projects/linux·blob master · refs

2428 lines · 61.1 KiB

1// SPDX-License-Identifier: GPL-2.0-or-later
2
3/*
4 * zsmalloc memory allocator
5 *
6 * Copyright (C) 2011 Nitin Gupta
7 * Copyright (C) 2012, 2013 Minchan Kim
8 *
9 * This code is released using a dual license strategy: BSD/GPL
10 * You can choose the license that better fits your requirements.
11 *
12 * Released under the terms of 3-clause BSD License
13 * Released under the terms of GNU General Public License Version 2.0
14 */
15
16#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17
18/*
19 * lock ordering:
20 * page_lock
21 * pool->lock
22 * class->lock
23 * zspage->lock
24 *
25 * When ZS_OBJ_CLASS_BITS > 0, zs_free() skips pool->lock; it picks
26 * the size_class from obj's encoded class_idx and serializes against
27 * page migration via class->lock.
28 */
29
30#include <linux/module.h>
31#include <linux/kernel.h>
32#include <linux/sched.h>
33#include <linux/errno.h>
34#include <linux/highmem.h>
35#include <linux/string.h>
36#include <linux/slab.h>
37#include <linux/scatterlist.h>
38#include <linux/spinlock.h>
39#include <linux/sprintf.h>
40#include <linux/shrinker.h>
41#include <linux/types.h>
42#include <linux/debugfs.h>
43#include <linux/zsmalloc.h>
44#include <linux/fs.h>
45#include <linux/workqueue.h>
46#include "zpdesc.h"
47
48#define ZSPAGE_MAGIC 0x58
49
50/*
51 * This must be power of 2 and greater than or equal to sizeof(link_free).
52 * These two conditions ensure that any 'struct link_free' itself doesn't
53 * span more than 1 page which avoids complex case of mapping 2 pages simply
54 * to restore link_free pointer values.
55 */
56#define ZS_ALIGN 8
57
58#define ZS_HANDLE_SIZE (sizeof(unsigned long))
59
60/*
61 * Object location (<PFN>, <obj_idx>) is encoded as
62 * a single (unsigned long) handle value.
63 *
64 * Note that object index <obj_idx> starts from 0.
65 *
66 * This is made more complicated by various memory models and PAE.
67 */
68
69#ifndef MAX_POSSIBLE_PHYSMEM_BITS
70#ifdef MAX_PHYSMEM_BITS
71#define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
72#else
73/*
74 * If this definition of MAX_PHYSMEM_BITS is used, ZS_OBJ_PFN_SHIFT will
75 * just be PAGE_SHIFT
76 */
77#define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
78#endif
79#endif
80
81#define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
82
83/*
84 * Head in allocated object should have OBJ_ALLOCATED_TAG
85 * to identify the object was allocated or not.
86 * It's okay to add the status bit in the least bit because
87 * header keeps handle which is 4byte-aligned address so we
88 * have room for two bit at least.
89 */
90#define OBJ_ALLOCATED_TAG 1
91
92#define OBJ_TAG_BITS 1
93#define OBJ_TAG_MASK OBJ_ALLOCATED_TAG
94
95/*
96 * obj is encoded as [PFN | class_idx | obj_idx] within an unsigned long:
97 *
98 * |<-- _PFN_BITS -->|<-- ZS_OBJ_CLASS_BITS -->|<-- ZS_OBJ_IDX_BITS -->|
99 * +-----------------+-------------------------+-----------------------+
100 * | PFN | class_idx | obj_idx |
101 * +-----------------+-------------------------+-----------------------+
102 * MSB ^ LSB
103 * |
104 * +-- ZS_OBJ_PFN_SHIFT
105 *
106 * Encoding class_idx into obj lets zs_free() locate the size_class
107 * without holding pool->lock; class_idx is invariant across page
108 * migration (only PFN changes), so a lockless read of the obj value
109 * always yields a valid class_idx.
110 */
111#define ZS_OBJ_PFN_SHIFT (BITS_PER_LONG - _PFN_BITS)
112
113#define HUGE_BITS 1
114#define FULLNESS_BITS 4
115#define CLASS_BITS 8
116#define MAGIC_VAL_BITS 8
117
118#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(CONFIG_ZSMALLOC_CHAIN_SIZE, UL))
119
120/*
121 * Bits to index a page within a zspage = ceil(log2(ZS_MAX_PAGES_PER_ZSPAGE)).
122 * Computed at preprocessor time, for use in #if below. Kconfig
123 * restricts ZSMALLOC_CHAIN_SIZE to [4, 16].
124 */
125#if ZS_MAX_PAGES_PER_ZSPAGE <= 4
126#define ZS_PAGES_PER_ZSPAGE_BITS 2
127#elif ZS_MAX_PAGES_PER_ZSPAGE <= 8
128#define ZS_PAGES_PER_ZSPAGE_BITS 3
129#elif ZS_MAX_PAGES_PER_ZSPAGE <= 16
130#define ZS_PAGES_PER_ZSPAGE_BITS 4
131#else
132#error "ZSMALLOC_CHAIN_SIZE out of expected range [4,16]"
133#endif
134
135/*
136 * Bits to index an object within a single PAGE_SIZE at the smallest
137 * possible object size: log2(PAGE_SIZE / 32) = PAGE_SHIFT - 5.
138 * 32 is the hard floor of ZS_MIN_ALLOC_SIZE.
139 */
140#define ZS_OBJS_PER_PAGE_BITS (PAGE_SHIFT - 5)
141
142/*
143 * Bits to index any object in the densest possible zspage. Below this,
144 * ZS_MIN_ALLOC_SIZE is auto-raised by the MAX(32, ...) formula -- still
145 * correct, but objects are coarser.
146 */
147#define ZS_OBJS_PER_ZSPAGE_BITS \
148 (ZS_PAGES_PER_ZSPAGE_BITS + ZS_OBJS_PER_PAGE_BITS)
149
150/*
151 * Encode class_idx only when obj has spare bits; otherwise
152 * ZS_OBJ_CLASS_BITS folds to 0 (32-bit, or 64-bit UML/fallback).
153 */
154#if BITS_PER_LONG >= 64 && \
155 ZS_OBJ_PFN_SHIFT >= (CLASS_BITS + 1) + ZS_OBJS_PER_ZSPAGE_BITS
156#define ZS_OBJ_CLASS_BITS (CLASS_BITS + 1)
157#else
158#define ZS_OBJ_CLASS_BITS 0
159#endif
160#define ZS_OBJ_CLASS_MASK ((_AC(1, UL) << ZS_OBJ_CLASS_BITS) - 1)
161
162#define ZS_OBJ_IDX_BITS (ZS_OBJ_PFN_SHIFT - ZS_OBJ_CLASS_BITS)
163#define ZS_OBJ_IDX_MASK ((_AC(1, UL) << ZS_OBJ_IDX_BITS) - 1)
164
165/*
166 * Belt-and-suspenders: the #if above already guarantees this when
167 * class_idx is enabled. Catches future tweaks that bypass it.
168 */
169static_assert(ZS_OBJ_IDX_BITS >= ZS_PAGES_PER_ZSPAGE_BITS,
170 "zsmalloc: ZS_MIN_ALLOC_SIZE would exceed ZS_MAX_ALLOC_SIZE");
171
172/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
173#define ZS_MIN_ALLOC_SIZE \
174 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> ZS_OBJ_IDX_BITS))
175/* each chunk includes extra space to keep handle */
176#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
177
178/*
179 * On systems with 4K page size, this gives 255 size classes! There is a
180 * trade-off here:
181 * - Large number of size classes is potentially wasteful as free page are
182 * spread across these classes
183 * - Small number of size classes causes large internal fragmentation
184 * - Probably its better to use specific size classes (empirically
185 * determined). NOTE: all those class sizes must be set as multiple of
186 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
187 *
188 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
189 * (reason above)
190 */
191#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
192#define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
193 ZS_SIZE_CLASS_DELTA) + 1)
194
195/*
196 * Pages are distinguished by the ratio of used memory (that is the ratio
197 * of ->inuse objects to all objects that page can store). For example,
198 * INUSE_RATIO_10 means that the ratio of used objects is > 0% and <= 10%.
199 *
200 * The number of fullness groups is not random. It allows us to keep
201 * difference between the least busy page in the group (minimum permitted
202 * number of ->inuse objects) and the most busy page (maximum permitted
203 * number of ->inuse objects) at a reasonable value.
204 */
205enum fullness_group {
206 ZS_INUSE_RATIO_0,
207 ZS_INUSE_RATIO_10,
208 /* NOTE: 8 more fullness groups here */
209 ZS_INUSE_RATIO_99 = 10,
210 ZS_INUSE_RATIO_100,
211 NR_FULLNESS_GROUPS,
212};
213
214enum class_stat_type {
215 /* NOTE: stats for 12 fullness groups here: from inuse 0 to 100 */
216 ZS_OBJS_ALLOCATED = NR_FULLNESS_GROUPS,
217 ZS_OBJS_INUSE,
218 NR_CLASS_STAT_TYPES,
219};
220
221struct zs_size_stat {
222 unsigned long objs[NR_CLASS_STAT_TYPES];
223};
224
225#ifdef CONFIG_ZSMALLOC_STAT
226static struct dentry *zs_stat_root;
227#endif
228
229static size_t huge_class_size;
230
231struct size_class {
232 spinlock_t lock;
233 struct list_head fullness_list[NR_FULLNESS_GROUPS];
234 /*
235 * Size of objects stored in this class. Must be multiple
236 * of ZS_ALIGN.
237 */
238 int size;
239 int objs_per_zspage;
240 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
241 int pages_per_zspage;
242
243 unsigned int index;
244 struct zs_size_stat stats;
245};
246
247/*
248 * Placed within free objects to form a singly linked list.
249 * For every zspage, zspage->freeobj gives head of this list.
250 *
251 * This must be power of 2 and less than or equal to ZS_ALIGN
252 */
253struct link_free {
254 union {
255 /*
256 * Free object index;
257 * It's valid for non-allocated object
258 */
259 unsigned long next;
260 /*
261 * Handle of allocated object.
