mm/zpool: implement common zpool api to zbud/zsmalloc
[deliverable/linux.git] / mm / zsmalloc.c
1 /*
2 * zsmalloc memory allocator
3 *
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
6 *
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
9 *
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
12 */
13
14 /*
15 * This allocator is designed for use with zram. Thus, the allocator is
16 * supposed to work well under low memory conditions. In particular, it
17 * never attempts higher order page allocation which is very likely to
18 * fail under memory pressure. On the other hand, if we just use single
19 * (0-order) pages, it would suffer from very high fragmentation --
20 * any object of size PAGE_SIZE/2 or larger would occupy an entire page.
21 * This was one of the major issues with its predecessor (xvmalloc).
22 *
23 * To overcome these issues, zsmalloc allocates a bunch of 0-order pages
24 * and links them together using various 'struct page' fields. These linked
25 * pages act as a single higher-order page i.e. an object can span 0-order
26 * page boundaries. The code refers to these linked pages as a single entity
27 * called zspage.
28 *
29 * For simplicity, zsmalloc can only allocate objects of size up to PAGE_SIZE
30 * since this satisfies the requirements of all its current users (in the
31 * worst case, page is incompressible and is thus stored "as-is" i.e. in
32 * uncompressed form). For allocation requests larger than this size, failure
33 * is returned (see zs_malloc).
34 *
35 * Additionally, zs_malloc() does not return a dereferenceable pointer.
36 * Instead, it returns an opaque handle (unsigned long) which encodes actual
37 * location of the allocated object. The reason for this indirection is that
38 * zsmalloc does not keep zspages permanently mapped since that would cause
39 * issues on 32-bit systems where the VA region for kernel space mappings
40 * is very small. So, before using the allocating memory, the object has to
41 * be mapped using zs_map_object() to get a usable pointer and subsequently
42 * unmapped using zs_unmap_object().
43 *
44 * Following is how we use various fields and flags of underlying
45 * struct page(s) to form a zspage.
46 *
47 * Usage of struct page fields:
48 * page->first_page: points to the first component (0-order) page
49 * page->index (union with page->freelist): offset of the first object
50 * starting in this page. For the first page, this is
51 * always 0, so we use this field (aka freelist) to point
52 * to the first free object in zspage.
53 * page->lru: links together all component pages (except the first page)
54 * of a zspage
55 *
56 * For _first_ page only:
57 *
58 * page->private (union with page->first_page): refers to the
59 * component page after the first page
60 * page->freelist: points to the first free object in zspage.
61 * Free objects are linked together using in-place
62 * metadata.
63 * page->objects: maximum number of objects we can store in this
64 * zspage (class->zspage_order * PAGE_SIZE / class->size)
65 * page->lru: links together first pages of various zspages.
66 * Basically forming list of zspages in a fullness group.
67 * page->mapping: class index and fullness group of the zspage
68 *
69 * Usage of struct page flags:
70 * PG_private: identifies the first component page
71 * PG_private2: identifies the last component page
72 *
73 */
74
75 #ifdef CONFIG_ZSMALLOC_DEBUG
76 #define DEBUG
77 #endif
78
79 #include <linux/module.h>
80 #include <linux/kernel.h>
81 #include <linux/bitops.h>
82 #include <linux/errno.h>
83 #include <linux/highmem.h>
84 #include <linux/string.h>
85 #include <linux/slab.h>
86 #include <asm/tlbflush.h>
87 #include <asm/pgtable.h>
88 #include <linux/cpumask.h>
89 #include <linux/cpu.h>
90 #include <linux/vmalloc.h>
91 #include <linux/hardirq.h>
92 #include <linux/spinlock.h>
93 #include <linux/types.h>
94 #include <linux/zsmalloc.h>
95
96 /*
97 * This must be power of 2 and greater than of equal to sizeof(link_free).
98 * These two conditions ensure that any 'struct link_free' itself doesn't
99 * span more than 1 page which avoids complex case of mapping 2 pages simply
100 * to restore link_free pointer values.
101 */
102 #define ZS_ALIGN 8
103
104 /*
105 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
106 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
107 */
108 #define ZS_MAX_ZSPAGE_ORDER 2
109 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
110
111 /*
112 * Object location (<PFN>, <obj_idx>) is encoded as
113 * as single (unsigned long) handle value.
