Mercurial > illumos > illumos-gate
annotate usr/src/uts/common/io/lvm/mirror/mirror.c @ 1623:7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
4521995 Allow descriptive names for metadevices and hot spare pools
6396962 metareplace -c is obsolete and no longer needed.
author | tw21770 |
---|---|
date | Tue, 14 Mar 2006 14:53:36 -0800 |
parents | 18ae7db30fe7 |
children | a6ebd483c3cf |
rev | line source |
---|---|
0 | 1 /* |
2 * CDDL HEADER START | |
3 * | |
4 * The contents of this file are subject to the terms of the | |
1366
18ae7db30fe7
6376469 Drivers are declaring _depends_on incorrectly leading to venus optimizing it out
petede
parents:
46
diff
changeset
|
5 * Common Development and Distribution License (the "License"). |
18ae7db30fe7
6376469 Drivers are declaring _depends_on incorrectly leading to venus optimizing it out
petede
parents:
46
diff
changeset
|
6 * You may not use this file except in compliance with the License. |
0 | 7 * |
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 * or http://www.opensolaris.org/os/licensing. | |
10 * See the License for the specific language governing permissions | |
11 * and limitations under the License. | |
12 * | |
13 * When distributing Covered Code, include this CDDL HEADER in each | |
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 * If applicable, add the following below this CDDL HEADER, with the | |
16 * fields enclosed by brackets "[]" replaced with your own identifying | |
17 * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 * | |
19 * CDDL HEADER END | |
20 */ | |
21 /* | |
1366
18ae7db30fe7
6376469 Drivers are declaring _depends_on incorrectly leading to venus optimizing it out
petede
parents:
46
diff
changeset
|
22 * Copyright 2006 Sun Microsystems, Inc. All rights reserved. |
0 | 23 * Use is subject to license terms. |
24 */ | |
25 | |
26 #pragma ident "%Z%%M% %I% %E% SMI" | |
27 | |
28 #include <sys/param.h> | |
29 #include <sys/systm.h> | |
30 #include <sys/conf.h> | |
31 #include <sys/file.h> | |
32 #include <sys/user.h> | |
33 #include <sys/uio.h> | |
34 #include <sys/t_lock.h> | |
35 #include <sys/buf.h> | |
36 #include <sys/dkio.h> | |
37 #include <sys/vtoc.h> | |
38 #include <sys/kmem.h> | |
39 #include <vm/page.h> | |
40 #include <sys/cmn_err.h> | |
41 #include <sys/sysmacros.h> | |
42 #include <sys/types.h> | |
43 #include <sys/mkdev.h> | |
44 #include <sys/stat.h> | |
45 #include <sys/open.h> | |
46 #include <sys/modctl.h> | |
47 #include <sys/ddi.h> | |
48 #include <sys/sunddi.h> | |
49 #include <sys/debug.h> | |
50 #include <sys/dklabel.h> | |
51 #include <vm/hat.h> | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
52 #include <sys/lvm/mdvar.h> |
0 | 53 #include <sys/lvm/md_mirror.h> |
54 #include <sys/lvm/md_convert.h> | |
55 #include <sys/lvm/md_mddb.h> | |
56 #include <sys/esunddi.h> | |
57 | |
58 #include <sys/sysevent/eventdefs.h> | |
59 #include <sys/sysevent/svm.h> | |
60 #include <sys/lvm/mdmn_commd.h> | |
61 | |
62 md_ops_t mirror_md_ops; | |
63 #ifndef lint | |
1366
18ae7db30fe7
6376469 Drivers are declaring _depends_on incorrectly leading to venus optimizing it out
petede
parents:
46
diff
changeset
|
64 char _depends_on[] = "drv/md"; |
0 | 65 md_ops_t *md_interface_ops = &mirror_md_ops; |
66 #endif | |
67 | |
68 extern mdq_anchor_t md_done_daemon; | |
69 extern mdq_anchor_t md_mstr_daemon; | |
70 extern mdq_anchor_t md_mirror_daemon; | |
71 extern mdq_anchor_t md_mirror_io_daemon; | |
72 extern mdq_anchor_t md_mirror_rs_daemon; | |
73 extern mdq_anchor_t md_mhs_daemon; | |
74 | |
75 extern unit_t md_nunits; | |
76 extern set_t md_nsets; | |
77 extern md_set_t md_set[]; | |
78 | |
79 extern int md_status; | |
80 extern clock_t md_hz; | |
81 | |
82 extern md_krwlock_t md_unit_array_rw; | |
83 extern kmutex_t md_mx; | |
84 extern kcondvar_t md_cv; | |
85 extern int md_mtioctl_cnt; | |
86 | |
87 daemon_request_t mirror_timeout; | |
88 static daemon_request_t hotspare_request; | |
89 static daemon_request_t mn_hs_request[MD_MAXSETS]; /* Multinode hs req */ | |
90 | |
91 int md_mirror_mcs_buf_off; | |
92 | |
93 /* Flags for mdmn_ksend_message to allow debugging */ | |
94 int md_mirror_msg_flags; | |
95 | |
96 #ifdef DEBUG | |
97 /* Flag to switch on debug messages */ | |
98 int mirror_debug_flag = 0; | |
99 #endif | |
100 | |
101 /* | |
102 * Struct used to hold count of DMR reads and the timestamp of last DMR read | |
103 * It is used to verify, using a debugger, that the DMR read ioctl has been | |
104 * executed. | |
105 */ | |
106 dmr_stats_t mirror_dmr_stats = {0, 0}; | |
107 | |
108 /* | |
109 * Mutex protecting list of non-failfast drivers. | |
110 */ | |
111 static kmutex_t non_ff_drv_mutex; | |
112 static char **non_ff_drivers = NULL; | |
113 | |
114 extern major_t md_major; | |
115 | |
116 /* | |
117 * Write-On-Write memory pool. | |
118 */ | |
119 static void copy_write_cont(wowhdr_t *wowhdr); | |
120 static kmem_cache_t *mirror_wowblk_cache = NULL; | |
121 static int md_wowbuf_size = 16384; | |
122 static size_t md_wowblk_size; | |
123 | |
124 /* | |
125 * This is a flag that allows: | |
126 * - disabling the write-on-write mechanism. | |
127 * - logging occurrences of write-on-write | |
128 * - switching wow handling procedure processing | |
129 * Counter for occurences of WOW. | |
130 */ | |
131 static uint_t md_mirror_wow_flg = 0; | |
132 static int md_mirror_wow_cnt = 0; | |
133 | |
134 /* | |
135 * Tunable to enable/disable dirty region | |
136 * processing when closing down a mirror. | |
137 */ | |
138 static int new_resync = 1; | |
139 kmem_cache_t *mirror_parent_cache = NULL; | |
140 kmem_cache_t *mirror_child_cache = NULL; | |
141 | |
142 extern int md_ff_disable; /* disable failfast */ | |
143 | |
144 static int mirror_map_write(mm_unit_t *, md_mcs_t *, md_mps_t *, int); | |
145 static void mirror_read_strategy(buf_t *, int, void *); | |
146 static void mirror_write_strategy(buf_t *, int, void *); | |
147 static void become_owner(daemon_queue_t *); | |
148 static int mirror_done(struct buf *cb); | |
149 static int mirror_done_common(struct buf *cb); | |
150 static void clear_retry_error(struct buf *cb); | |
151 | |
152 /* | |
153 * patchables | |
154 */ | |
155 int md_min_rr_size = 200; /* 2000 blocks, or 100k */ | |
156 int md_def_num_rr = 1000; /* Default number of dirty regions */ | |
157 | |
158 /* | |
159 * patchable to change delay before rescheduling mirror ownership request. | |
160 * Value is clock ticks, default 0.5 seconds | |
161 */ | |
162 clock_t md_mirror_owner_to = 500000; | |
163 | |
164 /*ARGSUSED1*/ | |
165 static int | |
166 mirror_parent_constructor(void *p, void *d1, int d2) | |
167 { | |
168 mutex_init(&((md_mps_t *)p)->ps_mx, NULL, MUTEX_DEFAULT, NULL); | |
169 return (0); | |
170 } | |
171 | |
172 static void | |
173 mirror_parent_init(md_mps_t *ps) | |
174 { | |
175 bzero(ps, offsetof(md_mps_t, ps_mx)); | |
176 } | |
177 | |
178 /*ARGSUSED1*/ | |
179 static void | |
180 mirror_parent_destructor(void *p, void *d) | |
181 { | |
182 mutex_destroy(&((md_mps_t *)p)->ps_mx); | |
183 } | |
184 | |
185 /*ARGSUSED1*/ | |
186 static int | |
187 mirror_child_constructor(void *p, void *d1, int d2) | |
188 { | |
189 bioinit(&((md_mcs_t *)p)->cs_buf); | |
190 return (0); | |
191 } | |
192 | |
193 void | |
194 mirror_child_init(md_mcs_t *cs) | |
195 { | |
196 cs->cs_ps = NULL; | |
197 cs->cs_mdunit = 0; | |
198 md_bioreset(&cs->cs_buf); | |
199 } | |
200 | |
201 /*ARGSUSED1*/ | |
202 static void | |
203 mirror_child_destructor(void *p, void *d) | |
204 { | |
205 biofini(&((md_mcs_t *)p)->cs_buf); | |
206 } | |
207 | |
208 static void | |
209 mirror_wowblk_init(wowhdr_t *p) | |
210 { | |
211 bzero(p, md_wowblk_size); | |
212 } | |
213 | |
214 static void | |
215 send_poke_hotspares_msg(daemon_request_t *drq) | |
216 { | |
217 int rval; | |
218 md_mn_msg_pokehsp_t pokehsp; | |
219 md_mn_kresult_t *kresult; | |
220 set_t setno = (set_t)drq->dq.qlen; | |
221 | |
222 pokehsp.pokehsp_setno = setno; | |
223 | |
224 kresult = kmem_alloc(sizeof (md_mn_kresult_t), KM_SLEEP); | |
225 rval = mdmn_ksend_message(setno, MD_MN_MSG_POKE_HOTSPARES, | |
226 MD_MSGF_NO_LOG | MD_MSGF_NO_BCAST, (char *)&pokehsp, | |
227 sizeof (pokehsp), kresult); | |
228 | |
229 if (!MDMN_KSEND_MSG_OK(rval, kresult)) { | |
230 mdmn_ksend_show_error(rval, kresult, "POKE_HOTSPARES"); | |
231 cmn_err(CE_PANIC, | |
232 "ksend_message failure: POKE_HOTSPARES"); | |
233 } | |
234 kmem_free(kresult, sizeof (md_mn_kresult_t)); | |
235 | |
236 /* Allow further requests to use this set's queue structure */ | |
237 mutex_enter(&drq->dr_mx); | |
238 drq->dr_pending = 0; | |
239 mutex_exit(&drq->dr_mx); | |
240 } | |
241 | |
242 /* | |
243 * Send a poke_hotspares message to the master node. To avoid swamping the | |
244 * commd handler with requests we only send a message if there is not one | |
245 * already outstanding. We punt the request to a separate thread context as | |
246 * cannot afford to block waiting on the request to be serviced. This is | |
247 * essential when a reconfig cycle is in progress as any open() of a multinode | |
248 * metadevice may result in a livelock. | |
249 */ | |
250 static void | |
251 send_poke_hotspares(set_t setno) | |
252 { | |
253 daemon_request_t *drq = &mn_hs_request[setno]; | |
254 | |
255 mutex_enter(&drq->dr_mx); | |
256 if (drq->dr_pending == 0) { | |
257 drq->dr_pending = 1; | |
258 drq->dq.qlen = (int)setno; | |
259 daemon_request(&md_mhs_daemon, | |
260 send_poke_hotspares_msg, (daemon_queue_t *)drq, REQ_OLD); | |
261 } | |
262 mutex_exit(&drq->dr_mx); | |
263 } | |
264 | |
265 void | |
266 mirror_set_sm_state( | |
267 mm_submirror_t *sm, | |
268 mm_submirror_ic_t *smic, | |
269 sm_state_t newstate, | |
270 int force) | |
271 { | |
272 int compcnt; | |
273 int i; | |
274 int errcnt; | |
275 sm_state_t origstate; | |
276 md_m_shared_t *shared; | |
277 | |
278 if (force) { | |
279 sm->sm_state = newstate; | |
280 uniqtime32(&sm->sm_timestamp); | |
281 return; | |
282 } | |
283 | |
284 origstate = newstate; | |
285 | |
286 compcnt = (*(smic->sm_get_component_count))(sm->sm_dev, sm); | |
287 for (i = 0, errcnt = 0; i < compcnt; i++) { | |
288 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx)) | |
289 (sm->sm_dev, sm, i); | |
290 if (shared->ms_state & (CS_ERRED | CS_LAST_ERRED)) | |
291 newstate |= SMS_COMP_ERRED; | |
292 if (shared->ms_state & (CS_RESYNC)) | |
293 newstate |= SMS_COMP_RESYNC; | |
294 if (shared->ms_state & CS_ERRED) | |
295 errcnt++; | |
296 } | |
297 | |
298 if ((newstate & (SMS_COMP_ERRED | SMS_COMP_RESYNC)) != 0) | |
299 newstate &= ~origstate; | |
300 | |
301 if (errcnt == compcnt) | |
302 newstate |= SMS_ALL_ERRED; | |
303 else | |
304 newstate &= ~SMS_ALL_ERRED; | |
305 | |
306 sm->sm_state = newstate; | |
307 uniqtime32(&sm->sm_timestamp); | |
308 } | |
309 | |
310 static int | |
311 mirror_geterror(mm_unit_t *un, int *smi, int *cip, int clr_error, | |
312 int frm_probe) | |
313 { | |
314 mm_submirror_t *sm; | |
315 mm_submirror_ic_t *smic; | |
316 md_m_shared_t *shared; | |
317 int ci; | |
318 int i; | |
319 int compcnt; | |
320 int open_comp; /* flag for open component */ | |
321 | |
322 for (i = *smi; i < NMIRROR; i++) { | |
323 sm = &un->un_sm[i]; | |
324 smic = &un->un_smic[i]; | |
325 | |
326 if (!SMS_IS(sm, SMS_INUSE)) | |
327 continue; | |
328 | |
329 compcnt = (*(smic->sm_get_component_count)) (sm->sm_dev, un); | |
330 for (ci = *cip; ci < compcnt; ci++) { | |
331 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx)) | |
332 (sm->sm_dev, sm, ci); | |
333 /* | |
334 * if called from any routine but probe, we check for | |
335 * MDM_S_ISOPEN flag. Since probe does a pseduo open, | |
336 * it sets MDM_S_PROBEOPEN flag and we test for this | |
337 * flag. They are both exclusive tests. | |
338 */ | |
339 open_comp = (frm_probe) ? | |
340 (shared->ms_flags & MDM_S_PROBEOPEN): | |
341 (shared->ms_flags & MDM_S_ISOPEN); | |
342 if ((shared->ms_flags & MDM_S_IOERR || !open_comp) && | |
343 ((shared->ms_state == CS_OKAY) || | |
344 (shared->ms_state == CS_RESYNC))) { | |
345 if (clr_error) { | |
346 shared->ms_flags &= ~MDM_S_IOERR; | |
347 } | |
348 *cip = ci; | |
349 *smi = i; | |
350 return (1); | |
351 } | |
352 | |
353 if (clr_error && (shared->ms_flags & MDM_S_IOERR)) { | |
354 shared->ms_flags &= ~MDM_S_IOERR; | |
355 } | |
356 } | |
357 | |
358 *cip = 0; | |
359 } | |
360 return (0); | |
361 } | |
362 | |
363 /*ARGSUSED*/ | |
364 static void | |
365 mirror_run_queue(void *d) | |
366 { | |
367 if (!(md_status & MD_GBL_DAEMONS_LIVE)) | |
368 md_daemon(1, &md_done_daemon); | |
369 } | |
370 /* | |
371 * check_comp_4_hotspares | |
372 * | |
373 * This function attempts to allocate a hotspare for this component if the | |
374 * component is in error. In a MN set, the function can be called in 2 modes. | |
375 * It can be called either when a component error has been detected or when a | |
376 * new hotspare has been allocated. In this case, MD_HOTSPARE_XMIT is set | |
377 * in flags and the request is sent to all nodes. | |
378 * The handler on each of the nodes then calls this function with | |
379 * MD_HOTSPARE_XMIT unset and the hotspare allocation is then performed. | |
380 * | |
381 * For non-MN sets the function simply attempts to allocate a hotspare. | |
382 * | |
383 * On entry, the following locks are held | |
384 * mirror_md_ops.md_link_rw (if flags has MD_HOTSPARE_LINKHELD set) | |
385 * md_unit_writerlock | |
386 * | |
387 * Returns 0 if ok | |
388 * 1 if the unit containing the component has been cleared while | |
389 * the mdmn_ksend_message() was being executed | |
390 */ | |
391 extern int | |
392 check_comp_4_hotspares( | |
393 mm_unit_t *un, | |
394 int smi, | |
395 int ci, | |
396 uint_t flags, | |
397 mddb_recid_t hs_id, /* Only used by MN disksets */ | |
398 IOLOCK *lockp /* can be NULL */ | |
399 ) | |
400 { | |
401 mm_submirror_t *sm; | |
402 mm_submirror_ic_t *smic; | |
403 md_m_shared_t *shared; | |
404 mddb_recid_t recids[6]; | |
405 minor_t mnum; | |
406 intptr_t (*hs_dev)(); | |
407 void (*hs_done)(); | |
408 void *hs_data; | |
409 md_error_t mde = mdnullerror; | |
410 set_t setno; | |
411 md_mn_msg_allochsp_t allochspmsg; | |
412 md_mn_kresult_t *kresult; | |
413 mm_unit_t *new_un; | |
414 int rval; | |
415 | |
416 mnum = MD_SID(un); | |
417 setno = MD_UN2SET(un); | |
418 sm = &un->un_sm[smi]; | |
419 smic = &un->un_smic[smi]; | |
420 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx)) | |
421 (sm->sm_dev, sm, ci); | |
422 | |
423 if (shared->ms_state != CS_ERRED) | |
424 return (0); | |
425 | |
426 /* Don't start a new component resync if a resync is already running. */ | |
427 if (MD_STATUS(un) & MD_UN_RESYNC_ACTIVE) | |
428 return (0); | |
429 | |
430 if (MD_MNSET_SETNO(setno) && (flags & MD_HOTSPARE_XMIT)) { | |
431 uint_t msgflags; | |
432 md_mn_msgtype_t msgtype; | |
433 | |
434 /* Send allocate hotspare message to all nodes */ | |
435 | |
436 allochspmsg.msg_allochsp_mnum = un->c.un_self_id; | |
437 allochspmsg.msg_allochsp_sm = smi; | |
438 allochspmsg.msg_allochsp_comp = ci; | |
439 allochspmsg.msg_allochsp_hs_id = shared->ms_hs_id; | |
440 | |
441 /* | |
442 * Before calling mdmn_ksend_message(), release locks | |
443 * Can never be in the context of an ioctl. | |
444 */ | |
445 md_unit_writerexit(MDI_UNIT(mnum)); | |
446 if (flags & MD_HOTSPARE_LINKHELD) | |
447 rw_exit(&mirror_md_ops.md_link_rw.lock); | |
448 #ifdef DEBUG | |
449 if (mirror_debug_flag) | |
450 printf("send alloc hotspare, flags=0x%x %x, %x, %x, %x\n", | |
451 flags, | |
452 allochspmsg.msg_allochsp_mnum, | |
453 allochspmsg.msg_allochsp_sm, | |
454 allochspmsg.msg_allochsp_comp, | |
455 allochspmsg.msg_allochsp_hs_id); | |
456 #endif | |
457 if (flags & MD_HOTSPARE_WMUPDATE) { | |
458 msgtype = MD_MN_MSG_ALLOCATE_HOTSPARE2; | |
459 /* | |
460 * When coming from an update of watermarks, there | |
461 * must already be a message logged that triggered | |
462 * this action. So, no need to log this message, too. | |
463 */ | |
464 msgflags = MD_MSGF_NO_LOG; | |
465 } else { | |
466 msgtype = MD_MN_MSG_ALLOCATE_HOTSPARE; | |
467 msgflags = MD_MSGF_DEFAULT_FLAGS; | |
468 } | |
469 | |
470 kresult = kmem_alloc(sizeof (md_mn_kresult_t), KM_SLEEP); | |
471 rval = mdmn_ksend_message(setno, msgtype, msgflags, | |
472 (char *)&allochspmsg, sizeof (allochspmsg), | |
473 kresult); | |
474 | |
475 if (!MDMN_KSEND_MSG_OK(rval, kresult)) { | |
476 #ifdef DEBUG | |
477 if (mirror_debug_flag) | |
478 mdmn_ksend_show_error(rval, kresult, | |
479 "ALLOCATE HOTSPARE"); | |
480 #endif | |
481 /* | |
482 * If message is sent ok but exitval indicates an error | |
483 * it must be because the mirror has been cleared. In | |
484 * this case re-obtain lock and return an error | |
485 */ | |
486 if ((rval == 0) && (kresult->kmmr_exitval != 0)) { | |
487 if (flags & MD_HOTSPARE_LINKHELD) { | |
488 rw_enter(&mirror_md_ops.md_link_rw.lock, | |
489 RW_READER); | |
490 } | |
491 kmem_free(kresult, sizeof (md_mn_kresult_t)); | |
492 return (1); | |
493 } | |
494 cmn_err(CE_PANIC, | |
495 "ksend_message failure: ALLOCATE_HOTSPARE"); | |
496 } | |
497 kmem_free(kresult, sizeof (md_mn_kresult_t)); | |
498 | |
499 /* | |
500 * re-obtain the locks | |
501 */ | |
502 if (flags & MD_HOTSPARE_LINKHELD) | |
503 rw_enter(&mirror_md_ops.md_link_rw.lock, RW_READER); | |
504 new_un = md_unit_writerlock(MDI_UNIT(mnum)); | |
505 | |
506 /* | |
507 * As we had to release the locks in order to send the | |
508 * message to all nodes, we need to check to see if the | |
509 * unit has changed. If it has we release the writerlock | |
510 * and return fail. | |
511 */ | |
512 if ((new_un != un) || (un->c.un_type != MD_METAMIRROR)) { | |
513 md_unit_writerexit(MDI_UNIT(mnum)); | |
514 return (1); | |
515 } | |
516 } else { | |
517 if (MD_MNSET_SETNO(setno)) { | |
518 /* | |
519 * If 2 or more nodes simultaneously see a | |
520 * component failure, these nodes will each | |
521 * send an ALLOCATE_HOTSPARE[2] message. | |
522 * The first message will allocate the hotspare | |
523 * and the subsequent messages should do nothing. | |
524 * | |
525 * If a slave node doesn't have a hotspare allocated | |
526 * at the time the message is initiated, then the | |
527 * passed in hs_id will be 0. If the node | |
528 * executing this routine has a component shared | |
529 * ms_hs_id of non-zero, but the message shows a | |
530 * hs_id of 0, then just return since a hotspare | |
531 * has already been allocated for this failing | |
532 * component. When the slave node returns from | |
533 * the ksend_message the hotspare will have | |
534 * already been allocated. | |
535 * | |
536 * If the slave node does send an hs_id of non-zero, | |
537 * and the slave node's hs_id matches this node's | |
538 * ms_hs_id, then the hotspare has error'd and | |
539 * should be replaced. | |
540 * | |
541 * If the slave node sends an hs_id of non-zero and | |
542 * this node has a different shared ms_hs_id, then | |
543 * just return since this hotspare has already | |
544 * been hotspared. | |
545 */ | |
546 if (shared->ms_hs_id != 0) { | |
547 if (hs_id == 0) { | |
548 #ifdef DEBUG | |
549 if (mirror_debug_flag) { | |
550 printf("check_comp_4_hotspares" | |
551 "(NOXMIT), short circuit " | |
552 "hs_id=0x%x, " | |
553 "ms_hs_id=0x%x\n", | |
554 hs_id, shared->ms_hs_id); | |
555 } | |
556 #endif | |
557 return (0); | |
558 } | |
559 if (hs_id != shared->ms_hs_id) { | |
560 #ifdef DEBUG | |
561 if (mirror_debug_flag) { | |
562 printf("check_comp_4_hotspares" | |
563 "(NOXMIT), short circuit2 " | |
564 "hs_id=0x%x, " | |
565 "ms_hs_id=0x%x\n", | |
566 hs_id, shared->ms_hs_id); | |
567 } | |
568 #endif | |
569 return (0); | |
570 } | |
571 } | |
572 } | |
573 | |
574 sm = &un->un_sm[smi]; | |
575 hs_dev = md_get_named_service(sm->sm_dev, 0, | |
576 "hotspare device", 0); | |
577 if ((*hs_dev)(sm->sm_dev, 0, ci, recids, 6, &hs_done, | |
578 &hs_data) != 0) | |
579 return (0); | |
580 | |
581 /* | |
582 * set_sm_comp_state() commits the modified records. | |
583 * As we don't transmit the changes, no need to drop the lock. | |
584 */ | |
585 set_sm_comp_state(un, smi, ci, CS_RESYNC, recids, | |
586 MD_STATE_NO_XMIT, (IOLOCK *)NULL); | |
587 | |
588 (*hs_done)(sm->sm_dev, hs_data); | |
589 | |
590 mirror_check_failfast(mnum); | |
591 | |
592 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_HOTSPARED, SVM_TAG_METADEVICE, | |
593 setno, MD_SID(un)); | |
594 | |
595 /* | |
596 * For a multi-node set we need to reset the un_rs_type, | |
597 * un_rs_resync_done and un_rs_resync_2_do fields as the | |
598 * hot-spare resync must copy all applicable data. | |
599 */ | |
600 if (MD_MNSET_SETNO(setno)) { | |
601 un->un_rs_type = MD_RS_NONE; | |
602 un->un_rs_resync_done = 0; | |
603 un->un_rs_resync_2_do = 0; | |
604 } | |
605 | |
606 /* | |
607 * Must drop writer lock since mirror_resync_unit will | |
608 * open devices and must be able to grab readerlock. | |
609 * Don't need to drop IOLOCK since any descendent routines | |
610 * calling ksend_messages will drop the IOLOCK as needed. | |
611 * | |
612 */ | |
613 if (lockp) { | |
614 md_ioctl_writerexit(lockp); | |
615 } else { | |
616 md_unit_writerexit(MDI_UNIT(mnum)); | |
617 } | |
618 | |
619 /* start resync */ | |
620 (void) mirror_resync_unit(mnum, NULL, &mde, lockp); | |
621 | |
622 if (lockp) { | |
623 new_un = md_ioctl_writerlock(lockp, MDI_UNIT(mnum)); | |
624 } else { | |
625 new_un = md_unit_writerlock(MDI_UNIT(mnum)); | |
626 } | |
627 } | |
628 return (0); | |
629 } | |
630 | |
631 /* | |
632 * check_unit_4_hotspares | |
633 * | |
634 * For a given mirror, allocate hotspares, if available for any components | |
635 * that are in error | |
636 * | |
637 * Returns 0 if ok | |
638 * 1 if check_comp_4_hotspares returns non-zero. This will only | |
639 * happen for a MN unit where the unit has been cleared while | |
640 * the allocate hotspare message is sent to all nodes. | |
641 */ | |
642 static int | |
643 check_unit_4_hotspares(mm_unit_t *un, int flags) | |
644 { | |
645 mm_submirror_t *sm; | |
646 mm_submirror_ic_t *smic; | |
647 int ci; | |
648 int i; | |
649 int compcnt; | |
650 | |
651 if (MD_STATUS(un) & MD_UN_RESYNC_ACTIVE) | |
652 return (0); | |
653 | |
654 for (i = 0; i < NMIRROR; i++) { | |
655 sm = &un->un_sm[i]; | |
656 smic = &un->un_smic[i]; | |
657 if (!SMS_IS(sm, SMS_INUSE)) | |
658 continue; | |
659 compcnt = (*(smic->sm_get_component_count)) (sm->sm_dev, sm); | |
660 for (ci = 0; ci < compcnt; ci++) { | |
661 md_m_shared_t *shared; | |
662 | |
663 shared = (md_m_shared_t *) | |
664 (*(smic->sm_shared_by_indx))(sm->sm_dev, | |
665 sm, ci); | |
666 /* | |
667 * Never called from ioctl context, so pass in | |
668 * (IOLOCK *)NULL. Pass through flags from calling | |
669 * routine, also setting XMIT flag. | |
670 */ | |
671 if (check_comp_4_hotspares(un, i, ci, | |
672 (MD_HOTSPARE_XMIT | flags), | |
673 shared->ms_hs_id, (IOLOCK *)NULL) != 0) | |
674 return (1); | |
675 } | |
676 } | |
677 return (0); | |
678 } | |
679 | |
680 static void | |
681 check_4_hotspares(daemon_request_t *drq) | |
682 { | |
683 mdi_unit_t *ui; | |
684 mm_unit_t *un; | |
685 md_link_t *next; | |
686 int x; | |
687 | |
688 mutex_enter(&drq->dr_mx); /* clear up front so can poke */ | |
689 drq->dr_pending = 0; /* again in low level routine if */ | |
690 mutex_exit(&drq->dr_mx); /* something found to do */ | |
691 | |
692 /* | |
693 * Used to have a problem here. The disksets weren't marked as being | |
694 * MNHOLD. This opened a window where we could be searching for | |
695 * hotspares and have the disk set unloaded (released) from under | |
696 * us causing a panic in stripe_component_count(). | |
697 * The way to prevent that is to mark the set MNHOLD which prevents | |
698 * any diskset from being released while we are scanning the mirrors, | |
699 * submirrors and components. | |
700 */ | |
701 | |
702 for (x = 0; x < md_nsets; x++) | |
703 md_holdset_enter(x); | |
704 | |
705 rw_enter(&mirror_md_ops.md_link_rw.lock, RW_READER); | |
706 for (next = mirror_md_ops.md_head; next != NULL; next = next->ln_next) { | |
707 ui = MDI_UNIT(next->ln_id); | |
708 | |
709 un = (mm_unit_t *)md_unit_readerlock(ui); | |
710 | |
711 /* | |
712 * Only check the unit if we are the master for this set | |
713 * For an MN set, poke_hotspares() is only effective on the | |
714 * master | |
715 */ | |
716 if (MD_MNSET_SETNO(MD_UN2SET(un)) && | |
717 md_set[MD_UN2SET(un)].s_am_i_master == 0) { | |
718 md_unit_readerexit(ui); | |
719 continue; | |
720 } | |
721 if (MD_STATUS(un) & MD_UN_RESYNC_ACTIVE) { | |
722 md_unit_readerexit(ui); | |
723 continue; | |
724 } | |
725 md_unit_readerexit(ui); | |
726 | |
727 un = (mm_unit_t *)md_unit_writerlock(ui); | |
728 /* | |
729 * check_unit_4_hotspares will exit 1 if the unit has been | |
730 * removed during the process of allocating the hotspare. | |
731 * This can only happen for a MN metadevice. If unit no longer | |
732 * exists, no need to release writerlock | |
733 */ | |
734 if (check_unit_4_hotspares(un, MD_HOTSPARE_LINKHELD) == 0) | |
735 md_unit_writerexit(ui); | |
736 else { | |
737 /* | |
738 * If check_unit_4_hotspares failed, queue another | |
739 * request and break out of this one | |
740 */ | |
