Loading...
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 254 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 */
163enum 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 */
186static const int fullness_threshold_frac = 4;
187
188struct 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 */
213struct link_free {
214 /* Handle of next free chunk (encodes <PFN, obj_idx>) */
215 void *next;
216};
217
218struct 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
233struct 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
244/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
245static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
246
247static int is_first_page(struct page *page)
248{
249 return PagePrivate(page);
250}
251
252static int is_last_page(struct page *page)
253{
254 return PagePrivate2(page);
255}
256
257static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
258 enum fullness_group *fullness)
259{
260 unsigned long m;
261 BUG_ON(!is_first_page(page));
262
263 m = (unsigned long)page->mapping;
264 *fullness = m & FULLNESS_MASK;
265 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
266}
267
268static void set_zspage_mapping(struct page *page, unsigned int class_idx,
269 enum fullness_group fullness)
270{
271 unsigned long m;
272 BUG_ON(!is_first_page(page));
273
274 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
275 (fullness & FULLNESS_MASK);
276 page->mapping = (struct address_space *)m;
277}
278
279/*
280 * zsmalloc divides the pool into various size classes where each
281 * class maintains a list of zspages where each zspage is divided
282 * into equal sized chunks. Each allocation falls into one of these
283 * classes depending on its size. This function returns index of the
284 * size class which has chunk size big enough to hold the give size.
285 */
286static int get_size_class_index(int size)
287{
288 int idx = 0;
289
290 if (likely(size > ZS_MIN_ALLOC_SIZE))
291 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
292 ZS_SIZE_CLASS_DELTA);
293
294 return idx;
295}
296
297/*
298 * For each size class, zspages are divided into different groups
299 * depending on how "full" they are. This was done so that we could
300 * easily find empty or nearly empty zspages when we try to shrink
301 * the pool (not yet implemented). This function returns fullness
302 * status of the given page.
303 */
304static enum fullness_group get_fullness_group(struct page *page)
305{
306 int inuse, max_objects;
307 enum fullness_group fg;
308 BUG_ON(!is_first_page(page));
309
310 inuse = page->inuse;
311 max_objects = page->objects;
312
313 if (inuse == 0)
314 fg = ZS_EMPTY;
315 else if (inuse == max_objects)
316 fg = ZS_FULL;
317 else if (inuse <= max_objects / fullness_threshold_frac)
318 fg = ZS_ALMOST_EMPTY;
319 else
320 fg = ZS_ALMOST_FULL;
321
322 return fg;
323}
324
325/*
326 * Each size class maintains various freelists and zspages are assigned
327 * to one of these freelists based on the number of live objects they
328 * have. This functions inserts the given zspage into the freelist
329 * identified by <class, fullness_group>.
330 */
331static void insert_zspage(struct page *page, struct size_class *class,
332 enum fullness_group fullness)
333{
334 struct page **head;
335
336 BUG_ON(!is_first_page(page));
337
338 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
339 return;
340
341 head = &class->fullness_list[fullness];
342 if (*head)
343 list_add_tail(&page->lru, &(*head)->lru);
344
345 *head = page;
346}
347
348/*
349 * This function removes the given zspage from the freelist identified
350 * by <class, fullness_group>.
351 */
352static void remove_zspage(struct page *page, struct size_class *class,
353 enum fullness_group fullness)
354{
355 struct page **head;
356
357 BUG_ON(!is_first_page(page));
358
359 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
360 return;
361
362 head = &class->fullness_list[fullness];
363 BUG_ON(!*head);
364 if (list_empty(&(*head)->lru))
365 *head = NULL;
366 else if (*head == page)
367 *head = (struct page *)list_entry((*head)->lru.next,
368 struct page, lru);
369
370 list_del_init(&page->lru);
371}
372
373/*
374 * Each size class maintains zspages in different fullness groups depending
375 * on the number of live objects they contain. When allocating or freeing
376 * objects, the fullness status of the page can change, say, from ALMOST_FULL
377 * to ALMOST_EMPTY when freeing an object. This function checks if such
378 * a status change has occurred for the given page and accordingly moves the
379 * page from the freelist of the old fullness group to that of the new
380 * fullness group.
381 */
382static enum fullness_group fix_fullness_group(struct zs_pool *pool,
383 struct page *page)
384{
385 int class_idx;
386 struct size_class *class;
387 enum fullness_group currfg, newfg;
388
389 BUG_ON(!is_first_page(page));
390
391 get_zspage_mapping(page, &class_idx, &currfg);
392 newfg = get_fullness_group(page);
393 if (newfg == currfg)
394 goto out;
395
396 class = &pool->size_class[class_idx];
397 remove_zspage(page, class, currfg);
398 insert_zspage(page, class, newfg);
399 set_zspage_mapping(page, class_idx, newfg);
400
401out:
402 return newfg;
403}
404
405/*
406 * We have to decide on how many pages to link together
407 * to form a zspage for each size class. This is important
408 * to reduce wastage due to unusable space left at end of
409 * each zspage which is given as:
410 * wastage = Zp - Zp % size_class
411 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
412 *
413 * For example, for size class of 3/8 * PAGE_SIZE, we should
414 * link together 3 PAGE_SIZE sized pages to form a zspage
415 * since then we can perfectly fit in 8 such objects.
416 */
417static int get_pages_per_zspage(int class_size)
418{
419 int i, max_usedpc = 0;
420 /* zspage order which gives maximum used size per KB */
421 int max_usedpc_order = 1;
422
423 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
424 int zspage_size;
425 int waste, usedpc;
426
427 zspage_size = i * PAGE_SIZE;
428 waste = zspage_size % class_size;
429 usedpc = (zspage_size - waste) * 100 / zspage_size;
430
431 if (usedpc > max_usedpc) {
432 max_usedpc = usedpc;
433 max_usedpc_order = i;
434 }
435 }
436
437 return max_usedpc_order;
438}
439
440/*
441 * A single 'zspage' is composed of many system pages which are
442 * linked together using fields in struct page. This function finds
443 * the first/head page, given any component page of a zspage.
444 */
445static struct page *get_first_page(struct page *page)
446{
447 if (is_first_page(page))
448 return page;
449 else
450 return page->first_page;
451}
452
453static struct page *get_next_page(struct page *page)
454{
455 struct page *next;
456
457 if (is_last_page(page))
458 next = NULL;
459 else if (is_first_page(page))
460 next = (struct page *)page_private(page);
461 else
462 next = list_entry(page->lru.next, struct page, lru);
463
464 return next;
465}
466
467/*
468 * Encode <page, obj_idx> as a single handle value.
469 * On hardware platforms with physical memory starting at 0x0 the pfn
470 * could be 0 so we ensure that the handle will never be 0 by adjusting the
471 * encoded obj_idx value before encoding.
472 */
473static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
474{
475 unsigned long handle;
476
477 if (!page) {
478 BUG_ON(obj_idx);
479 return NULL;
480 }
481
482 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
483 handle |= ((obj_idx + 1) & OBJ_INDEX_MASK);
484
485 return (void *)handle;
486}
487
488/*
489 * Decode <page, obj_idx> pair from the given object handle. We adjust the
490 * decoded obj_idx back to its original value since it was adjusted in
491 * obj_location_to_handle().
492 */
493static void obj_handle_to_location(unsigned long handle, struct page **page,
494 unsigned long *obj_idx)
495{
496 *page = pfn_to_page(handle >> OBJ_INDEX_BITS);
497 *obj_idx = (handle & OBJ_INDEX_MASK) - 1;
498}
499
500static unsigned long obj_idx_to_offset(struct page *page,
501 unsigned long obj_idx, int class_size)
502{
503 unsigned long off = 0;
504
505 if (!is_first_page(page))
506 off = page->index;
507
508 return off + obj_idx * class_size;
509}
510
511static void reset_page(struct page *page)
512{
513 clear_bit(PG_private, &page->flags);
514 clear_bit(PG_private_2, &page->flags);
515 set_page_private(page, 0);
516 page->mapping = NULL;
517 page->freelist = NULL;
518 page_mapcount_reset(page);
519}
520
521static void free_zspage(struct page *first_page)
522{
523 struct page *nextp, *tmp, *head_extra;
524
525 BUG_ON(!is_first_page(first_page));
526 BUG_ON(first_page->inuse);
527
528 head_extra = (struct page *)page_private(first_page);
529
530 reset_page(first_page);
531 __free_page(first_page);
532
533 /* zspage with only 1 system page */
534 if (!head_extra)
535 return;
536
537 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
538 list_del(&nextp->lru);
539 reset_page(nextp);
540 __free_page(nextp);
541 }
542 reset_page(head_extra);
543 __free_page(head_extra);
544}
545
546/* Initialize a newly allocated zspage */
547static void init_zspage(struct page *first_page, struct size_class *class)
548{
549 unsigned long off = 0;
550 struct page *page = first_page;
551
552 BUG_ON(!is_first_page(first_page));
553 while (page) {
554 struct page *next_page;
555 struct link_free *link;
556 unsigned int i, objs_on_page;
557
558 /*
559 * page->index stores offset of first object starting
560 * in the page. For the first page, this is always 0,
561 * so we use first_page->index (aka ->freelist) to store
562 * head of corresponding zspage's freelist.
563 */
564 if (page != first_page)
565 page->index = off;
566
567 link = (struct link_free *)kmap_atomic(page) +
568 off / sizeof(*link);
569 objs_on_page = (PAGE_SIZE - off) / class->size;
570
571 for (i = 1; i <= objs_on_page; i++) {
572 off += class->size;
573 if (off < PAGE_SIZE) {
574 link->next = obj_location_to_handle(page, i);
575 link += class->size / sizeof(*link);
576 }
577 }
578
579 /*
580 * We now come to the last (full or partial) object on this
581 * page, which must point to the first object on the next
582 * page (if present)
583 */
584 next_page = get_next_page(page);
585 link->next = obj_location_to_handle(next_page, 0);
586 kunmap_atomic(link);
587 page = next_page;
588 off = (off + class->size) % PAGE_SIZE;
589 }
590}
591
592/*
593 * Allocate a zspage for the given size class
594 */
595static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
596{
597 int i, error;
598 struct page *first_page = NULL, *uninitialized_var(prev_page);
599
600 /*
601 * Allocate individual pages and link them together as:
602 * 1. first page->private = first sub-page
603 * 2. all sub-pages are linked together using page->lru
604 * 3. each sub-page is linked to the first page using page->first_page
605 *
606 * For each size class, First/Head pages are linked together using
607 * page->lru. Also, we set PG_private to identify the first page
608 * (i.e. no other sub-page has this flag set) and PG_private_2 to
609 * identify the last page.
