1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
4 * Copyright 2004-2011 Red Hat, Inc.
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 #include <linux/dlm.h>
11 #include <linux/slab.h>
12 #include <linux/types.h>
13 #include <linux/delay.h>
14 #include <linux/gfs2_ondisk.h>
15 #include <linux/sched/signal.h>
23 #include "trace_gfs2.h"
26 * gfs2_update_stats - Update time based stats
27 * @s: The stats to update (local or global)
28 * @index: The index inside @s
29 * @sample: New data to include
31 static inline void gfs2_update_stats(struct gfs2_lkstats
*s
, unsigned index
,
35 * @delta is the difference between the current rtt sample and the
36 * running average srtt. We add 1/8 of that to the srtt in order to
37 * update the current srtt estimate. The variance estimate is a bit
38 * more complicated. We subtract the current variance estimate from
39 * the abs value of the @delta and add 1/4 of that to the running
40 * total. That's equivalent to 3/4 of the current variance
41 * estimate plus 1/4 of the abs of @delta.
43 * Note that the index points at the array entry containing the
44 * smoothed mean value, and the variance is always in the following
47 * Reference: TCP/IP Illustrated, vol 2, p. 831,832
48 * All times are in units of integer nanoseconds. Unlike the TCP/IP
49 * case, they are not scaled fixed point.
52 s64 delta
= sample
- s
->stats
[index
];
53 s
->stats
[index
] += (delta
>> 3);
55 s
->stats
[index
] += (s64
)(abs(delta
) - s
->stats
[index
]) >> 2;
59 * gfs2_update_reply_times - Update locking statistics
60 * @gl: The glock to update
62 * This assumes that gl->gl_dstamp has been set earlier.
64 * The rtt (lock round trip time) is an estimate of the time
65 * taken to perform a dlm lock request. We update it on each
68 * The blocking flag is set on the glock for all dlm requests
69 * which may potentially block due to lock requests from other nodes.
70 * DLM requests where the current lock state is exclusive, the
71 * requested state is null (or unlocked) or where the TRY or
72 * TRY_1CB flags are set are classified as non-blocking. All
73 * other DLM requests are counted as (potentially) blocking.
75 static inline void gfs2_update_reply_times(struct gfs2_glock
*gl
)
77 struct gfs2_pcpu_lkstats
*lks
;
78 const unsigned gltype
= gl
->gl_name
.ln_type
;
79 unsigned index
= test_bit(GLF_BLOCKING
, &gl
->gl_flags
) ?
80 GFS2_LKS_SRTTB
: GFS2_LKS_SRTT
;
84 rtt
= ktime_to_ns(ktime_sub(ktime_get_real(), gl
->gl_dstamp
));
85 lks
= this_cpu_ptr(gl
->gl_name
.ln_sbd
->sd_lkstats
);
86 gfs2_update_stats(&gl
->gl_stats
, index
, rtt
); /* Local */
87 gfs2_update_stats(&lks
->lkstats
[gltype
], index
, rtt
); /* Global */
90 trace_gfs2_glock_lock_time(gl
, rtt
);
94 * gfs2_update_request_times - Update locking statistics
95 * @gl: The glock to update
97 * The irt (lock inter-request times) measures the average time
98 * between requests to the dlm. It is updated immediately before
102 static inline void gfs2_update_request_times(struct gfs2_glock
*gl
)
104 struct gfs2_pcpu_lkstats
*lks
;
105 const unsigned gltype
= gl
->gl_name
.ln_type
;
110 dstamp
= gl
->gl_dstamp
;
111 gl
->gl_dstamp
= ktime_get_real();
112 irt
= ktime_to_ns(ktime_sub(gl
->gl_dstamp
, dstamp
));
113 lks
= this_cpu_ptr(gl
->gl_name
.ln_sbd
->sd_lkstats
);
114 gfs2_update_stats(&gl
->gl_stats
, GFS2_LKS_SIRT
, irt
); /* Local */
115 gfs2_update_stats(&lks
->lkstats
[gltype
], GFS2_LKS_SIRT
, irt
); /* Global */
119 static void gdlm_ast(void *arg
)
121 struct gfs2_glock
*gl
= arg
;
122 unsigned ret
= gl
->gl_state
;
124 gfs2_update_reply_times(gl
);
125 BUG_ON(gl
->gl_lksb
.sb_flags
& DLM_SBF_DEMOTED
);
127 if ((gl
->gl_lksb
.sb_flags
& DLM_SBF_VALNOTVALID
) && gl
->gl_lksb
.sb_lvbptr
)
128 memset(gl
->gl_lksb
.sb_lvbptr
, 0, GDLM_LVB_SIZE
);
130 switch (gl
->gl_lksb
.sb_status
) {
131 case -DLM_EUNLOCK
: /* Unlocked, so glock can be freed */
132 if (gl
->gl_ops
->go_free
)
133 gl
->gl_ops
->go_free(gl
);
136 case -DLM_ECANCEL
: /* Cancel while getting lock */
137 ret
|= LM_OUT_CANCELED
;
139 case -EAGAIN
: /* Try lock fails */
140 case -EDEADLK
: /* Deadlock detected */
142 case -ETIMEDOUT
: /* Canceled due to timeout */
145 case 0: /* Success */
147 default: /* Something unexpected */
152 if (gl
->gl_lksb
.sb_flags
& DLM_SBF_ALTMODE
) {
153 if (gl
->gl_req
== LM_ST_SHARED
)
154 ret
= LM_ST_DEFERRED
;
155 else if (gl
->gl_req
== LM_ST_DEFERRED
)
161 set_bit(GLF_INITIAL
, &gl
->gl_flags
);
162 gfs2_glock_complete(gl
, ret
);
165 if (!test_bit(GLF_INITIAL
, &gl
->gl_flags
))
166 gl
->gl_lksb
.sb_lkid
= 0;
167 gfs2_glock_complete(gl
