2 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
3 * Copyright 2004-2011 Red Hat, Inc.
5 * This copyrighted material is made available to anyone wishing to use,
6 * modify, copy, or redistribute it subject to the terms and conditions
7 * of the GNU General Public License version 2.
11 #include <linux/dlm.h>
12 #include <linux/slab.h>
13 #include <linux/types.h>
14 #include <linux/delay.h>
15 #include <linux/gfs2_ondisk.h>
21 #include "trace_gfs2.h"
23 extern struct workqueue_struct
*gfs2_control_wq
;
26 * gfs2_update_stats - Update time based stats
27 * @mv: Pointer to mean/variance structure to update
28 * @sample: New data to include
30 * @delta is the difference between the current rtt sample and the
31 * running average srtt. We add 1/8 of that to the srtt in order to
32 * update the current srtt estimate. The varience estimate is a bit
33 * more complicated. We subtract the abs value of the @delta from
34 * the current variance estimate and add 1/4 of that to the running
37 * Note that the index points at the array entry containing the smoothed
38 * mean value, and the variance is always in the following entry
40 * Reference: TCP/IP Illustrated, vol 2, p. 831,832
41 * All times are in units of integer nanoseconds. Unlike the TCP/IP case,
42 * they are not scaled fixed point.
45 static inline void gfs2_update_stats(struct gfs2_lkstats
*s
, unsigned index
,
48 s64 delta
= sample
- s
->stats
[index
];
49 s
->stats
[index
] += (delta
>> 3);
51 s
->stats
[index
] += ((abs64(delta
) - s
->stats
[index
]) >> 2);
55 * gfs2_update_reply_times - Update locking statistics
56 * @gl: The glock to update
58 * This assumes that gl->gl_dstamp has been set earlier.
60 * The rtt (lock round trip time) is an estimate of the time
61 * taken to perform a dlm lock request. We update it on each
64 * The blocking flag is set on the glock for all dlm requests
65 * which may potentially block due to lock requests from other nodes.
66 * DLM requests where the current lock state is exclusive, the
67 * requested state is null (or unlocked) or where the TRY or
68 * TRY_1CB flags are set are classified as non-blocking. All
69 * other DLM requests are counted as (potentially) blocking.
71 static inline void gfs2_update_reply_times(struct gfs2_glock
*gl
)
73 struct gfs2_pcpu_lkstats
*lks
;
74 const unsigned gltype
= gl
->gl_name
.ln_type
;
75 unsigned index
= test_bit(GLF_BLOCKING
, &gl
->gl_flags
) ?
76 GFS2_LKS_SRTTB
: GFS2_LKS_SRTT
;
80 rtt
= ktime_to_ns(ktime_sub(ktime_get_real(), gl
->gl_dstamp
));
81 lks
= this_cpu_ptr(gl
->gl_sbd
->sd_lkstats
);
82 gfs2_update_stats(&gl
->gl_stats
, index
, rtt
); /* Local */
83 gfs2_update_stats(&lks
->lkstats
[gltype
], index
, rtt
); /* Global */
86 trace_gfs2_glock_lock_time(gl
, rtt
);
90 * gfs2_update_request_times - Update locking statistics
91 * @gl: The glock to update
93 * The irt (lock inter-request times) measures the average time
94 * between requests to the dlm. It is updated immediately before
98 static inline void gfs2_update_request_times(struct gfs2_glock
*gl
)
100 struct gfs2_pcpu_lkstats
*lks
;
101 const unsigned gltype
= gl
->gl_name
.ln_type
;
106 dstamp
= gl
->gl_dstamp
;
107 gl
->gl_dstamp
= ktime_get_real();
108 irt
= ktime_to_ns(ktime_sub(gl
->gl_dstamp
, dstamp
));
109 lks
= this_cpu_ptr(gl
->gl_sbd
->sd_lkstats
);
110 gfs2_update_stats(&gl
->gl_stats
, GFS2_LKS_SIRT
, irt
); /* Local */
111 gfs2_update_stats(&lks
->lkstats
[gltype
], GFS2_LKS_SIRT
, irt
); /* Global */
115 static void gdlm_ast(void *arg
)
117 struct gfs2_glock
*gl
= arg
;
118 unsigned ret
= gl
->gl_state
;
120 gfs2_update_reply_times(gl
);
121 BUG_ON(gl
->gl_lksb
.sb_flags
& DLM_SBF_DEMOTED
);
123 if ((gl
->gl_lksb
.sb_flags
& DLM_SBF_VALNOTVALID
) && gl
->gl_lksb
.sb_lvbptr
)
124 memset(gl
->gl_lksb
.sb_lvbptr
, 0, GDLM_LVB_SIZE
);
126 switch (gl
->gl_lksb
.sb_status
) {
127 case -DLM_EUNLOCK
: /* Unlocked, so glock can be freed */
130 case -DLM_ECANCEL
: /* Cancel while getting lock */
131 ret
|= LM_OUT_CANCELED
;
133 case -EAGAIN
: /* Try lock fails */
134 case -EDEADLK
: /* Deadlock detected */
136 case -ETIMEDOUT
: /* Canceled due to timeout */
139 case 0: /* Success */
141 default: /* Something unexpected */
146 if (gl
->gl_lksb
.sb_flags
& DLM_SBF_ALTMODE
) {
147 if (gl
->gl_req
== LM_ST_SHARED
)
148 ret
= LM_ST_DEFERRED
;
149 else if (gl
->gl_req
== LM_ST_DEFERRED
)
155 set_bit(GLF_INITIAL
, &gl
->gl_flags
);
156 gfs2_glock_complete(gl
, ret
