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#include <linux/mm.h>
9#include <linux/sched.h>
10#include <linux/highmem.h>
11#include <linux/hugetlb.h>
12#include <linux/mmu_notifier.h>
13#include <linux/rmap.h>
14#include <linux/swap.h>
15#include <linux/shrinker.h>
16#include <linux/mm_inline.h>
17#include <linux/kthread.h>
18#include <linux/khugepaged.h>
19#include <linux/freezer.h>
20#include <linux/mman.h>
21#include <linux/pagemap.h>
22#include <linux/migrate.h>
23#include <linux/hashtable.h>
24
25#include <asm/tlb.h>
26#include <asm/pgalloc.h>
27#include "internal.h"
28
29/*
30 * By default transparent hugepage support is disabled in order that avoid
31 * to risk increase the memory footprint of applications without a guaranteed
32 * benefit. When transparent hugepage support is enabled, is for all mappings,
33 * and khugepaged scans all mappings.
34 * Defrag is invoked by khugepaged hugepage allocations and by page faults
35 * for all hugepage allocations.
36 */
37unsigned long transparent_hugepage_flags __read_mostly =
38#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
39 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
40#endif
41#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
42 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
43#endif
44 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
45 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
46 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
47
48/* default scan 8*512 pte (or vmas) every 30 second */
49static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
50static unsigned int khugepaged_pages_collapsed;
51static unsigned int khugepaged_full_scans;
52static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
53/* during fragmentation poll the hugepage allocator once every minute */
54static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
55static struct task_struct *khugepaged_thread __read_mostly;
56static DEFINE_MUTEX(khugepaged_mutex);
57static DEFINE_SPINLOCK(khugepaged_mm_lock);
58static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
59/*
60 * default collapse hugepages if there is at least one pte mapped like
61 * it would have happened if the vma was large enough during page
62 * fault.
63 */
64static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
65
66static int khugepaged(void *none);
67static int khugepaged_slab_init(void);
68
69#define MM_SLOTS_HASH_BITS 10
70static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
71
72static struct kmem_cache *mm_slot_cache __read_mostly;
73
74/**
75 * struct mm_slot - hash lookup from mm to mm_slot
76 * @hash: hash collision list
77 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
78 * @mm: the mm that this information is valid for
79 */
80struct mm_slot {
81 struct hlist_node hash;
82 struct list_head mm_node;
83 struct mm_struct *mm;
84};
85
86/**
87 * struct khugepaged_scan - cursor for scanning
88 * @mm_head: the head of the mm list to scan
89 * @mm_slot: the current mm_slot we are scanning
90 * @address: the next address inside that to be scanned
91 *
92 * There is only the one khugepaged_scan instance of this cursor structure.
93 */
94struct khugepaged_scan {
95 struct list_head mm_head;
96 struct mm_slot *mm_slot;
97 unsigned long address;
98};
99static struct khugepaged_scan khugepaged_scan = {
100 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
101};
102
103
104static int set_recommended_min_free_kbytes(void)
105{
106 struct zone *zone;
107 int nr_zones = 0;
108 unsigned long recommended_min;
109
110 if (!khugepaged_enabled())
111 return 0;
112
113 for_each_populated_zone(zone)
114 nr_zones++;
115
116 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
117 recommended_min = pageblock_nr_pages * nr_zones * 2;
118
119 /*
120 * Make sure that on average at least two pageblocks are almost free
121 * of another type, one for a migratetype to fall back to and a
122 * second to avoid subsequent fallbacks of other types There are 3
123 * MIGRATE_TYPES we care about.
124 */
125 recommended_min += pageblock_nr_pages * nr_zones *
126 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
127
128 /* don't ever allow to reserve more than 5% of the lowmem */
129 recommended_min = min(recommended_min,
130 (unsigned long) nr_free_buffer_pages() / 20);
131 recommended_min <<= (PAGE_SHIFT-10);
132
133 if (recommended_min > min_free_kbytes) {
134 if (user_min_free_kbytes >= 0)
135 pr_info("raising min_free_kbytes from %d to %lu "
136 "to help transparent hugepage allocations\n",
137 min_free_kbytes, recommended_min);
138
139 min_free_kbytes = recommended_min;
140 }
141 setup_per_zone_wmarks();
142 return 0;
143}
144late_initcall(set_recommended_min_free_kbytes);
145
146static int start_khugepaged(void)
147{
148 int err = 0;
149 if (khugepaged_enabled()) {
150 if (!khugepaged_thread)
151 khugepaged_thread = kthread_run(khugepaged, NULL,
152 "khugepaged");
153 if (unlikely(IS_ERR(khugepaged_thread))) {
154 printk(KERN_ERR
155 "khugepaged: kthread_run(khugepaged) failed\n");
156 err = PTR_ERR(khugepaged_thread);
157 khugepaged_thread = NULL;
158 }
159
160 if (!list_empty(&khugepaged_scan.mm_head))
161 wake_up_interruptible(&khugepaged_wait);
162
163 set_recommended_min_free_kbytes();
164 } else if (khugepaged_thread) {
165 kthread_stop(khugepaged_thread);
166 khugepaged_thread = NULL;
167 }
168
169 return err;
170}
171
172static atomic_t huge_zero_refcount;
173static struct page *huge_zero_page __read_mostly;
174
175static inline bool is_huge_zero_page(struct page *page)
176{
177 return ACCESS_ONCE(huge_zero_page) == page;
178}
179
180static inline bool is_huge_zero_pmd(pmd_t pmd)
181{
182 return is_huge_zero_page(pmd_page(pmd));
183}
184
185static struct page *get_huge_zero_page(void)
186{
187 struct page *zero_page;
188retry:
189 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
190 return ACCESS_ONCE(huge_zero_page);
191
192 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
193 HPAGE_PMD_ORDER);
194 if (!zero_page) {
195 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
196 return NULL;
197 }
198 count_vm_event(THP_ZERO_PAGE_ALLOC);
199 preempt_disable();
200 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
201 preempt_enable();
202 __free_page(zero_page);
203 goto retry;
204 }
205
206 /* We take additional reference here. It will be put back by shrinker */
207 atomic_set(&huge_zero_refcount, 2);
208 preempt_enable();
209 return ACCESS_ONCE(huge_zero_page);
210}
211
212static void put_huge_zero_page(void)
213{
214 /*
215 * Counter should never go to zero here. Only shrinker can put
216 * last reference.
217 */
218 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
219}
220
221static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
222 struct shrink_control *sc)
223{
224 /* we can free zero page only if last reference remains */
225 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
226}
227
228static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
229 struct shrink_control *sc)
230{
231 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
232 struct page *zero_page = xchg(&huge_zero_page, NULL);
233 BUG_ON(zero_page == NULL);
234 __free_page(zero_page);
235 return HPAGE_PMD_NR;
236 }
237
238 return 0;
239}
240
241static struct shrinker huge_zero_page_shrinker = {
242 .count_objects = shrink_huge_zero_page_count,
243 .scan_objects = shrink_huge_zero_page_scan,
244 .seeks = DEFAULT_SEEKS,
245};
246
247#ifdef CONFIG_SYSFS
248
249static ssize_t double_flag_show(struct kobject *kobj,
250 struct kobj_attribute *attr, char *buf,
251 enum transparent_hugepage_flag enabled,
252 enum transparent_hugepage_flag req_madv)
253{
254 if (test_bit(enabled, &transparent_hugepage_flags)) {
255 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
256 return sprintf(buf, "[always] madvise never\n");
257 } else if (test_bit(req_madv, &transparent_hugepage_flags))
258 return sprintf(buf, "always [madvise] never\n");
259 else
260 return sprintf(buf, "always madvise [never]\n");
261}
262static ssize_t double_flag_store(struct kobject *kobj,
263 struct kobj_attribute *attr,
264 const char *buf, size_t count,
265 enum transparent_hugepage_flag enabled,
266 enum transparent_hugepage_flag req_madv)
267{
268 if (!memcmp("always", buf,
269 min(sizeof("always")-1, count))) {
270 set_bit(enabled, &transparent_hugepage_flags);
271 clear_bit(req_madv, &transparent_hugepage_flags);
272 } else if (!memcmp("madvise", buf,
273 min(sizeof("madvise")-1, count))) {
274 clear_bit(enabled, &transparent_hugepage_flags);
275 set_bit(req_madv, &transparent_hugepage_flags);
276 } else if (!memcmp("never", buf,
277 min(sizeof("never")-1, count))) {
278 clear_bit(enabled, &transparent_hugepage_flags);
279 clear_bit(req_madv, &transparent_hugepage_flags);
280 } else
281 return -EINVAL;
282
283 return count;
284}
285
286static ssize_t enabled_show(struct kobject *kobj,
287 struct kobj_attribute *attr, char *buf)
288{
289 return double_flag_show(kobj, attr, buf,
290 TRANSPARENT_HUGEPAGE_FLAG,
291 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
292}
293static ssize_t enabled_store(struct kobject *kobj,
294 struct kobj_attribute *attr,
295 const char *buf, size_t count)
296{
297 ssize_t ret;
298
299 ret = double_flag_store(kobj, attr, buf, count,
300 TRANSPARENT_HUGEPAGE_FLAG,
301 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
302
303 if (ret > 0) {
304 int err;
305
306 mutex_lock(&khugepaged_mutex);
307 err = start_khugepaged();
308 mutex_unlock(&khugepaged_mutex);
309
310 if (err)
311 ret = err;
312 }
313
314 return ret;
315}
316static struct kobj_attribute enabled_attr =
317 __ATTR(enabled, 0644, enabled_show, enabled_store);
318
319static ssize_t single_flag_show(struct kobject *kobj,
320 struct kobj_attribute *attr, char *buf,
321 enum transparent_hugepage_flag flag)
322{
323 return sprintf(buf, "%d\n",
324 !!test_bit(flag, &transparent_hugepage_flags));
325}
326
327static ssize_t single_flag_store(struct kobject *kobj,
328 struct kobj_attribute *attr,
329 const char *buf, size_t count,
330 enum transparent_hugepage_flag flag)
331{
332 unsigned long value;
333 int ret;
334
335 ret = kstrtoul(buf, 10, &value);
336 if (ret < 0)
337 return ret;
338 if (value > 1)
339 return -EINVAL;
340
341 if (value)
342 set_bit(flag, &transparent_hugepage_flags);
343 else
344 clear_bit(flag, &transparent_hugepage_flags);
345
346 return count;
347}
348
349/*
350 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
351 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
352 * memory just to allocate one more hugepage.