114 *
115 * Note that object index <obj_idx> is relative to system
116 * page <PFN> it is stored in, so for each sub-page belonging
117 * to a zspage, obj_idx starts with 0.
118 *
119 * This is made more complicated by various memory models and PAE.
120 */
121
122 #ifndef MAX_PHYSMEM_BITS
123 #ifdef CONFIG_HIGHMEM64G
124 #define MAX_PHYSMEM_BITS 36
125 #else /* !CONFIG_HIGHMEM64G */
126 /*
127 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
128 * be PAGE_SHIFT
129 */
130 #define MAX_PHYSMEM_BITS BITS_PER_LONG
131 #endif
132 #endif
133 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
134 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
135 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
136
137 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
138 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
139 #define ZS_MIN_ALLOC_SIZE \
140 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
141 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
142
143 /*
144 * On systems with 4K page size, this gives 255 size classes! There is a
145 * trader-off here:
146 * - Large number of size classes is potentially wasteful as free page are
147 * spread across these classes
148 * - Small number of size classes causes large internal fragmentation
149 * - Probably its better to use specific size classes (empirically
150 * determined). NOTE: all those class sizes must be set as multiple of
151 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
152 *
153 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
154 * (reason above)
155 */
156 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
157 #define ZS_SIZE_CLASSES ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / \
158 ZS_SIZE_CLASS_DELTA + 1)
159
160 /*
161 * We do not maintain any list for completely empty or full pages
162 */
163 enum fullness_group {
164 ZS_ALMOST_FULL,
165 ZS_ALMOST_EMPTY,
166 _ZS_NR_FULLNESS_GROUPS,
167
168 ZS_EMPTY,
169 ZS_FULL
170 };
171
172 /*
173 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
174 * n <= N / f, where
175 * n = number of allocated objects
176 * N = total number of objects zspage can store
177 * f = 1/fullness_threshold_frac
178 *
179 * Similarly, we assign zspage to:
180 * ZS_ALMOST_FULL when n > N / f
181 * ZS_EMPTY when n == 0
182 * ZS_FULL when n == N
183 *
184 * (see: fix_fullness_group())
185 */
186 static const int fullness_threshold_frac = 4;
187
188 struct size_class {
189 /*
190 * Size of objects stored in this class. Must be multiple
191 * of ZS_ALIGN.
192 */
193 int size;
194 unsigned int index;
195
196 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
197 int pages_per_zspage;
198
199 spinlock_t lock;
200
201 /* stats */
202 u64 pages_allocated;
203
204 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
205 };
206
207 /*
208 * Placed within free objects to form a singly linked list.
209 * For every zspage, first_page->freelist gives head of this list.
210 *
211 * This must be power of 2 and less than or equal to ZS_ALIGN
212 */
213 struct link_free {
214 /* Handle of next free chunk (encodes <PFN, obj_idx>) */
215 void *next;
216 };
217
218 struct zs_pool {
219 struct size_class size_class[ZS_SIZE_CLASSES];
220
221 gfp_t flags; /* allocation flags used when growing pool */
222 };
223
224 /*
225 * A zspage's class index and fullness group
226 * are encoded in its (first)page->mapping
227 */
228 #define CLASS_IDX_BITS 28
229 #define FULLNESS_BITS 4
230 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
231 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
232
233 struct mapping_area {
234 #ifdef CONFIG_PGTABLE_MAPPING
235 struct vm_struct *vm; /* vm area for mapping object that span pages */
236 #else
237 char *vm_buf; /* copy buffer for objects that span pages */
238 #endif
239 char *vm_addr; /* address of kmap_atomic()'ed pages */
240 enum zs_mapmode vm_mm; /* mapping mode */
241 };
242
243 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
244 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
245
246 static int is_first_page(struct page *page)
247 {
248 return PagePrivate(page);
249 }
250
251 static int is_last_page(struct page *page)
252 {
253 return PagePrivate2(page);
254 }
255
256 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
257 enum fullness_group *fullness)
258 {
259 unsigned long m;
260 BUG_ON(!is_first_page(page));
261
262 m = (unsigned long)page->mapping;
263 *fullness = m & FULLNESS_MASK;
264 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
265 }
266
267 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
268 enum fullness_group fullness)
269 {
270 unsigned long m;
271 BUG_ON(!is_first_page(page));
272
273 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
274 (fullness & FULLNESS_MASK);
275 page->mapping = (struct address_space *)m;
276 }
277
278 /*
279 * zsmalloc divides the pool into various size classes where each
280 * class maintains a list of zspages where each zspage is divided
281 * into equal sized chunks. Each allocation falls into one of these
282 * classes depending on its size. This function returns index of the
283 * size class which has chunk size big enough to hold the give size.
