]>
Commit | Line | Data |
---|---|---|
f057f6cd SW |
1 | /* |
2 | * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved. | |
e0c2a9aa | 3 | * Copyright 2004-2011 Red Hat, Inc. |
f057f6cd SW |
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 | #include <linux/fs.h> | |
11 | #include <linux/dlm.h> | |
5a0e3ad6 | 12 | #include <linux/slab.h> |
f057f6cd | 13 | #include <linux/types.h> |
e0c2a9aa | 14 | #include <linux/delay.h> |
f057f6cd SW |
15 | #include <linux/gfs2_ondisk.h> |
16 | ||
17 | #include "incore.h" | |
18 | #include "glock.h" | |
19 | #include "util.h" | |
e0c2a9aa | 20 | #include "sys.h" |
a245769f | 21 | #include "trace_gfs2.h" |
f057f6cd | 22 | |
e0c2a9aa | 23 | extern struct workqueue_struct *gfs2_control_wq; |
f057f6cd | 24 | |
a245769f SW |
25 | /** |
26 | * gfs2_update_stats - Update time based stats | |
27 | * @mv: Pointer to mean/variance structure to update | |
28 | * @sample: New data to include | |
29 | * | |
30 | * @delta is the difference between the current rtt sample and the | |
31 | * running average srtt. We add 1/8 of that to the srtt in order to | |
32 | * update the current srtt estimate. The varience estimate is a bit | |
33 | * more complicated. We subtract the abs value of the @delta from | |
34 | * the current variance estimate and add 1/4 of that to the running | |
35 | * total. | |
36 | * | |
37 | * Note that the index points at the array entry containing the smoothed | |
38 | * mean value, and the variance is always in the following entry | |
39 | * | |
40 | * Reference: TCP/IP Illustrated, vol 2, p. 831,832 | |
41 | * All times are in units of integer nanoseconds. Unlike the TCP/IP case, | |
42 | * they are not scaled fixed point. | |
43 | */ | |
44 | ||
45 | static inline void gfs2_update_stats(struct gfs2_lkstats *s, unsigned index, | |
46 | s64 sample) | |
47 | { | |
48 | s64 delta = sample - s->stats[index]; | |
49 | s->stats[index] += (delta >> 3); | |
50 | index++; | |
51 | s->stats[index] += ((abs64(delta) - s->stats[index]) >> 2); | |
52 | } | |
53 | ||
54 | /** | |
55 | * gfs2_update_reply_times - Update locking statistics | |
56 | * @gl: The glock to update | |
57 | * | |
58 | * This assumes that gl->gl_dstamp has been set earlier. | |
59 | * | |
60 | * The rtt (lock round trip time) is an estimate of the time | |
61 | * taken to perform a dlm lock request. We update it on each | |
62 | * reply from the dlm. | |
63 | * | |
64 | * The blocking flag is set on the glock for all dlm requests | |
65 | * which may potentially block due to lock requests from other nodes. | |
66 | * DLM requests where the current lock state is exclusive, the | |
67 | * requested state is null (or unlocked) or where the TRY or | |
68 | * TRY_1CB flags are set are classified as non-blocking. All | |
69 | * other DLM requests are counted as (potentially) blocking. | |
70 | */ | |
71 | static inline void gfs2_update_reply_times(struct gfs2_glock *gl) | |
72 | { | |
73 | struct gfs2_pcpu_lkstats *lks; | |
74 | const unsigned gltype = gl->gl_name.ln_type; | |
75 | unsigned index = test_bit(GLF_BLOCKING, &gl->gl_flags) ? | |
76 | GFS2_LKS_SRTTB : GFS2_LKS_SRTT; | |
77 | s64 rtt; | |
78 | ||
79 | preempt_disable(); | |
80 | rtt = ktime_to_ns(ktime_sub(ktime_get_real(), gl->gl_dstamp)); | |
81 | lks = this_cpu_ptr(gl->gl_sbd->sd_lkstats); | |
82 | gfs2_update_stats(&gl->gl_stats, index, rtt); /* Local */ | |
83 | gfs2_update_stats(&lks->lkstats[gltype], index, rtt); /* Global */ | |
84 | preempt_enable(); | |
85 | ||
86 | trace_gfs2_glock_lock_time(gl, rtt); | |
87 | } | |
88 | ||
89 | /** | |
90 | * gfs2_update_request_times - Update locking statistics | |
91 | * @gl: The glock to update | |
92 | * | |
93 | * The irt (lock inter-request times) measures the average time | |
94 | * between requests to the dlm. It is updated immediately before | |
95 | * each dlm call. | |
96 | */ | |
97 | ||
98 | static inline void gfs2_update_request_times(struct gfs2_glock *gl) | |
99 | { | |
100 | struct gfs2_pcpu_lkstats *lks; | |
101 | const unsigned gltype = gl->gl_name.ln_type; | |
102 | ktime_t dstamp; | |
103 | s64 irt; | |
104 | ||
105 | preempt_disable(); | |
106 | dstamp = gl->gl_dstamp; | |
107 | gl->gl_dstamp = ktime_get_real(); | |
108 | irt = ktime_to_ns(ktime_sub(gl->gl_dstamp, dstamp)); | |
109 | lks = this_cpu_ptr(gl->gl_sbd->sd_lkstats); | |
110 | gfs2_update_stats(&gl->gl_stats, GFS2_LKS_SIRT, irt); /* Local */ | |
111 | gfs2_update_stats(&lks->lkstats[gltype], GFS2_LKS_SIRT, irt); /* Global */ | |
112 | preempt_enable(); | |
113 | } | |
114 | ||
f057f6cd SW |
115 | static void gdlm_ast(void *arg) |
116 | { | |
117 | struct gfs2_glock *gl = arg; | |
118 | unsigned ret = gl->gl_state; | |
119 | ||
a245769f | 120 | gfs2_update_reply_times(gl); |
f057f6cd SW |
121 | BUG_ON(gl->gl_lksb.sb_flags & DLM_SBF_DEMOTED); |
122 | ||
4e2f8849 DT |
123 | if ((gl->gl_lksb.sb_flags & DLM_SBF_VALNOTVALID) && gl->gl_lksb.sb_lvbptr) |
124 | memset(gl->gl_lksb.sb_lvbptr, 0, GDLM_LVB_SIZE); | |
f057f6cd SW |
125 | |
126 | switch (gl->gl_lksb.sb_status) { | |
127 | case -DLM_EUNLOCK: /* Unlocked, so glock can be freed */ | |
fc0e38da | 128 | gfs2_glock_free(gl); |
f057f6cd SW |
129 | return; |
130 | case -DLM_ECANCEL: /* Cancel while getting lock */ | |
131 | ret |= LM_OUT_CANCELED; | |
132 | goto out; | |
133 | case -EAGAIN: /* Try lock fails */ | |
1fea7c25 | 134 | case -EDEADLK: /* Deadlock detected */ |
f057f6cd | 135 | goto out; |
1fea7c25 | 136 | case -ETIMEDOUT: /* Canceled due to timeout */ |
f057f6cd SW |
137 | ret |= LM_OUT_ERROR; |
138 | goto out; | |
139 | case 0: /* Success */ | |
140 | break; | |
141 | default: /* Something unexpected */ | |
142 | BUG(); | |
143 | } | |
144 | ||
02ffad08 | 145 | ret = gl->gl_req; |
f057f6cd | 146 | if (gl->gl_lksb.sb_flags & DLM_SBF_ALTMODE) { |
02ffad08 | 147 | if (gl->gl_req == LM_ST_SHARED) |
f057f6cd | 148 | ret = LM_ST_DEFERRED; |
02ffad08 | 149 | else if (gl->gl_req == LM_ST_DEFERRED) |
f057f6cd SW |
150 | ret = LM_ST_SHARED; |
151 | else | |
152 | BUG(); | |
153 | } | |
154 | ||
155 | set_bit(GLF_INITIAL, &gl->gl_flags); | |
156 | gfs2_glock_complete(gl, ret); | |
157 | return; | |
158 | out: | |
159 | if (!test_bit(GLF_INITIAL, &gl->gl_flags)) | |
160 | gl->gl_lksb.sb_lkid = 0; | |
161 | gfs2_glock_complete(gl, ret); | |
162 | } | |
163 | ||
164 | static void gdlm_bast(void *arg, int mode) | |
165 | { | |
166 | struct gfs2_glock *gl = arg; | |
167 | ||
168 | switch (mode) { | |
169 | case DLM_LOCK_EX: | |
170 | gfs2_glock_cb(gl, LM_ST_UNLOCKED); | |
171 | break; | |
172 | case DLM_LOCK_CW: | |