353 */
354static ssize_t defrag_show(struct kobject *kobj,
355 struct kobj_attribute *attr, char *buf)
356{
357 return double_flag_show(kobj, attr, buf,
358 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
359 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
360}
361static ssize_t defrag_store(struct kobject *kobj,
362 struct kobj_attribute *attr,
363 const char *buf, size_t count)
364{
365 return double_flag_store(kobj, attr, buf, count,
366 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
367 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
368}
369static struct kobj_attribute defrag_attr =
370 __ATTR(defrag, 0644, defrag_show, defrag_store);
371
372static ssize_t use_zero_page_show(struct kobject *kobj,
373 struct kobj_attribute *attr, char *buf)
374{
375 return single_flag_show(kobj, attr, buf,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377}
378static ssize_t use_zero_page_store(struct kobject *kobj,
379 struct kobj_attribute *attr, const char *buf, size_t count)
380{
381 return single_flag_store(kobj, attr, buf, count,
382 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
383}
384static struct kobj_attribute use_zero_page_attr =
385 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
386#ifdef CONFIG_DEBUG_VM
387static ssize_t debug_cow_show(struct kobject *kobj,
388 struct kobj_attribute *attr, char *buf)
389{
390 return single_flag_show(kobj, attr, buf,
391 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
392}
393static ssize_t debug_cow_store(struct kobject *kobj,
394 struct kobj_attribute *attr,
395 const char *buf, size_t count)
396{
397 return single_flag_store(kobj, attr, buf, count,
398 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
399}
400static struct kobj_attribute debug_cow_attr =
401 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
402#endif /* CONFIG_DEBUG_VM */
403
404static struct attribute *hugepage_attr[] = {
405 &enabled_attr.attr,
406 &defrag_attr.attr,
407 &use_zero_page_attr.attr,
408#ifdef CONFIG_DEBUG_VM
409 &debug_cow_attr.attr,
410#endif
411 NULL,
412};
413
414static struct attribute_group hugepage_attr_group = {
415 .attrs = hugepage_attr,
416};
417
418static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
419 struct kobj_attribute *attr,
420 char *buf)
421{
422 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
423}
424
425static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
426 struct kobj_attribute *attr,
427 const char *buf, size_t count)
428{
429 unsigned long msecs;
430 int err;
431
432 err = kstrtoul(buf, 10, &msecs);
433 if (err || msecs > UINT_MAX)
434 return -EINVAL;
435
436 khugepaged_scan_sleep_millisecs = msecs;
437 wake_up_interruptible(&khugepaged_wait);
438
439 return count;
440}
441static struct kobj_attribute scan_sleep_millisecs_attr =
442 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
443 scan_sleep_millisecs_store);
444
445static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
446 struct kobj_attribute *attr,
447 char *buf)
448{
449 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
450}
451
452static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
453 struct kobj_attribute *attr,
454 const char *buf, size_t count)
455{
456 unsigned long msecs;
457 int err;
458
459 err = kstrtoul(buf, 10, &msecs);
460 if (err || msecs > UINT_MAX)
461 return -EINVAL;
462
463 khugepaged_alloc_sleep_millisecs = msecs;
464 wake_up_interruptible(&khugepaged_wait);
465
466 return count;
467}
468static struct kobj_attribute alloc_sleep_millisecs_attr =
469 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
470 alloc_sleep_millisecs_store);
471
472static ssize_t pages_to_scan_show(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 char *buf)
475{
476 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
477}
478static ssize_t pages_to_scan_store(struct kobject *kobj,
479 struct kobj_attribute *attr,
480 const char *buf, size_t count)
481{
482 int err;
483 unsigned long pages;
484
485 err = kstrtoul(buf, 10, &pages);
486 if (err || !pages || pages > UINT_MAX)
487 return -EINVAL;
488
489 khugepaged_pages_to_scan = pages;
490
491 return count;
492}
493static struct kobj_attribute pages_to_scan_attr =
494 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
495 pages_to_scan_store);
496
497static ssize_t pages_collapsed_show(struct kobject *kobj,
498 struct kobj_attribute *attr,
499 char *buf)
500{
501 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
502}
503static struct kobj_attribute pages_collapsed_attr =
504 __ATTR_RO(pages_collapsed);
505
506static ssize_t full_scans_show(struct kobject *kobj,
507 struct kobj_attribute *attr,
508 char *buf)
509{
510 return sprintf(buf, "%u\n", khugepaged_full_scans);
511}
512static struct kobj_attribute full_scans_attr =
513 __ATTR_RO(full_scans);
514
515static ssize_t khugepaged_defrag_show(struct kobject *kobj,
516 struct kobj_attribute *attr, char *buf)
517{
518 return single_flag_show(kobj, attr, buf,
519 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
520}
521static ssize_t khugepaged_defrag_store(struct kobject *kobj,
522 struct kobj_attribute *attr,
523 const char *buf, size_t count)
524{
525 return single_flag_store(kobj, attr, buf, count,
526 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
527}
528static struct kobj_attribute khugepaged_defrag_attr =
529 __ATTR(defrag, 0644, khugepaged_defrag_show,
530 khugepaged_defrag_store);
531
532/*
533 * max_ptes_none controls if khugepaged should collapse hugepages over
534 * any unmapped ptes in turn potentially increasing the memory
535 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
536 * reduce the available free memory in the system as it
537 * runs. Increasing max_ptes_none will instead potentially reduce the
538 * free memory in the system during the khugepaged scan.
539 */
540static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
541 struct kobj_attribute *attr,
542 char *buf)
543{
544 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
545}
546static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
547 struct kobj_attribute *attr,
548 const char *buf, size_t count)
549{
550 int err;
551 unsigned long max_ptes_none;
552
553 err = kstrtoul(buf, 10, &max_ptes_none);
554 if (err || max_ptes_none > HPAGE_PMD_NR-1)
555 return -EINVAL;
556
557 khugepaged_max_ptes_none = max_ptes_none;
558
559 return count;
560}
561static struct kobj_attribute khugepaged_max_ptes_none_attr =
562 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
563 khugepaged_max_ptes_none_store);
564
565static struct attribute *khugepaged_attr[] = {
566 &khugepaged_defrag_attr.attr,
567 &khugepaged_max_ptes_none_attr.attr,
568 &pages_to_scan_attr.attr,
569 &pages_collapsed_attr.attr,
570 &full_scans_attr.attr,
571 &scan_sleep_millisecs_attr.attr,
572 &alloc_sleep_millisecs_attr.attr,
573 NULL,
574};
575
576static struct attribute_group khugepaged_attr_group = {
577 .attrs = khugepaged_attr,
578 .name = "khugepaged",
579};
580
581static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
582{
583 int err;
584
585 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
586 if (unlikely(!*hugepage_kobj)) {
587 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
588 return -ENOMEM;
589 }
590
591 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
592 if (err) {
593 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
594 goto delete_obj;
595 }
596
597 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
598 if (err) {
599 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
600 goto remove_hp_group;
601 }
602
603 return 0;
604
605remove_hp_group:
606 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
607delete_obj:
608 kobject_put(*hugepage_kobj);
609 return err;
610}
611
612static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
613{
614 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
615 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
616 kobject_put(hugepage_kobj);
617}
618#else
619static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
620{
621 return 0;
622}
623
624static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
625{
626}
627#endif /* CONFIG_SYSFS */
628
629static int __init hugepage_init(void)
630{
631 int err;
632 struct kobject *hugepage_kobj;
633
634 if (!has_transparent_hugepage()) {
635 transparent_hugepage_flags = 0;
636 return -EINVAL;
637 }
638
639 err = hugepage_init_sysfs(&hugepage_kobj);
640 if (err)
641 return err;
642
643 err = khugepaged_slab_init();
644 if (err)
645 goto out;
646
647 register_shrinker(&huge_zero_page_shrinker);
648
649 /*
650 * By default disable transparent hugepages on smaller systems,
651 * where the extra memory used could hurt more than TLB overhead
652 * is likely to save. The admin can still enable it through /sys.
653 */
654 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
655 transparent_hugepage_flags = 0;
656
657 start_khugepaged();
658
659 return 0;
660out:
661 hugepage_exit_sysfs(hugepage_kobj);
662 return err;
663}
664subsys_initcall(hugepage_init);
665
666static int __init setup_transparent_hugepage(char *str)
667{
668 int ret = 0;
669 if (!str)
670 goto out;
671 if (!strcmp(str, "always")) {
672 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
676 ret = 1;
677 } else if (!strcmp(str, "madvise")) {
678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
679 &transparent_hugepage_flags);
680 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
681 &transparent_hugepage_flags);
682 ret = 1;
683 } else if (!strcmp(str, "never")) {
684 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
685 &transparent_hugepage_flags);
686 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
687 &transparent_hugepage_flags);
688 ret = 1;
689 }
690out:
691 if (!ret)
692 printk(KERN_WARNING
693 "transparent_hugepage= cannot parse, ignored\n");
694 return ret;
695}
696__setup("transparent_hugepage=", setup_transparent_hugepage);
697
698pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
699{
700 if (likely(vma->vm_flags & VM_WRITE))
701 pmd = pmd_mkwrite(pmd);
702 return pmd;
703}
704
705static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
706{
707 pmd_t entry;
708 entry = mk_pmd(page, prot);
709 entry = pmd_mkhuge(entry);
710 return entry;
711}
712
713static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
714 struct vm_area_struct *vma,
715 unsigned long haddr, pmd_t *pmd,
716 struct page *page)
717{
718 pgtable_t pgtable;
719 spinlock_t *ptl;
720
721 VM_BUG_ON_PAGE(!PageCompound(page), page);
722 pgtable = pte_alloc_one(mm, haddr);
723 if (unlikely(!pgtable))
724 return VM_FAULT_OOM;
725
726 clear_huge_page(page, haddr, HPAGE_PMD_NR);
727 /*
728 * The memory barrier inside __SetPageUptodate makes sure that
729 * clear_huge_page writes become visible before the set_pmd_at()
730 * write.
