) || !huge_pte_none(dst_entry)) | |
c5c99429 LW |
3832 | continue; |
3833 | ||
cb900f41 KS |
3834 | dst_ptl = huge_pte_lock(h, dst, dst_pte); |
3835 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3836 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
4a705fef | 3837 | entry = huge_ptep_get(src_pte); |
5e41540c | 3838 | dst_entry = huge_ptep_get(dst_pte); |
4eae4efa | 3839 | again: |
5e41540c MK |
3840 | if (huge_pte_none(entry) || !huge_pte_none(dst_entry)) { |
3841 | /* | |
3842 | * Skip if src entry none. Also, skip in the | |
3843 | * unlikely case dst entry !none as this implies | |
3844 | * sharing with another vma. | |
3845 | */ | |
4a705fef NH |
3846 | ; |
3847 | } else if (unlikely(is_hugetlb_entry_migration(entry) || | |
3848 | is_hugetlb_entry_hwpoisoned(entry))) { | |
3849 | swp_entry_t swp_entry = pte_to_swp_entry(entry); | |
3850 | ||
3851 | if (is_write_migration_entry(swp_entry) && cow) { | |
3852 | /* | |
3853 | * COW mappings require pages in both | |
3854 | * parent and child to be set to read. | |
3855 | */ | |
3856 | make_migration_entry_read(&swp_entry); | |
3857 | entry = swp_entry_to_pte(swp_entry); | |
e5251fd4 PA |
3858 | set_huge_swap_pte_at(src, addr, src_pte, |
3859 | entry, sz); | |
4a705fef | 3860 | } |
e5251fd4 | 3861 | set_huge_swap_pte_at(dst, addr, dst_pte, entry, sz); |
4a705fef | 3862 | } else { |
4eae4efa PX |
3863 | entry = huge_ptep_get(src_pte); |
3864 | ptepage = pte_page(entry); | |
3865 | get_page(ptepage); | |
3866 | ||
3867 | /* | |
3868 | * This is a rare case where we see pinned hugetlb | |
3869 | * pages while they're prone to COW. We need to do the | |
3870 | * COW earlier during fork. | |
3871 | * | |
3872 | * When pre-allocating the page or copying data, we | |
3873 | * need to be without the pgtable locks since we could | |
3874 | * sleep during the process. | |
3875 | */ | |
3876 | if (unlikely(page_needs_cow_for_dma(vma, ptepage))) { | |
3877 | pte_t src_pte_old = entry; | |
3878 | struct page *new; | |
3879 | ||
3880 | spin_unlock(src_ptl); | |
3881 | spin_unlock(dst_ptl); | |
3882 | /* Do not use reserve as it's private owned */ | |
3883 | new = alloc_huge_page(vma, addr, 1); | |
3884 | if (IS_ERR(new)) { | |
3885 | put_page(ptepage); | |
3886 | ret = PTR_ERR(new); | |
3887 | break; | |
3888 | } | |
3889 | copy_user_huge_page(new, ptepage, addr, vma, | |
3890 | npages); | |
3891 | put_page(ptepage); | |
3892 | ||
3893 | /* Install the new huge page if src pte stable */ | |
3894 | dst_ptl = huge_pte_lock(h, dst, dst_pte); | |
3895 | src_ptl = huge_pte_lockptr(h, src, src_pte); | |
3896 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
3897 | entry = huge_ptep_get(src_pte); | |
3898 | if (!pte_same(src_pte_old, entry)) { | |
3899 | put_page(new); | |
3900 | /* dst_entry won't change as in child */ | |
3901 | goto again; | |
3902 | } | |
3903 | hugetlb_install_page(vma, dst_pte, addr, new); | |
3904 | spin_unlock(src_ptl); | |
3905 | spin_unlock(dst_ptl); | |
3906 | continue; | |
3907 | } | |
3908 | ||
34ee645e | 3909 | if (cow) { |
0f10851e JG |
3910 | /* |
3911 | * No need to notify as we are downgrading page | |
3912 | * table protection not changing it to point | |
3913 | * to a new page. | |
3914 | * | |
ad56b738 | 3915 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 3916 | */ |
7f2e9525 | 3917 | huge_ptep_set_wrprotect(src, addr, src_pte); |
34ee645e | 3918 | } |
4eae4efa | 3919 | |
53f9263b | 3920 | page_dup_rmap(ptepage, true); |
1c59827d | 3921 | set_huge_pte_at(dst, addr, dst_pte, entry); |
4eae4efa | 3922 | hugetlb_count_add(npages, dst); |
1c59827d | 3923 | } |
cb900f41 KS |
3924 | spin_unlock(src_ptl); |
3925 | spin_unlock(dst_ptl); | |
63551ae0 | 3926 | } |
63551ae0 | 3927 | |
e8569dd2 | 3928 | if (cow) |
ac46d4f3 | 3929 | mmu_notifier_invalidate_range_end(&range); |
c0d0381a MK |
3930 | else |
3931 | i_mmap_unlock_read(mapping); | |
e8569dd2 AS |
3932 | |
3933 | return ret; | |
63551ae0 DG |
3934 | } |
3935 | ||
24669e58 AK |
3936 | void __unmap_hugepage_range(struct mmu_gather *tlb, struct vm_area_struct *vma, |
3937 | unsigned long start, unsigned long end, | |
3938 | struct page *ref_page) | |
63551ae0 DG |
3939 | { |
3940 | struct mm_struct *mm = vma->vm_mm; | |
3941 | unsigned long address; | |
c7546f8f | 3942 | pte_t *ptep; |
63551ae0 | 3943 | pte_t pte; |
cb900f41 | 3944 | spinlock_t *ptl; |
63551ae0 | 3945 | struct page *page; |
a5516438 AK |
3946 | struct hstate *h = hstate_vma(vma); |
3947 | unsigned long sz = huge_page_size(h); | |
ac46d4f3 | 3948 | struct mmu_notifier_range range; |
a5516438 | 3949 | |
63551ae0 | 3950 | WARN_ON(!is_vm_hugetlb_page(vma)); |
a5516438 AK |
3951 | BUG_ON(start & ~huge_page_mask(h)); |
3952 | BUG_ON(end & ~huge_page_mask(h)); | |
63551ae0 | 3953 | |
07e32661 AK |
3954 | /* |
3955 | * This is a hugetlb vma, all the pte entries should point | |
3956 | * to huge page. | |
3957 | */ | |
ed6a7935 | 3958 | tlb_change_page_size(tlb, sz); |
24669e58 | 3959 | tlb_start_vma(tlb, vma); |
dff11abe MK |
3960 | |
3961 | /* | |
3962 | * If sharing possible, alert mmu notifiers of worst case. | |
3963 | */ | |
6f4f13e8 JG |
3964 | mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, mm, start, |
3965 | end); | |
ac46d4f3 JG |
3966 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
3967 | mmu_notifier_invalidate_range_start(&range); | |
569f48b8 | 3968 | address = start; |
569f48b8 | 3969 | for (; address < end; address += sz) { |
7868a208 | 3970 | ptep = huge_pte_offset(mm, address, sz); |
4c887265 | 3971 | if (!ptep) |
c7546f8f DG |
3972 | continue; |
3973 | ||
cb900f41 | 3974 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 3975 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
31d49da5 | 3976 | spin_unlock(ptl); |
dff11abe MK |
3977 | /* |
3978 | * We just unmapped a page of PMDs by clearing a PUD. | |
3979 | * The caller's TLB flush range should cover this area. | |
3980 | */ | |
31d49da5 AK |
3981 | continue; |
3982 | } | |
39dde65c | 3983 | |
6629326b | 3984 | pte = huge_ptep_get(ptep); |
31d49da5 AK |
3985 | if (huge_pte_none(pte)) { |
3986 | spin_unlock(ptl); | |
3987 | continue; | |
3988 | } | |
6629326b HD |
3989 | |
3990 | /* | |
9fbc1f63 NH |
3991 | * Migrating hugepage or HWPoisoned hugepage is already |
3992 | * unmapped and its refcount is dropped, so just clear pte here. | |
6629326b | 3993 | */ |
9fbc1f63 | 3994 | if (unlikely(!pte_present(pte))) { |
9386fac3 | 3995 | huge_pte_clear(mm, address, ptep, sz); |
31d49da5 AK |
3996 | spin_unlock(ptl); |
3997 | continue; | |
8c4894c6 | 3998 | } |
6629326b HD |
3999 | |
4000 | page = pte_page(pte); | |
04f2cbe3 MG |
4001 | /* |
4002 | * If a reference page is supplied, it is because a specific | |
4003 | * page is being unmapped, not a range. Ensure the page we | |
4004 | * are about to unmap is the actual page of interest. | |
4005 | */ | |
4006 | if (ref_page) { | |
31d49da5 AK |
4007 | if (page != ref_page) { |
4008 | spin_unlock(ptl); | |
4009 | continue; | |
4010 | } | |
04f2cbe3 MG |
4011 | /* |
4012 | * Mark the VMA as having unmapped its page so that | |
4013 | * future faults in this VMA will fail rather than | |
4014 | * looking like data was lost | |
4015 | */ | |
4016 | set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED); | |
4017 | } | |
4018 | ||
c7546f8f | 4019 | pte = huge_ptep_get_and_clear(mm, address, ptep); |
b528e4b6 | 4020 | tlb_remove_huge_tlb_entry(h, tlb, ptep, address); |
106c992a | 4021 | if (huge_pte_dirty(pte)) |
6649a386 | 4022 | set_page_dirty(page); |
9e81130b | 4023 | |
5d317b2b | 4024 | hugetlb_count_sub(pages_per_huge_page(h), mm); |
d281ee61 | 4025 | page_remove_rmap(page, true); |
31d49da5 | 4026 | |
cb900f41 | 4027 | spin_unlock(ptl); |
e77b0852 | 4028 | tlb_remove_page_size(tlb, page, huge_page_size(h)); |
31d49da5 AK |
4029 | /* |
4030 | * Bail out after unmapping reference page if supplied | |
4031 | */ | |
4032 | if (ref_page) | |
4033 | break; | |
fe1668ae | 4034 | } |
ac46d4f3 | 4035 | mmu_notifier_invalidate_range_end(&range); |
24669e58 | 4036 | tlb_end_vma(tlb, vma); |
1da177e4 | 4037 | } |
63551ae0 | 4038 | |
d833352a MG |
4039 | void __unmap_hugepage_range_final(struct mmu_gather *tlb, |
4040 | struct vm_area_struct *vma, unsigned long start, | |
4041 | unsigned long end, struct page *ref_page) | |
4042 | { | |
4043 | __unmap_hugepage_range(tlb, vma, start, end, ref_page); | |
4044 | ||
4045 | /* | |
4046 | * Clear this flag so that x86's huge_pmd_share page_table_shareable | |
4047 | * test will fail on a vma being torn down, and not grab a page table | |
4048 | * on its way out. We're lucky that the flag has such an appropriate | |
4049 | * name, and can in fact be safely cleared here. We could clear it | |
4050 | * before the __unmap_hugepage_range above, but all that's necessary | |
c8c06efa | 4051 | * is to clear it before releasing the i_mmap_rwsem. This works |
d833352a | 4052 | * because in the context this is called, the VMA is about to be |
c8c06efa | 4053 | * destroyed and the i_mmap_rwsem is held. |
d833352a MG |
4054 | */ |
4055 | vma->vm_flags &= ~VM_MAYSHARE; | |
4056 | } | |
4057 | ||
502717f4 | 4058 | void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start, |
04f2cbe3 | 4059 | unsigned long end, struct page *ref_page) |
502717f4 | 4060 | { |
24669e58 | 4061 | struct mmu_gather tlb; |
dff11abe | 4062 | |
a72afd87 | 4063 | tlb_gather_mmu(&tlb, vma->vm_mm); |
24669e58 | 4064 | __unmap_hugepage_range(&tlb, vma, start, end, ref_page); |
ae8eba8b | 4065 | tlb_finish_mmu(&tlb); |
502717f4 KC |
4066 | } |
4067 | ||
04f2cbe3 MG |
4068 | /* |
4069 | * This is called when the original mapper is failing to COW a MAP_PRIVATE | |
578b7725 | 4070 | * mapping it owns the reserve page for. The intention is to unmap the page |
04f2cbe3 MG |
4071 | * from other VMAs and let the children be SIGKILLed if they are faulting the |
4072 | * same region. | |
4073 | */ | |
2f4612af DB |
4074 | static void unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma, |
4075 | struct page *page, unsigned long address) | |
04f2cbe3 | 4076 | { |
7526674d | 4077 | struct hstate *h = hstate_vma(vma); |
04f2cbe3 MG |
4078 | struct vm_area_struct *iter_vma; |
4079 | struct address_space *mapping; | |
04f2cbe3 MG |
4080 | pgoff_t pgoff; |
4081 | ||
4082 | /* | |
4083 | * vm_pgoff is in PAGE_SIZE units, hence the different calculation | |
