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
] = "";
266 req
= make_mode(gl
->gl_name
.ln_sbd
, req_state
);
267 lkf
= make_flags(gl
, flags
, req
);
268 gfs2_glstats_inc(gl
, GFS2_LKS_DCOUNT
);
269 gfs2_sbstats_inc(gl
, GFS2_LKS_DCOUNT
);
270 if (gl
->gl_lksb
.sb_lkid
) {
271 gfs2_update_request_times(gl
);
273 memset(strname
, ' ', GDLM_STRNAME_BYTES
- 1);
274 strname
[GDLM_STRNAME_BYTES
- 1] = '\0';
275 gfs2_reverse_hex(strname
+ 7, gl
->gl_name
.ln_type
);
276 gfs2_reverse_hex(strname
+ 23, gl
->gl_name
.ln_number
);
277 gl
->gl_dstamp
= ktime_get_real();
280 * Submit the actual lock request.
284 error
= dlm_lock(ls
->ls_dlm
, req
, &gl
->gl_lksb
, lkf
, strname
,
285 GDLM_STRNAME_BYTES
- 1, 0, gdlm_ast
, gl
, gdlm_bast
);
286 if (error
== -EBUSY
) {
293 static void gdlm_put_lock(struct gfs2_glock
*gl
)
295 struct gfs2_sbd
*sdp
= gl
->gl_name
.ln_sbd
;
296 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
299 if (gl
->gl_lksb
.sb_lkid
== 0) {
304 clear_bit(GLF_BLOCKING
, &gl
->gl_flags
);
305 gfs2_glstats_inc(gl
, GFS2_LKS_DCOUNT
);
306 gfs2_sbstats_inc(gl
, GFS2_LKS_DCOUNT
);
307 gfs2_update_request_times(gl
);
309 /* don't want to call dlm if we've unmounted the lock protocol */
310 if (test_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
)) {
314 /* don't want to skip dlm_unlock writing the lvb when lock has one */
316 if (test_bit(SDF_SKIP_DLM_UNLOCK
, &sdp
->sd_flags
) &&
317 !gl
->gl_lksb
.sb_lvbptr
) {
323 error
= dlm_unlock(ls
->ls_dlm
, gl
->gl_lksb
.sb_lkid
, DLM_LKF_VALBLK
,
325 if (error
== -EBUSY
) {
331 fs_err(sdp
, "gdlm_unlock %x,%llx err=%d\n",
333 (unsigned long long)gl
->gl_name
.ln_number
, error
);
338 static void gdlm_cancel(struct gfs2_glock
*gl
)
340 struct lm_lockstruct
*ls
= &gl
->gl_name
.ln_sbd
->sd_lockstruct
;
341 dlm_unlock(ls
->ls_dlm
, gl
->gl_lksb
.sb_lkid
, DLM_LKF_CANCEL
, NULL
, gl
);
345 * dlm/gfs2 recovery coordination using dlm_recover callbacks
347 * 0. gfs2 checks for another cluster node withdraw, needing journal replay
348 * 1. dlm_controld sees lockspace members change
349 * 2. dlm_controld blocks dlm-kernel locking activity
350 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
351 * 4. dlm_controld starts and finishes its own user level recovery
352 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
353 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
354 * 7. dlm_recoverd does its own lock recovery
355 * 8. dlm_recoverd unblocks dlm-kernel locking activity
356 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
357 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
358 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
359 * 12. gfs2_recover dequeues and recovers journals of failed nodes
360 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
361 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
362 * 15. gfs2_control unblocks normal locking when all journals are recovered
364 * - failures during recovery
366 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
367 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
368 * recovering for a prior failure. gfs2_control needs a way to detect
369 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
370 * the recover_block and recover_start values.
372 * recover_done() provides a new lockspace generation number each time it
373 * is called (step 9). This generation number is saved as recover_start.
374 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
375 * recover_block = recover_start. So, while recover_block is equal to
376 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
377 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
379 * - more specific gfs2 steps in sequence above
381 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
382 * 6. recover_slot records any failed jids (maybe none)
383 * 9. recover_done sets recover_start = new generation number
384 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
385 * 12. gfs2_recover does journal recoveries for failed jids identified above
386 * 14. gfs2_control clears control_lock lvb bits for recovered jids
387 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
388 * again) then do nothing, otherwise if recover_start > recover_block
389 * then clear BLOCK_LOCKS.
391 * - parallel recovery steps across all nodes
393 * All nodes attempt to update the control_lock lvb with the new generation
394 * number and jid bits, but only the first to get the control_lock EX will
395 * do so; others will see that it's already done (lvb already contains new
396 * generation number.)
398 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
399 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
400 * . One node gets control_lock first and writes the lvb, others see it's done
401 * . All nodes attempt to recover jids for which they see control_lock bits set
402 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
403 * . All nodes will eventually see all lvb bits clear and unblock locks
405 * - is there a problem with clearing an lvb bit that should be set
406 * and missing a journal recovery?
