Loading...
Note: File does not exist in v3.1.
1/*
2 * zsmalloc memory allocator
3 *
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
6 *
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
9 *
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
12 */
13
14/*
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
17 *
18 * Usage of struct page fields:
19 * page->private: points to the first component (0-order) page
20 * page->index (union with page->freelist): offset of the first object
21 * starting in this page. For the first page, this is
22 * always 0, so we use this field (aka freelist) to point
23 * to the first free object in zspage.
24 * page->lru: links together all component pages (except the first page)
25 * of a zspage
26 *
27 * For _first_ page only:
28 *
29 * page->private: refers to the component page after the first page
30 * If the page is first_page for huge object, it stores handle.
31 * Look at size_class->huge.
32 * page->freelist: points to the first free object in zspage.
33 * Free objects are linked together using in-place
34 * metadata.
35 * page->objects: maximum number of objects we can store in this
36 * zspage (class->zspage_order * PAGE_SIZE / class->size)
37 * page->lru: links together first pages of various zspages.
38 * Basically forming list of zspages in a fullness group.
39 * page->mapping: class index and fullness group of the zspage
40 * page->inuse: the number of objects that are used in this zspage
41 *
42 * Usage of struct page flags:
43 * PG_private: identifies the first component page
44 * PG_private2: identifies the last component page
45 *
46 */
47
48#include <linux/module.h>
49#include <linux/kernel.h>
50#include <linux/sched.h>
51#include <linux/bitops.h>
52#include <linux/errno.h>
53#include <linux/highmem.h>
54#include <linux/string.h>
55#include <linux/slab.h>
56#include <asm/tlbflush.h>
57#include <asm/pgtable.h>
58#include <linux/cpumask.h>
59#include <linux/cpu.h>
60#include <linux/vmalloc.h>
61#include <linux/preempt.h>
62#include <linux/spinlock.h>
63#include <linux/types.h>
64#include <linux/debugfs.h>
65#include <linux/zsmalloc.h>
66#include <linux/zpool.h>
67
68/*
69 * This must be power of 2 and greater than of equal to sizeof(link_free).
70 * These two conditions ensure that any 'struct link_free' itself doesn't
71 * span more than 1 page which avoids complex case of mapping 2 pages simply
72 * to restore link_free pointer values.
73 */
74#define ZS_ALIGN 8
75
76/*
77 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
78 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
79 */
80#define ZS_MAX_ZSPAGE_ORDER 2
81#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
82
83#define ZS_HANDLE_SIZE (sizeof(unsigned long))
84
85/*
86 * Object location (<PFN>, <obj_idx>) is encoded as
87 * as single (unsigned long) handle value.
88 *
89 * Note that object index <obj_idx> is relative to system
90 * page <PFN> it is stored in, so for each sub-page belonging
91 * to a zspage, obj_idx starts with 0.
92 *
93 * This is made more complicated by various memory models and PAE.
94 */
95
96#ifndef MAX_PHYSMEM_BITS
97#ifdef CONFIG_HIGHMEM64G
98#define MAX_PHYSMEM_BITS 36
99#else /* !CONFIG_HIGHMEM64G */
100/*
101 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
102 * be PAGE_SHIFT
103 */
104#define MAX_PHYSMEM_BITS BITS_PER_LONG
105#endif
106#endif
107#define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
108
109/*
110 * Memory for allocating for handle keeps object position by
111 * encoding <page, obj_idx> and the encoded value has a room
112 * in least bit(ie, look at obj_to_location).
113 * We use the bit to synchronize between object access by
114 * user and migration.
115 */
116#define HANDLE_PIN_BIT 0
117
118/*
119 * Head in allocated object should have OBJ_ALLOCATED_TAG
120 * to identify the object was allocated or not.
121 * It's okay to add the status bit in the least bit because
122 * header keeps handle which is 4byte-aligned address so we
123 * have room for two bit at least.
124 */
125#define OBJ_ALLOCATED_TAG 1
126#define OBJ_TAG_BITS 1
127#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
128#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
129
130#define MAX(a, b) ((a) >= (b) ? (a) : (b))
131/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
132#define ZS_MIN_ALLOC_SIZE \
133 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
134/* each chunk includes extra space to keep handle */
135#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
136
137/*
138 * On systems with 4K page size, this gives 255 size classes! There is a
139 * trader-off here:
140 * - Large number of size classes is potentially wasteful as free page are
141 * spread across these classes
142 * - Small number of size classes causes large internal fragmentation
143 * - Probably its better to use specific size classes (empirically
144 * determined). NOTE: all those class sizes must be set as multiple of
145 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
146 *
147 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
148 * (reason above)
149 */
150#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
151
152/*
153 * We do not maintain any list for completely empty or full pages
154 */
155enum fullness_group {
156 ZS_ALMOST_FULL,
157 ZS_ALMOST_EMPTY,
158 _ZS_NR_FULLNESS_GROUPS,
159
160 ZS_EMPTY,
161 ZS_FULL
162};
163
164enum zs_stat_type {
165 OBJ_ALLOCATED,
166 OBJ_USED,
167 CLASS_ALMOST_FULL,
168 CLASS_ALMOST_EMPTY,
169};
170
171#ifdef CONFIG_ZSMALLOC_STAT
172#define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1)
173#else
174#define NR_ZS_STAT_TYPE (OBJ_USED + 1)
175#endif
176
177struct zs_size_stat {
178 unsigned long objs[NR_ZS_STAT_TYPE];
179};
180
181#ifdef CONFIG_ZSMALLOC_STAT
182static struct dentry *zs_stat_root;
183#endif
184
185/*
186 * number of size_classes
187 */
188static int zs_size_classes;
189
190/*
191 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
192 * n <= N / f, where
193 * n = number of allocated objects
194 * N = total number of objects zspage can store
195 * f = fullness_threshold_frac
196 *
197 * Similarly, we assign zspage to:
198 * ZS_ALMOST_FULL when n > N / f
199 * ZS_EMPTY when n == 0
200 * ZS_FULL when n == N
201 *
202 * (see: fix_fullness_group())
203 */
204static const int fullness_threshold_frac = 4;
205
206struct size_class {
207 spinlock_t lock;
208 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
209 /*
210 * Size of objects stored in this class. Must be multiple
211 * of ZS_ALIGN.
212 */
213 int size;
214 unsigned int index;
215
216 struct zs_size_stat stats;
217
218 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
219 int pages_per_zspage;
220 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
221 bool huge;
222};
223
224/*
225 * Placed within free objects to form a singly linked list.
226 * For every zspage, first_page->freelist gives head of this list.
227 *
228 * This must be power of 2 and less than or equal to ZS_ALIGN
229 */
230struct link_free {
231 union {
232 /*
233 * Position of next free chunk (encodes <PFN, obj_idx>)
234 * It's valid for non-allocated object
235 */
236 void *next;
237 /*
238 * Handle of allocated object.
239 */
240 unsigned long handle;
241 };
242};
243
244struct zs_pool {
245 const char *name;
246
247 struct size_class **size_class;
248 struct kmem_cache *handle_cachep;
249
250 gfp_t flags; /* allocation flags used when growing pool */
251 atomic_long_t pages_allocated;
252
253 struct zs_pool_stats stats;
254
255 /* Compact classes */
256 struct shrinker shrinker;
257 /*
258 * To signify that register_shrinker() was successful
259 * and unregister_shrinker() will not Oops.