741 (void) poke_hotspares(); | |
742 break; | |
743 } | |
744 } | |
745 rw_exit(&mirror_md_ops.md_link_rw.lock); | |
746 | |
747 for (x = 0; x < md_nsets; x++) | |
748 md_holdset_exit(x); | |
749 } | |
750 | |
751 /* | |
752 * poke_hotspares | |
753 * | |
754 * If there is not a pending poke_hotspares request pending, queue a requent | |
755 * to call check_4_hotspares(). This will scan all mirrors and attempt to | |
756 * allocate hotspares for all components in error. | |
757 */ | |
758 int | |
759 poke_hotspares() | |
760 { | |
761 mutex_enter(&hotspare_request.dr_mx); | |
762 if (hotspare_request.dr_pending == 0) { | |
763 hotspare_request.dr_pending = 1; | |
764 daemon_request(&md_mhs_daemon, | |
765 check_4_hotspares, | |
766 (daemon_queue_t *)&hotspare_request, REQ_OLD); | |
767 } | |
768 mutex_exit(&hotspare_request.dr_mx); | |
769 return (0); | |
770 } | |
771 | |
772 static void | |
773 free_all_ecomps(err_comp_t *ecomp) | |
774 { | |
775 err_comp_t *d; | |
776 | |
777 while (ecomp != NULL) { | |
778 d = ecomp; | |
779 ecomp = ecomp->ec_next; | |
780 kmem_free(d, sizeof (err_comp_t)); | |
781 } | |
782 } | |
783 | |
784 /* | |
785 * NAME: mirror_openfail_console_info | |
786 * | |
787 * DESCRIPTION: Prints a informative message to the console when mirror | |
788 * cannot be opened. | |
789 * | |
790 * PARAMETERS: mm_unit_t un - pointer to mirror unit structure | |
791 * int smi - submirror index | |
792 * int ci - component index | |
793 */ | |
794 | |
795 void | |
796 mirror_openfail_console_info(mm_unit_t *un, int smi, int ci) | |
797 { | |
798 void (*get_dev)(); | |
799 ms_cd_info_t cd; | |
800 md_dev64_t tmpdev; | |
801 | |
802 tmpdev = un->un_sm[smi].sm_dev; | |
803 get_dev = (void (*)())md_get_named_service(tmpdev, 0, "get device", 0); | |
804 if (get_dev != NULL) { | |
805 (void) (*get_dev)(tmpdev, smi, ci, &cd); | |
806 cmn_err(CE_WARN, "md %s: open error on %s", | |
807 md_shortname(MD_SID(un)), | |
808 md_devname(MD_UN2SET(un), cd.cd_dev, | |
809 NULL, 0)); | |
810 } else { | |
811 cmn_err(CE_WARN, "md %s: open error", | |
812 md_shortname(MD_SID(un))); | |
813 } | |
814 } | |
815 | |
816 static int | |
817 mirror_close_all_devs(mm_unit_t *un, int md_cflags) | |
818 { | |
819 int i; | |
820 md_dev64_t dev; | |
821 | |
822 for (i = 0; i < NMIRROR; i++) { | |
823 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) | |
824 continue; | |
825 dev = un->un_sm[i].sm_dev; | |
826 md_layered_close(dev, md_cflags); | |
827 } | |
828 return (0); | |
829 } | |
830 | |
831 /* | |
832 * Keep track of drivers that don't support failfast. We use this so that | |
833 * we only log one diagnostic message for each of these drivers, no matter | |
834 * how many times we run the mirror_check_failfast function. | |
835 * Return 1 if this is a new driver that does not support failfast, | |
836 * return 0 if we have already seen this non-failfast driver. | |
837 */ | |
838 static int | |
839 new_non_ff_driver(const char *s) | |
840 { | |
841 mutex_enter(&non_ff_drv_mutex); | |
842 if (non_ff_drivers == NULL) { | |
843 non_ff_drivers = (char **)kmem_alloc(2 * sizeof (char *), | |
844 KM_NOSLEEP); | |
845 if (non_ff_drivers == NULL) { | |
846 mutex_exit(&non_ff_drv_mutex); | |
847 return (1); | |
848 } | |
849 | |
850 non_ff_drivers[0] = (char *)kmem_alloc(strlen(s) + 1, KM_NOSLEEP); | |
851 if (non_ff_drivers[0] == NULL) { | |
852 kmem_free(non_ff_drivers, 2 * sizeof (char *)); | |
853 non_ff_drivers = NULL; | |
854 mutex_exit(&non_ff_drv_mutex); | |
855 return (1); | |
856 } | |
857 | |
858 (void) strcpy(non_ff_drivers[0], s); | |
859 non_ff_drivers[1] = NULL; | |
860 | |
861 } else { | |
862 int i; | |
863 char **tnames; | |
864 char **tmp; | |
865 | |
866 for (i = 0; non_ff_drivers[i] != NULL; i++) { | |
867 if (strcmp(s, non_ff_drivers[i]) == 0) { | |
868 mutex_exit(&non_ff_drv_mutex); | |
869 return (0); | |
870 } | |
871 } | |
872 | |
873 /* allow for new element and null */ | |
874 i += 2; | |
875 tnames = (char **)kmem_alloc(i * sizeof (char *), KM_NOSLEEP); | |
876 if (tnames == NULL) { | |
877 mutex_exit(&non_ff_drv_mutex); | |
878 return (1); | |
879 } | |
880 | |
881 for (i = 0; non_ff_drivers[i] != NULL; i++) | |
882 tnames[i] = non_ff_drivers[i]; | |
883 | |
884 tnames[i] = (char *)kmem_alloc(strlen(s) + 1, KM_NOSLEEP); | |
885 if (tnames[i] == NULL) { | |
886 /* adjust i so that it is the right count to free */ | |
887 kmem_free(tnames, (i + 2) * sizeof (char *)); | |
888 mutex_exit(&non_ff_drv_mutex); | |
889 return (1); | |
890 } | |
891 | |
892 (void) strcpy(tnames[i++], s); | |
893 tnames[i] = NULL; | |
894 | |
895 tmp = non_ff_drivers; | |
896 non_ff_drivers = tnames; | |
897 /* i now represents the count we previously alloced */ | |
898 kmem_free(tmp, i * sizeof (char *)); | |
899 } | |
900 mutex_exit(&non_ff_drv_mutex); | |
901 | |
902 return (1); | |
903 } | |
904 | |
905 /* | |
906 * Check for the "ddi-failfast-supported" devtree property on each submirror | |
907 * component to indicate if we should do I/O to that submirror with the | |
908 * B_FAILFAST flag set or not. This check is made at various state transitions | |
909 * in the mirror code (e.g. open, enable, hotspare, etc.). Sometimes we | |
910 * only need to check one drive (e.g. hotspare) but since the check is | |
911 * fast and infrequent and sometimes needs to be done on all components we | |
912 * just check all components on each call. | |
913 */ | |
914 void | |
915 mirror_check_failfast(minor_t mnum) | |
916 { | |
917 int i; | |
918 mm_unit_t *un; | |
919 | |
920 if (md_ff_disable) | |
921 return; | |
922 | |
923 un = MD_UNIT(mnum); | |
924 | |
925 for (i = 0; i < NMIRROR; i++) { | |
926 int ci; | |
927 int cnt; | |
928 int ff = 1; | |
929 mm_submirror_t *sm; | |
930 mm_submirror_ic_t *smic; | |
931 void (*get_dev)(); | |
932 | |
933 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) | |
934 continue; | |
935 | |
936 sm = &un->un_sm[i]; | |
937 smic = &un->un_smic[i]; | |
938 | |
939 get_dev = (void (*)())md_get_named_service(sm->sm_dev, 0, | |
940 "get device", 0); | |
941 | |
942 cnt = (*(smic->sm_get_component_count))(sm->sm_dev, sm); | |
943 for (ci = 0; ci < cnt; ci++) { | |
944 int found = 0; | |
945 dev_t ci_dev; | |
946 major_t major; | |
947 dev_info_t *devi; | |
948 ms_cd_info_t cd; | |
949 | |
950 /* this already returns the hs dev if the device is spared */ | |
951 (void) (*get_dev)(sm->sm_dev, sm, ci, &cd); | |
952 | |
953 ci_dev = md_dev64_to_dev(cd.cd_dev); | |
954 major = getmajor(ci_dev); | |
955 | |
956 if (major == md_major) { | |
957 /* this component must be a soft partition; get real dev */ | |
958 minor_t dev_mnum; | |
959 mdi_unit_t *ui; | |
960 mp_unit_t *un; | |
961 set_t setno; | |
962 side_t side; | |
963 md_dev64_t tmpdev; | |
964 | |
965 ui = MDI_UNIT(getminor(ci_dev)); | |
966 | |
967 /* grab necessary lock */ | |
968 un = (mp_unit_t *)md_unit_readerlock(ui); | |
969 | |
970 dev_mnum = MD_SID(un); | |
971 setno = MD_MIN2SET(dev_mnum); | |
972 side = mddb_getsidenum(setno); | |
973 | |
974 tmpdev = un->un_dev; | |
975 | |
976 /* Get dev by device id */ | |
977 if (md_devid_found(setno, side, un->un_key) == 1) { | |
978 tmpdev = md_resolve_bydevid(dev_mnum, tmpdev, | |
979 un->un_key); | |
980 } | |
981 | |
982 md_unit_readerexit(ui); | |
983 | |
984 ci_dev = md_dev64_to_dev(tmpdev); | |
985 major = getmajor(ci_dev); | |
986 } | |
987 | |
988 if (ci_dev != NODEV32 && | |
989 (devi = e_ddi_hold_devi_by_dev(ci_dev, 0)) != NULL) { | |
990 ddi_prop_op_t prop_op = PROP_LEN_AND_VAL_BUF; | |
991 int propvalue = 0; | |
992 int proplength = sizeof (int); | |
993 int error; | |
994 struct cb_ops *cb; | |
995 | |
996 if ((cb = devopsp[major]->devo_cb_ops) != NULL) { | |
997 error = (*cb->cb_prop_op)(DDI_DEV_T_ANY, devi, prop_op, | |
998 DDI_PROP_NOTPROM|DDI_PROP_DONTPASS, | |
999 "ddi-failfast-supported", | |
1000 (caddr_t)&propvalue, &proplength); | |
1001 | |
1002 if (error == DDI_PROP_SUCCESS) | |
1003 found = 1; | |
1004 } | |
1005 | |
1006 if (!found && new_non_ff_driver(ddi_driver_name(devi))) | |
1007 cmn_err(CE_NOTE, "!md: B_FAILFAST I/O disabled on %s", | |
1008 ddi_driver_name(devi)); | |
1009 | |
1010 ddi_release_devi(devi); | |
1011 } | |
1012 | |
1013 /* All components must support failfast in the submirror. */ | |
1014 if (!found) { | |
1015 ff = 0; | |
1016 break; | |
1017 } | |
1018 } | |
1019 | |
1020 if (ff) { | |
1021 sm->sm_flags |= MD_SM_FAILFAST; | |
1022 } else { | |
1023 sm->sm_flags &= ~MD_SM_FAILFAST; | |
1024 } | |
1025 } | |
1026 } | |
1027 | |
1028 /* | |
1029 * Return true if the submirror is unavailable. | |
1030 * If any of the submirror components are opened then the submirror cannot | |
1031 * be unavailable (MD_INACCESSIBLE). | |
1032 * If any of the components are already in the errored state, then the submirror | |
1033 * cannot be unavailable (MD_INACCESSIBLE). | |
1034 */ | |
1035 static bool_t | |
1036 submirror_unavailable(mm_unit_t *un, int smi, int from_probe) | |
1037 { | |
1038 mm_submirror_t *sm; | |
1039 mm_submirror_ic_t *smic; | |
1040 md_m_shared_t *shared; | |
1041 int ci; | |
1042 int compcnt; | |
1043 | |
1044 sm = &un->un_sm[smi]; | |
1045 smic = &un->un_smic[smi]; | |
1046 | |
1047 compcnt = (*(smic->sm_get_component_count)) (sm->sm_dev, un); | |
1048 for (ci = 0; ci < compcnt; ci++) { | |
1049 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx)) | |
1050 (sm->sm_dev, sm, ci); | |
1051 if (from_probe) { | |
1052 if (shared->ms_flags & MDM_S_PROBEOPEN) | |
1053 return (B_FALSE); | |
1054 } else { | |
1055 if (shared->ms_flags & MDM_S_ISOPEN) | |
1056 return (B_FALSE); | |
1057 } | |
1058 if (shared->ms_state == CS_ERRED || | |
1059 shared->ms_state == CS_LAST_ERRED) | |
1060 return (B_FALSE); | |
1061 } | |
1062 | |
1063 return (B_TRUE); | |
1064 } | |
1065 | |
1066 static int | |
1067 mirror_open_all_devs(minor_t mnum, int md_oflags, IOLOCK *lockp) | |
1068 { | |
1069 int i; | |
1070 mm_unit_t *un; | |
1071 mdi_unit_t *ui; | |
1072 int err; | |
1073 int smi; | |
1074 int ci; | |
1075 err_comp_t *c; | |
1076 err_comp_t *ecomps = NULL; | |
1077 int smmask = 0; | |
1078 set_t setno; | |
1079 int sm_cnt; | |
1080 int sm_unavail_cnt; | |
1081 | |
1082 mirror_check_failfast(mnum); | |
1083 | |
1084 un = MD_UNIT(mnum); | |
1085 ui = MDI_UNIT(mnum); | |
1086 setno = MD_UN2SET(un); | |
1087 | |
1088 for (i = 0; i < NMIRROR; i++) { | |
1089 md_dev64_t tmpdev = un->un_sm[i].sm_dev; | |
1090 | |
1091 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) | |
1092 continue; | |
1093 if (md_layered_open(mnum, &tmpdev, md_oflags)) | |
1094 smmask |= SMI2BIT(i); | |
1095 un->un_sm[i].sm_dev = tmpdev; | |
1096 } | |
1097 | |
1098 /* | |
1099 * If smmask is clear, all submirrors are accessible. Clear the | |
1100 * MD_INACCESSIBLE bit in this case. This bit is also cleared for the | |
1101 * mirror device. If smmask is set, we have to determine which of the | |
1102 * submirrors are in error. If no submirror is accessible we mark the | |
1103 * whole mirror as MD_INACCESSIBLE. | |
1104 */ | |
1105 if (smmask == 0) { | |
1106 if (lockp) { | |
1107 md_ioctl_readerexit(lockp); | |
1108 (void) md_ioctl_writerlock(lockp, ui); | |
1109 } else { | |
1110 md_unit_readerexit(ui); | |
1111 (void) md_unit_writerlock(ui); | |
1112 } | |
1113 ui->ui_tstate &= ~MD_INACCESSIBLE; | |
1114 if (lockp) { | |
1115 md_ioctl_writerexit(lockp); | |
1116 (void) md_ioctl_readerlock(lockp, ui); | |
1117 } else { | |
1118 md_unit_writerexit(ui); | |
1119 (void) md_unit_readerlock(ui); | |
1120 } | |
1121 | |
1122 for (i = 0; i < NMIRROR; i++) { | |
1123 md_dev64_t tmpdev; | |
1124 mdi_unit_t *sm_ui; | |
1125 | |
1126 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) | |
1127 continue; | |
1128 | |
1129 tmpdev = un->un_sm[i].sm_dev; | |
1130 sm_ui = MDI_UNIT(getminor(md_dev64_to_dev(tmpdev))); | |
1131 (void) md_unit_writerlock(sm_ui); | |
1132 sm_ui->ui_tstate &= ~MD_INACCESSIBLE; | |
1133 md_unit_writerexit(sm_ui); | |
1134 } | |
1135 | |
1136 return (0); | |
1137 } | |
1138 | |
1139 for (i = 0; i < NMIRROR; i++) { | |
1140 md_dev64_t tmpdev; | |
1141 | |
1142 if (!(smmask & SMI2BIT(i))) | |
1143 continue; | |
1144 | |
1145 tmpdev = un->un_sm[i].sm_dev; | |
1146 err = md_layered_open(mnum, &tmpdev, MD_OFLG_CONT_ERRS); | |
1147 un->un_sm[i].sm_dev = tmpdev; | |
1148 ASSERT(err == 0); | |
1149 } | |
1150 | |
1151 if (lockp) { | |
1152 md_ioctl_readerexit(lockp); | |
1153 un = (mm_unit_t *)md_ioctl_writerlock(lockp, ui); | |
1154 } else { | |
1155 md_unit_readerexit(ui); | |
1156 un = (mm_unit_t *)md_unit_writerlock(ui); | |
1157 } | |
1158 | |
1159 /* | |
1160 * We want to make sure the unavailable flag is not masking a real | |
1161 * error on the submirror. | |
1162 * For each submirror, | |
1163 * if all of the submirror components couldn't be opened and there | |
1164 * are no errors on the submirror, then set the unavailable flag | |
1165 * otherwise, clear unavailable. | |
1166 */ | |
1167 sm_cnt = 0; | |
1168 sm_unavail_cnt = 0; | |
1169 for (i = 0; i < NMIRROR; i++) { | |
1170 md_dev64_t tmpdev; | |
1171 mdi_unit_t *sm_ui; | |
1172 | |
1173 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) | |
1174 continue; | |
1175 | |
1176 sm_cnt++; | |
1177 tmpdev = un->un_sm[i].sm_dev; | |
1178 sm_ui = MDI_UNIT(getminor(md_dev64_to_dev(tmpdev))); | |
1179 | |
1180 (void) md_unit_writerlock(sm_ui); | |
1181 if (submirror_unavailable(un, i, 0)) { | |
1182 sm_ui->ui_tstate |= MD_INACCESSIBLE; | |
1183 sm_unavail_cnt++; | |
1184 } else { | |
1185 sm_ui->ui_tstate &= ~MD_INACCESSIBLE; | |
1186 } | |
1187 md_unit_writerexit(sm_ui); | |
1188 } | |
1189 | |
1190 /* | |
1191 * If all of the submirrors are unavailable, the mirror is also | |
1192 * unavailable. | |
1193 */ | |
1194 if (sm_cnt == sm_unavail_cnt) { | |
1195 ui->ui_tstate |= MD_INACCESSIBLE; | |
1196 } else { | |
1197 ui->ui_tstate &= ~MD_INACCESSIBLE; | |
1198 } | |
1199 | |
1200 smi = 0; | |
1201 ci = 0; | |
1202 while (mirror_geterror(un, &smi, &ci, 1, 0) != 0) { | |
1203 if (mirror_other_sources(un, smi, ci, 1) == 1) { | |
1204 | |
1205 free_all_ecomps(ecomps); | |
1206 (void) mirror_close_all_devs(un, md_oflags); | |
1207 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_OPEN_FAIL, | |
1208 SVM_TAG_METADEVICE, setno, MD_SID(un)); | |
1209 mirror_openfail_console_info(un, smi, ci); | |
1210 if (lockp) { | |
1211 md_ioctl_writerexit(lockp); | |
1212 (void) md_ioctl_readerlock(lockp, ui); | |
1213 } else { | |
1214 md_unit_writerexit(ui); | |
1215 (void) md_unit_readerlock(ui); | |
1216 } | |
1217 return (ENXIO); | |
1218 } | |
1219 | |
1220 /* track all component states that need changing */ | |
1221 c = (err_comp_t *)kmem_alloc(sizeof (err_comp_t), KM_SLEEP); | |
1222 c->ec_next = ecomps; | |
1223 c->ec_smi = smi; | |
1224 c->ec_ci = ci; | |
1225 ecomps = c; | |
1226 ci++; | |
1227 } | |
1228 | |
1229 /* Make all state changes and commit them */ | |
1230 for (c = ecomps; c != NULL; c = c->ec_next) { | |
1231 /* | |
1232 * If lockp is set, then entering kernel through ioctl. | |
1233 * For a MN set, the only ioctl path is via a commd message | |
1234 * (ALLOCATE_HOTSPARE or *RESYNC* messages) that is already | |
1235 * being sent to each node. | |
1236 * In this case, set NO_XMIT so that set_sm_comp_state | |
1237 * won't attempt to send a message on a message. | |
1238 * | |
1239 * In !MN sets, the xmit flag is ignored, so it doesn't matter | |
1240 * which flag is passed. | |
1241 */ | |
1242 if (lockp) { | |
1243 set_sm_comp_state(un, c->ec_smi, c->ec_ci, CS_ERRED, 0, | |
1244 MD_STATE_NO_XMIT, lockp); | |
1245 } else { | |
1246 set_sm_comp_state(un, c->ec_smi, c->ec_ci, CS_ERRED, 0, | |
1247 (MD_STATE_XMIT | MD_STATE_OCHELD), lockp); | |
1248 } | |
1249 /* | |
1250 * For a MN set, the NOTIFY is done when the state change is | |
1251 * processed on each node | |
1252 */ | |
1253 if (!MD_MNSET_SETNO(setno)) { | |
1254 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_ERRED, | |
1255 SVM_TAG_METADEVICE, setno, MD_SID(un)); | |
1256 } | |
1257 } | |
1258 | |
1259 if (lockp) { | |
1260 md_ioctl_writerexit(lockp); | |
1261 (void) md_ioctl_readerlock(lockp, ui); | |
1262 } else { | |
1263 md_unit_writerexit(ui); | |
1264 (void) md_unit_readerlock(ui); | |
1265 } | |
1266 | |
1267 free_all_ecomps(ecomps); | |
1268 | |
1269 /* allocate hotspares for all errored components */ | |
1270 if (MD_MNSET_SETNO(setno)) { | |
1271 /* | |
1272 * If we're called from an ioctl (lockp set) then we cannot | |
1273 * directly call send_poke_hotspares as this will block until | |
1274 * the message gets despatched to all nodes. If the cluster is | |
1275 * going through a reconfig cycle then the message will block | |
1276 * until the cycle is complete, and as we originate from a | |
1277 * service call from commd we will livelock. | |
1278 */ | |
1279 if (lockp == NULL) { | |
1280 md_unit_readerexit(ui); | |
1281 send_poke_hotspares(setno); | |
1282 (void) md_unit_readerlock(ui); | |
1283 } | |
1284 } else { | |
1285 (void) poke_hotspares(); | |
1286 } | |
1287 return (0); | |
1288 } | |
1289 | |
1290 void | |
1291 mirror_overlap_chain_remove(md_mps_t *ps) | |
1292 { | |
1293 mm_unit_t *un; | |
1294 | |
1295 if (panicstr) | |
1296 return; | |
1297 | |
1298 ASSERT(ps->ps_flags & MD_MPS_ON_OVERLAP); | |
1299 | |
1300 un = ps->ps_un; | |
1301 | |
1302 mutex_enter(&un->un_ovrlap_chn_mx); | |
1303 if (ps->ps_ovrlap_prev != &un->un_ovrlap_chn) | |
1304 ps->ps_ovrlap_prev->ps_ovrlap_next = ps->ps_ovrlap_next; | |
1305 else | |
1306 un->un_ovrlap_chn.ps_ovrlap_next = ps->ps_ovrlap_next; | |
1307 if (ps->ps_ovrlap_next != &un->un_ovrlap_chn) | |
1308 ps->ps_ovrlap_next->ps_ovrlap_prev = ps->ps_ovrlap_prev; | |
1309 else | |
1310 un->un_ovrlap_chn.ps_ovrlap_prev = ps->ps_ovrlap_prev; | |
1311 /* Handle empty overlap chain */ | |
1312 if (un->un_ovrlap_chn.ps_ovrlap_prev == &un->un_ovrlap_chn) { | |
1313 un->un_ovrlap_chn.ps_ovrlap_prev = | |
1314 un->un_ovrlap_chn.ps_ovrlap_next = NULL; | |
1315 } | |
1316 if (un->un_ovrlap_chn_flg) { | |
1317 un->un_ovrlap_chn_flg = 0; | |
1318 cv_broadcast(&un->un_ovrlap_chn_cv); | |
1319 } | |
1320 ps->ps_flags &= ~MD_MPS_ON_OVERLAP; | |
1321 mutex_exit(&un->un_ovrlap_chn_mx); | |
1322 } | |
1323 | |
1324 | |
1325 /* | |
1326 * wait_for_overlaps: | |
1327 * ----------------- | |
1328 * Check that given i/o request does not cause an overlap with already pending | |
1329 * i/o. If it does, block until the overlapped i/o completes. | |
1330 * | |
1331 * Note: the overlap chain is held as a monotonically increasing doubly-linked | |
1332 * list with the sentinel contained in un->un_ovrlap_chn. We avoid a linear | |
1333 * search of the list by the following logic: | |
1334 * ps->ps_lastblk < un_ovrlap_chn.ps_ovrlap_next->ps_firstblk => No overlap | |
1335 * ps->ps_firstblk > un_ovrlap_chn.ps_ovrlap_prev->ps_lastblk => No overlap | |
1336 * otherwise | |
1337 * scan un_ovrlap_chn.ps_ovrlap_next for location where ps->ps_firstblk | |
1338 * > chain->ps_lastblk. This is the insertion point. As the list is | |
1339 * guaranteed to be ordered there is no need to continue scanning. | |
1340 * | |
1341 * The flag argument has MD_OVERLAP_ALLOW_REPEAT set if it is ok for the parent | |
1342 * structure to be already on the overlap chain and MD_OVERLAP_NO_REPEAT | |
1343 * if it must not already be on the chain | |
1344 */ | |
1345 static void | |
1346 wait_for_overlaps(md_mps_t *ps, int flags) | |
1347 { | |
1348 mm_unit_t *un; | |
1349 md_mps_t *ps1, **head, **tail; | |
1350 | |
1351 if (panicstr) | |
1352 return; | |
1353 | |
1354 | |
1355 un = ps->ps_un; | |
1356 | |
1357 mutex_enter(&un->un_ovrlap_chn_mx); | |
1358 if ((flags & MD_OVERLAP_ALLOW_REPEAT) && | |
1359 (ps->ps_flags & MD_MPS_ON_OVERLAP)) { | |
1360 mutex_exit(&un->un_ovrlap_chn_mx); | |
1361 return; | |
1362 } | |
1363 | |
1364 ASSERT(!(ps->ps_flags & MD_MPS_ON_OVERLAP)); | |
1365 head = &(un->un_ovrlap_chn.ps_ovrlap_next); | |
1366 tail = &(un->un_ovrlap_chn.ps_ovrlap_prev); | |
1367 ps1 = *head; | |
1368 /* | |
1369 * Check for simple limit cases: | |
1370 * *head == NULL | |
1371 * insert ps at head of list | |
1372 * lastblk < head->firstblk | |
1373 * insert at head of list | |
1374 * firstblk > tail->lastblk | |
1375 * insert at tail of list | |
1376 */ | |
1377 if (ps1 == NULL) { | |
1378 /* Insert at head */ | |
1379 ps->ps_ovrlap_next = &un->un_ovrlap_chn; | |
1380 ps->ps_ovrlap_prev = &un->un_ovrlap_chn; | |
1381 *head = ps; | |
1382 *tail = ps; | |
1383 ps->ps_flags |= MD_MPS_ON_OVERLAP; | |
1384 mutex_exit(&un->un_ovrlap_chn_mx); | |
1385 return; | |
1386 } else if (ps->ps_lastblk < (*head)->ps_firstblk) { | |
1387 /* Insert at head */ | |
1388 ps->ps_ovrlap_next = (*head); | |
1389 ps->ps_ovrlap_prev = &un->un_ovrlap_chn; | |
1390 (*head)->ps_ovrlap_prev = ps; | |
1391 *head = ps; | |
1392 ps->ps_flags |= MD_MPS_ON_OVERLAP; | |
1393 mutex_exit(&un->un_ovrlap_chn_mx); | |
1394 return; | |
1395 } else if (ps->ps_firstblk > (*tail)->ps_lastblk) { | |
1396 /* Insert at tail */ | |
1397 ps->ps_ovrlap_prev = (*tail); | |
1398 ps->ps_ovrlap_next = &un->un_ovrlap_chn; | |
1399 (*tail)->ps_ovrlap_next = ps; | |
1400 *tail = ps; | |
1401 ps->ps_flags |= MD_MPS_ON_OVERLAP; | |
1402 mutex_exit(&un->un_ovrlap_chn_mx); | |
1403 return; | |
1404 } | |
1405 /* Now we have to scan the list for possible overlaps */ | |
1406 while (ps1 != NULL) { | |
1407 /* | |
1408 * If this region has been put on the chain by another thread | |
1409 * just exit | |
1410 */ | |
1411 if ((flags & MD_OVERLAP_ALLOW_REPEAT) && | |
1412 (ps->ps_flags & MD_MPS_ON_OVERLAP)) { | |
1413 mutex_exit(&un->un_ovrlap_chn_mx); | |
1414 return; | |
1415 | |
1416 } | |
1417 for (ps1 = *head; ps1 && (ps1 != &un->un_ovrlap_chn); | |
1418 ps1 = ps1->ps_ovrlap_next) { | |
1419 if (ps->ps_firstblk > (*tail)->ps_lastblk) { | |
1420 /* Insert at tail */ | |
1421 ps->ps_ovrlap_prev = (*tail); | |
1422 ps->ps_ovrlap_next = &un->un_ovrlap_chn; | |
1423 (*tail)->ps_ovrlap_next = ps; | |
1424 *tail = ps; | |
1425 ps->ps_flags |= MD_MPS_ON_OVERLAP; | |
1426 mutex_exit(&un->un_ovrlap_chn_mx); | |
1427 return; | |
1428 } | |
1429 if (ps->ps_firstblk > ps1->ps_lastblk) | |
1430 continue; | |
1431 if (ps->ps_lastblk < ps1->ps_firstblk) { | |
1432 /* Insert into list at current 'ps1' position */ | |
1433 ps->ps_ovrlap_next = ps1; | |
1434 ps->ps_ovrlap_prev = ps1->ps_ovrlap_prev; | |
1435 ps1->ps_ovrlap_prev->ps_ovrlap_next = ps; | |
1436 ps1->ps_ovrlap_prev = ps; | |
1437 ps->ps_flags |= MD_MPS_ON_OVERLAP; | |
1438 mutex_exit(&un->un_ovrlap_chn_mx); | |
1439 return; | |
1440 } | |
1441 break; | |
1442 } | |
1443 if (ps1 != NULL) { | |
1444 un->un_ovrlap_chn_flg = 1; | |
1445 cv_wait(&un->un_ovrlap_chn_cv, &un->un_ovrlap_chn_mx); | |
1446 /* | |
1447 * Now ps1 refers to the old insertion point and we | |
1448 * have to check the whole chain to see if we're still | |
1449 * overlapping any other i/o. | |
1450 */ | |
1451 } | |
1452 } | |
1453 | |
1454 /* | |
1455 * Only get here if we had one overlapping i/o on the list and that | |
1456 * has now completed. In this case the list is empty so we insert <ps> | |
1457 * at the head of the chain. | |
1458 */ | |
1459 ASSERT(*head == NULL); | |
1460 *tail = *head = ps; | |
1461 ps->ps_ovrlap_next = ps->ps_ovrlap_prev = &un->un_ovrlap_chn; | |
1462 ps->ps_flags |= MD_MPS_ON_OVERLAP; | |
1463 mutex_exit(&un->un_ovrlap_chn_mx); | |
1464 } | |
1465 | |
1466 /* | |
1467 * This function is called from mirror_done to check whether any pages have | |
1468 * been modified while a mirrored write was in progress. Returns 0 if | |
1469 * all pages associated with bp are clean, 1 otherwise. | |
1470 */ | |
1471 static int | |
1472 any_pages_dirty(struct buf *bp) | |
1473 { | |
1474 int rval; | |
1475 | |
1476 rval = biomodified(bp); | |
1477 if (rval == -1) | |
1478 rval = 0; | |
1479 | |
1480 return (rval); | |
1481 } | |
1482 | |
1483 #define MAX_EXTRAS 10 | |
1484 | |
1485 void | |
1486 mirror_commit( | |
1487 mm_unit_t *un, | |
1488 int smmask, | |
1489 mddb_recid_t *extras | |
1490 ) | |
1491 { | |
1492 mm_submirror_t *sm; | |
1493 md_unit_t *su; | |
1494 int i; | |
1495 | |
1496 /* 2=mirror,null id */ | |
1497 mddb_recid_t recids[NMIRROR+2+MAX_EXTRAS]; | |
1498 | |
1499 int ri = 0; | |
1500 | |
1501 if (md_get_setstatus(MD_UN2SET(un)) & MD_SET_STALE) | |
1502 return; | |
1503 | |
1504 /* Add two, this includes the mirror unit and the null recid */ | |
1505 if (extras != NULL) { | |
1506 int nrecids = 0; | |
1507 while (extras[nrecids] != 0) { | |
1508 nrecids++; | |
1509 } | |
1510 ASSERT(nrecids <= MAX_EXTRAS); | |
1511 } | |
1512 | |
1513 if (un != NULL) | |
1514 recids[ri++] = un->c.un_record_id; | |
1515 for (i = 0; i < NMIRROR; i++) { | |
1516 if (!(smmask & SMI2BIT(i))) | |
1517 continue; | |
1518 sm = &un->un_sm[i]; | |
1519 if (!SMS_IS(sm, SMS_INUSE)) | |
1520 continue; | |
1521 if (md_getmajor(sm->sm_dev) != md_major) | |
1522 continue; | |
1523 su = MD_UNIT(md_getminor(sm->sm_dev)); | |
1524 recids[ri++] = su->c.un_record_id; | |
1525 } | |
1526 | |
1527 if (extras != NULL) | |
1528 while (*extras != 0) { | |
1529 recids[ri++] = *extras; | |
1530 extras++; | |
1531 } | |
1532 | |
1533 if (ri == 0) | |
1534 return; | |