409 * 2. lvb bit set for step 1
410 * 3. jid recovered for step 1
411 * 4. jid taken again (new mount)
412 * 5. jid fails (for step 4)
413 * 6. lvb bit set for step 5 (will already be set)
414 * 7. lvb bit cleared for step 3
416 * This is not a problem because the failure in step 5 does not
417 * require recovery, because the mount in step 4 could not have
418 * progressed far enough to unblock locks and access the fs. The
419 * control_mount() function waits for all recoveries to be complete
420 * for the latest lockspace generation before ever unblocking locks
421 * and returning. The mount in step 4 waits until the recovery in
424 * - special case of first mounter: first node to mount the fs
426 * The first node to mount a gfs2 fs needs to check all the journals
427 * and recover any that need recovery before other nodes are allowed
428 * to mount the fs. (Others may begin mounting, but they must wait
429 * for the first mounter to be done before taking locks on the fs
430 * or accessing the fs.) This has two parts:
432 * 1. The mounted_lock tells a node it's the first to mount the fs.
433 * Each node holds the mounted_lock in PR while it's mounted.
434 * Each node tries to acquire the mounted_lock in EX when it mounts.
435 * If a node is granted the mounted_lock EX it means there are no
436 * other mounted nodes (no PR locks exist), and it is the first mounter.
437 * The mounted_lock is demoted to PR when first recovery is done, so
438 * others will fail to get an EX lock, but will get a PR lock.
440 * 2. The control_lock blocks others in control_mount() while the first
441 * mounter is doing first mount recovery of all journals.
442 * A mounting node needs to acquire control_lock in EX mode before
443 * it can proceed. The first mounter holds control_lock in EX while doing
444 * the first mount recovery, blocking mounts from other nodes, then demotes
445 * control_lock to NL when it's done (others_may_mount/first_done),
446 * allowing other nodes to continue mounting.
449 * control_lock EX/NOQUEUE success
450 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
452 * do first mounter recovery
453 * mounted_lock EX->PR
454 * control_lock EX->NL, write lvb generation
457 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
458 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
459 * mounted_lock PR/NOQUEUE success
460 * read lvb generation
461 * control_lock EX->NL
464 * - mount during recovery
466 * If a node mounts while others are doing recovery (not first mounter),
467 * the mounting node will get its initial recover_done() callback without
468 * having seen any previous failures/callbacks.
470 * It must wait for all recoveries preceding its mount to be finished
471 * before it unblocks locks. It does this by repeating the "other mounter"
472 * steps above until the lvb generation number is >= its mount generation
473 * number (from initial recover_done) and all lvb bits are clear.
475 * - control_lock lvb format
477 * 4 bytes generation number: the latest dlm lockspace generation number
478 * from recover_done callback. Indicates the jid bitmap has been updated
479 * to reflect all slot failures through that generation.
481 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
482 * that jid N needs recovery.
485 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
487 static void control_lvb_read(struct lm_lockstruct
*ls
, uint32_t *lvb_gen
,
491 memcpy(lvb_bits
, ls
->ls_control_lvb
, GDLM_LVB_SIZE
);
492 memcpy(&gen
, lvb_bits
, sizeof(__le32
));
493 *lvb_gen
= le32_to_cpu(gen
);
496 static void control_lvb_write(struct lm_lockstruct
*ls
, uint32_t lvb_gen
,
500 memcpy(ls
->ls_control_lvb
, lvb_bits
, GDLM_LVB_SIZE
);
501 gen
= cpu_to_le32(lvb_gen
);
502 memcpy(ls
->ls_control_lvb
, &gen
, sizeof(__le32
));
505 static int all_jid_bits_clear(char *lvb
)
507 return !memchr_inv(lvb
+ JID_BITMAP_OFFSET
, 0,
508 GDLM_LVB_SIZE
- JID_BITMAP_OFFSET
);
511 static void sync_wait_cb(void *arg
)
513 struct lm_lockstruct
*ls
= arg
;
514 complete(&ls
->ls_sync_wait
);
517 static int sync_unlock(struct gfs2_sbd
*sdp
, struct dlm_lksb
*lksb
, char *name
)
519 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
522 error
= dlm_unlock(ls
->ls_dlm
, lksb
->sb_lkid
, 0, lksb
, ls
);
524 fs_err(sdp
, "%s lkid %x error %d\n",
525 name
, lksb
->sb_lkid
, error
);
529 wait_for_completion(&ls
->ls_sync_wait
);
531 if (lksb
->sb_status
!= -DLM_EUNLOCK
) {
532 fs_err(sdp
, "%s lkid %x status %d\n",
533 name
, lksb
->sb_lkid
, lksb
->sb_status
);
539 static int sync_lock(struct gfs2_sbd
