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)
246 *c-- = hex_asc[value & 0x0f];
251 static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state,
254 struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
257 char strname[GDLM_STRNAME_BYTES] = "";
259 req = make_mode(req_state);
260 lkf = make_flags(gl, flags, req);
261 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
262 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
263 if (gl->gl_lksb.sb_lkid) {
264 gfs2_update_request_times(gl);
266 memset(strname, ' ', GDLM_STRNAME_BYTES - 1);
267 strname[GDLM_STRNAME_BYTES - 1] = '\0';
268 gfs2_reverse_hex(strname + 7, gl->gl_name.ln_type);
269 gfs2_reverse_hex(strname + 23, gl->gl_name.ln_number);
270 gl->gl_dstamp = ktime_get_real();
273 * Submit the actual lock request.
276 return dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname,
277 GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast);
280 static void gdlm_put_lock(struct gfs2_glock *gl)
282 struct gfs2_sbd *sdp = gl->gl_sbd;
283 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
284 int lvb_needs_unlock = 0;
287 if (gl->gl_lksb.sb_lkid == 0) {
292 clear_bit(GLF_BLOCKING, &gl->gl_flags);
293 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
294 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
295 gfs2_update_request_times(gl);
297 /* don't want to skip dlm_unlock writing the lvb when lock is ex */
299 if (gl->gl_lksb.sb_lvbptr && (gl->gl_state == LM_ST_EXCLUSIVE))
300 lvb_needs_unlock = 1;
302 if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
308 error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_VALBLK,
311 printk(KERN_ERR "gdlm_unlock %x,%llx err=%d\n",
313 (unsigned long long)gl->gl_name.ln_number, error);
318 static void gdlm_cancel(struct gfs2_glock *gl)
320 struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
321 dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
325 * dlm/gfs2 recovery coordination using dlm_recover callbacks
327 * 1. dlm_controld sees lockspace members change
328 * 2. dlm_controld blocks dlm-kernel locking activity
329 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
330 * 4. dlm_controld starts and finishes its own user level recovery
331 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
332 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
333 * 7. dlm_recoverd does its own lock recovery
334 * 8. dlm_recoverd unblocks dlm-kernel locking activity
335 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
336 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
337 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
338 * 12. gfs2_recover dequeues and recovers journals of failed nodes
339 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
340 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
341 * 15. gfs2_control unblocks normal locking when all journals are recovered
343 * - failures during recovery
345 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
346 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
347 * recovering for a prior failure. gfs2_control needs a way to detect
348 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
349 * the recover_block and recover_start values.
351 * recover_done() provides a new lockspace generation number each time it
352 * is called (step 9). This generation number is saved as recover_start.
353 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
354 * recover_block = recover_start. So, while recover_block is equal to
355 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
356 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
358 * - more specific gfs2 steps in sequence above
360 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
361 * 6. recover_slot records any failed jids (maybe none)
362 * 9. recover_done sets recover_start = new generation number
363 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
364 * 12. gfs2_recover does journal recoveries for failed jids identified above
365 * 14. gfs2_control clears control_lock lvb bits for recovered jids
366 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
367 * again) then do nothing, otherwise if recover_start > recover_block
368 * then clear BLOCK_LOCKS.
370 * - parallel recovery steps across all nodes
372 * All nodes attempt to update the control_lock lvb with the new generation
373 * number and jid bits, but only the first to get the control_lock EX will
374 * do so; others will see that it's already done (lvb already contains new
375 * generation number.)
377 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
378 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
379 * . One node gets control_lock first and writes the lvb, others see it's done
380 * . All nodes attempt to recover jids for which they see control_lock bits set
381 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
382 * . All nodes will eventually see all lvb bits clear and unblock locks
384 * - is there a problem with clearing an lvb bit that should be set
385 * and missing a journal recovery?