284 */
285 static int get_size_class_index(int size)
286 {
287 int idx = 0;
288
289 if (likely(size > ZS_MIN_ALLOC_SIZE))
290 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
291 ZS_SIZE_CLASS_DELTA);
292
293 return idx;
294 }
295
296 /*
297 * For each size class, zspages are divided into different groups
298 * depending on how "full" they are. This was done so that we could
299 * easily find empty or nearly empty zspages when we try to shrink
300 * the pool (not yet implemented). This function returns fullness
301 * status of the given page.
302 */
303 static enum fullness_group get_fullness_group(struct page *page)
304 {
305 int inuse, max_objects;
306 enum fullness_group fg;
307 BUG_ON(!is_first_page(page));
308
309 inuse = page->inuse;
310 max_objects = page->objects;
311
312 if (inuse == 0)
313 fg = ZS_EMPTY;
314 else if (inuse == max_objects)
315 fg = ZS_FULL;
316 else if (inuse <= max_objects / fullness_threshold_frac)
317 fg = ZS_ALMOST_EMPTY;
318 else
319 fg = ZS_ALMOST_FULL;
320
321 return fg;
322 }
323
324 /*
325 * Each size class maintains various freelists and zspages are assigned
326 * to one of these freelists based on the number of live objects they
327 * have. This functions inserts the given zspage into the freelist
328 * identified by <class, fullness_group>.
329 */
330 static void insert_zspage(struct page *page, struct size_class *class,
331 enum fullness_group fullness)
332 {
333 struct page **head;
334
335 BUG_ON(!is_first_page(page));
336
337 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
338 return;
339
340 head = &class->fullness_list[fullness];
341 if (*head)
342 list_add_tail(&page->lru, &(*head)->lru);
343
344 *head = page;
345 }
346
347 /*
348 * This function removes the given zspage from the freelist identified
349 * by <class, fullness_group>.
350 */
351 static void remove_zspage(struct page *page, struct size_class *class,
352 enum fullness_group fullness)
353 {
354 struct page **head;
355
356 BUG_ON(!is_first_page(page));
357
358 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
359 return;
360
361 head = &class->fullness_list[fullness];
362 BUG_ON(!*head);
363 if (list_empty(&(*head)->lru))
364 *head = NULL;
365 else if (*head == page)
366 *head = (struct page *)list_entry((*head)->lru.next,
367 struct page, lru);
368
369 list_del_init(&page->lru);
370 }
371
372 /*
373 * Each size class maintains zspages in different fullness groups depending
374 * on the number of live objects they contain. When allocating or freeing
375 * objects, the fullness status of the page can change, say, from ALMOST_FULL
376 * to ALMOST_EMPTY when freeing an object. This function checks if such
377 * a status change has occurred for the given page and accordingly moves the
378 * page from the freelist of the old fullness group to that of the new
379 * fullness group.
380 */
381 static enum fullness_group fix_fullness_group(struct zs_pool *pool,
382 struct page *page)
383 {
384 int class_idx;
385 struct size_class *class;
386 enum fullness_group currfg, newfg;
387
388 BUG_ON(!is_first_page(page));
389
390 get_zspage_mapping(page, &class_idx, &currfg);
391 newfg = get_fullness_group(page);
392 if (newfg == currfg)
393 goto out;
394
395 class = &pool->size_class[class_idx];
396 remove_zspage(page, class, currfg);
397 insert_zspage(page, class, newfg);
398 set_zspage_mapping(page, class_idx, newfg);
399
400 out:
401 return newfg;
402 }
403
404 /*
405 * We have to decide on how many pages to link together
406 * to form a zspage for each size class. This is important
407 * to reduce wastage due to unusable space left at end of
408 * each zspage which is given as:
409 * wastage = Zp - Zp % size_class
410 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
411 *
412 * For example, for size class of 3/8 * PAGE_SIZE, we should
413 * link together 3 PAGE_SIZE sized pages to form a zspage
414 * since then we can perfectly fit in 8 such objects.