262 */
263 unsigned long handle;
264 };
265};
266
267static struct kmem_cache *handle_cachep;
268static struct kmem_cache *zspage_cachep;
269
270struct zs_pool {
271 const char *name;
272
273 struct size_class *size_class[ZS_SIZE_CLASSES];
274
275 atomic_long_t pages_allocated;
276
277 struct zs_pool_stats stats;
278
279 /* Compact classes */
280 struct shrinker *shrinker;
281
282#ifdef CONFIG_ZSMALLOC_STAT
283 struct dentry *stat_dentry;
284#endif
285#ifdef CONFIG_COMPACTION
286 struct work_struct free_work;
287#endif
288 /* protect zspage migration/compaction */
289 rwlock_t lock;
290 atomic_t compaction_in_progress;
291};
292
293static inline void zpdesc_set_first(struct zpdesc *zpdesc)
294{
295 SetPagePrivate(zpdesc_page(zpdesc));
296}
297
298static inline void zpdesc_inc_zone_page_state(struct zpdesc *zpdesc)
299{
300 inc_zone_page_state(zpdesc_page(zpdesc), NR_ZSPAGES);
301}
302
303static inline void zpdesc_dec_zone_page_state(struct zpdesc *zpdesc)
304{
305 dec_zone_page_state(zpdesc_page(zpdesc), NR_ZSPAGES);
306}
307
308static inline struct zpdesc *alloc_zpdesc(gfp_t gfp, const int nid)
309{
310 struct page *page = alloc_pages_node(nid, gfp, 0);
311
312 return page_zpdesc(page);
313}
314
315static inline void free_zpdesc(struct zpdesc *zpdesc)
316{
317 struct page *page = zpdesc_page(zpdesc);
318
319 /* PageZsmalloc is sticky until the page is freed to the buddy. */
320 __free_page(page);
321}
322
323#define ZS_PAGE_UNLOCKED 0
324#define ZS_PAGE_WRLOCKED -1
325
326struct zspage_lock {
327 spinlock_t lock;
328 int cnt;
329 struct lockdep_map dep_map;
330};
331
332struct zspage {
333 struct {
334 unsigned int huge:HUGE_BITS;
335 unsigned int fullness:FULLNESS_BITS;
336 unsigned int class:CLASS_BITS + 1;
337 unsigned int magic:MAGIC_VAL_BITS;
338 };
339 unsigned int inuse;
340 unsigned int freeobj;
341 struct zpdesc *first_zpdesc;
342 struct list_head list; /* fullness list */
343 struct zs_pool *pool;
344 struct zspage_lock zsl;
345};
346
347static void zspage_lock_init(struct zspage *zspage)
348{
349 static struct lock_class_key __key;
350 struct zspage_lock *zsl = &zspage->zsl;
351
352 lockdep_init_map(&zsl->dep_map, "zspage->lock", &__key, 0);
353 spin_lock_init(&zsl->lock);
354 zsl->cnt = ZS_PAGE_UNLOCKED;
355}
356
357/*
358 * The zspage lock can be held from atomic contexts, but it needs to remain
359 * preemptible when held for reading because it remains held outside of those
360 * atomic contexts, otherwise we unnecessarily lose preemptibility.
361 *
362 * To achieve this, the following rules are enforced on readers and writers:
363 *
364 * - Writers are blocked by both writers and readers, while readers are only
365 * blocked by writers (i.e. normal rwlock semantics).
366 *
367 * - Writers are always atomic (to allow readers to spin waiting for them).
368 *
369 * - Writers always use trylock (as the lock may be held be sleeping readers).
370 *
371 * - Readers may spin on the lock (as they can only wait for atomic writers).
372 *
373 * - Readers may sleep while holding the lock (as writes only use trylock).
374 */
375static void zspage_read_lock(struct zspage *zspage)
376{
377 struct zspage_lock *zsl = &zspage->zsl;
378
379 rwsem_acquire_read(&zsl->dep_map, 0, 0, _RET_IP_);
380
381 spin_lock(&zsl->lock);
382 zsl->cnt++;
383 spin_unlock(&zsl->lock);
384
385 lock_acquired(&zsl->dep_map, _RET_IP_);
386}
387
388static void zspage_read_unlock(struct zspage *zspage)
389{
390 struct zspage_lock *zsl = &zspage->zsl;
391
392 rwsem_release(&zsl->dep_map, _RET_IP_);
393
394 spin_lock(&zsl->lock);
395 zsl->cnt--;
396 spin_unlock(&zsl->lock);
397}
398
399static __must_check bool zspage_write_trylock(struct zspage *zspage)
400{
401 struct zspage_lock *zsl = &zspage->zsl;
402
403 spin_lock(&zsl->lock);
404 if (zsl->cnt == ZS_PAGE_UNLOCKED) {
405 zsl->cnt = ZS_PAGE_WRLOCKED;
406 rwsem_acquire(&zsl->dep_map, 0, 1, _RET_IP_);
407 lock_acquired(&zsl->dep_map, _RET_IP_);
408 return true;
409 }
410
411 spin_unlock(&zsl->lock);
412 return false;
413}
414
415static void zspage_write_unlock(struct zspage *zspage)
416{
417 struct zspage_lock *zsl = &zspage->zsl;
418
419 rwsem_release(&zsl->dep_map, _RET_IP_);
420
421 zsl->cnt = ZS_PAGE_UNLOCKED;
422 spin_unlock(&zsl->lock);
423}
424
425/* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
426static void SetZsHugePage(struct zspage *zspage)
427{
428 zspage->huge = 1;
429}
430
431static bool ZsHugePage(struct zspage *zspage)
432{
433 return zspage->huge;
434}
435
436#ifdef CONFIG_COMPACTION
437static void kick_deferred_free(struct zs_pool *pool);
438static void init_deferred_free(struct zs_pool *pool);
439static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
440#else
441static void kick_deferred_free(struct zs_pool *pool) {}
442static void init_deferred_free(struct zs_pool *pool) {}
443static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
444#endif
445
446static unsigned long cache_alloc_handle(gfp_t gfp)
447{
448 gfp = gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE);
449
450 return (unsigned long)kmem_cache_alloc(handle_cachep, gfp);
451}
452
453static void cache_free_handle(unsigned long handle)
454{
455 kmem_cache_free(handle_cachep, (void *)handle);
456}
457
458static struct zspage *cache_alloc_zspage(gfp_t gfp)
459{
460 gfp = gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE);
461
462 return kmem_cache_zalloc(zspage_cachep, gfp);
463}
464
465static void cache_free_zspage(struct zspage *zspage)
466{
467 kmem_cache_free(zspage_cachep, zspage);
468}
469
470/*
471 * Pairs with READ_ONCE() in handle_to_obj(): zs_free() may read the
472 * handle locklessly, so prevent store tearing here.
473 */
474static void record_obj(unsigned long handle, unsigned long obj)
475{
476 WRITE_ONCE(*(unsigned long *)handle, obj);
477}
478
479static inline bool __maybe_unused is_first_zpdesc(struct zpdesc *zpdesc)
480{
481 return PagePrivate(zpdesc_page(zpdesc));
482}
483
484/* Protected by class->lock */
485static inline int get_zspage_inuse(struct zspage *zspage)
486{
487 return zspage->inuse;
488}
489
490static inline void mod_zspage_inuse(struct zspage *zspage, int val)
491{
492 zspage->inuse += val;
493}
494
495static struct zpdesc *get_first_zpdesc(struct zspage *zspage)
496{
497 struct zpdesc *first_zpdesc = zspage->first_zpdesc;
498
499 VM_BUG_ON_PAGE(!is_first_zpdesc(first_zpdesc), zpdesc_page(first_zpdesc));
500 return first_zpdesc;
501}
502
503#define FIRST_OBJ_PAGE_TYPE_MASK 0xffffff
504
505static inline unsigned int get_first_obj_offset(struct zpdesc *zpdesc)
506{
507 VM_WARN_ON_ONCE(!PageZsmalloc(zpdesc_page(zpdesc)));
508 return zpdesc->first_obj_offset & FIRST_OBJ_PAGE_TYPE_MASK;
509}
510
511static inline void set_first_obj_offset(struct zpdesc *zpdesc, unsigned int offset)
512{
513 /* With 24 bits available, we can support offsets into 16 MiB pages. */
514 BUILD_BUG_ON(PAGE_SIZE > SZ_16M);
515 VM_WARN_ON_ONCE(!PageZsmalloc(zpdesc_page(zpdesc)));
516 VM_WARN_ON_ONCE(offset & ~FIRST_OBJ_PAGE_TYPE_MASK);
517 zpdesc->first_obj_offset &= ~FIRST_OBJ_PAGE_TYPE_MASK;
518 zpdesc->first_obj_offset |= offset & FIRST_OBJ_PAGE_TYPE_MASK;
519}
520
521static inline unsigned int get_freeobj(struct zspage *zspage)
522{
523 return zspage->freeobj;
524}
525
526static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
527{
528 zspage->freeobj = obj;
529}
530
531static struct size_class *zspage_class(struct zs_pool *pool,
532 struct zspage *zspage)
533{
534 return pool->size_class[zspage->class];
535}
536
537/*
538 * zsmalloc divides the pool into various size classes where each
539 * class maintains a list of zspages where each zspage is divided
540 * into equal sized chunks. Each allocation falls into one of these
541 * classes depending on its size. This function returns index of the
542 * size class which has chunk size big enough to hold the given size.
