1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
4 * Copyright 2004-2011 Red Hat, Inc.
5 */
6
7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
8
9 #include <linux/fs.h>
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>
16
17 #include "incore.h"
18 #include "glock.h"
19 #include "glops.h"
20 #include "recovery.h"
21 #include "util.h"
22 #include "sys.h"
23 #include "trace_gfs2.h"
24
25 /**
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
30 */
gfs2_update_stats(struct gfs2_lkstats * s,unsigned index,s64 sample)31 static inline void gfs2_update_stats(struct gfs2_lkstats *s, unsigned index,
32 s64 sample)
33 {
34 /*
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.
42 *
43 * Note that the index points at the array entry containing the
44 * smoothed mean value, and the variance is always in the following
45 * entry
46 *
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.
50 */
51
52 s64 delta = sample - s->stats[index];
53 s->stats[index] += (delta >> 3);
54 index++;
55 s->stats[index] += (s64)(abs(delta) - s->stats[index]) >> 2;
56 }
57
58 /**
59 * gfs2_update_reply_times - Update locking statistics
60 * @gl: The glock to update
61 * @blocking: The operation may have been blocking
62 *
63 * This assumes that gl->gl_dstamp has been set earlier.
64 *
65 * The rtt (lock round trip time) is an estimate of the time
66 * taken to perform a dlm lock request. We update it on each
67 * reply from the dlm.
68 *
69 * The blocking flag is set on the glock for all dlm requests
70 * which may potentially block due to lock requests from other nodes.
71 * DLM requests where the current lock state is exclusive, the
72 * requested state is null (or unlocked) or where the TRY or
73 * TRY_1CB flags are set are classified as non-blocking. All
74 * other DLM requests are counted as (potentially) blocking.
75 */
gfs2_update_reply_times(struct gfs2_glock * gl,bool blocking)76 static inline void gfs2_update_reply_times(struct gfs2_glock *gl,
77 bool blocking)
78 {
79 struct gfs2_pcpu_lkstats *lks;
80 const unsigned gltype = gl->gl_name.ln_type;
81 unsigned index = blocking ? GFS2_LKS_SRTTB : GFS2_LKS_SRTT;
82 s64 rtt;
83
84 preempt_disable();
85 rtt = ktime_to_ns(ktime_sub(ktime_get_real(), gl->gl_dstamp));
86 lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
87 gfs2_update_stats(&gl->gl_stats, index, rtt); /* Local */
88 gfs2_update_stats(&lks->lkstats[gltype], index, rtt); /* Global */
89 preempt_enable();
90
91 trace_gfs2_glock_lock_time(gl, rtt);
92 }
93
94 /**
95 * gfs2_update_request_times - Update locking statistics
96 * @gl: The glock to update
97 *
98 * The irt (lock inter-request times) measures the average time
99 * between requests to the dlm. It is updated immediately before
100 * each dlm call.
101 */
102
gfs2_update_request_times(struct gfs2_glock * gl)103 static inline void gfs2_update_request_times(struct gfs2_glock *gl)
104 {
105 struct gfs2_pcpu_lkstats *lks;
106 const unsigned gltype = gl->gl_name.ln_type;
107 ktime_t dstamp;
108 s64 irt;
109
110 preempt_disable();
111 dstamp = gl->gl_dstamp;
112 gl->gl_dstamp = ktime_get_real();
113 irt = ktime_to_ns(ktime_sub(gl->gl_dstamp, dstamp));
114 lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
115 gfs2_update_stats(&gl->gl_stats, GFS2_LKS_SIRT, irt); /* Local */
116 gfs2_update_stats(&lks->lkstats[gltype], GFS2_LKS_SIRT, irt); /* Global */
117 preempt_enable();
118 }
119
gdlm_ast(void * arg)120 static void gdlm_ast(void *arg)
121 {
122 struct gfs2_glock *gl = arg;
123 bool blocking;
124 unsigned ret;
125
126 blocking = test_bit(GLF_BLOCKING, &gl->gl_flags);
127 gfs2_update_reply_times(gl, blocking);
128 clear_bit(GLF_BLOCKING, &gl->gl_flags);
129
130 /* If the glock is dead, we only react to a dlm_unlock() reply. */
131 if (__lockref_is_dead(&gl->gl_lockref) &&
132 gl->gl_lksb.sb_status != -DLM_EUNLOCK)
133 return;
134
135 BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED);
136
137 if ((gl->gl_lksb.sb_flags & DLM_SBF_VALNOTVALID) && gl->gl_lksb.sb_lvbptr)
138 memset(gl->gl_lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
139
140 switch (gl->gl_lksb.sb_status) {
141 case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */
142 if (gl->gl_ops->go_unlocked)
143 gl->gl_ops->go_unlocked(gl);
144 gfs2_glock_free(gl);
145 return;
146 case -DLM_ECANCEL: /* Cancel while getting lock */
147 ret = LM_OUT_CANCELED;
148 goto out;
149 case -EAGAIN: /* Try lock fails */
150 ret = LM_OUT_TRY_AGAIN;
151 goto out;
152 case -EDEADLK: /* Deadlock detected */
153 ret = LM_OUT_DEADLOCK;
154 goto out;
155 case -ETIMEDOUT: /* Canceled due to timeout */
156 ret = LM_OUT_ERROR;
157 goto out;
158 case 0: /* Success */
159 break;
160 default: /* Something unexpected */
161 BUG();
162 }
163
164 ret = gl->gl_req;
165
166 /*
167 * The GLF_INITIAL flag is initially set for new glocks. Upon the
168 * first successful new (non-conversion) request, we clear this flag to
169 * indicate that a DLM lock exists and that gl->gl_lksb.sb_lkid is the
170 * identifier to use for identifying it.
171 *
172 * Any failed initial requests do not create a DLM lock, so we ignore
173 * the gl->gl_lksb.sb_lkid values that come with such requests.
