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(&first_peer_device(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(&first_peer_device(device
)->connection
->req_lock
, flags
);
114 drbd_queue_work(&first_peer_device(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(&first_peer_device(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(&first_peer_device(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(first_peer_device(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(&first_peer_device(device
)->connection
->req_lock
, flags
);
277 __req_mod(req
, what
, &m
);
278 spin_unlock_irqrestore(&first_peer_device(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(first_peer_device(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
, first_peer_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(&first_peer_device(device
)->connection
->req_lock
);
401 list_add(&peer_req
->w
.list
, &device
->read_ee
);
402 spin_unlock_irq(&first_peer_device(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(&first_peer_device(device
)->connection
->req_lock
);
413 list_del(&peer_req
->w
.list
);
414 spin_unlock_irq(&first_peer_device(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(&first_peer_device(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(&first_peer_device(device
)->connection
->data
.mutex
);
601 if (first_peer_device(device
)->connection
->data
.socket
) {
602 queued
= first_peer_device(device
)->connection
->data
.socket
->sk
->sk_wmem_queued
;
603 sndbuf
= first_peer_device(device
)->connection
->data
.socket
->sk
->sk_sndbuf
;
608 mutex_unlock(&first_peer_device(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 (first_peer_device(device
)->connection
->agreed_pro_version
>= 89 &&
679 first_peer_device(device
)->connection
->csums_tfm
) {
680 switch (read_for_csum(device
, sector
, size
)) {
681 case -EIO
: /* Disk failure */
684 case -EAGAIN
: /* allocation failed, or ldev busy */
685 drbd_rs_complete_io(device
, sector
);
686 device
->bm_resync_fo
= BM_SECT_TO_BIT(sector
);
698 inc_rs_pending(device
);
699 err
= drbd_send_drequest(device
, P_RS_DATA_REQUEST
,
700 sector
, size
, ID_SYNCER
);
702 dev_err(DEV
, "drbd_send_drequest() failed, aborting...\n");
703 dec_rs_pending(device
);
710 if (device
->bm_resync_fo
>= drbd_bm_bits(device
)) {
711 /* last syncer _request_ was sent,
712 * but the P_RS_DATA_REPLY not yet received. sync will end (and
713 * next sync group will resume), as soon as we receive the last
714 * resync data block, and the last bit is cleared.
715 * until then resync "work" is "inactive" ...
722 device
->rs_in_flight
+= (i
<< (BM_BLOCK_SHIFT
- 9));
723 mod_timer(&device
->resync_timer
, jiffies
+ SLEEP_TIME
);
728 static int w_make_ov_request(struct drbd_work
*w
, int cancel
)
730 struct drbd_device
*device
= w
->device
;
733 const sector_t capacity
= drbd_get_capacity(device
->this_bdev
);
734 bool stop_sector_reached
= false;
736 if (unlikely(cancel
))
739 number
= drbd_rs_number_requests(device
);
741 sector
= device
->ov_position
;
742 for (i
= 0; i
< number
; i
++) {
743 if (sector
>= capacity
)
746 /* We check for "finished" only in the reply path:
747 * w_e_end_ov_reply().
748 * We need to send at least one request out. */
749 stop_sector_reached
= i
> 0
750 && verify_can_do_stop_sector(device
)
751 && sector
>= device
->ov_stop_sector
;
752 if (stop_sector_reached
)
755 size
= BM_BLOCK_SIZE
;
757 if (drbd_rs_should_slow_down(device
, sector
) ||
758 drbd_try_rs_begin_io(device
, sector
)) {
759 device
->ov_position
= sector
;
763 if (sector
+ (size
>>9) > capacity
)
764 size
= (capacity
-sector
)<<9;
766 inc_rs_pending(device
);
767 if (drbd_send_ov_request(device
, sector
, size
)) {
768 dec_rs_pending(device
);
771 sector
+= BM_SECT_PER_BIT
;
773 device
->ov_position
= sector
;
776 device
->rs_in_flight
+= (i
<< (BM_BLOCK_SHIFT
- 9));
777 if (i
== 0 || !stop_sector_reached
)
778 mod_timer(&device
->resync_timer
, jiffies
+ SLEEP_TIME
);
782 int w_ov_finished(struct drbd_work
*w
, int cancel
)
784 struct drbd_device
*device
= w
->device
;
786 ov_out_of_sync_print(device
);
787 drbd_resync_finished(device
);
792 static int w_resync_finished(struct drbd_work
*w
, int cancel
)
794 struct drbd_device
*device
= w
->device
;
797 drbd_resync_finished(device
);
802 static void ping_peer(struct drbd_device
*device
)
804 struct drbd_connection
*connection
= first_peer_device(device
)->connection
;
806 clear_bit(GOT_PING_ACK
, &connection
->flags
);
807 request_ping(connection
);
808 wait_event(connection
->ping_wait
,
809 test_bit(GOT_PING_ACK
, &connection
->flags
) || device
->state
.conn
< C_CONNECTED
);
812 int drbd_resync_finished(struct drbd_device
*device
)
814 unsigned long db
, dt
, dbdt
;
816 union drbd_state os
, ns
;
818 char *khelper_cmd
= NULL
;
821 /* Remove all elements from the resync LRU. Since future actions
