Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
b46e756f KS |
2 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
3 | ||
4 | #include <linux/mm.h> | |
5 | #include <linux/sched.h> | |
6e84f315 | 6 | #include <linux/sched/mm.h> |
f7ccbae4 | 7 | #include <linux/sched/coredump.h> |
b46e756f KS |
8 | #include <linux/mmu_notifier.h> |
9 | #include <linux/rmap.h> | |
10 | #include <linux/swap.h> | |
11 | #include <linux/mm_inline.h> | |
12 | #include <linux/kthread.h> | |
13 | #include <linux/khugepaged.h> | |
14 | #include <linux/freezer.h> | |
15 | #include <linux/mman.h> | |
16 | #include <linux/hashtable.h> | |
17 | #include <linux/userfaultfd_k.h> | |
18 | #include <linux/page_idle.h> | |
80110bbf | 19 | #include <linux/page_table_check.h> |
b46e756f | 20 | #include <linux/swapops.h> |
f3f0e1d2 | 21 | #include <linux/shmem_fs.h> |
b46e756f KS |
22 | |
23 | #include <asm/tlb.h> | |
24 | #include <asm/pgalloc.h> | |
25 | #include "internal.h" | |
b26e2701 | 26 | #include "mm_slot.h" |
b46e756f KS |
27 | |
28 | enum scan_result { | |
29 | SCAN_FAIL, | |
30 | SCAN_SUCCEED, | |
31 | SCAN_PMD_NULL, | |
34488399 | 32 | SCAN_PMD_NONE, |
50722804 | 33 | SCAN_PMD_MAPPED, |
b46e756f | 34 | SCAN_EXCEED_NONE_PTE, |
71a2c112 KS |
35 | SCAN_EXCEED_SWAP_PTE, |
36 | SCAN_EXCEED_SHARED_PTE, | |
b46e756f | 37 | SCAN_PTE_NON_PRESENT, |
e1e267c7 | 38 | SCAN_PTE_UFFD_WP, |
58ac9a89 | 39 | SCAN_PTE_MAPPED_HUGEPAGE, |
b46e756f | 40 | SCAN_PAGE_RO, |
0db501f7 | 41 | SCAN_LACK_REFERENCED_PAGE, |
b46e756f KS |
42 | SCAN_PAGE_NULL, |
43 | SCAN_SCAN_ABORT, | |
44 | SCAN_PAGE_COUNT, | |
45 | SCAN_PAGE_LRU, | |
46 | SCAN_PAGE_LOCK, | |
47 | SCAN_PAGE_ANON, | |
48 | SCAN_PAGE_COMPOUND, | |
49 | SCAN_ANY_PROCESS, | |
50 | SCAN_VMA_NULL, | |
51 | SCAN_VMA_CHECK, | |
52 | SCAN_ADDRESS_RANGE, | |
b46e756f KS |
53 | SCAN_DEL_PAGE_LRU, |
54 | SCAN_ALLOC_HUGE_PAGE_FAIL, | |
55 | SCAN_CGROUP_CHARGE_FAIL, | |
f3f0e1d2 | 56 | SCAN_TRUNCATED, |
99cb0dbd | 57 | SCAN_PAGE_HAS_PRIVATE, |
2ce0bdfe | 58 | SCAN_STORE_FAILED, |
98c76c9f | 59 | SCAN_COPY_MC, |
ac492b9c | 60 | SCAN_PAGE_FILLED, |
b46e756f KS |
61 | }; |
62 | ||
63 | #define CREATE_TRACE_POINTS | |
64 | #include <trace/events/huge_memory.h> | |
65 | ||
4aab2be0 VB |
66 | static struct task_struct *khugepaged_thread __read_mostly; |
67 | static DEFINE_MUTEX(khugepaged_mutex); | |
68 | ||
b46e756f KS |
69 | /* default scan 8*512 pte (or vmas) every 30 second */ |
70 | static unsigned int khugepaged_pages_to_scan __read_mostly; | |
71 | static unsigned int khugepaged_pages_collapsed; | |
72 | static unsigned int khugepaged_full_scans; | |
73 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
74 | /* during fragmentation poll the hugepage allocator once every minute */ | |
75 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
76 | static unsigned long khugepaged_sleep_expire; | |
77 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
78 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
79 | /* | |
80 | * default collapse hugepages if there is at least one pte mapped like | |
81 | * it would have happened if the vma was large enough during page | |
82 | * fault. | |
d8ea7cc8 ZK |
83 | * |
84 | * Note that these are only respected if collapse was initiated by khugepaged. | |
b46e756f KS |
85 | */ |
86 | static unsigned int khugepaged_max_ptes_none __read_mostly; | |
87 | static unsigned int khugepaged_max_ptes_swap __read_mostly; | |
71a2c112 | 88 | static unsigned int khugepaged_max_ptes_shared __read_mostly; |
b46e756f KS |
89 | |
90 | #define MM_SLOTS_HASH_BITS 10 | |
e1ad3e66 | 91 | static DEFINE_READ_MOSTLY_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); |
b46e756f KS |
92 | |
93 | static struct kmem_cache *mm_slot_cache __read_mostly; | |
94 | ||
27e1f827 SL |
95 | #define MAX_PTE_MAPPED_THP 8 |
96 | ||
34d6b470 | 97 | struct collapse_control { |
d8ea7cc8 ZK |
98 | bool is_khugepaged; |
99 | ||
34d6b470 ZK |
100 | /* Num pages scanned per node */ |
101 | u32 node_load[MAX_NUMNODES]; | |
102 | ||
e031ff96 YS |
103 | /* nodemask for allocation fallback */ |
104 | nodemask_t alloc_nmask; | |
34d6b470 ZK |
105 | }; |
106 | ||
b46e756f | 107 | /** |
b26e2701 QZ |
108 | * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned |
109 | * @slot: hash lookup from mm to mm_slot | |
336e6b53 AS |
110 | * @nr_pte_mapped_thp: number of pte mapped THP |
111 | * @pte_mapped_thp: address array corresponding pte mapped THP | |
b46e756f | 112 | */ |
b26e2701 QZ |
113 | struct khugepaged_mm_slot { |
114 | struct mm_slot slot; | |
27e1f827 SL |
115 | |
116 | /* pte-mapped THP in this mm */ | |
117 | int nr_pte_mapped_thp; | |
118 | unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP]; | |
b46e756f KS |
119 | }; |
120 | ||
121 | /** | |
122 | * struct khugepaged_scan - cursor for scanning | |
123 | * @mm_head: the head of the mm list to scan | |
124 | * @mm_slot: the current mm_slot we are scanning | |
125 | * @address: the next address inside that to be scanned | |
126 | * | |
127 | * There is only the one khugepaged_scan instance of this cursor structure. | |
128 | */ | |
129 | struct khugepaged_scan { | |
130 | struct list_head mm_head; | |
b26e2701 | 131 | struct khugepaged_mm_slot *mm_slot; |
b46e756f KS |
132 | unsigned long address; |
133 | }; | |
134 | ||
135 | static struct khugepaged_scan khugepaged_scan = { | |
136 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), | |
137 | }; | |
138 | ||
e1465d12 | 139 | #ifdef CONFIG_SYSFS |
b46e756f KS |
140 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, |
141 | struct kobj_attribute *attr, | |
142 | char *buf) | |
143 | { | |
ae7a927d | 144 | return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs); |
b46e756f KS |
145 | } |
146 | ||
147 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
148 | struct kobj_attribute *attr, | |
149 | const char *buf, size_t count) | |
150 | { | |
dfefd226 | 151 | unsigned int msecs; |
b46e756f KS |
152 | int err; |
153 | ||
dfefd226 AD |
154 | err = kstrtouint(buf, 10, &msecs); |
155 | if (err) | |
b46e756f KS |
156 | return -EINVAL; |
157 | ||
158 | khugepaged_scan_sleep_millisecs = msecs; | |
159 | khugepaged_sleep_expire = 0; | |
160 | wake_up_interruptible(&khugepaged_wait); | |
161 | ||
162 | return count; | |
163 | } | |
164 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
6dcdc94d | 165 | __ATTR_RW(scan_sleep_millisecs); |
b46e756f KS |
166 | |
167 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
168 | struct kobj_attribute *attr, | |
169 | char *buf) | |
170 | { | |
ae7a927d | 171 | return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs); |
b46e756f KS |
172 | } |
173 | ||
174 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
175 | struct kobj_attribute *attr, | |
176 | const char *buf, size_t count) | |
177 | { | |
dfefd226 | 178 | unsigned int msecs; |
b46e756f KS |
179 | int err; |
180 | ||
dfefd226 AD |
181 | err = kstrtouint(buf, 10, &msecs); |
182 | if (err) | |
b46e756f KS |
183 | return -EINVAL; |
184 | ||
185 | khugepaged_alloc_sleep_millisecs = msecs; | |
186 | khugepaged_sleep_expire = 0; | |
187 | wake_up_interruptible(&khugepaged_wait); | |
188 | ||
189 | return count; | |
190 | } | |
191 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
6dcdc94d | 192 | __ATTR_RW(alloc_sleep_millisecs); |
b46e756f KS |
193 | |
194 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
195 | struct kobj_attribute *attr, | |
196 | char *buf) | |
197 | { | |
ae7a927d | 198 | return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan); |
b46e756f KS |
199 | } |
200 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
201 | struct kobj_attribute *attr, | |
202 | const char *buf, size_t count) | |
203 | { | |
dfefd226 | 204 | unsigned int pages; |
b46e756f | 205 | int err; |
b46e756f | 206 | |
dfefd226 AD |
207 | err = kstrtouint(buf, 10, &pages); |
208 | if (err || !pages) | |
b46e756f KS |
209 | return -EINVAL; |
210 | ||
211 | khugepaged_pages_to_scan = pages; | |
212 | ||
213 | return count; | |
214 | } | |
215 | static struct kobj_attribute pages_to_scan_attr = | |
6dcdc94d | 216 | __ATTR_RW(pages_to_scan); |
b46e756f KS |
217 | |
218 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
219 | struct kobj_attribute *attr, | |
220 | char *buf) | |
221 | { | |
ae7a927d | 222 | return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed); |
b46e756f KS |
223 | } |
224 | static struct kobj_attribute pages_collapsed_attr = | |
225 | __ATTR_RO(pages_collapsed); | |
226 | ||
227 | static ssize_t full_scans_show(struct kobject *kobj, | |
228 | struct kobj_attribute *attr, | |
229 | char *buf) | |
230 | { | |
ae7a927d | 231 | return sysfs_emit(buf, "%u\n", khugepaged_full_scans); |
b46e756f KS |
232 | } |
233 | static struct kobj_attribute full_scans_attr = | |
234 | __ATTR_RO(full_scans); | |
235 | ||
6dcdc94d ML |
236 | static ssize_t defrag_show(struct kobject *kobj, |
237 | struct kobj_attribute *attr, char *buf) | |
b46e756f KS |
238 | { |
239 | return single_hugepage_flag_show(kobj, attr, buf, | |
ae7a927d | 240 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); |
b46e756f | 241 | } |
6dcdc94d ML |
242 | static ssize_t defrag_store(struct kobject *kobj, |
243 | struct kobj_attribute *attr, | |
244 | const char *buf, size_t count) | |
b46e756f KS |
245 | { |
246 | return single_hugepage_flag_store(kobj, attr, buf, count, | |
247 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
248 | } | |
249 | static struct kobj_attribute khugepaged_defrag_attr = | |
6dcdc94d | 250 | __ATTR_RW(defrag); |
b46e756f KS |
251 | |
252 | /* | |
253 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
254 | * any unmapped ptes in turn potentially increasing the memory | |
255 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
256 | * reduce the available free memory in the system as it | |
257 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
258 | * free memory in the system during the khugepaged scan. | |
259 | */ | |
6dcdc94d ML |
260 | static ssize_t max_ptes_none_show(struct kobject *kobj, |
261 | struct kobj_attribute *attr, | |
262 | char *buf) | |
b46e756f | 263 | { |
ae7a927d | 264 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none); |
b46e756f | 265 | } |
6dcdc94d ML |
266 | static ssize_t max_ptes_none_store(struct kobject *kobj, |
267 | struct kobj_attribute *attr, | |
268 | const char *buf, size_t count) | |
b46e756f KS |
269 | { |
270 | int err; | |
271 | unsigned long max_ptes_none; | |
272 | ||
273 | err = kstrtoul(buf, 10, &max_ptes_none); | |
36ee2c78 | 274 | if (err || max_ptes_none > HPAGE_PMD_NR - 1) |
b46e756f KS |
275 | return -EINVAL; |
276 | ||
277 | khugepaged_max_ptes_none = max_ptes_none; | |
278 | ||
279 | return count; | |
280 | } | |
281 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
6dcdc94d | 282 | __ATTR_RW(max_ptes_none); |
b46e756f | 283 | |
6dcdc94d ML |
284 | static ssize_t max_ptes_swap_show(struct kobject *kobj, |
285 | struct kobj_attribute *attr, | |
286 | char *buf) | |
b46e756f | 287 | { |
ae7a927d | 288 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap); |
b46e756f KS |
289 | } |
290 | ||
6dcdc94d ML |
291 | static ssize_t max_ptes_swap_store(struct kobject *kobj, |
292 | struct kobj_attribute *attr, | |
293 | const char *buf, size_t count) | |
b46e756f KS |
294 | { |
295 | int err; | |
296 | unsigned long max_ptes_swap; | |
297 | ||
298 | err = kstrtoul(buf, 10, &max_ptes_swap); | |
36ee2c78 | 299 | if (err || max_ptes_swap > HPAGE_PMD_NR - 1) |
b46e756f KS |
300 | return -EINVAL; |
301 | ||
302 | khugepaged_max_ptes_swap = max_ptes_swap; | |
303 | ||
304 | return count; | |
305 | } | |
306 | ||
307 | static struct kobj_attribute khugepaged_max_ptes_swap_attr = | |
6dcdc94d | 308 | __ATTR_RW(max_ptes_swap); |
b46e756f | 309 | |
6dcdc94d ML |
310 | static ssize_t max_ptes_shared_show(struct kobject *kobj, |
311 | struct kobj_attribute *attr, | |
312 | char *buf) | |
71a2c112 | 313 | { |
ae7a927d | 314 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared); |
71a2c112 KS |
315 | } |
316 | ||
6dcdc94d ML |
317 | static ssize_t max_ptes_shared_store(struct kobject *kobj, |
318 | struct kobj_attribute *attr, | |
319 | const char *buf, size_t count) | |
71a2c112 KS |
320 | { |
321 | int err; | |
322 | unsigned long max_ptes_shared; | |
323 | ||
324 | err = kstrtoul(buf, 10, &max_ptes_shared); | |
36ee2c78 | 325 | if (err || max_ptes_shared > HPAGE_PMD_NR - 1) |
71a2c112 KS |
326 | return -EINVAL; |
327 | ||
328 | khugepaged_max_ptes_shared = max_ptes_shared; | |
329 | ||
330 | return count; | |
331 | } | |
332 | ||
333 | static struct kobj_attribute khugepaged_max_ptes_shared_attr = | |
6dcdc94d | 334 | __ATTR_RW(max_ptes_shared); |
71a2c112 | 335 | |
b46e756f KS |
336 | static struct attribute *khugepaged_attr[] = { |
337 | &khugepaged_defrag_attr.attr, | |
338 | &khugepaged_max_ptes_none_attr.attr, | |
71a2c112 KS |
339 | &khugepaged_max_ptes_swap_attr.attr, |
340 | &khugepaged_max_ptes_shared_attr.attr, | |
b46e756f KS |
341 | &pages_to_scan_attr.attr, |
342 | &pages_collapsed_attr.attr, | |
343 | &full_scans_attr.attr, | |
344 | &scan_sleep_millisecs_attr.attr, | |
345 | &alloc_sleep_millisecs_attr.attr, | |
b46e756f KS |
346 | NULL, |
347 | }; | |
348 | ||
349 | struct attribute_group khugepaged_attr_group = { | |
350 | .attrs = khugepaged_attr, | |
351 | .name = "khugepaged", | |
352 | }; | |
e1465d12 | 353 | #endif /* CONFIG_SYSFS */ |
b46e756f | 354 | |
b46e756f KS |
355 | int hugepage_madvise(struct vm_area_struct *vma, |
356 | unsigned long *vm_flags, int advice) | |
357 | { | |
358 | switch (advice) { | |
359 | case MADV_HUGEPAGE: | |
360 | #ifdef CONFIG_S390 | |
361 | /* | |
362 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
363 | * can't handle this properly after s390_enable_sie, so we simply | |
364 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
365 | */ | |
366 | if (mm_has_pgste(vma->vm_mm)) | |
367 | return 0; | |
368 | #endif | |
369 | *vm_flags &= ~VM_NOHUGEPAGE; | |
370 | *vm_flags |= VM_HUGEPAGE; | |
371 | /* | |
372 | * If the vma become good for khugepaged to scan, | |
373 | * register it here without waiting a page fault that | |