543 */
544static int get_size_class_index(int size)
545{
546 int idx = 0;
547
548 if (likely(size > ZS_MIN_ALLOC_SIZE))
549 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
550 ZS_SIZE_CLASS_DELTA);
551
552 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
553}
554
555static struct size_class *lookup_size_class(struct zs_pool *pool, size_t size)
556{
557 return pool->size_class[get_size_class_index(size + ZS_HANDLE_SIZE)];
558}
559
560static inline void class_stat_add(struct size_class *class, int type,
561 unsigned long cnt)
562{
563 class->stats.objs[type] += cnt;
564}
565
566static inline void class_stat_sub(struct size_class *class, int type,
567 unsigned long cnt)
568{
569 class->stats.objs[type] -= cnt;
570}
571
572static inline unsigned long class_stat_read(struct size_class *class, int type)
573{
574 return class->stats.objs[type];
575}
576
577#ifdef CONFIG_ZSMALLOC_STAT
578
579static void __init zs_stat_init(void)
580{
581 if (!debugfs_initialized()) {
582 pr_warn("debugfs not available, stat dir not created\n");
583 return;
584 }
585
586 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
587}
588
589static void __exit zs_stat_exit(void)
590{
591 debugfs_remove_recursive(zs_stat_root);
592}
593
594static unsigned long zs_can_compact(struct size_class *class);
595
596static int zs_stats_size_show(struct seq_file *s, void *v)
597{
598 int i, fg;
599 struct zs_pool *pool = s->private;
600 struct size_class *class;
601 int objs_per_zspage;
602 unsigned long obj_allocated, obj_used, pages_used, freeable;
603 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
604 unsigned long total_freeable = 0;
605 unsigned long inuse_totals[NR_FULLNESS_GROUPS] = {0, };
606
607 seq_printf(s, " %5s %5s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %13s %10s %10s %16s %8s\n",
608 "class", "size", "10%", "20%", "30%", "40%",
609 "50%", "60%", "70%", "80%", "90%", "99%", "100%",
610 "obj_allocated", "obj_used", "pages_used",
611 "pages_per_zspage", "freeable");
612
613 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
614
615 class = pool->size_class[i];
616
617 if (class->index != i)
618 continue;
619
620 spin_lock(&class->lock);
621
622 seq_printf(s, " %5u %5u ", i, class->size);
623 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++) {
624 inuse_totals[fg] += class_stat_read(class, fg);
625 seq_printf(s, "%9lu ", class_stat_read(class, fg));
626 }
627
628 obj_allocated = class_stat_read(class, ZS_OBJS_ALLOCATED);
629 obj_used = class_stat_read(class, ZS_OBJS_INUSE);
630 freeable = zs_can_compact(class);
631 spin_unlock(&class->lock);
632
633 objs_per_zspage = class->objs_per_zspage;
634 pages_used = obj_allocated / objs_per_zspage *
635 class->pages_per_zspage;
636
637 seq_printf(s, "%13lu %10lu %10lu %16d %8lu\n",
638 obj_allocated, obj_used, pages_used,
639 class->pages_per_zspage, freeable);
640
641 total_objs += obj_allocated;
642 total_used_objs += obj_used;
643 total_pages += pages_used;
644 total_freeable += freeable;
645 }
646
647 seq_printf(s, "\n %5s %5s ", "Total", "");
648
649 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++)
650 seq_printf(s, "%9lu ", inuse_totals[fg]);
651
652 seq_printf(s, "%13lu %10lu %10lu %16s %8lu\n",
653 total_objs, total_used_objs, total_pages, "",
654 total_freeable);
655
656 return 0;
657}
658DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
659
660static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
661{
662 if (!zs_stat_root) {
663 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
664 return;
665 }
666
667 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
668
669 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
670 &zs_stats_size_fops);
671}
672
673static void zs_pool_stat_destroy(struct zs_pool *pool)
674{
675 debugfs_remove_recursive(pool->stat_dentry);
676}
677
678#else /* CONFIG_ZSMALLOC_STAT */
679static void __init zs_stat_init(void)
680{
681}
682
683static void __exit zs_stat_exit(void)
684{
685}
686
687static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
688{
689}
690
691static inline void zs_pool_stat_destroy(struct zs_pool *pool)
692{
693}
694#endif
695
696
697/*
698 * For each size class, zspages are divided into different groups
699 * depending on their usage ratio. This function returns fullness
700 * status of the given page.
701 */
702static int get_fullness_group(struct size_class *class, struct zspage *zspage)
703{
704 int inuse, objs_per_zspage, ratio;
705
706 inuse = get_zspage_inuse(zspage);
707 objs_per_zspage = class->objs_per_zspage;
708
709 if (inuse == 0)
710 return ZS_INUSE_RATIO_0;
711 if (inuse == objs_per_zspage)
712 return ZS_INUSE_RATIO_100;
713
714 ratio = 100 * inuse / objs_per_zspage;
715 /*
716 * Take integer division into consideration: a page with one inuse
717 * object out of 127 possible, will end up having 0 usage ratio,
718 * which is wrong as it belongs in ZS_INUSE_RATIO_10 fullness group.
719 */
720 return ratio / 10 + 1;
721}
722
723/*
724 * Each size class maintains various freelists and zspages are assigned
725 * to one of these freelists based on the number of live objects they
726 * have. This functions inserts the given zspage into the freelist
727 * identified by <class, fullness_group>.
728 */
729static void insert_zspage(struct size_class *class,
730 struct zspage *zspage,
731 int fullness)
732{
733 class_stat_add(class, fullness, 1);
734 list_add(&zspage->list, &class->fullness_list[fullness]);
735 zspage->fullness = fullness;
736}
737
738/*
739 * This function removes the given zspage from the freelist identified
740 * by <class, fullness_group>.
741 */
742static void remove_zspage(struct size_class *class, struct zspage *zspage)
743{
744 int fullness = zspage->fullness;
745
746 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
747
748 list_del_init(&zspage->list);
749 class_stat_sub(class, fullness, 1);
750}
751
752/*
753 * Each size class maintains zspages in different fullness groups depending
754 * on the number of live objects they contain. When allocating or freeing
755 * objects, the fullness status of the page can change, for instance, from
756 * INUSE_RATIO_80 to INUSE_RATIO_70 when freeing an object. This function
757 * checks if such a status change has occurred for the given page and
758 * accordingly moves the page from the list of the old fullness group to that
759 * of the new fullness group.
760 */
761static int fix_fullness_group(struct size_class *class, struct zspage *zspage)
762{
763 int newfg;
764
765 newfg = get_fullness_group(class, zspage);
766 if (newfg == zspage->fullness)
767 goto out;
768
769 remove_zspage(class, zspage);
770 insert_zspage(class, zspage, newfg);
771out:
772 return newfg;
773}
774
775static struct zspage *get_zspage(struct zpdesc *zpdesc)
776{
777 struct zspage *zspage = zpdesc->zspage;
778
779 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
780 return zspage;
781}
782
783static struct zpdesc *get_next_zpdesc(struct zpdesc *zpdesc)
784{
785 struct zspage *zspage = get_zspage(zpdesc);
786
787 if (unlikely(ZsHugePage(zspage)))
788 return NULL;
789
790 return zpdesc->next;
791}
792
793/**
794 * obj_to_location - get (<zpdesc>, <obj_idx>) from encoded object value
795 * @obj: the encoded object value
796 * @zpdesc: zpdesc object resides in zspage
797 * @obj_idx: object index
798 */
799static void obj_to_location(unsigned long obj, struct zpdesc **zpdesc,
800 unsigned int *obj_idx)
801{
802 *zpdesc = pfn_zpdesc(obj >> ZS_OBJ_PFN_SHIFT);
803 *obj_idx = (obj & ZS_OBJ_IDX_MASK);
804}
805
806static void obj_to_zpdesc(unsigned long obj, struct zpdesc **zpdesc)
807{
808 *zpdesc = pfn_zpdesc(obj >> ZS_OBJ_PFN_SHIFT);
809}
810
811/**
812 * location_to_obj - encode (<zpdesc>, <obj_idx>, <class_idx>) into obj value
813 * @zpdesc: zpdesc object resides in zspage
814 * @obj_idx: object index
815 * @class_idx: size class index; ignored when ZS_OBJ_CLASS_BITS == 0
816 */
817static unsigned long location_to_obj(struct zpdesc *zpdesc, unsigned int obj_idx,
818 unsigned int class_idx)
819{
820 unsigned long obj;
821
822 obj = zpdesc_pfn(zpdesc) << ZS_OBJ_PFN_SHIFT;
823 obj |= (unsigned long)(class_idx & ZS_OBJ_CLASS_MASK) << ZS_OBJ_IDX_BITS;
824 obj |= obj_idx & ZS_OBJ_IDX_MASK;
825
826 return obj;
827}
828
829static unsigned long handle_to_obj(unsigned long handle)
830{
831 return READ_ONCE(*(unsigned long *)handle);
832}
833
834static inline bool obj_allocated(struct zpdesc *zpdesc, void *obj,
835 unsigned long *phandle)
836{
837 unsigned long handle;
838 struct zspage *zspage = get_zspage(zpdesc);
839
840 if (unlikely(ZsHugePage(zspage))) {
841 VM_BUG_ON_PAGE(!is_first_zpdesc(zpdesc), zpdesc_page(zpdesc));
842 handle = zpdesc->handle;
843 } else
844 handle = *(unsigned long *)obj;
845
846 if (!(handle & OBJ_ALLOCATED_TAG))
847 return false;
848
849 /* Clear all tags before returning the handle */
850 *phandle = handle & ~OBJ_TAG_MASK;
851 return true;
852}
853
854static void reset_zpdesc(struct zpdesc *zpdesc)
855{
856 struct page *page = zpdesc_page(zpdesc);
857
858 ClearPagePrivate(page);
859 zpdesc->zspage = NULL;
860 zpdesc->next = NULL;
861 /* PageZsmalloc is sticky until the page is freed to the buddy. */
862}
863
864static int trylock_zspage(struct zspage *zspage)
865{
866 struct zpdesc *cursor, *fail;
867
868 for (cursor = get_first_zpdesc(zspage); cursor != NULL; cursor =
869 get_next_zpdesc(cursor)) {
870 if (!zpdesc_trylock(cursor)) {
871 fail = cursor;
872 goto unlock;
873 }
874 }
875
876 return 1;
877unlock:
878 for (cursor = get_first_zpdesc(zspage); cursor != fail; cursor =
879 get_next_zpdesc(cursor))
880 zpdesc_unlock(cursor);
881
882 return 0;
883}
884
885/*
886 * Three free helpers, kept apart here:
887 *
888 * __free_zspage_lockless(): bare core; walks zpdescs and returns pages
889 * to the buddy allocator. Caller owns all zpdesc locks and has
890 * removed the zspage from its class list. Used by zs_free() outside
891 * class->lock so the buddy-side work does not stall the class.
892 *
893 * __free_zspage(): __free_zspage_lockless() + per-class accounting,
894 * under class->lock. Used by async_free_zspage(), the worker for
895 * zspages whose trylock_zspage() failed.
896 *
897 * free_zspage(): full wrapper - trylock zpdescs, remove from class
898 * list, call __free_zspage(); kicks deferred free on contention.
899 * Used by compaction.
