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 $
48 * Revive a block of a subdisk. Return an error
49 * indication. EAGAIN means successful copy, but
50 * that more blocks remain to be copied. EINVAL
51 * means that the subdisk isn't associated with a
52 * plex (which means a programming error if we get
53 * here at all; FIXME).
57 revive_block(int sdno)
65 int size; /* size of revive block, bytes */
66 vinum_off_t plexblkno; /* lblkno in plex */
67 int psd; /* parity subdisk number */
68 u_int64_t stripe; /* stripe number */
69 int paritysd = 0; /* set if this is the parity stripe */
70 struct rangelock *lock; /* for locking */
71 vinum_off_t stripeoffset; /* offset in stripe */
73 plexblkno = 0; /* to keep the compiler happy */
76 if (sd->plexno < 0) /* no plex? */
78 plex = &PLEX[sd->plexno]; /* point to plex */
80 vol = &VOL[plex->volno];
84 if ((sd->revive_blocksize == 0) /* no block size */
85 ||(sd->revive_blocksize & ((1 << DEV_BSHIFT) - 1))) /* or invalid block size */
86 sd->revive_blocksize = DEFAULT_REVIVE_BLOCKSIZE;
87 else if (sd->revive_blocksize > MAX_REVIVE_BLOCKSIZE)
88 sd->revive_blocksize = MAX_REVIVE_BLOCKSIZE;
89 size = u64min(sd->revive_blocksize >> DEV_BSHIFT, sd->sectors - sd->revived) << DEV_BSHIFT;
90 sd->reviver = curproc->p_pid; /* note who last had a bash at it */
92 /* Now decide where to read from */
93 switch (plex->organization) {
95 plexblkno = sd->revived + sd->plexoffset; /* corresponding address in plex */
99 stripeoffset = sd->revived % plex->stripesize; /* offset from beginning of stripe */
100 if (stripeoffset + (size >> DEV_BSHIFT) > plex->stripesize)
101 size = (plex->stripesize - stripeoffset) << DEV_BSHIFT;
102 plexblkno = sd->plexoffset /* base */
103 + (sd->revived - stripeoffset) * plex->subdisks /* offset to beginning of stripe */
104 + stripeoffset; /* offset from beginning of stripe */
109 stripeoffset = sd->revived % plex->stripesize; /* offset from beginning of stripe */
110 plexblkno = sd->plexoffset /* base */
111 + (sd->revived - stripeoffset) * (plex->subdisks - 1) /* offset to beginning of stripe */
112 +stripeoffset; /* offset from beginning of stripe */
113 stripe = (sd->revived / plex->stripesize); /* stripe number */
115 /* Make sure we don't go beyond the end of the band. */
116 size = u64min(size, (plex->stripesize - stripeoffset) << DEV_BSHIFT);
117 if (plex->organization == plex_raid4)
118 psd = plex->subdisks - 1; /* parity subdisk for this stripe */
120 psd = plex->subdisks - 1 - stripe % plex->subdisks; /* parity subdisk for this stripe */
121 paritysd = plex->sdnos[psd] == sdno; /* note if it's the parity subdisk */
124 * Now adjust for the strangenesses
125 * in RAID-4 and RAID-5 striping.
127 if (sd->plexsdno > psd) /* beyond the parity stripe, */
128 plexblkno -= plex->stripesize; /* one stripe less */
130 plexblkno -= plex->stripesize * sd->plexsdno; /* go back to the beginning of the band */
133 case plex_disorg: /* to keep the compiler happy */
137 if (paritysd) { /* we're reviving a parity block, */
138 bp = parityrebuild(plex, sd->revived, size, rebuildparity, &lock, NULL); /* do the grunt work */
139 if (bp == NULL) /* no buffer space */
140 return ENOMEM; /* chicken out */
141 } else { /* data block */
142 bp = getpbuf(&vinum_conf.physbufs); /* Get a buffer */
143 bp->b_data = Malloc(size);
146 * Amount to transfer: block size, unless it
147 * would overlap the end.
150 bp->b_resid = bp->b_bcount;
151 bp->b_bio1.bio_offset = (off_t)plexblkno << DEV_BSHIFT; /* start here */
152 bp->b_bio1.bio_done = biodone_sync;
153 bp->b_bio1.bio_flags |= BIO_SYNC;
154 if (isstriped(plex)) /* we need to lock striped plexes */
155 lock = lockrange(plexblkno << DEV_BSHIFT, bp, plex); /* lock it */
156 if (vol != NULL) /* it's part of a volume, */
158 * First, read the data from the volume. We
159 * don't care which plex, that's bre's job.
