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 * @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 variance estimate is a bit
33 * more complicated. We subtract the current variance estimate from
34 * the abs value of the @delta and add 1/4 of that to the running
35 * total. That's equivalent to 3/4 of the current variance
36 * estimate plus 1/4 of the abs of @delta.
38 * Note that the index points at the array entry containing the smoothed
39 * mean value, and the variance is always in the following entry
41 * Reference: TCP/IP Illustrated, vol 2, p. 831,832
42 * All times are in units of integer nanoseconds. Unlike the TCP/IP case,
43 * they are not scaled fixed point.
46 static inline void gfs2_update_stats(struct gfs2_lkstats *s, unsigned index,
49 s64 delta = sample - s->stats[index];
50 s->stats[index] += (delta >> 3);
52 s->stats[index] += (s64)(abs(delta) - s->stats[index]) >> 2;
56 * gfs2_update_reply_times - Update locking statistics
57 * @gl: The glock to update
59 * This assumes that gl->gl_dstamp has been set earlier.
61 * The rtt (lock round trip time) is an estimate of the time
62 * taken to perform a dlm lock request. We update it on each
65 * The blocking flag is set on the glock for all dlm requests
66 * which may potentially block due to lock requests from other nodes.
67 * DLM requests where the current lock state is exclusive, the
68 * requested state is null (or unlocked) or where the TRY or
69 * TRY_1CB flags are set are classified as non-blocking. All
70 * other DLM requests are counted as (potentially) blocking.
72 static inline void gfs2_update_reply_times(struct gfs2_glock *gl)
74 struct gfs2_pcpu_lkstats *lks;
75 const unsigned gltype = gl->gl_name.ln_type;
76 unsigned index = test_bit(GLF_BLOCKING, &gl->gl_flags) ?
77 GFS2_LKS_SRTTB : GFS2_LKS_SRTT;
81 rtt = ktime_to_ns(ktime_sub(ktime_get_real(), gl->gl_dstamp));
82 lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
83 gfs2_update_stats(&gl->gl_stats, index, rtt); /* Local */
84 gfs2_update_stats(&lks->lkstats[gltype], index, rtt); /* Global */
87 trace_gfs2_glock_lock_time(gl, rtt);
91 * gfs2_update_request_times - Update locking statistics
92 * @gl: The glock to update
94 * The irt (lock inter-request times) measures the average time
95 * between requests to the dlm. It is updated immediately before
99 static inline void gfs2_update_request_times(struct gfs2_glock *gl)
101 struct gfs2_pcpu_lkstats *lks;
102 const unsigned gltype = gl->gl_name.ln_type;
107 dstamp = gl->gl_dstamp;
108 gl->gl_dstamp = ktime_get_real();
109 irt = ktime_to_ns(ktime_sub(gl->gl_dstamp, dstamp));
110 lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
111 gfs2_update_stats(&gl->gl_stats, GFS2_LKS_SIRT, irt); /* Local */
112 gfs2_update_stats(&lks->lkstats[gltype], GFS2_LKS_SIRT, irt); /* Global */
116 static void gdlm_ast(void *arg)
118 struct gfs2_glock *gl = arg;
119 unsigned ret = gl->gl_state;
121 gfs2_update_reply_times(gl);
122 BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED);
124 if ((gl->gl_lksb.sb_flags & DLM_SBF_VALNOTVALID) && gl->gl_lksb.sb_lvbptr)
125 memset(gl->gl_lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
127 switch (gl->gl_lksb.sb_status) {
128 case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */
129 if (gl->gl_ops->go_free)
130 gl->gl_ops->go_free(gl);
133 case -DLM_ECANCEL: /* Cancel while getting lock */
134 ret |= LM_OUT_CANCELED;
136 case -EAGAIN: /* Try lock fails */
137 case -EDEADLK: /* Deadlock detected */
139 case -ETIMEDOUT: /* Canceled due to timeout */
142 case 0: /* Success */
144 default: /* Something unexpected */
149 if (gl->gl_lksb.sb_flags & DLM_SBF_ALTMODE) {
150 if (gl->gl_req == LM_ST_SHARED)
151 ret = LM_ST_DEFERRED;
152 else if (gl->gl_req == LM_ST_DEFERRED)
158 set_bit(GLF_INITIAL, &gl->gl_flags);
159 gfs2_glock_complete(gl, ret);
162 if (!test_bit(GLF_INITIAL, &gl->gl_flags))
163 gl->gl_lksb.sb_lkid = 0;
164 gfs2_glock_complete(gl, ret);
167 static void gdlm_bast(void *arg, int mode)
169 struct gfs2_glock *gl = arg;
173 gfs2_glock_cb(gl, LM_ST_UNLOCKED);
176 gfs2_glock_cb(gl, LM_ST_DEFERRED);
179 gfs2_glock_cb(gl, LM_ST_SHARED);
182 fs_err(gl->gl_name.ln_sbd, "unknown bast mode %d\n", mode);
