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)
302 clear_bit(GLF_BLOCKING, &gl->gl_flags);
303 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
304 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
305 gfs2_update_request_times(gl);
307 /* don't want to call dlm if we've unmounted the lock protocol */
308 if (test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
310 /* don't want to skip dlm_unlock writing the lvb when lock has one */
312 if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
313 !gl->gl_lksb.sb_lvbptr)
317 error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_VALBLK,
319 if (error == -EBUSY) {
325 fs_err(sdp, "gdlm_unlock %x,%llx err=%d\n",
327 (unsigned long long)gl->gl_name.ln_number, error);
335 static void gdlm_cancel(struct gfs2_glock *gl)
337 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
338 dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
342 * dlm/gfs2 recovery coordination using dlm_recover callbacks
344 * 0. gfs2 checks for another cluster node withdraw, needing journal replay
345 * 1. dlm_controld sees lockspace members change
346 * 2. dlm_controld blocks dlm-kernel locking activity
347 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
348 * 4. dlm_controld starts and finishes its own user level recovery
349 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
350 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
351 * 7. dlm_recoverd does its own lock recovery
352 * 8. dlm_recoverd unblocks dlm-kernel locking activity
353 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
354 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
355 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
356 * 12. gfs2_recover dequeues and recovers journals of failed nodes
357 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
358 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
359 * 15. gfs2_control unblocks normal locking when all journals are recovered
361 * - failures during recovery
363 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
364 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
365 * recovering for a prior failure. gfs2_control needs a way to detect
366 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
367 * the recover_block and recover_start values.
369 * recover_done() provides a new lockspace generation number each time it
370 * is called (step 9). This generation number is saved as recover_start.
371 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
372 * recover_block = recover_start. So, while recover_block is equal to
373 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
374 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
376 * - more specific gfs2 steps in sequence above
378 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
379 * 6. recover_slot records any failed jids (maybe none)
380 * 9. recover_done sets recover_start = new generation number
381 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
382 * 12. gfs2_recover does journal recoveries for failed jids identified above
383 * 14. gfs2_control clears control_lock lvb bits for recovered jids
384 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
385 * again) then do nothing, otherwise if recover_start > recover_block
386 * then clear BLOCK_LOCKS.
388 * - parallel recovery steps across all nodes
390 * All nodes attempt to update the control_lock lvb with the new generation
391 * number and jid bits, but only the first to get the control_lock EX will
392 * do so; others will see that it's already done (lvb already contains new
393 * generation number.)
395 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
396 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
397 * . One node gets control_lock first and writes the lvb, others see it's done
398 * . All nodes attempt to recover jids for which they see control_lock bits set
399 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
400 * . All nodes will eventually see all lvb bits clear and unblock locks
402 * - is there a problem with clearing an lvb bit that should be set
403 * and missing a journal recovery?
406 * 2. lvb bit set for step 1
407 * 3. jid recovered for step 1
408 * 4. jid taken again (new mount)
409 * 5. jid fails (for step 4)
410 * 6. lvb bit set for step 5 (will already be set)
411 * 7. lvb bit cleared for step 3
413 * This is not a problem because the failure in step 5 does not
414 * require recovery, because the mount in step 4 could not have
415 * progressed far enough to unblock locks and access the fs. The
416 * control_mount() function waits for all recoveries to be complete
417 * for the latest lockspace generation before ever unblocking locks
418 * and returning. The mount in step 4 waits until the recovery in
421 * - special case of first mounter: first node to mount the fs
423 * The first node to mount a gfs2 fs needs to check all the journals
424 * and recover any that need recovery before other nodes are allowed
425 * to mount the fs. (Others may begin mounting, but they must wait
426 * for the first mounter to be done before taking locks on the fs
427 * or accessing the fs.) This has two parts:
429 * 1. The mounted_lock tells a node it's the first to mount the fs.
430 * Each node holds the mounted_lock in PR while it's mounted.
431 * Each node tries to acquire the mounted_lock in EX when it mounts.