*sdp
, int mode
, uint32_t flags
,
540 unsigned int num
, struct dlm_lksb
*lksb
, char *name
)
542 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
543 char strname
[GDLM_STRNAME_BYTES
];
546 memset(strname
, 0, GDLM_STRNAME_BYTES
);
547 snprintf(strname
, GDLM_STRNAME_BYTES
, "%8x%16x", LM_TYPE_NONDISK
, num
);
549 error
= dlm_lock(ls
->ls_dlm
, mode
, lksb
, flags
,
550 strname
, GDLM_STRNAME_BYTES
- 1,
551 0, sync_wait_cb
, ls
, NULL
);
553 fs_err(sdp
, "%s lkid %x flags %x mode %d error %d\n",
554 name
, lksb
->sb_lkid
, flags
, mode
, error
);
558 wait_for_completion(&ls
->ls_sync_wait
);
560 status
= lksb
->sb_status
;
562 if (status
&& status
!= -EAGAIN
) {
563 fs_err(sdp
, "%s lkid %x flags %x mode %d status %d\n",
564 name
, lksb
->sb_lkid
, flags
, mode
, status
);
570 static int mounted_unlock(struct gfs2_sbd
*sdp
)
572 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
573 return sync_unlock(sdp
, &ls
->ls_mounted_lksb
, "mounted_lock");
576 static int mounted_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_MOUNTED_LOCK
,
580 &ls
->ls_mounted_lksb
, "mounted_lock");
583 static int control_unlock(struct gfs2_sbd
*sdp
)
585 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
586 return sync_unlock(sdp
, &ls
->ls_control_lksb
, "control_lock");
589 static int control_lock(struct gfs2_sbd
*sdp
, int mode
, uint32_t flags
)
591 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
592 return sync_lock(sdp
, mode
, flags
, GFS2_CONTROL_LOCK
,
593 &ls
->ls_control_lksb
, "control_lock");
597 * remote_withdraw - react to a node withdrawing from the file system
598 * @sdp: The superblock
600 static void remote_withdraw(struct gfs2_sbd
*sdp
)
602 struct gfs2_jdesc
*jd
;
603 int ret
= 0, count
= 0;
605 list_for_each_entry(jd
, &sdp
->sd_jindex_list
, jd_list
) {
606 if (jd
->jd_jid
== sdp
->sd_lockstruct
.ls_jid
)
608 ret
= gfs2_recover_journal(jd
, true);
614 /* Now drop the additional reference we acquired */
615 fs_err(sdp
, "Journals checked: %d, ret = %d.\n", count
, ret
);
618 static void gfs2_control_func(struct work_struct
*work
)
620 struct gfs2_sbd
*sdp
= container_of(work
, struct gfs2_sbd
, sd_control_work
.work
);
621 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
622 uint32_t block_gen
, start_gen
, lvb_gen
, flags
;
628 /* First check for other nodes that may have done a withdraw. */
629 if (test_bit(SDF_REMOTE_WITHDRAW
, &sdp
->sd_flags
)) {
630 remote_withdraw(sdp
);
631 clear_bit(SDF_REMOTE_WITHDRAW
, &sdp
->sd_flags
);
635 spin_lock(&ls
->ls_recover_spin
);
637 * No MOUNT_DONE means we're still mounting; control_mount()
638 * will set this flag, after which this thread will take over
639 * all further clearing of BLOCK_LOCKS.
641 * FIRST_MOUNT means this node is doing first mounter recovery,
642 * for which recovery control is handled by
643 * control_mount()/control_first_done(), not this thread.
645 if (!test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
646 test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
647 spin_unlock(&ls
->ls_recover_spin
);
650 block_gen
= ls
->ls_recover_block
;
651 start_gen
= ls
->ls_recover_start
;
652 spin_unlock(&ls
->ls_recover_spin
);
655 * Equal block_gen and start_gen implies we are between
656 * recover_prep and recover_done callbacks, which means
657 * dlm recovery is in progress and dlm locking is blocked.
658 * There's no point trying to do any work until recover_done.
661 if (block_gen
== start_gen
)
665 * Propagate recover_submit[] and recover_result[] to lvb:
666 * dlm_recoverd adds to recover_submit[] jids needing recovery
667 * gfs2_recover adds to recover_result[] journal recovery results
669 * set lvb bit for jids in recover_submit[] if the lvb has not
670 * yet been updated for the generation of the failure
672 * clear lvb bit for jids in recover_result[] if the result of
673 * the journal recovery is SUCCESS
676 error
= control_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_VALBLK
);
678 fs_err(sdp
, "control lock EX error %d\n", error
);
682 control_lvb_read(ls
, &lvb_gen
, ls
->ls_lvb_bits
);
684 spin_lock(&ls
->ls_recover_spin
);
685 if (block_gen
!= ls
->ls_recover_block
||
686 start_gen
!= ls
->ls_recover_start
) {
687 fs_info(sdp
, "recover generation %u block1 %u %u\n",
688 start_gen
, block_gen
, ls
->ls_recover_block
);
689 spin_unlock(&ls
->ls_recover_spin
);
690 control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
694 recover_size
= ls
->ls_recover_size
;
696 if (lvb_gen
<= start_gen
) {
698 * Clear lvb bits for jids we've successfully recovered.
