2 * Copyright (c) 1997, 1998, 1999
3 * Nan Yang Computer Services Limited. All rights reserved.
5 * Parts copyright (c) 1997, 1998 Cybernet Corporation, NetMAX project.
7 * Written by Greg Lehey
9 * This software is distributed under the so-called ``Berkeley
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by Nan Yang Computer
24 * 4. Neither the name of the Company nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * This software is provided ``as is'', and any express or implied
29 * warranties, including, but not limited to, the implied warranties of
30 * merchantability and fitness for a particular purpose are disclaimed.
31 * In no event shall the company or contributors be liable for any
32 * direct, indirect, incidental, special, exemplary, or consequential
33 * damages (including, but not limited to, procurement of substitute
34 * goods or services; loss of use, data, or profits; or business
35 * interruption) however caused and on any theory of liability, whether
36 * in contract, strict liability, or tort (including negligence or
37 * otherwise) arising in any way out of the use of this software, even if
38 * advised of the possibility of such damage.
40 * $Id: vinumrevive.c,v 1.14 2000/12/21 01:55:11 grog Exp grog $
41 * $FreeBSD: src/sys/dev/vinum/vinumrevive.c,v 1.22.2.5 2001/03/13 02:59:43 grog Exp $
42 * $DragonFly: src/sys/dev/raid/vinum/vinumrevive.c,v 1.14 2006/12/20 18:14:40 dillon Exp $
49 * Revive a block of a subdisk. Return an error
50 * indication. EAGAIN means successful copy, but
51 * that more blocks remain to be copied. EINVAL
52 * means that the subdisk isn't associated with a
53 * plex (which means a programming error if we get
54 * here at all; FIXME).
58 revive_block(int sdno)
66 int size; /* size of revive block, bytes */
67 daddr_t plexblkno; /* lblkno in plex */
68 int psd; /* parity subdisk number */
69 u_int64_t stripe; /* stripe number */
70 int paritysd = 0; /* set if this is the parity stripe */
71 struct rangelock *lock; /* for locking */
72 daddr_t stripeoffset; /* offset in stripe */
74 plexblkno = 0; /* to keep the compiler happy */
77 if (sd->plexno < 0) /* no plex? */
79 plex = &PLEX[sd->plexno]; /* point to plex */
81 vol = &VOL[plex->volno];
85 if ((sd->revive_blocksize == 0) /* no block size */
86 ||(sd->revive_blocksize & ((1 << DEV_BSHIFT) - 1))) /* or invalid block size */
87 sd->revive_blocksize = DEFAULT_REVIVE_BLOCKSIZE;
88 else if (sd->revive_blocksize > MAX_REVIVE_BLOCKSIZE)
89 sd->revive_blocksize = MAX_REVIVE_BLOCKSIZE;
90 size = min(sd->revive_blocksize >> DEV_BSHIFT, sd->sectors - sd->revived) << DEV_BSHIFT;
91 sd->reviver = curproc->p_pid; /* note who last had a bash at it */
93 /* Now decide where to read from */
94 switch (plex->organization) {
96 plexblkno = sd->revived + sd->plexoffset; /* corresponding address in plex */
100 stripeoffset = sd->revived % plex->stripesize; /* offset from beginning of stripe */
101 if (stripeoffset + (size >> DEV_BSHIFT) > plex->stripesize)
102 size = (plex->stripesize - stripeoffset) << DEV_BSHIFT;
103 plexblkno = sd->plexoffset /* base */
104 + (sd->revived - stripeoffset) * plex->subdisks /* offset to beginning of stripe */
105 + stripeoffset; /* offset from beginning of stripe */
110 stripeoffset = sd->revived % plex->stripesize; /* offset from beginning of stripe */
111 plexblkno = sd->plexoffset /* base */
112 + (sd->revived - stripeoffset) * (plex->subdisks - 1) /* offset to beginning of stripe */
113 +stripeoffset; /* offset from beginning of stripe */
114 stripe = (sd->revived / plex->stripesize); /* stripe number */
116 /* Make sure we don't go beyond the end of the band. */
117 size = min(size, (plex->stripesize - stripeoffset) << DEV_BSHIFT);
118 if (plex->organization == plex_raid4)
119 psd = plex->subdisks - 1; /* parity subdisk for this stripe */
121 psd = plex->subdisks - 1 - stripe % plex->subdisks; /* parity subdisk for this stripe */
122 paritysd = plex->sdnos[psd] == sdno; /* note if it's the parity subdisk */
125 * Now adjust for the strangenesses
126 * in RAID-4 and RAID-5 striping.