260 */
261 bool shrinker_enabled;
262#ifdef CONFIG_ZSMALLOC_STAT
263 struct dentry *stat_dentry;
264#endif
265};
266
267/*
268 * A zspage's class index and fullness group
269 * are encoded in its (first)page->mapping
270 */
271#define CLASS_IDX_BITS 28
272#define FULLNESS_BITS 4
273#define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
274#define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
275
276struct mapping_area {
277#ifdef CONFIG_PGTABLE_MAPPING
278 struct vm_struct *vm; /* vm area for mapping object that span pages */
279#else
280 char *vm_buf; /* copy buffer for objects that span pages */
281#endif
282 char *vm_addr; /* address of kmap_atomic()'ed pages */
283 enum zs_mapmode vm_mm; /* mapping mode */
284};
285
286static int create_handle_cache(struct zs_pool *pool)
287{
288 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
289 0, 0, NULL);
290 return pool->handle_cachep ? 0 : 1;
291}
292
293static void destroy_handle_cache(struct zs_pool *pool)
294{
295 kmem_cache_destroy(pool->handle_cachep);
296}
297
298static unsigned long alloc_handle(struct zs_pool *pool)
299{
300 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
301 pool->flags & ~__GFP_HIGHMEM);
302}
303
304static void free_handle(struct zs_pool *pool, unsigned long handle)
305{
306 kmem_cache_free(pool->handle_cachep, (void *)handle);
307}
308
309static void record_obj(unsigned long handle, unsigned long obj)
310{
311 /*
312 * lsb of @obj represents handle lock while other bits
313 * represent object value the handle is pointing so
314 * updating shouldn't do store tearing.
315 */
316 WRITE_ONCE(*(unsigned long *)handle, obj);
317}
318
319/* zpool driver */
320
321#ifdef CONFIG_ZPOOL
322
323static void *zs_zpool_create(const char *name, gfp_t gfp,
324 const struct zpool_ops *zpool_ops,
325 struct zpool *zpool)
326{
327 return zs_create_pool(name, gfp);
328}
329
330static void zs_zpool_destroy(void *pool)
331{
332 zs_destroy_pool(pool);
333}
334
335static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
336 unsigned long *handle)
337{
338 *handle = zs_malloc(pool, size);
339 return *handle ? 0 : -1;
340}
341static void zs_zpool_free(void *pool, unsigned long handle)
342{
343 zs_free(pool, handle);
344}
345
346static int zs_zpool_shrink(void *pool, unsigned int pages,
347 unsigned int *reclaimed)
348{
349 return -EINVAL;
350}
351
352static void *zs_zpool_map(void *pool, unsigned long handle,
353 enum zpool_mapmode mm)
354{
355 enum zs_mapmode zs_mm;
356
357 switch (mm) {
358 case ZPOOL_MM_RO:
359 zs_mm = ZS_MM_RO;
360 break;
361 case ZPOOL_MM_WO:
362 zs_mm = ZS_MM_WO;
363 break;
364 case ZPOOL_MM_RW: /* fallthru */
365 default:
366 zs_mm = ZS_MM_RW;
367 break;
368 }
369
370 return zs_map_object(pool, handle, zs_mm);
371}
372static void zs_zpool_unmap(void *pool, unsigned long handle)
373{
374 zs_unmap_object(pool, handle);
375}
376
377static u64 zs_zpool_total_size(void *pool)
378{
379 return zs_get_total_pages(pool) << PAGE_SHIFT;
380}
381
382static struct zpool_driver zs_zpool_driver = {
383 .type = "zsmalloc",
384 .owner = THIS_MODULE,
385 .create = zs_zpool_create,
386 .destroy = zs_zpool_destroy,
387 .malloc = zs_zpool_malloc,
388 .free = zs_zpool_free,
389 .shrink = zs_zpool_shrink,
390 .map = zs_zpool_map,
391 .unmap = zs_zpool_unmap,
392 .total_size = zs_zpool_total_size,
393};
394
395MODULE_ALIAS("zpool-zsmalloc");
396#endif /* CONFIG_ZPOOL */
397
398static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
399{
400 return pages_per_zspage * PAGE_SIZE / size;
401}
402
403/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
404static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
405
406static int is_first_page(struct page *page)
407{
408 return PagePrivate(page);
409}
410
411static int is_last_page(struct page *page)
412{
413 return PagePrivate2(page);
414}
415
416static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
417 enum fullness_group *fullness)
418{
419 unsigned long m;
420 BUG_ON(!is_first_page(page));
421
422 m = (unsigned long)page->mapping;
423 *fullness = m & FULLNESS_MASK;
424 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
425}
426
427static void set_zspage_mapping(struct page *page, unsigned int class_idx,
428 enum fullness_group fullness)
429{
430 unsigned long m;
431 BUG_ON(!is_first_page(page));
432
433 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
434 (fullness & FULLNESS_MASK);
435 page->mapping = (struct address_space *)m;
436}
437
438/*
439 * zsmalloc divides the pool into various size classes where each
440 * class maintains a list of zspages where each zspage is divided
441 * into equal sized chunks. Each allocation falls into one of these
442 * classes depending on its size. This function returns index of the
443 * size class which has chunk size big enough to hold the give size.
444 */
445static int get_size_class_index(int size)
446{
447 int idx = 0;
448
449 if (likely(size > ZS_MIN_ALLOC_SIZE))
450 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
451 ZS_SIZE_CLASS_DELTA);
452
453 return min(zs_size_classes - 1, idx);
454}
455
456static inline void zs_stat_inc(struct size_class *class,
457 enum zs_stat_type type, unsigned long cnt)
458{
459 if (type < NR_ZS_STAT_TYPE)
460 class->stats.objs[type] += cnt;
461}
462
463static inline void zs_stat_dec(struct size_class *class,
464 enum zs_stat_type type, unsigned long cnt)
465{
466 if (type < NR_ZS_STAT_TYPE)
467 class->stats.objs[type] -= cnt;
468}
469
470static inline unsigned long zs_stat_get(struct size_class *class,
471 enum zs_stat_type type)
472{
473 if (type < NR_ZS_STAT_TYPE)
474 return class->stats.objs[type];
475 return 0;
476}
477
478#ifdef CONFIG_ZSMALLOC_STAT
479
480static int __init zs_stat_init(void)
481{
482 if (!debugfs_initialized())
483 return -ENODEV;
484
485 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
486 if (!zs_stat_root)
487 return -ENOMEM;
488
489 return 0;
490}
491
492static void __exit zs_stat_exit(void)
493{
494 debugfs_remove_recursive(zs_stat_root);
495}
496
497static unsigned long zs_can_compact(struct size_class *class);
498
499static int zs_stats_size_show(struct seq_file *s, void *v)
500{
501 int i;
502 struct zs_pool *pool = s->private;
503 struct size_class *class;
504 int objs_per_zspage;
505 unsigned long class_almost_full, class_almost_empty;
506 unsigned long obj_allocated, obj_used, pages_used, freeable;
507 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
508 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
509 unsigned long total_freeable = 0;
510
511 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
512 "class", "size", "almost_full", "almost_empty",
513 "obj_allocated", "obj_used", "pages_used",
514 "pages_per_zspage", "freeable");
515
516 for (i = 0; i < zs_size_classes; i++) {
517 class = pool->size_class[i];
518
519 if (class->index != i)
520 continue;
521
522 spin_lock(&class->lock);
523 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
524 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
525 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
526 obj_used = zs_stat_get(class, OBJ_USED);
527 freeable = zs_can_compact(class);
528 spin_unlock(&class->lock);
