blob: c03ca5e9fe15c8b9725db0ac41b6ba78f64b0e5d [file] [log] [blame]
Nitin Gupta61989a82012-01-09 16:51:56 -06001/*
2 * zsmalloc memory allocator
3 *
4 * Copyright (C) 2011 Nitin Gupta
Minchan Kim31fc00b2014-01-30 15:45:55 -08005 * Copyright (C) 2012, 2013 Minchan Kim
Nitin Gupta61989a82012-01-09 16:51:56 -06006 *
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
Nitin Gupta2db51da2012-06-09 17:41:14 -070014/*
Nitin Cuptac3e3e882013-12-11 11:04:37 +090015 * 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).
Nitin Gupta2db51da2012-06-09 17:41:14 -070022 *
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 *
Nitin Cuptac3e3e882013-12-11 11:04:37 +090029 * 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 *
Nitin Gupta2db51da2012-06-09 17:41:14 -070044 * 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
Nitin Gupta61989a82012-01-09 16:51:56 -060075#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>
Nitin Gupta61989a82012-01-09 16:51:56 -060084#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>
Seth Jennings0cbb6132012-02-13 08:47:49 -060090#include <linux/vmalloc.h>
Seth Jenningsc60369f2012-07-18 11:55:55 -050091#include <linux/hardirq.h>
Seth Jennings0959c632012-08-08 15:12:17 +090092#include <linux/spinlock.h>
93#include <linux/types.h>
Minchan Kimbcf16472014-01-30 15:45:50 -080094#include <linux/zsmalloc.h>
Seth Jennings0959c632012-08-08 15:12:17 +090095
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
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900113 * as single (unsigned long) handle value.
Seth Jennings0959c632012-08-08 15:12:17 +0900114 *
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 */
Seth Jenningsd662b8e2013-01-25 11:46:18 -0600156#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
Seth Jennings0959c632012-08-08 15:12:17 +0900157#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 */
Seth Jennings0959c632012-08-08 15:12:17 +0900222};
Nitin Gupta61989a82012-01-09 16:51:56 -0600223
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
Seth Jenningsf5536462012-07-18 11:55:56 -0500233struct mapping_area {
Minchan Kim1b945ae2013-12-11 11:04:36 +0900234#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -0500235 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
Nitin Gupta61989a82012-01-09 16:51:56 -0600244/* 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{
Minchan Kima27545bf2012-04-25 15:23:09 +0900249 return PagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600250}
251
252static int is_last_page(struct page *page)
253{
Minchan Kima27545bf2012-04-25 15:23:09 +0900254 return PagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600255}
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
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900279/*
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 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600286static 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
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900297/*
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 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600304static 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
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900325/*
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 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600331static 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
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900348/*
349 * This function removes the given zspage from the freelist identified
350 * by <class, fullness_group>.
351 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600352static 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
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900373/*
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 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600382static 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 */
Minchan Kim2e3b61542012-05-03 15:40:39 +0900417static int get_pages_per_zspage(int class_size)
Nitin Gupta61989a82012-01-09 16:51:56 -0600418{
419 int i, max_usedpc = 0;
420 /* zspage order which gives maximum used size per KB */
421 int max_usedpc_order = 1;
422
Seth Jennings84d4faa2012-03-05 11:33:21 -0600423 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600424 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))
Sunghan Suhe842b972013-07-12 16:08:13 +0900460 next = (struct page *)page_private(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600461 else
462 next = list_entry(page->lru.next, struct page, lru);
463
464 return next;
465}
466
Olav Haugan67296872013-11-22 09:30:41 -0800467/*
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 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600473static 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;