174 */
175
176 clear_bit(GLF_INITIAL, &gl->gl_flags);
177 gfs2_glock_complete(gl, ret);
178 return;
179 out:
180 if (test_bit(GLF_INITIAL, &gl->gl_flags))
181 gl->gl_lksb.sb_lkid = 0;
182 gfs2_glock_complete(gl, ret);
183 }
184
gdlm_bast(void * arg,int mode)185 static void gdlm_bast(void *arg, int mode)
186 {
187 struct gfs2_glock *gl = arg;
188
189 if (__lockref_is_dead(&gl->gl_lockref))
190 return;
191
192 switch (mode) {
193 case DLM_LOCK_EX:
194 gfs2_glock_cb(gl, LM_ST_UNLOCKED);
195 break;
196 case DLM_LOCK_CW:
197 gfs2_glock_cb(gl, LM_ST_DEFERRED);
198 break;
199 case DLM_LOCK_PR:
200 gfs2_glock_cb(gl, LM_ST_SHARED);
201 break;
202 default:
203 fs_err(gl->gl_name.ln_sbd, "unknown bast mode %d\n", mode);
204 BUG();
205 }
206 }
207
208 /* convert gfs lock-state to dlm lock-mode */
209
make_mode(struct gfs2_sbd * sdp,const unsigned int lmstate)210 static int make_mode(struct gfs2_sbd *sdp, const unsigned int lmstate)
211 {
212 switch (lmstate) {
213 case LM_ST_UNLOCKED:
214 return DLM_LOCK_NL;
215 case LM_ST_EXCLUSIVE:
216 return DLM_LOCK_EX;
217 case LM_ST_DEFERRED:
218 return DLM_LOCK_CW;
219 case LM_ST_SHARED:
220 return DLM_LOCK_PR;
221 }
222 fs_err(sdp, "unknown LM state %d\n", lmstate);
223 BUG();
224 return -1;
225 }
226
227 /* Taken from fs/dlm/lock.c. */
228
middle_conversion(int cur,int req)229 static bool middle_conversion(int cur, int req)
230 {
231 return (cur == DLM_LOCK_PR && req == DLM_LOCK_CW) ||
232 (cur == DLM_LOCK_CW && req == DLM_LOCK_PR);
233 }
234
down_conversion(int cur,int req)235 static bool down_conversion(int cur, int req)
236 {
237 return !middle_conversion(cur, req) && req < cur;
238 }
239
make_flags(struct gfs2_glock * gl,const unsigned int gfs_flags,const int req,bool blocking)240 static u32 make_flags(struct gfs2_glock *gl, const unsigned int gfs_flags,
241 const int req, bool blocking)
242 {
243 u32 lkf = 0;
244
245 if (gl->gl_lksb.sb_lvbptr)
246 lkf |= DLM_LKF_VALBLK;
247
248 if (gfs_flags & LM_FLAG_TRY)
249 lkf |= DLM_LKF_NOQUEUE;
250
251 if (gfs_flags & LM_FLAG_TRY_1CB) {
252 lkf |= DLM_LKF_NOQUEUE;
253 lkf |= DLM_LKF_NOQUEUEBAST;
254 }
255
256 if (!test_bit(GLF_INITIAL, &gl->gl_flags)) {
257 lkf |= DLM_LKF_CONVERT;
258
259 /*
260 * The DLM_LKF_QUECVT flag needs to be set for "first come,
261 * first served" semantics, but it must only be set for
262 * "upward" lock conversions or else DLM will reject the
263 * request as invalid.
264 */
265 if (blocking)
266 lkf |= DLM_LKF_QUECVT;
267 }
268
269 return lkf;
270 }
271
gfs2_reverse_hex(char * c,u64 value)272 static void gfs2_reverse_hex(char *c, u64 value)
273 {
274 *c = '0';
275 while (value) {
276 *c-- = hex_asc[value & 0x0f];
277 value >>= 4;
278 }
279 }
280
gdlm_lock(struct gfs2_glock * gl,unsigned int req_state,unsigned int flags)281 static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state,
282 unsigned int flags)
283 {
284 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
285 bool blocking;
286 int cur, req;
287 u32 lkf;
288 char strname[GDLM_STRNAME_BYTES] = "";
289 int error;
290
291 gl->gl_req = req_state;
292 cur = make_mode(gl->gl_name.ln_sbd, gl->gl_state);
293 req = make_mode(gl->gl_name.ln_sbd, req_state);
294 blocking = !down_conversion(cur, req) &&
295 !(flags & (LM_FLAG_TRY|LM_FLAG_TRY_1CB));
296 lkf = make_flags(gl, flags, req, blocking);
297 if (blocking)
298 set_bit(GLF_BLOCKING, &gl->gl_flags);
299 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
300 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
301 if (test_bit(GLF_INITIAL, &gl->gl_flags)) {
302 memset(strname, ' ', GDLM_STRNAME_BYTES - 1);
303 strname[GDLM_STRNAME_BYTES - 1] = '\0';
304 gfs2_reverse_hex(strname + 7, gl->gl_name.ln_type);
305 gfs2_reverse_hex(strname + 23, gl->gl_name.ln_number);
306 gl->gl_dstamp = ktime_get_real();
307 } else {
308 gfs2_update_request_times(gl);
309 }
310 /*
311 * Submit the actual lock request.
312 */
313
314 again:
315 down_read(&ls->ls_sem);
316 error = -ENODEV;
317 if (likely(ls->ls_dlm != NULL)) {
318 error = dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname,
319 GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast);
320 }
321 up_read(&ls->ls_sem);
322 if (error == -EBUSY) {
323 msleep(20);
324 goto again;
325 }
326 return error;
327 }
328
gdlm_put_lock(struct gfs2_glock * gl)329 static void gdlm_put_lock(struct gfs2_glock *gl)
330 {
331 struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
332 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
333 uint32_t flags = 0;
334 int error;
335
336 BUG_ON(!__lockref_is_dead(&gl->gl_lockref));
337
338 if (test_bit(GLF_INITIAL, &gl->gl_flags)) {
339 gfs2_glock_free(gl);
340 return;
341 }
342
343 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
344 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
345 gfs2_update_request_times(gl);
346
347 /*
348 * When the lockspace is released, all remaining glocks will be
349 * unlocked automatically. This is more efficient than unlocking them
350 * individually, but when the lock is held in DLM_LOCK_EX or
351 * DLM_LOCK_PW mode, the lock value block (LVB) would be lost.
