]>
Commit | Line | Data |
---|---|---|
9a001fc1 VW |
1 | /* |
2 | * z3fold.c | |
3 | * | |
4 | * Author: Vitaly Wool <vitaly.wool@konsulko.com> | |
5 | * Copyright (C) 2016, Sony Mobile Communications Inc. | |
6 | * | |
7 | * This implementation is based on zbud written by Seth Jennings. | |
8 | * | |
9 | * z3fold is an special purpose allocator for storing compressed pages. It | |
10 | * can store up to three compressed pages per page which improves the | |
11 | * compression ratio of zbud while retaining its main concepts (e. g. always | |
12 | * storing an integral number of objects per page) and simplicity. | |
13 | * It still has simple and deterministic reclaim properties that make it | |
14 | * preferable to a higher density approach (with no requirement on integral | |
15 | * number of object per page) when reclaim is used. | |
16 | * | |
17 | * As in zbud, pages are divided into "chunks". The size of the chunks is | |
18 | * fixed at compile time and is determined by NCHUNKS_ORDER below. | |
19 | * | |
20 | * z3fold doesn't export any API and is meant to be used via zpool API. | |
21 | */ | |
22 | ||
23 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
24 | ||
25 | #include <linux/atomic.h> | |
d30561c5 | 26 | #include <linux/sched.h> |
9a001fc1 VW |
27 | #include <linux/list.h> |
28 | #include <linux/mm.h> | |
29 | #include <linux/module.h> | |
d30561c5 | 30 | #include <linux/percpu.h> |
9a001fc1 | 31 | #include <linux/preempt.h> |
d30561c5 | 32 | #include <linux/workqueue.h> |
9a001fc1 VW |
33 | #include <linux/slab.h> |
34 | #include <linux/spinlock.h> | |
35 | #include <linux/zpool.h> | |
36 | ||
7c2b8baa VW |
37 | /* |
38 | * NCHUNKS_ORDER determines the internal allocation granularity, effectively | |
39 | * adjusting internal fragmentation. It also determines the number of | |
40 | * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the | |
41 | * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks | |
42 | * in the beginning of an allocated page are occupied by z3fold header, so | |
43 | * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y), | |
44 | * which shows the max number of free chunks in z3fold page, also there will | |
45 | * be 63, or 62, respectively, freelists per pool. | |
46 | */ | |
47 | #define NCHUNKS_ORDER 6 | |
48 | ||
49 | #define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER) | |
50 | #define CHUNK_SIZE (1 << CHUNK_SHIFT) | |
51 | #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE) | |
52 | #define ZHDR_CHUNKS (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT) | |
53 | #define TOTAL_CHUNKS (PAGE_SIZE >> CHUNK_SHIFT) | |
54 | #define NCHUNKS ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT) | |
55 | ||
56 | #define BUDDY_MASK (0x3) | |
57 | #define BUDDY_SHIFT 2 | |
58 | #define SLOTS_ALIGN (0x40) | |
59 | ||
9a001fc1 VW |
60 | /***************** |
61 | * Structures | |
62 | *****************/ | |
ede93213 VW |
63 | struct z3fold_pool; |
64 | struct z3fold_ops { | |
65 | int (*evict)(struct z3fold_pool *pool, unsigned long handle); | |
66 | }; | |
67 | ||
68 | enum buddy { | |
69 | HEADLESS = 0, | |
70 | FIRST, | |
71 | MIDDLE, | |
72 | LAST, | |
7c2b8baa VW |
73 | BUDDIES_MAX = LAST |
74 | }; | |
75 | ||
76 | struct z3fold_buddy_slots { | |
77 | /* | |
78 | * we are using BUDDY_MASK in handle_to_buddy etc. so there should | |
79 | * be enough slots to hold all possible variants | |
80 | */ | |
81 | unsigned long slot[BUDDY_MASK + 1]; | |
82 | unsigned long pool; /* back link + flags */ | |
ede93213 | 83 | }; |
7c2b8baa | 84 | #define HANDLE_FLAG_MASK (0x03) |
ede93213 VW |
85 | |
86 | /* | |
d30561c5 | 87 | * struct z3fold_header - z3fold page metadata occupying first chunks of each |
ede93213 | 88 | * z3fold page, except for HEADLESS pages |
d30561c5 VW |
89 | * @buddy: links the z3fold page into the relevant list in the |
90 | * pool | |
2f1e5e4d | 91 | * @page_lock: per-page lock |
d30561c5 VW |
92 | * @refcount: reference count for the z3fold page |
93 | * @work: work_struct for page layout optimization | |
7c2b8baa | 94 | * @slots: pointer to the structure holding buddy slots |
d30561c5 | 95 | * @cpu: CPU which this page "belongs" to |
ede93213 VW |
96 | * @first_chunks: the size of the first buddy in chunks, 0 if free |
97 | * @middle_chunks: the size of the middle buddy in chunks, 0 if free | |
98 | * @last_chunks: the size of the last buddy in chunks, 0 if free | |
99 | * @first_num: the starting number (for the first handle) | |
100 | */ | |
101 | struct z3fold_header { | |
102 | struct list_head buddy; | |
2f1e5e4d | 103 | spinlock_t page_lock; |
5a27aa82 | 104 | struct kref refcount; |
d30561c5 | 105 | struct work_struct work; |
7c2b8baa | 106 | struct z3fold_buddy_slots *slots; |
d30561c5 | 107 | short cpu; |
ede93213 VW |
108 | unsigned short first_chunks; |
109 | unsigned short middle_chunks; | |
110 | unsigned short last_chunks; | |
111 | unsigned short start_middle; | |
112 | unsigned short first_num:2; | |
113 | }; | |
114 | ||
9a001fc1 VW |
115 | /** |
116 | * struct z3fold_pool - stores metadata for each z3fold pool | |
d30561c5 VW |
117 | * @name: pool name |
118 | * @lock: protects pool unbuddied/lru lists | |
119 | * @stale_lock: protects pool stale page list | |
120 | * @unbuddied: per-cpu array of lists tracking z3fold pages that contain 2- | |
121 | * buddies; the list each z3fold page is added to depends on | |
122 | * the size of its free region. | |
9a001fc1 VW |
123 | * @lru: list tracking the z3fold pages in LRU order by most recently |
124 | * added buddy. | |
d30561c5 | 125 | * @stale: list of pages marked for freeing |
9a001fc1 | 126 | * @pages_nr: number of z3fold pages in the pool. |
7c2b8baa | 127 | * @c_handle: cache for z3fold_buddy_slots allocation |
9a001fc1 VW |
128 | * @ops: pointer to a structure of user defined operations specified at |
129 | * pool creation time. | |
d30561c5 VW |
130 | * @compact_wq: workqueue for page layout background optimization |
131 | * @release_wq: workqueue for safe page release | |
132 | * @work: work_struct for safe page release | |
9a001fc1 VW |
133 | * |
