| 1 | /* $NetBSD: rf_paritymap.c,v 1.8 2011/04/27 07:55:15 mrg Exp $ */ |
| 2 | |
| 3 | /*- |
| 4 | * Copyright (c) 2009 Jed Davis. |
| 5 | * All rights reserved. |
| 6 | * |
| 7 | * Redistribution and use in source and binary forms, with or without |
| 8 | * modification, are permitted provided that the following conditions |
| 9 | * are met: |
| 10 | * 1. Redistributions of source code must retain the above copyright |
| 11 | * notice, this list of conditions and the following disclaimer. |
| 12 | * 2. Redistributions in binary form must reproduce the above copyright |
| 13 | * notice, this list of conditions and the following disclaimer in the |
| 14 | * documentation and/or other materials provided with the distribution. |
| 15 | * |
| 16 | * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS |
| 17 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED |
| 18 | * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
| 19 | * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS |
| 20 | * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
| 21 | * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 22 | * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS |
| 23 | * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN |
| 24 | * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) |
| 25 | * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
| 26 | * POSSIBILITY OF SUCH DAMAGE. |
| 27 | */ |
| 28 | |
| 29 | #include <sys/cdefs.h> |
| 30 | __KERNEL_RCSID(0, "$NetBSD: rf_paritymap.c,v 1.8 2011/04/27 07:55:15 mrg Exp $" ); |
| 31 | |
| 32 | #include <sys/param.h> |
| 33 | #include <sys/callout.h> |
| 34 | #include <sys/kmem.h> |
| 35 | #include <sys/mutex.h> |
| 36 | #include <sys/rwlock.h> |
| 37 | #include <sys/systm.h> |
| 38 | #include <sys/types.h> |
| 39 | |
| 40 | #include <dev/raidframe/rf_paritymap.h> |
| 41 | #include <dev/raidframe/rf_stripelocks.h> |
| 42 | #include <dev/raidframe/rf_layout.h> |
| 43 | #include <dev/raidframe/rf_raid.h> |
| 44 | #include <dev/raidframe/rf_parityscan.h> |
| 45 | #include <dev/raidframe/rf_kintf.h> |
| 46 | |
| 47 | /* Important parameters: */ |
| 48 | #define REGION_MINSIZE (25ULL << 20) |
| 49 | #define DFL_TICKMS 40000 |
| 50 | #define DFL_COOLDOWN 8 /* 7-8 intervals of 40s = 5min +/- 20s */ |
| 51 | |
| 52 | /* Internal-use flag bits. */ |
| 53 | #define TICKING 1 |
| 54 | #define TICKED 2 |
| 55 | |
| 56 | /* Prototypes! */ |
| 57 | static void rf_paritymap_write_locked(struct rf_paritymap *); |
| 58 | static void rf_paritymap_tick(void *); |
| 59 | static u_int rf_paritymap_nreg(RF_Raid_t *); |
| 60 | |
| 61 | /* Extract the current status of the parity map. */ |
| 62 | void |
| 63 | rf_paritymap_status(struct rf_paritymap *pm, struct rf_pmstat *ps) |
| 64 | { |
| 65 | memset(ps, 0, sizeof(*ps)); |
| 66 | if (pm == NULL) |
| 67 | ps->enabled = 0; |
| 68 | else { |
| 69 | ps->enabled = 1; |
| 70 | ps->region_size = pm->region_size; |
| 71 | mutex_enter(&pm->lock); |
| 72 | memcpy(&ps->params, &pm->params, sizeof(ps->params)); |
| 73 | memcpy(ps->dirty, pm->disk_now, sizeof(ps->dirty)); |
| 74 | memcpy(&ps->ctrs, &pm->ctrs, sizeof(ps->ctrs)); |
