1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
4 * Copyright 2004-2011 Red Hat, Inc.
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 #include <linux/dlm.h>
11 #include <linux/slab.h>
12 #include <linux/types.h>
13 #include <linux/delay.h>
14 #include <linux/gfs2_ondisk.h>
15 #include <linux/sched/signal.h>
23 #include "trace_gfs2.h"
26 * gfs2_update_stats - Update time based stats
27 * @s: The stats to update (local or global)
28 * @index: The index inside @s
29 * @sample: New data to include
31 static inline void gfs2_update_stats(struct gfs2_lkstats *s, unsigned index,
35 * @delta is the difference between the current rtt sample and the
36 * running average srtt. We add 1/8 of that to the srtt in order to
37 * update the current srtt estimate. The variance estimate is a bit
38 * more complicated. We subtract the current variance estimate from
39 * the abs value of the @delta and add 1/4 of that to the running
40 * total. That's equivalent to 3/4 of the current variance
41 * estimate plus 1/4 of the abs of @delta.
43 * Note that the index points at the array entry containing the
44 * smoothed mean value, and the variance is always in the following
47 * Reference: TCP/IP Illustrated, vol 2, p. 831,832
48 * All times are in units of integer nanoseconds. Unlike the TCP/IP
49 * case, they are not scaled fixed point.
52 s64 delta = sample - s->stats[index];
53 s->stats[index] += (delta >> 3);
55 s->stats[index] += (s64)(abs(delta) - s->stats[index]) >> 2;
59 * gfs2_update_reply_times - Update locking statistics
60 * @gl: The glock to update
62 * This assumes that gl->gl_dstamp has been set earlier.
64 * The rtt (lock round trip time) is an estimate of the time
65 * taken to perform a dlm lock request. We update it on each
68 * The blocking flag is set on the glock for all dlm requests
69 * which may potentially block due to lock requests from other nodes.
70 * DLM requests where the current lock state is exclusive, the
71 * requested state is null (or unlocked) or where the TRY or
72 * TRY_1CB flags are set are classified as non-blocking. All
73 * other DLM requests are counted as (potentially) blocking.
75 static inline void gfs2_update_reply_times(struct gfs2_glock *gl)
77 struct gfs2_pcpu_lkstats *lks;
78 const unsigned gltype = gl->gl_name.ln_type;
79 unsigned index = test_bit(GLF_BLOCKING, &gl->gl_flags) ?
80 GFS2_LKS_SRTTB : GFS2_LKS_SRTT;
84 rtt = ktime_to_ns(ktime_sub(ktime_get_real(), gl->gl_dstamp));
85 lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
86 gfs2_update_stats(&gl->gl_stats, index, rtt); /* Local */
87 gfs2_update_stats(&lks->lkstats[gltype], index, rtt); /* Global */
90 trace_gfs2_glock_lock_time(gl, rtt);
94 * gfs2_update_request_times - Update locking statistics
95 * @gl: The glock to update
97 * The irt (lock inter-request times) measures the average time
98 * between requests to the dlm. It is updated immediately before
102 static inline void gfs2_update_request_times(struct gfs2_glock *gl)
104 struct gfs2_pcpu_lkstats *lks;
105 const unsigned gltype = gl->gl_name.ln_type;
110 dstamp = gl->gl_dstamp;
111 gl->gl_dstamp = ktime_get_real();
112 irt = ktime_to_ns(ktime_sub(gl->gl_dstamp, dstamp));
113 lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
114 gfs2_update_stats(&gl->gl_stats, GFS2_LKS_SIRT, irt); /* Local */
115 gfs2_update_stats(&lks->lkstats[gltype], GFS2_LKS_SIRT, irt); /* Global */
119 static void gdlm_ast(void *arg)
121 struct gfs2_glock *gl = arg;
122 unsigned ret = gl->gl_state;
124 gfs2_update_reply_times(gl);
125 BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED);
127 if ((gl->gl_lksb.sb_flags & DLM_SBF_VALNOTVALID) && gl->gl_lksb.sb_lvbptr)
128 memset(gl->gl_lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
130 switch (gl->gl_lksb.sb_status) {
131 case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */
132 if (gl->gl_ops->go_free)
133 gl->gl_ops->go_free(gl);
136 case -DLM_ECANCEL: /* Cancel while getting lock */
137 ret |= LM_OUT_CANCELED;
139 case -EAGAIN: /* Try lock fails */
140 case -EDEADLK: /* Deadlock detected */
142 case -ETIMEDOUT: /* Canceled due to timeout */
145 case 0: /* Success */
147 default: /* Something unexpected */
152 if (gl->gl_lksb.sb_flags & DLM_SBF_ALTMODE) {
153 if (gl->gl_req == LM_ST_SHARED)
154 ret = LM_ST_DEFERRED;
155 else if (gl->gl_req == LM_ST_DEFERRED)
161 set_bit(GLF_INITIAL, &gl->gl_flags);
162 gfs2_glock_complete(gl, ret);
165 if (!test_bit(GLF_INITIAL, &gl->gl_flags))
