1 /*
2 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
3 * Copyright 2004-2011 Red Hat, Inc.
4 *
5 * This copyrighted material is made available to anyone wishing to use,
6 * modify, copy, or redistribute it subject to the terms and conditions
7 * of the GNU General Public License version 2.
8 */
9
10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12 #include <linux/fs.h>
13 #include <linux/dlm.h>
14 #include <linux/slab.h>
15 #include <linux/types.h>
16 #include <linux/delay.h>
17 #include <linux/gfs2_ondisk.h>
18 #include <linux/sched/signal.h>
19
20 #include "incore.h"
21 #include "glock.h"
22 #include "util.h"
23 #include "sys.h"
24 #include "trace_gfs2.h"
25
26 /**
27 * gfs2_update_stats - Update time based stats
28 * @mv: Pointer to mean/variance structure to update
29 * @sample: New data to include
30 *
31 * @delta is the difference between the current rtt sample and the
32 * running average srtt. We add 1/8 of that to the srtt in order to
33 * update the current srtt estimate. The variance estimate is a bit
34 * more complicated. We subtract the current variance estimate from
35 * the abs value of the @delta and add 1/4 of that to the running
36 * total. That's equivalent to 3/4 of the current variance
37 * estimate plus 1/4 of the abs of @delta.
38 *
39 * Note that the index points at the array entry containing the smoothed
40 * mean value, and the variance is always in the following entry
41 *
42 * Reference: TCP/IP Illustrated, vol 2, p. 831,832
43 * All times are in units of integer nanoseconds. Unlike the TCP/IP case,
44 * they are not scaled fixed point.
45 */
46
gfs2_update_stats(struct gfs2_lkstats * s,unsigned index,s64 sample)47 static inline void gfs2_update_stats(struct gfs2_lkstats *s, unsigned index,
48 s64 sample)
49 {
50 s64 delta = sample - s->stats[index];
51 s->stats[index] += (delta >> 3);
52 index++;
53 s->stats[index] += (s64)(abs(delta) - s->stats[index]) >> 2;
54 }
55
56 /**
57 * gfs2_update_reply_times - Update locking statistics
58 * @gl: The glock to update
59 *
60 * This assumes that gl->gl_dstamp has been set earlier.
61 *
62 * The rtt (lock round trip time) is an estimate of the time
63 * taken to perform a dlm lock request. We update it on each
64 * reply from the dlm.
65 *
66 * The blocking flag is set on the glock for all dlm requests
67 * which may potentially block due to lock requests from other nodes.
68 * DLM requests where the current lock state is exclusive, the
69 * requested state is null (or unlocked) or where the TRY or
70 * TRY_1CB flags are set are classified as non-blocking. All
71 * other DLM requests are counted as (potentially) blocking.
72 */
gfs2_update_reply_times(struct gfs2_glock * gl)73 static inline void gfs2_update_reply_times(struct gfs2_glock *gl)
74 {
75 struct gfs2_pcpu_lkstats *lks;
76 const unsigned gltype = gl->gl_name.ln_type;
77 unsigned index = test_bit(GLF_BLOCKING, &gl->gl_flags) ?
78 GFS2_LKS_SRTTB : GFS2_LKS_SRTT;
79 s64 rtt;
80
81 preempt_disable();
82 rtt = ktime_to_ns(ktime_sub(ktime_get_real(), gl->gl_dstamp));
83 lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
84 gfs2_update_stats(&gl->gl_stats, index, rtt); /* Local */
85 gfs2_update_stats(&lks->lkstats[gltype], index, rtt); /* Global */
86 preempt_enable();
87
88 trace_gfs2_glock_lock_time(gl, rtt);
89 }
90
91 /**
92 * gfs2_update_request_times - Update locking statistics
93 * @gl: The glock to update
94 *
95 * The irt (lock inter-request times) measures the average time
96 * between requests to the dlm. It is updated immediately before
97 * each dlm call.