388 * 2. lvb bit set for step 1
389 * 3. jid recovered for step 1
390 * 4. jid taken again (new mount)
391 * 5. jid fails (for step 4)
392 * 6. lvb bit set for step 5 (will already be set)
393 * 7. lvb bit cleared for step 3
395 * This is not a problem because the failure in step 5 does not
396 * require recovery, because the mount in step 4 could not have
397 * progressed far enough to unblock locks and access the fs. The
398 * control_mount() function waits for all recoveries to be complete
399 * for the latest lockspace generation before ever unblocking locks
400 * and returning. The mount in step 4 waits until the recovery in
403 * - special case of first mounter: first node to mount the fs
405 * The first node to mount a gfs2 fs needs to check all the journals
406 * and recover any that need recovery before other nodes are allowed
407 * to mount the fs. (Others may begin mounting, but they must wait
408 * for the first mounter to be done before taking locks on the fs
409 * or accessing the fs.) This has two parts:
411 * 1. The mounted_lock tells a node it's the first to mount the fs.
412 * Each node holds the mounted_lock in PR while it's mounted.
413 * Each node tries to acquire the mounted_lock in EX when it mounts.
414 * If a node is granted the mounted_lock EX it means there are no
415 * other mounted nodes (no PR locks exist), and it is the first mounter.
416 * The mounted_lock is demoted to PR when first recovery is done, so
417 * others will fail to get an EX lock, but will get a PR lock.
419 * 2. The control_lock blocks others in control_mount() while the first
420 * mounter is doing first mount recovery of all journals.
421 * A mounting node needs to acquire control_lock in EX mode before
422 * it can proceed. The first mounter holds control_lock in EX while doing
423 * the first mount recovery, blocking mounts from other nodes, then demotes
424 * control_lock to NL when it's done (others_may_mount/first_done),
425 * allowing other nodes to continue mounting.
428 * control_lock EX/NOQUEUE success
429 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
431 * do first mounter recovery
432 * mounted_lock EX->PR
433 * control_lock EX->NL, write lvb generation
436 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
437 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
438 * mounted_lock PR/NOQUEUE success
439 * read lvb generation
440 * control_lock EX->NL
443 * - mount during recovery
445 * If a node mounts while others are doing recovery (not first mounter),
446 * the mounting node will get its initial recover_done() callback without
447 * having seen any previous failures/callbacks.
449 * It must wait for all recoveries preceding its mount to be finished
450 * before it unblocks locks. It does this by repeating the "other mounter"
451 * steps above until the lvb generation number is >= its mount generation
452 * number (from initial recover_done) and all lvb bits are clear.
454 * - control_lock lvb format
456 * 4 bytes generation number: the latest dlm lockspace generation number
457 * from recover_done callback. Indicates the jid bitmap has been updated
458 * to reflect all slot failures through that generation.
460 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
461 * that jid N needs recovery.
464 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
466 static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
470 memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE);
471 memcpy(&gen, lvb_bits, sizeof(uint32_t));
472 *lvb_gen = le32_to_cpu(gen);
475 static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
479 memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE);
480 gen = cpu_to_le32(lvb_gen);
481 memcpy(ls->ls_control_lvb, &gen, sizeof(uint32_t));
484 static int all_jid_bits_clear(char *lvb)
487 for (i = JID_BITMAP_OFFSET; i < GDLM_LVB_SIZE; i++) {
494 static void sync_wait_cb(void *arg)
496 struct lm_lockstruct *ls = arg;
497 complete(&ls->ls_sync_wait);
500 static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
502 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
505 error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
507 fs_err(sdp, "%s lkid %x error %d\n",
508 name, lksb->sb_lkid, error);
512 wait_for_completion(&ls->ls_sync_wait);
514 if (lksb->sb_status != -DLM_EUNLOCK) {
515 fs_err(sdp, "%s lkid %x status %d\n",
516 name, lksb->sb_lkid, lksb->sb_status);
522 static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
523 unsigned int num, struct dlm_lksb *lksb, char *name)
525 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
