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