Loading...
Note: File does not exist in v3.1.
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/*
77 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
78 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
79 */
80#define ZS_MAX_ZSPAGE_ORDER 2
81#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
82
83#define ZS_HANDLE_SIZE (sizeof(unsigned long))
84
85/*
86 * Object location (<PFN>, <obj_idx>) is encoded as
87 * a single (unsigned long) handle value.
88 *
89 * Note that object index <obj_idx> starts from 0.
90 *
91 * This is made more complicated by various memory models and PAE.
92 */
93
94#ifndef MAX_POSSIBLE_PHYSMEM_BITS
95#ifdef MAX_PHYSMEM_BITS
96#define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
97#else
98/*
99 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
100 * be PAGE_SHIFT
101 */
102#define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
103#endif
104#endif
105
106#define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
107
108/*
109 * Head in allocated object should have OBJ_ALLOCATED_TAG
110 * to identify the object was allocated or not.
111 * It's okay to add the status bit in the least bit because
112 * header keeps handle which is 4byte-aligned address so we
113 * have room for two bit at least.
114 */
115#define OBJ_ALLOCATED_TAG 1
116
117#ifdef CONFIG_ZPOOL
118/*
119 * The second least-significant bit in the object's header identifies if the
120 * value stored at the header is a deferred handle from the last reclaim
121 * attempt.
122 *
123 * As noted above, this is valid because we have room for two bits.
124 */
125#define OBJ_DEFERRED_HANDLE_TAG 2
126#define OBJ_TAG_BITS 2
127#define OBJ_TAG_MASK (OBJ_ALLOCATED_TAG | OBJ_DEFERRED_HANDLE_TAG)
128#else
129#define OBJ_TAG_BITS 1
130#define OBJ_TAG_MASK OBJ_ALLOCATED_TAG
131#endif /* CONFIG_ZPOOL */
132
133#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
134#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
135
136#define HUGE_BITS 1
137#define FULLNESS_BITS 2
138#define CLASS_BITS 8
139#define ISOLATED_BITS 3
140#define MAGIC_VAL_BITS 8
141
142#define MAX(a, b) ((a) >= (b) ? (a) : (b))
143/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
144#define ZS_MIN_ALLOC_SIZE \
145 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
146/* each chunk includes extra space to keep handle */
147#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
148
149/*
150 * On systems with 4K page size, this gives 255 size classes! There is a
151 * trader-off here:
152 * - Large number of size classes is potentially wasteful as free page are
153 * spread across these classes
154 * - Small number of size classes causes large internal fragmentation
155 * - Probably its better to use specific size classes (empirically
156 * determined). NOTE: all those class sizes must be set as multiple of
157 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
158 *
159 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
160 * (reason above)
161 */
162#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
163#define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
164 ZS_SIZE_CLASS_DELTA) + 1)
165
166enum fullness_group {
167 ZS_EMPTY,
168 ZS_ALMOST_EMPTY,
169 ZS_ALMOST_FULL,
170 ZS_FULL,
171 NR_ZS_FULLNESS,
172};
173
174enum class_stat_type {
175 CLASS_EMPTY,
176 CLASS_ALMOST_EMPTY,
177 CLASS_ALMOST_FULL,
178 CLASS_FULL,
179 OBJ_ALLOCATED,
180 OBJ_USED,
181 NR_ZS_STAT_TYPE,
182};
183
184struct zs_size_stat {
185 unsigned long objs[NR_ZS_STAT_TYPE];
186};
187
188#ifdef CONFIG_ZSMALLOC_STAT
189static struct dentry *zs_stat_root;
190#endif
191
192/*
193 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
194 * n <= N / f, where
195 * n = number of allocated objects
196 * N = total number of objects zspage can store
197 * f = fullness_threshold_frac
198 *
199 * Similarly, we assign zspage to:
200 * ZS_ALMOST_FULL when n > N / f
201 * ZS_EMPTY when n == 0
202 * ZS_FULL when n == N
203 *
204 * (see: fix_fullness_group())
205 */
206static const int fullness_threshold_frac = 4;
207static size_t huge_class_size;
208
209struct size_class {
210 struct list_head fullness_list[NR_ZS_FULLNESS];
211 /*
212 * Size of objects stored in this class. Must be multiple
213 * of ZS_ALIGN.
214 */
215 int size;
216 int objs_per_zspage;
217 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
218 int pages_per_zspage;
219
220 unsigned int index;
221 struct zs_size_stat stats;
222};
223
224/*
225 * Placed within free objects to form a singly linked list.
226 * For every zspage, zspage->freeobj gives head of this list.
227 *
228 * This must be power of 2 and less than or equal to ZS_ALIGN
229 */
230struct link_free {
231 union {
232 /*
233 * Free object index;
234 * It's valid for non-allocated object
235 */
236 unsigned long next;
237 /*
238 * Handle of allocated object.
239 */
240 unsigned long handle;
241#ifdef CONFIG_ZPOOL
242 /*
243 * Deferred handle of a reclaimed object.
244 */
245 unsigned long deferred_handle;
246#endif
247 };
248};
249
250struct zs_pool {
251 const char *name;
252
253 struct size_class *size_class[ZS_SIZE_CLASSES];
254 struct kmem_cache *handle_cachep;
255 struct kmem_cache *zspage_cachep;
256
257 atomic_long_t pages_allocated;
258
259 struct zs_pool_stats stats;
260
261 /* Compact classes */
262 struct shrinker shrinker;
263
264#ifdef CONFIG_ZPOOL
265 /* List tracking the zspages in LRU order by most recently added object */
266 struct list_head lru;
267 struct zpool *zpool;
268 const struct zpool_ops *zpool_ops;
269#endif
270
271#ifdef CONFIG_ZSMALLOC_STAT
272 struct dentry *stat_dentry;
273#endif
274#ifdef CONFIG_COMPACTION
275 struct work_struct free_work;
276#endif
277 spinlock_t lock;
278};
279
280struct zspage {
281 struct {
282 unsigned int huge:HUGE_BITS;
283 unsigned int fullness:FULLNESS_BITS;
284 unsigned int class:CLASS_BITS + 1;
285 unsigned int isolated:ISOLATED_BITS;
286 unsigned int magic:MAGIC_VAL_BITS;
287 };
288 unsigned int inuse;
289 unsigned int freeobj;
290 struct page *first_page;
291 struct list_head list; /* fullness list */
292
293#ifdef CONFIG_ZPOOL
294 /* links the zspage to the lru list in the pool */
295 struct list_head lru;
296 bool under_reclaim;
297#endif
298
299 struct zs_pool *pool;
300 rwlock_t lock;
301};
302
303struct mapping_area {
304 local_lock_t lock;
305 char *vm_buf; /* copy buffer for objects that span pages */
306 char *vm_addr; /* address of kmap_atomic()'ed pages */
307 enum zs_mapmode vm_mm; /* mapping mode */
308};
309
310/* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
311static void SetZsHugePage(struct zspage *zspage)
312{
313 zspage->huge = 1;
314}
315
316static bool ZsHugePage(struct zspage *zspage)
317{
318 return zspage->huge;
319}
320
321static void migrate_lock_init(struct zspage *zspage);
322static void migrate_read_lock(struct zspage *zspage);
323static void migrate_read_unlock(struct zspage *zspage);
324
325#ifdef CONFIG_COMPACTION
326static void migrate_write_lock(struct zspage *zspage);
327static void migrate_write_lock_nested(struct zspage *zspage);
328static void migrate_write_unlock(struct zspage *zspage);
329static void kick_deferred_free(struct zs_pool *pool);
330static void init_deferred_free(struct zs_pool *pool);
331static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
332#else
333static void migrate_write_lock(struct zspage *zspage) {}
334static void migrate_write_lock_nested(struct zspage *zspage) {}
335static void migrate_write_unlock(struct zspage *zspage) {}
336static void kick_deferred_free(struct zs_pool *pool) {}
337static void init_deferred_free(struct zs_pool *pool) {}
338static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
339#endif
340
341static int create_cache(struct zs_pool *pool)
342{
343 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
344 0, 0, NULL);
345 if (!pool->handle_cachep)
346 return 1;
347
348 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
349 0, 0, NULL);
350 if (!pool->zspage_cachep) {
351 kmem_cache_destroy(pool->handle_cachep);
352 pool->handle_cachep = NULL;
353 return 1;
354 }
355
356 return 0;
357}
358
359static void destroy_cache(struct zs_pool *pool)
360{
361 kmem_cache_destroy(pool->handle_cachep);
362 kmem_cache_destroy(pool->zspage_cachep);
363}
364
365static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
366{
367 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
368 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
369}
370
371static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
372{
373 kmem_cache_free(pool->handle_cachep, (void *)handle);
374}
375
376static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
377{
378 return kmem_cache_zalloc(pool->zspage_cachep,
379 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
380}
381
382static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
383{
384 kmem_cache_free(pool->zspage_cachep, zspage);
385}
386
387/* pool->lock(which owns the handle) synchronizes races */
388static void record_obj(unsigned long handle, unsigned long obj)
389{
390 *(unsigned long *)handle = obj;
391}
392
393/* zpool driver */
394
395#ifdef CONFIG_ZPOOL
396
397static void *zs_zpool_create(const char *name, gfp_t gfp,
398 const struct zpool_ops *zpool_ops,
399 struct zpool *zpool)
400{
401 /*
402 * Ignore global gfp flags: zs_malloc() may be invoked from
403 * different contexts and its caller must provide a valid
404 * gfp mask.