822 * might set bits in the (main) bitmap, then the entries in the
823 * resync LRU would be wrong. */
824 if (drbd_rs_del_all(device
)) {
825 /* In case this is not possible now, most probably because
826 * there are P_RS_DATA_REPLY Packets lingering on the worker's
827 * queue (or even the read operations for those packets
828 * is not finished by now). Retry in 100ms. */
830 schedule_timeout_interruptible(HZ
/ 10);
831 w
= kmalloc(sizeof(struct drbd_work
), GFP_ATOMIC
);
833 w
->cb
= w_resync_finished
;
835 drbd_queue_work(&first_peer_device(device
)->connection
->sender_work
, w
);
838 dev_err(DEV
, "Warn failed to drbd_rs_del_all() and to kmalloc(w).\n");
841 dt
= (jiffies
- device
->rs_start
- device
->rs_paused
) / HZ
;
845 db
= device
->rs_total
;
846 /* adjust for verify start and stop sectors, respective reached position */
847 if (device
->state
.conn
== C_VERIFY_S
|| device
->state
.conn
== C_VERIFY_T
)
848 db
-= device
->ov_left
;
850 dbdt
= Bit2KB(db
/dt
);
851 device
->rs_paused
/= HZ
;
853 if (!get_ldev(device
))
858 spin_lock_irq(&first_peer_device(device
)->connection
->req_lock
);
859 os
= drbd_read_state(device
);
861 verify_done
= (os
.conn
== C_VERIFY_S
|| os
.conn
== C_VERIFY_T
);
863 /* This protects us against multiple calls (that can happen in the presence
864 of application IO), and against connectivity loss just before we arrive here. */
865 if (os
.conn
<= C_CONNECTED
)
869 ns
.conn
= C_CONNECTED
;
871 dev_info(DEV
, "%s done (total %lu sec; paused %lu sec; %lu K/sec)\n",
872 verify_done
? "Online verify" : "Resync",
873 dt
+ device
->rs_paused
, device
->rs_paused
, dbdt
);
875 n_oos
= drbd_bm_total_weight(device
);
877 if (os
.conn
== C_VERIFY_S
|| os
.conn
== C_VERIFY_T
) {
879 dev_alert(DEV
, "Online verify found %lu %dk block out of sync!\n",
881 khelper_cmd
= "out-of-sync";
884 D_ASSERT((n_oos
- device
->rs_failed
) == 0);
886 if (os
.conn
== C_SYNC_TARGET
|| os
.conn
== C_PAUSED_SYNC_T
)
887 khelper_cmd
= "after-resync-target";
889 if (first_peer_device(device
)->connection
->csums_tfm
&& device
->rs_total
) {
890 const unsigned long s
= device
->rs_same_csum
;
891 const unsigned long t
= device
->rs_total
;
894 (t
< 100000) ? ((s
*100)/t
) : (s
/(t
/100));
895 dev_info(DEV
, "%u %% had equal checksums, eliminated: %luK; "
896 "transferred %luK total %luK\n",
898 Bit2KB(device
->rs_same_csum
),
899 Bit2KB(device
->rs_total
- device
->rs_same_csum
),
900 Bit2KB(device
->rs_total
));
904 if (device
->rs_failed
) {
905 dev_info(DEV
, " %lu failed blocks\n", device
->rs_failed
);
907 if (os
.conn
== C_SYNC_TARGET
|| os
.conn
== C_PAUSED_SYNC_T
) {
908 ns
.disk
= D_INCONSISTENT
;
909 ns
.pdsk
= D_UP_TO_DATE
;
911 ns
.disk
= D_UP_TO_DATE
;
912 ns
.pdsk
= D_INCONSISTENT
;
915 ns
.disk
= D_UP_TO_DATE
;
916 ns
.pdsk
= D_UP_TO_DATE
;
918 if (os
.conn
== C_SYNC_TARGET
|| os
.conn
== C_PAUSED_SYNC_T
) {
919 if (device
->p_uuid
) {
921 for (i
= UI_BITMAP
; i
<= UI_HISTORY_END
; i
++)
922 _drbd_uuid_set(device
, i
, device
->p_uuid
[i
]);
923 drbd_uuid_set(device
, UI_BITMAP
, device
->ldev
->md
.uuid
[UI_CURRENT
]);
924 _drbd_uuid_set(device
, UI_CURRENT
, device
->p_uuid
[UI_CURRENT
]);
926 dev_err(DEV
, "device->p_uuid is NULL! BUG\n");
930 if (!(os
.conn
== C_VERIFY_S
|| os
.conn
== C_VERIFY_T
)) {
931 /* for verify runs, we don't update uuids here,
932 * so there would be nothing to report. */
933 drbd_uuid_set_bm(device
, 0UL);
934 drbd_print_uuids(device
, "updated UUIDs");
935 if (device
->p_uuid
) {
936 /* Now the two UUID sets are equal, update what we
937 * know of the peer. */
939 for (i
= UI_CURRENT
; i
<= UI_HISTORY_END
; i
++)
940 device
->p_uuid
[i
] = device
->ldev
->md
.uuid
[i
];
945 _drbd_set_state(device
, ns
, CS_VERBOSE
, NULL
);
947 spin_unlock_irq(&first_peer_device(device
)->connection
->req_lock
);
950 device
->rs_total
= 0;
951 device
->rs_failed
= 0;
952 device
->rs_paused
= 0;
954 /* reset start sector, if we reached end of device */
955 if (verify_done
&& device
->ov_left
== 0)
956 device
->ov_start_sector
= 0;
958 drbd_md_sync(device
);
961 drbd_khelper(device
, khelper_cmd
);
967 static void move_to_net_ee_or_free(struct drbd_device
*device
, struct drbd_peer_request
*peer_req
)
969 if (drbd_peer_req_has_active_page(peer_req
)) {
970 /* This might happen if sendpage() has not finished */
971 int i
= (peer_req
->i
.size
+ PAGE_SIZE
-1) >> PAGE_SHIFT
;
972 atomic_add(i
, &device
->pp_in_use_by_net
);
973 atomic_sub(i
, &device
->pp_in_use
);
974 spin_lock_irq(&first_peer_device(device
)->connection
->req_lock
);
975 list_add_tail(&peer_req
->w
.list
, &device
->net_ee
);
976 spin_unlock_irq(&first_peer_device(device
)->connection
->req_lock
);
977 wake_up(&drbd_pp_wait
);
979 drbd_free_peer_req(device
, peer_req
);
983 * w_e_end_data_req() - Worker callback, to send a P_DATA_REPLY packet in response to a P_DATA_REQUEST
984 * @device: DRBD device.