166 gl->gl_lksb.sb_lkid = 0;
167 gfs2_glock_complete(gl, ret);
170 static void gdlm_bast(void *arg, int mode)
172 struct gfs2_glock *gl = arg;
176 gfs2_glock_cb(gl, LM_ST_UNLOCKED);
179 gfs2_glock_cb(gl, LM_ST_DEFERRED);
182 gfs2_glock_cb(gl, LM_ST_SHARED);
185 fs_err(gl->gl_name.ln_sbd, "unknown bast mode %d\n", mode);
190 /* convert gfs lock-state to dlm lock-mode */
192 static int make_mode(struct gfs2_sbd *sdp, const unsigned int lmstate)
197 case LM_ST_EXCLUSIVE:
204 fs_err(sdp, "unknown LM state %d\n", lmstate);
209 static u32 make_flags(struct gfs2_glock *gl, const unsigned int gfs_flags,
214 if (gl->gl_lksb.sb_lvbptr)
215 lkf |= DLM_LKF_VALBLK;
217 if (gfs_flags & LM_FLAG_TRY)
218 lkf |= DLM_LKF_NOQUEUE;
220 if (gfs_flags & LM_FLAG_TRY_1CB) {
221 lkf |= DLM_LKF_NOQUEUE;
222 lkf |= DLM_LKF_NOQUEUEBAST;
225 if (gfs_flags & LM_FLAG_PRIORITY) {
226 lkf |= DLM_LKF_NOORDER;
227 lkf |= DLM_LKF_HEADQUE;
230 if (gfs_flags & LM_FLAG_ANY) {
231 if (req == DLM_LOCK_PR)
232 lkf |= DLM_LKF_ALTCW;
233 else if (req == DLM_LOCK_CW)
234 lkf |= DLM_LKF_ALTPR;
239 if (gl->gl_lksb.sb_lkid != 0) {
240 lkf |= DLM_LKF_CONVERT;
241 if (test_bit(GLF_BLOCKING, &gl->gl_flags))
242 lkf |= DLM_LKF_QUECVT;
248 static void gfs2_reverse_hex(char *c, u64 value)
252 *c-- = hex_asc[value & 0x0f];
257 static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state,
260 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
263 char strname[GDLM_STRNAME_BYTES] = "";
265 req = make_mode(gl->gl_name.ln_sbd, req_state);
266 lkf = make_flags(gl, flags, req);
267 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
268 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
269 if (gl->gl_lksb.sb_lkid) {
270 gfs2_update_request_times(gl);
272 memset(strname, ' ', GDLM_STRNAME_BYTES - 1);
273 strname[GDLM_STRNAME_BYTES - 1] = '\0';
274 gfs2_reverse_hex(strname + 7, gl->gl_name.ln_type);
275 gfs2_reverse_hex(strname + 23, gl->gl_name.ln_number);
276 gl->gl_dstamp = ktime_get_real();
279 * Submit the actual lock request.
282 return dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname,
283 GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast);
286 static void gdlm_put_lock(struct gfs2_glock *gl)
288 struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
289 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
292 if (gl->gl_lksb.sb_lkid == 0) {
297 clear_bit(GLF_BLOCKING, &gl->gl_flags);
298 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
299 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
300 gfs2_update_request_times(gl);
302 /* don't want to skip dlm_unlock writing the lvb when lock has one */
304 if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
305 !gl->gl_lksb.sb_lvbptr) {
310 error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_VALBLK,
313 fs_err(sdp, "gdlm_unlock %x,%llx err=%d\n",
315 (unsigned long long)gl->gl_name.ln_number, error);
320 static void gdlm_cancel(struct gfs2_glock *gl)
322 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
323 dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
327 * dlm/gfs2 recovery coordination using dlm_recover callbacks
329 * 0. gfs2 checks for another cluster node withdraw, needing journal replay
330 * 1. dlm_controld sees lockspace members change
331 * 2. dlm_controld blocks dlm-kernel locking activity
332 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
333 * 4. dlm_controld starts and finishes its own user level recovery
334 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
335 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
336 * 7. dlm_recoverd does its own lock recovery
337 * 8. dlm_recoverd unblocks dlm-kernel locking activity
338 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
339 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
340 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
341 * 12. gfs2_recover dequeues and recovers journals of failed nodes
342 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
343 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
344 * 15. gfs2_control unblocks normal locking when all journals are recovered
346 * - failures during recovery
348 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
349 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
350 * recovering for a prior failure. gfs2_control needs a way to detect
351 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
352 * the recover_block and recover_start values.