900 */
901static inline void __free_zspage_lockless(struct zspage *zspage)
902{
903 struct zpdesc *zpdesc, *next;
904
905 VM_BUG_ON(get_zspage_inuse(zspage));
906 VM_BUG_ON(zspage->fullness != ZS_INUSE_RATIO_0);
907
908 next = zpdesc = get_first_zpdesc(zspage);
909 do {
910 VM_BUG_ON_PAGE(!zpdesc_is_locked(zpdesc), zpdesc_page(zpdesc));
911 next = get_next_zpdesc(zpdesc);
912 reset_zpdesc(zpdesc);
913 zpdesc_unlock(zpdesc);
914 zpdesc_dec_zone_page_state(zpdesc);
915 zpdesc_put(zpdesc);
916 zpdesc = next;
917 } while (zpdesc != NULL);
918
919 cache_free_zspage(zspage);
920}
921
922static void __free_zspage(struct zs_pool *pool, struct size_class *class,
923 struct zspage *zspage)
924{
925 assert_spin_locked(&class->lock);
926 __free_zspage_lockless(zspage);
927 class_stat_sub(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
928 atomic_long_sub(class->pages_per_zspage, &pool->pages_allocated);
929}
930
931static void free_zspage(struct zs_pool *pool, struct size_class *class,
932 struct zspage *zspage)
933{
934 VM_BUG_ON(get_zspage_inuse(zspage));
935 VM_BUG_ON(list_empty(&zspage->list));
936
937 /*
938 * Since zs_free couldn't be sleepable, this function cannot call
939 * lock_page. The page locks trylock_zspage got will be released
940 * by __free_zspage.
941 */
942 if (!trylock_zspage(zspage)) {
943 kick_deferred_free(pool);
944 return;
945 }
946
947 remove_zspage(class, zspage);
948 __free_zspage(pool, class, zspage);
949}
950
951/* Initialize a newly allocated zspage */
952static void init_zspage(struct size_class *class, struct zspage *zspage)
953{
954 unsigned int freeobj = 1;
955 unsigned long off = 0;
956 struct zpdesc *zpdesc = get_first_zpdesc(zspage);
957
958 while (zpdesc) {
959 struct zpdesc *next_zpdesc;
960 struct link_free *link;
961 void *vaddr;
962
963 set_first_obj_offset(zpdesc, off);
964
965 vaddr = kmap_local_zpdesc(zpdesc);
966 link = (struct link_free *)vaddr + off / sizeof(*link);
967
968 while ((off += class->size) < PAGE_SIZE) {
969 link->next = freeobj++ << OBJ_TAG_BITS;
970 link += class->size / sizeof(*link);
971 }
972
973 /*
974 * We now come to the last (full or partial) object on this
975 * page, which must point to the first object on the next
976 * page (if present)
977 */
978 next_zpdesc = get_next_zpdesc(zpdesc);
979 if (next_zpdesc) {
980 link->next = freeobj++ << OBJ_TAG_BITS;
981 } else {
982 /*
983 * Reset OBJ_TAG_BITS bit to last link to tell
984 * whether it's allocated object or not.
985 */
986 link->next = -1UL << OBJ_TAG_BITS;
987 }
988 kunmap_local(vaddr);
989 zpdesc = next_zpdesc;
990 off %= PAGE_SIZE;
991 }
992
993 set_freeobj(zspage, 0);
994}
995
996static void create_page_chain(struct size_class *class, struct zspage *zspage,
997 struct zpdesc *zpdescs[])
998{
999 int i;
1000 struct zpdesc *zpdesc;
1001 struct zpdesc *prev_zpdesc = NULL;
1002 int nr_zpdescs = class->pages_per_zspage;
1003
1004 /*
1005 * Allocate individual pages and link them together as:
1006 * 1. all pages are linked together using zpdesc->next
1007 * 2. each sub-page point to zspage using zpdesc->zspage
1008 *
1009 * we set PG_private to identify the first zpdesc (i.e. no other zpdesc
1010 * has this flag set).
1011 */
1012 for (i = 0; i < nr_zpdescs; i++) {
1013 zpdesc = zpdescs[i];
1014 zpdesc->zspage = zspage;
1015 zpdesc->next = NULL;
1016 if (i == 0) {
1017 zspage->first_zpdesc = zpdesc;
1018 zpdesc_set_first(zpdesc);
1019 if (unlikely(class->objs_per_zspage == 1 &&
1020 class->pages_per_zspage == 1))
1021 SetZsHugePage(zspage);
1022 } else {
1023 prev_zpdesc->next = zpdesc;
1024 }
1025 prev_zpdesc = zpdesc;
1026 }
1027}
1028
1029/*
1030 * Allocate a zspage for the given size class
1031 */
1032static struct zspage *alloc_zspage(struct zs_pool *pool,
1033 struct size_class *class,
1034 gfp_t gfp, const int nid)
1035{
1036 int i;
1037 struct zpdesc *zpdescs[ZS_MAX_PAGES_PER_ZSPAGE];
1038 struct zspage *zspage = cache_alloc_zspage(gfp);
1039
1040 if (!zspage)
1041 return NULL;
1042
1043 if (!IS_ENABLED(CONFIG_COMPACTION))
1044 gfp &= ~__GFP_MOVABLE;
1045
1046 zspage->magic = ZSPAGE_MAGIC;
1047 zspage->pool = pool;
1048 zspage->class = class->index;
1049 zspage_lock_init(zspage);
1050
1051 for (i = 0; i < class->pages_per_zspage; i++) {
1052 struct zpdesc *zpdesc;
1053
1054 zpdesc = alloc_zpdesc(gfp, nid);
1055 if (!zpdesc) {
1056 while (--i >= 0) {
1057 zpdesc_dec_zone_page_state(zpdescs[i]);
1058 free_zpdesc(zpdescs[i]);
1059 }
1060 cache_free_zspage(zspage);
1061 return NULL;
1062 }
1063 __zpdesc_set_zsmalloc(zpdesc);
1064
1065 zpdesc_inc_zone_page_state(zpdesc);
1066 zpdescs[i] = zpdesc;
1067 }
1068
1069 create_page_chain(class, zspage, zpdescs);
1070 init_zspage(class, zspage);
1071
1072 return zspage;
1073}
1074
1075static struct zspage *find_get_zspage(struct size_class *class)
1076{
1077 int i;
1078 struct zspage *zspage;
1079
1080 for (i = ZS_INUSE_RATIO_99; i >= ZS_INUSE_RATIO_0; i--) {
1081 zspage = list_first_entry_or_null(&class->fullness_list[i],
1082 struct zspage, list);
1083 if (zspage)
1084 break;
1085 }
1086
1087 return zspage;
1088}
1089
1090static bool can_merge(struct size_class *prev, int pages_per_zspage,
1091 int objs_per_zspage)
1092{
1093 if (prev->pages_per_zspage == pages_per_zspage &&
1094 prev->objs_per_zspage == objs_per_zspage)
1095 return true;
1096
1097 return false;
1098}
1099
1100static bool zspage_full(struct size_class *class, struct zspage *zspage)
1101{
1102 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1103}
1104
1105static bool zspage_empty(struct zspage *zspage)
1106{
1107 return get_zspage_inuse(zspage) == 0;
1108}
1109
1110/**
1111 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1112 * that hold objects of the provided size.
1113 * @pool: zsmalloc pool to use
1114 * @size: object size
1115 *
1116 * Context: Any context.
1117 *
1118 * Return: the index of the zsmalloc &size_class that hold objects of the
1119 * provided size.
1120 */
1121unsigned int zs_lookup_class_index(struct zs_pool *pool, unsigned int size)
1122{
1123 struct size_class *class;
1124
1125 class = lookup_size_class(pool, size);
1126
1127 return class->index;
1128}
1129EXPORT_SYMBOL_GPL(zs_lookup_class_index);
1130
1131unsigned long zs_get_total_pages(struct zs_pool *pool)
1132{
1133 return atomic_long_read(&pool->pages_allocated);
1134}
1135EXPORT_SYMBOL_GPL(zs_get_total_pages);
1136
1137void *zs_obj_read_begin(struct zs_pool *pool, unsigned long handle,
1138 size_t mem_len, void *local_copy)
1139{
1140 struct zspage *zspage;
1141 struct zpdesc *zpdesc;
1142 unsigned long obj, off;
1143 unsigned int obj_idx;
1144 struct size_class *class;
1145 void *addr;
1146
1147 /* Guarantee we can get zspage from handle safely */
1148 read_lock(&pool->lock);
1149 obj = handle_to_obj(handle);
1150 obj_to_location(obj, &zpdesc, &obj_idx);
1151 zspage = get_zspage(zpdesc);
1152
1153 /* Make sure migration doesn't move any pages in this zspage */
1154 zspage_read_lock(zspage);
1155 read_unlock(&pool->lock);
1156
1157 class = zspage_class(pool, zspage);
1158 off = offset_in_page(class->size * obj_idx);
1159
1160 if (!ZsHugePage(zspage))
1161 off += ZS_HANDLE_SIZE;
1162
1163 if (off + mem_len <= PAGE_SIZE) {
1164 /* this object is contained entirely within a page */
1165 addr = kmap_local_zpdesc(zpdesc);
1166 addr += off;
1167 } else {
1168 size_t sizes[2];
1169
1170 /* this object spans two pages */
1171 sizes[0] = PAGE_SIZE - off;
1172 sizes[1] = mem_len - sizes[0];
1173 addr = local_copy;
1174
1175 memcpy_from_page(addr, zpdesc_page(zpdesc),
1176 off, sizes[0]);
1177 zpdesc = get_next_zpdesc(zpdesc);
1178 memcpy_from_page(addr + sizes[0],
1179 zpdesc_page(zpdesc),
1180 0, sizes[1]);
1181 }
1182
1183 return addr;
1184}
1185EXPORT_SYMBOL_GPL(zs_obj_read_begin);
1186
1187void zs_obj_read_end(struct zs_pool *pool, unsigned long handle,
1188 size_t mem_len, void *handle_mem)
1189{
1190 struct zspage *zspage;
1191 struct zpdesc *zpdesc;
1192 unsigned long obj, off;
1193 unsigned int obj_idx;
1194 struct size_class *class;
1195
1196 obj = handle_to_obj(handle);
1197 obj_to_location(obj, &zpdesc, &obj_idx);
1198 zspage = get_zspage(zpdesc);
1199 class = zspage_class(pool, zspage);
1200 off = offset_in_page(class->size * obj_idx);
1201
1202 if (!ZsHugePage(zspage))
1203 off += ZS_HANDLE_SIZE;
1204
1205 if (off + mem_len <= PAGE_SIZE) {
1206 handle_mem -= off;
1207 kunmap_local(handle_mem);
1208 }
1209
1210 zspage_read_unlock(zspage);
1211}
1212EXPORT_SYMBOL_GPL(zs_obj_read_end);
1213
