Merge tag 'sound-3.4' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound
[firefly-linux-kernel-4.4.55.git] / drivers / staging / zsmalloc / zsmalloc-main.c
1 /*
2  * zsmalloc memory allocator
3  *
4  * Copyright (C) 2011  Nitin Gupta
5  *
6  * This code is released using a dual license strategy: BSD/GPL
7  * You can choose the license that better fits your requirements.
8  *
9  * Released under the terms of 3-clause BSD License
10  * Released under the terms of GNU General Public License Version 2.0
11  */
12
13 #ifdef CONFIG_ZSMALLOC_DEBUG
14 #define DEBUG
15 #endif
16
17 #include <linux/module.h>
18 #include <linux/kernel.h>
19 #include <linux/bitops.h>
20 #include <linux/errno.h>
21 #include <linux/highmem.h>
22 #include <linux/init.h>
23 #include <linux/string.h>
24 #include <linux/slab.h>
25 #include <asm/tlbflush.h>
26 #include <asm/pgtable.h>
27 #include <linux/cpumask.h>
28 #include <linux/cpu.h>
29 #include <linux/vmalloc.h>
30
31 #include "zsmalloc.h"
32 #include "zsmalloc_int.h"
33
34 /*
35  * A zspage's class index and fullness group
36  * are encoded in its (first)page->mapping
37  */
38 #define CLASS_IDX_BITS  28
39 #define FULLNESS_BITS   4
40 #define CLASS_IDX_MASK  ((1 << CLASS_IDX_BITS) - 1)
41 #define FULLNESS_MASK   ((1 << FULLNESS_BITS) - 1)
42
43 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
44 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
45
46 static int is_first_page(struct page *page)
47 {
48         return test_bit(PG_private, &page->flags);
49 }
50
51 static int is_last_page(struct page *page)
52 {
53         return test_bit(PG_private_2, &page->flags);
54 }
55
56 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
57                                 enum fullness_group *fullness)
58 {
59         unsigned long m;
60         BUG_ON(!is_first_page(page));
61
62         m = (unsigned long)page->mapping;
63         *fullness = m & FULLNESS_MASK;
64         *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
65 }
66
67 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
68                                 enum fullness_group fullness)
69 {
70         unsigned long m;
71         BUG_ON(!is_first_page(page));
72
73         m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
74                         (fullness & FULLNESS_MASK);
75         page->mapping = (struct address_space *)m;
76 }
77
78 static int get_size_class_index(int size)
79 {
80         int idx = 0;
81
82         if (likely(size > ZS_MIN_ALLOC_SIZE))
83                 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
84                                 ZS_SIZE_CLASS_DELTA);
85
86         return idx;
87 }
88
89 static enum fullness_group get_fullness_group(struct page *page)
90 {
91         int inuse, max_objects;
92         enum fullness_group fg;
93         BUG_ON(!is_first_page(page));
94
95         inuse = page->inuse;
96         max_objects = page->objects;
97
98         if (inuse == 0)
99                 fg = ZS_EMPTY;
100         else if (inuse == max_objects)
101                 fg = ZS_FULL;
102         else if (inuse <= max_objects / fullness_threshold_frac)
103                 fg = ZS_ALMOST_EMPTY;
104         else
105                 fg = ZS_ALMOST_FULL;
106
107         return fg;
108 }
109
110 static void insert_zspage(struct page *page, struct size_class *class,
111                                 enum fullness_group fullness)
112 {
113         struct page **head;
114
115         BUG_ON(!is_first_page(page));
116
117         if (fullness >= _ZS_NR_FULLNESS_GROUPS)
118                 return;
119
120         head = &class->fullness_list[fullness];
121         if (*head)
122                 list_add_tail(&page->lru, &(*head)->lru);
123
124         *head = page;
125 }
126
127 static void remove_zspage(struct page *page, struct size_class *class,
128                                 enum fullness_group fullness)
129 {
130         struct page **head;
131
132         BUG_ON(!is_first_page(page));
133
134         if (fullness >= _ZS_NR_FULLNESS_GROUPS)
135                 return;
136
137         head = &class->fullness_list[fullness];
138         BUG_ON(!*head);
139         if (list_empty(&(*head)->lru))
140                 *head = NULL;
141         else if (*head == page)
142                 *head = (struct page *)list_entry((*head)->lru.next,
143                                         struct page, lru);
144
145         list_del_init(&page->lru);
146 }
147
148 static enum fullness_group fix_fullness_group(struct zs_pool *pool,
149                                                 struct page *page)
150 {
151         int class_idx;
152         struct size_class *class;
153         enum fullness_group currfg, newfg;
154
155         BUG_ON(!is_first_page(page));
156
157         get_zspage_mapping(page, &class_idx, &currfg);
158         newfg = get_fullness_group(page);
159         if (newfg == currfg)
160                 goto out;
161
162         class = &pool->size_class[class_idx];
163         remove_zspage(page, class, currfg);
164         insert_zspage(page, class, newfg);
165         set_zspage_mapping(page, class_idx, newfg);
166
167 out:
168         return newfg;
169 }
170
171 /*
172  * We have to decide on how many pages to link together
173  * to form a zspage for each size class. This is important
174  * to reduce wastage due to unusable space left at end of
175  * each zspage which is given as:
176  *      wastage = Zp - Zp % size_class
177  * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
178  *
179  * For example, for size class of 3/8 * PAGE_SIZE, we should
180  * link together 3 PAGE_SIZE sized pages to form a zspage
181  * since then we can perfectly fit in 8 such objects.
