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/kthread.h>
22#include <linux/khugepaged.h>
23#include <linux/freezer.h>
24#include <linux/pfn_t.h>
25#include <linux/mman.h>
26#include <linux/memremap.h>
27#include <linux/pagemap.h>
28#include <linux/debugfs.h>
29#include <linux/migrate.h>
30#include <linux/hashtable.h>
31#include <linux/userfaultfd_k.h>
32#include <linux/page_idle.h>
33
34#include <asm/tlb.h>
35#include <asm/pgalloc.h>
36#include "internal.h"
37
38enum scan_result {
39 SCAN_FAIL,
40 SCAN_SUCCEED,
41 SCAN_PMD_NULL,
42 SCAN_EXCEED_NONE_PTE,
43 SCAN_PTE_NON_PRESENT,
44 SCAN_PAGE_RO,
45 SCAN_NO_REFERENCED_PAGE,
46 SCAN_PAGE_NULL,
47 SCAN_SCAN_ABORT,
48 SCAN_PAGE_COUNT,
49 SCAN_PAGE_LRU,
50 SCAN_PAGE_LOCK,
51 SCAN_PAGE_ANON,
52 SCAN_PAGE_COMPOUND,
53 SCAN_ANY_PROCESS,
54 SCAN_VMA_NULL,
55 SCAN_VMA_CHECK,
56 SCAN_ADDRESS_RANGE,
57 SCAN_SWAP_CACHE_PAGE,
58 SCAN_DEL_PAGE_LRU,
59 SCAN_ALLOC_HUGE_PAGE_FAIL,
60 SCAN_CGROUP_CHARGE_FAIL
61};
62
63#define CREATE_TRACE_POINTS
64#include <trace/events/huge_memory.h>
65
66/*
67 * By default transparent hugepage support is disabled in order that avoid
68 * to risk increase the memory footprint of applications without a guaranteed
69 * benefit. When transparent hugepage support is enabled, is for all mappings,
70 * and khugepaged scans all mappings.
71 * Defrag is invoked by khugepaged hugepage allocations and by page faults
72 * for all hugepage allocations.
73 */
74unsigned long transparent_hugepage_flags __read_mostly =
75#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
76 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
77#endif
78#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
79 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
80#endif
81 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
82 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
83 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
84
85/* default scan 8*512 pte (or vmas) every 30 second */
86static unsigned int khugepaged_pages_to_scan __read_mostly;
87static unsigned int khugepaged_pages_collapsed;
88static unsigned int khugepaged_full_scans;
89static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
90/* during fragmentation poll the hugepage allocator once every minute */
91static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
92static struct task_struct *khugepaged_thread __read_mostly;
93static DEFINE_MUTEX(khugepaged_mutex);
94static DEFINE_SPINLOCK(khugepaged_mm_lock);
95static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
96/*
97 * default collapse hugepages if there is at least one pte mapped like
98 * it would have happened if the vma was large enough during page
99 * fault.
100 */
101static unsigned int khugepaged_max_ptes_none __read_mostly;
102
103static int khugepaged(void *none);
104static int khugepaged_slab_init(void);
105static void khugepaged_slab_exit(void);
106
107#define MM_SLOTS_HASH_BITS 10
108static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
109
110static struct kmem_cache *mm_slot_cache __read_mostly;
111
112/**
113 * struct mm_slot - hash lookup from mm to mm_slot
114 * @hash: hash collision list
115 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
116 * @mm: the mm that this information is valid for
117 */
118struct mm_slot {
119 struct hlist_node hash;
120 struct list_head mm_node;
121 struct mm_struct *mm;
122};
123
124/**
125 * struct khugepaged_scan - cursor for scanning
126 * @mm_head: the head of the mm list to scan
127 * @mm_slot: the current mm_slot we are scanning
128 * @address: the next address inside that to be scanned
129 *
130 * There is only the one khugepaged_scan instance of this cursor structure.
131 */
132struct khugepaged_scan {
133 struct list_head mm_head;
134 struct mm_slot *mm_slot;
135 unsigned long address;
136};
137static struct khugepaged_scan khugepaged_scan = {
138 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
139};
140
141static struct shrinker deferred_split_shrinker;
142
143static void set_recommended_min_free_kbytes(void)
144{
145 struct zone *zone;
146 int nr_zones = 0;
147 unsigned long recommended_min;
148
149 for_each_populated_zone(zone)
150 nr_zones++;
151
152 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
153 recommended_min = pageblock_nr_pages * nr_zones * 2;
154
155 /*
156 * Make sure that on average at least two pageblocks are almost free
157 * of another type, one for a migratetype to fall back to and a
158 * second to avoid subsequent fallbacks of other types There are 3
159 * MIGRATE_TYPES we care about.
160 */
161 recommended_min += pageblock_nr_pages * nr_zones *
162 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
163
164 /* don't ever allow to reserve more than 5% of the lowmem */
165 recommended_min = min(recommended_min,
166 (unsigned long) nr_free_buffer_pages() / 20);
167 recommended_min <<= (PAGE_SHIFT-10);
168
169 if (recommended_min > min_free_kbytes) {
170 if (user_min_free_kbytes >= 0)
171 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
172 min_free_kbytes, recommended_min);
173
174 min_free_kbytes = recommended_min;
175 }
176 setup_per_zone_wmarks();
177}
178
179static int start_stop_khugepaged(void)
180{
181 int err = 0;
182 if (khugepaged_enabled()) {
183 if (!khugepaged_thread)
184 khugepaged_thread = kthread_run(khugepaged, NULL,
185 "khugepaged");
186 if (IS_ERR(khugepaged_thread)) {
187 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
188 err = PTR_ERR(khugepaged_thread);
189 khugepaged_thread = NULL;
190 goto fail;
191 }
192
193 if (!list_empty(&khugepaged_scan.mm_head))
194 wake_up_interruptible(&khugepaged_wait);
195
196 set_recommended_min_free_kbytes();
197 } else if (khugepaged_thread) {
198 kthread_stop(khugepaged_thread);
199 khugepaged_thread = NULL;
200 }
201fail:
202 return err;
203}
204
205static atomic_t huge_zero_refcount;
206struct page *huge_zero_page __read_mostly;
207
208struct page *get_huge_zero_page(void)
209{
210 struct page *zero_page;
211retry:
212 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
213 return READ_ONCE(huge_zero_page);
214
215 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
216 HPAGE_PMD_ORDER);
217 if (!zero_page) {
218 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
219 return NULL;
220 }
221 count_vm_event(THP_ZERO_PAGE_ALLOC);
222 preempt_disable();
223 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
224 preempt_enable();
225 __free_pages(zero_page, compound_order(zero_page));
226 goto retry;
227 }
228
229 /* We take additional reference here. It will be put back by shrinker */
230 atomic_set(&huge_zero_refcount, 2);
231 preempt_enable();
232 return READ_ONCE(huge_zero_page);
233}
234
235void put_huge_zero_page(void)
236{
237 /*
238 * Counter should never go to zero here. Only shrinker can put
239 * last reference.
240 */
241 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
242}
243
244static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
245 struct shrink_control *sc)
246{
247 /* we can free zero page only if last reference remains */
248 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
249}
250
251static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
252 struct shrink_control *sc)
253{
254 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
255 struct page *zero_page = xchg(&huge_zero_page, NULL);
256 BUG_ON(zero_page == NULL);
257 __free_pages(zero_page, compound_order(zero_page));
258 return HPAGE_PMD_NR;
259 }
260
261 return 0;
262}
263
264static struct shrinker huge_zero_page_shrinker = {
265 .count_objects = shrink_huge_zero_page_count,
266 .scan_objects = shrink_huge_zero_page_scan,
267 .seeks = DEFAULT_SEEKS,
268};
269
270#ifdef CONFIG_SYSFS
271
272static ssize_t triple_flag_store(struct kobject *kobj,
273 struct kobj_attribute *attr,
274 const char *buf, size_t count,
275 enum transparent_hugepage_flag enabled,
276 enum transparent_hugepage_flag deferred,
277 enum transparent_hugepage_flag req_madv)
278{
279 if (!memcmp("defer", buf,
280 min(sizeof("defer")-1, count))) {
281 if (enabled == deferred)
282 return -EINVAL;
283 clear_bit(enabled, &transparent_hugepage_flags);
284 clear_bit(req_madv, &transparent_hugepage_flags);
285 set_bit(deferred, &transparent_hugepage_flags);
286 } else if (!memcmp("always", buf,
287 min(sizeof("always")-1, count))) {
288 clear_bit(deferred, &transparent_hugepage_flags);
289 clear_bit(req_madv, &transparent_hugepage_flags);
290 set_bit(enabled, &transparent_hugepage_flags);
291 } else if (!memcmp("madvise", buf,
292 min(sizeof("madvise")-1, count))) {
293 clear_bit(enabled, &transparent_hugepage_flags);
294 clear_bit(deferred, &transparent_hugepage_flags);
295 set_bit(req_madv, &transparent_hugepage_flags);
296 } else if (!memcmp("never", buf,
297 min(sizeof("never")-1, count))) {
298 clear_bit(enabled, &transparent_hugepage_flags);
299 clear_bit(req_madv, &transparent_hugepage_flags);
300 clear_bit(deferred, &transparent_hugepage_flags);
301 } else
302 return -EINVAL;
303
304 return count;
305}
306
307static ssize_t enabled_show(struct kobject *kobj,
308 struct kobj_attribute *attr, char *buf)
309{
310 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
311 return sprintf(buf, "[always] madvise never\n");
312 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
313 return sprintf(buf, "always [madvise] never\n");
314 else
315 return sprintf(buf, "always madvise [never]\n");
316}
317
318static ssize_t enabled_store(struct kobject *kobj,
319 struct kobj_attribute *attr,
320 const char *buf, size_t count)
321{
322 ssize_t ret;
323
324 ret = triple_flag_store(kobj, attr, buf, count,
325 TRANSPARENT_HUGEPAGE_FLAG,
326 TRANSPARENT_HUGEPAGE_FLAG,
327 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
328
329 if (ret > 0) {
330 int err;
331
332 mutex_lock(&khugepaged_mutex);
333 err = start_stop_khugepaged();
334 mutex_unlock(&khugepaged_mutex);
335
336 if (err)
337 ret = err;
338 }
339
340 return ret;
341}
342static struct kobj_attribute enabled_attr =
343 __ATTR(enabled, 0644, enabled_show, enabled_store);
344
345static ssize_t single_flag_show(struct kobject *kobj,
346 struct kobj_attribute *attr, char *buf,
347 enum transparent_hugepage_flag flag)
348{
349 return sprintf(buf, "%d\n",
350 !!test_bit(flag, &transparent_hugepage_flags));
351}
352
353static ssize_t single_flag_store(struct kobject *kobj,
354 struct kobj_attribute *attr,
355 const char *buf, size_t count,
356 enum transparent_hugepage_flag flag)
357{
358 unsigned long value;
359 int ret;
360
361 ret = kstrtoul(buf, 10, &value);
362 if (ret < 0)
363 return ret;
364 if (value > 1)
365 return -EINVAL;
366
367 if (value)
368 set_bit(flag, &transparent_hugepage_flags);
369 else
370 clear_bit(flag, &transparent_hugepage_flags);
371
372 return count;
373}
374
375/*
376 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
377 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
378 * memory just to allocate one more hugepage.
