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
{
98 struct drbd_device
*device
;
102 struct update_al_work
{
104 struct drbd_device
*device
;
105 struct completion event
;
110 void *drbd_md_get_buffer(struct drbd_device
*device
)
114 wait_event(device
->misc_wait
,
115 (r
= atomic_cmpxchg(&device
->md_io_in_use
, 0, 1)) == 0 ||
116 device
->state
.disk
<= D_FAILED
);
118 return r
? NULL
: page_address(device
->md_io_page
);
121 void drbd_md_put_buffer(struct drbd_device
*device
)
123 if (atomic_dec_and_test(&device
->md_io_in_use
))
124 wake_up(&device
->misc_wait
);
127 void wait_until_done_or_force_detached(struct drbd_device
*device
, struct drbd_backing_dev
*bdev
,
133 dt
= rcu_dereference(bdev
->disk_conf
)->disk_timeout
;
137 dt
= MAX_SCHEDULE_TIMEOUT
;
139 dt
= wait_event_timeout(device
->misc_wait
,
140 *done
|| test_bit(FORCE_DETACH
, &device
->flags
), dt
);
142 drbd_err(device
, "meta-data IO operation timed out\n");
143 drbd_chk_io_error(device
, 1, DRBD_FORCE_DETACH
);
147 static int _drbd_md_sync_page_io(struct drbd_device
*device
,
148 struct drbd_backing_dev
*bdev
,
149 struct page
*page
, sector_t sector
,
155 device
->md_io
.done
= 0;
156 device
->md_io
.error
= -ENODEV
;
158 if ((rw
& WRITE
) && !test_bit(MD_NO_FUA
, &device
->flags
))
159 rw
|= REQ_FUA
| REQ_FLUSH
;
162 bio
= bio_alloc_drbd(GFP_NOIO
);
163 bio
->bi_bdev
= bdev
->md_bdev
;
164 bio
->bi_iter
.bi_sector
= sector
;
166 if (bio_add_page(bio
, page
, size
, 0) != size
)
168 bio
->bi_private
= &device
->md_io
;
169 bio
->bi_end_io
= drbd_md_io_complete
;
172 if (!(rw
& WRITE
) && device
->state
.disk
== D_DISKLESS
&& device
->ldev
== NULL
)
173 /* special case, drbd_md_read() during drbd_adm_attach(): no get_ldev */
175 else if (!get_ldev_if_state(device
, D_ATTACHING
)) {
176 /* Corresponding put_ldev in drbd_md_io_complete() */
177 drbd_err(device
, "ASSERT FAILED: get_ldev_if_state() == 1 in _drbd_md_sync_page_io()\n");
182 bio_get(bio
); /* one bio_put() is in the completion handler */
183 atomic_inc(&device
->md_io_in_use
); /* drbd_md_put_buffer() is in the completion handler */
184 if (drbd_insert_fault(device
, (rw
& WRITE
) ? DRBD_FAULT_MD_WR
: DRBD_FAULT_MD_RD
))
185 bio_endio(bio
, -EIO
);
188 wait_until_done_or_force_detached(device
, bdev
, &device
->md_io
.done
);
189 if (bio_flagged(bio
, BIO_UPTODATE
))
190 err
= device
->md_io
.error
;
197 int drbd_md_sync_page_io(struct drbd_device
*device
, struct drbd_backing_dev
*bdev
,
198 sector_t sector
, int rw
)
201 struct page
*iop
= device
->md_io_page
;
203 D_ASSERT(device
, atomic_read(&device
->md_io_in_use
) == 1);
205 BUG_ON(!bdev
->md_bdev
);
207 drbd_dbg(device
, "meta_data io: %s [%d]:%s(,%llus,%s) %pS\n",
208 current
->comm
, current
->pid
, __func__
,
209 (unsigned long long)sector
, (rw
& WRITE
) ? "WRITE" : "READ",
212 if (sector
< drbd_md_first_sector(bdev
) ||
213 sector
+ 7 > drbd_md_last_sector(bdev
))
214 drbd_alert(device
, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
215 current
->comm
, current
->pid
, __func__
,
216 (unsigned long long)sector
, (rw
& WRITE
) ? "WRITE" : "READ");
218 /* we do all our meta data IO in aligned 4k blocks. */
219 err
= _drbd_md_sync_page_io(device
, bdev
, iop
, sector
, rw
, 4096);
221 drbd_err(device
, "drbd_md_sync_page_io(,%llus,%s) failed with error %d\n",
222 (unsigned long long)sector
, (rw
& WRITE
) ? "WRITE" : "READ", err
);
227 static struct bm_extent
*find_active_resync_extent(struct drbd_device
*device
, unsigned int enr
)
229 struct lc_element
*tmp
;
230 tmp
= lc_find(device
->resync
, enr
/AL_EXT_PER_BM_SECT
);
231 if (unlikely(tmp
!= NULL
)) {
232 struct bm_extent
*bm_ext
= lc_entry(tmp
, struct bm_extent
, lce
);
233 if (test_bit(BME_NO_WRITES
, &bm_ext
->flags
))
239 static struct lc_element
*_al_get(struct drbd_device
*device
, unsigned int enr
, bool nonblock
)
241 struct lc_element
*al_ext
;
242 struct bm_extent
*bm_ext
;
245 spin_lock_irq(&device
->al_lock
);
246 bm_ext
= find_active_resync_extent(device
, enr
);
248 wake
= !test_and_set_bit(BME_PRIORITY
, &bm_ext
->flags
);
249 spin_unlock_irq(&device
->al_lock
);
251 wake_up(&device
->al_wait
);
255 al_ext
= lc_try_get(device
->act_log
, enr
);
257 al_ext
= lc_get(device
->act_log
, enr
);
258 spin_unlock_irq(&device
->al_lock
);
262 bool drbd_al_begin_io_fastpath(struct drbd_device
*device
, struct drbd_interval
*i
)
264 /* for bios crossing activity log extent boundaries,
265 * we may need to activate two extents in one go */
