0
|
1 /*
|
|
2 * CDDL HEADER START
|
|
3 *
|
|
4 * The contents of this file are subject to the terms of the
|
|
5 * Common Development and Distribution License, Version 1.0 only
|
|
6 * (the "License"). You may not use this file except in compliance
|
|
7 * with the License.
|
|
8 *
|
|
9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
|
|
10 * or http://www.opensolaris.org/os/licensing.
|
|
11 * See the License for the specific language governing permissions
|
|
12 * and limitations under the License.
|
|
13 *
|
|
14 * When distributing Covered Code, include this CDDL HEADER in each
|
|
15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
|
|
16 * If applicable, add the following below this CDDL HEADER, with the
|
|
17 * fields enclosed by brackets "[]" replaced with your own identifying
|
|
18 * information: Portions Copyright [yyyy] [name of copyright owner]
|
|
19 *
|
|
20 * CDDL HEADER END
|
|
21 */
|
|
22 /*
|
|
23 * Copyright 2005 Sun Microsystems, Inc. All rights reserved.
|
|
24 * Use is subject to license terms.
|
|
25 */
|
|
26
|
|
27 #pragma ident "%Z%%M% %I% %E% SMI"
|
|
28
|
|
29 #include <sys/param.h>
|
|
30 #include <sys/systm.h>
|
|
31 #include <sys/conf.h>
|
|
32 #include <sys/file.h>
|
|
33 #include <sys/user.h>
|
|
34 #include <sys/uio.h>
|
|
35 #include <sys/t_lock.h>
|
|
36 #include <sys/buf.h>
|
|
37 #include <sys/dkio.h>
|
|
38 #include <sys/vtoc.h>
|
|
39 #include <sys/kmem.h>
|
|
40 #include <vm/page.h>
|
|
41 #include <sys/cmn_err.h>
|
|
42 #include <sys/sysmacros.h>
|
|
43 #include <sys/types.h>
|
|
44 #include <sys/mkdev.h>
|
|
45 #include <sys/stat.h>
|
|
46 #include <sys/open.h>
|
|
47 #include <sys/modctl.h>
|
|
48 #include <sys/ddi.h>
|
|
49 #include <sys/sunddi.h>
|
|
50 #include <sys/debug.h>
|
|
51 #include <sys/dklabel.h>
|
|
52 #include <vm/hat.h>
|
|
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
|
|
64 static char _depends_on[] = "drv/md";
|
|
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
|
|
1975 if (!removing)
|
|
1976 return;
|
|
1977
|
|
1978 for (smi = 0; smi < NMIRROR; smi++) {
|
|
1979 if (!SMS_BY_INDEX_IS(un, smi, SMS_INUSE))
|
|
1980 continue;
|
|
1981 /* reallow soft partitioning of submirror and reset parent */
|
|
1982 su = MD_UNIT(md_getminor(un->un_sm[smi].sm_dev));
|
|
1983 MD_CAPAB(su) |= MD_CAN_SP;
|
|
1984 md_reset_parent(un->un_sm[smi].sm_dev);
|
|
1985 reset_comp_states(&un->un_sm[smi], &un->un_smic[smi]);
|
|
1986
|
|
1987 sv[nsv].setno = MD_MIN2SET(mnum);
|
|
1988 sv[nsv++].key = un->un_sm[smi].sm_key;
|
|
1989 bits |= SMI2BIT(smi);
|
|
1990 }
|
|
1991
|
|
1992 MD_STATUS(un) |= MD_UN_BEING_RESET;
|
|
1993 recid = un->un_rr_dirty_recid;
|
|
1994 vtoc_id = un->c.un_vtoc_id;
|
|
1995 selfid = MD_SID(un);
|
|
1996
|
|
1997 mirror_commit(un, bits, 0);
|
|
1998
|
|
1999 /* Destroy all mutexes and condvars before returning. */
|
|
2000 mutex_destroy(&un->un_suspend_wr_mx);
|
|
2001 cv_destroy(&un->un_suspend_wr_cv);
|
|
2002 mutex_destroy(&un->un_ovrlap_chn_mx);
|
|
2003 cv_destroy(&un->un_ovrlap_chn_cv);
|
|
2004 mutex_destroy(&un->un_owner_mx);
|
|
2005 mutex_destroy(&un->un_rs_thread_mx);
|
|
2006 cv_destroy(&un->un_rs_thread_cv);
|
|
2007 mutex_destroy(&un->un_rs_progress_mx);
|
|
2008 cv_destroy(&un->un_rs_progress_cv);
|
|
2009 mutex_destroy(&un->un_dmr_mx);
|
|
2010 cv_destroy(&un->un_dmr_cv);
|
|
2011 mddb_deleterec_wrapper(un->c.un_record_id);
|
|
2012 if (recid != 0)
|
|
2013 mddb_deleterec_wrapper(recid);
|
|
2014
|
|
2015 /* Remove the vtoc, if present */
|
|
2016 if (vtoc_id)
|
|
2017 mddb_deleterec_wrapper(vtoc_id);
|
|
2018
|
|
2019 md_rem_names(sv, nsv);
|
|
2020
|
|
2021 SE_NOTIFY(EC_SVM_CONFIG, ESC_SVM_DELETE, SVM_TAG_METADEVICE,
|
|
2022 MD_MIN2SET(selfid), selfid);
|
|
2023 }
|
|
2024
|
|
2025 int
|
|
2026 mirror_internal_open(
|
|
2027 minor_t mnum,
|
|
2028 int flag,
|
|
2029 int otyp,
|
|
2030 int md_oflags,
|
|
2031 IOLOCK *lockp /* can be NULL */
|
|
2032 )
|
|
2033 {
|
|
2034 mdi_unit_t *ui = MDI_UNIT(mnum);
|
|
2035 int err = 0;
|
|
2036
|
|
2037 tryagain:
|
|
2038 /* single thread */
|
|
2039 if (lockp) {
|
|
2040 /*
|
|
2041 * If ioctl lock is held, use openclose_enter
|
|
2042 * routine that will set the ioctl flag when
|
|
2043 * grabbing the readerlock.
|
|
2044 */
|
|
2045 (void) md_ioctl_openclose_enter(lockp, ui);
|
|
2046 } else {
|
|
2047 (void) md_unit_openclose_enter(ui);
|
|
2048 }
|
|
2049
|
|
2050 /*
|
|
2051 * The mirror_open_all_devs routine may end up sending a STATE_UPDATE
|
|
2052 * message in a MN diskset and this requires that the openclose
|
|
2053 * lock is dropped in order to send this message. So, another
|
|
2054 * flag (MD_UL_OPENINPROGRESS) is used to keep another thread from
|
|
2055 * attempting an open while this thread has an open in progress.
|
|
2056 * Call the *_lh version of the lock exit routines since the ui_mx
|
|
2057 * mutex must be held from checking for OPENINPROGRESS until
|
|
2058 * after the cv_wait call.
|
|
2059 */
|
|
2060 mutex_enter(&ui->ui_mx);
|
|
2061 if (ui->ui_lock & MD_UL_OPENINPROGRESS) {
|
|
2062 if (lockp) {
|
|
2063 (void) md_ioctl_openclose_exit_lh(lockp);
|
|
2064 } else {
|
|
2065 md_unit_openclose_exit_lh(ui);
|
|
2066 }
|
|
2067 cv_wait(&ui->ui_cv, &ui->ui_mx);
|
|
2068 mutex_exit(&ui->ui_mx);
|
|
2069 goto tryagain;
|
|
2070 }
|
|
2071
|
|
2072 ui->ui_lock |= MD_UL_OPENINPROGRESS;
|
|
2073 mutex_exit(&ui->ui_mx);
|
|
2074
|
|
2075 /* open devices, if necessary */
|
|
2076 if (! md_unit_isopen(ui) || (ui->ui_tstate & MD_INACCESSIBLE)) {
|
|
2077 if ((err = mirror_open_all_devs(mnum, md_oflags, lockp)) != 0)
|
|
2078 goto out;
|
|
2079 }
|
|
2080
|
|
2081 /* count open */
|
|
2082 if ((err = md_unit_incopen(mnum, flag, otyp)) != 0)
|
|
2083 goto out;
|
|
2084
|
|
2085 /* unlock, return success */
|
|
2086 out:
|
|
2087 mutex_enter(&ui->ui_mx);
|
|
2088 ui->ui_lock &= ~MD_UL_OPENINPROGRESS;
|
|
2089 mutex_exit(&ui->ui_mx);
|
|
2090
|
|
2091 if (lockp) {
|
|
2092 /*
|
|
2093 * If ioctl lock is held, use openclose_exit
|
|
2094 * routine that will clear the lockp reader flag.
|
|
2095 */
|
|
2096 (void) md_ioctl_openclose_exit(lockp);
|
|
2097 } else {
|
|
2098 md_unit_openclose_exit(ui);
|
|
2099 }
|
|
2100 return (err);
|
|
2101 }
|
|
2102
|
|
2103 int
|
|
2104 mirror_internal_close(
|
|
2105 minor_t mnum,
|
|
2106 int otyp,
|
|
2107 int md_cflags,
|
|
2108 IOLOCK *lockp /* can be NULL */
|
|
2109 )
|
|
2110 {
|
|
2111 mdi_unit_t *ui = MDI_UNIT(mnum);
|
|
2112 mm_unit_t *un;
|
|
2113 int err = 0;
|
|
2114
|
|
2115 /* single thread */
|
|
2116 if (lockp) {
|
|
2117 /*
|
|
2118 * If ioctl lock is held, use openclose_enter
|
|
2119 * routine that will set the ioctl flag when
|
|
2120 * grabbing the readerlock.
|
|
2121 */
|
|
2122 un = (mm_unit_t *)md_ioctl_openclose_enter(lockp, ui);
|
|
2123 } else {
|
|
2124 un = (mm_unit_t *)md_unit_openclose_enter(ui);
|
|
2125 }
|
|
2126
|
|
2127 /* count closed */
|
|
2128 if ((err = md_unit_decopen(mnum, otyp)) != 0)
|
|
2129 goto out;
|
|
2130
|
|
2131 /* close devices, if necessary */
|
|
2132 if (! md_unit_isopen(ui) || (md_cflags & MD_OFLG_PROBEDEV)) {
|
|
2133 /*
|
|
2134 * Clean up dirty bitmap for this unit. Do this
|
|
2135 * before closing the underlying devices to avoid
|
|
2136 * race conditions with reset_mirror() as a
|
|
2137 * result of a 'metaset -r' command running in
|
|
2138 * parallel. This might cause deallocation of
|
|
2139 * dirty region bitmaps; with underlying metadevices
|
|
2140 * in place this can't happen.
|
|
2141 * Don't do this if a MN set and ABR not set
|
|
2142 */
|
|
2143 if (new_resync && !(MD_STATUS(un) & MD_UN_KEEP_DIRTY)) {
|
|
2144 if (!MD_MNSET_SETNO(MD_UN2SET(un)) ||
|
|
2145 !(ui->ui_tstate & MD_ABR_CAP))
|
|
2146 mirror_process_unit_resync(un);
|
|
2147 }
|
|
2148 (void) mirror_close_all_devs(un, md_cflags);
|
|
2149
|
|
2150 /*
|
|
2151 * For a MN set with transient capabilities (eg ABR/DMR) set,
|
|
2152 * clear these capabilities on the last open in the cluster.
|
|
2153 * To do this we send a message to all nodes to see of the
|
|
2154 * device is open.
|
|
2155 */
|
|
2156 if (MD_MNSET_SETNO(MD_UN2SET(un)) &&
|
|
2157 (ui->ui_tstate & (MD_ABR_CAP|MD_DMR_CAP))) {
|
|
2158 if (lockp) {
|
|
2159 (void) md_ioctl_openclose_exit(lockp);
|
|
2160 } else {
|
|
2161 md_unit_openclose_exit(ui);
|
|
2162 }
|
|
2163
|
|
2164 /*
|
|
2165 * if we are in the context of an ioctl, drop the
|
|
2166 * ioctl lock.
|
|
2167 * Otherwise, no other locks should be held.
|
|
2168 */
|
|
2169 if (lockp) {
|
|
2170 IOLOCK_RETURN_RELEASE(0, lockp);
|
|
2171 }
|
|
2172
|
|
2173 mdmn_clear_all_capabilities(mnum);
|
|
2174
|
|
2175 /* if dropped the lock previously, regain it */
|
|
2176 if (lockp) {
|
|
2177 IOLOCK_RETURN_REACQUIRE(lockp);
|
|
2178 }
|
|
2179 return (0);
|
|
2180 }
|
|
2181 /* unlock and return success */
|
|
2182 }
|
|
2183 out:
|
|
2184 /* Call whether lockp is NULL or not. */
|
|
2185 if (lockp) {
|
|
2186 md_ioctl_openclose_exit(lockp);
|
|
2187 } else {
|
|
2188 md_unit_openclose_exit(ui);
|
|
2189 }
|
|
2190 return (err);
|
|
2191 }
|
|
2192
|
|
2193 /*
|
|
2194 * When a component has completed resyncing and is now ok, check if the
|
|
2195 * corresponding component in the other submirrors is in the Last Erred
|
|
2196 * state. If it is, we want to change that to the Erred state so we stop
|
|
2197 * using that component and start using this good component instead.
|
|
2198 *
|
|
2199 * This is called from set_sm_comp_state and recursively calls
|
|
2200 * set_sm_comp_state if it needs to change the Last Erred state.
|
|
2201 */
|
|
2202 static void
|
|
2203 reset_lasterred(mm_unit_t *un, int smi, mddb_recid_t *extras, uint_t flags,
|
|
2204 IOLOCK *lockp)
|
|
2205 {
|
|
2206 mm_submirror_t *sm;
|
|
2207 mm_submirror_ic_t *smic;
|
|
2208 int ci;
|
|
2209 int i;
|
|
2210 int compcnt;
|
|
2211 int changed = 0;
|
|
2212
|
|
2213 for (i = 0; i < NMIRROR; i++) {
|
|
2214 sm = &un->un_sm[i];
|
|
2215 smic = &un->un_smic[i];
|
|
2216
|
|
2217 if (!SMS_IS(sm, SMS_INUSE))
|
|
2218 continue;
|
|
2219
|
|
2220 /* ignore the submirror that we just made ok */
|
|
2221 if (i == smi)
|
|
2222 continue;
|
|
2223
|
|
2224 compcnt = (*(smic->sm_get_component_count)) (sm->sm_dev, un);
|
|
2225 for (ci = 0; ci < compcnt; ci++) {
|
|
2226 md_m_shared_t *shared;
|
|
2227
|
|
2228 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx))
|
|
2229 (sm->sm_dev, sm, ci);
|
|
2230
|
|
2231 if ((shared->ms_state & CS_LAST_ERRED) &&
|
|
2232 !mirror_other_sources(un, i, ci, 1)) {
|
|
2233
|
|
2234 set_sm_comp_state(un, i, ci, CS_ERRED, extras,
|
|
2235 flags, lockp);
|
|
2236 changed = 1;
|
|
2237 }
|
|
2238 }
|
|
2239 }
|
|
2240
|
|
2241 /* maybe there is a hotspare for this newly erred component */
|
|
2242 if (changed) {
|
|
2243 set_t setno;
|
|
2244
|
|
2245 setno = MD_UN2SET(un);
|
|
2246 if (MD_MNSET_SETNO(setno)) {
|
|
2247 send_poke_hotspares(setno);
|
|
2248 } else {
|
|
2249 (void) poke_hotspares();
|
|
2250 }
|
|
2251 }
|
|
2252 }
|
|
2253
|
|
2254 /*
|
|
2255 * set_sm_comp_state
|
|
2256 *
|
|
2257 * Set the state of a submirror component to the specified new state.
|
|
2258 * If the mirror is in a multi-node set, send messages to all nodes to
|
|
2259 * block all writes to the mirror and then update the state and release the
|
|
2260 * writes. These messages are only sent if MD_STATE_XMIT is set in flags.
|
|
2261 * MD_STATE_XMIT will be unset in 2 cases:
|
|
2262 * 1. When the state is changed to CS_RESYNC as this state change
|
|
2263 * will already have been updated on each node by the processing of the
|
|
2264 * distributed metasync command, hence no need to xmit.
|
|
2265 * 2. When the state is change to CS_OKAY after a resync has completed. Again
|
|
2266 * the resync completion will already have been processed on each node by
|
|
2267 * the processing of the MD_MN_MSG_RESYNC_PHASE_DONE message for a component
|
|
2268 * resync, hence no need to xmit.
|
|
2269 *
|
|
2270 * In case we are called from the updates of a watermark,
|
|
2271 * (then MD_STATE_WMUPDATE will be set in the ps->flags) this is due to
|
|
2272 * a metainit or similar. In this case the message that we sent to propagate
|
|
2273 * the state change must not be a class1 message as that would deadlock with
|
|
2274 * the metainit command that is still being processed.
|
|
2275 * This we achieve by creating a class2 message MD_MN_MSG_STATE_UPDATE2
|
|
2276 * instead. This also makes the submessage generator to create a class2
|
|
2277 * submessage rather than a class1 (which would also block)
|
|
2278 *
|
|
2279 * On entry, unit_writerlock is held
|
|
2280 * If MD_STATE_OCHELD is set in flags, then unit_openclose lock is
|
|
2281 * also held.
|
|
2282 */
|
|
2283 void
|
|
2284 set_sm_comp_state(
|
|
2285 mm_unit_t *un,
|
|
2286 int smi,
|
|
2287 int ci,
|
|
2288 int newstate,
|
|
2289 mddb_recid_t *extras,
|
|
2290 uint_t flags,
|
|
2291 IOLOCK *lockp
|
|
2292 )
|
|
2293 {
|
|
2294 mm_submirror_t *sm;
|
|
2295 mm_submirror_ic_t *smic;
|
|
2296 md_m_shared_t *shared;
|
|
2297 int origstate;
|
|
2298 void (*get_dev)();
|
|
2299 ms_cd_info_t cd;
|
|
2300 char devname[MD_MAX_CTDLEN];
|
|
2301 int err;
|
|
2302 set_t setno = MD_UN2SET(un);
|
|
2303 md_mn_msg_stch_t stchmsg;
|
|
2304 mdi_unit_t *ui = MDI_UNIT(MD_SID(un));
|
|
2305 md_mn_kresult_t *kresult;
|
|
2306 int rval;
|
|
2307 uint_t msgflags;
|
|
2308 md_mn_msgtype_t msgtype;
|
|
2309 int save_lock = 0;
|
|
2310 mdi_unit_t *ui_sm;
|
|
2311
|
|
2312 sm = &un->un_sm[smi];
|
|
2313 smic = &un->un_smic[smi];
|
|
2314
|
|
2315 /* If we have a real error status then turn off MD_INACCESSIBLE. */
|
|
2316 ui_sm = MDI_UNIT(getminor(md_dev64_to_dev(sm->sm_dev)));
|
|
2317 if (newstate & (CS_ERRED | CS_RESYNC | CS_LAST_ERRED) &&
|
|
2318 ui_sm->ui_tstate & MD_INACCESSIBLE) {
|
|
2319 ui_sm->ui_tstate &= ~MD_INACCESSIBLE;
|
|
2320 }
|
|
2321
|
|
2322 shared = (md_m_shared_t *)
|
|
2323 (*(smic->sm_shared_by_indx))(sm->sm_dev, sm, ci);
|
|
2324 origstate = shared->ms_state;
|
|
2325
|
|
2326 /*
|
|
2327 * If the new state is an error and the old one wasn't, generate
|
|
2328 * a console message. We do this before we send the state to other
|
|
2329 * nodes in a MN set because the state change may change the component
|
|
2330 * name if a hotspare is allocated.
|
|
2331 */
|
|
2332 if ((! (origstate & (CS_ERRED|CS_LAST_ERRED))) &&
|
|
2333 (newstate & (CS_ERRED|CS_LAST_ERRED))) {
|
|
2334
|
|
2335 get_dev =
|
|
2336 (void (*)())md_get_named_service(sm->sm_dev, 0,
|
|
2337 "get device", 0);
|
|
2338 (void) (*get_dev)(sm->sm_dev, sm, ci, &cd);
|
|
2339
|
|
2340 err = md_getdevname(setno, mddb_getsidenum(setno), 0,
|
|
2341 cd.cd_dev, devname, sizeof (devname));
|
|
2342
|
|
2343 if (err == ENOENT) {
|
|
2344 (void) md_devname(setno, cd.cd_dev, devname,
|
|
2345 sizeof (devname));
|
|
2346 }
|
|
2347
|
|
2348 cmn_err(CE_WARN, "md: %s: %s needs maintenance",
|
|
2349 md_shortname(md_getminor(sm->sm_dev)), devname);
|
|
2350
|
|
2351 if (newstate & CS_LAST_ERRED) {
|
|
2352 cmn_err(CE_WARN, "md: %s: %s last erred",
|
|
2353 md_shortname(md_getminor(sm->sm_dev)),
|
|
2354 devname);
|
|
2355
|
|
2356 } else if (shared->ms_flags & MDM_S_ISOPEN) {
|
|
2357 /*
|
|
2358 * Close the broken device and clear the open flag on
|
|
2359 * it. Closing the device means the RCM framework will
|
|
2360 * be able to unconfigure the device if required.
|
|
2361 *
|
|
2362 * We have to check that the device is open, otherwise
|
|
2363 * the first open on it has resulted in the error that
|
|
2364 * is being processed and the actual cd.cd_dev will be
|
|
2365 * NODEV64.
|
|
2366 *
|
|
2367 * If this is a multi-node mirror, then the multinode
|
|
2368 * state checks following this code will cause the
|
|
2369 * slave nodes to close the mirror in the function
|
|
2370 * mirror_set_state().
|
|
2371 */
|
|
2372 md_layered_close(cd.cd_dev, MD_OFLG_NULL);
|
|
2373 shared->ms_flags &= ~MDM_S_ISOPEN;
|
|
2374 }
|
|
2375
|
|
2376 } else if ((origstate & CS_LAST_ERRED) && (newstate & CS_ERRED) &&
|
|
2377 (shared->ms_flags & MDM_S_ISOPEN)) {
|
|
2378 /*
|
|
2379 * Similar to logic above except no log messages since we
|
|
2380 * are just transitioning from Last Erred to Erred.