4084 | * from page cache lookup which is in HPAGE_SIZE units. | |
4085 | */ | |
7526674d | 4086 | address = address & huge_page_mask(h); |
36e4f20a MH |
4087 | pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + |
4088 | vma->vm_pgoff; | |
93c76a3d | 4089 | mapping = vma->vm_file->f_mapping; |
04f2cbe3 | 4090 | |
4eb2b1dc MG |
4091 | /* |
4092 | * Take the mapping lock for the duration of the table walk. As | |
4093 | * this mapping should be shared between all the VMAs, | |
4094 | * __unmap_hugepage_range() is called as the lock is already held | |
4095 | */ | |
83cde9e8 | 4096 | i_mmap_lock_write(mapping); |
6b2dbba8 | 4097 | vma_interval_tree_foreach(iter_vma, &mapping->i_mmap, pgoff, pgoff) { |
04f2cbe3 MG |
4098 | /* Do not unmap the current VMA */ |
4099 | if (iter_vma == vma) | |
4100 | continue; | |
4101 | ||
2f84a899 MG |
4102 | /* |
4103 | * Shared VMAs have their own reserves and do not affect | |
4104 | * MAP_PRIVATE accounting but it is possible that a shared | |
4105 | * VMA is using the same page so check and skip such VMAs. | |
4106 | */ | |
4107 | if (iter_vma->vm_flags & VM_MAYSHARE) | |
4108 | continue; | |
4109 | ||
04f2cbe3 MG |
4110 | /* |
4111 | * Unmap the page from other VMAs without their own reserves. | |
4112 | * They get marked to be SIGKILLed if they fault in these | |
4113 | * areas. This is because a future no-page fault on this VMA | |
4114 | * could insert a zeroed page instead of the data existing | |
4115 | * from the time of fork. This would look like data corruption | |
4116 | */ | |
4117 | if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER)) | |
24669e58 AK |
4118 | unmap_hugepage_range(iter_vma, address, |
4119 | address + huge_page_size(h), page); | |
04f2cbe3 | 4120 | } |
83cde9e8 | 4121 | i_mmap_unlock_write(mapping); |
04f2cbe3 MG |
4122 | } |
4123 | ||
0fe6e20b NH |
4124 | /* |
4125 | * Hugetlb_cow() should be called with page lock of the original hugepage held. | |
ef009b25 MH |
4126 | * Called with hugetlb_instantiation_mutex held and pte_page locked so we |
4127 | * cannot race with other handlers or page migration. | |
4128 | * Keep the pte_same checks anyway to make transition from the mutex easier. | |
0fe6e20b | 4129 | */ |
2b740303 | 4130 | static vm_fault_t hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma, |
974e6d66 | 4131 | unsigned long address, pte_t *ptep, |
3999f52e | 4132 | struct page *pagecache_page, spinlock_t *ptl) |
1e8f889b | 4133 | { |
3999f52e | 4134 | pte_t pte; |
a5516438 | 4135 | struct hstate *h = hstate_vma(vma); |
1e8f889b | 4136 | struct page *old_page, *new_page; |
2b740303 SJ |
4137 | int outside_reserve = 0; |
4138 | vm_fault_t ret = 0; | |
974e6d66 | 4139 | unsigned long haddr = address & huge_page_mask(h); |
ac46d4f3 | 4140 | struct mmu_notifier_range range; |
1e8f889b | 4141 | |
3999f52e | 4142 | pte = huge_ptep_get(ptep); |
1e8f889b DG |
4143 | old_page = pte_page(pte); |
4144 | ||
04f2cbe3 | 4145 | retry_avoidcopy: |
1e8f889b DG |
4146 | /* If no-one else is actually using this page, avoid the copy |
4147 | * and just make the page writable */ | |
37a2140d | 4148 | if (page_mapcount(old_page) == 1 && PageAnon(old_page)) { |
5a49973d | 4149 | page_move_anon_rmap(old_page, vma); |
5b7a1d40 | 4150 | set_huge_ptep_writable(vma, haddr, ptep); |
83c54070 | 4151 | return 0; |
1e8f889b DG |
4152 | } |
4153 | ||
04f2cbe3 MG |
4154 | /* |
4155 | * If the process that created a MAP_PRIVATE mapping is about to | |
4156 | * perform a COW due to a shared page count, attempt to satisfy | |
4157 | * the allocation without using the existing reserves. The pagecache | |
4158 | * page is used to determine if the reserve at this address was | |
4159 | * consumed or not. If reserves were used, a partial faulted mapping | |
4160 | * at the time of fork() could consume its reserves on COW instead | |
4161 | * of the full address range. | |
4162 | */ | |
5944d011 | 4163 | if (is_vma_resv_set(vma, HPAGE_RESV_OWNER) && |
04f2cbe3 MG |
4164 | old_page != pagecache_page) |
4165 | outside_reserve = 1; | |
4166 | ||
09cbfeaf | 4167 | get_page(old_page); |
b76c8cfb | 4168 | |
ad4404a2 DB |
4169 | /* |
4170 | * Drop page table lock as buddy allocator may be called. It will | |
4171 | * be acquired again before returning to the caller, as expected. | |
4172 | */ | |
cb900f41 | 4173 | spin_unlock(ptl); |
5b7a1d40 | 4174 | new_page = alloc_huge_page(vma, haddr, outside_reserve); |
1e8f889b | 4175 | |
2fc39cec | 4176 | if (IS_ERR(new_page)) { |
04f2cbe3 MG |
4177 | /* |
4178 | * If a process owning a MAP_PRIVATE mapping fails to COW, | |
4179 | * it is due to references held by a child and an insufficient | |
4180 | * huge page pool. To guarantee the original mappers | |
4181 | * reliability, unmap the page from child processes. The child | |
4182 | * may get SIGKILLed if it later faults. | |
4183 | */ | |
4184 | if (outside_reserve) { | |
e7dd91c4 MK |
4185 | struct address_space *mapping = vma->vm_file->f_mapping; |
4186 | pgoff_t idx; | |
4187 | u32 hash; | |
4188 | ||
09cbfeaf | 4189 | put_page(old_page); |
04f2cbe3 | 4190 | BUG_ON(huge_pte_none(pte)); |
e7dd91c4 MK |
4191 | /* |
4192 | * Drop hugetlb_fault_mutex and i_mmap_rwsem before | |
4193 | * unmapping. unmapping needs to hold i_mmap_rwsem | |
4194 | * in write mode. Dropping i_mmap_rwsem in read mode | |
4195 | * here is OK as COW mappings do not interact with | |
4196 | * PMD sharing. | |
4197 | * | |
4198 | * Reacquire both after unmap operation. | |
4199 | */ | |
4200 | idx = vma_hugecache_offset(h, vma, haddr); | |
4201 | hash = hugetlb_fault_mutex_hash(mapping, idx); | |
4202 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); | |
4203 | i_mmap_unlock_read(mapping); | |
4204 | ||
5b7a1d40 | 4205 | unmap_ref_private(mm, vma, old_page, haddr); |
e7dd91c4 MK |
4206 | |
4207 | i_mmap_lock_read(mapping); | |
4208 | mutex_lock(&hugetlb_fault_mutex_table[hash]); | |
2f4612af | 4209 | spin_lock(ptl); |
5b7a1d40 | 4210 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
2f4612af DB |
4211 | if (likely(ptep && |
4212 | pte_same(huge_ptep_get(ptep), pte))) | |
4213 | goto retry_avoidcopy; | |
4214 | /* | |
4215 | * race occurs while re-acquiring page table | |
4216 | * lock, and our job is done. | |
4217 | */ | |
4218 | return 0; | |
04f2cbe3 MG |
4219 | } |
4220 | ||
2b740303 | 4221 | ret = vmf_error(PTR_ERR(new_page)); |
ad4404a2 | 4222 | goto out_release_old; |
1e8f889b DG |
4223 | } |
4224 | ||
0fe6e20b NH |
4225 | /* |
4226 | * When the original hugepage is shared one, it does not have | |
4227 | * anon_vma prepared. | |
4228 | */ | |
44e2aa93 | 4229 | if (unlikely(anon_vma_prepare(vma))) { |
ad4404a2 DB |
4230 | ret = VM_FAULT_OOM; |
4231 | goto out_release_all; | |
44e2aa93 | 4232 | } |
0fe6e20b | 4233 | |
974e6d66 | 4234 | copy_user_huge_page(new_page, old_page, address, vma, |
47ad8475 | 4235 | pages_per_huge_page(h)); |
0ed361de | 4236 | __SetPageUptodate(new_page); |
1e8f889b | 4237 | |
7269f999 | 4238 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, haddr, |
6f4f13e8 | 4239 | haddr + huge_page_size(h)); |
ac46d4f3 | 4240 | mmu_notifier_invalidate_range_start(&range); |
ad4404a2 | 4241 | |
b76c8cfb | 4242 | /* |
cb900f41 | 4243 | * Retake the page table lock to check for racing updates |
b76c8cfb LW |
4244 | * before the page tables are altered |
4245 | */ | |
cb900f41 | 4246 | spin_lock(ptl); |
5b7a1d40 | 4247 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
a9af0c5d | 4248 | if (likely(ptep && pte_same(huge_ptep_get(ptep), pte))) { |
d6995da3 | 4249 | ClearHPageRestoreReserve(new_page); |
07443a85 | 4250 | |
1e8f889b | 4251 | /* Break COW */ |
5b7a1d40 | 4252 | huge_ptep_clear_flush(vma, haddr, ptep); |
ac46d4f3 | 4253 | mmu_notifier_invalidate_range(mm, range.start, range.end); |
5b7a1d40 | 4254 | set_huge_pte_at(mm, haddr, ptep, |
1e8f889b | 4255 | make_huge_pte(vma, new_page, 1)); |
d281ee61 | 4256 | page_remove_rmap(old_page, true); |
5b7a1d40 | 4257 | hugepage_add_new_anon_rmap(new_page, vma, haddr); |
8f251a3d | 4258 | SetHPageMigratable(new_page); |
1e8f889b DG |
4259 | /* Make the old page be freed below */ |
4260 | new_page = old_page; | |
4261 | } | |
cb900f41 | 4262 | spin_unlock(ptl); |
ac46d4f3 | 4263 | mmu_notifier_invalidate_range_end(&range); |
ad4404a2 | 4264 | out_release_all: |
5b7a1d40 | 4265 | restore_reserve_on_error(h, vma, haddr, new_page); |
09cbfeaf | 4266 | put_page(new_page); |
ad4404a2 | 4267 | out_release_old: |
09cbfeaf | 4268 | put_page(old_page); |
8312034f | 4269 | |
ad4404a2 DB |
4270 | spin_lock(ptl); /* Caller expects lock to be held */ |
4271 | return ret; | |
1e8f889b DG |
4272 | } |
4273 | ||
04f2cbe3 | 4274 | /* Return the pagecache page at a given address within a VMA */ |
a5516438 AK |
4275 | static struct page *hugetlbfs_pagecache_page(struct hstate *h, |
4276 | struct vm_area_struct *vma, unsigned long address) | |
04f2cbe3 MG |
4277 | { |
4278 | struct address_space *mapping; | |
e7c4b0bf | 4279 | pgoff_t idx; |
04f2cbe3 MG |
4280 | |
4281 | mapping = vma->vm_file->f_mapping; | |
a5516438 | 4282 | idx = vma_hugecache_offset(h, vma, address); |
04f2cbe3 MG |
4283 | |
4284 | return find_lock_page(mapping, idx); | |
4285 | } | |
4286 | ||
3ae77f43 HD |
4287 | /* |
4288 | * Return whether there is a pagecache page to back given address within VMA. | |
4289 | * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page. | |
4290 | */ | |
4291 | static bool hugetlbfs_pagecache_present(struct hstate *h, | |
2a15efc9 HD |
4292 | struct vm_area_struct *vma, unsigned long address) |
4293 | { | |
4294 | struct address_space *mapping; | |
4295 | pgoff_t idx; | |
4296 | struct page *page; | |
4297 | ||
4298 | mapping = vma->vm_file->f_mapping; | |
4299 | idx = vma_hugecache_offset(h, vma, address); | |
4300 | ||
4301 | page = find_get_page(mapping, idx); | |
4302 | if (page) | |
4303 | put_page(page); | |
4304 | return page != NULL; | |
4305 | } | |
4306 | ||
ab76ad54 MK |
4307 | int huge_add_to_page_cache(struct page *page, struct address_space *mapping, |
4308 | pgoff_t idx) | |
4309 | { | |
4310 | struct inode *inode = mapping->host; | |
4311 | struct hstate *h = hstate_inode(inode); | |
4312 | int err = add_to_page_cache(page, mapping, idx, GFP_KERNEL); | |
4313 | ||
4314 | if (err) | |
4315 | return err; | |
d6995da3 | 4316 | ClearHPageRestoreReserve(page); |
