?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Linux 4.9.199
[linux/fpc-iii.git] / mm / huge_memory.c
blobf32f73fa5d3a1faba037ea3b08c337b9e81b9d5f
1 /*
2 * Copyright (C) 2009 Red Hat, Inc.
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
6 */
7
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10 #include <linux/mm.h>
11 #include <linux/sched.h>
12 #include <linux/highmem.h>
13 #include <linux/hugetlb.h>
14 #include <linux/mmu_notifier.h>
15 #include <linux/rmap.h>
16 #include <linux/swap.h>
17 #include <linux/shrinker.h>
18 #include <linux/mm_inline.h>
19 #include <linux/swapops.h>
20 #include <linux/dax.h>
21 #include <linux/khugepaged.h>
22 #include <linux/freezer.h>
23 #include <linux/pfn_t.h>
24 #include <linux/mman.h>
25 #include <linux/memremap.h>
26 #include <linux/pagemap.h>
27 #include <linux/debugfs.h>
28 #include <linux/migrate.h>
29 #include <linux/hashtable.h>
30 #include <linux/userfaultfd_k.h>
31 #include <linux/page_idle.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/page_owner.h>
34
35 #include <asm/tlb.h>
36 #include <asm/pgalloc.h>
37 #include "internal.h"
38
39 /*
40 * By default transparent hugepage support is disabled in order that avoid
41 * to risk increase the memory footprint of applications without a guaranteed
42 * benefit. When transparent hugepage support is enabled, is for all mappings,
43 * and khugepaged scans all mappings.
44 * Defrag is invoked by khugepaged hugepage allocations and by page faults
45 * for all hugepage allocations.
46 */
47 unsigned long transparent_hugepage_flags __read_mostly =
48 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
49 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
50 #endif
51 #ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
52 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
53 #endif
54 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
55 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
56 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
57
58 static struct shrinker deferred_split_shrinker;
59
60 static atomic_t huge_zero_refcount;
61 struct page *huge_zero_page __read_mostly;
62
63 static struct page *get_huge_zero_page(void)
64 {
65 struct page *zero_page;
66 retry:
67 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
68 return READ_ONCE(huge_zero_page);
69
70 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
71 HPAGE_PMD_ORDER);
72 if (!zero_page) {
73 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
74 return NULL;
75 }
76 count_vm_event(THP_ZERO_PAGE_ALLOC);
77 preempt_disable();
78 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
79 preempt_enable();
80 __free_pages(zero_page, compound_order(zero_page));
81 goto retry;
82 }
83
84 /* We take additional reference here. It will be put back by shrinker */
85 atomic_set(&huge_zero_refcount, 2);
86 preempt_enable();
87 return READ_ONCE(huge_zero_page);
88 }
89
90 static void put_huge_zero_page(void)
91 {
92 /*
93 * Counter should never go to zero here. Only shrinker can put
94 * last reference.
95 */
96 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
97 }
98
99 struct page *mm_get_huge_zero_page(struct mm_struct *mm)
100 {
101 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
102 return READ_ONCE(huge_zero_page);
103
104 if (!get_huge_zero_page())
105 return NULL;
106
107 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
108 put_huge_zero_page();
109
110 return READ_ONCE(huge_zero_page);
111 }
112
113 void mm_put_huge_zero_page(struct mm_struct *mm)
114 {
115 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
116 put_huge_zero_page();
117 }
118
119 static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
120 struct shrink_control *sc)
121 {
122 /* we can free zero page only if last reference remains */
123 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
124 }
125
126 static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
127 struct shrink_control *sc)
128 {
129 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
130 struct page *zero_page = xchg(&huge_zero_page, NULL);
131 BUG_ON(zero_page == NULL);
132 __free_pages(zero_page, compound_order(zero_page));
133 return HPAGE_PMD_NR;
134 }
135
136 return 0;
137 }
138
139 static struct shrinker huge_zero_page_shrinker = {
140 .count_objects = shrink_huge_zero_page_count,
141 .scan_objects = shrink_huge_zero_page_scan,
142 .seeks = DEFAULT_SEEKS,
143 };
144
145 #ifdef CONFIG_SYSFS
146
147 static ssize_t triple_flag_store(struct kobject *kobj,
148 struct kobj_attribute *attr,
149 const char *buf, size_t count,
150 enum transparent_hugepage_flag enabled,
151 enum transparent_hugepage_flag deferred,
152 enum transparent_hugepage_flag req_madv)
153 {
154 if (!memcmp("defer", buf,
155 min(sizeof("defer")-1, count))) {
156 if (enabled == deferred)
157 return -EINVAL;
158 clear_bit(enabled, &transparent_hugepage_flags);
159 clear_bit(req_madv, &transparent_hugepage_flags);
160 set_bit(deferred, &transparent_hugepage_flags);
161 } else if (!memcmp("always", buf,
162 min(sizeof("always")-1, count))) {
163 clear_bit(deferred, &transparent_hugepage_flags);
164 clear_bit(req_madv, &transparent_hugepage_flags);
165 set_bit(enabled, &transparent_hugepage_flags);
166 } else if (!memcmp("madvise", buf,
167 min(sizeof("madvise")-1, count))) {
168 clear_bit(enabled, &transparent_hugepage_flags);
169 clear_bit(deferred, &transparent_hugepage_flags);
170 set_bit(req_madv, &transparent_hugepage_flags);
171 } else if (!memcmp("never", buf,
172 min(sizeof("never")-1, count))) {
173 clear_bit(enabled, &transparent_hugepage_flags);
174 clear_bit(req_madv, &transparent_hugepage_flags);
175 clear_bit(deferred, &transparent_hugepage_flags);
176 } else
177 return -EINVAL;
178
179 return count;
180 }
181
182 static ssize_t enabled_show(struct kobject *kobj,
183 struct kobj_attribute *attr, char *buf)
184 {
185 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
186 return sprintf(buf, "[always] madvise never\n");
187 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
188 return sprintf(buf, "always [madvise] never\n");
189 else
190 return sprintf(buf, "always madvise [never]\n");
191 }
192
193 static ssize_t enabled_store(struct kobject *kobj,
194 struct kobj_attribute *attr,
195 const char *buf, size_t count)
196 {
197 ssize_t ret;
198
199 ret = triple_flag_store(kobj, attr, buf, count,
200 TRANSPARENT_HUGEPAGE_FLAG,
201 TRANSPARENT_HUGEPAGE_FLAG,
202 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
203
204 if (ret > 0) {
205 int err = start_stop_khugepaged();
206 if (err)
207 ret = err;
208 }
209
210 return ret;
211 }
212 static struct kobj_attribute enabled_attr =
213 __ATTR(enabled, 0644, enabled_show, enabled_store);
214
215 ssize_t single_hugepage_flag_show(struct kobject *kobj,
216 struct kobj_attribute *attr, char *buf,
217 enum transparent_hugepage_flag flag)
218 {
219 return sprintf(buf, "%d\n",
220 !!test_bit(flag, &transparent_hugepage_flags));
221 }
222
223 ssize_t single_hugepage_flag_store(struct kobject *kobj,
224 struct kobj_attribute *attr,
225 const char *buf, size_t count,
226 enum transparent_hugepage_flag flag)
227 {
228 unsigned long value;
229 int ret;
230
231 ret = kstrtoul(buf, 10, &value);
232 if (ret < 0)
233 return ret;
234 if (value > 1)
235 return -EINVAL;
236
237 if (value)
238 set_bit(flag, &transparent_hugepage_flags);
239 else
240 clear_bit(flag, &transparent_hugepage_flags);
241
242 return count;
243 }
244
245 /*
246 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
247 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
248 * memory just to allocate one more hugepage.