134 | * This structure is allocated at pool creation time and maintains metadata | |
135 | * pertaining to a particular z3fold pool. | |
136 | */ | |
137 | struct z3fold_pool { | |
d30561c5 | 138 | const char *name; |
9a001fc1 | 139 | spinlock_t lock; |
d30561c5 VW |
140 | spinlock_t stale_lock; |
141 | struct list_head *unbuddied; | |
9a001fc1 | 142 | struct list_head lru; |
d30561c5 | 143 | struct list_head stale; |
12d59ae6 | 144 | atomic64_t pages_nr; |
7c2b8baa | 145 | struct kmem_cache *c_handle; |
9a001fc1 VW |
146 | const struct z3fold_ops *ops; |
147 | struct zpool *zpool; | |
148 | const struct zpool_ops *zpool_ops; | |
d30561c5 VW |
149 | struct workqueue_struct *compact_wq; |
150 | struct workqueue_struct *release_wq; | |
151 | struct work_struct work; | |
9a001fc1 VW |
152 | }; |
153 | ||
9a001fc1 VW |
154 | /* |
155 | * Internal z3fold page flags | |
156 | */ | |
157 | enum z3fold_page_flags { | |
5a27aa82 | 158 | PAGE_HEADLESS = 0, |
9a001fc1 | 159 | MIDDLE_CHUNK_MAPPED, |
d30561c5 | 160 | NEEDS_COMPACTING, |
6098d7e1 | 161 | PAGE_STALE, |
ca0246bb | 162 | PAGE_CLAIMED, /* by either reclaim or free */ |
9a001fc1 VW |
163 | }; |
164 | ||
165 | /***************** | |
166 | * Helpers | |
167 | *****************/ | |
168 | ||
169 | /* Converts an allocation size in bytes to size in z3fold chunks */ | |
170 | static int size_to_chunks(size_t size) | |
171 | { | |
172 | return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT; | |
173 | } | |
174 | ||
175 | #define for_each_unbuddied_list(_iter, _begin) \ | |
176 | for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++) | |
177 | ||
d30561c5 VW |
178 | static void compact_page_work(struct work_struct *w); |
179 | ||
7c2b8baa VW |
180 | static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool) |
181 | { | |
182 | struct z3fold_buddy_slots *slots = kmem_cache_alloc(pool->c_handle, | |
183 | GFP_KERNEL); | |
184 | ||
185 | if (slots) { | |
186 | memset(slots->slot, 0, sizeof(slots->slot)); | |
187 | slots->pool = (unsigned long)pool; | |
188 | } | |
189 | ||
190 | return slots; | |
191 | } | |
192 | ||
193 | static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s) | |
194 | { | |
195 | return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK); | |
196 | } | |
197 | ||
198 | static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle) | |
199 | { | |
200 | return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1)); | |
201 | } | |
202 | ||
203 | static inline void free_handle(unsigned long handle) | |
204 | { | |
205 | struct z3fold_buddy_slots *slots; | |
206 | int i; | |
207 | bool is_free; | |
208 | ||
209 | if (handle & (1 << PAGE_HEADLESS)) | |
210 | return; | |
211 | ||
212 | WARN_ON(*(unsigned long *)handle == 0); | |
213 | *(unsigned long *)handle = 0; | |
214 | slots = handle_to_slots(handle); | |
215 | is_free = true; | |
216 | for (i = 0; i <= BUDDY_MASK; i++) { | |
217 | if (slots->slot[i]) { | |
218 | is_free = false; | |
219 | break; | |
220 | } | |
221 | } | |
222 | ||
223 | if (is_free) { | |
224 | struct z3fold_pool *pool = slots_to_pool(slots); | |
225 | ||
226 | kmem_cache_free(pool->c_handle, slots); | |
227 | } | |
228 | } | |
229 | ||
9a001fc1 | 230 | /* Initializes the z3fold header of a newly allocated z3fold page */ |
d30561c5 VW |
231 | static struct z3fold_header *init_z3fold_page(struct page *page, |
232 | struct z3fold_pool *pool) | |
9a001fc1 VW |
233 | { |
234 | struct z3fold_header *zhdr = page_address(page); | |
7c2b8baa VW |
235 | struct z3fold_buddy_slots *slots = alloc_slots(pool); |
236 | ||
237 | if (!slots) | |
238 | return NULL; | |
9a001fc1 VW |
239 | |
240 | INIT_LIST_HEAD(&page->lru); | |
9a001fc1 VW |
241 | clear_bit(PAGE_HEADLESS, &page->private); |
242 | clear_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
d30561c5 VW |
243 | clear_bit(NEEDS_COMPACTING, &page->private); |
244 | clear_bit(PAGE_STALE, &page->private); | |
ca0246bb | 245 | clear_bit(PAGE_CLAIMED, &page->private); |
9a001fc1 | 246 | |
2f1e5e4d | 247 | spin_lock_init(&zhdr->page_lock); |
5a27aa82 | 248 | kref_init(&zhdr->refcount); |
9a001fc1 VW |
249 | zhdr->first_chunks = 0; |
250 | zhdr->middle_chunks = 0; | |
251 | zhdr->last_chunks = 0; | |
252 | zhdr->first_num = 0; | |
253 | zhdr->start_middle = 0; | |
d30561c5 | 254 | zhdr->cpu = -1; |
7c2b8baa | 255 | zhdr->slots = slots; |
9a001fc1 | 256 | INIT_LIST_HEAD(&zhdr->buddy); |
d30561c5 | 257 | INIT_WORK(&zhdr->work, compact_page_work); |
9a001fc1 VW |
258 | return zhdr; |
259 | } | |
260 | ||
261 | /* Resets the struct page fields and frees the page */ | |
5a27aa82 | 262 | static void free_z3fold_page(struct page *page) |
9a001fc1 | 263 | { |
5a27aa82 VW |
264 | __free_page(page); |
265 | } | |
266 | ||
2f1e5e4d VW |
267 | /* Lock a z3fold page */ |
268 | static inline void z3fold_page_lock(struct z3fold_header *zhdr) | |
269 | { | |
270 | spin_lock(&zhdr->page_lock); | |
271 | } | |
272 | ||
76e32a2a VW |
273 | /* Try to lock a z3fold page */ |
274 | static inline int z3fold_page_trylock(struct z3fold_header *zhdr) | |
275 | { | |
276 | return spin_trylock(&zhdr->page_lock); | |
277 | } | |
278 | ||
2f1e5e4d VW |
279 | /* Unlock a z3fold page */ |
280 | static inline void z3fold_page_unlock(struct z3fold_header *zhdr) | |
281 | { | |
282 | spin_unlock(&zhdr->page_lock); | |
283 | } | |
284 | ||
7c2b8baa VW |
285 | /* Helper function to build the index */ |
286 | static inline int __idx(struct z3fold_header *zhdr, enum buddy bud) | |
287 | { | |
288 | return (bud + zhdr->first_num) & BUDDY_MASK; | |
289 | } | |
290 | ||
9a001fc1 VW |
291 | /* |
292 | * Encodes the handle of a particular buddy within a z3fold page | |
293 | * Pool lock should be held as this function accesses first_num | |
294 | */ | |
295 | static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud) | |
296 | { | |
7c2b8baa VW |
297 | struct z3fold_buddy_slots *slots; |
298 | unsigned long h = (unsigned long)zhdr; | |
299 | int idx = 0; | |
9a001fc1 | 300 | |
7c2b8baa VW |
301 | /* |
302 | * For a headless page, its handle is its pointer with the extra | |
303 | * PAGE_HEADLESS bit set | |
304 | */ | |