173 | gfs2_glock_cb(gl, LM_ST_DEFERRED); | |
174 | break; | |
175 | case DLM_LOCK_PR: | |
176 | gfs2_glock_cb(gl, LM_ST_SHARED); | |
177 | break; | |
178 | default: | |
179 | printk(KERN_ERR "unknown bast mode %d", mode); | |
180 | BUG(); | |
181 | } | |
182 | } | |
183 | ||
184 | /* convert gfs lock-state to dlm lock-mode */ | |
185 | ||
186 | static int make_mode(const unsigned int lmstate) | |
187 | { | |
188 | switch (lmstate) { | |
189 | case LM_ST_UNLOCKED: | |
190 | return DLM_LOCK_NL; | |
191 | case LM_ST_EXCLUSIVE: | |
192 | return DLM_LOCK_EX; | |
193 | case LM_ST_DEFERRED: | |
194 | return DLM_LOCK_CW; | |
195 | case LM_ST_SHARED: | |
196 | return DLM_LOCK_PR; | |
197 | } | |
198 | printk(KERN_ERR "unknown LM state %d", lmstate); | |
199 | BUG(); | |
200 | return -1; | |
201 | } | |
202 | ||
4c569a72 | 203 | static u32 make_flags(struct gfs2_glock *gl, const unsigned int gfs_flags, |
f057f6cd SW |
204 | const int req) |
205 | { | |
dba2d70c DT |
206 | u32 lkf = 0; |
207 | ||
4e2f8849 | 208 | if (gl->gl_lksb.sb_lvbptr) |
dba2d70c | 209 | lkf |= DLM_LKF_VALBLK; |
f057f6cd SW |
210 | |
211 | if (gfs_flags & LM_FLAG_TRY) | |
212 | lkf |= DLM_LKF_NOQUEUE; | |
213 | ||
214 | if (gfs_flags & LM_FLAG_TRY_1CB) { | |
215 | lkf |= DLM_LKF_NOQUEUE; | |
216 | lkf |= DLM_LKF_NOQUEUEBAST; | |
217 | } | |
218 | ||
219 | if (gfs_flags & LM_FLAG_PRIORITY) { | |
220 | lkf |= DLM_LKF_NOORDER; | |
221 | lkf |= DLM_LKF_HEADQUE; | |
222 | } | |
223 | ||
224 | if (gfs_flags & LM_FLAG_ANY) { | |
225 | if (req == DLM_LOCK_PR) | |
226 | lkf |= DLM_LKF_ALTCW; | |
227 | else if (req == DLM_LOCK_CW) | |
228 | lkf |= DLM_LKF_ALTPR; | |
229 | else | |
230 | BUG(); | |
231 | } | |
232 | ||
dba2d70c | 233 | if (gl->gl_lksb.sb_lkid != 0) { |
f057f6cd | 234 | lkf |= DLM_LKF_CONVERT; |
4c569a72 BP |
235 | if (test_bit(GLF_BLOCKING, &gl->gl_flags)) |
236 | lkf |= DLM_LKF_QUECVT; | |
237 | } | |
f057f6cd | 238 | |
f057f6cd SW |
239 | return lkf; |
240 | } | |
241 | ||
a245769f SW |
242 | static void gfs2_reverse_hex(char *c, u64 value) |
243 | { | |
ec148752 | 244 | *c = '0'; |
a245769f SW |
245 | while (value) { |
246 | *c-- = hex_asc[value & 0x0f]; | |
247 | value >>= 4; | |
248 | } | |
249 | } | |
250 | ||
921169ca SW |
251 | static int gdlm_lock(struct gfs2_glock *gl, unsigned int req_state, |
252 | unsigned int flags) | |
f057f6cd SW |
253 | { |
254 | struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct; | |
f057f6cd SW |
255 | int req; |
256 | u32 lkf; | |
a245769f | 257 | char strname[GDLM_STRNAME_BYTES] = ""; |
f057f6cd SW |
258 | |
259 | req = make_mode(req_state); | |
4c569a72 | 260 | lkf = make_flags(gl, flags, req); |
a245769f SW |
261 | gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT); |
262 | gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT); | |
263 | if (gl->gl_lksb.sb_lkid) { | |
264 | gfs2_update_request_times(gl); | |
265 | } else { | |
266 | memset(strname, ' ', GDLM_STRNAME_BYTES - 1); | |
267 | strname[GDLM_STRNAME_BYTES - 1] = '\0'; | |
268 | gfs2_reverse_hex(strname + 7, gl->gl_name.ln_type); | |
269 | gfs2_reverse_hex(strname + 23, gl->gl_name.ln_number); | |
270 | gl->gl_dstamp = ktime_get_real(); | |
271 | } | |
f057f6cd SW |
272 | /* |
273 | * Submit the actual lock request. | |
274 | */ | |
275 | ||
a245769f | 276 | return dlm_lock(ls->ls_dlm, req, &gl->gl_lksb, lkf, strname, |
921169ca | 277 | GDLM_STRNAME_BYTES - 1, 0, gdlm_ast, gl, gdlm_bast); |
f057f6cd SW |
278 | } |
279 | ||
bc015cb8 | 280 | static void gdlm_put_lock(struct gfs2_glock *gl) |
f057f6cd | 281 | { |
e402746a SW |
282 | struct gfs2_sbd *sdp = gl->gl_sbd; |
283 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
d4e0bfec | 284 | int lvb_needs_unlock = 0; |
f057f6cd SW |
285 | int error; |
286 | ||
287 | if (gl->gl_lksb.sb_lkid == 0) { | |
fc0e38da | 288 | gfs2_glock_free(gl); |
f057f6cd SW |
289 | return; |
290 | } | |
291 | ||
a245769f SW |
292 | clear_bit(GLF_BLOCKING, &gl->gl_flags); |
293 | gfs2_glstats_inc(gl, GFS2_LKS_DCOUNT); | |
294 | gfs2_sbstats_inc(gl, GFS2_LKS_DCOUNT); | |
295 | gfs2_update_request_times(gl); | |
fb6791d1 DT |
296 | |
297 | /* don't want to skip dlm_unlock writing the lvb when lock is ex */ | |
d4e0bfec DT |
298 | |
299 | if (gl->gl_lksb.sb_lvbptr && (gl->gl_state == LM_ST_EXCLUSIVE)) | |
300 | lvb_needs_unlock = 1; | |
301 | ||
fb6791d1 | 302 | if (test_bit(SDF_SKIP_DLM_UNLOCK, &sdp->sd_flags) && |
d4e0bfec | 303 | !lvb_needs_unlock) { |
fb6791d1 DT |
304 | gfs2_glock_free(gl); |
305 | return; | |
306 | } | |
307 | ||
f057f6cd SW |
308 | error = dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_VALBLK, |
309 | NULL, gl); | |
310 | if (error) { | |
311 | printk(KERN_ERR "gdlm_unlock %x,%llx err=%d\n", | |
312 | gl->gl_name.ln_type, | |
313 | (unsigned long long)gl->gl_name.ln_number, error); | |
314 | return; | |
315 | } | |
316 | } | |
317 | ||
318 | static void gdlm_cancel(struct gfs2_glock *gl) | |
319 | { | |
320 | struct lm_lockstruct *ls = &gl->gl_sbd->sd_lockstruct; | |
321 | dlm_unlock(ls->ls_dlm, gl->gl_lksb.sb_lkid, DLM_LKF_CANCEL, NULL, gl); | |
322 | } | |
323 | ||
e0c2a9aa DT |
324 | /* |
325 | * dlm/gfs2 recovery coordination using dlm_recover callbacks | |
326 | * | |
327 | * 1. dlm_controld sees lockspace members change | |
328 | * 2. dlm_controld blocks dlm-kernel locking activity | |
329 | * 3. dlm_controld within dlm-kernel notifies gfs2 (recover_prep) | |
330 | * 4. dlm_controld starts and finishes its own user level recovery | |
331 | * 5. dlm_controld starts dlm-kernel dlm_recoverd to do kernel recovery | |
332 | * 6. dlm_recoverd notifies gfs2 of failed nodes (recover_slot) | |
333 | * 7. dlm_recoverd does its own lock recovery | |
334 | * 8. dlm_recoverd unblocks dlm-kernel locking activity | |
335 | * 9. dlm_recoverd notifies gfs2 when done (recover_done with new generation) | |
336 | * 10. gfs2_control updates control_lock lvb with new generation and jid bits | |
337 | * 11. gfs2_control enqueues journals for gfs2_recover to recover (maybe none) | |
338 | * 12. gfs2_recover dequeues and recovers journals of failed nodes | |
339 | * 13. gfs2_recover provides recovery results to gfs2_control (recovery_result) | |
340 | * 14. gfs2_control updates control_lock lvb jid bits for recovered journals | |
341 | * 15. gfs2_control unblocks normal locking when all journals are recovered | |
342 | * | |
343 | * - failures during recovery | |
344 | * | |
345 | * recover_prep() may set BLOCK_LOCKS (step 3) again before gfs2_control | |
346 | * clears BLOCK_LOCKS (step 15), e.g. another node fails while still | |
347 | * recovering for a prior failure. gfs2_control needs a way to detect | |