266 unsigned first
= i
->sector
>> (AL_EXTENT_SHIFT
-9);
267 unsigned last
= i
->size
== 0 ? first
: (i
->sector
+ (i
->size
>> 9) - 1) >> (AL_EXTENT_SHIFT
-9);
269 D_ASSERT(device
, (unsigned)(last
- first
) <= 1);
270 D_ASSERT(device
, atomic_read(&device
->local_cnt
) > 0);
272 /* FIXME figure out a fast path for bios crossing AL extent boundaries */
276 return _al_get(device
, first
, true);
280 bool drbd_al_begin_io_prepare(struct drbd_device
*device
, struct drbd_interval
*i
)
282 /* for bios crossing activity log extent boundaries,
283 * we may need to activate two extents in one go */
284 unsigned first
= i
->sector
>> (AL_EXTENT_SHIFT
-9);
285 unsigned last
= i
->size
== 0 ? first
: (i
->sector
+ (i
->size
>> 9) - 1) >> (AL_EXTENT_SHIFT
-9);
287 bool need_transaction
= false;
289 D_ASSERT(device
, first
<= last
);
290 D_ASSERT(device
, atomic_read(&device
->local_cnt
) > 0);
292 for (enr
= first
; enr
<= last
; enr
++) {
293 struct lc_element
*al_ext
;
294 wait_event(device
->al_wait
,
295 (al_ext
= _al_get(device
, enr
, false)) != NULL
);
296 if (al_ext
->lc_number
!= enr
)
297 need_transaction
= true;
299 return need_transaction
;
302 static int al_write_transaction(struct drbd_device
*device
, bool delegate
);
304 /* When called through generic_make_request(), we must delegate
305 * activity log I/O to the worker thread: a further request
306 * submitted via generic_make_request() within the same task
307 * would be queued on current->bio_list, and would only start
308 * after this function returns (see generic_make_request()).
310 * However, if we *are* the worker, we must not delegate to ourselves.
314 * @delegate: delegate activity log I/O to the worker thread
316 void drbd_al_begin_io_commit(struct drbd_device
*device
, bool delegate
)
320 BUG_ON(delegate
&& current
== first_peer_device(device
)->connection
->worker
.task
);
322 /* Serialize multiple transactions.
323 * This uses test_and_set_bit, memory barrier is implicit.
325 wait_event(device
->al_wait
,
326 device
->act_log
->pending_changes
== 0 ||
327 (locked
= lc_try_lock_for_transaction(device
->act_log
)));
330 /* Double check: it may have been committed by someone else,
331 * while we have been waiting for the lock. */
332 if (device
->act_log
->pending_changes
) {
333 bool write_al_updates
;
336 write_al_updates
= rcu_dereference(device
->ldev
->disk_conf
)->al_updates
;
339 if (write_al_updates
)
340 al_write_transaction(device
, delegate
);
341 spin_lock_irq(&device
->al_lock
);
344 we need an "lc_cancel" here;
346 lc_committed(device
->act_log
);
347 spin_unlock_irq(&device
->al_lock
);
349 lc_unlock(device
->act_log
);
350 wake_up(&device
->al_wait
);
355 * @delegate: delegate activity log I/O to the worker thread
357 void drbd_al_begin_io(struct drbd_device
*device
, struct drbd_interval
*i
, bool delegate
)
359 BUG_ON(delegate
&& current
== first_peer_device(device
)->connection
->worker
.task
);
361 if (drbd_al_begin_io_prepare(device
, i
))
362 drbd_al_begin_io_commit(device
, delegate
);
365 int drbd_al_begin_io_nonblock(struct drbd_device
*device
, struct drbd_interval
*i
)
367 struct lru_cache
*al
= device
->act_log
;
368 /* for bios crossing activity log extent boundaries,
369 * we may need to activate two extents in one go */
370 unsigned first
= i
->sector
>> (AL_EXTENT_SHIFT
-9);
371 unsigned last
= i
->size
== 0 ? first
: (i
->sector
+ (i
->size
>> 9) - 1) >> (AL_EXTENT_SHIFT
-9);
372 unsigned nr_al_extents
;
373 unsigned available_update_slots
;
376 D_ASSERT(device
, first
<= last
);
378 nr_al_extents
= 1 + last
- first
; /* worst case: all touched extends are cold. */
379 available_update_slots
= min(al
->nr_elements
- al
->used
,
380 al
->max_pending_changes
- al
->pending_changes
);
382 /* We want all necessary updates for a given request within the same transaction
383 * We could first check how many updates are *actually* needed,
384 * and use that instead of the worst-case nr_al_extents */
385 if (available_update_slots
< nr_al_extents
)
388 /* Is resync active in this area? */
389 for (enr
= first
; enr
<= last
; enr
++) {
390 struct lc_element
*tmp
;
391 tmp
= lc_find(device
->resync
, enr
/AL_EXT_PER_BM_SECT
);
392 if (unlikely(tmp
!= NULL
)) {
393 struct bm_extent
*bm_ext
= lc_entry(tmp
, struct bm_extent
, lce
);
394 if (test_bit(BME_NO_WRITES
, &bm_ext
->flags
)) {
395 if (!test_and_set_bit(BME_PRIORITY
, &bm_ext
->flags
))
402 /* Checkout the refcounts.
