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 1.5 2005/06/11 00:05:46 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)
65 int size; /* size of revive block, bytes */
66 daddr_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 daddr_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 = min(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 = min(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 */
143 bp = geteblk(size); /* Get a buffer */
149 * Amount to transfer: block size, unless it
150 * would overlap the end.
153 bp->b_resid = bp->b_bcount;
154 bp->b_blkno = plexblkno; /* start here */
155 if (isstriped(plex)) /* we need to lock striped plexes */
156 lock = lockrange(plexblkno << DEV_BSHIFT, bp, plex); /* lock it */
157 if (vol != NULL) /* it's part of a volume, */
159 * First, read the data from the volume. We
160 * don't care which plex, that's bre's job.
162 bp->b_dev = VINUMDEV(plex->volno, 0, 0, VINUM_VOLUME_TYPE); /* create the device number */
163 else /* it's an unattached plex */
164 bp->b_dev = VINUM_PLEX(sd->plexno); /* create the device number */
166 bp->b_flags = B_READ; /* either way, read it */
171 if (bp->b_flags & B_ERROR)
174 /* Now write to the subdisk */
176 bp->b_dev = VINUM_SD(sdno); /* create the device number */
177 bp->b_flags = B_ORDERED | B_WRITE; /* and make this an ordered write */
178 bp->b_resid = bp->b_bcount;
179 bp->b_blkno = sd->revived; /* write it to here */
180 sdio(bp); /* 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;
200 if (debug & DEBUG_REVIVECONFLICT)
202 "Relaunch revive conflict sd %d: %p\n%s dev %d.%d, offset 0x%x, length %ld\n",
205 rq->bp->b_flags & B_READ ? "Read" : "Write",
206 major(rq->bp->b_dev),
207 minor(rq->bp->b_dev),
211 launch_requests(sd->waitlist, 1); /* do them now */
212 sd->waitlist = sd->waitlist->next; /* and move on to the next */
215 if (bp->b_qindex == 0) { /* not on a queue, */
216 bp->b_flags |= B_INVAL;
217 bp->b_flags &= ~B_ERROR;
218 brelse(bp); /* is this kosher? */
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_flags &= ~B_READ;
294 pbp->b_resid = pbp->b_bcount;
295 sdio(pbp); /* 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;
317 pbp->b_flags |= B_INVAL;
318 pbp->b_flags &= ~B_ERROR;
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 = min(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. */
394 bpp[sdno] = geteblk(mysize); /* Get a buffer */
395 if (bpp[sdno] == NULL) {
396 while (sdno-- > 0) { /* release the ones we got */
397 bpp[sdno]->b_flags |= B_INVAL;
398 brelse(bpp[sdno]); /* give back our resources */
401 printf("vinum: can't allocate buffer space for parity op.\n");
402 return NULL; /* no bpps */
406 parity_buf = (int *) bpp[sdno]->b_data;
407 if (sdno == newpsd) /* the new one? */
408 bpp[sdno]->b_dev = VINUM_SD(plex->sdnos[psd]); /* write back to the parity SD */
410 bpp[sdno]->b_dev = VINUM_SD(plex->sdnos[sdno]); /* device number */
411 bpp[sdno]->b_flags = B_READ; /* either way, read it */
412 bpp[sdno]->b_bcount = mysize;
413 bpp[sdno]->b_resid = bpp[sdno]->b_bcount;
414 bpp[sdno]->b_blkno = pstripe; /* transfer from here */
418 /* Initialize result buffer */
420 newparity_buf = (int *) bpp[newpsd]->b_data;
421 bzero(newparity_buf, mysize);
424 * Now lock the stripe with the first non-parity
427 *lockp = lockrange(pstripe * plex->stripesize * (plex->subdisks - 1),
432 * Then issue requests for all subdisks in
433 * parallel. Don't transfer the parity stripe
434 * if we're rebuilding parity, unless we also
437 for (sdno = 0; sdno < plex->subdisks; sdno++) { /* for each real subdisk */
438 if ((sdno != psd) || (op != rebuildparity)) {
444 * Next, wait for the requests to complete.
445 * We wait in the order in which they were
446 * issued, which isn't necessarily the order in
447 * which they complete, but we don't have a
448 * convenient way of doing the latter, and the
451 for (sdno = 0; sdno < plex->subdisks; sdno++) { /* for each subdisk */
452 if ((sdno != psd) || (op != rebuildparity)) {
454 if (bpp[sdno]->b_flags & B_ERROR) /* can't read, */
455 error = bpp[sdno]->b_error;
456 else if (sdno != psd) { /* update parity */
457 sbuf = (int *) bpp[sdno]->b_data;
458 for (i = 0; i < isize; i++)
459 ((int *) newparity_buf)[i] ^= sbuf[i]; /* xor in the buffer */
462 if (sdno != psd) { /* release all bps except parity */
463 bpp[sdno]->b_flags |= B_INVAL;
464 brelse(bpp[sdno]); /* give back our resources */
469 * If we're checking, compare the calculated
470 * and the read parity block. If they're
471 * different, return the plex-relative offset;
472 * otherwise return -1.
