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, |
b46e756f KS |
58 | }; |
59 | ||
60 | #define CREATE_TRACE_POINTS | |
61 | #include <trace/events/huge_memory.h> | |
62 | ||
4aab2be0 VB |
63 | static struct task_struct *khugepaged_thread __read_mostly; |
64 | static DEFINE_MUTEX(khugepaged_mutex); | |
65 | ||
b46e756f KS |
66 | /* default scan 8*512 pte (or vmas) every 30 second */ |
67 | static unsigned int khugepaged_pages_to_scan __read_mostly; | |
68 | static unsigned int khugepaged_pages_collapsed; | |
69 | static unsigned int khugepaged_full_scans; | |
70 | static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000; | |
71 | /* during fragmentation poll the hugepage allocator once every minute */ | |
72 | static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000; | |
73 | static unsigned long khugepaged_sleep_expire; | |
74 | static DEFINE_SPINLOCK(khugepaged_mm_lock); | |
75 | static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait); | |
76 | /* | |
77 | * default collapse hugepages if there is at least one pte mapped like | |
78 | * it would have happened if the vma was large enough during page | |
79 | * fault. | |
d8ea7cc8 ZK |
80 | * |
81 | * Note that these are only respected if collapse was initiated by khugepaged. | |
b46e756f KS |
82 | */ |
83 | static unsigned int khugepaged_max_ptes_none __read_mostly; | |
84 | static unsigned int khugepaged_max_ptes_swap __read_mostly; | |
71a2c112 | 85 | static unsigned int khugepaged_max_ptes_shared __read_mostly; |
b46e756f KS |
86 | |
87 | #define MM_SLOTS_HASH_BITS 10 | |
88 | static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS); | |
89 | ||
90 | static struct kmem_cache *mm_slot_cache __read_mostly; | |
91 | ||
27e1f827 SL |
92 | #define MAX_PTE_MAPPED_THP 8 |
93 | ||
34d6b470 | 94 | struct collapse_control { |
d8ea7cc8 ZK |
95 | bool is_khugepaged; |
96 | ||
34d6b470 ZK |
97 | /* Num pages scanned per node */ |
98 | u32 node_load[MAX_NUMNODES]; | |
99 | ||
e031ff96 YS |
100 | /* nodemask for allocation fallback */ |
101 | nodemask_t alloc_nmask; | |
34d6b470 ZK |
102 | }; |
103 | ||
b46e756f | 104 | /** |
b26e2701 QZ |
105 | * struct khugepaged_mm_slot - khugepaged information per mm that is being scanned |
106 | * @slot: hash lookup from mm to mm_slot | |
336e6b53 AS |
107 | * @nr_pte_mapped_thp: number of pte mapped THP |
108 | * @pte_mapped_thp: address array corresponding pte mapped THP | |
b46e756f | 109 | */ |
b26e2701 QZ |
110 | struct khugepaged_mm_slot { |
111 | struct mm_slot slot; | |
27e1f827 SL |
112 | |
113 | /* pte-mapped THP in this mm */ | |
114 | int nr_pte_mapped_thp; | |
115 | unsigned long pte_mapped_thp[MAX_PTE_MAPPED_THP]; | |
b46e756f KS |
116 | }; |
117 | ||
118 | /** | |
119 | * struct khugepaged_scan - cursor for scanning | |
120 | * @mm_head: the head of the mm list to scan | |
121 | * @mm_slot: the current mm_slot we are scanning | |
122 | * @address: the next address inside that to be scanned | |
123 | * | |
124 | * There is only the one khugepaged_scan instance of this cursor structure. | |
125 | */ | |
126 | struct khugepaged_scan { | |
127 | struct list_head mm_head; | |
b26e2701 | 128 | struct khugepaged_mm_slot *mm_slot; |
b46e756f KS |
129 | unsigned long address; |
130 | }; | |
131 | ||
132 | static struct khugepaged_scan khugepaged_scan = { | |
133 | .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head), | |
134 | }; | |
135 | ||
e1465d12 | 136 | #ifdef CONFIG_SYSFS |
b46e756f KS |
137 | static ssize_t scan_sleep_millisecs_show(struct kobject *kobj, |
138 | struct kobj_attribute *attr, | |
139 | char *buf) | |
140 | { | |
ae7a927d | 141 | return sysfs_emit(buf, "%u\n", khugepaged_scan_sleep_millisecs); |
b46e756f KS |
142 | } |
143 | ||
144 | static ssize_t scan_sleep_millisecs_store(struct kobject *kobj, | |
145 | struct kobj_attribute *attr, | |
146 | const char *buf, size_t count) | |
147 | { | |
dfefd226 | 148 | unsigned int msecs; |
b46e756f KS |
149 | int err; |
150 | ||
dfefd226 AD |
151 | err = kstrtouint(buf, 10, &msecs); |
152 | if (err) | |
b46e756f KS |
153 | return -EINVAL; |
154 | ||
155 | khugepaged_scan_sleep_millisecs = msecs; | |
156 | khugepaged_sleep_expire = 0; | |
157 | wake_up_interruptible(&khugepaged_wait); | |
158 | ||
159 | return count; | |
160 | } | |
161 | static struct kobj_attribute scan_sleep_millisecs_attr = | |
6dcdc94d | 162 | __ATTR_RW(scan_sleep_millisecs); |
b46e756f KS |
163 | |
164 | static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj, | |
165 | struct kobj_attribute *attr, | |
166 | char *buf) | |
167 | { | |
ae7a927d | 168 | return sysfs_emit(buf, "%u\n", khugepaged_alloc_sleep_millisecs); |
b46e756f KS |
169 | } |
170 | ||
171 | static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj, | |
172 | struct kobj_attribute *attr, | |
173 | const char *buf, size_t count) | |
174 | { | |
dfefd226 | 175 | unsigned int msecs; |
b46e756f KS |
176 | int err; |
177 | ||
dfefd226 AD |
178 | err = kstrtouint(buf, 10, &msecs); |
179 | if (err) | |
b46e756f KS |
180 | return -EINVAL; |
181 | ||
182 | khugepaged_alloc_sleep_millisecs = msecs; | |
183 | khugepaged_sleep_expire = 0; | |
184 | wake_up_interruptible(&khugepaged_wait); | |
185 | ||
186 | return count; | |
187 | } | |
188 | static struct kobj_attribute alloc_sleep_millisecs_attr = | |
6dcdc94d | 189 | __ATTR_RW(alloc_sleep_millisecs); |
b46e756f KS |
190 | |
191 | static ssize_t pages_to_scan_show(struct kobject *kobj, | |
192 | struct kobj_attribute *attr, | |
193 | char *buf) | |
194 | { | |
ae7a927d | 195 | return sysfs_emit(buf, "%u\n", khugepaged_pages_to_scan); |
b46e756f KS |
196 | } |
197 | static ssize_t pages_to_scan_store(struct kobject *kobj, | |
198 | struct kobj_attribute *attr, | |
199 | const char *buf, size_t count) | |
200 | { | |
dfefd226 | 201 | unsigned int pages; |
b46e756f | 202 | int err; |
b46e756f | 203 | |
dfefd226 AD |
204 | err = kstrtouint(buf, 10, &pages); |
205 | if (err || !pages) | |
b46e756f KS |
206 | return -EINVAL; |
207 | ||
208 | khugepaged_pages_to_scan = pages; | |
209 | ||
210 | return count; | |
211 | } | |
212 | static struct kobj_attribute pages_to_scan_attr = | |
6dcdc94d | 213 | __ATTR_RW(pages_to_scan); |
b46e756f KS |
214 | |
215 | static ssize_t pages_collapsed_show(struct kobject *kobj, | |
216 | struct kobj_attribute *attr, | |
217 | char *buf) | |
218 | { | |
ae7a927d | 219 | return sysfs_emit(buf, "%u\n", khugepaged_pages_collapsed); |
b46e756f KS |
220 | } |
221 | static struct kobj_attribute pages_collapsed_attr = | |
222 | __ATTR_RO(pages_collapsed); | |
223 | ||
224 | static ssize_t full_scans_show(struct kobject *kobj, | |
225 | struct kobj_attribute *attr, | |
226 | char *buf) | |
227 | { | |
ae7a927d | 228 | return sysfs_emit(buf, "%u\n", khugepaged_full_scans); |
b46e756f KS |
229 | } |
230 | static struct kobj_attribute full_scans_attr = | |
231 | __ATTR_RO(full_scans); | |
232 | ||
6dcdc94d ML |
233 | static ssize_t defrag_show(struct kobject *kobj, |
234 | struct kobj_attribute *attr, char *buf) | |
b46e756f KS |
235 | { |
236 | return single_hugepage_flag_show(kobj, attr, buf, | |
ae7a927d | 237 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); |
b46e756f | 238 | } |
6dcdc94d ML |
239 | static ssize_t defrag_store(struct kobject *kobj, |
240 | struct kobj_attribute *attr, | |
241 | const char *buf, size_t count) | |
b46e756f KS |
242 | { |
243 | return single_hugepage_flag_store(kobj, attr, buf, count, | |
244 | TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG); | |
245 | } | |
246 | static struct kobj_attribute khugepaged_defrag_attr = | |
6dcdc94d | 247 | __ATTR_RW(defrag); |
b46e756f KS |
248 | |
249 | /* | |
250 | * max_ptes_none controls if khugepaged should collapse hugepages over | |
251 | * any unmapped ptes in turn potentially increasing the memory | |
252 | * footprint of the vmas. When max_ptes_none is 0 khugepaged will not | |
253 | * reduce the available free memory in the system as it | |
254 | * runs. Increasing max_ptes_none will instead potentially reduce the | |
255 | * free memory in the system during the khugepaged scan. | |
256 | */ | |
6dcdc94d ML |
257 | static ssize_t max_ptes_none_show(struct kobject *kobj, |
258 | struct kobj_attribute *attr, | |
259 | char *buf) | |
b46e756f | 260 | { |
ae7a927d | 261 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_none); |
b46e756f | 262 | } |
6dcdc94d ML |
263 | static ssize_t max_ptes_none_store(struct kobject *kobj, |
264 | struct kobj_attribute *attr, | |
265 | const char *buf, size_t count) | |
b46e756f KS |
266 | { |
267 | int err; | |
268 | unsigned long max_ptes_none; | |
269 | ||
270 | err = kstrtoul(buf, 10, &max_ptes_none); | |
36ee2c78 | 271 | if (err || max_ptes_none > HPAGE_PMD_NR - 1) |
b46e756f KS |
272 | return -EINVAL; |
273 | ||
274 | khugepaged_max_ptes_none = max_ptes_none; | |
275 | ||
276 | return count; | |
277 | } | |
278 | static struct kobj_attribute khugepaged_max_ptes_none_attr = | |
6dcdc94d | 279 | __ATTR_RW(max_ptes_none); |
b46e756f | 280 | |
6dcdc94d ML |
281 | static ssize_t max_ptes_swap_show(struct kobject *kobj, |
282 | struct kobj_attribute *attr, | |
283 | char *buf) | |
b46e756f | 284 | { |
ae7a927d | 285 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_swap); |
b46e756f KS |
286 | } |
287 | ||
6dcdc94d ML |
288 | static ssize_t max_ptes_swap_store(struct kobject *kobj, |
289 | struct kobj_attribute *attr, | |
290 | const char *buf, size_t count) | |
b46e756f KS |
291 | { |
292 | int err; | |
293 | unsigned long max_ptes_swap; | |
294 | ||
295 | err = kstrtoul(buf, 10, &max_ptes_swap); | |
36ee2c78 | 296 | if (err || max_ptes_swap > HPAGE_PMD_NR - 1) |
b46e756f KS |
297 | return -EINVAL; |
298 | ||
299 | khugepaged_max_ptes_swap = max_ptes_swap; | |
300 | ||
301 | return count; | |
302 | } | |
303 | ||
304 | static struct kobj_attribute khugepaged_max_ptes_swap_attr = | |
6dcdc94d | 305 | __ATTR_RW(max_ptes_swap); |
b46e756f | 306 | |
6dcdc94d ML |
307 | static ssize_t max_ptes_shared_show(struct kobject *kobj, |
308 | struct kobj_attribute *attr, | |
309 | char *buf) | |
71a2c112 | 310 | { |
ae7a927d | 311 | return sysfs_emit(buf, "%u\n", khugepaged_max_ptes_shared); |
71a2c112 KS |
312 | } |
313 | ||
6dcdc94d ML |
314 | static ssize_t max_ptes_shared_store(struct kobject *kobj, |
315 | struct kobj_attribute *attr, | |
316 | const char *buf, size_t count) | |
71a2c112 KS |
317 | { |
318 | int err; | |
319 | unsigned long max_ptes_shared; | |
320 | ||
321 | err = kstrtoul(buf, 10, &max_ptes_shared); | |
36ee2c78 | 322 | if (err || max_ptes_shared > HPAGE_PMD_NR - 1) |
71a2c112 KS |
323 | return -EINVAL; |
324 | ||
325 | khugepaged_max_ptes_shared = max_ptes_shared; | |
326 | ||
327 | return count; | |
328 | } | |
329 | ||
330 | static struct kobj_attribute khugepaged_max_ptes_shared_attr = | |
6dcdc94d | 331 | __ATTR_RW(max_ptes_shared); |
71a2c112 | 332 | |
b46e756f KS |
333 | static struct attribute *khugepaged_attr[] = { |
334 | &khugepaged_defrag_attr.attr, | |
335 | &khugepaged_max_ptes_none_attr.attr, | |
71a2c112 KS |
336 | &khugepaged_max_ptes_swap_attr.attr, |
337 | &khugepaged_max_ptes_shared_attr.attr, | |
b46e756f KS |
338 | &pages_to_scan_attr.attr, |
339 | &pages_collapsed_attr.attr, | |
340 | &full_scans_attr.attr, | |
341 | &scan_sleep_millisecs_attr.attr, | |
342 | &alloc_sleep_millisecs_attr.attr, | |
b46e756f KS |
343 | NULL, |
344 | }; | |
345 | ||
346 | struct attribute_group khugepaged_attr_group = { | |
347 | .attrs = khugepaged_attr, | |
348 | .name = "khugepaged", | |
349 | }; | |
e1465d12 | 350 | #endif /* CONFIG_SYSFS */ |
b46e756f | 351 | |
b46e756f KS |
352 | int hugepage_madvise(struct vm_area_struct *vma, |
353 | unsigned long *vm_flags, int advice) | |
354 | { | |
355 | switch (advice) { | |
356 | case MADV_HUGEPAGE: | |
357 | #ifdef CONFIG_S390 | |
358 | /* | |
359 | * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390 | |
360 | * can't handle this properly after s390_enable_sie, so we simply | |
361 | * ignore the madvise to prevent qemu from causing a SIGSEGV. | |
362 | */ | |
363 | if (mm_has_pgste(vma->vm_mm)) | |
364 | return 0; | |
365 | #endif | |
366 | *vm_flags &= ~VM_NOHUGEPAGE; | |
367 | *vm_flags |= VM_HUGEPAGE; | |
368 | /* | |
369 | * If the vma become good for khugepaged to scan, | |
370 | * register it here without waiting a page fault that | |
371 | * may not happen any time soon. | |
372 | */ | |
c791576c | 373 | khugepaged_enter_vma(vma, *vm_flags); |
b46e756f KS |
374 | break; |
375 | case MADV_NOHUGEPAGE: | |
376 | *vm_flags &= ~VM_HUGEPAGE; | |
377 | *vm_flags |= VM_NOHUGEPAGE; | |
378 | /* | |
379 | * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning | |
380 | * this vma even if we leave the mm registered in khugepaged if | |
381 | * it got registered before VM_NOHUGEPAGE was set. | |
382 | */ | |
383 | break; | |
384 | } | |
385 | ||
386 | return 0; | |
387 | } | |
388 | ||
389 | int __init khugepaged_init(void) | |
390 | { | |