432 * If a node is granted the mounted_lock EX it means there are no
433 * other mounted nodes (no PR locks exist), and it is the first mounter.
434 * The mounted_lock is demoted to PR when first recovery is done, so
435 * others will fail to get an EX lock, but will get a PR lock.
437 * 2. The control_lock blocks others in control_mount() while the first
438 * mounter is doing first mount recovery of all journals.
439 * A mounting node needs to acquire control_lock in EX mode before
440 * it can proceed. The first mounter holds control_lock in EX while doing
441 * the first mount recovery, blocking mounts from other nodes, then demotes
442 * control_lock to NL when it's done (others_may_mount/first_done),
443 * allowing other nodes to continue mounting.
446 * control_lock EX/NOQUEUE success
447 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
449 * do first mounter recovery
450 * mounted_lock EX->PR
451 * control_lock EX->NL, write lvb generation
454 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
455 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
456 * mounted_lock PR/NOQUEUE success
457 * read lvb generation
458 * control_lock EX->NL
461 * - mount during recovery
463 * If a node mounts while others are doing recovery (not first mounter),
464 * the mounting node will get its initial recover_done() callback without
465 * having seen any previous failures/callbacks.
467 * It must wait for all recoveries preceding its mount to be finished
468 * before it unblocks locks. It does this by repeating the "other mounter"
469 * steps above until the lvb generation number is >= its mount generation
470 * number (from initial recover_done) and all lvb bits are clear.
472 * - control_lock lvb format
474 * 4 bytes generation number: the latest dlm lockspace generation number
475 * from recover_done callback. Indicates the jid bitmap has been updated
476 * to reflect all slot failures through that generation.
478 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
479 * that jid N needs recovery.
482 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
484 static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
488 memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE);
489 memcpy(&gen, lvb_bits, sizeof(__le32));
490 *lvb_gen = le32_to_cpu(gen);
493 static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
497 memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE);
498 gen = cpu_to_le32(lvb_gen);
499 memcpy(ls->ls_control_lvb, &gen, sizeof(__le32));
502 static int all_jid_bits_clear(char *lvb)
504 return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0,
505 GDLM_LVB_SIZE - JID_BITMAP_OFFSET);
508 static void sync_wait_cb(void *arg)
510 struct lm_lockstruct *ls = arg;
511 complete(&ls->ls_sync_wait);
514 static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
516 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
519 error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
521 fs_err(sdp, "%s lkid %x error %d\n",
522 name, lksb->sb_lkid, error);
526 wait_for_completion(&ls->ls_sync_wait);
528 if (lksb->sb_status != -DLM_EUNLOCK) {
529 fs_err(sdp, "%s lkid %x status %d\n",
530 name, lksb->sb_lkid, lksb->sb_status);
536 static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
537 unsigned int num, struct dlm_lksb *lksb, char *name)
539 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
540 char strname[GDLM_STRNAME_BYTES];
543 memset(strname, 0, GDLM_STRNAME_BYTES);
544 snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
546 error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
547 strname, GDLM_STRNAME_BYTES - 1,
548 0, sync_wait_cb, ls, NULL);
550 fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
551 name, lksb->sb_lkid, flags, mode, error);
555 wait_for_completion(&ls->ls_sync_wait);
557 status = lksb->sb_status;
559 if (status && status != -EAGAIN) {
560 fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
561 name, lksb->sb_lkid, flags, mode, status);
567 static int mounted_unlock(struct gfs2_sbd *sdp)
569 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
570 return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
573 static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
575 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
576 return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
577 &ls->ls_mounted_lksb, "mounted_lock");
580 static int control_unlock(struct gfs2_sbd *sdp)
582 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
583 return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
586 static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
588 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
589 return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK,
590 &ls->ls_control_lksb, "control_lock");
594 * remote_withdraw - react to a node withdrawing from the file system
595 * @sdp: The superblock
597 static void remote_withdraw(struct gfs2_sbd *sdp)
599 struct gfs2_jdesc *jd;
600 int ret = 0, count = 0;
602 list_for_each_entry(jd, &sdp->sd_jindex_list, jd_list) {
603 if (jd->jd_jid == sdp->sd_lockstruct.ls_jid)
605 ret = gfs2_recover_journal(jd, true);
611 /* Now drop the additional reference we acquired */
612 fs_err(sdp, "Journals checked: %d, ret = %d.\n", count, ret);
615 static void gfs2_control_func(struct work_struct *work)
617 struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
618 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
619 uint32_t block_gen, start_gen, lvb_gen, flags;
625 /* First check for other nodes that may have done a withdraw. */
626 if (test_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags)) {
627 remote_withdraw(sdp);
628 clear_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags);
632 spin_lock(&ls->ls_recover_spin);
634 * No MOUNT_DONE means we're still mounting; control_mount()
635 * will set this flag, after which this thread will take over
636 * all further clearing of BLOCK_LOCKS.