);
159 if (!test_bit(GLF_INITIAL
, &gl
->gl_flags
))
160 gl
->gl_lksb
.sb_lkid
= 0;
161 gfs2_glock_complete(gl
, ret
);
164 static void gdlm_bast(void *arg
, int mode
)
166 struct gfs2_glock
*gl
= arg
;
170 gfs2_glock_cb(gl
, LM_ST_UNLOCKED
);
173 gfs2_glock_cb(gl
, LM_ST_DEFERRED
);
176 gfs2_glock_cb(gl
, LM_ST_SHARED
);
179 printk(KERN_ERR
"unknown bast mode %d", mode
);
184 /* convert gfs lock-state to dlm lock-mode */
186 static int make_mode(const unsigned int lmstate
)
191 case LM_ST_EXCLUSIVE
:
198 printk(KERN_ERR
"unknown LM state %d", lmstate
);
203 static u32
make_flags(struct gfs2_glock
*gl
, const unsigned int gfs_flags
,
208 if (gl
->gl_lksb
.sb_lvbptr
)
209 lkf
|= DLM_LKF_VALBLK
;
211 if (gfs_flags
& LM_FLAG_TRY
)
212 lkf
|= DLM_LKF_NOQUEUE
;
214 if (gfs_flags
& LM_FLAG_TRY_1CB
) {
215 lkf
|= DLM_LKF_NOQUEUE
;
216 lkf
|= DLM_LKF_NOQUEUEBAST
;
219 if (gfs_flags
& LM_FLAG_PRIORITY
) {
220 lkf
|= DLM_LKF_NOORDER
;
221 lkf
|= DLM_LKF_HEADQUE
;
224 if (gfs_flags
& LM_FLAG_ANY
) {
225 if (req
== DLM_LOCK_PR
)
226 lkf
|= DLM_LKF_ALTCW
;
227 else if (req
== DLM_LOCK_CW
)
228 lkf
|= DLM_LKF_ALTPR
;
233 if (gl
->gl_lksb
.sb_lkid
!= 0) {
234 lkf
|= DLM_LKF_CONVERT
;
235 if (test_bit(GLF_BLOCKING
, &gl
->gl_flags
))
236 lkf
|= DLM_LKF_QUECVT
;
242 static void gfs2_reverse_hex(char *c
, u64 value
)
245 *c
-- = hex_asc
[value
& 0x0f];
250 static int gdlm_lock(struct gfs2_glock
*gl
, unsigned int req_state
,
253 struct lm_lockstruct
*ls
= &gl
->gl_sbd
->sd_lockstruct
;
256 char strname
[GDLM_STRNAME_BYTES
] = "";
258 req
= make_mode(req_state
);
259 lkf
= make_flags(gl
, flags
, req
);
260 gfs2_glstats_inc(gl
, GFS2_LKS_DCOUNT
);
261 gfs2_sbstats_inc(gl
, GFS2_LKS_DCOUNT
);
262 if (gl
->gl_lksb
.sb_lkid
) {
263 gfs2_update_request_times(gl
);
265 memset(strname
, ' ', GDLM_STRNAME_BYTES
- 1);
266 strname
[GDLM_STRNAME_BYTES
- 1] = '\0';
267 gfs2_reverse_hex(strname
+ 7, gl
->gl_name
.ln_type
);
268 gfs2_reverse_hex(strname
+ 23, gl
->gl_name
.ln_number
);
269 gl
->gl_dstamp
= ktime_get_real();
272 * Submit the actual lock request.
275 return dlm_lock(ls
->ls_dlm
, req
, &gl
->gl_lksb
, lkf
, strname
,
276 GDLM_STRNAME_BYTES
- 1, 0, gdlm_ast
, gl
, gdlm_bast
);
279 static void gdlm_put_lock(struct gfs2_glock
*gl
)
281 struct gfs2_sbd
*sdp
= gl
->gl_sbd
;
282 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
285 if (gl
->gl_lksb
.sb_lkid
== 0) {
290 clear_bit(GLF_BLOCKING
, &gl
->gl_flags
);
291 gfs2_glstats_inc(gl
, GFS2_LKS_DCOUNT
);
292 gfs2_sbstats_inc(gl
, GFS2_LKS_DCOUNT
);
293 gfs2_update_request_times(gl
);
295 /* don't want to skip dlm_unlock writing the lvb when lock is ex */
296 if (test_bit(SDF_SKIP_DLM_UNLOCK
, &sdp
->sd_flags
) &&
297 gl
->gl_lksb
.sb_lvbptr
&& (gl
->gl_state
!= LM_ST_EXCLUSIVE
)) {
302 error
= dlm_unlock(ls
->ls_dlm
, gl
->gl_lksb
.sb_lkid
, DLM_LKF_VALBLK
,
305 printk(KERN_ERR
"gdlm_unlock %x,%llx err=%d\n",
307 (unsigned long long)gl
->gl_name
.ln_number
, error
);
312 static void gdlm_cancel(struct gfs2_glock
*gl
)
314 struct lm_lockstruct
*ls
= &gl
->gl_sbd
->sd_lockstruct
;
315 dlm_unlock(ls
->ls_dlm
, gl
->gl_lksb
.sb_lkid
, DLM_LKF_CANCEL
, NULL
, gl
);
319 * dlm/gfs2 recovery coordination using dlm_recover callbacks
321 * 1. dlm_controld sees lockspace members change
322 * 2. dlm_controld blocks dlm-kernel locking activity
323 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
324 * 4. dlm_controld starts and finishes its own user level recovery
325 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
326 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
327 * 7. dlm_recoverd does its own lock recovery
328 * 8. dlm_recoverd unblocks dlm-kernel locking activity
329 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
330 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
331 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
332 * 12. gfs2_recover dequeues and recovers journals of failed nodes
333 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
334 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
335 * 15. gfs2_control unblocks normal locking when all journals are recovered
337 * - failures during recovery
339 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
340 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
341 * recovering for a prior failure. gfs2_control needs a way to detect
342 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
343 * the recover_block and recover_start values.
