]>
git.proxmox.com Git - mirror_ubuntu-eoan-kernel.git/blob - drivers/block/drbd/drbd_actlog.c
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/drbd.h>
29 #include "drbd_wrappers.h"
31 /* We maintain a trivial check sum in our on disk activity log.
32 * With that we can ensure correct operation even when the storage
33 * device might do a partial (last) sector write while losing power.
35 struct __packed al_transaction
{
40 u32 extent
; } updates
[1 + AL_EXTENTS_PT
];
44 struct update_odbm_work
{
49 struct update_al_work
{
51 struct lc_element
*al_ext
;
52 struct completion event
;
54 /* if old_enr != LC_FREE, write corresponding bitmap sector, too */
58 struct drbd_atodb_wait
{
60 struct completion io_done
;
61 struct drbd_conf
*mdev
;
66 int w_al_write_transaction(struct drbd_conf
*, struct drbd_work
*, int);
68 static int _drbd_md_sync_page_io(struct drbd_conf
*mdev
,
69 struct drbd_backing_dev
*bdev
,
70 struct page
*page
, sector_t sector
,
74 struct drbd_md_io md_io
;
78 init_completion(&md_io
.event
);
81 if ((rw
& WRITE
) && !test_bit(MD_NO_FUA
, &mdev
->flags
))
85 bio
= bio_alloc(GFP_NOIO
, 1);
86 bio
->bi_bdev
= bdev
->md_bdev
;
87 bio
->bi_sector
= sector
;
88 ok
= (bio_add_page(bio
, page
, size
, 0) == size
);
91 bio
->bi_private
= &md_io
;
92 bio
->bi_end_io
= drbd_md_io_complete
;
95 if (drbd_insert_fault(mdev
, (rw
& WRITE
) ? DRBD_FAULT_MD_WR
: DRBD_FAULT_MD_RD
))
99 wait_for_completion(&md_io
.event
);
100 ok
= bio_flagged(bio
, BIO_UPTODATE
) && md_io
.error
== 0;
107 int drbd_md_sync_page_io(struct drbd_conf
*mdev
, struct drbd_backing_dev
*bdev
,
108 sector_t sector
, int rw
)
110 int logical_block_size
, mask
, ok
;
112 struct page
*iop
= mdev
->md_io_page
;
114 D_ASSERT(mutex_is_locked(&mdev
->md_io_mutex
));
116 BUG_ON(!bdev
->md_bdev
);
118 logical_block_size
= bdev_logical_block_size(bdev
->md_bdev
);
119 if (logical_block_size
== 0)
120 logical_block_size
= MD_SECTOR_SIZE
;
122 /* in case logical_block_size != 512 [ s390 only? ] */
123 if (logical_block_size
!= MD_SECTOR_SIZE
) {
124 mask
= (logical_block_size
/ MD_SECTOR_SIZE
) - 1;
125 D_ASSERT(mask
== 1 || mask
== 3 || mask
== 7);
126 D_ASSERT(logical_block_size
== (mask
+1) * MD_SECTOR_SIZE
);
127 offset
= sector
& mask
;
128 sector
= sector
& ~mask
;
129 iop
= mdev
->md_io_tmpp
;
132 /* these are GFP_KERNEL pages, pre-allocated
133 * on device initialization */
134 void *p
= page_address(mdev
->md_io_page
);
135 void *hp
= page_address(mdev
->md_io_tmpp
);
137 ok
= _drbd_md_sync_page_io(mdev
, bdev
, iop
, sector
,
138 READ
, logical_block_size
);
141 dev_err(DEV
, "drbd_md_sync_page_io(,%llus,"
142 "READ [logical_block_size!=512]) failed!\n",
143 (unsigned long long)sector
);
147 memcpy(hp
+ offset
*MD_SECTOR_SIZE
, p
, MD_SECTOR_SIZE
);
151 if (sector
< drbd_md_first_sector(bdev
) ||
152 sector
> drbd_md_last_sector(bdev
))
153 dev_alert(DEV
, "%s [%d]:%s(,%llus,%s) out of range md access!\n",
154 current
->comm
, current
->pid
, __func__
,
155 (unsigned long long)sector
, (rw
& WRITE
) ? "WRITE" : "READ");
157 ok
= _drbd_md_sync_page_io(mdev
, bdev
, iop
, sector
, rw
, logical_block_size
);
159 dev_err(DEV
, "drbd_md_sync_page_io(,%llus,%s) failed!\n",
160 (unsigned long long)sector
, (rw
& WRITE
) ? "WRITE" : "READ");
164 if (logical_block_size
!= MD_SECTOR_SIZE
&& !(rw
& WRITE
)) {
165 void *p
= page_address(mdev
->md_io_page
);
166 void *hp
= page_address(mdev
->md_io_tmpp
);
168 memcpy(p
, hp
+ offset
*MD_SECTOR_SIZE
, MD_SECTOR_SIZE
);
174 static struct lc_element
*_al_get(struct drbd_conf
*mdev
, unsigned int enr
)
176 struct lc_element
*al_ext
;
177 struct lc_element
*tmp
;
178 unsigned long al_flags
= 0;
181 spin_lock_irq(&mdev
->al_lock
);
182 tmp
= lc_find(mdev
->resync
, enr
/AL_EXT_PER_BM_SECT
);
183 if (unlikely(tmp
!= NULL
)) {
184 struct bm_extent
*bm_ext
= lc_entry(tmp
, struct bm_extent
, lce
);
185 if (test_bit(BME_NO_WRITES
, &bm_ext
->flags
)) {
186 wake
= !test_and_set_bit(BME_PRIORITY
, &bm_ext
->flags
);
187 spin_unlock_irq(&mdev
->al_lock
);
189 wake_up(&mdev
->al_wait
);
193 al_ext
= lc_get(mdev
->act_log
, enr
);
194 al_flags
= mdev
->act_log
->flags
;
195 spin_unlock_irq(&mdev
->al_lock
);
199 if (al_flags & LC_STARVING)
200 dev_warn(DEV, "Have to wait for LRU element (AL too small?)\n");
201 if (al_flags & LC_DIRTY)
202 dev_warn(DEV, "Ongoing AL update (AL device too slow?)\n");
209 void drbd_al_begin_io(struct drbd_conf
*mdev
, sector_t sector
)
211 unsigned int enr
= (sector
>> (AL_EXTENT_SHIFT
-9));
212 struct lc_element
*al_ext
;
213 struct update_al_work al_work
;
215 D_ASSERT(atomic_read(&mdev
->local_cnt
) > 0);
217 wait_event(mdev
->al_wait
, (al_ext
= _al_get(mdev
, enr
)));
219 if (al_ext
->lc_number
!= enr
) {
220 /* drbd_al_write_transaction(mdev,al_ext,enr);
221 * recurses into generic_make_request(), which
222 * disallows recursion, bios being serialized on the
223 * current->bio_tail list now.