379 */
380static ssize_t defrag_show(struct kobject *kobj,
381 struct kobj_attribute *attr, char *buf)
382{
383 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
384 return sprintf(buf, "[always] defer madvise never\n");
385 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
386 return sprintf(buf, "always [defer] madvise never\n");
387 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
388 return sprintf(buf, "always defer [madvise] never\n");
389 else
390 return sprintf(buf, "always defer madvise [never]\n");
391
392}
393static ssize_t defrag_store(struct kobject *kobj,
394 struct kobj_attribute *attr,
395 const char *buf, size_t count)
396{
397 return triple_flag_store(kobj, attr, buf, count,
398 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
399 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
400 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
401}
402static struct kobj_attribute defrag_attr =
403 __ATTR(defrag, 0644, defrag_show, defrag_store);
404
405static ssize_t use_zero_page_show(struct kobject *kobj,
406 struct kobj_attribute *attr, char *buf)
407{
408 return single_flag_show(kobj, attr, buf,
409 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
410}
411static ssize_t use_zero_page_store(struct kobject *kobj,
412 struct kobj_attribute *attr, const char *buf, size_t count)
413{
414 return single_flag_store(kobj, attr, buf, count,
415 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
416}
417static struct kobj_attribute use_zero_page_attr =
418 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
419#ifdef CONFIG_DEBUG_VM
420static ssize_t debug_cow_show(struct kobject *kobj,
421 struct kobj_attribute *attr, char *buf)
422{
423 return single_flag_show(kobj, attr, buf,
424 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
425}
426static ssize_t debug_cow_store(struct kobject *kobj,
427 struct kobj_attribute *attr,
428 const char *buf, size_t count)
429{
430 return single_flag_store(kobj, attr, buf, count,
431 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
432}
433static struct kobj_attribute debug_cow_attr =
434 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
435#endif /* CONFIG_DEBUG_VM */
436
437static struct attribute *hugepage_attr[] = {
438 &enabled_attr.attr,
439 &defrag_attr.attr,
440 &use_zero_page_attr.attr,
441#ifdef CONFIG_DEBUG_VM
442 &debug_cow_attr.attr,
443#endif
444 NULL,
445};
446
447static struct attribute_group hugepage_attr_group = {
448 .attrs = hugepage_attr,
449};
450
451static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
452 struct kobj_attribute *attr,
453 char *buf)
454{
455 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
456}
457
458static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
459 struct kobj_attribute *attr,
460 const char *buf, size_t count)
461{
462 unsigned long msecs;
463 int err;
464
465 err = kstrtoul(buf, 10, &msecs);
466 if (err || msecs > UINT_MAX)
467 return -EINVAL;
468
469 khugepaged_scan_sleep_millisecs = msecs;
470 wake_up_interruptible(&khugepaged_wait);
471
472 return count;
473}
474static struct kobj_attribute scan_sleep_millisecs_attr =
475 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
476 scan_sleep_millisecs_store);
477
478static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
479 struct kobj_attribute *attr,
480 char *buf)
481{
482 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
483}
484
485static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
486 struct kobj_attribute *attr,
487 const char *buf, size_t count)
488{
489 unsigned long msecs;
490 int err;
491
492 err = kstrtoul(buf, 10, &msecs);
493 if (err || msecs > UINT_MAX)
494 return -EINVAL;
495
496 khugepaged_alloc_sleep_millisecs = msecs;
497 wake_up_interruptible(&khugepaged_wait);
498
499 return count;
500}
501static struct kobj_attribute alloc_sleep_millisecs_attr =
502 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
503 alloc_sleep_millisecs_store);
504
505static ssize_t pages_to_scan_show(struct kobject *kobj,
506 struct kobj_attribute *attr,
507 char *buf)
508{
509 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
510}
511static ssize_t pages_to_scan_store(struct kobject *kobj,
512 struct kobj_attribute *attr,
513 const char *buf, size_t count)
514{
515 int err;
516 unsigned long pages;
517
518 err = kstrtoul(buf, 10, &pages);
519 if (err || !pages || pages > UINT_MAX)
520 return -EINVAL;
521
522 khugepaged_pages_to_scan = pages;
523
524 return count;
525}
526static struct kobj_attribute pages_to_scan_attr =
527 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
528 pages_to_scan_store);
529
530static ssize_t pages_collapsed_show(struct kobject *kobj,
531 struct kobj_attribute *attr,
532 char *buf)
533{
534 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
535}
536static struct kobj_attribute pages_collapsed_attr =
537 __ATTR_RO(pages_collapsed);
538
539static ssize_t full_scans_show(struct kobject *kobj,
540 struct kobj_attribute *attr,
541 char *buf)
542{
543 return sprintf(buf, "%u\n", khugepaged_full_scans);
544}
545static struct kobj_attribute full_scans_attr =
546 __ATTR_RO(full_scans);
547
548static ssize_t khugepaged_defrag_show(struct kobject *kobj,
549 struct kobj_attribute *attr, char *buf)
550{
551 return single_flag_show(kobj, attr, buf,
552 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
553}
554static ssize_t khugepaged_defrag_store(struct kobject *kobj,
555 struct kobj_attribute *attr,
556 const char *buf, size_t count)
557{
558 return single_flag_store(kobj, attr, buf, count,
559 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
560}
561static struct kobj_attribute khugepaged_defrag_attr =
562 __ATTR(defrag, 0644, khugepaged_defrag_show,
563 khugepaged_defrag_store);
564
565/*
566 * max_ptes_none controls if khugepaged should collapse hugepages over
567 * any unmapped ptes in turn potentially increasing the memory
568 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
569 * reduce the available free memory in the system as it
570 * runs. Increasing max_ptes_none will instead potentially reduce the
571 * free memory in the system during the khugepaged scan.
572 */
573static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
574 struct kobj_attribute *attr,
575 char *buf)
576{
577 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
578}
579static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
580 struct kobj_attribute *attr,
581 const char *buf, size_t count)
582{
583 int err;
584 unsigned long max_ptes_none;
585
586 err = kstrtoul(buf, 10, &max_ptes_none);
587 if (err || max_ptes_none > HPAGE_PMD_NR-1)
588 return -EINVAL;
589
590 khugepaged_max_ptes_none = max_ptes_none;
591
592 return count;
593}
594static struct kobj_attribute khugepaged_max_ptes_none_attr =
595 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
596 khugepaged_max_ptes_none_store);
597
598static struct attribute *khugepaged_attr[] = {
599 &khugepaged_defrag_attr.attr,
600 &khugepaged_max_ptes_none_attr.attr,
601 &pages_to_scan_attr.attr,
602 &pages_collapsed_attr.attr,
603 &full_scans_attr.attr,
604 &scan_sleep_millisecs_attr.attr,
605 &alloc_sleep_millisecs_attr.attr,
606 NULL,
607};
608
609static struct attribute_group khugepaged_attr_group = {
610 .attrs = khugepaged_attr,
611 .name = "khugepaged",
612};
613
614static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
615{
616 int err;
617
618 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
619 if (unlikely(!*hugepage_kobj)) {
620 pr_err("failed to create transparent hugepage kobject\n");
621 return -ENOMEM;
622 }
623
624 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
625 if (err) {
626 pr_err("failed to register transparent hugepage group\n");
627 goto delete_obj;
628 }
629
630 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
631 if (err) {
632 pr_err("failed to register transparent hugepage group\n");
633 goto remove_hp_group;
634 }
635
636 return 0;
637
638remove_hp_group:
639 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
640delete_obj:
641 kobject_put(*hugepage_kobj);
642 return err;
643}
644
645static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
646{
647 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
648 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
649 kobject_put(hugepage_kobj);
650}
651#else
652static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
653{
654 return 0;
655}
656
657static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
658{
659}
660#endif /* CONFIG_SYSFS */
661
662static int __init hugepage_init(void)
663{
664 int err;
665 struct kobject *hugepage_kobj;
666
667 if (!has_transparent_hugepage()) {
668 transparent_hugepage_flags = 0;
669 return -EINVAL;
670 }
671
672 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
673 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
674 /*
675 * hugepages can't be allocated by the buddy allocator
676 */
677 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
678 /*
679 * we use page->mapping and page->index in second tail page
680 * as list_head: assuming THP order >= 2
681 */
682 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
683
684 err = hugepage_init_sysfs(&hugepage_kobj);
685 if (err)
686 goto err_sysfs;
687
688 err = khugepaged_slab_init();
689 if (err)
690 goto err_slab;
691
692 err = register_shrinker(&huge_zero_page_shrinker);
693 if (err)
694 goto err_hzp_shrinker;
695 err = register_shrinker(&deferred_split_shrinker);
696 if (err)
697 goto err_split_shrinker;
698
699 /*
700 * By default disable transparent hugepages on smaller systems,
701 * where the extra memory used could hurt more than TLB overhead
702 * is likely to save. The admin can still enable it through /sys.
703 */
704 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
705 transparent_hugepage_flags = 0;
706 return 0;
707 }
708
709 err = start_stop_khugepaged();
710 if (err)
711 goto err_khugepaged;
712
713 return 0;
714err_khugepaged:
715 unregister_shrinker(&deferred_split_shrinker);
716err_split_shrinker:
717 unregister_shrinker(&huge_zero_page_shrinker);
718err_hzp_shrinker:
719 khugepaged_slab_exit();
720err_slab:
721 hugepage_exit_sysfs(hugepage_kobj);
722err_sysfs:
723 return err;
724}
725subsys_initcall(hugepage_init);
726
727static int __init setup_transparent_hugepage(char *str)
728{
729 int ret = 0;
730 if (!str)
731 goto out;
732 if (!strcmp(str, "always")) {
733 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
734 &transparent_hugepage_flags);
735 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
736 &transparent_hugepage_flags);
737 ret = 1;
738 } else if (!strcmp(str, "madvise")) {
739 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
740 &transparent_hugepage_flags);
741 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
742 &transparent_hugepage_flags);
743 ret = 1;
744 } else if (!strcmp(str, "never")) {
745 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
746 &transparent_hugepage_flags);
747 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
748 &transparent_hugepage_flags);
749 ret = 1;
750 }
751out:
752 if (!ret)
753 pr_warn("transparent_hugepage= cannot parse, ignored\n");
754 return ret;
755}
756__setup("transparent_hugepage=", setup_transparent_hugepage);
757
758pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
759{
760 if (likely(vma->vm_flags & VM_WRITE))
761 pmd = pmd_mkwrite(pmd);
762 return pmd;
763}
764
765static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
766{
767 pmd_t entry;
768 entry = mk_pmd(page, prot);
769 entry = pmd_mkhuge(entry);
770 return entry;
771}
772
773static inline struct list_head *page_deferred_list(struct page *page)
774{
775 /*
776 * ->lru in the tail pages is occupied by compound_head.
777 * Let's use ->mapping + ->index in the second tail page as list_head.
778 */
779 return (struct list_head *)&page[2].mapping;
780}
781
782void prep_transhuge_page(struct page *page)
783{
784 /*
785 * we use page->mapping and page->indexlru in second tail page
786 * as list_head: assuming THP order >= 2
787 */
788
789 INIT_LIST_HEAD(page_deferred_list(page));
790 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
791}
792
793static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
794 struct vm_area_struct *vma,
795 unsigned long address, pmd_t *pmd,
796 struct page *page, gfp_t gfp,
797 unsigned int flags)
798{
799 struct mem_cgroup *memcg;
800 pgtable_t pgtable;
801 spinlock_t *ptl;
802 unsigned long haddr = address & HPAGE_PMD_MASK;
803
804 VM_BUG_ON_PAGE(!PageCompound(page), page);
805
806 if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
807 put_page(page);
808 count_vm_event(THP_FAULT_FALLBACK);
809 return VM_FAULT_FALLBACK;
810 }
811
812 pgtable = pte_alloc_one(mm, haddr);
813 if (unlikely(!pgtable)) {
814 mem_cgroup_cancel_charge(page, memcg, true);
815 put_page(page);
816 return VM_FAULT_OOM;
817 }
818
819 clear_huge_page(page, haddr, HPAGE_PMD_NR);
820 /*
821 * The memory barrier inside __SetPageUptodate makes sure that
822 * clear_huge_page writes become visible before the set_pmd_at()
823 * write.
824 */
825 __SetPageUptodate(page);
826
827 ptl = pmd_lock(mm, pmd);
828 if (unlikely(!pmd_none(*pmd))) {
829 spin_unlock(ptl);
830 mem_cgroup_cancel_charge(page, memcg, true);
831 put_page(page);
832 pte_free(mm, pgtable);
833 } else {
834 pmd_t entry;
835
836 /* Deliver the page fault to userland */
837 if (userfaultfd_missing(vma)) {
838 int ret;
839
840 spin_unlock(ptl);
841 mem_cgroup_cancel_charge(page, memcg, true);
842 put_page(page);
843 pte_free(mm, pgtable);
844 ret = handle_userfault(vma, address, flags,
845 VM_UFFD_MISSING);
846 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
847 return ret;
848 }
849
850 entry = mk_huge_pmd(page, vma->vm_page_prot);
851 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
852 page_add_new_anon_rmap(page, vma, haddr, true);
853 mem_cgroup_commit_charge(page, memcg, false, true);
854 lru_cache_add_active_or_unevictable(page, vma);
855 pgtable_trans_huge_deposit(mm, pmd, pgtable);
856 set_pmd_at(mm, haddr, pmd, entry);
857 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
858 atomic_long_inc(&mm->nr_ptes);
859 spin_unlock(ptl);
860 count_vm_event(THP_FAULT_ALLOC);
861 }
862
863 return 0;
864}
865
866/*
867 * If THP is set to always then directly reclaim/compact as necessary
868 * If set to defer then do no reclaim and defer to khugepaged
869 * If set to madvise and the VMA is flagged then directly reclaim/compact
870 */
871static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
872{
873 gfp_t reclaim_flags = 0;
874
875 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags) &&
876 (vma->vm_flags & VM_HUGEPAGE))
877 reclaim_flags = __GFP_DIRECT_RECLAIM;
878 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
879 reclaim_flags = __GFP_KSWAPD_RECLAIM;
880 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
881 reclaim_flags = __GFP_DIRECT_RECLAIM;
882
883 return GFP_TRANSHUGE | reclaim_flags;
884}
885
886/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
887static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
888{
889 return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0);
890}
891
892/* Caller must hold page table lock. */
893static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
894 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
895 struct page *zero_page)
896{
897 pmd_t entry;
898 if (!pmd_none(*pmd))
899 return false;
900 entry = mk_pmd(zero_page, vma->vm_page_prot);
901 entry = pmd_mkhuge(entry);
902 if (pgtable)
903 pgtable_trans_huge_deposit(mm, pmd, pgtable);
904 set_pmd_at(mm, haddr, pmd, entry);
905 atomic_long_inc(&mm->nr_ptes);
906 return true;
907}
908
909int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
910 unsigned long address, pmd_t *pmd,
911 unsigned int flags)
912{
913 gfp_t gfp;
914 struct page *page;
915 unsigned long haddr = address & HPAGE_PMD_MASK;
916
917 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
918 return VM_FAULT_FALLBACK;
919 if (unlikely(anon_vma_prepare(vma)))
920 return VM_FAULT_OOM;
921 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
922 return VM_FAULT_OOM;
923 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
924 transparent_hugepage_use_zero_page()) {
925 spinlock_t *ptl;
926 pgtable_t pgtable;
927 struct page *zero_page;
928 bool set;
929 int ret;
930 pgtable = pte_alloc_one(mm, haddr);
931 if (unlikely(!pgtable))
932 return VM_FAULT_OOM;
933 zero_page = get_huge_zero_page();
934 if (unlikely(!zero_page)) {
935 pte_free(mm, pgtable);
936 count_vm_event(THP_FAULT_FALLBACK);
937 return VM_FAULT_FALLBACK;
938 }
939 ptl = pmd_lock(mm, pmd);
940 ret = 0;
941 set = false;
942 if (pmd_none(*pmd)) {
943 if (userfaultfd_missing(vma)) {
944 spin_unlock(ptl);
945 ret = handle_userfault(vma, address, flags,
946 VM_UFFD_MISSING);
947 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
948 } else {
949 set_huge_zero_page(pgtable, mm, vma,
950 haddr, pmd,
951 zero_page);
952 spin_unlock(ptl);
953 set = true;
954 }
955 } else
956 spin_unlock(ptl);
957 if (!set) {
958 pte_free(mm, pgtable);
959 put_huge_zero_page();
960 }
961 return ret;
962 }
963 gfp = alloc_hugepage_direct_gfpmask(vma);
964 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
965 if (unlikely(!page)) {
966 count_vm_event(THP_FAULT_FALLBACK);
967 return VM_FAULT_FALLBACK;
968 }
969 prep_transhuge_page(page);
970 return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
971 flags);
972}
973
974static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
975 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
976{
977 struct mm_struct *mm = vma->vm_mm;
978 pmd_t entry;
979 spinlock_t *ptl;
980
981 ptl = pmd_lock(mm, pmd);
982 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
983 if (pfn_t_devmap(pfn))
984 entry = pmd_mkdevmap(entry);
985 if (write) {
986 entry = pmd_mkyoung(pmd_mkdirty(entry));
987 entry = maybe_pmd_mkwrite(entry, vma);
988 }
989 set_pmd_at(mm, addr, pmd, entry);
990 update_mmu_cache_pmd(vma, addr, pmd);
991 spin_unlock(ptl);
992}
993
994int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
995 pmd_t *pmd, pfn_t pfn, bool write)
996{
997 pgprot_t pgprot = vma->vm_page_prot;
998 /*
999 * If we had pmd_special, we could avoid all these restrictions,
1000 * but we need to be consistent with PTEs and architectures that
1001 * can't support a 'special' bit.