ab76ad54 | 4317 | |
22146c3c MK |
4318 | /* |
4319 | * set page dirty so that it will not be removed from cache/file | |
4320 | * by non-hugetlbfs specific code paths. | |
4321 | */ | |
4322 | set_page_dirty(page); | |
4323 | ||
ab76ad54 MK |
4324 | spin_lock(&inode->i_lock); |
4325 | inode->i_blocks += blocks_per_huge_page(h); | |
4326 | spin_unlock(&inode->i_lock); | |
4327 | return 0; | |
4328 | } | |
4329 | ||
2b740303 SJ |
4330 | static vm_fault_t hugetlb_no_page(struct mm_struct *mm, |
4331 | struct vm_area_struct *vma, | |
4332 | struct address_space *mapping, pgoff_t idx, | |
4333 | unsigned long address, pte_t *ptep, unsigned int flags) | |
ac9b9c66 | 4334 | { |
a5516438 | 4335 | struct hstate *h = hstate_vma(vma); |
2b740303 | 4336 | vm_fault_t ret = VM_FAULT_SIGBUS; |
409eb8c2 | 4337 | int anon_rmap = 0; |
4c887265 | 4338 | unsigned long size; |
4c887265 | 4339 | struct page *page; |
1e8f889b | 4340 | pte_t new_pte; |
cb900f41 | 4341 | spinlock_t *ptl; |
285b8dca | 4342 | unsigned long haddr = address & huge_page_mask(h); |
cb6acd01 | 4343 | bool new_page = false; |
4c887265 | 4344 | |
04f2cbe3 MG |
4345 | /* |
4346 | * Currently, we are forced to kill the process in the event the | |
4347 | * original mapper has unmapped pages from the child due to a failed | |
25985edc | 4348 | * COW. Warn that such a situation has occurred as it may not be obvious |
04f2cbe3 MG |
4349 | */ |
4350 | if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) { | |
910154d5 | 4351 | pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n", |
ffb22af5 | 4352 | current->pid); |
04f2cbe3 MG |
4353 | return ret; |
4354 | } | |
4355 | ||
4c887265 | 4356 | /* |
87bf91d3 MK |
4357 | * We can not race with truncation due to holding i_mmap_rwsem. |
4358 | * i_size is modified when holding i_mmap_rwsem, so check here | |
4359 | * once for faults beyond end of file. | |
4c887265 | 4360 | */ |
87bf91d3 MK |
4361 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4362 | if (idx >= size) | |
4363 | goto out; | |
4364 | ||
6bda666a CL |
4365 | retry: |
4366 | page = find_lock_page(mapping, idx); | |
4367 | if (!page) { | |
1a1aad8a MK |
4368 | /* |
4369 | * Check for page in userfault range | |
4370 | */ | |
4371 | if (userfaultfd_missing(vma)) { | |
4372 | u32 hash; | |
4373 | struct vm_fault vmf = { | |
4374 | .vma = vma, | |
285b8dca | 4375 | .address = haddr, |
1a1aad8a MK |
4376 | .flags = flags, |
4377 | /* | |
4378 | * Hard to debug if it ends up being | |
4379 | * used by a callee that assumes | |
4380 | * something about the other | |
4381 | * uninitialized fields... same as in | |
4382 | * memory.c | |
4383 | */ | |
4384 | }; | |
4385 | ||
4386 | /* | |
c0d0381a MK |
4387 | * hugetlb_fault_mutex and i_mmap_rwsem must be |
4388 | * dropped before handling userfault. Reacquire | |
4389 | * after handling fault to make calling code simpler. | |
1a1aad8a | 4390 | */ |
188b04a7 | 4391 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
1a1aad8a | 4392 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4393 | i_mmap_unlock_read(mapping); |
1a1aad8a | 4394 | ret = handle_userfault(&vmf, VM_UFFD_MISSING); |
c0d0381a | 4395 | i_mmap_lock_read(mapping); |
1a1aad8a MK |
4396 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
4397 | goto out; | |
4398 | } | |
4399 | ||
285b8dca | 4400 | page = alloc_huge_page(vma, haddr, 0); |
2fc39cec | 4401 | if (IS_ERR(page)) { |
4643d67e MK |
4402 | /* |
4403 | * Returning error will result in faulting task being | |
4404 | * sent SIGBUS. The hugetlb fault mutex prevents two | |
4405 | * tasks from racing to fault in the same page which | |
4406 | * could result in false unable to allocate errors. | |
4407 | * Page migration does not take the fault mutex, but | |
4408 | * does a clear then write of pte's under page table | |
4409 | * lock. Page fault code could race with migration, | |
4410 | * notice the clear pte and try to allocate a page | |
4411 | * here. Before returning error, get ptl and make | |
4412 | * sure there really is no pte entry. | |
4413 | */ | |
4414 | ptl = huge_pte_lock(h, mm, ptep); | |
d83e6c8a ML |
4415 | ret = 0; |
4416 | if (huge_pte_none(huge_ptep_get(ptep))) | |
4417 | ret = vmf_error(PTR_ERR(page)); | |
4643d67e | 4418 | spin_unlock(ptl); |
6bda666a CL |
4419 | goto out; |
4420 | } | |
47ad8475 | 4421 | clear_huge_page(page, address, pages_per_huge_page(h)); |
0ed361de | 4422 | __SetPageUptodate(page); |
cb6acd01 | 4423 | new_page = true; |
ac9b9c66 | 4424 | |
f83a275d | 4425 | if (vma->vm_flags & VM_MAYSHARE) { |
ab76ad54 | 4426 | int err = huge_add_to_page_cache(page, mapping, idx); |
6bda666a CL |
4427 | if (err) { |
4428 | put_page(page); | |
6bda666a CL |
4429 | if (err == -EEXIST) |
4430 | goto retry; | |
4431 | goto out; | |
4432 | } | |
23be7468 | 4433 | } else { |
6bda666a | 4434 | lock_page(page); |
0fe6e20b NH |
4435 | if (unlikely(anon_vma_prepare(vma))) { |
4436 | ret = VM_FAULT_OOM; | |
4437 | goto backout_unlocked; | |
4438 | } | |
409eb8c2 | 4439 | anon_rmap = 1; |
23be7468 | 4440 | } |
0fe6e20b | 4441 | } else { |
998b4382 NH |
4442 | /* |
4443 | * If memory error occurs between mmap() and fault, some process | |
4444 | * don't have hwpoisoned swap entry for errored virtual address. | |
4445 | * So we need to block hugepage fault by PG_hwpoison bit check. | |
4446 | */ | |
4447 | if (unlikely(PageHWPoison(page))) { | |
0eb98f15 | 4448 | ret = VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4449 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
998b4382 NH |
4450 | goto backout_unlocked; |
4451 | } | |
6bda666a | 4452 | } |
1e8f889b | 4453 | |
57303d80 AW |
4454 | /* |
4455 | * If we are going to COW a private mapping later, we examine the | |
4456 | * pending reservations for this page now. This will ensure that | |
4457 | * any allocations necessary to record that reservation occur outside | |
4458 | * the spinlock. | |
4459 | */ | |
5e911373 | 4460 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
285b8dca | 4461 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c AW |
4462 | ret = VM_FAULT_OOM; |
4463 | goto backout_unlocked; | |
4464 | } | |
5e911373 | 4465 | /* Just decrements count, does not deallocate */ |
285b8dca | 4466 | vma_end_reservation(h, vma, haddr); |
5e911373 | 4467 | } |
57303d80 | 4468 | |
8bea8052 | 4469 | ptl = huge_pte_lock(h, mm, ptep); |
83c54070 | 4470 | ret = 0; |
7f2e9525 | 4471 | if (!huge_pte_none(huge_ptep_get(ptep))) |
4c887265 AL |
4472 | goto backout; |
4473 | ||
07443a85 | 4474 | if (anon_rmap) { |
d6995da3 | 4475 | ClearHPageRestoreReserve(page); |
285b8dca | 4476 | hugepage_add_new_anon_rmap(page, vma, haddr); |
ac714904 | 4477 | } else |
53f9263b | 4478 | page_dup_rmap(page, true); |
1e8f889b DG |
4479 | new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE) |
4480 | && (vma->vm_flags & VM_SHARED))); | |
285b8dca | 4481 | set_huge_pte_at(mm, haddr, ptep, new_pte); |
1e8f889b | 4482 | |
5d317b2b | 4483 | hugetlb_count_add(pages_per_huge_page(h), mm); |
788c7df4 | 4484 | if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED)) { |
1e8f889b | 4485 | /* Optimization, do the COW without a second fault */ |
974e6d66 | 4486 | ret = hugetlb_cow(mm, vma, address, ptep, page, ptl); |
1e8f889b DG |
4487 | } |
4488 | ||
cb900f41 | 4489 | spin_unlock(ptl); |
cb6acd01 MK |
4490 | |
4491 | /* | |
8f251a3d MK |
4492 | * Only set HPageMigratable in newly allocated pages. Existing pages |
4493 | * found in the pagecache may not have HPageMigratableset if they have | |
4494 | * been isolated for migration. | |
cb6acd01 MK |
4495 | */ |
4496 | if (new_page) | |
8f251a3d | 4497 | SetHPageMigratable(page); |
cb6acd01 | 4498 | |
4c887265 AL |
4499 | unlock_page(page); |
4500 | out: | |
ac9b9c66 | 4501 | return ret; |
4c887265 AL |
4502 | |
4503 | backout: | |
cb900f41 | 4504 | spin_unlock(ptl); |
2b26736c | 4505 | backout_unlocked: |
4c887265 | 4506 | unlock_page(page); |
285b8dca | 4507 | restore_reserve_on_error(h, vma, haddr, page); |
4c887265 AL |
4508 | put_page(page); |
4509 | goto out; | |
ac9b9c66 HD |
4510 | } |
4511 | ||
8382d914 | 4512 | #ifdef CONFIG_SMP |
188b04a7 | 4513 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4514 | { |
4515 | unsigned long key[2]; | |
4516 | u32 hash; | |
4517 | ||
1b426bac MK |
4518 | key[0] = (unsigned long) mapping; |
4519 | key[1] = idx; | |
8382d914 | 4520 | |
55254636 | 4521 | hash = jhash2((u32 *)&key, sizeof(key)/(sizeof(u32)), 0); |
8382d914 DB |
4522 | |
4523 | return hash & (num_fault_mutexes - 1); | |
4524 | } | |
4525 | #else | |
4526 | /* | |
6c26d310 | 4527 | * For uniprocessor systems we always use a single mutex, so just |
8382d914 DB |
4528 | * return 0 and avoid the hashing overhead. |
4529 | */ | |
188b04a7 | 4530 | u32 hugetlb_fault_mutex_hash(struct address_space *mapping, pgoff_t idx) |
8382d914 DB |
4531 | { |
4532 | return 0; | |
4533 | } | |
4534 | #endif | |
4535 | ||
2b740303 | 4536 | vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
788c7df4 | 4537 | unsigned long address, unsigned int flags) |
86e5216f | 4538 | { |
8382d914 | 4539 | pte_t *ptep, entry; |
cb900f41 | 4540 | spinlock_t *ptl; |
2b740303 | 4541 | vm_fault_t ret; |
8382d914 DB |
4542 | u32 hash; |
4543 | pgoff_t idx; | |
0fe6e20b | 4544 | struct page *page = NULL; |
57303d80 | 4545 | struct page *pagecache_page = NULL; |
a5516438 | 4546 | struct hstate *h = hstate_vma(vma); |
8382d914 | 4547 | struct address_space *mapping; |
0f792cf9 | 4548 | int need_wait_lock = 0; |
285b8dca | 4549 | unsigned long haddr = address & huge_page_mask(h); |
86e5216f | 4550 | |
285b8dca | 4551 | ptep = huge_pte_offset(mm, haddr, huge_page_size(h)); |
fd6a03ed | 4552 | if (ptep) { |
c0d0381a MK |
4553 | /* |
4554 | * Since we hold no locks, ptep could be stale. That is | |
4555 | * OK as we are only making decisions based on content and | |
4556 | * not actually modifying content here. | |
4557 | */ | |
fd6a03ed | 4558 | entry = huge_ptep_get(ptep); |
290408d4 | 4559 | if (unlikely(is_hugetlb_entry_migration(entry))) { |
cb900f41 | 4560 | migration_entry_wait_huge(vma, mm, ptep); |
290408d4 NH |
4561 | return 0; |
4562 | } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry))) | |
32f84528 | 4563 | return VM_FAULT_HWPOISON_LARGE | |
972dc4de | 4564 | VM_FAULT_SET_HINDEX(hstate_index(h)); |
fd6a03ed NH |