224 * we have to delegate updates to the activity log
225 * to the worker thread. */
226 init_completion(&al_work
.event
);
227 al_work
.al_ext
= al_ext
;
229 al_work
.old_enr
= al_ext
->lc_number
;
230 al_work
.w
.cb
= w_al_write_transaction
;
231 drbd_queue_work_front(&mdev
->data
.work
, &al_work
.w
);
232 wait_for_completion(&al_work
.event
);
236 spin_lock_irq(&mdev
->al_lock
);
237 lc_changed(mdev
->act_log
, al_ext
);
238 spin_unlock_irq(&mdev
->al_lock
);
239 wake_up(&mdev
->al_wait
);
243 void drbd_al_complete_io(struct drbd_conf
*mdev
, sector_t sector
)
245 unsigned int enr
= (sector
>> (AL_EXTENT_SHIFT
-9));
246 struct lc_element
*extent
;
249 spin_lock_irqsave(&mdev
->al_lock
, flags
);
251 extent
= lc_find(mdev
->act_log
, enr
);
254 spin_unlock_irqrestore(&mdev
->al_lock
, flags
);
255 dev_err(DEV
, "al_complete_io() called on inactive extent %u\n", enr
);
259 if (lc_put(mdev
->act_log
, extent
) == 0)
260 wake_up(&mdev
->al_wait
);
262 spin_unlock_irqrestore(&mdev
->al_lock
, flags
);
265 #if (PAGE_SHIFT + 3) < (AL_EXTENT_SHIFT - BM_BLOCK_SHIFT)
266 /* Currently BM_BLOCK_SHIFT, BM_EXT_SHIFT and AL_EXTENT_SHIFT
267 * are still coupled, or assume too much about their relation.
268 * Code below will not work if this is violated.
269 * Will be cleaned up with some followup patch.
274 static unsigned int al_extent_to_bm_page(unsigned int al_enr
)
279 /* al extent number to bit */
280 (AL_EXTENT_SHIFT
- BM_BLOCK_SHIFT
));
283 static unsigned int rs_extent_to_bm_page(unsigned int rs_enr
)
288 /* al extent number to bit */
289 (BM_EXT_SHIFT
- BM_BLOCK_SHIFT
));
293 w_al_write_transaction(struct drbd_conf
*mdev
, struct drbd_work
*w
, int unused
)
295 struct update_al_work
*aw
= container_of(w
, struct update_al_work
, w
);
296 struct lc_element
*updated
= aw
->al_ext
;
297 const unsigned int new_enr
= aw
->enr
;
298 const unsigned int evicted
= aw
->old_enr
;
299 struct al_transaction
*buffer
;
302 unsigned int extent_nr
;
305 if (!get_ldev(mdev
)) {
307 "disk is %s, cannot start al transaction (-%d +%d)\n",
308 drbd_disk_str(mdev
->state
.disk
), evicted
, new_enr
);
309 complete(&((struct update_al_work
*)w
)->event
);
312 /* do we have to do a bitmap write, first?
313 * TODO reduce maximum latency:
314 * submit both bios, then wait for both,
315 * instead of doing two synchronous sector writes.