1535 recids[ri] = 0; | |
1536 | |
1537 /* | |
1538 * Ok to hold ioctl lock across record commit to mddb as | |
1539 * long as the record(s) being committed aren't resync records. | |
1540 */ | |
1541 mddb_commitrecs_wrapper(recids); | |
1542 } | |
1543 | |
1544 | |
1545 /* | |
1546 * This routine is used to set a bit in the writable_bm bitmap | |
1547 * which represents each submirror in a metamirror which | |
1548 * is writable. The first writable submirror index is assigned | |
1549 * to the sm_index. The number of writable submirrors are returned in nunits. | |
1550 * | |
1551 * This routine returns the submirror's unit number. | |
1552 */ | |
1553 | |
1554 static void | |
1555 select_write_units(struct mm_unit *un, md_mps_t *ps) | |
1556 { | |
1557 | |
1558 int i; | |
1559 unsigned writable_bm = 0; | |
1560 unsigned nunits = 0; | |
1561 | |
1562 for (i = 0; i < NMIRROR; i++) { | |
1563 if (SUBMIRROR_IS_WRITEABLE(un, i)) { | |
1564 /* set bit of all writable units */ | |
1565 writable_bm |= SMI2BIT(i); | |
1566 nunits++; | |
1567 } | |
1568 } | |
1569 ps->ps_writable_sm = writable_bm; | |
1570 ps->ps_active_cnt = nunits; | |
1571 ps->ps_current_sm = 0; | |
1572 } | |
1573 | |
1574 static | |
1575 unsigned | |
1576 select_write_after_read_units(struct mm_unit *un, md_mps_t *ps) | |
1577 { | |
1578 | |
1579 int i; | |
1580 unsigned writable_bm = 0; | |
1581 unsigned nunits = 0; | |
1582 | |
1583 for (i = 0; i < NMIRROR; i++) { | |
1584 if (SUBMIRROR_IS_WRITEABLE(un, i) && | |
1585 un->un_sm[i].sm_flags & MD_SM_RESYNC_TARGET) { | |
1586 writable_bm |= SMI2BIT(i); | |
1587 nunits++; | |
1588 } | |
1589 } | |
1590 if ((writable_bm & ps->ps_allfrom_sm) != 0) { | |
1591 writable_bm &= ~ps->ps_allfrom_sm; | |
1592 nunits--; | |
1593 } | |
1594 ps->ps_writable_sm = writable_bm; | |
1595 ps->ps_active_cnt = nunits; | |
1596 ps->ps_current_sm = 0; | |
1597 return (nunits); | |
1598 } | |
1599 | |
1600 static md_dev64_t | |
1601 select_read_unit( | |
1602 mm_unit_t *un, | |
1603 diskaddr_t blkno, | |
1604 u_longlong_t reqcount, | |
1605 u_longlong_t *cando, | |
1606 int must_be_opened, | |
1607 md_m_shared_t **shared, | |
1608 md_mcs_t *cs) | |
1609 { | |
1610 int i; | |
1611 md_m_shared_t *s; | |
1612 uint_t lasterrcnt = 0; | |
1613 md_dev64_t dev = 0; | |
1614 u_longlong_t cnt; | |
1615 u_longlong_t mincnt; | |
1616 mm_submirror_t *sm; | |
1617 mm_submirror_ic_t *smic; | |
1618 mdi_unit_t *ui; | |
1619 | |
1620 mincnt = reqcount; | |
1621 for (i = 0; i < NMIRROR; i++) { | |
1622 if (!SUBMIRROR_IS_READABLE(un, i)) | |
1623 continue; | |
1624 sm = &un->un_sm[i]; | |
1625 smic = &un->un_smic[i]; | |
1626 cnt = reqcount; | |
1627 | |
1628 /* | |
1629 * If the current submirror is marked as inaccessible, do not | |
1630 * try to access it. | |
1631 */ | |
1632 ui = MDI_UNIT(getminor(expldev(sm->sm_dev))); | |
1633 (void) md_unit_readerlock(ui); | |
1634 if (ui->ui_tstate & MD_INACCESSIBLE) { | |
1635 md_unit_readerexit(ui); | |
1636 continue; | |
1637 } | |
1638 md_unit_readerexit(ui); | |
1639 | |
1640 s = (md_m_shared_t *)(*(smic->sm_shared_by_blk)) | |
1641 (sm->sm_dev, sm, blkno, &cnt); | |
1642 | |
1643 if (must_be_opened && !(s->ms_flags & MDM_S_ISOPEN)) | |
1644 continue; | |
1645 if (s->ms_state == CS_OKAY) { | |
1646 *cando = cnt; | |
1647 if (shared != NULL) | |
1648 *shared = s; | |
1649 | |
1650 if (un->un_sm[i].sm_flags & MD_SM_FAILFAST && | |
1651 cs != NULL) { | |
1652 cs->cs_buf.b_flags |= B_FAILFAST; | |
1653 } | |
1654 | |
1655 return (un->un_sm[i].sm_dev); | |
1656 } | |
1657 if (s->ms_state != CS_LAST_ERRED) | |
1658 continue; | |
1659 | |
1660 /* don't use B_FAILFAST since we're Last Erred */ | |
1661 | |
1662 if (mincnt > cnt) | |
1663 mincnt = cnt; | |
1664 if (s->ms_lasterrcnt > lasterrcnt) { | |
1665 lasterrcnt = s->ms_lasterrcnt; | |
1666 if (shared != NULL) | |
1667 *shared = s; | |
1668 dev = un->un_sm[i].sm_dev; | |
1669 } | |
1670 } | |
1671 *cando = mincnt; | |
1672 return (dev); | |
1673 } | |
1674 | |
1675 /* | |
1676 * Given a 32-bit bitmap, this routine will return the bit number | |
1677 * of the nth bit set. The nth bit set is passed via the index integer. | |
1678 * | |
1679 * This routine is used to run through the writable submirror bitmap | |
1680 * and starting all of the writes. See the value returned is the | |
1681 * index to appropriate submirror structure, in the md_sm | |
1682 * array for metamirrors. | |
1683 */ | |
1684 static int | |
1685 md_find_nth_unit(uint_t mask, int index) | |
1686 { | |
1687 int bit, nfound; | |
1688 | |
1689 for (bit = -1, nfound = -1; nfound != index; bit++) { | |
1690 ASSERT(mask != 0); | |
1691 nfound += (mask & 1); | |
1692 mask >>= 1; | |
1693 } | |
1694 return (bit); | |
1695 } | |
1696 | |
1697 static int | |
1698 fast_select_read_unit(md_mps_t *ps, md_mcs_t *cs) | |
1699 { | |
1700 mm_unit_t *un; | |
1701 buf_t *bp; | |
1702 int i; | |
1703 unsigned nunits = 0; | |
1704 int iunit; | |
1705 uint_t running_bm = 0; | |
1706 uint_t sm_index; | |
1707 | |
1708 bp = &cs->cs_buf; | |
1709 un = ps->ps_un; | |
1710 | |
1711 for (i = 0; i < NMIRROR; i++) { | |
1712 if (!SMS_BY_INDEX_IS(un, i, SMS_RUNNING)) | |
1713 continue; | |
1714 running_bm |= SMI2BIT(i); | |
1715 nunits++; | |
1716 } | |
1717 if (nunits == 0) | |
1718 return (1); | |
1719 | |
1720 /* | |
1721 * For directed mirror read (DMR) we only use the specified side and | |
1722 * do not compute the source of the read. | |
1723 */ | |
1724 if (ps->ps_flags & MD_MPS_DMR) { | |
1725 sm_index = un->un_dmr_last_read; | |
1726 } else { | |
1727 /* Normal (non-DMR) operation */ | |
1728 switch (un->un_read_option) { | |
1729 case RD_GEOMETRY: | |
1730 iunit = (int)(bp->b_lblkno / | |
1731 howmany(un->c.un_total_blocks, nunits)); | |
1732 sm_index = md_find_nth_unit(running_bm, iunit); | |
1733 break; | |
1734 case RD_FIRST: | |
1735 sm_index = md_find_nth_unit(running_bm, 0); | |
1736 break; | |
1737 case RD_LOAD_BAL: | |
1738 /* this is intentional to fall into the default */ | |
1739 default: | |
1740 un->un_last_read = (un->un_last_read + 1) % nunits; | |
1741 sm_index = md_find_nth_unit(running_bm, | |
1742 un->un_last_read); | |
1743 break; | |
1744 } | |
1745 } | |
1746 bp->b_edev = md_dev64_to_dev(un->un_sm[sm_index].sm_dev); | |
1747 ps->ps_allfrom_sm = SMI2BIT(sm_index); | |
1748 | |
1749 if (un->un_sm[sm_index].sm_flags & MD_SM_FAILFAST) { | |
1750 bp->b_flags |= B_FAILFAST; | |
1751 } | |
1752 | |
1753 return (0); | |
1754 } | |
1755 | |
1756 static | |
1757 int | |
1758 mirror_are_submirrors_available(mm_unit_t *un) | |
1759 { | |
1760 int i; | |
1761 for (i = 0; i < NMIRROR; i++) { | |
1762 md_dev64_t tmpdev = un->un_sm[i].sm_dev; | |
1763 | |
1764 if ((!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) || | |
1765 md_getmajor(tmpdev) != md_major) | |
1766 continue; | |
1767 | |
1768 if ((MD_MIN2SET(md_getminor(tmpdev)) >= md_nsets) || | |
1769 (MD_MIN2UNIT(md_getminor(tmpdev)) >= md_nunits)) | |
1770 return (0); | |
1771 | |
1772 if (MDI_UNIT(md_getminor(tmpdev)) == NULL) | |
1773 return (0); | |
1774 } | |
1775 return (1); | |
1776 } | |
1777 | |
1778 void | |
1779 build_submirror(mm_unit_t *un, int i, int snarfing) | |
1780 { | |
1781 struct mm_submirror *sm; | |
1782 struct mm_submirror_ic *smic; | |
1783 md_unit_t *su; | |
1784 set_t setno; | |
1785 | |
1786 sm = &un->un_sm[i]; | |
1787 smic = &un->un_smic[i]; | |
1788 | |
1789 sm->sm_flags = 0; /* sometime we may need to do more here */ | |
1790 | |
1791 setno = MD_UN2SET(un); | |
1792 | |
1793 if (!SMS_IS(sm, SMS_INUSE)) | |
1794 return; | |
1795 if (snarfing) { | |
1796 sm->sm_dev = md_getdevnum(setno, mddb_getsidenum(setno), | |
1797 sm->sm_key, MD_NOTRUST_DEVT); | |
1798 } else { | |
1799 if (md_getmajor(sm->sm_dev) == md_major) { | |
1800 su = MD_UNIT(md_getminor(sm->sm_dev)); | |
1801 un->c.un_flag |= (su->c.un_flag & MD_LABELED); | |
1802 /* submirror can no longer be soft partitioned */ | |
1803 MD_CAPAB(su) &= (~MD_CAN_SP); | |
1804 } | |
1805 } | |
1806 smic->sm_shared_by_blk = md_get_named_service(sm->sm_dev, | |
1807 0, "shared by blk", 0); | |
1808 smic->sm_shared_by_indx = md_get_named_service(sm->sm_dev, | |
1809 0, "shared by indx", 0); | |
1810 smic->sm_get_component_count = | |
1811 (int (*)())md_get_named_service(sm->sm_dev, 0, | |
1812 "get component count", 0); | |
1813 smic->sm_get_bcss = | |
1814 (int (*)())md_get_named_service(sm->sm_dev, 0, | |
1815 "get block count skip size", 0); | |
1816 sm->sm_state &= ~SMS_IGNORE; | |
1817 if (SMS_IS(sm, SMS_OFFLINE)) | |
1818 MD_STATUS(un) |= MD_UN_OFFLINE_SM; | |
1819 md_set_parent(sm->sm_dev, MD_SID(un)); | |
1820 } | |
1821 | |
1822 static void | |
1823 mirror_cleanup(mm_unit_t *un) | |
1824 { | |
1825 mddb_recid_t recid; | |
1826 int smi; | |
1827 sv_dev_t sv[NMIRROR]; | |
1828 int nsv = 0; | |
1829 | |
1830 /* | |
1831 * If a MN diskset and this node is not the master, do | |
1832 * not delete any records on snarf of the mirror records. | |
1833 */ | |
1834 if (MD_MNSET_SETNO(MD_UN2SET(un)) && | |
1835 md_set[MD_UN2SET(un)].s_am_i_master == 0) { | |
1836 return; | |
1837 } | |
1838 | |
1839 for (smi = 0; smi < NMIRROR; smi++) { | |
1840 if (!SMS_BY_INDEX_IS(un, smi, SMS_INUSE)) | |
1841 continue; | |
1842 sv[nsv].setno = MD_UN2SET(un); | |
1843 sv[nsv++].key = un->un_sm[smi].sm_key; | |
1844 } | |
1845 | |
1846 recid = un->un_rr_dirty_recid; | |
1847 mddb_deleterec_wrapper(un->c.un_record_id); | |
1848 if (recid > 0) | |
1849 mddb_deleterec_wrapper(recid); | |
1850 | |
1851 md_rem_names(sv, nsv); | |
1852 } | |
1853 | |
1854 /* Return a -1 if optimized record unavailable and set should be released */ | |
1855 int | |
1856 mirror_build_incore(mm_unit_t *un, int snarfing) | |
1857 { | |
1858 int i; | |
1859 | |
1860 if (MD_STATUS(un) & MD_UN_BEING_RESET) { | |
1861 mddb_setrecprivate(un->c.un_record_id, MD_PRV_PENDCLEAN); | |
1862 return (1); | |
1863 } | |
1864 | |
1865 if (mirror_are_submirrors_available(un) == 0) | |
1866 return (1); | |
1867 | |
1868 if (MD_UNIT(MD_SID(un)) != NULL) | |
1869 return (0); | |
1870 | |
1871 MD_STATUS(un) = 0; | |
1872 | |
1873 /* pre-4.1 didn't define CAN_META_CHILD capability */ | |
1874 MD_CAPAB(un) = MD_CAN_META_CHILD | MD_CAN_PARENT | MD_CAN_SP; | |
1875 | |
1876 un->un_ovrlap_chn_flg = 0; | |
1877 bzero(&un->un_ovrlap_chn, sizeof (un->un_ovrlap_chn)); | |
1878 | |
1879 for (i = 0; i < NMIRROR; i++) | |
1880 build_submirror(un, i, snarfing); | |
1881 | |
1882 if (unit_setup_resync(un, snarfing) != 0) { | |
1883 if (snarfing) { | |
1884 mddb_setrecprivate(un->c.un_record_id, MD_PRV_GOTIT); | |
1885 /* | |
1886 * If a MN set and set is not stale, then return -1 | |
1887 * which will force the caller to unload the set. | |
1888 * The MN diskset nodes will return failure if | |
1889 * unit_setup_resync fails so that nodes won't | |
1890 * get out of sync. | |
1891 * | |
1892 * If set is STALE, the master node can't allocate | |
1893 * a resync record (if needed), but node needs to | |
1894 * join the set so that user can delete broken mddbs. | |
1895 * So, if set is STALE, just continue on. | |
1896 */ | |
1897 if (MD_MNSET_SETNO(MD_UN2SET(un)) && | |
1898 !(md_get_setstatus(MD_UN2SET(un)) & MD_SET_STALE)) { | |
1899 return (-1); | |
1900 } | |
1901 } else | |
1902 return (1); | |
1903 } | |
1904 | |
1905 mutex_init(&un->un_ovrlap_chn_mx, NULL, MUTEX_DEFAULT, NULL); | |
1906 cv_init(&un->un_ovrlap_chn_cv, NULL, CV_DEFAULT, NULL); | |
1907 | |
1908 un->un_suspend_wr_flag = 0; | |
1909 mutex_init(&un->un_suspend_wr_mx, NULL, MUTEX_DEFAULT, NULL); | |
1910 cv_init(&un->un_suspend_wr_cv, NULL, CV_DEFAULT, NULL); | |
1911 | |
1912 /* | |
1913 * Allocate mutexes for mirror-owner and resync-owner changes. | |
1914 * All references to the owner message state field must be guarded | |
1915 * by this mutex. | |
1916 */ | |
1917 mutex_init(&un->un_owner_mx, NULL, MUTEX_DEFAULT, NULL); | |
1918 | |
1919 /* | |
1920 * Allocate mutex and condvar for resync thread manipulation. These | |
1921 * will be used by mirror_resync_unit/mirror_ioctl_resync | |
1922 */ | |
1923 mutex_init(&un->un_rs_thread_mx, NULL, MUTEX_DEFAULT, NULL); | |
1924 cv_init(&un->un_rs_thread_cv, NULL, CV_DEFAULT, NULL); | |
1925 | |
1926 /* | |
1927 * Allocate mutex and condvar for resync progress thread manipulation. | |
1928 * This allows resyncs to be continued across an intervening reboot. | |
1929 */ | |
1930 mutex_init(&un->un_rs_progress_mx, NULL, MUTEX_DEFAULT, NULL); | |
1931 cv_init(&un->un_rs_progress_cv, NULL, CV_DEFAULT, NULL); | |
1932 | |
1933 /* | |
1934 * Allocate mutex and condvar for Directed Mirror Reads (DMR). This | |
1935 * provides synchronization between a user-ioctl and the resulting | |
1936 * strategy() call that performs the read(). | |
1937 */ | |
1938 mutex_init(&un->un_dmr_mx, NULL, MUTEX_DEFAULT, NULL); | |
1939 cv_init(&un->un_dmr_cv, NULL, CV_DEFAULT, NULL); | |
1940 | |
1941 MD_UNIT(MD_SID(un)) = un; | |
1942 return (0); | |
1943 } | |
1944 | |
1945 | |
1946 void | |
1947 reset_mirror(struct mm_unit *un, minor_t mnum, int removing) | |
1948 { | |
1949 mddb_recid_t recid, vtoc_id; | |
1950 size_t bitcnt; | |
1951 size_t shortcnt; | |
1952 int smi; | |
1953 sv_dev_t sv[NMIRROR]; | |
1954 int nsv = 0; | |
1955 uint_t bits = 0; | |
1956 minor_t selfid; | |
1957 md_unit_t *su; | |
1958 | |
1959 md_destroy_unit_incore(mnum, &mirror_md_ops); | |
1960 | |
1961 shortcnt = un->un_rrd_num * sizeof (short); | |
1962 bitcnt = howmany(un->un_rrd_num, NBBY); | |
1963 | |
1964 if (un->un_outstanding_writes) | |
1965 kmem_free((caddr_t)un->un_outstanding_writes, shortcnt); | |
1966 if (un->un_goingclean_bm) | |
1967 kmem_free((caddr_t)un->un_goingclean_bm, bitcnt); | |
1968 if (un->un_goingdirty_bm) | |
1969 kmem_free((caddr_t)un->un_goingdirty_bm, bitcnt); | |
1970 if (un->un_resync_bm) | |
1971 kmem_free((caddr_t)un->un_resync_bm, bitcnt); | |
1972 | |
1973 MD_UNIT(mnum) = NULL; | |
1974 | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
1975 /* |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
1976 * Attempt release of its minor node |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
1977 */ |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
1978 (void) md_remove_minor_node(mnum); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
1979 |
0 | 1980 if (!removing) |
1981 return; | |
1982 | |
1983 for (smi = 0; smi < NMIRROR; smi++) { | |
1984 if (!SMS_BY_INDEX_IS(un, smi, SMS_INUSE)) | |
1985 continue; | |
1986 /* reallow soft partitioning of submirror and reset parent */ | |
1987 su = MD_UNIT(md_getminor(un->un_sm[smi].sm_dev)); | |
1988 MD_CAPAB(su) |= MD_CAN_SP; | |
1989 md_reset_parent(un->un_sm[smi].sm_dev); | |
1990 reset_comp_states(&un->un_sm[smi], &un->un_smic[smi]); | |
1991 | |
1992 sv[nsv].setno = MD_MIN2SET(mnum); | |
1993 sv[nsv++].key = un->un_sm[smi].sm_key; | |
1994 bits |= SMI2BIT(smi); | |
1995 } | |
1996 | |
1997 MD_STATUS(un) |= MD_UN_BEING_RESET; | |
1998 recid = un->un_rr_dirty_recid; | |
1999 vtoc_id = un->c.un_vtoc_id; | |
2000 selfid = MD_SID(un); | |
2001 | |
2002 mirror_commit(un, bits, 0); | |
2003 | |
2004 /* Destroy all mutexes and condvars before returning. */ | |
2005 mutex_destroy(&un->un_suspend_wr_mx); | |
2006 cv_destroy(&un->un_suspend_wr_cv); | |
2007 mutex_destroy(&un->un_ovrlap_chn_mx); | |
2008 cv_destroy(&un->un_ovrlap_chn_cv); | |
2009 mutex_destroy(&un->un_owner_mx); | |
2010 mutex_destroy(&un->un_rs_thread_mx); | |
2011 cv_destroy(&un->un_rs_thread_cv); | |
2012 mutex_destroy(&un->un_rs_progress_mx); | |
2013 cv_destroy(&un->un_rs_progress_cv); | |
2014 mutex_destroy(&un->un_dmr_mx); | |
2015 cv_destroy(&un->un_dmr_cv); | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
2016 |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
2017 /* |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
2018 * Remove self from the namespace |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
2019 */ |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
2020 if (un->c.un_revision & MD_FN_META_DEV) { |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
2021 (void) md_rem_selfname(un->c.un_self_id); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
2022 } |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
2023 |
0 | 2024 mddb_deleterec_wrapper(un->c.un_record_id); |
2025 if (recid != 0) | |
2026 mddb_deleterec_wrapper(recid); | |
2027 | |
2028 /* Remove the vtoc, if present */ | |
2029 if (vtoc_id) | |
2030 mddb_deleterec_wrapper(vtoc_id); | |
2031 | |
2032 md_rem_names(sv, nsv); | |
2033 | |
2034 SE_NOTIFY(EC_SVM_CONFIG, ESC_SVM_DELETE, SVM_TAG_METADEVICE, | |
2035 MD_MIN2SET(selfid), selfid); | |
2036 } | |
2037 | |
2038 int | |
2039 mirror_internal_open( | |
2040 minor_t mnum, | |
2041 int flag, | |
2042 int otyp, | |
2043 int md_oflags, | |
2044 IOLOCK *lockp /* can be NULL */ | |
2045 ) | |
2046 { | |
2047 mdi_unit_t *ui = MDI_UNIT(mnum); | |
2048 int err = 0; | |
2049 | |
2050 tryagain: | |
2051 /* single thread */ | |
2052 if (lockp) { | |
2053 /* | |
2054 * If ioctl lock is held, use openclose_enter | |
2055 * routine that will set the ioctl flag when | |
2056 * grabbing the readerlock. | |
2057 */ | |
2058 (void) md_ioctl_openclose_enter(lockp, ui); | |
2059 } else { | |
2060 (void) md_unit_openclose_enter(ui); | |
2061 } | |
2062 | |
2063 /* | |
2064 * The mirror_open_all_devs routine may end up sending a STATE_UPDATE | |
2065 * message in a MN diskset and this requires that the openclose | |
2066 * lock is dropped in order to send this message. So, another | |
2067 * flag (MD_UL_OPENINPROGRESS) is used to keep another thread from | |
2068 * attempting an open while this thread has an open in progress. | |
2069 * Call the *_lh version of the lock exit routines since the ui_mx | |
2070 * mutex must be held from checking for OPENINPROGRESS until | |
2071 * after the cv_wait call. | |
2072 */ | |
2073 mutex_enter(&ui->ui_mx); | |
2074 if (ui->ui_lock & MD_UL_OPENINPROGRESS) { | |
2075 if (lockp) { | |
2076 (void) md_ioctl_openclose_exit_lh(lockp); | |
2077 } else { | |
2078 md_unit_openclose_exit_lh(ui); | |
2079 } | |
2080 cv_wait(&ui->ui_cv, &ui->ui_mx); | |
2081 mutex_exit(&ui->ui_mx); | |
2082 goto tryagain; | |
2083 } | |
2084 | |
2085 ui->ui_lock |= MD_UL_OPENINPROGRESS; | |
2086 mutex_exit(&ui->ui_mx); | |
2087 | |
2088 /* open devices, if necessary */ | |
2089 if (! md_unit_isopen(ui) || (ui->ui_tstate & MD_INACCESSIBLE)) { | |
2090 if ((err = mirror_open_all_devs(mnum, md_oflags, lockp)) != 0) | |
2091 goto out; | |
2092 } | |
2093 | |
2094 /* count open */ | |
2095 if ((err = md_unit_incopen(mnum, flag, otyp)) != 0) | |
2096 goto out; | |
2097 | |
2098 /* unlock, return success */ | |
2099 out: | |
2100 mutex_enter(&ui->ui_mx); | |
2101 ui->ui_lock &= ~MD_UL_OPENINPROGRESS; | |
2102 mutex_exit(&ui->ui_mx); | |
2103 | |
2104 if (lockp) { | |
2105 /* | |
2106 * If ioctl lock is held, use openclose_exit | |
2107 * routine that will clear the lockp reader flag. | |
2108 */ | |
2109 (void) md_ioctl_openclose_exit(lockp); | |
2110 } else { | |
2111 md_unit_openclose_exit(ui); | |
2112 } | |
2113 return (err); | |
2114 } | |
2115 | |
2116 int | |
2117 mirror_internal_close( | |
2118 minor_t mnum, | |
2119 int otyp, | |
2120 int md_cflags, | |
2121 IOLOCK *lockp /* can be NULL */ | |
2122 ) | |
2123 { | |
2124 mdi_unit_t *ui = MDI_UNIT(mnum); | |
2125 mm_unit_t *un; | |
2126 int err = 0; | |
2127 | |
2128 /* single thread */ | |
2129 if (lockp) { | |
2130 /* | |
2131 * If ioctl lock is held, use openclose_enter | |
2132 * routine that will set the ioctl flag when | |
2133 * grabbing the readerlock. | |
2134 */ | |
2135 un = (mm_unit_t *)md_ioctl_openclose_enter(lockp, ui); | |
2136 } else { | |
2137 un = (mm_unit_t *)md_unit_openclose_enter(ui); | |
2138 } | |
2139 | |
2140 /* count closed */ | |
2141 if ((err = md_unit_decopen(mnum, otyp)) != 0) | |
2142 goto out; | |
2143 | |
2144 /* close devices, if necessary */ | |
2145 if (! md_unit_isopen(ui) || (md_cflags & MD_OFLG_PROBEDEV)) { | |
2146 /* | |
2147 * Clean up dirty bitmap for this unit. Do this | |
2148 * before closing the underlying devices to avoid | |
2149 * race conditions with reset_mirror() as a | |
2150 * result of a 'metaset -r' command running in | |
2151 * parallel. This might cause deallocation of | |
2152 * dirty region bitmaps; with underlying metadevices | |
2153 * in place this can't happen. | |
2154 * Don't do this if a MN set and ABR not set | |
2155 */ | |
2156 if (new_resync && !(MD_STATUS(un) & MD_UN_KEEP_DIRTY)) { | |
2157 if (!MD_MNSET_SETNO(MD_UN2SET(un)) || | |
2158 !(ui->ui_tstate & MD_ABR_CAP)) | |
2159 mirror_process_unit_resync(un); | |
2160 } | |
2161 (void) mirror_close_all_devs(un, md_cflags); | |
2162 | |
2163 /* | |
2164 * For a MN set with transient capabilities (eg ABR/DMR) set, | |
2165 * clear these capabilities on the last open in the cluster. | |
2166 * To do this we send a message to all nodes to see of the | |
2167 * device is open. | |
2168 */ | |
2169 if (MD_MNSET_SETNO(MD_UN2SET(un)) && | |
2170 (ui->ui_tstate & (MD_ABR_CAP|MD_DMR_CAP))) { | |
2171 if (lockp) { | |
2172 (void) md_ioctl_openclose_exit(lockp); | |
2173 } else { | |
2174 md_unit_openclose_exit(ui); | |
2175 } | |
2176 | |
2177 /* | |
2178 * if we are in the context of an ioctl, drop the | |
2179 * ioctl lock. | |
2180 * Otherwise, no other locks should be held. | |
2181 */ | |
2182 if (lockp) { | |
2183 IOLOCK_RETURN_RELEASE(0, lockp); | |
2184 } | |
2185 | |
2186 mdmn_clear_all_capabilities(mnum); | |
2187 | |
2188 /* if dropped the lock previously, regain it */ | |
2189 if (lockp) { | |
2190 IOLOCK_RETURN_REACQUIRE(lockp); | |
2191 } | |
2192 return (0); | |
2193 } | |
2194 /* unlock and return success */ | |
2195 } | |
2196 out: | |
2197 /* Call whether lockp is NULL or not. */ | |
2198 if (lockp) { | |
2199 md_ioctl_openclose_exit(lockp); | |
2200 } else { | |
2201 md_unit_openclose_exit(ui); | |
2202 } | |
2203 return (err); | |
2204 } | |
2205 | |
2206 /* | |
2207 * When a component has completed resyncing and is now ok, check if the | |
2208 * corresponding component in the other submirrors is in the Last Erred | |
2209 * state. If it is, we want to change that to the Erred state so we stop | |
2210 * using that component and start using this good component instead. | |
2211 * | |
2212 * This is called from set_sm_comp_state and recursively calls | |
2213 * set_sm_comp_state if it needs to change the Last Erred state. | |
2214 */ | |
2215 static void | |
2216 reset_lasterred(mm_unit_t *un, int smi, mddb_recid_t *extras, uint_t flags, | |
2217 IOLOCK *lockp) | |
2218 { | |
2219 mm_submirror_t *sm; | |
2220 mm_submirror_ic_t *smic; | |
2221 int ci; | |
2222 int i; | |
2223 int compcnt; | |
2224 int changed = 0; | |
2225 | |
2226 for (i = 0; i < NMIRROR; i++) { | |
2227 sm = &un->un_sm[i]; | |
2228 smic = &un->un_smic[i]; | |
2229 | |
2230 if (!SMS_IS(sm, SMS_INUSE)) | |
2231 continue; | |
2232 | |
2233 /* ignore the submirror that we just made ok */ | |
2234 if (i == smi) | |
2235 continue; | |
2236 | |
2237 compcnt = (*(smic->sm_get_component_count)) (sm->sm_dev, un); | |
2238 for (ci = 0; ci < compcnt; ci++) { | |
2239 md_m_shared_t *shared; | |
2240 | |
2241 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx)) | |
2242 (sm->sm_dev, sm, ci); | |
2243 | |
2244 if ((shared->ms_state & CS_LAST_ERRED) && | |
2245 !mirror_other_sources(un, i, ci, 1)) { | |
2246 | |
2247 set_sm_comp_state(un, i, ci, CS_ERRED, extras, | |
2248 flags, lockp); | |
2249 changed = 1; | |
2250 } | |
2251 } | |
2252 } | |
2253 | |
2254 /* maybe there is a hotspare for this newly erred component */ | |
2255 if (changed) { | |
2256 set_t setno; | |
2257 | |
2258 setno = MD_UN2SET(un); | |
2259 if (MD_MNSET_SETNO(setno)) { | |
2260 send_poke_hotspares(setno); | |
2261 } else { | |
2262 (void) poke_hotspares(); | |
2263 } | |
2264 } | |
2265 } | |
2266 | |
2267 /* | |
2268 * set_sm_comp_state | |
2269 * | |
2270 * Set the state of a submirror component to the specified new state. | |
2271 * If the mirror is in a multi-node set, send messages to all nodes to | |
2272 * block all writes to the mirror and then update the state and release the | |
2273 * writes. These messages are only sent if MD_STATE_XMIT is set in flags. | |
2274 * MD_STATE_XMIT will be unset in 2 cases: | |
2275 * 1. When the state is changed to CS_RESYNC as this state change | |
2276 * will already have been updated on each node by the processing of the | |
2277 * distributed metasync command, hence no need to xmit. | |
2278 * 2. When the state is change to CS_OKAY after a resync has completed. Again | |
2279 * the resync completion will already have been processed on each node by | |
2280 * the processing of the MD_MN_MSG_RESYNC_PHASE_DONE message for a component | |
2281 * resync, hence no need to xmit. | |
2282 * | |
2283 * In case we are called from the updates of a watermark, | |
2284 * (then MD_STATE_WMUPDATE will be set in the ps->flags) this is due to | |
2285 * a metainit or similar. In this case the message that we sent to propagate | |
2286 * the state change must not be a class1 message as that would deadlock with | |
2287 * the metainit command that is still being processed. | |
