Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/dev/vinum/vinumraid5.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
1 /*- 2 * Copyright (c) 1997, 1998 3 * Cybernet Corporation and Nan Yang Computer Services Limited. 4 * All rights reserved. 5 * 6 * This software was developed as part of the NetMAX project. 7 * 8 * Written by Greg Lehey 9 * 10 * This software is distributed under the so-called ``Berkeley 11 * License'': 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by Cybernet Corporation 24 * and Nan Yang Computer Services Limited 25 * 4. Neither the name of the Companies nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * This software is provided ``as is'', and any express or implied 30 * warranties, including, but not limited to, the implied warranties of 31 * merchantability and fitness for a particular purpose are disclaimed. 32 * In no event shall the company or contributors be liable for any 33 * direct, indirect, incidental, special, exemplary, or consequential 34 * damages (including, but not limited to, procurement of substitute 35 * goods or services; loss of use, data, or profits; or business 36 * interruption) however caused and on any theory of liability, whether 37 * in contract, strict liability, or tort (including negligence or 38 * otherwise) arising in any way out of the use of this software, even if 39 * advised of the possibility of such damage. 40 * 41 * $Id: vinumraid5.c,v 1.21 2001/01/09 04:21:27 grog Exp grog $ 42 * $FreeBSD: releng/5.0/sys/dev/vinum/vinumraid5.c 70866 2001-01-10 05:07:14Z grog $ 43 */ 44 #include <dev/vinum/vinumhdr.h> 45 #include <dev/vinum/request.h> 46 #include <sys/resourcevar.h> 47 48 /* 49 * Parameters which describe the current transfer. 50 * These are only used for calculation, but they 51 * need to be passed to other functions, so it's 52 * tidier to put them in a struct 53 */ 54 struct metrics { 55 daddr_t stripebase; /* base address of stripe (1st subdisk) */ 56 int stripeoffset; /* offset in stripe */ 57 int stripesectors; /* total sectors to transfer in this stripe */ 58 daddr_t sdbase; /* offset in subdisk of stripe base */ 59 int sdcount; /* number of disks involved in this transfer */ 60 daddr_t diskstart; /* remember where this transfer starts */ 61 int psdno; /* number of parity subdisk */ 62 int badsdno; /* number of down subdisk, if there is one */ 63 int firstsdno; /* first data subdisk number */ 64 /* These correspond to the fields in rqelement, sort of */ 65 int useroffset; 66 /* 67 * Initial offset and length values for the first 68 * data block 69 */ 70 int initoffset; /* start address of block to transfer */ 71 short initlen; /* length in sectors of data transfer */ 72 /* Define a normal operation */ 73 int dataoffset; /* start address of block to transfer */ 74 int datalen; /* length in sectors of data transfer */ 75 /* Define a group operation */ 76 int groupoffset; /* subdisk offset of group operation */ 77 int grouplen; /* length in sectors of group operation */ 78 /* Define a normal write operation */ 79 int writeoffset; /* subdisk offset of normal write */ 80 int writelen; /* length in sectors of write operation */ 81 enum xferinfo flags; /* to check what we're doing */ 82 int rqcount; /* number of elements in request */ 83 }; 84 85 enum requeststatus bre5(struct request *rq, 86 int plexno, 87 daddr_t * diskstart, 88 daddr_t diskend); 89 void complete_raid5_write(struct rqelement *); 90 enum requeststatus build_rq_buffer(struct rqelement *rqe, struct plex *plex); 91 void setrqebounds(struct rqelement *rqe, struct metrics *mp); 92 93 /* 94 * define the low-level requests needed to perform 95 * a high-level I/O operation for a specific plex 96 * 'plexno'. 97 * 98 * Return 0 if all subdisks involved in the 99 * request are up, 1 if some subdisks are not up, 100 * and -1 if the request is at least partially 101 * outside the bounds of the subdisks. 