316 * For now, we must not write the transaction,
317 * if we cannot write out the bitmap of the evicted extent. */
318 if (mdev
->state
.conn
< C_CONNECTED
&& evicted
!= LC_FREE
)
319 drbd_bm_write_page(mdev
, al_extent_to_bm_page(evicted
));
321 /* The bitmap write may have failed, causing a state change. */
322 if (mdev
->state
.disk
< D_INCONSISTENT
) {
324 "disk is %s, cannot write al transaction (-%d +%d)\n",
325 drbd_disk_str(mdev
->state
.disk
), evicted
, new_enr
);
326 complete(&((struct update_al_work
*)w
)->event
);
331 mutex_lock(&mdev
->md_io_mutex
); /* protects md_io_buffer, al_tr_cycle, ... */
332 buffer
= (struct al_transaction
*)page_address(mdev
->md_io_page
);
334 buffer
->magic
= __constant_cpu_to_be32(DRBD_MAGIC
);
335 buffer
->tr_number
= cpu_to_be32(mdev
->al_tr_number
);
337 n
= lc_index_of(mdev
->act_log
, updated
);
339 buffer
->updates
[0].pos
= cpu_to_be32(n
);
340 buffer
->updates
[0].extent
= cpu_to_be32(new_enr
);
344 mx
= min_t(int, AL_EXTENTS_PT
,
345 mdev
->act_log
->nr_elements
- mdev
->al_tr_cycle
);
346 for (i
= 0; i
< mx
; i
++) {
347 unsigned idx
= mdev
->al_tr_cycle
+ i
;
348 extent_nr
= lc_element_by_index(mdev
->act_log
, idx
)->lc_number
;
349 buffer
->updates
[i
+1].pos
= cpu_to_be32(idx
);
350 buffer
->updates
[i
+1].extent
= cpu_to_be32(extent_nr
);
351 xor_sum
^= extent_nr
;
353 for (; i
< AL_EXTENTS_PT
; i
++) {
354 buffer
->updates
[i
+1].pos
= __constant_cpu_to_be32(-1);
355 buffer
->updates
[i
+1].extent
= __constant_cpu_to_be32(LC_FREE
);
358 mdev
->al_tr_cycle
+= AL_EXTENTS_PT
;
359 if (mdev
->al_tr_cycle
>= mdev
->act_log
->nr_elements
)
360 mdev
->al_tr_cycle
= 0;
362 buffer
->xor_sum
= cpu_to_be32(xor_sum
);
364 sector
= mdev
->ldev
->md
.md_offset
365 + mdev
->ldev
->md
.al_offset
+ mdev
->al_tr_pos
;
367 if (!drbd_md_sync_page_io(mdev
, mdev
->ldev
, sector
, WRITE
))
368 drbd_chk_io_error(mdev
, 1, true);
370 if (++mdev
->al_tr_pos
>
371 div_ceil(mdev
->act_log
->nr_elements
, AL_EXTENTS_PT
))
374 D_ASSERT(mdev
->al_tr_pos
< MD_AL_MAX_SIZE
);
375 mdev
->al_tr_number
++;
377 mutex_unlock(&mdev
->md_io_mutex
);
379 complete(&((struct update_al_work
*)w
)->event
);
386 * drbd_al_read_tr() - Read a single transaction from the on disk activity log
387 * @mdev: DRBD device.
388 * @bdev: Block device to read form.
389 * @b: pointer to an al_transaction.
390 * @index: On disk slot of the transaction to read.
392 * Returns -1 on IO error, 0 on checksum error and 1 upon success.
394 static int drbd_al_read_tr(struct drbd_conf
*mdev
,
395 struct drbd_backing_dev
*bdev
,
396 struct al_transaction
*b
,
403 sector
= bdev
->md
.md_offset
+ bdev
->md
.al_offset
+ index
;
405 /* Dont process error normally,
406 * as this is done before disk is attached! */
407 if (!drbd_md_sync_page_io(mdev
, bdev
, sector
, READ
))
410 rv
= (be32_to_cpu(b
->magic
) == DRBD_MAGIC
);
412 for (i
= 0; i
< AL_EXTENTS_PT
+ 1; i
++)
413 xor_sum
^= be32_to_cpu(b
->updates
[i
].extent
);
414 rv
&= (xor_sum
== be32_to_cpu(b
->xor_sum
));
420 * drbd_al_read_log() - Restores the activity log from its on disk representation.
421 * @mdev: DRBD device.
422 * @bdev: Block device to read form.
424 * Returns 1 on success, returns 0 when reading the log failed due to IO errors.
426 int drbd_al_read_log(struct drbd_conf
*mdev
, struct drbd_backing_dev
*bdev
)
428 struct al_transaction
*buffer
;
432 int active_extents
= 0;
433 int transactions
= 0;
441 mx
= div_ceil(mdev
->act_log
->nr_elements
, AL_EXTENTS_PT
);
443 /* lock out all other meta data io for now,
444 * and make sure the page is mapped.
446 mutex_lock(&mdev
->md_io_mutex
);
447 buffer
= page_address(mdev
->md_io_page
);
449 /* Find the valid transaction in the log */
450 for (i
= 0; i
<= mx
; i
++) {
451 rv
= drbd_al_read_tr(mdev
, bdev
, buffer
, i
);
455 mutex_unlock(&mdev
->md_io_mutex
);
458 cnr
= be32_to_cpu(buffer
->tr_number
);
460 if (++found_valid
== 1) {
467 if ((int)cnr
- (int)from_tnr
< 0) {
468 D_ASSERT(from_tnr
- cnr
+ i
- from
== mx
+1);
472 if ((int)cnr
- (int)to_tnr
> 0) {
473 D_ASSERT(cnr
- to_tnr
== i
- to
);
480 dev_warn(DEV
, "No usable activity log found.\n");
481 mutex_unlock(&mdev
->md_io_mutex
);
485 /* Read the valid transactions.
