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: PG_zsmalloc, lower 16 bit locate the first object
24 * offset in a subpage of a zspage
25 *
26 * Usage of struct page flags:
27 * PG_private: identifies the first component page
28 * PG_owner_priv_1: identifies the huge component page
29 *
30 */
31
32 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33
34 /*
35 * lock ordering:
36 * page_lock
37 * pool->migrate_lock
38 * class->lock
39 * zspage->lock
40 */
41
42 #include <linux/module.h>
43 #include <linux/kernel.h>
44 #include <linux/sched.h>
45 #include <linux/bitops.h>
46 #include <linux/errno.h>
47 #include <linux/highmem.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/pgtable.h>
51 #include <asm/tlbflush.h>
52 #include <linux/cpumask.h>
53 #include <linux/cpu.h>
54 #include <linux/vmalloc.h>
55 #include <linux/preempt.h>
56 #include <linux/spinlock.h>
57 #include <linux/shrinker.h>
58 #include <linux/types.h>
59 #include <linux/debugfs.h>
60 #include <linux/zsmalloc.h>
61 #include <linux/zpool.h>
62 #include <linux/migrate.h>
63 #include <linux/wait.h>
64 #include <linux/pagemap.h>
65 #include <linux/fs.h>
66 #include <linux/local_lock.h>
67
68 #define ZSPAGE_MAGIC 0x58
69
70 /*
71 * This must be power of 2 and greater than or equal to sizeof(link_free).
72 * These two conditions ensure that any 'struct link_free' itself doesn't
73 * span more than 1 page which avoids complex case of mapping 2 pages simply
74 * to restore link_free pointer values.
75 */
76 #define ZS_ALIGN 8
77
78 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
79
80 /*
81 * Object location (<PFN>, <obj_idx>) is encoded as
82 * a single (unsigned long) handle value.
83 *
84 * Note that object index <obj_idx> starts from 0.
85 *
86 * This is made more complicated by various memory models and PAE.
87 */
88
89 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
90 #ifdef MAX_PHYSMEM_BITS
91 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
92 #else
93 /*
94 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
95 * be PAGE_SHIFT
96 */
97 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
98 #endif
99 #endif
100
101 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
102
103 /*
104 * Head in allocated object should have OBJ_ALLOCATED_TAG
105 * to identify the object was allocated or not.
106 * It's okay to add the status bit in the least bit because
107 * header keeps handle which is 4byte-aligned address so we
108 * have room for two bit at least.
109 */
110 #define OBJ_ALLOCATED_TAG 1
111
112 #define OBJ_TAG_BITS 1
113 #define OBJ_TAG_MASK OBJ_ALLOCATED_TAG
114
115 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
116 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
117
118 #define HUGE_BITS 1
119 #define FULLNESS_BITS 4
120 #define CLASS_BITS 8
121 #define MAGIC_VAL_BITS 8
122
123 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(CONFIG_ZSMALLOC_CHAIN_SIZE, UL))
124
125 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
126 #define ZS_MIN_ALLOC_SIZE \
127 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
128 /* each chunk includes extra space to keep handle */
129 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
130
131 /*
132 * On systems with 4K page size, this gives 255 size classes! There is a
133 * trader-off here:
134 * - Large number of size classes is potentially wasteful as free page are
135 * spread across these classes
136 * - Small number of size classes causes large internal fragmentation
137 * - Probably its better to use specific size classes (empirically
138 * determined). NOTE: all those class sizes must be set as multiple of
139 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
140 *
141 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
142 * (reason above)
143 */
144 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
145 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
146 ZS_SIZE_CLASS_DELTA) + 1)
147
148 /*
149 * Pages are distinguished by the ratio of used memory (that is the ratio
150 * of ->inuse objects to all objects that page can store). For example,
151 * INUSE_RATIO_10 means that the ratio of used objects is > 0% and <= 10%.
152 *
153 * The number of fullness groups is not random. It allows us to keep
154 * difference between the least busy page in the group (minimum permitted
155 * number of ->inuse objects) and the most busy page (maximum permitted
156 * number of ->inuse objects) at a reasonable value.
157 */
158 enum fullness_group {
159 ZS_INUSE_RATIO_0,
160 ZS_INUSE_RATIO_10,
161 /* NOTE: 8 more fullness groups here */
162 ZS_INUSE_RATIO_99 = 10,
163 ZS_INUSE_RATIO_100,
164 NR_FULLNESS_GROUPS,
165 };
166
167 enum class_stat_type {
168 /* NOTE: stats for 12 fullness groups here: from inuse 0 to 100 */
169 ZS_OBJS_ALLOCATED = NR_FULLNESS_GROUPS,
170 ZS_OBJS_INUSE,
171 NR_CLASS_STAT_TYPES,
172 };
173
174 struct zs_size_stat {
175 unsigned long objs[NR_CLASS_STAT_TYPES];
176 };
177
178 #ifdef CONFIG_ZSMALLOC_STAT
179 static struct dentry *zs_stat_root;
180 #endif
181
182 static size_t huge_class_size;
183
184 struct size_class {
185 spinlock_t lock;
186 struct list_head fullness_list[NR_FULLNESS_GROUPS];
187 /*
188 * Size of objects stored in this class. Must be multiple
189 * of ZS_ALIGN.
190 */
191 int size;
192 int objs_per_zspage;
193 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
194 int pages_per_zspage;
195
196 unsigned int index;
197 struct zs_size_stat stats;
198 };
199
200 /*
201 * Placed within free objects to form a singly linked list.
202 * For every zspage, zspage->freeobj gives head of this list.
203 *
204 * This must be power of 2 and less than or equal to ZS_ALIGN
205 */
206 struct link_free {
207 union {
208 /*
209 * Free object index;
210 * It's valid for non-allocated object
211 */
212 unsigned long next;
213 /*
214 * Handle of allocated object.
215 */
216 unsigned long handle;
217 };
218 };
219
220 struct zs_pool {
221 const char *name;
222
223 struct size_class *size_class[ZS_SIZE_CLASSES];
224 struct kmem_cache *handle_cachep;
225 struct kmem_cache *zspage_cachep;
226
227 atomic_long_t pages_allocated;
228
229 struct zs_pool_stats stats;
230
231 /* Compact classes */
232 struct shrinker *shrinker;
233
234 #ifdef CONFIG_ZSMALLOC_STAT
235 struct dentry *stat_dentry;
236 #endif
237 #ifdef CONFIG_COMPACTION
238 struct work_struct free_work;
239 #endif
240 /* protect page/zspage migration */
241 rwlock_t migrate_lock;
242 atomic_t compaction_in_progress;
243 };
244
245 struct zspage {
246 struct {
247 unsigned int huge:HUGE_BITS;
248 unsigned int fullness:FULLNESS_BITS;
249 unsigned int class:CLASS_BITS + 1;
250 unsigned int magic:MAGIC_VAL_BITS;
251 };
252 unsigned int inuse;
253 unsigned int freeobj;
254 struct page *first_page;
255 struct list_head list; /* fullness list */
256 struct zs_pool *pool;
257 rwlock_t lock;
258 };
259
260 struct mapping_area {
261 local_lock_t lock;
262 char *vm_buf; /* copy buffer for objects that span pages */
263 char *vm_addr; /* address of kmap_atomic()'ed pages */
264 enum zs_mapmode vm_mm; /* mapping mode */
265 };
266
267 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
268 static void SetZsHugePage(struct zspage *zspage)
269 {
270 zspage->huge = 1;
271 }
272
273 static bool ZsHugePage(struct zspage *zspage)
274 {
275 return zspage->huge;
276 }
277
278 static void migrate_lock_init(struct zspage *zspage);
279 static void migrate_read_lock(struct zspage *zspage);
280 static void migrate_read_unlock(struct zspage *zspage);
281 static void migrate_write_lock(struct zspage *zspage);
282 static void migrate_write_unlock(struct zspage *zspage);
283
284 #ifdef CONFIG_COMPACTION
285 static void kick_deferred_free(struct zs_pool *pool);
286 static void init_deferred_free(struct zs_pool *pool);
287 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
288 #else
289 static void kick_deferred_free(struct zs_pool *pool) {}
290 static void init_deferred_free(struct zs_pool *pool) {}
291 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
292 #endif
293
294 static int create_cache(struct zs_pool *pool)
295 {
296 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
297 0, 0, NULL);
298 if (!pool->handle_cachep)
299 return 1;
300
301 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
302 0, 0, NULL);
303 if (!pool->zspage_cachep) {
304 kmem_cache_destroy(pool->handle_cachep);
305 pool->handle_cachep = NULL;
306 return 1;
307 }
308
309 return 0;
310 }
311
312 static void destroy_cache(struct zs_pool *pool)
313 {
314 kmem_cache_destroy(pool->handle_cachep);
315 kmem_cache_destroy(pool->zspage_cachep);
316 }
317
318 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
319 {
320 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
321 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
322 }
323
324 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
325 {
326 kmem_cache_free(pool->handle_cachep, (void *)handle);
327 }
328
329 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
330 {
331 return kmem_cache_zalloc(pool->zspage_cachep,
332 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
333 }
334
335 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
336 {
337 kmem_cache_free(pool->zspage_cachep, zspage);
338 }
339
340 /* class->lock(which owns the handle) synchronizes races */
341 static void record_obj(unsigned long handle, unsigned long obj)
342 {
343 *(unsigned long *)handle = obj;
344 }
345
346 /* zpool driver */
347
348 #ifdef CONFIG_ZPOOL
349
350 static void *zs_zpool_create(const char *name, gfp_t gfp)
351 {
352 /*
353 * Ignore global gfp flags: zs_malloc() may be invoked from
354 * different contexts and its caller must provide a valid
355 * gfp mask.
