blob: 641383a9c1bbf5a346315ad3a4805b2ffaf788b4 [file] [log] [blame]
Steven Whitehousef057f6c2009-01-12 10:43:39 +00001/*
2 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
David Teiglande0c2a9a2012-01-09 17:18:05 -05003 * Copyright 2004-2011 Red Hat, Inc.
Steven Whitehousef057f6c2009-01-12 10:43:39 +00004 *
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
Joe Perchesd77d1b52014-03-06 12:10:45 -080010#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
Steven Whitehousef057f6c2009-01-12 10:43:39 +000012#include <linux/fs.h>
13#include <linux/dlm.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090014#include <linux/slab.h>
Steven Whitehousef057f6c2009-01-12 10:43:39 +000015#include <linux/types.h>
David Teiglande0c2a9a2012-01-09 17:18:05 -050016#include <linux/delay.h>
Steven Whitehousef057f6c2009-01-12 10:43:39 +000017#include <linux/gfs2_ondisk.h>
18
19#include "incore.h"
20#include "glock.h"
21#include "util.h"
David Teiglande0c2a9a2012-01-09 17:18:05 -050022#include "sys.h"
Steven Whitehousea2457692012-01-20 10:38:36 +000023#include "trace_gfs2.h"
Steven Whitehousef057f6c2009-01-12 10:43:39 +000024
David Teiglande0c2a9a2012-01-09 17:18:05 -050025extern struct workqueue_struct *gfs2_control_wq;
Steven Whitehousef057f6c2009-01-12 10:43:39 +000026
Steven Whitehousea2457692012-01-20 10:38:36 +000027/**
28 * gfs2_update_stats - Update time based stats
29 * @mv: Pointer to mean/variance structure to update
30 * @sample: New data to include
31 *
32 * @delta is the difference between the current rtt sample and the
33 * running average srtt. We add 1/8 of that to the srtt in order to
34 * update the current srtt estimate. The varience estimate is a bit
35 * more complicated. We subtract the abs value of the @delta from
36 * the current variance estimate and add 1/4 of that to the running
37 * total.
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
47static 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] += ((abs64(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 */
73static 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_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
100static 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_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
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000117static void gdlm_ast(void *arg)
118{
119 struct gfs2_glock *gl = arg;
120 unsigned ret = gl->gl_state;
121
Steven Whitehousea2457692012-01-20 10:38:36 +0000122 gfs2_update_reply_times(gl);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000123 BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED);
124
David Teigland4e2f8842012-11-14 13:47:37 -0500125 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);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000127
128 switch (gl->gl_lksb.sb_status) {
129 case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */
Steven Whitehousefc0e38d2011-03-09 10:58:04 +0000130 gfs2_glock_free(gl);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000131 return;
132 case -DLM_ECANCEL: /* Cancel while getting lock */
133 ret |= LM_OUT_CANCELED;
134 goto out;
135 case -EAGAIN: /* Try lock fails */
Steven Whitehouse1fea7c22010-09-08 10:09:25 +0100136 case -EDEADLK: /* Deadlock detected */
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000137 goto out;
Steven Whitehouse1fea7c22010-09-08 10:09:25 +0100138 case -ETIMEDOUT: /* Canceled due to timeout */
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000139 ret |= LM_OUT_ERROR;
140 goto out;
141 case 0: /* Success */
142 break;
143 default: /* Something unexpected */
144 BUG();
145 }
146
Benjamin Marzinski02ffad082009-03-06 10:03:20 -0600147 ret = gl->gl_req;
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000148 if (gl->gl_lksb.sb_flags & DLM_SBF_ALTMODE) {
Benjamin Marzinski02ffad082009-03-06 10:03:20 -0600149 if (gl->gl_req == LM_ST_SHARED)
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000150 ret = LM_ST_DEFERRED;
Benjamin Marzinski02ffad082009-03-06 10:03:20 -0600151 else if (gl->gl_req == LM_ST_DEFERRED)
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000152 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;
160out:
161 if (!test_bit(GLF_INITIAL, &gl->gl_flags))
162 gl->gl_lksb.sb_lkid = 0;
163 gfs2_glock_complete(gl, ret);
164}
165
166static 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:
Joe Perchesd77d1b52014-03-06 12:10:45 -0800181 pr_err("unknown bast mode %d\n", mode);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000182 BUG();
183 }
184}
185
186/* convert gfs lock-state to dlm lock-mode */
187
188static 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 }
Joe Perchesd77d1b52014-03-06 12:10:45 -0800200 pr_err("unknown LM state %d\n", lmstate);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000201 BUG();
202 return -1;
203}
204
Bob Peterson4c569a72012-04-10 14:45:24 -0400205static u32 make_flags(struct gfs2_glock *gl, const unsigned int gfs_flags,
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000206 const int req)
207{
David Teiglanddba2d702012-11-14 13:46:53 -0500208 u32 lkf = 0;
209
David Teigland4e2f8842012-11-14 13:47:37 -0500210 if (gl->gl_lksb.sb_lvbptr)
David Teiglanddba2d702012-11-14 13:46:53 -0500211 lkf |= DLM_LKF_VALBLK;
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000212
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
David Teiglanddba2d702012-11-14 13:46:53 -0500235 if (gl->gl_lksb.sb_lkid != 0) {
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000236 lkf |= DLM_LKF_CONVERT;
Bob Peterson4c569a72012-04-10 14:45:24 -0400237 if (test_bit(GLF_BLOCKING, &gl->gl_flags))
238 lkf |= DLM_LKF_QUECVT;
239 }
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000240
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000241 return lkf;
242}
243
Steven Whitehousea2457692012-01-20 10:38:36 +0000244static void gfs2_reverse_hex(char *c, u64 value)
245{
Nathan Strazec148752012-12-11 17:01:24 -0500246 *c = '0';
Steven Whitehousea2457692012-01-20 10:38:36 +0000247 while (value) {
248 *c-- = hex_asc[value & 0x0f];
249 value >>= 4;
250 }
251}
252
Steven Whitehouse921169c2010-11-29 12:50:38 +0000253static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state,
254 unsigned int flags)
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000255{
256 struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000257 int req;
258 u32 lkf;
Steven Whitehousea2457692012-01-20 10:38:36 +0000259 char strname[GDLM_STRNAME_BYTES] = "";
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000260
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000261 req = make_mode(req_state);
Bob Peterson4c569a72012-04-10 14:45:24 -0400262 lkf = make_flags(gl, flags, req);
Steven Whitehousea2457692012-01-20 10:38:36 +0000263 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 }
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000274 /*
275 * Submit the actual lock request.