4565 | } |
4566 | ||
c0d0381a MK |
4567 | /* |
4568 | * Acquire i_mmap_rwsem before calling huge_pte_alloc and hold | |
87bf91d3 MK |
4569 | * until finished with ptep. This serves two purposes: |
4570 | * 1) It prevents huge_pmd_unshare from being called elsewhere | |
4571 | * and making the ptep no longer valid. | |
4572 | * 2) It synchronizes us with i_size modifications during truncation. | |
c0d0381a MK |
4573 | * |
4574 | * ptep could have already be assigned via huge_pte_offset. That | |
4575 | * is OK, as huge_pte_alloc will return the same value unless | |
4576 | * something has changed. | |
4577 | */ | |
8382d914 | 4578 | mapping = vma->vm_file->f_mapping; |
c0d0381a | 4579 | i_mmap_lock_read(mapping); |
aec44e0f | 4580 | ptep = huge_pte_alloc(mm, vma, haddr, huge_page_size(h)); |
c0d0381a MK |
4581 | if (!ptep) { |
4582 | i_mmap_unlock_read(mapping); | |
4583 | return VM_FAULT_OOM; | |
4584 | } | |
8382d914 | 4585 | |
3935baa9 DG |
4586 | /* |
4587 | * Serialize hugepage allocation and instantiation, so that we don't | |
4588 | * get spurious allocation failures if two CPUs race to instantiate | |
4589 | * the same page in the page cache. | |
4590 | */ | |
c0d0381a | 4591 | idx = vma_hugecache_offset(h, vma, haddr); |
188b04a7 | 4592 | hash = hugetlb_fault_mutex_hash(mapping, idx); |
c672c7f2 | 4593 | mutex_lock(&hugetlb_fault_mutex_table[hash]); |
8382d914 | 4594 | |
7f2e9525 GS |
4595 | entry = huge_ptep_get(ptep); |
4596 | if (huge_pte_none(entry)) { | |
8382d914 | 4597 | ret = hugetlb_no_page(mm, vma, mapping, idx, address, ptep, flags); |
b4d1d99f | 4598 | goto out_mutex; |
3935baa9 | 4599 | } |
86e5216f | 4600 | |
83c54070 | 4601 | ret = 0; |
1e8f889b | 4602 | |
0f792cf9 NH |
4603 | /* |
4604 | * entry could be a migration/hwpoison entry at this point, so this | |
4605 | * check prevents the kernel from going below assuming that we have | |
7c8de358 EP |
4606 | * an active hugepage in pagecache. This goto expects the 2nd page |
4607 | * fault, and is_hugetlb_entry_(migration|hwpoisoned) check will | |
4608 | * properly handle it. | |
0f792cf9 NH |
4609 | */ |
4610 | if (!pte_present(entry)) | |
4611 | goto out_mutex; | |
4612 | ||
57303d80 AW |
4613 | /* |
4614 | * If we are going to COW the mapping later, we examine the pending | |
4615 | * reservations for this page now. This will ensure that any | |
4616 | * allocations necessary to record that reservation occur outside the | |
4617 | * spinlock. For private mappings, we also lookup the pagecache | |
4618 | * page now as it is used to determine if a reservation has been | |
4619 | * consumed. | |
4620 | */ | |
106c992a | 4621 | if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) { |
285b8dca | 4622 | if (vma_needs_reservation(h, vma, haddr) < 0) { |
2b26736c | 4623 | ret = VM_FAULT_OOM; |
b4d1d99f | 4624 | goto out_mutex; |
2b26736c | 4625 | } |
5e911373 | 4626 | /* Just decrements count, does not deallocate */ |
285b8dca | 4627 | vma_end_reservation(h, vma, haddr); |
57303d80 | 4628 | |
f83a275d | 4629 | if (!(vma->vm_flags & VM_MAYSHARE)) |
57303d80 | 4630 | pagecache_page = hugetlbfs_pagecache_page(h, |
285b8dca | 4631 | vma, haddr); |
57303d80 AW |
4632 | } |
4633 | ||
0f792cf9 NH |
4634 | ptl = huge_pte_lock(h, mm, ptep); |
4635 | ||
4636 | /* Check for a racing update before calling hugetlb_cow */ | |
4637 | if (unlikely(!pte_same(entry, huge_ptep_get(ptep)))) | |
4638 | goto out_ptl; | |
4639 | ||
56c9cfb1 NH |
4640 | /* |
4641 | * hugetlb_cow() requires page locks of pte_page(entry) and | |
4642 | * pagecache_page, so here we need take the former one | |
4643 | * when page != pagecache_page or !pagecache_page. | |
56c9cfb1 NH |
4644 | */ |
4645 | page = pte_page(entry); | |
4646 | if (page != pagecache_page) | |
0f792cf9 NH |
4647 | if (!trylock_page(page)) { |
4648 | need_wait_lock = 1; | |
4649 | goto out_ptl; | |
4650 | } | |
b4d1d99f | 4651 | |
0f792cf9 | 4652 | get_page(page); |
b4d1d99f | 4653 | |
788c7df4 | 4654 | if (flags & FAULT_FLAG_WRITE) { |
106c992a | 4655 | if (!huge_pte_write(entry)) { |
974e6d66 | 4656 | ret = hugetlb_cow(mm, vma, address, ptep, |
3999f52e | 4657 | pagecache_page, ptl); |
0f792cf9 | 4658 | goto out_put_page; |
b4d1d99f | 4659 | } |
106c992a | 4660 | entry = huge_pte_mkdirty(entry); |
b4d1d99f DG |
4661 | } |
4662 | entry = pte_mkyoung(entry); | |
285b8dca | 4663 | if (huge_ptep_set_access_flags(vma, haddr, ptep, entry, |
788c7df4 | 4664 | flags & FAULT_FLAG_WRITE)) |
285b8dca | 4665 | update_mmu_cache(vma, haddr, ptep); |
0f792cf9 NH |
4666 | out_put_page: |
4667 | if (page != pagecache_page) | |
4668 | unlock_page(page); | |
4669 | put_page(page); | |
cb900f41 KS |
4670 | out_ptl: |
4671 | spin_unlock(ptl); | |
57303d80 AW |
4672 | |
4673 | if (pagecache_page) { | |
4674 | unlock_page(pagecache_page); | |
4675 | put_page(pagecache_page); | |
4676 | } | |
b4d1d99f | 4677 | out_mutex: |
c672c7f2 | 4678 | mutex_unlock(&hugetlb_fault_mutex_table[hash]); |
c0d0381a | 4679 | i_mmap_unlock_read(mapping); |
0f792cf9 NH |
4680 | /* |
4681 | * Generally it's safe to hold refcount during waiting page lock. But | |
4682 | * here we just wait to defer the next page fault to avoid busy loop and | |
4683 | * the page is not used after unlocked before returning from the current | |
4684 | * page fault. So we are safe from accessing freed page, even if we wait | |
4685 | * here without taking refcount. | |
4686 | */ | |
4687 | if (need_wait_lock) | |
4688 | wait_on_page_locked(page); | |
1e8f889b | 4689 | return ret; |
86e5216f AL |
4690 | } |
4691 | ||
8fb5debc MK |
4692 | /* |
4693 | * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with | |
4694 | * modifications for huge pages. | |
4695 | */ | |
4696 | int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm, | |
4697 | pte_t *dst_pte, | |
4698 | struct vm_area_struct *dst_vma, | |
4699 | unsigned long dst_addr, | |
4700 | unsigned long src_addr, | |
4701 | struct page **pagep) | |
4702 | { | |
1e392147 AA |
4703 | struct address_space *mapping; |
4704 | pgoff_t idx; | |
4705 | unsigned long size; | |
1c9e8def | 4706 | int vm_shared = dst_vma->vm_flags & VM_SHARED; |
8fb5debc MK |
4707 | struct hstate *h = hstate_vma(dst_vma); |
4708 | pte_t _dst_pte; | |
4709 | spinlock_t *ptl; | |
4710 | int ret; | |
4711 | struct page *page; | |
4712 | ||
4713 | if (!*pagep) { | |
4714 | ret = -ENOMEM; | |
4715 | page = alloc_huge_page(dst_vma, dst_addr, 0); | |
4716 | if (IS_ERR(page)) | |
4717 | goto out; | |
4718 | ||
4719 | ret = copy_huge_page_from_user(page, | |
4720 | (const void __user *) src_addr, | |
810a56b9 | 4721 | pages_per_huge_page(h), false); |
8fb5debc | 4722 | |
c1e8d7c6 | 4723 | /* fallback to copy_from_user outside mmap_lock */ |
8fb5debc | 4724 | if (unlikely(ret)) { |
9e368259 | 4725 | ret = -ENOENT; |
8fb5debc MK |
4726 | *pagep = page; |
4727 | /* don't free the page */ | |
4728 | goto out; | |
4729 | } | |
4730 | } else { | |
4731 | page = *pagep; | |
4732 | *pagep = NULL; | |
4733 | } | |
4734 | ||
4735 | /* | |
4736 | * The memory barrier inside __SetPageUptodate makes sure that | |
4737 | * preceding stores to the page contents become visible before | |
4738 | * the set_pte_at() write. | |
4739 | */ | |
4740 | __SetPageUptodate(page); | |
8fb5debc | 4741 | |
1e392147 AA |
4742 | mapping = dst_vma->vm_file->f_mapping; |
4743 | idx = vma_hugecache_offset(h, dst_vma, dst_addr); | |
4744 | ||
1c9e8def MK |
4745 | /* |
4746 | * If shared, add to page cache | |
4747 | */ | |
4748 | if (vm_shared) { | |
1e392147 AA |
4749 | size = i_size_read(mapping->host) >> huge_page_shift(h); |
4750 | ret = -EFAULT; | |
4751 | if (idx >= size) | |
4752 | goto out_release_nounlock; | |
1c9e8def | 4753 | |
1e392147 AA |
4754 | /* |
4755 | * Serialization between remove_inode_hugepages() and | |
4756 | * huge_add_to_page_cache() below happens through the | |
4757 | * hugetlb_fault_mutex_table that here must be hold by | |
4758 | * the caller. | |
4759 | */ | |
1c9e8def MK |
4760 | ret = huge_add_to_page_cache(page, mapping, idx); |
4761 | if (ret) | |
4762 | goto out_release_nounlock; | |
4763 | } | |
4764 | ||
8fb5debc MK |
4765 | ptl = huge_pte_lockptr(h, dst_mm, dst_pte); |
4766 | spin_lock(ptl); | |
4767 | ||
1e392147 AA |
4768 | /* |
4769 | * Recheck the i_size after holding PT lock to make sure not | |
4770 | * to leave any page mapped (as page_mapped()) beyond the end | |
4771 | * of the i_size (remove_inode_hugepages() is strict about | |
4772 | * enforcing that). If we bail out here, we'll also leave a | |
4773 | * page in the radix tree in the vm_shared case beyond the end | |
4774 | * of the i_size, but remove_inode_hugepages() will take care | |
4775 | * of it as soon as we drop the hugetlb_fault_mutex_table. | |
4776 | */ | |
4777 | size = i_size_read(mapping->host) >> huge_page_shift(h); | |
4778 | ret = -EFAULT; | |
4779 | if (idx >= size) | |
4780 | goto out_release_unlock; | |
4781 | ||
8fb5debc MK |
4782 | ret = -EEXIST; |
4783 | if (!huge_pte_none(huge_ptep_get(dst_pte))) | |
4784 | goto out_release_unlock; | |
4785 | ||
1c9e8def MK |
4786 | if (vm_shared) { |
4787 | page_dup_rmap(page, true); | |
4788 | } else { | |
d6995da3 | 4789 | ClearHPageRestoreReserve(page); |
1c9e8def MK |
4790 | hugepage_add_new_anon_rmap(page, dst_vma, dst_addr); |
4791 | } | |
8fb5debc MK |
4792 | |
4793 | _dst_pte = make_huge_pte(dst_vma, page, dst_vma->vm_flags & VM_WRITE); | |
4794 | if (dst_vma->vm_flags & VM_WRITE) | |
4795 | _dst_pte = huge_pte_mkdirty(_dst_pte); | |
4796 | _dst_pte = pte_mkyoung(_dst_pte); | |
4797 | ||
4798 | set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte); | |
4799 | ||
4800 | (void)huge_ptep_set_access_flags(dst_vma, dst_addr, dst_pte, _dst_pte, | |
4801 | dst_vma->vm_flags & VM_WRITE); | |
4802 | hugetlb_count_add(pages_per_huge_page(h), dst_mm); | |
4803 | ||
4804 | /* No need to invalidate - it was non-present before */ | |
4805 | update_mmu_cache(dst_vma, dst_addr, dst_pte); | |
4806 | ||
4807 | spin_unlock(ptl); | |
8f251a3d | 4808 | SetHPageMigratable(page); |
1c9e8def MK |
4809 | if (vm_shared) |
4810 | unlock_page(page); | |
8fb5debc MK |
4811 | ret = 0; |
4812 | out: | |
4813 | return ret; | |
4814 | out_release_unlock: | |
4815 | spin_unlock(ptl); | |
1c9e8def MK |
4816 | if (vm_shared) |
4817 | unlock_page(page); | |
5af10dfd | 4818 | out_release_nounlock: |