182  */
183 static int get_zspage_order(int class_size)
184 {
185         int i, max_usedpc = 0;
186         /* zspage order which gives maximum used size per KB */
187         int max_usedpc_order = 1;
188
189         for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
190                 int zspage_size;
191                 int waste, usedpc;
192
193                 zspage_size = i * PAGE_SIZE;
194                 waste = zspage_size % class_size;
195                 usedpc = (zspage_size - waste) * 100 / zspage_size;
196
197                 if (usedpc > max_usedpc) {
198                         max_usedpc = usedpc;
199                         max_usedpc_order = i;
200                 }
201         }
202
203         return max_usedpc_order;
204 }
205
206 /*
207  * A single 'zspage' is composed of many system pages which are
208  * linked together using fields in struct page. This function finds
209  * the first/head page, given any component page of a zspage.
210  */
211 static struct page *get_first_page(struct page *page)
212 {
213         if (is_first_page(page))
214                 return page;
215         else
216                 return page->first_page;
217 }
218
219 static struct page *get_next_page(struct page *page)
220 {
221         struct page *next;
222
223         if (is_last_page(page))
224                 next = NULL;
225         else if (is_first_page(page))
226                 next = (struct page *)page->private;
227         else
228                 next = list_entry(page->lru.next, struct page, lru);
229
230         return next;
231 }
232
233 /* Encode <page, obj_idx> as a single handle value */
234 static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
235 {
236         unsigned long handle;
237
238         if (!page) {
239                 BUG_ON(obj_idx);
240                 return NULL;
241         }
242
243         handle = page_to_pfn(page) << OBJ_INDEX_BITS;
244         handle |= (obj_idx & OBJ_INDEX_MASK);
245
246         return (void *)handle;
247 }
248
249 /* Decode <page, obj_idx> pair from the given object handle */
250 static void obj_handle_to_location(void *handle, struct page **page,
251                                 unsigned long *obj_idx)
252 {
253         unsigned long hval = (unsigned long)handle;
254
255         *page = pfn_to_page(hval >> OBJ_INDEX_BITS);
256         *obj_idx = hval & OBJ_INDEX_MASK;
257 }
258
259 static unsigned long obj_idx_to_offset(struct page *page,
260                                 unsigned long obj_idx, int class_size)
261 {
262         unsigned long off = 0;
263
264         if (!is_first_page(page))
265                 off = page->index;
266
267         return off + obj_idx * class_size;
268 }
269
270 static void free_zspage(struct page *first_page)
271 {
272         struct page *nextp, *tmp;
273
274         BUG_ON(!is_first_page(first_page));
275         BUG_ON(first_page->inuse);
276
277         nextp = (struct page *)page_private(first_page);
278
279         clear_bit(PG_private, &first_page->flags);
280         clear_bit(PG_private_2, &first_page->flags);
281         set_page_private(first_page, 0);
282         first_page->mapping = NULL;
283         first_page->freelist = NULL;
284         reset_page_mapcount(first_page);
285         __free_page(first_page);
286
287         /* zspage with only 1 system page */
288         if (!nextp)
289                 return;
290
291         list_for_each_entry_safe(nextp, tmp, &nextp->lru, lru) {
292                 list_del(&nextp->lru);
293                 clear_bit(PG_private_2, &nextp->flags);
294                 nextp->index = 0;
295                 __free_page(nextp);
296         }
297 }
298
299 /* Initialize a newly allocated zspage */
300 static void init_zspage(struct page *first_page, struct size_class *class)
301 {
302         unsigned long off = 0;
303         struct page *page = first_page;
304
305         BUG_ON(!is_first_page(first_page));
306         while (page) {
307                 struct page *next_page;
308                 struct link_free *link;
309                 unsigned int i, objs_on_page;
310
311                 /*
312                  * page->index stores offset of first object starting
313                  * in the page. For the first page, this is always 0,
314                  * so we use first_page->index (aka ->freelist) to store
315                  * head of corresponding zspage's freelist.