356 */
357 return zs_create_pool(name);
358 }
359
360 static void zs_zpool_destroy(void *pool)
361 {
362 zs_destroy_pool(pool);
363 }
364
365 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
366 unsigned long *handle)
367 {
368 *handle = zs_malloc(pool, size, gfp);
369
370 if (IS_ERR_VALUE(*handle))
371 return PTR_ERR((void *)*handle);
372 return 0;
373 }
374 static void zs_zpool_free(void *pool, unsigned long handle)
375 {
376 zs_free(pool, handle);
377 }
378
379 static void *zs_zpool_map(void *pool, unsigned long handle,
380 enum zpool_mapmode mm)
381 {
382 enum zs_mapmode zs_mm;
383
384 switch (mm) {
385 case ZPOOL_MM_RO:
386 zs_mm = ZS_MM_RO;
387 break;
388 case ZPOOL_MM_WO:
389 zs_mm = ZS_MM_WO;
390 break;
391 case ZPOOL_MM_RW:
392 default:
393 zs_mm = ZS_MM_RW;
394 break;
395 }
396
397 return zs_map_object(pool, handle, zs_mm);
398 }
399 static void zs_zpool_unmap(void *pool, unsigned long handle)
400 {
401 zs_unmap_object(pool, handle);
402 }
403
404 static u64 zs_zpool_total_pages(void *pool)
405 {
406 return zs_get_total_pages(pool);
407 }
408
409 static struct zpool_driver zs_zpool_driver = {
410 .type = "zsmalloc",
411 .owner = THIS_MODULE,
412 .create = zs_zpool_create,
413 .destroy = zs_zpool_destroy,
414 .malloc_support_movable = true,
415 .malloc = zs_zpool_malloc,
416 .free = zs_zpool_free,
417 .map = zs_zpool_map,
418 .unmap = zs_zpool_unmap,
419 .total_pages = zs_zpool_total_pages,
420 };
421
422 MODULE_ALIAS("zpool-zsmalloc");
423 #endif /* CONFIG_ZPOOL */
424
425 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
426 static DEFINE_PER_CPU(struct mapping_area, zs_map_area) = {
427 .lock = INIT_LOCAL_LOCK(lock),
428 };
429
430 static __maybe_unused int is_first_page(struct page *page)
431 {
432 return PagePrivate(page);
433 }
434
435 /* Protected by class->lock */
436 static inline int get_zspage_inuse(struct zspage *zspage)
437 {
438 return zspage->inuse;
439 }
440
441
442 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
443 {
444 zspage->inuse += val;
445 }
446
447 static inline struct page *get_first_page(struct zspage *zspage)
448 {
449 struct page *first_page = zspage->first_page;
450
451 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
452 return first_page;
453 }
454
455 #define FIRST_OBJ_PAGE_TYPE_MASK 0xffff
456
457 static inline void reset_first_obj_offset(struct page *page)
458 {
459 VM_WARN_ON_ONCE(!PageZsmalloc(page));
460 page->page_type |= FIRST_OBJ_PAGE_TYPE_MASK;
461 }
462
463 static inline unsigned int get_first_obj_offset(struct page *page)
464 {
465 VM_WARN_ON_ONCE(!PageZsmalloc(page));
466 return page->page_type & FIRST_OBJ_PAGE_TYPE_MASK;
467 }
468
469 static inline void set_first_obj_offset(struct page *page, unsigned int offset)
470 {
471 /* With 16 bit available, we can support offsets into 64 KiB pages. */
472 BUILD_BUG_ON(PAGE_SIZE > SZ_64K);
473 VM_WARN_ON_ONCE(!PageZsmalloc(page));
474 VM_WARN_ON_ONCE(offset & ~FIRST_OBJ_PAGE_TYPE_MASK);
475 page->page_type &= ~FIRST_OBJ_PAGE_TYPE_MASK;
476 page->page_type |= offset & FIRST_OBJ_PAGE_TYPE_MASK;
477 }
478
479 static inline unsigned int get_freeobj(struct zspage *zspage)
480 {
481 return zspage->freeobj;
482 }
483
484 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
485 {
486 zspage->freeobj = obj;
487 }
488
489 static struct size_class *zspage_class(struct zs_pool *pool,
490 struct zspage *zspage)
491 {
492 return pool->size_class[zspage->class];
493 }
494
495 /*
496 * zsmalloc divides the pool into various size classes where each
497 * class maintains a list of zspages where each zspage is divided
498 * into equal sized chunks. Each allocation falls into one of these
499 * classes depending on its size. This function returns index of the
500 * size class which has chunk size big enough to hold the given size.
501 */
502 static int get_size_class_index(int size)
503 {
504 int idx = 0;
505
506 if (likely(size > ZS_MIN_ALLOC_SIZE))
507 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
508 ZS_SIZE_CLASS_DELTA);
509
510 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
511 }
512
513 static inline void class_stat_add(struct size_class *class, int type,
514 unsigned long cnt)
515 {
516 class->stats.objs[type] += cnt;
517 }
518
519 static inline void class_stat_sub(struct size_class *class, int type,
520 unsigned long cnt)
521 {
522 class->stats.objs[type] -= cnt;
523 }
524
525 static inline unsigned long class_stat_read(struct size_class *class, int type)
526 {
527 return class->stats.objs[type];
528 }
529
530 #ifdef CONFIG_ZSMALLOC_STAT
531
532 static void __init zs_stat_init(void)
533 {
534 if (!debugfs_initialized()) {
535 pr_warn("debugfs not available, stat dir not created\n");
536 return;
537 }
538
539 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
540 }
541
542 static void __exit zs_stat_exit(void)
543 {
544 debugfs_remove_recursive(zs_stat_root);
545 }
546
547 static unsigned long zs_can_compact(struct size_class *class);
548
549 static int zs_stats_size_show(struct seq_file *s, void *v)
550 {
551 int i, fg;
552 struct zs_pool *pool = s->private;
553 struct size_class *class;
554 int objs_per_zspage;
555 unsigned long obj_allocated, obj_used, pages_used, freeable;
556 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
557 unsigned long total_freeable = 0;
558 unsigned long inuse_totals[NR_FULLNESS_GROUPS] = {0, };
559
560 seq_printf(s, " %5s %5s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %13s %10s %10s %16s %8s\n",
561 "class", "size", "10%", "20%", "30%", "40%",
562 "50%", "60%", "70%", "80%", "90%", "99%", "100%",
563 "obj_allocated", "obj_used", "pages_used",
564 "pages_per_zspage", "freeable");
565
566 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
567
568 class = pool->size_class[i];
569
570 if (class->index != i)
571 continue;
572
573 spin_lock(&class->lock);
574
575 seq_printf(s, " %5u %5u ", i, class->size);
576 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++) {
577 inuse_totals[fg] += class_stat_read(class, fg);
578 seq_printf(s, "%9lu ", class_stat_read(class, fg));
579 }
580
581 obj_allocated = class_stat_read(class, ZS_OBJS_ALLOCATED);
582 obj_used = class_stat_read(class, ZS_OBJS_INUSE);
583 freeable = zs_can_compact(class);
584 spin_unlock(&class->lock);
585
586 objs_per_zspage = class->objs_per_zspage;
587 pages_used = obj_allocated / objs_per_zspage *
588 class->pages_per_zspage;
589
590 seq_printf(s, "%13lu %10lu %10lu %16d %8lu\n",
591 obj_allocated, obj_used, pages_used,
592 class->pages_per_zspage, freeable);
593
594 total_objs += obj_allocated;
595 total_used_objs += obj_used;
596 total_pages += pages_used;
597 total_freeable += freeable;
598 }
599
600 seq_puts(s, "\n");
601 seq_printf(s, " %5s %5s ", "Total", "");
602
603 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++)
604 seq_printf(s, "%9lu ", inuse_totals[fg]);
605
606 seq_printf(s, "%13lu %10lu %10lu %16s %8lu\n",
607 total_objs, total_used_objs, total_pages, "",
608 total_freeable);
609
610 return 0;
611 }
612 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
613
614 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
615 {
616 if (!zs_stat_root) {
617 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
618 return;
619 }
620
621 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
622
623 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
624 &zs_stats_size_fops);
625 }
626
627 static void zs_pool_stat_destroy(struct zs_pool *pool)
628 {
629 debugfs_remove_recursive(pool->stat_dentry);
630 }
631
632 #else /* CONFIG_ZSMALLOC_STAT */
633 static void __init zs_stat_init(void)
634 {
635 }
636
637 static void __exit zs_stat_exit(void)
638 {
639 }
640
641 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
642 {
643 }
644
645 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
646 {
647 }
648 #endif
649
650
651 /*
652 * For each size class, zspages are divided into different groups
653 * depending on their usage ratio. This function returns fullness
654 * status of the given page.