276 */
277
Steven Whitehousea2457692012-01-20 10:38:36 +0000278 return dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname,
Steven Whitehouse921169c2010-11-29 12:50:38 +0000279 GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000280}
281
Steven Whitehousebc015cb2011-01-19 09:30:01 +0000282static void gdlm_put_lock(struct gfs2_glock *gl)
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000283{
Steven Whitehousee4027462010-01-25 11:20:19 +0000284 struct gfs2_sbd *sdp = gl->gl_sbd;
285 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
David Teiglandd4e0bfe2013-01-03 17:52:07 -0500286 int lvb_needs_unlock = 0;
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000287 int error;
288
289 if (gl->gl_lksb.sb_lkid == 0) {
Steven Whitehousefc0e38d2011-03-09 10:58:04 +0000290 gfs2_glock_free(gl);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000291 return;
292 }
293
Steven Whitehousea2457692012-01-20 10:38:36 +0000294 clear_bit(GLF_BLOCKING, &gl->gl_flags);
295 gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT);
296 gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT);
297 gfs2_update_request_times(gl);
David Teiglandfb6791d2012-11-13 10:58:56 -0500298
299 /* don't want to skip dlm_unlock writing the lvb when lock is ex */
David Teiglandd4e0bfe2013-01-03 17:52:07 -0500300
301 if (gl->gl_lksb.sb_lvbptr && (gl->gl_state == LM_ST_EXCLUSIVE))
302 lvb_needs_unlock = 1;
303
David Teiglandfb6791d2012-11-13 10:58:56 -0500304 if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) &&
David Teiglandd4e0bfe2013-01-03 17:52:07 -0500305 !lvb_needs_unlock) {
David Teiglandfb6791d2012-11-13 10:58:56 -0500306 gfs2_glock_free(gl);
307 return;
308 }
309
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000310 error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_VALBLK,
311 NULL, gl);
312 if (error) {
Joe Perchesd77d1b52014-03-06 12:10:45 -0800313 pr_err("gdlm_unlock %x,%llx err=%d\n",
314 gl->gl_name.ln_type,
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000315 (unsigned long long)gl->gl_name.ln_number, error);
316 return;
317 }
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000318}
319
320static void gdlm_cancel(struct gfs2_glock *gl)
321{
322 struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct;
323 dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl);
324}
325
David Teiglande0c2a9a2012-01-09 17:18:05 -0500326/*
327 * dlm/gfs2 recovery coordination using dlm_recover callbacks
328 *
329 * 1. dlm_controld sees lockspace members change
330 * 2. dlm_controld blocks dlm-kernel locking activity
331 * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep)
332 * 4. dlm_controld starts and finishes its own user level recovery
333 * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery
334 * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot)
335 * 7. dlm_recoverd does its own lock recovery
336 * 8. dlm_recoverd unblocks dlm-kernel locking activity
337 * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation)
338 * 10. gfs2_control updates control_lock lvb with new generation and jid bits
339 * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none)
340 * 12. gfs2_recover dequeues and recovers journals of failed nodes
341 * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result)
342 * 14. gfs2_control updates control_lock lvb jid bits for recovered journals
343 * 15. gfs2_control unblocks normal locking when all journals are recovered
344 *
345 * - failures during recovery
346 *
347 * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control
348 * clears BLOCK_LOCKS (step 15), e.g. another node fails while still
349 * recovering for a prior failure. gfs2_control needs a way to detect
350 * this so it can leave BLOCK_LOCKS set in step 15. This is managed using
351 * the recover_block and recover_start values.
352 *
353 * recover_done() provides a new lockspace generation number each time it
354 * is called (step 9). This generation number is saved as recover_start.
355 * When recover_prep() is called, it sets BLOCK_LOCKS and sets
356 * recover_block = recover_start. So, while recover_block is equal to
357 * recover_start, BLOCK_LOCKS should remain set. (recover_spin must
358 * be held around the BLOCK_LOCKS/recover_block/recover_start logic.)
359 *
360 * - more specific gfs2 steps in sequence above
361 *
362 * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start
363 * 6. recover_slot records any failed jids (maybe none)
364 * 9. recover_done sets recover_start = new generation number
365 * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids
366 * 12. gfs2_recover does journal recoveries for failed jids identified above
367 * 14. gfs2_control clears control_lock lvb bits for recovered jids
368 * 15. gfs2_control checks if recover_block == recover_start (step 3 occured
369 * again) then do nothing, otherwise if recover_start > recover_block
370 * then clear BLOCK_LOCKS.
371 *
372 * - parallel recovery steps across all nodes
373 *
374 * All nodes attempt to update the control_lock lvb with the new generation
375 * number and jid bits, but only the first to get the control_lock EX will
376 * do so; others will see that it's already done (lvb already contains new
377 * generation number.)