731 */
732 __SetPageUptodate(page);
733
734 ptl = pmd_lock(mm, pmd);
735 if (unlikely(!pmd_none(*pmd))) {
736 spin_unlock(ptl);
737 mem_cgroup_uncharge_page(page);
738 put_page(page);
739 pte_free(mm, pgtable);
740 } else {
741 pmd_t entry;
742 entry = mk_huge_pmd(page, vma->vm_page_prot);
743 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
744 page_add_new_anon_rmap(page, vma, haddr);
745 pgtable_trans_huge_deposit(mm, pmd, pgtable);
746 set_pmd_at(mm, haddr, pmd, entry);
747 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
748 atomic_long_inc(&mm->nr_ptes);
749 spin_unlock(ptl);
750 }
751
752 return 0;
753}
754
755static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
756{
757 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
758}
759
760static inline struct page *alloc_hugepage_vma(int defrag,
761 struct vm_area_struct *vma,
762 unsigned long haddr, int nd,
763 gfp_t extra_gfp)
764{
765 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
766 HPAGE_PMD_ORDER, vma, haddr, nd);
767}
768
769/* Caller must hold page table lock. */
770static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
771 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
772 struct page *zero_page)
773{
774 pmd_t entry;
775 if (!pmd_none(*pmd))
776 return false;
777 entry = mk_pmd(zero_page, vma->vm_page_prot);
778 entry = pmd_wrprotect(entry);
779 entry = pmd_mkhuge(entry);
780 pgtable_trans_huge_deposit(mm, pmd, pgtable);
781 set_pmd_at(mm, haddr, pmd, entry);
782 atomic_long_inc(&mm->nr_ptes);
783 return true;
784}
785
786int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
787 unsigned long address, pmd_t *pmd,
788 unsigned int flags)
789{
790 struct page *page;
791 unsigned long haddr = address & HPAGE_PMD_MASK;
792
793 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
794 return VM_FAULT_FALLBACK;
795 if (unlikely(anon_vma_prepare(vma)))
796 return VM_FAULT_OOM;
797 if (unlikely(khugepaged_enter(vma)))
798 return VM_FAULT_OOM;
799 if (!(flags & FAULT_FLAG_WRITE) &&
800 transparent_hugepage_use_zero_page()) {
801 spinlock_t *ptl;
802 pgtable_t pgtable;
803 struct page *zero_page;
804 bool set;
805 pgtable = pte_alloc_one(mm, haddr);
806 if (unlikely(!pgtable))
807 return VM_FAULT_OOM;
808 zero_page = get_huge_zero_page();
809 if (unlikely(!zero_page)) {
810 pte_free(mm, pgtable);
811 count_vm_event(THP_FAULT_FALLBACK);
812 return VM_FAULT_FALLBACK;
813 }
814 ptl = pmd_lock(mm, pmd);
815 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
816 zero_page);
817 spin_unlock(ptl);
818 if (!set) {
819 pte_free(mm, pgtable);
820 put_huge_zero_page();
821 }
822 return 0;
823 }
824 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
825 vma, haddr, numa_node_id(), 0);
826 if (unlikely(!page)) {
827 count_vm_event(THP_FAULT_FALLBACK);
828 return VM_FAULT_FALLBACK;
829 }
830 if (unlikely(mem_cgroup_charge_anon(page, mm, GFP_KERNEL))) {
831 put_page(page);
832 count_vm_event(THP_FAULT_FALLBACK);
833 return VM_FAULT_FALLBACK;
834 }
835 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
836 mem_cgroup_uncharge_page(page);
837 put_page(page);
838 count_vm_event(THP_FAULT_FALLBACK);
839 return VM_FAULT_FALLBACK;
840 }
841
842 count_vm_event(THP_FAULT_ALLOC);
843 return 0;
844}
845
846int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
847 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
848 struct vm_area_struct *vma)
849{
850 spinlock_t *dst_ptl, *src_ptl;
851 struct page *src_page;
852 pmd_t pmd;
853 pgtable_t pgtable;
854 int ret;
855
856 ret = -ENOMEM;
857 pgtable = pte_alloc_one(dst_mm, addr);
858 if (unlikely(!pgtable))
859 goto out;
860
861 dst_ptl = pmd_lock(dst_mm, dst_pmd);
862 src_ptl = pmd_lockptr(src_mm, src_pmd);
863 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
864
865 ret = -EAGAIN;
866 pmd = *src_pmd;
867 if (unlikely(!pmd_trans_huge(pmd))) {
868 pte_free(dst_mm, pgtable);
869 goto out_unlock;
870 }
871 /*
872 * When page table lock is held, the huge zero pmd should not be
873 * under splitting since we don't split the page itself, only pmd to
874 * a page table.
875 */
876 if (is_huge_zero_pmd(pmd)) {
877 struct page *zero_page;
878 bool set;
879 /*
880 * get_huge_zero_page() will never allocate a new page here,
881 * since we already have a zero page to copy. It just takes a
882 * reference.
883 */
884 zero_page = get_huge_zero_page();
885 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
886 zero_page);
887 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
888 ret = 0;
889 goto out_unlock;
890 }
891
892 if (unlikely(pmd_trans_splitting(pmd))) {
893 /* split huge page running from under us */
894 spin_unlock(src_ptl);
895 spin_unlock(dst_ptl);
896 pte_free(dst_mm, pgtable);
897
898 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
899 goto out;
900 }
901 src_page = pmd_page(pmd);
902 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
903 get_page(src_page);
904 page_dup_rmap(src_page);
905 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
906
907 pmdp_set_wrprotect(src_mm, addr, src_pmd);
908 pmd = pmd_mkold(pmd_wrprotect(pmd));
909 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
910 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
911 atomic_long_inc(&dst_mm->nr_ptes);
912
913 ret = 0;
914out_unlock:
915 spin_unlock(src_ptl);
916 spin_unlock(dst_ptl);
917out:
918 return ret;
919}
920
921void huge_pmd_set_accessed(struct mm_struct *mm,
922 struct vm_area_struct *vma,
923 unsigned long address,
924 pmd_t *pmd, pmd_t orig_pmd,
925 int dirty)
926{
927 spinlock_t *ptl;
928 pmd_t entry;
929 unsigned long haddr;
930
931 ptl = pmd_lock(mm, pmd);
932 if (unlikely(!pmd_same(*pmd, orig_pmd)))
933 goto unlock;
934
935 entry = pmd_mkyoung(orig_pmd);
936 haddr = address & HPAGE_PMD_MASK;
937 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
938 update_mmu_cache_pmd(vma, address, pmd);
939
940unlock:
941 spin_unlock(ptl);
942}
943
944static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
945 struct vm_area_struct *vma,
946 unsigned long address,
947 pmd_t *pmd, pmd_t orig_pmd,
948 struct page *page,
949 unsigned long haddr)
950{
951 spinlock_t *ptl;
952 pgtable_t pgtable;
953 pmd_t _pmd;
954 int ret = 0, i;
955 struct page **pages;
956 unsigned long mmun_start; /* For mmu_notifiers */
957 unsigned long mmun_end; /* For mmu_notifiers */
958
959 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
960 GFP_KERNEL);
961 if (unlikely(!pages)) {
962 ret |= VM_FAULT_OOM;
963 goto out;
964 }
965
966 for (i = 0; i < HPAGE_PMD_NR; i++) {
967 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
968 __GFP_OTHER_NODE,
969 vma, address, page_to_nid(page));
970 if (unlikely(!pages[i] ||
971 mem_cgroup_charge_anon(pages[i], mm,
972 GFP_KERNEL))) {
973 if (pages[i])
974 put_page(pages[i]);
975 mem_cgroup_uncharge_start();
976 while (--i >= 0) {
977 mem_cgroup_uncharge_page(pages[i]);
978 put_page(pages[i]);
979 }
980 mem_cgroup_uncharge_end();
981 kfree(pages);
982 ret |= VM_FAULT_OOM;
983 goto out;
984 }
985 }
986
987 for (i = 0; i < HPAGE_PMD_NR; i++) {
988 copy_user_highpage(pages[i], page + i,
989 haddr + PAGE_SIZE * i, vma);
990 __SetPageUptodate(pages[i]);
991 cond_resched();
992 }
993
994 mmun_start = haddr;
995 mmun_end = haddr + HPAGE_PMD_SIZE;
996 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
997
998 ptl = pmd_lock(mm, pmd);
999 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1000 goto out_free_pages;
1001 VM_BUG_ON_PAGE(!PageHead(page), page);
1002
1003 pmdp_clear_flush(vma, haddr, pmd);
1004 /* leave pmd empty until pte is filled */
1005
1006 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1007 pmd_populate(mm, &_pmd, pgtable);
1008
1009 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1010 pte_t *pte, entry;
1011 entry = mk_pte(pages[i], vma->vm_page_prot);
1012 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1013 page_add_new_anon_rmap(pages[i], vma, haddr);
1014 pte = pte_offset_map(&_pmd, haddr);
1015 VM_BUG_ON(!pte_none(*pte));
1016 set_pte_at(mm, haddr, pte, entry);
1017 pte_unmap(pte);
1018 }
1019 kfree(pages);
1020
1021 smp_wmb(); /* make pte visible before pmd */
1022 pmd_populate(mm, pmd, pgtable);
1023 page_remove_rmap(page);
1024 spin_unlock(ptl);
1025
1026 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1027
1028 ret |= VM_FAULT_WRITE;
1029 put_page(page);
1030
1031out:
1032 return ret;
1033
1034out_free_pages:
1035 spin_unlock(ptl);
1036 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1037 mem_cgroup_uncharge_start();
1038 for (i = 0; i < HPAGE_PMD_NR; i++) {
1039 mem_cgroup_uncharge_page(pages[i]);
1040 put_page(pages[i]);
1041 }
1042 mem_cgroup_uncharge_end();
1043 kfree(pages);
1044 goto out;
1045}
1046
1047int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1048 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1049{
1050 spinlock_t *ptl;
1051 int ret = 0;
1052 struct page *page = NULL, *new_page;
1053 unsigned long haddr;
1054 unsigned long mmun_start; /* For mmu_notifiers */
1055 unsigned long mmun_end; /* For mmu_notifiers */
1056
1057 ptl = pmd_lockptr(mm, pmd);
1058 VM_BUG_ON(!vma->anon_vma);
1059 haddr = address & HPAGE_PMD_MASK;
1060 if (is_huge_zero_pmd(orig_pmd))
1061 goto alloc;
1062 spin_lock(ptl);
1063 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1064 goto out_unlock;
1065
1066 page = pmd_page(orig_pmd);
1067 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1068 if (page_mapcount(page) == 1) {
1069 pmd_t entry;
1070 entry = pmd_mkyoung(orig_pmd);
1071 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1072 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1073 update_mmu_cache_pmd(vma, address, pmd);
1074 ret |= VM_FAULT_WRITE;
1075 goto out_unlock;
1076 }
1077 get_page(page);
1078 spin_unlock(ptl);
1079alloc:
1080 if (transparent_hugepage_enabled(vma) &&
1081 !transparent_hugepage_debug_cow())
1082 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1083 vma, haddr, numa_node_id(), 0);
1084 else
1085 new_page = NULL;
1086
1087 if (unlikely(!new_page)) {
1088 if (!page) {
1089 split_huge_page_pmd(vma, address, pmd);
1090 ret |= VM_FAULT_FALLBACK;
1091 } else {
1092 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1093 pmd, orig_pmd, page, haddr);
1094 if (ret & VM_FAULT_OOM) {
1095 split_huge_page(page);
1096 ret |= VM_FAULT_FALLBACK;
1097 }
1098 put_page(page);
1099 }
1100 count_vm_event(THP_FAULT_FALLBACK);
1101 goto out;
1102 }
1103
1104 if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))) {
1105 put_page(new_page);
1106 if (page) {
1107 split_huge_page(page);
1108 put_page(page);
1109 } else
1110 split_huge_page_pmd(vma, address, pmd);
1111 ret |= VM_FAULT_FALLBACK;
1112 count_vm_event(THP_FAULT_FALLBACK);
1113 goto out;
1114 }
1115
1116 count_vm_event(THP_FAULT_ALLOC);
1117
1118 if (!page)
1119 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1120 else
1121 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1122 __SetPageUptodate(new_page);
1123
1124 mmun_start = haddr;
1125 mmun_end = haddr + HPAGE_PMD_SIZE;
1126 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1127
1128 spin_lock(ptl);
1129 if (page)
1130 put_page(page);
1131 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1132 spin_unlock(ptl);
1133 mem_cgroup_uncharge_page(new_page);
1134 put_page(new_page);
1135 goto out_mn;
1136 } else {
1137 pmd_t entry;
1138 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1139 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1140 pmdp_clear_flush(vma, haddr, pmd);
1141 page_add_new_anon_rmap(new_page, vma, haddr);
1142 set_pmd_at(mm, haddr, pmd, entry);
1143 update_mmu_cache_pmd(vma, address, pmd);
1144 if (!page) {
1145 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1146 put_huge_zero_page();
1147 } else {
1148 VM_BUG_ON_PAGE(!PageHead(page), page);
1149 page_remove_rmap(page);
1150 put_page(page);
1151 }
1152 ret |= VM_FAULT_WRITE;
1153 }
1154 spin_unlock(ptl);
1155out_mn:
1156 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1157out:
1158 return ret;
1159out_unlock:
1160 spin_unlock(ptl);
1161 return ret;
1162}
1163
1164struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1165 unsigned long addr,
1166 pmd_t *pmd,
1167 unsigned int flags)
1168{
1169 struct mm_struct *mm = vma->vm_mm;
1170 struct page *page = NULL;
1171
1172 assert_spin_locked(pmd_lockptr(mm, pmd));
1173
1174 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1175 goto out;
1176
1177 /* Avoid dumping huge zero page */
1178 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1179 return ERR_PTR(-EFAULT);
1180
1181 /* Full NUMA hinting faults to serialise migration in fault paths */
1182 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1183 goto out;
1184
1185 page = pmd_page(*pmd);
1186 VM_BUG_ON_PAGE(!PageHead(page), page);
1187 if (flags & FOLL_TOUCH) {
1188 pmd_t _pmd;
1189 /*
1190 * We should set the dirty bit only for FOLL_WRITE but
1191 * for now the dirty bit in the pmd is meaningless.