529
530 objs_per_zspage = get_maxobj_per_zspage(class->size,
531 class->pages_per_zspage);
532 pages_used = obj_allocated / objs_per_zspage *
533 class->pages_per_zspage;
534
535 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
536 " %10lu %10lu %16d %8lu\n",
537 i, class->size, class_almost_full, class_almost_empty,
538 obj_allocated, obj_used, pages_used,
539 class->pages_per_zspage, freeable);
540
541 total_class_almost_full += class_almost_full;
542 total_class_almost_empty += class_almost_empty;
543 total_objs += obj_allocated;
544 total_used_objs += obj_used;
545 total_pages += pages_used;
546 total_freeable += freeable;
547 }
548
549 seq_puts(s, "\n");
550 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
551 "Total", "", total_class_almost_full,
552 total_class_almost_empty, total_objs,
553 total_used_objs, total_pages, "", total_freeable);
554
555 return 0;
556}
557
558static int zs_stats_size_open(struct inode *inode, struct file *file)
559{
560 return single_open(file, zs_stats_size_show, inode->i_private);
561}
562
563static const struct file_operations zs_stat_size_ops = {
564 .open = zs_stats_size_open,
565 .read = seq_read,
566 .llseek = seq_lseek,
567 .release = single_release,
568};
569
570static int zs_pool_stat_create(const char *name, struct zs_pool *pool)
571{
572 struct dentry *entry;
573
574 if (!zs_stat_root)
575 return -ENODEV;
576
577 entry = debugfs_create_dir(name, zs_stat_root);
578 if (!entry) {
579 pr_warn("debugfs dir <%s> creation failed\n", name);
580 return -ENOMEM;
581 }
582 pool->stat_dentry = entry;
583
584 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
585 pool->stat_dentry, pool, &zs_stat_size_ops);
586 if (!entry) {
587 pr_warn("%s: debugfs file entry <%s> creation failed\n",
588 name, "classes");
589 return -ENOMEM;
590 }
591
592 return 0;
593}
594
595static void zs_pool_stat_destroy(struct zs_pool *pool)
596{
597 debugfs_remove_recursive(pool->stat_dentry);
598}
599
600#else /* CONFIG_ZSMALLOC_STAT */
601static int __init zs_stat_init(void)
602{
603 return 0;
604}
605
606static void __exit zs_stat_exit(void)
607{
608}
609
610static inline int zs_pool_stat_create(const char *name, struct zs_pool *pool)
611{
612 return 0;
613}
614
615static inline void zs_pool_stat_destroy(struct zs_pool *pool)
616{
617}
618#endif
619
620
621/*
622 * For each size class, zspages are divided into different groups
623 * depending on how "full" they are. This was done so that we could
624 * easily find empty or nearly empty zspages when we try to shrink
625 * the pool (not yet implemented). This function returns fullness
626 * status of the given page.
627 */
628static enum fullness_group get_fullness_group(struct page *page)
629{
630 int inuse, max_objects;
631 enum fullness_group fg;
632 BUG_ON(!is_first_page(page));
633
634 inuse = page->inuse;
635 max_objects = page->objects;
636
637 if (inuse == 0)
638 fg = ZS_EMPTY;
639 else if (inuse == max_objects)
640 fg = ZS_FULL;
641 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
642 fg = ZS_ALMOST_EMPTY;
643 else
644 fg = ZS_ALMOST_FULL;
645
646 return fg;
647}
648
649/*
650 * Each size class maintains various freelists and zspages are assigned
651 * to one of these freelists based on the number of live objects they
652 * have. This functions inserts the given zspage into the freelist
653 * identified by <class, fullness_group>.
654 */
655static void insert_zspage(struct page *page, struct size_class *class,
656 enum fullness_group fullness)
657{
658 struct page **head;
659
660 BUG_ON(!is_first_page(page));
661
662 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
663 return;
664
665 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
666 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
667
668 head = &class->fullness_list[fullness];
669 if (!*head) {
670 *head = page;
671 return;
672 }
673
674 /*
675 * We want to see more ZS_FULL pages and less almost
676 * empty/full. Put pages with higher ->inuse first.
677 */
678 list_add_tail(&page->lru, &(*head)->lru);
679 if (page->inuse >= (*head)->inuse)
680 *head = page;
681}
682
683/*
684 * This function removes the given zspage from the freelist identified
685 * by <class, fullness_group>.
686 */
687static void remove_zspage(struct page *page, struct size_class *class,
688 enum fullness_group fullness)
689{
690 struct page **head;
691
692 BUG_ON(!is_first_page(page));
693
694 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
695 return;
696
697 head = &class->fullness_list[fullness];
698 BUG_ON(!*head);
699 if (list_empty(&(*head)->lru))
700 *head = NULL;
701 else if (*head == page)
702 *head = (struct page *)list_entry((*head)->lru.next,
703 struct page, lru);
704
705 list_del_init(&page->lru);
706 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
707 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
708}
709
710/*
711 * Each size class maintains zspages in different fullness groups depending
712 * on the number of live objects they contain. When allocating or freeing
713 * objects, the fullness status of the page can change, say, from ALMOST_FULL
714 * to ALMOST_EMPTY when freeing an object. This function checks if such
715 * a status change has occurred for the given page and accordingly moves the
716 * page from the freelist of the old fullness group to that of the new
717 * fullness group.
718 */
719static enum fullness_group fix_fullness_group(struct size_class *class,
720 struct page *page)
721{
722 int class_idx;
723 enum fullness_group currfg, newfg;
724
725 BUG_ON(!is_first_page(page));
726
727 get_zspage_mapping(page, &class_idx, &currfg);
728 newfg = get_fullness_group(page);
729 if (newfg == currfg)
730 goto out;
731
732 remove_zspage(page, class, currfg);
733 insert_zspage(page, class, newfg);
734 set_zspage_mapping(page, class_idx, newfg);
735
736out:
737 return newfg;
738}
739
740/*
741 * We have to decide on how many pages to link together
742 * to form a zspage for each size class. This is important
743 * to reduce wastage due to unusable space left at end of
744 * each zspage which is given as:
745 * wastage = Zp % class_size
746 * usage = Zp - wastage
747 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
748 *
749 * For example, for size class of 3/8 * PAGE_SIZE, we should
750 * link together 3 PAGE_SIZE sized pages to form a zspage
751 * since then we can perfectly fit in 8 such objects.
752 */
753static int get_pages_per_zspage(int class_size)
754{
755 int i, max_usedpc = 0;
756 /* zspage order which gives maximum used size per KB */
757 int max_usedpc_order = 1;
758
759 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
760 int zspage_size;
761 int waste, usedpc;
762
763 zspage_size = i * PAGE_SIZE;
764 waste = zspage_size % class_size;
765 usedpc = (zspage_size - waste) * 100 / zspage_size;
766
767 if (usedpc > max_usedpc) {
768 max_usedpc = usedpc;
769 max_usedpc_order = i;
770 }
771 }
772
773 return max_usedpc_order;
774}
775
776/*
777 * A single 'zspage' is composed of many system pages which are
778 * linked together using fields in struct page. This function finds
779 * the first/head page, given any component page of a zspage.