526 char strname[GDLM_STRNAME_BYTES];
529 memset(strname, 0, GDLM_STRNAME_BYTES);
530 snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
532 error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
533 strname, GDLM_STRNAME_BYTES - 1,
534 0, sync_wait_cb, ls, NULL);
536 fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
537 name, lksb->sb_lkid, flags, mode, error);
541 wait_for_completion(&ls->ls_sync_wait);
543 status = lksb->sb_status;
545 if (status && status != -EAGAIN) {
546 fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
547 name, lksb->sb_lkid, flags, mode, status);
553 static int mounted_unlock(struct gfs2_sbd *sdp)
555 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
556 return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
559 static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
561 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
562 return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
563 &ls->ls_mounted_lksb, "mounted_lock");
566 static int control_unlock(struct gfs2_sbd *sdp)
568 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
569 return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
572 static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
574 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
575 return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK,
576 &ls->ls_control_lksb, "control_lock");
579 static void gfs2_control_func(struct work_struct *work)
581 struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
582 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
583 char lvb_bits[GDLM_LVB_SIZE];
584 uint32_t block_gen, start_gen, lvb_gen, flags;
590 spin_lock(&ls->ls_recover_spin);
592 * No MOUNT_DONE means we're still mounting; control_mount()
593 * will set this flag, after which this thread will take over
594 * all further clearing of BLOCK_LOCKS.
596 * FIRST_MOUNT means this node is doing first mounter recovery,
597 * for which recovery control is handled by
598 * control_mount()/control_first_done(), not this thread.
600 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
601 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
602 spin_unlock(&ls->ls_recover_spin);
605 block_gen = ls->ls_recover_block;
606 start_gen = ls->ls_recover_start;
607 spin_unlock(&ls->ls_recover_spin);
610 * Equal block_gen and start_gen implies we are between
611 * recover_prep and recover_done callbacks, which means
612 * dlm recovery is in progress and dlm locking is blocked.
613 * There's no point trying to do any work until recover_done.
616 if (block_gen == start_gen)
620 * Propagate recover_submit[] and recover_result[] to lvb:
621 * dlm_recoverd adds to recover_submit[] jids needing recovery
622 * gfs2_recover adds to recover_result[] journal recovery results
624 * set lvb bit for jids in recover_submit[] if the lvb has not
625 * yet been updated for the generation of the failure
627 * clear lvb bit for jids in recover_result[] if the result of
628 * the journal recovery is SUCCESS
631 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
633 fs_err(sdp, "control lock EX error %d\n", error);
637 control_lvb_read(ls, &lvb_gen, lvb_bits);
639 spin_lock(&ls->ls_recover_spin);
640 if (block_gen != ls->ls_recover_block ||
641 start_gen != ls->ls_recover_start) {
642 fs_info(sdp, "recover generation %u block1 %u %u\n",
643 start_gen, block_gen, ls->ls_recover_block);
644 spin_unlock(&ls->ls_recover_spin);
645 control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
649 recover_size = ls->ls_recover_size;
651 if (lvb_gen <= start_gen) {
653 * Clear lvb bits for jids we've successfully recovered.
654 * Because all nodes attempt to recover failed journals,
655 * a journal can be recovered multiple times successfully
656 * in succession. Only the first will really do recovery,
657 * the others find it clean, but still report a successful
658 * recovery. So, another node may have already recovered
659 * the jid and cleared the lvb bit for it.
661 for (i = 0; i < recover_size; i++) {
662 if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
665 ls->ls_recover_result[i] = 0;
667 if (!test_bit_le(i, lvb_bits + JID_BITMAP_OFFSET))
670 __clear_bit_le(i, lvb_bits + JID_BITMAP_OFFSET);
675 if (lvb_gen == start_gen) {
677 * Failed slots before start_gen are already set in lvb.
679 for (i = 0; i < recover_size; i++) {
680 if (!ls->ls_recover_submit[i])
682 if (ls->ls_recover_submit[i] < lvb_gen)
683 ls->ls_recover_submit[i] = 0;
685 } else if (lvb_gen < start_gen) {
687 * Failed slots before start_gen are not yet set in lvb.