|
|
2381 */
|
|
2382 get_dev = (void (*)())md_get_named_service(sm->sm_dev, 0,
|
|
2383 "get device", 0);
|
|
2384 (void) (*get_dev)(sm->sm_dev, sm, ci, &cd);
|
|
2385
|
|
2386 md_layered_close(cd.cd_dev, MD_OFLG_NULL);
|
|
2387 shared->ms_flags &= ~MDM_S_ISOPEN;
|
|
2388 }
|
|
2389
|
|
2390 if ((MD_MNSET_SETNO(setno)) && (origstate != newstate) &&
|
|
2391 (flags & MD_STATE_XMIT) && !(ui->ui_tstate & MD_ERR_PENDING)) {
|
|
2392 /*
|
|
2393 * For a multi-node mirror, send the state change to the
|
|
2394 * master, which broadcasts to all nodes, including this
|
|
2395 * one. Once the message is received, the state is set
|
|
2396 * in-core and the master commits the change to disk.
|
|
2397 * There is a case, comp_replace, where this function
|
|
2398 * can be called from within an ioctl and therefore in this
|
|
2399 * case, as the ioctl will already be called on each node,
|
|
2400 * there is no need to xmit the state change to the master for
|
|
2401 * distribution to the other nodes. MD_STATE_XMIT flag is used
|
|
2402 * to indicate whether a xmit is required. The mirror's
|
|
2403 * transient state is set to MD_ERR_PENDING to avoid sending
|
|
2404 * multiple messages.
|
|
2405 */
|
|
2406 if (newstate & (CS_ERRED|CS_LAST_ERRED))
|
|
2407 ui->ui_tstate |= MD_ERR_PENDING;
|
|
2408
|
|
2409 /*
|
|
2410 * Send a state update message to all nodes. This message
|
|
2411 * will generate 2 submessages, the first one to suspend
|
|
2412 * all writes to the mirror and the second to update the
|
|
2413 * state and resume writes.
|
|
2414 */
|
|
2415 stchmsg.msg_stch_mnum = un->c.un_self_id;
|
|
2416 stchmsg.msg_stch_sm = smi;
|
|
2417 stchmsg.msg_stch_comp = ci;
|
|
2418 stchmsg.msg_stch_new_state = newstate;
|
|
2419 stchmsg.msg_stch_hs_id = shared->ms_hs_id;
|
|
2420 #ifdef DEBUG
|
|
2421 if (mirror_debug_flag)
|
|
2422 printf("send set state, %x, %x, %x, %x, %x\n",
|
|
2423 stchmsg.msg_stch_mnum, stchmsg.msg_stch_sm,
|
|
2424 stchmsg.msg_stch_comp, stchmsg.msg_stch_new_state,
|
|
2425 stchmsg.msg_stch_hs_id);
|
|
2426 #endif
|
|
2427 if (flags & MD_STATE_WMUPDATE) {
|
|
2428 msgtype = MD_MN_MSG_STATE_UPDATE2;
|
|
2429 /*
|
|
2430 * When coming from an update of watermarks, there
|
|
2431 * must already be a message logged that triggered
|
|
2432 * this action. So, no need to log this message, too.
|
|
2433 */
|
|
2434 msgflags = MD_MSGF_NO_LOG;
|
|
2435 } else {
|
|
2436 msgtype = MD_MN_MSG_STATE_UPDATE;
|
|
2437 msgflags = MD_MSGF_DEFAULT_FLAGS;
|
|
2438 }
|
|
2439
|
|
2440 /*
|
|
2441 * If we are in the context of an ioctl, drop the ioctl lock.
|
|
2442 * lockp holds the list of locks held.
|
|
2443 *
|
|
2444 * Otherwise, increment the appropriate reacquire counters.
|
|
2445 * If openclose lock is *held, then must reacquire reader
|
|
2446 * lock before releasing the openclose lock.
|
|
2447 * Do not drop the ARRAY_WRITER lock as we may not be able
|
|
2448 * to reacquire it.
|
|
2449 */
|
|
2450 if (lockp) {
|
|
2451 if (lockp->l_flags & MD_ARRAY_WRITER) {
|
|
2452 save_lock = MD_ARRAY_WRITER;
|
|
2453 lockp->l_flags &= ~MD_ARRAY_WRITER;
|
|
2454 } else if (lockp->l_flags & MD_ARRAY_READER) {
|
|
2455 save_lock = MD_ARRAY_READER;
|
|
2456 lockp->l_flags &= ~MD_ARRAY_READER;
|
|
2457 }
|
|
2458 IOLOCK_RETURN_RELEASE(0, lockp);
|
|
2459 } else {
|
|
2460 if (flags & MD_STATE_OCHELD) {
|
|
2461 md_unit_writerexit(ui);
|
|
2462 (void) md_unit_readerlock(ui);
|
|
2463 md_unit_openclose_exit(ui);
|
|
2464 } else {
|
|
2465 md_unit_writerexit(ui);
|
|
2466 }
|
|
2467 }
|
|
2468
|
|
2469 kresult = kmem_alloc(sizeof (md_mn_kresult_t), KM_SLEEP);
|
|
2470 rval = mdmn_ksend_message(setno,
|
|
2471 msgtype,
|
|
2472 msgflags,
|
|
2473 (char *)&stchmsg,
|
|
2474 sizeof (stchmsg),
|
|
2475 kresult);
|
|
2476
|
|
2477 if (!MDMN_KSEND_MSG_OK(rval, kresult)) {
|
|
2478 mdmn_ksend_show_error(rval, kresult, "STATE UPDATE");
|
|
2479 cmn_err(CE_PANIC,
|
|
2480 "ksend_message failure: STATE_UPDATE");
|
|
2481 }
|
|
2482 kmem_free(kresult, sizeof (md_mn_kresult_t));
|
|
2483
|
|
2484 /* if dropped the lock previously, regain it */
|
|
2485 if (lockp) {
|
|
2486 IOLOCK_RETURN_REACQUIRE(lockp);
|
|
2487 lockp->l_flags |= save_lock;
|
|
2488 } else {
|
|
2489 /*
|
|
2490 * Reacquire dropped locks and update acquirecnts
|
|
2491 * appropriately.
|
|
2492 */
|
|
2493 if (flags & MD_STATE_OCHELD) {
|
|
2494 /*
|
|
2495 * openclose also grabs readerlock.
|
|
2496 */
|
|
2497 (void) md_unit_openclose_enter(ui);
|
|
2498 md_unit_readerexit(ui);
|
|
2499 (void) md_unit_writerlock(ui);
|
|
2500 } else {
|
|
2501 (void) md_unit_writerlock(ui);
|
|
2502 }
|
|
2503 }
|
|
2504
|
|
2505 ui->ui_tstate &= ~MD_ERR_PENDING;
|
|
2506 } else {
|
|
2507 shared->ms_state = newstate;
|
|
2508 uniqtime32(&shared->ms_timestamp);
|
|
2509
|
|
2510 if (newstate == CS_ERRED)
|
|
2511 shared->ms_flags |= MDM_S_NOWRITE;
|
|
2512 else
|
|
2513 shared->ms_flags &= ~MDM_S_NOWRITE;
|
|
2514
|
|
2515 shared->ms_flags &= ~MDM_S_IOERR;
|
|
2516 un->un_changecnt++;
|
|
2517 shared->ms_lasterrcnt = un->un_changecnt;
|
|
2518
|
|
2519 mirror_set_sm_state(sm, smic, SMS_RUNNING, 0);
|
|
2520 mirror_commit(un, SMI2BIT(smi), extras);
|
|
2521 }
|
|
2522
|
|
2523 if ((origstate & CS_RESYNC) && (newstate & CS_OKAY)) {
|
|
2524 /*
|
|
2525 * Resetting the Last Erred state will recursively call back
|
|
2526 * into this function (set_sm_comp_state) to update the state.
|
|
2527 */
|
|
2528 reset_lasterred(un, smi, extras, flags, lockp);
|
|
2529 }
|
|
2530 }
|
|
2531
|
|
2532 static int
|
|
2533 find_another_logical(
|
|
2534 mm_unit_t *un,
|
|
2535 mm_submirror_t *esm,
|
|
2536 diskaddr_t blk,
|
|
2537 u_longlong_t cnt,
|
|
2538 int must_be_open,
|
|
2539 int state,
|
|
2540 int err_cnt)
|
|
2541 {
|
|
2542 u_longlong_t cando;
|
|
2543 md_dev64_t dev;
|
|
2544 md_m_shared_t *s;
|
|
2545
|
|
2546 esm->sm_state |= SMS_IGNORE;
|
|
2547 while (cnt != 0) {
|
|
2548 u_longlong_t mcnt;
|
|
2549
|
|
2550 mcnt = MIN(cnt, lbtodb(1024 * 1024 * 1024)); /* 1 Gig Blks */
|
|
2551
|
|
2552 dev = select_read_unit(un, blk, mcnt, &cando, must_be_open, &s,
|
|
2553 NULL);
|
|
2554 if (dev == (md_dev64_t)0)
|
|
2555 break;
|
|
2556
|
|
2557 if ((state == CS_LAST_ERRED) &&
|
|
2558 (s->ms_state == CS_LAST_ERRED) &&
|
|
2559 (err_cnt > s->ms_lasterrcnt))
|
|
2560 break;
|
|
2561
|
|
2562 cnt -= cando;
|
|
2563 blk += cando;
|
|
2564 }
|
|
2565 esm->sm_state &= ~SMS_IGNORE;
|
|
2566 return (cnt != 0);
|
|
2567 }
|
|
2568
|
|
2569 int
|
|
2570 mirror_other_sources(mm_unit_t *un, int smi, int ci, int must_be_open)
|
|
2571 {
|
|
2572 mm_submirror_t *sm;
|
|
2573 mm_submirror_ic_t *smic;
|
|
2574 size_t count;
|
|
2575 diskaddr_t block;
|
|
2576 u_longlong_t skip;
|
|
2577 u_longlong_t size;
|
|
2578 md_dev64_t dev;
|
|
2579 int cnt;
|
|
2580 md_m_shared_t *s;
|
|
2581 int not_found;
|
|
2582
|
|
2583 sm = &un->un_sm[smi];
|
|
2584 smic = &un->un_smic[smi];
|
|
2585 dev = sm->sm_dev;
|
|
2586
|
|
2587 /*
|
|
2588 * Make sure every component of the submirror
|
|
2589 * has other sources.
|
|
2590 */
|
|
2591 if (ci < 0) {
|
|
2592 /* Find the highest lasterrcnt */
|
|
2593 cnt = (*(smic->sm_get_component_count))(dev, sm);
|
|
2594 for (ci = 0; ci < cnt; ci++) {
|
|
2595 not_found = mirror_other_sources(un, smi, ci,
|
|
2596 must_be_open);
|
|
2597 if (not_found)
|
|
2598 return (1);
|
|
2599 }
|
|
2600 return (0);
|
|
2601 }
|
|
2602
|
|
2603 /*
|
|
2604 * Make sure this component has other sources
|
|
2605 */
|
|
2606 (void) (*(smic->sm_get_bcss))
|
|
2607 (dev, sm, ci, &block, &count, &skip, &size);
|
|
2608
|
|
2609 if (count == 0)
|
|
2610 return (1);
|
|
2611
|
|
2612 s = (md_m_shared_t *)(*(smic->sm_shared_by_indx))(dev, sm, ci);
|
|
2613
|
|
2614 while (count--) {
|
|
2615 if (block >= un->c.un_total_blocks)
|
|
2616 return (0);
|
|
2617
|
|
2618 if ((block + size) > un->c.un_total_blocks)
|
|
2619 size = un->c.un_total_blocks - block;
|
|
2620
|
|
2621 not_found = find_another_logical(un, sm, block, size,
|
|
2622 must_be_open, s->ms_state, s->ms_lasterrcnt);
|
|
2623 if (not_found)
|
|
2624 return (1);
|
|
2625
|
|
2626 block += size + skip;
|
|
2627 }
|
|
2628 return (0);
|
|
2629 }
|
|
2630
|
|
2631 static void
|
|
2632 finish_error(md_mps_t *ps)
|
|
2633 {
|
|
2634 struct buf *pb;
|
|
2635 mm_unit_t *un;
|
|
2636 mdi_unit_t *ui;
|
|
2637 uint_t new_str_flags;
|
|
2638
|
|
2639 pb = ps->ps_bp;
|
|
2640 un = ps->ps_un;
|
|
2641 ui = ps->ps_ui;
|
|
2642
|
|
2643 /*
|
|
2644 * Must flag any error to the resync originator if we're performing
|
|
2645 * a Write-after-Read. This corresponds to an i/o error on a resync
|
|
2646 * target device and in this case we ought to abort the resync as there
|
|
2647 * is nothing that can be done to recover from this without operator
|
|
2648 * intervention. If we don't set the B_ERROR flag we will continue
|
|
2649 * reading from the mirror but won't write to the target (as it will
|
|
2650 * have been placed into an errored state).
|
|
2651 * To handle the case of multiple components within a submirror we only
|
|
2652 * set the B_ERROR bit if explicitly requested to via MD_MPS_FLAG_ERROR.
|
|
2653 * The originator of the resync read will cause this bit to be set if
|
|
2654 * the underlying component count is one for a submirror resync. All
|
|
2655 * other resync types will have the flag set as there is no underlying
|
|
2656 * resync which can be performed on a contained metadevice for these
|
|
2657 * resync types (optimized or component).
|
|
2658 */
|
|
2659
|
|
2660 if (ps->ps_flags & MD_MPS_WRITE_AFTER_READ) {
|
|
2661 if (ps->ps_flags & MD_MPS_FLAG_ERROR)
|
|
2662 pb->b_flags |= B_ERROR;
|
|
2663 md_kstat_done(ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ));
|
|
2664 MPS_FREE(mirror_parent_cache, ps);
|
|
2665 md_unit_readerexit(ui);
|
|
2666 md_biodone(pb);
|
|
2667 return;
|
|
2668 }
|
|
2669 /*
|
|
2670 * Set the MD_IO_COUNTED flag as we are retrying the same I/O
|
|
2671 * operation therefore this I/O request has already been counted,
|
|
2672 * the I/O count variable will be decremented by mirror_done()'s
|
|
2673 * call to md_biodone().
|
|
2674 */
|
|
2675 if (ps->ps_changecnt != un->un_changecnt) {
|
|
2676 new_str_flags = MD_STR_NOTTOP | MD_IO_COUNTED;
|
|
2677 if (ps->ps_flags & MD_MPS_WOW)
|
|
2678 new_str_flags |= MD_STR_WOW;
|
|
2679 if (ps->ps_flags & MD_MPS_MAPPED)
|
|
2680 new_str_flags |= MD_STR_MAPPED;
|
|
2681 /*
|
|
2682 * If this I/O request was a read that was part of a resync,
|
|
2683 * set MD_STR_WAR for the retried read to ensure that the
|
|
2684 * resync write (i.e. write-after-read) will be performed
|
|
2685 */
|
|
2686 if (ps->ps_flags & MD_MPS_RESYNC_READ)
|
|
2687 new_str_flags |= MD_STR_WAR;
|
|
2688 md_kstat_done(ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ));
|
|
2689 MPS_FREE(mirror_parent_cache, ps);
|
|
2690 md_unit_readerexit(ui);
|
|
2691 (void) md_mirror_strategy(pb, new_str_flags, NULL);
|
|
2692 return;
|
|
2693 }
|
|
2694
|
|
2695 pb->b_flags |= B_ERROR;
|
|
2696 md_kstat_done(ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ));
|
|
2697 MPS_FREE(mirror_parent_cache, ps);
|
|
2698 md_unit_readerexit(ui);
|
|
2699 md_biodone(pb);
|
|
2700 }
|
|
2701
|
|
2702 static void
|
|
2703 error_update_unit(md_mps_t *ps)
|
|
2704 {
|
|
2705 mm_unit_t *un;
|
|
2706 mdi_unit_t *ui;
|
|
2707 int smi; /* sub mirror index */
|
|
2708 int ci; /* errored component */
|
|
2709 set_t setno;
|
|
2710 uint_t flags; /* for set_sm_comp_state() */
|
|
2711 uint_t hspflags; /* for check_comp_4_hotspares() */
|
|
2712
|
|
2713 ui = ps->ps_ui;
|
|
2714 un = (mm_unit_t *)md_unit_writerlock(ui);
|
|
2715 setno = MD_UN2SET(un);
|
|
2716
|
|
2717 /* All of these updates have to propagated in case of MN set */
|
|
2718 flags = MD_STATE_XMIT;
|
|
2719 hspflags = MD_HOTSPARE_XMIT;
|
|
2720
|
|
2721 /* special treatment if we are called during updating watermarks */
|
|
2722 if (ps->ps_flags & MD_MPS_WMUPDATE) {
|
|
2723 flags |= MD_STATE_WMUPDATE;
|
|
2724 hspflags |= MD_HOTSPARE_WMUPDATE;
|
|
2725 }
|
|
2726 smi = 0;
|
|
2727 ci = 0;
|
|
2728 while (mirror_geterror(un, &smi, &ci, 1, 0) != 0) {
|
|
2729 if (mirror_other_sources(un, smi, ci, 0) == 1) {
|
|
2730
|
|
2731 /* Never called from ioctl context, so (IOLOCK *)NULL */
|
|
2732 set_sm_comp_state(un, smi, ci, CS_LAST_ERRED, 0, flags,
|
|
2733 (IOLOCK *)NULL);
|
|
2734 /*
|
|
2735 * For a MN set, the NOTIFY is done when the state
|
|
2736 * change is processed on each node
|
|
2737 */
|
|
2738 if (!MD_MNSET_SETNO(MD_UN2SET(un))) {
|
|
2739 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_LASTERRED,
|
|
2740 SVM_TAG_METADEVICE, setno, MD_SID(un));
|
|
2741 }
|
|
2742 continue;
|
|
2743 }
|
|
2744 /* Never called from ioctl context, so (IOLOCK *)NULL */
|
|
2745 set_sm_comp_state(un, smi, ci, CS_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_ERRED,
|
|
2753 SVM_TAG_METADEVICE, setno, MD_SID(un));
|
|
2754 }
|
|
2755 smi = 0;
|
|
2756 ci = 0;
|
|
2757 }
|
|
2758
|
|
2759 md_unit_writerexit(ui);
|
|
2760 if (MD_MNSET_SETNO(setno)) {
|
|
2761 send_poke_hotspares(setno);
|
|
2762 } else {
|
|
2763 (void) poke_hotspares();
|
|
2764 }
|
|
2765 (void) md_unit_readerlock(ui);
|
|
2766
|
|
2767 finish_error(ps);
|
|
2768 }
|
|
2769
|
|
2770 /*
|
|
2771 * When we have a B_FAILFAST IO error on a Last Erred component we need to
|
|
2772 * retry the IO without B_FAILFAST set so that we try to ensure that the
|
|
2773 * component "sees" each IO.
|
|
2774 */
|
|
2775 static void
|
|
2776 last_err_retry(md_mcs_t *cs)
|
|
2777 {
|
|
2778 struct buf *cb;
|
|
2779 md_mps_t *ps;
|
|
2780 uint_t flags;
|
|
2781
|
|
2782 cb = &cs->cs_buf;
|
|
2783 cb->b_flags &= ~B_FAILFAST;
|
|
2784
|
|
2785 /* if we're panicing just let this I/O error out */
|
|
2786 if (panicstr) {
|
|
2787 (void) mirror_done(cb);
|
|
2788 return;
|
|
2789 }
|
|
2790
|
|
2791 /* reissue the I/O */
|
|
2792
|
|
2793 ps = cs->cs_ps;
|
|
2794
|
|
2795 bioerror(cb, 0);
|
|
2796
|
|
2797 mutex_enter(&ps->ps_mx);
|
|
2798
|
|
2799 flags = MD_STR_NOTTOP;
|
|
2800 if (ps->ps_flags & MD_MPS_MAPPED)
|
|
2801 flags |= MD_STR_MAPPED;
|
|
2802 if (ps->ps_flags & MD_MPS_NOBLOCK)
|
|
2803 flags |= MD_NOBLOCK;
|
|
2804
|
|
2805 mutex_exit(&ps->ps_mx);
|
|
2806
|
|
2807 clear_retry_error(cb);
|
|
2808
|
|
2809 cmn_err(CE_NOTE, "!md: %s: Last Erred, retry I/O without B_FAILFAST",
|
|
2810 md_shortname(getminor(cb->b_edev)));
|
|
2811
|
|
2812 md_call_strategy(cb, flags, NULL);
|
|
2813 }
|
|
2814
|
|
2815 static void
|
|
2816 mirror_error(md_mps_t *ps)
|
|
2817 {
|
|
2818 int smi; /* sub mirror index */
|
|
2819 int ci; /* errored component */
|
|
2820
|
|
2821 if (panicstr) {
|
|
2822 finish_error(ps);
|
|
2823 return;
|
|
2824 }
|
|
2825
|
|
2826 if (ps->ps_flags & MD_MPS_ON_OVERLAP)
|
|
2827 mirror_overlap_chain_remove(ps);
|
|
2828
|
|
2829 smi = 0;
|
|
2830 ci = 0;
|
|
2831 if (mirror_geterror(ps->ps_un, &smi, &ci, 0, 0) != 0) {
|
|
2832 md_unit_readerexit(ps->ps_ui);
|
|
2833 daemon_request(&md_mstr_daemon, error_update_unit,
|
|
2834 (daemon_queue_t *)ps, REQ_OLD);
|
|
2835 return;
|
|
2836 }
|
|
2837
|
|
2838 finish_error(ps);
|
|
2839 }
|
|
2840
|
|
2841 static int
|
|
2842 copy_write_done(struct buf *cb)
|
|
2843 {
|
|
2844 md_mps_t *ps;
|
|
2845 buf_t *pb;
|
|
2846 char *wowbuf;
|
|
2847 wowhdr_t *wowhdr;
|
|
2848 ssize_t wow_resid;
|
|
2849
|
|
2850 /* get wowbuf ans save structure */
|
|
2851 wowbuf = cb->b_un.b_addr;
|
|
2852 wowhdr = WOWBUF_HDR(wowbuf);
|
|
2853 ps = wowhdr->wow_ps;
|
|
2854 pb = ps->ps_bp;
|
|
2855
|
|
2856 /* Save error information, then free cb */
|
|
2857 if (cb->b_flags & B_ERROR)
|
|
2858 pb->b_flags |= B_ERROR;
|
|
2859
|
|
2860 if (cb->b_flags & B_REMAPPED)
|
|
2861 bp_mapout(cb);
|
|
2862
|
|
2863 freerbuf(cb);
|
|
2864
|
|
2865 /* update residual and continue if needed */
|
|
2866 if ((pb->b_flags & B_ERROR) == 0) {
|
|
2867 wow_resid = pb->b_bcount - wowhdr->wow_offset;
|
|
2868 pb->b_resid = wow_resid;
|
|
2869 if (wow_resid > 0) {
|
|
2870 daemon_request(&md_mstr_daemon, copy_write_cont,
|
|
2871 (daemon_queue_t *)wowhdr, REQ_OLD);
|
|
2872 return (1);
|
|
2873 }
|
|
2874 }
|
|
2875
|
|
2876 /* Write is complete, release resources. */
|
|
2877 kmem_cache_free(mirror_wowblk_cache, wowhdr);
|
|
2878 ASSERT(!(ps->ps_flags & MD_MPS_ON_OVERLAP));
|
|
2879 md_kstat_done(ps->ps_ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ));
|
|
2880 MPS_FREE(mirror_parent_cache, ps);
|
|
2881 md_biodone(pb);
|
|
2882 return (0);
|
|
2883 }
|
|
2884
|
|
2885 static void
|
|
2886 copy_write_cont(wowhdr_t *wowhdr)
|
|
2887 {
|
|
2888 buf_t *pb;
|
|
2889 buf_t *cb;
|
|
2890 char *wowbuf;
|
|
2891 int wow_offset;
|
|
2892 size_t wow_resid;
|
|
2893 diskaddr_t wow_blkno;
|
|
2894
|
|
2895 wowbuf = WOWHDR_BUF(wowhdr);
|
|
2896 pb = wowhdr->wow_ps->ps_bp;
|
|
2897
|
|
2898 /* get data on current location */
|
|
2899 wow_offset = wowhdr->wow_offset;
|
|
2900 wow_resid = pb->b_bcount - wow_offset;
|
|
2901 wow_blkno = pb->b_lblkno + lbtodb(wow_offset);
|
|
2902
|
|
2903 /* setup child buffer */
|
|
2904 cb = getrbuf(KM_SLEEP);
|
|
2905 cb->b_flags = B_WRITE;
|
|
2906 cb->b_edev = pb->b_edev;
|
|
2907 cb->b_un.b_addr = wowbuf; /* change to point at WOWBUF */
|
|
2908 cb->b_bufsize = md_wowbuf_size; /* change to wowbuf_size */
|
|
2909 cb->b_iodone = copy_write_done;
|
|
2910 cb->b_bcount = MIN(md_wowbuf_size, wow_resid);
|
|
2911 cb->b_lblkno = wow_blkno;
|
|
2912
|
|
2913 /* move offset to next section */
|
|
2914 wowhdr->wow_offset += cb->b_bcount;
|
|
2915
|
|
2916 /* copy and setup write for current section */
|
|
2917 bcopy(&pb->b_un.b_addr[wow_offset], wowbuf, cb->b_bcount);
|
|
2918
|
|
2919 /* do it */
|
|
2920 /*
|
|
2921 * Do not set the MD_IO_COUNTED flag as this is a new I/O request
|
|
2922 * that handles the WOW condition. The resultant increment on the
|
|
2923 * I/O count variable is cleared by copy_write_done()'s call to
|
|
2924 * md_biodone().