403 * Given that we checked for available elements and update slots above,
404 * this has to be successful. */
405 for (enr
= first
; enr
<= last
; enr
++) {
406 struct lc_element
*al_ext
;
407 al_ext
= lc_get_cumulative(device
->act_log
, enr
);
409 drbd_info(device
, "LOGIC BUG for enr=%u\n", enr
);
414 void drbd_al_complete_io(struct drbd_device
*device
, struct drbd_interval
*i
)
416 /* for bios crossing activity log extent boundaries,
417 * we may need to activate two extents in one go */
418 unsigned first
= i
->sector
>> (AL_EXTENT_SHIFT
-9);
419 unsigned last
= i
->size
== 0 ? first
: (i
->sector
+ (i
->size
>> 9) - 1) >> (AL_EXTENT_SHIFT
-9);
421 struct lc_element
*extent
;
424 D_ASSERT(device
, first
<= last
);
425 spin_lock_irqsave(&device
->al_lock
, flags
);
427 for (enr
= first
; enr
<= last
; enr
++) {
428 extent
= lc_find(device
->act_log
, enr
);
430 drbd_err(device
, "al_complete_io() called on inactive extent %u\n", enr
);
433 lc_put(device
->act_log
, extent
);
435 spin_unlock_irqrestore(&device
->al_lock
, flags
);
436 wake_up(&device
->al_wait
);
439 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
440 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
441 * are still coupled, or assume too much about their relation.
442 * Code below will not work if this is violated.
443 * Will be cleaned up with some followup patch.
448 static unsigned int al_extent_to_bm_page(unsigned int al_enr
)
453 /* al extent number to bit */
454 (AL_EXTENT_SHIFT
- BM_BLOCK_SHIFT
));
457 static unsigned int rs_extent_to_bm_page(unsigned int rs_enr
)
462 /* resync extent number to bit */
463 (BM_EXT_SHIFT
- BM_BLOCK_SHIFT
));
466 static sector_t
al_tr_number_to_on_disk_sector(struct drbd_device
*device
)
468 const unsigned int stripes
= device
->ldev
->md
.al_stripes
;
469 const unsigned int stripe_size_4kB
= device
->ldev
->md
.al_stripe_size_4k
;
471 /* transaction number, modulo on-disk ring buffer wrap around */
472 unsigned int t
= device
->al_tr_number
% (device
->ldev
->md
.al_size_4k
);
474 /* ... to aligned 4k on disk block */
475 t
= ((t
% stripes
) * stripe_size_4kB
) + t
/stripes
;
477 /* ... to 512 byte sector in activity log */
480 /* ... plus offset to the on disk position */
481 return device
->ldev
->md
.md_offset
+ device
->ldev
->md
.al_offset
+ t
;
485 _al_write_transaction(struct drbd_device
*device
)
487 struct al_transaction_on_disk
*buffer
;
488 struct lc_element
*e
;
495 if (!get_ldev(device
)) {
496 drbd_err(device
, "disk is %s, cannot start al transaction\n",
497 drbd_disk_str(device
->state
.disk
));
501 /* The bitmap write may have failed, causing a state change. */
502 if (device
->state
.disk
< D_INCONSISTENT
) {
504 "disk is %s, cannot write al transaction\n",
505 drbd_disk_str(device
->state
.disk
));
510 buffer
= drbd_md_get_buffer(device
); /* protects md_io_buffer, al_tr_cycle, ... */
512 drbd_err(device
, "disk failed while waiting for md_io buffer\n");
517 memset(buffer
, 0, sizeof(*buffer
));
518 buffer
->magic
= cpu_to_be32(DRBD_AL_MAGIC
);
519 buffer
->tr_number
= cpu_to_be32(device
->al_tr_number
);
523 /* Even though no one can start to change this list
524 * once we set the LC_LOCKED -- from drbd_al_begin_io(),
525 * lc_try_lock_for_transaction() --, someone may still
526 * be in the process of changing it. */
527 spin_lock_irq(&device
->al_lock
);
528 list_for_each_entry(e
, &device
->act_log
->to_be_changed
, list
) {
529 if (i
== AL_UPDATES_PER_TRANSACTION
) {
533 buffer
->update_slot_nr
[i
] = cpu_to_be16(e
->lc_index
);
534 buffer
->update_extent_nr
[i
] = cpu_to_be32(e
->lc_new_number
);
535 if (e
->lc_number
!= LC_FREE
)
536 drbd_bm_mark_for_writeout(device
,
537 al_extent_to_bm_page(e
->lc_number
));
540 spin_unlock_irq(&device
->al_lock
);
541 BUG_ON(i
> AL_UPDATES_PER_TRANSACTION
);
543 buffer
->n_updates
= cpu_to_be16(i
);
544 for ( ; i
< AL_UPDATES_PER_TRANSACTION
; i
++) {
545 buffer
->update_slot_nr
[i
] = cpu_to_be16(-1);
546 buffer
->update_extent_nr
[i
] = cpu_to_be32(LC_FREE
);
549 buffer
->context_size
= cpu_to_be16(device
->act_log
->nr_elements
);
550 buffer
->context_start_slot_nr
= cpu_to_be16(device
->al_tr_cycle
);
552 mx
= min_t(int, AL_CONTEXT_PER_TRANSACTION
,
553 device
->act_log
->nr_elements
- device
->al_tr_cycle
);
554 for (i
= 0; i
< mx
; i
++) {