655 */
656 static int get_fullness_group(struct size_class *class, struct zspage *zspage)
657 {
658 int inuse, objs_per_zspage, ratio;
659
660 inuse = get_zspage_inuse(zspage);
661 objs_per_zspage = class->objs_per_zspage;
662
663 if (inuse == 0)
664 return ZS_INUSE_RATIO_0;
665 if (inuse == objs_per_zspage)
666 return ZS_INUSE_RATIO_100;
667
668 ratio = 100 * inuse / objs_per_zspage;
669 /*
670 * Take integer division into consideration: a page with one inuse
671 * object out of 127 possible, will end up having 0 usage ratio,
672 * which is wrong as it belongs in ZS_INUSE_RATIO_10 fullness group.
673 */
674 return ratio / 10 + 1;
675 }
676
677 /*
678 * Each size class maintains various freelists and zspages are assigned
679 * to one of these freelists based on the number of live objects they
680 * have. This functions inserts the given zspage into the freelist
681 * identified by <class, fullness_group>.
682 */
683 static void insert_zspage(struct size_class *class,
684 struct zspage *zspage,
685 int fullness)
686 {
687 class_stat_add(class, fullness, 1);
688 list_add(&zspage->list, &class->fullness_list[fullness]);
689 zspage->fullness = fullness;
690 }
691
692 /*
693 * This function removes the given zspage from the freelist identified
694 * by <class, fullness_group>.
695 */
696 static void remove_zspage(struct size_class *class, struct zspage *zspage)
697 {
698 int fullness = zspage->fullness;
699
700 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
701
702 list_del_init(&zspage->list);
703 class_stat_sub(class, fullness, 1);
704 }
705
706 /*
707 * Each size class maintains zspages in different fullness groups depending
708 * on the number of live objects they contain. When allocating or freeing
709 * objects, the fullness status of the page can change, for instance, from
710 * INUSE_RATIO_80 to INUSE_RATIO_70 when freeing an object. This function
711 * checks if such a status change has occurred for the given page and
712 * accordingly moves the page from the list of the old fullness group to that
713 * of the new fullness group.
714 */
715 static int fix_fullness_group(struct size_class *class, struct zspage *zspage)
716 {
717 int newfg;
718
719 newfg = get_fullness_group(class, zspage);
720 if (newfg == zspage->fullness)
721 goto out;
722
723 remove_zspage(class, zspage);
724 insert_zspage(class, zspage, newfg);
725 out:
726 return newfg;
727 }
728
729 static struct zspage *get_zspage(struct page *page)
730 {
731 struct zspage *zspage = (struct zspage *)page_private(page);
732
733 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
734 return zspage;
735 }
736
737 static struct page *get_next_page(struct page *page)
738 {
739 struct zspage *zspage = get_zspage(page);
740
741 if (unlikely(ZsHugePage(zspage)))
742 return NULL;
743
744 return (struct page *)page->index;
745 }
746
747 /**
748 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
749 * @obj: the encoded object value
750 * @page: page object resides in zspage
751 * @obj_idx: object index
752 */
753 static void obj_to_location(unsigned long obj, struct page **page,
754 unsigned int *obj_idx)
755 {
756 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
757 *obj_idx = (obj & OBJ_INDEX_MASK);
758 }
759
760 static void obj_to_page(unsigned long obj, struct page **page)
761 {
762 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
763 }
764
765 /**
766 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
767 * @page: page object resides in zspage
768 * @obj_idx: object index
769 */
770 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
771 {
772 unsigned long obj;
773
774 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
775 obj |= obj_idx & OBJ_INDEX_MASK;
776
777 return obj;
778 }
779
780 static unsigned long handle_to_obj(unsigned long handle)
781 {
782 return *(unsigned long *)handle;
783 }
784
785 static inline bool obj_allocated(struct page *page, void *obj,
786 unsigned long *phandle)
787 {
788 unsigned long handle;
789 struct zspage *zspage = get_zspage(page);
790
791 if (unlikely(ZsHugePage(zspage))) {
792 VM_BUG_ON_PAGE(!is_first_page(page), page);
793 handle = page->index;
794 } else
795 handle = *(unsigned long *)obj;
796
797 if (!(handle & OBJ_ALLOCATED_TAG))
798 return false;
799
800 /* Clear all tags before returning the handle */
801 *phandle = handle & ~OBJ_TAG_MASK;
802 return true;
803 }
804
805 static void reset_page(struct page *page)
806 {
807 __ClearPageMovable(page);
808 ClearPagePrivate(page);
809 set_page_private(page, 0);
810 page->index = 0;
811 reset_first_obj_offset(page);
812 __ClearPageZsmalloc(page);
813 }
814
815 static int trylock_zspage(struct zspage *zspage)
816 {
817 struct page *cursor, *fail;
818
819 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
820 get_next_page(cursor)) {
821 if (!trylock_page(cursor)) {
822 fail = cursor;
823 goto unlock;
824 }
825 }
826
827 return 1;
828 unlock:
829 for (cursor = get_first_page(zspage); cursor != fail; cursor =
830 get_next_page(cursor))
831 unlock_page(cursor);
832
833 return 0;
834 }
835
836 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
837 struct zspage *zspage)
838 {
839 struct page *page, *next;
840
841 assert_spin_locked(&class->lock);
842
843 VM_BUG_ON(get_zspage_inuse(zspage));
844 VM_BUG_ON(zspage->fullness != ZS_INUSE_RATIO_0);
845
846 next = page = get_first_page(zspage);
847 do {
848 VM_BUG_ON_PAGE(!PageLocked(page), page);
849 next = get_next_page(page);
850 reset_page(page);
851 unlock_page(page);
852 dec_zone_page_state(page, NR_ZSPAGES);
853 put_page(page);
854 page = next;
855 } while (page != NULL);
856
857 cache_free_zspage(pool, zspage);
858
859 class_stat_sub(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
860 atomic_long_sub(class->pages_per_zspage, &pool->pages_allocated);
861 }
862
863 static void free_zspage(struct zs_pool *pool, struct size_class *class,
864 struct zspage *zspage)
865 {
866 VM_BUG_ON(get_zspage_inuse(zspage));
867 VM_BUG_ON(list_empty(&zspage->list));
868
869 /*
870 * Since zs_free couldn't be sleepable, this function cannot call
871 * lock_page. The page locks trylock_zspage got will be released
872 * by __free_zspage.
873 */
874 if (!trylock_zspage(zspage)) {
875 kick_deferred_free(pool);
876 return;
877 }
878
879 remove_zspage(class, zspage);
880 __free_zspage(pool, class, zspage);
881 }
882
883 /* Initialize a newly allocated zspage */
884 static void init_zspage(struct size_class *class, struct zspage *zspage)
885 {
886 unsigned int freeobj = 1;
887 unsigned long off = 0;
888 struct page *page = get_first_page(zspage);
889
890 while (page) {
891 struct page *next_page;
892 struct link_free *link;
893 void *vaddr;
894
895 set_first_obj_offset(page, off);
896
897 vaddr = kmap_atomic(page);
898 link = (struct link_free *)vaddr + off / sizeof(*link);
899
900 while ((off += class->size) < PAGE_SIZE) {
901 link->next = freeobj++ << OBJ_TAG_BITS;
902 link += class->size / sizeof(*link);
903 }
904
905 /*
906 * We now come to the last (full or partial) object on this
907 * page, which must point to the first object on the next
908 * page (if present)
909 */
910 next_page = get_next_page(page);
911 if (next_page) {
912 link->next = freeobj++ << OBJ_TAG_BITS;
913 } else {
914 /*
915 * Reset OBJ_TAG_BITS bit to last link to tell
916 * whether it's allocated object or not.