474 if ((op == checkparity)
475 || (op == rebuildandcheckparity)) {
476 *errorloc = -1; /* no error yet */
477 for (i = 0; i < isize; i++) {
478 if (parity_buf[i] != newparity_buf[i]) {
479 *errorloc = (off_t) (pstripe << DEV_BSHIFT) * (plex->subdisks - 1)
484 bpp[psd]->b_flags |= B_INVAL;
485 brelse(bpp[psd]); /* give back our resources */
487 /* release our resources */
490 pbp->b_flags |= B_ERROR;
491 pbp->b_error = error;
497 * Initialize a subdisk by writing zeroes to the
498 * complete address space. If verify is set,
499 * check each transfer for correctness.
501 * Each call to this function writes (and maybe
502 * checks) a single block.
505 initsd(int sdno, int verify)
512 int size; /* size of init block, bytes */
513 daddr_t plexblkno; /* lblkno in plex */
514 int verified; /* set when we're happy with what we wrote */
517 plexblkno = 0; /* to keep the compiler happy */
519 if (sd->plexno < 0) /* no plex? */
521 plex = &PLEX[sd->plexno]; /* point to plex */
522 if (plex->volno >= 0)
523 vol = &VOL[plex->volno];
527 if (sd->init_blocksize == 0) {
528 if (plex->stripesize != 0) /* we're striped, don't init more than */
529 sd->init_blocksize = min(DEFAULT_REVIVE_BLOCKSIZE, /* one block at a time */
530 plex->stripesize << DEV_BSHIFT);
532 sd->init_blocksize = DEFAULT_REVIVE_BLOCKSIZE;
533 } else if (sd->init_blocksize > MAX_REVIVE_BLOCKSIZE)
534 sd->init_blocksize = MAX_REVIVE_BLOCKSIZE;
536 size = min(sd->init_blocksize >> DEV_BSHIFT, sd->sectors - sd->initialized) << DEV_BSHIFT;
539 while (!verified) { /* until we're happy with it, */
541 bp = geteblk(size); /* Get a buffer */
547 bp->b_resid = bp->b_bcount;
548 bp->b_blkno = sd->initialized; /* write it to here */
549 bzero(bp->b_data, bp->b_bcount);
550 bp->b_dev = VINUM_SD(sdno); /* create the device number */
551 bp->b_flags &= ~B_READ;
552 sdio(bp); /* perform the I/O */
554 if (bp->b_flags & B_ERROR)
556 if (bp->b_qindex == 0) { /* not on a queue, */
557 bp->b_flags |= B_INVAL;
558 bp->b_flags &= ~B_ERROR;
559 brelse(bp); /* is this kosher? */
561 if ((error == 0) && verify) { /* check that it got there */
563 bp = geteblk(size); /* get a buffer */
569 bp->b_resid = bp->b_bcount;
570 bp->b_blkno = sd->initialized; /* read from here */
571 bp->b_dev = VINUM_SD(sdno); /* create the device number */
572 bp->b_flags |= B_READ; /* read it back */
577 * XXX Bug fix code. This is hopefully no
578 * longer needed (21 February 2000).
580 if (bp->b_flags & B_ERROR)
582 else if ((*bp->b_data != 0) /* first word spammed */
583 ||(bcmp(bp->b_data, &bp->b_data[1], bp->b_bcount - 1))) { /* or one of the others */
584 printf("vinum: init error on %s, offset 0x%llx sectors\n",
586 (long long) sd->initialized);
590 if (bp->b_qindex == 0) { /* not on a queue, */
591 bp->b_flags |= B_INVAL;
592 bp->b_flags &= ~B_ERROR;
593 brelse(bp); /* is this kosher? */
599 if (error == 0) { /* did it, */
600 sd->initialized += size >> DEV_BSHIFT; /* moved this much further down */
601 if (sd->initialized >= sd->sectors) { /* finished */
603 set_sd_state(sdno, sd_initialized, setstate_force); /* bring the sd up */
604 log(LOG_INFO, "vinum: %s is %s\n", sd->name, sd_state(sd->state));
605 save_config(); /* and save the updated configuration */
606 } else /* more to go, */
607 error = EAGAIN; /* ya'll come back, see? */
612 /* Local Variables: */
613 /* fill-column: 50 */