102 * 103 * Modify the pointer *diskstart to point to the 104 * end address. On read, return on the first bad 105 * subdisk, so that the caller 106 * (build_read_request) can try alternatives. 107 * 108 * On entry to this routine, the prq structures 109 * are not assigned. The assignment is performed 110 * by expandrq(). Strictly speaking, the elements 111 * rqe->sdno of all entries should be set to -1, 112 * since 0 (from bzero) is a valid subdisk number. 113 * We avoid this problem by initializing the ones 114 * we use, and not looking at the others (index >= 115 * prq->requests). 116 */ 117 enum requeststatus 118 bre5(struct request *rq, 119 int plexno, 120 daddr_t * diskaddr, 121 daddr_t diskend) 122 { 123 struct metrics m; /* most of the information */ 124 struct sd *sd; 125 struct plex *plex; 126 struct buf *bp; /* user's bp */ 127 struct rqgroup *rqg; /* the request group that we will create */ 128 struct rqelement *rqe; /* point to this request information */ 129 int rsectors; /* sectors remaining in this stripe */ 130 int mysdno; /* another sd index in loops */ 131 int rqno; /* request number */ 132 133 rqg = NULL; /* shut up, damn compiler */ 134 m.diskstart = *diskaddr; /* start of transfer */ 135 bp = rq->bp; /* buffer pointer */ 136 plex = &PLEX[plexno]; /* point to the plex */ 137 138 139 while (*diskaddr < diskend) { /* until we get it all sorted out */ 140 if (*diskaddr >= plex->length) /* beyond the end of the plex */ 141 return REQUEST_EOF; /* can't continue */ 142 143 m.badsdno = -1; /* no bad subdisk yet */ 144 145 /* Part A: Define the request */ 146 /* 147 * First, calculate some sizes: 148 * The offset of the start address from 149 * the start of the stripe. 150 */ 151 m.stripeoffset = *diskaddr % (plex->stripesize * (plex->subdisks - 1)); 152 153 /* 154 * The plex-relative address of the 155 * start of the stripe. 156 */ 157 m.stripebase = *diskaddr - m.stripeoffset; 158 159 /* subdisk containing the parity stripe */ 160 if (plex->organization == plex_raid5) 161 m.psdno = plex->subdisks - 1 162 - (*diskaddr / (plex->stripesize * (plex->subdisks - 1))) 163 % plex->subdisks; 164 else /* RAID-4 */ 165 m.psdno = plex->subdisks - 1; 166 167 /* 168 * The number of the subdisk in which 169 * the start is located. 170 */ 171 m.firstsdno = m.stripeoffset / plex->stripesize; 172 if (m.firstsdno >= m.psdno) /* at or past parity sd */ 173 m.firstsdno++; /* increment it */ 174 175 /* 176 * The offset from the beginning of 177 * the stripe on this subdisk. 178 */ 179 m.initoffset = m.stripeoffset % plex->stripesize; 180 181 /* The offset of the stripe start relative to this subdisk */ 182 m.sdbase = m.stripebase / (plex->subdisks - 1); 183 184 m.useroffset = *diskaddr - m.diskstart; /* The offset of the start in the user buffer */ 185 186 /* 187 * The number of sectors to transfer in the 188 * current (first) subdisk. 189 */ 190 m.initlen = min(diskend - *diskaddr, /* the amount remaining to transfer */ 191 plex->stripesize - m.initoffset); /* and the amount left in this block */ 192 193 /* 194 * The number of sectors to transfer in this stripe 195 * is the minumum of the amount remaining to transfer 196 * and the amount left in this stripe. 197 */ 198 m.stripesectors = min(diskend - *diskaddr, 199 plex->stripesize * (plex->subdisks - 1) - m.stripeoffset); 200 201 /* The number of data subdisks involved in this request */ 202 m.sdcount = (m.stripesectors + m.initoffset + plex->stripesize - 1) / plex->stripesize; 203 204 /* Part B: decide what kind of transfer this will be. 205 206 * start and end addresses of the transfer in 207 * the current block. 208 * 209 * There are a number of different kinds of 210 * transfer, each of which relates to a 211 * specific subdisk: 212 * 213 * 1. Normal read. All participating subdisks 214 * are up, and the transfer can be made 215 * directly to the user buffer. The bounds 216 * of the transfer are described by 217 * m.dataoffset and m.datalen. We have 218 * already calculated m.initoffset and 219 * m.initlen, which define the parameters 220 * for the first data block. 