Merge branch 'uaccess.csum' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[linux-2.6-microblaze.git] / drivers / md / dm-bufio.c
1 /*
2  * Copyright (C) 2009-2011 Red Hat, Inc.
3  *
4  * Author: Mikulas Patocka <mpatocka@redhat.com>
5  *
6  * This file is released under the GPL.
7  */
8
9 #include <linux/dm-bufio.h>
10
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/slab.h>
14 #include <linux/sched/mm.h>
15 #include <linux/jiffies.h>
16 #include <linux/vmalloc.h>
17 #include <linux/shrinker.h>
18 #include <linux/module.h>
19 #include <linux/rbtree.h>
20 #include <linux/stacktrace.h>
21
22 #define DM_MSG_PREFIX "bufio"
23
24 /*
25  * Memory management policy:
26  *      Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
27  *      or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
28  *      Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
29  *      Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
30  *      dirty buffers.
31  */
32 #define DM_BUFIO_MIN_BUFFERS            8
33
34 #define DM_BUFIO_MEMORY_PERCENT         2
35 #define DM_BUFIO_VMALLOC_PERCENT        25
36 #define DM_BUFIO_WRITEBACK_RATIO        3
37 #define DM_BUFIO_LOW_WATERMARK_RATIO    16
38
39 /*
40  * Check buffer ages in this interval (seconds)
41  */
42 #define DM_BUFIO_WORK_TIMER_SECS        30
43
44 /*
45  * Free buffers when they are older than this (seconds)
46  */
47 #define DM_BUFIO_DEFAULT_AGE_SECS       300
48
49 /*
50  * The nr of bytes of cached data to keep around.
51  */
52 #define DM_BUFIO_DEFAULT_RETAIN_BYTES   (256 * 1024)
53
54 /*
55  * Align buffer writes to this boundary.
56  * Tests show that SSDs have the highest IOPS when using 4k writes.
57  */
58 #define DM_BUFIO_WRITE_ALIGN            4096
59
60 /*
61  * dm_buffer->list_mode
62  */
63 #define LIST_CLEAN      0
64 #define LIST_DIRTY      1
65 #define LIST_SIZE       2
66
67 /*
68  * Linking of buffers:
69  *      All buffers are linked to buffer_tree with their node field.
70  *
71  *      Clean buffers that are not being written (B_WRITING not set)
72  *      are linked to lru[LIST_CLEAN] with their lru_list field.
73  *
74  *      Dirty and clean buffers that are being written are linked to
75  *      lru[LIST_DIRTY] with their lru_list field. When the write
76  *      finishes, the buffer cannot be relinked immediately (because we
77  *      are in an interrupt context and relinking requires process
78  *      context), so some clean-not-writing buffers can be held on
79  *      dirty_lru too.  They are later added to lru in the process
80  *      context.
81  */
82 struct dm_bufio_client {
83         struct mutex lock;
84
85         struct list_head lru[LIST_SIZE];
86         unsigned long n_buffers[LIST_SIZE];
87
88         struct block_device *bdev;
89         unsigned block_size;
90         s8 sectors_per_block_bits;
91         void (*alloc_callback)(struct dm_buffer *);
92         void (*write_callback)(struct dm_buffer *);
93
94         struct kmem_cache *slab_buffer;
95         struct kmem_cache *slab_cache;
96         struct dm_io_client *dm_io;
97
98         struct list_head reserved_buffers;
99         unsigned need_reserved_buffers;
100
101         unsigned minimum_buffers;
102
103         struct rb_root buffer_tree;
104         wait_queue_head_t free_buffer_wait;
105
106         sector_t start;
107
108         int async_write_error;
109
110         struct list_head client_list;
111         struct shrinker shrinker;
112 };
113
114 /*
115  * Buffer state bits.
116  */
117 #define B_READING       0
118 #define B_WRITING       1
119 #define B_DIRTY         2
120
121 /*
122  * Describes how the block was allocated:
123  * kmem_cache_alloc(), __get_free_pages() or vmalloc().
124  * See the comment at alloc_buffer_data.
125  */
126 enum data_mode {
127         DATA_MODE_SLAB = 0,
128         DATA_MODE_GET_FREE_PAGES = 1,
129         DATA_MODE_VMALLOC = 2,
130         DATA_MODE_LIMIT = 3
131 };
132
133 struct dm_buffer {
134         struct rb_node node;
135         struct list_head lru_list;
136         struct list_head global_list;
137         sector_t block;
138         void *data;
139         unsigned char data_mode;                /* DATA_MODE_* */
140         unsigned char list_mode;                /* LIST_* */
141         blk_status_t read_error;
142         blk_status_t write_error;
143         unsigned accessed;
144         unsigned hold_count;
145         unsigned long state;
146         unsigned long last_accessed;
147         unsigned dirty_start;
148         unsigned dirty_end;
149         unsigned write_start;
150         unsigned write_end;
151         struct dm_bufio_client *c;
152         struct list_head write_list;
153         void (*end_io)(struct dm_buffer *, blk_status_t);
154 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
155 #define MAX_STACK 10
156         unsigned int stack_len;
157         unsigned long stack_entries[MAX_STACK];
158 #endif
159 };
160
161 /*----------------------------------------------------------------*/
162
163 #define dm_bufio_in_request()   (!!current->bio_list)
164
165 static void dm_bufio_lock(struct dm_bufio_client *c)
166 {
167         mutex_lock_nested(&c->lock, dm_bufio_in_request());
168 }
169
170 static int dm_bufio_trylock(struct dm_bufio_client *c)
171 {
172         return mutex_trylock(&c->lock);
173 }
174
175 static void dm_bufio_unlock(struct dm_bufio_client *c)
176 {
177         mutex_unlock(&c->lock);
178 }
179
180 /*----------------------------------------------------------------*/
181
182 /*
183  * Default cache size: available memory divided by the ratio.
184  */
185 static unsigned long dm_bufio_default_cache_size;
186
187 /*
188  * Total cache size set by the user.
189  */
190 static unsigned long dm_bufio_cache_size;
191
192 /*
193  * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
194  * at any time.  If it disagrees, the user has changed cache size.
195  */
196 static unsigned long dm_bufio_cache_size_latch;
197
198 static DEFINE_SPINLOCK(global_spinlock);
199
200 static LIST_HEAD(global_queue);
201
202 static unsigned long global_num = 0;
203
204 /*
205  * Buffers are freed after this timeout
206  */
207 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
208 static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
209
210 static unsigned long dm_bufio_peak_allocated;
211 static unsigned long dm_bufio_allocated_kmem_cache;
212 static unsigned long dm_bufio_allocated_get_free_pages;
213 static unsigned long dm_bufio_allocated_vmalloc;
214 static unsigned long dm_bufio_current_allocated;
215
216 /*----------------------------------------------------------------*/
217
218 /*
219  * The current number of clients.
220  */
221 static int dm_bufio_client_count;
222
223 /*
224  * The list of all clients.
225  */
226 static LIST_HEAD(dm_bufio_all_clients);
227
228 /*
229  * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
230  */
231 static DEFINE_MUTEX(dm_bufio_clients_lock);
232
233 static struct workqueue_struct *dm_bufio_wq;
234 static struct delayed_work dm_bufio_cleanup_old_work;
235 static struct work_struct dm_bufio_replacement_work;
236
237
238 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
239 static void buffer_record_stack(struct dm_buffer *b)
240 {
241         b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
242 }
243 #endif
244
245 /*----------------------------------------------------------------
246  * A red/black tree acts as an index for all the buffers.