| 75 | mutex_exit(&pm->lock); |
| 76 | } |
| 77 | } |
| 78 | |
| 79 | /* |
| 80 | * Test whether parity in a given sector is suspected of being inconsistent |
| 81 | * on disk (assuming that any pending I/O to it is allowed to complete). |
| 82 | * This may be of interest to future work on parity scrubbing. |
| 83 | */ |
| 84 | int |
| 85 | rf_paritymap_test(struct rf_paritymap *pm, daddr_t sector) |
| 86 | { |
| 87 | unsigned region = sector / pm->region_size; |
| 88 | int retval; |
| 89 | |
| 90 | mutex_enter(&pm->lock); |
| 91 | retval = isset(pm->disk_boot->bits, region) ? 1 : 0; |
| 92 | mutex_exit(&pm->lock); |
| 93 | return retval; |
| 94 | } |
| 95 | |
| 96 | /* To be called before a write to the RAID is submitted. */ |
| 97 | void |
| 98 | rf_paritymap_begin(struct rf_paritymap *pm, daddr_t offset, daddr_t size) |
| 99 | { |
| 100 | unsigned i, b, e; |
| 101 | |
| 102 | b = offset / pm->region_size; |
| 103 | e = (offset + size - 1) / pm->region_size; |
| 104 | |
| 105 | for (i = b; i <= e; i++) |
| 106 | rf_paritymap_begin_region(pm, i); |
| 107 | } |
| 108 | |
| 109 | /* To be called after a write to the RAID completes. */ |
| 110 | void |
| 111 | rf_paritymap_end(struct rf_paritymap *pm, daddr_t offset, daddr_t size) |
| 112 | { |
| 113 | unsigned i, b, e; |
| 114 | |
| 115 | b = offset / pm->region_size; |
| 116 | e = (offset + size - 1) / pm->region_size; |
| 117 | |
| 118 | for (i = b; i <= e; i++) |
| 119 | rf_paritymap_end_region(pm, i); |
| 120 | } |
| 121 | |
| 122 | void |
| 123 | rf_paritymap_begin_region(struct rf_paritymap *pm, unsigned region) |
| 124 | { |
| 125 | int needs_write; |
| 126 | |
| 127 | KASSERT(region < RF_PARITYMAP_NREG); |
| 128 | pm->ctrs.nwrite++; |
| 129 | |
| 130 | /* If it was being kept warm, deal with that. */ |
| 131 | mutex_enter(&pm->lock); |
| 132 | if (pm->current->state[region] < 0) |
| 133 | pm->current->state[region] = 0; |
| 134 | |
| 135 | /* This shouldn't happen unless RAIDOUTSTANDING is set too high. */ |
| 136 | KASSERT(pm->current->state[region] < 127); |
| 137 | pm->current->state[region]++; |
| 138 | |
| 139 | needs_write = isclr(pm->disk_now->bits, region); |
| 140 | |
| 141 | if (needs_write) { |
| 142 | KASSERT(pm->current->state[region] == 1); |
| 143 | rf_paritymap_write_locked(pm); |
| 144 | } |
| 145 | |
| 146 | mutex_exit(&pm->lock); |
| 147 | } |
| 148 | |
| 149 | void |
| 150 | rf_paritymap_end_region(struct rf_paritymap *pm, unsigned region) |
| 151 | { |
| 152 | KASSERT(region < RF_PARITYMAP_NREG); |
| 153 | |
| 154 | mutex_enter(&pm->lock); |
| 155 | KASSERT(pm->current->state[region] > 0); |
| 156 | --pm->current->state[region]; |
| 157 | |
| 158 | if (pm->current->state[region] <= 0) { |
| 159 | pm->current->state[region] = -pm->params.cooldown; |
| 160 | KASSERT(pm->current->state[region] <= 0); |
| 161 | mutex_enter(&pm->lk_flags); |
| 162 | if (!(pm->flags & TICKING)) { |
| 163 | pm->flags |= TICKING; |
| 164 | mutex_exit(&pm->lk_flags); |
| 165 | callout_schedule(&pm->ticker, |
| 166 | mstohz(pm->params.tickms)); |
| 167 | } else |
| 168 | mutex_exit(&pm->lk_flags); |
| 169 | } |
| 170 | mutex_exit(&pm->lock); |