610 */
611 error = -ENOMEM;
612 for (i = 0; i < class->pages_per_zspage; i++) {
613 struct page *page;
614
615 page = alloc_page(flags);
616 if (!page)
617 goto cleanup;
618
619 INIT_LIST_HEAD(&page->lru);
620 if (i == 0) { /* first page */
621 SetPagePrivate(page);
622 set_page_private(page, 0);
623 first_page = page;
624 first_page->inuse = 0;
625 }
626 if (i == 1)
627 set_page_private(first_page, (unsigned long)page);
628 if (i >= 1)
629 page->first_page = first_page;
630 if (i >= 2)
631 list_add(&page->lru, &prev_page->lru);
632 if (i == class->pages_per_zspage - 1) /* last page */
633 SetPagePrivate2(page);
634 prev_page = page;
635 }
636
637 init_zspage(first_page, class);
638
639 first_page->freelist = obj_location_to_handle(first_page, 0);
640 /* Maximum number of objects we can store in this zspage */
641 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
642
643 error = 0; /* Success */
644
645cleanup:
646 if (unlikely(error) && first_page) {
647 free_zspage(first_page);
648 first_page = NULL;
649 }
650
651 return first_page;
652}
653
654static struct page *find_get_zspage(struct size_class *class)
655{
656 int i;
657 struct page *page;
658
659 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
660 page = class->fullness_list[i];
661 if (page)
662 break;
663 }
664
665 return page;
666}
667
668#ifdef CONFIG_PGTABLE_MAPPING
669static inline int __zs_cpu_up(struct mapping_area *area)
670{
671 /*
672 * Make sure we don't leak memory if a cpu UP notification
673 * and zs_init() race and both call zs_cpu_up() on the same cpu
674 */
675 if (area->vm)
676 return 0;
677 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
678 if (!area->vm)
679 return -ENOMEM;
680 return 0;
681}
682
683static inline void __zs_cpu_down(struct mapping_area *area)
684{
685 if (area->vm)
686 free_vm_area(area->vm);
687 area->vm = NULL;
688}
689
690static inline void *__zs_map_object(struct mapping_area *area,
691 struct page *pages[2], int off, int size)
692{
693 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, &pages));
694 area->vm_addr = area->vm->addr;
695 return area->vm_addr + off;
696}
697
698static inline void __zs_unmap_object(struct mapping_area *area,
699 struct page *pages[2], int off, int size)
700{
701 unsigned long addr = (unsigned long)area->vm_addr;
702
703 unmap_kernel_range(addr, PAGE_SIZE * 2);
704}
705
706#else /* CONFIG_PGTABLE_MAPPING */
707
708static inline int __zs_cpu_up(struct mapping_area *area)
709{
710 /*
711 * Make sure we don't leak memory if a cpu UP notification
712 * and zs_init() race and both call zs_cpu_up() on the same cpu
713 */
714 if (area->vm_buf)
715 return 0;
716 area->vm_buf = (char *)__get_free_page(GFP_KERNEL);
717 if (!area->vm_buf)
718 return -ENOMEM;
719 return 0;
720}
721
722static inline void __zs_cpu_down(struct mapping_area *area)
723{
724 if (area->vm_buf)
725 free_page((unsigned long)area->vm_buf);
726 area->vm_buf = NULL;
727}
728
729static void *__zs_map_object(struct mapping_area *area,
730 struct page *pages[2], int off, int size)
731{
732 int sizes[2];
733 void *addr;
734 char *buf = area->vm_buf;
735
736 /* disable page faults to match kmap_atomic() return conditions */
737 pagefault_disable();
738
739 /* no read fastpath */
740 if (area->vm_mm == ZS_MM_WO)
741 goto out;
742
743 sizes[0] = PAGE_SIZE - off;
744 sizes[1] = size - sizes[0];
745
746 /* copy object to per-cpu buffer */
747 addr = kmap_atomic(pages[0]);
748 memcpy(buf, addr + off, sizes[0]);
749 kunmap_atomic(addr);
750 addr = kmap_atomic(pages[1]);
751 memcpy(buf + sizes[0], addr, sizes[1]);
752 kunmap_atomic(addr);
753out:
754 return area->vm_buf;
755}
756
757static void __zs_unmap_object(struct mapping_area *area,
758 struct page *pages[2], int off, int size)
759{
760 int sizes[2];
761 void *addr;
762 char *buf = area->vm_buf;
763
764 /* no write fastpath */
765 if (area->vm_mm == ZS_MM_RO)
766 goto out;
767
768 sizes[0] = PAGE_SIZE - off;
769 sizes[1] = size - sizes[0];
770
771 /* copy per-cpu buffer to object */
772 addr = kmap_atomic(pages[0]);
773 memcpy(addr + off, buf, sizes[0]);
774 kunmap_atomic(addr);
775 addr = kmap_atomic(pages[1]);
776 memcpy(addr, buf + sizes[0], sizes[1]);
777 kunmap_atomic(addr);
778
779out:
780 /* enable page faults to match kunmap_atomic() return conditions */
781 pagefault_enable();
782}
783
784#endif /* CONFIG_PGTABLE_MAPPING */
785
786static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
787 void *pcpu)
788{
789 int ret, cpu = (long)pcpu;
790 struct mapping_area *area;
791
792 switch (action) {
793 case CPU_UP_PREPARE:
794 area = &per_cpu(zs_map_area, cpu);
795 ret = __zs_cpu_up(area);
796 if (ret)
797 return notifier_from_errno(ret);
798 break;
799 case CPU_DEAD:
800 case CPU_UP_CANCELED:
801 area = &per_cpu(zs_map_area, cpu);
802 __zs_cpu_down(area);
803 break;
804 }
805
806 return NOTIFY_OK;
807}
808
809static struct notifier_block zs_cpu_nb = {
810 .notifier_call = zs_cpu_notifier
811};
812
813static void zs_exit(void)
814{
815 int cpu;
816
817 cpu_notifier_register_begin();
818
819 for_each_online_cpu(cpu)
820 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
821 __unregister_cpu_notifier(&zs_cpu_nb);
822
823 cpu_notifier_register_done();
824}
825
826static int zs_init(void)
827{
828 int cpu, ret;
829
830 cpu_notifier_register_begin();
831
832 __register_cpu_notifier(&zs_cpu_nb);
833 for_each_online_cpu(cpu) {
834 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
835 if (notifier_to_errno(ret)) {
836 cpu_notifier_register_done();
837 goto fail;
838 }
839 }
840
841 cpu_notifier_register_done();
842
843 return 0;
844fail:
845 zs_exit();
846 return notifier_to_errno(ret);
847}
848
849/**
850 * zs_create_pool - Creates an allocation pool to work from.
851 * @flags: allocation flags used to allocate pool metadata
852 *
853 * This function must be called before anything when using
854 * the zsmalloc allocator.
855 *
856 * On success, a pointer to the newly created pool is returned,
857 * otherwise NULL.
858 */
859struct zs_pool *zs_create_pool(gfp_t flags)
860{
861 int i, ovhd_size;
862 struct zs_pool *pool;
863
864 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
865 pool = kzalloc(ovhd_size, GFP_KERNEL);
866 if (!pool)
867 return NULL;
868
869 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
870 int size;
871 struct size_class *class;
872
873 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
874 if (size > ZS_MAX_ALLOC_SIZE)
875 size = ZS_MAX_ALLOC_SIZE;
876
877 class = &pool->size_class[i];
878 class->size = size;
879 class->index = i;
880 spin_lock_init(&class->lock);
881 class->pages_per_zspage = get_pages_per_zspage(size);
882
883 }
884
885 pool->flags = flags;
886
887 return pool;
888}
889EXPORT_SYMBOL_GPL(zs_create_pool);
890
891void zs_destroy_pool(struct zs_pool *pool)
892{
893 int i;
894
895 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
896 int fg;
897 struct size_class *class = &pool->size_class[i];
898
899 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
900 if (class->fullness_list[fg]) {
901 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
902 class->size, fg);
903 }
904 }
905 }
906 kfree(pool);
907}
908EXPORT_SYMBOL_GPL(zs_destroy_pool);
909
910/**
911 * zs_malloc - Allocate block of given size from pool.
912 * @pool: pool to allocate from
913 * @size: size of block to allocate
914 *
915 * On success, handle to the allocated object is returned,
916 * otherwise 0.
917 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
918 */
919unsigned long zs_malloc(struct zs_pool *pool, size_t size)
920{
921 unsigned long obj;
922 struct link_free *link;
923 int class_idx;
924 struct size_class *class;
925
926 struct page *first_page, *m_page;
927 unsigned long m_objidx, m_offset;
928
929 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
930 return 0;
931
932 class_idx = get_size_class_index(size);
933 class = &pool->size_class[class_idx];
934 BUG_ON(class_idx != class->index);
935
936 spin_lock(&class->lock);
937 first_page = find_get_zspage(class);
938
939 if (!first_page) {
940 spin_unlock(&class->lock);
941 first_page = alloc_zspage(class, pool->flags);
942 if (unlikely(!first_page))
943 return 0;
944
945 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
946 spin_lock(&class->lock);
947 class->pages_allocated += class->pages_per_zspage;
948 }
949
950 obj = (unsigned long)first_page->freelist;
951 obj_handle_to_location(obj, &m_page, &m_objidx);
952 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
953
954 link = (struct link_free *)kmap_atomic(m_page) +
955 m_offset / sizeof(*link);
956 first_page->freelist = link->next;
957 memset(link, POISON_INUSE, sizeof(*link));
958 kunmap_atomic(link);
959
960 first_page->inuse++;
961 /* Now move the zspage to another fullness group, if required */
962 fix_fullness_group(pool, first_page);
963 spin_unlock(&class->lock);
964
965 return obj;
966}
967EXPORT_SYMBOL_GPL(zs_malloc);
968
969void zs_free(struct zs_pool *pool, unsigned long obj)
970{
971 struct link_free *link;
972 struct page *first_page, *f_page;
973 unsigned long f_objidx, f_offset;
974
975 int class_idx;
976 struct size_class *class;
977 enum fullness_group fullness;
978
979 if (unlikely(!obj))
980 return;
981
982 obj_handle_to_location(obj, &f_page, &f_objidx);
983 first_page = get_first_page(f_page);
984
985 get_zspage_mapping(first_page, &class_idx, &fullness);
986 class = &pool->size_class[class_idx];
987 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
988
989 spin_lock(&class->lock);
990
991 /* Insert this object in containing zspage's freelist */
992 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
993 + f_offset);
994 link->next = first_page->freelist;
995 kunmap_atomic(link);
996 first_page->freelist = (void *)obj;
997
998 first_page->inuse--;
999 fullness = fix_fullness_group(pool, first_page);
1000
1001 if (fullness == ZS_EMPTY)
1002 class->pages_allocated -= class->pages_per_zspage;
1003
1004 spin_unlock(&class->lock);
1005
1006 if (fullness == ZS_EMPTY)
1007 free_zspage(first_page);
1008}
1009EXPORT_SYMBOL_GPL(zs_free);
1010
1011/**
1012 * zs_map_object - get address of allocated object from handle.
1013 * @pool: pool from which the object was allocated
1014 * @handle: handle returned from zs_malloc
1015 *
1016 * Before using an object allocated from zs_malloc, it must be mapped using
1017 * this function. When done with the object, it must be unmapped using
1018 * zs_unmap_object.
1019 *
1020 * Only one object can be mapped per cpu at a time. There is no protection
1021 * against nested mappings.
1022 *
1023 * This function returns with preemption and page faults disabled.
1024 */
1025void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1026 enum zs_mapmode mm)
1027{
1028 struct page *page;
1029 unsigned long obj_idx, off;
1030
1031 unsigned int class_idx;
1032 enum fullness_group fg;
1033 struct size_class *class;
1034 struct mapping_area *area;
1035 struct page *pages[2];
1036
1037 BUG_ON(!handle);
1038
1039 /*
1040 * Because we use per-cpu mapping areas shared among the
1041 * pools/users, we can't allow mapping in interrupt context
1042 * because it can corrupt another users mappings.