, ret
);
170 static void gdlm_bast(void *arg
, int mode
)
172 struct gfs2_glock
*gl
= arg
;
176 gfs2_glock_cb(gl
, LM_ST_UNLOCKED
);
179 gfs2_glock_cb(gl
, LM_ST_DEFERRED
);
182 gfs2_glock_cb(gl
, LM_ST_SHARED
);
185 fs_err(gl
->gl_name
.ln_sbd
, "unknown bast mode %d\n", mode
);
190 /* convert gfs lock-state to dlm lock-mode */
192 static int make_mode(struct gfs2_sbd
*sdp
, const unsigned int lmstate
)
197 case LM_ST_EXCLUSIVE
:
204 fs_err(sdp
, "unknown LM state %d\n", lmstate
);
209 static u32
make_flags(struct gfs2_glock
*gl
, const unsigned int gfs_flags
,
214 if (gl
->gl_lksb
.sb_lvbptr
)
215 lkf
|= DLM_LKF_VALBLK
;
217 if (gfs_flags
& LM_FLAG_TRY
)
218 lkf
|= DLM_LKF_NOQUEUE
;
220 if (gfs_flags
& LM_FLAG_TRY_1CB
) {
221 lkf
|= DLM_LKF_NOQUEUE
;
222 lkf
|= DLM_LKF_NOQUEUEBAST
;
225 if (gfs_flags
& LM_FLAG_PRIORITY
) {
226 lkf
|= DLM_LKF_NOORDER
;
227 lkf
|= DLM_LKF_HEADQUE
;
230 if (gfs_flags
& LM_FLAG_ANY
) {
231 if (req
== DLM_LOCK_PR
)
232 lkf
|= DLM_LKF_ALTCW
;
233 else if (req
== DLM_LOCK_CW
)
234 lkf
|= DLM_LKF_ALTPR
;
239 if (gl
->gl_lksb
.sb_lkid
!= 0) {
240 lkf
|= DLM_LKF_CONVERT
;
241 if (test_bit(GLF_BLOCKING
, &gl
->gl_flags
))
242 lkf
|= DLM_LKF_QUECVT
;
248 static void gfs2_reverse_hex(char *c
, u64 value
)
252 *c
-- = hex_asc
[value
& 0x0f];
257 static int gdlm_lock(struct gfs2_glock
*gl
, unsigned int req_state
,
260 struct lm_lockstruct
*ls
= &gl
->gl_name
.ln_sbd
->sd_lockstruct
;
263 char strname
[GDLM_STRNAME_BYTES
] = "";
265 req
= make_mode(gl
->gl_name
.ln_sbd
, req_state
);
266 lkf
= make_flags(gl
, flags
, req
);
267 gfs2_glstats_inc(gl
, GFS2_LKS_DCOUNT
);
268 gfs2_sbstats_inc(gl
, GFS2_LKS_DCOUNT
);
269 if (gl
->gl_lksb
.sb_lkid
) {
270 gfs2_update_request_times(gl
);
272 memset(strname
, ' ', GDLM_STRNAME_BYTES
- 1);
273 strname
[GDLM_STRNAME_BYTES
- 1] = '\0';
274 gfs2_reverse_hex(strname
+ 7, gl
->gl_name
.ln_type
);
275 gfs2_reverse_hex(strname
+ 23, gl
->gl_name
.ln_number
);
276 gl
->gl_dstamp
= ktime_get_real();
279 * Submit the actual lock request.
282 return dlm_lock(ls
->ls_dlm
, req
, &gl
->gl_lksb
, lkf
, strname
,
283 GDLM_STRNAME_BYTES
- 1, 0, gdlm_ast
, gl
, gdlm_bast
);
286 static void gdlm_put_lock(struct gfs2_glock
*gl
)
288 struct gfs2_sbd
*sdp
= gl
->gl_name
.ln_sbd
;
289 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
292 if (gl
->gl_lksb
.sb_lkid
== 0) {
297 clear_bit(GLF_BLOCKING
, &gl
->gl_flags
);
298 gfs2_glstats_inc(gl
, GFS2_LKS_DCOUNT
);
299 gfs2_sbstats_inc(gl
, GFS2_LKS_DCOUNT
);
300 gfs2_update_request_times(gl
);
302 /* don't want to call dlm if we've unmounted the lock protocol */
303 if (test_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
)) {
307 /* don't want to skip dlm_unlock writing the lvb when lock has one */
309 if (test_bit(SDF_SKIP_DLM_UNLOCK
, &sdp
->sd_flags
) &&
310 !gl
->gl_lksb
.sb_lvbptr
) {
315 error
= dlm_unlock(ls
->ls_dlm
, gl
->gl_lksb
.sb_lkid
, DLM_LKF_VALBLK
,
318 fs_err(sdp
, "gdlm_unlock %x,%llx err=%d\n",
320 (unsigned long long)gl
->gl_name
.ln_number
, error
);
325 static void gdlm_cancel(struct gfs2_glock
*gl
)
327 struct lm_lockstruct
*ls
= &gl
->gl_name
.ln_sbd
->sd_lockstruct
;
328 dlm_unlock(ls
->ls_dlm
, gl
->gl_lksb
.sb_lkid
, DLM_LKF_CANCEL
, NULL
, gl
);
332 * dlm/gfs2 recovery coordination using dlm_recover callbacks
334 * 0. gfs2 checks for another cluster node withdraw, needing journal replay
335 * 1. dlm_controld sees lockspace members change
336 * 2. dlm_controld blocks dlm-kernel locking activity
337 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
338 * 4. dlm_controld starts and finishes its own user level recovery
339 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
340 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
341 * 7. dlm_recoverd does its own lock recovery
342 * 8. dlm_recoverd unblocks dlm-kernel locking activity
343 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
344 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
345 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
346 * 12. gfs2_recover dequeues and recovers journals of failed nodes
347 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
348 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
349 * 15. gfs2_control unblocks normal locking when all journals are recovered
351 * - failures during recovery
353 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
354 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
355 * recovering for a prior failure. gfs2_control needs a way to detect
356 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
357 * the recover_block and recover_start values.
