]>
Commit | Line | Data |
---|---|---|
09c434b8 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
9a001fc1 VW |
2 | /* |
3 | * z3fold.c | |
4 | * | |
5 | * Author: Vitaly Wool <vitaly.wool@konsulko.com> | |
6 | * Copyright (C) 2016, Sony Mobile Communications Inc. | |
7 | * | |
8 | * This implementation is based on zbud written by Seth Jennings. | |
9 | * | |
10 | * z3fold is an special purpose allocator for storing compressed pages. It | |
11 | * can store up to three compressed pages per page which improves the | |
12 | * compression ratio of zbud while retaining its main concepts (e. g. always | |
13 | * storing an integral number of objects per page) and simplicity. | |
14 | * It still has simple and deterministic reclaim properties that make it | |
15 | * preferable to a higher density approach (with no requirement on integral | |
16 | * number of object per page) when reclaim is used. | |
17 | * | |
18 | * As in zbud, pages are divided into "chunks". The size of the chunks is | |
19 | * fixed at compile time and is determined by NCHUNKS_ORDER below. | |
20 | * | |
21 | * z3fold doesn't export any API and is meant to be used via zpool API. | |
22 | */ | |
23 | ||
24 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
25 | ||
26 | #include <linux/atomic.h> | |
d30561c5 | 27 | #include <linux/sched.h> |
1f862989 | 28 | #include <linux/cpumask.h> |
9a001fc1 VW |
29 | #include <linux/list.h> |
30 | #include <linux/mm.h> | |
31 | #include <linux/module.h> | |
1f862989 VW |
32 | #include <linux/page-flags.h> |
33 | #include <linux/migrate.h> | |
34 | #include <linux/node.h> | |
35 | #include <linux/compaction.h> | |
d30561c5 | 36 | #include <linux/percpu.h> |
1f862989 | 37 | #include <linux/mount.h> |
ea8157ab | 38 | #include <linux/pseudo_fs.h> |
1f862989 | 39 | #include <linux/fs.h> |
9a001fc1 | 40 | #include <linux/preempt.h> |
d30561c5 | 41 | #include <linux/workqueue.h> |
9a001fc1 VW |
42 | #include <linux/slab.h> |
43 | #include <linux/spinlock.h> | |
44 | #include <linux/zpool.h> | |
ea8157ab | 45 | #include <linux/magic.h> |
af4798a5 | 46 | #include <linux/kmemleak.h> |
9a001fc1 | 47 | |
7c2b8baa VW |
48 | /* |
49 | * NCHUNKS_ORDER determines the internal allocation granularity, effectively | |
50 | * adjusting internal fragmentation. It also determines the number of | |
51 | * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the | |
52 | * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks | |
53 | * in the beginning of an allocated page are occupied by z3fold header, so | |
54 | * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y), | |
55 | * which shows the max number of free chunks in z3fold page, also there will | |
56 | * be 63, or 62, respectively, freelists per pool. | |
57 | */ | |
58 | #define NCHUNKS_ORDER 6 | |
59 | ||
60 | #define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER) | |
61 | #define CHUNK_SIZE (1 << CHUNK_SHIFT) | |
62 | #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE) | |
63 | #define ZHDR_CHUNKS (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT) | |
64 | #define TOTAL_CHUNKS (PAGE_SIZE >> CHUNK_SHIFT) | |
e3c0db4f | 65 | #define NCHUNKS (TOTAL_CHUNKS - ZHDR_CHUNKS) |
7c2b8baa VW |
66 | |
67 | #define BUDDY_MASK (0x3) | |
68 | #define BUDDY_SHIFT 2 | |
69 | #define SLOTS_ALIGN (0x40) | |
70 | ||
9a001fc1 VW |
71 | /***************** |
72 | * Structures | |
73 | *****************/ | |
ede93213 VW |
74 | struct z3fold_pool; |
75 | struct z3fold_ops { | |
76 | int (*evict)(struct z3fold_pool *pool, unsigned long handle); | |
77 | }; | |
78 | ||
79 | enum buddy { | |
80 | HEADLESS = 0, | |
81 | FIRST, | |
82 | MIDDLE, | |
83 | LAST, | |
7c2b8baa VW |
84 | BUDDIES_MAX = LAST |
85 | }; | |
86 | ||
87 | struct z3fold_buddy_slots { | |
88 | /* | |
89 | * we are using BUDDY_MASK in handle_to_buddy etc. so there should | |
90 | * be enough slots to hold all possible variants | |
91 | */ | |
92 | unsigned long slot[BUDDY_MASK + 1]; | |
fc548865 | 93 | unsigned long pool; /* back link */ |
4a3ac931 | 94 | rwlock_t lock; |
ede93213 | 95 | }; |
7c2b8baa | 96 | #define HANDLE_FLAG_MASK (0x03) |
ede93213 VW |
97 | |
98 | /* | |
d30561c5 | 99 | * struct z3fold_header - z3fold page metadata occupying first chunks of each |
ede93213 | 100 | * z3fold page, except for HEADLESS pages |
d30561c5 VW |
101 | * @buddy: links the z3fold page into the relevant list in the |
102 | * pool | |
2f1e5e4d | 103 | * @page_lock: per-page lock |
d30561c5 VW |
104 | * @refcount: reference count for the z3fold page |
105 | * @work: work_struct for page layout optimization | |
7c2b8baa | 106 | * @slots: pointer to the structure holding buddy slots |
bb9a374d | 107 | * @pool: pointer to the containing pool |
d30561c5 | 108 | * @cpu: CPU which this page "belongs" to |
ede93213 VW |
109 | * @first_chunks: the size of the first buddy in chunks, 0 if free |
110 | * @middle_chunks: the size of the middle buddy in chunks, 0 if free | |
111 | * @last_chunks: the size of the last buddy in chunks, 0 if free | |
112 | * @first_num: the starting number (for the first handle) | |
1f862989 | 113 | * @mapped_count: the number of objects currently mapped |
ede93213 VW |
114 | */ |
115 | struct z3fold_header { | |
116 | struct list_head buddy; | |
2f1e5e4d | 117 | spinlock_t page_lock; |
5a27aa82 | 118 | struct kref refcount; |
d30561c5 | 119 | struct work_struct work; |
7c2b8baa | 120 | struct z3fold_buddy_slots *slots; |
bb9a374d | 121 | struct z3fold_pool *pool; |
d30561c5 | 122 | short cpu; |
ede93213 VW |
123 | unsigned short first_chunks; |
124 | unsigned short middle_chunks; | |
125 | unsigned short last_chunks; | |
126 | unsigned short start_middle; | |
127 | unsigned short first_num:2; | |
1f862989 | 128 | unsigned short mapped_count:2; |
4a3ac931 | 129 | unsigned short foreign_handles:2; |
ede93213 VW |
130 | }; |
131 | ||
9a001fc1 VW |
132 | /** |
133 | * struct z3fold_pool - stores metadata for each z3fold pool | |
d30561c5 VW |
134 | * @name: pool name |
135 | * @lock: protects pool unbuddied/lru lists | |
136 | * @stale_lock: protects pool stale page list | |
137 | * @unbuddied: per-cpu array of lists tracking z3fold pages that contain 2- | |
138 | * buddies; the list each z3fold page is added to depends on | |
139 | * the size of its free region. | |
9a001fc1 VW |
140 | * @lru: list tracking the z3fold pages in LRU order by most recently |
141 | * added buddy. | |
d30561c5 | 142 | * @stale: list of pages marked for freeing |
9a001fc1 | 143 | * @pages_nr: number of z3fold pages in the pool. |
7c2b8baa | 144 | * @c_handle: cache for z3fold_buddy_slots allocation |
9a001fc1 VW |
145 | * @ops: pointer to a structure of user defined operations specified at |
146 | * pool creation time. | |
d30561c5 VW |
147 | * @compact_wq: workqueue for page layout background optimization |
148 | * @release_wq: workqueue for safe page release | |
149 | * @work: work_struct for safe page release | |
1f862989 | 150 | * @inode: inode for z3fold pseudo filesystem |
9a001fc1 VW |
151 | * |
152 | * This structure is allocated at pool creation time and maintains metadata | |
153 | * pertaining to a particular z3fold pool. | |
154 | */ | |
155 | struct z3fold_pool { | |
d30561c5 | 156 | const char *name; |
9a001fc1 | 157 | spinlock_t lock; |
d30561c5 VW |
158 | spinlock_t stale_lock; |
159 | struct list_head *unbuddied; | |
9a001fc1 | 160 | struct list_head lru; |
d30561c5 | 161 | struct list_head stale; |
12d59ae6 | 162 | atomic64_t pages_nr; |
7c2b8baa | 163 | struct kmem_cache *c_handle; |
9a001fc1 VW |
164 | const struct z3fold_ops *ops; |
165 | struct zpool *zpool; | |
166 | const struct zpool_ops *zpool_ops; | |
d30561c5 VW |
167 | struct workqueue_struct *compact_wq; |
168 | struct workqueue_struct *release_wq; | |
169 | struct work_struct work; | |
1f862989 | 170 | struct inode *inode; |
9a001fc1 VW |
171 | }; |
172 | ||
9a001fc1 VW |
173 | /* |
174 | * Internal z3fold page flags | |
175 | */ | |
176 | enum z3fold_page_flags { | |
5a27aa82 | 177 | PAGE_HEADLESS = 0, |
9a001fc1 | 178 | MIDDLE_CHUNK_MAPPED, |
d30561c5 | 179 | NEEDS_COMPACTING, |
6098d7e1 | 180 | PAGE_STALE, |
ca0246bb | 181 | PAGE_CLAIMED, /* by either reclaim or free */ |
9a001fc1 VW |
182 | }; |
183 | ||
dcf5aedb VW |
184 | /* |
185 | * handle flags, go under HANDLE_FLAG_MASK | |
186 | */ | |
187 | enum z3fold_handle_flags { | |
188 | HANDLES_NOFREE = 0, | |
189 | }; | |
190 | ||
4a3ac931 VW |
191 | /* |
192 | * Forward declarations | |
193 | */ | |
194 | static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool); | |
195 | static void compact_page_work(struct work_struct *w); | |
