4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 Copyright (C) 2001-2008, LINBIT Information Technologies GmbH.
7 Copyright (C) 1999-2008, Philipp Reisner <philipp.reisner@linbit.com>.
8 Copyright (C) 2002-2008, 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/module.h>
27 #include <linux/drbd.h>
28 #include <linux/sched.h>
29 #include <linux/wait.h>
31 #include <linux/memcontrol.h>
32 #include <linux/mm_inline.h>
33 #include <linux/slab.h>
34 #include <linux/random.h>
35 #include <linux/string.h>
36 #include <linux/scatterlist.h>
39 #include "drbd_protocol.h"
42 static int w_make_ov_request(struct drbd_work
*w
, int cancel
);
46 * drbd_md_io_complete (defined here)
47 * drbd_request_endio (defined here)
48 * drbd_peer_request_endio (defined here)
49 * bm_async_io_complete (defined in drbd_bitmap.c)
51 * For all these callbacks, note the following:
52 * The callbacks will be called in irq context by the IDE drivers,
53 * and in Softirqs/Tasklets/BH context by the SCSI drivers.
54 * Try to get the locking right :)
59 /* About the global_state_lock
60 Each state transition on an device holds a read lock. In case we have
61 to evaluate the resync after dependencies, we grab a write lock, because
62 we need stable states on all devices for that. */
63 rwlock_t global_state_lock
;
65 /* used for synchronous meta data and bitmap IO
66 * submitted by drbd_md_sync_page_io()
68 void drbd_md_io_complete(struct bio
*bio
, int error
)
70 struct drbd_md_io
*md_io
;
71 struct drbd_device
*device
;
73 md_io
= (struct drbd_md_io
*)bio
->bi_private
;
74 device
= container_of(md_io
, struct drbd_device
, md_io
);
78 /* We grabbed an extra reference in _drbd_md_sync_page_io() to be able
79 * to timeout on the lower level device, and eventually detach from it.
80 * If this io completion runs after that timeout expired, this
81 * drbd_md_put_buffer() may allow us to finally try and re-attach.
82 * During normal operation, this only puts that extra reference
84 * Make sure we first drop the reference, and only then signal
85 * completion, or we may (in drbd_al_read_log()) cycle so fast into the
86 * next drbd_md_sync_page_io(), that we trigger the
87 * ASSERT(atomic_read(&device->md_io_in_use) == 1) there.
89 drbd_md_put_buffer(device
);
91 wake_up(&device
->misc_wait
);
93 if (device
->ldev
) /* special case: drbd_md_read() during drbd_adm_attach() */
97 /* reads on behalf of the partner,
98 * "submitted" by the receiver
100 static void drbd_endio_read_sec_final(struct drbd_peer_request
*peer_req
) __releases(local
)
102 unsigned long flags
= 0;
103 struct drbd_device
*device
= peer_req
->w
.device
;
105 spin_lock_irqsave(&device
->connection
->req_lock
, flags
);
106 device
->read_cnt
+= peer_req
->i
.size
>> 9;
107 list_del(&peer_req
->w
.list
);
108 if (list_empty(&device
->read_ee
))
109 wake_up(&device
->ee_wait
);
110 if (test_bit(__EE_WAS_ERROR
, &peer_req
->flags
))
111 __drbd_chk_io_error(device
, DRBD_READ_ERROR
);
112 spin_unlock_irqrestore(&device
->connection
->req_lock
, flags
);
114 drbd_queue_work(&device
->connection
->sender_work
, &peer_req
->w
);
118 /* writes on behalf of the partner, or resync writes,
119 * "submitted" by the receiver, final stage. */
120 static void drbd_endio_write_sec_final(struct drbd_peer_request
*peer_req
) __releases(local
)
122 unsigned long flags
= 0;
123 struct drbd_device
*device
= peer_req
->w
.device
;
124 struct drbd_interval i
;
127 int do_al_complete_io
;
129 /* after we moved peer_req to done_ee,
130 * we may no longer access it,
131 * it may be freed/reused already!
132 * (as soon as we release the req_lock) */
134 do_al_complete_io
= peer_req
->flags
& EE_CALL_AL_COMPLETE_IO
;
135 block_id
= peer_req
->block_id
;
137 spin_lock_irqsave(&device
->connection
->req_lock
, flags
);
138 device
->writ_cnt
+= peer_req
->i
.size
>> 9;
139 list_move_tail(&peer_req
->w
.list
, &device
->done_ee
);
142 * Do not remove from the write_requests tree here: we did not send the
143 * Ack yet and did not wake possibly waiting conflicting requests.
144 * Removed from the tree from "drbd_process_done_ee" within the
145 * appropriate w.cb (e_end_block/e_end_resync_block) or from
146 * _drbd_clear_done_ee.
149 do_wake
= list_empty(block_id
== ID_SYNCER
? &device
->sync_ee
: &device
->active_ee
);
151 if (test_bit(__EE_WAS_ERROR
, &peer_req
->flags
))
152 __drbd_chk_io_error(device
, DRBD_WRITE_ERROR
);
153 spin_unlock_irqrestore(&device
->connection
->req_lock
, flags
);
155 if (block_id
== ID_SYNCER
)
156 drbd_rs_complete_io(device
, i
.sector
);
159 wake_up(&device
->ee_wait
);
161 if (do_al_complete_io
)
162 drbd_al_complete_io(device
, &i
);
164 wake_asender(device
->connection
);
168 /* writes on behalf of the partner, or resync writes,
169 * "submitted" by the receiver.
171 void drbd_peer_request_endio(struct bio
*bio
, int error
)
173 struct drbd_peer_request
*peer_req
= bio
->bi_private
;
174 struct drbd_device
*device
= peer_req
->w
.device
;
175 int uptodate
= bio_flagged(bio
, BIO_UPTODATE
);
176 int is_write
= bio_data_dir(bio
) == WRITE
;
178 if (error
&& __ratelimit(&drbd_ratelimit_state
))
179 dev_warn(DEV
, "%s: error=%d s=%llus\n",
180 is_write
? "write" : "read", error
,
181 (unsigned long long)peer_req
->i
.sector
);
182 if (!error
&& !uptodate
) {
183 if (__ratelimit(&drbd_ratelimit_state
))
184 dev_warn(DEV
, "%s: setting error to -EIO s=%llus\n",
185 is_write
? "write" : "read",
186 (unsigned long long)peer_req
->i
.sector
);
187 /* strange behavior of some lower level drivers...
188 * fail the request by clearing the uptodate flag,
189 * but do not return any error?! */
194 set_bit(__EE_WAS_ERROR
, &peer_req
->flags
);
196 bio_put(bio
); /* no need for the bio anymore */
197 if (atomic_dec_and_test(&peer_req
->pending_bios
)) {
199 drbd_endio_write_sec_final(peer_req
);
201 drbd_endio_read_sec_final(peer_req
);
205 /* read, readA or write requests on R_PRIMARY coming from drbd_make_request
207 void drbd_request_endio(struct bio
*bio
, int error
)
210 struct drbd_request
*req
= bio
->bi_private
;
211 struct drbd_device
*device
= req
->w
.device
;
212 struct bio_and_error m
;
213 enum drbd_req_event what
;
214 int uptodate
= bio_flagged(bio
, BIO_UPTODATE
);
216 if (!error
&& !uptodate
) {
217 dev_warn(DEV
, "p %s: setting error to -EIO\n",
218 bio_data_dir(bio
) == WRITE
? "write" : "read");
219 /* strange behavior of some lower level drivers...
220 * fail the request by clearing the uptodate flag,
221 * but do not return any error?! */
226 /* If this request was aborted locally before,
227 * but now was completed "successfully",
228 * chances are that this caused arbitrary data corruption.
230 * "aborting" requests, or force-detaching the disk, is intended for
231 * completely blocked/hung local backing devices which do no longer
232 * complete requests at all, not even do error completions. In this
233 * situation, usually a hard-reset and failover is the only way out.
235 * By "aborting", basically faking a local error-completion,
236 * we allow for a more graceful swichover by cleanly migrating services.
237 * Still the affected node has to be rebooted "soon".
239 * By completing these requests, we allow the upper layers to re-use
240 * the associated data pages.