555 unsigned idx
= device
->al_tr_cycle
+ i
;
556 extent_nr
= lc_element_by_index(device
->act_log
, idx
)->lc_number
;
557 buffer
->context
[i
] = cpu_to_be32(extent_nr
);
559 for (; i
< AL_CONTEXT_PER_TRANSACTION
; i
++)
560 buffer
->context
[i
] = cpu_to_be32(LC_FREE
);
562 device
->al_tr_cycle
+= AL_CONTEXT_PER_TRANSACTION
;
563 if (device
->al_tr_cycle
>= device
->act_log
->nr_elements
)
564 device
->al_tr_cycle
= 0;
566 sector
= al_tr_number_to_on_disk_sector(device
);
568 crc
= crc32c(0, buffer
, 4096);
569 buffer
->crc32c
= cpu_to_be32(crc
);
571 if (drbd_bm_write_hinted(device
))
574 bool write_al_updates
;
576 write_al_updates
= rcu_dereference(device
->ldev
->disk_conf
)->al_updates
;
578 if (write_al_updates
) {
579 if (drbd_md_sync_page_io(device
, device
->ldev
, sector
, WRITE
)) {
581 drbd_chk_io_error(device
, 1, DRBD_META_IO_ERROR
);
583 device
->al_tr_number
++;
584 device
->al_writ_cnt
++;
589 drbd_md_put_buffer(device
);
596 static int w_al_write_transaction(struct drbd_work
*w
, int unused
)
598 struct update_al_work
*aw
= container_of(w
, struct update_al_work
, w
);
599 struct drbd_device
*device
= aw
->device
;
602 err
= _al_write_transaction(device
);
604 complete(&aw
->event
);
606 return err
!= -EIO
? err
: 0;
609 /* Calls from worker context (see w_restart_disk_io()) need to write the
610 transaction directly. Others came through generic_make_request(),
611 those need to delegate it to the worker. */
612 static int al_write_transaction(struct drbd_device
*device
, bool delegate
)
615 struct update_al_work al_work
;
616 init_completion(&al_work
.event
);
617 al_work
.w
.cb
= w_al_write_transaction
;
618 al_work
.device
= device
;
619 drbd_queue_work_front(&first_peer_device(device
)->connection
->sender_work
,
621 wait_for_completion(&al_work
.event
);
624 return _al_write_transaction(device
);
627 static int _try_lc_del(struct drbd_device
*device
, struct lc_element
*al_ext
)
631 spin_lock_irq(&device
->al_lock
);
632 rv
= (al_ext
->refcnt
== 0);
634 lc_del(device
->act_log
, al_ext
);
635 spin_unlock_irq(&device
->al_lock
);
641 * drbd_al_shrink() - Removes all active extents form the activity log
642 * @device: DRBD device.
644 * Removes all active extents form the activity log, waiting until
645 * the reference count of each entry dropped to 0 first, of course.
647 * You need to lock device->act_log with lc_try_lock() / lc_unlock()
649 void drbd_al_shrink(struct drbd_device
*device
)
651 struct lc_element
*al_ext
;
654 D_ASSERT(device
, test_bit(__LC_LOCKED
, &device
->act_log
->flags
));
656 for (i
= 0; i
< device
->act_log
->nr_elements
; i
++) {
657 al_ext
= lc_element_by_index(device
->act_log
, i
);
658 if (al_ext
->lc_number
== LC_FREE
)
660 wait_event(device
->al_wait
, _try_lc_del(device
, al_ext
));
663 wake_up(&device
->al_wait
);
666 int drbd_initialize_al(struct drbd_device
*device
, void *buffer
)
668 struct al_transaction_on_disk
*al
= buffer
;
669 struct drbd_md
*md
= &device
->ldev
->md
;
670 sector_t al_base
= md
->md_offset
+ md
->al_offset
;
671 int al_size_4k
= md
->al_stripes
* md
->al_stripe_size_4k
;
675 al
->magic
= cpu_to_be32(DRBD_AL_MAGIC
);
676 al
->transaction_type
= cpu_to_be16(AL_TR_INITIALIZED
);
677 al
->crc32c
= cpu_to_be32(crc32c(0, al
, 4096));
679 for (i
= 0; i
< al_size_4k
; i
++) {
680 int err
= drbd_md_sync_page_io(device
, device
->ldev
, al_base
+ i
* 8, WRITE
);
687 static int w_update_odbm(struct drbd_work
*w
, int unused
)
689 struct update_odbm_work
*udw
= container_of(w
, struct update_odbm_work
, w
);
690 struct drbd_device
*device
= udw
->device
;
691 struct sib_info sib
= { .sib_reason
= SIB_SYNC_PROGRESS
, };
693 if (!get_ldev(device
)) {
694 if (__ratelimit(&drbd_ratelimit_state
))
695 drbd_warn(device
, "Can not update on disk bitmap, local IO disabled.\n");
700 drbd_bm_write_page(device
, rs_extent_to_bm_page(udw
->enr
));
705 if (drbd_bm_total_weight(device
) <= device
->rs_failed
) {
706 switch (device
->state
.conn
) {
707 case C_SYNC_SOURCE
: case C_SYNC_TARGET
:
708 case C_PAUSED_SYNC_S
: case C_PAUSED_SYNC_T
:
709 drbd_resync_finished(device
);
715 drbd_bcast_event(device
, &sib
);
721 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
722 * resync LRU-cache are 16MB each.
723 * The caller of this function has to hold an get_ldev() reference.