8fb5debc MK |
4819 | put_page(page); |
4820 | goto out; | |
4821 | } | |
4822 | ||
82e5d378 JM |
4823 | static void record_subpages_vmas(struct page *page, struct vm_area_struct *vma, |
4824 | int refs, struct page **pages, | |
4825 | struct vm_area_struct **vmas) | |
4826 | { | |
4827 | int nr; | |
4828 | ||
4829 | for (nr = 0; nr < refs; nr++) { | |
4830 | if (likely(pages)) | |
4831 | pages[nr] = mem_map_offset(page, nr); | |
4832 | if (vmas) | |
4833 | vmas[nr] = vma; | |
4834 | } | |
4835 | } | |
4836 | ||
28a35716 ML |
4837 | long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
4838 | struct page **pages, struct vm_area_struct **vmas, | |
4839 | unsigned long *position, unsigned long *nr_pages, | |
4f6da934 | 4840 | long i, unsigned int flags, int *locked) |
63551ae0 | 4841 | { |
d5d4b0aa KC |
4842 | unsigned long pfn_offset; |
4843 | unsigned long vaddr = *position; | |
28a35716 | 4844 | unsigned long remainder = *nr_pages; |
a5516438 | 4845 | struct hstate *h = hstate_vma(vma); |
0fa5bc40 | 4846 | int err = -EFAULT, refs; |
63551ae0 | 4847 | |
63551ae0 | 4848 | while (vaddr < vma->vm_end && remainder) { |
4c887265 | 4849 | pte_t *pte; |
cb900f41 | 4850 | spinlock_t *ptl = NULL; |
2a15efc9 | 4851 | int absent; |
4c887265 | 4852 | struct page *page; |
63551ae0 | 4853 | |
02057967 DR |
4854 | /* |
4855 | * If we have a pending SIGKILL, don't keep faulting pages and | |
4856 | * potentially allocating memory. | |
4857 | */ | |
fa45f116 | 4858 | if (fatal_signal_pending(current)) { |
02057967 DR |
4859 | remainder = 0; |
4860 | break; | |
4861 | } | |
4862 | ||
4c887265 AL |
4863 | /* |
4864 | * Some archs (sparc64, sh*) have multiple pte_ts to | |
2a15efc9 | 4865 | * each hugepage. We have to make sure we get the |
4c887265 | 4866 | * first, for the page indexing below to work. |
cb900f41 KS |
4867 | * |
4868 | * Note that page table lock is not held when pte is null. | |
4c887265 | 4869 | */ |
7868a208 PA |
4870 | pte = huge_pte_offset(mm, vaddr & huge_page_mask(h), |
4871 | huge_page_size(h)); | |
cb900f41 KS |
4872 | if (pte) |
4873 | ptl = huge_pte_lock(h, mm, pte); | |
2a15efc9 HD |
4874 | absent = !pte || huge_pte_none(huge_ptep_get(pte)); |
4875 | ||
4876 | /* | |
4877 | * When coredumping, it suits get_dump_page if we just return | |
3ae77f43 HD |
4878 | * an error where there's an empty slot with no huge pagecache |
4879 | * to back it. This way, we avoid allocating a hugepage, and | |
4880 | * the sparse dumpfile avoids allocating disk blocks, but its | |
4881 | * huge holes still show up with zeroes where they need to be. | |
2a15efc9 | 4882 | */ |
3ae77f43 HD |
4883 | if (absent && (flags & FOLL_DUMP) && |
4884 | !hugetlbfs_pagecache_present(h, vma, vaddr)) { | |
cb900f41 KS |
4885 | if (pte) |
4886 | spin_unlock(ptl); | |
2a15efc9 HD |
4887 | remainder = 0; |
4888 | break; | |
4889 | } | |
63551ae0 | 4890 | |
9cc3a5bd NH |
4891 | /* |
4892 | * We need call hugetlb_fault for both hugepages under migration | |
4893 | * (in which case hugetlb_fault waits for the migration,) and | |
4894 | * hwpoisoned hugepages (in which case we need to prevent the | |
4895 | * caller from accessing to them.) In order to do this, we use | |
4896 | * here is_swap_pte instead of is_hugetlb_entry_migration and | |
4897 | * is_hugetlb_entry_hwpoisoned. This is because it simply covers | |
4898 | * both cases, and because we can't follow correct pages | |
4899 | * directly from any kind of swap entries. | |
4900 | */ | |
4901 | if (absent || is_swap_pte(huge_ptep_get(pte)) || | |
106c992a GS |
4902 | ((flags & FOLL_WRITE) && |
4903 | !huge_pte_write(huge_ptep_get(pte)))) { | |
2b740303 | 4904 | vm_fault_t ret; |
87ffc118 | 4905 | unsigned int fault_flags = 0; |
63551ae0 | 4906 | |
cb900f41 KS |
4907 | if (pte) |
4908 | spin_unlock(ptl); | |
87ffc118 AA |
4909 | if (flags & FOLL_WRITE) |
4910 | fault_flags |= FAULT_FLAG_WRITE; | |
4f6da934 | 4911 | if (locked) |
71335f37 PX |
4912 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | |
4913 | FAULT_FLAG_KILLABLE; | |
87ffc118 AA |
4914 | if (flags & FOLL_NOWAIT) |
4915 | fault_flags |= FAULT_FLAG_ALLOW_RETRY | | |
4916 | FAULT_FLAG_RETRY_NOWAIT; | |
4917 | if (flags & FOLL_TRIED) { | |
4426e945 PX |
4918 | /* |
4919 | * Note: FAULT_FLAG_ALLOW_RETRY and | |
4920 | * FAULT_FLAG_TRIED can co-exist | |
4921 | */ | |
87ffc118 AA |
4922 | fault_flags |= FAULT_FLAG_TRIED; |
4923 | } | |
4924 | ret = hugetlb_fault(mm, vma, vaddr, fault_flags); | |
4925 | if (ret & VM_FAULT_ERROR) { | |
2be7cfed | 4926 | err = vm_fault_to_errno(ret, flags); |
87ffc118 AA |
4927 | remainder = 0; |
4928 | break; | |
4929 | } | |
4930 | if (ret & VM_FAULT_RETRY) { | |
4f6da934 | 4931 | if (locked && |
1ac25013 | 4932 | !(fault_flags & FAULT_FLAG_RETRY_NOWAIT)) |
4f6da934 | 4933 | *locked = 0; |
87ffc118 AA |
4934 | *nr_pages = 0; |
4935 | /* | |
4936 | * VM_FAULT_RETRY must not return an | |
4937 | * error, it will return zero | |
4938 | * instead. | |
4939 | * | |
4940 | * No need to update "position" as the | |
4941 | * caller will not check it after | |
4942 | * *nr_pages is set to 0. | |
4943 | */ | |
4944 | return i; | |
4945 | } | |
4946 | continue; | |
4c887265 AL |
4947 | } |
4948 | ||
a5516438 | 4949 | pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT; |
7f2e9525 | 4950 | page = pte_page(huge_ptep_get(pte)); |
8fde12ca | 4951 | |
acbfb087 ZL |
4952 | /* |
4953 | * If subpage information not requested, update counters | |
4954 | * and skip the same_page loop below. | |
4955 | */ | |
4956 | if (!pages && !vmas && !pfn_offset && | |
4957 | (vaddr + huge_page_size(h) < vma->vm_end) && | |
4958 | (remainder >= pages_per_huge_page(h))) { | |
4959 | vaddr += huge_page_size(h); | |
4960 | remainder -= pages_per_huge_page(h); | |
4961 | i += pages_per_huge_page(h); | |
4962 | spin_unlock(ptl); | |
4963 | continue; | |
4964 | } | |
4965 | ||
82e5d378 JM |
4966 | refs = min3(pages_per_huge_page(h) - pfn_offset, |
4967 | (vma->vm_end - vaddr) >> PAGE_SHIFT, remainder); | |
0fa5bc40 | 4968 | |
82e5d378 JM |
4969 | if (pages || vmas) |
4970 | record_subpages_vmas(mem_map_offset(page, pfn_offset), | |
4971 | vma, refs, | |
4972 | likely(pages) ? pages + i : NULL, | |
4973 | vmas ? vmas + i : NULL); | |
63551ae0 | 4974 | |
82e5d378 | 4975 | if (pages) { |
0fa5bc40 JM |
4976 | /* |
4977 | * try_grab_compound_head() should always succeed here, | |
4978 | * because: a) we hold the ptl lock, and b) we've just | |
4979 | * checked that the huge page is present in the page | |
4980 | * tables. If the huge page is present, then the tail | |
4981 | * pages must also be present. The ptl prevents the | |
4982 | * head page and tail pages from being rearranged in | |
4983 | * any way. So this page must be available at this | |
4984 | * point, unless the page refcount overflowed: | |
4985 | */ | |
82e5d378 | 4986 | if (WARN_ON_ONCE(!try_grab_compound_head(pages[i], |
0fa5bc40 JM |
4987 | refs, |
4988 | flags))) { | |
4989 | spin_unlock(ptl); | |
4990 | remainder = 0; | |
4991 | err = -ENOMEM; | |
4992 | break; | |
4993 | } | |
d5d4b0aa | 4994 | } |
82e5d378 JM |
4995 | |
4996 | vaddr += (refs << PAGE_SHIFT); | |
4997 | remainder -= refs; | |
4998 | i += refs; | |
4999 | ||
cb900f41 | 5000 | spin_unlock(ptl); |
63551ae0 | 5001 | } |
28a35716 | 5002 | *nr_pages = remainder; |
87ffc118 AA |
5003 | /* |
5004 | * setting position is actually required only if remainder is | |
5005 | * not zero but it's faster not to add a "if (remainder)" | |
5006 | * branch. | |
5007 | */ | |
63551ae0 DG |
5008 | *position = vaddr; |
5009 | ||
2be7cfed | 5010 | return i ? i : err; |
63551ae0 | 5011 | } |
8f860591 | 5012 | |
7da4d641 | 5013 | unsigned long hugetlb_change_protection(struct vm_area_struct *vma, |
8f860591 ZY |
5014 | unsigned long address, unsigned long end, pgprot_t newprot) |
5015 | { | |
5016 | struct mm_struct *mm = vma->vm_mm; | |
5017 | unsigned long start = address; | |
5018 | pte_t *ptep; | |
5019 | pte_t pte; | |
a5516438 | 5020 | struct hstate *h = hstate_vma(vma); |
7da4d641 | 5021 | unsigned long pages = 0; |
dff11abe | 5022 | bool shared_pmd = false; |
ac46d4f3 | 5023 | struct mmu_notifier_range range; |
dff11abe MK |
5024 | |
5025 | /* | |
5026 | * In the case of shared PMDs, the area to flush could be beyond | |
ac46d4f3 | 5027 | * start/end. Set range.start/range.end to cover the maximum possible |
dff11abe MK |
5028 | * range if PMD sharing is possible. |
5029 | */ | |
7269f999 JG |
5030 | mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_VMA, |
5031 | 0, vma, mm, start, end); | |
ac46d4f3 | 5032 | adjust_range_if_pmd_sharing_possible(vma, &range.start, &range.end); |
8f860591 ZY |
5033 | |
5034 | BUG_ON(address >= end); | |
ac46d4f3 | 5035 | flush_cache_range(vma, range.start, range.end); |
8f860591 | 5036 | |
ac46d4f3 | 5037 | mmu_notifier_invalidate_range_start(&range); |
83cde9e8 | 5038 | i_mmap_lock_write(vma->vm_file->f_mapping); |
a5516438 | 5039 | for (; address < end; address += huge_page_size(h)) { |
cb900f41 | 5040 | spinlock_t *ptl; |
7868a208 | 5041 | ptep = huge_pte_offset(mm, address, huge_page_size(h)); |
8f860591 ZY |
5042 | if (!ptep) |
5043 | continue; | |
cb900f41 | 5044 | ptl = huge_pte_lock(h, mm, ptep); |
34ae204f | 5045 | if (huge_pmd_unshare(mm, vma, &address, ptep)) { |
7da4d641 | 5046 | pages++; |
cb900f41 | 5047 | spin_unlock(ptl); |
dff11abe | 5048 | shared_pmd = true; |
39dde65c | 5049 | continue; |
7da4d641 | 5050 | } |
a8bda28d NH |
5051 | pte = huge_ptep_get(ptep); |
5052 | if (unlikely(is_hugetlb_entry_hwpoisoned(pte))) { | |
5053 | spin_unlock(ptl); | |
5054 | continue; | |
5055 | } | |
5056 | if (unlikely(is_hugetlb_entry_migration(pte))) { | |
5057 | swp_entry_t entry = pte_to_swp_entry(pte); | |
5058 | ||
5059 | if (is_write_migration_entry(entry)) { | |
5060 | pte_t newpte; | |
5061 | ||
5062 | make_migration_entry_read(&entry); | |
5063 | newpte = swp_entry_to_pte(entry); | |
e5251fd4 PA |
5064 | set_huge_swap_pte_at(mm, address, ptep, |
5065 | newpte, huge_page_size(h)); | |
a8bda28d NH |
5066 | pages++; |
5067 | } | |
5068 | spin_unlock(ptl); | |
5069 | continue; | |
5070 | } | |
5071 | if (!huge_pte_none(pte)) { | |
023bdd00 AK |
5072 | pte_t old_pte; |
5073 | ||
5074 | old_pte = huge_ptep_modify_prot_start(vma, address, ptep); | |
5075 | pte = pte_mkhuge(huge_pte_modify(old_pte, newprot)); | |
