4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
7 Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>.
8 Copyright (C) 2003-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10 drbd is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation; either version 2, or (at your option)
15 drbd is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 GNU General Public License for more details.
20 You should have received a copy of the GNU General Public License
21 along with drbd; see the file COPYING. If not, write to
22 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
26 #include <linux/slab.h>
27 #include <linux/crc32c.h>
28 #include <linux/drbd.h>
29 #include <linux/drbd_limits.h>
30 #include <linux/dynamic_debug.h>
32 #include "drbd_wrappers.h"
35 enum al_transaction_types
{
37 AL_TR_INITIALIZED
= 0xffff
39 /* all fields on disc in big endian */
40 struct __packed al_transaction_on_disk
{
41 /* don't we all like magic */
44 /* to identify the most recent transaction block
45 * in the on disk ring buffer */
48 /* checksum on the full 4k block, with this field set to 0. */
51 /* type of transaction, special transaction types like:
52 * purge-all, set-all-idle, set-all-active, ... to-be-defined
53 * see also enum al_transaction_types */
54 __be16 transaction_type
;
56 /* we currently allow only a few thousand extents,
57 * so 16bit will be enough for the slot number. */
59 /* how many updates in this transaction */
62 /* maximum slot number, "al-extents" in drbd.conf speak.
63 * Having this in each transaction should make reconfiguration
64 * of that parameter easier. */
67 /* slot number the context starts with */
68 __be16 context_start_slot_nr
;
70 /* Some reserved bytes. Expected usage is a 64bit counter of
71 * sectors-written since device creation, and other data generation tag
75 /* --- 36 byte used --- */
77 /* Reserve space for up to AL_UPDATES_PER_TRANSACTION changes
78 * in one transaction, then use the remaining byte in the 4k block for
79 * context information. "Flexible" number of updates per transaction
80 * does not help, as we have to account for the case when all update
81 * slots are used anyways, so it would only complicate code without
84 __be16 update_slot_nr
[AL_UPDATES_PER_TRANSACTION
];
86 /* but the extent number is 32bit, which at an extent size of 4 MiB
87 * allows to cover device sizes of up to 2**54 Byte (16 PiB) */
88 __be32 update_extent_nr
[AL_UPDATES_PER_TRANSACTION
];
90 /* --- 420 bytes used (36 + 64*6) --- */
92 /* 4096 - 420 = 3676 = 919 * 4 */
93 __be32 context
[AL_CONTEXT_PER_TRANSACTION
];
96 struct update_odbm_work
{
101 struct update_al_work
{
103 struct completion event
;
108 void *drbd_md_get_buffer(struct drbd_device
*device
)
112 wait_event(device
->misc_wait
,
113 (r
= atomic_cmpxchg(&device
->md_io_in_use
, 0, 1)) == 0 ||
114 device
->state
.disk
<= D_FAILED
);
116 return r
? NULL
: page_address(device
->md_io_page
);
119 void drbd_md_put_buffer(struct drbd_device
*device
)
121 if (atomic_dec_and_test(&device
->md_io_in_use
))
122 wake_up(&device
->misc_wait
);
125 void wait_until_done_or_force_detached(struct drbd_device
*device
, struct drbd_backing_dev
*bdev
,
131 dt
= rcu_dereference(bdev
->disk_conf
)->disk_timeout
;
135 dt
= MAX_SCHEDULE_TIMEOUT
;
137 dt
= wait_event_timeout(device
->misc_wait
,
138 *done
|| test_bit(FORCE_DETACH
, &device
->flags
), dt
);
140 drbd_err(device
, "meta-data IO operation timed out\n");
141 drbd_chk_io_error(device
, 1, DRBD_FORCE_DETACH
);
145 static int _drbd_md_sync_page_io(struct drbd_device
*device
,
146 struct drbd_backing_dev
*bdev
,
147 struct page
*page
, sector_t sector
,
153 device
->md_io
.done
= 0;
154 device
->md_io
.error
= -ENODEV
;
156 if ((rw
& WRITE
) && !test_bit(MD_NO_FUA
, &device
->flags
))
157 rw
|= REQ_FUA
| REQ_FLUSH
;
160 bio
= bio_alloc_drbd(GFP_NOIO
);
161 bio
->bi_bdev
= bdev
->md_bdev
;
162 bio
->bi_iter
.bi_sector
= sector
;
164 if (bio_add_page(bio
, page
, size
, 0) != size
)
166 bio
->bi_private
= &device
->md_io
;
167 bio
->bi_end_io
= drbd_md_io_complete
;
170 if (!(rw
& WRITE
) && device
->state
.disk
== D_DISKLESS
&& device
->ldev
== NULL
)
171 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
173 else if (!get_ldev_if_state(device
, D_ATTACHING
)) {
174 /* Corresponding put_ldev in drbd_md_io_complete() */
175 drbd_err(device
, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
180 bio_get(bio
); /* one bio_put() is in the completion handler */