221 * 222 * 2. Recovery read. One participating 223 * subdisk is down. To recover data, all 224 * the other subdisks, including the parity 225 * subdisk, must be read. The data is 226 * recovered by exclusive-oring all the 227 * other blocks. The bounds of the 228 * transfer are described by m.groupoffset 229 * and m.grouplen. 230 * 231 * 3. A read request may request reading both 232 * available data (normal read) and 233 * non-available data (recovery read). 234 * This can be a problem if the address 235 * ranges of the two reads do not coincide: 236 * in this case, the normal read needs to 237 * be extended to cover the address range 238 * of the recovery read, and must thus be 239 * performed out of malloced memory. 240 * 241 * 4. Normal write. All the participating 242 * subdisks are up. The bounds of the 243 * transfer are described by m.dataoffset 244 * and m.datalen. Since these values 245 * differ for each block, we calculate the 246 * bounds for the parity block 247 * independently as the maximum of the 248 * individual blocks and store these values 249 * in m.writeoffset and m.writelen. This 250 * write proceeds in four phases: 251 * 252 * i. Read the old contents of each block 253 * and the parity block. 254 * ii. ``Remove'' the old contents from 255 * the parity block with exclusive or. 256 * iii. ``Insert'' the new contents of the 257 * block in the parity block, again 258 * with exclusive or. 259 * 260 * iv. Write the new contents of the data 261 * blocks and the parity block. The data 262 * block transfers can be made directly from 263 * the user buffer. 264 * 265 * 5. Degraded write where the data block is 266 * not available. The bounds of the 267 * transfer are described by m.groupoffset 268 * and m.grouplen. This requires the 269 * following steps: 270 * 271 * i. Read in all the other data blocks, 272 * excluding the parity block. 273 * 274 * ii. Recreate the parity block from the 275 * other data blocks and the data to be 276 * written. 277 * 278 * iii. Write the parity block. 279 * 280 * 6. Parityless write, a write where the 281 * parity block is not available. This is 282 * in fact the simplest: just write the 283 * data blocks. This can proceed directly 284 * from the user buffer. The bounds of the 285 * transfer are described by m.dataoffset 286 * and m.datalen. 287 * 288 * 7. Combination of degraded data block write 289 * and normal write. In this case the 290 * address ranges of the reads may also 291 * need to be extended to cover all 292 * participating blocks. 293 * 294 * All requests in a group transfer transfer 295 * the same address range relative to their 296 * subdisk. The individual transfers may 297 * vary, but since our group of requests is 298 * all in a single slice, we can define a 299 * range in which they all fall. 300 * 301 * In the following code section, we determine 302 * which kind of transfer we will perform. If 303 * there is a group transfer, we also decide 304 * its bounds relative to the subdisks. At 305 * the end, we have the following values: 306 * 307 * m.flags indicates the kinds of transfers 308 * we will perform. 309 * m.initoffset indicates the offset of the 310 * beginning of any data operation relative 311 * to the beginning of the stripe base. 312 * m.initlen specifies the length of any data 313 * operation. 314 * m.dataoffset contains the same value as 315 * m.initoffset. 316 * m.datalen contains the same value as 317 * m.initlen. Initially dataoffset and 318 * datalen describe the parameters for the 319 * first data block; while building the data 320 * block requests, they are updated for each 321 * block. 322 * m.groupoffset indicates the offset of any 323 * group operation relative to the beginning 324 * of the stripe base. 325 * m.grouplen specifies the length of any 326 * group operation. 