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