1043 */
1044 BUG_ON(in_interrupt());
1045
1046 obj_handle_to_location(handle, &page, &obj_idx);
1047 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1048 class = &pool->size_class[class_idx];
1049 off = obj_idx_to_offset(page, obj_idx, class->size);
1050
1051 area = &get_cpu_var(zs_map_area);
1052 area->vm_mm = mm;
1053 if (off + class->size <= PAGE_SIZE) {
1054 /* this object is contained entirely within a page */
1055 area->vm_addr = kmap_atomic(page);
1056 return area->vm_addr + off;
1057 }
1058
1059 /* this object spans two pages */
1060 pages[0] = page;
1061 pages[1] = get_next_page(page);
1062 BUG_ON(!pages[1]);
1063
1064 return __zs_map_object(area, pages, off, class->size);
1065}
1066EXPORT_SYMBOL_GPL(zs_map_object);
1067
1068void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1069{
1070 struct page *page;
1071 unsigned long obj_idx, off;
1072
1073 unsigned int class_idx;
1074 enum fullness_group fg;
1075 struct size_class *class;
1076 struct mapping_area *area;
1077
1078 BUG_ON(!handle);
1079
1080 obj_handle_to_location(handle, &page, &obj_idx);
1081 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1082 class = &pool->size_class[class_idx];
1083 off = obj_idx_to_offset(page, obj_idx, class->size);
1084
1085 area = &__get_cpu_var(zs_map_area);
1086 if (off + class->size <= PAGE_SIZE)
1087 kunmap_atomic(area->vm_addr);
1088 else {
1089 struct page *pages[2];
1090
1091 pages[0] = page;
1092 pages[1] = get_next_page(page);
1093 BUG_ON(!pages[1]);
1094
1095 __zs_unmap_object(area, pages, off, class->size);
1096 }
1097 put_cpu_var(zs_map_area);
1098}
1099EXPORT_SYMBOL_GPL(zs_unmap_object);
1100
1101u64 zs_get_total_size_bytes(struct zs_pool *pool)
1102{
1103 int i;
1104 u64 npages = 0;
1105
1106 for (i = 0; i < ZS_SIZE_CLASSES; i++)
1107 npages += pool->size_class[i].pages_allocated;
1108
1109 return npages << PAGE_SHIFT;
1110}
1111EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);
1112
1113module_init(zs_init);
1114module_exit(zs_exit);
1115
1116MODULE_LICENSE("Dual BSD/GPL");
1117MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");
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 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
17 *
18 * Usage of struct page fields:
19 * page->private: points to zspage
20 * page->index: links together all component pages of a zspage
21 * For the huge page, this is always 0, so we use this field
22 * to store handle.
23 * page->page_type: first object offset in a subpage of zspage
24 *
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
27 * PG_owner_priv_1: identifies the huge component page
28 *
29 */
30
31#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
32
33/*
34 * lock ordering:
35 * page_lock
36 * pool->lock
37 * zspage->lock
38 */
39
40#include <linux/module.h>
41#include <linux/kernel.h>
42#include <linux/sched.h>
43#include <linux/bitops.h>
44#include <linux/errno.h>
45#include <linux/highmem.h>
46#include <linux/string.h>
47#include <linux/slab.h>
48#include <linux/pgtable.h>
49#include <asm/tlbflush.h>
50#include <linux/cpumask.h>
51#include <linux/cpu.h>
52#include <linux/vmalloc.h>
53#include <linux/preempt.h>
54#include <linux/spinlock.h>
55#include <linux/shrinker.h>
56#include <linux/types.h>
57#include <linux/debugfs.h>
58#include <linux/zsmalloc.h>
59#include <linux/zpool.h>
60#include <linux/migrate.h>
61#include <linux/wait.h>
62#include <linux/pagemap.h>
63#include <linux/fs.h>
64#include <linux/local_lock.h>
65
66#define ZSPAGE_MAGIC 0x58
67
68/*
69 * This must be power of 2 and greater than or equal to sizeof(link_free).
70 * These two conditions ensure that any 'struct link_free' itself doesn't
71 * span more than 1 page which avoids complex case of mapping 2 pages simply
72 * to restore link_free pointer values.
73 */
74#define ZS_ALIGN 8
75
76#define ZS_HANDLE_SIZE (sizeof(unsigned long))
77
78/*
79 * Object location (<PFN>, <obj_idx>) is encoded as
80 * a single (unsigned long) handle value.
81 *
82 * Note that object index <obj_idx> starts from 0.
83 *
84 * This is made more complicated by various memory models and PAE.
85 */
86
87#ifndef MAX_POSSIBLE_PHYSMEM_BITS
88#ifdef MAX_PHYSMEM_BITS
89#define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
90#else
91/*
92 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
93 * be PAGE_SHIFT
94 */
95#define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
96#endif
97#endif
98
99#define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
100
101/*
102 * Head in allocated object should have OBJ_ALLOCATED_TAG
103 * to identify the object was allocated or not.
104 * It's okay to add the status bit in the least bit because
105 * header keeps handle which is 4byte-aligned address so we
106 * have room for two bit at least.
107 */
108#define OBJ_ALLOCATED_TAG 1
109
110#define OBJ_TAG_BITS 1
111#define OBJ_TAG_MASK OBJ_ALLOCATED_TAG
112
113#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
114#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
115
116#define HUGE_BITS 1
117#define FULLNESS_BITS 4
118#define CLASS_BITS 8
119#define ISOLATED_BITS 5
120#define MAGIC_VAL_BITS 8
121
122#define MAX(a, b) ((a) >= (b) ? (a) : (b))
123
124#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(CONFIG_ZSMALLOC_CHAIN_SIZE, UL))
125
126/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
127#define ZS_MIN_ALLOC_SIZE \
128 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
129/* each chunk includes extra space to keep handle */
130#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
131
132/*
133 * On systems with 4K page size, this gives 255 size classes! There is a
134 * trader-off here:
135 * - Large number of size classes is potentially wasteful as free page are
136 * spread across these classes
137 * - Small number of size classes causes large internal fragmentation
138 * - Probably its better to use specific size classes (empirically
139 * determined). NOTE: all those class sizes must be set as multiple of
140 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
141 *
142 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
143 * (reason above)
144 */
145#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
146#define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
147 ZS_SIZE_CLASS_DELTA) + 1)
148
149/*
150 * Pages are distinguished by the ratio of used memory (that is the ratio
151 * of ->inuse objects to all objects that page can store). For example,
152 * INUSE_RATIO_10 means that the ratio of used objects is > 0% and <= 10%.
153 *
154 * The number of fullness groups is not random. It allows us to keep
155 * difference between the least busy page in the group (minimum permitted
156 * number of ->inuse objects) and the most busy page (maximum permitted
157 * number of ->inuse objects) at a reasonable value.
158 */
159enum fullness_group {
160 ZS_INUSE_RATIO_0,
161 ZS_INUSE_RATIO_10,
162 /* NOTE: 8 more fullness groups here */
163 ZS_INUSE_RATIO_99 = 10,
164 ZS_INUSE_RATIO_100,
165 NR_FULLNESS_GROUPS,
166};
167
168enum class_stat_type {
169 /* NOTE: stats for 12 fullness groups here: from inuse 0 to 100 */
170 ZS_OBJS_ALLOCATED = NR_FULLNESS_GROUPS,
171 ZS_OBJS_INUSE,
172 NR_CLASS_STAT_TYPES,
173};
174
175struct zs_size_stat {
176 unsigned long objs[NR_CLASS_STAT_TYPES];
177};
178
179#ifdef CONFIG_ZSMALLOC_STAT
180static struct dentry *zs_stat_root;
181#endif
182
183static size_t huge_class_size;
184
185struct size_class {
186 struct list_head fullness_list[NR_FULLNESS_GROUPS];
187 /*
188 * Size of objects stored in this class. Must be multiple
189 * of ZS_ALIGN.
190 */
191 int size;
192 int objs_per_zspage;
193 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
194 int pages_per_zspage;
195
196 unsigned int index;
197 struct zs_size_stat stats;
198};
199
200/*
201 * Placed within free objects to form a singly linked list.
202 * For every zspage, zspage->freeobj gives head of this list.
203 *
204 * This must be power of 2 and less than or equal to ZS_ALIGN
205 */
206struct link_free {
207 union {
208 /*
209 * Free object index;
210 * It's valid for non-allocated object
211 */
212 unsigned long next;
213 /*
214 * Handle of allocated object.
215 */
216 unsigned long handle;
217 };
218};
219
220struct zs_pool {
221 const char *name;
222
223 struct size_class *size_class[ZS_SIZE_CLASSES];
224 struct kmem_cache *handle_cachep;
225 struct kmem_cache *zspage_cachep;
226
227 atomic_long_t pages_allocated;
228
229 struct zs_pool_stats stats;
230
231 /* Compact classes */
232 struct shrinker *shrinker;
233
234#ifdef CONFIG_ZSMALLOC_STAT
235 struct dentry *stat_dentry;
236#endif
237#ifdef CONFIG_COMPACTION
238 struct work_struct free_work;
239#endif
240 spinlock_t lock;
241 atomic_t compaction_in_progress;
242};
243
244struct zspage {
245 struct {
246 unsigned int huge:HUGE_BITS;
247 unsigned int fullness:FULLNESS_BITS;
248 unsigned int class:CLASS_BITS + 1;
249 unsigned int isolated:ISOLATED_BITS;
250 unsigned int magic:MAGIC_VAL_BITS;
251 };
252 unsigned int inuse;
253 unsigned int freeobj;
254 struct page *first_page;
255 struct list_head list; /* fullness list */
256 struct zs_pool *pool;
257 rwlock_t lock;
258};
259
260struct mapping_area {
261 local_lock_t lock;
262 char *vm_buf; /* copy buffer for objects that span pages */
263 char *vm_addr; /* address of kmap_atomic()'ed pages */
264 enum zs_mapmode vm_mm; /* mapping mode */
265};
266
267/* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
268static void SetZsHugePage(struct zspage *zspage)
269{
270 zspage->huge = 1;
271}
272
273static bool ZsHugePage(struct zspage *zspage)
274{
275 return zspage->huge;
276}
277
278static void migrate_lock_init(struct zspage *zspage);
279static void migrate_read_lock(struct zspage *zspage);
280static void migrate_read_unlock(struct zspage *zspage);
281
282#ifdef CONFIG_COMPACTION
283static void migrate_write_lock(struct zspage *zspage);
284static void migrate_write_lock_nested(struct zspage *zspage);
285static void migrate_write_unlock(struct zspage *zspage);
286static void kick_deferred_free(struct zs_pool *pool);
287static void init_deferred_free(struct zs_pool *pool);
288static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
289#else
290static void migrate_write_lock(struct zspage *zspage) {}
291static void migrate_write_lock_nested(struct zspage *zspage) {}
292static void migrate_write_unlock(struct zspage *zspage) {}
293static void kick_deferred_free(struct zs_pool *pool) {}
294static void init_deferred_free(struct zs_pool *pool) {}
295static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
296#endif
297
298static int create_cache(struct zs_pool *pool)
299{
300 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
301 0, 0, NULL);
302 if (!pool->handle_cachep)
303 return 1;
304
305 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
306 0, 0, NULL);
307 if (!pool->zspage_cachep) {
308 kmem_cache_destroy(pool->handle_cachep);
309 pool->handle_cachep = NULL;
310 return 1;
311 }
312
313 return 0;
314}
315
316static void destroy_cache(struct zs_pool *pool)
317{
318 kmem_cache_destroy(pool->handle_cachep);
319 kmem_cache_destroy(pool->zspage_cachep);
320}
321
322static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
323{
324 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
325 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
326}
327
328static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
329{
330 kmem_cache_free(pool->handle_cachep, (void *)handle);
331}
332
333static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
334{
335 return kmem_cache_zalloc(pool->zspage_cachep,
336 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
337}
338
339static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
340{
341 kmem_cache_free(pool->zspage_cachep, zspage);
342}
343
344/* pool->lock(which owns the handle) synchronizes races */
345static void record_obj(unsigned long handle, unsigned long obj)
346{
347 *(unsigned long *)handle = obj;
348}
349
350/* zpool driver */
351
352#ifdef CONFIG_ZPOOL
353
354static void *zs_zpool_create(const char *name, gfp_t gfp)
355{
356 /*
357 * Ignore global gfp flags: zs_malloc() may be invoked from
358 * different contexts and its caller must provide a valid
359 * gfp mask.