1214void zs_obj_read_sg_begin(struct zs_pool *pool, unsigned long handle,
1215 struct scatterlist *sg, size_t mem_len)
1216{
1217 struct zspage *zspage;
1218 struct zpdesc *zpdesc;
1219 unsigned long obj, off;
1220 unsigned int obj_idx;
1221 struct size_class *class;
1222
1223 /* Guarantee we can get zspage from handle safely */
1224 read_lock(&pool->lock);
1225 obj = handle_to_obj(handle);
1226 obj_to_location(obj, &zpdesc, &obj_idx);
1227 zspage = get_zspage(zpdesc);
1228
1229 /* Make sure migration doesn't move any pages in this zspage */
1230 zspage_read_lock(zspage);
1231 read_unlock(&pool->lock);
1232
1233 class = zspage_class(pool, zspage);
1234 off = offset_in_page(class->size * obj_idx);
1235
1236 if (!ZsHugePage(zspage))
1237 off += ZS_HANDLE_SIZE;
1238
1239 if (off + mem_len <= PAGE_SIZE) {
1240 /* this object is contained entirely within a page */
1241 sg_init_table(sg, 1);
1242 sg_set_page(sg, zpdesc_page(zpdesc), mem_len, off);
1243 } else {
1244 size_t sizes[2];
1245
1246 /* this object spans two pages */
1247 sizes[0] = PAGE_SIZE - off;
1248 sizes[1] = mem_len - sizes[0];
1249
1250 sg_init_table(sg, 2);
1251 sg_set_page(sg, zpdesc_page(zpdesc), sizes[0], off);
1252
1253 zpdesc = get_next_zpdesc(zpdesc);
1254 sg = sg_next(sg);
1255
1256 sg_set_page(sg, zpdesc_page(zpdesc), sizes[1], 0);
1257 }
1258}
1259EXPORT_SYMBOL_GPL(zs_obj_read_sg_begin);
1260
1261void zs_obj_read_sg_end(struct zs_pool *pool, unsigned long handle)
1262{
1263 struct zspage *zspage;
1264 struct zpdesc *zpdesc;
1265 unsigned long obj;
1266 unsigned int obj_idx;
1267
1268 obj = handle_to_obj(handle);
1269 obj_to_location(obj, &zpdesc, &obj_idx);
1270 zspage = get_zspage(zpdesc);
1271
1272 zspage_read_unlock(zspage);
1273}
1274EXPORT_SYMBOL_GPL(zs_obj_read_sg_end);
1275
1276void zs_obj_write(struct zs_pool *pool, unsigned long handle,
1277 void *handle_mem, size_t mem_len)
1278{
1279 struct zspage *zspage;
1280 struct zpdesc *zpdesc;
1281 unsigned long obj, off;
1282 unsigned int obj_idx;
1283 struct size_class *class;
1284
1285 /* Guarantee we can get zspage from handle safely */
1286 read_lock(&pool->lock);
1287 obj = handle_to_obj(handle);
1288 obj_to_location(obj, &zpdesc, &obj_idx);
1289 zspage = get_zspage(zpdesc);
1290
1291 /* Make sure migration doesn't move any pages in this zspage */
1292 zspage_read_lock(zspage);
1293 read_unlock(&pool->lock);
1294
1295 class = zspage_class(pool, zspage);
1296 off = offset_in_page(class->size * obj_idx);
1297
1298 if (!ZsHugePage(zspage))
1299 off += ZS_HANDLE_SIZE;
1300
1301 if (off + mem_len <= PAGE_SIZE) {
1302 /* this object is contained entirely within a page */
1303 void *dst = kmap_local_zpdesc(zpdesc);
1304
1305 memcpy(dst + off, handle_mem, mem_len);
1306 kunmap_local(dst);
1307 } else {
1308 /* this object spans two pages */
1309 size_t sizes[2];
1310
1311 sizes[0] = PAGE_SIZE - off;
1312 sizes[1] = mem_len - sizes[0];
1313
1314 memcpy_to_page(zpdesc_page(zpdesc), off,
1315 handle_mem, sizes[0]);
1316 zpdesc = get_next_zpdesc(zpdesc);
1317 memcpy_to_page(zpdesc_page(zpdesc), 0,
1318 handle_mem + sizes[0], sizes[1]);
1319 }
1320
1321 zspage_read_unlock(zspage);
1322}
1323EXPORT_SYMBOL_GPL(zs_obj_write);
1324
1325/**
1326 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1327 * zsmalloc &size_class.
1328 * @pool: zsmalloc pool to use
1329 *
1330 * The function returns the size of the first huge class - any object of equal
1331 * or bigger size will be stored in zspage consisting of a single physical
1332 * page.
1333 *
1334 * Context: Any context.
1335 *
1336 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1337 */
1338size_t zs_huge_class_size(struct zs_pool *pool)
1339{
1340 return huge_class_size;
1341}
1342EXPORT_SYMBOL_GPL(zs_huge_class_size);
1343
1344static unsigned long obj_malloc(struct zs_pool *pool,
1345 struct zspage *zspage, unsigned long handle)
1346{
1347 int i, nr_zpdesc, offset;
1348 unsigned long obj;
1349 struct link_free *link;
1350 struct size_class *class;
1351
1352 struct zpdesc *m_zpdesc;
1353 unsigned long m_offset;
1354 void *vaddr;
1355
1356 class = pool->size_class[zspage->class];
1357 obj = get_freeobj(zspage);
1358
1359 offset = obj * class->size;
1360 nr_zpdesc = offset >> PAGE_SHIFT;
1361 m_offset = offset_in_page(offset);
1362 m_zpdesc = get_first_zpdesc(zspage);
1363
1364 for (i = 0; i < nr_zpdesc; i++)
1365 m_zpdesc = get_next_zpdesc(m_zpdesc);
1366
1367 vaddr = kmap_local_zpdesc(m_zpdesc);
1368 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1369 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1370 if (likely(!ZsHugePage(zspage)))
1371 /* record handle in the header of allocated chunk */
1372 link->handle = handle | OBJ_ALLOCATED_TAG;
1373 else
1374 zspage->first_zpdesc->handle = handle | OBJ_ALLOCATED_TAG;
1375
1376 kunmap_local(vaddr);
1377 mod_zspage_inuse(zspage, 1);
1378
1379 obj = location_to_obj(m_zpdesc, obj, zspage->class);
1380 record_obj(handle, obj);
1381
1382 return obj;
1383}
1384
1385
1386/**
1387 * zs_malloc - Allocate block of given size from pool.
1388 * @pool: pool to allocate from
1389 * @size: size of block to allocate
1390 * @gfp: gfp flags when allocating object
1391 * @nid: The preferred node id to allocate new zspage (if needed)
1392 *
1393 * On success, handle to the allocated object is returned,
1394 * otherwise an ERR_PTR().
1395 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1396 */
1397unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp,
1398 const int nid)
1399{
1400 unsigned long handle;
1401 struct size_class *class;
1402 int newfg;
1403 struct zspage *zspage;
1404
1405 if (unlikely(!size))
1406 return (unsigned long)ERR_PTR(-EINVAL);
1407
1408 if (unlikely(size > ZS_MAX_ALLOC_SIZE))
1409 return (unsigned long)ERR_PTR(-ENOSPC);
1410
1411 handle = cache_alloc_handle(gfp);
1412 if (!handle)
1413 return (unsigned long)ERR_PTR(-ENOMEM);
1414
1415 class = lookup_size_class(pool, size);
1416
1417 /* class->lock effectively protects the zpage migration */
1418 spin_lock(&class->lock);
1419 zspage = find_get_zspage(class);
1420 if (likely(zspage)) {
1421 obj_malloc(pool, zspage, handle);
1422 /* Now move the zspage to another fullness group, if required */
1423 fix_fullness_group(class, zspage);
1424 class_stat_add(class, ZS_OBJS_INUSE, 1);
1425
1426 goto out;
1427 }
1428
1429 spin_unlock(&class->lock);
1430
1431 zspage = alloc_zspage(pool, class, gfp, nid);
1432 if (!zspage) {
1433 cache_free_handle(handle);
1434 return (unsigned long)ERR_PTR(-ENOMEM);
1435 }
1436
1437 spin_lock(&class->lock);
1438 obj_malloc(pool, zspage, handle);
1439 newfg = get_fullness_group(class, zspage);
1440 insert_zspage(class, zspage, newfg);
1441 atomic_long_add(class->pages_per_zspage, &pool->pages_allocated);
1442 class_stat_add(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
1443 class_stat_add(class, ZS_OBJS_INUSE, 1);
1444
1445 /* We completely set up zspage so mark them as movable */
1446 SetZsPageMovable(pool, zspage);
1447out:
1448 spin_unlock(&class->lock);
1449
1450 return handle;
1451}
1452EXPORT_SYMBOL_GPL(zs_malloc);
1453
1454static void obj_free(int class_size, unsigned long obj)
1455{
1456 struct link_free *link;
1457 struct zspage *zspage;
1458 struct zpdesc *f_zpdesc;
1459 unsigned long f_offset;
1460 unsigned int f_objidx;
1461 void *vaddr;
1462
1463
1464 obj_to_location(obj, &f_zpdesc, &f_objidx);
1465 f_offset = offset_in_page(class_size * f_objidx);
1466 zspage = get_zspage(f_zpdesc);
1467
1468 vaddr = kmap_local_zpdesc(f_zpdesc);
1469 link = (struct link_free *)(vaddr + f_offset);
1470
1471 /* Insert this object in containing zspage's freelist */
1472 if (likely(!ZsHugePage(zspage)))
1473 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1474 else
1475 f_zpdesc->handle = 0;
1476 set_freeobj(zspage, f_objidx);
1477
1478 kunmap_local(vaddr);
1479 mod_zspage_inuse(zspage, -1);
1480}
1481
1482#if (ZS_OBJ_CLASS_BITS > 0) || defined(CONFIG_COMPACTION)
1483/* Folds to 0 when ZS_OBJ_CLASS_BITS == 0; no ifdef needed at callers. */
1484static unsigned int obj_to_class_idx(unsigned long obj)
1485{
1486 return (obj >> ZS_OBJ_IDX_BITS) & ZS_OBJ_CLASS_MASK;
1487}
1488#endif
1489
1490/*
1491 * Resolve @handle to its zspage / size_class and acquire class->lock.
1492 *
1493 * When class_idx is encoded in obj (ZS_OBJ_CLASS_BITS > 0), it is
1494 * invariant under page migration, so the handle can be read locklessly
1495 * to pick the size_class. Once class->lock is held migration is
1496 * blocked and the handle is re-read to obtain a stable PFN.