Olav Haugan67296872013-11-22 09:30:41 -0800483 handle |= ((obj_idx + 1) & OBJ_INDEX_MASK);
Nitin Gupta61989a82012-01-09 16:51:56 -0600484
485 return (void *)handle;
486}
487
Olav Haugan67296872013-11-22 09:30:41 -0800488/*
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 */
Minchan Kimc2344342012-06-08 15:39:25 +0900493static void obj_handle_to_location(unsigned long handle, struct page **page,
Nitin Gupta61989a82012-01-09 16:51:56 -0600494 unsigned long *obj_idx)
495{
Minchan Kimc2344342012-06-08 15:39:25 +0900496 *page = pfn_to_page(handle >> OBJ_INDEX_BITS);
Olav Haugan67296872013-11-22 09:30:41 -0800497 *obj_idx = (handle & OBJ_INDEX_MASK) - 1;
Nitin Gupta61989a82012-01-09 16:51:56 -0600498}
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
Nitin Guptaf4477e92012-04-02 09:13:56 -0500511static 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;
Mel Gorman22b751c2013-02-22 16:34:59 -0800518 page_mapcount_reset(page);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500519}
520
Nitin Gupta61989a82012-01-09 16:51:56 -0600521static void free_zspage(struct page *first_page)
522{
Nitin Guptaf4477e92012-04-02 09:13:56 -0500523 struct page *nextp, *tmp, *head_extra;
Nitin Gupta61989a82012-01-09 16:51:56 -0600524
525 BUG_ON(!is_first_page(first_page));
526 BUG_ON(first_page->inuse);
527
Nitin Guptaf4477e92012-04-02 09:13:56 -0500528 head_extra = (struct page *)page_private(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600529
Nitin Guptaf4477e92012-04-02 09:13:56 -0500530 reset_page(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600531 __free_page(first_page);
532
533 /* zspage with only 1 system page */
Nitin Guptaf4477e92012-04-02 09:13:56 -0500534 if (!head_extra)
Nitin Gupta61989a82012-01-09 16:51:56 -0600535 return;
536
Nitin Guptaf4477e92012-04-02 09:13:56 -0500537 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600538 list_del(&nextp->lru);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500539 reset_page(nextp);
Nitin Gupta61989a82012-01-09 16:51:56 -0600540 __free_page(nextp);
541 }
Nitin Guptaf4477e92012-04-02 09:13:56 -0500542 reset_page(head_extra);
543 __free_page(head_extra);
Nitin Gupta61989a82012-01-09 16:51:56 -0600544}
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;
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500598 struct page *first_page = NULL, *uninitialized_var(prev_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600599
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;
Minchan Kim2e3b61542012-05-03 15:40:39 +0900612 for (i = 0; i < class->pages_per_zspage; i++) {
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500613 struct page *page;
Nitin Gupta61989a82012-01-09 16:51:56 -0600614
615 page = alloc_page(flags);
616 if (!page)
617 goto cleanup;
618
619 INIT_LIST_HEAD(&page->lru);
620 if (i == 0) { /* first page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900621 SetPagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600622 set_page_private(page, 0);
623 first_page = page;
624 first_page->inuse = 0;
625 }
626 if (i == 1)
Sunghan Suhe842b972013-07-12 16:08:13 +0900627 set_page_private(first_page, (unsigned long)page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600628 if (i >= 1)
629 page->first_page = first_page;
630 if (i >= 2)
631 list_add(&page->lru, &prev_page->lru);
Minchan Kim2e3b61542012-05-03 15:40:39 +0900632 if (i == class->pages_per_zspage - 1) /* last page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900633 SetPagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600634 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 */
Minchan Kim2e3b61542012-05-03 15:40:39 +0900641 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
Nitin Gupta61989a82012-01-09 16:51:56 -0600642
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
Minchan Kim1b945ae2013-12-11 11:04:36 +0900668#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -0500669static inline int __zs_cpu_up(struct mapping_area *area)
Seth Jennings5f601902012-07-02 16:15:49 -0500670{
Seth Jenningsf5536462012-07-18 11:55:56 -0500671 /*
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;
Seth Jenningsf5536462012-07-18 11:55:56 -0500702
Joerg Roedeld95abbb2013-03-27 01:43:14 +0100703 unmap_kernel_range(addr, PAGE_SIZE * 2);
Seth Jenningsf5536462012-07-18 11:55:56 -0500704}
705
Minchan Kim1b945ae2013-12-11 11:04:36 +0900706#else /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -0500707
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{
Seth Jennings5f601902012-07-02 16:15:49 -0500732 int sizes[2];
733 void *addr;
Seth Jenningsf5536462012-07-18 11:55:56 -0500734 char *buf = area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -0500735
Seth Jenningsf5536462012-07-18 11:55:56 -0500736 /* 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;