725 * TODO will be obsoleted once we have a caching lru of the on disk bitmap
727 static void drbd_try_clear_on_disk_bm(struct drbd_device
*device
, sector_t sector
,
728 int count
, int success
)
730 struct lc_element
*e
;
731 struct update_odbm_work
*udw
;
735 D_ASSERT(device
, atomic_read(&device
->local_cnt
));
737 /* I simply assume that a sector/size pair never crosses
738 * a 16 MB extent border. (Currently this is true...) */
739 enr
= BM_SECT_TO_EXT(sector
);
741 e
= lc_get(device
->resync
, enr
);
743 struct bm_extent
*ext
= lc_entry(e
, struct bm_extent
, lce
);
744 if (ext
->lce
.lc_number
== enr
) {
746 ext
->rs_left
-= count
;
748 ext
->rs_failed
+= count
;
749 if (ext
->rs_left
< ext
->rs_failed
) {
750 drbd_warn(device
, "BAD! sector=%llus enr=%u rs_left=%d "
751 "rs_failed=%d count=%d cstate=%s\n",
752 (unsigned long long)sector
,
753 ext
->lce
.lc_number
, ext
->rs_left
,
754 ext
->rs_failed
, count
,
755 drbd_conn_str(device
->state
.conn
));
757 /* We don't expect to be able to clear more bits
758 * than have been set when we originally counted
759 * the set bits to cache that value in ext->rs_left.
760 * Whatever the reason (disconnect during resync,
761 * delayed local completion of an application write),
762 * try to fix it up by recounting here. */
763 ext
->rs_left
= drbd_bm_e_weight(device
, enr
);
766 /* Normally this element should be in the cache,
767 * since drbd_rs_begin_io() pulled it already in.
769 * But maybe an application write finished, and we set
770 * something outside the resync lru_cache in sync.
772 int rs_left
= drbd_bm_e_weight(device
, enr
);
773 if (ext
->flags
!= 0) {
774 drbd_warn(device
, "changing resync lce: %d[%u;%02lx]"
776 ext
->lce
.lc_number
, ext
->rs_left
,
777 ext
->flags
, enr
, rs_left
);
780 if (ext
->rs_failed
) {
781 drbd_warn(device
, "Kicking resync_lru element enr=%u "
782 "out with rs_failed=%d\n",
783 ext
->lce
.lc_number
, ext
->rs_failed
);
785 ext
->rs_left
= rs_left
;
786 ext
->rs_failed
= success
? 0 : count
;
787 /* we don't keep a persistent log of the resync lru,
788 * we can commit any change right away. */
789 lc_committed(device
->resync
);
791 lc_put(device
->resync
, &ext
->lce
);
792 /* no race, we are within the al_lock! */
794 if (ext
->rs_left
== ext
->rs_failed
) {
797 udw
= kmalloc(sizeof(*udw
), GFP_ATOMIC
);
799 udw
->enr
= ext
->lce
.lc_number
;
800 udw
->w
.cb
= w_update_odbm
;
801 udw
->device
= device
;
802 drbd_queue_work_front(&first_peer_device(device
)->connection
->sender_work
,
805 drbd_warn(device
, "Could not kmalloc an udw\n");
809 drbd_err(device
, "lc_get() failed! locked=%d/%d flags=%lu\n",
810 device
->resync_locked
,
811 device
->resync
->nr_elements
,
812 device
->resync
->flags
);
816 void drbd_advance_rs_marks(struct drbd_device
*device
, unsigned long still_to_go
)
818 unsigned long now
= jiffies
;
819 unsigned long last
= device
->rs_mark_time
[device
->rs_last_mark
];
820 int next
= (device
->rs_last_mark
+ 1) % DRBD_SYNC_MARKS
;
821 if (time_after_eq(now
, last
+ DRBD_SYNC_MARK_STEP
)) {
822 if (device
->rs_mark_left
[device
->rs_last_mark
] != still_to_go
&&
823 device
->state
.conn
!= C_PAUSED_SYNC_T
&&
824 device
->state
.conn
!= C_PAUSED_SYNC_S
) {
825 device
->rs_mark_time
[next
] = now
;
826 device
->rs_mark_left
[next
] = still_to_go
;
827 device
->rs_last_mark
= next
;
832 /* clear the bit corresponding to the piece of storage in question:
833 * size byte of data starting from sector. Only clear a bits of the affected
834 * one ore more _aligned_ BM_BLOCK_SIZE blocks.
836 * called by worker on C_SYNC_TARGET and receiver on SyncSource.
839 void __drbd_set_in_sync(struct drbd_device
*device
, sector_t sector
, int size
,
840 const char *file
, const unsigned int line
)
842 /* Is called from worker and receiver context _only_ */
843 unsigned long sbnr
, ebnr
, lbnr
;
844 unsigned long count
= 0;
845 sector_t esector
, nr_sectors
;
849 if (size
<= 0 || !IS_ALIGNED(size
, 512) || size
> DRBD_MAX_BIO_SIZE
) {
850 drbd_err(device
, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n",
851 (unsigned long long)sector
, size
);
855 if (!get_ldev(device
))
856 return; /* no disk, no metadata, no bitmap to clear bits in */
858 nr_sectors
= drbd_get_capacity(device
->this_bdev
);
859 esector
= sector
+ (size
>> 9) - 1;
861 if (!expect(sector
< nr_sectors
))
863 if (!expect(esector
< nr_sectors
))
864 esector
= nr_sectors
- 1;
866 lbnr
= BM_SECT_TO_BIT(nr_sectors
-1);
868 /* we clear it (in sync).