378 *
379 * . All nodes get the same recover_prep/recover_slot/recover_done callbacks
380 * . All nodes attempt to set control_lock lvb gen + bits for the new gen
381 * . One node gets control_lock first and writes the lvb, others see it's done
382 * . All nodes attempt to recover jids for which they see control_lock bits set
383 * . One node succeeds for a jid, and that one clears the jid bit in the lvb
384 * . All nodes will eventually see all lvb bits clear and unblock locks
385 *
386 * - is there a problem with clearing an lvb bit that should be set
387 * and missing a journal recovery?
388 *
389 * 1. jid fails
390 * 2. lvb bit set for step 1
391 * 3. jid recovered for step 1
392 * 4. jid taken again (new mount)
393 * 5. jid fails (for step 4)
394 * 6. lvb bit set for step 5 (will already be set)
395 * 7. lvb bit cleared for step 3
396 *
397 * This is not a problem because the failure in step 5 does not
398 * require recovery, because the mount in step 4 could not have
399 * progressed far enough to unblock locks and access the fs. The
400 * control_mount() function waits for all recoveries to be complete
401 * for the latest lockspace generation before ever unblocking locks
402 * and returning. The mount in step 4 waits until the recovery in
403 * step 1 is done.
404 *
405 * - special case of first mounter: first node to mount the fs
406 *
407 * The first node to mount a gfs2 fs needs to check all the journals
408 * and recover any that need recovery before other nodes are allowed
409 * to mount the fs. (Others may begin mounting, but they must wait
410 * for the first mounter to be done before taking locks on the fs
411 * or accessing the fs.) This has two parts:
412 *
413 * 1. The mounted_lock tells a node it's the first to mount the fs.
414 * Each node holds the mounted_lock in PR while it's mounted.
415 * Each node tries to acquire the mounted_lock in EX when it mounts.
416 * If a node is granted the mounted_lock EX it means there are no
417 * other mounted nodes (no PR locks exist), and it is the first mounter.
418 * The mounted_lock is demoted to PR when first recovery is done, so
419 * others will fail to get an EX lock, but will get a PR lock.
420 *
421 * 2. The control_lock blocks others in control_mount() while the first
422 * mounter is doing first mount recovery of all journals.
423 * A mounting node needs to acquire control_lock in EX mode before
424 * it can proceed. The first mounter holds control_lock in EX while doing
425 * the first mount recovery, blocking mounts from other nodes, then demotes
426 * control_lock to NL when it's done (others_may_mount/first_done),
427 * allowing other nodes to continue mounting.
428 *
429 * first mounter:
430 * control_lock EX/NOQUEUE success
431 * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters)
432 * set first=1
433 * do first mounter recovery
434 * mounted_lock EX->PR
435 * control_lock EX->NL, write lvb generation
436 *
437 * other mounter:
438 * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry)
439 * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR)
440 * mounted_lock PR/NOQUEUE success
441 * read lvb generation
442 * control_lock EX->NL
443 * set first=0
444 *
445 * - mount during recovery
446 *
447 * If a node mounts while others are doing recovery (not first mounter),
448 * the mounting node will get its initial recover_done() callback without
449 * having seen any previous failures/callbacks.
450 *
451 * It must wait for all recoveries preceding its mount to be finished
452 * before it unblocks locks. It does this by repeating the "other mounter"
453 * steps above until the lvb generation number is >= its mount generation
454 * number (from initial recover_done) and all lvb bits are clear.
455 *
456 * - control_lock lvb format
457 *
458 * 4 bytes generation number: the latest dlm lockspace generation number
459 * from recover_done callback. Indicates the jid bitmap has been updated
460 * to reflect all slot failures through that generation.
461 * 4 bytes unused.
462 * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates
463 * that jid N needs recovery.
464 */
465
466#define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */
467
468static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen,
469 char *lvb_bits)
470{
Al Viro951b4bd2013-06-02 19:53:40 -0400471 __le32 gen;
David Teiglande0c2a9a2012-01-09 17:18:05 -0500472 memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE);
Al Viro951b4bd2013-06-02 19:53:40 -0400473 memcpy(&gen, lvb_bits, sizeof(__le32));
David Teiglande0c2a9a2012-01-09 17:18:05 -0500474 *lvb_gen = le32_to_cpu(gen);
475}
476
477static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen,
478 char *lvb_bits)
479{
Al Viro951b4bd2013-06-02 19:53:40 -0400480 __le32 gen;
David Teiglande0c2a9a2012-01-09 17:18:05 -0500481 memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE);
482 gen = cpu_to_le32(lvb_gen);
Al Viro951b4bd2013-06-02 19:53:40 -0400483 memcpy(ls->ls_control_lvb, &gen, sizeof(__le32));
David Teiglande0c2a9a2012-01-09 17:18:05 -0500484}
485
486static int all_jid_bits_clear(char *lvb)
487{
Akinobu Mita4146c3d2013-03-07 23:42:52 +0900488 return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0,
489 GDLM_LVB_SIZE - JID_BITMAP_OFFSET);
David Teiglande0c2a9a2012-01-09 17:18:05 -0500490}
491
492static void sync_wait_cb(void *arg)
493{
494 struct lm_lockstruct *ls = arg;
495 complete(&ls->ls_sync_wait);
496}
497
498static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name)
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000499{
500 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
501 int error;
502
David Teiglande0c2a9a2012-01-09 17:18:05 -0500503 error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls);
504 if (error) {
505 fs_err(sdp, "%s lkid %x error %d\n",
506 name, lksb->sb_lkid, error);
507 return error;
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000508 }
509