187 /* convert gfs lock-state to dlm lock-mode */
189 static int make_mode(struct gfs2_sbd *sdp, const unsigned int lmstate)
194 case LM_ST_EXCLUSIVE:
201 fs_err(sdp, "unknown LM state %d\n", lmstate);
206 static u32 make_flags(struct gfs2_glock *gl, const unsigned int gfs_flags,
211 if (gl->gl_lksb.sb_lvbptr)
212 lkf |= DLM_LKF_VALBLK;
214 if (gfs_flags & LM_FLAG_TRY)
215 lkf |= DLM_LKF_NOQUEUE;
217 if (gfs_flags & LM_FLAG_TRY_1CB) {
218 lkf |= DLM_LKF_NOQUEUE;
219 lkf |= DLM_LKF_NOQUEUEBAST;
222 if (gfs_flags & LM_FLAG_PRIORITY) {
223 lkf |= DLM_LKF_NOORDER;
224 lkf |= DLM_LKF_HEADQUE;
227 if (gfs_flags & LM_FLAG_ANY) {
228 if (req == DLM_LOCK_PR)
229 lkf |= DLM_LKF_ALTCW;
230 else if (req == DLM_LOCK_CW)
231 lkf |= DLM_LKF_ALTPR;
236 if (gl->gl_lksb.sb_lkid != 0) {
237 lkf |= DLM_LKF_CONVERT;
238 if (test_bit(GLF_BLOCKING, &gl->gl_flags))
239 lkf |= DLM_LKF_QUECVT;
245 static void gfs2_reverse_hex(char *c, u64 value)
249 *c-- = hex_asc[value & 0x0f];
254 static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state,
257 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
260 char strname[GDLM_STRNAME_BYTES] = "";
262 req = make_mode(gl->gl_name.ln_sbd, req_state);
263 lkf = make_flags(gl, flags, req);
264 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
265 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
266 if (gl->gl_lksb.sb_lkid) {
267 gfs2_update_request_times(gl);
269 memset(strname, ' ', GDLM_STRNAME_BYTES - 1);
270 strname[GDLM_STRNAME_BYTES - 1] = '\0';
271 gfs2_reverse_hex(strname + 7, gl->gl_name.ln_type);
272 gfs2_reverse_hex(strname + 23, gl->gl_name.ln_number);
273 gl->gl_dstamp = ktime_get_real();
276 * Submit the actual lock request.
279 return dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname,
280 GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast);
283 static void gdlm_put_lock(struct gfs2_glock *gl)
285 struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
286 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
289 if (gl->gl_lksb.sb_lkid == 0) {
294 clear_bit(GLF_BLOCKING, &gl->gl_flags);
295 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
296 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
297 gfs2_update_request_times(gl);
299 /* don't want to call dlm if we've unmounted the lock protocol */
300 if (test_bit(DFL_UNMOUNT, &ls->ls_recover_flags)) {
304 /* don't want to skip dlm_unlock writing the lvb when lock has one */
306 if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
307 !gl->gl_lksb.sb_lvbptr) {
312 error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_VALBLK,
315 fs_err(sdp, "gdlm_unlock %x,%llx err=%d\n",
317 (unsigned long long)gl->gl_name.ln_number, error);
322 static void gdlm_cancel(struct gfs2_glock *gl)
324 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
325 dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
329 * dlm/gfs2 recovery coordination using dlm_recover callbacks
331 * 0. gfs2 checks for another cluster node withdraw, needing journal replay
332 * 1. dlm_controld sees lockspace members change
333 * 2. dlm_controld blocks dlm-kernel locking activity
334 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
335 * 4. dlm_controld starts and finishes its own user level recovery
336 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
337 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
338 * 7. dlm_recoverd does its own lock recovery
339 * 8. dlm_recoverd unblocks dlm-kernel locking activity
340 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
341 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
342 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
343 * 12. gfs2_recover dequeues and recovers journals of failed nodes
344 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
345 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
346 * 15. gfs2_control unblocks normal locking when all journals are recovered
348 * - failures during recovery
350 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
351 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
352 * recovering for a prior failure. gfs2_control needs a way to detect
353 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
354 * the recover_block and recover_start values.