249 */
250 static ssize_t defrag_show(struct kobject *kobj,
251 struct kobj_attribute *attr, char *buf)
252 {
253 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
254 return sprintf(buf, "[always] defer madvise never\n");
255 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
256 return sprintf(buf, "always [defer] madvise never\n");
257 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
258 return sprintf(buf, "always defer [madvise] never\n");
259 else
260 return sprintf(buf, "always defer madvise [never]\n");
261
262 }
263 static ssize_t defrag_store(struct kobject *kobj,
264 struct kobj_attribute *attr,
265 const char *buf, size_t count)
266 {
267 return triple_flag_store(kobj, attr, buf, count,
268 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
269 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
270 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
271 }
272 static struct kobj_attribute defrag_attr =
273 __ATTR(defrag, 0644, defrag_show, defrag_store);
274
275 static ssize_t use_zero_page_show(struct kobject *kobj,
276 struct kobj_attribute *attr, char *buf)
277 {
278 return single_hugepage_flag_show(kobj, attr, buf,
279 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
280 }
281 static ssize_t use_zero_page_store(struct kobject *kobj,
282 struct kobj_attribute *attr, const char *buf, size_t count)
283 {
284 return single_hugepage_flag_store(kobj, attr, buf, count,
285 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
286 }
287 static struct kobj_attribute use_zero_page_attr =
288 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
289 #ifdef CONFIG_DEBUG_VM
290 static ssize_t debug_cow_show(struct kobject *kobj,
291 struct kobj_attribute *attr, char *buf)
292 {
293 return single_hugepage_flag_show(kobj, attr, buf,
294 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
295 }
296 static ssize_t debug_cow_store(struct kobject *kobj,
297 struct kobj_attribute *attr,
298 const char *buf, size_t count)
299 {
300 return single_hugepage_flag_store(kobj, attr, buf, count,
301 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
302 }
303 static struct kobj_attribute debug_cow_attr =
304 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
305 #endif /* CONFIG_DEBUG_VM */
306
307 static struct attribute *hugepage_attr[] = {
308 &enabled_attr.attr,
309 &defrag_attr.attr,
310 &use_zero_page_attr.attr,
311 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
312 &shmem_enabled_attr.attr,
313 #endif
314 #ifdef CONFIG_DEBUG_VM
315 &debug_cow_attr.attr,
316 #endif
317 NULL,
318 };
319
320 static struct attribute_group hugepage_attr_group = {
321 .attrs = hugepage_attr,
322 };
323
324 static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
325 {
326 int err;
327
328 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
329 if (unlikely(!*hugepage_kobj)) {
330 pr_err("failed to create transparent hugepage kobject\n");
331 return -ENOMEM;
332 }
333
334 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
335 if (err) {
336 pr_err("failed to register transparent hugepage group\n");
337 goto delete_obj;
338 }
339
340 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
341 if (err) {
342 pr_err("failed to register transparent hugepage group\n");
343 goto remove_hp_group;
344 }
345
346 return 0;
347
348 remove_hp_group:
349 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
350 delete_obj:
351 kobject_put(*hugepage_kobj);
352 return err;
353 }
354
355 static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
356 {
357 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
358 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
359 kobject_put(hugepage_kobj);
360 }
361 #else
362 static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
363 {
364 return 0;
365 }
366
367 static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
368 {
369 }
370 #endif /* CONFIG_SYSFS */
371
372 static int __init hugepage_init(void)
373 {
374 int err;
375 struct kobject *hugepage_kobj;
376
377 if (!has_transparent_hugepage()) {
378 transparent_hugepage_flags = 0;
379 return -EINVAL;
380 }
381
382 /*
383 * hugepages can't be allocated by the buddy allocator
384 */
385 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
386 /*
387 * we use page->mapping and page->index in second tail page
388 * as list_head: assuming THP order >= 2
389 */
390 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
391
392 err = hugepage_init_sysfs(&hugepage_kobj);
393 if (err)
394 goto err_sysfs;
395
396 err = khugepaged_init();
397 if (err)
398 goto err_slab;
399
400 err = register_shrinker(&huge_zero_page_shrinker);
401 if (err)
402 goto err_hzp_shrinker;
403 err = register_shrinker(&deferred_split_shrinker);
404 if (err)
405 goto err_split_shrinker;
406
407 /*
408 * By default disable transparent hugepages on smaller systems,
409 * where the extra memory used could hurt more than TLB overhead
410 * is likely to save. The admin can still enable it through /sys.
411 */
412 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
413 transparent_hugepage_flags = 0;
414 return 0;
415 }
416
417 err = start_stop_khugepaged();
418 if (err)
419 goto err_khugepaged;
420
421 return 0;
422 err_khugepaged:
423 unregister_shrinker(&deferred_split_shrinker);
424 err_split_shrinker:
425 unregister_shrinker(&huge_zero_page_shrinker);
426 err_hzp_shrinker:
427 khugepaged_destroy();
428 err_slab:
429 hugepage_exit_sysfs(hugepage_kobj);
430 err_sysfs:
431 return err;
432 }
433 subsys_initcall(hugepage_init);
434
435 static int __init setup_transparent_hugepage(char *str)
436 {
437 int ret = 0;
438 if (!str)
439 goto out;
440 if (!strcmp(str, "always")) {
441 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
442 &transparent_hugepage_flags);
443 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
444 &transparent_hugepage_flags);
445 ret = 1;
446 } else if (!strcmp(str, "madvise")) {
447 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
448 &transparent_hugepage_flags);
449 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
450 &transparent_hugepage_flags);
451 ret = 1;
452 } else if (!strcmp(str, "never")) {
453 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
454 &transparent_hugepage_flags);
455 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
456 &transparent_hugepage_flags);
457 ret = 1;
458 }
459 out:
460 if (!ret)
461 pr_warn("transparent_hugepage= cannot parse, ignored\n");
462 return ret;
463 }
464 __setup("transparent_hugepage=", setup_transparent_hugepage);
465
466 pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
467 {
468 if (likely(vma->vm_flags & VM_WRITE))
469 pmd = pmd_mkwrite(pmd);
470 return pmd;
471 }
472
473 static inline struct list_head *page_deferred_list(struct page *page)
474 {
475 /*
476 * ->lru in the tail pages is occupied by compound_head.
477 * Let's use ->mapping + ->index in the second tail page as list_head.
478 */
479 return (struct list_head *)&page[2].mapping;
480 }
481
482 void prep_transhuge_page(struct page *page)
483 {
484 /*
485 * we use page->mapping and page->indexlru in second tail page
486 * as list_head: assuming THP order >= 2
487 */
488
489 INIT_LIST_HEAD(page_deferred_list(page));
490 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
491 }
492
493 unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
494 loff_t off, unsigned long flags, unsigned long size)
495 {
496 unsigned long addr;
497 loff_t off_end = off + len;
498 loff_t off_align = round_up(off, size);
499 unsigned long len_pad;
500
501 if (off_end <= off_align || (off_end - off_align) < size)
502 return 0;
503
504 len_pad = len + size;
505 if (len_pad < len || (off + len_pad) < off)
506 return 0;
507
508 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
509 off >> PAGE_SHIFT, flags);
510 if (IS_ERR_VALUE(addr))
511 return 0;
512
513 addr += (off - addr) & (size - 1);
514 return addr;
515 }
516
517 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
518 unsigned long len, unsigned long pgoff, unsigned long flags)
519 {
520 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
521
522 if (addr)
523 goto out;
524 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
525 goto out;
526
527 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
528 if (addr)
529 return addr;
530
531 out:
532 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
533 }
534 EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
535
536 static int __do_huge_pmd_anonymous_page(struct fault_env *fe, struct page *page,
537 gfp_t gfp)
538 {
539 struct vm_area_struct *vma = fe->vma;
540 struct mem_cgroup *memcg;
541 pgtable_t pgtable;
542 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
543
544 VM_BUG_ON_PAGE(!PageCompound(page), page);
545
546 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp | __GFP_NORETRY, &memcg,
547 true)) {
548 put_page(page);
549 count_vm_event(THP_FAULT_FALLBACK);
550 return VM_FAULT_FALLBACK;
551 }
552
553 pgtable = pte_alloc_one(vma->vm_mm, haddr);
554 if (unlikely(!pgtable)) {
555 mem_cgroup_cancel_charge(page, memcg, true);
556 put_page(page);
557 return VM_FAULT_OOM;
558 }
559
560 clear_huge_page(page, haddr, HPAGE_PMD_NR);
561 /*
562 * The memory barrier inside __SetPageUptodate makes sure that
563 * clear_huge_page writes become visible before the set_pmd_at()
564 * write.