917 */
918 link->next = -1UL << OBJ_TAG_BITS;
919 }
920 kunmap_atomic(vaddr);
921 page = next_page;
922 off %= PAGE_SIZE;
923 }
924
925 set_freeobj(zspage, 0);
926 }
927
928 static void create_page_chain(struct size_class *class, struct zspage *zspage,
929 struct page *pages[])
930 {
931 int i;
932 struct page *page;
933 struct page *prev_page = NULL;
934 int nr_pages = class->pages_per_zspage;
935
936 /*
937 * Allocate individual pages and link them together as:
938 * 1. all pages are linked together using page->index
939 * 2. each sub-page point to zspage using page->private
940 *
941 * we set PG_private to identify the first page (i.e. no other sub-page
942 * has this flag set).
943 */
944 for (i = 0; i < nr_pages; i++) {
945 page = pages[i];
946 set_page_private(page, (unsigned long)zspage);
947 page->index = 0;
948 if (i == 0) {
949 zspage->first_page = page;
950 SetPagePrivate(page);
951 if (unlikely(class->objs_per_zspage == 1 &&
952 class->pages_per_zspage == 1))
953 SetZsHugePage(zspage);
954 } else {
955 prev_page->index = (unsigned long)page;
956 }
957 prev_page = page;
958 }
959 }
960
961 /*
962 * Allocate a zspage for the given size class
963 */
964 static struct zspage *alloc_zspage(struct zs_pool *pool,
965 struct size_class *class,
966 gfp_t gfp)
967 {
968 int i;
969 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
970 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
971
972 if (!zspage)
973 return NULL;
974
975 zspage->magic = ZSPAGE_MAGIC;
976 migrate_lock_init(zspage);
977
978 for (i = 0; i < class->pages_per_zspage; i++) {
979 struct page *page;
980
981 page = alloc_page(gfp);
982 if (!page) {
983 while (--i >= 0) {
984 dec_zone_page_state(pages[i], NR_ZSPAGES);
985 __ClearPageZsmalloc(pages[i]);
986 __free_page(pages[i]);
987 }
988 cache_free_zspage(pool, zspage);
989 return NULL;
990 }
991 __SetPageZsmalloc(page);
992
993 inc_zone_page_state(page, NR_ZSPAGES);
994 pages[i] = page;
995 }
996
997 create_page_chain(class, zspage, pages);
998 init_zspage(class, zspage);
999 zspage->pool = pool;
1000 zspage->class = class->index;
1001
1002 return zspage;
1003 }
1004
1005 static struct zspage *find_get_zspage(struct size_class *class)
1006 {
1007 int i;
1008 struct zspage *zspage;
1009
1010 for (i = ZS_INUSE_RATIO_99; i >= ZS_INUSE_RATIO_0; i--) {
1011 zspage = list_first_entry_or_null(&class->fullness_list[i],
1012 struct zspage, list);
1013 if (zspage)
1014 break;
1015 }
1016
1017 return zspage;
1018 }
1019
1020 static inline int __zs_cpu_up(struct mapping_area *area)
1021 {
1022 /*
1023 * Make sure we don't leak memory if a cpu UP notification
1024 * and zs_init() race and both call zs_cpu_up() on the same cpu
1025 */
1026 if (area->vm_buf)
1027 return 0;
1028 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1029 if (!area->vm_buf)
1030 return -ENOMEM;
1031 return 0;
1032 }
1033
1034 static inline void __zs_cpu_down(struct mapping_area *area)
1035 {
1036 kfree(area->vm_buf);
1037 area->vm_buf = NULL;
1038 }
1039
1040 static void *__zs_map_object(struct mapping_area *area,
1041 struct page *pages[2], int off, int size)
1042 {
1043 int sizes[2];
1044 void *addr;
1045 char *buf = area->vm_buf;
1046
1047 /* disable page faults to match kmap_atomic() return conditions */
1048 pagefault_disable();
1049
1050 /* no read fastpath */
1051 if (area->vm_mm == ZS_MM_WO)
1052 goto out;
1053
1054 sizes[0] = PAGE_SIZE - off;
1055 sizes[1] = size - sizes[0];
1056
1057 /* copy object to per-cpu buffer */
1058 addr = kmap_atomic(pages[0]);
1059 memcpy(buf, addr + off, sizes[0]);
1060 kunmap_atomic(addr);
1061 addr = kmap_atomic(pages[1]);
1062 memcpy(buf + sizes[0], addr, sizes[1]);
1063 kunmap_atomic(addr);
1064 out:
1065 return area->vm_buf;
1066 }
1067
1068 static void __zs_unmap_object(struct mapping_area *area,
1069 struct page *pages[2], int off, int size)
1070 {
1071 int sizes[2];
1072 void *addr;
1073 char *buf;
1074
1075 /* no write fastpath */
1076 if (area->vm_mm == ZS_MM_RO)
1077 goto out;
1078
1079 buf = area->vm_buf;
1080 buf = buf + ZS_HANDLE_SIZE;
1081 size -= ZS_HANDLE_SIZE;
1082 off += ZS_HANDLE_SIZE;
1083
1084 sizes[0] = PAGE_SIZE - off;
1085 sizes[1] = size - sizes[0];
1086
1087 /* copy per-cpu buffer to object */
1088 addr = kmap_atomic(pages[0]);
1089 memcpy(addr + off, buf, sizes[0]);
1090 kunmap_atomic(addr);
1091 addr = kmap_atomic(pages[1]);
1092 memcpy(addr, buf + sizes[0], sizes[1]);
1093 kunmap_atomic(addr);
1094
1095 out:
1096 /* enable page faults to match kunmap_atomic() return conditions */
1097 pagefault_enable();
1098 }
1099
1100 static int zs_cpu_prepare(unsigned int cpu)
1101 {
1102 struct mapping_area *area;
1103
1104 area = &per_cpu(zs_map_area, cpu);
1105 return __zs_cpu_up(area);
1106 }
1107
1108 static int zs_cpu_dead(unsigned int cpu)
1109 {
1110 struct mapping_area *area;
1111
1112 area = &per_cpu(zs_map_area, cpu);
1113 __zs_cpu_down(area);
1114 return 0;
1115 }
1116
1117 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1118 int objs_per_zspage)
1119 {
1120 if (prev->pages_per_zspage == pages_per_zspage &&
1121 prev->objs_per_zspage == objs_per_zspage)
1122 return true;
1123
1124 return false;
1125 }
1126
1127 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1128 {
1129 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1130 }
1131
1132 static bool zspage_empty(struct zspage *zspage)
1133 {
1134 return get_zspage_inuse(zspage) == 0;
1135 }
1136
1137 /**
1138 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1139 * that hold objects of the provided size.
1140 * @pool: zsmalloc pool to use
1141 * @size: object size
1142 *
1143 * Context: Any context.
1144 *
1145 * Return: the index of the zsmalloc &size_class that hold objects of the
1146 * provided size.
1147 */
1148 unsigned int zs_lookup_class_index(struct zs_pool *pool, unsigned int size)
1149 {
1150 struct size_class *class;
1151
1152 class = pool->size_class[get_size_class_index(size)];
1153
1154 return class->index;
1155 }
1156 EXPORT_SYMBOL_GPL(zs_lookup_class_index);
1157
1158 unsigned long zs_get_total_pages(struct zs_pool *pool)
1159 {
1160 return atomic_long_read(&pool->pages_allocated);
1161 }
1162 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1163
1164 /**
1165 * zs_map_object - get address of allocated object from handle.
1166 * @pool: pool from which the object was allocated
1167 * @handle: handle returned from zs_malloc
1168 * @mm: mapping mode to use
1169 *
1170 * Before using an object allocated from zs_malloc, it must be mapped using
1171 * this function. When done with the object, it must be unmapped using
1172 * zs_unmap_object.
1173 *
1174 * Only one object can be mapped per cpu at a time. There is no protection
1175 * against nested mappings.
1176 *
1177 * This function returns with preemption and page faults disabled.
1178 */
1179 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1180 enum zs_mapmode mm)
1181 {
1182 struct zspage *zspage;
1183 struct page *page;
1184 unsigned long obj, off;
1185 unsigned int obj_idx;
1186
1187 struct size_class *class;
1188 struct mapping_area *area;
1189 struct page *pages[2];
1190 void *ret;
1191
1192 /*
1193 * Because we use per-cpu mapping areas shared among the
1194 * pools/users, we can't allow mapping in interrupt context
1195 * because it can corrupt another users mappings.