348 | * this so it can leave BLOCK_LOCKS set in step 15. This is managed using | |
349 | * the recover_block and recover_start values. | |
350 | * | |
351 | * recover_done() provides a new lockspace generation number each time it | |
352 | * is called (step 9). This generation number is saved as recover_start. | |
353 | * When recover_prep() is called, it sets BLOCK_LOCKS and sets | |
354 | * recover_block = recover_start. So, while recover_block is equal to | |
355 | * recover_start, BLOCK_LOCKS should remain set. (recover_spin must | |
356 | * be held around the BLOCK_LOCKS/recover_block/recover_start logic.) | |
357 | * | |
358 | * - more specific gfs2 steps in sequence above | |
359 | * | |
360 | * 3. recover_prep sets BLOCK_LOCKS and sets recover_block = recover_start | |
361 | * 6. recover_slot records any failed jids (maybe none) | |
362 | * 9. recover_done sets recover_start = new generation number | |
363 | * 10. gfs2_control sets control_lock lvb = new gen + bits for failed jids | |
364 | * 12. gfs2_recover does journal recoveries for failed jids identified above | |
365 | * 14. gfs2_control clears control_lock lvb bits for recovered jids | |
366 | * 15. gfs2_control checks if recover_block == recover_start (step 3 occured | |
367 | * again) then do nothing, otherwise if recover_start > recover_block | |
368 | * then clear BLOCK_LOCKS. | |
369 | * | |
370 | * - parallel recovery steps across all nodes | |
371 | * | |
372 | * All nodes attempt to update the control_lock lvb with the new generation | |
373 | * number and jid bits, but only the first to get the control_lock EX will | |
374 | * do so; others will see that it's already done (lvb already contains new | |
375 | * generation number.) | |
376 | * | |
377 | * . All nodes get the same recover_prep/recover_slot/recover_done callbacks | |
378 | * . All nodes attempt to set control_lock lvb gen + bits for the new gen | |
379 | * . One node gets control_lock first and writes the lvb, others see it's done | |
380 | * . All nodes attempt to recover jids for which they see control_lock bits set | |
381 | * . One node succeeds for a jid, and that one clears the jid bit in the lvb | |
382 | * . All nodes will eventually see all lvb bits clear and unblock locks | |
383 | * | |
384 | * - is there a problem with clearing an lvb bit that should be set | |
385 | * and missing a journal recovery? | |
386 | * | |
387 | * 1. jid fails | |
388 | * 2. lvb bit set for step 1 | |
389 | * 3. jid recovered for step 1 | |
390 | * 4. jid taken again (new mount) | |
391 | * 5. jid fails (for step 4) | |
392 | * 6. lvb bit set for step 5 (will already be set) | |
393 | * 7. lvb bit cleared for step 3 | |
394 | * | |
395 | * This is not a problem because the failure in step 5 does not | |
396 | * require recovery, because the mount in step 4 could not have | |
397 | * progressed far enough to unblock locks and access the fs. The | |
398 | * control_mount() function waits for all recoveries to be complete | |
399 | * for the latest lockspace generation before ever unblocking locks | |
400 | * and returning. The mount in step 4 waits until the recovery in | |
401 | * step 1 is done. | |
402 | * | |
403 | * - special case of first mounter: first node to mount the fs | |
404 | * | |
405 | * The first node to mount a gfs2 fs needs to check all the journals | |
406 | * and recover any that need recovery before other nodes are allowed | |
407 | * to mount the fs. (Others may begin mounting, but they must wait | |
408 | * for the first mounter to be done before taking locks on the fs | |
409 | * or accessing the fs.) This has two parts: | |
410 | * | |
411 | * 1. The mounted_lock tells a node it's the first to mount the fs. | |
412 | * Each node holds the mounted_lock in PR while it's mounted. | |
413 | * Each node tries to acquire the mounted_lock in EX when it mounts. | |
414 | * If a node is granted the mounted_lock EX it means there are no | |
415 | * other mounted nodes (no PR locks exist), and it is the first mounter. | |
416 | * The mounted_lock is demoted to PR when first recovery is done, so | |
417 | * others will fail to get an EX lock, but will get a PR lock. | |
418 | * | |
419 | * 2. The control_lock blocks others in control_mount() while the first | |
420 | * mounter is doing first mount recovery of all journals. | |
421 | * A mounting node needs to acquire control_lock in EX mode before | |
422 | * it can proceed. The first mounter holds control_lock in EX while doing | |
423 | * the first mount recovery, blocking mounts from other nodes, then demotes | |
424 | * control_lock to NL when it's done (others_may_mount/first_done), | |
425 | * allowing other nodes to continue mounting. | |
426 | * | |
427 | * first mounter: | |
428 | * control_lock EX/NOQUEUE success | |
429 | * mounted_lock EX/NOQUEUE success (no other PR, so no other mounters) | |
430 | * set first=1 | |
431 | * do first mounter recovery | |
432 | * mounted_lock EX->PR | |
433 | * control_lock EX->NL, write lvb generation | |
434 | * | |
435 | * other mounter: | |
436 | * control_lock EX/NOQUEUE success (if fail -EAGAIN, retry) | |
437 | * mounted_lock EX/NOQUEUE fail -EAGAIN (expected due to other mounters PR) | |
438 | * mounted_lock PR/NOQUEUE success | |
439 | * read lvb generation | |
440 | * control_lock EX->NL | |
441 | * set first=0 | |
442 | * | |
443 | * - mount during recovery | |
444 | * | |
445 | * If a node mounts while others are doing recovery (not first mounter), | |
446 | * the mounting node will get its initial recover_done() callback without | |
447 | * having seen any previous failures/callbacks. | |
448 | * | |
449 | * It must wait for all recoveries preceding its mount to be finished | |
450 | * before it unblocks locks. It does this by repeating the "other mounter" | |
451 | * steps above until the lvb generation number is >= its mount generation | |
452 | * number (from initial recover_done) and all lvb bits are clear. | |
453 | * | |
454 | * - control_lock lvb format | |
455 | * | |
456 | * 4 bytes generation number: the latest dlm lockspace generation number | |
457 | * from recover_done callback. Indicates the jid bitmap has been updated | |
458 | * to reflect all slot failures through that generation. | |
459 | * 4 bytes unused. | |
460 | * GDLM_LVB_SIZE-8 bytes of jid bit map. If bit N is set, it indicates | |
461 | * that jid N needs recovery. | |
462 | */ | |
463 | ||
464 | #define JID_BITMAP_OFFSET 8 /* 4 byte generation number + 4 byte unused */ | |
465 | ||
466 | static void control_lvb_read(struct lm_lockstruct *ls, uint32_t *lvb_gen, | |