359 * recover_done() provides a new lockspace generation number each time it
360 * is called (step 9). This generation number is saved as recover_start.
361 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
362 * recover_block = recover_start. So, while recover_block is equal to
363 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
364 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
366 * - more specific gfs2 steps in sequence above
368 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
369 * 6. recover_slot records any failed jids (maybe none)
370 * 9. recover_done sets recover_start = new generation number
371 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
372 * 12. gfs2_recover does journal recoveries for failed jids identified above
373 * 14. gfs2_control clears control_lock lvb bits for recovered jids
374 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
375 * again) then do nothing, otherwise if recover_start > recover_block
376 * then clear BLOCK_LOCKS.
378 * - parallel recovery steps across all nodes
380 * All nodes attempt to update the control_lock lvb with the new generation
381 * number and jid bits, but only the first to get the control_lock EX will
382 * do so; others will see that it's already done (lvb already contains new
383 * generation number.)
385 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
386 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
387 * . One node gets control_lock first and writes the lvb, others see it's done
388 * . All nodes attempt to recover jids for which they see control_lock bits set
389 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
390 * . All nodes will eventually see all lvb bits clear and unblock locks
392 * - is there a problem with clearing an lvb bit that should be set
393 * and missing a journal recovery?
396 * 2. lvb bit set for step 1
397 * 3. jid recovered for step 1
398 * 4. jid taken again (new mount)
399 * 5. jid fails (for step 4)
400 * 6. lvb bit set for step 5 (will already be set)
401 * 7. lvb bit cleared for step 3
403 * This is not a problem because the failure in step 5 does not
404 * require recovery, because the mount in step 4 could not have
405 * progressed far enough to unblock locks and access the fs. The
406 * control_mount() function waits for all recoveries to be complete
407 * for the latest lockspace generation before ever unblocking locks
408 * and returning. The mount in step 4 waits until the recovery in
411 * - special case of first mounter: first node to mount the fs
413 * The first node to mount a gfs2 fs needs to check all the journals
414 * and recover any that need recovery before other nodes are allowed
415 * to mount the fs. (Others may begin mounting, but they must wait
416 * for the first mounter to be done before taking locks on the fs
417 * or accessing the fs.) This has two parts:
419 * 1. The mounted_lock tells a node it's the first to mount the fs.
420 * Each node holds the mounted_lock in PR while it's mounted.
421 * Each node tries to acquire the mounted_lock in EX when it mounts.
422 * If a node is granted the mounted_lock EX it means there are no
423 * other mounted nodes (no PR locks exist), and it is the first mounter.
424 * The mounted_lock is demoted to PR when first recovery is done, so
425 * others will fail to get an EX lock, but will get a PR lock.
427 * 2. The control_lock blocks others in control_mount() while the first
428 * mounter is doing first mount recovery of all journals.
429 * A mounting node needs to acquire control_lock in EX mode before
430 * it can proceed. The first mounter holds control_lock in EX while doing
431 * the first mount recovery, blocking mounts from other nodes, then demotes
432 * control_lock to NL when it's done (others_may_mount/first_done),
433 * allowing other nodes to continue mounting.
436 * control_lock EX/NOQUEUE success
437 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
439 * do first mounter recovery
440 * mounted_lock EX->PR
441 * control_lock EX->NL, write lvb generation
444 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
445 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
446 * mounted_lock PR/NOQUEUE success
447 * read lvb generation
448 * control_lock EX->NL
451 * - mount during recovery
453 * If a node mounts while others are doing recovery (not first mounter),
454 * the mounting node will get its initial recover_done() callback without
455 * having seen any previous failures/callbacks.
457 * It must wait for all recoveries preceding its mount to be finished
458 * before it unblocks locks. It does this by repeating the "other mounter"
459 * steps above until the lvb generation number is >= its mount generation
460 * number (from initial recover_done) and all lvb bits are clear.
462 * - control_lock lvb format
464 * 4 bytes generation number: the latest dlm lockspace generation number
465 * from recover_done callback. Indicates the jid bitmap has been updated
466 * to reflect all slot failures through that generation.
468 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
469 * that jid N needs recovery.