1002 */
1003 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1004 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1005 (VM_PFNMAP|VM_MIXEDMAP));
1006 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1007 BUG_ON(!pfn_t_devmap(pfn));
1008
1009 if (addr < vma->vm_start || addr >= vma->vm_end)
1010 return VM_FAULT_SIGBUS;
1011 if (track_pfn_insert(vma, &pgprot, pfn))
1012 return VM_FAULT_SIGBUS;
1013 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
1014 return VM_FAULT_NOPAGE;
1015}
1016
1017static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1018 pmd_t *pmd)
1019{
1020 pmd_t _pmd;
1021
1022 /*
1023 * We should set the dirty bit only for FOLL_WRITE but for now
1024 * the dirty bit in the pmd is meaningless. And if the dirty
1025 * bit will become meaningful and we'll only set it with
1026 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
1027 * set the young bit, instead of the current set_pmd_at.
1028 */
1029 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1030 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1031 pmd, _pmd, 1))
1032 update_mmu_cache_pmd(vma, addr, pmd);
1033}
1034
1035struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1036 pmd_t *pmd, int flags)
1037{
1038 unsigned long pfn = pmd_pfn(*pmd);
1039 struct mm_struct *mm = vma->vm_mm;
1040 struct dev_pagemap *pgmap;
1041 struct page *page;
1042
1043 assert_spin_locked(pmd_lockptr(mm, pmd));
1044
1045 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1046 return NULL;
1047
1048 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1049 /* pass */;
1050 else
1051 return NULL;
1052
1053 if (flags & FOLL_TOUCH)
1054 touch_pmd(vma, addr, pmd);
1055
1056 /*
1057 * device mapped pages can only be returned if the
1058 * caller will manage the page reference count.
1059 */
1060 if (!(flags & FOLL_GET))
1061 return ERR_PTR(-EEXIST);
1062
1063 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1064 pgmap = get_dev_pagemap(pfn, NULL);
1065 if (!pgmap)
1066 return ERR_PTR(-EFAULT);
1067 page = pfn_to_page(pfn);
1068 get_page(page);
1069 put_dev_pagemap(pgmap);
1070
1071 return page;
1072}
1073
1074int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1075 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1076 struct vm_area_struct *vma)
1077{
1078 spinlock_t *dst_ptl, *src_ptl;
1079 struct page *src_page;
1080 pmd_t pmd;
1081 pgtable_t pgtable = NULL;
1082 int ret;
1083
1084 if (!vma_is_dax(vma)) {
1085 ret = -ENOMEM;
1086 pgtable = pte_alloc_one(dst_mm, addr);
1087 if (unlikely(!pgtable))
1088 goto out;
1089 }
1090
1091 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1092 src_ptl = pmd_lockptr(src_mm, src_pmd);
1093 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1094
1095 ret = -EAGAIN;
1096 pmd = *src_pmd;
1097 if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1098 pte_free(dst_mm, pgtable);
1099 goto out_unlock;
1100 }
1101 /*
1102 * When page table lock is held, the huge zero pmd should not be
1103 * under splitting since we don't split the page itself, only pmd to
1104 * a page table.
1105 */
1106 if (is_huge_zero_pmd(pmd)) {
1107 struct page *zero_page;
1108 /*
1109 * get_huge_zero_page() will never allocate a new page here,
1110 * since we already have a zero page to copy. It just takes a
1111 * reference.
1112 */
1113 zero_page = get_huge_zero_page();
1114 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1115 zero_page);
1116 ret = 0;
1117 goto out_unlock;
1118 }
1119
1120 if (!vma_is_dax(vma)) {
1121 /* thp accounting separate from pmd_devmap accounting */
1122 src_page = pmd_page(pmd);
1123 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1124 get_page(src_page);
1125 page_dup_rmap(src_page, true);
1126 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1127 atomic_long_inc(&dst_mm->nr_ptes);
1128 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1129 }
1130
1131 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1132 pmd = pmd_mkold(pmd_wrprotect(pmd));
1133 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1134
1135 ret = 0;
1136out_unlock:
1137 spin_unlock(src_ptl);
1138 spin_unlock(dst_ptl);
1139out:
1140 return ret;
1141}
1142
1143void huge_pmd_set_accessed(struct mm_struct *mm,
1144 struct vm_area_struct *vma,
1145 unsigned long address,
1146 pmd_t *pmd, pmd_t orig_pmd,
1147 int dirty)
1148{
1149 spinlock_t *ptl;
1150 pmd_t entry;
1151 unsigned long haddr;
1152
1153 ptl = pmd_lock(mm, pmd);
1154 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1155 goto unlock;
1156
1157 entry = pmd_mkyoung(orig_pmd);
1158 haddr = address & HPAGE_PMD_MASK;
1159 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1160 update_mmu_cache_pmd(vma, address, pmd);
1161
1162unlock:
1163 spin_unlock(ptl);
1164}
1165
1166static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1167 struct vm_area_struct *vma,
1168 unsigned long address,
1169 pmd_t *pmd, pmd_t orig_pmd,
1170 struct page *page,
1171 unsigned long haddr)
1172{
1173 struct mem_cgroup *memcg;
1174 spinlock_t *ptl;
1175 pgtable_t pgtable;
1176 pmd_t _pmd;
1177 int ret = 0, i;
1178 struct page **pages;
1179 unsigned long mmun_start; /* For mmu_notifiers */
1180 unsigned long mmun_end; /* For mmu_notifiers */
1181
1182 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1183 GFP_KERNEL);
1184 if (unlikely(!pages)) {
1185 ret |= VM_FAULT_OOM;
1186 goto out;
1187 }
1188
1189 for (i = 0; i < HPAGE_PMD_NR; i++) {
1190 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1191 __GFP_OTHER_NODE,
1192 vma, address, page_to_nid(page));
1193 if (unlikely(!pages[i] ||
1194 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1195 &memcg, false))) {
1196 if (pages[i])
1197 put_page(pages[i]);
1198 while (--i >= 0) {
1199 memcg = (void *)page_private(pages[i]);
1200 set_page_private(pages[i], 0);
1201 mem_cgroup_cancel_charge(pages[i], memcg,
1202 false);
1203 put_page(pages[i]);
1204 }
1205 kfree(pages);
1206 ret |= VM_FAULT_OOM;
1207 goto out;
1208 }
1209 set_page_private(pages[i], (unsigned long)memcg);
1210 }
1211
1212 for (i = 0; i < HPAGE_PMD_NR; i++) {
1213 copy_user_highpage(pages[i], page + i,
1214 haddr + PAGE_SIZE * i, vma);
1215 __SetPageUptodate(pages[i]);
1216 cond_resched();
1217 }
1218
1219 mmun_start = haddr;
1220 mmun_end = haddr + HPAGE_PMD_SIZE;
1221 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1222
1223 ptl = pmd_lock(mm, pmd);
1224 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1225 goto out_free_pages;
1226 VM_BUG_ON_PAGE(!PageHead(page), page);
1227
1228 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1229 /* leave pmd empty until pte is filled */
1230
1231 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1232 pmd_populate(mm, &_pmd, pgtable);
1233
1234 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1235 pte_t *pte, entry;
1236 entry = mk_pte(pages[i], vma->vm_page_prot);
1237 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1238 memcg = (void *)page_private(pages[i]);
1239 set_page_private(pages[i], 0);
1240 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1241 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1242 lru_cache_add_active_or_unevictable(pages[i], vma);
1243 pte = pte_offset_map(&_pmd, haddr);
1244 VM_BUG_ON(!pte_none(*pte));
1245 set_pte_at(mm, haddr, pte, entry);
1246 pte_unmap(pte);
1247 }
1248 kfree(pages);
1249
1250 smp_wmb(); /* make pte visible before pmd */
1251 pmd_populate(mm, pmd, pgtable);
1252 page_remove_rmap(page, true);
1253 spin_unlock(ptl);
1254
1255 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1256
1257 ret |= VM_FAULT_WRITE;
1258 put_page(page);
1259
1260out:
1261 return ret;
1262
1263out_free_pages:
1264 spin_unlock(ptl);
1265 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1266 for (i = 0; i < HPAGE_PMD_NR; i++) {
1267 memcg = (void *)page_private(pages[i]);
1268 set_page_private(pages[i], 0);
1269 mem_cgroup_cancel_charge(pages[i], memcg, false);
1270 put_page(pages[i]);
1271 }
1272 kfree(pages);
1273 goto out;
1274}
1275
1276int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1277 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1278{
1279 spinlock_t *ptl;
1280 int ret = 0;
1281 struct page *page = NULL, *new_page;
1282 struct mem_cgroup *memcg;
1283 unsigned long haddr;
1284 unsigned long mmun_start; /* For mmu_notifiers */
1285 unsigned long mmun_end; /* For mmu_notifiers */
1286 gfp_t huge_gfp; /* for allocation and charge */
1287
1288 ptl = pmd_lockptr(mm, pmd);
1289 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1290 haddr = address & HPAGE_PMD_MASK;
1291 if (is_huge_zero_pmd(orig_pmd))
1292 goto alloc;
1293 spin_lock(ptl);
1294 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1295 goto out_unlock;
1296
1297 page = pmd_page(orig_pmd);
1298 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1299 /*
1300 * We can only reuse the page if nobody else maps the huge page or it's
1301 * part.
1302 */
1303 if (page_trans_huge_mapcount(page, NULL) == 1) {
1304 pmd_t entry;
1305 entry = pmd_mkyoung(orig_pmd);
1306 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1307 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1308 update_mmu_cache_pmd(vma, address, pmd);
1309 ret |= VM_FAULT_WRITE;
1310 goto out_unlock;
1311 }
1312 get_page(page);
1313 spin_unlock(ptl);
1314alloc:
1315 if (transparent_hugepage_enabled(vma) &&
1316 !transparent_hugepage_debug_cow()) {
1317 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1318 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1319 } else
1320 new_page = NULL;
1321
1322 if (likely(new_page)) {
1323 prep_transhuge_page(new_page);
1324 } else {
1325 if (!page) {
1326 split_huge_pmd(vma, pmd, address);
1327 ret |= VM_FAULT_FALLBACK;
1328 } else {
1329 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1330 pmd, orig_pmd, page, haddr);
1331 if (ret & VM_FAULT_OOM) {
1332 split_huge_pmd(vma, pmd, address);
1333 ret |= VM_FAULT_FALLBACK;
1334 }
1335 put_page(page);
1336 }
1337 count_vm_event(THP_FAULT_FALLBACK);
1338 goto out;
1339 }
1340
1341 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1342 true))) {
1343 put_page(new_page);
1344 if (page) {
1345 split_huge_pmd(vma, pmd, address);
1346 put_page(page);
1347 } else
1348 split_huge_pmd(vma, pmd, address);
1349 ret |= VM_FAULT_FALLBACK;
1350 count_vm_event(THP_FAULT_FALLBACK);
1351 goto out;
1352 }
1353
1354 count_vm_event(THP_FAULT_ALLOC);
1355
1356 if (!page)
1357 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1358 else
1359 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1360 __SetPageUptodate(new_page);
1361
1362 mmun_start = haddr;
1363 mmun_end = haddr + HPAGE_PMD_SIZE;
1364 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1365
1366 spin_lock(ptl);
1367 if (page)
1368 put_page(page);
1369 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1370 spin_unlock(ptl);
1371 mem_cgroup_cancel_charge(new_page, memcg, true);
1372 put_page(new_page);
1373 goto out_mn;
1374 } else {
1375 pmd_t entry;
1376 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1377 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1378 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1379 page_add_new_anon_rmap(new_page, vma, haddr, true);
1380 mem_cgroup_commit_charge(new_page, memcg, false, true);
1381 lru_cache_add_active_or_unevictable(new_page, vma);
1382 set_pmd_at(mm, haddr, pmd, entry);
1383 update_mmu_cache_pmd(vma, address, pmd);
1384 if (!page) {
1385 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1386 put_huge_zero_page();
1387 } else {
1388 VM_BUG_ON_PAGE(!PageHead(page), page);
1389 page_remove_rmap(page, true);
1390 put_page(page);
1391 }
1392 ret |= VM_FAULT_WRITE;
1393 }
1394 spin_unlock(ptl);
1395out_mn:
1396 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1397out:
1398 return ret;
1399out_unlock:
1400 spin_unlock(ptl);
1401 return ret;
1402}
1403
1404struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1405 unsigned long addr,
1406 pmd_t *pmd,
1407 unsigned int flags)
1408{
1409 struct mm_struct *mm = vma->vm_mm;
1410 struct page *page = NULL;
1411
1412 assert_spin_locked(pmd_lockptr(mm, pmd));
1413
1414 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1415 goto out;
1416
1417 /* Avoid dumping huge zero page */
1418 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1419 return ERR_PTR(-EFAULT);
1420
1421 /* Full NUMA hinting faults to serialise migration in fault paths */
1422 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1423 goto out;
1424
1425 page = pmd_page(*pmd);
1426 VM_BUG_ON_PAGE(!PageHead(page), page);
1427 if (flags & FOLL_TOUCH)
1428 touch_pmd(vma, addr, pmd);
1429 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1430 /*
1431 * We don't mlock() pte-mapped THPs. This way we can avoid
1432 * leaking mlocked pages into non-VM_LOCKED VMAs.