869 * round up start sector, round down end sector. we make sure we only
870 * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */
871 if (unlikely(esector
< BM_SECT_PER_BIT
-1))
873 if (unlikely(esector
== (nr_sectors
-1)))
876 ebnr
= BM_SECT_TO_BIT(esector
- (BM_SECT_PER_BIT
-1));
877 sbnr
= BM_SECT_TO_BIT(sector
+ BM_SECT_PER_BIT
-1);
883 * ok, (capacity & 7) != 0 sometimes, but who cares...
884 * we count rs_{total,left} in bits, not sectors.
886 count
= drbd_bm_clear_bits(device
, sbnr
, ebnr
);
888 drbd_advance_rs_marks(device
, drbd_bm_total_weight(device
));
889 spin_lock_irqsave(&device
->al_lock
, flags
);
890 drbd_try_clear_on_disk_bm(device
, sector
, count
, true);
891 spin_unlock_irqrestore(&device
->al_lock
, flags
);
893 /* just wake_up unconditional now, various lc_chaged(),
894 * lc_put() in drbd_try_clear_on_disk_bm(). */
900 wake_up(&device
->al_wait
);
904 * this is intended to set one request worth of data out of sync.
905 * affects at least 1 bit,
906 * and at most 1+DRBD_MAX_BIO_SIZE/BM_BLOCK_SIZE bits.
908 * called by tl_clear and drbd_send_dblock (==drbd_make_request).
909 * so this can be _any_ process.
911 int __drbd_set_out_of_sync(struct drbd_device
*device
, sector_t sector
, int size
,
912 const char *file
, const unsigned int line
)
914 unsigned long sbnr
, ebnr
, flags
;
915 sector_t esector
, nr_sectors
;
916 unsigned int enr
, count
= 0;
917 struct lc_element
*e
;
919 /* this should be an empty REQ_FLUSH */
923 if (size
< 0 || !IS_ALIGNED(size
, 512) || size
> DRBD_MAX_BIO_SIZE
) {
924 drbd_err(device
, "sector: %llus, size: %d\n",
925 (unsigned long long)sector
, size
);
929 if (!get_ldev(device
))
930 return 0; /* no disk, no metadata, no bitmap to set bits in */
932 nr_sectors
= drbd_get_capacity(device
->this_bdev
);
933 esector
= sector
+ (size
>> 9) - 1;
935 if (!expect(sector
< nr_sectors
))
937 if (!expect(esector
< nr_sectors
))
938 esector
= nr_sectors
- 1;
940 /* we set it out of sync,
941 * we do not need to round anything here */
942 sbnr
= BM_SECT_TO_BIT(sector
);
943 ebnr
= BM_SECT_TO_BIT(esector
);
945 /* ok, (capacity & 7) != 0 sometimes, but who cares...
946 * we count rs_{total,left} in bits, not sectors. */
947 spin_lock_irqsave(&device
->al_lock
, flags
);
948 count
= drbd_bm_set_bits(device
, sbnr
, ebnr
);
950 enr
= BM_SECT_TO_EXT(sector
);
951 e
= lc_find(device
->resync
, enr
);
953 lc_entry(e
, struct bm_extent
, lce
)->rs_left
+= count
;
954 spin_unlock_irqrestore(&device
->al_lock
, flags
);
963 struct bm_extent
*_bme_get(struct drbd_device
*device
, unsigned int enr
)
965 struct lc_element
*e
;
966 struct bm_extent
*bm_ext
;
968 unsigned long rs_flags
;
970 spin_lock_irq(&device
->al_lock
);
971 if (device
->resync_locked
> device
->resync
->nr_elements
/2) {
972 spin_unlock_irq(&device
->al_lock
);
975 e
= lc_get(device
->resync
, enr
);
976 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
978 if (bm_ext
->lce
.lc_number
!= enr
) {
979 bm_ext
->rs_left
= drbd_bm_e_weight(device
, enr
);
980 bm_ext
->rs_failed
= 0;
981 lc_committed(device
->resync
);
984 if (bm_ext
->lce
.refcnt
== 1)
985 device
->resync_locked
++;
986 set_bit(BME_NO_WRITES
, &bm_ext
->flags
);
988 rs_flags
= device
->resync
->flags
;
989 spin_unlock_irq(&device
->al_lock
);
991 wake_up(&device
->al_wait
);
994 if (rs_flags
& LC_STARVING
)
995 drbd_warn(device
, "Have to wait for element"
996 " (resync LRU too small?)\n");
997 BUG_ON(rs_flags
& LC_LOCKED
);
1003 static int _is_in_al(struct drbd_device
*device
, unsigned int enr
)
1007 spin_lock_irq(&device
->al_lock
);
1008 rv
= lc_is_used(device
->act_log
, enr
);
1009 spin_unlock_irq(&device
->al_lock
);
1015 * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
1016 * @device: DRBD device.
1017 * @sector: The sector number.
1019 * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
1021 int drbd_rs_begin_io(struct drbd_device
*device
, sector_t sector
)
1023 unsigned int enr
= BM_SECT_TO_EXT(sector
);
1024 struct bm_extent
*bm_ext
;
1026 int sa
= 200; /* Step aside 200 times, then grab the extent and let app-IO wait.
