Loading...
1/*
2 * zsmalloc memory allocator
3 *
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
6 *
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
9 *
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
12 */
13
14/*
15 * This allocator is designed for use with zram. Thus, the allocator is
16 * supposed to work well under low memory conditions. In particular, it
17 * never attempts higher order page allocation which is very likely to
18 * fail under memory pressure. On the other hand, if we just use single
19 * (0-order) pages, it would suffer from very high fragmentation --
20 * any object of size PAGE_SIZE/2 or larger would occupy an entire page.
21 * This was one of the major issues with its predecessor (xvmalloc).
22 *
23 * To overcome these issues, zsmalloc allocates a bunch of 0-order pages
24 * and links them together using various 'struct page' fields. These linked
25 * pages act as a single higher-order page i.e. an object can span 0-order
26 * page boundaries. The code refers to these linked pages as a single entity
27 * called zspage.
28 *
29 * For simplicity, zsmalloc can only allocate objects of size up to PAGE_SIZE
30 * since this satisfies the requirements of all its current users (in the
31 * worst case, page is incompressible and is thus stored "as-is" i.e. in
32 * uncompressed form). For allocation requests larger than this size, failure
33 * is returned (see zs_malloc).
34 *
35 * Additionally, zs_malloc() does not return a dereferenceable pointer.
36 * Instead, it returns an opaque handle (unsigned long) which encodes actual
37 * location of the allocated object. The reason for this indirection is that
38 * zsmalloc does not keep zspages permanently mapped since that would cause
39 * issues on 32-bit systems where the VA region for kernel space mappings
40 * is very small. So, before using the allocating memory, the object has to
41 * be mapped using zs_map_object() to get a usable pointer and subsequently
42 * unmapped using zs_unmap_object().
43 *
44 * Following is how we use various fields and flags of underlying
45 * struct page(s) to form a zspage.
46 *
47 * Usage of struct page fields:
48 * page->first_page: points to the first component (0-order) page
49 * page->index (union with page->freelist): offset of the first object
50 * starting in this page. For the first page, this is
51 * always 0, so we use this field (aka freelist) to point
52 * to the first free object in zspage.
53 * page->lru: links together all component pages (except the first page)
54 * of a zspage
55 *
56 * For _first_ page only:
57 *
58 * page->private (union with page->first_page): refers to the
59 * component page after the first page
60 * page->freelist: points to the first free object in zspage.
61 * Free objects are linked together using in-place
62 * metadata.
63 * page->objects: maximum number of objects we can store in this
64 * zspage (class->zspage_order * PAGE_SIZE / class->size)
65 * page->lru: links together first pages of various zspages.
66 * Basically forming list of zspages in a fullness group.
67 * page->mapping: class index and fullness group of the zspage
68 *
69 * Usage of struct page flags:
70 * PG_private: identifies the first component page
71 * PG_private2: identifies the last component page
72 *
73 */
74
75#ifdef CONFIG_ZSMALLOC_DEBUG
76#define DEBUG
77#endif
78
79#include <linux/module.h>
80#include <linux/kernel.h>
81#include <linux/bitops.h>
82#include <linux/errno.h>
83#include <linux/highmem.h>
84#include <linux/string.h>
85#include <linux/slab.h>
86#include <asm/tlbflush.h>
87#include <asm/pgtable.h>
88#include <linux/cpumask.h>
89#include <linux/cpu.h>
90#include <linux/vmalloc.h>
91#include <linux/hardirq.h>
92#include <linux/spinlock.h>
93#include <linux/types.h>
94#include <linux/zsmalloc.h>
95
96/*
97 * This must be power of 2 and greater than of equal to sizeof(link_free).
98 * These two conditions ensure that any 'struct link_free' itself doesn't
99 * span more than 1 page which avoids complex case of mapping 2 pages simply
100 * to restore link_free pointer values.
101 */
102#define ZS_ALIGN 8
103
104/*
105 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
106 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
107 */
108#define ZS_MAX_ZSPAGE_ORDER 2
109#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
110
111/*
112 * Object location (<PFN>, <obj_idx>) is encoded as
113 * as single (unsigned long) handle value.
114 *
115 * Note that object index <obj_idx> is relative to system
116 * page <PFN> it is stored in, so for each sub-page belonging
117 * to a zspage, obj_idx starts with 0.
118 *
119 * This is made more complicated by various memory models and PAE.
120 */
121
122#ifndef MAX_PHYSMEM_BITS
123#ifdef CONFIG_HIGHMEM64G
124#define MAX_PHYSMEM_BITS 36
125#else /* !CONFIG_HIGHMEM64G */
126/*
127 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
128 * be PAGE_SHIFT
129 */
130#define MAX_PHYSMEM_BITS BITS_PER_LONG
131#endif
132#endif
133#define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
134#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
135#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
136
137#define MAX(a, b) ((a) >= (b) ? (a) : (b))
138/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
139#define ZS_MIN_ALLOC_SIZE \
140 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
141#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
142
143/*
144 * On systems with 4K page size, this gives 254 size classes! There is a
145 * trader-off here:
146 * - Large number of size classes is potentially wasteful as free page are
147 * spread across these classes
148 * - Small number of size classes causes large internal fragmentation
149 * - Probably its better to use specific size classes (empirically
150 * determined). NOTE: all those class sizes must be set as multiple of
151 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
152 *
153 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
154 * (reason above)
155 */
156#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
157#define ZS_SIZE_CLASSES ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / \
158 ZS_SIZE_CLASS_DELTA + 1)
159
160/*
161 * We do not maintain any list for completely empty or full pages
162 */
163enum fullness_group {
164 ZS_ALMOST_FULL,
165 ZS_ALMOST_EMPTY,
166 _ZS_NR_FULLNESS_GROUPS,
167
168 ZS_EMPTY,
169 ZS_FULL
170};
171
172/*
173 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
174 * n <= N / f, where
175 * n = number of allocated objects
176 * N = total number of objects zspage can store
177 * f = 1/fullness_threshold_frac
178 *
179 * Similarly, we assign zspage to:
180 * ZS_ALMOST_FULL when n > N / f
181 * ZS_EMPTY when n == 0
182 * ZS_FULL when n == N
183 *
184 * (see: fix_fullness_group())
185 */
186static const int fullness_threshold_frac = 4;
187
188struct size_class {
189 /*
190 * Size of objects stored in this class. Must be multiple
191 * of ZS_ALIGN.
192 */
193 int size;
194 unsigned int index;
195
196 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
197 int pages_per_zspage;
198
199 spinlock_t lock;
200
201 /* stats */
202 u64 pages_allocated;
203
204 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
205};
206
207/*
208 * Placed within free objects to form a singly linked list.
209 * For every zspage, first_page->freelist gives head of this list.
210 *
211 * This must be power of 2 and less than or equal to ZS_ALIGN
212 */
213struct link_free {
214 /* Handle of next free chunk (encodes <PFN, obj_idx>) */
215 void *next;
216};
217
218struct zs_pool {
219 struct size_class size_class[ZS_SIZE_CLASSES];
220
221 gfp_t flags; /* allocation flags used when growing pool */
222};
223
224/*
225 * A zspage's class index and fullness group
226 * are encoded in its (first)page->mapping
227 */
228#define CLASS_IDX_BITS 28
229#define FULLNESS_BITS 4
230#define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
231#define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
232
233struct mapping_area {
234#ifdef CONFIG_PGTABLE_MAPPING
235 struct vm_struct *vm; /* vm area for mapping object that span pages */
236#else
237 char *vm_buf; /* copy buffer for objects that span pages */
238#endif
239 char *vm_addr; /* address of kmap_atomic()'ed pages */
240 enum zs_mapmode vm_mm; /* mapping mode */
241};
242
243
244/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
245static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
246
247static int is_first_page(struct page *page)
248{
249 return PagePrivate(page);
250}
251
252static int is_last_page(struct page *page)
253{
254 return PagePrivate2(page);
255}
256
257static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
258 enum fullness_group *fullness)
259{
260 unsigned long m;
261 BUG_ON(!is_first_page(page));
262
263 m = (unsigned long)page->mapping;
264 *fullness = m & FULLNESS_MASK;
265 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
266}
267
268static void set_zspage_mapping(struct page *page, unsigned int class_idx,
269 enum fullness_group fullness)
270{
271 unsigned long m;
272 BUG_ON(!is_first_page(page));
273
274 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
275 (fullness & FULLNESS_MASK);
276 page->mapping = (struct address_space *)m;
277}
278
279/*
280 * zsmalloc divides the pool into various size classes where each
281 * class maintains a list of zspages where each zspage is divided
282 * into equal sized chunks. Each allocation falls into one of these
283 * classes depending on its size. This function returns index of the
284 * size class which has chunk size big enough to hold the give size.
285 */
286static int get_size_class_index(int size)
287{
288 int idx = 0;
289
290 if (likely(size > ZS_MIN_ALLOC_SIZE))
291 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
292 ZS_SIZE_CLASS_DELTA);
293
294 return idx;
295}
296
297/*
298 * For each size class, zspages are divided into different groups
299 * depending on how "full" they are. This was done so that we could
300 * easily find empty or nearly empty zspages when we try to shrink
301 * the pool (not yet implemented). This function returns fullness
302 * status of the given page.
303 */
304static enum fullness_group get_fullness_group(struct page *page)
305{
306 int inuse, max_objects;
307 enum fullness_group fg;
308 BUG_ON(!is_first_page(page));
309
310 inuse = page->inuse;
311 max_objects = page->objects;
312
313 if (inuse == 0)
314 fg = ZS_EMPTY;
315 else if (inuse == max_objects)
316 fg = ZS_FULL;
317 else if (inuse <= max_objects / fullness_threshold_frac)
318 fg = ZS_ALMOST_EMPTY;
319 else
320 fg = ZS_ALMOST_FULL;
321
322 return fg;
323}
324
325/*
326 * Each size class maintains various freelists and zspages are assigned
327 * to one of these freelists based on the number of live objects they
328 * have. This functions inserts the given zspage into the freelist
329 * identified by <class, fullness_group>.
330 */
331static void insert_zspage(struct page *page, struct size_class *class,
332 enum fullness_group fullness)
333{
334 struct page **head;
335
336 BUG_ON(!is_first_page(page));
337
338 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
339 return;
340
341 head = &class->fullness_list[fullness];
342 if (*head)
343 list_add_tail(&page->lru, &(*head)->lru);
344
345 *head = page;
346}
347
348/*
349 * This function removes the given zspage from the freelist identified
350 * by <class, fullness_group>.
351 */
352static void remove_zspage(struct page *page, struct size_class *class,
353 enum fullness_group fullness)
354{
355 struct page **head;
356
357 BUG_ON(!is_first_page(page));
358
359 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
360 return;
361
362 head = &class->fullness_list[fullness];
363 BUG_ON(!*head);
364 if (list_empty(&(*head)->lru))
365 *head = NULL;
366 else if (*head == page)
367 *head = (struct page *)list_entry((*head)->lru.next,
368 struct page, lru);
369
370 list_del_init(&page->lru);
371}
372
373/*
374 * Each size class maintains zspages in different fullness groups depending
375 * on the number of live objects they contain. When allocating or freeing
376 * objects, the fullness status of the page can change, say, from ALMOST_FULL
377 * to ALMOST_EMPTY when freeing an object. This function checks if such
378 * a status change has occurred for the given page and accordingly moves the
379 * page from the freelist of the old fullness group to that of the new
380 * fullness group.
381 */
382static enum fullness_group fix_fullness_group(struct zs_pool *pool,
383 struct page *page)
384{
385 int class_idx;
386 struct size_class *class;
387 enum fullness_group currfg, newfg;
388
389 BUG_ON(!is_first_page(page));
390
391 get_zspage_mapping(page, &class_idx, &currfg);
392 newfg = get_fullness_group(page);
393 if (newfg == currfg)
394 goto out;
395
396 class = &pool->size_class[class_idx];
397 remove_zspage(page, class, currfg);
398 insert_zspage(page, class, newfg);
399 set_zspage_mapping(page, class_idx, newfg);
400
401out:
402 return newfg;
403}
404
405/*
406 * We have to decide on how many pages to link together
407 * to form a zspage for each size class. This is important
408 * to reduce wastage due to unusable space left at end of
409 * each zspage which is given as:
410 * wastage = Zp - Zp % size_class
411 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
412 *
413 * For example, for size class of 3/8 * PAGE_SIZE, we should
414 * link together 3 PAGE_SIZE sized pages to form a zspage
415 * since then we can perfectly fit in 8 such objects.
416 */
417static int get_pages_per_zspage(int class_size)
418{
419 int i, max_usedpc = 0;
420 /* zspage order which gives maximum used size per KB */
421 int max_usedpc_order = 1;
422
423 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
424 int zspage_size;
425 int waste, usedpc;
426
427 zspage_size = i * PAGE_SIZE;
428 waste = zspage_size % class_size;
429 usedpc = (zspage_size - waste) * 100 / zspage_size;
430
431 if (usedpc > max_usedpc) {
432 max_usedpc = usedpc;
433 max_usedpc_order = i;
434 }
435 }
436
437 return max_usedpc_order;
438}
439
440/*
441 * A single 'zspage' is composed of many system pages which are
442 * linked together using fields in struct page. This function finds
443 * the first/head page, given any component page of a zspage.