780 */
781static struct page *get_first_page(struct page *page)
782{
783 if (is_first_page(page))
784 return page;
785 else
786 return (struct page *)page_private(page);
787}
788
789static struct page *get_next_page(struct page *page)
790{
791 struct page *next;
792
793 if (is_last_page(page))
794 next = NULL;
795 else if (is_first_page(page))
796 next = (struct page *)page_private(page);
797 else
798 next = list_entry(page->lru.next, struct page, lru);
799
800 return next;
801}
802
803/*
804 * Encode <page, obj_idx> as a single handle value.
805 * We use the least bit of handle for tagging.
806 */
807static void *location_to_obj(struct page *page, unsigned long obj_idx)
808{
809 unsigned long obj;
810
811 if (!page) {
812 BUG_ON(obj_idx);
813 return NULL;
814 }
815
816 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
817 obj |= ((obj_idx) & OBJ_INDEX_MASK);
818 obj <<= OBJ_TAG_BITS;
819
820 return (void *)obj;
821}
822
823/*
824 * Decode <page, obj_idx> pair from the given object handle. We adjust the
825 * decoded obj_idx back to its original value since it was adjusted in
826 * location_to_obj().
827 */
828static void obj_to_location(unsigned long obj, struct page **page,
829 unsigned long *obj_idx)
830{
831 obj >>= OBJ_TAG_BITS;
832 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
833 *obj_idx = (obj & OBJ_INDEX_MASK);
834}
835
836static unsigned long handle_to_obj(unsigned long handle)
837{
838 return *(unsigned long *)handle;
839}
840
841static unsigned long obj_to_head(struct size_class *class, struct page *page,
842 void *obj)
843{
844 if (class->huge) {
845 VM_BUG_ON(!is_first_page(page));
846 return page_private(page);
847 } else
848 return *(unsigned long *)obj;
849}
850
851static unsigned long obj_idx_to_offset(struct page *page,
852 unsigned long obj_idx, int class_size)
853{
854 unsigned long off = 0;
855
856 if (!is_first_page(page))
857 off = page->index;
858
859 return off + obj_idx * class_size;
860}
861
862static inline int trypin_tag(unsigned long handle)
863{
864 unsigned long *ptr = (unsigned long *)handle;
865
866 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
867}
868
869static void pin_tag(unsigned long handle)
870{
871 while (!trypin_tag(handle));
872}
873
874static void unpin_tag(unsigned long handle)
875{
876 unsigned long *ptr = (unsigned long *)handle;
877
878 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
879}
880
881static void reset_page(struct page *page)
882{
883 clear_bit(PG_private, &page->flags);
884 clear_bit(PG_private_2, &page->flags);
885 set_page_private(page, 0);
886 page->mapping = NULL;
887 page->freelist = NULL;
888 page_mapcount_reset(page);
889}
890
891static void free_zspage(struct page *first_page)
892{
893 struct page *nextp, *tmp, *head_extra;
894
895 BUG_ON(!is_first_page(first_page));
896 BUG_ON(first_page->inuse);
897
898 head_extra = (struct page *)page_private(first_page);
899
900 reset_page(first_page);
901 __free_page(first_page);
902
903 /* zspage with only 1 system page */
904 if (!head_extra)
905 return;
906
907 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
908 list_del(&nextp->lru);
909 reset_page(nextp);
910 __free_page(nextp);
911 }
912 reset_page(head_extra);
913 __free_page(head_extra);
914}
915
916/* Initialize a newly allocated zspage */
917static void init_zspage(struct page *first_page, struct size_class *class)
918{
919 unsigned long off = 0;
920 struct page *page = first_page;
921
922 BUG_ON(!is_first_page(first_page));
923 while (page) {
924 struct page *next_page;
925 struct link_free *link;
926 unsigned int i = 1;
927 void *vaddr;
928
929 /*
930 * page->index stores offset of first object starting
931 * in the page. For the first page, this is always 0,
932 * so we use first_page->index (aka ->freelist) to store
933 * head of corresponding zspage's freelist.
934 */
935 if (page != first_page)
936 page->index = off;
937
938 vaddr = kmap_atomic(page);
939 link = (struct link_free *)vaddr + off / sizeof(*link);
940
941 while ((off += class->size) < PAGE_SIZE) {
942 link->next = location_to_obj(page, i++);
943 link += class->size / sizeof(*link);
944 }
945
946 /*
947 * We now come to the last (full or partial) object on this
948 * page, which must point to the first object on the next
949 * page (if present)
950 */
951 next_page = get_next_page(page);
952 link->next = location_to_obj(next_page, 0);
953 kunmap_atomic(vaddr);
954 page = next_page;
955 off %= PAGE_SIZE;
956 }
957}
958
959/*
960 * Allocate a zspage for the given size class
961 */
962static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
963{
964 int i, error;
965 struct page *first_page = NULL, *uninitialized_var(prev_page);
966
967 /*
968 * Allocate individual pages and link them together as:
969 * 1. first page->private = first sub-page
970 * 2. all sub-pages are linked together using page->lru
971 * 3. each sub-page is linked to the first page using page->private
972 *
973 * For each size class, First/Head pages are linked together using
974 * page->lru. Also, we set PG_private to identify the first page
975 * (i.e. no other sub-page has this flag set) and PG_private_2 to
976 * identify the last page.
977 */
978 error = -ENOMEM;
979 for (i = 0; i < class->pages_per_zspage; i++) {
980 struct page *page;
981
982 page = alloc_page(flags);
983 if (!page)
984 goto cleanup;
985
986 INIT_LIST_HEAD(&page->lru);
987 if (i == 0) { /* first page */
988 SetPagePrivate(page);
989 set_page_private(page, 0);
990 first_page = page;
991 first_page->inuse = 0;
992 }
993 if (i == 1)
994 set_page_private(first_page, (unsigned long)page);
995 if (i >= 1)
996 set_page_private(page, (unsigned long)first_page);
997 if (i >= 2)
998 list_add(&page->lru, &prev_page->lru);
999 if (i == class->pages_per_zspage - 1) /* last page */
1000 SetPagePrivate2(page);
1001 prev_page = page;
1002 }
1003
1004 init_zspage(first_page, class);
1005
1006 first_page->freelist = location_to_obj(first_page, 0);
1007 /* Maximum number of objects we can store in this zspage */
1008 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
1009
1010 error = 0; /* Success */
1011
1012cleanup:
1013 if (unlikely(error) && first_page) {
1014 free_zspage(first_page);
1015 first_page = NULL;
1016 }
1017
1018 return first_page;
1019}
1020
1021static struct page *find_get_zspage(struct size_class *class)
1022{
1023 int i;
1024 struct page *page;
1025
1026 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1027 page = class->fullness_list[i];
1028 if (page)
1029 break;
1030 }
1031