345 * recover_done() provides a new lockspace generation number each time it
346 * is called (step 9). This generation number is saved as recover_start.
347 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
348 * recover_block = recover_start. So, while recover_block is equal to
349 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
350 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
352 * - more specific gfs2 steps in sequence above
354 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
355 * 6. recover_slot records any failed jids (maybe none)
356 * 9. recover_done sets recover_start = new generation number
357 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
358 * 12. gfs2_recover does journal recoveries for failed jids identified above
359 * 14. gfs2_control clears control_lock lvb bits for recovered jids
360 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
361 * again) then do nothing, otherwise if recover_start > recover_block
362 * then clear BLOCK_LOCKS.
364 * - parallel recovery steps across all nodes
366 * All nodes attempt to update the control_lock lvb with the new generation
367 * number and jid bits, but only the first to get the control_lock EX will
368 * do so; others will see that it's already done (lvb already contains new
369 * generation number.)
371 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
372 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
373 * . One node gets control_lock first and writes the lvb, others see it's done
374 * . All nodes attempt to recover jids for which they see control_lock bits set
375 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
376 * . All nodes will eventually see all lvb bits clear and unblock locks
378 * - is there a problem with clearing an lvb bit that should be set
379 * and missing a journal recovery?
382 * 2. lvb bit set for step 1
383 * 3. jid recovered for step 1
384 * 4. jid taken again (new mount)
385 * 5. jid fails (for step 4)
386 * 6. lvb bit set for step 5 (will already be set)
387 * 7. lvb bit cleared for step 3
389 * This is not a problem because the failure in step 5 does not
390 * require recovery, because the mount in step 4 could not have
391 * progressed far enough to unblock locks and access the fs. The
392 * control_mount() function waits for all recoveries to be complete
393 * for the latest lockspace generation before ever unblocking locks
394 * and returning. The mount in step 4 waits until the recovery in
397 * - special case of first mounter: first node to mount the fs
399 * The first node to mount a gfs2 fs needs to check all the journals
400 * and recover any that need recovery before other nodes are allowed
401 * to mount the fs. (Others may begin mounting, but they must wait
402 * for the first mounter to be done before taking locks on the fs
403 * or accessing the fs.) This has two parts:
405 * 1. The mounted_lock tells a node it's the first to mount the fs.
406 * Each node holds the mounted_lock in PR while it's mounted.
407 * Each node tries to acquire the mounted_lock in EX when it mounts.
408 * If a node is granted the mounted_lock EX it means there are no
409 * other mounted nodes (no PR locks exist), and it is the first mounter.
410 * The mounted_lock is demoted to PR when first recovery is done, so
411 * others will fail to get an EX lock, but will get a PR lock.
413 * 2. The control_lock blocks others in control_mount() while the first
414 * mounter is doing first mount recovery of all journals.
415 * A mounting node needs to acquire control_lock in EX mode before
416 * it can proceed. The first mounter holds control_lock in EX while doing
417 * the first mount recovery, blocking mounts from other nodes, then demotes
418 * control_lock to NL when it's done (others_may_mount/first_done),
419 * allowing other nodes to continue mounting.
422 * control_lock EX/NOQUEUE success
423 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
425 * do first mounter recovery
426 * mounted_lock EX->PR
427 * control_lock EX->NL, write lvb generation
430 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
431 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
432 * mounted_lock PR/NOQUEUE success
433 * read lvb generation
434 * control_lock EX->NL
437 * - mount during recovery
439 * If a node mounts while others are doing recovery (not first mounter),
440 * the mounting node will get its initial recover_done() callback without
441 * having seen any previous failures/callbacks.
443 * It must wait for all recoveries preceding its mount to be finished
444 * before it unblocks locks. It does this by repeating the "other mounter"
445 * steps above until the lvb generation number is >= its mount generation
446 * number (from initial recover_done) and all lvb bits are clear.
448 * - control_lock lvb format
450 * 4 bytes generation number: the latest dlm lockspace generation number
451 * from recover_done callback. Indicates the jid bitmap has been updated
452 * to reflect all slot failures through that generation.
454 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
455 * that jid N needs recovery.