405 */
406 struct zs_pool *pool = zs_create_pool(name);
407
408 if (pool) {
409 pool->zpool = zpool;
410 pool->zpool_ops = zpool_ops;
411 }
412
413 return pool;
414}
415
416static void zs_zpool_destroy(void *pool)
417{
418 zs_destroy_pool(pool);
419}
420
421static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
422 unsigned long *handle)
423{
424 *handle = zs_malloc(pool, size, gfp);
425
426 if (IS_ERR_VALUE(*handle))
427 return PTR_ERR((void *)*handle);
428 return 0;
429}
430static void zs_zpool_free(void *pool, unsigned long handle)
431{
432 zs_free(pool, handle);
433}
434
435static int zs_reclaim_page(struct zs_pool *pool, unsigned int retries);
436
437static int zs_zpool_shrink(void *pool, unsigned int pages,
438 unsigned int *reclaimed)
439{
440 unsigned int total = 0;
441 int ret = -EINVAL;
442
443 while (total < pages) {
444 ret = zs_reclaim_page(pool, 8);
445 if (ret < 0)
446 break;
447 total++;
448 }
449
450 if (reclaimed)
451 *reclaimed = total;
452
453 return ret;
454}
455
456static void *zs_zpool_map(void *pool, unsigned long handle,
457 enum zpool_mapmode mm)
458{
459 enum zs_mapmode zs_mm;
460
461 switch (mm) {
462 case ZPOOL_MM_RO:
463 zs_mm = ZS_MM_RO;
464 break;
465 case ZPOOL_MM_WO:
466 zs_mm = ZS_MM_WO;
467 break;
468 case ZPOOL_MM_RW:
469 default:
470 zs_mm = ZS_MM_RW;
471 break;
472 }
473
474 return zs_map_object(pool, handle, zs_mm);
475}
476static void zs_zpool_unmap(void *pool, unsigned long handle)
477{
478 zs_unmap_object(pool, handle);
479}
480
481static u64 zs_zpool_total_size(void *pool)
482{
483 return zs_get_total_pages(pool) << PAGE_SHIFT;
484}
485
486static struct zpool_driver zs_zpool_driver = {
487 .type = "zsmalloc",
488 .owner = THIS_MODULE,
489 .create = zs_zpool_create,
490 .destroy = zs_zpool_destroy,
491 .malloc_support_movable = true,
492 .malloc = zs_zpool_malloc,
493 .free = zs_zpool_free,
494 .shrink = zs_zpool_shrink,
495 .map = zs_zpool_map,
496 .unmap = zs_zpool_unmap,
497 .total_size = zs_zpool_total_size,
498};
499
500MODULE_ALIAS("zpool-zsmalloc");
501#endif /* CONFIG_ZPOOL */
502
503/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
504static DEFINE_PER_CPU(struct mapping_area, zs_map_area) = {
505 .lock = INIT_LOCAL_LOCK(lock),
506};
507
508static __maybe_unused int is_first_page(struct page *page)
509{
510 return PagePrivate(page);
511}
512
513/* Protected by pool->lock */
514static inline int get_zspage_inuse(struct zspage *zspage)
515{
516 return zspage->inuse;
517}
518
519
520static inline void mod_zspage_inuse(struct zspage *zspage, int val)
521{
522 zspage->inuse += val;
523}
524
525static inline struct page *get_first_page(struct zspage *zspage)
526{
527 struct page *first_page = zspage->first_page;
528
529 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
530 return first_page;
531}
532
533static inline unsigned int get_first_obj_offset(struct page *page)
534{
535 return page->page_type;
536}
537
538static inline void set_first_obj_offset(struct page *page, unsigned int offset)
539{
540 page->page_type = offset;
541}
542
543static inline unsigned int get_freeobj(struct zspage *zspage)
544{
545 return zspage->freeobj;
546}
547
548static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
549{
550 zspage->freeobj = obj;
551}
552
553static void get_zspage_mapping(struct zspage *zspage,
554 unsigned int *class_idx,
555 enum fullness_group *fullness)
556{
557 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
558
559 *fullness = zspage->fullness;
560 *class_idx = zspage->class;
561}
562
563static struct size_class *zspage_class(struct zs_pool *pool,
564 struct zspage *zspage)
565{
566 return pool->size_class[zspage->class];
567}
568
569static void set_zspage_mapping(struct zspage *zspage,
570 unsigned int class_idx,
571 enum fullness_group fullness)
572{
573 zspage->class = class_idx;
574 zspage->fullness = fullness;
575}
576
577/*
578 * zsmalloc divides the pool into various size classes where each
579 * class maintains a list of zspages where each zspage is divided
580 * into equal sized chunks. Each allocation falls into one of these
581 * classes depending on its size. This function returns index of the
582 * size class which has chunk size big enough to hold the given size.
583 */
584static int get_size_class_index(int size)
585{
586 int idx = 0;
587
588 if (likely(size > ZS_MIN_ALLOC_SIZE))
589 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
590 ZS_SIZE_CLASS_DELTA);
591
592 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
593}
594
595/* type can be of enum type class_stat_type or fullness_group */
596static inline void class_stat_inc(struct size_class *class,
597 int type, unsigned long cnt)
598{
599 class->stats.objs[type] += cnt;
600}
601
602/* type can be of enum type class_stat_type or fullness_group */
603static inline void class_stat_dec(struct size_class *class,
604 int type, unsigned long cnt)
605{
606 class->stats.objs[type] -= cnt;
607}
608
609/* type can be of enum type class_stat_type or fullness_group */
610static inline unsigned long zs_stat_get(struct size_class *class,
611 int type)
612{
613 return class->stats.objs[type];
614}
615
616#ifdef CONFIG_ZSMALLOC_STAT
617
618static void __init zs_stat_init(void)
619{
620 if (!debugfs_initialized()) {
621 pr_warn("debugfs not available, stat dir not created\n");
622 return;
623 }
624
625 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
626}
627
628static void __exit zs_stat_exit(void)
629{
630 debugfs_remove_recursive(zs_stat_root);
631}
632
633static unsigned long zs_can_compact(struct size_class *class);
634
635static int zs_stats_size_show(struct seq_file *s, void *v)
636{
637 int i;
638 struct zs_pool *pool = s->private;
639 struct size_class *class;
640 int objs_per_zspage;
641 unsigned long class_almost_full, class_almost_empty;
642 unsigned long obj_allocated, obj_used, pages_used, freeable;
643 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
644 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
645 unsigned long total_freeable = 0;
646
647 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
648 "class", "size", "almost_full", "almost_empty",
649 "obj_allocated", "obj_used", "pages_used",
650 "pages_per_zspage", "freeable");
651
652 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
653 class = pool->size_class[i];
654
655 if (class->index != i)
656 continue;
657
658 spin_lock(&pool->lock);
659 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
660 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
661 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
662 obj_used = zs_stat_get(class, OBJ_USED);
663 freeable = zs_can_compact(class);
664 spin_unlock(&pool->lock);
665
666 objs_per_zspage = class->objs_per_zspage;
667 pages_used = obj_allocated / objs_per_zspage *
668 class->pages_per_zspage;
669
670 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
671 " %10lu %10lu %16d %8lu\n",
672 i, class->size, class_almost_full, class_almost_empty,
673 obj_allocated, obj_used, pages_used,
674 class->pages_per_zspage, freeable);
675
676 total_class_almost_full += class_almost_full;
677 total_class_almost_empty += class_almost_empty;
678 total_objs += obj_allocated;
679 total_used_objs += obj_used;
680 total_pages += pages_used;
681 total_freeable += freeable;
682 }
683
684 seq_puts(s, "\n");
685 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
686 "Total", "", total_class_almost_full,
687 total_class_almost_empty, total_objs,
688 total_used_objs, total_pages, "", total_freeable);
689
690 return 0;
691}
692DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
693
694static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
695{
696 if (!zs_stat_root) {
697 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
698 return;
699 }
700
701 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
702
703 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
704 &zs_stats_size_fops);
705}
706
707static void zs_pool_stat_destroy(struct zs_pool *pool)
708{
709 debugfs_remove_recursive(pool->stat_dentry);
710}
711
712#else /* CONFIG_ZSMALLOC_STAT */
713static void __init zs_stat_init(void)
714{
715}
716
717static void __exit zs_stat_exit(void)
718{
719}
720
721static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
722{
723}
724
725static inline void zs_pool_stat_destroy(struct zs_pool *pool)
726{
727}
728#endif
729
730
731/*
732 * For each size class, zspages are divided into different groups
733 * depending on how "full" they are. This was done so that we could
734 * easily find empty or nearly empty zspages when we try to shrink
735 * the pool (not yet implemented). This function returns fullness
736 * status of the given page.
737 */
738static enum fullness_group get_fullness_group(struct size_class *class,
739 struct zspage *zspage)
740{
741 int inuse, objs_per_zspage;
742 enum fullness_group fg;
743
744 inuse = get_zspage_inuse(zspage);
745 objs_per_zspage = class->objs_per_zspage;
746
747 if (inuse == 0)
748 fg = ZS_EMPTY;
749 else if (inuse == objs_per_zspage)
750 fg = ZS_FULL;
751 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
752 fg = ZS_ALMOST_EMPTY;
753 else
754 fg = ZS_ALMOST_FULL;
755
756 return fg;
757}
758
759/*
760 * Each size class maintains various freelists and zspages are assigned
761 * to one of these freelists based on the number of live objects they
762 * have. This functions inserts the given zspage into the freelist
763 * identified by <class, fullness_group>.
764 */
765static void insert_zspage(struct size_class *class,
766 struct zspage *zspage,
767 enum fullness_group fullness)
768{
769 struct zspage *head;
770
771 class_stat_inc(class, fullness, 1);
772 head = list_first_entry_or_null(&class->fullness_list[fullness],
773 struct zspage, list);
774 /*
775 * We want to see more ZS_FULL pages and less almost empty/full.
776 * Put pages with higher ->inuse first.
777 */
778 if (head && get_zspage_inuse(zspage) < get_zspage_inuse(head))
779 list_add(&zspage->list, &head->list);
780 else
781 list_add(&zspage->list, &class->fullness_list[fullness]);
782}
783
784/*
785 * This function removes the given zspage from the freelist identified
786 * by <class, fullness_group>.
787 */
788static void remove_zspage(struct size_class *class,
789 struct zspage *zspage,
790 enum fullness_group fullness)
791{
792 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
793
794 list_del_init(&zspage->list);
795 class_stat_dec(class, fullness, 1);
796}
797
798/*
799 * Each size class maintains zspages in different fullness groups depending
800 * on the number of live objects they contain. When allocating or freeing
801 * objects, the fullness status of the page can change, say, from ALMOST_FULL
802 * to ALMOST_EMPTY when freeing an object. This function checks if such
803 * a status change has occurred for the given page and accordingly moves the
804 * page from the freelist of the old fullness group to that of the new
805 * fullness group.