486 * dev_info(DEV, "Reading from %d to %d.\n",from,to); */
490 unsigned int extent_nr
;
493 rv
= drbd_al_read_tr(mdev
, bdev
, buffer
, i
);
494 ERR_IF(rv
== 0) goto cancel
;
496 mutex_unlock(&mdev
->md_io_mutex
);
500 trn
= be32_to_cpu(buffer
->tr_number
);
502 spin_lock_irq(&mdev
->al_lock
);
504 /* This loop runs backwards because in the cyclic
505 elements there might be an old version of the
506 updated element (in slot 0). So the element in slot 0
507 can overwrite old versions. */
508 for (j
= AL_EXTENTS_PT
; j
>= 0; j
--) {
509 pos
= be32_to_cpu(buffer
->updates
[j
].pos
);
510 extent_nr
= be32_to_cpu(buffer
->updates
[j
].extent
);
512 if (extent_nr
== LC_FREE
)
515 lc_set(mdev
->act_log
, extent_nr
, pos
);
518 spin_unlock_irq(&mdev
->al_lock
);
530 mdev
->al_tr_number
= to_tnr
+1;
531 mdev
->al_tr_pos
= to
;
532 if (++mdev
->al_tr_pos
>
533 div_ceil(mdev
->act_log
->nr_elements
, AL_EXTENTS_PT
))
536 /* ok, we are done with it */
537 mutex_unlock(&mdev
->md_io_mutex
);
539 dev_info(DEV
, "Found %d transactions (%d active extents) in activity log.\n",
540 transactions
, active_extents
);
546 * drbd_al_apply_to_bm() - Sets the bitmap to diry(1) where covered ba active AL extents
547 * @mdev: DRBD device.
549 void drbd_al_apply_to_bm(struct drbd_conf
*mdev
)
552 unsigned long add
= 0;
556 wait_event(mdev
->al_wait
, lc_try_lock(mdev
->act_log
));
558 for (i
= 0; i
< mdev
->act_log
->nr_elements
; i
++) {
559 enr
= lc_element_by_index(mdev
->act_log
, i
)->lc_number
;
562 tmp
= drbd_bm_ALe_set_all(mdev
, enr
);
563 dynamic_dev_dbg(DEV
, "AL: set %d bits in extent %u\n", tmp
, enr
);
567 lc_unlock(mdev
->act_log
);
568 wake_up(&mdev
->al_wait
);
570 dev_info(DEV
, "Marked additional %s as out-of-sync based on AL.\n",
571 ppsize(ppb
, Bit2KB(add
)));
574 static int _try_lc_del(struct drbd_conf
*mdev
, struct lc_element
*al_ext
)
578 spin_lock_irq(&mdev
->al_lock
);
579 rv
= (al_ext
->refcnt
== 0);
581 lc_del(mdev
->act_log
, al_ext
);
582 spin_unlock_irq(&mdev
->al_lock
);
588 * drbd_al_shrink() - Removes all active extents form the activity log
589 * @mdev: DRBD device.
591 * Removes all active extents form the activity log, waiting until
592 * the reference count of each entry dropped to 0 first, of course.
594 * You need to lock mdev->act_log with lc_try_lock() / lc_unlock()
596 void drbd_al_shrink(struct drbd_conf
*mdev
)
598 struct lc_element
*al_ext
;
601 D_ASSERT(test_bit(__LC_DIRTY
, &mdev
->act_log
->flags
));
603 for (i
= 0; i
< mdev
->act_log
->nr_elements
; i
++) {
604 al_ext
= lc_element_by_index(mdev
->act_log
, i
);
605 if (al_ext
->lc_number
== LC_FREE
)
607 wait_event(mdev
->al_wait
, _try_lc_del(mdev
, al_ext
));
610 wake_up(&mdev
->al_wait
);
613 static int w_update_odbm(struct drbd_conf
*mdev
, struct drbd_work
*w
, int unused
)
615 struct update_odbm_work
*udw
= container_of(w
, struct update_odbm_work
, w
);
617 if (!get_ldev(mdev
)) {
618 if (__ratelimit(&drbd_ratelimit_state
))
619 dev_warn(DEV
, "Can not update on disk bitmap, local IO disabled.\n");
624 drbd_bm_write_page(mdev
, rs_extent_to_bm_page(udw
->enr
));
629 if (drbd_bm_total_weight(mdev
) <= mdev
->rs_failed
) {
630 switch (mdev
->state
.conn
) {
631 case C_SYNC_SOURCE
: case C_SYNC_TARGET
:
632 case C_PAUSED_SYNC_S
: case C_PAUSED_SYNC_T
:
633 drbd_resync_finished(mdev
);
639 drbd_bcast_sync_progress(mdev
);
645 /* ATTENTION. The AL's extents are 4MB each, while the extents in the
646 * resync LRU-cache are 16MB each.
647 * The caller of this function has to hold an get_ldev() reference.
649 * TODO will be obsoleted once we have a caching lru of the on disk bitmap
651 static void drbd_try_clear_on_disk_bm(struct drbd_conf
*mdev
, sector_t sector
,
652 int count
, int success
)
654 struct lc_element
*e
;
655 struct update_odbm_work
*udw
;
659 D_ASSERT(atomic_read(&mdev
->local_cnt
));
661 /* I simply assume that a sector/size pair never crosses
662 * a 16 MB extent border. (Currently this is true...) */
663 enr
= BM_SECT_TO_EXT(sector
);
665 e
= lc_get(mdev
->resync
, enr
);
667 struct bm_extent
*ext
= lc_entry(e
, struct bm_extent
, lce
);
668 if (ext
->lce
.lc_number
== enr
) {
670 ext
->rs_left
-= count
;
672 ext
->rs_failed
+= count
;
673 if (ext
->rs_left
< ext
->rs_failed
) {
674 dev_err(DEV
, "BAD! sector=%llus enr=%u rs_left=%d "
675 "rs_failed=%d count=%d\n",
676 (unsigned long long)sector
,
677 ext
->lce
.lc_number
, ext
->rs_left
,
678 ext
->rs_failed
, count
);
681 lc_put(mdev
->resync
, &ext
->lce
);
682 drbd_force_state(mdev
, NS(conn
, C_DISCONNECTING
));
686 /* Normally this element should be in the cache,
687 * since drbd_rs_begin_io() pulled it already in.