1027 200 times -> 20 seconds. */
1030 sig
= wait_event_interruptible(device
->al_wait
,
1031 (bm_ext
= _bme_get(device
, enr
)));
1035 if (test_bit(BME_LOCKED
, &bm_ext
->flags
))
1038 for (i
= 0; i
< AL_EXT_PER_BM_SECT
; i
++) {
1039 sig
= wait_event_interruptible(device
->al_wait
,
1040 !_is_in_al(device
, enr
* AL_EXT_PER_BM_SECT
+ i
) ||
1041 test_bit(BME_PRIORITY
, &bm_ext
->flags
));
1043 if (sig
|| (test_bit(BME_PRIORITY
, &bm_ext
->flags
) && sa
)) {
1044 spin_lock_irq(&device
->al_lock
);
1045 if (lc_put(device
->resync
, &bm_ext
->lce
) == 0) {
1046 bm_ext
->flags
= 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
1047 device
->resync_locked
--;
1048 wake_up(&device
->al_wait
);
1050 spin_unlock_irq(&device
->al_lock
);
1053 if (schedule_timeout_interruptible(HZ
/10))
1055 if (sa
&& --sa
== 0)
1056 drbd_warn(device
, "drbd_rs_begin_io() stepped aside for 20sec."
1057 "Resync stalled?\n");
1061 set_bit(BME_LOCKED
, &bm_ext
->flags
);
1066 * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
1067 * @device: DRBD device.
1068 * @sector: The sector number.
1070 * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
1071 * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
1072 * if there is still application IO going on in this area.
1074 int drbd_try_rs_begin_io(struct drbd_device
*device
, sector_t sector
)
1076 unsigned int enr
= BM_SECT_TO_EXT(sector
);
1077 const unsigned int al_enr
= enr
*AL_EXT_PER_BM_SECT
;
1078 struct lc_element
*e
;
1079 struct bm_extent
*bm_ext
;
1082 spin_lock_irq(&device
->al_lock
);
1083 if (device
->resync_wenr
!= LC_FREE
&& device
->resync_wenr
!= enr
) {
1084 /* in case you have very heavy scattered io, it may
1085 * stall the syncer undefined if we give up the ref count
1086 * when we try again and requeue.
1088 * if we don't give up the refcount, but the next time
1089 * we are scheduled this extent has been "synced" by new
1090 * application writes, we'd miss the lc_put on the
1091 * extent we keep the refcount on.
1092 * so we remembered which extent we had to try again, and
1093 * if the next requested one is something else, we do
1094 * the lc_put here...
1095 * we also have to wake_up
1097 e
= lc_find(device
->resync
, device
->resync_wenr
);
1098 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1100 D_ASSERT(device
, !test_bit(BME_LOCKED
, &bm_ext
->flags
));
1101 D_ASSERT(device
, test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1102 clear_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1103 device
->resync_wenr
= LC_FREE
;
1104 if (lc_put(device
->resync
, &bm_ext
->lce
) == 0)
1105 device
->resync_locked
--;
1106 wake_up(&device
->al_wait
);
1108 drbd_alert(device
, "LOGIC BUG\n");
1112 e
= lc_try_get(device
->resync
, enr
);
1113 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1115 if (test_bit(BME_LOCKED
, &bm_ext
->flags
))
1117 if (!test_and_set_bit(BME_NO_WRITES
, &bm_ext
->flags
)) {
1118 device
->resync_locked
++;
1120 /* we did set the BME_NO_WRITES,
1121 * but then could not set BME_LOCKED,
1122 * so we tried again.
1123 * drop the extra reference. */
1124 bm_ext
->lce
.refcnt
--;
1125 D_ASSERT(device
, bm_ext
->lce
.refcnt
> 0);
1129 /* do we rather want to try later? */
1130 if (device
->resync_locked
> device
->resync
->nr_elements
-3)
1132 /* Do or do not. There is no try. -- Yoda */
1133 e
= lc_get(device
->resync
, enr
);
1134 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1136 const unsigned long rs_flags
= device
->resync
->flags
;
1137 if (rs_flags
& LC_STARVING
)
1138 drbd_warn(device
, "Have to wait for element"
1139 " (resync LRU too small?)\n");
1140 BUG_ON(rs_flags
& LC_LOCKED
);
1143 if (bm_ext
->lce
.lc_number
!= enr
) {
1144 bm_ext
->rs_left
= drbd_bm_e_weight(device
, enr
);
1145 bm_ext
->rs_failed
= 0;
1146 lc_committed(device
->resync
);
1147 wake_up(&device
->al_wait
);
1148 D_ASSERT(device
, test_bit(BME_LOCKED
, &bm_ext
->flags
) == 0);
1150 set_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1151 D_ASSERT(device
, bm_ext
->lce
.refcnt
== 1);
1152 device
->resync_locked
++;
1156 for (i
= 0; i
< AL_EXT_PER_BM_SECT
; i
++) {
1157 if (lc_is_used(device
->act_log
, al_enr
+i
))
1160 set_bit(BME_LOCKED
, &bm_ext
->flags
);
1162 device
->resync_wenr
= LC_FREE
;
1163 spin_unlock_irq(&device
->al_lock
);
1168 device
->resync_wenr
= enr
;
1169 spin_unlock_irq(&device
->al_lock
);
1173 void drbd_rs_complete_io(struct drbd_device
*device
, sector_t sector
)
1175 unsigned int enr
= BM_SECT_TO_EXT(sector
);
1176 struct lc_element
*e
;
1177 struct bm_extent
*bm_ext
;
1178 unsigned long flags
;
1180 spin_lock_irqsave(&device
->al_lock
, flags
);
1181 e
= lc_find(device
->resync
, enr
);
1182 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1184 spin_unlock_irqrestore(&device
->al_lock
, flags
);
1185 if (__ratelimit(&drbd_ratelimit_state
))
1186 drbd_err(device
, "drbd_rs_complete_io() called, but extent not found\n");
1190 if (bm_ext
->lce
.refcnt
== 0) {
1191 spin_unlock_irqrestore(&device
->al_lock
, flags
);
1192 drbd_err(device
, "drbd_rs_complete_io(,%llu [=%u]) called, "
1193 "but refcnt is 0!?\n",
1194 (unsigned long long)sector
, enr
);
1198 if (lc_put(device
->resync
, &bm_ext
->lce
) == 0) {
1199 bm_ext
->flags
= 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1200 device
->resync_locked
--;
1201 wake_up(&device
->al_wait
);
1204 spin_unlock_irqrestore(&device
->al_lock
, flags
);
1208 * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1209 * @device: DRBD device.