247  *--------------------------------------------------------------*/
248 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
249 {
250         struct rb_node *n = c->buffer_tree.rb_node;
251         struct dm_buffer *b;
252
253         while (n) {
254                 b = container_of(n, struct dm_buffer, node);
255
256                 if (b->block == block)
257                         return b;
258
259                 n = (b->block < block) ? n->rb_left : n->rb_right;
260         }
261
262         return NULL;
263 }
264
265 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
266 {
267         struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
268         struct dm_buffer *found;
269
270         while (*new) {
271                 found = container_of(*new, struct dm_buffer, node);
272
273                 if (found->block == b->block) {
274                         BUG_ON(found != b);
275                         return;
276                 }
277
278                 parent = *new;
279                 new = (found->block < b->block) ?
280                         &((*new)->rb_left) : &((*new)->rb_right);
281         }
282
283         rb_link_node(&b->node, parent, new);
284         rb_insert_color(&b->node, &c->buffer_tree);
285 }
286
287 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
288 {
289         rb_erase(&b->node, &c->buffer_tree);
290 }
291
292 /*----------------------------------------------------------------*/
293
294 static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
295 {
296         unsigned char data_mode;
297         long diff;
298
299         static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
300                 &dm_bufio_allocated_kmem_cache,
301                 &dm_bufio_allocated_get_free_pages,
302                 &dm_bufio_allocated_vmalloc,
303         };
304
305         data_mode = b->data_mode;
306         diff = (long)b->c->block_size;
307         if (unlink)
308                 diff = -diff;
309
310         spin_lock(&global_spinlock);
311
312         *class_ptr[data_mode] += diff;
313
314         dm_bufio_current_allocated += diff;
315
316         if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
317                 dm_bufio_peak_allocated = dm_bufio_current_allocated;
318
319         b->accessed = 1;
320
321         if (!unlink) {
322                 list_add(&b->global_list, &global_queue);
323                 global_num++;
324                 if (dm_bufio_current_allocated > dm_bufio_cache_size)
325                         queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
326         } else {
327                 list_del(&b->global_list);
328                 global_num--;
329         }
330
331         spin_unlock(&global_spinlock);
332 }
333
334 /*
335  * Change the number of clients and recalculate per-client limit.
336  */
337 static void __cache_size_refresh(void)
338 {
339         BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
340         BUG_ON(dm_bufio_client_count < 0);
341
342         dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
343
344         /*
345          * Use default if set to 0 and report the actual cache size used.
346          */
347         if (!dm_bufio_cache_size_latch) {
348                 (void)cmpxchg(&dm_bufio_cache_size, 0,
349                               dm_bufio_default_cache_size);
350                 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
351         }
352 }
353
354 /*
355  * Allocating buffer data.
356  *
357  * Small buffers are allocated with kmem_cache, to use space optimally.
358  *
359  * For large buffers, we choose between get_free_pages and vmalloc.
360  * Each has advantages and disadvantages.
361  *
362  * __get_free_pages can randomly fail if the memory is fragmented.
363  * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
364  * as low as 128M) so using it for caching is not appropriate.
365  *
366  * If the allocation may fail we use __get_free_pages. Memory fragmentation
367  * won't have a fatal effect here, but it just causes flushes of some other
368  * buffers and more I/O will be performed. Don't use __get_free_pages if it
369  * always fails (i.e. order >= MAX_ORDER).
370  *
371  * If the allocation shouldn't fail we use __vmalloc. This is only for the
372  * initial reserve allocation, so there's no risk of wasting all vmalloc
373  * space.
374  */
375 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
376                                unsigned char *data_mode)
377 {
378         if (unlikely(c->slab_cache != NULL)) {
379                 *data_mode = DATA_MODE_SLAB;
380                 return kmem_cache_alloc(c->slab_cache, gfp_mask);
381         }
382
383         if (c->block_size <= KMALLOC_MAX_SIZE &&
384             gfp_mask & __GFP_NORETRY) {
385                 *data_mode = DATA_MODE_GET_FREE_PAGES;
386                 return (void *)__get_free_pages(gfp_mask,
387                                                 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
388         }
389
390         *data_mode = DATA_MODE_VMALLOC;
391
392         /*
393          * __vmalloc allocates the data pages and auxiliary structures with
394          * gfp_flags that were specified, but pagetables are always allocated
395          * with GFP_KERNEL, no matter what was specified as gfp_mask.
396          *
397          * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
398          * all allocations done by this process (including pagetables) are done
399          * as if GFP_NOIO was specified.
400          */
401         if (gfp_mask & __GFP_NORETRY) {
402                 unsigned noio_flag = memalloc_noio_save();
403                 void *ptr = __vmalloc(c->block_size, gfp_mask);
404
405                 memalloc_noio_restore(noio_flag);
406                 return ptr;
407         }
408
409         return __vmalloc(c->block_size, gfp_mask);
410 }
411
412 /*
413  * Free buffer's data.
414  */
415 static void free_buffer_data(struct dm_bufio_client *c,
416                              void *data, unsigned char data_mode)
417 {
418         switch (data_mode) {
419         case DATA_MODE_SLAB:
420                 kmem_cache_free(c->slab_cache, data);
421                 break;
422
423         case DATA_MODE_GET_FREE_PAGES:
424                 free_pages((unsigned long)data,
425                            c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
426                 break;
427
428         case DATA_MODE_VMALLOC:
429                 vfree(data);
430                 break;
431
432         default:
433                 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
434                        data_mode);
435                 BUG();
436         }
437 }
438
439 /*
440  * Allocate buffer and its data.
441  */
442 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
443 {
444         struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
445
446         if (!b)
447                 return NULL;
448
449         b->c = c;
450
451         b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
452         if (!b->data) {
453                 kmem_cache_free(c->slab_buffer, b);
454                 return NULL;
455         }
456
457 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
458         b->stack_len = 0;
459 #endif
460         return b;
461 }
462
463 /*
464  * Free buffer and its data.
465  */
466 static void free_buffer(struct dm_buffer *b)
467 {
468         struct dm_bufio_client *c = b->c;
469
470         free_buffer_data(c, b->data, b->data_mode);
471         kmem_cache_free(c->slab_buffer, b);
472 }
473
474 /*
475  * Link buffer to the buffer tree and clean or dirty queue.
476  */
477 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
478 {
479         struct dm_bufio_client *c = b->c;
480
481         c->n_buffers[dirty]++;
482         b->block = block;
483         b->list_mode = dirty;
484         list_add(&b->lru_list, &c->lru[dirty]);
485         __insert(b->c, b);
486         b->last_accessed = jiffies;
487
488         adjust_total_allocated(b, false);
489 }
490
491 /*
492  * Unlink buffer from the buffer tree and dirty or clean queue.
493  */
494 static void __unlink_buffer(struct dm_buffer *b)
495 {
496         struct dm_bufio_client *c = b->c;
497
498         BUG_ON(!c->n_buffers[b->list_mode]);
499
500         c->n_buffers[b->list_mode]--;
501         __remove(b->c, b);
502         list_del(&b->lru_list);
503
504         adjust_total_allocated(b, true);
505 }
506
507 /*
508  * Place the buffer to the head of dirty or clean LRU queue.
509  */
510 static void __relink_lru(struct dm_buffer *b, int dirty)
511 {
512         struct dm_bufio_client *c = b->c;
513
514         b->accessed = 1;
515
516         BUG_ON(!c->n_buffers[b->list_mode]);
517
518         c->n_buffers[b->list_mode]--;
519         c->n_buffers[dirty]++;
520         b->list_mode = dirty;
521         list_move(&b->lru_list, &c->lru[dirty]);
522         b->last_accessed = jiffies;
523 }
524
525 /*----------------------------------------------------------------
526  * Submit I/O on the buffer.