638 * FIRST_MOUNT means this node is doing first mounter recovery,
639 * for which recovery control is handled by
640 * control_mount()/control_first_done(), not this thread.
642 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
643 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
644 spin_unlock(&ls->ls_recover_spin);
647 block_gen = ls->ls_recover_block;
648 start_gen = ls->ls_recover_start;
649 spin_unlock(&ls->ls_recover_spin);
652 * Equal block_gen and start_gen implies we are between
653 * recover_prep and recover_done callbacks, which means
654 * dlm recovery is in progress and dlm locking is blocked.
655 * There's no point trying to do any work until recover_done.
658 if (block_gen == start_gen)
662 * Propagate recover_submit[] and recover_result[] to lvb:
663 * dlm_recoverd adds to recover_submit[] jids needing recovery
664 * gfs2_recover adds to recover_result[] journal recovery results
666 * set lvb bit for jids in recover_submit[] if the lvb has not
667 * yet been updated for the generation of the failure
669 * clear lvb bit for jids in recover_result[] if the result of
670 * the journal recovery is SUCCESS
673 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
675 fs_err(sdp, "control lock EX error %d\n", error);
679 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
681 spin_lock(&ls->ls_recover_spin);
682 if (block_gen != ls->ls_recover_block ||
683 start_gen != ls->ls_recover_start) {
684 fs_info(sdp, "recover generation %u block1 %u %u\n",
685 start_gen, block_gen, ls->ls_recover_block);
686 spin_unlock(&ls->ls_recover_spin);
687 control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
691 recover_size = ls->ls_recover_size;
693 if (lvb_gen <= start_gen) {
695 * Clear lvb bits for jids we've successfully recovered.
696 * Because all nodes attempt to recover failed journals,
697 * a journal can be recovered multiple times successfully
698 * in succession. Only the first will really do recovery,
699 * the others find it clean, but still report a successful
700 * recovery. So, another node may have already recovered
701 * the jid and cleared the lvb bit for it.
703 for (i = 0; i < recover_size; i++) {
704 if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
707 ls->ls_recover_result[i] = 0;
709 if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
712 __clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
717 if (lvb_gen == start_gen) {
719 * Failed slots before start_gen are already set in lvb.
721 for (i = 0; i < recover_size; i++) {
722 if (!ls->ls_recover_submit[i])
724 if (ls->ls_recover_submit[i] < lvb_gen)
725 ls->ls_recover_submit[i] = 0;
727 } else if (lvb_gen < start_gen) {
729 * Failed slots before start_gen are not yet set in lvb.
731 for (i = 0; i < recover_size; i++) {
732 if (!ls->ls_recover_submit[i])
734 if (ls->ls_recover_submit[i] < start_gen) {
735 ls->ls_recover_submit[i] = 0;
736 __set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
739 /* even if there are no bits to set, we need to write the
740 latest generation to the lvb */
744 * we should be getting a recover_done() for lvb_gen soon
747 spin_unlock(&ls->ls_recover_spin);
750 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
751 flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
753 flags = DLM_LKF_CONVERT;
756 error = control_lock(sdp, DLM_LOCK_NL, flags);
758 fs_err(sdp, "control lock NL error %d\n", error);
763 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
764 * and clear a jid bit in the lvb if the recovery is a success.