806 */
807static enum fullness_group fix_fullness_group(struct size_class *class,
808 struct zspage *zspage)
809{
810 int class_idx;
811 enum fullness_group currfg, newfg;
812
813 get_zspage_mapping(zspage, &class_idx, &currfg);
814 newfg = get_fullness_group(class, zspage);
815 if (newfg == currfg)
816 goto out;
817
818 remove_zspage(class, zspage, currfg);
819 insert_zspage(class, zspage, newfg);
820 set_zspage_mapping(zspage, class_idx, newfg);
821out:
822 return newfg;
823}
824
825/*
826 * We have to decide on how many pages to link together
827 * to form a zspage for each size class. This is important
828 * to reduce wastage due to unusable space left at end of
829 * each zspage which is given as:
830 * wastage = Zp % class_size
831 * usage = Zp - wastage
832 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
833 *
834 * For example, for size class of 3/8 * PAGE_SIZE, we should
835 * link together 3 PAGE_SIZE sized pages to form a zspage
836 * since then we can perfectly fit in 8 such objects.
837 */
838static int get_pages_per_zspage(int class_size)
839{
840 int i, max_usedpc = 0;
841 /* zspage order which gives maximum used size per KB */
842 int max_usedpc_order = 1;
843
844 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
845 int zspage_size;
846 int waste, usedpc;
847
848 zspage_size = i * PAGE_SIZE;
849 waste = zspage_size % class_size;
850 usedpc = (zspage_size - waste) * 100 / zspage_size;
851
852 if (usedpc > max_usedpc) {
853 max_usedpc = usedpc;
854 max_usedpc_order = i;
855 }
856 }
857
858 return max_usedpc_order;
859}
860
861static struct zspage *get_zspage(struct page *page)
862{
863 struct zspage *zspage = (struct zspage *)page_private(page);
864
865 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
866 return zspage;
867}
868
869static struct page *get_next_page(struct page *page)
870{
871 struct zspage *zspage = get_zspage(page);
872
873 if (unlikely(ZsHugePage(zspage)))
874 return NULL;
875
876 return (struct page *)page->index;
877}
878
879/**
880 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
881 * @obj: the encoded object value
882 * @page: page object resides in zspage
883 * @obj_idx: object index
884 */
885static void obj_to_location(unsigned long obj, struct page **page,
886 unsigned int *obj_idx)
887{
888 obj >>= OBJ_TAG_BITS;
889 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
890 *obj_idx = (obj & OBJ_INDEX_MASK);
891}
892
893static void obj_to_page(unsigned long obj, struct page **page)
894{
895 obj >>= OBJ_TAG_BITS;
896 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
897}
898
899/**
900 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
901 * @page: page object resides in zspage
902 * @obj_idx: object index
903 */
904static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
905{
906 unsigned long obj;
907
908 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
909 obj |= obj_idx & OBJ_INDEX_MASK;
910 obj <<= OBJ_TAG_BITS;
911
912 return obj;
913}
914
915static unsigned long handle_to_obj(unsigned long handle)
916{
917 return *(unsigned long *)handle;
918}
919
920static bool obj_tagged(struct page *page, void *obj, unsigned long *phandle,
921 int tag)
922{
923 unsigned long handle;
924 struct zspage *zspage = get_zspage(page);
925
926 if (unlikely(ZsHugePage(zspage))) {
927 VM_BUG_ON_PAGE(!is_first_page(page), page);
928 handle = page->index;
929 } else
930 handle = *(unsigned long *)obj;
931
932 if (!(handle & tag))
933 return false;
934
935 /* Clear all tags before returning the handle */
936 *phandle = handle & ~OBJ_TAG_MASK;
937 return true;
938}
939
940static inline bool obj_allocated(struct page *page, void *obj, unsigned long *phandle)
941{
942 return obj_tagged(page, obj, phandle, OBJ_ALLOCATED_TAG);
943}
944
945#ifdef CONFIG_ZPOOL
946static bool obj_stores_deferred_handle(struct page *page, void *obj,
947 unsigned long *phandle)
948{
949 return obj_tagged(page, obj, phandle, OBJ_DEFERRED_HANDLE_TAG);
950}
951#endif
952
953static void reset_page(struct page *page)
954{
955 __ClearPageMovable(page);
956 ClearPagePrivate(page);
957 set_page_private(page, 0);
958 page_mapcount_reset(page);
959 page->index = 0;
960}
961
962static int trylock_zspage(struct zspage *zspage)
963{
964 struct page *cursor, *fail;
965
966 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
967 get_next_page(cursor)) {
968 if (!trylock_page(cursor)) {
969 fail = cursor;
970 goto unlock;
971 }
972 }
973
974 return 1;
975unlock:
976 for (cursor = get_first_page(zspage); cursor != fail; cursor =
977 get_next_page(cursor))
978 unlock_page(cursor);
979
980 return 0;
981}
982
983#ifdef CONFIG_ZPOOL
984static unsigned long find_deferred_handle_obj(struct size_class *class,
985 struct page *page, int *obj_idx);
986
987/*
988 * Free all the deferred handles whose objects are freed in zs_free.
989 */
990static void free_handles(struct zs_pool *pool, struct size_class *class,
991 struct zspage *zspage)
992{
993 int obj_idx = 0;
994 struct page *page = get_first_page(zspage);
995 unsigned long handle;
996
997 while (1) {
998 handle = find_deferred_handle_obj(class, page, &obj_idx);
999 if (!handle) {
1000 page = get_next_page(page);
1001 if (!page)
1002 break;
1003 obj_idx = 0;
1004 continue;
1005 }
1006
1007 cache_free_handle(pool, handle);
1008 obj_idx++;
1009 }
1010}
1011#else
1012static inline void free_handles(struct zs_pool *pool, struct size_class *class,
1013 struct zspage *zspage) {}
1014#endif
1015
1016static void __free_zspage(struct zs_pool *pool, struct size_class *class,
1017 struct zspage *zspage)
1018{
1019 struct page *page, *next;
1020 enum fullness_group fg;
1021 unsigned int class_idx;
1022
1023 get_zspage_mapping(zspage, &class_idx, &fg);
1024
1025 assert_spin_locked(&pool->lock);
1026
1027 VM_BUG_ON(get_zspage_inuse(zspage));
1028 VM_BUG_ON(fg != ZS_EMPTY);
1029
1030 /* Free all deferred handles from zs_free */
1031 free_handles(pool, class, zspage);
1032
1033 next = page = get_first_page(zspage);
1034 do {
1035 VM_BUG_ON_PAGE(!PageLocked(page), page);
1036 next = get_next_page(page);
1037 reset_page(page);
1038 unlock_page(page);
1039 dec_zone_page_state(page, NR_ZSPAGES);
1040 put_page(page);
1041 page = next;
1042 } while (page != NULL);
1043
1044 cache_free_zspage(pool, zspage);
1045
1046 class_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
1047 atomic_long_sub(class->pages_per_zspage,
1048 &pool->pages_allocated);
1049}
1050
1051static void free_zspage(struct zs_pool *pool, struct size_class *class,
1052 struct zspage *zspage)
1053{
1054 VM_BUG_ON(get_zspage_inuse(zspage));
1055 VM_BUG_ON(list_empty(&zspage->list));
1056
1057 /*
1058 * Since zs_free couldn't be sleepable, this function cannot call
1059 * lock_page. The page locks trylock_zspage got will be released
1060 * by __free_zspage.
1061 */
1062 if (!trylock_zspage(zspage)) {
1063 kick_deferred_free(pool);
1064 return;
1065 }
1066
1067 remove_zspage(class, zspage, ZS_EMPTY);
1068#ifdef CONFIG_ZPOOL
1069 list_del(&zspage->lru);
1070#endif
1071 __free_zspage(pool, class, zspage);
1072}
1073
1074/* Initialize a newly allocated zspage */
1075static void init_zspage(struct size_class *class, struct zspage *zspage)
1076{
1077 unsigned int freeobj = 1;
1078 unsigned long off = 0;
1079 struct page *page = get_first_page(zspage);
1080
1081 while (page) {
1082 struct page *next_page;
1083 struct link_free *link;
1084 void *vaddr;
1085
1086 set_first_obj_offset(page, off);
1087
1088 vaddr = kmap_atomic(page);
1089 link = (struct link_free *)vaddr + off / sizeof(*link);
1090
1091 while ((off += class->size) < PAGE_SIZE) {
1092 link->next = freeobj++ << OBJ_TAG_BITS;
1093 link += class->size / sizeof(*link);
1094 }
1095
1096 /*
1097 * We now come to the last (full or partial) object on this
1098 * page, which must point to the first object on the next
1099 * page (if present)
1100 */
1101 next_page = get_next_page(page);
1102 if (next_page) {
1103 link->next = freeobj++ << OBJ_TAG_BITS;
1104 } else {
1105 /*
1106 * Reset OBJ_TAG_BITS bit to last link to tell
1107 * whether it's allocated object or not.
1108 */
1109 link->next = -1UL << OBJ_TAG_BITS;
1110 }
1111 kunmap_atomic(vaddr);
1112 page = next_page;
1113 off %= PAGE_SIZE;
1114 }
1115
1116#ifdef CONFIG_ZPOOL
1117 INIT_LIST_HEAD(&zspage->lru);
1118 zspage->under_reclaim = false;
1119#endif
1120
1121 set_freeobj(zspage, 0);
1122}
1123
1124static void create_page_chain(struct size_class *class, struct zspage *zspage,
1125 struct page *pages[])
1126{
1127 int i;
1128 struct page *page;
1129 struct page *prev_page = NULL;
1130 int nr_pages = class->pages_per_zspage;
1131
1132 /*
1133 * Allocate individual pages and link them together as:
1134 * 1. all pages are linked together using page->index
1135 * 2. each sub-page point to zspage using page->private
1136 *
1137 * we set PG_private to identify the first page (i.e. no other sub-page
1138 * has this flag set).