527  *
528  * Bio interface is faster but it has some problems:
529  *      the vector list is limited (increasing this limit increases
530  *      memory-consumption per buffer, so it is not viable);
531  *
532  *      the memory must be direct-mapped, not vmalloced;
533  *
534  * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
535  * it is not vmalloced, try using the bio interface.
536  *
537  * If the buffer is big, if it is vmalloced or if the underlying device
538  * rejects the bio because it is too large, use dm-io layer to do the I/O.
539  * The dm-io layer splits the I/O into multiple requests, avoiding the above
540  * shortcomings.
541  *--------------------------------------------------------------*/
542
543 /*
544  * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
545  * that the request was handled directly with bio interface.
546  */
547 static void dmio_complete(unsigned long error, void *context)
548 {
549         struct dm_buffer *b = context;
550
551         b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
552 }
553
554 static void use_dmio(struct dm_buffer *b, int rw, sector_t sector,
555                      unsigned n_sectors, unsigned offset)
556 {
557         int r;
558         struct dm_io_request io_req = {
559                 .bi_op = rw,
560                 .bi_op_flags = 0,
561                 .notify.fn = dmio_complete,
562                 .notify.context = b,
563                 .client = b->c->dm_io,
564         };
565         struct dm_io_region region = {
566                 .bdev = b->c->bdev,
567                 .sector = sector,
568                 .count = n_sectors,
569         };
570
571         if (b->data_mode != DATA_MODE_VMALLOC) {
572                 io_req.mem.type = DM_IO_KMEM;
573                 io_req.mem.ptr.addr = (char *)b->data + offset;
574         } else {
575                 io_req.mem.type = DM_IO_VMA;
576                 io_req.mem.ptr.vma = (char *)b->data + offset;
577         }
578
579         r = dm_io(&io_req, 1, &region, NULL);
580         if (unlikely(r))
581                 b->end_io(b, errno_to_blk_status(r));
582 }
583
584 static void bio_complete(struct bio *bio)
585 {
586         struct dm_buffer *b = bio->bi_private;
587         blk_status_t status = bio->bi_status;
588         bio_put(bio);
589         b->end_io(b, status);
590 }
591
592 static void use_bio(struct dm_buffer *b, int rw, sector_t sector,
593                     unsigned n_sectors, unsigned offset)
594 {
595         struct bio *bio;
596         char *ptr;
597         unsigned vec_size, len;
598
599         vec_size = b->c->block_size >> PAGE_SHIFT;
600         if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
601                 vec_size += 2;
602
603         bio = bio_kmalloc(GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN, vec_size);
604         if (!bio) {
605 dmio:
606                 use_dmio(b, rw, sector, n_sectors, offset);
607                 return;
608         }
609
610         bio->bi_iter.bi_sector = sector;
611         bio_set_dev(bio, b->c->bdev);
612         bio_set_op_attrs(bio, rw, 0);
613         bio->bi_end_io = bio_complete;
614         bio->bi_private = b;
615
616         ptr = (char *)b->data + offset;
617         len = n_sectors << SECTOR_SHIFT;
618
619         do {
620                 unsigned this_step = min((unsigned)(PAGE_SIZE - offset_in_page(ptr)), len);
621                 if (!bio_add_page(bio, virt_to_page(ptr), this_step,
622                                   offset_in_page(ptr))) {
623                         bio_put(bio);
624                         goto dmio;
625                 }
626
627                 len -= this_step;
628                 ptr += this_step;
629         } while (len > 0);
630
631         submit_bio(bio);
632 }
633
634 static void submit_io(struct dm_buffer *b, int rw, void (*end_io)(struct dm_buffer *, blk_status_t))
635 {
636         unsigned n_sectors;
637         sector_t sector;
638         unsigned offset, end;
639
640         b->end_io = end_io;
641
642         if (likely(b->c->sectors_per_block_bits >= 0))
643                 sector = b->block << b->c->sectors_per_block_bits;
644         else
645                 sector = b->block * (b->c->block_size >> SECTOR_SHIFT);
646         sector += b->c->start;
647
648         if (rw != REQ_OP_WRITE) {
649                 n_sectors = b->c->block_size >> SECTOR_SHIFT;
650                 offset = 0;
651         } else {
652                 if (b->c->write_callback)
653                         b->c->write_callback(b);
654                 offset = b->write_start;
655                 end = b->write_end;
656                 offset &= -DM_BUFIO_WRITE_ALIGN;
657                 end += DM_BUFIO_WRITE_ALIGN - 1;
658                 end &= -DM_BUFIO_WRITE_ALIGN;
659                 if (unlikely(end > b->c->block_size))
660                         end = b->c->block_size;
661
662                 sector += offset >> SECTOR_SHIFT;
663                 n_sectors = (end - offset) >> SECTOR_SHIFT;
664         }
665
666         if (b->data_mode != DATA_MODE_VMALLOC)
667                 use_bio(b, rw, sector, n_sectors, offset);
668         else
669                 use_dmio(b, rw, sector, n_sectors, offset);
670 }
671
672 /*----------------------------------------------------------------
673  * Writing dirty buffers
674  *--------------------------------------------------------------*/
675
676 /*
677  * The endio routine for write.
678  *
679  * Set the error, clear B_WRITING bit and wake anyone who was waiting on
680  * it.
681  */
682 static void write_endio(struct dm_buffer *b, blk_status_t status)
683 {
684         b->write_error = status;
685         if (unlikely(status)) {
686                 struct dm_bufio_client *c = b->c;
687
688                 (void)cmpxchg(&c->async_write_error, 0,
689                                 blk_status_to_errno(status));
690         }
691
692         BUG_ON(!test_bit(B_WRITING, &b->state));
693
694         smp_mb__before_atomic();
695         clear_bit(B_WRITING, &b->state);
696         smp_mb__after_atomic();
697
698         wake_up_bit(&b->state, B_WRITING);
699 }
700
701 /*
702  * Initiate a write on a dirty buffer, but don't wait for it.
703  *
704  * - If the buffer is not dirty, exit.
705  * - If there some previous write going on, wait for it to finish (we can't
706  *   have two writes on the same buffer simultaneously).
707  * - Submit our write and don't wait on it. We set B_WRITING indicating
708  *   that there is a write in progress.
709  */
710 static void __write_dirty_buffer(struct dm_buffer *b,
711                                  struct list_head *write_list)
712 {
713         if (!test_bit(B_DIRTY, &b->state))
714                 return;
715
716         clear_bit(B_DIRTY, &b->state);
717         wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
718
719         b->write_start = b->dirty_start;
720         b->write_end = b->dirty_end;
721
722         if (!write_list)
723                 submit_io(b, REQ_OP_WRITE, write_endio);
724         else
725                 list_add_tail(&b->write_list, write_list);
726 }
727
728 static void __flush_write_list(struct list_head *write_list)
729 {
730         struct blk_plug plug;
731         blk_start_plug(&plug);
732         while (!list_empty(write_list)) {
733                 struct dm_buffer *b =
734                         list_entry(write_list->next, struct dm_buffer, write_list);
735                 list_del(&b->write_list);
736                 submit_io(b, REQ_OP_WRITE, write_endio);
737                 cond_resched();
738         }
739         blk_finish_plug(&plug);
740 }
741
742 /*
743  * Wait until any activity on the buffer finishes.  Possibly write the
744  * buffer if it is dirty.  When this function finishes, there is no I/O
745  * running on the buffer and the buffer is not dirty.