David Teiglande0c2a9a2012-01-09 17:18:05 -0500510 wait_for_completion(&ls->ls_sync_wait);
511
512 if (lksb->sb_status != -DLM_EUNLOCK) {
513 fs_err(sdp, "%s lkid %x status %d\n",
514 name, lksb->sb_lkid, lksb->sb_status);
515 return -1;
516 }
517 return 0;
518}
519
520static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags,
521 unsigned int num, struct dlm_lksb *lksb, char *name)
522{
523 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
524 char strname[GDLM_STRNAME_BYTES];
525 int error, status;
526
527 memset(strname, 0, GDLM_STRNAME_BYTES);
528 snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num);
529
530 error = dlm_lock(ls->ls_dlm, mode, lksb, flags,
531 strname, GDLM_STRNAME_BYTES - 1,
532 0, sync_wait_cb, ls, NULL);
533 if (error) {
534 fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n",
535 name, lksb->sb_lkid, flags, mode, error);
536 return error;
537 }
538
539 wait_for_completion(&ls->ls_sync_wait);
540
541 status = lksb->sb_status;
542
543 if (status && status != -EAGAIN) {
544 fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n",
545 name, lksb->sb_lkid, flags, mode, status);
546 }
547
548 return status;
549}
550
551static int mounted_unlock(struct gfs2_sbd *sdp)
552{
553 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
554 return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock");
555}
556
557static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
558{
559 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
560 return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK,
561 &ls->ls_mounted_lksb, "mounted_lock");
562}
563
564static int control_unlock(struct gfs2_sbd *sdp)
565{
566 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
567 return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock");
568}
569
570static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags)
571{
572 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
573 return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK,
574 &ls->ls_control_lksb, "control_lock");
575}
576
577static void gfs2_control_func(struct work_struct *work)
578{
579 struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work);
580 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
David Teiglande0c2a9a2012-01-09 17:18:05 -0500581 uint32_t block_gen, start_gen, lvb_gen, flags;
582 int recover_set = 0;
583 int write_lvb = 0;
584 int recover_size;
585 int i, error;
586
587 spin_lock(&ls->ls_recover_spin);
588 /*
589 * No MOUNT_DONE means we're still mounting; control_mount()
590 * will set this flag, after which this thread will take over
591 * all further clearing of BLOCK_LOCKS.
592 *
593 * FIRST_MOUNT means this node is doing first mounter recovery,
594 * for which recovery control is handled by
595 * control_mount()/control_first_done(), not this thread.
596 */
597 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
598 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
599 spin_unlock(&ls->ls_recover_spin);
600 return;
601 }
602 block_gen = ls->ls_recover_block;
603 start_gen = ls->ls_recover_start;
604 spin_unlock(&ls->ls_recover_spin);
605
606 /*
607 * Equal block_gen and start_gen implies we are between
608 * recover_prep and recover_done callbacks, which means
609 * dlm recovery is in progress and dlm locking is blocked.
610 * There's no point trying to do any work until recover_done.
611 */
612
613 if (block_gen == start_gen)
614 return;
615
616 /*
617 * Propagate recover_submit[] and recover_result[] to lvb:
618 * dlm_recoverd adds to recover_submit[] jids needing recovery
619 * gfs2_recover adds to recover_result[] journal recovery results
620 *
621 * set lvb bit for jids in recover_submit[] if the lvb has not
622 * yet been updated for the generation of the failure
623 *
624 * clear lvb bit for jids in recover_result[] if the result of
625 * the journal recovery is SUCCESS
626 */
627
628 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
629 if (error) {
630 fs_err(sdp, "control lock EX error %d\n", error);
631 return;
632 }
633
David Teigland57c73102013-03-05 16:01:47 -0500634 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
David Teiglande0c2a9a2012-01-09 17:18:05 -0500635
636 spin_lock(&ls->ls_recover_spin);
637 if (block_gen != ls->ls_recover_block ||
638 start_gen != ls->ls_recover_start) {
639 fs_info(sdp, "recover generation %u block1 %u %u\n",
640 start_gen, block_gen, ls->ls_recover_block);
641 spin_unlock(&ls->ls_recover_spin);
642 control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
643 return;
644 }
645
646 recover_size = ls->ls_recover_size;
647
648 if (lvb_gen <= start_gen) {
649 /*
650 * Clear lvb bits for jids we've successfully recovered.
651 * Because all nodes attempt to recover failed journals,
652 * a journal can be recovered multiple times successfully
653 * in succession. Only the first will really do recovery,
654 * the others find it clean, but still report a successful
655 * recovery. So, another node may have already recovered
656 * the jid and cleared the lvb bit for it.
657 */
658 for (i = 0; i < recover_size; i++) {
659 if (ls->ls_recover_result[i] != LM_RD_SUCCESS)
660 continue;
661
662 ls->ls_recover_result[i] = 0;
663
David Teigland57c73102013-03-05 16:01:47 -0500664 if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET))
David Teiglande0c2a9a2012-01-09 17:18:05 -0500665 continue;
666
David Teigland57c73102013-03-05 16:01:47 -0500667 __clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
David Teiglande0c2a9a2012-01-09 17:18:05 -0500668 write_lvb = 1;
669 }
670 }
671
672 if (lvb_gen == start_gen) {
673 /*
674 * Failed slots before start_gen are already set in lvb.
675 */
676 for (i = 0; i < recover_size; i++) {
677 if (!ls->ls_recover_submit[i])
678 continue;
679 if (ls->ls_recover_submit[i] < lvb_gen)
680 ls->ls_recover_submit[i] = 0;
681 }
682 } else if (lvb_gen < start_gen) {
683 /*
684 * Failed slots before start_gen are not yet set in lvb.