|
|
2925 */
|
|
2926 (void) md_mirror_strategy(cb, MD_STR_NOTTOP | MD_STR_WOW
|
|
2927 | MD_STR_MAPPED, NULL);
|
|
2928 }
|
|
2929
|
|
2930 static void
|
|
2931 md_mirror_copy_write(md_mps_t *ps)
|
|
2932 {
|
|
2933 wowhdr_t *wowhdr;
|
|
2934
|
|
2935 wowhdr = kmem_cache_alloc(mirror_wowblk_cache, MD_ALLOCFLAGS);
|
|
2936 mirror_wowblk_init(wowhdr);
|
|
2937 wowhdr->wow_ps = ps;
|
|
2938 wowhdr->wow_offset = 0;
|
|
2939 copy_write_cont(wowhdr);
|
|
2940 }
|
|
2941
|
|
2942 static void
|
|
2943 handle_wow(md_mps_t *ps)
|
|
2944 {
|
|
2945 buf_t *pb;
|
|
2946
|
|
2947 pb = ps->ps_bp;
|
|
2948
|
|
2949 bp_mapin(pb);
|
|
2950
|
|
2951 md_mirror_wow_cnt++;
|
|
2952 if (!(pb->b_flags & B_PHYS) && (md_mirror_wow_flg & WOW_LOGIT)) {
|
|
2953 cmn_err(CE_NOTE,
|
|
2954 "md: %s, blk %lld, cnt %ld: Write on write %d occurred",
|
|
2955 md_shortname(getminor(pb->b_edev)),
|
|
2956 (longlong_t)pb->b_lblkno, pb->b_bcount, md_mirror_wow_cnt);
|
|
2957 }
|
|
2958
|
|
2959 /*
|
|
2960 * Set the MD_IO_COUNTED flag as we are retrying the same I/O
|
|
2961 * operation therefore this I/O request has already been counted,
|
|
2962 * the I/O count variable will be decremented by mirror_done()'s
|
|
2963 * call to md_biodone().
|
|
2964 */
|
|
2965 if (md_mirror_wow_flg & WOW_NOCOPY)
|
|
2966 (void) md_mirror_strategy(pb, MD_STR_NOTTOP | MD_STR_WOW |
|
|
2967 MD_STR_MAPPED | MD_IO_COUNTED, ps);
|
|
2968 else
|
|
2969 md_mirror_copy_write(ps);
|
|
2970 }
|
|
2971
|
|
2972 /*
|
|
2973 * Return true if the specified submirror is either in the Last Erred
|
|
2974 * state or is transitioning into the Last Erred state.
|
|
2975 */
|
|
2976 static bool_t
|
|
2977 submirror_is_lasterred(mm_unit_t *un, int smi)
|
|
2978 {
|
|
2979 mm_submirror_t *sm;
|
|
2980 mm_submirror_ic_t *smic;
|
|
2981 md_m_shared_t *shared;
|
|
2982 int ci;
|
|
2983 int compcnt;
|
|
2984
|
|
2985 sm = &un->un_sm[smi];
|
|
2986 smic = &un->un_smic[smi];
|
|
2987
|
|
2988 compcnt = (*(smic->sm_get_component_count)) (sm->sm_dev, un);
|
|
2989 for (ci = 0; ci < compcnt; ci++) {
|
|
2990 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx))
|
|
2991 (sm->sm_dev, sm, ci);
|
|
2992
|
|
2993 if (shared->ms_state == CS_LAST_ERRED)
|
|
2994 return (B_TRUE);
|
|
2995
|
|
2996 /*
|
|
2997 * It is not currently Last Erred, check if entering Last Erred.
|
|
2998 */
|
|
2999 if ((shared->ms_flags & MDM_S_IOERR) &&
|
|
3000 ((shared->ms_state == CS_OKAY) ||
|
|
3001 (shared->ms_state == CS_RESYNC))) {
|
|
3002 if (mirror_other_sources(un, smi, ci, 0) == 1)
|
|
3003 return (B_TRUE);
|
|
3004 }
|
|
3005 }
|
|
3006
|
|
3007 return (B_FALSE);
|
|
3008 }
|
|
3009
|
|
3010
|
|
3011 static int
|
|
3012 mirror_done(struct buf *cb)
|
|
3013 {
|
|
3014 md_mps_t *ps;
|
|
3015 md_mcs_t *cs;
|
|
3016
|
|
3017 /*LINTED*/
|
|
3018 cs = (md_mcs_t *)((caddr_t)cb - md_mirror_mcs_buf_off);
|
|
3019 ps = cs->cs_ps;
|
|
3020
|
|
3021 mutex_enter(&ps->ps_mx);
|
|
3022
|
|
3023 /* check if we need to retry an errored failfast I/O */
|
|
3024 if (cb->b_flags & B_ERROR) {
|
|
3025 struct buf *pb = ps->ps_bp;
|
|
3026
|
|
3027 if (cb->b_flags & B_FAILFAST) {
|
|
3028 int i;
|
|
3029 mm_unit_t *un = ps->ps_un;
|
|
3030
|
|
3031 for (i = 0; i < NMIRROR; i++) {
|
|
3032 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE))
|
|
3033 continue;
|
|
3034
|
|
3035 if (cb->b_edev ==
|
|
3036 md_dev64_to_dev(un->un_sm[i].sm_dev)) {
|
|
3037
|
|
3038 /*
|
|
3039 * This is the submirror that had the
|
|
3040 * error. Check if it is Last Erred.
|
|
3041 */
|
|
3042 if (submirror_is_lasterred(un, i)) {
|
|
3043 daemon_queue_t *dqp;
|
|
3044
|
|
3045 mutex_exit(&ps->ps_mx);
|
|
3046 dqp = (daemon_queue_t *)cs;
|
|
3047 dqp->dq_prev = NULL;
|
|
3048 dqp->dq_next = NULL;
|
|
3049 daemon_request(&md_done_daemon,
|
|
3050 last_err_retry, dqp,
|
|
3051 REQ_OLD);
|
|
3052 return (1);
|
|
3053 }
|
|
3054 break;
|
|
3055 }
|
|
3056 }
|
|
3057 }
|
|
3058
|
|
3059 /* continue to process the buf without doing a retry */
|
|
3060 ps->ps_flags |= MD_MPS_ERROR;
|
|
3061 pb->b_error = cb->b_error;
|
|
3062 }
|
|
3063
|
|
3064 return (mirror_done_common(cb));
|
|
3065 }
|
|
3066
|
|
3067 /*
|
|
3068 * Split from the original mirror_done function so we can handle bufs after a
|
|
3069 * retry.
|
|
3070 * ps->ps_mx is already held in the caller of this function and the cb error
|
|
3071 * has already been checked and handled in the caller.
|
|
3072 */
|
|
3073 static int
|
|
3074 mirror_done_common(struct buf *cb)
|
|
3075 {
|
|
3076 struct buf *pb;
|
|
3077 mm_unit_t *un;
|
|
3078 mdi_unit_t *ui;
|
|
3079 md_mps_t *ps;
|
|
3080 md_mcs_t *cs;
|
|
3081 size_t end_rr, start_rr, current_rr;
|
|
3082
|
|
3083 /*LINTED*/
|
|
3084 cs = (md_mcs_t *)((caddr_t)cb - md_mirror_mcs_buf_off);
|
|
3085 ps = cs->cs_ps;
|
|
3086 pb = ps->ps_bp;
|
|
3087
|
|
3088 if (cb->b_flags & B_REMAPPED)
|
|
3089 bp_mapout(cb);
|
|
3090
|
|
3091 ps->ps_frags--;
|
|
3092 if (ps->ps_frags != 0) {
|
|
3093 mutex_exit(&ps->ps_mx);
|
|
3094 kmem_cache_free(mirror_child_cache, cs);
|
|
3095 return (1);
|
|
3096 }
|
|
3097 un = ps->ps_un;
|
|
3098 ui = ps->ps_ui;
|
|
3099
|
|
3100 /*
|
|
3101 * Do not update outstanding_writes if we're running with ABR
|
|
3102 * set for this mirror or the write() was issued with MD_STR_ABR set.
|
|
3103 * Also a resync initiated write() has no outstanding_writes update
|
|
3104 * either.
|
|
3105 */
|
|
3106 if (((cb->b_flags & B_READ) == 0) &&
|
|
3107 (un->un_nsm >= 2) &&
|
|
3108 (ps->ps_call == NULL) &&
|
|
3109 !((ui->ui_tstate & MD_ABR_CAP) || (ps->ps_flags & MD_MPS_ABR)) &&
|
|
3110 !(ps->ps_flags & MD_MPS_WRITE_AFTER_READ)) {
|
|
3111 BLK_TO_RR(end_rr, ps->ps_lastblk, un);
|
|
3112 BLK_TO_RR(start_rr, ps->ps_firstblk, un);
|
|
3113 mutex_enter(&un->un_resync_mx);
|
|
3114 for (current_rr = start_rr; current_rr <= end_rr; current_rr++)
|
|
3115 un->un_outstanding_writes[current_rr]--;
|
|
3116 mutex_exit(&un->un_resync_mx);
|
|
3117 }
|
|
3118 kmem_cache_free(mirror_child_cache, cs);
|
|
3119 mutex_exit(&ps->ps_mx);
|
|
3120
|
|
3121 if (ps->ps_call != NULL) {
|
|
3122 daemon_request(&md_done_daemon, ps->ps_call,
|
|
3123 (daemon_queue_t *)ps, REQ_OLD);
|
|
3124 return (1);
|
|
3125 }
|
|
3126
|
|
3127 if ((ps->ps_flags & MD_MPS_ERROR)) {
|
|
3128 daemon_request(&md_done_daemon, mirror_error,
|
|
3129 (daemon_queue_t *)ps, REQ_OLD);
|
|
3130 return (1);
|
|
3131 }
|
|
3132
|
|
3133 if (ps->ps_flags & MD_MPS_ON_OVERLAP)
|
|
3134 mirror_overlap_chain_remove(ps);
|
|
3135
|
|
3136 /*
|
|
3137 * Handle Write-on-Write problem.
|
|
3138 * Skip In case of Raw and Direct I/O as they are
|
|
3139 * handled earlier.
|
|
3140 *
|
|
3141 */
|
|
3142 if (!(md_mirror_wow_flg & WOW_DISABLE) &&
|
|
3143 !(pb->b_flags & B_READ) &&
|
|
3144 !(ps->ps_flags & MD_MPS_WOW) &&
|
|
3145 !(pb->b_flags & B_PHYS) &&
|
|
3146 any_pages_dirty(pb)) {
|
|
3147 md_unit_readerexit(ps->ps_ui);
|
|
3148 daemon_request(&md_mstr_daemon, handle_wow,
|
|
3149 (daemon_queue_t *)ps, REQ_OLD);
|
|
3150 return (1);
|
|
3151 }
|
|
3152
|
|
3153 md_kstat_done(ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ));
|
|
3154 MPS_FREE(mirror_parent_cache, ps);
|
|
3155 md_unit_readerexit(ui);
|
|
3156 md_biodone(pb);
|
|
3157 return (0);
|
|
3158 }
|
|
3159
|
|
3160 /*
|
|
3161 * Clear error state in submirror component if the retry worked after
|
|
3162 * a failfast error.
|
|
3163 */
|
|
3164 static void
|
|
3165 clear_retry_error(struct buf *cb)
|
|
3166 {
|
|
3167 int smi;
|
|
3168 md_mcs_t *cs;
|
|
3169 mm_unit_t *un;
|
|
3170 mdi_unit_t *ui_sm;
|
|
3171 mm_submirror_t *sm;
|
|
3172 mm_submirror_ic_t *smic;
|
|
3173 u_longlong_t cnt;
|
|
3174 md_m_shared_t *shared;
|
|
3175
|
|
3176 /*LINTED*/
|
|
3177 cs = (md_mcs_t *)((caddr_t)cb - md_mirror_mcs_buf_off);
|
|
3178 un = cs->cs_ps->ps_un;
|
|
3179
|
|
3180 for (smi = 0; smi < NMIRROR; smi++) {
|
|
3181 if (!SMS_BY_INDEX_IS(un, smi, SMS_INUSE))
|
|
3182 continue;
|
|
3183
|
|
3184 if (cb->b_edev == md_dev64_to_dev(un->un_sm[smi].sm_dev)) {
|
|
3185 break;
|
|
3186 }
|
|
3187 }
|
|
3188
|
|
3189 if (smi >= NMIRROR)
|
|
3190 return;
|
|
3191
|
|
3192 sm = &un->un_sm[smi];
|
|
3193 smic = &un->un_smic[smi];
|
|
3194 cnt = cb->b_bcount;
|
|
3195
|
|
3196 ui_sm = MDI_UNIT(getminor(cb->b_edev));
|
|
3197 (void) md_unit_writerlock(ui_sm);
|
|
3198
|
|
3199 shared = (md_m_shared_t *)(*(smic->sm_shared_by_blk))(sm->sm_dev, sm,
|
|
3200 cb->b_blkno, &cnt);
|
|
3201
|
|
3202 if (shared->ms_flags & MDM_S_IOERR) {
|
|
3203 shared->ms_flags &= ~MDM_S_IOERR;
|
|
3204
|
|
3205 } else {
|
|
3206 /* the I/O buf spans components and the first one is not erred */
|
|
3207 int cnt;
|
|
3208 int i;
|
|
3209
|
|
3210 cnt = (*(smic->sm_get_component_count))(sm->sm_dev, un);
|
|
3211 for (i = 0; i < cnt; i++) {
|
|
3212 shared = (md_m_shared_t *)(*(smic->sm_shared_by_indx))
|
|
3213 (sm->sm_dev, sm, i);
|
|
3214
|
|
3215 if (shared->ms_flags & MDM_S_IOERR &&
|
|
3216 shared->ms_state == CS_OKAY) {
|
|
3217
|
|
3218 shared->ms_flags &= ~MDM_S_IOERR;
|
|
3219 break;
|
|
3220 }
|
|
3221 }
|
|
3222 }
|
|
3223
|
|
3224 md_unit_writerexit(ui_sm);
|
|
3225 }
|
|
3226
|
|
3227 static size_t
|
|
3228 mirror_map_read(
|
|
3229 md_mps_t *ps,
|
|
3230 md_mcs_t *cs,
|
|
3231 diskaddr_t blkno,
|
|
3232 u_longlong_t count
|
|
3233 )
|
|
3234 {
|
|
3235 mm_unit_t *un;
|
|
3236 buf_t *bp;
|
|
3237 u_longlong_t cando;
|
|
3238
|
|
3239 bp = &cs->cs_buf;
|
|
3240 un = ps->ps_un;
|
|
3241
|
|
3242 bp->b_lblkno = blkno;
|
|
3243 if (fast_select_read_unit(ps, cs) == 0) {
|
|
3244 bp->b_bcount = ldbtob(count);
|
|
3245 return (0);
|
|
3246 }
|
|
3247 bp->b_edev = md_dev64_to_dev(select_read_unit(un, blkno, count, &cando,
|
|
3248 0, NULL, cs));
|
|
3249 bp->b_bcount = ldbtob(cando);
|
|
3250 if (count != cando)
|
|
3251 return (cando);
|
|
3252 return (0);
|
|
3253 }
|
|
3254
|
|
3255 static void
|
|
3256 write_after_read(md_mps_t *ps)
|
|
3257 {
|
|
3258 struct buf *pb;
|
|
3259 int flags;
|
|
3260
|
|
3261 if (ps->ps_flags & MD_MPS_ERROR) {
|
|
3262 mirror_error(ps);
|
|
3263 return;
|
|
3264 }
|
|
3265
|
|
3266 pb = ps->ps_bp;
|
|
3267 md_kstat_done(ps->ps_ui, pb, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ));
|
|
3268 ps->ps_call = NULL;
|
|
3269 ps->ps_flags |= MD_MPS_WRITE_AFTER_READ;
|
|
3270 flags = MD_STR_NOTTOP | MD_STR_WAR;
|
|
3271 if (ps->ps_flags & MD_MPS_MAPPED)
|
|
3272 flags |= MD_STR_MAPPED;
|
|
3273 if (ps->ps_flags & MD_MPS_NOBLOCK)
|
|
3274 flags |= MD_NOBLOCK;
|
|
3275 if (ps->ps_flags & MD_MPS_DIRTY_RD)
|
|
3276 flags |= MD_STR_DIRTY_RD;
|
|
3277 (void) mirror_write_strategy(pb, flags, ps);
|
|
3278 }
|
|
3279
|
|
3280 static void
|
|
3281 continue_serial(md_mps_t *ps)
|
|
3282 {
|
|
3283 md_mcs_t *cs;
|
|
3284 buf_t *cb;
|
|
3285 mm_unit_t *un;
|
|
3286 int flags;
|
|
3287
|
|
3288 un = ps->ps_un;
|
|
3289 cs = kmem_cache_alloc(mirror_child_cache, MD_ALLOCFLAGS);
|
|
3290 mirror_child_init(cs);
|
|
3291 cb = &cs->cs_buf;
|
|
3292 ps->ps_call = NULL;
|
|
3293 ps->ps_frags = 1;
|
|
3294 (void) mirror_map_write(un, cs, ps, 0);
|
|
3295 flags = MD_STR_NOTTOP;
|
|
3296 if (ps->ps_flags & MD_MPS_MAPPED)
|
|
3297 flags |= MD_STR_MAPPED;
|
|
3298 md_call_strategy(cb, flags, NULL);
|
|
3299 }
|
|
3300
|
|
3301 static int
|
|
3302 mirror_map_write(mm_unit_t *un, md_mcs_t *cs, md_mps_t *ps, int war)
|
|
3303 {
|
|
3304 int i;
|
|
3305 dev_t dev; /* needed for bioclone, so not md_dev64_t */
|
|
3306 buf_t *cb;
|
|
3307 buf_t *pb;
|
|
3308 diskaddr_t blkno;
|
|
3309 size_t bcount;
|
|
3310 off_t offset;
|
|
3311
|
|
3312 pb = ps->ps_bp;
|
|
3313 cb = &cs->cs_buf;
|
|
3314 cs->cs_ps = ps;
|
|
3315
|
|
3316 i = md_find_nth_unit(ps->ps_writable_sm, ps->ps_current_sm);
|
|
3317
|
|
3318 dev = md_dev64_to_dev(un->un_sm[i].sm_dev);
|
|
3319
|
|
3320 blkno = pb->b_lblkno;
|
|
3321 bcount = pb->b_bcount;
|
|
3322 offset = 0;
|
|
3323 if (war && (blkno == 0) && (un->c.un_flag & MD_LABELED)) {
|
|
3324 blkno = DK_LABEL_LOC + 1;
|
|
3325 /*
|
|
3326 * This handles the case where we're requesting
|
|
3327 * a write to block 0 on a label partition
|
|
3328 * and the request size was smaller than the
|
|
3329 * size of the label. If this is the case
|
|
3330 * then we'll return -1. Failure to do so will
|
|
3331 * either cause the calling thread to hang due to
|
|
3332 * an ssd bug, or worse if the bcount were allowed
|
|
3333 * to go negative (ie large).
|
|
3334 */
|
|
3335 if (bcount <= DEV_BSIZE*(DK_LABEL_LOC + 1))
|
|
3336 return (-1);
|
|
3337 bcount -= (DEV_BSIZE*(DK_LABEL_LOC + 1));
|
|
3338 offset = (DEV_BSIZE*(DK_LABEL_LOC + 1));
|
|
3339 }
|
|
3340
|
|
3341 cb = md_bioclone(pb, offset, bcount, dev, blkno, mirror_done,
|
|
3342 cb, KM_NOSLEEP);
|
|
3343 if (war)
|
|
3344 cb->b_flags = (cb->b_flags & ~B_READ) | B_WRITE;
|
|
3345
|
|
3346 /*
|
|
3347 * If the submirror is in the erred stated, check if any component is
|
|
3348 * in the Last Erred state. If so, we don't want to use the B_FAILFAST
|
|
3349 * flag on the IO.