128 if (sd->plexsdno > psd) /* beyond the parity stripe, */
129 plexblkno -= plex->stripesize; /* one stripe less */
131 plexblkno -= plex->stripesize * sd->plexsdno; /* go back to the beginning of the band */
134 case plex_disorg: /* to keep the compiler happy */
138 if (paritysd) { /* we're reviving a parity block, */
139 bp = parityrebuild(plex, sd->revived, size, rebuildparity, &lock, NULL); /* do the grunt work */
140 if (bp == NULL) /* no buffer space */
141 return ENOMEM; /* chicken out */
142 } else { /* data block */
143 bp = getpbuf(&vinum_conf.physbufs); /* Get a buffer */
144 bp->b_data = Malloc(size);
147 * Amount to transfer: block size, unless it
148 * would overlap the end.
151 bp->b_resid = bp->b_bcount;
152 bp->b_bio1.bio_offset = (off_t)plexblkno << DEV_BSHIFT; /* start here */
153 if (isstriped(plex)) /* we need to lock striped plexes */
154 lock = lockrange(plexblkno << DEV_BSHIFT, bp, plex); /* lock it */
155 if (vol != NULL) /* it's part of a volume, */
157 * First, read the data from the volume. We
158 * don't care which plex, that's bre's job.
160 dev = VINUMDEV(plex->volno, 0, 0, VINUM_VOLUME_TYPE); /* create the device number */
161 else /* it's an unattached plex */
162 dev = VINUM_PLEX(sd->plexno); /* create the device number */
164 bp->b_cmd = BUF_CMD_READ;
165 vinumstart(dev, &bp->b_bio1, 1);
169 if (bp->b_flags & B_ERROR)
172 /* Now write to the subdisk */
174 dev = VINUM_SD(sdno); /* create the device number */
175 bp->b_flags |= B_ORDERED; /* and make this an ordered write */
176 bp->b_cmd = BUF_CMD_WRITE;
177 bp->b_resid = bp->b_bcount;
178 bp->b_bio1.bio_offset = (off_t)sd->revived << DEV_BSHIFT; /* write it to here */
179 bp->b_bio1.bio_driver_info = dev;
180 sdio(&bp->b_bio1); /* perform the I/O */
182 if (bp->b_flags & B_ERROR)
185 sd->revived += bp->b_bcount >> DEV_BSHIFT; /* moved this much further down */
186 if (sd->revived >= sd->sectors) { /* finished */
188 set_sd_state(sdno, sd_up, setstate_force); /* bring the sd up */
189 log(LOG_INFO, "vinum: %s is %s\n", sd->name, sd_state(sd->state));
190 save_config(); /* and save the updated configuration */
191 error = 0; /* we're done */
194 if (lock) /* we took a lock, */
195 unlockrange(sd->plexno, lock); /* give it back */
196 while (sd->waitlist) { /* we have waiting requests */
198 struct request *rq = sd->waitlist;
201 if (debug & DEBUG_REVIVECONFLICT) {
202 dev = rq->bio->bio_driver_info;
204 "Relaunch revive conflict sd %d: %p\n%s dev %d.%d, offset 0x%llx, length %d\n",
207 (rq->bio->bio_buf->b_cmd == BUF_CMD_READ) ? "Read" : "Write",
211 rq->bio->bio_buf->b_bcount);
214 launch_requests(sd->waitlist, 1); /* do them now */
215 sd->waitlist = sd->waitlist->next; /* and move on to the next */
219 relpbuf(bp, &vinum_conf.physbufs);
224 * Check or rebuild the parity blocks of a RAID-4
227 * The variables plex->checkblock and
228 * plex->rebuildblock represent the
229 * subdisk-relative address of the stripe we're
230 * looking at, not the plex-relative address. We
231 * store it in the plex and not as a local
232 * variable because this function could be
233 * stopped, and we don't want to repeat the part
234 * we've already done. This is also the reason
235 * why we don't initialize it here except at the
236 * end. It gets initialized with the plex on
239 * Each call to this function processes at most
240 * one stripe. We can't loop in this function,
241 * because we're unstoppable, so we have to be
242 * called repeatedly from userland.