1139 */
1140 for (i = 0; i < nr_pages; i++) {
1141 page = pages[i];
1142 set_page_private(page, (unsigned long)zspage);
1143 page->index = 0;
1144 if (i == 0) {
1145 zspage->first_page = page;
1146 SetPagePrivate(page);
1147 if (unlikely(class->objs_per_zspage == 1 &&
1148 class->pages_per_zspage == 1))
1149 SetZsHugePage(zspage);
1150 } else {
1151 prev_page->index = (unsigned long)page;
1152 }
1153 prev_page = page;
1154 }
1155}
1156
1157/*
1158 * Allocate a zspage for the given size class
1159 */
1160static struct zspage *alloc_zspage(struct zs_pool *pool,
1161 struct size_class *class,
1162 gfp_t gfp)
1163{
1164 int i;
1165 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1166 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1167
1168 if (!zspage)
1169 return NULL;
1170
1171 zspage->magic = ZSPAGE_MAGIC;
1172 migrate_lock_init(zspage);
1173
1174 for (i = 0; i < class->pages_per_zspage; i++) {
1175 struct page *page;
1176
1177 page = alloc_page(gfp);
1178 if (!page) {
1179 while (--i >= 0) {
1180 dec_zone_page_state(pages[i], NR_ZSPAGES);
1181 __free_page(pages[i]);
1182 }
1183 cache_free_zspage(pool, zspage);
1184 return NULL;
1185 }
1186
1187 inc_zone_page_state(page, NR_ZSPAGES);
1188 pages[i] = page;
1189 }
1190
1191 create_page_chain(class, zspage, pages);
1192 init_zspage(class, zspage);
1193 zspage->pool = pool;
1194
1195 return zspage;
1196}
1197
1198static struct zspage *find_get_zspage(struct size_class *class)
1199{
1200 int i;
1201 struct zspage *zspage;
1202
1203 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1204 zspage = list_first_entry_or_null(&class->fullness_list[i],
1205 struct zspage, list);
1206 if (zspage)
1207 break;
1208 }
1209
1210 return zspage;
1211}
1212
1213static inline int __zs_cpu_up(struct mapping_area *area)
1214{
1215 /*
1216 * Make sure we don't leak memory if a cpu UP notification
1217 * and zs_init() race and both call zs_cpu_up() on the same cpu
1218 */
1219 if (area->vm_buf)
1220 return 0;
1221 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1222 if (!area->vm_buf)
1223 return -ENOMEM;
1224 return 0;
1225}
1226
1227static inline void __zs_cpu_down(struct mapping_area *area)
1228{
1229 kfree(area->vm_buf);
1230 area->vm_buf = NULL;
1231}
1232
1233static void *__zs_map_object(struct mapping_area *area,
1234 struct page *pages[2], int off, int size)
1235{
1236 int sizes[2];
1237 void *addr;
1238 char *buf = area->vm_buf;
1239
1240 /* disable page faults to match kmap_atomic() return conditions */
1241 pagefault_disable();
1242
1243 /* no read fastpath */
1244 if (area->vm_mm == ZS_MM_WO)
1245 goto out;
1246
1247 sizes[0] = PAGE_SIZE - off;
1248 sizes[1] = size - sizes[0];
1249
1250 /* copy object to per-cpu buffer */
1251 addr = kmap_atomic(pages[0]);
1252 memcpy(buf, addr + off, sizes[0]);
1253 kunmap_atomic(addr);
1254 addr = kmap_atomic(pages[1]);
1255 memcpy(buf + sizes[0], addr, sizes[1]);
1256 kunmap_atomic(addr);
1257out:
1258 return area->vm_buf;
1259}
1260
1261static void __zs_unmap_object(struct mapping_area *area,
1262 struct page *pages[2], int off, int size)
1263{
1264 int sizes[2];
1265 void *addr;
1266 char *buf;
1267
1268 /* no write fastpath */
1269 if (area->vm_mm == ZS_MM_RO)
1270 goto out;
1271
1272 buf = area->vm_buf;
1273 buf = buf + ZS_HANDLE_SIZE;
1274 size -= ZS_HANDLE_SIZE;
1275 off += ZS_HANDLE_SIZE;
1276
1277 sizes[0] = PAGE_SIZE - off;
1278 sizes[1] = size - sizes[0];
1279
1280 /* copy per-cpu buffer to object */
1281 addr = kmap_atomic(pages[0]);
1282 memcpy(addr + off, buf, sizes[0]);
1283 kunmap_atomic(addr);
1284 addr = kmap_atomic(pages[1]);
1285 memcpy(addr, buf + sizes[0], sizes[1]);
1286 kunmap_atomic(addr);
1287
1288out:
1289 /* enable page faults to match kunmap_atomic() return conditions */
1290 pagefault_enable();
1291}
1292
1293static int zs_cpu_prepare(unsigned int cpu)
1294{
1295 struct mapping_area *area;
1296
1297 area = &per_cpu(zs_map_area, cpu);
1298 return __zs_cpu_up(area);
1299}
1300
1301static int zs_cpu_dead(unsigned int cpu)
1302{
1303 struct mapping_area *area;
1304
1305 area = &per_cpu(zs_map_area, cpu);
1306 __zs_cpu_down(area);
1307 return 0;
1308}
1309
1310static bool can_merge(struct size_class *prev, int pages_per_zspage,
1311 int objs_per_zspage)
1312{
1313 if (prev->pages_per_zspage == pages_per_zspage &&
1314 prev->objs_per_zspage == objs_per_zspage)
1315 return true;
1316
1317 return false;
1318}
1319
1320static bool zspage_full(struct size_class *class, struct zspage *zspage)
1321{
1322 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1323}
1324
1325/**
1326 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1327 * that hold objects of the provided size.
1328 * @pool: zsmalloc pool to use
1329 * @size: object size
1330 *
1331 * Context: Any context.
1332 *
1333 * Return: the index of the zsmalloc &size_class that hold objects of the
1334 * provided size.
1335 */
1336unsigned int zs_lookup_class_index(struct zs_pool *pool, unsigned int size)
1337{
1338 struct size_class *class;
1339
1340 class = pool->size_class[get_size_class_index(size)];
1341
1342 return class->index;
1343}
1344EXPORT_SYMBOL_GPL(zs_lookup_class_index);
1345
1346unsigned long zs_get_total_pages(struct zs_pool *pool)
1347{
1348 return atomic_long_read(&pool->pages_allocated);
1349}
1350EXPORT_SYMBOL_GPL(zs_get_total_pages);
1351
1352/**
1353 * zs_map_object - get address of allocated object from handle.
1354 * @pool: pool from which the object was allocated
1355 * @handle: handle returned from zs_malloc
1356 * @mm: mapping mode to use
1357 *
1358 * Before using an object allocated from zs_malloc, it must be mapped using
1359 * this function. When done with the object, it must be unmapped using
1360 * zs_unmap_object.
1361 *
1362 * Only one object can be mapped per cpu at a time. There is no protection
1363 * against nested mappings.
1364 *
1365 * This function returns with preemption and page faults disabled.
1366 */
1367void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1368 enum zs_mapmode mm)
1369{
1370 struct zspage *zspage;
1371 struct page *page;
1372 unsigned long obj, off;
1373 unsigned int obj_idx;
1374
1375 struct size_class *class;
1376 struct mapping_area *area;
1377 struct page *pages[2];
1378 void *ret;
1379
1380 /*
1381 * Because we use per-cpu mapping areas shared among the
1382 * pools/users, we can't allow mapping in interrupt context
1383 * because it can corrupt another users mappings.
1384 */
1385 BUG_ON(in_interrupt());
1386
1387 /* It guarantees it can get zspage from handle safely */
1388 spin_lock(&pool->lock);
1389 obj = handle_to_obj(handle);
1390 obj_to_location(obj, &page, &obj_idx);
1391 zspage = get_zspage(page);
1392
1393#ifdef CONFIG_ZPOOL
1394 /*
1395 * Move the zspage to front of pool's LRU.
1396 *
1397 * Note that this is swap-specific, so by definition there are no ongoing
1398 * accesses to the memory while the page is swapped out that would make
1399 * it "hot". A new entry is hot, then ages to the tail until it gets either
1400 * written back or swaps back in.
1401 *
1402 * Furthermore, map is also called during writeback. We must not put an
1403 * isolated page on the LRU mid-reclaim.
1404 *
1405 * As a result, only update the LRU when the page is mapped for write
1406 * when it's first instantiated.
1407 *
1408 * This is a deviation from the other backends, which perform this update
1409 * in the allocation function (zbud_alloc, z3fold_alloc).
1410 */
1411 if (mm == ZS_MM_WO) {
1412 if (!list_empty(&zspage->lru))
1413 list_del(&zspage->lru);
1414 list_add(&zspage->lru, &pool->lru);
1415 }
1416#endif
1417
1418 /*
1419 * migration cannot move any zpages in this zspage. Here, pool->lock
1420 * is too heavy since callers would take some time until they calls
1421 * zs_unmap_object API so delegate the locking from class to zspage
1422 * which is smaller granularity.
1423 */
1424 migrate_read_lock(zspage);
1425 spin_unlock(&pool->lock);
1426
1427 class = zspage_class(pool, zspage);
1428 off = (class->size * obj_idx) & ~PAGE_MASK;
1429
1430 local_lock(&zs_map_area.lock);
1431 area = this_cpu_ptr(&zs_map_area);
1432 area->vm_mm = mm;
1433 if (off + class->size <= PAGE_SIZE) {
1434 /* this object is contained entirely within a page */
1435 area->vm_addr = kmap_atomic(page);
1436 ret = area->vm_addr + off;
1437 goto out;
1438 }
1439
1440 /* this object spans two pages */
1441 pages[0] = page;
1442 pages[1] = get_next_page(page);
1443 BUG_ON(!pages[1]);
1444
1445 ret = __zs_map_object(area, pages, off, class->size);
1446out:
1447 if (likely(!ZsHugePage(zspage)))
1448 ret += ZS_HANDLE_SIZE;
1449
1450 return ret;
1451}
1452EXPORT_SYMBOL_GPL(zs_map_object);
1453
1454void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1455{
1456 struct zspage *zspage;
1457 struct page *page;
1458 unsigned long obj, off;
1459 unsigned int obj_idx;
1460
1461 struct size_class *class;
1462 struct mapping_area *area;
1463
1464 obj = handle_to_obj(handle);
1465 obj_to_location(obj, &page, &obj_idx);
1466 zspage = get_zspage(page);
1467 class = zspage_class(pool, zspage);
1468 off = (class->size * obj_idx) & ~PAGE_MASK;
1469
1470 area = this_cpu_ptr(&zs_map_area);
1471 if (off + class->size <= PAGE_SIZE)
1472 kunmap_atomic(area->vm_addr);
1473 else {
1474 struct page *pages[2];
1475
1476 pages[0] = page;
1477 pages[1] = get_next_page(page);
1478 BUG_ON(!pages[1]);
1479
1480 __zs_unmap_object(area, pages, off, class->size);
1481 }
1482 local_unlock(&zs_map_area.lock);
1483
1484 migrate_read_unlock(zspage);
1485}
1486EXPORT_SYMBOL_GPL(zs_unmap_object);
1487
1488/**
1489 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1490 * zsmalloc &size_class.
