Loading...
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/sched/coredump.h>
13#include <linux/sched/numa_balancing.h>
14#include <linux/highmem.h>
15#include <linux/hugetlb.h>
16#include <linux/mmu_notifier.h>
17#include <linux/rmap.h>
18#include <linux/swap.h>
19#include <linux/shrinker.h>
20#include <linux/mm_inline.h>
21#include <linux/swapops.h>
22#include <linux/dax.h>
23#include <linux/khugepaged.h>
24#include <linux/freezer.h>
25#include <linux/pfn_t.h>
26#include <linux/mman.h>
27#include <linux/memremap.h>
28#include <linux/pagemap.h>
29#include <linux/debugfs.h>
30#include <linux/migrate.h>
31#include <linux/hashtable.h>
32#include <linux/userfaultfd_k.h>
33#include <linux/page_idle.h>
34#include <linux/shmem_fs.h>
35#include <linux/oom.h>
36
37#include <asm/tlb.h>
38#include <asm/pgalloc.h>
39#include "internal.h"
40
41/*
42 * By default, transparent hugepage support is disabled in order to avoid
43 * risking an increased memory footprint for applications that are not
44 * guaranteed to benefit from it. When transparent hugepage support is
45 * enabled, it is for all mappings, and khugepaged scans all mappings.
46 * Defrag is invoked by khugepaged hugepage allocations and by page faults
47 * for all hugepage allocations.
48 */
49unsigned long transparent_hugepage_flags __read_mostly =
50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
51 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
52#endif
53#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
54 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
55#endif
56 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
57 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
58 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
59
60static struct shrinker deferred_split_shrinker;
61
62static atomic_t huge_zero_refcount;
63struct page *huge_zero_page __read_mostly;
64
65static struct page *get_huge_zero_page(void)
66{
67 struct page *zero_page;
68retry:
69 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
70 return READ_ONCE(huge_zero_page);
71
72 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
73 HPAGE_PMD_ORDER);
74 if (!zero_page) {
75 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
76 return NULL;
77 }
78 count_vm_event(THP_ZERO_PAGE_ALLOC);
79 preempt_disable();
80 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
81 preempt_enable();
82 __free_pages(zero_page, compound_order(zero_page));
83 goto retry;
84 }
85
86 /* We take additional reference here. It will be put back by shrinker */
87 atomic_set(&huge_zero_refcount, 2);
88 preempt_enable();
89 return READ_ONCE(huge_zero_page);
90}
91
92static void put_huge_zero_page(void)
93{
94 /*
95 * Counter should never go to zero here. Only shrinker can put
96 * last reference.
97 */
98 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
99}
100
101struct page *mm_get_huge_zero_page(struct mm_struct *mm)
102{
103 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
104 return READ_ONCE(huge_zero_page);
105
106 if (!get_huge_zero_page())
107 return NULL;
108
109 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
110 put_huge_zero_page();
111
112 return READ_ONCE(huge_zero_page);
113}
114
115void mm_put_huge_zero_page(struct mm_struct *mm)
116{
117 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
118 put_huge_zero_page();
119}
120
121static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
122 struct shrink_control *sc)
123{
124 /* we can free zero page only if last reference remains */
125 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
126}
127
128static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
129 struct shrink_control *sc)
130{
131 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
132 struct page *zero_page = xchg(&huge_zero_page, NULL);
133 BUG_ON(zero_page == NULL);
134 __free_pages(zero_page, compound_order(zero_page));
135 return HPAGE_PMD_NR;
136 }
137
138 return 0;
139}
140
141static struct shrinker huge_zero_page_shrinker = {
142 .count_objects = shrink_huge_zero_page_count,
143 .scan_objects = shrink_huge_zero_page_scan,
144 .seeks = DEFAULT_SEEKS,
145};
146
147#ifdef CONFIG_SYSFS
148static ssize_t enabled_show(struct kobject *kobj,
149 struct kobj_attribute *attr, char *buf)
150{
151 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
152 return sprintf(buf, "[always] madvise never\n");
153 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
154 return sprintf(buf, "always [madvise] never\n");
155 else
156 return sprintf(buf, "always madvise [never]\n");
157}
158
159static ssize_t enabled_store(struct kobject *kobj,
160 struct kobj_attribute *attr,
161 const char *buf, size_t count)
162{
163 ssize_t ret = count;
164
165 if (!memcmp("always", buf,
166 min(sizeof("always")-1, count))) {
167 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
168 set_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
169 } else if (!memcmp("madvise", buf,
170 min(sizeof("madvise")-1, count))) {
171 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
172 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
173 } else if (!memcmp("never", buf,
174 min(sizeof("never")-1, count))) {
175 clear_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags);
176 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags);
177 } else
178 ret = -EINVAL;
179
180 if (ret > 0) {
181 int err = start_stop_khugepaged();
182 if (err)
183 ret = err;
184 }
185 return ret;
186}
187static struct kobj_attribute enabled_attr =
188 __ATTR(enabled, 0644, enabled_show, enabled_store);
189
190ssize_t single_hugepage_flag_show(struct kobject *kobj,
191 struct kobj_attribute *attr, char *buf,
192 enum transparent_hugepage_flag flag)
193{
194 return sprintf(buf, "%d\n",
195 !!test_bit(flag, &transparent_hugepage_flags));
196}
197
198ssize_t single_hugepage_flag_store(struct kobject *kobj,
199 struct kobj_attribute *attr,
200 const char *buf, size_t count,
201 enum transparent_hugepage_flag flag)
202{
203 unsigned long value;
204 int ret;
205
206 ret = kstrtoul(buf, 10, &value);
207 if (ret < 0)
208 return ret;
209 if (value > 1)
210 return -EINVAL;
211
212 if (value)
213 set_bit(flag, &transparent_hugepage_flags);
214 else
215 clear_bit(flag, &transparent_hugepage_flags);
216
217 return count;
218}
219
220static ssize_t defrag_show(struct kobject *kobj,
221 struct kobj_attribute *attr, char *buf)
222{
223 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
224 return sprintf(buf, "[always] defer defer+madvise madvise never\n");
225 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
226 return sprintf(buf, "always [defer] defer+madvise madvise never\n");
227 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
228 return sprintf(buf, "always defer [defer+madvise] madvise never\n");
229 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
230 return sprintf(buf, "always defer defer+madvise [madvise] never\n");
231 return sprintf(buf, "always defer defer+madvise madvise [never]\n");
232}
233
234static ssize_t defrag_store(struct kobject *kobj,
235 struct kobj_attribute *attr,
236 const char *buf, size_t count)
237{
238 if (!memcmp("always", buf,
239 min(sizeof("always")-1, count))) {
240 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
241 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
242 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
243 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
244 } else if (!memcmp("defer+madvise", buf,
245 min(sizeof("defer+madvise")-1, count))) {
246 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
247 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
248 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
249 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
250 } else if (!memcmp("defer", buf,
251 min(sizeof("defer")-1, count))) {
252 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
253 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
254 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
255 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
256 } else if (!memcmp("madvise", buf,
257 min(sizeof("madvise")-1, count))) {
258 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
259 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
260 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
261 set_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
262 } else if (!memcmp("never", buf,
263 min(sizeof("never")-1, count))) {
264 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags);
265 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
266 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
267 clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags);
268 } else
269 return -EINVAL;
270
271 return count;
272}
273static struct kobj_attribute defrag_attr =
274 __ATTR(defrag, 0644, defrag_show, defrag_store);
275
276static ssize_t use_zero_page_show(struct kobject *kobj,
277 struct kobj_attribute *attr, char *buf)
278{
279 return single_hugepage_flag_show(kobj, attr, buf,
280 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
281}
282static ssize_t use_zero_page_store(struct kobject *kobj,
283 struct kobj_attribute *attr, const char *buf, size_t count)
284{
285 return single_hugepage_flag_store(kobj, attr, buf, count,
286 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
287}
288static struct kobj_attribute use_zero_page_attr =
289 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
290
291static ssize_t hpage_pmd_size_show(struct kobject *kobj,
292 struct kobj_attribute *attr, char *buf)
293{
294 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
295}
296static struct kobj_attribute hpage_pmd_size_attr =
297 __ATTR_RO(hpage_pmd_size);
298
299#ifdef CONFIG_DEBUG_VM
300static ssize_t debug_cow_show(struct kobject *kobj,
301 struct kobj_attribute *attr, char *buf)
302{
303 return single_hugepage_flag_show(kobj, attr, buf,
304 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
305}
306static ssize_t debug_cow_store(struct kobject *kobj,
307 struct kobj_attribute *attr,
308 const char *buf, size_t count)
309{
310 return single_hugepage_flag_store(kobj, attr, buf, count,
311 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
312}
313static struct kobj_attribute debug_cow_attr =
314 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
315#endif /* CONFIG_DEBUG_VM */
316
317static struct attribute *hugepage_attr[] = {
318 &enabled_attr.attr,
319 &defrag_attr.attr,
320 &use_zero_page_attr.attr,
321 &hpage_pmd_size_attr.attr,
322#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
323 &shmem_enabled_attr.attr,
324#endif
325#ifdef CONFIG_DEBUG_VM
326 &debug_cow_attr.attr,
327#endif
328 NULL,
329};
330
331static const struct attribute_group hugepage_attr_group = {
332 .attrs = hugepage_attr,
333};
334
335static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
336{
337 int err;
338
339 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
340 if (unlikely(!*hugepage_kobj)) {
341 pr_err("failed to create transparent hugepage kobject\n");
342 return -ENOMEM;
343 }
344
345 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
346 if (err) {
347 pr_err("failed to register transparent hugepage group\n");
348 goto delete_obj;
349 }
350
351 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
352 if (err) {
353 pr_err("failed to register transparent hugepage group\n");
354 goto remove_hp_group;
355 }
356
357 return 0;
358
359remove_hp_group:
360 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
361delete_obj:
362 kobject_put(*hugepage_kobj);
363 return err;
364}
365
366static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
367{
368 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
369 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
370 kobject_put(hugepage_kobj);
371}
372#else
373static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
374{
375 return 0;
376}
377
378static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
379{
380}
381#endif /* CONFIG_SYSFS */
382
383static int __init hugepage_init(void)
384{
385 int err;
386 struct kobject *hugepage_kobj;
387
388 if (!has_transparent_hugepage()) {
389 transparent_hugepage_flags = 0;
390 return -EINVAL;
391 }
392
393 /*
394 * hugepages can't be allocated by the buddy allocator
395 */
396 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
397 /*
398 * we use page->mapping and page->index in second tail page
399 * as list_head: assuming THP order >= 2
400 */
401 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
402
403 err = hugepage_init_sysfs(&hugepage_kobj);
404 if (err)
405 goto err_sysfs;
406
407 err = khugepaged_init();
408 if (err)
409 goto err_slab;
410
411 err = register_shrinker(&huge_zero_page_shrinker);
412 if (err)
413 goto err_hzp_shrinker;
414 err = register_shrinker(&deferred_split_shrinker);
415 if (err)
416 goto err_split_shrinker;
417
418 /*
419 * By default disable transparent hugepages on smaller systems,
420 * where the extra memory used could hurt more than TLB overhead
421 * is likely to save. The admin can still enable it through /sys.
422 */
423 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
424 transparent_hugepage_flags = 0;
425 return 0;
426 }
427
428 err = start_stop_khugepaged();
429 if (err)
430 goto err_khugepaged;
431
432 return 0;
433err_khugepaged:
434 unregister_shrinker(&deferred_split_shrinker);
435err_split_shrinker:
436 unregister_shrinker(&huge_zero_page_shrinker);
437err_hzp_shrinker:
438 khugepaged_destroy();
439err_slab:
440 hugepage_exit_sysfs(hugepage_kobj);
441err_sysfs:
442 return err;
443}
444subsys_initcall(hugepage_init);
445
446static int __init setup_transparent_hugepage(char *str)
447{
448 int ret = 0;
449 if (!str)
450 goto out;
451 if (!strcmp(str, "always")) {
452 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
453 &transparent_hugepage_flags);
454 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
455 &transparent_hugepage_flags);
456 ret = 1;
457 } else if (!strcmp(str, "madvise")) {
458 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
459 &transparent_hugepage_flags);
460 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
461 &transparent_hugepage_flags);
462 ret = 1;
463 } else if (!strcmp(str, "never")) {
464 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
465 &transparent_hugepage_flags);
466 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
467 &transparent_hugepage_flags);
468 ret = 1;
469 }
470out:
471 if (!ret)
472 pr_warn("transparent_hugepage= cannot parse, ignored\n");
473 return ret;
474}
475__setup("transparent_hugepage=", setup_transparent_hugepage);
476
477pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
478{
479 if (likely(vma->vm_flags & VM_WRITE))
480 pmd = pmd_mkwrite(pmd);
481 return pmd;
482}
483
484static inline struct list_head *page_deferred_list(struct page *page)
485{
486 /*
487 * ->lru in the tail pages is occupied by compound_head.
488 * Let's use ->mapping + ->index in the second tail page as list_head.
489 */
490 return (struct list_head *)&page[2].mapping;
491}
492
493void prep_transhuge_page(struct page *page)
494{
495 /*
496 * we use page->mapping and page->indexlru in second tail page
497 * as list_head: assuming THP order >= 2
498 */
499
500 INIT_LIST_HEAD(page_deferred_list(page));
501 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
502}
503
504unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
505 loff_t off, unsigned long flags, unsigned long size)
506{
507 unsigned long addr;
508 loff_t off_end = off + len;
509 loff_t off_align = round_up(off, size);
510 unsigned long len_pad;
511
512 if (off_end <= off_align || (off_end - off_align) < size)
513 return 0;
514
515 len_pad = len + size;
516 if (len_pad < len || (off + len_pad) < off)
517 return 0;
518
519 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
520 off >> PAGE_SHIFT, flags);
521 if (IS_ERR_VALUE(addr))
522 return 0;
523
524 addr += (off - addr) & (size - 1);
525 return addr;
526}
527
528unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
529 unsigned long len, unsigned long pgoff, unsigned long flags)
530{
531 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
532
533 if (addr)
534 goto out;
535 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
536 goto out;
537
538 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
539 if (addr)
540 return addr;
541
542 out:
543 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
544}
545EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
546
547static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
548 gfp_t gfp)
549{
550 struct vm_area_struct *vma = vmf->vma;
551 struct mem_cgroup *memcg;
552 pgtable_t pgtable;
553 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
554 int ret = 0;
555
556 VM_BUG_ON_PAGE(!PageCompound(page), page);
557
558 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
559 put_page(page);
560 count_vm_event(THP_FAULT_FALLBACK);
561 return VM_FAULT_FALLBACK;
562 }
563
564 pgtable = pte_alloc_one(vma->vm_mm, haddr);
565 if (unlikely(!pgtable)) {
566 ret = VM_FAULT_OOM;
567 goto release;
568 }
569
570 clear_huge_page(page, vmf->address, HPAGE_PMD_NR);
571 /*
572 * The memory barrier inside __SetPageUptodate makes sure that
573 * clear_huge_page writes become visible before the set_pmd_at()
574 * write.
575 */
576 __SetPageUptodate(page);
577
578 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
579 if (unlikely(!pmd_none(*vmf->pmd))) {
580 goto unlock_release;
581 } else {
582 pmd_t entry;
583
584 ret = check_stable_address_space(vma->vm_mm);
585 if (ret)
586 goto unlock_release;
587
588 /* Deliver the page fault to userland */
589 if (userfaultfd_missing(vma)) {
590 int ret;
591
592 spin_unlock(vmf->ptl);
593 mem_cgroup_cancel_charge(page, memcg, true);
594 put_page(page);
595 pte_free(vma->vm_mm, pgtable);
596 ret = handle_userfault(vmf, VM_UFFD_MISSING);
597 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
598 return ret;
599 }
600
601 entry = mk_huge_pmd(page, vma->vm_page_prot);
602 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
603 page_add_new_anon_rmap(page, vma, haddr, true);
604 mem_cgroup_commit_charge(page, memcg, false, true);
605 lru_cache_add_active_or_unevictable(page, vma);
606 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
607 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
608 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
609 mm_inc_nr_ptes(vma->vm_mm);
610 spin_unlock(vmf->ptl);
611 count_vm_event(THP_FAULT_ALLOC);
612 }
613
614 return 0;
615unlock_release:
616 spin_unlock(vmf->ptl);
617release:
618 if (pgtable)
619 pte_free(vma->vm_mm, pgtable);
620 mem_cgroup_cancel_charge(page, memcg, true);
621 put_page(page);
622 return ret;
623
624}
625
626/*
627 * always: directly stall for all thp allocations
628 * defer: wake kswapd and fail if not immediately available
629 * defer+madvise: wake kswapd and directly stall for MADV_HUGEPAGE, otherwise
630 * fail if not immediately available
631 * madvise: directly stall for MADV_HUGEPAGE, otherwise fail if not immediately
632 * available
633 * never: never stall for any thp allocation
634 */
635static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
636{
637 const bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
638
639 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
640 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
641 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
642 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
643 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags))
644 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
645 __GFP_KSWAPD_RECLAIM);
646 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
647 return GFP_TRANSHUGE_LIGHT | (vma_madvised ? __GFP_DIRECT_RECLAIM :
648 0);
649 return GFP_TRANSHUGE_LIGHT;
650}
651
652/* Caller must hold page table lock. */
653static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
654 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
655 struct page *zero_page)
656{
657 pmd_t entry;
658 if (!pmd_none(*pmd))
659 return false;
660 entry = mk_pmd(zero_page, vma->vm_page_prot);
661 entry = pmd_mkhuge(entry);
662 if (pgtable)
663 pgtable_trans_huge_deposit(mm, pmd, pgtable);
664 set_pmd_at(mm, haddr, pmd, entry);
665 mm_inc_nr_ptes(mm);
666 return true;
667}
668
669int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
670{
671 struct vm_area_struct *vma = vmf->vma;
672 gfp_t gfp;
673 struct page *page;
674 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
675
676 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
677 return VM_FAULT_FALLBACK;
678 if (unlikely(anon_vma_prepare(vma)))
679 return VM_FAULT_OOM;
680 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
681 return VM_FAULT_OOM;
682 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
683 !mm_forbids_zeropage(vma->vm_mm) &&
684 transparent_hugepage_use_zero_page()) {
685 pgtable_t pgtable;
686 struct page *zero_page;
687 bool set;
688 int ret;
689 pgtable = pte_alloc_one(vma->vm_mm, haddr);
690 if (unlikely(!pgtable))
691 return VM_FAULT_OOM;
692 zero_page = mm_get_huge_zero_page(vma->vm_mm);
693 if (unlikely(!zero_page)) {
694 pte_free(vma->vm_mm, pgtable);
695 count_vm_event(THP_FAULT_FALLBACK);
696 return VM_FAULT_FALLBACK;
697 }
698 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
699 ret = 0;
700 set = false;
701 if (pmd_none(*vmf->pmd)) {
702 ret = check_stable_address_space(vma->vm_mm);
703 if (ret) {
704 spin_unlock(vmf->ptl);
705 } else if (userfaultfd_missing(vma)) {
706 spin_unlock(vmf->ptl);
707 ret = handle_userfault(vmf, VM_UFFD_MISSING);
708 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
709 } else {
710 set_huge_zero_page(pgtable, vma->vm_mm, vma,
711 haddr, vmf->pmd, zero_page);
712 spin_unlock(vmf->ptl);
713 set = true;
714 }
715 } else
716 spin_unlock(vmf->ptl);
717 if (!set)
718 pte_free(vma->vm_mm, pgtable);
719 return ret;
720 }
721 gfp = alloc_hugepage_direct_gfpmask(vma);
722 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
723 if (unlikely(!page)) {
724 count_vm_event(THP_FAULT_FALLBACK);
725 return VM_FAULT_FALLBACK;
726 }
727 prep_transhuge_page(page);
728 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
729}
730
731static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
732 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write,
733 pgtable_t pgtable)
734{
735 struct mm_struct *mm = vma->vm_mm;
736 pmd_t entry;
737 spinlock_t *ptl;
738
739 ptl = pmd_lock(mm, pmd);
740 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
741 if (pfn_t_devmap(pfn))
742 entry = pmd_mkdevmap(entry);
743 if (write) {
744 entry = pmd_mkyoung(pmd_mkdirty(entry));
745 entry = maybe_pmd_mkwrite(entry, vma);
746 }
747
748 if (pgtable) {
749 pgtable_trans_huge_deposit(mm, pmd, pgtable);
750 mm_inc_nr_ptes(mm);
751 }
752
753 set_pmd_at(mm, addr, pmd, entry);
754 update_mmu_cache_pmd(vma, addr, pmd);
755 spin_unlock(ptl);
756}
757
758int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
759 pmd_t *pmd, pfn_t pfn, bool write)
760{
761 pgprot_t pgprot = vma->vm_page_prot;
762 pgtable_t pgtable = NULL;
763 /*
764 * If we had pmd_special, we could avoid all these restrictions,
765 * but we need to be consistent with PTEs and architectures that
766 * can't support a 'special' bit.