Seth Jennings5f601902012-07-02 16:15:49 -0500742
743 sizes[0] = PAGE_SIZE - off;
744 sizes[1] = size - sizes[0];
745
Seth Jennings5f601902012-07-02 16:15:49 -0500746 /* 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);
Seth Jenningsf5536462012-07-18 11:55:56 -0500753out:
754 return area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -0500755}
756
Seth Jenningsf5536462012-07-18 11:55:56 -0500757static void __zs_unmap_object(struct mapping_area *area,
758 struct page *pages[2], int off, int size)
Seth Jennings5f601902012-07-02 16:15:49 -0500759{
Seth Jennings5f601902012-07-02 16:15:49 -0500760 int sizes[2];
761 void *addr;
Seth Jenningsf5536462012-07-18 11:55:56 -0500762 char *buf = area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -0500763
Seth Jenningsf5536462012-07-18 11:55:56 -0500764 /* no write fastpath */
765 if (area->vm_mm == ZS_MM_RO)
766 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -0500767
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);
Seth Jenningsf5536462012-07-18 11:55:56 -0500778
779out:
780 /* enable page faults to match kunmap_atomic() return conditions */
781 pagefault_enable();
Seth Jennings5f601902012-07-02 16:15:49 -0500782}
Nitin Gupta61989a82012-01-09 16:51:56 -0600783
Minchan Kim1b945ae2013-12-11 11:04:36 +0900784#endif /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -0500785
Nitin Gupta61989a82012-01-09 16:51:56 -0600786static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
787 void *pcpu)
788{
Seth Jenningsf5536462012-07-18 11:55:56 -0500789 int ret, cpu = (long)pcpu;
Nitin Gupta61989a82012-01-09 16:51:56 -0600790 struct mapping_area *area;
791
792 switch (action) {
793 case CPU_UP_PREPARE:
794 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -0500795 ret = __zs_cpu_up(area);
796 if (ret)
797 return notifier_from_errno(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -0600798 break;
799 case CPU_DEAD:
800 case CPU_UP_CANCELED:
801 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -0500802 __zs_cpu_down(area);
Nitin Gupta61989a82012-01-09 16:51:56 -0600803 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 for_each_online_cpu(cpu)
818 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
819 unregister_cpu_notifier(&zs_cpu_nb);
820}
821
822static int zs_init(void)
823{
824 int cpu, ret;
825
826 register_cpu_notifier(&zs_cpu_nb);
827 for_each_online_cpu(cpu) {
828 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
829 if (notifier_to_errno(ret))
830 goto fail;
831 }
832 return 0;
833fail:
834 zs_exit();
835 return notifier_to_errno(ret);
836}
837
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -0800838/**
839 * zs_create_pool - Creates an allocation pool to work from.
Seth Jennings0d145a52013-01-30 09:36:52 -0600840 * @flags: allocation flags used to allocate pool metadata
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -0800841 *
842 * This function must be called before anything when using
843 * the zsmalloc allocator.
844 *
845 * On success, a pointer to the newly created pool is returned,
846 * otherwise NULL.
847 */
Seth Jennings0d145a52013-01-30 09:36:52 -0600848struct zs_pool *zs_create_pool(gfp_t flags)
Nitin Gupta61989a82012-01-09 16:51:56 -0600849{
Ben Hutchings069f1012012-06-20 02:31:11 +0100850 int i, ovhd_size;
Nitin Gupta61989a82012-01-09 16:51:56 -0600851 struct zs_pool *pool;
852
Nitin Gupta61989a82012-01-09 16:51:56 -0600853 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
854 pool = kzalloc(ovhd_size, GFP_KERNEL);
855 if (!pool)
856 return NULL;
857
858 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
859 int size;
860 struct size_class *class;
861
862 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
863 if (size > ZS_MAX_ALLOC_SIZE)
864 size = ZS_MAX_ALLOC_SIZE;
865
866 class = &pool->size_class[i];
867 class->size = size;
868 class->index = i;
869 spin_lock_init(&class->lock);
Minchan Kim2e3b61542012-05-03 15:40:39 +0900870 class->pages_per_zspage = get_pages_per_zspage(size);
Nitin Gupta61989a82012-01-09 16:51:56 -0600871
872 }
873
Nitin Gupta61989a82012-01-09 16:51:56 -0600874 pool->flags = flags;
Nitin Gupta61989a82012-01-09 16:51:56 -0600875
Nitin Gupta61989a82012-01-09 16:51:56 -0600876 return pool;
877}
878EXPORT_SYMBOL_GPL(zs_create_pool);
879
880void zs_destroy_pool(struct zs_pool *pool)
881{
882 int i;
883
884 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
885 int fg;
886 struct size_class *class = &pool->size_class[i];
887
888 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
889 if (class->fullness_list[fg]) {
Marlies Ruck93ad5ab2013-05-15 16:56:49 -0400890 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
Nitin Gupta61989a82012-01-09 16:51:56 -0600891 class->size, fg);
892 }
893 }
894 }
895 kfree(pool);
896}
897EXPORT_SYMBOL_GPL(zs_destroy_pool);
898
899/**
900 * zs_malloc - Allocate block of given size from pool.
