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. | |
30522175 MG |
147 | * @zpool: zpool driver |
148 | * @zpool_ops: zpool operations structure with an evict callback | |
d30561c5 VW |
149 | * @compact_wq: workqueue for page layout background optimization |
150 | * @release_wq: workqueue for safe page release | |
151 | * @work: work_struct for safe page release | |
1f862989 | 152 | * @inode: inode for z3fold pseudo filesystem |
9a001fc1 VW |
153 | * |
154 | * This structure is allocated at pool creation time and maintains metadata | |
155 | * pertaining to a particular z3fold pool. | |
156 | */ | |
157 | struct z3fold_pool { | |
d30561c5 | 158 | const char *name; |
9a001fc1 | 159 | spinlock_t lock; |
d30561c5 VW |
160 | spinlock_t stale_lock; |
161 | struct list_head *unbuddied; | |
9a001fc1 | 162 | struct list_head lru; |
d30561c5 | 163 | struct list_head stale; |
12d59ae6 | 164 | atomic64_t pages_nr; |
7c2b8baa | 165 | struct kmem_cache *c_handle; |
9a001fc1 VW |
166 | const struct z3fold_ops *ops; |
167 | struct zpool *zpool; | |
168 | const struct zpool_ops *zpool_ops; | |
d30561c5 VW |
169 | struct workqueue_struct *compact_wq; |
170 | struct workqueue_struct *release_wq; | |
171 | struct work_struct work; | |
1f862989 | 172 | struct inode *inode; |
9a001fc1 VW |
173 | }; |
174 | ||
9a001fc1 VW |
175 | /* |
176 | * Internal z3fold page flags | |
177 | */ | |
178 | enum z3fold_page_flags { | |
5a27aa82 | 179 | PAGE_HEADLESS = 0, |
9a001fc1 | 180 | MIDDLE_CHUNK_MAPPED, |
d30561c5 | 181 | NEEDS_COMPACTING, |
6098d7e1 | 182 | PAGE_STALE, |
ca0246bb | 183 | PAGE_CLAIMED, /* by either reclaim or free */ |
9a001fc1 VW |
184 | }; |
185 | ||
dcf5aedb VW |
186 | /* |
187 | * handle flags, go under HANDLE_FLAG_MASK | |
188 | */ | |
189 | enum z3fold_handle_flags { | |
190 | HANDLES_NOFREE = 0, | |
191 | }; | |
192 | ||
4a3ac931 VW |
193 | /* |
194 | * Forward declarations | |
195 | */ | |
196 | static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool); | |
197 | static void compact_page_work(struct work_struct *w); | |
198 | ||
9a001fc1 VW |
199 | /***************** |
200 | * Helpers | |
201 | *****************/ | |
202 | ||
203 | /* Converts an allocation size in bytes to size in z3fold chunks */ | |
204 | static int size_to_chunks(size_t size) | |
205 | { | |
206 | return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT; | |
207 | } | |
208 | ||
209 | #define for_each_unbuddied_list(_iter, _begin) \ | |
210 | for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++) | |
211 | ||
bb9f6f63 VW |
212 | static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool, |
213 | gfp_t gfp) | |
7c2b8baa | 214 | { |
f1549cb5 HB |
215 | struct z3fold_buddy_slots *slots; |
216 | ||
f94afee9 | 217 | slots = kmem_cache_zalloc(pool->c_handle, |
f1549cb5 | 218 | (gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE))); |
7c2b8baa VW |
219 | |
220 | if (slots) { | |
af4798a5 QC |
221 | /* It will be freed separately in free_handle(). */ |
222 | kmemleak_not_leak(slots); | |
7c2b8baa | 223 | slots->pool = (unsigned long)pool; |
4a3ac931 | 224 | rwlock_init(&slots->lock); |
7c2b8baa VW |
225 | } |
226 | ||
227 | return slots; | |
228 | } | |
229 | ||
230 | static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s) | |
231 | { | |
232 | return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK); | |
233 | } | |
234 | ||
235 | static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle) | |
236 | { | |
237 | return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1)); | |
238 | } | |
239 | ||
4a3ac931 VW |
240 | /* Lock a z3fold page */ |
241 | static inline void z3fold_page_lock(struct z3fold_header *zhdr) | |
242 | { | |
243 | spin_lock(&zhdr->page_lock); | |
244 | } | |
245 | ||
246 | /* Try to lock a z3fold page */ | |
247 | static inline int z3fold_page_trylock(struct z3fold_header *zhdr) | |
248 | { | |
249 | return spin_trylock(&zhdr->page_lock); | |
250 | } | |
251 | ||
252 | /* Unlock a z3fold page */ | |
253 | static inline void z3fold_page_unlock(struct z3fold_header *zhdr) | |
254 | { | |
255 | spin_unlock(&zhdr->page_lock); | |
256 | } | |
257 | ||
767cc6c5 ML |
258 | /* return locked z3fold page if it's not headless */ |
259 | static inline struct z3fold_header *get_z3fold_header(unsigned long handle) | |
4a3ac931 VW |
260 | { |
261 | struct z3fold_buddy_slots *slots; | |
262 | struct z3fold_header *zhdr; | |
263 | int locked = 0; | |
264 | ||
265 | if (!(handle & (1 << PAGE_HEADLESS))) { | |
266 | slots = handle_to_slots(handle); | |
267 | do { | |
268 | unsigned long addr; | |
269 | ||
270 | read_lock(&slots->lock); | |
271 | addr = *(unsigned long *)handle; | |
272 | zhdr = (struct z3fold_header *)(addr & PAGE_MASK); | |
767cc6c5 | 273 | locked = z3fold_page_trylock(zhdr); |
4a3ac931 VW |
274 | read_unlock(&slots->lock); |
275 | if (locked) | |
276 | break; | |
277 | cpu_relax(); | |
767cc6c5 | 278 | } while (true); |
4a3ac931 VW |
279 | } else { |
280 | zhdr = (struct z3fold_header *)(handle & PAGE_MASK); | |
281 | } | |
282 | ||
283 | return zhdr; | |
284 | } | |
285 | ||
4a3ac931 VW |
286 | static inline void put_z3fold_header(struct z3fold_header *zhdr) |
287 | { | |
288 | struct page *page = virt_to_page(zhdr); | |
289 | ||
290 | if (!test_bit(PAGE_HEADLESS, &page->private)) | |
291 | z3fold_page_unlock(zhdr); | |
292 | } | |
293 | ||
fc548865 | 294 | static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr) |
7c2b8baa VW |
295 | { |
296 | struct z3fold_buddy_slots *slots; | |
297 | int i; | |
298 | bool is_free; | |
299 | ||
300 | if (handle & (1 << PAGE_HEADLESS)) | |
301 | return; | |
302 | ||
4a3ac931 VW |
303 | if (WARN_ON(*(unsigned long *)handle == 0)) |
304 | return; | |
305 | ||
7c2b8baa | 306 | slots = handle_to_slots(handle); |
4a3ac931 VW |
307 | write_lock(&slots->lock); |
308 | *(unsigned long *)handle = 0; | |
dcf5aedb VW |
309 | |
310 | if (test_bit(HANDLES_NOFREE, &slots->pool)) { | |
311 | write_unlock(&slots->lock); | |
312 | return; /* simple case, nothing else to do */ | |
313 | } | |
314 | ||
fc548865 VW |
315 | if (zhdr->slots != slots) |
316 | zhdr->foreign_handles--; | |
4a3ac931 | 317 | |
7c2b8baa VW |
318 | is_free = true; |
319 | for (i = 0; i <= BUDDY_MASK; i++) { | |
320 | if (slots->slot[i]) { | |
321 | is_free = false; | |
322 | break; | |
323 | } | |
324 | } | |
d8f117ab | 325 | write_unlock(&slots->lock); |
7c2b8baa VW |
326 | |
327 | if (is_free) { | |
328 | struct z3fold_pool *pool = slots_to_pool(slots); | |
329 | ||
fc548865 VW |
330 | if (zhdr->slots == slots) |
331 | zhdr->slots = NULL; | |
7c2b8baa VW |
332 | kmem_cache_free(pool->c_handle, slots); |
333 | } | |
334 | } | |
335 | ||
ea8157ab | 336 | static int z3fold_init_fs_context(struct fs_context *fc) |
1f862989 | 337 | { |
ea8157ab | 338 | return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM; |
1f862989 VW |
339 | } |
340 | ||
341 | static struct file_system_type z3fold_fs = { | |