305 | if (bud == HEADLESS) | |
306 | return h | (1 << PAGE_HEADLESS); | |
307 | ||
308 | /* otherwise, return pointer to encoded handle */ | |
309 | idx = __idx(zhdr, bud); | |
310 | h += idx; | |
311 | if (bud == LAST) | |
312 | h |= (zhdr->last_chunks << BUDDY_SHIFT); | |
313 | ||
314 | slots = zhdr->slots; | |
315 | slots->slot[idx] = h; | |
316 | return (unsigned long)&slots->slot[idx]; | |
9a001fc1 VW |
317 | } |
318 | ||
319 | /* Returns the z3fold page where a given handle is stored */ | |
7c2b8baa | 320 | static inline struct z3fold_header *handle_to_z3fold_header(unsigned long handle) |
9a001fc1 | 321 | { |
7c2b8baa VW |
322 | unsigned long addr = handle; |
323 | ||
324 | if (!(addr & (1 << PAGE_HEADLESS))) | |
325 | addr = *(unsigned long *)handle; | |
326 | ||
327 | return (struct z3fold_header *)(addr & PAGE_MASK); | |
9a001fc1 VW |
328 | } |
329 | ||
ca0246bb VW |
330 | /* only for LAST bud, returns zero otherwise */ |
331 | static unsigned short handle_to_chunks(unsigned long handle) | |
332 | { | |
7c2b8baa VW |
333 | unsigned long addr = *(unsigned long *)handle; |
334 | ||
335 | return (addr & ~PAGE_MASK) >> BUDDY_SHIFT; | |
ca0246bb VW |
336 | } |
337 | ||
f201ebd8 | 338 | /* |
339 | * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle | |
340 | * but that doesn't matter. because the masking will result in the | |
341 | * correct buddy number. | |
342 | */ | |
9a001fc1 VW |
343 | static enum buddy handle_to_buddy(unsigned long handle) |
344 | { | |
7c2b8baa VW |
345 | struct z3fold_header *zhdr; |
346 | unsigned long addr; | |
347 | ||
348 | WARN_ON(handle & (1 << PAGE_HEADLESS)); | |
349 | addr = *(unsigned long *)handle; | |
350 | zhdr = (struct z3fold_header *)(addr & PAGE_MASK); | |
351 | return (addr - zhdr->first_num) & BUDDY_MASK; | |
9a001fc1 VW |
352 | } |
353 | ||
9050cce1 VW |
354 | static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr) |
355 | { | |
7c2b8baa | 356 | return slots_to_pool(zhdr->slots); |
9050cce1 VW |
357 | } |
358 | ||
d30561c5 VW |
359 | static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked) |
360 | { | |
361 | struct page *page = virt_to_page(zhdr); | |
9050cce1 | 362 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
d30561c5 VW |
363 | |
364 | WARN_ON(!list_empty(&zhdr->buddy)); | |
365 | set_bit(PAGE_STALE, &page->private); | |
35529357 | 366 | clear_bit(NEEDS_COMPACTING, &page->private); |
d30561c5 VW |
367 | spin_lock(&pool->lock); |
368 | if (!list_empty(&page->lru)) | |
369 | list_del(&page->lru); | |
370 | spin_unlock(&pool->lock); | |
371 | if (locked) | |
372 | z3fold_page_unlock(zhdr); | |
373 | spin_lock(&pool->stale_lock); | |
374 | list_add(&zhdr->buddy, &pool->stale); | |
375 | queue_work(pool->release_wq, &pool->work); | |
376 | spin_unlock(&pool->stale_lock); | |
377 | } | |
378 | ||
379 | static void __attribute__((__unused__)) | |
380 | release_z3fold_page(struct kref *ref) | |
381 | { | |
382 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
383 | refcount); | |
384 | __release_z3fold_page(zhdr, false); | |
385 | } | |
386 | ||
387 | static void release_z3fold_page_locked(struct kref *ref) | |
388 | { | |
389 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
390 | refcount); | |
391 | WARN_ON(z3fold_page_trylock(zhdr)); | |
392 | __release_z3fold_page(zhdr, true); | |
393 | } | |
394 | ||
395 | static void release_z3fold_page_locked_list(struct kref *ref) | |
396 | { | |
397 | struct z3fold_header *zhdr = container_of(ref, struct z3fold_header, | |
398 | refcount); | |
9050cce1 VW |
399 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
400 | spin_lock(&pool->lock); | |
d30561c5 | 401 | list_del_init(&zhdr->buddy); |
9050cce1 | 402 | spin_unlock(&pool->lock); |
d30561c5 VW |
403 | |
404 | WARN_ON(z3fold_page_trylock(zhdr)); | |
405 | __release_z3fold_page(zhdr, true); | |
406 | } | |
407 | ||
408 | static void free_pages_work(struct work_struct *w) | |
409 | { | |
410 | struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work); | |
411 | ||
412 | spin_lock(&pool->stale_lock); | |
413 | while (!list_empty(&pool->stale)) { | |
414 | struct z3fold_header *zhdr = list_first_entry(&pool->stale, | |
415 | struct z3fold_header, buddy); | |
416 | struct page *page = virt_to_page(zhdr); | |
417 | ||
418 | list_del(&zhdr->buddy); | |
419 | if (WARN_ON(!test_bit(PAGE_STALE, &page->private))) | |
420 | continue; | |
d30561c5 VW |
421 | spin_unlock(&pool->stale_lock); |
422 | cancel_work_sync(&zhdr->work); | |
423 | free_z3fold_page(page); | |
424 | cond_resched(); | |
425 | spin_lock(&pool->stale_lock); | |
426 | } | |
427 | spin_unlock(&pool->stale_lock); | |
428 | } | |
429 | ||
9a001fc1 VW |
430 | /* |
431 | * Returns the number of free chunks in a z3fold page. | |
432 | * NB: can't be used with HEADLESS pages. | |
433 | */ | |
434 | static int num_free_chunks(struct z3fold_header *zhdr) | |
435 | { | |
436 | int nfree; | |
437 | /* | |
438 | * If there is a middle object, pick up the bigger free space | |
439 | * either before or after it. Otherwise just subtract the number | |
440 | * of chunks occupied by the first and the last objects. | |
441 | */ | |
442 | if (zhdr->middle_chunks != 0) { | |
443 | int nfree_before = zhdr->first_chunks ? | |
ede93213 | 444 | 0 : zhdr->start_middle - ZHDR_CHUNKS; |
9a001fc1 | 445 | int nfree_after = zhdr->last_chunks ? |
ede93213 VW |
446 | 0 : TOTAL_CHUNKS - |
447 | (zhdr->start_middle + zhdr->middle_chunks); | |
9a001fc1 VW |
448 | nfree = max(nfree_before, nfree_after); |
449 | } else | |
450 | nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks; | |
451 | return nfree; | |
452 | } | |
453 | ||
9050cce1 VW |
454 | /* Add to the appropriate unbuddied list */ |
455 | static inline void add_to_unbuddied(struct z3fold_pool *pool, | |
456 | struct z3fold_header *zhdr) | |
457 | { | |
458 | if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 || | |
459 | zhdr->middle_chunks == 0) { | |
460 | struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied); | |
461 | ||
462 | int freechunks = num_free_chunks(zhdr); | |
463 | spin_lock(&pool->lock); | |
464 | list_add(&zhdr->buddy, &unbuddied[freechunks]); | |