1491 * @pool: zsmalloc pool to use
1492 *
1493 * The function returns the size of the first huge class - any object of equal
1494 * or bigger size will be stored in zspage consisting of a single physical
1495 * page.
1496 *
1497 * Context: Any context.
1498 *
1499 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1500 */
1501size_t zs_huge_class_size(struct zs_pool *pool)
1502{
1503 return huge_class_size;
1504}
1505EXPORT_SYMBOL_GPL(zs_huge_class_size);
1506
1507static unsigned long obj_malloc(struct zs_pool *pool,
1508 struct zspage *zspage, unsigned long handle)
1509{
1510 int i, nr_page, offset;
1511 unsigned long obj;
1512 struct link_free *link;
1513 struct size_class *class;
1514
1515 struct page *m_page;
1516 unsigned long m_offset;
1517 void *vaddr;
1518
1519 class = pool->size_class[zspage->class];
1520 handle |= OBJ_ALLOCATED_TAG;
1521 obj = get_freeobj(zspage);
1522
1523 offset = obj * class->size;
1524 nr_page = offset >> PAGE_SHIFT;
1525 m_offset = offset & ~PAGE_MASK;
1526 m_page = get_first_page(zspage);
1527
1528 for (i = 0; i < nr_page; i++)
1529 m_page = get_next_page(m_page);
1530
1531 vaddr = kmap_atomic(m_page);
1532 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1533 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1534 if (likely(!ZsHugePage(zspage)))
1535 /* record handle in the header of allocated chunk */
1536 link->handle = handle;
1537 else
1538 /* record handle to page->index */
1539 zspage->first_page->index = handle;
1540
1541 kunmap_atomic(vaddr);
1542 mod_zspage_inuse(zspage, 1);
1543
1544 obj = location_to_obj(m_page, obj);
1545
1546 return obj;
1547}
1548
1549
1550/**
1551 * zs_malloc - Allocate block of given size from pool.
1552 * @pool: pool to allocate from
1553 * @size: size of block to allocate
1554 * @gfp: gfp flags when allocating object
1555 *
1556 * On success, handle to the allocated object is returned,
1557 * otherwise an ERR_PTR().
1558 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1559 */
1560unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1561{
1562 unsigned long handle, obj;
1563 struct size_class *class;
1564 enum fullness_group newfg;
1565 struct zspage *zspage;
1566
1567 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1568 return (unsigned long)ERR_PTR(-EINVAL);
1569
1570 handle = cache_alloc_handle(pool, gfp);
1571 if (!handle)
1572 return (unsigned long)ERR_PTR(-ENOMEM);
1573
1574 /* extra space in chunk to keep the handle */
1575 size += ZS_HANDLE_SIZE;
1576 class = pool->size_class[get_size_class_index(size)];
1577
1578 /* pool->lock effectively protects the zpage migration */
1579 spin_lock(&pool->lock);
1580 zspage = find_get_zspage(class);
1581 if (likely(zspage)) {
1582 obj = obj_malloc(pool, zspage, handle);
1583 /* Now move the zspage to another fullness group, if required */
1584 fix_fullness_group(class, zspage);
1585 record_obj(handle, obj);
1586 class_stat_inc(class, OBJ_USED, 1);
1587 spin_unlock(&pool->lock);
1588
1589 return handle;
1590 }
1591
1592 spin_unlock(&pool->lock);
1593
1594 zspage = alloc_zspage(pool, class, gfp);
1595 if (!zspage) {
1596 cache_free_handle(pool, handle);
1597 return (unsigned long)ERR_PTR(-ENOMEM);
1598 }
1599
1600 spin_lock(&pool->lock);
1601 obj = obj_malloc(pool, zspage, handle);
1602 newfg = get_fullness_group(class, zspage);
1603 insert_zspage(class, zspage, newfg);
1604 set_zspage_mapping(zspage, class->index, newfg);
1605 record_obj(handle, obj);
1606 atomic_long_add(class->pages_per_zspage,
1607 &pool->pages_allocated);
1608 class_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1609 class_stat_inc(class, OBJ_USED, 1);
1610
1611 /* We completely set up zspage so mark them as movable */
1612 SetZsPageMovable(pool, zspage);
1613 spin_unlock(&pool->lock);
1614
1615 return handle;
1616}
1617EXPORT_SYMBOL_GPL(zs_malloc);
1618
1619static void obj_free(int class_size, unsigned long obj, unsigned long *handle)
1620{
1621 struct link_free *link;
1622 struct zspage *zspage;
1623 struct page *f_page;
1624 unsigned long f_offset;
1625 unsigned int f_objidx;
1626 void *vaddr;
1627
1628 obj_to_location(obj, &f_page, &f_objidx);
1629 f_offset = (class_size * f_objidx) & ~PAGE_MASK;
1630 zspage = get_zspage(f_page);
1631
1632 vaddr = kmap_atomic(f_page);
1633 link = (struct link_free *)(vaddr + f_offset);
1634
1635 if (handle) {
1636#ifdef CONFIG_ZPOOL
1637 /* Stores the (deferred) handle in the object's header */
1638 *handle |= OBJ_DEFERRED_HANDLE_TAG;
1639 *handle &= ~OBJ_ALLOCATED_TAG;
1640
1641 if (likely(!ZsHugePage(zspage)))
1642 link->deferred_handle = *handle;
1643 else
1644 f_page->index = *handle;
1645#endif
1646 } else {
1647 /* Insert this object in containing zspage's freelist */
1648 if (likely(!ZsHugePage(zspage)))
1649 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1650 else
1651 f_page->index = 0;
1652 set_freeobj(zspage, f_objidx);
1653 }
1654
1655 kunmap_atomic(vaddr);
1656 mod_zspage_inuse(zspage, -1);
1657}
1658
1659void zs_free(struct zs_pool *pool, unsigned long handle)
1660{
1661 struct zspage *zspage;
1662 struct page *f_page;
1663 unsigned long obj;
1664 struct size_class *class;
1665 enum fullness_group fullness;
1666
1667 if (IS_ERR_OR_NULL((void *)handle))
1668 return;
1669
1670 /*
1671 * The pool->lock protects the race with zpage's migration
1672 * so it's safe to get the page from handle.
1673 */
1674 spin_lock(&pool->lock);
1675 obj = handle_to_obj(handle);
1676 obj_to_page(obj, &f_page);
1677 zspage = get_zspage(f_page);
1678 class = zspage_class(pool, zspage);
1679
1680 class_stat_dec(class, OBJ_USED, 1);
1681
1682#ifdef CONFIG_ZPOOL
1683 if (zspage->under_reclaim) {
1684 /*
1685 * Reclaim needs the handles during writeback. It'll free
1686 * them along with the zspage when it's done with them.
1687 *
1688 * Record current deferred handle in the object's header.
1689 */
1690 obj_free(class->size, obj, &handle);
1691 spin_unlock(&pool->lock);
1692 return;
1693 }
1694#endif
1695 obj_free(class->size, obj, NULL);
1696
1697 fullness = fix_fullness_group(class, zspage);
1698 if (fullness == ZS_EMPTY)
1699 free_zspage(pool, class, zspage);
1700
1701 spin_unlock(&pool->lock);
1702 cache_free_handle(pool, handle);
1703}
1704EXPORT_SYMBOL_GPL(zs_free);
1705
1706static void zs_object_copy(struct size_class *class, unsigned long dst,
1707 unsigned long src)
1708{
1709 struct page *s_page, *d_page;
1710 unsigned int s_objidx, d_objidx;
1711 unsigned long s_off, d_off;
1712 void *s_addr, *d_addr;
1713 int s_size, d_size, size;
1714 int written = 0;
1715
1716 s_size = d_size = class->size;
1717
1718 obj_to_location(src, &s_page, &s_objidx);
1719 obj_to_location(dst, &d_page, &d_objidx);
1720
1721 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1722 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1723
1724 if (s_off + class->size > PAGE_SIZE)
1725 s_size = PAGE_SIZE - s_off;
1726
1727 if (d_off + class->size > PAGE_SIZE)
1728 d_size = PAGE_SIZE - d_off;
1729
1730 s_addr = kmap_atomic(s_page);
1731 d_addr = kmap_atomic(d_page);
1732
1733 while (1) {
1734 size = min(s_size, d_size);
1735 memcpy(d_addr + d_off, s_addr + s_off, size);
1736 written += size;
1737
1738 if (written == class->size)
1739 break;
1740
1741 s_off += size;
1742 s_size -= size;
1743 d_off += size;
1744 d_size -= size;
1745
1746 /*
1747 * Calling kunmap_atomic(d_addr) is necessary. kunmap_atomic()
1748 * calls must occurs in reverse order of calls to kmap_atomic().
1749 * So, to call kunmap_atomic(s_addr) we should first call
1750 * kunmap_atomic(d_addr). For more details see
1751 * Documentation/mm/highmem.rst.
1752 */
1753 if (s_off >= PAGE_SIZE) {
1754 kunmap_atomic(d_addr);
1755 kunmap_atomic(s_addr);
1756 s_page = get_next_page(s_page);
1757 s_addr = kmap_atomic(s_page);
1758 d_addr = kmap_atomic(d_page);
1759 s_size = class->size - written;
1760 s_off = 0;
1761 }
1762
1763 if (d_off >= PAGE_SIZE) {
1764 kunmap_atomic(d_addr);
1765 d_page = get_next_page(d_page);
1766 d_addr = kmap_atomic(d_page);
1767 d_size = class->size - written;
1768 d_off = 0;
1769 }
1770 }
1771
1772 kunmap_atomic(d_addr);
1773 kunmap_atomic(s_addr);
1774}
1775
1776/*
1777 * Find object with a certain tag in zspage from index object and
1778 * return handle.