565 */
566 __SetPageUptodate(page);
567
568 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
569 if (unlikely(!pmd_none(*fe->pmd))) {
570 spin_unlock(fe->ptl);
571 mem_cgroup_cancel_charge(page, memcg, true);
572 put_page(page);
573 pte_free(vma->vm_mm, pgtable);
574 } else {
575 pmd_t entry;
576
577 /* Deliver the page fault to userland */
578 if (userfaultfd_missing(vma)) {
579 int ret;
580
581 spin_unlock(fe->ptl);
582 mem_cgroup_cancel_charge(page, memcg, true);
583 put_page(page);
584 pte_free(vma->vm_mm, pgtable);
585 ret = handle_userfault(fe, VM_UFFD_MISSING);
586 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
587 return ret;
588 }
589
590 entry = mk_huge_pmd(page, vma->vm_page_prot);
591 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
592 page_add_new_anon_rmap(page, vma, haddr, true);
593 mem_cgroup_commit_charge(page, memcg, false, true);
594 lru_cache_add_active_or_unevictable(page, vma);
595 pgtable_trans_huge_deposit(vma->vm_mm, fe->pmd, pgtable);
596 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
597 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
598 atomic_long_inc(&vma->vm_mm->nr_ptes);
599 spin_unlock(fe->ptl);
600 count_vm_event(THP_FAULT_ALLOC);
601 }
602
603 return 0;
604 }
605
606 /*
607 * If THP defrag is set to always then directly reclaim/compact as necessary
608 * If set to defer then do only background reclaim/compact and defer to khugepaged
609 * If set to madvise and the VMA is flagged then directly reclaim/compact
610 * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
611 */
612 static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
613 {
614 bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
615
616 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
617 &transparent_hugepage_flags) && vma_madvised)
618 return GFP_TRANSHUGE;
619 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
620 &transparent_hugepage_flags))
621 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
622 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
623 &transparent_hugepage_flags))
624 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
625
626 return GFP_TRANSHUGE_LIGHT;
627 }
628
629 /* Caller must hold page table lock. */
630 static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
631 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
632 struct page *zero_page)
633 {
634 pmd_t entry;
635 if (!pmd_none(*pmd))
636 return false;
637 entry = mk_pmd(zero_page, vma->vm_page_prot);
638 entry = pmd_mkhuge(entry);
639 if (pgtable)
640 pgtable_trans_huge_deposit(mm, pmd, pgtable);
641 set_pmd_at(mm, haddr, pmd, entry);
642 atomic_long_inc(&mm->nr_ptes);
643 return true;
644 }
645
646 int do_huge_pmd_anonymous_page(struct fault_env *fe)
647 {
648 struct vm_area_struct *vma = fe->vma;
649 gfp_t gfp;
650 struct page *page;
651 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
652
653 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
654 return VM_FAULT_FALLBACK;
655 if (unlikely(anon_vma_prepare(vma)))
656 return VM_FAULT_OOM;
657 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
658 return VM_FAULT_OOM;
659 if (!(fe->flags & FAULT_FLAG_WRITE) &&
660 !mm_forbids_zeropage(vma->vm_mm) &&
661 transparent_hugepage_use_zero_page()) {
662 pgtable_t pgtable;
663 struct page *zero_page;
664 bool set;
665 int ret;
666 pgtable = pte_alloc_one(vma->vm_mm, haddr);
667 if (unlikely(!pgtable))
668 return VM_FAULT_OOM;
669 zero_page = mm_get_huge_zero_page(vma->vm_mm);
670 if (unlikely(!zero_page)) {
671 pte_free(vma->vm_mm, pgtable);
672 count_vm_event(THP_FAULT_FALLBACK);
673 return VM_FAULT_FALLBACK;
674 }
675 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
676 ret = 0;
677 set = false;
678 if (pmd_none(*fe->pmd)) {
679 if (userfaultfd_missing(vma)) {
680 spin_unlock(fe->ptl);
681 ret = handle_userfault(fe, VM_UFFD_MISSING);
682 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
683 } else {
684 set_huge_zero_page(pgtable, vma->vm_mm, vma,
685 haddr, fe->pmd, zero_page);
686 spin_unlock(fe->ptl);
687 set = true;
688 }
689 } else
690 spin_unlock(fe->ptl);
691 if (!set)
692 pte_free(vma->vm_mm, pgtable);
693 return ret;
694 }
695 gfp = alloc_hugepage_direct_gfpmask(vma);
696 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
697 if (unlikely(!page)) {
698 count_vm_event(THP_FAULT_FALLBACK);
699 return VM_FAULT_FALLBACK;
700 }
701 prep_transhuge_page(page);
702 return __do_huge_pmd_anonymous_page(fe, page, gfp);
703 }
704
705 static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
706 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
707 {
708 struct mm_struct *mm = vma->vm_mm;
709 pmd_t entry;
710 spinlock_t *ptl;
711
712 ptl = pmd_lock(mm, pmd);
713 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
714 if (pfn_t_devmap(pfn))
715 entry = pmd_mkdevmap(entry);
716 if (write) {
717 entry = pmd_mkyoung(pmd_mkdirty(entry));
718 entry = maybe_pmd_mkwrite(entry, vma);
719 }
720 set_pmd_at(mm, addr, pmd, entry);
721 update_mmu_cache_pmd(vma, addr, pmd);
722 spin_unlock(ptl);
723 }
724
725 int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
726 pmd_t *pmd, pfn_t pfn, bool write)
727 {
728 pgprot_t pgprot = vma->vm_page_prot;
729 /*
730 * If we had pmd_special, we could avoid all these restrictions,
731 * but we need to be consistent with PTEs and architectures that
732 * can't support a 'special' bit.
733 */
734 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
735 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
736 (VM_PFNMAP|VM_MIXEDMAP));
737 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
738 BUG_ON(!pfn_t_devmap(pfn));
739
740 if (addr < vma->vm_start || addr >= vma->vm_end)
741 return VM_FAULT_SIGBUS;
742 if (track_pfn_insert(vma, &pgprot, pfn))
743 return VM_FAULT_SIGBUS;
744 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
745 return VM_FAULT_NOPAGE;
746 }
747 EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
748
749 static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
750 pmd_t *pmd, int flags)
751 {
752 pmd_t _pmd;
753
754 _pmd = pmd_mkyoung(*pmd);
755 if (flags & FOLL_WRITE)
756 _pmd = pmd_mkdirty(_pmd);
757 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
758 pmd, _pmd, flags & FOLL_WRITE))
759 update_mmu_cache_pmd(vma, addr, pmd);
760 }
761
762 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
763 pmd_t *pmd, int flags)
764 {
765 unsigned long pfn = pmd_pfn(*pmd);
766 struct mm_struct *mm = vma->vm_mm;
767 struct dev_pagemap *pgmap;
768 struct page *page;
769
770 assert_spin_locked(pmd_lockptr(mm, pmd));
771
772 /*
773 * When we COW a devmap PMD entry, we split it into PTEs, so we should
774 * not be in this function with `flags & FOLL_COW` set.
775 */
776 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
777
778 if (flags & FOLL_WRITE && !pmd_write(*pmd))
779 return NULL;
780
781 if (pmd_present(*pmd) && pmd_devmap(*pmd))
782 /* pass */;
783 else
784 return NULL;
785
786 if (flags & FOLL_TOUCH)
787 touch_pmd(vma, addr, pmd, flags);
788
789 /*
790 * device mapped pages can only be returned if the
791 * caller will manage the page reference count.
792 */
793 if (!(flags & FOLL_GET))
794 return ERR_PTR(-EEXIST);
795
796 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
797 pgmap = get_dev_pagemap(pfn, NULL);
798 if (!pgmap)
799 return ERR_PTR(-EFAULT);
800 page = pfn_to_page(pfn);
801 get_page(page);
802 put_dev_pagemap(pgmap);
803
804 return page;
805 }
806
807 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
808 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
809 struct vm_area_struct *vma)
810 {
811 spinlock_t *dst_ptl, *src_ptl;
812 struct page *src_page;
813 pmd_t pmd;
814 pgtable_t pgtable = NULL;
815 int ret = -ENOMEM;
816
817 /* Skip if can be re-fill on fault */
818 if (!vma_is_anonymous(vma))
819 return 0;
820
821 pgtable = pte_alloc_one(dst_mm, addr);
822 if (unlikely(!pgtable))
823 goto out;
824
825 dst_ptl = pmd_lock(dst_mm, dst_pmd);
826 src_ptl = pmd_lockptr(src_mm, src_pmd);
827 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
828
829 ret = -EAGAIN;
830 pmd = *src_pmd;
831 if (unlikely(!pmd_trans_huge(pmd))) {
832 pte_free(dst_mm, pgtable);
833 goto out_unlock;
834 }
835 /*
836 * When page table lock is held, the huge zero pmd should not be
837 * under splitting since we don't split the page itself, only pmd to
838 * a page table.
839 */
840 if (is_huge_zero_pmd(pmd)) {
841 struct page *zero_page;
842 /*
843 * get_huge_zero_page() will never allocate a new page here,
844 * since we already have a zero page to copy. It just takes a
845 * reference.