467 | char *lvb_bits) | |
468 | { | |
951b4bd5 | 469 | __le32 gen; |
e0c2a9aa | 470 | memcpy(lvb_bits, ls->ls_control_lvb, GDLM_LVB_SIZE); |
951b4bd5 | 471 | memcpy(&gen, lvb_bits, sizeof(__le32)); |
e0c2a9aa DT |
472 | *lvb_gen = le32_to_cpu(gen); |
473 | } | |
474 | ||
475 | static void control_lvb_write(struct lm_lockstruct *ls, uint32_t lvb_gen, | |
476 | char *lvb_bits) | |
477 | { | |
951b4bd5 | 478 | __le32 gen; |
e0c2a9aa DT |
479 | memcpy(ls->ls_control_lvb, lvb_bits, GDLM_LVB_SIZE); |
480 | gen = cpu_to_le32(lvb_gen); | |
951b4bd5 | 481 | memcpy(ls->ls_control_lvb, &gen, sizeof(__le32)); |
e0c2a9aa DT |
482 | } |
483 | ||
484 | static int all_jid_bits_clear(char *lvb) | |
485 | { | |
4146c3d4 AM |
486 | return !memchr_inv(lvb + JID_BITMAP_OFFSET, 0, |
487 | GDLM_LVB_SIZE - JID_BITMAP_OFFSET); | |
e0c2a9aa DT |
488 | } |
489 | ||
490 | static void sync_wait_cb(void *arg) | |
491 | { | |
492 | struct lm_lockstruct *ls = arg; | |
493 | complete(&ls->ls_sync_wait); | |
494 | } | |
495 | ||
496 | static int sync_unlock(struct gfs2_sbd *sdp, struct dlm_lksb *lksb, char *name) | |
f057f6cd SW |
497 | { |
498 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
499 | int error; | |
500 | ||
e0c2a9aa DT |
501 | error = dlm_unlock(ls->ls_dlm, lksb->sb_lkid, 0, lksb, ls); |
502 | if (error) { | |
503 | fs_err(sdp, "%s lkid %x error %d\n", | |
504 | name, lksb->sb_lkid, error); | |
505 | return error; | |
506 | } | |
507 | ||
508 | wait_for_completion(&ls->ls_sync_wait); | |
509 | ||
510 | if (lksb->sb_status != -DLM_EUNLOCK) { | |
511 | fs_err(sdp, "%s lkid %x status %d\n", | |
512 | name, lksb->sb_lkid, lksb->sb_status); | |
513 | return -1; | |
514 | } | |
515 | return 0; | |
516 | } | |
517 | ||
518 | static int sync_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags, | |
519 | unsigned int num, struct dlm_lksb *lksb, char *name) | |
520 | { | |
521 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
522 | char strname[GDLM_STRNAME_BYTES]; | |
523 | int error, status; | |
524 | ||
525 | memset(strname, 0, GDLM_STRNAME_BYTES); | |
526 | snprintf(strname, GDLM_STRNAME_BYTES, "%8x%16x", LM_TYPE_NONDISK, num); | |
527 | ||
528 | error = dlm_lock(ls->ls_dlm, mode, lksb, flags, | |
529 | strname, GDLM_STRNAME_BYTES - 1, | |
530 | 0, sync_wait_cb, ls, NULL); | |
531 | if (error) { | |
532 | fs_err(sdp, "%s lkid %x flags %x mode %d error %d\n", | |
533 | name, lksb->sb_lkid, flags, mode, error); | |
534 | return error; | |
535 | } | |
536 | ||
537 | wait_for_completion(&ls->ls_sync_wait); | |
538 | ||
539 | status = lksb->sb_status; | |
540 | ||
541 | if (status && status != -EAGAIN) { | |
542 | fs_err(sdp, "%s lkid %x flags %x mode %d status %d\n", | |
543 | name, lksb->sb_lkid, flags, mode, status); | |
544 | } | |
545 | ||
546 | return status; | |
547 | } | |
548 | ||
549 | static int mounted_unlock(struct gfs2_sbd *sdp) | |
550 | { | |
551 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
552 | return sync_unlock(sdp, &ls->ls_mounted_lksb, "mounted_lock"); | |
553 | } | |
554 | ||
555 | static int mounted_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags) | |
556 | { | |
557 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
558 | return sync_lock(sdp, mode, flags, GFS2_MOUNTED_LOCK, | |
559 | &ls->ls_mounted_lksb, "mounted_lock"); | |
560 | } | |
561 | ||
562 | static int control_unlock(struct gfs2_sbd *sdp) | |
563 | { | |
564 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
565 | return sync_unlock(sdp, &ls->ls_control_lksb, "control_lock"); | |
566 | } | |
567 | ||
568 | static int control_lock(struct gfs2_sbd *sdp, int mode, uint32_t flags) | |
569 | { | |
570 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
571 | return sync_lock(sdp, mode, flags, GFS2_CONTROL_LOCK, | |
572 | &ls->ls_control_lksb, "control_lock"); | |
573 | } | |
574 | ||
575 | static void gfs2_control_func(struct work_struct *work) | |
576 | { | |
577 | struct gfs2_sbd *sdp = container_of(work, struct gfs2_sbd, sd_control_work.work); | |
578 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
e0c2a9aa DT |
579 | uint32_t block_gen, start_gen, lvb_gen, flags; |
580 | int recover_set = 0; | |
581 | int write_lvb = 0; | |
582 | int recover_size; | |
583 | int i, error; | |
584 | ||
585 | spin_lock(&ls->ls_recover_spin); | |
586 | /* | |
587 | * No MOUNT_DONE means we're still mounting; control_mount() | |
588 | * will set this flag, after which this thread will take over | |
589 | * all further clearing of BLOCK_LOCKS. | |
590 | * | |
591 | * FIRST_MOUNT means this node is doing first mounter recovery, | |
592 | * for which recovery control is handled by | |
593 | * control_mount()/control_first_done(), not this thread. | |
594 | */ | |
595 | if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) || | |
596 | test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) { | |
597 | spin_unlock(&ls->ls_recover_spin); | |
598 | return; | |
599 | } | |
600 | block_gen = ls->ls_recover_block; | |
601 | start_gen = ls->ls_recover_start; | |
602 | spin_unlock(&ls->ls_recover_spin); | |
603 | ||
604 | /* | |
605 | * Equal block_gen and start_gen implies we are between | |
606 | * recover_prep and recover_done callbacks, which means | |
607 | * dlm recovery is in progress and dlm locking is blocked. | |
608 | * There's no point trying to do any work until recover_done. | |
609 | */ | |
610 | ||
611 | if (block_gen == start_gen) | |
612 | return; | |
613 | ||
614 | /* | |
615 | * Propagate recover_submit[] and recover_result[] to lvb: | |
616 | * dlm_recoverd adds to recover_submit[] jids needing recovery | |
617 | * gfs2_recover adds to recover_result[] journal recovery results | |
618 | * | |
619 | * set lvb bit for jids in recover_submit[] if the lvb has not | |
620 | * yet been updated for the generation of the failure | |
621 | * | |
622 | * clear lvb bit for jids in recover_result[] if the result of | |
623 | * the journal recovery is SUCCESS | |
624 | */ | |
625 | ||
626 | error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_VALBLK); | |
627 | if (error) { | |
628 | fs_err(sdp, "control lock EX error %d\n", error); | |
629 | return; | |
630 | } | |
631 | ||
57c7310b | 632 | control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits); |
e0c2a9aa DT |
633 | |
634 | spin_lock(&ls->ls_recover_spin); | |
635 | if (block_gen != ls->ls_recover_block || | |
636 | start_gen != ls->ls_recover_start) { | |
637 | fs_info(sdp, "recover generation %u block1 %u %u\n", | |
638 | start_gen, block_gen, ls->ls_recover_block); | |