986 * @cancel: The connection will be closed anyways
988 int w_e_end_data_req(struct drbd_work
*w
, int cancel
)
990 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
991 struct drbd_device
*device
= w
->device
;
994 if (unlikely(cancel
)) {
995 drbd_free_peer_req(device
, peer_req
);
1000 if (likely((peer_req
->flags
& EE_WAS_ERROR
) == 0)) {
1001 err
= drbd_send_block(device
, P_DATA_REPLY
, peer_req
);
1003 if (__ratelimit(&drbd_ratelimit_state
))
1004 dev_err(DEV
, "Sending NegDReply. sector=%llus.\n",
1005 (unsigned long long)peer_req
->i
.sector
);
1007 err
= drbd_send_ack(device
, P_NEG_DREPLY
, peer_req
);
1010 dec_unacked(device
);
1012 move_to_net_ee_or_free(device
, peer_req
);
1015 dev_err(DEV
, "drbd_send_block() failed\n");
1020 * w_e_end_rsdata_req() - Worker callback to send a P_RS_DATA_REPLY packet in response to a P_RS_DATA_REQUEST
1021 * @device: DRBD device.
1023 * @cancel: The connection will be closed anyways
1025 int w_e_end_rsdata_req(struct drbd_work
*w
, int cancel
)
1027 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
1028 struct drbd_device
*device
= w
->device
;
1031 if (unlikely(cancel
)) {
1032 drbd_free_peer_req(device
, peer_req
);
1033 dec_unacked(device
);
1037 if (get_ldev_if_state(device
, D_FAILED
)) {
1038 drbd_rs_complete_io(device
, peer_req
->i
.sector
);
1042 if (device
->state
.conn
== C_AHEAD
) {
1043 err
= drbd_send_ack(device
, P_RS_CANCEL
, peer_req
);
1044 } else if (likely((peer_req
->flags
& EE_WAS_ERROR
) == 0)) {
1045 if (likely(device
->state
.pdsk
>= D_INCONSISTENT
)) {
1046 inc_rs_pending(device
);
1047 err
= drbd_send_block(device
, P_RS_DATA_REPLY
, peer_req
);
1049 if (__ratelimit(&drbd_ratelimit_state
))
1050 dev_err(DEV
, "Not sending RSDataReply, "
1051 "partner DISKLESS!\n");
1055 if (__ratelimit(&drbd_ratelimit_state
))
1056 dev_err(DEV
, "Sending NegRSDReply. sector %llus.\n",
1057 (unsigned long long)peer_req
->i
.sector
);
1059 err
= drbd_send_ack(device
, P_NEG_RS_DREPLY
, peer_req
);
1061 /* update resync data with failure */
1062 drbd_rs_failed_io(device
, peer_req
->i
.sector
, peer_req
->i
.size
);
1065 dec_unacked(device
);
1067 move_to_net_ee_or_free(device
, peer_req
);
1070 dev_err(DEV
, "drbd_send_block() failed\n");
1074 int w_e_end_csum_rs_req(struct drbd_work
*w
, int cancel
)
1076 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
1077 struct drbd_device
*device
= w
->device
;
1078 struct digest_info
*di
;
1080 void *digest
= NULL
;
1083 if (unlikely(cancel
)) {
1084 drbd_free_peer_req(device
, peer_req
);
1085 dec_unacked(device
);
1089 if (get_ldev(device
)) {
1090 drbd_rs_complete_io(device
, peer_req
->i
.sector
);
1094 di
= peer_req
->digest
;
1096 if (likely((peer_req
->flags
& EE_WAS_ERROR
) == 0)) {
1097 /* quick hack to try to avoid a race against reconfiguration.
1098 * a real fix would be much more involved,
1099 * introducing more locking mechanisms */
1100 if (first_peer_device(device
)->connection
->csums_tfm
) {
1101 digest_size
= crypto_hash_digestsize(first_peer_device(device
)->connection
->csums_tfm
);
1102 D_ASSERT(digest_size
== di
->digest_size
);
1103 digest
= kmalloc(digest_size
, GFP_NOIO
);
1106 drbd_csum_ee(device
, first_peer_device(device
)->connection
->csums_tfm
, peer_req
, digest
);
1107 eq
= !memcmp(digest
, di
->digest
, digest_size
);
1112 drbd_set_in_sync(device
, peer_req
->i
.sector
, peer_req
->i
.size
);
1113 /* rs_same_csums unit is BM_BLOCK_SIZE */
1114 device
->rs_same_csum
+= peer_req
->i
.size
>> BM_BLOCK_SHIFT
;
1115 err
= drbd_send_ack(device
, P_RS_IS_IN_SYNC
, peer_req
);
1117 inc_rs_pending(device
);
1118 peer_req
->block_id
= ID_SYNCER
; /* By setting block_id, digest pointer becomes invalid! */
1119 peer_req
->flags
&= ~EE_HAS_DIGEST
; /* This peer request no longer has a digest pointer */
1121 err
= drbd_send_block(device
, P_RS_DATA_REPLY
, peer_req
);
1124 err
= drbd_send_ack(device
, P_NEG_RS_DREPLY
, peer_req
);
1125 if (__ratelimit(&drbd_ratelimit_state
))
1126 dev_err(DEV
, "Sending NegDReply. I guess it gets messy.\n");
1129 dec_unacked(device
);
1130 move_to_net_ee_or_free(device
, peer_req
);
1133 dev_err(DEV
, "drbd_send_block/ack() failed\n");
1137 int w_e_end_ov_req(struct drbd_work
*w
, int cancel
)
1139 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
1140 struct drbd_device
*device
= w
->device
;
1141 sector_t sector
= peer_req
->i
.sector
;
1142 unsigned int size
= peer_req
->i
.size
;
1147 if (unlikely(cancel
))
1150 digest_size
= crypto_hash_digestsize(first_peer_device(device
)->connection
->verify_tfm
);
1151 digest
= kmalloc(digest_size
, GFP_NOIO
);
1153 err
= 1; /* terminate the connection in case the allocation failed */
1157 if (likely(!(peer_req
->flags
& EE_WAS_ERROR
)))
1158 drbd_csum_ee(device
, first_peer_device(device
)->connection
->verify_tfm
, peer_req
, digest
);
1160 memset(digest
, 0, digest_size
);
1162 /* Free e and pages before send.