352 */
353
354 if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
355 (!gl->gl_lksb.sb_lvbptr || gl->gl_state != LM_ST_EXCLUSIVE)) {
356 gfs2_glock_free_later(gl);
357 return;
358 }
359
360 if (gl->gl_lksb.sb_lvbptr)
361 flags |= DLM_LKF_VALBLK;
362
363 again:
364 down_read(&ls->ls_sem);
365 error = -ENODEV;
366 if (likely(ls->ls_dlm != NULL)) {
367 error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, flags,
368 NULL, gl);
369 }
370 up_read(&ls->ls_sem);
371 if (error == -EBUSY) {
372 msleep(20);
373 goto again;
374 }
375
376 if (error == -ENODEV) {
377 gfs2_glock_free(gl);
378 return;
379 }
380
381 if (error) {
382 fs_err(sdp, "gdlm_unlock %x,%llx err=%d\n",
383 gl->gl_name.ln_type,
384 (unsigned long long)gl->gl_name.ln_number, error);
385 }
386 }
387
gdlm_cancel(struct gfs2_glock * gl)388 static void gdlm_cancel(struct gfs2_glock *gl)
389 {
390 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
391
392 down_read(&ls->ls_sem);
393 if (likely(ls->ls_dlm != NULL)) {
394 dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
395 }
396 up_read(&ls->ls_sem);
397 }
398
399 /*
400 * dlm/gfs2 recovery coordination using dlm_recover callbacks
401 *
402 * 0. gfs2 checks for another cluster node withdraw, needing journal replay
403 * 1. dlm_controld sees lockspace members change
404 * 2. dlm_controld blocks dlm-kernel locking activity
405 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
406 * 4. dlm_controld starts and finishes its own user level recovery
407 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
408 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
409 * 7. dlm_recoverd does its own lock recovery
410 * 8. dlm_recoverd unblocks dlm-kernel locking activity
411 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
412 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
413 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
414 * 12. gfs2_recover dequeues and recovers journals of failed nodes
415 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
416 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
417 * 15. gfs2_control unblocks normal locking when all journals are recovered
418 *
419 * - failures during recovery
420 *
421 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
422 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
423 * recovering for a prior failure. gfs2_control needs a way to detect
424 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
425 * the recover_block and recover_start values.
426 *
427 * recover_done() provides a new lockspace generation number each time it
428 * is called (step 9). This generation number is saved as recover_start.
429 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
430 * recover_block = recover_start. So, while recover_block is equal to
431 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
432 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
433 *
434 * - more specific gfs2 steps in sequence above
435 *
436 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
437 * 6. recover_slot records any failed jids (maybe none)
438 * 9. recover_done sets recover_start = new generation number
439 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
440 * 12. gfs2_recover does journal recoveries for failed jids identified above
441 * 14. gfs2_control clears control_lock lvb bits for recovered jids
442 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
443 * again) then do nothing, otherwise if recover_start > recover_block
444 * then clear BLOCK_LOCKS.
445 *
446 * - parallel recovery steps across all nodes
447 *
448 * All nodes attempt to update the control_lock lvb with the new generation
449 * number and jid bits, but only the first to get the control_lock EX will
450 * do so; others will see that it's already done (lvb already contains new
451 * generation number.)
452 *
453 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
454 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
455 * . One node gets control_lock first and writes the lvb, others see it's done
456 * . All nodes attempt to recover jids for which they see control_lock bits set
457 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
458 * . All nodes will eventually see all lvb bits clear and unblock locks
459 *
460 * - is there a problem with clearing an lvb bit that should be set
461 * and missing a journal recovery?
462 *
463 * 1. jid fails
464 * 2. lvb bit set for step 1
465 * 3. jid recovered for step 1
466 * 4. jid taken again (new mount)
467 * 5. jid fails (for step 4)
468 * 6. lvb bit set for step 5 (will already be set)
469 * 7. lvb bit cleared for step 3
470 *
471 * This is not a problem because the failure in step 5 does not
472 * require recovery, because the mount in step 4 could not have
473 * progressed far enough to unblock locks and access the fs. The
474 * control_mount() function waits for all recoveries to be complete
475 * for the latest lockspace generation before ever unblocking locks
476 * and returning. The mount in step 4 waits until the recovery in
477 * step 1 is done.
478 *
479 * - special case of first mounter: first node to mount the fs
480 *
481 * The first node to mount a gfs2 fs needs to check all the journals
482 * and recover any that need recovery before other nodes are allowed
483 * to mount the fs. (Others may begin mounting, but they must wait
484 * for the first mounter to be done before taking locks on the fs
485 * or accessing the fs.) This has two parts:
486 *
487 * 1. The mounted_lock tells a node it's the first to mount the fs.
488 * Each node holds the mounted_lock in PR while it's mounted.
489 * Each node tries to acquire the mounted_lock in EX when it mounts.
490 * If a node is granted the mounted_lock EX it means there are no
491 * other mounted nodes (no PR locks exist), and it is the first mounter.
492 * The mounted_lock is demoted to PR when first recovery is done, so
493 * others will fail to get an EX lock, but will get a PR lock.
494 *
495 * 2. The control_lock blocks others in control_mount() while the first
496 * mounter is doing first mount recovery of all journals.
497 * A mounting node needs to acquire control_lock in EX mode before
498 * it can proceed. The first mounter holds control_lock in EX while doing
499 * the first mount recovery, blocking mounts from other nodes, then demotes
500 * control_lock to NL when it's done (others_may_mount/first_done),
501 * allowing other nodes to continue mounting.
502 *
503 * first mounter:
504 * control_lock EX/NOQUEUE success
505 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
506 * set first=1
507 * do first mounter recovery
508 * mounted_lock EX->PR
509 * control_lock EX->NL, write lvb generation
510 *
511 * other mounter:
512 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
513 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
514 * mounted_lock PR/NOQUEUE success
515 * read lvb generation
516 * control_lock EX->NL
517 * set first=0
518 *
519 * - mount during recovery
520 *
521 * If a node mounts while others are doing recovery (not first mounter),
522 * the mounting node will get its initial recover_done() callback without
523 * having seen any previous failures/callbacks.