354 * recover_done() provides a new lockspace generation number each time it
355 * is called (step 9). This generation number is saved as recover_start.
356 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
357 * recover_block = recover_start. So, while recover_block is equal to
358 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
359 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
361 * - more specific gfs2 steps in sequence above
363 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
364 * 6. recover_slot records any failed jids (maybe none)
365 * 9. recover_done sets recover_start = new generation number
366 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
367 * 12. gfs2_recover does journal recoveries for failed jids identified above
368 * 14. gfs2_control clears control_lock lvb bits for recovered jids
369 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
370 * again) then do nothing, otherwise if recover_start > recover_block
371 * then clear BLOCK_LOCKS.
373 * - parallel recovery steps across all nodes
375 * All nodes attempt to update the control_lock lvb with the new generation
376 * number and jid bits, but only the first to get the control_lock EX will
377 * do so; others will see that it's already done (lvb already contains new
378 * generation number.)
380 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
381 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
382 * . One node gets control_lock first and writes the lvb, others see it's done
383 * . All nodes attempt to recover jids for which they see control_lock bits set
384 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
385 * . All nodes will eventually see all lvb bits clear and unblock locks
387 * - is there a problem with clearing an lvb bit that should be set
388 * and missing a journal recovery?
391 * 2. lvb bit set for step 1
392 * 3. jid recovered for step 1
393 * 4. jid taken again (new mount)
394 * 5. jid fails (for step 4)
395 * 6. lvb bit set for step 5 (will already be set)
396 * 7. lvb bit cleared for step 3
398 * This is not a problem because the failure in step 5 does not
399 * require recovery, because the mount in step 4 could not have
400 * progressed far enough to unblock locks and access the fs. The
401 * control_mount() function waits for all recoveries to be complete
402 * for the latest lockspace generation before ever unblocking locks
403 * and returning. The mount in step 4 waits until the recovery in
406 * - special case of first mounter: first node to mount the fs
408 * The first node to mount a gfs2 fs needs to check all the journals
409 * and recover any that need recovery before other nodes are allowed
410 * to mount the fs. (Others may begin mounting, but they must wait
411 * for the first mounter to be done before taking locks on the fs
412 * or accessing the fs.) This has two parts:
414 * 1. The mounted_lock tells a node it's the first to mount the fs.
415 * Each node holds the mounted_lock in PR while it's mounted.
416 * Each node tries to acquire the mounted_lock in EX when it mounts.
417 * If a node is granted the mounted_lock EX it means there are no
418 * other mounted nodes (no PR locks exist), and it is the first mounter.
419 * The mounted_lock is demoted to PR when first recovery is done, so
420 * others will fail to get an EX lock, but will get a PR lock.
422 * 2. The control_lock blocks others in control_mount() while the first
423 * mounter is doing first mount recovery of all journals.
424 * A mounting node needs to acquire control_lock in EX mode before
425 * it can proceed. The first mounter holds control_lock in EX while doing
426 * the first mount recovery, blocking mounts from other nodes, then demotes
427 * control_lock to NL when it's done (others_may_mount/first_done),
428 * allowing other nodes to continue mounting.
431 * control_lock EX/NOQUEUE success
432 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
434 * do first mounter recovery
435 * mounted_lock EX->PR
436 * control_lock EX->NL, write lvb generation
439 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
440 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
441 * mounted_lock PR/NOQUEUE success
442 * read lvb generation
443 * control_lock EX->NL
446 * - mount during recovery
448 * If a node mounts while others are doing recovery (not first mounter),
449 * the mounting node will get its initial recover_done() callback without
450 * having seen any previous failures/callbacks.
452 * It must wait for all recoveries preceding its mount to be finished
453 * before it unblocks locks. It does this by repeating the "other mounter"
454 * steps above until the lvb generation number is >= its mount generation
455 * number (from initial recover_done) and all lvb bits are clear.