1196 */
1197 BUG_ON(in_interrupt());
1198
1199 /* It guarantees it can get zspage from handle safely */
1200 read_lock(&pool->migrate_lock);
1201 obj = handle_to_obj(handle);
1202 obj_to_location(obj, &page, &obj_idx);
1203 zspage = get_zspage(page);
1204
1205 /*
1206 * migration cannot move any zpages in this zspage. Here, class->lock
1207 * is too heavy since callers would take some time until they calls
1208 * zs_unmap_object API so delegate the locking from class to zspage
1209 * which is smaller granularity.
1210 */
1211 migrate_read_lock(zspage);
1212 read_unlock(&pool->migrate_lock);
1213
1214 class = zspage_class(pool, zspage);
1215 off = offset_in_page(class->size * obj_idx);
1216
1217 local_lock(&zs_map_area.lock);
1218 area = this_cpu_ptr(&zs_map_area);
1219 area->vm_mm = mm;
1220 if (off + class->size <= PAGE_SIZE) {
1221 /* this object is contained entirely within a page */
1222 area->vm_addr = kmap_atomic(page);
1223 ret = area->vm_addr + off;
1224 goto out;
1225 }
1226
1227 /* this object spans two pages */
1228 pages[0] = page;
1229 pages[1] = get_next_page(page);
1230 BUG_ON(!pages[1]);
1231
1232 ret = __zs_map_object(area, pages, off, class->size);
1233 out:
1234 if (likely(!ZsHugePage(zspage)))
1235 ret += ZS_HANDLE_SIZE;
1236
1237 return ret;
1238 }
1239 EXPORT_SYMBOL_GPL(zs_map_object);
1240
1241 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1242 {
1243 struct zspage *zspage;
1244 struct page *page;
1245 unsigned long obj, off;
1246 unsigned int obj_idx;
1247
1248 struct size_class *class;
1249 struct mapping_area *area;
1250
1251 obj = handle_to_obj(handle);
1252 obj_to_location(obj, &page, &obj_idx);
1253 zspage = get_zspage(page);
1254 class = zspage_class(pool, zspage);
1255 off = offset_in_page(class->size * obj_idx);
1256
1257 area = this_cpu_ptr(&zs_map_area);
1258 if (off + class->size <= PAGE_SIZE)
1259 kunmap_atomic(area->vm_addr);
1260 else {
1261 struct page *pages[2];
1262
1263 pages[0] = page;
1264 pages[1] = get_next_page(page);
1265 BUG_ON(!pages[1]);
1266
1267 __zs_unmap_object(area, pages, off, class->size);
1268 }
1269 local_unlock(&zs_map_area.lock);
1270
1271 migrate_read_unlock(zspage);
1272 }
1273 EXPORT_SYMBOL_GPL(zs_unmap_object);
1274
1275 /**
1276 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1277 * zsmalloc &size_class.
1278 * @pool: zsmalloc pool to use
1279 *
1280 * The function returns the size of the first huge class - any object of equal
1281 * or bigger size will be stored in zspage consisting of a single physical
1282 * page.
1283 *
1284 * Context: Any context.
1285 *
1286 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1287 */
1288 size_t zs_huge_class_size(struct zs_pool *pool)
1289 {
1290 return huge_class_size;
1291 }
1292 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1293
1294 static unsigned long obj_malloc(struct zs_pool *pool,
1295 struct zspage *zspage, unsigned long handle)
1296 {
1297 int i, nr_page, offset;
1298 unsigned long obj;
1299 struct link_free *link;
1300 struct size_class *class;
1301
1302 struct page *m_page;
1303 unsigned long m_offset;
1304 void *vaddr;
1305
1306 class = pool->size_class[zspage->class];
1307 obj = get_freeobj(zspage);
1308
1309 offset = obj * class->size;
1310 nr_page = offset >> PAGE_SHIFT;
1311 m_offset = offset_in_page(offset);
1312 m_page = get_first_page(zspage);
1313
1314 for (i = 0; i < nr_page; i++)
1315 m_page = get_next_page(m_page);
1316
1317 vaddr = kmap_atomic(m_page);
1318 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1319 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1320 if (likely(!ZsHugePage(zspage)))
1321 /* record handle in the header of allocated chunk */
1322 link->handle = handle | OBJ_ALLOCATED_TAG;
1323 else
1324 /* record handle to page->index */
1325 zspage->first_page->index = handle | OBJ_ALLOCATED_TAG;
1326
1327 kunmap_atomic(vaddr);
1328 mod_zspage_inuse(zspage, 1);
1329
1330 obj = location_to_obj(m_page, obj);
1331 record_obj(handle, obj);
1332
1333 return obj;
1334 }
1335
1336
1337 /**
1338 * zs_malloc - Allocate block of given size from pool.
1339 * @pool: pool to allocate from
1340 * @size: size of block to allocate
1341 * @gfp: gfp flags when allocating object
1342 *
1343 * On success, handle to the allocated object is returned,
1344 * otherwise an ERR_PTR().
1345 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1346 */
1347 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1348 {
1349 unsigned long handle;
1350 struct size_class *class;
1351 int newfg;
1352 struct zspage *zspage;
1353
1354 if (unlikely(!size))
1355 return (unsigned long)ERR_PTR(-EINVAL);
1356
1357 if (unlikely(size > ZS_MAX_ALLOC_SIZE))
1358 return (unsigned long)ERR_PTR(-ENOSPC);
1359
1360 handle = cache_alloc_handle(pool, gfp);
1361 if (!handle)
1362 return (unsigned long)ERR_PTR(-ENOMEM);
1363
1364 /* extra space in chunk to keep the handle */
1365 size += ZS_HANDLE_SIZE;
1366 class = pool->size_class[get_size_class_index(size)];
1367
1368 /* class->lock effectively protects the zpage migration */
1369 spin_lock(&class->lock);
1370 zspage = find_get_zspage(class);
1371 if (likely(zspage)) {
1372 obj_malloc(pool, zspage, handle);
1373 /* Now move the zspage to another fullness group, if required */
1374 fix_fullness_group(class, zspage);
1375 class_stat_add(class, ZS_OBJS_INUSE, 1);
1376
1377 goto out;
1378 }
1379
1380 spin_unlock(&class->lock);
1381
1382 zspage = alloc_zspage(pool, class, gfp);
1383 if (!zspage) {
1384 cache_free_handle(pool, handle);
1385 return (unsigned long)ERR_PTR(-ENOMEM);
1386 }
1387
1388 spin_lock(&class->lock);
1389 obj_malloc(pool, zspage, handle);
1390 newfg = get_fullness_group(class, zspage);
1391 insert_zspage(class, zspage, newfg);
1392 atomic_long_add(class->pages_per_zspage, &pool->pages_allocated);
1393 class_stat_add(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
1394 class_stat_add(class, ZS_OBJS_INUSE, 1);
1395
1396 /* We completely set up zspage so mark them as movable */
1397 SetZsPageMovable(pool, zspage);
1398 out:
1399 spin_unlock(&class->lock);
1400
1401 return handle;
1402 }
1403 EXPORT_SYMBOL_GPL(zs_malloc);
1404
1405 static void obj_free(int class_size, unsigned long obj)
1406 {
1407 struct link_free *link;
1408 struct zspage *zspage;
1409 struct page *f_page;
1410 unsigned long f_offset;
1411 unsigned int f_objidx;
1412 void *vaddr;
1413
1414 obj_to_location(obj, &f_page, &f_objidx);
1415 f_offset = offset_in_page(class_size * f_objidx);
1416 zspage = get_zspage(f_page);
1417
1418 vaddr = kmap_atomic(f_page);
1419 link = (struct link_free *)(vaddr + f_offset);
1420
1421 /* Insert this object in containing zspage's freelist */
1422 if (likely(!ZsHugePage(zspage)))
1423 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1424 else
1425 f_page->index = 0;
1426 set_freeobj(zspage, f_objidx);
1427
1428 kunmap_atomic(vaddr);
1429 mod_zspage_inuse(zspage, -1);
1430 }
1431
1432 void zs_free(struct zs_pool *pool, unsigned long handle)
1433 {
1434 struct zspage *zspage;
1435 struct page *f_page;
1436 unsigned long obj;
1437 struct size_class *class;
1438 int fullness;
1439
1440 if (IS_ERR_OR_NULL((void *)handle))
1441 return;
1442
1443 /*
1444 * The pool->migrate_lock protects the race with zpage's migration
1445 * so it's safe to get the page from handle.