415 */
416 static int get_pages_per_zspage(int class_size)
417 {
418 int i, max_usedpc = 0;
419 /* zspage order which gives maximum used size per KB */
420 int max_usedpc_order = 1;
421
422 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
423 int zspage_size;
424 int waste, usedpc;
425
426 zspage_size = i * PAGE_SIZE;
427 waste = zspage_size % class_size;
428 usedpc = (zspage_size - waste) * 100 / zspage_size;
429
430 if (usedpc > max_usedpc) {
431 max_usedpc = usedpc;
432 max_usedpc_order = i;
433 }
434 }
435
436 return max_usedpc_order;
437 }
438
439 /*
440 * A single 'zspage' is composed of many system pages which are
441 * linked together using fields in struct page. This function finds
442 * the first/head page, given any component page of a zspage.
443 */
444 static struct page *get_first_page(struct page *page)
445 {
446 if (is_first_page(page))
447 return page;
448 else
449 return page->first_page;
450 }
451
452 static struct page *get_next_page(struct page *page)
453 {
454 struct page *next;
455
456 if (is_last_page(page))
457 next = NULL;
458 else if (is_first_page(page))
459 next = (struct page *)page_private(page);
460 else
461 next = list_entry(page->lru.next, struct page, lru);
462
463 return next;
464 }
465
466 /*
467 * Encode <page, obj_idx> as a single handle value.
468 * On hardware platforms with physical memory starting at 0x0 the pfn
469 * could be 0 so we ensure that the handle will never be 0 by adjusting the
470 * encoded obj_idx value before encoding.
471 */
472 static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
473 {
474 unsigned long handle;
475
476 if (!page) {
477 BUG_ON(obj_idx);
478 return NULL;
479 }
480
481 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
482 handle |= ((obj_idx + 1) & OBJ_INDEX_MASK);
483
484 return (void *)handle;
485 }
486
487 /*
488 * Decode <page, obj_idx> pair from the given object handle. We adjust the
489 * decoded obj_idx back to its original value since it was adjusted in
490 * obj_location_to_handle().
491 */
492 static void obj_handle_to_location(unsigned long handle, struct page **page,
493 unsigned long *obj_idx)
494 {
495 *page = pfn_to_page(handle >> OBJ_INDEX_BITS);
496 *obj_idx = (handle & OBJ_INDEX_MASK) - 1;
497 }
498
499 static unsigned long obj_idx_to_offset(struct page *page,
500 unsigned long obj_idx, int class_size)
501 {
502 unsigned long off = 0;
503
504 if (!is_first_page(page))
505 off = page->index;
506
507 return off + obj_idx * class_size;
508 }
509
510 static void reset_page(struct page *page)
511 {
512 clear_bit(PG_private, &page->flags);
513 clear_bit(PG_private_2, &page->flags);
514 set_page_private(page, 0);
515 page->mapping = NULL;
516 page->freelist = NULL;
517 page_mapcount_reset(page);
518 }
519
520 static void free_zspage(struct page *first_page)
521 {
522 struct page *nextp, *tmp, *head_extra;
523
524 BUG_ON(!is_first_page(first_page));
525 BUG_ON(first_page->inuse);
526
527 head_extra = (struct page *)page_private(first_page);
528
529 reset_page(first_page);
530 __free_page(first_page);
531
532 /* zspage with only 1 system page */
533 if (!head_extra)
534 return;
535
536 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
537 list_del(&nextp->lru);
538 reset_page(nextp);
539 __free_page(nextp);
540 }
541 reset_page(head_extra);
542 __free_page(head_extra);
543 }
544
545 /* Initialize a newly allocated zspage */
546 static void init_zspage(struct page *first_page, struct size_class *class)
547 {
548 unsigned long off = 0;
549 struct page *page = first_page;
550
551 BUG_ON(!is_first_page(first_page));
552 while (page) {
553 struct page *next_page;
554 struct link_free *link;
555 unsigned int i, objs_on_page;
556
557 /*
558 * page->index stores offset of first object starting
559 * in the page. For the first page, this is always 0,
560 * so we use first_page->index (aka ->freelist) to store
561 * head of corresponding zspage's freelist.