457 * - control_lock lvb format
459 * 4 bytes generation number: the latest dlm lockspace generation number
460 * from recover_done callback. Indicates the jid bitmap has been updated
461 * to reflect all slot failures through that generation.
463 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
464 * that jid N needs recovery.
467 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
469 static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
473 memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE);
474 memcpy(&gen, lvb_bits, sizeof(__le32));
475 *lvb_gen = le32_to_cpu(gen);
478 static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
482 memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE);
483 gen = cpu_to_le32(lvb_gen);
484 memcpy(ls->ls_control_lvb, &gen, sizeof(__le32));
487 static int all_jid_bits_clear(char *lvb)
489 return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0,
490 GDLM_LVB_SIZE - JID_BITMAP_OFFSET);
493 static void sync_wait_cb(void *arg)
495 struct lm_lockstruct *ls = arg;
496 complete(&ls->ls_sync_wait);
499 static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
501 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
504 error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
506 fs_err(sdp, "%s lkid %x error %d\n",
507 name, lksb->sb_lkid, error);
511 wait_for_completion(&ls->ls_sync_wait);
513 if (lksb->sb_status != -DLM_EUNLOCK) {
514 fs_err(sdp, "%s lkid %x status %d\n",
515 name, lksb->sb_lkid, lksb->sb_status);
521 static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
522 unsigned int num, struct dlm_lksb *lksb, char *name)
524 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
525 char strname[GDLM_STRNAME_BYTES];
528 memset(strname, 0, GDLM_STRNAME_BYTES);
529 snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
531 error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
532 strname, GDLM_STRNAME_BYTES - 1,
533 0, sync_wait_cb, ls, NULL);
535 fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
536 name, lksb->sb_lkid, flags, mode, error);
540 wait_for_completion(&ls->ls_sync_wait);
542 status = lksb->sb_status;
544 if (status && status != -EAGAIN) {
545 fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
546 name, lksb->sb_lkid, flags, mode, status);
552 static int mounted_unlock(struct gfs2_sbd *sdp)
554 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
555 return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
558 static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
560 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
561 return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
562 &ls->ls_mounted_lksb, "mounted_lock");
565 static int control_unlock(struct gfs2_sbd *sdp)
567 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
568 return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
571 static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
573 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
574 return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK,
575 &ls->ls_control_lksb, "control_lock");
579 * remote_withdraw - react to a node withdrawing from the file system
580 * @sdp: The superblock
582 static void remote_withdraw(struct gfs2_sbd *sdp)
584 struct gfs2_jdesc *jd;
585 int ret = 0, count = 0;
587 list_for_each_entry(jd, &sdp->sd_jindex_list, jd_list) {
588 if (jd->jd_jid == sdp->sd_lockstruct.ls_jid)
590 ret = gfs2_recover_journal(jd, true);
596 /* Now drop the additional reference we acquired */
597 fs_err(sdp, "Journals checked: %d, ret = %d.\n", count, ret);
600 static void gfs2_control_func(struct work_struct *work)
602 struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
603 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
604 uint32_t block_gen, start_gen, lvb_gen, flags;
610 /* First check for other nodes that may have done a withdraw. */
611 if (test_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags)) {
612 remote_withdraw(sdp);
613 clear_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags);
617 spin_lock(&ls->ls_recover_spin);
619 * No MOUNT_DONE means we're still mounting; control_mount()
620 * will set this flag, after which this thread will take over
621 * all further clearing of BLOCK_LOCKS.
623 * FIRST_MOUNT means this node is doing first mounter recovery,
624 * for which recovery control is handled by
625 * control_mount()/control_first_done(), not this thread.
627 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
628 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
629 spin_unlock(&ls->ls_recover_spin);
632 block_gen = ls->ls_recover_block;
633 start_gen = ls->ls_recover_start;
634 spin_unlock(&ls->ls_recover_spin);
637 * Equal block_gen and start_gen implies we are between
638 * recover_prep and recover_done callbacks, which means
639 * dlm recovery is in progress and dlm locking is blocked.
640 * There's no point trying to do any work until recover_done.