465 | spin_unlock(&pool->lock); | |
466 | zhdr->cpu = smp_processor_id(); | |
467 | put_cpu_ptr(pool->unbuddied); | |
468 | } | |
469 | } | |
470 | ||
ede93213 VW |
471 | static inline void *mchunk_memmove(struct z3fold_header *zhdr, |
472 | unsigned short dst_chunk) | |
473 | { | |
474 | void *beg = zhdr; | |
475 | return memmove(beg + (dst_chunk << CHUNK_SHIFT), | |
476 | beg + (zhdr->start_middle << CHUNK_SHIFT), | |
477 | zhdr->middle_chunks << CHUNK_SHIFT); | |
478 | } | |
479 | ||
1b096e5a | 480 | #define BIG_CHUNK_GAP 3 |
9a001fc1 VW |
481 | /* Has to be called with lock held */ |
482 | static int z3fold_compact_page(struct z3fold_header *zhdr) | |
483 | { | |
484 | struct page *page = virt_to_page(zhdr); | |
9a001fc1 | 485 | |
ede93213 VW |
486 | if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private)) |
487 | return 0; /* can't move middle chunk, it's used */ | |
9a001fc1 | 488 | |
ede93213 VW |
489 | if (zhdr->middle_chunks == 0) |
490 | return 0; /* nothing to compact */ | |
491 | ||
492 | if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) { | |
493 | /* move to the beginning */ | |
494 | mchunk_memmove(zhdr, ZHDR_CHUNKS); | |
9a001fc1 VW |
495 | zhdr->first_chunks = zhdr->middle_chunks; |
496 | zhdr->middle_chunks = 0; | |
497 | zhdr->start_middle = 0; | |
498 | zhdr->first_num++; | |
1b096e5a | 499 | return 1; |
9a001fc1 | 500 | } |
1b096e5a VW |
501 | |
502 | /* | |
503 | * moving data is expensive, so let's only do that if | |
504 | * there's substantial gain (at least BIG_CHUNK_GAP chunks) | |
505 | */ | |
506 | if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 && | |
507 | zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >= | |
508 | BIG_CHUNK_GAP) { | |
509 | mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS); | |
510 | zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS; | |
511 | return 1; | |
512 | } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 && | |
513 | TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle | |
514 | + zhdr->middle_chunks) >= | |
515 | BIG_CHUNK_GAP) { | |
516 | unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks - | |
517 | zhdr->middle_chunks; | |
518 | mchunk_memmove(zhdr, new_start); | |
519 | zhdr->start_middle = new_start; | |
520 | return 1; | |
521 | } | |
522 | ||
523 | return 0; | |
9a001fc1 VW |
524 | } |
525 | ||
d30561c5 VW |
526 | static void do_compact_page(struct z3fold_header *zhdr, bool locked) |
527 | { | |
9050cce1 | 528 | struct z3fold_pool *pool = zhdr_to_pool(zhdr); |
d30561c5 | 529 | struct page *page; |
d30561c5 VW |
530 | |
531 | page = virt_to_page(zhdr); | |
532 | if (locked) | |
533 | WARN_ON(z3fold_page_trylock(zhdr)); | |
534 | else | |
535 | z3fold_page_lock(zhdr); | |
5d03a661 | 536 | if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) { |
d30561c5 VW |
537 | z3fold_page_unlock(zhdr); |
538 | return; | |
539 | } | |
540 | spin_lock(&pool->lock); | |
541 | list_del_init(&zhdr->buddy); | |
542 | spin_unlock(&pool->lock); | |
543 | ||
5d03a661 VW |
544 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) { |
545 | atomic64_dec(&pool->pages_nr); | |
546 | return; | |
547 | } | |
548 | ||
d30561c5 | 549 | z3fold_compact_page(zhdr); |
9050cce1 | 550 | add_to_unbuddied(pool, zhdr); |
d30561c5 VW |
551 | z3fold_page_unlock(zhdr); |
552 | } | |
553 | ||
554 | static void compact_page_work(struct work_struct *w) | |
555 | { | |
556 | struct z3fold_header *zhdr = container_of(w, struct z3fold_header, | |
557 | work); | |
558 | ||
559 | do_compact_page(zhdr, false); | |
560 | } | |
561 | ||
9050cce1 VW |
562 | /* returns _locked_ z3fold page header or NULL */ |
563 | static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool, | |
564 | size_t size, bool can_sleep) | |
565 | { | |
566 | struct z3fold_header *zhdr = NULL; | |
567 | struct page *page; | |
568 | struct list_head *unbuddied; | |
569 | int chunks = size_to_chunks(size), i; | |
570 | ||
571 | lookup: | |
572 | /* First, try to find an unbuddied z3fold page. */ | |
573 | unbuddied = get_cpu_ptr(pool->unbuddied); | |
574 | for_each_unbuddied_list(i, chunks) { | |
575 | struct list_head *l = &unbuddied[i]; | |
576 | ||
577 | zhdr = list_first_entry_or_null(READ_ONCE(l), | |
578 | struct z3fold_header, buddy); | |
579 | ||
580 | if (!zhdr) | |
581 | continue; | |
582 | ||
583 | /* Re-check under lock. */ | |
584 | spin_lock(&pool->lock); | |
585 | l = &unbuddied[i]; | |
586 | if (unlikely(zhdr != list_first_entry(READ_ONCE(l), | |
587 | struct z3fold_header, buddy)) || | |
588 | !z3fold_page_trylock(zhdr)) { | |
589 | spin_unlock(&pool->lock); | |
590 | zhdr = NULL; | |
591 | put_cpu_ptr(pool->unbuddied); | |
592 | if (can_sleep) | |
593 | cond_resched(); | |
594 | goto lookup; | |
595 | } | |
596 | list_del_init(&zhdr->buddy); | |
597 | zhdr->cpu = -1; | |
598 | spin_unlock(&pool->lock); | |
599 | ||
600 | page = virt_to_page(zhdr); | |
601 | if (test_bit(NEEDS_COMPACTING, &page->private)) { | |
602 | z3fold_page_unlock(zhdr); | |
603 | zhdr = NULL; | |
604 | put_cpu_ptr(pool->unbuddied); | |
605 | if (can_sleep) | |
606 | cond_resched(); | |
607 | goto lookup; | |
608 | } | |
609 | ||
610 | /* | |
611 | * this page could not be removed from its unbuddied | |
612 | * list while pool lock was held, and then we've taken | |
613 | * page lock so kref_put could not be called before | |
614 | * we got here, so it's safe to just call kref_get() | |
615 | */ | |
616 | kref_get(&zhdr->refcount); | |
617 | break; | |
618 | } | |
619 | put_cpu_ptr(pool->unbuddied); | |
620 | ||
351618b2 VW |
621 | if (!zhdr) { |
622 | int cpu; | |
623 | ||
624 | /* look for _exact_ match on other cpus' lists */ | |
625 | for_each_online_cpu(cpu) { | |
626 | struct list_head *l; | |
627 | ||
628 | unbuddied = per_cpu_ptr(pool->unbuddied, cpu); | |
629 | spin_lock(&pool->lock); | |
630 | l = &unbuddied[chunks]; | |
631 | ||
632 | zhdr = list_first_entry_or_null(READ_ONCE(l), | |
633 | struct z3fold_header, buddy); | |
634 | ||
635 | if (!zhdr || !z3fold_page_trylock(zhdr)) { | |
636 | spin_unlock(&pool->lock); | |
637 | zhdr = NULL; | |
638 | continue; | |