|
|
3350 *
|
|
3351 * Provide a fast path for the non-erred case (which should be the
|
|
3352 * normal case).
|
|
3353 */
|
|
3354 if (un->un_sm[i].sm_flags & MD_SM_FAILFAST) {
|
|
3355 if (un->un_sm[i].sm_state & SMS_COMP_ERRED) {
|
|
3356 mm_submirror_t *sm;
|
|
3357 mm_submirror_ic_t *smic;
|
|
3358 int ci;
|
|
3359 int compcnt;
|
|
3360
|
|
3361 sm = &un->un_sm[i];
|
|
3362 smic = &un->un_smic[i];
|
|
3363
|
|
3364 compcnt = (*(smic->sm_get_component_count))
|
|
3365 (sm->sm_dev, un);
|
|
3366 for (ci = 0; ci < compcnt; ci++) {
|
|
3367 md_m_shared_t *shared;
|
|
3368
|
|
3369 shared = (md_m_shared_t *)
|
|
3370 (*(smic->sm_shared_by_indx))(sm->sm_dev,
|
|
3371 sm, ci);
|
|
3372
|
|
3373 if (shared->ms_state == CS_LAST_ERRED)
|
|
3374 break;
|
|
3375 }
|
|
3376 if (ci >= compcnt)
|
|
3377 cb->b_flags |= B_FAILFAST;
|
|
3378
|
|
3379 } else {
|
|
3380 cb->b_flags |= B_FAILFAST;
|
|
3381 }
|
|
3382 }
|
|
3383
|
|
3384 ps->ps_current_sm++;
|
|
3385 if (ps->ps_current_sm != ps->ps_active_cnt) {
|
|
3386 if (un->un_write_option == WR_SERIAL) {
|
|
3387 ps->ps_call = continue_serial;
|
|
3388 return (0);
|
|
3389 }
|
|
3390 return (1);
|
|
3391 }
|
|
3392 return (0);
|
|
3393 }
|
|
3394
|
|
3395 /*
|
|
3396 * directed_read_done:
|
|
3397 * ------------------
|
|
3398 * Completion routine called when a DMR request has been returned from the
|
|
3399 * underlying driver. Wake-up the original ioctl() and return the data to
|
|
3400 * the user.
|
|
3401 */
|
|
3402 static void
|
|
3403 directed_read_done(md_mps_t *ps)
|
|
3404 {
|
|
3405 mm_unit_t *un;
|
|
3406 mdi_unit_t *ui;
|
|
3407
|
|
3408 un = ps->ps_un;
|
|
3409 ui = ps->ps_ui;
|
|
3410
|
|
3411 md_unit_readerexit(ui);
|
|
3412 md_kstat_done(ui, ps->ps_bp, (ps->ps_flags & MD_MPS_WRITE_AFTER_READ));
|
|
3413 ps->ps_call = NULL;
|
|
3414
|
|
3415 mutex_enter(&un->un_dmr_mx);
|
|
3416 cv_signal(&un->un_dmr_cv);
|
|
3417 mutex_exit(&un->un_dmr_mx);
|
|
3418
|
|
3419 /* release the parent structure */
|
|
3420 kmem_cache_free(mirror_parent_cache, ps);
|
|
3421 }
|
|
3422
|
|
3423 /*
|
|
3424 * daemon_io:
|
|
3425 * ------------
|
|
3426 * Called to issue a mirror_write_strategy() or mirror_read_strategy
|
|
3427 * call from a blockable context. NOTE: no mutex can be held on entry to this
|
|
3428 * routine
|
|
3429 */
|
|
3430 static void
|
|
3431 daemon_io(daemon_queue_t *dq)
|
|
3432 {
|
|
3433 md_mps_t *ps = (md_mps_t *)dq;
|
|
3434 int flag = MD_STR_NOTTOP;
|
|
3435 buf_t *pb = ps->ps_bp;
|
|
3436
|
|
3437 if (ps->ps_flags & MD_MPS_MAPPED)
|
|
3438 flag |= MD_STR_MAPPED;
|
|
3439 if (ps->ps_flags & MD_MPS_WOW)
|
|
3440 flag |= MD_STR_WOW;
|
|
3441 if (ps->ps_flags & MD_MPS_WRITE_AFTER_READ)
|
|
3442 flag |= MD_STR_WAR;
|
|
3443 if (ps->ps_flags & MD_MPS_ABR)
|
|
3444 flag |= MD_STR_ABR;
|
|
3445
|
|
3446 /*
|
|
3447 * If this is a resync read, ie MD_STR_DIRTY_RD not set, set
|
|
3448 * MD_STR_WAR before calling mirror_read_strategy
|
|
3449 */
|
|
3450 if (pb->b_flags & B_READ) {
|
|
3451 if (!(ps->ps_flags & MD_MPS_DIRTY_RD))
|
|
3452 flag |= MD_STR_WAR;
|
|
3453 mirror_read_strategy(pb, flag, ps);
|
|
3454 } else
|
|
3455 mirror_write_strategy(pb, flag, ps);
|
|
3456 }
|
|
3457
|
|
3458 /*
|
|
3459 * update_resync:
|
|
3460 * -------------
|
|
3461 * Called to update the in-core version of the resync record with the latest
|
|
3462 * version that was committed to disk when the previous mirror owner
|
|
3463 * relinquished ownership. This call is likely to block as we must hold-off
|
|
3464 * any current resync processing that may be occurring.
|
|
3465 * On completion of the resync record update we issue the mirror_write_strategy
|
|
3466 * call to complete the i/o that first started this sequence. To remove a race
|
|
3467 * condition between a new write() request which is submitted and the resync
|
|
3468 * record update we acquire the writerlock. This will hold off all i/o to the
|
|
3469 * mirror until the resync update has completed.
|
|
3470 * NOTE: no mutex can be held on entry to this routine
|
|
3471 */
|
|
3472 static void
|
|
3473 update_resync(daemon_queue_t *dq)
|
|
3474 {
|
|
3475 md_mps_t *ps = (md_mps_t *)dq;
|
|
3476 buf_t *pb = ps->ps_bp;
|
|
3477 mdi_unit_t *ui = ps->ps_ui;
|
|
3478 mm_unit_t *un;
|
|
3479 set_t setno;
|
|
3480 int restart_resync;
|
|
3481
|
|
3482 un = md_unit_writerlock(ui);
|
|
3483 ps->ps_un = un;
|
|
3484 setno = MD_MIN2SET(getminor(pb->b_edev));
|
|
3485 if (mddb_reread_rr(setno, un->un_rr_dirty_recid) == 0) {
|
|
3486 /*
|
|
3487 * Synchronize our in-core view of what regions need to be
|
|
3488 * resync'd with the on-disk version.
|
|
3489 */
|
|
3490 mutex_enter(&un->un_rrp_inflight_mx);
|
|
3491 mirror_copy_rr(howmany(un->un_rrd_num, NBBY), un->un_resync_bm,
|
|
3492 un->un_dirty_bm);
|
|
3493 mutex_exit(&un->un_rrp_inflight_mx);
|
|
3494
|
|
3495 /* Region dirty map is now up to date */
|
|
3496 }
|
|
3497 restart_resync = (un->un_rs_thread_flags & MD_RI_BLOCK_OWNER) ? 1 : 0;
|
|
3498 md_unit_writerexit(ui);
|
|
3499
|
|
3500 /* Restart the resync thread if it was previously blocked */
|
|
3501 if (restart_resync) {
|
|
3502 mutex_enter(&un->un_rs_thread_mx);
|
|
3503 un->un_rs_thread_flags &= ~MD_RI_BLOCK_OWNER;
|
|
3504 cv_signal(&un->un_rs_thread_cv);
|
|
3505 mutex_exit(&un->un_rs_thread_mx);
|
|
3506 }
|
|
3507 /* Continue with original deferred i/o */
|
|
3508 daemon_io(dq);
|
|
3509 }
|
|
3510
|
|
3511 /*
|
|
3512 * owner_timeout:
|
|
3513 * -------------
|
|
3514 * Called if the original mdmn_ksend_message() failed and the request is to be
|
|
3515 * retried. Reattempt the original ownership change.
|
|
3516 *
|
|
3517 * NOTE: called at interrupt context (see timeout(9f)).
|
|
3518 */
|
|
3519 static void
|
|
3520 owner_timeout(void *arg)
|
|
3521 {
|
|
3522 daemon_queue_t *dq = (daemon_queue_t *)arg;
|
|
3523
|
|
3524 daemon_request(&md_mirror_daemon, become_owner, dq, REQ_OLD);
|
|
3525 }
|
|
3526
|
|
3527 /*
|
|
3528 * become_owner:
|
|
3529 * ------------
|
|
3530 * Called to issue RPC request to become the owner of the mirror
|
|
3531 * associated with this i/o request. We assume that the ownership request
|
|
3532 * is synchronous, so if it succeeds we will issue the request via
|
|
3533 * mirror_write_strategy().
|
|
3534 * If multiple i/o's are outstanding we will be called from the mirror_daemon
|
|
3535 * service thread.
|
|
3536 * NOTE: no mutex should be held on entry to this routine.
|
|
3537 */
|
|
3538 static void
|
|
3539 become_owner(daemon_queue_t *dq)
|
|
3540 {
|
|
3541 md_mps_t *ps = (md_mps_t *)dq;
|
|
3542 mm_unit_t *un = ps->ps_un;
|
|
3543 buf_t *pb = ps->ps_bp;
|
|
3544 set_t setno;
|
|
3545 md_mn_kresult_t *kres;
|
|
3546 int msg_flags = md_mirror_msg_flags;
|
|
3547 md_mps_t *ps1;
|
|
3548
|
|
3549 ASSERT(dq->dq_next == NULL && dq->dq_prev == NULL);
|
|
3550
|
|
3551 /*
|
|
3552 * If we're already the mirror owner we do not need to send a message
|
|
3553 * but can simply process the i/o request immediately.
|
|
3554 * If we've already sent the request to become owner we requeue the
|
|
3555 * request as we're waiting for the synchronous ownership message to
|
|
3556 * be processed.
|
|
3557 */
|
|
3558 if (MD_MN_MIRROR_OWNER(un)) {
|
|
3559 /*
|
|
3560 * As the strategy() call will potentially block we need to
|
|
3561 * punt this to a separate thread and complete this request
|
|
3562 * as quickly as possible. Note: if we're a read request
|
|
3563 * this must be a resync, we cannot afford to be queued
|
|
3564 * behind any intervening i/o requests. In this case we put the
|
|
3565 * request on the md_mirror_rs_daemon queue.
|
|
3566 */
|
|
3567 if (pb->b_flags & B_READ) {
|
|
3568 daemon_request(&md_mirror_rs_daemon, daemon_io, dq,
|
|
3569 REQ_OLD);
|
|
3570 } else {
|
|
3571 daemon_request(&md_mirror_io_daemon, daemon_io, dq,
|
|
3572 REQ_OLD);
|
|
3573 }
|
|
3574 } else {
|
|
3575 mutex_enter(&un->un_owner_mx);
|
|
3576 if ((un->un_owner_state & MM_MN_OWNER_SENT) == 0) {
|
|
3577 md_mn_req_owner_t *msg;
|
|
3578 int rval = 0;
|
|
3579
|
|
3580 /*
|
|
3581 * Check to see that we haven't exceeded the maximum
|
|
3582 * retry count. If we have we fail the i/o as the
|
|
3583 * comms mechanism has become wedged beyond recovery.
|
|
3584 */
|
|
3585 if (dq->qlen++ >= MD_OWNER_RETRIES) {
|
|
3586 mutex_exit(&un->un_owner_mx);
|
|
3587 cmn_err(CE_WARN,
|
|
3588 "md_mirror: Request exhausted ownership "
|
|
3589 "retry limit of %d attempts", dq->qlen);
|
|
3590 pb->b_error = EIO;
|
|
3591 pb->b_flags |= B_ERROR;
|
|
3592 pb->b_resid = pb->b_bcount;
|
|
3593 kmem_cache_free(mirror_parent_cache, ps);
|
|
3594 md_biodone(pb);
|
|
3595 return;
|
|
3596 }
|
|
3597
|
|
3598 /*
|
|
3599 * Issue request to change ownership. The call is
|
|
3600 * synchronous so when it returns we can complete the
|
|
3601 * i/o (if successful), or enqueue it again so that
|
|
3602 * the operation will be retried.
|
|
3603 */
|
|
3604 un->un_owner_state |= MM_MN_OWNER_SENT;
|
|
3605 mutex_exit(&un->un_owner_mx);
|
|
3606
|
|
3607 msg = kmem_zalloc(sizeof (md_mn_req_owner_t), KM_SLEEP);
|
|
3608 setno = MD_MIN2SET(getminor(pb->b_edev));
|
|
3609 msg->mnum = MD_SID(un);
|
|
3610 msg->owner = md_mn_mynode_id;
|
|
3611 msg_flags |= MD_MSGF_NO_LOG;
|
|
3612 /*
|
|
3613 * If this IO is triggered by updating a watermark,
|
|
3614 * it might be issued by the creation of a softpartition
|
|
3615 * while the commd subsystem is suspended.
|
|
3616 * We don't want this message to block.
|
|
3617 */
|
|
3618 if (ps->ps_flags & MD_MPS_WMUPDATE) {
|
|
3619 msg_flags |= MD_MSGF_OVERRIDE_SUSPEND;
|
|
3620 }
|
|
3621
|
|
3622 kres = kmem_alloc(sizeof (md_mn_kresult_t), KM_SLEEP);
|
|
3623 rval = mdmn_ksend_message(setno,
|
|
3624 MD_MN_MSG_REQUIRE_OWNER,
|
|
3625 msg_flags, /* flags */
|
|
3626 (char *)msg,
|
|
3627 sizeof (md_mn_req_owner_t),
|
|
3628 kres);
|
|
3629
|
|
3630 kmem_free(msg, sizeof (md_mn_req_owner_t));
|
|
3631
|
|
3632 if (MDMN_KSEND_MSG_OK(rval, kres)) {
|
|
3633 dq->qlen = 0;
|
|
3634 /*
|
|
3635 * Successfully changed owner, reread the
|
|
3636 * resync record so that we have a valid idea of
|
|
3637 * any previously committed incomplete write()s.
|
|
3638 * NOTE: As we need to acquire the resync mutex
|
|
3639 * this may block, so we defer it to a separate
|
|
3640 * thread handler. This makes us (effectively)
|
|
3641 * non-blocking once the ownership message
|
|
3642 * handling has completed.
|
|
3643 */
|
|
3644 mutex_enter(&un->un_owner_mx);
|
|
3645 if (un->un_owner_state & MM_MN_BECOME_OWNER) {
|
|
3646 un->un_mirror_owner = md_mn_mynode_id;
|
|
3647 /* Sets owner of un_rr_dirty record */
|
|
3648 if (un->un_rr_dirty_recid)
|
|
3649 (void) mddb_setowner(
|
|
3650 un->un_rr_dirty_recid,
|
|
3651 md_mn_mynode_id);
|
|
3652 un->un_owner_state &=
|
|
3653 ~MM_MN_BECOME_OWNER;
|
|
3654 /*
|
|
3655 * Release the block on the current
|
|
3656 * resync region if it is blocked
|
|
3657 */
|
|
3658 ps1 = un->un_rs_prev_ovrlap;
|
|
3659 if ((ps1 != NULL) &&
|
|
3660 (ps1->ps_flags & MD_MPS_ON_OVERLAP))
|
|
3661 mirror_overlap_chain_remove(
|
|
3662 ps1);
|
|
3663 mutex_exit(&un->un_owner_mx);
|
|
3664
|
|
3665 /*
|
|
3666 * If we're a read, this must be a
|
|
3667 * resync request, issue
|
|
3668 * the i/o request on the
|
|
3669 * md_mirror_rs_daemon queue. This is
|
|
3670 * to avoid a deadlock between the
|
|
3671 * resync_unit thread and
|
|
3672 * subsequent i/o requests that may
|
|
3673 * block on the resync region.
|
|
3674 */
|
|
3675 if (pb->b_flags & B_READ) {
|
|
3676 daemon_request(
|
|
3677 &md_mirror_rs_daemon,
|
|
3678 update_resync, dq, REQ_OLD);
|
|
3679 } else {
|
|
3680 daemon_request(
|
|
3681 &md_mirror_io_daemon,
|
|
3682 update_resync, dq, REQ_OLD);
|
|
3683 }
|
|
3684 kmem_free(kres,
|
|
3685 sizeof (md_mn_kresult_t));
|
|
3686 return;
|
|
3687 } else {
|
|
3688 /*
|
|
3689 * Some other node has beaten us to
|
|
3690 * obtain ownership. We need to
|
|
3691 * reschedule our ownership request
|
|
3692 */
|
|
3693 mutex_exit(&un->un_owner_mx);
|
|
3694 }
|
|
3695 } else {
|
|
3696 mdmn_ksend_show_error(rval, kres,
|
|
3697 "MD_MN_MSG_REQUIRE_OWNER");
|
|
3698 /*
|
|
3699 * Message transport failure is handled by the
|
|
3700 * comms layer. If the ownership change request
|
|
3701 * does not succeed we need to flag the error to
|
|
3702 * the initiator of the i/o. This is handled by
|
|
3703 * the retry logic above. As the request failed
|
|
3704 * we do not know _who_ the owner of the mirror
|
|
3705 * currently is. We reset our idea of the owner
|
|
3706 * to None so that any further write()s will
|
|
3707 * attempt to become the owner again. This stops
|
|
3708 * multiple nodes writing to the same mirror
|
|
3709 * simultaneously.
|
|
3710 */
|
|
3711 mutex_enter(&un->un_owner_mx);
|
|
3712 un->un_owner_state &=
|
|
3713 ~(MM_MN_OWNER_SENT|MM_MN_BECOME_OWNER);
|
|
3714 un->un_mirror_owner = MD_MN_MIRROR_UNOWNED;
|
|
3715 mutex_exit(&un->un_owner_mx);
|
|
3716 }
|
|
3717 kmem_free(kres, sizeof (md_mn_kresult_t));
|
|
3718 } else
|
|
3719 mutex_exit(&un->un_owner_mx);
|
|
3720
|
|
3721 /*
|
|
3722 * Re-enqueue this request on the deferred i/o list. Delay the
|
|
3723 * request for md_mirror_owner_to usecs to stop thrashing.
|
|
3724 */
|
|
3725 (void) timeout(owner_timeout, dq,
|
|
3726 drv_usectohz(md_mirror_owner_to));
|
|
3727 }
|
|
3728 }
|
|
3729
|
|
3730 static void
|
|
3731 mirror_write_strategy(buf_t *pb, int flag, void *private)
|
|
3732 {
|
|
3733 md_mps_t *ps;
|
|
3734 md_mcs_t *cs;
|
|
3735 int more;
|
|
3736 mm_unit_t *un;
|
|
3737 mdi_unit_t *ui;
|
|
3738 buf_t *cb; /* child buf pointer */
|
|
3739 set_t setno;
|
|
3740 int rs_on_overlap = 0;
|
|
3741
|
|
3742 ui = MDI_UNIT(getminor(pb->b_edev));
|
|
3743 un = (mm_unit_t *)MD_UNIT(getminor(pb->b_edev));
|
|
3744
|
|
3745
|
|
3746 md_kstat_waitq_enter(ui);
|
|
3747
|
|
3748 /*
|
|
3749 * If a state change is in progress for this mirror in a MN set,
|
|
3750 * suspend all non-resync writes until the state change is complete.
|
|
3751 * The objective of this suspend is to ensure that it is not
|
|
3752 * possible for one node to read data from a submirror that another node
|
|
3753 * has not written to because of the state change. Therefore we
|
|
3754 * suspend all writes until the state change has been made. As it is
|
|
3755 * not possible to read from the target of a resync, there is no need
|
|
3756 * to suspend resync writes.
|
|
3757 */
|
|
3758
|
|
3759 if (!(flag & MD_STR_WAR)) {
|
|
3760 mutex_enter(&un->un_suspend_wr_mx);
|
|
3761 while (un->un_suspend_wr_flag) {
|
|
3762 cv_wait(&un->un_suspend_wr_cv, &un->un_suspend_wr_mx);
|
|
3763 }
|
|
3764 mutex_exit(&un->un_suspend_wr_mx);
|
|
3765 (void) md_unit_readerlock(ui);
|
|
3766 }
|
|
3767
|
|
3768 if (!(flag & MD_STR_NOTTOP)) {
|
|
3769 if (md_checkbuf(ui, (md_unit_t *)un, pb)) {
|
|
3770 md_kstat_waitq_exit(ui);
|
|
3771 return;
|
|
3772 }
|
|
3773 }
|
|
3774
|
|
3775 setno = MD_MIN2SET(getminor(pb->b_edev));
|
|
3776
|
|
3777 /* If an ABR write has been requested, set MD_STR_ABR flag */
|
|
3778 if (MD_MNSET_SETNO(setno) && (pb->b_flags & B_ABRWRITE))
|
|
3779 flag |= MD_STR_ABR;
|
|
3780
|
|
3781 if (private == NULL) {
|
|
3782 ps = kmem_cache_alloc(mirror_parent_cache, MD_ALLOCFLAGS);
|
|
3783 mirror_parent_init(ps);
|
|
3784 } else {
|
|
3785 ps = private;
|
|
3786 private = NULL;
|
|
3787 }
|
|
3788 if (flag & MD_STR_MAPPED)
|
|
3789 ps->ps_flags |= MD_MPS_MAPPED;
|
|
3790
|
|
3791 if (flag & MD_STR_WOW)
|
|
3792 ps->ps_flags |= MD_MPS_WOW;
|
|
3793
|
|
3794 if (flag & MD_STR_ABR)
|
|
3795 ps->ps_flags |= MD_MPS_ABR;
|
|
3796
|
|
3797 if (flag & MD_STR_WMUPDATE)
|
|
3798 ps->ps_flags |= MD_MPS_WMUPDATE;
|
|
3799
|
|
3800 /*
|
|
3801 * Save essential information from the original buffhdr
|
|
3802 * in the md_save structure.