1192 * And if the dirty bit will become meaningful and
1193 * we'll only set it with FOLL_WRITE, an atomic
1194 * set_bit will be required on the pmd to set the
1195 * young bit, instead of the current set_pmd_at.
1196 */
1197 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1198 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1199 pmd, _pmd, 1))
1200 update_mmu_cache_pmd(vma, addr, pmd);
1201 }
1202 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1203 if (page->mapping && trylock_page(page)) {
1204 lru_add_drain();
1205 if (page->mapping)
1206 mlock_vma_page(page);
1207 unlock_page(page);
1208 }
1209 }
1210 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1211 VM_BUG_ON_PAGE(!PageCompound(page), page);
1212 if (flags & FOLL_GET)
1213 get_page_foll(page);
1214
1215out:
1216 return page;
1217}
1218
1219/* NUMA hinting page fault entry point for trans huge pmds */
1220int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1221 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1222{
1223 spinlock_t *ptl;
1224 struct anon_vma *anon_vma = NULL;
1225 struct page *page;
1226 unsigned long haddr = addr & HPAGE_PMD_MASK;
1227 int page_nid = -1, this_nid = numa_node_id();
1228 int target_nid, last_cpupid = -1;
1229 bool page_locked;
1230 bool migrated = false;
1231 int flags = 0;
1232
1233 ptl = pmd_lock(mm, pmdp);
1234 if (unlikely(!pmd_same(pmd, *pmdp)))
1235 goto out_unlock;
1236
1237 /*
1238 * If there are potential migrations, wait for completion and retry
1239 * without disrupting NUMA hinting information. Do not relock and
1240 * check_same as the page may no longer be mapped.
1241 */
1242 if (unlikely(pmd_trans_migrating(*pmdp))) {
1243 spin_unlock(ptl);
1244 wait_migrate_huge_page(vma->anon_vma, pmdp);
1245 goto out;
1246 }
1247
1248 page = pmd_page(pmd);
1249 BUG_ON(is_huge_zero_page(page));
1250 page_nid = page_to_nid(page);
1251 last_cpupid = page_cpupid_last(page);
1252 count_vm_numa_event(NUMA_HINT_FAULTS);
1253 if (page_nid == this_nid) {
1254 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1255 flags |= TNF_FAULT_LOCAL;
1256 }
1257
1258 /*
1259 * Avoid grouping on DSO/COW pages in specific and RO pages
1260 * in general, RO pages shouldn't hurt as much anyway since
1261 * they can be in shared cache state.
1262 */
1263 if (!pmd_write(pmd))
1264 flags |= TNF_NO_GROUP;
1265
1266 /*
1267 * Acquire the page lock to serialise THP migrations but avoid dropping
1268 * page_table_lock if at all possible
1269 */
1270 page_locked = trylock_page(page);
1271 target_nid = mpol_misplaced(page, vma, haddr);
1272 if (target_nid == -1) {
1273 /* If the page was locked, there are no parallel migrations */
1274 if (page_locked)
1275 goto clear_pmdnuma;
1276 }
1277
1278 /* Migration could have started since the pmd_trans_migrating check */
1279 if (!page_locked) {
1280 spin_unlock(ptl);
1281 wait_on_page_locked(page);
1282 page_nid = -1;
1283 goto out;
1284 }
1285
1286 /*
1287 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1288 * to serialises splits
1289 */
1290 get_page(page);
1291 spin_unlock(ptl);
1292 anon_vma = page_lock_anon_vma_read(page);
1293
1294 /* Confirm the PMD did not change while page_table_lock was released */
1295 spin_lock(ptl);
1296 if (unlikely(!pmd_same(pmd, *pmdp))) {
1297 unlock_page(page);
1298 put_page(page);
1299 page_nid = -1;
1300 goto out_unlock;
1301 }
1302
1303 /* Bail if we fail to protect against THP splits for any reason */
1304 if (unlikely(!anon_vma)) {
1305 put_page(page);
1306 page_nid = -1;
1307 goto clear_pmdnuma;
1308 }
1309
1310 /*
1311 * Migrate the THP to the requested node, returns with page unlocked
1312 * and pmd_numa cleared.
1313 */
1314 spin_unlock(ptl);
1315 migrated = migrate_misplaced_transhuge_page(mm, vma,
1316 pmdp, pmd, addr, page, target_nid);
1317 if (migrated) {
1318 flags |= TNF_MIGRATED;
1319 page_nid = target_nid;
1320 }
1321
1322 goto out;
1323clear_pmdnuma:
1324 BUG_ON(!PageLocked(page));
1325 pmd = pmd_mknonnuma(pmd);
1326 set_pmd_at(mm, haddr, pmdp, pmd);
1327 VM_BUG_ON(pmd_numa(*pmdp));
1328 update_mmu_cache_pmd(vma, addr, pmdp);
1329 unlock_page(page);
1330out_unlock:
1331 spin_unlock(ptl);
1332
1333out:
1334 if (anon_vma)
1335 page_unlock_anon_vma_read(anon_vma);
1336
1337 if (page_nid != -1)
1338 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1339
1340 return 0;
1341}
1342
1343int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1344 pmd_t *pmd, unsigned long addr)
1345{
1346 spinlock_t *ptl;
1347 int ret = 0;
1348
1349 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1350 struct page *page;
1351 pgtable_t pgtable;
1352 pmd_t orig_pmd;
1353 /*
1354 * For architectures like ppc64 we look at deposited pgtable
1355 * when calling pmdp_get_and_clear. So do the
1356 * pgtable_trans_huge_withdraw after finishing pmdp related
1357 * operations.
1358 */
1359 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1360 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1361 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1362 if (is_huge_zero_pmd(orig_pmd)) {
1363 atomic_long_dec(&tlb->mm->nr_ptes);
1364 spin_unlock(ptl);
1365 put_huge_zero_page();
1366 } else {
1367 page = pmd_page(orig_pmd);
1368 page_remove_rmap(page);
1369 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1370 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1371 VM_BUG_ON_PAGE(!PageHead(page), page);
1372 atomic_long_dec(&tlb->mm->nr_ptes);
1373 spin_unlock(ptl);
1374 tlb_remove_page(tlb, page);
1375 }
1376 pte_free(tlb->mm, pgtable);
1377 ret = 1;
1378 }
1379 return ret;
1380}
1381
1382int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1383 unsigned long addr, unsigned long end,
1384 unsigned char *vec)
1385{
1386 spinlock_t *ptl;
1387 int ret = 0;
1388
1389 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1390 /*
1391 * All logical pages in the range are present
1392 * if backed by a huge page.
1393 */
1394 spin_unlock(ptl);
1395 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1396 ret = 1;
1397 }
1398
1399 return ret;
1400}
1401
1402int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1403 unsigned long old_addr,
1404 unsigned long new_addr, unsigned long old_end,
1405 pmd_t *old_pmd, pmd_t *new_pmd)
1406{
1407 spinlock_t *old_ptl, *new_ptl;
1408 int ret = 0;
1409 pmd_t pmd;
1410
1411 struct mm_struct *mm = vma->vm_mm;
1412
1413 if ((old_addr & ~HPAGE_PMD_MASK) ||
1414 (new_addr & ~HPAGE_PMD_MASK) ||
1415 old_end - old_addr < HPAGE_PMD_SIZE ||
1416 (new_vma->vm_flags & VM_NOHUGEPAGE))
1417 goto out;
1418
1419 /*
1420 * The destination pmd shouldn't be established, free_pgtables()
1421 * should have release it.
1422 */
1423 if (WARN_ON(!pmd_none(*new_pmd))) {
1424 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1425 goto out;
1426 }
1427
1428 /*
1429 * We don't have to worry about the ordering of src and dst
1430 * ptlocks because exclusive mmap_sem prevents deadlock.
1431 */
1432 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1433 if (ret == 1) {
1434 new_ptl = pmd_lockptr(mm, new_pmd);
1435 if (new_ptl != old_ptl)
1436 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1437 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1438 VM_BUG_ON(!pmd_none(*new_pmd));
1439
1440 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1441 pgtable_t pgtable;
1442 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1443 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1444 }
1445 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1446 if (new_ptl != old_ptl)
1447 spin_unlock(new_ptl);
1448 spin_unlock(old_ptl);
1449 }
1450out:
1451 return ret;
1452}
1453
1454/*
1455 * Returns
1456 * - 0 if PMD could not be locked
1457 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1458 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1459 */
1460int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1461 unsigned long addr, pgprot_t newprot, int prot_numa)
1462{
1463 struct mm_struct *mm = vma->vm_mm;
1464 spinlock_t *ptl;
1465 int ret = 0;
1466
1467 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1468 pmd_t entry;
1469 ret = 1;
1470 if (!prot_numa) {
1471 entry = pmdp_get_and_clear(mm, addr, pmd);
1472 if (pmd_numa(entry))
1473 entry = pmd_mknonnuma(entry);
1474 entry = pmd_modify(entry, newprot);
1475 ret = HPAGE_PMD_NR;
1476 set_pmd_at(mm, addr, pmd, entry);
1477 BUG_ON(pmd_write(entry));
1478 } else {
1479 struct page *page = pmd_page(*pmd);
1480
1481 /*
1482 * Do not trap faults against the zero page. The
1483 * read-only data is likely to be read-cached on the
1484 * local CPU cache and it is less useful to know about
1485 * local vs remote hits on the zero page.
1486 */
1487 if (!is_huge_zero_page(page) &&
1488 !pmd_numa(*pmd)) {
1489 pmdp_set_numa(mm, addr, pmd);
1490 ret = HPAGE_PMD_NR;
1491 }
1492 }
1493 spin_unlock(ptl);
1494 }
1495
1496 return ret;
1497}
1498
1499/*
1500 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1501 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1502 *
1503 * Note that if it returns 1, this routine returns without unlocking page
1504 * table locks. So callers must unlock them.