643 if (block_gen == start_gen)
647 * Propagate recover_submit[] and recover_result[] to lvb:
648 * dlm_recoverd adds to recover_submit[] jids needing recovery
649 * gfs2_recover adds to recover_result[] journal recovery results
651 * set lvb bit for jids in recover_submit[] if the lvb has not
652 * yet been updated for the generation of the failure
654 * clear lvb bit for jids in recover_result[] if the result of
655 * the journal recovery is SUCCESS
658 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
660 fs_err(sdp, "control lock EX error %d\n", error);
664 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
666 spin_lock(&ls->ls_recover_spin);
667 if (block_gen != ls->ls_recover_block ||
668 start_gen != ls->ls_recover_start) {
669 fs_info(sdp, "recover generation %u block1 %u %u\n",
670 start_gen, block_gen, ls->ls_recover_block);
671 spin_unlock(&ls->ls_recover_spin);
672 control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
676 recover_size = ls->ls_recover_size;
678 if (lvb_gen <= start_gen) {
680 * Clear lvb bits for jids we've successfully recovered.
681 * Because all nodes attempt to recover failed journals,
682 * a journal can be recovered multiple times successfully
683 * in succession. Only the first will really do recovery,
684 * the others find it clean, but still report a successful
685 * recovery. So, another node may have already recovered
686 * the jid and cleared the lvb bit for it.
688 for (i = 0; i < recover_size; i++) {
689 if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
692 ls->ls_recover_result[i] = 0;
694 if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
697 __clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
702 if (lvb_gen == start_gen) {
704 * Failed slots before start_gen are already set in lvb.
706 for (i = 0; i < recover_size; i++) {
707 if (!ls->ls_recover_submit[i])
709 if (ls->ls_recover_submit[i] < lvb_gen)
710 ls->ls_recover_submit[i] = 0;
712 } else if (lvb_gen < start_gen) {
714 * Failed slots before start_gen are not yet set in lvb.
716 for (i = 0; i < recover_size; i++) {
717 if (!ls->ls_recover_submit[i])
719 if (ls->ls_recover_submit[i] < start_gen) {
720 ls->ls_recover_submit[i] = 0;
721 __set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
724 /* even if there are no bits to set, we need to write the
725 latest generation to the lvb */
729 * we should be getting a recover_done() for lvb_gen soon
732 spin_unlock(&ls->ls_recover_spin);
735 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
736 flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
738 flags = DLM_LKF_CONVERT;
741 error = control_lock(sdp, DLM_LOCK_NL, flags);
743 fs_err(sdp, "control lock NL error %d\n", error);
748 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
749 * and clear a jid bit in the lvb if the recovery is a success.
750 * Eventually all journals will be recovered, all jid bits will
751 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
754 for (i = 0; i < recover_size; i++) {
755 if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
756 fs_info(sdp, "recover generation %u jid %d\n",
758 gfs2_recover_set(sdp, i);
766 * No more jid bits set in lvb, all recovery is done, unblock locks
767 * (unless a new recover_prep callback has occured blocking locks
768 * again while working above)
771 spin_lock(&ls->ls_recover_spin);
772 if (ls->ls_recover_block == block_gen &&
773 ls->ls_recover_start == start_gen) {
774 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
775 spin_unlock(&ls->ls_recover_spin);
776 fs_info(sdp, "recover generation %u done\n", start_gen);
777 gfs2_glock_thaw(sdp);
779 fs_info(sdp, "recover generation %u block2 %u %u\n",
780 start_gen, block_gen, ls->ls_recover_block);
781 spin_unlock(&ls->ls_recover_spin);
785 static int control_mount(struct gfs2_sbd *sdp)
787 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
788 uint32_t start_gen, block_gen, mount_gen, lvb_gen;
793 memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
794 memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
795 memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
796 ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
797 init_completion(&ls->ls_sync_wait);
799 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
801 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
803 fs_err(sdp, "control_mount control_lock NL error %d\n", error);
807 error = mounted_lock(sdp, DLM_LOCK_NL, 0);
809 fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
813 mounted_mode = DLM_LOCK_NL;
816 if (retries++ && signal_pending(current)) {
822 * We always start with both locks in NL. control_lock is
823 * demoted to NL below so we don't need to do it here.
826 if (mounted_mode != DLM_LOCK_NL) {
827 error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
830 mounted_mode = DLM_LOCK_NL;
834 * Other nodes need to do some work in dlm recovery and gfs2_control
835 * before the recover_done and control_lock will be ready for us below.
836 * A delay here is not required but often avoids having to retry.
839 msleep_interruptible(500);
842 * Acquire control_lock in EX and mounted_lock in either EX or PR.
843 * control_lock lvb keeps track of any pending journal recoveries.
844 * mounted_lock indicates if any other nodes have the fs mounted.