562 */
563 if (page != first_page)
564 page->index = off;
565
566 link = (struct link_free *)kmap_atomic(page) +
567 off / sizeof(*link);
568 objs_on_page = (PAGE_SIZE - off) / class->size;
569
570 for (i = 1; i <= objs_on_page; i++) {
571 off += class->size;
572 if (off < PAGE_SIZE) {
573 link->next = obj_location_to_handle(page, i);
574 link += class->size / sizeof(*link);
575 }
576 }
577
578 /*
579 * We now come to the last (full or partial) object on this
580 * page, which must point to the first object on the next
581 * page (if present)
582 */
583 next_page = get_next_page(page);
584 link->next = obj_location_to_handle(next_page, 0);
585 kunmap_atomic(link);
586 page = next_page;
587 off = (off + class->size) % PAGE_SIZE;
588 }
589 }
590
591 /*
592 * Allocate a zspage for the given size class
593 */
594 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
595 {
596 int i, error;
597 struct page *first_page = NULL, *uninitialized_var(prev_page);
598
599 /*
600 * Allocate individual pages and link them together as:
601 * 1. first page->private = first sub-page
602 * 2. all sub-pages are linked together using page->lru
603 * 3. each sub-page is linked to the first page using page->first_page
604 *
605 * For each size class, First/Head pages are linked together using
606 * page->lru. Also, we set PG_private to identify the first page
607 * (i.e. no other sub-page has this flag set) and PG_private_2 to
608 * identify the last page.
609 */
610 error = -ENOMEM;
611 for (i = 0; i < class->pages_per_zspage; i++) {
612 struct page *page;
613
614 page = alloc_page(flags);
615 if (!page)
616 goto cleanup;
617
618 INIT_LIST_HEAD(&page->lru);
619 if (i == 0) { /* first page */
620 SetPagePrivate(page);
621 set_page_private(page, 0);
622 first_page = page;
623 first_page->inuse = 0;
624 }
625 if (i == 1)
626 set_page_private(first_page, (unsigned long)page);
627 if (i >= 1)
628 page->first_page = first_page;
629 if (i >= 2)
630 list_add(&page->lru, &prev_page->lru);
631 if (i == class->pages_per_zspage - 1) /* last page */
632 SetPagePrivate2(page);
633 prev_page = page;
634 }
635
636 init_zspage(first_page, class);
637
638 first_page->freelist = obj_location_to_handle(first_page, 0);
639 /* Maximum number of objects we can store in this zspage */
640 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
641
642 error = 0; /* Success */
643
644 cleanup:
645 if (unlikely(error) && first_page) {
646 free_zspage(first_page);
647 first_page = NULL;
648 }
649
650 return first_page;
651 }
652
653 static struct page *find_get_zspage(struct size_class *class)
654 {
655 int i;
656 struct page *page;
657
658 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
659 page = class->fullness_list[i];
660 if (page)
661 break;
662 }
663
664 return page;
665 }
666
667 #ifdef CONFIG_PGTABLE_MAPPING
668 static inline int __zs_cpu_up(struct mapping_area *area)
669 {
670 /*
671 * Make sure we don't leak memory if a cpu UP notification
672 * and zs_init() race and both call zs_cpu_up() on the same cpu
673 */
674 if (area->vm)
675 return 0;
676 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
677 if (!area->vm)
678 return -ENOMEM;
679 return 0;
680 }
681
682 static inline void __zs_cpu_down(struct mapping_area *area)
683 {
684 if (area->vm)
685 free_vm_area(area->vm);
686 area->vm = NULL;
687 }
688
689 static inline void *__zs_map_object(struct mapping_area *area,
690 struct page *pages[2], int off, int size)
691 {
692 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
693 area->vm_addr = area->vm->addr;
694 return area->vm_addr + off;
695 }
696