1505 */
1506int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1507 spinlock_t **ptl)
1508{
1509 *ptl = pmd_lock(vma->vm_mm, pmd);
1510 if (likely(pmd_trans_huge(*pmd))) {
1511 if (unlikely(pmd_trans_splitting(*pmd))) {
1512 spin_unlock(*ptl);
1513 wait_split_huge_page(vma->anon_vma, pmd);
1514 return -1;
1515 } else {
1516 /* Thp mapped by 'pmd' is stable, so we can
1517 * handle it as it is. */
1518 return 1;
1519 }
1520 }
1521 spin_unlock(*ptl);
1522 return 0;
1523}
1524
1525/*
1526 * This function returns whether a given @page is mapped onto the @address
1527 * in the virtual space of @mm.
1528 *
1529 * When it's true, this function returns *pmd with holding the page table lock
1530 * and passing it back to the caller via @ptl.
1531 * If it's false, returns NULL without holding the page table lock.
1532 */
1533pmd_t *page_check_address_pmd(struct page *page,
1534 struct mm_struct *mm,
1535 unsigned long address,
1536 enum page_check_address_pmd_flag flag,
1537 spinlock_t **ptl)
1538{
1539 pgd_t *pgd;
1540 pud_t *pud;
1541 pmd_t *pmd;
1542
1543 if (address & ~HPAGE_PMD_MASK)
1544 return NULL;
1545
1546 pgd = pgd_offset(mm, address);
1547 if (!pgd_present(*pgd))
1548 return NULL;
1549 pud = pud_offset(pgd, address);
1550 if (!pud_present(*pud))
1551 return NULL;
1552 pmd = pmd_offset(pud, address);
1553
1554 *ptl = pmd_lock(mm, pmd);
1555 if (!pmd_present(*pmd))
1556 goto unlock;
1557 if (pmd_page(*pmd) != page)
1558 goto unlock;
1559 /*
1560 * split_vma() may create temporary aliased mappings. There is
1561 * no risk as long as all huge pmd are found and have their
1562 * splitting bit set before __split_huge_page_refcount
1563 * runs. Finding the same huge pmd more than once during the
1564 * same rmap walk is not a problem.
1565 */
1566 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1567 pmd_trans_splitting(*pmd))
1568 goto unlock;
1569 if (pmd_trans_huge(*pmd)) {
1570 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1571 !pmd_trans_splitting(*pmd));
1572 return pmd;
1573 }
1574unlock:
1575 spin_unlock(*ptl);
1576 return NULL;
1577}
1578
1579static int __split_huge_page_splitting(struct page *page,
1580 struct vm_area_struct *vma,
1581 unsigned long address)
1582{
1583 struct mm_struct *mm = vma->vm_mm;
1584 spinlock_t *ptl;
1585 pmd_t *pmd;
1586 int ret = 0;
1587 /* For mmu_notifiers */
1588 const unsigned long mmun_start = address;
1589 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1590
1591 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1592 pmd = page_check_address_pmd(page, mm, address,
1593 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1594 if (pmd) {
1595 /*
1596 * We can't temporarily set the pmd to null in order
1597 * to split it, the pmd must remain marked huge at all
1598 * times or the VM won't take the pmd_trans_huge paths
1599 * and it won't wait on the anon_vma->root->rwsem to
1600 * serialize against split_huge_page*.
1601 */
1602 pmdp_splitting_flush(vma, address, pmd);
1603 ret = 1;
1604 spin_unlock(ptl);
1605 }
1606 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1607
1608 return ret;
1609}
1610
1611static void __split_huge_page_refcount(struct page *page,
1612 struct list_head *list)
1613{
1614 int i;
1615 struct zone *zone = page_zone(page);
1616 struct lruvec *lruvec;
1617 int tail_count = 0;
1618
1619 /* prevent PageLRU to go away from under us, and freeze lru stats */
1620 spin_lock_irq(&zone->lru_lock);
1621 lruvec = mem_cgroup_page_lruvec(page, zone);
1622
1623 compound_lock(page);
1624 /* complete memcg works before add pages to LRU */
1625 mem_cgroup_split_huge_fixup(page);
1626
1627 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1628 struct page *page_tail = page + i;
1629
1630 /* tail_page->_mapcount cannot change */
1631 BUG_ON(page_mapcount(page_tail) < 0);
1632 tail_count += page_mapcount(page_tail);
1633 /* check for overflow */
1634 BUG_ON(tail_count < 0);
1635 BUG_ON(atomic_read(&page_tail->_count) != 0);
1636 /*
1637 * tail_page->_count is zero and not changing from
1638 * under us. But get_page_unless_zero() may be running
1639 * from under us on the tail_page. If we used
1640 * atomic_set() below instead of atomic_add(), we
1641 * would then run atomic_set() concurrently with
1642 * get_page_unless_zero(), and atomic_set() is
1643 * implemented in C not using locked ops. spin_unlock
1644 * on x86 sometime uses locked ops because of PPro
1645 * errata 66, 92, so unless somebody can guarantee
1646 * atomic_set() here would be safe on all archs (and
1647 * not only on x86), it's safer to use atomic_add().
1648 */
1649 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1650 &page_tail->_count);
1651
1652 /* after clearing PageTail the gup refcount can be released */
1653 smp_mb();
1654
1655 /*
1656 * retain hwpoison flag of the poisoned tail page:
1657 * fix for the unsuitable process killed on Guest Machine(KVM)
1658 * by the memory-failure.
1659 */
1660 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1661 page_tail->flags |= (page->flags &
1662 ((1L << PG_referenced) |
1663 (1L << PG_swapbacked) |
1664 (1L << PG_mlocked) |
1665 (1L << PG_uptodate) |
1666 (1L << PG_active) |
1667 (1L << PG_unevictable)));
1668 page_tail->flags |= (1L << PG_dirty);
1669
1670 /* clear PageTail before overwriting first_page */
1671 smp_wmb();
1672
1673 /*
1674 * __split_huge_page_splitting() already set the
1675 * splitting bit in all pmd that could map this
1676 * hugepage, that will ensure no CPU can alter the
1677 * mapcount on the head page. The mapcount is only
1678 * accounted in the head page and it has to be
1679 * transferred to all tail pages in the below code. So
1680 * for this code to be safe, the split the mapcount
1681 * can't change. But that doesn't mean userland can't
1682 * keep changing and reading the page contents while
1683 * we transfer the mapcount, so the pmd splitting
1684 * status is achieved setting a reserved bit in the
1685 * pmd, not by clearing the present bit.
1686 */
1687 page_tail->_mapcount = page->_mapcount;
1688
1689 BUG_ON(page_tail->mapping);
1690 page_tail->mapping = page->mapping;
1691
1692 page_tail->index = page->index + i;
1693 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1694
1695 BUG_ON(!PageAnon(page_tail));
1696 BUG_ON(!PageUptodate(page_tail));
1697 BUG_ON(!PageDirty(page_tail));
1698 BUG_ON(!PageSwapBacked(page_tail));
1699
1700 lru_add_page_tail(page, page_tail, lruvec, list);
1701 }
1702 atomic_sub(tail_count, &page->_count);
1703 BUG_ON(atomic_read(&page->_count) <= 0);
1704
1705 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1706
1707 ClearPageCompound(page);
1708 compound_unlock(page);
1709 spin_unlock_irq(&zone->lru_lock);
1710
1711 for (i = 1; i < HPAGE_PMD_NR; i++) {
1712 struct page *page_tail = page + i;
1713 BUG_ON(page_count(page_tail) <= 0);
1714 /*
1715 * Tail pages may be freed if there wasn't any mapping
1716 * like if add_to_swap() is running on a lru page that
1717 * had its mapping zapped. And freeing these pages
1718 * requires taking the lru_lock so we do the put_page
1719 * of the tail pages after the split is complete.
1720 */
1721 put_page(page_tail);
1722 }
1723
1724 /*
1725 * Only the head page (now become a regular page) is required
1726 * to be pinned by the caller.
1727 */
1728 BUG_ON(page_count(page) <= 0);
1729}
1730
1731static int __split_huge_page_map(struct page *page,
1732 struct vm_area_struct *vma,
1733 unsigned long address)
1734{
1735 struct mm_struct *mm = vma->vm_mm;
1736 spinlock_t *ptl;
1737 pmd_t *pmd, _pmd;
1738 int ret = 0, i;
1739 pgtable_t pgtable;
1740 unsigned long haddr;
1741
1742 pmd = page_check_address_pmd(page, mm, address,
1743 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1744 if (pmd) {
1745 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1746 pmd_populate(mm, &_pmd, pgtable);
1747
1748 haddr = address;
1749 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1750 pte_t *pte, entry;
1751 BUG_ON(PageCompound(page+i));
1752 entry = mk_pte(page + i, vma->vm_page_prot);
1753 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1754 if (!pmd_write(*pmd))
1755 entry = pte_wrprotect(entry);
1756 else
1757 BUG_ON(page_mapcount(page) != 1);
1758 if (!pmd_young(*pmd))
1759 entry = pte_mkold(entry);
1760 if (pmd_numa(*pmd))
1761 entry = pte_mknuma(entry);
1762 pte = pte_offset_map(&_pmd, haddr);
1763 BUG_ON(!pte_none(*pte));
1764 set_pte_at(mm, haddr, pte, entry);
1765 pte_unmap(pte);
1766 }
1767
1768 smp_wmb(); /* make pte visible before pmd */
1769 /*
1770 * Up to this point the pmd is present and huge and
1771 * userland has the whole access to the hugepage
1772 * during the split (which happens in place). If we
1773 * overwrite the pmd with the not-huge version
1774 * pointing to the pte here (which of course we could
1775 * if all CPUs were bug free), userland could trigger
1776 * a small page size TLB miss on the small sized TLB
1777 * while the hugepage TLB entry is still established
1778 * in the huge TLB. Some CPU doesn't like that. See
1779 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1780 * Erratum 383 on page 93. Intel should be safe but is
1781 * also warns that it's only safe if the permission
1782 * and cache attributes of the two entries loaded in
1783 * the two TLB is identical (which should be the case
1784 * here). But it is generally safer to never allow
1785 * small and huge TLB entries for the same virtual
1786 * address to be loaded simultaneously. So instead of
1787 * doing "pmd_populate(); flush_tlb_range();" we first
1788 * mark the current pmd notpresent (atomically because
1789 * here the pmd_trans_huge and pmd_trans_splitting
1790 * must remain set at all times on the pmd until the
1791 * split is complete for this pmd), then we flush the
1792 * SMP TLB and finally we write the non-huge version
1793 * of the pmd entry with pmd_populate.
1794 */
1795 pmdp_invalidate(vma, address, pmd);
1796 pmd_populate(mm, pmd, pgtable);
1797 ret = 1;
1798 spin_unlock(ptl);
1799 }
1800
1801 return ret;
1802}
1803
1804/* must be called with anon_vma->root->rwsem held */
1805static void __split_huge_page(struct page *page,
1806 struct anon_vma *anon_vma,
1807 struct list_head *list)
1808{
1809 int mapcount, mapcount2;
1810 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1811 struct anon_vma_chain *avc;
1812
1813 BUG_ON(!PageHead(page));
1814 BUG_ON(PageTail(page));
1815
1816 mapcount = 0;
1817 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1818 struct vm_area_struct *vma = avc->vma;
1819 unsigned long addr = vma_address(page, vma);
1820 BUG_ON(is_vma_temporary_stack(vma));
1821 mapcount += __split_huge_page_splitting(page, vma, addr);
1822 }
1823 /*
1824 * It is critical that new vmas are added to the tail of the
1825 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1826 * and establishes a child pmd before
1827 * __split_huge_page_splitting() freezes the parent pmd (so if
1828 * we fail to prevent copy_huge_pmd() from running until the
1829 * whole __split_huge_page() is complete), we will still see
1830 * the newly established pmd of the child later during the
1831 * walk, to be able to set it as pmd_trans_splitting too.