327 * m.writeoffset indicates the offset of a 328 * normal write relative to the beginning of 329 * the stripe base. This value differs from 330 * m.dataoffset in that it applies to the 331 * entire operation, and not just the first 332 * block. 333 * m.writelen specifies the total span of a 334 * normal write operation. writeoffset and 335 * writelen are used to define the parity 336 * block. 337 */ 338 m.groupoffset = 0; /* assume no group... */ 339 m.grouplen = 0; /* until we know we have one */ 340 m.writeoffset = m.initoffset; /* start offset of transfer */ 341 m.writelen = 0; /* nothing to write yet */ 342 m.flags = 0; /* no flags yet */ 343 rsectors = m.stripesectors; /* remaining sectors to examine */ 344 m.dataoffset = m.initoffset; /* start at the beginning of the transfer */ 345 m.datalen = m.initlen; 346 347 if (m.sdcount > 1) { 348 plex->multiblock++; /* more than one block for the request */ 349 /* 350 * If we have two transfers that don't overlap, 351 * (one at the end of the first block, the other 352 * at the beginning of the second block), 353 * it's cheaper to split them. 354 */ 355 if (rsectors < plex->stripesize) { 356 m.sdcount = 1; /* just one subdisk */ 357 m.stripesectors = m.initlen; /* and just this many sectors */ 358 rsectors = m.initlen; /* and in the loop counter */ 359 } 360 } 361 if (SD[plex->sdnos[m.psdno]].state < sd_reborn) /* is our parity subdisk down? */ 362 m.badsdno = m.psdno; /* note that it's down */ 363 if (bp->b_iocmd == BIO_READ) { /* read operation */ 364 for (mysdno = m.firstsdno; rsectors > 0; mysdno++) { 365 if (mysdno == m.psdno) /* ignore parity on read */ 366 mysdno++; 367 if (mysdno == plex->subdisks) /* wraparound */ 368 mysdno = 0; 369 if (mysdno == m.psdno) /* parity, */ 370 mysdno++; /* we've given already */ 371 372 if (SD[plex->sdnos[mysdno]].state < sd_reborn) { /* got a bad subdisk, */ 373 if (m.badsdno >= 0) /* we had one already, */ 374 return REQUEST_DOWN; /* we can't take a second */ 375 m.badsdno = mysdno; /* got the first */ 376 m.groupoffset = m.dataoffset; /* define the bounds */ 377 m.grouplen = m.datalen; 378 m.flags |= XFR_RECOVERY_READ; /* we need recovery */ 379 plex->recovered_reads++; /* count another one */ 380 } else 381 m.flags |= XFR_NORMAL_READ; /* normal read */ 382 383 /* Update the pointers for the next block */ 384 m.dataoffset = 0; /* back to the start of the stripe */ 385 rsectors -= m.datalen; /* remaining sectors to examine */ 386 m.datalen = min(rsectors, plex->stripesize); /* amount that will fit in this block */ 387 } 388 } else { /* write operation */ 389 for (mysdno = m.firstsdno; rsectors > 0; mysdno++) { 390 if (mysdno == m.psdno) /* parity stripe, we've dealt with that */ 391 mysdno++; 392 if (mysdno == plex->subdisks) /* wraparound */ 393 mysdno = 0; 394 if (mysdno == m.psdno) /* parity, */ 395 mysdno++; /* we've given already */ 396 397 sd = &SD[plex->sdnos[mysdno]]; 398 if (sd->state != sd_up) { 399 enum requeststatus s; 400 401 s = checksdstate(sd, rq, *diskaddr, diskend); /* do we need to change state? */ 402 if (s && (m.badsdno >= 0)) { /* second bad disk, */ 403 int sdno; 404 /* 405 * If the parity disk is down, there's 406 * no recovery. We make all involved 407 * subdisks stale. Otherwise, we 408 * should be able to recover, but it's 409 * like pulling teeth. Fix it later. 