374 | * may not happen any time soon. | |
375 | */ | |
c791576c | 376 | khugepaged_enter_vma(vma, *vm_flags); |
b46e756f KS |
377 | break; |
378 | case MADV_NOHUGEPAGE: | |
379 | *vm_flags &= ~VM_HUGEPAGE; | |
380 | *vm_flags |= VM_NOHUGEPAGE; | |
381 | /* | |
382 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
383 | * this vma even if we leave the mm registered in khugepaged if | |
384 | * it got registered before VM_NOHUGEPAGE was set. | |
385 | */ | |
386 | break; | |
387 | } | |
388 | ||
389 | return 0; | |
390 | } | |
391 | ||
392 | int __init khugepaged_init(void) | |
393 | { | |
394 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
b26e2701 QZ |
395 | sizeof(struct khugepaged_mm_slot), |
396 | __alignof__(struct khugepaged_mm_slot), | |
397 | 0, NULL); | |
b46e756f KS |
398 | if (!mm_slot_cache) |
399 | return -ENOMEM; | |
400 | ||
401 | khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; | |
402 | khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; | |
403 | khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; | |
71a2c112 | 404 | khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; |
b46e756f KS |
405 | |
406 | return 0; | |
407 | } | |
408 | ||
409 | void __init khugepaged_destroy(void) | |
410 | { | |
411 | kmem_cache_destroy(mm_slot_cache); | |
412 | } | |
413 | ||
7d2c4385 | 414 | static inline int hpage_collapse_test_exit(struct mm_struct *mm) |
b46e756f | 415 | { |
4d45e75a | 416 | return atomic_read(&mm->mm_users) == 0; |
b46e756f KS |
417 | } |
418 | ||
d2081b2b | 419 | void __khugepaged_enter(struct mm_struct *mm) |
b46e756f | 420 | { |
b26e2701 QZ |
421 | struct khugepaged_mm_slot *mm_slot; |
422 | struct mm_slot *slot; | |
b46e756f KS |
423 | int wakeup; |
424 | ||
16618670 XH |
425 | /* __khugepaged_exit() must not run from under us */ |
426 | VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm); | |
427 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) | |
428 | return; | |
429 | ||
b26e2701 | 430 | mm_slot = mm_slot_alloc(mm_slot_cache); |
b46e756f | 431 | if (!mm_slot) |
d2081b2b | 432 | return; |
b46e756f | 433 | |
b26e2701 QZ |
434 | slot = &mm_slot->slot; |
435 | ||
b46e756f | 436 | spin_lock(&khugepaged_mm_lock); |
b26e2701 | 437 | mm_slot_insert(mm_slots_hash, mm, slot); |
b46e756f KS |
438 | /* |
439 | * Insert just behind the scanning cursor, to let the area settle | |
440 | * down a little. | |
441 | */ | |
442 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
b26e2701 | 443 | list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head); |
b46e756f KS |
444 | spin_unlock(&khugepaged_mm_lock); |
445 | ||
f1f10076 | 446 | mmgrab(mm); |
b46e756f KS |
447 | if (wakeup) |
448 | wake_up_interruptible(&khugepaged_wait); | |
b46e756f KS |
449 | } |
450 | ||
c791576c YS |
451 | void khugepaged_enter_vma(struct vm_area_struct *vma, |
452 | unsigned long vm_flags) | |
b46e756f | 453 | { |
2647d11b | 454 | if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && |
1064026b | 455 | hugepage_flags_enabled()) { |
a7f4e6e4 | 456 | if (hugepage_vma_check(vma, vm_flags, false, false, true)) |
2647d11b YS |
457 | __khugepaged_enter(vma->vm_mm); |
458 | } | |
b46e756f KS |
459 | } |
460 | ||
461 | void __khugepaged_exit(struct mm_struct *mm) | |
462 | { | |
b26e2701 QZ |
463 | struct khugepaged_mm_slot *mm_slot; |
464 | struct mm_slot *slot; | |
b46e756f KS |
465 | int free = 0; |
466 | ||
467 | spin_lock(&khugepaged_mm_lock); | |
b26e2701 QZ |
468 | slot = mm_slot_lookup(mm_slots_hash, mm); |
469 | mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); | |
b46e756f | 470 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { |
b26e2701 QZ |
471 | hash_del(&slot->hash); |
472 | list_del(&slot->mm_node); | |
b46e756f KS |
473 | free = 1; |
474 | } | |
475 | spin_unlock(&khugepaged_mm_lock); | |
476 | ||
477 | if (free) { | |
478 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
b26e2701 | 479 | mm_slot_free(mm_slot_cache, mm_slot); |
b46e756f KS |
480 | mmdrop(mm); |
481 | } else if (mm_slot) { | |
482 | /* | |
483 | * This is required to serialize against | |
7d2c4385 ZK |
484 | * hpage_collapse_test_exit() (which is guaranteed to run |
485 | * under mmap sem read mode). Stop here (after we return all | |
486 | * pagetables will be destroyed) until khugepaged has finished | |
487 | * working on the pagetables under the mmap_lock. | |
b46e756f | 488 | */ |
d8ed45c5 ML |
489 | mmap_write_lock(mm); |
490 | mmap_write_unlock(mm); | |
b46e756f KS |
491 | } |
492 | } | |
493 | ||
92644f58 VMO |
494 | static void release_pte_folio(struct folio *folio) |
495 | { | |
496 | node_stat_mod_folio(folio, | |
497 | NR_ISOLATED_ANON + folio_is_file_lru(folio), | |
498 | -folio_nr_pages(folio)); | |
499 | folio_unlock(folio); | |
500 | folio_putback_lru(folio); | |
501 | } | |
502 | ||
b46e756f KS |
503 | static void release_pte_page(struct page *page) |
504 | { | |
92644f58 | 505 | release_pte_folio(page_folio(page)); |
b46e756f KS |
506 | } |
507 | ||
5503fbf2 KS |
508 | static void release_pte_pages(pte_t *pte, pte_t *_pte, |
509 | struct list_head *compound_pagelist) | |
b46e756f | 510 | { |
9bdfeea4 | 511 | struct folio *folio, *tmp; |
5503fbf2 | 512 | |
b46e756f | 513 | while (--_pte >= pte) { |
c33c7948 | 514 | pte_t pteval = ptep_get(_pte); |
f528260b | 515 | unsigned long pfn; |
5503fbf2 | 516 | |
f528260b VMO |
517 | if (pte_none(pteval)) |
518 | continue; | |
519 | pfn = pte_pfn(pteval); | |
520 | if (is_zero_pfn(pfn)) | |
521 | continue; | |
522 | folio = pfn_folio(pfn); | |
523 | if (folio_test_large(folio)) | |
524 | continue; | |
525 | release_pte_folio(folio); | |
5503fbf2 KS |
526 | } |
527 | ||
9bdfeea4 VMO |
528 | list_for_each_entry_safe(folio, tmp, compound_pagelist, lru) { |
529 | list_del(&folio->lru); | |
530 | release_pte_folio(folio); | |
b46e756f KS |
531 | } |
532 | } | |
533 | ||
9445689f KS |
534 | static bool is_refcount_suitable(struct page *page) |
535 | { | |
536 | int expected_refcount; | |
537 | ||
538 | expected_refcount = total_mapcount(page); | |
539 | if (PageSwapCache(page)) | |
540 | expected_refcount += compound_nr(page); | |
541 | ||
542 | return page_count(page) == expected_refcount; | |
543 | } | |
544 | ||
b46e756f KS |
545 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
546 | unsigned long address, | |
5503fbf2 | 547 | pte_t *pte, |
d8ea7cc8 | 548 | struct collapse_control *cc, |
5503fbf2 | 549 | struct list_head *compound_pagelist) |
b46e756f KS |
550 | { |
551 | struct page *page = NULL; | |
552 | pte_t *_pte; | |
50ad2f24 | 553 | int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0; |
0db501f7 | 554 | bool writable = false; |
b46e756f | 555 | |
36ee2c78 | 556 | for (_pte = pte; _pte < pte + HPAGE_PMD_NR; |
b46e756f | 557 | _pte++, address += PAGE_SIZE) { |
c33c7948 | 558 | pte_t pteval = ptep_get(_pte); |
b46e756f KS |
559 | if (pte_none(pteval) || (pte_present(pteval) && |
560 | is_zero_pfn(pte_pfn(pteval)))) { | |
d8ea7cc8 | 561 | ++none_or_zero; |
b46e756f | 562 | if (!userfaultfd_armed(vma) && |
d8ea7cc8 ZK |
563 | (!cc->is_khugepaged || |
564 | none_or_zero <= khugepaged_max_ptes_none)) { | |
b46e756f KS |
565 | continue; |
566 | } else { | |
567 | result = SCAN_EXCEED_NONE_PTE; | |
e9ea874a | 568 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
b46e756f KS |
569 | goto out; |
570 | } | |
571 | } | |
572 | if (!pte_present(pteval)) { | |
573 | result = SCAN_PTE_NON_PRESENT; | |
574 | goto out; | |
575 | } | |
dd47ac42 PX |
576 | if (pte_uffd_wp(pteval)) { |
577 | result = SCAN_PTE_UFFD_WP; | |
578 | goto out; | |
579 | } | |
b46e756f | 580 | page = vm_normal_page(vma, address, pteval); |
3218f871 | 581 | if (unlikely(!page) || unlikely(is_zone_device_page(page))) { |
b46e756f KS |
582 | result = SCAN_PAGE_NULL; |
583 | goto out; | |
584 | } | |
585 | ||
5503fbf2 KS |
586 | VM_BUG_ON_PAGE(!PageAnon(page), page); |
587 | ||
d8ea7cc8 ZK |
588 | if (page_mapcount(page) > 1) { |
589 | ++shared; | |
590 | if (cc->is_khugepaged && | |
591 | shared > khugepaged_max_ptes_shared) { | |
592 | result = SCAN_EXCEED_SHARED_PTE; | |
593 | count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); | |
594 | goto out; | |
595 | } | |
71a2c112 KS |
596 | } |
597 | ||
fece2029 | 598 | if (PageCompound(page)) { |
5503fbf2 KS |
599 | struct page *p; |
600 | page = compound_head(page); | |
fece2029 | 601 | |
5503fbf2 KS |
602 | /* |
603 | * Check if we have dealt with the compound page | |
604 | * already | |
605 | */ | |
606 | list_for_each_entry(p, compound_pagelist, lru) { | |
607 | if (page == p) | |
608 | goto next; | |
609 | } | |
610 | } | |
b46e756f KS |
611 | |
612 | /* | |
613 | * We can do it before isolate_lru_page because the | |
614 | * page can't be freed from under us. NOTE: PG_lock | |
615 | * is needed to serialize against split_huge_page | |
616 | * when invoked from the VM. | |
617 | */ | |
618 | if (!trylock_page(page)) { | |
619 | result = SCAN_PAGE_LOCK; | |
620 | goto out; | |
621 | } | |
622 | ||
623 | /* | |
9445689f KS |
624 | * Check if the page has any GUP (or other external) pins. |
625 | * | |
626 | * The page table that maps the page has been already unlinked | |
627 | * from the page table tree and this process cannot get | |
f0953a1b | 628 | * an additional pin on the page. |
9445689f KS |
629 | * |
630 | * New pins can come later if the page is shared across fork, | |
631 | * but not from this process. The other process cannot write to | |
632 | * the page, only trigger CoW. | |
b46e756f | 633 | */ |
9445689f | 634 | if (!is_refcount_suitable(page)) { |
b46e756f KS |
635 | unlock_page(page); |
636 | result = SCAN_PAGE_COUNT; | |
637 | goto out; | |
638 | } | |
b46e756f KS |
639 | |
640 | /* | |
641 | * Isolate the page to avoid collapsing an hugepage | |
642 | * currently in use by the VM. | |
643 | */ | |
f7f9c00d | 644 | if (!isolate_lru_page(page)) { |
b46e756f KS |
645 | unlock_page(page); |
646 | result = SCAN_DEL_PAGE_LRU; | |
647 | goto out; | |
648 | } | |
5503fbf2 KS |
649 | mod_node_page_state(page_pgdat(page), |
650 | NR_ISOLATED_ANON + page_is_file_lru(page), | |
651 | compound_nr(page)); | |
b46e756f KS |
652 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
653 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
654 | ||
5503fbf2 KS |
655 | if (PageCompound(page)) |
656 | list_add_tail(&page->lru, compound_pagelist); | |
657 | next: | |
d8ea7cc8 ZK |
658 | /* |
659 | * If collapse was initiated by khugepaged, check that there is | |
660 | * enough young pte to justify collapsing the page | |
661 | */ | |
662 | if (cc->is_khugepaged && | |
663 | (pte_young(pteval) || page_is_young(page) || | |
664 | PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm, | |
665 | address))) | |
0db501f7 | 666 | referenced++; |
5503fbf2 KS |
667 | |
668 | if (pte_write(pteval)) | |
669 | writable = true; | |
b46e756f | 670 | } |
74e579bf ML |
671 | |
672 | if (unlikely(!writable)) { | |
b46e756f | 673 | result = SCAN_PAGE_RO; |
d8ea7cc8 | 674 | } else if (unlikely(cc->is_khugepaged && !referenced)) { |
74e579bf ML |
675 | result = SCAN_LACK_REFERENCED_PAGE; |
676 | } else { | |
677 | result = SCAN_SUCCEED; | |
678 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, | |
679 | referenced, writable, result); | |
50ad2f24 | 680 | return result; |
b46e756f | 681 | } |
b46e756f | 682 | out: |
5503fbf2 | 683 | release_pte_pages(pte, _pte, compound_pagelist); |
b46e756f KS |
684 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, |
685 | referenced, writable, result); | |
50ad2f24 | 686 | return result; |
b46e756f KS |
687 | } |
688 | ||
98c76c9f JY |
689 | static void __collapse_huge_page_copy_succeeded(pte_t *pte, |
690 | struct vm_area_struct *vma, | |
691 | unsigned long address, | |
692 | spinlock_t *ptl, | |
693 | struct list_head *compound_pagelist) | |
b46e756f | 694 | { |
98c76c9f JY |
695 | struct page *src_page; |
696 | struct page *tmp; | |
b46e756f | 697 | pte_t *_pte; |
98c76c9f | 698 | pte_t pteval; |
b46e756f | 699 | |
98c76c9f JY |
700 | for (_pte = pte; _pte < pte + HPAGE_PMD_NR; |
701 | _pte++, address += PAGE_SIZE) { | |
c33c7948 | 702 | pteval = ptep_get(_pte); |
b46e756f | 703 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
b46e756f KS |
704 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); |
705 | if (is_zero_pfn(pte_pfn(pteval))) { | |
706 | /* | |
707 | * ptl mostly unnecessary. | |
708 | */ | |
709 | spin_lock(ptl); | |
08d5b29e | 710 | ptep_clear(vma->vm_mm, address, _pte); |
b46e756f KS |
711 | spin_unlock(ptl); |
712 | } | |
713 | } else { | |
714 | src_page = pte_page(pteval); | |
5503fbf2 KS |
715 | if (!PageCompound(src_page)) |
716 | release_pte_page(src_page); | |
b46e756f KS |
717 | /* |
718 | * ptl mostly unnecessary, but preempt has to | |
719 | * be disabled to update the per-cpu stats | |
720 | * inside page_remove_rmap(). | |
721 | */ | |
722 | spin_lock(ptl); | |
08d5b29e | 723 | ptep_clear(vma->vm_mm, address, _pte); |
cea86fe2 | 724 | page_remove_rmap(src_page, vma, false); |
b46e756f KS |
725 | spin_unlock(ptl); |
726 | free_page_and_swap_cache(src_page); | |
727 | } | |
b46e756f | 728 | } |
5503fbf2 KS |
729 | |
730 | list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { | |
731 | list_del(&src_page->lru); | |
1baec203 ML |
732 | mod_node_page_state(page_pgdat(src_page), |
733 | NR_ISOLATED_ANON + page_is_file_lru(src_page), | |
734 | -compound_nr(src_page)); | |
735 | unlock_page(src_page); | |
736 | free_swap_cache(src_page); | |
737 | putback_lru_page(src_page); | |
5503fbf2 | 738 | } |
b46e756f KS |
739 | } |
740 | ||
98c76c9f JY |
741 | static void __collapse_huge_page_copy_failed(pte_t *pte, |
742 | pmd_t *pmd, | |
743 | pmd_t orig_pmd, | |
744 | struct vm_area_struct *vma, | |
745 | struct list_head *compound_pagelist) | |
746 | { | |
747 | spinlock_t *pmd_ptl; | |
748 | ||
749 | /* | |
750 | * Re-establish the PMD to point to the original page table | |
751 | * entry. Restoring PMD needs to be done prior to releasing | |
752 | * pages. Since pages are still isolated and locked here, | |
753 | * acquiring anon_vma_lock_write is unnecessary. | |