1779 */
1780static unsigned long find_tagged_obj(struct size_class *class,
1781 struct page *page, int *obj_idx, int tag)
1782{
1783 unsigned int offset;
1784 int index = *obj_idx;
1785 unsigned long handle = 0;
1786 void *addr = kmap_atomic(page);
1787
1788 offset = get_first_obj_offset(page);
1789 offset += class->size * index;
1790
1791 while (offset < PAGE_SIZE) {
1792 if (obj_tagged(page, addr + offset, &handle, tag))
1793 break;
1794
1795 offset += class->size;
1796 index++;
1797 }
1798
1799 kunmap_atomic(addr);
1800
1801 *obj_idx = index;
1802
1803 return handle;
1804}
1805
1806/*
1807 * Find alloced object in zspage from index object and
1808 * return handle.
1809 */
1810static unsigned long find_alloced_obj(struct size_class *class,
1811 struct page *page, int *obj_idx)
1812{
1813 return find_tagged_obj(class, page, obj_idx, OBJ_ALLOCATED_TAG);
1814}
1815
1816#ifdef CONFIG_ZPOOL
1817/*
1818 * Find object storing a deferred handle in header in zspage from index object
1819 * and return handle.
1820 */
1821static unsigned long find_deferred_handle_obj(struct size_class *class,
1822 struct page *page, int *obj_idx)
1823{
1824 return find_tagged_obj(class, page, obj_idx, OBJ_DEFERRED_HANDLE_TAG);
1825}
1826#endif
1827
1828struct zs_compact_control {
1829 /* Source spage for migration which could be a subpage of zspage */
1830 struct page *s_page;
1831 /* Destination page for migration which should be a first page
1832 * of zspage. */
1833 struct page *d_page;
1834 /* Starting object index within @s_page which used for live object
1835 * in the subpage. */
1836 int obj_idx;
1837};
1838
1839static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1840 struct zs_compact_control *cc)
1841{
1842 unsigned long used_obj, free_obj;
1843 unsigned long handle;
1844 struct page *s_page = cc->s_page;
1845 struct page *d_page = cc->d_page;
1846 int obj_idx = cc->obj_idx;
1847 int ret = 0;
1848
1849 while (1) {
1850 handle = find_alloced_obj(class, s_page, &obj_idx);
1851 if (!handle) {
1852 s_page = get_next_page(s_page);
1853 if (!s_page)
1854 break;
1855 obj_idx = 0;
1856 continue;
1857 }
1858
1859 /* Stop if there is no more space */
1860 if (zspage_full(class, get_zspage(d_page))) {
1861 ret = -ENOMEM;
1862 break;
1863 }
1864
1865 used_obj = handle_to_obj(handle);
1866 free_obj = obj_malloc(pool, get_zspage(d_page), handle);
1867 zs_object_copy(class, free_obj, used_obj);
1868 obj_idx++;
1869 record_obj(handle, free_obj);
1870 obj_free(class->size, used_obj, NULL);
1871 }
1872
1873 /* Remember last position in this iteration */
1874 cc->s_page = s_page;
1875 cc->obj_idx = obj_idx;
1876
1877 return ret;
1878}
1879
1880static struct zspage *isolate_zspage(struct size_class *class, bool source)
1881{
1882 int i;
1883 struct zspage *zspage;
1884 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1885
1886 if (!source) {
1887 fg[0] = ZS_ALMOST_FULL;
1888 fg[1] = ZS_ALMOST_EMPTY;
1889 }
1890
1891 for (i = 0; i < 2; i++) {
1892 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1893 struct zspage, list);
1894 if (zspage) {
1895 remove_zspage(class, zspage, fg[i]);
1896 return zspage;
1897 }
1898 }
1899
1900 return zspage;
1901}
1902
1903/*
1904 * putback_zspage - add @zspage into right class's fullness list
1905 * @class: destination class
1906 * @zspage: target page
1907 *
1908 * Return @zspage's fullness_group
1909 */
1910static enum fullness_group putback_zspage(struct size_class *class,
1911 struct zspage *zspage)
1912{
1913 enum fullness_group fullness;
1914
1915 fullness = get_fullness_group(class, zspage);
1916 insert_zspage(class, zspage, fullness);
1917 set_zspage_mapping(zspage, class->index, fullness);
1918
1919 return fullness;
1920}
1921
1922#if defined(CONFIG_ZPOOL) || defined(CONFIG_COMPACTION)
1923/*
1924 * To prevent zspage destroy during migration, zspage freeing should
1925 * hold locks of all pages in the zspage.
1926 */
1927static void lock_zspage(struct zspage *zspage)
1928{
1929 struct page *curr_page, *page;
1930
1931 /*
1932 * Pages we haven't locked yet can be migrated off the list while we're
1933 * trying to lock them, so we need to be careful and only attempt to
1934 * lock each page under migrate_read_lock(). Otherwise, the page we lock
1935 * may no longer belong to the zspage. This means that we may wait for
1936 * the wrong page to unlock, so we must take a reference to the page
1937 * prior to waiting for it to unlock outside migrate_read_lock().
1938 */
1939 while (1) {
1940 migrate_read_lock(zspage);
1941 page = get_first_page(zspage);
1942 if (trylock_page(page))
1943 break;
1944 get_page(page);
1945 migrate_read_unlock(zspage);
1946 wait_on_page_locked(page);
1947 put_page(page);
1948 }
1949
1950 curr_page = page;
1951 while ((page = get_next_page(curr_page))) {
1952 if (trylock_page(page)) {
1953 curr_page = page;
1954 } else {
1955 get_page(page);
1956 migrate_read_unlock(zspage);
1957 wait_on_page_locked(page);
1958 put_page(page);
1959 migrate_read_lock(zspage);
1960 }
1961 }
1962 migrate_read_unlock(zspage);
1963}
1964#endif /* defined(CONFIG_ZPOOL) || defined(CONFIG_COMPACTION) */
1965
1966#ifdef CONFIG_ZPOOL
1967/*
1968 * Unlocks all the pages of the zspage.
1969 *
1970 * pool->lock must be held before this function is called
1971 * to prevent the underlying pages from migrating.
1972 */
1973static void unlock_zspage(struct zspage *zspage)
1974{
1975 struct page *page = get_first_page(zspage);
1976
1977 do {
1978 unlock_page(page);
1979 } while ((page = get_next_page(page)) != NULL);
1980}
1981#endif /* CONFIG_ZPOOL */
1982
1983static void migrate_lock_init(struct zspage *zspage)
1984{
1985 rwlock_init(&zspage->lock);
1986}
1987
1988static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1989{
1990 read_lock(&zspage->lock);
1991}
1992
1993static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1994{
1995 read_unlock(&zspage->lock);
1996}
1997
1998#ifdef CONFIG_COMPACTION
1999static void migrate_write_lock(struct zspage *zspage)
2000{
2001 write_lock(&zspage->lock);
2002}
2003
2004static void migrate_write_lock_nested(struct zspage *zspage)
2005{
2006 write_lock_nested(&zspage->lock, SINGLE_DEPTH_NESTING);
2007}
2008
2009static void migrate_write_unlock(struct zspage *zspage)
2010{
2011 write_unlock(&zspage->lock);
2012}
2013
2014/* Number of isolated subpage for *page migration* in this zspage */
2015static void inc_zspage_isolation(struct zspage *zspage)
2016{
2017 zspage->isolated++;
2018}
2019
2020static void dec_zspage_isolation(struct zspage *zspage)
2021{
2022 VM_BUG_ON(zspage->isolated == 0);
2023 zspage->isolated--;
2024}
2025
2026static const struct movable_operations zsmalloc_mops;
2027
2028static void replace_sub_page(struct size_class *class, struct zspage *zspage,
2029 struct page *newpage, struct page *oldpage)
2030{
2031 struct page *page;
2032 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
2033 int idx = 0;
2034
2035 page = get_first_page(zspage);
2036 do {
2037 if (page == oldpage)
2038 pages[idx] = newpage;
2039 else
2040 pages[idx] = page;
2041 idx++;
2042 } while ((page = get_next_page(page)) != NULL);
2043
2044 create_page_chain(class, zspage, pages);
2045 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
2046 if (unlikely(ZsHugePage(zspage)))
2047 newpage->index = oldpage->index;
2048 __SetPageMovable(newpage, &zsmalloc_mops);
2049}
2050
2051static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
2052{
2053 struct zspage *zspage;
2054
2055 /*
2056 * Page is locked so zspage couldn't be destroyed. For detail, look at
2057 * lock_zspage in free_zspage.
2058 */
2059 VM_BUG_ON_PAGE(!PageMovable(page), page);
2060 VM_BUG_ON_PAGE(PageIsolated(page), page);
2061
2062 zspage = get_zspage(page);
2063 migrate_write_lock(zspage);
2064 inc_zspage_isolation(zspage);
2065 migrate_write_unlock(zspage);
2066
2067 return true;
2068}
2069
2070static int zs_page_migrate(struct page *newpage, struct page *page,
2071 enum migrate_mode mode)
2072{
2073 struct zs_pool *pool;
2074 struct size_class *class;
2075 struct zspage *zspage;
2076 struct page *dummy;
2077 void *s_addr, *d_addr, *addr;
2078 unsigned int offset;
2079 unsigned long handle;
2080 unsigned long old_obj, new_obj;
2081 unsigned int obj_idx;
2082
2083 /*
2084 * We cannot support the _NO_COPY case here, because copy needs to
2085 * happen under the zs lock, which does not work with
2086 * MIGRATE_SYNC_NO_COPY workflow.
2087 */
2088 if (mode == MIGRATE_SYNC_NO_COPY)
2089 return -EINVAL;
2090
2091 VM_BUG_ON_PAGE(!PageMovable(page), page);
2092 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2093
2094 /* The page is locked, so this pointer must remain valid */
2095 zspage = get_zspage(page);
2096 pool = zspage->pool;
2097
2098 /*
2099 * The pool's lock protects the race between zpage migration
2100 * and zs_free.