196 | ||
9a001fc1 VW |
197 | /***************** |
198 | * Helpers | |
199 | *****************/ | |
200 | ||
201 | /* Converts an allocation size in bytes to size in z3fold chunks */ | |
202 | static int size_to_chunks(size_t size) | |
203 | { | |
204 | return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT; | |
205 | } | |
206 | ||
207 | #define for_each_unbuddied_list(_iter, _begin) \ | |
208 | for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++) | |
209 | ||
bb9f6f63 VW |
210 | static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool, |
211 | gfp_t gfp) | |
7c2b8baa | 212 | { |
f1549cb5 HB |
213 | struct z3fold_buddy_slots *slots; |
214 | ||
f94afee9 | 215 | slots = kmem_cache_zalloc(pool->c_handle, |
f1549cb5 | 216 | (gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE))); |
7c2b8baa VW |
217 | |
218 | if (slots) { | |
af4798a5 QC |
219 | /* It will be freed separately in free_handle(). */ |
220 | kmemleak_not_leak(slots); | |
7c2b8baa | 221 | slots->pool = (unsigned long)pool; |
4a3ac931 | 222 | rwlock_init(&slots->lock); |
7c2b8baa VW |
223 | } |
224 | ||
225 | return slots; | |
226 | } | |
227 | ||
228 | static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s) | |
229 | { | |
230 | return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK); | |
231 | } | |
232 | ||
233 | static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle) | |
234 | { | |
235 | return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1)); | |
236 | } | |
237 | ||
4a3ac931 VW |
238 | /* Lock a z3fold page */ |
239 | static inline void z3fold_page_lock(struct z3fold_header *zhdr) | |
240 | { | |
241 | spin_lock(&zhdr->page_lock); | |
242 | } | |
243 | ||
244 | /* Try to lock a z3fold page */ | |
245 | static inline int z3fold_page_trylock(struct z3fold_header *zhdr) | |
246 | { | |
247 | return spin_trylock(&zhdr->page_lock); | |
248 | } | |
249 | ||
250 | /* Unlock a z3fold page */ | |
251 | static inline void z3fold_page_unlock(struct z3fold_header *zhdr) | |
252 | { | |
253 | spin_unlock(&zhdr->page_lock); | |
254 | } | |
255 | ||
767cc6c5 ML |
256 | /* return locked z3fold page if it's not headless */ |
257 | static inline struct z3fold_header *get_z3fold_header(unsigned long handle) | |
4a3ac931 VW |
258 | { |
259 | struct z3fold_buddy_slots *slots; | |
260 | struct z3fold_header *zhdr; | |
261 | int locked = 0; | |
262 | ||
263 | if (!(handle & (1 << PAGE_HEADLESS))) { | |
264 | slots = handle_to_slots(handle); | |
265 | do { | |
266 | unsigned long addr; | |
267 | ||
268 | read_lock(&slots->lock); | |
269 | addr = *(unsigned long *)handle; | |
270 | zhdr = (struct z3fold_header *)(addr & PAGE_MASK); | |
767cc6c5 | 271 | locked = z3fold_page_trylock(zhdr); |
4a3ac931 VW |
272 | read_unlock(&slots->lock); |
273 | if (locked) | |
274 | break; | |
275 | cpu_relax(); | |
767cc6c5 | 276 | } while (true); |
4a3ac931 VW |
277 | } else { |
278 | zhdr = (struct z3fold_header *)(handle & PAGE_MASK); | |
279 | } | |
280 | ||
281 | return zhdr; | |
282 | } | |
283 | ||
4a3ac931 VW |
284 | static inline void put_z3fold_header(struct z3fold_header *zhdr) |
285 | { | |
286 | struct page *page = virt_to_page(zhdr); | |
287 | ||
288 | if (!test_bit(PAGE_HEADLESS, &page->private)) | |
289 | z3fold_page_unlock(zhdr); | |
290 | } | |
291 | ||
fc548865 | 292 | static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr) |
7c2b8baa VW |
293 | { |
294 | struct z3fold_buddy_slots *slots; | |
295 | int i; | |
296 | bool is_free; | |
297 | ||
298 | if (handle & (1 << PAGE_HEADLESS)) | |
299 | return; | |
300 | ||
4a3ac931 VW |
301 | if (WARN_ON(*(unsigned long *)handle == 0)) |
302 | return; | |
303 | ||
7c2b8baa | 304 | slots = handle_to_slots(handle); |
4a3ac931 VW |
305 | write_lock(&slots->lock); |
306 | *(unsigned long *)handle = 0; | |
dcf5aedb VW |
307 | |
308 | if (test_bit(HANDLES_NOFREE, &slots->pool)) { | |
309 | write_unlock(&slots->lock); | |
310 | return; /* simple case, nothing else to do */ | |
311 | } | |
312 | ||
fc548865 VW |
313 | if (zhdr->slots != slots) |
314 | zhdr->foreign_handles--; | |
4a3ac931 | 315 | |
7c2b8baa VW |
316 | is_free = true; |
317 | for (i = 0; i <= BUDDY_MASK; i++) { | |
318 | if (slots->slot[i]) { | |
319 | is_free = false; | |
320 | break; | |
321 | } | |
322 | } | |
d8f117ab | 323 | write_unlock(&slots->lock); |
7c2b8baa VW |
324 | |
325 | if (is_free) { | |
326 | struct z3fold_pool *pool = slots_to_pool(slots); | |
327 | ||
fc548865 VW |
328 | if (zhdr->slots == slots) |
329 | zhdr->slots = NULL; | |
7c2b8baa VW |
330 | kmem_cache_free(pool->c_handle, slots); |
331 | } | |
332 | } | |
333 | ||
ea8157ab | 334 | static int z3fold_init_fs_context(struct fs_context *fc) |
1f862989 | 335 | { |
ea8157ab | 336 | return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM; |
1f862989 VW |
337 | } |
338 | ||
339 | static struct file_system_type z3fold_fs = { | |
340 | .name = "z3fold", | |
ea8157ab | 341 | .init_fs_context = z3fold_init_fs_context, |
1f862989 VW |
342 | .kill_sb = kill_anon_super, |
343 | }; | |
344 | ||
345 | static struct vfsmount *z3fold_mnt; | |
346 | static int z3fold_mount(void) | |
347 | { | |
348 | int ret = 0; | |
349 | ||
350 | z3fold_mnt = kern_mount(&z3fold_fs); | |
351 | if (IS_ERR(z3fold_mnt)) | |
352 | ret = PTR_ERR(z3fold_mnt); | |
353 | ||
354 | return ret; | |
355 | } | |
356 | ||
357 | static void z3fold_unmount(void) | |
358 | { | |
359 | kern_unmount(z3fold_mnt); | |
360 | } | |
361 | ||
362 | static const struct address_space_operations z3fold_aops; | |
363 | static int z3fold_register_migration(struct z3fold_pool *pool) | |
364 | { | |
365 | pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb); | |
366 | if (IS_ERR(pool->inode)) { | |
367 | pool->inode = NULL; | |
368 | return 1; | |
369 | } | |
370 | ||
371 | pool->inode->i_mapping->private_data = pool; | |
372 | pool->inode->i_mapping->a_ops = &z3fold_aops; | |
373 | return 0; | |
374 | } | |
375 | ||
376 | static void z3fold_unregister_migration(struct z3fold_pool *pool) | |
377 | { | |
378 | if (pool->inode) | |
379 | iput(pool->inode); | |
cb152a1a | 380 | } |
1f862989 | 381 | |
9a001fc1 | 382 | /* Initializes the z3fold header of a newly allocated z3fold page */ |
63398413 | 383 | static struct z3fold_header *init_z3fold_page(struct page *page, bool headless, |
bb9f6f63 | 384 | struct z3fold_pool *pool, gfp_t gfp) |
9a001fc1 VW |
385 | { |
386 | struct z3fold_header *zhdr = page_address(page); | |
63398413 | 387 | struct z3fold_buddy_slots *slots; |
9a001fc1 VW |
388 | |
389 | INIT_LIST_HEAD(&page->lru); | |
9a001fc1 VW |
390 | clear_bit(PAGE_HEADLESS, &page->private); |
391 | clear_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
d30561c5 VW |
392 | clear_bit(NEEDS_COMPACTING, &page->private); |
393 | clear_bit(PAGE_STALE, &page->private); | |
ca0246bb | 394 | clear_bit(PAGE_CLAIMED, &page->private); |
63398413 VW |
395 | if (headless) |
396 | return zhdr; | |
397 | ||
398 | slots = alloc_slots(pool, gfp); | |
399 | if (!slots) | |
400 | return NULL; | |
9a001fc1 | 401 | |
c457cd96 | 402 | memset(zhdr, 0, sizeof(*zhdr)); |
2f1e5e4d | 403 | spin_lock_init(&zhdr->page_lock); |
5a27aa82 | 404 | kref_init(&zhdr->refcount); |
d30561c5 | 405 | zhdr->cpu = -1; |
7c2b8baa | 406 | zhdr->slots = slots; |
bb9a374d | 407 | zhdr->pool = pool; |
9a001fc1 | 408 | INIT_LIST_HEAD(&zhdr->buddy); |
d30561c5 | 409 | INIT_WORK(&zhdr->work, compact_page_work); |
9a001fc1 VW |
410 | return zhdr; |
411 | } | |
412 | ||
413 | /* Resets the struct page fields and frees the page */ | |
1f862989 | 414 | static void free_z3fold_page(struct page *page, bool headless) |
9a001fc1 | 415 | { |
1f862989 VW |
416 | if (!headless) { |
417 | lock_page(page); | |
418 | __ClearPageMovable(page); | |
419 | unlock_page(page); | |
420 | } | |
421 | ClearPagePrivate(page); | |
5a27aa82 VW |
422 | __free_page(page); |
423 | } | |
424 | ||
7c2b8baa VW |
425 | /* Helper function to build the index */ |
426 | static inline int __idx(struct z3fold_header *zhdr, enum buddy bud) | |
427 | { | |
428 | return (bud + zhdr->first_num) & BUDDY_MASK; | |
429 | } | |
430 | ||
9a001fc1 VW |
431 | /* |
432 | * Encodes the handle of a particular buddy within a z3fold page | |
433 | * Pool lock should be held as this function accesses first_num | |
434 | */ | |
3f9d2b57 VW |
435 | static unsigned long __encode_handle(struct z3fold_header *zhdr, |
436 | struct z3fold_buddy_slots *slots, | |
437 | enum buddy bud) | |
9a001fc1 | 438 | { |
7c2b8baa VW |
439 | unsigned long h = (unsigned long)zhdr; |
440 | int idx = 0; | |
9a001fc1 | 441 | |
7c2b8baa VW |
442 | /* |