458 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
460 static void control_lvb_read(struct lm_lockstruct
*ls
, uint32_t *lvb_gen
,
464 memcpy(lvb_bits
, ls
->ls_control_lvb
, GDLM_LVB_SIZE
);
465 memcpy(&gen
, lvb_bits
, sizeof(uint32_t));
466 *lvb_gen
= le32_to_cpu(gen
);
469 static void control_lvb_write(struct lm_lockstruct
*ls
, uint32_t lvb_gen
,
473 memcpy(ls
->ls_control_lvb
, lvb_bits
, GDLM_LVB_SIZE
);
474 gen
= cpu_to_le32(lvb_gen
);
475 memcpy(ls
->ls_control_lvb
, &gen
, sizeof(uint32_t));
478 static int all_jid_bits_clear(char *lvb
)
481 for (i
= JID_BITMAP_OFFSET
; i
< GDLM_LVB_SIZE
; i
++) {
488 static void sync_wait_cb(void *arg
)
490 struct lm_lockstruct
*ls
= arg
;
491 complete(&ls
->ls_sync_wait
);
494 static int sync_unlock(struct gfs2_sbd
*sdp
, struct dlm_lksb
*lksb
, char *name
)
496 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
499 error
= dlm_unlock(ls
->ls_dlm
, lksb
->sb_lkid
, 0, lksb
, ls
);
501 fs_err(sdp
, "%s lkid %x error %d\n",
502 name
, lksb
->sb_lkid
, error
);
506 wait_for_completion(&ls
->ls_sync_wait
);
508 if (lksb
->sb_status
!= -DLM_EUNLOCK
) {
509 fs_err(sdp
, "%s lkid %x status %d\n",
510 name
, lksb
->sb_lkid
, lksb
->sb_status
);
516 static int sync_lock(struct gfs2_sbd
*sdp
, int mode
, uint32_t flags
,
517 unsigned int num
, struct dlm_lksb
*lksb
, char *name
)
519 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
520 char strname
[GDLM_STRNAME_BYTES
];
523 memset(strname
, 0, GDLM_STRNAME_BYTES
);
524 snprintf(strname
, GDLM_STRNAME_BYTES
, "%8x%16x", LM_TYPE_NONDISK
, num
);
526 error
= dlm_lock(ls
->ls_dlm
, mode
, lksb
, flags
,
527 strname
, GDLM_STRNAME_BYTES
- 1,
528 0, sync_wait_cb
, ls
, NULL
);
530 fs_err(sdp
, "%s lkid %x flags %x mode %d error %d\n",
531 name
, lksb
->sb_lkid
, flags
, mode
, error
);
535 wait_for_completion(&ls
->ls_sync_wait
);
537 status
= lksb
->sb_status
;
539 if (status
&& status
!= -EAGAIN
) {
540 fs_err(sdp
, "%s lkid %x flags %x mode %d status %d\n",
541 name
, lksb
->sb_lkid
, flags
, mode
, status
);
547 static int mounted_unlock(struct gfs2_sbd
*sdp
)
549 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
550 return sync_unlock(sdp
, &ls
->ls_mounted_lksb
, "mounted_lock");
553 static int mounted_lock(struct gfs2_sbd
*sdp
, int mode
, uint32_t flags
)
555 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
556 return sync_lock(sdp
, mode
, flags
, GFS2_MOUNTED_LOCK
,
557 &ls
->ls_mounted_lksb
, "mounted_lock");
560 static int control_unlock(struct gfs2_sbd
*sdp
)
562 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
563 return sync_unlock(sdp
, &ls
->ls_control_lksb
, "control_lock");
566 static int control_lock(struct gfs2_sbd
*sdp
, int mode
, uint32_t flags
)
568 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
569 return sync_lock(sdp
, mode
, flags
, GFS2_CONTROL_LOCK
,
570 &ls
->ls_control_lksb
, "control_lock");
573 static void gfs2_control_func(struct work_struct
*work
)
575 struct gfs2_sbd
*sdp
= container_of(work
, struct gfs2_sbd
, sd_control_work
.work
);
576 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
577 char lvb_bits
[GDLM_LVB_SIZE
];
578 uint32_t block_gen
, start_gen
, lvb_gen
, flags
;
584 spin_lock(&ls
->ls_recover_spin
);
586 * No MOUNT_DONE means we're still mounting; control_mount()
587 * will set this flag, after which this thread will take over
588 * all further clearing of BLOCK_LOCKS.
590 * FIRST_MOUNT means this node is doing first mounter recovery,
591 * for which recovery control is handled by
592 * control_mount()/control_first_done(), not this thread.
594 if (!test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
595 test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
596 spin_unlock(&ls
->ls_recover_spin
);
599 block_gen
= ls
->ls_recover_block
;
600 start_gen
= ls
->ls_recover_start
;
601 spin_unlock(&ls
->ls_recover_spin
);
604 * Equal block_gen and start_gen implies we are between
605 * recover_prep and recover_done callbacks, which means
606 * dlm recovery is in progress and dlm locking is blocked.
607 * There's no point trying to do any work until recover_done.
610 if (block_gen
== start_gen
)
614 * Propagate recover_submit[] and recover_result[] to lvb:
615 * dlm_recoverd adds to recover_submit[] jids needing recovery
616 * gfs2_recover adds to recover_result[] journal recovery results
618 * set lvb bit for jids in recover_submit[] if the lvb has not
619 * yet been updated for the generation of the failure
621 * clear lvb bit for jids in recover_result[] if the result of
622 * the journal recovery is SUCCESS
625 error
= control_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_VALBLK
);
627 fs_err(sdp
, "control lock EX error %d\n", error
);
631 control_lvb_read(ls
, &lvb_gen
, lvb_bits
);
633 spin_lock(&ls
->ls_recover_spin
);
634 if (block_gen
!= ls
->ls_recover_block
||
635 start_gen
!= ls
->ls_recover_start
) {
636 fs_info(sdp
, "recover generation %u block1 %u %u\n",
637 start_gen
, block_gen
, ls
->ls_recover_block
);
638 spin_unlock(&ls
->ls_recover_spin
);
639 control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
643 recover_size
= ls
->ls_recover_size
;
645 if (lvb_gen
<= start_gen
) {
647 * Clear lvb bits for jids we've successfully recovered.
648 * Because all nodes attempt to recover failed journals,
649 * a journal can be recovered multiple times successfully
650 * in succession. Only the first will really do recovery,
651 * the others find it clean, but still report a successful
652 * recovery. So, another node may have already recovered
653 * the jid and cleared the lvb bit for it.