245 parityops(struct vinum_ioctl_msg *data)
249 int size; /* I/O transfer size, bytes */
250 int stripe; /* stripe number in plex */
251 int psd; /* parity subdisk number */
252 struct rangelock *lock; /* lock on stripe */
253 struct _ioctl_reply *reply;
254 off_t pstripe; /* pointer to our stripe counter */
256 off_t errorloc; /* offset of parity error */
257 enum parityop op; /* operation to perform */
259 plexno = data->index;
262 reply = (struct _ioctl_reply *) data;
263 reply->error = EAGAIN; /* expect to repeat this call */
264 plex = &PLEX[plexno];
265 if (!isparity(plex)) { /* not RAID-4 or RAID-5 */
266 reply->error = EINVAL;
268 } else if (plex->state < plex_flaky) {
270 strcpy(reply->msg, "Plex is not completely accessible\n");
273 pstripe = data->offset;
274 stripe = pstripe / plex->stripesize; /* stripe number */
275 psd = plex->subdisks - 1 - stripe % plex->subdisks; /* parity subdisk for this stripe */
276 size = min(DEFAULT_REVIVE_BLOCKSIZE, /* one block at a time */
277 plex->stripesize << DEV_BSHIFT);
279 pbp = parityrebuild(plex, pstripe, size, op, &lock, &errorloc); /* do the grunt work */
280 if (pbp == NULL) { /* no buffer space */
281 reply->error = ENOMEM;
282 return; /* chicken out */
285 * Now we have a result in the data buffer of
286 * the parity buffer header, which we have kept.
287 * Decide what to do with it.
289 reply->msg[0] = '\0'; /* until shown otherwise */
290 if ((pbp->b_flags & B_ERROR) == 0) { /* no error */
291 if ((op == rebuildparity)
292 || (op == rebuildandcheckparity)) {
293 pbp->b_cmd = BUF_CMD_WRITE;
294 pbp->b_resid = pbp->b_bcount;
295 sdio(&pbp->b_bio1); /* write the parity block */
298 if (((op == checkparity)
299 || (op == rebuildandcheckparity))
300 && (errorloc != -1)) {
301 if (op == checkparity)
304 "Parity incorrect at offset 0x%llx\n",
307 if (reply->error == EAGAIN) { /* still OK, */
308 plex->checkblock = pstripe + (pbp->b_bcount >> DEV_BSHIFT); /* moved this much further down */
309 if (plex->checkblock >= SD[plex->sdnos[0]].sectors) { /* finished */
310 plex->checkblock = 0;
315 if (pbp->b_flags & B_ERROR)
316 reply->error = pbp->b_error;
318 relpbuf(pbp, &vinum_conf.physbufs);
319 unlockrange(plexno, lock);
323 * Rebuild a parity stripe. Return pointer to
324 * parity bp. On return,
326 * 1. The band is locked. The caller must unlock
327 * the band and release the buffer header.
329 * 2. All buffer headers except php have been
330 * released. The caller must release pbp.
332 * 3. For checkparity and rebuildandcheckparity,
333 * the parity is compared with the current
334 * parity block. If it's different, the
335 * offset of the error is returned to
336 * errorloc. The caller can set the value of
337 * the pointer to NULL if this is called for
340 * pstripe is the subdisk-relative base address of
341 * the data to be reconstructed, size is the size
342 * of the transfer in bytes.