639 | } | |
640 | list_del_init(&zhdr->buddy); | |
641 | zhdr->cpu = -1; | |
642 | spin_unlock(&pool->lock); | |
643 | ||
644 | page = virt_to_page(zhdr); | |
645 | if (test_bit(NEEDS_COMPACTING, &page->private)) { | |
646 | z3fold_page_unlock(zhdr); | |
647 | zhdr = NULL; | |
648 | if (can_sleep) | |
649 | cond_resched(); | |
650 | continue; | |
651 | } | |
652 | kref_get(&zhdr->refcount); | |
653 | break; | |
654 | } | |
655 | } | |
656 | ||
9050cce1 VW |
657 | return zhdr; |
658 | } | |
d30561c5 VW |
659 | |
660 | /* | |
661 | * API Functions | |
662 | */ | |
663 | ||
664 | /** | |
665 | * z3fold_create_pool() - create a new z3fold pool | |
666 | * @name: pool name | |
667 | * @gfp: gfp flags when allocating the z3fold pool structure | |
668 | * @ops: user-defined operations for the z3fold pool | |
669 | * | |
670 | * Return: pointer to the new z3fold pool or NULL if the metadata allocation | |
671 | * failed. | |
672 | */ | |
673 | static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp, | |
674 | const struct z3fold_ops *ops) | |
675 | { | |
676 | struct z3fold_pool *pool = NULL; | |
677 | int i, cpu; | |
678 | ||
679 | pool = kzalloc(sizeof(struct z3fold_pool), gfp); | |
680 | if (!pool) | |
681 | goto out; | |
7c2b8baa VW |
682 | pool->c_handle = kmem_cache_create("z3fold_handle", |
683 | sizeof(struct z3fold_buddy_slots), | |
684 | SLOTS_ALIGN, 0, NULL); | |
685 | if (!pool->c_handle) | |
686 | goto out_c; | |
d30561c5 VW |
687 | spin_lock_init(&pool->lock); |
688 | spin_lock_init(&pool->stale_lock); | |
689 | pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2); | |
1ec6995d XW |
690 | if (!pool->unbuddied) |
691 | goto out_pool; | |
d30561c5 VW |
692 | for_each_possible_cpu(cpu) { |
693 | struct list_head *unbuddied = | |
694 | per_cpu_ptr(pool->unbuddied, cpu); | |
695 | for_each_unbuddied_list(i, 0) | |
696 | INIT_LIST_HEAD(&unbuddied[i]); | |
697 | } | |
698 | INIT_LIST_HEAD(&pool->lru); | |
699 | INIT_LIST_HEAD(&pool->stale); | |
700 | atomic64_set(&pool->pages_nr, 0); | |
701 | pool->name = name; | |
702 | pool->compact_wq = create_singlethread_workqueue(pool->name); | |
703 | if (!pool->compact_wq) | |
1ec6995d | 704 | goto out_unbuddied; |
d30561c5 VW |
705 | pool->release_wq = create_singlethread_workqueue(pool->name); |
706 | if (!pool->release_wq) | |
707 | goto out_wq; | |
708 | INIT_WORK(&pool->work, free_pages_work); | |
709 | pool->ops = ops; | |
710 | return pool; | |
711 | ||
712 | out_wq: | |
713 | destroy_workqueue(pool->compact_wq); | |
1ec6995d XW |
714 | out_unbuddied: |
715 | free_percpu(pool->unbuddied); | |
716 | out_pool: | |
7c2b8baa VW |
717 | kmem_cache_destroy(pool->c_handle); |
718 | out_c: | |
d30561c5 | 719 | kfree(pool); |
1ec6995d | 720 | out: |
d30561c5 VW |
721 | return NULL; |
722 | } | |
723 | ||
724 | /** | |
725 | * z3fold_destroy_pool() - destroys an existing z3fold pool | |
726 | * @pool: the z3fold pool to be destroyed | |
727 | * | |
728 | * The pool should be emptied before this function is called. | |
729 | */ | |
730 | static void z3fold_destroy_pool(struct z3fold_pool *pool) | |
731 | { | |
7c2b8baa | 732 | kmem_cache_destroy(pool->c_handle); |
d30561c5 VW |
733 | destroy_workqueue(pool->release_wq); |
734 | destroy_workqueue(pool->compact_wq); | |
735 | kfree(pool); | |
736 | } | |
737 | ||
9a001fc1 VW |
738 | /** |
739 | * z3fold_alloc() - allocates a region of a given size | |
740 | * @pool: z3fold pool from which to allocate | |
741 | * @size: size in bytes of the desired allocation | |
742 | * @gfp: gfp flags used if the pool needs to grow | |
743 | * @handle: handle of the new allocation | |
744 | * | |
745 | * This function will attempt to find a free region in the pool large enough to | |
746 | * satisfy the allocation request. A search of the unbuddied lists is | |
747 | * performed first. If no suitable free region is found, then a new page is | |
748 | * allocated and added to the pool to satisfy the request. | |
749 | * | |
750 | * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used | |
751 | * as z3fold pool pages. | |
752 | * | |
753 | * Return: 0 if success and handle is set, otherwise -EINVAL if the size or | |
754 | * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate | |
755 | * a new page. | |
756 | */ | |
757 | static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp, | |
758 | unsigned long *handle) | |
759 | { | |
9050cce1 | 760 | int chunks = size_to_chunks(size); |
9a001fc1 | 761 | struct z3fold_header *zhdr = NULL; |
d30561c5 | 762 | struct page *page = NULL; |
9a001fc1 | 763 | enum buddy bud; |
8a97ea54 | 764 | bool can_sleep = gfpflags_allow_blocking(gfp); |
9a001fc1 VW |
765 | |
766 | if (!size || (gfp & __GFP_HIGHMEM)) | |
767 | return -EINVAL; | |
768 | ||
769 | if (size > PAGE_SIZE) | |
770 | return -ENOSPC; | |
771 | ||
772 | if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE) | |
773 | bud = HEADLESS; | |
774 | else { | |
9050cce1 VW |
775 | retry: |
776 | zhdr = __z3fold_alloc(pool, size, can_sleep); | |
d30561c5 | 777 | if (zhdr) { |
2f1e5e4d VW |
778 | if (zhdr->first_chunks == 0) { |
779 | if (zhdr->middle_chunks != 0 && | |
780 | chunks >= zhdr->start_middle) | |
9a001fc1 | 781 | bud = LAST; |
2f1e5e4d VW |
782 | else |
783 | bud = FIRST; | |
784 | } else if (zhdr->last_chunks == 0) | |
785 | bud = LAST; | |
786 | else if (zhdr->middle_chunks == 0) | |
787 | bud = MIDDLE; | |
788 | else { | |
5a27aa82 | 789 | if (kref_put(&zhdr->refcount, |
d30561c5 | 790 | release_z3fold_page_locked)) |
5a27aa82 | 791 | atomic64_dec(&pool->pages_nr); |
d30561c5 VW |
792 | else |
793 | z3fold_page_unlock(zhdr); | |
2f1e5e4d VW |
794 | pr_err("No free chunks in unbuddied\n"); |
795 | WARN_ON(1); | |
9050cce1 | 796 | goto retry; |
9a001fc1 | 797 | } |
9050cce1 | 798 | page = virt_to_page(zhdr); |
2f1e5e4d | 799 | goto found; |
9a001fc1 VW |
800 | } |
801 | bud = FIRST; | |
9a001fc1 VW |
802 | } |
803 | ||
5c9bab59 VW |
804 | page = NULL; |
805 | if (can_sleep) { | |
806 | spin_lock(&pool->stale_lock); | |
807 | zhdr = list_first_entry_or_null(&pool->stale, | |
808 | struct z3fold_header, buddy); | |
809 | /* | |