697 static inline void __zs_unmap_object(struct mapping_area *area,
698 struct page *pages[2], int off, int size)
699 {
700 unsigned long addr = (unsigned long)area->vm_addr;
701
702 unmap_kernel_range(addr, PAGE_SIZE * 2);
703 }
704
705 #else /* CONFIG_PGTABLE_MAPPING */
706
707 static inline int __zs_cpu_up(struct mapping_area *area)
708 {
709 /*
710 * Make sure we don't leak memory if a cpu UP notification
711 * and zs_init() race and both call zs_cpu_up() on the same cpu
712 */
713 if (area->vm_buf)
714 return 0;
715 area->vm_buf = (char *)__get_free_page(GFP_KERNEL);
716 if (!area->vm_buf)
717 return -ENOMEM;
718 return 0;
719 }
720
721 static inline void __zs_cpu_down(struct mapping_area *area)
722 {
723 if (area->vm_buf)
724 free_page((unsigned long)area->vm_buf);
725 area->vm_buf = NULL;
726 }
727
728 static void *__zs_map_object(struct mapping_area *area,
729 struct page *pages[2], int off, int size)
730 {
731 int sizes[2];
732 void *addr;
733 char *buf = area->vm_buf;
734
735 /* disable page faults to match kmap_atomic() return conditions */
736 pagefault_disable();
737
738 /* no read fastpath */
739 if (area->vm_mm == ZS_MM_WO)
740 goto out;
741
742 sizes[0] = PAGE_SIZE - off;
743 sizes[1] = size - sizes[0];
744
745 /* copy object to per-cpu buffer */
746 addr = kmap_atomic(pages[0]);
747 memcpy(buf, addr + off, sizes[0]);
748 kunmap_atomic(addr);
749 addr = kmap_atomic(pages[1]);
750 memcpy(buf + sizes[0], addr, sizes[1]);
751 kunmap_atomic(addr);
752 out:
753 return area->vm_buf;
754 }
755
756 static void __zs_unmap_object(struct mapping_area *area,
757 struct page *pages[2], int off, int size)
758 {
759 int sizes[2];
760 void *addr;
761 char *buf = area->vm_buf;
762
763 /* no write fastpath */
764 if (area->vm_mm == ZS_MM_RO)
765 goto out;
766
767 sizes[0] = PAGE_SIZE - off;
768 sizes[1] = size - sizes[0];
769
770 /* copy per-cpu buffer to object */
771 addr = kmap_atomic(pages[0]);
772 memcpy(addr + off, buf, sizes[0]);
773 kunmap_atomic(addr);
774 addr = kmap_atomic(pages[1]);
775 memcpy(addr, buf + sizes[0], sizes[1]);
776 kunmap_atomic(addr);
777
778 out:
779 /* enable page faults to match kunmap_atomic() return conditions */
780 pagefault_enable();
781 }
782
783 #endif /* CONFIG_PGTABLE_MAPPING */
784
785 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
786 void *pcpu)
787 {
788 int ret, cpu = (long)pcpu;
789 struct mapping_area *area;
790
791 switch (action) {
792 case CPU_UP_PREPARE:
793 area = &per_cpu(zs_map_area, cpu);
794 ret = __zs_cpu_up(area);
795 if (ret)
796 return notifier_from_errno(ret);
797 break;
798 case CPU_DEAD:
799 case CPU_UP_CANCELED:
800 area = &per_cpu(zs_map_area, cpu);
801 __zs_cpu_down(area);
802 break;
803 }
804
805 return NOTIFY_OK;
806 }
807
808 static struct notifier_block zs_cpu_nb = {
809 .notifier_call = zs_cpu_notifier
810 };
811
812 static void zs_exit(void)
813 {
814 int cpu;
815
816 cpu_notifier_register_begin();
817
818 for_each_online_cpu(cpu)
819 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
820 __unregister_cpu_notifier(&zs_cpu_nb);
821
822 cpu_notifier_register_done();
823 }
824
825 static int zs_init(void)
826 {
827 int cpu, ret;
828
829 cpu_notifier_register_begin();
830
831 __register_cpu_notifier(&zs_cpu_nb);
832 for_each_online_cpu(cpu) {
833 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
834 if (notifier_to_errno(ret)) {
835 cpu_notifier_register_done();
836 goto fail;
837 }
838 }
839
840 cpu_notifier_register_done();
841
842 return 0;
843 fail:
844 zs_exit();
845 return notifier_to_errno(ret);
846 }
847
848 /**
849 * zs_create_pool - Creates an allocation pool to work from.