1446 */
1447 read_lock(&pool->migrate_lock);
1448 obj = handle_to_obj(handle);
1449 obj_to_page(obj, &f_page);
1450 zspage = get_zspage(f_page);
1451 class = zspage_class(pool, zspage);
1452 spin_lock(&class->lock);
1453 read_unlock(&pool->migrate_lock);
1454
1455 class_stat_sub(class, ZS_OBJS_INUSE, 1);
1456 obj_free(class->size, obj);
1457
1458 fullness = fix_fullness_group(class, zspage);
1459 if (fullness == ZS_INUSE_RATIO_0)
1460 free_zspage(pool, class, zspage);
1461
1462 spin_unlock(&class->lock);
1463 cache_free_handle(pool, handle);
1464 }
1465 EXPORT_SYMBOL_GPL(zs_free);
1466
1467 static void zs_object_copy(struct size_class *class, unsigned long dst,
1468 unsigned long src)
1469 {
1470 struct page *s_page, *d_page;
1471 unsigned int s_objidx, d_objidx;
1472 unsigned long s_off, d_off;
1473 void *s_addr, *d_addr;
1474 int s_size, d_size, size;
1475 int written = 0;
1476
1477 s_size = d_size = class->size;
1478
1479 obj_to_location(src, &s_page, &s_objidx);
1480 obj_to_location(dst, &d_page, &d_objidx);
1481
1482 s_off = offset_in_page(class->size * s_objidx);
1483 d_off = offset_in_page(class->size * d_objidx);
1484
1485 if (s_off + class->size > PAGE_SIZE)
1486 s_size = PAGE_SIZE - s_off;
1487
1488 if (d_off + class->size > PAGE_SIZE)
1489 d_size = PAGE_SIZE - d_off;
1490
1491 s_addr = kmap_atomic(s_page);
1492 d_addr = kmap_atomic(d_page);
1493
1494 while (1) {
1495 size = min(s_size, d_size);
1496 memcpy(d_addr + d_off, s_addr + s_off, size);
1497 written += size;
1498
1499 if (written == class->size)
1500 break;
1501
1502 s_off += size;
1503 s_size -= size;
1504 d_off += size;
1505 d_size -= size;
1506
1507 /*
1508 * Calling kunmap_atomic(d_addr) is necessary. kunmap_atomic()
1509 * calls must occurs in reverse order of calls to kmap_atomic().
1510 * So, to call kunmap_atomic(s_addr) we should first call
1511 * kunmap_atomic(d_addr). For more details see
1512 * Documentation/mm/highmem.rst.
1513 */
1514 if (s_off >= PAGE_SIZE) {
1515 kunmap_atomic(d_addr);
1516 kunmap_atomic(s_addr);
1517 s_page = get_next_page(s_page);
1518 s_addr = kmap_atomic(s_page);
1519 d_addr = kmap_atomic(d_page);
1520 s_size = class->size - written;
1521 s_off = 0;
1522 }
1523
1524 if (d_off >= PAGE_SIZE) {
1525 kunmap_atomic(d_addr);
1526 d_page = get_next_page(d_page);
1527 d_addr = kmap_atomic(d_page);
1528 d_size = class->size - written;
1529 d_off = 0;
1530 }
1531 }
1532
1533 kunmap_atomic(d_addr);
1534 kunmap_atomic(s_addr);
1535 }
1536
1537 /*
1538 * Find alloced object in zspage from index object and
1539 * return handle.
1540 */
1541 static unsigned long find_alloced_obj(struct size_class *class,
1542 struct page *page, int *obj_idx)
1543 {
1544 unsigned int offset;
1545 int index = *obj_idx;
1546 unsigned long handle = 0;
1547 void *addr = kmap_atomic(page);
1548
1549 offset = get_first_obj_offset(page);
1550 offset += class->size * index;
1551
1552 while (offset < PAGE_SIZE) {
1553 if (obj_allocated(page, addr + offset, &handle))
1554 break;
1555
1556 offset += class->size;
1557 index++;
1558 }
1559
1560 kunmap_atomic(addr);
1561
1562 *obj_idx = index;
1563
1564 return handle;
1565 }
1566
1567 static void migrate_zspage(struct zs_pool *pool, struct zspage *src_zspage,
1568 struct zspage *dst_zspage)
1569 {
1570 unsigned long used_obj, free_obj;
1571 unsigned long handle;
1572 int obj_idx = 0;
1573 struct page *s_page = get_first_page(src_zspage);
1574 struct size_class *class = pool->size_class[src_zspage->class];
1575
1576 while (1) {
1577 handle = find_alloced_obj(class, s_page, &obj_idx);
1578 if (!handle) {
1579 s_page = get_next_page(s_page);
1580 if (!s_page)
1581 break;
1582 obj_idx = 0;
1583 continue;
1584 }
1585
1586 used_obj = handle_to_obj(handle);
1587 free_obj = obj_malloc(pool, dst_zspage, handle);
1588 zs_object_copy(class, free_obj, used_obj);
1589 obj_idx++;
1590 obj_free(class->size, used_obj);
1591
1592 /* Stop if there is no more space */
1593 if (zspage_full(class, dst_zspage))
1594 break;
1595
1596 /* Stop if there are no more objects to migrate */
1597 if (zspage_empty(src_zspage))
1598 break;
1599 }
1600 }
1601
1602 static struct zspage *isolate_src_zspage(struct size_class *class)
1603 {
1604 struct zspage *zspage;
1605 int fg;
1606
1607 for (fg = ZS_INUSE_RATIO_10; fg <= ZS_INUSE_RATIO_99; fg++) {
1608 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1609 struct zspage, list);
1610 if (zspage) {
1611 remove_zspage(class, zspage);
1612 return zspage;
1613 }
1614 }
1615
1616 return zspage;
1617 }
1618
1619 static struct zspage *isolate_dst_zspage(struct size_class *class)
1620 {
1621 struct zspage *zspage;
1622 int fg;
1623
1624 for (fg = ZS_INUSE_RATIO_99; fg >= ZS_INUSE_RATIO_10; fg--) {
1625 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1626 struct zspage, list);
1627 if (zspage) {
1628 remove_zspage(class, zspage);
1629 return zspage;
1630 }
1631 }
1632
1633 return zspage;
1634 }
1635
1636 /*
1637 * putback_zspage - add @zspage into right class's fullness list
1638 * @class: destination class
1639 * @zspage: target page
1640 *
1641 * Return @zspage's fullness status
1642 */
1643 static int putback_zspage(struct size_class *class, struct zspage *zspage)
1644 {
1645 int fullness;
1646
1647 fullness = get_fullness_group(class, zspage);
1648 insert_zspage(class, zspage, fullness);
1649
1650 return fullness;
1651 }
1652
1653 #ifdef CONFIG_COMPACTION
1654 /*
1655 * To prevent zspage destroy during migration, zspage freeing should
1656 * hold locks of all pages in the zspage.
1657 */
1658 static void lock_zspage(struct zspage *zspage)
1659 {
1660 struct page *curr_page, *page;
1661
1662 /*
1663 * Pages we haven't locked yet can be migrated off the list while we're
1664 * trying to lock them, so we need to be careful and only attempt to
1665 * lock each page under migrate_read_lock(). Otherwise, the page we lock
1666 * may no longer belong to the zspage. This means that we may wait for
1667 * the wrong page to unlock, so we must take a reference to the page
1668 * prior to waiting for it to unlock outside migrate_read_lock().
1669 */
1670 while (1) {
1671 migrate_read_lock(zspage);
1672 page = get_first_page(zspage);
1673 if (trylock_page(page))
1674 break;
1675 get_page(page);
1676 migrate_read_unlock(zspage);
1677 wait_on_page_locked(page);
1678 put_page(page);
1679 }
1680
1681 curr_page = page;
1682 while ((page = get_next_page(curr_page))) {
1683 if (trylock_page(page)) {
1684 curr_page = page;
1685 } else {
1686 get_page(page);
1687 migrate_read_unlock(zspage);
1688 wait_on_page_locked(page);
1689 put_page(page);
1690 migrate_read_lock(zspage);
1691 }
1692 }
1693 migrate_read_unlock(zspage);
1694 }
1695 #endif /* CONFIG_COMPACTION */
1696
1697 static void migrate_lock_init(struct zspage *zspage)
1698 {
1699 rwlock_init(&zspage->lock);
1700 }
1701
1702 static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1703 {
1704 read_lock(&zspage->lock);
1705 }
1706
1707 static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1708 {
1709 read_unlock(&zspage->lock);
1710 }
1711
1712 static void migrate_write_lock(struct zspage *zspage)
1713 {
1714 write_lock(&zspage->lock);
1715 }
1716
1717 static void migrate_write_unlock(struct zspage *zspage)
1718 {
1719 write_unlock(&zspage->lock);
1720 }
1721
1722 #ifdef CONFIG_COMPACTION
1723
1724 static const struct movable_operations zsmalloc_mops;
1725
1726 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1727 struct page *newpage, struct page *oldpage)
1728 {
1729 struct page *page;
1730 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1731 int idx = 0;
1732
1733 page = get_first_page(zspage);
1734 do {
1735 if (page == oldpage)
1736 pages[idx] = newpage;
1737 else
1738 pages[idx] = page;
1739 idx++;
1740 } while ((page = get_next_page(page)) != NULL);
1741
1742 create_page_chain(class, zspage, pages);
1743 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1744 if (unlikely(ZsHugePage(zspage)))
1745 newpage->index = oldpage->index;
1746 __SetPageMovable(newpage, &zsmalloc_mops);
1747 }
1748
1749 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1750 {
1751 /*
1752 * Page is locked so zspage couldn't be destroyed. For detail, look at
1753 * lock_zspage in free_zspage.