2101 */
2102 spin_lock(&pool->lock);
2103 class = zspage_class(pool, zspage);
2104
2105 /* the migrate_write_lock protects zpage access via zs_map_object */
2106 migrate_write_lock(zspage);
2107
2108 offset = get_first_obj_offset(page);
2109 s_addr = kmap_atomic(page);
2110
2111 /*
2112 * Here, any user cannot access all objects in the zspage so let's move.
2113 */
2114 d_addr = kmap_atomic(newpage);
2115 memcpy(d_addr, s_addr, PAGE_SIZE);
2116 kunmap_atomic(d_addr);
2117
2118 for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE;
2119 addr += class->size) {
2120 if (obj_allocated(page, addr, &handle)) {
2121
2122 old_obj = handle_to_obj(handle);
2123 obj_to_location(old_obj, &dummy, &obj_idx);
2124 new_obj = (unsigned long)location_to_obj(newpage,
2125 obj_idx);
2126 record_obj(handle, new_obj);
2127 }
2128 }
2129 kunmap_atomic(s_addr);
2130
2131 replace_sub_page(class, zspage, newpage, page);
2132 /*
2133 * Since we complete the data copy and set up new zspage structure,
2134 * it's okay to release the pool's lock.
2135 */
2136 spin_unlock(&pool->lock);
2137 dec_zspage_isolation(zspage);
2138 migrate_write_unlock(zspage);
2139
2140 get_page(newpage);
2141 if (page_zone(newpage) != page_zone(page)) {
2142 dec_zone_page_state(page, NR_ZSPAGES);
2143 inc_zone_page_state(newpage, NR_ZSPAGES);
2144 }
2145
2146 reset_page(page);
2147 put_page(page);
2148
2149 return MIGRATEPAGE_SUCCESS;
2150}
2151
2152static void zs_page_putback(struct page *page)
2153{
2154 struct zspage *zspage;
2155
2156 VM_BUG_ON_PAGE(!PageMovable(page), page);
2157 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2158
2159 zspage = get_zspage(page);
2160 migrate_write_lock(zspage);
2161 dec_zspage_isolation(zspage);
2162 migrate_write_unlock(zspage);
2163}
2164
2165static const struct movable_operations zsmalloc_mops = {
2166 .isolate_page = zs_page_isolate,
2167 .migrate_page = zs_page_migrate,
2168 .putback_page = zs_page_putback,
2169};
2170
2171/*
2172 * Caller should hold page_lock of all pages in the zspage
2173 * In here, we cannot use zspage meta data.
2174 */
2175static void async_free_zspage(struct work_struct *work)
2176{
2177 int i;
2178 struct size_class *class;
2179 unsigned int class_idx;
2180 enum fullness_group fullness;
2181 struct zspage *zspage, *tmp;
2182 LIST_HEAD(free_pages);
2183 struct zs_pool *pool = container_of(work, struct zs_pool,
2184 free_work);
2185
2186 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2187 class = pool->size_class[i];
2188 if (class->index != i)
2189 continue;
2190
2191 spin_lock(&pool->lock);
2192 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2193 spin_unlock(&pool->lock);
2194 }
2195
2196 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2197 list_del(&zspage->list);
2198 lock_zspage(zspage);
2199
2200 get_zspage_mapping(zspage, &class_idx, &fullness);
2201 VM_BUG_ON(fullness != ZS_EMPTY);
2202 class = pool->size_class[class_idx];
2203 spin_lock(&pool->lock);
2204#ifdef CONFIG_ZPOOL
2205 list_del(&zspage->lru);
2206#endif
2207 __free_zspage(pool, class, zspage);
2208 spin_unlock(&pool->lock);
2209 }
2210};
2211
2212static void kick_deferred_free(struct zs_pool *pool)
2213{
2214 schedule_work(&pool->free_work);
2215}
2216
2217static void zs_flush_migration(struct zs_pool *pool)
2218{
2219 flush_work(&pool->free_work);
2220}
2221
2222static void init_deferred_free(struct zs_pool *pool)
2223{
2224 INIT_WORK(&pool->free_work, async_free_zspage);
2225}
2226
2227static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2228{
2229 struct page *page = get_first_page(zspage);
2230
2231 do {
2232 WARN_ON(!trylock_page(page));
2233 __SetPageMovable(page, &zsmalloc_mops);
2234 unlock_page(page);
2235 } while ((page = get_next_page(page)) != NULL);
2236}
2237#else
2238static inline void zs_flush_migration(struct zs_pool *pool) { }
2239#endif
2240
2241/*
2242 *
2243 * Based on the number of unused allocated objects calculate
2244 * and return the number of pages that we can free.
2245 */
2246static unsigned long zs_can_compact(struct size_class *class)
2247{
2248 unsigned long obj_wasted;
2249 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2250 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2251
2252 if (obj_allocated <= obj_used)
2253 return 0;
2254
2255 obj_wasted = obj_allocated - obj_used;
2256 obj_wasted /= class->objs_per_zspage;
2257
2258 return obj_wasted * class->pages_per_zspage;
2259}
2260
2261static unsigned long __zs_compact(struct zs_pool *pool,
2262 struct size_class *class)
2263{
2264 struct zs_compact_control cc;
2265 struct zspage *src_zspage;
2266 struct zspage *dst_zspage = NULL;
2267 unsigned long pages_freed = 0;
2268
2269 /*
2270 * protect the race between zpage migration and zs_free
2271 * as well as zpage allocation/free
2272 */
2273 spin_lock(&pool->lock);
2274 while ((src_zspage = isolate_zspage(class, true))) {
2275 /* protect someone accessing the zspage(i.e., zs_map_object) */
2276 migrate_write_lock(src_zspage);
2277
2278 if (!zs_can_compact(class))
2279 break;
2280
2281 cc.obj_idx = 0;
2282 cc.s_page = get_first_page(src_zspage);
2283
2284 while ((dst_zspage = isolate_zspage(class, false))) {
2285 migrate_write_lock_nested(dst_zspage);
2286
2287 cc.d_page = get_first_page(dst_zspage);
2288 /*
2289 * If there is no more space in dst_page, resched
2290 * and see if anyone had allocated another zspage.
2291 */
2292 if (!migrate_zspage(pool, class, &cc))
2293 break;
2294
2295 putback_zspage(class, dst_zspage);
2296 migrate_write_unlock(dst_zspage);
2297 dst_zspage = NULL;
2298 if (spin_is_contended(&pool->lock))
2299 break;
2300 }
2301
2302 /* Stop if we couldn't find slot */
2303 if (dst_zspage == NULL)
2304 break;
2305
2306 putback_zspage(class, dst_zspage);
2307 migrate_write_unlock(dst_zspage);
2308
2309 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2310 migrate_write_unlock(src_zspage);
2311 free_zspage(pool, class, src_zspage);
2312 pages_freed += class->pages_per_zspage;
2313 } else
2314 migrate_write_unlock(src_zspage);
2315 spin_unlock(&pool->lock);
2316 cond_resched();
2317 spin_lock(&pool->lock);
2318 }
2319
2320 if (src_zspage) {
2321 putback_zspage(class, src_zspage);
2322 migrate_write_unlock(src_zspage);
2323 }
2324
2325 spin_unlock(&pool->lock);
2326
2327 return pages_freed;
2328}
2329
2330unsigned long zs_compact(struct zs_pool *pool)
2331{
2332 int i;
2333 struct size_class *class;
2334 unsigned long pages_freed = 0;
2335
2336 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2337 class = pool->size_class[i];
2338 if (class->index != i)
2339 continue;
2340 pages_freed += __zs_compact(pool, class);
2341 }
2342 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2343
2344 return pages_freed;
2345}
2346EXPORT_SYMBOL_GPL(zs_compact);
2347
2348void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2349{
2350 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2351}
2352EXPORT_SYMBOL_GPL(zs_pool_stats);
2353
2354static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2355 struct shrink_control *sc)
2356{
2357 unsigned long pages_freed;
2358 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2359 shrinker);
2360
2361 /*
2362 * Compact classes and calculate compaction delta.
2363 * Can run concurrently with a manually triggered
2364 * (by user) compaction.
2365 */
2366 pages_freed = zs_compact(pool);
2367
2368 return pages_freed ? pages_freed : SHRINK_STOP;
2369}
2370
2371static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2372 struct shrink_control *sc)
2373{
2374 int i;
2375 struct size_class *class;
2376 unsigned long pages_to_free = 0;
2377 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2378 shrinker);
2379
2380 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2381 class = pool->size_class[i];
2382 if (class->index != i)
2383 continue;
2384
2385 pages_to_free += zs_can_compact(class);
2386 }
2387
2388 return pages_to_free;
2389}
2390
2391static void zs_unregister_shrinker(struct zs_pool *pool)
2392{
2393 unregister_shrinker(&pool->shrinker);
2394}
2395
2396static int zs_register_shrinker(struct zs_pool *pool)
2397{
2398 pool->shrinker.scan_objects = zs_shrinker_scan;
2399 pool->shrinker.count_objects = zs_shrinker_count;
2400 pool->shrinker.batch = 0;
2401 pool->shrinker.seeks = DEFAULT_SEEKS;
2402
2403 return register_shrinker(&pool->shrinker, "mm-zspool:%s",
2404 pool->name);
2405}
2406
2407/**
2408 * zs_create_pool - Creates an allocation pool to work from.
2409 * @name: pool name to be created
2410 *
2411 * This function must be called before anything when using
2412 * the zsmalloc allocator.
2413 *
2414 * On success, a pointer to the newly created pool is returned,
2415 * otherwise NULL.
2416 */
2417struct zs_pool *zs_create_pool(const char *name)
2418{
2419 int i;
2420 struct zs_pool *pool;
2421 struct size_class *prev_class = NULL;
2422
2423 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2424 if (!pool)
2425 return NULL;
2426
2427 init_deferred_free(pool);
2428 spin_lock_init(&pool->lock);
2429
2430 pool->name = kstrdup(name, GFP_KERNEL);
2431 if (!pool->name)
2432 goto err;
2433
2434 if (create_cache(pool))
2435 goto err;
2436
2437 /*
2438 * Iterate reversely, because, size of size_class that we want to use
2439 * for merging should be larger or equal to current size.