444 */
445static struct page *get_first_page(struct page *page)
446{
447 if (is_first_page(page))
448 return page;
449 else
450 return page->first_page;
451}
452
453static struct page *get_next_page(struct page *page)
454{
455 struct page *next;
456
457 if (is_last_page(page))
458 next = NULL;
459 else if (is_first_page(page))
460 next = (struct page *)page_private(page);
461 else
462 next = list_entry(page->lru.next, struct page, lru);
463
464 return next;
465}
466
467/*
468 * Encode <page, obj_idx> as a single handle value.
469 * On hardware platforms with physical memory starting at 0x0 the pfn
470 * could be 0 so we ensure that the handle will never be 0 by adjusting the
471 * encoded obj_idx value before encoding.
472 */
473static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
474{
475 unsigned long handle;
476
477 if (!page) {
478 BUG_ON(obj_idx);
479 return NULL;
480 }
481
482 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
483 handle |= ((obj_idx + 1) & OBJ_INDEX_MASK);
484
485 return (void *)handle;
486}
487
488/*
489 * Decode <page, obj_idx> pair from the given object handle. We adjust the
490 * decoded obj_idx back to its original value since it was adjusted in
491 * obj_location_to_handle().
492 */
493static void obj_handle_to_location(unsigned long handle, struct page **page,
494 unsigned long *obj_idx)
495{
496 *page = pfn_to_page(handle >> OBJ_INDEX_BITS);
497 *obj_idx = (handle & OBJ_INDEX_MASK) - 1;
498}
499
500static unsigned long obj_idx_to_offset(struct page *page,
501 unsigned long obj_idx, int class_size)
502{
503 unsigned long off = 0;
504
505 if (!is_first_page(page))
506 off = page->index;
507
508 return off + obj_idx * class_size;
509}
510
511static void reset_page(struct page *page)
512{
513 clear_bit(PG_private, &page->flags);
514 clear_bit(PG_private_2, &page->flags);
515 set_page_private(page, 0);
516 page->mapping = NULL;
517 page->freelist = NULL;
518 page_mapcount_reset(page);
519}
520
521static void free_zspage(struct page *first_page)
522{
523 struct page *nextp, *tmp, *head_extra;
524
525 BUG_ON(!is_first_page(first_page));
526 BUG_ON(first_page->inuse);
527
528 head_extra = (struct page *)page_private(first_page);
529
530 reset_page(first_page);
531 __free_page(first_page);
532
533 /* zspage with only 1 system page */
534 if (!head_extra)
535 return;
536
537 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
538 list_del(&nextp->lru);
539 reset_page(nextp);
540 __free_page(nextp);
541 }
542 reset_page(head_extra);
543 __free_page(head_extra);
544}
545
546/* Initialize a newly allocated zspage */
547static void init_zspage(struct page *first_page, struct size_class *class)
548{
549 unsigned long off = 0;
550 struct page *page = first_page;
551
552 BUG_ON(!is_first_page(first_page));
553 while (page) {
554 struct page *next_page;
555 struct link_free *link;
556 unsigned int i, objs_on_page;
557
558 /*
559 * page->index stores offset of first object starting
560 * in the page. For the first page, this is always 0,
561 * so we use first_page->index (aka ->freelist) to store
562 * head of corresponding zspage's freelist.
563 */
564 if (page != first_page)
565 page->index = off;
566
567 link = (struct link_free *)kmap_atomic(page) +
568 off / sizeof(*link);
569 objs_on_page = (PAGE_SIZE - off) / class->size;
570
571 for (i = 1; i <= objs_on_page; i++) {
572 off += class->size;
573 if (off < PAGE_SIZE) {
574 link->next = obj_location_to_handle(page, i);
575 link += class->size / sizeof(*link);
576 }
577 }
578
579 /*
580 * We now come to the last (full or partial) object on this
581 * page, which must point to the first object on the next
582 * page (if present)
583 */
584 next_page = get_next_page(page);
585 link->next = obj_location_to_handle(next_page, 0);
586 kunmap_atomic(link);
587 page = next_page;
588 off = (off + class->size) % PAGE_SIZE;
589 }
590}
591
592/*
593 * Allocate a zspage for the given size class
594 */
595static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
596{
597 int i, error;
598 struct page *first_page = NULL, *uninitialized_var(prev_page);
599
600 /*
601 * Allocate individual pages and link them together as:
602 * 1. first page->private = first sub-page
603 * 2. all sub-pages are linked together using page->lru
604 * 3. each sub-page is linked to the first page using page->first_page
605 *
606 * For each size class, First/Head pages are linked together using
607 * page->lru. Also, we set PG_private to identify the first page
608 * (i.e. no other sub-page has this flag set) and PG_private_2 to
609 * identify the last page.
610 */
611 error = -ENOMEM;
612 for (i = 0; i < class->pages_per_zspage; i++) {
613 struct page *page;
614
615 page = alloc_page(flags);
616 if (!page)
617 goto cleanup;
618
619 INIT_LIST_HEAD(&page->lru);
620 if (i == 0) { /* first page */
621 SetPagePrivate(page);
622 set_page_private(page, 0);
623 first_page = page;
624 first_page->inuse = 0;
625 }
626 if (i == 1)
627 set_page_private(first_page, (unsigned long)page);
628 if (i >= 1)
629 page->first_page = first_page;
630 if (i >= 2)
631 list_add(&page->lru, &prev_page->lru);
632 if (i == class->pages_per_zspage - 1) /* last page */
633 SetPagePrivate2(page);
634 prev_page = page;
635 }
636
637 init_zspage(first_page, class);
638
639 first_page->freelist = obj_location_to_handle(first_page, 0);
640 /* Maximum number of objects we can store in this zspage */
641 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
642
643 error = 0; /* Success */
644
645cleanup:
646 if (unlikely(error) && first_page) {
647 free_zspage(first_page);
648 first_page = NULL;
649 }
650
651 return first_page;
652}
653
654static struct page *find_get_zspage(struct size_class *class)
655{
656 int i;
657 struct page *page;
658
659 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
660 page = class->fullness_list[i];
661 if (page)
662 break;
663 }
664
665 return page;
666}
667
668#ifdef CONFIG_PGTABLE_MAPPING
669static inline int __zs_cpu_up(struct mapping_area *area)
670{
671 /*
672 * Make sure we don't leak memory if a cpu UP notification
673 * and zs_init() race and both call zs_cpu_up() on the same cpu
674 */
675 if (area->vm)
676 return 0;
677 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
678 if (!area->vm)
679 return -ENOMEM;
680 return 0;
681}
682
683static inline void __zs_cpu_down(struct mapping_area *area)
684{
685 if (area->vm)
686 free_vm_area(area->vm);
687 area->vm = NULL;
688}
689
690static inline void *__zs_map_object(struct mapping_area *area,
691 struct page *pages[2], int off, int size)
692{
693 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, &pages));
694 area->vm_addr = area->vm->addr;
695 return area->vm_addr + off;
696}
697
698static inline void __zs_unmap_object(struct mapping_area *area,
699 struct page *pages[2], int off, int size)
700{
701 unsigned long addr = (unsigned long)area->vm_addr;
702
703 unmap_kernel_range(addr, PAGE_SIZE * 2);
704}
705
706#else /* CONFIG_PGTABLE_MAPPING */
707
708static inline int __zs_cpu_up(struct mapping_area *area)
709{
710 /*
711 * Make sure we don't leak memory if a cpu UP notification
712 * and zs_init() race and both call zs_cpu_up() on the same cpu
713 */
714 if (area->vm_buf)
715 return 0;
716 area->vm_buf = (char *)__get_free_page(GFP_KERNEL);
717 if (!area->vm_buf)
718 return -ENOMEM;
719 return 0;
720}
721
722static inline void __zs_cpu_down(struct mapping_area *area)
723{
724 if (area->vm_buf)
725 free_page((unsigned long)area->vm_buf);
726 area->vm_buf = NULL;
727}
728
729static void *__zs_map_object(struct mapping_area *area,
730 struct page *pages[2], int off, int size)
731{
732 int sizes[2];
733 void *addr;
734 char *buf = area->vm_buf;
735
736 /* disable page faults to match kmap_atomic() return conditions */
737 pagefault_disable();
738
739 /* no read fastpath */
740 if (area->vm_mm == ZS_MM_WO)
741 goto out;
742
743 sizes[0] = PAGE_SIZE - off;
744 sizes[1] = size - sizes[0];
745
746 /* copy object to per-cpu buffer */
747 addr = kmap_atomic(pages[0]);
748 memcpy(buf, addr + off, sizes[0]);
749 kunmap_atomic(addr);
750 addr = kmap_atomic(pages[1]);
751 memcpy(buf + sizes[0], addr, sizes[1]);
752 kunmap_atomic(addr);
753out:
754 return area->vm_buf;
755}
756
757static void __zs_unmap_object(struct mapping_area *area,
758 struct page *pages[2], int off, int size)
759{
760 int sizes[2];
761 void *addr;
762 char *buf = area->vm_buf;
763
764 /* no write fastpath */
765 if (area->vm_mm == ZS_MM_RO)
766 goto out;
767
768 sizes[0] = PAGE_SIZE - off;
769 sizes[1] = size - sizes[0];
770
771 /* copy per-cpu buffer to object */
772 addr = kmap_atomic(pages[0]);
773 memcpy(addr + off, buf, sizes[0]);
774 kunmap_atomic(addr);
775 addr = kmap_atomic(pages[1]);
776 memcpy(addr, buf + sizes[0], sizes[1]);
777 kunmap_atomic(addr);
778
779out:
780 /* enable page faults to match kunmap_atomic() return conditions */
781 pagefault_enable();
782}
783
784#endif /* CONFIG_PGTABLE_MAPPING */
785
786static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
787 void *pcpu)
788{
789 int ret, cpu = (long)pcpu;
790 struct mapping_area *area;
791
792 switch (action) {
793 case CPU_UP_PREPARE:
794 area = &per_cpu(zs_map_area, cpu);
795 ret = __zs_cpu_up(area);
796 if (ret)
797 return notifier_from_errno(ret);
798 break;
799 case CPU_DEAD:
800 case CPU_UP_CANCELED:
801 area = &per_cpu(zs_map_area, cpu);
802 __zs_cpu_down(area);
803 break;
804 }
805
806 return NOTIFY_OK;
807}
808
809static struct notifier_block zs_cpu_nb = {
810 .notifier_call = zs_cpu_notifier
811};
812
813static void zs_exit(void)
814{
815 int cpu;
816
817 cpu_notifier_register_begin();
818
819 for_each_online_cpu(cpu)
820 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
821 __unregister_cpu_notifier(&zs_cpu_nb);
822
823 cpu_notifier_register_done();
824}
825
826static int zs_init(void)
827{
828 int cpu, ret;
829
830 cpu_notifier_register_begin();
831
832 __register_cpu_notifier(&zs_cpu_nb);
833 for_each_online_cpu(cpu) {
834 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
835 if (notifier_to_errno(ret)) {
836 cpu_notifier_register_done();
837 goto fail;
838 }
839 }
840
841 cpu_notifier_register_done();
842
843 return 0;
844fail:
845 zs_exit();
846 return notifier_to_errno(ret);
847}
848
849/**
850 * zs_create_pool - Creates an allocation pool to work from.
851 * @flags: allocation flags used to allocate pool metadata
852 *
853 * This function must be called before anything when using
854 * the zsmalloc allocator.
855 *
856 * On success, a pointer to the newly created pool is returned,
857 * otherwise NULL.
858 */
859struct zs_pool *zs_create_pool(gfp_t flags)
860{
861 int i, ovhd_size;
862 struct zs_pool *pool;
863
864 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
865 pool = kzalloc(ovhd_size, GFP_KERNEL);
866 if (!pool)
867 return NULL;
868
869 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
870 int size;
871 struct size_class *class;
872
873 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
874 if (size > ZS_MAX_ALLOC_SIZE)
875 size = ZS_MAX_ALLOC_SIZE;
876
877 class = &pool->size_class[i];
878 class->size = size;
879 class->index = i;
880 spin_lock_init(&class->lock);
881 class->pages_per_zspage = get_pages_per_zspage(size);
882
883 }
884
885 pool->flags = flags;
886
887 return pool;
888}
889EXPORT_SYMBOL_GPL(zs_create_pool);
890
891void zs_destroy_pool(struct zs_pool *pool)
892{
893 int i;
894
895 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
896 int fg;
897 struct size_class *class = &pool->size_class[i];
898
899 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
900 if (class->fullness_list[fg]) {
901 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
902 class->size, fg);
903 }
904 }
905 }
906 kfree(pool);
907}
908EXPORT_SYMBOL_GPL(zs_destroy_pool);
909
910/**
911 * zs_malloc - Allocate block of given size from pool.
912 * @pool: pool to allocate from
913 * @size: size of block to allocate
914 *
915 * On success, handle to the allocated object is returned,
916 * otherwise 0.