410 */ 411 for (sdno = 0; sdno < m.sdcount; sdno++) { 412 struct sd *sd = &SD[plex->sdnos[sdno]]; 413 if (sd->state >= sd_reborn) /* sort of up, */ 414 set_sd_state(sd->sdno, sd_stale, setstate_force); /* make it stale */ 415 } 416 return s; /* and crap out */ 417 } 418 m.badsdno = mysdno; /* note which one is bad */ 419 m.flags |= XFR_DEGRADED_WRITE; /* we need recovery */ 420 plex->degraded_writes++; /* count another one */ 421 m.groupoffset = m.dataoffset; /* define the bounds */ 422 m.grouplen = m.datalen; 423 } else { 424 m.flags |= XFR_NORMAL_WRITE; /* normal write operation */ 425 if (m.writeoffset > m.dataoffset) { /* move write operation lower */ 426 m.writelen = max(m.writeoffset + m.writelen, 427 m.dataoffset + m.datalen) 428 - m.dataoffset; 429 m.writeoffset = m.dataoffset; 430 } else 431 m.writelen = max(m.writeoffset + m.writelen, 432 m.dataoffset + m.datalen) 433 - m.writeoffset; 434 } 435 436 /* Update the pointers for the next block */ 437 m.dataoffset = 0; /* back to the start of the stripe */ 438 rsectors -= m.datalen; /* remaining sectors to examine */ 439 m.datalen = min(rsectors, plex->stripesize); /* amount that will fit in this block */ 440 } 441 if (m.badsdno == m.psdno) { /* got a bad parity block, */ 442 struct sd *psd = &SD[plex->sdnos[m.psdno]]; 443 444 if (psd->state == sd_down) 445 set_sd_state(psd->sdno, sd_obsolete, setstate_force); /* it's obsolete now */ 446 else if (psd->state == sd_crashed) 447 set_sd_state(psd->sdno, sd_stale, setstate_force); /* it's stale now */ 448 m.flags &= ~XFR_NORMAL_WRITE; /* this write isn't normal, */ 449 m.flags |= XFR_PARITYLESS_WRITE; /* it's parityless */ 450 plex->parityless_writes++; /* count another one */ 451 } 452 } 453 454 /* reset the initial transfer values */ 455 m.dataoffset = m.initoffset; /* start at the beginning of the transfer */ 456 m.datalen = m.initlen; 457 458 /* decide how many requests we need */ 459 if (m.flags & (XFR_RECOVERY_READ | XFR_DEGRADED_WRITE)) 460 /* doing a recovery read or degraded write, */ 461 m.rqcount = plex->subdisks; /* all subdisks */ 462 else if (m.flags & XFR_NORMAL_WRITE) /* normal write, */ 463 m.rqcount = m.sdcount + 1; /* all data blocks and the parity block */ 464 else /* parityless write or normal read */ 465 m.rqcount = m.sdcount; /* just the data blocks */ 466 467 /* Part C: build the requests */ 468 rqg = allocrqg(rq, m.rqcount); /* get a request group */ 469 if (rqg == NULL) { /* malloc failed */ 470 bp->b_error = ENOMEM; 471 bp->b_ioflags |= BIO_ERROR; 472 return REQUEST_ENOMEM; 473 } 474 rqg->plexno = plexno; 475 rqg->flags = m.flags; 476 rqno = 0; /* index in the request group */ 477 478 /* 1: PARITY BLOCK */ 479 /* 480 * Are we performing an operation which requires parity? In that case, 481 * work out the parameters and define the parity block. 482 * XFR_PARITYOP is XFR_NORMAL_WRITE | XFR_RECOVERY_READ | XFR_DEGRADED_WRITE 483 */ 484 if (m.flags & XFR_PARITYOP) { /* need parity */ 485 rqe = &rqg->rqe[rqno]; /* point to element */ 486 sd = &SD[plex->sdnos[m.psdno]]; /* the subdisk in question */ 487 rqe->rqg = rqg; /* point back to group */ 488 rqe->flags = (m.flags | XFR_PARITY_BLOCK | XFR_MALLOCED) /* always malloc parity block */ 489 &~(XFR_NORMAL_READ | XFR_PARITYLESS_WRITE); /* transfer flags without data op stuf */ 490 setrqebounds(rqe, &m); /* set up the bounds of the transfer */ 491 rqe->sdno = sd->sdno; /* subdisk number */ 492 rqe->driveno = sd->driveno; 493 if (build_rq_buffer(rqe, plex)) /* build the buffer */ 494 return REQUEST_ENOMEM; /* can't do it */ 495 rqe->b.b_iocmd = BIO_READ; /* we must read first */ 496 m.sdcount++; /* adjust the subdisk count */ 497 rqno++; /* and point to the next request */ 498 } 499 /* 500 * 2: DATA BLOCKS 501 * Now build up requests for the blocks required 502 * for individual transfers 503 */ 504 for (mysdno = m.firstsdno; rqno < m.sdcount; mysdno++, rqno++) { 505 if (mysdno == m.psdno) /* parity, */ 506 mysdno++; /* we've given already */ 507 if (mysdno == plex->subdisks) /* got to the end, */ 508 mysdno = 0; /* wrap around */ 509 if (mysdno == m.psdno) /* parity, */ 510 mysdno++; /* we've given already */ 511 512 rqe = &rqg->rqe[rqno]; /* point to element */ 513 sd = &SD[plex->sdnos[mysdno]]; /* the subdisk in question */ 514 rqe->rqg = rqg; /* point to group */ 515 if (m.flags & XFR_NEEDS_MALLOC) /* we need a malloced buffer first */ 516 rqe->flags = m.flags | XFR_DATA_BLOCK | XFR_MALLOCED; /* transfer flags */ 517 else 518 rqe->flags = m.flags | XFR_DATA_BLOCK; /* transfer flags */ 519 if (mysdno == m.badsdno) { /* this is the bad subdisk */ 520 rqg->badsdno = rqno; /* note which one */ 521 rqe->flags |= XFR_BAD_SUBDISK; /* note that it's dead */ 522 /* 523 * we can't read or write from/to it, 524 * but we don't need to malloc 525 */ 526 rqe->flags &= ~(XFR_MALLOCED | XFR_NORMAL_READ | XFR_NORMAL_WRITE); 527 } 528 setrqebounds(rqe, &m); /* set up the bounds of the transfer */ 529 rqe->useroffset = m.useroffset; /* offset in user buffer */ 530 rqe->sdno = sd->sdno; /* subdisk number */ 531 rqe->driveno = sd->driveno; 532 if (build_rq_buffer(rqe, plex)) /* build the buffer */ 533 return REQUEST_ENOMEM; /* can't do it */ 534 if ((m.flags & XFR_PARITYOP) /* parity operation, */ 535 &&((m.flags & XFR_BAD_SUBDISK) == 0)) /* and not the bad subdisk, */ 536 rqe->b.b_iocmd = BIO_READ; /* we must read first */ 537 538 /* Now update pointers for the next block */ 539 *diskaddr += m.datalen; /* skip past what we've done */ 540 m.stripesectors -= m.datalen; /* deduct from what's left */ 541 m.useroffset += m.datalen; /* and move on in the user buffer */ 542 m.datalen = min(m.stripesectors, plex->stripesize); /* and recalculate */ 543 m.dataoffset = 0; /* start at the beginning of next block */ 544 } 545 546 /* 547 * 3: REMAINING BLOCKS FOR RECOVERY 548 * Finally, if we have a recovery operation, build 549 * up transfers for the other subdisks. Follow the 550 * subdisks around until we get to where we started. 551 * These requests use only the group parameters. 552 */ 553 if ((rqno < m.rqcount) /* haven't done them all already */ 554 &&(m.flags & (XFR_RECOVERY_READ | XFR_DEGRADED_WRITE))) { 555 for (; rqno < m.rqcount; rqno++, mysdno++) { 556 if (mysdno == m.psdno) /* parity, */ 557 mysdno++; /* we've given already */ 558 if (mysdno == plex->subdisks) /* got to the end, */ 559 mysdno = 0; /* wrap around */ 560 if (mysdno == m.psdno) /* parity, */ 561 mysdno++; /* we've given already */ 562 563 rqe = &rqg->rqe[rqno]; /* point to element */ 564 sd = &SD[plex->sdnos[mysdno]]; /* the subdisk in question */ 565 rqe->rqg = rqg; /* point to group */ 566 567 rqe->sdoffset = m.sdbase + m.groupoffset; /* start of transfer */ 568 rqe->dataoffset = 0; /* for tidiness' sake */ 569 rqe->groupoffset = 0; /* group starts at the beginining */ 570 rqe->datalen = 0; 571 rqe->grouplen = m.grouplen; 572 rqe->buflen = m.grouplen; 573 rqe->flags = (m.flags | XFR_MALLOCED) /* transfer flags without data op stuf */ 574 &~XFR_DATAOP; 575 rqe->sdno = sd->sdno; /* subdisk number */ 576 rqe->driveno = sd->driveno; 577 if (build_rq_buffer(rqe, plex)) /* build the buffer */ 578 return REQUEST_ENOMEM; /* can't do it */ 579 rqe->b.b_iocmd = BIO_READ; /* we must read first */ 580 } 581 } 582 /* 583 * We need to lock the address range before 584 * doing anything. We don't have to be 585 * performing a recovery operation: somebody 586 * else could be doing so, and the results could 587 * influence us. Note the fact here, we'll perform 588 * the lock in launch_requests. 589 */ 590 rqg->lockbase = m.stripebase; 591 if (*diskaddr < diskend) /* didn't finish the request on this stripe */ 592 plex->multistripe++; /* count another one */ 593 } 594 return REQUEST_OK; 595 } 596 597 /* 598 * Helper function for rqe5: adjust the bounds of 599 * the transfers to minimize the buffer 600 * allocation. 601 * 602 * Each request can handle two of three different 603 * data ranges: 604 * 605 * 1. The range described by the parameters 606 * dataoffset and datalen, for normal read or 607 * parityless write. 608 * 2. The range described by the parameters 609 * groupoffset and grouplen, for recovery read 610 * and degraded write. 611 * 3. For normal write, the range depends on the 612 * kind of block. For data blocks, the range 613 * is defined by dataoffset and datalen. For 614 * parity blocks, it is defined by writeoffset 615 * and writelen. 