699 * Because all nodes attempt to recover failed journals,
700 * a journal can be recovered multiple times successfully
701 * in succession. Only the first will really do recovery,
702 * the others find it clean, but still report a successful
703 * recovery. So, another node may have already recovered
704 * the jid and cleared the lvb bit for it.
706 for (i
= 0; i
< recover_size
; i
++) {
707 if (ls
->ls_recover_result
[i
] != LM_RD_SUCCESS
)
710 ls
->ls_recover_result
[i
] = 0;
712 if (!test_bit_le(i
, ls
->ls_lvb_bits
+ JID_BITMAP_OFFSET
))
715 __clear_bit_le(i
, ls
->ls_lvb_bits
+ JID_BITMAP_OFFSET
);
720 if (lvb_gen
== start_gen
) {
722 * Failed slots before start_gen are already set in lvb.
724 for (i
= 0; i
< recover_size
; i
++) {
725 if (!ls
->ls_recover_submit
[i
])
727 if (ls
->ls_recover_submit
[i
] < lvb_gen
)
728 ls
->ls_recover_submit
[i
] = 0;
730 } else if (lvb_gen
< start_gen
) {
732 * Failed slots before start_gen are not yet set in lvb.
734 for (i
= 0; i
< recover_size
; i
++) {
735 if (!ls
->ls_recover_submit
[i
])
737 if (ls
->ls_recover_submit
[i
] < start_gen
) {
738 ls
->ls_recover_submit
[i
] = 0;
739 __set_bit_le(i
, ls
->ls_lvb_bits
+ JID_BITMAP_OFFSET
);
742 /* even if there are no bits to set, we need to write the
743 latest generation to the lvb */
747 * we should be getting a recover_done() for lvb_gen soon
750 spin_unlock(&ls
->ls_recover_spin
);
753 control_lvb_write(ls
, start_gen
, ls
->ls_lvb_bits
);
754 flags
= DLM_LKF_CONVERT
| DLM_LKF_VALBLK
;
756 flags
= DLM_LKF_CONVERT
;
759 error
= control_lock(sdp
, DLM_LOCK_NL
, flags
);
761 fs_err(sdp
, "control lock NL error %d\n", error
);
766 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
767 * and clear a jid bit in the lvb if the recovery is a success.
768 * Eventually all journals will be recovered, all jid bits will
769 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
772 for (i
= 0; i
< recover_size
; i
++) {
773 if (test_bit_le(i
, ls
->ls_lvb_bits
+ JID_BITMAP_OFFSET
)) {
774 fs_info(sdp
, "recover generation %u jid %d\n",
776 gfs2_recover_set(sdp
, i
);
784 * No more jid bits set in lvb, all recovery is done, unblock locks
785 * (unless a new recover_prep callback has occured blocking locks
786 * again while working above)
789 spin_lock(&ls
->ls_recover_spin
);
790 if (ls
->ls_recover_block
== block_gen
&&
791 ls
->ls_recover_start
== start_gen
) {
792 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
793 spin_unlock(&ls
->ls_recover_spin
);
794 fs_info(sdp
, "recover generation %u done\n", start_gen
);
795 gfs2_glock_thaw(sdp
);
797 fs_info(sdp
, "recover generation %u block2 %u %u\n",
798 start_gen
, block_gen
, ls
->ls_recover_block
);
799 spin_unlock(&ls
->ls_recover_spin
);
803 static int control_mount(struct gfs2_sbd
*sdp
)
805 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
806 uint32_t start_gen
, block_gen
, mount_gen
, lvb_gen
;
811 memset(&ls
->ls_mounted_lksb
, 0, sizeof(struct dlm_lksb
));
812 memset(&ls
->ls_control_lksb
, 0, sizeof(struct dlm_lksb
));
813 memset(&ls
->ls_control_lvb
, 0, GDLM_LVB_SIZE
);
814 ls
->ls_control_lksb
.sb_lvbptr
= ls
->ls_control_lvb
;
815 init_completion(&ls
->ls_sync_wait
);
817 set_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
819 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_VALBLK
);
821 fs_err(sdp
, "control_mount control_lock NL error %d\n", error
);
825 error
= mounted_lock(sdp
, DLM_LOCK_NL
, 0);
827 fs_err(sdp
, "control_mount mounted_lock NL error %d\n", error
);
831 mounted_mode
= DLM_LOCK_NL
;
834 if (retries
++ && signal_pending(current
)) {
840 * We always start with both locks in NL. control_lock is
841 * demoted to NL below so we don't need to do it here.