1433 *
1434 * In most cases the pmd is the only mapping of the page as we
1435 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1436 * writable private mappings in populate_vma_page_range().
1437 *
1438 * The only scenario when we have the page shared here is if we
1439 * mlocking read-only mapping shared over fork(). We skip
1440 * mlocking such pages.
1441 */
1442 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1443 page->mapping && trylock_page(page)) {
1444 lru_add_drain();
1445 if (page->mapping)
1446 mlock_vma_page(page);
1447 unlock_page(page);
1448 }
1449 }
1450 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1451 VM_BUG_ON_PAGE(!PageCompound(page), page);
1452 if (flags & FOLL_GET)
1453 get_page(page);
1454
1455out:
1456 return page;
1457}
1458
1459/* NUMA hinting page fault entry point for trans huge pmds */
1460int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1461 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1462{
1463 spinlock_t *ptl;
1464 struct anon_vma *anon_vma = NULL;
1465 struct page *page;
1466 unsigned long haddr = addr & HPAGE_PMD_MASK;
1467 int page_nid = -1, this_nid = numa_node_id();
1468 int target_nid, last_cpupid = -1;
1469 bool page_locked;
1470 bool migrated = false;
1471 bool was_writable;
1472 int flags = 0;
1473
1474 /* A PROT_NONE fault should not end up here */
1475 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1476
1477 ptl = pmd_lock(mm, pmdp);
1478 if (unlikely(!pmd_same(pmd, *pmdp)))
1479 goto out_unlock;
1480
1481 /*
1482 * If there are potential migrations, wait for completion and retry
1483 * without disrupting NUMA hinting information. Do not relock and
1484 * check_same as the page may no longer be mapped.
1485 */
1486 if (unlikely(pmd_trans_migrating(*pmdp))) {
1487 page = pmd_page(*pmdp);
1488 spin_unlock(ptl);
1489 wait_on_page_locked(page);
1490 goto out;
1491 }
1492
1493 page = pmd_page(pmd);
1494 BUG_ON(is_huge_zero_page(page));
1495 page_nid = page_to_nid(page);
1496 last_cpupid = page_cpupid_last(page);
1497 count_vm_numa_event(NUMA_HINT_FAULTS);
1498 if (page_nid == this_nid) {
1499 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1500 flags |= TNF_FAULT_LOCAL;
1501 }
1502
1503 /* See similar comment in do_numa_page for explanation */
1504 if (!(vma->vm_flags & VM_WRITE))
1505 flags |= TNF_NO_GROUP;
1506
1507 /*
1508 * Acquire the page lock to serialise THP migrations but avoid dropping
1509 * page_table_lock if at all possible
1510 */
1511 page_locked = trylock_page(page);
1512 target_nid = mpol_misplaced(page, vma, haddr);
1513 if (target_nid == -1) {
1514 /* If the page was locked, there are no parallel migrations */
1515 if (page_locked)
1516 goto clear_pmdnuma;
1517 }
1518
1519 /* Migration could have started since the pmd_trans_migrating check */
1520 if (!page_locked) {
1521 spin_unlock(ptl);
1522 wait_on_page_locked(page);
1523 page_nid = -1;
1524 goto out;
1525 }
1526
1527 /*
1528 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1529 * to serialises splits
1530 */
1531 get_page(page);
1532 spin_unlock(ptl);
1533 anon_vma = page_lock_anon_vma_read(page);
1534
1535 /* Confirm the PMD did not change while page_table_lock was released */
1536 spin_lock(ptl);
1537 if (unlikely(!pmd_same(pmd, *pmdp))) {
1538 unlock_page(page);
1539 put_page(page);
1540 page_nid = -1;
1541 goto out_unlock;
1542 }
1543
1544 /* Bail if we fail to protect against THP splits for any reason */
1545 if (unlikely(!anon_vma)) {
1546 put_page(page);
1547 page_nid = -1;
1548 goto clear_pmdnuma;
1549 }
1550
1551 /*
1552 * Migrate the THP to the requested node, returns with page unlocked
1553 * and access rights restored.
1554 */
1555 spin_unlock(ptl);
1556 migrated = migrate_misplaced_transhuge_page(mm, vma,
1557 pmdp, pmd, addr, page, target_nid);
1558 if (migrated) {
1559 flags |= TNF_MIGRATED;
1560 page_nid = target_nid;
1561 } else
1562 flags |= TNF_MIGRATE_FAIL;
1563
1564 goto out;
1565clear_pmdnuma:
1566 BUG_ON(!PageLocked(page));
1567 was_writable = pmd_write(pmd);
1568 pmd = pmd_modify(pmd, vma->vm_page_prot);
1569 pmd = pmd_mkyoung(pmd);
1570 if (was_writable)
1571 pmd = pmd_mkwrite(pmd);
1572 set_pmd_at(mm, haddr, pmdp, pmd);
1573 update_mmu_cache_pmd(vma, addr, pmdp);
1574 unlock_page(page);
1575out_unlock:
1576 spin_unlock(ptl);
1577
1578out:
1579 if (anon_vma)
1580 page_unlock_anon_vma_read(anon_vma);
1581
1582 if (page_nid != -1)
1583 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1584
1585 return 0;
1586}
1587
1588int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1589 pmd_t *pmd, unsigned long addr, unsigned long next)
1590
1591{
1592 spinlock_t *ptl;
1593 pmd_t orig_pmd;
1594 struct page *page;
1595 struct mm_struct *mm = tlb->mm;
1596 int ret = 0;
1597
1598 ptl = pmd_trans_huge_lock(pmd, vma);
1599 if (!ptl)
1600 goto out_unlocked;
1601
1602 orig_pmd = *pmd;
1603 if (is_huge_zero_pmd(orig_pmd)) {
1604 ret = 1;
1605 goto out;
1606 }
1607
1608 page = pmd_page(orig_pmd);
1609 /*
1610 * If other processes are mapping this page, we couldn't discard
1611 * the page unless they all do MADV_FREE so let's skip the page.
1612 */
1613 if (page_mapcount(page) != 1)
1614 goto out;
1615
1616 if (!trylock_page(page))
1617 goto out;
1618
1619 /*
1620 * If user want to discard part-pages of THP, split it so MADV_FREE
1621 * will deactivate only them.
1622 */
1623 if (next - addr != HPAGE_PMD_SIZE) {
1624 get_page(page);
1625 spin_unlock(ptl);
1626 if (split_huge_page(page)) {
1627 put_page(page);
1628 unlock_page(page);
1629 goto out_unlocked;
1630 }
1631 put_page(page);
1632 unlock_page(page);
1633 ret = 1;
1634 goto out_unlocked;
1635 }
1636
1637 if (PageDirty(page))
1638 ClearPageDirty(page);
1639 unlock_page(page);
1640
1641 if (PageActive(page))
1642 deactivate_page(page);
1643
1644 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1645 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1646 tlb->fullmm);
1647 orig_pmd = pmd_mkold(orig_pmd);
1648 orig_pmd = pmd_mkclean(orig_pmd);
1649
1650 set_pmd_at(mm, addr, pmd, orig_pmd);
1651 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1652 }
1653 ret = 1;
1654out:
1655 spin_unlock(ptl);
1656out_unlocked:
1657 return ret;
1658}
1659
1660int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1661 pmd_t *pmd, unsigned long addr)
1662{
1663 pmd_t orig_pmd;
1664 spinlock_t *ptl;
1665
1666 ptl = __pmd_trans_huge_lock(pmd, vma);
1667 if (!ptl)
1668 return 0;
1669 /*
1670 * For architectures like ppc64 we look at deposited pgtable
1671 * when calling pmdp_huge_get_and_clear. So do the
1672 * pgtable_trans_huge_withdraw after finishing pmdp related
1673 * operations.
1674 */
1675 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1676 tlb->fullmm);
1677 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1678 if (vma_is_dax(vma)) {
1679 spin_unlock(ptl);
1680 if (is_huge_zero_pmd(orig_pmd))
1681 tlb_remove_page(tlb, pmd_page(orig_pmd));
1682 } else if (is_huge_zero_pmd(orig_pmd)) {
1683 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1684 atomic_long_dec(&tlb->mm->nr_ptes);
1685 spin_unlock(ptl);
1686 tlb_remove_page(tlb, pmd_page(orig_pmd));
1687 } else {
1688 struct page *page = pmd_page(orig_pmd);
1689 page_remove_rmap(page, true);
1690 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1691 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1692 VM_BUG_ON_PAGE(!PageHead(page), page);
1693 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1694 atomic_long_dec(&tlb->mm->nr_ptes);
1695 spin_unlock(ptl);
1696 tlb_remove_page(tlb, page);
1697 }
1698 return 1;
1699}
1700
1701bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1702 unsigned long old_addr,
1703 unsigned long new_addr, unsigned long old_end,
1704 pmd_t *old_pmd, pmd_t *new_pmd)
1705{
1706 spinlock_t *old_ptl, *new_ptl;
1707 pmd_t pmd;
1708
1709 struct mm_struct *mm = vma->vm_mm;
1710
1711 if ((old_addr & ~HPAGE_PMD_MASK) ||
1712 (new_addr & ~HPAGE_PMD_MASK) ||
1713 old_end - old_addr < HPAGE_PMD_SIZE ||
1714 (new_vma->vm_flags & VM_NOHUGEPAGE))
1715 return false;
1716
1717 /*
1718 * The destination pmd shouldn't be established, free_pgtables()
1719 * should have release it.
1720 */
1721 if (WARN_ON(!pmd_none(*new_pmd))) {
1722 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1723 return false;
1724 }
1725
1726 /*
1727 * We don't have to worry about the ordering of src and dst
1728 * ptlocks because exclusive mmap_sem prevents deadlock.
1729 */
1730 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1731 if (old_ptl) {
1732 new_ptl = pmd_lockptr(mm, new_pmd);
1733 if (new_ptl != old_ptl)
1734 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1735 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1736 VM_BUG_ON(!pmd_none(*new_pmd));
1737
1738 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1739 vma_is_anonymous(vma)) {
1740 pgtable_t pgtable;
1741 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1742 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1743 }
1744 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1745 if (new_ptl != old_ptl)
1746 spin_unlock(new_ptl);
1747 spin_unlock(old_ptl);
1748 return true;
1749 }
1750 return false;
1751}
1752
1753/*
1754 * Returns
1755 * - 0 if PMD could not be locked
1756 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1757 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1758 */
1759int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1760 unsigned long addr, pgprot_t newprot, int prot_numa)
1761{
1762 struct mm_struct *mm = vma->vm_mm;
1763 spinlock_t *ptl;
1764 int ret = 0;
1765
1766 ptl = __pmd_trans_huge_lock(pmd, vma);
1767 if (ptl) {
1768 pmd_t entry;
1769 bool preserve_write = prot_numa && pmd_write(*pmd);
1770 ret = 1;
1771
1772 /*
1773 * Avoid trapping faults against the zero page. The read-only
1774 * data is likely to be read-cached on the local CPU and
1775 * local/remote hits to the zero page are not interesting.
1776 */
1777 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1778 spin_unlock(ptl);
1779 return ret;
1780 }
1781
1782 if (!prot_numa || !pmd_protnone(*pmd)) {
1783 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1784 entry = pmd_modify(entry, newprot);
1785 if (preserve_write)
1786 entry = pmd_mkwrite(entry);
1787 ret = HPAGE_PMD_NR;
1788 set_pmd_at(mm, addr, pmd, entry);
1789 BUG_ON(!preserve_write && pmd_write(entry));
1790 }
1791 spin_unlock(ptl);
1792 }
1793
1794 return ret;
1795}
1796
1797/*
1798 * Returns true if a given pmd maps a thp, false otherwise.
1799 *
1800 * Note that if it returns true, this routine returns without unlocking page
1801 * table lock. So callers must unlock it.