689 for (i = 0; i < recover_size; i++) {
690 if (!ls->ls_recover_submit[i])
692 if (ls->ls_recover_submit[i] < start_gen) {
693 ls->ls_recover_submit[i] = 0;
694 __set_bit_le(i, lvb_bits + JID_BITMAP_OFFSET);
697 /* even if there are no bits to set, we need to write the
698 latest generation to the lvb */
702 * we should be getting a recover_done() for lvb_gen soon
705 spin_unlock(&ls->ls_recover_spin);
708 control_lvb_write(ls, start_gen, lvb_bits);
709 flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
711 flags = DLM_LKF_CONVERT;
714 error = control_lock(sdp, DLM_LOCK_NL, flags);
716 fs_err(sdp, "control lock NL error %d\n", error);
721 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
722 * and clear a jid bit in the lvb if the recovery is a success.
723 * Eventually all journals will be recovered, all jid bits will
724 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
727 for (i = 0; i < recover_size; i++) {
728 if (test_bit_le(i, lvb_bits + JID_BITMAP_OFFSET)) {
729 fs_info(sdp, "recover generation %u jid %d\n",
731 gfs2_recover_set(sdp, i);
739 * No more jid bits set in lvb, all recovery is done, unblock locks
740 * (unless a new recover_prep callback has occured blocking locks
741 * again while working above)
744 spin_lock(&ls->ls_recover_spin);
745 if (ls->ls_recover_block == block_gen &&
746 ls->ls_recover_start == start_gen) {
747 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
748 spin_unlock(&ls->ls_recover_spin);
749 fs_info(sdp, "recover generation %u done\n", start_gen);
750 gfs2_glock_thaw(sdp);
752 fs_info(sdp, "recover generation %u block2 %u %u\n",
753 start_gen, block_gen, ls->ls_recover_block);
754 spin_unlock(&ls->ls_recover_spin);
758 static int control_mount(struct gfs2_sbd *sdp)
760 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
761 char lvb_bits[GDLM_LVB_SIZE];
762 uint32_t start_gen, block_gen, mount_gen, lvb_gen;
767 memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
768 memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
769 memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
770 ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
771 init_completion(&ls->ls_sync_wait);
773 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
775 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
777 fs_err(sdp, "control_mount control_lock NL error %d\n", error);
781 error = mounted_lock(sdp, DLM_LOCK_NL, 0);
783 fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
787 mounted_mode = DLM_LOCK_NL;
790 if (retries++ && signal_pending(current)) {
796 * We always start with both locks in NL. control_lock is
797 * demoted to NL below so we don't need to do it here.
800 if (mounted_mode != DLM_LOCK_NL) {
801 error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
804 mounted_mode = DLM_LOCK_NL;
808 * Other nodes need to do some work in dlm recovery and gfs2_control
809 * before the recover_done and control_lock will be ready for us below.
810 * A delay here is not required but often avoids having to retry.
813 msleep_interruptible(500);
816 * Acquire control_lock in EX and mounted_lock in either EX or PR.
817 * control_lock lvb keeps track of any pending journal recoveries.
818 * mounted_lock indicates if any other nodes have the fs mounted.
821 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
822 if (error == -EAGAIN) {
825 fs_err(sdp, "control_mount control_lock EX error %d\n", error);
829 error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
831 mounted_mode = DLM_LOCK_EX;
833 } else if (error != -EAGAIN) {
834 fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
838 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
840 mounted_mode = DLM_LOCK_PR;
843 /* not even -EAGAIN should happen here */
844 fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
850 * If we got both locks above in EX, then we're the first mounter.
851 * If not, then we need to wait for the control_lock lvb to be
852 * updated by other mounted nodes to reflect our mount generation.