685 */
686 for (i = 0; i < recover_size; i++) {
687 if (!ls->ls_recover_submit[i])
688 continue;
689 if (ls->ls_recover_submit[i] < start_gen) {
690 ls->ls_recover_submit[i] = 0;
David Teigland57c73102013-03-05 16:01:47 -0500691 __set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET);
David Teiglande0c2a9a2012-01-09 17:18:05 -0500692 }
693 }
694 /* even if there are no bits to set, we need to write the
695 latest generation to the lvb */
696 write_lvb = 1;
697 } else {
698 /*
699 * we should be getting a recover_done() for lvb_gen soon
700 */
701 }
702 spin_unlock(&ls->ls_recover_spin);
703
704 if (write_lvb) {
David Teigland57c73102013-03-05 16:01:47 -0500705 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
David Teiglande0c2a9a2012-01-09 17:18:05 -0500706 flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK;
707 } else {
708 flags = DLM_LKF_CONVERT;
709 }
710
711 error = control_lock(sdp, DLM_LOCK_NL, flags);
712 if (error) {
713 fs_err(sdp, "control lock NL error %d\n", error);
714 return;
715 }
716
717 /*
718 * Everyone will see jid bits set in the lvb, run gfs2_recover_set(),
719 * and clear a jid bit in the lvb if the recovery is a success.
720 * Eventually all journals will be recovered, all jid bits will
721 * be cleared in the lvb, and everyone will clear BLOCK_LOCKS.
722 */
723
724 for (i = 0; i < recover_size; i++) {
David Teigland57c73102013-03-05 16:01:47 -0500725 if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) {
David Teiglande0c2a9a2012-01-09 17:18:05 -0500726 fs_info(sdp, "recover generation %u jid %d\n",
727 start_gen, i);
728 gfs2_recover_set(sdp, i);
729 recover_set++;
730 }
731 }
732 if (recover_set)
733 return;
734
735 /*
736 * No more jid bits set in lvb, all recovery is done, unblock locks
737 * (unless a new recover_prep callback has occured blocking locks
738 * again while working above)
739 */
740
741 spin_lock(&ls->ls_recover_spin);
742 if (ls->ls_recover_block == block_gen &&
743 ls->ls_recover_start == start_gen) {
744 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
745 spin_unlock(&ls->ls_recover_spin);
746 fs_info(sdp, "recover generation %u done\n", start_gen);
747 gfs2_glock_thaw(sdp);
748 } else {
749 fs_info(sdp, "recover generation %u block2 %u %u\n",
750 start_gen, block_gen, ls->ls_recover_block);
751 spin_unlock(&ls->ls_recover_spin);
752 }
753}
754
755static int control_mount(struct gfs2_sbd *sdp)
756{
757 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
David Teiglande0c2a9a2012-01-09 17:18:05 -0500758 uint32_t start_gen, block_gen, mount_gen, lvb_gen;
759 int mounted_mode;
760 int retries = 0;
761 int error;
762
763 memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb));
764 memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb));
765 memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE);
766 ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb;
767 init_completion(&ls->ls_sync_wait);
768
769 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
770
771 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK);
772 if (error) {
773 fs_err(sdp, "control_mount control_lock NL error %d\n", error);
774 return error;
775 }
776
777 error = mounted_lock(sdp, DLM_LOCK_NL, 0);
778 if (error) {
779 fs_err(sdp, "control_mount mounted_lock NL error %d\n", error);
780 control_unlock(sdp);
781 return error;
782 }
783 mounted_mode = DLM_LOCK_NL;
784
785restart:
786 if (retries++ && signal_pending(current)) {
787 error = -EINTR;
788 goto fail;
789 }
790
791 /*
792 * We always start with both locks in NL. control_lock is
793 * demoted to NL below so we don't need to do it here.
794 */
795
796 if (mounted_mode != DLM_LOCK_NL) {
797 error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
798 if (error)
799 goto fail;
800 mounted_mode = DLM_LOCK_NL;
801 }
802
803 /*
804 * Other nodes need to do some work in dlm recovery and gfs2_control
805 * before the recover_done and control_lock will be ready for us below.
806 * A delay here is not required but often avoids having to retry.
807 */
808
809 msleep_interruptible(500);
810
811 /*
812 * Acquire control_lock in EX and mounted_lock in either EX or PR.
813 * control_lock lvb keeps track of any pending journal recoveries.
814 * mounted_lock indicates if any other nodes have the fs mounted.
815 */
816
817 error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK);
818 if (error == -EAGAIN) {
819 goto restart;
820 } else if (error) {
821 fs_err(sdp, "control_mount control_lock EX error %d\n", error);
822 goto fail;
823 }
824
825 error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
826 if (!error) {
827 mounted_mode = DLM_LOCK_EX;
828 goto locks_done;
829 } else if (error != -EAGAIN) {
830 fs_err(sdp, "control_mount mounted_lock EX error %d\n", error);
831 goto fail;
832 }
833
834 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE);
835 if (!error) {
836 mounted_mode = DLM_LOCK_PR;
837 goto locks_done;
838 } else {
839 /* not even -EAGAIN should happen here */
840 fs_err(sdp, "control_mount mounted_lock PR error %d\n", error);
841 goto fail;
842 }
843
844locks_done:
845 /*
846 * If we got both locks above in EX, then we're the first mounter.
847 * If not, then we need to wait for the control_lock lvb to be
848 * updated by other mounted nodes to reflect our mount generation.
849 *
850 * In simple first mounter cases, first mounter will see zero lvb_gen,
851 * but in cases where all existing nodes leave/fail before mounting
852 * nodes finish control_mount, then all nodes will be mounting and
853 * lvb_gen will be non-zero.
854 */
855
David Teigland57c73102013-03-05 16:01:47 -0500856 control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits);
David Teiglande0c2a9a2012-01-09 17:18:05 -0500857
858 if (lvb_gen == 0xFFFFFFFF) {
859 /* special value to force mount attempts to fail */
860 fs_err(sdp, "control_mount control_lock disabled\n");
861 error = -EINVAL;
862 goto fail;
863 }
864
865 if (mounted_mode == DLM_LOCK_EX) {
866 /* first mounter, keep both EX while doing first recovery */
867 spin_lock(&ls->ls_recover_spin);
868 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
869 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
870 set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
871 spin_unlock(&ls->ls_recover_spin);
872 fs_info(sdp, "first mounter control generation %u\n", lvb_gen);
873 return 0;
874 }
875
876 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000877 if (error)
David Teiglande0c2a9a2012-01-09 17:18:05 -0500878 goto fail;
879
880 /*
881 * We are not first mounter, now we need to wait for the control_lock
882 * lvb generation to be >= the generation from our first recover_done
883 * and all lvb bits to be clear (no pending journal recoveries.)