746  */
747 static void __make_buffer_clean(struct dm_buffer *b)
748 {
749         BUG_ON(b->hold_count);
750
751         if (!b->state)  /* fast case */
752                 return;
753
754         wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
755         __write_dirty_buffer(b, NULL);
756         wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
757 }
758
759 /*
760  * Find some buffer that is not held by anybody, clean it, unlink it and
761  * return it.
762  */
763 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
764 {
765         struct dm_buffer *b;
766
767         list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
768                 BUG_ON(test_bit(B_WRITING, &b->state));
769                 BUG_ON(test_bit(B_DIRTY, &b->state));
770
771                 if (!b->hold_count) {
772                         __make_buffer_clean(b);
773                         __unlink_buffer(b);
774                         return b;
775                 }
776                 cond_resched();
777         }
778
779         list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
780                 BUG_ON(test_bit(B_READING, &b->state));
781
782                 if (!b->hold_count) {
783                         __make_buffer_clean(b);
784                         __unlink_buffer(b);
785                         return b;
786                 }
787                 cond_resched();
788         }
789
790         return NULL;
791 }
792
793 /*
794  * Wait until some other threads free some buffer or release hold count on
795  * some buffer.
796  *
797  * This function is entered with c->lock held, drops it and regains it
798  * before exiting.
799  */
800 static void __wait_for_free_buffer(struct dm_bufio_client *c)
801 {
802         DECLARE_WAITQUEUE(wait, current);
803
804         add_wait_queue(&c->free_buffer_wait, &wait);
805         set_current_state(TASK_UNINTERRUPTIBLE);
806         dm_bufio_unlock(c);
807
808         io_schedule();
809
810         remove_wait_queue(&c->free_buffer_wait, &wait);
811
812         dm_bufio_lock(c);
813 }
814
815 enum new_flag {
816         NF_FRESH = 0,
817         NF_READ = 1,
818         NF_GET = 2,
819         NF_PREFETCH = 3
820 };
821
822 /*
823  * Allocate a new buffer. If the allocation is not possible, wait until
824  * some other thread frees a buffer.
825  *
826  * May drop the lock and regain it.
827  */
828 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
829 {
830         struct dm_buffer *b;
831         bool tried_noio_alloc = false;
832
833         /*
834          * dm-bufio is resistant to allocation failures (it just keeps
835          * one buffer reserved in cases all the allocations fail).
836          * So set flags to not try too hard:
837          *      GFP_NOWAIT: don't wait; if we need to sleep we'll release our
838          *                  mutex and wait ourselves.
839          *      __GFP_NORETRY: don't retry and rather return failure
840          *      __GFP_NOMEMALLOC: don't use emergency reserves
841          *      __GFP_NOWARN: don't print a warning in case of failure
842          *
843          * For debugging, if we set the cache size to 1, no new buffers will
844          * be allocated.
845          */
846         while (1) {
847                 if (dm_bufio_cache_size_latch != 1) {
848                         b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
849                         if (b)
850                                 return b;
851                 }
852
853                 if (nf == NF_PREFETCH)
854                         return NULL;
855
856                 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
857                         dm_bufio_unlock(c);
858                         b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
859                         dm_bufio_lock(c);
860                         if (b)
861                                 return b;
862                         tried_noio_alloc = true;
863                 }
864
865                 if (!list_empty(&c->reserved_buffers)) {
866                         b = list_entry(c->reserved_buffers.next,
867                                        struct dm_buffer, lru_list);
868                         list_del(&b->lru_list);
869                         c->need_reserved_buffers++;
870
871                         return b;
872                 }
873
874                 b = __get_unclaimed_buffer(c);
875                 if (b)
876                         return b;
877
878                 __wait_for_free_buffer(c);
879         }
880 }
881
882 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
883 {
884         struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
885
886         if (!b)
887                 return NULL;
888
889         if (c->alloc_callback)
890                 c->alloc_callback(b);
891
892         return b;
893 }
894
895 /*
896  * Free a buffer and wake other threads waiting for free buffers.
897  */
898 static void __free_buffer_wake(struct dm_buffer *b)
899 {
900         struct dm_bufio_client *c = b->c;
901
902         if (!c->need_reserved_buffers)
903                 free_buffer(b);
904         else {
905                 list_add(&b->lru_list, &c->reserved_buffers);
906                 c->need_reserved_buffers--;
907         }
908
909         wake_up(&c->free_buffer_wait);
910 }
911
912 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
913                                         struct list_head *write_list)
914 {
915         struct dm_buffer *b, *tmp;
916
917         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
918                 BUG_ON(test_bit(B_READING, &b->state));
919
920                 if (!test_bit(B_DIRTY, &b->state) &&
921                     !test_bit(B_WRITING, &b->state)) {
922                         __relink_lru(b, LIST_CLEAN);
923                         continue;
924                 }
925
926                 if (no_wait && test_bit(B_WRITING, &b->state))
927                         return;
928
929                 __write_dirty_buffer(b, write_list);
930                 cond_resched();
931         }
932 }
933
934 /*
935  * Check if we're over watermark.
936  * If we are over threshold_buffers, start freeing buffers.
937  * If we're over "limit_buffers", block until we get under the limit.
938  */
939 static void __check_watermark(struct dm_bufio_client *c,
940                               struct list_head *write_list)
941 {
942         if (c->n_buffers[LIST_DIRTY] > c->n_buffers[LIST_CLEAN] * DM_BUFIO_WRITEBACK_RATIO)
943                 __write_dirty_buffers_async(c, 1, write_list);
944 }
945
946 /*----------------------------------------------------------------
947  * Getting a buffer
948  *--------------------------------------------------------------*/
949
950 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
951                                      enum new_flag nf, int *need_submit,
952                                      struct list_head *write_list)
953 {
954         struct dm_buffer *b, *new_b = NULL;
955
956         *need_submit = 0;
957
958         b = __find(c, block);
959         if (b)
960                 goto found_buffer;
961
962         if (nf == NF_GET)
963                 return NULL;
964
965         new_b = __alloc_buffer_wait(c, nf);
966         if (!new_b)
967                 return NULL;
968
969         /*
970          * We've had a period where the mutex was unlocked, so need to
971          * recheck the buffer tree.
972          */
973         b = __find(c, block);
974         if (b) {
975                 __free_buffer_wake(new_b);
976                 goto found_buffer;
977         }
978
979         __check_watermark(c, write_list);
980
981         b = new_b;
982         b->hold_count = 1;
983         b->read_error = 0;
984         b->write_error = 0;
985         __link_buffer(b, block, LIST_CLEAN);
986
987         if (nf == NF_FRESH) {
988                 b->state = 0;
989                 return b;
990         }
991
992         b->state = 1 << B_READING;
993         *need_submit = 1;
994
995         return b;
996
997 found_buffer:
998         if (nf == NF_PREFETCH)
999                 return NULL;
1000         /*
1001          * Note: it is essential that we don't wait for the buffer to be
1002          * read if dm_bufio_get function is used. Both dm_bufio_get and
1003          * dm_bufio_prefetch can be used in the driver request routine.
1004          * If the user called both dm_bufio_prefetch and dm_bufio_get on
1005          * the same buffer, it would deadlock if we waited.
1006          */
1007         if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1008                 return NULL;
1009
1010         b->hold_count++;
1011         __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1012                      test_bit(B_WRITING, &b->state));
1013         return b;
1014 }
1015
1016 /*
1017  * The endio routine for reading: set the error, clear the bit and wake up
1018  * anyone waiting on the buffer.