524 *
525 * It must wait for all recoveries preceding its mount to be finished
526 * before it unblocks locks. It does this by repeating the "other mounter"
527 * steps above until the lvb generation number is >= its mount generation
528 * number (from initial recover_done) and all lvb bits are clear.
529 *
530 * - control_lock lvb format
531 *
532 * 4 bytes generation number: the latest dlm lockspace generation number
533 * from recover_done callback. Indicates the jid bitmap has been updated
534 * to reflect all slot failures through that generation.
535 * 4 bytes unused.
536 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
537 * that jid N needs recovery.
538 */
539
540 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
541
control_lvb_read(struct lm_lockstruct * ls,uint32_t * lvb_gen,char * lvb_bits)542 static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
543 char *lvb_bits)
544 {
545 __le32 gen;
546 memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE);
547 memcpy(&gen, lvb_bits, sizeof(__le32));
548 *lvb_gen = le32_to_cpu(gen);
549 }
550
control_lvb_write(struct lm_lockstruct * ls,uint32_t lvb_gen,char * lvb_bits)551 static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
552 char *lvb_bits)
553 {
554 __le32 gen;
555 memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE);
556 gen = cpu_to_le32(lvb_gen);
557 memcpy(ls->ls_control_lvb, &gen, sizeof(__le32));
558 }
559
all_jid_bits_clear(char * lvb)560 static int all_jid_bits_clear(char *lvb)
561 {
562 return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0,
563 GDLM_LVB_SIZE - JID_BITMAP_OFFSET);
564 }
565
sync_wait_cb(void * arg)566 static void sync_wait_cb(void *arg)
567 {
568 struct lm_lockstruct *ls = arg;
569 complete(&ls->ls_sync_wait);
570 }
571
sync_unlock(struct gfs2_sbd * sdp,struct dlm_lksb * lksb,char * name)572 static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
573 {
574 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
575 int error;
576
577 down_read(&ls->ls_sem);
578 error = -ENODEV;
579 if (likely(ls->ls_dlm != NULL))
580 error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
581 up_read(&ls->ls_sem);
582 if (error) {
583 fs_err(sdp, "%s lkid %x error %d\n",
584 name, lksb->sb_lkid, error);
585 return error;
586 }
587
588 wait_for_completion(&ls->ls_sync_wait);
589
590 if (lksb->sb_status != -DLM_EUNLOCK) {
591 fs_err(sdp, "%s lkid %x status %d\n",
592 name, lksb->sb_lkid, lksb->sb_status);
593 return -1;
594 }
595 return 0;
596 }
597
sync_lock(struct gfs2_sbd * sdp,int mode,uint32_t flags,unsigned int num,struct dlm_lksb * lksb,char * name)598 static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
599 unsigned int num, struct dlm_lksb *lksb, char *name)
600 {
601 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
602 char strname[GDLM_STRNAME_BYTES];
603 int error, status;
604
605 memset(strname, 0, GDLM_STRNAME_BYTES);
606 snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
607
608 down_read(&ls->ls_sem);
609 error = -ENODEV;
610 if (likely(ls->ls_dlm != NULL)) {
611 error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
612 strname, GDLM_STRNAME_BYTES - 1,
613 0, sync_wait_cb, ls, NULL);
614 }
615 up_read(&ls->ls_sem);
616 if (error) {
617 fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
618 name, lksb->sb_lkid, flags, mode, error);
619 return error;
620 }
621
622 wait_for_completion(&ls->ls_sync_wait);
623
624 status = lksb->sb_status;
625
626 if (status && status != -EAGAIN) {
627 fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
628 name, lksb->sb_lkid, flags, mode, status);
629 }
630
631 return status;
632 }
633
mounted_unlock(struct gfs2_sbd * sdp)634 static int mounted_unlock(struct gfs2_sbd *sdp)
635 {
636 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
637 return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
638 }
639
mounted_lock(struct gfs2_sbd * sdp,int mode,uint32_t flags)640 static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
641 {
642 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
643 return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
644 &ls->ls_mounted_lksb, "mounted_lock");
645 }
646
control_unlock(struct gfs2_sbd * sdp)647 static int control_unlock(struct gfs2_sbd *sdp)
648 {
649 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
650 return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
651 }
652
control_lock(struct gfs2_sbd * sdp,int mode,uint32_t flags)653 static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
654 {
655 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
656 return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK,
657 &ls->ls_control_lksb, "control_lock");
658 }
659
660 /**
661 * remote_withdraw - react to a node withdrawing from the file system
662 * @sdp: The superblock
663 */
remote_withdraw(struct gfs2_sbd * sdp)664 static void remote_withdraw(struct gfs2_sbd *sdp)
665 {
666 struct gfs2_jdesc *jd;
667 int ret = 0, count = 0;
668
669 list_for_each_entry(jd, &sdp->sd_jindex_list, jd_list) {
670 if (jd->jd_jid == sdp->sd_lockstruct.ls_jid)
671 continue;
672 ret = gfs2_recover_journal(jd, true);
673 if (ret)
674 break;
675 count++;
676 }
677
678 /* Now drop the additional reference we acquired */
679 fs_err(sdp, "Journals checked: %d, ret = %d.\n", count, ret);
680 }
681
gfs2_control_func(struct work_struct * work)682 static void gfs2_control_func(struct work_struct *work)
683 {
684 struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
685 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
686 uint32_t block_gen, start_gen, lvb_gen, flags;
687 int recover_set = 0;
688 int write_lvb = 0;
689 int recover_size;
690 int i, error;
691
692 /* First check for other nodes that may have done a withdraw. */
693 if (test_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags)) {
694 remote_withdraw(sdp);
695 clear_bit(SDF_REMOTE_WITHDRAW, &sdp->sd_flags);
696 return;
697 }
698
699 spin_lock(&ls->ls_recover_spin);
700 /*
701 * No MOUNT_DONE means we're still mounting; control_mount()
702 * will set this flag, after which this thread will take over
703 * all further clearing of BLOCK_LOCKS.
704 *
705 * FIRST_MOUNT means this node is doing first mounter recovery,
706 * for which recovery control is handled by
707 * control_mount()/control_first_done(), not this thread.