884 */
885
David Teigland57c73102013-03-05 16:01:47 -0500886 if (!all_jid_bits_clear(ls->ls_lvb_bits)) {
David Teiglande0c2a9a2012-01-09 17:18:05 -0500887 /* journals need recovery, wait until all are clear */
888 fs_info(sdp, "control_mount wait for journal recovery\n");
889 goto restart;
890 }
891
892 spin_lock(&ls->ls_recover_spin);
893 block_gen = ls->ls_recover_block;
894 start_gen = ls->ls_recover_start;
895 mount_gen = ls->ls_recover_mount;
896
897 if (lvb_gen < mount_gen) {
898 /* wait for mounted nodes to update control_lock lvb to our
899 generation, which might include new recovery bits set */
900 fs_info(sdp, "control_mount wait1 block %u start %u mount %u "
901 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
902 lvb_gen, ls->ls_recover_flags);
903 spin_unlock(&ls->ls_recover_spin);
904 goto restart;
905 }
906
907 if (lvb_gen != start_gen) {
908 /* wait for mounted nodes to update control_lock lvb to the
909 latest recovery generation */
910 fs_info(sdp, "control_mount wait2 block %u start %u mount %u "
911 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
912 lvb_gen, ls->ls_recover_flags);
913 spin_unlock(&ls->ls_recover_spin);
914 goto restart;
915 }
916
917 if (block_gen == start_gen) {
918 /* dlm recovery in progress, wait for it to finish */
919 fs_info(sdp, "control_mount wait3 block %u start %u mount %u "
920 "lvb %u flags %lx\n", block_gen, start_gen, mount_gen,
921 lvb_gen, ls->ls_recover_flags);
922 spin_unlock(&ls->ls_recover_spin);
923 goto restart;
924 }
925
926 clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
927 set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags);
928 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
929 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
930 spin_unlock(&ls->ls_recover_spin);
931 return 0;
932
933fail:
934 mounted_unlock(sdp);
935 control_unlock(sdp);
936 return error;
937}
938
David Teiglande0c2a9a2012-01-09 17:18:05 -0500939static int control_first_done(struct gfs2_sbd *sdp)
940{
941 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
David Teiglande0c2a9a2012-01-09 17:18:05 -0500942 uint32_t start_gen, block_gen;
943 int error;
944
945restart:
946 spin_lock(&ls->ls_recover_spin);
947 start_gen = ls->ls_recover_start;
948 block_gen = ls->ls_recover_block;
949
950 if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) ||
951 !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
952 !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
953 /* sanity check, should not happen */
954 fs_err(sdp, "control_first_done start %u block %u flags %lx\n",
955 start_gen, block_gen, ls->ls_recover_flags);
956 spin_unlock(&ls->ls_recover_spin);
957 control_unlock(sdp);
958 return -1;
959 }
960
961 if (start_gen == block_gen) {
962 /*
963 * Wait for the end of a dlm recovery cycle to switch from
964 * first mounter recovery. We can ignore any recover_slot
965 * callbacks between the recover_prep and next recover_done
966 * because we are still the first mounter and any failed nodes
967 * have not fully mounted, so they don't need recovery.
968 */
969 spin_unlock(&ls->ls_recover_spin);
970 fs_info(sdp, "control_first_done wait gen %u\n", start_gen);
971
972 wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY,
NeilBrown74316202014-07-07 15:16:04 +1000973 TASK_UNINTERRUPTIBLE);
David Teiglande0c2a9a2012-01-09 17:18:05 -0500974 goto restart;
975 }
976
977 clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
978 set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags);
979 memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t));
980 memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t));
981 spin_unlock(&ls->ls_recover_spin);
982
David Teigland57c73102013-03-05 16:01:47 -0500983 memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE);
984 control_lvb_write(ls, start_gen, ls->ls_lvb_bits);
David Teiglande0c2a9a2012-01-09 17:18:05 -0500985
986 error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT);
987 if (error)
988 fs_err(sdp, "control_first_done mounted PR error %d\n", error);
989
990 error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK);
991 if (error)
992 fs_err(sdp, "control_first_done control NL error %d\n", error);
Steven Whitehousef057f6c2009-01-12 10:43:39 +0000993
994 return error;
995}
996
David Teiglande0c2a9a2012-01-09 17:18:05 -0500997/*
998 * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC)
999 * to accomodate the largest slot number. (NB dlm slot numbers start at 1,
1000 * gfs2 jids start at 0, so jid = slot - 1)
1001 */
1002
1003#define RECOVER_SIZE_INC 16
1004
1005static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots,
1006 int num_slots)
1007{
1008 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1009 uint32_t *submit = NULL;
1010 uint32_t *result = NULL;
1011 uint32_t old_size, new_size;
1012 int i, max_jid;
1013
David Teigland57c73102013-03-05 16:01:47 -05001014 if (!ls->ls_lvb_bits) {
1015 ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS);
1016 if (!ls->ls_lvb_bits)
1017 return -ENOMEM;
1018 }
1019
David Teiglande0c2a9a2012-01-09 17:18:05 -05001020 max_jid = 0;
1021 for (i = 0; i < num_slots; i++) {