342 | .name = "z3fold", | |
ea8157ab | 343 | .init_fs_context = z3fold_init_fs_context, |
1f862989 VW |
344 | .kill_sb = kill_anon_super, |
345 | }; | |
346 | ||
347 | static struct vfsmount *z3fold_mnt; | |
dc3a1f30 | 348 | static int __init z3fold_mount(void) |
1f862989 VW |
349 | { |
350 | int ret = 0; | |
351 | ||
352 | z3fold_mnt = kern_mount(&z3fold_fs); | |
353 | if (IS_ERR(z3fold_mnt)) | |
354 | ret = PTR_ERR(z3fold_mnt); | |
355 | ||
356 | return ret; | |
357 | } | |
358 | ||
359 | static void z3fold_unmount(void) | |
360 | { | |
361 | kern_unmount(z3fold_mnt); | |
362 | } | |
363 | ||
364 | static const struct address_space_operations z3fold_aops; | |
365 | static int z3fold_register_migration(struct z3fold_pool *pool) | |
366 | { | |
367 | pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb); | |
368 | if (IS_ERR(pool->inode)) { | |
369 | pool->inode = NULL; | |
370 | return 1; | |
371 | } | |
372 | ||
373 | pool->inode->i_mapping->private_data = pool; | |
374 | pool->inode->i_mapping->a_ops = &z3fold_aops; | |
375 | return 0; | |
376 | } | |
377 | ||
378 | static void z3fold_unregister_migration(struct z3fold_pool *pool) | |
379 | { | |
380 | if (pool->inode) | |
381 | iput(pool->inode); | |
cb152a1a | 382 | } |
1f862989 | 383 | |
9a001fc1 | 384 | /* Initializes the z3fold header of a newly allocated z3fold page */ |
63398413 | 385 | static struct z3fold_header *init_z3fold_page(struct page *page, bool headless, |
bb9f6f63 | 386 | struct z3fold_pool *pool, gfp_t gfp) |
9a001fc1 VW |
387 | { |
388 | struct z3fold_header *zhdr = page_address(page); | |
63398413 | 389 | struct z3fold_buddy_slots *slots; |
9a001fc1 VW |
390 | |
391 | INIT_LIST_HEAD(&page->lru); | |
9a001fc1 VW |
392 | clear_bit(PAGE_HEADLESS, &page->private); |
393 | clear_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
d30561c5 VW |
394 | clear_bit(NEEDS_COMPACTING, &page->private); |
395 | clear_bit(PAGE_STALE, &page->private); | |
ca0246bb | 396 | clear_bit(PAGE_CLAIMED, &page->private); |
63398413 VW |
397 | if (headless) |
398 | return zhdr; | |
399 | ||
400 | slots = alloc_slots(pool, gfp); | |
401 | if (!slots) | |
402 | return NULL; | |
9a001fc1 | 403 | |
c457cd96 | 404 | memset(zhdr, 0, sizeof(*zhdr)); |
2f1e5e4d | 405 | spin_lock_init(&zhdr->page_lock); |
5a27aa82 | 406 | kref_init(&zhdr->refcount); |
d30561c5 | 407 | zhdr->cpu = -1; |
7c2b8baa | 408 | zhdr->slots = slots; |
bb9a374d | 409 | zhdr->pool = pool; |
9a001fc1 | 410 | INIT_LIST_HEAD(&zhdr->buddy); |
d30561c5 | 411 | INIT_WORK(&zhdr->work, compact_page_work); |
9a001fc1 VW |
412 | return zhdr; |
413 | } | |
414 | ||
415 | /* Resets the struct page fields and frees the page */ | |
1f862989 | 416 | static void free_z3fold_page(struct page *page, bool headless) |
9a001fc1 | 417 | { |
1f862989 VW |
418 | if (!headless) { |
419 | lock_page(page); | |
420 | __ClearPageMovable(page); | |
421 | unlock_page(page); | |
422 | } | |
5a27aa82 VW |
423 | __free_page(page); |
424 | } | |
425 | ||
7c2b8baa VW |
426 | /* Helper function to build the index */ |
427 | static inline int __idx(struct z3fold_header *zhdr, enum buddy bud) | |
428 | { | |
429 | return (bud + zhdr->first_num) & BUDDY_MASK; | |
430 | } | |
431 | ||
9a001fc1 VW |
432 | /* |
433 | * Encodes the handle of a particular buddy within a z3fold page | |
434 | * Pool lock should be held as this function accesses first_num | |
435 | */ | |
3f9d2b57 VW |
436 | static unsigned long __encode_handle(struct z3fold_header *zhdr, |
437 | struct z3fold_buddy_slots *slots, | |
438 | enum buddy bud) | |
9a001fc1 | 439 | { |
7c2b8baa VW |
440 | unsigned long h = (unsigned long)zhdr; |
441 | int idx = 0; | |
9a001fc1 | 442 | |
7c2b8baa VW |
443 | /* |
444 | * For a headless page, its handle is its pointer with the extra | |
445 | * PAGE_HEADLESS bit set | |
446 | */ | |
447 | if (bud == HEADLESS) | |
448 | return h | (1 << PAGE_HEADLESS); | |
449 | ||
450 | /* otherwise, return pointer to encoded handle */ | |
451 | idx = __idx(zhdr, bud); | |
452 | h += idx; | |
453 | if (bud == LAST) | |
454 | h |= (zhdr->last_chunks << BUDDY_SHIFT); | |
455 | ||
4a3ac931 | 456 | write_lock(&slots->lock); |
7c2b8baa | 457 | slots->slot[idx] = h; |
4a3ac931 | 458 | write_unlock(&slots->lock); |
7c2b8baa | 459 | return (unsigned long)&slots->slot[idx]; |
9a001fc1 VW |
460 | } |
461 | ||
3f9d2b57 VW |
462 | static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud) |
463 | { | |
464 | return __encode_handle(zhdr, zhdr->slots, bud); | |
465 | } | |
466 | ||
ca0246bb VW |
467 | /* only for LAST bud, returns zero otherwise */ |
468 | static unsigned short handle_to_chunks(unsigned long handle) | |
469 | { | |
4a3ac931 VW |
470 | struct z3fold_buddy_slots *slots = handle_to_slots(handle); |
471 | unsigned long addr; | |
7c2b8baa | 472 | |
4a3ac931 VW |
473 | read_lock(&slots->lock); |
474 | addr = *(unsigned long *)handle; | |
475 | read_unlock(&slots->lock); | |
7c2b8baa | 476 | return (addr & ~PAGE_MASK) >> BUDDY_SHIFT; |
ca0246bb VW |
477 | } |
478 | ||
f201ebd8 | 479 | /* |
480 | * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle | |
481 | * but that doesn't matter. because the masking will result in the | |
482 | * correct buddy number. | |
483 | */ | |
9a001fc1 VW |
484 | static enum buddy handle_to_buddy(unsigned long handle) |
485 | { | |
7c2b8baa | 486 | struct z3fold_header *zhdr; |
4a3ac931 | 487 | struct z3fold_buddy_slots *slots = handle_to_slots(handle); |
7c2b8baa VW |
488 | unsigned long addr; |
489 | ||
4a3ac931 | 490 | read_lock(&slots->lock); |
7c2b8baa VW |
491 | WARN_ON(handle & (1 << PAGE_HEADLESS)); |
492 | addr = *(unsigned long *)handle; | |
4a3ac931 | 493 | read_unlock(&slots->lock); |
7c2b8baa VW |
494 | zhdr = (struct z3fold_header *)(addr & PAGE_MASK); |
495 | return (addr - zhdr->first_num) & BUDDY_MASK; | |
9a001fc1 VW |
496 | } |
497 | ||
9050cce1 VW |
498 | static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr) |
499 | { | |
bb9a374d | 500 | return zhdr->pool; |
9050cce1 VW |
501 | } |
502 | ||
d30561c5 VW |
503 | static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked) |
504 | { | |
505 | struct page *page = virt_to_page(zhdr); | |
9050cce1 | 506 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
d30561c5 VW |
507 | |
508 | WARN_ON(!list_empty(&zhdr->buddy)); | |
509 | set_bit(PAGE_STALE, &page->private); | |
35529357 | 510 | clear_bit(NEEDS_COMPACTING, &page->private); |
d30561c5 VW |
511 | spin_lock(&pool->lock); |
512 | if (!list_empty(&page->lru)) | |
1f862989 | 513 | list_del_init(&page->lru); |
d30561c5 | 514 | spin_unlock(&pool->lock); |
4a3ac931 | 515 | |
d30561c5 VW |
516 | if (locked) |
517 | z3fold_page_unlock(zhdr); | |
4a3ac931 | 518 | |
d30561c5 VW |
519 | spin_lock(&pool->stale_lock); |
520 | list_add(&zhdr->buddy, &pool->stale); | |
521 | queue_work(pool->release_wq, &pool->work); | |
522 | spin_unlock(&pool->stale_lock); | |
523 | } | |
524 | ||