316                  */
317                 if (page != first_page)
318                         page->index = off;
319
320                 link = (struct link_free *)kmap_atomic(page) +
321                                                 off / sizeof(*link);
322                 objs_on_page = (PAGE_SIZE - off) / class->size;
323
324                 for (i = 1; i <= objs_on_page; i++) {
325                         off += class->size;
326                         if (off < PAGE_SIZE) {
327                                 link->next = obj_location_to_handle(page, i);
328                                 link += class->size / sizeof(*link);
329                         }
330                 }
331
332                 /*
333                  * We now come to the last (full or partial) object on this
334                  * page, which must point to the first object on the next
335                  * page (if present)
336                  */
337                 next_page = get_next_page(page);
338                 link->next = obj_location_to_handle(next_page, 0);
339                 kunmap_atomic(link);
340                 page = next_page;
341                 off = (off + class->size) % PAGE_SIZE;
342         }
343 }
344
345 /*
346  * Allocate a zspage for the given size class
347  */
348 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
349 {
350         int i, error;
351         struct page *first_page = NULL;
352
353         /*
354          * Allocate individual pages and link them together as:
355          * 1. first page->private = first sub-page
356          * 2. all sub-pages are linked together using page->lru
357          * 3. each sub-page is linked to the first page using page->first_page
358          *
359          * For each size class, First/Head pages are linked together using
360          * page->lru. Also, we set PG_private to identify the first page
361          * (i.e. no other sub-page has this flag set) and PG_private_2 to
362          * identify the last page.
363          */
364         error = -ENOMEM;
365         for (i = 0; i < class->zspage_order; i++) {
366                 struct page *page, *prev_page;
367
368                 page = alloc_page(flags);
369                 if (!page)
370                         goto cleanup;
371
372                 INIT_LIST_HEAD(&page->lru);
373                 if (i == 0) {   /* first page */
374                         set_bit(PG_private, &page->flags);
375                         set_page_private(page, 0);
376                         first_page = page;
377                         first_page->inuse = 0;
378                 }
379                 if (i == 1)
380                         first_page->private = (unsigned long)page;
381                 if (i >= 1)
382                         page->first_page = first_page;
383                 if (i >= 2)
384                         list_add(&page->lru, &prev_page->lru);
385                 if (i == class->zspage_order - 1)       /* last page */
386                         set_bit(PG_private_2, &page->flags);
387
388                 prev_page = page;
389         }
390
391         init_zspage(first_page, class);
392
393         first_page->freelist = obj_location_to_handle(first_page, 0);
394         /* Maximum number of objects we can store in this zspage */
395         first_page->objects = class->zspage_order * PAGE_SIZE / class->size;
396
397         error = 0; /* Success */
398
399 cleanup:
400         if (unlikely(error) && first_page) {
401                 free_zspage(first_page);
402                 first_page = NULL;
403         }
404
405         return first_page;
406 }
407
408 static struct page *find_get_zspage(struct size_class *class)
409 {
410         int i;
411         struct page *page;
412
413         for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
414                 page = class->fullness_list[i];
415                 if (page)
416                         break;
417         }
418
419         return page;
420 }
421
422
423 /*
424  * If this becomes a separate module, register zs_init() with
425  * module_init(), zs_exit with module_exit(), and remove zs_initialized
426 */
427 static int zs_initialized;
428
429 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
430                                 void *pcpu)
431 {