162 else /* it's an unattached plex */
163 dev = PLEX[sd->plexno].plex_dev;
165 bp->b_cmd = BUF_CMD_READ;
166 vinumstart(dev, &bp->b_bio1, 1);
167 biowait(&bp->b_bio1, "drvrd");
170 if (bp->b_flags & B_ERROR)
173 /* Now write to the subdisk */
175 dev = SD[sdno].sd_dev;
176 KKASSERT(dev != NULL);
177 bp->b_flags |= B_ORDERED; /* and make this an ordered write */
178 bp->b_cmd = BUF_CMD_WRITE;
179 bp->b_resid = bp->b_bcount;
180 bp->b_bio1.bio_offset = (off_t)sd->revived << DEV_BSHIFT; /* write it to here */
181 bp->b_bio1.bio_driver_info = dev;
182 bp->b_bio1.bio_done = biodone_sync;
183 sdio(&bp->b_bio1); /* perform the I/O */
184 biowait(&bp->b_bio1, "drvwr");
185 if (bp->b_flags & B_ERROR)
188 sd->revived += bp->b_bcount >> DEV_BSHIFT; /* moved this much further down */
189 if (sd->revived >= sd->sectors) { /* finished */
191 set_sd_state(sdno, sd_up, setstate_force); /* bring the sd up */
192 log(LOG_INFO, "vinum: %s is %s\n", sd->name, sd_state(sd->state));
193 save_config(); /* and save the updated configuration */
194 error = 0; /* we're done */
197 if (lock) /* we took a lock, */
198 unlockrange(sd->plexno, lock); /* give it back */
199 while (sd->waitlist) { /* we have waiting requests */
201 struct request *rq = sd->waitlist;
204 if (debug & DEBUG_REVIVECONFLICT) {
205 dev = rq->bio->bio_driver_info;
207 "Relaunch revive conflict sd %d: %p\n%s dev %d.%d, offset 0x%jx, length %d\n",
210 (rq->bio->bio_buf->b_cmd == BUF_CMD_READ) ? "Read" : "Write",
213 (uintmax_t)rq->bio->bio_offset,
214 rq->bio->bio_buf->b_bcount);
217 launch_requests(sd->waitlist, 1); /* do them now */
218 sd->waitlist = sd->waitlist->next; /* and move on to the next */
222 relpbuf(bp, &vinum_conf.physbufs);
227 * Check or rebuild the parity blocks of a RAID-4
230 * The variables plex->checkblock and
231 * plex->rebuildblock represent the
232 * subdisk-relative address of the stripe we're
233 * looking at, not the plex-relative address. We
234 * store it in the plex and not as a local
235 * variable because this function could be
236 * stopped, and we don't want to repeat the part
237 * we've already done. This is also the reason
238 * why we don't initialize it here except at the
239 * end. It gets initialized with the plex on
242 * Each call to this function processes at most
243 * one stripe. We can't loop in this function,
244 * because we're unstoppable, so we have to be
245 * called repeatedly from userland.
248 parityops(struct vinum_ioctl_msg *data)
252 int size; /* I/O transfer size, bytes */
253 struct rangelock *lock; /* lock on stripe */
254 struct _ioctl_reply *reply;
255 off_t pstripe; /* pointer to our stripe counter */
257 off_t errorloc; /* offset of parity error */
258 enum parityop op; /* operation to perform */
260 plexno = data->index;
263 reply = (struct _ioctl_reply *) data;
264 reply->error = EAGAIN; /* expect to repeat this call */
265 plex = &PLEX[plexno];
266 if (!isparity(plex)) { /* not RAID-4 or RAID-5 */
267 reply->error = EINVAL;
269 } else if (plex->state < plex_flaky) {
271 strcpy(reply->msg, "Plex is not completely accessible\n");
274 pstripe = data->offset;
275 size = imin(DEFAULT_REVIVE_BLOCKSIZE, /* one block at a time */
276 plex->stripesize << DEV_BSHIFT);
278 errorloc = 0; /* avoid gcc warnings */
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 pbp->b_bio1.bio_done = biodone_sync;
296 sdio(&pbp->b_bio1); /* write the parity block */
297 biowait(&pbp->b_bio1, "drvwr");
299 if (((op == checkparity)
300 || (op == rebuildandcheckparity))
301 && (errorloc != -1)) {
302 if (op == checkparity)
305 "Parity incorrect at offset 0x%llx\n",
306 (long long)errorloc);
308 if (reply->error == EAGAIN) { /* still OK, */
309 plex->checkblock = pstripe + (pbp->b_bcount >> DEV_BSHIFT); /* moved this much further down */
310 if (plex->checkblock >= SD[plex->sdnos[0]].sectors) { /* finished */
311 plex->checkblock = 0;
316 if (pbp->b_flags & B_ERROR)
317 reply->error = pbp->b_error;
319 relpbuf(pbp, &vinum_conf.physbufs);
320 unlockrange(plexno, lock);
324 * Rebuild a parity stripe. Return pointer to
325 * parity bp. On return,
327 * 1. The band is locked. The caller must unlock
328 * the band and release the buffer header.