639 | spin_unlock(&ls->ls_recover_spin); | |
640 | control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT); | |
641 | return; | |
642 | } | |
643 | ||
644 | recover_size = ls->ls_recover_size; | |
645 | ||
646 | if (lvb_gen <= start_gen) { | |
647 | /* | |
648 | * Clear lvb bits for jids we've successfully recovered. | |
649 | * Because all nodes attempt to recover failed journals, | |
650 | * a journal can be recovered multiple times successfully | |
651 | * in succession. Only the first will really do recovery, | |
652 | * the others find it clean, but still report a successful | |
653 | * recovery. So, another node may have already recovered | |
654 | * the jid and cleared the lvb bit for it. | |
655 | */ | |
656 | for (i = 0; i < recover_size; i++) { | |
657 | if (ls->ls_recover_result[i] != LM_RD_SUCCESS) | |
658 | continue; | |
659 | ||
660 | ls->ls_recover_result[i] = 0; | |
661 | ||
57c7310b | 662 | if (!test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) |
e0c2a9aa DT |
663 | continue; |
664 | ||
57c7310b | 665 | __clear_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET); |
e0c2a9aa DT |
666 | write_lvb = 1; |
667 | } | |
668 | } | |
669 | ||
670 | if (lvb_gen == start_gen) { | |
671 | /* | |
672 | * Failed slots before start_gen are already set in lvb. | |
673 | */ | |
674 | for (i = 0; i < recover_size; i++) { | |
675 | if (!ls->ls_recover_submit[i]) | |
676 | continue; | |
677 | if (ls->ls_recover_submit[i] < lvb_gen) | |
678 | ls->ls_recover_submit[i] = 0; | |
679 | } | |
680 | } else if (lvb_gen < start_gen) { | |
681 | /* | |
682 | * Failed slots before start_gen are not yet set in lvb. | |
683 | */ | |
684 | for (i = 0; i < recover_size; i++) { | |
685 | if (!ls->ls_recover_submit[i]) | |
686 | continue; | |
687 | if (ls->ls_recover_submit[i] < start_gen) { | |
688 | ls->ls_recover_submit[i] = 0; | |
57c7310b | 689 | __set_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET); |
e0c2a9aa DT |
690 | } |
691 | } | |
692 | /* even if there are no bits to set, we need to write the | |
693 | latest generation to the lvb */ | |
694 | write_lvb = 1; | |
695 | } else { | |
696 | /* | |
697 | * we should be getting a recover_done() for lvb_gen soon | |
698 | */ | |
699 | } | |
700 | spin_unlock(&ls->ls_recover_spin); | |
701 | ||
702 | if (write_lvb) { | |
57c7310b | 703 | control_lvb_write(ls, start_gen, ls->ls_lvb_bits); |
e0c2a9aa DT |
704 | flags = DLM_LKF_CONVERT | DLM_LKF_VALBLK; |
705 | } else { | |
706 | flags = DLM_LKF_CONVERT; | |
707 | } | |
708 | ||
709 | error = control_lock(sdp, DLM_LOCK_NL, flags); | |
710 | if (error) { | |
711 | fs_err(sdp, "control lock NL error %d\n", error); | |
712 | return; | |
713 | } | |
714 | ||
715 | /* | |
716 | * Everyone will see jid bits set in the lvb, run gfs2_recover_set(), | |
717 | * and clear a jid bit in the lvb if the recovery is a success. | |
718 | * Eventually all journals will be recovered, all jid bits will | |
719 | * be cleared in the lvb, and everyone will clear BLOCK_LOCKS. | |
720 | */ | |
721 | ||
722 | for (i = 0; i < recover_size; i++) { | |
57c7310b | 723 | if (test_bit_le(i, ls->ls_lvb_bits + JID_BITMAP_OFFSET)) { |
e0c2a9aa DT |
724 | fs_info(sdp, "recover generation %u jid %d\n", |
725 | start_gen, i); | |
726 | gfs2_recover_set(sdp, i); | |
727 | recover_set++; | |
728 | } | |
729 | } | |
730 | if (recover_set) | |
731 | return; | |
732 | ||
733 | /* | |
734 | * No more jid bits set in lvb, all recovery is done, unblock locks | |
735 | * (unless a new recover_prep callback has occured blocking locks | |
736 | * again while working above) | |
737 | */ | |
738 | ||
739 | spin_lock(&ls->ls_recover_spin); | |
740 | if (ls->ls_recover_block == block_gen && | |
741 | ls->ls_recover_start == start_gen) { | |
742 | clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags); | |
743 | spin_unlock(&ls->ls_recover_spin); | |
744 | fs_info(sdp, "recover generation %u done\n", start_gen); | |
745 | gfs2_glock_thaw(sdp); | |
746 | } else { | |
747 | fs_info(sdp, "recover generation %u block2 %u %u\n", | |
748 | start_gen, block_gen, ls->ls_recover_block); | |
749 | spin_unlock(&ls->ls_recover_spin); | |
750 | } | |
751 | } | |
752 | ||
753 | static int control_mount(struct gfs2_sbd *sdp) | |
754 | { | |
755 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
e0c2a9aa DT |
756 | uint32_t start_gen, block_gen, mount_gen, lvb_gen; |
757 | int mounted_mode; | |
758 | int retries = 0; | |
759 | int error; | |
760 | ||
761 | memset(&ls->ls_mounted_lksb, 0, sizeof(struct dlm_lksb)); | |
762 | memset(&ls->ls_control_lksb, 0, sizeof(struct dlm_lksb)); | |
763 | memset(&ls->ls_control_lvb, 0, GDLM_LVB_SIZE); | |
764 | ls->ls_control_lksb.sb_lvbptr = ls->ls_control_lvb; | |
765 | init_completion(&ls->ls_sync_wait); | |
766 | ||
767 | set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags); | |
768 | ||
769 | error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_VALBLK); | |
770 | if (error) { | |
771 | fs_err(sdp, "control_mount control_lock NL error %d\n", error); | |
772 | return error; | |
773 | } | |
774 | ||
775 | error = mounted_lock(sdp, DLM_LOCK_NL, 0); | |
776 | if (error) { | |
777 | fs_err(sdp, "control_mount mounted_lock NL error %d\n", error); | |
778 | control_unlock(sdp); | |
779 | return error; | |
780 | } | |
781 | mounted_mode = DLM_LOCK_NL; | |
782 | ||
783 | restart: | |
784 | if (retries++ && signal_pending(current)) { | |
785 | error = -EINTR; | |
786 | goto fail; | |
787 | } | |
788 | ||
789 | /* | |
790 | * We always start with both locks in NL. control_lock is | |
791 | * demoted to NL below so we don't need to do it here. | |
792 | */ | |
793 | ||
794 | if (mounted_mode != DLM_LOCK_NL) { | |
795 | error = mounted_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT); | |
796 | if (error) | |
797 | goto fail; | |
798 | mounted_mode = DLM_LOCK_NL; | |
799 | } | |
800 | ||
801 | /* | |
802 | * Other nodes need to do some work in dlm recovery and gfs2_control | |
803 | * before the recover_done and control_lock will be ready for us below. | |
804 | * A delay here is not required but often avoids having to retry. | |
805 | */ | |
806 | ||
807 | msleep_interruptible(500); | |
808 | ||
809 | /* | |
810 | * Acquire control_lock in EX and mounted_lock in either EX or PR. | |
811 | * control_lock lvb keeps track of any pending journal recoveries. | |
812 | * mounted_lock indicates if any other nodes have the fs mounted. | |
813 | */ | |
814 | ||
815 | error = control_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE|DLM_LKF_VALBLK); | |
816 | if (error == -EAGAIN) { | |