708 */
709 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
710 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
711 spin_unlock(&ls->ls_recover_spin);
712 return;
713 }
714 block_gen = ls->ls_recover_block;
715 start_gen = ls->ls_recover_start;
716 spin_unlock(&ls->ls_recover_spin);
717
718 /*
719 * Equal block_gen and start_gen implies we are between
720 * recover_prep and recover_done callbacks, which means
721 * dlm recovery is in progress and dlm locking is blocked.
722 * There's no point trying to do any work until recover_done.
723 */
724
725 if (block_gen == start_gen)
726 return;
727
728 /*
729 * Propagate recover_submit[] and recover_result[] to lvb:
730 * dlm_recoverd adds to recover_submit[] jids needing recovery
731 * gfs2_recover adds to recover_result[] journal recovery results
732 *
733 * set lvb bit for jids in recover_submit[] if the lvb has not
734 * yet been updated for the generation of the failure
735 *
736 * clear lvb bit for jids in recover_result[] if the result of
737 * the journal recovery is SUCCESS
738 */
739
740 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
741 if (error) {
742 fs_err(sdp, "control lock EX error %d\n", error);
743 return;
744 }
745
746 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
747
748 spin_lock(&ls->ls_recover_spin);
749 if (block_gen != ls->ls_recover_block ||
750 start_gen != ls->ls_recover_start) {
751 fs_info(sdp, "recover generation %u block1 %u %u\n",
752 start_gen, block_gen, ls->ls_recover_block);
753 spin_unlock(&ls->ls_recover_spin);
754 control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
755 return;
756 }
757
758 recover_size = ls->ls_recover_size;
759
760 if (lvb_gen <= start_gen) {
761 /*
762 * Clear lvb bits for jids we've successfully recovered.
763 * Because all nodes attempt to recover failed journals,
764 * a journal can be recovered multiple times successfully
765 * in succession. Only the first will really do recovery,
766 * the others find it clean, but still report a successful
767 * recovery. So, another node may have already recovered
768 * the jid and cleared the lvb bit for it.
769 */
770 for (i = 0; i < recover_size; i++) {
771 if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
772 continue;
773
774 ls->ls_recover_result[i] = 0;
775
776 if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
777 continue;
778
779 __clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
780 write_lvb = 1;
781 }
782 }
783
784 if (lvb_gen == start_gen) {
785 /*
786 * Failed slots before start_gen are already set in lvb.
787 */
788 for (i = 0; i < recover_size; i++) {
789 if (!ls->ls_recover_submit[i])
790 continue;
791 if (ls->ls_recover_submit[i] < lvb_gen)
792 ls->ls_recover_submit[i] = 0;
793 }
794 } else if (lvb_gen < start_gen) {
795 /*
796 * Failed slots before start_gen are not yet set in lvb.
797 */
798 for (i = 0; i < recover_size; i++) {
799 if (!ls->ls_recover_submit[i])
800 continue;
801 if (ls->ls_recover_submit[i] < start_gen) {
802 ls->ls_recover_submit[i] = 0;
803 __set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
804 }
805 }
806 /* even if there are no bits to set, we need to write the
807 latest generation to the lvb */
808 write_lvb = 1;
809 } else {
810 /*
811 * we should be getting a recover_done() for lvb_gen soon
812 */
813 }
814 spin_unlock(&ls->ls_recover_spin);
815
816 if (write_lvb) {
817 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
818 flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
819 } else {
820 flags = DLM_LKF_CONVERT;
821 }
822
823 error = control_lock(sdp, DLM_LOCK_NL, flags);
824 if (error) {
825 fs_err(sdp, "control lock NL error %d\n", error);
826 return;
827 }
828
829 /*
830 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
831 * and clear a jid bit in the lvb if the recovery is a success.
832 * Eventually all journals will be recovered, all jid bits will
833 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
834 */
835
836 for (i = 0; i < recover_size; i++) {
837 if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
838 fs_info(sdp, "recover generation %u jid %d\n",
839 start_gen, i);
840 gfs2_recover_set(sdp, i);
841 recover_set++;
842 }
843 }
844 if (recover_set)
845 return;
846
847 /*
848 * No more jid bits set in lvb, all recovery is done, unblock locks
849 * (unless a new recover_prep callback has occured blocking locks
850 * again while working above)
851 */
852
853 spin_lock(&ls->ls_recover_spin);
854 if (ls->ls_recover_block == block_gen &&
855 ls->ls_recover_start == start_gen) {
856 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
857 spin_unlock(&ls->ls_recover_spin);
858 fs_info(sdp, "recover generation %u done\n", start_gen);
859 gfs2_glock_thaw(sdp);
860 } else {
861 fs_info(sdp, "recover generation %u block2 %u %u\n",
862 start_gen, block_gen, ls->ls_recover_block);
863 spin_unlock(&ls->ls_recover_spin);
864 }
865 }
866
control_mount(struct gfs2_sbd * sdp)867 static int control_mount(struct gfs2_sbd *sdp)
868 {
869 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
870 uint32_t start_gen, block_gen, mount_gen, lvb_gen;
871 int mounted_mode;
872 int retries = 0;
873 int error;
874
875 memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
876 memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
877 memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
878 ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
879 init_completion(&ls->ls_sync_wait);
880
881 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
882
883 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
884 if (error) {
885 fs_err(sdp, "control_mount control_lock NL error %d\n", error);
886 return error;
887 }
888
889 error = mounted_lock(sdp, DLM_LOCK_NL, 0);
890 if (error) {
891 fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
892 control_unlock(sdp);
893 return error;
894 }
895 mounted_mode = DLM_LOCK_NL;
896
897 restart:
898 if (retries++ && signal_pending(current)) {
899 error = -EINTR;
900 goto fail;
901 }
902
903 /*
904 * We always start with both locks in NL. control_lock is
905 * demoted to NL below so we don't need to do it here.
906 */
907
908 if (mounted_mode != DLM_LOCK_NL) {
909 error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
910 if (error)
911 goto fail;
912 mounted_mode = DLM_LOCK_NL;
913 }
914
915 /*
916 * Other nodes need to do some work in dlm recovery and gfs2_control
917 * before the recover_done and control_lock will be ready for us below.