356 * recover_done() provides a new lockspace generation number each time it
357 * is called (step 9). This generation number is saved as recover_start.
358 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
359 * recover_block = recover_start. So, while recover_block is equal to
360 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
361 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
363 * - more specific gfs2 steps in sequence above
365 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
366 * 6. recover_slot records any failed jids (maybe none)
367 * 9. recover_done sets recover_start = new generation number
368 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
369 * 12. gfs2_recover does journal recoveries for failed jids identified above
370 * 14. gfs2_control clears control_lock lvb bits for recovered jids
371 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
372 * again) then do nothing, otherwise if recover_start > recover_block
373 * then clear BLOCK_LOCKS.
375 * - parallel recovery steps across all nodes
377 * All nodes attempt to update the control_lock lvb with the new generation
378 * number and jid bits, but only the first to get the control_lock EX will
379 * do so; others will see that it's already done (lvb already contains new
380 * generation number.)
382 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
383 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
384 * . One node gets control_lock first and writes the lvb, others see it's done
385 * . All nodes attempt to recover jids for which they see control_lock bits set
386 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
387 * . All nodes will eventually see all lvb bits clear and unblock locks
389 * - is there a problem with clearing an lvb bit that should be set
390 * and missing a journal recovery?
393 * 2. lvb bit set for step 1
394 * 3. jid recovered for step 1
395 * 4. jid taken again (new mount)
396 * 5. jid fails (for step 4)
397 * 6. lvb bit set for step 5 (will already be set)
398 * 7. lvb bit cleared for step 3
400 * This is not a problem because the failure in step 5 does not
401 * require recovery, because the mount in step 4 could not have
402 * progressed far enough to unblock locks and access the fs. The
403 * control_mount() function waits for all recoveries to be complete
404 * for the latest lockspace generation before ever unblocking locks
405 * and returning. The mount in step 4 waits until the recovery in
408 * - special case of first mounter: first node to mount the fs
410 * The first node to mount a gfs2 fs needs to check all the journals
411 * and recover any that need recovery before other nodes are allowed
412 * to mount the fs. (Others may begin mounting, but they must wait
413 * for the first mounter to be done before taking locks on the fs
414 * or accessing the fs.) This has two parts:
416 * 1. The mounted_lock tells a node it's the first to mount the fs.
417 * Each node holds the mounted_lock in PR while it's mounted.
418 * Each node tries to acquire the mounted_lock in EX when it mounts.
419 * If a node is granted the mounted_lock EX it means there are no
420 * other mounted nodes (no PR locks exist), and it is the first mounter.
421 * The mounted_lock is demoted to PR when first recovery is done, so
422 * others will fail to get an EX lock, but will get a PR lock.
424 * 2. The control_lock blocks others in control_mount() while the first
425 * mounter is doing first mount recovery of all journals.
426 * A mounting node needs to acquire control_lock in EX mode before
427 * it can proceed. The first mounter holds control_lock in EX while doing
428 * the first mount recovery, blocking mounts from other nodes, then demotes
429 * control_lock to NL when it's done (others_may_mount/first_done),
430 * allowing other nodes to continue mounting.
433 * control_lock EX/NOQUEUE success
434 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
436 * do first mounter recovery
437 * mounted_lock EX->PR
438 * control_lock EX->NL, write lvb generation
441 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
442 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
443 * mounted_lock PR/NOQUEUE success
444 * read lvb generation
445 * control_lock EX->NL
448 * - mount during recovery
450 * If a node mounts while others are doing recovery (not first mounter),
451 * the mounting node will get its initial recover_done() callback without
452 * having seen any previous failures/callbacks.
454 * It must wait for all recoveries preceding its mount to be finished
455 * before it unblocks locks. It does this by repeating the "other mounter"
456 * steps above until the lvb generation number is >= its mount generation
457 * number (from initial recover_done) and all lvb bits are clear.
459 * - control_lock lvb format
461 * 4 bytes generation number: the latest dlm lockspace generation number
462 * from recover_done callback. Indicates the jid bitmap has been updated
463 * to reflect all slot failures through that generation.
465 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
466 * that jid N needs recovery.