844 if (mounted_mode
!= DLM_LOCK_NL
) {
845 error
= mounted_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
848 mounted_mode
= DLM_LOCK_NL
;
852 * Other nodes need to do some work in dlm recovery and gfs2_control
853 * before the recover_done and control_lock will be ready for us below.
854 * A delay here is not required but often avoids having to retry.
857 msleep_interruptible(500);
860 * Acquire control_lock in EX and mounted_lock in either EX or PR.
861 * control_lock lvb keeps track of any pending journal recoveries.
862 * mounted_lock indicates if any other nodes have the fs mounted.
865 error
= control_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
|DLM_LKF_VALBLK
);
866 if (error
== -EAGAIN
) {
869 fs_err(sdp
, "control_mount control_lock EX error %d\n", error
);
874 * If we're a spectator, we don't want to take the lock in EX because
875 * we cannot do the first-mount responsibility it implies: recovery.
877 if (sdp
->sd_args
.ar_spectator
)
880 error
= mounted_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
);
882 mounted_mode
= DLM_LOCK_EX
;
884 } else if (error
!= -EAGAIN
) {
885 fs_err(sdp
, "control_mount mounted_lock EX error %d\n", error
);
889 error
= mounted_lock(sdp
, DLM_LOCK_PR
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
);
891 mounted_mode
= DLM_LOCK_PR
;
894 /* not even -EAGAIN should happen here */
895 fs_err(sdp
, "control_mount mounted_lock PR error %d\n", error
);
901 * If we got both locks above in EX, then we're the first mounter.
902 * If not, then we need to wait for the control_lock lvb to be
903 * updated by other mounted nodes to reflect our mount generation.
905 * In simple first mounter cases, first mounter will see zero lvb_gen,
906 * but in cases where all existing nodes leave/fail before mounting
907 * nodes finish control_mount, then all nodes will be mounting and
908 * lvb_gen will be non-zero.
911 control_lvb_read(ls
, &lvb_gen
, ls
->ls_lvb_bits
);
913 if (lvb_gen
== 0xFFFFFFFF) {
914 /* special value to force mount attempts to fail */
915 fs_err(sdp
, "control_mount control_lock disabled\n");
920 if (mounted_mode
== DLM_LOCK_EX
) {
921 /* first mounter, keep both EX while doing first recovery */
922 spin_lock(&ls
->ls_recover_spin
);
923 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
924 set_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
);
925 set_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
926 spin_unlock(&ls
->ls_recover_spin
);
927 fs_info(sdp
, "first mounter control generation %u\n", lvb_gen
);
931 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
936 * We are not first mounter, now we need to wait for the control_lock
937 * lvb generation to be >= the generation from our first recover_done
938 * and all lvb bits to be clear (no pending journal recoveries.)
941 if (!all_jid_bits_clear(ls
->ls_lvb_bits
)) {
942 /* journals need recovery, wait until all are clear */
943 fs_info(sdp
, "control_mount wait for journal recovery\n");
947 spin_lock(&ls
->ls_recover_spin
);
948 block_gen
= ls
->ls_recover_block
;
949 start_gen
= ls
->ls_recover_start
;
950 mount_gen
= ls
->ls_recover_mount
;
952 if (lvb_gen
< mount_gen
) {
953 /* wait for mounted nodes to update control_lock lvb to our
954 generation, which might include new recovery bits set */
955 if (sdp
->sd_args
.ar_spectator
) {
956 fs_info(sdp
, "Recovery is required. Waiting for a "
957 "non-spectator to mount.\n");
958 msleep_interruptible(1000);
960 fs_info(sdp
, "control_mount wait1 block %u start %u "
961 "mount %u lvb %u flags %lx\n", block_gen
,
962 start_gen
, mount_gen
, lvb_gen
,
963 ls
->ls_recover_flags
);
965 spin_unlock(&ls
->ls_recover_spin
);
969 if (lvb_gen
!= start_gen
) {
970 /* wait for mounted nodes to update control_lock lvb to the
971 latest recovery generation */
972 fs_info(sdp
, "control_mount wait2 block %u start %u mount %u "
973 "lvb %u flags %lx\n", block_gen
, start_gen
, mount_gen
,
974 lvb_gen
, ls
->ls_recover_flags
);
975 spin_unlock(&ls
->ls_recover_spin
);
979 if (block_gen
== start_gen
) {
980 /* dlm recovery in progress, wait for it to finish */
981 fs_info(sdp
, "control_mount wait3 block %u start %u mount %u "
982 "lvb %u flags %lx\n", block_gen
, start_gen
, mount_gen
,
983 lvb_gen
, ls
->ls_recover_flags
);
984 spin_unlock(&ls
->ls_recover_spin
);
988 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
989 set_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
);
990 memset(ls
->ls_recover_submit
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
991 memset(ls
->ls_recover_result
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
992 spin_unlock(&ls
->ls_recover_spin
);
1001 static int control_first_done(struct gfs2_sbd
*sdp
)
1003 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1004 uint32_t start_gen
, block_gen
;
1008 spin_lock(&ls
->ls_recover_spin
);
1009 start_gen
= ls
->ls_recover_start
;
1010 block_gen
= ls
->ls_recover_block
;
1012 if (test_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
) ||
1013 !test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
1014 !test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
1015 /* sanity check, should not happen */
1016 fs_err(sdp
, "control_first_done start %u block %u flags %lx\n",
1017 start_gen
, block_gen
, ls
->ls_recover_flags
);
1018 spin_unlock(&ls
->ls_recover_spin
);
1019 control_unlock(sdp
);
1023 if (start_gen
== block_gen
) {
1025 * Wait for the end of a dlm recovery cycle to switch from
1026 * first mounter recovery. We can ignore any recover_slot
1027 * callbacks between the recover_prep and next recover_done
1028 * because we are still the first mounter and any failed nodes
1029 * have not fully mounted, so they don't need recovery.