443 | * For a headless page, its handle is its pointer with the extra | |
444 | * PAGE_HEADLESS bit set | |
445 | */ | |
446 | if (bud == HEADLESS) | |
447 | return h | (1 << PAGE_HEADLESS); | |
448 | ||
449 | /* otherwise, return pointer to encoded handle */ | |
450 | idx = __idx(zhdr, bud); | |
451 | h += idx; | |
452 | if (bud == LAST) | |
453 | h |= (zhdr->last_chunks << BUDDY_SHIFT); | |
454 | ||
4a3ac931 | 455 | write_lock(&slots->lock); |
7c2b8baa | 456 | slots->slot[idx] = h; |
4a3ac931 | 457 | write_unlock(&slots->lock); |
7c2b8baa | 458 | return (unsigned long)&slots->slot[idx]; |
9a001fc1 VW |
459 | } |
460 | ||
3f9d2b57 VW |
461 | static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud) |
462 | { | |
463 | return __encode_handle(zhdr, zhdr->slots, bud); | |
464 | } | |
465 | ||
ca0246bb VW |
466 | /* only for LAST bud, returns zero otherwise */ |
467 | static unsigned short handle_to_chunks(unsigned long handle) | |
468 | { | |
4a3ac931 VW |
469 | struct z3fold_buddy_slots *slots = handle_to_slots(handle); |
470 | unsigned long addr; | |
7c2b8baa | 471 | |
4a3ac931 VW |
472 | read_lock(&slots->lock); |
473 | addr = *(unsigned long *)handle; | |
474 | read_unlock(&slots->lock); | |
7c2b8baa | 475 | return (addr & ~PAGE_MASK) >> BUDDY_SHIFT; |
ca0246bb VW |
476 | } |
477 | ||
f201ebd8 | 478 | /* |
479 | * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle | |
480 | * but that doesn't matter. because the masking will result in the | |
481 | * correct buddy number. | |
482 | */ | |
9a001fc1 VW |
483 | static enum buddy handle_to_buddy(unsigned long handle) |
484 | { | |
7c2b8baa | 485 | struct z3fold_header *zhdr; |
4a3ac931 | 486 | struct z3fold_buddy_slots *slots = handle_to_slots(handle); |
7c2b8baa VW |
487 | unsigned long addr; |
488 | ||
4a3ac931 | 489 | read_lock(&slots->lock); |
7c2b8baa VW |
490 | WARN_ON(handle & (1 << PAGE_HEADLESS)); |
491 | addr = *(unsigned long *)handle; | |
4a3ac931 | 492 | read_unlock(&slots->lock); |
7c2b8baa VW |
493 | zhdr = (struct z3fold_header *)(addr & PAGE_MASK); |
494 | return (addr - zhdr->first_num) & BUDDY_MASK; | |
9a001fc1 VW |
495 | } |
496 | ||
9050cce1 VW |
497 | static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr) |
498 | { | |
bb9a374d | 499 | return zhdr->pool; |
9050cce1 VW |
500 | } |
501 | ||
d30561c5 VW |
502 | static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked) |
503 | { | |
504 | struct page *page = virt_to_page(zhdr); | |
9050cce1 | 505 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
d30561c5 VW |
506 | |
507 | WARN_ON(!list_empty(&zhdr->buddy)); | |
508 | set_bit(PAGE_STALE, &page->private); | |
35529357 | 509 | clear_bit(NEEDS_COMPACTING, &page->private); |
d30561c5 VW |
510 | spin_lock(&pool->lock); |
511 | if (!list_empty(&page->lru)) | |
1f862989 | 512 | list_del_init(&page->lru); |
d30561c5 | 513 | spin_unlock(&pool->lock); |
4a3ac931 | 514 | |
d30561c5 VW |
515 | if (locked) |
516 | z3fold_page_unlock(zhdr); | |
4a3ac931 | 517 | |
d30561c5 VW |
518 | spin_lock(&pool->stale_lock); |
519 | list_add(&zhdr->buddy, &pool->stale); | |
520 | queue_work(pool->release_wq, &pool->work); | |
521 | spin_unlock(&pool->stale_lock); | |
522 | } | |
523 | ||
70ad3196 | 524 | static void release_z3fold_page(struct kref *ref) |
d30561c5 VW |
525 | { |
526 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
527 | refcount); | |
528 | __release_z3fold_page(zhdr, false); | |
529 | } | |
530 | ||
531 | static void release_z3fold_page_locked(struct kref *ref) | |
532 | { | |
533 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
534 | refcount); | |
535 | WARN_ON(z3fold_page_trylock(zhdr)); | |
536 | __release_z3fold_page(zhdr, true); | |
537 | } | |
538 | ||
539 | static void release_z3fold_page_locked_list(struct kref *ref) | |
540 | { | |
541 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
542 | refcount); | |
9050cce1 | 543 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
4a3ac931 | 544 | |
9050cce1 | 545 | spin_lock(&pool->lock); |
d30561c5 | 546 | list_del_init(&zhdr->buddy); |
9050cce1 | 547 | spin_unlock(&pool->lock); |
d30561c5 VW |
548 | |
549 | WARN_ON(z3fold_page_trylock(zhdr)); | |
550 | __release_z3fold_page(zhdr, true); | |
551 | } | |
552 | ||
553 | static void free_pages_work(struct work_struct *w) | |
554 | { | |
555 | struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work); | |
556 | ||
557 | spin_lock(&pool->stale_lock); | |
558 | while (!list_empty(&pool->stale)) { | |
559 | struct z3fold_header *zhdr = list_first_entry(&pool->stale, | |
560 | struct z3fold_header, buddy); | |
561 | struct page *page = virt_to_page(zhdr); | |
562 | ||
563 | list_del(&zhdr->buddy); | |
564 | if (WARN_ON(!test_bit(PAGE_STALE, &page->private))) | |
565 | continue; | |
d30561c5 VW |
566 | spin_unlock(&pool->stale_lock); |
567 | cancel_work_sync(&zhdr->work); | |
1f862989 | 568 | free_z3fold_page(page, false); |
d30561c5 VW |
569 | cond_resched(); |
570 | spin_lock(&pool->stale_lock); | |
571 | } | |
572 | spin_unlock(&pool->stale_lock); | |
573 | } | |
574 | ||
9a001fc1 VW |
575 | /* |
576 | * Returns the number of free chunks in a z3fold page. | |
577 | * NB: can't be used with HEADLESS pages. | |
578 | */ | |
579 | static int num_free_chunks(struct z3fold_header *zhdr) | |
580 | { | |
581 | int nfree; | |
582 | /* | |
583 | * If there is a middle object, pick up the bigger free space | |
584 | * either before or after it. Otherwise just subtract the number | |
585 | * of chunks occupied by the first and the last objects. | |
586 | */ | |
587 | if (zhdr->middle_chunks != 0) { | |
588 | int nfree_before = zhdr->first_chunks ? | |
ede93213 | 589 | 0 : zhdr->start_middle - ZHDR_CHUNKS; |
9a001fc1 | 590 | int nfree_after = zhdr->last_chunks ? |
ede93213 VW |
591 | 0 : TOTAL_CHUNKS - |
592 | (zhdr->start_middle + zhdr->middle_chunks); | |
9a001fc1 VW |
593 | nfree = max(nfree_before, nfree_after); |
594 | } else | |
595 | nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks; | |
596 | return nfree; | |
597 | } | |
598 | ||
9050cce1 VW |
599 | /* Add to the appropriate unbuddied list */ |
600 | static inline void add_to_unbuddied(struct z3fold_pool *pool, | |
601 | struct z3fold_header *zhdr) | |
602 | { | |
603 | if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 || | |
604 | zhdr->middle_chunks == 0) { | |
135f97fd | 605 | struct list_head *unbuddied; |
9050cce1 | 606 | int freechunks = num_free_chunks(zhdr); |
135f97fd VW |
607 | |
608 | migrate_disable(); | |
609 | unbuddied = this_cpu_ptr(pool->unbuddied); | |
9050cce1 VW |
610 | spin_lock(&pool->lock); |
611 | list_add(&zhdr->buddy, &unbuddied[freechunks]); | |
612 | spin_unlock(&pool->lock); | |
613 | zhdr->cpu = smp_processor_id(); | |
135f97fd | 614 | migrate_enable(); |
9050cce1 VW |
615 | } |
616 | } | |
617 | ||
dcf5aedb VW |
618 | static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks) |
619 | { | |
620 | enum buddy bud = HEADLESS; | |
621 | ||
622 | if (zhdr->middle_chunks) { | |
623 | if (!zhdr->first_chunks && | |
624 | chunks <= zhdr->start_middle - ZHDR_CHUNKS) | |
625 | bud = FIRST; | |
626 | else if (!zhdr->last_chunks) | |
627 | bud = LAST; | |
628 | } else { | |
629 | if (!zhdr->first_chunks) | |
630 | bud = FIRST; | |
631 | else if (!zhdr->last_chunks) | |
632 | bud = LAST; | |
633 | else | |
634 | bud = MIDDLE; | |
635 | } | |
636 | ||
637 | return bud; | |
638 | } | |
639 | ||
ede93213 VW |
640 | static inline void *mchunk_memmove(struct z3fold_header *zhdr, |
641 | unsigned short dst_chunk) | |
642 | { | |
643 | void *beg = zhdr; | |
644 | return memmove(beg + (dst_chunk << CHUNK_SHIFT), | |
645 | beg + (zhdr->start_middle << CHUNK_SHIFT), | |
646 | zhdr->middle_chunks << CHUNK_SHIFT); | |
647 | } | |
648 | ||
4a3ac931 VW |
649 | static inline bool buddy_single(struct z3fold_header *zhdr) |
650 | { | |
651 | return !((zhdr->first_chunks && zhdr->middle_chunks) || | |
652 | (zhdr->first_chunks && zhdr->last_chunks) || | |
653 | (zhdr->middle_chunks && zhdr->last_chunks)); | |
654 | } | |
655 | ||
656 | static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr) | |
657 | { | |
658 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); | |
659 | void *p = zhdr; | |
660 | unsigned long old_handle = 0; | |
661 | size_t sz = 0; | |
662 | struct z3fold_header *new_zhdr = NULL; | |
663 | int first_idx = __idx(zhdr, FIRST); | |
664 | int middle_idx = __idx(zhdr, MIDDLE); | |
665 | int last_idx = __idx(zhdr, LAST); | |
666 | unsigned short *moved_chunks = NULL; | |
667 | ||
668 | /* | |
669 | * No need to protect slots here -- all the slots are "local" and | |
670 | * the page lock is already taken | |