1497 *
1498 * Otherwise (32-bit, or 64-bit fallback paths like UML where the
1499 * encoding is disabled), fall back to pool->lock for the lookup.
1500 */
1501#if ZS_OBJ_CLASS_BITS > 0
1502static inline void obj_class_get_and_lock(struct zs_pool *pool, unsigned long handle,
1503 unsigned long *objp, struct zspage **zspagep,
1504 struct size_class **classp)
1505 __acquires(&(*classp)->lock)
1506{
1507 struct zpdesc *f_zpdesc;
1508 unsigned long obj;
1509
1510 obj = handle_to_obj(handle);
1511 *classp = pool->size_class[obj_to_class_idx(obj)];
1512 spin_lock(&(*classp)->lock);
1513 /* Re-read under class->lock: PFN is now stable vs migration. */
1514 obj = handle_to_obj(handle);
1515 obj_to_zpdesc(obj, &f_zpdesc);
1516 *zspagep = get_zspage(f_zpdesc);
1517 *objp = obj;
1518}
1519#else
1520static inline void obj_class_get_and_lock(struct zs_pool *pool, unsigned long handle,
1521 unsigned long *objp, struct zspage **zspagep,
1522 struct size_class **classp)
1523 __acquires(&(*classp)->lock)
1524{
1525 struct zpdesc *f_zpdesc;
1526 unsigned long obj;
1527
1528 read_lock(&pool->lock);
1529 obj = handle_to_obj(handle);
1530 obj_to_zpdesc(obj, &f_zpdesc);
1531 *zspagep = get_zspage(f_zpdesc);
1532 *classp = zspage_class(pool, *zspagep);
1533 spin_lock(&(*classp)->lock);
1534 read_unlock(&pool->lock);
1535 *objp = obj;
1536}
1537#endif
1538
1539void zs_free(struct zs_pool *pool, unsigned long handle)
1540{
1541 struct zspage *zspage;
1542 unsigned long obj;
1543 struct size_class *class;
1544 int fullness;
1545 struct zspage *zspage_to_free = NULL;
1546
1547 if (IS_ERR_OR_NULL((void *)handle))
1548 return;
1549
1550 obj_class_get_and_lock(pool, handle, &obj, &zspage, &class);
1551
1552 class_stat_sub(class, ZS_OBJS_INUSE, 1);
1553 obj_free(class->size, obj);
1554
1555 fullness = fix_fullness_group(class, zspage);
1556 if (fullness == ZS_INUSE_RATIO_0) {
1557 if (trylock_zspage(zspage)) {
1558 remove_zspage(class, zspage);
1559 class_stat_sub(class, ZS_OBJS_ALLOCATED,
1560 class->objs_per_zspage);
1561 zspage_to_free = zspage;
1562 } else {
1563 kick_deferred_free(pool);
1564 }
1565 }
1566
1567 spin_unlock(&class->lock);
1568
1569 if (zspage_to_free) {
1570 __free_zspage_lockless(zspage_to_free);
1571 atomic_long_sub(class->pages_per_zspage, &pool->pages_allocated);
1572 }
1573 cache_free_handle(handle);
1574}
1575EXPORT_SYMBOL_GPL(zs_free);
1576
1577static void zs_object_copy(struct size_class *class, unsigned long dst,
1578 unsigned long src)
1579{
1580 struct zpdesc *s_zpdesc, *d_zpdesc;
1581 unsigned int s_objidx, d_objidx;
1582 unsigned long s_off, d_off;
1583 void *s_addr, *d_addr;
1584 int s_size, d_size, size;
1585 int written = 0;
1586
1587 s_size = d_size = class->size;
1588
1589 obj_to_location(src, &s_zpdesc, &s_objidx);
1590 obj_to_location(dst, &d_zpdesc, &d_objidx);
1591
1592 s_off = offset_in_page(class->size * s_objidx);
1593 d_off = offset_in_page(class->size * d_objidx);
1594
1595 if (s_off + class->size > PAGE_SIZE)
1596 s_size = PAGE_SIZE - s_off;
1597
1598 if (d_off + class->size > PAGE_SIZE)
1599 d_size = PAGE_SIZE - d_off;
1600
1601 s_addr = kmap_local_zpdesc(s_zpdesc);
1602 d_addr = kmap_local_zpdesc(d_zpdesc);
1603
1604 while (1) {
1605 size = min(s_size, d_size);
1606 memcpy(d_addr + d_off, s_addr + s_off, size);
1607 written += size;
1608
1609 if (written == class->size)
1610 break;
1611
1612 s_off += size;
1613 s_size -= size;
1614 d_off += size;
1615 d_size -= size;
1616
1617 /*
1618 * Calling kunmap_local(d_addr) is necessary. kunmap_local()
1619 * calls must occurs in reverse order of calls to kmap_local_page().
1620 * So, to call kunmap_local(s_addr) we should first call
1621 * kunmap_local(d_addr). For more details see
1622 * Documentation/mm/highmem.rst.
1623 */
1624 if (s_off >= PAGE_SIZE) {
1625 kunmap_local(d_addr);
1626 kunmap_local(s_addr);
1627 s_zpdesc = get_next_zpdesc(s_zpdesc);
1628 s_addr = kmap_local_zpdesc(s_zpdesc);
1629 d_addr = kmap_local_zpdesc(d_zpdesc);
1630 s_size = class->size - written;
1631 s_off = 0;
1632 }
1633
1634 if (d_off >= PAGE_SIZE) {
1635 kunmap_local(d_addr);
1636 d_zpdesc = get_next_zpdesc(d_zpdesc);
1637 d_addr = kmap_local_zpdesc(d_zpdesc);
1638 d_size = class->size - written;
1639 d_off = 0;
1640 }
1641 }
1642
1643 kunmap_local(d_addr);
1644 kunmap_local(s_addr);
1645}
1646
1647/*
1648 * Find alloced object in zspage from index object and
1649 * return handle.
1650 */
1651static unsigned long find_alloced_obj(struct size_class *class,
1652 struct zpdesc *zpdesc, int *obj_idx)
1653{
1654 unsigned int offset;
1655 int index = *obj_idx;
1656 unsigned long handle = 0;
1657 void *addr = kmap_local_zpdesc(zpdesc);
1658
1659 offset = get_first_obj_offset(zpdesc);
1660 offset += class->size * index;
1661
1662 while (offset < PAGE_SIZE) {
1663 if (obj_allocated(zpdesc, addr + offset, &handle))
1664 break;
1665
1666 offset += class->size;
1667 index++;
1668 }
1669
1670 kunmap_local(addr);
1671
1672 *obj_idx = index;
1673
1674 return handle;
1675}
1676
1677static void migrate_zspage(struct zs_pool *pool, struct zspage *src_zspage,
1678 struct zspage *dst_zspage)
1679{
1680 unsigned long used_obj, free_obj;
1681 unsigned long handle;
1682 int obj_idx = 0;
1683 struct zpdesc *s_zpdesc = get_first_zpdesc(src_zspage);
1684 struct size_class *class = pool->size_class[src_zspage->class];
1685
1686 while (1) {
1687 handle = find_alloced_obj(class, s_zpdesc, &obj_idx);
1688 if (!handle) {
1689 s_zpdesc = get_next_zpdesc(s_zpdesc);
1690 if (!s_zpdesc)
1691 break;
1692 obj_idx = 0;
1693 continue;
1694 }
1695
1696 used_obj = handle_to_obj(handle);
1697 free_obj = obj_malloc(pool, dst_zspage, handle);
1698 zs_object_copy(class, free_obj, used_obj);
1699 obj_idx++;
1700 obj_free(class->size, used_obj);
1701
1702 /* Stop if there is no more space */
1703 if (zspage_full(class, dst_zspage))
1704 break;
1705
1706 /* Stop if there are no more objects to migrate */
1707 if (zspage_empty(src_zspage))
1708 break;
1709 }
1710}
1711
1712static struct zspage *isolate_src_zspage(struct size_class *class)
1713{
1714 struct zspage *zspage;
1715 int fg;
1716
1717 for (fg = ZS_INUSE_RATIO_10; fg <= ZS_INUSE_RATIO_99; fg++) {
1718 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1719 struct zspage, list);
1720 if (zspage) {
1721 remove_zspage(class, zspage);
1722 return zspage;
1723 }
1724 }
1725
1726 return zspage;
1727}
1728
1729static struct zspage *isolate_dst_zspage(struct size_class *class)
1730{
1731 struct zspage *zspage;
1732 int fg;
1733
1734 for (fg = ZS_INUSE_RATIO_99; fg >= ZS_INUSE_RATIO_10; fg--) {
1735 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1736 struct zspage, list);
1737 if (zspage) {
1738 remove_zspage(class, zspage);
1739 return zspage;
1740 }
1741 }
1742
1743 return zspage;
1744}
1745
1746/*
1747 * putback_zspage - add @zspage into right class's fullness list
1748 * @class: destination class
1749 * @zspage: target page
1750 *
1751 * Return @zspage's fullness status
1752 */
1753static int putback_zspage(struct size_class *class, struct zspage *zspage)
1754{
1755 int fullness;
1756
1757 fullness = get_fullness_group(class, zspage);
1758 insert_zspage(class, zspage, fullness);
1759
1760 return fullness;
1761}
1762
1763#ifdef CONFIG_COMPACTION
1764/*
1765 * To prevent zspage destroy during migration, zspage freeing should
1766 * hold locks of all pages in the zspage.
1767 */
1768static void lock_zspage(struct zspage *zspage)
1769{
1770 struct zpdesc *curr_zpdesc, *zpdesc;
1771
1772 /*
1773 * Pages we haven't locked yet can be migrated off the list while we're
1774 * trying to lock them, so we need to be careful and only attempt to
1775 * lock each page under zspage_read_lock(). Otherwise, the page we lock
1776 * may no longer belong to the zspage. This means that we may wait for
1777 * the wrong page to unlock, so we must take a reference to the page
1778 * prior to waiting for it to unlock outside zspage_read_lock().