754 | */ | |
755 | pmd_ptl = pmd_lock(vma->vm_mm, pmd); | |
756 | pmd_populate(vma->vm_mm, pmd, pmd_pgtable(orig_pmd)); | |
757 | spin_unlock(pmd_ptl); | |
758 | /* | |
759 | * Release both raw and compound pages isolated | |
760 | * in __collapse_huge_page_isolate. | |
761 | */ | |
762 | release_pte_pages(pte, pte + HPAGE_PMD_NR, compound_pagelist); | |
763 | } | |
764 | ||
765 | /* | |
766 | * __collapse_huge_page_copy - attempts to copy memory contents from raw | |
767 | * pages to a hugepage. Cleans up the raw pages if copying succeeds; | |
768 | * otherwise restores the original page table and releases isolated raw pages. | |
769 | * Returns SCAN_SUCCEED if copying succeeds, otherwise returns SCAN_COPY_MC. | |
770 | * | |
771 | * @pte: starting of the PTEs to copy from | |
772 | * @page: the new hugepage to copy contents to | |
773 | * @pmd: pointer to the new hugepage's PMD | |
774 | * @orig_pmd: the original raw pages' PMD | |
775 | * @vma: the original raw pages' virtual memory area | |
776 | * @address: starting address to copy | |
777 | * @ptl: lock on raw pages' PTEs | |
778 | * @compound_pagelist: list that stores compound pages | |
779 | */ | |
780 | static int __collapse_huge_page_copy(pte_t *pte, | |
781 | struct page *page, | |
782 | pmd_t *pmd, | |
783 | pmd_t orig_pmd, | |
784 | struct vm_area_struct *vma, | |
785 | unsigned long address, | |
786 | spinlock_t *ptl, | |
787 | struct list_head *compound_pagelist) | |
788 | { | |
789 | struct page *src_page; | |
790 | pte_t *_pte; | |
791 | pte_t pteval; | |
792 | unsigned long _address; | |
793 | int result = SCAN_SUCCEED; | |
794 | ||
795 | /* | |
796 | * Copying pages' contents is subject to memory poison at any iteration. | |
797 | */ | |
798 | for (_pte = pte, _address = address; _pte < pte + HPAGE_PMD_NR; | |
799 | _pte++, page++, _address += PAGE_SIZE) { | |
c33c7948 | 800 | pteval = ptep_get(_pte); |
98c76c9f JY |
801 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { |
802 | clear_user_highpage(page, _address); | |
803 | continue; | |
804 | } | |
805 | src_page = pte_page(pteval); | |
806 | if (copy_mc_user_highpage(page, src_page, _address, vma) > 0) { | |
807 | result = SCAN_COPY_MC; | |
808 | break; | |
809 | } | |
810 | } | |
811 | ||
812 | if (likely(result == SCAN_SUCCEED)) | |
813 | __collapse_huge_page_copy_succeeded(pte, vma, address, ptl, | |
814 | compound_pagelist); | |
815 | else | |
816 | __collapse_huge_page_copy_failed(pte, pmd, orig_pmd, vma, | |
817 | compound_pagelist); | |
818 | ||
819 | return result; | |
820 | } | |
821 | ||
b46e756f KS |
822 | static void khugepaged_alloc_sleep(void) |
823 | { | |
824 | DEFINE_WAIT(wait); | |
825 | ||
826 | add_wait_queue(&khugepaged_wait, &wait); | |
f5d39b02 PZ |
827 | __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE); |
828 | schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
b46e756f KS |
829 | remove_wait_queue(&khugepaged_wait, &wait); |
830 | } | |
831 | ||
34d6b470 | 832 | struct collapse_control khugepaged_collapse_control = { |
d8ea7cc8 | 833 | .is_khugepaged = true, |
34d6b470 | 834 | }; |
b46e756f | 835 | |
7d2c4385 | 836 | static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc) |
b46e756f KS |
837 | { |
838 | int i; | |
839 | ||
840 | /* | |
a5f5f91d | 841 | * If node_reclaim_mode is disabled, then no extra effort is made to |
b46e756f KS |
842 | * allocate memory locally. |
843 | */ | |
202e35db | 844 | if (!node_reclaim_enabled()) |
b46e756f KS |
845 | return false; |
846 | ||
847 | /* If there is a count for this node already, it must be acceptable */ | |
34d6b470 | 848 | if (cc->node_load[nid]) |
b46e756f KS |
849 | return false; |
850 | ||
851 | for (i = 0; i < MAX_NUMNODES; i++) { | |
34d6b470 | 852 | if (!cc->node_load[i]) |
b46e756f | 853 | continue; |
a55c7454 | 854 | if (node_distance(nid, i) > node_reclaim_distance) |
b46e756f KS |
855 | return true; |
856 | } | |
857 | return false; | |
858 | } | |
859 | ||
1064026b YS |
860 | #define khugepaged_defrag() \ |
861 | (transparent_hugepage_flags & \ | |
862 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)) | |
863 | ||
b46e756f KS |
864 | /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ |
865 | static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) | |
866 | { | |
25160354 | 867 | return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; |
b46e756f KS |
868 | } |
869 | ||
870 | #ifdef CONFIG_NUMA | |
7d2c4385 | 871 | static int hpage_collapse_find_target_node(struct collapse_control *cc) |
b46e756f | 872 | { |
b46e756f KS |
873 | int nid, target_node = 0, max_value = 0; |
874 | ||
875 | /* find first node with max normal pages hit */ | |
876 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
34d6b470 ZK |
877 | if (cc->node_load[nid] > max_value) { |
878 | max_value = cc->node_load[nid]; | |
b46e756f KS |
879 | target_node = nid; |
880 | } | |
881 | ||
e031ff96 YS |
882 | for_each_online_node(nid) { |
883 | if (max_value == cc->node_load[nid]) | |
884 | node_set(nid, cc->alloc_nmask); | |
885 | } | |
b46e756f | 886 | |
b46e756f KS |
887 | return target_node; |
888 | } | |
c6a7f445 | 889 | #else |
7d2c4385 | 890 | static int hpage_collapse_find_target_node(struct collapse_control *cc) |
b46e756f | 891 | { |
c6a7f445 | 892 | return 0; |
b46e756f | 893 | } |
c6a7f445 | 894 | #endif |
b46e756f | 895 | |
e031ff96 YS |
896 | static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node, |
897 | nodemask_t *nmask) | |
b46e756f | 898 | { |
e031ff96 | 899 | *hpage = __alloc_pages(gfp, HPAGE_PMD_ORDER, node, nmask); |
b46e756f KS |
900 | if (unlikely(!*hpage)) { |
901 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
9710a78a | 902 | return false; |
b46e756f KS |
903 | } |
904 | ||
905 | prep_transhuge_page(*hpage); | |
906 | count_vm_event(THP_COLLAPSE_ALLOC); | |
b46e756f KS |
907 | return true; |
908 | } | |
909 | ||
b46e756f | 910 | /* |
c1e8d7c6 ML |
911 | * If mmap_lock temporarily dropped, revalidate vma |
912 | * before taking mmap_lock. | |
50ad2f24 | 913 | * Returns enum scan_result value. |
b46e756f KS |
914 | */ |
915 | ||
c131f751 | 916 | static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, |
34488399 | 917 | bool expect_anon, |
a7f4e6e4 ZK |
918 | struct vm_area_struct **vmap, |
919 | struct collapse_control *cc) | |
b46e756f KS |
920 | { |
921 | struct vm_area_struct *vma; | |
b46e756f | 922 | |
7d2c4385 | 923 | if (unlikely(hpage_collapse_test_exit(mm))) |
b46e756f KS |
924 | return SCAN_ANY_PROCESS; |
925 | ||
c131f751 | 926 | *vmap = vma = find_vma(mm, address); |
b46e756f KS |
927 | if (!vma) |
928 | return SCAN_VMA_NULL; | |
929 | ||
4fa6893f | 930 | if (!transhuge_vma_suitable(vma, address)) |
b46e756f | 931 | return SCAN_ADDRESS_RANGE; |
a7f4e6e4 ZK |
932 | if (!hugepage_vma_check(vma, vma->vm_flags, false, false, |
933 | cc->is_khugepaged)) | |
b46e756f | 934 | return SCAN_VMA_CHECK; |
f707fa49 YS |
935 | /* |
936 | * Anon VMA expected, the address may be unmapped then | |
937 | * remapped to file after khugepaged reaquired the mmap_lock. | |
938 | * | |
939 | * hugepage_vma_check may return true for qualified file | |
940 | * vmas. | |
941 | */ | |
34488399 ZK |
942 | if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap))) |
943 | return SCAN_PAGE_ANON; | |
50ad2f24 | 944 | return SCAN_SUCCEED; |
b46e756f KS |
945 | } |
946 | ||
50722804 ZK |
947 | static int find_pmd_or_thp_or_none(struct mm_struct *mm, |
948 | unsigned long address, | |
949 | pmd_t **pmd) | |
950 | { | |
951 | pmd_t pmde; | |
952 | ||
953 | *pmd = mm_find_pmd(mm, address); | |
954 | if (!*pmd) | |
955 | return SCAN_PMD_NULL; | |
956 | ||
dab6e717 | 957 | pmde = pmdp_get_lockless(*pmd); |
34488399 ZK |
958 | if (pmd_none(pmde)) |
959 | return SCAN_PMD_NONE; | |
edb5d0cf ZK |
960 | if (!pmd_present(pmde)) |
961 | return SCAN_PMD_NULL; | |
50722804 ZK |
962 | if (pmd_trans_huge(pmde)) |
963 | return SCAN_PMD_MAPPED; | |
edb5d0cf ZK |
964 | if (pmd_devmap(pmde)) |
965 | return SCAN_PMD_NULL; | |
50722804 ZK |
966 | if (pmd_bad(pmde)) |
967 | return SCAN_PMD_NULL; | |
968 | return SCAN_SUCCEED; | |
969 | } | |
970 | ||
971 | static int check_pmd_still_valid(struct mm_struct *mm, | |
972 | unsigned long address, | |
973 | pmd_t *pmd) | |
974 | { | |
975 | pmd_t *new_pmd; | |
976 | int result = find_pmd_or_thp_or_none(mm, address, &new_pmd); | |
977 | ||
978 | if (result != SCAN_SUCCEED) | |
979 | return result; | |
980 | if (new_pmd != pmd) | |
981 | return SCAN_FAIL; | |
982 | return SCAN_SUCCEED; | |
b46e756f KS |
983 | } |
984 | ||
985 | /* | |
986 | * Bring missing pages in from swap, to complete THP collapse. | |
7d2c4385 | 987 | * Only done if hpage_collapse_scan_pmd believes it is worthwhile. |
b46e756f | 988 | * |
4d928e20 | 989 | * Called and returns without pte mapped or spinlocks held. |
895f5ee4 | 990 | * Returns result: if not SCAN_SUCCEED, mmap_lock has been released. |
b46e756f | 991 | */ |
50ad2f24 ZK |
992 | static int __collapse_huge_page_swapin(struct mm_struct *mm, |
993 | struct vm_area_struct *vma, | |
994 | unsigned long haddr, pmd_t *pmd, | |
995 | int referenced) | |
b46e756f | 996 | { |
2b740303 SJ |
997 | int swapped_in = 0; |
998 | vm_fault_t ret = 0; | |
2b635dd3 | 999 | unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); |
895f5ee4 HD |
1000 | int result; |
1001 | pte_t *pte = NULL; | |
c7ad0880 | 1002 | spinlock_t *ptl; |
2b635dd3 WD |
1003 | |
1004 | for (address = haddr; address < end; address += PAGE_SIZE) { | |
1005 | struct vm_fault vmf = { | |
1006 | .vma = vma, | |
1007 | .address = address, | |
895f5ee4 | 1008 | .pgoff = linear_page_index(vma, address), |
2b635dd3 WD |
1009 | .flags = FAULT_FLAG_ALLOW_RETRY, |
1010 | .pmd = pmd, | |
1011 | }; | |
1012 | ||
895f5ee4 | 1013 | if (!pte++) { |
c7ad0880 | 1014 | pte = pte_offset_map_nolock(mm, pmd, address, &ptl); |
895f5ee4 HD |
1015 | if (!pte) { |
1016 | mmap_read_unlock(mm); | |
1017 | result = SCAN_PMD_NULL; | |
1018 | goto out; | |
1019 | } | |
2b635dd3 | 1020 | } |
895f5ee4 | 1021 | |
c7ad0880 | 1022 | vmf.orig_pte = ptep_get_lockless(pte); |
895f5ee4 HD |
1023 | if (!is_swap_pte(vmf.orig_pte)) |
1024 | continue; | |
1025 | ||
1026 | vmf.pte = pte; | |
c7ad0880 | 1027 | vmf.ptl = ptl; |
2994302b | 1028 | ret = do_swap_page(&vmf); |
895f5ee4 HD |
1029 | /* Which unmaps pte (after perhaps re-checking the entry) */ |
1030 | pte = NULL; | |
0db501f7 | 1031 | |
4d928e20 ML |
1032 | /* |
1033 | * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. | |
1034 | * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because | |
1035 | * we do not retry here and swap entry will remain in pagetable | |
1036 | * resulting in later failure. | |
1037 | */ | |
b46e756f | 1038 | if (ret & VM_FAULT_RETRY) { |
50ad2f24 | 1039 | /* Likely, but not guaranteed, that page lock failed */ |
895f5ee4 HD |
1040 | result = SCAN_PAGE_LOCK; |
1041 | goto out; | |
b46e756f KS |
1042 | } |
1043 | if (ret & VM_FAULT_ERROR) { | |
4d928e20 | 1044 | mmap_read_unlock(mm); |
895f5ee4 HD |
1045 | result = SCAN_FAIL; |
1046 | goto out; | |
b46e756f | 1047 | } |
4d928e20 | 1048 | swapped_in++; |
b46e756f | 1049 | } |
ae2c5d80 | 1050 | |
895f5ee4 HD |
1051 | if (pte) |
1052 | pte_unmap(pte); | |
1053 | ||
1fec6890 | 1054 | /* Drain LRU cache to remove extra pin on the swapped in pages */ |
ae2c5d80 KS |
1055 | if (swapped_in) |
1056 | lru_add_drain(); | |
1057 | ||
895f5ee4 HD |
1058 | result = SCAN_SUCCEED; |
1059 | out: | |
1060 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, result); | |
1061 | return result; | |
b46e756f KS |
1062 | } |
1063 | ||
9710a78a ZK |
1064 | static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm, |
1065 | struct collapse_control *cc) | |
1066 | { | |
7d8faaf1 | 1067 | gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() : |
e031ff96 | 1068 | GFP_TRANSHUGE); |
7d2c4385 | 1069 | int node = hpage_collapse_find_target_node(cc); |
94c02ad7 | 1070 | struct folio *folio; |
9710a78a | 1071 | |
e031ff96 | 1072 | if (!hpage_collapse_alloc_page(hpage, gfp, node, &cc->alloc_nmask)) |
9710a78a | 1073 | return SCAN_ALLOC_HUGE_PAGE_FAIL; |
94c02ad7 PX |
1074 | |
1075 | folio = page_folio(*hpage); | |
1076 | if (unlikely(mem_cgroup_charge(folio, mm, gfp))) { | |
1077 | folio_put(folio); | |
1078 | *hpage = NULL; | |
9710a78a | 1079 | return SCAN_CGROUP_CHARGE_FAIL; |
94c02ad7 | 1080 | } |
9710a78a | 1081 | count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC); |
94c02ad7 | 1082 | |
9710a78a ZK |
1083 | return SCAN_SUCCEED; |
1084 | } | |
1085 | ||
50ad2f24 ZK |
1086 | static int collapse_huge_page(struct mm_struct *mm, unsigned long address, |
1087 | int referenced, int unmapped, | |
1088 | struct collapse_control *cc) | |
b46e756f | 1089 | { |
5503fbf2 | 1090 | LIST_HEAD(compound_pagelist); |
b46e756f KS |
1091 | pmd_t *pmd, _pmd; |
1092 | pte_t *pte; | |
1093 | pgtable_t pgtable; | |
50ad2f24 | 1094 | struct page *hpage; |
b46e756f | 1095 | spinlock_t *pmd_ptl, *pte_ptl; |
50ad2f24 | 1096 | int result = SCAN_FAIL; |
c131f751 | 1097 | struct vm_area_struct *vma; |
ac46d4f3 | 1098 | struct mmu_notifier_range range; |
b46e756f KS |
1099 | |
1100 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
1101 | ||
988ddb71 | 1102 | /* |
c1e8d7c6 | 1103 | * Before allocating the hugepage, release the mmap_lock read lock. |
988ddb71 | 1104 | * The allocation can take potentially a long time if it involves |
c1e8d7c6 | 1105 | * sync compaction, and we do not need to hold the mmap_lock during |
988ddb71 KS |
1106 | * that. We will recheck the vma after taking it again in write mode. |
1107 | */ | |
d8ed45c5 | 1108 | mmap_read_unlock(mm); |
b46e756f | 1109 | |
50ad2f24 | 1110 | result = alloc_charge_hpage(&hpage, mm, cc); |
9710a78a | 1111 | if (result != SCAN_SUCCEED) |
b46e756f | 1112 | goto out_nolock; |
b46e756f | 1113 | |
d8ed45c5 | 1114 | mmap_read_lock(mm); |
34488399 | 1115 | result = hugepage_vma_revalidate(mm, address, true, &vma, cc); |
50ad2f24 | 1116 | if (result != SCAN_SUCCEED) { |
d8ed45c5 | 1117 | mmap_read_unlock(mm); |
b46e756f KS |
1118 | goto out_nolock; |
1119 | } | |
1120 | ||
50722804 ZK |