98 */
99
gfs2_update_request_times(struct gfs2_glock * gl)100 static inline void gfs2_update_request_times(struct gfs2_glock *gl)
101 {
102 struct gfs2_pcpu_lkstats *lks;
103 const unsigned gltype = gl->gl_name.ln_type;
104 ktime_t dstamp;
105 s64 irt;
106
107 preempt_disable();
108 dstamp = gl->gl_dstamp;
109 gl->gl_dstamp = ktime_get_real();
110 irt = ktime_to_ns(ktime_sub(gl->gl_dstamp, dstamp));
111 lks = this_cpu_ptr(gl->gl_name.ln_sbd->sd_lkstats);
112 gfs2_update_stats(&gl->gl_stats, GFS2_LKS_SIRT, irt); /* Local */
113 gfs2_update_stats(&lks->lkstats[gltype], GFS2_LKS_SIRT, irt); /* Global */
114 preempt_enable();
115 }
116
gdlm_ast(void * arg)117 static void gdlm_ast(void *arg)
118 {
119 struct gfs2_glock *gl = arg;
120 unsigned ret = gl->gl_state;
121
122 gfs2_update_reply_times(gl);
123 BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED);
124
125 if ((gl->gl_lksb.sb_flags & DLM_SBF_VALNOTVALID) && gl->gl_lksb.sb_lvbptr)
126 memset(gl->gl_lksb.sb_lvbptr, 0, GDLM_LVB_SIZE);
127
128 switch (gl->gl_lksb.sb_status) {
129 case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */
130 gfs2_glock_free(gl);
131 return;
132 case -DLM_ECANCEL: /* Cancel while getting lock */
133 ret |= LM_OUT_CANCELED;
134 goto out;
135 case -EAGAIN: /* Try lock fails */
136 case -EDEADLK: /* Deadlock detected */
137 goto out;
138 case -ETIMEDOUT: /* Canceled due to timeout */
139 ret |= LM_OUT_ERROR;
140 goto out;
141 case 0: /* Success */
142 break;
143 default: /* Something unexpected */
144 BUG();
145 }
146
147 ret = gl->gl_req;
148 if (gl->gl_lksb.sb_flags & DLM_SBF_ALTMODE) {
149 if (gl->gl_req == LM_ST_SHARED)
150 ret = LM_ST_DEFERRED;
151 else if (gl->gl_req == LM_ST_DEFERRED)
152 ret = LM_ST_SHARED;
153 else
154 BUG();
155 }
156
157 set_bit(GLF_INITIAL, &gl->gl_flags);
158 gfs2_glock_complete(gl, ret);
159 return;
160 out:
161 if (!test_bit(GLF_INITIAL, &gl->gl_flags))
162 gl->gl_lksb.sb_lkid = 0;
163 gfs2_glock_complete(gl, ret);
164 }
165
gdlm_bast(void * arg,int mode)166 static void gdlm_bast(void *arg, int mode)
167 {
168 struct gfs2_glock *gl = arg;
169
170 switch (mode) {
171 case DLM_LOCK_EX:
172 gfs2_glock_cb(gl, LM_ST_UNLOCKED);
173 break;
174 case DLM_LOCK_CW:
175 gfs2_glock_cb(gl, LM_ST_DEFERRED);
176 break;
177 case DLM_LOCK_PR:
178 gfs2_glock_cb(gl, LM_ST_SHARED);
179 break;
180 default:
181 pr_err("unknown bast mode %d\n", mode);
182 BUG();
183 }
184 }
185
186 /* convert gfs lock-state to dlm lock-mode */
187
make_mode(const unsigned int lmstate)188 static int make_mode(const unsigned int lmstate)
189 {
190 switch (lmstate) {
191 case LM_ST_UNLOCKED:
192 return DLM_LOCK_NL;
193 case LM_ST_EXCLUSIVE:
194 return DLM_LOCK_EX;
195 case LM_ST_DEFERRED:
196 return DLM_LOCK_CW;
197 case LM_ST_SHARED:
198 return DLM_LOCK_PR;
199 }
200 pr_err("unknown LM state %d\n", lmstate);
201 BUG();
202 return -1;
203 }
204
make_flags(struct gfs2_glock * gl,const unsigned int gfs_flags,const int req)205 static u32 make_flags(struct gfs2_glock *gl, const unsigned int gfs_flags,
206 const int req)
207 {
208 u32 lkf = 0;
209
210 if (gl->gl_lksb.sb_lvbptr)
211 lkf |= DLM_LKF_VALBLK;
212
213 if (gfs_flags & LM_FLAG_TRY)
214 lkf |= DLM_LKF_NOQUEUE;
215
216 if (gfs_flags & LM_FLAG_TRY_1CB) {
217 lkf |= DLM_LKF_NOQUEUE;
218 lkf |= DLM_LKF_NOQUEUEBAST;
219 }
220
221 if (gfs_flags & LM_FLAG_PRIORITY) {
222 lkf |= DLM_LKF_NOORDER;
223 lkf |= DLM_LKF_HEADQUE;
224 }
225
226 if (gfs_flags & LM_FLAG_ANY) {
227 if (req == DLM_LOCK_PR)
228 lkf |= DLM_LKF_ALTCW;
229 else if (req == DLM_LOCK_CW)
230 lkf |= DLM_LKF_ALTPR;
231 else
232 BUG();
233 }
234
235 if (gl->gl_lksb.sb_lkid != 0) {
236 lkf |= DLM_LKF_CONVERT;
237 if (test_bit(GLF_BLOCKING, &gl->gl_flags))
238 lkf |= DLM_LKF_QUECVT;
239 }
240
241 return lkf;
242 }
243
gfs2_reverse_hex(char * c,u64 value)244 static void gfs2_reverse_hex(char *c, u64 value)
245 {
246 *c = '0';
247 while (value) {
248 *c-- = hex_asc[value & 0x0f];
249 value >>= 4;
250 }
251 }
252
gdlm_lock(struct gfs2_glock * gl,unsigned int req_state,unsigned int flags)253 static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state,
254 unsigned int flags)
255 {
256 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
257 int req;
258 u32 lkf;
259 char strname[GDLM_STRNAME_BYTES] = "";
260
261 req = make_mode(req_state);
262 lkf = make_flags(gl, flags, req);
263 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
264 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
265 if (gl->gl_lksb.sb_lkid) {
266 gfs2_update_request_times(gl);
267 } else {
268 memset(strname, ' ', GDLM_STRNAME_BYTES - 1);
269 strname[GDLM_STRNAME_BYTES - 1] = '\0';
270 gfs2_reverse_hex(strname + 7, gl->gl_name.ln_type);
271 gfs2_reverse_hex(strname + 23, gl->gl_name.ln_number);
272 gl->gl_dstamp = ktime_get_real();
273 }
274 /*
275 * Submit the actual lock request.