70ad3196 | 525 | static void release_z3fold_page(struct kref *ref) |
d30561c5 VW |
526 | { |
527 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
528 | refcount); | |
529 | __release_z3fold_page(zhdr, false); | |
530 | } | |
531 | ||
532 | static void release_z3fold_page_locked(struct kref *ref) | |
533 | { | |
534 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
535 | refcount); | |
536 | WARN_ON(z3fold_page_trylock(zhdr)); | |
537 | __release_z3fold_page(zhdr, true); | |
538 | } | |
539 | ||
540 | static void release_z3fold_page_locked_list(struct kref *ref) | |
541 | { | |
542 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
543 | refcount); | |
9050cce1 | 544 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
4a3ac931 | 545 | |
9050cce1 | 546 | spin_lock(&pool->lock); |
d30561c5 | 547 | list_del_init(&zhdr->buddy); |
9050cce1 | 548 | spin_unlock(&pool->lock); |
d30561c5 VW |
549 | |
550 | WARN_ON(z3fold_page_trylock(zhdr)); | |
551 | __release_z3fold_page(zhdr, true); | |
552 | } | |
553 | ||
554 | static void free_pages_work(struct work_struct *w) | |
555 | { | |
556 | struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work); | |
557 | ||
558 | spin_lock(&pool->stale_lock); | |
559 | while (!list_empty(&pool->stale)) { | |
560 | struct z3fold_header *zhdr = list_first_entry(&pool->stale, | |
561 | struct z3fold_header, buddy); | |
562 | struct page *page = virt_to_page(zhdr); | |
563 | ||
564 | list_del(&zhdr->buddy); | |
565 | if (WARN_ON(!test_bit(PAGE_STALE, &page->private))) | |
566 | continue; | |
d30561c5 VW |
567 | spin_unlock(&pool->stale_lock); |
568 | cancel_work_sync(&zhdr->work); | |
1f862989 | 569 | free_z3fold_page(page, false); |
d30561c5 VW |
570 | cond_resched(); |
571 | spin_lock(&pool->stale_lock); | |
572 | } | |
573 | spin_unlock(&pool->stale_lock); | |
574 | } | |
575 | ||
9a001fc1 VW |
576 | /* |
577 | * Returns the number of free chunks in a z3fold page. | |
578 | * NB: can't be used with HEADLESS pages. | |
579 | */ | |
580 | static int num_free_chunks(struct z3fold_header *zhdr) | |
581 | { | |
582 | int nfree; | |
583 | /* | |
584 | * If there is a middle object, pick up the bigger free space | |
585 | * either before or after it. Otherwise just subtract the number | |
586 | * of chunks occupied by the first and the last objects. | |
587 | */ | |
588 | if (zhdr->middle_chunks != 0) { | |
589 | int nfree_before = zhdr->first_chunks ? | |
ede93213 | 590 | 0 : zhdr->start_middle - ZHDR_CHUNKS; |
9a001fc1 | 591 | int nfree_after = zhdr->last_chunks ? |
ede93213 VW |
592 | 0 : TOTAL_CHUNKS - |
593 | (zhdr->start_middle + zhdr->middle_chunks); | |
9a001fc1 VW |
594 | nfree = max(nfree_before, nfree_after); |
595 | } else | |
596 | nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks; | |
597 | return nfree; | |
598 | } | |
599 | ||
9050cce1 VW |
600 | /* Add to the appropriate unbuddied list */ |
601 | static inline void add_to_unbuddied(struct z3fold_pool *pool, | |
602 | struct z3fold_header *zhdr) | |
603 | { | |
604 | if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 || | |
605 | zhdr->middle_chunks == 0) { | |
135f97fd | 606 | struct list_head *unbuddied; |
9050cce1 | 607 | int freechunks = num_free_chunks(zhdr); |
135f97fd VW |
608 | |
609 | migrate_disable(); | |
610 | unbuddied = this_cpu_ptr(pool->unbuddied); | |
9050cce1 VW |
611 | spin_lock(&pool->lock); |
612 | list_add(&zhdr->buddy, &unbuddied[freechunks]); | |
613 | spin_unlock(&pool->lock); | |
614 | zhdr->cpu = smp_processor_id(); | |
135f97fd | 615 | migrate_enable(); |
9050cce1 VW |
616 | } |
617 | } | |
618 | ||
dcf5aedb VW |
619 | static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks) |
620 | { | |
621 | enum buddy bud = HEADLESS; | |
622 | ||
623 | if (zhdr->middle_chunks) { | |
624 | if (!zhdr->first_chunks && | |
625 | chunks <= zhdr->start_middle - ZHDR_CHUNKS) | |
626 | bud = FIRST; | |
627 | else if (!zhdr->last_chunks) | |
628 | bud = LAST; | |
629 | } else { | |
630 | if (!zhdr->first_chunks) | |
631 | bud = FIRST; | |
632 | else if (!zhdr->last_chunks) | |
633 | bud = LAST; | |
634 | else | |
635 | bud = MIDDLE; | |
636 | } | |
637 | ||
638 | return bud; | |
639 | } | |
640 | ||
ede93213 VW |
641 | static inline void *mchunk_memmove(struct z3fold_header *zhdr, |
642 | unsigned short dst_chunk) | |
643 | { | |
644 | void *beg = zhdr; | |
645 | return memmove(beg + (dst_chunk << CHUNK_SHIFT), | |
646 | beg + (zhdr->start_middle << CHUNK_SHIFT), | |
647 | zhdr->middle_chunks << CHUNK_SHIFT); | |
648 | } | |
649 | ||
4a3ac931 VW |
650 | static inline bool buddy_single(struct z3fold_header *zhdr) |
651 | { | |
652 | return !((zhdr->first_chunks && zhdr->middle_chunks) || | |
653 | (zhdr->first_chunks && zhdr->last_chunks) || | |
654 | (zhdr->middle_chunks && zhdr->last_chunks)); | |
655 | } | |
656 | ||
657 | static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr) | |
658 | { | |
659 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); | |
660 | void *p = zhdr; | |
661 | unsigned long old_handle = 0; | |
662 | size_t sz = 0; | |
663 | struct z3fold_header *new_zhdr = NULL; | |
664 | int first_idx = __idx(zhdr, FIRST); | |
665 | int middle_idx = __idx(zhdr, MIDDLE); | |
666 | int last_idx = __idx(zhdr, LAST); | |
667 | unsigned short *moved_chunks = NULL; | |
668 | ||
669 | /* | |
670 | * No need to protect slots here -- all the slots are "local" and | |
671 | * the page lock is already taken | |
672 | */ | |
673 | if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) { | |
674 | p += ZHDR_SIZE_ALIGNED; | |
675 | sz = zhdr->first_chunks << CHUNK_SHIFT; | |
676 | old_handle = (unsigned long)&zhdr->slots->slot[first_idx]; | |
677 | moved_chunks = &zhdr->first_chunks; | |
678 | } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) { | |
679 | p += zhdr->start_middle << CHUNK_SHIFT; | |
680 | sz = zhdr->middle_chunks << CHUNK_SHIFT; | |
681 | old_handle = (unsigned long)&zhdr->slots->slot[middle_idx]; | |
682 | moved_chunks = &zhdr->middle_chunks; | |
683 | } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) { | |
684 | p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT); | |
685 | sz = zhdr->last_chunks << CHUNK_SHIFT; | |
686 | old_handle = (unsigned long)&zhdr->slots->slot[last_idx]; | |
687 | moved_chunks = &zhdr->last_chunks; | |
688 | } | |
689 | ||
690 | if (sz > 0) { | |
691 | enum buddy new_bud = HEADLESS; | |
692 | short chunks = size_to_chunks(sz); | |
693 | void *q; | |
694 | ||
695 | new_zhdr = __z3fold_alloc(pool, sz, false); | |
696 | if (!new_zhdr) | |
697 | return NULL; | |
698 | ||
699 | if (WARN_ON(new_zhdr == zhdr)) | |
700 | goto out_fail; | |
701 | ||
dcf5aedb | 702 | new_bud = get_free_buddy(new_zhdr, chunks); |
4a3ac931 VW |
703 | q = new_zhdr; |
704 | switch (new_bud) { | |
705 | case FIRST: | |
706 | new_zhdr->first_chunks = chunks; | |
707 | q += ZHDR_SIZE_ALIGNED; | |
708 | break; | |
709 | case MIDDLE: | |
710 | new_zhdr->middle_chunks = chunks; | |