2440 */
2441 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2442 int size;
2443 int pages_per_zspage;
2444 int objs_per_zspage;
2445 struct size_class *class;
2446 int fullness = 0;
2447
2448 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2449 if (size > ZS_MAX_ALLOC_SIZE)
2450 size = ZS_MAX_ALLOC_SIZE;
2451 pages_per_zspage = get_pages_per_zspage(size);
2452 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2453
2454 /*
2455 * We iterate from biggest down to smallest classes,
2456 * so huge_class_size holds the size of the first huge
2457 * class. Any object bigger than or equal to that will
2458 * endup in the huge class.
2459 */
2460 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2461 !huge_class_size) {
2462 huge_class_size = size;
2463 /*
2464 * The object uses ZS_HANDLE_SIZE bytes to store the
2465 * handle. We need to subtract it, because zs_malloc()
2466 * unconditionally adds handle size before it performs
2467 * size class search - so object may be smaller than
2468 * huge class size, yet it still can end up in the huge
2469 * class because it grows by ZS_HANDLE_SIZE extra bytes
2470 * right before class lookup.
2471 */
2472 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2473 }
2474
2475 /*
2476 * size_class is used for normal zsmalloc operation such
2477 * as alloc/free for that size. Although it is natural that we
2478 * have one size_class for each size, there is a chance that we
2479 * can get more memory utilization if we use one size_class for
2480 * many different sizes whose size_class have same
2481 * characteristics. So, we makes size_class point to
2482 * previous size_class if possible.
2483 */
2484 if (prev_class) {
2485 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2486 pool->size_class[i] = prev_class;
2487 continue;
2488 }
2489 }
2490
2491 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2492 if (!class)
2493 goto err;
2494
2495 class->size = size;
2496 class->index = i;
2497 class->pages_per_zspage = pages_per_zspage;
2498 class->objs_per_zspage = objs_per_zspage;
2499 pool->size_class[i] = class;
2500 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2501 fullness++)
2502 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2503
2504 prev_class = class;
2505 }
2506
2507 /* debug only, don't abort if it fails */
2508 zs_pool_stat_create(pool, name);
2509
2510 /*
2511 * Not critical since shrinker is only used to trigger internal
2512 * defragmentation of the pool which is pretty optional thing. If
2513 * registration fails we still can use the pool normally and user can
2514 * trigger compaction manually. Thus, ignore return code.
2515 */
2516 zs_register_shrinker(pool);
2517
2518#ifdef CONFIG_ZPOOL
2519 INIT_LIST_HEAD(&pool->lru);
2520#endif
2521
2522 return pool;
2523
2524err:
2525 zs_destroy_pool(pool);
2526 return NULL;
2527}
2528EXPORT_SYMBOL_GPL(zs_create_pool);
2529
2530void zs_destroy_pool(struct zs_pool *pool)
2531{
2532 int i;
2533
2534 zs_unregister_shrinker(pool);
2535 zs_flush_migration(pool);
2536 zs_pool_stat_destroy(pool);
2537
2538 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2539 int fg;
2540 struct size_class *class = pool->size_class[i];
2541
2542 if (!class)
2543 continue;
2544
2545 if (class->index != i)
2546 continue;
2547
2548 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2549 if (!list_empty(&class->fullness_list[fg])) {
2550 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2551 class->size, fg);
2552 }
2553 }
2554 kfree(class);
2555 }
2556
2557 destroy_cache(pool);
2558 kfree(pool->name);
2559 kfree(pool);
2560}
2561EXPORT_SYMBOL_GPL(zs_destroy_pool);
2562
2563#ifdef CONFIG_ZPOOL
2564static void restore_freelist(struct zs_pool *pool, struct size_class *class,
2565 struct zspage *zspage)
2566{
2567 unsigned int obj_idx = 0;
2568 unsigned long handle, off = 0; /* off is within-page offset */
2569 struct page *page = get_first_page(zspage);
2570 struct link_free *prev_free = NULL;
2571 void *prev_page_vaddr = NULL;
2572
2573 /* in case no free object found */
2574 set_freeobj(zspage, (unsigned int)(-1UL));
2575
2576 while (page) {
2577 void *vaddr = kmap_atomic(page);
2578 struct page *next_page;
2579
2580 while (off < PAGE_SIZE) {
2581 void *obj_addr = vaddr + off;
2582
2583 /* skip allocated object */
2584 if (obj_allocated(page, obj_addr, &handle)) {
2585 obj_idx++;
2586 off += class->size;
2587 continue;
2588 }
2589
2590 /* free deferred handle from reclaim attempt */
2591 if (obj_stores_deferred_handle(page, obj_addr, &handle))
2592 cache_free_handle(pool, handle);
2593
2594 if (prev_free)
2595 prev_free->next = obj_idx << OBJ_TAG_BITS;
2596 else /* first free object found */
2597 set_freeobj(zspage, obj_idx);
2598
2599 prev_free = (struct link_free *)vaddr + off / sizeof(*prev_free);
2600 /* if last free object in a previous page, need to unmap */
2601 if (prev_page_vaddr) {
2602 kunmap_atomic(prev_page_vaddr);
2603 prev_page_vaddr = NULL;
2604 }
2605
2606 obj_idx++;
2607 off += class->size;
2608 }
2609
2610 /*
2611 * Handle the last (full or partial) object on this page.
2612 */
2613 next_page = get_next_page(page);
2614 if (next_page) {
2615 if (!prev_free || prev_page_vaddr) {
2616 /*
2617 * There is no free object in this page, so we can safely
2618 * unmap it.
2619 */
2620 kunmap_atomic(vaddr);
2621 } else {
2622 /* update prev_page_vaddr since prev_free is on this page */
2623 prev_page_vaddr = vaddr;
2624 }
2625 } else { /* this is the last page */
2626 if (prev_free) {
2627 /*
2628 * Reset OBJ_TAG_BITS bit to last link to tell
2629 * whether it's allocated object or not.
2630 */
2631 prev_free->next = -1UL << OBJ_TAG_BITS;
2632 }
2633
2634 /* unmap previous page (if not done yet) */
2635 if (prev_page_vaddr) {
2636 kunmap_atomic(prev_page_vaddr);
2637 prev_page_vaddr = NULL;
2638 }
2639
2640 kunmap_atomic(vaddr);
2641 }
2642
2643 page = next_page;
2644 off %= PAGE_SIZE;
2645 }
2646}
2647
2648static int zs_reclaim_page(struct zs_pool *pool, unsigned int retries)
2649{
2650 int i, obj_idx, ret = 0;
2651 unsigned long handle;
2652 struct zspage *zspage;
2653 struct page *page;
2654 enum fullness_group fullness;
2655
2656 /* Lock LRU and fullness list */
2657 spin_lock(&pool->lock);
2658 if (list_empty(&pool->lru)) {
2659 spin_unlock(&pool->lock);
2660 return -EINVAL;
2661 }
2662
2663 for (i = 0; i < retries; i++) {
2664 struct size_class *class;
2665
2666 zspage = list_last_entry(&pool->lru, struct zspage, lru);
2667 list_del(&zspage->lru);
2668
2669 /* zs_free may free objects, but not the zspage and handles */
2670 zspage->under_reclaim = true;
2671
2672 class = zspage_class(pool, zspage);
2673 fullness = get_fullness_group(class, zspage);
2674
2675 /* Lock out object allocations and object compaction */
2676 remove_zspage(class, zspage, fullness);
2677
2678 spin_unlock(&pool->lock);
2679 cond_resched();
2680
2681 /* Lock backing pages into place */
2682 lock_zspage(zspage);
2683
2684 obj_idx = 0;
2685 page = get_first_page(zspage);
2686 while (1) {
2687 handle = find_alloced_obj(class, page, &obj_idx);
2688 if (!handle) {
2689 page = get_next_page(page);
2690 if (!page)
2691 break;
2692 obj_idx = 0;
2693 continue;
2694 }
2695
2696 /*
2697 * This will write the object and call zs_free.
2698 *
2699 * zs_free will free the object, but the
2700 * under_reclaim flag prevents it from freeing
2701 * the zspage altogether. This is necessary so
2702 * that we can continue working with the
2703 * zspage potentially after the last object
2704 * has been freed.
2705 */
2706 ret = pool->zpool_ops->evict(pool->zpool, handle);
2707 if (ret)
2708 goto next;
2709
2710 obj_idx++;
2711 }
2712
2713next:
2714 /* For freeing the zspage, or putting it back in the pool and LRU list. */
2715 spin_lock(&pool->lock);
2716 zspage->under_reclaim = false;
2717
2718 if (!get_zspage_inuse(zspage)) {
2719 /*
2720 * Fullness went stale as zs_free() won't touch it
2721 * while the page is removed from the pool. Fix it
2722 * up for the check in __free_zspage().
2723 */
2724 zspage->fullness = ZS_EMPTY;
2725
2726 __free_zspage(pool, class, zspage);
2727 spin_unlock(&pool->lock);
2728 return 0;
2729 }
2730
2731 /*
2732 * Eviction fails on one of the handles, so we need to restore zspage.
2733 * We need to rebuild its freelist (and free stored deferred handles),
2734 * put it back to the correct size class, and add it to the LRU list.
2735 */
2736 restore_freelist(pool, class, zspage);
2737 putback_zspage(class, zspage);
2738 list_add(&zspage->lru, &pool->lru);
2739 unlock_zspage(zspage);
2740 }
2741
2742 spin_unlock(&pool->lock);
2743 return -EAGAIN;
2744}
2745#endif /* CONFIG_ZPOOL */
2746
2747static int __init zs_init(void)
2748{
2749 int ret;
2750
2751 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2752 zs_cpu_prepare, zs_cpu_dead);
2753 if (ret)
2754 goto out;
2755
2756#ifdef CONFIG_ZPOOL
2757 zpool_register_driver(&zs_zpool_driver);
2758#endif
2759
2760 zs_stat_init();
2761
2762 return 0;
2763
2764out:
2765 return ret;
2766}
2767
2768static void __exit zs_exit(void)
2769{
2770#ifdef CONFIG_ZPOOL
2771 zpool_unregister_driver(&zs_zpool_driver);
2772#endif
2773 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2774
2775 zs_stat_exit();
2776}
2777
2778module_init(zs_init);
2779module_exit(zs_exit);
2780
2781MODULE_LICENSE("Dual BSD/GPL");
2782MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");