655 for (i
= 0; i
< recover_size
; i
++) {
656 if (ls
->ls_recover_result
[i
] != LM_RD_SUCCESS
)
659 ls
->ls_recover_result
[i
] = 0;
661 if (!test_bit_le(i
, lvb_bits
+ JID_BITMAP_OFFSET
))
664 __clear_bit_le(i
, lvb_bits
+ JID_BITMAP_OFFSET
);
669 if (lvb_gen
== start_gen
) {
671 * Failed slots before start_gen are already set in lvb.
673 for (i
= 0; i
< recover_size
; i
++) {
674 if (!ls
->ls_recover_submit
[i
])
676 if (ls
->ls_recover_submit
[i
] < lvb_gen
)
677 ls
->ls_recover_submit
[i
] = 0;
679 } else if (lvb_gen
< start_gen
) {
681 * Failed slots before start_gen are not yet set in lvb.
683 for (i
= 0; i
< recover_size
; i
++) {
684 if (!ls
->ls_recover_submit
[i
])
686 if (ls
->ls_recover_submit
[i
] < start_gen
) {
687 ls
->ls_recover_submit
[i
] = 0;
688 __set_bit_le(i
, lvb_bits
+ JID_BITMAP_OFFSET
);
691 /* even if there are no bits to set, we need to write the
692 latest generation to the lvb */
696 * we should be getting a recover_done() for lvb_gen soon
699 spin_unlock(&ls
->ls_recover_spin
);
702 control_lvb_write(ls
, start_gen
, lvb_bits
);
703 flags
= DLM_LKF_CONVERT
| DLM_LKF_VALBLK
;
705 flags
= DLM_LKF_CONVERT
;
708 error
= control_lock(sdp
, DLM_LOCK_NL
, flags
);
710 fs_err(sdp
, "control lock NL error %d\n", error
);
715 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
716 * and clear a jid bit in the lvb if the recovery is a success.
717 * Eventually all journals will be recovered, all jid bits will
718 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
721 for (i
= 0; i
< recover_size
; i
++) {
722 if (test_bit_le(i
, lvb_bits
+ JID_BITMAP_OFFSET
)) {
723 fs_info(sdp
, "recover generation %u jid %d\n",
725 gfs2_recover_set(sdp
, i
);
733 * No more jid bits set in lvb, all recovery is done, unblock locks
734 * (unless a new recover_prep callback has occured blocking locks
735 * again while working above)
738 spin_lock(&ls
->ls_recover_spin
);
739 if (ls
->ls_recover_block
== block_gen
&&
740 ls
->ls_recover_start
== start_gen
) {
741 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
742 spin_unlock(&ls
->ls_recover_spin
);
743 fs_info(sdp
, "recover generation %u done\n", start_gen
);
744 gfs2_glock_thaw(sdp
);
746 fs_info(sdp
, "recover generation %u block2 %u %u\n",
747 start_gen
, block_gen
, ls
->ls_recover_block
);
748 spin_unlock(&ls
->ls_recover_spin
);
752 static int control_mount(struct gfs2_sbd
*sdp
)
754 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
755 char lvb_bits
[GDLM_LVB_SIZE
];
756 uint32_t start_gen
, block_gen
, mount_gen
, lvb_gen
;
761 memset(&ls
->ls_mounted_lksb
, 0, sizeof(struct dlm_lksb
));
762 memset(&ls
->ls_control_lksb
, 0, sizeof(struct dlm_lksb
));
763 memset(&ls
->ls_control_lvb
, 0, GDLM_LVB_SIZE
);
764 ls
->ls_control_lksb
.sb_lvbptr
= ls
->ls_control_lvb
;
765 init_completion(&ls
->ls_sync_wait
);
767 set_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
769 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_VALBLK
);
771 fs_err(sdp
, "control_mount control_lock NL error %d\n", error
);
775 error
= mounted_lock(sdp
, DLM_LOCK_NL
, 0);
777 fs_err(sdp
, "control_mount mounted_lock NL error %d\n", error
);
781 mounted_mode
= DLM_LOCK_NL
;
784 if (retries
++ && signal_pending(current
)) {
790 * We always start with both locks in NL. control_lock is
791 * demoted to NL below so we don't need to do it here.
794 if (mounted_mode
!= DLM_LOCK_NL
) {
795 error
= mounted_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
798 mounted_mode
= DLM_LOCK_NL
;
802 * Other nodes need to do some work in dlm recovery and gfs2_control
803 * before the recover_done and control_lock will be ready for us below.
804 * A delay here is not required but often avoids having to retry.
807 msleep_interruptible(500);
810 * Acquire control_lock in EX and mounted_lock in either EX or PR.
811 * control_lock lvb keeps track of any pending journal recoveries.
812 * mounted_lock indicates if any other nodes have the fs mounted.
815 error
= control_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
|DLM_LKF_VALBLK
);
816 if (error
== -EAGAIN
) {
819 fs_err(sdp
, "control_mount control_lock EX error %d\n", error
);
823 error
= mounted_lock(sdp
, DLM_LOCK_EX
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
);
825 mounted_mode
= DLM_LOCK_EX
;
827 } else if (error
!= -EAGAIN
) {
828 fs_err(sdp
, "control_mount mounted_lock EX error %d\n", error
);
832 error
= mounted_lock(sdp
, DLM_LOCK_PR
, DLM_LKF_CONVERT
|DLM_LKF_NOQUEUE
);
834 mounted_mode
= DLM_LOCK_PR
;
837 /* not even -EAGAIN should happen here */
838 fs_err(sdp
, "control_mount mounted_lock PR error %d\n", error
);
844 * If we got both locks above in EX, then we're the first mounter.
