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

Archive Download this file



interactive