391 | mm_slot_cache = kmem_cache_create("khugepaged_mm_slot", | |
b26e2701 QZ |
392 | sizeof(struct khugepaged_mm_slot), |
393 | __alignof__(struct khugepaged_mm_slot), | |
394 | 0, NULL); | |
b46e756f KS |
395 | if (!mm_slot_cache) |
396 | return -ENOMEM; | |
397 | ||
398 | khugepaged_pages_to_scan = HPAGE_PMD_NR * 8; | |
399 | khugepaged_max_ptes_none = HPAGE_PMD_NR - 1; | |
400 | khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8; | |
71a2c112 | 401 | khugepaged_max_ptes_shared = HPAGE_PMD_NR / 2; |
b46e756f KS |
402 | |
403 | return 0; | |
404 | } | |
405 | ||
406 | void __init khugepaged_destroy(void) | |
407 | { | |
408 | kmem_cache_destroy(mm_slot_cache); | |
409 | } | |
410 | ||
7d2c4385 | 411 | static inline int hpage_collapse_test_exit(struct mm_struct *mm) |
b46e756f | 412 | { |
4d45e75a | 413 | return atomic_read(&mm->mm_users) == 0; |
b46e756f KS |
414 | } |
415 | ||
d2081b2b | 416 | void __khugepaged_enter(struct mm_struct *mm) |
b46e756f | 417 | { |
b26e2701 QZ |
418 | struct khugepaged_mm_slot *mm_slot; |
419 | struct mm_slot *slot; | |
b46e756f KS |
420 | int wakeup; |
421 | ||
b26e2701 | 422 | mm_slot = mm_slot_alloc(mm_slot_cache); |
b46e756f | 423 | if (!mm_slot) |
d2081b2b | 424 | return; |
b46e756f | 425 | |
b26e2701 QZ |
426 | slot = &mm_slot->slot; |
427 | ||
b46e756f | 428 | /* __khugepaged_exit() must not run from under us */ |
7d2c4385 | 429 | VM_BUG_ON_MM(hpage_collapse_test_exit(mm), mm); |
b46e756f | 430 | if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) { |
b26e2701 | 431 | mm_slot_free(mm_slot_cache, mm_slot); |
d2081b2b | 432 | return; |
b46e756f KS |
433 | } |
434 | ||
435 | spin_lock(&khugepaged_mm_lock); | |
b26e2701 | 436 | mm_slot_insert(mm_slots_hash, mm, slot); |
b46e756f KS |
437 | /* |
438 | * Insert just behind the scanning cursor, to let the area settle | |
439 | * down a little. | |
440 | */ | |
441 | wakeup = list_empty(&khugepaged_scan.mm_head); | |
b26e2701 | 442 | list_add_tail(&slot->mm_node, &khugepaged_scan.mm_head); |
b46e756f KS |
443 | spin_unlock(&khugepaged_mm_lock); |
444 | ||
f1f10076 | 445 | mmgrab(mm); |
b46e756f KS |
446 | if (wakeup) |
447 | wake_up_interruptible(&khugepaged_wait); | |
b46e756f KS |
448 | } |
449 | ||
c791576c YS |
450 | void khugepaged_enter_vma(struct vm_area_struct *vma, |
451 | unsigned long vm_flags) | |
b46e756f | 452 | { |
2647d11b | 453 | if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags) && |
1064026b | 454 | hugepage_flags_enabled()) { |
a7f4e6e4 | 455 | if (hugepage_vma_check(vma, vm_flags, false, false, true)) |
2647d11b YS |
456 | __khugepaged_enter(vma->vm_mm); |
457 | } | |
b46e756f KS |
458 | } |
459 | ||
460 | void __khugepaged_exit(struct mm_struct *mm) | |
461 | { | |
b26e2701 QZ |
462 | struct khugepaged_mm_slot *mm_slot; |
463 | struct mm_slot *slot; | |
b46e756f KS |
464 | int free = 0; |
465 | ||
466 | spin_lock(&khugepaged_mm_lock); | |
b26e2701 QZ |
467 | slot = mm_slot_lookup(mm_slots_hash, mm); |
468 | mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); | |
b46e756f | 469 | if (mm_slot && khugepaged_scan.mm_slot != mm_slot) { |
b26e2701 QZ |
470 | hash_del(&slot->hash); |
471 | list_del(&slot->mm_node); | |
b46e756f KS |
472 | free = 1; |
473 | } | |
474 | spin_unlock(&khugepaged_mm_lock); | |
475 | ||
476 | if (free) { | |
477 | clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
b26e2701 | 478 | mm_slot_free(mm_slot_cache, mm_slot); |
b46e756f KS |
479 | mmdrop(mm); |
480 | } else if (mm_slot) { | |
481 | /* | |
482 | * This is required to serialize against | |
7d2c4385 ZK |
483 | * hpage_collapse_test_exit() (which is guaranteed to run |
484 | * under mmap sem read mode). Stop here (after we return all | |
485 | * pagetables will be destroyed) until khugepaged has finished | |
486 | * working on the pagetables under the mmap_lock. | |
b46e756f | 487 | */ |
d8ed45c5 ML |
488 | mmap_write_lock(mm); |
489 | mmap_write_unlock(mm); | |
b46e756f KS |
490 | } |
491 | } | |
492 | ||
493 | static void release_pte_page(struct page *page) | |
494 | { | |
5503fbf2 KS |
495 | mod_node_page_state(page_pgdat(page), |
496 | NR_ISOLATED_ANON + page_is_file_lru(page), | |
497 | -compound_nr(page)); | |
b46e756f KS |
498 | unlock_page(page); |
499 | putback_lru_page(page); | |
500 | } | |
501 | ||
5503fbf2 KS |
502 | static void release_pte_pages(pte_t *pte, pte_t *_pte, |
503 | struct list_head *compound_pagelist) | |
b46e756f | 504 | { |
5503fbf2 KS |
505 | struct page *page, *tmp; |
506 | ||
b46e756f KS |
507 | while (--_pte >= pte) { |
508 | pte_t pteval = *_pte; | |
5503fbf2 KS |
509 | |
510 | page = pte_page(pteval); | |
511 | if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)) && | |
512 | !PageCompound(page)) | |
513 | release_pte_page(page); | |
514 | } | |
515 | ||
516 | list_for_each_entry_safe(page, tmp, compound_pagelist, lru) { | |
517 | list_del(&page->lru); | |
518 | release_pte_page(page); | |
b46e756f KS |
519 | } |
520 | } | |
521 | ||
9445689f KS |
522 | static bool is_refcount_suitable(struct page *page) |
523 | { | |
524 | int expected_refcount; | |
525 | ||
526 | expected_refcount = total_mapcount(page); | |
527 | if (PageSwapCache(page)) | |
528 | expected_refcount += compound_nr(page); | |
529 | ||
530 | return page_count(page) == expected_refcount; | |
531 | } | |
532 | ||
b46e756f KS |
533 | static int __collapse_huge_page_isolate(struct vm_area_struct *vma, |
534 | unsigned long address, | |
5503fbf2 | 535 | pte_t *pte, |
d8ea7cc8 | 536 | struct collapse_control *cc, |
5503fbf2 | 537 | struct list_head *compound_pagelist) |
b46e756f KS |
538 | { |
539 | struct page *page = NULL; | |
540 | pte_t *_pte; | |
50ad2f24 | 541 | int none_or_zero = 0, shared = 0, result = SCAN_FAIL, referenced = 0; |
0db501f7 | 542 | bool writable = false; |
b46e756f | 543 | |
36ee2c78 | 544 | for (_pte = pte; _pte < pte + HPAGE_PMD_NR; |
b46e756f KS |
545 | _pte++, address += PAGE_SIZE) { |
546 | pte_t pteval = *_pte; | |
547 | if (pte_none(pteval) || (pte_present(pteval) && | |
548 | is_zero_pfn(pte_pfn(pteval)))) { | |
d8ea7cc8 | 549 | ++none_or_zero; |
b46e756f | 550 | if (!userfaultfd_armed(vma) && |
d8ea7cc8 ZK |
551 | (!cc->is_khugepaged || |
552 | none_or_zero <= khugepaged_max_ptes_none)) { | |
b46e756f KS |
553 | continue; |
554 | } else { | |
555 | result = SCAN_EXCEED_NONE_PTE; | |
e9ea874a | 556 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
b46e756f KS |
557 | goto out; |
558 | } | |
559 | } | |
560 | if (!pte_present(pteval)) { | |
561 | result = SCAN_PTE_NON_PRESENT; | |
562 | goto out; | |
563 | } | |
564 | page = vm_normal_page(vma, address, pteval); | |
3218f871 | 565 | if (unlikely(!page) || unlikely(is_zone_device_page(page))) { |
b46e756f KS |
566 | result = SCAN_PAGE_NULL; |
567 | goto out; | |
568 | } | |
569 | ||
5503fbf2 KS |
570 | VM_BUG_ON_PAGE(!PageAnon(page), page); |
571 | ||
d8ea7cc8 ZK |
572 | if (page_mapcount(page) > 1) { |
573 | ++shared; | |
574 | if (cc->is_khugepaged && | |
575 | shared > khugepaged_max_ptes_shared) { | |
576 | result = SCAN_EXCEED_SHARED_PTE; | |
577 | count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); | |
578 | goto out; | |
579 | } | |
71a2c112 KS |
580 | } |
581 | ||
fece2029 | 582 | if (PageCompound(page)) { |
5503fbf2 KS |
583 | struct page *p; |
584 | page = compound_head(page); | |
fece2029 | 585 | |
5503fbf2 KS |
586 | /* |
587 | * Check if we have dealt with the compound page | |
588 | * already | |
589 | */ | |
590 | list_for_each_entry(p, compound_pagelist, lru) { | |
591 | if (page == p) | |
592 | goto next; | |
593 | } | |
594 | } | |
b46e756f KS |
595 | |
596 | /* | |
597 | * We can do it before isolate_lru_page because the | |
598 | * page can't be freed from under us. NOTE: PG_lock | |
599 | * is needed to serialize against split_huge_page | |
600 | * when invoked from the VM. | |
601 | */ | |
602 | if (!trylock_page(page)) { | |
603 | result = SCAN_PAGE_LOCK; | |
604 | goto out; | |
605 | } | |
606 | ||
607 | /* | |
9445689f KS |
608 | * Check if the page has any GUP (or other external) pins. |
609 | * | |
610 | * The page table that maps the page has been already unlinked | |
611 | * from the page table tree and this process cannot get | |
f0953a1b | 612 | * an additional pin on the page. |
9445689f KS |
613 | * |
614 | * New pins can come later if the page is shared across fork, | |
615 | * but not from this process. The other process cannot write to | |
616 | * the page, only trigger CoW. | |
b46e756f | 617 | */ |
9445689f | 618 | if (!is_refcount_suitable(page)) { |
b46e756f KS |
619 | unlock_page(page); |
620 | result = SCAN_PAGE_COUNT; | |
621 | goto out; | |
622 | } | |
b46e756f KS |
623 | |
624 | /* | |
625 | * Isolate the page to avoid collapsing an hugepage | |
626 | * currently in use by the VM. | |
627 | */ | |
628 | if (isolate_lru_page(page)) { | |
629 | unlock_page(page); | |
630 | result = SCAN_DEL_PAGE_LRU; | |
631 | goto out; | |
632 | } | |
5503fbf2 KS |
633 | mod_node_page_state(page_pgdat(page), |
634 | NR_ISOLATED_ANON + page_is_file_lru(page), | |
635 | compound_nr(page)); | |
b46e756f KS |
636 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
637 | VM_BUG_ON_PAGE(PageLRU(page), page); | |
638 | ||
5503fbf2 KS |
639 | if (PageCompound(page)) |
640 | list_add_tail(&page->lru, compound_pagelist); | |
641 | next: | |
d8ea7cc8 ZK |
642 | /* |
643 | * If collapse was initiated by khugepaged, check that there is | |
644 | * enough young pte to justify collapsing the page | |
645 | */ | |
646 | if (cc->is_khugepaged && | |
647 | (pte_young(pteval) || page_is_young(page) || | |
648 | PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm, | |
649 | address))) | |
0db501f7 | 650 | referenced++; |
5503fbf2 KS |
651 | |
652 | if (pte_write(pteval)) | |
653 | writable = true; | |
b46e756f | 654 | } |
74e579bf ML |
655 | |
656 | if (unlikely(!writable)) { | |
b46e756f | 657 | result = SCAN_PAGE_RO; |
d8ea7cc8 | 658 | } else if (unlikely(cc->is_khugepaged && !referenced)) { |
74e579bf ML |
659 | result = SCAN_LACK_REFERENCED_PAGE; |
660 | } else { | |
661 | result = SCAN_SUCCEED; | |
662 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, | |
663 | referenced, writable, result); | |
50ad2f24 | 664 | return result; |
b46e756f | 665 | } |
b46e756f | 666 | out: |
5503fbf2 | 667 | release_pte_pages(pte, _pte, compound_pagelist); |
b46e756f KS |
668 | trace_mm_collapse_huge_page_isolate(page, none_or_zero, |
669 | referenced, writable, result); | |
50ad2f24 | 670 | return result; |
b46e756f KS |
671 | } |
672 | ||
673 | static void __collapse_huge_page_copy(pte_t *pte, struct page *page, | |
674 | struct vm_area_struct *vma, | |
675 | unsigned long address, | |
5503fbf2 KS |
676 | spinlock_t *ptl, |
677 | struct list_head *compound_pagelist) | |
b46e756f | 678 | { |
5503fbf2 | 679 | struct page *src_page, *tmp; |
b46e756f | 680 | pte_t *_pte; |
338a16ba DR |
681 | for (_pte = pte; _pte < pte + HPAGE_PMD_NR; |
682 | _pte++, page++, address += PAGE_SIZE) { | |
b46e756f | 683 | pte_t pteval = *_pte; |
b46e756f KS |
684 | |
685 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
686 | clear_user_highpage(page, address); | |
687 | add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1); | |
688 | if (is_zero_pfn(pte_pfn(pteval))) { | |
689 | /* | |
690 | * ptl mostly unnecessary. | |
691 | */ | |
692 | spin_lock(ptl); | |
08d5b29e | 693 | ptep_clear(vma->vm_mm, address, _pte); |
b46e756f KS |
694 | spin_unlock(ptl); |
695 | } | |
696 | } else { | |
697 | src_page = pte_page(pteval); | |
698 | copy_user_highpage(page, src_page, address, vma); | |
5503fbf2 KS |
699 | if (!PageCompound(src_page)) |
700 | release_pte_page(src_page); | |
b46e756f KS |
701 | /* |
702 | * ptl mostly unnecessary, but preempt has to | |
703 | * be disabled to update the per-cpu stats | |
704 | * inside page_remove_rmap(). | |
705 | */ | |
706 | spin_lock(ptl); | |
08d5b29e | 707 | ptep_clear(vma->vm_mm, address, _pte); |
cea86fe2 | 708 | page_remove_rmap(src_page, vma, false); |
b46e756f KS |
709 | spin_unlock(ptl); |
710 | free_page_and_swap_cache(src_page); | |
711 | } | |
b46e756f | 712 | } |
5503fbf2 KS |
713 | |
714 | list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) { | |
715 | list_del(&src_page->lru); | |
1baec203 ML |
716 | mod_node_page_state(page_pgdat(src_page), |
717 | NR_ISOLATED_ANON + page_is_file_lru(src_page), | |
718 | -compound_nr(src_page)); | |
719 | unlock_page(src_page); | |
720 | free_swap_cache(src_page); | |
721 | putback_lru_page(src_page); | |
5503fbf2 | 722 | } |
b46e756f KS |
723 | } |
724 | ||
725 | static void khugepaged_alloc_sleep(void) | |
726 | { | |
727 | DEFINE_WAIT(wait); | |
728 | ||
729 | add_wait_queue(&khugepaged_wait, &wait); | |
f5d39b02 PZ |
730 | __set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE); |
731 | schedule_timeout(msecs_to_jiffies(khugepaged_alloc_sleep_millisecs)); | |
b46e756f KS |
732 | remove_wait_queue(&khugepaged_wait, &wait); |
733 | } | |
734 | ||