810 | * Before allocating a page, let's see if we can take one from | |
811 | * the stale pages list. cancel_work_sync() can sleep so we | |
812 | * limit this case to the contexts where we can sleep | |
813 | */ | |
814 | if (zhdr) { | |
815 | list_del(&zhdr->buddy); | |
816 | spin_unlock(&pool->stale_lock); | |
d30561c5 | 817 | cancel_work_sync(&zhdr->work); |
5c9bab59 VW |
818 | page = virt_to_page(zhdr); |
819 | } else { | |
820 | spin_unlock(&pool->stale_lock); | |
821 | } | |
d30561c5 | 822 | } |
5c9bab59 VW |
823 | if (!page) |
824 | page = alloc_page(gfp); | |
d30561c5 | 825 | |
9a001fc1 VW |
826 | if (!page) |
827 | return -ENOMEM; | |
2f1e5e4d | 828 | |
d30561c5 | 829 | zhdr = init_z3fold_page(page, pool); |
9050cce1 VW |
830 | if (!zhdr) { |
831 | __free_page(page); | |
832 | return -ENOMEM; | |
833 | } | |
834 | atomic64_inc(&pool->pages_nr); | |
9a001fc1 VW |
835 | |
836 | if (bud == HEADLESS) { | |
837 | set_bit(PAGE_HEADLESS, &page->private); | |
838 | goto headless; | |
839 | } | |
2f1e5e4d | 840 | z3fold_page_lock(zhdr); |
9a001fc1 VW |
841 | |
842 | found: | |
843 | if (bud == FIRST) | |
844 | zhdr->first_chunks = chunks; | |
845 | else if (bud == LAST) | |
846 | zhdr->last_chunks = chunks; | |
847 | else { | |
848 | zhdr->middle_chunks = chunks; | |
ede93213 | 849 | zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS; |
9a001fc1 | 850 | } |
9050cce1 | 851 | add_to_unbuddied(pool, zhdr); |
9a001fc1 VW |
852 | |
853 | headless: | |
d30561c5 | 854 | spin_lock(&pool->lock); |
9a001fc1 VW |
855 | /* Add/move z3fold page to beginning of LRU */ |
856 | if (!list_empty(&page->lru)) | |
857 | list_del(&page->lru); | |
858 | ||
859 | list_add(&page->lru, &pool->lru); | |
860 | ||
861 | *handle = encode_handle(zhdr, bud); | |
862 | spin_unlock(&pool->lock); | |
2f1e5e4d VW |
863 | if (bud != HEADLESS) |
864 | z3fold_page_unlock(zhdr); | |
9a001fc1 VW |
865 | |
866 | return 0; | |
867 | } | |
868 | ||
869 | /** | |
870 | * z3fold_free() - frees the allocation associated with the given handle | |
871 | * @pool: pool in which the allocation resided | |
872 | * @handle: handle associated with the allocation returned by z3fold_alloc() | |
873 | * | |
874 | * In the case that the z3fold page in which the allocation resides is under | |
875 | * reclaim, as indicated by the PG_reclaim flag being set, this function | |
876 | * only sets the first|last_chunks to 0. The page is actually freed | |
877 | * once both buddies are evicted (see z3fold_reclaim_page() below). | |
878 | */ | |
879 | static void z3fold_free(struct z3fold_pool *pool, unsigned long handle) | |
880 | { | |
881 | struct z3fold_header *zhdr; | |
9a001fc1 VW |
882 | struct page *page; |
883 | enum buddy bud; | |
884 | ||
9a001fc1 VW |
885 | zhdr = handle_to_z3fold_header(handle); |
886 | page = virt_to_page(zhdr); | |
887 | ||
888 | if (test_bit(PAGE_HEADLESS, &page->private)) { | |
ca0246bb VW |
889 | /* if a headless page is under reclaim, just leave. |
890 | * NB: we use test_and_set_bit for a reason: if the bit | |
891 | * has not been set before, we release this page | |
892 | * immediately so we don't care about its value any more. | |
893 | */ | |
894 | if (!test_and_set_bit(PAGE_CLAIMED, &page->private)) { | |
895 | spin_lock(&pool->lock); | |
896 | list_del(&page->lru); | |
897 | spin_unlock(&pool->lock); | |
898 | free_z3fold_page(page); | |
899 | atomic64_dec(&pool->pages_nr); | |
9a001fc1 | 900 | } |
ca0246bb | 901 | return; |
9a001fc1 VW |
902 | } |
903 | ||
ca0246bb VW |
904 | /* Non-headless case */ |
905 | z3fold_page_lock(zhdr); | |
906 | bud = handle_to_buddy(handle); | |
907 | ||
908 | switch (bud) { | |
909 | case FIRST: | |
910 | zhdr->first_chunks = 0; | |
911 | break; | |
912 | case MIDDLE: | |
913 | zhdr->middle_chunks = 0; | |
914 | break; | |
915 | case LAST: | |
916 | zhdr->last_chunks = 0; | |
917 | break; | |
918 | default: | |
919 | pr_err("%s: unknown bud %d\n", __func__, bud); | |
920 | WARN_ON(1); | |
921 | z3fold_page_unlock(zhdr); | |
d30561c5 VW |
922 | return; |
923 | } | |
924 | ||
7c2b8baa | 925 | free_handle(handle); |
d30561c5 VW |
926 | if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) { |
927 | atomic64_dec(&pool->pages_nr); | |
928 | return; | |
929 | } | |
ca0246bb | 930 | if (test_bit(PAGE_CLAIMED, &page->private)) { |
6098d7e1 VW |
931 | z3fold_page_unlock(zhdr); |
932 | return; | |
933 | } | |
d30561c5 | 934 | if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) { |
5a27aa82 | 935 | z3fold_page_unlock(zhdr); |
d30561c5 VW |
936 | return; |
937 | } | |
938 | if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) { | |
2f1e5e4d | 939 | spin_lock(&pool->lock); |
d30561c5 | 940 | list_del_init(&zhdr->buddy); |
2f1e5e4d | 941 | spin_unlock(&pool->lock); |
d30561c5 | 942 | zhdr->cpu = -1; |
5d03a661 | 943 | kref_get(&zhdr->refcount); |
d30561c5 VW |
944 | do_compact_page(zhdr, true); |
945 | return; | |
9a001fc1 | 946 | } |
5d03a661 | 947 | kref_get(&zhdr->refcount); |
d30561c5 VW |
948 | queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work); |
949 | z3fold_page_unlock(zhdr); | |
9a001fc1 VW |
950 | } |
951 | ||
952 | /** | |
953 | * z3fold_reclaim_page() - evicts allocations from a pool page and frees it | |
954 | * @pool: pool from which a page will attempt to be evicted | |
f144c390 | 955 | * @retries: number of pages on the LRU list for which eviction will |
9a001fc1 VW |
956 | * be attempted before failing |
957 | * | |
958 | * z3fold reclaim is different from normal system reclaim in that it is done | |
959 | * from the bottom, up. This is because only the bottom layer, z3fold, has | |
960 | * information on how the allocations are organized within each z3fold page. | |
961 | * This has the potential to create interesting locking situations between | |
962 | * z3fold and the user, however. | |
963 | * | |
964 | * To avoid these, this is how z3fold_reclaim_page() should be called: | |
f144c390 | 965 | * |
9a001fc1 VW |
966 | * The user detects a page should be reclaimed and calls z3fold_reclaim_page(). |
967 | * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and | |