| 171 | } |
| 172 | |
| 173 | /* |
| 174 | * Updates the parity map to account for any changes in current activity |
| 175 | * and/or an ongoing parity scan, then writes it to disk with appropriate |
| 176 | * synchronization. |
| 177 | */ |
| 178 | void |
| 179 | rf_paritymap_write(struct rf_paritymap *pm) |
| 180 | { |
| 181 | mutex_enter(&pm->lock); |
| 182 | rf_paritymap_write_locked(pm); |
| 183 | mutex_exit(&pm->lock); |
| 184 | } |
| 185 | |
| 186 | /* As above, but to be used when pm->lock is already held. */ |
| 187 | static void |
| 188 | rf_paritymap_write_locked(struct rf_paritymap *pm) |
| 189 | { |
| 190 | char w, w0; |
| 191 | int i, j, setting, clearing; |
| 192 | |
| 193 | setting = clearing = 0; |
| 194 | for (i = 0; i < RF_PARITYMAP_NBYTE; i++) { |
| 195 | w0 = pm->disk_now->bits[i]; |
| 196 | w = pm->disk_boot->bits[i]; |
| 197 | |
| 198 | for (j = 0; j < NBBY; j++) |
| 199 | if (pm->current->state[i * NBBY + j] != 0) |
| 200 | w |= 1 << j; |
| 201 | |
| 202 | if (w & ~w0) |
| 203 | setting = 1; |
| 204 | if (w0 & ~w) |
| 205 | clearing = 1; |
| 206 | |
| 207 | pm->disk_now->bits[i] = w; |
| 208 | } |
| 209 | pm->ctrs.ncachesync += setting + clearing; |
| 210 | pm->ctrs.nclearing += clearing; |
| 211 | |
| 212 | /* |
| 213 | * If bits are being set in the parity map, then a sync is |
| 214 | * required afterwards, so that the regions are marked dirty |
| 215 | * on disk before any writes to them take place. If bits are |
| 216 | * being cleared, then a sync is required before the write, so |
| 217 | * that any writes to those regions are processed before the |
| 218 | * region is marked clean. (Synchronization is somewhat |
| 219 | * overkill; a write ordering barrier would suffice, but we |
| 220 | * currently have no way to express that directly.) |
| 221 | */ |
| 222 | if (clearing) |
| 223 | rf_sync_component_caches(pm->raid); |
| 224 | rf_paritymap_kern_write(pm->raid, pm->disk_now); |
| 225 | if (setting) |
| 226 | rf_sync_component_caches(pm->raid); |
| 227 | } |
| 228 | |
| 229 | /* Mark all parity as being in need of rewrite. */ |
| 230 | void |
| 231 | rf_paritymap_invalidate(struct rf_paritymap *pm) |
| 232 | { |
| 233 | mutex_enter(&pm->lock); |
| 234 | memset(pm->disk_boot, ~(unsigned char)0, |
| 235 | sizeof(struct rf_paritymap_ondisk)); |
| 236 | mutex_exit(&pm->lock); |
| 237 | } |
| 238 | |
| 239 | /* Mark all parity as being correct. */ |
| 240 | void |
| 241 | rf_paritymap_forceclean(struct rf_paritymap *pm) |
| 242 | { |
| 243 | mutex_enter(&pm->lock); |
| 244 | memset(pm->disk_boot, (unsigned char)0, |
| 245 | sizeof(struct rf_paritymap_ondisk)); |
| 246 | mutex_exit(&pm->lock); |
| 247 | } |
| 248 | |
| 249 | /* |
| 250 | * The cooldown callout routine just defers its work to a thread; it can't do |
| 251 | * the parity map write itself as it would block, and although mutex-induced |
| 252 | * blocking is permitted it seems wise to avoid tying up the softint. |
| 253 | */ |
| 254 | static void |
| 255 | rf_paritymap_tick(void *arg) |
| 256 | { |
| 257 | struct rf_paritymap *pm = arg; |
| 258 | |
| 259 | mutex_enter(&pm->lk_flags); |