360 */
361 return zs_create_pool(name);
362}
363
364static void zs_zpool_destroy(void *pool)
365{
366 zs_destroy_pool(pool);
367}
368
369static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
370 unsigned long *handle)
371{
372 *handle = zs_malloc(pool, size, gfp);
373
374 if (IS_ERR_VALUE(*handle))
375 return PTR_ERR((void *)*handle);
376 return 0;
377}
378static void zs_zpool_free(void *pool, unsigned long handle)
379{
380 zs_free(pool, handle);
381}
382
383static void *zs_zpool_map(void *pool, unsigned long handle,
384 enum zpool_mapmode mm)
385{
386 enum zs_mapmode zs_mm;
387
388 switch (mm) {
389 case ZPOOL_MM_RO:
390 zs_mm = ZS_MM_RO;
391 break;
392 case ZPOOL_MM_WO:
393 zs_mm = ZS_MM_WO;
394 break;
395 case ZPOOL_MM_RW:
396 default:
397 zs_mm = ZS_MM_RW;
398 break;
399 }
400
401 return zs_map_object(pool, handle, zs_mm);
402}
403static void zs_zpool_unmap(void *pool, unsigned long handle)
404{
405 zs_unmap_object(pool, handle);
406}
407
408static u64 zs_zpool_total_size(void *pool)
409{
410 return zs_get_total_pages(pool) << PAGE_SHIFT;
411}
412
413static struct zpool_driver zs_zpool_driver = {
414 .type = "zsmalloc",
415 .owner = THIS_MODULE,
416 .create = zs_zpool_create,
417 .destroy = zs_zpool_destroy,
418 .malloc_support_movable = true,
419 .malloc = zs_zpool_malloc,
420 .free = zs_zpool_free,
421 .map = zs_zpool_map,
422 .unmap = zs_zpool_unmap,
423 .total_size = zs_zpool_total_size,
424};
425
426MODULE_ALIAS("zpool-zsmalloc");
427#endif /* CONFIG_ZPOOL */
428
429/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
430static DEFINE_PER_CPU(struct mapping_area, zs_map_area) = {
431 .lock = INIT_LOCAL_LOCK(lock),
432};
433
434static __maybe_unused int is_first_page(struct page *page)
435{
436 return PagePrivate(page);
437}
438
439/* Protected by pool->lock */
440static inline int get_zspage_inuse(struct zspage *zspage)
441{
442 return zspage->inuse;
443}
444
445
446static inline void mod_zspage_inuse(struct zspage *zspage, int val)
447{
448 zspage->inuse += val;
449}
450
451static inline struct page *get_first_page(struct zspage *zspage)
452{
453 struct page *first_page = zspage->first_page;
454
455 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
456 return first_page;
457}
458
459static inline unsigned int get_first_obj_offset(struct page *page)
460{
461 return page->page_type;
462}
463
464static inline void set_first_obj_offset(struct page *page, unsigned int offset)
465{
466 page->page_type = offset;
467}
468
469static inline unsigned int get_freeobj(struct zspage *zspage)
470{
471 return zspage->freeobj;
472}
473
474static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
475{
476 zspage->freeobj = obj;
477}
478
479static void get_zspage_mapping(struct zspage *zspage,
480 unsigned int *class_idx,
481 int *fullness)
482{
483 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
484
485 *fullness = zspage->fullness;
486 *class_idx = zspage->class;
487}
488
489static struct size_class *zspage_class(struct zs_pool *pool,
490 struct zspage *zspage)
491{
492 return pool->size_class[zspage->class];
493}
494
495static void set_zspage_mapping(struct zspage *zspage,
496 unsigned int class_idx,
497 int fullness)
498{
499 zspage->class = class_idx;
500 zspage->fullness = fullness;
501}
502
503/*
504 * zsmalloc divides the pool into various size classes where each
505 * class maintains a list of zspages where each zspage is divided
506 * into equal sized chunks. Each allocation falls into one of these
507 * classes depending on its size. This function returns index of the
508 * size class which has chunk size big enough to hold the given size.
509 */
510static int get_size_class_index(int size)
511{
512 int idx = 0;
513
514 if (likely(size > ZS_MIN_ALLOC_SIZE))
515 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
516 ZS_SIZE_CLASS_DELTA);
517
518 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
519}
520
521static inline void class_stat_inc(struct size_class *class,
522 int type, unsigned long cnt)
523{
524 class->stats.objs[type] += cnt;
525}
526
527static inline void class_stat_dec(struct size_class *class,
528 int type, unsigned long cnt)
529{
530 class->stats.objs[type] -= cnt;
531}
532
533static inline unsigned long zs_stat_get(struct size_class *class, int type)
534{
535 return class->stats.objs[type];
536}
537
538#ifdef CONFIG_ZSMALLOC_STAT
539
540static void __init zs_stat_init(void)
541{
542 if (!debugfs_initialized()) {
543 pr_warn("debugfs not available, stat dir not created\n");
544 return;
545 }
546
547 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
548}
549
550static void __exit zs_stat_exit(void)
551{
552 debugfs_remove_recursive(zs_stat_root);
553}
554
555static unsigned long zs_can_compact(struct size_class *class);
556
557static int zs_stats_size_show(struct seq_file *s, void *v)
558{
559 int i, fg;
560 struct zs_pool *pool = s->private;
561 struct size_class *class;
562 int objs_per_zspage;
563 unsigned long obj_allocated, obj_used, pages_used, freeable;
564 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
565 unsigned long total_freeable = 0;
566 unsigned long inuse_totals[NR_FULLNESS_GROUPS] = {0, };
567
568 seq_printf(s, " %5s %5s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %13s %10s %10s %16s %8s\n",
569 "class", "size", "10%", "20%", "30%", "40%",
570 "50%", "60%", "70%", "80%", "90%", "99%", "100%",
571 "obj_allocated", "obj_used", "pages_used",
572 "pages_per_zspage", "freeable");
573
574 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
575
576 class = pool->size_class[i];
577
578 if (class->index != i)
579 continue;
580
581 spin_lock(&pool->lock);
582
583 seq_printf(s, " %5u %5u ", i, class->size);
584 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++) {
585 inuse_totals[fg] += zs_stat_get(class, fg);
586 seq_printf(s, "%9lu ", zs_stat_get(class, fg));
587 }
588
589 obj_allocated = zs_stat_get(class, ZS_OBJS_ALLOCATED);
590 obj_used = zs_stat_get(class, ZS_OBJS_INUSE);
591 freeable = zs_can_compact(class);
592 spin_unlock(&pool->lock);
593
594 objs_per_zspage = class->objs_per_zspage;
595 pages_used = obj_allocated / objs_per_zspage *
596 class->pages_per_zspage;
597
598 seq_printf(s, "%13lu %10lu %10lu %16d %8lu\n",
599 obj_allocated, obj_used, pages_used,
600 class->pages_per_zspage, freeable);
601
602 total_objs += obj_allocated;
603 total_used_objs += obj_used;
604 total_pages += pages_used;
605 total_freeable += freeable;
606 }
607
608 seq_puts(s, "\n");
609 seq_printf(s, " %5s %5s ", "Total", "");
610
611 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++)
612 seq_printf(s, "%9lu ", inuse_totals[fg]);
613
614 seq_printf(s, "%13lu %10lu %10lu %16s %8lu\n",
615 total_objs, total_used_objs, total_pages, "",
616 total_freeable);
617
618 return 0;
619}
620DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
621
622static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
623{
624 if (!zs_stat_root) {
625 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
626 return;
627 }
628
629 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
630
631 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
632 &zs_stats_size_fops);
633}
634
635static void zs_pool_stat_destroy(struct zs_pool *pool)
636{
637 debugfs_remove_recursive(pool->stat_dentry);
638}
639
640#else /* CONFIG_ZSMALLOC_STAT */
641static void __init zs_stat_init(void)
642{
643}
644
645static void __exit zs_stat_exit(void)
646{
647}
648
649static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
650{
651}
652
653static inline void zs_pool_stat_destroy(struct zs_pool *pool)
654{
655}
656#endif
657
658
659/*
660 * For each size class, zspages are divided into different groups
661 * depending on their usage ratio. This function returns fullness
662 * status of the given page.
663 */
664static int get_fullness_group(struct size_class *class, struct zspage *zspage)
665{
666 int inuse, objs_per_zspage, ratio;
667
668 inuse = get_zspage_inuse(zspage);
669 objs_per_zspage = class->objs_per_zspage;
670
671 if (inuse == 0)
672 return ZS_INUSE_RATIO_0;
673 if (inuse == objs_per_zspage)
674 return ZS_INUSE_RATIO_100;
675
676 ratio = 100 * inuse / objs_per_zspage;
677 /*
678 * Take integer division into consideration: a page with one inuse
679 * object out of 127 possible, will end up having 0 usage ratio,
680 * which is wrong as it belongs in ZS_INUSE_RATIO_10 fullness group.
681 */
682 return ratio / 10 + 1;
683}
684
685/*
686 * Each size class maintains various freelists and zspages are assigned
687 * to one of these freelists based on the number of live objects they
688 * have. This functions inserts the given zspage into the freelist
689 * identified by <class, fullness_group>.
690 */
691static void insert_zspage(struct size_class *class,
692 struct zspage *zspage,
693 int fullness)
694{
695 class_stat_inc(class, fullness, 1);
696 list_add(&zspage->list, &class->fullness_list[fullness]);
697}
698
699/*
700 * This function removes the given zspage from the freelist identified
701 * by <class, fullness_group>.
702 */
703static void remove_zspage(struct size_class *class,
704 struct zspage *zspage,
705 int fullness)
706{
707 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
708
709 list_del_init(&zspage->list);
710 class_stat_dec(class, fullness, 1);
711}
712
713/*
714 * Each size class maintains zspages in different fullness groups depending
715 * on the number of live objects they contain. When allocating or freeing
716 * objects, the fullness status of the page can change, for instance, from
717 * INUSE_RATIO_80 to INUSE_RATIO_70 when freeing an object. This function
718 * checks if such a status change has occurred for the given page and
719 * accordingly moves the page from the list of the old fullness group to that
720 * of the new fullness group.
721 */
722static int fix_fullness_group(struct size_class *class, struct zspage *zspage)
723{
724 int class_idx;
725 int currfg, newfg;
726
727 get_zspage_mapping(zspage, &class_idx, &currfg);
728 newfg = get_fullness_group(class, zspage);
729 if (newfg == currfg)
730 goto out;
731
732 remove_zspage(class, zspage, currfg);
733 insert_zspage(class, zspage, newfg);
734 set_zspage_mapping(zspage, class_idx, newfg);
735out:
736 return newfg;
737}
738
739static struct zspage *get_zspage(struct page *page)
740{
741 struct zspage *zspage = (struct zspage *)page_private(page);
742
743 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
744 return zspage;
745}
746
747static struct page *get_next_page(struct page *page)
748{
749 struct zspage *zspage = get_zspage(page);
750
751 if (unlikely(ZsHugePage(zspage)))
752 return NULL;
753
754 return (struct page *)page->index;
755}
756
757/**
758 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
759 * @obj: the encoded object value
760 * @page: page object resides in zspage
761 * @obj_idx: object index
762 */
763static void obj_to_location(unsigned long obj, struct page **page,
764 unsigned int *obj_idx)
765{
766 obj >>= OBJ_TAG_BITS;
767 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
768 *obj_idx = (obj & OBJ_INDEX_MASK);
769}
770
771static void obj_to_page(unsigned long obj, struct page **page)
772{
773 obj >>= OBJ_TAG_BITS;
774 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
775}
776
777/**
778 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
779 * @page: page object resides in zspage
780 * @obj_idx: object index
781 */
782static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
783{
784 unsigned long obj;
785
786 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
787 obj |= obj_idx & OBJ_INDEX_MASK;
788 obj <<= OBJ_TAG_BITS;
789
790 return obj;
791}
792
793static unsigned long handle_to_obj(unsigned long handle)
794{
795 return *(unsigned long *)handle;
796}
797
798static inline bool obj_allocated(struct page *page, void *obj,
799 unsigned long *phandle)
800{
801 unsigned long handle;
802 struct zspage *zspage = get_zspage(page);
803
804 if (unlikely(ZsHugePage(zspage))) {
805 VM_BUG_ON_PAGE(!is_first_page(page), page);
806 handle = page->index;
807 } else
808 handle = *(unsigned long *)obj;
809
810 if (!(handle & OBJ_ALLOCATED_TAG))
811 return false;
812
813 /* Clear all tags before returning the handle */
814 *phandle = handle & ~OBJ_TAG_MASK;
815 return true;
816}
817
818static void reset_page(struct page *page)
819{
820 __ClearPageMovable(page);
821 ClearPagePrivate(page);
822 set_page_private(page, 0);
823 page_mapcount_reset(page);
824 page->index = 0;
825}
826
827static int trylock_zspage(struct zspage *zspage)
828{
829 struct page *cursor, *fail;
830
831 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
832 get_next_page(cursor)) {
833 if (!trylock_page(cursor)) {
834 fail = cursor;
835 goto unlock;
836 }
837 }
838
839 return 1;
840unlock:
841 for (cursor = get_first_page(zspage); cursor != fail; cursor =
842 get_next_page(cursor))
843 unlock_page(cursor);
844
845 return 0;
846}
847
848static void __free_zspage(struct zs_pool *pool, struct size_class *class,
849 struct zspage *zspage)
850{
851 struct page *page, *next;
852 int fg;
853 unsigned int class_idx;
854
855 get_zspage_mapping(zspage, &class_idx, &fg);
856
857 assert_spin_locked(&pool->lock);
858
859 VM_BUG_ON(get_zspage_inuse(zspage));
860 VM_BUG_ON(fg != ZS_INUSE_RATIO_0);
861
862 next = page = get_first_page(zspage);
863 do {
864 VM_BUG_ON_PAGE(!PageLocked(page), page);
865 next = get_next_page(page);
866 reset_page(page);
867 unlock_page(page);
868 dec_zone_page_state(page, NR_ZSPAGES);
869 put_page(page);
870 page = next;
871 } while (page != NULL);
872
873 cache_free_zspage(pool, zspage);
874
875 class_stat_dec(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
876 atomic_long_sub(class->pages_per_zspage, &pool->pages_allocated);
877}
878
879static void free_zspage(struct zs_pool *pool, struct size_class *class,
880 struct zspage *zspage)
881{
882 VM_BUG_ON(get_zspage_inuse(zspage));
883 VM_BUG_ON(list_empty(&zspage->list));
884
885 /*
886 * Since zs_free couldn't be sleepable, this function cannot call
887 * lock_page. The page locks trylock_zspage got will be released
888 * by __free_zspage.