765 * Eventually all journals will be recovered, all jid bits will
766 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
769 for (i = 0; i < recover_size; i++) {
770 if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
771 fs_info(sdp, "recover generation %u jid %d\n",
773 gfs2_recover_set(sdp, i);
781 * No more jid bits set in lvb, all recovery is done, unblock locks
782 * (unless a new recover_prep callback has occured blocking locks
783 * again while working above)
786 spin_lock(&ls->ls_recover_spin);
787 if (ls->ls_recover_block == block_gen &&
788 ls->ls_recover_start == start_gen) {
789 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
790 spin_unlock(&ls->ls_recover_spin);
791 fs_info(sdp, "recover generation %u done\n", start_gen);
792 gfs2_glock_thaw(sdp);
794 fs_info(sdp, "recover generation %u block2 %u %u\n",
795 start_gen, block_gen, ls->ls_recover_block);
796 spin_unlock(&ls->ls_recover_spin);
800 static int control_mount(struct gfs2_sbd *sdp)
802 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
803 uint32_t start_gen, block_gen, mount_gen, lvb_gen;
808 memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
809 memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
810 memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
811 ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
812 init_completion(&ls->ls_sync_wait);
814 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
816 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
818 fs_err(sdp, "control_mount control_lock NL error %d\n", error);
822 error = mounted_lock(sdp, DLM_LOCK_NL, 0);
824 fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
828 mounted_mode = DLM_LOCK_NL;
831 if (retries++ && signal_pending(current)) {
837 * We always start with both locks in NL. control_lock is
838 * demoted to NL below so we don't need to do it here.
841 if (mounted_mode != DLM_LOCK_NL) {
842 error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
845 mounted_mode = DLM_LOCK_NL;
849 * Other nodes need to do some work in dlm recovery and gfs2_control
850 * before the recover_done and control_lock will be ready for us below.
851 * A delay here is not required but often avoids having to retry.
854 msleep_interruptible(500);
857 * Acquire control_lock in EX and mounted_lock in either EX or PR.
858 * control_lock lvb keeps track of any pending journal recoveries.
859 * mounted_lock indicates if any other nodes have the fs mounted.
862 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
863 if (error == -EAGAIN) {
866 fs_err(sdp, "control_mount control_lock EX error %d\n", error);
871 * If we're a spectator, we don't want to take the lock in EX because
872 * we cannot do the first-mount responsibility it implies: recovery.
874 if (sdp->sd_args.ar_spectator)
877 error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
879 mounted_mode = DLM_LOCK_EX;
881 } else if (error != -EAGAIN) {
882 fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
886 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
888 mounted_mode = DLM_LOCK_PR;
891 /* not even -EAGAIN should happen here */
892 fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
898 * If we got both locks above in EX, then we're the first mounter.
899 * If not, then we need to wait for the control_lock lvb to be
900 * updated by other mounted nodes to reflect our mount generation.
902 * In simple first mounter cases, first mounter will see zero lvb_gen,
903 * but in cases where all existing nodes leave/fail before mounting
904 * nodes finish control_mount, then all nodes will be mounting and
905 * lvb_gen will be non-zero.
908 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
910 if (lvb_gen == 0xFFFFFFFF) {
911 /* special value to force mount attempts to fail */
912 fs_err(sdp, "control_mount control_lock disabled\n");
917 if (mounted_mode == DLM_LOCK_EX) {
918 /* first mounter, keep both EX while doing first recovery */
919 spin_lock(&ls->ls_recover_spin);
920 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
921 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
922 set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
923 spin_unlock(&ls->ls_recover_spin);
924 fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
928 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
933 * We are not first mounter, now we need to wait for the control_lock
934 * lvb generation to be >= the generation from our first recover_done
935 * and all lvb bits to be clear (no pending journal recoveries.)