330 * 2. All buffer headers except php have been
331 * released. The caller must release pbp.
333 * 3. For checkparity and rebuildandcheckparity,
334 * the parity is compared with the current
335 * parity block. If it's different, the
336 * offset of the error is returned to
337 * errorloc. The caller can set the value of
338 * the pointer to NULL if this is called for
341 * pstripe is the subdisk-relative base address of
342 * the data to be reconstructed, size is the size
343 * of the transfer in bytes.
346 parityrebuild(struct plex *plex,
350 struct rangelock **lockp,
355 u_int64_t stripe; /* stripe number */
356 int *parity_buf; /* buffer address for current parity block */
357 int *newparity_buf; /* and for new parity block */
358 int mysize; /* I/O transfer size for this transfer */
359 int isize; /* mysize in ints */
361 int psd; /* parity subdisk number */
362 int newpsd; /* and "subdisk number" of new parity */
363 struct buf **bpp; /* pointers to our bps */
364 struct buf *pbp; /* buffer header for parity stripe */
366 int bufcount; /* number of buffers we need */
368 stripe = pstripe / plex->stripesize; /* stripe number */
369 psd = plex->subdisks - 1 - stripe % plex->subdisks; /* parity subdisk for this stripe */
370 parity_buf = NULL; /* to keep the compiler happy */
374 * It's possible that the default transfer size
375 * we chose is not a factor of the stripe size.
376 * We *must* limit this operation to a single
377 * stripe, at least for RAID-5 rebuild, since
378 * the parity subdisk changes between stripes,
379 * so in this case we need to perform a short
380 * transfer. Set variable mysize to reflect
383 mysize = u64min(size, (plex->stripesize * (stripe + 1) - pstripe) << DEV_BSHIFT);
384 isize = mysize / (sizeof(int)); /* number of ints in the buffer */
385 bufcount = plex->subdisks + 1; /* sd buffers plus result buffer */
386 newpsd = plex->subdisks;
387 bpp = (struct buf **) Malloc(bufcount * sizeof(struct buf *)); /* array of pointers to bps */
389 /* First, build requests for all subdisks */
390 for (sdno = 0; sdno < bufcount; sdno++) { /* for each subdisk */
391 if ((sdno != psd) || (op != rebuildparity)) {
392 /* Get a buffer header and initialize it. */
393 bpp[sdno] = getpbuf(&vinum_conf.physbufs); /* Get a buffer */
394 bpp[sdno]->b_data = Malloc(mysize);
396 parity_buf = (int *) bpp[sdno]->b_data;
397 if (sdno == newpsd) /* the new one? */
398 bpp[sdno]->b_bio1.bio_driver_info = SD[plex->sdnos[psd]].sd_dev; /* write back to the parity SD */
400 bpp[sdno]->b_bio1.bio_driver_info = SD[plex->sdnos[sdno]].sd_dev; /* device number */
401 KKASSERT(bpp[sdno]->b_bio1.bio_driver_info);
402 bpp[sdno]->b_cmd = BUF_CMD_READ; /* either way, read it */
403 bpp[sdno]->b_bcount = mysize;
404 bpp[sdno]->b_resid = bpp[sdno]->b_bcount;
405 bpp[sdno]->b_bio1.bio_offset = (off_t)pstripe << DEV_BSHIFT; /* transfer from here */
406 bpp[sdno]->b_bio1.bio_done = biodone_sync;
410 /* Initialize result buffer */
412 newparity_buf = (int *) bpp[newpsd]->b_data;
413 bzero(newparity_buf, mysize);
416 * Now lock the stripe with the first non-parity
419 *lockp = lockrange(pstripe * plex->stripesize * (plex->subdisks - 1),
424 * Then issue requests for all subdisks in
425 * parallel. Don't transfer the parity stripe
426 * if we're rebuilding parity, unless we also
429 for (sdno = 0; sdno < plex->subdisks; sdno++) { /* for each real subdisk */
430 if ((sdno != psd) || (op != rebuildparity)) {
431 sdio(&bpp[sdno]->b_bio1);
436 * Next, wait for the requests to complete.