889 */
890 if (!trylock_zspage(zspage)) {
891 kick_deferred_free(pool);
892 return;
893 }
894
895 remove_zspage(class, zspage, ZS_INUSE_RATIO_0);
896 __free_zspage(pool, class, zspage);
897}
898
899/* Initialize a newly allocated zspage */
900static void init_zspage(struct size_class *class, struct zspage *zspage)
901{
902 unsigned int freeobj = 1;
903 unsigned long off = 0;
904 struct page *page = get_first_page(zspage);
905
906 while (page) {
907 struct page *next_page;
908 struct link_free *link;
909 void *vaddr;
910
911 set_first_obj_offset(page, off);
912
913 vaddr = kmap_atomic(page);
914 link = (struct link_free *)vaddr + off / sizeof(*link);
915
916 while ((off += class->size) < PAGE_SIZE) {
917 link->next = freeobj++ << OBJ_TAG_BITS;
918 link += class->size / sizeof(*link);
919 }
920
921 /*
922 * We now come to the last (full or partial) object on this
923 * page, which must point to the first object on the next
924 * page (if present)
925 */
926 next_page = get_next_page(page);
927 if (next_page) {
928 link->next = freeobj++ << OBJ_TAG_BITS;
929 } else {
930 /*
931 * Reset OBJ_TAG_BITS bit to last link to tell
932 * whether it's allocated object or not.
933 */
934 link->next = -1UL << OBJ_TAG_BITS;
935 }
936 kunmap_atomic(vaddr);
937 page = next_page;
938 off %= PAGE_SIZE;
939 }
940
941 set_freeobj(zspage, 0);
942}
943
944static void create_page_chain(struct size_class *class, struct zspage *zspage,
945 struct page *pages[])
946{
947 int i;
948 struct page *page;
949 struct page *prev_page = NULL;
950 int nr_pages = class->pages_per_zspage;
951
952 /*
953 * Allocate individual pages and link them together as:
954 * 1. all pages are linked together using page->index
955 * 2. each sub-page point to zspage using page->private
956 *
957 * we set PG_private to identify the first page (i.e. no other sub-page
958 * has this flag set).
959 */
960 for (i = 0; i < nr_pages; i++) {
961 page = pages[i];
962 set_page_private(page, (unsigned long)zspage);
963 page->index = 0;
964 if (i == 0) {
965 zspage->first_page = page;
966 SetPagePrivate(page);
967 if (unlikely(class->objs_per_zspage == 1 &&
968 class->pages_per_zspage == 1))
969 SetZsHugePage(zspage);
970 } else {
971 prev_page->index = (unsigned long)page;
972 }
973 prev_page = page;
974 }
975}
976
977/*
978 * Allocate a zspage for the given size class
979 */
980static struct zspage *alloc_zspage(struct zs_pool *pool,
981 struct size_class *class,
982 gfp_t gfp)
983{
984 int i;
985 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
986 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
987
988 if (!zspage)
989 return NULL;
990
991 zspage->magic = ZSPAGE_MAGIC;
992 migrate_lock_init(zspage);
993
994 for (i = 0; i < class->pages_per_zspage; i++) {
995 struct page *page;
996
997 page = alloc_page(gfp);
998 if (!page) {
999 while (--i >= 0) {
1000 dec_zone_page_state(pages[i], NR_ZSPAGES);
1001 __free_page(pages[i]);
1002 }
1003 cache_free_zspage(pool, zspage);
1004 return NULL;
1005 }
1006
1007 inc_zone_page_state(page, NR_ZSPAGES);
1008 pages[i] = page;
1009 }
1010
1011 create_page_chain(class, zspage, pages);
1012 init_zspage(class, zspage);
1013 zspage->pool = pool;
1014
1015 return zspage;
1016}
1017
1018static struct zspage *find_get_zspage(struct size_class *class)
1019{
1020 int i;
1021 struct zspage *zspage;
1022
1023 for (i = ZS_INUSE_RATIO_99; i >= ZS_INUSE_RATIO_0; i--) {
1024 zspage = list_first_entry_or_null(&class->fullness_list[i],
1025 struct zspage, list);
1026 if (zspage)
1027 break;
1028 }
1029
1030 return zspage;
1031}
1032
1033static inline int __zs_cpu_up(struct mapping_area *area)
1034{
1035 /*
1036 * Make sure we don't leak memory if a cpu UP notification
1037 * and zs_init() race and both call zs_cpu_up() on the same cpu
1038 */
1039 if (area->vm_buf)
1040 return 0;
1041 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1042 if (!area->vm_buf)
1043 return -ENOMEM;
1044 return 0;
1045}
1046
1047static inline void __zs_cpu_down(struct mapping_area *area)
1048{
1049 kfree(area->vm_buf);
1050 area->vm_buf = NULL;
1051}
1052
1053static void *__zs_map_object(struct mapping_area *area,
1054 struct page *pages[2], int off, int size)
1055{
1056 int sizes[2];
1057 void *addr;
1058 char *buf = area->vm_buf;
1059
1060 /* disable page faults to match kmap_atomic() return conditions */
1061 pagefault_disable();
1062
1063 /* no read fastpath */
1064 if (area->vm_mm == ZS_MM_WO)
1065 goto out;
1066
1067 sizes[0] = PAGE_SIZE - off;
1068 sizes[1] = size - sizes[0];
1069
1070 /* copy object to per-cpu buffer */
1071 addr = kmap_atomic(pages[0]);
1072 memcpy(buf, addr + off, sizes[0]);
1073 kunmap_atomic(addr);
1074 addr = kmap_atomic(pages[1]);
1075 memcpy(buf + sizes[0], addr, sizes[1]);
1076 kunmap_atomic(addr);
1077out:
1078 return area->vm_buf;
1079}
1080
1081static void __zs_unmap_object(struct mapping_area *area,
1082 struct page *pages[2], int off, int size)
1083{
1084 int sizes[2];
1085 void *addr;
1086 char *buf;
1087
1088 /* no write fastpath */
1089 if (area->vm_mm == ZS_MM_RO)
1090 goto out;
1091
1092 buf = area->vm_buf;
1093 buf = buf + ZS_HANDLE_SIZE;
1094 size -= ZS_HANDLE_SIZE;
1095 off += ZS_HANDLE_SIZE;
1096
1097 sizes[0] = PAGE_SIZE - off;
1098 sizes[1] = size - sizes[0];
1099
1100 /* copy per-cpu buffer to object */
1101 addr = kmap_atomic(pages[0]);
1102 memcpy(addr + off, buf, sizes[0]);
1103 kunmap_atomic(addr);
1104 addr = kmap_atomic(pages[1]);
1105 memcpy(addr, buf + sizes[0], sizes[1]);
1106 kunmap_atomic(addr);
1107
1108out:
1109 /* enable page faults to match kunmap_atomic() return conditions */
1110 pagefault_enable();
1111}
1112
1113static int zs_cpu_prepare(unsigned int cpu)
1114{
1115 struct mapping_area *area;
1116
1117 area = &per_cpu(zs_map_area, cpu);
1118 return __zs_cpu_up(area);
1119}
1120
1121static int zs_cpu_dead(unsigned int cpu)
1122{
1123 struct mapping_area *area;
1124
1125 area = &per_cpu(zs_map_area, cpu);
1126 __zs_cpu_down(area);
1127 return 0;
1128}
1129
1130static bool can_merge(struct size_class *prev, int pages_per_zspage,
1131 int objs_per_zspage)
1132{
1133 if (prev->pages_per_zspage == pages_per_zspage &&
1134 prev->objs_per_zspage == objs_per_zspage)
1135 return true;
1136
1137 return false;
1138}
1139
1140static bool zspage_full(struct size_class *class, struct zspage *zspage)
1141{
1142 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1143}
1144
1145static bool zspage_empty(struct zspage *zspage)
1146{
1147 return get_zspage_inuse(zspage) == 0;
1148}
1149
1150/**
1151 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1152 * that hold objects of the provided size.
1153 * @pool: zsmalloc pool to use
1154 * @size: object size
1155 *
1156 * Context: Any context.
1157 *
1158 * Return: the index of the zsmalloc &size_class that hold objects of the
1159 * provided size.
1160 */
1161unsigned int zs_lookup_class_index(struct zs_pool *pool, unsigned int size)
1162{
1163 struct size_class *class;
1164
1165 class = pool->size_class[get_size_class_index(size)];
1166
1167 return class->index;
1168}
1169EXPORT_SYMBOL_GPL(zs_lookup_class_index);
1170
1171unsigned long zs_get_total_pages(struct zs_pool *pool)
1172{
1173 return atomic_long_read(&pool->pages_allocated);
1174}
1175EXPORT_SYMBOL_GPL(zs_get_total_pages);
1176
1177/**
1178 * zs_map_object - get address of allocated object from handle.
1179 * @pool: pool from which the object was allocated
1180 * @handle: handle returned from zs_malloc
1181 * @mm: mapping mode to use
1182 *
1183 * Before using an object allocated from zs_malloc, it must be mapped using
1184 * this function. When done with the object, it must be unmapped using
1185 * zs_unmap_object.
1186 *
1187 * Only one object can be mapped per cpu at a time. There is no protection
1188 * against nested mappings.
1189 *
1190 * This function returns with preemption and page faults disabled.
1191 */
1192void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1193 enum zs_mapmode mm)
1194{
1195 struct zspage *zspage;
1196 struct page *page;
1197 unsigned long obj, off;
1198 unsigned int obj_idx;
1199
1200 struct size_class *class;
1201 struct mapping_area *area;
1202 struct page *pages[2];
1203 void *ret;
1204
1205 /*
1206 * Because we use per-cpu mapping areas shared among the
1207 * pools/users, we can't allow mapping in interrupt context
1208 * because it can corrupt another users mappings.