|
|
3803 */
|
|
3804 ps->ps_un = un;
|
|
3805 ps->ps_ui = ui;
|
|
3806 ps->ps_bp = pb;
|
|
3807 ps->ps_addr = pb->b_un.b_addr;
|
|
3808 ps->ps_firstblk = pb->b_lblkno;
|
|
3809 ps->ps_lastblk = pb->b_lblkno + lbtodb(pb->b_bcount) - 1;
|
|
3810 ps->ps_changecnt = un->un_changecnt;
|
|
3811
|
|
3812 /*
|
|
3813 * If not MN owner and this is an ABR write, make sure the current
|
|
3814 * resync region is on the overlaps chain
|
|
3815 */
|
|
3816 mutex_enter(&un->un_owner_mx);
|
|
3817 if (MD_MNSET_SETNO(setno) && (!(MD_MN_MIRROR_OWNER(un))) &&
|
|
3818 ((ui->ui_tstate & MD_ABR_CAP) || (flag & MD_STR_ABR))) {
|
|
3819 md_mps_t *ps1;
|
|
3820 /* Block the current resync region, if not already blocked */
|
|
3821 ps1 = un->un_rs_prev_ovrlap;
|
|
3822
|
|
3823 if ((ps1 != NULL) && ((ps1->ps_firstblk != 0) ||
|
|
3824 (ps1->ps_lastblk != 0))) {
|
|
3825 /* Drop locks to avoid deadlock */
|
|
3826 mutex_exit(&un->un_owner_mx);
|
|
3827 md_unit_readerexit(ui);
|
|
3828 wait_for_overlaps(ps1, MD_OVERLAP_ALLOW_REPEAT);
|
|
3829 rs_on_overlap = 1;
|
|
3830 (void) md_unit_readerlock(ui);
|
|
3831 mutex_enter(&un->un_owner_mx);
|
|
3832 /*
|
|
3833 * Check to see if we have obtained ownership
|
|
3834 * while waiting for overlaps. If we have, remove
|
|
3835 * the resync_region entry from the overlap chain
|
|
3836 */
|
|
3837 if (MD_MN_MIRROR_OWNER(un) &&
|
|
3838 (ps1->ps_flags & MD_MPS_ON_OVERLAP)) {
|
|
3839 mirror_overlap_chain_remove(ps1);
|
|
3840 rs_on_overlap = 0;
|
|
3841 }
|
|
3842 }
|
|
3843 }
|
|
3844 mutex_exit(&un->un_owner_mx);
|
|
3845
|
|
3846
|
|
3847 /*
|
|
3848 * following keep write after read from writing to the
|
|
3849 * source in the case where it all came from one place
|
|
3850 */
|
|
3851 if (flag & MD_STR_WAR) {
|
|
3852 int abort_write = 0;
|
|
3853 /*
|
|
3854 * We are perfoming a write-after-read. This is either as a
|
|
3855 * result of a resync read or as a result of a read in a
|
|
3856 * dirty resync region when the optimized resync is not
|
|
3857 * complete. If in a MN set and a resync generated i/o,
|
|
3858 * if the current block is not in the current
|
|
3859 * resync region terminate the write as another node must have
|
|
3860 * completed this resync region
|
|
3861 */
|
|
3862 if ((MD_MNSET_SETNO(MD_UN2SET(un))) &&
|
|
3863 (!flag & MD_STR_DIRTY_RD)) {
|
|
3864 if (!IN_RESYNC_REGION(un, ps))
|
|
3865 abort_write = 1;
|
|
3866 }
|
|
3867 if ((select_write_after_read_units(un, ps) == 0) ||
|
|
3868 (abort_write)) {
|
|
3869 #ifdef DEBUG
|
|
3870 if (mirror_debug_flag)
|
|
3871 printf("Abort resync write on %x, block %lld\n",
|
|
3872 MD_SID(un), ps->ps_firstblk);
|
|
3873 #endif
|
|
3874 if (ps->ps_flags & MD_MPS_ON_OVERLAP)
|
|
3875 mirror_overlap_chain_remove(ps);
|
|
3876 kmem_cache_free(mirror_parent_cache, ps);
|
|
3877 md_kstat_waitq_exit(ui);
|
|
3878 md_unit_readerexit(ui);
|
|
3879 md_biodone(pb);
|
|
3880 return;
|
|
3881 }
|
|
3882 } else {
|
|
3883 select_write_units(un, ps);
|
|
3884
|
|
3885 /* Drop readerlock to avoid deadlock */
|
|
3886 md_unit_readerexit(ui);
|
|
3887 wait_for_overlaps(ps, MD_OVERLAP_NO_REPEAT);
|
|
3888 un = md_unit_readerlock(ui);
|
|
3889 /*
|
|
3890 * For a MN set with an ABR write, if we are now the
|
|
3891 * owner and we have a resync region on the overlap
|
|
3892 * chain, remove the entry from overlaps and retry the write.
|
|
3893 */
|
|
3894
|
|
3895 if (MD_MNSET_SETNO(setno) &&
|
|
3896 ((ui->ui_tstate & MD_ABR_CAP) || (flag & MD_STR_ABR))) {
|
|
3897 mutex_enter(&un->un_owner_mx);
|
|
3898 if (((MD_MN_MIRROR_OWNER(un))) && rs_on_overlap) {
|
|
3899 mirror_overlap_chain_remove(ps);
|
|
3900 md_kstat_waitq_exit(ui);
|
|
3901 mutex_exit(&un->un_owner_mx);
|
|
3902 md_unit_readerexit(ui);
|
|
3903 daemon_request(&md_mirror_daemon, daemon_io,
|
|
3904 (daemon_queue_t *)ps, REQ_OLD);
|
|
3905 return;
|
|
3906 }
|
|
3907 mutex_exit(&un->un_owner_mx);
|
|
3908 }
|
|
3909 }
|
|
3910
|
|
3911 /*
|
|
3912 * For Multinode mirrors with a Resync Region (not ABR) we need to
|
|
3913 * become the mirror owner before continuing with the write(). For ABR
|
|
3914 * mirrors we check that we 'own' the resync if we're in
|
|
3915 * write-after-read mode. We do this _after_ ensuring that there are no
|
|
3916 * overlaps to ensure that the once we know that we are the owner, the
|
|
3917 * readerlock will not released until the write is complete. As a
|
|
3918 * change of ownership in a MN set requires the writerlock, this
|
|
3919 * ensures that ownership cannot be changed until the write is
|
|
3920 * complete
|
|
3921 */
|
|
3922 if (MD_MNSET_SETNO(setno) && (!((ui->ui_tstate & MD_ABR_CAP) ||
|
|
3923 (flag & MD_STR_ABR)) || (flag & MD_STR_WAR))) {
|
|
3924 if (!MD_MN_MIRROR_OWNER(un)) {
|
|
3925 if (ps->ps_flags & MD_MPS_ON_OVERLAP)
|
|
3926 mirror_overlap_chain_remove(ps);
|
|
3927 md_kstat_waitq_exit(ui);
|
|
3928 ASSERT(!(flag & MD_STR_WAR));
|
|
3929 md_unit_readerexit(ui);
|
|
3930 daemon_request(&md_mirror_daemon, become_owner,
|
|
3931 (daemon_queue_t *)ps, REQ_OLD);
|
|
3932 return;
|
|
3933 }
|
|
3934 }
|
|
3935
|
|
3936 /*
|
|
3937 * Mark resync region if mirror has a Resync Region _and_ we are not
|
|
3938 * a resync initiated write(). Don't mark region if we're flagged as
|
|
3939 * an ABR write.
|
|
3940 */
|
|
3941 if (!((ui->ui_tstate & MD_ABR_CAP) || (flag & MD_STR_ABR)) &&
|
|
3942 !(flag & MD_STR_WAR)) {
|
|
3943 if (mirror_mark_resync_region(un, ps->ps_firstblk,
|
|
3944 ps->ps_lastblk)) {
|
|
3945 pb->b_flags |= B_ERROR;
|
|
3946 pb->b_resid = pb->b_bcount;
|
|
3947 ASSERT(!(ps->ps_flags & MD_MPS_ON_OVERLAP));
|
|
3948 kmem_cache_free(mirror_parent_cache, ps);
|
|
3949 md_kstat_waitq_exit(ui);
|
|
3950 md_unit_readerexit(ui);
|
|
3951 md_biodone(pb);
|
|
3952 return;
|
|
3953 }
|
|
3954 }
|
|
3955
|
|
3956 ps->ps_childbflags = pb->b_flags | B_WRITE;
|
|
3957 ps->ps_childbflags &= ~B_READ;
|
|
3958 if (flag & MD_STR_MAPPED)
|
|
3959 ps->ps_childbflags &= ~B_PAGEIO;
|
|
3960
|
|
3961 if (!(flag & MD_STR_NOTTOP) && panicstr)
|
|
3962 /* Disable WOW and don't free ps */
|
|
3963 ps->ps_flags |= (MD_MPS_WOW|MD_MPS_DONTFREE);
|
|
3964
|
|
3965 md_kstat_waitq_to_runq(ui);
|
|
3966
|
|
3967 /*
|
|
3968 * Treat Raw and Direct I/O as Write-on-Write always
|
|
3969 */
|
|
3970
|
|
3971 if (!(md_mirror_wow_flg & WOW_DISABLE) &&
|
|
3972 (md_mirror_wow_flg & WOW_PHYS_ENABLE) &&
|
|
3973 (pb->b_flags & B_PHYS) &&
|
|
3974 !(ps->ps_flags & MD_MPS_WOW)) {
|
|
3975 if (ps->ps_flags & MD_MPS_ON_OVERLAP)
|
|
3976 mirror_overlap_chain_remove(ps);
|
|
3977 md_unit_readerexit(ui);
|
|
3978 daemon_request(&md_mstr_daemon, handle_wow,
|
|
3979 (daemon_queue_t *)ps, REQ_OLD);
|
|
3980 return;
|
|
3981 }
|
|
3982
|
|
3983 ps->ps_frags = 1;
|
|
3984 do {
|
|
3985 cs = kmem_cache_alloc(mirror_child_cache, MD_ALLOCFLAGS);
|
|
3986 mirror_child_init(cs);
|
|
3987 cb = &cs->cs_buf;
|
|
3988 more = mirror_map_write(un, cs, ps, (flag & MD_STR_WAR));
|
|
3989
|
|
3990 /*
|
|
3991 * This handles the case where we're requesting
|
|
3992 * a write to block 0 on a label partition. (more < 0)
|
|
3993 * means that the request size was smaller than the
|
|
3994 * size of the label. If so this request is done.
|
|
3995 */
|
|
3996 if (more < 0) {
|
|
3997 if (ps->ps_flags & MD_MPS_ON_OVERLAP)
|
|
3998 mirror_overlap_chain_remove(ps);
|
|
3999 md_kstat_runq_exit(ui);
|
|
4000 kmem_cache_free(mirror_child_cache, cs);
|
|
4001 kmem_cache_free(mirror_parent_cache, ps);
|
|
4002 md_unit_readerexit(ui);
|
|
4003 md_biodone(pb);
|
|
4004 return;
|
|
4005 }
|
|
4006 if (more) {
|
|
4007 mutex_enter(&ps->ps_mx);
|
|
4008 ps->ps_frags++;
|
|
4009 mutex_exit(&ps->ps_mx);
|
|
4010 }
|
|
4011 md_call_strategy(cb, flag, private);
|
|
4012 } while (more);
|
|
4013
|
|
4014 if (!(flag & MD_STR_NOTTOP) && panicstr) {
|
|
4015 while (!(ps->ps_flags & MD_MPS_DONE)) {
|
|
4016 md_daemon(1, &md_done_daemon);
|
|
4017 drv_usecwait(10);
|
|
4018 }
|
|
4019 kmem_cache_free(mirror_parent_cache, ps);
|
|
4020 }
|
|
4021 }
|
|
4022
|
|
4023 static void
|
|
4024 mirror_read_strategy(buf_t *pb, int flag, void *private)
|
|
4025 {
|
|
4026 md_mps_t *ps;
|
|
4027 md_mcs_t *cs;
|
|
4028 size_t more;
|
|
4029 mm_unit_t *un;
|
|
4030 mdi_unit_t *ui;
|
|
4031 size_t current_count;
|
|
4032 diskaddr_t current_blkno;
|
|
4033 off_t current_offset;
|
|
4034 buf_t *cb; /* child buf pointer */
|
|
4035 set_t setno;
|
|
4036
|
|
4037 ui = MDI_UNIT(getminor(pb->b_edev));
|
|
4038
|
|
4039 md_kstat_waitq_enter(ui);
|
|
4040
|
|
4041 un = (mm_unit_t *)md_unit_readerlock(ui);
|
|
4042
|
|
4043 if (!(flag & MD_STR_NOTTOP)) {
|
|
4044 if (md_checkbuf(ui, (md_unit_t *)un, pb)) {
|
|
4045 md_kstat_waitq_exit(ui);
|
|
4046 return;
|
|
4047 }
|
|
4048 }
|
|
4049
|
|
4050 if (private == NULL) {
|
|
4051 ps = kmem_cache_alloc(mirror_parent_cache, MD_ALLOCFLAGS);
|
|
4052 mirror_parent_init(ps);
|
|
4053 } else {
|
|
4054 ps = private;
|
|
4055 private = NULL;
|
|
4056 }
|
|
4057
|
|
4058 if (flag & MD_STR_MAPPED)
|
|
4059 ps->ps_flags |= MD_MPS_MAPPED;
|
|
4060 if (flag & MD_NOBLOCK)
|
|
4061 ps->ps_flags |= MD_MPS_NOBLOCK;
|
|
4062 if (flag & MD_STR_WMUPDATE)
|
|
4063 ps->ps_flags |= MD_MPS_WMUPDATE;
|
|
4064
|
|
4065 /*
|
|
4066 * Check to see if this is a DMR driven read. If so we need to use the
|
|
4067 * specified side (in un->un_dmr_last_read) for the source of the data.
|
|
4068 */
|
|
4069 if (flag & MD_STR_DMR)
|
|
4070 ps->ps_flags |= MD_MPS_DMR;
|
|
4071
|
|
4072 /*
|
|
4073 * Save essential information from the original buffhdr
|
|
4074 * in the md_save structure.
|
|
4075 */
|
|
4076 ps->ps_un = un;
|
|
4077 ps->ps_ui = ui;
|
|
4078 ps->ps_bp = pb;
|
|
4079 ps->ps_addr = pb->b_un.b_addr;
|
|
4080 ps->ps_firstblk = pb->b_lblkno;
|
|
4081 ps->ps_lastblk = pb->b_lblkno + lbtodb(pb->b_bcount) - 1;
|
|
4082 ps->ps_changecnt = un->un_changecnt;
|
|
4083
|
|
4084 current_count = btodb(pb->b_bcount);
|
|
4085 current_blkno = pb->b_lblkno;
|
|
4086 current_offset = 0;
|
|
4087
|
|
4088 /*
|
|
4089 * If flag has MD_STR_WAR set this means that the read is issued by a
|
|
4090 * resync thread which may or may not be an optimised resync.
|
|
4091 *
|
|
4092 * If MD_UN_OPT_NOT_DONE is set this means that the optimized resync
|
|
4093 * code has not completed; either a resync has not started since snarf,
|
|
4094 * or there is an optimized resync in progress.
|
|
4095 *
|
|
4096 * We need to generate a write after this read in the following two
|
|
4097 * cases,
|
|
4098 *
|
|
4099 * 1. Any Resync-Generated read
|
|
4100 *
|
|
4101 * 2. Any read to a DIRTY REGION if there is an optimized resync
|
|
4102 * pending or in progress.
|
|
4103 *
|
|
4104 * The write after read is done in these cases to ensure that all sides
|
|
4105 * of the mirror are in sync with the read data and that it is not
|
|
4106 * possible for an application to read the same block multiple times
|
|
4107 * and get different data.
|
|
4108 *
|
|
4109 * This would be possible if the block was in a dirty region.
|
|
4110 *
|
|
4111 * If we're performing a directed read we don't write the data out as
|
|
4112 * the application is responsible for restoring the mirror to a known
|
|
4113 * state.
|
|
4114 */
|
|
4115 if (((MD_STATUS(un) & MD_UN_OPT_NOT_DONE) || (flag & MD_STR_WAR)) &&
|
|
4116 !(flag & MD_STR_DMR)) {
|
|
4117 size_t start_rr, i, end_rr;
|
|
4118 int region_dirty = 1;
|
|
4119
|
|
4120 /*
|
|
4121 * We enter here under three circumstances,
|
|
4122 *
|
|
4123 * MD_UN_OPT_NOT_DONE MD_STR_WAR
|
|
4124 * 0 1
|
|
4125 * 1 0
|
|
4126 * 1 1
|
|
4127 *
|
|
4128 * To be optimal we only care to explicitly check for dirty
|
|
4129 * regions in the second case since if MD_STR_WAR is set we
|
|
4130 * always do the write after read.
|
|
4131 */
|
|
4132 if (!(flag & MD_STR_WAR)) {
|
|
4133 BLK_TO_RR(end_rr, ps->ps_lastblk, un);
|
|
4134 BLK_TO_RR(start_rr, ps->ps_firstblk, un);
|
|
4135
|
|
4136 for (i = start_rr; i <= end_rr; i++)
|
|
4137 if ((region_dirty = IS_KEEPDIRTY(i, un)) != 0)
|
|
4138 break;
|
|
4139 }
|
|
4140
|
|
4141 if ((region_dirty) &&
|
|
4142 !(md_get_setstatus(MD_UN2SET(un)) & MD_SET_STALE)) {
|
|
4143 ps->ps_call = write_after_read;
|
|
4144 /*
|
|
4145 * Mark this as a RESYNC_READ in ps_flags.
|
|
4146 * This is used if the read fails during a
|
|
4147 * resync of a 3-way mirror to ensure that
|
|
4148 * the retried read to the remaining
|
|
4149 * good submirror has MD_STR_WAR set. This
|
|
4150 * is needed to ensure that the resync write
|
|
4151 * (write-after-read) takes place.
|
|
4152 */
|
|
4153 ps->ps_flags |= MD_MPS_RESYNC_READ;
|
|
4154
|
|
4155 /*
|
|
4156 * If MD_STR_FLAG_ERR is set in the flags we
|
|
4157 * set MD_MPS_FLAG_ERROR so that an error on the resync
|
|
4158 * write (issued by write_after_read) will be flagged
|
|
4159 * to the biowait'ing resync thread. This allows us to
|
|
4160 * avoid issuing further resync requests to a device
|
|
4161 * that has had a write failure.
|
|
4162 */
|
|
4163 if (flag & MD_STR_FLAG_ERR)
|
|
4164 ps->ps_flags |= MD_MPS_FLAG_ERROR;
|
|
4165
|
|
4166 setno = MD_UN2SET(un);
|
|
4167 /*
|
|
4168 * Drop the readerlock to avoid
|
|
4169 * deadlock
|
|
4170 */
|
|
4171 md_unit_readerexit(ui);
|
|
4172 wait_for_overlaps(ps, MD_OVERLAP_NO_REPEAT);
|
|
4173 un = md_unit_readerlock(ui);
|
|
4174 /*
|
|
4175 * Ensure that we are owner
|
|
4176 */
|
|
4177 if (MD_MNSET_SETNO(setno)) {
|
|
4178 /*
|
|
4179 * For a non-resync read that requires a
|
|
4180 * write-after-read to be done, set a flag
|
|
4181 * in the parent structure, so that the
|
|
4182 * write_strategy routine can omit the
|
|
4183 * test that the write is still within the
|
|
4184 * resync region
|
|
4185 */
|
|
4186 if (!(flag & MD_STR_WAR))
|
|
4187 ps->ps_flags |= MD_MPS_DIRTY_RD;
|
|
4188
|
|
4189 /*
|
|
4190 * Before reading the buffer, see if
|
|
4191 * we are the owner
|
|
4192 */
|
|
4193 if (!MD_MN_MIRROR_OWNER(un)) {
|
|
4194 ps->ps_call = NULL;
|
|
4195 mirror_overlap_chain_remove(ps);
|
|
4196 md_kstat_waitq_exit(ui);
|
|
4197 md_unit_readerexit(ui);
|
|
4198 daemon_request(
|
|
4199 &md_mirror_daemon,
|
|
4200 become_owner,
|
|
4201 (daemon_queue_t *)ps,
|
|
4202 REQ_OLD);
|
|
4203 return;
|
|
4204 }
|
|
4205 /*
|
|
4206 * For a resync read, check to see if I/O is
|
|
4207 * outside of the current resync region, or
|
|
4208 * the resync has finished. If so
|
|
4209 * just terminate the I/O
|
|
4210 */
|
|
4211 if ((flag & MD_STR_WAR) &&
|
|
4212 (!(un->c.un_status & MD_UN_WAR) ||
|
|
4213 (!IN_RESYNC_REGION(un, ps)))) {
|
|
4214 #ifdef DEBUG
|
|
4215 if (mirror_debug_flag)
|
|
4216 printf("Abort resync read "
|
|
4217 "%x: %lld\n",
|
|
4218 MD_SID(un),
|
|
4219 ps->ps_firstblk);
|
|
4220 #endif
|
|
4221 mirror_overlap_chain_remove(ps);
|
|
4222 kmem_cache_free(mirror_parent_cache,
|
|
4223 ps);
|
|
4224 md_kstat_waitq_exit(ui);
|
|
4225 md_unit_readerexit(ui);
|
|
4226 md_biodone(pb);
|
|
4227 return;
|
|
4228 }
|
|
4229 }
|
|
4230 }
|
|
4231 }
|
|
4232
|
|
4233 if (flag & MD_STR_DMR) {
|
|
4234 ps->ps_call = directed_read_done;
|
|
4235 }
|
|
4236
|
|
4237 if (!(flag & MD_STR_NOTTOP) && panicstr)
|
|
4238 ps->ps_flags |= MD_MPS_DONTFREE;
|
|
4239
|
|
4240 md_kstat_waitq_to_runq(ui);
|
|
4241
|
|
4242 ps->ps_frags++;
|
|
4243 do {
|
|
4244 cs = kmem_cache_alloc(mirror_child_cache, MD_ALLOCFLAGS);
|
|
4245 mirror_child_init(cs);
|
|
4246 cb = &cs->cs_buf;
|
|
4247 cs->cs_ps = ps;
|
|
4248
|
|
4249 cb = md_bioclone(pb, current_offset, current_count, NODEV,
|
|
4250 current_blkno, mirror_done, cb, KM_NOSLEEP);
|
|
4251
|
|
4252 more = mirror_map_read(ps, cs, current_blkno,
|
|
4253 (u_longlong_t)current_count);
|
|
4254 if (more) {
|
|
4255 mutex_enter(&ps->ps_mx);
|
|
4256 ps->ps_frags++;
|
|
4257 mutex_exit(&ps->ps_mx);
|
|
4258 }
|
|
4259
|
|
4260 /*
|
|
4261 * Do these calculations now,
|
|
4262 * so that we pickup a valid b_bcount from the chld_bp.