1163 * In case we block on congestion, we could otherwise run into
1164 * some distributed deadlock, if the other side blocks on
1165 * congestion as well, because our receiver blocks in
1166 * drbd_alloc_pages due to pp_in_use > max_buffers. */
1167 drbd_free_peer_req(device
, peer_req
);
1169 inc_rs_pending(device
);
1170 err
= drbd_send_drequest_csum(device
, sector
, size
, digest
, digest_size
, P_OV_REPLY
);
1172 dec_rs_pending(device
);
1177 drbd_free_peer_req(device
, peer_req
);
1178 dec_unacked(device
);
1182 void drbd_ov_out_of_sync_found(struct drbd_device
*device
, sector_t sector
, int size
)
1184 if (device
->ov_last_oos_start
+ device
->ov_last_oos_size
== sector
) {
1185 device
->ov_last_oos_size
+= size
>>9;
1187 device
->ov_last_oos_start
= sector
;
1188 device
->ov_last_oos_size
= size
>>9;
1190 drbd_set_out_of_sync(device
, sector
, size
);
1193 int w_e_end_ov_reply(struct drbd_work
*w
, int cancel
)
1195 struct drbd_peer_request
*peer_req
= container_of(w
, struct drbd_peer_request
, w
);
1196 struct drbd_device
*device
= w
->device
;
1197 struct digest_info
*di
;
1199 sector_t sector
= peer_req
->i
.sector
;
1200 unsigned int size
= peer_req
->i
.size
;
1203 bool stop_sector_reached
= false;
1205 if (unlikely(cancel
)) {
1206 drbd_free_peer_req(device
, peer_req
);
1207 dec_unacked(device
);
1211 /* after "cancel", because after drbd_disconnect/drbd_rs_cancel_all
1212 * the resync lru has been cleaned up already */
1213 if (get_ldev(device
)) {
1214 drbd_rs_complete_io(device
, peer_req
->i
.sector
);
1218 di
= peer_req
->digest
;
1220 if (likely((peer_req
->flags
& EE_WAS_ERROR
) == 0)) {
1221 digest_size
= crypto_hash_digestsize(first_peer_device(device
)->connection
->verify_tfm
);
1222 digest
= kmalloc(digest_size
, GFP_NOIO
);
1224 drbd_csum_ee(device
, first_peer_device(device
)->connection
->verify_tfm
, peer_req
, digest
);
1226 D_ASSERT(digest_size
== di
->digest_size
);
1227 eq
= !memcmp(digest
, di
->digest
, digest_size
);
1232 /* Free peer_req and pages before send.
1233 * In case we block on congestion, we could otherwise run into
1234 * some distributed deadlock, if the other side blocks on
1235 * congestion as well, because our receiver blocks in
1236 * drbd_alloc_pages due to pp_in_use > max_buffers. */
1237 drbd_free_peer_req(device
, peer_req
);
1239 drbd_ov_out_of_sync_found(device
, sector
, size
);
1241 ov_out_of_sync_print(device
);
1243 err
= drbd_send_ack_ex(device
, P_OV_RESULT
, sector
, size
,
1244 eq
? ID_IN_SYNC
: ID_OUT_OF_SYNC
);
1246 dec_unacked(device
);
1250 /* let's advance progress step marks only for every other megabyte */
1251 if ((device
->ov_left
& 0x200) == 0x200)
1252 drbd_advance_rs_marks(device
, device
->ov_left
);
1254 stop_sector_reached
= verify_can_do_stop_sector(device
) &&
1255 (sector
+ (size
>>9)) >= device
->ov_stop_sector
;
1257 if (device
->ov_left
== 0 || stop_sector_reached
) {
1258 ov_out_of_sync_print(device
);
1259 drbd_resync_finished(device
);
1265 int w_prev_work_done(struct drbd_work
*w
, int cancel
)
1267 struct drbd_wq_barrier
*b
= container_of(w
, struct drbd_wq_barrier
, w
);
1274 * We need to track the number of pending barrier acks,
1275 * and to be able to wait for them.
1276 * See also comment in drbd_adm_attach before drbd_suspend_io.