1121 | result = find_pmd_or_thp_or_none(mm, address, &pmd); |
1122 | if (result != SCAN_SUCCEED) { | |
d8ed45c5 | 1123 | mmap_read_unlock(mm); |
b46e756f KS |
1124 | goto out_nolock; |
1125 | } | |
1126 | ||
50ad2f24 ZK |
1127 | if (unmapped) { |
1128 | /* | |
1129 | * __collapse_huge_page_swapin will return with mmap_lock | |
1130 | * released when it fails. So we jump out_nolock directly in | |
1131 | * that case. Continuing to collapse causes inconsistency. | |
1132 | */ | |
1133 | result = __collapse_huge_page_swapin(mm, vma, address, pmd, | |
1134 | referenced); | |
1135 | if (result != SCAN_SUCCEED) | |
1136 | goto out_nolock; | |
b46e756f KS |
1137 | } |
1138 | ||
d8ed45c5 | 1139 | mmap_read_unlock(mm); |
b46e756f KS |
1140 | /* |
1141 | * Prevent all access to pagetables with the exception of | |
1142 | * gup_fast later handled by the ptep_clear_flush and the VM | |
1143 | * handled by the anon_vma lock + PG_lock. | |
1144 | */ | |
d8ed45c5 | 1145 | mmap_write_lock(mm); |
34488399 | 1146 | result = hugepage_vma_revalidate(mm, address, true, &vma, cc); |
50ad2f24 | 1147 | if (result != SCAN_SUCCEED) |
18d24a7c | 1148 | goto out_up_write; |
b46e756f | 1149 | /* check if the pmd is still valid */ |
50722804 ZK |
1150 | result = check_pmd_still_valid(mm, address, pmd); |
1151 | if (result != SCAN_SUCCEED) | |
18d24a7c | 1152 | goto out_up_write; |
b46e756f | 1153 | |
55fd6fcc | 1154 | vma_start_write(vma); |
b46e756f KS |
1155 | anon_vma_lock_write(vma->anon_vma); |
1156 | ||
7d4a8be0 AP |
1157 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, address, |
1158 | address + HPAGE_PMD_SIZE); | |
ac46d4f3 | 1159 | mmu_notifier_invalidate_range_start(&range); |
ec649c9d | 1160 | |
b46e756f KS |
1161 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
1162 | /* | |
70cbc3cc YS |
1163 | * This removes any huge TLB entry from the CPU so we won't allow |
1164 | * huge and small TLB entries for the same virtual address to | |
1165 | * avoid the risk of CPU bugs in that area. | |
1166 | * | |
1167 | * Parallel fast GUP is fine since fast GUP will back off when | |
1168 | * it detects PMD is changed. | |
b46e756f KS |
1169 | */ |
1170 | _pmd = pmdp_collapse_flush(vma, address, pmd); | |
1171 | spin_unlock(pmd_ptl); | |
ac46d4f3 | 1172 | mmu_notifier_invalidate_range_end(&range); |
2ba99c5e | 1173 | tlb_remove_table_sync_one(); |
b46e756f | 1174 | |
895f5ee4 HD |
1175 | pte = pte_offset_map_lock(mm, &_pmd, address, &pte_ptl); |
1176 | if (pte) { | |
1177 | result = __collapse_huge_page_isolate(vma, address, pte, cc, | |
1178 | &compound_pagelist); | |
1179 | spin_unlock(pte_ptl); | |
1180 | } else { | |
1181 | result = SCAN_PMD_NULL; | |
1182 | } | |
b46e756f | 1183 | |
50ad2f24 | 1184 | if (unlikely(result != SCAN_SUCCEED)) { |
895f5ee4 HD |
1185 | if (pte) |
1186 | pte_unmap(pte); | |
b46e756f KS |
1187 | spin_lock(pmd_ptl); |
1188 | BUG_ON(!pmd_none(*pmd)); | |
1189 | /* | |
1190 | * We can only use set_pmd_at when establishing | |
1191 | * hugepmds and never for establishing regular pmds that | |
1192 | * points to regular pagetables. Use pmd_populate for that | |
1193 | */ | |
1194 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
1195 | spin_unlock(pmd_ptl); | |
1196 | anon_vma_unlock_write(vma->anon_vma); | |
18d24a7c | 1197 | goto out_up_write; |
b46e756f KS |
1198 | } |
1199 | ||
1200 | /* | |
1201 | * All pages are isolated and locked so anon_vma rmap | |
1202 | * can't run anymore. | |
1203 | */ | |
1204 | anon_vma_unlock_write(vma->anon_vma); | |
1205 | ||
98c76c9f JY |
1206 | result = __collapse_huge_page_copy(pte, hpage, pmd, _pmd, |
1207 | vma, address, pte_ptl, | |
1208 | &compound_pagelist); | |
b46e756f | 1209 | pte_unmap(pte); |
98c76c9f JY |
1210 | if (unlikely(result != SCAN_SUCCEED)) |
1211 | goto out_up_write; | |
1212 | ||
588d01f9 ML |
1213 | /* |
1214 | * spin_lock() below is not the equivalent of smp_wmb(), but | |
1215 | * the smp_wmb() inside __SetPageUptodate() can be reused to | |
1216 | * avoid the copy_huge_page writes to become visible after | |
1217 | * the set_pmd_at() write. | |
1218 | */ | |
50ad2f24 | 1219 | __SetPageUptodate(hpage); |
b46e756f KS |
1220 | pgtable = pmd_pgtable(_pmd); |
1221 | ||
50ad2f24 | 1222 | _pmd = mk_huge_pmd(hpage, vma->vm_page_prot); |
f55e1014 | 1223 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); |
b46e756f | 1224 | |
b46e756f KS |
1225 | spin_lock(pmd_ptl); |
1226 | BUG_ON(!pmd_none(*pmd)); | |
50ad2f24 ZK |
1227 | page_add_new_anon_rmap(hpage, vma, address); |
1228 | lru_cache_add_inactive_or_unevictable(hpage, vma); | |
b46e756f KS |
1229 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
1230 | set_pmd_at(mm, address, pmd, _pmd); | |
1231 | update_mmu_cache_pmd(vma, address, pmd); | |
1232 | spin_unlock(pmd_ptl); | |
1233 | ||
50ad2f24 | 1234 | hpage = NULL; |
b46e756f | 1235 | |
b46e756f KS |
1236 | result = SCAN_SUCCEED; |
1237 | out_up_write: | |
d8ed45c5 | 1238 | mmap_write_unlock(mm); |
b46e756f | 1239 | out_nolock: |
7cb1d7ef | 1240 | if (hpage) |
50ad2f24 | 1241 | put_page(hpage); |
50ad2f24 ZK |
1242 | trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result); |
1243 | return result; | |
b46e756f KS |
1244 | } |
1245 | ||
7d2c4385 ZK |
1246 | static int hpage_collapse_scan_pmd(struct mm_struct *mm, |
1247 | struct vm_area_struct *vma, | |
1248 | unsigned long address, bool *mmap_locked, | |
1249 | struct collapse_control *cc) | |
b46e756f KS |
1250 | { |
1251 | pmd_t *pmd; | |
1252 | pte_t *pte, *_pte; | |
50ad2f24 | 1253 | int result = SCAN_FAIL, referenced = 0; |
71a2c112 | 1254 | int none_or_zero = 0, shared = 0; |
b46e756f KS |
1255 | struct page *page = NULL; |
1256 | unsigned long _address; | |
1257 | spinlock_t *ptl; | |
1258 | int node = NUMA_NO_NODE, unmapped = 0; | |
0db501f7 | 1259 | bool writable = false; |
b46e756f KS |
1260 | |
1261 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
1262 | ||
50722804 ZK |
1263 | result = find_pmd_or_thp_or_none(mm, address, &pmd); |
1264 | if (result != SCAN_SUCCEED) | |
b46e756f | 1265 | goto out; |
b46e756f | 1266 | |
34d6b470 | 1267 | memset(cc->node_load, 0, sizeof(cc->node_load)); |
e031ff96 | 1268 | nodes_clear(cc->alloc_nmask); |
b46e756f | 1269 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
895f5ee4 HD |
1270 | if (!pte) { |
1271 | result = SCAN_PMD_NULL; | |
1272 | goto out; | |
1273 | } | |
1274 | ||
36ee2c78 | 1275 | for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; |
b46e756f | 1276 | _pte++, _address += PAGE_SIZE) { |
c33c7948 | 1277 | pte_t pteval = ptep_get(_pte); |
b46e756f | 1278 | if (is_swap_pte(pteval)) { |
d8ea7cc8 ZK |
1279 | ++unmapped; |
1280 | if (!cc->is_khugepaged || | |
1281 | unmapped <= khugepaged_max_ptes_swap) { | |
e1e267c7 PX |
1282 | /* |
1283 | * Always be strict with uffd-wp | |
1284 | * enabled swap entries. Please see | |
1285 | * comment below for pte_uffd_wp(). | |
1286 | */ | |
2bad466c | 1287 | if (pte_swp_uffd_wp_any(pteval)) { |
e1e267c7 PX |
1288 | result = SCAN_PTE_UFFD_WP; |
1289 | goto out_unmap; | |
1290 | } | |
b46e756f KS |
1291 | continue; |
1292 | } else { | |
1293 | result = SCAN_EXCEED_SWAP_PTE; | |
e9ea874a | 1294 | count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); |
b46e756f KS |
1295 | goto out_unmap; |
1296 | } | |
1297 | } | |
1298 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
d8ea7cc8 | 1299 | ++none_or_zero; |
b46e756f | 1300 | if (!userfaultfd_armed(vma) && |
d8ea7cc8 ZK |
1301 | (!cc->is_khugepaged || |
1302 | none_or_zero <= khugepaged_max_ptes_none)) { | |
b46e756f KS |
1303 | continue; |
1304 | } else { | |
1305 | result = SCAN_EXCEED_NONE_PTE; | |
e9ea874a | 1306 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
b46e756f KS |
1307 | goto out_unmap; |
1308 | } | |
1309 | } | |
e1e267c7 PX |
1310 | if (pte_uffd_wp(pteval)) { |
1311 | /* | |
1312 | * Don't collapse the page if any of the small | |
1313 | * PTEs are armed with uffd write protection. | |
1314 | * Here we can also mark the new huge pmd as | |
1315 | * write protected if any of the small ones is | |
8958b249 | 1316 | * marked but that could bring unknown |
e1e267c7 PX |
1317 | * userfault messages that falls outside of |
1318 | * the registered range. So, just be simple. | |
1319 | */ | |
1320 | result = SCAN_PTE_UFFD_WP; | |
1321 | goto out_unmap; | |
1322 | } | |
b46e756f KS |
1323 | if (pte_write(pteval)) |
1324 | writable = true; | |
1325 | ||
1326 | page = vm_normal_page(vma, _address, pteval); | |
3218f871 | 1327 | if (unlikely(!page) || unlikely(is_zone_device_page(page))) { |
b46e756f KS |
1328 | result = SCAN_PAGE_NULL; |
1329 | goto out_unmap; | |
1330 | } | |
1331 | ||
d8ea7cc8 ZK |
1332 | if (page_mapcount(page) > 1) { |
1333 | ++shared; | |
1334 | if (cc->is_khugepaged && | |
1335 | shared > khugepaged_max_ptes_shared) { | |
1336 | result = SCAN_EXCEED_SHARED_PTE; | |
1337 | count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); | |
1338 | goto out_unmap; | |
1339 | } | |
71a2c112 KS |
1340 | } |
1341 | ||
5503fbf2 | 1342 | page = compound_head(page); |
b46e756f KS |
1343 | |
1344 | /* | |
1345 | * Record which node the original page is from and save this | |
34d6b470 | 1346 | * information to cc->node_load[]. |
0b8f0d87 | 1347 | * Khugepaged will allocate hugepage from the node has the max |
b46e756f KS |
1348 | * hit record. |
1349 | */ | |
1350 | node = page_to_nid(page); | |
7d2c4385 | 1351 | if (hpage_collapse_scan_abort(node, cc)) { |
b46e756f KS |
1352 | result = SCAN_SCAN_ABORT; |
1353 | goto out_unmap; | |
1354 | } | |
34d6b470 | 1355 | cc->node_load[node]++; |
b46e756f KS |
1356 | if (!PageLRU(page)) { |
1357 | result = SCAN_PAGE_LRU; | |
1358 | goto out_unmap; | |
1359 | } | |
1360 | if (PageLocked(page)) { | |
1361 | result = SCAN_PAGE_LOCK; | |
1362 | goto out_unmap; | |
1363 | } | |
1364 | if (!PageAnon(page)) { | |
1365 | result = SCAN_PAGE_ANON; | |
1366 | goto out_unmap; | |
1367 | } | |
1368 | ||
1369 | /* | |
9445689f KS |
1370 | * Check if the page has any GUP (or other external) pins. |
1371 | * | |
cb67f428 HD |
1372 | * Here the check may be racy: |
1373 | * it may see total_mapcount > refcount in some cases? | |
9445689f KS |
1374 | * But such case is ephemeral we could always retry collapse |
1375 | * later. However it may report false positive if the page | |
1376 | * has excessive GUP pins (i.e. 512). Anyway the same check | |
1377 | * will be done again later the risk seems low. | |
b46e756f | 1378 | */ |
9445689f | 1379 | if (!is_refcount_suitable(page)) { |
b46e756f KS |
1380 | result = SCAN_PAGE_COUNT; |
1381 | goto out_unmap; | |
1382 | } | |
d8ea7cc8 ZK |
1383 | |
1384 | /* | |
1385 | * If collapse was initiated by khugepaged, check that there is | |
1386 | * enough young pte to justify collapsing the page | |
1387 | */ | |
1388 | if (cc->is_khugepaged && | |
1389 | (pte_young(pteval) || page_is_young(page) || | |
1390 | PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm, | |
1391 | address))) | |
0db501f7 | 1392 | referenced++; |
b46e756f | 1393 | } |
ffe945e6 | 1394 | if (!writable) { |
b46e756f | 1395 | result = SCAN_PAGE_RO; |
d8ea7cc8 ZK |
1396 | } else if (cc->is_khugepaged && |
1397 | (!referenced || | |
1398 | (unmapped && referenced < HPAGE_PMD_NR / 2))) { | |
ffe945e6 KS |
1399 | result = SCAN_LACK_REFERENCED_PAGE; |
1400 | } else { | |
1401 | result = SCAN_SUCCEED; | |
b46e756f KS |
1402 | } |
1403 | out_unmap: | |
1404 | pte_unmap_unlock(pte, ptl); | |
50ad2f24 ZK |
1405 | if (result == SCAN_SUCCEED) { |
1406 | result = collapse_huge_page(mm, address, referenced, | |
1407 | unmapped, cc); | |
c1e8d7c6 | 1408 | /* collapse_huge_page will return with the mmap_lock released */ |
50ad2f24 | 1409 | *mmap_locked = false; |
b46e756f KS |
1410 | } |
1411 | out: | |
1412 | trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, | |
1413 | none_or_zero, result, unmapped); | |
50ad2f24 | 1414 | return result; |
b46e756f KS |
1415 | } |
1416 | ||
b26e2701 | 1417 | static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot) |
b46e756f | 1418 | { |
b26e2701 QZ |
1419 | struct mm_slot *slot = &mm_slot->slot; |
1420 | struct mm_struct *mm = slot->mm; | |
b46e756f | 1421 | |
35f3aa39 | 1422 | lockdep_assert_held(&khugepaged_mm_lock); |
b46e756f | 1423 | |
7d2c4385 | 1424 | if (hpage_collapse_test_exit(mm)) { |
b46e756f | 1425 | /* free mm_slot */ |
b26e2701 QZ |
1426 | hash_del(&slot->hash); |
1427 | list_del(&slot->mm_node); | |
b46e756f KS |
1428 | |
1429 | /* | |
1430 | * Not strictly needed because the mm exited already. | |
1431 | * | |
1432 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
1433 | */ | |
1434 | ||
1435 | /* khugepaged_mm_lock actually not necessary for the below */ | |
b26e2701 | 1436 | mm_slot_free(mm_slot_cache, mm_slot); |
b46e756f KS |
1437 | mmdrop(mm); |
1438 | } | |
1439 | } | |
1440 | ||
396bcc52 | 1441 | #ifdef CONFIG_SHMEM |
27e1f827 SL |
1442 | /* |
1443 | * Notify khugepaged that given addr of the mm is pte-mapped THP. Then | |
1444 | * khugepaged should try to collapse the page table. | |
34488399 ZK |
1445 | * |
1446 | * Note that following race exists: | |
1447 | * (1) khugepaged calls khugepaged_collapse_pte_mapped_thps() for mm_struct A, | |
1448 | * emptying the A's ->pte_mapped_thp[] array. | |
1449 | * (2) MADV_COLLAPSE collapses some file extent with target mm_struct B, and | |
1450 | * retract_page_tables() finds a VMA in mm_struct A mapping the same extent | |
1451 | * (at virtual address X) and adds an entry (for X) into mm_struct A's | |
1452 | * ->pte-mapped_thp[] array. | |
1453 | * (3) khugepaged calls khugepaged_collapse_scan_file() for mm_struct A at X, | |
1454 | * sees a pte-mapped THP (SCAN_PTE_MAPPED_HUGEPAGE) and adds an entry | |
1455 | * (for X) into mm_struct A's ->pte-mapped_thp[] array. | |
1456 | * Thus, it's possible the same address is added multiple times for the same | |
1457 | * mm_struct. Should this happen, we'll simply attempt | |
1458 | * collapse_pte_mapped_thp() multiple times for the same address, under the same | |
1459 | * exclusive mmap_lock, and assuming the first call is successful, subsequent | |