846 */
847 zero_page = mm_get_huge_zero_page(dst_mm);
848 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
849 zero_page);
850 ret = 0;
851 goto out_unlock;
852 }
853
854 src_page = pmd_page(pmd);
855 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
856 get_page(src_page);
857 page_dup_rmap(src_page, true);
858 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
859 atomic_long_inc(&dst_mm->nr_ptes);
860 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
861
862 pmdp_set_wrprotect(src_mm, addr, src_pmd);
863 pmd = pmd_mkold(pmd_wrprotect(pmd));
864 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
865
866 ret = 0;
867 out_unlock:
868 spin_unlock(src_ptl);
869 spin_unlock(dst_ptl);
870 out:
871 return ret;
872 }
873
874 void huge_pmd_set_accessed(struct fault_env *fe, pmd_t orig_pmd)
875 {
876 pmd_t entry;
877 unsigned long haddr;
878 bool write = fe->flags & FAULT_FLAG_WRITE;
879
880 fe->ptl = pmd_lock(fe->vma->vm_mm, fe->pmd);
881 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
882 goto unlock;
883
884 entry = pmd_mkyoung(orig_pmd);
885 if (write)
886 entry = pmd_mkdirty(entry);
887 haddr = fe->address & HPAGE_PMD_MASK;
888 if (pmdp_set_access_flags(fe->vma, haddr, fe->pmd, entry, write))
889 update_mmu_cache_pmd(fe->vma, fe->address, fe->pmd);
890
891 unlock:
892 spin_unlock(fe->ptl);
893 }
894
895 static int do_huge_pmd_wp_page_fallback(struct fault_env *fe, pmd_t orig_pmd,
896 struct page *page)
897 {
898 struct vm_area_struct *vma = fe->vma;
899 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
900 struct mem_cgroup *memcg;
901 pgtable_t pgtable;
902 pmd_t _pmd;
903 int ret = 0, i;
904 struct page **pages;
905 unsigned long mmun_start; /* For mmu_notifiers */
906 unsigned long mmun_end; /* For mmu_notifiers */
907
908 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
909 GFP_KERNEL);
910 if (unlikely(!pages)) {
911 ret |= VM_FAULT_OOM;
912 goto out;
913 }
914
915 for (i = 0; i < HPAGE_PMD_NR; i++) {
916 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
917 __GFP_OTHER_NODE, vma,
918 fe->address, page_to_nid(page));
919 if (unlikely(!pages[i] ||
920 mem_cgroup_try_charge(pages[i], vma->vm_mm,
921 GFP_KERNEL, &memcg, false))) {
922 if (pages[i])
923 put_page(pages[i]);
924 while (--i >= 0) {
925 memcg = (void *)page_private(pages[i]);
926 set_page_private(pages[i], 0);
927 mem_cgroup_cancel_charge(pages[i], memcg,
928 false);
929 put_page(pages[i]);
930 }
931 kfree(pages);
932 ret |= VM_FAULT_OOM;
933 goto out;
934 }
935 set_page_private(pages[i], (unsigned long)memcg);
936 }
937
938 for (i = 0; i < HPAGE_PMD_NR; i++) {
939 copy_user_highpage(pages[i], page + i,
940 haddr + PAGE_SIZE * i, vma);
941 __SetPageUptodate(pages[i]);
942 cond_resched();
943 }
944
945 mmun_start = haddr;
946 mmun_end = haddr + HPAGE_PMD_SIZE;
947 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
948
949 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
950 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
951 goto out_free_pages;
952 VM_BUG_ON_PAGE(!PageHead(page), page);
953
954 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
955 /* leave pmd empty until pte is filled */
956
957 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, fe->pmd);
958 pmd_populate(vma->vm_mm, &_pmd, pgtable);
959
960 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
961 pte_t entry;
962 entry = mk_pte(pages[i], vma->vm_page_prot);
963 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
964 memcg = (void *)page_private(pages[i]);
965 set_page_private(pages[i], 0);
966 page_add_new_anon_rmap(pages[i], fe->vma, haddr, false);
967 mem_cgroup_commit_charge(pages[i], memcg, false, false);
968 lru_cache_add_active_or_unevictable(pages[i], vma);
969 fe->pte = pte_offset_map(&_pmd, haddr);
970 VM_BUG_ON(!pte_none(*fe->pte));
971 set_pte_at(vma->vm_mm, haddr, fe->pte, entry);
972 pte_unmap(fe->pte);
973 }
974 kfree(pages);
975
976 smp_wmb(); /* make pte visible before pmd */
977 pmd_populate(vma->vm_mm, fe->pmd, pgtable);
978 page_remove_rmap(page, true);
979 spin_unlock(fe->ptl);
980
981 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
982
983 ret |= VM_FAULT_WRITE;
984 put_page(page);
985
986 out:
987 return ret;
988
989 out_free_pages:
990 spin_unlock(fe->ptl);
991 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
992 for (i = 0; i < HPAGE_PMD_NR; i++) {
993 memcg = (void *)page_private(pages[i]);
994 set_page_private(pages[i], 0);
995 mem_cgroup_cancel_charge(pages[i], memcg, false);
996 put_page(pages[i]);
997 }
998 kfree(pages);
999 goto out;
1000 }
1001
1002 int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd)
1003 {
1004 struct vm_area_struct *vma = fe->vma;
1005 struct page *page = NULL, *new_page;
1006 struct mem_cgroup *memcg;
1007 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1008 unsigned long mmun_start; /* For mmu_notifiers */
1009 unsigned long mmun_end; /* For mmu_notifiers */
1010 gfp_t huge_gfp; /* for allocation and charge */
1011 int ret = 0;
1012
1013 fe->ptl = pmd_lockptr(vma->vm_mm, fe->pmd);
1014 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1015 if (is_huge_zero_pmd(orig_pmd))
1016 goto alloc;
1017 spin_lock(fe->ptl);
1018 if (unlikely(!pmd_same(*fe->pmd, orig_pmd)))
1019 goto out_unlock;
1020
1021 page = pmd_page(orig_pmd);
1022 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1023 /*
1024 * We can only reuse the page if nobody else maps the huge page or it's
1025 * part.
1026 */
1027 if (page_trans_huge_mapcount(page, NULL) == 1) {
1028 pmd_t entry;
1029 entry = pmd_mkyoung(orig_pmd);
1030 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1031 if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1))
1032 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1033 ret |= VM_FAULT_WRITE;
1034 goto out_unlock;
1035 }
1036 get_page(page);
1037 spin_unlock(fe->ptl);
1038 alloc:
1039 if (transparent_hugepage_enabled(vma) &&
1040 !transparent_hugepage_debug_cow()) {
1041 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1042 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1043 } else
1044 new_page = NULL;
1045
1046 if (likely(new_page)) {
1047 prep_transhuge_page(new_page);
1048 } else {
1049 if (!page) {
1050 split_huge_pmd(vma, fe->pmd, fe->address);
1051 ret |= VM_FAULT_FALLBACK;
1052 } else {
1053 ret = do_huge_pmd_wp_page_fallback(fe, orig_pmd, page);
1054 if (ret & VM_FAULT_OOM) {
1055 split_huge_pmd(vma, fe->pmd, fe->address);
1056 ret |= VM_FAULT_FALLBACK;
1057 }
1058 put_page(page);
1059 }
1060 count_vm_event(THP_FAULT_FALLBACK);
1061 goto out;
1062 }
1063
1064 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1065 huge_gfp | __GFP_NORETRY, &memcg, true))) {
1066 put_page(new_page);
1067 split_huge_pmd(vma, fe->pmd, fe->address);
1068 if (page)
1069 put_page(page);
1070 ret |= VM_FAULT_FALLBACK;
1071 count_vm_event(THP_FAULT_FALLBACK);
1072 goto out;
1073 }
1074
1075 count_vm_event(THP_FAULT_ALLOC);
1076
1077 if (!page)
1078 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1079 else
1080 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1081 __SetPageUptodate(new_page);
1082
1083 mmun_start = haddr;
1084 mmun_end = haddr + HPAGE_PMD_SIZE;
1085 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1086
1087 spin_lock(fe->ptl);
1088 if (page)
1089 put_page(page);
1090 if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) {
1091 spin_unlock(fe->ptl);
1092 mem_cgroup_cancel_charge(new_page, memcg, true);
1093 put_page(new_page);
1094 goto out_mn;
1095 } else {
1096 pmd_t entry;
1097 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1098 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1099 pmdp_huge_clear_flush_notify(vma, haddr, fe->pmd);
1100 page_add_new_anon_rmap(new_page, vma, haddr, true);
1101 mem_cgroup_commit_charge(new_page, memcg, false, true);
1102 lru_cache_add_active_or_unevictable(new_page, vma);
1103 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
1104 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1105 if (!page) {
1106 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1107 } else {
1108 VM_BUG_ON_PAGE(!PageHead(page), page);
1109 page_remove_rmap(page, true);
1110 put_page(page);
1111 }
1112 ret |= VM_FAULT_WRITE;
1113 }
1114 spin_unlock(fe->ptl);
1115 out_mn:
1116 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1117 out:
1118 return ret;
1119 out_unlock:
1120 spin_unlock(fe->ptl);
1121 return ret;
1122 }
1123
1124 /*
1125 * FOLL_FORCE can write to even unwritable pmd's, but only
1126 * after we've gone through a COW cycle and they are dirty.
1127 */
1128 static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1129 {
1130 return pmd_write(pmd) ||
1131 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1132 }
1133
1134 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1135 unsigned long addr,
1136 pmd_t *pmd,
1137 unsigned int flags)
1138 {
1139 struct mm_struct *mm = vma->vm_mm;
1140 struct page *page = NULL;
1141
1142 assert_spin_locked(pmd_lockptr(mm, pmd));
1143
1144 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1145 goto out;
1146
1147 /* Avoid dumping huge zero page */
1148 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1149 return ERR_PTR(-EFAULT);
1150
1151 /* Full NUMA hinting faults to serialise migration in fault paths */
1152 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1153 goto out;
1154
1155 page = pmd_page(*pmd);
1156 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1157 if (flags & FOLL_TOUCH)
1158 touch_pmd(vma, addr, pmd, flags);
1159 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1160 /*
1161 * We don't mlock() pte-mapped THPs. This way we can avoid
1162 * leaking mlocked pages into non-VM_LOCKED VMAs.
1163 *
1164 * For anon THP:
1165 *
1166 * In most cases the pmd is the only mapping of the page as we
1167 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1168 * writable private mappings in populate_vma_page_range().
1169 *
1170 * The only scenario when we have the page shared here is if we
1171 * mlocking read-only mapping shared over fork(). We skip
1172 * mlocking such pages.
1173 *
1174 * For file THP:
1175 *
1176 * We can expect PageDoubleMap() to be stable under page lock:
1177 * for file pages we set it in page_add_file_rmap(), which
1178 * requires page to be locked.
1179 */
1180
1181 if (PageAnon(page) && compound_mapcount(page) != 1)
1182 goto skip_mlock;
1183 if (PageDoubleMap(page) || !page->mapping)
1184 goto skip_mlock;
1185 if (!trylock_page(page))
1186 goto skip_mlock;
1187 lru_add_drain();
1188 if (page->mapping && !PageDoubleMap(page))
1189 mlock_vma_page(page);
1190 unlock_page(page);
1191 }
1192 skip_mlock:
1193 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1194 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1195 if (flags & FOLL_GET)
1196 get_page(page);
1197
1198 out:
1199 return page;
1200 }
1201
1202 /* NUMA hinting page fault entry point for trans huge pmds */
1203 int do_huge_pmd_numa_page(struct fault_env *fe, pmd_t pmd)
1204 {
1205 struct vm_area_struct *vma = fe->vma;
1206 struct anon_vma *anon_vma = NULL;
1207 struct page *page;
1208 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
1209 int page_nid = -1, this_nid = numa_node_id();
1210 int target_nid, last_cpupid = -1;
1211 bool page_locked;
1212 bool migrated = false;
1213 bool was_writable;
1214 int flags = 0;
1215
1216 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
1217 if (unlikely(!pmd_same(pmd, *fe->pmd)))
1218 goto out_unlock;
1219
1220 /*
1221 * If there are potential migrations, wait for completion and retry
1222 * without disrupting NUMA hinting information. Do not relock and
1223 * check_same as the page may no longer be mapped.