918 * A delay here is not required but often avoids having to retry.
919 */
920
921 msleep_interruptible(500);
922
923 /*
924 * Acquire control_lock in EX and mounted_lock in either EX or PR.
925 * control_lock lvb keeps track of any pending journal recoveries.
926 * mounted_lock indicates if any other nodes have the fs mounted.
927 */
928
929 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
930 if (error == -EAGAIN) {
931 goto restart;
932 } else if (error) {
933 fs_err(sdp, "control_mount control_lock EX error %d\n", error);
934 goto fail;
935 }
936
937 /**
938 * If we're a spectator, we don't want to take the lock in EX because
939 * we cannot do the first-mount responsibility it implies: recovery.
940 */
941 if (sdp->sd_args.ar_spectator)
942 goto locks_done;
943
944 error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
945 if (!error) {
946 mounted_mode = DLM_LOCK_EX;
947 goto locks_done;
948 } else if (error != -EAGAIN) {
949 fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
950 goto fail;
951 }
952
953 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
954 if (!error) {
955 mounted_mode = DLM_LOCK_PR;
956 goto locks_done;
957 } else {
958 /* not even -EAGAIN should happen here */
959 fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
960 goto fail;
961 }
962
963 locks_done:
964 /*
965 * If we got both locks above in EX, then we're the first mounter.
966 * If not, then we need to wait for the control_lock lvb to be
967 * updated by other mounted nodes to reflect our mount generation.
968 *
969 * In simple first mounter cases, first mounter will see zero lvb_gen,
970 * but in cases where all existing nodes leave/fail before mounting
971 * nodes finish control_mount, then all nodes will be mounting and
972 * lvb_gen will be non-zero.
973 */
974
975 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
976
977 if (lvb_gen == 0xFFFFFFFF) {
978 /* special value to force mount attempts to fail */
979 fs_err(sdp, "control_mount control_lock disabled\n");
980 error = -EINVAL;
981 goto fail;
982 }
983
984 if (mounted_mode == DLM_LOCK_EX) {
985 /* first mounter, keep both EX while doing first recovery */
986 spin_lock(&ls->ls_recover_spin);
987 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
988 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
989 set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
990 spin_unlock(&ls->ls_recover_spin);
991 fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
992 return 0;
993 }
994
995 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
996 if (error)
997 goto fail;
998
999 /*
1000 * We are not first mounter, now we need to wait for the control_lock
1001 * lvb generation to be >= the generation from our first recover_done
1002 * and all lvb bits to be clear (no pending journal recoveries.)
1003 */
1004
1005 if (!all_jid_bits_clear(ls->ls_lvb_bits)) {
1006 /* journals need recovery, wait until all are clear */
1007 fs_info(sdp, "control_mount wait for journal recovery\n");
1008 goto restart;
1009 }
1010
1011 spin_lock(&ls->ls_recover_spin);
1012 block_gen = ls->ls_recover_block;
1013 start_gen = ls->ls_recover_start;
1014 mount_gen = ls->ls_recover_mount;
1015
1016 if (lvb_gen < mount_gen) {
1017 /* wait for mounted nodes to update control_lock lvb to our
1018 generation, which might include new recovery bits set */
1019 if (sdp->sd_args.ar_spectator) {
1020 fs_info(sdp, "Recovery is required. Waiting for a "
1021 "non-spectator to mount.\n");
1022 spin_unlock(&ls->ls_recover_spin);
1023 msleep_interruptible(1000);
1024 } else {
1025 fs_info(sdp, "control_mount wait1 block %u start %u "
1026 "mount %u lvb %u flags %lx\n", block_gen,
1027 start_gen, mount_gen, lvb_gen,
1028 ls->ls_recover_flags);
1029 spin_unlock(&ls->ls_recover_spin);
1030 }
1031 goto restart;
1032 }
1033
1034 if (lvb_gen != start_gen) {
1035 /* wait for mounted nodes to update control_lock lvb to the
1036 latest recovery generation */
1037 fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
1038 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
1039 lvb_gen, ls->ls_recover_flags);
1040 spin_unlock(&ls->ls_recover_spin);
1041 goto restart;
1042 }
1043
1044 if (block_gen == start_gen) {
1045 /* dlm recovery in progress, wait for it to finish */
1046 fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
1047 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
1048 lvb_gen, ls->ls_recover_flags);
1049 spin_unlock(&ls->ls_recover_spin);
1050 goto restart;
1051 }
1052
1053 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1054 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
1055 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
1056 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
1057 spin_unlock(&ls->ls_recover_spin);
1058 return 0;
1059
1060 fail:
1061 mounted_unlock(sdp);
1062 control_unlock(sdp);
1063 return error;
1064 }
1065
control_first_done(struct gfs2_sbd * sdp)1066 static int control_first_done(struct gfs2_sbd *sdp)
1067 {
1068 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1069 uint32_t start_gen, block_gen;
1070 int error;
1071
1072 restart:
1073 spin_lock(&ls->ls_recover_spin);
1074 start_gen = ls->ls_recover_start;
1075 block_gen = ls->ls_recover_block;
1076
1077 if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
1078 !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1079 !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1080 /* sanity check, should not happen */
1081 fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
1082 start_gen, block_gen, ls->ls_recover_flags);
1083 spin_unlock(&ls->ls_recover_spin);
1084 control_unlock(sdp);
1085 return -1;
1086 }
1087
1088 if (start_gen == block_gen) {
1089 /*
1090 * Wait for the end of a dlm recovery cycle to switch from
1091 * first mounter recovery. We can ignore any recover_slot
1092 * callbacks between the recover_prep and next recover_done
1093 * because we are still the first mounter and any failed nodes
1094 * have not fully mounted, so they don't need recovery.