1802 */
1803spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1804{
1805 spinlock_t *ptl;
1806 ptl = pmd_lock(vma->vm_mm, pmd);
1807 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1808 return ptl;
1809 spin_unlock(ptl);
1810 return NULL;
1811}
1812
1813#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1814
1815int hugepage_madvise(struct vm_area_struct *vma,
1816 unsigned long *vm_flags, int advice)
1817{
1818 switch (advice) {
1819 case MADV_HUGEPAGE:
1820#ifdef CONFIG_S390
1821 /*
1822 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1823 * can't handle this properly after s390_enable_sie, so we simply
1824 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1825 */
1826 if (mm_has_pgste(vma->vm_mm))
1827 return 0;
1828#endif
1829 /*
1830 * Be somewhat over-protective like KSM for now!
1831 */
1832 if (*vm_flags & VM_NO_THP)
1833 return -EINVAL;
1834 *vm_flags &= ~VM_NOHUGEPAGE;
1835 *vm_flags |= VM_HUGEPAGE;
1836 /*
1837 * If the vma become good for khugepaged to scan,
1838 * register it here without waiting a page fault that
1839 * may not happen any time soon.
1840 */
1841 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1842 return -ENOMEM;
1843 break;
1844 case MADV_NOHUGEPAGE:
1845 /*
1846 * Be somewhat over-protective like KSM for now!
1847 */
1848 if (*vm_flags & VM_NO_THP)
1849 return -EINVAL;
1850 *vm_flags &= ~VM_HUGEPAGE;
1851 *vm_flags |= VM_NOHUGEPAGE;
1852 /*
1853 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1854 * this vma even if we leave the mm registered in khugepaged if
1855 * it got registered before VM_NOHUGEPAGE was set.
1856 */
1857 break;
1858 }
1859
1860 return 0;
1861}
1862
1863static int __init khugepaged_slab_init(void)
1864{
1865 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1866 sizeof(struct mm_slot),
1867 __alignof__(struct mm_slot), 0, NULL);
1868 if (!mm_slot_cache)
1869 return -ENOMEM;
1870
1871 return 0;
1872}
1873
1874static void __init khugepaged_slab_exit(void)
1875{
1876 kmem_cache_destroy(mm_slot_cache);
1877}
1878
1879static inline struct mm_slot *alloc_mm_slot(void)
1880{
1881 if (!mm_slot_cache) /* initialization failed */
1882 return NULL;
1883 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1884}
1885
1886static inline void free_mm_slot(struct mm_slot *mm_slot)
1887{
1888 kmem_cache_free(mm_slot_cache, mm_slot);
1889}
1890
1891static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1892{
1893 struct mm_slot *mm_slot;
1894
1895 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1896 if (mm == mm_slot->mm)
1897 return mm_slot;
1898
1899 return NULL;
1900}
1901
1902static void insert_to_mm_slots_hash(struct mm_struct *mm,
1903 struct mm_slot *mm_slot)
1904{
1905 mm_slot->mm = mm;
1906 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1907}
1908
1909static inline int khugepaged_test_exit(struct mm_struct *mm)
1910{
1911 return atomic_read(&mm->mm_users) == 0;
1912}
1913
1914int __khugepaged_enter(struct mm_struct *mm)
1915{
1916 struct mm_slot *mm_slot;
1917 int wakeup;
1918
1919 mm_slot = alloc_mm_slot();
1920 if (!mm_slot)
1921 return -ENOMEM;
1922
1923 /* __khugepaged_exit() must not run from under us */
1924 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1925 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1926 free_mm_slot(mm_slot);
1927 return 0;
1928 }
1929
1930 spin_lock(&khugepaged_mm_lock);
1931 insert_to_mm_slots_hash(mm, mm_slot);
1932 /*
1933 * Insert just behind the scanning cursor, to let the area settle
1934 * down a little.
1935 */
1936 wakeup = list_empty(&khugepaged_scan.mm_head);
1937 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1938 spin_unlock(&khugepaged_mm_lock);
1939
1940 atomic_inc(&mm->mm_count);
1941 if (wakeup)
1942 wake_up_interruptible(&khugepaged_wait);
1943
1944 return 0;
1945}
1946
1947int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1948 unsigned long vm_flags)
1949{
1950 unsigned long hstart, hend;
1951 if (!vma->anon_vma)
1952 /*
1953 * Not yet faulted in so we will register later in the
1954 * page fault if needed.
1955 */
1956 return 0;
1957 if (vma->vm_ops || (vm_flags & VM_NO_THP))
1958 /* khugepaged not yet working on file or special mappings */
1959 return 0;
1960 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1961 hend = vma->vm_end & HPAGE_PMD_MASK;
1962 if (hstart < hend)
1963 return khugepaged_enter(vma, vm_flags);
1964 return 0;
1965}
1966
1967void __khugepaged_exit(struct mm_struct *mm)
1968{
1969 struct mm_slot *mm_slot;
1970 int free = 0;
1971
1972 spin_lock(&khugepaged_mm_lock);
1973 mm_slot = get_mm_slot(mm);
1974 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1975 hash_del(&mm_slot->hash);
1976 list_del(&mm_slot->mm_node);
1977 free = 1;
1978 }
1979 spin_unlock(&khugepaged_mm_lock);
1980
1981 if (free) {
1982 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1983 free_mm_slot(mm_slot);
1984 mmdrop(mm);
1985 } else if (mm_slot) {
1986 /*
1987 * This is required to serialize against
1988 * khugepaged_test_exit() (which is guaranteed to run
1989 * under mmap sem read mode). Stop here (after we
1990 * return all pagetables will be destroyed) until
1991 * khugepaged has finished working on the pagetables
1992 * under the mmap_sem.
1993 */
1994 down_write(&mm->mmap_sem);
1995 up_write(&mm->mmap_sem);
1996 }
1997}
1998
1999static void release_pte_page(struct page *page)
2000{
2001 /* 0 stands for page_is_file_cache(page) == false */
2002 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2003 unlock_page(page);
2004 putback_lru_page(page);
2005}
2006
2007static void release_pte_pages(pte_t *pte, pte_t *_pte)
2008{
2009 while (--_pte >= pte) {
2010 pte_t pteval = *_pte;
2011 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2012 release_pte_page(pte_page(pteval));
2013 }
2014}
2015
2016static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2017 unsigned long address,
2018 pte_t *pte)
2019{
2020 struct page *page = NULL;
2021 pte_t *_pte;
2022 int none_or_zero = 0, result = 0;
2023 bool referenced = false, writable = false;
2024
2025 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2026 _pte++, address += PAGE_SIZE) {
2027 pte_t pteval = *_pte;
2028 if (pte_none(pteval) || (pte_present(pteval) &&
2029 is_zero_pfn(pte_pfn(pteval)))) {
2030 if (!userfaultfd_armed(vma) &&
2031 ++none_or_zero <= khugepaged_max_ptes_none) {
2032 continue;
2033 } else {
2034 result = SCAN_EXCEED_NONE_PTE;
2035 goto out;
2036 }
2037 }
2038 if (!pte_present(pteval)) {
2039 result = SCAN_PTE_NON_PRESENT;
2040 goto out;
2041 }
2042 page = vm_normal_page(vma, address, pteval);
2043 if (unlikely(!page)) {
2044 result = SCAN_PAGE_NULL;
2045 goto out;
2046 }
2047
2048 VM_BUG_ON_PAGE(PageCompound(page), page);
2049 VM_BUG_ON_PAGE(!PageAnon(page), page);
2050 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2051
2052 /*
2053 * We can do it before isolate_lru_page because the
2054 * page can't be freed from under us. NOTE: PG_lock
2055 * is needed to serialize against split_huge_page
2056 * when invoked from the VM.
2057 */
2058 if (!trylock_page(page)) {
2059 result = SCAN_PAGE_LOCK;
2060 goto out;
2061 }
2062
2063 /*
2064 * cannot use mapcount: can't collapse if there's a gup pin.
2065 * The page must only be referenced by the scanned process
2066 * and page swap cache.
2067 */
2068 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2069 unlock_page(page);
2070 result = SCAN_PAGE_COUNT;
2071 goto out;
2072 }
2073 if (pte_write(pteval)) {
2074 writable = true;
2075 } else {
2076 if (PageSwapCache(page) &&
2077 !reuse_swap_page(page, NULL)) {
2078 unlock_page(page);
2079 result = SCAN_SWAP_CACHE_PAGE;
2080 goto out;
2081 }
2082 /*
2083 * Page is not in the swap cache. It can be collapsed
2084 * into a THP.
2085 */
2086 }
2087
2088 /*
2089 * Isolate the page to avoid collapsing an hugepage
2090 * currently in use by the VM.
2091 */
2092 if (isolate_lru_page(page)) {
2093 unlock_page(page);
2094 result = SCAN_DEL_PAGE_LRU;
2095 goto out;
2096 }
2097 /* 0 stands for page_is_file_cache(page) == false */
2098 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2099 VM_BUG_ON_PAGE(!PageLocked(page), page);
2100 VM_BUG_ON_PAGE(PageLRU(page), page);
2101
2102 /* If there is no mapped pte young don't collapse the page */
2103 if (pte_young(pteval) ||
2104 page_is_young(page) || PageReferenced(page) ||
2105 mmu_notifier_test_young(vma->vm_mm, address))
2106 referenced = true;
2107 }
2108 if (likely(writable)) {
2109 if (likely(referenced)) {
2110 result = SCAN_SUCCEED;
2111 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2112 referenced, writable, result);
2113 return 1;
2114 }
2115 } else {
2116 result = SCAN_PAGE_RO;
2117 }
2118
2119out:
2120 release_pte_pages(pte, _pte);
2121 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2122 referenced, writable, result);
2123 return 0;
2124}
2125
2126static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2127 struct vm_area_struct *vma,
2128 unsigned long address,
2129 spinlock_t *ptl)
2130{
2131 pte_t *_pte;
2132 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2133 pte_t pteval = *_pte;
2134 struct page *src_page;
2135
2136 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2137 clear_user_highpage(page, address);
2138 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2139 if (is_zero_pfn(pte_pfn(pteval))) {
2140 /*
2141 * ptl mostly unnecessary.
2142 */
2143 spin_lock(ptl);
2144 /*
2145 * paravirt calls inside pte_clear here are
2146 * superfluous.
2147 */
2148 pte_clear(vma->vm_mm, address, _pte);
2149 spin_unlock(ptl);
2150 }
2151 } else {
2152 src_page = pte_page(pteval);
2153 copy_user_highpage(page, src_page, address, vma);
2154 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2155 release_pte_page(src_page);
2156 /*
2157 * ptl mostly unnecessary, but preempt has to
2158 * be disabled to update the per-cpu stats
2159 * inside page_remove_rmap().
2160 */
2161 spin_lock(ptl);
2162 /*
2163 * paravirt calls inside pte_clear here are
2164 * superfluous.
2165 */
2166 pte_clear(vma->vm_mm, address, _pte);
2167 page_remove_rmap(src_page, false);
2168 spin_unlock(ptl);
2169 free_page_and_swap_cache(src_page);
2170 }
2171
2172 address += PAGE_SIZE;
2173 page++;
2174 }
2175}
2176
2177static void khugepaged_alloc_sleep(void)
2178{
2179 DEFINE_WAIT(wait);
2180
2181 add_wait_queue(&khugepaged_wait, &wait);
2182 freezable_schedule_timeout_interruptible(
2183 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2184 remove_wait_queue(&khugepaged_wait, &wait);
2185}
2186
2187static int khugepaged_node_load[MAX_NUMNODES];
2188
2189static bool khugepaged_scan_abort(int nid)
2190{
2191 int i;
2192
2193 /*
2194 * If zone_reclaim_mode is disabled, then no extra effort is made to
2195 * allocate memory locally.
2196 */
2197 if (!zone_reclaim_mode)
2198 return false;
2199
2200 /* If there is a count for this node already, it must be acceptable */
2201 if (khugepaged_node_load[nid])
2202 return false;
2203
2204 for (i = 0; i < MAX_NUMNODES; i++) {
2205 if (!khugepaged_node_load[i])
2206 continue;
2207 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2208 return true;
2209 }
2210 return false;
2211}
2212
2213#ifdef CONFIG_NUMA
2214static int khugepaged_find_target_node(void)
2215{
2216 static int last_khugepaged_target_node = NUMA_NO_NODE;
2217 int nid, target_node = 0, max_value = 0;
2218
2219 /* find first node with max normal pages hit */
2220 for (nid = 0; nid < MAX_NUMNODES; nid++)
2221 if (khugepaged_node_load[nid] > max_value) {
2222 max_value = khugepaged_node_load[nid];
2223 target_node = nid;
2224 }
2225
2226 /* do some balance if several nodes have the same hit record */
2227 if (target_node <= last_khugepaged_target_node)
2228 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2229 nid++)
2230 if (max_value == khugepaged_node_load[nid]) {
2231 target_node = nid;
2232 break;
2233 }
2234
2235 last_khugepaged_target_node = target_node;
2236 return target_node;
2237}
2238
2239static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2240{
2241 if (IS_ERR(*hpage)) {
2242 if (!*wait)
2243 return false;
2244
2245 *wait = false;
2246 *hpage = NULL;
2247 khugepaged_alloc_sleep();
2248 } else if (*hpage) {
2249 put_page(*hpage);
2250 *hpage = NULL;
2251 }
2252
2253 return true;
2254}
2255
2256static struct page *
2257khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2258 unsigned long address, int node)
2259{
2260 VM_BUG_ON_PAGE(*hpage, *hpage);
2261
2262 /*
2263 * Before allocating the hugepage, release the mmap_sem read lock.