938 if (!all_jid_bits_clear(ls->ls_lvb_bits)) {
939 /* journals need recovery, wait until all are clear */
940 fs_info(sdp, "control_mount wait for journal recovery\n");
944 spin_lock(&ls->ls_recover_spin);
945 block_gen = ls->ls_recover_block;
946 start_gen = ls->ls_recover_start;
947 mount_gen = ls->ls_recover_mount;
949 if (lvb_gen < mount_gen) {
950 /* wait for mounted nodes to update control_lock lvb to our
951 generation, which might include new recovery bits set */
952 if (sdp->sd_args.ar_spectator) {
953 fs_info(sdp, "Recovery is required. Waiting for a "
954 "non-spectator to mount.\n");
955 msleep_interruptible(1000);
957 fs_info(sdp, "control_mount wait1 block %u start %u "
958 "mount %u lvb %u flags %lx\n", block_gen,
959 start_gen, mount_gen, lvb_gen,
960 ls->ls_recover_flags);
962 spin_unlock(&ls->ls_recover_spin);
966 if (lvb_gen != start_gen) {
967 /* wait for mounted nodes to update control_lock lvb to the
968 latest recovery generation */
969 fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
970 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
971 lvb_gen, ls->ls_recover_flags);
972 spin_unlock(&ls->ls_recover_spin);
976 if (block_gen == start_gen) {
977 /* dlm recovery in progress, wait for it to finish */
978 fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
979 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
980 lvb_gen, ls->ls_recover_flags);
981 spin_unlock(&ls->ls_recover_spin);
985 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
986 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
987 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
988 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
989 spin_unlock(&ls->ls_recover_spin);
998 static int control_first_done(struct gfs2_sbd *sdp)
1000 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1001 uint32_t start_gen, block_gen;
1005 spin_lock(&ls->ls_recover_spin);
1006 start_gen = ls->ls_recover_start;
1007 block_gen = ls->ls_recover_block;
1009 if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
1010 !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1011 !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1012 /* sanity check, should not happen */
1013 fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
1014 start_gen, block_gen, ls->ls_recover_flags);
1015 spin_unlock(&ls->ls_recover_spin);
1016 control_unlock(sdp);
1020 if (start_gen == block_gen) {
1022 * Wait for the end of a dlm recovery cycle to switch from
1023 * first mounter recovery. We can ignore any recover_slot
1024 * callbacks between the recover_prep and next recover_done
1025 * because we are still the first mounter and any failed nodes
1026 * have not fully mounted, so they don't need recovery.
1028 spin_unlock(&ls->ls_recover_spin);
1029 fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
1031 wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
1032 TASK_UNINTERRUPTIBLE);
1036 clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1037 set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
1038 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
1039 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
1040 spin_unlock(&ls->ls_recover_spin);
1042 memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE);
1043 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
1045 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
1047 fs_err(sdp, "control_first_done mounted PR error %d\n", error);
1049 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
1051 fs_err(sdp, "control_first_done control NL error %d\n", error);
1057 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1058 * to accommodate the largest slot number. (NB dlm slot numbers start at 1,
1059 * gfs2 jids start at 0, so jid = slot - 1)
1062 #define RECOVER_SIZE_INC 16
1064 static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1067 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1068 uint32_t *submit = NULL;
1069 uint32_t *result = NULL;
1070 uint32_t old_size, new_size;
1073 if (!ls->ls_lvb_bits) {
1074 ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1075 if (!ls->ls_lvb_bits)
1080 for (i = 0; i < num_slots; i++) {