276 */
277
278 return dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname,
279 GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast);
280 }
281
gdlm_put_lock(struct gfs2_glock * gl)282 static void gdlm_put_lock(struct gfs2_glock *gl)
283 {
284 struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
285 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
286 int error;
287
288 if (gl->gl_lksb.sb_lkid == 0) {
289 gfs2_glock_free(gl);
290 return;
291 }
292
293 clear_bit(GLF_BLOCKING, &gl->gl_flags);
294 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
295 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
296 gfs2_update_request_times(gl);
297
298 /* don't want to call dlm if we've unmounted the lock protocol */
299 if (test_bit(DFL_UNMOUNT, &ls->ls_recover_flags)) {
300 gfs2_glock_free(gl);
301 return;
302 }
303 /* don't want to skip dlm_unlock writing the lvb when lock has one */
304
305 if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
306 !gl->gl_lksb.sb_lvbptr) {
307 gfs2_glock_free(gl);
308 return;
309 }
310
311 error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_VALBLK,
312 NULL, gl);
313 if (error) {
314 pr_err("gdlm_unlock %x,%llx err=%d\n",
315 gl->gl_name.ln_type,
316 (unsigned long long)gl->gl_name.ln_number, error);
317 return;
318 }
319 }
320
gdlm_cancel(struct gfs2_glock * gl)321 static void gdlm_cancel(struct gfs2_glock *gl)
322 {
323 struct lm_lockstruct *ls = &gl->gl_name.ln_sbd->sd_lockstruct;
324 dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
325 }
326
327 /*
328 * dlm/gfs2 recovery coordination using dlm_recover callbacks
329 *
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
345 *
346 * - failures during recovery
347 *
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.
353 *
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.)
360 *
361 * - more specific gfs2 steps in sequence above
362 *
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.
372 *
373 * - parallel recovery steps across all nodes
374 *
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.)
379 *
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
386 *
387 * - is there a problem with clearing an lvb bit that should be set
388 * and missing a journal recovery?
389 *
390 * 1. jid fails
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
397 *
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
404 * step 1 is done.
405 *
406 * - special case of first mounter: first node to mount the fs
407 *
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:
413 *
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.
421 *
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.
429 *
430 * first mounter:
431 * control_lock EX/NOQUEUE success
432 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
433 * set first=1
434 * do first mounter recovery
435 * mounted_lock EX->PR
436 * control_lock EX->NL, write lvb generation
437 *
438 * other mounter:
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
444 * set first=0
445 *
446 * - mount during recovery
447 *
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.
451 *
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.
456 *
457 * - control_lock lvb format
458 *
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.
462 * 4 bytes unused.
463 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
464 * that jid N needs recovery.
465 */
466
467 #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
468
control_lvb_read(struct lm_lockstruct * ls,uint32_t * lvb_gen,char * lvb_bits)469 static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
470 char *lvb_bits)
471 {
472 __le32 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);
476 }
477
control_lvb_write(struct lm_lockstruct * ls,uint32_t lvb_gen,char * lvb_bits)478 static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
479 char *lvb_bits)
480 {
481 __le32 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));
485 }
486
all_jid_bits_clear(char * lvb)487 static int all_jid_bits_clear(char *lvb)
488 {
489 return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0,
490 GDLM_LVB_SIZE - JID_BITMAP_OFFSET);
491 }
492
sync_wait_cb(void * arg)493 static void sync_wait_cb(void *arg)
494 {
495 struct lm_lockstruct *ls = arg;
496 complete(&ls->ls_sync_wait);
497 }
498
sync_unlock(struct gfs2_sbd * sdp,struct dlm_lksb * lksb,char * name)499 static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
500 {
501 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
502 int error;
503
504 error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
505 if (error) {
506 fs_err(sdp, "%s lkid %x error %d\n",
507 name, lksb->sb_lkid, error);
508 return error;
509 }
510
511 wait_for_completion(&ls->ls_sync_wait);
512
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);
516 return -1;
517 }
518 return 0;
519 }
520
sync_lock(struct gfs2_sbd * sdp,int mode,uint32_t flags,unsigned int num,struct dlm_lksb * lksb,char * name)521 static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
522 unsigned int num, struct dlm_lksb *lksb, char *name)
523 {
524 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
525 char strname[GDLM_STRNAME_BYTES];
526 int error, status;
527
528 memset(strname, 0, GDLM_STRNAME_BYTES);
529 snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
530
531 error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
532 strname, GDLM_STRNAME_BYTES - 1,
533 0, sync_wait_cb, ls, NULL);
534 if (error) {
535 fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
536 name, lksb->sb_lkid, flags, mode, error);
537 return error;
538 }
539