2264 * The allocation can take potentially a long time if it involves
2265 * sync compaction, and we do not need to hold the mmap_sem during
2266 * that. We will recheck the vma after taking it again in write mode.
2267 */
2268 up_read(&mm->mmap_sem);
2269
2270 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2271 if (unlikely(!*hpage)) {
2272 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2273 *hpage = ERR_PTR(-ENOMEM);
2274 return NULL;
2275 }
2276
2277 prep_transhuge_page(*hpage);
2278 count_vm_event(THP_COLLAPSE_ALLOC);
2279 return *hpage;
2280}
2281#else
2282static int khugepaged_find_target_node(void)
2283{
2284 return 0;
2285}
2286
2287static inline struct page *alloc_khugepaged_hugepage(void)
2288{
2289 struct page *page;
2290
2291 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2292 HPAGE_PMD_ORDER);
2293 if (page)
2294 prep_transhuge_page(page);
2295 return page;
2296}
2297
2298static struct page *khugepaged_alloc_hugepage(bool *wait)
2299{
2300 struct page *hpage;
2301
2302 do {
2303 hpage = alloc_khugepaged_hugepage();
2304 if (!hpage) {
2305 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2306 if (!*wait)
2307 return NULL;
2308
2309 *wait = false;
2310 khugepaged_alloc_sleep();
2311 } else
2312 count_vm_event(THP_COLLAPSE_ALLOC);
2313 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2314
2315 return hpage;
2316}
2317
2318static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2319{
2320 if (!*hpage)
2321 *hpage = khugepaged_alloc_hugepage(wait);
2322
2323 if (unlikely(!*hpage))
2324 return false;
2325
2326 return true;
2327}
2328
2329static struct page *
2330khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2331 unsigned long address, int node)
2332{
2333 up_read(&mm->mmap_sem);
2334 VM_BUG_ON(!*hpage);
2335
2336 return *hpage;
2337}
2338#endif
2339
2340static bool hugepage_vma_check(struct vm_area_struct *vma)
2341{
2342 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2343 (vma->vm_flags & VM_NOHUGEPAGE))
2344 return false;
2345 if (!vma->anon_vma || vma->vm_ops)
2346 return false;
2347 if (is_vma_temporary_stack(vma))
2348 return false;
2349 return !(vma->vm_flags & VM_NO_THP);
2350}
2351
2352static void collapse_huge_page(struct mm_struct *mm,
2353 unsigned long address,
2354 struct page **hpage,
2355 struct vm_area_struct *vma,
2356 int node)
2357{
2358 pmd_t *pmd, _pmd;
2359 pte_t *pte;
2360 pgtable_t pgtable;
2361 struct page *new_page;
2362 spinlock_t *pmd_ptl, *pte_ptl;
2363 int isolated = 0, result = 0;
2364 unsigned long hstart, hend;
2365 struct mem_cgroup *memcg;
2366 unsigned long mmun_start; /* For mmu_notifiers */
2367 unsigned long mmun_end; /* For mmu_notifiers */
2368 gfp_t gfp;
2369
2370 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2371
2372 /* Only allocate from the target node */
2373 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
2374
2375 /* release the mmap_sem read lock. */
2376 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2377 if (!new_page) {
2378 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2379 goto out_nolock;
2380 }
2381
2382 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2383 result = SCAN_CGROUP_CHARGE_FAIL;
2384 goto out_nolock;
2385 }
2386
2387 /*
2388 * Prevent all access to pagetables with the exception of
2389 * gup_fast later hanlded by the ptep_clear_flush and the VM
2390 * handled by the anon_vma lock + PG_lock.
2391 */
2392 down_write(&mm->mmap_sem);
2393 if (unlikely(khugepaged_test_exit(mm))) {
2394 result = SCAN_ANY_PROCESS;
2395 goto out;
2396 }
2397
2398 vma = find_vma(mm, address);
2399 if (!vma) {
2400 result = SCAN_VMA_NULL;
2401 goto out;
2402 }
2403 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2404 hend = vma->vm_end & HPAGE_PMD_MASK;
2405 if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2406 result = SCAN_ADDRESS_RANGE;
2407 goto out;
2408 }
2409 if (!hugepage_vma_check(vma)) {
2410 result = SCAN_VMA_CHECK;
2411 goto out;
2412 }
2413 pmd = mm_find_pmd(mm, address);
2414 if (!pmd) {
2415 result = SCAN_PMD_NULL;
2416 goto out;
2417 }
2418
2419 anon_vma_lock_write(vma->anon_vma);
2420
2421 pte = pte_offset_map(pmd, address);
2422 pte_ptl = pte_lockptr(mm, pmd);
2423
2424 mmun_start = address;
2425 mmun_end = address + HPAGE_PMD_SIZE;
2426 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2427 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2428 /*
2429 * After this gup_fast can't run anymore. This also removes
2430 * any huge TLB entry from the CPU so we won't allow
2431 * huge and small TLB entries for the same virtual address
2432 * to avoid the risk of CPU bugs in that area.
2433 */
2434 _pmd = pmdp_collapse_flush(vma, address, pmd);
2435 spin_unlock(pmd_ptl);
2436 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2437
2438 spin_lock(pte_ptl);
2439 isolated = __collapse_huge_page_isolate(vma, address, pte);
2440 spin_unlock(pte_ptl);
2441
2442 if (unlikely(!isolated)) {
2443 pte_unmap(pte);
2444 spin_lock(pmd_ptl);
2445 BUG_ON(!pmd_none(*pmd));
2446 /*
2447 * We can only use set_pmd_at when establishing
2448 * hugepmds and never for establishing regular pmds that
2449 * points to regular pagetables. Use pmd_populate for that
2450 */
2451 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2452 spin_unlock(pmd_ptl);
2453 anon_vma_unlock_write(vma->anon_vma);
2454 result = SCAN_FAIL;
2455 goto out;
2456 }
2457
2458 /*
2459 * All pages are isolated and locked so anon_vma rmap
2460 * can't run anymore.
2461 */
2462 anon_vma_unlock_write(vma->anon_vma);
2463
2464 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2465 pte_unmap(pte);
2466 __SetPageUptodate(new_page);
2467 pgtable = pmd_pgtable(_pmd);
2468
2469 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2470 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2471
2472 /*
2473 * spin_lock() below is not the equivalent of smp_wmb(), so
2474 * this is needed to avoid the copy_huge_page writes to become
2475 * visible after the set_pmd_at() write.
2476 */
2477 smp_wmb();
2478
2479 spin_lock(pmd_ptl);
2480 BUG_ON(!pmd_none(*pmd));
2481 page_add_new_anon_rmap(new_page, vma, address, true);
2482 mem_cgroup_commit_charge(new_page, memcg, false, true);
2483 lru_cache_add_active_or_unevictable(new_page, vma);
2484 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2485 set_pmd_at(mm, address, pmd, _pmd);
2486 update_mmu_cache_pmd(vma, address, pmd);
2487 spin_unlock(pmd_ptl);
2488
2489 *hpage = NULL;
2490
2491 khugepaged_pages_collapsed++;
2492 result = SCAN_SUCCEED;
2493out_up_write:
2494 up_write(&mm->mmap_sem);
2495 trace_mm_collapse_huge_page(mm, isolated, result);
2496 return;
2497
2498out_nolock:
2499 trace_mm_collapse_huge_page(mm, isolated, result);
2500 return;
2501out:
2502 mem_cgroup_cancel_charge(new_page, memcg, true);
2503 goto out_up_write;
2504}
2505
2506static int khugepaged_scan_pmd(struct mm_struct *mm,
2507 struct vm_area_struct *vma,
2508 unsigned long address,
2509 struct page **hpage)
2510{
2511 pmd_t *pmd;
2512 pte_t *pte, *_pte;
2513 int ret = 0, none_or_zero = 0, result = 0;
2514 struct page *page = NULL;
2515 unsigned long _address;
2516 spinlock_t *ptl;
2517 int node = NUMA_NO_NODE;
2518 bool writable = false, referenced = false;
2519
2520 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2521
2522 pmd = mm_find_pmd(mm, address);
2523 if (!pmd) {
2524 result = SCAN_PMD_NULL;
2525 goto out;
2526 }
2527
2528 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2529 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2530 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2531 _pte++, _address += PAGE_SIZE) {
2532 pte_t pteval = *_pte;
2533 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2534 if (!userfaultfd_armed(vma) &&
2535 ++none_or_zero <= khugepaged_max_ptes_none) {
2536 continue;
2537 } else {
2538 result = SCAN_EXCEED_NONE_PTE;
2539 goto out_unmap;
2540 }
2541 }
2542 if (!pte_present(pteval)) {
2543 result = SCAN_PTE_NON_PRESENT;
2544 goto out_unmap;
2545 }
2546 if (pte_write(pteval))
2547 writable = true;
2548
2549 page = vm_normal_page(vma, _address, pteval);
2550 if (unlikely(!page)) {
2551 result = SCAN_PAGE_NULL;
2552 goto out_unmap;
2553 }
2554
2555 /* TODO: teach khugepaged to collapse THP mapped with pte */
2556 if (PageCompound(page)) {
2557 result = SCAN_PAGE_COMPOUND;
2558 goto out_unmap;
2559 }
2560
2561 /*
2562 * Record which node the original page is from and save this
2563 * information to khugepaged_node_load[].
2564 * Khupaged will allocate hugepage from the node has the max
2565 * hit record.
2566 */
2567 node = page_to_nid(page);
2568 if (khugepaged_scan_abort(node)) {
2569 result = SCAN_SCAN_ABORT;
2570 goto out_unmap;
2571 }
2572 khugepaged_node_load[node]++;
2573 if (!PageLRU(page)) {
2574 result = SCAN_PAGE_LRU;
2575 goto out_unmap;
2576 }
2577 if (PageLocked(page)) {
2578 result = SCAN_PAGE_LOCK;
2579 goto out_unmap;
2580 }
2581 if (!PageAnon(page)) {
2582 result = SCAN_PAGE_ANON;
2583 goto out_unmap;
2584 }
2585
2586 /*
2587 * cannot use mapcount: can't collapse if there's a gup pin.
2588 * The page must only be referenced by the scanned process
2589 * and page swap cache.
2590 */
2591 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2592 result = SCAN_PAGE_COUNT;
2593 goto out_unmap;
2594 }
2595 if (pte_young(pteval) ||
2596 page_is_young(page) || PageReferenced(page) ||
2597 mmu_notifier_test_young(vma->vm_mm, address))
2598 referenced = true;
2599 }
2600 if (writable) {
2601 if (referenced) {
2602 result = SCAN_SUCCEED;
2603 ret = 1;
2604 } else {
2605 result = SCAN_NO_REFERENCED_PAGE;
2606 }
2607 } else {
2608 result = SCAN_PAGE_RO;
2609 }
2610out_unmap:
2611 pte_unmap_unlock(pte, ptl);
2612 if (ret) {
2613 node = khugepaged_find_target_node();
2614 /* collapse_huge_page will return with the mmap_sem released */
2615 collapse_huge_page(mm, address, hpage, vma, node);
2616 }
2617out:
2618 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2619 none_or_zero, result);
2620 return ret;
2621}
2622
2623static void collect_mm_slot(struct mm_slot *mm_slot)
2624{
2625 struct mm_struct *mm = mm_slot->mm;
2626
2627 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2628
2629 if (khugepaged_test_exit(mm)) {
2630 /* free mm_slot */
2631 hash_del(&mm_slot->hash);
2632 list_del(&mm_slot->mm_node);
2633
2634 /*
2635 * Not strictly needed because the mm exited already.
2636 *
2637 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2638 */
2639
2640 /* khugepaged_mm_lock actually not necessary for the below */
2641 free_mm_slot(mm_slot);
2642 mmdrop(mm);
2643 }
2644}
2645
2646static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2647 struct page **hpage)
2648 __releases(&khugepaged_mm_lock)
2649 __acquires(&khugepaged_mm_lock)
2650{
2651 struct mm_slot *mm_slot;
2652 struct mm_struct *mm;
2653 struct vm_area_struct *vma;
2654 int progress = 0;
2655
2656 VM_BUG_ON(!pages);
2657 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2658
2659 if (khugepaged_scan.mm_slot)
2660 mm_slot = khugepaged_scan.mm_slot;
2661 else {
2662 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2663 struct mm_slot, mm_node);
2664 khugepaged_scan.address = 0;
2665 khugepaged_scan.mm_slot = mm_slot;
2666 }
2667 spin_unlock(&khugepaged_mm_lock);
2668
2669 mm = mm_slot->mm;
2670 down_read(&mm->mmap_sem);
2671 if (unlikely(khugepaged_test_exit(mm)))
2672 vma = NULL;
2673 else
2674 vma = find_vma(mm, khugepaged_scan.address);
2675
2676 progress++;
2677 for (; vma; vma = vma->vm_next) {
2678 unsigned long hstart, hend;
2679
2680 cond_resched();
2681 if (unlikely(khugepaged_test_exit(mm))) {
2682 progress++;
2683 break;
2684 }
2685 if (!hugepage_vma_check(vma)) {
2686skip:
2687 progress++;
2688 continue;
2689 }
2690 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2691 hend = vma->vm_end & HPAGE_PMD_MASK;
2692 if (hstart >= hend)
2693 goto skip;
2694 if (khugepaged_scan.address > hend)
2695 goto skip;
2696 if (khugepaged_scan.address < hstart)
2697 khugepaged_scan.address = hstart;
2698 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2699
2700 while (khugepaged_scan.address < hend) {
2701 int ret;
2702 cond_resched();
2703 if (unlikely(khugepaged_test_exit(mm)))
2704 goto breakouterloop;
2705
2706 VM_BUG_ON(khugepaged_scan.address < hstart ||
2707 khugepaged_scan.address + HPAGE_PMD_SIZE >
2708 hend);
2709 ret = khugepaged_scan_pmd(mm, vma,
2710 khugepaged_scan.address,
2711 hpage);
2712 /* move to next address */
2713 khugepaged_scan.address += HPAGE_PMD_SIZE;
2714 progress += HPAGE_PMD_NR;
2715 if (ret)
2716 /* we released mmap_sem so break loop */
2717 goto breakouterloop_mmap_sem;
2718 if (progress >= pages)
2719 goto breakouterloop;
2720 }
2721 }
2722breakouterloop:
2723 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2724breakouterloop_mmap_sem:
2725
2726 spin_lock(&khugepaged_mm_lock);
2727 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2728 /*
2729 * Release the current mm_slot if this mm is about to die, or
2730 * if we scanned all vmas of this mm.