1031 spin_unlock(&ls
->ls_recover_spin
);
1032 fs_info(sdp
, "control_first_done wait gen %u\n", start_gen
);
1034 wait_on_bit(&ls
->ls_recover_flags
, DFL_DLM_RECOVERY
,
1035 TASK_UNINTERRUPTIBLE
);
1039 clear_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
1040 set_bit(DFL_FIRST_MOUNT_DONE
, &ls
->ls_recover_flags
);
1041 memset(ls
->ls_recover_submit
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
1042 memset(ls
->ls_recover_result
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
1043 spin_unlock(&ls
->ls_recover_spin
);
1045 memset(ls
->ls_lvb_bits
, 0, GDLM_LVB_SIZE
);
1046 control_lvb_write(ls
, start_gen
, ls
->ls_lvb_bits
);
1048 error
= mounted_lock(sdp
, DLM_LOCK_PR
, DLM_LKF_CONVERT
);
1050 fs_err(sdp
, "control_first_done mounted PR error %d\n", error
);
1052 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
|DLM_LKF_VALBLK
);
1054 fs_err(sdp
, "control_first_done control NL error %d\n", error
);
1060 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1061 * to accommodate the largest slot number. (NB dlm slot numbers start at 1,
1062 * gfs2 jids start at 0, so jid = slot - 1)
1065 #define RECOVER_SIZE_INC 16
1067 static int set_recover_size(struct gfs2_sbd
*sdp
, struct dlm_slot
*slots
,
1070 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1071 uint32_t *submit
= NULL
;
1072 uint32_t *result
= NULL
;
1073 uint32_t old_size
, new_size
;
1076 if (!ls
->ls_lvb_bits
) {
1077 ls
->ls_lvb_bits
= kzalloc(GDLM_LVB_SIZE
, GFP_NOFS
);
1078 if (!ls
->ls_lvb_bits
)
1083 for (i
= 0; i
< num_slots
; i
++) {
1084 if (max_jid
< slots
[i
].slot
- 1)
1085 max_jid
= slots
[i
].slot
- 1;
1088 old_size
= ls
->ls_recover_size
;
1089 new_size
= old_size
;
1090 while (new_size
< max_jid
+ 1)
1091 new_size
+= RECOVER_SIZE_INC
;
1092 if (new_size
== old_size
)
1095 submit
= kcalloc(new_size
, sizeof(uint32_t), GFP_NOFS
);
1096 result
= kcalloc(new_size
, sizeof(uint32_t), GFP_NOFS
);
1097 if (!submit
|| !result
) {
1103 spin_lock(&ls
->ls_recover_spin
);
1104 memcpy(submit
, ls
->ls_recover_submit
, old_size
* sizeof(uint32_t));
1105 memcpy(result
, ls
->ls_recover_result
, old_size
* sizeof(uint32_t));
1106 kfree(ls
->ls_recover_submit
);
1107 kfree(ls
->ls_recover_result
);
1108 ls
->ls_recover_submit
= submit
;
1109 ls
->ls_recover_result
= result
;
1110 ls
->ls_recover_size
= new_size
;
1111 spin_unlock(&ls
->ls_recover_spin
);
1115 static void free_recover_size(struct lm_lockstruct
*ls
)
1117 kfree(ls
->ls_lvb_bits
);
1118 kfree(ls
->ls_recover_submit
);
1119 kfree(ls
->ls_recover_result
);
1120 ls
->ls_recover_submit
= NULL
;
1121 ls
->ls_recover_result
= NULL
;
1122 ls
->ls_recover_size
= 0;
1123 ls
->ls_lvb_bits
= NULL
;
1126 /* dlm calls before it does lock recovery */
1128 static void gdlm_recover_prep(void *arg
)
1130 struct gfs2_sbd
*sdp
= arg
;
1131 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1133 if (gfs2_withdrawn(sdp
)) {
1134 fs_err(sdp
, "recover_prep ignored due to withdraw.\n");
1137 spin_lock(&ls
->ls_recover_spin
);
1138 ls
->ls_recover_block
= ls
->ls_recover_start
;
1139 set_bit(DFL_DLM_RECOVERY
, &ls
->ls_recover_flags
);
1141 if (!test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
1142 test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
1143 spin_unlock(&ls
->ls_recover_spin
);
1146 set_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
1147 spin_unlock(&ls
->ls_recover_spin
);