767 */
768 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
769 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
770 (VM_PFNMAP|VM_MIXEDMAP));
771 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
772 BUG_ON(!pfn_t_devmap(pfn));
773
774 if (addr < vma->vm_start || addr >= vma->vm_end)
775 return VM_FAULT_SIGBUS;
776
777 if (arch_needs_pgtable_deposit()) {
778 pgtable = pte_alloc_one(vma->vm_mm, addr);
779 if (!pgtable)
780 return VM_FAULT_OOM;
781 }
782
783 track_pfn_insert(vma, &pgprot, pfn);
784
785 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write, pgtable);
786 return VM_FAULT_NOPAGE;
787}
788EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
789
790#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
791static pud_t maybe_pud_mkwrite(pud_t pud, struct vm_area_struct *vma)
792{
793 if (likely(vma->vm_flags & VM_WRITE))
794 pud = pud_mkwrite(pud);
795 return pud;
796}
797
798static void insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
799 pud_t *pud, pfn_t pfn, pgprot_t prot, bool write)
800{
801 struct mm_struct *mm = vma->vm_mm;
802 pud_t entry;
803 spinlock_t *ptl;
804
805 ptl = pud_lock(mm, pud);
806 entry = pud_mkhuge(pfn_t_pud(pfn, prot));
807 if (pfn_t_devmap(pfn))
808 entry = pud_mkdevmap(entry);
809 if (write) {
810 entry = pud_mkyoung(pud_mkdirty(entry));
811 entry = maybe_pud_mkwrite(entry, vma);
812 }
813 set_pud_at(mm, addr, pud, entry);
814 update_mmu_cache_pud(vma, addr, pud);
815 spin_unlock(ptl);
816}
817
818int vmf_insert_pfn_pud(struct vm_area_struct *vma, unsigned long addr,
819 pud_t *pud, pfn_t pfn, bool write)
820{
821 pgprot_t pgprot = vma->vm_page_prot;
822 /*
823 * If we had pud_special, we could avoid all these restrictions,
824 * but we need to be consistent with PTEs and architectures that
825 * can't support a 'special' bit.
826 */
827 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
828 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
829 (VM_PFNMAP|VM_MIXEDMAP));
830 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
831 BUG_ON(!pfn_t_devmap(pfn));
832
833 if (addr < vma->vm_start || addr >= vma->vm_end)
834 return VM_FAULT_SIGBUS;
835
836 track_pfn_insert(vma, &pgprot, pfn);
837
838 insert_pfn_pud(vma, addr, pud, pfn, pgprot, write);
839 return VM_FAULT_NOPAGE;
840}
841EXPORT_SYMBOL_GPL(vmf_insert_pfn_pud);
842#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
843
844static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
845 pmd_t *pmd, int flags)
846{
847 pmd_t _pmd;
848
849 _pmd = pmd_mkyoung(*pmd);
850 if (flags & FOLL_WRITE)
851 _pmd = pmd_mkdirty(_pmd);
852 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
853 pmd, _pmd, flags & FOLL_WRITE))
854 update_mmu_cache_pmd(vma, addr, pmd);
855}
856
857struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
858 pmd_t *pmd, int flags)
859{
860 unsigned long pfn = pmd_pfn(*pmd);
861 struct mm_struct *mm = vma->vm_mm;
862 struct dev_pagemap *pgmap;
863 struct page *page;
864
865 assert_spin_locked(pmd_lockptr(mm, pmd));
866
867 /*
868 * When we COW a devmap PMD entry, we split it into PTEs, so we should
869 * not be in this function with `flags & FOLL_COW` set.
870 */
871 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
872
873 if (flags & FOLL_WRITE && !pmd_write(*pmd))
874 return NULL;
875
876 if (pmd_present(*pmd) && pmd_devmap(*pmd))
877 /* pass */;
878 else
879 return NULL;
880
881 if (flags & FOLL_TOUCH)
882 touch_pmd(vma, addr, pmd, flags);
883
884 /*
885 * device mapped pages can only be returned if the
886 * caller will manage the page reference count.
887 */
888 if (!(flags & FOLL_GET))
889 return ERR_PTR(-EEXIST);
890
891 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
892 pgmap = get_dev_pagemap(pfn, NULL);
893 if (!pgmap)
894 return ERR_PTR(-EFAULT);
895 page = pfn_to_page(pfn);
896 get_page(page);
897 put_dev_pagemap(pgmap);
898
899 return page;
900}
901
902int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
903 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
904 struct vm_area_struct *vma)
905{
906 spinlock_t *dst_ptl, *src_ptl;
907 struct page *src_page;
908 pmd_t pmd;
909 pgtable_t pgtable = NULL;
910 int ret = -ENOMEM;
911
912 /* Skip if can be re-fill on fault */
913 if (!vma_is_anonymous(vma))
914 return 0;
915
916 pgtable = pte_alloc_one(dst_mm, addr);
917 if (unlikely(!pgtable))
918 goto out;
919
920 dst_ptl = pmd_lock(dst_mm, dst_pmd);
921 src_ptl = pmd_lockptr(src_mm, src_pmd);
922 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
923
924 ret = -EAGAIN;
925 pmd = *src_pmd;
926
927#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
928 if (unlikely(is_swap_pmd(pmd))) {
929 swp_entry_t entry = pmd_to_swp_entry(pmd);
930
931 VM_BUG_ON(!is_pmd_migration_entry(pmd));
932 if (is_write_migration_entry(entry)) {
933 make_migration_entry_read(&entry);
934 pmd = swp_entry_to_pmd(entry);
935 if (pmd_swp_soft_dirty(*src_pmd))
936 pmd = pmd_swp_mksoft_dirty(pmd);
937 set_pmd_at(src_mm, addr, src_pmd, pmd);
938 }
939 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
940 mm_inc_nr_ptes(dst_mm);
941 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
942 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
943 ret = 0;
944 goto out_unlock;
945 }
946#endif
947
948 if (unlikely(!pmd_trans_huge(pmd))) {
949 pte_free(dst_mm, pgtable);
950 goto out_unlock;
951 }
952 /*
953 * When page table lock is held, the huge zero pmd should not be
954 * under splitting since we don't split the page itself, only pmd to
955 * a page table.
956 */
957 if (is_huge_zero_pmd(pmd)) {
958 struct page *zero_page;
959 /*
960 * get_huge_zero_page() will never allocate a new page here,
961 * since we already have a zero page to copy. It just takes a
962 * reference.
963 */
964 zero_page = mm_get_huge_zero_page(dst_mm);
965 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
966 zero_page);
967 ret = 0;
968 goto out_unlock;
969 }
970
971 src_page = pmd_page(pmd);
972 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
973 get_page(src_page);
974 page_dup_rmap(src_page, true);
975 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
976 mm_inc_nr_ptes(dst_mm);
977 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
978
979 pmdp_set_wrprotect(src_mm, addr, src_pmd);
980 pmd = pmd_mkold(pmd_wrprotect(pmd));
981 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
982
983 ret = 0;
984out_unlock:
985 spin_unlock(src_ptl);
986 spin_unlock(dst_ptl);
987out:
988 return ret;
989}
990
991#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
992static void touch_pud(struct vm_area_struct *vma, unsigned long addr,
993 pud_t *pud, int flags)
994{
995 pud_t _pud;
996
997 _pud = pud_mkyoung(*pud);
998 if (flags & FOLL_WRITE)
999 _pud = pud_mkdirty(_pud);
1000 if (pudp_set_access_flags(vma, addr & HPAGE_PUD_MASK,
1001 pud, _pud, flags & FOLL_WRITE))
1002 update_mmu_cache_pud(vma, addr, pud);
1003}
1004
1005struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
1006 pud_t *pud, int flags)
1007{
1008 unsigned long pfn = pud_pfn(*pud);
1009 struct mm_struct *mm = vma->vm_mm;
1010 struct dev_pagemap *pgmap;
1011 struct page *page;
1012
1013 assert_spin_locked(pud_lockptr(mm, pud));
1014
1015 if (flags & FOLL_WRITE && !pud_write(*pud))
1016 return NULL;
1017
1018 if (pud_present(*pud) && pud_devmap(*pud))
1019 /* pass */;
1020 else
1021 return NULL;
1022
1023 if (flags & FOLL_TOUCH)
1024 touch_pud(vma, addr, pud, flags);
1025
1026 /*
1027 * device mapped pages can only be returned if the
1028 * caller will manage the page reference count.
1029 */
1030 if (!(flags & FOLL_GET))
1031 return ERR_PTR(-EEXIST);
1032
1033 pfn += (addr & ~PUD_MASK) >> PAGE_SHIFT;
1034 pgmap = get_dev_pagemap(pfn, NULL);
1035 if (!pgmap)
1036 return ERR_PTR(-EFAULT);
1037 page = pfn_to_page(pfn);
1038 get_page(page);
1039 put_dev_pagemap(pgmap);
1040
1041 return page;
1042}
1043
1044int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1046 struct vm_area_struct *vma)
1047{
1048 spinlock_t *dst_ptl, *src_ptl;
1049 pud_t pud;
1050 int ret;
1051
1052 dst_ptl = pud_lock(dst_mm, dst_pud);
1053 src_ptl = pud_lockptr(src_mm, src_pud);
1054 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1055
1056 ret = -EAGAIN;
1057 pud = *src_pud;
1058 if (unlikely(!pud_trans_huge(pud) && !pud_devmap(pud)))
1059 goto out_unlock;
1060
1061 /*
1062 * When page table lock is held, the huge zero pud should not be
1063 * under splitting since we don't split the page itself, only pud to
1064 * a page table.
1065 */
1066 if (is_huge_zero_pud(pud)) {
1067 /* No huge zero pud yet */
1068 }
1069
1070 pudp_set_wrprotect(src_mm, addr, src_pud);
1071 pud = pud_mkold(pud_wrprotect(pud));
1072 set_pud_at(dst_mm, addr, dst_pud, pud);
1073
1074 ret = 0;
1075out_unlock:
1076 spin_unlock(src_ptl);
1077 spin_unlock(dst_ptl);
1078 return ret;
1079}
1080
1081void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
1082{
1083 pud_t entry;
1084 unsigned long haddr;
1085 bool write = vmf->flags & FAULT_FLAG_WRITE;
1086
1087 vmf->ptl = pud_lock(vmf->vma->vm_mm, vmf->pud);
1088 if (unlikely(!pud_same(*vmf->pud, orig_pud)))
1089 goto unlock;
1090
1091 entry = pud_mkyoung(orig_pud);
1092 if (write)
1093 entry = pud_mkdirty(entry);
1094 haddr = vmf->address & HPAGE_PUD_MASK;
1095 if (pudp_set_access_flags(vmf->vma, haddr, vmf->pud, entry, write))
1096 update_mmu_cache_pud(vmf->vma, vmf->address, vmf->pud);
1097
1098unlock:
1099 spin_unlock(vmf->ptl);
1100}
1101#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
1102
1103void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
1104{
1105 pmd_t entry;
1106 unsigned long haddr;
1107 bool write = vmf->flags & FAULT_FLAG_WRITE;
1108
1109 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
1110 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1111 goto unlock;
1112
1113 entry = pmd_mkyoung(orig_pmd);
1114 if (write)
1115 entry = pmd_mkdirty(entry);
1116 haddr = vmf->address & HPAGE_PMD_MASK;
1117 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
1118 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
1119
1120unlock:
1121 spin_unlock(vmf->ptl);
1122}
1123
1124static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
1125 struct page *page)
1126{
1127 struct vm_area_struct *vma = vmf->vma;
1128 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1129 struct mem_cgroup *memcg;
1130 pgtable_t pgtable;
1131 pmd_t _pmd;
1132 int ret = 0, i;
1133 struct page **pages;
1134 unsigned long mmun_start; /* For mmu_notifiers */
1135 unsigned long mmun_end; /* For mmu_notifiers */
1136
1137 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1138 GFP_KERNEL);
1139 if (unlikely(!pages)) {
1140 ret |= VM_FAULT_OOM;
1141 goto out;
1142 }
1143
1144 for (i = 0; i < HPAGE_PMD_NR; i++) {
1145 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
1146 vmf->address, page_to_nid(page));
1147 if (unlikely(!pages[i] ||
1148 mem_cgroup_try_charge(pages[i], vma->vm_mm,
1149 GFP_KERNEL, &memcg, false))) {
1150 if (pages[i])
1151 put_page(pages[i]);
1152 while (--i >= 0) {
1153 memcg = (void *)page_private(pages[i]);
1154 set_page_private(pages[i], 0);
1155 mem_cgroup_cancel_charge(pages[i], memcg,
1156 false);
1157 put_page(pages[i]);
1158 }
1159 kfree(pages);
1160 ret |= VM_FAULT_OOM;
1161 goto out;
1162 }
1163 set_page_private(pages[i], (unsigned long)memcg);
1164 }
1165
1166 for (i = 0; i < HPAGE_PMD_NR; i++) {
1167 copy_user_highpage(pages[i], page + i,
1168 haddr + PAGE_SIZE * i, vma);
1169 __SetPageUptodate(pages[i]);
1170 cond_resched();
1171 }
1172
1173 mmun_start = haddr;
1174 mmun_end = haddr + HPAGE_PMD_SIZE;
1175 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1176
1177 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1178 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1179 goto out_free_pages;
1180 VM_BUG_ON_PAGE(!PageHead(page), page);
1181
1182 /*
1183 * Leave pmd empty until pte is filled note we must notify here as
1184 * concurrent CPU thread might write to new page before the call to
1185 * mmu_notifier_invalidate_range_end() happens which can lead to a
1186 * device seeing memory write in different order than CPU.
1187 *
1188 * See Documentation/vm/mmu_notifier.txt
1189 */
1190 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1191
1192 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
1193 pmd_populate(vma->vm_mm, &_pmd, pgtable);
1194
1195 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1196 pte_t entry;
1197 entry = mk_pte(pages[i], vma->vm_page_prot);
1198 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1199 memcg = (void *)page_private(pages[i]);
1200 set_page_private(pages[i], 0);
1201 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
1202 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1203 lru_cache_add_active_or_unevictable(pages[i], vma);
1204 vmf->pte = pte_offset_map(&_pmd, haddr);
1205 VM_BUG_ON(!pte_none(*vmf->pte));
1206 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
1207 pte_unmap(vmf->pte);
1208 }
1209 kfree(pages);
1210
1211 smp_wmb(); /* make pte visible before pmd */
1212 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
1213 page_remove_rmap(page, true);
1214 spin_unlock(vmf->ptl);
1215
1216 /*
1217 * No need to double call mmu_notifier->invalidate_range() callback as
1218 * the above pmdp_huge_clear_flush_notify() did already call it.
1219 */
1220 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1221 mmun_end);
1222
1223 ret |= VM_FAULT_WRITE;
1224 put_page(page);
1225
1226out:
1227 return ret;
1228
1229out_free_pages:
1230 spin_unlock(vmf->ptl);
1231 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1232 for (i = 0; i < HPAGE_PMD_NR; i++) {
1233 memcg = (void *)page_private(pages[i]);
1234 set_page_private(pages[i], 0);
1235 mem_cgroup_cancel_charge(pages[i], memcg, false);
1236 put_page(pages[i]);
1237 }
1238 kfree(pages);
1239 goto out;
1240}
1241
1242int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1243{
1244 struct vm_area_struct *vma = vmf->vma;
1245 struct page *page = NULL, *new_page;
1246 struct mem_cgroup *memcg;
1247 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1248 unsigned long mmun_start; /* For mmu_notifiers */
1249 unsigned long mmun_end; /* For mmu_notifiers */
1250 gfp_t huge_gfp; /* for allocation and charge */
1251 int ret = 0;
1252
1253 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1254 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1255 if (is_huge_zero_pmd(orig_pmd))
1256 goto alloc;
1257 spin_lock(vmf->ptl);
1258 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1259 goto out_unlock;
1260
1261 page = pmd_page(orig_pmd);
1262 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1263 /*
1264 * We can only reuse the page if nobody else maps the huge page or it's
1265 * part.
1266 */
1267 if (!trylock_page(page)) {
1268 get_page(page);
1269 spin_unlock(vmf->ptl);
1270 lock_page(page);
1271 spin_lock(vmf->ptl);
1272 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1273 unlock_page(page);
1274 put_page(page);
1275 goto out_unlock;
1276 }
1277 put_page(page);
1278 }
1279 if (reuse_swap_page(page, NULL)) {
1280 pmd_t entry;
1281 entry = pmd_mkyoung(orig_pmd);
1282 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1283 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1284 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1285 ret |= VM_FAULT_WRITE;
1286 unlock_page(page);
1287 goto out_unlock;
1288 }
1289 unlock_page(page);
1290 get_page(page);
1291 spin_unlock(vmf->ptl);
1292alloc:
1293 if (transparent_hugepage_enabled(vma) &&
1294 !transparent_hugepage_debug_cow()) {
1295 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1296 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1297 } else
1298 new_page = NULL;
1299
1300 if (likely(new_page)) {
1301 prep_transhuge_page(new_page);
1302 } else {
1303 if (!page) {
1304 split_huge_pmd(vma, vmf->pmd, vmf->address);
1305 ret |= VM_FAULT_FALLBACK;
1306 } else {
1307 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1308 if (ret & VM_FAULT_OOM) {
1309 split_huge_pmd(vma, vmf->pmd, vmf->address);
1310 ret |= VM_FAULT_FALLBACK;
1311 }
1312 put_page(page);
1313 }
1314 count_vm_event(THP_FAULT_FALLBACK);
1315 goto out;
1316 }
1317
1318 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1319 huge_gfp, &memcg, true))) {
1320 put_page(new_page);
1321 split_huge_pmd(vma, vmf->pmd, vmf->address);
1322 if (page)
1323 put_page(page);
1324 ret |= VM_FAULT_FALLBACK;
1325 count_vm_event(THP_FAULT_FALLBACK);
1326 goto out;
1327 }
1328
1329 count_vm_event(THP_FAULT_ALLOC);
1330
1331 if (!page)
1332 clear_huge_page(new_page, vmf->address, HPAGE_PMD_NR);
1333 else
1334 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1335 __SetPageUptodate(new_page);
1336
1337 mmun_start = haddr;
1338 mmun_end = haddr + HPAGE_PMD_SIZE;
1339 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1340
1341 spin_lock(vmf->ptl);
1342 if (page)
1343 put_page(page);
1344 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1345 spin_unlock(vmf->ptl);
1346 mem_cgroup_cancel_charge(new_page, memcg, true);
1347 put_page(new_page);
1348 goto out_mn;
1349 } else {
1350 pmd_t entry;
1351 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1352 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1353 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1354 page_add_new_anon_rmap(new_page, vma, haddr, true);
1355 mem_cgroup_commit_charge(new_page, memcg, false, true);
1356 lru_cache_add_active_or_unevictable(new_page, vma);
1357 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1358 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1359 if (!page) {
1360 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1361 } else {
1362 VM_BUG_ON_PAGE(!PageHead(page), page);
1363 page_remove_rmap(page, true);
1364 put_page(page);
1365 }
1366 ret |= VM_FAULT_WRITE;
1367 }
1368 spin_unlock(vmf->ptl);
1369out_mn:
1370 /*
1371 * No need to double call mmu_notifier->invalidate_range() callback as
1372 * the above pmdp_huge_clear_flush_notify() did already call it.
1373 */
1374 mmu_notifier_invalidate_range_only_end(vma->vm_mm, mmun_start,
1375 mmun_end);
1376out:
1377 return ret;
1378out_unlock:
1379 spin_unlock(vmf->ptl);
1380 return ret;
1381}
1382
1383/*
1384 * FOLL_FORCE can write to even unwritable pmd's, but only
1385 * after we've gone through a COW cycle and they are dirty.
1386 */
1387static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1388{
1389 return pmd_write(pmd) ||
1390 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1391}
1392
1393struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1394 unsigned long addr,
1395 pmd_t *pmd,
1396 unsigned int flags)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct page *page = NULL;
1400
1401 assert_spin_locked(pmd_lockptr(mm, pmd));
1402
1403 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1404 goto out;
1405
1406 /* Avoid dumping huge zero page */
1407 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1408 return ERR_PTR(-EFAULT);
1409
1410 /* Full NUMA hinting faults to serialise migration in fault paths */
1411 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1412 goto out;
1413
1414 page = pmd_page(*pmd);
1415 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1416 if (flags & FOLL_TOUCH)
1417 touch_pmd(vma, addr, pmd, flags);
1418 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1419 /*
1420 * We don't mlock() pte-mapped THPs. This way we can avoid
1421 * leaking mlocked pages into non-VM_LOCKED VMAs.
1422 *
1423 * For anon THP:
1424 *
1425 * In most cases the pmd is the only mapping of the page as we
1426 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1427 * writable private mappings in populate_vma_page_range().
1428 *
1429 * The only scenario when we have the page shared here is if we
1430 * mlocking read-only mapping shared over fork(). We skip
1431 * mlocking such pages.