1460 | * attempts will return quickly (without grabbing any additional locks) when | |
1461 | * a huge pmd is found in find_pmd_or_thp_or_none(). Since this is a cheap | |
1462 | * check, and since this is a rare occurrence, the cost of preventing this | |
1463 | * "multiple-add" is thought to be more expensive than just handling it, should | |
1464 | * it occur. | |
27e1f827 | 1465 | */ |
58ac9a89 | 1466 | static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm, |
081c3256 | 1467 | unsigned long addr) |
27e1f827 | 1468 | { |
b26e2701 QZ |
1469 | struct khugepaged_mm_slot *mm_slot; |
1470 | struct mm_slot *slot; | |
58ac9a89 | 1471 | bool ret = false; |
27e1f827 SL |
1472 | |
1473 | VM_BUG_ON(addr & ~HPAGE_PMD_MASK); | |
1474 | ||
1475 | spin_lock(&khugepaged_mm_lock); | |
b26e2701 QZ |
1476 | slot = mm_slot_lookup(mm_slots_hash, mm); |
1477 | mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); | |
58ac9a89 | 1478 | if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) { |
27e1f827 | 1479 | mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; |
58ac9a89 ZK |
1480 | ret = true; |
1481 | } | |
27e1f827 | 1482 | spin_unlock(&khugepaged_mm_lock); |
58ac9a89 | 1483 | return ret; |
27e1f827 SL |
1484 | } |
1485 | ||
34488399 ZK |
1486 | /* hpage must be locked, and mmap_lock must be held in write */ |
1487 | static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, | |
1488 | pmd_t *pmdp, struct page *hpage) | |
1489 | { | |
1490 | struct vm_fault vmf = { | |
1491 | .vma = vma, | |
1492 | .address = addr, | |
1493 | .flags = 0, | |
1494 | .pmd = pmdp, | |
1495 | }; | |
1496 | ||
1497 | VM_BUG_ON(!PageTransHuge(hpage)); | |
1498 | mmap_assert_write_locked(vma->vm_mm); | |
1499 | ||
1500 | if (do_set_pmd(&vmf, hpage)) | |
1501 | return SCAN_FAIL; | |
1502 | ||
1503 | get_page(hpage); | |
1504 | return SCAN_SUCCEED; | |
27e1f827 SL |
1505 | } |
1506 | ||
8d3c106e JH |
1507 | /* |
1508 | * A note about locking: | |
1509 | * Trying to take the page table spinlocks would be useless here because those | |
1510 | * are only used to synchronize: | |
1511 | * | |
1512 | * - modifying terminal entries (ones that point to a data page, not to another | |
1513 | * page table) | |
1514 | * - installing *new* non-terminal entries | |
1515 | * | |
1516 | * Instead, we need roughly the same kind of protection as free_pgtables() or | |
1517 | * mm_take_all_locks() (but only for a single VMA): | |
1518 | * The mmap lock together with this VMA's rmap locks covers all paths towards | |
1519 | * the page table entries we're messing with here, except for hardware page | |
1520 | * table walks and lockless_pages_from_mm(). | |
1521 | */ | |
e59a47b8 PT |
1522 | static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma, |
1523 | unsigned long addr, pmd_t *pmdp) | |
1524 | { | |
e59a47b8 | 1525 | pmd_t pmd; |
f268f6cf | 1526 | struct mmu_notifier_range range; |
e59a47b8 | 1527 | |
80110bbf | 1528 | mmap_assert_write_locked(mm); |
8d3c106e JH |
1529 | if (vma->vm_file) |
1530 | lockdep_assert_held_write(&vma->vm_file->f_mapping->i_mmap_rwsem); | |
1531 | /* | |
1532 | * All anon_vmas attached to the VMA have the same root and are | |
1533 | * therefore locked by the same lock. | |
1534 | */ | |
1535 | if (vma->anon_vma) | |
1536 | lockdep_assert_held_write(&vma->anon_vma->root->rwsem); | |
1537 | ||
7d4a8be0 | 1538 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm, addr, |
f268f6cf JH |
1539 | addr + HPAGE_PMD_SIZE); |
1540 | mmu_notifier_invalidate_range_start(&range); | |
e59a47b8 | 1541 | pmd = pmdp_collapse_flush(vma, addr, pmdp); |
2ba99c5e | 1542 | tlb_remove_table_sync_one(); |
f268f6cf | 1543 | mmu_notifier_invalidate_range_end(&range); |
e59a47b8 | 1544 | mm_dec_nr_ptes(mm); |
80110bbf | 1545 | page_table_check_pte_clear_range(mm, addr, pmd); |
e59a47b8 PT |
1546 | pte_free(mm, pmd_pgtable(pmd)); |
1547 | } | |
1548 | ||
27e1f827 | 1549 | /** |
336e6b53 AS |
1550 | * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at |
1551 | * address haddr. | |
1552 | * | |
1553 | * @mm: process address space where collapse happens | |
1554 | * @addr: THP collapse address | |
34488399 | 1555 | * @install_pmd: If a huge PMD should be installed |
27e1f827 SL |
1556 | * |
1557 | * This function checks whether all the PTEs in the PMD are pointing to the | |
1558 | * right THP. If so, retract the page table so the THP can refault in with | |
34488399 | 1559 | * as pmd-mapped. Possibly install a huge PMD mapping the THP. |
27e1f827 | 1560 | */ |
34488399 ZK |
1561 | int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, |
1562 | bool install_pmd) | |
27e1f827 SL |
1563 | { |
1564 | unsigned long haddr = addr & HPAGE_PMD_MASK; | |
94d815b2 | 1565 | struct vm_area_struct *vma = vma_lookup(mm, haddr); |
119a5fc1 | 1566 | struct page *hpage; |
27e1f827 | 1567 | pte_t *start_pte, *pte; |
e59a47b8 | 1568 | pmd_t *pmd; |
27e1f827 | 1569 | spinlock_t *ptl; |
58ac9a89 | 1570 | int count = 0, result = SCAN_FAIL; |
27e1f827 SL |
1571 | int i; |
1572 | ||
58ac9a89 ZK |
1573 | mmap_assert_write_locked(mm); |
1574 | ||
34488399 | 1575 | /* Fast check before locking page if already PMD-mapped */ |
58ac9a89 | 1576 | result = find_pmd_or_thp_or_none(mm, haddr, &pmd); |
34488399 ZK |
1577 | if (result == SCAN_PMD_MAPPED) |
1578 | return result; | |
58ac9a89 | 1579 | |
27e1f827 | 1580 | if (!vma || !vma->vm_file || |
fef792a4 | 1581 | !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) |
34488399 | 1582 | return SCAN_VMA_CHECK; |
27e1f827 SL |
1583 | |
1584 | /* | |
a7f4e6e4 ZK |
1585 | * If we are here, we've succeeded in replacing all the native pages |
1586 | * in the page cache with a single hugepage. If a mm were to fault-in | |
1587 | * this memory (mapped by a suitably aligned VMA), we'd get the hugepage | |
1588 | * and map it by a PMD, regardless of sysfs THP settings. As such, let's | |
1589 | * analogously elide sysfs THP settings here. | |
27e1f827 | 1590 | */ |
a7f4e6e4 | 1591 | if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) |
34488399 | 1592 | return SCAN_VMA_CHECK; |
27e1f827 | 1593 | |
deb4c93a PX |
1594 | /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ |
1595 | if (userfaultfd_wp(vma)) | |
34488399 | 1596 | return SCAN_PTE_UFFD_WP; |
deb4c93a | 1597 | |
119a5fc1 HD |
1598 | hpage = find_lock_page(vma->vm_file->f_mapping, |
1599 | linear_page_index(vma, haddr)); | |
1600 | if (!hpage) | |
34488399 | 1601 | return SCAN_PAGE_NULL; |
119a5fc1 | 1602 | |
34488399 ZK |
1603 | if (!PageHead(hpage)) { |
1604 | result = SCAN_FAIL; | |
119a5fc1 | 1605 | goto drop_hpage; |
34488399 | 1606 | } |
119a5fc1 | 1607 | |
34488399 ZK |
1608 | if (compound_order(hpage) != HPAGE_PMD_ORDER) { |
1609 | result = SCAN_PAGE_COMPOUND; | |
119a5fc1 | 1610 | goto drop_hpage; |
34488399 | 1611 | } |
119a5fc1 | 1612 | |
34488399 ZK |
1613 | switch (result) { |
1614 | case SCAN_SUCCEED: | |
1615 | break; | |
1616 | case SCAN_PMD_NONE: | |
1617 | /* | |
1618 | * In MADV_COLLAPSE path, possible race with khugepaged where | |
1619 | * all pte entries have been removed and pmd cleared. If so, | |
1620 | * skip all the pte checks and just update the pmd mapping. | |
1621 | */ | |
1622 | goto maybe_install_pmd; | |
1623 | default: | |
119a5fc1 | 1624 | goto drop_hpage; |
34488399 | 1625 | } |
27e1f827 | 1626 | |
55fd6fcc SB |
1627 | /* Lock the vma before taking i_mmap and page table locks */ |
1628 | vma_start_write(vma); | |
1629 | ||
8d3c106e JH |
1630 | /* |
1631 | * We need to lock the mapping so that from here on, only GUP-fast and | |
1632 | * hardware page walks can access the parts of the page tables that | |
1633 | * we're operating on. | |
1634 | * See collapse_and_free_pmd(). | |
1635 | */ | |
1636 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
1637 | ||
1638 | /* | |
1639 | * This spinlock should be unnecessary: Nobody else should be accessing | |
1640 | * the page tables under spinlock protection here, only | |
1641 | * lockless_pages_from_mm() and the hardware page walker can access page | |
1642 | * tables while all the high-level locks are held in write mode. | |
1643 | */ | |
34488399 | 1644 | result = SCAN_FAIL; |
895f5ee4 HD |
1645 | start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); |
1646 | if (!start_pte) | |
1647 | goto drop_immap; | |
27e1f827 SL |
1648 | |
1649 | /* step 1: check all mapped PTEs are to the right huge page */ | |
1650 | for (i = 0, addr = haddr, pte = start_pte; | |
1651 | i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { | |
1652 | struct page *page; | |
c33c7948 | 1653 | pte_t ptent = ptep_get(pte); |
27e1f827 SL |
1654 | |
1655 | /* empty pte, skip */ | |
c33c7948 | 1656 | if (pte_none(ptent)) |
27e1f827 SL |
1657 | continue; |
1658 | ||
1659 | /* page swapped out, abort */ | |
c33c7948 | 1660 | if (!pte_present(ptent)) { |
34488399 | 1661 | result = SCAN_PTE_NON_PRESENT; |
27e1f827 | 1662 | goto abort; |
34488399 | 1663 | } |
27e1f827 | 1664 | |
c33c7948 | 1665 | page = vm_normal_page(vma, addr, ptent); |
3218f871 AS |
1666 | if (WARN_ON_ONCE(page && is_zone_device_page(page))) |
1667 | page = NULL; | |
27e1f827 | 1668 | /* |
119a5fc1 HD |
1669 | * Note that uprobe, debugger, or MAP_PRIVATE may change the |
1670 | * page table, but the new page will not be a subpage of hpage. | |
27e1f827 | 1671 | */ |
119a5fc1 | 1672 | if (hpage + i != page) |
27e1f827 SL |
1673 | goto abort; |
1674 | count++; | |
1675 | } | |
1676 | ||
1677 | /* step 2: adjust rmap */ | |
1678 | for (i = 0, addr = haddr, pte = start_pte; | |
1679 | i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { | |
1680 | struct page *page; | |
c33c7948 | 1681 | pte_t ptent = ptep_get(pte); |
27e1f827 | 1682 | |
c33c7948 | 1683 | if (pte_none(ptent)) |
27e1f827 | 1684 | continue; |
c33c7948 | 1685 | page = vm_normal_page(vma, addr, ptent); |
3218f871 AS |
1686 | if (WARN_ON_ONCE(page && is_zone_device_page(page))) |
1687 | goto abort; | |
cea86fe2 | 1688 | page_remove_rmap(page, vma, false); |
27e1f827 SL |
1689 | } |
1690 | ||
1691 | pte_unmap_unlock(start_pte, ptl); | |
1692 | ||
1693 | /* step 3: set proper refcount and mm_counters. */ | |
119a5fc1 | 1694 | if (count) { |
27e1f827 SL |
1695 | page_ref_sub(hpage, count); |
1696 | add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); | |
1697 | } | |
1698 | ||
34488399 | 1699 | /* step 4: remove pte entries */ |
ab0c3f12 HD |
1700 | /* we make no change to anon, but protect concurrent anon page lookup */ |
1701 | if (vma->anon_vma) | |
1702 | anon_vma_lock_write(vma->anon_vma); | |
1703 | ||
e59a47b8 | 1704 | collapse_and_free_pmd(mm, vma, haddr, pmd); |
34488399 | 1705 | |
ab0c3f12 HD |
1706 | if (vma->anon_vma) |
1707 | anon_vma_unlock_write(vma->anon_vma); | |
8d3c106e JH |
1708 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
1709 | ||
34488399 ZK |
1710 | maybe_install_pmd: |
1711 | /* step 5: install pmd entry */ | |
1712 | result = install_pmd | |
1713 | ? set_huge_pmd(vma, haddr, pmd, hpage) | |
1714 | : SCAN_SUCCEED; | |
1715 | ||
119a5fc1 HD |
1716 | drop_hpage: |
1717 | unlock_page(hpage); | |
1718 | put_page(hpage); | |
34488399 | 1719 | return result; |
27e1f827 SL |
1720 | |
1721 | abort: | |
1722 | pte_unmap_unlock(start_pte, ptl); | |
895f5ee4 | 1723 | drop_immap: |
8d3c106e | 1724 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
119a5fc1 | 1725 | goto drop_hpage; |
27e1f827 SL |
1726 | } |
1727 | ||
b26e2701 | 1728 | static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot) |
27e1f827 | 1729 | { |
b26e2701 QZ |
1730 | struct mm_slot *slot = &mm_slot->slot; |
1731 | struct mm_struct *mm = slot->mm; | |
27e1f827 SL |
1732 | int i; |
1733 | ||
1734 | if (likely(mm_slot->nr_pte_mapped_thp == 0)) | |
0edf61e5 | 1735 | return; |
27e1f827 | 1736 | |
d8ed45c5 | 1737 | if (!mmap_write_trylock(mm)) |
0edf61e5 | 1738 | return; |
27e1f827 | 1739 | |
7d2c4385 | 1740 | if (unlikely(hpage_collapse_test_exit(mm))) |
27e1f827 SL |
1741 | goto out; |
1742 | ||
1743 | for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) | |
34488399 | 1744 | collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i], false); |
27e1f827 SL |
1745 | |
1746 | out: | |
1747 | mm_slot->nr_pte_mapped_thp = 0; | |
d8ed45c5 | 1748 | mmap_write_unlock(mm); |
27e1f827 SL |
1749 | } |
1750 | ||
34488399 ZK |
1751 | static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff, |
1752 | struct mm_struct *target_mm, | |
1753 | unsigned long target_addr, struct page *hpage, | |
1754 | struct collapse_control *cc) | |
f3f0e1d2 KS |
1755 | { |
1756 | struct vm_area_struct *vma; | |
34488399 | 1757 | int target_result = SCAN_FAIL; |
f3f0e1d2 KS |
1758 | |
1759 | i_mmap_lock_write(mapping); | |
1760 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { | |
34488399 ZK |
1761 | int result = SCAN_FAIL; |
1762 | struct mm_struct *mm = NULL; | |
1763 | unsigned long addr = 0; | |
1764 | pmd_t *pmd; | |
1765 | bool is_target = false; | |
1766 | ||
27e1f827 SL |
1767 | /* |
1768 | * Check vma->anon_vma to exclude MAP_PRIVATE mappings that | |
1769 | * got written to. These VMAs are likely not worth investing | |
3e4e28c5 | 1770 | * mmap_write_lock(mm) as PMD-mapping is likely to be split |
27e1f827 SL |
1771 | * later. |
1772 | * | |
36ee2c78 | 1773 | * Note that vma->anon_vma check is racy: it can be set up after |
c1e8d7c6 | 1774 | * the check but before we took mmap_lock by the fault path. |
27e1f827 SL |
1775 | * But page lock would prevent establishing any new ptes of the |
1776 | * page, so we are safe. | |
1777 | * | |
1778 | * An alternative would be drop the check, but check that page | |
1779 | * table is clear before calling pmdp_collapse_flush() under | |
1780 | * ptl. It has higher chance to recover THP for the VMA, but | |
8d3c106e JH |
1781 | * has higher cost too. It would also probably require locking |
1782 | * the anon_vma. | |
27e1f827 | 1783 | */ |
023f47a8 | 1784 | if (READ_ONCE(vma->anon_vma)) { |
34488399 ZK |
1785 | result = SCAN_PAGE_ANON; |
1786 | goto next; | |
1787 | } | |
f3f0e1d2 | 1788 | addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); |
34488399 ZK |
1789 | if (addr & ~HPAGE_PMD_MASK || |