181 atomic_inc(&device
->md_io_in_use
); /* drbd_md_put_buffer() is in the completion handler */
182 if (drbd_insert_fault(device
, (rw
& WRITE
) ? DRBD_FAULT_MD_WR
: DRBD_FAULT_MD_RD
))
183 bio_endio(bio
, -EIO
);
186 wait_until_done_or_force_detached(device
, bdev
, &device
->md_io
.done
);
187 if (bio_flagged(bio
, BIO_UPTODATE
))
188 err
= device
->md_io
.error
;
195 int drbd_md_sync_page_io(struct drbd_device
*device
, struct drbd_backing_dev
*bdev
,
196 sector_t sector
, int rw
)
199 struct page
*iop
= device
->md_io_page
;
201 D_ASSERT(device
, atomic_read(&device
->md_io_in_use
) == 1);
203 BUG_ON(!bdev
->md_bdev
);
205 drbd_dbg(device
, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
206 current
->comm
, current
->pid
, __func__
,
207 (unsigned long long)sector
, (rw
& WRITE
) ? "WRITE" : "READ",
210 if (sector
< drbd_md_first_sector(bdev
) ||
211 sector
+ 7 > drbd_md_last_sector(bdev
))
212 drbd_alert(device
, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
213 current
->comm
, current
->pid
, __func__
,
214 (unsigned long long)sector
, (rw
& WRITE
) ? "WRITE" : "READ");
216 /* we do all our meta data IO in aligned 4k blocks. */
217 err
= _drbd_md_sync_page_io(device
, bdev
, iop
, sector
, rw
, 4096);
219 drbd_err(device
, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
220 (unsigned long long)sector
, (rw
& WRITE
) ? "WRITE" : "READ", err
);
225 static struct bm_extent
*find_active_resync_extent(struct drbd_device
*device
, unsigned int enr
)
227 struct lc_element
*tmp
;
228 tmp
= lc_find(device
->resync
, enr
/AL_EXT_PER_BM_SECT
);
229 if (unlikely(tmp
!= NULL
)) {
230 struct bm_extent
*bm_ext
= lc_entry(tmp
, struct bm_extent
, lce
);
231 if (test_bit(BME_NO_WRITES
, &bm_ext
->flags
))
237 static struct lc_element
*_al_get(struct drbd_device
*device
, unsigned int enr
, bool nonblock
)
239 struct lc_element
*al_ext
;
240 struct bm_extent
*bm_ext
;
243 spin_lock_irq(&device
->al_lock
);
244 bm_ext
= find_active_resync_extent(device
, enr
);
246 wake
= !test_and_set_bit(BME_PRIORITY
, &bm_ext
->flags
);
247 spin_unlock_irq(&device
->al_lock
);
249 wake_up(&device
->al_wait
);
253 al_ext
= lc_try_get(device
->act_log
, enr
);
255 al_ext
= lc_get(device
->act_log
, enr
);
256 spin_unlock_irq(&device
->al_lock
);
260 bool drbd_al_begin_io_fastpath(struct drbd_device
*device
, struct drbd_interval
*i
)
262 /* for bios crossing activity log extent boundaries,
263 * we may need to activate two extents in one go */
264 unsigned first
= i
->sector
>> (AL_EXTENT_SHIFT
-9);
265 unsigned last
= i
->size
== 0 ? first
: (i
->sector
+ (i
->size
>> 9) - 1) >> (AL_EXTENT_SHIFT
-9);
267 D_ASSERT(device
, (unsigned)(last
- first
) <= 1);
268 D_ASSERT(device
, atomic_read(&device
->local_cnt
) > 0);
270 /* FIXME figure out a fast path for bios crossing AL extent boundaries */
274 return _al_get(device
, first
, true);
278 bool drbd_al_begin_io_prepare(struct drbd_device
*device
, struct drbd_interval
*i
)
280 /* for bios crossing activity log extent boundaries,
281 * we may need to activate two extents in one go */
282 unsigned first
= i
->sector
>> (AL_EXTENT_SHIFT
-9);
283 unsigned last
= i
->size
== 0 ? first
: (i
->sector
+ (i
->size
>> 9) - 1) >> (AL_EXTENT_SHIFT
-9);
285 bool need_transaction
= false;
287 D_ASSERT(device
, first
<= last
);
288 D_ASSERT(device
, atomic_read(&device
->local_cnt
) > 0);
290 for (enr
= first
; enr
<= last
; enr
++) {
291 struct lc_element
*al_ext
;
292 wait_event(device
->al_wait
,
293 (al_ext
= _al_get(device
, enr
, false)) != NULL
);
294 if (al_ext
->lc_number
!= enr
)
295 need_transaction
= true;
297 return need_transaction
;
300 static int al_write_transaction(struct drbd_device
*device
, bool delegate
);
302 /* When called through generic_make_request(), we must delegate
303 * activity log I/O to the worker thread: a further request
304 * submitted via generic_make_request() within the same task
305 * would be queued on current->bio_list, and would only start
306 * after this function returns (see generic_make_request()).
308 * However, if we *are* the worker, we must not delegate to ourselves.
312 * @delegate: delegate activity log I/O to the worker thread
314 void drbd_al_begin_io_commit(struct drbd_device
*device
, bool delegate
)
318 BUG_ON(delegate
&& current
== first_peer_device(device
)->connection
->worker
.task
);
320 /* Serialize multiple transactions.
321 * This uses test_and_set_bit, memory barrier is implicit.