| 260 | pm->flags |= TICKED; |
| 261 | mutex_exit(&pm->lk_flags); |
| 262 | |
| 263 | rf_lock_mutex2(pm->raid->iodone_lock); |
| 264 | rf_signal_cond2(pm->raid->iodone_cv); /* XXX */ |
| 265 | rf_unlock_mutex2(pm->raid->iodone_lock); |
| 266 | } |
| 267 | |
| 268 | /* |
| 269 | * This is where the parity cooling work (and rearming the callout if needed) |
| 270 | * is done; the raidio thread calls it when woken up, as by the above. |
| 271 | */ |
| 272 | void |
| 273 | rf_paritymap_checkwork(struct rf_paritymap *pm) |
| 274 | { |
| 275 | int i, zerop, progressp; |
| 276 | |
| 277 | mutex_enter(&pm->lk_flags); |
| 278 | if (pm->flags & TICKED) { |
| 279 | zerop = progressp = 0; |
| 280 | |
| 281 | pm->flags &= ~TICKED; |
| 282 | mutex_exit(&pm->lk_flags); |
| 283 | |
| 284 | mutex_enter(&pm->lock); |
| 285 | for (i = 0; i < RF_PARITYMAP_NREG; i++) { |
| 286 | if (pm->current->state[i] < 0) { |
| 287 | progressp = 1; |
| 288 | pm->current->state[i]++; |
| 289 | if (pm->current->state[i] == 0) |
| 290 | zerop = 1; |
| 291 | } |
| 292 | } |
| 293 | |
| 294 | if (progressp) |
| 295 | callout_schedule(&pm->ticker, |
| 296 | mstohz(pm->params.tickms)); |
| 297 | else { |
| 298 | mutex_enter(&pm->lk_flags); |
| 299 | pm->flags &= ~TICKING; |
| 300 | mutex_exit(&pm->lk_flags); |
| 301 | } |
| 302 | |
| 303 | if (zerop) |
| 304 | rf_paritymap_write_locked(pm); |
| 305 | mutex_exit(&pm->lock); |
| 306 | } else |
| 307 | mutex_exit(&pm->lk_flags); |
| 308 | } |
| 309 | |
| 310 | /* |
| 311 | * Set parity map parameters; used both to alter parameters on the fly and to |
| 312 | * establish their initial values. Note that setting a parameter to 0 means |
| 313 | * to leave the previous setting unchanged, and that if this is done for the |
| 314 | * initial setting of "regions", then a default value will be computed based |
| 315 | * on the RAID component size. |
| 316 | */ |
| 317 | int |
| 318 | rf_paritymap_set_params(struct rf_paritymap *pm, |
| 319 | const struct rf_pmparams *params, int todisk) |
| 320 | { |
| 321 | int cooldown, tickms; |
| 322 | u_int regions; |
| 323 | RF_RowCol_t col; |
| 324 | RF_ComponentLabel_t *clabel; |
| 325 | RF_Raid_t *raidPtr; |
| 326 | |
| 327 | cooldown = params->cooldown != 0 |
| 328 | ? params->cooldown : pm->params.cooldown; |
| 329 | tickms = params->tickms != 0 |
| 330 | ? params->tickms : pm->params.tickms; |
| 331 | regions = params->regions != 0 |
| 332 | ? params->regions : pm->params.regions; |
| 333 | |
| 334 | if (cooldown < 1 || cooldown > 128) { |
| 335 | printf("raid%d: cooldown %d out of range\n" , pm->raid->raidid, |
| 336 | cooldown); |
| 337 | return (-1); |
| 338 | } |
| 339 | if (tickms < 10) { |
| 340 | printf("raid%d: tick time %dms out of range\n" , |
| 341 | pm->raid->raidid, tickms); |
| 342 | return (-1); |
| 343 | } |
| 344 | if (regions == 0) { |
| 345 | regions = rf_paritymap_nreg(pm->raid); |
| 346 | } else if (regions > RF_PARITYMAP_NREG) { |
| 347 | printf("raid%d: region count %u too large (more than %u)\n" , |
| 348 | pm->raid->raidid, regions, RF_PARITYMAP_NREG); |
| 349 | return (-1); |
| 350 | } |
| 351 | |
| 352 | /* XXX any currently warm parity will be used with the new tickms! */ |