901 * @pool: pool to allocate from
902 * @size: size of block to allocate
Nitin Gupta61989a82012-01-09 16:51:56 -0600903 *
Minchan Kim00a61d82012-05-03 15:40:40 +0900904 * On success, handle to the allocated object is returned,
Minchan Kimc2344342012-06-08 15:39:25 +0900905 * otherwise 0.
Nitin Gupta61989a82012-01-09 16:51:56 -0600906 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
907 */
Minchan Kimc2344342012-06-08 15:39:25 +0900908unsigned long zs_malloc(struct zs_pool *pool, size_t size)
Nitin Gupta61989a82012-01-09 16:51:56 -0600909{
Minchan Kimc2344342012-06-08 15:39:25 +0900910 unsigned long obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600911 struct link_free *link;
912 int class_idx;
913 struct size_class *class;
914
915 struct page *first_page, *m_page;
916 unsigned long m_objidx, m_offset;
917
918 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
Minchan Kimc2344342012-06-08 15:39:25 +0900919 return 0;
Nitin Gupta61989a82012-01-09 16:51:56 -0600920
921 class_idx = get_size_class_index(size);
922 class = &pool->size_class[class_idx];
923 BUG_ON(class_idx != class->index);
924
925 spin_lock(&class->lock);
926 first_page = find_get_zspage(class);
927
928 if (!first_page) {
929 spin_unlock(&class->lock);
930 first_page = alloc_zspage(class, pool->flags);
931 if (unlikely(!first_page))
Minchan Kimc2344342012-06-08 15:39:25 +0900932 return 0;
Nitin Gupta61989a82012-01-09 16:51:56 -0600933
934 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
935 spin_lock(&class->lock);
Minchan Kim2e3b61542012-05-03 15:40:39 +0900936 class->pages_allocated += class->pages_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -0600937 }
938
Minchan Kimc2344342012-06-08 15:39:25 +0900939 obj = (unsigned long)first_page->freelist;
Nitin Gupta61989a82012-01-09 16:51:56 -0600940 obj_handle_to_location(obj, &m_page, &m_objidx);
941 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
942
943 link = (struct link_free *)kmap_atomic(m_page) +
944 m_offset / sizeof(*link);
945 first_page->freelist = link->next;
946 memset(link, POISON_INUSE, sizeof(*link));
947 kunmap_atomic(link);
948
949 first_page->inuse++;
950 /* Now move the zspage to another fullness group, if required */
951 fix_fullness_group(pool, first_page);
952 spin_unlock(&class->lock);
953
954 return obj;
955}
956EXPORT_SYMBOL_GPL(zs_malloc);
957
Minchan Kimc2344342012-06-08 15:39:25 +0900958void zs_free(struct zs_pool *pool, unsigned long obj)
Nitin Gupta61989a82012-01-09 16:51:56 -0600959{
960 struct link_free *link;
961 struct page *first_page, *f_page;
962 unsigned long f_objidx, f_offset;
963
964 int class_idx;
965 struct size_class *class;
966 enum fullness_group fullness;
967
968 if (unlikely(!obj))
969 return;
970
971 obj_handle_to_location(obj, &f_page, &f_objidx);
972 first_page = get_first_page(f_page);
973
974 get_zspage_mapping(first_page, &class_idx, &fullness);
975 class = &pool->size_class[class_idx];
976 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
977
978 spin_lock(&class->lock);
979
980 /* Insert this object in containing zspage's freelist */
981 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
982 + f_offset);
983 link->next = first_page->freelist;
984 kunmap_atomic(link);
Minchan Kimc2344342012-06-08 15:39:25 +0900985 first_page->freelist = (void *)obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600986
987 first_page->inuse--;
988 fullness = fix_fullness_group(pool, first_page);
989
990 if (fullness == ZS_EMPTY)
Minchan Kim2e3b61542012-05-03 15:40:39 +0900991 class->pages_allocated -= class->pages_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -0600992
993 spin_unlock(&class->lock);
994
995 if (fullness == ZS_EMPTY)
996 free_zspage(first_page);
997}
998EXPORT_SYMBOL_GPL(zs_free);
999
Minchan Kim00a61d82012-05-03 15:40:40 +09001000/**
1001 * zs_map_object - get address of allocated object from handle.
1002 * @pool: pool from which the object was allocated
1003 * @handle: handle returned from zs_malloc
1004 *
1005 * Before using an object allocated from zs_malloc, it must be mapped using
1006 * this function. When done with the object, it must be unmapped using
Seth Jennings166cfda2012-07-02 16:15:51 -05001007 * zs_unmap_object.
1008 *
1009 * Only one object can be mapped per cpu at a time. There is no protection
1010 * against nested mappings.