1211 void drbd_rs_cancel_all(struct drbd_device
*device
)
1213 spin_lock_irq(&device
->al_lock
);
1215 if (get_ldev_if_state(device
, D_FAILED
)) { /* Makes sure ->resync is there. */
1216 lc_reset(device
->resync
);
1219 device
->resync_locked
= 0;
1220 device
->resync_wenr
= LC_FREE
;
1221 spin_unlock_irq(&device
->al_lock
);
1222 wake_up(&device
->al_wait
);
1226 * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1227 * @device: DRBD device.
1229 * Returns 0 upon success, -EAGAIN if at least one reference count was
1232 int drbd_rs_del_all(struct drbd_device
*device
)
1234 struct lc_element
*e
;
1235 struct bm_extent
*bm_ext
;
1238 spin_lock_irq(&device
->al_lock
);
1240 if (get_ldev_if_state(device
, D_FAILED
)) {
1241 /* ok, ->resync is there. */
1242 for (i
= 0; i
< device
->resync
->nr_elements
; i
++) {
1243 e
= lc_element_by_index(device
->resync
, i
);
1244 bm_ext
= lc_entry(e
, struct bm_extent
, lce
);
1245 if (bm_ext
->lce
.lc_number
== LC_FREE
)
1247 if (bm_ext
->lce
.lc_number
== device
->resync_wenr
) {
1248 drbd_info(device
, "dropping %u in drbd_rs_del_all, apparently"
1249 " got 'synced' by application io\n",
1250 device
->resync_wenr
);
1251 D_ASSERT(device
, !test_bit(BME_LOCKED
, &bm_ext
->flags
));
1252 D_ASSERT(device
, test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1253 clear_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1254 device
->resync_wenr
= LC_FREE
;
1255 lc_put(device
->resync
, &bm_ext
->lce
);
1257 if (bm_ext
->lce
.refcnt
!= 0) {
1258 drbd_info(device
, "Retrying drbd_rs_del_all() later. "
1259 "refcnt=%d\n", bm_ext
->lce
.refcnt
);
1261 spin_unlock_irq(&device
->al_lock
);
1264 D_ASSERT(device
, !test_bit(BME_LOCKED
, &bm_ext
->flags
));
1265 D_ASSERT(device
, !test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1266 lc_del(device
->resync
, &bm_ext
->lce
);
1268 D_ASSERT(device
, device
->resync
->used
== 0);
1271 spin_unlock_irq(&device
->al_lock
);
1272 wake_up(&device
->al_wait
);
1278 * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks
1279 * @device: DRBD device.
1280 * @sector: The sector number.
1281 * @size: Size of failed IO operation, in byte.
1283 void drbd_rs_failed_io(struct drbd_device
*device
, sector_t sector
, int size
)
1285 /* Is called from worker and receiver context _only_ */
1286 unsigned long sbnr
, ebnr
, lbnr
;
1287 unsigned long count
;
1288 sector_t esector
, nr_sectors
;
1291 if (size
<= 0 || !IS_ALIGNED(size
, 512) || size
> DRBD_MAX_BIO_SIZE
) {
1292 drbd_err(device
, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n",
1293 (unsigned long long)sector
, size
);
1296 nr_sectors
= drbd_get_capacity(device
->this_bdev
);
1297 esector
= sector
+ (size
>> 9) - 1;
1299 if (!expect(sector
< nr_sectors
))
1301 if (!expect(esector
< nr_sectors
))
1302 esector
= nr_sectors
- 1;
1304 lbnr
= BM_SECT_TO_BIT(nr_sectors
-1);
1307 * round up start sector, round down end sector. we make sure we only
1308 * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */
1309 if (unlikely(esector
< BM_SECT_PER_BIT
-1))
1311 if (unlikely(esector
== (nr_sectors
-1)))
1314 ebnr
= BM_SECT_TO_BIT(esector
- (BM_SECT_PER_BIT
-1));
1315 sbnr
= BM_SECT_TO_BIT(sector
+ BM_SECT_PER_BIT
-1);
1321 * ok, (capacity & 7) != 0 sometimes, but who cares...
1322 * we count rs_{total,left} in bits, not sectors.
1324 spin_lock_irq(&device
->al_lock
);
1325 count
= drbd_bm_count_bits(device
, sbnr
, ebnr
);
1327 device
->rs_failed
+= count
;
1329 if (get_ldev(device
)) {
1330 drbd_try_clear_on_disk_bm(device
, sector
, count
, false);
1334 /* just wake_up unconditional now, various lc_chaged(),
1335 * lc_put() in drbd_try_clear_on_disk_bm(). */
1338 spin_unlock_irq(&device
->al_lock
);
1340 wake_up(&device
->al_wait
);