472 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
474 static void control_lvb_read(struct lm_lockstruct
*ls
, uint32_t *lvb_gen
,
478 memcpy(lvb_bits
, ls
->ls_control_lvb
, GDLM_LVB_SIZE
);
479 memcpy(&gen
, lvb_bits
, sizeof(__le32
));
480 *lvb_gen
= le32_to_cpu(gen
);
483 static void control_lvb_write(struct lm_lockstruct
*ls
, uint32_t lvb_gen
,
487 memcpy(ls
->ls_control_lvb
, lvb_bits
, GDLM_LVB_SIZE
);
488 gen
= cpu_to_le32(lvb_gen
);
489 memcpy(ls
->ls_control_lvb
, &gen
, sizeof(__le32
));
492 static int all_jid_bits_clear(char *lvb
)
494 return !memchr_inv(lvb
+ JID_BITMAP_OFFSET
, 0,
495 GDLM_LVB_SIZE
- JID_BITMAP_OFFSET
);
498 static void sync_wait_cb(void *arg
)
500 struct lm_lockstruct
*ls
= arg
;
501 complete(&ls
->ls_sync_wait
);
504 static int sync_unlock(struct gfs2_sbd
*sdp
, struct dlm_lksb
*lksb
, char *name
)
506 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
509 error
= dlm_unlock(ls
->ls_dlm
, lksb
->sb_lkid
, 0, lksb
, ls
);
511 fs_err(sdp
, "%s lkid %x error %d\n",
512 name
, lksb
->sb_lkid
, error
);
516 wait_for_completion(&ls
->ls_sync_wait
);
518 if (lksb
->sb_status
!= -DLM_EUNLOCK
) {
519 fs_err(sdp
, "%s lkid %x status %d\n",
520 name
, lksb
->sb_lkid
, lksb
->sb_status
);
526 static int sync_lock(struct gfs2_sbd
*sdp
, int mode
, uint32_t flags
,
527 unsigned int num
, struct dlm_lksb
*lksb
, char *name
)
529 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
530 char strname
[GDLM_STRNAME_BYTES
];
533 memset(strname
, 0, GDLM_STRNAME_BYTES
);
534 snprintf(strname
, GDLM_STRNAME_BYTES
, "%8x%16x", LM_TYPE_NONDISK
, num
);
536 error
= dlm_lock(ls
->ls_dlm
, mode
, lksb
, flags
,
537 strname
, GDLM_STRNAME_BYTES
- 1,
538 0, sync_wait_cb
, ls
, NULL
);
540 fs_err(sdp
, "%s lkid %x flags %x mode %d error %d\n",
541 name
, lksb
->sb_lkid
, flags
, mode
, error
);
545 wait_for_completion(&ls
->ls_sync_wait
);
547 status
= lksb
->sb_status
;
549 if (status
&& status
!= -EAGAIN
) {
550 fs_err(sdp
, "%s lkid %x flags %x mode %d status %d\n",
551 name
, lksb
->sb_lkid
, flags
, mode
, status
);
557 static int mounted_unlock(struct gfs2_sbd
*sdp
)
559 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
560 return sync_unlock(sdp
, &ls
->ls_mounted_lksb
, "mounted_lock");
563 static int mounted_lock(struct gfs2_sbd
*sdp
, int mode
, uint32_t flags
)
565 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
566 return sync_lock(sdp
, mode
, flags
, GFS2_MOUNTED_LOCK
,
567 &ls
->ls_mounted_lksb
, "mounted_lock");
570 static int control_unlock(struct gfs2_sbd
*sdp
)
572 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
573 return sync_unlock(sdp
, &ls
->ls_control_lksb
, "control_lock");
576 static int control_lock(struct gfs2_sbd
*sdp
, int mode
, uint32_t flags
)
578 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
579 return sync_lock(sdp
, mode
, flags
, GFS2_CONTROL_LOCK
,
580 &ls
->ls_control_lksb
, "control_lock");
584 * remote_withdraw - react to a node withdrawing from the file system
585 * @sdp: The superblock
587 static void remote_withdraw(struct gfs2_sbd
*sdp
)
589 struct gfs2_jdesc
*jd
;
590 int ret
= 0, count
= 0;
592 list_for_each_entry(jd
, &sdp
->sd_jindex_list
, jd_list
) {
593 if (jd
->jd_jid
== sdp
->sd_lockstruct
.ls_jid
)
595 ret
= gfs2_recover_journal(jd
, true);
601 /* Now drop the additional reference we acquired */
602 fs_err(sdp
, "Journals checked: %d, ret = %d.\n", count
, ret
);
605 static void gfs2_control_func(struct work_struct
*work
)
607 struct gfs2_sbd
*sdp
= container_of(work
, struct gfs2_sbd
, sd_control_work
.work
);
608 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
609 uint32_t block_gen
, start_gen
, lvb_gen
, flags
;
615 /* First check for other nodes that may have done a withdraw. */
616 if (test_bit(SDF_REMOTE_WITHDRAW
, &sdp
->sd_flags
)) {
617 remote_withdraw(sdp
);
618 clear_bit(SDF_REMOTE_WITHDRAW
, &sdp
->sd_flags
);
622 spin_lock(&ls
->ls_recover_spin
);
624 * No MOUNT_DONE means we're still mounting; control_mount()
625 * will set this flag, after which this thread will take over
626 * all further clearing of BLOCK_LOCKS.
628 * FIRST_MOUNT means this node is doing first mounter recovery,
629 * for which recovery control is handled by
630 * control_mount()/control_first_done(), not this thread.
632 if (!test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
633 test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
634 spin_unlock(&ls
->ls_recover_spin
);
637 block_gen
= ls
->ls_recover_block
;
638 start_gen
= ls
->ls_recover_start
;
639 spin_unlock(&ls
->ls_recover_spin
);
642 * Equal block_gen and start_gen implies we are between
643 * recover_prep and recover_done callbacks, which means
644 * dlm recovery is in progress and dlm locking is blocked.
645 * There's no point trying to do any work until recover_done.
648 if (block_gen
== start_gen
)
652 * Propagate recover_submit[] and recover_result[] to lvb:
653 * dlm_recoverd adds to recover_submit[] jids needing recovery
654 * gfs2_recover adds to recover_result[] journal recovery results
656 * set lvb bit for jids in recover_submit[] if the lvb has not
657 * yet been updated for the generation of the failure
659 * clear lvb bit for jids in recover_result[] if the result of
660 * the journal recovery is SUCCESS
663 error
= control_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_VALBLK
);
665 fs_err(sdp
, "control lock EX error %d\n", error
);
669 control_lvb_read(ls
, &lvb_gen
, ls
->ls_lvb_bits
);
671 spin_lock(&ls
->ls_recover_spin
);
672 if (block_gen
!= ls
->ls_recover_block
||
673 start_gen
!= ls
->ls_recover_start
) {
674 fs_info(sdp
, "recover generation %u block1 %u %u\n",
675 start_gen
, block_gen
, ls
->ls_recover_block
);
676 spin_unlock(&ls
->ls_recover_spin
);
677 control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
681 recover_size
= ls
->ls_recover_size
;
683 if (lvb_gen
<= start_gen
) {
685 * Clear lvb bits for jids we've successfully recovered.