34d6b470 | 735 | struct collapse_control khugepaged_collapse_control = { |
d8ea7cc8 | 736 | .is_khugepaged = true, |
34d6b470 | 737 | }; |
b46e756f | 738 | |
7d2c4385 | 739 | static bool hpage_collapse_scan_abort(int nid, struct collapse_control *cc) |
b46e756f KS |
740 | { |
741 | int i; | |
742 | ||
743 | /* | |
a5f5f91d | 744 | * If node_reclaim_mode is disabled, then no extra effort is made to |
b46e756f KS |
745 | * allocate memory locally. |
746 | */ | |
202e35db | 747 | if (!node_reclaim_enabled()) |
b46e756f KS |
748 | return false; |
749 | ||
750 | /* If there is a count for this node already, it must be acceptable */ | |
34d6b470 | 751 | if (cc->node_load[nid]) |
b46e756f KS |
752 | return false; |
753 | ||
754 | for (i = 0; i < MAX_NUMNODES; i++) { | |
34d6b470 | 755 | if (!cc->node_load[i]) |
b46e756f | 756 | continue; |
a55c7454 | 757 | if (node_distance(nid, i) > node_reclaim_distance) |
b46e756f KS |
758 | return true; |
759 | } | |
760 | return false; | |
761 | } | |
762 | ||
1064026b YS |
763 | #define khugepaged_defrag() \ |
764 | (transparent_hugepage_flags & \ | |
765 | (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)) | |
766 | ||
b46e756f KS |
767 | /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */ |
768 | static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void) | |
769 | { | |
25160354 | 770 | return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT; |
b46e756f KS |
771 | } |
772 | ||
773 | #ifdef CONFIG_NUMA | |
7d2c4385 | 774 | static int hpage_collapse_find_target_node(struct collapse_control *cc) |
b46e756f | 775 | { |
b46e756f KS |
776 | int nid, target_node = 0, max_value = 0; |
777 | ||
778 | /* find first node with max normal pages hit */ | |
779 | for (nid = 0; nid < MAX_NUMNODES; nid++) | |
34d6b470 ZK |
780 | if (cc->node_load[nid] > max_value) { |
781 | max_value = cc->node_load[nid]; | |
b46e756f KS |
782 | target_node = nid; |
783 | } | |
784 | ||
e031ff96 YS |
785 | for_each_online_node(nid) { |
786 | if (max_value == cc->node_load[nid]) | |
787 | node_set(nid, cc->alloc_nmask); | |
788 | } | |
b46e756f | 789 | |
b46e756f KS |
790 | return target_node; |
791 | } | |
c6a7f445 | 792 | #else |
7d2c4385 | 793 | static int hpage_collapse_find_target_node(struct collapse_control *cc) |
b46e756f | 794 | { |
c6a7f445 | 795 | return 0; |
b46e756f | 796 | } |
c6a7f445 | 797 | #endif |
b46e756f | 798 | |
e031ff96 YS |
799 | static bool hpage_collapse_alloc_page(struct page **hpage, gfp_t gfp, int node, |
800 | nodemask_t *nmask) | |
b46e756f | 801 | { |
e031ff96 | 802 | *hpage = __alloc_pages(gfp, HPAGE_PMD_ORDER, node, nmask); |
b46e756f KS |
803 | if (unlikely(!*hpage)) { |
804 | count_vm_event(THP_COLLAPSE_ALLOC_FAILED); | |
9710a78a | 805 | return false; |
b46e756f KS |
806 | } |
807 | ||
808 | prep_transhuge_page(*hpage); | |
809 | count_vm_event(THP_COLLAPSE_ALLOC); | |
b46e756f KS |
810 | return true; |
811 | } | |
812 | ||
b46e756f | 813 | /* |
c1e8d7c6 ML |
814 | * If mmap_lock temporarily dropped, revalidate vma |
815 | * before taking mmap_lock. | |
50ad2f24 | 816 | * Returns enum scan_result value. |
b46e756f KS |
817 | */ |
818 | ||
c131f751 | 819 | static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, |
34488399 | 820 | bool expect_anon, |
a7f4e6e4 ZK |
821 | struct vm_area_struct **vmap, |
822 | struct collapse_control *cc) | |
b46e756f KS |
823 | { |
824 | struct vm_area_struct *vma; | |
b46e756f | 825 | |
7d2c4385 | 826 | if (unlikely(hpage_collapse_test_exit(mm))) |
b46e756f KS |
827 | return SCAN_ANY_PROCESS; |
828 | ||
c131f751 | 829 | *vmap = vma = find_vma(mm, address); |
b46e756f KS |
830 | if (!vma) |
831 | return SCAN_VMA_NULL; | |
832 | ||
4fa6893f | 833 | if (!transhuge_vma_suitable(vma, address)) |
b46e756f | 834 | return SCAN_ADDRESS_RANGE; |
a7f4e6e4 ZK |
835 | if (!hugepage_vma_check(vma, vma->vm_flags, false, false, |
836 | cc->is_khugepaged)) | |
b46e756f | 837 | return SCAN_VMA_CHECK; |
f707fa49 YS |
838 | /* |
839 | * Anon VMA expected, the address may be unmapped then | |
840 | * remapped to file after khugepaged reaquired the mmap_lock. | |
841 | * | |
842 | * hugepage_vma_check may return true for qualified file | |
843 | * vmas. | |
844 | */ | |
34488399 ZK |
845 | if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap))) |
846 | return SCAN_PAGE_ANON; | |
50ad2f24 | 847 | return SCAN_SUCCEED; |
b46e756f KS |
848 | } |
849 | ||
50722804 ZK |
850 | static int find_pmd_or_thp_or_none(struct mm_struct *mm, |
851 | unsigned long address, | |
852 | pmd_t **pmd) | |
853 | { | |
854 | pmd_t pmde; | |
855 | ||
856 | *pmd = mm_find_pmd(mm, address); | |
857 | if (!*pmd) | |
858 | return SCAN_PMD_NULL; | |
859 | ||
dab6e717 | 860 | pmde = pmdp_get_lockless(*pmd); |
50722804 ZK |
861 | |
862 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
863 | /* See comments in pmd_none_or_trans_huge_or_clear_bad() */ | |
864 | barrier(); | |
865 | #endif | |
34488399 ZK |
866 | if (pmd_none(pmde)) |
867 | return SCAN_PMD_NONE; | |
50722804 ZK |
868 | if (pmd_trans_huge(pmde)) |
869 | return SCAN_PMD_MAPPED; | |
870 | if (pmd_bad(pmde)) | |
871 | return SCAN_PMD_NULL; | |
872 | return SCAN_SUCCEED; | |
873 | } | |
874 | ||
875 | static int check_pmd_still_valid(struct mm_struct *mm, | |
876 | unsigned long address, | |
877 | pmd_t *pmd) | |
878 | { | |
879 | pmd_t *new_pmd; | |
880 | int result = find_pmd_or_thp_or_none(mm, address, &new_pmd); | |
881 | ||
882 | if (result != SCAN_SUCCEED) | |
883 | return result; | |
884 | if (new_pmd != pmd) | |
885 | return SCAN_FAIL; | |
886 | return SCAN_SUCCEED; | |
b46e756f KS |
887 | } |
888 | ||
889 | /* | |
890 | * Bring missing pages in from swap, to complete THP collapse. | |
7d2c4385 | 891 | * Only done if hpage_collapse_scan_pmd believes it is worthwhile. |
b46e756f | 892 | * |
4d928e20 ML |
893 | * Called and returns without pte mapped or spinlocks held. |
894 | * Note that if false is returned, mmap_lock will be released. | |
b46e756f KS |
895 | */ |
896 | ||
50ad2f24 ZK |
897 | static int __collapse_huge_page_swapin(struct mm_struct *mm, |
898 | struct vm_area_struct *vma, | |
899 | unsigned long haddr, pmd_t *pmd, | |
900 | int referenced) | |
b46e756f | 901 | { |
2b740303 SJ |
902 | int swapped_in = 0; |
903 | vm_fault_t ret = 0; | |
2b635dd3 WD |
904 | unsigned long address, end = haddr + (HPAGE_PMD_NR * PAGE_SIZE); |
905 | ||
906 | for (address = haddr; address < end; address += PAGE_SIZE) { | |
907 | struct vm_fault vmf = { | |
908 | .vma = vma, | |
909 | .address = address, | |
910 | .pgoff = linear_page_index(vma, haddr), | |
911 | .flags = FAULT_FLAG_ALLOW_RETRY, | |
912 | .pmd = pmd, | |
913 | }; | |
914 | ||
915 | vmf.pte = pte_offset_map(pmd, address); | |
2994302b | 916 | vmf.orig_pte = *vmf.pte; |
2b635dd3 WD |
917 | if (!is_swap_pte(vmf.orig_pte)) { |
918 | pte_unmap(vmf.pte); | |
b46e756f | 919 | continue; |
2b635dd3 | 920 | } |
2994302b | 921 | ret = do_swap_page(&vmf); |
0db501f7 | 922 | |
4d928e20 ML |
923 | /* |
924 | * do_swap_page returns VM_FAULT_RETRY with released mmap_lock. | |
925 | * Note we treat VM_FAULT_RETRY as VM_FAULT_ERROR here because | |
926 | * we do not retry here and swap entry will remain in pagetable | |
927 | * resulting in later failure. | |
928 | */ | |
b46e756f | 929 | if (ret & VM_FAULT_RETRY) { |
4d928e20 | 930 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); |
50ad2f24 ZK |
931 | /* Likely, but not guaranteed, that page lock failed */ |
932 | return SCAN_PAGE_LOCK; | |
b46e756f KS |
933 | } |
934 | if (ret & VM_FAULT_ERROR) { | |
4d928e20 | 935 | mmap_read_unlock(mm); |
0db501f7 | 936 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0); |
50ad2f24 | 937 | return SCAN_FAIL; |
b46e756f | 938 | } |
4d928e20 | 939 | swapped_in++; |
b46e756f | 940 | } |
ae2c5d80 KS |
941 | |
942 | /* Drain LRU add pagevec to remove extra pin on the swapped in pages */ | |
943 | if (swapped_in) | |
944 | lru_add_drain(); | |
945 | ||
0db501f7 | 946 | trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1); |
50ad2f24 | 947 | return SCAN_SUCCEED; |
b46e756f KS |
948 | } |
949 | ||
9710a78a ZK |
950 | static int alloc_charge_hpage(struct page **hpage, struct mm_struct *mm, |
951 | struct collapse_control *cc) | |
952 | { | |
7d8faaf1 | 953 | gfp_t gfp = (cc->is_khugepaged ? alloc_hugepage_khugepaged_gfpmask() : |
e031ff96 | 954 | GFP_TRANSHUGE); |
7d2c4385 | 955 | int node = hpage_collapse_find_target_node(cc); |
9710a78a | 956 | |
e031ff96 | 957 | if (!hpage_collapse_alloc_page(hpage, gfp, node, &cc->alloc_nmask)) |
9710a78a ZK |
958 | return SCAN_ALLOC_HUGE_PAGE_FAIL; |
959 | if (unlikely(mem_cgroup_charge(page_folio(*hpage), mm, gfp))) | |
960 | return SCAN_CGROUP_CHARGE_FAIL; | |
961 | count_memcg_page_event(*hpage, THP_COLLAPSE_ALLOC); | |
962 | return SCAN_SUCCEED; | |
963 | } | |
964 | ||
50ad2f24 ZK |
965 | static int collapse_huge_page(struct mm_struct *mm, unsigned long address, |
966 | int referenced, int unmapped, | |
967 | struct collapse_control *cc) | |
b46e756f | 968 | { |
5503fbf2 | 969 | LIST_HEAD(compound_pagelist); |
b46e756f KS |
970 | pmd_t *pmd, _pmd; |
971 | pte_t *pte; | |
972 | pgtable_t pgtable; | |
50ad2f24 | 973 | struct page *hpage; |
b46e756f | 974 | spinlock_t *pmd_ptl, *pte_ptl; |
50ad2f24 | 975 | int result = SCAN_FAIL; |
c131f751 | 976 | struct vm_area_struct *vma; |
ac46d4f3 | 977 | struct mmu_notifier_range range; |
b46e756f KS |
978 | |
979 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
980 | ||
988ddb71 | 981 | /* |
c1e8d7c6 | 982 | * Before allocating the hugepage, release the mmap_lock read lock. |
988ddb71 | 983 | * The allocation can take potentially a long time if it involves |
c1e8d7c6 | 984 | * sync compaction, and we do not need to hold the mmap_lock during |
988ddb71 KS |
985 | * that. We will recheck the vma after taking it again in write mode. |
986 | */ | |
d8ed45c5 | 987 | mmap_read_unlock(mm); |
b46e756f | 988 | |
50ad2f24 | 989 | result = alloc_charge_hpage(&hpage, mm, cc); |
9710a78a | 990 | if (result != SCAN_SUCCEED) |
b46e756f | 991 | goto out_nolock; |
b46e756f | 992 | |
d8ed45c5 | 993 | mmap_read_lock(mm); |
34488399 | 994 | result = hugepage_vma_revalidate(mm, address, true, &vma, cc); |
50ad2f24 | 995 | if (result != SCAN_SUCCEED) { |
d8ed45c5 | 996 | mmap_read_unlock(mm); |
b46e756f KS |
997 | goto out_nolock; |
998 | } | |
999 | ||
50722804 ZK |
1000 | result = find_pmd_or_thp_or_none(mm, address, &pmd); |
1001 | if (result != SCAN_SUCCEED) { | |
d8ed45c5 | 1002 | mmap_read_unlock(mm); |
b46e756f KS |
1003 | goto out_nolock; |
1004 | } | |
1005 | ||
50ad2f24 ZK |
1006 | if (unmapped) { |
1007 | /* | |
1008 | * __collapse_huge_page_swapin will return with mmap_lock | |
1009 | * released when it fails. So we jump out_nolock directly in | |
1010 | * that case. Continuing to collapse causes inconsistency. | |
1011 | */ | |
1012 | result = __collapse_huge_page_swapin(mm, vma, address, pmd, | |
1013 | referenced); | |
1014 | if (result != SCAN_SUCCEED) | |
1015 | goto out_nolock; | |
b46e756f KS |
1016 | } |
1017 | ||
d8ed45c5 | 1018 | mmap_read_unlock(mm); |
b46e756f KS |
1019 | /* |
1020 | * Prevent all access to pagetables with the exception of | |
1021 | * gup_fast later handled by the ptep_clear_flush and the VM | |
1022 | * handled by the anon_vma lock + PG_lock. | |
1023 | */ | |
d8ed45c5 | 1024 | mmap_write_lock(mm); |
34488399 | 1025 | result = hugepage_vma_revalidate(mm, address, true, &vma, cc); |
50ad2f24 | 1026 | if (result != SCAN_SUCCEED) |
18d24a7c | 1027 | goto out_up_write; |
b46e756f | 1028 | /* check if the pmd is still valid */ |
50722804 ZK |
1029 | result = check_pmd_still_valid(mm, address, pmd); |
1030 | if (result != SCAN_SUCCEED) | |
18d24a7c | 1031 | goto out_up_write; |
b46e756f KS |
1032 | |
1033 | anon_vma_lock_write(vma->anon_vma); | |
1034 | ||
7269f999 | 1035 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, |
6f4f13e8 | 1036 | address, address + HPAGE_PMD_SIZE); |
ac46d4f3 | 1037 | mmu_notifier_invalidate_range_start(&range); |
ec649c9d VS |
1038 | |
1039 | pte = pte_offset_map(pmd, address); | |
1040 | pte_ptl = pte_lockptr(mm, pmd); | |
1041 | ||
b46e756f KS |
1042 | pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */ |
1043 | /* | |
70cbc3cc YS |
1044 | * This removes any huge TLB entry from the CPU so we won't allow |
1045 | * huge and small TLB entries for the same virtual address to | |
1046 | * avoid the risk of CPU bugs in that area. | |
1047 | * | |
1048 | * Parallel fast GUP is fine since fast GUP will back off when | |
1049 | * it detects PMD is changed. | |
b46e756f KS |
1050 | */ |
1051 | _pmd = pmdp_collapse_flush(vma, address, pmd); | |
1052 | spin_unlock(pmd_ptl); | |
ac46d4f3 | 1053 | mmu_notifier_invalidate_range_end(&range); |
2ba99c5e | 1054 | tlb_remove_table_sync_one(); |
b46e756f KS |
1055 | |
1056 | spin_lock(pte_ptl); | |
d8ea7cc8 | 1057 | result = __collapse_huge_page_isolate(vma, address, pte, cc, |
50ad2f24 | 1058 | &compound_pagelist); |