be7517d6 | 5076 | pte = arch_make_huge_pte(pte, vma, NULL, 0); |
023bdd00 | 5077 | huge_ptep_modify_prot_commit(vma, address, ptep, old_pte, pte); |
7da4d641 | 5078 | pages++; |
8f860591 | 5079 | } |
cb900f41 | 5080 | spin_unlock(ptl); |
8f860591 | 5081 | } |
d833352a | 5082 | /* |
c8c06efa | 5083 | * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare |
d833352a | 5084 | * may have cleared our pud entry and done put_page on the page table: |
c8c06efa | 5085 | * once we release i_mmap_rwsem, another task can do the final put_page |
dff11abe MK |
5086 | * and that page table be reused and filled with junk. If we actually |
5087 | * did unshare a page of pmds, flush the range corresponding to the pud. | |
d833352a | 5088 | */ |
dff11abe | 5089 | if (shared_pmd) |
ac46d4f3 | 5090 | flush_hugetlb_tlb_range(vma, range.start, range.end); |
dff11abe MK |
5091 | else |
5092 | flush_hugetlb_tlb_range(vma, start, end); | |
0f10851e JG |
5093 | /* |
5094 | * No need to call mmu_notifier_invalidate_range() we are downgrading | |
5095 | * page table protection not changing it to point to a new page. | |
5096 | * | |
ad56b738 | 5097 | * See Documentation/vm/mmu_notifier.rst |
0f10851e | 5098 | */ |
83cde9e8 | 5099 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
ac46d4f3 | 5100 | mmu_notifier_invalidate_range_end(&range); |
7da4d641 PZ |
5101 | |
5102 | return pages << h->order; | |
8f860591 ZY |
5103 | } |
5104 | ||
33b8f84a MK |
5105 | /* Return true if reservation was successful, false otherwise. */ |
5106 | bool hugetlb_reserve_pages(struct inode *inode, | |
a1e78772 | 5107 | long from, long to, |
5a6fe125 | 5108 | struct vm_area_struct *vma, |
ca16d140 | 5109 | vm_flags_t vm_flags) |
e4e574b7 | 5110 | { |
33b8f84a | 5111 | long chg, add = -1; |
a5516438 | 5112 | struct hstate *h = hstate_inode(inode); |
90481622 | 5113 | struct hugepage_subpool *spool = subpool_inode(inode); |
9119a41e | 5114 | struct resv_map *resv_map; |
075a61d0 | 5115 | struct hugetlb_cgroup *h_cg = NULL; |
0db9d74e | 5116 | long gbl_reserve, regions_needed = 0; |
e4e574b7 | 5117 | |
63489f8e MK |
5118 | /* This should never happen */ |
5119 | if (from > to) { | |
5120 | VM_WARN(1, "%s called with a negative range\n", __func__); | |
33b8f84a | 5121 | return false; |
63489f8e MK |
5122 | } |
5123 | ||
17c9d12e MG |
5124 | /* |
5125 | * Only apply hugepage reservation if asked. At fault time, an | |
5126 | * attempt will be made for VM_NORESERVE to allocate a page | |
90481622 | 5127 | * without using reserves |
17c9d12e | 5128 | */ |
ca16d140 | 5129 | if (vm_flags & VM_NORESERVE) |
33b8f84a | 5130 | return true; |
17c9d12e | 5131 | |
a1e78772 MG |
5132 | /* |
5133 | * Shared mappings base their reservation on the number of pages that | |
5134 | * are already allocated on behalf of the file. Private mappings need | |
5135 | * to reserve the full area even if read-only as mprotect() may be | |
5136 | * called to make the mapping read-write. Assume !vma is a shm mapping | |
5137 | */ | |
9119a41e | 5138 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
f27a5136 MK |
5139 | /* |
5140 | * resv_map can not be NULL as hugetlb_reserve_pages is only | |
5141 | * called for inodes for which resv_maps were created (see | |
5142 | * hugetlbfs_get_inode). | |
5143 | */ | |
4e35f483 | 5144 | resv_map = inode_resv_map(inode); |
9119a41e | 5145 | |
0db9d74e | 5146 | chg = region_chg(resv_map, from, to, ®ions_needed); |
9119a41e JK |
5147 | |
5148 | } else { | |
e9fe92ae | 5149 | /* Private mapping. */ |
9119a41e | 5150 | resv_map = resv_map_alloc(); |
17c9d12e | 5151 | if (!resv_map) |
33b8f84a | 5152 | return false; |
17c9d12e | 5153 | |
a1e78772 | 5154 | chg = to - from; |
84afd99b | 5155 | |
17c9d12e MG |
5156 | set_vma_resv_map(vma, resv_map); |
5157 | set_vma_resv_flags(vma, HPAGE_RESV_OWNER); | |
5158 | } | |
5159 | ||
33b8f84a | 5160 | if (chg < 0) |
c50ac050 | 5161 | goto out_err; |
8a630112 | 5162 | |
33b8f84a MK |
5163 | if (hugetlb_cgroup_charge_cgroup_rsvd(hstate_index(h), |
5164 | chg * pages_per_huge_page(h), &h_cg) < 0) | |
075a61d0 | 5165 | goto out_err; |
075a61d0 MA |
5166 | |
5167 | if (vma && !(vma->vm_flags & VM_MAYSHARE) && h_cg) { | |
5168 | /* For private mappings, the hugetlb_cgroup uncharge info hangs | |
5169 | * of the resv_map. | |
5170 | */ | |
5171 | resv_map_set_hugetlb_cgroup_uncharge_info(resv_map, h_cg, h); | |
5172 | } | |
5173 | ||
1c5ecae3 MK |
5174 | /* |
5175 | * There must be enough pages in the subpool for the mapping. If | |
5176 | * the subpool has a minimum size, there may be some global | |
5177 | * reservations already in place (gbl_reserve). | |
5178 | */ | |
5179 | gbl_reserve = hugepage_subpool_get_pages(spool, chg); | |
33b8f84a | 5180 | if (gbl_reserve < 0) |
075a61d0 | 5181 | goto out_uncharge_cgroup; |
5a6fe125 MG |
5182 | |
5183 | /* | |
17c9d12e | 5184 | * Check enough hugepages are available for the reservation. |
90481622 | 5185 | * Hand the pages back to the subpool if there are not |
5a6fe125 | 5186 | */ |
33b8f84a | 5187 | if (hugetlb_acct_memory(h, gbl_reserve) < 0) |
075a61d0 | 5188 | goto out_put_pages; |
17c9d12e MG |
5189 | |
5190 | /* | |
5191 | * Account for the reservations made. Shared mappings record regions | |
5192 | * that have reservations as they are shared by multiple VMAs. | |
5193 | * When the last VMA disappears, the region map says how much | |
5194 | * the reservation was and the page cache tells how much of | |
5195 | * the reservation was consumed. Private mappings are per-VMA and | |
5196 | * only the consumed reservations are tracked. When the VMA | |
5197 | * disappears, the original reservation is the VMA size and the | |
5198 | * consumed reservations are stored in the map. Hence, nothing | |
5199 | * else has to be done for private mappings here | |
5200 | */ | |
33039678 | 5201 | if (!vma || vma->vm_flags & VM_MAYSHARE) { |
075a61d0 | 5202 | add = region_add(resv_map, from, to, regions_needed, h, h_cg); |
0db9d74e MA |
5203 | |
5204 | if (unlikely(add < 0)) { | |
5205 | hugetlb_acct_memory(h, -gbl_reserve); | |
075a61d0 | 5206 | goto out_put_pages; |
0db9d74e | 5207 | } else if (unlikely(chg > add)) { |
33039678 MK |
5208 | /* |
5209 | * pages in this range were added to the reserve | |
5210 | * map between region_chg and region_add. This | |
5211 | * indicates a race with alloc_huge_page. Adjust | |
5212 | * the subpool and reserve counts modified above | |
5213 | * based on the difference. | |
5214 | */ | |
5215 | long rsv_adjust; | |
5216 | ||
d85aecf2 ML |
5217 | /* |
5218 | * hugetlb_cgroup_uncharge_cgroup_rsvd() will put the | |
5219 | * reference to h_cg->css. See comment below for detail. | |
5220 | */ | |
075a61d0 MA |
5221 | hugetlb_cgroup_uncharge_cgroup_rsvd( |
5222 | hstate_index(h), | |
5223 | (chg - add) * pages_per_huge_page(h), h_cg); | |
5224 | ||
33039678 MK |
5225 | rsv_adjust = hugepage_subpool_put_pages(spool, |
5226 | chg - add); | |
5227 | hugetlb_acct_memory(h, -rsv_adjust); | |
d85aecf2 ML |
5228 | } else if (h_cg) { |
5229 | /* | |
5230 | * The file_regions will hold their own reference to | |
5231 | * h_cg->css. So we should release the reference held | |
5232 | * via hugetlb_cgroup_charge_cgroup_rsvd() when we are | |
5233 | * done. | |
5234 | */ | |
5235 | hugetlb_cgroup_put_rsvd_cgroup(h_cg); | |
33039678 MK |
5236 | } |
5237 | } | |
33b8f84a MK |
5238 | return true; |
5239 | ||
075a61d0 MA |
5240 | out_put_pages: |
5241 | /* put back original number of pages, chg */ | |
5242 | (void)hugepage_subpool_put_pages(spool, chg); | |
5243 | out_uncharge_cgroup: | |
5244 | hugetlb_cgroup_uncharge_cgroup_rsvd(hstate_index(h), | |
5245 | chg * pages_per_huge_page(h), h_cg); | |
c50ac050 | 5246 | out_err: |
5e911373 | 5247 | if (!vma || vma->vm_flags & VM_MAYSHARE) |
0db9d74e MA |
5248 | /* Only call region_abort if the region_chg succeeded but the |
5249 | * region_add failed or didn't run. | |
5250 | */ | |
5251 | if (chg >= 0 && add < 0) | |
5252 | region_abort(resv_map, from, to, regions_needed); | |
f031dd27 JK |
5253 | if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER)) |
5254 | kref_put(&resv_map->refs, resv_map_release); | |
33b8f84a | 5255 | return false; |
a43a8c39 KC |
5256 | } |
5257 | ||
b5cec28d MK |
5258 | long hugetlb_unreserve_pages(struct inode *inode, long start, long end, |
5259 | long freed) | |
a43a8c39 | 5260 | { |
a5516438 | 5261 | struct hstate *h = hstate_inode(inode); |
4e35f483 | 5262 | struct resv_map *resv_map = inode_resv_map(inode); |
9119a41e | 5263 | long chg = 0; |
90481622 | 5264 | struct hugepage_subpool *spool = subpool_inode(inode); |
1c5ecae3 | 5265 | long gbl_reserve; |
45c682a6 | 5266 | |
f27a5136 MK |
5267 | /* |
5268 | * Since this routine can be called in the evict inode path for all | |
5269 | * hugetlbfs inodes, resv_map could be NULL. | |
5270 | */ | |
b5cec28d MK |
5271 | if (resv_map) { |
5272 | chg = region_del(resv_map, start, end); | |
5273 | /* | |
5274 | * region_del() can fail in the rare case where a region | |
5275 | * must be split and another region descriptor can not be | |
5276 | * allocated. If end == LONG_MAX, it will not fail. | |
5277 | */ | |
5278 | if (chg < 0) | |
5279 | return chg; | |
5280 | } | |
5281 | ||
45c682a6 | 5282 | spin_lock(&inode->i_lock); |
e4c6f8be | 5283 | inode->i_blocks -= (blocks_per_huge_page(h) * freed); |
45c682a6 KC |
5284 | spin_unlock(&inode->i_lock); |
5285 | ||
1c5ecae3 MK |
5286 | /* |
5287 | * If the subpool has a minimum size, the number of global | |
5288 | * reservations to be released may be adjusted. | |
dddf31a4 ML |
5289 | * |
5290 | * Note that !resv_map implies freed == 0. So (chg - freed) | |
5291 | * won't go negative. | |
1c5ecae3 MK |
5292 | */ |
5293 | gbl_reserve = hugepage_subpool_put_pages(spool, (chg - freed)); | |
5294 | hugetlb_acct_memory(h, -gbl_reserve); | |
b5cec28d MK |
5295 | |
5296 | return 0; | |
a43a8c39 | 5297 | } |
93f70f90 | 5298 | |
3212b535 SC |
5299 | #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE |
5300 | static unsigned long page_table_shareable(struct vm_area_struct *svma, | |
5301 | struct vm_area_struct *vma, | |
5302 | unsigned long addr, pgoff_t idx) | |
5303 | { | |
5304 | unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) + | |
5305 | svma->vm_start; | |
5306 | unsigned long sbase = saddr & PUD_MASK; | |