242 * If later the local backing device "recovers", and now DMAs some data
243 * from disk into the original request pages, in the best case it will
244 * just put random data into unused pages; but typically it will corrupt
245 * meanwhile completely unrelated data, causing all sorts of damage.
247 * Which means delayed successful completion,
248 * especially for READ requests,
249 * is a reason to panic().
251 * We assume that a delayed *error* completion is OK,
252 * though we still will complain noisily about it.
254 if (unlikely(req
->rq_state
& RQ_LOCAL_ABORTED
)) {
255 if (__ratelimit(&drbd_ratelimit_state
))
256 dev_emerg(DEV
, "delayed completion of aborted local request; disk-timeout may be too aggressive\n");
259 panic("possible random memory corruption caused by delayed completion of aborted local request\n");
262 /* to avoid recursion in __req_mod */
263 if (unlikely(error
)) {
264 what
= (bio_data_dir(bio
) == WRITE
)
265 ? WRITE_COMPLETED_WITH_ERROR
266 : (bio_rw(bio
) == READ
)
267 ? READ_COMPLETED_WITH_ERROR
268 : READ_AHEAD_COMPLETED_WITH_ERROR
;
272 bio_put(req
->private_bio
);
273 req
->private_bio
= ERR_PTR(error
);
275 /* not req_mod(), we need irqsave here! */
276 spin_lock_irqsave(&device
->connection
->req_lock
, flags
);
277 __req_mod(req
, what
, &m
);
278 spin_unlock_irqrestore(&device
->connection
->req_lock
, flags
);
282 complete_master_bio(device
, &m
);
285 void drbd_csum_ee(struct drbd_device
*device
, struct crypto_hash
*tfm
,
286 struct drbd_peer_request
*peer_req
, void *digest
)
288 struct hash_desc desc
;
289 struct scatterlist sg
;
290 struct page
*page
= peer_req
->pages
;
297 sg_init_table(&sg
, 1);
298 crypto_hash_init(&desc
);
300 while ((tmp
= page_chain_next(page
))) {
301 /* all but the last page will be fully used */
302 sg_set_page(&sg
, page
, PAGE_SIZE
, 0);
303 crypto_hash_update(&desc
, &sg
, sg
.length
);
306 /* and now the last, possibly only partially used page */
307 len
= peer_req
->i
.size
& (PAGE_SIZE
- 1);
308 sg_set_page(&sg
, page
, len
?: PAGE_SIZE
, 0);
309 crypto_hash_update(&desc
, &sg
, sg
.length
);
310 crypto_hash_final(&desc
, digest
);
313 void drbd_csum_bio(struct drbd_device
*device
, struct crypto_hash
*tfm
, struct bio
*bio
, void *digest
)
315 struct hash_desc desc
;
316 struct scatterlist sg
;
318 struct bvec_iter iter
;
323 sg_init_table(&sg
, 1);
324 crypto_hash_init(&desc
);
326 bio_for_each_segment(bvec
, bio
, iter
) {
327 sg_set_page(&sg
, bvec
.bv_page
, bvec
.bv_len
, bvec
.bv_offset
);
328 crypto_hash_update(&desc
, &sg
, sg
.length
);
330 crypto_hash_final(&desc
, digest
);
333 /* MAYBE merge common code with w_e_end_ov_req */
334 static int w_e_send_csum(struct drbd_work
*w
, int cancel
)
336 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
337 struct drbd_device
*device
= w
->device
;
342 if (unlikely(cancel
))
345 if (unlikely((peer_req
->flags
& EE_WAS_ERROR
) != 0))
348 digest_size
= crypto_hash_digestsize(device
->connection
->csums_tfm
);
349 digest
= kmalloc(digest_size
, GFP_NOIO
);
351 sector_t sector
= peer_req
->i
.sector
;
352 unsigned int size
= peer_req
->i
.size
;
353 drbd_csum_ee(device
, device
->connection
->csums_tfm
, peer_req
, digest
);
354 /* Free peer_req and pages before send.
355 * In case we block on congestion, we could otherwise run into
356 * some distributed deadlock, if the other side blocks on
357 * congestion as well, because our receiver blocks in
358 * drbd_alloc_pages due to pp_in_use > max_buffers. */
359 drbd_free_peer_req(device
, peer_req
);
361 inc_rs_pending(device
);
362 err
= drbd_send_drequest_csum(device
, sector
, size
,
367 dev_err(DEV
, "kmalloc() of digest failed.\n");
373 drbd_free_peer_req(device
, peer_req
);
376 dev_err(DEV
, "drbd_send_drequest(..., csum) failed\n");
380 #define GFP_TRY (__GFP_HIGHMEM | __GFP_NOWARN)
382 static int read_for_csum(struct drbd_device
*device
, sector_t sector
, int size
)
384 struct drbd_peer_request
*peer_req
;
386 if (!get_ldev(device
))
389 if (drbd_rs_should_slow_down(device
, sector
))
392 /* GFP_TRY, because if there is no memory available right now, this may
393 * be rescheduled for later. It is "only" background resync, after all. */
394 peer_req
= drbd_alloc_peer_req(device
, ID_SYNCER
/* unused */, sector
,
399 peer_req
->w
.cb
= w_e_send_csum
;
400 spin_lock_irq(&device
->connection
->req_lock
);
401 list_add(&peer_req
->w
.list
, &device
->read_ee
);
402 spin_unlock_irq(&device
->connection
->req_lock
);
404 atomic_add(size
>> 9, &device
->rs_sect_ev
);
405 if (drbd_submit_peer_request(device
, peer_req
, READ
, DRBD_FAULT_RS_RD
) == 0)
408 /* If it failed because of ENOMEM, retry should help. If it failed
409 * because bio_add_page failed (probably broken lower level driver),
410 * retry may or may not help.
411 * If it does not, you may need to force disconnect. */
412 spin_lock_irq(&device
->connection
->req_lock
);
413 list_del(&peer_req
->w
.list
);
414 spin_unlock_irq(&device
->connection
->req_lock
);
416 drbd_free_peer_req(device
, peer_req
);
422 int w_resync_timer(struct drbd_work
*w
, int cancel
)
424 struct drbd_device
*device
= w
->device
;
425 switch (device
->state
.conn
) {
427 w_make_ov_request(w
, cancel
);
430 w_make_resync_request(w
, cancel
);
437 void resync_timer_fn(unsigned long data
)
439 struct drbd_device
*device
= (struct drbd_device
*) data
;
441 if (list_empty(&device
->resync_work
.list
))
442 drbd_queue_work(&device
->connection
->sender_work
, &device
->resync_work
);
445 static void fifo_set(struct fifo_buffer
*fb
, int value
)
449 for (i
= 0; i
< fb
->size
; i
++)
450 fb
->values
[i
] = value
;
453 static int fifo_push(struct fifo_buffer
*fb
, int value
)
457 ov
= fb
->values
[fb
->head_index
];
458 fb
->values
[fb
->head_index
++] = value
;
460 if (fb
->head_index
>= fb
->size
)
466 static void fifo_add_val(struct fifo_buffer
*fb
, int value
)
470 for (i
= 0; i
< fb
->size
; i
++)
471 fb
->values
[i
] += value
;
474 struct fifo_buffer
*fifo_alloc(int fifo_size
)
476 struct fifo_buffer
*fb
;
478 fb
= kzalloc(sizeof(struct fifo_buffer
) + sizeof(int) * fifo_size
, GFP_NOIO
);
483 fb
->size
= fifo_size
;
489 static int drbd_rs_controller(struct drbd_device
*device
)
491 struct disk_conf
*dc
;
492 unsigned int sect_in
; /* Number of sectors that came in since the last turn */
493 unsigned int want