1432 *
1433 * For file THP:
1434 *
1435 * We can expect PageDoubleMap() to be stable under page lock:
1436 * for file pages we set it in page_add_file_rmap(), which
1437 * requires page to be locked.
1438 */
1439
1440 if (PageAnon(page) && compound_mapcount(page) != 1)
1441 goto skip_mlock;
1442 if (PageDoubleMap(page) || !page->mapping)
1443 goto skip_mlock;
1444 if (!trylock_page(page))
1445 goto skip_mlock;
1446 lru_add_drain();
1447 if (page->mapping && !PageDoubleMap(page))
1448 mlock_vma_page(page);
1449 unlock_page(page);
1450 }
1451skip_mlock:
1452 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1453 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1454 if (flags & FOLL_GET)
1455 get_page(page);
1456
1457out:
1458 return page;
1459}
1460
1461/* NUMA hinting page fault entry point for trans huge pmds */
1462int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1463{
1464 struct vm_area_struct *vma = vmf->vma;
1465 struct anon_vma *anon_vma = NULL;
1466 struct page *page;
1467 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1468 int page_nid = -1, this_nid = numa_node_id();
1469 int target_nid, last_cpupid = -1;
1470 bool page_locked;
1471 bool migrated = false;
1472 bool was_writable;
1473 int flags = 0;
1474
1475 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1476 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1477 goto out_unlock;
1478
1479 /*
1480 * If there are potential migrations, wait for completion and retry
1481 * without disrupting NUMA hinting information. Do not relock and
1482 * check_same as the page may no longer be mapped.
1483 */
1484 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1485 page = pmd_page(*vmf->pmd);
1486 if (!get_page_unless_zero(page))
1487 goto out_unlock;
1488 spin_unlock(vmf->ptl);
1489 wait_on_page_locked(page);
1490 put_page(page);
1491 goto out;
1492 }
1493
1494 page = pmd_page(pmd);
1495 BUG_ON(is_huge_zero_page(page));
1496 page_nid = page_to_nid(page);
1497 last_cpupid = page_cpupid_last(page);
1498 count_vm_numa_event(NUMA_HINT_FAULTS);
1499 if (page_nid == this_nid) {
1500 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1501 flags |= TNF_FAULT_LOCAL;
1502 }
1503
1504 /* See similar comment in do_numa_page for explanation */
1505 if (!pmd_savedwrite(pmd))
1506 flags |= TNF_NO_GROUP;
1507
1508 /*
1509 * Acquire the page lock to serialise THP migrations but avoid dropping
1510 * page_table_lock if at all possible
1511 */
1512 page_locked = trylock_page(page);
1513 target_nid = mpol_misplaced(page, vma, haddr);
1514 if (target_nid == -1) {
1515 /* If the page was locked, there are no parallel migrations */
1516 if (page_locked)
1517 goto clear_pmdnuma;
1518 }
1519
1520 /* Migration could have started since the pmd_trans_migrating check */
1521 if (!page_locked) {
1522 page_nid = -1;
1523 if (!get_page_unless_zero(page))
1524 goto out_unlock;
1525 spin_unlock(vmf->ptl);
1526 wait_on_page_locked(page);
1527 put_page(page);
1528 goto out;
1529 }
1530
1531 /*
1532 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1533 * to serialises splits
1534 */
1535 get_page(page);
1536 spin_unlock(vmf->ptl);
1537 anon_vma = page_lock_anon_vma_read(page);
1538
1539 /* Confirm the PMD did not change while page_table_lock was released */
1540 spin_lock(vmf->ptl);
1541 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1542 unlock_page(page);
1543 put_page(page);
1544 page_nid = -1;
1545 goto out_unlock;
1546 }
1547
1548 /* Bail if we fail to protect against THP splits for any reason */
1549 if (unlikely(!anon_vma)) {
1550 put_page(page);
1551 page_nid = -1;
1552 goto clear_pmdnuma;
1553 }
1554
1555 /*
1556 * Since we took the NUMA fault, we must have observed the !accessible
1557 * bit. Make sure all other CPUs agree with that, to avoid them
1558 * modifying the page we're about to migrate.
1559 *
1560 * Must be done under PTL such that we'll observe the relevant
1561 * inc_tlb_flush_pending().
1562 *
1563 * We are not sure a pending tlb flush here is for a huge page
1564 * mapping or not. Hence use the tlb range variant
1565 */
1566 if (mm_tlb_flush_pending(vma->vm_mm))
1567 flush_tlb_range(vma, haddr, haddr + HPAGE_PMD_SIZE);
1568
1569 /*
1570 * Migrate the THP to the requested node, returns with page unlocked
1571 * and access rights restored.
1572 */
1573 spin_unlock(vmf->ptl);
1574
1575 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1576 vmf->pmd, pmd, vmf->address, page, target_nid);
1577 if (migrated) {
1578 flags |= TNF_MIGRATED;
1579 page_nid = target_nid;
1580 } else
1581 flags |= TNF_MIGRATE_FAIL;
1582
1583 goto out;
1584clear_pmdnuma:
1585 BUG_ON(!PageLocked(page));
1586 was_writable = pmd_savedwrite(pmd);
1587 pmd = pmd_modify(pmd, vma->vm_page_prot);
1588 pmd = pmd_mkyoung(pmd);
1589 if (was_writable)
1590 pmd = pmd_mkwrite(pmd);
1591 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1592 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1593 unlock_page(page);
1594out_unlock:
1595 spin_unlock(vmf->ptl);
1596
1597out:
1598 if (anon_vma)
1599 page_unlock_anon_vma_read(anon_vma);
1600
1601 if (page_nid != -1)
1602 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1603 flags);
1604
1605 return 0;
1606}
1607
1608/*
1609 * Return true if we do MADV_FREE successfully on entire pmd page.
1610 * Otherwise, return false.
1611 */
1612bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1613 pmd_t *pmd, unsigned long addr, unsigned long next)
1614{
1615 spinlock_t *ptl;
1616 pmd_t orig_pmd;
1617 struct page *page;
1618 struct mm_struct *mm = tlb->mm;
1619 bool ret = false;
1620
1621 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1622
1623 ptl = pmd_trans_huge_lock(pmd, vma);
1624 if (!ptl)
1625 goto out_unlocked;
1626
1627 orig_pmd = *pmd;
1628 if (is_huge_zero_pmd(orig_pmd))
1629 goto out;
1630
1631 if (unlikely(!pmd_present(orig_pmd))) {
1632 VM_BUG_ON(thp_migration_supported() &&
1633 !is_pmd_migration_entry(orig_pmd));
1634 goto out;
1635 }
1636
1637 page = pmd_page(orig_pmd);
1638 /*
1639 * If other processes are mapping this page, we couldn't discard
1640 * the page unless they all do MADV_FREE so let's skip the page.
1641 */
1642 if (page_mapcount(page) != 1)
1643 goto out;
1644
1645 if (!trylock_page(page))
1646 goto out;
1647
1648 /*
1649 * If user want to discard part-pages of THP, split it so MADV_FREE
1650 * will deactivate only them.
1651 */
1652 if (next - addr != HPAGE_PMD_SIZE) {
1653 get_page(page);
1654 spin_unlock(ptl);
1655 split_huge_page(page);
1656 unlock_page(page);
1657 put_page(page);
1658 goto out_unlocked;
1659 }
1660
1661 if (PageDirty(page))
1662 ClearPageDirty(page);
1663 unlock_page(page);
1664
1665 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1666 pmdp_invalidate(vma, addr, pmd);
1667 orig_pmd = pmd_mkold(orig_pmd);
1668 orig_pmd = pmd_mkclean(orig_pmd);
1669
1670 set_pmd_at(mm, addr, pmd, orig_pmd);
1671 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1672 }
1673
1674 mark_page_lazyfree(page);
1675 ret = true;
1676out:
1677 spin_unlock(ptl);
1678out_unlocked:
1679 return ret;
1680}
1681
1682static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1683{
1684 pgtable_t pgtable;
1685
1686 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1687 pte_free(mm, pgtable);
1688 mm_dec_nr_ptes(mm);
1689}
1690
1691int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1692 pmd_t *pmd, unsigned long addr)
1693{
1694 pmd_t orig_pmd;
1695 spinlock_t *ptl;
1696
1697 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1698
1699 ptl = __pmd_trans_huge_lock(pmd, vma);
1700 if (!ptl)
1701 return 0;
1702 /*
1703 * For architectures like ppc64 we look at deposited pgtable
1704 * when calling pmdp_huge_get_and_clear. So do the
1705 * pgtable_trans_huge_withdraw after finishing pmdp related
1706 * operations.
1707 */
1708 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1709 tlb->fullmm);
1710 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1711 if (vma_is_dax(vma)) {
1712 if (arch_needs_pgtable_deposit())
1713 zap_deposited_table(tlb->mm, pmd);
1714 spin_unlock(ptl);
1715 if (is_huge_zero_pmd(orig_pmd))
1716 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1717 } else if (is_huge_zero_pmd(orig_pmd)) {
1718 zap_deposited_table(tlb->mm, pmd);
1719 spin_unlock(ptl);
1720 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1721 } else {
1722 struct page *page = NULL;
1723 int flush_needed = 1;
1724
1725 if (pmd_present(orig_pmd)) {
1726 page = pmd_page(orig_pmd);
1727 page_remove_rmap(page, true);
1728 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1729 VM_BUG_ON_PAGE(!PageHead(page), page);
1730 } else if (thp_migration_supported()) {
1731 swp_entry_t entry;
1732
1733 VM_BUG_ON(!is_pmd_migration_entry(orig_pmd));
1734 entry = pmd_to_swp_entry(orig_pmd);
1735 page = pfn_to_page(swp_offset(entry));
1736 flush_needed = 0;
1737 } else
1738 WARN_ONCE(1, "Non present huge pmd without pmd migration enabled!");
1739
1740 if (PageAnon(page)) {
1741 zap_deposited_table(tlb->mm, pmd);
1742 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1743 } else {
1744 if (arch_needs_pgtable_deposit())
1745 zap_deposited_table(tlb->mm, pmd);
1746 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1747 }
1748
1749 spin_unlock(ptl);
1750 if (flush_needed)
1751 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1752 }
1753 return 1;
1754}
1755
1756#ifndef pmd_move_must_withdraw
1757static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1758 spinlock_t *old_pmd_ptl,
1759 struct vm_area_struct *vma)
1760{
1761 /*
1762 * With split pmd lock we also need to move preallocated
1763 * PTE page table if new_pmd is on different PMD page table.
1764 *
1765 * We also don't deposit and withdraw tables for file pages.
1766 */
1767 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1768}
1769#endif
1770
1771static pmd_t move_soft_dirty_pmd(pmd_t pmd)
1772{
1773#ifdef CONFIG_MEM_SOFT_DIRTY
1774 if (unlikely(is_pmd_migration_entry(pmd)))
1775 pmd = pmd_swp_mksoft_dirty(pmd);
1776 else if (pmd_present(pmd))
1777 pmd = pmd_mksoft_dirty(pmd);
1778#endif
1779 return pmd;
1780}
1781
1782bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1783 unsigned long new_addr, unsigned long old_end,
1784 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1785{
1786 spinlock_t *old_ptl, *new_ptl;
1787 pmd_t pmd;
1788 struct mm_struct *mm = vma->vm_mm;
1789 bool force_flush = false;
1790
1791 if ((old_addr & ~HPAGE_PMD_MASK) ||
1792 (new_addr & ~HPAGE_PMD_MASK) ||
1793 old_end - old_addr < HPAGE_PMD_SIZE)
1794 return false;
1795
1796 /*
1797 * The destination pmd shouldn't be established, free_pgtables()
1798 * should have release it.
1799 */
1800 if (WARN_ON(!pmd_none(*new_pmd))) {
1801 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1802 return false;
1803 }
1804
1805 /*
1806 * We don't have to worry about the ordering of src and dst
1807 * ptlocks because exclusive mmap_sem prevents deadlock.
1808 */
1809 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1810 if (old_ptl) {
1811 new_ptl = pmd_lockptr(mm, new_pmd);
1812 if (new_ptl != old_ptl)
1813 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1814 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1815 if (pmd_present(pmd) && pmd_dirty(pmd))
1816 force_flush = true;
1817 VM_BUG_ON(!pmd_none(*new_pmd));
1818
1819 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1820 pgtable_t pgtable;
1821 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1822 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1823 }
1824 pmd = move_soft_dirty_pmd(pmd);
1825 set_pmd_at(mm, new_addr, new_pmd, pmd);
1826 if (new_ptl != old_ptl)
1827 spin_unlock(new_ptl);
1828 if (force_flush)
1829 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1830 else
1831 *need_flush = true;
1832 spin_unlock(old_ptl);
1833 return true;
1834 }
1835 return false;
1836}
1837
1838/*
1839 * Returns
1840 * - 0 if PMD could not be locked
1841 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1842 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1843 */
1844int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1845 unsigned long addr, pgprot_t newprot, int prot_numa)
1846{
1847 struct mm_struct *mm = vma->vm_mm;
1848 spinlock_t *ptl;
1849 pmd_t entry;
1850 bool preserve_write;
1851 int ret;
1852
1853 ptl = __pmd_trans_huge_lock(pmd, vma);
1854 if (!ptl)
1855 return 0;
1856
1857 preserve_write = prot_numa && pmd_write(*pmd);
1858 ret = 1;
1859
1860#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
1861 if (is_swap_pmd(*pmd)) {
1862 swp_entry_t entry = pmd_to_swp_entry(*pmd);
1863
1864 VM_BUG_ON(!is_pmd_migration_entry(*pmd));
1865 if (is_write_migration_entry(entry)) {
1866 pmd_t newpmd;
1867 /*
1868 * A protection check is difficult so
1869 * just be safe and disable write
1870 */
1871 make_migration_entry_read(&entry);
1872 newpmd = swp_entry_to_pmd(entry);
1873 if (pmd_swp_soft_dirty(*pmd))
1874 newpmd = pmd_swp_mksoft_dirty(newpmd);
1875 set_pmd_at(mm, addr, pmd, newpmd);
1876 }
1877 goto unlock;
1878 }
1879#endif
1880
1881 /*
1882 * Avoid trapping faults against the zero page. The read-only
1883 * data is likely to be read-cached on the local CPU and
1884 * local/remote hits to the zero page are not interesting.
1885 */
1886 if (prot_numa && is_huge_zero_pmd(*pmd))
1887 goto unlock;
1888
1889 if (prot_numa && pmd_protnone(*pmd))
1890 goto unlock;
1891
1892 /*
1893 * In case prot_numa, we are under down_read(mmap_sem). It's critical
1894 * to not clear pmd intermittently to avoid race with MADV_DONTNEED
1895 * which is also under down_read(mmap_sem):
1896 *
1897 * CPU0: CPU1:
1898 * change_huge_pmd(prot_numa=1)
1899 * pmdp_huge_get_and_clear_notify()
1900 * madvise_dontneed()
1901 * zap_pmd_range()
1902 * pmd_trans_huge(*pmd) == 0 (without ptl)
1903 * // skip the pmd
1904 * set_pmd_at();
1905 * // pmd is re-established
1906 *
1907 * The race makes MADV_DONTNEED miss the huge pmd and don't clear it
1908 * which may break userspace.
1909 *
1910 * pmdp_invalidate() is required to make sure we don't miss
1911 * dirty/young flags set by hardware.
1912 */
1913 entry = pmdp_invalidate(vma, addr, pmd);
1914
1915 entry = pmd_modify(entry, newprot);
1916 if (preserve_write)
1917 entry = pmd_mk_savedwrite(entry);
1918 ret = HPAGE_PMD_NR;
1919 set_pmd_at(mm, addr, pmd, entry);
1920 BUG_ON(vma_is_anonymous(vma) && !preserve_write && pmd_write(entry));
1921unlock:
1922 spin_unlock(ptl);
1923 return ret;
1924}
1925
1926/*
1927 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1928 *
1929 * Note that if it returns page table lock pointer, this routine returns without
1930 * unlocking page table lock. So callers must unlock it.
1931 */
1932spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1933{
1934 spinlock_t *ptl;
1935 ptl = pmd_lock(vma->vm_mm, pmd);
1936 if (likely(is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) ||
1937 pmd_devmap(*pmd)))
1938 return ptl;
1939 spin_unlock(ptl);
1940 return NULL;
1941}
1942
1943/*
1944 * Returns true if a given pud maps a thp, false otherwise.
1945 *
1946 * Note that if it returns true, this routine returns without unlocking page
1947 * table lock. So callers must unlock it.
1948 */
1949spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma)
1950{
1951 spinlock_t *ptl;
1952
1953 ptl = pud_lock(vma->vm_mm, pud);
1954 if (likely(pud_trans_huge(*pud) || pud_devmap(*pud)))
1955 return ptl;
1956 spin_unlock(ptl);
1957 return NULL;
1958}
1959
1960#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
1961int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
1962 pud_t *pud, unsigned long addr)
1963{
1964 pud_t orig_pud;
1965 spinlock_t *ptl;
1966
1967 ptl = __pud_trans_huge_lock(pud, vma);
1968 if (!ptl)
1969 return 0;
1970 /*
1971 * For architectures like ppc64 we look at deposited pgtable
1972 * when calling pudp_huge_get_and_clear. So do the
1973 * pgtable_trans_huge_withdraw after finishing pudp related
1974 * operations.
1975 */
1976 orig_pud = pudp_huge_get_and_clear_full(tlb->mm, addr, pud,
1977 tlb->fullmm);
1978 tlb_remove_pud_tlb_entry(tlb, pud, addr);
1979 if (vma_is_dax(vma)) {
1980 spin_unlock(ptl);
1981 /* No zero page support yet */
1982 } else {
1983 /* No support for anonymous PUD pages yet */
1984 BUG();
1985 }
1986 return 1;
1987}
1988
1989static void __split_huge_pud_locked(struct vm_area_struct *vma, pud_t *pud,
1990 unsigned long haddr)
1991{
1992 VM_BUG_ON(haddr & ~HPAGE_PUD_MASK);
1993 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1994 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PUD_SIZE, vma);
1995 VM_BUG_ON(!pud_trans_huge(*pud) && !pud_devmap(*pud));
1996
1997 count_vm_event(THP_SPLIT_PUD);
1998
1999 pudp_huge_clear_flush_notify(vma, haddr, pud);
2000}
2001
2002void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
2003 unsigned long address)
2004{
2005 spinlock_t *ptl;
2006 struct mm_struct *mm = vma->vm_mm;
2007 unsigned long haddr = address & HPAGE_PUD_MASK;
2008
2009 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PUD_SIZE);
2010 ptl = pud_lock(mm, pud);
2011 if (unlikely(!pud_trans_huge(*pud) && !pud_devmap(*pud)))
2012 goto out;
2013 __split_huge_pud_locked(vma, pud, haddr);
2014
2015out:
2016 spin_unlock(ptl);
2017 /*
2018 * No need to double call mmu_notifier->invalidate_range() callback as
2019 * the above pudp_huge_clear_flush_notify() did already call it.
2020 */
2021 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2022 HPAGE_PUD_SIZE);
2023}
2024#endif /* CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
2025
2026static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2027 unsigned long haddr, pmd_t *pmd)
2028{
2029 struct mm_struct *mm = vma->vm_mm;
2030 pgtable_t pgtable;
2031 pmd_t _pmd;
2032 int i;
2033
2034 /*
2035 * Leave pmd empty until pte is filled note that it is fine to delay
2036 * notification until mmu_notifier_invalidate_range_end() as we are
2037 * replacing a zero pmd write protected page with a zero pte write
2038 * protected page.
2039 *
2040 * See Documentation/vm/mmu_notifier.txt
2041 */
2042 pmdp_huge_clear_flush(vma, haddr, pmd);
2043
2044 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2045 pmd_populate(mm, &_pmd, pgtable);
2046
2047 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2048 pte_t *pte, entry;
2049 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2050 entry = pte_mkspecial(entry);
2051 pte = pte_offset_map(&_pmd, haddr);
2052 VM_BUG_ON(!pte_none(*pte));
2053 set_pte_at(mm, haddr, pte, entry);
2054 pte_unmap(pte);
2055 }
2056 smp_wmb(); /* make pte visible before pmd */
2057 pmd_populate(mm, pmd, pgtable);
2058}
2059
2060static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2061 unsigned long haddr, bool freeze)
2062{
2063 struct mm_struct *mm = vma->vm_mm;
2064 struct page *page;
2065 pgtable_t pgtable;
2066 pmd_t old_pmd, _pmd;
2067 bool young, write, soft_dirty, pmd_migration = false;
2068 unsigned long addr;
2069 int i;
2070
2071 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2072 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2073 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2074 VM_BUG_ON(!is_pmd_migration_entry(*pmd) && !pmd_trans_huge(*pmd)
2075 && !pmd_devmap(*pmd));
2076
2077 count_vm_event(THP_SPLIT_PMD);
2078
2079 if (!vma_is_anonymous(vma)) {
2080 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2081 /*
2082 * We are going to unmap this huge page. So
2083 * just go ahead and zap it
2084 */
2085 if (arch_needs_pgtable_deposit())
2086 zap_deposited_table(mm, pmd);
2087 if (vma_is_dax(vma))
2088 return;
2089 page = pmd_page(_pmd);
2090 if (!PageReferenced(page) && pmd_young(_pmd))
2091 SetPageReferenced(page);
2092 page_remove_rmap(page, true);
2093 put_page(page);
2094 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
2095 return;
2096 } else if (is_huge_zero_pmd(*pmd)) {
2097 /*
2098 * FIXME: Do we want to invalidate secondary mmu by calling
2099 * mmu_notifier_invalidate_range() see comments below inside
2100 * __split_huge_pmd() ?