1832 */
1833 if (mapcount != page_mapcount(page))
1834 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1835 mapcount, page_mapcount(page));
1836 BUG_ON(mapcount != page_mapcount(page));
1837
1838 __split_huge_page_refcount(page, list);
1839
1840 mapcount2 = 0;
1841 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1842 struct vm_area_struct *vma = avc->vma;
1843 unsigned long addr = vma_address(page, vma);
1844 BUG_ON(is_vma_temporary_stack(vma));
1845 mapcount2 += __split_huge_page_map(page, vma, addr);
1846 }
1847 if (mapcount != mapcount2)
1848 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1849 mapcount, mapcount2, page_mapcount(page));
1850 BUG_ON(mapcount != mapcount2);
1851}
1852
1853/*
1854 * Split a hugepage into normal pages. This doesn't change the position of head
1855 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1856 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1857 * from the hugepage.
1858 * Return 0 if the hugepage is split successfully otherwise return 1.
1859 */
1860int split_huge_page_to_list(struct page *page, struct list_head *list)
1861{
1862 struct anon_vma *anon_vma;
1863 int ret = 1;
1864
1865 BUG_ON(is_huge_zero_page(page));
1866 BUG_ON(!PageAnon(page));
1867
1868 /*
1869 * The caller does not necessarily hold an mmap_sem that would prevent
1870 * the anon_vma disappearing so we first we take a reference to it
1871 * and then lock the anon_vma for write. This is similar to
1872 * page_lock_anon_vma_read except the write lock is taken to serialise
1873 * against parallel split or collapse operations.
1874 */
1875 anon_vma = page_get_anon_vma(page);
1876 if (!anon_vma)
1877 goto out;
1878 anon_vma_lock_write(anon_vma);
1879
1880 ret = 0;
1881 if (!PageCompound(page))
1882 goto out_unlock;
1883
1884 BUG_ON(!PageSwapBacked(page));
1885 __split_huge_page(page, anon_vma, list);
1886 count_vm_event(THP_SPLIT);
1887
1888 BUG_ON(PageCompound(page));
1889out_unlock:
1890 anon_vma_unlock_write(anon_vma);
1891 put_anon_vma(anon_vma);
1892out:
1893 return ret;
1894}
1895
1896#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1897
1898int hugepage_madvise(struct vm_area_struct *vma,
1899 unsigned long *vm_flags, int advice)
1900{
1901 switch (advice) {
1902 case MADV_HUGEPAGE:
1903#ifdef CONFIG_S390
1904 /*
1905 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1906 * can't handle this properly after s390_enable_sie, so we simply
1907 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1908 */
1909 if (mm_has_pgste(vma->vm_mm))
1910 return 0;
1911#endif
1912 /*
1913 * Be somewhat over-protective like KSM for now!
1914 */
1915 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1916 return -EINVAL;
1917 *vm_flags &= ~VM_NOHUGEPAGE;
1918 *vm_flags |= VM_HUGEPAGE;
1919 /*
1920 * If the vma become good for khugepaged to scan,
1921 * register it here without waiting a page fault that
1922 * may not happen any time soon.
1923 */
1924 if (unlikely(khugepaged_enter_vma_merge(vma)))
1925 return -ENOMEM;
1926 break;
1927 case MADV_NOHUGEPAGE:
1928 /*
1929 * Be somewhat over-protective like KSM for now!
1930 */
1931 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1932 return -EINVAL;
1933 *vm_flags &= ~VM_HUGEPAGE;
1934 *vm_flags |= VM_NOHUGEPAGE;
1935 /*
1936 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1937 * this vma even if we leave the mm registered in khugepaged if
1938 * it got registered before VM_NOHUGEPAGE was set.
1939 */
1940 break;
1941 }
1942
1943 return 0;
1944}
1945
1946static int __init khugepaged_slab_init(void)
1947{
1948 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1949 sizeof(struct mm_slot),
1950 __alignof__(struct mm_slot), 0, NULL);
1951 if (!mm_slot_cache)
1952 return -ENOMEM;
1953
1954 return 0;
1955}
1956
1957static inline struct mm_slot *alloc_mm_slot(void)
1958{
1959 if (!mm_slot_cache) /* initialization failed */
1960 return NULL;
1961 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1962}
1963
1964static inline void free_mm_slot(struct mm_slot *mm_slot)
1965{
1966 kmem_cache_free(mm_slot_cache, mm_slot);
1967}
1968
1969static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1970{
1971 struct mm_slot *mm_slot;
1972
1973 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1974 if (mm == mm_slot->mm)
1975 return mm_slot;
1976
1977 return NULL;
1978}
1979
1980static void insert_to_mm_slots_hash(struct mm_struct *mm,
1981 struct mm_slot *mm_slot)
1982{
1983 mm_slot->mm = mm;
1984 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1985}
1986
1987static inline int khugepaged_test_exit(struct mm_struct *mm)
1988{
1989 return atomic_read(&mm->mm_users) == 0;
1990}
1991
1992int __khugepaged_enter(struct mm_struct *mm)
1993{
1994 struct mm_slot *mm_slot;
1995 int wakeup;
1996
1997 mm_slot = alloc_mm_slot();
1998 if (!mm_slot)
1999 return -ENOMEM;
2000
2001 /* __khugepaged_exit() must not run from under us */
2002 VM_BUG_ON(khugepaged_test_exit(mm));
2003 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2004 free_mm_slot(mm_slot);
2005 return 0;
2006 }
2007
2008 spin_lock(&khugepaged_mm_lock);
2009 insert_to_mm_slots_hash(mm, mm_slot);
2010 /*
2011 * Insert just behind the scanning cursor, to let the area settle
2012 * down a little.
2013 */
2014 wakeup = list_empty(&khugepaged_scan.mm_head);
2015 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2016 spin_unlock(&khugepaged_mm_lock);
2017
2018 atomic_inc(&mm->mm_count);
2019 if (wakeup)
2020 wake_up_interruptible(&khugepaged_wait);
2021
2022 return 0;
2023}
2024
2025int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2026{
2027 unsigned long hstart, hend;
2028 if (!vma->anon_vma)
2029 /*
2030 * Not yet faulted in so we will register later in the
2031 * page fault if needed.
2032 */
2033 return 0;
2034 if (vma->vm_ops)
2035 /* khugepaged not yet working on file or special mappings */
2036 return 0;
2037 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2038 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2039 hend = vma->vm_end & HPAGE_PMD_MASK;
2040 if (hstart < hend)
2041 return khugepaged_enter(vma);
2042 return 0;
2043}
2044
2045void __khugepaged_exit(struct mm_struct *mm)
2046{
2047 struct mm_slot *mm_slot;
2048 int free = 0;
2049
2050 spin_lock(&khugepaged_mm_lock);
2051 mm_slot = get_mm_slot(mm);
2052 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2053 hash_del(&mm_slot->hash);
2054 list_del(&mm_slot->mm_node);
2055 free = 1;
2056 }
2057 spin_unlock(&khugepaged_mm_lock);
2058
2059 if (free) {
2060 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2061 free_mm_slot(mm_slot);
2062 mmdrop(mm);
2063 } else if (mm_slot) {
2064 /*
2065 * This is required to serialize against
2066 * khugepaged_test_exit() (which is guaranteed to run
2067 * under mmap sem read mode). Stop here (after we
2068 * return all pagetables will be destroyed) until
2069 * khugepaged has finished working on the pagetables
2070 * under the mmap_sem.
2071 */
2072 down_write(&mm->mmap_sem);
2073 up_write(&mm->mmap_sem);
2074 }
2075}
2076
2077static void release_pte_page(struct page *page)
2078{
2079 /* 0 stands for page_is_file_cache(page) == false */
2080 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2081 unlock_page(page);
2082 putback_lru_page(page);
2083}
2084
2085static void release_pte_pages(pte_t *pte, pte_t *_pte)
2086{
2087 while (--_pte >= pte) {
2088 pte_t pteval = *_pte;
2089 if (!pte_none(pteval))
2090 release_pte_page(pte_page(pteval));
2091 }
2092}
2093
2094static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2095 unsigned long address,
2096 pte_t *pte)
2097{
2098 struct page *page;
2099 pte_t *_pte;
2100 int referenced = 0, none = 0;
2101 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2102 _pte++, address += PAGE_SIZE) {
2103 pte_t pteval = *_pte;
2104 if (pte_none(pteval)) {
2105 if (++none <= khugepaged_max_ptes_none)
2106 continue;
2107 else
2108 goto out;
2109 }
2110 if (!pte_present(pteval) || !pte_write(pteval))
2111 goto out;
2112 page = vm_normal_page(vma, address, pteval);
2113 if (unlikely(!page))
2114 goto out;
2115
2116 VM_BUG_ON_PAGE(PageCompound(page), page);
2117 VM_BUG_ON_PAGE(!PageAnon(page), page);
2118 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2119
2120 /* cannot use mapcount: can't collapse if there's a gup pin */
2121 if (page_count(page) != 1)
2122 goto out;
2123 /*
2124 * We can do it before isolate_lru_page because the
2125 * page can't be freed from under us. NOTE: PG_lock
2126 * is needed to serialize against split_huge_page
2127 * when invoked from the VM.
2128 */
2129 if (!trylock_page(page))
2130 goto out;
2131 /*
2132 * Isolate the page to avoid collapsing an hugepage
2133 * currently in use by the VM.
2134 */
2135 if (isolate_lru_page(page)) {
2136 unlock_page(page);
2137 goto out;
2138 }
2139 /* 0 stands for page_is_file_cache(page) == false */
2140 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2141 VM_BUG_ON_PAGE(!PageLocked(page), page);
2142 VM_BUG_ON_PAGE(PageLRU(page), page);
2143
2144 /* If there is no mapped pte young don't collapse the page */
2145 if (pte_young(pteval) || PageReferenced(page) ||
2146 mmu_notifier_test_young(vma->vm_mm, address))
2147 referenced = 1;
2148 }
2149 if (likely(referenced))
2150 return 1;
2151out:
2152 release_pte_pages(pte, _pte);
2153 return 0;
2154}
2155
2156static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2157 struct vm_area_struct *vma,
2158 unsigned long address,
2159 spinlock_t *ptl)
2160{
2161 pte_t *_pte;
2162 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2163 pte_t pteval = *_pte;
2164 struct page *src_page;
2165
2166 if (pte_none(pteval)) {
2167 clear_user_highpage(page, address);
2168 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2169 } else {
2170 src_page = pte_page(pteval);
2171 copy_user_highpage(page, src_page, address, vma);
2172 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2173 release_pte_page(src_page);
2174 /*
2175 * ptl mostly unnecessary, but preempt has to
2176 * be disabled to update the per-cpu stats
2177 * inside page_remove_rmap().