1224 */
1225 if (unlikely(pmd_trans_migrating(*fe->pmd))) {
1226 page = pmd_page(*fe->pmd);
1227 if (!get_page_unless_zero(page))
1228 goto out_unlock;
1229 spin_unlock(fe->ptl);
1230 wait_on_page_locked(page);
1231 put_page(page);
1232 goto out;
1233 }
1234
1235 page = pmd_page(pmd);
1236 BUG_ON(is_huge_zero_page(page));
1237 page_nid = page_to_nid(page);
1238 last_cpupid = page_cpupid_last(page);
1239 count_vm_numa_event(NUMA_HINT_FAULTS);
1240 if (page_nid == this_nid) {
1241 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1242 flags |= TNF_FAULT_LOCAL;
1243 }
1244
1245 /* See similar comment in do_numa_page for explanation */
1246 if (!pmd_write(pmd))
1247 flags |= TNF_NO_GROUP;
1248
1249 /*
1250 * Acquire the page lock to serialise THP migrations but avoid dropping
1251 * page_table_lock if at all possible
1252 */
1253 page_locked = trylock_page(page);
1254 target_nid = mpol_misplaced(page, vma, haddr);
1255 if (target_nid == -1) {
1256 /* If the page was locked, there are no parallel migrations */
1257 if (page_locked)
1258 goto clear_pmdnuma;
1259 }
1260
1261 /* Migration could have started since the pmd_trans_migrating check */
1262 if (!page_locked) {
1263 page_nid = -1;
1264 if (!get_page_unless_zero(page))
1265 goto out_unlock;
1266 spin_unlock(fe->ptl);
1267 wait_on_page_locked(page);
1268 put_page(page);
1269 goto out;
1270 }
1271
1272 /*
1273 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1274 * to serialises splits
1275 */
1276 get_page(page);
1277 spin_unlock(fe->ptl);
1278 anon_vma = page_lock_anon_vma_read(page);
1279
1280 /* Confirm the PMD did not change while page_table_lock was released */
1281 spin_lock(fe->ptl);
1282 if (unlikely(!pmd_same(pmd, *fe->pmd))) {
1283 unlock_page(page);
1284 put_page(page);
1285 page_nid = -1;
1286 goto out_unlock;
1287 }
1288
1289 /* Bail if we fail to protect against THP splits for any reason */
1290 if (unlikely(!anon_vma)) {
1291 put_page(page);
1292 page_nid = -1;
1293 goto clear_pmdnuma;
1294 }
1295
1296 /*
1297 * Migrate the THP to the requested node, returns with page unlocked
1298 * and access rights restored.
1299 */
1300 spin_unlock(fe->ptl);
1301 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1302 fe->pmd, pmd, fe->address, page, target_nid);
1303 if (migrated) {
1304 flags |= TNF_MIGRATED;
1305 page_nid = target_nid;
1306 } else
1307 flags |= TNF_MIGRATE_FAIL;
1308
1309 goto out;
1310 clear_pmdnuma:
1311 BUG_ON(!PageLocked(page));
1312 was_writable = pmd_write(pmd);
1313 pmd = pmd_modify(pmd, vma->vm_page_prot);
1314 pmd = pmd_mkyoung(pmd);
1315 if (was_writable)
1316 pmd = pmd_mkwrite(pmd);
1317 set_pmd_at(vma->vm_mm, haddr, fe->pmd, pmd);
1318 update_mmu_cache_pmd(vma, fe->address, fe->pmd);
1319 unlock_page(page);
1320 out_unlock:
1321 spin_unlock(fe->ptl);
1322
1323 out:
1324 if (anon_vma)
1325 page_unlock_anon_vma_read(anon_vma);
1326
1327 if (page_nid != -1)
1328 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, fe->flags);
1329
1330 return 0;
1331 }
1332
1333 /*
1334 * Return true if we do MADV_FREE successfully on entire pmd page.
1335 * Otherwise, return false.
1336 */
1337 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1338 pmd_t *pmd, unsigned long addr, unsigned long next)
1339 {
1340 spinlock_t *ptl;
1341 pmd_t orig_pmd;
1342 struct page *page;
1343 struct mm_struct *mm = tlb->mm;
1344 bool ret = false;
1345
1346 ptl = pmd_trans_huge_lock(pmd, vma);
1347 if (!ptl)
1348 goto out_unlocked;
1349
1350 orig_pmd = *pmd;
1351 if (is_huge_zero_pmd(orig_pmd))
1352 goto out;
1353
1354 page = pmd_page(orig_pmd);
1355 /*
1356 * If other processes are mapping this page, we couldn't discard
1357 * the page unless they all do MADV_FREE so let's skip the page.
1358 */
1359 if (page_mapcount(page) != 1)
1360 goto out;
1361
1362 if (!trylock_page(page))
1363 goto out;
1364
1365 /*
1366 * If user want to discard part-pages of THP, split it so MADV_FREE
1367 * will deactivate only them.
1368 */
1369 if (next - addr != HPAGE_PMD_SIZE) {
1370 get_page(page);
1371 spin_unlock(ptl);
1372 split_huge_page(page);
1373 unlock_page(page);
1374 put_page(page);
1375 goto out_unlocked;
1376 }
1377
1378 if (PageDirty(page))
1379 ClearPageDirty(page);
1380 unlock_page(page);
1381
1382 if (PageActive(page))
1383 deactivate_page(page);
1384
1385 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1386 pmdp_invalidate(vma, addr, pmd);
1387 orig_pmd = pmd_mkold(orig_pmd);
1388 orig_pmd = pmd_mkclean(orig_pmd);
1389
1390 set_pmd_at(mm, addr, pmd, orig_pmd);
1391 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1392 }
1393 ret = true;
1394 out:
1395 spin_unlock(ptl);
1396 out_unlocked:
1397 return ret;
1398 }
1399
1400 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1401 pmd_t *pmd, unsigned long addr)
1402 {
1403 pmd_t orig_pmd;
1404 spinlock_t *ptl;
1405
1406 ptl = __pmd_trans_huge_lock(pmd, vma);
1407 if (!ptl)
1408 return 0;
1409 /*
1410 * For architectures like ppc64 we look at deposited pgtable
1411 * when calling pmdp_huge_get_and_clear. So do the
1412 * pgtable_trans_huge_withdraw after finishing pmdp related
1413 * operations.
1414 */
1415 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1416 tlb->fullmm);
1417 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1418 if (vma_is_dax(vma)) {
1419 spin_unlock(ptl);
1420 if (is_huge_zero_pmd(orig_pmd))
1421 tlb_remove_page(tlb, pmd_page(orig_pmd));
1422 } else if (is_huge_zero_pmd(orig_pmd)) {
1423 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1424 atomic_long_dec(&tlb->mm->nr_ptes);
1425 spin_unlock(ptl);
1426 tlb_remove_page(tlb, pmd_page(orig_pmd));
1427 } else {
1428 struct page *page = pmd_page(orig_pmd);
1429 page_remove_rmap(page, true);
1430 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1431 VM_BUG_ON_PAGE(!PageHead(page), page);
1432 if (PageAnon(page)) {
1433 pgtable_t pgtable;
1434 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1435 pte_free(tlb->mm, pgtable);
1436 atomic_long_dec(&tlb->mm->nr_ptes);
1437 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1438 } else {
1439 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1440 }
1441 spin_unlock(ptl);
1442 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1443 }
1444 return 1;
1445 }
1446
1447 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1448 unsigned long new_addr, unsigned long old_end,
1449 pmd_t *old_pmd, pmd_t *new_pmd)
1450 {
1451 spinlock_t *old_ptl, *new_ptl;
1452 pmd_t pmd;
1453 struct mm_struct *mm = vma->vm_mm;
1454 bool force_flush = false;
1455
1456 if ((old_addr & ~HPAGE_PMD_MASK) ||
1457 (new_addr & ~HPAGE_PMD_MASK) ||
1458 old_end - old_addr < HPAGE_PMD_SIZE)
1459 return false;
1460
1461 /*
1462 * The destination pmd shouldn't be established, free_pgtables()
1463 * should have release it.
1464 */
1465 if (WARN_ON(!pmd_none(*new_pmd))) {
1466 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1467 return false;
1468 }
1469
1470 /*
1471 * We don't have to worry about the ordering of src and dst
1472 * ptlocks because exclusive mmap_sem prevents deadlock.
1473 */
1474 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1475 if (old_ptl) {
1476 new_ptl = pmd_lockptr(mm, new_pmd);
1477 if (new_ptl != old_ptl)
1478 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1479 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1480 if (pmd_present(pmd))
1481 force_flush = true;
1482 VM_BUG_ON(!pmd_none(*new_pmd));
1483
1484 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1485 vma_is_anonymous(vma)) {
1486 pgtable_t pgtable;
1487 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1488 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1489 }
1490 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1491 if (force_flush)
1492 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1493 if (new_ptl != old_ptl)
1494 spin_unlock(new_ptl);
1495 spin_unlock(old_ptl);
1496 return true;
1497 }
1498 return false;
1499 }
1500
1501 /*
1502 * Returns
1503 * - 0 if PMD could not be locked
1504 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1505 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1506 */
1507 int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1508 unsigned long addr, pgprot_t newprot, int prot_numa)
1509 {
1510 struct mm_struct *mm = vma->vm_mm;
1511 spinlock_t *ptl;
1512 pmd_t entry;
1513 bool preserve_write;
1514 int ret;
1515
1516 ptl = __pmd_trans_huge_lock(pmd, vma);
1517 if (!ptl)
1518 return 0;
1519
1520 preserve_write = prot_numa && pmd_write(*pmd);
1521 ret = 1;
1522
1523 /*
1524 * Avoid trapping faults against the zero page. The read-only
1525 * data is likely to be read-cached on the local CPU and
1526 * local/remote hits to the zero page are not interesting.