1032 return page;
1033}
1034
1035#ifdef CONFIG_PGTABLE_MAPPING
1036static inline int __zs_cpu_up(struct mapping_area *area)
1037{
1038 /*
1039 * Make sure we don't leak memory if a cpu UP notification
1040 * and zs_init() race and both call zs_cpu_up() on the same cpu
1041 */
1042 if (area->vm)
1043 return 0;
1044 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1045 if (!area->vm)
1046 return -ENOMEM;
1047 return 0;
1048}
1049
1050static inline void __zs_cpu_down(struct mapping_area *area)
1051{
1052 if (area->vm)
1053 free_vm_area(area->vm);
1054 area->vm = NULL;
1055}
1056
1057static inline void *__zs_map_object(struct mapping_area *area,
1058 struct page *pages[2], int off, int size)
1059{
1060 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1061 area->vm_addr = area->vm->addr;
1062 return area->vm_addr + off;
1063}
1064
1065static inline void __zs_unmap_object(struct mapping_area *area,
1066 struct page *pages[2], int off, int size)
1067{
1068 unsigned long addr = (unsigned long)area->vm_addr;
1069
1070 unmap_kernel_range(addr, PAGE_SIZE * 2);
1071}
1072
1073#else /* CONFIG_PGTABLE_MAPPING */
1074
1075static inline int __zs_cpu_up(struct mapping_area *area)
1076{
1077 /*
1078 * Make sure we don't leak memory if a cpu UP notification
1079 * and zs_init() race and both call zs_cpu_up() on the same cpu
1080 */
1081 if (area->vm_buf)
1082 return 0;
1083 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1084 if (!area->vm_buf)
1085 return -ENOMEM;
1086 return 0;
1087}
1088
1089static inline void __zs_cpu_down(struct mapping_area *area)
1090{
1091 kfree(area->vm_buf);
1092 area->vm_buf = NULL;
1093}
1094
1095static void *__zs_map_object(struct mapping_area *area,
1096 struct page *pages[2], int off, int size)
1097{
1098 int sizes[2];
1099 void *addr;
1100 char *buf = area->vm_buf;
1101
1102 /* disable page faults to match kmap_atomic() return conditions */
1103 pagefault_disable();
1104
1105 /* no read fastpath */
1106 if (area->vm_mm == ZS_MM_WO)
1107 goto out;
1108
1109 sizes[0] = PAGE_SIZE - off;
1110 sizes[1] = size - sizes[0];
1111
1112 /* copy object to per-cpu buffer */
1113 addr = kmap_atomic(pages[0]);
1114 memcpy(buf, addr + off, sizes[0]);
1115 kunmap_atomic(addr);
1116 addr = kmap_atomic(pages[1]);
1117 memcpy(buf + sizes[0], addr, sizes[1]);
1118 kunmap_atomic(addr);
1119out:
1120 return area->vm_buf;
1121}
1122
1123static void __zs_unmap_object(struct mapping_area *area,
1124 struct page *pages[2], int off, int size)
1125{
1126 int sizes[2];
1127 void *addr;
1128 char *buf;
1129
1130 /* no write fastpath */
1131 if (area->vm_mm == ZS_MM_RO)
1132 goto out;
1133
1134 buf = area->vm_buf;
1135 buf = buf + ZS_HANDLE_SIZE;
1136 size -= ZS_HANDLE_SIZE;
1137 off += ZS_HANDLE_SIZE;
1138
1139 sizes[0] = PAGE_SIZE - off;
1140 sizes[1] = size - sizes[0];
1141
1142 /* copy per-cpu buffer to object */
1143 addr = kmap_atomic(pages[0]);
1144 memcpy(addr + off, buf, sizes[0]);
1145 kunmap_atomic(addr);
1146 addr = kmap_atomic(pages[1]);
1147 memcpy(addr, buf + sizes[0], sizes[1]);
1148 kunmap_atomic(addr);
1149
1150out:
1151 /* enable page faults to match kunmap_atomic() return conditions */
1152 pagefault_enable();
1153}
1154
1155#endif /* CONFIG_PGTABLE_MAPPING */
1156
1157static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1158 void *pcpu)
1159{
1160 int ret, cpu = (long)pcpu;
1161 struct mapping_area *area;
1162
1163 switch (action) {
1164 case CPU_UP_PREPARE:
1165 area = &per_cpu(zs_map_area, cpu);
1166 ret = __zs_cpu_up(area);
1167 if (ret)
1168 return notifier_from_errno(ret);
1169 break;
1170 case CPU_DEAD:
1171 case CPU_UP_CANCELED:
1172 area = &per_cpu(zs_map_area, cpu);
1173 __zs_cpu_down(area);
1174 break;
1175 }
1176
1177 return NOTIFY_OK;
1178}
1179
1180static struct notifier_block zs_cpu_nb = {
1181 .notifier_call = zs_cpu_notifier
1182};
1183
1184static int zs_register_cpu_notifier(void)
1185{
1186 int cpu, uninitialized_var(ret);
1187
1188 cpu_notifier_register_begin();
1189
1190 __register_cpu_notifier(&zs_cpu_nb);
1191 for_each_online_cpu(cpu) {
1192 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1193 if (notifier_to_errno(ret))
1194 break;
1195 }
1196
1197 cpu_notifier_register_done();
1198 return notifier_to_errno(ret);
1199}
1200
1201static void zs_unregister_cpu_notifier(void)
1202{
1203 int cpu;
1204
1205 cpu_notifier_register_begin();
1206
1207 for_each_online_cpu(cpu)
1208 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1209 __unregister_cpu_notifier(&zs_cpu_nb);
1210
1211 cpu_notifier_register_done();
1212}
1213
1214static void init_zs_size_classes(void)
1215{
1216 int nr;
1217
1218 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1219 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1220 nr += 1;
1221
1222 zs_size_classes = nr;
1223}
1224
1225static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1226{
1227 if (prev->pages_per_zspage != pages_per_zspage)
1228 return false;
1229
1230 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1231 != get_maxobj_per_zspage(size, pages_per_zspage))
1232 return false;
1233
1234 return true;
1235}
1236
1237static bool zspage_full(struct page *page)
1238{
1239 BUG_ON(!is_first_page(page));
1240
1241 return page->inuse == page->objects;
1242}
1243
1244unsigned long zs_get_total_pages(struct zs_pool *pool)
1245{
1246 return atomic_long_read(&pool->pages_allocated);
1247}
1248EXPORT_SYMBOL_GPL(zs_get_total_pages);
1249
1250/**
1251 * zs_map_object - get address of allocated object from handle.
1252 * @pool: pool from which the object was allocated
1253 * @handle: handle returned from zs_malloc
1254 *
1255 * Before using an object allocated from zs_malloc, it must be mapped using
1256 * this function. When done with the object, it must be unmapped using
1257 * zs_unmap_object.
1258 *
1259 * Only one object can be mapped per cpu at a time. There is no protection
1260 * against nested mappings.
1261 *
1262 * This function returns with preemption and page faults disabled.
1263 */
1264void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1265 enum zs_mapmode mm)
1266{
1267 struct page *page;
1268 unsigned long obj, obj_idx, off;
1269
1270 unsigned int class_idx;
1271 enum fullness_group fg;
1272 struct size_class *class;
1273 struct mapping_area *area;
1274 struct page *pages[2];
1275 void *ret;
1276
1277 BUG_ON(!handle);
1278
1279 /*
1280 * Because we use per-cpu mapping areas shared among the
1281 * pools/users, we can't allow mapping in interrupt context
1282 * because it can corrupt another users mappings.