2288 * This we achieve by creating a class2 message MD_MN_MSG_STATE_UPDATE2 | |
2289 * instead. This also makes the submessage generator to create a class2 | |
2290 * submessage rather than a class1 (which would also block) | |
2291 * | |
2292 * On entry, unit_writerlock is held | |
2293 * If MD_STATE_OCHELD is set in flags, then unit_openclose lock is | |
2294 * also held. | |
2295 */ | |
2296 void | |
2297 set_sm_comp_state( | |
2298 mm_unit_t *un, | |
2299 int smi, | |
2300 int ci, | |
2301 int newstate, | |
2302 mddb_recid_t *extras, | |
2303 uint_t flags, | |
2304 IOLOCK *lockp | |
2305 ) | |
2306 { | |
2307 mm_submirror_t *sm; | |
2308 mm_submirror_ic_t *smic; | |
2309 md_m_shared_t *shared; | |
2310 int origstate; | |
2311 void (*get_dev)(); | |
2312 ms_cd_info_t cd; | |
2313 char devname[MD_MAX_CTDLEN]; | |
2314 int err; | |
2315 set_t setno = MD_UN2SET(un); | |
2316 md_mn_msg_stch_t stchmsg; | |
2317 mdi_unit_t *ui = MDI_UNIT(MD_SID(un)); | |
2318 md_mn_kresult_t *kresult; | |
2319 int rval; | |
2320 uint_t msgflags; | |
2321 md_mn_msgtype_t msgtype; | |
2322 int save_lock = 0; | |
2323 mdi_unit_t *ui_sm; | |
2324 | |
2325 sm = &un->un_sm[smi]; | |
2326 smic = &un->un_smic[smi]; | |
2327 | |
2328 /* If we have a real error status then turn off MD_INACCESSIBLE. */ | |
2329 ui_sm = MDI_UNIT(getminor(md_dev64_to_dev(sm->sm_dev))); | |
2330 if (newstate & (CS_ERRED | CS_RESYNC | CS_LAST_ERRED) && | |
2331 ui_sm->ui_tstate & MD_INACCESSIBLE) { | |
2332 ui_sm->ui_tstate &= ~MD_INACCESSIBLE; | |
2333 } | |
2334 | |
2335 shared = (md_m_shared_t *) | |
2336 (*(smic->sm_shared_by_indx))(sm->sm_dev, sm, ci); | |
2337 origstate = shared->ms_state; | |
2338 | |
2339 /* | |
2340 * If the new state is an error and the old one wasn't, generate | |
2341 * a console message. We do this before we send the state to other | |
2342 * nodes in a MN set because the state change may change the component | |
2343 * name if a hotspare is allocated. | |
2344 */ | |
2345 if ((! (origstate & (CS_ERRED|CS_LAST_ERRED))) && | |
2346 (newstate & (CS_ERRED|CS_LAST_ERRED))) { | |
2347 | |
2348 get_dev = | |
2349 (void (*)())md_get_named_service(sm->sm_dev, 0, | |
2350 "get device", 0); | |
2351 (void) (*get_dev)(sm->sm_dev, sm, ci, &cd); | |
2352 | |
2353 err = md_getdevname(setno, mddb_getsidenum(setno), 0, | |
2354 cd.cd_dev, devname, sizeof (devname)); | |
2355 | |
2356 if (err == ENOENT) { | |
2357 (void) md_devname(setno, cd.cd_dev, devname, | |
2358 sizeof (devname)); | |
2359 } | |
2360 | |
2361 cmn_err(CE_WARN, "md: %s: %s needs maintenance", | |
2362 md_shortname(md_getminor(sm->sm_dev)), devname); | |
2363 | |
2364 if (newstate & CS_LAST_ERRED) { | |
2365 cmn_err(CE_WARN, "md: %s: %s last erred", | |
2366 md_shortname(md_getminor(sm->sm_dev)), | |
2367 devname); | |
2368 | |
2369 } else if (shared->ms_flags & MDM_S_ISOPEN) { | |
2370 /* | |
2371 * Close the broken device and clear the open flag on | |
2372 * it. Closing the device means the RCM framework will | |
2373 * be able to unconfigure the device if required. | |
2374 * | |
2375 * We have to check that the device is open, otherwise | |
2376 * the first open on it has resulted in the error that | |
2377 * is being processed and the actual cd.cd_dev will be | |
2378 * NODEV64. | |
2379 * | |
2380 * If this is a multi-node mirror, then the multinode | |
2381 * state checks following this code will cause the | |
2382 * slave nodes to close the mirror in the function | |
2383 * mirror_set_state(). | |
2384 */ | |
2385 md_layered_close(cd.cd_dev, MD_OFLG_NULL); | |
2386 shared->ms_flags &= ~MDM_S_ISOPEN; | |
2387 } | |
2388 | |
2389 } else if ((origstate & CS_LAST_ERRED) && (newstate & CS_ERRED) && | |
2390 (shared->ms_flags & MDM_S_ISOPEN)) { | |
2391 /* | |
2392 * Similar to logic above except no log messages since we | |
2393 * are just transitioning from Last Erred to Erred. | |
2394 */ | |
2395 get_dev = (void (*)())md_get_named_service(sm->sm_dev, 0, | |
2396 "get device", 0); | |
2397 (void) (*get_dev)(sm->sm_dev, sm, ci, &cd); | |
2398 | |
2399 md_layered_close(cd.cd_dev, MD_OFLG_NULL); | |
2400 shared->ms_flags &= ~MDM_S_ISOPEN; | |
2401 } | |
2402 | |
2403 if ((MD_MNSET_SETNO(setno)) && (origstate != newstate) && | |
2404 (flags & MD_STATE_XMIT) && !(ui->ui_tstate & MD_ERR_PENDING)) { | |
2405 /* | |
2406 * For a multi-node mirror, send the state change to the | |
2407 * master, which broadcasts to all nodes, including this | |
2408 * one. Once the message is received, the state is set | |
2409 * in-core and the master commits the change to disk. | |
2410 * There is a case, comp_replace, where this function | |
2411 * can be called from within an ioctl and therefore in this | |
2412 * case, as the ioctl will already be called on each node, | |
2413 * there is no need to xmit the state change to the master for | |
2414 * distribution to the other nodes. MD_STATE_XMIT flag is used | |
2415 * to indicate whether a xmit is required. The mirror's | |
2416 * transient state is set to MD_ERR_PENDING to avoid sending | |
2417 * multiple messages. | |
2418 */ | |
2419 if (newstate & (CS_ERRED|CS_LAST_ERRED)) | |
2420 ui->ui_tstate |= MD_ERR_PENDING; | |
2421 | |
2422 /* | |
2423 * Send a state update message to all nodes. This message | |
2424 * will generate 2 submessages, the first one to suspend | |
2425 * all writes to the mirror and the second to update the | |
2426 * state and resume writes. | |
2427 */ | |
2428 stchmsg.msg_stch_mnum = un->c.un_self_id; | |
2429 stchmsg.msg_stch_sm = smi; | |
2430 stchmsg.msg_stch_comp = ci; | |
2431 stchmsg.msg_stch_new_state = newstate; | |
2432 stchmsg.msg_stch_hs_id = shared->ms_hs_id; | |
2433 #ifdef DEBUG | |
2434 if (mirror_debug_flag) | |
2435 printf("send set state, %x, %x, %x, %x, %x\n", | |
2436 stchmsg.msg_stch_mnum, stchmsg.msg_stch_sm, | |
2437 stchmsg.msg_stch_comp, stchmsg.msg_stch_new_state, | |
2438 stchmsg.msg_stch_hs_id); | |
2439 #endif | |
2440 if (flags & MD_STATE_WMUPDATE) { | |
2441 msgtype = MD_MN_MSG_STATE_UPDATE2; | |
2442 /* | |
2443 * When coming from an update of watermarks, there | |
2444 * must already be a message logged that triggered | |
2445 * this action. So, no need to log this message, too. | |
2446 */ | |
2447 msgflags = MD_MSGF_NO_LOG; | |
2448 } else { | |
2449 msgtype = MD_MN_MSG_STATE_UPDATE; | |
2450 msgflags = MD_MSGF_DEFAULT_FLAGS; | |
2451 } | |
2452 | |
2453 /* | |
2454 * If we are in the context of an ioctl, drop the ioctl lock. | |
2455 * lockp holds the list of locks held. | |
2456 * | |
2457 * Otherwise, increment the appropriate reacquire counters. | |
2458 * If openclose lock is *held, then must reacquire reader | |
2459 * lock before releasing the openclose lock. | |
2460 * Do not drop the ARRAY_WRITER lock as we may not be able | |
2461 * to reacquire it. | |
2462 */ | |
2463 if (lockp) { | |
2464 if (lockp->l_flags & MD_ARRAY_WRITER) { | |
2465 save_lock = MD_ARRAY_WRITER; | |
2466 lockp->l_flags &= ~MD_ARRAY_WRITER; | |
2467 } else if (lockp->l_flags & MD_ARRAY_READER) { | |
2468 save_lock = MD_ARRAY_READER; | |
2469 lockp->l_flags &= ~MD_ARRAY_READER; | |
2470 } | |
2471 IOLOCK_RETURN_RELEASE(0, lockp); | |
2472 } else { | |
2473 if (flags & MD_STATE_OCHELD) { | |
2474 md_unit_writerexit(ui); | |
2475 (void) md_unit_readerlock(ui); | |
2476 md_unit_openclose_exit(ui); | |
2477 } else { | |
2478 md_unit_writerexit(ui); | |
2479 } | |
2480 } | |
2481 | |
2482 kresult = kmem_alloc(sizeof (md_mn_kresult_t), KM_SLEEP); | |
2483 rval = mdmn_ksend_message(setno, | |
2484 msgtype, | |
2485 msgflags, | |
2486 (char *)&stchmsg, | |
2487 sizeof (stchmsg), | |
2488 kresult); | |
2489 | |
2490 if (!MDMN_KSEND_MSG_OK(rval, kresult)) { | |
2491 mdmn_ksend_show_error(rval, kresult, "STATE UPDATE"); | |
2492 cmn_err(CE_PANIC, | |
2493 "ksend_message failure: STATE_UPDATE"); | |
2494 } | |
2495 kmem_free(kresult, sizeof (md_mn_kresult_t)); | |
2496 | |
2497 /* if dropped the lock previously, regain it */ | |
2498 if (lockp) { | |
2499 IOLOCK_RETURN_REACQUIRE(lockp); | |
2500 lockp->l_flags |= save_lock; | |
2501 } else { | |
2502 /* | |
2503 * Reacquire dropped locks and update acquirecnts | |
2504 * appropriately. | |
2505 */ | |
2506 if (flags & MD_STATE_OCHELD) { | |
2507 /* | |
2508 * openclose also grabs readerlock. | |
2509 */ | |
2510 (void) md_unit_openclose_enter(ui); | |
2511 md_unit_readerexit(ui); | |
2512 (void) md_unit_writerlock(ui); | |
2513 } else { | |
2514 (void) md_unit_writerlock(ui); | |
2515 } | |
2516 } | |
2517 | |
2518 ui->ui_tstate &= ~MD_ERR_PENDING; | |
2519 } else { | |
2520 shared->ms_state = newstate; | |
2521 uniqtime32(&shared->ms_timestamp); | |
2522 | |
2523 if (newstate == CS_ERRED) | |
2524 shared->ms_flags |= MDM_S_NOWRITE; | |
2525 else | |
2526 shared->ms_flags &= ~MDM_S_NOWRITE; | |
2527 | |
2528 shared->ms_flags &= ~MDM_S_IOERR; | |
2529 un->un_changecnt++; | |
2530 shared->ms_lasterrcnt = un->un_changecnt; | |
2531 | |
2532 mirror_set_sm_state(sm, smic, SMS_RUNNING, 0); | |
2533 mirror_commit(un, SMI2BIT(smi), extras); | |
2534 } | |
2535 | |
2536 if ((origstate & CS_RESYNC) && (newstate & CS_OKAY)) { | |
2537 /* | |
2538 * Resetting the Last Erred state will recursively call back | |
2539 * into this function (set_sm_comp_state) to update the state. | |
2540 */ | |
2541 reset_lasterred(un, smi, extras, flags, lockp); | |
2542 } | |
2543 } | |
2544 | |
2545 static int | |
2546 find_another_logical( | |
2547 mm_unit_t *un, | |
2548 mm_submirror_t *esm, | |
2549 diskaddr_t blk, | |
2550 u_longlong_t cnt, | |
2551 int must_be_open, | |
2552 int state, | |
2553 int err_cnt) | |
2554 { | |
2555 u_longlong_t cando; | |
2556 md_dev64_t dev; | |
2557 md_m_shared_t *s; | |
2558 | |
2559 esm->sm_state |= SMS_IGNORE; | |
2560 while (cnt != 0) { | |
2561 u_longlong_t mcnt; | |
2562 | |
2563 mcnt = MIN(cnt, lbtodb(1024 * 1024 * 1024)); /* 1 Gig Blks */ | |
2564 | |
2565 dev = select_read_unit(un, blk, mcnt, &cando, must_be_open, &s, | |
2566 NULL); | |
2567 if (dev == (md_dev64_t)0) | |
2568 break; | |
2569 | |
2570 if ((state == CS_LAST_ERRED) && | |
2571 (s->ms_state == CS_LAST_ERRED) && | |
2572 (err_cnt > s->ms_lasterrcnt)) | |
2573 break; | |
2574 | |
2575 cnt -= cando; | |
2576 blk += cando; | |
2577 } | |
2578 esm->sm_state &= ~SMS_IGNORE; | |
2579 return (cnt != 0); | |
2580 } | |
2581 | |
2582 int | |
2583 mirror_other_sources(mm_unit_t *un, int smi, int ci, int must_be_open) | |
2584 { | |
2585 mm_submirror_t *sm; | |
2586 mm_submirror_ic_t *smic; | |
2587 size_t count; | |
2588 diskaddr_t block; | |
2589 u_longlong_t skip; | |
2590 u_longlong_t size; | |
2591 md_dev64_t dev; | |
2592 int cnt; | |
2593 md_m_shared_t *s; | |
2594 int not_found; | |
2595 | |
2596 sm = &un->un_sm[smi]; | |
2597 smic = &un->un_smic[smi]; | |
2598 dev = sm->sm_dev; | |
2599 | |
2600 /* | |
2601 * Make sure every component of the submirror | |
2602 * has other sources. | |
2603 */ | |
2604 if (ci < 0) { | |
2605 /* Find the highest lasterrcnt */ | |
2606 cnt = (*(smic->sm_get_component_count))(dev, sm); | |
2607 for (ci = 0; ci < cnt; ci++) { | |
2608 not_found = mirror_other_sources(un, smi, ci, | |
2609 must_be_open); | |
2610 if (not_found) | |
2611 return (1); | |
2612 } | |
2613 return (0); | |
2614 } | |
2615 | |
2616 /* | |
2617 * Make sure this component has other sources | |
2618 */ | |
2619 (void) (*(smic->sm_get_bcss)) | |
2620 (dev, sm, ci, &block, &count, &skip, &size); | |
2621 | |
2622 if (count == 0) | |
2623 return (1); | |
2624 | |
2625 s = (md_m_shared_t *)(*(smic->sm_shared_by_indx))(dev, sm, ci); | |
2626 | |
2627 while (count--) { | |
2628 if (block >= un->c.un_total_blocks) | |
2629 return (0); | |
2630 | |
2631 if ((block + size) > un->c.un_total_blocks) | |
2632 size = un->c.un_total_blocks - block; | |
2633 | |
2634 not_found = find_another_logical(un, sm, block, size, | |
2635 must_be_open, s->ms_state, s->ms_lasterrcnt); | |
2636 if (not_found) | |
2637 return (1); | |
2638 | |
2639 block += size + skip; | |
2640 } | |
2641 return (0); | |
2642 } | |
2643 | |
2644 static void | |
2645 finish_error(md_mps_t *ps) | |
2646 { | |
2647 struct buf *pb; | |
2648 mm_unit_t *un; | |
2649 mdi_unit_t *ui; | |
2650 uint_t new_str_flags; | |
2651 | |
2652 pb = ps->ps_bp; | |
2653 un = ps->ps_un; | |
2654 ui = ps->ps_ui; | |
2655 | |
2656 /* | |
2657 * Must flag any error to the resync originator if we're performing | |
2658 * a Write-after-Read. This corresponds to an i/o error on a resync | |
2659 * target device and in this case we ought to abort the resync as there | |
2660 * is nothing that can be done to recover from this without operator | |
2661 * intervention. If we don't set the B_ERROR flag we will continue | |
2662 * reading from the mirror but won't write to the target (as it will | |
2663 * have been placed into an errored state). | |
2664 * To handle the case of multiple components within a submirror we only | |
2665 * set the B_ERROR bit if explicitly requested to via MD_MPS_FLAG_ERROR. | |
2666 * The originator of the resync read will cause this bit to be set if | |
2667 * the underlying component count is one for a submirror resync. All | |
2668 * other resync types will have the flag set as there is no underlying | |
2669 * resync which can be performed on a contained metadevice for these | |
2670 * resync types (optimized or component). | |
2671 */ | |
2672 | |
2673 if (ps->ps_flags & MD_MPS_WRITE_AFTER_READ) { | |
2674 if (ps->ps_flags & MD_MPS_FLAG_ERROR) | |
2675 pb->b_flags |= B_ERROR; | |
2676 md_kstat_done(ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ)); | |
2677 MPS_FREE(mirror_parent_cache, ps); | |
2678 md_unit_readerexit(ui); | |
2679 md_biodone(pb); | |
2680 return; | |
2681 } | |
2682 /* | |
2683 * Set the MD_IO_COUNTED flag as we are retrying the same I/O | |
2684 * operation therefore this I/O request has already been counted, | |
2685 * the I/O count variable will be decremented by mirror_done()'s | |
2686 * call to md_biodone(). | |
2687 */ | |
2688 if (ps->ps_changecnt != un->un_changecnt) { | |
2689 new_str_flags = MD_STR_NOTTOP | MD_IO_COUNTED; | |
2690 if (ps->ps_flags & MD_MPS_WOW) | |
2691 new_str_flags |= MD_STR_WOW; | |
2692 if (ps->ps_flags & MD_MPS_MAPPED) | |
2693 new_str_flags |= MD_STR_MAPPED; | |
2694 /* | |
2695 * If this I/O request was a read that was part of a resync, | |
2696 * set MD_STR_WAR for the retried read to ensure that the | |
2697 * resync write (i.e. write-after-read) will be performed | |
2698 */ | |
2699 if (ps->ps_flags & MD_MPS_RESYNC_READ) | |
2700 new_str_flags |= MD_STR_WAR; | |
2701 md_kstat_done(ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ)); | |
2702 MPS_FREE(mirror_parent_cache, ps); | |
2703 md_unit_readerexit(ui); | |
2704 (void) md_mirror_strategy(pb, new_str_flags, NULL); | |
2705 return; | |
2706 } | |
2707 | |
2708 pb->b_flags |= B_ERROR; | |
2709 md_kstat_done(ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ)); | |
2710 MPS_FREE(mirror_parent_cache, ps); | |
2711 md_unit_readerexit(ui); | |
2712 md_biodone(pb); | |
2713 } | |
2714 | |
2715 static void | |
2716 error_update_unit(md_mps_t *ps) | |
2717 { | |
2718 mm_unit_t *un; | |
2719 mdi_unit_t *ui; | |
2720 int smi; /* sub mirror index */ | |
2721 int ci; /* errored component */ | |
2722 set_t setno; | |
2723 uint_t flags; /* for set_sm_comp_state() */ | |
2724 uint_t hspflags; /* for check_comp_4_hotspares() */ | |
2725 | |
2726 ui = ps->ps_ui; | |
2727 un = (mm_unit_t *)md_unit_writerlock(ui); | |
2728 setno = MD_UN2SET(un); | |
2729 | |
2730 /* All of these updates have to propagated in case of MN set */ | |
2731 flags = MD_STATE_XMIT; | |
2732 hspflags = MD_HOTSPARE_XMIT; | |
2733 | |
2734 /* special treatment if we are called during updating watermarks */ | |
2735 if (ps->ps_flags & MD_MPS_WMUPDATE) { | |
2736 flags |= MD_STATE_WMUPDATE; | |
2737 hspflags |= MD_HOTSPARE_WMUPDATE; | |
2738 } | |
2739 smi = 0; | |
2740 ci = 0; | |
2741 while (mirror_geterror(un, &smi, &ci, 1, 0) != 0) { | |
2742 if (mirror_other_sources(un, smi, ci, 0) == 1) { | |
2743 | |
2744 /* Never called from ioctl context, so (IOLOCK *)NULL */ | |
2745 set_sm_comp_state(un, smi, ci, CS_LAST_ERRED, 0, flags, | |
2746 (IOLOCK *)NULL); | |
2747 /* | |
2748 * For a MN set, the NOTIFY is done when the state | |
2749 * change is processed on each node | |
2750 */ | |
2751 if (!MD_MNSET_SETNO(MD_UN2SET(un))) { | |
2752 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_LASTERRED, | |
2753 SVM_TAG_METADEVICE, setno, MD_SID(un)); | |
2754 } | |
2755 continue; | |
2756 } | |
2757 /* Never called from ioctl context, so (IOLOCK *)NULL */ | |
2758 set_sm_comp_state(un, smi, ci, CS_ERRED, 0, flags, | |
2759 (IOLOCK *)NULL); | |
2760 /* | |
2761 * For a MN set, the NOTIFY is done when the state | |
2762 * change is processed on each node | |
2763 */ | |
2764 if (!MD_MNSET_SETNO(MD_UN2SET(un))) { | |
2765 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_ERRED, | |
2766 SVM_TAG_METADEVICE, setno, MD_SID(un)); | |
2767 } | |
2768 smi = 0; | |
2769 ci = 0; | |
2770 } | |
2771 | |
2772 md_unit_writerexit(ui); | |
2773 if (MD_MNSET_SETNO(setno)) { | |
2774 send_poke_hotspares(setno); | |
2775 } else { | |
2776 (void) poke_hotspares(); | |
2777 } | |
2778 (void) md_unit_readerlock(ui); | |
2779 | |
2780 finish_error(ps); | |
2781 } | |
2782 | |
2783 /* | |
2784 * When we have a B_FAILFAST IO error on a Last Erred component we need to | |
2785 * retry the IO without B_FAILFAST set so that we try to ensure that the | |
2786 * component "sees" each IO. | |
2787 */ | |
2788 static void | |
2789 last_err_retry(md_mcs_t *cs) | |
2790 { | |
2791 struct buf *cb; | |
2792 md_mps_t *ps; | |
2793 uint_t flags; | |
2794 | |
2795 cb = &cs->cs_buf; | |
2796 cb->b_flags &= ~B_FAILFAST; | |
2797 | |
2798 /* if we're panicing just let this I/O error out */ | |
2799 if (panicstr) { | |
2800 (void) mirror_done(cb); | |
2801 return; | |
2802 } | |
2803 | |
2804 /* reissue the I/O */ | |
2805 | |
2806 ps = cs->cs_ps; | |
2807 | |
2808 bioerror(cb, 0); | |
2809 | |
2810 mutex_enter(&ps->ps_mx); | |
2811 | |
2812 flags = MD_STR_NOTTOP; | |
2813 if (ps->ps_flags & MD_MPS_MAPPED) | |
2814 flags |= MD_STR_MAPPED; | |
2815 if (ps->ps_flags & MD_MPS_NOBLOCK) | |
2816 flags |= MD_NOBLOCK; | |
2817 | |
2818 mutex_exit(&ps->ps_mx); | |
2819 | |
2820 clear_retry_error(cb); | |
2821 | |
2822 cmn_err(CE_NOTE, "!md: %s: Last Erred, retry I/O without B_FAILFAST", | |
2823 md_shortname(getminor(cb->b_edev))); | |
2824 | |
2825 md_call_strategy(cb, flags, NULL); | |
2826 } | |
2827 | |
2828 static void | |
2829 mirror_error(md_mps_t *ps) | |
2830 { | |
2831 int smi; /* sub mirror index */ | |
2832 int ci; /* errored component */ | |
2833 | |
2834 if (panicstr) { | |
2835 finish_error(ps); | |
2836 return; | |
2837 } | |
2838 | |
2839 if (ps->ps_flags & MD_MPS_ON_OVERLAP) | |
2840 mirror_overlap_chain_remove(ps); | |
2841 | |
2842 smi = 0; | |
2843 ci = 0; | |
2844 if (mirror_geterror(ps->ps_un, &smi, &ci, 0, 0) != 0) { | |
2845 md_unit_readerexit(ps->ps_ui); | |
2846 daemon_request(&md_mstr_daemon, error_update_unit, | |
2847 (daemon_queue_t *)ps, REQ_OLD); | |
2848 return; | |
2849 } | |
2850 | |
2851 finish_error(ps); | |
2852 } | |
2853 | |
2854 static int | |
2855 copy_write_done(struct buf *cb) | |
2856 { | |
2857 md_mps_t *ps; | |
2858 buf_t *pb; | |
2859 char *wowbuf; | |
2860 wowhdr_t *wowhdr; | |
2861 ssize_t wow_resid; | |
2862 | |
2863 /* get wowbuf ans save structure */ | |
2864 wowbuf = cb->b_un.b_addr; | |
2865 wowhdr = WOWBUF_HDR(wowbuf); | |
2866 ps = wowhdr->wow_ps; | |
2867 pb = ps->ps_bp; | |
2868 | |
2869 /* Save error information, then free cb */ | |
2870 if (cb->b_flags & B_ERROR) | |
2871 pb->b_flags |= B_ERROR; | |
2872 | |
2873 if (cb->b_flags & B_REMAPPED) | |
2874 bp_mapout(cb); | |
2875 | |
2876 freerbuf(cb); | |
2877 | |
2878 /* update residual and continue if needed */ | |
2879 if ((pb->b_flags & B_ERROR) == 0) { | |
2880 wow_resid = pb->b_bcount - wowhdr->wow_offset; | |
2881 pb->b_resid = wow_resid; | |
2882 if (wow_resid > 0) { | |
2883 daemon_request(&md_mstr_daemon, copy_write_cont, | |
2884 (daemon_queue_t *)wowhdr, REQ_OLD); | |
2885 return (1); | |
2886 } | |
2887 } | |
2888 | |
2889 /* Write is complete, release resources. */ | |
2890 kmem_cache_free(mirror_wowblk_cache, wowhdr); | |
2891 ASSERT(!(ps->ps_flags & MD_MPS_ON_OVERLAP)); | |
2892 md_kstat_done(ps->ps_ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ)); | |
2893 MPS_FREE(mirror_parent_cache, ps); | |
2894 md_biodone(pb); | |
2895 return (0); | |
2896 } | |
2897 | |
2898 static void | |
2899 copy_write_cont(wowhdr_t *wowhdr) | |
2900 { | |
2901 buf_t *pb; | |
2902 buf_t *cb; | |
2903 char *wowbuf; | |
2904 int wow_offset; | |
2905 size_t wow_resid; | |
2906 diskaddr_t wow_blkno; | |
2907 | |
2908 wowbuf = WOWHDR_BUF(wowhdr); | |
2909 pb = wowhdr->wow_ps->ps_bp; | |
2910 | |
2911 /* get data on current location */ | |
2912 wow_offset = wowhdr->wow_offset; | |
2913 wow_resid = pb->b_bcount - wow_offset; | |
2914 wow_blkno = pb->b_lblkno + lbtodb(wow_offset); | |
2915 | |
2916 /* setup child buffer */ | |
2917 cb = getrbuf(KM_SLEEP); | |
2918 cb->b_flags = B_WRITE; | |
2919 cb->b_edev = pb->b_edev; | |
2920 cb->b_un.b_addr = wowbuf; /* change to point at WOWBUF */ | |
2921 cb->b_bufsize = md_wowbuf_size; /* change to wowbuf_size */ | |
2922 cb->b_iodone = copy_write_done; | |
2923 cb->b_bcount = MIN(md_wowbuf_size, wow_resid); | |
2924 cb->b_lblkno = wow_blkno; | |
2925 | |
2926 /* move offset to next section */ | |
2927 wowhdr->wow_offset += cb->b_bcount; | |
2928 | |
2929 /* copy and setup write for current section */ | |
2930 bcopy(&pb->b_un.b_addr[wow_offset], wowbuf, cb->b_bcount); | |
2931 | |
2932 /* do it */ | |
2933 /* | |
2934 * Do not set the MD_IO_COUNTED flag as this is a new I/O request | |
2935 * that handles the WOW condition. The resultant increment on the | |
2936 * I/O count variable is cleared by copy_write_done()'s call to | |
2937 * md_biodone(). | |
2938 */ | |
2939 (void) md_mirror_strategy(cb, MD_STR_NOTTOP | MD_STR_WOW | |
2940 | MD_STR_MAPPED, NULL); | |
2941 } | |
2942 | |
2943 static void | |
2944 md_mirror_copy_write(md_mps_t *ps) | |
2945 { | |
2946 wowhdr_t *wowhdr; | |
2947 | |
2948 wowhdr = kmem_cache_alloc(mirror_wowblk_cache, MD_ALLOCFLAGS); | |
2949 mirror_wowblk_init(wowhdr); | |
2950 wowhdr->wow_ps = ps; | |
2951 wowhdr->wow_offset = 0; | |
2952 copy_write_cont(wowhdr); | |
2953 } | |
2954 | |
2955 static void | |
2956 handle_wow(md_mps_t *ps) | |
2957 { | |
2958 buf_t *pb; | |
2959 | |
2960 pb = ps->ps_bp; | |
2961 | |
2962 bp_mapin(pb); | |
2963 | |
2964 md_mirror_wow_cnt++; | |
2965 if (!(pb->b_flags & B_PHYS) && (md_mirror_wow_flg & WOW_LOGIT)) { | |
2966 cmn_err(CE_NOTE, | |
2967 "md: %s, blk %lld, cnt %ld: Write on write %d occurred", | |
2968 md_shortname(getminor(pb->b_edev)), | |
2969 (longlong_t)pb->b_lblkno, pb->b_bcount, md_mirror_wow_cnt); | |
2970 } | |
2971 | |
2972 /* | |
2973 * Set the MD_IO_COUNTED flag as we are retrying the same I/O | |
2974 * operation therefore this I/O request has already been counted, | |
2975 * the I/O count variable will be decremented by mirror_done()'s | |
2976 * call to md_biodone(). | |
2977 */ | |
2978 if (md_mirror_wow_flg & WOW_NOCOPY) | |
2979 (void) md_mirror_strategy(pb, MD_STR_NOTTOP | MD_STR_WOW | | |
2980 MD_STR_MAPPED | MD_IO_COUNTED, ps); | |
2981 else | |
2982 md_mirror_copy_write(ps); | |
2983 } | |
2984 | |
2985 /* | |
2986 * Return true if the specified submirror is either in the Last Erred | |
2987 * state or is transitioning into the Last Erred state. | |
2988 */ | |
2989 static bool_t | |
2990 submirror_is_lasterred(mm_unit_t *un, int smi) | |
2991 { | |
2992 mm_submirror_t *sm; | |
2993 mm_submirror_ic_t *smic; | |
2994 md_m_shared_t *shared; | |
2995 int ci; | |
2996 int compcnt; | |
2997 | |
2998 sm = &un->un_sm[smi]; | |
2999 smic = &un->un_smic[smi]; | |
3000 | |
3001 compcnt = (*(smic->sm_get_component_count)) (sm->sm_dev, un); | |
3002 for (ci = 0; ci < compcnt; ci++) { | |
3003 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx)) | |
3004 (sm->sm_dev, sm, ci); | |
3005 | |
3006 if (shared->ms_state == CS_LAST_ERRED) | |
3007 return (B_TRUE); | |
3008 | |
3009 /* | |
3010 * It is not currently Last Erred, check if entering Last Erred. | |
3011 */ | |
3012 if ((shared->ms_flags & MDM_S_IOERR) && | |
3013 ((shared->ms_state == CS_OKAY) || | |
3014 (shared->ms_state == CS_RESYNC))) { | |
3015 if (mirror_other_sources(un, smi, ci, 0) == 1) | |
3016 return (B_TRUE); | |
3017 } | |
3018 } | |
3019 | |
3020 return (B_FALSE); | |
3021 } | |
3022 | |
3023 | |
3024 static int | |
3025 mirror_done(struct buf *cb) | |
3026 { | |
3027 md_mps_t *ps; | |
3028 md_mcs_t *cs; | |
3029 | |
3030 /*LINTED*/ | |
3031 cs = (md_mcs_t *)((caddr_t)cb - md_mirror_mcs_buf_off); | |
3032 ps = cs->cs_ps; | |
3033 | |
3034 mutex_enter(&ps->ps_mx); | |
3035 | |
3036 /* check if we need to retry an errored failfast I/O */ | |
3037 if (cb->b_flags & B_ERROR) { | |
3038 struct buf *pb = ps->ps_bp; | |
3039 | |
3040 if (cb->b_flags & B_FAILFAST) { | |
3041 int i; | |
3042 mm_unit_t *un = ps->ps_un; | |
3043 | |
3044 for (i = 0; i < NMIRROR; i++) { | |
3045 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) | |