854 * In simple first mounter cases, first mounter will see zero lvb_gen,
855 * but in cases where all existing nodes leave/fail before mounting
856 * nodes finish control_mount, then all nodes will be mounting and
857 * lvb_gen will be non-zero.
860 control_lvb_read(ls, &lvb_gen, lvb_bits);
862 if (lvb_gen == 0xFFFFFFFF) {
863 /* special value to force mount attempts to fail */
864 fs_err(sdp, "control_mount control_lock disabled\n");
869 if (mounted_mode == DLM_LOCK_EX) {
870 /* first mounter, keep both EX while doing first recovery */
871 spin_lock(&ls->ls_recover_spin);
872 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
873 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
874 set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
875 spin_unlock(&ls->ls_recover_spin);
876 fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
880 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
885 * We are not first mounter, now we need to wait for the control_lock
886 * lvb generation to be >= the generation from our first recover_done
887 * and all lvb bits to be clear (no pending journal recoveries.)
890 if (!all_jid_bits_clear(lvb_bits)) {
891 /* journals need recovery, wait until all are clear */
892 fs_info(sdp, "control_mount wait for journal recovery\n");
896 spin_lock(&ls->ls_recover_spin);
897 block_gen = ls->ls_recover_block;
898 start_gen = ls->ls_recover_start;
899 mount_gen = ls->ls_recover_mount;
901 if (lvb_gen < mount_gen) {
902 /* wait for mounted nodes to update control_lock lvb to our
903 generation, which might include new recovery bits set */
904 fs_info(sdp, "control_mount wait1 block %u start %u mount %u "
905 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
906 lvb_gen, ls->ls_recover_flags);
907 spin_unlock(&ls->ls_recover_spin);
911 if (lvb_gen != start_gen) {
912 /* wait for mounted nodes to update control_lock lvb to the
913 latest recovery generation */
914 fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
915 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
916 lvb_gen, ls->ls_recover_flags);
917 spin_unlock(&ls->ls_recover_spin);
921 if (block_gen == start_gen) {
922 /* dlm recovery in progress, wait for it to finish */
923 fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
924 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
925 lvb_gen, ls->ls_recover_flags);
926 spin_unlock(&ls->ls_recover_spin);
930 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
931 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
932 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
933 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
934 spin_unlock(&ls->ls_recover_spin);
943 static int dlm_recovery_wait(void *word)
949 static int control_first_done(struct gfs2_sbd *sdp)
951 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
952 char lvb_bits[GDLM_LVB_SIZE];
953 uint32_t start_gen, block_gen;
957 spin_lock(&ls->ls_recover_spin);
958 start_gen = ls->ls_recover_start;
959 block_gen = ls->ls_recover_block;
961 if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
962 !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
963 !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
964 /* sanity check, should not happen */
965 fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
966 start_gen, block_gen, ls->ls_recover_flags);
967 spin_unlock(&ls->ls_recover_spin);
972 if (start_gen == block_gen) {
974 * Wait for the end of a dlm recovery cycle to switch from
975 * first mounter recovery. We can ignore any recover_slot
976 * callbacks between the recover_prep and next recover_done
977 * because we are still the first mounter and any failed nodes
978 * have not fully mounted, so they don't need recovery.