1790 | vma->vm_end < addr + HPAGE_PMD_SIZE) { | |
1791 | result = SCAN_VMA_CHECK; | |
1792 | goto next; | |
1793 | } | |
18e77600 | 1794 | mm = vma->vm_mm; |
34488399 ZK |
1795 | is_target = mm == target_mm && addr == target_addr; |
1796 | result = find_pmd_or_thp_or_none(mm, addr, &pmd); | |
1797 | if (result != SCAN_SUCCEED) | |
1798 | goto next; | |
f3f0e1d2 | 1799 | /* |
c1e8d7c6 | 1800 | * We need exclusive mmap_lock to retract page table. |
27e1f827 SL |
1801 | * |
1802 | * We use trylock due to lock inversion: we need to acquire | |
c1e8d7c6 | 1803 | * mmap_lock while holding page lock. Fault path does it in |
27e1f827 | 1804 | * reverse order. Trylock is a way to avoid deadlock. |
34488399 ZK |
1805 | * |
1806 | * Also, it's not MADV_COLLAPSE's job to collapse other | |
1807 | * mappings - let khugepaged take care of them later. | |
f3f0e1d2 | 1808 | */ |
34488399 ZK |
1809 | result = SCAN_PTE_MAPPED_HUGEPAGE; |
1810 | if ((cc->is_khugepaged || is_target) && | |
1811 | mmap_write_trylock(mm)) { | |
55fd6fcc SB |
1812 | /* trylock for the same lock inversion as above */ |
1813 | if (!vma_try_start_write(vma)) | |
1814 | goto unlock_next; | |
1815 | ||
023f47a8 JH |
1816 | /* |
1817 | * Re-check whether we have an ->anon_vma, because | |
1818 | * collapse_and_free_pmd() requires that either no | |
1819 | * ->anon_vma exists or the anon_vma is locked. | |
1820 | * We already checked ->anon_vma above, but that check | |
1821 | * is racy because ->anon_vma can be populated under the | |
1822 | * mmap lock in read mode. | |
1823 | */ | |
1824 | if (vma->anon_vma) { | |
1825 | result = SCAN_PAGE_ANON; | |
1826 | goto unlock_next; | |
1827 | } | |
deb4c93a PX |
1828 | /* |
1829 | * When a vma is registered with uffd-wp, we can't | |
1830 | * recycle the pmd pgtable because there can be pte | |
1831 | * markers installed. Skip it only, so the rest mm/vma | |
1832 | * can still have the same file mapped hugely, however | |
1833 | * it'll always mapped in small page size for uffd-wp | |
1834 | * registered ranges. | |
1835 | */ | |
34488399 ZK |
1836 | if (hpage_collapse_test_exit(mm)) { |
1837 | result = SCAN_ANY_PROCESS; | |
1838 | goto unlock_next; | |
1839 | } | |
1840 | if (userfaultfd_wp(vma)) { | |
1841 | result = SCAN_PTE_UFFD_WP; | |
1842 | goto unlock_next; | |
1843 | } | |
1844 | collapse_and_free_pmd(mm, vma, addr, pmd); | |
1845 | if (!cc->is_khugepaged && is_target) | |
1846 | result = set_huge_pmd(vma, addr, pmd, hpage); | |
1847 | else | |
1848 | result = SCAN_SUCCEED; | |
1849 | ||
1850 | unlock_next: | |
18e77600 | 1851 | mmap_write_unlock(mm); |
34488399 ZK |
1852 | goto next; |
1853 | } | |
1854 | /* | |
1855 | * Calling context will handle target mm/addr. Otherwise, let | |
1856 | * khugepaged try again later. | |
1857 | */ | |
1858 | if (!is_target) { | |
18e77600 | 1859 | khugepaged_add_pte_mapped_thp(mm, addr); |
34488399 | 1860 | continue; |
f3f0e1d2 | 1861 | } |
34488399 ZK |
1862 | next: |
1863 | if (is_target) | |
1864 | target_result = result; | |
f3f0e1d2 KS |
1865 | } |
1866 | i_mmap_unlock_write(mapping); | |
34488399 | 1867 | return target_result; |
f3f0e1d2 KS |
1868 | } |
1869 | ||
1870 | /** | |
99cb0dbd | 1871 | * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. |
f3f0e1d2 | 1872 | * |
336e6b53 | 1873 | * @mm: process address space where collapse happens |
34488399 | 1874 | * @addr: virtual collapse start address |
336e6b53 AS |
1875 | * @file: file that collapse on |
1876 | * @start: collapse start address | |
9710a78a | 1877 | * @cc: collapse context and scratchpad |
336e6b53 | 1878 | * |
f3f0e1d2 | 1879 | * Basic scheme is simple, details are more complex: |
87c460a0 | 1880 | * - allocate and lock a new huge page; |
a2e17cc2 | 1881 | * - scan page cache, locking old pages |
99cb0dbd | 1882 | * + swap/gup in pages if necessary; |
a2e17cc2 DS |
1883 | * - copy data to new page |
1884 | * - handle shmem holes | |
1885 | * + re-validate that holes weren't filled by someone else | |
1886 | * + check for userfaultfd | |
ac492b9c | 1887 | * - finalize updates to the page cache; |
77da9389 | 1888 | * - if replacing succeeds: |
87c460a0 | 1889 | * + unlock huge page; |
a2e17cc2 | 1890 | * + free old pages; |
f3f0e1d2 | 1891 | * - if replacing failed; |
a2e17cc2 | 1892 | * + unlock old pages |
87c460a0 | 1893 | * + unlock and free huge page; |
f3f0e1d2 | 1894 | */ |
34488399 ZK |
1895 | static int collapse_file(struct mm_struct *mm, unsigned long addr, |
1896 | struct file *file, pgoff_t start, | |
1897 | struct collapse_control *cc) | |
f3f0e1d2 | 1898 | { |
579c571e | 1899 | struct address_space *mapping = file->f_mapping; |
50ad2f24 | 1900 | struct page *hpage; |
12904d95 JY |
1901 | struct page *page; |
1902 | struct page *tmp; | |
1903 | struct folio *folio; | |
4c9473e8 | 1904 | pgoff_t index = 0, end = start + HPAGE_PMD_NR; |
f3f0e1d2 | 1905 | LIST_HEAD(pagelist); |
77da9389 | 1906 | XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); |
f3f0e1d2 | 1907 | int nr_none = 0, result = SCAN_SUCCEED; |
99cb0dbd | 1908 | bool is_shmem = shmem_file(file); |
4c9473e8 | 1909 | int nr = 0; |
f3f0e1d2 | 1910 | |
99cb0dbd | 1911 | VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); |
f3f0e1d2 KS |
1912 | VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); |
1913 | ||
50ad2f24 | 1914 | result = alloc_charge_hpage(&hpage, mm, cc); |
9710a78a | 1915 | if (result != SCAN_SUCCEED) |
f3f0e1d2 | 1916 | goto out; |
f3f0e1d2 | 1917 | |
cae106dd DS |
1918 | __SetPageLocked(hpage); |
1919 | if (is_shmem) | |
1920 | __SetPageSwapBacked(hpage); | |
1921 | hpage->index = start; | |
1922 | hpage->mapping = mapping; | |
1923 | ||
6b24ca4a MWO |
1924 | /* |
1925 | * Ensure we have slots for all the pages in the range. This is | |
1926 | * almost certainly a no-op because most of the pages must be present | |
1927 | */ | |
95feeabb HD |
1928 | do { |
1929 | xas_lock_irq(&xas); | |
1930 | xas_create_range(&xas); | |
1931 | if (!xas_error(&xas)) | |
1932 | break; | |
1933 | xas_unlock_irq(&xas); | |
1934 | if (!xas_nomem(&xas, GFP_KERNEL)) { | |
95feeabb | 1935 | result = SCAN_FAIL; |
cae106dd | 1936 | goto rollback; |
95feeabb HD |
1937 | } |
1938 | } while (1); | |
1939 | ||
77da9389 | 1940 | for (index = start; index < end; index++) { |
e8c716bc HD |
1941 | xas_set(&xas, index); |
1942 | page = xas_load(&xas); | |
77da9389 MW |
1943 | |
1944 | VM_BUG_ON(index != xas.xa_index); | |
99cb0dbd SL |
1945 | if (is_shmem) { |
1946 | if (!page) { | |
1947 | /* | |
1948 | * Stop if extent has been truncated or | |
1949 | * hole-punched, and is now completely | |
1950 | * empty. | |
1951 | */ | |
1952 | if (index == start) { | |
1953 | if (!xas_next_entry(&xas, end - 1)) { | |
1954 | result = SCAN_TRUNCATED; | |
1955 | goto xa_locked; | |
1956 | } | |
99cb0dbd SL |
1957 | } |
1958 | if (!shmem_charge(mapping->host, 1)) { | |
1959 | result = SCAN_FAIL; | |
042a3082 | 1960 | goto xa_locked; |
701270fa | 1961 | } |
99cb0dbd SL |
1962 | nr_none++; |
1963 | continue; | |
701270fa | 1964 | } |
99cb0dbd SL |
1965 | |
1966 | if (xa_is_value(page) || !PageUptodate(page)) { | |
1967 | xas_unlock_irq(&xas); | |
1968 | /* swap in or instantiate fallocated page */ | |
7459c149 MWO |
1969 | if (shmem_get_folio(mapping->host, index, |
1970 | &folio, SGP_NOALLOC)) { | |
99cb0dbd SL |
1971 | result = SCAN_FAIL; |
1972 | goto xa_unlocked; | |
1973 | } | |
1fec6890 | 1974 | /* drain lru cache to help isolate_lru_page() */ |
efa3d814 | 1975 | lru_add_drain(); |
7459c149 | 1976 | page = folio_file_page(folio, index); |
99cb0dbd SL |
1977 | } else if (trylock_page(page)) { |
1978 | get_page(page); | |
1979 | xas_unlock_irq(&xas); | |
1980 | } else { | |
1981 | result = SCAN_PAGE_LOCK; | |
042a3082 | 1982 | goto xa_locked; |
77da9389 | 1983 | } |
99cb0dbd SL |
1984 | } else { /* !is_shmem */ |
1985 | if (!page || xa_is_value(page)) { | |
1986 | xas_unlock_irq(&xas); | |
1987 | page_cache_sync_readahead(mapping, &file->f_ra, | |
1988 | file, index, | |
e5a59d30 | 1989 | end - index); |
1fec6890 | 1990 | /* drain lru cache to help isolate_lru_page() */ |
99cb0dbd SL |
1991 | lru_add_drain(); |
1992 | page = find_lock_page(mapping, index); | |
1993 | if (unlikely(page == NULL)) { | |
1994 | result = SCAN_FAIL; | |
1995 | goto xa_unlocked; | |
1996 | } | |
75f36069 SL |
1997 | } else if (PageDirty(page)) { |
1998 | /* | |
1999 | * khugepaged only works on read-only fd, | |
2000 | * so this page is dirty because it hasn't | |
2001 | * been flushed since first write. There | |
2002 | * won't be new dirty pages. | |
2003 | * | |
2004 | * Trigger async flush here and hope the | |
2005 | * writeback is done when khugepaged | |
2006 | * revisits this page. | |
2007 | * | |
2008 | * This is a one-off situation. We are not | |
2009 | * forcing writeback in loop. | |
2010 | */ | |
2011 | xas_unlock_irq(&xas); | |
2012 | filemap_flush(mapping); | |
2013 | result = SCAN_FAIL; | |
2014 | goto xa_unlocked; | |
74c42e1b RW |
2015 | } else if (PageWriteback(page)) { |
2016 | xas_unlock_irq(&xas); | |
2017 | result = SCAN_FAIL; | |
2018 | goto xa_unlocked; | |
99cb0dbd SL |
2019 | } else if (trylock_page(page)) { |
2020 | get_page(page); | |
2021 | xas_unlock_irq(&xas); | |
2022 | } else { | |
2023 | result = SCAN_PAGE_LOCK; | |
2024 | goto xa_locked; | |
f3f0e1d2 | 2025 | } |
f3f0e1d2 KS |
2026 | } |
2027 | ||
2028 | /* | |
b93b0163 | 2029 | * The page must be locked, so we can drop the i_pages lock |
f3f0e1d2 KS |
2030 | * without racing with truncate. |
2031 | */ | |
2032 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
4655e5e5 SL |
2033 | |
2034 | /* make sure the page is up to date */ | |
2035 | if (unlikely(!PageUptodate(page))) { | |
2036 | result = SCAN_FAIL; | |
2037 | goto out_unlock; | |
2038 | } | |
06a5e126 HD |
2039 | |
2040 | /* | |
2041 | * If file was truncated then extended, or hole-punched, before | |
2042 | * we locked the first page, then a THP might be there already. | |
58ac9a89 | 2043 | * This will be discovered on the first iteration. |
06a5e126 HD |
2044 | */ |
2045 | if (PageTransCompound(page)) { | |
58ac9a89 ZK |
2046 | struct page *head = compound_head(page); |
2047 | ||
2048 | result = compound_order(head) == HPAGE_PMD_ORDER && | |
2049 | head->index == start | |
2050 | /* Maybe PMD-mapped */ | |
2051 | ? SCAN_PTE_MAPPED_HUGEPAGE | |
2052 | : SCAN_PAGE_COMPOUND; | |
06a5e126 HD |
2053 | goto out_unlock; |
2054 | } | |
f3f0e1d2 | 2055 | |
64ab3195 VMO |
2056 | folio = page_folio(page); |
2057 | ||
2058 | if (folio_mapping(folio) != mapping) { | |
f3f0e1d2 KS |
2059 | result = SCAN_TRUNCATED; |
2060 | goto out_unlock; | |
2061 | } | |
f3f0e1d2 | 2062 | |
64ab3195 VMO |
2063 | if (!is_shmem && (folio_test_dirty(folio) || |
2064 | folio_test_writeback(folio))) { | |
4655e5e5 SL |
2065 | /* |
2066 | * khugepaged only works on read-only fd, so this | |
2067 | * page is dirty because it hasn't been flushed | |
2068 | * since first write. | |
2069 | */ | |
2070 | result = SCAN_FAIL; | |
2071 | goto out_unlock; | |
2072 | } | |
2073 | ||
be2d5756 | 2074 | if (!folio_isolate_lru(folio)) { |
f3f0e1d2 | 2075 | result = SCAN_DEL_PAGE_LRU; |
042a3082 | 2076 | goto out_unlock; |
f3f0e1d2 KS |
2077 | } |
2078 | ||
64ab3195 VMO |
2079 | if (folio_has_private(folio) && |
2080 | !filemap_release_folio(folio, GFP_KERNEL)) { | |
99cb0dbd | 2081 | result = SCAN_PAGE_HAS_PRIVATE; |
64ab3195 | 2082 | folio_putback_lru(folio); |
99cb0dbd SL |
2083 | goto out_unlock; |
2084 | } | |
2085 | ||
64ab3195 VMO |
2086 | if (folio_mapped(folio)) |
2087 | try_to_unmap(folio, | |
869f7ee6 | 2088 | TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); |
f3f0e1d2 | 2089 | |
77da9389 | 2090 | xas_lock_irq(&xas); |
f3f0e1d2 | 2091 | |
e8c716bc | 2092 | VM_BUG_ON_PAGE(page != xa_load(xas.xa, index), page); |
f3f0e1d2 KS |
2093 | |
2094 | /* | |
a2e17cc2 | 2095 | * We control three references to the page: |
f3f0e1d2 | 2096 | * - we hold a pin on it; |
77da9389 | 2097 | * - one reference from page cache; |
f3f0e1d2 | 2098 | * - one from isolate_lru_page; |
a2e17cc2 DS |
2099 | * If those are the only references, then any new usage of the |
2100 | * page will have to fetch it from the page cache. That requires | |
2101 | * locking the page to handle truncate, so any new usage will be | |
2102 | * blocked until we unlock page after collapse/during rollback. | |
f3f0e1d2 | 2103 | */ |
a2e17cc2 | 2104 | if (page_count(page) != 3) { |
f3f0e1d2 | 2105 | result = SCAN_PAGE_COUNT; |
042a3082 HD |
2106 | xas_unlock_irq(&xas); |
2107 | putback_lru_page(page); | |
2108 | goto out_unlock; | |
f3f0e1d2 KS |
2109 | } |
2110 | ||
2111 | /* | |
a2e17cc2 | 2112 | * Accumulate the pages that are being collapsed. |
f3f0e1d2 KS |
2113 | */ |
2114 | list_add_tail(&page->lru, &pagelist); | |
f3f0e1d2 | 2115 | continue; |
f3f0e1d2 KS |
2116 | out_unlock: |
2117 | unlock_page(page); | |
2118 | put_page(page); | |
042a3082 | 2119 | goto xa_unlocked; |
f3f0e1d2 KS |
2120 | } |
2121 | ||
12904d95 | 2122 | if (!is_shmem) { |
09d91cda | 2123 | filemap_nr_thps_inc(mapping); |
eb6ecbed CF |
2124 | /* |
2125 | * Paired with smp_mb() in do_dentry_open() to ensure | |
2126 | * i_writecount is up to date and the update to nr_thps is | |
2127 | * visible. Ensures the page cache will be truncated if the | |
2128 | * file is opened writable. | |
2129 | */ | |
2130 | smp_mb(); | |
2131 | if (inode_is_open_for_write(mapping->host)) { | |
2132 | result = SCAN_FAIL; | |
eb6ecbed | 2133 | filemap_nr_thps_dec(mapping); |
eb6ecbed | 2134 | } |
09d91cda | 2135 | } |
99cb0dbd | 2136 | |
042a3082 HD |
2137 | xa_locked: |
2138 | xas_unlock_irq(&xas); | |
77da9389 | 2139 | xa_unlocked: |
042a3082 | 2140 | |
6d9df8a5 HD |
2141 | /* |
2142 | * If collapse is successful, flush must be done now before copying. | |
2143 | * If collapse is unsuccessful, does flush actually need to be done? | |
2144 | * Do it anyway, to clear the state. | |
2145 | */ | |
2146 | try_to_unmap_flush(); | |
2147 | ||
cae106dd DS |
2148 | if (result != SCAN_SUCCEED) |
2149 | goto rollback; | |
2150 | ||
2151 | /* | |
a2e17cc2 | 2152 | * The old pages are locked, so they won't change anymore. |