1081 if (max_jid < slots[i].slot - 1)
1082 max_jid = slots[i].slot - 1;
1085 old_size = ls->ls_recover_size;
1086 new_size = old_size;
1087 while (new_size < max_jid + 1)
1088 new_size += RECOVER_SIZE_INC;
1089 if (new_size == old_size)
1092 submit = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1093 result = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1094 if (!submit || !result) {
1100 spin_lock(&ls->ls_recover_spin);
1101 memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1102 memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1103 kfree(ls->ls_recover_submit);
1104 kfree(ls->ls_recover_result);
1105 ls->ls_recover_submit = submit;
1106 ls->ls_recover_result = result;
1107 ls->ls_recover_size = new_size;
1108 spin_unlock(&ls->ls_recover_spin);
1112 static void free_recover_size(struct lm_lockstruct *ls)
1114 kfree(ls->ls_lvb_bits);
1115 kfree(ls->ls_recover_submit);
1116 kfree(ls->ls_recover_result);
1117 ls->ls_recover_submit = NULL;
1118 ls->ls_recover_result = NULL;
1119 ls->ls_recover_size = 0;
1120 ls->ls_lvb_bits = NULL;
1123 /* dlm calls before it does lock recovery */
1125 static void gdlm_recover_prep(void *arg)
1127 struct gfs2_sbd *sdp = arg;
1128 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1130 if (gfs2_withdrawn(sdp)) {
1131 fs_err(sdp, "recover_prep ignored due to withdraw.\n");
1134 spin_lock(&ls->ls_recover_spin);
1135 ls->ls_recover_block = ls->ls_recover_start;
1136 set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1138 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1139 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1140 spin_unlock(&ls->ls_recover_spin);
1143 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1144 spin_unlock(&ls->ls_recover_spin);
1147 /* dlm calls after recover_prep has been completed on all lockspace members;
1148 identifies slot/jid of failed member */
1150 static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1152 struct gfs2_sbd *sdp = arg;
1153 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1154 int jid = slot->slot - 1;
1156 if (gfs2_withdrawn(sdp)) {
1157 fs_err(sdp, "recover_slot jid %d ignored due to withdraw.\n",
1161 spin_lock(&ls->ls_recover_spin);
1162 if (ls->ls_recover_size < jid + 1) {
1163 fs_err(sdp, "recover_slot jid %d gen %u short size %d\n",
1164 jid, ls->ls_recover_block, ls->ls_recover_size);
1165 spin_unlock(&ls->ls_recover_spin);
1169 if (ls->ls_recover_submit[jid]) {
1170 fs_info(sdp, "recover_slot jid %d gen %u prev %u\n",
1171 jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1173 ls->ls_recover_submit[jid] = ls->ls_recover_block;
1174 spin_unlock(&ls->ls_recover_spin);
1177 /* dlm calls after recover_slot and after it completes lock recovery */
1179 static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1180 int our_slot, uint32_t generation)
1182 struct gfs2_sbd *sdp = arg;
1183 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1185 if (gfs2_withdrawn(sdp)) {
1186 fs_err(sdp, "recover_done ignored due to withdraw.\n");
1189 /* ensure the ls jid arrays are large enough */
1190 set_recover_size(sdp, slots, num_slots);
1192 spin_lock(&ls->ls_recover_spin);
1193 ls->ls_recover_start = generation;
1195 if (!ls->ls_recover_mount) {
1196 ls->ls_recover_mount = generation;
1197 ls->ls_jid = our_slot - 1;
1200 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1201 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1203 clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1204 smp_mb__after_atomic();
1205 wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1206 spin_unlock(&ls->ls_recover_spin);
1209 /* gfs2_recover thread has a journal recovery result */
1211 static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1212 unsigned int result)
1214 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1216 if (gfs2_withdrawn(sdp)) {
1217 fs_err(sdp, "recovery_result jid %d ignored due to withdraw.\n",
1221 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1224 /* don't care about the recovery of own journal during mount */
1225 if (jid == ls->ls_jid)
1228 spin_lock(&ls->ls_recover_spin);
1229 if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1230 spin_unlock(&ls->ls_recover_spin);