b46e756f KS |
1059 | spin_unlock(pte_ptl); |
1060 | ||
50ad2f24 | 1061 | if (unlikely(result != SCAN_SUCCEED)) { |
b46e756f KS |
1062 | pte_unmap(pte); |
1063 | spin_lock(pmd_ptl); | |
1064 | BUG_ON(!pmd_none(*pmd)); | |
1065 | /* | |
1066 | * We can only use set_pmd_at when establishing | |
1067 | * hugepmds and never for establishing regular pmds that | |
1068 | * points to regular pagetables. Use pmd_populate for that | |
1069 | */ | |
1070 | pmd_populate(mm, pmd, pmd_pgtable(_pmd)); | |
1071 | spin_unlock(pmd_ptl); | |
1072 | anon_vma_unlock_write(vma->anon_vma); | |
18d24a7c | 1073 | goto out_up_write; |
b46e756f KS |
1074 | } |
1075 | ||
1076 | /* | |
1077 | * All pages are isolated and locked so anon_vma rmap | |
1078 | * can't run anymore. | |
1079 | */ | |
1080 | anon_vma_unlock_write(vma->anon_vma); | |
1081 | ||
50ad2f24 ZK |
1082 | __collapse_huge_page_copy(pte, hpage, vma, address, pte_ptl, |
1083 | &compound_pagelist); | |
b46e756f | 1084 | pte_unmap(pte); |
588d01f9 ML |
1085 | /* |
1086 | * spin_lock() below is not the equivalent of smp_wmb(), but | |
1087 | * the smp_wmb() inside __SetPageUptodate() can be reused to | |
1088 | * avoid the copy_huge_page writes to become visible after | |
1089 | * the set_pmd_at() write. | |
1090 | */ | |
50ad2f24 | 1091 | __SetPageUptodate(hpage); |
b46e756f KS |
1092 | pgtable = pmd_pgtable(_pmd); |
1093 | ||
50ad2f24 | 1094 | _pmd = mk_huge_pmd(hpage, vma->vm_page_prot); |
f55e1014 | 1095 | _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma); |
b46e756f | 1096 | |
b46e756f KS |
1097 | spin_lock(pmd_ptl); |
1098 | BUG_ON(!pmd_none(*pmd)); | |
50ad2f24 ZK |
1099 | page_add_new_anon_rmap(hpage, vma, address); |
1100 | lru_cache_add_inactive_or_unevictable(hpage, vma); | |
b46e756f KS |
1101 | pgtable_trans_huge_deposit(mm, pmd, pgtable); |
1102 | set_pmd_at(mm, address, pmd, _pmd); | |
1103 | update_mmu_cache_pmd(vma, address, pmd); | |
1104 | spin_unlock(pmd_ptl); | |
1105 | ||
50ad2f24 | 1106 | hpage = NULL; |
b46e756f | 1107 | |
b46e756f KS |
1108 | result = SCAN_SUCCEED; |
1109 | out_up_write: | |
d8ed45c5 | 1110 | mmap_write_unlock(mm); |
b46e756f | 1111 | out_nolock: |
50ad2f24 ZK |
1112 | if (hpage) { |
1113 | mem_cgroup_uncharge(page_folio(hpage)); | |
1114 | put_page(hpage); | |
c6a7f445 | 1115 | } |
50ad2f24 ZK |
1116 | trace_mm_collapse_huge_page(mm, result == SCAN_SUCCEED, result); |
1117 | return result; | |
b46e756f KS |
1118 | } |
1119 | ||
7d2c4385 ZK |
1120 | static int hpage_collapse_scan_pmd(struct mm_struct *mm, |
1121 | struct vm_area_struct *vma, | |
1122 | unsigned long address, bool *mmap_locked, | |
1123 | struct collapse_control *cc) | |
b46e756f KS |
1124 | { |
1125 | pmd_t *pmd; | |
1126 | pte_t *pte, *_pte; | |
50ad2f24 | 1127 | int result = SCAN_FAIL, referenced = 0; |
71a2c112 | 1128 | int none_or_zero = 0, shared = 0; |
b46e756f KS |
1129 | struct page *page = NULL; |
1130 | unsigned long _address; | |
1131 | spinlock_t *ptl; | |
1132 | int node = NUMA_NO_NODE, unmapped = 0; | |
0db501f7 | 1133 | bool writable = false; |
b46e756f KS |
1134 | |
1135 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
1136 | ||
50722804 ZK |
1137 | result = find_pmd_or_thp_or_none(mm, address, &pmd); |
1138 | if (result != SCAN_SUCCEED) | |
b46e756f | 1139 | goto out; |
b46e756f | 1140 | |
34d6b470 | 1141 | memset(cc->node_load, 0, sizeof(cc->node_load)); |
e031ff96 | 1142 | nodes_clear(cc->alloc_nmask); |
b46e756f | 1143 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
36ee2c78 | 1144 | for (_address = address, _pte = pte; _pte < pte + HPAGE_PMD_NR; |
b46e756f KS |
1145 | _pte++, _address += PAGE_SIZE) { |
1146 | pte_t pteval = *_pte; | |
1147 | if (is_swap_pte(pteval)) { | |
d8ea7cc8 ZK |
1148 | ++unmapped; |
1149 | if (!cc->is_khugepaged || | |
1150 | unmapped <= khugepaged_max_ptes_swap) { | |
e1e267c7 PX |
1151 | /* |
1152 | * Always be strict with uffd-wp | |
1153 | * enabled swap entries. Please see | |
1154 | * comment below for pte_uffd_wp(). | |
1155 | */ | |
1156 | if (pte_swp_uffd_wp(pteval)) { | |
1157 | result = SCAN_PTE_UFFD_WP; | |
1158 | goto out_unmap; | |
1159 | } | |
b46e756f KS |
1160 | continue; |
1161 | } else { | |
1162 | result = SCAN_EXCEED_SWAP_PTE; | |
e9ea874a | 1163 | count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); |
b46e756f KS |
1164 | goto out_unmap; |
1165 | } | |
1166 | } | |
1167 | if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) { | |
d8ea7cc8 | 1168 | ++none_or_zero; |
b46e756f | 1169 | if (!userfaultfd_armed(vma) && |
d8ea7cc8 ZK |
1170 | (!cc->is_khugepaged || |
1171 | none_or_zero <= khugepaged_max_ptes_none)) { | |
b46e756f KS |
1172 | continue; |
1173 | } else { | |
1174 | result = SCAN_EXCEED_NONE_PTE; | |
e9ea874a | 1175 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
b46e756f KS |
1176 | goto out_unmap; |
1177 | } | |
1178 | } | |
e1e267c7 PX |
1179 | if (pte_uffd_wp(pteval)) { |
1180 | /* | |
1181 | * Don't collapse the page if any of the small | |
1182 | * PTEs are armed with uffd write protection. | |
1183 | * Here we can also mark the new huge pmd as | |
1184 | * write protected if any of the small ones is | |
8958b249 | 1185 | * marked but that could bring unknown |
e1e267c7 PX |
1186 | * userfault messages that falls outside of |
1187 | * the registered range. So, just be simple. | |
1188 | */ | |
1189 | result = SCAN_PTE_UFFD_WP; | |
1190 | goto out_unmap; | |
1191 | } | |
b46e756f KS |
1192 | if (pte_write(pteval)) |
1193 | writable = true; | |
1194 | ||
1195 | page = vm_normal_page(vma, _address, pteval); | |
3218f871 | 1196 | if (unlikely(!page) || unlikely(is_zone_device_page(page))) { |
b46e756f KS |
1197 | result = SCAN_PAGE_NULL; |
1198 | goto out_unmap; | |
1199 | } | |
1200 | ||
d8ea7cc8 ZK |
1201 | if (page_mapcount(page) > 1) { |
1202 | ++shared; | |
1203 | if (cc->is_khugepaged && | |
1204 | shared > khugepaged_max_ptes_shared) { | |
1205 | result = SCAN_EXCEED_SHARED_PTE; | |
1206 | count_vm_event(THP_SCAN_EXCEED_SHARED_PTE); | |
1207 | goto out_unmap; | |
1208 | } | |
71a2c112 KS |
1209 | } |
1210 | ||
5503fbf2 | 1211 | page = compound_head(page); |
b46e756f KS |
1212 | |
1213 | /* | |
1214 | * Record which node the original page is from and save this | |
34d6b470 | 1215 | * information to cc->node_load[]. |
0b8f0d87 | 1216 | * Khugepaged will allocate hugepage from the node has the max |
b46e756f KS |
1217 | * hit record. |
1218 | */ | |
1219 | node = page_to_nid(page); | |
7d2c4385 | 1220 | if (hpage_collapse_scan_abort(node, cc)) { |
b46e756f KS |
1221 | result = SCAN_SCAN_ABORT; |
1222 | goto out_unmap; | |
1223 | } | |
34d6b470 | 1224 | cc->node_load[node]++; |
b46e756f KS |
1225 | if (!PageLRU(page)) { |
1226 | result = SCAN_PAGE_LRU; | |
1227 | goto out_unmap; | |
1228 | } | |
1229 | if (PageLocked(page)) { | |
1230 | result = SCAN_PAGE_LOCK; | |
1231 | goto out_unmap; | |
1232 | } | |
1233 | if (!PageAnon(page)) { | |
1234 | result = SCAN_PAGE_ANON; | |
1235 | goto out_unmap; | |
1236 | } | |
1237 | ||
1238 | /* | |
9445689f KS |
1239 | * Check if the page has any GUP (or other external) pins. |
1240 | * | |
cb67f428 HD |
1241 | * Here the check may be racy: |
1242 | * it may see total_mapcount > refcount in some cases? | |
9445689f KS |
1243 | * But such case is ephemeral we could always retry collapse |
1244 | * later. However it may report false positive if the page | |
1245 | * has excessive GUP pins (i.e. 512). Anyway the same check | |
1246 | * will be done again later the risk seems low. | |
b46e756f | 1247 | */ |
9445689f | 1248 | if (!is_refcount_suitable(page)) { |
b46e756f KS |
1249 | result = SCAN_PAGE_COUNT; |
1250 | goto out_unmap; | |
1251 | } | |
d8ea7cc8 ZK |
1252 | |
1253 | /* | |
1254 | * If collapse was initiated by khugepaged, check that there is | |
1255 | * enough young pte to justify collapsing the page | |
1256 | */ | |
1257 | if (cc->is_khugepaged && | |
1258 | (pte_young(pteval) || page_is_young(page) || | |
1259 | PageReferenced(page) || mmu_notifier_test_young(vma->vm_mm, | |
1260 | address))) | |
0db501f7 | 1261 | referenced++; |
b46e756f | 1262 | } |
ffe945e6 | 1263 | if (!writable) { |
b46e756f | 1264 | result = SCAN_PAGE_RO; |
d8ea7cc8 ZK |
1265 | } else if (cc->is_khugepaged && |
1266 | (!referenced || | |
1267 | (unmapped && referenced < HPAGE_PMD_NR / 2))) { | |
ffe945e6 KS |
1268 | result = SCAN_LACK_REFERENCED_PAGE; |
1269 | } else { | |
1270 | result = SCAN_SUCCEED; | |
b46e756f KS |
1271 | } |
1272 | out_unmap: | |
1273 | pte_unmap_unlock(pte, ptl); | |
50ad2f24 ZK |
1274 | if (result == SCAN_SUCCEED) { |
1275 | result = collapse_huge_page(mm, address, referenced, | |
1276 | unmapped, cc); | |
c1e8d7c6 | 1277 | /* collapse_huge_page will return with the mmap_lock released */ |
50ad2f24 | 1278 | *mmap_locked = false; |
b46e756f KS |
1279 | } |
1280 | out: | |
1281 | trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced, | |
1282 | none_or_zero, result, unmapped); | |
50ad2f24 | 1283 | return result; |
b46e756f KS |
1284 | } |
1285 | ||
b26e2701 | 1286 | static void collect_mm_slot(struct khugepaged_mm_slot *mm_slot) |
b46e756f | 1287 | { |
b26e2701 QZ |
1288 | struct mm_slot *slot = &mm_slot->slot; |
1289 | struct mm_struct *mm = slot->mm; | |
b46e756f | 1290 | |
35f3aa39 | 1291 | lockdep_assert_held(&khugepaged_mm_lock); |
b46e756f | 1292 | |
7d2c4385 | 1293 | if (hpage_collapse_test_exit(mm)) { |
b46e756f | 1294 | /* free mm_slot */ |
b26e2701 QZ |
1295 | hash_del(&slot->hash); |
1296 | list_del(&slot->mm_node); | |
b46e756f KS |
1297 | |
1298 | /* | |
1299 | * Not strictly needed because the mm exited already. | |
1300 | * | |
1301 | * clear_bit(MMF_VM_HUGEPAGE, &mm->flags); | |
1302 | */ | |
1303 | ||
1304 | /* khugepaged_mm_lock actually not necessary for the below */ | |
b26e2701 | 1305 | mm_slot_free(mm_slot_cache, mm_slot); |
b46e756f KS |
1306 | mmdrop(mm); |
1307 | } | |
1308 | } | |
1309 | ||
396bcc52 | 1310 | #ifdef CONFIG_SHMEM |
27e1f827 SL |
1311 | /* |
1312 | * Notify khugepaged that given addr of the mm is pte-mapped THP. Then | |
1313 | * khugepaged should try to collapse the page table. | |
34488399 ZK |
1314 | * |
1315 | * Note that following race exists: | |
1316 | * (1) khugepaged calls khugepaged_collapse_pte_mapped_thps() for mm_struct A, | |
1317 | * emptying the A's ->pte_mapped_thp[] array. | |
1318 | * (2) MADV_COLLAPSE collapses some file extent with target mm_struct B, and | |
1319 | * retract_page_tables() finds a VMA in mm_struct A mapping the same extent | |
1320 | * (at virtual address X) and adds an entry (for X) into mm_struct A's | |
1321 | * ->pte-mapped_thp[] array. | |
1322 | * (3) khugepaged calls khugepaged_collapse_scan_file() for mm_struct A at X, | |
1323 | * sees a pte-mapped THP (SCAN_PTE_MAPPED_HUGEPAGE) and adds an entry | |
1324 | * (for X) into mm_struct A's ->pte-mapped_thp[] array. | |
1325 | * Thus, it's possible the same address is added multiple times for the same | |
1326 | * mm_struct. Should this happen, we'll simply attempt | |
1327 | * collapse_pte_mapped_thp() multiple times for the same address, under the same | |
1328 | * exclusive mmap_lock, and assuming the first call is successful, subsequent | |
1329 | * attempts will return quickly (without grabbing any additional locks) when | |
1330 | * a huge pmd is found in find_pmd_or_thp_or_none(). Since this is a cheap | |
1331 | * check, and since this is a rare occurrence, the cost of preventing this | |
1332 | * "multiple-add" is thought to be more expensive than just handling it, should | |
1333 | * it occur. | |
27e1f827 | 1334 | */ |
58ac9a89 | 1335 | static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm, |
081c3256 | 1336 | unsigned long addr) |
27e1f827 | 1337 | { |
b26e2701 QZ |
1338 | struct khugepaged_mm_slot *mm_slot; |
1339 | struct mm_slot *slot; | |
58ac9a89 | 1340 | bool ret = false; |
27e1f827 SL |
1341 | |
1342 | VM_BUG_ON(addr & ~HPAGE_PMD_MASK); | |
1343 | ||
1344 | spin_lock(&khugepaged_mm_lock); | |
b26e2701 QZ |
1345 | slot = mm_slot_lookup(mm_slots_hash, mm); |
1346 | mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); | |
58ac9a89 | 1347 | if (likely(mm_slot && mm_slot->nr_pte_mapped_thp < MAX_PTE_MAPPED_THP)) { |
27e1f827 | 1348 | mm_slot->pte_mapped_thp[mm_slot->nr_pte_mapped_thp++] = addr; |
58ac9a89 ZK |
1349 | ret = true; |
1350 | } | |
27e1f827 | 1351 | spin_unlock(&khugepaged_mm_lock); |
58ac9a89 | 1352 | return ret; |
27e1f827 SL |
1353 | } |
1354 | ||
34488399 ZK |
1355 | /* hpage must be locked, and mmap_lock must be held in write */ |
1356 | static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, | |
1357 | pmd_t *pmdp, struct page *hpage) | |
1358 | { | |
1359 | struct vm_fault vmf = { | |
1360 | .vma = vma, | |
1361 | .address = addr, | |
1362 | .flags = 0, | |
1363 | .pmd = pmdp, | |
1364 | }; | |
1365 | ||
1366 | VM_BUG_ON(!PageTransHuge(hpage)); | |
1367 | mmap_assert_write_locked(vma->vm_mm); | |
1368 | ||
1369 | if (do_set_pmd(&vmf, hpage)) | |