917 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
918 */
919unsigned long zs_malloc(struct zs_pool *pool, size_t size)
920{
921 unsigned long obj;
922 struct link_free *link;
923 int class_idx;
924 struct size_class *class;
925
926 struct page *first_page, *m_page;
927 unsigned long m_objidx, m_offset;
928
929 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
930 return 0;
931
932 class_idx = get_size_class_index(size);
933 class = &pool->size_class[class_idx];
934 BUG_ON(class_idx != class->index);
935
936 spin_lock(&class->lock);
937 first_page = find_get_zspage(class);
938
939 if (!first_page) {
940 spin_unlock(&class->lock);
941 first_page = alloc_zspage(class, pool->flags);
942 if (unlikely(!first_page))
943 return 0;
944
945 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
946 spin_lock(&class->lock);
947 class->pages_allocated += class->pages_per_zspage;
948 }
949
950 obj = (unsigned long)first_page->freelist;
951 obj_handle_to_location(obj, &m_page, &m_objidx);
952 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
953
954 link = (struct link_free *)kmap_atomic(m_page) +
955 m_offset / sizeof(*link);
956 first_page->freelist = link->next;
957 memset(link, POISON_INUSE, sizeof(*link));
958 kunmap_atomic(link);
959
960 first_page->inuse++;
961 /* Now move the zspage to another fullness group, if required */
962 fix_fullness_group(pool, first_page);
963 spin_unlock(&class->lock);
964
965 return obj;
966}
967EXPORT_SYMBOL_GPL(zs_malloc);
968
969void zs_free(struct zs_pool *pool, unsigned long obj)
970{
971 struct link_free *link;
972 struct page *first_page, *f_page;
973 unsigned long f_objidx, f_offset;
974
975 int class_idx;
976 struct size_class *class;
977 enum fullness_group fullness;
978
979 if (unlikely(!obj))
980 return;
981
982 obj_handle_to_location(obj, &f_page, &f_objidx);
983 first_page = get_first_page(f_page);
984
985 get_zspage_mapping(first_page, &class_idx, &fullness);
986 class = &pool->size_class[class_idx];
987 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
988
989 spin_lock(&class->lock);
990
991 /* Insert this object in containing zspage's freelist */
992 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
993 + f_offset);
994 link->next = first_page->freelist;
995 kunmap_atomic(link);
996 first_page->freelist = (void *)obj;
997
998 first_page->inuse--;
999 fullness = fix_fullness_group(pool, first_page);
1000
1001 if (fullness == ZS_EMPTY)
1002 class->pages_allocated -= class->pages_per_zspage;
1003
1004 spin_unlock(&class->lock);
1005
1006 if (fullness == ZS_EMPTY)
1007 free_zspage(first_page);
1008}
1009EXPORT_SYMBOL_GPL(zs_free);
1010
1011/**
1012 * zs_map_object - get address of allocated object from handle.
1013 * @pool: pool from which the object was allocated
1014 * @handle: handle returned from zs_malloc
1015 *
1016 * Before using an object allocated from zs_malloc, it must be mapped using
1017 * this function. When done with the object, it must be unmapped using
1018 * zs_unmap_object.
1019 *
1020 * Only one object can be mapped per cpu at a time. There is no protection
1021 * against nested mappings.
1022 *
1023 * This function returns with preemption and page faults disabled.
1024 */
1025void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1026 enum zs_mapmode mm)
1027{
1028 struct page *page;
1029 unsigned long obj_idx, off;
1030
1031 unsigned int class_idx;
1032 enum fullness_group fg;
1033 struct size_class *class;
1034 struct mapping_area *area;
1035 struct page *pages[2];
1036
1037 BUG_ON(!handle);
1038
1039 /*
1040 * Because we use per-cpu mapping areas shared among the
1041 * pools/users, we can't allow mapping in interrupt context
1042 * because it can corrupt another users mappings.
1043 */
1044 BUG_ON(in_interrupt());
1045
1046 obj_handle_to_location(handle, &page, &obj_idx);
1047 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1048 class = &pool->size_class[class_idx];
1049 off = obj_idx_to_offset(page, obj_idx, class->size);
1050
1051 area = &get_cpu_var(zs_map_area);
1052 area->vm_mm = mm;
1053 if (off + class->size <= PAGE_SIZE) {
1054 /* this object is contained entirely within a page */
1055 area->vm_addr = kmap_atomic(page);
1056 return area->vm_addr + off;
1057 }
1058
1059 /* this object spans two pages */
1060 pages[0] = page;
1061 pages[1] = get_next_page(page);
1062 BUG_ON(!pages[1]);
1063
1064 return __zs_map_object(area, pages, off, class->size);
1065}
1066EXPORT_SYMBOL_GPL(zs_map_object);
1067
1068void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1069{
1070 struct page *page;
1071 unsigned long obj_idx, off;
1072
1073 unsigned int class_idx;
1074 enum fullness_group fg;
1075 struct size_class *class;
1076 struct mapping_area *area;
1077
1078 BUG_ON(!handle);
1079
1080 obj_handle_to_location(handle, &page, &obj_idx);
1081 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1082 class = &pool->size_class[class_idx];
1083 off = obj_idx_to_offset(page, obj_idx, class->size);
1084
1085 area = &__get_cpu_var(zs_map_area);
1086 if (off + class->size <= PAGE_SIZE)
1087 kunmap_atomic(area->vm_addr);
1088 else {
1089 struct page *pages[2];
1090
1091 pages[0] = page;
1092 pages[1] = get_next_page(page);
1093 BUG_ON(!pages[1]);
1094
1095 __zs_unmap_object(area, pages, off, class->size);
1096 }
1097 put_cpu_var(zs_map_area);
1098}
1099EXPORT_SYMBOL_GPL(zs_unmap_object);
1100
1101u64 zs_get_total_size_bytes(struct zs_pool *pool)
1102{
1103 int i;
1104 u64 npages = 0;
1105
1106 for (i = 0; i < ZS_SIZE_CLASSES; i++)
1107 npages += pool->size_class[i].pages_allocated;
1108
1109 return npages << PAGE_SHIFT;
1110}
1111EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);
1112
1113module_init(zs_init);
1114module_exit(zs_exit);
1115
1116MODULE_LICENSE("Dual BSD/GPL");
1117MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");
1/*
2 * zsmalloc memory allocator
3 *
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
6 *
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
9 *
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
12 */
13
14/*
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
17 *
18 * Usage of struct page fields:
19 * page->private: points to zspage
20 * page->freelist(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->units: first object offset in a subpage of zspage
24 *
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
27 * PG_private2: identifies the last component page
28 * PG_owner_priv_1: indentifies the huge component page
29 *
30 */
31
32#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33
34#include <linux/module.h>
35#include <linux/kernel.h>
36#include <linux/sched.h>
37#include <linux/bitops.h>
38#include <linux/errno.h>
39#include <linux/highmem.h>
40#include <linux/string.h>
41#include <linux/slab.h>
42#include <asm/tlbflush.h>
43#include <asm/pgtable.h>
44#include <linux/cpumask.h>
45#include <linux/cpu.h>
46#include <linux/vmalloc.h>
47#include <linux/preempt.h>
48#include <linux/spinlock.h>
49#include <linux/types.h>
50#include <linux/debugfs.h>
51#include <linux/zsmalloc.h>
52#include <linux/zpool.h>
53#include <linux/mount.h>
54#include <linux/migrate.h>
55#include <linux/pagemap.h>
56
57#define ZSPAGE_MAGIC 0x58
58
59/*
60 * This must be power of 2 and greater than of equal to sizeof(link_free).
61 * These two conditions ensure that any 'struct link_free' itself doesn't
62 * span more than 1 page which avoids complex case of mapping 2 pages simply
63 * to restore link_free pointer values.
64 */
65#define ZS_ALIGN 8
66
67/*
68 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
69 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
70 */
71#define ZS_MAX_ZSPAGE_ORDER 2
72#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
73
74#define ZS_HANDLE_SIZE (sizeof(unsigned long))
75
76/*
77 * Object location (<PFN>, <obj_idx>) is encoded as
78 * as single (unsigned long) handle value.
79 *
80 * Note that object index <obj_idx> starts from 0.
81 *
82 * This is made more complicated by various memory models and PAE.
83 */
84
85#ifndef MAX_PHYSMEM_BITS
86#ifdef CONFIG_HIGHMEM64G
87#define MAX_PHYSMEM_BITS 36
88#else /* !CONFIG_HIGHMEM64G */
89/*
90 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
91 * be PAGE_SHIFT
92 */
93#define MAX_PHYSMEM_BITS BITS_PER_LONG
94#endif
95#endif
96#define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
97
98/*
99 * Memory for allocating for handle keeps object position by
100 * encoding <page, obj_idx> and the encoded value has a room
101 * in least bit(ie, look at obj_to_location).
102 * We use the bit to synchronize between object access by
103 * user and migration.
104 */
105#define HANDLE_PIN_BIT 0
106
107/*
108 * Head in allocated object should have OBJ_ALLOCATED_TAG
109 * to identify the object was allocated or not.
110 * It's okay to add the status bit in the least bit because
111 * header keeps handle which is 4byte-aligned address so we
112 * have room for two bit at least.
113 */
114#define OBJ_ALLOCATED_TAG 1
115#define OBJ_TAG_BITS 1
116#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
117#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
118
119#define MAX(a, b) ((a) >= (b) ? (a) : (b))
120/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
121#define ZS_MIN_ALLOC_SIZE \
122 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
123/* each chunk includes extra space to keep handle */
124#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
125
126/*
127 * On systems with 4K page size, this gives 255 size classes! There is a
128 * trader-off here:
129 * - Large number of size classes is potentially wasteful as free page are
130 * spread across these classes
131 * - Small number of size classes causes large internal fragmentation
132 * - Probably its better to use specific size classes (empirically
133 * determined). NOTE: all those class sizes must be set as multiple of
134 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
135 *
136 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
137 * (reason above)
138 */
139#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
140
141enum fullness_group {
142 ZS_EMPTY,
143 ZS_ALMOST_EMPTY,
144 ZS_ALMOST_FULL,
145 ZS_FULL,
146 NR_ZS_FULLNESS,
147};
148
149enum zs_stat_type {
150 CLASS_EMPTY,
151 CLASS_ALMOST_EMPTY,
152 CLASS_ALMOST_FULL,
153 CLASS_FULL,
154 OBJ_ALLOCATED,
155 OBJ_USED,
156 NR_ZS_STAT_TYPE,
157};
158
159struct zs_size_stat {
160 unsigned long objs[NR_ZS_STAT_TYPE];
161};
162
163#ifdef CONFIG_ZSMALLOC_STAT
164static struct dentry *zs_stat_root;
165#endif
166
167#ifdef CONFIG_COMPACTION
168static struct vfsmount *zsmalloc_mnt;
169#endif
170
171/*
172 * number of size_classes
173 */
174static int zs_size_classes;
175
176/*
177 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
178 * n <= N / f, where
179 * n = number of allocated objects
180 * N = total number of objects zspage can store
181 * f = fullness_threshold_frac
182 *
183 * Similarly, we assign zspage to:
184 * ZS_ALMOST_FULL when n > N / f
185 * ZS_EMPTY when n == 0
186 * ZS_FULL when n == N
187 *
188 * (see: fix_fullness_group())
189 */
190static const int fullness_threshold_frac = 4;
191
192struct size_class {
193 spinlock_t lock;
194 struct list_head fullness_list[NR_ZS_FULLNESS];
195 /*
196 * Size of objects stored in this class. Must be multiple
197 * of ZS_ALIGN.
198 */
199 int size;
200 int objs_per_zspage;
201 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
202 int pages_per_zspage;
203
204 unsigned int index;
205 struct zs_size_stat stats;
206};
207
208/* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
209static void SetPageHugeObject(struct page *page)
210{
211 SetPageOwnerPriv1(page);
212}
213
214static void ClearPageHugeObject(struct page *page)
215{
216 ClearPageOwnerPriv1(page);
217}
218
219static int PageHugeObject(struct page *page)
220{
221 return PageOwnerPriv1(page);
222}
223
224/*
225 * Placed within free objects to form a singly linked list.
226 * For every zspage, zspage->freeobj gives head of this list.
227 *
228 * This must be power of 2 and less than or equal to ZS_ALIGN
229 */
230struct link_free {
231 union {
232 /*
233 * Free object index;
234 * It's valid for non-allocated object
235 */
236 unsigned long next;
237 /*
238 * Handle of allocated object.