689 * But maybe an application write finished, and we set
690 * something outside the resync lru_cache in sync.
692 int rs_left
= drbd_bm_e_weight(mdev
, enr
);
693 if (ext
->flags
!= 0) {
694 dev_warn(DEV
, "changing resync lce: %d[%u;%02lx]"
696 ext
->lce
.lc_number
, ext
->rs_left
,
697 ext
->flags
, enr
, rs_left
);
700 if (ext
->rs_failed
) {
701 dev_warn(DEV
, "Kicking resync_lru element enr=%u "
702 "out with rs_failed=%d\n",
703 ext
->lce
.lc_number
, ext
->rs_failed
);
705 ext
->rs_left
= rs_left
;
706 ext
->rs_failed
= success
? 0 : count
;
707 lc_changed(mdev
->resync
, &ext
->lce
);
709 lc_put(mdev
->resync
, &ext
->lce
);
710 /* no race, we are within the al_lock! */
712 if (ext
->rs_left
== ext
->rs_failed
) {
715 udw
= kmalloc(sizeof(*udw
), GFP_ATOMIC
);
717 udw
->enr
= ext
->lce
.lc_number
;
718 udw
->w
.cb
= w_update_odbm
;
719 drbd_queue_work_front(&mdev
->data
.work
, &udw
->w
);
721 dev_warn(DEV
, "Could not kmalloc an udw\n");
725 dev_err(DEV
, "lc_get() failed! locked=%d/%d flags=%lu\n",
727 mdev
->resync
->nr_elements
,
728 mdev
->resync
->flags
);
732 void drbd_advance_rs_marks(struct drbd_conf
*mdev
, unsigned long still_to_go
)
734 unsigned long now
= jiffies
;
735 unsigned long last
= mdev
->rs_mark_time
[mdev
->rs_last_mark
];
736 int next
= (mdev
->rs_last_mark
+ 1) % DRBD_SYNC_MARKS
;
737 if (time_after_eq(now
, last
+ DRBD_SYNC_MARK_STEP
)) {
738 if (mdev
->rs_mark_left
[mdev
->rs_last_mark
] != still_to_go
&&
739 mdev
->state
.conn
!= C_PAUSED_SYNC_T
&&
740 mdev
->state
.conn
!= C_PAUSED_SYNC_S
) {
741 mdev
->rs_mark_time
[next
] = now
;
742 mdev
->rs_mark_left
[next
] = still_to_go
;
743 mdev
->rs_last_mark
= next
;
748 /* clear the bit corresponding to the piece of storage in question:
749 * size byte of data starting from sector. Only clear a bits of the affected
750 * one ore more _aligned_ BM_BLOCK_SIZE blocks.
752 * called by worker on C_SYNC_TARGET and receiver on SyncSource.
755 void __drbd_set_in_sync(struct drbd_conf
*mdev
, sector_t sector
, int size
,
756 const char *file
, const unsigned int line
)
758 /* Is called from worker and receiver context _only_ */
759 unsigned long sbnr
, ebnr
, lbnr
;
760 unsigned long count
= 0;
761 sector_t esector
, nr_sectors
;
765 if (size
<= 0 || (size
& 0x1ff) != 0 || size
> DRBD_MAX_BIO_SIZE
) {
766 dev_err(DEV
, "drbd_set_in_sync: sector=%llus size=%d nonsense!\n",
767 (unsigned long long)sector
, size
);
770 nr_sectors
= drbd_get_capacity(mdev
->this_bdev
);
771 esector
= sector
+ (size
>> 9) - 1;
773 ERR_IF(sector
>= nr_sectors
) return;
774 ERR_IF(esector
>= nr_sectors
) esector
= (nr_sectors
-1);
776 lbnr
= BM_SECT_TO_BIT(nr_sectors
-1);
778 /* we clear it (in sync).
779 * round up start sector, round down end sector. we make sure we only
780 * clear full, aligned, BM_BLOCK_SIZE (4K) blocks */
781 if (unlikely(esector
< BM_SECT_PER_BIT
-1))
783 if (unlikely(esector
== (nr_sectors
-1)))
786 ebnr
= BM_SECT_TO_BIT(esector
- (BM_SECT_PER_BIT
-1));
787 sbnr
= BM_SECT_TO_BIT(sector
+ BM_SECT_PER_BIT
-1);
793 * ok, (capacity & 7) != 0 sometimes, but who cares...
794 * we count rs_{total,left} in bits, not sectors.
796 count
= drbd_bm_clear_bits(mdev
, sbnr
, ebnr
);
797 if (count
&& get_ldev(mdev
)) {
798 drbd_advance_rs_marks(mdev
, drbd_bm_total_weight(mdev
));
799 spin_lock_irqsave(&mdev
->al_lock
, flags
);
800 drbd_try_clear_on_disk_bm(mdev
, sector
, count
, true);
801 spin_unlock_irqrestore(&mdev
->al_lock
, flags
);
803 /* just wake_up unconditional now, various lc_chaged(),
804 * lc_put() in drbd_try_clear_on_disk_bm(). */
809 wake_up(&mdev
->al_wait
);
813 * this is intended to set one request worth of data out of sync.
814 * affects at least 1 bit,
815 * and at most 1+DRBD_MAX_BIO_SIZE/BM_BLOCK_SIZE bits.
817 * called by tl_clear and drbd_send_dblock (==drbd_make_request).
818 * so this can be _any_ process.