5307 | unsigned long s_end = sbase + PUD_SIZE; | |
5308 | ||
5309 | /* Allow segments to share if only one is marked locked */ | |
de60f5f1 EM |
5310 | unsigned long vm_flags = vma->vm_flags & VM_LOCKED_CLEAR_MASK; |
5311 | unsigned long svm_flags = svma->vm_flags & VM_LOCKED_CLEAR_MASK; | |
3212b535 SC |
5312 | |
5313 | /* | |
5314 | * match the virtual addresses, permission and the alignment of the | |
5315 | * page table page. | |
5316 | */ | |
5317 | if (pmd_index(addr) != pmd_index(saddr) || | |
5318 | vm_flags != svm_flags || | |
07e51edf | 5319 | !range_in_vma(svma, sbase, s_end)) |
3212b535 SC |
5320 | return 0; |
5321 | ||
5322 | return saddr; | |
5323 | } | |
5324 | ||
31aafb45 | 5325 | static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr) |
3212b535 SC |
5326 | { |
5327 | unsigned long base = addr & PUD_MASK; | |
5328 | unsigned long end = base + PUD_SIZE; | |
5329 | ||
5330 | /* | |
5331 | * check on proper vm_flags and page table alignment | |
5332 | */ | |
017b1660 | 5333 | if (vma->vm_flags & VM_MAYSHARE && range_in_vma(vma, base, end)) |
31aafb45 NK |
5334 | return true; |
5335 | return false; | |
3212b535 SC |
5336 | } |
5337 | ||
c1991e07 PX |
5338 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) |
5339 | { | |
5340 | #ifdef CONFIG_USERFAULTFD | |
5341 | if (uffd_disable_huge_pmd_share(vma)) | |
5342 | return false; | |
5343 | #endif | |
5344 | return vma_shareable(vma, addr); | |
5345 | } | |
5346 | ||
017b1660 MK |
5347 | /* |
5348 | * Determine if start,end range within vma could be mapped by shared pmd. | |
5349 | * If yes, adjust start and end to cover range associated with possible | |
5350 | * shared pmd mappings. | |
5351 | */ | |
5352 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5353 | unsigned long *start, unsigned long *end) | |
5354 | { | |
a1ba9da8 LX |
5355 | unsigned long v_start = ALIGN(vma->vm_start, PUD_SIZE), |
5356 | v_end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
017b1660 | 5357 | |
a1ba9da8 LX |
5358 | /* |
5359 | * vma need span at least one aligned PUD size and the start,end range | |
5360 | * must at least partialy within it. | |
5361 | */ | |
5362 | if (!(vma->vm_flags & VM_MAYSHARE) || !(v_end > v_start) || | |
5363 | (*end <= v_start) || (*start >= v_end)) | |
017b1660 MK |
5364 | return; |
5365 | ||
75802ca6 | 5366 | /* Extend the range to be PUD aligned for a worst case scenario */ |
a1ba9da8 LX |
5367 | if (*start > v_start) |
5368 | *start = ALIGN_DOWN(*start, PUD_SIZE); | |
017b1660 | 5369 | |
a1ba9da8 LX |
5370 | if (*end < v_end) |
5371 | *end = ALIGN(*end, PUD_SIZE); | |
017b1660 MK |
5372 | } |
5373 | ||
3212b535 SC |
5374 | /* |
5375 | * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc() | |
5376 | * and returns the corresponding pte. While this is not necessary for the | |
5377 | * !shared pmd case because we can allocate the pmd later as well, it makes the | |
c0d0381a MK |
5378 | * code much cleaner. |
5379 | * | |
0bf7b64e MK |
5380 | * This routine must be called with i_mmap_rwsem held in at least read mode if |
5381 | * sharing is possible. For hugetlbfs, this prevents removal of any page | |
5382 | * table entries associated with the address space. This is important as we | |
5383 | * are setting up sharing based on existing page table entries (mappings). | |
5384 | * | |
5385 | * NOTE: This routine is only called from huge_pte_alloc. Some callers of | |
5386 | * huge_pte_alloc know that sharing is not possible and do not take | |
5387 | * i_mmap_rwsem as a performance optimization. This is handled by the | |
5388 | * if !vma_shareable check at the beginning of the routine. i_mmap_rwsem is | |
5389 | * only required for subsequent processing. | |
3212b535 | 5390 | */ |
aec44e0f PX |
5391 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
5392 | unsigned long addr, pud_t *pud) | |
3212b535 | 5393 | { |
3212b535 SC |
5394 | struct address_space *mapping = vma->vm_file->f_mapping; |
5395 | pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) + | |
5396 | vma->vm_pgoff; | |
5397 | struct vm_area_struct *svma; | |
5398 | unsigned long saddr; | |
5399 | pte_t *spte = NULL; | |
5400 | pte_t *pte; | |
cb900f41 | 5401 | spinlock_t *ptl; |
3212b535 | 5402 | |
0bf7b64e | 5403 | i_mmap_assert_locked(mapping); |
3212b535 SC |
5404 | vma_interval_tree_foreach(svma, &mapping->i_mmap, idx, idx) { |
5405 | if (svma == vma) | |
5406 | continue; | |
5407 | ||
5408 | saddr = page_table_shareable(svma, vma, addr, idx); | |
5409 | if (saddr) { | |
7868a208 PA |
5410 | spte = huge_pte_offset(svma->vm_mm, saddr, |
5411 | vma_mmu_pagesize(svma)); | |
3212b535 SC |
5412 | if (spte) { |
5413 | get_page(virt_to_page(spte)); | |
5414 | break; | |
5415 | } | |
5416 | } | |
5417 | } | |
5418 | ||
5419 | if (!spte) | |
5420 | goto out; | |
5421 | ||
8bea8052 | 5422 | ptl = huge_pte_lock(hstate_vma(vma), mm, spte); |
dc6c9a35 | 5423 | if (pud_none(*pud)) { |
3212b535 SC |
5424 | pud_populate(mm, pud, |
5425 | (pmd_t *)((unsigned long)spte & PAGE_MASK)); | |
c17b1f42 | 5426 | mm_inc_nr_pmds(mm); |
dc6c9a35 | 5427 | } else { |
3212b535 | 5428 | put_page(virt_to_page(spte)); |
dc6c9a35 | 5429 | } |
cb900f41 | 5430 | spin_unlock(ptl); |
3212b535 SC |
5431 | out: |
5432 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
3212b535 SC |
5433 | return pte; |
5434 | } | |
5435 | ||
5436 | /* | |
5437 | * unmap huge page backed by shared pte. | |
5438 | * | |
5439 | * Hugetlb pte page is ref counted at the time of mapping. If pte is shared | |
5440 | * indicated by page_count > 1, unmap is achieved by clearing pud and | |
5441 | * decrementing the ref count. If count == 1, the pte page is not shared. | |
5442 | * | |
c0d0381a | 5443 | * Called with page table lock held and i_mmap_rwsem held in write mode. |
3212b535 SC |
5444 | * |
5445 | * returns: 1 successfully unmapped a shared pte page | |
5446 | * 0 the underlying pte page is not shared, or it is the last user | |
5447 | */ | |
34ae204f MK |
5448 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5449 | unsigned long *addr, pte_t *ptep) | |
3212b535 SC |
5450 | { |
5451 | pgd_t *pgd = pgd_offset(mm, *addr); | |
c2febafc KS |
5452 | p4d_t *p4d = p4d_offset(pgd, *addr); |
5453 | pud_t *pud = pud_offset(p4d, *addr); | |
3212b535 | 5454 | |
34ae204f | 5455 | i_mmap_assert_write_locked(vma->vm_file->f_mapping); |
3212b535 SC |
5456 | BUG_ON(page_count(virt_to_page(ptep)) == 0); |
5457 | if (page_count(virt_to_page(ptep)) == 1) | |
5458 | return 0; | |
5459 | ||
5460 | pud_clear(pud); | |
5461 | put_page(virt_to_page(ptep)); | |
dc6c9a35 | 5462 | mm_dec_nr_pmds(mm); |
3212b535 SC |
5463 | *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE; |
5464 | return 1; | |
5465 | } | |
c1991e07 | 5466 | |
9e5fc74c | 5467 | #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
aec44e0f PX |
5468 | pte_t *huge_pmd_share(struct mm_struct *mm, struct vm_area_struct *vma, |
5469 | unsigned long addr, pud_t *pud) | |
9e5fc74c SC |
5470 | { |
5471 | return NULL; | |
5472 | } | |
e81f2d22 | 5473 | |
34ae204f MK |
5474 | int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma, |
5475 | unsigned long *addr, pte_t *ptep) | |
e81f2d22 ZZ |
5476 | { |
5477 | return 0; | |
5478 | } | |
017b1660 MK |
5479 | |
5480 | void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma, | |
5481 | unsigned long *start, unsigned long *end) | |
5482 | { | |
5483 | } | |
c1991e07 PX |
5484 | |
5485 | bool want_pmd_share(struct vm_area_struct *vma, unsigned long addr) | |
5486 | { | |
5487 | return false; | |
5488 | } | |
3212b535 SC |
5489 | #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */ |
5490 | ||
9e5fc74c | 5491 | #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB |
aec44e0f | 5492 | pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, |
9e5fc74c SC |
5493 | unsigned long addr, unsigned long sz) |
5494 | { | |
5495 | pgd_t *pgd; | |
c2febafc | 5496 | p4d_t *p4d; |
9e5fc74c SC |
5497 | pud_t *pud; |
5498 | pte_t *pte = NULL; | |
5499 | ||
5500 | pgd = pgd_offset(mm, addr); | |
f4f0a3d8 KS |
5501 | p4d = p4d_alloc(mm, pgd, addr); |
5502 | if (!p4d) | |
5503 | return NULL; | |
c2febafc | 5504 | pud = pud_alloc(mm, p4d, addr); |
9e5fc74c SC |
5505 | if (pud) { |
5506 | if (sz == PUD_SIZE) { | |
5507 | pte = (pte_t *)pud; | |
5508 | } else { | |
5509 | BUG_ON(sz != PMD_SIZE); | |
c1991e07 | 5510 | if (want_pmd_share(vma, addr) && pud_none(*pud)) |
aec44e0f | 5511 | pte = huge_pmd_share(mm, vma, addr, pud); |
9e5fc74c SC |
5512 | else |
5513 | pte = (pte_t *)pmd_alloc(mm, pud, addr); | |
5514 | } | |
5515 | } | |
4e666314 | 5516 | BUG_ON(pte && pte_present(*pte) && !pte_huge(*pte)); |
9e5fc74c SC |
5517 | |
5518 | return pte; | |
5519 | } | |
5520 | ||
9b19df29 PA |
5521 | /* |
5522 | * huge_pte_offset() - Walk the page table to resolve the hugepage | |
5523 | * entry at address @addr | |
5524 | * | |
8ac0b81a LX |
5525 | * Return: Pointer to page table entry (PUD or PMD) for |
5526 | * address @addr, or NULL if a !p*d_present() entry is encountered and the | |
9b19df29 PA |
5527 | * size @sz doesn't match the hugepage size at this level of the page |
5528 | * table. | |
5529 | */ | |
7868a208 PA |
5530 | pte_t *huge_pte_offset(struct mm_struct *mm, |
5531 | unsigned long addr, unsigned long sz) | |
9e5fc74c SC |
5532 | { |
5533 | pgd_t *pgd; | |
c2febafc | 5534 | p4d_t *p4d; |
8ac0b81a LX |
5535 | pud_t *pud; |
5536 | pmd_t *pmd; | |
9e5fc74c SC |
5537 | |
5538 | pgd = pgd_offset(mm, addr); | |
c2febafc KS |
5539 | if (!pgd_present(*pgd)) |
5540 | return NULL; | |
5541 | p4d = p4d_offset(pgd, addr); | |
5542 | if (!p4d_present(*p4d)) | |
5543 | return NULL; | |
9b19df29 | 5544 | |
c2febafc | 5545 | pud = pud_offset(p4d, addr); |
8ac0b81a LX |
5546 | if (sz == PUD_SIZE) |
5547 | /* must be pud huge, non-present or none */ | |
c2febafc | 5548 | return (pte_t *)pud; |
8ac0b81a | 5549 | if (!pud_present(*pud)) |
9b19df29 | 5550 | return NULL; |
8ac0b81a | 5551 | /* must have a valid entry and size to go further */ |
9b19df29 | 5552 | |
8ac0b81a LX |
5553 | pmd = pmd_offset(pud, addr); |
5554 | /* must be pmd huge, non-present or none */ | |
5555 | return (pte_t *)pmd; | |
9e5fc74c SC |
5556 | } |
5557 | ||
61f77eda NH |
5558 | #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */ |
5559 | ||
5560 | /* | |