686 * Because all nodes attempt to recover failed journals,
687 * a journal can be recovered multiple times successfully
688 * in succession. Only the first will really do recovery,
689 * the others find it clean, but still report a successful
690 * recovery. So, another node may have already recovered
691 * the jid and cleared the lvb bit for it.
693 for (i
= 0; i
< recover_size
; i
++) {
694 if (ls
->ls_recover_result
[i
] != LM_RD_SUCCESS
)
697 ls
->ls_recover_result
[i
] = 0;
699 if (!test_bit_le(i
, ls
->ls_lvb_bits
+ JID_BITMAP_OFFSET
))
702 __clear_bit_le(i
, ls
->ls_lvb_bits
+ JID_BITMAP_OFFSET
);
707 if (lvb_gen
== start_gen
) {
709 * Failed slots before start_gen are already set in lvb.
711 for (i
= 0; i
< recover_size
; i
++) {
712 if (!ls
->ls_recover_submit
[i
])
714 if (ls
->ls_recover_submit
[i
] < lvb_gen
)
715 ls
->ls_recover_submit
[i
] = 0;
717 } else if (lvb_gen
< start_gen
) {
719 * Failed slots before start_gen are not yet set in lvb.
721 for (i
= 0; i
< recover_size
; i
++) {
722 if (!ls
->ls_recover_submit
[i
])
724 if (ls
->ls_recover_submit
[i
] < start_gen
) {
725 ls
->ls_recover_submit
[i
] = 0;
726 __set_bit_le(i
, ls
->ls_lvb_bits
+ JID_BITMAP_OFFSET
);
729 /* even if there are no bits to set, we need to write the
730 latest generation to the lvb */
734 * we should be getting a recover_done() for lvb_gen soon
737 spin_unlock(&ls
->ls_recover_spin
);
740 control_lvb_write(ls
, start_gen
, ls
->ls_lvb_bits
);
741 flags
= DLM_LKF_CONVERT
| DLM_LKF_VALBLK
;
743 flags
= DLM_LKF_CONVERT
;
746 error
= control_lock(sdp
, DLM_LOCK_NL
, flags
);
748 fs_err(sdp
, "control lock NL error %d\n", error
);
753 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
754 * and clear a jid bit in the lvb if the recovery is a success.
755 * Eventually all journals will be recovered, all jid bits will
756 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
759 for (i
= 0; i
< recover_size
; i
++) {
760 if (test_bit_le(i
, ls
->ls_lvb_bits
+ JID_BITMAP_OFFSET
)) {
761 fs_info(sdp
, "recover generation %u jid %d\n",
763 gfs2_recover_set(sdp
, i
);
771 * No more jid bits set in lvb, all recovery is done, unblock locks
772 * (unless a new recover_prep callback has occured blocking locks
773 * again while working above)
776 spin_lock(&ls
->ls_recover_spin
);
777 if (ls
->ls_recover_block
== block_gen
&&
778 ls
->ls_recover_start
== start_gen
) {
779 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
780 spin_unlock(&ls
->ls_recover_spin
);
781 fs_info(sdp
, "recover generation %u done\n", start_gen
);
782 gfs2_glock_thaw(sdp
);
784 fs_info(sdp
, "recover generation %u block2 %u %u\n",
785 start_gen
, block_gen
, ls
->ls_recover_block
);
786 spin_unlock(&ls
->ls_recover_spin
);
790 static int control_mount(struct gfs2_sbd
*sdp
)
792 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
793 uint32_t start_gen
, block_gen
, mount_gen
, lvb_gen
;
798 memset(&ls
->ls_mounted_lksb
, 0, sizeof(struct dlm_lksb
));
799 memset(&ls
->ls_control_lksb
, 0, sizeof(struct dlm_lksb
));
800 memset(&ls
->ls_control_lvb
, 0, GDLM_LVB_SIZE
);
801 ls
->ls_control_lksb
.sb_lvbptr
= ls
->ls_control_lvb
;
802 init_completion(&ls
->ls_sync_wait
);
804 set_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
806 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_VALBLK
);
808 fs_err(sdp
, "control_mount control_lock NL error %d\n", error
);
812 error
= mounted_lock(sdp
, DLM_LOCK_NL
, 0);
814 fs_err(sdp
, "control_mount mounted_lock NL error %d\n", error
);
818 mounted_mode
= DLM_LOCK_NL
;
821 if (retries
++ && signal_pending(current
)) {
827 * We always start with both locks in NL. control_lock is
828 * demoted to NL below so we don't need to do it here.
831 if (mounted_mode
!= DLM_LOCK_NL
) {
832 error
= mounted_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
835 mounted_mode
= DLM_LOCK_NL
;
839 * Other nodes need to do some work in dlm recovery and gfs2_control
840 * before the recover_done and control_lock will be ready for us below.
841 * A delay here is not required but often avoids having to retry.
844 msleep_interruptible(500);
847 * Acquire control_lock in EX and mounted_lock in either EX or PR.
848 * control_lock lvb keeps track of any pending journal recoveries.
849 * mounted_lock indicates if any other nodes have the fs mounted.
852 error
= control_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
|DLM_LKF_VALBLK
);
853 if (error
== -EAGAIN
) {
856 fs_err(sdp
, "control_mount control_lock EX error %d\n", error
);
861 * If we're a spectator, we don't want to take the lock in EX because
862 * we cannot do the first-mount responsibility it implies: recovery.
864 if (sdp
->sd_args
.ar_spectator
)
867 error
= mounted_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
);
869 mounted_mode
= DLM_LOCK_EX
;
871 } else if (error
!= -EAGAIN
) {
872 fs_err(sdp
, "control_mount mounted_lock EX error %d\n", error
);
876 error
= mounted_lock(sdp
, DLM_LOCK_PR
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
);
878 mounted_mode
= DLM_LOCK_PR
;
881 /* not even -EAGAIN should happen here */
882 fs_err(sdp
, "control_mount mounted_lock PR error %d\n", error
);
888 * If we got both locks above in EX, then we're the first mounter.