| 353 | pm->params.cooldown = cooldown; |
| 354 | pm->params.tickms = tickms; |
| 355 | /* Apply the initial region count, but do not change it after that. */ |
| 356 | if (pm->params.regions == 0) |
| 357 | pm->params.regions = regions; |
| 358 | |
| 359 | /* So that the newly set parameters can be tested: */ |
| 360 | pm->ctrs.nwrite = pm->ctrs.ncachesync = pm->ctrs.nclearing = 0; |
| 361 | |
| 362 | if (todisk) { |
| 363 | raidPtr = pm->raid; |
| 364 | for (col = 0; col < raidPtr->numCol; col++) { |
| 365 | if (RF_DEAD_DISK(raidPtr->Disks[col].status)) |
| 366 | continue; |
| 367 | |
| 368 | clabel = raidget_component_label(raidPtr, col); |
| 369 | clabel->parity_map_ntick = cooldown; |
| 370 | clabel->parity_map_tickms = tickms; |
| 371 | clabel->parity_map_regions = regions; |
| 372 | |
| 373 | /* Don't touch the disk if it's been spared */ |
| 374 | if (clabel->status == rf_ds_spared) |
| 375 | continue; |
| 376 | |
| 377 | raidflush_component_label(raidPtr, col); |
| 378 | } |
| 379 | |
| 380 | /* handle the spares too... */ |
| 381 | for (col = 0; col < raidPtr->numSpare; col++) { |
| 382 | if (raidPtr->Disks[raidPtr->numCol+col].status == rf_ds_used_spare) { |
| 383 | clabel = raidget_component_label(raidPtr, raidPtr->numCol+col); |
| 384 | clabel->parity_map_ntick = cooldown; |
| 385 | clabel->parity_map_tickms = tickms; |
| 386 | clabel->parity_map_regions = regions; |
| 387 | raidflush_component_label(raidPtr, raidPtr->numCol+col); |
| 388 | } |
| 389 | } |
| 390 | } |
| 391 | return 0; |
| 392 | } |
| 393 | |
| 394 | /* |
| 395 | * The number of regions may not be as many as can fit into the map, because |
| 396 | * when regions are too small, the overhead of setting parity map bits |
| 397 | * becomes significant in comparison to the actual I/O, while the |
| 398 | * corresponding gains in parity verification time become negligible. Thus, |
| 399 | * a minimum region size (defined above) is imposed. |
| 400 | * |
| 401 | * Note that, if the number of regions is less than the maximum, then some of |
| 402 | * the regions will be "fictional", corresponding to no actual disk; some |
| 403 | * parts of the code may process them as normal, but they can not ever be |
| 404 | * written to. |
| 405 | */ |
| 406 | static u_int |
| 407 | rf_paritymap_nreg(RF_Raid_t *raid) |
| 408 | { |
| 409 | daddr_t bytes_per_disk, nreg; |
| 410 | |
| 411 | bytes_per_disk = raid->sectorsPerDisk << raid->logBytesPerSector; |
| 412 | nreg = bytes_per_disk / REGION_MINSIZE; |
| 413 | if (nreg > RF_PARITYMAP_NREG) |
| 414 | nreg = RF_PARITYMAP_NREG; |
| 415 | if (nreg < 1) |
| 416 | nreg = 1; |
| 417 | |
| 418 | return (u_int)nreg; |
| 419 | } |
| 420 | |
| 421 | /* |
| 422 | * Initialize a parity map given specific parameters. This neither reads nor |
| 423 | * writes the parity map config in the component labels; for that, see below. |
| 424 | */ |
| 425 | int |
| 426 | rf_paritymap_init(struct rf_paritymap *pm, RF_Raid_t *raid, |
| 427 | const struct rf_pmparams *params) |
| 428 | { |
| 429 | daddr_t rstripes; |
| 430 | struct rf_pmparams safe; |
| 431 | |
| 432 | pm->raid = raid; |