817 | goto restart; | |
818 | } else if (error) { | |
819 | fs_err(sdp, "control_mount control_lock EX error %d\n", error); | |
820 | goto fail; | |
821 | } | |
822 | ||
823 | error = mounted_lock(sdp, DLM_LOCK_EX, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE); | |
824 | if (!error) { | |
825 | mounted_mode = DLM_LOCK_EX; | |
826 | goto locks_done; | |
827 | } else if (error != -EAGAIN) { | |
828 | fs_err(sdp, "control_mount mounted_lock EX error %d\n", error); | |
829 | goto fail; | |
830 | } | |
831 | ||
832 | error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT|DLM_LKF_NOQUEUE); | |
833 | if (!error) { | |
834 | mounted_mode = DLM_LOCK_PR; | |
835 | goto locks_done; | |
836 | } else { | |
837 | /* not even -EAGAIN should happen here */ | |
838 | fs_err(sdp, "control_mount mounted_lock PR error %d\n", error); | |
839 | goto fail; | |
840 | } | |
841 | ||
842 | locks_done: | |
843 | /* | |
844 | * If we got both locks above in EX, then we're the first mounter. | |
845 | * If not, then we need to wait for the control_lock lvb to be | |
846 | * updated by other mounted nodes to reflect our mount generation. | |
847 | * | |
848 | * In simple first mounter cases, first mounter will see zero lvb_gen, | |
849 | * but in cases where all existing nodes leave/fail before mounting | |
850 | * nodes finish control_mount, then all nodes will be mounting and | |
851 | * lvb_gen will be non-zero. | |
852 | */ | |
853 | ||
57c7310b | 854 | control_lvb_read(ls, &lvb_gen, ls->ls_lvb_bits); |
e0c2a9aa DT |
855 | |
856 | if (lvb_gen == 0xFFFFFFFF) { | |
857 | /* special value to force mount attempts to fail */ | |
858 | fs_err(sdp, "control_mount control_lock disabled\n"); | |
859 | error = -EINVAL; | |
860 | goto fail; | |
861 | } | |
862 | ||
863 | if (mounted_mode == DLM_LOCK_EX) { | |
864 | /* first mounter, keep both EX while doing first recovery */ | |
865 | spin_lock(&ls->ls_recover_spin); | |
866 | clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags); | |
867 | set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags); | |
868 | set_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags); | |
869 | spin_unlock(&ls->ls_recover_spin); | |
870 | fs_info(sdp, "first mounter control generation %u\n", lvb_gen); | |
871 | return 0; | |
872 | } | |
873 | ||
874 | error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT); | |
875 | if (error) | |
876 | goto fail; | |
877 | ||
878 | /* | |
879 | * We are not first mounter, now we need to wait for the control_lock | |
880 | * lvb generation to be >= the generation from our first recover_done | |
881 | * and all lvb bits to be clear (no pending journal recoveries.) | |
882 | */ | |
883 | ||
57c7310b | 884 | if (!all_jid_bits_clear(ls->ls_lvb_bits)) { |
e0c2a9aa DT |
885 | /* journals need recovery, wait until all are clear */ |
886 | fs_info(sdp, "control_mount wait for journal recovery\n"); | |
887 | goto restart; | |
888 | } | |
889 | ||
890 | spin_lock(&ls->ls_recover_spin); | |
891 | block_gen = ls->ls_recover_block; | |
892 | start_gen = ls->ls_recover_start; | |
893 | mount_gen = ls->ls_recover_mount; | |
894 | ||
895 | if (lvb_gen < mount_gen) { | |
896 | /* wait for mounted nodes to update control_lock lvb to our | |
897 | generation, which might include new recovery bits set */ | |
898 | fs_info(sdp, "control_mount wait1 block %u start %u mount %u " | |
899 | "lvb %u flags %lx\n", block_gen, start_gen, mount_gen, | |
900 | lvb_gen, ls->ls_recover_flags); | |
901 | spin_unlock(&ls->ls_recover_spin); | |
902 | goto restart; | |
903 | } | |
904 | ||
905 | if (lvb_gen != start_gen) { | |
906 | /* wait for mounted nodes to update control_lock lvb to the | |
907 | latest recovery generation */ | |
908 | fs_info(sdp, "control_mount wait2 block %u start %u mount %u " | |
909 | "lvb %u flags %lx\n", block_gen, start_gen, mount_gen, | |
910 | lvb_gen, ls->ls_recover_flags); | |
911 | spin_unlock(&ls->ls_recover_spin); | |
912 | goto restart; | |
913 | } | |
914 | ||
915 | if (block_gen == start_gen) { | |
916 | /* dlm recovery in progress, wait for it to finish */ | |
917 | fs_info(sdp, "control_mount wait3 block %u start %u mount %u " | |
918 | "lvb %u flags %lx\n", block_gen, start_gen, mount_gen, | |
919 | lvb_gen, ls->ls_recover_flags); | |
920 | spin_unlock(&ls->ls_recover_spin); | |
921 | goto restart; | |
f057f6cd SW |
922 | } |
923 | ||
e0c2a9aa DT |
924 | clear_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags); |
925 | set_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags); | |
926 | memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t)); | |
927 | memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t)); | |
928 | spin_unlock(&ls->ls_recover_spin); | |
929 | return 0; | |
930 | ||
931 | fail: | |
932 | mounted_unlock(sdp); | |
933 | control_unlock(sdp); | |
934 | return error; | |
935 | } | |
936 | ||
937 | static int dlm_recovery_wait(void *word) | |
938 | { | |
939 | schedule(); | |
940 | return 0; | |
941 | } | |
942 | ||
943 | static int control_first_done(struct gfs2_sbd *sdp) | |
944 | { | |
945 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
e0c2a9aa DT |
946 | uint32_t start_gen, block_gen; |
947 | int error; | |
948 | ||
949 | restart: | |
950 | spin_lock(&ls->ls_recover_spin); | |
951 | start_gen = ls->ls_recover_start; | |
952 | block_gen = ls->ls_recover_block; | |
953 | ||
954 | if (test_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags) || | |
955 | !test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) || | |
956 | !test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) { | |
957 | /* sanity check, should not happen */ | |
958 | fs_err(sdp, "control_first_done start %u block %u flags %lx\n", | |
959 | start_gen, block_gen, ls->ls_recover_flags); | |
960 | spin_unlock(&ls->ls_recover_spin); | |
961 | control_unlock(sdp); | |
962 | return -1; | |
963 | } | |
964 | ||
965 | if (start_gen == block_gen) { | |
966 | /* | |
967 | * Wait for the end of a dlm recovery cycle to switch from | |
968 | * first mounter recovery. We can ignore any recover_slot | |
969 | * callbacks between the recover_prep and next recover_done | |
970 | * because we are still the first mounter and any failed nodes | |
971 | * have not fully mounted, so they don't need recovery. | |
972 | */ | |
973 | spin_unlock(&ls->ls_recover_spin); | |
974 | fs_info(sdp, "control_first_done wait gen %u\n", start_gen); | |
975 | ||
976 | wait_on_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY, | |
977 | dlm_recovery_wait, TASK_UNINTERRUPTIBLE); | |
978 | goto restart; | |
979 | } | |
980 | ||