; /* The number of sectors we want in the proxy */
494 int req_sect
; /* Number of sectors to request in this turn */
495 int correction
; /* Number of sectors more we need in the proxy*/
496 int cps
; /* correction per invocation of drbd_rs_controller() */
497 int steps
; /* Number of time steps to plan ahead */
500 struct fifo_buffer
*plan
;
502 sect_in
= atomic_xchg(&device
->rs_sect_in
, 0); /* Number of sectors that came in */
503 device
->rs_in_flight
-= sect_in
;
505 dc
= rcu_dereference(device
->ldev
->disk_conf
);
506 plan
= rcu_dereference(device
->rs_plan_s
);
508 steps
= plan
->size
; /* (dc->c_plan_ahead * 10 * SLEEP_TIME) / HZ; */
510 if (device
->rs_in_flight
+ sect_in
== 0) { /* At start of resync */
511 want
= ((dc
->resync_rate
* 2 * SLEEP_TIME
) / HZ
) * steps
;
512 } else { /* normal path */
513 want
= dc
->c_fill_target
? dc
->c_fill_target
:
514 sect_in
* dc
->c_delay_target
* HZ
/ (SLEEP_TIME
* 10);
517 correction
= want
- device
->rs_in_flight
- plan
->total
;
520 cps
= correction
/ steps
;
521 fifo_add_val(plan
, cps
);
522 plan
->total
+= cps
* steps
;
524 /* What we do in this step */
525 curr_corr
= fifo_push(plan
, 0);
526 plan
->total
-= curr_corr
;
528 req_sect
= sect_in
+ curr_corr
;
532 max_sect
= (dc
->c_max_rate
* 2 * SLEEP_TIME
) / HZ
;
533 if (req_sect
> max_sect
)
537 dev_warn(DEV, "si=%u if=%d wa=%u co=%d st=%d cps=%d pl=%d cc=%d rs=%d\n",
538 sect_in, device->rs_in_flight, want, correction,
539 steps, cps, device->rs_planed, curr_corr, req_sect);
545 static int drbd_rs_number_requests(struct drbd_device
*device
)
550 if (rcu_dereference(device
->rs_plan_s
)->size
) {
551 number
= drbd_rs_controller(device
) >> (BM_BLOCK_SHIFT
- 9);
552 device
->c_sync_rate
= number
* HZ
* (BM_BLOCK_SIZE
/ 1024) / SLEEP_TIME
;
554 device
->c_sync_rate
= rcu_dereference(device
->ldev
->disk_conf
)->resync_rate
;
555 number
= SLEEP_TIME
* device
->c_sync_rate
/ ((BM_BLOCK_SIZE
/ 1024) * HZ
);
559 /* ignore the amount of pending requests, the resync controller should
560 * throttle down to incoming reply rate soon enough anyways. */
564 int w_make_resync_request(struct drbd_work
*w
, int cancel
)
566 struct drbd_device
*device
= w
->device
;
569 const sector_t capacity
= drbd_get_capacity(device
->this_bdev
);
571 int number
, rollback_i
, size
;
572 int align
, queued
, sndbuf
;
575 if (unlikely(cancel
))
578 if (device
->rs_total
== 0) {
580 drbd_resync_finished(device
);
584 if (!get_ldev(device
)) {
585 /* Since we only need to access device->rsync a
586 get_ldev_if_state(device,D_FAILED) would be sufficient, but
587 to continue resync with a broken disk makes no sense at
589 dev_err(DEV
, "Disk broke down during resync!\n");
593 max_bio_size
= queue_max_hw_sectors(device
->rq_queue
) << 9;
594 number
= drbd_rs_number_requests(device
);
598 for (i
= 0; i
< number
; i
++) {
599 /* Stop generating RS requests, when half of the send buffer is filled */
600 mutex_lock(&device
->connection
->data
.mutex
);
601 if (device
->connection
->data
.socket
) {
602 queued
= device
->connection
->data
.socket
->sk
->sk_wmem_queued
;
603 sndbuf
= device
->connection
->data
.socket
->sk
->sk_sndbuf
;
608 mutex_unlock(&device
->connection
->data
.mutex
);
609 if (queued
> sndbuf
/ 2)
613 size
= BM_BLOCK_SIZE
;
614 bit
= drbd_bm_find_next(device
, device
->bm_resync_fo
);
616 if (bit
== DRBD_END_OF_BITMAP
) {
617 device
->bm_resync_fo
= drbd_bm_bits(device
);
622 sector
= BM_BIT_TO_SECT(bit
);
624 if (drbd_rs_should_slow_down(device
, sector
) ||
625 drbd_try_rs_begin_io(device
, sector
)) {
626 device
->bm_resync_fo
= bit
;
629 device
->bm_resync_fo
= bit
+ 1;
631 if (unlikely(drbd_bm_test_bit(device
, bit
) == 0)) {
632 drbd_rs_complete_io(device
, sector
);
636 #if DRBD_MAX_BIO_SIZE > BM_BLOCK_SIZE
637 /* try to find some adjacent bits.
638 * we stop if we have already the maximum req size.
640 * Additionally always align bigger requests, in order to
641 * be prepared for all stripe sizes of software RAIDs.
646 if (size
+ BM_BLOCK_SIZE
> max_bio_size
)
649 /* Be always aligned */
650 if (sector
& ((1<<(align
+3))-1))
653 /* do not cross extent boundaries */
654 if (((bit
+1) & BM_BLOCKS_PER_BM_EXT_MASK
) == 0)
656 /* now, is it actually dirty, after all?
657 * caution, drbd_bm_test_bit is tri-state for some
658 * obscure reason; ( b == 0 ) would get the out-of-band
659 * only accidentally right because of the "oddly sized"
660 * adjustment below */
661 if (drbd_bm_test_bit(device
, bit
+1) != 1)
664 size
+= BM_BLOCK_SIZE
;
665 if ((BM_BLOCK_SIZE
<< align
) <= size
)
669 /* if we merged some,
670 * reset the offset to start the next drbd_bm_find_next from */
671 if (size
> BM_BLOCK_SIZE
)
672 device
->bm_resync_fo
= bit
+ 1;
675 /* adjust very last sectors, in case we are oddly sized */
676 if (sector
+ (size
>>9) > capacity
)
677 size
= (capacity
-sector
)<<9;
678 if (device
->connection
->agreed_pro_version
>= 89 && device
->connection
->csums_tfm
) {
679 switch (read_for_csum(device
, sector
, size
)) {
680 case -EIO
: /* Disk failure */
683 case -EAGAIN
: /* allocation failed, or ldev busy */
684 drbd_rs_complete_io(device
, sector
);
685 device
->bm_resync_fo
= BM_SECT_TO_BIT(sector
);
697 inc_rs_pending(device
);
698 err
= drbd_send_drequest(device
, P_RS_DATA_REQUEST
,
699 sector
, size
, ID_SYNCER
);
701 dev_err(DEV
, "drbd_send_drequest() failed, aborting...\n");
702 dec_rs_pending(device
);
709 if (device
->bm_resync_fo
>= drbd_bm_bits(device
)) {
710 /* last syncer _request_ was sent,
711 * but the P_RS_DATA_REPLY not yet received. sync will end (and
712 * next sync group will resume), as soon as we receive the last
713 * resync data block, and the last bit is cleared.
714 * until then resync "work" is "inactive" ...
721 device
->rs_in_flight
+= (i
<< (BM_BLOCK_SHIFT
- 9));
722 mod_timer(&device
->resync_timer
, jiffies
+ SLEEP_TIME
);
727 static int w_make_ov_request(struct drbd_work
*w
, int cancel
)
729 struct drbd_device
*device
= w
->device
;
732 const sector_t capacity
= drbd_get_capacity(device
->this_bdev
);
733 bool stop_sector_reached
= false;
735 if (unlikely(cancel
))
738 number
= drbd_rs_number_requests(device
);
740 sector
= device
->ov_position
;
741 for (i
= 0; i
< number
; i
++) {
742 if (sector
>= capacity
)
745 /* We check for "finished" only in the reply path:
746 * w_e_end_ov_reply().