345 parityrebuild(struct plex *plex,
349 struct rangelock **lockp,
354 u_int64_t stripe; /* stripe number */
355 int *parity_buf; /* buffer address for current parity block */
356 int *newparity_buf; /* and for new parity block */
357 int mysize; /* I/O transfer size for this transfer */
358 int isize; /* mysize in ints */
360 int psd; /* parity subdisk number */
361 int newpsd; /* and "subdisk number" of new parity */
362 struct buf **bpp; /* pointers to our bps */
363 struct buf *pbp; /* buffer header for parity stripe */
365 int bufcount; /* number of buffers we need */
367 stripe = pstripe / plex->stripesize; /* stripe number */
368 psd = plex->subdisks - 1 - stripe % plex->subdisks; /* parity subdisk for this stripe */
369 parity_buf = NULL; /* to keep the compiler happy */
373 * It's possible that the default transfer size
374 * we chose is not a factor of the stripe size.
375 * We *must* limit this operation to a single
376 * stripe, at least for RAID-5 rebuild, since
377 * the parity subdisk changes between stripes,
378 * so in this case we need to perform a short
379 * transfer. Set variable mysize to reflect
382 mysize = min(size, (plex->stripesize * (stripe + 1) - pstripe) << DEV_BSHIFT);
383 isize = mysize / (sizeof(int)); /* number of ints in the buffer */
384 bufcount = plex->subdisks + 1; /* sd buffers plus result buffer */
385 newpsd = plex->subdisks;
386 bpp = (struct buf **) Malloc(bufcount * sizeof(struct buf *)); /* array of pointers to bps */
388 /* First, build requests for all subdisks */
389 for (sdno = 0; sdno < bufcount; sdno++) { /* for each subdisk */
390 if ((sdno != psd) || (op != rebuildparity)) {
391 /* Get a buffer header and initialize it. */
392 bpp[sdno] = getpbuf(&vinum_conf.physbufs); /* Get a buffer */
393 bpp[sdno]->b_data = Malloc(mysize);
395 parity_buf = (int *) bpp[sdno]->b_data;
396 if (sdno == newpsd) /* the new one? */
397 bpp[sdno]->b_bio1.bio_driver_info = VINUM_SD(plex->sdnos[psd]); /* write back to the parity SD */
399 bpp[sdno]->b_bio1.bio_driver_info = VINUM_SD(plex->sdnos[sdno]); /* device number */
400 bpp[sdno]->b_cmd = BUF_CMD_READ; /* either way, read it */
401 bpp[sdno]->b_bcount = mysize;
402 bpp[sdno]->b_resid = bpp[sdno]->b_bcount;
403 bpp[sdno]->b_bio1.bio_offset = (off_t)pstripe << DEV_BSHIFT; /* transfer from here */
407 /* Initialize result buffer */
409 newparity_buf = (int *) bpp[newpsd]->b_data;
410 bzero(newparity_buf, mysize);
413 * Now lock the stripe with the first non-parity
416 *lockp = lockrange(pstripe * plex->stripesize * (plex->subdisks - 1),
421 * Then issue requests for all subdisks in
422 * parallel. Don't transfer the parity stripe
423 * if we're rebuilding parity, unless we also
426 for (sdno = 0; sdno < plex->subdisks; sdno++) { /* for each real subdisk */
427 if ((sdno != psd) || (op != rebuildparity)) {
428 sdio(&bpp[sdno]->b_bio1);
433 * Next, wait for the requests to complete.
434 * We wait in the order in which they were
435 * issued, which isn't necessarily the order in
436 * which they complete, but we don't have a
437 * convenient way of doing the latter, and the
440 for (sdno = 0; sdno < plex->subdisks; sdno++) { /* for each subdisk */
441 if ((sdno != psd) || (op != rebuildparity)) {
443 if (bpp[sdno]->b_flags & B_ERROR) /* can't read, */
444 error = bpp[sdno]->b_error;
445 else if (sdno != psd) { /* update parity */
446 sbuf = (int *) bpp[sdno]->b_data;
447 for (i = 0; i < isize; i++)
448 ((int *) newparity_buf)[i] ^= sbuf[i]; /* xor in the buffer */
451 if (sdno != psd) { /* release all bps except parity */
452 Free(bpp[sdno]->b_data);
453 relpbuf(bpp[sdno], &vinum_conf.physbufs); /* give back our resources */
458 * If we're checking, compare the calculated
459 * and the read parity block. If they're
460 * different, return the plex-relative offset;
461 * otherwise return -1.