850 * @flags: allocation flags used to allocate pool metadata
851 *
852 * This function must be called before anything when using
853 * the zsmalloc allocator.
854 *
855 * On success, a pointer to the newly created pool is returned,
856 * otherwise NULL.
857 */
858 struct zs_pool *zs_create_pool(gfp_t flags)
859 {
860 int i, ovhd_size;
861 struct zs_pool *pool;
862
863 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
864 pool = kzalloc(ovhd_size, GFP_KERNEL);
865 if (!pool)
866 return NULL;
867
868 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
869 int size;
870 struct size_class *class;
871
872 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
873 if (size > ZS_MAX_ALLOC_SIZE)
874 size = ZS_MAX_ALLOC_SIZE;
875
876 class = &pool->size_class[i];
877 class->size = size;
878 class->index = i;
879 spin_lock_init(&class->lock);
880 class->pages_per_zspage = get_pages_per_zspage(size);
881
882 }
883
884 pool->flags = flags;
885
886 return pool;
887 }
888 EXPORT_SYMBOL_GPL(zs_create_pool);
889
890 void zs_destroy_pool(struct zs_pool *pool)
891 {
892 int i;
893
894 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
895 int fg;
896 struct size_class *class = &pool->size_class[i];
897
898 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
899 if (class->fullness_list[fg]) {
900 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
901 class->size, fg);
902 }
903 }
904 }
905 kfree(pool);
906 }
907 EXPORT_SYMBOL_GPL(zs_destroy_pool);
908
909 /**
910 * zs_malloc - Allocate block of given size from pool.
911 * @pool: pool to allocate from
912 * @size: size of block to allocate
913 *
914 * On success, handle to the allocated object is returned,
915 * otherwise 0.
916 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
917 */
918 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
919 {
920 unsigned long obj;
921 struct link_free *link;
922 int class_idx;
923 struct size_class *class;
924
925 struct page *first_page, *m_page;
926 unsigned long m_objidx, m_offset;
927
928 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
929 return 0;
930
931 class_idx = get_size_class_index(size);
932 class = &pool->size_class[class_idx];
933 BUG_ON(class_idx != class->index);
934
935 spin_lock(&class->lock);
936 first_page = find_get_zspage(class);
937
938 if (!first_page) {
939 spin_unlock(&class->lock);
940 first_page = alloc_zspage(class, pool->flags);
941 if (unlikely(!first_page))
942 return 0;
943
944 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
945 spin_lock(&class->lock);
946 class->pages_allocated += class->pages_per_zspage;
947 }
948
949 obj = (unsigned long)first_page->freelist;
950 obj_handle_to_location(obj, &m_page, &m_objidx);
951 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
952
953 link = (struct link_free *)kmap_atomic(m_page) +
954 m_offset / sizeof(*link);
955 first_page->freelist = link->next;
956 memset(link, POISON_INUSE, sizeof(*link));
957 kunmap_atomic(link);
958
959 first_page->inuse++;
960 /* Now move the zspage to another fullness group, if required */
961 fix_fullness_group(pool, first_page);
962 spin_unlock(&class->lock);
963
964 return obj;
965 }
966 EXPORT_SYMBOL_GPL(zs_malloc);
967
968 void zs_free(struct zs_pool *pool, unsigned long obj)
969 {
970 struct link_free *link;
971 struct page *first_page, *f_page;
972 unsigned long f_objidx, f_offset;
973
974 int class_idx;
975 struct size_class *class;
976 enum fullness_group fullness;
977
978 if (unlikely(!obj))
979 return;
980
981 obj_handle_to_location(obj, &f_page, &f_objidx);
982 first_page = get_first_page(f_page);
983
984 get_zspage_mapping(first_page, &class_idx, &fullness);
985 class = &pool->size_class[class_idx];
986 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
987
988 spin_lock(&class->lock);
989
990 /* Insert this object in containing zspage's freelist */
991 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
992 + f_offset);
993 link->next = first_page->freelist;
994 kunmap_atomic(link);
995 first_page->freelist = (void *)obj;
996
997 first_page->inuse--;
998 fullness = fix_fullness_group(pool, first_page);
999
1000 if (fullness == ZS_EMPTY)
1001 class->pages_allocated -= class->pages_per_zspage;
1002
1003 spin_unlock(&class->lock);
1004
1005 if (fullness == ZS_EMPTY)
1006 free_zspage(first_page);
1007 }
1008 EXPORT_SYMBOL_GPL(zs_free);
1009
1010 /**
1011 * zs_map_object - get address of allocated object from handle.