1233 if (ls->ls_recover_size < jid + 1) {
1234 fs_err(sdp, "recovery_result jid %d short size %d\n",
1235 jid, ls->ls_recover_size);
1236 spin_unlock(&ls->ls_recover_spin);
1240 fs_info(sdp, "recover jid %d result %s\n", jid,
1241 result == LM_RD_GAVEUP ? "busy" : "success");
1243 ls->ls_recover_result[jid] = result;
1245 /* GAVEUP means another node is recovering the journal; delay our
1246 next attempt to recover it, to give the other node a chance to
1247 finish before trying again */
1249 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1250 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1251 result == LM_RD_GAVEUP ? HZ : 0);
1252 spin_unlock(&ls->ls_recover_spin);
1255 static const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1256 .recover_prep = gdlm_recover_prep,
1257 .recover_slot = gdlm_recover_slot,
1258 .recover_done = gdlm_recover_done,
1261 static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1263 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1264 char cluster[GFS2_LOCKNAME_LEN];
1267 int error, ops_result;
1270 * initialize everything
1273 INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1274 spin_lock_init(&ls->ls_recover_spin);
1275 ls->ls_recover_flags = 0;
1276 ls->ls_recover_mount = 0;
1277 ls->ls_recover_start = 0;
1278 ls->ls_recover_block = 0;
1279 ls->ls_recover_size = 0;
1280 ls->ls_recover_submit = NULL;
1281 ls->ls_recover_result = NULL;
1282 ls->ls_lvb_bits = NULL;
1284 error = set_recover_size(sdp, NULL, 0);
1289 * prepare dlm_new_lockspace args
1292 fsname = strchr(table, ':');
1294 fs_info(sdp, "no fsname found\n");
1298 memset(cluster, 0, sizeof(cluster));
1299 memcpy(cluster, table, strlen(table) - strlen(fsname));
1302 flags = DLM_LSFL_NEWEXCL;
1305 * create/join lockspace
1308 error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1309 &gdlm_lockspace_ops, sdp, &ops_result,
1312 fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1316 if (ops_result < 0) {
1318 * dlm does not support ops callbacks,
1319 * old dlm_controld/gfs_controld are used, try without ops.
1321 fs_info(sdp, "dlm lockspace ops not used\n");
1322 free_recover_size(ls);
1323 set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1327 if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1328 fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1334 * control_mount() uses control_lock to determine first mounter,
1335 * and for later mounts, waits for any recoveries to be cleared.
1338 error = control_mount(sdp);
1340 fs_err(sdp, "mount control error %d\n", error);
1344 ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1345 clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
1346 smp_mb__after_atomic();
1347 wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1351 dlm_release_lockspace(ls->ls_dlm, 2);
1353 free_recover_size(ls);
1358 static void gdlm_first_done(struct gfs2_sbd *sdp)
1360 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1363 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1366 error = control_first_done(sdp);
1368 fs_err(sdp, "mount first_done error %d\n", error);
1371 static void gdlm_unmount(struct gfs2_sbd *sdp)
1373 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1375 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1378 /* wait for gfs2_control_wq to be done with this mount */
1380 spin_lock(&ls->ls_recover_spin);
1381 set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1382 spin_unlock(&ls->ls_recover_spin);
1383 flush_delayed_work(&sdp->sd_control_work);
1385 /* mounted_lock and control_lock will be purged in dlm recovery */
1388 dlm_release_lockspace(ls->ls_dlm, 2);
1392 free_recover_size(ls);
1395 static const match_table_t dlm_tokens = {
1396 { Opt_jid, "jid=%d"},
1398 { Opt_first, "first=%d"},
1399 { Opt_nodir, "nodir=%d"},
1403 const struct lm_lockops gfs2_dlm_ops = {
1404 .lm_proto_name = "lock_dlm",
1405 .lm_mount = gdlm_mount,
1406 .lm_first_done = gdlm_first_done,
1407 .lm_recovery_result = gdlm_recovery_result,
1408 .lm_unmount = gdlm_unmount,
1409 .lm_put_lock = gdlm_put_lock,
1410 .lm_lock = gdlm_lock,
1411 .lm_cancel = gdlm_cancel,
1412 .lm_tokens = &dlm_tokens,