540 wait_for_completion(&ls->ls_sync_wait);
541
542 status = lksb->sb_status;
543
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);
547 }
548
549 return status;
550 }
551
mounted_unlock(struct gfs2_sbd * sdp)552 static int mounted_unlock(struct gfs2_sbd *sdp)
553 {
554 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
555 return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
556 }
557
mounted_lock(struct gfs2_sbd * sdp,int mode,uint32_t flags)558 static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
559 {
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");
563 }
564
control_unlock(struct gfs2_sbd * sdp)565 static int control_unlock(struct gfs2_sbd *sdp)
566 {
567 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
568 return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
569 }
570
control_lock(struct gfs2_sbd * sdp,int mode,uint32_t flags)571 static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
572 {
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");
576 }
577
gfs2_control_func(struct work_struct * work)578 static void gfs2_control_func(struct work_struct *work)
579 {
580 struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
581 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
582 uint32_t block_gen, start_gen, lvb_gen, flags;
583 int recover_set = 0;
584 int write_lvb = 0;
585 int recover_size;
586 int i, error;
587
588 spin_lock(&ls->ls_recover_spin);
589 /*
590 * No MOUNT_DONE means we're still mounting; control_mount()
591 * will set this flag, after which this thread will take over
592 * all further clearing of BLOCK_LOCKS.
593 *
594 * FIRST_MOUNT means this node is doing first mounter recovery,
595 * for which recovery control is handled by
596 * control_mount()/control_first_done(), not this thread.
597 */
598 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
599 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
600 spin_unlock(&ls->ls_recover_spin);
601 return;
602 }
603 block_gen = ls->ls_recover_block;
604 start_gen = ls->ls_recover_start;
605 spin_unlock(&ls->ls_recover_spin);
606
607 /*
608 * Equal block_gen and start_gen implies we are between
609 * recover_prep and recover_done callbacks, which means
610 * dlm recovery is in progress and dlm locking is blocked.
611 * There's no point trying to do any work until recover_done.
612 */
613
614 if (block_gen == start_gen)
615 return;
616
617 /*
618 * Propagate recover_submit[] and recover_result[] to lvb:
619 * dlm_recoverd adds to recover_submit[] jids needing recovery
620 * gfs2_recover adds to recover_result[] journal recovery results
621 *
622 * set lvb bit for jids in recover_submit[] if the lvb has not
623 * yet been updated for the generation of the failure
624 *
625 * clear lvb bit for jids in recover_result[] if the result of
626 * the journal recovery is SUCCESS
627 */
628
629 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
630 if (error) {
631 fs_err(sdp, "control lock EX error %d\n", error);
632 return;
633 }
634
635 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
636
637 spin_lock(&ls->ls_recover_spin);
638 if (block_gen != ls->ls_recover_block ||
639 start_gen != ls->ls_recover_start) {
640 fs_info(sdp, "recover generation %u block1 %u %u\n",
641 start_gen, block_gen, ls->ls_recover_block);
642 spin_unlock(&ls->ls_recover_spin);
643 control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
644 return;
645 }
646
647 recover_size = ls->ls_recover_size;
648
649 if (lvb_gen <= start_gen) {
650 /*
651 * Clear lvb bits for jids we've successfully recovered.
652 * Because all nodes attempt to recover failed journals,
653 * a journal can be recovered multiple times successfully
654 * in succession. Only the first will really do recovery,
655 * the others find it clean, but still report a successful
656 * recovery. So, another node may have already recovered
657 * the jid and cleared the lvb bit for it.
658 */
659 for (i = 0; i < recover_size; i++) {
660 if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
661 continue;
662
663 ls->ls_recover_result[i] = 0;
664
665 if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
666 continue;
667
668 __clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
669 write_lvb = 1;
670 }
671 }
672
673 if (lvb_gen == start_gen) {
674 /*
675 * Failed slots before start_gen are already set in lvb.
676 */
677 for (i = 0; i < recover_size; i++) {
678 if (!ls->ls_recover_submit[i])
679 continue;
680 if (ls->ls_recover_submit[i] < lvb_gen)
681 ls->ls_recover_submit[i] = 0;
682 }
683 } else if (lvb_gen < start_gen) {
684 /*
685 * Failed slots before start_gen are not yet set in lvb.
686 */
687 for (i = 0; i < recover_size; i++) {
688 if (!ls->ls_recover_submit[i])
689 continue;
690 if (ls->ls_recover_submit[i] < start_gen) {
691 ls->ls_recover_submit[i] = 0;
692 __set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
693 }
694 }
695 /* even if there are no bits to set, we need to write the
696 latest generation to the lvb */
697 write_lvb = 1;
698 } else {
699 /*
700 * we should be getting a recover_done() for lvb_gen soon
701 */
702 }
703 spin_unlock(&ls->ls_recover_spin);
704
705 if (write_lvb) {
706 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
707 flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
708 } else {
709 flags = DLM_LKF_CONVERT;
710 }
711
712 error = control_lock(sdp, DLM_LOCK_NL, flags);
713 if (error) {
714 fs_err(sdp, "control lock NL error %d\n", error);
715 return;
716 }
717
718 /*
719 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
720 * and clear a jid bit in the lvb if the recovery is a success.