2731 */
2732 if (khugepaged_test_exit(mm) || !vma) {
2733 /*
2734 * Make sure that if mm_users is reaching zero while
2735 * khugepaged runs here, khugepaged_exit will find
2736 * mm_slot not pointing to the exiting mm.
2737 */
2738 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2739 khugepaged_scan.mm_slot = list_entry(
2740 mm_slot->mm_node.next,
2741 struct mm_slot, mm_node);
2742 khugepaged_scan.address = 0;
2743 } else {
2744 khugepaged_scan.mm_slot = NULL;
2745 khugepaged_full_scans++;
2746 }
2747
2748 collect_mm_slot(mm_slot);
2749 }
2750
2751 return progress;
2752}
2753
2754static int khugepaged_has_work(void)
2755{
2756 return !list_empty(&khugepaged_scan.mm_head) &&
2757 khugepaged_enabled();
2758}
2759
2760static int khugepaged_wait_event(void)
2761{
2762 return !list_empty(&khugepaged_scan.mm_head) ||
2763 kthread_should_stop();
2764}
2765
2766static void khugepaged_do_scan(void)
2767{
2768 struct page *hpage = NULL;
2769 unsigned int progress = 0, pass_through_head = 0;
2770 unsigned int pages = khugepaged_pages_to_scan;
2771 bool wait = true;
2772
2773 barrier(); /* write khugepaged_pages_to_scan to local stack */
2774
2775 while (progress < pages) {
2776 if (!khugepaged_prealloc_page(&hpage, &wait))
2777 break;
2778
2779 cond_resched();
2780
2781 if (unlikely(kthread_should_stop() || try_to_freeze()))
2782 break;
2783
2784 spin_lock(&khugepaged_mm_lock);
2785 if (!khugepaged_scan.mm_slot)
2786 pass_through_head++;
2787 if (khugepaged_has_work() &&
2788 pass_through_head < 2)
2789 progress += khugepaged_scan_mm_slot(pages - progress,
2790 &hpage);
2791 else
2792 progress = pages;
2793 spin_unlock(&khugepaged_mm_lock);
2794 }
2795
2796 if (!IS_ERR_OR_NULL(hpage))
2797 put_page(hpage);
2798}
2799
2800static void khugepaged_wait_work(void)
2801{
2802 if (khugepaged_has_work()) {
2803 if (!khugepaged_scan_sleep_millisecs)
2804 return;
2805
2806 wait_event_freezable_timeout(khugepaged_wait,
2807 kthread_should_stop(),
2808 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2809 return;
2810 }
2811
2812 if (khugepaged_enabled())
2813 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2814}
2815
2816static int khugepaged(void *none)
2817{
2818 struct mm_slot *mm_slot;
2819
2820 set_freezable();
2821 set_user_nice(current, MAX_NICE);
2822
2823 while (!kthread_should_stop()) {
2824 khugepaged_do_scan();
2825 khugepaged_wait_work();
2826 }
2827
2828 spin_lock(&khugepaged_mm_lock);
2829 mm_slot = khugepaged_scan.mm_slot;
2830 khugepaged_scan.mm_slot = NULL;
2831 if (mm_slot)
2832 collect_mm_slot(mm_slot);
2833 spin_unlock(&khugepaged_mm_lock);
2834 return 0;
2835}
2836
2837static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2838 unsigned long haddr, pmd_t *pmd)
2839{
2840 struct mm_struct *mm = vma->vm_mm;
2841 pgtable_t pgtable;
2842 pmd_t _pmd;
2843 int i;
2844
2845 /* leave pmd empty until pte is filled */
2846 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2847
2848 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2849 pmd_populate(mm, &_pmd, pgtable);
2850
2851 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2852 pte_t *pte, entry;
2853 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2854 entry = pte_mkspecial(entry);
2855 pte = pte_offset_map(&_pmd, haddr);
2856 VM_BUG_ON(!pte_none(*pte));
2857 set_pte_at(mm, haddr, pte, entry);
2858 pte_unmap(pte);
2859 }
2860 smp_wmb(); /* make pte visible before pmd */
2861 pmd_populate(mm, pmd, pgtable);
2862 put_huge_zero_page();
2863}
2864
2865static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2866 unsigned long haddr, bool freeze)
2867{
2868 struct mm_struct *mm = vma->vm_mm;
2869 struct page *page;
2870 pgtable_t pgtable;
2871 pmd_t _pmd;
2872 bool young, write, dirty;
2873 unsigned long addr;
2874 int i;
2875
2876 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2877 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2878 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2879 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2880
2881 count_vm_event(THP_SPLIT_PMD);
2882
2883 if (vma_is_dax(vma)) {
2884 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2885 if (is_huge_zero_pmd(_pmd))
2886 put_huge_zero_page();
2887 return;
2888 } else if (is_huge_zero_pmd(*pmd)) {
2889 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2890 }
2891
2892 page = pmd_page(*pmd);
2893 VM_BUG_ON_PAGE(!page_count(page), page);
2894 page_ref_add(page, HPAGE_PMD_NR - 1);
2895 write = pmd_write(*pmd);
2896 young = pmd_young(*pmd);
2897 dirty = pmd_dirty(*pmd);
2898
2899 pmdp_huge_split_prepare(vma, haddr, pmd);
2900 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2901 pmd_populate(mm, &_pmd, pgtable);
2902
2903 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2904 pte_t entry, *pte;
2905 /*
2906 * Note that NUMA hinting access restrictions are not
2907 * transferred to avoid any possibility of altering
2908 * permissions across VMAs.
2909 */
2910 if (freeze) {
2911 swp_entry_t swp_entry;
2912 swp_entry = make_migration_entry(page + i, write);
2913 entry = swp_entry_to_pte(swp_entry);
2914 } else {
2915 entry = mk_pte(page + i, vma->vm_page_prot);
2916 entry = maybe_mkwrite(entry, vma);
2917 if (!write)
2918 entry = pte_wrprotect(entry);
2919 if (!young)
2920 entry = pte_mkold(entry);
2921 }
2922 if (dirty)
2923 SetPageDirty(page + i);
2924 pte = pte_offset_map(&_pmd, addr);
2925 BUG_ON(!pte_none(*pte));
2926 set_pte_at(mm, addr, pte, entry);
2927 atomic_inc(&page[i]._mapcount);
2928 pte_unmap(pte);
2929 }
2930
2931 /*
2932 * Set PG_double_map before dropping compound_mapcount to avoid
2933 * false-negative page_mapped().
2934 */
2935 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2936 for (i = 0; i < HPAGE_PMD_NR; i++)
2937 atomic_inc(&page[i]._mapcount);
2938 }
2939
2940 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2941 /* Last compound_mapcount is gone. */
2942 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2943 if (TestClearPageDoubleMap(page)) {
2944 /* No need in mapcount reference anymore */
2945 for (i = 0; i < HPAGE_PMD_NR; i++)
2946 atomic_dec(&page[i]._mapcount);
2947 }
2948 }
2949
2950 smp_wmb(); /* make pte visible before pmd */
2951 /*
2952 * Up to this point the pmd is present and huge and userland has the
2953 * whole access to the hugepage during the split (which happens in
2954 * place). If we overwrite the pmd with the not-huge version pointing
2955 * to the pte here (which of course we could if all CPUs were bug
2956 * free), userland could trigger a small page size TLB miss on the
2957 * small sized TLB while the hugepage TLB entry is still established in
2958 * the huge TLB. Some CPU doesn't like that.
2959 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2960 * 383 on page 93. Intel should be safe but is also warns that it's
2961 * only safe if the permission and cache attributes of the two entries
2962 * loaded in the two TLB is identical (which should be the case here).
2963 * But it is generally safer to never allow small and huge TLB entries
2964 * for the same virtual address to be loaded simultaneously. So instead
2965 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2966 * current pmd notpresent (atomically because here the pmd_trans_huge
2967 * and pmd_trans_splitting must remain set at all times on the pmd
2968 * until the split is complete for this pmd), then we flush the SMP TLB
2969 * and finally we write the non-huge version of the pmd entry with
2970 * pmd_populate.
2971 */
2972 pmdp_invalidate(vma, haddr, pmd);
2973 pmd_populate(mm, pmd, pgtable);
2974
2975 if (freeze) {
2976 for (i = 0; i < HPAGE_PMD_NR; i++) {
2977 page_remove_rmap(page + i, false);
2978 put_page(page + i);
2979 }
2980 }
2981}
2982
2983void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2984 unsigned long address, bool freeze)
2985{
2986 spinlock_t *ptl;
2987 struct mm_struct *mm = vma->vm_mm;
2988 unsigned long haddr = address & HPAGE_PMD_MASK;
2989
2990 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2991 ptl = pmd_lock(mm, pmd);
2992 if (pmd_trans_huge(*pmd)) {
2993 struct page *page = pmd_page(*pmd);
2994 if (PageMlocked(page))
2995 clear_page_mlock(page);
2996 } else if (!pmd_devmap(*pmd))
2997 goto out;
2998 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2999out:
3000 spin_unlock(ptl);
3001 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
3002}
3003
3004void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
3005 bool freeze, struct page *page)
3006{
3007 pgd_t *pgd;
3008 pud_t *pud;
3009 pmd_t *pmd;
3010
3011 pgd = pgd_offset(vma->vm_mm, address);
3012 if (!pgd_present(*pgd))
3013 return;
3014
3015 pud = pud_offset(pgd, address);
3016 if (!pud_present(*pud))
3017 return;
3018
3019 pmd = pmd_offset(pud, address);
3020 if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
3021 return;
3022
3023 /*
3024 * If caller asks to setup a migration entries, we need a page to check
3025 * pmd against. Otherwise we can end up replacing wrong page.
3026 */
3027 VM_BUG_ON(freeze && !page);
3028 if (page && page != pmd_page(*pmd))
3029 return;
3030
3031 /*
3032 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3033 * materialize from under us.
3034 */
3035 __split_huge_pmd(vma, pmd, address, freeze);
3036}
3037
3038void vma_adjust_trans_huge(struct vm_area_struct *vma,
3039 unsigned long start,
3040 unsigned long end,
3041 long adjust_next)
3042{
3043 /*
3044 * If the new start address isn't hpage aligned and it could
3045 * previously contain an hugepage: check if we need to split
3046 * an huge pmd.
3047 */
3048 if (start & ~HPAGE_PMD_MASK &&
3049 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3050 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3051 split_huge_pmd_address(vma, start, false, NULL);
3052
3053 /*
3054 * If the new end address isn't hpage aligned and it could
3055 * previously contain an hugepage: check if we need to split
3056 * an huge pmd.
3057 */
3058 if (end & ~HPAGE_PMD_MASK &&
3059 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3060 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3061 split_huge_pmd_address(vma, end, false, NULL);
3062
3063 /*
3064 * If we're also updating the vma->vm_next->vm_start, if the new
3065 * vm_next->vm_start isn't page aligned and it could previously
3066 * contain an hugepage: check if we need to split an huge pmd.
3067 */
3068 if (adjust_next > 0) {
3069 struct vm_area_struct *next = vma->vm_next;
3070 unsigned long nstart = next->vm_start;
3071 nstart += adjust_next << PAGE_SHIFT;
3072 if (nstart & ~HPAGE_PMD_MASK &&
3073 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3074 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3075 split_huge_pmd_address(next, nstart, false, NULL);
3076 }
3077}
3078
3079static void freeze_page(struct page *page)
3080{
3081 enum ttu_flags ttu_flags = TTU_MIGRATION | TTU_IGNORE_MLOCK |
3082 TTU_IGNORE_ACCESS | TTU_RMAP_LOCKED;
3083 int i, ret;
3084
3085 VM_BUG_ON_PAGE(!PageHead(page), page);
3086
3087 /* We only need TTU_SPLIT_HUGE_PMD once */
3088 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
3089 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
3090 /* Cut short if the page is unmapped */
3091 if (page_count(page) == 1)
3092 return;
3093
3094 ret = try_to_unmap(page + i, ttu_flags);
3095 }
3096 VM_BUG_ON(ret);
3097}
3098
3099static void unfreeze_page(struct page *page)
3100{
3101 int i;
3102
3103 for (i = 0; i < HPAGE_PMD_NR; i++)
3104 remove_migration_ptes(page + i, page + i, true);
3105}
3106
3107static void __split_huge_page_tail(struct page *head, int tail,
3108 struct lruvec *lruvec, struct list_head *list)
3109{
3110 struct page *page_tail = head + tail;
3111
3112 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
3113 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
3114
3115 /*
3116 * tail_page->_count is zero and not changing from under us. But
3117 * get_page_unless_zero() may be running from under us on the
3118 * tail_page. If we used atomic_set() below instead of atomic_inc(), we
3119 * would then run atomic_set() concurrently with
3120 * get_page_unless_zero(), and atomic_set() is implemented in C not
3121 * using locked ops. spin_unlock on x86 sometime uses locked ops
3122 * because of PPro errata 66, 92, so unless somebody can guarantee
3123 * atomic_set() here would be safe on all archs (and not only on x86),
3124 * it's safer to use atomic_inc().