747 * We need to send at least one request out. */
748 stop_sector_reached
= i
> 0
749 && verify_can_do_stop_sector(device
)
750 && sector
>= device
->ov_stop_sector
;
751 if (stop_sector_reached
)
754 size
= BM_BLOCK_SIZE
;
756 if (drbd_rs_should_slow_down(device
, sector
) ||
757 drbd_try_rs_begin_io(device
, sector
)) {
758 device
->ov_position
= sector
;
762 if (sector
+ (size
>>9) > capacity
)
763 size
= (capacity
-sector
)<<9;
765 inc_rs_pending(device
);
766 if (drbd_send_ov_request(device
, sector
, size
)) {
767 dec_rs_pending(device
);
770 sector
+= BM_SECT_PER_BIT
;
772 device
->ov_position
= sector
;
775 device
->rs_in_flight
+= (i
<< (BM_BLOCK_SHIFT
- 9));
776 if (i
== 0 || !stop_sector_reached
)
777 mod_timer(&device
->resync_timer
, jiffies
+ SLEEP_TIME
);
781 int w_ov_finished(struct drbd_work
*w
, int cancel
)
783 struct drbd_device
*device
= w
->device
;
785 ov_out_of_sync_print(device
);
786 drbd_resync_finished(device
);
791 static int w_resync_finished(struct drbd_work
*w
, int cancel
)
793 struct drbd_device
*device
= w
->device
;
796 drbd_resync_finished(device
);
801 static void ping_peer(struct drbd_device
*device
)
803 struct drbd_connection
*connection
= device
->connection
;
805 clear_bit(GOT_PING_ACK
, &connection
->flags
);
806 request_ping(connection
);
807 wait_event(connection
->ping_wait
,
808 test_bit(GOT_PING_ACK
, &connection
->flags
) || device
->state
.conn
< C_CONNECTED
);
811 int drbd_resync_finished(struct drbd_device
*device
)
813 unsigned long db
, dt
, dbdt
;
815 union drbd_state os
, ns
;
817 char *khelper_cmd
= NULL
;
820 /* Remove all elements from the resync LRU. Since future actions
821 * might set bits in the (main) bitmap, then the entries in the
822 * resync LRU would be wrong. */
823 if (drbd_rs_del_all(device
)) {
824 /* In case this is not possible now, most probably because
825 * there are P_RS_DATA_REPLY Packets lingering on the worker's
826 * queue (or even the read operations for those packets
827 * is not finished by now). Retry in 100ms. */
829 schedule_timeout_interruptible(HZ
/ 10);
830 w
= kmalloc(sizeof(struct drbd_work
), GFP_ATOMIC
);
832 w
->cb
= w_resync_finished
;
834 drbd_queue_work(&device
->connection
->sender_work
, w
);
837 dev_err(DEV
, "Warn failed to drbd_rs_del_all() and to kmalloc(w).\n");
840 dt
= (jiffies
- device
->rs_start
- device
->rs_paused
) / HZ
;
844 db
= device
->rs_total
;
845 /* adjust for verify start and stop sectors, respective reached position */
846 if (device
->state
.conn
== C_VERIFY_S
|| device
->state
.conn
== C_VERIFY_T
)
847 db
-= device
->ov_left
;
849 dbdt
= Bit2KB(db
/dt
);
850 device
->rs_paused
/= HZ
;
852 if (!get_ldev(device
))
857 spin_lock_irq(&device
->connection
->req_lock
);
858 os
= drbd_read_state(device
);
860 verify_done
= (os
.conn
== C_VERIFY_S
|| os
.conn
== C_VERIFY_T
);
862 /* This protects us against multiple calls (that can happen in the presence
863 of application IO), and against connectivity loss just before we arrive here. */
864 if (os
.conn
<= C_CONNECTED
)
868 ns
.conn
= C_CONNECTED
;
870 dev_info(DEV
, "%s done (total %lu sec; paused %lu sec; %lu K/sec)\n",
871 verify_done
? "Online verify" : "Resync",
872 dt
+ device
->rs_paused
, device
->rs_paused
, dbdt
);
874 n_oos
= drbd_bm_total_weight(device
);
876 if (os
.conn
== C_VERIFY_S
|| os
.conn
== C_VERIFY_T
) {
878 dev_alert(DEV
, "Online verify found %lu %dk block out of sync!\n",
880 khelper_cmd
= "out-of-sync";
883 D_ASSERT((n_oos
- device
->rs_failed
) == 0);
885 if (os
.conn
== C_SYNC_TARGET
|| os
.conn
== C_PAUSED_SYNC_T
)
886 khelper_cmd
= "after-resync-target";
888 if (device
->connection
->csums_tfm
&& device
->rs_total
) {
889 const unsigned long s
= device
->rs_same_csum
;
890 const unsigned long t
= device
->rs_total
;
893 (t
< 100000) ? ((s
*100)/t
) : (s
/(t
/100));
894 dev_info(DEV
, "%u %% had equal checksums, eliminated: %luK; "
895 "transferred %luK total %luK\n",
897 Bit2KB(device
->rs_same_csum
),
898 Bit2KB(device
->rs_total
- device
->rs_same_csum
),
899 Bit2KB(device
->rs_total
));
903 if (device
->rs_failed
) {
904 dev_info(DEV
, " %lu failed blocks\n", device
->rs_failed
);
906 if (os
.conn
== C_SYNC_TARGET
|| os
.conn
== C_PAUSED_SYNC_T
) {
907 ns
.disk
= D_INCONSISTENT
;
908 ns
.pdsk
= D_UP_TO_DATE
;
910 ns
.disk
= D_UP_TO_DATE
;
911 ns
.pdsk
= D_INCONSISTENT
;
914 ns
.disk
= D_UP_TO_DATE
;
915 ns
.pdsk
= D_UP_TO_DATE
;
917 if (os
.conn
== C_SYNC_TARGET
|| os
.conn
== C_PAUSED_SYNC_T
) {
918 if (device
->p_uuid
) {
920 for (i
= UI_BITMAP
; i
<= UI_HISTORY_END
; i
++)
921 _drbd_uuid_set(device
, i
, device
->p_uuid
[i
]);
922 drbd_uuid_set(device
, UI_BITMAP
, device
->ldev
->md
.uuid
[UI_CURRENT
]);
923 _drbd_uuid_set(device
, UI_CURRENT
, device
->p_uuid
[UI_CURRENT
]);
925 dev_err(DEV
, "device->p_uuid is NULL! BUG\n");
929 if (!(os
.conn
== C_VERIFY_S
|| os
.conn
== C_VERIFY_T
)) {
930 /* for verify runs, we don't update uuids here,
931 * so there would be nothing to report. */
932 drbd_uuid_set_bm(device
, 0UL);
933 drbd_print_uuids(device
, "updated UUIDs");
934 if (device
->p_uuid
) {
935 /* Now the two UUID sets are equal, update what we
936 * know of the peer. */
938 for (i
= UI_CURRENT
; i
<= UI_HISTORY_END
; i
++)
939 device
->p_uuid
[i
] = device
->ldev
->md
.uuid
[i
];
944 _drbd_set_state(device
, ns
, CS_VERBOSE
, NULL
);
946 spin_unlock_irq(&device
->connection
->req_lock
);
949 device
->rs_total
= 0;
950 device
->rs_failed
= 0;
951 device
->rs_paused
= 0;
953 /* reset start sector, if we reached end of device */
954 if (verify_done
&& device
->ov_left
== 0)
955 device
->ov_start_sector
= 0;
957 drbd_md_sync(device
);
960 drbd_khelper(device
, khelper_cmd
);
966 static void move_to_net_ee_or_free(struct drbd_device
*device
, struct drbd_peer_request
*peer_req
)
968 if (drbd_peer_req_has_active_page(peer_req
)) {
969 /* This might happen if sendpage() has not finished */
970 int i
= (peer_req
->i
.size
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
971 atomic_add(i
, &device
->pp_in_use_by_net
);
972 atomic_sub(i
, &device
->pp_in_use
);
973 spin_lock_irq(&device
->connection
->req_lock
);
974 list_add_tail(&peer_req
->w
.list
, &device
->net_ee
);
975 spin_unlock_irq(&device
->connection
->req_lock
);
976 wake_up(&drbd_pp_wait
);
978 drbd_free_peer_req(device
, peer_req
);
982 * w_e_end_data_req() - Worker callback, to send a P_DATA_REPLY packet in response to a P_DATA_REQUEST
983 * @device: DRBD device.