847 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
848 if (error == -EAGAIN) {
851 fs_err(sdp, "control_mount control_lock EX error %d\n", error);
856 * If we're a spectator, we don't want to take the lock in EX because
857 * we cannot do the first-mount responsibility it implies: recovery.
859 if (sdp->sd_args.ar_spectator)
862 error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
864 mounted_mode = DLM_LOCK_EX;
866 } else if (error != -EAGAIN) {
867 fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
871 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
873 mounted_mode = DLM_LOCK_PR;
876 /* not even -EAGAIN should happen here */
877 fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
883 * If we got both locks above in EX, then we're the first mounter.
884 * If not, then we need to wait for the control_lock lvb to be
885 * updated by other mounted nodes to reflect our mount generation.
887 * In simple first mounter cases, first mounter will see zero lvb_gen,
888 * but in cases where all existing nodes leave/fail before mounting
889 * nodes finish control_mount, then all nodes will be mounting and
890 * lvb_gen will be non-zero.
893 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
895 if (lvb_gen == 0xFFFFFFFF) {
896 /* special value to force mount attempts to fail */
897 fs_err(sdp, "control_mount control_lock disabled\n");
902 if (mounted_mode == DLM_LOCK_EX) {
903 /* first mounter, keep both EX while doing first recovery */
904 spin_lock(&ls->ls_recover_spin);
905 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
906 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
907 set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
908 spin_unlock(&ls->ls_recover_spin);
909 fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
913 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
918 * We are not first mounter, now we need to wait for the control_lock
919 * lvb generation to be >= the generation from our first recover_done
920 * and all lvb bits to be clear (no pending journal recoveries.)
923 if (!all_jid_bits_clear(ls->ls_lvb_bits)) {
924 /* journals need recovery, wait until all are clear */
925 fs_info(sdp, "control_mount wait for journal recovery\n");
929 spin_lock(&ls->ls_recover_spin);
930 block_gen = ls->ls_recover_block;
931 start_gen = ls->ls_recover_start;
932 mount_gen = ls->ls_recover_mount;
934 if (lvb_gen < mount_gen) {
935 /* wait for mounted nodes to update control_lock lvb to our
936 generation, which might include new recovery bits set */
937 if (sdp->sd_args.ar_spectator) {
938 fs_info(sdp, "Recovery is required. Waiting for a "
939 "non-spectator to mount.\n");
940 msleep_interruptible(1000);
942 fs_info(sdp, "control_mount wait1 block %u start %u "
943 "mount %u lvb %u flags %lx\n", block_gen,
944 start_gen, mount_gen, lvb_gen,
945 ls->ls_recover_flags);
947 spin_unlock(&ls->ls_recover_spin);
951 if (lvb_gen != start_gen) {
952 /* wait for mounted nodes to update control_lock lvb to the
953 latest recovery generation */
954 fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
955 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
956 lvb_gen, ls->ls_recover_flags);
957 spin_unlock(&ls->ls_recover_spin);
961 if (block_gen == start_gen) {
962 /* dlm recovery in progress, wait for it to finish */
963 fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
964 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
965 lvb_gen, ls->ls_recover_flags);
966 spin_unlock(&ls->ls_recover_spin);
970 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
971 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
972 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
973 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
974 spin_unlock(&ls->ls_recover_spin);
983 static int control_first_done(struct gfs2_sbd *sdp)
985 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
986 uint32_t start_gen, block_gen;
990 spin_lock(&ls->ls_recover_spin);
991 start_gen = ls->ls_recover_start;
992 block_gen = ls->ls_recover_block;
994 if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
995 !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
996 !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
997 /* sanity check, should not happen */
998 fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
999 start_gen, block_gen, ls->ls_recover_flags);
1000 spin_unlock(&ls->ls_recover_spin);
1001 control_unlock(sdp);
1005 if (start_gen == block_gen) {
1007 * Wait for the end of a dlm recovery cycle to switch from
1008 * first mounter recovery. We can ignore any recover_slot
1009 * callbacks between the recover_prep and next recover_done
1010 * because we are still the first mounter and any failed nodes
1011 * have not fully mounted, so they don't need recovery.