1527 */
1528 if (prot_numa && is_huge_zero_pmd(*pmd))
1529 goto unlock;
1530
1531 if (prot_numa && pmd_protnone(*pmd))
1532 goto unlock;
1533
1534 /*
1535 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1536 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1537 * which is also under down_read(mmap_sem):
1538 *
1539 * CPU0: CPU1:
1540 * change_huge_pmd(prot_numa=1)
1541 * pmdp_huge_get_and_clear_notify()
1542 * madvise_dontneed()
1543 * zap_pmd_range()
1544 * pmd_trans_huge(*pmd) == 0 (without ptl)
1545 * // skip the pmd
1546 * set_pmd_at();
1547 * // pmd is re-established
1548 *
1549 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1550 * which may break userspace.
1551 *
1552 * pmdp_invalidate() is required to make sure we don't miss
1553 * dirty/young flags set by hardware.
1554 */
1555 entry = *pmd;
1556 pmdp_invalidate(vma, addr, pmd);
1557
1558 /*
1559 * Recover dirty/young flags. It relies on pmdp_invalidate to not
1560 * corrupt them.
1561 */
1562 if (pmd_dirty(*pmd))
1563 entry = pmd_mkdirty(entry);
1564 if (pmd_young(*pmd))
1565 entry = pmd_mkyoung(entry);
1566
1567 entry = pmd_modify(entry, newprot);
1568 if (preserve_write)
1569 entry = pmd_mkwrite(entry);
1570 ret = HPAGE_PMD_NR;
1571 set_pmd_at(mm, addr, pmd, entry);
1572 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1573 unlock:
1574 spin_unlock(ptl);
1575 return ret;
1576 }
1577
1578 /*
1579 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1580 *
1581 * Note that if it returns page table lock pointer, this routine returns without
1582 * unlocking page table lock. So callers must unlock it.
1583 */
1584 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1585 {
1586 spinlock_t *ptl;
1587 ptl = pmd_lock(vma->vm_mm, pmd);
1588 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1589 return ptl;
1590 spin_unlock(ptl);
1591 return NULL;
1592 }
1593
1594 static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1595 unsigned long haddr, pmd_t *pmd)
1596 {
1597 struct mm_struct *mm = vma->vm_mm;
1598 pgtable_t pgtable;
1599 pmd_t _pmd;
1600 int i;
1601
1602 /* leave pmd empty until pte is filled */
1603 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1604
1605 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1606 pmd_populate(mm, &_pmd, pgtable);
1607
1608 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1609 pte_t *pte, entry;
1610 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1611 entry = pte_mkspecial(entry);
1612 pte = pte_offset_map(&_pmd, haddr);
1613 VM_BUG_ON(!pte_none(*pte));
1614 set_pte_at(mm, haddr, pte, entry);
1615 pte_unmap(pte);
1616 }
1617 smp_wmb(); /* make pte visible before pmd */
1618 pmd_populate(mm, pmd, pgtable);
1619 }
1620
1621 static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1622 unsigned long haddr, bool freeze)
1623 {
1624 struct mm_struct *mm = vma->vm_mm;
1625 struct page *page;
1626 pgtable_t pgtable;
1627 pmd_t _pmd;
1628 bool young, write, dirty, soft_dirty;
1629 unsigned long addr;
1630 int i;
1631
1632 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1633 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1634 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1635 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1636
1637 count_vm_event(THP_SPLIT_PMD);
1638
1639 if (!vma_is_anonymous(vma)) {
1640 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1641 if (vma_is_dax(vma))
1642 return;
1643 page = pmd_page(_pmd);
1644 if (!PageDirty(page) && pmd_dirty(_pmd))
1645 set_page_dirty(page);
1646 if (!PageReferenced(page) && pmd_young(_pmd))
1647 SetPageReferenced(page);
1648 page_remove_rmap(page, true);
1649 put_page(page);
1650 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1651 return;
1652 } else if (is_huge_zero_pmd(*pmd)) {
1653 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1654 }
1655
1656 page = pmd_page(*pmd);
1657 VM_BUG_ON_PAGE(!page_count(page), page);
1658 page_ref_add(page, HPAGE_PMD_NR - 1);
1659 write = pmd_write(*pmd);
1660 young = pmd_young(*pmd);
1661 dirty = pmd_dirty(*pmd);
1662 soft_dirty = pmd_soft_dirty(*pmd);
1663
1664 pmdp_huge_split_prepare(vma, haddr, pmd);
1665 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1666 pmd_populate(mm, &_pmd, pgtable);
1667
1668 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1669 pte_t entry, *pte;
1670 /*
1671 * Note that NUMA hinting access restrictions are not
1672 * transferred to avoid any possibility of altering
1673 * permissions across VMAs.
1674 */
1675 if (freeze) {
1676 swp_entry_t swp_entry;
1677 swp_entry = make_migration_entry(page + i, write);
1678 entry = swp_entry_to_pte(swp_entry);
1679 if (soft_dirty)
1680 entry = pte_swp_mksoft_dirty(entry);
1681 } else {
1682 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1683 entry = maybe_mkwrite(entry, vma);
1684 if (!write)
1685 entry = pte_wrprotect(entry);
1686 if (!young)
1687 entry = pte_mkold(entry);
1688 if (soft_dirty)
1689 entry = pte_mksoft_dirty(entry);
1690 }
1691 if (dirty)
1692 SetPageDirty(page + i);
1693 pte = pte_offset_map(&_pmd, addr);
1694 BUG_ON(!pte_none(*pte));
1695 set_pte_at(mm, addr, pte, entry);
1696 atomic_inc(&page[i]._mapcount);
1697 pte_unmap(pte);
1698 }
1699
1700 /*
1701 * Set PG_double_map before dropping compound_mapcount to avoid
1702 * false-negative page_mapped().
1703 */
1704 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1705 for (i = 0; i < HPAGE_PMD_NR; i++)
1706 atomic_inc(&page[i]._mapcount);
1707 }
1708
1709 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1710 /* Last compound_mapcount is gone. */
1711 __dec_node_page_state(page, NR_ANON_THPS);
1712 if (TestClearPageDoubleMap(page)) {
1713 /* No need in mapcount reference anymore */
1714 for (i = 0; i < HPAGE_PMD_NR; i++)
1715 atomic_dec(&page[i]._mapcount);
1716 }
1717 }
1718
1719 smp_wmb(); /* make pte visible before pmd */
1720 /*
1721 * Up to this point the pmd is present and huge and userland has the
1722 * whole access to the hugepage during the split (which happens in
1723 * place). If we overwrite the pmd with the not-huge version pointing
1724 * to the pte here (which of course we could if all CPUs were bug
1725 * free), userland could trigger a small page size TLB miss on the
1726 * small sized TLB while the hugepage TLB entry is still established in
1727 * the huge TLB. Some CPU doesn't like that.
1728 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1729 * 383 on page 93. Intel should be safe but is also warns that it's
1730 * only safe if the permission and cache attributes of the two entries
1731 * loaded in the two TLB is identical (which should be the case here).
1732 * But it is generally safer to never allow small and huge TLB entries
1733 * for the same virtual address to be loaded simultaneously. So instead
1734 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1735 * current pmd notpresent (atomically because here the pmd_trans_huge
1736 * and pmd_trans_splitting must remain set at all times on the pmd
1737 * until the split is complete for this pmd), then we flush the SMP TLB
1738 * and finally we write the non-huge version of the pmd entry with
1739 * pmd_populate.
1740 */
1741 pmdp_invalidate(vma, haddr, pmd);
1742 pmd_populate(mm, pmd, pgtable);
1743
1744 if (freeze) {
1745 for (i = 0; i < HPAGE_PMD_NR; i++) {
1746 page_remove_rmap(page + i, false);
1747 put_page(page + i);
1748 }
1749 }
1750 }
1751
1752 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1753 unsigned long address, bool freeze, struct page *page)
1754 {
1755 spinlock_t *ptl;
1756 struct mm_struct *mm = vma->vm_mm;
1757 unsigned long haddr = address & HPAGE_PMD_MASK;
1758
1759 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1760 ptl = pmd_lock(mm, pmd);
1761
1762 /*
1763 * If caller asks to setup a migration entries, we need a page to check
1764 * pmd against. Otherwise we can end up replacing wrong page.