721 * Eventually all journals will be recovered, all jid bits will
722 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
723 */
724
725 for (i = 0; i < recover_size; i++) {
726 if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
727 fs_info(sdp, "recover generation %u jid %d\n",
728 start_gen, i);
729 gfs2_recover_set(sdp, i);
730 recover_set++;
731 }
732 }
733 if (recover_set)
734 return;
735
736 /*
737 * No more jid bits set in lvb, all recovery is done, unblock locks
738 * (unless a new recover_prep callback has occured blocking locks
739 * again while working above)
740 */
741
742 spin_lock(&ls->ls_recover_spin);
743 if (ls->ls_recover_block == block_gen &&
744 ls->ls_recover_start == start_gen) {
745 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
746 spin_unlock(&ls->ls_recover_spin);
747 fs_info(sdp, "recover generation %u done\n", start_gen);
748 gfs2_glock_thaw(sdp);
749 } else {
750 fs_info(sdp, "recover generation %u block2 %u %u\n",
751 start_gen, block_gen, ls->ls_recover_block);
752 spin_unlock(&ls->ls_recover_spin);
753 }
754 }
755
control_mount(struct gfs2_sbd * sdp)756 static int control_mount(struct gfs2_sbd *sdp)
757 {
758 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
759 uint32_t start_gen, block_gen, mount_gen, lvb_gen;
760 int mounted_mode;
761 int retries = 0;
762 int error;
763
764 memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
765 memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
766 memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
767 ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
768 init_completion(&ls->ls_sync_wait);
769
770 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
771
772 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
773 if (error) {
774 fs_err(sdp, "control_mount control_lock NL error %d\n", error);
775 return error;
776 }
777
778 error = mounted_lock(sdp, DLM_LOCK_NL, 0);
779 if (error) {
780 fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
781 control_unlock(sdp);
782 return error;
783 }
784 mounted_mode = DLM_LOCK_NL;
785
786 restart:
787 if (retries++ && signal_pending(current)) {
788 error = -EINTR;
789 goto fail;
790 }
791
792 /*
793 * We always start with both locks in NL. control_lock is
794 * demoted to NL below so we don't need to do it here.
795 */
796
797 if (mounted_mode != DLM_LOCK_NL) {
798 error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
799 if (error)
800 goto fail;
801 mounted_mode = DLM_LOCK_NL;
802 }
803
804 /*
805 * Other nodes need to do some work in dlm recovery and gfs2_control
806 * before the recover_done and control_lock will be ready for us below.
807 * A delay here is not required but often avoids having to retry.
808 */
809
810 msleep_interruptible(500);
811
812 /*
813 * Acquire control_lock in EX and mounted_lock in either EX or PR.
814 * control_lock lvb keeps track of any pending journal recoveries.
815 * mounted_lock indicates if any other nodes have the fs mounted.
816 */
817
818 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
819 if (error == -EAGAIN) {
820 goto restart;
821 } else if (error) {
822 fs_err(sdp, "control_mount control_lock EX error %d\n", error);
823 goto fail;
824 }
825
826 /**
827 * If we're a spectator, we don't want to take the lock in EX because
828 * we cannot do the first-mount responsibility it implies: recovery.
829 */
830 if (sdp->sd_args.ar_spectator)
831 goto locks_done;
832
833 error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
834 if (!error) {
835 mounted_mode = DLM_LOCK_EX;
836 goto locks_done;
837 } else if (error != -EAGAIN) {
838 fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
839 goto fail;
840 }
841
842 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
843 if (!error) {
844 mounted_mode = DLM_LOCK_PR;
845 goto locks_done;
846 } else {
847 /* not even -EAGAIN should happen here */
848 fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
849 goto fail;
850 }
851
852 locks_done:
853 /*
854 * If we got both locks above in EX, then we're the first mounter.
855 * If not, then we need to wait for the control_lock lvb to be
856 * updated by other mounted nodes to reflect our mount generation.
857 *
858 * In simple first mounter cases, first mounter will see zero lvb_gen,
859 * but in cases where all existing nodes leave/fail before mounting
860 * nodes finish control_mount, then all nodes will be mounting and
861 * lvb_gen will be non-zero.