980 spin_unlock(&ls->ls_recover_spin);
981 fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
983 wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
984 dlm_recovery_wait, TASK_UNINTERRUPTIBLE);
988 clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
989 set_bit(DFL_FIRST_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);
994 memset(lvb_bits, 0, sizeof(lvb_bits));
995 control_lvb_write(ls, start_gen, lvb_bits);
997 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
999 fs_err(sdp, "control_first_done mounted PR error %d\n", error);
1001 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
1003 fs_err(sdp, "control_first_done control NL error %d\n", error);
1009 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1010 * to accomodate the largest slot number. (NB dlm slot numbers start at 1,
1011 * gfs2 jids start at 0, so jid = slot - 1)
1014 #define RECOVER_SIZE_INC 16
1016 static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1019 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1020 uint32_t *submit = NULL;
1021 uint32_t *result = NULL;
1022 uint32_t old_size, new_size;
1026 for (i = 0; i < num_slots; i++) {
1027 if (max_jid < slots[i].slot - 1)
1028 max_jid = slots[i].slot - 1;
1031 old_size = ls->ls_recover_size;
1033 if (old_size >= max_jid + 1)
1036 new_size = old_size + RECOVER_SIZE_INC;
1038 submit = kzalloc(new_size * sizeof(uint32_t), GFP_NOFS);
1039 result = kzalloc(new_size * sizeof(uint32_t), GFP_NOFS);
1040 if (!submit || !result) {
1046 spin_lock(&ls->ls_recover_spin);
1047 memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1048 memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1049 kfree(ls->ls_recover_submit);
1050 kfree(ls->ls_recover_result);
1051 ls->ls_recover_submit = submit;
1052 ls->ls_recover_result = result;
1053 ls->ls_recover_size = new_size;
1054 spin_unlock(&ls->ls_recover_spin);
1058 static void free_recover_size(struct lm_lockstruct *ls)
1060 kfree(ls->ls_recover_submit);
1061 kfree(ls->ls_recover_result);
1062 ls->ls_recover_submit = NULL;
1063 ls->ls_recover_result = NULL;
1064 ls->ls_recover_size = 0;
1067 /* dlm calls before it does lock recovery */
1069 static void gdlm_recover_prep(void *arg)
1071 struct gfs2_sbd *sdp = arg;
1072 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1074 spin_lock(&ls->ls_recover_spin);
1075 ls->ls_recover_block = ls->ls_recover_start;
1076 set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1078 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1079 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1080 spin_unlock(&ls->ls_recover_spin);
1083 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1084 spin_unlock(&ls->ls_recover_spin);
1087 /* dlm calls after recover_prep has been completed on all lockspace members;
1088 identifies slot/jid of failed member */
1090 static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1092 struct gfs2_sbd *sdp = arg;
1093 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1094 int jid = slot->slot - 1;
1096 spin_lock(&ls->ls_recover_spin);
1097 if (ls->ls_recover_size < jid + 1) {
1098 fs_err(sdp, "recover_slot jid %d gen %u short size %d",
1099 jid, ls->ls_recover_block, ls->ls_recover_size);
1100 spin_unlock(&ls->ls_recover_spin);
1104 if (ls->ls_recover_submit[jid]) {
1105 fs_info(sdp, "recover_slot jid %d gen %u prev %u",
1106 jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1108 ls->ls_recover_submit[jid] = ls->ls_recover_block;
1109 spin_unlock(&ls->ls_recover_spin);
1112 /* dlm calls after recover_slot and after it completes lock recovery */
1114 static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1115 int our_slot, uint32_t generation)
1117 struct gfs2_sbd *sdp = arg;
1118 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1120 /* ensure the ls jid arrays are large enough */
1121 set_recover_size(sdp, slots, num_slots);
1123 spin_lock(&ls->ls_recover_spin);
1124 ls->ls_recover_start = generation;
1126 if (!ls->ls_recover_mount) {
1127 ls->ls_recover_mount = generation;
1128 ls->ls_jid = our_slot - 1;
1131 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1132 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1134 clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1135 smp_mb__after_clear_bit();
1136 wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1137 spin_unlock(&ls->ls_recover_spin);
1140 /* gfs2_recover thread has a journal recovery result */
1142 static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1143 unsigned int result)