1022 if (max_jid < slots[i].slot - 1)
1023 max_jid = slots[i].slot - 1;
1024 }
1025
1026 old_size = ls->ls_recover_size;
1027
1028 if (old_size >= max_jid + 1)
1029 return 0;
1030
1031 new_size = old_size + RECOVER_SIZE_INC;
1032
Fabian Frederick6ec43b12014-06-25 20:40:45 +02001033 submit = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
1034 result = kcalloc(new_size, sizeof(uint32_t), GFP_NOFS);
David Teiglande0c2a9a2012-01-09 17:18:05 -05001035 if (!submit || !result) {
1036 kfree(submit);
1037 kfree(result);
1038 return -ENOMEM;
1039 }
1040
1041 spin_lock(&ls->ls_recover_spin);
1042 memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t));
1043 memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t));
1044 kfree(ls->ls_recover_submit);
1045 kfree(ls->ls_recover_result);
1046 ls->ls_recover_submit = submit;
1047 ls->ls_recover_result = result;
1048 ls->ls_recover_size = new_size;
1049 spin_unlock(&ls->ls_recover_spin);
1050 return 0;
1051}
1052
1053static void free_recover_size(struct lm_lockstruct *ls)
1054{
David Teigland57c73102013-03-05 16:01:47 -05001055 kfree(ls->ls_lvb_bits);
David Teiglande0c2a9a2012-01-09 17:18:05 -05001056 kfree(ls->ls_recover_submit);
1057 kfree(ls->ls_recover_result);
1058 ls->ls_recover_submit = NULL;
1059 ls->ls_recover_result = NULL;
1060 ls->ls_recover_size = 0;
1061}
1062
1063/* dlm calls before it does lock recovery */
1064
1065static void gdlm_recover_prep(void *arg)
1066{
1067 struct gfs2_sbd *sdp = arg;
1068 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1069
1070 spin_lock(&ls->ls_recover_spin);
1071 ls->ls_recover_block = ls->ls_recover_start;
1072 set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
1073
1074 if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) ||
1075 test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1076 spin_unlock(&ls->ls_recover_spin);
1077 return;
1078 }
1079 set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags);
1080 spin_unlock(&ls->ls_recover_spin);
1081}
1082
1083/* dlm calls after recover_prep has been completed on all lockspace members;
1084 identifies slot/jid of failed member */
1085
1086static void gdlm_recover_slot(void *arg, struct dlm_slot *slot)
1087{
1088 struct gfs2_sbd *sdp = arg;
1089 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1090 int jid = slot->slot - 1;
1091
1092 spin_lock(&ls->ls_recover_spin);
1093 if (ls->ls_recover_size < jid + 1) {
1094 fs_err(sdp, "recover_slot jid %d gen %u short size %d",
1095 jid, ls->ls_recover_block, ls->ls_recover_size);
1096 spin_unlock(&ls->ls_recover_spin);
1097 return;
1098 }
1099
1100 if (ls->ls_recover_submit[jid]) {
David Teiglandad781972014-02-14 12:10:19 -06001101 fs_info(sdp, "recover_slot jid %d gen %u prev %u\n",
David Teiglande0c2a9a2012-01-09 17:18:05 -05001102 jid, ls->ls_recover_block, ls->ls_recover_submit[jid]);
1103 }
1104 ls->ls_recover_submit[jid] = ls->ls_recover_block;
1105 spin_unlock(&ls->ls_recover_spin);
1106}
1107
1108/* dlm calls after recover_slot and after it completes lock recovery */
1109
1110static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots,
1111 int our_slot, uint32_t generation)
1112{
1113 struct gfs2_sbd *sdp = arg;
1114 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1115
1116 /* ensure the ls jid arrays are large enough */
1117 set_recover_size(sdp, slots, num_slots);
1118
1119 spin_lock(&ls->ls_recover_spin);
1120 ls->ls_recover_start = generation;
1121
1122 if (!ls->ls_recover_mount) {
1123 ls->ls_recover_mount = generation;
1124 ls->ls_jid = our_slot - 1;
1125 }
1126
1127 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1128 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0);
1129
1130 clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags);
Peter Zijlstra4e857c52014-03-17 18:06:10 +01001131 smp_mb__after_atomic();
David Teiglande0c2a9a2012-01-09 17:18:05 -05001132 wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY);
1133 spin_unlock(&ls->ls_recover_spin);
1134}
1135
1136/* gfs2_recover thread has a journal recovery result */
1137
1138static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid,
1139 unsigned int result)
1140{
1141 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1142
1143 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1144 return;
1145
1146 /* don't care about the recovery of own journal during mount */
1147 if (jid == ls->ls_jid)
1148 return;
1149
1150 spin_lock(&ls->ls_recover_spin);
1151 if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) {
1152 spin_unlock(&ls->ls_recover_spin);
1153 return;
1154 }
1155 if (ls->ls_recover_size < jid + 1) {
1156 fs_err(sdp, "recovery_result jid %d short size %d",
1157 jid, ls->ls_recover_size);
1158 spin_unlock(&ls->ls_recover_spin);
1159 return;
1160 }
1161
1162 fs_info(sdp, "recover jid %d result %s\n", jid,
1163 result == LM_RD_GAVEUP ? "busy" : "success");
1164
1165 ls->ls_recover_result[jid] = result;
1166
1167 /* GAVEUP means another node is recovering the journal; delay our
1168 next attempt to recover it, to give the other node a chance to
1169 finish before trying again */
1170
1171 if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags))