2101 *
2102 * We are going from a zero huge page write protected to zero
2103 * small page also write protected so it does not seems useful
2104 * to invalidate secondary mmu at this time.
2105 */
2106 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2107 }
2108
2109 /*
2110 * Up to this point the pmd is present and huge and userland has the
2111 * whole access to the hugepage during the split (which happens in
2112 * place). If we overwrite the pmd with the not-huge version pointing
2113 * to the pte here (which of course we could if all CPUs were bug
2114 * free), userland could trigger a small page size TLB miss on the
2115 * small sized TLB while the hugepage TLB entry is still established in
2116 * the huge TLB. Some CPU doesn't like that.
2117 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2118 * 383 on page 93. Intel should be safe but is also warns that it's
2119 * only safe if the permission and cache attributes of the two entries
2120 * loaded in the two TLB is identical (which should be the case here).
2121 * But it is generally safer to never allow small and huge TLB entries
2122 * for the same virtual address to be loaded simultaneously. So instead
2123 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2124 * current pmd notpresent (atomically because here the pmd_trans_huge
2125 * must remain set at all times on the pmd until the split is complete
2126 * for this pmd), then we flush the SMP TLB and finally we write the
2127 * non-huge version of the pmd entry with pmd_populate.
2128 */
2129 old_pmd = pmdp_invalidate(vma, haddr, pmd);
2130
2131#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2132 pmd_migration = is_pmd_migration_entry(old_pmd);
2133 if (pmd_migration) {
2134 swp_entry_t entry;
2135
2136 entry = pmd_to_swp_entry(old_pmd);
2137 page = pfn_to_page(swp_offset(entry));
2138 } else
2139#endif
2140 page = pmd_page(old_pmd);
2141 VM_BUG_ON_PAGE(!page_count(page), page);
2142 page_ref_add(page, HPAGE_PMD_NR - 1);
2143 if (pmd_dirty(old_pmd))
2144 SetPageDirty(page);
2145 write = pmd_write(old_pmd);
2146 young = pmd_young(old_pmd);
2147 soft_dirty = pmd_soft_dirty(old_pmd);
2148
2149 /*
2150 * Withdraw the table only after we mark the pmd entry invalid.
2151 * This's critical for some architectures (Power).
2152 */
2153 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2154 pmd_populate(mm, &_pmd, pgtable);
2155
2156 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2157 pte_t entry, *pte;
2158 /*
2159 * Note that NUMA hinting access restrictions are not
2160 * transferred to avoid any possibility of altering
2161 * permissions across VMAs.
2162 */
2163 if (freeze || pmd_migration) {
2164 swp_entry_t swp_entry;
2165 swp_entry = make_migration_entry(page + i, write);
2166 entry = swp_entry_to_pte(swp_entry);
2167 if (soft_dirty)
2168 entry = pte_swp_mksoft_dirty(entry);
2169 } else {
2170 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
2171 entry = maybe_mkwrite(entry, vma);
2172 if (!write)
2173 entry = pte_wrprotect(entry);
2174 if (!young)
2175 entry = pte_mkold(entry);
2176 if (soft_dirty)
2177 entry = pte_mksoft_dirty(entry);
2178 }
2179 pte = pte_offset_map(&_pmd, addr);
2180 BUG_ON(!pte_none(*pte));
2181 set_pte_at(mm, addr, pte, entry);
2182 atomic_inc(&page[i]._mapcount);
2183 pte_unmap(pte);
2184 }
2185
2186 /*
2187 * Set PG_double_map before dropping compound_mapcount to avoid
2188 * false-negative page_mapped().
2189 */
2190 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2191 for (i = 0; i < HPAGE_PMD_NR; i++)
2192 atomic_inc(&page[i]._mapcount);
2193 }
2194
2195 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2196 /* Last compound_mapcount is gone. */
2197 __dec_node_page_state(page, NR_ANON_THPS);
2198 if (TestClearPageDoubleMap(page)) {
2199 /* No need in mapcount reference anymore */
2200 for (i = 0; i < HPAGE_PMD_NR; i++)
2201 atomic_dec(&page[i]._mapcount);
2202 }
2203 }
2204
2205 smp_wmb(); /* make pte visible before pmd */
2206 pmd_populate(mm, pmd, pgtable);
2207
2208 if (freeze) {
2209 for (i = 0; i < HPAGE_PMD_NR; i++) {
2210 page_remove_rmap(page + i, false);
2211 put_page(page + i);
2212 }
2213 }
2214}
2215
2216void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2217 unsigned long address, bool freeze, struct page *page)
2218{
2219 spinlock_t *ptl;
2220 struct mm_struct *mm = vma->vm_mm;
2221 unsigned long haddr = address & HPAGE_PMD_MASK;
2222
2223 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2224 ptl = pmd_lock(mm, pmd);
2225
2226 /*
2227 * If caller asks to setup a migration entries, we need a page to check
2228 * pmd against. Otherwise we can end up replacing wrong page.
2229 */
2230 VM_BUG_ON(freeze && !page);
2231 if (page && page != pmd_page(*pmd))
2232 goto out;
2233
2234 if (pmd_trans_huge(*pmd)) {
2235 page = pmd_page(*pmd);
2236 if (PageMlocked(page))
2237 clear_page_mlock(page);
2238 } else if (!(pmd_devmap(*pmd) || is_pmd_migration_entry(*pmd)))
2239 goto out;
2240 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2241out:
2242 spin_unlock(ptl);
2243 /*
2244 * No need to double call mmu_notifier->invalidate_range() callback.
2245 * They are 3 cases to consider inside __split_huge_pmd_locked():
2246 * 1) pmdp_huge_clear_flush_notify() call invalidate_range() obvious
2247 * 2) __split_huge_zero_page_pmd() read only zero page and any write
2248 * fault will trigger a flush_notify before pointing to a new page
2249 * (it is fine if the secondary mmu keeps pointing to the old zero
2250 * page in the meantime)
2251 * 3) Split a huge pmd into pte pointing to the same page. No need
2252 * to invalidate secondary tlb entry they are all still valid.
2253 * any further changes to individual pte will notify. So no need
2254 * to call mmu_notifier->invalidate_range()
2255 */
2256 mmu_notifier_invalidate_range_only_end(mm, haddr, haddr +
2257 HPAGE_PMD_SIZE);
2258}
2259
2260void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
2261 bool freeze, struct page *page)
2262{
2263 pgd_t *pgd;
2264 p4d_t *p4d;
2265 pud_t *pud;
2266 pmd_t *pmd;
2267
2268 pgd = pgd_offset(vma->vm_mm, address);
2269 if (!pgd_present(*pgd))
2270 return;
2271
2272 p4d = p4d_offset(pgd, address);
2273 if (!p4d_present(*p4d))
2274 return;
2275
2276 pud = pud_offset(p4d, address);
2277 if (!pud_present(*pud))
2278 return;
2279
2280 pmd = pmd_offset(pud, address);
2281
2282 __split_huge_pmd(vma, pmd, address, freeze, page);
2283}
2284
2285void vma_adjust_trans_huge(struct vm_area_struct *vma,
2286 unsigned long start,
2287 unsigned long end,
2288 long adjust_next)
2289{
2290 /*
2291 * If the new start address isn't hpage aligned and it could
2292 * previously contain an hugepage: check if we need to split
2293 * an huge pmd.
2294 */
2295 if (start & ~HPAGE_PMD_MASK &&
2296 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2297 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2298 split_huge_pmd_address(vma, start, false, NULL);
2299
2300 /*
2301 * If the new end address isn't hpage aligned and it could
2302 * previously contain an hugepage: check if we need to split
2303 * an huge pmd.
2304 */
2305 if (end & ~HPAGE_PMD_MASK &&
2306 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2307 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2308 split_huge_pmd_address(vma, end, false, NULL);
2309
2310 /*
2311 * If we're also updating the vma->vm_next->vm_start, if the new
2312 * vm_next->vm_start isn't page aligned and it could previously
2313 * contain an hugepage: check if we need to split an huge pmd.
2314 */
2315 if (adjust_next > 0) {
2316 struct vm_area_struct *next = vma->vm_next;
2317 unsigned long nstart = next->vm_start;
2318 nstart += adjust_next << PAGE_SHIFT;
2319 if (nstart & ~HPAGE_PMD_MASK &&
2320 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2321 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2322 split_huge_pmd_address(next, nstart, false, NULL);
2323 }
2324}
2325
2326static void freeze_page(struct page *page)
2327{
2328 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
2329 TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD;
2330 bool unmap_success;
2331
2332 VM_BUG_ON_PAGE(!PageHead(page), page);
2333
2334 if (PageAnon(page))
2335 ttu_flags |= TTU_SPLIT_FREEZE;
2336
2337 unmap_success = try_to_unmap(page, ttu_flags);
2338 VM_BUG_ON_PAGE(!unmap_success, page);
2339}
2340
2341static void unfreeze_page(struct page *page)
2342{
2343 int i;
2344 if (PageTransHuge(page)) {
2345 remove_migration_ptes(page, page, true);
2346 } else {
2347 for (i = 0; i < HPAGE_PMD_NR; i++)
2348 remove_migration_ptes(page + i, page + i, true);
2349 }
2350}
2351
2352static void __split_huge_page_tail(struct page *head, int tail,
2353 struct lruvec *lruvec, struct list_head *list)
2354{
2355 struct page *page_tail = head + tail;
2356
2357 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
2358
2359 /*
2360 * Clone page flags before unfreezing refcount.
2361 *
2362 * After successful get_page_unless_zero() might follow flags change,
2363 * for exmaple lock_page() which set PG_waiters.
2364 */
2365 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
2366 page_tail->flags |= (head->flags &
2367 ((1L << PG_referenced) |
2368 (1L << PG_swapbacked) |
2369 (1L << PG_swapcache) |
2370 (1L << PG_mlocked) |
2371 (1L << PG_uptodate) |
2372 (1L << PG_active) |
2373 (1L << PG_locked) |
2374 (1L << PG_unevictable) |
2375 (1L << PG_dirty)));
2376
2377 /* Page flags must be visible before we make the page non-compound. */
2378 smp_wmb();
2379
2380 /*
2381 * Clear PageTail before unfreezing page refcount.
2382 *
2383 * After successful get_page_unless_zero() might follow put_page()
2384 * which needs correct compound_head().
2385 */
2386 clear_compound_head(page_tail);
2387
2388 /* Finally unfreeze refcount. Additional reference from page cache. */
2389 page_ref_unfreeze(page_tail, 1 + (!PageAnon(head) ||
2390 PageSwapCache(head)));
2391
2392 if (page_is_young(head))
2393 set_page_young(page_tail);
2394 if (page_is_idle(head))
2395 set_page_idle(page_tail);
2396
2397 /* ->mapping in first tail page is compound_mapcount */
2398 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
2399 page_tail);
2400 page_tail->mapping = head->mapping;
2401
2402 page_tail->index = head->index + tail;
2403 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
2404
2405 /*
2406 * always add to the tail because some iterators expect new
2407 * pages to show after the currently processed elements - e.g.
2408 * migrate_pages
2409 */
2410 lru_add_page_tail(head, page_tail, lruvec, list);
2411}
2412
2413static void __split_huge_page(struct page *page, struct list_head *list,
2414 unsigned long flags)
2415{
2416 struct page *head = compound_head(page);
2417 struct zone *zone = page_zone(head);
2418 struct lruvec *lruvec;
2419 pgoff_t end = -1;
2420 int i;
2421
2422 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
2423
2424 /* complete memcg works before add pages to LRU */
2425 mem_cgroup_split_huge_fixup(head);
2426
2427 if (!PageAnon(page))
2428 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
2429
2430 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
2431 __split_huge_page_tail(head, i, lruvec, list);
2432 /* Some pages can be beyond i_size: drop them from page cache */
2433 if (head[i].index >= end) {
2434 ClearPageDirty(head + i);
2435 __delete_from_page_cache(head + i, NULL);
2436 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
2437 shmem_uncharge(head->mapping->host, 1);
2438 put_page(head + i);
2439 }
2440 }
2441
2442 ClearPageCompound(head);
2443 /* See comment in __split_huge_page_tail() */
2444 if (PageAnon(head)) {
2445 /* Additional pin to radix tree of swap cache */
2446 if (PageSwapCache(head))
2447 page_ref_add(head, 2);
2448 else
2449 page_ref_inc(head);
2450 } else {
2451 /* Additional pin to radix tree */
2452 page_ref_add(head, 2);
2453 xa_unlock(&head->mapping->i_pages);
2454 }
2455
2456 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2457
2458 unfreeze_page(head);
2459
2460 for (i = 0; i < HPAGE_PMD_NR; i++) {
2461 struct page *subpage = head + i;
2462 if (subpage == page)
2463 continue;
2464 unlock_page(subpage);
2465
2466 /*
2467 * Subpages may be freed if there wasn't any mapping
2468 * like if add_to_swap() is running on a lru page that
2469 * had its mapping zapped. And freeing these pages
2470 * requires taking the lru_lock so we do the put_page
2471 * of the tail pages after the split is complete.
2472 */
2473 put_page(subpage);
2474 }
2475}
2476
2477int total_mapcount(struct page *page)
2478{
2479 int i, compound, ret;
2480
2481 VM_BUG_ON_PAGE(PageTail(page), page);
2482
2483 if (likely(!PageCompound(page)))
2484 return atomic_read(&page->_mapcount) + 1;
2485
2486 compound = compound_mapcount(page);
2487 if (PageHuge(page))
2488 return compound;
2489 ret = compound;
2490 for (i = 0; i < HPAGE_PMD_NR; i++)
2491 ret += atomic_read(&page[i]._mapcount) + 1;
2492 /* File pages has compound_mapcount included in _mapcount */
2493 if (!PageAnon(page))
2494 return ret - compound * HPAGE_PMD_NR;
2495 if (PageDoubleMap(page))
2496 ret -= HPAGE_PMD_NR;
2497 return ret;
2498}
2499
2500/*
2501 * This calculates accurately how many mappings a transparent hugepage
2502 * has (unlike page_mapcount() which isn't fully accurate). This full
2503 * accuracy is primarily needed to know if copy-on-write faults can
2504 * reuse the page and change the mapping to read-write instead of
2505 * copying them. At the same time this returns the total_mapcount too.
2506 *
2507 * The function returns the highest mapcount any one of the subpages
2508 * has. If the return value is one, even if different processes are
2509 * mapping different subpages of the transparent hugepage, they can
2510 * all reuse it, because each process is reusing a different subpage.
2511 *
2512 * The total_mapcount is instead counting all virtual mappings of the
2513 * subpages. If the total_mapcount is equal to "one", it tells the
2514 * caller all mappings belong to the same "mm" and in turn the
2515 * anon_vma of the transparent hugepage can become the vma->anon_vma
2516 * local one as no other process may be mapping any of the subpages.
2517 *
2518 * It would be more accurate to replace page_mapcount() with
2519 * page_trans_huge_mapcount(), however we only use
2520 * page_trans_huge_mapcount() in the copy-on-write faults where we
2521 * need full accuracy to avoid breaking page pinning, because
2522 * page_trans_huge_mapcount() is slower than page_mapcount().
2523 */
2524int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2525{
2526 int i, ret, _total_mapcount, mapcount;
2527
2528 /* hugetlbfs shouldn't call it */
2529 VM_BUG_ON_PAGE(PageHuge(page), page);
2530
2531 if (likely(!PageTransCompound(page))) {
2532 mapcount = atomic_read(&page->_mapcount) + 1;
2533 if (total_mapcount)
2534 *total_mapcount = mapcount;
2535 return mapcount;
2536 }
2537
2538 page = compound_head(page);
2539
2540 _total_mapcount = ret = 0;
2541 for (i = 0; i < HPAGE_PMD_NR; i++) {
2542 mapcount = atomic_read(&page[i]._mapcount) + 1;
2543 ret = max(ret, mapcount);
2544 _total_mapcount += mapcount;
2545 }
2546 if (PageDoubleMap(page)) {
2547 ret -= 1;
2548 _total_mapcount -= HPAGE_PMD_NR;
2549 }
2550 mapcount = compound_mapcount(page);
2551 ret += mapcount;
2552 _total_mapcount += mapcount;
2553 if (total_mapcount)
2554 *total_mapcount = _total_mapcount;
2555 return ret;
2556}
2557
2558/* Racy check whether the huge page can be split */
2559bool can_split_huge_page(struct page *page, int *pextra_pins)
2560{
2561 int extra_pins;
2562
2563 /* Additional pins from radix tree */
2564 if (PageAnon(page))
2565 extra_pins = PageSwapCache(page) ? HPAGE_PMD_NR : 0;
2566 else
2567 extra_pins = HPAGE_PMD_NR;
2568 if (pextra_pins)
2569 *pextra_pins = extra_pins;
2570 return total_mapcount(page) == page_count(page) - extra_pins - 1;
2571}
2572
2573/*
2574 * This function splits huge page into normal pages. @page can point to any
2575 * subpage of huge page to split. Split doesn't change the position of @page.
2576 *
2577 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2578 * The huge page must be locked.
2579 *
2580 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2581 *
2582 * Both head page and tail pages will inherit mapping, flags, and so on from
2583 * the hugepage.
2584 *
2585 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2586 * they are not mapped.
2587 *
2588 * Returns 0 if the hugepage is split successfully.
2589 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2590 * us.
2591 */
2592int split_huge_page_to_list(struct page *page, struct list_head *list)
2593{
2594 struct page *head = compound_head(page);
2595 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2596 struct anon_vma *anon_vma = NULL;
2597 struct address_space *mapping = NULL;
2598 int count, mapcount, extra_pins, ret;
2599 bool mlocked;
2600 unsigned long flags;
2601
2602 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2603 VM_BUG_ON_PAGE(!PageLocked(page), page);
2604 VM_BUG_ON_PAGE(!PageCompound(page), page);
2605
2606 if (PageWriteback(page))
2607 return -EBUSY;
2608
2609 if (PageAnon(head)) {
2610 /*
2611 * The caller does not necessarily hold an mmap_sem that would
2612 * prevent the anon_vma disappearing so we first we take a
2613 * reference to it and then lock the anon_vma for write. This
2614 * is similar to page_lock_anon_vma_read except the write lock
2615 * is taken to serialise against parallel split or collapse
2616 * operations.
2617 */
2618 anon_vma = page_get_anon_vma(head);
2619 if (!anon_vma) {
2620 ret = -EBUSY;
2621 goto out;
2622 }
2623 mapping = NULL;
2624 anon_vma_lock_write(anon_vma);
2625 } else {
2626 mapping = head->mapping;
2627
2628 /* Truncated ? */
2629 if (!mapping) {
2630 ret = -EBUSY;
2631 goto out;
2632 }
2633
2634 anon_vma = NULL;
2635 i_mmap_lock_read(mapping);
2636 }
2637
2638 /*
2639 * Racy check if we can split the page, before freeze_page() will
2640 * split PMDs
2641 */
2642 if (!can_split_huge_page(head, &extra_pins)) {
2643 ret = -EBUSY;
2644 goto out_unlock;
2645 }
2646
2647 mlocked = PageMlocked(page);
2648 freeze_page(head);
2649 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2650
2651 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2652 if (mlocked)
2653 lru_add_drain();
2654
2655 /* prevent PageLRU to go away from under us, and freeze lru stats */
2656 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2657
2658 if (mapping) {
2659 void **pslot;
2660
2661 xa_lock(&mapping->i_pages);
2662 pslot = radix_tree_lookup_slot(&mapping->i_pages,
2663 page_index(head));
2664 /*
2665 * Check if the head page is present in radix tree.
2666 * We assume all tail are present too, if head is there.