1012 * @pool: pool from which the object was allocated
1013 * @handle: handle returned from zs_malloc
1014 *
1015 * Before using an object allocated from zs_malloc, it must be mapped using
1016 * this function. When done with the object, it must be unmapped using
1017 * zs_unmap_object.
1018 *
1019 * Only one object can be mapped per cpu at a time. There is no protection
1020 * against nested mappings.
1021 *
1022 * This function returns with preemption and page faults disabled.
1023 */
1024 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1025 enum zs_mapmode mm)
1026 {
1027 struct page *page;
1028 unsigned long obj_idx, off;
1029
1030 unsigned int class_idx;
1031 enum fullness_group fg;
1032 struct size_class *class;
1033 struct mapping_area *area;
1034 struct page *pages[2];
1035
1036 BUG_ON(!handle);
1037
1038 /*
1039 * Because we use per-cpu mapping areas shared among the
1040 * pools/users, we can't allow mapping in interrupt context
1041 * because it can corrupt another users mappings.
1042 */
1043 BUG_ON(in_interrupt());
1044
1045 obj_handle_to_location(handle, &page, &obj_idx);
1046 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1047 class = &pool->size_class[class_idx];
1048 off = obj_idx_to_offset(page, obj_idx, class->size);
1049
1050 area = &get_cpu_var(zs_map_area);
1051 area->vm_mm = mm;
1052 if (off + class->size <= PAGE_SIZE) {
1053 /* this object is contained entirely within a page */
1054 area->vm_addr = kmap_atomic(page);
1055 return area->vm_addr + off;
1056 }
1057
1058 /* this object spans two pages */
1059 pages[0] = page;
1060 pages[1] = get_next_page(page);
1061 BUG_ON(!pages[1]);
1062
1063 return __zs_map_object(area, pages, off, class->size);
1064 }
1065 EXPORT_SYMBOL_GPL(zs_map_object);
1066
1067 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1068 {
1069 struct page *page;
1070 unsigned long obj_idx, off;
1071
1072 unsigned int class_idx;
1073 enum fullness_group fg;
1074 struct size_class *class;
1075 struct mapping_area *area;
1076
1077 BUG_ON(!handle);
1078
1079 obj_handle_to_location(handle, &page, &obj_idx);
1080 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1081 class = &pool->size_class[class_idx];
1082 off = obj_idx_to_offset(page, obj_idx, class->size);
1083
1084 area = this_cpu_ptr(&zs_map_area);
1085 if (off + class->size <= PAGE_SIZE)
1086 kunmap_atomic(area->vm_addr);
1087 else {
1088 struct page *pages[2];
1089
1090 pages[0] = page;
1091 pages[1] = get_next_page(page);
1092 BUG_ON(!pages[1]);
1093
1094 __zs_unmap_object(area, pages, off, class->size);
1095 }
1096 put_cpu_var(zs_map_area);
1097 }
1098 EXPORT_SYMBOL_GPL(zs_unmap_object);
1099
1100 u64 zs_get_total_size_bytes(struct zs_pool *pool)
1101 {
1102 int i;
1103 u64 npages = 0;
1104
1105 for (i = 0; i < ZS_SIZE_CLASSES; i++)
1106 npages += pool->size_class[i].pages_allocated;
1107
1108 return npages << PAGE_SHIFT;
1109 }
1110 EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);
1111
1112 module_init(zs_init);
1113 module_exit(zs_exit);
1114
1115 MODULE_LICENSE("Dual BSD/GPL");
1116 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");
This page took 0.067633 seconds and 5 git commands to generate.