1278 static int drbd_send_barrier(struct drbd_connection
*connection
)
1280 struct p_barrier
*p
;
1281 struct drbd_socket
*sock
;
1283 sock
= &connection
->data
;
1284 p
= conn_prepare_command(connection
, sock
);
1287 p
->barrier
= connection
->send
.current_epoch_nr
;
1289 connection
->send
.current_epoch_writes
= 0;
1291 return conn_send_command(connection
, sock
, P_BARRIER
, sizeof(*p
), NULL
, 0);
1294 int w_send_write_hint(struct drbd_work
*w
, int cancel
)
1296 struct drbd_device
*device
= w
->device
;
1297 struct drbd_socket
*sock
;
1301 sock
= &first_peer_device(device
)->connection
->data
;
1302 if (!drbd_prepare_command(device
, sock
))
1304 return drbd_send_command(device
, sock
, P_UNPLUG_REMOTE
, 0, NULL
, 0);
1307 static void re_init_if_first_write(struct drbd_connection
*connection
, unsigned int epoch
)
1309 if (!connection
->send
.seen_any_write_yet
) {
1310 connection
->send
.seen_any_write_yet
= true;
1311 connection
->send
.current_epoch_nr
= epoch
;
1312 connection
->send
.current_epoch_writes
= 0;
1316 static void maybe_send_barrier(struct drbd_connection
*connection
, unsigned int epoch
)
1318 /* re-init if first write on this connection */
1319 if (!connection
->send
.seen_any_write_yet
)
1321 if (connection
->send
.current_epoch_nr
!= epoch
) {
1322 if (connection
->send
.current_epoch_writes
)
1323 drbd_send_barrier(connection
);
1324 connection
->send
.current_epoch_nr
= epoch
;
1328 int w_send_out_of_sync(struct drbd_work
*w
, int cancel
)
1330 struct drbd_request
*req
= container_of(w
, struct drbd_request
, w
);
1331 struct drbd_device
*device
= w
->device
;
1332 struct drbd_connection
*connection
= first_peer_device(device
)->connection
;
1335 if (unlikely(cancel
)) {
1336 req_mod(req
, SEND_CANCELED
);
1340 /* this time, no connection->send.current_epoch_writes++;
1341 * If it was sent, it was the closing barrier for the last
1342 * replicated epoch, before we went into AHEAD mode.
1343 * No more barriers will be sent, until we leave AHEAD mode again. */
1344 maybe_send_barrier(connection
, req
->epoch
);
1346 err
= drbd_send_out_of_sync(device
, req
);
1347 req_mod(req
, OOS_HANDED_TO_NETWORK
);
1353 * w_send_dblock() - Worker callback to send a P_DATA packet in order to mirror a write request
1354 * @device: DRBD device.
1356 * @cancel: The connection will be closed anyways
1358 int w_send_dblock(struct drbd_work
*w
, int cancel
)
1360 struct drbd_request
*req
= container_of(w
, struct drbd_request
, w
);
1361 struct drbd_device
*device
= w
->device
;
1362 struct drbd_connection
*connection
= first_peer_device(device
)->connection
;
1365 if (unlikely(cancel
)) {
1366 req_mod(req
, SEND_CANCELED
);
1370 re_init_if_first_write(connection
, req
->epoch
);
1371 maybe_send_barrier(connection
, req
->epoch
);
1372 connection
->send
.current_epoch_writes
++;
1374 err
= drbd_send_dblock(device
, req
);
1375 req_mod(req
, err
? SEND_FAILED
: HANDED_OVER_TO_NETWORK
);
1381 * w_send_read_req() - Worker callback to send a read request (P_DATA_REQUEST) packet
1382 * @device: DRBD device.
1384 * @cancel: The connection will be closed anyways
1386 int w_send_read_req(struct drbd_work
*w
, int cancel
)
1388 struct drbd_request
*req
= container_of(w
, struct drbd_request
, w
);
1389 struct drbd_device
*device
= w
->device
;
1390 struct drbd_connection
*connection
= first_peer_device(device
)->connection
;
1393 if (unlikely(cancel
)) {
1394 req_mod(req
, SEND_CANCELED
);
1398 /* Even read requests may close a write epoch,
1399 * if there was any yet. */
1400 maybe_send_barrier(connection
, req
->epoch
);
1402 err
= drbd_send_drequest(device
, P_DATA_REQUEST
, req
->i
.sector
, req
->i
.size
,
1403 (unsigned long)req
);
1405 req_mod(req
, err
? SEND_FAILED
: HANDED_OVER_TO_NETWORK
);
1410 int w_restart_disk_io(struct drbd_work
*w
, int cancel
)
1412 struct drbd_request
*req
= container_of(w
, struct drbd_request
, w
);
1413 struct drbd_device
*device
= w
->device
;
1415 if (bio_data_dir(req
->master_bio
) == WRITE
&& req
->rq_state
& RQ_IN_ACT_LOG
)
1416 drbd_al_begin_io(device
, &req
->i
, false);
1418 drbd_req_make_private_bio(req
, req
->master_bio
);
1419 req
->private_bio
->bi_bdev
= device
->ldev
->backing_bdev
;
1420 generic_make_request(req
->private_bio
);
1425 static int _drbd_may_sync_now(struct drbd_device
*device
)
1427 struct drbd_device
*odev
= device
;
1431 if (!odev
->ldev
|| odev
->state
.disk
== D_DISKLESS
)
1434 resync_after
= rcu_dereference(odev
->ldev
->disk_conf
)->resync_after
;
1436 if (resync_after
== -1)
1438 odev
= minor_to_device(resync_after
);
1441 if ((odev
->state
.conn
>= C_SYNC_SOURCE
&&
1442 odev
->state
.conn
<= C_PAUSED_SYNC_T
) ||
1443 odev
->state
.aftr_isp
|| odev
->state
.peer_isp
||
1444 odev
->state
.user_isp
)
1450 * _drbd_pause_after() - Pause resync on all devices that may not resync now
1451 * @device: DRBD device.
1453 * Called from process context only (admin command and after_state_ch).
1455 static int _drbd_pause_after(struct drbd_device
*device
)
1457 struct drbd_device
*odev
;
1461 idr_for_each_entry(&drbd_devices
, odev
, i
) {
1462 if (odev
->state
.conn
== C_STANDALONE
&& odev
->state
.disk
== D_DISKLESS
)
1464 if (!_drbd_may_sync_now(odev
))
1465 rv
|= (__drbd_set_state(_NS(odev
, aftr_isp
, 1), CS_HARD
, NULL
)
1466 != SS_NOTHING_TO_DO
);
1474 * _drbd_resume_next() - Resume resync on all devices that may resync now
1475 * @device: DRBD device.