| 433 | pm->params.regions = 0; |
| 434 | if (0 != rf_paritymap_set_params(pm, params, 0)) { |
| 435 | /* |
| 436 | * If the parameters are out-of-range, then bring the |
| 437 | * parity map up with something reasonable, so that |
| 438 | * the admin can at least go and fix it (or ignore it |
| 439 | * entirely). |
| 440 | */ |
| 441 | safe.cooldown = DFL_COOLDOWN; |
| 442 | safe.tickms = DFL_TICKMS; |
| 443 | safe.regions = 0; |
| 444 | |
| 445 | if (0 != rf_paritymap_set_params(pm, &safe, 0)) |
| 446 | return (-1); |
| 447 | } |
| 448 | |
| 449 | rstripes = howmany(raid->Layout.numStripe, pm->params.regions); |
| 450 | pm->region_size = rstripes * raid->Layout.dataSectorsPerStripe; |
| 451 | |
| 452 | callout_init(&pm->ticker, CALLOUT_MPSAFE); |
| 453 | callout_setfunc(&pm->ticker, rf_paritymap_tick, pm); |
| 454 | pm->flags = 0; |
| 455 | |
| 456 | pm->disk_boot = kmem_alloc(sizeof(struct rf_paritymap_ondisk), |
| 457 | KM_SLEEP); |
| 458 | pm->disk_now = kmem_alloc(sizeof(struct rf_paritymap_ondisk), |
| 459 | KM_SLEEP); |
| 460 | pm->current = kmem_zalloc(sizeof(struct rf_paritymap_current), |
| 461 | KM_SLEEP); |
| 462 | |
| 463 | rf_paritymap_kern_read(pm->raid, pm->disk_boot); |
| 464 | memcpy(pm->disk_now, pm->disk_boot, sizeof(*pm->disk_now)); |
| 465 | |
| 466 | mutex_init(&pm->lock, MUTEX_DEFAULT, IPL_NONE); |
| 467 | mutex_init(&pm->lk_flags, MUTEX_DEFAULT, IPL_SOFTCLOCK); |
| 468 | |
| 469 | return 0; |
| 470 | } |
| 471 | |
| 472 | /* |
| 473 | * Destroys a parity map; unless "force" is set, also cleans parity for any |
| 474 | * regions which were still in cooldown (but are not dirty on disk). |
| 475 | */ |
| 476 | void |
| 477 | rf_paritymap_destroy(struct rf_paritymap *pm, int force) |
| 478 | { |
| 479 | int i; |
| 480 | |
| 481 | callout_halt(&pm->ticker, NULL); /* XXX stop? halt? */ |
| 482 | callout_destroy(&pm->ticker); |
| 483 | |
| 484 | if (!force) { |
| 485 | for (i = 0; i < RF_PARITYMAP_NREG; i++) { |
| 486 | /* XXX check for > 0 ? */ |
| 487 | if (pm->current->state[i] < 0) |
| 488 | pm->current->state[i] = 0; |
| 489 | } |
| 490 | |
| 491 | rf_paritymap_write_locked(pm); |
| 492 | } |
| 493 | |
| 494 | mutex_destroy(&pm->lock); |
| 495 | mutex_destroy(&pm->lk_flags); |
| 496 | |
| 497 | kmem_free(pm->disk_boot, sizeof(struct rf_paritymap_ondisk)); |
| 498 | kmem_free(pm->disk_now, sizeof(struct rf_paritymap_ondisk)); |
| 499 | kmem_free(pm->current, sizeof(struct rf_paritymap_current)); |
| 500 | } |
| 501 | |
| 502 | /* |
| 503 | * Rewrite parity, taking parity map into account; this is the equivalent of |
| 504 | * the old rf_RewriteParity, and is likewise to be called from a suitable |
| 505 | * thread and shouldn't have multiple copies running in parallel and so on. |
| 506 | * |
| 507 | * Note that the fictional regions are "cleaned" in one shot, so that very |
| 508 | * small RAIDs (useful for testing) will not experience potentially severe |
| 509 | * regressions in rewrite time. |
| 510 | */ |
| 511 | int |
| 512 | rf_paritymap_rewrite(struct rf_paritymap *pm) |
| 513 | { |
| 514 | int i, ret_val = 0; |
| 515 | daddr_t reg_b, reg_e; |
| 516 | |