1283 */
1284 BUG_ON(in_interrupt());
1285
1286 /* From now on, migration cannot move the object */
1287 pin_tag(handle);
1288
1289 obj = handle_to_obj(handle);
1290 obj_to_location(obj, &page, &obj_idx);
1291 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1292 class = pool->size_class[class_idx];
1293 off = obj_idx_to_offset(page, obj_idx, class->size);
1294
1295 area = &get_cpu_var(zs_map_area);
1296 area->vm_mm = mm;
1297 if (off + class->size <= PAGE_SIZE) {
1298 /* this object is contained entirely within a page */
1299 area->vm_addr = kmap_atomic(page);
1300 ret = area->vm_addr + off;
1301 goto out;
1302 }
1303
1304 /* this object spans two pages */
1305 pages[0] = page;
1306 pages[1] = get_next_page(page);
1307 BUG_ON(!pages[1]);
1308
1309 ret = __zs_map_object(area, pages, off, class->size);
1310out:
1311 if (!class->huge)
1312 ret += ZS_HANDLE_SIZE;
1313
1314 return ret;
1315}
1316EXPORT_SYMBOL_GPL(zs_map_object);
1317
1318void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1319{
1320 struct page *page;
1321 unsigned long obj, obj_idx, off;
1322
1323 unsigned int class_idx;
1324 enum fullness_group fg;
1325 struct size_class *class;
1326 struct mapping_area *area;
1327
1328 BUG_ON(!handle);
1329
1330 obj = handle_to_obj(handle);
1331 obj_to_location(obj, &page, &obj_idx);
1332 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1333 class = pool->size_class[class_idx];
1334 off = obj_idx_to_offset(page, obj_idx, class->size);
1335
1336 area = this_cpu_ptr(&zs_map_area);
1337 if (off + class->size <= PAGE_SIZE)
1338 kunmap_atomic(area->vm_addr);
1339 else {
1340 struct page *pages[2];
1341
1342 pages[0] = page;
1343 pages[1] = get_next_page(page);
1344 BUG_ON(!pages[1]);
1345
1346 __zs_unmap_object(area, pages, off, class->size);
1347 }
1348 put_cpu_var(zs_map_area);
1349 unpin_tag(handle);
1350}
1351EXPORT_SYMBOL_GPL(zs_unmap_object);
1352
1353static unsigned long obj_malloc(struct page *first_page,
1354 struct size_class *class, unsigned long handle)
1355{
1356 unsigned long obj;
1357 struct link_free *link;
1358
1359 struct page *m_page;
1360 unsigned long m_objidx, m_offset;
1361 void *vaddr;
1362
1363 handle |= OBJ_ALLOCATED_TAG;
1364 obj = (unsigned long)first_page->freelist;
1365 obj_to_location(obj, &m_page, &m_objidx);
1366 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1367
1368 vaddr = kmap_atomic(m_page);
1369 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1370 first_page->freelist = link->next;
1371 if (!class->huge)
1372 /* record handle in the header of allocated chunk */
1373 link->handle = handle;
1374 else
1375 /* record handle in first_page->private */
1376 set_page_private(first_page, handle);
1377 kunmap_atomic(vaddr);
1378 first_page->inuse++;
1379 zs_stat_inc(class, OBJ_USED, 1);
1380
1381 return obj;
1382}
1383
1384
1385/**
1386 * zs_malloc - Allocate block of given size from pool.
1387 * @pool: pool to allocate from
1388 * @size: size of block to allocate
1389 *
1390 * On success, handle to the allocated object is returned,
1391 * otherwise 0.
1392 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1393 */
1394unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1395{
1396 unsigned long handle, obj;
1397 struct size_class *class;
1398 struct page *first_page;
1399
1400 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1401 return 0;
1402
1403 handle = alloc_handle(pool);
1404 if (!handle)
1405 return 0;
1406
1407 /* extra space in chunk to keep the handle */
1408 size += ZS_HANDLE_SIZE;
1409 class = pool->size_class[get_size_class_index(size)];
1410
1411 spin_lock(&class->lock);
1412 first_page = find_get_zspage(class);
1413
1414 if (!first_page) {
1415 spin_unlock(&class->lock);
1416 first_page = alloc_zspage(class, pool->flags);
1417 if (unlikely(!first_page)) {
1418 free_handle(pool, handle);
1419 return 0;
1420 }
1421
1422 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1423 atomic_long_add(class->pages_per_zspage,
1424 &pool->pages_allocated);
1425
1426 spin_lock(&class->lock);
1427 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1428 class->size, class->pages_per_zspage));
1429 }
1430
1431 obj = obj_malloc(first_page, class, handle);
1432 /* Now move the zspage to another fullness group, if required */
1433 fix_fullness_group(class, first_page);
1434 record_obj(handle, obj);
1435 spin_unlock(&class->lock);
1436
1437 return handle;
1438}
1439EXPORT_SYMBOL_GPL(zs_malloc);
1440
1441static void obj_free(struct zs_pool *pool, struct size_class *class,
1442 unsigned long obj)
1443{
1444 struct link_free *link;
1445 struct page *first_page, *f_page;
1446 unsigned long f_objidx, f_offset;
1447 void *vaddr;
1448
1449 BUG_ON(!obj);
1450
1451 obj &= ~OBJ_ALLOCATED_TAG;
1452 obj_to_location(obj, &f_page, &f_objidx);
1453 first_page = get_first_page(f_page);
1454
1455 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1456
1457 vaddr = kmap_atomic(f_page);
1458
1459 /* Insert this object in containing zspage's freelist */
1460 link = (struct link_free *)(vaddr + f_offset);
1461 link->next = first_page->freelist;
1462 if (class->huge)
1463 set_page_private(first_page, 0);
1464 kunmap_atomic(vaddr);
1465 first_page->freelist = (void *)obj;
1466 first_page->inuse--;
1467 zs_stat_dec(class, OBJ_USED, 1);
1468}
1469
1470void zs_free(struct zs_pool *pool, unsigned long handle)
1471{
1472 struct page *first_page, *f_page;
1473 unsigned long obj, f_objidx;
1474 int class_idx;
1475 struct size_class *class;
1476 enum fullness_group fullness;
1477
1478 if (unlikely(!handle))
1479 return;
1480
1481 pin_tag(handle);
1482 obj = handle_to_obj(handle);
1483 obj_to_location(obj, &f_page, &f_objidx);
1484 first_page = get_first_page(f_page);
1485
1486 get_zspage_mapping(first_page, &class_idx, &fullness);
1487 class = pool->size_class[class_idx];
1488
1489 spin_lock(&class->lock);
1490 obj_free(pool, class, obj);
1491 fullness = fix_fullness_group(class, first_page);
1492 if (fullness == ZS_EMPTY) {
1493 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1494 class->size, class->pages_per_zspage));
1495 atomic_long_sub(class->pages_per_zspage,
1496 &pool->pages_allocated);
1497 free_zspage(first_page);
1498 }
1499 spin_unlock(&class->lock);
1500 unpin_tag(handle);
1501
1502 free_handle(pool, handle);
1503}
1504EXPORT_SYMBOL_GPL(zs_free);
1505
1506static void zs_object_copy(unsigned long dst, unsigned long src,
1507 struct size_class *class)
1508{
1509 struct page *s_page, *d_page;
1510 unsigned long s_objidx, d_objidx;
1511 unsigned long s_off, d_off;
1512 void *s_addr, *d_addr;
1513 int s_size, d_size, size;
1514 int written = 0;
1515
1516 s_size = d_size = class->size;
1517
1518 obj_to_location(src, &s_page, &s_objidx);
1519 obj_to_location(dst, &d_page, &d_objidx);
1520
1521 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1522 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1523
1524 if (s_off + class->size > PAGE_SIZE)
1525 s_size = PAGE_SIZE - s_off;
1526
1527 if (d_off + class->size > PAGE_SIZE)
1528 d_size = PAGE_SIZE - d_off;
1529
1530 s_addr = kmap_atomic(s_page);
1531 d_addr = kmap_atomic(d_page);
1532
1533 while (1) {
1534 size = min(s_size, d_size);
1535 memcpy(d_addr + d_off, s_addr + s_off, size);
1536 written += size;
1537
1538 if (written == class->size)
1539 break;
1540
1541 s_off += size;
1542 s_size -= size;
1543 d_off += size;
1544 d_size -= size;
1545
1546 if (s_off >= PAGE_SIZE) {
1547 kunmap_atomic(d_addr);
1548 kunmap_atomic(s_addr);
1549 s_page = get_next_page(s_page);
1550 BUG_ON(!s_page);
1551 s_addr = kmap_atomic(s_page);
1552 d_addr = kmap_atomic(d_page);
1553 s_size = class->size - written;
1554 s_off = 0;
1555 }
1556
1557 if (d_off >= PAGE_SIZE) {
1558 kunmap_atomic(d_addr);
1559 d_page = get_next_page(d_page);
1560 BUG_ON(!d_page);
1561 d_addr = kmap_atomic(d_page);
1562 d_size = class->size - written;
1563 d_off = 0;
1564 }
1565 }
1566
1567 kunmap_atomic(d_addr);
1568 kunmap_atomic(s_addr);
1569}
1570
1571/*
1572 * Find alloced object in zspage from index object and
1573 * return handle.