1477 * Called from process context only (admin command and worker).
1479 static int _drbd_resume_next(struct drbd_device
*device
)
1481 struct drbd_device
*odev
;
1485 idr_for_each_entry(&drbd_devices
, odev
, i
) {
1486 if (odev
->state
.conn
== C_STANDALONE
&& odev
->state
.disk
== D_DISKLESS
)
1488 if (odev
->state
.aftr_isp
) {
1489 if (_drbd_may_sync_now(odev
))
1490 rv
|= (__drbd_set_state(_NS(odev
, aftr_isp
, 0),
1492 != SS_NOTHING_TO_DO
) ;
1499 void resume_next_sg(struct drbd_device
*device
)
1501 write_lock_irq(&global_state_lock
);
1502 _drbd_resume_next(device
);
1503 write_unlock_irq(&global_state_lock
);
1506 void suspend_other_sg(struct drbd_device
*device
)
1508 write_lock_irq(&global_state_lock
);
1509 _drbd_pause_after(device
);
1510 write_unlock_irq(&global_state_lock
);
1513 /* caller must hold global_state_lock */
1514 enum drbd_ret_code
drbd_resync_after_valid(struct drbd_device
*device
, int o_minor
)
1516 struct drbd_device
*odev
;
1521 if (o_minor
< -1 || o_minor
> MINORMASK
)
1522 return ERR_RESYNC_AFTER
;
1524 /* check for loops */
1525 odev
= minor_to_device(o_minor
);
1528 return ERR_RESYNC_AFTER_CYCLE
;
1530 /* You are free to depend on diskless, non-existing,
1531 * or not yet/no longer existing minors.
1532 * We only reject dependency loops.
1533 * We cannot follow the dependency chain beyond a detached or
1536 if (!odev
|| !odev
->ldev
|| odev
->state
.disk
== D_DISKLESS
)
1540 resync_after
= rcu_dereference(odev
->ldev
->disk_conf
)->resync_after
;
1542 /* dependency chain ends here, no cycles. */
1543 if (resync_after
== -1)
1546 /* follow the dependency chain */
1547 odev
= minor_to_device(resync_after
);
1551 /* caller must hold global_state_lock */
1552 void drbd_resync_after_changed(struct drbd_device
*device
)
1557 changes
= _drbd_pause_after(device
);
1558 changes
|= _drbd_resume_next(device
);
1562 void drbd_rs_controller_reset(struct drbd_device
*device
)
1564 struct fifo_buffer
*plan
;
1566 atomic_set(&device
->rs_sect_in
, 0);
1567 atomic_set(&device
->rs_sect_ev
, 0);
1568 device
->rs_in_flight
= 0;
1570 /* Updating the RCU protected object in place is necessary since
1571 this function gets called from atomic context.
1572 It is valid since all other updates also lead to an completely
1575 plan
= rcu_dereference(device
->rs_plan_s
);
1581 void start_resync_timer_fn(unsigned long data
)
1583 struct drbd_device
*device
= (struct drbd_device
*) data
;
1585 drbd_queue_work(&first_peer_device(device
)->connection
->sender_work
, &device
->start_resync_work
);
1588 int w_start_resync(struct drbd_work
*w
, int cancel
)
1590 struct drbd_device
*device
= w
->device
;
1592 if (atomic_read(&device
->unacked_cnt
) || atomic_read(&device
->rs_pending_cnt
)) {
1593 dev_warn(DEV
, "w_start_resync later...\n");
1594 device
->start_resync_timer
.expires
= jiffies
+ HZ
/10;
1595 add_timer(&device
->start_resync_timer
);
1599 drbd_start_resync(device
, C_SYNC_SOURCE
);
1600 clear_bit(AHEAD_TO_SYNC_SOURCE
, &device
->flags
);
1605 * drbd_start_resync() - Start the resync process
1606 * @device: DRBD device.
1607 * @side: Either C_SYNC_SOURCE or C_SYNC_TARGET
1609 * This function might bring you directly into one of the
1610 * C_PAUSED_SYNC_* states.