5561 | * These functions are overwritable if your architecture needs its own | |
5562 | * behavior. | |
5563 | */ | |
5564 | struct page * __weak | |
5565 | follow_huge_addr(struct mm_struct *mm, unsigned long address, | |
5566 | int write) | |
5567 | { | |
5568 | return ERR_PTR(-EINVAL); | |
5569 | } | |
5570 | ||
4dc71451 AK |
5571 | struct page * __weak |
5572 | follow_huge_pd(struct vm_area_struct *vma, | |
5573 | unsigned long address, hugepd_t hpd, int flags, int pdshift) | |
5574 | { | |
5575 | WARN(1, "hugepd follow called with no support for hugepage directory format\n"); | |
5576 | return NULL; | |
5577 | } | |
5578 | ||
61f77eda | 5579 | struct page * __weak |
9e5fc74c | 5580 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5581 | pmd_t *pmd, int flags) |
9e5fc74c | 5582 | { |
e66f17ff NH |
5583 | struct page *page = NULL; |
5584 | spinlock_t *ptl; | |
c9d398fa | 5585 | pte_t pte; |
3faa52c0 JH |
5586 | |
5587 | /* FOLL_GET and FOLL_PIN are mutually exclusive. */ | |
5588 | if (WARN_ON_ONCE((flags & (FOLL_PIN | FOLL_GET)) == | |
5589 | (FOLL_PIN | FOLL_GET))) | |
5590 | return NULL; | |
5591 | ||
e66f17ff NH |
5592 | retry: |
5593 | ptl = pmd_lockptr(mm, pmd); | |
5594 | spin_lock(ptl); | |
5595 | /* | |
5596 | * make sure that the address range covered by this pmd is not | |
5597 | * unmapped from other threads. | |
5598 | */ | |
5599 | if (!pmd_huge(*pmd)) | |
5600 | goto out; | |
c9d398fa NH |
5601 | pte = huge_ptep_get((pte_t *)pmd); |
5602 | if (pte_present(pte)) { | |
97534127 | 5603 | page = pmd_page(*pmd) + ((address & ~PMD_MASK) >> PAGE_SHIFT); |
3faa52c0 JH |
5604 | /* |
5605 | * try_grab_page() should always succeed here, because: a) we | |
5606 | * hold the pmd (ptl) lock, and b) we've just checked that the | |
5607 | * huge pmd (head) page is present in the page tables. The ptl | |
5608 | * prevents the head page and tail pages from being rearranged | |
5609 | * in any way. So this page must be available at this point, | |
5610 | * unless the page refcount overflowed: | |
5611 | */ | |
5612 | if (WARN_ON_ONCE(!try_grab_page(page, flags))) { | |
5613 | page = NULL; | |
5614 | goto out; | |
5615 | } | |
e66f17ff | 5616 | } else { |
c9d398fa | 5617 | if (is_hugetlb_entry_migration(pte)) { |
e66f17ff NH |
5618 | spin_unlock(ptl); |
5619 | __migration_entry_wait(mm, (pte_t *)pmd, ptl); | |
5620 | goto retry; | |
5621 | } | |
5622 | /* | |
5623 | * hwpoisoned entry is treated as no_page_table in | |
5624 | * follow_page_mask(). | |
5625 | */ | |
5626 | } | |
5627 | out: | |
5628 | spin_unlock(ptl); | |
9e5fc74c SC |
5629 | return page; |
5630 | } | |
5631 | ||
61f77eda | 5632 | struct page * __weak |
9e5fc74c | 5633 | follow_huge_pud(struct mm_struct *mm, unsigned long address, |
e66f17ff | 5634 | pud_t *pud, int flags) |
9e5fc74c | 5635 | { |
3faa52c0 | 5636 | if (flags & (FOLL_GET | FOLL_PIN)) |
e66f17ff | 5637 | return NULL; |
9e5fc74c | 5638 | |
e66f17ff | 5639 | return pte_page(*(pte_t *)pud) + ((address & ~PUD_MASK) >> PAGE_SHIFT); |
9e5fc74c SC |
5640 | } |
5641 | ||
faaa5b62 AK |
5642 | struct page * __weak |
5643 | follow_huge_pgd(struct mm_struct *mm, unsigned long address, pgd_t *pgd, int flags) | |
5644 | { | |
3faa52c0 | 5645 | if (flags & (FOLL_GET | FOLL_PIN)) |
faaa5b62 AK |
5646 | return NULL; |
5647 | ||
5648 | return pte_page(*(pte_t *)pgd) + ((address & ~PGDIR_MASK) >> PAGE_SHIFT); | |
5649 | } | |
5650 | ||
31caf665 NH |
5651 | bool isolate_huge_page(struct page *page, struct list_head *list) |
5652 | { | |
bcc54222 NH |
5653 | bool ret = true; |
5654 | ||
db71ef79 | 5655 | spin_lock_irq(&hugetlb_lock); |
8f251a3d MK |
5656 | if (!PageHeadHuge(page) || |
5657 | !HPageMigratable(page) || | |
0eb2df2b | 5658 | !get_page_unless_zero(page)) { |
bcc54222 NH |
5659 | ret = false; |
5660 | goto unlock; | |
5661 | } | |
8f251a3d | 5662 | ClearHPageMigratable(page); |
31caf665 | 5663 | list_move_tail(&page->lru, list); |
bcc54222 | 5664 | unlock: |
db71ef79 | 5665 | spin_unlock_irq(&hugetlb_lock); |
bcc54222 | 5666 | return ret; |
31caf665 NH |
5667 | } |
5668 | ||
5669 | void putback_active_hugepage(struct page *page) | |
5670 | { | |
db71ef79 | 5671 | spin_lock_irq(&hugetlb_lock); |
8f251a3d | 5672 | SetHPageMigratable(page); |
31caf665 | 5673 | list_move_tail(&page->lru, &(page_hstate(page))->hugepage_activelist); |
db71ef79 | 5674 | spin_unlock_irq(&hugetlb_lock); |
31caf665 NH |
5675 | put_page(page); |
5676 | } | |
ab5ac90a MH |
5677 | |
5678 | void move_hugetlb_state(struct page *oldpage, struct page *newpage, int reason) | |
5679 | { | |
5680 | struct hstate *h = page_hstate(oldpage); | |
5681 | ||
5682 | hugetlb_cgroup_migrate(oldpage, newpage); | |
5683 | set_page_owner_migrate_reason(newpage, reason); | |
5684 | ||
5685 | /* | |
5686 | * transfer temporary state of the new huge page. This is | |
5687 | * reverse to other transitions because the newpage is going to | |
5688 | * be final while the old one will be freed so it takes over | |
5689 | * the temporary status. | |
5690 | * | |
5691 | * Also note that we have to transfer the per-node surplus state | |
5692 | * here as well otherwise the global surplus count will not match | |
5693 | * the per-node's. | |
5694 | */ | |
9157c311 | 5695 | if (HPageTemporary(newpage)) { |
ab5ac90a MH |
5696 | int old_nid = page_to_nid(oldpage); |
5697 | int new_nid = page_to_nid(newpage); | |
5698 | ||
9157c311 MK |
5699 | SetHPageTemporary(oldpage); |
5700 | ClearHPageTemporary(newpage); | |
ab5ac90a | 5701 | |
5af1ab1d ML |
5702 | /* |
5703 | * There is no need to transfer the per-node surplus state | |
5704 | * when we do not cross the node. | |
5705 | */ | |
5706 | if (new_nid == old_nid) | |
5707 | return; | |
db71ef79 | 5708 | spin_lock_irq(&hugetlb_lock); |
ab5ac90a MH |
5709 | if (h->surplus_huge_pages_node[old_nid]) { |
5710 | h->surplus_huge_pages_node[old_nid]--; | |
5711 | h->surplus_huge_pages_node[new_nid]++; | |
5712 | } | |
db71ef79 | 5713 | spin_unlock_irq(&hugetlb_lock); |
ab5ac90a MH |
5714 | } |
5715 | } | |
cf11e85f | 5716 | |
6dfeaff9 PX |
5717 | /* |
5718 | * This function will unconditionally remove all the shared pmd pgtable entries | |
5719 | * within the specific vma for a hugetlbfs memory range. | |
5720 | */ | |
5721 | void hugetlb_unshare_all_pmds(struct vm_area_struct *vma) | |
5722 | { | |
5723 | struct hstate *h = hstate_vma(vma); | |
5724 | unsigned long sz = huge_page_size(h); | |
5725 | struct mm_struct *mm = vma->vm_mm; | |
5726 | struct mmu_notifier_range range; | |
5727 | unsigned long address, start, end; | |
5728 | spinlock_t *ptl; | |
5729 | pte_t *ptep; | |
5730 | ||
5731 | if (!(vma->vm_flags & VM_MAYSHARE)) | |
5732 | return; | |
5733 | ||
5734 | start = ALIGN(vma->vm_start, PUD_SIZE); | |
5735 | end = ALIGN_DOWN(vma->vm_end, PUD_SIZE); | |
5736 | ||
5737 | if (start >= end) | |
5738 | return; | |
5739 | ||
5740 | /* | |
5741 | * No need to call adjust_range_if_pmd_sharing_possible(), because | |
5742 | * we have already done the PUD_SIZE alignment. | |
5743 | */ | |
5744 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm, | |
5745 | start, end); | |
5746 | mmu_notifier_invalidate_range_start(&range); | |
5747 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
5748 | for (address = start; address < end; address += PUD_SIZE) { | |
5749 | unsigned long tmp = address; | |
5750 | ||
5751 | ptep = huge_pte_offset(mm, address, sz); | |
5752 | if (!ptep) | |
5753 | continue; | |
5754 | ptl = huge_pte_lock(h, mm, ptep); | |
5755 | /* We don't want 'address' to be changed */ | |
5756 | huge_pmd_unshare(mm, vma, &tmp, ptep); | |
5757 | spin_unlock(ptl); | |
5758 | } | |
5759 | flush_hugetlb_tlb_range(vma, start, end); | |
5760 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
5761 | /* | |
5762 | * No need to call mmu_notifier_invalidate_range(), see | |
5763 | * Documentation/vm/mmu_notifier.rst. | |
5764 | */ | |
5765 | mmu_notifier_invalidate_range_end(&range); | |
5766 | } | |
5767 | ||
cf11e85f | 5768 | #ifdef CONFIG_CMA |
cf11e85f RG |
5769 | static bool cma_reserve_called __initdata; |
5770 | ||
5771 | static int __init cmdline_parse_hugetlb_cma(char *p) | |
5772 | { | |
5773 | hugetlb_cma_size = memparse(p, &p); | |
5774 | return 0; | |
5775 | } | |
5776 | ||
5777 | early_param("hugetlb_cma", cmdline_parse_hugetlb_cma); | |
5778 | ||
5779 | void __init hugetlb_cma_reserve(int order) | |
5780 | { | |
5781 | unsigned long size, reserved, per_node; | |
5782 | int nid; | |
5783 | ||
5784 | cma_reserve_called = true; | |
5785 | ||
5786 | if (!hugetlb_cma_size) | |
5787 | return; | |
5788 | ||
5789 | if (hugetlb_cma_size < (PAGE_SIZE << order)) { | |
5790 | pr_warn("hugetlb_cma: cma area should be at least %lu MiB\n", | |
5791 | (PAGE_SIZE << order) / SZ_1M); | |
5792 | return; | |
5793 | } | |
5794 | ||
5795 | /* | |
5796 | * If 3 GB area is requested on a machine with 4 numa nodes, | |
5797 | * let's allocate 1 GB on first three nodes and ignore the last one. | |
5798 | */ | |
5799 | per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes); | |
5800 | pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n", | |
5801 | hugetlb_cma_size / SZ_1M, per_node / SZ_1M); | |
5802 | ||
5803 | reserved = 0; | |
5804 | for_each_node_state(nid, N_ONLINE) { | |
5805 | int res; | |
2281f797 | 5806 | char name[CMA_MAX_NAME]; |
cf11e85f RG |
5807 | |
5808 | size = min(per_node, hugetlb_cma_size - reserved); | |
5809 | size = round_up(size, PAGE_SIZE << order); | |
5810 | ||
2281f797 | 5811 | snprintf(name, sizeof(name), "hugetlb%d", nid); |
cf11e85f | 5812 | res = cma_declare_contiguous_nid(0, size, 0, PAGE_SIZE << order, |
29d0f41d | 5813 | 0, false, name, |
cf11e85f RG |
5814 | &hugetlb_cma[nid], nid); |
5815 | if (res) { | |
5816 | pr_warn("hugetlb_cma: reservation failed: err %d, node %d", | |
5817 | res, nid); | |
5818 | continue; | |
5819 | } | |
5820 | ||
5821 | reserved += size; | |
5822 | pr_info("hugetlb_cma: reserved %lu MiB on node %d\n", | |
5823 | size / SZ_1M, nid); | |
5824 | ||
5825 | if (reserved >= hugetlb_cma_size) | |
5826 | break; | |
5827 | } | |
5828 | } | |
5829 | ||
5830 | void __init hugetlb_cma_check(void) | |
5831 | { | |
5832 | if (!hugetlb_cma_size || cma_reserve_called) | |
5833 | return; | |
5834 | ||
5835 | pr_warn("hugetlb_cma: the option isn't supported by current arch\n"); | |
5836 | } | |
5837 | ||
5838 | #endif /* CONFIG_CMA */ |