1209 */
1210 BUG_ON(in_interrupt());
1211
1212 /* It guarantees it can get zspage from handle safely */
1213 spin_lock(&pool->lock);
1214 obj = handle_to_obj(handle);
1215 obj_to_location(obj, &page, &obj_idx);
1216 zspage = get_zspage(page);
1217
1218 /*
1219 * migration cannot move any zpages in this zspage. Here, pool->lock
1220 * is too heavy since callers would take some time until they calls
1221 * zs_unmap_object API so delegate the locking from class to zspage
1222 * which is smaller granularity.
1223 */
1224 migrate_read_lock(zspage);
1225 spin_unlock(&pool->lock);
1226
1227 class = zspage_class(pool, zspage);
1228 off = offset_in_page(class->size * obj_idx);
1229
1230 local_lock(&zs_map_area.lock);
1231 area = this_cpu_ptr(&zs_map_area);
1232 area->vm_mm = mm;
1233 if (off + class->size <= PAGE_SIZE) {
1234 /* this object is contained entirely within a page */
1235 area->vm_addr = kmap_atomic(page);
1236 ret = area->vm_addr + off;
1237 goto out;
1238 }
1239
1240 /* this object spans two pages */
1241 pages[0] = page;
1242 pages[1] = get_next_page(page);
1243 BUG_ON(!pages[1]);
1244
1245 ret = __zs_map_object(area, pages, off, class->size);
1246out:
1247 if (likely(!ZsHugePage(zspage)))
1248 ret += ZS_HANDLE_SIZE;
1249
1250 return ret;
1251}
1252EXPORT_SYMBOL_GPL(zs_map_object);
1253
1254void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1255{
1256 struct zspage *zspage;
1257 struct page *page;
1258 unsigned long obj, off;
1259 unsigned int obj_idx;
1260
1261 struct size_class *class;
1262 struct mapping_area *area;
1263
1264 obj = handle_to_obj(handle);
1265 obj_to_location(obj, &page, &obj_idx);
1266 zspage = get_zspage(page);
1267 class = zspage_class(pool, zspage);
1268 off = offset_in_page(class->size * obj_idx);
1269
1270 area = this_cpu_ptr(&zs_map_area);
1271 if (off + class->size <= PAGE_SIZE)
1272 kunmap_atomic(area->vm_addr);
1273 else {
1274 struct page *pages[2];
1275
1276 pages[0] = page;
1277 pages[1] = get_next_page(page);
1278 BUG_ON(!pages[1]);
1279
1280 __zs_unmap_object(area, pages, off, class->size);
1281 }
1282 local_unlock(&zs_map_area.lock);
1283
1284 migrate_read_unlock(zspage);
1285}
1286EXPORT_SYMBOL_GPL(zs_unmap_object);
1287
1288/**
1289 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1290 * zsmalloc &size_class.
1291 * @pool: zsmalloc pool to use
1292 *
1293 * The function returns the size of the first huge class - any object of equal
1294 * or bigger size will be stored in zspage consisting of a single physical
1295 * page.
1296 *
1297 * Context: Any context.
1298 *
1299 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1300 */
1301size_t zs_huge_class_size(struct zs_pool *pool)
1302{
1303 return huge_class_size;
1304}
1305EXPORT_SYMBOL_GPL(zs_huge_class_size);
1306
1307static unsigned long obj_malloc(struct zs_pool *pool,
1308 struct zspage *zspage, unsigned long handle)
1309{
1310 int i, nr_page, offset;
1311 unsigned long obj;
1312 struct link_free *link;
1313 struct size_class *class;
1314
1315 struct page *m_page;
1316 unsigned long m_offset;
1317 void *vaddr;
1318
1319 class = pool->size_class[zspage->class];
1320 handle |= OBJ_ALLOCATED_TAG;
1321 obj = get_freeobj(zspage);
1322
1323 offset = obj * class->size;
1324 nr_page = offset >> PAGE_SHIFT;
1325 m_offset = offset_in_page(offset);
1326 m_page = get_first_page(zspage);
1327
1328 for (i = 0; i < nr_page; i++)
1329 m_page = get_next_page(m_page);
1330
1331 vaddr = kmap_atomic(m_page);
1332 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1333 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1334 if (likely(!ZsHugePage(zspage)))
1335 /* record handle in the header of allocated chunk */
1336 link->handle = handle;
1337 else
1338 /* record handle to page->index */
1339 zspage->first_page->index = handle;
1340
1341 kunmap_atomic(vaddr);
1342 mod_zspage_inuse(zspage, 1);
1343
1344 obj = location_to_obj(m_page, obj);
1345
1346 return obj;
1347}
1348
1349
1350/**
1351 * zs_malloc - Allocate block of given size from pool.
1352 * @pool: pool to allocate from
1353 * @size: size of block to allocate
1354 * @gfp: gfp flags when allocating object
1355 *
1356 * On success, handle to the allocated object is returned,
1357 * otherwise an ERR_PTR().
1358 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1359 */
1360unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1361{
1362 unsigned long handle, obj;
1363 struct size_class *class;
1364 int newfg;
1365 struct zspage *zspage;
1366
1367 if (unlikely(!size))
1368 return (unsigned long)ERR_PTR(-EINVAL);
1369
1370 if (unlikely(size > ZS_MAX_ALLOC_SIZE))
1371 return (unsigned long)ERR_PTR(-ENOSPC);
1372
1373 handle = cache_alloc_handle(pool, gfp);
1374 if (!handle)
1375 return (unsigned long)ERR_PTR(-ENOMEM);
1376
1377 /* extra space in chunk to keep the handle */
1378 size += ZS_HANDLE_SIZE;
1379 class = pool->size_class[get_size_class_index(size)];
1380
1381 /* pool->lock effectively protects the zpage migration */
1382 spin_lock(&pool->lock);
1383 zspage = find_get_zspage(class);
1384 if (likely(zspage)) {
1385 obj = obj_malloc(pool, zspage, handle);
1386 /* Now move the zspage to another fullness group, if required */
1387 fix_fullness_group(class, zspage);
1388 record_obj(handle, obj);
1389 class_stat_inc(class, ZS_OBJS_INUSE, 1);
1390
1391 goto out;
1392 }
1393
1394 spin_unlock(&pool->lock);
1395
1396 zspage = alloc_zspage(pool, class, gfp);
1397 if (!zspage) {
1398 cache_free_handle(pool, handle);
1399 return (unsigned long)ERR_PTR(-ENOMEM);
1400 }
1401
1402 spin_lock(&pool->lock);
1403 obj = obj_malloc(pool, zspage, handle);
1404 newfg = get_fullness_group(class, zspage);
1405 insert_zspage(class, zspage, newfg);
1406 set_zspage_mapping(zspage, class->index, newfg);
1407 record_obj(handle, obj);
1408 atomic_long_add(class->pages_per_zspage, &pool->pages_allocated);
1409 class_stat_inc(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
1410 class_stat_inc(class, ZS_OBJS_INUSE, 1);
1411
1412 /* We completely set up zspage so mark them as movable */
1413 SetZsPageMovable(pool, zspage);
1414out:
1415 spin_unlock(&pool->lock);
1416
1417 return handle;
1418}
1419EXPORT_SYMBOL_GPL(zs_malloc);
1420
1421static void obj_free(int class_size, unsigned long obj)
1422{
1423 struct link_free *link;
1424 struct zspage *zspage;
1425 struct page *f_page;
1426 unsigned long f_offset;
1427 unsigned int f_objidx;
1428 void *vaddr;
1429
1430 obj_to_location(obj, &f_page, &f_objidx);
1431 f_offset = offset_in_page(class_size * f_objidx);
1432 zspage = get_zspage(f_page);
1433
1434 vaddr = kmap_atomic(f_page);
1435 link = (struct link_free *)(vaddr + f_offset);
1436
1437 /* Insert this object in containing zspage's freelist */
1438 if (likely(!ZsHugePage(zspage)))
1439 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1440 else
1441 f_page->index = 0;
1442 set_freeobj(zspage, f_objidx);
1443
1444 kunmap_atomic(vaddr);
1445 mod_zspage_inuse(zspage, -1);
1446}
1447
1448void zs_free(struct zs_pool *pool, unsigned long handle)
1449{
1450 struct zspage *zspage;
1451 struct page *f_page;
1452 unsigned long obj;
1453 struct size_class *class;
1454 int fullness;
1455
1456 if (IS_ERR_OR_NULL((void *)handle))
1457 return;
1458
1459 /*
1460 * The pool->lock protects the race with zpage's migration
1461 * so it's safe to get the page from handle.
1462 */
1463 spin_lock(&pool->lock);
1464 obj = handle_to_obj(handle);
1465 obj_to_page(obj, &f_page);
1466 zspage = get_zspage(f_page);
1467 class = zspage_class(pool, zspage);
1468
1469 class_stat_dec(class, ZS_OBJS_INUSE, 1);
1470 obj_free(class->size, obj);
1471
1472 fullness = fix_fullness_group(class, zspage);
1473 if (fullness == ZS_INUSE_RATIO_0)
1474 free_zspage(pool, class, zspage);
1475
1476 spin_unlock(&pool->lock);
1477 cache_free_handle(pool, handle);
1478}
1479EXPORT_SYMBOL_GPL(zs_free);
1480
1481static void zs_object_copy(struct size_class *class, unsigned long dst,
1482 unsigned long src)
1483{
1484 struct page *s_page, *d_page;
1485 unsigned int s_objidx, d_objidx;
1486 unsigned long s_off, d_off;
1487 void *s_addr, *d_addr;
1488 int s_size, d_size, size;
1489 int written = 0;
1490
1491 s_size = d_size = class->size;
1492
1493 obj_to_location(src, &s_page, &s_objidx);
1494 obj_to_location(dst, &d_page, &d_objidx);
1495
1496 s_off = offset_in_page(class->size * s_objidx);
1497 d_off = offset_in_page(class->size * d_objidx);
1498
1499 if (s_off + class->size > PAGE_SIZE)
1500 s_size = PAGE_SIZE - s_off;
1501
1502 if (d_off + class->size > PAGE_SIZE)
1503 d_size = PAGE_SIZE - d_off;
1504
1505 s_addr = kmap_atomic(s_page);
1506 d_addr = kmap_atomic(d_page);
1507
1508 while (1) {
1509 size = min(s_size, d_size);
1510 memcpy(d_addr + d_off, s_addr + s_off, size);
1511 written += size;
1512
1513 if (written == class->size)
1514 break;
1515
1516 s_off += size;
1517 s_size -= size;
1518 d_off += size;
1519 d_size -= size;
1520
1521 /*
1522 * Calling kunmap_atomic(d_addr) is necessary. kunmap_atomic()
1523 * calls must occurs in reverse order of calls to kmap_atomic().
1524 * So, to call kunmap_atomic(s_addr) we should first call
1525 * kunmap_atomic(d_addr). For more details see
1526 * Documentation/mm/highmem.rst.
1527 */
1528 if (s_off >= PAGE_SIZE) {
1529 kunmap_atomic(d_addr);
1530 kunmap_atomic(s_addr);
1531 s_page = get_next_page(s_page);
1532 s_addr = kmap_atomic(s_page);
1533 d_addr = kmap_atomic(d_page);
1534 s_size = class->size - written;
1535 s_off = 0;
1536 }
1537
1538 if (d_off >= PAGE_SIZE) {
1539 kunmap_atomic(d_addr);
1540 d_page = get_next_page(d_page);
1541 d_addr = kmap_atomic(d_page);
1542 d_size = class->size - written;
1543 d_off = 0;
1544 }
1545 }
1546
1547 kunmap_atomic(d_addr);
1548 kunmap_atomic(s_addr);
1549}
1550
1551/*
1552 * Find alloced object in zspage from index object and
1553 * return handle.