469 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
471 static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
475 memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE);
476 memcpy(&gen, lvb_bits, sizeof(__le32));
477 *lvb_gen = le32_to_cpu(gen);
480 static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
484 memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE);
485 gen = cpu_to_le32(lvb_gen);
486 memcpy(ls->ls_control_lvb, &gen, sizeof(__le32));
489 static int all_jid_bits_clear(char *lvb)
491 return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0,
492 GDLM_LVB_SIZE - JID_BITMAP_OFFSET);
495 static void sync_wait_cb(void *arg)
497 struct lm_lockstruct *ls = arg;
498 complete(&ls->ls_sync_wait);
501 static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
503 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
506 error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
508 fs_err(sdp, "%s lkid %x error %d\n",
509 name, lksb->sb_lkid, error);
513 wait_for_completion(&ls->ls_sync_wait);
515 if (lksb->sb_status != -DLM_EUNLOCK) {
516 fs_err(sdp, "%s lkid %x status %d\n",
517 name, lksb->sb_lkid, lksb->sb_status);
523 static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
524 unsigned int num, struct dlm_lksb *lksb, char *name)
526 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
527 char strname[GDLM_STRNAME_BYTES];
530 memset(strname, 0, GDLM_STRNAME_BYTES);
531 snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
533 error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
534 strname, GDLM_STRNAME_BYTES - 1,
535 0, sync_wait_cb, ls, NULL);
537 fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
538 name, lksb->sb_lkid, flags, mode, error);
542 wait_for_completion(&ls->ls_sync_wait);
544 status = lksb->sb_status;
546 if (status && status != -EAGAIN) {
547 fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
548 name, lksb->sb_lkid, flags, mode, status);
554 static int mounted_unlock(struct gfs2_sbd *sdp)
556 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
557 return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
560 static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
562 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
563 return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
564 &ls->ls_mounted_lksb, "mounted_lock");
567 static int control_unlock(struct gfs2_sbd *sdp)
569 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
570 return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
573 static int control_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_CONTROL_LOCK,
577 &ls->ls_control_lksb, "control_lock");
581 * remote_withdraw - react to a node withdrawing from the file system
582 * @sdp: The superblock
584 static void remote_withdraw(struct gfs2_sbd *sdp)
586 struct gfs2_jdesc *jd;
587 int ret = 0, count = 0;
589 list_for_each_entry(jd, &sdp->sd_jindex_list, jd_list) {
590 if (jd->jd_jid == sdp->sd_lockstruct.ls_jid)
592 ret = gfs2_recover_journal(jd, true);
598 /* Now drop the additional reference we acquired */
599 fs_err(sdp, "Journals checked: %d, ret = %d.\n", count, ret);
602 static void gfs2_control_func(struct work_struct *work)
604 struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
605 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
606 uint32_t block_gen, start_gen, lvb_gen, flags;
612 /* First check for other nodes that may have done a withdraw. */
613 if (test_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags)) {
614 remote_withdraw(sdp);
615 clear_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags);
619 spin_lock(&ls->ls_recover_spin);
621 * No MOUNT_DONE means we're still mounting; control_mount()
622 * will set this flag, after which this thread will take over
623 * all further clearing of BLOCK_LOCKS.
625 * FIRST_MOUNT means this node is doing first mounter recovery,
626 * for which recovery control is handled by
627 * control_mount()/control_first_done(), not this thread.
629 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
630 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
631 spin_unlock(&ls->ls_recover_spin);
634 block_gen = ls->ls_recover_block;
635 start_gen = ls->ls_recover_start;
636 spin_unlock(&ls->ls_recover_spin);
639 * Equal block_gen and start_gen implies we are between
640 * recover_prep and recover_done callbacks, which means
641 * dlm recovery is in progress and dlm locking is blocked.
642 * There's no point trying to do any work until recover_done.
645 if (block_gen == start_gen)
649 * Propagate recover_submit[] and recover_result[] to lvb:
650 * dlm_recoverd adds to recover_submit[] jids needing recovery
651 * gfs2_recover adds to recover_result[] journal recovery results
653 * set lvb bit for jids in recover_submit[] if the lvb has not
654 * yet been updated for the generation of the failure
656 * clear lvb bit for jids in recover_result[] if the result of
657 * the journal recovery is SUCCESS
660 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
662 fs_err(sdp, "control lock EX error %d\n", error);
666 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
668 spin_lock(&ls->ls_recover_spin);
669 if (block_gen != ls->ls_recover_block ||
670 start_gen != ls->ls_recover_start) {
671 fs_info(sdp, "recover generation %u block1 %u %u\n",
672 start_gen, block_gen, ls->ls_recover_block);
673 spin_unlock(&ls->ls_recover_spin);
674 control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
678 recover_size = ls->ls_recover_size;
680 if (lvb_gen <= start_gen) {
682 * Clear lvb bits for jids we've successfully recovered.