323 wait_event(device
->al_wait
,
324 device
->act_log
->pending_changes
== 0 ||
325 (locked
= lc_try_lock_for_transaction(device
->act_log
)));
328 /* Double check: it may have been committed by someone else,
329 * while we have been waiting for the lock. */
330 if (device
->act_log
->pending_changes
) {
331 bool write_al_updates
;
334 write_al_updates
= rcu_dereference(device
->ldev
->disk_conf
)->al_updates
;
337 if (write_al_updates
)
338 al_write_transaction(device
, delegate
);
339 spin_lock_irq(&device
->al_lock
);
342 we need an "lc_cancel" here;
344 lc_committed(device
->act_log
);
345 spin_unlock_irq(&device
->al_lock
);
347 lc_unlock(device
->act_log
);
348 wake_up(&device
->al_wait
);
353 * @delegate: delegate activity log I/O to the worker thread
355 void drbd_al_begin_io(struct drbd_device
*device
, struct drbd_interval
*i
, bool delegate
)
357 BUG_ON(delegate
&& current
== first_peer_device(device
)->connection
->worker
.task
);
359 if (drbd_al_begin_io_prepare(device
, i
))
360 drbd_al_begin_io_commit(device
, delegate
);
363 int drbd_al_begin_io_nonblock(struct drbd_device
*device
, struct drbd_interval
*i
)
365 struct lru_cache
*al
= device
->act_log
;
366 /* for bios crossing activity log extent boundaries,
367 * we may need to activate two extents in one go */
368 unsigned first
= i
->sector
>> (AL_EXTENT_SHIFT
-9);
369 unsigned last
= i
->size
== 0 ? first
: (i
->sector
+ (i
->size
>> 9) - 1) >> (AL_EXTENT_SHIFT
-9);
370 unsigned nr_al_extents
;
371 unsigned available_update_slots
;
374 D_ASSERT(device
, first
<= last
);
376 nr_al_extents
= 1 + last
- first
; /* worst case: all touched extends are cold. */
377 available_update_slots
= min(al
->nr_elements
- al
->used
,
378 al
->max_pending_changes
- al
->pending_changes
);
380 /* We want all necessary updates for a given request within the same transaction
381 * We could first check how many updates are *actually* needed,
382 * and use that instead of the worst-case nr_al_extents */
383 if (available_update_slots
< nr_al_extents
)
386 /* Is resync active in this area? */
387 for (enr
= first
; enr
<= last
; enr
++) {
388 struct lc_element
*tmp
;
389 tmp
= lc_find(device
->resync
, enr
/AL_EXT_PER_BM_SECT
);
390 if (unlikely(tmp
!= NULL
)) {
391 struct bm_extent
*bm_ext
= lc_entry(tmp
, struct bm_extent
, lce
);
392 if (test_bit(BME_NO_WRITES
, &bm_ext
->flags
)) {
393 if (!test_and_set_bit(BME_PRIORITY
, &bm_ext
->flags
))
400 /* Checkout the refcounts.
401 * Given that we checked for available elements and update slots above,
402 * this has to be successful. */
403 for (enr
= first
; enr
<= last
; enr
++) {
404 struct lc_element
*al_ext
;
405 al_ext
= lc_get_cumulative(device
->act_log
, enr
);
407 drbd_info(device
, "LOGIC BUG for enr=%u\n", enr
);
412 void drbd_al_complete_io(struct drbd_device
*device
, struct drbd_interval
*i
)
414 /* for bios crossing activity log extent boundaries,
415 * we may need to activate two extents in one go */
416 unsigned first
= i
->sector
>> (AL_EXTENT_SHIFT
-9);
417 unsigned last
= i
->size
== 0 ? first
: (i
->sector
+ (i
->size
>> 9) - 1) >> (AL_EXTENT_SHIFT
-9);
419 struct lc_element
*extent
;
422 D_ASSERT(device
, first
<= last
);
423 spin_lock_irqsave(&device
->al_lock
, flags
);
425 for (enr
= first
; enr
<= last
; enr
++) {
426 extent
= lc_find(device
->act_log
, enr
);
428 drbd_err(device
, "al_complete_io() called on inactive extent %u\n", enr
);
431 lc_put(device
->act_log
, extent
);
433 spin_unlock_irqrestore(&device
->al_lock
, flags
);
434 wake_up(&device
->al_wait
);
437 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
438 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
439 * are still coupled, or assume too much about their relation.
440 * Code below will not work if this is violated.
441 * Will be cleaned up with some followup patch.
446 static unsigned int al_extent_to_bm_page(unsigned int al_enr
)
451 /* al extent number to bit */
452 (AL_EXTENT_SHIFT
- BM_BLOCK_SHIFT
));
455 static unsigned int rs_extent_to_bm_page(unsigned int rs_enr
)
460 /* resync extent number to bit */
461 (BM_EXT_SHIFT
- BM_BLOCK_SHIFT
));
464 static sector_t
al_tr_number_to_on_disk_sector(struct drbd_device
*device
)
466 const unsigned int stripes
= device
->ldev
->md
.al_stripes
;
467 const unsigned int stripe_size_4kB
= device
->ldev
->md
.al_stripe_size_4k
;
469 /* transaction number, modulo on-disk ring buffer wrap around */
470 unsigned int t
= device
->al_tr_number
% (device
->ldev
->md
.al_size_4k
);
472 /* ... to aligned 4k on disk block */
473 t
= ((t
% stripes
) * stripe_size_4kB
) + t
/stripes
;
475 /* ... to 512 byte sector in activity log */
478 /* ... plus offset to the on disk position */
479 return device
->ldev
->md
.md_offset
+ device
->ldev
->md
.al_offset
+ t
;
483 _al_write_transaction(struct drbd_device
*device
)
485 struct al_transaction_on_disk
*buffer
;
486 struct lc_element
*e
;
493 if (!get_ldev(device
)) {
494 drbd_err(device
, "disk is %s, cannot start al transaction\n",
495 drbd_disk_str(device
->state
.disk
));
499 /* The bitmap write may have failed, causing a state change. */
500 if (device
->state
.disk
< D_INCONSISTENT
) {
502 "disk is %s, cannot write al transaction\n",
503 drbd_disk_str(device
->state
.disk
));
508 buffer
= drbd_md_get_buffer(device
); /* protects md_io_buffer, al_tr_cycle, ... */
510 drbd_err(device
, "disk failed while waiting for md_io buffer\n");
515 memset(buffer
, 0, sizeof(*buffer
));
516 buffer
->magic
= cpu_to_be32(DRBD_AL_MAGIC
);
517 buffer