862 */
863
864 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
865
866 if (lvb_gen == 0xFFFFFFFF) {
867 /* special value to force mount attempts to fail */
868 fs_err(sdp, "control_mount control_lock disabled\n");
869 error = -EINVAL;
870 goto fail;
871 }
872
873 if (mounted_mode == DLM_LOCK_EX) {
874 /* first mounter, keep both EX while doing first recovery */
875 spin_lock(&ls->ls_recover_spin);
876 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
877 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
878 set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
879 spin_unlock(&ls->ls_recover_spin);
880 fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
881 return 0;
882 }
883
884 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
885 if (error)
886 goto fail;
887
888 /*
889 * We are not first mounter, now we need to wait for the control_lock
890 * lvb generation to be >= the generation from our first recover_done
891 * and all lvb bits to be clear (no pending journal recoveries.)
892 */
893
894 if (!all_jid_bits_clear(ls->ls_lvb_bits)) {
895 /* journals need recovery, wait until all are clear */
896 fs_info(sdp, "control_mount wait for journal recovery\n");
897 goto restart;
898 }
899
900 spin_lock(&ls->ls_recover_spin);
901 block_gen = ls->ls_recover_block;
902 start_gen = ls->ls_recover_start;
903 mount_gen = ls->ls_recover_mount;
904
905 if (lvb_gen < mount_gen) {
906 /* wait for mounted nodes to update control_lock lvb to our
907 generation, which might include new recovery bits set */
908 if (sdp->sd_args.ar_spectator) {
909 fs_info(sdp, "Recovery is required. Waiting for a "
910 "non-spectator to mount.\n");
911 msleep_interruptible(1000);
912 } else {
913 fs_info(sdp, "control_mount wait1 block %u start %u "
914 "mount %u lvb %u flags %lx\n", block_gen,
915 start_gen, mount_gen, lvb_gen,
916 ls->ls_recover_flags);
917 }
918 spin_unlock(&ls->ls_recover_spin);
919 goto restart;
920 }
921
922 if (lvb_gen != start_gen) {
923 /* wait for mounted nodes to update control_lock lvb to the
924 latest recovery generation */
925 fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
926 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
927 lvb_gen, ls->ls_recover_flags);
928 spin_unlock(&ls->ls_recover_spin);
929 goto restart;
930 }
931
932 if (block_gen == start_gen) {
933 /* dlm recovery in progress, wait for it to finish */
934 fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
935 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
936 lvb_gen, ls->ls_recover_flags);
937 spin_unlock(&ls->ls_recover_spin);
938 goto restart;
939 }
940
941 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
942 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
943 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
944 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
945 spin_unlock(&ls->ls_recover_spin);
946 return 0;
947
948 fail:
949 mounted_unlock(sdp);
950 control_unlock(sdp);
951 return error;
952 }
953
control_first_done(struct gfs2_sbd * sdp)954 static int control_first_done(struct gfs2_sbd *sdp)
955 {
956 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
957 uint32_t start_gen, block_gen;
958 int error;
959
960 restart:
961 spin_lock(&ls->ls_recover_spin);
962 start_gen = ls->ls_recover_start;
963 block_gen = ls->ls_recover_block;
964
965 if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
966 !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
967 !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
968 /* sanity check, should not happen */
969 fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
970 start_gen, block_gen, ls->ls_recover_flags);
971 spin_unlock(&ls->ls_recover_spin);
972 control_unlock(sdp);
973 return -1;
974 }
975
976 if (start_gen == block_gen) {
977 /*
978 * Wait for the end of a dlm recovery cycle to switch from
979 * first mounter recovery. We can ignore any recover_slot
980 * callbacks between the recover_prep and next recover_done
981 * because we are still the first mounter and any failed nodes
982 * have not fully mounted, so they don't need recovery.
983 */
984 spin_unlock(&ls->ls_recover_spin);
985 fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
986
987 wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
988 TASK_UNINTERRUPTIBLE);
989 goto restart;
990 }
991
992 clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
993 set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
994 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
995 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
996 spin_unlock(&ls->ls_recover_spin);
997
998 memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE);
999 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
1000
1001 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
1002 if (error)
1003 fs_err(sdp, "control_first_done mounted PR error %d\n", error);
1004
1005 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
1006 if (error)
1007 fs_err(sdp, "control_first_done control NL error %d\n", error);
1008
1009 return error;
1010 }
1011
1012 /*
1013 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
1014 * to accomodate the largest slot number. (NB dlm slot numbers start at 1,