1150 /* dlm calls after recover_prep has been completed on all lockspace members;
1151 identifies slot/jid of failed member */
1153 static void gdlm_recover_slot(void *arg
, struct dlm_slot
*slot
)
1155 struct gfs2_sbd
*sdp
= arg
;
1156 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1157 int jid
= slot
->slot
- 1;
1159 if (gfs2_withdrawn(sdp
)) {
1160 fs_err(sdp
, "recover_slot jid %d ignored due to withdraw.\n",
1164 spin_lock(&ls
->ls_recover_spin
);
1165 if (ls
->ls_recover_size
< jid
+ 1) {
1166 fs_err(sdp
, "recover_slot jid %d gen %u short size %d\n",
1167 jid
, ls
->ls_recover_block
, ls
->ls_recover_size
);
1168 spin_unlock(&ls
->ls_recover_spin
);
1172 if (ls
->ls_recover_submit
[jid
]) {
1173 fs_info(sdp
, "recover_slot jid %d gen %u prev %u\n",
1174 jid
, ls
->ls_recover_block
, ls
->ls_recover_submit
[jid
]);
1176 ls
->ls_recover_submit
[jid
] = ls
->ls_recover_block
;
1177 spin_unlock(&ls
->ls_recover_spin
);
1180 /* dlm calls after recover_slot and after it completes lock recovery */
1182 static void gdlm_recover_done(void *arg
, struct dlm_slot
*slots
, int num_slots
,
1183 int our_slot
, uint32_t generation
)
1185 struct gfs2_sbd
*sdp
= arg
;
1186 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1188 if (gfs2_withdrawn(sdp
)) {
1189 fs_err(sdp
, "recover_done ignored due to withdraw.\n");
1192 /* ensure the ls jid arrays are large enough */
1193 set_recover_size(sdp
, slots
, num_slots
);
1195 spin_lock(&ls
->ls_recover_spin
);
1196 ls
->ls_recover_start
= generation
;
1198 if (!ls
->ls_recover_mount
) {
1199 ls
->ls_recover_mount
= generation
;
1200 ls
->ls_jid
= our_slot
- 1;
1203 if (!test_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
))
1204 queue_delayed_work(gfs2_control_wq
, &sdp
->sd_control_work
, 0);
1206 clear_bit(DFL_DLM_RECOVERY
, &ls
->ls_recover_flags
);
1207 smp_mb__after_atomic();
1208 wake_up_bit(&ls
->ls_recover_flags
, DFL_DLM_RECOVERY
);
1209 spin_unlock(&ls
->ls_recover_spin
);
1212 /* gfs2_recover thread has a journal recovery result */
1214 static void gdlm_recovery_result(struct gfs2_sbd
*sdp
, unsigned int jid
,
1215 unsigned int result
)
1217 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1219 if (gfs2_withdrawn(sdp
)) {
1220 fs_err(sdp
, "recovery_result jid %d ignored due to withdraw.\n",
1224 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1227 /* don't care about the recovery of own journal during mount */
1228 if (jid
== ls
->ls_jid
)
1231 spin_lock(&ls
->ls_recover_spin
);
1232 if (test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
1233 spin_unlock(&ls
->ls_recover_spin
);
1236 if (ls
->ls_recover_size
< jid
+ 1) {
1237 fs_err(sdp
, "recovery_result jid %d short size %d\n",
1238 jid
, ls
->ls_recover_size
);
1239 spin_unlock(&ls
->ls_recover_spin
);
1243 fs_info(sdp
, "recover jid %d result %s\n", jid
,
1244 result
== LM_RD_GAVEUP
? "busy" : "success");
1246 ls
->ls_recover_result
[jid
] = result
;
1248 /* GAVEUP means another node is recovering the journal; delay our
1249 next attempt to recover it, to give the other node a chance to
1250 finish before trying again */
1252 if (!test_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
))
1253 queue_delayed_work(gfs2_control_wq
, &sdp
->sd_control_work
,
1254 result
== LM_RD_GAVEUP
? HZ
: 0);
1255 spin_unlock(&ls
->ls_recover_spin
);
1258 static const struct dlm_lockspace_ops gdlm_lockspace_ops
= {
1259 .recover_prep
= gdlm_recover_prep
,
1260 .recover_slot
= gdlm_recover_slot