845 * If not, then we need to wait for the control_lock lvb to be
846 * updated by other mounted nodes to reflect our mount generation.
848 * In simple first mounter cases, first mounter will see zero lvb_gen,
849 * but in cases where all existing nodes leave/fail before mounting
850 * nodes finish control_mount, then all nodes will be mounting and
851 * lvb_gen will be non-zero.
854 control_lvb_read(ls
, &lvb_gen
, lvb_bits
);
856 if (lvb_gen
== 0xFFFFFFFF) {
857 /* special value to force mount attempts to fail */
858 fs_err(sdp
, "control_mount control_lock disabled\n");
863 if (mounted_mode
== DLM_LOCK_EX
) {
864 /* first mounter, keep both EX while doing first recovery */
865 spin_lock(&ls
->ls_recover_spin
);
866 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
867 set_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
);
868 set_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
869 spin_unlock(&ls
->ls_recover_spin
);
870 fs_info(sdp
, "first mounter control generation %u\n", lvb_gen
);
874 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
);
879 * We are not first mounter, now we need to wait for the control_lock
880 * lvb generation to be >= the generation from our first recover_done
881 * and all lvb bits to be clear (no pending journal recoveries.)
884 if (!all_jid_bits_clear(lvb_bits
)) {
885 /* journals need recovery, wait until all are clear */
886 fs_info(sdp
, "control_mount wait for journal recovery\n");
890 spin_lock(&ls
->ls_recover_spin
);
891 block_gen
= ls
->ls_recover_block
;
892 start_gen
= ls
->ls_recover_start
;
893 mount_gen
= ls
->ls_recover_mount
;
895 if (lvb_gen
< mount_gen
) {
896 /* wait for mounted nodes to update control_lock lvb to our
897 generation, which might include new recovery bits set */
898 fs_info(sdp
, "control_mount wait1 block %u start %u mount %u "
899 "lvb %u flags %lx\n", block_gen
, start_gen
, mount_gen
,
900 lvb_gen
, ls
->ls_recover_flags
);
901 spin_unlock(&ls
->ls_recover_spin
);
905 if (lvb_gen
!= start_gen
) {
906 /* wait for mounted nodes to update control_lock lvb to the
907 latest recovery generation */
908 fs_info(sdp
, "control_mount wait2 block %u start %u mount %u "
909 "lvb %u flags %lx\n", block_gen
, start_gen
, mount_gen
,
910 lvb_gen
, ls
->ls_recover_flags
);
911 spin_unlock(&ls
->ls_recover_spin
);
915 if (block_gen
== start_gen
) {
916 /* dlm recovery in progress, wait for it to finish */
917 fs_info(sdp
, "control_mount wait3 block %u start %u mount %u "
918 "lvb %u flags %lx\n", block_gen
, start_gen
, mount_gen
,
919 lvb_gen
, ls
->ls_recover_flags
);
920 spin_unlock(&ls
->ls_recover_spin
);
924 clear_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
925 set_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
);
926 memset(ls
->ls_recover_submit
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
927 memset(ls
->ls_recover_result
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
928 spin_unlock(&ls
->ls_recover_spin
);
937 static int dlm_recovery_wait(void *word
)
943 static int control_first_done(struct gfs2_sbd
*sdp
)
945 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
946 char lvb_bits
[GDLM_LVB_SIZE
];
947 uint32_t start_gen
, block_gen
;
951 spin_lock(&ls
->ls_recover_spin
);
952 start_gen
= ls
->ls_recover_start
;
953 block_gen
= ls
->ls_recover_block
;
955 if (test_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
) ||
956 !test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
957 !test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
958 /* sanity check, should not happen */
959 fs_err(sdp
, "control_first_done start %u block %u flags %lx\n",
960 start_gen
, block_gen
, ls
->ls_recover_flags
);
961 spin_unlock(&ls
->ls_recover_spin
);
966 if (start_gen
== block_gen
) {
968 * Wait for the end of a dlm recovery cycle to switch from
969 * first mounter recovery. We can ignore any recover_slot
970 * callbacks between the recover_prep and next recover_done
971 * because we are still the first mounter and any failed nodes
972 * have not fully mounted, so they don't need recovery.