968 | * call the user-defined eviction handler with the pool and handle as | |
969 | * arguments. | |
970 | * | |
971 | * If the handle can not be evicted, the eviction handler should return | |
972 | * non-zero. z3fold_reclaim_page() will add the z3fold page back to the | |
973 | * appropriate list and try the next z3fold page on the LRU up to | |
974 | * a user defined number of retries. | |
975 | * | |
976 | * If the handle is successfully evicted, the eviction handler should | |
977 | * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free() | |
978 | * contains logic to delay freeing the page if the page is under reclaim, | |
979 | * as indicated by the setting of the PG_reclaim flag on the underlying page. | |
980 | * | |
981 | * If all buddies in the z3fold page are successfully evicted, then the | |
982 | * z3fold page can be freed. | |
983 | * | |
984 | * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are | |
985 | * no pages to evict or an eviction handler is not registered, -EAGAIN if | |
986 | * the retry limit was hit. | |
987 | */ | |
988 | static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries) | |
989 | { | |
d30561c5 VW |
990 | int i, ret = 0; |
991 | struct z3fold_header *zhdr = NULL; | |
992 | struct page *page = NULL; | |
993 | struct list_head *pos; | |
9a001fc1 VW |
994 | unsigned long first_handle = 0, middle_handle = 0, last_handle = 0; |
995 | ||
996 | spin_lock(&pool->lock); | |
2f1e5e4d | 997 | if (!pool->ops || !pool->ops->evict || retries == 0) { |
9a001fc1 VW |
998 | spin_unlock(&pool->lock); |
999 | return -EINVAL; | |
1000 | } | |
1001 | for (i = 0; i < retries; i++) { | |
2f1e5e4d VW |
1002 | if (list_empty(&pool->lru)) { |
1003 | spin_unlock(&pool->lock); | |
1004 | return -EINVAL; | |
1005 | } | |
d30561c5 VW |
1006 | list_for_each_prev(pos, &pool->lru) { |
1007 | page = list_entry(pos, struct page, lru); | |
ca0246bb VW |
1008 | |
1009 | /* this bit could have been set by free, in which case | |
1010 | * we pass over to the next page in the pool. | |
1011 | */ | |
1012 | if (test_and_set_bit(PAGE_CLAIMED, &page->private)) | |
1013 | continue; | |
1014 | ||
1015 | zhdr = page_address(page); | |
d30561c5 | 1016 | if (test_bit(PAGE_HEADLESS, &page->private)) |
d30561c5 VW |
1017 | break; |
1018 | ||
ca0246bb VW |
1019 | if (!z3fold_page_trylock(zhdr)) { |
1020 | zhdr = NULL; | |
d30561c5 | 1021 | continue; /* can't evict at this point */ |
ca0246bb | 1022 | } |
d30561c5 VW |
1023 | kref_get(&zhdr->refcount); |
1024 | list_del_init(&zhdr->buddy); | |
1025 | zhdr->cpu = -1; | |
6098d7e1 | 1026 | break; |
d30561c5 VW |
1027 | } |
1028 | ||
ca0246bb VW |
1029 | if (!zhdr) |
1030 | break; | |
1031 | ||
5a27aa82 | 1032 | list_del_init(&page->lru); |
d30561c5 | 1033 | spin_unlock(&pool->lock); |
9a001fc1 | 1034 | |
9a001fc1 | 1035 | if (!test_bit(PAGE_HEADLESS, &page->private)) { |
9a001fc1 VW |
1036 | /* |
1037 | * We need encode the handles before unlocking, since | |
1038 | * we can race with free that will set | |
1039 | * (first|last)_chunks to 0 | |
1040 | */ | |
1041 | first_handle = 0; | |
1042 | last_handle = 0; | |
1043 | middle_handle = 0; | |
1044 | if (zhdr->first_chunks) | |
1045 | first_handle = encode_handle(zhdr, FIRST); | |
1046 | if (zhdr->middle_chunks) | |
1047 | middle_handle = encode_handle(zhdr, MIDDLE); | |
1048 | if (zhdr->last_chunks) | |
1049 | last_handle = encode_handle(zhdr, LAST); | |
d30561c5 VW |
1050 | /* |
1051 | * it's safe to unlock here because we hold a | |
1052 | * reference to this page | |
1053 | */ | |
2f1e5e4d | 1054 | z3fold_page_unlock(zhdr); |
9a001fc1 VW |
1055 | } else { |
1056 | first_handle = encode_handle(zhdr, HEADLESS); | |
1057 | last_handle = middle_handle = 0; | |
1058 | } | |
1059 | ||
9a001fc1 VW |
1060 | /* Issue the eviction callback(s) */ |
1061 | if (middle_handle) { | |
1062 | ret = pool->ops->evict(pool, middle_handle); | |
1063 | if (ret) | |
1064 | goto next; | |
1065 | } | |
1066 | if (first_handle) { | |
1067 | ret = pool->ops->evict(pool, first_handle); | |
1068 | if (ret) | |
1069 | goto next; | |
1070 | } | |
1071 | if (last_handle) { | |
1072 | ret = pool->ops->evict(pool, last_handle); | |
1073 | if (ret) | |
1074 | goto next; | |
1075 | } | |
1076 | next: | |
5a27aa82 VW |
1077 | if (test_bit(PAGE_HEADLESS, &page->private)) { |
1078 | if (ret == 0) { | |
1079 | free_z3fold_page(page); | |
ca0246bb | 1080 | atomic64_dec(&pool->pages_nr); |
5a27aa82 | 1081 | return 0; |
5a27aa82 | 1082 | } |
6098d7e1 VW |
1083 | spin_lock(&pool->lock); |
1084 | list_add(&page->lru, &pool->lru); | |
1085 | spin_unlock(&pool->lock); | |
1086 | } else { | |
1087 | z3fold_page_lock(zhdr); | |
ca0246bb | 1088 | clear_bit(PAGE_CLAIMED, &page->private); |
6098d7e1 VW |
1089 | if (kref_put(&zhdr->refcount, |
1090 | release_z3fold_page_locked)) { | |
1091 | atomic64_dec(&pool->pages_nr); | |
1092 | return 0; | |
1093 | } | |
1094 | /* | |
1095 | * if we are here, the page is still not completely | |
1096 | * free. Take the global pool lock then to be able | |
1097 | * to add it back to the lru list | |
1098 | */ | |
1099 | spin_lock(&pool->lock); | |
1100 | list_add(&page->lru, &pool->lru); | |
d5567c9d | 1101 | spin_unlock(&pool->lock); |
6098d7e1 | 1102 | z3fold_page_unlock(zhdr); |
5a27aa82 | 1103 | } |
2f1e5e4d | 1104 | |
6098d7e1 VW |
1105 | /* We started off locked to we need to lock the pool back */ |
1106 | spin_lock(&pool->lock); | |
9a001fc1 VW |
1107 | } |
1108 | spin_unlock(&pool->lock); | |
1109 | return -EAGAIN; | |
1110 | } | |
1111 | ||
1112 | /** | |
1113 | * z3fold_map() - maps the allocation associated with the given handle | |
1114 | * @pool: pool in which the allocation resides | |
1115 | * @handle: handle associated with the allocation to be mapped | |
1116 | * | |
1117 | * Extracts the buddy number from handle and constructs the pointer to the | |
1118 | * correct starting chunk within the page. | |
1119 | * | |
1120 | * Returns: a pointer to the mapped allocation | |
1121 | */ | |
1122 | static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle) | |
1123 | { | |
1124 | struct z3fold_header *zhdr; | |
1125 | struct page *page; | |