671 | */ | |
672 | if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) { | |
673 | p += ZHDR_SIZE_ALIGNED; | |
674 | sz = zhdr->first_chunks << CHUNK_SHIFT; | |
675 | old_handle = (unsigned long)&zhdr->slots->slot[first_idx]; | |
676 | moved_chunks = &zhdr->first_chunks; | |
677 | } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) { | |
678 | p += zhdr->start_middle << CHUNK_SHIFT; | |
679 | sz = zhdr->middle_chunks << CHUNK_SHIFT; | |
680 | old_handle = (unsigned long)&zhdr->slots->slot[middle_idx]; | |
681 | moved_chunks = &zhdr->middle_chunks; | |
682 | } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) { | |
683 | p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT); | |
684 | sz = zhdr->last_chunks << CHUNK_SHIFT; | |
685 | old_handle = (unsigned long)&zhdr->slots->slot[last_idx]; | |
686 | moved_chunks = &zhdr->last_chunks; | |
687 | } | |
688 | ||
689 | if (sz > 0) { | |
690 | enum buddy new_bud = HEADLESS; | |
691 | short chunks = size_to_chunks(sz); | |
692 | void *q; | |
693 | ||
694 | new_zhdr = __z3fold_alloc(pool, sz, false); | |
695 | if (!new_zhdr) | |
696 | return NULL; | |
697 | ||
698 | if (WARN_ON(new_zhdr == zhdr)) | |
699 | goto out_fail; | |
700 | ||
dcf5aedb | 701 | new_bud = get_free_buddy(new_zhdr, chunks); |
4a3ac931 VW |
702 | q = new_zhdr; |
703 | switch (new_bud) { | |
704 | case FIRST: | |
705 | new_zhdr->first_chunks = chunks; | |
706 | q += ZHDR_SIZE_ALIGNED; | |
707 | break; | |
708 | case MIDDLE: | |
709 | new_zhdr->middle_chunks = chunks; | |
710 | new_zhdr->start_middle = | |
711 | new_zhdr->first_chunks + ZHDR_CHUNKS; | |
712 | q += new_zhdr->start_middle << CHUNK_SHIFT; | |
713 | break; | |
714 | case LAST: | |
715 | new_zhdr->last_chunks = chunks; | |
716 | q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT); | |
717 | break; | |
718 | default: | |
719 | goto out_fail; | |
720 | } | |
721 | new_zhdr->foreign_handles++; | |
722 | memcpy(q, p, sz); | |
723 | write_lock(&zhdr->slots->lock); | |
724 | *(unsigned long *)old_handle = (unsigned long)new_zhdr + | |
725 | __idx(new_zhdr, new_bud); | |
726 | if (new_bud == LAST) | |
727 | *(unsigned long *)old_handle |= | |
728 | (new_zhdr->last_chunks << BUDDY_SHIFT); | |
729 | write_unlock(&zhdr->slots->lock); | |
730 | add_to_unbuddied(pool, new_zhdr); | |
731 | z3fold_page_unlock(new_zhdr); | |
732 | ||
733 | *moved_chunks = 0; | |
734 | } | |
735 | ||
736 | return new_zhdr; | |
737 | ||
738 | out_fail: | |
739 | if (new_zhdr) { | |
740 | if (kref_put(&new_zhdr->refcount, release_z3fold_page_locked)) | |
741 | atomic64_dec(&pool->pages_nr); | |
742 | else { | |
743 | add_to_unbuddied(pool, new_zhdr); | |
744 | z3fold_page_unlock(new_zhdr); | |
745 | } | |
746 | } | |
747 | return NULL; | |
748 | ||
749 | } | |
750 | ||
1b096e5a | 751 | #define BIG_CHUNK_GAP 3 |
9a001fc1 VW |
752 | /* Has to be called with lock held */ |
753 | static int z3fold_compact_page(struct z3fold_header *zhdr) | |
754 | { | |
755 | struct page *page = virt_to_page(zhdr); | |
9a001fc1 | 756 | |
ede93213 VW |
757 | if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private)) |
758 | return 0; /* can't move middle chunk, it's used */ | |
9a001fc1 | 759 | |
1f862989 VW |
760 | if (unlikely(PageIsolated(page))) |
761 | return 0; | |
762 | ||
ede93213 VW |
763 | if (zhdr->middle_chunks == 0) |
764 | return 0; /* nothing to compact */ | |
765 | ||
766 | if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) { | |
767 | /* move to the beginning */ | |
768 | mchunk_memmove(zhdr, ZHDR_CHUNKS); | |
9a001fc1 VW |
769 | zhdr->first_chunks = zhdr->middle_chunks; |
770 | zhdr->middle_chunks = 0; | |
771 | zhdr->start_middle = 0; | |
772 | zhdr->first_num++; | |
1b096e5a | 773 | return 1; |
9a001fc1 | 774 | } |
1b096e5a VW |
775 | |
776 | /* | |
777 | * moving data is expensive, so let's only do that if | |
778 | * there's substantial gain (at least BIG_CHUNK_GAP chunks) | |
779 | */ | |
780 | if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 && | |
781 | zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >= | |
782 | BIG_CHUNK_GAP) { | |
783 | mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS); | |
784 | zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS; | |
785 | return 1; | |
786 | } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 && | |
787 | TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle | |
788 | + zhdr->middle_chunks) >= | |
789 | BIG_CHUNK_GAP) { | |
790 | unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks - | |
791 | zhdr->middle_chunks; | |
792 | mchunk_memmove(zhdr, new_start); | |
793 | zhdr->start_middle = new_start; | |
794 | return 1; | |
795 | } | |
796 | ||
797 | return 0; | |
9a001fc1 VW |
798 | } |
799 | ||
d30561c5 VW |
800 | static void do_compact_page(struct z3fold_header *zhdr, bool locked) |
801 | { | |
9050cce1 | 802 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
d30561c5 | 803 | struct page *page; |
d30561c5 VW |
804 | |
805 | page = virt_to_page(zhdr); | |
806 | if (locked) | |
807 | WARN_ON(z3fold_page_trylock(zhdr)); | |
808 | else | |
809 | z3fold_page_lock(zhdr); | |
5d03a661 | 810 | if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) { |
d30561c5 VW |
811 | z3fold_page_unlock(zhdr); |
812 | return; | |
813 | } | |
814 | spin_lock(&pool->lock); | |
815 | list_del_init(&zhdr->buddy); | |
816 | spin_unlock(&pool->lock); | |
817 | ||
5d03a661 VW |
818 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) { |
819 | atomic64_dec(&pool->pages_nr); | |
820 | return; | |
821 | } | |
822 | ||
dcf5aedb VW |
823 | if (test_bit(PAGE_STALE, &page->private) || |
824 | test_and_set_bit(PAGE_CLAIMED, &page->private)) { | |
1f862989 VW |
825 | z3fold_page_unlock(zhdr); |
826 | return; | |
827 | } | |
828 | ||
4a3ac931 VW |
829 | if (!zhdr->foreign_handles && buddy_single(zhdr) && |
830 | zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) { | |
831 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) | |
832 | atomic64_dec(&pool->pages_nr); | |
dcf5aedb VW |
833 | else { |
834 | clear_bit(PAGE_CLAIMED, &page->private); | |
4a3ac931 | 835 | z3fold_page_unlock(zhdr); |
dcf5aedb | 836 | } |
4a3ac931 VW |
837 | return; |
838 | } | |
839 | ||
d30561c5 | 840 | z3fold_compact_page(zhdr); |
9050cce1 | 841 | add_to_unbuddied(pool, zhdr); |
dcf5aedb | 842 | clear_bit(PAGE_CLAIMED, &page->private); |
d30561c5 VW |
843 | z3fold_page_unlock(zhdr); |
844 | } | |
845 | ||
846 | static void compact_page_work(struct work_struct *w) | |
847 | { | |
848 | struct z3fold_header *zhdr = container_of(w, struct z3fold_header, | |
849 | work); | |
850 | ||
851 | do_compact_page(zhdr, false); | |
852 | } | |
853 | ||
9050cce1 VW |
854 | /* returns _locked_ z3fold page header or NULL */ |
855 | static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool, | |
856 | size_t size, bool can_sleep) | |
857 | { | |
858 | struct z3fold_header *zhdr = NULL; | |
859 | struct page *page; | |
860 | struct list_head *unbuddied; | |
861 | int chunks = size_to_chunks(size), i; | |
862 | ||
863 | lookup: | |
135f97fd | 864 | migrate_disable(); |
9050cce1 | 865 | /* First, try to find an unbuddied z3fold page. */ |
135f97fd | 866 | unbuddied = this_cpu_ptr(pool->unbuddied); |
9050cce1 VW |
867 | for_each_unbuddied_list(i, chunks) { |
868 | struct list_head *l = &unbuddied[i]; | |
869 | ||
870 | zhdr = list_first_entry_or_null(READ_ONCE(l), | |
871 | struct z3fold_header, buddy); | |
872 | ||
873 | if (!zhdr) | |
874 | continue; | |
875 | ||
876 | /* Re-check under lock. */ | |
877 | spin_lock(&pool->lock); | |
878 | l = &unbuddied[i]; | |
879 | if (unlikely(zhdr != list_first_entry(READ_ONCE(l), | |
880 | struct z3fold_header, buddy)) || | |
881 | !z3fold_page_trylock(zhdr)) { | |
882 | spin_unlock(&pool->lock); | |
883 | zhdr = NULL; | |
135f97fd | 884 | migrate_enable(); |
9050cce1 VW |
885 | if (can_sleep) |
886 | cond_resched(); | |
887 | goto lookup; | |
888 | } | |
889 | list_del_init(&zhdr->buddy); | |
890 | zhdr->cpu = -1; | |
891 | spin_unlock(&pool->lock); | |
892 | ||
893 | page = virt_to_page(zhdr); | |
4a3ac931 VW |
894 | if (test_bit(NEEDS_COMPACTING, &page->private) || |
895 | test_bit(PAGE_CLAIMED, &page->private)) { | |
9050cce1 VW |
896 | z3fold_page_unlock(zhdr); |
897 | zhdr = NULL; | |
135f97fd | 898 | migrate_enable(); |
9050cce1 VW |
899 | if (can_sleep) |
900 | cond_resched(); | |
901 | goto lookup; | |
902 | } | |
903 | ||
904 | /* | |
905 | * this page could not be removed from its unbuddied | |
906 | * list while pool lock was held, and then we've taken | |
907 | * page lock so kref_put could not be called before | |