239 */
240 unsigned long handle;
241 };
242};
243
244struct zs_pool {
245 const char *name;
246
247 struct size_class **size_class;
248 struct kmem_cache *handle_cachep;
249 struct kmem_cache *zspage_cachep;
250
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#ifdef CONFIG_COMPACTION
266 struct inode *inode;
267 struct work_struct free_work;
268#endif
269};
270
271/*
272 * A zspage's class index and fullness group
273 * are encoded in its (first)page->mapping
274 */
275#define FULLNESS_BITS 2
276#define CLASS_BITS 8
277#define ISOLATED_BITS 3
278#define MAGIC_VAL_BITS 8
279
280struct zspage {
281 struct {
282 unsigned int fullness:FULLNESS_BITS;
283 unsigned int class:CLASS_BITS;
284 unsigned int isolated:ISOLATED_BITS;
285 unsigned int magic:MAGIC_VAL_BITS;
286 };
287 unsigned int inuse;
288 unsigned int freeobj;
289 struct page *first_page;
290 struct list_head list; /* fullness list */
291#ifdef CONFIG_COMPACTION
292 rwlock_t lock;
293#endif
294};
295
296struct mapping_area {
297#ifdef CONFIG_PGTABLE_MAPPING
298 struct vm_struct *vm; /* vm area for mapping object that span pages */
299#else
300 char *vm_buf; /* copy buffer for objects that span pages */
301#endif
302 char *vm_addr; /* address of kmap_atomic()'ed pages */
303 enum zs_mapmode vm_mm; /* mapping mode */
304};
305
306#ifdef CONFIG_COMPACTION
307static int zs_register_migration(struct zs_pool *pool);
308static void zs_unregister_migration(struct zs_pool *pool);
309static void migrate_lock_init(struct zspage *zspage);
310static void migrate_read_lock(struct zspage *zspage);
311static void migrate_read_unlock(struct zspage *zspage);
312static void kick_deferred_free(struct zs_pool *pool);
313static void init_deferred_free(struct zs_pool *pool);
314static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
315#else
316static int zsmalloc_mount(void) { return 0; }
317static void zsmalloc_unmount(void) {}
318static int zs_register_migration(struct zs_pool *pool) { return 0; }
319static void zs_unregister_migration(struct zs_pool *pool) {}
320static void migrate_lock_init(struct zspage *zspage) {}
321static void migrate_read_lock(struct zspage *zspage) {}
322static void migrate_read_unlock(struct zspage *zspage) {}
323static void kick_deferred_free(struct zs_pool *pool) {}
324static void init_deferred_free(struct zs_pool *pool) {}
325static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
326#endif
327
328static int create_cache(struct zs_pool *pool)
329{
330 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
331 0, 0, NULL);
332 if (!pool->handle_cachep)
333 return 1;
334
335 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
336 0, 0, NULL);
337 if (!pool->zspage_cachep) {
338 kmem_cache_destroy(pool->handle_cachep);
339 pool->handle_cachep = NULL;
340 return 1;
341 }
342
343 return 0;
344}
345
346static void destroy_cache(struct zs_pool *pool)
347{
348 kmem_cache_destroy(pool->handle_cachep);
349 kmem_cache_destroy(pool->zspage_cachep);
350}
351
352static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
353{
354 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
355 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
356}
357
358static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
359{
360 kmem_cache_free(pool->handle_cachep, (void *)handle);
361}
362
363static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
364{
365 return kmem_cache_alloc(pool->zspage_cachep,
366 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
367};
368
369static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
370{
371 kmem_cache_free(pool->zspage_cachep, zspage);
372}
373
374static void record_obj(unsigned long handle, unsigned long obj)
375{
376 /*
377 * lsb of @obj represents handle lock while other bits
378 * represent object value the handle is pointing so
379 * updating shouldn't do store tearing.
380 */
381 WRITE_ONCE(*(unsigned long *)handle, obj);
382}
383
384/* zpool driver */
385
386#ifdef CONFIG_ZPOOL
387
388static void *zs_zpool_create(const char *name, gfp_t gfp,
389 const struct zpool_ops *zpool_ops,
390 struct zpool *zpool)
391{
392 /*
393 * Ignore global gfp flags: zs_malloc() may be invoked from
394 * different contexts and its caller must provide a valid
395 * gfp mask.
396 */
397 return zs_create_pool(name);
398}
399
400static void zs_zpool_destroy(void *pool)
401{
402 zs_destroy_pool(pool);
403}
404
405static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
406 unsigned long *handle)
407{
408 *handle = zs_malloc(pool, size, gfp);
409 return *handle ? 0 : -1;
410}
411static void zs_zpool_free(void *pool, unsigned long handle)
412{
413 zs_free(pool, handle);
414}
415
416static int zs_zpool_shrink(void *pool, unsigned int pages,
417 unsigned int *reclaimed)
418{
419 return -EINVAL;
420}
421
422static void *zs_zpool_map(void *pool, unsigned long handle,
423 enum zpool_mapmode mm)
424{
425 enum zs_mapmode zs_mm;
426
427 switch (mm) {
428 case ZPOOL_MM_RO:
429 zs_mm = ZS_MM_RO;
430 break;
431 case ZPOOL_MM_WO:
432 zs_mm = ZS_MM_WO;
433 break;
434 case ZPOOL_MM_RW: /* fallthru */
435 default:
436 zs_mm = ZS_MM_RW;
437 break;
438 }
439
440 return zs_map_object(pool, handle, zs_mm);
441}
442static void zs_zpool_unmap(void *pool, unsigned long handle)
443{
444 zs_unmap_object(pool, handle);
445}
446
447static u64 zs_zpool_total_size(void *pool)
448{
449 return zs_get_total_pages(pool) << PAGE_SHIFT;
450}
451
452static struct zpool_driver zs_zpool_driver = {
453 .type = "zsmalloc",
454 .owner = THIS_MODULE,
455 .create = zs_zpool_create,
456 .destroy = zs_zpool_destroy,
457 .malloc = zs_zpool_malloc,
458 .free = zs_zpool_free,
459 .shrink = zs_zpool_shrink,
460 .map = zs_zpool_map,
461 .unmap = zs_zpool_unmap,
462 .total_size = zs_zpool_total_size,
463};
464
465MODULE_ALIAS("zpool-zsmalloc");
466#endif /* CONFIG_ZPOOL */
467
468/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
469static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
470
471static bool is_zspage_isolated(struct zspage *zspage)
472{
473 return zspage->isolated;
474}
475
476static int is_first_page(struct page *page)
477{
478 return PagePrivate(page);
479}
480
481/* Protected by class->lock */
482static inline int get_zspage_inuse(struct zspage *zspage)
483{
484 return zspage->inuse;
485}
486
487static inline void set_zspage_inuse(struct zspage *zspage, int val)
488{
489 zspage->inuse = val;
490}
491
492static inline void mod_zspage_inuse(struct zspage *zspage, int val)
493{
494 zspage->inuse += val;
495}
496
497static inline struct page *get_first_page(struct zspage *zspage)
498{
499 struct page *first_page = zspage->first_page;
500
501 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
502 return first_page;
503}
504
505static inline int get_first_obj_offset(struct page *page)
506{
507 return page->units;
508}
509
510static inline void set_first_obj_offset(struct page *page, int offset)
511{
512 page->units = offset;
513}
514
515static inline unsigned int get_freeobj(struct zspage *zspage)
516{
517 return zspage->freeobj;
518}
519
520static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
521{
522 zspage->freeobj = obj;
523}
524
525static void get_zspage_mapping(struct zspage *zspage,
526 unsigned int *class_idx,
527 enum fullness_group *fullness)
528{
529 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
530
531 *fullness = zspage->fullness;
532 *class_idx = zspage->class;
533}
534
535static void set_zspage_mapping(struct zspage *zspage,
536 unsigned int class_idx,
537 enum fullness_group fullness)
538{
539 zspage->class = class_idx;
540 zspage->fullness = fullness;
541}
542
543/*
544 * zsmalloc divides the pool into various size classes where each
545 * class maintains a list of zspages where each zspage is divided
546 * into equal sized chunks. Each allocation falls into one of these
547 * classes depending on its size. This function returns index of the
548 * size class which has chunk size big enough to hold the give size.
549 */
550static int get_size_class_index(int size)
551{
552 int idx = 0;
553
554 if (likely(size > ZS_MIN_ALLOC_SIZE))
555 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
556 ZS_SIZE_CLASS_DELTA);
557
558 return min(zs_size_classes - 1, idx);
559}
560
561static inline void zs_stat_inc(struct size_class *class,
562 enum zs_stat_type type, unsigned long cnt)
563{
564 class->stats.objs[type] += cnt;
565}
566
567static inline void zs_stat_dec(struct size_class *class,
568 enum zs_stat_type type, unsigned long cnt)
569{
570 class->stats.objs[type] -= cnt;
571}
572
573static inline unsigned long zs_stat_get(struct size_class *class,
574 enum zs_stat_type type)
575{
576 return class->stats.objs[type];
577}
578
579#ifdef CONFIG_ZSMALLOC_STAT
580
581static void __init zs_stat_init(void)
582{
583 if (!debugfs_initialized()) {
584 pr_warn("debugfs not available, stat dir not created\n");
585 return;
586 }
587
588 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
589 if (!zs_stat_root)
590 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
591}
592
593static void __exit zs_stat_exit(void)
594{
595 debugfs_remove_recursive(zs_stat_root);
596}
597
598static unsigned long zs_can_compact(struct size_class *class);
599
600static int zs_stats_size_show(struct seq_file *s, void *v)
601{
602 int i;
603 struct zs_pool *pool = s->private;
604 struct size_class *class;
605 int objs_per_zspage;
606 unsigned long class_almost_full, class_almost_empty;
607 unsigned long obj_allocated, obj_used, pages_used, freeable;
608 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
609 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
610 unsigned long total_freeable = 0;
611
612 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
613 "class", "size", "almost_full", "almost_empty",
614 "obj_allocated", "obj_used", "pages_used",
615 "pages_per_zspage", "freeable");
616
617 for (i = 0; i < zs_size_classes; i++) {
618 class = pool->size_class[i];
619
620 if (class->index != i)
621 continue;
622
623 spin_lock(&class->lock);
624 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
625 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
626 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
627 obj_used = zs_stat_get(class, OBJ_USED);
628 freeable = zs_can_compact(class);
629 spin_unlock(&class->lock);
630
631 objs_per_zspage = class->objs_per_zspage;
632 pages_used = obj_allocated / objs_per_zspage *
633 class->pages_per_zspage;
634
635 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
636 " %10lu %10lu %16d %8lu\n",
637 i, class->size, class_almost_full, class_almost_empty,
638 obj_allocated, obj_used, pages_used,
639 class->pages_per_zspage, freeable);
640
641 total_class_almost_full += class_almost_full;
642 total_class_almost_empty += class_almost_empty;
643 total_objs += obj_allocated;
644 total_used_objs += obj_used;
645 total_pages += pages_used;
646 total_freeable += freeable;
647 }
648
649 seq_puts(s, "\n");
650 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
651 "Total", "", total_class_almost_full,
652 total_class_almost_empty, total_objs,
653 total_used_objs, total_pages, "", total_freeable);
654
655 return 0;
656}
657
658static int zs_stats_size_open(struct inode *inode, struct file *file)
659{
660 return single_open(file, zs_stats_size_show, inode->i_private);
661}
662
663static const struct file_operations zs_stat_size_ops = {
664 .open = zs_stats_size_open,
665 .read = seq_read,
666 .llseek = seq_lseek,
667 .release = single_release,
668};
669
670static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
671{
672 struct dentry *entry;
673
674 if (!zs_stat_root) {
675 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
676 return;
677 }
678
679 entry = debugfs_create_dir(name, zs_stat_root);
680 if (!entry) {
681 pr_warn("debugfs dir <%s> creation failed\n", name);
682 return;
683 }
684 pool->stat_dentry = entry;
685
686 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
687 pool->stat_dentry, pool, &zs_stat_size_ops);
688 if (!entry) {
689 pr_warn("%s: debugfs file entry <%s> creation failed\n",
690 name, "classes");
691 debugfs_remove_recursive(pool->stat_dentry);
692 pool->stat_dentry = NULL;
693 }
694}
695
696static void zs_pool_stat_destroy(struct zs_pool *pool)
697{
698 debugfs_remove_recursive(pool->stat_dentry);
699}
700
701#else /* CONFIG_ZSMALLOC_STAT */
702static void __init zs_stat_init(void)
703{
704}
705
706static void __exit zs_stat_exit(void)
707{
708}
709
710static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
711{
712}
713
714static inline void zs_pool_stat_destroy(struct zs_pool *pool)
715{
716}
717#endif
718
719
720/*
721 * For each size class, zspages are divided into different groups
722 * depending on how "full" they are. This was done so that we could
723 * easily find empty or nearly empty zspages when we try to shrink
724 * the pool (not yet implemented). This function returns fullness
725 * status of the given page.
726 */
727static enum fullness_group get_fullness_group(struct size_class *class,
728 struct zspage *zspage)
729{
730 int inuse, objs_per_zspage;
731 enum fullness_group fg;
732
733 inuse = get_zspage_inuse(zspage);
734 objs_per_zspage = class->objs_per_zspage;
735
736 if (inuse == 0)
737 fg = ZS_EMPTY;
738 else if (inuse == objs_per_zspage)
739 fg = ZS_FULL;
740 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
741 fg = ZS_ALMOST_EMPTY;
742 else
743 fg = ZS_ALMOST_FULL;
744
745 return fg;
746}
747
748/*
749 * Each size class maintains various freelists and zspages are assigned
750 * to one of these freelists based on the number of live objects they
751 * have. This functions inserts the given zspage into the freelist
752 * identified by <class, fullness_group>.
753 */
754static void insert_zspage(struct size_class *class,
755 struct zspage *zspage,
756 enum fullness_group fullness)
757{
758 struct zspage *head;
759
760 zs_stat_inc(class, fullness, 1);
761 head = list_first_entry_or_null(&class->fullness_list[fullness],
762 struct zspage, list);
763 /*
764 * We want to see more ZS_FULL pages and less almost empty/full.