1554 */
1555static unsigned long find_alloced_obj(struct size_class *class,
1556 struct page *page, int *obj_idx)
1557{
1558 unsigned int offset;
1559 int index = *obj_idx;
1560 unsigned long handle = 0;
1561 void *addr = kmap_atomic(page);
1562
1563 offset = get_first_obj_offset(page);
1564 offset += class->size * index;
1565
1566 while (offset < PAGE_SIZE) {
1567 if (obj_allocated(page, addr + offset, &handle))
1568 break;
1569
1570 offset += class->size;
1571 index++;
1572 }
1573
1574 kunmap_atomic(addr);
1575
1576 *obj_idx = index;
1577
1578 return handle;
1579}
1580
1581static void migrate_zspage(struct zs_pool *pool, struct zspage *src_zspage,
1582 struct zspage *dst_zspage)
1583{
1584 unsigned long used_obj, free_obj;
1585 unsigned long handle;
1586 int obj_idx = 0;
1587 struct page *s_page = get_first_page(src_zspage);
1588 struct size_class *class = pool->size_class[src_zspage->class];
1589
1590 while (1) {
1591 handle = find_alloced_obj(class, s_page, &obj_idx);
1592 if (!handle) {
1593 s_page = get_next_page(s_page);
1594 if (!s_page)
1595 break;
1596 obj_idx = 0;
1597 continue;
1598 }
1599
1600 used_obj = handle_to_obj(handle);
1601 free_obj = obj_malloc(pool, dst_zspage, handle);
1602 zs_object_copy(class, free_obj, used_obj);
1603 obj_idx++;
1604 record_obj(handle, free_obj);
1605 obj_free(class->size, used_obj);
1606
1607 /* Stop if there is no more space */
1608 if (zspage_full(class, dst_zspage))
1609 break;
1610
1611 /* Stop if there are no more objects to migrate */
1612 if (zspage_empty(src_zspage))
1613 break;
1614 }
1615}
1616
1617static struct zspage *isolate_src_zspage(struct size_class *class)
1618{
1619 struct zspage *zspage;
1620 int fg;
1621
1622 for (fg = ZS_INUSE_RATIO_10; fg <= ZS_INUSE_RATIO_99; fg++) {
1623 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1624 struct zspage, list);
1625 if (zspage) {
1626 remove_zspage(class, zspage, fg);
1627 return zspage;
1628 }
1629 }
1630
1631 return zspage;
1632}
1633
1634static struct zspage *isolate_dst_zspage(struct size_class *class)
1635{
1636 struct zspage *zspage;
1637 int fg;
1638
1639 for (fg = ZS_INUSE_RATIO_99; fg >= ZS_INUSE_RATIO_10; fg--) {
1640 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1641 struct zspage, list);
1642 if (zspage) {
1643 remove_zspage(class, zspage, fg);
1644 return zspage;
1645 }
1646 }
1647
1648 return zspage;
1649}
1650
1651/*
1652 * putback_zspage - add @zspage into right class's fullness list
1653 * @class: destination class
1654 * @zspage: target page
1655 *
1656 * Return @zspage's fullness status
1657 */
1658static int putback_zspage(struct size_class *class, struct zspage *zspage)
1659{
1660 int fullness;
1661
1662 fullness = get_fullness_group(class, zspage);
1663 insert_zspage(class, zspage, fullness);
1664 set_zspage_mapping(zspage, class->index, fullness);
1665
1666 return fullness;
1667}
1668
1669#ifdef CONFIG_COMPACTION
1670/*
1671 * To prevent zspage destroy during migration, zspage freeing should
1672 * hold locks of all pages in the zspage.
1673 */
1674static void lock_zspage(struct zspage *zspage)
1675{
1676 struct page *curr_page, *page;
1677
1678 /*
1679 * Pages we haven't locked yet can be migrated off the list while we're
1680 * trying to lock them, so we need to be careful and only attempt to
1681 * lock each page under migrate_read_lock(). Otherwise, the page we lock
1682 * may no longer belong to the zspage. This means that we may wait for
1683 * the wrong page to unlock, so we must take a reference to the page
1684 * prior to waiting for it to unlock outside migrate_read_lock().
1685 */
1686 while (1) {
1687 migrate_read_lock(zspage);
1688 page = get_first_page(zspage);
1689 if (trylock_page(page))
1690 break;
1691 get_page(page);
1692 migrate_read_unlock(zspage);
1693 wait_on_page_locked(page);
1694 put_page(page);
1695 }
1696
1697 curr_page = page;
1698 while ((page = get_next_page(curr_page))) {
1699 if (trylock_page(page)) {
1700 curr_page = page;
1701 } else {
1702 get_page(page);
1703 migrate_read_unlock(zspage);
1704 wait_on_page_locked(page);
1705 put_page(page);
1706 migrate_read_lock(zspage);
1707 }
1708 }
1709 migrate_read_unlock(zspage);
1710}
1711#endif /* CONFIG_COMPACTION */
1712
1713static void migrate_lock_init(struct zspage *zspage)
1714{
1715 rwlock_init(&zspage->lock);
1716}
1717
1718static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1719{
1720 read_lock(&zspage->lock);
1721}
1722
1723static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1724{
1725 read_unlock(&zspage->lock);
1726}
1727
1728#ifdef CONFIG_COMPACTION
1729static void migrate_write_lock(struct zspage *zspage)
1730{
1731 write_lock(&zspage->lock);
1732}
1733
1734static void migrate_write_lock_nested(struct zspage *zspage)
1735{
1736 write_lock_nested(&zspage->lock, SINGLE_DEPTH_NESTING);
1737}
1738
1739static void migrate_write_unlock(struct zspage *zspage)
1740{
1741 write_unlock(&zspage->lock);
1742}
1743
1744/* Number of isolated subpage for *page migration* in this zspage */
1745static void inc_zspage_isolation(struct zspage *zspage)
1746{
1747 zspage->isolated++;
1748}
1749
1750static void dec_zspage_isolation(struct zspage *zspage)
1751{
1752 VM_BUG_ON(zspage->isolated == 0);
1753 zspage->isolated--;
1754}
1755
1756static const struct movable_operations zsmalloc_mops;
1757
1758static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1759 struct page *newpage, struct page *oldpage)
1760{
1761 struct page *page;
1762 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1763 int idx = 0;
1764
1765 page = get_first_page(zspage);
1766 do {
1767 if (page == oldpage)
1768 pages[idx] = newpage;
1769 else
1770 pages[idx] = page;
1771 idx++;
1772 } while ((page = get_next_page(page)) != NULL);
1773
1774 create_page_chain(class, zspage, pages);
1775 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1776 if (unlikely(ZsHugePage(zspage)))
1777 newpage->index = oldpage->index;
1778 __SetPageMovable(newpage, &zsmalloc_mops);
1779}
1780
1781static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1782{
1783 struct zs_pool *pool;
1784 struct zspage *zspage;
1785
1786 /*
1787 * Page is locked so zspage couldn't be destroyed. For detail, look at
1788 * lock_zspage in free_zspage.
1789 */
1790 VM_BUG_ON_PAGE(PageIsolated(page), page);
1791
1792 zspage = get_zspage(page);
1793 pool = zspage->pool;
1794 spin_lock(&pool->lock);
1795 inc_zspage_isolation(zspage);
1796 spin_unlock(&pool->lock);
1797
1798 return true;
1799}
1800
1801static int zs_page_migrate(struct page *newpage, struct page *page,
1802 enum migrate_mode mode)
1803{
1804 struct zs_pool *pool;
1805 struct size_class *class;
1806 struct zspage *zspage;
1807 struct page *dummy;
1808 void *s_addr, *d_addr, *addr;
1809 unsigned int offset;
1810 unsigned long handle;
1811 unsigned long old_obj, new_obj;
1812 unsigned int obj_idx;
1813
1814 /*
1815 * We cannot support the _NO_COPY case here, because copy needs to
1816 * happen under the zs lock, which does not work with
1817 * MIGRATE_SYNC_NO_COPY workflow.
1818 */
1819 if (mode == MIGRATE_SYNC_NO_COPY)
1820 return -EINVAL;
1821
1822 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1823
1824 /* The page is locked, so this pointer must remain valid */
1825 zspage = get_zspage(page);
1826 pool = zspage->pool;
1827
1828 /*
1829 * The pool's lock protects the race between zpage migration
1830 * and zs_free.
1831 */
1832 spin_lock(&pool->lock);
1833 class = zspage_class(pool, zspage);
1834
1835 /* the migrate_write_lock protects zpage access via zs_map_object */
1836 migrate_write_lock(zspage);
1837
1838 offset = get_first_obj_offset(page);
1839 s_addr = kmap_atomic(page);
1840
1841 /*
1842 * Here, any user cannot access all objects in the zspage so let's move.
1843 */
1844 d_addr = kmap_atomic(newpage);
1845 copy_page(d_addr, s_addr);
1846 kunmap_atomic(d_addr);
1847
1848 for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE;
1849 addr += class->size) {
1850 if (obj_allocated(page, addr, &handle)) {
1851
1852 old_obj = handle_to_obj(handle);
1853 obj_to_location(old_obj, &dummy, &obj_idx);
1854 new_obj = (unsigned long)location_to_obj(newpage,
1855 obj_idx);
1856 record_obj(handle, new_obj);
1857 }
1858 }
1859 kunmap_atomic(s_addr);
1860
1861 replace_sub_page(class, zspage, newpage, page);
1862 dec_zspage_isolation(zspage);
1863 /*
1864 * Since we complete the data copy and set up new zspage structure,
1865 * it's okay to release the pool's lock.
1866 */
1867 spin_unlock(&pool->lock);
1868 migrate_write_unlock(zspage);
1869
1870 get_page(newpage);
1871 if (page_zone(newpage) != page_zone(page)) {
1872 dec_zone_page_state(page, NR_ZSPAGES);
1873 inc_zone_page_state(newpage, NR_ZSPAGES);
1874 }
1875
1876 reset_page(page);
1877 put_page(page);
1878
1879 return MIGRATEPAGE_SUCCESS;
1880}
1881
1882static void zs_page_putback(struct page *page)
1883{
1884 struct zs_pool *pool;
1885 struct zspage *zspage;
1886
1887 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1888
1889 zspage = get_zspage(page);
1890 pool = zspage->pool;
1891 spin_lock(&pool->lock);
1892 dec_zspage_isolation(zspage);
1893 spin_unlock(&pool->lock);
1894}
1895
1896static const struct movable_operations zsmalloc_mops = {
1897 .isolate_page = zs_page_isolate,
1898 .migrate_page = zs_page_migrate,
1899 .putback_page = zs_page_putback,
1900};
1901
1902/*
1903 * Caller should hold page_lock of all pages in the zspage
1904 * In here, we cannot use zspage meta data.
1905 */
1906static void async_free_zspage(struct work_struct *work)
1907{
1908 int i;
1909 struct size_class *class;
1910 unsigned int class_idx;
1911 int fullness;
1912 struct zspage *zspage, *tmp;
1913 LIST_HEAD(free_pages);
1914 struct zs_pool *pool = container_of(work, struct zs_pool,
1915 free_work);
1916
1917 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
1918 class = pool->size_class[i];
1919 if (class->index != i)
1920 continue;
1921
1922 spin_lock(&pool->lock);
1923 list_splice_init(&class->fullness_list[ZS_INUSE_RATIO_0],
1924 &free_pages);
1925 spin_unlock(&pool->lock);
1926 }
1927
1928 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
1929 list_del(&zspage->list);
1930 lock_zspage(zspage);
1931
1932 get_zspage_mapping(zspage, &class_idx, &fullness);
1933 VM_BUG_ON(fullness != ZS_INUSE_RATIO_0);
1934 class = pool->size_class[class_idx];
1935 spin_lock(&pool->lock);
1936 __free_zspage(pool, class, zspage);
1937 spin_unlock(&pool->lock);
1938 }
1939};
1940
1941static void kick_deferred_free(struct zs_pool *pool)
1942{
1943 schedule_work(&pool->free_work);
1944}
1945
1946static void zs_flush_migration(struct zs_pool *pool)
1947{
1948 flush_work(&pool->free_work);
1949}
1950
1951static void init_deferred_free(struct zs_pool *pool)
1952{
1953 INIT_WORK(&pool->free_work, async_free_zspage);
1954}
1955
1956static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
1957{
1958 struct page *page = get_first_page(zspage);
1959
1960 do {
1961 WARN_ON(!trylock_page(page));
1962 __SetPageMovable(page, &zsmalloc_mops);
1963 unlock_page(page);
1964 } while ((page = get_next_page(page)) != NULL);
1965}
1966#else
1967static inline void zs_flush_migration(struct zs_pool *pool) { }
1968#endif
1969
1970/*
1971 *
1972 * Based on the number of unused allocated objects calculate
1973 * and return the number of pages that we can free.