711 | new_zhdr->start_middle = | |
712 | new_zhdr->first_chunks + ZHDR_CHUNKS; | |
713 | q += new_zhdr->start_middle << CHUNK_SHIFT; | |
714 | break; | |
715 | case LAST: | |
716 | new_zhdr->last_chunks = chunks; | |
717 | q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT); | |
718 | break; | |
719 | default: | |
720 | goto out_fail; | |
721 | } | |
722 | new_zhdr->foreign_handles++; | |
723 | memcpy(q, p, sz); | |
724 | write_lock(&zhdr->slots->lock); | |
725 | *(unsigned long *)old_handle = (unsigned long)new_zhdr + | |
726 | __idx(new_zhdr, new_bud); | |
727 | if (new_bud == LAST) | |
728 | *(unsigned long *)old_handle |= | |
729 | (new_zhdr->last_chunks << BUDDY_SHIFT); | |
730 | write_unlock(&zhdr->slots->lock); | |
731 | add_to_unbuddied(pool, new_zhdr); | |
732 | z3fold_page_unlock(new_zhdr); | |
733 | ||
734 | *moved_chunks = 0; | |
735 | } | |
736 | ||
737 | return new_zhdr; | |
738 | ||
739 | out_fail: | |
740 | if (new_zhdr) { | |
741 | if (kref_put(&new_zhdr->refcount, release_z3fold_page_locked)) | |
742 | atomic64_dec(&pool->pages_nr); | |
743 | else { | |
744 | add_to_unbuddied(pool, new_zhdr); | |
745 | z3fold_page_unlock(new_zhdr); | |
746 | } | |
747 | } | |
748 | return NULL; | |
749 | ||
750 | } | |
751 | ||
1b096e5a | 752 | #define BIG_CHUNK_GAP 3 |
9a001fc1 VW |
753 | /* Has to be called with lock held */ |
754 | static int z3fold_compact_page(struct z3fold_header *zhdr) | |
755 | { | |
756 | struct page *page = virt_to_page(zhdr); | |
9a001fc1 | 757 | |
ede93213 VW |
758 | if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private)) |
759 | return 0; /* can't move middle chunk, it's used */ | |
9a001fc1 | 760 | |
1f862989 VW |
761 | if (unlikely(PageIsolated(page))) |
762 | return 0; | |
763 | ||
ede93213 VW |
764 | if (zhdr->middle_chunks == 0) |
765 | return 0; /* nothing to compact */ | |
766 | ||
767 | if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) { | |
768 | /* move to the beginning */ | |
769 | mchunk_memmove(zhdr, ZHDR_CHUNKS); | |
9a001fc1 VW |
770 | zhdr->first_chunks = zhdr->middle_chunks; |
771 | zhdr->middle_chunks = 0; | |
772 | zhdr->start_middle = 0; | |
773 | zhdr->first_num++; | |
1b096e5a | 774 | return 1; |
9a001fc1 | 775 | } |
1b096e5a VW |
776 | |
777 | /* | |
778 | * moving data is expensive, so let's only do that if | |
779 | * there's substantial gain (at least BIG_CHUNK_GAP chunks) | |
780 | */ | |
781 | if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 && | |
782 | zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >= | |
783 | BIG_CHUNK_GAP) { | |
784 | mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS); | |
785 | zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS; | |
786 | return 1; | |
787 | } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 && | |
788 | TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle | |
789 | + zhdr->middle_chunks) >= | |
790 | BIG_CHUNK_GAP) { | |
791 | unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks - | |
792 | zhdr->middle_chunks; | |
793 | mchunk_memmove(zhdr, new_start); | |
794 | zhdr->start_middle = new_start; | |
795 | return 1; | |
796 | } | |
797 | ||
798 | return 0; | |
9a001fc1 VW |
799 | } |
800 | ||
d30561c5 VW |
801 | static void do_compact_page(struct z3fold_header *zhdr, bool locked) |
802 | { | |
9050cce1 | 803 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
d30561c5 | 804 | struct page *page; |
d30561c5 VW |
805 | |
806 | page = virt_to_page(zhdr); | |
807 | if (locked) | |
808 | WARN_ON(z3fold_page_trylock(zhdr)); | |
809 | else | |
810 | z3fold_page_lock(zhdr); | |
5d03a661 | 811 | if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) { |
d30561c5 VW |
812 | z3fold_page_unlock(zhdr); |
813 | return; | |
814 | } | |
815 | spin_lock(&pool->lock); | |
816 | list_del_init(&zhdr->buddy); | |
817 | spin_unlock(&pool->lock); | |
818 | ||
5d03a661 VW |
819 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) { |
820 | atomic64_dec(&pool->pages_nr); | |
821 | return; | |
822 | } | |
823 | ||
dcf5aedb VW |
824 | if (test_bit(PAGE_STALE, &page->private) || |
825 | test_and_set_bit(PAGE_CLAIMED, &page->private)) { | |
1f862989 VW |
826 | z3fold_page_unlock(zhdr); |
827 | return; | |
828 | } | |
829 | ||
4a3ac931 VW |
830 | if (!zhdr->foreign_handles && buddy_single(zhdr) && |
831 | zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) { | |
832 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) | |
833 | atomic64_dec(&pool->pages_nr); | |
dcf5aedb VW |
834 | else { |
835 | clear_bit(PAGE_CLAIMED, &page->private); | |
4a3ac931 | 836 | z3fold_page_unlock(zhdr); |
dcf5aedb | 837 | } |
4a3ac931 VW |
838 | return; |
839 | } | |
840 | ||
d30561c5 | 841 | z3fold_compact_page(zhdr); |
9050cce1 | 842 | add_to_unbuddied(pool, zhdr); |
dcf5aedb | 843 | clear_bit(PAGE_CLAIMED, &page->private); |
d30561c5 VW |
844 | z3fold_page_unlock(zhdr); |
845 | } | |
846 | ||
847 | static void compact_page_work(struct work_struct *w) | |
848 | { | |
849 | struct z3fold_header *zhdr = container_of(w, struct z3fold_header, | |
850 | work); | |
851 | ||
852 | do_compact_page(zhdr, false); | |
853 | } | |
854 | ||
9050cce1 VW |
855 | /* returns _locked_ z3fold page header or NULL */ |
856 | static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool, | |
857 | size_t size, bool can_sleep) | |
858 | { | |
859 | struct z3fold_header *zhdr = NULL; | |
860 | struct page *page; | |
861 | struct list_head *unbuddied; | |
862 | int chunks = size_to_chunks(size), i; | |
863 | ||
864 | lookup: | |
135f97fd | 865 | migrate_disable(); |
9050cce1 | 866 | /* First, try to find an unbuddied z3fold page. */ |
135f97fd | 867 | unbuddied = this_cpu_ptr(pool->unbuddied); |
9050cce1 VW |
868 | for_each_unbuddied_list(i, chunks) { |
869 | struct list_head *l = &unbuddied[i]; | |
870 | ||
871 | zhdr = list_first_entry_or_null(READ_ONCE(l), | |
872 | struct z3fold_header, buddy); | |
873 | ||
874 | if (!zhdr) | |
875 | continue; | |
876 | ||
877 | /* Re-check under lock. */ | |
878 | spin_lock(&pool->lock); | |
9050cce1 VW |
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 | * | |
9a001fc1 VW |
1065 | * Return: 0 if success and handle is set, otherwise -EINVAL if the size or |
1066 | * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate | |
1067 | * a new page. | |
1068 | */ | |
1069 | static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp, | |
1070 | unsigned long *handle) | |
1071 | { | |
9050cce1 | 1072 | int chunks = size_to_chunks(size); |
9a001fc1 | 1073 | struct z3fold_header *zhdr = NULL; |
d30561c5 | 1074 | struct page *page = NULL; |
9a001fc1 | 1075 | enum buddy bud; |
8a97ea54 | 1076 | bool can_sleep = gfpflags_allow_blocking(gfp); |
9a001fc1 | 1077 | |
f1549cb5 | 1078 | if (!size) |
9a001fc1 VW |
1079 | return -EINVAL; |
1080 | ||
1081 | if (size > PAGE_SIZE) | |
1082 | return -ENOSPC; | |
1083 | ||
1084 | if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE) | |