908 | * we got here, so it's safe to just call kref_get() | |
909 | */ | |
910 | kref_get(&zhdr->refcount); | |
911 | break; | |
912 | } | |
135f97fd | 913 | migrate_enable(); |
9050cce1 | 914 | |
351618b2 VW |
915 | if (!zhdr) { |
916 | int cpu; | |
917 | ||
918 | /* look for _exact_ match on other cpus' lists */ | |
919 | for_each_online_cpu(cpu) { | |
920 | struct list_head *l; | |
921 | ||
922 | unbuddied = per_cpu_ptr(pool->unbuddied, cpu); | |
923 | spin_lock(&pool->lock); | |
924 | l = &unbuddied[chunks]; | |
925 | ||
926 | zhdr = list_first_entry_or_null(READ_ONCE(l), | |
927 | struct z3fold_header, buddy); | |
928 | ||
929 | if (!zhdr || !z3fold_page_trylock(zhdr)) { | |
930 | spin_unlock(&pool->lock); | |
931 | zhdr = NULL; | |
932 | continue; | |
933 | } | |
934 | list_del_init(&zhdr->buddy); | |
935 | zhdr->cpu = -1; | |
936 | spin_unlock(&pool->lock); | |
937 | ||
938 | page = virt_to_page(zhdr); | |
4a3ac931 VW |
939 | if (test_bit(NEEDS_COMPACTING, &page->private) || |
940 | test_bit(PAGE_CLAIMED, &page->private)) { | |
351618b2 VW |
941 | z3fold_page_unlock(zhdr); |
942 | zhdr = NULL; | |
943 | if (can_sleep) | |
944 | cond_resched(); | |
945 | continue; | |
946 | } | |
947 | kref_get(&zhdr->refcount); | |
948 | break; | |
949 | } | |
950 | } | |
951 | ||
fc548865 VW |
952 | if (zhdr && !zhdr->slots) |
953 | zhdr->slots = alloc_slots(pool, | |
954 | can_sleep ? GFP_NOIO : GFP_ATOMIC); | |
9050cce1 VW |
955 | return zhdr; |
956 | } | |
d30561c5 VW |
957 | |
958 | /* | |
959 | * API Functions | |
960 | */ | |
961 | ||
962 | /** | |
963 | * z3fold_create_pool() - create a new z3fold pool | |
964 | * @name: pool name | |
965 | * @gfp: gfp flags when allocating the z3fold pool structure | |
966 | * @ops: user-defined operations for the z3fold pool | |
967 | * | |
968 | * Return: pointer to the new z3fold pool or NULL if the metadata allocation | |
969 | * failed. | |
970 | */ | |
971 | static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp, | |
972 | const struct z3fold_ops *ops) | |
973 | { | |
974 | struct z3fold_pool *pool = NULL; | |
975 | int i, cpu; | |
976 | ||
977 | pool = kzalloc(sizeof(struct z3fold_pool), gfp); | |
978 | if (!pool) | |
979 | goto out; | |
7c2b8baa VW |
980 | pool->c_handle = kmem_cache_create("z3fold_handle", |
981 | sizeof(struct z3fold_buddy_slots), | |
982 | SLOTS_ALIGN, 0, NULL); | |
983 | if (!pool->c_handle) | |
984 | goto out_c; | |
d30561c5 VW |
985 | spin_lock_init(&pool->lock); |
986 | spin_lock_init(&pool->stale_lock); | |
e891f60e ML |
987 | pool->unbuddied = __alloc_percpu(sizeof(struct list_head) * NCHUNKS, |
988 | __alignof__(struct list_head)); | |
1ec6995d XW |
989 | if (!pool->unbuddied) |
990 | goto out_pool; | |
d30561c5 VW |
991 | for_each_possible_cpu(cpu) { |
992 | struct list_head *unbuddied = | |
993 | per_cpu_ptr(pool->unbuddied, cpu); | |
994 | for_each_unbuddied_list(i, 0) | |
995 | INIT_LIST_HEAD(&unbuddied[i]); | |
996 | } | |
997 | INIT_LIST_HEAD(&pool->lru); | |
998 | INIT_LIST_HEAD(&pool->stale); | |
999 | atomic64_set(&pool->pages_nr, 0); | |
1000 | pool->name = name; | |
1001 | pool->compact_wq = create_singlethread_workqueue(pool->name); | |
1002 | if (!pool->compact_wq) | |
1ec6995d | 1003 | goto out_unbuddied; |
d30561c5 VW |
1004 | pool->release_wq = create_singlethread_workqueue(pool->name); |
1005 | if (!pool->release_wq) | |
1006 | goto out_wq; | |
1f862989 VW |
1007 | if (z3fold_register_migration(pool)) |
1008 | goto out_rwq; | |
d30561c5 VW |
1009 | INIT_WORK(&pool->work, free_pages_work); |
1010 | pool->ops = ops; | |
1011 | return pool; | |
1012 | ||
1f862989 VW |
1013 | out_rwq: |
1014 | destroy_workqueue(pool->release_wq); | |
d30561c5 VW |
1015 | out_wq: |
1016 | destroy_workqueue(pool->compact_wq); | |
1ec6995d XW |
1017 | out_unbuddied: |
1018 | free_percpu(pool->unbuddied); | |
1019 | out_pool: | |
7c2b8baa VW |
1020 | kmem_cache_destroy(pool->c_handle); |
1021 | out_c: | |
d30561c5 | 1022 | kfree(pool); |
1ec6995d | 1023 | out: |
d30561c5 VW |
1024 | return NULL; |
1025 | } | |
1026 | ||
1027 | /** | |
1028 | * z3fold_destroy_pool() - destroys an existing z3fold pool | |
1029 | * @pool: the z3fold pool to be destroyed | |
1030 | * | |
1031 | * The pool should be emptied before this function is called. | |
1032 | */ | |
1033 | static void z3fold_destroy_pool(struct z3fold_pool *pool) | |
1034 | { | |
7c2b8baa | 1035 | kmem_cache_destroy(pool->c_handle); |
6051d3bd HB |
1036 | |
1037 | /* | |
1038 | * We need to destroy pool->compact_wq before pool->release_wq, | |
1039 | * as any pending work on pool->compact_wq will call | |
1040 | * queue_work(pool->release_wq, &pool->work). | |
b997052b HB |
1041 | * |
1042 | * There are still outstanding pages until both workqueues are drained, | |
1043 | * so we cannot unregister migration until then. | |
6051d3bd HB |
1044 | */ |
1045 | ||
d30561c5 | 1046 | destroy_workqueue(pool->compact_wq); |
6051d3bd | 1047 | destroy_workqueue(pool->release_wq); |
b997052b | 1048 | z3fold_unregister_migration(pool); |
dac0d1cf | 1049 | free_percpu(pool->unbuddied); |
d30561c5 VW |
1050 | kfree(pool); |
1051 | } | |
1052 | ||
9a001fc1 VW |
1053 | /** |
1054 | * z3fold_alloc() - allocates a region of a given size | |
1055 | * @pool: z3fold pool from which to allocate | |
1056 | * @size: size in bytes of the desired allocation | |
1057 | * @gfp: gfp flags used if the pool needs to grow | |
1058 | * @handle: handle of the new allocation | |
1059 | * | |
1060 | * This function will attempt to find a free region in the pool large enough to | |
1061 | * satisfy the allocation request. A search of the unbuddied lists is | |
1062 | * performed first. If no suitable free region is found, then a new page is | |
1063 | * allocated and added to the pool to satisfy the request. | |
1064 | * | |
1065 | * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used | |
1066 | * as z3fold pool pages. | |
1067 | * | |
1068 | * Return: 0 if success and handle is set, otherwise -EINVAL if the size or | |
1069 | * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate | |
1070 | * a new page. | |
1071 | */ | |
1072 | static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp, | |
1073 | unsigned long *handle) | |
1074 | { | |
9050cce1 | 1075 | int chunks = size_to_chunks(size); |
9a001fc1 | 1076 | struct z3fold_header *zhdr = NULL; |
d30561c5 | 1077 | struct page *page = NULL; |
9a001fc1 | 1078 | enum buddy bud; |
8a97ea54 | 1079 | bool can_sleep = gfpflags_allow_blocking(gfp); |
9a001fc1 | 1080 | |
f1549cb5 | 1081 | if (!size) |
9a001fc1 VW |
1082 | return -EINVAL; |
1083 | ||
1084 | if (size > PAGE_SIZE) | |
1085 | return -ENOSPC; | |
1086 | ||
1087 | if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE) | |
1088 | bud = HEADLESS; | |
1089 | else { | |
9050cce1 VW |
1090 | retry: |
1091 | zhdr = __z3fold_alloc(pool, size, can_sleep); | |
d30561c5 | 1092 | if (zhdr) { |
dcf5aedb VW |
1093 | bud = get_free_buddy(zhdr, chunks); |
1094 | if (bud == HEADLESS) { | |
5a27aa82 | 1095 | if (kref_put(&zhdr->refcount, |
d30561c5 | 1096 | release_z3fold_page_locked)) |
5a27aa82 | 1097 | atomic64_dec(&pool->pages_nr); |
d30561c5 VW |
1098 | else |
1099 | z3fold_page_unlock(zhdr); | |
2f1e5e4d VW |
1100 | pr_err("No free chunks in unbuddied\n"); |
1101 | WARN_ON(1); | |
9050cce1 | 1102 | goto retry; |
9a001fc1 | 1103 | } |
9050cce1 | 1104 | page = virt_to_page(zhdr); |
2f1e5e4d | 1105 | goto found; |
9a001fc1 VW |
1106 | } |
1107 | bud = FIRST; | |
9a001fc1 VW |
1108 | } |
1109 | ||
5c9bab59 VW |
1110 | page = NULL; |
1111 | if (can_sleep) { | |
1112 | spin_lock(&pool->stale_lock); | |
1113 | zhdr = list_first_entry_or_null(&pool->stale, | |
1114 | struct z3fold_header, buddy); | |
1115 | /* | |
1116 | * Before allocating a page, let's see if we can take one from | |
1117 | * the stale pages list. cancel_work_sync() can sleep so we | |
1118 | * limit this case to the contexts where we can sleep | |
1119 | */ | |
1120 | if (zhdr) { | |
1121 | list_del(&zhdr->buddy); | |
1122 | spin_unlock(&pool->stale_lock); | |
d30561c5 | 1123 | cancel_work_sync(&zhdr->work); |
5c9bab59 VW |
1124 | page = virt_to_page(zhdr); |
1125 | } else { | |
1126 | spin_unlock(&pool->stale_lock); | |
1127 | } | |
d30561c5 | 1128 | } |
5c9bab59 VW |
1129 | if (!page) |
1130 | page = alloc_page(gfp); | |
d30561c5 | 1131 | |
9a001fc1 VW |
1132 | if (!page) |
1133 | return -ENOMEM; | |
2f1e5e4d | 1134 | |
63398413 | 1135 | zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp); |
9050cce1 VW |
1136 | if (!zhdr) { |
1137 | __free_page(page); | |
1138 | return -ENOMEM; | |
1139 | } | |