1370 | return SCAN_FAIL; | |
1371 | ||
1372 | get_page(hpage); | |
1373 | return SCAN_SUCCEED; | |
27e1f827 SL |
1374 | } |
1375 | ||
8d3c106e JH |
1376 | /* |
1377 | * A note about locking: | |
1378 | * Trying to take the page table spinlocks would be useless here because those | |
1379 | * are only used to synchronize: | |
1380 | * | |
1381 | * - modifying terminal entries (ones that point to a data page, not to another | |
1382 | * page table) | |
1383 | * - installing *new* non-terminal entries | |
1384 | * | |
1385 | * Instead, we need roughly the same kind of protection as free_pgtables() or | |
1386 | * mm_take_all_locks() (but only for a single VMA): | |
1387 | * The mmap lock together with this VMA's rmap locks covers all paths towards | |
1388 | * the page table entries we're messing with here, except for hardware page | |
1389 | * table walks and lockless_pages_from_mm(). | |
1390 | */ | |
e59a47b8 PT |
1391 | static void collapse_and_free_pmd(struct mm_struct *mm, struct vm_area_struct *vma, |
1392 | unsigned long addr, pmd_t *pmdp) | |
1393 | { | |
e59a47b8 | 1394 | pmd_t pmd; |
f268f6cf | 1395 | struct mmu_notifier_range range; |
e59a47b8 | 1396 | |
80110bbf | 1397 | mmap_assert_write_locked(mm); |
8d3c106e JH |
1398 | if (vma->vm_file) |
1399 | lockdep_assert_held_write(&vma->vm_file->f_mapping->i_mmap_rwsem); | |
1400 | /* | |
1401 | * All anon_vmas attached to the VMA have the same root and are | |
1402 | * therefore locked by the same lock. | |
1403 | */ | |
1404 | if (vma->anon_vma) | |
1405 | lockdep_assert_held_write(&vma->anon_vma->root->rwsem); | |
1406 | ||
f268f6cf JH |
1407 | mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm, addr, |
1408 | addr + HPAGE_PMD_SIZE); | |
1409 | mmu_notifier_invalidate_range_start(&range); | |
e59a47b8 | 1410 | pmd = pmdp_collapse_flush(vma, addr, pmdp); |
2ba99c5e | 1411 | tlb_remove_table_sync_one(); |
f268f6cf | 1412 | mmu_notifier_invalidate_range_end(&range); |
e59a47b8 | 1413 | mm_dec_nr_ptes(mm); |
80110bbf | 1414 | page_table_check_pte_clear_range(mm, addr, pmd); |
e59a47b8 PT |
1415 | pte_free(mm, pmd_pgtable(pmd)); |
1416 | } | |
1417 | ||
27e1f827 | 1418 | /** |
336e6b53 AS |
1419 | * collapse_pte_mapped_thp - Try to collapse a pte-mapped THP for mm at |
1420 | * address haddr. | |
1421 | * | |
1422 | * @mm: process address space where collapse happens | |
1423 | * @addr: THP collapse address | |
34488399 | 1424 | * @install_pmd: If a huge PMD should be installed |
27e1f827 SL |
1425 | * |
1426 | * This function checks whether all the PTEs in the PMD are pointing to the | |
1427 | * right THP. If so, retract the page table so the THP can refault in with | |
34488399 | 1428 | * as pmd-mapped. Possibly install a huge PMD mapping the THP. |
27e1f827 | 1429 | */ |
34488399 ZK |
1430 | int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, |
1431 | bool install_pmd) | |
27e1f827 SL |
1432 | { |
1433 | unsigned long haddr = addr & HPAGE_PMD_MASK; | |
94d815b2 | 1434 | struct vm_area_struct *vma = vma_lookup(mm, haddr); |
119a5fc1 | 1435 | struct page *hpage; |
27e1f827 | 1436 | pte_t *start_pte, *pte; |
e59a47b8 | 1437 | pmd_t *pmd; |
27e1f827 | 1438 | spinlock_t *ptl; |
58ac9a89 | 1439 | int count = 0, result = SCAN_FAIL; |
27e1f827 SL |
1440 | int i; |
1441 | ||
58ac9a89 ZK |
1442 | mmap_assert_write_locked(mm); |
1443 | ||
34488399 | 1444 | /* Fast check before locking page if already PMD-mapped */ |
58ac9a89 | 1445 | result = find_pmd_or_thp_or_none(mm, haddr, &pmd); |
34488399 ZK |
1446 | if (result == SCAN_PMD_MAPPED) |
1447 | return result; | |
58ac9a89 | 1448 | |
27e1f827 | 1449 | if (!vma || !vma->vm_file || |
fef792a4 | 1450 | !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE)) |
34488399 | 1451 | return SCAN_VMA_CHECK; |
27e1f827 SL |
1452 | |
1453 | /* | |
a7f4e6e4 ZK |
1454 | * If we are here, we've succeeded in replacing all the native pages |
1455 | * in the page cache with a single hugepage. If a mm were to fault-in | |
1456 | * this memory (mapped by a suitably aligned VMA), we'd get the hugepage | |
1457 | * and map it by a PMD, regardless of sysfs THP settings. As such, let's | |
1458 | * analogously elide sysfs THP settings here. | |
27e1f827 | 1459 | */ |
a7f4e6e4 | 1460 | if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) |
34488399 | 1461 | return SCAN_VMA_CHECK; |
27e1f827 | 1462 | |
deb4c93a PX |
1463 | /* Keep pmd pgtable for uffd-wp; see comment in retract_page_tables() */ |
1464 | if (userfaultfd_wp(vma)) | |
34488399 | 1465 | return SCAN_PTE_UFFD_WP; |
deb4c93a | 1466 | |
119a5fc1 HD |
1467 | hpage = find_lock_page(vma->vm_file->f_mapping, |
1468 | linear_page_index(vma, haddr)); | |
1469 | if (!hpage) | |
34488399 | 1470 | return SCAN_PAGE_NULL; |
119a5fc1 | 1471 | |
34488399 ZK |
1472 | if (!PageHead(hpage)) { |
1473 | result = SCAN_FAIL; | |
119a5fc1 | 1474 | goto drop_hpage; |
34488399 | 1475 | } |
119a5fc1 | 1476 | |
34488399 ZK |
1477 | if (compound_order(hpage) != HPAGE_PMD_ORDER) { |
1478 | result = SCAN_PAGE_COMPOUND; | |
119a5fc1 | 1479 | goto drop_hpage; |
34488399 | 1480 | } |
119a5fc1 | 1481 | |
34488399 ZK |
1482 | switch (result) { |
1483 | case SCAN_SUCCEED: | |
1484 | break; | |
1485 | case SCAN_PMD_NONE: | |
1486 | /* | |
1487 | * In MADV_COLLAPSE path, possible race with khugepaged where | |
1488 | * all pte entries have been removed and pmd cleared. If so, | |
1489 | * skip all the pte checks and just update the pmd mapping. | |
1490 | */ | |
1491 | goto maybe_install_pmd; | |
1492 | default: | |
119a5fc1 | 1493 | goto drop_hpage; |
34488399 | 1494 | } |
27e1f827 | 1495 | |
8d3c106e JH |
1496 | /* |
1497 | * We need to lock the mapping so that from here on, only GUP-fast and | |
1498 | * hardware page walks can access the parts of the page tables that | |
1499 | * we're operating on. | |
1500 | * See collapse_and_free_pmd(). | |
1501 | */ | |
1502 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
1503 | ||
1504 | /* | |
1505 | * This spinlock should be unnecessary: Nobody else should be accessing | |
1506 | * the page tables under spinlock protection here, only | |
1507 | * lockless_pages_from_mm() and the hardware page walker can access page | |
1508 | * tables while all the high-level locks are held in write mode. | |
1509 | */ | |
27e1f827 | 1510 | start_pte = pte_offset_map_lock(mm, pmd, haddr, &ptl); |
34488399 | 1511 | result = SCAN_FAIL; |
27e1f827 SL |
1512 | |
1513 | /* step 1: check all mapped PTEs are to the right huge page */ | |
1514 | for (i = 0, addr = haddr, pte = start_pte; | |
1515 | i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { | |
1516 | struct page *page; | |
1517 | ||
1518 | /* empty pte, skip */ | |
1519 | if (pte_none(*pte)) | |
1520 | continue; | |
1521 | ||
1522 | /* page swapped out, abort */ | |
34488399 ZK |
1523 | if (!pte_present(*pte)) { |
1524 | result = SCAN_PTE_NON_PRESENT; | |
27e1f827 | 1525 | goto abort; |
34488399 | 1526 | } |
27e1f827 SL |
1527 | |
1528 | page = vm_normal_page(vma, addr, *pte); | |
3218f871 AS |
1529 | if (WARN_ON_ONCE(page && is_zone_device_page(page))) |
1530 | page = NULL; | |
27e1f827 | 1531 | /* |
119a5fc1 HD |
1532 | * Note that uprobe, debugger, or MAP_PRIVATE may change the |
1533 | * page table, but the new page will not be a subpage of hpage. | |
27e1f827 | 1534 | */ |
119a5fc1 | 1535 | if (hpage + i != page) |
27e1f827 SL |
1536 | goto abort; |
1537 | count++; | |
1538 | } | |
1539 | ||
1540 | /* step 2: adjust rmap */ | |
1541 | for (i = 0, addr = haddr, pte = start_pte; | |
1542 | i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) { | |
1543 | struct page *page; | |
1544 | ||
1545 | if (pte_none(*pte)) | |
1546 | continue; | |
1547 | page = vm_normal_page(vma, addr, *pte); | |
3218f871 AS |
1548 | if (WARN_ON_ONCE(page && is_zone_device_page(page))) |
1549 | goto abort; | |
cea86fe2 | 1550 | page_remove_rmap(page, vma, false); |
27e1f827 SL |
1551 | } |
1552 | ||
1553 | pte_unmap_unlock(start_pte, ptl); | |
1554 | ||
1555 | /* step 3: set proper refcount and mm_counters. */ | |
119a5fc1 | 1556 | if (count) { |
27e1f827 SL |
1557 | page_ref_sub(hpage, count); |
1558 | add_mm_counter(vma->vm_mm, mm_counter_file(hpage), -count); | |
1559 | } | |
1560 | ||
34488399 | 1561 | /* step 4: remove pte entries */ |
ab0c3f12 HD |
1562 | /* we make no change to anon, but protect concurrent anon page lookup */ |
1563 | if (vma->anon_vma) | |
1564 | anon_vma_lock_write(vma->anon_vma); | |
1565 | ||
e59a47b8 | 1566 | collapse_and_free_pmd(mm, vma, haddr, pmd); |
34488399 | 1567 | |
ab0c3f12 HD |
1568 | if (vma->anon_vma) |
1569 | anon_vma_unlock_write(vma->anon_vma); | |
8d3c106e JH |
1570 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
1571 | ||
34488399 ZK |
1572 | maybe_install_pmd: |
1573 | /* step 5: install pmd entry */ | |
1574 | result = install_pmd | |
1575 | ? set_huge_pmd(vma, haddr, pmd, hpage) | |
1576 | : SCAN_SUCCEED; | |
1577 | ||
119a5fc1 HD |
1578 | drop_hpage: |
1579 | unlock_page(hpage); | |
1580 | put_page(hpage); | |
34488399 | 1581 | return result; |
27e1f827 SL |
1582 | |
1583 | abort: | |
1584 | pte_unmap_unlock(start_pte, ptl); | |
8d3c106e | 1585 | i_mmap_unlock_write(vma->vm_file->f_mapping); |
119a5fc1 | 1586 | goto drop_hpage; |
27e1f827 SL |
1587 | } |
1588 | ||
b26e2701 | 1589 | static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot) |
27e1f827 | 1590 | { |
b26e2701 QZ |
1591 | struct mm_slot *slot = &mm_slot->slot; |
1592 | struct mm_struct *mm = slot->mm; | |
27e1f827 SL |
1593 | int i; |
1594 | ||
1595 | if (likely(mm_slot->nr_pte_mapped_thp == 0)) | |
0edf61e5 | 1596 | return; |
27e1f827 | 1597 | |
d8ed45c5 | 1598 | if (!mmap_write_trylock(mm)) |
0edf61e5 | 1599 | return; |
27e1f827 | 1600 | |
7d2c4385 | 1601 | if (unlikely(hpage_collapse_test_exit(mm))) |
27e1f827 SL |
1602 | goto out; |
1603 | ||
1604 | for (i = 0; i < mm_slot->nr_pte_mapped_thp; i++) | |
34488399 | 1605 | collapse_pte_mapped_thp(mm, mm_slot->pte_mapped_thp[i], false); |
27e1f827 SL |
1606 | |
1607 | out: | |
1608 | mm_slot->nr_pte_mapped_thp = 0; | |
d8ed45c5 | 1609 | mmap_write_unlock(mm); |
27e1f827 SL |
1610 | } |
1611 | ||
34488399 ZK |
1612 | static int retract_page_tables(struct address_space *mapping, pgoff_t pgoff, |
1613 | struct mm_struct *target_mm, | |
1614 | unsigned long target_addr, struct page *hpage, | |
1615 | struct collapse_control *cc) | |
f3f0e1d2 KS |
1616 | { |
1617 | struct vm_area_struct *vma; | |
34488399 | 1618 | int target_result = SCAN_FAIL; |
f3f0e1d2 KS |
1619 | |
1620 | i_mmap_lock_write(mapping); | |
1621 | vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) { | |
34488399 ZK |
1622 | int result = SCAN_FAIL; |
1623 | struct mm_struct *mm = NULL; | |
1624 | unsigned long addr = 0; | |
1625 | pmd_t *pmd; | |
1626 | bool is_target = false; | |
1627 | ||
27e1f827 SL |
1628 | /* |
1629 | * Check vma->anon_vma to exclude MAP_PRIVATE mappings that | |
1630 | * got written to. These VMAs are likely not worth investing | |
3e4e28c5 | 1631 | * mmap_write_lock(mm) as PMD-mapping is likely to be split |
27e1f827 SL |
1632 | * later. |
1633 | * | |
36ee2c78 | 1634 | * Note that vma->anon_vma check is racy: it can be set up after |
c1e8d7c6 | 1635 | * the check but before we took mmap_lock by the fault path. |
27e1f827 SL |
1636 | * But page lock would prevent establishing any new ptes of the |
1637 | * page, so we are safe. | |
1638 | * | |
1639 | * An alternative would be drop the check, but check that page | |
1640 | * table is clear before calling pmdp_collapse_flush() under | |
1641 | * ptl. It has higher chance to recover THP for the VMA, but | |
8d3c106e JH |
1642 | * has higher cost too. It would also probably require locking |
1643 | * the anon_vma. | |
27e1f827 | 1644 | */ |
34488399 ZK |
1645 | if (vma->anon_vma) { |
1646 | result = SCAN_PAGE_ANON; | |
1647 | goto next; | |
1648 | } | |
f3f0e1d2 | 1649 | addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT); |
34488399 ZK |
1650 | if (addr & ~HPAGE_PMD_MASK || |
1651 | vma->vm_end < addr + HPAGE_PMD_SIZE) { | |
1652 | result = SCAN_VMA_CHECK; | |
1653 | goto next; | |
1654 | } | |
18e77600 | 1655 | mm = vma->vm_mm; |
34488399 ZK |
1656 | is_target = mm == target_mm && addr == target_addr; |
1657 | result = find_pmd_or_thp_or_none(mm, addr, &pmd); | |
1658 | if (result != SCAN_SUCCEED) | |
1659 | goto next; | |
f3f0e1d2 | 1660 | /* |
c1e8d7c6 | 1661 | * We need exclusive mmap_lock to retract page table. |
27e1f827 SL |
1662 | * |
1663 | * We use trylock due to lock inversion: we need to acquire | |
c1e8d7c6 | 1664 | * mmap_lock while holding page lock. Fault path does it in |
27e1f827 | 1665 | * reverse order. Trylock is a way to avoid deadlock. |
34488399 ZK |
1666 | * |
1667 | * Also, it's not MADV_COLLAPSE's job to collapse other | |
1668 | * mappings - let khugepaged take care of them later. | |
f3f0e1d2 | 1669 | */ |
34488399 ZK |
1670 | result = SCAN_PTE_MAPPED_HUGEPAGE; |
1671 | if ((cc->is_khugepaged || is_target) && | |