683 * Because all nodes attempt to recover failed journals,
684 * a journal can be recovered multiple times successfully
685 * in succession. Only the first will really do recovery,
686 * the others find it clean, but still report a successful
687 * recovery. So, another node may have already recovered
688 * the jid and cleared the lvb bit for it.
690 for (i = 0; i < recover_size; i++) {
691 if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
694 ls->ls_recover_result[i] = 0;
696 if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
699 __clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
704 if (lvb_gen == start_gen) {
706 * Failed slots before start_gen are already set in lvb.
708 for (i = 0; i < recover_size; i++) {
709 if (!ls->ls_recover_submit[i])
711 if (ls->ls_recover_submit[i] < lvb_gen)
712 ls->ls_recover_submit[i] = 0;
714 } else if (lvb_gen < start_gen) {
716 * Failed slots before start_gen are not yet set in lvb.
718 for (i = 0; i < recover_size; i++) {
719 if (!ls->ls_recover_submit[i])
721 if (ls->ls_recover_submit[i] < start_gen) {
722 ls->ls_recover_submit[i] = 0;
723 __set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
726 /* even if there are no bits to set, we need to write the
727 latest generation to the lvb */
731 * we should be getting a recover_done() for lvb_gen soon
734 spin_unlock(&ls->ls_recover_spin);
737 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
738 flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
740 flags = DLM_LKF_CONVERT;
743 error = control_lock(sdp, DLM_LOCK_NL, flags);
745 fs_err(sdp, "control lock NL error %d\n", error);
750 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
751 * and clear a jid bit in the lvb if the recovery is a success.
752 * Eventually all journals will be recovered, all jid bits will
753 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
756 for (i = 0; i < recover_size; i++) {
757 if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
758 fs_info(sdp, "recover generation %u jid %d\n",
760 gfs2_recover_set(sdp, i);
768 * No more jid bits set in lvb, all recovery is done, unblock locks
769 * (unless a new recover_prep callback has occured blocking locks
770 * again while working above)
773 spin_lock(&ls->ls_recover_spin);
774 if (ls->ls_recover_block == block_gen &&
775 ls->ls_recover_start == start_gen) {
776 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
777 spin_unlock(&ls->ls_recover_spin);
778 fs_info(sdp, "recover generation %u done\n", start_gen);
779 gfs2_glock_thaw(sdp);
781 fs_info(sdp, "recover generation %u block2 %u %u\n",
782 start_gen, block_gen, ls->ls_recover_block);
783 spin_unlock(&ls->ls_recover_spin);
787 static int control_mount(struct gfs2_sbd *sdp)
789 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
790 uint32_t start_gen, block_gen, mount_gen, lvb_gen;
795 memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
796 memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
797 memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
798 ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
799 init_completion(&ls->ls_sync_wait);
801 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
803 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
805 fs_err(sdp, "control_mount control_lock NL error %d\n", error);
809 error = mounted_lock(sdp, DLM_LOCK_NL, 0);
811 fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
815 mounted_mode = DLM_LOCK_NL;
818 if (retries++ && signal_pending(current)) {
824 * We always start with both locks in NL. control_lock is
825 * demoted to NL below so we don't need to do it here.
828 if (mounted_mode != DLM_LOCK_NL) {
829 error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
832 mounted_mode = DLM_LOCK_NL;
836 * Other nodes need to do some work in dlm recovery and gfs2_control
837 * before the recover_done and control_lock will be ready for us below.
838 * A delay here is not required but often avoids having to retry.
841 msleep_interruptible(500);
844 * Acquire control_lock in EX and mounted_lock in either EX or PR.
845 * control_lock lvb keeps track of any pending journal recoveries.
846 * mounted_lock indicates if any other nodes have the fs mounted.
849 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
850 if (error == -EAGAIN) {
853 fs_err(sdp, "control_mount control_lock EX error %d\n", error);
858 * If we're a spectator, we don't want to take the lock in EX because
859 * we cannot do the first-mount responsibility it implies: recovery.