1612 void drbd_start_resync(struct drbd_device
*device
, enum drbd_conns side
)
1614 union drbd_state ns
;
1617 if (device
->state
.conn
>= C_SYNC_SOURCE
&& device
->state
.conn
< C_AHEAD
) {
1618 dev_err(DEV
, "Resync already running!\n");
1622 if (!test_bit(B_RS_H_DONE
, &device
->flags
)) {
1623 if (side
== C_SYNC_TARGET
) {
1624 /* Since application IO was locked out during C_WF_BITMAP_T and
1625 C_WF_SYNC_UUID we are still unmodified. Before going to C_SYNC_TARGET
1626 we check that we might make the data inconsistent. */
1627 r
= drbd_khelper(device
, "before-resync-target");
1628 r
= (r
>> 8) & 0xff;
1630 dev_info(DEV
, "before-resync-target handler returned %d, "
1631 "dropping connection.\n", r
);
1632 conn_request_state(first_peer_device(device
)->connection
, NS(conn
, C_DISCONNECTING
), CS_HARD
);
1635 } else /* C_SYNC_SOURCE */ {
1636 r
= drbd_khelper(device
, "before-resync-source");
1637 r
= (r
>> 8) & 0xff;
1640 dev_info(DEV
, "before-resync-source handler returned %d, "
1641 "ignoring. Old userland tools?", r
);
1643 dev_info(DEV
, "before-resync-source handler returned %d, "
1644 "dropping connection.\n", r
);
1645 conn_request_state(first_peer_device(device
)->connection
,
1646 NS(conn
, C_DISCONNECTING
), CS_HARD
);
1653 if (current
== first_peer_device(device
)->connection
->worker
.task
) {
1654 /* The worker should not sleep waiting for state_mutex,
1655 that can take long */
1656 if (!mutex_trylock(device
->state_mutex
)) {
1657 set_bit(B_RS_H_DONE
, &device
->flags
);
1658 device
->start_resync_timer
.expires
= jiffies
+ HZ
/5;
1659 add_timer(&device
->start_resync_timer
);
1663 mutex_lock(device
->state_mutex
);
1665 clear_bit(B_RS_H_DONE
, &device
->flags
);
1667 write_lock_irq(&global_state_lock
);
1668 /* Did some connection breakage or IO error race with us? */
1669 if (device
->state
.conn
< C_CONNECTED
1670 || !get_ldev_if_state(device
, D_NEGOTIATING
)) {
1671 write_unlock_irq(&global_state_lock
);
1672 mutex_unlock(device
->state_mutex
);
1676 ns
= drbd_read_state(device
);
1678 ns
.aftr_isp
= !_drbd_may_sync_now(device
);
1682 if (side
== C_SYNC_TARGET
)
1683 ns
.disk
= D_INCONSISTENT
;
1684 else /* side == C_SYNC_SOURCE */
1685 ns
.pdsk
= D_INCONSISTENT
;
1687 r
= __drbd_set_state(device
, ns
, CS_VERBOSE
, NULL
);
1688 ns
= drbd_read_state(device
);
1690 if (ns
.conn
< C_CONNECTED
)
1691 r
= SS_UNKNOWN_ERROR
;
1693 if (r
== SS_SUCCESS
) {
1694 unsigned long tw
= drbd_bm_total_weight(device
);
1695 unsigned long now
= jiffies
;
1698 device
->rs_failed
= 0;
1699 device
->rs_paused
= 0;
1700 device
->rs_same_csum
= 0;
1701 device
->rs_last_events
= 0;
1702 device
->rs_last_sect_ev
= 0;
1703 device
->rs_total
= tw
;
1704 device
->rs_start
= now
;
1705 for (i
= 0; i
< DRBD_SYNC_MARKS
; i
++) {
1706 device
->rs_mark_left
[i
] = tw
;
1707 device
->rs_mark_time
[i
] = now
;
1709 _drbd_pause_after(device
);
1711 write_unlock_irq(&global_state_lock
);
1713 if (r
== SS_SUCCESS
) {
1714 /* reset rs_last_bcast when a resync or verify is started,
1715 * to deal with potential jiffies wrap. */
1716 device
->rs_last_bcast
= jiffies
- HZ
;
1718 dev_info(DEV
, "Began resync as %s (will sync %lu KB [%lu bits set]).\n",
1719 drbd_conn_str(ns
.conn
),
1720 (unsigned long) device
->rs_total
<< (BM_BLOCK_SHIFT
-10),
1721 (unsigned long) device
->rs_total
);
1722 if (side
== C_SYNC_TARGET
)
1723 device
->bm_resync_fo
= 0;
1725 /* Since protocol 96, we must serialize drbd_gen_and_send_sync_uuid
1726 * with w_send_oos, or the sync target will get confused as to
1727 * how much bits to resync. We cannot do that always, because for an
1728 * empty resync and protocol < 95, we need to do it here, as we call
1729 * drbd_resync_finished from here in that case.
1730 * We drbd_gen_and_send_sync_uuid here for protocol < 96,
1731 * and from after_state_ch otherwise. */
1732 if (side
== C_SYNC_SOURCE
&&
1733 first_peer_device(device
)->connection
->agreed_pro_version
< 96)
1734 drbd_gen_and_send_sync_uuid(device
);
1736 if (first_peer_device(device
)->connection
->agreed_pro_version
< 95 &&
1737 device
->rs_total
== 0) {
1738 /* This still has a race (about when exactly the peers
1739 * detect connection loss) that can lead to a full sync
1740 * on next handshake. In 8.3.9 we fixed this with explicit
1741 * resync-finished notifications, but the fix
1742 * introduces a protocol change. Sleeping for some
1743 * time longer than the ping interval + timeout on the
1744 * SyncSource, to give the SyncTarget the chance to
1745 * detect connection loss, then waiting for a ping
1746 * response (implicit in drbd_resync_finished) reduces
1747 * the race considerably, but does not solve it. */
1748 if (side
== C_SYNC_SOURCE
) {
1749 struct net_conf
*nc
;
1753 nc
= rcu_dereference(first_peer_device(device
)->connection
->net_conf
);
1754 timeo
= nc
->ping_int
* HZ
+ nc
->ping_timeo
* HZ
/ 9;
1756 schedule_timeout_interruptible(timeo
);
1758 drbd_resync_finished(device
);
1761 drbd_rs_controller_reset(device
);
1762 /* ns.conn may already be != device->state.conn,
1763 * we may have been paused in between, or become paused until
1764 * the timer triggers.
1765 * No matter, that is handled in resync_timer_fn() */
1766 if (ns
.conn
== C_SYNC_TARGET
)
1767 mod_timer(&device
->resync_timer
, jiffies
);
1769 drbd_md_sync(device
);
1772 mutex_unlock(device
->state_mutex
);
1775 /* If the resource already closed the current epoch, but we did not
1776 * (because we have not yet seen new requests), we should send the
1777 * corresponding barrier now. Must be checked within the same spinlock
1778 * that is used to check for new requests. */
1779 static bool need_to_send_barrier(struct drbd_connection
*connection
)
1781 if (!connection
->send
.seen_any_write_yet
)
1784 /* Skip barriers that do not contain any writes.