1085 | bud = HEADLESS; | |
1086 | else { | |
9050cce1 VW |
1087 | retry: |
1088 | zhdr = __z3fold_alloc(pool, size, can_sleep); | |
d30561c5 | 1089 | if (zhdr) { |
dcf5aedb VW |
1090 | bud = get_free_buddy(zhdr, chunks); |
1091 | if (bud == HEADLESS) { | |
5a27aa82 | 1092 | if (kref_put(&zhdr->refcount, |
d30561c5 | 1093 | release_z3fold_page_locked)) |
5a27aa82 | 1094 | atomic64_dec(&pool->pages_nr); |
d30561c5 VW |
1095 | else |
1096 | z3fold_page_unlock(zhdr); | |
2f1e5e4d VW |
1097 | pr_err("No free chunks in unbuddied\n"); |
1098 | WARN_ON(1); | |
9050cce1 | 1099 | goto retry; |
9a001fc1 | 1100 | } |
9050cce1 | 1101 | page = virt_to_page(zhdr); |
2f1e5e4d | 1102 | goto found; |
9a001fc1 VW |
1103 | } |
1104 | bud = FIRST; | |
9a001fc1 VW |
1105 | } |
1106 | ||
5c9bab59 VW |
1107 | page = NULL; |
1108 | if (can_sleep) { | |
1109 | spin_lock(&pool->stale_lock); | |
1110 | zhdr = list_first_entry_or_null(&pool->stale, | |
1111 | struct z3fold_header, buddy); | |
1112 | /* | |
1113 | * Before allocating a page, let's see if we can take one from | |
1114 | * the stale pages list. cancel_work_sync() can sleep so we | |
1115 | * limit this case to the contexts where we can sleep | |
1116 | */ | |
1117 | if (zhdr) { | |
1118 | list_del(&zhdr->buddy); | |
1119 | spin_unlock(&pool->stale_lock); | |
d30561c5 | 1120 | cancel_work_sync(&zhdr->work); |
5c9bab59 VW |
1121 | page = virt_to_page(zhdr); |
1122 | } else { | |
1123 | spin_unlock(&pool->stale_lock); | |
1124 | } | |
d30561c5 | 1125 | } |
5c9bab59 VW |
1126 | if (!page) |
1127 | page = alloc_page(gfp); | |
d30561c5 | 1128 | |
9a001fc1 VW |
1129 | if (!page) |
1130 | return -ENOMEM; | |
2f1e5e4d | 1131 | |
63398413 | 1132 | zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp); |
9050cce1 VW |
1133 | if (!zhdr) { |
1134 | __free_page(page); | |
1135 | return -ENOMEM; | |
1136 | } | |
1137 | atomic64_inc(&pool->pages_nr); | |
9a001fc1 VW |
1138 | |
1139 | if (bud == HEADLESS) { | |
1140 | set_bit(PAGE_HEADLESS, &page->private); | |
1141 | goto headless; | |
1142 | } | |
810481a2 HB |
1143 | if (can_sleep) { |
1144 | lock_page(page); | |
1145 | __SetPageMovable(page, pool->inode->i_mapping); | |
1146 | unlock_page(page); | |
1147 | } else { | |
1148 | if (trylock_page(page)) { | |
1149 | __SetPageMovable(page, pool->inode->i_mapping); | |
1150 | unlock_page(page); | |
1151 | } | |
1152 | } | |
2f1e5e4d | 1153 | z3fold_page_lock(zhdr); |
9a001fc1 VW |
1154 | |
1155 | found: | |
1156 | if (bud == FIRST) | |
1157 | zhdr->first_chunks = chunks; | |
1158 | else if (bud == LAST) | |
1159 | zhdr->last_chunks = chunks; | |
1160 | else { | |
1161 | zhdr->middle_chunks = chunks; | |
ede93213 | 1162 | zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS; |
9a001fc1 | 1163 | } |
9050cce1 | 1164 | add_to_unbuddied(pool, zhdr); |
9a001fc1 VW |
1165 | |
1166 | headless: | |
d30561c5 | 1167 | spin_lock(&pool->lock); |
9a001fc1 VW |
1168 | /* Add/move z3fold page to beginning of LRU */ |
1169 | if (!list_empty(&page->lru)) | |
1170 | list_del(&page->lru); | |
1171 | ||
1172 | list_add(&page->lru, &pool->lru); | |
1173 | ||
1174 | *handle = encode_handle(zhdr, bud); | |
1175 | spin_unlock(&pool->lock); | |
2f1e5e4d VW |
1176 | if (bud != HEADLESS) |
1177 | z3fold_page_unlock(zhdr); | |
9a001fc1 VW |
1178 | |
1179 | return 0; | |
1180 | } | |
1181 | ||
1182 | /** | |
1183 | * z3fold_free() - frees the allocation associated with the given handle | |
1184 | * @pool: pool in which the allocation resided | |
1185 | * @handle: handle associated with the allocation returned by z3fold_alloc() | |
1186 | * | |
1187 | * In the case that the z3fold page in which the allocation resides is under | |
ed0e5dca ML |
1188 | * reclaim, as indicated by the PAGE_CLAIMED flag being set, this function |
1189 | * only sets the first|middle|last_chunks to 0. The page is actually freed | |
1190 | * once all buddies are evicted (see z3fold_reclaim_page() below). | |
9a001fc1 VW |
1191 | */ |
1192 | static void z3fold_free(struct z3fold_pool *pool, unsigned long handle) | |
1193 | { | |
1194 | struct z3fold_header *zhdr; | |
9a001fc1 VW |
1195 | struct page *page; |
1196 | enum buddy bud; | |
5b6807de | 1197 | bool page_claimed; |
9a001fc1 | 1198 | |
4a3ac931 | 1199 | zhdr = get_z3fold_header(handle); |
9a001fc1 | 1200 | page = virt_to_page(zhdr); |
5b6807de | 1201 | page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private); |
9a001fc1 VW |
1202 | |
1203 | if (test_bit(PAGE_HEADLESS, &page->private)) { | |
ca0246bb VW |
1204 | /* if a headless page is under reclaim, just leave. |
1205 | * NB: we use test_and_set_bit for a reason: if the bit | |
1206 | * has not been set before, we release this page | |
1207 | * immediately so we don't care about its value any more. | |
1208 | */ | |
5b6807de | 1209 | if (!page_claimed) { |
ca0246bb VW |
1210 | spin_lock(&pool->lock); |
1211 | list_del(&page->lru); | |
1212 | spin_unlock(&pool->lock); | |
4a3ac931 | 1213 | put_z3fold_header(zhdr); |
1f862989 | 1214 | free_z3fold_page(page, true); |
ca0246bb | 1215 | atomic64_dec(&pool->pages_nr); |
9a001fc1 | 1216 | } |
ca0246bb | 1217 | return; |
9a001fc1 VW |
1218 | } |
1219 | ||
ca0246bb | 1220 | /* Non-headless case */ |
ca0246bb VW |
1221 | bud = handle_to_buddy(handle); |
1222 | ||
1223 | switch (bud) { | |
1224 | case FIRST: | |
1225 | zhdr->first_chunks = 0; | |
1226 | break; | |
1227 | case MIDDLE: | |
1228 | zhdr->middle_chunks = 0; | |
1229 | break; | |
1230 | case LAST: | |
1231 | zhdr->last_chunks = 0; | |
1232 | break; | |
1233 | default: | |
1234 | pr_err("%s: unknown bud %d\n", __func__, bud); | |
1235 | WARN_ON(1); | |
4a3ac931 | 1236 | put_z3fold_header(zhdr); |
d30561c5 VW |
1237 | return; |
1238 | } | |
1239 | ||
4a3ac931 | 1240 | if (!page_claimed) |
fc548865 | 1241 | free_handle(handle, zhdr); |
d30561c5 VW |
1242 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) { |
1243 | atomic64_dec(&pool->pages_nr); | |
1244 | return; | |
1245 | } | |
5b6807de VW |
1246 | if (page_claimed) { |
1247 | /* the page has not been claimed by us */ | |
ed0e5dca | 1248 | put_z3fold_header(zhdr); |
6098d7e1 VW |
1249 | return; |
1250 | } | |
dcf5aedb | 1251 | if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) { |
4a3ac931 | 1252 | put_z3fold_header(zhdr); |
5b6807de | 1253 | clear_bit(PAGE_CLAIMED, &page->private); |
d30561c5 VW |
1254 | return; |
1255 | } | |
1256 | if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) { | |
2f1e5e4d | 1257 | spin_lock(&pool->lock); |
d30561c5 | 1258 | list_del_init(&zhdr->buddy); |
2f1e5e4d | 1259 | spin_unlock(&pool->lock); |
d30561c5 | 1260 | zhdr->cpu = -1; |
5d03a661 | 1261 | kref_get(&zhdr->refcount); |
5b6807de | 1262 | clear_bit(PAGE_CLAIMED, &page->private); |
4a3ac931 | 1263 | do_compact_page(zhdr, true); |
d30561c5 | 1264 | return; |
9a001fc1 | 1265 | } |
5d03a661 | 1266 | kref_get(&zhdr->refcount); |
5b6807de | 1267 | clear_bit(PAGE_CLAIMED, &page->private); |