861 if (sdp->sd_args.ar_spectator)
864 error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
866 mounted_mode = DLM_LOCK_EX;
868 } else if (error != -EAGAIN) {
869 fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
873 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
875 mounted_mode = DLM_LOCK_PR;
878 /* not even -EAGAIN should happen here */
879 fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
885 * If we got both locks above in EX, then we're the first mounter.
886 * If not, then we need to wait for the control_lock lvb to be
887 * updated by other mounted nodes to reflect our mount generation.
889 * In simple first mounter cases, first mounter will see zero lvb_gen,
890 * but in cases where all existing nodes leave/fail before mounting
891 * nodes finish control_mount, then all nodes will be mounting and
892 * lvb_gen will be non-zero.
895 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
897 if (lvb_gen == 0xFFFFFFFF) {
898 /* special value to force mount attempts to fail */
899 fs_err(sdp, "control_mount control_lock disabled\n");
904 if (mounted_mode == DLM_LOCK_EX) {
905 /* first mounter, keep both EX while doing first recovery */
906 spin_lock(&ls->ls_recover_spin);
907 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
908 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
909 set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
910 spin_unlock(&ls->ls_recover_spin);
911 fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
915 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
920 * We are not first mounter, now we need to wait for the control_lock
921 * lvb generation to be >= the generation from our first recover_done
922 * and all lvb bits to be clear (no pending journal recoveries.)
925 if (!all_jid_bits_clear(ls->ls_lvb_bits)) {
926 /* journals need recovery, wait until all are clear */
927 fs_info(sdp, "control_mount wait for journal recovery\n");
931 spin_lock(&ls->ls_recover_spin);
932 block_gen = ls->ls_recover_block;
933 start_gen = ls->ls_recover_start;
934 mount_gen = ls->ls_recover_mount;
936 if (lvb_gen < mount_gen) {
937 /* wait for mounted nodes to update control_lock lvb to our
938 generation, which might include new recovery bits set */
939 if (sdp->sd_args.ar_spectator) {
940 fs_info(sdp, "Recovery is required. Waiting for a "
941 "non-spectator to mount.\n");
942 msleep_interruptible(1000);
944 fs_info(sdp, "control_mount wait1 block %u start %u "
945 "mount %u lvb %u flags %lx\n", block_gen,
946 start_gen, mount_gen, lvb_gen,
947 ls->ls_recover_flags);
949 spin_unlock(&ls->ls_recover_spin);
953 if (lvb_gen != start_gen) {
954 /* wait for mounted nodes to update control_lock lvb to the
955 latest recovery generation */
956 fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
957 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
958 lvb_gen, ls->ls_recover_flags);
959 spin_unlock(&ls->ls_recover_spin);
963 if (block_gen == start_gen) {
964 /* dlm recovery in progress, wait for it to finish */
965 fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
966 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
967 lvb_gen, ls->ls_recover_flags);
968 spin_unlock(&ls->ls_recover_spin);
972 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
973 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
974 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
975 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
976 spin_unlock(&ls->ls_recover_spin);
985 static int control_first_done(struct gfs2_sbd *sdp)
987 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
988 uint32_t start_gen, block_gen;
992 spin_lock(&ls->ls_recover_spin);
993 start_gen = ls->ls_recover_start;
994 block_gen = ls->ls_recover_block;
996 if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
997 !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
998 !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
999 /* sanity check, should not happen */
1000 fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
1001 start_gen, block_gen, ls->ls_recover_flags);
1002 spin_unlock(&ls->ls_recover_spin);
1003 control_unlock(sdp);
1007 if (start_gen == block_gen) {
1009 * Wait for the end of a dlm recovery cycle to switch from
1010 * first mounter recovery. We can ignore any recover_slot
1011 * callbacks between the recover_prep and next recover_done
1012 * because we are still the first mounter and any failed nodes
1013 * have not fully mounted, so they don't need recovery.