1095 */
1096 spin_unlock(&ls->ls_recover_spin);
1097 fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
1098
1099 wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
1100 TASK_UNINTERRUPTIBLE);
1101 goto restart;
1102 }
1103
1104 clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1105 set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
1106 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
1107 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
1108 spin_unlock(&ls->ls_recover_spin);
1109
1110 memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE);
1111 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
1112
1113 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
1114 if (error)
1115 fs_err(sdp, "control_first_done mounted PR error %d\n", error);
1116
1117 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
1118 if (error)
1119 fs_err(sdp, "control_first_done control NL error %d\n", error);
1120
1121 return error;
1122 }
1123
1124 /*
1125 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1126 * to accommodate the largest slot number. (NB dlm slot numbers start at 1,
1127 * gfs2 jids start at 0, so jid = slot - 1)
1128 */
1129
1130 #define RECOVER_SIZE_INC 16
1131
set_recover_size(struct gfs2_sbd * sdp,struct dlm_slot * slots,int num_slots)1132 static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1133 int num_slots)
1134 {
1135 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1136 uint32_t *submit = NULL;
1137 uint32_t *result = NULL;
1138 uint32_t old_size, new_size;
1139 int i, max_jid;
1140
1141 if (!ls->ls_lvb_bits) {
1142 ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1143 if (!ls->ls_lvb_bits)
1144 return -ENOMEM;
1145 }
1146
1147 max_jid = 0;
1148 for (i = 0; i < num_slots; i++) {
1149 if (max_jid < slots[i].slot - 1)
1150 max_jid = slots[i].slot - 1;
1151 }
1152
1153 old_size = ls->ls_recover_size;
1154 new_size = old_size;
1155 while (new_size < max_jid + 1)
1156 new_size += RECOVER_SIZE_INC;
1157 if (new_size == old_size)
1158 return 0;
1159
1160 submit = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1161 result = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1162 if (!submit || !result) {
1163 kfree(submit);
1164 kfree(result);
1165 return -ENOMEM;
1166 }
1167
1168 spin_lock(&ls->ls_recover_spin);
1169 memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1170 memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1171 kfree(ls->ls_recover_submit);
1172 kfree(ls->ls_recover_result);
1173 ls->ls_recover_submit = submit;
1174 ls->ls_recover_result = result;
1175 ls->ls_recover_size = new_size;
1176 spin_unlock(&ls->ls_recover_spin);
1177 return 0;
1178 }
1179
free_recover_size(struct lm_lockstruct * ls)1180 static void free_recover_size(struct lm_lockstruct *ls)
1181 {
1182 kfree(ls->ls_lvb_bits);
1183 kfree(ls->ls_recover_submit);
1184 kfree(ls->ls_recover_result);
1185 ls->ls_recover_submit = NULL;
1186 ls->ls_recover_result = NULL;
1187 ls->ls_recover_size = 0;
1188 ls->ls_lvb_bits = NULL;
1189 }
1190
1191 /* dlm calls before it does lock recovery */
1192
gdlm_recover_prep(void * arg)1193 static void gdlm_recover_prep(void *arg)
1194 {
1195 struct gfs2_sbd *sdp = arg;
1196 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1197
1198 if (gfs2_withdrawing_or_withdrawn(sdp)) {
1199 fs_err(sdp, "recover_prep ignored due to withdraw.\n");
1200 return;
1201 }
1202 spin_lock(&ls->ls_recover_spin);
1203 ls->ls_recover_block = ls->ls_recover_start;
1204 set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1205
1206 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1207 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1208 spin_unlock(&ls->ls_recover_spin);
1209 return;
1210 }
1211 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1212 spin_unlock(&ls->ls_recover_spin);
1213 }
1214
1215 /* dlm calls after recover_prep has been completed on all lockspace members;
1216 identifies slot/jid of failed member */
1217
gdlm_recover_slot(void * arg,struct dlm_slot * slot)1218 static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1219 {
1220 struct gfs2_sbd *sdp = arg;
1221 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1222 int jid = slot->slot - 1;
1223
1224 if (gfs2_withdrawing_or_withdrawn(sdp)) {
1225 fs_err(sdp, "recover_slot jid %d ignored due to withdraw.\n",
1226 jid);
1227 return;
1228 }
1229 spin_lock(&ls->ls_recover_spin);
1230 if (ls->ls_recover_size < jid + 1) {
1231 fs_err(sdp, "recover_slot jid %d gen %u short size %d\n",
1232 jid, ls->ls_recover_block, ls->ls_recover_size);
1233 spin_unlock(&ls->ls_recover_spin);
1234 return;
1235 }
1236
1237 if (ls->ls_recover_submit[jid]) {
1238 fs_info(sdp, "recover_slot jid %d gen %u prev %u\n",
1239 jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1240 }
1241 ls->ls_recover_submit[jid] = ls->ls_recover_block;
1242 spin_unlock(&ls->ls_recover_spin);
1243 }
1244
1245 /* dlm calls after recover_slot and after it completes lock recovery */
1246
gdlm_recover_done(void * arg,struct dlm_slot * slots,int num_slots,int our_slot,uint32_t generation)1247 static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1248 int our_slot, uint32_t generation)
1249 {
1250 struct gfs2_sbd *sdp = arg;
1251 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1252
1253 if (gfs2_withdrawing_or_withdrawn(sdp)) {
1254 fs_err(sdp, "recover_done ignored due to withdraw.\n");
1255 return;
1256 }
1257 /* ensure the ls jid arrays are large enough */
1258 set_recover_size(sdp, slots, num_slots);
1259
1260 spin_lock(&ls->ls_recover_spin);
1261 ls->ls_recover_start = generation;
1262
1263 if (!ls->ls_recover_mount) {
1264 ls->ls_recover_mount = generation;
1265 ls->ls_jid = our_slot - 1;
1266 }
1267
1268 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1269 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1270
1271 clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1272 smp_mb__after_atomic();
1273 wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1274 spin_unlock(&ls->ls_recover_spin);
1275 }
1276
1277 /* gfs2_recover thread has a journal recovery result */
1278