974 spin_unlock(&ls
->ls_recover_spin
);
975 fs_info(sdp
, "control_first_done wait gen %u\n", start_gen
);
977 wait_on_bit(&ls
->ls_recover_flags
, DFL_DLM_RECOVERY
,
978 dlm_recovery_wait
, TASK_UNINTERRUPTIBLE
);
982 clear_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
983 set_bit(DFL_FIRST_MOUNT_DONE
, &ls
->ls_recover_flags
);
984 memset(ls
->ls_recover_submit
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
985 memset(ls
->ls_recover_result
, 0, ls
->ls_recover_size
*sizeof(uint32_t));
986 spin_unlock(&ls
->ls_recover_spin
);
988 memset(lvb_bits
, 0, sizeof(lvb_bits
));
989 control_lvb_write(ls
, start_gen
, lvb_bits
);
991 error
= mounted_lock(sdp
, DLM_LOCK_PR
, DLM_LKF_CONVERT
);
993 fs_err(sdp
, "control_first_done mounted PR error %d\n", error
);
995 error
= control_lock(sdp
, DLM_LOCK_NL
, DLM_LKF_CONVERT
|DLM_LKF_VALBLK
);
997 fs_err(sdp
, "control_first_done control NL error %d\n", error
);
1003 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1004 * to accomodate the largest slot number. (NB dlm slot numbers start at 1,
1005 * gfs2 jids start at 0, so jid = slot - 1)
1008 #define RECOVER_SIZE_INC 16
1010 static int set_recover_size(struct gfs2_sbd
*sdp
, struct dlm_slot
*slots
,
1013 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1014 uint32_t *submit
= NULL
;
1015 uint32_t *result
= NULL
;
1016 uint32_t old_size
, new_size
;
1020 for (i
= 0; i
< num_slots
; i
++) {
1021 if (max_jid
< slots
[i
].slot
- 1)
1022 max_jid
= slots
[i
].slot
- 1;
1025 old_size
= ls
->ls_recover_size
;
1027 if (old_size
>= max_jid
+ 1)
1030 new_size
= old_size
+ RECOVER_SIZE_INC
;
1032 submit
= kzalloc(new_size
* sizeof(uint32_t), GFP_NOFS
);
1033 result
= kzalloc(new_size
* sizeof(uint32_t), GFP_NOFS
);
1034 if (!submit
|| !result
) {
1040 spin_lock(&ls
->ls_recover_spin
);
1041 memcpy(submit
, ls
->ls_recover_submit
, old_size
* sizeof(uint32_t));
1042 memcpy(result
, ls
->ls_recover_result
, old_size
* sizeof(uint32_t));
1043 kfree(ls
->ls_recover_submit
);
1044 kfree(ls
->ls_recover_result
);
1045 ls
->ls_recover_submit
= submit
;
1046 ls
->ls_recover_result
= result
;
1047 ls
->ls_recover_size
= new_size
;
1048 spin_unlock(&ls
->ls_recover_spin
);
1052 static void free_recover_size(struct lm_lockstruct
*ls
)
1054 kfree(ls
->ls_recover_submit
);
1055 kfree(ls
->ls_recover_result
);
1056 ls
->ls_recover_submit
= NULL
;
1057 ls
->ls_recover_result
= NULL
;
1058 ls
->ls_recover_size
= 0;
1061 /* dlm calls before it does lock recovery */
1063 static void gdlm_recover_prep(void *arg
)
1065 struct gfs2_sbd
*sdp
= arg
;
1066 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1068 spin_lock(&ls
->ls_recover_spin
);
1069 ls
->ls_recover_block
= ls
->ls_recover_start
;
1070 set_bit(DFL_DLM_RECOVERY
, &ls
->ls_recover_flags
);
1072 if (!test_bit(DFL_MOUNT_DONE
, &ls
->ls_recover_flags
) ||
1073 test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
1074 spin_unlock(&ls
->ls_recover_spin
);
1077 set_bit(DFL_BLOCK_LOCKS
, &ls
->ls_recover_flags
);
1078 spin_unlock(&ls
->ls_recover_spin
);
1081 /* dlm calls after recover_prep has been completed on all lockspace members;
1082 identifies slot/jid of failed member */
1084 static void gdlm_recover_slot(void *arg
, struct dlm_slot
*slot
)
1086 struct gfs2_sbd
*sdp
= arg
;
1087 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1088 int jid
= slot
->slot
- 1;
1090 spin_lock(&ls
->ls_recover_spin
);
1091 if (ls
->ls_recover_size
< jid
+ 1) {
1092 fs_err(sdp
, "recover_slot jid %d gen %u short size %d",
1093 jid
, ls
->ls_recover_block
, ls
->ls_recover_size
);
1094 spin_unlock(&ls
->ls_recover_spin
);
1098 if (ls
->ls_recover_submit
[jid
]) {
1099 fs_info(sdp
, "recover_slot jid %d gen %u prev %u",
1100 jid
, ls
->ls_recover_block
, ls
->ls_recover_submit
[jid
]);
1102 ls
->ls_recover_submit
[jid
] = ls
->ls_recover_block
;
1103 spin_unlock(&ls
->ls_recover_spin
);
1106 /* dlm calls after recover_slot and after it completes lock recovery */
1108 static void gdlm_recover_done(void *arg
, struct dlm_slot
*slots
, int num_slots
,
1109 int our_slot
, uint32_t generation
)
1111 struct gfs2_sbd
*sdp
= arg
;
1112 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1114 /* ensure the ls jid arrays are large enough */
1115 set_recover_size(sdp
, slots
, num_slots
);
1117 spin_lock(&ls
->ls_recover_spin
);
1118 ls
->ls_recover_start
= generation
;
1120 if (!ls
->ls_recover_mount
) {
1121 ls
->ls_recover_mount
= generation
;
1122 ls
->ls_jid
= our_slot
- 1;
1125 if (!test_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
))
1126 queue_delayed_work(gfs2_control_wq
, &sdp
->sd_control_work
, 0);
1128 clear_bit(DFL_DLM_RECOVERY
, &ls
->ls_recover_flags
);
1129 smp_mb__after_clear_bit();
1130 wake_up_bit(&ls
->ls_recover_flags
, DFL_DLM_RECOVERY
);
1131 spin_unlock(&ls