4a3ac931 VW |
1268 | queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work); |
1269 | put_z3fold_header(zhdr); | |
9a001fc1 VW |
1270 | } |
1271 | ||
1272 | /** | |
1273 | * z3fold_reclaim_page() - evicts allocations from a pool page and frees it | |
1274 | * @pool: pool from which a page will attempt to be evicted | |
f144c390 | 1275 | * @retries: number of pages on the LRU list for which eviction will |
9a001fc1 VW |
1276 | * be attempted before failing |
1277 | * | |
1278 | * z3fold reclaim is different from normal system reclaim in that it is done | |
1279 | * from the bottom, up. This is because only the bottom layer, z3fold, has | |
1280 | * information on how the allocations are organized within each z3fold page. | |
1281 | * This has the potential to create interesting locking situations between | |
1282 | * z3fold and the user, however. | |
1283 | * | |
1284 | * To avoid these, this is how z3fold_reclaim_page() should be called: | |
f144c390 | 1285 | * |
9a001fc1 VW |
1286 | * The user detects a page should be reclaimed and calls z3fold_reclaim_page(). |
1287 | * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and | |
1288 | * call the user-defined eviction handler with the pool and handle as | |
1289 | * arguments. | |
1290 | * | |
1291 | * If the handle can not be evicted, the eviction handler should return | |
1292 | * non-zero. z3fold_reclaim_page() will add the z3fold page back to the | |
1293 | * appropriate list and try the next z3fold page on the LRU up to | |
1294 | * a user defined number of retries. | |
1295 | * | |
1296 | * If the handle is successfully evicted, the eviction handler should | |
1297 | * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free() | |
1298 | * contains logic to delay freeing the page if the page is under reclaim, | |
1299 | * as indicated by the setting of the PG_reclaim flag on the underlying page. | |
1300 | * | |
1301 | * If all buddies in the z3fold page are successfully evicted, then the | |
1302 | * z3fold page can be freed. | |
1303 | * | |
1304 | * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are | |
1305 | * no pages to evict or an eviction handler is not registered, -EAGAIN if | |
1306 | * the retry limit was hit. | |
1307 | */ | |
1308 | static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries) | |
1309 | { | |
4a3ac931 | 1310 | int i, ret = -1; |
d30561c5 VW |
1311 | struct z3fold_header *zhdr = NULL; |
1312 | struct page *page = NULL; | |
1313 | struct list_head *pos; | |
9a001fc1 | 1314 | unsigned long first_handle = 0, middle_handle = 0, last_handle = 0; |
dcf5aedb VW |
1315 | struct z3fold_buddy_slots slots __attribute__((aligned(SLOTS_ALIGN))); |
1316 | ||
1317 | rwlock_init(&slots.lock); | |
1318 | slots.pool = (unsigned long)pool | (1 << HANDLES_NOFREE); | |
9a001fc1 VW |
1319 | |
1320 | spin_lock(&pool->lock); | |
2f1e5e4d | 1321 | if (!pool->ops || !pool->ops->evict || retries == 0) { |
9a001fc1 VW |
1322 | spin_unlock(&pool->lock); |
1323 | return -EINVAL; | |
1324 | } | |
1325 | for (i = 0; i < retries; i++) { | |
2f1e5e4d VW |
1326 | if (list_empty(&pool->lru)) { |
1327 | spin_unlock(&pool->lock); | |
1328 | return -EINVAL; | |
1329 | } | |
d30561c5 VW |
1330 | list_for_each_prev(pos, &pool->lru) { |
1331 | page = list_entry(pos, struct page, lru); | |
ca0246bb | 1332 | |
3f9d2b57 | 1333 | zhdr = page_address(page); |
6d679578 TH |
1334 | if (test_bit(PAGE_HEADLESS, &page->private)) { |
1335 | /* | |
1336 | * For non-headless pages, we wait to do this | |
1337 | * until we have the page lock to avoid racing | |
1338 | * with __z3fold_alloc(). Headless pages don't | |
1339 | * have a lock (and __z3fold_alloc() will never | |
1340 | * see them), but we still need to test and set | |
1341 | * PAGE_CLAIMED to avoid racing with | |
1342 | * z3fold_free(), so just do it now before | |
1343 | * leaving the loop. | |
1344 | */ | |
1345 | if (test_and_set_bit(PAGE_CLAIMED, &page->private)) | |
1346 | continue; | |
1347 | ||
d30561c5 | 1348 | break; |
6d679578 | 1349 | } |
d30561c5 | 1350 | |
dcf5aedb VW |
1351 | if (kref_get_unless_zero(&zhdr->refcount) == 0) { |
1352 | zhdr = NULL; | |
1353 | break; | |
1354 | } | |
ca0246bb | 1355 | if (!z3fold_page_trylock(zhdr)) { |
dcf5aedb VW |
1356 | if (kref_put(&zhdr->refcount, |
1357 | release_z3fold_page)) | |
1358 | atomic64_dec(&pool->pages_nr); | |
ca0246bb | 1359 | zhdr = NULL; |
d30561c5 | 1360 | continue; /* can't evict at this point */ |
ca0246bb | 1361 | } |
dcf5aedb VW |
1362 | |
1363 | /* test_and_set_bit is of course atomic, but we still | |
1364 | * need to do it under page lock, otherwise checking | |
1365 | * that bit in __z3fold_alloc wouldn't make sense | |
1366 | */ | |
1367 | if (zhdr->foreign_handles || | |
1368 | test_and_set_bit(PAGE_CLAIMED, &page->private)) { | |
1369 | if (kref_put(&zhdr->refcount, | |
28473d91 | 1370 | release_z3fold_page_locked)) |
dcf5aedb VW |
1371 | atomic64_dec(&pool->pages_nr); |
1372 | else | |
1373 | z3fold_page_unlock(zhdr); | |
4a3ac931 VW |
1374 | zhdr = NULL; |
1375 | continue; /* can't evict such page */ | |
1376 | } | |
d30561c5 VW |
1377 | list_del_init(&zhdr->buddy); |
1378 | zhdr->cpu = -1; | |
6098d7e1 | 1379 | break; |
d30561c5 VW |
1380 | } |
1381 | ||
ca0246bb VW |
1382 | if (!zhdr) |
1383 | break; | |
1384 | ||
5a27aa82 | 1385 | list_del_init(&page->lru); |
d30561c5 | 1386 | spin_unlock(&pool->lock); |
9a001fc1 | 1387 | |
9a001fc1 | 1388 | if (!test_bit(PAGE_HEADLESS, &page->private)) { |
9a001fc1 | 1389 | /* |
3f9d2b57 VW |
1390 | * We need encode the handles before unlocking, and |
1391 | * use our local slots structure because z3fold_free | |
1392 | * can zero out zhdr->slots and we can't do much | |
1393 | * about that | |
9a001fc1 VW |
1394 | */ |
1395 | first_handle = 0; | |
1396 | last_handle = 0; | |
1397 | middle_handle = 0; | |
dcf5aedb | 1398 | memset(slots.slot, 0, sizeof(slots.slot)); |
9a001fc1 | 1399 | if (zhdr->first_chunks) |
dcf5aedb VW |
1400 | first_handle = __encode_handle(zhdr, &slots, |
1401 | FIRST); | |
9a001fc1 | 1402 | if (zhdr->middle_chunks) |
dcf5aedb VW |
1403 | middle_handle = __encode_handle(zhdr, &slots, |
1404 | MIDDLE); | |
9a001fc1 | 1405 | if (zhdr->last_chunks) |
dcf5aedb VW |
1406 | last_handle = __encode_handle(zhdr, &slots, |
1407 | LAST); | |
d30561c5 VW |
1408 | /* |
1409 | * it's safe to unlock here because we hold a | |
1410 | * reference to this page | |
1411 | */ | |
2f1e5e4d | 1412 | z3fold_page_unlock(zhdr); |
9a001fc1 | 1413 | } else { |
4a3ac931 | 1414 | first_handle = encode_handle(zhdr, HEADLESS); |
9a001fc1 VW |
1415 | last_handle = middle_handle = 0; |
1416 | } | |
9a001fc1 VW |
1417 | /* Issue the eviction callback(s) */ |
1418 | if (middle_handle) { | |
1419 | ret = pool->ops->evict(pool, middle_handle); | |
1420 | if (ret) | |
1421 | goto next; | |
1422 | } | |
1423 | if (first_handle) { | |
1424 | ret = pool->ops->evict(pool, first_handle); | |