1140 | atomic64_inc(&pool->pages_nr); | |
9a001fc1 VW |
1141 | |
1142 | if (bud == HEADLESS) { | |
1143 | set_bit(PAGE_HEADLESS, &page->private); | |
1144 | goto headless; | |
1145 | } | |
810481a2 HB |
1146 | if (can_sleep) { |
1147 | lock_page(page); | |
1148 | __SetPageMovable(page, pool->inode->i_mapping); | |
1149 | unlock_page(page); | |
1150 | } else { | |
1151 | if (trylock_page(page)) { | |
1152 | __SetPageMovable(page, pool->inode->i_mapping); | |
1153 | unlock_page(page); | |
1154 | } | |
1155 | } | |
2f1e5e4d | 1156 | z3fold_page_lock(zhdr); |
9a001fc1 VW |
1157 | |
1158 | found: | |
1159 | if (bud == FIRST) | |
1160 | zhdr->first_chunks = chunks; | |
1161 | else if (bud == LAST) | |
1162 | zhdr->last_chunks = chunks; | |
1163 | else { | |
1164 | zhdr->middle_chunks = chunks; | |
ede93213 | 1165 | zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS; |
9a001fc1 | 1166 | } |
9050cce1 | 1167 | add_to_unbuddied(pool, zhdr); |
9a001fc1 VW |
1168 | |
1169 | headless: | |
d30561c5 | 1170 | spin_lock(&pool->lock); |
9a001fc1 VW |
1171 | /* Add/move z3fold page to beginning of LRU */ |
1172 | if (!list_empty(&page->lru)) | |
1173 | list_del(&page->lru); | |
1174 | ||
1175 | list_add(&page->lru, &pool->lru); | |
1176 | ||
1177 | *handle = encode_handle(zhdr, bud); | |
1178 | spin_unlock(&pool->lock); | |
2f1e5e4d VW |
1179 | if (bud != HEADLESS) |
1180 | z3fold_page_unlock(zhdr); | |
9a001fc1 VW |
1181 | |
1182 | return 0; | |
1183 | } | |
1184 | ||
1185 | /** | |
1186 | * z3fold_free() - frees the allocation associated with the given handle | |
1187 | * @pool: pool in which the allocation resided | |
1188 | * @handle: handle associated with the allocation returned by z3fold_alloc() | |
1189 | * | |
1190 | * In the case that the z3fold page in which the allocation resides is under | |
1191 | * reclaim, as indicated by the PG_reclaim flag being set, this function | |
1192 | * only sets the first|last_chunks to 0. The page is actually freed | |
1193 | * once both buddies are evicted (see z3fold_reclaim_page() below). | |
1194 | */ | |
1195 | static void z3fold_free(struct z3fold_pool *pool, unsigned long handle) | |
1196 | { | |
1197 | struct z3fold_header *zhdr; | |
9a001fc1 VW |
1198 | struct page *page; |
1199 | enum buddy bud; | |
5b6807de | 1200 | bool page_claimed; |
9a001fc1 | 1201 | |
4a3ac931 | 1202 | zhdr = get_z3fold_header(handle); |
9a001fc1 | 1203 | page = virt_to_page(zhdr); |
5b6807de | 1204 | page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private); |
9a001fc1 VW |
1205 | |
1206 | if (test_bit(PAGE_HEADLESS, &page->private)) { | |
ca0246bb VW |
1207 | /* if a headless page is under reclaim, just leave. |
1208 | * NB: we use test_and_set_bit for a reason: if the bit | |
1209 | * has not been set before, we release this page | |
1210 | * immediately so we don't care about its value any more. | |
1211 | */ | |
5b6807de | 1212 | if (!page_claimed) { |
ca0246bb VW |
1213 | spin_lock(&pool->lock); |
1214 | list_del(&page->lru); | |
1215 | spin_unlock(&pool->lock); | |
4a3ac931 | 1216 | put_z3fold_header(zhdr); |
1f862989 | 1217 | free_z3fold_page(page, true); |
ca0246bb | 1218 | atomic64_dec(&pool->pages_nr); |
9a001fc1 | 1219 | } |
ca0246bb | 1220 | return; |
9a001fc1 VW |
1221 | } |
1222 | ||
ca0246bb | 1223 | /* Non-headless case */ |
ca0246bb VW |
1224 | bud = handle_to_buddy(handle); |
1225 | ||
1226 | switch (bud) { | |
1227 | case FIRST: | |
1228 | zhdr->first_chunks = 0; | |
1229 | break; | |
1230 | case MIDDLE: | |
1231 | zhdr->middle_chunks = 0; | |
1232 | break; | |
1233 | case LAST: | |
1234 | zhdr->last_chunks = 0; | |
1235 | break; | |
1236 | default: | |
1237 | pr_err("%s: unknown bud %d\n", __func__, bud); | |
1238 | WARN_ON(1); | |
4a3ac931 | 1239 | put_z3fold_header(zhdr); |
d30561c5 VW |
1240 | return; |
1241 | } | |
1242 | ||
4a3ac931 | 1243 | if (!page_claimed) |
fc548865 | 1244 | free_handle(handle, zhdr); |
d30561c5 VW |
1245 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) { |
1246 | atomic64_dec(&pool->pages_nr); | |
1247 | return; | |
1248 | } | |
5b6807de VW |
1249 | if (page_claimed) { |
1250 | /* the page has not been claimed by us */ | |
6098d7e1 VW |
1251 | z3fold_page_unlock(zhdr); |
1252 | return; | |
1253 | } | |
dcf5aedb | 1254 | if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) { |
4a3ac931 | 1255 | put_z3fold_header(zhdr); |
5b6807de | 1256 | clear_bit(PAGE_CLAIMED, &page->private); |
d30561c5 VW |
1257 | return; |
1258 | } | |
1259 | if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) { | |
2f1e5e4d | 1260 | spin_lock(&pool->lock); |
d30561c5 | 1261 | list_del_init(&zhdr->buddy); |
2f1e5e4d | 1262 | spin_unlock(&pool->lock); |
d30561c5 | 1263 | zhdr->cpu = -1; |
5d03a661 | 1264 | kref_get(&zhdr->refcount); |
5b6807de | 1265 | clear_bit(PAGE_CLAIMED, &page->private); |
4a3ac931 | 1266 | do_compact_page(zhdr, true); |
d30561c5 | 1267 | return; |
9a001fc1 | 1268 | } |
5d03a661 | 1269 | kref_get(&zhdr->refcount); |
5b6807de | 1270 | clear_bit(PAGE_CLAIMED, &page->private); |
4a3ac931 VW |
1271 | queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work); |
1272 | put_z3fold_header(zhdr); | |
9a001fc1 VW |
1273 | } |
1274 | ||
1275 | /** | |
1276 | * z3fold_reclaim_page() - evicts allocations from a pool page and frees it | |
1277 | * @pool: pool from which a page will attempt to be evicted | |
f144c390 | 1278 | * @retries: number of pages on the LRU list for which eviction will |
9a001fc1 VW |
1279 | * be attempted before failing |
1280 | * | |
1281 | * z3fold reclaim is different from normal system reclaim in that it is done | |
1282 | * from the bottom, up. This is because only the bottom layer, z3fold, has | |
1283 | * information on how the allocations are organized within each z3fold page. | |
1284 | * This has the potential to create interesting locking situations between | |
1285 | * z3fold and the user, however. | |
1286 | * | |
1287 | * To avoid these, this is how z3fold_reclaim_page() should be called: | |
f144c390 | 1288 | * |
9a001fc1 VW |
1289 | * The user detects a page should be reclaimed and calls z3fold_reclaim_page(). |
1290 | * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and | |
1291 | * call the user-defined eviction handler with the pool and handle as | |
1292 | * arguments. | |
1293 | * | |
1294 | * If the handle can not be evicted, the eviction handler should return | |
1295 | * non-zero. z3fold_reclaim_page() will add the z3fold page back to the | |
1296 | * appropriate list and try the next z3fold page on the LRU up to | |
1297 | * a user defined number of retries. | |
1298 | * | |
1299 | * If the handle is successfully evicted, the eviction handler should | |
1300 | * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free() | |
1301 | * contains logic to delay freeing the page if the page is under reclaim, | |
1302 | * as indicated by the setting of the PG_reclaim flag on the underlying page. | |
1303 | * | |
1304 | * If all buddies in the z3fold page are successfully evicted, then the | |
1305 | * z3fold page can be freed. | |
1306 | * | |
1307 | * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are | |
1308 | * no pages to evict or an eviction handler is not registered, -EAGAIN if | |
1309 | * the retry limit was hit. | |
1310 | */ | |
1311 | static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries) | |
1312 | { | |
4a3ac931 | 1313 | int i, ret = -1; |
d30561c5 VW |
1314 | struct z3fold_header *zhdr = NULL; |
1315 | struct page *page = NULL; | |
1316 | struct list_head *pos; | |
9a001fc1 | 1317 | unsigned long first_handle = 0, middle_handle = 0, last_handle = 0; |
dcf5aedb VW |
1318 | struct z3fold_buddy_slots slots __attribute__((aligned(SLOTS_ALIGN))); |
1319 | ||
1320 | rwlock_init(&slots.lock); | |
1321 | slots.pool = (unsigned long)pool | (1 << HANDLES_NOFREE); | |
9a001fc1 VW |
1322 | |
1323 | spin_lock(&pool->lock); | |
2f1e5e4d | 1324 | if (!pool->ops || !pool->ops->evict || retries == 0) { |
9a001fc1 VW |
1325 | spin_unlock(&pool->lock); |
1326 | return -EINVAL; | |
1327 | } | |
1328 | for (i = 0; i < retries; i++) { | |
2f1e5e4d VW |
1329 | if (list_empty(&pool->lru)) { |
1330 | spin_unlock(&pool->lock); | |
1331 | return -EINVAL; | |
1332 | } | |
d30561c5 VW |
1333 | list_for_each_prev(pos, &pool->lru) { |
1334 | page = list_entry(pos, struct page, lru); | |
ca0246bb | 1335 | |
3f9d2b57 | 1336 | zhdr = page_address(page); |
6d679578 TH |
1337 | if (test_bit(PAGE_HEADLESS, &page->private)) { |
1338 | /* | |
1339 | * For non-headless pages, we wait to do this | |
1340 | * until we have the page lock to avoid racing | |