1145 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1147 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1150 /* don't care about the recovery of own journal during mount */
1151 if (jid == ls->ls_jid)
1154 spin_lock(&ls->ls_recover_spin);
1155 if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1156 spin_unlock(&ls->ls_recover_spin);
1159 if (ls->ls_recover_size < jid + 1) {
1160 fs_err(sdp, "recovery_result jid %d short size %d",
1161 jid, ls->ls_recover_size);
1162 spin_unlock(&ls->ls_recover_spin);
1166 fs_info(sdp, "recover jid %d result %s\n", jid,
1167 result == LM_RD_GAVEUP ? "busy" : "success");
1169 ls->ls_recover_result[jid] = result;
1171 /* GAVEUP means another node is recovering the journal; delay our
1172 next attempt to recover it, to give the other node a chance to
1173 finish before trying again */
1175 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1176 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1177 result == LM_RD_GAVEUP ? HZ : 0);
1178 spin_unlock(&ls->ls_recover_spin);
1181 const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1182 .recover_prep = gdlm_recover_prep,
1183 .recover_slot = gdlm_recover_slot,
1184 .recover_done = gdlm_recover_done,
1187 static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1189 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1190 char cluster[GFS2_LOCKNAME_LEN];
1193 int error, ops_result;
1196 * initialize everything
1199 INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1200 spin_lock_init(&ls->ls_recover_spin);
1201 ls->ls_recover_flags = 0;
1202 ls->ls_recover_mount = 0;
1203 ls->ls_recover_start = 0;
1204 ls->ls_recover_block = 0;
1205 ls->ls_recover_size = 0;
1206 ls->ls_recover_submit = NULL;
1207 ls->ls_recover_result = NULL;
1209 error = set_recover_size(sdp, NULL, 0);
1214 * prepare dlm_new_lockspace args
1217 fsname = strchr(table, ':');
1219 fs_info(sdp, "no fsname found\n");
1223 memset(cluster, 0, sizeof(cluster));
1224 memcpy(cluster, table, strlen(table) - strlen(fsname));
1227 flags = DLM_LSFL_FS | DLM_LSFL_NEWEXCL;
1230 * create/join lockspace
1233 error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1234 &gdlm_lockspace_ops, sdp, &ops_result,
1237 fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1241 if (ops_result < 0) {
1243 * dlm does not support ops callbacks,
1244 * old dlm_controld/gfs_controld are used, try without ops.
1246 fs_info(sdp, "dlm lockspace ops not used\n");
1247 free_recover_size(ls);
1248 set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1252 if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1253 fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1259 * control_mount() uses control_lock to determine first mounter,
1260 * and for later mounts, waits for any recoveries to be cleared.
1263 error = control_mount(sdp);
1265 fs_err(sdp, "mount control error %d\n", error);
1269 ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1270 clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
1271 smp_mb__after_clear_bit();
1272 wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1276 dlm_release_lockspace(ls->ls_dlm, 2);
1278 free_recover_size(ls);
1283 static void gdlm_first_done(struct gfs2_sbd *sdp)
1285 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1288 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1291 error = control_first_done(sdp);
1293 fs_err(sdp, "mount first_done error %d\n", error);
1296 static void gdlm_unmount(struct gfs2_sbd *sdp)
1298 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1300 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1303 /* wait for gfs2_control_wq to be done with this mount */
1305 spin_lock(&ls->ls_recover_spin);
1306 set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1307 spin_unlock(&ls->ls_recover_spin);
1308 flush_delayed_work(&sdp->sd_control_work);
1310 /* mounted_lock and control_lock will be purged in dlm recovery */
1313 dlm_release_lockspace(ls->ls_dlm, 2);
1317 free_recover_size(ls);
1320 static const match_table_t dlm_tokens = {
1321 { Opt_jid, "jid=%d"},
1323 { Opt_first, "first=%d"},
1324 { Opt_nodir, "nodir=%d"},
1328 const struct lm_lockops gfs2_dlm_ops = {
1329 .lm_proto_name = "lock_dlm",
1330 .lm_mount = gdlm_mount,
1331 .lm_first_done = gdlm_first_done,
1332 .lm_recovery_result = gdlm_recovery_result,
1333 .lm_unmount = gdlm_unmount,
1334 .lm_put_lock = gdlm_put_lock,
1335 .lm_lock = gdlm_lock,
1336 .lm_cancel = gdlm_cancel,
1337 .lm_tokens = &dlm_tokens,