1974 */
1975static unsigned long zs_can_compact(struct size_class *class)
1976{
1977 unsigned long obj_wasted;
1978 unsigned long obj_allocated = zs_stat_get(class, ZS_OBJS_ALLOCATED);
1979 unsigned long obj_used = zs_stat_get(class, ZS_OBJS_INUSE);
1980
1981 if (obj_allocated <= obj_used)
1982 return 0;
1983
1984 obj_wasted = obj_allocated - obj_used;
1985 obj_wasted /= class->objs_per_zspage;
1986
1987 return obj_wasted * class->pages_per_zspage;
1988}
1989
1990static unsigned long __zs_compact(struct zs_pool *pool,
1991 struct size_class *class)
1992{
1993 struct zspage *src_zspage = NULL;
1994 struct zspage *dst_zspage = NULL;
1995 unsigned long pages_freed = 0;
1996
1997 /*
1998 * protect the race between zpage migration and zs_free
1999 * as well as zpage allocation/free
2000 */
2001 spin_lock(&pool->lock);
2002 while (zs_can_compact(class)) {
2003 int fg;
2004
2005 if (!dst_zspage) {
2006 dst_zspage = isolate_dst_zspage(class);
2007 if (!dst_zspage)
2008 break;
2009 migrate_write_lock(dst_zspage);
2010 }
2011
2012 src_zspage = isolate_src_zspage(class);
2013 if (!src_zspage)
2014 break;
2015
2016 migrate_write_lock_nested(src_zspage);
2017
2018 migrate_zspage(pool, src_zspage, dst_zspage);
2019 fg = putback_zspage(class, src_zspage);
2020 migrate_write_unlock(src_zspage);
2021
2022 if (fg == ZS_INUSE_RATIO_0) {
2023 free_zspage(pool, class, src_zspage);
2024 pages_freed += class->pages_per_zspage;
2025 }
2026 src_zspage = NULL;
2027
2028 if (get_fullness_group(class, dst_zspage) == ZS_INUSE_RATIO_100
2029 || spin_is_contended(&pool->lock)) {
2030 putback_zspage(class, dst_zspage);
2031 migrate_write_unlock(dst_zspage);
2032 dst_zspage = NULL;
2033
2034 spin_unlock(&pool->lock);
2035 cond_resched();
2036 spin_lock(&pool->lock);
2037 }
2038 }
2039
2040 if (src_zspage) {
2041 putback_zspage(class, src_zspage);
2042 migrate_write_unlock(src_zspage);
2043 }
2044
2045 if (dst_zspage) {
2046 putback_zspage(class, dst_zspage);
2047 migrate_write_unlock(dst_zspage);
2048 }
2049 spin_unlock(&pool->lock);
2050
2051 return pages_freed;
2052}
2053
2054unsigned long zs_compact(struct zs_pool *pool)
2055{
2056 int i;
2057 struct size_class *class;
2058 unsigned long pages_freed = 0;
2059
2060 /*
2061 * Pool compaction is performed under pool->lock so it is basically
2062 * single-threaded. Having more than one thread in __zs_compact()
2063 * will increase pool->lock contention, which will impact other
2064 * zsmalloc operations that need pool->lock.
2065 */
2066 if (atomic_xchg(&pool->compaction_in_progress, 1))
2067 return 0;
2068
2069 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2070 class = pool->size_class[i];
2071 if (class->index != i)
2072 continue;
2073 pages_freed += __zs_compact(pool, class);
2074 }
2075 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2076 atomic_set(&pool->compaction_in_progress, 0);
2077
2078 return pages_freed;
2079}
2080EXPORT_SYMBOL_GPL(zs_compact);
2081
2082void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2083{
2084 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2085}
2086EXPORT_SYMBOL_GPL(zs_pool_stats);
2087
2088static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2089 struct shrink_control *sc)
2090{
2091 unsigned long pages_freed;
2092 struct zs_pool *pool = shrinker->private_data;
2093
2094 /*
2095 * Compact classes and calculate compaction delta.
2096 * Can run concurrently with a manually triggered
2097 * (by user) compaction.
2098 */
2099 pages_freed = zs_compact(pool);
2100
2101 return pages_freed ? pages_freed : SHRINK_STOP;
2102}
2103
2104static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2105 struct shrink_control *sc)
2106{
2107 int i;
2108 struct size_class *class;
2109 unsigned long pages_to_free = 0;
2110 struct zs_pool *pool = shrinker->private_data;
2111
2112 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2113 class = pool->size_class[i];
2114 if (class->index != i)
2115 continue;
2116
2117 pages_to_free += zs_can_compact(class);
2118 }
2119
2120 return pages_to_free;
2121}
2122
2123static void zs_unregister_shrinker(struct zs_pool *pool)
2124{
2125 shrinker_free(pool->shrinker);
2126}
2127
2128static int zs_register_shrinker(struct zs_pool *pool)
2129{
2130 pool->shrinker = shrinker_alloc(0, "mm-zspool:%s", pool->name);
2131 if (!pool->shrinker)
2132 return -ENOMEM;
2133
2134 pool->shrinker->scan_objects = zs_shrinker_scan;
2135 pool->shrinker->count_objects = zs_shrinker_count;
2136 pool->shrinker->batch = 0;
2137 pool->shrinker->private_data = pool;
2138
2139 shrinker_register(pool->shrinker);
2140
2141 return 0;
2142}
2143
2144static int calculate_zspage_chain_size(int class_size)
2145{
2146 int i, min_waste = INT_MAX;
2147 int chain_size = 1;
2148
2149 if (is_power_of_2(class_size))
2150 return chain_size;
2151
2152 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
2153 int waste;
2154
2155 waste = (i * PAGE_SIZE) % class_size;
2156 if (waste < min_waste) {
2157 min_waste = waste;
2158 chain_size = i;
2159 }
2160 }
2161
2162 return chain_size;
2163}
2164
2165/**
2166 * zs_create_pool - Creates an allocation pool to work from.
2167 * @name: pool name to be created
2168 *
2169 * This function must be called before anything when using
2170 * the zsmalloc allocator.
2171 *
2172 * On success, a pointer to the newly created pool is returned,
2173 * otherwise NULL.
2174 */
2175struct zs_pool *zs_create_pool(const char *name)
2176{
2177 int i;
2178 struct zs_pool *pool;
2179 struct size_class *prev_class = NULL;
2180
2181 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2182 if (!pool)
2183 return NULL;
2184
2185 init_deferred_free(pool);
2186 spin_lock_init(&pool->lock);
2187 atomic_set(&pool->compaction_in_progress, 0);
2188
2189 pool->name = kstrdup(name, GFP_KERNEL);
2190 if (!pool->name)
2191 goto err;
2192
2193 if (create_cache(pool))
2194 goto err;
2195
2196 /*
2197 * Iterate reversely, because, size of size_class that we want to use
2198 * for merging should be larger or equal to current size.
2199 */
2200 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2201 int size;
2202 int pages_per_zspage;
2203 int objs_per_zspage;
2204 struct size_class *class;
2205 int fullness;
2206
2207 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2208 if (size > ZS_MAX_ALLOC_SIZE)
2209 size = ZS_MAX_ALLOC_SIZE;
2210 pages_per_zspage = calculate_zspage_chain_size(size);
2211 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2212
2213 /*
2214 * We iterate from biggest down to smallest classes,
2215 * so huge_class_size holds the size of the first huge
2216 * class. Any object bigger than or equal to that will
2217 * endup in the huge class.
2218 */
2219 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2220 !huge_class_size) {
2221 huge_class_size = size;
2222 /*
2223 * The object uses ZS_HANDLE_SIZE bytes to store the
2224 * handle. We need to subtract it, because zs_malloc()
2225 * unconditionally adds handle size before it performs
2226 * size class search - so object may be smaller than
2227 * huge class size, yet it still can end up in the huge
2228 * class because it grows by ZS_HANDLE_SIZE extra bytes
2229 * right before class lookup.
2230 */
2231 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2232 }
2233
2234 /*
2235 * size_class is used for normal zsmalloc operation such
2236 * as alloc/free for that size. Although it is natural that we
2237 * have one size_class for each size, there is a chance that we
2238 * can get more memory utilization if we use one size_class for
2239 * many different sizes whose size_class have same
2240 * characteristics. So, we makes size_class point to
2241 * previous size_class if possible.
2242 */
2243 if (prev_class) {
2244 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2245 pool->size_class[i] = prev_class;
2246 continue;
2247 }
2248 }
2249
2250 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2251 if (!class)
2252 goto err;
2253
2254 class->size = size;
2255 class->index = i;
2256 class->pages_per_zspage = pages_per_zspage;
2257 class->objs_per_zspage = objs_per_zspage;
2258 pool->size_class[i] = class;
2259
2260 fullness = ZS_INUSE_RATIO_0;
2261 while (fullness < NR_FULLNESS_GROUPS) {
2262 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2263 fullness++;
2264 }
2265
2266 prev_class = class;
2267 }
2268
2269 /* debug only, don't abort if it fails */
2270 zs_pool_stat_create(pool, name);
2271
2272 /*
2273 * Not critical since shrinker is only used to trigger internal
2274 * defragmentation of the pool which is pretty optional thing. If
2275 * registration fails we still can use the pool normally and user can
2276 * trigger compaction manually. Thus, ignore return code.
2277 */
2278 zs_register_shrinker(pool);
2279
2280 return pool;
2281
2282err:
2283 zs_destroy_pool(pool);
2284 return NULL;
2285}
2286EXPORT_SYMBOL_GPL(zs_create_pool);
2287
2288void zs_destroy_pool(struct zs_pool *pool)
2289{
2290 int i;
2291
2292 zs_unregister_shrinker(pool);
2293 zs_flush_migration(pool);
2294 zs_pool_stat_destroy(pool);
2295
2296 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2297 int fg;
2298 struct size_class *class = pool->size_class[i];
2299
2300 if (!class)
2301 continue;
2302
2303 if (class->index != i)
2304 continue;
2305
2306 for (fg = ZS_INUSE_RATIO_0; fg < NR_FULLNESS_GROUPS; fg++) {
2307 if (list_empty(&class->fullness_list[fg]))
2308 continue;
2309
2310 pr_err("Class-%d fullness group %d is not empty\n",
2311 class->size, fg);
2312 }
2313 kfree(class);
2314 }
2315
2316 destroy_cache(pool);
2317 kfree(pool->name);
2318 kfree(pool);
2319}
2320EXPORT_SYMBOL_GPL(zs_destroy_pool);
2321
2322static int __init zs_init(void)
2323{
2324 int ret;
2325
2326 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2327 zs_cpu_prepare, zs_cpu_dead);
2328 if (ret)
2329 goto out;
2330
2331#ifdef CONFIG_ZPOOL
2332 zpool_register_driver(&zs_zpool_driver);
2333#endif
2334
2335 zs_stat_init();
2336
2337 return 0;
2338
2339out:
2340 return ret;
2341}
2342
2343static void __exit zs_exit(void)
2344{
2345#ifdef CONFIG_ZPOOL
2346 zpool_unregister_driver(&zs_zpool_driver);
2347#endif
2348 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2349
2350 zs_stat_exit();
2351}
2352
2353module_init(zs_init);
2354module_exit(zs_exit);
2355
2356MODULE_LICENSE("Dual BSD/GPL");
2357MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");