1011 *
1012 * This function returns with preemption and page faults disabled.
Sara Bird396b7fd2013-05-20 15:18:14 -04001013 */
Seth Jenningsb7418512012-07-02 16:15:52 -05001014void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1015 enum zs_mapmode mm)
Nitin Gupta61989a82012-01-09 16:51:56 -06001016{
1017 struct page *page;
1018 unsigned long obj_idx, off;
1019
1020 unsigned int class_idx;
1021 enum fullness_group fg;
1022 struct size_class *class;
1023 struct mapping_area *area;
Seth Jenningsf5536462012-07-18 11:55:56 -05001024 struct page *pages[2];
Nitin Gupta61989a82012-01-09 16:51:56 -06001025
1026 BUG_ON(!handle);
1027
Seth Jenningsc60369f2012-07-18 11:55:55 -05001028 /*
1029 * Because we use per-cpu mapping areas shared among the
1030 * pools/users, we can't allow mapping in interrupt context
1031 * because it can corrupt another users mappings.
1032 */
1033 BUG_ON(in_interrupt());
1034
Nitin Gupta61989a82012-01-09 16:51:56 -06001035 obj_handle_to_location(handle, &page, &obj_idx);
1036 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1037 class = &pool->size_class[class_idx];
1038 off = obj_idx_to_offset(page, obj_idx, class->size);
1039
1040 area = &get_cpu_var(zs_map_area);
Seth Jenningsf5536462012-07-18 11:55:56 -05001041 area->vm_mm = mm;
Nitin Gupta61989a82012-01-09 16:51:56 -06001042 if (off + class->size <= PAGE_SIZE) {
1043 /* this object is contained entirely within a page */
1044 area->vm_addr = kmap_atomic(page);
Seth Jennings5f601902012-07-02 16:15:49 -05001045 return area->vm_addr + off;
Nitin Gupta61989a82012-01-09 16:51:56 -06001046 }
1047
Seth Jenningsf5536462012-07-18 11:55:56 -05001048 /* this object spans two pages */
1049 pages[0] = page;
1050 pages[1] = get_next_page(page);
1051 BUG_ON(!pages[1]);
Seth Jenningsb7418512012-07-02 16:15:52 -05001052
Seth Jenningsf5536462012-07-18 11:55:56 -05001053 return __zs_map_object(area, pages, off, class->size);
Nitin Gupta61989a82012-01-09 16:51:56 -06001054}
1055EXPORT_SYMBOL_GPL(zs_map_object);
1056
Minchan Kimc2344342012-06-08 15:39:25 +09001057void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
Nitin Gupta61989a82012-01-09 16:51:56 -06001058{
1059 struct page *page;
1060 unsigned long obj_idx, off;
1061
1062 unsigned int class_idx;
1063 enum fullness_group fg;
1064 struct size_class *class;
1065 struct mapping_area *area;
1066
1067 BUG_ON(!handle);
1068
1069 obj_handle_to_location(handle, &page, &obj_idx);
1070 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1071 class = &pool->size_class[class_idx];
1072 off = obj_idx_to_offset(page, obj_idx, class->size);
1073
Seth Jenningsf5536462012-07-18 11:55:56 -05001074 area = &__get_cpu_var(zs_map_area);
1075 if (off + class->size <= PAGE_SIZE)
1076 kunmap_atomic(area->vm_addr);
1077 else {
1078 struct page *pages[2];
Seth Jenningsb7418512012-07-02 16:15:52 -05001079
Seth Jenningsf5536462012-07-18 11:55:56 -05001080 pages[0] = page;
1081 pages[1] = get_next_page(page);
1082 BUG_ON(!pages[1]);
1083
1084 __zs_unmap_object(area, pages, off, class->size);
1085 }
Nitin Gupta61989a82012-01-09 16:51:56 -06001086 put_cpu_var(zs_map_area);
1087}
1088EXPORT_SYMBOL_GPL(zs_unmap_object);
1089
1090u64 zs_get_total_size_bytes(struct zs_pool *pool)
1091{
1092 int i;
1093 u64 npages = 0;
1094
1095 for (i = 0; i < ZS_SIZE_CLASSES; i++)
1096 npages += pool->size_class[i].pages_allocated;
1097
1098 return npages << PAGE_SHIFT;
1099}
1100EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);
Ben Hutchings069f1012012-06-20 02:31:11 +01001101
1102module_init(zs_init);
1103module_exit(zs_exit);
1104
1105MODULE_LICENSE("Dual BSD/GPL");
1106MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");