3125 */
3126 page_ref_inc(page_tail);
3127
3128 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3129 page_tail->flags |= (head->flags &
3130 ((1L << PG_referenced) |
3131 (1L << PG_swapbacked) |
3132 (1L << PG_mlocked) |
3133 (1L << PG_uptodate) |
3134 (1L << PG_active) |
3135 (1L << PG_locked) |
3136 (1L << PG_unevictable) |
3137 (1L << PG_dirty)));
3138
3139 /*
3140 * After clearing PageTail the gup refcount can be released.
3141 * Page flags also must be visible before we make the page non-compound.
3142 */
3143 smp_wmb();
3144
3145 clear_compound_head(page_tail);
3146
3147 if (page_is_young(head))
3148 set_page_young(page_tail);
3149 if (page_is_idle(head))
3150 set_page_idle(page_tail);
3151
3152 /* ->mapping in first tail page is compound_mapcount */
3153 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3154 page_tail);
3155 page_tail->mapping = head->mapping;
3156
3157 page_tail->index = head->index + tail;
3158 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3159 lru_add_page_tail(head, page_tail, lruvec, list);
3160}
3161
3162static void __split_huge_page(struct page *page, struct list_head *list)
3163{
3164 struct page *head = compound_head(page);
3165 struct zone *zone = page_zone(head);
3166 struct lruvec *lruvec;
3167 int i;
3168
3169 /* prevent PageLRU to go away from under us, and freeze lru stats */
3170 spin_lock_irq(&zone->lru_lock);
3171 lruvec = mem_cgroup_page_lruvec(head, zone);
3172
3173 /* complete memcg works before add pages to LRU */
3174 mem_cgroup_split_huge_fixup(head);
3175
3176 for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3177 __split_huge_page_tail(head, i, lruvec, list);
3178
3179 ClearPageCompound(head);
3180 spin_unlock_irq(&zone->lru_lock);
3181
3182 unfreeze_page(head);
3183
3184 for (i = 0; i < HPAGE_PMD_NR; i++) {
3185 struct page *subpage = head + i;
3186 if (subpage == page)
3187 continue;
3188 unlock_page(subpage);
3189
3190 /*
3191 * Subpages may be freed if there wasn't any mapping
3192 * like if add_to_swap() is running on a lru page that
3193 * had its mapping zapped. And freeing these pages
3194 * requires taking the lru_lock so we do the put_page
3195 * of the tail pages after the split is complete.
3196 */
3197 put_page(subpage);
3198 }
3199}
3200
3201int total_mapcount(struct page *page)
3202{
3203 int i, ret;
3204
3205 VM_BUG_ON_PAGE(PageTail(page), page);
3206
3207 if (likely(!PageCompound(page)))
3208 return atomic_read(&page->_mapcount) + 1;
3209
3210 ret = compound_mapcount(page);
3211 if (PageHuge(page))
3212 return ret;
3213 for (i = 0; i < HPAGE_PMD_NR; i++)
3214 ret += atomic_read(&page[i]._mapcount) + 1;
3215 if (PageDoubleMap(page))
3216 ret -= HPAGE_PMD_NR;
3217 return ret;
3218}
3219
3220/*
3221 * This calculates accurately how many mappings a transparent hugepage
3222 * has (unlike page_mapcount() which isn't fully accurate). This full
3223 * accuracy is primarily needed to know if copy-on-write faults can
3224 * reuse the page and change the mapping to read-write instead of
3225 * copying them. At the same time this returns the total_mapcount too.
3226 *
3227 * The function returns the highest mapcount any one of the subpages
3228 * has. If the return value is one, even if different processes are
3229 * mapping different subpages of the transparent hugepage, they can
3230 * all reuse it, because each process is reusing a different subpage.
3231 *
3232 * The total_mapcount is instead counting all virtual mappings of the
3233 * subpages. If the total_mapcount is equal to "one", it tells the
3234 * caller all mappings belong to the same "mm" and in turn the
3235 * anon_vma of the transparent hugepage can become the vma->anon_vma
3236 * local one as no other process may be mapping any of the subpages.
3237 *
3238 * It would be more accurate to replace page_mapcount() with
3239 * page_trans_huge_mapcount(), however we only use
3240 * page_trans_huge_mapcount() in the copy-on-write faults where we
3241 * need full accuracy to avoid breaking page pinning, because
3242 * page_trans_huge_mapcount() is slower than page_mapcount().
3243 */
3244int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
3245{
3246 int i, ret, _total_mapcount, mapcount;
3247
3248 /* hugetlbfs shouldn't call it */
3249 VM_BUG_ON_PAGE(PageHuge(page), page);
3250
3251 if (likely(!PageTransCompound(page))) {
3252 mapcount = atomic_read(&page->_mapcount) + 1;
3253 if (total_mapcount)
3254 *total_mapcount = mapcount;
3255 return mapcount;
3256 }
3257
3258 page = compound_head(page);
3259
3260 _total_mapcount = ret = 0;
3261 for (i = 0; i < HPAGE_PMD_NR; i++) {
3262 mapcount = atomic_read(&page[i]._mapcount) + 1;
3263 ret = max(ret, mapcount);
3264 _total_mapcount += mapcount;
3265 }
3266 if (PageDoubleMap(page)) {
3267 ret -= 1;
3268 _total_mapcount -= HPAGE_PMD_NR;
3269 }
3270 mapcount = compound_mapcount(page);
3271 ret += mapcount;
3272 _total_mapcount += mapcount;
3273 if (total_mapcount)
3274 *total_mapcount = _total_mapcount;
3275 return ret;
3276}
3277
3278/*
3279 * This function splits huge page into normal pages. @page can point to any
3280 * subpage of huge page to split. Split doesn't change the position of @page.
3281 *
3282 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3283 * The huge page must be locked.
3284 *
3285 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3286 *
3287 * Both head page and tail pages will inherit mapping, flags, and so on from
3288 * the hugepage.
3289 *
3290 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3291 * they are not mapped.
3292 *
3293 * Returns 0 if the hugepage is split successfully.
3294 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3295 * us.
3296 */
3297int split_huge_page_to_list(struct page *page, struct list_head *list)
3298{
3299 struct page *head = compound_head(page);
3300 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3301 struct anon_vma *anon_vma;
3302 int count, mapcount, ret;
3303 bool mlocked;
3304 unsigned long flags;
3305
3306 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3307 VM_BUG_ON_PAGE(!PageAnon(page), page);
3308 VM_BUG_ON_PAGE(!PageLocked(page), page);
3309 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3310 VM_BUG_ON_PAGE(!PageCompound(page), page);
3311
3312 /*
3313 * The caller does not necessarily hold an mmap_sem that would prevent
3314 * the anon_vma disappearing so we first we take a reference to it
3315 * and then lock the anon_vma for write. This is similar to
3316 * page_lock_anon_vma_read except the write lock is taken to serialise
3317 * against parallel split or collapse operations.
3318 */
3319 anon_vma = page_get_anon_vma(head);
3320 if (!anon_vma) {
3321 ret = -EBUSY;
3322 goto out;
3323 }
3324 anon_vma_lock_write(anon_vma);
3325
3326 /*
3327 * Racy check if we can split the page, before freeze_page() will
3328 * split PMDs
3329 */
3330 if (total_mapcount(head) != page_count(head) - 1) {
3331 ret = -EBUSY;
3332 goto out_unlock;
3333 }
3334
3335 mlocked = PageMlocked(page);
3336 freeze_page(head);
3337 VM_BUG_ON_PAGE(compound_mapcount(head), head);
3338
3339 /* Make sure the page is not on per-CPU pagevec as it takes pin */
3340 if (mlocked)
3341 lru_add_drain();
3342
3343 /* Prevent deferred_split_scan() touching ->_count */
3344 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3345 count = page_count(head);
3346 mapcount = total_mapcount(head);
3347 if (!mapcount && count == 1) {
3348 if (!list_empty(page_deferred_list(head))) {
3349 pgdata->split_queue_len--;
3350 list_del(page_deferred_list(head));
3351 }
3352 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3353 __split_huge_page(page, list);
3354 ret = 0;
3355 } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3356 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3357 pr_alert("total_mapcount: %u, page_count(): %u\n",
3358 mapcount, count);
3359 if (PageTail(page))
3360 dump_page(head, NULL);
3361 dump_page(page, "total_mapcount(head) > 0");
3362 BUG();
3363 } else {
3364 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3365 unfreeze_page(head);
3366 ret = -EBUSY;
3367 }
3368
3369out_unlock:
3370 anon_vma_unlock_write(anon_vma);
3371 put_anon_vma(anon_vma);
3372out:
3373 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3374 return ret;
3375}
3376
3377void free_transhuge_page(struct page *page)
3378{
3379 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3380 unsigned long flags;
3381
3382 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3383 if (!list_empty(page_deferred_list(page))) {
3384 pgdata->split_queue_len--;
3385 list_del(page_deferred_list(page));
3386 }
3387 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3388 free_compound_page(page);
3389}
3390
3391void deferred_split_huge_page(struct page *page)
3392{
3393 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3394 unsigned long flags;
3395
3396 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3397
3398 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3399 if (list_empty(page_deferred_list(page))) {
3400 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3401 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3402 pgdata->split_queue_len++;
3403 }
3404 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3405}
3406
3407static unsigned long deferred_split_count(struct shrinker *shrink,
3408 struct shrink_control *sc)
3409{
3410 struct pglist_data *pgdata = NODE_DATA(sc->nid);
3411 return ACCESS_ONCE(pgdata->split_queue_len);
3412}
3413
3414static unsigned long deferred_split_scan(struct shrinker *shrink,
3415 struct shrink_control *sc)
3416{
3417 struct pglist_data *pgdata = NODE_DATA(sc->nid);
3418 unsigned long flags;
3419 LIST_HEAD(list), *pos, *next;
3420 struct page *page;
3421 int split = 0;
3422
3423 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3424 /* Take pin on all head pages to avoid freeing them under us */
3425 list_for_each_safe(pos, next, &pgdata->split_queue) {
3426 page = list_entry((void *)pos, struct page, mapping);
3427 page = compound_head(page);
3428 if (get_page_unless_zero(page)) {
3429 list_move(page_deferred_list(page), &list);
3430 } else {
3431 /* We lost race with put_compound_page() */
3432 list_del_init(page_deferred_list(page));
3433 pgdata->split_queue_len--;
3434 }
3435 if (!--sc->nr_to_scan)
3436 break;
3437 }
3438 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3439
3440 list_for_each_safe(pos, next, &list) {
3441 page = list_entry((void *)pos, struct page, mapping);
3442 lock_page(page);
3443 /* split_huge_page() removes page from list on success */
3444 if (!split_huge_page(page))
3445 split++;
3446 unlock_page(page);
3447 put_page(page);
3448 }
3449
3450 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3451 list_splice_tail(&list, &pgdata->split_queue);
3452 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3453
3454 /*
3455 * Stop shrinker if we didn't split any page, but the queue is empty.
3456 * This can happen if pages were freed under us.
3457 */
3458 if (!split && list_empty(&pgdata->split_queue))
3459 return SHRINK_STOP;
3460 return split;
3461}
3462
3463static struct shrinker deferred_split_shrinker = {
3464 .count_objects = deferred_split_count,
3465 .scan_objects = deferred_split_scan,
3466 .seeks = DEFAULT_SEEKS,
3467 .flags = SHRINKER_NUMA_AWARE,
3468};
3469
3470#ifdef CONFIG_DEBUG_FS
3471static int split_huge_pages_set(void *data, u64 val)
3472{
3473 struct zone *zone;
3474 struct page *page;
3475 unsigned long pfn, max_zone_pfn;
3476 unsigned long total = 0, split = 0;
3477
3478 if (val != 1)
3479 return -EINVAL;
3480
3481 for_each_populated_zone(zone) {
3482 max_zone_pfn = zone_end_pfn(zone);
3483 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3484 if (!pfn_valid(pfn))
3485 continue;
3486
3487 page = pfn_to_page(pfn);
3488 if (!get_page_unless_zero(page))
3489 continue;
3490
3491 if (zone != page_zone(page))
3492 goto next;
3493
3494 if (!PageHead(page) || !PageAnon(page) ||
3495 PageHuge(page))
3496 goto next;
3497
3498 total++;
3499 lock_page(page);
3500 if (!split_huge_page(page))
3501 split++;
3502 unlock_page(page);
3503next:
3504 put_page(page);
3505 }
3506 }
3507
3508 pr_info("%lu of %lu THP split\n", split, total);
3509
3510 return 0;
3511}
3512DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3513 "%llu\n");
3514
3515static int __init split_huge_pages_debugfs(void)
3516{
3517 void *ret;
3518
3519 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3520 &split_huge_pages_fops);
3521 if (!ret)
3522 pr_warn("Failed to create split_huge_pages in debugfs");
3523 return 0;
3524}
3525late_initcall(split_huge_pages_debugfs);
3526#endif