765 * Put pages with higher ->inuse first.
766 */
767 if (head) {
768 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
769 list_add(&zspage->list, &head->list);
770 return;
771 }
772 }
773 list_add(&zspage->list, &class->fullness_list[fullness]);
774}
775
776/*
777 * This function removes the given zspage from the freelist identified
778 * by <class, fullness_group>.
779 */
780static void remove_zspage(struct size_class *class,
781 struct zspage *zspage,
782 enum fullness_group fullness)
783{
784 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
785 VM_BUG_ON(is_zspage_isolated(zspage));
786
787 list_del_init(&zspage->list);
788 zs_stat_dec(class, fullness, 1);
789}
790
791/*
792 * Each size class maintains zspages in different fullness groups depending
793 * on the number of live objects they contain. When allocating or freeing
794 * objects, the fullness status of the page can change, say, from ALMOST_FULL
795 * to ALMOST_EMPTY when freeing an object. This function checks if such
796 * a status change has occurred for the given page and accordingly moves the
797 * page from the freelist of the old fullness group to that of the new
798 * fullness group.
799 */
800static enum fullness_group fix_fullness_group(struct size_class *class,
801 struct zspage *zspage)
802{
803 int class_idx;
804 enum fullness_group currfg, newfg;
805
806 get_zspage_mapping(zspage, &class_idx, &currfg);
807 newfg = get_fullness_group(class, zspage);
808 if (newfg == currfg)
809 goto out;
810
811 if (!is_zspage_isolated(zspage)) {
812 remove_zspage(class, zspage, currfg);
813 insert_zspage(class, zspage, newfg);
814 }
815
816 set_zspage_mapping(zspage, class_idx, newfg);
817
818out:
819 return newfg;
820}
821
822/*
823 * We have to decide on how many pages to link together
824 * to form a zspage for each size class. This is important
825 * to reduce wastage due to unusable space left at end of
826 * each zspage which is given as:
827 * wastage = Zp % class_size
828 * usage = Zp - wastage
829 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
830 *
831 * For example, for size class of 3/8 * PAGE_SIZE, we should
832 * link together 3 PAGE_SIZE sized pages to form a zspage
833 * since then we can perfectly fit in 8 such objects.
834 */
835static int get_pages_per_zspage(int class_size)
836{
837 int i, max_usedpc = 0;
838 /* zspage order which gives maximum used size per KB */
839 int max_usedpc_order = 1;
840
841 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
842 int zspage_size;
843 int waste, usedpc;
844
845 zspage_size = i * PAGE_SIZE;
846 waste = zspage_size % class_size;
847 usedpc = (zspage_size - waste) * 100 / zspage_size;
848
849 if (usedpc > max_usedpc) {
850 max_usedpc = usedpc;
851 max_usedpc_order = i;
852 }
853 }
854
855 return max_usedpc_order;
856}
857
858static struct zspage *get_zspage(struct page *page)
859{
860 struct zspage *zspage = (struct zspage *)page->private;
861
862 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
863 return zspage;
864}
865
866static struct page *get_next_page(struct page *page)
867{
868 if (unlikely(PageHugeObject(page)))
869 return NULL;
870
871 return page->freelist;
872}
873
874/**
875 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
876 * @page: page object resides in zspage
877 * @obj_idx: object index
878 */
879static void obj_to_location(unsigned long obj, struct page **page,
880 unsigned int *obj_idx)
881{
882 obj >>= OBJ_TAG_BITS;
883 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
884 *obj_idx = (obj & OBJ_INDEX_MASK);
885}
886
887/**
888 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
889 * @page: page object resides in zspage
890 * @obj_idx: object index
891 */
892static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
893{
894 unsigned long obj;
895
896 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
897 obj |= obj_idx & OBJ_INDEX_MASK;
898 obj <<= OBJ_TAG_BITS;
899
900 return obj;
901}
902
903static unsigned long handle_to_obj(unsigned long handle)
904{
905 return *(unsigned long *)handle;
906}
907
908static unsigned long obj_to_head(struct page *page, void *obj)
909{
910 if (unlikely(PageHugeObject(page))) {
911 VM_BUG_ON_PAGE(!is_first_page(page), page);
912 return page->index;
913 } else
914 return *(unsigned long *)obj;
915}
916
917static inline int testpin_tag(unsigned long handle)
918{
919 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
920}
921
922static inline int trypin_tag(unsigned long handle)
923{
924 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
925}
926
927static void pin_tag(unsigned long handle)
928{
929 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
930}
931
932static void unpin_tag(unsigned long handle)
933{
934 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
935}
936
937static void reset_page(struct page *page)
938{
939 __ClearPageMovable(page);
940 ClearPagePrivate(page);
941 ClearPagePrivate2(page);
942 set_page_private(page, 0);
943 page_mapcount_reset(page);
944 ClearPageHugeObject(page);
945 page->freelist = NULL;
946}
947
948/*
949 * To prevent zspage destroy during migration, zspage freeing should
950 * hold locks of all pages in the zspage.
951 */
952void lock_zspage(struct zspage *zspage)
953{
954 struct page *page = get_first_page(zspage);
955
956 do {
957 lock_page(page);
958 } while ((page = get_next_page(page)) != NULL);
959}
960
961int trylock_zspage(struct zspage *zspage)
962{
963 struct page *cursor, *fail;
964
965 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
966 get_next_page(cursor)) {
967 if (!trylock_page(cursor)) {
968 fail = cursor;
969 goto unlock;
970 }
971 }
972
973 return 1;
974unlock:
975 for (cursor = get_first_page(zspage); cursor != fail; cursor =
976 get_next_page(cursor))
977 unlock_page(cursor);
978
979 return 0;
980}
981
982static void __free_zspage(struct zs_pool *pool, struct size_class *class,
983 struct zspage *zspage)
984{
985 struct page *page, *next;
986 enum fullness_group fg;
987 unsigned int class_idx;
988
989 get_zspage_mapping(zspage, &class_idx, &fg);
990
991 assert_spin_locked(&class->lock);
992
993 VM_BUG_ON(get_zspage_inuse(zspage));
994 VM_BUG_ON(fg != ZS_EMPTY);
995
996 next = page = get_first_page(zspage);
997 do {
998 VM_BUG_ON_PAGE(!PageLocked(page), page);
999 next = get_next_page(page);
1000 reset_page(page);
1001 unlock_page(page);
1002 dec_zone_page_state(page, NR_ZSPAGES);
1003 put_page(page);
1004 page = next;
1005 } while (page != NULL);
1006
1007 cache_free_zspage(pool, zspage);
1008
1009 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
1010 atomic_long_sub(class->pages_per_zspage,
1011 &pool->pages_allocated);
1012}
1013
1014static void free_zspage(struct zs_pool *pool, struct size_class *class,
1015 struct zspage *zspage)
1016{
1017 VM_BUG_ON(get_zspage_inuse(zspage));
1018 VM_BUG_ON(list_empty(&zspage->list));
1019
1020 if (!trylock_zspage(zspage)) {
1021 kick_deferred_free(pool);
1022 return;
1023 }
1024
1025 remove_zspage(class, zspage, ZS_EMPTY);
1026 __free_zspage(pool, class, zspage);
1027}
1028
1029/* Initialize a newly allocated zspage */
1030static void init_zspage(struct size_class *class, struct zspage *zspage)
1031{
1032 unsigned int freeobj = 1;
1033 unsigned long off = 0;
1034 struct page *page = get_first_page(zspage);
1035
1036 while (page) {
1037 struct page *next_page;
1038 struct link_free *link;
1039 void *vaddr;
1040
1041 set_first_obj_offset(page, off);
1042
1043 vaddr = kmap_atomic(page);
1044 link = (struct link_free *)vaddr + off / sizeof(*link);
1045
1046 while ((off += class->size) < PAGE_SIZE) {
1047 link->next = freeobj++ << OBJ_TAG_BITS;
1048 link += class->size / sizeof(*link);
1049 }
1050
1051 /*
1052 * We now come to the last (full or partial) object on this
1053 * page, which must point to the first object on the next
1054 * page (if present)
1055 */
1056 next_page = get_next_page(page);
1057 if (next_page) {
1058 link->next = freeobj++ << OBJ_TAG_BITS;
1059 } else {
1060 /*
1061 * Reset OBJ_TAG_BITS bit to last link to tell
1062 * whether it's allocated object or not.
1063 */
1064 link->next = -1 << OBJ_TAG_BITS;
1065 }
1066 kunmap_atomic(vaddr);
1067 page = next_page;
1068 off %= PAGE_SIZE;
1069 }
1070
1071 set_freeobj(zspage, 0);
1072}
1073
1074static void create_page_chain(struct size_class *class, struct zspage *zspage,
1075 struct page *pages[])
1076{
1077 int i;
1078 struct page *page;
1079 struct page *prev_page = NULL;
1080 int nr_pages = class->pages_per_zspage;
1081
1082 /*
1083 * Allocate individual pages and link them together as:
1084 * 1. all pages are linked together using page->freelist
1085 * 2. each sub-page point to zspage using page->private
1086 *
1087 * we set PG_private to identify the first page (i.e. no other sub-page
1088 * has this flag set) and PG_private_2 to identify the last page.
1089 */
1090 for (i = 0; i < nr_pages; i++) {
1091 page = pages[i];
1092 set_page_private(page, (unsigned long)zspage);
1093 page->freelist = NULL;
1094 if (i == 0) {
1095 zspage->first_page = page;
1096 SetPagePrivate(page);
1097 if (unlikely(class->objs_per_zspage == 1 &&
1098 class->pages_per_zspage == 1))
1099 SetPageHugeObject(page);
1100 } else {
1101 prev_page->freelist = page;
1102 }
1103 if (i == nr_pages - 1)
1104 SetPagePrivate2(page);
1105 prev_page = page;
1106 }
1107}
1108
1109/*
1110 * Allocate a zspage for the given size class
1111 */
1112static struct zspage *alloc_zspage(struct zs_pool *pool,
1113 struct size_class *class,
1114 gfp_t gfp)
1115{
1116 int i;
1117 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1118 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1119
1120 if (!zspage)
1121 return NULL;
1122
1123 memset(zspage, 0, sizeof(struct zspage));
1124 zspage->magic = ZSPAGE_MAGIC;
1125 migrate_lock_init(zspage);
1126
1127 for (i = 0; i < class->pages_per_zspage; i++) {
1128 struct page *page;
1129
1130 page = alloc_page(gfp);
1131 if (!page) {
1132 while (--i >= 0) {
1133 dec_zone_page_state(pages[i], NR_ZSPAGES);
1134 __free_page(pages[i]);
1135 }
1136 cache_free_zspage(pool, zspage);
1137 return NULL;
1138 }
1139
1140 inc_zone_page_state(page, NR_ZSPAGES);
1141 pages[i] = page;
1142 }
1143
1144 create_page_chain(class, zspage, pages);
1145 init_zspage(class, zspage);
1146
1147 return zspage;
1148}
1149
1150static struct zspage *find_get_zspage(struct size_class *class)
1151{
1152 int i;
1153 struct zspage *zspage;
1154
1155 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1156 zspage = list_first_entry_or_null(&class->fullness_list[i],
1157 struct zspage, list);
1158 if (zspage)
1159 break;
1160 }
1161
1162 return zspage;
1163}
1164
1165#ifdef CONFIG_PGTABLE_MAPPING
1166static inline int __zs_cpu_up(struct mapping_area *area)
1167{
1168 /*
1169 * Make sure we don't leak memory if a cpu UP notification
1170 * and zs_init() race and both call zs_cpu_up() on the same cpu
1171 */
1172 if (area->vm)
1173 return 0;
1174 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1175 if (!area->vm)
1176 return -ENOMEM;
1177 return 0;
1178}
1179
1180static inline void __zs_cpu_down(struct mapping_area *area)
1181{
1182 if (area->vm)
1183 free_vm_area(area->vm);
1184 area->vm = NULL;
1185}
1186
1187static inline void *__zs_map_object(struct mapping_area *area,
1188 struct page *pages[2], int off, int size)
1189{
1190 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1191 area->vm_addr = area->vm->addr;
1192 return area->vm_addr + off;
1193}
1194
1195static inline void __zs_unmap_object(struct mapping_area *area,
1196 struct page *pages[2], int off, int size)
1197{
1198 unsigned long addr = (unsigned long)area->vm_addr;
1199
1200 unmap_kernel_range(addr, PAGE_SIZE * 2);
1201}
1202
1203#else /* CONFIG_PGTABLE_MAPPING */
1204
1205static inline int __zs_cpu_up(struct mapping_area *area)
1206{
1207 /*
1208 * Make sure we don't leak memory if a cpu UP notification
1209 * and zs_init() race and both call zs_cpu_up() on the same cpu
1210 */
1211 if (area->vm_buf)
1212 return 0;
1213 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1214 if (!area->vm_buf)
1215 return -ENOMEM;
1216 return 0;
1217}
1218
1219static inline void __zs_cpu_down(struct mapping_area *area)
1220{
1221 kfree(area->vm_buf);
1222 area->vm_buf = NULL;
1223}
1224
1225static void *__zs_map_object(struct mapping_area *area,
1226 struct page *pages[2], int off, int size)
1227{
1228 int sizes[2];
1229 void *addr;
1230 char *buf = area->vm_buf;
1231
1232 /* disable page faults to match kmap_atomic() return conditions */
1233 pagefault_disable();
1234
1235 /* no read fastpath */
1236 if (area->vm_mm == ZS_MM_WO)
1237 goto out;
1238
1239 sizes[0] = PAGE_SIZE - off;
1240 sizes[1] = size - sizes[0];
1241
1242 /* copy object to per-cpu buffer */
1243 addr = kmap_atomic(pages[0]);
1244 memcpy(buf, addr + off, sizes[0]);
1245 kunmap_atomic(addr);
1246 addr = kmap_atomic(pages[1]);
1247 memcpy(buf + sizes[0], addr, sizes[1]);
1248 kunmap_atomic(addr);
1249out:
1250 return area->vm_buf;
1251}
1252
1253static void __zs_unmap_object(struct mapping_area *area,
1254 struct page *pages[2], int off, int size)
1255{
1256 int sizes[2];
1257 void *addr;
1258 char *buf;
1259
1260 /* no write fastpath */
1261 if (area->vm_mm == ZS_MM_RO)
1262 goto out;
1263
1264 buf = area->vm_buf;
1265 buf = buf + ZS_HANDLE_SIZE;
1266 size -= ZS_HANDLE_SIZE;
1267 off += ZS_HANDLE_SIZE;
1268
1269 sizes[0] = PAGE_SIZE - off;
1270 sizes[1] = size - sizes[0];
1271
1272 /* copy per-cpu buffer to object */
1273 addr = kmap_atomic(pages[0]);
1274 memcpy(addr + off, buf, sizes[0]);
1275 kunmap_atomic(addr);
1276 addr = kmap_atomic(pages[1]);
1277 memcpy(addr, buf + sizes[0], sizes[1]);
1278 kunmap_atomic(addr);
1279
1280out:
1281 /* enable page faults to match kunmap_atomic() return conditions */
1282 pagefault_enable();
1283}
1284
1285#endif /* CONFIG_PGTABLE_MAPPING */
1286
1287static int zs_cpu_prepare(unsigned int cpu)
1288{
1289 struct mapping_area *area;
1290
1291 area = &per_cpu(zs_map_area, cpu);
1292 return __zs_cpu_up(area);
1293}
1294
1295static int zs_cpu_dead(unsigned int cpu)
1296{
1297 struct mapping_area *area;
1298
1299 area = &per_cpu(zs_map_area, cpu);
1300 __zs_cpu_down(area);
1301 return 0;
1302}
1303
1304static void __init init_zs_size_classes(void)
1305{
1306 int nr;
1307
1308 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1309 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1310 nr += 1;
1311
1312 zs_size_classes = nr;
1313}
1314
1315static bool can_merge(struct size_class *prev, int pages_per_zspage,
1316 int objs_per_zspage)
1317{
1318 if (prev->pages_per_zspage == pages_per_zspage &&
1319 prev->objs_per_zspage == objs_per_zspage)
1320 return true;
1321
1322 return false;
1323}
1324
1325static bool zspage_full(struct size_class *class, struct zspage *zspage)
1326{
1327 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1328}
1329
1330unsigned long zs_get_total_pages(struct zs_pool *pool)
1331{
1332 return atomic_long_read(&pool->pages_allocated);
1333}
1334EXPORT_SYMBOL_GPL(zs_get_total_pages);
1335
1336/**
1337 * zs_map_object - get address of allocated object from handle.