985 * @cancel: The connection will be closed anyways
987 int w_e_end_data_req(struct drbd_work
*w
, int cancel
)
989 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
990 struct drbd_device
*device
= w
->device
;
993 if (unlikely(cancel
)) {
994 drbd_free_peer_req(device
, peer_req
);
999 if (likely((peer_req
->flags
& EE_WAS_ERROR
) == 0)) {
1000 err
= drbd_send_block(device
, P_DATA_REPLY
, peer_req
);
1002 if (__ratelimit(&drbd_ratelimit_state
))
1003 dev_err(DEV
, "Sending NegDReply. sector=%llus.\n",
1004 (unsigned long long)peer_req
->i
.sector
);
1006 err
= drbd_send_ack(device
, P_NEG_DREPLY
, peer_req
);
1009 dec_unacked(device
);
1011 move_to_net_ee_or_free(device
, peer_req
);
1014 dev_err(DEV
, "drbd_send_block() failed\n");
1019 * w_e_end_rsdata_req() - Worker callback to send a P_RS_DATA_REPLY packet in response to a P_RS_DATA_REQUEST
1020 * @device: DRBD device.
1022 * @cancel: The connection will be closed anyways
1024 int w_e_end_rsdata_req(struct drbd_work
*w
, int cancel
)
1026 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
1027 struct drbd_device
*device
= w
->device
;
1030 if (unlikely(cancel
)) {
1031 drbd_free_peer_req(device
, peer_req
);
1032 dec_unacked(device
);
1036 if (get_ldev_if_state(device
, D_FAILED
)) {
1037 drbd_rs_complete_io(device
, peer_req
->i
.sector
);
1041 if (device
->state
.conn
== C_AHEAD
) {
1042 err
= drbd_send_ack(device
, P_RS_CANCEL
, peer_req
);
1043 } else if (likely((peer_req
->flags
& EE_WAS_ERROR
) == 0)) {
1044 if (likely(device
->state
.pdsk
>= D_INCONSISTENT
)) {
1045 inc_rs_pending(device
);
1046 err
= drbd_send_block(device
, P_RS_DATA_REPLY
, peer_req
);
1048 if (__ratelimit(&drbd_ratelimit_state
))
1049 dev_err(DEV
, "Not sending RSDataReply, "
1050 "partner DISKLESS!\n");
1054 if (__ratelimit(&drbd_ratelimit_state
))
1055 dev_err(DEV
, "Sending NegRSDReply. sector %llus.\n",
1056 (unsigned long long)peer_req
->i
.sector
);
1058 err
= drbd_send_ack(device
, P_NEG_RS_DREPLY
, peer_req
);
1060 /* update resync data with failure */
1061 drbd_rs_failed_io(device
, peer_req
->i
.sector
, peer_req
->i
.size
);
1064 dec_unacked(device
);
1066 move_to_net_ee_or_free(device
, peer_req
);
1069 dev_err(DEV
, "drbd_send_block() failed\n");
1073 int w_e_end_csum_rs_req(struct drbd_work
*w
, int cancel
)
1075 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
1076 struct drbd_device
*device
= w
->device
;
1077 struct digest_info
*di
;
1079 void *digest
= NULL
;
1082 if (unlikely(cancel
)) {
1083 drbd_free_peer_req(device
, peer_req
);
1084 dec_unacked(device
);
1088 if (get_ldev(device
)) {
1089 drbd_rs_complete_io(device
, peer_req
->i
.sector
);
1093 di
= peer_req
->digest
;
1095 if (likely((peer_req
->flags
& EE_WAS_ERROR
) == 0)) {
1096 /* quick hack to try to avoid a race against reconfiguration.
1097 * a real fix would be much more involved,
1098 * introducing more locking mechanisms */
1099 if (device
->connection
->csums_tfm
) {
1100 digest_size
= crypto_hash_digestsize(device
->connection
->csums_tfm
);
1101 D_ASSERT(digest_size
== di
->digest_size
);
1102 digest
= kmalloc(digest_size
, GFP_NOIO
);
1105 drbd_csum_ee(device
, device
->connection
->csums_tfm
, peer_req
, digest
);
1106 eq
= !memcmp(digest
, di
->digest
, digest_size
);
1111 drbd_set_in_sync(device
, peer_req
->i
.sector
, peer_req
->i
.size
);
1112 /* rs_same_csums unit is BM_BLOCK_SIZE */
1113 device
->rs_same_csum
+= peer_req
->i
.size
>> BM_BLOCK_SHIFT
;
1114 err
= drbd_send_ack(device
, P_RS_IS_IN_SYNC
, peer_req
);
1116 inc_rs_pending(device
);
1117 peer_req
->block_id
= ID_SYNCER
; /* By setting block_id, digest pointer becomes invalid! */
1118 peer_req
->flags
&= ~EE_HAS_DIGEST
; /* This peer request no longer has a digest pointer */
1120 err
= drbd_send_block(device
, P_RS_DATA_REPLY
, peer_req
);
1123 err
= drbd_send_ack(device
, P_NEG_RS_DREPLY
, peer_req
);
1124 if (__ratelimit(&drbd_ratelimit_state
))
1125 dev_err(DEV
, "Sending NegDReply. I guess it gets messy.\n");
1128 dec_unacked(device
);
1129 move_to_net_ee_or_free(device
, peer_req
);
1132 dev_err(DEV
, "drbd_send_block/ack() failed\n");
1136 int w_e_end_ov_req(struct drbd_work
*w
, int cancel
)
1138 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
1139 struct drbd_device
*device
= w
->device
;
1140 sector_t sector
= peer_req
->i
.sector
;
1141 unsigned int size
= peer_req
->i
.size
;
1146 if (unlikely(cancel
))
1149 digest_size
= crypto_hash_digestsize(device
->connection
->verify_tfm
);
1150 digest
= kmalloc(digest_size
, GFP_NOIO
);
1152 err
= 1; /* terminate the connection in case the allocation failed */
1156 if (likely(!(peer_req
->flags
& EE_WAS_ERROR
)))
1157 drbd_csum_ee(device
, device
->connection
->verify_tfm
, peer_req
, digest
);
1159 memset(digest
, 0, digest_size
);
1161 /* Free e and pages before send.
1162 * In case we block on congestion, we could otherwise run into
1163 * some distributed deadlock, if the other side blocks on
1164 * congestion as well, because our receiver blocks in
1165 * drbd_alloc_pages due to pp_in_use > max_buffers. */
1166 drbd_free_peer_req(device
, peer_req
);
1168 inc_rs_pending(device
);
1169 err
= drbd_send_drequest_csum(device
, sector
, size
, digest
, digest_size
, P_OV_REPLY
);
1171 dec_rs_pending(device
);
1176 drbd_free_peer_req(device
, peer_req
);
1177 dec_unacked(device
);
1181 void drbd_ov_out_of_sync_found(struct drbd_device
*device
, sector_t sector
, int size
)
1183 if (device
->ov_last_oos_start
+ device
->ov_last_oos_size
== sector
) {
1184 device
->ov_last_oos_size
+= size
>>9;
1186 device
->ov_last_oos_start
= sector
;
1187 device
->ov_last_oos_size
= size
>>9;
1189 drbd_set_out_of_sync(device
, sector
, size
);
1192 int w_e_end_ov_reply(struct drbd_work
*w
, int cancel
)
1194 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
1195 struct drbd_device
*device
= w
->device
;
1196 struct digest_info
*di
;
1198 sector_t sector
= peer_req
->i
.sector
;
1199 unsigned int size
= peer_req
->i
.size
;
1202 bool stop_sector_reached
= false;
1204 if (unlikely(cancel
)) {
1205 drbd_free_peer_req(device
, peer_req
);
1206 dec_unacked(device
);
1210 /* after "cancel", because after drbd_disconnect/drbd_rs_cancel_all
1211 * the resync lru has been cleaned up already */
1212 if (get_ldev(device
)) {
1213 drbd_rs_complete_io(device
, peer_req
->i
.sector
);
1217 di
= peer_req
->digest
;
1219 if (likely((peer_req
->flags
& EE_WAS_ERROR
) == 0)) {
1220 digest_size
= crypto_hash_digestsize(device
->connection
->verify_tfm
);
1221 digest
= kmalloc(digest_size
, GFP_NOIO
);
1223 drbd_csum_ee(device
, device
->connection
->verify_tfm
, peer_req
, digest
);
1225 D_ASSERT(digest_size
== di
->digest_size
);
1226 eq
= !memcmp(digest
, di
->digest
, digest_size
);
1231 /* Free peer_req and pages before send.