820 int __drbd_set_out_of_sync(struct drbd_conf
*mdev
, sector_t sector
, int size
,
821 const char *file
, const unsigned int line
)
823 unsigned long sbnr
, ebnr
, lbnr
, flags
;
824 sector_t esector
, nr_sectors
;
825 unsigned int enr
, count
= 0;
826 struct lc_element
*e
;
828 if (size
<= 0 || (size
& 0x1ff) != 0 || size
> DRBD_MAX_BIO_SIZE
) {
829 dev_err(DEV
, "sector: %llus, size: %d\n",
830 (unsigned long long)sector
, size
);
835 return 0; /* no disk, no metadata, no bitmap to set bits in */
837 nr_sectors
= drbd_get_capacity(mdev
->this_bdev
);
838 esector
= sector
+ (size
>> 9) - 1;
840 ERR_IF(sector
>= nr_sectors
)
842 ERR_IF(esector
>= nr_sectors
)
843 esector
= (nr_sectors
-1);
845 lbnr
= BM_SECT_TO_BIT(nr_sectors
-1);
847 /* we set it out of sync,
848 * we do not need to round anything here */
849 sbnr
= BM_SECT_TO_BIT(sector
);
850 ebnr
= BM_SECT_TO_BIT(esector
);
852 /* ok, (capacity & 7) != 0 sometimes, but who cares...
853 * we count rs_{total,left} in bits, not sectors. */
854 spin_lock_irqsave(&mdev
->al_lock
, flags
);
855 count
= drbd_bm_set_bits(mdev
, sbnr
, ebnr
);
857 enr
= BM_SECT_TO_EXT(sector
);
858 e
= lc_find(mdev
->resync
, enr
);
860 lc_entry(e
, struct bm_extent
, lce
)->rs_left
+= count
;
861 spin_unlock_irqrestore(&mdev
->al_lock
, flags
);
870 struct bm_extent
*_bme_get(struct drbd_conf
*mdev
, unsigned int enr
)
872 struct lc_element
*e
;
873 struct bm_extent
*bm_ext
;
875 unsigned long rs_flags
;
877 spin_lock_irq(&mdev
->al_lock
);
878 if (mdev
->resync_locked
> mdev
->resync
->nr_elements
/2) {
879 spin_unlock_irq(&mdev
->al_lock
);
882 e
= lc_get(mdev
->resync
, enr
);
883 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
885 if (bm_ext
->lce
.lc_number
!= enr
) {
886 bm_ext
->rs_left
= drbd_bm_e_weight(mdev
, enr
);
887 bm_ext
->rs_failed
= 0;
888 lc_changed(mdev
->resync
, &bm_ext
->lce
);
891 if (bm_ext
->lce
.refcnt
== 1)
892 mdev
->resync_locked
++;
893 set_bit(BME_NO_WRITES
, &bm_ext
->flags
);
895 rs_flags
= mdev
->resync
->flags
;
896 spin_unlock_irq(&mdev
->al_lock
);
898 wake_up(&mdev
->al_wait
);
901 if (rs_flags
& LC_STARVING
)
902 dev_warn(DEV
, "Have to wait for element"
903 " (resync LRU too small?)\n");
904 BUG_ON(rs_flags
& LC_DIRTY
);
910 static int _is_in_al(struct drbd_conf
*mdev
, unsigned int enr
)
912 struct lc_element
*al_ext
;
915 spin_lock_irq(&mdev
->al_lock
);
916 if (unlikely(enr
== mdev
->act_log
->new_number
))
919 al_ext
= lc_find(mdev
->act_log
, enr
);
925 spin_unlock_irq(&mdev
->al_lock
);
929 dev_info(DEV, "Delaying sync read until app's write is done\n");
936 * drbd_rs_begin_io() - Gets an extent in the resync LRU cache and sets it to BME_LOCKED
937 * @mdev: DRBD device.
938 * @sector: The sector number.
940 * This functions sleeps on al_wait. Returns 0 on success, -EINTR if interrupted.
942 int drbd_rs_begin_io(struct drbd_conf
*mdev
, sector_t sector
)
944 unsigned int enr
= BM_SECT_TO_EXT(sector
);
945 struct bm_extent
*bm_ext
;
947 int sa
= 200; /* Step aside 200 times, then grab the extent and let app-IO wait.
948 200 times -> 20 seconds. */
951 sig
= wait_event_interruptible(mdev
->al_wait
,
952 (bm_ext
= _bme_get(mdev
, enr
)));
956 if (test_bit(BME_LOCKED
, &bm_ext
->flags
))
959 for (i
= 0; i
< AL_EXT_PER_BM_SECT
; i
++) {
960 sig
= wait_event_interruptible(mdev
->al_wait
,
961 !_is_in_al(mdev
, enr
* AL_EXT_PER_BM_SECT
+ i
) ||
962 test_bit(BME_PRIORITY
, &bm_ext
->flags
));
964 if (sig
|| (test_bit(BME_PRIORITY
, &bm_ext
->flags
) && sa
)) {
965 spin_lock_irq(&mdev
->al_lock
);
966 if (lc_put(mdev
->resync
, &bm_ext
->lce
) == 0) {
967 bm_ext
->flags
= 0; /* clears BME_NO_WRITES and eventually BME_PRIORITY */
968 mdev
->resync_locked
--;
969 wake_up(&mdev
->al_wait
);
971 spin_unlock_irq(&mdev
->al_lock
);
974 if (schedule_timeout_interruptible(HZ
/10))
977 dev_warn(DEV
,"drbd_rs_begin_io() stepped aside for 20sec."
978 "Resync stalled?\n");
982 set_bit(BME_LOCKED
, &bm_ext
->flags
);
987 * drbd_try_rs_begin_io() - Gets an extent in the resync LRU cache, does not sleep
988 * @mdev: DRBD device.