,
1261 .recover_done
= gdlm_recover_done
,
1264 static int gdlm_mount(struct gfs2_sbd
*sdp
, const char *table
)
1266 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1267 char cluster
[GFS2_LOCKNAME_LEN
];
1270 int error
, ops_result
;
1273 * initialize everything
1276 INIT_DELAYED_WORK(&sdp
->sd_control_work
, gfs2_control_func
);
1277 spin_lock_init(&ls
->ls_recover_spin
);
1278 ls
->ls_recover_flags
= 0;
1279 ls
->ls_recover_mount
= 0;
1280 ls
->ls_recover_start
= 0;
1281 ls
->ls_recover_block
= 0;
1282 ls
->ls_recover_size
= 0;
1283 ls
->ls_recover_submit
= NULL
;
1284 ls
->ls_recover_result
= NULL
;
1285 ls
->ls_lvb_bits
= NULL
;
1287 error
= set_recover_size(sdp
, NULL
, 0);
1292 * prepare dlm_new_lockspace args
1295 fsname
= strchr(table
, ':');
1297 fs_info(sdp
, "no fsname found\n");
1301 memset(cluster
, 0, sizeof(cluster
));
1302 memcpy(cluster
, table
, strlen(table
) - strlen(fsname
));
1305 flags
= DLM_LSFL_FS
| DLM_LSFL_NEWEXCL
;
1308 * create/join lockspace
1311 error
= dlm_new_lockspace(fsname
, cluster
, flags
, GDLM_LVB_SIZE
,
1312 &gdlm_lockspace_ops
, sdp
, &ops_result
,
1315 fs_err(sdp
, "dlm_new_lockspace error %d\n", error
);
1319 if (ops_result
< 0) {
1321 * dlm does not support ops callbacks,
1322 * old dlm_controld/gfs_controld are used, try without ops.
1324 fs_info(sdp
, "dlm lockspace ops not used\n");
1325 free_recover_size(ls
);
1326 set_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
);
1330 if (!test_bit(SDF_NOJOURNALID
, &sdp
->sd_flags
)) {
1331 fs_err(sdp
, "dlm lockspace ops disallow jid preset\n");
1337 * control_mount() uses control_lock to determine first mounter,
1338 * and for later mounts, waits for any recoveries to be cleared.
1341 error
= control_mount(sdp
);
1343 fs_err(sdp
, "mount control error %d\n", error
);
1347 ls
->ls_first
= !!test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
1348 clear_bit(SDF_NOJOURNALID
, &sdp
->sd_flags
);
1349 smp_mb__after_atomic();
1350 wake_up_bit(&sdp
->sd_flags
, SDF_NOJOURNALID
);
1354 dlm_release_lockspace(ls
->ls_dlm
, 2);
1356 free_recover_size(ls
);
1361 static void gdlm_first_done(struct gfs2_sbd
*sdp
)
1363 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1366 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1369 error
= control_first_done(sdp
);
1371 fs_err(sdp
, "mount first_done error %d\n", error
);
1374 static void gdlm_unmount(struct gfs2_sbd
*sdp
)
1376 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1378 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1381 /* wait for gfs2_control_wq to be done with this mount */
1383 spin_lock(&ls
->ls_recover_spin
);
1384 set_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
);
1385 spin_unlock(&ls
->ls_recover_spin
);
1386 flush_delayed_work(&sdp
->sd_control_work
);
1388 /* mounted_lock and control_lock will be purged in dlm recovery */
1391 dlm_release_lockspace(ls
->ls_dlm
, 2);
1395 free_recover_size(ls
);
1398 static const match_table_t dlm_tokens
= {
1399 { Opt_jid
, "jid=%d"},
1401 { Opt_first
, "first=%d"},
1402 { Opt_nodir
, "nodir=%d"},
1406 const struct lm_lockops gfs2_dlm_ops
= {
1407 .lm_proto_name
= "lock_dlm",
1408 .lm_mount
= gdlm_mount
,
1409 .lm_first_done
= gdlm_first_done
,
1410 .lm_recovery_result
= gdlm_recovery_result
,
1411 .lm_unmount
= gdlm_unmount
,
1412 .lm_put_lock
= gdlm_put_lock
,
1413 .lm_lock
= gdlm_lock
,
1414 .lm_cancel
= gdlm_cancel
,
1415 .lm_tokens
= &dlm_tokens
,