1765 */
1766 VM_BUG_ON(freeze && !page);
1767 if (page && page != pmd_page(*pmd))
1768 goto out;
1769
1770 if (pmd_trans_huge(*pmd)) {
1771 page = pmd_page(*pmd);
1772 if (PageMlocked(page))
1773 clear_page_mlock(page);
1774 } else if (!pmd_devmap(*pmd))
1775 goto out;
1776 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1777 out:
1778 spin_unlock(ptl);
1779 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1780 }
1781
1782 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1783 bool freeze, struct page *page)
1784 {
1785 pgd_t *pgd;
1786 pud_t *pud;
1787 pmd_t *pmd;
1788
1789 pgd = pgd_offset(vma->vm_mm, address);
1790 if (!pgd_present(*pgd))
1791 return;
1792
1793 pud = pud_offset(pgd, address);
1794 if (!pud_present(*pud))
1795 return;
1796
1797 pmd = pmd_offset(pud, address);
1798
1799 __split_huge_pmd(vma, pmd, address, freeze, page);
1800 }
1801
1802 void vma_adjust_trans_huge(struct vm_area_struct *vma,
1803 unsigned long start,
1804 unsigned long end,
1805 long adjust_next)
1806 {
1807 /*
1808 * If the new start address isn't hpage aligned and it could
1809 * previously contain an hugepage: check if we need to split
1810 * an huge pmd.
1811 */
1812 if (start & ~HPAGE_PMD_MASK &&
1813 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1814 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1815 split_huge_pmd_address(vma, start, false, NULL);
1816
1817 /*
1818 * If the new end address isn't hpage aligned and it could
1819 * previously contain an hugepage: check if we need to split
1820 * an huge pmd.
1821 */
1822 if (end & ~HPAGE_PMD_MASK &&
1823 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1824 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1825 split_huge_pmd_address(vma, end, false, NULL);
1826
1827 /*
1828 * If we're also updating the vma->vm_next->vm_start, if the new
1829 * vm_next->vm_start isn't page aligned and it could previously
1830 * contain an hugepage: check if we need to split an huge pmd.
1831 */
1832 if (adjust_next > 0) {
1833 struct vm_area_struct *next = vma->vm_next;
1834 unsigned long nstart = next->vm_start;
1835 nstart += adjust_next << PAGE_SHIFT;
1836 if (nstart & ~HPAGE_PMD_MASK &&
1837 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1838 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1839 split_huge_pmd_address(next, nstart, false, NULL);
1840 }
1841 }
1842
1843 static void unmap_page(struct page *page)
1844 {
1845 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1846 TTU_RMAP_LOCKED;
1847 int i, ret;
1848
1849 VM_BUG_ON_PAGE(!PageHead(page), page);
1850
1851 if (PageAnon(page))
1852 ttu_flags |= TTU_MIGRATION;
1853
1854 /* We only need TTU_SPLIT_HUGE_PMD once */
1855 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1856 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1857 /* Cut short if the page is unmapped */
1858 if (page_count(page) == 1)
1859 return;
1860
1861 ret = try_to_unmap(page + i, ttu_flags);
1862 }
1863 VM_BUG_ON_PAGE(ret, page + i - 1);
1864 }
1865
1866 static void remap_page(struct page *page)
1867 {
1868 int i;
1869
1870 for (i = 0; i < HPAGE_PMD_NR; i++)
1871 remove_migration_ptes(page + i, page + i, true);
1872 }
1873
1874 static void __split_huge_page_tail(struct page *head, int tail,
1875 struct lruvec *lruvec, struct list_head *list)
1876 {
1877 struct page *page_tail = head + tail;
1878
1879 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1880
1881 /*
1882 * Clone page flags before unfreezing refcount.
1883 *
1884 * After successful get_page_unless_zero() might follow flags change,
1885 * for exmaple lock_page() which set PG_waiters.
1886 */
1887 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1888 page_tail->flags |= (head->flags &
1889 ((1L << PG_referenced) |
1890 (1L << PG_swapbacked) |
1891 (1L << PG_mlocked) |
1892 (1L << PG_uptodate) |
1893 (1L << PG_active) |
1894 (1L << PG_locked) |
1895 (1L << PG_unevictable) |
1896 (1L << PG_dirty)));
1897
1898 /* ->mapping in first tail page is compound_mapcount */
1899 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1900 page_tail);
1901 page_tail->mapping = head->mapping;
1902 page_tail->index = head->index + tail;
1903
1904 /* Page flags must be visible before we make the page non-compound. */
1905 smp_wmb();
1906
1907 /*
1908 * Clear PageTail before unfreezing page refcount.
1909 *
1910 * After successful get_page_unless_zero() might follow put_page()
1911 * which needs correct compound_head().
1912 */
1913 clear_compound_head(page_tail);
1914
1915 /* Finally unfreeze refcount. Additional reference from page cache. */
1916 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
1917 PageSwapCache(head)));
1918
1919 if (page_is_young(head))
1920 set_page_young(page_tail);
1921 if (page_is_idle(head))
1922 set_page_idle(page_tail);
1923
1924 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1925 lru_add_page_tail(head, page_tail, lruvec, list);
1926 }
1927
1928 static void __split_huge_page(struct page *page, struct list_head *list,
1929 pgoff_t end, unsigned long flags)
1930 {
1931 struct page *head = compound_head(page);
1932 struct zone *zone = page_zone(head);
1933 struct lruvec *lruvec;
1934 int i;
1935
1936 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1937
1938 /* complete memcg works before add pages to LRU */
1939 mem_cgroup_split_huge_fixup(head);
1940
1941 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1942 __split_huge_page_tail(head, i, lruvec, list);
1943 /* Some pages can be beyond i_size: drop them from page cache */
1944 if (head[i].index >= end) {
1945 __ClearPageDirty(head + i);
1946 __delete_from_page_cache(head + i, NULL);
1947 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1948 shmem_uncharge(head->mapping->host, 1);
1949 put_page(head + i);
1950 }
1951 }
1952
1953 ClearPageCompound(head);
1954
1955 split_page_owner(head, HPAGE_PMD_ORDER);
1956
1957 /* See comment in __split_huge_page_tail() */
1958 if (PageAnon(head)) {
1959 page_ref_inc(head);
1960 } else {
1961 /* Additional pin to radix tree */
1962 page_ref_add(head, 2);
1963 spin_unlock(&head->mapping->tree_lock);
1964 }
1965
1966 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1967
1968 remap_page(head);
1969
1970 for (i = 0; i < HPAGE_PMD_NR; i++) {
1971 struct page *subpage = head + i;
1972 if (subpage == page)
1973 continue;
1974 unlock_page(subpage);
1975
1976 /*
1977 * Subpages may be freed if there wasn't any mapping
1978 * like if add_to_swap() is running on a lru page that
1979 * had its mapping zapped. And freeing these pages
1980 * requires taking the lru_lock so we do the put_page
1981 * of the tail pages after the split is complete.
1982 */
1983 put_page(subpage);
1984 }
1985 }
1986
1987 int total_mapcount(struct page *page)
1988 {
1989 int i, compound, ret;
1990
1991 VM_BUG_ON_PAGE(PageTail(page), page);
1992
1993 if (likely(!PageCompound(page)))
1994 return atomic_read(&page->_mapcount) + 1;
1995
1996 compound = compound_mapcount(page);
1997 if (PageHuge(page))
1998 return compound;
1999 ret = compound;
2000 for (i = 0; i < HPAGE_PMD_NR; i++)
2001 ret += atomic_read(&page[i]._mapcount) + 1;
2002 /* File pages has compound_mapcount included in _mapcount */
2003 if (!PageAnon(page))
2004 return ret - compound * HPAGE_PMD_NR;
2005 if (PageDoubleMap(page))
2006 ret -= HPAGE_PMD_NR;
2007 return ret;
2008 }
2009
2010 /*
2011 * This calculates accurately how many mappings a transparent hugepage
2012 * has (unlike page_mapcount() which isn't fully accurate). This full
2013 * accuracy is primarily needed to know if copy-on-write faults can
2014 * reuse the page and change the mapping to read-write instead of
2015 * copying them. At the same time this returns the total_mapcount too.
2016 *
2017 * The function returns the highest mapcount any one of the subpages
2018 * has. If the return value is one, even if different processes are
2019 * mapping different subpages of the transparent hugepage, they can
2020 * all reuse it, because each process is reusing a different subpage.
2021 *
2022 * The total_mapcount is instead counting all virtual mappings of the
2023 * subpages. If the total_mapcount is equal to "one", it tells the
2024 * caller all mappings belong to the same "mm" and in turn the
2025 * anon_vma of the transparent hugepage can become the vma->anon_vma
2026 * local one as no other process may be mapping any of the subpages.
2027 *
2028 * It would be more accurate to replace page_mapcount() with
2029 * page_trans_huge_mapcount(), however we only use
2030 * page_trans_huge_mapcount() in the copy-on-write faults where we
2031 * need full accuracy to avoid breaking page pinning, because
2032 * page_trans_huge_mapcount() is slower than page_mapcount().
2033 */
2034 int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2035 {
2036 int i, ret, _total_mapcount, mapcount;
2037
2038 /* hugetlbfs shouldn't call it */
2039 VM_BUG_ON_PAGE(PageHuge(page), page);
2040
2041 if (likely(!PageTransCompound(page))) {
2042 mapcount = atomic_read(&page->_mapcount) + 1;
2043 if (total_mapcount)
2044 *total_mapcount = mapcount;
2045 return mapcount;
2046 }
2047
2048 page = compound_head(page);
2049
2050 _total_mapcount = ret = 0;
2051 for (i = 0; i < HPAGE_PMD_NR; i++) {
2052 mapcount = atomic_read(&page[i]._mapcount) + 1;
2053 ret = max(ret, mapcount);
2054 _total_mapcount += mapcount;
2055 }
2056 if (PageDoubleMap(page)) {
2057 ret -= 1;
2058 _total_mapcount -= HPAGE_PMD_NR;
2059 }
2060 mapcount = compound_mapcount(page);
2061 ret += mapcount;
2062 _total_mapcount += mapcount;
2063 if (total_mapcount)
2064 *total_mapcount = _total_mapcount;
2065 return ret;
2066 }
2067
2068 /*
2069 * This function splits huge page into normal pages. @page can point to any
2070 * subpage of huge page to split. Split doesn't change the position of @page.