1672 | mmap_write_trylock(mm)) { | |
deb4c93a PX |
1673 | /* |
1674 | * When a vma is registered with uffd-wp, we can't | |
1675 | * recycle the pmd pgtable because there can be pte | |
1676 | * markers installed. Skip it only, so the rest mm/vma | |
1677 | * can still have the same file mapped hugely, however | |
1678 | * it'll always mapped in small page size for uffd-wp | |
1679 | * registered ranges. | |
1680 | */ | |
34488399 ZK |
1681 | if (hpage_collapse_test_exit(mm)) { |
1682 | result = SCAN_ANY_PROCESS; | |
1683 | goto unlock_next; | |
1684 | } | |
1685 | if (userfaultfd_wp(vma)) { | |
1686 | result = SCAN_PTE_UFFD_WP; | |
1687 | goto unlock_next; | |
1688 | } | |
1689 | collapse_and_free_pmd(mm, vma, addr, pmd); | |
1690 | if (!cc->is_khugepaged && is_target) | |
1691 | result = set_huge_pmd(vma, addr, pmd, hpage); | |
1692 | else | |
1693 | result = SCAN_SUCCEED; | |
1694 | ||
1695 | unlock_next: | |
18e77600 | 1696 | mmap_write_unlock(mm); |
34488399 ZK |
1697 | goto next; |
1698 | } | |
1699 | /* | |
1700 | * Calling context will handle target mm/addr. Otherwise, let | |
1701 | * khugepaged try again later. | |
1702 | */ | |
1703 | if (!is_target) { | |
18e77600 | 1704 | khugepaged_add_pte_mapped_thp(mm, addr); |
34488399 | 1705 | continue; |
f3f0e1d2 | 1706 | } |
34488399 ZK |
1707 | next: |
1708 | if (is_target) | |
1709 | target_result = result; | |
f3f0e1d2 KS |
1710 | } |
1711 | i_mmap_unlock_write(mapping); | |
34488399 | 1712 | return target_result; |
f3f0e1d2 KS |
1713 | } |
1714 | ||
1715 | /** | |
99cb0dbd | 1716 | * collapse_file - collapse filemap/tmpfs/shmem pages into huge one. |
f3f0e1d2 | 1717 | * |
336e6b53 | 1718 | * @mm: process address space where collapse happens |
34488399 | 1719 | * @addr: virtual collapse start address |
336e6b53 AS |
1720 | * @file: file that collapse on |
1721 | * @start: collapse start address | |
9710a78a | 1722 | * @cc: collapse context and scratchpad |
336e6b53 | 1723 | * |
f3f0e1d2 | 1724 | * Basic scheme is simple, details are more complex: |
87c460a0 | 1725 | * - allocate and lock a new huge page; |
77da9389 | 1726 | * - scan page cache replacing old pages with the new one |
99cb0dbd | 1727 | * + swap/gup in pages if necessary; |
f3f0e1d2 | 1728 | * + fill in gaps; |
77da9389 MW |
1729 | * + keep old pages around in case rollback is required; |
1730 | * - if replacing succeeds: | |
f3f0e1d2 KS |
1731 | * + copy data over; |
1732 | * + free old pages; | |
87c460a0 | 1733 | * + unlock huge page; |
f3f0e1d2 KS |
1734 | * - if replacing failed; |
1735 | * + put all pages back and unfreeze them; | |
77da9389 | 1736 | * + restore gaps in the page cache; |
87c460a0 | 1737 | * + unlock and free huge page; |
f3f0e1d2 | 1738 | */ |
34488399 ZK |
1739 | static int collapse_file(struct mm_struct *mm, unsigned long addr, |
1740 | struct file *file, pgoff_t start, | |
1741 | struct collapse_control *cc) | |
f3f0e1d2 | 1742 | { |
579c571e | 1743 | struct address_space *mapping = file->f_mapping; |
50ad2f24 | 1744 | struct page *hpage; |
4c9473e8 | 1745 | pgoff_t index = 0, end = start + HPAGE_PMD_NR; |
f3f0e1d2 | 1746 | LIST_HEAD(pagelist); |
77da9389 | 1747 | XA_STATE_ORDER(xas, &mapping->i_pages, start, HPAGE_PMD_ORDER); |
f3f0e1d2 | 1748 | int nr_none = 0, result = SCAN_SUCCEED; |
99cb0dbd | 1749 | bool is_shmem = shmem_file(file); |
4c9473e8 | 1750 | int nr = 0; |
f3f0e1d2 | 1751 | |
99cb0dbd | 1752 | VM_BUG_ON(!IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS) && !is_shmem); |
f3f0e1d2 KS |
1753 | VM_BUG_ON(start & (HPAGE_PMD_NR - 1)); |
1754 | ||
50ad2f24 | 1755 | result = alloc_charge_hpage(&hpage, mm, cc); |
9710a78a | 1756 | if (result != SCAN_SUCCEED) |
f3f0e1d2 | 1757 | goto out; |
f3f0e1d2 | 1758 | |
6b24ca4a MWO |
1759 | /* |
1760 | * Ensure we have slots for all the pages in the range. This is | |
1761 | * almost certainly a no-op because most of the pages must be present | |
1762 | */ | |
95feeabb HD |
1763 | do { |
1764 | xas_lock_irq(&xas); | |
1765 | xas_create_range(&xas); | |
1766 | if (!xas_error(&xas)) | |
1767 | break; | |
1768 | xas_unlock_irq(&xas); | |
1769 | if (!xas_nomem(&xas, GFP_KERNEL)) { | |
95feeabb HD |
1770 | result = SCAN_FAIL; |
1771 | goto out; | |
1772 | } | |
1773 | } while (1); | |
1774 | ||
50ad2f24 | 1775 | __SetPageLocked(hpage); |
99cb0dbd | 1776 | if (is_shmem) |
50ad2f24 ZK |
1777 | __SetPageSwapBacked(hpage); |
1778 | hpage->index = start; | |
1779 | hpage->mapping = mapping; | |
f3f0e1d2 | 1780 | |
f3f0e1d2 | 1781 | /* |
50ad2f24 | 1782 | * At this point the hpage is locked and not up-to-date. |
87c460a0 HD |
1783 | * It's safe to insert it into the page cache, because nobody would |
1784 | * be able to map it or use it in another way until we unlock it. | |
f3f0e1d2 KS |
1785 | */ |
1786 | ||
77da9389 MW |
1787 | xas_set(&xas, start); |
1788 | for (index = start; index < end; index++) { | |
1789 | struct page *page = xas_next(&xas); | |
64ab3195 | 1790 | struct folio *folio; |
77da9389 MW |
1791 | |
1792 | VM_BUG_ON(index != xas.xa_index); | |
99cb0dbd SL |
1793 | if (is_shmem) { |
1794 | if (!page) { | |
1795 | /* | |
1796 | * Stop if extent has been truncated or | |
1797 | * hole-punched, and is now completely | |
1798 | * empty. | |
1799 | */ | |
1800 | if (index == start) { | |
1801 | if (!xas_next_entry(&xas, end - 1)) { | |
1802 | result = SCAN_TRUNCATED; | |
1803 | goto xa_locked; | |
1804 | } | |
1805 | xas_set(&xas, index); | |
1806 | } | |
1807 | if (!shmem_charge(mapping->host, 1)) { | |
1808 | result = SCAN_FAIL; | |
042a3082 | 1809 | goto xa_locked; |
701270fa | 1810 | } |
50ad2f24 | 1811 | xas_store(&xas, hpage); |
99cb0dbd SL |
1812 | nr_none++; |
1813 | continue; | |
701270fa | 1814 | } |
99cb0dbd SL |
1815 | |
1816 | if (xa_is_value(page) || !PageUptodate(page)) { | |
1817 | xas_unlock_irq(&xas); | |
1818 | /* swap in or instantiate fallocated page */ | |
7459c149 MWO |
1819 | if (shmem_get_folio(mapping->host, index, |
1820 | &folio, SGP_NOALLOC)) { | |
99cb0dbd SL |
1821 | result = SCAN_FAIL; |
1822 | goto xa_unlocked; | |
1823 | } | |
7459c149 | 1824 | page = folio_file_page(folio, index); |
99cb0dbd SL |
1825 | } else if (trylock_page(page)) { |
1826 | get_page(page); | |
1827 | xas_unlock_irq(&xas); | |
1828 | } else { | |
1829 | result = SCAN_PAGE_LOCK; | |
042a3082 | 1830 | goto xa_locked; |
77da9389 | 1831 | } |
99cb0dbd SL |
1832 | } else { /* !is_shmem */ |
1833 | if (!page || xa_is_value(page)) { | |
1834 | xas_unlock_irq(&xas); | |
1835 | page_cache_sync_readahead(mapping, &file->f_ra, | |
1836 | file, index, | |
e5a59d30 | 1837 | end - index); |
99cb0dbd SL |
1838 | /* drain pagevecs to help isolate_lru_page() */ |
1839 | lru_add_drain(); | |
1840 | page = find_lock_page(mapping, index); | |
1841 | if (unlikely(page == NULL)) { | |
1842 | result = SCAN_FAIL; | |
1843 | goto xa_unlocked; | |
1844 | } | |
75f36069 SL |
1845 | } else if (PageDirty(page)) { |
1846 | /* | |
1847 | * khugepaged only works on read-only fd, | |
1848 | * so this page is dirty because it hasn't | |
1849 | * been flushed since first write. There | |
1850 | * won't be new dirty pages. | |
1851 | * | |
1852 | * Trigger async flush here and hope the | |
1853 | * writeback is done when khugepaged | |
1854 | * revisits this page. | |
1855 | * | |
1856 | * This is a one-off situation. We are not | |
1857 | * forcing writeback in loop. | |
1858 | */ | |
1859 | xas_unlock_irq(&xas); | |
1860 | filemap_flush(mapping); | |
1861 | result = SCAN_FAIL; | |
1862 | goto xa_unlocked; | |
74c42e1b RW |
1863 | } else if (PageWriteback(page)) { |
1864 | xas_unlock_irq(&xas); | |
1865 | result = SCAN_FAIL; | |
1866 | goto xa_unlocked; | |
99cb0dbd SL |
1867 | } else if (trylock_page(page)) { |
1868 | get_page(page); | |
1869 | xas_unlock_irq(&xas); | |
1870 | } else { | |
1871 | result = SCAN_PAGE_LOCK; | |
1872 | goto xa_locked; | |
f3f0e1d2 | 1873 | } |
f3f0e1d2 KS |
1874 | } |
1875 | ||
1876 | /* | |
b93b0163 | 1877 | * The page must be locked, so we can drop the i_pages lock |
f3f0e1d2 KS |
1878 | * without racing with truncate. |
1879 | */ | |
1880 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
4655e5e5 SL |
1881 | |
1882 | /* make sure the page is up to date */ | |
1883 | if (unlikely(!PageUptodate(page))) { | |
1884 | result = SCAN_FAIL; | |
1885 | goto out_unlock; | |
1886 | } | |
06a5e126 HD |
1887 | |
1888 | /* | |
1889 | * If file was truncated then extended, or hole-punched, before | |
1890 | * we locked the first page, then a THP might be there already. | |
58ac9a89 | 1891 | * This will be discovered on the first iteration. |
06a5e126 HD |
1892 | */ |
1893 | if (PageTransCompound(page)) { | |
58ac9a89 ZK |
1894 | struct page *head = compound_head(page); |
1895 | ||
1896 | result = compound_order(head) == HPAGE_PMD_ORDER && | |
1897 | head->index == start | |
1898 | /* Maybe PMD-mapped */ | |
1899 | ? SCAN_PTE_MAPPED_HUGEPAGE | |
1900 | : SCAN_PAGE_COMPOUND; | |
06a5e126 HD |
1901 | goto out_unlock; |
1902 | } | |
f3f0e1d2 | 1903 | |
64ab3195 VMO |
1904 | folio = page_folio(page); |
1905 | ||
1906 | if (folio_mapping(folio) != mapping) { | |
f3f0e1d2 KS |
1907 | result = SCAN_TRUNCATED; |
1908 | goto out_unlock; | |
1909 | } | |
f3f0e1d2 | 1910 | |
64ab3195 VMO |
1911 | if (!is_shmem && (folio_test_dirty(folio) || |
1912 | folio_test_writeback(folio))) { | |
4655e5e5 SL |
1913 | /* |
1914 | * khugepaged only works on read-only fd, so this | |
1915 | * page is dirty because it hasn't been flushed | |
1916 | * since first write. | |
1917 | */ | |
1918 | result = SCAN_FAIL; | |
1919 | goto out_unlock; | |
1920 | } | |
1921 | ||
64ab3195 | 1922 | if (folio_isolate_lru(folio)) { |
f3f0e1d2 | 1923 | result = SCAN_DEL_PAGE_LRU; |
042a3082 | 1924 | goto out_unlock; |
f3f0e1d2 KS |
1925 | } |
1926 | ||
64ab3195 VMO |
1927 | if (folio_has_private(folio) && |
1928 | !filemap_release_folio(folio, GFP_KERNEL)) { | |
99cb0dbd | 1929 | result = SCAN_PAGE_HAS_PRIVATE; |
64ab3195 | 1930 | folio_putback_lru(folio); |
99cb0dbd SL |
1931 | goto out_unlock; |
1932 | } | |
1933 | ||
64ab3195 VMO |
1934 | if (folio_mapped(folio)) |
1935 | try_to_unmap(folio, | |
869f7ee6 | 1936 | TTU_IGNORE_MLOCK | TTU_BATCH_FLUSH); |
f3f0e1d2 | 1937 | |
77da9389 MW |
1938 | xas_lock_irq(&xas); |
1939 | xas_set(&xas, index); | |
f3f0e1d2 | 1940 | |
77da9389 | 1941 | VM_BUG_ON_PAGE(page != xas_load(&xas), page); |
f3f0e1d2 KS |
1942 | |
1943 | /* | |
1944 | * The page is expected to have page_count() == 3: | |
1945 | * - we hold a pin on it; | |
77da9389 | 1946 | * - one reference from page cache; |
f3f0e1d2 KS |
1947 | * - one from isolate_lru_page; |
1948 | */ | |
1949 | if (!page_ref_freeze(page, 3)) { | |
1950 | result = SCAN_PAGE_COUNT; | |
042a3082 HD |
1951 | xas_unlock_irq(&xas); |
1952 | putback_lru_page(page); | |
1953 | goto out_unlock; | |
f3f0e1d2 KS |
1954 | } |
1955 | ||
1956 | /* | |
1957 | * Add the page to the list to be able to undo the collapse if | |
1958 | * something go wrong. | |
1959 | */ | |
1960 | list_add_tail(&page->lru, &pagelist); | |
1961 | ||
1962 | /* Finally, replace with the new page. */ | |
50ad2f24 | 1963 | xas_store(&xas, hpage); |
f3f0e1d2 | 1964 | continue; |
f3f0e1d2 KS |
1965 | out_unlock: |
1966 | unlock_page(page); | |
1967 | put_page(page); | |
042a3082 | 1968 | goto xa_unlocked; |
f3f0e1d2 | 1969 | } |
50ad2f24 | 1970 | nr = thp_nr_pages(hpage); |
f3f0e1d2 | 1971 | |
99cb0dbd | 1972 | if (is_shmem) |
50ad2f24 | 1973 | __mod_lruvec_page_state(hpage, NR_SHMEM_THPS, nr); |
09d91cda | 1974 | else { |
50ad2f24 | 1975 | __mod_lruvec_page_state(hpage, NR_FILE_THPS, nr); |
09d91cda | 1976 | filemap_nr_thps_inc(mapping); |
eb6ecbed CF |
1977 | /* |
1978 | * Paired with smp_mb() in do_dentry_open() to ensure | |
1979 | * i_writecount is up to date and the update to nr_thps is | |
1980 | * visible. Ensures the page cache will be truncated if the | |
1981 | * file is opened writable. | |
1982 | */ | |
1983 | smp_mb(); | |
1984 | if (inode_is_open_for_write(mapping->host)) { | |
1985 | result = SCAN_FAIL; | |
50ad2f24 | 1986 | __mod_lruvec_page_state(hpage, NR_FILE_THPS, -nr); |
eb6ecbed CF |
1987 | filemap_nr_thps_dec(mapping); |
1988 | goto xa_locked; | |
1989 | } | |
09d91cda | 1990 | } |
99cb0dbd | 1991 | |
042a3082 | 1992 | if (nr_none) { |
50ad2f24 | 1993 | __mod_lruvec_page_state(hpage, NR_FILE_PAGES, nr_none); |
2f55f070 | 1994 | /* nr_none is always 0 for non-shmem. */ |
50ad2f24 | 1995 | __mod_lruvec_page_state(hpage, NR_SHMEM, nr_none); |
042a3082 HD |
1996 | } |
1997 | ||
6b24ca4a MWO |
1998 | /* Join all the small entries into a single multi-index entry */ |
1999 | xas_set_order(&xas, start, HPAGE_PMD_ORDER); | |
50ad2f24 | 2000 | xas_store(&xas, hpage); |
042a3082 HD |
2001 | xa_locked: |
2002 | xas_unlock_irq(&xas); | |
77da9389 | 2003 | xa_unlocked: |
042a3082 | 2004 | |
6d9df8a5 HD |
2005 | /* |
2006 | * If collapse is successful, flush must be done now before copying. | |
2007 | * If collapse is unsuccessful, does flush actually need to be done? | |