3046 continue; | |
3047 | |
3048 if (cb->b_edev == | |
3049 md_dev64_to_dev(un->un_sm[i].sm_dev)) { | |
3050 | |
3051 /* | |
3052 * This is the submirror that had the | |
3053 * error. Check if it is Last Erred. | |
3054 */ | |
3055 if (submirror_is_lasterred(un, i)) { | |
3056 daemon_queue_t *dqp; | |
3057 | |
3058 mutex_exit(&ps->ps_mx); | |
3059 dqp = (daemon_queue_t *)cs; | |
3060 dqp->dq_prev = NULL; | |
3061 dqp->dq_next = NULL; | |
3062 daemon_request(&md_done_daemon, | |
3063 last_err_retry, dqp, | |
3064 REQ_OLD); | |
3065 return (1); | |
3066 } | |
3067 break; | |
3068 } | |
3069 } | |
3070 } | |
3071 | |
3072 /* continue to process the buf without doing a retry */ | |
3073 ps->ps_flags |= MD_MPS_ERROR; | |
3074 pb->b_error = cb->b_error; | |
3075 } | |
3076 | |
3077 return (mirror_done_common(cb)); | |
3078 } | |
3079 | |
3080 /* | |
3081 * Split from the original mirror_done function so we can handle bufs after a | |
3082 * retry. | |
3083 * ps->ps_mx is already held in the caller of this function and the cb error | |
3084 * has already been checked and handled in the caller. | |
3085 */ | |
3086 static int | |
3087 mirror_done_common(struct buf *cb) | |
3088 { | |
3089 struct buf *pb; | |
3090 mm_unit_t *un; | |
3091 mdi_unit_t *ui; | |
3092 md_mps_t *ps; | |
3093 md_mcs_t *cs; | |
3094 size_t end_rr, start_rr, current_rr; | |
3095 | |
3096 /*LINTED*/ | |
3097 cs = (md_mcs_t *)((caddr_t)cb - md_mirror_mcs_buf_off); | |
3098 ps = cs->cs_ps; | |
3099 pb = ps->ps_bp; | |
3100 | |
3101 if (cb->b_flags & B_REMAPPED) | |
3102 bp_mapout(cb); | |
3103 | |
3104 ps->ps_frags--; | |
3105 if (ps->ps_frags != 0) { | |
3106 mutex_exit(&ps->ps_mx); | |
3107 kmem_cache_free(mirror_child_cache, cs); | |
3108 return (1); | |
3109 } | |
3110 un = ps->ps_un; | |
3111 ui = ps->ps_ui; | |
3112 | |
3113 /* | |
3114 * Do not update outstanding_writes if we're running with ABR | |
3115 * set for this mirror or the write() was issued with MD_STR_ABR set. | |
3116 * Also a resync initiated write() has no outstanding_writes update | |
3117 * either. | |
3118 */ | |
3119 if (((cb->b_flags & B_READ) == 0) && | |
3120 (un->un_nsm >= 2) && | |
3121 (ps->ps_call == NULL) && | |
3122 !((ui->ui_tstate & MD_ABR_CAP) || (ps->ps_flags & MD_MPS_ABR)) && | |
3123 !(ps->ps_flags & MD_MPS_WRITE_AFTER_READ)) { | |
3124 BLK_TO_RR(end_rr, ps->ps_lastblk, un); | |
3125 BLK_TO_RR(start_rr, ps->ps_firstblk, un); | |
3126 mutex_enter(&un->un_resync_mx); | |
3127 for (current_rr = start_rr; current_rr <= end_rr; current_rr++) | |
3128 un->un_outstanding_writes[current_rr]--; | |
3129 mutex_exit(&un->un_resync_mx); | |
3130 } | |
3131 kmem_cache_free(mirror_child_cache, cs); | |
3132 mutex_exit(&ps->ps_mx); | |
3133 | |
3134 if (ps->ps_call != NULL) { | |
3135 daemon_request(&md_done_daemon, ps->ps_call, | |
3136 (daemon_queue_t *)ps, REQ_OLD); | |
3137 return (1); | |
3138 } | |
3139 | |
3140 if ((ps->ps_flags & MD_MPS_ERROR)) { | |
3141 daemon_request(&md_done_daemon, mirror_error, | |
3142 (daemon_queue_t *)ps, REQ_OLD); | |
3143 return (1); | |
3144 } | |
3145 | |
3146 if (ps->ps_flags & MD_MPS_ON_OVERLAP) | |
3147 mirror_overlap_chain_remove(ps); | |
3148 | |
3149 /* | |
3150 * Handle Write-on-Write problem. | |
3151 * Skip In case of Raw and Direct I/O as they are | |
3152 * handled earlier. | |
3153 * | |
3154 */ | |
3155 if (!(md_mirror_wow_flg & WOW_DISABLE) && | |
3156 !(pb->b_flags & B_READ) && | |
3157 !(ps->ps_flags & MD_MPS_WOW) && | |
3158 !(pb->b_flags & B_PHYS) && | |
3159 any_pages_dirty(pb)) { | |
3160 md_unit_readerexit(ps->ps_ui); | |
3161 daemon_request(&md_mstr_daemon, handle_wow, | |
3162 (daemon_queue_t *)ps, REQ_OLD); | |
3163 return (1); | |
3164 } | |
3165 | |
3166 md_kstat_done(ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ)); | |
3167 MPS_FREE(mirror_parent_cache, ps); | |
3168 md_unit_readerexit(ui); | |
3169 md_biodone(pb); | |
3170 return (0); | |
3171 } | |
3172 | |
3173 /* | |
3174 * Clear error state in submirror component if the retry worked after | |
3175 * a failfast error. | |
3176 */ | |
3177 static void | |
3178 clear_retry_error(struct buf *cb) | |
3179 { | |
3180 int smi; | |
3181 md_mcs_t *cs; | |
3182 mm_unit_t *un; | |
3183 mdi_unit_t *ui_sm; | |
3184 mm_submirror_t *sm; | |
3185 mm_submirror_ic_t *smic; | |
3186 u_longlong_t cnt; | |
3187 md_m_shared_t *shared; | |
3188 | |
3189 /*LINTED*/ | |
3190 cs = (md_mcs_t *)((caddr_t)cb - md_mirror_mcs_buf_off); | |
3191 un = cs->cs_ps->ps_un; | |
3192 | |
3193 for (smi = 0; smi < NMIRROR; smi++) { | |
3194 if (!SMS_BY_INDEX_IS(un, smi, SMS_INUSE)) | |
3195 continue; | |
3196 | |
3197 if (cb->b_edev == md_dev64_to_dev(un->un_sm[smi].sm_dev)) { | |
3198 break; | |
3199 } | |
3200 } | |
3201 | |
3202 if (smi >= NMIRROR) | |
3203 return; | |
3204 | |
3205 sm = &un->un_sm[smi]; | |
3206 smic = &un->un_smic[smi]; | |
3207 cnt = cb->b_bcount; | |
3208 | |
3209 ui_sm = MDI_UNIT(getminor(cb->b_edev)); | |
3210 (void) md_unit_writerlock(ui_sm); | |
3211 | |
3212 shared = (md_m_shared_t *)(*(smic->sm_shared_by_blk))(sm->sm_dev, sm, | |
3213 cb->b_blkno, &cnt); | |
3214 | |
3215 if (shared->ms_flags & MDM_S_IOERR) { | |
3216 shared->ms_flags &= ~MDM_S_IOERR; | |
3217 | |
3218 } else { | |
3219 /* the I/O buf spans components and the first one is not erred */ | |
3220 int cnt; | |
3221 int i; | |
3222 | |
3223 cnt = (*(smic->sm_get_component_count))(sm->sm_dev, un); | |
3224 for (i = 0; i < cnt; i++) { | |
3225 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx)) | |
3226 (sm->sm_dev, sm, i); | |
3227 | |
3228 if (shared->ms_flags & MDM_S_IOERR && | |
3229 shared->ms_state == CS_OKAY) { | |
3230 | |
3231 shared->ms_flags &= ~MDM_S_IOERR; | |
3232 break; | |
3233 } | |
3234 } | |
3235 } | |
3236 | |
3237 md_unit_writerexit(ui_sm); | |
3238 } | |
3239 | |
3240 static size_t | |
3241 mirror_map_read( | |
3242 md_mps_t *ps, | |
3243 md_mcs_t *cs, | |
3244 diskaddr_t blkno, | |
3245 u_longlong_t count | |
3246 ) | |
3247 { | |
3248 mm_unit_t *un; | |
3249 buf_t *bp; | |
3250 u_longlong_t cando; | |
3251 | |
3252 bp = &cs->cs_buf; | |
3253 un = ps->ps_un; | |
3254 | |
3255 bp->b_lblkno = blkno; | |
3256 if (fast_select_read_unit(ps, cs) == 0) { | |
3257 bp->b_bcount = ldbtob(count); | |
3258 return (0); | |
3259 } | |
3260 bp->b_edev = md_dev64_to_dev(select_read_unit(un, blkno, count, &cando, | |
3261 0, NULL, cs)); | |
3262 bp->b_bcount = ldbtob(cando); | |
3263 if (count != cando) | |
3264 return (cando); | |
3265 return (0); | |
3266 } | |
3267 | |
3268 static void | |
3269 write_after_read(md_mps_t *ps) | |
3270 { | |
3271 struct buf *pb; | |
3272 int flags; | |
3273 | |
3274 if (ps->ps_flags & MD_MPS_ERROR) { | |
3275 mirror_error(ps); | |
3276 return; | |
3277 } | |
3278 | |
3279 pb = ps->ps_bp; | |
3280 md_kstat_done(ps->ps_ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ)); | |
3281 ps->ps_call = NULL; | |
3282 ps->ps_flags |= MD_MPS_WRITE_AFTER_READ; | |
3283 flags = MD_STR_NOTTOP | MD_STR_WAR; | |
3284 if (ps->ps_flags & MD_MPS_MAPPED) | |
3285 flags |= MD_STR_MAPPED; | |
3286 if (ps->ps_flags & MD_MPS_NOBLOCK) | |
3287 flags |= MD_NOBLOCK; | |
3288 if (ps->ps_flags & MD_MPS_DIRTY_RD) | |
3289 flags |= MD_STR_DIRTY_RD; | |
3290 (void) mirror_write_strategy(pb, flags, ps); | |
3291 } | |
3292 | |
3293 static void | |
3294 continue_serial(md_mps_t *ps) | |
3295 { | |
3296 md_mcs_t *cs; | |
3297 buf_t *cb; | |
3298 mm_unit_t *un; | |
3299 int flags; | |
3300 | |
3301 un = ps->ps_un; | |
3302 cs = kmem_cache_alloc(mirror_child_cache, MD_ALLOCFLAGS); | |
3303 mirror_child_init(cs); | |
3304 cb = &cs->cs_buf; | |
3305 ps->ps_call = NULL; | |
3306 ps->ps_frags = 1; | |
3307 (void) mirror_map_write(un, cs, ps, 0); | |
3308 flags = MD_STR_NOTTOP; | |
3309 if (ps->ps_flags & MD_MPS_MAPPED) | |
3310 flags |= MD_STR_MAPPED; | |
3311 md_call_strategy(cb, flags, NULL); | |
3312 } | |
3313 | |
3314 static int | |
3315 mirror_map_write(mm_unit_t *un, md_mcs_t *cs, md_mps_t *ps, int war) | |
3316 { | |
3317 int i; | |
3318 dev_t dev; /* needed for bioclone, so not md_dev64_t */ | |
3319 buf_t *cb; | |
3320 buf_t *pb; | |
3321 diskaddr_t blkno; | |
3322 size_t bcount; | |
3323 off_t offset; | |
3324 | |
3325 pb = ps->ps_bp; | |
3326 cb = &cs->cs_buf; | |
3327 cs->cs_ps = ps; | |
3328 | |
3329 i = md_find_nth_unit(ps->ps_writable_sm, ps->ps_current_sm); | |
3330 | |
3331 dev = md_dev64_to_dev(un->un_sm[i].sm_dev); | |
3332 | |
3333 blkno = pb->b_lblkno; | |
3334 bcount = pb->b_bcount; | |
3335 offset = 0; | |
3336 if (war && (blkno == 0) && (un->c.un_flag & MD_LABELED)) { | |
3337 blkno = DK_LABEL_LOC + 1; | |
3338 /* | |
3339 * This handles the case where we're requesting | |
3340 * a write to block 0 on a label partition | |
3341 * and the request size was smaller than the | |
3342 * size of the label. If this is the case | |
3343 * then we'll return -1. Failure to do so will | |
3344 * either cause the calling thread to hang due to | |
3345 * an ssd bug, or worse if the bcount were allowed | |
3346 * to go negative (ie large). | |
3347 */ | |
3348 if (bcount <= DEV_BSIZE*(DK_LABEL_LOC + 1)) | |
3349 return (-1); | |
3350 bcount -= (DEV_BSIZE*(DK_LABEL_LOC + 1)); | |
3351 offset = (DEV_BSIZE*(DK_LABEL_LOC + 1)); | |
3352 } | |
3353 | |
3354 cb = md_bioclone(pb, offset, bcount, dev, blkno, mirror_done, | |
3355 cb, KM_NOSLEEP); | |
3356 if (war) | |
3357 cb->b_flags = (cb->b_flags & ~B_READ) | B_WRITE; | |
3358 | |
3359 /* | |
3360 * If the submirror is in the erred stated, check if any component is | |
3361 * in the Last Erred state. If so, we don't want to use the B_FAILFAST | |
3362 * flag on the IO. | |
3363 * | |
3364 * Provide a fast path for the non-erred case (which should be the | |
3365 * normal case). | |
3366 */ | |
3367 if (un->un_sm[i].sm_flags & MD_SM_FAILFAST) { | |
3368 if (un->un_sm[i].sm_state & SMS_COMP_ERRED) { | |
3369 mm_submirror_t *sm; | |
3370 mm_submirror_ic_t *smic; | |
3371 int ci; | |
3372 int compcnt; | |
3373 | |
3374 sm = &un->un_sm[i]; | |
3375 smic = &un->un_smic[i]; | |
3376 | |
3377 compcnt = (*(smic->sm_get_component_count)) | |
3378 (sm->sm_dev, un); | |
3379 for (ci = 0; ci < compcnt; ci++) { | |
3380 md_m_shared_t *shared; | |
3381 | |
3382 shared = (md_m_shared_t *) | |
3383 (*(smic->sm_shared_by_indx))(sm->sm_dev, | |
3384 sm, ci); | |
3385 | |
3386 if (shared->ms_state == CS_LAST_ERRED) | |
3387 break; | |
3388 } | |
3389 if (ci >= compcnt) | |
3390 cb->b_flags |= B_FAILFAST; | |
3391 | |
3392 } else { | |
3393 cb->b_flags |= B_FAILFAST; | |
3394 } | |
3395 } | |
3396 | |
3397 ps->ps_current_sm++; | |
3398 if (ps->ps_current_sm != ps->ps_active_cnt) { | |
3399 if (un->un_write_option == WR_SERIAL) { | |
3400 ps->ps_call = continue_serial; | |
3401 return (0); | |
3402 } | |
3403 return (1); | |
3404 } | |
3405 return (0); | |
3406 } | |
3407 | |
3408 /* | |
3409 * directed_read_done: | |
3410 * ------------------ | |
3411 * Completion routine called when a DMR request has been returned from the | |
3412 * underlying driver. Wake-up the original ioctl() and return the data to | |
3413 * the user. | |
3414 */ | |
3415 static void | |
3416 directed_read_done(md_mps_t *ps) | |
3417 { | |
3418 mm_unit_t *un; | |
3419 mdi_unit_t *ui; | |
3420 | |
3421 un = ps->ps_un; | |
3422 ui = ps->ps_ui; | |
3423 | |
3424 md_unit_readerexit(ui); | |
3425 md_kstat_done(ui, ps->ps_bp, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ)); | |
3426 ps->ps_call = NULL; | |
3427 | |
3428 mutex_enter(&un->un_dmr_mx); | |
3429 cv_signal(&un->un_dmr_cv); | |
3430 mutex_exit(&un->un_dmr_mx); | |
3431 | |
3432 /* release the parent structure */ | |
3433 kmem_cache_free(mirror_parent_cache, ps); | |
3434 } | |
3435 | |
3436 /* | |
3437 * daemon_io: | |
3438 * ------------ | |
3439 * Called to issue a mirror_write_strategy() or mirror_read_strategy | |
3440 * call from a blockable context. NOTE: no mutex can be held on entry to this | |
3441 * routine | |
3442 */ | |
3443 static void | |
3444 daemon_io(daemon_queue_t *dq) | |
3445 { | |
3446 md_mps_t *ps = (md_mps_t *)dq; | |
3447 int flag = MD_STR_NOTTOP; | |
3448 buf_t *pb = ps->ps_bp; | |
3449 | |
3450 if (ps->ps_flags & MD_MPS_MAPPED) | |
3451 flag |= MD_STR_MAPPED; | |
3452 if (ps->ps_flags & MD_MPS_WOW) | |
3453 flag |= MD_STR_WOW; | |
3454 if (ps->ps_flags & MD_MPS_WRITE_AFTER_READ) | |
3455 flag |= MD_STR_WAR; | |
3456 if (ps->ps_flags & MD_MPS_ABR) | |
3457 flag |= MD_STR_ABR; | |
3458 | |
3459 /* | |
3460 * If this is a resync read, ie MD_STR_DIRTY_RD not set, set | |
3461 * MD_STR_WAR before calling mirror_read_strategy | |
3462 */ | |
3463 if (pb->b_flags & B_READ) { | |
3464 if (!(ps->ps_flags & MD_MPS_DIRTY_RD)) | |
3465 flag |= MD_STR_WAR; | |
3466 mirror_read_strategy(pb, flag, ps); | |
3467 } else | |
3468 mirror_write_strategy(pb, flag, ps); | |
3469 } | |
3470 | |
3471 /* | |
3472 * update_resync: | |
3473 * ------------- | |
3474 * Called to update the in-core version of the resync record with the latest | |
3475 * version that was committed to disk when the previous mirror owner | |
3476 * relinquished ownership. This call is likely to block as we must hold-off | |
3477 * any current resync processing that may be occurring. | |
3478 * On completion of the resync record update we issue the mirror_write_strategy | |
3479 * call to complete the i/o that first started this sequence. To remove a race | |
3480 * condition between a new write() request which is submitted and the resync | |
3481 * record update we acquire the writerlock. This will hold off all i/o to the | |
3482 * mirror until the resync update has completed. | |
3483 * NOTE: no mutex can be held on entry to this routine | |
3484 */ | |
3485 static void | |
3486 update_resync(daemon_queue_t *dq) | |
3487 { | |
3488 md_mps_t *ps = (md_mps_t *)dq; | |
3489 buf_t *pb = ps->ps_bp; | |
3490 mdi_unit_t *ui = ps->ps_ui; | |
3491 mm_unit_t *un; | |
3492 set_t setno; | |
3493 int restart_resync; | |
3494 | |
3495 un = md_unit_writerlock(ui); | |
3496 ps->ps_un = un; | |
3497 setno = MD_MIN2SET(getminor(pb->b_edev)); | |
3498 if (mddb_reread_rr(setno, un->un_rr_dirty_recid) == 0) { | |
3499 /* | |
3500 * Synchronize our in-core view of what regions need to be | |
3501 * resync'd with the on-disk version. | |
3502 */ | |
3503 mutex_enter(&un->un_rrp_inflight_mx); | |
3504 mirror_copy_rr(howmany(un->un_rrd_num, NBBY), un->un_resync_bm, | |
3505 un->un_dirty_bm); | |
3506 mutex_exit(&un->un_rrp_inflight_mx); | |
3507 | |
3508 /* Region dirty map is now up to date */ | |
3509 } | |
3510 restart_resync = (un->un_rs_thread_flags & MD_RI_BLOCK_OWNER) ? 1 : 0; | |
3511 md_unit_writerexit(ui); | |
3512 | |
3513 /* Restart the resync thread if it was previously blocked */ | |
3514 if (restart_resync) { | |
3515 mutex_enter(&un->un_rs_thread_mx); | |
3516 un->un_rs_thread_flags &= ~MD_RI_BLOCK_OWNER; | |
3517 cv_signal(&un->un_rs_thread_cv); | |
3518 mutex_exit(&un->un_rs_thread_mx); | |
3519 } | |
3520 /* Continue with original deferred i/o */ | |
3521 daemon_io(dq); | |
3522 } | |
3523 | |
3524 /* | |
3525 * owner_timeout: | |
3526 * ------------- | |
3527 * Called if the original mdmn_ksend_message() failed and the request is to be | |
3528 * retried. Reattempt the original ownership change. | |
3529 * | |
3530 * NOTE: called at interrupt context (see timeout(9f)). | |
3531 */ | |
3532 static void | |
3533 owner_timeout(void *arg) | |
3534 { | |
3535 daemon_queue_t *dq = (daemon_queue_t *)arg; | |
3536 | |
3537 daemon_request(&md_mirror_daemon, become_owner, dq, REQ_OLD); | |
3538 } | |
3539 | |
3540 /* | |
3541 * become_owner: | |
3542 * ------------ | |
3543 * Called to issue RPC request to become the owner of the mirror | |
3544 * associated with this i/o request. We assume that the ownership request | |
3545 * is synchronous, so if it succeeds we will issue the request via | |
3546 * mirror_write_strategy(). | |
3547 * If multiple i/o's are outstanding we will be called from the mirror_daemon | |
3548 * service thread. | |
3549 * NOTE: no mutex should be held on entry to this routine. | |
3550 */ | |
3551 static void | |
3552 become_owner(daemon_queue_t *dq) | |
3553 { | |
3554 md_mps_t *ps = (md_mps_t *)dq; | |
3555 mm_unit_t *un = ps->ps_un; | |
3556 buf_t *pb = ps->ps_bp; | |
3557 set_t setno; | |
3558 md_mn_kresult_t *kres; | |
3559 int msg_flags = md_mirror_msg_flags; | |
3560 md_mps_t *ps1; | |
3561 | |
3562 ASSERT(dq->dq_next == NULL && dq->dq_prev == NULL); | |
3563 | |
3564 /* | |
3565 * If we're already the mirror owner we do not need to send a message | |
3566 * but can simply process the i/o request immediately. | |
3567 * If we've already sent the request to become owner we requeue the | |
3568 * request as we're waiting for the synchronous ownership message to | |
3569 * be processed. | |
3570 */ | |
3571 if (MD_MN_MIRROR_OWNER(un)) { | |
3572 /* | |
3573 * As the strategy() call will potentially block we need to | |
3574 * punt this to a separate thread and complete this request | |
3575 * as quickly as possible. Note: if we're a read request | |
3576 * this must be a resync, we cannot afford to be queued | |
3577 * behind any intervening i/o requests. In this case we put the | |
3578 * request on the md_mirror_rs_daemon queue. | |
3579 */ | |
3580 if (pb->b_flags & B_READ) { | |
3581 daemon_request(&md_mirror_rs_daemon, daemon_io, dq, | |
3582 REQ_OLD); | |
3583 } else { | |
3584 daemon_request(&md_mirror_io_daemon, daemon_io, dq, | |
3585 REQ_OLD); | |
3586 } | |
3587 } else { | |
3588 mutex_enter(&un->un_owner_mx); | |
3589 if ((un->un_owner_state & MM_MN_OWNER_SENT) == 0) { | |
3590 md_mn_req_owner_t *msg; | |
3591 int rval = 0; | |
3592 | |
3593 /* | |
3594 * Check to see that we haven't exceeded the maximum | |
3595 * retry count. If we have we fail the i/o as the | |
3596 * comms mechanism has become wedged beyond recovery. | |
3597 */ | |
3598 if (dq->qlen++ >= MD_OWNER_RETRIES) { | |
3599 mutex_exit(&un->un_owner_mx); | |
3600 cmn_err(CE_WARN, | |
3601 "md_mirror: Request exhausted ownership " | |
3602 "retry limit of %d attempts", dq->qlen); | |
3603 pb->b_error = EIO; | |
3604 pb->b_flags |= B_ERROR; | |
3605 pb->b_resid = pb->b_bcount; | |
3606 kmem_cache_free(mirror_parent_cache, ps); | |
3607 md_biodone(pb); | |
3608 return; | |
3609 } | |
3610 | |
3611 /* | |
3612 * Issue request to change ownership. The call is | |
3613 * synchronous so when it returns we can complete the | |
3614 * i/o (if successful), or enqueue it again so that | |
3615 * the operation will be retried. | |
3616 */ | |
3617 un->un_owner_state |= MM_MN_OWNER_SENT; | |
3618 mutex_exit(&un->un_owner_mx); | |
3619 | |
3620 msg = kmem_zalloc(sizeof (md_mn_req_owner_t), KM_SLEEP); | |
3621 setno = MD_MIN2SET(getminor(pb->b_edev)); | |
3622 msg->mnum = MD_SID(un); | |
3623 msg->owner = md_mn_mynode_id; | |
3624 msg_flags |= MD_MSGF_NO_LOG; | |
3625 /* | |
3626 * If this IO is triggered by updating a watermark, | |
3627 * it might be issued by the creation of a softpartition | |
3628 * while the commd subsystem is suspended. | |
3629 * We don't want this message to block. | |
3630 */ | |
3631 if (ps->ps_flags & MD_MPS_WMUPDATE) { | |
3632 msg_flags |= MD_MSGF_OVERRIDE_SUSPEND; | |
3633 } | |
3634 | |
3635 kres = kmem_alloc(sizeof (md_mn_kresult_t), KM_SLEEP); | |
3636 rval = mdmn_ksend_message(setno, | |
3637 MD_MN_MSG_REQUIRE_OWNER, | |
3638 msg_flags, /* flags */ | |
3639 (char *)msg, | |
3640 sizeof (md_mn_req_owner_t), | |
3641 kres); | |
3642 | |
3643 kmem_free(msg, sizeof (md_mn_req_owner_t)); | |
3644 | |
3645 if (MDMN_KSEND_MSG_OK(rval, kres)) { | |
3646 dq->qlen = 0; | |
3647 /* | |
3648 * Successfully changed owner, reread the | |
3649 * resync record so that we have a valid idea of | |
3650 * any previously committed incomplete write()s. | |
3651 * NOTE: As we need to acquire the resync mutex | |
3652 * this may block, so we defer it to a separate | |
3653 * thread handler. This makes us (effectively) | |
3654 * non-blocking once the ownership message | |
3655 * handling has completed. | |
3656 */ | |
3657 mutex_enter(&un->un_owner_mx); | |
3658 if (un->un_owner_state & MM_MN_BECOME_OWNER) { | |
3659 un->un_mirror_owner = md_mn_mynode_id; | |
3660 /* Sets owner of un_rr_dirty record */ | |
3661 if (un->un_rr_dirty_recid) | |
3662 (void) mddb_setowner( | |
3663 un->un_rr_dirty_recid, | |
3664 md_mn_mynode_id); | |
3665 un->un_owner_state &= | |
3666 ~MM_MN_BECOME_OWNER; | |
3667 /* | |
3668 * Release the block on the current | |
3669 * resync region if it is blocked | |
3670 */ | |
3671 ps1 = un->un_rs_prev_ovrlap; | |
3672 if ((ps1 != NULL) && | |
3673 (ps1->ps_flags & MD_MPS_ON_OVERLAP)) | |
3674 mirror_overlap_chain_remove( | |
3675 ps1); | |
3676 mutex_exit(&un->un_owner_mx); | |
3677 | |
3678 /* | |
3679 * If we're a read, this must be a | |
3680 * resync request, issue | |
3681 * the i/o request on the | |
3682 * md_mirror_rs_daemon queue. This is | |
3683 * to avoid a deadlock between the | |
3684 * resync_unit thread and | |
3685 * subsequent i/o requests that may | |
3686 * block on the resync region. | |
3687 */ | |
3688 if (pb->b_flags & B_READ) { | |
3689 daemon_request( | |
3690 &md_mirror_rs_daemon, | |
3691 update_resync, dq, REQ_OLD); | |
3692 } else { | |
3693 daemon_request( | |
3694 &md_mirror_io_daemon, | |
3695 update_resync, dq, REQ_OLD); | |
3696 } | |
3697 kmem_free(kres, | |
3698 sizeof (md_mn_kresult_t)); | |
3699 return; | |
3700 } else { | |
3701 /* | |
3702 * Some other node has beaten us to | |
3703 * obtain ownership. We need to | |
3704 * reschedule our ownership request | |
3705 */ | |
3706 mutex_exit(&un->un_owner_mx); | |
3707 } | |
3708 } else { | |
3709 mdmn_ksend_show_error(rval, kres, | |
3710 "MD_MN_MSG_REQUIRE_OWNER"); | |
3711 /* | |
3712 * Message transport failure is handled by the | |
3713 * comms layer. If the ownership change request | |
3714 * does not succeed we need to flag the error to | |
3715 * the initiator of the i/o. This is handled by | |
3716 * the retry logic above. As the request failed | |
3717 * we do not know _who_ the owner of the mirror | |
3718 * currently is. We reset our idea of the owner | |
3719 * to None so that any further write()s will | |
3720 * attempt to become the owner again. This stops | |
3721 * multiple nodes writing to the same mirror | |
3722 * simultaneously. | |
3723 */ | |
3724 mutex_enter(&un->un_owner_mx); | |
3725 un->un_owner_state &= | |
3726 ~(MM_MN_OWNER_SENT|MM_MN_BECOME_OWNER); | |
3727 un->un_mirror_owner = MD_MN_MIRROR_UNOWNED; | |
3728 mutex_exit(&un->un_owner_mx); | |
3729 } | |
3730 kmem_free(kres, sizeof (md_mn_kresult_t)); | |
3731 } else | |
3732 mutex_exit(&un->un_owner_mx); | |
3733 | |
3734 /* | |
3735 * Re-enqueue this request on the deferred i/o list. Delay the | |
3736 * request for md_mirror_owner_to usecs to stop thrashing. | |
3737 */ | |
3738 (void) timeout(owner_timeout, dq, | |
3739 drv_usectohz(md_mirror_owner_to)); | |
3740 } | |
3741 } | |
3742 | |
3743 static void | |
3744 mirror_write_strategy(buf_t *pb, int flag, void *private) | |
3745 { | |
3746 md_mps_t *ps; | |
3747 md_mcs_t *cs; | |
3748 int more; | |
3749 mm_unit_t *un; | |
3750 mdi_unit_t *ui; | |
3751 buf_t *cb; /* child buf pointer */ | |
3752 set_t setno; | |
3753 int rs_on_overlap = 0; | |
3754 | |
3755 ui = MDI_UNIT(getminor(pb->b_edev)); | |
3756 un = (mm_unit_t *)MD_UNIT(getminor(pb->b_edev)); | |
3757 | |
3758 | |
3759 md_kstat_waitq_enter(ui); | |
3760 | |
3761 /* | |
3762 * If a state change is in progress for this mirror in a MN set, | |
3763 * suspend all non-resync writes until the state change is complete. | |
3764 * The objective of this suspend is to ensure that it is not | |
3765 * possible for one node to read data from a submirror that another node | |
3766 * has not written to because of the state change. Therefore we | |
3767 * suspend all writes until the state change has been made. As it is | |
3768 * not possible to read from the target of a resync, there is no need | |
3769 * to suspend resync writes. | |
3770 */ | |
3771 | |
3772 if (!(flag & MD_STR_WAR)) { | |
3773 mutex_enter(&un->un_suspend_wr_mx); | |
3774 while (un->un_suspend_wr_flag) { | |
3775 cv_wait(&un->un_suspend_wr_cv, &un->un_suspend_wr_mx); | |
3776 } | |
3777 mutex_exit(&un->un_suspend_wr_mx); | |
3778 (void) md_unit_readerlock(ui); | |
3779 } | |
3780 | |
3781 if (!(flag & MD_STR_NOTTOP)) { | |
3782 if (md_checkbuf(ui, (md_unit_t *)un, pb)) { | |
3783 md_kstat_waitq_exit(ui); | |
3784 return; | |
3785 } | |
3786 } | |
3787 | |
3788 setno = MD_MIN2SET(getminor(pb->b_edev)); | |
3789 | |
3790 /* If an ABR write has been requested, set MD_STR_ABR flag */ | |
3791 if (MD_MNSET_SETNO(setno) && (pb->b_flags & B_ABRWRITE)) | |
3792 flag |= MD_STR_ABR; | |
3793 | |
3794 if (private == NULL) { | |
3795 ps = kmem_cache_alloc(mirror_parent_cache, MD_ALLOCFLAGS); | |
3796 mirror_parent_init(ps); | |
3797 } else { | |
3798 ps = private; | |
3799 private = NULL; | |
3800 } | |
3801 if (flag & MD_STR_MAPPED) | |
3802 ps->ps_flags |= MD_MPS_MAPPED; | |
3803 | |
3804 if (flag & MD_STR_WOW) | |
3805 ps->ps_flags |= MD_MPS_WOW; | |
3806 | |
3807 if (flag & MD_STR_ABR) | |
3808 ps->ps_flags |= MD_MPS_ABR; | |
3809 | |