1754 */
1755 VM_BUG_ON_PAGE(PageIsolated(page), page);
1756
1757 return true;
1758 }
1759
1760 static int zs_page_migrate(struct page *newpage, struct page *page,
1761 enum migrate_mode mode)
1762 {
1763 struct zs_pool *pool;
1764 struct size_class *class;
1765 struct zspage *zspage;
1766 struct page *dummy;
1767 void *s_addr, *d_addr, *addr;
1768 unsigned int offset;
1769 unsigned long handle;
1770 unsigned long old_obj, new_obj;
1771 unsigned int obj_idx;
1772
1773 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1774
1775 /* We're committed, tell the world that this is a Zsmalloc page. */
1776 __SetPageZsmalloc(newpage);
1777
1778 /* The page is locked, so this pointer must remain valid */
1779 zspage = get_zspage(page);
1780 pool = zspage->pool;
1781
1782 /*
1783 * The pool migrate_lock protects the race between zpage migration
1784 * and zs_free.
1785 */
1786 write_lock(&pool->migrate_lock);
1787 class = zspage_class(pool, zspage);
1788
1789 /*
1790 * the class lock protects zpage alloc/free in the zspage.
1791 */
1792 spin_lock(&class->lock);
1793 /* the migrate_write_lock protects zpage access via zs_map_object */
1794 migrate_write_lock(zspage);
1795
1796 offset = get_first_obj_offset(page);
1797 s_addr = kmap_atomic(page);
1798
1799 /*
1800 * Here, any user cannot access all objects in the zspage so let's move.
1801 */
1802 d_addr = kmap_atomic(newpage);
1803 copy_page(d_addr, s_addr);
1804 kunmap_atomic(d_addr);
1805
1806 for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE;
1807 addr += class->size) {
1808 if (obj_allocated(page, addr, &handle)) {
1809
1810 old_obj = handle_to_obj(handle);
1811 obj_to_location(old_obj, &dummy, &obj_idx);
1812 new_obj = (unsigned long)location_to_obj(newpage,
1813 obj_idx);
1814 record_obj(handle, new_obj);
1815 }
1816 }
1817 kunmap_atomic(s_addr);
1818
1819 replace_sub_page(class, zspage, newpage, page);
1820 /*
1821 * Since we complete the data copy and set up new zspage structure,
1822 * it's okay to release migration_lock.
1823 */
1824 write_unlock(&pool->migrate_lock);
1825 spin_unlock(&class->lock);
1826 migrate_write_unlock(zspage);
1827
1828 get_page(newpage);
1829 if (page_zone(newpage) != page_zone(page)) {
1830 dec_zone_page_state(page, NR_ZSPAGES);
1831 inc_zone_page_state(newpage, NR_ZSPAGES);
1832 }
1833
1834 reset_page(page);
1835 put_page(page);
1836
1837 return MIGRATEPAGE_SUCCESS;
1838 }
1839
1840 static void zs_page_putback(struct page *page)
1841 {
1842 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1843 }
1844
1845 static const struct movable_operations zsmalloc_mops = {
1846 .isolate_page = zs_page_isolate,
1847 .migrate_page = zs_page_migrate,
1848 .putback_page = zs_page_putback,
1849 };
1850
1851 /*
1852 * Caller should hold page_lock of all pages in the zspage
1853 * In here, we cannot use zspage meta data.
1854 */
1855 static void async_free_zspage(struct work_struct *work)
1856 {
1857 int i;
1858 struct size_class *class;
1859 struct zspage *zspage, *tmp;
1860 LIST_HEAD(free_pages);
1861 struct zs_pool *pool = container_of(work, struct zs_pool,
1862 free_work);
1863
1864 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
1865 class = pool->size_class[i];
1866 if (class->index != i)
1867 continue;
1868
1869 spin_lock(&class->lock);
1870 list_splice_init(&class->fullness_list[ZS_INUSE_RATIO_0],
1871 &free_pages);
1872 spin_unlock(&class->lock);
1873 }
1874
1875 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
1876 list_del(&zspage->list);
1877 lock_zspage(zspage);
1878
1879 class = zspage_class(pool, zspage);
1880 spin_lock(&class->lock);
1881 class_stat_sub(class, ZS_INUSE_RATIO_0, 1);
1882 __free_zspage(pool, class, zspage);
1883 spin_unlock(&class->lock);
1884 }
1885 };
1886
1887 static void kick_deferred_free(struct zs_pool *pool)
1888 {
1889 schedule_work(&pool->free_work);
1890 }
1891
1892 static void zs_flush_migration(struct zs_pool *pool)
1893 {
1894 flush_work(&pool->free_work);
1895 }
1896
1897 static void init_deferred_free(struct zs_pool *pool)
1898 {
1899 INIT_WORK(&pool->free_work, async_free_zspage);
1900 }
1901
1902 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
1903 {
1904 struct page *page = get_first_page(zspage);
1905
1906 do {
1907 WARN_ON(!trylock_page(page));
1908 __SetPageMovable(page, &zsmalloc_mops);
1909 unlock_page(page);
1910 } while ((page = get_next_page(page)) != NULL);
1911 }
1912 #else
1913 static inline void zs_flush_migration(struct zs_pool *pool) { }
1914 #endif
1915
1916 /*
1917 *
1918 * Based on the number of unused allocated objects calculate
1919 * and return the number of pages that we can free.
1920 */
1921 static unsigned long zs_can_compact(struct size_class *class)
1922 {
1923 unsigned long obj_wasted;
1924 unsigned long obj_allocated = class_stat_read(class, ZS_OBJS_ALLOCATED);
1925 unsigned long obj_used = class_stat_read(class, ZS_OBJS_INUSE);
1926
1927 if (obj_allocated <= obj_used)
1928 return 0;
1929
1930 obj_wasted = obj_allocated - obj_used;
1931 obj_wasted /= class->objs_per_zspage;
1932
1933 return obj_wasted * class->pages_per_zspage;
1934 }
1935
1936 static unsigned long __zs_compact(struct zs_pool *pool,
1937 struct size_class *class)
1938 {
1939 struct zspage *src_zspage = NULL;
1940 struct zspage *dst_zspage = NULL;
1941 unsigned long pages_freed = 0;
1942
1943 /*
1944 * protect the race between zpage migration and zs_free
1945 * as well as zpage allocation/free
1946 */
1947 write_lock(&pool->migrate_lock);
1948 spin_lock(&class->lock);
1949 while (zs_can_compact(class)) {
1950 int fg;
1951
1952 if (!dst_zspage) {
1953 dst_zspage = isolate_dst_zspage(class);
1954 if (!dst_zspage)
1955 break;
1956 }
1957
1958 src_zspage = isolate_src_zspage(class);
1959 if (!src_zspage)
1960 break;
1961
1962 migrate_write_lock(src_zspage);
1963 migrate_zspage(pool, src_zspage, dst_zspage);
1964 migrate_write_unlock(src_zspage);
1965
1966 fg = putback_zspage(class, src_zspage);
1967 if (fg == ZS_INUSE_RATIO_0) {
1968 free_zspage(pool, class, src_zspage);
1969 pages_freed += class->pages_per_zspage;
1970 }
1971 src_zspage = NULL;
1972
1973 if (get_fullness_group(class, dst_zspage) == ZS_INUSE_RATIO_100
1974 || rwlock_is_contended(&pool->migrate_lock)) {
1975 putback_zspage(class, dst_zspage);
1976 dst_zspage = NULL;
1977
1978 spin_unlock(&class->lock);
1979 write_unlock(&pool->migrate_lock);
1980 cond_resched();
1981 write_lock(&pool->migrate_lock);
1982 spin_lock(&class->lock);
1983 }
1984 }
1985
1986 if (src_zspage)
1987 putback_zspage(class, src_zspage);
1988
1989 if (dst_zspage)
1990 putback_zspage(class, dst_zspage);
1991
1992 spin_unlock(&class->lock);
1993 write_unlock(&pool->migrate_lock);
1994
1995 return pages_freed;
1996 }
1997
1998 unsigned long zs_compact(struct zs_pool *pool)
1999 {
2000 int i;
2001 struct size_class *class;
2002 unsigned long pages_freed = 0;
2003
2004 /*
2005 * Pool compaction is performed under pool->migrate_lock so it is basically
2006 * single-threaded. Having more than one thread in __zs_compact()
2007 * will increase pool->migrate_lock contention, which will impact other
2008 * zsmalloc operations that need pool->migrate_lock.