2178 */
2179 spin_lock(ptl);
2180 /*
2181 * paravirt calls inside pte_clear here are
2182 * superfluous.
2183 */
2184 pte_clear(vma->vm_mm, address, _pte);
2185 page_remove_rmap(src_page);
2186 spin_unlock(ptl);
2187 free_page_and_swap_cache(src_page);
2188 }
2189
2190 address += PAGE_SIZE;
2191 page++;
2192 }
2193}
2194
2195static void khugepaged_alloc_sleep(void)
2196{
2197 wait_event_freezable_timeout(khugepaged_wait, false,
2198 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2199}
2200
2201static int khugepaged_node_load[MAX_NUMNODES];
2202
2203#ifdef CONFIG_NUMA
2204static int khugepaged_find_target_node(void)
2205{
2206 static int last_khugepaged_target_node = NUMA_NO_NODE;
2207 int nid, target_node = 0, max_value = 0;
2208
2209 /* find first node with max normal pages hit */
2210 for (nid = 0; nid < MAX_NUMNODES; nid++)
2211 if (khugepaged_node_load[nid] > max_value) {
2212 max_value = khugepaged_node_load[nid];
2213 target_node = nid;
2214 }
2215
2216 /* do some balance if several nodes have the same hit record */
2217 if (target_node <= last_khugepaged_target_node)
2218 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2219 nid++)
2220 if (max_value == khugepaged_node_load[nid]) {
2221 target_node = nid;
2222 break;
2223 }
2224
2225 last_khugepaged_target_node = target_node;
2226 return target_node;
2227}
2228
2229static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2230{
2231 if (IS_ERR(*hpage)) {
2232 if (!*wait)
2233 return false;
2234
2235 *wait = false;
2236 *hpage = NULL;
2237 khugepaged_alloc_sleep();
2238 } else if (*hpage) {
2239 put_page(*hpage);
2240 *hpage = NULL;
2241 }
2242
2243 return true;
2244}
2245
2246static struct page
2247*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2248 struct vm_area_struct *vma, unsigned long address,
2249 int node)
2250{
2251 VM_BUG_ON_PAGE(*hpage, *hpage);
2252 /*
2253 * Allocate the page while the vma is still valid and under
2254 * the mmap_sem read mode so there is no memory allocation
2255 * later when we take the mmap_sem in write mode. This is more
2256 * friendly behavior (OTOH it may actually hide bugs) to
2257 * filesystems in userland with daemons allocating memory in
2258 * the userland I/O paths. Allocating memory with the
2259 * mmap_sem in read mode is good idea also to allow greater
2260 * scalability.
2261 */
2262 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2263 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2264 /*
2265 * After allocating the hugepage, release the mmap_sem read lock in
2266 * preparation for taking it in write mode.
2267 */
2268 up_read(&mm->mmap_sem);
2269 if (unlikely(!*hpage)) {
2270 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2271 *hpage = ERR_PTR(-ENOMEM);
2272 return NULL;
2273 }
2274
2275 count_vm_event(THP_COLLAPSE_ALLOC);
2276 return *hpage;
2277}
2278#else
2279static int khugepaged_find_target_node(void)
2280{
2281 return 0;
2282}
2283
2284static inline struct page *alloc_hugepage(int defrag)
2285{
2286 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2287 HPAGE_PMD_ORDER);
2288}
2289
2290static struct page *khugepaged_alloc_hugepage(bool *wait)
2291{
2292 struct page *hpage;
2293
2294 do {
2295 hpage = alloc_hugepage(khugepaged_defrag());
2296 if (!hpage) {
2297 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2298 if (!*wait)
2299 return NULL;
2300
2301 *wait = false;
2302 khugepaged_alloc_sleep();
2303 } else
2304 count_vm_event(THP_COLLAPSE_ALLOC);
2305 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2306
2307 return hpage;
2308}
2309
2310static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2311{
2312 if (!*hpage)
2313 *hpage = khugepaged_alloc_hugepage(wait);
2314
2315 if (unlikely(!*hpage))
2316 return false;
2317
2318 return true;
2319}
2320
2321static struct page
2322*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2323 struct vm_area_struct *vma, unsigned long address,
2324 int node)
2325{
2326 up_read(&mm->mmap_sem);
2327 VM_BUG_ON(!*hpage);
2328 return *hpage;
2329}
2330#endif
2331
2332static bool hugepage_vma_check(struct vm_area_struct *vma)
2333{
2334 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2335 (vma->vm_flags & VM_NOHUGEPAGE))
2336 return false;
2337
2338 if (!vma->anon_vma || vma->vm_ops)
2339 return false;
2340 if (is_vma_temporary_stack(vma))
2341 return false;
2342 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2343 return true;
2344}
2345
2346static void collapse_huge_page(struct mm_struct *mm,
2347 unsigned long address,
2348 struct page **hpage,
2349 struct vm_area_struct *vma,
2350 int node)
2351{
2352 pmd_t *pmd, _pmd;
2353 pte_t *pte;
2354 pgtable_t pgtable;
2355 struct page *new_page;
2356 spinlock_t *pmd_ptl, *pte_ptl;
2357 int isolated;
2358 unsigned long hstart, hend;
2359 unsigned long mmun_start; /* For mmu_notifiers */
2360 unsigned long mmun_end; /* For mmu_notifiers */
2361
2362 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2363
2364 /* release the mmap_sem read lock. */
2365 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2366 if (!new_page)
2367 return;
2368
2369 if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)))
2370 return;
2371
2372 /*
2373 * Prevent all access to pagetables with the exception of
2374 * gup_fast later hanlded by the ptep_clear_flush and the VM
2375 * handled by the anon_vma lock + PG_lock.
2376 */
2377 down_write(&mm->mmap_sem);
2378 if (unlikely(khugepaged_test_exit(mm)))
2379 goto out;
2380
2381 vma = find_vma(mm, address);
2382 if (!vma)
2383 goto out;
2384 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2385 hend = vma->vm_end & HPAGE_PMD_MASK;
2386 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2387 goto out;
2388 if (!hugepage_vma_check(vma))
2389 goto out;
2390 pmd = mm_find_pmd(mm, address);
2391 if (!pmd)
2392 goto out;
2393 if (pmd_trans_huge(*pmd))
2394 goto out;
2395
2396 anon_vma_lock_write(vma->anon_vma);
2397
2398 pte = pte_offset_map(pmd, address);
2399 pte_ptl = pte_lockptr(mm, pmd);
2400
2401 mmun_start = address;
2402 mmun_end = address + HPAGE_PMD_SIZE;
2403 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2404 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2405 /*
2406 * After this gup_fast can't run anymore. This also removes
2407 * any huge TLB entry from the CPU so we won't allow
2408 * huge and small TLB entries for the same virtual address
2409 * to avoid the risk of CPU bugs in that area.
2410 */
2411 _pmd = pmdp_clear_flush(vma, address, pmd);
2412 spin_unlock(pmd_ptl);
2413 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2414
2415 spin_lock(pte_ptl);
2416 isolated = __collapse_huge_page_isolate(vma, address, pte);
2417 spin_unlock(pte_ptl);
2418
2419 if (unlikely(!isolated)) {
2420 pte_unmap(pte);
2421 spin_lock(pmd_ptl);
2422 BUG_ON(!pmd_none(*pmd));
2423 /*
2424 * We can only use set_pmd_at when establishing
2425 * hugepmds and never for establishing regular pmds that
2426 * points to regular pagetables. Use pmd_populate for that
2427 */
2428 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2429 spin_unlock(pmd_ptl);
2430 anon_vma_unlock_write(vma->anon_vma);
2431 goto out;
2432 }
2433
2434 /*
2435 * All pages are isolated and locked so anon_vma rmap
2436 * can't run anymore.
2437 */
2438 anon_vma_unlock_write(vma->anon_vma);
2439
2440 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2441 pte_unmap(pte);
2442 __SetPageUptodate(new_page);
2443 pgtable = pmd_pgtable(_pmd);
2444
2445 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2446 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2447
2448 /*
2449 * spin_lock() below is not the equivalent of smp_wmb(), so
2450 * this is needed to avoid the copy_huge_page writes to become
2451 * visible after the set_pmd_at() write.
2452 */
2453 smp_wmb();
2454
2455 spin_lock(pmd_ptl);
2456 BUG_ON(!pmd_none(*pmd));
2457 page_add_new_anon_rmap(new_page, vma, address);
2458 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2459 set_pmd_at(mm, address, pmd, _pmd);
2460 update_mmu_cache_pmd(vma, address, pmd);
2461 spin_unlock(pmd_ptl);
2462
2463 *hpage = NULL;
2464
2465 khugepaged_pages_collapsed++;
2466out_up_write:
2467 up_write(&mm->mmap_sem);
2468 return;
2469
2470out:
2471 mem_cgroup_uncharge_page(new_page);
2472 goto out_up_write;
2473}
2474
2475static int khugepaged_scan_pmd(struct mm_struct *mm,
2476 struct vm_area_struct *vma,
2477 unsigned long address,
2478 struct page **hpage)
2479{
2480 pmd_t *pmd;
2481 pte_t *pte, *_pte;
2482 int ret = 0, referenced = 0, none = 0;
2483 struct page *page;
2484 unsigned long _address;
2485 spinlock_t *ptl;
2486 int node = NUMA_NO_NODE;
2487
2488 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2489
2490 pmd = mm_find_pmd(mm, address);
2491 if (!pmd)
2492 goto out;
2493 if (pmd_trans_huge(*pmd))
2494 goto out;
2495
2496 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2497 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2498 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2499 _pte++, _address += PAGE_SIZE) {
2500 pte_t pteval = *_pte;
2501 if (pte_none(pteval)) {
2502 if (++none <= khugepaged_max_ptes_none)
2503 continue;
2504 else
2505 goto out_unmap;
2506 }
2507 if (!pte_present(pteval) || !pte_write(pteval))
2508 goto out_unmap;
2509 page = vm_normal_page(vma, _address, pteval);
2510 if (unlikely(!page))
2511 goto out_unmap;
2512 /*
2513 * Record which node the original page is from and save this
2514 * information to khugepaged_node_load[].
2515 * Khupaged will allocate hugepage from the node has the max
2516 * hit record.
2517 */
2518 node = page_to_nid(page);
2519 khugepaged_node_load[node]++;
2520 VM_BUG_ON_PAGE(PageCompound(page), page);
2521 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2522 goto out_unmap;
2523 /* cannot use mapcount: can't collapse if there's a gup pin */
2524 if (page_count(page) != 1)
2525 goto out_unmap;
2526 if (pte_young(pteval) || PageReferenced(page) ||
2527 mmu_notifier_test_young(vma->vm_mm, address))
2528 referenced = 1;
2529 }
2530 if (referenced)
2531 ret = 1;
2532out_unmap:
2533 pte_unmap_unlock(pte, ptl);
2534 if (ret) {
2535 node = khugepaged_find_target_node();
2536 /* collapse_huge_page will return with the mmap_sem released */
2537 collapse_huge_page(mm, address, hpage, vma, node);
2538 }
2539out:
2540 return ret;
2541}
2542
2543static void collect_mm_slot(struct mm_slot *mm_slot)
2544{
2545 struct mm_struct *mm = mm_slot->mm;
2546
2547 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2548
2549 if (khugepaged_test_exit(mm)) {
2550 /* free mm_slot */
2551 hash_del(&mm_slot->hash);
2552 list_del(&mm_slot->mm_node);
2553
2554 /*
2555 * Not strictly needed because the mm exited already.