->tr_number
= cpu_to_be32(device
->al_tr_number
);
521 /* Even though no one can start to change this list
522 * once we set the LC_LOCKED -- from drbd_al_begin_io(),
523 * lc_try_lock_for_transaction() --, someone may still
524 * be in the process of changing it. */
525 spin_lock_irq(&device
->al_lock
);
526 list_for_each_entry(e
, &device
->act_log
->to_be_changed
, list
) {
527 if (i
== AL_UPDATES_PER_TRANSACTION
) {
531 buffer
->update_slot_nr
[i
] = cpu_to_be16(e
->lc_index
);
532 buffer
->update_extent_nr
[i
] = cpu_to_be32(e
->lc_new_number
);
533 if (e
->lc_number
!= LC_FREE
)
534 drbd_bm_mark_for_writeout(device
,
535 al_extent_to_bm_page(e
->lc_number
));
538 spin_unlock_irq(&device
->al_lock
);
539 BUG_ON(i
> AL_UPDATES_PER_TRANSACTION
);
541 buffer
->n_updates
= cpu_to_be16(i
);
542 for ( ; i
< AL_UPDATES_PER_TRANSACTION
; i
++) {
543 buffer
->update_slot_nr
[i
] = cpu_to_be16(-1);
544 buffer
->update_extent_nr
[i
] = cpu_to_be32(LC_FREE
);
547 buffer
->context_size
= cpu_to_be16(device
->act_log
->nr_elements
);
548 buffer
->context_start_slot_nr
= cpu_to_be16(device
->al_tr_cycle
);
550 mx
= min_t(int, AL_CONTEXT_PER_TRANSACTION
,
551 device
->act_log
->nr_elements
- device
->al_tr_cycle
);
552 for (i
= 0; i
< mx
; i
++) {
553 unsigned idx
= device
->al_tr_cycle
+ i
;
554 extent_nr
= lc_element_by_index(device
->act_log
, idx
)->lc_number
;
555 buffer
->context
[i
] = cpu_to_be32(extent_nr
);
557 for (; i
< AL_CONTEXT_PER_TRANSACTION
; i
++)
558 buffer
->context
[i
] = cpu_to_be32(LC_FREE
);
560 device
->al_tr_cycle
+= AL_CONTEXT_PER_TRANSACTION
;
561 if (device
->al_tr_cycle
>= device
->act_log
->nr_elements
)
562 device
->al_tr_cycle
= 0;
564 sector
= al_tr_number_to_on_disk_sector(device
);
566 crc
= crc32c(0, buffer
, 4096);
567 buffer
->crc32c
= cpu_to_be32(crc
);
569 if (drbd_bm_write_hinted(device
))
572 bool write_al_updates
;
574 write_al_updates
= rcu_dereference(device
->ldev
->disk_conf
)->al_updates
;
576 if (write_al_updates
) {
577 if (drbd_md_sync_page_io(device
, device
->ldev
, sector
, WRITE
)) {
579 drbd_chk_io_error(device
, 1, DRBD_META_IO_ERROR
);
581 device
->al_tr_number
++;
582 device
->al_writ_cnt
++;
587 drbd_md_put_buffer(device
);
594 static int w_al_write_transaction(struct drbd_work
*w
, int unused
)
596 struct update_al_work
*aw
= container_of(w
, struct update_al_work
, w
);
597 struct drbd_device
*device
= w
->device
;
600 err
= _al_write_transaction(device
);
602 complete(&aw
->event
);
604 return err
!= -EIO
? err
: 0;
607 /* Calls from worker context (see w_restart_disk_io()) need to write the
608 transaction directly. Others came through generic_make_request(),
609 those need to delegate it to the worker. */
610 static int al_write_transaction(struct drbd_device
*device
, bool delegate
)
613 struct update_al_work al_work
;
614 init_completion(&al_work
.event
);
615 al_work
.w
.cb
= w_al_write_transaction
;
616 al_work
.w
.device
= device
;
617 drbd_queue_work_front(&first_peer_device(device
)->connection
->sender_work
, &al_work
.w
);
618 wait_for_completion(&al_work
.event
);
621 return _al_write_transaction(device
);
624 static int _try_lc_del(struct drbd_device
*device
, struct lc_element
*al_ext
)
628 spin_lock_irq(&device
->al_lock
);
629 rv
= (al_ext
->refcnt
== 0);
631 lc_del(device
->act_log
, al_ext
);
632 spin_unlock_irq(&device
->al_lock
);
638 * drbd_al_shrink() - Removes all active extents form the activity log
639 * @device: DRBD device.
641 * Removes all active extents form the activity log, waiting until
642 * the reference count of each entry dropped to 0 first, of course.
644 * You need to lock device->act_log with lc_try_lock() / lc_unlock()
646 void drbd_al_shrink(struct drbd_device
*device
)
648 struct lc_element
*al_ext
;
651 D_ASSERT(device
, test_bit(__LC_LOCKED
, &device
->act_log
->flags
));
653 for (i
= 0; i
< device
->act_log
->nr_elements
; i
++) {
654 al_ext
= lc_element_by_index(device
->act_log
, i
);
655 if (al_ext
->lc_number
== LC_FREE
)
657 wait_event(device
->al_wait
, _try_lc_del(device
, al_ext
));
660 wake_up(&device
->al_wait
);
663 int drbd_initialize_al(struct drbd_device
*device
, void *buffer
)
665 struct al_transaction_on_disk
*al
= buffer
;
666 struct drbd_md
*md
= &device
->ldev
->md
;
667 sector_t al_base
= md
->md_offset
+ md
->al_offset
;
668 int al_size_4k
= md
->al_stripes
* md
->al_stripe_size_4k
;
672 al
->magic
= cpu_to_be32(DRBD_AL_MAGIC
);
673 al
->transaction_type
= cpu_to_be16(AL_TR_INITIALIZED
);
674 al
->crc32c
= cpu_to_be32(crc32c(0, al
, 4096));
676 for (i
= 0; i
< al_size_4k
; i
++) {
677 int err
= drbd_md_sync_page_io(device
, device
->ldev
, al_base
+ i
* 8, WRITE
);
684 static int w_update_odbm(struct drbd_work
*w
, int unused
)
686 struct update_odbm_work
*udw
= container_of(w
, struct update_odbm_work
, w
);
687 struct drbd_device
*device
= w
->device
;
688 struct sib_info sib
= { .sib_reason
= SIB_SYNC_PROGRESS
, };
690 if (!get_ldev(device
)) {
691 if (__ratelimit(&drbd_ratelimit_state
))
692 drbd_warn(device
, "Can not update on disk bitmap, local IO disabled.\n");
697 drbd_bm_write_page(device
, rs_extent_to_bm_page(udw
->enr
));
702 if (drbd_bm_total_weight(device
) <= device
->rs_failed
) {
703 switch (device
->state
.conn
) {
704 case C_SYNC_SOURCE
: case C_SYNC_TARGET
:
705 case C_PAUSED_SYNC_S
: case C_PAUSED_SYNC_T
:
706 drbd_resync_finished(device
);
712 drbd_bcast_event(device
, &sib
);
718 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
719 * resync LRU-cache are 16MB each.