2009 */
2010 if (atomic_xchg(&pool->compaction_in_progress, 1))
2011 return 0;
2012
2013 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2014 class = pool->size_class[i];
2015 if (class->index != i)
2016 continue;
2017 pages_freed += __zs_compact(pool, class);
2018 }
2019 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2020 atomic_set(&pool->compaction_in_progress, 0);
2021
2022 return pages_freed;
2023 }
2024 EXPORT_SYMBOL_GPL(zs_compact);
2025
2026 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2027 {
2028 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2029 }
2030 EXPORT_SYMBOL_GPL(zs_pool_stats);
2031
2032 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2033 struct shrink_control *sc)
2034 {
2035 unsigned long pages_freed;
2036 struct zs_pool *pool = shrinker->private_data;
2037
2038 /*
2039 * Compact classes and calculate compaction delta.
2040 * Can run concurrently with a manually triggered
2041 * (by user) compaction.
2042 */
2043 pages_freed = zs_compact(pool);
2044
2045 return pages_freed ? pages_freed : SHRINK_STOP;
2046 }
2047
2048 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2049 struct shrink_control *sc)
2050 {
2051 int i;
2052 struct size_class *class;
2053 unsigned long pages_to_free = 0;
2054 struct zs_pool *pool = shrinker->private_data;
2055
2056 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2057 class = pool->size_class[i];
2058 if (class->index != i)
2059 continue;
2060
2061 pages_to_free += zs_can_compact(class);
2062 }
2063
2064 return pages_to_free;
2065 }
2066
2067 static void zs_unregister_shrinker(struct zs_pool *pool)
2068 {
2069 shrinker_free(pool->shrinker);
2070 }
2071
2072 static int zs_register_shrinker(struct zs_pool *pool)
2073 {
2074 pool->shrinker = shrinker_alloc(0, "mm-zspool:%s", pool->name);
2075 if (!pool->shrinker)
2076 return -ENOMEM;
2077
2078 pool->shrinker->scan_objects = zs_shrinker_scan;
2079 pool->shrinker->count_objects = zs_shrinker_count;
2080 pool->shrinker->batch = 0;
2081 pool->shrinker->private_data = pool;
2082
2083 shrinker_register(pool->shrinker);
2084
2085 return 0;
2086 }
2087
2088 static int calculate_zspage_chain_size(int class_size)
2089 {
2090 int i, min_waste = INT_MAX;
2091 int chain_size = 1;
2092
2093 if (is_power_of_2(class_size))
2094 return chain_size;
2095
2096 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
2097 int waste;
2098
2099 waste = (i * PAGE_SIZE) % class_size;
2100 if (waste < min_waste) {
2101 min_waste = waste;
2102 chain_size = i;
2103 }
2104 }
2105
2106 return chain_size;
2107 }
2108
2109 /**
2110 * zs_create_pool - Creates an allocation pool to work from.
2111 * @name: pool name to be created
2112 *
2113 * This function must be called before anything when using
2114 * the zsmalloc allocator.
2115 *
2116 * On success, a pointer to the newly created pool is returned,
2117 * otherwise NULL.
2118 */
2119 struct zs_pool *zs_create_pool(const char *name)
2120 {
2121 int i;
2122 struct zs_pool *pool;
2123 struct size_class *prev_class = NULL;
2124
2125 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2126 if (!pool)
2127 return NULL;
2128
2129 init_deferred_free(pool);
2130 rwlock_init(&pool->migrate_lock);
2131 atomic_set(&pool->compaction_in_progress, 0);
2132
2133 pool->name = kstrdup(name, GFP_KERNEL);
2134 if (!pool->name)
2135 goto err;
2136
2137 if (create_cache(pool))
2138 goto err;
2139
2140 /*
2141 * Iterate reversely, because, size of size_class that we want to use
2142 * for merging should be larger or equal to current size.
2143 */
2144 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2145 int size;
2146 int pages_per_zspage;
2147 int objs_per_zspage;
2148 struct size_class *class;
2149 int fullness;
2150
2151 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2152 if (size > ZS_MAX_ALLOC_SIZE)
2153 size = ZS_MAX_ALLOC_SIZE;
2154 pages_per_zspage = calculate_zspage_chain_size(size);
2155 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2156
2157 /*
2158 * We iterate from biggest down to smallest classes,
2159 * so huge_class_size holds the size of the first huge
2160 * class. Any object bigger than or equal to that will
2161 * endup in the huge class.
2162 */
2163 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2164 !huge_class_size) {
2165 huge_class_size = size;
2166 /*
2167 * The object uses ZS_HANDLE_SIZE bytes to store the
2168 * handle. We need to subtract it, because zs_malloc()
2169 * unconditionally adds handle size before it performs
2170 * size class search - so object may be smaller than
2171 * huge class size, yet it still can end up in the huge
2172 * class because it grows by ZS_HANDLE_SIZE extra bytes
2173 * right before class lookup.
2174 */
2175 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2176 }
2177
2178 /*
2179 * size_class is used for normal zsmalloc operation such
2180 * as alloc/free for that size. Although it is natural that we
2181 * have one size_class for each size, there is a chance that we
2182 * can get more memory utilization if we use one size_class for
2183 * many different sizes whose size_class have same
2184 * characteristics. So, we makes size_class point to
2185 * previous size_class if possible.
2186 */
2187 if (prev_class) {
2188 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2189 pool->size_class[i] = prev_class;
2190 continue;
2191 }
2192 }
2193
2194 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2195 if (!class)
2196 goto err;
2197
2198 class->size = size;
2199 class->index = i;
2200 class->pages_per_zspage = pages_per_zspage;
2201 class->objs_per_zspage = objs_per_zspage;
2202 spin_lock_init(&class->lock);
2203 pool->size_class[i] = class;
2204
2205 fullness = ZS_INUSE_RATIO_0;
2206 while (fullness < NR_FULLNESS_GROUPS) {
2207 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2208 fullness++;
2209 }
2210
2211 prev_class = class;
2212 }
2213
2214 /* debug only, don't abort if it fails */
2215 zs_pool_stat_create(pool, name);
2216
2217 /*
2218 * Not critical since shrinker is only used to trigger internal
2219 * defragmentation of the pool which is pretty optional thing. If
2220 * registration fails we still can use the pool normally and user can
2221 * trigger compaction manually. Thus, ignore return code.
2222 */
2223 zs_register_shrinker(pool);
2224
2225 return pool;
2226
2227 err:
2228 zs_destroy_pool(pool);
2229 return NULL;
2230 }
2231 EXPORT_SYMBOL_GPL(zs_create_pool);
2232
2233 void zs_destroy_pool(struct zs_pool *pool)
2234 {
2235 int i;
2236
2237 zs_unregister_shrinker(pool);
2238 zs_flush_migration(pool);
2239 zs_pool_stat_destroy(pool);
2240
2241 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2242 int fg;
2243 struct size_class *class = pool->size_class[i];
2244
2245 if (!class)
2246 continue;
2247
2248 if (class->index != i)
2249 continue;
2250
2251 for (fg = ZS_INUSE_RATIO_0; fg < NR_FULLNESS_GROUPS; fg++) {
2252 if (list_empty(&class->fullness_list[fg]))
2253 continue;
2254
2255 pr_err("Class-%d fullness group %d is not empty\n",
2256 class->size, fg);
2257 }
2258 kfree(class);
2259 }
2260
2261 destroy_cache(pool);
2262 kfree(pool->name);
2263 kfree(pool);
2264 }
2265 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2266
2267 static int __init zs_init(void)
2268 {
2269 int ret;
2270
2271 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2272 zs_cpu_prepare, zs_cpu_dead);
2273 if (ret)
2274 goto out;
2275
2276 #ifdef CONFIG_ZPOOL
2277 zpool_register_driver(&zs_zpool_driver);
2278 #endif
2279
2280 zs_stat_init();
2281
2282 return 0;
2283
2284 out:
2285 return ret;
2286 }
2287
2288 static void __exit zs_exit(void)
2289 {
2290 #ifdef CONFIG_ZPOOL
2291 zpool_unregister_driver(&zs_zpool_driver);
2292 #endif
2293 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2294
2295 zs_stat_exit();
2296 }
2297
2298 module_init(zs_init);
2299 module_exit(zs_exit);
2300
2301 MODULE_LICENSE("Dual BSD/GPL");
2302 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");
2303 MODULE_DESCRIPTION("zsmalloc memory allocator");
2304
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.