1341 | * with __z3fold_alloc(). Headless pages don't | |
1342 | * have a lock (and __z3fold_alloc() will never | |
1343 | * see them), but we still need to test and set | |
1344 | * PAGE_CLAIMED to avoid racing with | |
1345 | * z3fold_free(), so just do it now before | |
1346 | * leaving the loop. | |
1347 | */ | |
1348 | if (test_and_set_bit(PAGE_CLAIMED, &page->private)) | |
1349 | continue; | |
1350 | ||
d30561c5 | 1351 | break; |
6d679578 | 1352 | } |
d30561c5 | 1353 | |
dcf5aedb VW |
1354 | if (kref_get_unless_zero(&zhdr->refcount) == 0) { |
1355 | zhdr = NULL; | |
1356 | break; | |
1357 | } | |
ca0246bb | 1358 | if (!z3fold_page_trylock(zhdr)) { |
dcf5aedb VW |
1359 | if (kref_put(&zhdr->refcount, |
1360 | release_z3fold_page)) | |
1361 | atomic64_dec(&pool->pages_nr); | |
ca0246bb | 1362 | zhdr = NULL; |
d30561c5 | 1363 | continue; /* can't evict at this point */ |
ca0246bb | 1364 | } |
dcf5aedb VW |
1365 | |
1366 | /* test_and_set_bit is of course atomic, but we still | |
1367 | * need to do it under page lock, otherwise checking | |
1368 | * that bit in __z3fold_alloc wouldn't make sense | |
1369 | */ | |
1370 | if (zhdr->foreign_handles || | |
1371 | test_and_set_bit(PAGE_CLAIMED, &page->private)) { | |
1372 | if (kref_put(&zhdr->refcount, | |
1373 | release_z3fold_page)) | |
1374 | atomic64_dec(&pool->pages_nr); | |
1375 | else | |
1376 | z3fold_page_unlock(zhdr); | |
4a3ac931 VW |
1377 | zhdr = NULL; |
1378 | continue; /* can't evict such page */ | |
1379 | } | |
d30561c5 VW |
1380 | list_del_init(&zhdr->buddy); |
1381 | zhdr->cpu = -1; | |
6098d7e1 | 1382 | break; |
d30561c5 VW |
1383 | } |
1384 | ||
ca0246bb VW |
1385 | if (!zhdr) |
1386 | break; | |
1387 | ||
5a27aa82 | 1388 | list_del_init(&page->lru); |
d30561c5 | 1389 | spin_unlock(&pool->lock); |
9a001fc1 | 1390 | |
9a001fc1 | 1391 | if (!test_bit(PAGE_HEADLESS, &page->private)) { |
9a001fc1 | 1392 | /* |
3f9d2b57 VW |
1393 | * We need encode the handles before unlocking, and |
1394 | * use our local slots structure because z3fold_free | |
1395 | * can zero out zhdr->slots and we can't do much | |
1396 | * about that | |
9a001fc1 VW |
1397 | */ |
1398 | first_handle = 0; | |
1399 | last_handle = 0; | |
1400 | middle_handle = 0; | |
dcf5aedb | 1401 | memset(slots.slot, 0, sizeof(slots.slot)); |
9a001fc1 | 1402 | if (zhdr->first_chunks) |
dcf5aedb VW |
1403 | first_handle = __encode_handle(zhdr, &slots, |
1404 | FIRST); | |
9a001fc1 | 1405 | if (zhdr->middle_chunks) |
dcf5aedb VW |
1406 | middle_handle = __encode_handle(zhdr, &slots, |
1407 | MIDDLE); | |
9a001fc1 | 1408 | if (zhdr->last_chunks) |
dcf5aedb VW |
1409 | last_handle = __encode_handle(zhdr, &slots, |
1410 | LAST); | |
d30561c5 VW |
1411 | /* |
1412 | * it's safe to unlock here because we hold a | |
1413 | * reference to this page | |
1414 | */ | |
2f1e5e4d | 1415 | z3fold_page_unlock(zhdr); |
9a001fc1 | 1416 | } else { |
4a3ac931 | 1417 | first_handle = encode_handle(zhdr, HEADLESS); |
9a001fc1 VW |
1418 | last_handle = middle_handle = 0; |
1419 | } | |
9a001fc1 VW |
1420 | /* Issue the eviction callback(s) */ |
1421 | if (middle_handle) { | |
1422 | ret = pool->ops->evict(pool, middle_handle); | |
1423 | if (ret) | |
1424 | goto next; | |
1425 | } | |
1426 | if (first_handle) { | |
1427 | ret = pool->ops->evict(pool, first_handle); | |
1428 | if (ret) | |
1429 | goto next; | |
1430 | } | |
1431 | if (last_handle) { | |
1432 | ret = pool->ops->evict(pool, last_handle); | |
1433 | if (ret) | |
1434 | goto next; | |
1435 | } | |
1436 | next: | |
5a27aa82 VW |
1437 | if (test_bit(PAGE_HEADLESS, &page->private)) { |
1438 | if (ret == 0) { | |
1f862989 | 1439 | free_z3fold_page(page, true); |
ca0246bb | 1440 | atomic64_dec(&pool->pages_nr); |
5a27aa82 | 1441 | return 0; |
5a27aa82 | 1442 | } |
6098d7e1 VW |
1443 | spin_lock(&pool->lock); |
1444 | list_add(&page->lru, &pool->lru); | |
1445 | spin_unlock(&pool->lock); | |
3f9d2b57 | 1446 | clear_bit(PAGE_CLAIMED, &page->private); |
6098d7e1 | 1447 | } else { |
dcf5aedb | 1448 | struct z3fold_buddy_slots *slots = zhdr->slots; |
6098d7e1 | 1449 | z3fold_page_lock(zhdr); |
6098d7e1 VW |
1450 | if (kref_put(&zhdr->refcount, |
1451 | release_z3fold_page_locked)) { | |
dcf5aedb | 1452 | kmem_cache_free(pool->c_handle, slots); |
6098d7e1 VW |
1453 | atomic64_dec(&pool->pages_nr); |
1454 | return 0; | |
1455 | } | |
1456 | /* | |
1457 | * if we are here, the page is still not completely | |
1458 | * free. Take the global pool lock then to be able | |
1459 | * to add it back to the lru list | |
1460 | */ | |
1461 | spin_lock(&pool->lock); | |
1462 | list_add(&page->lru, &pool->lru); | |
d5567c9d | 1463 | spin_unlock(&pool->lock); |
6098d7e1 | 1464 | z3fold_page_unlock(zhdr); |
3f9d2b57 | 1465 | clear_bit(PAGE_CLAIMED, &page->private); |
5a27aa82 | 1466 | } |
2f1e5e4d | 1467 | |
6098d7e1 VW |
1468 | /* We started off locked to we need to lock the pool back */ |
1469 | spin_lock(&pool->lock); | |
9a001fc1 VW |
1470 | } |
1471 | spin_unlock(&pool->lock); | |
1472 | return -EAGAIN; | |
1473 | } | |
1474 | ||
1475 | /** | |
1476 | * z3fold_map() - maps the allocation associated with the given handle | |
1477 | * @pool: pool in which the allocation resides | |
1478 | * @handle: handle associated with the allocation to be mapped | |
1479 | * | |
1480 | * Extracts the buddy number from handle and constructs the pointer to the | |
1481 | * correct starting chunk within the page. | |
1482 | * | |
1483 | * Returns: a pointer to the mapped allocation | |
1484 | */ | |
1485 | static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle) | |
1486 | { | |
1487 | struct z3fold_header *zhdr; | |
1488 | struct page *page; | |
1489 | void *addr; | |
1490 | enum buddy buddy; | |
1491 | ||
4a3ac931 | 1492 | zhdr = get_z3fold_header(handle); |
9a001fc1 VW |
1493 | addr = zhdr; |
1494 | page = virt_to_page(zhdr); | |
1495 | ||
1496 | if (test_bit(PAGE_HEADLESS, &page->private)) | |
1497 | goto out; | |
1498 | ||
1499 | buddy = handle_to_buddy(handle); | |
1500 | switch (buddy) { | |
1501 | case FIRST: | |
1502 | addr += ZHDR_SIZE_ALIGNED; | |
1503 | break; | |
1504 | case MIDDLE: | |
1505 | addr += zhdr->start_middle << CHUNK_SHIFT; | |
1506 | set_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
1507 | break; | |
1508 | case LAST: | |
ca0246bb | 1509 | addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT); |
9a001fc1 VW |
1510 | break; |
1511 | default: | |
1512 | pr_err("unknown buddy id %d\n", buddy); | |
1513 | WARN_ON(1); | |
1514 | addr = NULL; | |
1515 | break; | |
1516 | } | |
2f1e5e4d | 1517 | |
1f862989 VW |
1518 | if (addr) |
1519 | zhdr->mapped_count++; | |
9a001fc1 | 1520 | out: |
4a3ac931 | 1521 | put_z3fold_header(zhdr); |
9a001fc1 VW |
1522 | return addr; |
1523 | } | |
1524 | ||
1525 | /** | |
1526 | * z3fold_unmap() - unmaps the allocation associated with the given handle | |
1527 | * @pool: pool in which the allocation resides | |
1528 | * @handle: handle associated with the allocation to be unmapped | |
1529 | */ | |
1530 | static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle) | |
1531 | { | |
1532 | struct z3fold_header *zhdr; | |
1533 | struct page *page; | |
1534 | enum buddy buddy; | |
1535 | ||
4a3ac931 | 1536 | zhdr = get_z3fold_header(handle); |
9a001fc1 VW |
1537 | page = virt_to_page(zhdr); |
1538 | ||
2f1e5e4d | 1539 | if (test_bit(PAGE_HEADLESS, &page->private)) |
9a001fc1 | 1540 | return; |
9a001fc1 VW |
1541 | |
1542 | buddy = handle_to_buddy(handle); | |
1543 | if (buddy == MIDDLE) | |
1544 | clear_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
1f862989 | 1545 | zhdr->mapped_count--; |
4a3ac931 | 1546 | put_z3fold_header(zhdr); |
9a001fc1 VW |
1547 | } |
1548 | ||
1549 | /** | |
1550 | * z3fold_get_pool_size() - gets the z3fold pool size in pages | |
1551 | * @pool: pool whose size is being queried | |
1552 | * | |
12d59ae6 | 1553 | * Returns: size in pages of the given pool. |
9a001fc1 VW |
1554 | */ |
1555 | static u64 z3fold_get_pool_size(struct z3fold_pool *pool) | |
1556 | { | |
12d59ae6 | 1557 | return atomic64_read(&pool->pages_nr); |
9a001fc1 VW |
1558 | } |
1559 | ||
1f862989 VW |
1560 | static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode) |
1561 | { | |
1562 | struct z3fold_header *zhdr; | |
1563 | struct z3fold_pool *pool; | |
1564 | ||
1565 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1566 | VM_BUG_ON_PAGE(PageIsolated(page), page); | |
1567 | ||
dcf5aedb | 1568 | if (test_bit(PAGE_HEADLESS, &page->private)) |
1f862989 VW |
1569 | return false; |
1570 | ||
1571 | zhdr = page_address(page); | |
1572 | z3fold_page_lock(zhdr); | |