1019  */
1020 static void read_endio(struct dm_buffer *b, blk_status_t status)
1021 {
1022         b->read_error = status;
1023
1024         BUG_ON(!test_bit(B_READING, &b->state));
1025
1026         smp_mb__before_atomic();
1027         clear_bit(B_READING, &b->state);
1028         smp_mb__after_atomic();
1029
1030         wake_up_bit(&b->state, B_READING);
1031 }
1032
1033 /*
1034  * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1035  * functions is similar except that dm_bufio_new doesn't read the
1036  * buffer from the disk (assuming that the caller overwrites all the data
1037  * and uses dm_bufio_mark_buffer_dirty to write new data back).
1038  */
1039 static void *new_read(struct dm_bufio_client *c, sector_t block,
1040                       enum new_flag nf, struct dm_buffer **bp)
1041 {
1042         int need_submit;
1043         struct dm_buffer *b;
1044
1045         LIST_HEAD(write_list);
1046
1047         dm_bufio_lock(c);
1048         b = __bufio_new(c, block, nf, &need_submit, &write_list);
1049 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1050         if (b && b->hold_count == 1)
1051                 buffer_record_stack(b);
1052 #endif
1053         dm_bufio_unlock(c);
1054
1055         __flush_write_list(&write_list);
1056
1057         if (!b)
1058                 return NULL;
1059
1060         if (need_submit)
1061                 submit_io(b, REQ_OP_READ, read_endio);
1062
1063         wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1064
1065         if (b->read_error) {
1066                 int error = blk_status_to_errno(b->read_error);
1067
1068                 dm_bufio_release(b);
1069
1070                 return ERR_PTR(error);
1071         }
1072
1073         *bp = b;
1074
1075         return b->data;
1076 }
1077
1078 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1079                    struct dm_buffer **bp)
1080 {
1081         return new_read(c, block, NF_GET, bp);
1082 }
1083 EXPORT_SYMBOL_GPL(dm_bufio_get);
1084
1085 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1086                     struct dm_buffer **bp)
1087 {
1088         BUG_ON(dm_bufio_in_request());
1089
1090         return new_read(c, block, NF_READ, bp);
1091 }
1092 EXPORT_SYMBOL_GPL(dm_bufio_read);
1093
1094 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1095                    struct dm_buffer **bp)
1096 {
1097         BUG_ON(dm_bufio_in_request());
1098
1099         return new_read(c, block, NF_FRESH, bp);
1100 }
1101 EXPORT_SYMBOL_GPL(dm_bufio_new);
1102
1103 void dm_bufio_prefetch(struct dm_bufio_client *c,
1104                        sector_t block, unsigned n_blocks)
1105 {
1106         struct blk_plug plug;
1107
1108         LIST_HEAD(write_list);
1109
1110         BUG_ON(dm_bufio_in_request());
1111
1112         blk_start_plug(&plug);
1113         dm_bufio_lock(c);
1114
1115         for (; n_blocks--; block++) {
1116                 int need_submit;
1117                 struct dm_buffer *b;
1118                 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1119                                 &write_list);
1120                 if (unlikely(!list_empty(&write_list))) {
1121                         dm_bufio_unlock(c);
1122                         blk_finish_plug(&plug);
1123                         __flush_write_list(&write_list);
1124                         blk_start_plug(&plug);
1125                         dm_bufio_lock(c);
1126                 }
1127                 if (unlikely(b != NULL)) {
1128                         dm_bufio_unlock(c);
1129
1130                         if (need_submit)
1131                                 submit_io(b, REQ_OP_READ, read_endio);
1132                         dm_bufio_release(b);
1133
1134                         cond_resched();
1135
1136                         if (!n_blocks)
1137                                 goto flush_plug;
1138                         dm_bufio_lock(c);
1139                 }
1140         }
1141
1142         dm_bufio_unlock(c);
1143
1144 flush_plug:
1145         blk_finish_plug(&plug);
1146 }
1147 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1148
1149 void dm_bufio_release(struct dm_buffer *b)
1150 {
1151         struct dm_bufio_client *c = b->c;
1152
1153         dm_bufio_lock(c);
1154
1155         BUG_ON(!b->hold_count);
1156
1157         b->hold_count--;
1158         if (!b->hold_count) {
1159                 wake_up(&c->free_buffer_wait);
1160
1161                 /*
1162                  * If there were errors on the buffer, and the buffer is not
1163                  * to be written, free the buffer. There is no point in caching
1164                  * invalid buffer.
1165                  */
1166                 if ((b->read_error || b->write_error) &&
1167                     !test_bit(B_READING, &b->state) &&
1168                     !test_bit(B_WRITING, &b->state) &&
1169                     !test_bit(B_DIRTY, &b->state)) {
1170                         __unlink_buffer(b);
1171                         __free_buffer_wake(b);
1172                 }
1173         }
1174
1175         dm_bufio_unlock(c);
1176 }
1177 EXPORT_SYMBOL_GPL(dm_bufio_release);
1178
1179 void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1180                                         unsigned start, unsigned end)
1181 {
1182         struct dm_bufio_client *c = b->c;
1183
1184         BUG_ON(start >= end);
1185         BUG_ON(end > b->c->block_size);
1186
1187         dm_bufio_lock(c);
1188
1189         BUG_ON(test_bit(B_READING, &b->state));
1190
1191         if (!test_and_set_bit(B_DIRTY, &b->state)) {
1192                 b->dirty_start = start;
1193                 b->dirty_end = end;
1194                 __relink_lru(b, LIST_DIRTY);
1195         } else {
1196                 if (start < b->dirty_start)
1197                         b->dirty_start = start;
1198                 if (end > b->dirty_end)
1199                         b->dirty_end = end;
1200         }
1201
1202         dm_bufio_unlock(c);
1203 }
1204 EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1205
1206 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1207 {
1208         dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1209 }
1210 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1211
1212 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1213 {
1214         LIST_HEAD(write_list);
1215
1216         BUG_ON(dm_bufio_in_request());
1217
1218         dm_bufio_lock(c);
1219         __write_dirty_buffers_async(c, 0, &write_list);
1220         dm_bufio_unlock(c);
1221         __flush_write_list(&write_list);
1222 }
1223 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1224
1225 /*
1226  * For performance, it is essential that the buffers are written asynchronously
1227  * and simultaneously (so that the block layer can merge the writes) and then
1228  * waited upon.
1229  *
1230  * Finally, we flush hardware disk cache.
1231  */
1232 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1233 {
1234         int a, f;
1235         unsigned long buffers_processed = 0;
1236         struct dm_buffer *b, *tmp;
1237
1238         LIST_HEAD(write_list);
1239
1240         dm_bufio_lock(c);
1241         __write_dirty_buffers_async(c, 0, &write_list);
1242         dm_bufio_unlock(c);
1243         __flush_write_list(&write_list);
1244         dm_bufio_lock(c);
1245
1246 again:
1247         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1248                 int dropped_lock = 0;
1249
1250                 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1251                         buffers_processed++;
1252
1253                 BUG_ON(test_bit(B_READING, &b->state));
1254
1255                 if (test_bit(B_WRITING, &b->state)) {
1256                         if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1257                                 dropped_lock = 1;
1258                                 b->hold_count++;
1259                                 dm_bufio_unlock(c);
1260                                 wait_on_bit_io(&b->state, B_WRITING,
1261                                                TASK_UNINTERRUPTIBLE);
1262                                 dm_bufio_lock(c);
1263                                 b->hold_count--;
1264                         } else
1265                                 wait_on_bit_io(&b->state, B_WRITING,
1266                                                TASK_UNINTERRUPTIBLE);
1267                 }
1268
1269                 if (!test_bit(B_DIRTY, &b->state) &&
1270                     !test_bit(B_WRITING, &b->state))
1271                         __relink_lru(b, LIST_CLEAN);
1272
1273                 cond_resched();
1274
1275                 /*
1276                  * If we dropped the lock, the list is no longer consistent,
1277                  * so we must restart the search.