1015 * gfs2 jids start at 0, so jid = slot - 1)
1016 */
1017
1018 #define RECOVER_SIZE_INC 16
1019
set_recover_size(struct gfs2_sbd * sdp,struct dlm_slot * slots,int num_slots)1020 static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1021 int num_slots)
1022 {
1023 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1024 uint32_t *submit = NULL;
1025 uint32_t *result = NULL;
1026 uint32_t old_size, new_size;
1027 int i, max_jid;
1028
1029 if (!ls->ls_lvb_bits) {
1030 ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1031 if (!ls->ls_lvb_bits)
1032 return -ENOMEM;
1033 }
1034
1035 max_jid = 0;
1036 for (i = 0; i < num_slots; i++) {
1037 if (max_jid < slots[i].slot - 1)
1038 max_jid = slots[i].slot - 1;
1039 }
1040
1041 old_size = ls->ls_recover_size;
1042
1043 if (old_size >= max_jid + 1)
1044 return 0;
1045
1046 new_size = old_size + RECOVER_SIZE_INC;
1047
1048 submit = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1049 result = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1050 if (!submit || !result) {
1051 kfree(submit);
1052 kfree(result);
1053 return -ENOMEM;
1054 }
1055
1056 spin_lock(&ls->ls_recover_spin);
1057 memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1058 memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1059 kfree(ls->ls_recover_submit);
1060 kfree(ls->ls_recover_result);
1061 ls->ls_recover_submit = submit;
1062 ls->ls_recover_result = result;
1063 ls->ls_recover_size = new_size;
1064 spin_unlock(&ls->ls_recover_spin);
1065 return 0;
1066 }
1067
free_recover_size(struct lm_lockstruct * ls)1068 static void free_recover_size(struct lm_lockstruct *ls)
1069 {
1070 kfree(ls->ls_lvb_bits);
1071 kfree(ls->ls_recover_submit);
1072 kfree(ls->ls_recover_result);
1073 ls->ls_recover_submit = NULL;
1074 ls->ls_recover_result = NULL;
1075 ls->ls_recover_size = 0;
1076 ls->ls_lvb_bits = NULL;
1077 }
1078
1079 /* dlm calls before it does lock recovery */
1080
gdlm_recover_prep(void * arg)1081 static void gdlm_recover_prep(void *arg)
1082 {
1083 struct gfs2_sbd *sdp = arg;
1084 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1085
1086 spin_lock(&ls->ls_recover_spin);
1087 ls->ls_recover_block = ls->ls_recover_start;
1088 set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1089
1090 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1091 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1092 spin_unlock(&ls->ls_recover_spin);
1093 return;
1094 }
1095 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1096 spin_unlock(&ls->ls_recover_spin);
1097 }
1098
1099 /* dlm calls after recover_prep has been completed on all lockspace members;
1100 identifies slot/jid of failed member */
1101
gdlm_recover_slot(void * arg,struct dlm_slot * slot)1102 static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1103 {
1104 struct gfs2_sbd *sdp = arg;
1105 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1106 int jid = slot->slot - 1;
1107
1108 spin_lock(&ls->ls_recover_spin);
1109 if (ls->ls_recover_size < jid + 1) {
1110 fs_err(sdp, "recover_slot jid %d gen %u short size %d\n",
1111 jid, ls->ls_recover_block, ls->ls_recover_size);
1112 spin_unlock(&ls->ls_recover_spin);
1113 return;
1114 }
1115
1116 if (ls->ls_recover_submit[jid]) {
1117 fs_info(sdp, "recover_slot jid %d gen %u prev %u\n",
1118 jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1119 }
1120 ls->ls_recover_submit[jid] = ls->ls_recover_block;
1121 spin_unlock(&ls->ls_recover_spin);
1122 }
1123
1124 /* dlm calls after recover_slot and after it completes lock recovery */
1125
gdlm_recover_done(void * arg,struct dlm_slot * slots,int num_slots,int our_slot,uint32_t generation)1126 static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1127 int our_slot, uint32_t generation)
1128 {
1129 struct gfs2_sbd *sdp = arg;
1130 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1131
1132 /* ensure the ls jid arrays are large enough */
1133 set_recover_size(sdp, slots, num_slots);
1134
1135 spin_lock(&ls->ls_recover_spin);
1136 ls->ls_recover_start = generation;
1137
1138 if (!ls->ls_recover_mount) {
1139 ls->ls_recover_mount = generation;
1140 ls->ls_jid = our_slot - 1;
1141 }
1142
1143 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1144 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1145
1146 clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1147 smp_mb__after_atomic();
1148 wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1149 spin_unlock(&ls->ls_recover_spin);
1150 }
1151
1152 /* gfs2_recover thread has a journal recovery result */
1153
gdlm_recovery_result(struct gfs2_sbd * sdp,unsigned int jid,unsigned int result)1154 static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1155 unsigned int result)
1156 {
1157 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1158
1159 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1160 return;
1161
1162 /* don't care about the recovery of own journal during mount */
1163 if (jid == ls->ls_jid)
1164 return;
1165
1166 spin_lock(&ls->ls_recover_spin);
1167 if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1168 spin_unlock(&ls->ls_recover_spin);
1169 return;
1170 }
1171 if (ls->ls_recover_size < jid + 1) {