1779 */
1780 while (1) {
1781 zspage_read_lock(zspage);
1782 zpdesc = get_first_zpdesc(zspage);
1783 if (zpdesc_trylock(zpdesc))
1784 break;
1785 zpdesc_get(zpdesc);
1786 zspage_read_unlock(zspage);
1787 zpdesc_wait_locked(zpdesc);
1788 zpdesc_put(zpdesc);
1789 }
1790
1791 curr_zpdesc = zpdesc;
1792 while ((zpdesc = get_next_zpdesc(curr_zpdesc))) {
1793 if (zpdesc_trylock(zpdesc)) {
1794 curr_zpdesc = zpdesc;
1795 } else {
1796 zpdesc_get(zpdesc);
1797 zspage_read_unlock(zspage);
1798 zpdesc_wait_locked(zpdesc);
1799 zpdesc_put(zpdesc);
1800 zspage_read_lock(zspage);
1801 }
1802 }
1803 zspage_read_unlock(zspage);
1804}
1805
1806static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1807 struct zpdesc *newzpdesc, struct zpdesc *oldzpdesc)
1808{
1809 struct zpdesc *zpdesc;
1810 struct zpdesc *zpdescs[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1811 unsigned int first_obj_offset;
1812 int idx = 0;
1813
1814 zpdesc = get_first_zpdesc(zspage);
1815 do {
1816 if (zpdesc == oldzpdesc)
1817 zpdescs[idx] = newzpdesc;
1818 else
1819 zpdescs[idx] = zpdesc;
1820 idx++;
1821 } while ((zpdesc = get_next_zpdesc(zpdesc)) != NULL);
1822
1823 create_page_chain(class, zspage, zpdescs);
1824 first_obj_offset = get_first_obj_offset(oldzpdesc);
1825 set_first_obj_offset(newzpdesc, first_obj_offset);
1826 if (unlikely(ZsHugePage(zspage)))
1827 newzpdesc->handle = oldzpdesc->handle;
1828 __zpdesc_set_movable(newzpdesc);
1829}
1830
1831static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1832{
1833 /*
1834 * Page is locked so zspage can't be destroyed concurrently
1835 * (see free_zspage()). But if the page was already destroyed
1836 * (see reset_zpdesc()), refuse isolation here.
1837 */
1838 return page_zpdesc(page)->zspage;
1839}
1840
1841static int zs_page_migrate(struct page *newpage, struct page *page,
1842 enum migrate_mode mode)
1843{
1844 struct zs_pool *pool;
1845 struct size_class *class;
1846 struct zspage *zspage;
1847 struct zpdesc *dummy;
1848 struct zpdesc *newzpdesc = page_zpdesc(newpage);
1849 struct zpdesc *zpdesc = page_zpdesc(page);
1850 void *s_addr, *d_addr, *addr;
1851 unsigned int offset;
1852 unsigned long handle;
1853 unsigned long old_obj, new_obj;
1854 unsigned int obj_idx;
1855
1856 /*
1857 * TODO: nothing prevents a zspage from getting destroyed while
1858 * it is isolated for migration, as the page lock is temporarily
1859 * dropped after zs_page_isolate() succeeded: we should rework that
1860 * and defer destroying such pages once they are un-isolated (putback)
1861 * instead.
1862 */
1863 if (!zpdesc->zspage)
1864 return 0;
1865
1866 /* The page is locked, so this pointer must remain valid */
1867 zspage = get_zspage(zpdesc);
1868 pool = zspage->pool;
1869
1870 /*
1871 * The pool migrate_lock protects against races between zpage migration
1872 * and zs_free(), but only when ZS_OBJ_CLASS_BITS does not apply.
1873 */
1874 write_lock(&pool->lock);
1875 class = zspage_class(pool, zspage);
1876
1877 /*
1878 * the class lock protects zpage alloc/free in the zspage.
1879 */
1880 spin_lock(&class->lock);
1881 /* the zspage write_lock protects zpage access via zs_obj_read/write() */
1882 if (!zspage_write_trylock(zspage)) {
1883 spin_unlock(&class->lock);
1884 write_unlock(&pool->lock);
1885 /*
1886 * Return -EBUSY but not -EAGAIN: the zspage's reader-lock
1887 * owner may hold the lock for an unbounded duration due to a
1888 * slow decompression or reader-lock owner preemption.
1889 * Since migration retries are bounded by
1890 * NR_MAX_MIGRATE_PAGES_RETRY and performed with virtually no
1891 * delay between attempts, there is no guarantee the lock will
1892 * be released in time for a retry to succeed.
1893 * -EAGAIN implies "try again soon", which does not hold here.
1894 * -EBUSY more accurately conveys "resource is occupied,
1895 * migration cannot proceed".
1896 */
1897 return -EBUSY;
1898 }
1899
1900 /* We're committed, tell the world that this is a Zsmalloc page. */
1901 __zpdesc_set_zsmalloc(newzpdesc);
1902
1903 offset = get_first_obj_offset(zpdesc);
1904 s_addr = kmap_local_zpdesc(zpdesc);
1905
1906 /*
1907 * Here, any user cannot access all objects in the zspage so let's move.
1908 */
1909 d_addr = kmap_local_zpdesc(newzpdesc);
1910 copy_page(d_addr, s_addr);
1911 kmsan_copy_page_meta(zpdesc_page(newzpdesc), zpdesc_page(zpdesc));
1912 kunmap_local(d_addr);
1913
1914 for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE;
1915 addr += class->size) {
1916 if (obj_allocated(zpdesc, addr, &handle)) {
1917
1918 old_obj = handle_to_obj(handle);
1919 obj_to_location(old_obj, &dummy, &obj_idx);
1920 new_obj = location_to_obj(newzpdesc, obj_idx,
1921 obj_to_class_idx(old_obj));
1922 record_obj(handle, new_obj);
1923 }
1924 }
1925 kunmap_local(s_addr);
1926
1927 replace_sub_page(class, zspage, newzpdesc, zpdesc);
1928 /*
1929 * Since we complete the data copy and set up new zspage structure,
1930 * it's okay to release migration_lock.
1931 */
1932 zspage_write_unlock(zspage);
1933 spin_unlock(&class->lock);
1934 write_unlock(&pool->lock);
1935
1936 zpdesc_get(newzpdesc);
1937 if (zpdesc_zone(newzpdesc) != zpdesc_zone(zpdesc)) {
1938 zpdesc_dec_zone_page_state(zpdesc);
1939 zpdesc_inc_zone_page_state(newzpdesc);
1940 }
1941
1942 reset_zpdesc(zpdesc);
1943 zpdesc_put(zpdesc);
1944
1945 return 0;
1946}
1947
1948static void zs_page_putback(struct page *page)
1949{
1950}
1951
1952const struct movable_operations zsmalloc_mops = {
1953 .isolate_page = zs_page_isolate,
1954 .migrate_page = zs_page_migrate,
1955 .putback_page = zs_page_putback,
1956};
1957
1958/*
1959 * Caller should hold page_lock of all pages in the zspage
1960 * In here, we cannot use zspage meta data.
1961 */
1962static void async_free_zspage(struct work_struct *work)
1963{
1964 int i;
1965 struct size_class *class;
1966 struct zspage *zspage, *tmp;
1967 LIST_HEAD(free_pages);
1968 struct zs_pool *pool = container_of(work, struct zs_pool,
1969 free_work);
1970
1971 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
1972 class = pool->size_class[i];
1973 if (class->index != i)
1974 continue;
1975
1976 spin_lock(&class->lock);
1977 list_splice_init(&class->fullness_list[ZS_INUSE_RATIO_0],
1978 &free_pages);
1979 spin_unlock(&class->lock);
1980 }
1981
1982 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
1983 list_del(&zspage->list);
1984 lock_zspage(zspage);
1985
1986 class = zspage_class(pool, zspage);
1987 spin_lock(&class->lock);
1988 class_stat_sub(class, ZS_INUSE_RATIO_0, 1);
1989 __free_zspage(pool, class, zspage);
1990 spin_unlock(&class->lock);
1991 }
1992};
1993
1994static void kick_deferred_free(struct zs_pool *pool)
1995{
1996 schedule_work(&pool->free_work);
1997}
1998
1999static void zs_flush_migration(struct zs_pool *pool)
2000{
2001 flush_work(&pool->free_work);
2002}
2003
2004static void init_deferred_free(struct zs_pool *pool)
2005{
2006 INIT_WORK(&pool->free_work, async_free_zspage);
2007}
2008
2009static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2010{
2011 struct zpdesc *zpdesc = get_first_zpdesc(zspage);
2012
2013 do {
2014 WARN_ON(!zpdesc_trylock(zpdesc));
2015 __zpdesc_set_movable(zpdesc);
2016 zpdesc_unlock(zpdesc);
2017 } while ((zpdesc = get_next_zpdesc(zpdesc)) != NULL);
2018}
2019#else
2020static inline void zs_flush_migration(struct zs_pool *pool) { }
2021#endif
2022
2023/*
2024 *
2025 * Based on the number of unused allocated objects calculate
2026 * and return the number of pages that we can free.
2027 */
2028static unsigned long zs_can_compact(struct size_class *class)
2029{
2030 unsigned long obj_wasted;
2031 unsigned long obj_allocated = class_stat_read(class, ZS_OBJS_ALLOCATED);
2032 unsigned long obj_used = class_stat_read(class, ZS_OBJS_INUSE);
2033
2034 if (obj_allocated <= obj_used)
2035 return 0;
2036
2037 obj_wasted = obj_allocated - obj_used;
2038 obj_wasted /= class->objs_per_zspage;
2039
2040 return obj_wasted * class->pages_per_zspage;
2041}
2042
2043static unsigned long __zs_compact(struct zs_pool *pool,
2044 struct size_class *class)
2045{
2046 struct zspage *src_zspage = NULL;
2047 struct zspage *dst_zspage = NULL;
2048 unsigned long pages_freed = 0;
2049
2050 /*
2051 * Protect against races between zpage migration and zs_free()
2052 * (only when ZS_OBJ_CLASS_BITS does not apply), as well as
2053 * zpage allocation and free.