1278                  *
1279                  * In the most common case, the buffer just processed is
1280                  * relinked to the clean list, so we won't loop scanning the
1281                  * same buffer again and again.
1282                  *
1283                  * This may livelock if there is another thread simultaneously
1284                  * dirtying buffers, so we count the number of buffers walked
1285                  * and if it exceeds the total number of buffers, it means that
1286                  * someone is doing some writes simultaneously with us.  In
1287                  * this case, stop, dropping the lock.
1288                  */
1289                 if (dropped_lock)
1290                         goto again;
1291         }
1292         wake_up(&c->free_buffer_wait);
1293         dm_bufio_unlock(c);
1294
1295         a = xchg(&c->async_write_error, 0);
1296         f = dm_bufio_issue_flush(c);
1297         if (a)
1298                 return a;
1299
1300         return f;
1301 }
1302 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1303
1304 /*
1305  * Use dm-io to send an empty barrier to flush the device.
1306  */
1307 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1308 {
1309         struct dm_io_request io_req = {
1310                 .bi_op = REQ_OP_WRITE,
1311                 .bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
1312                 .mem.type = DM_IO_KMEM,
1313                 .mem.ptr.addr = NULL,
1314                 .client = c->dm_io,
1315         };
1316         struct dm_io_region io_reg = {
1317                 .bdev = c->bdev,
1318                 .sector = 0,
1319                 .count = 0,
1320         };
1321
1322         BUG_ON(dm_bufio_in_request());
1323
1324         return dm_io(&io_req, 1, &io_reg, NULL);
1325 }
1326 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1327
1328 /*
1329  * We first delete any other buffer that may be at that new location.
1330  *
1331  * Then, we write the buffer to the original location if it was dirty.
1332  *
1333  * Then, if we are the only one who is holding the buffer, relink the buffer
1334  * in the buffer tree for the new location.
1335  *
1336  * If there was someone else holding the buffer, we write it to the new
1337  * location but not relink it, because that other user needs to have the buffer
1338  * at the same place.
1339  */
1340 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1341 {
1342         struct dm_bufio_client *c = b->c;
1343         struct dm_buffer *new;
1344
1345         BUG_ON(dm_bufio_in_request());
1346
1347         dm_bufio_lock(c);
1348
1349 retry:
1350         new = __find(c, new_block);
1351         if (new) {
1352                 if (new->hold_count) {
1353                         __wait_for_free_buffer(c);
1354                         goto retry;
1355                 }
1356
1357                 /*
1358                  * FIXME: Is there any point waiting for a write that's going
1359                  * to be overwritten in a bit?
1360                  */
1361                 __make_buffer_clean(new);
1362                 __unlink_buffer(new);
1363                 __free_buffer_wake(new);
1364         }
1365
1366         BUG_ON(!b->hold_count);
1367         BUG_ON(test_bit(B_READING, &b->state));
1368
1369         __write_dirty_buffer(b, NULL);
1370         if (b->hold_count == 1) {
1371                 wait_on_bit_io(&b->state, B_WRITING,
1372                                TASK_UNINTERRUPTIBLE);
1373                 set_bit(B_DIRTY, &b->state);
1374                 b->dirty_start = 0;
1375                 b->dirty_end = c->block_size;
1376                 __unlink_buffer(b);
1377                 __link_buffer(b, new_block, LIST_DIRTY);
1378         } else {
1379                 sector_t old_block;
1380                 wait_on_bit_lock_io(&b->state, B_WRITING,
1381                                     TASK_UNINTERRUPTIBLE);
1382                 /*
1383                  * Relink buffer to "new_block" so that write_callback
1384                  * sees "new_block" as a block number.
1385                  * After the write, link the buffer back to old_block.
1386                  * All this must be done in bufio lock, so that block number
1387                  * change isn't visible to other threads.
1388                  */
1389                 old_block = b->block;
1390                 __unlink_buffer(b);
1391                 __link_buffer(b, new_block, b->list_mode);
1392                 submit_io(b, REQ_OP_WRITE, write_endio);
1393                 wait_on_bit_io(&b->state, B_WRITING,
1394                                TASK_UNINTERRUPTIBLE);
1395                 __unlink_buffer(b);
1396                 __link_buffer(b, old_block, b->list_mode);
1397         }
1398
1399         dm_bufio_unlock(c);
1400         dm_bufio_release(b);
1401 }
1402 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1403
1404 /*
1405  * Free the given buffer.
1406  *
1407  * This is just a hint, if the buffer is in use or dirty, this function
1408  * does nothing.
1409  */
1410 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1411 {
1412         struct dm_buffer *b;
1413
1414         dm_bufio_lock(c);
1415
1416         b = __find(c, block);
1417         if (b && likely(!b->hold_count) && likely(!b->state)) {
1418                 __unlink_buffer(b);
1419                 __free_buffer_wake(b);
1420         }
1421
1422         dm_bufio_unlock(c);
1423 }
1424 EXPORT_SYMBOL_GPL(dm_bufio_forget);
1425
1426 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1427 {
1428         c->minimum_buffers = n;
1429 }
1430 EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1431
1432 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1433 {
1434         return c->block_size;
1435 }
1436 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1437
1438 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1439 {
1440         sector_t s = i_size_read(c->bdev->bd_inode) >> SECTOR_SHIFT;
1441         if (likely(c->sectors_per_block_bits >= 0))
1442                 s >>= c->sectors_per_block_bits;
1443         else
1444                 sector_div(s, c->block_size >> SECTOR_SHIFT);
1445         return s;
1446 }
1447 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1448
1449 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1450 {
1451         return b->block;
1452 }
1453 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1454
1455 void *dm_bufio_get_block_data(struct dm_buffer *b)
1456 {
1457         return b->data;
1458 }
1459 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1460
1461 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1462 {
1463         return b + 1;
1464 }
1465 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1466
1467 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1468 {
1469         return b->c;
1470 }
1471 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1472
1473 static void drop_buffers(struct dm_bufio_client *c)
1474 {
1475         struct dm_buffer *b;
1476         int i;
1477         bool warned = false;
1478
1479         BUG_ON(dm_bufio_in_request());
1480
1481         /*
1482          * An optimization so that the buffers are not written one-by-one.
1483          */
1484         dm_bufio_write_dirty_buffers_async(c);
1485
1486         dm_bufio_lock(c);
1487
1488         while ((b = __get_unclaimed_buffer(c)))
1489                 __free_buffer_wake(b);
1490
1491         for (i = 0; i < LIST_SIZE; i++)
1492                 list_for_each_entry(b, &c->lru[i], lru_list) {
1493                         WARN_ON(!warned);
1494                         warned = true;
1495                         DMERR("leaked buffer %llx, hold count %u, list %d",
1496                               (unsigned long long)b->block, b->hold_count, i);
1497 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1498                         stack_trace_print(b->stack_entries, b->stack_len, 1);
1499                         /* mark unclaimed to avoid BUG_ON below */
1500                         b->hold_count = 0;
1501 #endif
1502                 }
1503
1504 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1505         while ((b = __get_unclaimed_buffer(c)))
1506                 __free_buffer_wake(b);
1507 #endif
1508
1509         for (i = 0; i < LIST_SIZE; i++)
1510                 BUG_ON(!list_empty(&c->lru[i]));
1511
1512         dm_bufio_unlock(c);
1513 }
1514
1515 /*
1516  * We may not be able to evict this buffer if IO pending or the client
1517  * is still using it.  Caller is expected to know buffer is too old.