1172 queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work,
1173 result == LM_RD_GAVEUP ? HZ : 0);
1174 spin_unlock(&ls->ls_recover_spin);
1175}
1176
1177const struct dlm_lockspace_ops gdlm_lockspace_ops = {
1178 .recover_prep = gdlm_recover_prep,
1179 .recover_slot = gdlm_recover_slot,
1180 .recover_done = gdlm_recover_done,
1181};
1182
1183static int gdlm_mount(struct gfs2_sbd *sdp, const char *table)
1184{
1185 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1186 char cluster[GFS2_LOCKNAME_LEN];
1187 const char *fsname;
1188 uint32_t flags;
1189 int error, ops_result;
1190
1191 /*
1192 * initialize everything
1193 */
1194
1195 INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func);
1196 spin_lock_init(&ls->ls_recover_spin);
1197 ls->ls_recover_flags = 0;
1198 ls->ls_recover_mount = 0;
1199 ls->ls_recover_start = 0;
1200 ls->ls_recover_block = 0;
1201 ls->ls_recover_size = 0;
1202 ls->ls_recover_submit = NULL;
1203 ls->ls_recover_result = NULL;
David Teigland57c73102013-03-05 16:01:47 -05001204 ls->ls_lvb_bits = NULL;
David Teiglande0c2a9a2012-01-09 17:18:05 -05001205
1206 error = set_recover_size(sdp, NULL, 0);
1207 if (error)
1208 goto fail;
1209
1210 /*
1211 * prepare dlm_new_lockspace args
1212 */
1213
1214 fsname = strchr(table, ':');
1215 if (!fsname) {
1216 fs_info(sdp, "no fsname found\n");
1217 error = -EINVAL;
1218 goto fail_free;
1219 }
1220 memset(cluster, 0, sizeof(cluster));
1221 memcpy(cluster, table, strlen(table) - strlen(fsname));
1222 fsname++;
1223
1224 flags = DLM_LSFL_FS | DLM_LSFL_NEWEXCL;
David Teiglande0c2a9a2012-01-09 17:18:05 -05001225
1226 /*
1227 * create/join lockspace
1228 */
1229
1230 error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE,
1231 &gdlm_lockspace_ops, sdp, &ops_result,
1232 &ls->ls_dlm);
1233 if (error) {
1234 fs_err(sdp, "dlm_new_lockspace error %d\n", error);
1235 goto fail_free;
1236 }
1237
1238 if (ops_result < 0) {
1239 /*
1240 * dlm does not support ops callbacks,
1241 * old dlm_controld/gfs_controld are used, try without ops.
1242 */
1243 fs_info(sdp, "dlm lockspace ops not used\n");
1244 free_recover_size(ls);
1245 set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags);
1246 return 0;
1247 }
1248
1249 if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) {
1250 fs_err(sdp, "dlm lockspace ops disallow jid preset\n");
1251 error = -EINVAL;
1252 goto fail_release;
1253 }
1254
1255 /*
1256 * control_mount() uses control_lock to determine first mounter,
1257 * and for later mounts, waits for any recoveries to be cleared.
1258 */
1259
1260 error = control_mount(sdp);
1261 if (error) {
1262 fs_err(sdp, "mount control error %d\n", error);
1263 goto fail_release;
1264 }
1265
1266 ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags);
1267 clear_bit(SDF_NOJOURNALID, &sdp->sd_flags);
Peter Zijlstra4e857c52014-03-17 18:06:10 +01001268 smp_mb__after_atomic();
David Teiglande0c2a9a2012-01-09 17:18:05 -05001269 wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID);
1270 return 0;
1271
1272fail_release:
1273 dlm_release_lockspace(ls->ls_dlm, 2);
1274fail_free:
1275 free_recover_size(ls);
1276fail:
1277 return error;
1278}
1279
1280static void gdlm_first_done(struct gfs2_sbd *sdp)
1281{
1282 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1283 int error;
1284
1285 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1286 return;
1287
1288 error = control_first_done(sdp);
1289 if (error)
1290 fs_err(sdp, "mount first_done error %d\n", error);
1291}
1292
Steven Whitehousef057f6c2009-01-12 10:43:39 +00001293static void gdlm_unmount(struct gfs2_sbd *sdp)
1294{
1295 struct lm_lockstruct *ls = &sdp->sd_lockstruct;
1296
David Teiglande0c2a9a2012-01-09 17:18:05 -05001297 if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags))
1298 goto release;
1299
1300 /* wait for gfs2_control_wq to be done with this mount */
1301
1302 spin_lock(&ls->ls_recover_spin);
1303 set_bit(DFL_UNMOUNT, &ls->ls_recover_flags);
1304 spin_unlock(&ls->ls_recover_spin);
Tejun Heo43829732012-08-20 14:51:24 -07001305 flush_delayed_work(&sdp->sd_control_work);
David Teiglande0c2a9a2012-01-09 17:18:05 -05001306
1307 /* mounted_lock and control_lock will be purged in dlm recovery */
1308release:
Steven Whitehousef057f6c2009-01-12 10:43:39 +00001309 if (ls->ls_dlm) {
1310 dlm_release_lockspace(ls->ls_dlm, 2);
1311 ls->ls_dlm = NULL;
1312 }
David Teiglande0c2a9a2012-01-09 17:18:05 -05001313
1314 free_recover_size(ls);
Steven Whitehousef057f6c2009-01-12 10:43:39 +00001315}
1316
1317static const match_table_t dlm_tokens = {
1318 { Opt_jid, "jid=%d"},
1319 { Opt_id, "id=%d"},
1320 { Opt_first, "first=%d"},
1321 { Opt_nodir, "nodir=%d"},
1322 { Opt_err, NULL },
1323};
1324
1325const struct lm_lockops gfs2_dlm_ops = {
1326 .lm_proto_name = "lock_dlm",
1327 .lm_mount = gdlm_mount,
David Teiglande0c2a9a2012-01-09 17:18:05 -05001328 .lm_first_done = gdlm_first_done,
1329 .lm_recovery_result = gdlm_recovery_result,
Steven Whitehousef057f6c2009-01-12 10:43:39 +00001330 .lm_unmount = gdlm_unmount,
1331 .lm_put_lock = gdlm_put_lock,
1332 .lm_lock = gdlm_lock,
1333 .lm_cancel = gdlm_cancel,
1334 .lm_tokens = &dlm_tokens,
1335};
1336