3810 if (flag & MD_STR_WMUPDATE) | |
3811 ps->ps_flags |= MD_MPS_WMUPDATE; | |
3812 | |
3813 /* | |
3814 * Save essential information from the original buffhdr | |
3815 * in the md_save structure. | |
3816 */ | |
3817 ps->ps_un = un; | |
3818 ps->ps_ui = ui; | |
3819 ps->ps_bp = pb; | |
3820 ps->ps_addr = pb->b_un.b_addr; | |
3821 ps->ps_firstblk = pb->b_lblkno; | |
3822 ps->ps_lastblk = pb->b_lblkno + lbtodb(pb->b_bcount) - 1; | |
3823 ps->ps_changecnt = un->un_changecnt; | |
3824 | |
3825 /* | |
3826 * If not MN owner and this is an ABR write, make sure the current | |
3827 * resync region is on the overlaps chain | |
3828 */ | |
3829 mutex_enter(&un->un_owner_mx); | |
3830 if (MD_MNSET_SETNO(setno) && (!(MD_MN_MIRROR_OWNER(un))) && | |
3831 ((ui->ui_tstate & MD_ABR_CAP) || (flag & MD_STR_ABR))) { | |
3832 md_mps_t *ps1; | |
3833 /* Block the current resync region, if not already blocked */ | |
3834 ps1 = un->un_rs_prev_ovrlap; | |
3835 | |
3836 if ((ps1 != NULL) && ((ps1->ps_firstblk != 0) || | |
3837 (ps1->ps_lastblk != 0))) { | |
3838 /* Drop locks to avoid deadlock */ | |
3839 mutex_exit(&un->un_owner_mx); | |
3840 md_unit_readerexit(ui); | |
3841 wait_for_overlaps(ps1, MD_OVERLAP_ALLOW_REPEAT); | |
3842 rs_on_overlap = 1; | |
3843 (void) md_unit_readerlock(ui); | |
3844 mutex_enter(&un->un_owner_mx); | |
3845 /* | |
3846 * Check to see if we have obtained ownership | |
3847 * while waiting for overlaps. If we have, remove | |
3848 * the resync_region entry from the overlap chain | |
3849 */ | |
3850 if (MD_MN_MIRROR_OWNER(un) && | |
3851 (ps1->ps_flags & MD_MPS_ON_OVERLAP)) { | |
3852 mirror_overlap_chain_remove(ps1); | |
3853 rs_on_overlap = 0; | |
3854 } | |
3855 } | |
3856 } | |
3857 mutex_exit(&un->un_owner_mx); | |
3858 | |
3859 | |
3860 /* | |
3861 * following keep write after read from writing to the | |
3862 * source in the case where it all came from one place | |
3863 */ | |
3864 if (flag & MD_STR_WAR) { | |
3865 int abort_write = 0; | |
3866 /* | |
3867 * We are perfoming a write-after-read. This is either as a | |
3868 * result of a resync read or as a result of a read in a | |
3869 * dirty resync region when the optimized resync is not | |
3870 * complete. If in a MN set and a resync generated i/o, | |
3871 * if the current block is not in the current | |
3872 * resync region terminate the write as another node must have | |
3873 * completed this resync region | |
3874 */ | |
3875 if ((MD_MNSET_SETNO(MD_UN2SET(un))) && | |
3876 (!flag & MD_STR_DIRTY_RD)) { | |
3877 if (!IN_RESYNC_REGION(un, ps)) | |
3878 abort_write = 1; | |
3879 } | |
3880 if ((select_write_after_read_units(un, ps) == 0) || | |
3881 (abort_write)) { | |
3882 #ifdef DEBUG | |
3883 if (mirror_debug_flag) | |
3884 printf("Abort resync write on %x, block %lld\n", | |
3885 MD_SID(un), ps->ps_firstblk); | |
3886 #endif | |
3887 if (ps->ps_flags & MD_MPS_ON_OVERLAP) | |
3888 mirror_overlap_chain_remove(ps); | |
3889 kmem_cache_free(mirror_parent_cache, ps); | |
3890 md_kstat_waitq_exit(ui); | |
3891 md_unit_readerexit(ui); | |
3892 md_biodone(pb); | |
3893 return; | |
3894 } | |
3895 } else { | |
3896 select_write_units(un, ps); | |
3897 | |
3898 /* Drop readerlock to avoid deadlock */ | |
3899 md_unit_readerexit(ui); | |
3900 wait_for_overlaps(ps, MD_OVERLAP_NO_REPEAT); | |
3901 un = md_unit_readerlock(ui); | |
3902 /* | |
3903 * For a MN set with an ABR write, if we are now the | |
3904 * owner and we have a resync region on the overlap | |
3905 * chain, remove the entry from overlaps and retry the write. | |
3906 */ | |
3907 | |
3908 if (MD_MNSET_SETNO(setno) && | |
3909 ((ui->ui_tstate & MD_ABR_CAP) || (flag & MD_STR_ABR))) { | |
3910 mutex_enter(&un->un_owner_mx); | |
3911 if (((MD_MN_MIRROR_OWNER(un))) && rs_on_overlap) { | |
3912 mirror_overlap_chain_remove(ps); | |
3913 md_kstat_waitq_exit(ui); | |
3914 mutex_exit(&un->un_owner_mx); | |
3915 md_unit_readerexit(ui); | |
3916 daemon_request(&md_mirror_daemon, daemon_io, | |
3917 (daemon_queue_t *)ps, REQ_OLD); | |
3918 return; | |
3919 } | |
3920 mutex_exit(&un->un_owner_mx); | |
3921 } | |
3922 } | |
3923 | |
3924 /* | |
3925 * For Multinode mirrors with a Resync Region (not ABR) we need to | |
3926 * become the mirror owner before continuing with the write(). For ABR | |
3927 * mirrors we check that we 'own' the resync if we're in | |
3928 * write-after-read mode. We do this _after_ ensuring that there are no | |
3929 * overlaps to ensure that the once we know that we are the owner, the | |
3930 * readerlock will not released until the write is complete. As a | |
3931 * change of ownership in a MN set requires the writerlock, this | |
3932 * ensures that ownership cannot be changed until the write is | |
3933 * complete | |
3934 */ | |
3935 if (MD_MNSET_SETNO(setno) && (!((ui->ui_tstate & MD_ABR_CAP) || | |
3936 (flag & MD_STR_ABR)) || (flag & MD_STR_WAR))) { | |
3937 if (!MD_MN_MIRROR_OWNER(un)) { | |
3938 if (ps->ps_flags & MD_MPS_ON_OVERLAP) | |
3939 mirror_overlap_chain_remove(ps); | |
3940 md_kstat_waitq_exit(ui); | |
3941 ASSERT(!(flag & MD_STR_WAR)); | |
3942 md_unit_readerexit(ui); | |
3943 daemon_request(&md_mirror_daemon, become_owner, | |
3944 (daemon_queue_t *)ps, REQ_OLD); | |
3945 return; | |
3946 } | |
3947 } | |
3948 | |
3949 /* | |
3950 * Mark resync region if mirror has a Resync Region _and_ we are not | |
3951 * a resync initiated write(). Don't mark region if we're flagged as | |
3952 * an ABR write. | |
3953 */ | |
3954 if (!((ui->ui_tstate & MD_ABR_CAP) || (flag & MD_STR_ABR)) && | |
3955 !(flag & MD_STR_WAR)) { | |
3956 if (mirror_mark_resync_region(un, ps->ps_firstblk, | |
3957 ps->ps_lastblk)) { | |
3958 pb->b_flags |= B_ERROR; | |
3959 pb->b_resid = pb->b_bcount; | |
3960 ASSERT(!(ps->ps_flags & MD_MPS_ON_OVERLAP)); | |
3961 kmem_cache_free(mirror_parent_cache, ps); | |
3962 md_kstat_waitq_exit(ui); | |
3963 md_unit_readerexit(ui); | |
3964 md_biodone(pb); | |
3965 return; | |
3966 } | |
3967 } | |
3968 | |
3969 ps->ps_childbflags = pb->b_flags | B_WRITE; | |
3970 ps->ps_childbflags &= ~B_READ; | |
3971 if (flag & MD_STR_MAPPED) | |
3972 ps->ps_childbflags &= ~B_PAGEIO; | |
3973 | |
3974 if (!(flag & MD_STR_NOTTOP) && panicstr) | |
3975 /* Disable WOW and don't free ps */ | |
3976 ps->ps_flags |= (MD_MPS_WOW|MD_MPS_DONTFREE); | |
3977 | |
3978 md_kstat_waitq_to_runq(ui); | |
3979 | |
3980 /* | |
3981 * Treat Raw and Direct I/O as Write-on-Write always | |
3982 */ | |
3983 | |
3984 if (!(md_mirror_wow_flg & WOW_DISABLE) && | |
3985 (md_mirror_wow_flg & WOW_PHYS_ENABLE) && | |
3986 (pb->b_flags & B_PHYS) && | |
3987 !(ps->ps_flags & MD_MPS_WOW)) { | |
3988 if (ps->ps_flags & MD_MPS_ON_OVERLAP) | |
3989 mirror_overlap_chain_remove(ps); | |
3990 md_unit_readerexit(ui); | |
3991 daemon_request(&md_mstr_daemon, handle_wow, | |
3992 (daemon_queue_t *)ps, REQ_OLD); | |
3993 return; | |
3994 } | |
3995 | |
3996 ps->ps_frags = 1; | |
3997 do { | |
3998 cs = kmem_cache_alloc(mirror_child_cache, MD_ALLOCFLAGS); | |
3999 mirror_child_init(cs); | |
4000 cb = &cs->cs_buf; | |
4001 more = mirror_map_write(un, cs, ps, (flag & MD_STR_WAR)); | |
4002 | |
4003 /* | |
4004 * This handles the case where we're requesting | |
4005 * a write to block 0 on a label partition. (more < 0) | |
4006 * means that the request size was smaller than the | |
4007 * size of the label. If so this request is done. | |
4008 */ | |
4009 if (more < 0) { | |
4010 if (ps->ps_flags & MD_MPS_ON_OVERLAP) | |
4011 mirror_overlap_chain_remove(ps); | |
4012 md_kstat_runq_exit(ui); | |
4013 kmem_cache_free(mirror_child_cache, cs); | |
4014 kmem_cache_free(mirror_parent_cache, ps); | |
4015 md_unit_readerexit(ui); | |
4016 md_biodone(pb); | |
4017 return; | |
4018 } | |
4019 if (more) { | |
4020 mutex_enter(&ps->ps_mx); | |
4021 ps->ps_frags++; | |
4022 mutex_exit(&ps->ps_mx); | |
4023 } | |
4024 md_call_strategy(cb, flag, private); | |
4025 } while (more); | |
4026 | |
4027 if (!(flag & MD_STR_NOTTOP) && panicstr) { | |
4028 while (!(ps->ps_flags & MD_MPS_DONE)) { | |
4029 md_daemon(1, &md_done_daemon); | |
4030 drv_usecwait(10); | |
4031 } | |
4032 kmem_cache_free(mirror_parent_cache, ps); | |
4033 } | |
4034 } | |
4035 | |
4036 static void | |
4037 mirror_read_strategy(buf_t *pb, int flag, void *private) | |
4038 { | |
4039 md_mps_t *ps; | |
4040 md_mcs_t *cs; | |
4041 size_t more; | |
4042 mm_unit_t *un; | |
4043 mdi_unit_t *ui; | |
4044 size_t current_count; | |
4045 diskaddr_t current_blkno; | |
4046 off_t current_offset; | |
4047 buf_t *cb; /* child buf pointer */ | |
4048 set_t setno; | |
4049 | |
4050 ui = MDI_UNIT(getminor(pb->b_edev)); | |
4051 | |
4052 md_kstat_waitq_enter(ui); | |
4053 | |
4054 un = (mm_unit_t *)md_unit_readerlock(ui); | |
4055 | |
4056 if (!(flag & MD_STR_NOTTOP)) { | |
4057 if (md_checkbuf(ui, (md_unit_t *)un, pb)) { | |
4058 md_kstat_waitq_exit(ui); | |
4059 return; | |
4060 } | |
4061 } | |
4062 | |
4063 if (private == NULL) { | |
4064 ps = kmem_cache_alloc(mirror_parent_cache, MD_ALLOCFLAGS); | |
4065 mirror_parent_init(ps); | |
4066 } else { | |
4067 ps = private; | |
4068 private = NULL; | |
4069 } | |
4070 | |
4071 if (flag & MD_STR_MAPPED) | |
4072 ps->ps_flags |= MD_MPS_MAPPED; | |
4073 if (flag & MD_NOBLOCK) | |
4074 ps->ps_flags |= MD_MPS_NOBLOCK; | |
4075 if (flag & MD_STR_WMUPDATE) | |
4076 ps->ps_flags |= MD_MPS_WMUPDATE; | |
4077 | |
4078 /* | |
4079 * Check to see if this is a DMR driven read. If so we need to use the | |
4080 * specified side (in un->un_dmr_last_read) for the source of the data. | |
4081 */ | |
4082 if (flag & MD_STR_DMR) | |
4083 ps->ps_flags |= MD_MPS_DMR; | |
4084 | |
4085 /* | |
4086 * Save essential information from the original buffhdr | |
4087 * in the md_save structure. | |
4088 */ | |
4089 ps->ps_un = un; | |
4090 ps->ps_ui = ui; | |
4091 ps->ps_bp = pb; | |
4092 ps->ps_addr = pb->b_un.b_addr; | |
4093 ps->ps_firstblk = pb->b_lblkno; | |
4094 ps->ps_lastblk = pb->b_lblkno + lbtodb(pb->b_bcount) - 1; | |
4095 ps->ps_changecnt = un->un_changecnt; | |
4096 | |
4097 current_count = btodb(pb->b_bcount); | |
4098 current_blkno = pb->b_lblkno; | |
4099 current_offset = 0; | |
4100 | |
4101 /* | |
4102 * If flag has MD_STR_WAR set this means that the read is issued by a | |
4103 * resync thread which may or may not be an optimised resync. | |
4104 * | |
4105 * If MD_UN_OPT_NOT_DONE is set this means that the optimized resync | |
4106 * code has not completed; either a resync has not started since snarf, | |
4107 * or there is an optimized resync in progress. | |
4108 * | |
4109 * We need to generate a write after this read in the following two | |
4110 * cases, | |
4111 * | |
4112 * 1. Any Resync-Generated read | |
4113 * | |
4114 * 2. Any read to a DIRTY REGION if there is an optimized resync | |
4115 * pending or in progress. | |
4116 * | |
4117 * The write after read is done in these cases to ensure that all sides | |
4118 * of the mirror are in sync with the read data and that it is not | |
4119 * possible for an application to read the same block multiple times | |
4120 * and get different data. | |
4121 * | |
4122 * This would be possible if the block was in a dirty region. | |
4123 * | |
4124 * If we're performing a directed read we don't write the data out as | |
4125 * the application is responsible for restoring the mirror to a known | |
4126 * state. | |
4127 */ | |
4128 if (((MD_STATUS(un) & MD_UN_OPT_NOT_DONE) || (flag & MD_STR_WAR)) && | |
4129 !(flag & MD_STR_DMR)) { | |
4130 size_t start_rr, i, end_rr; | |
4131 int region_dirty = 1; | |
4132 | |
4133 /* | |
4134 * We enter here under three circumstances, | |
4135 * | |
4136 * MD_UN_OPT_NOT_DONE MD_STR_WAR | |
4137 * 0 1 | |
4138 * 1 0 | |
4139 * 1 1 | |
4140 * | |
4141 * To be optimal we only care to explicitly check for dirty | |
4142 * regions in the second case since if MD_STR_WAR is set we | |
4143 * always do the write after read. | |
4144 */ | |
4145 if (!(flag & MD_STR_WAR)) { | |
4146 BLK_TO_RR(end_rr, ps->ps_lastblk, un); | |
4147 BLK_TO_RR(start_rr, ps->ps_firstblk, un); | |
4148 | |
4149 for (i = start_rr; i <= end_rr; i++) | |
4150 if ((region_dirty = IS_KEEPDIRTY(i, un)) != 0) | |
4151 break; | |
4152 } | |
4153 | |
4154 if ((region_dirty) && | |
4155 !(md_get_setstatus(MD_UN2SET(un)) & MD_SET_STALE)) { | |
4156 ps->ps_call = write_after_read; | |
4157 /* | |
4158 * Mark this as a RESYNC_READ in ps_flags. | |
4159 * This is used if the read fails during a | |
4160 * resync of a 3-way mirror to ensure that | |
4161 * the retried read to the remaining | |
4162 * good submirror has MD_STR_WAR set. This | |
4163 * is needed to ensure that the resync write | |
4164 * (write-after-read) takes place. | |
4165 */ | |
4166 ps->ps_flags |= MD_MPS_RESYNC_READ; | |
4167 | |
4168 /* | |
4169 * If MD_STR_FLAG_ERR is set in the flags we | |
4170 * set MD_MPS_FLAG_ERROR so that an error on the resync | |
4171 * write (issued by write_after_read) will be flagged | |
4172 * to the biowait'ing resync thread. This allows us to | |
4173 * avoid issuing further resync requests to a device | |
4174 * that has had a write failure. | |
4175 */ | |
4176 if (flag & MD_STR_FLAG_ERR) | |
4177 ps->ps_flags |= MD_MPS_FLAG_ERROR; | |
4178 | |
4179 setno = MD_UN2SET(un); | |
4180 /* | |
4181 * Drop the readerlock to avoid | |
4182 * deadlock | |
4183 */ | |
4184 md_unit_readerexit(ui); | |
4185 wait_for_overlaps(ps, MD_OVERLAP_NO_REPEAT); | |
4186 un = md_unit_readerlock(ui); | |
4187 /* | |
4188 * Ensure that we are owner | |
4189 */ | |
4190 if (MD_MNSET_SETNO(setno)) { | |
4191 /* | |
4192 * For a non-resync read that requires a | |
4193 * write-after-read to be done, set a flag | |
4194 * in the parent structure, so that the | |
4195 * write_strategy routine can omit the | |
4196 * test that the write is still within the | |
4197 * resync region | |
4198 */ | |
4199 if (!(flag & MD_STR_WAR)) | |
4200 ps->ps_flags |= MD_MPS_DIRTY_RD; | |
4201 | |
4202 /* | |
4203 * Before reading the buffer, see if | |
4204 * we are the owner | |
4205 */ | |
4206 if (!MD_MN_MIRROR_OWNER(un)) { | |
4207 ps->ps_call = NULL; | |
4208 mirror_overlap_chain_remove(ps); | |
4209 md_kstat_waitq_exit(ui); | |
4210 md_unit_readerexit(ui); | |
4211 daemon_request( | |
4212 &md_mirror_daemon, | |
4213 become_owner, | |
4214 (daemon_queue_t *)ps, | |
4215 REQ_OLD); | |
4216 return; | |
4217 } | |
4218 /* | |
4219 * For a resync read, check to see if I/O is | |
4220 * outside of the current resync region, or | |
4221 * the resync has finished. If so | |
4222 * just terminate the I/O | |
4223 */ | |
4224 if ((flag & MD_STR_WAR) && | |
4225 (!(un->c.un_status & MD_UN_WAR) || | |
4226 (!IN_RESYNC_REGION(un, ps)))) { | |
4227 #ifdef DEBUG | |
4228 if (mirror_debug_flag) | |
4229 printf("Abort resync read " | |
4230 "%x: %lld\n", | |
4231 MD_SID(un), | |
4232 ps->ps_firstblk); | |
4233 #endif | |
4234 mirror_overlap_chain_remove(ps); | |
4235 kmem_cache_free(mirror_parent_cache, | |
4236 ps); | |
4237 md_kstat_waitq_exit(ui); | |
4238 md_unit_readerexit(ui); | |
4239 md_biodone(pb); | |
4240 return; | |
4241 } | |
4242 } | |
4243 } | |
4244 } | |
4245 | |
4246 if (flag & MD_STR_DMR) { | |
4247 ps->ps_call = directed_read_done; | |
4248 } | |
4249 | |
4250 if (!(flag & MD_STR_NOTTOP) && panicstr) | |
4251 ps->ps_flags |= MD_MPS_DONTFREE; | |
4252 | |
4253 md_kstat_waitq_to_runq(ui); | |
4254 | |
4255 ps->ps_frags++; | |
4256 do { | |
4257 cs = kmem_cache_alloc(mirror_child_cache, MD_ALLOCFLAGS); | |
4258 mirror_child_init(cs); | |
4259 cb = &cs->cs_buf; | |
4260 cs->cs_ps = ps; | |
4261 | |
4262 cb = md_bioclone(pb, current_offset, current_count, NODEV, | |
4263 current_blkno, mirror_done, cb, KM_NOSLEEP); | |
4264 | |
4265 more = mirror_map_read(ps, cs, current_blkno, | |
4266 (u_longlong_t)current_count); | |
4267 if (more) { | |
4268 mutex_enter(&ps->ps_mx); | |
4269 ps->ps_frags++; | |
4270 mutex_exit(&ps->ps_mx); | |
4271 } | |
4272 | |
4273 /* | |
4274 * Do these calculations now, | |
4275 * so that we pickup a valid b_bcount from the chld_bp. | |
4276 */ | |
4277 current_count -= more; | |
4278 current_offset += cb->b_bcount; | |
4279 current_blkno += more; | |
4280 md_call_strategy(cb, flag, private); | |
4281 } while (more); | |
4282 | |
4283 if (!(flag & MD_STR_NOTTOP) && panicstr) { | |
4284 while (!(ps->ps_flags & MD_MPS_DONE)) { | |
4285 md_daemon(1, &md_done_daemon); | |
4286 drv_usecwait(10); | |
4287 } | |
4288 kmem_cache_free(mirror_parent_cache, ps); | |
4289 } | |
4290 } | |
4291 | |
4292 void | |
4293 md_mirror_strategy(buf_t *bp, int flag, void *private) | |
4294 { | |
4295 set_t setno = MD_MIN2SET(getminor(bp->b_edev)); | |
4296 | |
4297 /* | |
4298 * When doing IO to a multi owner meta device, check if set is halted. | |
4299 * We do this check without the needed lock held, for performance | |
4300 * reasons. | |
4301 * If an IO just slips through while the set is locked via an | |
4302 * MD_MN_SUSPEND_SET, we don't care about it. | |
4303 * Only check for suspension if we are a top-level i/o request | |
4304 * (MD_STR_NOTTOP is cleared in 'flag'). | |
4305 */ | |
4306 if ((md_set[setno].s_status & (MD_SET_HALTED | MD_SET_MNSET)) == | |
4307 (MD_SET_HALTED | MD_SET_MNSET)) { | |
4308 if ((flag & MD_STR_NOTTOP) == 0) { | |
4309 mutex_enter(&md_mx); | |
4310 /* Here we loop until the set is no longer halted */ | |
4311 while (md_set[setno].s_status & MD_SET_HALTED) { | |
4312 cv_wait(&md_cv, &md_mx); | |
4313 } | |
4314 mutex_exit(&md_mx); | |
4315 } | |
4316 } | |
4317 | |
4318 if ((flag & MD_IO_COUNTED) == 0) { | |
4319 if ((flag & MD_NOBLOCK) == 0) { | |
4320 if (md_inc_iocount(setno) != 0) { | |
4321 bp->b_flags |= B_ERROR; | |
4322 bp->b_error = ENXIO; | |
4323 bp->b_resid = bp->b_bcount; | |
4324 biodone(bp); | |
4325 return; | |
4326 } | |
4327 } else { | |
4328 md_inc_iocount_noblock(setno); | |
4329 } | |
4330 } | |
4331 | |
4332 if (bp->b_flags & B_READ) | |
4333 mirror_read_strategy(bp, flag, private); | |
4334 else | |
4335 mirror_write_strategy(bp, flag, private); | |
4336 } | |
4337 | |
4338 /* | |
4339 * mirror_directed_read: | |
4340 * -------------------- | |
4341 * Entry-point for the DKIOCDMR ioctl. We issue a read to a specified sub-mirror | |
4342 * so that the application can determine what (if any) resync needs to be | |
4343 * performed. The data is copied out to the user-supplied buffer. | |
4344 * | |
4345 * Parameters: | |
4346 * mdev - dev_t for the mirror device | |
4347 * vdr - directed read parameters specifying location and submirror | |
4348 * to perform the read from | |
4349 * mode - used to ddi_copyout() any resulting data from the read | |
4350 * | |
4351 * Returns: | |
4352 * 0 success | |
4353 * !0 error code | |
4354 * EINVAL - invalid request format | |
4355 */ | |
4356 int | |
4357 mirror_directed_read(dev_t mdev, vol_directed_rd_t *vdr, int mode) | |
4358 { | |
4359 buf_t *bp; | |
4360 minor_t mnum = getminor(mdev); | |
4361 mdi_unit_t *ui = MDI_UNIT(mnum); | |
4362 mm_unit_t *un; | |
4363 mm_submirror_t *sm; | |
4364 char *sm_nm; | |
4365 uint_t next_side; | |
4366 void *kbuffer; | |
4367 | |
4368 if (ui == NULL) | |
4369 return (ENXIO); | |
4370 | |
4371 if (!(vdr->vdr_flags & DKV_DMR_NEXT_SIDE)) { | |
4372 return (EINVAL); | |
4373 } | |
4374 | |
4375 /* Check for aligned block access. We disallow non-aligned requests. */ | |
4376 if (vdr->vdr_offset % DEV_BSIZE) { | |
4377 return (EINVAL); | |
4378 } | |
4379 | |
4380 /* | |
4381 * Allocate kernel buffer for target of read(). If we had a reliable | |
4382 * (sorry functional) DDI this wouldn't be needed. | |
4383 */ | |
4384 kbuffer = kmem_alloc(vdr->vdr_nbytes, KM_NOSLEEP); | |
4385 if (kbuffer == NULL) { | |
4386 cmn_err(CE_WARN, "mirror_directed_read: couldn't allocate %lx" | |
4387 " bytes\n", vdr->vdr_nbytes); | |
4388 return (ENOMEM); | |
4389 } | |
4390 | |
4391 bp = getrbuf(KM_SLEEP); | |
4392 | |
4393 bp->b_un.b_addr = kbuffer; | |
4394 bp->b_flags = B_READ; | |
4395 bp->b_bcount = vdr->vdr_nbytes; | |
4396 bp->b_lblkno = lbtodb(vdr->vdr_offset); | |
4397 bp->b_edev = mdev; | |
4398 | |
4399 un = md_unit_readerlock(ui); | |
4400 | |
4401 /* | |
4402 * If DKV_SIDE_INIT is set we need to determine the first available | |
4403 * side to start reading from. If it isn't set we increment to the | |
4404 * next readable submirror. | |
4405 * If there are no readable submirrors we error out with DKV_DMR_ERROR. | |
4406 * Note: we check for a readable submirror on completion of the i/o so | |
4407 * we should _always_ have one available. If this becomes unavailable | |
4408 * we have missed the 'DKV_DMR_DONE' opportunity. This could happen if | |
4409 * a metadetach is made between the completion of one DKIOCDMR ioctl | |
4410 * and the start of the next (i.e. a sys-admin 'accident' occurred). | |
4411 * The chance of this is small, but not non-existent. | |
4412 */ | |
4413 if (vdr->vdr_side == DKV_SIDE_INIT) { | |
4414 next_side = 0; | |
4415 } else { | |
4416 next_side = vdr->vdr_side + 1; | |
4417 } | |
4418 while ((next_side < NMIRROR) && | |
4419 !SUBMIRROR_IS_READABLE(un, next_side)) | |
4420 next_side++; | |
4421 if (next_side >= NMIRROR) { | |
4422 vdr->vdr_flags |= DKV_DMR_ERROR; | |
4423 freerbuf(bp); | |
4424 vdr->vdr_bytesread = 0; | |
4425 md_unit_readerexit(ui); | |
4426 return (0); | |
4427 } | |
4428 | |
4429 /* Set the side to read from */ | |
4430 un->un_dmr_last_read = next_side; | |
4431 | |
4432 md_unit_readerexit(ui); | |
4433 | |
4434 /* | |
4435 * Save timestamp for verification purposes. Can be read by debugger | |
4436 * to verify that this ioctl has been executed and to find the number | |
4437 * of DMR reads and the time of the last DMR read. | |
4438 */ | |
4439 uniqtime(&mirror_dmr_stats.dmr_timestamp); | |
4440 mirror_dmr_stats.dmr_count++; | |
4441 | |
4442 /* Issue READ request and wait for completion */ | |
4443 mirror_read_strategy(bp, MD_STR_DMR|MD_NOBLOCK|MD_STR_NOTTOP, NULL); | |
4444 | |
4445 mutex_enter(&un->un_dmr_mx); | |
4446 cv_wait(&un->un_dmr_cv, &un->un_dmr_mx); | |
4447 mutex_exit(&un->un_dmr_mx); | |
4448 | |
4449 /* | |
4450 * Check to see if we encountered an error during the read. If so we | |
4451 * can make no guarantee about any possibly returned data. | |
4452 */ | |
4453 if ((bp->b_flags & B_ERROR) == 0) { | |
4454 vdr->vdr_flags &= ~DKV_DMR_ERROR; | |
4455 if (bp->b_resid) { | |
4456 vdr->vdr_flags |= DKV_DMR_SHORT; | |
4457 vdr->vdr_bytesread = vdr->vdr_nbytes - bp->b_resid; | |
4458 } else { | |
4459 vdr->vdr_flags |= DKV_DMR_SUCCESS; | |
4460 vdr->vdr_bytesread = vdr->vdr_nbytes; | |
4461 } | |
4462 /* Copy the data read back out to the user supplied buffer */ | |
4463 if (ddi_copyout(kbuffer, vdr->vdr_data, vdr->vdr_bytesread, | |
4464 mode)) { | |
4465 kmem_free(kbuffer, vdr->vdr_nbytes); | |
4466 return (EFAULT); | |
4467 } | |
4468 | |
4469 } else { | |
4470 /* Error out with DKV_DMR_ERROR */ | |
4471 vdr->vdr_flags |= DKV_DMR_ERROR; | |
4472 vdr->vdr_flags &= ~(DKV_DMR_SUCCESS|DKV_DMR_SHORT|DKV_DMR_DONE); | |
4473 } | |
4474 /* | |
4475 * Update the DMR parameters with the side and name of submirror that | |
4476 * we have just read from (un->un_dmr_last_read) | |
4477 */ | |
4478 un = md_unit_readerlock(ui); | |
4479 | |
4480 vdr->vdr_side = un->un_dmr_last_read; | |
4481 sm = &un->un_sm[un->un_dmr_last_read]; | |
4482 sm_nm = md_shortname(md_getminor(sm->sm_dev)); | |
4483 | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
4484 (void) strncpy(vdr->vdr_side_name, sm_nm, sizeof (vdr->vdr_side_name)); |
0 | 4485 |
4486 /* | |
4487 * Determine if we've completed the read cycle. This is true iff the | |
4488 * next computed submirror (side) equals or exceeds NMIRROR. We cannot | |
4489 * use un_nsm as we need to handle a sparse array of submirrors (which | |
4490 * can occur if a submirror is metadetached). | |
4491 */ | |
4492 next_side = un->un_dmr_last_read + 1; | |
4493 while ((next_side < NMIRROR) && | |
4494 !SUBMIRROR_IS_READABLE(un, next_side)) | |
4495 next_side++; | |
4496 if (next_side >= NMIRROR) { | |
4497 /* We've finished */ | |
4498 vdr->vdr_flags |= DKV_DMR_DONE; | |
4499 } | |
4500 | |
4501 md_unit_readerexit(ui); | |
4502 freerbuf(bp); | |
4503 kmem_free(kbuffer, vdr->vdr_nbytes); | |
4504 | |
4505 return (0); | |
4506 } | |
4507 | |
4508 /* | |
4509 * mirror_resync_message: | |
4510 * --------------------- | |
4511 * Handle the multi-node resync messages that keep all nodes within a given | |
4512 * disk-set in sync with their view of a mirror's resync status. | |
4513 * | |
4514 * The message types dealt with are: | |
4515 * MD_MN_MSG_RESYNC_STARTING - start a resync thread for a unit | |
4516 * MD_MN_MSG_RESYNC_NEXT - specified next region to be resynced | |
4517 * MD_MN_MSG_RESYNC_FINISH - stop the resync thread for a unit | |
4518 * MD_MN_MSG_RESYNC_PHASE_DONE - end of a resync phase, opt, submirror or comp | |
4519 * | |
4520 * Returns: | |
4521 * 0 Success | |
4522 * >0 Failure error number | |
4523 */ | |
4524 int | |
4525 mirror_resync_message(md_mn_rs_params_t *p, IOLOCK *lockp) | |
4526 { | |
4527 mdi_unit_t *ui; | |
4528 mm_unit_t *un; | |
4529 set_t setno; | |
4530 int is_ABR; | |
4531 int smi; | |
4532 int ci; | |
4533 sm_state_t state; | |
4534 int broke_out; | |