616 * 617 * In order not to allocate more memory than 618 * necessary, this function adjusts the bounds 619 * parameter for each request to cover just the 620 * minimum necessary for the function it performs. 621 * This will normally vary from one request to the 622 * next. 623 * 624 * Things are slightly different for the parity 625 * block. In this case, the bounds defined by 626 * mp->writeoffset and mp->writelen also play a 627 * rôle. Select this case by setting the 628 * parameter forparity != 0 629 */ 630 void 631 setrqebounds(struct rqelement *rqe, struct metrics *mp) 632 { 633 /* parity block of a normal write */ 634 if ((rqe->flags & (XFR_NORMAL_WRITE | XFR_PARITY_BLOCK)) 635 == (XFR_NORMAL_WRITE | XFR_PARITY_BLOCK)) { /* case 3 */ 636 if (rqe->flags & XFR_DEGRADED_WRITE) { /* also degraded write */ 637 /* 638 * With a combined normal and degraded write, we 639 * will zero out the area of the degraded write 640 * in the second phase, so we don't need to read 641 * it in. Unfortunately, we need a way to tell 642 * build_request_buffer the size of the buffer, 643 * and currently that's the length of the read. 644 * As a result, we read everything, even the stuff 645 * that we're going to nuke. 646 * FIXME XXX 647 */ 648 if (mp->groupoffset < mp->writeoffset) { /* group operation starts lower */ 649 rqe->sdoffset = mp->sdbase + mp->groupoffset; /* start of transfer */ 650 rqe->dataoffset = mp->writeoffset - mp->groupoffset; /* data starts here */ 651 rqe->groupoffset = 0; /* and the group at the beginning */ 652 } else { /* individual data starts first */ 653 rqe->sdoffset = mp->sdbase + mp->writeoffset; /* start of transfer */ 654 rqe->dataoffset = 0; /* individual data starts at the beginning */ 655 rqe->groupoffset = mp->groupoffset - mp->writeoffset; /* group starts here */ 656 } 657 rqe->datalen = mp->writelen; 658 rqe->grouplen = mp->grouplen; 659 } else { /* just normal write (case 3) */ 660 rqe->sdoffset = mp->sdbase + mp->writeoffset; /* start of transfer */ 661 rqe->dataoffset = 0; /* degradation starts at the beginning */ 662 rqe->groupoffset = 0; /* for tidiness' sake */ 663 rqe->datalen = mp->writelen; 664 rqe->grouplen = 0; 665 } 666 } else if (rqe->flags & XFR_DATAOP) { /* data operation (case 1 or 3) */ 667 if (rqe->flags & XFR_GROUPOP) { /* also a group operation (case 2) */ 668 if (mp->groupoffset < mp->dataoffset) { /* group operation starts lower */ 669 rqe->sdoffset = mp->sdbase + mp->groupoffset; /* start of transfer */ 670 rqe->dataoffset = mp->dataoffset - mp->groupoffset; /* data starts here */ 671 rqe->groupoffset = 0; /* and the group at the beginning */ 672 } else { /* individual data starts first */ 673 rqe->sdoffset = mp->sdbase + mp->dataoffset; /* start of transfer */ 674 rqe->dataoffset = 0; /* individual data starts at the beginning */ 675 rqe->groupoffset = mp->groupoffset - mp->dataoffset; /* group starts here */ 676 } 677 rqe->datalen = mp->datalen; 678 rqe->grouplen = mp->grouplen; 679 } else { /* just data operation (case 1) */ 680 rqe->sdoffset = mp->sdbase + mp->dataoffset; /* start of transfer */ 681 rqe->dataoffset = 0; /* degradation starts at the beginning */ 682 rqe->groupoffset = 0; /* for tidiness' sake */ 683 rqe->datalen = mp->datalen; 684 rqe->grouplen = 0; 685 } 686 } else { /* just group operations (case 2) */ 687 rqe->sdoffset = mp->sdbase + mp->groupoffset; /* start of transfer */ 688 rqe->dataoffset = 0; /* for tidiness' sake */ 689 rqe->groupoffset = 0; /* group starts at the beginining */ 690 rqe->datalen = 0; 691 rqe->grouplen = mp->grouplen; 692 } 693 rqe->buflen = max(rqe->dataoffset + rqe->datalen, /* total buffer length */ 694 rqe->groupoffset + rqe->grouplen); 695 } 696 /* Local Variables: */ 697 /* fill-column: 50 */ 698 /* End: */
Cache object: e893c4d0ef932e8df03216539461f8a7
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]