720 * The caller of this function has to hold an get_ldev() reference.
722 * TODO will be obsoleted once we have a caching lru of the on disk bitmap
724 static void drbd_try_clear_on_disk_bm(struct drbd_device
*device
, sector_t sector
,
725 int count
, int success
)
727 struct lc_element
*e
;
728 struct update_odbm_work
*udw
;
732 D_ASSERT(device
, atomic_read(&device
->local_cnt
));
734 /* I simply assume that a sector/size pair never crosses
735 * a 16 MB extent border. (Currently this is true...) */
736 enr
= BM_SECT_TO_EXT(sector
);
738 e
= lc_get(device
->resync
, enr
);
740 struct bm_extent
*ext
= lc_entry(e
, struct bm_extent
, lce
);
741 if (ext
->lce
.lc_number
== enr
) {
743 ext
->rs_left
-= count
;
745 ext
->rs_failed
+= count
;
746 if (ext
->rs_left
< ext
->rs_failed
) {
747 drbd_warn(device
, "BAD! sector=%llus enr=%u rs_left=%d "
748 "rs_failed=%d count=%d cstate=%s\n",
749 (unsigned long long)sector
,
750 ext
->lce
.lc_number
, ext
->rs_left
,
751 ext
->rs_failed
, count
,
752 drbd_conn_str(device
->state
.conn
));
754 /* We don't expect to be able to clear more bits
755 * than have been set when we originally counted
756 * the set bits to cache that value in ext->rs_left.
757 * Whatever the reason (disconnect during resync,
758 * delayed local completion of an application write),
759 * try to fix it up by recounting here. */
760 ext
->rs_left
= drbd_bm_e_weight(device
, enr
);
763 /* Normally this element should be in the cache,
764 * since drbd_rs_begin_io() pulled it already in.
766 * But maybe an application write finished, and we set
767 * something outside the resync lru_cache in sync.
769 int rs_left
= drbd_bm_e_weight(device
, enr
);
770 if (ext
->flags
!= 0) {
771 drbd_warn(device
, "changing resync lce: %d[%u;%02lx]"
773 ext
->lce
.lc_number
, ext
->rs_left
,
774 ext
->flags
, enr
, rs_left
);
777 if (ext
->rs_failed
) {
778 drbd_warn(device
, "Kicking resync_lru element enr=%u "
779 "out with rs_failed=%d\n",
780 ext
->lce
.lc_number
, ext
->rs_failed
);
782 ext
->rs_left
= rs_left
;
783 ext
->rs_failed
= success
? 0 : count
;
784 /* we don't keep a persistent log of the resync lru,
785 * we can commit any change right away. */
786 lc_committed(device
->resync
);
788 lc_put(device
->resync
, &ext
->lce
);
789 /* no race, we are within the al_lock! */
791 if (ext
->rs_left
== ext
->rs_failed
) {
794 udw
= kmalloc(sizeof(*udw
), GFP_ATOMIC
);
796 udw
->enr
= ext
->lce
.lc_number
;
797 udw
->w
.cb
= w_update_odbm
;
798 udw
->w
.device
= device
;
799 drbd_queue_work_front(&first_peer_device(device
)->connection
->sender_work
, &udw
->w
);
801 drbd_warn(device
, "Could not kmalloc an udw\n");
805 drbd_err(device
, "lc_get() failed! locked=%d/%d flags=%lu\n",
806 device
->resync_locked
,
807 device
->resync
->nr_elements
,
808 device
->resync
->flags
);
812 void drbd_advance_rs_marks(struct drbd_device
*device
, unsigned long still_to_go
)
814 unsigned long now
= jiffies
;
815 unsigned long last
= device
->rs_mark_time
[device
->rs_last_mark
];
816 int next
= (device
->rs_last_mark
+ 1) % DRBD_SYNC_MARKS
;
817 if (time_after_eq(now
, last
+ DRBD_SYNC_MARK_STEP
)) {
818 if (device
->rs_mark_left
[device
->rs_last_mark
] != still_to_go
&&
819 device
->state
.conn
!= C_PAUSED_SYNC_T
&&
820 device
->state
.conn
!= C_PAUSED_SYNC_S
) {
821 device
->rs_mark_time
[next
] = now
;
822 device
->rs_mark_left
[next
] = still_to_go
;
823 device
->rs_last_mark
= next
;
828 /* clear the bit corresponding to the piece of storage in question:
829 * size byte of data starting from sector. Only clear a bits of the affected
830 * one ore more _aligned_ BM_BLOCK_SIZE blocks.
832 * called by worker on C_SYNC_TARGET and receiver on SyncSource.