1338 * @pool: pool from which the object was allocated
1339 * @handle: handle returned from zs_malloc
1340 *
1341 * Before using an object allocated from zs_malloc, it must be mapped using
1342 * this function. When done with the object, it must be unmapped using
1343 * zs_unmap_object.
1344 *
1345 * Only one object can be mapped per cpu at a time. There is no protection
1346 * against nested mappings.
1347 *
1348 * This function returns with preemption and page faults disabled.
1349 */
1350void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1351 enum zs_mapmode mm)
1352{
1353 struct zspage *zspage;
1354 struct page *page;
1355 unsigned long obj, off;
1356 unsigned int obj_idx;
1357
1358 unsigned int class_idx;
1359 enum fullness_group fg;
1360 struct size_class *class;
1361 struct mapping_area *area;
1362 struct page *pages[2];
1363 void *ret;
1364
1365 /*
1366 * Because we use per-cpu mapping areas shared among the
1367 * pools/users, we can't allow mapping in interrupt context
1368 * because it can corrupt another users mappings.
1369 */
1370 WARN_ON_ONCE(in_interrupt());
1371
1372 /* From now on, migration cannot move the object */
1373 pin_tag(handle);
1374
1375 obj = handle_to_obj(handle);
1376 obj_to_location(obj, &page, &obj_idx);
1377 zspage = get_zspage(page);
1378
1379 /* migration cannot move any subpage in this zspage */
1380 migrate_read_lock(zspage);
1381
1382 get_zspage_mapping(zspage, &class_idx, &fg);
1383 class = pool->size_class[class_idx];
1384 off = (class->size * obj_idx) & ~PAGE_MASK;
1385
1386 area = &get_cpu_var(zs_map_area);
1387 area->vm_mm = mm;
1388 if (off + class->size <= PAGE_SIZE) {
1389 /* this object is contained entirely within a page */
1390 area->vm_addr = kmap_atomic(page);
1391 ret = area->vm_addr + off;
1392 goto out;
1393 }
1394
1395 /* this object spans two pages */
1396 pages[0] = page;
1397 pages[1] = get_next_page(page);
1398 BUG_ON(!pages[1]);
1399
1400 ret = __zs_map_object(area, pages, off, class->size);
1401out:
1402 if (likely(!PageHugeObject(page)))
1403 ret += ZS_HANDLE_SIZE;
1404
1405 return ret;
1406}
1407EXPORT_SYMBOL_GPL(zs_map_object);
1408
1409void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1410{
1411 struct zspage *zspage;
1412 struct page *page;
1413 unsigned long obj, off;
1414 unsigned int obj_idx;
1415
1416 unsigned int class_idx;
1417 enum fullness_group fg;
1418 struct size_class *class;
1419 struct mapping_area *area;
1420
1421 obj = handle_to_obj(handle);
1422 obj_to_location(obj, &page, &obj_idx);
1423 zspage = get_zspage(page);
1424 get_zspage_mapping(zspage, &class_idx, &fg);
1425 class = pool->size_class[class_idx];
1426 off = (class->size * obj_idx) & ~PAGE_MASK;
1427
1428 area = this_cpu_ptr(&zs_map_area);
1429 if (off + class->size <= PAGE_SIZE)
1430 kunmap_atomic(area->vm_addr);
1431 else {
1432 struct page *pages[2];
1433
1434 pages[0] = page;
1435 pages[1] = get_next_page(page);
1436 BUG_ON(!pages[1]);
1437
1438 __zs_unmap_object(area, pages, off, class->size);
1439 }
1440 put_cpu_var(zs_map_area);
1441
1442 migrate_read_unlock(zspage);
1443 unpin_tag(handle);
1444}
1445EXPORT_SYMBOL_GPL(zs_unmap_object);
1446
1447static unsigned long obj_malloc(struct size_class *class,
1448 struct zspage *zspage, unsigned long handle)
1449{
1450 int i, nr_page, offset;
1451 unsigned long obj;
1452 struct link_free *link;
1453
1454 struct page *m_page;
1455 unsigned long m_offset;
1456 void *vaddr;
1457
1458 handle |= OBJ_ALLOCATED_TAG;
1459 obj = get_freeobj(zspage);
1460
1461 offset = obj * class->size;
1462 nr_page = offset >> PAGE_SHIFT;
1463 m_offset = offset & ~PAGE_MASK;
1464 m_page = get_first_page(zspage);
1465
1466 for (i = 0; i < nr_page; i++)
1467 m_page = get_next_page(m_page);
1468
1469 vaddr = kmap_atomic(m_page);
1470 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1471 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1472 if (likely(!PageHugeObject(m_page)))
1473 /* record handle in the header of allocated chunk */
1474 link->handle = handle;
1475 else
1476 /* record handle to page->index */
1477 zspage->first_page->index = handle;
1478
1479 kunmap_atomic(vaddr);
1480 mod_zspage_inuse(zspage, 1);
1481 zs_stat_inc(class, OBJ_USED, 1);
1482
1483 obj = location_to_obj(m_page, obj);
1484
1485 return obj;
1486}
1487
1488
1489/**
1490 * zs_malloc - Allocate block of given size from pool.
1491 * @pool: pool to allocate from
1492 * @size: size of block to allocate
1493 * @gfp: gfp flags when allocating object
1494 *
1495 * On success, handle to the allocated object is returned,
1496 * otherwise 0.
1497 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1498 */
1499unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1500{
1501 unsigned long handle, obj;
1502 struct size_class *class;
1503 enum fullness_group newfg;
1504 struct zspage *zspage;
1505
1506 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1507 return 0;
1508
1509 handle = cache_alloc_handle(pool, gfp);
1510 if (!handle)
1511 return 0;
1512
1513 /* extra space in chunk to keep the handle */
1514 size += ZS_HANDLE_SIZE;
1515 class = pool->size_class[get_size_class_index(size)];
1516
1517 spin_lock(&class->lock);
1518 zspage = find_get_zspage(class);
1519 if (likely(zspage)) {
1520 obj = obj_malloc(class, zspage, handle);
1521 /* Now move the zspage to another fullness group, if required */
1522 fix_fullness_group(class, zspage);
1523 record_obj(handle, obj);
1524 spin_unlock(&class->lock);
1525
1526 return handle;
1527 }
1528
1529 spin_unlock(&class->lock);
1530
1531 zspage = alloc_zspage(pool, class, gfp);
1532 if (!zspage) {
1533 cache_free_handle(pool, handle);
1534 return 0;
1535 }
1536
1537 spin_lock(&class->lock);
1538 obj = obj_malloc(class, zspage, handle);
1539 newfg = get_fullness_group(class, zspage);
1540 insert_zspage(class, zspage, newfg);
1541 set_zspage_mapping(zspage, class->index, newfg);
1542 record_obj(handle, obj);
1543 atomic_long_add(class->pages_per_zspage,
1544 &pool->pages_allocated);
1545 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1546
1547 /* We completely set up zspage so mark them as movable */
1548 SetZsPageMovable(pool, zspage);
1549 spin_unlock(&class->lock);
1550
1551 return handle;
1552}
1553EXPORT_SYMBOL_GPL(zs_malloc);
1554
1555static void obj_free(struct size_class *class, unsigned long obj)
1556{
1557 struct link_free *link;
1558 struct zspage *zspage;
1559 struct page *f_page;
1560 unsigned long f_offset;
1561 unsigned int f_objidx;
1562 void *vaddr;
1563
1564 obj &= ~OBJ_ALLOCATED_TAG;
1565 obj_to_location(obj, &f_page, &f_objidx);
1566 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1567 zspage = get_zspage(f_page);
1568
1569 vaddr = kmap_atomic(f_page);
1570
1571 /* Insert this object in containing zspage's freelist */
1572 link = (struct link_free *)(vaddr + f_offset);
1573 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1574 kunmap_atomic(vaddr);
1575 set_freeobj(zspage, f_objidx);
1576 mod_zspage_inuse(zspage, -1);
1577 zs_stat_dec(class, OBJ_USED, 1);
1578}
1579
1580void zs_free(struct zs_pool *pool, unsigned long handle)
1581{
1582 struct zspage *zspage;
1583 struct page *f_page;
1584 unsigned long obj;
1585 unsigned int f_objidx;
1586 int class_idx;
1587 struct size_class *class;
1588 enum fullness_group fullness;
1589 bool isolated;
1590
1591 if (unlikely(!handle))
1592 return;
1593
1594 pin_tag(handle);
1595 obj = handle_to_obj(handle);
1596 obj_to_location(obj, &f_page, &f_objidx);
1597 zspage = get_zspage(f_page);
1598
1599 migrate_read_lock(zspage);
1600
1601 get_zspage_mapping(zspage, &class_idx, &fullness);
1602 class = pool->size_class[class_idx];
1603
1604 spin_lock(&class->lock);
1605 obj_free(class, obj);
1606 fullness = fix_fullness_group(class, zspage);
1607 if (fullness != ZS_EMPTY) {
1608 migrate_read_unlock(zspage);
1609 goto out;
1610 }
1611
1612 isolated = is_zspage_isolated(zspage);
1613 migrate_read_unlock(zspage);
1614 /* If zspage is isolated, zs_page_putback will free the zspage */
1615 if (likely(!isolated))
1616 free_zspage(pool, class, zspage);
1617out:
1618
1619 spin_unlock(&class->lock);
1620 unpin_tag(handle);
1621 cache_free_handle(pool, handle);
1622}
1623EXPORT_SYMBOL_GPL(zs_free);
1624
1625static void zs_object_copy(struct size_class *class, unsigned long dst,
1626 unsigned long src)
1627{
1628 struct page *s_page, *d_page;
1629 unsigned int s_objidx, d_objidx;
1630 unsigned long s_off, d_off;
1631 void *s_addr, *d_addr;
1632 int s_size, d_size, size;
1633 int written = 0;
1634
1635 s_size = d_size = class->size;
1636
1637 obj_to_location(src, &s_page, &s_objidx);
1638 obj_to_location(dst, &d_page, &d_objidx);
1639
1640 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1641 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1642
1643 if (s_off + class->size > PAGE_SIZE)
1644 s_size = PAGE_SIZE - s_off;
1645
1646 if (d_off + class->size > PAGE_SIZE)
1647 d_size = PAGE_SIZE - d_off;
1648
1649 s_addr = kmap_atomic(s_page);
1650 d_addr = kmap_atomic(d_page);
1651
1652 while (1) {
1653 size = min(s_size, d_size);
1654 memcpy(d_addr + d_off, s_addr + s_off, size);
1655 written += size;
1656
1657 if (written == class->size)
1658 break;
1659
1660 s_off += size;
1661 s_size -= size;
1662 d_off += size;
1663 d_size -= size;
1664
1665 if (s_off >= PAGE_SIZE) {
1666 kunmap_atomic(d_addr);
1667 kunmap_atomic(s_addr);
1668 s_page = get_next_page(s_page);
1669 s_addr = kmap_atomic(s_page);
1670 d_addr = kmap_atomic(d_page);
1671 s_size = class->size - written;
1672 s_off = 0;
1673 }
1674
1675 if (d_off >= PAGE_SIZE) {
1676 kunmap_atomic(d_addr);
1677 d_page = get_next_page(d_page);
1678 d_addr = kmap_atomic(d_page);
1679 d_size = class->size - written;
1680 d_off = 0;
1681 }
1682 }
1683
1684 kunmap_atomic(d_addr);
1685 kunmap_atomic(s_addr);
1686}
1687
1688/*
1689 * Find alloced object in zspage from index object and
1690 * return handle.