->ls_recover_spin
);
1134 /* gfs2_recover thread has a journal recovery result */
1136 static void gdlm_recovery_result(struct gfs2_sbd
*sdp
, unsigned int jid
,
1137 unsigned int result
)
1139 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1141 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1144 /* don't care about the recovery of own journal during mount */
1145 if (jid
== ls
->ls_jid
)
1148 spin_lock(&ls
->ls_recover_spin
);
1149 if (test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
)) {
1150 spin_unlock(&ls
->ls_recover_spin
);
1153 if (ls
->ls_recover_size
< jid
+ 1) {
1154 fs_err(sdp
, "recovery_result jid %d short size %d",
1155 jid
, ls
->ls_recover_size
);
1156 spin_unlock(&ls
->ls_recover_spin
);
1160 fs_info(sdp
, "recover jid %d result %s\n", jid
,
1161 result
== LM_RD_GAVEUP
? "busy" : "success");
1163 ls
->ls_recover_result
[jid
] = result
;
1165 /* GAVEUP means another node is recovering the journal; delay our
1166 next attempt to recover it, to give the other node a chance to
1167 finish before trying again */
1169 if (!test_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
))
1170 queue_delayed_work(gfs2_control_wq
, &sdp
->sd_control_work
,
1171 result
== LM_RD_GAVEUP
? HZ
: 0);
1172 spin_unlock(&ls
->ls_recover_spin
);
1175 const struct dlm_lockspace_ops gdlm_lockspace_ops
= {
1176 .recover_prep
= gdlm_recover_prep
,
1177 .recover_slot
= gdlm_recover_slot
,
1178 .recover_done
= gdlm_recover_done
,
1181 static int gdlm_mount(struct gfs2_sbd
*sdp
, const char *table
)
1183 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1184 char cluster
[GFS2_LOCKNAME_LEN
];
1187 int error
, ops_result
;
1190 * initialize everything
1193 INIT_DELAYED_WORK(&sdp
->sd_control_work
, gfs2_control_func
);
1194 spin_lock_init(&ls
->ls_recover_spin
);
1195 ls
->ls_recover_flags
= 0;
1196 ls
->ls_recover_mount
= 0;
1197 ls
->ls_recover_start
= 0;
1198 ls
->ls_recover_block
= 0;
1199 ls
->ls_recover_size
= 0;
1200 ls
->ls_recover_submit
= NULL
;
1201 ls
->ls_recover_result
= NULL
;
1203 error
= set_recover_size(sdp
, NULL
, 0);
1208 * prepare dlm_new_lockspace args
1211 fsname
= strchr(table
, ':');
1213 fs_info(sdp
, "no fsname found\n");
1217 memset(cluster
, 0, sizeof(cluster
));
1218 memcpy(cluster
, table
, strlen(table
) - strlen(fsname
));
1221 flags
= DLM_LSFL_FS
| DLM_LSFL_NEWEXCL
;
1224 * create/join lockspace
1227 error
= dlm_new_lockspace(fsname
, cluster
, flags
, GDLM_LVB_SIZE
,
1228 &gdlm_lockspace_ops
, sdp
, &ops_result
,
1231 fs_err(sdp
, "dlm_new_lockspace error %d\n", error
);
1235 if (ops_result
< 0) {
1237 * dlm does not support ops callbacks,
1238 * old dlm_controld/gfs_controld are used, try without ops.
1240 fs_info(sdp
, "dlm lockspace ops not used\n");
1241 free_recover_size(ls
);
1242 set_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
);
1246 if (!test_bit(SDF_NOJOURNALID
, &sdp
->sd_flags
)) {
1247 fs_err(sdp
, "dlm lockspace ops disallow jid preset\n");
1253 * control_mount() uses control_lock to determine first mounter,
1254 * and for later mounts, waits for any recoveries to be cleared.
1257 error
= control_mount(sdp
);
1259 fs_err(sdp
, "mount control error %d\n", error
);
1263 ls
->ls_first
= !!test_bit(DFL_FIRST_MOUNT
, &ls
->ls_recover_flags
);
1264 clear_bit(SDF_NOJOURNALID
, &sdp
->sd_flags
);
1265 smp_mb__after_clear_bit();
1266 wake_up_bit(&sdp
->sd_flags
, SDF_NOJOURNALID
);
1270 dlm_release_lockspace(ls
->ls_dlm
, 2);
1272 free_recover_size(ls
);
1277 static void gdlm_first_done(struct gfs2_sbd
*sdp
)
1279 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1282 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1285 error
= control_first_done(sdp
);
1287 fs_err(sdp
, "mount first_done error %d\n", error
);
1290 static void gdlm_unmount(struct gfs2_sbd
*sdp
)
1292 struct lm_lockstruct
*ls
= &sdp
->sd_lockstruct
;
1294 if (test_bit(DFL_NO_DLM_OPS
, &ls
->ls_recover_flags
))
1297 /* wait for gfs2_control_wq to be done with this mount */
1299 spin_lock(&ls
->ls_recover_spin
);
1300 set_bit(DFL_UNMOUNT
, &ls
->ls_recover_flags
);
1301 spin_unlock(&ls
->ls_recover_spin
);
1302 flush_delayed_work(&sdp
->sd_control_work
);
1304 /* mounted_lock and control_lock will be purged in dlm recovery */
1307 dlm_release_lockspace(ls
->ls_dlm
, 2);
1311 free_recover_size(ls
);
1314 static const match_table_t dlm_tokens
= {
1315 { Opt_jid
, "jid=%d"},
1317 { Opt_first
, "first=%d"},
1318 { Opt_nodir
, "nodir=%d"},
1322 const struct lm_lockops gfs2_dlm_ops
= {
1323 .lm_proto_name
= "lock_dlm",
1324 .lm_mount
= gdlm_mount
,
1325 .lm_first_done
= gdlm_first_done
,
1326 .lm_recovery_result
= gdlm_recovery_result
,
1327 .lm_unmount
= gdlm_unmount
,
1328 .lm_put_lock
= gdlm_put_lock
,
1329 .lm_lock
= gdlm_lock
,
1330 .lm_cancel
= gdlm_cancel
,
1331 .lm_tokens
= &dlm_tokens
,