2008 | * Do it anyway, to clear the state. | |
2009 | */ | |
2010 | try_to_unmap_flush(); | |
2011 | ||
f3f0e1d2 | 2012 | if (result == SCAN_SUCCEED) { |
77da9389 | 2013 | struct page *page, *tmp; |
284a344e | 2014 | struct folio *folio; |
f3f0e1d2 KS |
2015 | |
2016 | /* | |
77da9389 MW |
2017 | * Replacing old pages with new one has succeeded, now we |
2018 | * need to copy the content and free the old pages. | |
f3f0e1d2 | 2019 | */ |
2af8ff29 | 2020 | index = start; |
f3f0e1d2 | 2021 | list_for_each_entry_safe(page, tmp, &pagelist, lru) { |
2af8ff29 | 2022 | while (index < page->index) { |
50ad2f24 | 2023 | clear_highpage(hpage + (index % HPAGE_PMD_NR)); |
2af8ff29 HD |
2024 | index++; |
2025 | } | |
50ad2f24 ZK |
2026 | copy_highpage(hpage + (page->index % HPAGE_PMD_NR), |
2027 | page); | |
f3f0e1d2 | 2028 | list_del(&page->lru); |
f3f0e1d2 | 2029 | page->mapping = NULL; |
042a3082 | 2030 | page_ref_unfreeze(page, 1); |
f3f0e1d2 KS |
2031 | ClearPageActive(page); |
2032 | ClearPageUnevictable(page); | |
042a3082 | 2033 | unlock_page(page); |
f3f0e1d2 | 2034 | put_page(page); |
2af8ff29 HD |
2035 | index++; |
2036 | } | |
2037 | while (index < end) { | |
50ad2f24 | 2038 | clear_highpage(hpage + (index % HPAGE_PMD_NR)); |
2af8ff29 | 2039 | index++; |
f3f0e1d2 KS |
2040 | } |
2041 | ||
284a344e VMO |
2042 | folio = page_folio(hpage); |
2043 | folio_mark_uptodate(folio); | |
2044 | folio_ref_add(folio, HPAGE_PMD_NR - 1); | |
2045 | ||
6058eaec | 2046 | if (is_shmem) |
284a344e VMO |
2047 | folio_mark_dirty(folio); |
2048 | folio_add_lru(folio); | |
f3f0e1d2 | 2049 | |
042a3082 HD |
2050 | /* |
2051 | * Remove pte page tables, so we can re-fault the page as huge. | |
2052 | */ | |
34488399 ZK |
2053 | result = retract_page_tables(mapping, start, mm, addr, hpage, |
2054 | cc); | |
50ad2f24 ZK |
2055 | unlock_page(hpage); |
2056 | hpage = NULL; | |
f3f0e1d2 | 2057 | } else { |
77da9389 | 2058 | struct page *page; |
aaa52e34 | 2059 | |
77da9389 | 2060 | /* Something went wrong: roll back page cache changes */ |
77da9389 | 2061 | xas_lock_irq(&xas); |
2f55f070 ML |
2062 | if (nr_none) { |
2063 | mapping->nrpages -= nr_none; | |
99cb0dbd | 2064 | shmem_uncharge(mapping->host, nr_none); |
2f55f070 | 2065 | } |
aaa52e34 | 2066 | |
77da9389 MW |
2067 | xas_set(&xas, start); |
2068 | xas_for_each(&xas, page, end - 1) { | |
f3f0e1d2 KS |
2069 | page = list_first_entry_or_null(&pagelist, |
2070 | struct page, lru); | |
77da9389 | 2071 | if (!page || xas.xa_index < page->index) { |
f3f0e1d2 KS |
2072 | if (!nr_none) |
2073 | break; | |
f3f0e1d2 | 2074 | nr_none--; |
59749e6c | 2075 | /* Put holes back where they were */ |
77da9389 | 2076 | xas_store(&xas, NULL); |
f3f0e1d2 KS |
2077 | continue; |
2078 | } | |
2079 | ||
77da9389 | 2080 | VM_BUG_ON_PAGE(page->index != xas.xa_index, page); |
f3f0e1d2 KS |
2081 | |
2082 | /* Unfreeze the page. */ | |
2083 | list_del(&page->lru); | |
2084 | page_ref_unfreeze(page, 2); | |
77da9389 MW |
2085 | xas_store(&xas, page); |
2086 | xas_pause(&xas); | |
2087 | xas_unlock_irq(&xas); | |
f3f0e1d2 | 2088 | unlock_page(page); |
042a3082 | 2089 | putback_lru_page(page); |
77da9389 | 2090 | xas_lock_irq(&xas); |
f3f0e1d2 KS |
2091 | } |
2092 | VM_BUG_ON(nr_none); | |
77da9389 | 2093 | xas_unlock_irq(&xas); |
f3f0e1d2 | 2094 | |
50ad2f24 | 2095 | hpage->mapping = NULL; |
f3f0e1d2 | 2096 | } |
042a3082 | 2097 | |
50ad2f24 ZK |
2098 | if (hpage) |
2099 | unlock_page(hpage); | |
f3f0e1d2 KS |
2100 | out: |
2101 | VM_BUG_ON(!list_empty(&pagelist)); | |
50ad2f24 ZK |
2102 | if (hpage) { |
2103 | mem_cgroup_uncharge(page_folio(hpage)); | |
2104 | put_page(hpage); | |
c6a7f445 | 2105 | } |
4c9473e8 GM |
2106 | |
2107 | trace_mm_khugepaged_collapse_file(mm, hpage, index, is_shmem, addr, file, nr, result); | |
50ad2f24 | 2108 | return result; |
f3f0e1d2 KS |
2109 | } |
2110 | ||
34488399 ZK |
2111 | static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, |
2112 | struct file *file, pgoff_t start, | |
2113 | struct collapse_control *cc) | |
f3f0e1d2 KS |
2114 | { |
2115 | struct page *page = NULL; | |
579c571e | 2116 | struct address_space *mapping = file->f_mapping; |
85b392db | 2117 | XA_STATE(xas, &mapping->i_pages, start); |
f3f0e1d2 KS |
2118 | int present, swap; |
2119 | int node = NUMA_NO_NODE; | |
2120 | int result = SCAN_SUCCEED; | |
2121 | ||
2122 | present = 0; | |
2123 | swap = 0; | |
34d6b470 | 2124 | memset(cc->node_load, 0, sizeof(cc->node_load)); |
e031ff96 | 2125 | nodes_clear(cc->alloc_nmask); |
f3f0e1d2 | 2126 | rcu_read_lock(); |
85b392db MW |
2127 | xas_for_each(&xas, page, start + HPAGE_PMD_NR - 1) { |
2128 | if (xas_retry(&xas, page)) | |
f3f0e1d2 | 2129 | continue; |
f3f0e1d2 | 2130 | |
85b392db | 2131 | if (xa_is_value(page)) { |
d8ea7cc8 ZK |
2132 | ++swap; |
2133 | if (cc->is_khugepaged && | |
2134 | swap > khugepaged_max_ptes_swap) { | |
f3f0e1d2 | 2135 | result = SCAN_EXCEED_SWAP_PTE; |
e9ea874a | 2136 | count_vm_event(THP_SCAN_EXCEED_SWAP_PTE); |
f3f0e1d2 KS |
2137 | break; |
2138 | } | |
2139 | continue; | |
2140 | } | |
2141 | ||
6b24ca4a | 2142 | /* |
58ac9a89 | 2143 | * TODO: khugepaged should compact smaller compound pages |
6b24ca4a MWO |
2144 | * into a PMD sized page |
2145 | */ | |
f3f0e1d2 | 2146 | if (PageTransCompound(page)) { |
58ac9a89 ZK |
2147 | struct page *head = compound_head(page); |
2148 | ||
2149 | result = compound_order(head) == HPAGE_PMD_ORDER && | |
2150 | head->index == start | |
2151 | /* Maybe PMD-mapped */ | |
2152 | ? SCAN_PTE_MAPPED_HUGEPAGE | |
2153 | : SCAN_PAGE_COMPOUND; | |
2154 | /* | |
2155 | * For SCAN_PTE_MAPPED_HUGEPAGE, further processing | |
2156 | * by the caller won't touch the page cache, and so | |
2157 | * it's safe to skip LRU and refcount checks before | |
2158 | * returning. | |
2159 | */ | |
f3f0e1d2 KS |
2160 | break; |
2161 | } | |
2162 | ||
2163 | node = page_to_nid(page); | |
7d2c4385 | 2164 | if (hpage_collapse_scan_abort(node, cc)) { |
f3f0e1d2 KS |
2165 | result = SCAN_SCAN_ABORT; |
2166 | break; | |
2167 | } | |
34d6b470 | 2168 | cc->node_load[node]++; |
f3f0e1d2 KS |
2169 | |
2170 | if (!PageLRU(page)) { | |
2171 | result = SCAN_PAGE_LRU; | |
2172 | break; | |
2173 | } | |
2174 | ||
99cb0dbd SL |
2175 | if (page_count(page) != |
2176 | 1 + page_mapcount(page) + page_has_private(page)) { | |
f3f0e1d2 KS |
2177 | result = SCAN_PAGE_COUNT; |
2178 | break; | |
2179 | } | |
2180 | ||
2181 | /* | |
2182 | * We probably should check if the page is referenced here, but | |
2183 | * nobody would transfer pte_young() to PageReferenced() for us. | |
2184 | * And rmap walk here is just too costly... | |
2185 | */ | |
2186 | ||
2187 | present++; | |
2188 | ||
2189 | if (need_resched()) { | |
85b392db | 2190 | xas_pause(&xas); |
f3f0e1d2 | 2191 | cond_resched_rcu(); |
f3f0e1d2 KS |
2192 | } |
2193 | } | |
2194 | rcu_read_unlock(); | |
2195 | ||
2196 | if (result == SCAN_SUCCEED) { | |
d8ea7cc8 ZK |
2197 | if (cc->is_khugepaged && |
2198 | present < HPAGE_PMD_NR - khugepaged_max_ptes_none) { | |
f3f0e1d2 | 2199 | result = SCAN_EXCEED_NONE_PTE; |
e9ea874a | 2200 | count_vm_event(THP_SCAN_EXCEED_NONE_PTE); |
f3f0e1d2 | 2201 | } else { |
34488399 | 2202 | result = collapse_file(mm, addr, file, start, cc); |
f3f0e1d2 KS |
2203 | } |
2204 | } | |
2205 | ||
045634ff | 2206 | trace_mm_khugepaged_scan_file(mm, page, file, present, swap, result); |
50ad2f24 | 2207 | return result; |
f3f0e1d2 KS |
2208 | } |
2209 | #else | |
34488399 ZK |
2210 | static int hpage_collapse_scan_file(struct mm_struct *mm, unsigned long addr, |
2211 | struct file *file, pgoff_t start, | |
2212 | struct collapse_control *cc) | |
f3f0e1d2 KS |
2213 | { |
2214 | BUILD_BUG(); | |
2215 | } | |
27e1f827 | 2216 | |
b26e2701 | 2217 | static void khugepaged_collapse_pte_mapped_thps(struct khugepaged_mm_slot *mm_slot) |
27e1f827 | 2218 | { |
27e1f827 | 2219 | } |
58ac9a89 ZK |
2220 | |
2221 | static bool khugepaged_add_pte_mapped_thp(struct mm_struct *mm, | |
2222 | unsigned long addr) | |
2223 | { | |
2224 | return false; | |
2225 | } | |
f3f0e1d2 KS |
2226 | #endif |
2227 | ||
50ad2f24 | 2228 | static unsigned int khugepaged_scan_mm_slot(unsigned int pages, int *result, |
34d6b470 | 2229 | struct collapse_control *cc) |
b46e756f KS |
2230 | __releases(&khugepaged_mm_lock) |
2231 | __acquires(&khugepaged_mm_lock) | |
2232 | { | |
68540502 | 2233 | struct vma_iterator vmi; |
b26e2701 QZ |
2234 | struct khugepaged_mm_slot *mm_slot; |
2235 | struct mm_slot *slot; | |
b46e756f KS |
2236 | struct mm_struct *mm; |
2237 | struct vm_area_struct *vma; | |
2238 | int progress = 0; | |
2239 | ||
2240 | VM_BUG_ON(!pages); | |
35f3aa39 | 2241 | lockdep_assert_held(&khugepaged_mm_lock); |
50ad2f24 | 2242 | *result = SCAN_FAIL; |
b46e756f | 2243 | |
b26e2701 | 2244 | if (khugepaged_scan.mm_slot) { |
b46e756f | 2245 | mm_slot = khugepaged_scan.mm_slot; |
b26e2701 QZ |
2246 | slot = &mm_slot->slot; |
2247 | } else { | |
2248 | slot = list_entry(khugepaged_scan.mm_head.next, | |
b46e756f | 2249 | struct mm_slot, mm_node); |
b26e2701 | 2250 | mm_slot = mm_slot_entry(slot, struct khugepaged_mm_slot, slot); |
b46e756f KS |
2251 | khugepaged_scan.address = 0; |
2252 | khugepaged_scan.mm_slot = mm_slot; | |
2253 | } | |
2254 | spin_unlock(&khugepaged_mm_lock); | |
27e1f827 | 2255 | khugepaged_collapse_pte_mapped_thps(mm_slot); |
b46e756f | 2256 | |
b26e2701 | 2257 | mm = slot->mm; |
3b454ad3 YS |
2258 | /* |
2259 | * Don't wait for semaphore (to avoid long wait times). Just move to | |
2260 | * the next mm on the list. | |
2261 | */ | |
2262 | vma = NULL; | |
d8ed45c5 | 2263 | if (unlikely(!mmap_read_trylock(mm))) |
c1e8d7c6 | 2264 | goto breakouterloop_mmap_lock; |
b46e756f KS |
2265 | |
2266 | progress++; | |
68540502 MWO |
2267 | if (unlikely(hpage_collapse_test_exit(mm))) |
2268 | goto breakouterloop; | |
2269 | ||
2270 | vma_iter_init(&vmi, mm, khugepaged_scan.address); | |
2271 | for_each_vma(vmi, vma) { | |
b46e756f KS |
2272 | unsigned long hstart, hend; |
2273 | ||
2274 | cond_resched(); | |
7d2c4385 | 2275 | if (unlikely(hpage_collapse_test_exit(mm))) { |
b46e756f KS |
2276 | progress++; |
2277 | break; | |
2278 | } | |
a7f4e6e4 | 2279 | if (!hugepage_vma_check(vma, vma->vm_flags, false, false, true)) { |
b46e756f KS |
2280 | skip: |
2281 | progress++; | |
2282 | continue; | |
2283 | } | |
4fa6893f YS |
2284 | hstart = round_up(vma->vm_start, HPAGE_PMD_SIZE); |
2285 | hend = round_down(vma->vm_end, HPAGE_PMD_SIZE); | |
b46e756f KS |
2286 | if (khugepaged_scan.address > hend) |
2287 | goto skip; | |
2288 | if (khugepaged_scan.address < hstart) | |
2289 | khugepaged_scan.address = hstart; | |
2290 | VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK); | |
2291 | ||
2292 | while (khugepaged_scan.address < hend) { | |
50ad2f24 ZK |
2293 | bool mmap_locked = true; |
2294 | ||
b46e756f | 2295 | cond_resched(); |
7d2c4385 | 2296 | if (unlikely(hpage_collapse_test_exit(mm))) |
b46e756f KS |
2297 | goto breakouterloop; |
2298 | ||
2299 | VM_BUG_ON(khugepaged_scan.address < hstart || | |
2300 | khugepaged_scan.address + HPAGE_PMD_SIZE > | |
2301 | hend); | |
99cb0dbd | 2302 | if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { |
396bcc52 | 2303 | struct file *file = get_file(vma->vm_file); |
f3f0e1d2 KS |
2304 | pgoff_t pgoff = linear_page_index(vma, |
2305 | khugepaged_scan.address); | |
99cb0dbd | 2306 | |
d8ed45c5 | 2307 | mmap_read_unlock(mm); |
34488399 ZK |
2308 | *result = hpage_collapse_scan_file(mm, |
2309 | khugepaged_scan.address, | |
2310 | file, pgoff, cc); | |
50ad2f24 | 2311 | mmap_locked = false; |
f3f0e1d2 KS |
2312 | fput(file); |
2313 | } else { | |
7d2c4385 ZK |
2314 | *result = hpage_collapse_scan_pmd(mm, vma, |
2315 | khugepaged_scan.address, | |
2316 | &mmap_locked, | |
2317 | cc); | |
f3f0e1d2 | 2318 | } |
58ac9a89 ZK |
2319 | switch (*result) { |
2320 | case SCAN_PTE_MAPPED_HUGEPAGE: { | |
2321 | pmd_t *pmd; | |
2322 | ||
2323 | *result = find_pmd_or_thp_or_none(mm, | |
2324 | khugepaged_scan.address, | |
2325 | &pmd); | |
2326 | if (*result != SCAN_SUCCEED) | |
2327 | break; | |
2328 | if (!khugepaged_add_pte_mapped_thp(mm, | |
2329 | khugepaged_scan.address)) | |
2330 | break; | |
2331 | } fallthrough; | |
2332 | case SCAN_SUCCEED: | |
50ad2f24 | 2333 | ++khugepaged_pages_collapsed; |
58ac9a89 ZK |
2334 | break; |
2335 | default: | |
2336 | break; | |
f3f0e1d2 | 2337 | } |
58ac9a89 | 2338 | |
b46e756f KS |
2339 | /* move to next address */ |
2340 | khugepaged_scan.address += HPAGE_PMD_SIZE; | |
2341 | progress += HPAGE_PMD_NR; | |
50ad2f24 ZK |
2342 | if (!mmap_locked) |
2343 | /* | |
2344 | * We released mmap_lock so break loop. Note | |
2345 | * that we drop mmap_lock before all hugepage | |
2346 | * allocations, so if allocation fails, we are | |
2347 | * guaranteed to break here and report the | |
2348 | * correct result back to caller. | |
2349 | */ | |
c1e8d7c6 | 2350 | goto breakouterloop_mmap_lock; |
b46e756f KS |
2351 | if (progress >= pages) |