|
|
4263 */
|
|
4264 current_count -= more;
|
|
4265 current_offset += cb->b_bcount;
|
|
4266 current_blkno += more;
|
|
4267 md_call_strategy(cb, flag, private);
|
|
4268 } while (more);
|
|
4269
|
|
4270 if (!(flag & MD_STR_NOTTOP) && panicstr) {
|
|
4271 while (!(ps->ps_flags & MD_MPS_DONE)) {
|
|
4272 md_daemon(1, &md_done_daemon);
|
|
4273 drv_usecwait(10);
|
|
4274 }
|
|
4275 kmem_cache_free(mirror_parent_cache, ps);
|
|
4276 }
|
|
4277 }
|
|
4278
|
|
4279 void
|
|
4280 md_mirror_strategy(buf_t *bp, int flag, void *private)
|
|
4281 {
|
|
4282 set_t setno = MD_MIN2SET(getminor(bp->b_edev));
|
|
4283
|
|
4284 /*
|
|
4285 * When doing IO to a multi owner meta device, check if set is halted.
|
|
4286 * We do this check without the needed lock held, for performance
|
|
4287 * reasons.
|
|
4288 * If an IO just slips through while the set is locked via an
|
|
4289 * MD_MN_SUSPEND_SET, we don't care about it.
|
|
4290 * Only check for suspension if we are a top-level i/o request
|
|
4291 * (MD_STR_NOTTOP is cleared in 'flag').
|
|
4292 */
|
|
4293 if ((md_set[setno].s_status & (MD_SET_HALTED | MD_SET_MNSET)) ==
|
|
4294 (MD_SET_HALTED | MD_SET_MNSET)) {
|
|
4295 if ((flag & MD_STR_NOTTOP) == 0) {
|
|
4296 mutex_enter(&md_mx);
|
|
4297 /* Here we loop until the set is no longer halted */
|
|
4298 while (md_set[setno].s_status & MD_SET_HALTED) {
|
|
4299 cv_wait(&md_cv, &md_mx);
|
|
4300 }
|
|
4301 mutex_exit(&md_mx);
|
|
4302 }
|
|
4303 }
|
|
4304
|
|
4305 if ((flag & MD_IO_COUNTED) == 0) {
|
|
4306 if ((flag & MD_NOBLOCK) == 0) {
|
|
4307 if (md_inc_iocount(setno) != 0) {
|
|
4308 bp->b_flags |= B_ERROR;
|
|
4309 bp->b_error = ENXIO;
|
|
4310 bp->b_resid = bp->b_bcount;
|
|
4311 biodone(bp);
|
|
4312 return;
|
|
4313 }
|
|
4314 } else {
|
|
4315 md_inc_iocount_noblock(setno);
|
|
4316 }
|
|
4317 }
|
|
4318
|
|
4319 if (bp->b_flags & B_READ)
|
|
4320 mirror_read_strategy(bp, flag, private);
|
|
4321 else
|
|
4322 mirror_write_strategy(bp, flag, private);
|
|
4323 }
|
|
4324
|
|
4325 /*
|
|
4326 * mirror_directed_read:
|
|
4327 * --------------------
|
|
4328 * Entry-point for the DKIOCDMR ioctl. We issue a read to a specified sub-mirror
|
|
4329 * so that the application can determine what (if any) resync needs to be
|
|
4330 * performed. The data is copied out to the user-supplied buffer.
|
|
4331 *
|
|
4332 * Parameters:
|
|
4333 * mdev - dev_t for the mirror device
|
|
4334 * vdr - directed read parameters specifying location and submirror
|
|
4335 * to perform the read from
|
|
4336 * mode - used to ddi_copyout() any resulting data from the read
|
|
4337 *
|
|
4338 * Returns:
|
|
4339 * 0 success
|
|
4340 * !0 error code
|
|
4341 * EINVAL - invalid request format
|
|
4342 */
|
|
4343 int
|
|
4344 mirror_directed_read(dev_t mdev, vol_directed_rd_t *vdr, int mode)
|
|
4345 {
|
|
4346 buf_t *bp;
|
|
4347 minor_t mnum = getminor(mdev);
|
|
4348 mdi_unit_t *ui = MDI_UNIT(mnum);
|
|
4349 mm_unit_t *un;
|
|
4350 mm_submirror_t *sm;
|
|
4351 char *sm_nm;
|
|
4352 size_t namelen;
|
|
4353 uint_t next_side;
|
|
4354 void *kbuffer;
|
|
4355
|
|
4356 if (ui == NULL)
|
|
4357 return (ENXIO);
|
|
4358
|
|
4359 if (!(vdr->vdr_flags & DKV_DMR_NEXT_SIDE)) {
|
|
4360 return (EINVAL);
|
|
4361 }
|
|
4362
|
|
4363 /* Check for aligned block access. We disallow non-aligned requests. */
|
|
4364 if (vdr->vdr_offset % DEV_BSIZE) {
|
|
4365 return (EINVAL);
|
|
4366 }
|
|
4367
|
|
4368 /*
|
|
4369 * Allocate kernel buffer for target of read(). If we had a reliable
|
|
4370 * (sorry functional) DDI this wouldn't be needed.
|
|
4371 */
|
|
4372 kbuffer = kmem_alloc(vdr->vdr_nbytes, KM_NOSLEEP);
|
|
4373 if (kbuffer == NULL) {
|
|
4374 cmn_err(CE_WARN, "mirror_directed_read: couldn't allocate %lx"
|
|
4375 " bytes\n", vdr->vdr_nbytes);
|
|
4376 return (ENOMEM);
|
|
4377 }
|
|
4378
|
|
4379 bp = getrbuf(KM_SLEEP);
|
|
4380
|
|
4381 bp->b_un.b_addr = kbuffer;
|
|
4382 bp->b_flags = B_READ;
|
|
4383 bp->b_bcount = vdr->vdr_nbytes;
|
|
4384 bp->b_lblkno = lbtodb(vdr->vdr_offset);
|
|
4385 bp->b_edev = mdev;
|
|
4386
|
|
4387 un = md_unit_readerlock(ui);
|
|
4388
|
|
4389 /*
|
|
4390 * If DKV_SIDE_INIT is set we need to determine the first available
|
|
4391 * side to start reading from. If it isn't set we increment to the
|
|
4392 * next readable submirror.
|
|
4393 * If there are no readable submirrors we error out with DKV_DMR_ERROR.
|
|
4394 * Note: we check for a readable submirror on completion of the i/o so
|
|
4395 * we should _always_ have one available. If this becomes unavailable
|
|
4396 * we have missed the 'DKV_DMR_DONE' opportunity. This could happen if
|
|
4397 * a metadetach is made between the completion of one DKIOCDMR ioctl
|
|
4398 * and the start of the next (i.e. a sys-admin 'accident' occurred).
|
|
4399 * The chance of this is small, but not non-existent.
|
|
4400 */
|
|
4401 if (vdr->vdr_side == DKV_SIDE_INIT) {
|
|
4402 next_side = 0;
|
|
4403 } else {
|
|
4404 next_side = vdr->vdr_side + 1;
|
|
4405 }
|
|
4406 while ((next_side < NMIRROR) &&
|
|
4407 !SUBMIRROR_IS_READABLE(un, next_side))
|
|
4408 next_side++;
|
|
4409 if (next_side >= NMIRROR) {
|
|
4410 vdr->vdr_flags |= DKV_DMR_ERROR;
|
|
4411 freerbuf(bp);
|
|
4412 vdr->vdr_bytesread = 0;
|
|
4413 md_unit_readerexit(ui);
|
|
4414 return (0);
|
|
4415 }
|
|
4416
|
|
4417 /* Set the side to read from */
|
|
4418 un->un_dmr_last_read = next_side;
|
|
4419
|
|
4420 md_unit_readerexit(ui);
|
|
4421
|
|
4422 /*
|
|
4423 * Save timestamp for verification purposes. Can be read by debugger
|
|
4424 * to verify that this ioctl has been executed and to find the number
|
|
4425 * of DMR reads and the time of the last DMR read.
|
|
4426 */
|
|
4427 uniqtime(&mirror_dmr_stats.dmr_timestamp);
|
|
4428 mirror_dmr_stats.dmr_count++;
|
|
4429
|
|
4430 /* Issue READ request and wait for completion */
|
|
4431 mirror_read_strategy(bp, MD_STR_DMR|MD_NOBLOCK|MD_STR_NOTTOP, NULL);
|
|
4432
|
|
4433 mutex_enter(&un->un_dmr_mx);
|
|
4434 cv_wait(&un->un_dmr_cv, &un->un_dmr_mx);
|
|
4435 mutex_exit(&un->un_dmr_mx);
|
|
4436
|
|
4437 /*
|
|
4438 * Check to see if we encountered an error during the read. If so we
|
|
4439 * can make no guarantee about any possibly returned data.
|
|
4440 */
|
|
4441 if ((bp->b_flags & B_ERROR) == 0) {
|
|
4442 vdr->vdr_flags &= ~DKV_DMR_ERROR;
|
|
4443 if (bp->b_resid) {
|
|
4444 vdr->vdr_flags |= DKV_DMR_SHORT;
|
|
4445 vdr->vdr_bytesread = vdr->vdr_nbytes - bp->b_resid;
|
|
4446 } else {
|
|
4447 vdr->vdr_flags |= DKV_DMR_SUCCESS;
|
|
4448 vdr->vdr_bytesread = vdr->vdr_nbytes;
|
|
4449 }
|
|
4450 /* Copy the data read back out to the user supplied buffer */
|
|
4451 if (ddi_copyout(kbuffer, vdr->vdr_data, vdr->vdr_bytesread,
|
|
4452 mode)) {
|
|
4453 kmem_free(kbuffer, vdr->vdr_nbytes);
|
|
4454 return (EFAULT);
|
|
4455 }
|
|
4456
|
|
4457 } else {
|
|
4458 /* Error out with DKV_DMR_ERROR */
|
|
4459 vdr->vdr_flags |= DKV_DMR_ERROR;
|
|
4460 vdr->vdr_flags &= ~(DKV_DMR_SUCCESS|DKV_DMR_SHORT|DKV_DMR_DONE);
|
|
4461 }
|
|
4462 /*
|
|
4463 * Update the DMR parameters with the side and name of submirror that
|
|
4464 * we have just read from (un->un_dmr_last_read)
|
|
4465 */
|
|
4466 un = md_unit_readerlock(ui);
|
|
4467
|
|
4468 vdr->vdr_side = un->un_dmr_last_read;
|
|
4469 sm = &un->un_sm[un->un_dmr_last_read];
|
|
4470 sm_nm = md_shortname(md_getminor(sm->sm_dev));
|
|
4471
|
|
4472 namelen = MIN(MD_MAX_SIDENAME_LEN, VOL_SIDENAME);
|
|
4473 (void) strncpy(vdr->vdr_side_name, sm_nm, namelen);
|
|
4474
|
|
4475 /*
|
|
4476 * Determine if we've completed the read cycle. This is true iff the
|
|
4477 * next computed submirror (side) equals or exceeds NMIRROR. We cannot
|
|
4478 * use un_nsm as we need to handle a sparse array of submirrors (which
|
|
4479 * can occur if a submirror is metadetached).
|
|
4480 */
|
|
4481 next_side = un->un_dmr_last_read + 1;
|
|
4482 while ((next_side < NMIRROR) &&
|
|
4483 !SUBMIRROR_IS_READABLE(un, next_side))
|
|
4484 next_side++;
|
|
4485 if (next_side >= NMIRROR) {
|
|
4486 /* We've finished */
|
|
4487 vdr->vdr_flags |= DKV_DMR_DONE;
|
|
4488 }
|
|
4489
|
|
4490 md_unit_readerexit(ui);
|
|
4491 freerbuf(bp);
|
|
4492 kmem_free(kbuffer, vdr->vdr_nbytes);
|
|
4493
|
|
4494 return (0);
|
|
4495 }
|
|
4496
|
|
4497 /*
|
|
4498 * mirror_resync_message:
|
|
4499 * ---------------------
|
|
4500 * Handle the multi-node resync messages that keep all nodes within a given
|
|
4501 * disk-set in sync with their view of a mirror's resync status.
|
|
4502 *
|
|
4503 * The message types dealt with are:
|
|
4504 * MD_MN_MSG_RESYNC_STARTING - start a resync thread for a unit
|
|
4505 * MD_MN_MSG_RESYNC_NEXT - specified next region to be resynced
|
|
4506 * MD_MN_MSG_RESYNC_FINISH - stop the resync thread for a unit
|
|
4507 * MD_MN_MSG_RESYNC_PHASE_DONE - end of a resync phase, opt, submirror or comp
|
|
4508 *
|
|
4509 * Returns:
|
|
4510 * 0 Success
|
|
4511 * >0 Failure error number
|
|
4512 */
|
|
4513 int
|
|
4514 mirror_resync_message(md_mn_rs_params_t *p, IOLOCK *lockp)
|
|
4515 {
|
|
4516 mdi_unit_t *ui;
|
|
4517 mm_unit_t *un;
|
|
4518 set_t setno;
|
|
4519 int is_ABR;
|
|
4520 int smi;
|
|
4521 int ci;
|
|
4522 sm_state_t state;
|
|
4523 int broke_out;
|
|
4524 mm_submirror_t *sm;
|
|
4525 mm_submirror_ic_t *smic;
|
|
4526 md_m_shared_t *shared;
|
|
4527 md_error_t mde = mdnullerror;
|
|
4528 md_mps_t *ps;
|
|
4529 int rs_active;
|
|
4530
|
|
4531 /* Check that the given device is part of a multi-node set */
|
|
4532 setno = MD_MIN2SET(p->mnum);
|
|
4533 if (setno >= md_nsets) {
|
|
4534 return (ENXIO);
|
|
4535 }
|
|
4536 if (!MD_MNSET_SETNO(setno)) {
|
|
4537 return (EINVAL);
|
|
4538 }
|
|
4539
|
|
4540 if ((un = mirror_getun(p->mnum, &p->mde, NO_LOCK, NULL)) == NULL)
|
|
4541 return (EINVAL);
|
|
4542 if ((ui = MDI_UNIT(p->mnum)) == NULL)
|
|
4543 return (EINVAL);
|
|
4544 is_ABR = (ui->ui_tstate & MD_ABR_CAP);
|
|
4545
|
|
4546 /* Obtain the current resync status */
|
|
4547 (void) md_ioctl_readerlock(lockp, ui);
|
|
4548 rs_active = (MD_STATUS(un) & MD_UN_RESYNC_ACTIVE) ? 1 : 0;
|
|
4549 md_ioctl_readerexit(lockp);
|
|
4550
|
|
4551 switch ((md_mn_msgtype_t)p->msg_type) {
|
|
4552 case MD_MN_MSG_RESYNC_STARTING:
|
|
4553 /* Start the resync thread for the mirror */
|
|
4554 (void) mirror_resync_unit(p->mnum, NULL, &p->mde, lockp);
|
|
4555 break;
|
|
4556
|
|
4557 case MD_MN_MSG_RESYNC_NEXT:
|
|
4558 /*
|
|
4559 * We have to release any previously marked overlap regions
|
|
4560 * so that i/o can resume. Then we need to block the region
|
|
4561 * from [rs_start..rs_start+rs_size) * so that no i/o is issued.
|
|
4562 * Update un_rs_resync_done and un_rs_resync_2_do.
|
|
4563 */
|
|
4564 (void) md_ioctl_readerlock(lockp, ui);
|
|
4565 /*
|
|
4566 * Ignore the message if there is no active resync thread or
|
|
4567 * if it is for a resync type that we have already completed.
|
|
4568 * un_resync_completed is set to the last resync completed
|
|
4569 * when processing a PHASE_DONE message.
|
|
4570 */
|
|
4571 if (!rs_active || (p->rs_type == un->un_resync_completed))
|
|
4572 break;
|
|
4573 /*
|
|
4574 * If this message is for the same resync and is for an earlier
|
|
4575 * resync region, just ignore it. This can only occur if this
|
|
4576 * node has progressed on to the next resync region before
|
|
4577 * we receive this message. This can occur if the class for
|
|
4578 * this message is busy and the originator has to retry thus
|
|
4579 * allowing this node to move onto the next resync_region.
|
|
4580 */
|
|
4581 if ((p->rs_type == un->un_rs_type) &&
|
|
4582 (p->rs_start < un->un_resync_startbl))
|
|
4583 break;
|
|
4584 ps = un->un_rs_prev_ovrlap;
|
|
4585
|
|
4586 /* Allocate previous overlap reference if needed */
|
|
4587 if (ps == NULL) {
|
|
4588 ps = kmem_cache_alloc(mirror_parent_cache,
|
|
4589 MD_ALLOCFLAGS);
|
|
4590 ps->ps_un = un;
|
|
4591 ps->ps_ui = ui;
|
|
4592 ps->ps_firstblk = 0;
|
|
4593 ps->ps_lastblk = 0;
|
|
4594 ps->ps_flags = 0;
|
|
4595 md_ioctl_readerexit(lockp);
|
|
4596 (void) md_ioctl_writerlock(lockp, ui);
|
|
4597 un->un_rs_prev_ovrlap = ps;
|
|
4598 md_ioctl_writerexit(lockp);
|
|
4599 } else
|
|
4600 md_ioctl_readerexit(lockp);
|
|
4601
|
|
4602 if (p->rs_originator != md_mn_mynode_id) {
|
|
4603 /*
|
|
4604 * On all but the originating node, first update
|
|
4605 * the resync state, then unblock the previous
|
|
4606 * region and block the next one. No need
|
|
4607 * to do this if the region is already blocked.
|
|
4608 * Update the submirror state and flags from the
|
|
4609 * originator. This keeps the cluster in sync with
|
|
4610 * regards to the resync status.
|
|
4611 */
|
|
4612
|
|
4613 (void) md_ioctl_writerlock(lockp, ui);
|
|
4614 un->un_rs_resync_done = p->rs_done;
|
|
4615 un->un_rs_resync_2_do = p->rs_2_do;
|
|
4616 un->un_rs_type = p->rs_type;
|
|
4617 un->un_resync_startbl = p->rs_start;
|
|
4618 md_ioctl_writerexit(lockp);
|
|
4619 /*
|
|
4620 * Use un_owner_mx to ensure that an ownership change
|
|
4621 * cannot happen at the same time as this message
|
|
4622 */
|
|
4623 mutex_enter(&un->un_owner_mx);
|
|
4624 if (MD_MN_MIRROR_OWNER(un)) {
|
|
4625 ps->ps_firstblk = p->rs_start;
|
|
4626 ps->ps_lastblk = ps->ps_firstblk +
|
|
4627 p->rs_size - 1;
|
|
4628 } else {
|
|
4629 if ((ps->ps_firstblk != p->rs_start) ||
|
|
4630 (ps->ps_lastblk != p->rs_start +
|
|
4631 p->rs_size - 1)) {
|
|
4632 /* Remove previous overlap range */
|
|
4633 if (ps->ps_flags & MD_MPS_ON_OVERLAP)
|
|
4634 mirror_overlap_chain_remove(ps);
|
|
4635
|
|
4636 ps->ps_firstblk = p->rs_start;
|
|
4637 ps->ps_lastblk = ps->ps_firstblk +
|
|
4638 p->rs_size - 1;
|
|
4639
|
|
4640 mutex_exit(&un->un_owner_mx);
|
|
4641 /* Block this range from all i/o. */
|
|
4642 if (ps->ps_firstblk != 0 ||
|
|
4643 ps->ps_lastblk != 0)
|
|
4644 wait_for_overlaps(ps,
|
|
4645 MD_OVERLAP_ALLOW_REPEAT);
|
|
4646 mutex_enter(&un->un_owner_mx);
|
|
4647 /*
|
|
4648 * Check to see if we have obtained
|
|
4649 * ownership while waiting for
|
|
4650 * overlaps. If we have, remove
|
|
4651 * the resync_region entry from the
|
|
4652 * overlap chain
|
|
4653 */
|
|
4654 if (MD_MN_MIRROR_OWNER(un) &&
|
|
4655 (ps->ps_flags & MD_MPS_ON_OVERLAP))
|
|
4656 mirror_overlap_chain_remove(ps);
|
|
4657 }
|
|
4658 }
|
|
4659 mutex_exit(&un->un_owner_mx);
|
|
4660
|
|
4661 /*
|
|
4662 * If this is the first RESYNC_NEXT message (i.e.
|
|
4663 * MD_MN_RS_FIRST_RESYNC_NEXT set in p->rs_flags),
|
|
4664 * issue RESYNC_START NOTIFY event
|
|
4665 */
|
|
4666 if (p->rs_flags & MD_MN_RS_FIRST_RESYNC_NEXT) {
|
|
4667 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_RESYNC_START,
|
|
4668 SVM_TAG_METADEVICE, MD_UN2SET(un),
|
|
4669 MD_SID(un));
|
|
4670 }
|
|
4671
|
|
4672 /* Ensure that our local resync thread is running */
|
|
4673 if (un->un_rs_thread == NULL) {
|
|
4674 (void) mirror_resync_unit(p->mnum, NULL,
|
|
4675 &p->mde, lockp);
|
|
4676 }
|
|
4677 }
|
|
4678 break;
|
|
4679 case MD_MN_MSG_RESYNC_FINISH:
|
|
4680 /*
|
|
4681 * Complete the resync by stopping the resync thread.
|
|
4682 * Also release the previous overlap region field.
|
|
4683 * Update the resync_progress_thread by cv_signal'ing it so
|
|
4684 * that we mark the end of the resync as soon as possible. This
|
|
4685 * stops an unnecessary delay should be panic after resync
|
|
4686 * completion.