gdlm_recovery_result(struct gfs2_sbd * sdp,unsigned int jid,unsigned int result)1279 static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1280 unsigned int result)
1281 {
1282 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1283
1284 if (gfs2_withdrawing_or_withdrawn(sdp)) {
1285 fs_err(sdp, "recovery_result jid %d ignored due to withdraw.\n",
1286 jid);
1287 return;
1288 }
1289 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1290 return;
1291
1292 /* don't care about the recovery of own journal during mount */
1293 if (jid == ls->ls_jid)
1294 return;
1295
1296 spin_lock(&ls->ls_recover_spin);
1297 if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1298 spin_unlock(&ls->ls_recover_spin);
1299 return;
1300 }
1301 if (ls->ls_recover_size < jid + 1) {
1302 fs_err(sdp, "recovery_result jid %d short size %d\n",
1303 jid, ls->ls_recover_size);
1304 spin_unlock(&ls->ls_recover_spin);
1305 return;
1306 }
1307
1308 fs_info(sdp, "recover jid %d result %s\n", jid,
1309 result == LM_RD_GAVEUP ? "busy" : "success");
1310
1311 ls->ls_recover_result[jid] = result;
1312
1313 /* GAVEUP means another node is recovering the journal; delay our
1314 next attempt to recover it, to give the other node a chance to
1315 finish before trying again */
1316
1317 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1318 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1319 result == LM_RD_GAVEUP ? HZ : 0);
1320 spin_unlock(&ls->ls_recover_spin);
1321 }
1322
1323 static const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1324 .recover_prep = gdlm_recover_prep,
1325 .recover_slot = gdlm_recover_slot,
1326 .recover_done = gdlm_recover_done,
1327 };
1328
gdlm_mount(struct gfs2_sbd * sdp,const char * table)1329 static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1330 {
1331 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1332 char cluster[GFS2_LOCKNAME_LEN];
1333 const char *fsname;
1334 uint32_t flags;
1335 int error, ops_result;
1336
1337 /*
1338 * initialize everything
1339 */
1340
1341 INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1342 ls->ls_dlm = NULL;
1343 spin_lock_init(&ls->ls_recover_spin);
1344 ls->ls_recover_flags = 0;
1345 ls->ls_recover_mount = 0;
1346 ls->ls_recover_start = 0;
1347 ls->ls_recover_block = 0;
1348 ls->ls_recover_size = 0;
1349 ls->ls_recover_submit = NULL;
1350 ls->ls_recover_result = NULL;
1351 ls->ls_lvb_bits = NULL;
1352
1353 error = set_recover_size(sdp, NULL, 0);
1354 if (error)
1355 goto fail;
1356
1357 /*
1358 * prepare dlm_new_lockspace args
1359 */
1360
1361 fsname = strchr(table, ':');
1362 if (!fsname) {
1363 fs_info(sdp, "no fsname found\n");
1364 error = -EINVAL;
1365 goto fail_free;
1366 }
1367 memset(cluster, 0, sizeof(cluster));
1368 memcpy(cluster, table, strlen(table) - strlen(fsname));
1369 fsname++;
1370
1371 flags = DLM_LSFL_NEWEXCL;
1372
1373 /*
1374 * create/join lockspace
1375 */
1376
1377 init_rwsem(&ls->ls_sem);
1378 error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1379 &gdlm_lockspace_ops, sdp, &ops_result,
1380 &ls->ls_dlm);
1381 if (error) {
1382 fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1383 goto fail_free;
1384 }
1385
1386 if (ops_result < 0) {
1387 /*
1388 * dlm does not support ops callbacks,
1389 * old dlm_controld/gfs_controld are used, try without ops.
1390 */
1391 fs_info(sdp, "dlm lockspace ops not used\n");
1392 free_recover_size(ls);
1393 set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1394 return 0;
1395 }
1396
1397 if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1398 fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1399 error = -EINVAL;
1400 goto fail_release;
1401 }
1402
1403 /*
1404 * control_mount() uses control_lock to determine first mounter,
1405 * and for later mounts, waits for any recoveries to be cleared.
1406 */
1407
1408 error = control_mount(sdp);
1409 if (error) {
1410 fs_err(sdp, "mount control error %d\n", error);
1411 goto fail_release;
1412 }
1413
1414 ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1415 clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
1416 smp_mb__after_atomic();
1417 wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1418 return 0;
1419
1420 fail_release:
1421 dlm_release_lockspace(ls->ls_dlm, DLM_RELEASE_NORMAL);
1422 fail_free:
1423 free_recover_size(ls);
1424 fail:
1425 return error;
1426 }
1427
gdlm_first_done(struct gfs2_sbd * sdp)1428 static void gdlm_first_done(struct gfs2_sbd *sdp)
1429 {
1430 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1431 int error;
1432
1433 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1434 return;
1435
1436 error = control_first_done(sdp);
1437 if (error)
1438 fs_err(sdp, "mount first_done error %d\n", error);
1439 }
1440
gdlm_unmount(struct gfs2_sbd * sdp)1441 static void gdlm_unmount(struct gfs2_sbd *sdp)
1442 {
1443 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1444
1445 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1446 goto release;
1447
1448 /* wait for gfs2_control_wq to be done with this mount */
1449
1450 spin_lock(&ls->ls_recover_spin);
1451 set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1452 spin_unlock(&ls->ls_recover_spin);
1453 flush_delayed_work(&sdp->sd_control_work);
1454
1455 /* mounted_lock and control_lock will be purged in dlm recovery */
1456 release:
1457 down_write(&ls->ls_sem);
1458 if (ls->ls_dlm) {
1459 dlm_release_lockspace(ls->ls_dlm, DLM_RELEASE_NORMAL);
1460 ls->ls_dlm = NULL;
1461 }
1462 up_write(&ls->ls_sem);
1463
1464 free_recover_size(ls);
1465 }
1466
1467 static const match_table_t dlm_tokens = {
1468 { Opt_jid, "jid=%d"},
1469 { Opt_id, "id=%d"},
1470 { Opt_first, "first=%d"},
1471 { Opt_nodir, "nodir=%d"},
1472 { Opt_err, NULL },
1473 };
1474
1475 const struct lm_lockops gfs2_dlm_ops = {
1476 .lm_proto_name = "lock_dlm",
1477 .lm_mount = gdlm_mount,
1478 .lm_first_done = gdlm_first_done,
1479 .lm_recovery_result = gdlm_recovery_result,
1480 .lm_unmount = gdlm_unmount,
1481 .lm_put_lock = gdlm_put_lock,
1482 .lm_lock = gdlm_lock,
1483 .lm_cancel = gdlm_cancel,
1484 .lm_tokens = &dlm_tokens,
1485 };
1486
1487