1425 | if (ret) | |
1426 | goto next; | |
1427 | } | |
1428 | if (last_handle) { | |
1429 | ret = pool->ops->evict(pool, last_handle); | |
1430 | if (ret) | |
1431 | goto next; | |
1432 | } | |
1433 | next: | |
5a27aa82 VW |
1434 | if (test_bit(PAGE_HEADLESS, &page->private)) { |
1435 | if (ret == 0) { | |
1f862989 | 1436 | free_z3fold_page(page, true); |
ca0246bb | 1437 | atomic64_dec(&pool->pages_nr); |
5a27aa82 | 1438 | return 0; |
5a27aa82 | 1439 | } |
6098d7e1 VW |
1440 | spin_lock(&pool->lock); |
1441 | list_add(&page->lru, &pool->lru); | |
1442 | spin_unlock(&pool->lock); | |
3f9d2b57 | 1443 | clear_bit(PAGE_CLAIMED, &page->private); |
6098d7e1 | 1444 | } else { |
dcf5aedb | 1445 | struct z3fold_buddy_slots *slots = zhdr->slots; |
6098d7e1 | 1446 | z3fold_page_lock(zhdr); |
6098d7e1 VW |
1447 | if (kref_put(&zhdr->refcount, |
1448 | release_z3fold_page_locked)) { | |
dcf5aedb | 1449 | kmem_cache_free(pool->c_handle, slots); |
6098d7e1 VW |
1450 | atomic64_dec(&pool->pages_nr); |
1451 | return 0; | |
1452 | } | |
1453 | /* | |
1454 | * if we are here, the page is still not completely | |
1455 | * free. Take the global pool lock then to be able | |
1456 | * to add it back to the lru list | |
1457 | */ | |
1458 | spin_lock(&pool->lock); | |
1459 | list_add(&page->lru, &pool->lru); | |
d5567c9d | 1460 | spin_unlock(&pool->lock); |
6098d7e1 | 1461 | z3fold_page_unlock(zhdr); |
3f9d2b57 | 1462 | clear_bit(PAGE_CLAIMED, &page->private); |
5a27aa82 | 1463 | } |
2f1e5e4d | 1464 | |
6098d7e1 VW |
1465 | /* We started off locked to we need to lock the pool back */ |
1466 | spin_lock(&pool->lock); | |
9a001fc1 VW |
1467 | } |
1468 | spin_unlock(&pool->lock); | |
1469 | return -EAGAIN; | |
1470 | } | |
1471 | ||
1472 | /** | |
1473 | * z3fold_map() - maps the allocation associated with the given handle | |
1474 | * @pool: pool in which the allocation resides | |
1475 | * @handle: handle associated with the allocation to be mapped | |
1476 | * | |
1477 | * Extracts the buddy number from handle and constructs the pointer to the | |
1478 | * correct starting chunk within the page. | |
1479 | * | |
1480 | * Returns: a pointer to the mapped allocation | |
1481 | */ | |
1482 | static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle) | |
1483 | { | |
1484 | struct z3fold_header *zhdr; | |
1485 | struct page *page; | |
1486 | void *addr; | |
1487 | enum buddy buddy; | |
1488 | ||
4a3ac931 | 1489 | zhdr = get_z3fold_header(handle); |
9a001fc1 VW |
1490 | addr = zhdr; |
1491 | page = virt_to_page(zhdr); | |
1492 | ||
1493 | if (test_bit(PAGE_HEADLESS, &page->private)) | |
1494 | goto out; | |
1495 | ||
1496 | buddy = handle_to_buddy(handle); | |
1497 | switch (buddy) { | |
1498 | case FIRST: | |
1499 | addr += ZHDR_SIZE_ALIGNED; | |
1500 | break; | |
1501 | case MIDDLE: | |
1502 | addr += zhdr->start_middle << CHUNK_SHIFT; | |
1503 | set_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
1504 | break; | |
1505 | case LAST: | |
ca0246bb | 1506 | addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT); |
9a001fc1 VW |
1507 | break; |
1508 | default: | |
1509 | pr_err("unknown buddy id %d\n", buddy); | |
1510 | WARN_ON(1); | |
1511 | addr = NULL; | |
1512 | break; | |
1513 | } | |
2f1e5e4d | 1514 | |
1f862989 VW |
1515 | if (addr) |
1516 | zhdr->mapped_count++; | |
9a001fc1 | 1517 | out: |
4a3ac931 | 1518 | put_z3fold_header(zhdr); |
9a001fc1 VW |
1519 | return addr; |
1520 | } | |
1521 | ||
1522 | /** | |
1523 | * z3fold_unmap() - unmaps the allocation associated with the given handle | |
1524 | * @pool: pool in which the allocation resides | |
1525 | * @handle: handle associated with the allocation to be unmapped | |
1526 | */ | |
1527 | static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle) | |
1528 | { | |
1529 | struct z3fold_header *zhdr; | |
1530 | struct page *page; | |
1531 | enum buddy buddy; | |
1532 | ||
4a3ac931 | 1533 | zhdr = get_z3fold_header(handle); |
9a001fc1 VW |
1534 | page = virt_to_page(zhdr); |
1535 | ||
2f1e5e4d | 1536 | if (test_bit(PAGE_HEADLESS, &page->private)) |
9a001fc1 | 1537 | return; |
9a001fc1 VW |
1538 | |
1539 | buddy = handle_to_buddy(handle); | |
1540 | if (buddy == MIDDLE) | |
1541 | clear_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
1f862989 | 1542 | zhdr->mapped_count--; |
4a3ac931 | 1543 | put_z3fold_header(zhdr); |
9a001fc1 VW |
1544 | } |
1545 | ||
1546 | /** | |
1547 | * z3fold_get_pool_size() - gets the z3fold pool size in pages | |
1548 | * @pool: pool whose size is being queried | |
1549 | * | |
12d59ae6 | 1550 | * Returns: size in pages of the given pool. |
9a001fc1 VW |
1551 | */ |
1552 | static u64 z3fold_get_pool_size(struct z3fold_pool *pool) | |
1553 | { | |
12d59ae6 | 1554 | return atomic64_read(&pool->pages_nr); |
9a001fc1 VW |
1555 | } |
1556 | ||
1f862989 VW |
1557 | static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode) |
1558 | { | |
1559 | struct z3fold_header *zhdr; | |
1560 | struct z3fold_pool *pool; | |
1561 | ||
1562 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1563 | VM_BUG_ON_PAGE(PageIsolated(page), page); | |
1564 | ||
dcf5aedb | 1565 | if (test_bit(PAGE_HEADLESS, &page->private)) |
1f862989 VW |
1566 | return false; |
1567 | ||
1568 | zhdr = page_address(page); | |
1569 | z3fold_page_lock(zhdr); | |
1570 | if (test_bit(NEEDS_COMPACTING, &page->private) || | |
1571 | test_bit(PAGE_STALE, &page->private)) | |
1572 | goto out; | |
1573 | ||
4a3ac931 VW |
1574 | if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) |
1575 | goto out; | |
1576 | ||
dcf5aedb VW |
1577 | if (test_and_set_bit(PAGE_CLAIMED, &page->private)) |
1578 | goto out; | |
1f862989 | 1579 | pool = zhdr_to_pool(zhdr); |
4a3ac931 VW |
1580 | spin_lock(&pool->lock); |
1581 | if (!list_empty(&zhdr->buddy)) | |
1582 | list_del_init(&zhdr->buddy); | |
1583 | if (!list_empty(&page->lru)) | |
1584 | list_del_init(&page->lru); | |
1585 | spin_unlock(&pool->lock); | |
1586 | ||
1587 | kref_get(&zhdr->refcount); | |
1588 | z3fold_page_unlock(zhdr); | |
1589 | return true; | |
1f862989 | 1590 | |
1f862989 VW |
1591 | out: |
1592 | z3fold_page_unlock(zhdr); | |
1593 | return false; | |
1594 | } | |
1595 | ||
1596 | static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage, | |
1597 | struct page *page, enum migrate_mode mode) | |
1598 | { | |
1599 | struct z3fold_header *zhdr, *new_zhdr; | |
1600 | struct z3fold_pool *pool; | |
1601 | struct address_space *new_mapping; | |
1602 | ||
1603 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1604 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
dcf5aedb | 1605 | VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page); |
810481a2 | 1606 | VM_BUG_ON_PAGE(!PageLocked(newpage), newpage); |
1f862989 VW |
1607 | |
1608 | zhdr = page_address(page); | |
1609 | pool = zhdr_to_pool(zhdr); | |
1610 | ||
dcf5aedb | 1611 | if (!z3fold_page_trylock(zhdr)) |
1f862989 | 1612 | return -EAGAIN; |