1573 | if (test_bit(NEEDS_COMPACTING, &page->private) || | |
1574 | test_bit(PAGE_STALE, &page->private)) | |
1575 | goto out; | |
1576 | ||
4a3ac931 VW |
1577 | if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) |
1578 | goto out; | |
1579 | ||
dcf5aedb VW |
1580 | if (test_and_set_bit(PAGE_CLAIMED, &page->private)) |
1581 | goto out; | |
1f862989 | 1582 | pool = zhdr_to_pool(zhdr); |
4a3ac931 VW |
1583 | spin_lock(&pool->lock); |
1584 | if (!list_empty(&zhdr->buddy)) | |
1585 | list_del_init(&zhdr->buddy); | |
1586 | if (!list_empty(&page->lru)) | |
1587 | list_del_init(&page->lru); | |
1588 | spin_unlock(&pool->lock); | |
1589 | ||
1590 | kref_get(&zhdr->refcount); | |
1591 | z3fold_page_unlock(zhdr); | |
1592 | return true; | |
1f862989 | 1593 | |
1f862989 VW |
1594 | out: |
1595 | z3fold_page_unlock(zhdr); | |
1596 | return false; | |
1597 | } | |
1598 | ||
1599 | static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage, | |
1600 | struct page *page, enum migrate_mode mode) | |
1601 | { | |
1602 | struct z3fold_header *zhdr, *new_zhdr; | |
1603 | struct z3fold_pool *pool; | |
1604 | struct address_space *new_mapping; | |
1605 | ||
1606 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1607 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
dcf5aedb | 1608 | VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page); |
810481a2 | 1609 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); |
1f862989 VW |
1610 | |
1611 | zhdr = page_address(page); | |
1612 | pool = zhdr_to_pool(zhdr); | |
1613 | ||
dcf5aedb | 1614 | if (!z3fold_page_trylock(zhdr)) |
1f862989 | 1615 | return -EAGAIN; |
4a3ac931 | 1616 | if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) { |
1f862989 | 1617 | z3fold_page_unlock(zhdr); |
dcf5aedb | 1618 | clear_bit(PAGE_CLAIMED, &page->private); |
1f862989 VW |
1619 | return -EBUSY; |
1620 | } | |
c92d2f38 HB |
1621 | if (work_pending(&zhdr->work)) { |
1622 | z3fold_page_unlock(zhdr); | |
1623 | return -EAGAIN; | |
1624 | } | |
1f862989 VW |
1625 | new_zhdr = page_address(newpage); |
1626 | memcpy(new_zhdr, zhdr, PAGE_SIZE); | |
1627 | newpage->private = page->private; | |
1628 | page->private = 0; | |
1629 | z3fold_page_unlock(zhdr); | |
1630 | spin_lock_init(&new_zhdr->page_lock); | |
c92d2f38 HB |
1631 | INIT_WORK(&new_zhdr->work, compact_page_work); |
1632 | /* | |
1633 | * z3fold_page_isolate() ensures that new_zhdr->buddy is empty, | |
1634 | * so we only have to reinitialize it. | |
1635 | */ | |
1636 | INIT_LIST_HEAD(&new_zhdr->buddy); | |
1f862989 VW |
1637 | new_mapping = page_mapping(page); |
1638 | __ClearPageMovable(page); | |
1639 | ClearPagePrivate(page); | |
1640 | ||
1641 | get_page(newpage); | |
1642 | z3fold_page_lock(new_zhdr); | |
1643 | if (new_zhdr->first_chunks) | |
1644 | encode_handle(new_zhdr, FIRST); | |
1645 | if (new_zhdr->last_chunks) | |
1646 | encode_handle(new_zhdr, LAST); | |
1647 | if (new_zhdr->middle_chunks) | |
1648 | encode_handle(new_zhdr, MIDDLE); | |
1649 | set_bit(NEEDS_COMPACTING, &newpage->private); | |
1650 | new_zhdr->cpu = smp_processor_id(); | |
1651 | spin_lock(&pool->lock); | |
1652 | list_add(&newpage->lru, &pool->lru); | |
1653 | spin_unlock(&pool->lock); | |
1654 | __SetPageMovable(newpage, new_mapping); | |
1655 | z3fold_page_unlock(new_zhdr); | |
1656 | ||
1657 | queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work); | |
1658 | ||
1659 | page_mapcount_reset(page); | |
dcf5aedb | 1660 | clear_bit(PAGE_CLAIMED, &page->private); |
1f862989 VW |
1661 | put_page(page); |
1662 | return 0; | |
1663 | } | |
1664 | ||
1665 | static void z3fold_page_putback(struct page *page) | |
1666 | { | |
1667 | struct z3fold_header *zhdr; | |
1668 | struct z3fold_pool *pool; | |
1669 | ||
1670 | zhdr = page_address(page); | |
1671 | pool = zhdr_to_pool(zhdr); | |
1672 | ||
1673 | z3fold_page_lock(zhdr); | |
1674 | if (!list_empty(&zhdr->buddy)) | |
1675 | list_del_init(&zhdr->buddy); | |
1676 | INIT_LIST_HEAD(&page->lru); | |
1677 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) { | |
1678 | atomic64_dec(&pool->pages_nr); | |
1679 | return; | |
1680 | } | |
1681 | spin_lock(&pool->lock); | |
1682 | list_add(&page->lru, &pool->lru); | |
1683 | spin_unlock(&pool->lock); | |
dcf5aedb | 1684 | clear_bit(PAGE_CLAIMED, &page->private); |
1f862989 VW |
1685 | z3fold_page_unlock(zhdr); |
1686 | } | |
1687 | ||
1688 | static const struct address_space_operations z3fold_aops = { | |
1689 | .isolate_page = z3fold_page_isolate, | |
1690 | .migratepage = z3fold_page_migrate, | |
1691 | .putback_page = z3fold_page_putback, | |
1692 | }; | |
1693 | ||
9a001fc1 VW |
1694 | /***************** |
1695 | * zpool | |
1696 | ****************/ | |
1697 | ||
1698 | static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle) | |
1699 | { | |
1700 | if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict) | |
1701 | return pool->zpool_ops->evict(pool->zpool, handle); | |
1702 | else | |
1703 | return -ENOENT; | |
1704 | } | |
1705 | ||
1706 | static const struct z3fold_ops z3fold_zpool_ops = { | |
1707 | .evict = z3fold_zpool_evict | |
1708 | }; | |
1709 | ||
1710 | static void *z3fold_zpool_create(const char *name, gfp_t gfp, | |
1711 | const struct zpool_ops *zpool_ops, | |
1712 | struct zpool *zpool) | |
1713 | { | |
1714 | struct z3fold_pool *pool; | |
1715 | ||
d30561c5 VW |
1716 | pool = z3fold_create_pool(name, gfp, |
1717 | zpool_ops ? &z3fold_zpool_ops : NULL); | |
9a001fc1 VW |
1718 | if (pool) { |
1719 | pool->zpool = zpool; | |
1720 | pool->zpool_ops = zpool_ops; | |
1721 | } | |
1722 | return pool; | |
1723 | } | |
1724 | ||
1725 | static void z3fold_zpool_destroy(void *pool) | |
1726 | { | |
1727 | z3fold_destroy_pool(pool); | |
1728 | } | |
1729 | ||
1730 | static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
1731 | unsigned long *handle) | |
1732 | { | |
1733 | return z3fold_alloc(pool, size, gfp, handle); | |
1734 | } | |
1735 | static void z3fold_zpool_free(void *pool, unsigned long handle) | |
1736 | { | |
1737 | z3fold_free(pool, handle); | |
1738 | } | |
1739 | ||
1740 | static int z3fold_zpool_shrink(void *pool, unsigned int pages, | |
1741 | unsigned int *reclaimed) | |
1742 | { | |
1743 | unsigned int total = 0; | |
1744 | int ret = -EINVAL; | |
1745 | ||
1746 | while (total < pages) { | |
1747 | ret = z3fold_reclaim_page(pool, 8); | |
1748 | if (ret < 0) | |
1749 | break; | |
1750 | total++; | |
1751 | } | |
1752 | ||
1753 | if (reclaimed) | |
1754 | *reclaimed = total; | |
1755 | ||
1756 | return ret; | |
1757 | } | |
1758 | ||
1759 | static void *z3fold_zpool_map(void *pool, unsigned long handle, | |
1760 | enum zpool_mapmode mm) | |
1761 | { | |
1762 | return z3fold_map(pool, handle); | |
1763 | } | |
1764 | static void z3fold_zpool_unmap(void *pool, unsigned long handle) | |
1765 | { | |
1766 | z3fold_unmap(pool, handle); | |
1767 | } | |
1768 | ||
1769 | static u64 z3fold_zpool_total_size(void *pool) | |
1770 | { | |
1771 | return z3fold_get_pool_size(pool) * PAGE_SIZE; | |
1772 | } | |
1773 | ||
1774 | static struct zpool_driver z3fold_zpool_driver = { | |
1775 | .type = "z3fold", | |
e818e820 | 1776 | .sleep_mapped = true, |
9a001fc1 VW |
1777 | .owner = THIS_MODULE, |
1778 | .create = z3fold_zpool_create, | |
1779 | .destroy = z3fold_zpool_destroy, | |
1780 | .malloc = z3fold_zpool_malloc, | |
1781 | .free = z3fold_zpool_free, | |
1782 | .shrink = z3fold_zpool_shrink, | |
1783 | .map = z3fold_zpool_map, | |
1784 | .unmap = z3fold_zpool_unmap, | |
1785 | .total_size = z3fold_zpool_total_size, | |
1786 | }; | |
1787 | ||
1788 | MODULE_ALIAS("zpool-z3fold"); | |
1789 | ||
1790 | static int __init init_z3fold(void) | |
1791 | { | |
1f862989 VW |
1792 | int ret; |
1793 | ||
014284a0 ML |
1794 | /* |
1795 | * Make sure the z3fold header is not larger than the page size and | |
1796 | * there has remaining spaces for its buddy. | |
1797 | */ | |
1798 | BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE); | |
1f862989 VW |
1799 | ret = z3fold_mount(); |
1800 | if (ret) | |
1801 | return ret; | |
1802 | ||
9a001fc1 VW |
1803 | zpool_register_driver(&z3fold_zpool_driver); |
1804 | ||
1805 | return 0; | |
1806 | } | |
1807 | ||
1808 | static void __exit exit_z3fold(void) | |
1809 | { | |
1f862989 | 1810 | z3fold_unmount(); |
9a001fc1 VW |
1811 | zpool_unregister_driver(&z3fold_zpool_driver); |
1812 | } | |
1813 | ||
1814 | module_init(init_z3fold); | |
1815 | module_exit(exit_z3fold); | |
1816 | ||
1817 | MODULE_LICENSE("GPL"); | |
1818 | MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>"); | |
1819 | MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages"); |