cae106dd DS |
2153 | */ |
2154 | index = start; | |
2155 | list_for_each_entry(page, &pagelist, lru) { | |
2156 | while (index < page->index) { | |
12904d95 JY |
2157 | clear_highpage(hpage + (index % HPAGE_PMD_NR)); |
2158 | index++; | |
2159 | } | |
cae106dd DS |
2160 | if (copy_mc_highpage(hpage + (page->index % HPAGE_PMD_NR), page) > 0) { |
2161 | result = SCAN_COPY_MC; | |
2162 | goto rollback; | |
2163 | } | |
2164 | index++; | |
2165 | } | |
2166 | while (index < end) { | |
2167 | clear_highpage(hpage + (index % HPAGE_PMD_NR)); | |
2168 | index++; | |
2169 | } | |
2170 | ||
ac492b9c DS |
2171 | if (nr_none) { |
2172 | struct vm_area_struct *vma; | |
2173 | int nr_none_check = 0; | |
2174 | ||
2175 | i_mmap_lock_read(mapping); | |
2176 | xas_lock_irq(&xas); | |
2177 | ||
2178 | xas_set(&xas, start); | |
2179 | for (index = start; index < end; index++) { | |
2180 | if (!xas_next(&xas)) { | |
2181 | xas_store(&xas, XA_RETRY_ENTRY); | |
2182 | if (xas_error(&xas)) { | |
2183 | result = SCAN_STORE_FAILED; | |
2184 | goto immap_locked; | |
2185 | } | |
2186 | nr_none_check++; | |
2187 | } | |
2188 | } | |
2189 | ||
2190 | if (nr_none != nr_none_check) { | |
2191 | result = SCAN_PAGE_FILLED; | |
2192 | goto immap_locked; | |
2193 | } | |
2194 | ||
2195 | /* | |
2196 | * If userspace observed a missing page in a VMA with a MODE_MISSING | |
2197 | * userfaultfd, then it might expect a UFFD_EVENT_PAGEFAULT for that | |
2198 | * page. If so, we need to roll back to avoid suppressing such an | |
2199 | * event. Since wp/minor userfaultfds don't give userspace any | |
2200 | * guarantees that the kernel doesn't fill a missing page with a zero | |
2201 | * page, so they don't matter here. | |
2202 | * | |
2203 | * Any userfaultfds registered after this point will not be able to | |
2204 | * observe any missing pages due to the previously inserted retry | |
2205 | * entries. | |
2206 | */ | |
2207 | vma_interval_tree_foreach(vma, &mapping->i_mmap, start, end) { | |
2208 | if (userfaultfd_missing(vma)) { | |
2209 | result = SCAN_EXCEED_NONE_PTE; | |
2210 | goto immap_locked; | |
2211 | } | |
2212 | } | |
2213 | ||
2214 | immap_locked: | |
2215 | i_mmap_unlock_read(mapping); | |
2216 | if (result != SCAN_SUCCEED) { | |
2217 | xas_set(&xas, start); | |
2218 | for (index = start; index < end; index++) { | |
2219 | if (xas_next(&xas) == XA_RETRY_ENTRY) | |
2220 | xas_store(&xas, NULL); | |
2221 | } | |
2222 | ||
2223 | xas_unlock_irq(&xas); | |
2224 | goto rollback; | |
2225 | } | |
2226 | } else { | |
2227 | xas_lock_irq(&xas); | |
12904d95 JY |
2228 | } |
2229 | ||
2230 | nr = thp_nr_pages(hpage); | |
cae106dd DS |
2231 | if (is_shmem) |
2232 | __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr); | |
2233 | else | |
2234 | __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr); | |
12904d95 | 2235 | |
cae106dd DS |
2236 | if (nr_none) { |
2237 | __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none); | |
2238 | /* nr_none is always 0 for non-shmem. */ | |
2239 | __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none); | |
2240 | } | |
f3f0e1d2 | 2241 | |
a2e17cc2 DS |
2242 | /* |
2243 | * Mark hpage as uptodate before inserting it into the page cache so | |
2244 | * that it isn't mistaken for an fallocated but unwritten page. | |
2245 | */ | |
cae106dd DS |
2246 | folio = page_folio(hpage); |
2247 | folio_mark_uptodate(folio); | |
2248 | folio_ref_add(folio, HPAGE_PMD_NR - 1); | |
284a344e | 2249 | |
cae106dd DS |
2250 | if (is_shmem) |
2251 | folio_mark_dirty(folio); | |
2252 | folio_add_lru(folio); | |
f3f0e1d2 | 2253 | |
a2e17cc2 DS |
2254 | /* Join all the small entries into a single multi-index entry. */ |
2255 | xas_set_order(&xas, start, HPAGE_PMD_ORDER); | |
2256 | xas_store(&xas, hpage); | |
0175ab61 | 2257 | WARN_ON_ONCE(xas_error(&xas)); |
a2e17cc2 DS |
2258 | xas_unlock_irq(&xas); |
2259 | ||
cae106dd DS |
2260 | /* |
2261 | * Remove pte page tables, so we can re-fault the page as huge. | |
2262 | */ | |
2263 | result = retract_page_tables(mapping, start, mm, addr, hpage, | |
2264 | cc); | |
2265 | unlock_page(hpage); | |
ac492b9c DS |
2266 | |
2267 | /* | |
2268 | * The collapse has succeeded, so free the old pages. | |
2269 | */ | |
2270 | list_for_each_entry_safe(page, tmp, &pagelist, lru) { | |
2271 | list_del(&page->lru); | |
2272 | page->mapping = NULL; | |
ac492b9c DS |
2273 | ClearPageActive(page); |
2274 | ClearPageUnevictable(page); | |
2275 | unlock_page(page); | |
a2e17cc2 | 2276 | folio_put_refs(page_folio(page), 3); |
ac492b9c DS |
2277 | } |
2278 | ||
cae106dd DS |
2279 | goto out; |
2280 | ||
2281 | rollback: | |
2282 | /* Something went wrong: roll back page cache changes */ | |
cae106dd | 2283 | if (nr_none) { |
a2e17cc2 | 2284 | xas_lock_irq(&xas); |
cae106dd DS |
2285 | mapping->nrpages -= nr_none; |
2286 | shmem_uncharge(mapping->host, nr_none); | |
a2e17cc2 | 2287 | xas_unlock_irq(&xas); |
cae106dd | 2288 | } |
aaa52e34 | 2289 | |
a2e17cc2 | 2290 | list_for_each_entry_safe(page, tmp, &pagelist, lru) { |
cae106dd | 2291 | list_del(&page->lru); |
cae106dd DS |
2292 | unlock_page(page); |
2293 | putback_lru_page(page); | |
a2e17cc2 | 2294 | put_page(page); |
cae106dd | 2295 | } |
cae106dd DS |
2296 | /* |
2297 | * Undo the updates of filemap_nr_thps_inc for non-SHMEM | |
2298 | * file only. This undo is not needed unless failure is | |
2299 | * due to SCAN_COPY_MC. | |
2300 | */ | |
2301 | if (!is_shmem && result == SCAN_COPY_MC) { | |
2302 | filemap_nr_thps_dec(mapping); | |
12904d95 | 2303 | /* |
cae106dd DS |
2304 | * Paired with smp_mb() in do_dentry_open() to |
2305 | * ensure the update to nr_thps is visible. | |
12904d95 | 2306 | */ |
cae106dd DS |
2307 | smp_mb(); |
2308 | } | |
12904d95 | 2309 | |
cae106dd | 2310 | hpage->mapping = NULL; |
042a3082 | 2311 | |
cae106dd DS |
2312 | unlock_page(hpage); |
2313 | put_page(hpage); | |
f3f0e1d2 KS |
2314 | out: |
2315 | VM_BUG_ON(!list_empty(&pagelist)); | |
4c9473e8 | 2316 | trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result); |
50ad2f24 | 2317 | return result; |
f3f0e1d2 KS |
2318 | } |
2319 | ||
34488399 ZK |
2320 | static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, |
2321 | struct file *file, pgoff_t start, | |
2322 | struct collapse_control *cc) | |
f3f0e1d2 KS |
2323 | { |
2324 | struct page *page = NULL; | |
579c571e | 2325 | struct address_space *mapping = file->f_mapping; |
85b392db | 2326 | XA_STATE(xas, &mapping->i_pages, start); |
f3f0e1d2 KS |
2327 | int present, swap; |
2328 | int node = NUMA_NO_NODE; | |
2329 | int result = SCAN_SUCCEED; | |
2330 | ||
2331 | present = 0; | |
2332 | swap = 0; | |
34d6b470 | 2333 | memset(cc->node_load, 0, sizeof(cc->node_load)); |
e031ff96 | 2334 | nodes_clear(cc->alloc_nmask); |
f3f0e1d2 | 2335 | rcu_read_lock(); |
85b392db MW |
2336 | xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { |
2337 | if (xas_retry(&xas, page)) | |
f3f0e1d2 | 2338 | continue; |
f3f0e1d2 | 2339 | |
85b392db | 2340 | if (xa_is_value(page)) { |
d8ea7cc8 ZK |
2341 | ++swap; |
2342 | if (cc->is_khugepaged && | |
2343 | swap > khugepaged_max_ptes_swap) { | |
f3f0e1d2 | 2344 | result = SCAN_EXCEED_SWAP_PTE; |
e9ea874a | 2345 | count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); |
f3f0e1d2 KS |
2346 | break; |
2347 | } | |
2348 | continue; | |
2349 | } | |
2350 | ||
6b24ca4a | 2351 | /* |
58ac9a89 | 2352 | * TODO: khugepaged should compact smaller compound pages |
6b24ca4a MWO |
2353 | * into a PMD sized page |
2354 | */ | |
f3f0e1d2 | 2355 | if (PageTransCompound(page)) { |
58ac9a89 ZK |
2356 | struct page *head = compound_head(page); |
2357 | ||
2358 | result = compound_order(head) == HPAGE_PMD_ORDER && | |
2359 | head->index == start | |
2360 | /* Maybe PMD-mapped */ | |
2361 | ? SCAN_PTE_MAPPED_HUGEPAGE | |
2362 | : SCAN_PAGE_COMPOUND; | |
2363 | /* | |
2364 | * For SCAN_PTE_MAPPED_HUGEPAGE, further processing | |
2365 | * by the caller won't touch the page cache, and so | |
2366 | * it's safe to skip LRU and refcount checks before | |
2367 | * returning. | |
2368 | */ | |
f3f0e1d2 KS |
2369 | break; |
2370 | } | |
2371 | ||
2372 | node = page_to_nid(page); | |
7d2c4385 | 2373 | if (hpage_collapse_scan_abort(node, cc)) { |
f3f0e1d2 KS |
2374 | result = SCAN_SCAN_ABORT; |
2375 | break; | |
2376 | } | |
34d6b470 | 2377 | cc->node_load[node]++; |
f3f0e1d2 KS |
2378 | |
2379 | if (!PageLRU(page)) { | |
2380 | result = SCAN_PAGE_LRU; | |
2381 | break; | |
2382 | } | |
2383 | ||
99cb0dbd SL |
2384 | if (page_count(page) != |
2385 | 1 + page_mapcount(page) + page_has_private(page)) { | |
f3f0e1d2 KS |
2386 | result = SCAN_PAGE_COUNT; |
2387 | break; | |
2388 | } | |
2389 | ||
2390 | /* | |
2391 | * We probably should check if the page is referenced here, but | |
2392 | * nobody would transfer pte_young() to PageReferenced() for us. | |
2393 | * And rmap walk here is just too costly... | |
2394 | */ | |
2395 | ||
2396 | present++; | |
2397 | ||
2398 | if (need_resched()) { | |
85b392db | 2399 | xas_pause(&xas); |
f3f0e1d2 | 2400 | cond_resched_rcu(); |
f3f0e1d2 KS |
2401 | } |
2402 | } | |
2403 | rcu_read_unlock(); | |
2404 | ||
2405 | if (result == SCAN_SUCCEED) { | |
d8ea7cc8 ZK |
2406 | if (cc->is_khugepaged && |
2407 | present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { | |
f3f0e1d2 | 2408 | result = SCAN_EXCEED_NONE_PTE; |
e9ea874a | 2409 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
f3f0e1d2 | 2410 | } else { |
34488399 | 2411 | result = collapse_file(mm, addr, file, start, cc); |
f3f0e1d2 KS |
2412 | } |
2413 | } | |
2414 | ||
045634ff | 2415 | trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result); |
50ad2f24 | 2416 | return result; |
f3f0e1d2 KS |
2417 | } |
2418 | #else | |
34488399 ZK |
2419 | static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, |
2420 | struct file *file, pgoff_t start, | |
2421 | struct collapse_control *cc) | |
f3f0e1d2 KS |
2422 | { |
2423 | BUILD_BUG(); | |
2424 | } | |
27e1f827 | 2425 | |
b26e2701 | 2426 | static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot) |
27e1f827 | 2427 | { |
27e1f827 | 2428 | } |
58ac9a89 ZK |
2429 | |
2430 | static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm, | |
2431 | unsigned long addr) | |
2432 | { | |
2433 | return false; | |
2434 | } | |
f3f0e1d2 KS |
2435 | #endif |
2436 | ||
50ad2f24 | 2437 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result, |
34d6b470 | 2438 | struct collapse_control *cc) |
b46e756f KS |
2439 | __releases(&khugepaged_mm_lock) |
2440 | __acquires(&khugepaged_mm_lock) | |
2441 | { | |
68540502 | 2442 | struct vma_iterator vmi; |
b26e2701 QZ |
2443 | struct khugepaged_mm_slot *mm_slot; |
2444 | struct mm_slot *slot; | |
b46e756f KS |
2445 | struct mm_struct *mm; |
2446 | struct vm_area_struct *vma; | |
2447 | int progress = 0; | |
2448 | ||
2449 | VM_BUG_ON(!pages); | |
35f3aa39 | 2450 | lockdep_assert_held(&khugepaged_mm_lock); |
50ad2f24 | 2451 | *result = SCAN_FAIL; |
b46e756f | 2452 | |
b26e2701 | 2453 | if (khugepaged_scan.mm_slot) { |
b46e756f | 2454 | mm_slot = khugepaged_scan.mm_slot; |
b26e2701 QZ |
2455 | slot = &mm_slot->slot; |
2456 | } else { | |
2457 | slot = list_entry(khugepaged_scan.mm_head.next, | |
b46e756f | 2458 | struct mm_slot, mm_node); |
b26e2701 | 2459 | mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); |
b46e756f KS |
2460 | khugepaged_scan.address = 0; |
2461 | khugepaged_scan.mm_slot = mm_slot; | |
2462 | } | |
2463 | spin_unlock(&khugepaged_mm_lock); | |
27e1f827 | 2464 | khugepaged_collapse_pte_mapped_thps(mm_slot); |
b46e756f | 2465 | |
b26e2701 | 2466 | mm = slot->mm; |
3b454ad3 YS |
2467 | /* |
2468 | * Don't wait for semaphore (to avoid long wait times). Just move to | |
2469 | * the next mm on the list. | |
2470 | */ | |
2471 | vma = NULL; | |
d8ed45c5 | 2472 | if (unlikely(!mmap_read_trylock(mm))) |
c1e8d7c6 | 2473 | goto breakouterloop_mmap_lock; |
b46e756f KS |
2474 | |
2475 | progress++; | |
68540502 MWO |
2476 | if (unlikely(hpage_collapse_test_exit(mm))) |
2477 | goto breakouterloop; | |
2478 | ||
2479 | vma_iter_init(&vmi, mm, khugepaged_scan.address); | |
2480 | for_each_vma(vmi, vma) { | |
b46e756f KS |
2481 | unsigned long hstart, hend; |
2482 | ||
2483 | cond_resched(); | |
7d2c4385 | 2484 | if (unlikely(hpage_collapse_test_exit(mm))) { |
b46e756f KS |
2485 | progress++; |
2486 | break; | |
2487 | } | |
a7f4e6e4 | 2488 | if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) { |
b46e756f KS |
2489 | skip: |
2490 | progress++; | |
2491 | continue; | |
2492 | } | |
4fa6893f YS |
2493 | hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE); |
2494 | hend = round_down(vma->vm_end, HPAGE_PMD_SIZE); | |
b46e756f KS |
2495 | if (khugepaged_scan.address > hend) |
2496 | goto skip; | |
2497 | if (khugepaged_scan.address < hstart) | |
2498 | khugepaged_scan.address = hstart; | |
2499 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); | |
2500 | ||
2501 | while (khugepaged_scan.address < hend) { | |
50ad2f24 ZK |
2502 | bool mmap_locked = true; |
2503 | ||
b46e756f | 2504 | cond_resched(); |
7d2c4385 | 2505 | if (unlikely(hpage_collapse_test_exit(mm))) |
b46e756f KS |
2506 | goto breakouterloop; |
2507 | ||
2508 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2509 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2510 | hend); | |
99cb0dbd | 2511 | if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { |
396bcc52 | 2512 | struct file *file = get_file(vma->vm_file); |
f3f0e1d2 KS |
2513 | pgoff_t pgoff = linear_page_index(vma, |
2514 | khugepaged_scan.address); | |
99cb0dbd | 2515 | |
d8ed45c5 | 2516 | mmap_read_unlock(mm); |
34488399 ZK |
2517 | *result = hpage_collapse_scan_file(mm, |
2518 | khugepaged_scan.address, | |
2519 | file, pgoff, cc); | |
50ad2f24 | 2520 | mmap_locked = false; |
f3f0e1d2 KS |
2521 | fput(file); |
2522 | } else { | |
7d2c4385 ZK |
2523 | *result = hpage_collapse_scan_pmd(mm, vma, |
2524 | khugepaged_scan.address, | |
2525 | &mmap_locked, | |
2526 | cc); | |
f3f0e1d2 | 2527 | } |
58ac9a89 ZK |
2528 | switch (*result) { |
2529 | case SCAN_PTE_MAPPED_HUGEPAGE: { | |