4a3ac931 | 1613 | if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) { |
1f862989 | 1614 | z3fold_page_unlock(zhdr); |
dcf5aedb | 1615 | clear_bit(PAGE_CLAIMED, &page->private); |
1f862989 VW |
1616 | return -EBUSY; |
1617 | } | |
c92d2f38 HB |
1618 | if (work_pending(&zhdr->work)) { |
1619 | z3fold_page_unlock(zhdr); | |
1620 | return -EAGAIN; | |
1621 | } | |
1f862989 VW |
1622 | new_zhdr = page_address(newpage); |
1623 | memcpy(new_zhdr, zhdr, PAGE_SIZE); | |
1624 | newpage->private = page->private; | |
1625 | page->private = 0; | |
1626 | z3fold_page_unlock(zhdr); | |
1627 | spin_lock_init(&new_zhdr->page_lock); | |
c92d2f38 HB |
1628 | INIT_WORK(&new_zhdr->work, compact_page_work); |
1629 | /* | |
1630 | * z3fold_page_isolate() ensures that new_zhdr->buddy is empty, | |
1631 | * so we only have to reinitialize it. | |
1632 | */ | |
1633 | INIT_LIST_HEAD(&new_zhdr->buddy); | |
1f862989 VW |
1634 | new_mapping = page_mapping(page); |
1635 | __ClearPageMovable(page); | |
1f862989 VW |
1636 | |
1637 | get_page(newpage); | |
1638 | z3fold_page_lock(new_zhdr); | |
1639 | if (new_zhdr->first_chunks) | |
1640 | encode_handle(new_zhdr, FIRST); | |
1641 | if (new_zhdr->last_chunks) | |
1642 | encode_handle(new_zhdr, LAST); | |
1643 | if (new_zhdr->middle_chunks) | |
1644 | encode_handle(new_zhdr, MIDDLE); | |
1645 | set_bit(NEEDS_COMPACTING, &newpage->private); | |
1646 | new_zhdr->cpu = smp_processor_id(); | |
1647 | spin_lock(&pool->lock); | |
1648 | list_add(&newpage->lru, &pool->lru); | |
1649 | spin_unlock(&pool->lock); | |
1650 | __SetPageMovable(newpage, new_mapping); | |
1651 | z3fold_page_unlock(new_zhdr); | |
1652 | ||
1653 | queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work); | |
1654 | ||
dcf5aedb | 1655 | clear_bit(PAGE_CLAIMED, &page->private); |
1f862989 VW |
1656 | put_page(page); |
1657 | return 0; | |
1658 | } | |
1659 | ||
1660 | static void z3fold_page_putback(struct page *page) | |
1661 | { | |
1662 | struct z3fold_header *zhdr; | |
1663 | struct z3fold_pool *pool; | |
1664 | ||
1665 | zhdr = page_address(page); | |
1666 | pool = zhdr_to_pool(zhdr); | |
1667 | ||
1668 | z3fold_page_lock(zhdr); | |
1669 | if (!list_empty(&zhdr->buddy)) | |
1670 | list_del_init(&zhdr->buddy); | |
1671 | INIT_LIST_HEAD(&page->lru); | |
1672 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) { | |
1673 | atomic64_dec(&pool->pages_nr); | |
1674 | return; | |
1675 | } | |
1676 | spin_lock(&pool->lock); | |
1677 | list_add(&page->lru, &pool->lru); | |
1678 | spin_unlock(&pool->lock); | |
dcf5aedb | 1679 | clear_bit(PAGE_CLAIMED, &page->private); |
1f862989 VW |
1680 | z3fold_page_unlock(zhdr); |
1681 | } | |
1682 | ||
1683 | static const struct address_space_operations z3fold_aops = { | |
1684 | .isolate_page = z3fold_page_isolate, | |
1685 | .migratepage = z3fold_page_migrate, | |
1686 | .putback_page = z3fold_page_putback, | |
1687 | }; | |
1688 | ||
9a001fc1 VW |
1689 | /***************** |
1690 | * zpool | |
1691 | ****************/ | |
1692 | ||
1693 | static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle) | |
1694 | { | |
1695 | if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict) | |
1696 | return pool->zpool_ops->evict(pool->zpool, handle); | |
1697 | else | |
1698 | return -ENOENT; | |
1699 | } | |
1700 | ||
1701 | static const struct z3fold_ops z3fold_zpool_ops = { | |
1702 | .evict = z3fold_zpool_evict | |
1703 | }; | |
1704 | ||
1705 | static void *z3fold_zpool_create(const char *name, gfp_t gfp, | |
1706 | const struct zpool_ops *zpool_ops, | |
1707 | struct zpool *zpool) | |
1708 | { | |
1709 | struct z3fold_pool *pool; | |
1710 | ||
d30561c5 VW |
1711 | pool = z3fold_create_pool(name, gfp, |
1712 | zpool_ops ? &z3fold_zpool_ops : NULL); | |
9a001fc1 VW |
1713 | if (pool) { |
1714 | pool->zpool = zpool; | |
1715 | pool->zpool_ops = zpool_ops; | |
1716 | } | |
1717 | return pool; | |
1718 | } | |
1719 | ||
1720 | static void z3fold_zpool_destroy(void *pool) | |
1721 | { | |
1722 | z3fold_destroy_pool(pool); | |
1723 | } | |
1724 | ||
1725 | static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
1726 | unsigned long *handle) | |
1727 | { | |
1728 | return z3fold_alloc(pool, size, gfp, handle); | |
1729 | } | |
1730 | static void z3fold_zpool_free(void *pool, unsigned long handle) | |
1731 | { | |
1732 | z3fold_free(pool, handle); | |
1733 | } | |
1734 | ||
1735 | static int z3fold_zpool_shrink(void *pool, unsigned int pages, | |
1736 | unsigned int *reclaimed) | |
1737 | { | |
1738 | unsigned int total = 0; | |
1739 | int ret = -EINVAL; | |
1740 | ||
1741 | while (total < pages) { | |
1742 | ret = z3fold_reclaim_page(pool, 8); | |
1743 | if (ret < 0) | |
1744 | break; | |
1745 | total++; | |
1746 | } | |
1747 | ||
1748 | if (reclaimed) | |
1749 | *reclaimed = total; | |
1750 | ||
1751 | return ret; | |
1752 | } | |
1753 | ||
1754 | static void *z3fold_zpool_map(void *pool, unsigned long handle, | |
1755 | enum zpool_mapmode mm) | |
1756 | { | |
1757 | return z3fold_map(pool, handle); | |
1758 | } | |
1759 | static void z3fold_zpool_unmap(void *pool, unsigned long handle) | |
1760 | { | |
1761 | z3fold_unmap(pool, handle); | |
1762 | } | |
1763 | ||
1764 | static u64 z3fold_zpool_total_size(void *pool) | |
1765 | { | |
1766 | return z3fold_get_pool_size(pool) * PAGE_SIZE; | |
1767 | } | |
1768 | ||
1769 | static struct zpool_driver z3fold_zpool_driver = { | |
1770 | .type = "z3fold", | |
e818e820 | 1771 | .sleep_mapped = true, |
9a001fc1 VW |
1772 | .owner = THIS_MODULE, |
1773 | .create = z3fold_zpool_create, | |
1774 | .destroy = z3fold_zpool_destroy, | |
1775 | .malloc = z3fold_zpool_malloc, | |
1776 | .free = z3fold_zpool_free, | |
1777 | .shrink = z3fold_zpool_shrink, | |
1778 | .map = z3fold_zpool_map, | |
1779 | .unmap = z3fold_zpool_unmap, | |
1780 | .total_size = z3fold_zpool_total_size, | |
1781 | }; | |
1782 | ||
1783 | MODULE_ALIAS("zpool-z3fold"); | |
1784 | ||
1785 | static int __init init_z3fold(void) | |
1786 | { | |
1f862989 VW |
1787 | int ret; |
1788 | ||
014284a0 ML |
1789 | /* |
1790 | * Make sure the z3fold header is not larger than the page size and | |
1791 | * there has remaining spaces for its buddy. | |
1792 | */ | |
1793 | BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE); | |
1f862989 VW |
1794 | ret = z3fold_mount(); |
1795 | if (ret) | |
1796 | return ret; | |
1797 | ||
9a001fc1 VW |
1798 | zpool_register_driver(&z3fold_zpool_driver); |
1799 | ||
1800 | return 0; | |
1801 | } | |
1802 | ||
1803 | static void __exit exit_z3fold(void) | |
1804 | { | |
1f862989 | 1805 | z3fold_unmount(); |
9a001fc1 VW |
1806 | zpool_unregister_driver(&z3fold_zpool_driver); |
1807 | } | |
1808 | ||
1809 | module_init(init_z3fold); | |
1810 | module_exit(exit_z3fold); | |
1811 | ||
1812 | MODULE_LICENSE("GPL"); | |
1813 | MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>"); | |
1814 | MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages"); |