2667 */
2668 if (radix_tree_deref_slot_protected(pslot,
2669 &mapping->i_pages.xa_lock) != head)
2670 goto fail;
2671 }
2672
2673 /* Prevent deferred_split_scan() touching ->_refcount */
2674 spin_lock(&pgdata->split_queue_lock);
2675 count = page_count(head);
2676 mapcount = total_mapcount(head);
2677 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2678 if (!list_empty(page_deferred_list(head))) {
2679 pgdata->split_queue_len--;
2680 list_del(page_deferred_list(head));
2681 }
2682 if (mapping)
2683 __dec_node_page_state(page, NR_SHMEM_THPS);
2684 spin_unlock(&pgdata->split_queue_lock);
2685 __split_huge_page(page, list, flags);
2686 if (PageSwapCache(head)) {
2687 swp_entry_t entry = { .val = page_private(head) };
2688
2689 ret = split_swap_cluster(entry);
2690 } else
2691 ret = 0;
2692 } else {
2693 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2694 pr_alert("total_mapcount: %u, page_count(): %u\n",
2695 mapcount, count);
2696 if (PageTail(page))
2697 dump_page(head, NULL);
2698 dump_page(page, "total_mapcount(head) > 0");
2699 BUG();
2700 }
2701 spin_unlock(&pgdata->split_queue_lock);
2702fail: if (mapping)
2703 xa_unlock(&mapping->i_pages);
2704 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2705 unfreeze_page(head);
2706 ret = -EBUSY;
2707 }
2708
2709out_unlock:
2710 if (anon_vma) {
2711 anon_vma_unlock_write(anon_vma);
2712 put_anon_vma(anon_vma);
2713 }
2714 if (mapping)
2715 i_mmap_unlock_read(mapping);
2716out:
2717 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2718 return ret;
2719}
2720
2721void free_transhuge_page(struct page *page)
2722{
2723 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2724 unsigned long flags;
2725
2726 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2727 if (!list_empty(page_deferred_list(page))) {
2728 pgdata->split_queue_len--;
2729 list_del(page_deferred_list(page));
2730 }
2731 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2732 free_compound_page(page);
2733}
2734
2735void deferred_split_huge_page(struct page *page)
2736{
2737 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2738 unsigned long flags;
2739
2740 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2741
2742 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2743 if (list_empty(page_deferred_list(page))) {
2744 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2745 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2746 pgdata->split_queue_len++;
2747 }
2748 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2749}
2750
2751static unsigned long deferred_split_count(struct shrinker *shrink,
2752 struct shrink_control *sc)
2753{
2754 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2755 return READ_ONCE(pgdata->split_queue_len);
2756}
2757
2758static unsigned long deferred_split_scan(struct shrinker *shrink,
2759 struct shrink_control *sc)
2760{
2761 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2762 unsigned long flags;
2763 LIST_HEAD(list), *pos, *next;
2764 struct page *page;
2765 int split = 0;
2766
2767 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2768 /* Take pin on all head pages to avoid freeing them under us */
2769 list_for_each_safe(pos, next, &pgdata->split_queue) {
2770 page = list_entry((void *)pos, struct page, mapping);
2771 page = compound_head(page);
2772 if (get_page_unless_zero(page)) {
2773 list_move(page_deferred_list(page), &list);
2774 } else {
2775 /* We lost race with put_compound_page() */
2776 list_del_init(page_deferred_list(page));
2777 pgdata->split_queue_len--;
2778 }
2779 if (!--sc->nr_to_scan)
2780 break;
2781 }
2782 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2783
2784 list_for_each_safe(pos, next, &list) {
2785 page = list_entry((void *)pos, struct page, mapping);
2786 if (!trylock_page(page))
2787 goto next;
2788 /* split_huge_page() removes page from list on success */
2789 if (!split_huge_page(page))
2790 split++;
2791 unlock_page(page);
2792next:
2793 put_page(page);
2794 }
2795
2796 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2797 list_splice_tail(&list, &pgdata->split_queue);
2798 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2799
2800 /*
2801 * Stop shrinker if we didn't split any page, but the queue is empty.
2802 * This can happen if pages were freed under us.
2803 */
2804 if (!split && list_empty(&pgdata->split_queue))
2805 return SHRINK_STOP;
2806 return split;
2807}
2808
2809static struct shrinker deferred_split_shrinker = {
2810 .count_objects = deferred_split_count,
2811 .scan_objects = deferred_split_scan,
2812 .seeks = DEFAULT_SEEKS,
2813 .flags = SHRINKER_NUMA_AWARE,
2814};
2815
2816#ifdef CONFIG_DEBUG_FS
2817static int split_huge_pages_set(void *data, u64 val)
2818{
2819 struct zone *zone;
2820 struct page *page;
2821 unsigned long pfn, max_zone_pfn;
2822 unsigned long total = 0, split = 0;
2823
2824 if (val != 1)
2825 return -EINVAL;
2826
2827 for_each_populated_zone(zone) {
2828 max_zone_pfn = zone_end_pfn(zone);
2829 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2830 if (!pfn_valid(pfn))
2831 continue;
2832
2833 page = pfn_to_page(pfn);
2834 if (!get_page_unless_zero(page))
2835 continue;
2836
2837 if (zone != page_zone(page))
2838 goto next;
2839
2840 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2841 goto next;
2842
2843 total++;
2844 lock_page(page);
2845 if (!split_huge_page(page))
2846 split++;
2847 unlock_page(page);
2848next:
2849 put_page(page);
2850 }
2851 }
2852
2853 pr_info("%lu of %lu THP split\n", split, total);
2854
2855 return 0;
2856}
2857DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2858 "%llu\n");
2859
2860static int __init split_huge_pages_debugfs(void)
2861{
2862 void *ret;
2863
2864 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2865 &split_huge_pages_fops);
2866 if (!ret)
2867 pr_warn("Failed to create split_huge_pages in debugfs");
2868 return 0;
2869}
2870late_initcall(split_huge_pages_debugfs);
2871#endif
2872
2873#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION
2874void set_pmd_migration_entry(struct page_vma_mapped_walk *pvmw,
2875 struct page *page)
2876{
2877 struct vm_area_struct *vma = pvmw->vma;
2878 struct mm_struct *mm = vma->vm_mm;
2879 unsigned long address = pvmw->address;
2880 pmd_t pmdval;
2881 swp_entry_t entry;
2882 pmd_t pmdswp;
2883
2884 if (!(pvmw->pmd && !pvmw->pte))
2885 return;
2886
2887 mmu_notifier_invalidate_range_start(mm, address,
2888 address + HPAGE_PMD_SIZE);
2889
2890 flush_cache_range(vma, address, address + HPAGE_PMD_SIZE);
2891 pmdval = *pvmw->pmd;
2892 pmdp_invalidate(vma, address, pvmw->pmd);
2893 if (pmd_dirty(pmdval))
2894 set_page_dirty(page);
2895 entry = make_migration_entry(page, pmd_write(pmdval));
2896 pmdswp = swp_entry_to_pmd(entry);
2897 if (pmd_soft_dirty(pmdval))
2898 pmdswp = pmd_swp_mksoft_dirty(pmdswp);
2899 set_pmd_at(mm, address, pvmw->pmd, pmdswp);
2900 page_remove_rmap(page, true);
2901 put_page(page);
2902
2903 mmu_notifier_invalidate_range_end(mm, address,
2904 address + HPAGE_PMD_SIZE);
2905}
2906
2907void remove_migration_pmd(struct page_vma_mapped_walk *pvmw, struct page *new)
2908{
2909 struct vm_area_struct *vma = pvmw->vma;
2910 struct mm_struct *mm = vma->vm_mm;
2911 unsigned long address = pvmw->address;
2912 unsigned long mmun_start = address & HPAGE_PMD_MASK;
2913 pmd_t pmde;
2914 swp_entry_t entry;
2915
2916 if (!(pvmw->pmd && !pvmw->pte))
2917 return;
2918
2919 entry = pmd_to_swp_entry(*pvmw->pmd);
2920 get_page(new);
2921 pmde = pmd_mkold(mk_huge_pmd(new, vma->vm_page_prot));
2922 if (pmd_swp_soft_dirty(*pvmw->pmd))
2923 pmde = pmd_mksoft_dirty(pmde);
2924 if (is_write_migration_entry(entry))
2925 pmde = maybe_pmd_mkwrite(pmde, vma);
2926
2927 flush_cache_range(vma, mmun_start, mmun_start + HPAGE_PMD_SIZE);
2928 if (PageAnon(new))
2929 page_add_anon_rmap(new, vma, mmun_start, true);
2930 else
2931 page_add_file_rmap(new, true);
2932 set_pmd_at(mm, mmun_start, pvmw->pmd, pmde);
2933 if (vma->vm_flags & VM_LOCKED)
2934 mlock_vma_page(new);
2935 update_mmu_cache_pmd(vma, address, pvmw->pmd);
2936}
2937#endif
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
34#include <asm/tlb.h>
35#include <asm/pgalloc.h>
36#include "internal.h"
37
38/*
39 * By default transparent hugepage support is disabled in order that avoid
40 * to risk increase the memory footprint of applications without a guaranteed
41 * benefit. When transparent hugepage support is enabled, is for all mappings,
42 * and khugepaged scans all mappings.
43 * Defrag is invoked by khugepaged hugepage allocations and by page faults
44 * for all hugepage allocations.
45 */
46unsigned long transparent_hugepage_flags __read_mostly =
47#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
48 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
49#endif
50#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
51 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
52#endif
53 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
54 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
55 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
56
57static struct shrinker deferred_split_shrinker;
58
59static atomic_t huge_zero_refcount;
60struct page *huge_zero_page __read_mostly;
61
62static struct page *get_huge_zero_page(void)
63{
64 struct page *zero_page;
65retry:
66 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
67 return READ_ONCE(huge_zero_page);
68
69 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
70 HPAGE_PMD_ORDER);
71 if (!zero_page) {
72 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
73 return NULL;
74 }
75 count_vm_event(THP_ZERO_PAGE_ALLOC);
76 preempt_disable();
77 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
78 preempt_enable();
79 __free_pages(zero_page, compound_order(zero_page));
80 goto retry;
81 }
82
83 /* We take additional reference here. It will be put back by shrinker */
84 atomic_set(&huge_zero_refcount, 2);
85 preempt_enable();
86 return READ_ONCE(huge_zero_page);
87}
88
89static void put_huge_zero_page(void)
90{
91 /*
92 * Counter should never go to zero here. Only shrinker can put
93 * last reference.
94 */
95 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
96}
97
98struct page *mm_get_huge_zero_page(struct mm_struct *mm)
99{
100 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
101 return READ_ONCE(huge_zero_page);
102
103 if (!get_huge_zero_page())
104 return NULL;
105
106 if (test_and_set_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
107 put_huge_zero_page();
108
109 return READ_ONCE(huge_zero_page);
110}
111
112void mm_put_huge_zero_page(struct mm_struct *mm)
113{
114 if (test_bit(MMF_HUGE_ZERO_PAGE, &mm->flags))
115 put_huge_zero_page();
116}
117
118static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
119 struct shrink_control *sc)
120{
121 /* we can free zero page only if last reference remains */
122 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
123}
124
125static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
126 struct shrink_control *sc)
127{
128 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
129 struct page *zero_page = xchg(&huge_zero_page, NULL);
130 BUG_ON(zero_page == NULL);
131 __free_pages(zero_page, compound_order(zero_page));
132 return HPAGE_PMD_NR;
133 }
134
135 return 0;
136}
137
138static struct shrinker huge_zero_page_shrinker = {
139 .count_objects = shrink_huge_zero_page_count,
140 .scan_objects = shrink_huge_zero_page_scan,
141 .seeks = DEFAULT_SEEKS,
142};
143
144#ifdef CONFIG_SYSFS
145
146static ssize_t triple_flag_store(struct kobject *kobj,
147 struct kobj_attribute *attr,
148 const char *buf, size_t count,
149 enum transparent_hugepage_flag enabled,
150 enum transparent_hugepage_flag deferred,
151 enum transparent_hugepage_flag req_madv)
152{
153 if (!memcmp("defer", buf,
154 min(sizeof("defer")-1, count))) {
155 if (enabled == deferred)
156 return -EINVAL;
157 clear_bit(enabled, &transparent_hugepage_flags);
158 clear_bit(req_madv, &transparent_hugepage_flags);
159 set_bit(deferred, &transparent_hugepage_flags);
160 } else if (!memcmp("always", buf,
161 min(sizeof("always")-1, count))) {
162 clear_bit(deferred, &transparent_hugepage_flags);
163 clear_bit(req_madv, &transparent_hugepage_flags);
164 set_bit(enabled, &transparent_hugepage_flags);
165 } else if (!memcmp("madvise", buf,
166 min(sizeof("madvise")-1, count))) {
167 clear_bit(enabled, &transparent_hugepage_flags);
168 clear_bit(deferred, &transparent_hugepage_flags);
169 set_bit(req_madv, &transparent_hugepage_flags);
170 } else if (!memcmp("never", buf,
171 min(sizeof("never")-1, count))) {
172 clear_bit(enabled, &transparent_hugepage_flags);
173 clear_bit(req_madv, &transparent_hugepage_flags);
174 clear_bit(deferred, &transparent_hugepage_flags);
175 } else
176 return -EINVAL;
177
178 return count;
179}
180
181static ssize_t enabled_show(struct kobject *kobj,
182 struct kobj_attribute *attr, char *buf)
183{
184 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
185 return sprintf(buf, "[always] madvise never\n");
186 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
187 return sprintf(buf, "always [madvise] never\n");
188 else
189 return sprintf(buf, "always madvise [never]\n");
190}
191
192static ssize_t enabled_store(struct kobject *kobj,
193 struct kobj_attribute *attr,
194 const char *buf, size_t count)
195{
196 ssize_t ret;
197
198 ret = triple_flag_store(kobj, attr, buf, count,
199 TRANSPARENT_HUGEPAGE_FLAG,
200 TRANSPARENT_HUGEPAGE_FLAG,
201 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
202
203 if (ret > 0) {
204 int err = start_stop_khugepaged();
205 if (err)
206 ret = err;
207 }
208
209 return ret;
210}
211static struct kobj_attribute enabled_attr =
212 __ATTR(enabled, 0644, enabled_show, enabled_store);
213
214ssize_t single_hugepage_flag_show(struct kobject *kobj,
215 struct kobj_attribute *attr, char *buf,
216 enum transparent_hugepage_flag flag)
217{
218 return sprintf(buf, "%d\n",
219 !!test_bit(flag, &transparent_hugepage_flags));
220}
221
222ssize_t single_hugepage_flag_store(struct kobject *kobj,
223 struct kobj_attribute *attr,
224 const char *buf, size_t count,
225 enum transparent_hugepage_flag flag)
226{
227 unsigned long value;
228 int ret;
229
230 ret = kstrtoul(buf, 10, &value);
231 if (ret < 0)
232 return ret;
233 if (value > 1)
234 return -EINVAL;
235
236 if (value)
237 set_bit(flag, &transparent_hugepage_flags);
238 else
239 clear_bit(flag, &transparent_hugepage_flags);
240
241 return count;
242}
243
244/*
245 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
246 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
247 * memory just to allocate one more hugepage.
248 */
249static ssize_t defrag_show(struct kobject *kobj,
250 struct kobj_attribute *attr, char *buf)
251{
252 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
253 return sprintf(buf, "[always] defer madvise never\n");
254 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
255 return sprintf(buf, "always [defer] madvise never\n");
256 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
257 return sprintf(buf, "always defer [madvise] never\n");
258 else
259 return sprintf(buf, "always defer madvise [never]\n");
260
261}
262static ssize_t defrag_store(struct kobject *kobj,
263 struct kobj_attribute *attr,
264 const char *buf, size_t count)
265{
266 return triple_flag_store(kobj, attr, buf, count,
267 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
268 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
269 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
270}
271static struct kobj_attribute defrag_attr =
272 __ATTR(defrag, 0644, defrag_show, defrag_store);
273
274static ssize_t use_zero_page_show(struct kobject *kobj,
275 struct kobj_attribute *attr, char *buf)
276{
277 return single_hugepage_flag_show(kobj, attr, buf,
278 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
279}
280static ssize_t use_zero_page_store(struct kobject *kobj,
281 struct kobj_attribute *attr, const char *buf, size_t count)
282{
283 return single_hugepage_flag_store(kobj, attr, buf, count,
284 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
285}
286static struct kobj_attribute use_zero_page_attr =
287 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
288
289static ssize_t hpage_pmd_size_show(struct kobject *kobj,
290 struct kobj_attribute *attr, char *buf)
291{
292 return sprintf(buf, "%lu\n", HPAGE_PMD_SIZE);
293}
294static struct kobj_attribute hpage_pmd_size_attr =
295 __ATTR_RO(hpage_pmd_size);
296
297#ifdef CONFIG_DEBUG_VM
298static ssize_t debug_cow_show(struct kobject *kobj,
299 struct kobj_attribute *attr, char *buf)
300{
301 return single_hugepage_flag_show(kobj, attr, buf,
302 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
303}
304static ssize_t debug_cow_store(struct kobject *kobj,
305 struct kobj_attribute *attr,
306 const char *buf, size_t count)
307{
308 return single_hugepage_flag_store(kobj, attr, buf, count,
309 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
310}
311static struct kobj_attribute debug_cow_attr =
312 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
313#endif /* CONFIG_DEBUG_VM */
314
315static struct attribute *hugepage_attr[] = {
316 &enabled_attr.attr,
317 &defrag_attr.attr,
318 &use_zero_page_attr.attr,
319 &hpage_pmd_size_attr.attr,
320#if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
321 &shmem_enabled_attr.attr,
322#endif
323#ifdef CONFIG_DEBUG_VM
324 &debug_cow_attr.attr,
325#endif
326 NULL,
327};
328
329static struct attribute_group hugepage_attr_group = {
330 .attrs = hugepage_attr,
331};
332
333static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
334{
335 int err;
336
337 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
338 if (unlikely(!*hugepage_kobj)) {
339 pr_err("failed to create transparent hugepage kobject\n");
340 return -ENOMEM;
341 }
342
343 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
344 if (err) {
345 pr_err("failed to register transparent hugepage group\n");
346 goto delete_obj;
347 }
348
349 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
350 if (err) {
351 pr_err("failed to register transparent hugepage group\n");
352 goto remove_hp_group;
353 }
354
355 return 0;
356
357remove_hp_group:
358 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
359delete_obj:
360 kobject_put(*hugepage_kobj);
361 return err;
362}
363
364static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
365{
366 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
367 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
368 kobject_put(hugepage_kobj);
369}
370#else
371static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
372{
373 return 0;
374}
375
376static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
377{
378}
379#endif /* CONFIG_SYSFS */
380
381static int __init hugepage_init(void)
382{
383 int err;
384 struct kobject *hugepage_kobj;
385
386 if (!has_transparent_hugepage()) {
387 transparent_hugepage_flags = 0;
388 return -EINVAL;
389 }
390
391 /*
392 * hugepages can't be allocated by the buddy allocator
393 */
394 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
395 /*
396 * we use page->mapping and page->index in second tail page
397 * as list_head: assuming THP order >= 2
398 */
399 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
400
401 err = hugepage_init_sysfs(&hugepage_kobj);
402 if (err)
403 goto err_sysfs;
404
405 err = khugepaged_init();
406 if (err)
407 goto err_slab;
408
409 err = register_shrinker(&huge_zero_page_shrinker);
410 if (err)
411 goto err_hzp_shrinker;
412 err = register_shrinker(&deferred_split_shrinker);
413 if (err)
414 goto err_split_shrinker;
415
416 /*
417 * By default disable transparent hugepages on smaller systems,
418 * where the extra memory used could hurt more than TLB overhead
419 * is likely to save. The admin can still enable it through /sys.
420 */
421 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
422 transparent_hugepage_flags = 0;
423 return 0;
424 }
425
426 err = start_stop_khugepaged();
427 if (err)
428 goto err_khugepaged;
429
430 return 0;
431err_khugepaged:
432 unregister_shrinker(&deferred_split_shrinker);
433err_split_shrinker:
434 unregister_shrinker(&huge_zero_page_shrinker);
435err_hzp_shrinker:
436 khugepaged_destroy();
437err_slab:
438 hugepage_exit_sysfs(hugepage_kobj);
439err_sysfs:
440 return err;
441}
442subsys_initcall(hugepage_init);
443
444static int __init setup_transparent_hugepage(char *str)
445{
446 int ret = 0;
447 if (!str)
448 goto out;
449 if (!strcmp(str, "always")) {
450 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
451 &transparent_hugepage_flags);
452 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
453 &transparent_hugepage_flags);
454 ret = 1;
455 } else if (!strcmp(str, "madvise")) {
456 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
457 &transparent_hugepage_flags);
458 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
459 &transparent_hugepage_flags);
460 ret = 1;
461 } else if (!strcmp(str, "never")) {
462 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
463 &transparent_hugepage_flags);
464 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
465 &transparent_hugepage_flags);
466 ret = 1;
467 }
468out:
469 if (!ret)
470 pr_warn("transparent_hugepage= cannot parse, ignored\n");
471 return ret;
472}
473__setup("transparent_hugepage=", setup_transparent_hugepage);
474
475pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
476{
477 if (likely(vma->vm_flags & VM_WRITE))
478 pmd = pmd_mkwrite(pmd);
479 return pmd;
480}
481
482static inline struct list_head *page_deferred_list(struct page *page)
483{
484 /*
485 * ->lru in the tail pages is occupied by compound_head.
486 * Let's use ->mapping + ->index in the second tail page as list_head.