2071 *
2072 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2073 * The huge page must be locked.
2074 *
2075 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2076 *
2077 * Both head page and tail pages will inherit mapping, flags, and so on from
2078 * the hugepage.
2079 *
2080 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2081 * they are not mapped.
2082 *
2083 * Returns 0 if the hugepage is split successfully.
2084 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2085 * us.
2086 */
2087 int split_huge_page_to_list(struct page *page, struct list_head *list)
2088 {
2089 struct page *head = compound_head(page);
2090 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2091 struct anon_vma *anon_vma = NULL;
2092 struct address_space *mapping = NULL;
2093 int count, mapcount, extra_pins, ret;
2094 bool mlocked;
2095 unsigned long flags;
2096 pgoff_t end;
2097
2098 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2099 VM_BUG_ON_PAGE(!PageLocked(page), page);
2100 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2101 VM_BUG_ON_PAGE(!PageCompound(page), page);
2102
2103 if (PageAnon(head)) {
2104 /*
2105 * The caller does not necessarily hold an mmap_sem that would
2106 * prevent the anon_vma disappearing so we first we take a
2107 * reference to it and then lock the anon_vma for write. This
2108 * is similar to page_lock_anon_vma_read except the write lock
2109 * is taken to serialise against parallel split or collapse
2110 * operations.
2111 */
2112 anon_vma = page_get_anon_vma(head);
2113 if (!anon_vma) {
2114 ret = -EBUSY;
2115 goto out;
2116 }
2117 extra_pins = 0;
2118 end = -1;
2119 mapping = NULL;
2120 anon_vma_lock_write(anon_vma);
2121 } else {
2122 mapping = head->mapping;
2123
2124 /* Truncated ? */
2125 if (!mapping) {
2126 ret = -EBUSY;
2127 goto out;
2128 }
2129
2130 /* Addidional pins from radix tree */
2131 extra_pins = HPAGE_PMD_NR;
2132 anon_vma = NULL;
2133 i_mmap_lock_read(mapping);
2134
2135 /*
2136 *__split_huge_page() may need to trim off pages beyond EOF:
2137 * but on 32-bit, i_size_read() takes an irq-unsafe seqlock,
2138 * which cannot be nested inside the page tree lock. So note
2139 * end now: i_size itself may be changed at any moment, but
2140 * head page lock is good enough to serialize the trimming.
2141 */
2142 end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
2143 }
2144
2145 /*
2146 * Racy check if we can split the page, before unmap_page() will
2147 * split PMDs
2148 */
2149 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2150 ret = -EBUSY;
2151 goto out_unlock;
2152 }
2153
2154 mlocked = PageMlocked(page);
2155 unmap_page(head);
2156 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2157
2158 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2159 if (mlocked)
2160 lru_add_drain();
2161
2162 /* prevent PageLRU to go away from under us, and freeze lru stats */
2163 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2164
2165 if (mapping) {
2166 void **pslot;
2167
2168 spin_lock(&mapping->tree_lock);
2169 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2170 page_index(head));
2171 /*
2172 * Check if the head page is present in radix tree.
2173 * We assume all tail are present too, if head is there.
2174 */
2175 if (radix_tree_deref_slot_protected(pslot,
2176 &mapping->tree_lock) != head)
2177 goto fail;
2178 }
2179
2180 /* Prevent deferred_split_scan() touching ->_refcount */
2181 spin_lock(&pgdata->split_queue_lock);
2182 count = page_count(head);
2183 mapcount = total_mapcount(head);
2184 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2185 if (!list_empty(page_deferred_list(head))) {
2186 pgdata->split_queue_len--;
2187 list_del(page_deferred_list(head));
2188 }
2189 if (mapping)
2190 __dec_node_page_state(page, NR_SHMEM_THPS);
2191 spin_unlock(&pgdata->split_queue_lock);
2192 __split_huge_page(page, list, end, flags);
2193 ret = 0;
2194 } else {
2195 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2196 pr_alert("total_mapcount: %u, page_count(): %u\n",
2197 mapcount, count);
2198 if (PageTail(page))
2199 dump_page(head, NULL);
2200 dump_page(page, "total_mapcount(head) > 0");
2201 BUG();
2202 }
2203 spin_unlock(&pgdata->split_queue_lock);
2204 fail: if (mapping)
2205 spin_unlock(&mapping->tree_lock);
2206 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2207 remap_page(head);
2208 ret = -EBUSY;
2209 }
2210
2211 out_unlock:
2212 if (anon_vma) {
2213 anon_vma_unlock_write(anon_vma);
2214 put_anon_vma(anon_vma);
2215 }
2216 if (mapping)
2217 i_mmap_unlock_read(mapping);
2218 out:
2219 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2220 return ret;
2221 }
2222
2223 void free_transhuge_page(struct page *page)
2224 {
2225 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2226 unsigned long flags;
2227
2228 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2229 if (!list_empty(page_deferred_list(page))) {
2230 pgdata->split_queue_len--;
2231 list_del(page_deferred_list(page));
2232 }
2233 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2234 free_compound_page(page);
2235 }
2236
2237 void deferred_split_huge_page(struct page *page)
2238 {
2239 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2240 unsigned long flags;
2241
2242 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2243
2244 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2245 if (list_empty(page_deferred_list(page))) {
2246 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2247 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2248 pgdata->split_queue_len++;
2249 }
2250 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2251 }
2252
2253 static unsigned long deferred_split_count(struct shrinker *shrink,
2254 struct shrink_control *sc)
2255 {
2256 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2257 return ACCESS_ONCE(pgdata->split_queue_len);
2258 }
2259
2260 static unsigned long deferred_split_scan(struct shrinker *shrink,
2261 struct shrink_control *sc)
2262 {
2263 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2264 unsigned long flags;
2265 LIST_HEAD(list), *pos, *next;
2266 struct page *page;
2267 int split = 0;
2268
2269 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2270 /* Take pin on all head pages to avoid freeing them under us */
2271 list_for_each_safe(pos, next, &pgdata->split_queue) {
2272 page = list_entry((void *)pos, struct page, mapping);
2273 page = compound_head(page);
2274 if (get_page_unless_zero(page)) {
2275 list_move(page_deferred_list(page), &list);
2276 } else {
2277 /* We lost race with put_compound_page() */
2278 list_del_init(page_deferred_list(page));
2279 pgdata->split_queue_len--;
2280 }
2281 if (!--sc->nr_to_scan)
2282 break;
2283 }
2284 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2285
2286 list_for_each_safe(pos, next, &list) {
2287 page = list_entry((void *)pos, struct page, mapping);
2288 if (!trylock_page(page))
2289 goto next;
2290 /* split_huge_page() removes page from list on success */
2291 if (!split_huge_page(page))
2292 split++;
2293 unlock_page(page);
2294 next:
2295 put_page(page);
2296 }
2297
2298 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2299 list_splice_tail(&list, &pgdata->split_queue);
2300 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2301
2302 /*
2303 * Stop shrinker if we didn't split any page, but the queue is empty.
2304 * This can happen if pages were freed under us.
2305 */
2306 if (!split && list_empty(&pgdata->split_queue))
2307 return SHRINK_STOP;
2308 return split;
2309 }
2310
2311 static struct shrinker deferred_split_shrinker = {
2312 .count_objects = deferred_split_count,
2313 .scan_objects = deferred_split_scan,
2314 .seeks = DEFAULT_SEEKS,
2315 .flags = SHRINKER_NUMA_AWARE,
2316 };
2317
2318 #ifdef CONFIG_DEBUG_FS
2319 static int split_huge_pages_set(void *data, u64 val)
2320 {
2321 struct zone *zone;
2322 struct page *page;
2323 unsigned long pfn, max_zone_pfn;
2324 unsigned long total = 0, split = 0;
2325
2326 if (val != 1)
2327 return -EINVAL;
2328
2329 for_each_populated_zone(zone) {
2330 max_zone_pfn = zone_end_pfn(zone);
2331 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2332 if (!pfn_valid(pfn))
2333 continue;
2334
2335 page = pfn_to_page(pfn);
2336 if (!get_page_unless_zero(page))
2337 continue;
2338
2339 if (zone != page_zone(page))
2340 goto next;
2341
2342 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2343 goto next;
2344
2345 total++;
2346 lock_page(page);
2347 if (!split_huge_page(page))
2348 split++;
2349 unlock_page(page);
2350 next:
2351 put_page(page);
2352 }
2353 }
2354
2355 pr_info("%lu of %lu THP split\n", split, total);
2356
2357 return 0;
2358 }
2359 DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2360 "%llu\n");
2361
2362 static int __init split_huge_pages_debugfs(void)
2363 {
2364 void *ret;
2365
2366 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2367 &split_huge_pages_fops);
2368 if (!ret)
2369 pr_warn("Failed to create split_huge_pages in debugfs");
2370 return 0;
2371 }
2372 late_initcall(split_huge_pages_debugfs);
2373 #endif
Linux with added FPC-III support