1013 spin_unlock(&ls->ls_recover_spin);
1014 fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
1016 wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
1017 TASK_UNINTERRUPTIBLE);
1021 clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1022 set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
1023 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
1024 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
1025 spin_unlock(&ls->ls_recover_spin);
1027 memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE);
1028 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
1030 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
1032 fs_err(sdp, "control_first_done mounted PR error %d\n", error);
1034 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
1036 fs_err(sdp, "control_first_done control NL error %d\n", error);
1042 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1043 * to accomodate the largest slot number. (NB dlm slot numbers start at 1,
1044 * gfs2 jids start at 0, so jid = slot - 1)
1047 #define RECOVER_SIZE_INC 16
1049 static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1052 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1053 uint32_t *submit = NULL;
1054 uint32_t *result = NULL;
1055 uint32_t old_size, new_size;
1058 if (!ls->ls_lvb_bits) {
1059 ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1060 if (!ls->ls_lvb_bits)
1065 for (i = 0; i < num_slots; i++) {
1066 if (max_jid < slots[i].slot - 1)
1067 max_jid = slots[i].slot - 1;
1070 old_size = ls->ls_recover_size;
1071 new_size = old_size;
1072 while (new_size < max_jid + 1)
1073 new_size += RECOVER_SIZE_INC;
1074 if (new_size == old_size)
1077 submit = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1078 result = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1079 if (!submit || !result) {
1085 spin_lock(&ls->ls_recover_spin);
1086 memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1087 memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1088 kfree(ls->ls_recover_submit);
1089 kfree(ls->ls_recover_result);
1090 ls->ls_recover_submit = submit;
1091 ls->ls_recover_result = result;
1092 ls->ls_recover_size = new_size;
1093 spin_unlock(&ls->ls_recover_spin);
1097 static void free_recover_size(struct lm_lockstruct *ls)
1099 kfree(ls->ls_lvb_bits);
1100 kfree(ls->ls_recover_submit);
1101 kfree(ls->ls_recover_result);
1102 ls->ls_recover_submit = NULL;
1103 ls->ls_recover_result = NULL;
1104 ls->ls_recover_size = 0;
1105 ls->ls_lvb_bits = NULL;
1108 /* dlm calls before it does lock recovery */
1110 static void gdlm_recover_prep(void *arg)
1112 struct gfs2_sbd *sdp = arg;
1113 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1115 if (gfs2_withdrawn(sdp)) {
1116 fs_err(sdp, "recover_prep ignored due to withdraw.\n");
1119 spin_lock(&ls->ls_recover_spin);
1120 ls->ls_recover_block = ls->ls_recover_start;
1121 set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1123 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1124 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1125 spin_unlock(&ls->ls_recover_spin);
1128 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1129 spin_unlock(&ls->ls_recover_spin);
1132 /* dlm calls after recover_prep has been completed on all lockspace members;
1133 identifies slot/jid of failed member */
1135 static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1137 struct gfs2_sbd *sdp = arg;
1138 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1139 int jid = slot->slot - 1;
1141 if (gfs2_withdrawn(sdp)) {
1142 fs_err(sdp, "recover_slot jid %d ignored due to withdraw.\n",
1146 spin_lock(&ls->ls_recover_spin);
1147 if (ls->ls_recover_size < jid + 1) {
1148 fs_err(sdp, "recover_slot jid %d gen %u short size %d\n",
1149 jid, ls->ls_recover_block, ls->ls_recover_size);
1150 spin_unlock(&ls->ls_recover_spin);
1154 if (ls->ls_recover_submit[jid]) {
1155 fs_info(sdp, "recover_slot jid %d gen %u prev %u\n",
1156 jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1158 ls->ls_recover_submit[jid] = ls->ls_recover_block;
1159 spin_unlock(&ls->ls_recover_spin);
1162 /* dlm calls after recover_slot and after it completes lock recovery */
1164 static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1165 int our_slot, uint32_t generation)
1167 struct gfs2_sbd *sdp = arg;
1168 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1170 if (gfs2_withdrawn(sdp)) {
1171 fs_err(sdp, "recover_done ignored due to withdraw.\n");
1174 /* ensure the ls jid arrays are large enough */
1175 set_recover_size(sdp, slots, num_slots);
1177 spin_lock(&ls->ls_recover_spin);
1178 ls->ls_recover_start = generation;
1180 if (!ls->ls_recover_mount) {
1181 ls->ls_recover_mount = generation;
1182 ls->ls_jid = our_slot - 1;
1185 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1186 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1188 clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1189 smp_mb__after_atomic();