835 void __drbd_set_in_sync(struct drbd_device
*device
, sector_t sector
, int size
,
836 const char *file
, const unsigned int line
)
838 /* Is called from worker and receiver context _only_ */
839 unsigned long sbnr
, ebnr
, lbnr
;
840 unsigned long count
= 0;
841 sector_t esector
, nr_sectors
;
845 if (size
<= 0 || !IS_ALIGNED(size
, 512) || size
> DRBD_MAX_BIO_SIZE
) {
846 drbd_err(device
, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n",
847 (unsigned long long)sector
, size
);
851 if (!get_ldev(device
))
852 return; /* no disk, no metadata, no bitmap to clear bits in */
854 nr_sectors
= drbd_get_capacity(device
->this_bdev
);
855 esector
= sector
+ (size
>> 9) - 1;
857 if (!expect(sector
< nr_sectors
))
859 if (!expect(esector
< nr_sectors
))
860 esector
= nr_sectors
- 1;
862 lbnr
= BM_SECT_TO_BIT(nr_sectors
-1);
864 /* we clear it (in sync).
865 * round up start sector, round down end sector. we make sure we only
866 * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */
867 if (unlikely(esector
< BM_SECT_PER_BIT
-1))
869 if (unlikely(esector
== (nr_sectors
-1)))
872 ebnr
= BM_SECT_TO_BIT(esector
- (BM_SECT_PER_BIT
-1));
873 sbnr
= BM_SECT_TO_BIT(sector
+ BM_SECT_PER_BIT
-1);
879 * ok, (capacity & 7) != 0 sometimes, but who cares...
880 * we count rs_{total,left} in bits, not sectors.
882 count
= drbd_bm_clear_bits(device
, sbnr
, ebnr
);
884 drbd_advance_rs_marks(device
, drbd_bm_total_weight(device
));
885 spin_lock_irqsave(&device
->al_lock
, flags
);
886 drbd_try_clear_on_disk_bm(device
, sector
, count
, true);
887 spin_unlock_irqrestore(&device
->al_lock
, flags
);
889 /* just wake_up unconditional now, various lc_chaged(),
890 * lc_put() in drbd_try_clear_on_disk_bm(). */
896 wake_up(&device
->al_wait
);
900 * this is intended to set one request worth of data out of sync.
901 * affects at least 1 bit,
902 * and at most 1+DRBD_MAX_BIO_SIZE/BM_BLOCK_SIZE bits.
904 * called by tl_clear and drbd_send_dblock (==drbd_make_request).
905 * so this can be _any_ process.
907 int __drbd_set_out_of_sync(struct drbd_device
*device
, sector_t sector
, int size
,
908 const char *file
, const unsigned int line
)
910 unsigned long sbnr
, ebnr
, flags
;
911 sector_t esector
, nr_sectors
;
912 unsigned int enr
, count
= 0;
913 struct lc_element
*e
;
915 /* this should be an empty REQ_FLUSH */
919 if (size
< 0 || !IS_ALIGNED(size
, 512) || size
> DRBD_MAX_BIO_SIZE
) {
920 drbd_err(device
, "sector: %llus, size: %d\n",
921 (unsigned long long)sector
, size
);
925 if (!get_ldev(device
))
926 return 0; /* no disk, no metadata, no bitmap to set bits in */
928 nr_sectors
= drbd_get_capacity(device
->this_bdev
);
929 esector
= sector
+ (size
>> 9) - 1;
931 if (!expect(sector
< nr_sectors
))
933 if (!expect(esector
< nr_sectors
))
934 esector
= nr_sectors
- 1;
936 /* we set it out of sync,
937 * we do not need to round anything here */
938 sbnr
= BM_SECT_TO_BIT(sector
);
939 ebnr
= BM_SECT_TO_BIT(esector
);
941 /* ok, (capacity & 7) != 0 sometimes, but who cares...
942 * we count rs_{total,left} in bits, not sectors. */
943 spin_lock_irqsave(&device
->al_lock
, flags
);
944 count
= drbd_bm_set_bits(device
, sbnr
, ebnr
);
946 enr
= BM_SECT_TO_EXT(sector
);
947 e
= lc_find(device
->resync
, enr
);
949 lc_entry(e
, struct bm_extent
, lce
)->rs_left
+= count
;
950 spin_unlock_irqrestore(&device
->al_lock
, flags
);
959 struct bm_extent
*_bme_get(struct drbd_device
*device
, unsigned int enr
)
961 struct lc_element
*e
;
962 struct bm_extent
*bm_ext
;
964 unsigned long rs_flags
;
966 spin_lock_irq(&device
->al_lock
);
967 if (device
->resync_locked
> device
->resync
->nr_elements
/2) {
968 spin_unlock_irq(&device
->al_lock
);
971 e
= lc_get(device
->resync
, enr
);
972 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
974 if (bm_ext
->lce
.lc_number
!= enr
) {
975 bm_ext
->rs_left
= drbd_bm_e_weight(device
, enr
);
976 bm_ext
->rs_failed
= 0;
977 lc_committed(device
->resync
);
980 if (bm_ext
->lce
.refcnt
== 1)
981 device
->resync_locked
++;
982 set_bit(BME_NO_WRITES
, &bm_ext
->flags
);
984 rs_flags
= device
->resync
->flags
;
985 spin_unlock_irq(&device
->al_lock
);
987 wake_up(&device
->al_wait
);
990 if (rs_flags
& LC_STARVING
)
991 drbd_warn(device
, "Have to wait for element"
992 " (resync LRU too small?)\n");
993 BUG_ON(rs_flags
& LC_LOCKED
);
999 static int _is_in_al(struct drbd_device
*device
, unsigned int enr
)
1003 spin_lock_irq(&device
->al_lock
);
1004 rv
= lc_is_used(device
->act_log
, enr
);
1005 spin_unlock_irq(&device
->al_lock
);
1011 * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
1012 * @device: DRBD device.
1013 * @sector: The sector number.