1232 * In case we block on congestion, we could otherwise run into
1233 * some distributed deadlock, if the other side blocks on
1234 * congestion as well, because our receiver blocks in
1235 * drbd_alloc_pages due to pp_in_use > max_buffers. */
1236 drbd_free_peer_req(device
, peer_req
);
1238 drbd_ov_out_of_sync_found(device
, sector
, size
);
1240 ov_out_of_sync_print(device
);
1242 err
= drbd_send_ack_ex(device
, P_OV_RESULT
, sector
, size
,
1243 eq
? ID_IN_SYNC
: ID_OUT_OF_SYNC
);
1245 dec_unacked(device
);
1249 /* let's advance progress step marks only for every other megabyte */
1250 if ((device
->ov_left
& 0x200) == 0x200)
1251 drbd_advance_rs_marks(device
, device
->ov_left
);
1253 stop_sector_reached
= verify_can_do_stop_sector(device
) &&
1254 (sector
+ (size
>>9)) >= device
->ov_stop_sector
;
1256 if (device
->ov_left
== 0 || stop_sector_reached
) {
1257 ov_out_of_sync_print(device
);
1258 drbd_resync_finished(device
);
1264 int w_prev_work_done(struct drbd_work
*w
, int cancel
)
1266 struct drbd_wq_barrier
*b
= container_of(w
, struct drbd_wq_barrier
, w
);
1273 * We need to track the number of pending barrier acks,
1274 * and to be able to wait for them.
1275 * See also comment in drbd_adm_attach before drbd_suspend_io.
1277 static int drbd_send_barrier(struct drbd_connection
*connection
)
1279 struct p_barrier
*p
;
1280 struct drbd_socket
*sock
;
1282 sock
= &connection
->data
;
1283 p
= conn_prepare_command(connection
, sock
);
1286 p
->barrier
= connection
->send
.current_epoch_nr
;
1288 connection
->send
.current_epoch_writes
= 0;
1290 return conn_send_command(connection
, sock
, P_BARRIER
, sizeof(*p
), NULL
, 0);
1293 int w_send_write_hint(struct drbd_work
*w
, int cancel
)
1295 struct drbd_device
*device
= w
->device
;
1296 struct drbd_socket
*sock
;
1300 sock
= &device
->connection
->data
;
1301 if (!drbd_prepare_command(device
, sock
))
1303 return drbd_send_command(device
, sock
, P_UNPLUG_REMOTE
, 0, NULL
, 0);
1306 static void re_init_if_first_write(struct drbd_connection
*connection
, unsigned int epoch
)
1308 if (!connection
->send
.seen_any_write_yet
) {
1309 connection
->send
.seen_any_write_yet
= true;
1310 connection
->send
.current_epoch_nr
= epoch
;
1311 connection
->send
.current_epoch_writes
= 0;
1315 static void maybe_send_barrier(struct drbd_connection
*connection
, unsigned int epoch
)
1317 /* re-init if first write on this connection */
1318 if (!connection
->send
.seen_any_write_yet
)
1320 if (connection
->send
.current_epoch_nr
!= epoch
) {
1321 if (connection
->send
.current_epoch_writes
)
1322 drbd_send_barrier(connection
);
1323 connection
->send
.current_epoch_nr
= epoch
;
1327 int w_send_out_of_sync(struct drbd_work
*w
, int cancel
)
1329 struct drbd_request
*req
= container_of(w
, struct drbd_request
, w
);
1330 struct drbd_device
*device
= w
->device
;
1331 struct drbd_connection
*connection
= device
->connection
;
1334 if (unlikely(cancel
)) {
1335 req_mod(req
, SEND_CANCELED
);
1339 /* this time, no connection->send.current_epoch_writes++;
1340 * If it was sent, it was the closing barrier for the last
1341 * replicated epoch, before we went into AHEAD mode.
1342 * No more barriers will be sent, until we leave AHEAD mode again. */
1343 maybe_send_barrier(connection
, req
->epoch
);
1345 err
= drbd_send_out_of_sync(device
, req
);
1346 req_mod(req
, OOS_HANDED_TO_NETWORK
);
1352 * w_send_dblock() - Worker callback to send a P_DATA packet in order to mirror a write request
1353 * @device: DRBD device.
1355 * @cancel: The connection will be closed anyways
1357 int w_send_dblock(struct drbd_work
*w
, int cancel
)
1359 struct drbd_request
*req
= container_of(w
, struct drbd_request
, w
);
1360 struct drbd_device
*device
= w
->device
;
1361 struct drbd_connection
*connection
= device
->connection
;
1364 if (unlikely(cancel
)) {
1365 req_mod(req
, SEND_CANCELED
);
1369 re_init_if_first_write(connection
, req
->epoch
);
1370 maybe_send_barrier(connection
, req
->epoch
);
1371 connection
->send
.current_epoch_writes
++;
1373 err
= drbd_send_dblock(device
, req
);
1374 req_mod(req
, err
? SEND_FAILED
: HANDED_OVER_TO_NETWORK
);
1380 * w_send_read_req() - Worker callback to send a read request (P_DATA_REQUEST) packet
1381 * @device: DRBD device.
1383 * @cancel: The connection will be closed anyways
1385 int w_send_read_req(struct drbd_work
*w
, int cancel
)
1387 struct drbd_request
*req
= container_of(w
, struct drbd_request
, w
);
1388 struct drbd_device
*device
= w
->device
;
1389 struct drbd_connection
*connection
= device
->connection
;
1392 if (unlikely(cancel
)) {
1393 req_mod(req
, SEND_CANCELED
);
1397 /* Even read requests may close a write epoch,
1398 * if there was any yet. */
1399 maybe_send_barrier(connection
, req
->epoch
);
1401 err
= drbd_send_drequest(device
, P_DATA_REQUEST
, req
->i
.sector
, req
->i
.size
,
1402 (unsigned long)req
);
1404 req_mod(req
, err
? SEND_FAILED
: HANDED_OVER_TO_NETWORK
);
1409 int w_restart_disk_io(struct drbd_work
*w
, int cancel
)
1411 struct drbd_request
*req
= container_of(w
, struct drbd_request
, w
);
1412 struct drbd_device
*device
= w
->device
;
1414 if (bio_data_dir(req
->master_bio
) == WRITE
&& req
->rq_state
& RQ_IN_ACT_LOG
)
1415 drbd_al_begin_io(device
, &req
->i
, false);
1417 drbd_req_make_private_bio(req
, req
->master_bio
);
1418 req
->private_bio
->bi_bdev
= device
->ldev
->backing_bdev
;
1419 generic_make_request(req
->private_bio
);
1424 static int _drbd_may_sync_now(struct drbd_device
*device
)
1426 struct drbd_device
*odev
= device
;
1430 if (!odev
->ldev
|| odev
->state
.disk
== D_DISKLESS
)
1433 resync_after
= rcu_dereference(odev
->ldev
->disk_conf
)->resync_after
;
1435 if (resync_after
== -1)
1437 odev
= minor_to_device(resync_after
);
1440 if ((odev
->state
.conn
>= C_SYNC_SOURCE
&&
1441 odev
->state
.conn
<= C_PAUSED_SYNC_T
) ||
1442 odev
->state
.aftr_isp
|| odev
->state
.peer_isp
||
1443 odev
->state
.user_isp
)
1449 * _drbd_pause_after() - Pause resync on all devices that may not resync now
1450 * @device: DRBD device.
1452 * Called from process context only (admin command and after_state_ch).
1454 static int _drbd_pause_after(struct drbd_device
*device
)
1456 struct drbd_device
*odev
;
1460 idr_for_each_entry(&minors
, odev
, i
) {
1461 if (odev
->state
.conn
== C_STANDALONE
&& odev
->state
.disk
== D_DISKLESS
)
1463 if (!_drbd_may_sync_now(odev
))
1464 rv
|= (__drbd_set_state(_NS(odev
, aftr_isp
, 1), CS_HARD
, NULL
)
1465 != SS_NOTHING_TO_DO
);
1473 * _drbd_resume_next() - Resume resync on all devices that may resync now
1474 * @device: DRBD device.
1476 * Called from process context only (admin command and worker).