989 * @sector: The sector number.
991 * Gets an extent in the resync LRU cache, sets it to BME_NO_WRITES, then
992 * tries to set it to BME_LOCKED. Returns 0 upon success, and -EAGAIN
993 * if there is still application IO going on in this area.
995 int drbd_try_rs_begin_io(struct drbd_conf
*mdev
, sector_t sector
)
997 unsigned int enr
= BM_SECT_TO_EXT(sector
);
998 const unsigned int al_enr
= enr
*AL_EXT_PER_BM_SECT
;
999 struct lc_element
*e
;
1000 struct bm_extent
*bm_ext
;
1003 spin_lock_irq(&mdev
->al_lock
);
1004 if (mdev
->resync_wenr
!= LC_FREE
&& mdev
->resync_wenr
!= enr
) {
1005 /* in case you have very heavy scattered io, it may
1006 * stall the syncer undefined if we give up the ref count
1007 * when we try again and requeue.
1009 * if we don't give up the refcount, but the next time
1010 * we are scheduled this extent has been "synced" by new
1011 * application writes, we'd miss the lc_put on the
1012 * extent we keep the refcount on.
1013 * so we remembered which extent we had to try again, and
1014 * if the next requested one is something else, we do
1015 * the lc_put here...
1016 * we also have to wake_up
1018 e
= lc_find(mdev
->resync
, mdev
->resync_wenr
);
1019 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1021 D_ASSERT(!test_bit(BME_LOCKED
, &bm_ext
->flags
));
1022 D_ASSERT(test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1023 clear_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1024 mdev
->resync_wenr
= LC_FREE
;
1025 if (lc_put(mdev
->resync
, &bm_ext
->lce
) == 0)
1026 mdev
->resync_locked
--;
1027 wake_up(&mdev
->al_wait
);
1029 dev_alert(DEV
, "LOGIC BUG\n");
1033 e
= lc_try_get(mdev
->resync
, enr
);
1034 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1036 if (test_bit(BME_LOCKED
, &bm_ext
->flags
))
1038 if (!test_and_set_bit(BME_NO_WRITES
, &bm_ext
->flags
)) {
1039 mdev
->resync_locked
++;
1041 /* we did set the BME_NO_WRITES,
1042 * but then could not set BME_LOCKED,
1043 * so we tried again.
1044 * drop the extra reference. */
1045 bm_ext
->lce
.refcnt
--;
1046 D_ASSERT(bm_ext
->lce
.refcnt
> 0);
1050 /* do we rather want to try later? */
1051 if (mdev
->resync_locked
> mdev
->resync
->nr_elements
-3)
1053 /* Do or do not. There is no try. -- Yoda */
1054 e
= lc_get(mdev
->resync
, enr
);
1055 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1057 const unsigned long rs_flags
= mdev
->resync
->flags
;
1058 if (rs_flags
& LC_STARVING
)
1059 dev_warn(DEV
, "Have to wait for element"
1060 " (resync LRU too small?)\n");
1061 BUG_ON(rs_flags
& LC_DIRTY
);
1064 if (bm_ext
->lce
.lc_number
!= enr
) {
1065 bm_ext
->rs_left
= drbd_bm_e_weight(mdev
, enr
);
1066 bm_ext
->rs_failed
= 0;
1067 lc_changed(mdev
->resync
, &bm_ext
->lce
);
1068 wake_up(&mdev
->al_wait
);
1069 D_ASSERT(test_bit(BME_LOCKED
, &bm_ext
->flags
) == 0);
1071 set_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1072 D_ASSERT(bm_ext
->lce
.refcnt
== 1);
1073 mdev
->resync_locked
++;
1077 for (i
= 0; i
< AL_EXT_PER_BM_SECT
; i
++) {
1078 if (unlikely(al_enr
+i
== mdev
->act_log
->new_number
))
1080 if (lc_is_used(mdev
->act_log
, al_enr
+i
))
1083 set_bit(BME_LOCKED
, &bm_ext
->flags
);
1085 mdev
->resync_wenr
= LC_FREE
;
1086 spin_unlock_irq(&mdev
->al_lock
);
1091 mdev
->resync_wenr
= enr
;
1092 spin_unlock_irq(&mdev
->al_lock
);
1096 void drbd_rs_complete_io(struct drbd_conf
*mdev
, sector_t sector
)
1098 unsigned int enr
= BM_SECT_TO_EXT(sector
);
1099 struct lc_element
*e
;
1100 struct bm_extent
*bm_ext
;
1101 unsigned long flags
;
1103 spin_lock_irqsave(&mdev
->al_lock
, flags
);
1104 e
= lc_find(mdev
->resync
, enr
);
1105 bm_ext
= e
? lc_entry(e
, struct bm_extent
, lce
) : NULL
;
1107 spin_unlock_irqrestore(&mdev
->al_lock
, flags
);
1108 if (__ratelimit(&drbd_ratelimit_state
))
1109 dev_err(DEV
, "drbd_rs_complete_io() called, but extent not found\n");
1113 if (bm_ext
->lce
.refcnt
== 0) {
1114 spin_unlock_irqrestore(&mdev
->al_lock
, flags
);
1115 dev_err(DEV
, "drbd_rs_complete_io(,%llu [=%u]) called, "
1116 "but refcnt is 0!?\n",
1117 (unsigned long long)sector
, enr
);
1121 if (lc_put(mdev
->resync
, &bm_ext
->lce
) == 0) {
1122 bm_ext
->flags
= 0; /* clear BME_LOCKED, BME_NO_WRITES and BME_PRIORITY */
1123 mdev
->resync_locked
--;
1124 wake_up(&mdev
->al_wait
);
1127 spin_unlock_irqrestore(&mdev
->al_lock
, flags
);
1131 * drbd_rs_cancel_all() - Removes all extents from the resync LRU (even BME_LOCKED)
1132 * @mdev: DRBD device.