1015 * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
1017 int drbd_rs_begin_io(struct drbd_device
*device
, sector_t sector
)
1019 unsigned int enr
= BM_SECT_TO_EXT(sector
);
1020 struct bm_extent
*bm_ext
;
1022 int sa
= 200; /* Step aside 200 times, then grab the extent and let app-IO wait.
1023 200 times -> 20 seconds. */
1026 sig
= wait_event_interruptible(device
->al_wait
,
1027 (bm_ext
= _bme_get(device
, enr
)));
1031 if (test_bit(BME_LOCKED
, &bm_ext
->flags
))
1034 for (i
= 0; i
< AL_EXT_PER_BM_SECT
; i
++) {
1035 sig
= wait_event_interruptible(device
->al_wait
,
1036 !_is_in_al(device
, enr
* AL_EXT_PER_BM_SECT
+ i
) ||
1037 test_bit(BME_PRIORITY
, &bm_ext
->flags
));
1039 if (sig
|| (test_bit(BME_PRIORITY
, &bm_ext
->flags
) && sa
)) {
1040 spin_lock_irq(&device
->al_lock
);
1041 if (lc_put(device
->resync
, &bm_ext
->lce
) == 0) {
1042 bm_ext
->flags
= 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
1043 device
->resync_locked
--;
1044 wake_up(&device
->al_wait
);
1046 spin_unlock_irq(&device
->al_lock
);
1049 if (schedule_timeout_interruptible(HZ
/10))
1051 if (sa
&& --sa
== 0)
1052 drbd_warn(device
, "drbd_rs_begin_io() stepped aside for 20sec."
1053 "Resync stalled?\n");
1057 set_bit(BME_LOCKED
, &bm_ext
->flags
);
1062 * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1063 * @device: DRBD device.
1064 * @sector: The sector number.
1066 * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1067 * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1068 * if there is still application IO going on in this area.
1070 int drbd_try_rs_begin_io(struct drbd_device
*device
, sector_t sector
)
1072 unsigned int enr
= BM_SECT_TO_EXT(sector
);
1073 const unsigned int al_enr
= enr
*AL_EXT_PER_BM_SECT
;
1074 struct lc_element
*e
;
1075 struct bm_extent
*bm_ext
;
1078 spin_lock_irq(&device
->al_lock
);
1079 if (device
->resync_wenr
!= LC_FREE
&& device
->resync_wenr
!= enr
) {
1080 /* in case you have very heavy scattered io, it may
1081 * stall the syncer undefined if we give up the ref count
1082 * when we try again and requeue.
1084 * if we don't give up the refcount, but the next time
1085 * we are scheduled this extent has been "synced" by new
1086 * application writes, we'd miss the lc_put on the
1087 * extent we keep the refcount on.
1088 * so we remembered which extent we had to try again, and
1089 * if the next requested one is something else, we do
1090 * the lc_put here...
1091 * we also have to wake_up
1093 e
= lc_find(device
->resync
, device
->resync_wenr
);
1094 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1096 D_ASSERT(device
, !test_bit(BME_LOCKED
, &bm_ext
->flags
));
1097 D_ASSERT(device
, test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1098 clear_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1099 device
->resync_wenr
= LC_FREE
;
1100 if (lc_put(device
->resync
, &bm_ext
->lce
) == 0)
1101 device
->resync_locked
--;
1102 wake_up(&device
->al_wait
);
1104 drbd_alert(device
, "LOGIC BUG\n");
1108 e
= lc_try_get(device
->resync
, enr
);
1109 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1111 if (test_bit(BME_LOCKED
, &bm_ext
->flags
))
1113 if (!test_and_set_bit(BME_NO_WRITES
, &bm_ext
->flags
)) {
1114 device
->resync_locked
++;
1116 /* we did set the BME_NO_WRITES,
1117 * but then could not set BME_LOCKED,
1118 * so we tried again.
1119 * drop the extra reference. */
1120 bm_ext
->lce
.refcnt
--;
1121 D_ASSERT(device
, bm_ext
->lce
.refcnt
> 0);
1125 /* do we rather want to try later? */
1126 if (device
->resync_locked
> device
->resync
->nr_elements
-3)
1128 /* Do or do not. There is no try. -- Yoda */
1129 e
= lc_get(device
->resync
, enr
);
1130 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1132 const unsigned long rs_flags
= device
->resync
->flags
;
1133 if (rs_flags
& LC_STARVING
)
1134 drbd_warn(device
, "Have to wait for element"
1135 " (resync LRU too small?)\n");
1136 BUG_ON(rs_flags
& LC_LOCKED
);
1139 if (bm_ext
->lce
.lc_number
!= enr
) {
1140 bm_ext
->rs_left
= drbd_bm_e_weight(device
, enr
);
1141 bm_ext
->rs_failed
= 0;
1142 lc_committed(device
->resync
);
1143 wake_up(&device
->al_wait
);
1144 D_ASSERT(device
, test_bit(BME_LOCKED
, &bm_ext
->flags
) == 0);
1146 set_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1147 D_ASSERT(device
, bm_ext
->lce
.refcnt
== 1);
1148 device
->resync_locked
++;
1152 for (i
= 0; i
< AL_EXT_PER_BM_SECT
; i
++) {
1153 if (lc_is_used(device
->act_log
, al_enr
+i
))
1156 set_bit(BME_LOCKED
, &bm_ext
->flags
);
1158 device
->resync_wenr
= LC_FREE
;
1159 spin_unlock_irq(&device
->al_lock
);
1164 device
->resync_wenr
= enr
;
1165 spin_unlock_irq(&device
->al_lock
);
1169 void drbd_rs_complete_io(struct drbd_device
*device
, sector_t sector
)
1171 unsigned int enr
= BM_SECT_TO_EXT(sector
);
1172 struct lc_element
*e
;
1173 struct bm_extent
*bm_ext
;
1174 unsigned long flags
;
1176 spin_lock_irqsave(&device
->al_lock
, flags
);
1177 e
= lc_find(device
->resync
, enr
);
1178 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1180 spin_unlock_irqrestore(&device
->al_lock
, flags
);
1181 if (__ratelimit(&drbd_ratelimit_state
))
1182 drbd_err(device
, "drbd_rs_complete_io() called, but extent not found\n");
1186 if (bm_ext
->lce
.refcnt
== 0) {
1187 spin_unlock_irqrestore(&device
->al_lock
, flags
);
1188 drbd_err(device
, "drbd_rs_complete_io(,%llu [=%u]) called, "
1189 "but refcnt is 0!?\n",
1190 (unsigned long long)sector
, enr
);
1194 if (lc_put(device
->resync
, &bm_ext
->lce
) == 0) {
1195 bm_ext
->flags
= 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1196 device
->resync_locked
--;
1197 wake_up(&device
->al_wait
);
1200 spin_unlock_irqrestore(&device
->al_lock
, flags
);
1204 * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1205 * @device: DRBD device.