2352 | goto breakouterloop; | |
2353 | } | |
2354 | } | |
2355 | breakouterloop: | |
d8ed45c5 | 2356 | mmap_read_unlock(mm); /* exit_mmap will destroy ptes after this */ |
c1e8d7c6 | 2357 | breakouterloop_mmap_lock: |
b46e756f KS |
2358 | |
2359 | spin_lock(&khugepaged_mm_lock); | |
2360 | VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot); | |
2361 | /* | |
2362 | * Release the current mm_slot if this mm is about to die, or | |
2363 | * if we scanned all vmas of this mm. | |
2364 | */ | |
7d2c4385 | 2365 | if (hpage_collapse_test_exit(mm) || !vma) { |
b46e756f KS |
2366 | /* |
2367 | * Make sure that if mm_users is reaching zero while | |
2368 | * khugepaged runs here, khugepaged_exit will find | |
2369 | * mm_slot not pointing to the exiting mm. | |
2370 | */ | |
b26e2701 QZ |
2371 | if (slot->mm_node.next != &khugepaged_scan.mm_head) { |
2372 | slot = list_entry(slot->mm_node.next, | |
2373 | struct mm_slot, mm_node); | |
2374 | khugepaged_scan.mm_slot = | |
2375 | mm_slot_entry(slot, struct khugepaged_mm_slot, slot); | |
b46e756f KS |
2376 | khugepaged_scan.address = 0; |
2377 | } else { | |
2378 | khugepaged_scan.mm_slot = NULL; | |
2379 | khugepaged_full_scans++; | |
2380 | } | |
2381 | ||
2382 | collect_mm_slot(mm_slot); | |
2383 | } | |
2384 | ||
2385 | return progress; | |
2386 | } | |
2387 | ||
2388 | static int khugepaged_has_work(void) | |
2389 | { | |
2390 | return !list_empty(&khugepaged_scan.mm_head) && | |
1064026b | 2391 | hugepage_flags_enabled(); |
b46e756f KS |
2392 | } |
2393 | ||
2394 | static int khugepaged_wait_event(void) | |
2395 | { | |
2396 | return !list_empty(&khugepaged_scan.mm_head) || | |
2397 | kthread_should_stop(); | |
2398 | } | |
2399 | ||
34d6b470 | 2400 | static void khugepaged_do_scan(struct collapse_control *cc) |
b46e756f | 2401 | { |
b46e756f | 2402 | unsigned int progress = 0, pass_through_head = 0; |
89dc6a96 | 2403 | unsigned int pages = READ_ONCE(khugepaged_pages_to_scan); |
b46e756f | 2404 | bool wait = true; |
50ad2f24 | 2405 | int result = SCAN_SUCCEED; |
b46e756f | 2406 | |
a980df33 KS |
2407 | lru_add_drain_all(); |
2408 | ||
c6a7f445 | 2409 | while (true) { |
b46e756f KS |
2410 | cond_resched(); |
2411 | ||
2412 | if (unlikely(kthread_should_stop() || try_to_freeze())) | |
2413 | break; | |
2414 | ||
2415 | spin_lock(&khugepaged_mm_lock); | |
2416 | if (!khugepaged_scan.mm_slot) | |
2417 | pass_through_head++; | |
2418 | if (khugepaged_has_work() && | |
2419 | pass_through_head < 2) | |
2420 | progress += khugepaged_scan_mm_slot(pages - progress, | |
50ad2f24 | 2421 | &result, cc); |
b46e756f KS |
2422 | else |
2423 | progress = pages; | |
2424 | spin_unlock(&khugepaged_mm_lock); | |
b46e756f | 2425 | |
c6a7f445 YS |
2426 | if (progress >= pages) |
2427 | break; | |
2428 | ||
50ad2f24 | 2429 | if (result == SCAN_ALLOC_HUGE_PAGE_FAIL) { |
c6a7f445 YS |
2430 | /* |
2431 | * If fail to allocate the first time, try to sleep for | |
2432 | * a while. When hit again, cancel the scan. | |
2433 | */ | |
2434 | if (!wait) | |
2435 | break; | |
2436 | wait = false; | |
c6a7f445 YS |
2437 | khugepaged_alloc_sleep(); |
2438 | } | |
2439 | } | |
b46e756f KS |
2440 | } |
2441 | ||
2442 | static bool khugepaged_should_wakeup(void) | |
2443 | { | |
2444 | return kthread_should_stop() || | |
2445 | time_after_eq(jiffies, khugepaged_sleep_expire); | |
2446 | } | |
2447 | ||
2448 | static void khugepaged_wait_work(void) | |
2449 | { | |
2450 | if (khugepaged_has_work()) { | |
2451 | const unsigned long scan_sleep_jiffies = | |
2452 | msecs_to_jiffies(khugepaged_scan_sleep_millisecs); | |
2453 | ||
2454 | if (!scan_sleep_jiffies) | |
2455 | return; | |
2456 | ||
2457 | khugepaged_sleep_expire = jiffies + scan_sleep_jiffies; | |
2458 | wait_event_freezable_timeout(khugepaged_wait, | |
2459 | khugepaged_should_wakeup(), | |
2460 | scan_sleep_jiffies); | |
2461 | return; | |
2462 | } | |
2463 | ||
1064026b | 2464 | if (hugepage_flags_enabled()) |
b46e756f KS |
2465 | wait_event_freezable(khugepaged_wait, khugepaged_wait_event()); |
2466 | } | |
2467 | ||
2468 | static int khugepaged(void *none) | |
2469 | { | |
b26e2701 | 2470 | struct khugepaged_mm_slot *mm_slot; |
b46e756f KS |
2471 | |
2472 | set_freezable(); | |
2473 | set_user_nice(current, MAX_NICE); | |
2474 | ||
2475 | while (!kthread_should_stop()) { | |
34d6b470 | 2476 | khugepaged_do_scan(&khugepaged_collapse_control); |
b46e756f KS |
2477 | khugepaged_wait_work(); |
2478 | } | |
2479 | ||
2480 | spin_lock(&khugepaged_mm_lock); | |
2481 | mm_slot = khugepaged_scan.mm_slot; | |
2482 | khugepaged_scan.mm_slot = NULL; | |
2483 | if (mm_slot) | |
2484 | collect_mm_slot(mm_slot); | |
2485 | spin_unlock(&khugepaged_mm_lock); | |
2486 | return 0; | |
2487 | } | |
2488 | ||
2489 | static void set_recommended_min_free_kbytes(void) | |
2490 | { | |
2491 | struct zone *zone; | |
2492 | int nr_zones = 0; | |
2493 | unsigned long recommended_min; | |
2494 | ||
1064026b | 2495 | if (!hugepage_flags_enabled()) { |
bd3400ea LF |
2496 | calculate_min_free_kbytes(); |
2497 | goto update_wmarks; | |
2498 | } | |
2499 | ||
b7d349c7 JK |
2500 | for_each_populated_zone(zone) { |
2501 | /* | |
2502 | * We don't need to worry about fragmentation of | |
2503 | * ZONE_MOVABLE since it only has movable pages. | |
2504 | */ | |
2505 | if (zone_idx(zone) > gfp_zone(GFP_USER)) | |
2506 | continue; | |
2507 | ||
b46e756f | 2508 | nr_zones++; |
b7d349c7 | 2509 | } |
b46e756f KS |
2510 | |
2511 | /* Ensure 2 pageblocks are free to assist fragmentation avoidance */ | |
2512 | recommended_min = pageblock_nr_pages * nr_zones * 2; | |
2513 | ||
2514 | /* | |
2515 | * Make sure that on average at least two pageblocks are almost free | |
2516 | * of another type, one for a migratetype to fall back to and a | |
2517 | * second to avoid subsequent fallbacks of other types There are 3 | |
2518 | * MIGRATE_TYPES we care about. | |
2519 | */ | |
2520 | recommended_min += pageblock_nr_pages * nr_zones * | |
2521 | MIGRATE_PCPTYPES * MIGRATE_PCPTYPES; | |
2522 | ||
2523 | /* don't ever allow to reserve more than 5% of the lowmem */ | |
2524 | recommended_min = min(recommended_min, | |
2525 | (unsigned long) nr_free_buffer_pages() / 20); | |
2526 | recommended_min <<= (PAGE_SHIFT-10); | |
2527 | ||
2528 | if (recommended_min > min_free_kbytes) { | |
2529 | if (user_min_free_kbytes >= 0) | |
2530 | pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n", | |
2531 | min_free_kbytes, recommended_min); | |
2532 | ||
2533 | min_free_kbytes = recommended_min; | |
2534 | } | |
bd3400ea LF |
2535 | |
2536 | update_wmarks: | |
b46e756f KS |
2537 | setup_per_zone_wmarks(); |
2538 | } | |
2539 | ||
2540 | int start_stop_khugepaged(void) | |
2541 | { | |
b46e756f KS |
2542 | int err = 0; |
2543 | ||
2544 | mutex_lock(&khugepaged_mutex); | |
1064026b | 2545 | if (hugepage_flags_enabled()) { |
b46e756f KS |
2546 | if (!khugepaged_thread) |
2547 | khugepaged_thread = kthread_run(khugepaged, NULL, | |
2548 | "khugepaged"); | |
2549 | if (IS_ERR(khugepaged_thread)) { | |
2550 | pr_err("khugepaged: kthread_run(khugepaged) failed\n"); | |
2551 | err = PTR_ERR(khugepaged_thread); | |
2552 | khugepaged_thread = NULL; | |
2553 | goto fail; | |
2554 | } | |
2555 | ||
2556 | if (!list_empty(&khugepaged_scan.mm_head)) | |
2557 | wake_up_interruptible(&khugepaged_wait); | |
b46e756f KS |
2558 | } else if (khugepaged_thread) { |
2559 | kthread_stop(khugepaged_thread); | |
2560 | khugepaged_thread = NULL; | |
2561 | } | |
bd3400ea | 2562 | set_recommended_min_free_kbytes(); |
b46e756f KS |
2563 | fail: |
2564 | mutex_unlock(&khugepaged_mutex); | |
2565 | return err; | |
2566 | } | |
4aab2be0 VB |
2567 | |
2568 | void khugepaged_min_free_kbytes_update(void) | |
2569 | { | |
2570 | mutex_lock(&khugepaged_mutex); | |
1064026b | 2571 | if (hugepage_flags_enabled() && khugepaged_thread) |
4aab2be0 VB |
2572 | set_recommended_min_free_kbytes(); |
2573 | mutex_unlock(&khugepaged_mutex); | |
2574 | } | |
7d8faaf1 | 2575 | |
57e9cc50 JW |
2576 | bool current_is_khugepaged(void) |
2577 | { | |
2578 | return kthread_func(current) == khugepaged; | |
2579 | } | |
2580 | ||
7d8faaf1 ZK |
2581 | static int madvise_collapse_errno(enum scan_result r) |
2582 | { | |
2583 | /* | |
2584 | * MADV_COLLAPSE breaks from existing madvise(2) conventions to provide | |
2585 | * actionable feedback to caller, so they may take an appropriate | |
2586 | * fallback measure depending on the nature of the failure. | |
2587 | */ | |
2588 | switch (r) { | |
2589 | case SCAN_ALLOC_HUGE_PAGE_FAIL: | |
2590 | return -ENOMEM; | |
2591 | case SCAN_CGROUP_CHARGE_FAIL: | |
2592 | return -EBUSY; | |
2593 | /* Resource temporary unavailable - trying again might succeed */ | |
2594 | case SCAN_PAGE_LOCK: | |
2595 | case SCAN_PAGE_LRU: | |
0f3e2a2c | 2596 | case SCAN_DEL_PAGE_LRU: |
7d8faaf1 ZK |
2597 | return -EAGAIN; |
2598 | /* | |
2599 | * Other: Trying again likely not to succeed / error intrinsic to | |
2600 | * specified memory range. khugepaged likely won't be able to collapse | |
2601 | * either. | |
2602 | */ | |
2603 | default: | |
2604 | return -EINVAL; | |
2605 | } | |
2606 | } | |
2607 | ||
2608 | int madvise_collapse(struct vm_area_struct *vma, struct vm_area_struct **prev, | |
2609 | unsigned long start, unsigned long end) | |
2610 | { | |
2611 | struct collapse_control *cc; | |
2612 | struct mm_struct *mm = vma->vm_mm; | |
2613 | unsigned long hstart, hend, addr; | |
2614 | int thps = 0, last_fail = SCAN_FAIL; | |
2615 | bool mmap_locked = true; | |
2616 | ||
2617 | BUG_ON(vma->vm_start > start); | |
2618 | BUG_ON(vma->vm_end < end); | |
2619 | ||
2620 | *prev = vma; | |
2621 | ||
7d8faaf1 ZK |
2622 | if (!hugepage_vma_check(vma, vma->vm_flags, false, false, false)) |
2623 | return -EINVAL; | |
2624 | ||
2625 | cc = kmalloc(sizeof(*cc), GFP_KERNEL); | |
2626 | if (!cc) | |
2627 | return -ENOMEM; | |
2628 | cc->is_khugepaged = false; | |
7d8faaf1 ZK |
2629 | |
2630 | mmgrab(mm); | |
2631 | lru_add_drain_all(); | |
2632 | ||
2633 | hstart = (start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK; | |
2634 | hend = end & HPAGE_PMD_MASK; | |
2635 | ||
2636 | for (addr = hstart; addr < hend; addr += HPAGE_PMD_SIZE) { | |
2637 | int result = SCAN_FAIL; | |
2638 | ||
2639 | if (!mmap_locked) { | |
2640 | cond_resched(); | |
2641 | mmap_read_lock(mm); | |
2642 | mmap_locked = true; | |
34488399 ZK |
2643 | result = hugepage_vma_revalidate(mm, addr, false, &vma, |
2644 | cc); | |
7d8faaf1 ZK |
2645 | if (result != SCAN_SUCCEED) { |
2646 | last_fail = result; | |
2647 | goto out_nolock; | |
2648 | } | |
4d24de94 | 2649 | |
52dc0310 | 2650 | hend = min(hend, vma->vm_end & HPAGE_PMD_MASK); |
7d8faaf1 ZK |
2651 | } |
2652 | mmap_assert_locked(mm); | |
2653 | memset(cc->node_load, 0, sizeof(cc->node_load)); | |
e031ff96 | 2654 | nodes_clear(cc->alloc_nmask); |
34488399 ZK |
2655 | if (IS_ENABLED(CONFIG_SHMEM) && vma->vm_file) { |
2656 | struct file *file = get_file(vma->vm_file); | |
2657 | pgoff_t pgoff = linear_page_index(vma, addr); | |
2658 | ||
2659 | mmap_read_unlock(mm); | |
2660 | mmap_locked = false; | |
2661 | result = hpage_collapse_scan_file(mm, addr, file, pgoff, | |
2662 | cc); | |
2663 | fput(file); | |
2664 | } else { | |
2665 | result = hpage_collapse_scan_pmd(mm, vma, addr, | |
2666 | &mmap_locked, cc); | |
2667 | } | |
7d8faaf1 ZK |
2668 | if (!mmap_locked) |
2669 | *prev = NULL; /* Tell caller we dropped mmap_lock */ | |
2670 | ||
34488399 | 2671 | handle_result: |
7d8faaf1 ZK |
2672 | switch (result) { |
2673 | case SCAN_SUCCEED: | |
2674 | case SCAN_PMD_MAPPED: | |
2675 | ++thps; | |
2676 | break; | |
34488399 ZK |
2677 | case SCAN_PTE_MAPPED_HUGEPAGE: |
2678 | BUG_ON(mmap_locked); | |
2679 | BUG_ON(*prev); | |
2680 | mmap_write_lock(mm); | |
2681 | result = collapse_pte_mapped_thp(mm, addr, true); | |
2682 | mmap_write_unlock(mm); | |
2683 | goto handle_result; | |
7d8faaf1 ZK |
2684 | /* Whitelisted set of results where continuing OK */ |
2685 | case SCAN_PMD_NULL: | |
2686 | case SCAN_PTE_NON_PRESENT: | |
2687 | case SCAN_PTE_UFFD_WP: | |
2688 | case SCAN_PAGE_RO: | |
2689 | case SCAN_LACK_REFERENCED_PAGE: | |
2690 | case SCAN_PAGE_NULL: | |
2691 | case SCAN_PAGE_COUNT: | |
2692 | case SCAN_PAGE_LOCK: | |
2693 | case SCAN_PAGE_COMPOUND: | |
2694 | case SCAN_PAGE_LRU: | |
0f3e2a2c | 2695 | case SCAN_DEL_PAGE_LRU: |
7d8faaf1 ZK |
2696 | last_fail = result; |
2697 | break; | |
2698 | default: | |
2699 | last_fail = result; | |
2700 | /* Other error, exit */ | |
2701 | goto out_maybelock; | |
2702 | } | |
2703 | } | |
2704 | ||
2705 | out_maybelock: | |
2706 | /* Caller expects us to hold mmap_lock on return */ | |
2707 | if (!mmap_locked) | |
2708 | mmap_read_lock(mm); | |
2709 | out_nolock: | |
2710 | mmap_assert_locked(mm); | |
2711 | mmdrop(mm); | |
2712 | kfree(cc); | |
2713 | ||
2714 | return thps == ((hend - hstart) >> HPAGE_PMD_SHIFT) ? 0 | |
2715 | : madvise_collapse_errno(last_fail); | |
2716 | } |