487 */
488 return (struct list_head *)&page[2].mapping;
489}
490
491void prep_transhuge_page(struct page *page)
492{
493 /*
494 * we use page->mapping and page->indexlru in second tail page
495 * as list_head: assuming THP order >= 2
496 */
497
498 INIT_LIST_HEAD(page_deferred_list(page));
499 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
500}
501
502unsigned long __thp_get_unmapped_area(struct file *filp, unsigned long len,
503 loff_t off, unsigned long flags, unsigned long size)
504{
505 unsigned long addr;
506 loff_t off_end = off + len;
507 loff_t off_align = round_up(off, size);
508 unsigned long len_pad;
509
510 if (off_end <= off_align || (off_end - off_align) < size)
511 return 0;
512
513 len_pad = len + size;
514 if (len_pad < len || (off + len_pad) < off)
515 return 0;
516
517 addr = current->mm->get_unmapped_area(filp, 0, len_pad,
518 off >> PAGE_SHIFT, flags);
519 if (IS_ERR_VALUE(addr))
520 return 0;
521
522 addr += (off - addr) & (size - 1);
523 return addr;
524}
525
526unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
527 unsigned long len, unsigned long pgoff, unsigned long flags)
528{
529 loff_t off = (loff_t)pgoff << PAGE_SHIFT;
530
531 if (addr)
532 goto out;
533 if (!IS_DAX(filp->f_mapping->host) || !IS_ENABLED(CONFIG_FS_DAX_PMD))
534 goto out;
535
536 addr = __thp_get_unmapped_area(filp, len, off, flags, PMD_SIZE);
537 if (addr)
538 return addr;
539
540 out:
541 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
542}
543EXPORT_SYMBOL_GPL(thp_get_unmapped_area);
544
545static int __do_huge_pmd_anonymous_page(struct vm_fault *vmf, struct page *page,
546 gfp_t gfp)
547{
548 struct vm_area_struct *vma = vmf->vma;
549 struct mem_cgroup *memcg;
550 pgtable_t pgtable;
551 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
552
553 VM_BUG_ON_PAGE(!PageCompound(page), page);
554
555 if (mem_cgroup_try_charge(page, vma->vm_mm, gfp, &memcg, true)) {
556 put_page(page);
557 count_vm_event(THP_FAULT_FALLBACK);
558 return VM_FAULT_FALLBACK;
559 }
560
561 pgtable = pte_alloc_one(vma->vm_mm, haddr);
562 if (unlikely(!pgtable)) {
563 mem_cgroup_cancel_charge(page, memcg, true);
564 put_page(page);
565 return VM_FAULT_OOM;
566 }
567
568 clear_huge_page(page, haddr, HPAGE_PMD_NR);
569 /*
570 * The memory barrier inside __SetPageUptodate makes sure that
571 * clear_huge_page writes become visible before the set_pmd_at()
572 * write.
573 */
574 __SetPageUptodate(page);
575
576 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
577 if (unlikely(!pmd_none(*vmf->pmd))) {
578 spin_unlock(vmf->ptl);
579 mem_cgroup_cancel_charge(page, memcg, true);
580 put_page(page);
581 pte_free(vma->vm_mm, pgtable);
582 } else {
583 pmd_t entry;
584
585 /* Deliver the page fault to userland */
586 if (userfaultfd_missing(vma)) {
587 int ret;
588
589 spin_unlock(vmf->ptl);
590 mem_cgroup_cancel_charge(page, memcg, true);
591 put_page(page);
592 pte_free(vma->vm_mm, pgtable);
593 ret = handle_userfault(vmf, VM_UFFD_MISSING);
594 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
595 return ret;
596 }
597
598 entry = mk_huge_pmd(page, vma->vm_page_prot);
599 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
600 page_add_new_anon_rmap(page, vma, haddr, true);
601 mem_cgroup_commit_charge(page, memcg, false, true);
602 lru_cache_add_active_or_unevictable(page, vma);
603 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, pgtable);
604 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
605 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
606 atomic_long_inc(&vma->vm_mm->nr_ptes);
607 spin_unlock(vmf->ptl);
608 count_vm_event(THP_FAULT_ALLOC);
609 }
610
611 return 0;
612}
613
614/*
615 * If THP defrag is set to always then directly reclaim/compact as necessary
616 * If set to defer then do only background reclaim/compact and defer to khugepaged
617 * If set to madvise and the VMA is flagged then directly reclaim/compact
618 * When direct reclaim/compact is allowed, don't retry except for flagged VMA's
619 */
620static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
621{
622 bool vma_madvised = !!(vma->vm_flags & VM_HUGEPAGE);
623
624 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
625 &transparent_hugepage_flags) && vma_madvised)
626 return GFP_TRANSHUGE;
627 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
628 &transparent_hugepage_flags))
629 return GFP_TRANSHUGE_LIGHT | __GFP_KSWAPD_RECLAIM;
630 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
631 &transparent_hugepage_flags))
632 return GFP_TRANSHUGE | (vma_madvised ? 0 : __GFP_NORETRY);
633
634 return GFP_TRANSHUGE_LIGHT;
635}
636
637/* Caller must hold page table lock. */
638static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
639 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
640 struct page *zero_page)
641{
642 pmd_t entry;
643 if (!pmd_none(*pmd))
644 return false;
645 entry = mk_pmd(zero_page, vma->vm_page_prot);
646 entry = pmd_mkhuge(entry);
647 if (pgtable)
648 pgtable_trans_huge_deposit(mm, pmd, pgtable);
649 set_pmd_at(mm, haddr, pmd, entry);
650 atomic_long_inc(&mm->nr_ptes);
651 return true;
652}
653
654int do_huge_pmd_anonymous_page(struct vm_fault *vmf)
655{
656 struct vm_area_struct *vma = vmf->vma;
657 gfp_t gfp;
658 struct page *page;
659 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
660
661 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
662 return VM_FAULT_FALLBACK;
663 if (unlikely(anon_vma_prepare(vma)))
664 return VM_FAULT_OOM;
665 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
666 return VM_FAULT_OOM;
667 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
668 !mm_forbids_zeropage(vma->vm_mm) &&
669 transparent_hugepage_use_zero_page()) {
670 pgtable_t pgtable;
671 struct page *zero_page;
672 bool set;
673 int ret;
674 pgtable = pte_alloc_one(vma->vm_mm, haddr);
675 if (unlikely(!pgtable))
676 return VM_FAULT_OOM;
677 zero_page = mm_get_huge_zero_page(vma->vm_mm);
678 if (unlikely(!zero_page)) {
679 pte_free(vma->vm_mm, pgtable);
680 count_vm_event(THP_FAULT_FALLBACK);
681 return VM_FAULT_FALLBACK;
682 }
683 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
684 ret = 0;
685 set = false;
686 if (pmd_none(*vmf->pmd)) {
687 if (userfaultfd_missing(vma)) {
688 spin_unlock(vmf->ptl);
689 ret = handle_userfault(vmf, VM_UFFD_MISSING);
690 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
691 } else {
692 set_huge_zero_page(pgtable, vma->vm_mm, vma,
693 haddr, vmf->pmd, zero_page);
694 spin_unlock(vmf->ptl);
695 set = true;
696 }
697 } else
698 spin_unlock(vmf->ptl);
699 if (!set)
700 pte_free(vma->vm_mm, pgtable);
701 return ret;
702 }
703 gfp = alloc_hugepage_direct_gfpmask(vma);
704 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
705 if (unlikely(!page)) {
706 count_vm_event(THP_FAULT_FALLBACK);
707 return VM_FAULT_FALLBACK;
708 }
709 prep_transhuge_page(page);
710 return __do_huge_pmd_anonymous_page(vmf, page, gfp);
711}
712
713static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
714 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
715{
716 struct mm_struct *mm = vma->vm_mm;
717 pmd_t entry;
718 spinlock_t *ptl;
719
720 ptl = pmd_lock(mm, pmd);
721 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
722 if (pfn_t_devmap(pfn))
723 entry = pmd_mkdevmap(entry);
724 if (write) {
725 entry = pmd_mkyoung(pmd_mkdirty(entry));
726 entry = maybe_pmd_mkwrite(entry, vma);
727 }
728 set_pmd_at(mm, addr, pmd, entry);
729 update_mmu_cache_pmd(vma, addr, pmd);
730 spin_unlock(ptl);
731}
732
733int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
734 pmd_t *pmd, pfn_t pfn, bool write)
735{
736 pgprot_t pgprot = vma->vm_page_prot;
737 /*
738 * If we had pmd_special, we could avoid all these restrictions,
739 * but we need to be consistent with PTEs and architectures that
740 * can't support a 'special' bit.
741 */
742 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
743 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
744 (VM_PFNMAP|VM_MIXEDMAP));
745 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
746 BUG_ON(!pfn_t_devmap(pfn));
747
748 if (addr < vma->vm_start || addr >= vma->vm_end)
749 return VM_FAULT_SIGBUS;
750
751 track_pfn_insert(vma, &pgprot, pfn);
752
753 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
754 return VM_FAULT_NOPAGE;
755}
756EXPORT_SYMBOL_GPL(vmf_insert_pfn_pmd);
757
758static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
759 pmd_t *pmd)
760{
761 pmd_t _pmd;
762
763 /*
764 * We should set the dirty bit only for FOLL_WRITE but for now
765 * the dirty bit in the pmd is meaningless. And if the dirty
766 * bit will become meaningful and we'll only set it with
767 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
768 * set the young bit, instead of the current set_pmd_at.
769 */
770 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
771 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
772 pmd, _pmd, 1))
773 update_mmu_cache_pmd(vma, addr, pmd);
774}
775
776struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
777 pmd_t *pmd, int flags)
778{
779 unsigned long pfn = pmd_pfn(*pmd);
780 struct mm_struct *mm = vma->vm_mm;
781 struct dev_pagemap *pgmap;
782 struct page *page;
783
784 assert_spin_locked(pmd_lockptr(mm, pmd));
785
786 /*
787 * When we COW a devmap PMD entry, we split it into PTEs, so we should
788 * not be in this function with `flags & FOLL_COW` set.
789 */
790 WARN_ONCE(flags & FOLL_COW, "mm: In follow_devmap_pmd with FOLL_COW set");
791
792 if (flags & FOLL_WRITE && !pmd_write(*pmd))
793 return NULL;
794
795 if (pmd_present(*pmd) && pmd_devmap(*pmd))
796 /* pass */;
797 else
798 return NULL;
799
800 if (flags & FOLL_TOUCH)
801 touch_pmd(vma, addr, pmd);
802
803 /*
804 * device mapped pages can only be returned if the
805 * caller will manage the page reference count.
806 */
807 if (!(flags & FOLL_GET))
808 return ERR_PTR(-EEXIST);
809
810 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
811 pgmap = get_dev_pagemap(pfn, NULL);
812 if (!pgmap)
813 return ERR_PTR(-EFAULT);
814 page = pfn_to_page(pfn);
815 get_page(page);
816 put_dev_pagemap(pgmap);
817
818 return page;
819}
820
821int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
822 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
823 struct vm_area_struct *vma)
824{
825 spinlock_t *dst_ptl, *src_ptl;
826 struct page *src_page;
827 pmd_t pmd;
828 pgtable_t pgtable = NULL;
829 int ret = -ENOMEM;
830
831 /* Skip if can be re-fill on fault */
832 if (!vma_is_anonymous(vma))
833 return 0;
834
835 pgtable = pte_alloc_one(dst_mm, addr);
836 if (unlikely(!pgtable))
837 goto out;
838
839 dst_ptl = pmd_lock(dst_mm, dst_pmd);
840 src_ptl = pmd_lockptr(src_mm, src_pmd);
841 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
842
843 ret = -EAGAIN;
844 pmd = *src_pmd;
845 if (unlikely(!pmd_trans_huge(pmd))) {
846 pte_free(dst_mm, pgtable);
847 goto out_unlock;
848 }
849 /*
850 * When page table lock is held, the huge zero pmd should not be
851 * under splitting since we don't split the page itself, only pmd to
852 * a page table.
853 */
854 if (is_huge_zero_pmd(pmd)) {
855 struct page *zero_page;
856 /*
857 * get_huge_zero_page() will never allocate a new page here,
858 * since we already have a zero page to copy. It just takes a
859 * reference.
860 */
861 zero_page = mm_get_huge_zero_page(dst_mm);
862 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
863 zero_page);
864 ret = 0;
865 goto out_unlock;
866 }
867
868 src_page = pmd_page(pmd);
869 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
870 get_page(src_page);
871 page_dup_rmap(src_page, true);
872 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
873 atomic_long_inc(&dst_mm->nr_ptes);
874 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
875
876 pmdp_set_wrprotect(src_mm, addr, src_pmd);
877 pmd = pmd_mkold(pmd_wrprotect(pmd));
878 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
879
880 ret = 0;
881out_unlock:
882 spin_unlock(src_ptl);
883 spin_unlock(dst_ptl);
884out:
885 return ret;
886}
887
888void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd)
889{
890 pmd_t entry;
891 unsigned long haddr;
892 bool write = vmf->flags & FAULT_FLAG_WRITE;
893
894 vmf->ptl = pmd_lock(vmf->vma->vm_mm, vmf->pmd);
895 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
896 goto unlock;
897
898 entry = pmd_mkyoung(orig_pmd);
899 if (write)
900 entry = pmd_mkdirty(entry);
901 haddr = vmf->address & HPAGE_PMD_MASK;
902 if (pmdp_set_access_flags(vmf->vma, haddr, vmf->pmd, entry, write))
903 update_mmu_cache_pmd(vmf->vma, vmf->address, vmf->pmd);
904
905unlock:
906 spin_unlock(vmf->ptl);
907}
908
909static int do_huge_pmd_wp_page_fallback(struct vm_fault *vmf, pmd_t orig_pmd,
910 struct page *page)
911{
912 struct vm_area_struct *vma = vmf->vma;
913 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
914 struct mem_cgroup *memcg;
915 pgtable_t pgtable;
916 pmd_t _pmd;
917 int ret = 0, i;
918 struct page **pages;
919 unsigned long mmun_start; /* For mmu_notifiers */
920 unsigned long mmun_end; /* For mmu_notifiers */
921
922 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
923 GFP_KERNEL);
924 if (unlikely(!pages)) {
925 ret |= VM_FAULT_OOM;
926 goto out;
927 }
928
929 for (i = 0; i < HPAGE_PMD_NR; i++) {
930 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE, vma,
931 vmf->address, page_to_nid(page));
932 if (unlikely(!pages[i] ||
933 mem_cgroup_try_charge(pages[i], vma->vm_mm,
934 GFP_KERNEL, &memcg, false))) {
935 if (pages[i])
936 put_page(pages[i]);
937 while (--i >= 0) {
938 memcg = (void *)page_private(pages[i]);
939 set_page_private(pages[i], 0);
940 mem_cgroup_cancel_charge(pages[i], memcg,
941 false);
942 put_page(pages[i]);
943 }
944 kfree(pages);
945 ret |= VM_FAULT_OOM;
946 goto out;
947 }
948 set_page_private(pages[i], (unsigned long)memcg);
949 }
950
951 for (i = 0; i < HPAGE_PMD_NR; i++) {
952 copy_user_highpage(pages[i], page + i,
953 haddr + PAGE_SIZE * i, vma);
954 __SetPageUptodate(pages[i]);
955 cond_resched();
956 }
957
958 mmun_start = haddr;
959 mmun_end = haddr + HPAGE_PMD_SIZE;
960 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
961
962 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
963 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
964 goto out_free_pages;
965 VM_BUG_ON_PAGE(!PageHead(page), page);
966
967 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
968 /* leave pmd empty until pte is filled */
969
970 pgtable = pgtable_trans_huge_withdraw(vma->vm_mm, vmf->pmd);
971 pmd_populate(vma->vm_mm, &_pmd, pgtable);
972
973 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
974 pte_t entry;
975 entry = mk_pte(pages[i], vma->vm_page_prot);
976 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
977 memcg = (void *)page_private(pages[i]);
978 set_page_private(pages[i], 0);
979 page_add_new_anon_rmap(pages[i], vmf->vma, haddr, false);
980 mem_cgroup_commit_charge(pages[i], memcg, false, false);
981 lru_cache_add_active_or_unevictable(pages[i], vma);
982 vmf->pte = pte_offset_map(&_pmd, haddr);
983 VM_BUG_ON(!pte_none(*vmf->pte));
984 set_pte_at(vma->vm_mm, haddr, vmf->pte, entry);
985 pte_unmap(vmf->pte);
986 }
987 kfree(pages);
988
989 smp_wmb(); /* make pte visible before pmd */
990 pmd_populate(vma->vm_mm, vmf->pmd, pgtable);
991 page_remove_rmap(page, true);
992 spin_unlock(vmf->ptl);
993
994 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
995
996 ret |= VM_FAULT_WRITE;
997 put_page(page);
998
999out:
1000 return ret;
1001
1002out_free_pages:
1003 spin_unlock(vmf->ptl);
1004 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1005 for (i = 0; i < HPAGE_PMD_NR; i++) {
1006 memcg = (void *)page_private(pages[i]);
1007 set_page_private(pages[i], 0);
1008 mem_cgroup_cancel_charge(pages[i], memcg, false);
1009 put_page(pages[i]);
1010 }
1011 kfree(pages);
1012 goto out;
1013}
1014
1015int do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd)
1016{
1017 struct vm_area_struct *vma = vmf->vma;
1018 struct page *page = NULL, *new_page;
1019 struct mem_cgroup *memcg;
1020 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1021 unsigned long mmun_start; /* For mmu_notifiers */
1022 unsigned long mmun_end; /* For mmu_notifiers */
1023 gfp_t huge_gfp; /* for allocation and charge */
1024 int ret = 0;
1025
1026 vmf->ptl = pmd_lockptr(vma->vm_mm, vmf->pmd);
1027 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1028 if (is_huge_zero_pmd(orig_pmd))
1029 goto alloc;
1030 spin_lock(vmf->ptl);
1031 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd)))
1032 goto out_unlock;
1033
1034 page = pmd_page(orig_pmd);
1035 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1036 /*
1037 * We can only reuse the page if nobody else maps the huge page or it's
1038 * part.
1039 */
1040 if (page_trans_huge_mapcount(page, NULL) == 1) {
1041 pmd_t entry;
1042 entry = pmd_mkyoung(orig_pmd);
1043 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1044 if (pmdp_set_access_flags(vma, haddr, vmf->pmd, entry, 1))
1045 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1046 ret |= VM_FAULT_WRITE;
1047 goto out_unlock;
1048 }
1049 get_page(page);
1050 spin_unlock(vmf->ptl);
1051alloc:
1052 if (transparent_hugepage_enabled(vma) &&
1053 !transparent_hugepage_debug_cow()) {
1054 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1055 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1056 } else
1057 new_page = NULL;
1058
1059 if (likely(new_page)) {
1060 prep_transhuge_page(new_page);
1061 } else {
1062 if (!page) {
1063 split_huge_pmd(vma, vmf->pmd, vmf->address);
1064 ret |= VM_FAULT_FALLBACK;
1065 } else {
1066 ret = do_huge_pmd_wp_page_fallback(vmf, orig_pmd, page);
1067 if (ret & VM_FAULT_OOM) {
1068 split_huge_pmd(vma, vmf->pmd, vmf->address);
1069 ret |= VM_FAULT_FALLBACK;
1070 }
1071 put_page(page);
1072 }
1073 count_vm_event(THP_FAULT_FALLBACK);
1074 goto out;
1075 }
1076
1077 if (unlikely(mem_cgroup_try_charge(new_page, vma->vm_mm,
1078 huge_gfp, &memcg, true))) {
1079 put_page(new_page);
1080 split_huge_pmd(vma, vmf->pmd, vmf->address);
1081 if (page)
1082 put_page(page);
1083 ret |= VM_FAULT_FALLBACK;
1084 count_vm_event(THP_FAULT_FALLBACK);
1085 goto out;
1086 }
1087
1088 count_vm_event(THP_FAULT_ALLOC);
1089
1090 if (!page)
1091 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1092 else
1093 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1094 __SetPageUptodate(new_page);
1095
1096 mmun_start = haddr;
1097 mmun_end = haddr + HPAGE_PMD_SIZE;
1098 mmu_notifier_invalidate_range_start(vma->vm_mm, mmun_start, mmun_end);
1099
1100 spin_lock(vmf->ptl);
1101 if (page)
1102 put_page(page);
1103 if (unlikely(!pmd_same(*vmf->pmd, orig_pmd))) {
1104 spin_unlock(vmf->ptl);
1105 mem_cgroup_cancel_charge(new_page, memcg, true);
1106 put_page(new_page);
1107 goto out_mn;
1108 } else {
1109 pmd_t entry;
1110 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1111 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1112 pmdp_huge_clear_flush_notify(vma, haddr, vmf->pmd);
1113 page_add_new_anon_rmap(new_page, vma, haddr, true);
1114 mem_cgroup_commit_charge(new_page, memcg, false, true);
1115 lru_cache_add_active_or_unevictable(new_page, vma);
1116 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
1117 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1118 if (!page) {
1119 add_mm_counter(vma->vm_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1120 } else {
1121 VM_BUG_ON_PAGE(!PageHead(page), page);
1122 page_remove_rmap(page, true);
1123 put_page(page);
1124 }
1125 ret |= VM_FAULT_WRITE;
1126 }
1127 spin_unlock(vmf->ptl);
1128out_mn:
1129 mmu_notifier_invalidate_range_end(vma->vm_mm, mmun_start, mmun_end);
1130out:
1131 return ret;
1132out_unlock:
1133 spin_unlock(vmf->ptl);
1134 return ret;
1135}
1136
1137/*
1138 * FOLL_FORCE can write to even unwritable pmd's, but only
1139 * after we've gone through a COW cycle and they are dirty.