1207 void drbd_rs_cancel_all(struct drbd_device
*device
)
1209 spin_lock_irq(&device
->al_lock
);
1211 if (get_ldev_if_state(device
, D_FAILED
)) { /* Makes sure ->resync is there. */
1212 lc_reset(device
->resync
);
1215 device
->resync_locked
= 0;
1216 device
->resync_wenr
= LC_FREE
;
1217 spin_unlock_irq(&device
->al_lock
);
1218 wake_up(&device
->al_wait
);
1222 * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1223 * @device: DRBD device.
1225 * Returns 0 upon success, -EAGAIN if at least one reference count was
1228 int drbd_rs_del_all(struct drbd_device
*device
)
1230 struct lc_element
*e
;
1231 struct bm_extent
*bm_ext
;
1234 spin_lock_irq(&device
->al_lock
);
1236 if (get_ldev_if_state(device
, D_FAILED
)) {
1237 /* ok, ->resync is there. */
1238 for (i
= 0; i
< device
->resync
->nr_elements
; i
++) {
1239 e
= lc_element_by_index(device
->resync
, i
);
1240 bm_ext
= lc_entry(e
, struct bm_extent
, lce
);
1241 if (bm_ext
->lce
.lc_number
== LC_FREE
)
1243 if (bm_ext
->lce
.lc_number
== device
->resync_wenr
) {
1244 drbd_info(device
, "dropping %u in drbd_rs_del_all, apparently"
1245 " got 'synced' by application io\n",
1246 device
->resync_wenr
);
1247 D_ASSERT(device
, !test_bit(BME_LOCKED
, &bm_ext
->flags
));
1248 D_ASSERT(device
, test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1249 clear_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1250 device
->resync_wenr
= LC_FREE
;
1251 lc_put(device
->resync
, &bm_ext
->lce
);
1253 if (bm_ext
->lce
.refcnt
!= 0) {
1254 drbd_info(device
, "Retrying drbd_rs_del_all() later. "
1255 "refcnt=%d\n", bm_ext
->lce
.refcnt
);
1257 spin_unlock_irq(&device
->al_lock
);
1260 D_ASSERT(device
, !test_bit(BME_LOCKED
, &bm_ext
->flags
));
1261 D_ASSERT(device
, !test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1262 lc_del(device
->resync
, &bm_ext
->lce
);
1264 D_ASSERT(device
, device
->resync
->used
== 0);
1267 spin_unlock_irq(&device
->al_lock
);
1268 wake_up(&device
->al_wait
);
1274 * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks
1275 * @device: DRBD device.
1276 * @sector: The sector number.
1277 * @size: Size of failed IO operation, in byte.
1279 void drbd_rs_failed_io(struct drbd_device
*device
, sector_t sector
, int size
)
1281 /* Is called from worker and receiver context _only_ */
1282 unsigned long sbnr
, ebnr
, lbnr
;
1283 unsigned long count
;
1284 sector_t esector
, nr_sectors
;
1287 if (size
<= 0 || !IS_ALIGNED(size
, 512) || size
> DRBD_MAX_BIO_SIZE
) {
1288 drbd_err(device
, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n",
1289 (unsigned long long)sector
, size
);
1292 nr_sectors
= drbd_get_capacity(device
->this_bdev
);
1293 esector
= sector
+ (size
>> 9) - 1;
1295 if (!expect(sector
< nr_sectors
))
1297 if (!expect(esector
< nr_sectors
))
1298 esector
= nr_sectors
- 1;
1300 lbnr
= BM_SECT_TO_BIT(nr_sectors
-1);
1303 * round up start sector, round down end sector. we make sure we only
1304 * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */
1305 if (unlikely(esector
< BM_SECT_PER_BIT
-1))
1307 if (unlikely(esector
== (nr_sectors
-1)))
1310 ebnr
= BM_SECT_TO_BIT(esector
- (BM_SECT_PER_BIT
-1));
1311 sbnr
= BM_SECT_TO_BIT(sector
+ BM_SECT_PER_BIT
-1);
1317 * ok, (capacity & 7) != 0 sometimes, but who cares...
1318 * we count rs_{total,left} in bits, not sectors.
1320 spin_lock_irq(&device
->al_lock
);
1321 count
= drbd_bm_count_bits(device
, sbnr
, ebnr
);
1323 device
->rs_failed
+= count
;
1325 if (get_ldev(device
)) {
1326 drbd_try_clear_on_disk_bm(device
, sector
, count
, false);
1330 /* just wake_up unconditional now, various lc_chaged(),
1331 * lc_put() in drbd_try_clear_on_disk_bm(). */
1334 spin_unlock_irq(&device
->al_lock
);
1336 wake_up(&device
->al_wait
);