889 * If not, then we need to wait for the control_lock lvb to be
890 * updated by other mounted nodes to reflect our mount generation.
892 * In simple first mounter cases, first mounter will see zero lvb_gen,
893 * but in cases where all existing nodes leave/fail before mounting
894 * nodes finish control_mount, then all nodes will be mounting and
895 * lvb_gen will be non-zero.
898 control_lvb_read(ls
, &lvb_gen
, ls
->ls_lvb_bits
);
900 if (lvb_gen
== 0xFFFFFFFF) {
901 /* special value to force mount attempts to fail */
902 fs_err(sdp
, "control_mount control_lock disabled\n");
907 if (mounted_mode
== DLM_LOCK_EX
) {
908 /* first mounter, keep both EX while doing first recovery */
909 spin_lock(&ls
->ls_recover_spin
);
910 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
911 set_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
);
912 set_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
913 spin_unlock(&ls
->ls_recover_spin
);
914 fs_info(sdp
, "first mounter control generation %u\n", lvb_gen
);
918 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
923 * We are not first mounter, now we need to wait for the control_lock
924 * lvb generation to be >= the generation from our first recover_done
925 * and all lvb bits to be clear (no pending journal recoveries.)
928 if (!all_jid_bits_clear(ls
->ls_lvb_bits
)) {
929 /* journals need recovery, wait until all are clear */
930 fs_info(sdp
, "control_mount wait for journal recovery\n");
934 spin_lock(&ls
->ls_recover_spin
);
935 block_gen
= ls
->ls_recover_block
;
936 start_gen
= ls
->ls_recover_start
;
937 mount_gen
= ls
->ls_recover_mount
;
939 if (lvb_gen
< mount_gen
) {
940 /* wait for mounted nodes to update control_lock lvb to our
941 generation, which might include new recovery bits set */
942 if (sdp
->sd_args
.ar_spectator
) {
943 fs_info(sdp
, "Recovery is required. Waiting for a "
944 "non-spectator to mount.\n");
945 msleep_interruptible(1000);
947 fs_info(sdp
, "control_mount wait1 block %u start %u "
948 "mount %u lvb %u flags %lx\n", block_gen
,
949 start_gen
, mount_gen
, lvb_gen
,
950 ls
->ls_recover_flags
);
952 spin_unlock(&ls
->ls_recover_spin
);
956 if (lvb_gen
!= start_gen
) {
957 /* wait for mounted nodes to update control_lock lvb to the
958 latest recovery generation */
959 fs_info(sdp
, "control_mount wait2 block %u start %u mount %u "
960 "lvb %u flags %lx\n", block_gen
, start_gen
, mount_gen
,
961 lvb_gen
, ls
->ls_recover_flags
);
962 spin_unlock(&ls
->ls_recover_spin
);
966 if (block_gen
== start_gen
) {
967 /* dlm recovery in progress, wait for it to finish */
968 fs_info(sdp
, "control_mount wait3 block %u start %u mount %u "
969 "lvb %u flags %lx\n", block_gen
, start_gen
, mount_gen
,
970 lvb_gen
, ls
->ls_recover_flags
);
971 spin_unlock(&ls
->ls_recover_spin
);
975 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
976 set_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
);
977 memset(ls
->ls_recover_submit
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
978 memset(ls
->ls_recover_result
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
979 spin_unlock(&ls
->ls_recover_spin
);
988 static int control_first_done(struct gfs2_sbd
*sdp
)
990 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
991 uint32_t start_gen
, block_gen
;
995 spin_lock(&ls
->ls_recover_spin
);
996 start_gen
= ls
->ls_recover_start
;
997 block_gen
= ls
->ls_recover_block
;
999 if (test_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
) ||
1000 !test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
1001 !test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
1002 /* sanity check, should not happen */
1003 fs_err(sdp
, "control_first_done start %u block %u flags %lx\n",
1004 start_gen
, block_gen
, ls
->ls_recover_flags
);
1005 spin_unlock(&ls
->ls_recover_spin
);
1006 control_unlock(sdp
);
1010 if (start_gen
== block_gen
) {
1012 * Wait for the end of a dlm recovery cycle to switch from
1013 * first mounter recovery. We can ignore any recover_slot
1014 * callbacks between the recover_prep and next recover_done
1015 * because we are still the first mounter and any failed nodes
1016 * have not fully mounted, so they don't need recovery.