981 | clear_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags); | |
982 | set_bit(DFL_FIRST_MOUNT_DONE, &ls->ls_recover_flags); | |
983 | memset(ls->ls_recover_submit, 0, ls->ls_recover_size*sizeof(uint32_t)); | |
984 | memset(ls->ls_recover_result, 0, ls->ls_recover_size*sizeof(uint32_t)); | |
985 | spin_unlock(&ls->ls_recover_spin); | |
986 | ||
57c7310b DT |
987 | memset(ls->ls_lvb_bits, 0, GDLM_LVB_SIZE); |
988 | control_lvb_write(ls, start_gen, ls->ls_lvb_bits); | |
e0c2a9aa DT |
989 | |
990 | error = mounted_lock(sdp, DLM_LOCK_PR, DLM_LKF_CONVERT); | |
991 | if (error) | |
992 | fs_err(sdp, "control_first_done mounted PR error %d\n", error); | |
993 | ||
994 | error = control_lock(sdp, DLM_LOCK_NL, DLM_LKF_CONVERT|DLM_LKF_VALBLK); | |
f057f6cd | 995 | if (error) |
e0c2a9aa | 996 | fs_err(sdp, "control_first_done control NL error %d\n", error); |
f057f6cd SW |
997 | |
998 | return error; | |
999 | } | |
1000 | ||
e0c2a9aa DT |
1001 | /* |
1002 | * Expand static jid arrays if necessary (by increments of RECOVER_SIZE_INC) | |
1003 | * to accomodate the largest slot number. (NB dlm slot numbers start at 1, | |
1004 | * gfs2 jids start at 0, so jid = slot - 1) | |
1005 | */ | |
1006 | ||
1007 | #define RECOVER_SIZE_INC 16 | |
1008 | ||
1009 | static int set_recover_size(struct gfs2_sbd *sdp, struct dlm_slot *slots, | |
1010 | int num_slots) | |
1011 | { | |
1012 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
1013 | uint32_t *submit = NULL; | |
1014 | uint32_t *result = NULL; | |
1015 | uint32_t old_size, new_size; | |
1016 | int i, max_jid; | |
1017 | ||
57c7310b DT |
1018 | if (!ls->ls_lvb_bits) { |
1019 | ls->ls_lvb_bits = kzalloc(GDLM_LVB_SIZE, GFP_NOFS); | |
1020 | if (!ls->ls_lvb_bits) | |
1021 | return -ENOMEM; | |
1022 | } | |
1023 | ||
e0c2a9aa DT |
1024 | max_jid = 0; |
1025 | for (i = 0; i < num_slots; i++) { | |
1026 | if (max_jid < slots[i].slot - 1) | |
1027 | max_jid = slots[i].slot - 1; | |
1028 | } | |
1029 | ||
1030 | old_size = ls->ls_recover_size; | |
1031 | ||
1032 | if (old_size >= max_jid + 1) | |
1033 | return 0; | |
1034 | ||
1035 | new_size = old_size + RECOVER_SIZE_INC; | |
1036 | ||
1037 | submit = kzalloc(new_size * sizeof(uint32_t), GFP_NOFS); | |
1038 | result = kzalloc(new_size * sizeof(uint32_t), GFP_NOFS); | |
1039 | if (!submit || !result) { | |
1040 | kfree(submit); | |
1041 | kfree(result); | |
1042 | return -ENOMEM; | |
1043 | } | |
1044 | ||
1045 | spin_lock(&ls->ls_recover_spin); | |
1046 | memcpy(submit, ls->ls_recover_submit, old_size * sizeof(uint32_t)); | |
1047 | memcpy(result, ls->ls_recover_result, old_size * sizeof(uint32_t)); | |
1048 | kfree(ls->ls_recover_submit); | |
1049 | kfree(ls->ls_recover_result); | |
1050 | ls->ls_recover_submit = submit; | |
1051 | ls->ls_recover_result = result; | |
1052 | ls->ls_recover_size = new_size; | |
1053 | spin_unlock(&ls->ls_recover_spin); | |
1054 | return 0; | |
1055 | } | |
1056 | ||
1057 | static void free_recover_size(struct lm_lockstruct *ls) | |
1058 | { | |
57c7310b | 1059 | kfree(ls->ls_lvb_bits); |
e0c2a9aa DT |
1060 | kfree(ls->ls_recover_submit); |
1061 | kfree(ls->ls_recover_result); | |
1062 | ls->ls_recover_submit = NULL; | |
1063 | ls->ls_recover_result = NULL; | |
1064 | ls->ls_recover_size = 0; | |
1065 | } | |
1066 | ||
1067 | /* dlm calls before it does lock recovery */ | |
1068 | ||
1069 | static void gdlm_recover_prep(void *arg) | |
1070 | { | |
1071 | struct gfs2_sbd *sdp = arg; | |
1072 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
1073 | ||
1074 | spin_lock(&ls->ls_recover_spin); | |
1075 | ls->ls_recover_block = ls->ls_recover_start; | |
1076 | set_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags); | |
1077 | ||
1078 | if (!test_bit(DFL_MOUNT_DONE, &ls->ls_recover_flags) || | |
1079 | test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) { | |
1080 | spin_unlock(&ls->ls_recover_spin); | |
1081 | return; | |
1082 | } | |
1083 | set_bit(DFL_BLOCK_LOCKS, &ls->ls_recover_flags); | |
1084 | spin_unlock(&ls->ls_recover_spin); | |
1085 | } | |
1086 | ||
1087 | /* dlm calls after recover_prep has been completed on all lockspace members; | |
1088 | identifies slot/jid of failed member */ | |
1089 | ||
1090 | static void gdlm_recover_slot(void *arg, struct dlm_slot *slot) | |
1091 | { | |
1092 | struct gfs2_sbd *sdp = arg; | |
1093 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
1094 | int jid = slot->slot - 1; | |
1095 | ||
1096 | spin_lock(&ls->ls_recover_spin); | |
1097 | if (ls->ls_recover_size < jid + 1) { | |
1098 | fs_err(sdp, "recover_slot jid %d gen %u short size %d", | |
1099 | jid, ls->ls_recover_block, ls->ls_recover_size); | |
1100 | spin_unlock(&ls->ls_recover_spin); | |
1101 | return; | |
1102 | } | |
1103 | ||
1104 | if (ls->ls_recover_submit[jid]) { | |
1105 | fs_info(sdp, "recover_slot jid %d gen %u prev %u", | |
1106 | jid, ls->ls_recover_block, ls->ls_recover_submit[jid]); | |
1107 | } | |
1108 | ls->ls_recover_submit[jid] = ls->ls_recover_block; | |
1109 | spin_unlock(&ls->ls_recover_spin); | |
1110 | } | |
1111 | ||
1112 | /* dlm calls after recover_slot and after it completes lock recovery */ | |
1113 | ||
1114 | static void gdlm_recover_done(void *arg, struct dlm_slot *slots, int num_slots, | |
1115 | int our_slot, uint32_t generation) | |
1116 | { | |
1117 | struct gfs2_sbd *sdp = arg; | |
1118 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
1119 | ||
1120 | /* ensure the ls jid arrays are large enough */ | |
1121 | set_recover_size(sdp, slots, num_slots); | |
1122 | ||
1123 | spin_lock(&ls->ls_recover_spin); | |
1124 | ls->ls_recover_start = generation; | |
1125 | ||
1126 | if (!ls->ls_recover_mount) { | |
1127 | ls->ls_recover_mount = generation; | |
1128 | ls->ls_jid = our_slot - 1; | |
1129 | } | |
1130 | ||
1131 | if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags)) | |
1132 | queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, 0); | |
1133 | ||
1134 | clear_bit(DFL_DLM_RECOVERY, &ls->ls_recover_flags); | |
1135 | smp_mb__after_clear_bit(); | |
1136 | wake_up_bit(&ls->ls_recover_flags, DFL_DLM_RECOVERY); | |
1137 | spin_unlock(&ls->ls_recover_spin); | |
1138 | } | |
1139 | ||
1140 | /* gfs2_recover thread has a journal recovery result */ | |
1141 | ||
1142 | static void gdlm_recovery_result(struct gfs2_sbd *sdp, unsigned int jid, | |
1143 | unsigned int result) | |
1144 | { | |
1145 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
1146 | ||
1147 | if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags)) | |
1148 | return; | |
1149 | ||