437 * We wait in the order in which they were
438 * issued, which isn't necessarily the order in
439 * which they complete, but we don't have a
440 * convenient way of doing the latter, and the
443 for (sdno = 0; sdno < plex->subdisks; sdno++) { /* for each subdisk */
444 if ((sdno != psd) || (op != rebuildparity)) {
445 biowait(&bpp[sdno]->b_bio1, "drvio");
446 if (bpp[sdno]->b_flags & B_ERROR) /* can't read, */
447 error = bpp[sdno]->b_error;
448 else if (sdno != psd) { /* update parity */
449 sbuf = (int *) bpp[sdno]->b_data;
450 for (i = 0; i < isize; i++)
451 newparity_buf[i] ^= sbuf[i]; /* xor in the buffer */
454 if (sdno != psd) { /* release all bps except parity */
455 Free(bpp[sdno]->b_data);
456 relpbuf(bpp[sdno], &vinum_conf.physbufs); /* give back our resources */
461 * If we're checking, compare the calculated
462 * and the read parity block. If they're
463 * different, return the plex-relative offset;
464 * otherwise return -1.
466 if ((op == checkparity)
467 || (op == rebuildandcheckparity)) {
468 *errorloc = -1; /* no error yet */
469 for (i = 0; i < isize; i++) {
470 if (parity_buf[i] != newparity_buf[i]) {
471 *errorloc = (off_t) (pstripe << DEV_BSHIFT) * (plex->subdisks - 1)
476 Free(bpp[psd]->b_data);
477 relpbuf(bpp[psd], &vinum_conf.physbufs); /* give back our resources */
479 /* release our resources */
482 pbp->b_flags |= B_ERROR;
483 pbp->b_error = error;
489 * Initialize a subdisk by writing zeroes to the
490 * complete address space. If verify is set,
491 * check each transfer for correctness.
493 * Each call to this function writes (and maybe
494 * checks) a single block.
497 initsd(int sdno, int verify)
503 int size; /* size of init block, bytes */
504 int verified; /* set when we're happy with what we wrote */
508 if (sd->plexno < 0) /* no plex? */
510 plex = &PLEX[sd->plexno]; /* point to plex */
512 if (sd->init_blocksize == 0) {
513 if (plex->stripesize != 0) /* we're striped, don't init more than */
514 sd->init_blocksize = u64min(DEFAULT_REVIVE_BLOCKSIZE, /* one block at a time */
515 plex->stripesize << DEV_BSHIFT);
517 sd->init_blocksize = DEFAULT_REVIVE_BLOCKSIZE;
518 } else if (sd->init_blocksize > MAX_REVIVE_BLOCKSIZE)
519 sd->init_blocksize = MAX_REVIVE_BLOCKSIZE;
521 size = u64min(sd->init_blocksize >> DEV_BSHIFT, sd->sectors - sd->initialized) << DEV_BSHIFT;
523 bp = getpbuf(&vinum_conf.physbufs); /* Get a buffer */
524 bp->b_data = Malloc(size);
527 while (!verified) { /* until we're happy with it, */
529 bp->b_resid = bp->b_bcount;
530 bp->b_bio1.bio_offset = (off_t)sd->initialized << DEV_BSHIFT; /* write it to here */
531 bp->b_bio1.bio_driver_info = SD[sdno].sd_dev;
532 bp->b_bio1.bio_done = biodone_sync;
533 KKASSERT(bp->b_bio1.bio_driver_info);
534 bzero(bp->b_data, bp->b_bcount);
535 bp->b_cmd = BUF_CMD_WRITE;
536 sdio(&bp->b_bio1); /* perform the I/O */
537 biowait(&bp->b_bio1, "drvwr");
538 if (bp->b_flags & B_ERROR)
540 if ((error == 0) && verify) { /* check that it got there */
542 bp->b_resid = bp->b_bcount;
543 bp->b_bio1.bio_offset = (off_t)sd->initialized << DEV_BSHIFT; /* read from here */
544 bp->b_bio1.bio_driver_info = SD[sdno].sd_dev;
545 bp->b_bio1.bio_done = biodone_sync;
546 KKASSERT(bp->b_bio1.bio_driver_info);
547 bp->b_cmd = BUF_CMD_READ; /* read it back */
549 biowait(&bp->b_bio1, "drvrd");
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 kprintf("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? */