1691 */
1692static unsigned long find_alloced_obj(struct size_class *class,
1693 struct page *page, int *obj_idx)
1694{
1695 unsigned long head;
1696 int offset = 0;
1697 int index = *obj_idx;
1698 unsigned long handle = 0;
1699 void *addr = kmap_atomic(page);
1700
1701 offset = get_first_obj_offset(page);
1702 offset += class->size * index;
1703
1704 while (offset < PAGE_SIZE) {
1705 head = obj_to_head(page, addr + offset);
1706 if (head & OBJ_ALLOCATED_TAG) {
1707 handle = head & ~OBJ_ALLOCATED_TAG;
1708 if (trypin_tag(handle))
1709 break;
1710 handle = 0;
1711 }
1712
1713 offset += class->size;
1714 index++;
1715 }
1716
1717 kunmap_atomic(addr);
1718
1719 *obj_idx = index;
1720
1721 return handle;
1722}
1723
1724struct zs_compact_control {
1725 /* Source spage for migration which could be a subpage of zspage */
1726 struct page *s_page;
1727 /* Destination page for migration which should be a first page
1728 * of zspage. */
1729 struct page *d_page;
1730 /* Starting object index within @s_page which used for live object
1731 * in the subpage. */
1732 int obj_idx;
1733};
1734
1735static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1736 struct zs_compact_control *cc)
1737{
1738 unsigned long used_obj, free_obj;
1739 unsigned long handle;
1740 struct page *s_page = cc->s_page;
1741 struct page *d_page = cc->d_page;
1742 int obj_idx = cc->obj_idx;
1743 int ret = 0;
1744
1745 while (1) {
1746 handle = find_alloced_obj(class, s_page, &obj_idx);
1747 if (!handle) {
1748 s_page = get_next_page(s_page);
1749 if (!s_page)
1750 break;
1751 obj_idx = 0;
1752 continue;
1753 }
1754
1755 /* Stop if there is no more space */
1756 if (zspage_full(class, get_zspage(d_page))) {
1757 unpin_tag(handle);
1758 ret = -ENOMEM;
1759 break;
1760 }
1761
1762 used_obj = handle_to_obj(handle);
1763 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1764 zs_object_copy(class, free_obj, used_obj);
1765 obj_idx++;
1766 /*
1767 * record_obj updates handle's value to free_obj and it will
1768 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1769 * breaks synchronization using pin_tag(e,g, zs_free) so
1770 * let's keep the lock bit.
1771 */
1772 free_obj |= BIT(HANDLE_PIN_BIT);
1773 record_obj(handle, free_obj);
1774 unpin_tag(handle);
1775 obj_free(class, used_obj);
1776 }
1777
1778 /* Remember last position in this iteration */
1779 cc->s_page = s_page;
1780 cc->obj_idx = obj_idx;
1781
1782 return ret;
1783}
1784
1785static struct zspage *isolate_zspage(struct size_class *class, bool source)
1786{
1787 int i;
1788 struct zspage *zspage;
1789 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1790
1791 if (!source) {
1792 fg[0] = ZS_ALMOST_FULL;
1793 fg[1] = ZS_ALMOST_EMPTY;
1794 }
1795
1796 for (i = 0; i < 2; i++) {
1797 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1798 struct zspage, list);
1799 if (zspage) {
1800 VM_BUG_ON(is_zspage_isolated(zspage));
1801 remove_zspage(class, zspage, fg[i]);
1802 return zspage;
1803 }
1804 }
1805
1806 return zspage;
1807}
1808
1809/*
1810 * putback_zspage - add @zspage into right class's fullness list
1811 * @class: destination class
1812 * @zspage: target page
1813 *
1814 * Return @zspage's fullness_group
1815 */
1816static enum fullness_group putback_zspage(struct size_class *class,
1817 struct zspage *zspage)
1818{
1819 enum fullness_group fullness;
1820
1821 VM_BUG_ON(is_zspage_isolated(zspage));
1822
1823 fullness = get_fullness_group(class, zspage);
1824 insert_zspage(class, zspage, fullness);
1825 set_zspage_mapping(zspage, class->index, fullness);
1826
1827 return fullness;
1828}
1829
1830#ifdef CONFIG_COMPACTION
1831static struct dentry *zs_mount(struct file_system_type *fs_type,
1832 int flags, const char *dev_name, void *data)
1833{
1834 static const struct dentry_operations ops = {
1835 .d_dname = simple_dname,
1836 };
1837
1838 return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1839}
1840
1841static struct file_system_type zsmalloc_fs = {
1842 .name = "zsmalloc",
1843 .mount = zs_mount,
1844 .kill_sb = kill_anon_super,
1845};
1846
1847static int zsmalloc_mount(void)
1848{
1849 int ret = 0;
1850
1851 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1852 if (IS_ERR(zsmalloc_mnt))
1853 ret = PTR_ERR(zsmalloc_mnt);
1854
1855 return ret;
1856}
1857
1858static void zsmalloc_unmount(void)
1859{
1860 kern_unmount(zsmalloc_mnt);
1861}
1862
1863static void migrate_lock_init(struct zspage *zspage)
1864{
1865 rwlock_init(&zspage->lock);
1866}
1867
1868static void migrate_read_lock(struct zspage *zspage)
1869{
1870 read_lock(&zspage->lock);
1871}
1872
1873static void migrate_read_unlock(struct zspage *zspage)
1874{
1875 read_unlock(&zspage->lock);
1876}
1877
1878static void migrate_write_lock(struct zspage *zspage)
1879{
1880 write_lock(&zspage->lock);
1881}
1882
1883static void migrate_write_unlock(struct zspage *zspage)
1884{
1885 write_unlock(&zspage->lock);
1886}
1887
1888/* Number of isolated subpage for *page migration* in this zspage */
1889static void inc_zspage_isolation(struct zspage *zspage)
1890{
1891 zspage->isolated++;
1892}
1893
1894static void dec_zspage_isolation(struct zspage *zspage)
1895{
1896 zspage->isolated--;
1897}
1898
1899static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1900 struct page *newpage, struct page *oldpage)
1901{
1902 struct page *page;
1903 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1904 int idx = 0;
1905
1906 page = get_first_page(zspage);
1907 do {
1908 if (page == oldpage)
1909 pages[idx] = newpage;
1910 else
1911 pages[idx] = page;
1912 idx++;
1913 } while ((page = get_next_page(page)) != NULL);
1914
1915 create_page_chain(class, zspage, pages);
1916 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1917 if (unlikely(PageHugeObject(oldpage)))
1918 newpage->index = oldpage->index;
1919 __SetPageMovable(newpage, page_mapping(oldpage));
1920}
1921
1922bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1923{
1924 struct zs_pool *pool;
1925 struct size_class *class;
1926 int class_idx;
1927 enum fullness_group fullness;
1928 struct zspage *zspage;
1929 struct address_space *mapping;
1930
1931 /*
1932 * Page is locked so zspage couldn't be destroyed. For detail, look at
1933 * lock_zspage in free_zspage.
1934 */
1935 VM_BUG_ON_PAGE(!PageMovable(page), page);
1936 VM_BUG_ON_PAGE(PageIsolated(page), page);
1937
1938 zspage = get_zspage(page);
1939
1940 /*
1941 * Without class lock, fullness could be stale while class_idx is okay
1942 * because class_idx is constant unless page is freed so we should get
1943 * fullness again under class lock.
1944 */
1945 get_zspage_mapping(zspage, &class_idx, &fullness);
1946 mapping = page_mapping(page);
1947 pool = mapping->private_data;
1948 class = pool->size_class[class_idx];
1949
1950 spin_lock(&class->lock);
1951 if (get_zspage_inuse(zspage) == 0) {
1952 spin_unlock(&class->lock);
1953 return false;
1954 }
1955
1956 /* zspage is isolated for object migration */
1957 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1958 spin_unlock(&class->lock);
1959 return false;
1960 }
1961
1962 /*
1963 * If this is first time isolation for the zspage, isolate zspage from
1964 * size_class to prevent further object allocation from the zspage.
1965 */
1966 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1967 get_zspage_mapping(zspage, &class_idx, &fullness);
1968 remove_zspage(class, zspage, fullness);
1969 }
1970
1971 inc_zspage_isolation(zspage);
1972 spin_unlock(&class->lock);
1973
1974 return true;
1975}
1976
1977int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1978 struct page *page, enum migrate_mode mode)
1979{
1980 struct zs_pool *pool;
1981 struct size_class *class;
1982 int class_idx;
1983 enum fullness_group fullness;
1984 struct zspage *zspage;
1985 struct page *dummy;
1986 void *s_addr, *d_addr, *addr;
1987 int offset, pos;
1988 unsigned long handle, head;
1989 unsigned long old_obj, new_obj;
1990 unsigned int obj_idx;
1991 int ret = -EAGAIN;
1992
1993 VM_BUG_ON_PAGE(!PageMovable(page), page);
1994 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1995
1996 zspage = get_zspage(page);
1997
1998 /* Concurrent compactor cannot migrate any subpage in zspage */
1999 migrate_write_lock(zspage);
2000 get_zspage_mapping(zspage, &class_idx, &fullness);
2001 pool = mapping->private_data;
2002 class = pool->size_class[class_idx];
2003 offset = get_first_obj_offset(page);
2004
2005 spin_lock(&class->lock);
2006 if (!get_zspage_inuse(zspage)) {
2007 ret = -EBUSY;
2008 goto unlock_class;
2009 }
2010
2011 pos = offset;
2012 s_addr = kmap_atomic(page);
2013 while (pos < PAGE_SIZE) {
2014 head = obj_to_head(page, s_addr + pos);
2015 if (head & OBJ_ALLOCATED_TAG) {
2016 handle = head & ~OBJ_ALLOCATED_TAG;
2017 if (!trypin_tag(handle))
2018 goto unpin_objects;
2019 }
2020 pos += class->size;
2021 }
2022
2023 /*
2024 * Here, any user cannot access all objects in the zspage so let's move.
2025 */
2026 d_addr = kmap_atomic(newpage);
2027 memcpy(d_addr, s_addr, PAGE_SIZE);
2028 kunmap_atomic(d_addr);
2029
2030 for (addr = s_addr + offset; addr < s_addr + pos;
2031 addr += class->size) {
2032 head = obj_to_head(page, addr);
2033 if (head & OBJ_ALLOCATED_TAG) {
2034 handle = head & ~OBJ_ALLOCATED_TAG;
2035 if (!testpin_tag(handle))
2036 BUG();
2037
2038 old_obj = handle_to_obj(handle);
2039 obj_to_location(old_obj, &dummy, &obj_idx);
2040 new_obj = (unsigned long)location_to_obj(newpage,
2041 obj_idx);
2042 new_obj |= BIT(HANDLE_PIN_BIT);
2043 record_obj(handle, new_obj);
2044 }
2045 }
2046
2047 replace_sub_page(class, zspage, newpage, page);
2048 get_page(newpage);
2049
2050 dec_zspage_isolation(zspage);
2051
2052 /*
2053 * Page migration is done so let's putback isolated zspage to
2054 * the list if @page is final isolated subpage in the zspage.