1018 spin_unlock(&ls
->ls_recover_spin
);
1019 fs_info(sdp
, "control_first_done wait gen %u\n", start_gen
);
1021 wait_on_bit(&ls
->ls_recover_flags
, DFL_DLM_RECOVERY
,
1022 TASK_UNINTERRUPTIBLE
);
1026 clear_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
1027 set_bit(DFL_FIRST_MOUNT_DONE
, &ls
->ls_recover_flags
);
1028 memset(ls
->ls_recover_submit
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
1029 memset(ls
->ls_recover_result
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
1030 spin_unlock(&ls
->ls_recover_spin
);
1032 memset(ls
->ls_lvb_bits
, 0, GDLM_LVB_SIZE
);
1033 control_lvb_write(ls
, start_gen
, ls
->ls_lvb_bits
);
1035 error
= mounted_lock(sdp
, DLM_LOCK_PR
, DLM_LKF_CONVERT
);
1037 fs_err(sdp
, "control_first_done mounted PR error %d\n", error
);
1039 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
|DLM_LKF_VALBLK
);
1041 fs_err(sdp
, "control_first_done control NL error %d\n", error
);
1047 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1048 * to accomodate the largest slot number. (NB dlm slot numbers start at 1,
1049 * gfs2 jids start at 0, so jid = slot - 1)
1052 #define RECOVER_SIZE_INC 16
1054 static int set_recover_size(struct gfs2_sbd
*sdp
, struct dlm_slot
*slots
,
1057 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1058 uint32_t *submit
= NULL
;
1059 uint32_t *result
= NULL
;
1060 uint32_t old_size
, new_size
;
1063 if (!ls
->ls_lvb_bits
) {
1064 ls
->ls_lvb_bits
= kzalloc(GDLM_LVB_SIZE
, GFP_NOFS
);
1065 if (!ls
->ls_lvb_bits
)
1070 for (i
= 0; i
< num_slots
; i
++) {
1071 if (max_jid
< slots
[i
].slot
- 1)
1072 max_jid
= slots
[i
].slot
- 1;
1075 old_size
= ls
->ls_recover_size
;
1076 new_size
= old_size
;
1077 while (new_size
< max_jid
+ 1)
1078 new_size
+= RECOVER_SIZE_INC
;
1079 if (new_size
== old_size
)
1082 submit
= kcalloc(new_size
, sizeof(uint32_t), GFP_NOFS
);
1083 result
= kcalloc(new_size
, sizeof(uint32_t), GFP_NOFS
);
1084 if (!submit
|| !result
) {
1090 spin_lock(&ls
->ls_recover_spin
);
1091 memcpy(submit
, ls
->ls_recover_submit
, old_size
* sizeof(uint32_t));
1092 memcpy(result
, ls
->ls_recover_result
, old_size
* sizeof(uint32_t));
1093 kfree(ls
->ls_recover_submit
);
1094 kfree(ls
->ls_recover_result
);
1095 ls
->ls_recover_submit
= submit
;
1096 ls
->ls_recover_result
= result
;
1097 ls
->ls_recover_size
= new_size
;
1098 spin_unlock(&ls
->ls_recover_spin
);
1102 static void free_recover_size(struct lm_lockstruct
*ls
)
1104 kfree(ls
->ls_lvb_bits
);
1105 kfree(ls
->ls_recover_submit
);
1106 kfree(ls
->ls_recover_result
);
1107 ls
->ls_recover_submit
= NULL
;
1108 ls
->ls_recover_result
= NULL
;
1109 ls
->ls_recover_size
= 0;
1110 ls
->ls_lvb_bits
= NULL
;
1113 /* dlm calls before it does lock recovery */
1115 static void gdlm_recover_prep(void *arg
)
1117 struct gfs2_sbd
*sdp
= arg
;
1118 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1120 if (gfs2_withdrawn(sdp
)) {
1121 fs_err(sdp
, "recover_prep ignored due to withdraw.\n");
1124 spin_lock(&ls
->ls_recover_spin
);
1125 ls
->ls_recover_block
= ls
->ls_recover_start
;
1126 set_bit(DFL_DLM_RECOVERY
, &ls
->ls_recover_flags
);
1128 if (!test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
1129 test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
1130 spin_unlock(&ls
->ls_recover_spin
);
1133 set_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
1134 spin_unlock(&ls
->ls_recover_spin
);
1137 /* dlm calls after recover_prep has been completed on all lockspace members;
1138 identifies slot/jid of failed member */
1140 static void gdlm_recover_slot(void *arg
, struct dlm_slot
*slot
)
1142 struct gfs2_sbd
*sdp
= arg
;
1143 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1144 int jid
= slot
->slot
- 1;
1146 if (gfs2_withdrawn(sdp
)) {
1147 fs_err(sdp
, "recover_slot jid %d ignored due to withdraw.\n",
1151 spin_lock(&ls
->ls_recover_spin
);
1152 if (ls
->ls_recover_size
< jid
+ 1) {
1153 fs_err(sdp
, "recover_slot jid %d gen %u short size %d\n",
1154 jid
, ls
->ls_recover_block
, ls
->ls_recover_size
);
1155 spin_unlock(&ls
->ls_recover_spin
);
1159 if (ls
->ls_recover_submit
[jid
]) {
1160 fs_info(sdp
, "recover_slot jid %d gen %u prev %u\n",
1161 jid
, ls
->ls_recover_block
, ls
->ls_recover_submit
[jid
]);
1163 ls
->ls_recover_submit
[jid
] = ls
->ls_recover_block
;
1164 spin_unlock(&ls
->ls_recover_spin
);
1167 /* dlm calls after recover_slot and after it completes lock recovery */
1169 static void gdlm_recover_done(void *arg
, struct dlm_slot
*slots
, int num_slots
,
1170 int our_slot
, uint32_t generation
)
1172 struct gfs2_sbd
*sdp
= arg
;
1173 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1175 if (gfs2_withdrawn(sdp
)) {
1176 fs_err(sdp
, "recover_done ignored due to withdraw.\n");
1179 /* ensure the ls jid arrays are large enough */
1180 set_recover_size(sdp
, slots
, num_slots
);
1182 spin_lock(&ls
->ls_recover_spin
);
1183 ls
->ls_recover_start
= generation
;
1185 if (!ls
->ls_recover_mount
) {
1186 ls
->ls_recover_mount
= generation
;
1187 ls
->ls_jid
= our_slot
- 1;
1190 if (!test_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
))
1191 queue_delayed_work(gfs2_control_wq
, &sdp