1478 static int _drbd_resume_next(struct drbd_device
*device
)
1480 struct drbd_device
*odev
;
1484 idr_for_each_entry(&minors
, odev
, i
) {
1485 if (odev
->state
.conn
== C_STANDALONE
&& odev
->state
.disk
== D_DISKLESS
)
1487 if (odev
->state
.aftr_isp
) {
1488 if (_drbd_may_sync_now(odev
))
1489 rv
|= (__drbd_set_state(_NS(odev
, aftr_isp
, 0),
1491 != SS_NOTHING_TO_DO
) ;
1498 void resume_next_sg(struct drbd_device
*device
)
1500 write_lock_irq(&global_state_lock
);
1501 _drbd_resume_next(device
);
1502 write_unlock_irq(&global_state_lock
);
1505 void suspend_other_sg(struct drbd_device
*device
)
1507 write_lock_irq(&global_state_lock
);
1508 _drbd_pause_after(device
);
1509 write_unlock_irq(&global_state_lock
);
1512 /* caller must hold global_state_lock */
1513 enum drbd_ret_code
drbd_resync_after_valid(struct drbd_device
*device
, int o_minor
)
1515 struct drbd_device
*odev
;
1520 if (o_minor
< -1 || o_minor
> MINORMASK
)
1521 return ERR_RESYNC_AFTER
;
1523 /* check for loops */
1524 odev
= minor_to_device(o_minor
);
1527 return ERR_RESYNC_AFTER_CYCLE
;
1529 /* You are free to depend on diskless, non-existing,
1530 * or not yet/no longer existing minors.
1531 * We only reject dependency loops.
1532 * We cannot follow the dependency chain beyond a detached or
1535 if (!odev
|| !odev
->ldev
|| odev
->state
.disk
== D_DISKLESS
)
1539 resync_after
= rcu_dereference(odev
->ldev
->disk_conf
)->resync_after
;
1541 /* dependency chain ends here, no cycles. */
1542 if (resync_after
== -1)
1545 /* follow the dependency chain */
1546 odev
= minor_to_device(resync_after
);
1550 /* caller must hold global_state_lock */
1551 void drbd_resync_after_changed(struct drbd_device
*device
)
1556 changes
= _drbd_pause_after(device
);
1557 changes
|= _drbd_resume_next(device
);
1561 void drbd_rs_controller_reset(struct drbd_device
*device
)
1563 struct fifo_buffer
*plan
;
1565 atomic_set(&device
->rs_sect_in
, 0);
1566 atomic_set(&device
->rs_sect_ev
, 0);
1567 device
->rs_in_flight
= 0;
1569 /* Updating the RCU protected object in place is necessary since
1570 this function gets called from atomic context.
1571 It is valid since all other updates also lead to an completely
1574 plan
= rcu_dereference(device
->rs_plan_s
);
1580 void start_resync_timer_fn(unsigned long data
)
1582 struct drbd_device
*device
= (struct drbd_device
*) data
;
1584 drbd_queue_work(&device
->connection
->sender_work
, &device
->start_resync_work
);
1587 int w_start_resync(struct drbd_work
*w
, int cancel
)
1589 struct drbd_device
*device
= w
->device
;
1591 if (atomic_read(&device
->unacked_cnt
) || atomic_read(&device
->rs_pending_cnt
)) {
1592 dev_warn(DEV
, "w_start_resync later...\n");
1593 device
->start_resync_timer
.expires
= jiffies
+ HZ
/10;
1594 add_timer(&device
->start_resync_timer
);
1598 drbd_start_resync(device
, C_SYNC_SOURCE
);
1599 clear_bit(AHEAD_TO_SYNC_SOURCE
, &device
->flags
);
1604 * drbd_start_resync() - Start the resync process
1605 * @device: DRBD device.
1606 * @side: Either C_SYNC_SOURCE or C_SYNC_TARGET
1608 * This function might bring you directly into one of the
1609 * C_PAUSED_SYNC_* states.
1611 void drbd_start_resync(struct drbd_device
*device
, enum drbd_conns side
)
1613 union drbd_state ns
;
1616 if (device
->state
.conn
>= C_SYNC_SOURCE
&& device
->state
.conn
< C_AHEAD
) {
1617 dev_err(DEV
, "Resync already running!\n");
1621 if (!test_bit(B_RS_H_DONE
, &device
->flags
)) {
1622 if (side
== C_SYNC_TARGET
) {
1623 /* Since application IO was locked out during C_WF_BITMAP_T and
1624 C_WF_SYNC_UUID we are still unmodified. Before going to C_SYNC_TARGET
1625 we check that we might make the data inconsistent. */
1626 r
= drbd_khelper(device
, "before-resync-target");
1627 r
= (r
>> 8) & 0xff;
1629 dev_info(DEV
, "before-resync-target handler returned %d, "
1630 "dropping connection.\n", r
);
1631 conn_request_state(device
->connection
, NS(conn
, C_DISCONNECTING
), CS_HARD
);
1634 } else /* C_SYNC_SOURCE */ {
1635 r
= drbd_khelper(device
, "before-resync-source");
1636 r
= (r
>> 8) & 0xff;
1639 dev_info(DEV
, "before-resync-source handler returned %d, "
1640 "ignoring. Old userland tools?", r
);
1642 dev_info(DEV
, "before-resync-source handler returned %d, "
1643 "dropping connection.\n", r
);
1644 conn_request_state(device
->connection
, NS(conn
, C_DISCONNECTING
), CS_HARD
);
1651 if (current
== device
->connection
->worker
.task
) {
1652 /* The worker should not sleep waiting for state_mutex,
1653 that can take long */
1654 if (!mutex_trylock(device
->state_mutex
)) {
1655 set_bit(B_RS_H_DONE
, &device
->flags
);
1656 device
->start_resync_timer
.expires
= jiffies
+ HZ
/5;
1657 add_timer(&device
->start_resync_timer
);
1661 mutex_lock(device
->state_mutex
);
1663 clear_bit(B_RS_H_DONE
, &device
->flags
);
1665 write_lock_irq(&global_state_lock
);
1666 /* Did some connection breakage or IO error race with us? */
1667 if (device
->state
.conn
< C_CONNECTED
1668 || !get_ldev_if_state(device
, D_NEGOTIATING
)) {
1669 write_unlock_irq(&global_state_lock
);
1670 mutex_unlock(device
->state_mutex
);
1674 ns
= drbd_read_state(device
);
1676 ns
.aftr_isp
= !_drbd_may_sync_now(device
);
1680 if (side
== C_SYNC_TARGET
)
1681 ns
.disk
= D_INCONSISTENT
;
1682 else /* side == C_SYNC_SOURCE */
1683 ns
.pdsk
= D_INCONSISTENT
;
1685 r
= __drbd_set_state(device
, ns
, CS_VERBOSE
, NULL
);
1686 ns
= drbd_read_state(device
);
1688 if (ns
.conn
< C_CONNECTED
)
1689 r
= SS_UNKNOWN_ERROR
;
1691 if (r
== SS_SUCCESS
) {
1692 unsigned long tw
= drbd_bm_total_weight(device
);
1693 unsigned long now
= jiffies
;
1696 device
->rs_failed
= 0;
1697 device
->rs_paused
= 0;
1698 device
->rs_same_csum
= 0;
1699 device
->rs_last_events
= 0;
1700 device
->rs_last_sect_ev
= 0;
1701 device
->rs_total
= tw
;
1702 device
->rs_start
= now
;
1703 for (i
= 0; i
< DRBD_SYNC_MARKS
; i
++) {
1704 device
->rs_mark_left
[i
] = tw
;
1705 device
->rs_mark_time
[i
] = now
;
1707 _drbd_pause_after(device
);
1709 write_unlock_irq(&global_state_lock
);
1711 if (r
== SS_SUCCESS
) {
1712 /* reset rs_last_bcast when a resync or verify is started,
1713 * to deal with potential jiffies wrap. */
1714 device
->rs_last_bcast
= jiffies
- HZ
;
1716 dev_info(DEV
, "Began resync as %s (will sync %lu KB [%lu bits set]).\n",
1717 drbd_conn_str(ns
.conn
),
1718 (unsigned long) device
->rs_total
<< (BM_BLOCK_SHIFT
-10),
1719 (unsigned long) device
->rs_total
);
1720 if (side
== C_SYNC_TARGET
)
1721 device
->bm_resync_fo
= 0;
1723 /* Since protocol 96, we must serialize drbd_gen_and_send_sync_uuid
1724 * with w_send_oos, or the sync target will get confused as to
1725 * how much bits to resync. We cannot do that always, because for an
1726 * empty resync and protocol < 95, we need to do it here, as we call
1727 * drbd_resync_finished from here in that case.