|
|
4687 */
|
|
4688 #ifdef DEBUG
|
|
4689 if (!rs_active) {
|
|
4690 if (mirror_debug_flag)
|
|
4691 printf("RESYNC_FINISH (mnum = %x), "
|
|
4692 "Resync *NOT* active",
|
|
4693 p->mnum);
|
|
4694 }
|
|
4695 #endif
|
|
4696
|
|
4697 if ((un->c.un_status & MD_UN_RESYNC_ACTIVE) &&
|
|
4698 (p->rs_originator != md_mn_mynode_id)) {
|
|
4699 mutex_enter(&un->un_rs_thread_mx);
|
|
4700 un->c.un_status &= ~MD_UN_RESYNC_CANCEL;
|
|
4701 un->un_rs_thread_flags |= MD_RI_SHUTDOWN;
|
|
4702 un->un_rs_thread_flags &=
|
|
4703 ~(MD_RI_BLOCK|MD_RI_BLOCK_OWNER);
|
|
4704 cv_signal(&un->un_rs_thread_cv);
|
|
4705 mutex_exit(&un->un_rs_thread_mx);
|
|
4706 }
|
|
4707 if (is_ABR) {
|
|
4708 /* Resync finished, if ABR set owner to NULL */
|
|
4709 mutex_enter(&un->un_owner_mx);
|
|
4710 un->un_mirror_owner = 0;
|
|
4711 mutex_exit(&un->un_owner_mx);
|
|
4712 }
|
|
4713 (void) md_ioctl_writerlock(lockp, ui);
|
|
4714 ps = un->un_rs_prev_ovrlap;
|
|
4715 if (ps != NULL) {
|
|
4716 /* Remove previous overlap range */
|
|
4717 if (ps->ps_flags & MD_MPS_ON_OVERLAP)
|
|
4718 mirror_overlap_chain_remove(ps);
|
|
4719 /*
|
|
4720 * Release the overlap range reference
|
|
4721 */
|
|
4722 un->un_rs_prev_ovrlap = NULL;
|
|
4723 kmem_cache_free(mirror_parent_cache,
|
|
4724 ps);
|
|
4725 }
|
|
4726 md_ioctl_writerexit(lockp);
|
|
4727
|
|
4728 /* Mark the resync as complete in the metadb */
|
|
4729 un->un_rs_resync_done = p->rs_done;
|
|
4730 un->un_rs_resync_2_do = p->rs_2_do;
|
|
4731 un->un_rs_type = p->rs_type;
|
|
4732 mutex_enter(&un->un_rs_progress_mx);
|
|
4733 cv_signal(&un->un_rs_progress_cv);
|
|
4734 mutex_exit(&un->un_rs_progress_mx);
|
|
4735
|
|
4736 un = md_ioctl_writerlock(lockp, ui);
|
|
4737 un->c.un_status &= ~MD_UN_RESYNC_ACTIVE;
|
|
4738 /* Deal with any pending grow_unit */
|
|
4739 if (un->c.un_status & MD_UN_GROW_PENDING) {
|
|
4740 if ((mirror_grow_unit(un, &mde) != 0) ||
|
|
4741 (! mdismderror(&mde, MDE_GROW_DELAYED))) {
|
|
4742 un->c.un_status &= ~MD_UN_GROW_PENDING;
|
|
4743 }
|
|
4744 }
|
|
4745 md_ioctl_writerexit(lockp);
|
|
4746 break;
|
|
4747
|
|
4748 case MD_MN_MSG_RESYNC_PHASE_DONE:
|
|
4749 /*
|
|
4750 * A phase of the resync, optimized. component or
|
|
4751 * submirror is complete. Update mirror status.
|
|
4752 * If the flag CLEAR_OPT_NOT_DONE is set, it means that the
|
|
4753 * mirror owner is peforming a resync. If we have just snarfed
|
|
4754 * this set, then we must clear any of the flags set at snarf
|
|
4755 * time by unit_setup_resync().
|
|
4756 * Note that unit_setup_resync() sets up these flags to
|
|
4757 * indicate that an optimized resync is required. These flags
|
|
4758 * need to be reset because if we get here, the mirror owner
|
|
4759 * will have handled the optimized resync.
|
|
4760 * The flags that must be cleared are MD_UN_OPT_NOT_DONE and
|
|
4761 * MD_UN_WAR. In addition, for each submirror,
|
|
4762 * MD_SM_RESYNC_TARGET must be cleared and SMS_OFFLINE_RESYNC
|
|
4763 * set to SMS_OFFLINE.
|
|
4764 */
|
|
4765 #ifdef DEBUG
|
|
4766 if (mirror_debug_flag)
|
|
4767 printf("phase done mess received from %d, mnum=%x,"
|
|
4768 "type=%x, flags=%x\n", p->rs_originator, p->mnum,
|
|
4769 p->rs_type, p->rs_flags);
|
|
4770 #endif
|
|
4771 /*
|
|
4772 * Ignore the message if there is no active resync thread.
|
|
4773 */
|
|
4774 if (!rs_active)
|
|
4775 break;
|
|
4776
|
|
4777 broke_out = p->rs_flags & MD_MN_RS_ERR;
|
|
4778 switch (RS_TYPE(p->rs_type)) {
|
|
4779 case MD_RS_OPTIMIZED:
|
|
4780 un = md_ioctl_writerlock(lockp, ui);
|
|
4781 if (p->rs_flags & MD_MN_RS_CLEAR_OPT_NOT_DONE) {
|
|
4782 /* If we are originator, just clear rs_type */
|
|
4783 if (p->rs_originator == md_mn_mynode_id) {
|
|
4784 SET_RS_TYPE_NONE(un->un_rs_type);
|
|
4785 md_ioctl_writerexit(lockp);
|
|
4786 break;
|
|
4787 }
|
|
4788 /*
|
|
4789 * If CLEAR_OPT_NOT_DONE is set, only clear the
|
|
4790 * flags if OPT_NOT_DONE is set *and* rs_type
|
|
4791 * is MD_RS_NONE.
|
|
4792 */
|
|
4793 if ((un->c.un_status & MD_UN_OPT_NOT_DONE) &&
|
|
4794 (RS_TYPE(un->un_rs_type) == MD_RS_NONE)) {
|
|
4795 /* No resync in progress */
|
|
4796 un->c.un_status &= ~MD_UN_OPT_NOT_DONE;
|
|
4797 un->c.un_status &= ~MD_UN_WAR;
|
|
4798 } else {
|
|
4799 /*
|
|
4800 * We are in the middle of an
|
|
4801 * optimized resync and this message
|
|
4802 * should be ignored.
|
|
4803 */
|
|
4804 md_ioctl_writerexit(lockp);
|
|
4805 break;
|
|
4806 }
|
|
4807 } else {
|
|
4808 /*
|
|
4809 * This is the end of an optimized resync,
|
|
4810 * clear the OPT_NOT_DONE and OFFLINE_SM flags
|
|
4811 */
|
|
4812
|
|
4813 un->c.un_status &= ~MD_UN_KEEP_DIRTY;
|
|
4814 if (!broke_out)
|
|
4815 un->c.un_status &= ~MD_UN_WAR;
|
|
4816 }
|
|
4817
|
|
4818 /*
|
|
4819 * Set resync_completed to last resync type and then
|
|
4820 * clear resync_type to indicate no resync in progress
|
|
4821 */
|
|
4822 un->un_resync_completed = un->un_rs_type;
|
|
4823 SET_RS_TYPE_NONE(un->un_rs_type);
|
|
4824
|
|
4825 /*
|
|
4826 * If resync is as a result of a submirror ONLINE,
|
|
4827 * reset the submirror state to SMS_RUNNING if the
|
|
4828 * resync was ok else set back to SMS_OFFLINE.
|
|
4829 */
|
|
4830 for (smi = 0; smi < NMIRROR; smi++) {
|
|
4831 un->un_sm[smi].sm_flags &=
|
|
4832 ~MD_SM_RESYNC_TARGET;
|
|
4833 if (SMS_BY_INDEX_IS(un, smi,
|
|
4834 SMS_OFFLINE_RESYNC)) {
|
|
4835 if (p->rs_flags &
|
|
4836 MD_MN_RS_CLEAR_OPT_NOT_DONE) {
|
|
4837 state = SMS_OFFLINE;
|
|
4838 } else {
|
|
4839 state = (broke_out ?
|
|
4840 SMS_OFFLINE : SMS_RUNNING);
|
|
4841 }
|
|
4842 mirror_set_sm_state(
|
|
4843 &un->un_sm[smi],
|
|
4844 &un->un_smic[smi], state,
|
|
4845 broke_out);
|
|
4846 mirror_commit(un, NO_SUBMIRRORS,
|
|
4847 0);
|
|
4848 }
|
|
4849 /*
|
|
4850 * If we still have an offline submirror, reset
|
|
4851 * the OFFLINE_SM flag in the mirror status
|
|
4852 */
|
|
4853 if (SMS_BY_INDEX_IS(un, smi,
|
|
4854 SMS_OFFLINE))
|
|
4855 un->c.un_status |=
|
|
4856 MD_UN_OFFLINE_SM;
|
|
4857 }
|
|
4858 md_ioctl_writerexit(lockp);
|
|
4859 break;
|
|
4860 case MD_RS_SUBMIRROR:
|
|
4861 un = md_ioctl_writerlock(lockp, ui);
|
|
4862 smi = RS_SMI(p->rs_type);
|
|
4863 sm = &un->un_sm[smi];
|
|
4864 smic = &un->un_smic[smi];
|
|
4865 /* Clear RESYNC target */
|
|
4866 un->un_sm[smi].sm_flags &= ~MD_SM_RESYNC_TARGET;
|
|
4867 /*
|
|
4868 * Set resync_completed to last resync type and then
|
|
4869 * clear resync_type to indicate no resync in progress
|
|
4870 */
|
|
4871 un->un_resync_completed = un->un_rs_type;
|
|
4872 SET_RS_TYPE_NONE(un->un_rs_type);
|
|
4873 /*
|
|
4874 * If the resync completed ok reset the submirror
|
|
4875 * state to SMS_RUNNING else reset it to SMS_ATTACHED
|
|
4876 */
|
|
4877 state = (broke_out ?
|
|
4878 SMS_ATTACHED : SMS_RUNNING);
|
|
4879 mirror_set_sm_state(sm, smic, state, broke_out);
|
|
4880 un->c.un_status &= ~MD_UN_WAR;
|
|
4881 mirror_commit(un, SMI2BIT(smi), 0);
|
|
4882 md_ioctl_writerexit(lockp);
|
|
4883 break;
|
|
4884 case MD_RS_COMPONENT:
|
|
4885 un = md_ioctl_writerlock(lockp, ui);
|
|
4886 smi = RS_SMI(p->rs_type);
|
|
4887 ci = RS_CI(p->rs_type);
|
|
4888 sm = &un->un_sm[smi];
|
|
4889 smic = &un->un_smic[smi];
|
|
4890 shared = (md_m_shared_t *)
|
|
4891 (*(smic->sm_shared_by_indx))
|
|
4892 (sm->sm_dev, sm, ci);
|
|
4893 un->c.un_status &= ~MD_UN_WAR;
|
|
4894 /* Clear RESYNC target */
|
|
4895 un->un_sm[smi].sm_flags &= ~MD_SM_RESYNC_TARGET;
|
|
4896 /*
|
|
4897 * Set resync_completed to last resync type and then
|
|
4898 * clear resync_type to indicate no resync in progress
|
|
4899 */
|
|
4900 un->un_resync_completed = un->un_rs_type;
|
|
4901 SET_RS_TYPE_NONE(un->un_rs_type);
|
|
4902
|
|
4903 /*
|
|
4904 * If the resync completed ok, set the component state
|
|
4905 * to CS_OKAY.
|
|
4906 */
|
|
4907 if (broke_out)
|
|
4908 shared->ms_flags |= MDM_S_RS_TRIED;
|
|
4909 else {
|
|
4910 /*
|
|
4911 * As we don't transmit the changes,
|
|
4912 * no need to drop the lock.
|
|
4913 */
|
|
4914 set_sm_comp_state(un, smi, ci, CS_OKAY, 0,
|
|
4915 MD_STATE_NO_XMIT, (IOLOCK *)NULL);
|
|
4916 }
|
|
4917 md_ioctl_writerexit(lockp);
|
|
4918 default:
|
|
4919 break;
|
|
4920 }
|
|
4921 /*
|
|
4922 * If the purpose of this PHASE_DONE message is just to
|
|
4923 * indicate to all other nodes that the optimized resync
|
|
4924 * required (OPT_NOT_DONE) flag is to be cleared, there is
|
|
4925 * no need to generate a notify event as there has not
|
|
4926 * actually been a resync.
|
|
4927 */
|
|
4928 if (!(p->rs_flags & MD_MN_RS_CLEAR_OPT_NOT_DONE)) {
|
|
4929 if (broke_out) {
|
|
4930 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_RESYNC_FAILED,
|
|
4931 SVM_TAG_METADEVICE, MD_UN2SET(un),
|
|
4932 MD_SID(un));
|
|
4933 } else {
|
|
4934 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_RESYNC_DONE,
|
|
4935 SVM_TAG_METADEVICE, MD_UN2SET(un),
|
|
4936 MD_SID(un));
|
|
4937 }
|
|
4938 }
|
|
4939 break;
|
|
4940
|
|
4941 default:
|
|
4942 #ifdef DEBUG
|
|
4943 cmn_err(CE_PANIC, "mirror_resync_message: Unknown message type"
|
|
4944 " %x\n", p->msg_type);
|
|
4945 #endif
|
|
4946 return (EINVAL);
|
|
4947 }
|
|
4948 return (0);
|
|
4949 }
|
|
4950
|
|
4951 /* Return a -1 if snarf of optimized record failed and set should be released */
|
|
4952 static int
|
|
4953 mirror_snarf(md_snarfcmd_t cmd, set_t setno)
|
|
4954 {
|
|
4955 mddb_recid_t recid;
|
|
4956 int gotsomething;
|
|
4957 int all_mirrors_gotten;
|
|
4958 mm_unit_t *un;
|
|
4959 mddb_type_t typ1;
|
|
4960 mddb_de_ic_t *dep;
|
|
4961 mddb_rb32_t *rbp;
|
|
4962 size_t newreqsize;
|
|
4963 mm_unit_t *big_un;
|
|
4964 mm_unit32_od_t *small_un;
|
|
4965 int retval;
|
|
4966 mdi_unit_t *ui;
|
|
4967
|
|
4968 if (cmd == MD_SNARF_CLEANUP) {
|
|
4969 if (md_get_setstatus(setno) & MD_SET_STALE)
|
|
4970 return (0);
|
|
4971
|
|
4972 recid = mddb_makerecid(setno, 0);
|
|
4973 typ1 = (mddb_type_t)md_getshared_key(setno,
|
|
4974 mirror_md_ops.md_driver.md_drivername);
|
|
4975 while ((recid = mddb_getnextrec(recid, typ1, MIRROR_REC)) > 0) {
|
|
4976 if (mddb_getrecprivate(recid) & MD_PRV_CLEANUP) {
|
|
4977 un = (mm_unit_t *)mddb_getrecaddr(recid);
|
|
4978 mirror_cleanup(un);
|
|
4979 recid = mddb_makerecid(setno, 0);
|
|
4980 }
|
|
4981 }
|
|
4982 return (0);
|
|
4983 }
|
|
4984
|
|
4985 all_mirrors_gotten = 1;
|
|
4986 gotsomething = 0;
|
|
4987
|
|
4988 recid = mddb_makerecid(setno, 0);
|
|
4989 typ1 = (mddb_type_t)md_getshared_key(setno,
|
|
4990 mirror_md_ops.md_driver.md_drivername);
|
|
4991
|
|
4992 while ((recid = mddb_getnextrec(recid, typ1, MIRROR_REC)) > 0) {
|
|
4993 if (mddb_getrecprivate(recid) & MD_PRV_GOTIT)
|
|
4994 continue;
|
|
4995
|
|
4996 dep = mddb_getrecdep(recid);
|
|
4997 dep->de_flags = MDDB_F_MIRROR;
|
|
4998 rbp = dep->de_rb;
|
|
4999
|
|
5000 if ((rbp->rb_revision == MDDB_REV_RB) &&
|
|
5001 ((rbp->rb_private & MD_PRV_CONVD) == 0)) {
|
|
5002 /*
|
|
5003 * This means, we have an old and small record
|
|
5004 * and this record hasn't already been converted.
|
|
5005 * Before we create an incore metadevice from this
|
|
5006 * we have to convert it to a big record.
|
|
5007 */
|
|
5008 small_un = (mm_unit32_od_t *)mddb_getrecaddr(recid);
|
|
5009 newreqsize = sizeof (mm_unit_t);
|
|
5010 big_un = (mm_unit_t *)kmem_zalloc(newreqsize, KM_SLEEP);
|
|
5011 mirror_convert((caddr_t)small_un, (caddr_t)big_un,
|
|
5012 SMALL_2_BIG);
|
|
5013 kmem_free(small_un, dep->de_reqsize);
|
|
5014
|
|
5015 /*
|
|
5016 * Update userdata and incore userdata
|
|
5017 * incores are at the end of un
|
|
5018 */
|
|
5019 dep->de_rb_userdata_ic = big_un;
|
|
5020 dep->de_rb_userdata = big_un;
|
|
5021 dep->de_icreqsize = newreqsize;
|
|
5022 un = big_un;
|
|
5023 rbp->rb_private |= MD_PRV_CONVD;
|
|
5024 } else {
|
|
5025 /* Big device */
|
|
5026 un = (mm_unit_t *)mddb_getrecaddr_resize(recid,
|
|
5027 sizeof (*un), 0);
|
|
5028 }
|
|
5029
|
|
5030 /* Set revision and flag accordingly */
|
|
5031 if (rbp->rb_revision == MDDB_REV_RB) {
|
|
5032 un->c.un_revision = MD_32BIT_META_DEV;
|
|
5033 } else {
|
|
5034 un->c.un_revision = MD_64BIT_META_DEV;
|
|
5035 un->c.un_flag |= MD_EFILABEL;
|
|
5036 }
|
|
5037
|
|
5038 /*
|
|
5039 * Create minor device node for snarfed entry.
|
|
5040 */
|
|
5041 (void) md_create_minor_node(setno, MD_SID(un));
|
|
5042
|
|
5043 if (MD_UNIT(MD_SID(un)) != NULL) {
|
|
5044 mddb_setrecprivate(recid, MD_PRV_PENDDEL);
|
|
5045 continue;
|
|
5046 }
|
|
5047 all_mirrors_gotten = 0;
|
|
5048 retval = mirror_build_incore(un, 1);
|
|
5049 if (retval == 0) {
|
|
5050 mddb_setrecprivate(recid, MD_PRV_GOTIT);
|
|
5051 md_create_unit_incore(MD_SID(un), &mirror_md_ops, 0);
|
|
5052 resync_start_timeout(setno);
|
|
5053 gotsomething = 1;
|
|
5054 } else if (retval == -1) {
|
|
5055 return (-1);
|
|
5056 }
|
|
5057 /*
|
|
5058 * Set flag to indicate that the mirror has not yet
|
|
5059 * been through a reconfig. This flag is used for MN sets
|
|
5060 * when determining whether to update the mirror state from
|
|
5061 * the Master node.
|
|
5062 */
|
|
5063 if (MD_MNSET_SETNO(setno)) {
|
|
5064 ui = MDI_UNIT(MD_SID(un));
|
|
5065 ui->ui_tstate |= MD_RESYNC_NOT_DONE;
|
|
5066 }
|
|
5067 }
|
|
5068
|
|
5069 if (!all_mirrors_gotten)
|
|
5070 return (gotsomething);
|
|
5071
|
|
5072 recid = mddb_makerecid(setno, 0);
|
|
5073 while ((recid = mddb_getnextrec(recid, typ1, RESYNC_REC)) > 0)
|
|
5074 if (!(mddb_getrecprivate(recid) & MD_PRV_GOTIT))
|
|
5075 mddb_setrecprivate(recid, MD_PRV_PENDDEL);
|
|
5076
|
|
5077 return (0);
|
|
5078 }
|
|
5079
|
|
5080 static int
|
|
5081 mirror_halt(md_haltcmd_t cmd, set_t setno)
|
|
5082 {
|
|
5083 unit_t i;
|
|
5084 mdi_unit_t *ui;
|
|
5085 minor_t mnum;
|
|
5086 int reset_mirror_flag = 0;
|
|
5087
|
|
5088 if (cmd == MD_HALT_CLOSE)
|
|
5089 return (0);
|
|
5090
|
|
5091 if (cmd == MD_HALT_OPEN)
|
|
5092 return (0);
|
|
5093
|
|
5094 if (cmd == MD_HALT_UNLOAD)
|
|
5095 return (0);
|
|
5096
|
|
5097 if (cmd == MD_HALT_CHECK) {
|
|
5098 for (i = 0; i < md_nunits; i++) {
|
|
5099 mnum = MD_MKMIN(setno, i);
|
|
5100 if ((ui = MDI_UNIT(mnum)) == NULL)
|
|
5101 continue;
|
|
5102 if (ui->ui_opsindex != mirror_md_ops.md_selfindex)
|
|
5103 continue;
|
|
5104 if (md_unit_isopen(ui))
|
|
5105 return (1);
|
|
5106 }
|
|
5107 return (0);
|
|
5108 }
|
|
5109
|
|
5110 if (cmd != MD_HALT_DOIT)
|
|
5111 return (1);
|
|
5112
|
|
5113 for (i = 0; i < md_nunits; i++) {
|
|
5114 mnum = MD_MKMIN(setno, i);
|
|
5115 if ((ui = MDI_UNIT(mnum)) == NULL)
|
|
5116 continue;
|
|
5117 if (ui->ui_opsindex != mirror_md_ops.md_selfindex)
|
|
5118 continue;
|
|
5119 reset_mirror((mm_unit_t *)MD_UNIT(mnum), mnum, 0);
|
|
5120
|
|
5121 /* Set a flag if there is at least one mirror metadevice. */
|
|
5122 reset_mirror_flag = 1;
|
|
5123 }
|
|
5124
|
|
5125 /*
|
|
5126 * Only wait for the global dr_timeout to finish
|
|
5127 * - if there are mirror metadevices in this diskset or
|
|
5128 * - if this is the local set since an unload of the md_mirror
|
|
5129 * driver could follow a successful mirror halt in the local set.
|
|
5130 */
|
|
5131 if ((reset_mirror_flag != 0) || (setno == MD_LOCAL_SET)) {
|
|
5132 while ((mirror_md_ops.md_head == NULL) &&
|
|
5133 (mirror_timeout.dr_timeout_id != 0))
|
|
5134 delay(md_hz);
|
|
5135 }
|
|
5136
|
|
5137 return (0);
|
|
5138 }
|
|
5139
|
|
5140 /*ARGSUSED3*/
|
|
5141 static int
|
|
5142 mirror_open(dev_t *dev, int flag, int otyp, cred_t *cred_p, int md_oflags)
|
|
5143 {
|
|
5144 IOLOCK lock;
|
|
5145
|
|
5146 if (md_oflags & MD_OFLG_FROMIOCTL) {
|
|
5147 /*
|
|
5148 * This indicates that the caller is an ioctl service routine.
|
|
5149 * In this case we initialise our stack-based IOLOCK and pass
|
|
5150 * this into the internal open routine. This allows multi-owner
|
|
5151 * metadevices to avoid deadlocking if an error is encountered
|
|
5152 * during the open() attempt. The failure case is:
|
|
5153 * s-p -> mirror -> s-p (with error). Attempting to metaclear
|
|
5154 * this configuration would deadlock as the mirror code has to
|
|
5155 * send a state-update to the other nodes when it detects the
|
|
5156 * failure of the underlying submirror with an errored soft-part
|
|
5157 * on it. As there is a class1 message in progress (metaclear)
|
|
5158 * set_sm_comp_state() cannot send another class1 message;
|
|
5159 * instead we do not send a state_update message as the
|
|
5160 * metaclear is distributed and the failed submirror will be
|
|
5161 * cleared from the configuration by the metaclear.