432         int cpu = (long)pcpu;
433         struct mapping_area *area;
434
435         switch (action) {
436         case CPU_UP_PREPARE:
437                 area = &per_cpu(zs_map_area, cpu);
438                 if (area->vm)
439                         break;
440                 area->vm = alloc_vm_area(2 * PAGE_SIZE, area->vm_ptes);
441                 if (!area->vm)
442                         return notifier_from_errno(-ENOMEM);
443                 break;
444         case CPU_DEAD:
445         case CPU_UP_CANCELED:
446                 area = &per_cpu(zs_map_area, cpu);
447                 if (area->vm)
448                         free_vm_area(area->vm);
449                 area->vm = NULL;
450                 break;
451         }
452
453         return NOTIFY_OK;
454 }
455
456 static struct notifier_block zs_cpu_nb = {
457         .notifier_call = zs_cpu_notifier
458 };
459
460 static void zs_exit(void)
461 {
462         int cpu;
463
464         for_each_online_cpu(cpu)
465                 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
466         unregister_cpu_notifier(&zs_cpu_nb);
467 }
468
469 static int zs_init(void)
470 {
471         int cpu, ret;
472
473         register_cpu_notifier(&zs_cpu_nb);
474         for_each_online_cpu(cpu) {
475                 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
476                 if (notifier_to_errno(ret))
477                         goto fail;
478         }
479         return 0;
480 fail:
481         zs_exit();
482         return notifier_to_errno(ret);
483 }
484
485 struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
486 {
487         int i, error, ovhd_size;
488         struct zs_pool *pool;
489
490         if (!name)
491                 return NULL;
492
493         ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
494         pool = kzalloc(ovhd_size, GFP_KERNEL);
495         if (!pool)
496                 return NULL;
497
498         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
499                 int size;
500                 struct size_class *class;
501
502                 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
503                 if (size > ZS_MAX_ALLOC_SIZE)
504                         size = ZS_MAX_ALLOC_SIZE;
505
506                 class = &pool->size_class[i];
507                 class->size = size;
508                 class->index = i;
509                 spin_lock_init(&class->lock);
510                 class->zspage_order = get_zspage_order(size);
511
512         }
513
514         /*
515          * If this becomes a separate module, register zs_init with
516          * module_init, and remove this block
517         */
518         if (!zs_initialized) {
519                 error = zs_init();
520                 if (error)
521                         goto cleanup;
522                 zs_initialized = 1;
523         }
524
525         pool->flags = flags;
526         pool->name = name;
527
528         error = 0; /* Success */
529
530 cleanup:
531         if (error) {
532                 zs_destroy_pool(pool);
533                 pool = NULL;
534         }
535
536         return pool;
537 }
538 EXPORT_SYMBOL_GPL(zs_create_pool);
539
540 void zs_destroy_pool(struct zs_pool *pool)
541 {
542         int i;
543
544         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
545                 int fg;
546                 struct size_class *class = &pool->size_class[i];
547
548                 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
549                         if (class->fullness_list[fg]) {
550                                 pr_info("Freeing non-empty class with size "
551                                         "%db, fullness group %d\n",
552                                         class->size, fg);
553                         }
554                 }
555         }
556         kfree(pool);
557 }
558 EXPORT_SYMBOL_GPL(zs_destroy_pool);
559
560 /**
561  * zs_malloc - Allocate block of given size from pool.
562  * @pool: pool to allocate from
563  * @size: size of block to allocate
564  * @page: page no. that holds the object
565  * @offset: location of object within page
566  *
567  * On success, <page, offset> identifies block allocated
568  * and 0 is returned. On failure, <page, offset> is set to
569  * 0 and -ENOMEM is returned.