463 if ((op == checkparity)
464 || (op == rebuildandcheckparity)) {
465 *errorloc = -1; /* no error yet */
466 for (i = 0; i < isize; i++) {
467 if (parity_buf[i] != newparity_buf[i]) {
468 *errorloc = (off_t) (pstripe << DEV_BSHIFT) * (plex->subdisks - 1)
473 Free(bpp[psd]->b_data);
474 relpbuf(bpp[psd], &vinum_conf.physbufs); /* give back our resources */
476 /* release our resources */
479 pbp->b_flags |= B_ERROR;
480 pbp->b_error = error;
486 * Initialize a subdisk by writing zeroes to the
487 * complete address space. If verify is set,
488 * check each transfer for correctness.
490 * Each call to this function writes (and maybe
491 * checks) a single block.
494 initsd(int sdno, int verify)
501 int size; /* size of init block, bytes */
502 daddr_t plexblkno; /* lblkno in plex */
503 int verified; /* set when we're happy with what we wrote */
506 plexblkno = 0; /* to keep the compiler happy */
508 if (sd->plexno < 0) /* no plex? */
510 plex = &PLEX[sd->plexno]; /* point to plex */
511 if (plex->volno >= 0)
512 vol = &VOL[plex->volno];
516 if (sd->init_blocksize == 0) {
517 if (plex->stripesize != 0) /* we're striped, don't init more than */
518 sd->init_blocksize = min(DEFAULT_REVIVE_BLOCKSIZE, /* one block at a time */
519 plex->stripesize << DEV_BSHIFT);
521 sd->init_blocksize = DEFAULT_REVIVE_BLOCKSIZE;
522 } else if (sd->init_blocksize > MAX_REVIVE_BLOCKSIZE)
523 sd->init_blocksize = MAX_REVIVE_BLOCKSIZE;
525 size = min(sd->init_blocksize >> DEV_BSHIFT, sd->sectors - sd->initialized) << DEV_BSHIFT;
527 bp = getpbuf(&vinum_conf.physbufs); /* Get a buffer */
528 bp->b_data = Malloc(size);
531 while (!verified) { /* until we're happy with it, */
533 bp->b_resid = bp->b_bcount;
534 bp->b_bio1.bio_offset = (off_t)sd->initialized << DEV_BSHIFT; /* write it to here */
535 bp->b_bio1.bio_driver_info = VINUM_SD(sdno);
536 bzero(bp->b_data, bp->b_bcount);
537 bp->b_cmd = BUF_CMD_WRITE;
538 sdio(&bp->b_bio1); /* perform the I/O */
540 if (bp->b_flags & B_ERROR)
542 if ((error == 0) && verify) { /* check that it got there */
544 bp->b_resid = bp->b_bcount;
545 bp->b_bio1.bio_offset = (off_t)sd->initialized << DEV_BSHIFT; /* read from here */
546 bp->b_bio1.bio_driver_info = VINUM_SD(sdno);
547 bp->b_cmd = BUF_CMD_READ; /* read it back */
551 * XXX Bug fix code. This is hopefully no
552 * longer needed (21 February 2000).
554 if (bp->b_flags & B_ERROR)
556 else if ((*bp->b_data != 0) /* first word spammed */
557 ||(bcmp(bp->b_data, &bp->b_data[1], bp->b_bcount - 1))) { /* or one of the others */
558 printf("vinum: init error on %s, offset 0x%llx sectors\n",
560 (long long) sd->initialized);
568 relpbuf(bp, &vinum_conf.physbufs);
569 if (error == 0) { /* did it, */
570 sd->initialized += size >> DEV_BSHIFT; /* moved this much further down */
571 if (sd->initialized >= sd->sectors) { /* finished */
573 set_sd_state(sdno, sd_initialized, setstate_force); /* bring the sd up */
574 log(LOG_INFO, "vinum: %s is %s\n", sd->name, sd_state(sd->state));
575 save_config(); /* and save the updated configuration */
576 } else /* more to go, */
577 error = EAGAIN; /* ya'll come back, see? */
582 /* Local Variables: */
583 /* fill-column: 50 */