1140 */
1141static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
1142{
1143 return pmd_write(pmd) ||
1144 ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
1145}
1146
1147struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1148 unsigned long addr,
1149 pmd_t *pmd,
1150 unsigned int flags)
1151{
1152 struct mm_struct *mm = vma->vm_mm;
1153 struct page *page = NULL;
1154
1155 assert_spin_locked(pmd_lockptr(mm, pmd));
1156
1157 if (flags & FOLL_WRITE && !can_follow_write_pmd(*pmd, flags))
1158 goto out;
1159
1160 /* Avoid dumping huge zero page */
1161 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1162 return ERR_PTR(-EFAULT);
1163
1164 /* Full NUMA hinting faults to serialise migration in fault paths */
1165 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1166 goto out;
1167
1168 page = pmd_page(*pmd);
1169 VM_BUG_ON_PAGE(!PageHead(page) && !is_zone_device_page(page), page);
1170 if (flags & FOLL_TOUCH)
1171 touch_pmd(vma, addr, pmd);
1172 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1173 /*
1174 * We don't mlock() pte-mapped THPs. This way we can avoid
1175 * leaking mlocked pages into non-VM_LOCKED VMAs.
1176 *
1177 * For anon THP:
1178 *
1179 * In most cases the pmd is the only mapping of the page as we
1180 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1181 * writable private mappings in populate_vma_page_range().
1182 *
1183 * The only scenario when we have the page shared here is if we
1184 * mlocking read-only mapping shared over fork(). We skip
1185 * mlocking such pages.
1186 *
1187 * For file THP:
1188 *
1189 * We can expect PageDoubleMap() to be stable under page lock:
1190 * for file pages we set it in page_add_file_rmap(), which
1191 * requires page to be locked.
1192 */
1193
1194 if (PageAnon(page) && compound_mapcount(page) != 1)
1195 goto skip_mlock;
1196 if (PageDoubleMap(page) || !page->mapping)
1197 goto skip_mlock;
1198 if (!trylock_page(page))
1199 goto skip_mlock;
1200 lru_add_drain();
1201 if (page->mapping && !PageDoubleMap(page))
1202 mlock_vma_page(page);
1203 unlock_page(page);
1204 }
1205skip_mlock:
1206 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1207 VM_BUG_ON_PAGE(!PageCompound(page) && !is_zone_device_page(page), page);
1208 if (flags & FOLL_GET)
1209 get_page(page);
1210
1211out:
1212 return page;
1213}
1214
1215/* NUMA hinting page fault entry point for trans huge pmds */
1216int do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t pmd)
1217{
1218 struct vm_area_struct *vma = vmf->vma;
1219 struct anon_vma *anon_vma = NULL;
1220 struct page *page;
1221 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
1222 int page_nid = -1, this_nid = numa_node_id();
1223 int target_nid, last_cpupid = -1;
1224 bool page_locked;
1225 bool migrated = false;
1226 bool was_writable;
1227 int flags = 0;
1228
1229 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
1230 if (unlikely(!pmd_same(pmd, *vmf->pmd)))
1231 goto out_unlock;
1232
1233 /*
1234 * If there are potential migrations, wait for completion and retry
1235 * without disrupting NUMA hinting information. Do not relock and
1236 * check_same as the page may no longer be mapped.
1237 */
1238 if (unlikely(pmd_trans_migrating(*vmf->pmd))) {
1239 page = pmd_page(*vmf->pmd);
1240 spin_unlock(vmf->ptl);
1241 wait_on_page_locked(page);
1242 goto out;
1243 }
1244
1245 page = pmd_page(pmd);
1246 BUG_ON(is_huge_zero_page(page));
1247 page_nid = page_to_nid(page);
1248 last_cpupid = page_cpupid_last(page);
1249 count_vm_numa_event(NUMA_HINT_FAULTS);
1250 if (page_nid == this_nid) {
1251 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1252 flags |= TNF_FAULT_LOCAL;
1253 }
1254
1255 /* See similar comment in do_numa_page for explanation */
1256 if (!pmd_write(pmd))
1257 flags |= TNF_NO_GROUP;
1258
1259 /*
1260 * Acquire the page lock to serialise THP migrations but avoid dropping
1261 * page_table_lock if at all possible
1262 */
1263 page_locked = trylock_page(page);
1264 target_nid = mpol_misplaced(page, vma, haddr);
1265 if (target_nid == -1) {
1266 /* If the page was locked, there are no parallel migrations */
1267 if (page_locked)
1268 goto clear_pmdnuma;
1269 }
1270
1271 /* Migration could have started since the pmd_trans_migrating check */
1272 if (!page_locked) {
1273 spin_unlock(vmf->ptl);
1274 wait_on_page_locked(page);
1275 page_nid = -1;
1276 goto out;
1277 }
1278
1279 /*
1280 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1281 * to serialises splits
1282 */
1283 get_page(page);
1284 spin_unlock(vmf->ptl);
1285 anon_vma = page_lock_anon_vma_read(page);
1286
1287 /* Confirm the PMD did not change while page_table_lock was released */
1288 spin_lock(vmf->ptl);
1289 if (unlikely(!pmd_same(pmd, *vmf->pmd))) {
1290 unlock_page(page);
1291 put_page(page);
1292 page_nid = -1;
1293 goto out_unlock;
1294 }
1295
1296 /* Bail if we fail to protect against THP splits for any reason */
1297 if (unlikely(!anon_vma)) {
1298 put_page(page);
1299 page_nid = -1;
1300 goto clear_pmdnuma;
1301 }
1302
1303 /*
1304 * Migrate the THP to the requested node, returns with page unlocked
1305 * and access rights restored.
1306 */
1307 spin_unlock(vmf->ptl);
1308 migrated = migrate_misplaced_transhuge_page(vma->vm_mm, vma,
1309 vmf->pmd, pmd, vmf->address, page, target_nid);
1310 if (migrated) {
1311 flags |= TNF_MIGRATED;
1312 page_nid = target_nid;
1313 } else
1314 flags |= TNF_MIGRATE_FAIL;
1315
1316 goto out;
1317clear_pmdnuma:
1318 BUG_ON(!PageLocked(page));
1319 was_writable = pmd_write(pmd);
1320 pmd = pmd_modify(pmd, vma->vm_page_prot);
1321 pmd = pmd_mkyoung(pmd);
1322 if (was_writable)
1323 pmd = pmd_mkwrite(pmd);
1324 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, pmd);
1325 update_mmu_cache_pmd(vma, vmf->address, vmf->pmd);
1326 unlock_page(page);
1327out_unlock:
1328 spin_unlock(vmf->ptl);
1329
1330out:
1331 if (anon_vma)
1332 page_unlock_anon_vma_read(anon_vma);
1333
1334 if (page_nid != -1)
1335 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR,
1336 vmf->flags);
1337
1338 return 0;
1339}
1340
1341/*
1342 * Return true if we do MADV_FREE successfully on entire pmd page.
1343 * Otherwise, return false.
1344 */
1345bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1346 pmd_t *pmd, unsigned long addr, unsigned long next)
1347{
1348 spinlock_t *ptl;
1349 pmd_t orig_pmd;
1350 struct page *page;
1351 struct mm_struct *mm = tlb->mm;
1352 bool ret = false;
1353
1354 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1355
1356 ptl = pmd_trans_huge_lock(pmd, vma);
1357 if (!ptl)
1358 goto out_unlocked;
1359
1360 orig_pmd = *pmd;
1361 if (is_huge_zero_pmd(orig_pmd))
1362 goto out;
1363
1364 page = pmd_page(orig_pmd);
1365 /*
1366 * If other processes are mapping this page, we couldn't discard
1367 * the page unless they all do MADV_FREE so let's skip the page.
1368 */
1369 if (page_mapcount(page) != 1)
1370 goto out;
1371
1372 if (!trylock_page(page))
1373 goto out;
1374
1375 /*
1376 * If user want to discard part-pages of THP, split it so MADV_FREE
1377 * will deactivate only them.
1378 */
1379 if (next - addr != HPAGE_PMD_SIZE) {
1380 get_page(page);
1381 spin_unlock(ptl);
1382 split_huge_page(page);
1383 put_page(page);
1384 unlock_page(page);
1385 goto out_unlocked;
1386 }
1387
1388 if (PageDirty(page))
1389 ClearPageDirty(page);
1390 unlock_page(page);
1391
1392 if (PageActive(page))
1393 deactivate_page(page);
1394
1395 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1396 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1397 tlb->fullmm);
1398 orig_pmd = pmd_mkold(orig_pmd);
1399 orig_pmd = pmd_mkclean(orig_pmd);
1400
1401 set_pmd_at(mm, addr, pmd, orig_pmd);
1402 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1403 }
1404 ret = true;
1405out:
1406 spin_unlock(ptl);
1407out_unlocked:
1408 return ret;
1409}
1410
1411static inline void zap_deposited_table(struct mm_struct *mm, pmd_t *pmd)
1412{
1413 pgtable_t pgtable;
1414
1415 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1416 pte_free(mm, pgtable);
1417 atomic_long_dec(&mm->nr_ptes);
1418}
1419
1420int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1421 pmd_t *pmd, unsigned long addr)
1422{
1423 pmd_t orig_pmd;
1424 spinlock_t *ptl;
1425
1426 tlb_remove_check_page_size_change(tlb, HPAGE_PMD_SIZE);
1427
1428 ptl = __pmd_trans_huge_lock(pmd, vma);
1429 if (!ptl)
1430 return 0;
1431 /*
1432 * For architectures like ppc64 we look at deposited pgtable
1433 * when calling pmdp_huge_get_and_clear. So do the
1434 * pgtable_trans_huge_withdraw after finishing pmdp related
1435 * operations.
1436 */
1437 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1438 tlb->fullmm);
1439 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1440 if (vma_is_dax(vma)) {
1441 spin_unlock(ptl);
1442 if (is_huge_zero_pmd(orig_pmd))
1443 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1444 } else if (is_huge_zero_pmd(orig_pmd)) {
1445 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1446 atomic_long_dec(&tlb->mm->nr_ptes);
1447 spin_unlock(ptl);
1448 tlb_remove_page_size(tlb, pmd_page(orig_pmd), HPAGE_PMD_SIZE);
1449 } else {
1450 struct page *page = pmd_page(orig_pmd);
1451 page_remove_rmap(page, true);
1452 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1453 VM_BUG_ON_PAGE(!PageHead(page), page);
1454 if (PageAnon(page)) {
1455 pgtable_t pgtable;
1456 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1457 pte_free(tlb->mm, pgtable);
1458 atomic_long_dec(&tlb->mm->nr_ptes);
1459 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1460 } else {
1461 if (arch_needs_pgtable_deposit())
1462 zap_deposited_table(tlb->mm, pmd);
1463 add_mm_counter(tlb->mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1464 }
1465 spin_unlock(ptl);
1466 tlb_remove_page_size(tlb, page, HPAGE_PMD_SIZE);
1467 }
1468 return 1;
1469}
1470
1471#ifndef pmd_move_must_withdraw
1472static inline int pmd_move_must_withdraw(spinlock_t *new_pmd_ptl,
1473 spinlock_t *old_pmd_ptl,
1474 struct vm_area_struct *vma)
1475{
1476 /*
1477 * With split pmd lock we also need to move preallocated
1478 * PTE page table if new_pmd is on different PMD page table.
1479 *
1480 * We also don't deposit and withdraw tables for file pages.
1481 */
1482 return (new_pmd_ptl != old_pmd_ptl) && vma_is_anonymous(vma);
1483}
1484#endif
1485
1486bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
1487 unsigned long new_addr, unsigned long old_end,
1488 pmd_t *old_pmd, pmd_t *new_pmd, bool *need_flush)
1489{
1490 spinlock_t *old_ptl, *new_ptl;
1491 pmd_t pmd;
1492 struct mm_struct *mm = vma->vm_mm;
1493 bool force_flush = false;
1494
1495 if ((old_addr & ~HPAGE_PMD_MASK) ||
1496 (new_addr & ~HPAGE_PMD_MASK) ||
1497 old_end - old_addr < HPAGE_PMD_SIZE)
1498 return false;
1499
1500 /*
1501 * The destination pmd shouldn't be established, free_pgtables()
1502 * should have release it.
1503 */
1504 if (WARN_ON(!pmd_none(*new_pmd))) {
1505 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1506 return false;
1507 }
1508
1509 /*
1510 * We don't have to worry about the ordering of src and dst
1511 * ptlocks because exclusive mmap_sem prevents deadlock.
1512 */
1513 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1514 if (old_ptl) {
1515 new_ptl = pmd_lockptr(mm, new_pmd);
1516 if (new_ptl != old_ptl)
1517 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1518 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1519 if (pmd_present(pmd) && pmd_dirty(pmd))
1520 force_flush = true;
1521 VM_BUG_ON(!pmd_none(*new_pmd));
1522
1523 if (pmd_move_must_withdraw(new_ptl, old_ptl, vma)) {
1524 pgtable_t pgtable;
1525 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1526 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1527 }
1528 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1529 if (new_ptl != old_ptl)
1530 spin_unlock(new_ptl);
1531 if (force_flush)
1532 flush_tlb_range(vma, old_addr, old_addr + PMD_SIZE);
1533 else
1534 *need_flush = true;
1535 spin_unlock(old_ptl);
1536 return true;
1537 }
1538 return false;
1539}
1540
1541/*
1542 * Returns
1543 * - 0 if PMD could not be locked
1544 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1545 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1546 */
1547int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1548 unsigned long addr, pgprot_t newprot, int prot_numa)
1549{
1550 struct mm_struct *mm = vma->vm_mm;
1551 spinlock_t *ptl;
1552 int ret = 0;
1553
1554 ptl = __pmd_trans_huge_lock(pmd, vma);
1555 if (ptl) {
1556 pmd_t entry;
1557 bool preserve_write = prot_numa && pmd_write(*pmd);
1558 ret = 1;
1559
1560 /*
1561 * Avoid trapping faults against the zero page. The read-only
1562 * data is likely to be read-cached on the local CPU and
1563 * local/remote hits to the zero page are not interesting.
1564 */
1565 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1566 spin_unlock(ptl);
1567 return ret;
1568 }
1569
1570 if (!prot_numa || !pmd_protnone(*pmd)) {
1571 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1572 entry = pmd_modify(entry, newprot);
1573 if (preserve_write)
1574 entry = pmd_mkwrite(entry);
1575 ret = HPAGE_PMD_NR;
1576 set_pmd_at(mm, addr, pmd, entry);
1577 BUG_ON(vma_is_anonymous(vma) && !preserve_write &&
1578 pmd_write(entry));
1579 }
1580 spin_unlock(ptl);
1581 }
1582
1583 return ret;
1584}
1585
1586/*
1587 * Returns page table lock pointer if a given pmd maps a thp, NULL otherwise.
1588 *
1589 * Note that if it returns page table lock pointer, this routine returns without
1590 * unlocking page table lock. So callers must unlock it.
1591 */
1592spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1593{
1594 spinlock_t *ptl;
1595 ptl = pmd_lock(vma->vm_mm, pmd);
1596 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1597 return ptl;
1598 spin_unlock(ptl);
1599 return NULL;
1600}
1601
1602static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
1603 unsigned long haddr, pmd_t *pmd)
1604{
1605 struct mm_struct *mm = vma->vm_mm;
1606 pgtable_t pgtable;
1607 pmd_t _pmd;
1608 int i;
1609
1610 /* leave pmd empty until pte is filled */
1611 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1612
1613 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1614 pmd_populate(mm, &_pmd, pgtable);
1615
1616 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1617 pte_t *pte, entry;
1618 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
1619 entry = pte_mkspecial(entry);
1620 pte = pte_offset_map(&_pmd, haddr);
1621 VM_BUG_ON(!pte_none(*pte));
1622 set_pte_at(mm, haddr, pte, entry);
1623 pte_unmap(pte);
1624 }
1625 smp_wmb(); /* make pte visible before pmd */
1626 pmd_populate(mm, pmd, pgtable);
1627}
1628
1629static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
1630 unsigned long haddr, bool freeze)
1631{
1632 struct mm_struct *mm = vma->vm_mm;
1633 struct page *page;
1634 pgtable_t pgtable;
1635 pmd_t _pmd;
1636 bool young, write, dirty, soft_dirty;
1637 unsigned long addr;
1638 int i;
1639
1640 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
1641 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
1642 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
1643 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
1644
1645 count_vm_event(THP_SPLIT_PMD);
1646
1647 if (!vma_is_anonymous(vma)) {
1648 _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1649 /*
1650 * We are going to unmap this huge page. So
1651 * just go ahead and zap it
1652 */
1653 if (arch_needs_pgtable_deposit())
1654 zap_deposited_table(mm, pmd);
1655 if (vma_is_dax(vma))
1656 return;
1657 page = pmd_page(_pmd);
1658 if (!PageReferenced(page) && pmd_young(_pmd))
1659 SetPageReferenced(page);
1660 page_remove_rmap(page, true);
1661 put_page(page);
1662 add_mm_counter(mm, MM_FILEPAGES, -HPAGE_PMD_NR);
1663 return;
1664 } else if (is_huge_zero_pmd(*pmd)) {
1665 return __split_huge_zero_page_pmd(vma, haddr, pmd);
1666 }
1667
1668 page = pmd_page(*pmd);
1669 VM_BUG_ON_PAGE(!page_count(page), page);
1670 page_ref_add(page, HPAGE_PMD_NR - 1);
1671 write = pmd_write(*pmd);
1672 young = pmd_young(*pmd);
1673 dirty = pmd_dirty(*pmd);
1674 soft_dirty = pmd_soft_dirty(*pmd);
1675
1676 pmdp_huge_split_prepare(vma, haddr, pmd);
1677 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1678 pmd_populate(mm, &_pmd, pgtable);
1679
1680 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
1681 pte_t entry, *pte;
1682 /*
1683 * Note that NUMA hinting access restrictions are not
1684 * transferred to avoid any possibility of altering
1685 * permissions across VMAs.
1686 */
1687 if (freeze) {
1688 swp_entry_t swp_entry;
1689 swp_entry = make_migration_entry(page + i, write);
1690 entry = swp_entry_to_pte(swp_entry);
1691 if (soft_dirty)
1692 entry = pte_swp_mksoft_dirty(entry);
1693 } else {
1694 entry = mk_pte(page + i, READ_ONCE(vma->vm_page_prot));
1695 entry = maybe_mkwrite(entry, vma);
1696 if (!write)
1697 entry = pte_wrprotect(entry);
1698 if (!young)
1699 entry = pte_mkold(entry);
1700 if (soft_dirty)
1701 entry = pte_mksoft_dirty(entry);
1702 }
1703 if (dirty)
1704 SetPageDirty(page + i);
1705 pte = pte_offset_map(&_pmd, addr);
1706 BUG_ON(!pte_none(*pte));
1707 set_pte_at(mm, addr, pte, entry);
1708 atomic_inc(&page[i]._mapcount);
1709 pte_unmap(pte);
1710 }
1711
1712 /*
1713 * Set PG_double_map before dropping compound_mapcount to avoid
1714 * false-negative page_mapped().
1715 */
1716 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
1717 for (i = 0; i < HPAGE_PMD_NR; i++)
1718 atomic_inc(&page[i]._mapcount);
1719 }
1720
1721 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
1722 /* Last compound_mapcount is gone. */
1723 __dec_node_page_state(page, NR_ANON_THPS);
1724 if (TestClearPageDoubleMap(page)) {
1725 /* No need in mapcount reference anymore */
1726 for (i = 0; i < HPAGE_PMD_NR; i++)
1727 atomic_dec(&page[i]._mapcount);
1728 }
1729 }
1730
1731 smp_wmb(); /* make pte visible before pmd */
1732 /*
1733 * Up to this point the pmd is present and huge and userland has the
1734 * whole access to the hugepage during the split (which happens in
1735 * place). If we overwrite the pmd with the not-huge version pointing
1736 * to the pte here (which of course we could if all CPUs were bug
1737 * free), userland could trigger a small page size TLB miss on the
1738 * small sized TLB while the hugepage TLB entry is still established in
1739 * the huge TLB. Some CPU doesn't like that.
1740 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
1741 * 383 on page 93. Intel should be safe but is also warns that it's
1742 * only safe if the permission and cache attributes of the two entries
1743 * loaded in the two TLB is identical (which should be the case here).