|
|
5162 */
|
|
5163 IOLOCK_INIT(&lock);
|
|
5164 return (mirror_internal_open(getminor(*dev), flag, otyp,
|
|
5165 md_oflags, &lock));
|
|
5166 } else {
|
|
5167 return (mirror_internal_open(getminor(*dev), flag, otyp,
|
|
5168 md_oflags, (IOLOCK *)NULL));
|
|
5169 }
|
|
5170 }
|
|
5171
|
|
5172
|
|
5173 /*ARGSUSED1*/
|
|
5174 static int
|
|
5175 mirror_close(dev_t dev, int flag, int otyp, cred_t *cred_p, int md_cflags)
|
|
5176 {
|
|
5177 return (mirror_internal_close(getminor(dev), otyp, md_cflags,
|
|
5178 (IOLOCK *)NULL));
|
|
5179 }
|
|
5180
|
|
5181
|
|
5182 /*
|
|
5183 * This routine dumps memory to the disk. It assumes that the memory has
|
|
5184 * already been mapped into mainbus space. It is called at disk interrupt
|
|
5185 * priority when the system is in trouble.
|
|
5186 *
|
|
5187 */
|
|
5188 static int
|
|
5189 mirror_dump(dev_t dev, caddr_t addr, daddr_t blkno, int nblk)
|
|
5190 {
|
|
5191 mm_unit_t *un;
|
|
5192 dev_t mapdev;
|
|
5193 int result;
|
|
5194 int smi;
|
|
5195 int any_succeed = 0;
|
|
5196 int save_result = 0;
|
|
5197
|
|
5198 /*
|
|
5199 * Don't need to grab the unit lock.
|
|
5200 * Cause nothing else is suppose to be happenning.
|
|
5201 * Also dump is not suppose to sleep.
|
|
5202 */
|
|
5203 un = (mm_unit_t *)MD_UNIT(getminor(dev));
|
|
5204
|
|
5205 if ((diskaddr_t)blkno >= un->c.un_total_blocks)
|
|
5206 return (EINVAL);
|
|
5207
|
|
5208 if ((diskaddr_t)blkno + nblk > un->c.un_total_blocks)
|
|
5209 return (EINVAL);
|
|
5210
|
|
5211 for (smi = 0; smi < NMIRROR; smi++) {
|
|
5212 if (!SUBMIRROR_IS_WRITEABLE(un, smi))
|
|
5213 continue;
|
|
5214 mapdev = md_dev64_to_dev(un->un_sm[smi].sm_dev);
|
|
5215 result = bdev_dump(mapdev, addr, blkno, nblk);
|
|
5216 if (result)
|
|
5217 save_result = result;
|
|
5218
|
|
5219 if (result == 0)
|
|
5220 any_succeed++;
|
|
5221 }
|
|
5222
|
|
5223 if (any_succeed)
|
|
5224 return (0);
|
|
5225
|
|
5226 return (save_result);
|
|
5227 }
|
|
5228
|
|
5229 /*
|
|
5230 * NAME: mirror_probe_dev
|
|
5231 *
|
|
5232 * DESCRITPION: force opens every component of a mirror.
|
|
5233 *
|
|
5234 * On entry the unit writerlock is held
|
|
5235 */
|
|
5236 static int
|
|
5237 mirror_probe_dev(mdi_unit_t *ui, minor_t mnum)
|
|
5238 {
|
|
5239 int i;
|
|
5240 int smi;
|
|
5241 int ci;
|
|
5242 mm_unit_t *un;
|
|
5243 int md_devopen = 0;
|
|
5244 set_t setno;
|
|
5245 int sm_cnt;
|
|
5246 int sm_unavail_cnt;
|
|
5247
|
|
5248 if (md_unit_isopen(ui))
|
|
5249 md_devopen++;
|
|
5250
|
|
5251 un = MD_UNIT(mnum);
|
|
5252 setno = MD_UN2SET(un);
|
|
5253
|
|
5254 sm_cnt = 0;
|
|
5255 sm_unavail_cnt = 0;
|
|
5256 for (i = 0; i < NMIRROR; i++) {
|
|
5257 md_dev64_t tmpdev;
|
|
5258 mdi_unit_t *sm_ui;
|
|
5259
|
|
5260 if (!SMS_BY_INDEX_IS(un, i, SMS_INUSE)) {
|
|
5261 continue;
|
|
5262 }
|
|
5263
|
|
5264 sm_cnt++;
|
|
5265 tmpdev = un->un_sm[i].sm_dev;
|
|
5266 (void) md_layered_open(mnum, &tmpdev,
|
|
5267 MD_OFLG_CONT_ERRS | MD_OFLG_PROBEDEV);
|
|
5268 un->un_sm[i].sm_dev = tmpdev;
|
|
5269
|
|
5270 sm_ui = MDI_UNIT(getminor(md_dev64_to_dev(tmpdev)));
|
|
5271
|
|
5272 /*
|
|
5273 * Logic similar to that in mirror_open_all_devs. We set or
|
|
5274 * clear the submirror Unavailable bit.
|
|
5275 */
|
|
5276 (void) md_unit_writerlock(sm_ui);
|
|
5277 if (submirror_unavailable(un, i, 1)) {
|
|
5278 sm_ui->ui_tstate |= MD_INACCESSIBLE;
|
|
5279 sm_unavail_cnt++;
|
|
5280 } else {
|
|
5281 sm_ui->ui_tstate &= ~MD_INACCESSIBLE;
|
|
5282 }
|
|
5283 md_unit_writerexit(sm_ui);
|
|
5284 }
|
|
5285
|
|
5286 /*
|
|
5287 * If all of the submirrors are unavailable, the mirror is also
|
|
5288 * unavailable.
|
|
5289 */
|
|
5290 if (sm_cnt == sm_unavail_cnt) {
|
|
5291 ui->ui_tstate |= MD_INACCESSIBLE;
|
|
5292 } else {
|
|
5293 ui->ui_tstate &= ~MD_INACCESSIBLE;
|
|
5294 }
|
|
5295
|
|
5296 /*
|
|
5297 * Start checking from probe failures. If failures occur we
|
|
5298 * set the appropriate erred state only if the metadevice is in
|
|
5299 * use. This is specifically to prevent unnecessary resyncs.
|
|
5300 * For instance if the disks were accidentally disconnected when
|
|
5301 * the system booted up then until the metadevice is accessed
|
|
5302 * (like file system mount) the user can shutdown, recable and
|
|
5303 * reboot w/o incurring a potentially huge resync.
|
|
5304 */
|
|
5305
|
|
5306 smi = 0;
|
|
5307 ci = 0;
|
|
5308 while (mirror_geterror(un, &smi, &ci, 1, 1) != 0) {
|
|
5309
|
|
5310 if (mirror_other_sources(un, smi, ci, 0) == 1) {
|
|
5311 /*
|
|
5312 * Note that for a MN set, there is no need to call
|
|
5313 * SE_NOTIFY as that is done when processing the
|
|
5314 * state change
|
|
5315 */
|
|
5316 if (md_devopen) {
|
|
5317 /*
|
|
5318 * Never called from ioctl context,
|
|
5319 * so (IOLOCK *)NULL
|
|
5320 */
|
|
5321 set_sm_comp_state(un, smi, ci, CS_LAST_ERRED,
|
|
5322 0, MD_STATE_XMIT, (IOLOCK *)NULL);
|
|
5323 if (!MD_MNSET_SETNO(setno)) {
|
|
5324 SE_NOTIFY(EC_SVM_STATE,
|
|
5325 ESC_SVM_LASTERRED,
|
|
5326 SVM_TAG_METADEVICE, setno,
|
|
5327 MD_SID(un));
|
|
5328 }
|
|
5329 continue;
|
|
5330 } else {
|
|
5331 (void) mirror_close_all_devs(un,
|
|
5332 MD_OFLG_PROBEDEV);
|
|
5333 if (!MD_MNSET_SETNO(setno)) {
|
|
5334 SE_NOTIFY(EC_SVM_STATE,
|
|
5335 ESC_SVM_OPEN_FAIL,
|
|
5336 SVM_TAG_METADEVICE, setno,
|
|
5337 MD_SID(un));
|
|
5338 }
|
|
5339 mirror_openfail_console_info(un, smi, ci);
|
|
5340 return (ENXIO);
|
|
5341 }
|
|
5342 }
|
|
5343
|
|
5344 /*
|
|
5345 * Note that for a MN set, there is no need to call
|
|
5346 * SE_NOTIFY as that is done when processing the
|
|
5347 * state change
|
|
5348 */
|
|
5349 if (md_devopen) {
|
|
5350 /* Never called from ioctl context, so (IOLOCK *)NULL */
|
|
5351 set_sm_comp_state(un, smi, ci, CS_ERRED, 0,
|
|
5352 MD_STATE_XMIT, (IOLOCK *)NULL);
|
|
5353 if (!MD_MNSET_SETNO(setno)) {
|
|
5354 SE_NOTIFY(EC_SVM_STATE, ESC_SVM_ERRED,
|
|
5355 SVM_TAG_METADEVICE, setno,
|
|
5356 MD_SID(un));
|
|
5357 }
|
|
5358 }
|
|
5359 mirror_openfail_console_info(un, smi, ci);
|
|
5360 ci++;
|
|
5361 }
|
|
5362
|
|
5363 if (MD_MNSET_SETNO(setno)) {
|
|
5364 send_poke_hotspares(setno);
|
|
5365 } else {
|
|
5366 (void) poke_hotspares();
|
|
5367 }
|
|
5368 (void) mirror_close_all_devs(un, MD_OFLG_PROBEDEV);
|
|
5369
|
|
5370 return (0);
|
|
5371 }
|
|
5372
|
|
5373
|
|
5374 static int
|
|
5375 mirror_imp_set(
|
|
5376 set_t setno
|
|
5377 )
|
|
5378 {
|
|
5379
|
|
5380 mddb_recid_t recid;
|
|
5381 int gotsomething, i;
|
|
5382 mddb_type_t typ1;
|
|
5383 mddb_de_ic_t *dep;
|
|
5384 mddb_rb32_t *rbp;
|
|
5385 mm_unit32_od_t *un32;
|
|
5386 mm_unit_t *un64;
|
|
5387 minor_t *self_id; /* minor needs to be updated */
|
|
5388 md_parent_t *parent_id; /* parent needs to be updated */
|
|
5389 mddb_recid_t *record_id; /* record id needs to be updated */
|
|
5390 mddb_recid_t *optrec_id;
|
|
5391 md_dev64_t tmpdev;
|
|
5392
|
|
5393
|
|
5394 gotsomething = 0;
|
|
5395
|
|
5396 typ1 = (mddb_type_t)md_getshared_key(setno,
|
|
5397 mirror_md_ops.md_driver.md_drivername);
|
|
5398 recid = mddb_makerecid(setno, 0);
|
|
5399
|
|
5400 while ((recid = mddb_getnextrec(recid, typ1, MIRROR_REC)) > 0) {
|
|
5401 if (mddb_getrecprivate(recid) & MD_PRV_GOTIT)
|
|
5402 continue;
|
|
5403
|
|
5404 dep = mddb_getrecdep(recid);
|
|
5405 rbp = dep->de_rb;
|
|
5406
|
|
5407 if (rbp->rb_revision == MDDB_REV_RB) {
|
|
5408 /*
|
|
5409 * Small device
|
|
5410 */
|
|
5411 un32 = (mm_unit32_od_t *)mddb_getrecaddr(recid);
|
|
5412 self_id = &(un32->c.un_self_id);
|
|
5413 parent_id = &(un32->c.un_parent);
|
|
5414 record_id = &(un32->c.un_record_id);
|
|
5415 optrec_id = &(un32->un_rr_dirty_recid);
|
|
5416
|
|
5417 for (i = 0; i < un32->un_nsm; i++) {
|
|
5418 tmpdev = md_expldev(un32->un_sm[i].sm_dev);
|
|
5419 un32->un_sm[i].sm_dev = md_cmpldev
|
|
5420 (md_makedevice(md_major, MD_MKMIN(setno,
|
|
5421 MD_MIN2UNIT(md_getminor(tmpdev)))));
|
|
5422
|
|
5423 if (!md_update_minor(setno, mddb_getsidenum
|
|
5424 (setno), un32->un_sm[i].sm_key))
|
|
5425 goto out;
|
|
5426 }
|
|
5427 } else {
|
|
5428 un64 = (mm_unit_t *)mddb_getrecaddr(recid);
|
|
5429 self_id = &(un64->c.un_self_id);
|
|
5430 parent_id = &(un64->c.un_parent);
|
|
5431 record_id = &(un64->c.un_record_id);
|
|
5432 optrec_id = &(un64->un_rr_dirty_recid);
|
|
5433
|
|
5434 for (i = 0; i < un64->un_nsm; i++) {
|
|
5435 tmpdev = un64->un_sm[i].sm_dev;
|
|
5436 un64->un_sm[i].sm_dev = md_makedevice
|
|
5437 (md_major, MD_MKMIN(setno, MD_MIN2UNIT
|
|
5438 (md_getminor(tmpdev))));
|
|
5439
|
|
5440 if (!md_update_minor(setno, mddb_getsidenum
|
|
5441 (setno), un64->un_sm[i].sm_key))
|
|
5442 goto out;
|
|
5443 }
|
|
5444 }
|
|
5445
|
|
5446 /*
|
|
5447 * Update unit with the imported setno
|
|
5448 *
|
|
5449 */
|
|
5450 mddb_setrecprivate(recid, MD_PRV_GOTIT);
|
|
5451
|
|
5452 *self_id = MD_MKMIN(setno, MD_MIN2UNIT(*self_id));
|
|
5453 if (*parent_id != MD_NO_PARENT)
|
|
5454 *parent_id = MD_MKMIN(setno, MD_MIN2UNIT(*parent_id));
|
|
5455 *record_id = MAKERECID(setno, DBID(*record_id));
|
|
5456 *optrec_id = MAKERECID(setno, DBID(*optrec_id));
|
|
5457
|
|
5458 gotsomething = 1;
|
|
5459 }
|
|
5460
|
|
5461 out:
|
|
5462 return (gotsomething);
|
|
5463 }
|
|
5464
|
|
5465 /*
|
|
5466 * NAME: mirror_check_offline
|
|
5467 *
|
|
5468 * DESCRIPTION: return offline_status = 1 if any submirrors are offline
|
|
5469 *
|
|
5470 * Called from ioctl, so access to MD_UN_OFFLINE_SM in un_status is
|
|
5471 * protected by the global ioctl lock as it is only set by the MD_IOCOFFLINE
|
|
5472 * ioctl.
|
|
5473 */
|
|
5474 int
|
|
5475 mirror_check_offline(md_dev64_t dev, int *offline_status)
|
|
5476 {
|
|
5477 mm_unit_t *un;
|
|
5478 md_error_t mde = mdnullerror;
|
|
5479
|
|
5480 if ((un = mirror_getun(getminor(dev), &mde, NO_LOCK, NULL)) == NULL)
|
|
5481 return (EINVAL);
|
|
5482 *offline_status = 0;
|
|
5483 if (un->c.un_status & MD_UN_OFFLINE_SM)
|
|
5484 *offline_status = 1;
|
|
5485 return (0);
|
|
5486 }
|
|
5487
|
|
5488 /*
|
|
5489 * NAME: mirror_inc_abr_count
|
|
5490 *
|
|
5491 * DESCRIPTION: increment the count of layered soft parts with ABR set
|
|
5492 *
|
|
5493 * Called from ioctl, so access to un_abr_count is protected by the global
|
|
5494 * ioctl lock. It is only referenced in the MD_IOCOFFLINE ioctl.
|
|
5495 */
|
|
5496 int
|
|
5497 mirror_inc_abr_count(md_dev64_t dev)
|
|
5498 {
|
|
5499 mm_unit_t *un;
|
|
5500 md_error_t mde = mdnullerror;
|
|
5501
|
|
5502 if ((un = mirror_getun(getminor(dev), &mde, NO_LOCK, NULL)) == NULL)
|
|
5503 return (EINVAL);
|
|
5504 un->un_abr_count++;
|
|
5505 return (0);
|
|
5506 }
|
|
5507
|
|
5508 /*
|
|
5509 * NAME: mirror_dec_abr_count
|
|
5510 *
|
|
5511 * DESCRIPTION: decrement the count of layered soft parts with ABR set
|
|
5512 *
|
|
5513 * Called from ioctl, so access to un_abr_count is protected by the global
|
|
5514 * ioctl lock. It is only referenced in the MD_IOCOFFLINE ioctl.
|
|
5515 */
|
|
5516 int
|
|
5517 mirror_dec_abr_count(md_dev64_t dev)
|
|
5518 {
|
|
5519 mm_unit_t *un;
|
|
5520 md_error_t mde = mdnullerror;
|
|
5521
|
|
5522 if ((un = mirror_getun(getminor(dev), &mde, NO_LOCK, NULL)) == NULL)
|
|
5523 return (EINVAL);
|
|
5524 un->un_abr_count--;
|
|
5525 return (0);
|
|
5526 }
|
|
5527
|
|
5528 static md_named_services_t mirror_named_services[] = {
|
|
5529 {(intptr_t (*)()) poke_hotspares, "poke hotspares" },
|
|
5530 {(intptr_t (*)()) mirror_rename_listkids, MDRNM_LIST_URKIDS },
|
|
5531 {mirror_rename_check, MDRNM_CHECK },
|
|
5532 {(intptr_t (*)()) mirror_renexch_update_kids, MDRNM_UPDATE_KIDS },
|
|
5533 {(intptr_t (*)()) mirror_exchange_parent_update_to,
|
|
5534 MDRNM_PARENT_UPDATE_TO},
|
|
5535 {(intptr_t (*)()) mirror_exchange_self_update_from_down,
|
|
5536 MDRNM_SELF_UPDATE_FROM_DOWN },
|
|
5537 {(intptr_t (*)())mirror_probe_dev, "probe open test" },
|
|
5538 {(intptr_t (*)())mirror_check_offline, MD_CHECK_OFFLINE },
|
|
5539 {(intptr_t (*)())mirror_inc_abr_count, MD_INC_ABR_COUNT },
|
|
5540 {(intptr_t (*)())mirror_dec_abr_count, MD_DEC_ABR_COUNT },
|
|
5541 { NULL, 0 }
|
|
5542 };
|
|
5543
|
|
5544 md_ops_t mirror_md_ops = {
|
|
5545 mirror_open, /* open */
|
|
5546 mirror_close, /* close */
|
|
5547 md_mirror_strategy, /* strategy */
|
|
5548 NULL, /* print */
|
|
5549 mirror_dump, /* dump */
|
|
5550 NULL, /* read */
|
|
5551 NULL, /* write */
|
|
5552 md_mirror_ioctl, /* mirror_ioctl, */
|
|
5553 mirror_snarf, /* mirror_snarf */
|
|
5554 mirror_halt, /* mirror_halt */
|
|
5555 NULL, /* aread */
|
|
5556 NULL, /* awrite */
|
|
5557 mirror_imp_set, /* import set */
|
|
5558 mirror_named_services
|
|
5559 };
|
|
5560
|
|
5561 /* module specific initilization */
|
|
5562 static void
|
|
5563 init_init()
|
|
5564 {
|
|
5565 md_mirror_mcs_buf_off = sizeof (md_mcs_t) - sizeof (buf_t);
|
|
5566
|
|
5567 /* Initialize the parent and child save memory pools */
|
|
5568 mirror_parent_cache = kmem_cache_create("md_mirror_parent",
|
|
5569 sizeof (md_mps_t), 0, mirror_parent_constructor,
|
|
5570 mirror_parent_destructor, mirror_run_queue, NULL, NULL,
|
|
5571 0);
|
|
5572
|
|
5573 mirror_child_cache = kmem_cache_create("md_mirror_child",
|
|
5574 sizeof (md_mcs_t) - sizeof (buf_t) + biosize(), 0,
|
|
5575 mirror_child_constructor, mirror_child_destructor,
|
|
5576 mirror_run_queue, NULL, NULL, 0);
|
|
5577
|
|
5578 /*
|
|
5579 * Insure wowbuf_size is a multiple of DEV_BSIZE,
|
|
5580 * then initialize wowbuf memory pool.
|
|
5581 */
|
|
5582 md_wowbuf_size = roundup(md_wowbuf_size, DEV_BSIZE);
|
|
5583 if (md_wowbuf_size <= 0)
|
|
5584 md_wowbuf_size = 2 * DEV_BSIZE;
|
|
5585 if (md_wowbuf_size > (32 * DEV_BSIZE))
|
|
5586 md_wowbuf_size = (32 * DEV_BSIZE);
|
|
5587
|
|
5588 md_wowblk_size = md_wowbuf_size + sizeof (wowhdr_t);
|
|
5589 mirror_wowblk_cache = kmem_cache_create("md_mirror_wow",
|
|
5590 md_wowblk_size, 0, NULL, NULL, NULL, NULL, NULL, 0);
|
|
5591
|
|
5592 mutex_init(&mirror_timeout.dr_mx, NULL, MUTEX_DEFAULT, NULL);
|
|
5593 mutex_init(&hotspare_request.dr_mx, NULL, MUTEX_DEFAULT, NULL);
|
|
5594
|
|
5595 mutex_init(&non_ff_drv_mutex, NULL, MUTEX_DEFAULT, NULL);
|
|
5596 }
|
|
5597
|
|
5598 /* module specific uninitilization (undo init_init()) */
|
|
5599 static void
|
|
5600 fini_uninit()
|
|
5601 {
|
|
5602 kmem_cache_destroy(mirror_parent_cache);
|
|
5603 kmem_cache_destroy(mirror_child_cache);
|
|
5604 kmem_cache_destroy(mirror_wowblk_cache);
|
|
5605 mirror_parent_cache = mirror_child_cache =
|
|
5606 mirror_wowblk_cache = NULL;
|
|
5607
|
|
5608 mutex_destroy(&mirror_timeout.dr_mx);
|
|
5609 mutex_destroy(&hotspare_request.dr_mx);
|
|
5610 mutex_destroy(&non_ff_drv_mutex);
|
|
5611 }
|
|
5612
|
|
5613 /* define the module linkage */
|
|
5614 MD_PLUGIN_MISC_MODULE("mirrors module %I%", init_init(), fini_uninit())
|