1728 * We drbd_gen_and_send_sync_uuid here for protocol < 96,
1729 * and from after_state_ch otherwise. */
1730 if (side
== C_SYNC_SOURCE
&& device
->connection
->agreed_pro_version
< 96)
1731 drbd_gen_and_send_sync_uuid(device
);
1733 if (device
->connection
->agreed_pro_version
< 95 && device
->rs_total
== 0) {
1734 /* This still has a race (about when exactly the peers
1735 * detect connection loss) that can lead to a full sync
1736 * on next handshake. In 8.3.9 we fixed this with explicit
1737 * resync-finished notifications, but the fix
1738 * introduces a protocol change. Sleeping for some
1739 * time longer than the ping interval + timeout on the
1740 * SyncSource, to give the SyncTarget the chance to
1741 * detect connection loss, then waiting for a ping
1742 * response (implicit in drbd_resync_finished) reduces
1743 * the race considerably, but does not solve it. */
1744 if (side
== C_SYNC_SOURCE
) {
1745 struct net_conf
*nc
;
1749 nc
= rcu_dereference(device
->connection
->net_conf
);
1750 timeo
= nc
->ping_int
* HZ
+ nc
->ping_timeo
* HZ
/ 9;
1752 schedule_timeout_interruptible(timeo
);
1754 drbd_resync_finished(device
);
1757 drbd_rs_controller_reset(device
);
1758 /* ns.conn may already be != device->state.conn,
1759 * we may have been paused in between, or become paused until
1760 * the timer triggers.
1761 * No matter, that is handled in resync_timer_fn() */
1762 if (ns
.conn
== C_SYNC_TARGET
)
1763 mod_timer(&device
->resync_timer
, jiffies
);
1765 drbd_md_sync(device
);
1768 mutex_unlock(device
->state_mutex
);
1771 /* If the resource already closed the current epoch, but we did not
1772 * (because we have not yet seen new requests), we should send the
1773 * corresponding barrier now. Must be checked within the same spinlock
1774 * that is used to check for new requests. */
1775 static bool need_to_send_barrier(struct drbd_connection
*connection
)
1777 if (!connection
->send
.seen_any_write_yet
)
1780 /* Skip barriers that do not contain any writes.
1781 * This may happen during AHEAD mode. */
1782 if (!connection
->send
.current_epoch_writes
)
1785 /* ->req_lock is held when requests are queued on
1786 * connection->sender_work, and put into ->transfer_log.
1787 * It is also held when ->current_tle_nr is increased.
1788 * So either there are already new requests queued,
1789 * and corresponding barriers will be send there.
1790 * Or nothing new is queued yet, so the difference will be 1.
1792 if (atomic_read(&connection
->current_tle_nr
) !=
1793 connection
->send
.current_epoch_nr
+ 1)
1799 static bool dequeue_work_batch(struct drbd_work_queue
*queue
, struct list_head
*work_list
)
1801 spin_lock_irq(&queue
->q_lock
);
1802 list_splice_init(&queue
->q
, work_list
);
1803 spin_unlock_irq(&queue
->q_lock
);
1804 return !list_empty(work_list
);
1807 static bool dequeue_work_item(struct drbd_work_queue
*queue
, struct list_head
*work_list
)
1809 spin_lock_irq(&queue
->q_lock
);
1810 if (!list_empty(&queue
->q
))
1811 list_move(queue
->q
.next
, work_list
);
1812 spin_unlock_irq(&queue
->q_lock
);
1813 return !list_empty(work_list
);
1816 static void wait_for_work(struct drbd_connection
*connection
, struct list_head
*work_list
)
1819 struct net_conf
*nc
;
1822 dequeue_work_item(&connection
->sender_work
, work_list
);
1823 if (!list_empty(work_list
))
1826 /* Still nothing to do?
1827 * Maybe we still need to close the current epoch,
1828 * even if no new requests are queued yet.
1830 * Also, poke TCP, just in case.
1831 * Then wait for new work (or signal). */
1833 nc
= rcu_dereference(connection
->net_conf
);
1834 uncork
= nc
? nc
->tcp_cork
: 0;
1837 mutex_lock(&connection
->data
.mutex
);
1838 if (connection
->data
.socket
)
1839 drbd_tcp_uncork(connection
->data
.socket
);
1840 mutex_unlock(&connection
->data
.mutex
);
1845 prepare_to_wait(&connection
->sender_work
.q_wait
, &wait
, TASK_INTERRUPTIBLE
);
1846 spin_lock_irq(&connection
->req_lock
);
1847 spin_lock(&connection
->sender_work
.q_lock
); /* FIXME get rid of this one? */
1848 /* dequeue single item only,
1849 * we still use drbd_queue_work_front() in some places */
1850 if (!list_empty(&connection
->sender_work
.q
))
1851 list_move(connection
->sender_work
.q
.next
, work_list
);
1852 spin_unlock(&connection
->sender_work
.q_lock
); /* FIXME get rid of this one? */
1853 if (!list_empty(work_list
) || signal_pending(current
)) {
1854 spin_unlock_irq(&connection
->req_lock
);
1857 send_barrier
= need_to_send_barrier(connection
);
1858 spin_unlock_irq(&connection
->req_lock
);
1860 drbd_send_barrier(connection
);
1861 connection
->send
.current_epoch_nr
++;
1864 /* may be woken up for other things but new work, too,
1865 * e.g. if the current epoch got closed.
1866 * In which case we send the barrier above. */
1868 finish_wait(&connection
->sender_work
.q_wait
, &wait
);
1870 /* someone may have changed the config while we have been waiting above. */
1872 nc
= rcu_dereference(connection
->net_conf
);
1873 cork
= nc
? nc
->tcp_cork
: 0;
1875 mutex_lock(&connection
->data
.mutex
);
1876 if (connection
->data
.socket
) {
1878 drbd_tcp_cork(connection
->data
.socket
);
1880 drbd_tcp_uncork(connection
->data
.socket
);
1882 mutex_unlock(&connection
->data
.mutex
);
1885 int drbd_worker(struct drbd_thread
*thi
)
1887 struct drbd_connection
*connection
= thi
->connection
;
1888 struct drbd_work
*w
= NULL
;
1889 struct drbd_device
*device
;
1890 LIST_HEAD(work_list
);
1893 while (get_t_state(thi
) == RUNNING
) {
1894 drbd_thread_current_set_cpu(thi
);
1896 /* as long as we use drbd_queue_work_front(),
1897 * we may only dequeue single work items here, not batches. */
1898 if (list_empty(&work_list
))
1899 wait_for_work(connection
, &work_list
);
1901 if (signal_pending(current
)) {
1902 flush_signals(current
);
1903 if (get_t_state(thi
) == RUNNING
) {
1904 conn_warn(connection
, "Worker got an unexpected signal\n");
1910 if (get_t_state(thi
) != RUNNING
)
1913 while (!list_empty(&work_list
)) {
1914 w
= list_first_entry(&work_list
, struct drbd_work
, list
);
1915 list_del_init(&w
->list
);
1916 if (w
->cb(w
, connection
->cstate
< C_WF_REPORT_PARAMS
) == 0)
1918 if (connection
->cstate
>= C_WF_REPORT_PARAMS
)
1919 conn_request_state(connection
, NS(conn
, C_NETWORK_FAILURE
), CS_HARD
);
1924 while (!list_empty(&work_list
)) {
1925 w
= list_first_entry(&work_list
, struct drbd_work
, list
);
1926 list_del_init(&w
->list
);
1929 dequeue_work_batch(&connection
->sender_work
, &work_list
);
1930 } while (!list_empty(&work_list
));
1933 idr_for_each_entry(&connection
->volumes
, device
, vnr
) {
1934 D_ASSERT(device
->state
.disk
== D_DISKLESS
&& device
->state
.conn
== C_STANDALONE
);
1935 kref_get(&device
->kref
);
1937 drbd_device_cleanup(device
);
1938 kref_put(&device
->kref
, &drbd_minor_destroy
);