2556 *
2557 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2558 */
2559
2560 /* khugepaged_mm_lock actually not necessary for the below */
2561 free_mm_slot(mm_slot);
2562 mmdrop(mm);
2563 }
2564}
2565
2566static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2567 struct page **hpage)
2568 __releases(&khugepaged_mm_lock)
2569 __acquires(&khugepaged_mm_lock)
2570{
2571 struct mm_slot *mm_slot;
2572 struct mm_struct *mm;
2573 struct vm_area_struct *vma;
2574 int progress = 0;
2575
2576 VM_BUG_ON(!pages);
2577 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2578
2579 if (khugepaged_scan.mm_slot)
2580 mm_slot = khugepaged_scan.mm_slot;
2581 else {
2582 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2583 struct mm_slot, mm_node);
2584 khugepaged_scan.address = 0;
2585 khugepaged_scan.mm_slot = mm_slot;
2586 }
2587 spin_unlock(&khugepaged_mm_lock);
2588
2589 mm = mm_slot->mm;
2590 down_read(&mm->mmap_sem);
2591 if (unlikely(khugepaged_test_exit(mm)))
2592 vma = NULL;
2593 else
2594 vma = find_vma(mm, khugepaged_scan.address);
2595
2596 progress++;
2597 for (; vma; vma = vma->vm_next) {
2598 unsigned long hstart, hend;
2599
2600 cond_resched();
2601 if (unlikely(khugepaged_test_exit(mm))) {
2602 progress++;
2603 break;
2604 }
2605 if (!hugepage_vma_check(vma)) {
2606skip:
2607 progress++;
2608 continue;
2609 }
2610 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2611 hend = vma->vm_end & HPAGE_PMD_MASK;
2612 if (hstart >= hend)
2613 goto skip;
2614 if (khugepaged_scan.address > hend)
2615 goto skip;
2616 if (khugepaged_scan.address < hstart)
2617 khugepaged_scan.address = hstart;
2618 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2619
2620 while (khugepaged_scan.address < hend) {
2621 int ret;
2622 cond_resched();
2623 if (unlikely(khugepaged_test_exit(mm)))
2624 goto breakouterloop;
2625
2626 VM_BUG_ON(khugepaged_scan.address < hstart ||
2627 khugepaged_scan.address + HPAGE_PMD_SIZE >
2628 hend);
2629 ret = khugepaged_scan_pmd(mm, vma,
2630 khugepaged_scan.address,
2631 hpage);
2632 /* move to next address */
2633 khugepaged_scan.address += HPAGE_PMD_SIZE;
2634 progress += HPAGE_PMD_NR;
2635 if (ret)
2636 /* we released mmap_sem so break loop */
2637 goto breakouterloop_mmap_sem;
2638 if (progress >= pages)
2639 goto breakouterloop;
2640 }
2641 }
2642breakouterloop:
2643 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2644breakouterloop_mmap_sem:
2645
2646 spin_lock(&khugepaged_mm_lock);
2647 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2648 /*
2649 * Release the current mm_slot if this mm is about to die, or
2650 * if we scanned all vmas of this mm.
2651 */
2652 if (khugepaged_test_exit(mm) || !vma) {
2653 /*
2654 * Make sure that if mm_users is reaching zero while
2655 * khugepaged runs here, khugepaged_exit will find
2656 * mm_slot not pointing to the exiting mm.
2657 */
2658 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2659 khugepaged_scan.mm_slot = list_entry(
2660 mm_slot->mm_node.next,
2661 struct mm_slot, mm_node);
2662 khugepaged_scan.address = 0;
2663 } else {
2664 khugepaged_scan.mm_slot = NULL;
2665 khugepaged_full_scans++;
2666 }
2667
2668 collect_mm_slot(mm_slot);
2669 }
2670
2671 return progress;
2672}
2673
2674static int khugepaged_has_work(void)
2675{
2676 return !list_empty(&khugepaged_scan.mm_head) &&
2677 khugepaged_enabled();
2678}
2679
2680static int khugepaged_wait_event(void)
2681{
2682 return !list_empty(&khugepaged_scan.mm_head) ||
2683 kthread_should_stop();
2684}
2685
2686static void khugepaged_do_scan(void)
2687{
2688 struct page *hpage = NULL;
2689 unsigned int progress = 0, pass_through_head = 0;
2690 unsigned int pages = khugepaged_pages_to_scan;
2691 bool wait = true;
2692
2693 barrier(); /* write khugepaged_pages_to_scan to local stack */
2694
2695 while (progress < pages) {
2696 if (!khugepaged_prealloc_page(&hpage, &wait))
2697 break;
2698
2699 cond_resched();
2700
2701 if (unlikely(kthread_should_stop() || freezing(current)))
2702 break;
2703
2704 spin_lock(&khugepaged_mm_lock);
2705 if (!khugepaged_scan.mm_slot)
2706 pass_through_head++;
2707 if (khugepaged_has_work() &&
2708 pass_through_head < 2)
2709 progress += khugepaged_scan_mm_slot(pages - progress,
2710 &hpage);
2711 else
2712 progress = pages;
2713 spin_unlock(&khugepaged_mm_lock);
2714 }
2715
2716 if (!IS_ERR_OR_NULL(hpage))
2717 put_page(hpage);
2718}
2719
2720static void khugepaged_wait_work(void)
2721{
2722 try_to_freeze();
2723
2724 if (khugepaged_has_work()) {
2725 if (!khugepaged_scan_sleep_millisecs)
2726 return;
2727
2728 wait_event_freezable_timeout(khugepaged_wait,
2729 kthread_should_stop(),
2730 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2731 return;
2732 }
2733
2734 if (khugepaged_enabled())
2735 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2736}
2737
2738static int khugepaged(void *none)
2739{
2740 struct mm_slot *mm_slot;
2741
2742 set_freezable();
2743 set_user_nice(current, 19);
2744
2745 while (!kthread_should_stop()) {
2746 khugepaged_do_scan();
2747 khugepaged_wait_work();
2748 }
2749
2750 spin_lock(&khugepaged_mm_lock);
2751 mm_slot = khugepaged_scan.mm_slot;
2752 khugepaged_scan.mm_slot = NULL;
2753 if (mm_slot)
2754 collect_mm_slot(mm_slot);
2755 spin_unlock(&khugepaged_mm_lock);
2756 return 0;
2757}
2758
2759static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2760 unsigned long haddr, pmd_t *pmd)
2761{
2762 struct mm_struct *mm = vma->vm_mm;
2763 pgtable_t pgtable;
2764 pmd_t _pmd;
2765 int i;
2766
2767 pmdp_clear_flush(vma, haddr, pmd);
2768 /* leave pmd empty until pte is filled */
2769
2770 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2771 pmd_populate(mm, &_pmd, pgtable);
2772
2773 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2774 pte_t *pte, entry;
2775 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2776 entry = pte_mkspecial(entry);
2777 pte = pte_offset_map(&_pmd, haddr);
2778 VM_BUG_ON(!pte_none(*pte));
2779 set_pte_at(mm, haddr, pte, entry);
2780 pte_unmap(pte);
2781 }
2782 smp_wmb(); /* make pte visible before pmd */
2783 pmd_populate(mm, pmd, pgtable);
2784 put_huge_zero_page();
2785}
2786
2787void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2788 pmd_t *pmd)
2789{
2790 spinlock_t *ptl;
2791 struct page *page;
2792 struct mm_struct *mm = vma->vm_mm;
2793 unsigned long haddr = address & HPAGE_PMD_MASK;
2794 unsigned long mmun_start; /* For mmu_notifiers */
2795 unsigned long mmun_end; /* For mmu_notifiers */
2796
2797 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2798
2799 mmun_start = haddr;
2800 mmun_end = haddr + HPAGE_PMD_SIZE;
2801again:
2802 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2803 ptl = pmd_lock(mm, pmd);
2804 if (unlikely(!pmd_trans_huge(*pmd))) {
2805 spin_unlock(ptl);
2806 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2807 return;
2808 }
2809 if (is_huge_zero_pmd(*pmd)) {
2810 __split_huge_zero_page_pmd(vma, haddr, pmd);
2811 spin_unlock(ptl);
2812 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2813 return;
2814 }
2815 page = pmd_page(*pmd);
2816 VM_BUG_ON_PAGE(!page_count(page), page);
2817 get_page(page);
2818 spin_unlock(ptl);
2819 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2820
2821 split_huge_page(page);
2822
2823 put_page(page);
2824
2825 /*
2826 * We don't always have down_write of mmap_sem here: a racing
2827 * do_huge_pmd_wp_page() might have copied-on-write to another
2828 * huge page before our split_huge_page() got the anon_vma lock.
2829 */
2830 if (unlikely(pmd_trans_huge(*pmd)))
2831 goto again;
2832}
2833
2834void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2835 pmd_t *pmd)
2836{
2837 struct vm_area_struct *vma;
2838
2839 vma = find_vma(mm, address);
2840 BUG_ON(vma == NULL);
2841 split_huge_page_pmd(vma, address, pmd);
2842}
2843
2844static void split_huge_page_address(struct mm_struct *mm,
2845 unsigned long address)
2846{
2847 pmd_t *pmd;
2848
2849 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2850
2851 pmd = mm_find_pmd(mm, address);
2852 if (!pmd)
2853 return;
2854 /*
2855 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2856 * materialize from under us.
2857 */
2858 split_huge_page_pmd_mm(mm, address, pmd);
2859}
2860
2861void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2862 unsigned long start,
2863 unsigned long end,
2864 long adjust_next)
2865{
2866 /*
2867 * If the new start address isn't hpage aligned and it could
2868 * previously contain an hugepage: check if we need to split
2869 * an huge pmd.
2870 */
2871 if (start & ~HPAGE_PMD_MASK &&
2872 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2873 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2874 split_huge_page_address(vma->vm_mm, start);
2875
2876 /*
2877 * If the new end address isn't hpage aligned and it could
2878 * previously contain an hugepage: check if we need to split
2879 * an huge pmd.
2880 */
2881 if (end & ~HPAGE_PMD_MASK &&
2882 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2883 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2884 split_huge_page_address(vma->vm_mm, end);
2885
2886 /*
2887 * If we're also updating the vma->vm_next->vm_start, if the new
2888 * vm_next->vm_start isn't page aligned and it could previously
2889 * contain an hugepage: check if we need to split an huge pmd.
2890 */
2891 if (adjust_next > 0) {
2892 struct vm_area_struct *next = vma->vm_next;
2893 unsigned long nstart = next->vm_start;
2894 nstart += adjust_next << PAGE_SHIFT;
2895 if (nstart & ~HPAGE_PMD_MASK &&
2896 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2897 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2898 split_huge_page_address(next->vm_mm, nstart);
2899 }
2900}