| 517 | /* Process only the actual regions. */ |
| 518 | for (i = 0; i < pm->params.regions; i++) { |
| 519 | mutex_enter(&pm->lock); |
| 520 | if (isset(pm->disk_boot->bits, i)) { |
| 521 | mutex_exit(&pm->lock); |
| 522 | |
| 523 | reg_b = i * pm->region_size; |
| 524 | reg_e = reg_b + pm->region_size; |
| 525 | if (reg_e > pm->raid->totalSectors) |
| 526 | reg_e = pm->raid->totalSectors; |
| 527 | |
| 528 | if (rf_RewriteParityRange(pm->raid, reg_b, |
| 529 | reg_e - reg_b)) { |
| 530 | ret_val = 1; |
| 531 | if (pm->raid->waitShutdown) |
| 532 | return ret_val; |
| 533 | } else { |
| 534 | mutex_enter(&pm->lock); |
| 535 | clrbit(pm->disk_boot->bits, i); |
| 536 | rf_paritymap_write_locked(pm); |
| 537 | mutex_exit(&pm->lock); |
| 538 | } |
| 539 | } else { |
| 540 | mutex_exit(&pm->lock); |
| 541 | } |
| 542 | } |
| 543 | |
| 544 | /* Now, clear the fictional regions, if any. */ |
| 545 | rf_paritymap_forceclean(pm); |
| 546 | rf_paritymap_write(pm); |
| 547 | |
| 548 | return ret_val; |
| 549 | } |
| 550 | |
| 551 | /* |
| 552 | * How to merge the on-disk parity maps when reading them in from the |
| 553 | * various components; returns whether they differ. In the case that |
| 554 | * they do differ, sets *dst to the union of *dst and *src. |
| 555 | * |
| 556 | * In theory, it should be safe to take the intersection (or just pick |
| 557 | * a single component arbitrarily), but the paranoid approach costs |
| 558 | * little. |
| 559 | * |
| 560 | * Appropriate locking, if any, is the responsibility of the caller. |
| 561 | */ |
| 562 | int |
| 563 | rf_paritymap_merge(struct rf_paritymap_ondisk *dst, |
| 564 | struct rf_paritymap_ondisk *src) |
| 565 | { |
| 566 | int i, discrep = 0; |
| 567 | |
| 568 | for (i = 0; i < RF_PARITYMAP_NBYTE; i++) { |
| 569 | if (dst->bits[i] != src->bits[i]) |
| 570 | discrep = 1; |
| 571 | dst->bits[i] |= src->bits[i]; |
| 572 | } |
| 573 | |
| 574 | return discrep; |
| 575 | } |
| 576 | |
| 577 | /* |
| 578 | * Detach a parity map from its RAID. This is not meant to be applied except |
| 579 | * when unconfiguring the RAID after all I/O has been resolved, as otherwise |
| 580 | * an out-of-date parity map could be treated as current. |
| 581 | */ |
| 582 | void |
| 583 | rf_paritymap_detach(RF_Raid_t *raidPtr) |
| 584 | { |
| 585 | if (raidPtr->parity_map == NULL) |
| 586 | return; |
| 587 | |
| 588 | rf_lock_mutex2(raidPtr->iodone_lock); |
| 589 | struct rf_paritymap *pm = raidPtr->parity_map; |
| 590 | raidPtr->parity_map = NULL; |
| 591 | rf_unlock_mutex2(raidPtr->iodone_lock); |
| 592 | /* XXXjld is that enough locking? Or too much? */ |
| 593 | rf_paritymap_destroy(pm, 0); |
| 594 | kmem_free(pm, sizeof(*pm)); |
| 595 | } |
| 596 | |
| 597 | /* |
| 598 | * Is this RAID set ineligible for parity-map use due to not actually |
| 599 | * having any parity? (If so, rf_paritymap_attach is a no-op, but |
| 600 | * rf_paritymap_{get,set}_disable will still pointlessly act on the |
| 601 | * component labels.) |
| 602 | */ |
| 603 | int |
| 604 | rf_paritymap_ineligible(RF_Raid_t *raidPtr) |
| 605 | { |
| 606 | return raidPtr->Layout.map->faultsTolerated == 0; |