1015 spin_unlock(&ls->ls_recover_spin);
1016 fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
1018 wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
1019 TASK_UNINTERRUPTIBLE);
1023 clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1024 set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
1025 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
1026 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
1027 spin_unlock(&ls->ls_recover_spin);
1029 memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE);
1030 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
1032 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
1034 fs_err(sdp, "control_first_done mounted PR error %d\n", error);
1036 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
1038 fs_err(sdp, "control_first_done control NL error %d\n", error);
1044 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1045 * to accomodate the largest slot number. (NB dlm slot numbers start at 1,
1046 * gfs2 jids start at 0, so jid = slot - 1)
1049 #define RECOVER_SIZE_INC 16
1051 static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1054 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1055 uint32_t *submit = NULL;
1056 uint32_t *result = NULL;
1057 uint32_t old_size, new_size;
1060 if (!ls->ls_lvb_bits) {
1061 ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1062 if (!ls->ls_lvb_bits)
1067 for (i = 0; i < num_slots; i++) {
1068 if (max_jid < slots[i].slot - 1)
1069 max_jid = slots[i].slot - 1;
1072 old_size = ls->ls_recover_size;
1073 new_size = old_size;
1074 while (new_size < max_jid + 1)
1075 new_size += RECOVER_SIZE_INC;
1076 if (new_size == old_size)
1079 submit = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1080 result = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1081 if (!submit || !result) {
1087 spin_lock(&ls->ls_recover_spin);
1088 memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1089 memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1090 kfree(ls->ls_recover_submit);
1091 kfree(ls->ls_recover_result);
1092 ls->ls_recover_submit = submit;
1093 ls->ls_recover_result = result;
1094 ls->ls_recover_size = new_size;
1095 spin_unlock(&ls->ls_recover_spin);
1099 static void free_recover_size(struct lm_lockstruct *ls)
1101 kfree(ls->ls_lvb_bits);
1102 kfree(ls->ls_recover_submit);
1103 kfree(ls->ls_recover_result);
1104 ls->ls_recover_submit = NULL;
1105 ls->ls_recover_result = NULL;
1106 ls->ls_recover_size = 0;
1107 ls->ls_lvb_bits = NULL;
1110 /* dlm calls before it does lock recovery */
1112 static void gdlm_recover_prep(void *arg)
1114 struct gfs2_sbd *sdp = arg;
1115 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1117 if (gfs2_withdrawn(sdp)) {
1118 fs_err(sdp, "recover_prep ignored due to withdraw.\n");
1121 spin_lock(&ls->ls_recover_spin);
1122 ls->ls_recover_block = ls->ls_recover_start;
1123 set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1125 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1126 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1127 spin_unlock(&ls->ls_recover_spin);
1130 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1131 spin_unlock(&ls->ls_recover_spin);
1134 /* dlm calls after recover_prep has been completed on all lockspace members;
1135 identifies slot/jid of failed member */
1137 static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1139 struct gfs2_sbd *sdp = arg;
1140 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1141 int jid = slot->slot - 1;
1143 if (gfs2_withdrawn(sdp)) {
1144 fs_err(sdp, "recover_slot jid %d ignored due to withdraw.\n",
1148 spin_lock(&ls->ls_recover_spin);
1149 if (ls->ls_recover_size < jid + 1) {
1150 fs_err(sdp, "recover_slot jid %d gen %u short size %d\n",
1151 jid, ls->ls_recover_block, ls->ls_recover_size);
1152 spin_unlock(&ls->ls_recover_spin);
1156 if (ls->ls_recover_submit[jid]) {
1157 fs_info(sdp, "recover_slot jid %d gen %u prev %u\n",
1158 jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1160 ls->ls_recover_submit[jid] = ls->ls_recover_block;
1161 spin_unlock(&ls->ls_recover_spin);
1164 /* dlm calls after recover_slot and after it completes lock recovery */
1166 static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1167 int our_slot, uint32_t generation)
1169 struct gfs2_sbd *sdp = arg;
1170 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1172 if (gfs2_withdrawn(sdp)) {
1173 fs_err(sdp, "recover_done ignored due to withdraw.\n");
1176 /* ensure the ls jid arrays are large enough */
1177 set_recover_size(sdp, slots, num_slots);
1179 spin_lock(&ls->ls_recover_spin);
1180 ls->ls_recover_start = generation;
1182 if (!ls->ls_recover_mount) {
1183 ls->ls_recover_mount = generation;
1184 ls->ls_jid = our_slot - 1;
1187 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1188 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1190 clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1191 smp_mb__after_atomic();