1744 * But it is generally safer to never allow small and huge TLB entries
1745 * for the same virtual address to be loaded simultaneously. So instead
1746 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
1747 * current pmd notpresent (atomically because here the pmd_trans_huge
1748 * and pmd_trans_splitting must remain set at all times on the pmd
1749 * until the split is complete for this pmd), then we flush the SMP TLB
1750 * and finally we write the non-huge version of the pmd entry with
1751 * pmd_populate.
1752 */
1753 pmdp_invalidate(vma, haddr, pmd);
1754 pmd_populate(mm, pmd, pgtable);
1755
1756 if (freeze) {
1757 for (i = 0; i < HPAGE_PMD_NR; i++) {
1758 page_remove_rmap(page + i, false);
1759 put_page(page + i);
1760 }
1761 }
1762}
1763
1764void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1765 unsigned long address, bool freeze, struct page *page)
1766{
1767 spinlock_t *ptl;
1768 struct mm_struct *mm = vma->vm_mm;
1769 unsigned long haddr = address & HPAGE_PMD_MASK;
1770
1771 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
1772 ptl = pmd_lock(mm, pmd);
1773
1774 /*
1775 * If caller asks to setup a migration entries, we need a page to check
1776 * pmd against. Otherwise we can end up replacing wrong page.
1777 */
1778 VM_BUG_ON(freeze && !page);
1779 if (page && page != pmd_page(*pmd))
1780 goto out;
1781
1782 if (pmd_trans_huge(*pmd)) {
1783 page = pmd_page(*pmd);
1784 if (PageMlocked(page))
1785 clear_page_mlock(page);
1786 } else if (!pmd_devmap(*pmd))
1787 goto out;
1788 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
1789out:
1790 spin_unlock(ptl);
1791 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
1792}
1793
1794void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
1795 bool freeze, struct page *page)
1796{
1797 pgd_t *pgd;
1798 pud_t *pud;
1799 pmd_t *pmd;
1800
1801 pgd = pgd_offset(vma->vm_mm, address);
1802 if (!pgd_present(*pgd))
1803 return;
1804
1805 pud = pud_offset(pgd, address);
1806 if (!pud_present(*pud))
1807 return;
1808
1809 pmd = pmd_offset(pud, address);
1810
1811 __split_huge_pmd(vma, pmd, address, freeze, page);
1812}
1813
1814void vma_adjust_trans_huge(struct vm_area_struct *vma,
1815 unsigned long start,
1816 unsigned long end,
1817 long adjust_next)
1818{
1819 /*
1820 * If the new start address isn't hpage aligned and it could
1821 * previously contain an hugepage: check if we need to split
1822 * an huge pmd.
1823 */
1824 if (start & ~HPAGE_PMD_MASK &&
1825 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
1826 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1827 split_huge_pmd_address(vma, start, false, NULL);
1828
1829 /*
1830 * If the new end address isn't hpage aligned and it could
1831 * previously contain an hugepage: check if we need to split
1832 * an huge pmd.
1833 */
1834 if (end & ~HPAGE_PMD_MASK &&
1835 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
1836 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
1837 split_huge_pmd_address(vma, end, false, NULL);
1838
1839 /*
1840 * If we're also updating the vma->vm_next->vm_start, if the new
1841 * vm_next->vm_start isn't page aligned and it could previously
1842 * contain an hugepage: check if we need to split an huge pmd.
1843 */
1844 if (adjust_next > 0) {
1845 struct vm_area_struct *next = vma->vm_next;
1846 unsigned long nstart = next->vm_start;
1847 nstart += adjust_next << PAGE_SHIFT;
1848 if (nstart & ~HPAGE_PMD_MASK &&
1849 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
1850 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
1851 split_huge_pmd_address(next, nstart, false, NULL);
1852 }
1853}
1854
1855static void freeze_page(struct page *page)
1856{
1857 enum ttu_flags ttu_flags = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS |
1858 TTU_RMAP_LOCKED;
1859 int i, ret;
1860
1861 VM_BUG_ON_PAGE(!PageHead(page), page);
1862
1863 if (PageAnon(page))
1864 ttu_flags |= TTU_MIGRATION;
1865
1866 /* We only need TTU_SPLIT_HUGE_PMD once */
1867 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
1868 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
1869 /* Cut short if the page is unmapped */
1870 if (page_count(page) == 1)
1871 return;
1872
1873 ret = try_to_unmap(page + i, ttu_flags);
1874 }
1875 VM_BUG_ON_PAGE(ret, page + i - 1);
1876}
1877
1878static void unfreeze_page(struct page *page)
1879{
1880 int i;
1881
1882 for (i = 0; i < HPAGE_PMD_NR; i++)
1883 remove_migration_ptes(page + i, page + i, true);
1884}
1885
1886static void __split_huge_page_tail(struct page *head, int tail,
1887 struct lruvec *lruvec, struct list_head *list)
1888{
1889 struct page *page_tail = head + tail;
1890
1891 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
1892 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
1893
1894 /*
1895 * tail_page->_refcount is zero and not changing from under us. But
1896 * get_page_unless_zero() may be running from under us on the
1897 * tail_page. If we used atomic_set() below instead of atomic_inc() or
1898 * atomic_add(), we would then run atomic_set() concurrently with
1899 * get_page_unless_zero(), and atomic_set() is implemented in C not
1900 * using locked ops. spin_unlock on x86 sometime uses locked ops
1901 * because of PPro errata 66, 92, so unless somebody can guarantee
1902 * atomic_set() here would be safe on all archs (and not only on x86),
1903 * it's safer to use atomic_inc()/atomic_add().
1904 */
1905 if (PageAnon(head)) {
1906 page_ref_inc(page_tail);
1907 } else {
1908 /* Additional pin to radix tree */
1909 page_ref_add(page_tail, 2);
1910 }
1911
1912 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1913 page_tail->flags |= (head->flags &
1914 ((1L << PG_referenced) |
1915 (1L << PG_swapbacked) |
1916 (1L << PG_mlocked) |
1917 (1L << PG_uptodate) |
1918 (1L << PG_active) |
1919 (1L << PG_locked) |
1920 (1L << PG_unevictable) |
1921 (1L << PG_dirty)));
1922
1923 /*
1924 * After clearing PageTail the gup refcount can be released.
1925 * Page flags also must be visible before we make the page non-compound.
1926 */
1927 smp_wmb();
1928
1929 clear_compound_head(page_tail);
1930
1931 if (page_is_young(head))
1932 set_page_young(page_tail);
1933 if (page_is_idle(head))
1934 set_page_idle(page_tail);
1935
1936 /* ->mapping in first tail page is compound_mapcount */
1937 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
1938 page_tail);
1939 page_tail->mapping = head->mapping;
1940
1941 page_tail->index = head->index + tail;
1942 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
1943 lru_add_page_tail(head, page_tail, lruvec, list);
1944}
1945
1946static void __split_huge_page(struct page *page, struct list_head *list,
1947 unsigned long flags)
1948{
1949 struct page *head = compound_head(page);
1950 struct zone *zone = page_zone(head);
1951 struct lruvec *lruvec;
1952 pgoff_t end = -1;
1953 int i;
1954
1955 lruvec = mem_cgroup_page_lruvec(head, zone->zone_pgdat);
1956
1957 /* complete memcg works before add pages to LRU */
1958 mem_cgroup_split_huge_fixup(head);
1959
1960 if (!PageAnon(page))
1961 end = DIV_ROUND_UP(i_size_read(head->mapping->host), PAGE_SIZE);
1962
1963 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1964 __split_huge_page_tail(head, i, lruvec, list);
1965 /* Some pages can be beyond i_size: drop them from page cache */
1966 if (head[i].index >= end) {
1967 __ClearPageDirty(head + i);
1968 __delete_from_page_cache(head + i, NULL);
1969 if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
1970 shmem_uncharge(head->mapping->host, 1);
1971 put_page(head + i);
1972 }
1973 }
1974
1975 ClearPageCompound(head);
1976 /* See comment in __split_huge_page_tail() */
1977 if (PageAnon(head)) {
1978 page_ref_inc(head);
1979 } else {
1980 /* Additional pin to radix tree */
1981 page_ref_add(head, 2);
1982 spin_unlock(&head->mapping->tree_lock);
1983 }
1984
1985 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
1986
1987 unfreeze_page(head);
1988
1989 for (i = 0; i < HPAGE_PMD_NR; i++) {
1990 struct page *subpage = head + i;
1991 if (subpage == page)
1992 continue;
1993 unlock_page(subpage);
1994
1995 /*
1996 * Subpages may be freed if there wasn't any mapping
1997 * like if add_to_swap() is running on a lru page that
1998 * had its mapping zapped. And freeing these pages
1999 * requires taking the lru_lock so we do the put_page
2000 * of the tail pages after the split is complete.
2001 */
2002 put_page(subpage);
2003 }
2004}
2005
2006int total_mapcount(struct page *page)
2007{
2008 int i, compound, ret;
2009
2010 VM_BUG_ON_PAGE(PageTail(page), page);
2011
2012 if (likely(!PageCompound(page)))
2013 return atomic_read(&page->_mapcount) + 1;
2014
2015 compound = compound_mapcount(page);
2016 if (PageHuge(page))
2017 return compound;
2018 ret = compound;
2019 for (i = 0; i < HPAGE_PMD_NR; i++)
2020 ret += atomic_read(&page[i]._mapcount) + 1;
2021 /* File pages has compound_mapcount included in _mapcount */
2022 if (!PageAnon(page))
2023 return ret - compound * HPAGE_PMD_NR;
2024 if (PageDoubleMap(page))
2025 ret -= HPAGE_PMD_NR;
2026 return ret;
2027}
2028
2029/*
2030 * This calculates accurately how many mappings a transparent hugepage
2031 * has (unlike page_mapcount() which isn't fully accurate). This full
2032 * accuracy is primarily needed to know if copy-on-write faults can
2033 * reuse the page and change the mapping to read-write instead of
2034 * copying them. At the same time this returns the total_mapcount too.
2035 *
2036 * The function returns the highest mapcount any one of the subpages
2037 * has. If the return value is one, even if different processes are
2038 * mapping different subpages of the transparent hugepage, they can
2039 * all reuse it, because each process is reusing a different subpage.
2040 *
2041 * The total_mapcount is instead counting all virtual mappings of the
2042 * subpages. If the total_mapcount is equal to "one", it tells the
2043 * caller all mappings belong to the same "mm" and in turn the
2044 * anon_vma of the transparent hugepage can become the vma->anon_vma
2045 * local one as no other process may be mapping any of the subpages.
2046 *
2047 * It would be more accurate to replace page_mapcount() with
2048 * page_trans_huge_mapcount(), however we only use
2049 * page_trans_huge_mapcount() in the copy-on-write faults where we
2050 * need full accuracy to avoid breaking page pinning, because
2051 * page_trans_huge_mapcount() is slower than page_mapcount().
2052 */
2053int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
2054{
2055 int i, ret, _total_mapcount, mapcount;
2056
2057 /* hugetlbfs shouldn't call it */
2058 VM_BUG_ON_PAGE(PageHuge(page), page);
2059
2060 if (likely(!PageTransCompound(page))) {
2061 mapcount = atomic_read(&page->_mapcount) + 1;
2062 if (total_mapcount)
2063 *total_mapcount = mapcount;
2064 return mapcount;
2065 }
2066
2067 page = compound_head(page);
2068
2069 _total_mapcount = ret = 0;
2070 for (i = 0; i < HPAGE_PMD_NR; i++) {
2071 mapcount = atomic_read(&page[i]._mapcount) + 1;
2072 ret = max(ret, mapcount);
2073 _total_mapcount += mapcount;
2074 }
2075 if (PageDoubleMap(page)) {
2076 ret -= 1;
2077 _total_mapcount -= HPAGE_PMD_NR;
2078 }
2079 mapcount = compound_mapcount(page);
2080 ret += mapcount;
2081 _total_mapcount += mapcount;
2082 if (total_mapcount)
2083 *total_mapcount = _total_mapcount;
2084 return ret;
2085}
2086
2087/*
2088 * This function splits huge page into normal pages. @page can point to any
2089 * subpage of huge page to split. Split doesn't change the position of @page.
2090 *
2091 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
2092 * The huge page must be locked.
2093 *
2094 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
2095 *
2096 * Both head page and tail pages will inherit mapping, flags, and so on from
2097 * the hugepage.
2098 *
2099 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
2100 * they are not mapped.
2101 *
2102 * Returns 0 if the hugepage is split successfully.
2103 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
2104 * us.
2105 */
2106int split_huge_page_to_list(struct page *page, struct list_head *list)
2107{
2108 struct page *head = compound_head(page);
2109 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
2110 struct anon_vma *anon_vma = NULL;
2111 struct address_space *mapping = NULL;
2112 int count, mapcount, extra_pins, ret;
2113 bool mlocked;
2114 unsigned long flags;
2115
2116 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
2117 VM_BUG_ON_PAGE(!PageLocked(page), page);
2118 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2119 VM_BUG_ON_PAGE(!PageCompound(page), page);
2120
2121 if (PageAnon(head)) {
2122 /*
2123 * The caller does not necessarily hold an mmap_sem that would
2124 * prevent the anon_vma disappearing so we first we take a
2125 * reference to it and then lock the anon_vma for write. This
2126 * is similar to page_lock_anon_vma_read except the write lock
2127 * is taken to serialise against parallel split or collapse
2128 * operations.
2129 */
2130 anon_vma = page_get_anon_vma(head);
2131 if (!anon_vma) {
2132 ret = -EBUSY;
2133 goto out;
2134 }
2135 extra_pins = 0;
2136 mapping = NULL;
2137 anon_vma_lock_write(anon_vma);
2138 } else {
2139 mapping = head->mapping;
2140
2141 /* Truncated ? */
2142 if (!mapping) {
2143 ret = -EBUSY;
2144 goto out;
2145 }
2146
2147 /* Addidional pins from radix tree */
2148 extra_pins = HPAGE_PMD_NR;
2149 anon_vma = NULL;
2150 i_mmap_lock_read(mapping);
2151 }
2152
2153 /*
2154 * Racy check if we can split the page, before freeze_page() will
2155 * split PMDs
2156 */
2157 if (total_mapcount(head) != page_count(head) - extra_pins - 1) {
2158 ret = -EBUSY;
2159 goto out_unlock;
2160 }
2161
2162 mlocked = PageMlocked(page);
2163 freeze_page(head);
2164 VM_BUG_ON_PAGE(compound_mapcount(head), head);
2165
2166 /* Make sure the page is not on per-CPU pagevec as it takes pin */
2167 if (mlocked)
2168 lru_add_drain();
2169
2170 /* prevent PageLRU to go away from under us, and freeze lru stats */
2171 spin_lock_irqsave(zone_lru_lock(page_zone(head)), flags);
2172
2173 if (mapping) {
2174 void **pslot;
2175
2176 spin_lock(&mapping->tree_lock);
2177 pslot = radix_tree_lookup_slot(&mapping->page_tree,
2178 page_index(head));
2179 /*
2180 * Check if the head page is present in radix tree.
2181 * We assume all tail are present too, if head is there.
2182 */
2183 if (radix_tree_deref_slot_protected(pslot,
2184 &mapping->tree_lock) != head)
2185 goto fail;
2186 }
2187
2188 /* Prevent deferred_split_scan() touching ->_refcount */
2189 spin_lock(&pgdata->split_queue_lock);
2190 count = page_count(head);
2191 mapcount = total_mapcount(head);
2192 if (!mapcount && page_ref_freeze(head, 1 + extra_pins)) {
2193 if (!list_empty(page_deferred_list(head))) {
2194 pgdata->split_queue_len--;
2195 list_del(page_deferred_list(head));
2196 }
2197 if (mapping)
2198 __dec_node_page_state(page, NR_SHMEM_THPS);
2199 spin_unlock(&pgdata->split_queue_lock);
2200 __split_huge_page(page, list, flags);
2201 ret = 0;
2202 } else {
2203 if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
2204 pr_alert("total_mapcount: %u, page_count(): %u\n",
2205 mapcount, count);
2206 if (PageTail(page))
2207 dump_page(head, NULL);
2208 dump_page(page, "total_mapcount(head) > 0");
2209 BUG();
2210 }
2211 spin_unlock(&pgdata->split_queue_lock);
2212fail: if (mapping)
2213 spin_unlock(&mapping->tree_lock);
2214 spin_unlock_irqrestore(zone_lru_lock(page_zone(head)), flags);
2215 unfreeze_page(head);
2216 ret = -EBUSY;
2217 }
2218
2219out_unlock:
2220 if (anon_vma) {
2221 anon_vma_unlock_write(anon_vma);
2222 put_anon_vma(anon_vma);
2223 }
2224 if (mapping)
2225 i_mmap_unlock_read(mapping);
2226out:
2227 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
2228 return ret;
2229}
2230
2231void free_transhuge_page(struct page *page)
2232{
2233 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2234 unsigned long flags;
2235
2236 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2237 if (!list_empty(page_deferred_list(page))) {
2238 pgdata->split_queue_len--;
2239 list_del(page_deferred_list(page));
2240 }
2241 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2242 free_compound_page(page);
2243}
2244
2245void deferred_split_huge_page(struct page *page)
2246{
2247 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
2248 unsigned long flags;
2249
2250 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
2251
2252 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2253 if (list_empty(page_deferred_list(page))) {
2254 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
2255 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
2256 pgdata->split_queue_len++;
2257 }
2258 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2259}
2260
2261static unsigned long deferred_split_count(struct shrinker *shrink,
2262 struct shrink_control *sc)
2263{
2264 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2265 return ACCESS_ONCE(pgdata->split_queue_len);
2266}
2267
2268static unsigned long deferred_split_scan(struct shrinker *shrink,
2269 struct shrink_control *sc)
2270{
2271 struct pglist_data *pgdata = NODE_DATA(sc->nid);
2272 unsigned long flags;
2273 LIST_HEAD(list), *pos, *next;
2274 struct page *page;
2275 int split = 0;
2276
2277 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2278 /* Take pin on all head pages to avoid freeing them under us */
2279 list_for_each_safe(pos, next, &pgdata->split_queue) {
2280 page = list_entry((void *)pos, struct page, mapping);
2281 page = compound_head(page);
2282 if (get_page_unless_zero(page)) {
2283 list_move(page_deferred_list(page), &list);
2284 } else {
2285 /* We lost race with put_compound_page() */
2286 list_del_init(page_deferred_list(page));
2287 pgdata->split_queue_len--;
2288 }
2289 if (!--sc->nr_to_scan)
2290 break;
2291 }
2292 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2293
2294 list_for_each_safe(pos, next, &list) {
2295 page = list_entry((void *)pos, struct page, mapping);
2296 lock_page(page);
2297 /* split_huge_page() removes page from list on success */
2298 if (!split_huge_page(page))
2299 split++;
2300 unlock_page(page);
2301 put_page(page);
2302 }
2303
2304 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
2305 list_splice_tail(&list, &pgdata->split_queue);
2306 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
2307
2308 /*
2309 * Stop shrinker if we didn't split any page, but the queue is empty.
2310 * This can happen if pages were freed under us.
2311 */
2312 if (!split && list_empty(&pgdata->split_queue))
2313 return SHRINK_STOP;
2314 return split;
2315}
2316
2317static struct shrinker deferred_split_shrinker = {
2318 .count_objects = deferred_split_count,
2319 .scan_objects = deferred_split_scan,
2320 .seeks = DEFAULT_SEEKS,
2321 .flags = SHRINKER_NUMA_AWARE,
2322};
2323
2324#ifdef CONFIG_DEBUG_FS
2325static int split_huge_pages_set(void *data, u64 val)
2326{
2327 struct zone *zone;
2328 struct page *page;
2329 unsigned long pfn, max_zone_pfn;
2330 unsigned long total = 0, split = 0;
2331
2332 if (val != 1)
2333 return -EINVAL;
2334
2335 for_each_populated_zone(zone) {
2336 max_zone_pfn = zone_end_pfn(zone);
2337 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
2338 if (!pfn_valid(pfn))
2339 continue;
2340
2341 page = pfn_to_page(pfn);
2342 if (!get_page_unless_zero(page))
2343 continue;
2344
2345 if (zone != page_zone(page))
2346 goto next;
2347
2348 if (!PageHead(page) || PageHuge(page) || !PageLRU(page))
2349 goto next;
2350
2351 total++;
2352 lock_page(page);
2353 if (!split_huge_page(page))
2354 split++;
2355 unlock_page(page);
2356next:
2357 put_page(page);
2358 }
2359 }
2360
2361 pr_info("%lu of %lu THP split\n", split, total);
2362
2363 return 0;
2364}
2365DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
2366 "%llu\n");
2367
2368static int __init split_huge_pages_debugfs(void)
2369{
2370 void *ret;
2371
2372 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
2373 &split_huge_pages_fops);
2374 if (!ret)
2375 pr_warn("Failed to create split_huge_pages in debugfs");
2376 return 0;
2377}
2378late_initcall(split_huge_pages_debugfs);
2379#endif