1172 fs_err(sdp, "recovery_result jid %d short size %d\n",
1173 jid, ls->ls_recover_size);
1174 spin_unlock(&ls->ls_recover_spin);
1175 return;
1176 }
1177
1178 fs_info(sdp, "recover jid %d result %s\n", jid,
1179 result == LM_RD_GAVEUP ? "busy" : "success");
1180
1181 ls->ls_recover_result[jid] = result;
1182
1183 /* GAVEUP means another node is recovering the journal; delay our
1184 next attempt to recover it, to give the other node a chance to
1185 finish before trying again */
1186
1187 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1188 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1189 result == LM_RD_GAVEUP ? HZ : 0);
1190 spin_unlock(&ls->ls_recover_spin);
1191 }
1192
1193 static const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1194 .recover_prep = gdlm_recover_prep,
1195 .recover_slot = gdlm_recover_slot,
1196 .recover_done = gdlm_recover_done,
1197 };
1198
gdlm_mount(struct gfs2_sbd * sdp,const char * table)1199 static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1200 {
1201 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1202 char cluster[GFS2_LOCKNAME_LEN];
1203 const char *fsname;
1204 uint32_t flags;
1205 int error, ops_result;
1206
1207 /*
1208 * initialize everything
1209 */
1210
1211 INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1212 spin_lock_init(&ls->ls_recover_spin);
1213 ls->ls_recover_flags = 0;
1214 ls->ls_recover_mount = 0;
1215 ls->ls_recover_start = 0;
1216 ls->ls_recover_block = 0;
1217 ls->ls_recover_size = 0;
1218 ls->ls_recover_submit = NULL;
1219 ls->ls_recover_result = NULL;
1220 ls->ls_lvb_bits = NULL;
1221
1222 error = set_recover_size(sdp, NULL, 0);
1223 if (error)
1224 goto fail;
1225
1226 /*
1227 * prepare dlm_new_lockspace args
1228 */
1229
1230 fsname = strchr(table, ':');
1231 if (!fsname) {
1232 fs_info(sdp, "no fsname found\n");
1233 error = -EINVAL;
1234 goto fail_free;
1235 }
1236 memset(cluster, 0, sizeof(cluster));
1237 memcpy(cluster, table, strlen(table) - strlen(fsname));
1238 fsname++;
1239
1240 flags = DLM_LSFL_FS | DLM_LSFL_NEWEXCL;
1241
1242 /*
1243 * create/join lockspace
1244 */
1245
1246 error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1247 &gdlm_lockspace_ops, sdp, &ops_result,
1248 &ls->ls_dlm);
1249 if (error) {
1250 fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1251 goto fail_free;
1252 }
1253
1254 if (ops_result < 0) {
1255 /*
1256 * dlm does not support ops callbacks,
1257 * old dlm_controld/gfs_controld are used, try without ops.
1258 */
1259 fs_info(sdp, "dlm lockspace ops not used\n");
1260 free_recover_size(ls);
1261 set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1262 return 0;
1263 }
1264
1265 if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1266 fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1267 error = -EINVAL;
1268 goto fail_release;
1269 }
1270
1271 /*
1272 * control_mount() uses control_lock to determine first mounter,
1273 * and for later mounts, waits for any recoveries to be cleared.
1274 */
1275
1276 error = control_mount(sdp);
1277 if (error) {
1278 fs_err(sdp, "mount control error %d\n", error);
1279 goto fail_release;
1280 }
1281
1282 ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1283 clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
1284 smp_mb__after_atomic();
1285 wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1286 return 0;
1287
1288 fail_release:
1289 dlm_release_lockspace(ls->ls_dlm, 2);
1290 fail_free:
1291 free_recover_size(ls);
1292 fail:
1293 return error;
1294 }
1295
gdlm_first_done(struct gfs2_sbd * sdp)1296 static void gdlm_first_done(struct gfs2_sbd *sdp)
1297 {
1298 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1299 int error;
1300
1301 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1302 return;
1303
1304 error = control_first_done(sdp);
1305 if (error)
1306 fs_err(sdp, "mount first_done error %d\n", error);
1307 }
1308
gdlm_unmount(struct gfs2_sbd * sdp)1309 static void gdlm_unmount(struct gfs2_sbd *sdp)
1310 {
1311 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1312
1313 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1314 goto release;
1315
1316 /* wait for gfs2_control_wq to be done with this mount */
1317
1318 spin_lock(&ls->ls_recover_spin);
1319 set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1320 spin_unlock(&ls->ls_recover_spin);
1321 flush_delayed_work(&sdp->sd_control_work);
1322
1323 /* mounted_lock and control_lock will be purged in dlm recovery */
1324 release:
1325 if (ls->ls_dlm) {
1326 dlm_release_lockspace(ls->ls_dlm, 2);
1327 ls->ls_dlm = NULL;
1328 }
1329
1330 free_recover_size(ls);
1331 }
1332
1333 static const match_table_t dlm_tokens = {
1334 { Opt_jid, "jid=%d"},
1335 { Opt_id, "id=%d"},
1336 { Opt_first, "first=%d"},
1337 { Opt_nodir, "nodir=%d"},
1338 { Opt_err, NULL },
1339 };
1340
1341 const struct lm_lockops gfs2_dlm_ops = {
1342 .lm_proto_name = "lock_dlm",
1343 .lm_mount = gdlm_mount,
1344 .lm_first_done = gdlm_first_done,
1345 .lm_recovery_result = gdlm_recovery_result,
1346 .lm_unmount = gdlm_unmount,
1347 .lm_put_lock = gdlm_put_lock,
1348 .lm_lock = gdlm_lock,
1349 .lm_cancel = gdlm_cancel,
1350 .lm_tokens = &dlm_tokens,
1351 };
1352
1353