2054 */
2055 write_lock(&pool->lock);
2056 spin_lock(&class->lock);
2057 while (zs_can_compact(class)) {
2058 int fg;
2059
2060 if (!dst_zspage) {
2061 dst_zspage = isolate_dst_zspage(class);
2062 if (!dst_zspage)
2063 break;
2064 }
2065
2066 src_zspage = isolate_src_zspage(class);
2067 if (!src_zspage)
2068 break;
2069
2070 if (!zspage_write_trylock(src_zspage))
2071 break;
2072
2073 migrate_zspage(pool, src_zspage, dst_zspage);
2074 zspage_write_unlock(src_zspage);
2075
2076 fg = putback_zspage(class, src_zspage);
2077 if (fg == ZS_INUSE_RATIO_0) {
2078 free_zspage(pool, class, src_zspage);
2079 pages_freed += class->pages_per_zspage;
2080 }
2081 src_zspage = NULL;
2082
2083 if (get_fullness_group(class, dst_zspage) == ZS_INUSE_RATIO_100
2084 || rwlock_is_contended(&pool->lock)) {
2085 putback_zspage(class, dst_zspage);
2086 dst_zspage = NULL;
2087
2088 spin_unlock(&class->lock);
2089 write_unlock(&pool->lock);
2090 cond_resched();
2091 write_lock(&pool->lock);
2092 spin_lock(&class->lock);
2093 }
2094 }
2095
2096 if (src_zspage)
2097 putback_zspage(class, src_zspage);
2098
2099 if (dst_zspage)
2100 putback_zspage(class, dst_zspage);
2101
2102 spin_unlock(&class->lock);
2103 write_unlock(&pool->lock);
2104
2105 return pages_freed;
2106}
2107
2108unsigned long zs_compact(struct zs_pool *pool)
2109{
2110 int i;
2111 struct size_class *class;
2112 unsigned long pages_freed = 0;
2113
2114 /*
2115 * Pool compaction is performed under pool->lock so it is basically
2116 * single-threaded. Having more than one thread in __zs_compact()
2117 * will increase pool->lock contention, which will impact other
2118 * zsmalloc operations that need pool->lock.
2119 */
2120 if (atomic_xchg(&pool->compaction_in_progress, 1))
2121 return 0;
2122
2123 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2124 class = pool->size_class[i];
2125 if (class->index != i)
2126 continue;
2127 pages_freed += __zs_compact(pool, class);
2128 }
2129 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2130 atomic_set(&pool->compaction_in_progress, 0);
2131
2132 return pages_freed;
2133}
2134EXPORT_SYMBOL_GPL(zs_compact);
2135
2136void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2137{
2138 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2139}
2140EXPORT_SYMBOL_GPL(zs_pool_stats);
2141
2142static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2143 struct shrink_control *sc)
2144{
2145 unsigned long pages_freed;
2146 struct zs_pool *pool = shrinker->private_data;
2147
2148 /*
2149 * Compact classes and calculate compaction delta.
2150 * Can run concurrently with a manually triggered
2151 * (by user) compaction.
2152 */
2153 pages_freed = zs_compact(pool);
2154
2155 return pages_freed ? pages_freed : SHRINK_STOP;
2156}
2157
2158static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2159 struct shrink_control *sc)
2160{
2161 int i;
2162 struct size_class *class;
2163 unsigned long pages_to_free = 0;
2164 struct zs_pool *pool = shrinker->private_data;
2165
2166 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2167 class = pool->size_class[i];
2168 if (class->index != i)
2169 continue;
2170
2171 pages_to_free += zs_can_compact(class);
2172 }
2173
2174 return pages_to_free;
2175}
2176
2177static void zs_unregister_shrinker(struct zs_pool *pool)
2178{
2179 shrinker_free(pool->shrinker);
2180}
2181
2182static int zs_register_shrinker(struct zs_pool *pool)
2183{
2184 pool->shrinker = shrinker_alloc(0, "mm-zspool:%s", pool->name);
2185 if (!pool->shrinker)
2186 return -ENOMEM;
2187
2188 pool->shrinker->scan_objects = zs_shrinker_scan;
2189 pool->shrinker->count_objects = zs_shrinker_count;
2190 pool->shrinker->batch = 0;
2191 pool->shrinker->private_data = pool;
2192
2193 shrinker_register(pool->shrinker);
2194
2195 return 0;
2196}
2197
2198static int calculate_zspage_chain_size(int class_size)
2199{
2200 int i, min_waste = INT_MAX;
2201 int chain_size = 1;
2202
2203 if (is_power_of_2(class_size))
2204 return chain_size;
2205
2206 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
2207 int waste;
2208
2209 waste = (i * PAGE_SIZE) % class_size;
2210 if (waste < min_waste) {
2211 min_waste = waste;
2212 chain_size = i;
2213 }
2214 }
2215
2216 return chain_size;
2217}
2218
2219/**
2220 * zs_create_pool - Creates an allocation pool to work from.
2221 * @name: pool name to be created
2222 *
2223 * This function must be called before anything when using
2224 * the zsmalloc allocator.
2225 *
2226 * On success, a pointer to the newly created pool is returned,
2227 * otherwise NULL.
2228 */
2229struct zs_pool *zs_create_pool(const char *name)
2230{
2231 int i;
2232 struct zs_pool *pool;
2233 struct size_class *prev_class = NULL;
2234
2235 pool = kzalloc_obj(*pool);
2236 if (!pool)
2237 return NULL;
2238
2239 init_deferred_free(pool);
2240 rwlock_init(&pool->lock);
2241 atomic_set(&pool->compaction_in_progress, 0);
2242
2243 pool->name = kstrdup(name, GFP_KERNEL);
2244 if (!pool->name)
2245 goto err;
2246
2247 /*
2248 * Iterate reversely, because, size of size_class that we want to use
2249 * for merging should be larger or equal to current size.
2250 */
2251 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2252 int size;
2253 int pages_per_zspage;
2254 int objs_per_zspage;
2255 struct size_class *class;
2256 int fullness;
2257
2258 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2259 if (size > ZS_MAX_ALLOC_SIZE)
2260 size = ZS_MAX_ALLOC_SIZE;
2261 pages_per_zspage = calculate_zspage_chain_size(size);
2262 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2263
2264 /*
2265 * We iterate from biggest down to smallest classes,
2266 * so huge_class_size holds the size of the first huge
2267 * class. Any object bigger than or equal to that will
2268 * endup in the huge class.
2269 */
2270 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2271 !huge_class_size) {
2272 huge_class_size = size;
2273 /*
2274 * The object uses ZS_HANDLE_SIZE bytes to store the
2275 * handle. We need to subtract it, because zs_malloc()
2276 * unconditionally adds handle size before it performs
2277 * size class search - so object may be smaller than
2278 * huge class size, yet it still can end up in the huge
2279 * class because it grows by ZS_HANDLE_SIZE extra bytes
2280 * right before class lookup.
2281 */
2282 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2283 }
2284
2285 /*
2286 * size_class is used for normal zsmalloc operation such
2287 * as alloc/free for that size. Although it is natural that we
2288 * have one size_class for each size, there is a chance that we
2289 * can get more memory utilization if we use one size_class for
2290 * many different sizes whose size_class have same
2291 * characteristics. So, we makes size_class point to
2292 * previous size_class if possible.
2293 */
2294 if (prev_class) {
2295 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2296 pool->size_class[i] = prev_class;
2297 continue;
2298 }
2299 }
2300
2301 class = kzalloc_obj(struct size_class);
2302 if (!class)
2303 goto err;
2304
2305 class->size = size;
2306 class->index = i;
2307 class->pages_per_zspage = pages_per_zspage;
2308 class->objs_per_zspage = objs_per_zspage;
2309 spin_lock_init(&class->lock);
2310 pool->size_class[i] = class;
2311
2312 fullness = ZS_INUSE_RATIO_0;
2313 while (fullness < NR_FULLNESS_GROUPS) {
2314 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2315 fullness++;
2316 }
2317
2318 prev_class = class;
2319 }
2320
2321 /* debug only, don't abort if it fails */
2322 zs_pool_stat_create(pool, name);
2323
2324 /*
2325 * Not critical since shrinker is only used to trigger internal
2326 * defragmentation of the pool which is pretty optional thing. If
2327 * registration fails we still can use the pool normally and user can
2328 * trigger compaction manually. Thus, ignore return code.
2329 */
2330 zs_register_shrinker(pool);
2331
2332 return pool;
2333
2334err:
2335 zs_destroy_pool(pool);
2336 return NULL;
2337}
2338EXPORT_SYMBOL_GPL(zs_create_pool);
2339
2340void zs_destroy_pool(struct zs_pool *pool)
2341{
2342 int i;
2343
2344 zs_unregister_shrinker(pool);
2345 zs_flush_migration(pool);
2346 zs_pool_stat_destroy(pool);
2347
2348 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2349 int fg;
2350 struct size_class *class = pool->size_class[i];
2351
2352 if (!class)
2353 continue;
2354
2355 if (class->index != i)
2356 continue;
2357
2358 for (fg = ZS_INUSE_RATIO_0; fg < NR_FULLNESS_GROUPS; fg++) {
2359 if (list_empty(&class->fullness_list[fg]))
2360 continue;
2361
2362 pr_err("Class-%d fullness group %d is not empty\n",
2363 class->size, fg);
2364 }
2365 kfree(class);
2366 }
2367
2368 kfree(pool->name);
2369 kfree(pool);
2370}
2371EXPORT_SYMBOL_GPL(zs_destroy_pool);
2372
2373static void zs_destroy_caches(void)
2374{
2375 kmem_cache_destroy(handle_cachep);
2376 handle_cachep = NULL;
2377 kmem_cache_destroy(zspage_cachep);
2378 zspage_cachep = NULL;
2379}
2380
2381static int __init zs_init_caches(void)
2382{
2383 handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
2384 0, 0, NULL);
2385 zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
2386 0, 0, NULL);
2387
2388 if (!handle_cachep || !zspage_cachep) {
2389 zs_destroy_caches();
2390 return -ENOMEM;
2391 }
2392 return 0;
2393}
2394
2395static int __init zs_init(void)
2396{
2397 int rc;
2398
2399 rc = zs_init_caches();
2400 if (rc)
2401 return rc;
2402
2403#ifdef CONFIG_COMPACTION
2404 rc = set_movable_ops(&zsmalloc_mops, PGTY_zsmalloc);
2405 if (rc) {
2406 zs_destroy_caches();
2407 return rc;
2408 }
2409#endif
2410 zs_stat_init();
2411 return 0;
2412}
2413
2414static void __exit zs_exit(void)
2415{
2416#ifdef CONFIG_COMPACTION
2417 set_movable_ops(NULL, PGTY_zsmalloc);
2418#endif
2419 zs_stat_exit();
2420 zs_destroy_caches();
2421}
2422
2423module_init(zs_init);
2424module_exit(zs_exit);
2425
2426MODULE_LICENSE("Dual BSD/GPL");
2427MODULE_AUTHOR("Nitin Gupta <[email protected]>");
2428MODULE_DESCRIPTION("zsmalloc memory allocator");