4535 mm_submirror_t *sm; | |
4536 mm_submirror_ic_t *smic; | |
4537 md_m_shared_t *shared; | |
4538 md_error_t mde = mdnullerror; | |
4539 md_mps_t *ps; | |
4540 int rs_active; | |
4541 | |
4542 /* Check that the given device is part of a multi-node set */ | |
4543 setno = MD_MIN2SET(p->mnum); | |
4544 if (setno >= md_nsets) { | |
4545 return (ENXIO); | |
4546 } | |
4547 if (!MD_MNSET_SETNO(setno)) { | |
4548 return (EINVAL); | |
4549 } | |
4550 | |
4551 if ((un = mirror_getun(p->mnum, &p->mde, NO_LOCK, NULL)) == NULL) | |
4552 return (EINVAL); | |
4553 if ((ui = MDI_UNIT(p->mnum)) == NULL) | |
4554 return (EINVAL); | |
4555 is_ABR = (ui->ui_tstate & MD_ABR_CAP); | |
4556 | |
4557 /* Obtain the current resync status */ | |
4558 (void) md_ioctl_readerlock(lockp, ui); | |
4559 rs_active = (MD_STATUS(un) & MD_UN_RESYNC_ACTIVE) ? 1 : 0; | |
4560 md_ioctl_readerexit(lockp); | |
4561 | |
4562 switch ((md_mn_msgtype_t)p->msg_type) { | |
4563 case MD_MN_MSG_RESYNC_STARTING: | |
4564 /* Start the resync thread for the mirror */ | |
4565 (void) mirror_resync_unit(p->mnum, NULL, &p->mde, lockp); | |
4566 break; | |
4567 | |
4568 case MD_MN_MSG_RESYNC_NEXT: | |
4569 /* | |
4570 * We have to release any previously marked overlap regions | |
4571 * so that i/o can resume. Then we need to block the region | |
4572 * from [rs_start..rs_start+rs_size) * so that no i/o is issued. | |
4573 * Update un_rs_resync_done and un_rs_resync_2_do. | |
4574 */ | |
4575 (void) md_ioctl_readerlock(lockp, ui); | |
4576 /* | |
4577 * Ignore the message if there is no active resync thread or | |
4578 * if it is for a resync type that we have already completed. | |
4579 * un_resync_completed is set to the last resync completed | |
4580 * when processing a PHASE_DONE message. | |
4581 */ | |
4582 if (!rs_active || (p->rs_type == un->un_resync_completed)) | |
4583 break; | |
4584 /* | |
4585 * If this message is for the same resync and is for an earlier | |
4586 * resync region, just ignore it. This can only occur if this | |
4587 * node has progressed on to the next resync region before | |
4588 * we receive this message. This can occur if the class for | |
4589 * this message is busy and the originator has to retry thus | |
4590 * allowing this node to move onto the next resync_region. | |
4591 */ | |
4592 if ((p->rs_type == un->un_rs_type) && | |
4593 (p->rs_start < un->un_resync_startbl)) | |
4594 break; | |
4595 ps = un->un_rs_prev_ovrlap; | |
4596 | |
4597 /* Allocate previous overlap reference if needed */ | |
4598 if (ps == NULL) { | |
4599 ps = kmem_cache_alloc(mirror_parent_cache, | |
4600 MD_ALLOCFLAGS); | |
4601 ps->ps_un = un; | |
4602 ps->ps_ui = ui; | |
4603 ps->ps_firstblk = 0; | |
4604 ps->ps_lastblk = 0; | |
4605 ps->ps_flags = 0; | |
4606 md_ioctl_readerexit(lockp); | |
4607 (void) md_ioctl_writerlock(lockp, ui); | |
4608 un->un_rs_prev_ovrlap = ps; | |
4609 md_ioctl_writerexit(lockp); | |
4610 } else | |
4611 md_ioctl_readerexit(lockp); | |
4612 | |
4613 if (p->rs_originator != md_mn_mynode_id) { | |
4614 /* | |
4615 * On all but the originating node, first update | |
4616 * the resync state, then unblock the previous | |
4617 * region and block the next one. No need | |
4618 * to do this if the region is already blocked. | |
4619 * Update the submirror state and flags from the | |
4620 * originator. This keeps the cluster in sync with | |
4621 * regards to the resync status. | |
4622 */ | |
4623 | |
4624 (void) md_ioctl_writerlock(lockp, ui); | |
4625 un->un_rs_resync_done = p->rs_done; | |
4626 un->un_rs_resync_2_do = p->rs_2_do; | |
4627 un->un_rs_type = p->rs_type; | |
4628 un->un_resync_startbl = p->rs_start; | |
4629 md_ioctl_writerexit(lockp); | |
4630 /* | |
4631 * Use un_owner_mx to ensure that an ownership change | |
4632 * cannot happen at the same time as this message | |
4633 */ | |
4634 mutex_enter(&un->un_owner_mx); | |
4635 if (MD_MN_MIRROR_OWNER(un)) { | |
4636 ps->ps_firstblk = p->rs_start; | |
4637 ps->ps_lastblk = ps->ps_firstblk + | |
4638 p->rs_size - 1; | |
4639 } else { | |
4640 if ((ps->ps_firstblk != p->rs_start) || | |
4641 (ps->ps_lastblk != p->rs_start + | |
4642 p->rs_size - 1)) { | |
4643 /* Remove previous overlap range */ | |
4644 if (ps->ps_flags & MD_MPS_ON_OVERLAP) | |
4645 mirror_overlap_chain_remove(ps); | |
4646 | |
4647 ps->ps_firstblk = p->rs_start; | |
4648 ps->ps_lastblk = ps->ps_firstblk + | |
4649 p->rs_size - 1; | |
4650 | |
4651 mutex_exit(&un->un_owner_mx); | |
4652 /* Block this range from all i/o. */ | |
4653 if (ps->ps_firstblk != 0 || | |
4654 ps->ps_lastblk != 0) | |
4655 wait_for_overlaps(ps, | |
4656 MD_OVERLAP_ALLOW_REPEAT); | |
4657 mutex_enter(&un->un_owner_mx); | |
4658 /* | |
4659 * Check to see if we have obtained | |
4660 * ownership while waiting for | |
4661 * overlaps. If we have, remove | |
4662 * the resync_region entry from the | |
4663 * overlap chain | |
4664 */ | |
4665 if (MD_MN_MIRROR_OWNER(un) && | |
4666 (ps->ps_flags & MD_MPS_ON_OVERLAP)) | |
4667 mirror_overlap_chain_remove(ps); | |
4668 } | |
4669 } | |
4670 mutex_exit(&un->un_owner_mx); | |
4671 | |
4672 /* | |
4673 * If this is the first RESYNC_NEXT message (i.e. | |
4674 * MD_MN_RS_FIRST_RESYNC_NEXT set in p->rs_flags), | |
4675 * issue RESYNC_START NOTIFY event | |
4676 */ | |
4677 if (p->rs_flags & MD_MN_RS_FIRST_RESYNC_NEXT) { | |
4678 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_RESYNC_START, | |
4679 SVM_TAG_METADEVICE, MD_UN2SET(un), | |
4680 MD_SID(un)); | |
4681 } | |
4682 | |
4683 /* Ensure that our local resync thread is running */ | |
4684 if (un->un_rs_thread == NULL) { | |
4685 (void) mirror_resync_unit(p->mnum, NULL, | |
4686 &p->mde, lockp); | |
4687 } | |
4688 } | |
4689 break; | |
4690 case MD_MN_MSG_RESYNC_FINISH: | |
4691 /* | |
4692 * Complete the resync by stopping the resync thread. | |
4693 * Also release the previous overlap region field. | |
4694 * Update the resync_progress_thread by cv_signal'ing it so | |
4695 * that we mark the end of the resync as soon as possible. This | |
4696 * stops an unnecessary delay should be panic after resync | |
4697 * completion. | |
4698 */ | |
4699 #ifdef DEBUG | |
4700 if (!rs_active) { | |
4701 if (mirror_debug_flag) | |
4702 printf("RESYNC_FINISH (mnum = %x), " | |
4703 "Resync *NOT* active", | |
4704 p->mnum); | |
4705 } | |
4706 #endif | |
4707 | |
4708 if ((un->c.un_status & MD_UN_RESYNC_ACTIVE) && | |
4709 (p->rs_originator != md_mn_mynode_id)) { | |
4710 mutex_enter(&un->un_rs_thread_mx); | |
4711 un->c.un_status &= ~MD_UN_RESYNC_CANCEL; | |
4712 un->un_rs_thread_flags |= MD_RI_SHUTDOWN; | |
4713 un->un_rs_thread_flags &= | |
4714 ~(MD_RI_BLOCK|MD_RI_BLOCK_OWNER); | |
4715 cv_signal(&un->un_rs_thread_cv); | |
4716 mutex_exit(&un->un_rs_thread_mx); | |
4717 } | |
4718 if (is_ABR) { | |
4719 /* Resync finished, if ABR set owner to NULL */ | |
4720 mutex_enter(&un->un_owner_mx); | |
4721 un->un_mirror_owner = 0; | |
4722 mutex_exit(&un->un_owner_mx); | |
4723 } | |
4724 (void) md_ioctl_writerlock(lockp, ui); | |
4725 ps = un->un_rs_prev_ovrlap; | |
4726 if (ps != NULL) { | |
4727 /* Remove previous overlap range */ | |
4728 if (ps->ps_flags & MD_MPS_ON_OVERLAP) | |
4729 mirror_overlap_chain_remove(ps); | |
4730 /* | |
4731 * Release the overlap range reference | |
4732 */ | |
4733 un->un_rs_prev_ovrlap = NULL; | |
4734 kmem_cache_free(mirror_parent_cache, | |
4735 ps); | |
4736 } | |
4737 md_ioctl_writerexit(lockp); | |
4738 | |
4739 /* Mark the resync as complete in the metadb */ | |
4740 un->un_rs_resync_done = p->rs_done; | |
4741 un->un_rs_resync_2_do = p->rs_2_do; | |
4742 un->un_rs_type = p->rs_type; | |
4743 mutex_enter(&un->un_rs_progress_mx); | |
4744 cv_signal(&un->un_rs_progress_cv); | |
4745 mutex_exit(&un->un_rs_progress_mx); | |
4746 | |
4747 un = md_ioctl_writerlock(lockp, ui); | |
4748 un->c.un_status &= ~MD_UN_RESYNC_ACTIVE; | |
4749 /* Deal with any pending grow_unit */ | |
4750 if (un->c.un_status & MD_UN_GROW_PENDING) { | |
4751 if ((mirror_grow_unit(un, &mde) != 0) || | |
4752 (! mdismderror(&mde, MDE_GROW_DELAYED))) { | |
4753 un->c.un_status &= ~MD_UN_GROW_PENDING; | |
4754 } | |
4755 } | |
4756 md_ioctl_writerexit(lockp); | |
4757 break; | |
4758 | |
4759 case MD_MN_MSG_RESYNC_PHASE_DONE: | |
4760 /* | |
4761 * A phase of the resync, optimized. component or | |
4762 * submirror is complete. Update mirror status. | |
4763 * If the flag CLEAR_OPT_NOT_DONE is set, it means that the | |
4764 * mirror owner is peforming a resync. If we have just snarfed | |
4765 * this set, then we must clear any of the flags set at snarf | |
4766 * time by unit_setup_resync(). | |
4767 * Note that unit_setup_resync() sets up these flags to | |
4768 * indicate that an optimized resync is required. These flags | |
4769 * need to be reset because if we get here, the mirror owner | |
4770 * will have handled the optimized resync. | |
4771 * The flags that must be cleared are MD_UN_OPT_NOT_DONE and | |
4772 * MD_UN_WAR. In addition, for each submirror, | |
4773 * MD_SM_RESYNC_TARGET must be cleared and SMS_OFFLINE_RESYNC | |
4774 * set to SMS_OFFLINE. | |
4775 */ | |
4776 #ifdef DEBUG | |
4777 if (mirror_debug_flag) | |
4778 printf("phase done mess received from %d, mnum=%x," | |
4779 "type=%x, flags=%x\n", p->rs_originator, p->mnum, | |
4780 p->rs_type, p->rs_flags); | |
4781 #endif | |
4782 /* | |
4783 * Ignore the message if there is no active resync thread. | |
4784 */ | |
4785 if (!rs_active) | |
4786 break; | |
4787 | |
4788 broke_out = p->rs_flags & MD_MN_RS_ERR; | |
4789 switch (RS_TYPE(p->rs_type)) { | |
4790 case MD_RS_OPTIMIZED: | |
4791 un = md_ioctl_writerlock(lockp, ui); | |
4792 if (p->rs_flags & MD_MN_RS_CLEAR_OPT_NOT_DONE) { | |
4793 /* If we are originator, just clear rs_type */ | |
4794 if (p->rs_originator == md_mn_mynode_id) { | |
4795 SET_RS_TYPE_NONE(un->un_rs_type); | |
4796 md_ioctl_writerexit(lockp); | |
4797 break; | |
4798 } | |
4799 /* | |
4800 * If CLEAR_OPT_NOT_DONE is set, only clear the | |
4801 * flags if OPT_NOT_DONE is set *and* rs_type | |
4802 * is MD_RS_NONE. | |
4803 */ | |
4804 if ((un->c.un_status & MD_UN_OPT_NOT_DONE) && | |
4805 (RS_TYPE(un->un_rs_type) == MD_RS_NONE)) { | |
4806 /* No resync in progress */ | |
4807 un->c.un_status &= ~MD_UN_OPT_NOT_DONE; | |
4808 un->c.un_status &= ~MD_UN_WAR; | |
4809 } else { | |
4810 /* | |
4811 * We are in the middle of an | |
4812 * optimized resync and this message | |
4813 * should be ignored. | |
4814 */ | |
4815 md_ioctl_writerexit(lockp); | |
4816 break; | |
4817 } | |
4818 } else { | |
4819 /* | |
4820 * This is the end of an optimized resync, | |
4821 * clear the OPT_NOT_DONE and OFFLINE_SM flags | |
4822 */ | |
4823 | |
4824 un->c.un_status &= ~MD_UN_KEEP_DIRTY; | |
4825 if (!broke_out) | |
4826 un->c.un_status &= ~MD_UN_WAR; | |
4827 } | |
4828 | |
4829 /* | |
4830 * Set resync_completed to last resync type and then | |
4831 * clear resync_type to indicate no resync in progress | |
4832 */ | |
4833 un->un_resync_completed = un->un_rs_type; | |
4834 SET_RS_TYPE_NONE(un->un_rs_type); | |
4835 | |
4836 /* | |
4837 * If resync is as a result of a submirror ONLINE, | |
4838 * reset the submirror state to SMS_RUNNING if the | |
4839 * resync was ok else set back to SMS_OFFLINE. | |
4840 */ | |
4841 for (smi = 0; smi < NMIRROR; smi++) { | |
4842 un->un_sm[smi].sm_flags &= | |
4843 ~MD_SM_RESYNC_TARGET; | |
4844 if (SMS_BY_INDEX_IS(un, smi, | |
4845 SMS_OFFLINE_RESYNC)) { | |
4846 if (p->rs_flags & | |
4847 MD_MN_RS_CLEAR_OPT_NOT_DONE) { | |
4848 state = SMS_OFFLINE; | |
4849 } else { | |
4850 state = (broke_out ? | |
4851 SMS_OFFLINE : SMS_RUNNING); | |
4852 } | |
4853 mirror_set_sm_state( | |
4854 &un->un_sm[smi], | |
4855 &un->un_smic[smi], state, | |
4856 broke_out); | |
4857 mirror_commit(un, NO_SUBMIRRORS, | |
4858 0); | |
4859 } | |
4860 /* | |
4861 * If we still have an offline submirror, reset | |
4862 * the OFFLINE_SM flag in the mirror status | |
4863 */ | |
4864 if (SMS_BY_INDEX_IS(un, smi, | |
4865 SMS_OFFLINE)) | |
4866 un->c.un_status |= | |
4867 MD_UN_OFFLINE_SM; | |
4868 } | |
4869 md_ioctl_writerexit(lockp); | |
4870 break; | |
4871 case MD_RS_SUBMIRROR: | |
4872 un = md_ioctl_writerlock(lockp, ui); | |
4873 smi = RS_SMI(p->rs_type); | |
4874 sm = &un->un_sm[smi]; | |
4875 smic = &un->un_smic[smi]; | |
4876 /* Clear RESYNC target */ | |
4877 un->un_sm[smi].sm_flags &= ~MD_SM_RESYNC_TARGET; | |
4878 /* | |
4879 * Set resync_completed to last resync type and then | |
4880 * clear resync_type to indicate no resync in progress | |
4881 */ | |
4882 un->un_resync_completed = un->un_rs_type; | |
4883 SET_RS_TYPE_NONE(un->un_rs_type); | |
4884 /* | |
4885 * If the resync completed ok reset the submirror | |
4886 * state to SMS_RUNNING else reset it to SMS_ATTACHED | |
4887 */ | |
4888 state = (broke_out ? | |
4889 SMS_ATTACHED : SMS_RUNNING); | |
4890 mirror_set_sm_state(sm, smic, state, broke_out); | |
4891 un->c.un_status &= ~MD_UN_WAR; | |
4892 mirror_commit(un, SMI2BIT(smi), 0); | |
4893 md_ioctl_writerexit(lockp); | |
4894 break; | |
4895 case MD_RS_COMPONENT: | |
4896 un = md_ioctl_writerlock(lockp, ui); | |
4897 smi = RS_SMI(p->rs_type); | |
4898 ci = RS_CI(p->rs_type); | |
4899 sm = &un->un_sm[smi]; | |
4900 smic = &un->un_smic[smi]; | |
4901 shared = (md_m_shared_t *) | |
4902 (*(smic->sm_shared_by_indx)) | |
4903 (sm->sm_dev, sm, ci); | |
4904 un->c.un_status &= ~MD_UN_WAR; | |
4905 /* Clear RESYNC target */ | |
4906 un->un_sm[smi].sm_flags &= ~MD_SM_RESYNC_TARGET; | |
4907 /* | |
4908 * Set resync_completed to last resync type and then | |
4909 * clear resync_type to indicate no resync in progress | |
4910 */ | |
4911 un->un_resync_completed = un->un_rs_type; | |
4912 SET_RS_TYPE_NONE(un->un_rs_type); | |
4913 | |
4914 /* | |
4915 * If the resync completed ok, set the component state | |
4916 * to CS_OKAY. | |
4917 */ | |
4918 if (broke_out) | |
4919 shared->ms_flags |= MDM_S_RS_TRIED; | |
4920 else { | |
4921 /* | |
4922 * As we don't transmit the changes, | |
4923 * no need to drop the lock. | |
4924 */ | |
4925 set_sm_comp_state(un, smi, ci, CS_OKAY, 0, | |
4926 MD_STATE_NO_XMIT, (IOLOCK *)NULL); | |
4927 } | |
4928 md_ioctl_writerexit(lockp); | |
4929 default: | |
4930 break; | |
4931 } | |
4932 /* | |
4933 * If the purpose of this PHASE_DONE message is just to | |
4934 * indicate to all other nodes that the optimized resync | |
4935 * required (OPT_NOT_DONE) flag is to be cleared, there is | |
4936 * no need to generate a notify event as there has not | |
4937 * actually been a resync. | |
4938 */ | |
4939 if (!(p->rs_flags & MD_MN_RS_CLEAR_OPT_NOT_DONE)) { | |
4940 if (broke_out) { | |
4941 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_RESYNC_FAILED, | |
4942 SVM_TAG_METADEVICE, MD_UN2SET(un), | |
4943 MD_SID(un)); | |
4944 } else { | |
4945 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_RESYNC_DONE, | |
4946 SVM_TAG_METADEVICE, MD_UN2SET(un), | |
4947 MD_SID(un)); | |
4948 } | |
4949 } | |
4950 break; | |
4951 | |
4952 default: | |
4953 #ifdef DEBUG | |
4954 cmn_err(CE_PANIC, "mirror_resync_message: Unknown message type" | |
4955 " %x\n", p->msg_type); | |
4956 #endif | |
4957 return (EINVAL); | |
4958 } | |
4959 return (0); | |
4960 } | |
4961 | |
4962 /* Return a -1 if snarf of optimized record failed and set should be released */ | |
4963 static int | |
4964 mirror_snarf(md_snarfcmd_t cmd, set_t setno) | |
4965 { | |
4966 mddb_recid_t recid; | |
4967 int gotsomething; | |
4968 int all_mirrors_gotten; | |
4969 mm_unit_t *un; | |
4970 mddb_type_t typ1; | |
4971 mddb_de_ic_t *dep; | |
4972 mddb_rb32_t *rbp; | |
4973 size_t newreqsize; | |
4974 mm_unit_t *big_un; | |
4975 mm_unit32_od_t *small_un; | |
4976 int retval; | |
4977 mdi_unit_t *ui; | |
4978 | |
4979 if (cmd == MD_SNARF_CLEANUP) { | |
4980 if (md_get_setstatus(setno) & MD_SET_STALE) | |
4981 return (0); | |
4982 | |
4983 recid = mddb_makerecid(setno, 0); | |
4984 typ1 = (mddb_type_t)md_getshared_key(setno, | |
4985 mirror_md_ops.md_driver.md_drivername); | |
4986 while ((recid = mddb_getnextrec(recid, typ1, MIRROR_REC)) > 0) { | |
4987 if (mddb_getrecprivate(recid) & MD_PRV_CLEANUP) { | |
4988 un = (mm_unit_t *)mddb_getrecaddr(recid); | |
4989 mirror_cleanup(un); | |
4990 recid = mddb_makerecid(setno, 0); | |
4991 } | |
4992 } | |
4993 return (0); | |
4994 } | |
4995 | |
4996 all_mirrors_gotten = 1; | |
4997 gotsomething = 0; | |
4998 | |
4999 recid = mddb_makerecid(setno, 0); | |
5000 typ1 = (mddb_type_t)md_getshared_key(setno, | |
5001 mirror_md_ops.md_driver.md_drivername); | |
5002 | |
5003 while ((recid = mddb_getnextrec(recid, typ1, MIRROR_REC)) > 0) { | |
5004 if (mddb_getrecprivate(recid) & MD_PRV_GOTIT) | |
5005 continue; | |
5006 | |
5007 dep = mddb_getrecdep(recid); | |
5008 dep->de_flags = MDDB_F_MIRROR; | |
5009 rbp = dep->de_rb; | |
5010 | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5011 switch (rbp->rb_revision) { |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5012 case MDDB_REV_RB: |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5013 case MDDB_REV_RBFN: |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5014 if ((rbp->rb_private & MD_PRV_CONVD) == 0) { |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5015 /* |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5016 * This means, we have an old and small |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5017 * record and this record hasn't already |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5018 * been converted. Before we create an |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5019 * incore metadevice from this we have to |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5020 * convert it to a big record. |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5021 */ |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5022 small_un = |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5023 (mm_unit32_od_t *)mddb_getrecaddr(recid); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5024 newreqsize = sizeof (mm_unit_t); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5025 big_un = (mm_unit_t *)kmem_zalloc(newreqsize, |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5026 KM_SLEEP); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5027 mirror_convert((caddr_t)small_un, |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5028 (caddr_t)big_un, SMALL_2_BIG); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5029 kmem_free(small_un, dep->de_reqsize); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5030 |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5031 /* |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5032 * Update userdata and incore userdata |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5033 * incores are at the end of un |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5034 */ |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5035 dep->de_rb_userdata_ic = big_un; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5036 dep->de_rb_userdata = big_un; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5037 dep->de_icreqsize = newreqsize; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5038 un = big_un; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5039 rbp->rb_private |= MD_PRV_CONVD; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5040 } else { |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5041 /* |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5042 * Unit already converted, just get the |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5043 * record address. |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5044 */ |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5045 un = (mm_unit_t *)mddb_getrecaddr_resize(recid, |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5046 sizeof (*un), 0); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5047 } |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5048 un->c.un_revision &= ~MD_64BIT_META_DEV; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5049 break; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5050 case MDDB_REV_RB64: |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5051 case MDDB_REV_RB64FN: |
0 | 5052 /* Big device */ |
5053 un = (mm_unit_t *)mddb_getrecaddr_resize(recid, | |
5054 sizeof (*un), 0); | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5055 un->c.un_revision |= MD_64BIT_META_DEV; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5056 un->c.un_flag |= MD_EFILABEL; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5057 break; |
0 | 5058 } |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5059 NOTE_FN(rbp->rb_revision, un->c.un_revision); |
0 | 5060 |
5061 /* | |
5062 * Create minor device node for snarfed entry. | |
5063 */ | |
5064 (void) md_create_minor_node(setno, MD_SID(un)); | |
5065 | |
5066 if (MD_UNIT(MD_SID(un)) != NULL) { | |
5067 mddb_setrecprivate(recid, MD_PRV_PENDDEL); | |
5068 continue; | |
5069 } | |
5070 all_mirrors_gotten = 0; | |
5071 retval = mirror_build_incore(un, 1); | |
5072 if (retval == 0) { | |
5073 mddb_setrecprivate(recid, MD_PRV_GOTIT); | |
5074 md_create_unit_incore(MD_SID(un), &mirror_md_ops, 0); | |
5075 resync_start_timeout(setno); | |
5076 gotsomething = 1; | |
5077 } else if (retval == -1) { | |
5078 return (-1); | |
5079 } | |
5080 /* | |
5081 * Set flag to indicate that the mirror has not yet | |
5082 * been through a reconfig. This flag is used for MN sets | |
5083 * when determining whether to update the mirror state from | |
5084 * the Master node. | |
5085 */ | |
5086 if (MD_MNSET_SETNO(setno)) { | |
5087 ui = MDI_UNIT(MD_SID(un)); | |
5088 ui->ui_tstate |= MD_RESYNC_NOT_DONE; | |
5089 } | |
5090 } | |
5091 | |
5092 if (!all_mirrors_gotten) | |
5093 return (gotsomething); | |
5094 | |
5095 recid = mddb_makerecid(setno, 0); | |
5096 while ((recid = mddb_getnextrec(recid, typ1, RESYNC_REC)) > 0) | |
5097 if (!(mddb_getrecprivate(recid) & MD_PRV_GOTIT)) | |
5098 mddb_setrecprivate(recid, MD_PRV_PENDDEL); | |
5099 | |
5100 return (0); | |
5101 } | |
5102 | |
5103 static int | |
5104 mirror_halt(md_haltcmd_t cmd, set_t setno) | |
5105 { | |
5106 unit_t i; | |
5107 mdi_unit_t *ui; | |
5108 minor_t mnum; | |
5109 int reset_mirror_flag = 0; | |
5110 | |
5111 if (cmd == MD_HALT_CLOSE) | |
5112 return (0); | |
5113 | |
5114 if (cmd == MD_HALT_OPEN) | |
5115 return (0); | |
5116 | |
5117 if (cmd == MD_HALT_UNLOAD) | |
5118 return (0); | |
5119 | |
5120 if (cmd == MD_HALT_CHECK) { | |
5121 for (i = 0; i < md_nunits; i++) { | |
5122 mnum = MD_MKMIN(setno, i); | |
5123 if ((ui = MDI_UNIT(mnum)) == NULL) | |
5124 continue; | |
5125 if (ui->ui_opsindex != mirror_md_ops.md_selfindex) | |
5126 continue; | |
5127 if (md_unit_isopen(ui)) | |
5128 return (1); | |
5129 } | |
5130 return (0); | |
5131 } | |
5132 | |
5133 if (cmd != MD_HALT_DOIT) | |
5134 return (1); | |
5135 | |
5136 for (i = 0; i < md_nunits; i++) { | |
5137 mnum = MD_MKMIN(setno, i); | |
5138 if ((ui = MDI_UNIT(mnum)) == NULL) | |
5139 continue; | |
5140 if (ui->ui_opsindex != mirror_md_ops.md_selfindex) | |
5141 continue; | |
5142 reset_mirror((mm_unit_t *)MD_UNIT(mnum), mnum, 0); | |
5143 | |
5144 /* Set a flag if there is at least one mirror metadevice. */ | |
5145 reset_mirror_flag = 1; | |
5146 } | |
5147 | |
5148 /* | |
5149 * Only wait for the global dr_timeout to finish | |
5150 * - if there are mirror metadevices in this diskset or | |
5151 * - if this is the local set since an unload of the md_mirror | |
5152 * driver could follow a successful mirror halt in the local set. | |
5153 */ | |
5154 if ((reset_mirror_flag != 0) || (setno == MD_LOCAL_SET)) { | |
5155 while ((mirror_md_ops.md_head == NULL) && | |
5156 (mirror_timeout.dr_timeout_id != 0)) | |
5157 delay(md_hz); | |
5158 } | |
5159 | |
5160 return (0); | |
5161 } | |
5162 | |
5163 /*ARGSUSED3*/ | |
5164 static int | |
5165 mirror_open(dev_t *dev, int flag, int otyp, cred_t *cred_p, int md_oflags) | |
5166 { | |
5167 IOLOCK lock; | |
46
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5168 minor_t mnum = getminor(*dev); |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5169 set_t setno; |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5170 |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5171 /* |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5172 * When doing an open of a multi owner metadevice, check to see if this |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5173 * node is a starting node and if a reconfig cycle is underway. |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5174 * If so, the system isn't sufficiently set up enough to handle the |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5175 * open (which involves I/O during sp_validate), so fail with ENXIO. |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5176 */ |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5177 setno = MD_MIN2SET(mnum); |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5178 if ((md_set[setno].s_status & (MD_SET_MNSET | MD_SET_MN_START_RC)) == |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5179 (MD_SET_MNSET | MD_SET_MN_START_RC)) { |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5180 return (ENXIO); |