2055 */
2056 if (!is_zspage_isolated(zspage))
2057 putback_zspage(class, zspage);
2058
2059 reset_page(page);
2060 put_page(page);
2061 page = newpage;
2062
2063 ret = MIGRATEPAGE_SUCCESS;
2064unpin_objects:
2065 for (addr = s_addr + offset; addr < s_addr + pos;
2066 addr += class->size) {
2067 head = obj_to_head(page, addr);
2068 if (head & OBJ_ALLOCATED_TAG) {
2069 handle = head & ~OBJ_ALLOCATED_TAG;
2070 if (!testpin_tag(handle))
2071 BUG();
2072 unpin_tag(handle);
2073 }
2074 }
2075 kunmap_atomic(s_addr);
2076unlock_class:
2077 spin_unlock(&class->lock);
2078 migrate_write_unlock(zspage);
2079
2080 return ret;
2081}
2082
2083void zs_page_putback(struct page *page)
2084{
2085 struct zs_pool *pool;
2086 struct size_class *class;
2087 int class_idx;
2088 enum fullness_group fg;
2089 struct address_space *mapping;
2090 struct zspage *zspage;
2091
2092 VM_BUG_ON_PAGE(!PageMovable(page), page);
2093 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2094
2095 zspage = get_zspage(page);
2096 get_zspage_mapping(zspage, &class_idx, &fg);
2097 mapping = page_mapping(page);
2098 pool = mapping->private_data;
2099 class = pool->size_class[class_idx];
2100
2101 spin_lock(&class->lock);
2102 dec_zspage_isolation(zspage);
2103 if (!is_zspage_isolated(zspage)) {
2104 fg = putback_zspage(class, zspage);
2105 /*
2106 * Due to page_lock, we cannot free zspage immediately
2107 * so let's defer.
2108 */
2109 if (fg == ZS_EMPTY)
2110 schedule_work(&pool->free_work);
2111 }
2112 spin_unlock(&class->lock);
2113}
2114
2115const struct address_space_operations zsmalloc_aops = {
2116 .isolate_page = zs_page_isolate,
2117 .migratepage = zs_page_migrate,
2118 .putback_page = zs_page_putback,
2119};
2120
2121static int zs_register_migration(struct zs_pool *pool)
2122{
2123 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2124 if (IS_ERR(pool->inode)) {
2125 pool->inode = NULL;
2126 return 1;
2127 }
2128
2129 pool->inode->i_mapping->private_data = pool;
2130 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2131 return 0;
2132}
2133
2134static void zs_unregister_migration(struct zs_pool *pool)
2135{
2136 flush_work(&pool->free_work);
2137 iput(pool->inode);
2138}
2139
2140/*
2141 * Caller should hold page_lock of all pages in the zspage
2142 * In here, we cannot use zspage meta data.
2143 */
2144static void async_free_zspage(struct work_struct *work)
2145{
2146 int i;
2147 struct size_class *class;
2148 unsigned int class_idx;
2149 enum fullness_group fullness;
2150 struct zspage *zspage, *tmp;
2151 LIST_HEAD(free_pages);
2152 struct zs_pool *pool = container_of(work, struct zs_pool,
2153 free_work);
2154
2155 for (i = 0; i < zs_size_classes; i++) {
2156 class = pool->size_class[i];
2157 if (class->index != i)
2158 continue;
2159
2160 spin_lock(&class->lock);
2161 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2162 spin_unlock(&class->lock);
2163 }
2164
2165
2166 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2167 list_del(&zspage->list);
2168 lock_zspage(zspage);
2169
2170 get_zspage_mapping(zspage, &class_idx, &fullness);
2171 VM_BUG_ON(fullness != ZS_EMPTY);
2172 class = pool->size_class[class_idx];
2173 spin_lock(&class->lock);
2174 __free_zspage(pool, pool->size_class[class_idx], zspage);
2175 spin_unlock(&class->lock);
2176 }
2177};
2178
2179static void kick_deferred_free(struct zs_pool *pool)
2180{
2181 schedule_work(&pool->free_work);
2182}
2183
2184static void init_deferred_free(struct zs_pool *pool)
2185{
2186 INIT_WORK(&pool->free_work, async_free_zspage);
2187}
2188
2189static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2190{
2191 struct page *page = get_first_page(zspage);
2192
2193 do {
2194 WARN_ON(!trylock_page(page));
2195 __SetPageMovable(page, pool->inode->i_mapping);
2196 unlock_page(page);
2197 } while ((page = get_next_page(page)) != NULL);
2198}
2199#endif
2200
2201/*
2202 *
2203 * Based on the number of unused allocated objects calculate
2204 * and return the number of pages that we can free.
2205 */
2206static unsigned long zs_can_compact(struct size_class *class)
2207{
2208 unsigned long obj_wasted;
2209 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2210 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2211
2212 if (obj_allocated <= obj_used)
2213 return 0;
2214
2215 obj_wasted = obj_allocated - obj_used;
2216 obj_wasted /= class->objs_per_zspage;
2217
2218 return obj_wasted * class->pages_per_zspage;
2219}
2220
2221static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2222{
2223 struct zs_compact_control cc;
2224 struct zspage *src_zspage;
2225 struct zspage *dst_zspage = NULL;
2226
2227 spin_lock(&class->lock);
2228 while ((src_zspage = isolate_zspage(class, true))) {
2229
2230 if (!zs_can_compact(class))
2231 break;
2232
2233 cc.obj_idx = 0;
2234 cc.s_page = get_first_page(src_zspage);
2235
2236 while ((dst_zspage = isolate_zspage(class, false))) {
2237 cc.d_page = get_first_page(dst_zspage);
2238 /*
2239 * If there is no more space in dst_page, resched
2240 * and see if anyone had allocated another zspage.
2241 */
2242 if (!migrate_zspage(pool, class, &cc))
2243 break;
2244
2245 putback_zspage(class, dst_zspage);
2246 }
2247
2248 /* Stop if we couldn't find slot */
2249 if (dst_zspage == NULL)
2250 break;
2251
2252 putback_zspage(class, dst_zspage);
2253 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2254 free_zspage(pool, class, src_zspage);
2255 pool->stats.pages_compacted += class->pages_per_zspage;
2256 }
2257 spin_unlock(&class->lock);
2258 cond_resched();
2259 spin_lock(&class->lock);
2260 }
2261
2262 if (src_zspage)
2263 putback_zspage(class, src_zspage);
2264
2265 spin_unlock(&class->lock);
2266}
2267
2268unsigned long zs_compact(struct zs_pool *pool)
2269{
2270 int i;
2271 struct size_class *class;
2272
2273 for (i = zs_size_classes - 1; i >= 0; i--) {
2274 class = pool->size_class[i];
2275 if (!class)
2276 continue;
2277 if (class->index != i)
2278 continue;
2279 __zs_compact(pool, class);
2280 }
2281
2282 return pool->stats.pages_compacted;
2283}
2284EXPORT_SYMBOL_GPL(zs_compact);
2285
2286void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2287{
2288 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2289}
2290EXPORT_SYMBOL_GPL(zs_pool_stats);
2291
2292static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2293 struct shrink_control *sc)
2294{
2295 unsigned long pages_freed;
2296 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2297 shrinker);
2298
2299 pages_freed = pool->stats.pages_compacted;
2300 /*
2301 * Compact classes and calculate compaction delta.
2302 * Can run concurrently with a manually triggered
2303 * (by user) compaction.
2304 */
2305 pages_freed = zs_compact(pool) - pages_freed;
2306
2307 return pages_freed ? pages_freed : SHRINK_STOP;
2308}
2309
2310static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2311 struct shrink_control *sc)
2312{
2313 int i;
2314 struct size_class *class;
2315 unsigned long pages_to_free = 0;
2316 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2317 shrinker);
2318
2319 for (i = zs_size_classes - 1; i >= 0; i--) {
2320 class = pool->size_class[i];
2321 if (!class)
2322 continue;
2323 if (class->index != i)
2324 continue;
2325
2326 pages_to_free += zs_can_compact(class);
2327 }
2328
2329 return pages_to_free;
2330}
2331
2332static void zs_unregister_shrinker(struct zs_pool *pool)
2333{
2334 if (pool->shrinker_enabled) {
2335 unregister_shrinker(&pool->shrinker);
2336 pool->shrinker_enabled = false;
2337 }
2338}
2339
2340static int zs_register_shrinker(struct zs_pool *pool)
2341{
2342 pool->shrinker.scan_objects = zs_shrinker_scan;
2343 pool->shrinker.count_objects = zs_shrinker_count;
2344 pool->shrinker.batch = 0;
2345 pool->shrinker.seeks = DEFAULT_SEEKS;
2346
2347 return register_shrinker(&pool->shrinker);
2348}
2349
2350/**
2351 * zs_create_pool - Creates an allocation pool to work from.
2352 * @name: pool name to be created
2353 *
2354 * This function must be called before anything when using
2355 * the zsmalloc allocator.
2356 *
2357 * On success, a pointer to the newly created pool is returned,
2358 * otherwise NULL.
2359 */
2360struct zs_pool *zs_create_pool(const char *name)
2361{
2362 int i;
2363 struct zs_pool *pool;
2364 struct size_class *prev_class = NULL;
2365
2366 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2367 if (!pool)
2368 return NULL;
2369
2370 init_deferred_free(pool);
2371 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
2372 GFP_KERNEL);
2373 if (!pool->size_class) {
2374 kfree(pool);
2375 return NULL;
2376 }
2377
2378 pool->name = kstrdup(name, GFP_KERNEL);
2379 if (!pool->name)
2380 goto err;
2381
2382 if (create_cache(pool))
2383 goto err;
2384
2385 /*
2386 * Iterate reversly, because, size of size_class that we want to use
2387 * for merging should be larger or equal to current size.
2388 */
2389 for (i = zs_size_classes - 1; i >= 0; i--) {
2390 int size;
2391 int pages_per_zspage;
2392 int objs_per_zspage;
2393 struct size_class *class;
2394 int fullness = 0;
2395
2396 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2397 if (size > ZS_MAX_ALLOC_SIZE)
2398 size = ZS_MAX_ALLOC_SIZE;
2399 pages_per_zspage = get_pages_per_zspage(size);
2400 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2401
2402 /*
2403 * size_class is used for normal zsmalloc operation such
2404 * as alloc/free for that size. Although it is natural that we
2405 * have one size_class for each size, there is a chance that we
2406 * can get more memory utilization if we use one size_class for
2407 * many different sizes whose size_class have same
2408 * characteristics. So, we makes size_class point to
2409 * previous size_class if possible.
2410 */
2411 if (prev_class) {
2412 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2413 pool->size_class[i] = prev_class;
2414 continue;
2415 }
2416 }
2417
2418 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2419 if (!class)
2420 goto err;
2421
2422 class->size = size;
2423 class->index = i;
2424 class->pages_per_zspage = pages_per_zspage;
2425 class->objs_per_zspage = objs_per_zspage;
2426 spin_lock_init(&class->lock);
2427 pool->size_class[i] = class;
2428 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2429 fullness++)
2430 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2431
2432 prev_class = class;
2433 }
2434
2435 /* debug only, don't abort if it fails */
2436 zs_pool_stat_create(pool, name);
2437
2438 if (zs_register_migration(pool))
2439 goto err;
2440
2441 /*
2442 * Not critical, we still can use the pool
2443 * and user can trigger compaction manually.
2444 */
2445 if (zs_register_shrinker(pool) == 0)
2446 pool->shrinker_enabled = true;
2447 return pool;
2448
2449err:
2450 zs_destroy_pool(pool);
2451 return NULL;
2452}
2453EXPORT_SYMBOL_GPL(zs_create_pool);
2454
2455void zs_destroy_pool(struct zs_pool *pool)
2456{
2457 int i;
2458
2459 zs_unregister_shrinker(pool);
2460 zs_unregister_migration(pool);
2461 zs_pool_stat_destroy(pool);
2462
2463 for (i = 0; i < zs_size_classes; i++) {
2464 int fg;
2465 struct size_class *class = pool->size_class[i];
2466
2467 if (!class)
2468 continue;
2469
2470 if (class->index != i)
2471 continue;
2472
2473 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2474 if (!list_empty(&class->fullness_list[fg])) {
2475 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2476 class->size, fg);
2477 }
2478 }
2479 kfree(class);
2480 }
2481
2482 destroy_cache(pool);
2483 kfree(pool->size_class);
2484 kfree(pool->name);
2485 kfree(pool);
2486}
2487EXPORT_SYMBOL_GPL(zs_destroy_pool);
2488
2489static int __init zs_init(void)
2490{
2491 int ret;
2492
2493 ret = zsmalloc_mount();
2494 if (ret)
2495 goto out;
2496
2497 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2498 zs_cpu_prepare, zs_cpu_dead);
2499 if (ret)
2500 goto hp_setup_fail;
2501
2502 init_zs_size_classes();
2503
2504#ifdef CONFIG_ZPOOL
2505 zpool_register_driver(&zs_zpool_driver);
2506#endif
2507
2508 zs_stat_init();
2509
2510 return 0;
2511
2512hp_setup_fail:
2513 zsmalloc_unmount();
2514out:
2515 return ret;
2516}
2517
2518static void __exit zs_exit(void)
2519{
2520#ifdef CONFIG_ZPOOL
2521 zpool_unregister_driver(&zs_zpool_driver);
2522#endif
2523 zsmalloc_unmount();
2524 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2525
2526 zs_stat_exit();
2527}
2528
2529module_init(zs_init);
2530module_exit(zs_exit);
2531
2532MODULE_LICENSE("Dual BSD/GPL");
2533MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");