->sd_control_work
, 0);
1193 clear_bit(DFL_DLM_RECOVERY
, &ls
->ls_recover_flags
);
1194 smp_mb__after_atomic();
1195 wake_up_bit(&ls
->ls_recover_flags
, DFL_DLM_RECOVERY
);
1196 spin_unlock(&ls
->ls_recover_spin
);
1199 /* gfs2_recover thread has a journal recovery result */
1201 static void gdlm_recovery_result(struct gfs2_sbd
*sdp
, unsigned int jid
,
1202 unsigned int result
)
1204 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1206 if (gfs2_withdrawn(sdp
)) {
1207 fs_err(sdp
, "recovery_result jid %d ignored due to withdraw.\n",
1211 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1214 /* don't care about the recovery of own journal during mount */
1215 if (jid
== ls
->ls_jid
)
1218 spin_lock(&ls
->ls_recover_spin
);
1219 if (test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
1220 spin_unlock(&ls
->ls_recover_spin
);
1223 if (ls
->ls_recover_size
< jid
+ 1) {
1224 fs_err(sdp
, "recovery_result jid %d short size %d\n",
1225 jid
, ls
->ls_recover_size
);
1226 spin_unlock(&ls
->ls_recover_spin
);
1230 fs_info(sdp
, "recover jid %d result %s\n", jid
,
1231 result
== LM_RD_GAVEUP
? "busy" : "success");
1233 ls
->ls_recover_result
[jid
] = result
;
1235 /* GAVEUP means another node is recovering the journal; delay our
1236 next attempt to recover it, to give the other node a chance to
1237 finish before trying again */
1239 if (!test_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
))
1240 queue_delayed_work(gfs2_control_wq
, &sdp
->sd_control_work
,
1241 result
== LM_RD_GAVEUP
? HZ
: 0);
1242 spin_unlock(&ls
->ls_recover_spin
);
1245 static const struct dlm_lockspace_ops gdlm_lockspace_ops
= {
1246 .recover_prep
= gdlm_recover_prep
,
1247 .recover_slot
= gdlm_recover_slot
,
1248 .recover_done
= gdlm_recover_done
,
1251 static int gdlm_mount(struct gfs2_sbd
*sdp
, const char *table
)
1253 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1254 char cluster
[GFS2_LOCKNAME_LEN
];
1257 int error
, ops_result
;
1260 * initialize everything
1263 INIT_DELAYED_WORK(&sdp
->sd_control_work
, gfs2_control_func
);
1264 spin_lock_init(&ls
->ls_recover_spin
);
1265 ls
->ls_recover_flags
= 0;
1266 ls
->ls_recover_mount
= 0;
1267 ls
->ls_recover_start
= 0;
1268 ls
->ls_recover_block
= 0;
1269 ls
->ls_recover_size
= 0;
1270 ls
->ls_recover_submit
= NULL
;
1271 ls
->ls_recover_result
= NULL
;
1272 ls
->ls_lvb_bits
= NULL
;
1274 error
= set_recover_size(sdp
, NULL
, 0);
1279 * prepare dlm_new_lockspace args
1282 fsname
= strchr(table
, ':');
1284 fs_info(sdp
, "no fsname found\n");
1288 memset(cluster
, 0, sizeof(cluster
));
1289 memcpy(cluster
, table
, strlen(table
) - strlen(fsname
));
1292 flags
= DLM_LSFL_FS
| DLM_LSFL_NEWEXCL
;
1295 * create/join lockspace
1298 error
= dlm_new_lockspace(fsname
, cluster
, flags
, GDLM_LVB_SIZE
,
1299 &gdlm_lockspace_ops
, sdp
, &ops_result
,
1302 fs_err(sdp
, "dlm_new_lockspace error %d\n", error
);
1306 if (ops_result
< 0) {
1308 * dlm does not support ops callbacks,
1309 * old dlm_controld/gfs_controld are used, try without ops.
1311 fs_info(sdp
, "dlm lockspace ops not used\n");
1312 free_recover_size(ls
);
1313 set_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
);
1317 if (!test_bit(SDF_NOJOURNALID
, &sdp
->sd_flags
)) {
1318 fs_err(sdp
, "dlm lockspace ops disallow jid preset\n");
1324 * control_mount() uses control_lock to determine first mounter,
1325 * and for later mounts, waits for any recoveries to be cleared.
1328 error
= control_mount(sdp
);
1330 fs_err(sdp
, "mount control error %d\n", error
);
1334 ls
->ls_first
= !!test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
1335 clear_bit(SDF_NOJOURNALID
, &sdp
->sd_flags
);
1336 smp_mb__after_atomic();
1337 wake_up_bit(&sdp
->sd_flags
, SDF_NOJOURNALID
);
1341 dlm_release_lockspace(ls
->ls_dlm
, 2);
1343 free_recover_size(ls
);
1348 static void gdlm_first_done(struct gfs2_sbd
*sdp
)
1350 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1353 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1356 error
= control_first_done(sdp
);
1358 fs_err(sdp
, "mount first_done error %d\n", error
);
1361 static void gdlm_unmount(struct gfs2_sbd
*sdp
)
1363 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1365 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1368 /* wait for gfs2_control_wq to be done with this mount */
1370 spin_lock(&ls
->ls_recover_spin
);
1371 set_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
);
1372 spin_unlock(&ls
->ls_recover_spin
);
1373 flush_delayed_work(&sdp
->sd_control_work
);
1375 /* mounted_lock and control_lock will be purged in dlm recovery */
1378 dlm_release_lockspace(ls
->ls_dlm
, 2);
1382 free_recover_size(ls
);
1385 static const match_table_t dlm_tokens
= {
1386 { Opt_jid
, "jid=%d"},
1388 { Opt_first
, "first=%d"},
1389 { Opt_nodir
, "nodir=%d"},
1393 const struct lm_lockops gfs2_dlm_ops
= {
1394 .lm_proto_name
= "lock_dlm",
1395 .lm_mount
= gdlm_mount
,
1396 .lm_first_done
= gdlm_first_done
,
1397 .lm_recovery_result
= gdlm_recovery_result
,
1398 .lm_unmount
= gdlm_unmount
,
1399 .lm_put_lock
= gdlm_put_lock
,
1400 .lm_lock
= gdlm_lock
,
1401 .lm_cancel
= gdlm_cancel
,
1402 .lm_tokens
= &dlm_tokens
,