1785 * This may happen during AHEAD mode. */
1786 if (!connection
->send
.current_epoch_writes
)
1789 /* ->req_lock is held when requests are queued on
1790 * connection->sender_work, and put into ->transfer_log.
1791 * It is also held when ->current_tle_nr is increased.
1792 * So either there are already new requests queued,
1793 * and corresponding barriers will be send there.
1794 * Or nothing new is queued yet, so the difference will be 1.
1796 if (atomic_read(&connection
->current_tle_nr
) !=
1797 connection
->send
.current_epoch_nr
+ 1)
1803 static bool dequeue_work_batch(struct drbd_work_queue
*queue
, struct list_head
*work_list
)
1805 spin_lock_irq(&queue
->q_lock
);
1806 list_splice_init(&queue
->q
, work_list
);
1807 spin_unlock_irq(&queue
->q_lock
);
1808 return !list_empty(work_list
);
1811 static bool dequeue_work_item(struct drbd_work_queue
*queue
, struct list_head
*work_list
)
1813 spin_lock_irq(&queue
->q_lock
);
1814 if (!list_empty(&queue
->q
))
1815 list_move(queue
->q
.next
, work_list
);
1816 spin_unlock_irq(&queue
->q_lock
);
1817 return !list_empty(work_list
);
1820 static void wait_for_work(struct drbd_connection
*connection
, struct list_head
*work_list
)
1823 struct net_conf
*nc
;
1826 dequeue_work_item(&connection
->sender_work
, work_list
);
1827 if (!list_empty(work_list
))
1830 /* Still nothing to do?
1831 * Maybe we still need to close the current epoch,
1832 * even if no new requests are queued yet.
1834 * Also, poke TCP, just in case.
1835 * Then wait for new work (or signal). */
1837 nc
= rcu_dereference(connection
->net_conf
);
1838 uncork
= nc
? nc
->tcp_cork
: 0;
1841 mutex_lock(&connection
->data
.mutex
);
1842 if (connection
->data
.socket
)
1843 drbd_tcp_uncork(connection
->data
.socket
);
1844 mutex_unlock(&connection
->data
.mutex
);
1849 prepare_to_wait(&connection
->sender_work
.q_wait
, &wait
, TASK_INTERRUPTIBLE
);
1850 spin_lock_irq(&connection
->req_lock
);
1851 spin_lock(&connection
->sender_work
.q_lock
); /* FIXME get rid of this one? */
1852 /* dequeue single item only,
1853 * we still use drbd_queue_work_front() in some places */
1854 if (!list_empty(&connection
->sender_work
.q
))
1855 list_move(connection
->sender_work
.q
.next
, work_list
);
1856 spin_unlock(&connection
->sender_work
.q_lock
); /* FIXME get rid of this one? */
1857 if (!list_empty(work_list
) || signal_pending(current
)) {
1858 spin_unlock_irq(&connection
->req_lock
);
1861 send_barrier
= need_to_send_barrier(connection
);
1862 spin_unlock_irq(&connection
->req_lock
);
1864 drbd_send_barrier(connection
);
1865 connection
->send
.current_epoch_nr
++;
1868 /* may be woken up for other things but new work, too,
1869 * e.g. if the current epoch got closed.
1870 * In which case we send the barrier above. */
1872 finish_wait(&connection
->sender_work
.q_wait
, &wait
);
1874 /* someone may have changed the config while we have been waiting above. */
1876 nc
= rcu_dereference(connection
->net_conf
);
1877 cork
= nc
? nc
->tcp_cork
: 0;
1879 mutex_lock(&connection
->data
.mutex
);
1880 if (connection
->data
.socket
) {
1882 drbd_tcp_cork(connection
->data
.socket
);
1884 drbd_tcp_uncork(connection
->data
.socket
);
1886 mutex_unlock(&connection
->data
.mutex
);
1889 int drbd_worker(struct drbd_thread
*thi
)
1891 struct drbd_connection
*connection
= thi
->connection
;
1892 struct drbd_work
*w
= NULL
;
1893 struct drbd_device
*device
;
1894 LIST_HEAD(work_list
);
1897 while (get_t_state(thi
) == RUNNING
) {
1898 drbd_thread_current_set_cpu(thi
);
1900 /* as long as we use drbd_queue_work_front(),
1901 * we may only dequeue single work items here, not batches. */
1902 if (list_empty(&work_list
))
1903 wait_for_work(connection
, &work_list
);
1905 if (signal_pending(current
)) {
1906 flush_signals(current
);
1907 if (get_t_state(thi
) == RUNNING
) {
1908 conn_warn(connection
, "Worker got an unexpected signal\n");
1914 if (get_t_state(thi
) != RUNNING
)
1917 while (!list_empty(&work_list
)) {
1918 w
= list_first_entry(&work_list
, struct drbd_work
, list
);
1919 list_del_init(&w
->list
);
1920 if (w
->cb(w
, connection
->cstate
< C_WF_REPORT_PARAMS
) == 0)
1922 if (connection
->cstate
>= C_WF_REPORT_PARAMS
)
1923 conn_request_state(connection
, NS(conn
, C_NETWORK_FAILURE
), CS_HARD
);
1928 while (!list_empty(&work_list
)) {
1929 w
= list_first_entry(&work_list
, struct drbd_work
, list
);
1930 list_del_init(&w
->list
);
1933 dequeue_work_batch(&connection
->sender_work
, &work_list
);
1934 } while (!list_empty(&work_list
));
1937 idr_for_each_entry(&connection
->volumes
, device
, vnr
) {
1938 D_ASSERT(device
->state
.disk
== D_DISKLESS
&& device
->state
.conn
== C_STANDALONE
);
1939 kref_get(&device
->kref
);
1941 drbd_device_cleanup(device
);
1942 kref_put(&device
->kref
, drbd_destroy_device
);