1150 | /* don't care about the recovery of own journal during mount */ | |
1151 | if (jid == ls->ls_jid) | |
1152 | return; | |
1153 | ||
1154 | spin_lock(&ls->ls_recover_spin); | |
1155 | if (test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags)) { | |
1156 | spin_unlock(&ls->ls_recover_spin); | |
1157 | return; | |
1158 | } | |
1159 | if (ls->ls_recover_size < jid + 1) { | |
1160 | fs_err(sdp, "recovery_result jid %d short size %d", | |
1161 | jid, ls->ls_recover_size); | |
1162 | spin_unlock(&ls->ls_recover_spin); | |
1163 | return; | |
1164 | } | |
1165 | ||
1166 | fs_info(sdp, "recover jid %d result %s\n", jid, | |
1167 | result == LM_RD_GAVEUP ? "busy" : "success"); | |
1168 | ||
1169 | ls->ls_recover_result[jid] = result; | |
1170 | ||
1171 | /* GAVEUP means another node is recovering the journal; delay our | |
1172 | next attempt to recover it, to give the other node a chance to | |
1173 | finish before trying again */ | |
1174 | ||
1175 | if (!test_bit(DFL_UNMOUNT, &ls->ls_recover_flags)) | |
1176 | queue_delayed_work(gfs2_control_wq, &sdp->sd_control_work, | |
1177 | result == LM_RD_GAVEUP ? HZ : 0); | |
1178 | spin_unlock(&ls->ls_recover_spin); | |
1179 | } | |
1180 | ||
1181 | const struct dlm_lockspace_ops gdlm_lockspace_ops = { | |
1182 | .recover_prep = gdlm_recover_prep, | |
1183 | .recover_slot = gdlm_recover_slot, | |
1184 | .recover_done = gdlm_recover_done, | |
1185 | }; | |
1186 | ||
1187 | static int gdlm_mount(struct gfs2_sbd *sdp, const char *table) | |
1188 | { | |
1189 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
1190 | char cluster[GFS2_LOCKNAME_LEN]; | |
1191 | const char *fsname; | |
1192 | uint32_t flags; | |
1193 | int error, ops_result; | |
1194 | ||
1195 | /* | |
1196 | * initialize everything | |
1197 | */ | |
1198 | ||
1199 | INIT_DELAYED_WORK(&sdp->sd_control_work, gfs2_control_func); | |
1200 | spin_lock_init(&ls->ls_recover_spin); | |
1201 | ls->ls_recover_flags = 0; | |
1202 | ls->ls_recover_mount = 0; | |
1203 | ls->ls_recover_start = 0; | |
1204 | ls->ls_recover_block = 0; | |
1205 | ls->ls_recover_size = 0; | |
1206 | ls->ls_recover_submit = NULL; | |
1207 | ls->ls_recover_result = NULL; | |
57c7310b | 1208 | ls->ls_lvb_bits = NULL; |
e0c2a9aa DT |
1209 | |
1210 | error = set_recover_size(sdp, NULL, 0); | |
1211 | if (error) | |
1212 | goto fail; | |
1213 | ||
1214 | /* | |
1215 | * prepare dlm_new_lockspace args | |
1216 | */ | |
1217 | ||
1218 | fsname = strchr(table, ':'); | |
1219 | if (!fsname) { | |
1220 | fs_info(sdp, "no fsname found\n"); | |
1221 | error = -EINVAL; | |
1222 | goto fail_free; | |
1223 | } | |
1224 | memset(cluster, 0, sizeof(cluster)); | |
1225 | memcpy(cluster, table, strlen(table) - strlen(fsname)); | |
1226 | fsname++; | |
1227 | ||
1228 | flags = DLM_LSFL_FS | DLM_LSFL_NEWEXCL; | |
e0c2a9aa DT |
1229 | |
1230 | /* | |
1231 | * create/join lockspace | |
1232 | */ | |
1233 | ||
1234 | error = dlm_new_lockspace(fsname, cluster, flags, GDLM_LVB_SIZE, | |
1235 | &gdlm_lockspace_ops, sdp, &ops_result, | |
1236 | &ls->ls_dlm); | |
1237 | if (error) { | |
1238 | fs_err(sdp, "dlm_new_lockspace error %d\n", error); | |
1239 | goto fail_free; | |
1240 | } | |
1241 | ||
1242 | if (ops_result < 0) { | |
1243 | /* | |
1244 | * dlm does not support ops callbacks, | |
1245 | * old dlm_controld/gfs_controld are used, try without ops. | |
1246 | */ | |
1247 | fs_info(sdp, "dlm lockspace ops not used\n"); | |
1248 | free_recover_size(ls); | |
1249 | set_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags); | |
1250 | return 0; | |
1251 | } | |
1252 | ||
1253 | if (!test_bit(SDF_NOJOURNALID, &sdp->sd_flags)) { | |
1254 | fs_err(sdp, "dlm lockspace ops disallow jid preset\n"); | |
1255 | error = -EINVAL; | |
1256 | goto fail_release; | |
1257 | } | |
1258 | ||
1259 | /* | |
1260 | * control_mount() uses control_lock to determine first mounter, | |
1261 | * and for later mounts, waits for any recoveries to be cleared. | |
1262 | */ | |
1263 | ||
1264 | error = control_mount(sdp); | |
1265 | if (error) { | |
1266 | fs_err(sdp, "mount control error %d\n", error); | |
1267 | goto fail_release; | |
1268 | } | |
1269 | ||
1270 | ls->ls_first = !!test_bit(DFL_FIRST_MOUNT, &ls->ls_recover_flags); | |
1271 | clear_bit(SDF_NOJOURNALID, &sdp->sd_flags); | |
1272 | smp_mb__after_clear_bit(); | |
1273 | wake_up_bit(&sdp->sd_flags, SDF_NOJOURNALID); | |
1274 | return 0; | |
1275 | ||
1276 | fail_release: | |
1277 | dlm_release_lockspace(ls->ls_dlm, 2); | |
1278 | fail_free: | |
1279 | free_recover_size(ls); | |
1280 | fail: | |
1281 | return error; | |
1282 | } | |
1283 | ||
1284 | static void gdlm_first_done(struct gfs2_sbd *sdp) | |
1285 | { | |
1286 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
1287 | int error; | |
1288 | ||
1289 | if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags)) | |
1290 | return; | |
1291 | ||
1292 | error = control_first_done(sdp); | |
1293 | if (error) | |
1294 | fs_err(sdp, "mount first_done error %d\n", error); | |
1295 | } | |
1296 | ||
f057f6cd SW |
1297 | static void gdlm_unmount(struct gfs2_sbd *sdp) |
1298 | { | |
1299 | struct lm_lockstruct *ls = &sdp->sd_lockstruct; | |
1300 | ||
e0c2a9aa DT |
1301 | if (test_bit(DFL_NO_DLM_OPS, &ls->ls_recover_flags)) |
1302 | goto release; | |
1303 | ||
1304 | /* wait for gfs2_control_wq to be done with this mount */ | |
1305 | ||
1306 | spin_lock(&ls->ls_recover_spin); | |
1307 | set_bit(DFL_UNMOUNT, &ls->ls_recover_flags); | |
1308 | spin_unlock(&ls->ls_recover_spin); | |
43829731 | 1309 | flush_delayed_work(&sdp->sd_control_work); |
e0c2a9aa DT |
1310 | |
1311 | /* mounted_lock and control_lock will be purged in dlm recovery */ | |
1312 | release: | |
f057f6cd SW |
1313 | if (ls->ls_dlm) { |
1314 | dlm_release_lockspace(ls->ls_dlm, 2); | |
1315 | ls->ls_dlm = NULL; | |
1316 | } | |
e0c2a9aa DT |
1317 | |
1318 | free_recover_size(ls); | |
f057f6cd SW |
1319 | } |
1320 | ||
1321 | static const match_table_t dlm_tokens = { | |
1322 | { Opt_jid, "jid=%d"}, | |
1323 | { Opt_id, "id=%d"}, | |
1324 | { Opt_first, "first=%d"}, | |
1325 | { Opt_nodir, "nodir=%d"}, | |
1326 | { Opt_err, NULL }, | |
1327 | }; | |
1328 | ||
1329 | const struct lm_lockops gfs2_dlm_ops = { | |
1330 | .lm_proto_name = "lock_dlm", | |
1331 | .lm_mount = gdlm_mount, | |
e0c2a9aa DT |
1332 | .lm_first_done = gdlm_first_done, |
1333 | .lm_recovery_result = gdlm_recovery_result, | |
f057f6cd SW |
1334 | .lm_unmount = gdlm_unmount, |
1335 | .lm_put_lock = gdlm_put_lock, | |
1336 | .lm_lock = gdlm_lock, | |
1337 | .lm_cancel = gdlm_cancel, | |
1338 | .lm_tokens = &dlm_tokens, | |
1339 | }; | |
1340 |