042bf15ebd92
6274840 Cluster node(s) panic when I/O is starting on nodes during reconfig.
skamm
parents:
0
diff
changeset
|
5181 } |
0 | 5182 |
5183 if (md_oflags & MD_OFLG_FROMIOCTL) { | |
5184 /* | |
5185 * This indicates that the caller is an ioctl service routine. | |
5186 * In this case we initialise our stack-based IOLOCK and pass | |
5187 * this into the internal open routine. This allows multi-owner | |
5188 * metadevices to avoid deadlocking if an error is encountered | |
5189 * during the open() attempt. The failure case is: | |
5190 * s-p -> mirror -> s-p (with error). Attempting to metaclear | |
5191 * this configuration would deadlock as the mirror code has to | |
5192 * send a state-update to the other nodes when it detects the | |
5193 * failure of the underlying submirror with an errored soft-part | |
5194 * on it. As there is a class1 message in progress (metaclear) | |
5195 * set_sm_comp_state() cannot send another class1 message; | |
5196 * instead we do not send a state_update message as the | |
5197 * metaclear is distributed and the failed submirror will be | |
5198 * cleared from the configuration by the metaclear. | |
5199 */ | |
5200 IOLOCK_INIT(&lock); | |
5201 return (mirror_internal_open(getminor(*dev), flag, otyp, | |
5202 md_oflags, &lock)); | |
5203 } else { | |
5204 return (mirror_internal_open(getminor(*dev), flag, otyp, | |
5205 md_oflags, (IOLOCK *)NULL)); | |
5206 } | |
5207 } | |
5208 | |
5209 | |
5210 /*ARGSUSED1*/ | |
5211 static int | |
5212 mirror_close(dev_t dev, int flag, int otyp, cred_t *cred_p, int md_cflags) | |
5213 { | |
5214 return (mirror_internal_close(getminor(dev), otyp, md_cflags, | |
5215 (IOLOCK *)NULL)); | |
5216 } | |
5217 | |
5218 | |
5219 /* | |
5220 * This routine dumps memory to the disk. It assumes that the memory has | |
5221 * already been mapped into mainbus space. It is called at disk interrupt | |
5222 * priority when the system is in trouble. | |
5223 * | |
5224 */ | |
5225 static int | |
5226 mirror_dump(dev_t dev, caddr_t addr, daddr_t blkno, int nblk) | |
5227 { | |
5228 mm_unit_t *un; | |
5229 dev_t mapdev; | |
5230 int result; | |
5231 int smi; | |
5232 int any_succeed = 0; | |
5233 int save_result = 0; | |
5234 | |
5235 /* | |
5236 * Don't need to grab the unit lock. | |
5237 * Cause nothing else is suppose to be happenning. | |
5238 * Also dump is not suppose to sleep. | |
5239 */ | |
5240 un = (mm_unit_t *)MD_UNIT(getminor(dev)); | |
5241 | |
5242 if ((diskaddr_t)blkno >= un->c.un_total_blocks) | |
5243 return (EINVAL); | |
5244 | |
5245 if ((diskaddr_t)blkno + nblk > un->c.un_total_blocks) | |
5246 return (EINVAL); | |
5247 | |
5248 for (smi = 0; smi < NMIRROR; smi++) { | |
5249 if (!SUBMIRROR_IS_WRITEABLE(un, smi)) | |
5250 continue; | |
5251 mapdev = md_dev64_to_dev(un->un_sm[smi].sm_dev); | |
5252 result = bdev_dump(mapdev, addr, blkno, nblk); | |
5253 if (result) | |
5254 save_result = result; | |
5255 | |
5256 if (result == 0) | |
5257 any_succeed++; | |
5258 } | |
5259 | |
5260 if (any_succeed) | |
5261 return (0); | |
5262 | |
5263 return (save_result); | |
5264 } | |
5265 | |
5266 /* | |
5267 * NAME: mirror_probe_dev | |
5268 * | |
5269 * DESCRITPION: force opens every component of a mirror. | |
5270 * | |
5271 * On entry the unit writerlock is held | |
5272 */ | |
5273 static int | |
5274 mirror_probe_dev(mdi_unit_t *ui, minor_t mnum) | |
5275 { | |
5276 int i; | |
5277 int smi; | |
5278 int ci; | |
5279 mm_unit_t *un; | |
5280 int md_devopen = 0; | |
5281 set_t setno; | |
5282 int sm_cnt; | |
5283 int sm_unavail_cnt; | |
5284 | |
5285 if (md_unit_isopen(ui)) | |
5286 md_devopen++; | |
5287 | |
5288 un = MD_UNIT(mnum); | |
5289 setno = MD_UN2SET(un); | |
5290 | |
5291 sm_cnt = 0; | |
5292 sm_unavail_cnt = 0; | |
5293 for (i = 0; i < NMIRROR; i++) { | |
5294 md_dev64_t tmpdev; | |
5295 mdi_unit_t *sm_ui; | |
5296 | |
5297 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) { | |
5298 continue; | |
5299 } | |
5300 | |
5301 sm_cnt++; | |
5302 tmpdev = un->un_sm[i].sm_dev; | |
5303 (void) md_layered_open(mnum, &tmpdev, | |
5304 MD_OFLG_CONT_ERRS | MD_OFLG_PROBEDEV); | |
5305 un->un_sm[i].sm_dev = tmpdev; | |
5306 | |
5307 sm_ui = MDI_UNIT(getminor(md_dev64_to_dev(tmpdev))); | |
5308 | |
5309 /* | |
5310 * Logic similar to that in mirror_open_all_devs. We set or | |
5311 * clear the submirror Unavailable bit. | |
5312 */ | |
5313 (void) md_unit_writerlock(sm_ui); | |
5314 if (submirror_unavailable(un, i, 1)) { | |
5315 sm_ui->ui_tstate |= MD_INACCESSIBLE; | |
5316 sm_unavail_cnt++; | |
5317 } else { | |
5318 sm_ui->ui_tstate &= ~MD_INACCESSIBLE; | |
5319 } | |
5320 md_unit_writerexit(sm_ui); | |
5321 } | |
5322 | |
5323 /* | |
5324 * If all of the submirrors are unavailable, the mirror is also | |
5325 * unavailable. | |
5326 */ | |
5327 if (sm_cnt == sm_unavail_cnt) { | |
5328 ui->ui_tstate |= MD_INACCESSIBLE; | |
5329 } else { | |
5330 ui->ui_tstate &= ~MD_INACCESSIBLE; | |
5331 } | |
5332 | |
5333 /* | |
5334 * Start checking from probe failures. If failures occur we | |
5335 * set the appropriate erred state only if the metadevice is in | |
5336 * use. This is specifically to prevent unnecessary resyncs. | |
5337 * For instance if the disks were accidentally disconnected when | |
5338 * the system booted up then until the metadevice is accessed | |
5339 * (like file system mount) the user can shutdown, recable and | |
5340 * reboot w/o incurring a potentially huge resync. | |
5341 */ | |
5342 | |
5343 smi = 0; | |
5344 ci = 0; | |
5345 while (mirror_geterror(un, &smi, &ci, 1, 1) != 0) { | |
5346 | |
5347 if (mirror_other_sources(un, smi, ci, 0) == 1) { | |
5348 /* | |
5349 * Note that for a MN set, there is no need to call | |
5350 * SE_NOTIFY as that is done when processing the | |
5351 * state change | |
5352 */ | |
5353 if (md_devopen) { | |
5354 /* | |
5355 * Never called from ioctl context, | |
5356 * so (IOLOCK *)NULL | |
5357 */ | |
5358 set_sm_comp_state(un, smi, ci, CS_LAST_ERRED, | |
5359 0, MD_STATE_XMIT, (IOLOCK *)NULL); | |
5360 if (!MD_MNSET_SETNO(setno)) { | |
5361 SE_NOTIFY(EC_SVM_STATE, | |
5362 ESC_SVM_LASTERRED, | |
5363 SVM_TAG_METADEVICE, setno, | |
5364 MD_SID(un)); | |
5365 } | |
5366 continue; | |
5367 } else { | |
5368 (void) mirror_close_all_devs(un, | |
5369 MD_OFLG_PROBEDEV); | |
5370 if (!MD_MNSET_SETNO(setno)) { | |
5371 SE_NOTIFY(EC_SVM_STATE, | |
5372 ESC_SVM_OPEN_FAIL, | |
5373 SVM_TAG_METADEVICE, setno, | |
5374 MD_SID(un)); | |
5375 } | |
5376 mirror_openfail_console_info(un, smi, ci); | |
5377 return (ENXIO); | |
5378 } | |
5379 } | |
5380 | |
5381 /* | |
5382 * Note that for a MN set, there is no need to call | |
5383 * SE_NOTIFY as that is done when processing the | |
5384 * state change | |
5385 */ | |
5386 if (md_devopen) { | |
5387 /* Never called from ioctl context, so (IOLOCK *)NULL */ | |
5388 set_sm_comp_state(un, smi, ci, CS_ERRED, 0, | |
5389 MD_STATE_XMIT, (IOLOCK *)NULL); | |
5390 if (!MD_MNSET_SETNO(setno)) { | |
5391 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_ERRED, | |
5392 SVM_TAG_METADEVICE, setno, | |
5393 MD_SID(un)); | |
5394 } | |
5395 } | |
5396 mirror_openfail_console_info(un, smi, ci); | |
5397 ci++; | |
5398 } | |
5399 | |
5400 if (MD_MNSET_SETNO(setno)) { | |
5401 send_poke_hotspares(setno); | |
5402 } else { | |
5403 (void) poke_hotspares(); | |
5404 } | |
5405 (void) mirror_close_all_devs(un, MD_OFLG_PROBEDEV); | |
5406 | |
5407 return (0); | |
5408 } | |
5409 | |
5410 | |
5411 static int | |
5412 mirror_imp_set( | |
5413 set_t setno | |
5414 ) | |
5415 { | |
5416 | |
5417 mddb_recid_t recid; | |
5418 int gotsomething, i; | |
5419 mddb_type_t typ1; | |
5420 mddb_de_ic_t *dep; | |
5421 mddb_rb32_t *rbp; | |
5422 mm_unit32_od_t *un32; | |
5423 mm_unit_t *un64; | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5424 md_dev64_t self_devt; |
0 | 5425 minor_t *self_id; /* minor needs to be updated */ |
5426 md_parent_t *parent_id; /* parent needs to be updated */ | |
5427 mddb_recid_t *record_id; /* record id needs to be updated */ | |
5428 mddb_recid_t *optrec_id; | |
5429 md_dev64_t tmpdev; | |
5430 | |
5431 | |
5432 gotsomething = 0; | |
5433 | |
5434 typ1 = (mddb_type_t)md_getshared_key(setno, | |
5435 mirror_md_ops.md_driver.md_drivername); | |
5436 recid = mddb_makerecid(setno, 0); | |
5437 | |
5438 while ((recid = mddb_getnextrec(recid, typ1, MIRROR_REC)) > 0) { | |
5439 if (mddb_getrecprivate(recid) & MD_PRV_GOTIT) | |
5440 continue; | |
5441 | |
5442 dep = mddb_getrecdep(recid); | |
5443 rbp = dep->de_rb; | |
5444 | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5445 switch (rbp->rb_revision) { |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5446 case MDDB_REV_RB: |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5447 case MDDB_REV_RBFN: |
0 | 5448 /* |
5449 * Small device | |
5450 */ | |
5451 un32 = (mm_unit32_od_t *)mddb_getrecaddr(recid); | |
5452 self_id = &(un32->c.un_self_id); | |
5453 parent_id = &(un32->c.un_parent); | |
5454 record_id = &(un32->c.un_record_id); | |
5455 optrec_id = &(un32->un_rr_dirty_recid); | |
5456 | |
5457 for (i = 0; i < un32->un_nsm; i++) { | |
5458 tmpdev = md_expldev(un32->un_sm[i].sm_dev); | |
5459 un32->un_sm[i].sm_dev = md_cmpldev | |
5460 (md_makedevice(md_major, MD_MKMIN(setno, | |
5461 MD_MIN2UNIT(md_getminor(tmpdev))))); | |
5462 | |
5463 if (!md_update_minor(setno, mddb_getsidenum | |
5464 (setno), un32->un_sm[i].sm_key)) | |
5465 goto out; | |
5466 } | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5467 break; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5468 case MDDB_REV_RB64: |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5469 case MDDB_REV_RB64FN: |
0 | 5470 un64 = (mm_unit_t *)mddb_getrecaddr(recid); |
5471 self_id = &(un64->c.un_self_id); | |
5472 parent_id = &(un64->c.un_parent); | |
5473 record_id = &(un64->c.un_record_id); | |
5474 optrec_id = &(un64->un_rr_dirty_recid); | |
5475 | |
5476 for (i = 0; i < un64->un_nsm; i++) { | |
5477 tmpdev = un64->un_sm[i].sm_dev; | |
5478 un64->un_sm[i].sm_dev = md_makedevice | |
5479 (md_major, MD_MKMIN(setno, MD_MIN2UNIT | |
5480 (md_getminor(tmpdev)))); | |
5481 | |
5482 if (!md_update_minor(setno, mddb_getsidenum | |
5483 (setno), un64->un_sm[i].sm_key)) | |
5484 goto out; | |
5485 } | |
1623
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5486 break; |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5487 } |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5488 |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5489 /* |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5490 * If this is a top level and a friendly name metadevice, |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5491 * update its minor in the namespace. |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5492 */ |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5493 if ((*parent_id == MD_NO_PARENT) && |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5494 ((rbp->rb_revision == MDDB_REV_RBFN) || |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5495 (rbp->rb_revision == MDDB_REV_RB64FN))) { |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5496 |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5497 self_devt = md_makedevice(md_major, *self_id); |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5498 if (!md_update_top_device_minor(setno, |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5499 mddb_getsidenum(setno), self_devt)) |
7bac4a816ebe
PSARC/2005/153 Bunnahabhain: Descriptive Name Support in SVM
tw21770
parents:
1366
diff
changeset
|
5500 goto out; |
0 | 5501 } |
5502 | |
5503 /* | |
5504 * Update unit with the imported setno | |
5505 * | |
5506 */ | |
5507 mddb_setrecprivate(recid, MD_PRV_GOTIT); | |
5508 | |
5509 *self_id = MD_MKMIN(setno, MD_MIN2UNIT(*self_id)); | |
5510 if (*parent_id != MD_NO_PARENT) | |
5511 *parent_id = MD_MKMIN(setno, MD_MIN2UNIT(*parent_id)); | |
5512 *record_id = MAKERECID(setno, DBID(*record_id)); | |
5513 *optrec_id = MAKERECID(setno, DBID(*optrec_id)); | |
5514 | |
5515 gotsomething = 1; | |
5516 } | |
5517 | |
5518 out: | |
5519 return (gotsomething); | |
5520 } | |
5521 | |
5522 /* | |
5523 * NAME: mirror_check_offline | |
5524 * | |
5525 * DESCRIPTION: return offline_status = 1 if any submirrors are offline | |
5526 * | |
5527 * Called from ioctl, so access to MD_UN_OFFLINE_SM in un_status is | |
5528 * protected by the global ioctl lock as it is only set by the MD_IOCOFFLINE | |
5529 * ioctl. | |
5530 */ | |
5531 int | |
5532 mirror_check_offline(md_dev64_t dev, int *offline_status) | |
5533 { | |
5534 mm_unit_t *un; | |
5535 md_error_t mde = mdnullerror; | |
5536 | |
5537 if ((un = mirror_getun(getminor(dev), &mde, NO_LOCK, NULL)) == NULL) | |
5538 return (EINVAL); | |
5539 *offline_status = 0; | |
5540 if (un->c.un_status & MD_UN_OFFLINE_SM) | |
5541 *offline_status = 1; | |
5542 return (0); | |
5543 } | |
5544 | |
5545 /* | |
5546 * NAME: mirror_inc_abr_count | |
5547 * | |
5548 * DESCRIPTION: increment the count of layered soft parts with ABR set | |
5549 * | |
5550 * Called from ioctl, so access to un_abr_count is protected by the global | |
5551 * ioctl lock. It is only referenced in the MD_IOCOFFLINE ioctl. | |
5552 */ | |
5553 int | |
5554 mirror_inc_abr_count(md_dev64_t dev) | |
5555 { | |
5556 mm_unit_t *un; | |
5557 md_error_t mde = mdnullerror; | |
5558 | |
5559 if ((un = mirror_getun(getminor(dev), &mde, NO_LOCK, NULL)) == NULL) | |
5560 return (EINVAL); | |
5561 un->un_abr_count++; | |
5562 return (0); | |
5563 } | |
5564 | |
5565 /* | |
5566 * NAME: mirror_dec_abr_count | |
5567 * | |
5568 * DESCRIPTION: decrement the count of layered soft parts with ABR set | |
5569 * | |
5570 * Called from ioctl, so access to un_abr_count is protected by the global | |
5571 * ioctl lock. It is only referenced in the MD_IOCOFFLINE ioctl. | |
5572 */ | |
5573 int | |
5574 mirror_dec_abr_count(md_dev64_t dev) | |
5575 { | |
5576 mm_unit_t *un; | |
5577 md_error_t mde = mdnullerror; | |
5578 | |
5579 if ((un = mirror_getun(getminor(dev), &mde, NO_LOCK, NULL)) == NULL) | |
5580 return (EINVAL); | |
5581 un->un_abr_count--; | |
5582 return (0); | |
5583 } | |
5584 | |
5585 static md_named_services_t mirror_named_services[] = { | |
5586 {(intptr_t (*)()) poke_hotspares, "poke hotspares" }, | |
5587 {(intptr_t (*)()) mirror_rename_listkids, MDRNM_LIST_URKIDS }, | |
5588 {mirror_rename_check, MDRNM_CHECK }, | |
5589 {(intptr_t (*)()) mirror_renexch_update_kids, MDRNM_UPDATE_KIDS }, | |
5590 {(intptr_t (*)()) mirror_exchange_parent_update_to, | |
5591 MDRNM_PARENT_UPDATE_TO}, | |
5592 {(intptr_t (*)()) mirror_exchange_self_update_from_down, | |
5593 MDRNM_SELF_UPDATE_FROM_DOWN }, | |
5594 {(intptr_t (*)())mirror_probe_dev, "probe open test" }, | |
5595 {(intptr_t (*)())mirror_check_offline, MD_CHECK_OFFLINE }, | |
5596 {(intptr_t (*)())mirror_inc_abr_count, MD_INC_ABR_COUNT }, | |
5597 {(intptr_t (*)())mirror_dec_abr_count, MD_DEC_ABR_COUNT }, | |
5598 { NULL, 0 } | |
5599 }; | |
5600 | |
5601 md_ops_t mirror_md_ops = { | |
5602 mirror_open, /* open */ | |
5603 mirror_close, /* close */ | |
5604 md_mirror_strategy, /* strategy */ | |
5605 NULL, /* print */ | |
5606 mirror_dump, /* dump */ | |
5607 NULL, /* read */ | |
5608 NULL, /* write */ | |
5609 md_mirror_ioctl, /* mirror_ioctl, */ | |
5610 mirror_snarf, /* mirror_snarf */ | |
5611 mirror_halt, /* mirror_halt */ | |
5612 NULL, /* aread */ | |
5613 NULL, /* awrite */ | |
5614 mirror_imp_set, /* import set */ | |
5615 mirror_named_services | |
5616 }; | |
5617 | |
5618 /* module specific initilization */ | |
5619 static void | |
5620 init_init() | |
5621 { | |
5622 md_mirror_mcs_buf_off = sizeof (md_mcs_t) - sizeof (buf_t); | |
5623 | |
5624 /* Initialize the parent and child save memory pools */ | |
5625 mirror_parent_cache = kmem_cache_create("md_mirror_parent", | |
5626 sizeof (md_mps_t), 0, mirror_parent_constructor, | |
5627 mirror_parent_destructor, mirror_run_queue, NULL, NULL, | |
5628 0); | |
5629 | |
5630 mirror_child_cache = kmem_cache_create("md_mirror_child", | |
5631 sizeof (md_mcs_t) - sizeof (buf_t) + biosize(), 0, | |
5632 mirror_child_constructor, mirror_child_destructor, | |
5633 mirror_run_queue, NULL, NULL, 0); | |
5634 | |
5635 /* | |
5636 * Insure wowbuf_size is a multiple of DEV_BSIZE, | |
5637 * then initialize wowbuf memory pool. | |
5638 */ | |
5639 md_wowbuf_size = roundup(md_wowbuf_size, DEV_BSIZE); | |
5640 if (md_wowbuf_size <= 0) | |
5641 md_wowbuf_size = 2 * DEV_BSIZE; | |
5642 if (md_wowbuf_size > (32 * DEV_BSIZE)) | |
5643 md_wowbuf_size = (32 * DEV_BSIZE); | |
5644 | |
5645 md_wowblk_size = md_wowbuf_size + sizeof (wowhdr_t); | |
5646 mirror_wowblk_cache = kmem_cache_create("md_mirror_wow", | |
5647 md_wowblk_size, 0, NULL, NULL, NULL, NULL, NULL, 0); | |
5648 | |
5649 mutex_init(&mirror_timeout.dr_mx, NULL, MUTEX_DEFAULT, NULL); | |
5650 mutex_init(&hotspare_request.dr_mx, NULL, MUTEX_DEFAULT, NULL); | |
5651 | |
5652 mutex_init(&non_ff_drv_mutex, NULL, MUTEX_DEFAULT, NULL); | |
5653 } | |
5654 | |
5655 /* module specific uninitilization (undo init_init()) */ | |
5656 static void | |
5657 fini_uninit() | |
5658 { | |
5659 kmem_cache_destroy(mirror_parent_cache); | |
5660 kmem_cache_destroy(mirror_child_cache); | |
5661 kmem_cache_destroy(mirror_wowblk_cache); | |
5662 mirror_parent_cache = mirror_child_cache = | |
5663 mirror_wowblk_cache = NULL; | |
5664 | |
5665 mutex_destroy(&mirror_timeout.dr_mx); | |
5666 mutex_destroy(&hotspare_request.dr_mx); | |
5667 mutex_destroy(&non_ff_drv_mutex); | |
5668 } | |
5669 | |
5670 /* define the module linkage */ | |
5671 MD_PLUGIN_MISC_MODULE("mirrors module %I%", init_init(), fini_uninit()) |