2530 | pmd_t *pmd; | |
2531 | ||
2532 | *result = find_pmd_or_thp_or_none(mm, | |
2533 | khugepaged_scan.address, | |
2534 | &pmd); | |
2535 | if (*result != SCAN_SUCCEED) | |
2536 | break; | |
2537 | if (!khugepaged_add_pte_mapped_thp(mm, | |
2538 | khugepaged_scan.address)) | |
2539 | break; | |
2540 | } fallthrough; | |
2541 | case SCAN_SUCCEED: | |
50ad2f24 | 2542 | ++khugepaged_pages_collapsed; |
58ac9a89 ZK |
2543 | break; |
2544 | default: | |
2545 | break; | |
f3f0e1d2 | 2546 | } |
58ac9a89 | 2547 | |
b46e756f KS |
2548 | /* move to next address */ |
2549 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2550 | progress += HPAGE_PMD_NR; | |
50ad2f24 ZK |
2551 | if (!mmap_locked) |
2552 | /* | |
2553 | * We released mmap_lock so break loop. Note | |
2554 | * that we drop mmap_lock before all hugepage | |
2555 | * allocations, so if allocation fails, we are | |
2556 | * guaranteed to break here and report the | |
2557 | * correct result back to caller. | |
2558 | */ | |
c1e8d7c6 | 2559 | goto breakouterloop_mmap_lock; |
b46e756f KS |
2560 | if (progress >= pages) |
2561 | goto breakouterloop; | |
2562 | } | |
2563 | } | |
2564 | breakouterloop: | |
d8ed45c5 | 2565 | mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ |
c1e8d7c6 | 2566 | breakouterloop_mmap_lock: |
b46e756f KS |
2567 | |
2568 | spin_lock(&khugepaged_mm_lock); | |
2569 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); | |
2570 | /* | |
2571 | * Release the current mm_slot if this mm is about to die, or | |
2572 | * if we scanned all vmas of this mm. | |
2573 | */ | |
7d2c4385 | 2574 | if (hpage_collapse_test_exit(mm) || !vma) { |
b46e756f KS |
2575 | /* |
2576 | * Make sure that if mm_users is reaching zero while | |
2577 | * khugepaged runs here, khugepaged_exit will find | |
2578 | * mm_slot not pointing to the exiting mm. | |
2579 | */ | |
b26e2701 QZ |
2580 | if (slot->mm_node.next != &khugepaged_scan.mm_head) { |
2581 | slot = list_entry(slot->mm_node.next, | |
2582 | struct mm_slot, mm_node); | |
2583 | khugepaged_scan.mm_slot = | |
2584 | mm_slot_entry(slot, struct khugepaged_mm_slot, slot); | |
b46e756f KS |
2585 | khugepaged_scan.address = 0; |
2586 | } else { | |
2587 | khugepaged_scan.mm_slot = NULL; | |
2588 | khugepaged_full_scans++; | |
2589 | } | |
2590 | ||
2591 | collect_mm_slot(mm_slot); | |
2592 | } | |
2593 | ||
2594 | return progress; | |
2595 | } | |
2596 | ||
2597 | static int khugepaged_has_work(void) | |
2598 | { | |
2599 | return !list_empty(&khugepaged_scan.mm_head) && | |
1064026b | 2600 | hugepage_flags_enabled(); |
b46e756f KS |
2601 | } |
2602 | ||
2603 | static int khugepaged_wait_event(void) | |
2604 | { | |
2605 | return !list_empty(&khugepaged_scan.mm_head) || | |
2606 | kthread_should_stop(); | |
2607 | } | |
2608 | ||
34d6b470 | 2609 | static void khugepaged_do_scan(struct collapse_control *cc) |
b46e756f | 2610 | { |
b46e756f | 2611 | unsigned int progress = 0, pass_through_head = 0; |
89dc6a96 | 2612 | unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); |
b46e756f | 2613 | bool wait = true; |
50ad2f24 | 2614 | int result = SCAN_SUCCEED; |
b46e756f | 2615 | |
a980df33 KS |
2616 | lru_add_drain_all(); |
2617 | ||
c6a7f445 | 2618 | while (true) { |
b46e756f KS |
2619 | cond_resched(); |
2620 | ||
2621 | if (unlikely(kthread_should_stop() || try_to_freeze())) | |
2622 | break; | |
2623 | ||
2624 | spin_lock(&khugepaged_mm_lock); | |
2625 | if (!khugepaged_scan.mm_slot) | |
2626 | pass_through_head++; | |
2627 | if (khugepaged_has_work() && | |
2628 | pass_through_head < 2) | |
2629 | progress += khugepaged_scan_mm_slot(pages - progress, | |
50ad2f24 | 2630 | &result, cc); |
b46e756f KS |
2631 | else |
2632 | progress = pages; | |
2633 | spin_unlock(&khugepaged_mm_lock); | |
b46e756f | 2634 | |
c6a7f445 YS |
2635 | if (progress >= pages) |
2636 | break; | |
2637 | ||
50ad2f24 | 2638 | if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) { |
c6a7f445 YS |
2639 | /* |
2640 | * If fail to allocate the first time, try to sleep for | |
2641 | * a while. When hit again, cancel the scan. | |
2642 | */ | |
2643 | if (!wait) | |
2644 | break; | |
2645 | wait = false; | |
c6a7f445 YS |
2646 | khugepaged_alloc_sleep(); |
2647 | } | |
2648 | } | |
b46e756f KS |
2649 | } |
2650 | ||
2651 | static bool khugepaged_should_wakeup(void) | |
2652 | { | |
2653 | return kthread_should_stop() || | |
2654 | time_after_eq(jiffies, khugepaged_sleep_expire); | |
2655 | } | |
2656 | ||
2657 | static void khugepaged_wait_work(void) | |
2658 | { | |
2659 | if (khugepaged_has_work()) { | |
2660 | const unsigned long scan_sleep_jiffies = | |
2661 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs); | |
2662 | ||
2663 | if (!scan_sleep_jiffies) | |
2664 | return; | |
2665 | ||
2666 | khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; | |
2667 | wait_event_freezable_timeout(khugepaged_wait, | |
2668 | khugepaged_should_wakeup(), | |
2669 | scan_sleep_jiffies); | |
2670 | return; | |
2671 | } | |
2672 | ||
1064026b | 2673 | if (hugepage_flags_enabled()) |
b46e756f KS |
2674 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); |
2675 | } | |
2676 | ||
2677 | static int khugepaged(void *none) | |
2678 | { | |
b26e2701 | 2679 | struct khugepaged_mm_slot *mm_slot; |
b46e756f KS |
2680 | |
2681 | set_freezable(); | |
2682 | set_user_nice(current, MAX_NICE); | |
2683 | ||
2684 | while (!kthread_should_stop()) { | |
34d6b470 | 2685 | khugepaged_do_scan(&khugepaged_collapse_control); |
b46e756f KS |
2686 | khugepaged_wait_work(); |
2687 | } | |
2688 | ||
2689 | spin_lock(&khugepaged_mm_lock); | |
2690 | mm_slot = khugepaged_scan.mm_slot; | |
2691 | khugepaged_scan.mm_slot = NULL; | |
2692 | if (mm_slot) | |
2693 | collect_mm_slot(mm_slot); | |
2694 | spin_unlock(&khugepaged_mm_lock); | |
2695 | return 0; | |
2696 | } | |
2697 | ||
2698 | static void set_recommended_min_free_kbytes(void) | |
2699 | { | |
2700 | struct zone *zone; | |
2701 | int nr_zones = 0; | |
2702 | unsigned long recommended_min; | |
2703 | ||
1064026b | 2704 | if (!hugepage_flags_enabled()) { |
bd3400ea LF |
2705 | calculate_min_free_kbytes(); |
2706 | goto update_wmarks; | |
2707 | } | |
2708 | ||
b7d349c7 JK |
2709 | for_each_populated_zone(zone) { |
2710 | /* | |
2711 | * We don't need to worry about fragmentation of | |
2712 | * ZONE_MOVABLE since it only has movable pages. | |
2713 | */ | |
2714 | if (zone_idx(zone) > gfp_zone(GFP_USER)) | |
2715 | continue; | |
2716 | ||
b46e756f | 2717 | nr_zones++; |
b7d349c7 | 2718 | } |
b46e756f KS |
2719 | |
2720 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ | |
2721 | recommended_min = pageblock_nr_pages * nr_zones * 2; | |
2722 | ||
2723 | /* | |
2724 | * Make sure that on average at least two pageblocks are almost free | |
2725 | * of another type, one for a migratetype to fall back to and a | |
2726 | * second to avoid subsequent fallbacks of other types There are 3 | |
2727 | * MIGRATE_TYPES we care about. | |
2728 | */ | |
2729 | recommended_min += pageblock_nr_pages * nr_zones * | |
2730 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
2731 | ||
2732 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
2733 | recommended_min = min(recommended_min, | |
2734 | (unsigned long) nr_free_buffer_pages() / 20); | |
2735 | recommended_min <<= (PAGE_SHIFT-10); | |
2736 | ||
2737 | if (recommended_min > min_free_kbytes) { | |
2738 | if (user_min_free_kbytes >= 0) | |
2739 | pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", | |
2740 | min_free_kbytes, recommended_min); | |
2741 | ||
2742 | min_free_kbytes = recommended_min; | |
2743 | } | |
bd3400ea LF |
2744 | |
2745 | update_wmarks: | |
b46e756f KS |
2746 | setup_per_zone_wmarks(); |
2747 | } | |
2748 | ||
2749 | int start_stop_khugepaged(void) | |
2750 | { | |
b46e756f KS |
2751 | int err = 0; |
2752 | ||
2753 | mutex_lock(&khugepaged_mutex); | |
1064026b | 2754 | if (hugepage_flags_enabled()) { |
b46e756f KS |
2755 | if (!khugepaged_thread) |
2756 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
2757 | "khugepaged"); | |
2758 | if (IS_ERR(khugepaged_thread)) { | |
2759 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); | |
2760 | err = PTR_ERR(khugepaged_thread); | |
2761 | khugepaged_thread = NULL; | |
2762 | goto fail; | |
2763 | } | |
2764 | ||
2765 | if (!list_empty(&khugepaged_scan.mm_head)) | |
2766 | wake_up_interruptible(&khugepaged_wait); | |
b46e756f KS |
2767 | } else if (khugepaged_thread) { |
2768 | kthread_stop(khugepaged_thread); | |
2769 | khugepaged_thread = NULL; | |
2770 | } | |
bd3400ea | 2771 | set_recommended_min_free_kbytes(); |
b46e756f KS |
2772 | fail: |
2773 | mutex_unlock(&khugepaged_mutex); | |
2774 | return err; | |
2775 | } | |
4aab2be0 VB |
2776 | |
2777 | void khugepaged_min_free_kbytes_update(void) | |
2778 | { | |
2779 | mutex_lock(&khugepaged_mutex); | |
1064026b | 2780 | if (hugepage_flags_enabled() && khugepaged_thread) |
4aab2be0 VB |
2781 | set_recommended_min_free_kbytes(); |
2782 | mutex_unlock(&khugepaged_mutex); | |
2783 | } | |
7d8faaf1 | 2784 | |
57e9cc50 JW |
2785 | bool current_is_khugepaged(void) |
2786 | { | |
2787 | return kthread_func(current) == khugepaged; | |
2788 | } | |
2789 | ||
7d8faaf1 ZK |
2790 | static int madvise_collapse_errno(enum scan_result r) |
2791 | { | |
2792 | /* | |
2793 | * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide | |
2794 | * actionable feedback to caller, so they may take an appropriate | |
2795 | * fallback measure depending on the nature of the failure. | |
2796 | */ | |
2797 | switch (r) { | |
2798 | case SCAN_ALLOC_HUGE_PAGE_FAIL: | |
2799 | return -ENOMEM; | |
2800 | case SCAN_CGROUP_CHARGE_FAIL: | |
ac492b9c | 2801 | case SCAN_EXCEED_NONE_PTE: |
7d8faaf1 ZK |
2802 | return -EBUSY; |
2803 | /* Resource temporary unavailable - trying again might succeed */ | |
ae63c898 | 2804 | case SCAN_PAGE_COUNT: |
7d8faaf1 ZK |
2805 | case SCAN_PAGE_LOCK: |
2806 | case SCAN_PAGE_LRU: | |
0f3e2a2c | 2807 | case SCAN_DEL_PAGE_LRU: |
ac492b9c | 2808 | case SCAN_PAGE_FILLED: |
7d8faaf1 ZK |
2809 | return -EAGAIN; |
2810 | /* | |
2811 | * Other: Trying again likely not to succeed / error intrinsic to | |
2812 | * specified memory range. khugepaged likely won't be able to collapse | |
2813 | * either. | |
2814 | */ | |
2815 | default: | |
2816 | return -EINVAL; | |
2817 | } | |
2818 | } | |
2819 | ||
2820 | int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev, | |
2821 | unsigned long start, unsigned long end) | |
2822 | { | |
2823 | struct collapse_control *cc; | |
2824 | struct mm_struct *mm = vma->vm_mm; | |
2825 | unsigned long hstart, hend, addr; | |
2826 | int thps = 0, last_fail = SCAN_FAIL; | |
2827 | bool mmap_locked = true; | |
2828 | ||
2829 | BUG_ON(vma->vm_start > start); | |
2830 | BUG_ON(vma->vm_end < end); | |
2831 | ||
2832 | *prev = vma; | |
2833 | ||
7d8faaf1 ZK |
2834 | if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) |
2835 | return -EINVAL; | |
2836 | ||
2837 | cc = kmalloc(sizeof(*cc), GFP_KERNEL); | |
2838 | if (!cc) | |
2839 | return -ENOMEM; | |
2840 | cc->is_khugepaged = false; | |
7d8faaf1 ZK |
2841 | |
2842 | mmgrab(mm); | |
2843 | lru_add_drain_all(); | |
2844 | ||
2845 | hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
2846 | hend = end & HPAGE_PMD_MASK; | |
2847 | ||
2848 | for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) { | |
2849 | int result = SCAN_FAIL; | |
2850 | ||
2851 | if (!mmap_locked) { | |
2852 | cond_resched(); | |
2853 | mmap_read_lock(mm); | |
2854 | mmap_locked = true; | |
34488399 ZK |
2855 | result = hugepage_vma_revalidate(mm, addr, false, &vma, |
2856 | cc); | |
7d8faaf1 ZK |
2857 | if (result != SCAN_SUCCEED) { |
2858 | last_fail = result; | |
2859 | goto out_nolock; | |
2860 | } | |
4d24de94 | 2861 | |
52dc0310 | 2862 | hend = min(hend, vma->vm_end & HPAGE_PMD_MASK); |
7d8faaf1 ZK |
2863 | } |
2864 | mmap_assert_locked(mm); | |
2865 | memset(cc->node_load, 0, sizeof(cc->node_load)); | |
e031ff96 | 2866 | nodes_clear(cc->alloc_nmask); |
34488399 ZK |
2867 | if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { |
2868 | struct file *file = get_file(vma->vm_file); | |
2869 | pgoff_t pgoff = linear_page_index(vma, addr); | |
2870 | ||
2871 | mmap_read_unlock(mm); | |
2872 | mmap_locked = false; | |
2873 | result = hpage_collapse_scan_file(mm, addr, file, pgoff, | |
2874 | cc); | |
2875 | fput(file); | |
2876 | } else { | |
2877 | result = hpage_collapse_scan_pmd(mm, vma, addr, | |
2878 | &mmap_locked, cc); | |
2879 | } | |
7d8faaf1 ZK |
2880 | if (!mmap_locked) |
2881 | *prev = NULL; /* Tell caller we dropped mmap_lock */ | |
2882 | ||
34488399 | 2883 | handle_result: |
7d8faaf1 ZK |
2884 | switch (result) { |
2885 | case SCAN_SUCCEED: | |
2886 | case SCAN_PMD_MAPPED: | |
2887 | ++thps; | |
2888 | break; | |
34488399 ZK |
2889 | case SCAN_PTE_MAPPED_HUGEPAGE: |
2890 | BUG_ON(mmap_locked); | |
2891 | BUG_ON(*prev); | |
2892 | mmap_write_lock(mm); | |
2893 | result = collapse_pte_mapped_thp(mm, addr, true); | |
2894 | mmap_write_unlock(mm); | |
2895 | goto handle_result; | |
7d8faaf1 ZK |
2896 | /* Whitelisted set of results where continuing OK */ |
2897 | case SCAN_PMD_NULL: | |
2898 | case SCAN_PTE_NON_PRESENT: | |
2899 | case SCAN_PTE_UFFD_WP: | |
2900 | case SCAN_PAGE_RO: | |
2901 | case SCAN_LACK_REFERENCED_PAGE: | |
2902 | case SCAN_PAGE_NULL: | |
2903 | case SCAN_PAGE_COUNT: | |
2904 | case SCAN_PAGE_LOCK: | |
2905 | case SCAN_PAGE_COMPOUND: | |
2906 | case SCAN_PAGE_LRU: | |
0f3e2a2c | 2907 | case SCAN_DEL_PAGE_LRU: |
7d8faaf1 ZK |
2908 | last_fail = result; |
2909 | break; | |
2910 | default: | |
2911 | last_fail = result; | |
2912 | /* Other error, exit */ | |
2913 | goto out_maybelock; | |
2914 | } | |
2915 | } | |
2916 | ||
2917 | out_maybelock: | |
2918 | /* Caller expects us to hold mmap_lock on return */ | |
2919 | if (!mmap_locked) | |
2920 | mmap_read_lock(mm); | |
2921 | out_nolock: | |
2922 | mmap_assert_locked(mm); | |
2923 | mmdrop(mm); | |
2924 | kfree(cc); | |
2925 | ||
2926 | return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0 | |
2927 | : madvise_collapse_errno(last_fail); | |
2928 | } |