1126 | void *addr; | |
1127 | enum buddy buddy; | |
1128 | ||
9a001fc1 VW |
1129 | zhdr = handle_to_z3fold_header(handle); |
1130 | addr = zhdr; | |
1131 | page = virt_to_page(zhdr); | |
1132 | ||
1133 | if (test_bit(PAGE_HEADLESS, &page->private)) | |
1134 | goto out; | |
1135 | ||
2f1e5e4d | 1136 | z3fold_page_lock(zhdr); |
9a001fc1 VW |
1137 | buddy = handle_to_buddy(handle); |
1138 | switch (buddy) { | |
1139 | case FIRST: | |
1140 | addr += ZHDR_SIZE_ALIGNED; | |
1141 | break; | |
1142 | case MIDDLE: | |
1143 | addr += zhdr->start_middle << CHUNK_SHIFT; | |
1144 | set_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
1145 | break; | |
1146 | case LAST: | |
ca0246bb | 1147 | addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT); |
9a001fc1 VW |
1148 | break; |
1149 | default: | |
1150 | pr_err("unknown buddy id %d\n", buddy); | |
1151 | WARN_ON(1); | |
1152 | addr = NULL; | |
1153 | break; | |
1154 | } | |
2f1e5e4d VW |
1155 | |
1156 | z3fold_page_unlock(zhdr); | |
9a001fc1 | 1157 | out: |
9a001fc1 VW |
1158 | return addr; |
1159 | } | |
1160 | ||
1161 | /** | |
1162 | * z3fold_unmap() - unmaps the allocation associated with the given handle | |
1163 | * @pool: pool in which the allocation resides | |
1164 | * @handle: handle associated with the allocation to be unmapped | |
1165 | */ | |
1166 | static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle) | |
1167 | { | |
1168 | struct z3fold_header *zhdr; | |
1169 | struct page *page; | |
1170 | enum buddy buddy; | |
1171 | ||
9a001fc1 VW |
1172 | zhdr = handle_to_z3fold_header(handle); |
1173 | page = virt_to_page(zhdr); | |
1174 | ||
2f1e5e4d | 1175 | if (test_bit(PAGE_HEADLESS, &page->private)) |
9a001fc1 | 1176 | return; |
9a001fc1 | 1177 | |
2f1e5e4d | 1178 | z3fold_page_lock(zhdr); |
9a001fc1 VW |
1179 | buddy = handle_to_buddy(handle); |
1180 | if (buddy == MIDDLE) | |
1181 | clear_bit(MIDDLE_CHUNK_MAPPED, &page->private); | |
2f1e5e4d | 1182 | z3fold_page_unlock(zhdr); |
9a001fc1 VW |
1183 | } |
1184 | ||
1185 | /** | |
1186 | * z3fold_get_pool_size() - gets the z3fold pool size in pages | |
1187 | * @pool: pool whose size is being queried | |
1188 | * | |
12d59ae6 | 1189 | * Returns: size in pages of the given pool. |
9a001fc1 VW |
1190 | */ |
1191 | static u64 z3fold_get_pool_size(struct z3fold_pool *pool) | |
1192 | { | |
12d59ae6 | 1193 | return atomic64_read(&pool->pages_nr); |
9a001fc1 VW |
1194 | } |
1195 | ||
1196 | /***************** | |
1197 | * zpool | |
1198 | ****************/ | |
1199 | ||
1200 | static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle) | |
1201 | { | |
1202 | if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict) | |
1203 | return pool->zpool_ops->evict(pool->zpool, handle); | |
1204 | else | |
1205 | return -ENOENT; | |
1206 | } | |
1207 | ||
1208 | static const struct z3fold_ops z3fold_zpool_ops = { | |
1209 | .evict = z3fold_zpool_evict | |
1210 | }; | |
1211 | ||
1212 | static void *z3fold_zpool_create(const char *name, gfp_t gfp, | |
1213 | const struct zpool_ops *zpool_ops, | |
1214 | struct zpool *zpool) | |
1215 | { | |
1216 | struct z3fold_pool *pool; | |
1217 | ||
d30561c5 VW |
1218 | pool = z3fold_create_pool(name, gfp, |
1219 | zpool_ops ? &z3fold_zpool_ops : NULL); | |
9a001fc1 VW |
1220 | if (pool) { |
1221 | pool->zpool = zpool; | |
1222 | pool->zpool_ops = zpool_ops; | |
1223 | } | |
1224 | return pool; | |
1225 | } | |
1226 | ||
1227 | static void z3fold_zpool_destroy(void *pool) | |
1228 | { | |
1229 | z3fold_destroy_pool(pool); | |
1230 | } | |
1231 | ||
1232 | static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
1233 | unsigned long *handle) | |
1234 | { | |
1235 | return z3fold_alloc(pool, size, gfp, handle); | |
1236 | } | |
1237 | static void z3fold_zpool_free(void *pool, unsigned long handle) | |
1238 | { | |
1239 | z3fold_free(pool, handle); | |
1240 | } | |
1241 | ||
1242 | static int z3fold_zpool_shrink(void *pool, unsigned int pages, | |
1243 | unsigned int *reclaimed) | |
1244 | { | |
1245 | unsigned int total = 0; | |
1246 | int ret = -EINVAL; | |
1247 | ||
1248 | while (total < pages) { | |
1249 | ret = z3fold_reclaim_page(pool, 8); | |
1250 | if (ret < 0) | |
1251 | break; | |
1252 | total++; | |
1253 | } | |
1254 | ||
1255 | if (reclaimed) | |
1256 | *reclaimed = total; | |
1257 | ||
1258 | return ret; | |
1259 | } | |
1260 | ||
1261 | static void *z3fold_zpool_map(void *pool, unsigned long handle, | |
1262 | enum zpool_mapmode mm) | |
1263 | { | |
1264 | return z3fold_map(pool, handle); | |
1265 | } | |
1266 | static void z3fold_zpool_unmap(void *pool, unsigned long handle) | |
1267 | { | |
1268 | z3fold_unmap(pool, handle); | |
1269 | } | |
1270 | ||
1271 | static u64 z3fold_zpool_total_size(void *pool) | |
1272 | { | |
1273 | return z3fold_get_pool_size(pool) * PAGE_SIZE; | |
1274 | } | |
1275 | ||
1276 | static struct zpool_driver z3fold_zpool_driver = { | |
1277 | .type = "z3fold", | |
1278 | .owner = THIS_MODULE, | |
1279 | .create = z3fold_zpool_create, | |
1280 | .destroy = z3fold_zpool_destroy, | |
1281 | .malloc = z3fold_zpool_malloc, | |
1282 | .free = z3fold_zpool_free, | |
1283 | .shrink = z3fold_zpool_shrink, | |
1284 | .map = z3fold_zpool_map, | |
1285 | .unmap = z3fold_zpool_unmap, | |
1286 | .total_size = z3fold_zpool_total_size, | |
1287 | }; | |
1288 | ||
1289 | MODULE_ALIAS("zpool-z3fold"); | |
1290 | ||
1291 | static int __init init_z3fold(void) | |
1292 | { | |
ede93213 VW |
1293 | /* Make sure the z3fold header is not larger than the page size */ |
1294 | BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE); | |
9a001fc1 VW |
1295 | zpool_register_driver(&z3fold_zpool_driver); |
1296 | ||
1297 | return 0; | |
1298 | } | |
1299 | ||
1300 | static void __exit exit_z3fold(void) | |
1301 | { | |
1302 | zpool_unregister_driver(&z3fold_zpool_driver); | |
1303 | } | |
1304 | ||
1305 | module_init(init_z3fold); | |
1306 | module_exit(exit_z3fold); | |
1307 | ||
1308 | MODULE_LICENSE("GPL"); | |
1309 | MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>"); | |
1310 | MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages"); |