1134 void drbd_rs_cancel_all(struct drbd_conf
*mdev
)
1136 spin_lock_irq(&mdev
->al_lock
);
1138 if (get_ldev_if_state(mdev
, D_FAILED
)) { /* Makes sure ->resync is there. */
1139 lc_reset(mdev
->resync
);
1142 mdev
->resync_locked
= 0;
1143 mdev
->resync_wenr
= LC_FREE
;
1144 spin_unlock_irq(&mdev
->al_lock
);
1145 wake_up(&mdev
->al_wait
);
1149 * drbd_rs_del_all() - Gracefully remove all extents from the resync LRU
1150 * @mdev: DRBD device.
1152 * Returns 0 upon success, -EAGAIN if at least one reference count was
1155 int drbd_rs_del_all(struct drbd_conf
*mdev
)
1157 struct lc_element
*e
;
1158 struct bm_extent
*bm_ext
;
1161 spin_lock_irq(&mdev
->al_lock
);
1163 if (get_ldev_if_state(mdev
, D_FAILED
)) {
1164 /* ok, ->resync is there. */
1165 for (i
= 0; i
< mdev
->resync
->nr_elements
; i
++) {
1166 e
= lc_element_by_index(mdev
->resync
, i
);
1167 bm_ext
= lc_entry(e
, struct bm_extent
, lce
);
1168 if (bm_ext
->lce
.lc_number
== LC_FREE
)
1170 if (bm_ext
->lce
.lc_number
== mdev
->resync_wenr
) {
1171 dev_info(DEV
, "dropping %u in drbd_rs_del_all, apparently"
1172 " got 'synced' by application io\n",
1174 D_ASSERT(!test_bit(BME_LOCKED
, &bm_ext
->flags
));
1175 D_ASSERT(test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1176 clear_bit(BME_NO_WRITES
, &bm_ext
->flags
);
1177 mdev
->resync_wenr
= LC_FREE
;
1178 lc_put(mdev
->resync
, &bm_ext
->lce
);
1180 if (bm_ext
->lce
.refcnt
!= 0) {
1181 dev_info(DEV
, "Retrying drbd_rs_del_all() later. "
1182 "refcnt=%d\n", bm_ext
->lce
.refcnt
);
1184 spin_unlock_irq(&mdev
->al_lock
);
1187 D_ASSERT(!test_bit(BME_LOCKED
, &bm_ext
->flags
));
1188 D_ASSERT(!test_bit(BME_NO_WRITES
, &bm_ext
->flags
));
1189 lc_del(mdev
->resync
, &bm_ext
->lce
);
1191 D_ASSERT(mdev
->resync
->used
== 0);
1194 spin_unlock_irq(&mdev
->al_lock
);
1200 * drbd_rs_failed_io() - Record information on a failure to resync the specified blocks
1201 * @mdev: DRBD device.
1202 * @sector: The sector number.
1203 * @size: Size of failed IO operation, in byte.
1205 void drbd_rs_failed_io(struct drbd_conf
*mdev
, sector_t sector
, int size
)
1207 /* Is called from worker and receiver context _only_ */
1208 unsigned long sbnr
, ebnr
, lbnr
;
1209 unsigned long count
;
1210 sector_t esector
, nr_sectors
;
1213 if (size
<= 0 || (size
& 0x1ff) != 0 || size
> DRBD_MAX_BIO_SIZE
) {
1214 dev_err(DEV
, "drbd_rs_failed_io: sector=%llus size=%d nonsense!\n",
1215 (unsigned long long)sector
, size
);
1218 nr_sectors
= drbd_get_capacity(mdev
->this_bdev
);
1219 esector
= sector
+ (size
>> 9) - 1;
1221 ERR_IF(sector
>= nr_sectors
) return;
1222 ERR_IF(esector
>= nr_sectors
) esector
= (nr_sectors
-1);
1224 lbnr
= BM_SECT_TO_BIT(nr_sectors
-1);
1227 * round up start sector, round down end sector. we make sure we only
1228 * handle full, aligned, BM_BLOCK_SIZE (4K) blocks */
1229 if (unlikely(esector
< BM_SECT_PER_BIT
-1))
1231 if (unlikely(esector
== (nr_sectors
-1)))
1234 ebnr
= BM_SECT_TO_BIT(esector
- (BM_SECT_PER_BIT
-1));
1235 sbnr
= BM_SECT_TO_BIT(sector
+ BM_SECT_PER_BIT
-1);
1241 * ok, (capacity & 7) != 0 sometimes, but who cares...
1242 * we count rs_{total,left} in bits, not sectors.
1244 spin_lock_irq(&mdev
->al_lock
);
1245 count
= drbd_bm_count_bits(mdev
, sbnr
, ebnr
);
1247 mdev
->rs_failed
+= count
;
1249 if (get_ldev(mdev
)) {
1250 drbd_try_clear_on_disk_bm(mdev
, sector
, count
, false);
1254 /* just wake_up unconditional now, various lc_chaged(),
1255 * lc_put() in drbd_try_clear_on_disk_bm(). */
1258 spin_unlock_irq(&mdev
->al_lock
);
1260 wake_up(&mdev
->al_wait
);