1574 */
1575static unsigned long find_alloced_obj(struct page *page, int index,
1576 struct size_class *class)
1577{
1578 unsigned long head;
1579 int offset = 0;
1580 unsigned long handle = 0;
1581 void *addr = kmap_atomic(page);
1582
1583 if (!is_first_page(page))
1584 offset = page->index;
1585 offset += class->size * index;
1586
1587 while (offset < PAGE_SIZE) {
1588 head = obj_to_head(class, page, addr + offset);
1589 if (head & OBJ_ALLOCATED_TAG) {
1590 handle = head & ~OBJ_ALLOCATED_TAG;
1591 if (trypin_tag(handle))
1592 break;
1593 handle = 0;
1594 }
1595
1596 offset += class->size;
1597 index++;
1598 }
1599
1600 kunmap_atomic(addr);
1601 return handle;
1602}
1603
1604struct zs_compact_control {
1605 /* Source page for migration which could be a subpage of zspage. */
1606 struct page *s_page;
1607 /* Destination page for migration which should be a first page
1608 * of zspage. */
1609 struct page *d_page;
1610 /* Starting object index within @s_page which used for live object
1611 * in the subpage. */
1612 int index;
1613};
1614
1615static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1616 struct zs_compact_control *cc)
1617{
1618 unsigned long used_obj, free_obj;
1619 unsigned long handle;
1620 struct page *s_page = cc->s_page;
1621 struct page *d_page = cc->d_page;
1622 unsigned long index = cc->index;
1623 int ret = 0;
1624
1625 while (1) {
1626 handle = find_alloced_obj(s_page, index, class);
1627 if (!handle) {
1628 s_page = get_next_page(s_page);
1629 if (!s_page)
1630 break;
1631 index = 0;
1632 continue;
1633 }
1634
1635 /* Stop if there is no more space */
1636 if (zspage_full(d_page)) {
1637 unpin_tag(handle);
1638 ret = -ENOMEM;
1639 break;
1640 }
1641
1642 used_obj = handle_to_obj(handle);
1643 free_obj = obj_malloc(d_page, class, handle);
1644 zs_object_copy(free_obj, used_obj, class);
1645 index++;
1646 /*
1647 * record_obj updates handle's value to free_obj and it will
1648 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1649 * breaks synchronization using pin_tag(e,g, zs_free) so
1650 * let's keep the lock bit.
1651 */
1652 free_obj |= BIT(HANDLE_PIN_BIT);
1653 record_obj(handle, free_obj);
1654 unpin_tag(handle);
1655 obj_free(pool, class, used_obj);
1656 }
1657
1658 /* Remember last position in this iteration */
1659 cc->s_page = s_page;
1660 cc->index = index;
1661
1662 return ret;
1663}
1664
1665static struct page *isolate_target_page(struct size_class *class)
1666{
1667 int i;
1668 struct page *page;
1669
1670 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1671 page = class->fullness_list[i];
1672 if (page) {
1673 remove_zspage(page, class, i);
1674 break;
1675 }
1676 }
1677
1678 return page;
1679}
1680
1681/*
1682 * putback_zspage - add @first_page into right class's fullness list
1683 * @pool: target pool
1684 * @class: destination class
1685 * @first_page: target page
1686 *
1687 * Return @fist_page's fullness_group
1688 */
1689static enum fullness_group putback_zspage(struct zs_pool *pool,
1690 struct size_class *class,
1691 struct page *first_page)
1692{
1693 enum fullness_group fullness;
1694
1695 BUG_ON(!is_first_page(first_page));
1696
1697 fullness = get_fullness_group(first_page);
1698 insert_zspage(first_page, class, fullness);
1699 set_zspage_mapping(first_page, class->index, fullness);
1700
1701 if (fullness == ZS_EMPTY) {
1702 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1703 class->size, class->pages_per_zspage));
1704 atomic_long_sub(class->pages_per_zspage,
1705 &pool->pages_allocated);
1706
1707 free_zspage(first_page);
1708 }
1709
1710 return fullness;
1711}
1712
1713static struct page *isolate_source_page(struct size_class *class)
1714{
1715 int i;
1716 struct page *page = NULL;
1717
1718 for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) {
1719 page = class->fullness_list[i];
1720 if (!page)
1721 continue;
1722
1723 remove_zspage(page, class, i);
1724 break;
1725 }
1726
1727 return page;
1728}
1729
1730/*
1731 *
1732 * Based on the number of unused allocated objects calculate
1733 * and return the number of pages that we can free.
1734 */
1735static unsigned long zs_can_compact(struct size_class *class)
1736{
1737 unsigned long obj_wasted;
1738 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
1739 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
1740
1741 if (obj_allocated <= obj_used)
1742 return 0;
1743
1744 obj_wasted = obj_allocated - obj_used;
1745 obj_wasted /= get_maxobj_per_zspage(class->size,
1746 class->pages_per_zspage);
1747
1748 return obj_wasted * class->pages_per_zspage;
1749}
1750
1751static void __zs_compact(struct zs_pool *pool, struct size_class *class)
1752{
1753 struct zs_compact_control cc;
1754 struct page *src_page;
1755 struct page *dst_page = NULL;
1756
1757 spin_lock(&class->lock);
1758 while ((src_page = isolate_source_page(class))) {
1759
1760 BUG_ON(!is_first_page(src_page));
1761
1762 if (!zs_can_compact(class))
1763 break;
1764
1765 cc.index = 0;
1766 cc.s_page = src_page;
1767
1768 while ((dst_page = isolate_target_page(class))) {
1769 cc.d_page = dst_page;
1770 /*
1771 * If there is no more space in dst_page, resched
1772 * and see if anyone had allocated another zspage.