1518  *
1519  * And if GFP_NOFS is used, we must not do any I/O because we hold
1520  * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1521  * rerouted to different bufio client.
1522  */
1523 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1524 {
1525         if (!(gfp & __GFP_FS)) {
1526                 if (test_bit(B_READING, &b->state) ||
1527                     test_bit(B_WRITING, &b->state) ||
1528                     test_bit(B_DIRTY, &b->state))
1529                         return false;
1530         }
1531
1532         if (b->hold_count)
1533                 return false;
1534
1535         __make_buffer_clean(b);
1536         __unlink_buffer(b);
1537         __free_buffer_wake(b);
1538
1539         return true;
1540 }
1541
1542 static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1543 {
1544         unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
1545         if (likely(c->sectors_per_block_bits >= 0))
1546                 retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
1547         else
1548                 retain_bytes /= c->block_size;
1549         return retain_bytes;
1550 }
1551
1552 static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1553                             gfp_t gfp_mask)
1554 {
1555         int l;
1556         struct dm_buffer *b, *tmp;
1557         unsigned long freed = 0;
1558         unsigned long count = c->n_buffers[LIST_CLEAN] +
1559                               c->n_buffers[LIST_DIRTY];
1560         unsigned long retain_target = get_retain_buffers(c);
1561
1562         for (l = 0; l < LIST_SIZE; l++) {
1563                 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1564                         if (__try_evict_buffer(b, gfp_mask))
1565                                 freed++;
1566                         if (!--nr_to_scan || ((count - freed) <= retain_target))
1567                                 return freed;
1568                         cond_resched();
1569                 }
1570         }
1571         return freed;
1572 }
1573
1574 static unsigned long
1575 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1576 {
1577         struct dm_bufio_client *c;
1578         unsigned long freed;
1579
1580         c = container_of(shrink, struct dm_bufio_client, shrinker);
1581         if (sc->gfp_mask & __GFP_FS)
1582                 dm_bufio_lock(c);
1583         else if (!dm_bufio_trylock(c))
1584                 return SHRINK_STOP;
1585
1586         freed  = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1587         dm_bufio_unlock(c);
1588         return freed;
1589 }
1590
1591 static unsigned long
1592 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1593 {
1594         struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1595         unsigned long count = READ_ONCE(c->n_buffers[LIST_CLEAN]) +
1596                               READ_ONCE(c->n_buffers[LIST_DIRTY]);
1597         unsigned long retain_target = get_retain_buffers(c);
1598
1599         return (count < retain_target) ? 0 : (count - retain_target);
1600 }
1601
1602 /*
1603  * Create the buffering interface
1604  */
1605 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1606                                                unsigned reserved_buffers, unsigned aux_size,
1607                                                void (*alloc_callback)(struct dm_buffer *),
1608                                                void (*write_callback)(struct dm_buffer *))
1609 {
1610         int r;
1611         struct dm_bufio_client *c;
1612         unsigned i;
1613         char slab_name[27];
1614
1615         if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1616                 DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1617                 r = -EINVAL;
1618                 goto bad_client;
1619         }
1620
1621         c = kzalloc(sizeof(*c), GFP_KERNEL);
1622         if (!c) {
1623                 r = -ENOMEM;
1624                 goto bad_client;
1625         }
1626         c->buffer_tree = RB_ROOT;
1627
1628         c->bdev = bdev;
1629         c->block_size = block_size;
1630         if (is_power_of_2(block_size))
1631                 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1632         else
1633                 c->sectors_per_block_bits = -1;
1634
1635         c->alloc_callback = alloc_callback;
1636         c->write_callback = write_callback;
1637
1638         for (i = 0; i < LIST_SIZE; i++) {
1639                 INIT_LIST_HEAD(&c->lru[i]);
1640                 c->n_buffers[i] = 0;
1641         }
1642
1643         mutex_init(&c->lock);
1644         INIT_LIST_HEAD(&c->reserved_buffers);
1645         c->need_reserved_buffers = reserved_buffers;
1646
1647         dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1648
1649         init_waitqueue_head(&c->free_buffer_wait);
1650         c->async_write_error = 0;
1651
1652         c->dm_io = dm_io_client_create();
1653         if (IS_ERR(c->dm_io)) {
1654                 r = PTR_ERR(c->dm_io);
1655                 goto bad_dm_io;
1656         }
1657
1658         if (block_size <= KMALLOC_MAX_SIZE &&
1659             (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
1660                 unsigned align = min(1U << __ffs(block_size), (unsigned)PAGE_SIZE);
1661                 snprintf(slab_name, sizeof slab_name, "dm_bufio_cache-%u", block_size);
1662                 c->slab_cache = kmem_cache_create(slab_name, block_size, align,
1663                                                   SLAB_RECLAIM_ACCOUNT, NULL);
1664                 if (!c->slab_cache) {
1665                         r = -ENOMEM;
1666                         goto bad;
1667                 }
1668         }
1669         if (aux_size)
1670                 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer-%u", aux_size);
1671         else
1672                 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer");
1673         c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
1674                                            0, SLAB_RECLAIM_ACCOUNT, NULL);
1675         if (!c->slab_buffer) {
1676                 r = -ENOMEM;
1677                 goto bad;
1678         }
1679
1680         while (c->need_reserved_buffers) {
1681                 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1682
1683                 if (!b) {
1684                         r = -ENOMEM;
1685                         goto bad;
1686                 }
1687                 __free_buffer_wake(b);
1688         }
1689
1690         c->shrinker.count_objects = dm_bufio_shrink_count;
1691         c->shrinker.scan_objects = dm_bufio_shrink_scan;
1692         c->shrinker.seeks = 1;
1693         c->shrinker.batch = 0;
1694         r = register_shrinker(&c->shrinker);
1695         if (r)
1696                 goto bad;
1697
1698         mutex_lock(&dm_bufio_clients_lock);
1699         dm_bufio_client_count++;
1700         list_add(&c->client_list, &dm_bufio_all_clients);
1701         __cache_size_refresh();
1702         mutex_unlock(&dm_bufio_clients_lock);
1703
1704         return c;
1705
1706 bad:
1707         while (!list_empty(&c->reserved_buffers)) {
1708                 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1709                                                  struct dm_buffer, lru_list);
1710                 list_del(&b->lru_list);
1711                 free_buffer(b);
1712         }
1713         kmem_cache_destroy(c->slab_cache);
1714         kmem_cache_destroy(c->slab_buffer);
1715         dm_io_client_destroy(c->dm_io);
1716 bad_dm_io:
1717         mutex_destroy(&c->lock);
1718         kfree(c);
1719 bad_client:
1720         return ERR_PTR(r);
1721 }
1722 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1723
1724 /*
1725  * Free the buffering interface.
1726  * It is required that there are no references on any buffers.