1190 wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1191 spin_unlock(&ls->ls_recover_spin);
1194 /* gfs2_recover thread has a journal recovery result */
1196 static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1197 unsigned int result)
1199 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1201 if (gfs2_withdrawn(sdp)) {
1202 fs_err(sdp, "recovery_result jid %d ignored due to withdraw.\n",
1206 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1209 /* don't care about the recovery of own journal during mount */
1210 if (jid == ls->ls_jid)
1213 spin_lock(&ls->ls_recover_spin);
1214 if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1215 spin_unlock(&ls->ls_recover_spin);
1218 if (ls->ls_recover_size < jid + 1) {
1219 fs_err(sdp, "recovery_result jid %d short size %d\n",
1220 jid, ls->ls_recover_size);
1221 spin_unlock(&ls->ls_recover_spin);
1225 fs_info(sdp, "recover jid %d result %s\n", jid,
1226 result == LM_RD_GAVEUP ? "busy" : "success");
1228 ls->ls_recover_result[jid] = result;
1230 /* GAVEUP means another node is recovering the journal; delay our
1231 next attempt to recover it, to give the other node a chance to
1232 finish before trying again */
1234 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1235 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1236 result == LM_RD_GAVEUP ? HZ : 0);
1237 spin_unlock(&ls->ls_recover_spin);
1240 static const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1241 .recover_prep = gdlm_recover_prep,
1242 .recover_slot = gdlm_recover_slot,
1243 .recover_done = gdlm_recover_done,
1246 static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1248 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1249 char cluster[GFS2_LOCKNAME_LEN];
1252 int error, ops_result;
1255 * initialize everything
1258 INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1259 spin_lock_init(&ls->ls_recover_spin);
1260 ls->ls_recover_flags = 0;
1261 ls->ls_recover_mount = 0;
1262 ls->ls_recover_start = 0;
1263 ls->ls_recover_block = 0;
1264 ls->ls_recover_size = 0;
1265 ls->ls_recover_submit = NULL;
1266 ls->ls_recover_result = NULL;
1267 ls->ls_lvb_bits = NULL;
1269 error = set_recover_size(sdp, NULL, 0);
1274 * prepare dlm_new_lockspace args
1277 fsname = strchr(table, ':');
1279 fs_info(sdp, "no fsname found\n");
1283 memset(cluster, 0, sizeof(cluster));
1284 memcpy(cluster, table, strlen(table) - strlen(fsname));
1287 flags = DLM_LSFL_FS | DLM_LSFL_NEWEXCL;
1290 * create/join lockspace
1293 error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1294 &gdlm_lockspace_ops, sdp, &ops_result,
1297 fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1301 if (ops_result < 0) {
1303 * dlm does not support ops callbacks,
1304 * old dlm_controld/gfs_controld are used, try without ops.
1306 fs_info(sdp, "dlm lockspace ops not used\n");
1307 free_recover_size(ls);
1308 set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1312 if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1313 fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1319 * control_mount() uses control_lock to determine first mounter,
1320 * and for later mounts, waits for any recoveries to be cleared.
1323 error = control_mount(sdp);
1325 fs_err(sdp, "mount control error %d\n", error);
1329 ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1330 clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
1331 smp_mb__after_atomic();
1332 wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1336 dlm_release_lockspace(ls->ls_dlm, 2);
1338 free_recover_size(ls);
1343 static void gdlm_first_done(struct gfs2_sbd *sdp)
1345 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1348 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1351 error = control_first_done(sdp);
1353 fs_err(sdp, "mount first_done error %d\n", error);
1356 static void gdlm_unmount(struct gfs2_sbd *sdp)
1358 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1360 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1363 /* wait for gfs2_control_wq to be done with this mount */
1365 spin_lock(&ls->ls_recover_spin);
1366 set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1367 spin_unlock(&ls->ls_recover_spin);
1368 flush_delayed_work(&sdp->sd_control_work);
1370 /* mounted_lock and control_lock will be purged in dlm recovery */
1373 dlm_release_lockspace(ls->ls_dlm, 2);
1377 free_recover_size(ls);
1380 static const match_table_t dlm_tokens = {
1381 { Opt_jid, "jid=%d"},
1383 { Opt_first, "first=%d"},
1384 { Opt_nodir, "nodir=%d"},
1388 const struct lm_lockops gfs2_dlm_ops = {
1389 .lm_proto_name = "lock_dlm",
1390 .lm_mount = gdlm_mount,
1391 .lm_first_done = gdlm_first_done,
1392 .lm_recovery_result = gdlm_recovery_result,
1393 .lm_unmount = gdlm_unmount,
1394 .lm_put_lock = gdlm_put_lock,
1395 .lm_lock = gdlm_lock,
1396 .lm_cancel = gdlm_cancel,
1397 .lm_tokens = &dlm_tokens,