1773 */
1774 if (!migrate_zspage(pool, class, &cc))
1775 break;
1776
1777 putback_zspage(pool, class, dst_page);
1778 }
1779
1780 /* Stop if we couldn't find slot */
1781 if (dst_page == NULL)
1782 break;
1783
1784 putback_zspage(pool, class, dst_page);
1785 if (putback_zspage(pool, class, src_page) == ZS_EMPTY)
1786 pool->stats.pages_compacted += class->pages_per_zspage;
1787 spin_unlock(&class->lock);
1788 cond_resched();
1789 spin_lock(&class->lock);
1790 }
1791
1792 if (src_page)
1793 putback_zspage(pool, class, src_page);
1794
1795 spin_unlock(&class->lock);
1796}
1797
1798unsigned long zs_compact(struct zs_pool *pool)
1799{
1800 int i;
1801 struct size_class *class;
1802
1803 for (i = zs_size_classes - 1; i >= 0; i--) {
1804 class = pool->size_class[i];
1805 if (!class)
1806 continue;
1807 if (class->index != i)
1808 continue;
1809 __zs_compact(pool, class);
1810 }
1811
1812 return pool->stats.pages_compacted;
1813}
1814EXPORT_SYMBOL_GPL(zs_compact);
1815
1816void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
1817{
1818 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
1819}
1820EXPORT_SYMBOL_GPL(zs_pool_stats);
1821
1822static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
1823 struct shrink_control *sc)
1824{
1825 unsigned long pages_freed;
1826 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1827 shrinker);
1828
1829 pages_freed = pool->stats.pages_compacted;
1830 /*
1831 * Compact classes and calculate compaction delta.
1832 * Can run concurrently with a manually triggered
1833 * (by user) compaction.
1834 */
1835 pages_freed = zs_compact(pool) - pages_freed;
1836
1837 return pages_freed ? pages_freed : SHRINK_STOP;
1838}
1839
1840static unsigned long zs_shrinker_count(struct shrinker *shrinker,
1841 struct shrink_control *sc)
1842{
1843 int i;
1844 struct size_class *class;
1845 unsigned long pages_to_free = 0;
1846 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1847 shrinker);
1848
1849 for (i = zs_size_classes - 1; i >= 0; i--) {
1850 class = pool->size_class[i];
1851 if (!class)
1852 continue;
1853 if (class->index != i)
1854 continue;
1855
1856 pages_to_free += zs_can_compact(class);
1857 }
1858
1859 return pages_to_free;
1860}
1861
1862static void zs_unregister_shrinker(struct zs_pool *pool)
1863{
1864 if (pool->shrinker_enabled) {
1865 unregister_shrinker(&pool->shrinker);
1866 pool->shrinker_enabled = false;
1867 }
1868}
1869
1870static int zs_register_shrinker(struct zs_pool *pool)
1871{
1872 pool->shrinker.scan_objects = zs_shrinker_scan;
1873 pool->shrinker.count_objects = zs_shrinker_count;
1874 pool->shrinker.batch = 0;
1875 pool->shrinker.seeks = DEFAULT_SEEKS;
1876
1877 return register_shrinker(&pool->shrinker);
1878}
1879
1880/**
1881 * zs_create_pool - Creates an allocation pool to work from.
1882 * @flags: allocation flags used to allocate pool metadata
1883 *
1884 * This function must be called before anything when using
1885 * the zsmalloc allocator.
1886 *
1887 * On success, a pointer to the newly created pool is returned,
1888 * otherwise NULL.
1889 */
1890struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
1891{
1892 int i;
1893 struct zs_pool *pool;
1894 struct size_class *prev_class = NULL;
1895
1896 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
1897 if (!pool)
1898 return NULL;
1899
1900 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1901 GFP_KERNEL);
1902 if (!pool->size_class) {
1903 kfree(pool);
1904 return NULL;
1905 }
1906
1907 pool->name = kstrdup(name, GFP_KERNEL);
1908 if (!pool->name)
1909 goto err;
1910
1911 if (create_handle_cache(pool))
1912 goto err;
1913
1914 /*
1915 * Iterate reversly, because, size of size_class that we want to use
1916 * for merging should be larger or equal to current size.
1917 */
1918 for (i = zs_size_classes - 1; i >= 0; i--) {
1919 int size;
1920 int pages_per_zspage;
1921 struct size_class *class;
1922
1923 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1924 if (size > ZS_MAX_ALLOC_SIZE)
1925 size = ZS_MAX_ALLOC_SIZE;
1926 pages_per_zspage = get_pages_per_zspage(size);
1927
1928 /*
1929 * size_class is used for normal zsmalloc operation such
1930 * as alloc/free for that size. Although it is natural that we
1931 * have one size_class for each size, there is a chance that we
1932 * can get more memory utilization if we use one size_class for
1933 * many different sizes whose size_class have same
1934 * characteristics. So, we makes size_class point to
1935 * previous size_class if possible.
1936 */
1937 if (prev_class) {
1938 if (can_merge(prev_class, size, pages_per_zspage)) {
1939 pool->size_class[i] = prev_class;
1940 continue;
1941 }
1942 }
1943
1944 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1945 if (!class)
1946 goto err;
1947
1948 class->size = size;
1949 class->index = i;
1950 class->pages_per_zspage = pages_per_zspage;
1951 if (pages_per_zspage == 1 &&
1952 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1953 class->huge = true;
1954 spin_lock_init(&class->lock);
1955 pool->size_class[i] = class;
1956
1957 prev_class = class;
1958 }
1959
1960 pool->flags = flags;
1961
1962 if (zs_pool_stat_create(name, pool))
1963 goto err;
1964
1965 /*
1966 * Not critical, we still can use the pool
1967 * and user can trigger compaction manually.
1968 */
1969 if (zs_register_shrinker(pool) == 0)
1970 pool->shrinker_enabled = true;
1971 return pool;
1972
1973err:
1974 zs_destroy_pool(pool);
1975 return NULL;
1976}
1977EXPORT_SYMBOL_GPL(zs_create_pool);
1978
1979void zs_destroy_pool(struct zs_pool *pool)
1980{
1981 int i;
1982
1983 zs_unregister_shrinker(pool);
1984 zs_pool_stat_destroy(pool);
1985
1986 for (i = 0; i < zs_size_classes; i++) {
1987 int fg;
1988 struct size_class *class = pool->size_class[i];
1989
1990 if (!class)
1991 continue;
1992
1993 if (class->index != i)
1994 continue;
1995
1996 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1997 if (class->fullness_list[fg]) {
1998 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1999 class->size, fg);
2000 }
2001 }
2002 kfree(class);
2003 }
2004
2005 destroy_handle_cache(pool);
2006 kfree(pool->size_class);
2007 kfree(pool->name);
2008 kfree(pool);
2009}
2010EXPORT_SYMBOL_GPL(zs_destroy_pool);
2011
2012static int __init zs_init(void)
2013{
2014 int ret = zs_register_cpu_notifier();
2015
2016 if (ret)
2017 goto notifier_fail;
2018
2019 init_zs_size_classes();
2020
2021#ifdef CONFIG_ZPOOL
2022 zpool_register_driver(&zs_zpool_driver);
2023#endif
2024
2025 ret = zs_stat_init();
2026 if (ret) {
2027 pr_err("zs stat initialization failed\n");
2028 goto stat_fail;
2029 }
2030 return 0;
2031
2032stat_fail:
2033#ifdef CONFIG_ZPOOL
2034 zpool_unregister_driver(&zs_zpool_driver);
2035#endif
2036notifier_fail:
2037 zs_unregister_cpu_notifier();
2038
2039 return ret;
2040}
2041
2042static void __exit zs_exit(void)
2043{
2044#ifdef CONFIG_ZPOOL
2045 zpool_unregister_driver(&zs_zpool_driver);
2046#endif
2047 zs_unregister_cpu_notifier();
2048
2049 zs_stat_exit();
2050}
2051
2052module_init(zs_init);
2053module_exit(zs_exit);
2054
2055MODULE_LICENSE("Dual BSD/GPL");
2056MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");