570  *
571  * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
572  */
573 void *zs_malloc(struct zs_pool *pool, size_t size)
574 {
575         void *obj;
576         struct link_free *link;
577         int class_idx;
578         struct size_class *class;
579
580         struct page *first_page, *m_page;
581         unsigned long m_objidx, m_offset;
582
583         if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
584                 return NULL;
585
586         class_idx = get_size_class_index(size);
587         class = &pool->size_class[class_idx];
588         BUG_ON(class_idx != class->index);
589
590         spin_lock(&class->lock);
591         first_page = find_get_zspage(class);
592
593         if (!first_page) {
594                 spin_unlock(&class->lock);
595                 first_page = alloc_zspage(class, pool->flags);
596                 if (unlikely(!first_page))
597                         return NULL;
598
599                 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
600                 spin_lock(&class->lock);
601                 class->pages_allocated += class->zspage_order;
602         }
603
604         obj = first_page->freelist;
605         obj_handle_to_location(obj, &m_page, &m_objidx);
606         m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
607
608         link = (struct link_free *)kmap_atomic(m_page) +
609                                         m_offset / sizeof(*link);
610         first_page->freelist = link->next;
611         memset(link, POISON_INUSE, sizeof(*link));
612         kunmap_atomic(link);
613
614         first_page->inuse++;
615         /* Now move the zspage to another fullness group, if required */
616         fix_fullness_group(pool, first_page);
617         spin_unlock(&class->lock);
618
619         return obj;
620 }
621 EXPORT_SYMBOL_GPL(zs_malloc);
622
623 void zs_free(struct zs_pool *pool, void *obj)
624 {
625         struct link_free *link;
626         struct page *first_page, *f_page;
627         unsigned long f_objidx, f_offset;
628
629         int class_idx;
630         struct size_class *class;
631         enum fullness_group fullness;
632
633         if (unlikely(!obj))
634                 return;
635
636         obj_handle_to_location(obj, &f_page, &f_objidx);
637         first_page = get_first_page(f_page);
638
639         get_zspage_mapping(first_page, &class_idx, &fullness);
640         class = &pool->size_class[class_idx];
641         f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
642
643         spin_lock(&class->lock);
644
645         /* Insert this object in containing zspage's freelist */
646         link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
647                                                         + f_offset);
648         link->next = first_page->freelist;
649         kunmap_atomic(link);
650         first_page->freelist = obj;
651
652         first_page->inuse--;
653         fullness = fix_fullness_group(pool, first_page);
654
655         if (fullness == ZS_EMPTY)
656                 class->pages_allocated -= class->zspage_order;
657
658         spin_unlock(&class->lock);
659
660         if (fullness == ZS_EMPTY)
661                 free_zspage(first_page);
662 }
663 EXPORT_SYMBOL_GPL(zs_free);
664
665 void *zs_map_object(struct zs_pool *pool, void *handle)
666 {
667         struct page *page;
668         unsigned long obj_idx, off;
669
670         unsigned int class_idx;
671         enum fullness_group fg;
672         struct size_class *class;
673         struct mapping_area *area;
674
675         BUG_ON(!handle);
676
677         obj_handle_to_location(handle, &page, &obj_idx);
678         get_zspage_mapping(get_first_page(page), &class_idx, &fg);
679         class = &pool->size_class[class_idx];
680         off = obj_idx_to_offset(page, obj_idx, class->size);
681
682         area = &get_cpu_var(zs_map_area);
683         if (off + class->size <= PAGE_SIZE) {
684                 /* this object is contained entirely within a page */
685                 area->vm_addr = kmap_atomic(page);
686         } else {
687                 /* this object spans two pages */
688                 struct page *nextp;
689
690                 nextp = get_next_page(page);
691                 BUG_ON(!nextp);
692
693
694                 set_pte(area->vm_ptes[0], mk_pte(page, PAGE_KERNEL));
695                 set_pte(area->vm_ptes[1], mk_pte(nextp, PAGE_KERNEL));
696
697                 /* We pre-allocated VM area so mapping can never fail */
698                 area->vm_addr = area->vm->addr;
699         }
700
701         return area->vm_addr + off;
702 }
703 EXPORT_SYMBOL_GPL(zs_map_object);
704
705 void zs_unmap_object(struct zs_pool *pool, void *handle)
706 {
707         struct page *page;
708         unsigned long obj_idx, off;
709
710         unsigned int class_idx;
711         enum fullness_group fg;
712         struct size_class *class;
713         struct mapping_area *area;
714
715         BUG_ON(!handle);
716
717         obj_handle_to_location(handle, &page, &obj_idx);
718         get_zspage_mapping(get_first_page(page), &class_idx, &fg);
719         class = &pool->size_class[class_idx];
720         off = obj_idx_to_offset(page, obj_idx, class->size);
721
722         area = &__get_cpu_var(zs_map_area);
723         if (off + class->size <= PAGE_SIZE) {
724                 kunmap_atomic(area->vm_addr);
725         } else {
726                 set_pte(area->vm_ptes[0], __pte(0));
727                 set_pte(area->vm_ptes[1], __pte(0));
728                 __flush_tlb_one((unsigned long)area->vm_addr);
729                 __flush_tlb_one((unsigned long)area->vm_addr + PAGE_SIZE);
730         }
731         put_cpu_var(zs_map_area);
732 }
733 EXPORT_SYMBOL_GPL(zs_unmap_object);
734
735 u64 zs_get_total_size_bytes(struct zs_pool *pool)
736 {
737         int i;
738         u64 npages = 0;
739
740         for (i = 0; i < ZS_SIZE_CLASSES; i++)
741                 npages += pool->size_class[i].pages_allocated;
742
743         return npages << PAGE_SHIFT;
744 }
745 EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);