1192 wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1193 spin_unlock(&ls->ls_recover_spin);
1196 /* gfs2_recover thread has a journal recovery result */
1198 static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1199 unsigned int result)
1201 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1203 if (gfs2_withdrawn(sdp)) {
1204 fs_err(sdp, "recovery_result jid %d ignored due to withdraw.\n",
1208 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1211 /* don't care about the recovery of own journal during mount */
1212 if (jid == ls->ls_jid)
1215 spin_lock(&ls->ls_recover_spin);
1216 if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1217 spin_unlock(&ls->ls_recover_spin);
1220 if (ls->ls_recover_size < jid + 1) {
1221 fs_err(sdp, "recovery_result jid %d short size %d\n",
1222 jid, ls->ls_recover_size);
1223 spin_unlock(&ls->ls_recover_spin);
1227 fs_info(sdp, "recover jid %d result %s\n", jid,
1228 result == LM_RD_GAVEUP ? "busy" : "success");
1230 ls->ls_recover_result[jid] = result;
1232 /* GAVEUP means another node is recovering the journal; delay our
1233 next attempt to recover it, to give the other node a chance to
1234 finish before trying again */
1236 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1237 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1238 result == LM_RD_GAVEUP ? HZ : 0);
1239 spin_unlock(&ls->ls_recover_spin);
1242 static const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1243 .recover_prep = gdlm_recover_prep,
1244 .recover_slot = gdlm_recover_slot,
1245 .recover_done = gdlm_recover_done,
1248 static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1250 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1251 char cluster[GFS2_LOCKNAME_LEN];
1254 int error, ops_result;
1257 * initialize everything
1260 INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1261 spin_lock_init(&ls->ls_recover_spin);
1262 ls->ls_recover_flags = 0;
1263 ls->ls_recover_mount = 0;
1264 ls->ls_recover_start = 0;
1265 ls->ls_recover_block = 0;
1266 ls->ls_recover_size = 0;
1267 ls->ls_recover_submit = NULL;
1268 ls->ls_recover_result = NULL;
1269 ls->ls_lvb_bits = NULL;
1271 error = set_recover_size(sdp, NULL, 0);
1276 * prepare dlm_new_lockspace args
1279 fsname = strchr(table, ':');
1281 fs_info(sdp, "no fsname found\n");
1285 memset(cluster, 0, sizeof(cluster));
1286 memcpy(cluster, table, strlen(table) - strlen(fsname));
1289 flags = DLM_LSFL_FS | DLM_LSFL_NEWEXCL;
1292 * create/join lockspace
1295 error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1296 &gdlm_lockspace_ops, sdp, &ops_result,
1299 fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1303 if (ops_result < 0) {
1305 * dlm does not support ops callbacks,
1306 * old dlm_controld/gfs_controld are used, try without ops.
1308 fs_info(sdp, "dlm lockspace ops not used\n");
1309 free_recover_size(ls);
1310 set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1314 if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1315 fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1321 * control_mount() uses control_lock to determine first mounter,
1322 * and for later mounts, waits for any recoveries to be cleared.
1325 error = control_mount(sdp);
1327 fs_err(sdp, "mount control error %d\n", error);
1331 ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1332 clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
1333 smp_mb__after_atomic();
1334 wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1338 dlm_release_lockspace(ls->ls_dlm, 2);
1340 free_recover_size(ls);
1345 static void gdlm_first_done(struct gfs2_sbd *sdp)
1347 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1350 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1353 error = control_first_done(sdp);
1355 fs_err(sdp, "mount first_done error %d\n", error);
1358 static void gdlm_unmount(struct gfs2_sbd *sdp)
1360 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1362 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1365 /* wait for gfs2_control_wq to be done with this mount */
1367 spin_lock(&ls->ls_recover_spin);
1368 set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1369 spin_unlock(&ls->ls_recover_spin);
1370 flush_delayed_work(&sdp->sd_control_work);
1372 /* mounted_lock and control_lock will be purged in dlm recovery */
1375 dlm_release_lockspace(ls->ls_dlm, 2);
1379 free_recover_size(ls);
1382 static const match_table_t dlm_tokens = {
1383 { Opt_jid, "jid=%d"},
1385 { Opt_first, "first=%d"},
1386 { Opt_nodir, "nodir=%d"},
1390 const struct lm_lockops gfs2_dlm_ops = {
1391 .lm_proto_name = "lock_dlm",
1392 .lm_mount = gdlm_mount,
1393 .lm_first_done = gdlm_first_done,
1394 .lm_recovery_result = gdlm_recovery_result,
1395 .lm_unmount = gdlm_unmount,
1396 .lm_put_lock = gdlm_put_lock,
1397 .lm_lock = gdlm_lock,
1398 .lm_cancel = gdlm_cancel,
1399 .lm_tokens = &dlm_tokens,