1727  */
1728 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1729 {
1730         unsigned i;
1731
1732         drop_buffers(c);
1733
1734         unregister_shrinker(&c->shrinker);
1735
1736         mutex_lock(&dm_bufio_clients_lock);
1737
1738         list_del(&c->client_list);
1739         dm_bufio_client_count--;
1740         __cache_size_refresh();
1741
1742         mutex_unlock(&dm_bufio_clients_lock);
1743
1744         BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1745         BUG_ON(c->need_reserved_buffers);
1746
1747         while (!list_empty(&c->reserved_buffers)) {
1748                 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1749                                                  struct dm_buffer, lru_list);
1750                 list_del(&b->lru_list);
1751                 free_buffer(b);
1752         }
1753
1754         for (i = 0; i < LIST_SIZE; i++)
1755                 if (c->n_buffers[i])
1756                         DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1757
1758         for (i = 0; i < LIST_SIZE; i++)
1759                 BUG_ON(c->n_buffers[i]);
1760
1761         kmem_cache_destroy(c->slab_cache);
1762         kmem_cache_destroy(c->slab_buffer);
1763         dm_io_client_destroy(c->dm_io);
1764         mutex_destroy(&c->lock);
1765         kfree(c);
1766 }
1767 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1768
1769 void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1770 {
1771         c->start = start;
1772 }
1773 EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1774
1775 static unsigned get_max_age_hz(void)
1776 {
1777         unsigned max_age = READ_ONCE(dm_bufio_max_age);
1778
1779         if (max_age > UINT_MAX / HZ)
1780                 max_age = UINT_MAX / HZ;
1781
1782         return max_age * HZ;
1783 }
1784
1785 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1786 {
1787         return time_after_eq(jiffies, b->last_accessed + age_hz);
1788 }
1789
1790 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1791 {
1792         struct dm_buffer *b, *tmp;
1793         unsigned long retain_target = get_retain_buffers(c);
1794         unsigned long count;
1795         LIST_HEAD(write_list);
1796
1797         dm_bufio_lock(c);
1798
1799         __check_watermark(c, &write_list);
1800         if (unlikely(!list_empty(&write_list))) {
1801                 dm_bufio_unlock(c);
1802                 __flush_write_list(&write_list);
1803                 dm_bufio_lock(c);
1804         }
1805
1806         count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1807         list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1808                 if (count <= retain_target)
1809                         break;
1810
1811                 if (!older_than(b, age_hz))
1812                         break;
1813
1814                 if (__try_evict_buffer(b, 0))
1815                         count--;
1816
1817                 cond_resched();
1818         }
1819
1820         dm_bufio_unlock(c);
1821 }
1822
1823 static void do_global_cleanup(struct work_struct *w)
1824 {
1825         struct dm_bufio_client *locked_client = NULL;
1826         struct dm_bufio_client *current_client;
1827         struct dm_buffer *b;
1828         unsigned spinlock_hold_count;
1829         unsigned long threshold = dm_bufio_cache_size -
1830                 dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
1831         unsigned long loops = global_num * 2;
1832
1833         mutex_lock(&dm_bufio_clients_lock);
1834
1835         while (1) {
1836                 cond_resched();
1837
1838                 spin_lock(&global_spinlock);
1839                 if (unlikely(dm_bufio_current_allocated <= threshold))
1840                         break;
1841
1842                 spinlock_hold_count = 0;
1843 get_next:
1844                 if (!loops--)
1845                         break;
1846                 if (unlikely(list_empty(&global_queue)))
1847                         break;
1848                 b = list_entry(global_queue.prev, struct dm_buffer, global_list);
1849
1850                 if (b->accessed) {
1851                         b->accessed = 0;
1852                         list_move(&b->global_list, &global_queue);
1853                         if (likely(++spinlock_hold_count < 16))
1854                                 goto get_next;
1855                         spin_unlock(&global_spinlock);
1856                         continue;
1857                 }
1858
1859                 current_client = b->c;
1860                 if (unlikely(current_client != locked_client)) {
1861                         if (locked_client)
1862                                 dm_bufio_unlock(locked_client);
1863
1864                         if (!dm_bufio_trylock(current_client)) {
1865                                 spin_unlock(&global_spinlock);
1866                                 dm_bufio_lock(current_client);
1867                                 locked_client = current_client;
1868                                 continue;
1869                         }
1870
1871                         locked_client = current_client;
1872                 }
1873
1874                 spin_unlock(&global_spinlock);
1875
1876                 if (unlikely(!__try_evict_buffer(b, GFP_KERNEL))) {
1877                         spin_lock(&global_spinlock);
1878                         list_move(&b->global_list, &global_queue);
1879                         spin_unlock(&global_spinlock);
1880                 }
1881         }
1882
1883         spin_unlock(&global_spinlock);
1884
1885         if (locked_client)
1886                 dm_bufio_unlock(locked_client);
1887
1888         mutex_unlock(&dm_bufio_clients_lock);
1889 }
1890
1891 static void cleanup_old_buffers(void)
1892 {
1893         unsigned long max_age_hz = get_max_age_hz();
1894         struct dm_bufio_client *c;
1895
1896         mutex_lock(&dm_bufio_clients_lock);
1897
1898         __cache_size_refresh();
1899
1900         list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1901                 __evict_old_buffers(c, max_age_hz);
1902
1903         mutex_unlock(&dm_bufio_clients_lock);
1904 }
1905
1906 static void work_fn(struct work_struct *w)
1907 {
1908         cleanup_old_buffers();
1909
1910         queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
1911                            DM_BUFIO_WORK_TIMER_SECS * HZ);
1912 }
1913
1914 /*----------------------------------------------------------------
1915  * Module setup
1916  *--------------------------------------------------------------*/
1917
1918 /*
1919  * This is called only once for the whole dm_bufio module.
1920  * It initializes memory limit.
1921  */
1922 static int __init dm_bufio_init(void)
1923 {
1924         __u64 mem;
1925
1926         dm_bufio_allocated_kmem_cache = 0;
1927         dm_bufio_allocated_get_free_pages = 0;
1928         dm_bufio_allocated_vmalloc = 0;
1929         dm_bufio_current_allocated = 0;
1930
1931         mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
1932                                DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
1933
1934         if (mem > ULONG_MAX)
1935                 mem = ULONG_MAX;
1936
1937 #ifdef CONFIG_MMU
1938         if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
1939                 mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
1940 #endif
1941
1942         dm_bufio_default_cache_size = mem;
1943
1944         mutex_lock(&dm_bufio_clients_lock);
1945         __cache_size_refresh();
1946         mutex_unlock(&dm_bufio_clients_lock);
1947
1948         dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
1949         if (!dm_bufio_wq)
1950                 return -ENOMEM;
1951
1952         INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
1953         INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
1954         queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
1955                            DM_BUFIO_WORK_TIMER_SECS * HZ);
1956
1957         return 0;
1958 }
1959
1960 /*
1961  * This is called once when unloading the dm_bufio module.
1962  */
1963 static void __exit dm_bufio_exit(void)
1964 {
1965         int bug = 0;
1966
1967         cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
1968         flush_workqueue(dm_bufio_wq);
1969         destroy_workqueue(dm_bufio_wq);
1970
1971         if (dm_bufio_client_count) {
1972                 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1973                         __func__, dm_bufio_client_count);
1974                 bug = 1;
1975         }
1976
1977         if (dm_bufio_current_allocated) {
1978                 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1979                         __func__, dm_bufio_current_allocated);
1980                 bug = 1;
1981         }
1982
1983         if (dm_bufio_allocated_get_free_pages) {
1984                 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1985                        __func__, dm_bufio_allocated_get_free_pages);
1986                 bug = 1;
1987         }
1988
1989         if (dm_bufio_allocated_vmalloc) {
1990                 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1991                        __func__, dm_bufio_allocated_vmalloc);
1992                 bug = 1;
1993         }
1994
1995         BUG_ON(bug);
1996 }
1997
1998 module_init(dm_bufio_init)
1999 module_exit(dm_bufio_exit)
2000
2001 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
2002 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2003
2004 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
2005 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2006
2007 module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
2008 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2009
2010 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
2011 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2012
2013 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
2014 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2015
2016 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
2017 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2018
2019 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
2020 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2021
2022 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
2023 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2024
2025 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2026 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2027 MODULE_LICENSE("GPL");