| 607 | } |
| 608 | |
| 609 | /* |
| 610 | * Attach a parity map to a RAID set if appropriate. Includes |
| 611 | * configure-time processing of parity-map fields of component label. |
| 612 | */ |
| 613 | void |
| 614 | rf_paritymap_attach(RF_Raid_t *raidPtr, int force) |
| 615 | { |
| 616 | RF_RowCol_t col; |
| 617 | int pm_use, pm_zap; |
| 618 | int g_tickms, g_ntick, g_regions; |
| 619 | int good; |
| 620 | RF_ComponentLabel_t *clabel; |
| 621 | u_int flags, regions; |
| 622 | struct rf_pmparams params; |
| 623 | |
| 624 | if (rf_paritymap_ineligible(raidPtr)) { |
| 625 | /* There isn't any parity. */ |
| 626 | return; |
| 627 | } |
| 628 | |
| 629 | pm_use = 1; |
| 630 | pm_zap = 0; |
| 631 | g_tickms = DFL_TICKMS; |
| 632 | g_ntick = DFL_COOLDOWN; |
| 633 | g_regions = 0; |
| 634 | |
| 635 | /* |
| 636 | * Collect opinions on the set config. If this is the initial |
| 637 | * config (raidctl -C), treat all labels as invalid, since |
| 638 | * there may be random data present. |
| 639 | */ |
| 640 | if (!force) { |
| 641 | for (col = 0; col < raidPtr->numCol; col++) { |
| 642 | if (RF_DEAD_DISK(raidPtr->Disks[col].status)) |
| 643 | continue; |
| 644 | clabel = raidget_component_label(raidPtr, col); |
| 645 | flags = clabel->parity_map_flags; |
| 646 | /* Check for use by non-parity-map kernel. */ |
| 647 | if (clabel->parity_map_modcount |
| 648 | != clabel->mod_counter) { |
| 649 | flags &= ~RF_PMLABEL_WASUSED; |
| 650 | } |
| 651 | |
| 652 | if (flags & RF_PMLABEL_VALID) { |
| 653 | g_tickms = clabel->parity_map_tickms; |
| 654 | g_ntick = clabel->parity_map_ntick; |
| 655 | regions = clabel->parity_map_regions; |
| 656 | if (g_regions == 0) |
| 657 | g_regions = regions; |
| 658 | else if (g_regions != regions) { |
| 659 | pm_zap = 1; /* important! */ |
| 660 | } |
| 661 | |
| 662 | if (flags & RF_PMLABEL_DISABLE) { |
| 663 | pm_use = 0; |
| 664 | } |
| 665 | if (!(flags & RF_PMLABEL_WASUSED)) { |
| 666 | pm_zap = 1; |
| 667 | } |
| 668 | } else { |
| 669 | pm_zap = 1; |
| 670 | } |
| 671 | } |
| 672 | } else { |
| 673 | pm_zap = 1; |
| 674 | } |
| 675 | |
| 676 | /* Finally, create and attach the parity map. */ |
| 677 | if (pm_use) { |
| 678 | params.cooldown = g_ntick; |
| 679 | params.tickms = g_tickms; |
| 680 | params.regions = g_regions; |
| 681 | |
| 682 | raidPtr->parity_map = kmem_alloc(sizeof(struct rf_paritymap), |
| 683 | KM_SLEEP); |
| 684 | if (0 != rf_paritymap_init(raidPtr->parity_map, raidPtr, |
| 685 | ¶ms)) { |
| 686 | /* It failed; do without. */ |
| 687 | kmem_free(raidPtr->parity_map, |
| 688 | sizeof(struct rf_paritymap)); |
| 689 | raidPtr->parity_map = NULL; |
| 690 | return; |
| 691 | } |
| 692 | |
| 693 | if (g_regions == 0) |
| 694 | /* Pick up the autoconfigured region count. */ |
| 695 | g_regions = raidPtr->parity_map->params.regions; |
| 696 | |
| 697 | if (pm_zap) { |
| 698 | good = raidPtr->parity_good && !force; |
| 699 | |
| 700 | if (good) |
| 701 | rf_paritymap_forceclean(raidPtr->parity_map); |
| 702 | else |
| 703 | rf_paritymap_invalidate(raidPtr->parity_map); |
| 704 | /* This needs to be on disk before WASUSED is set. */ |
| 705 | rf_paritymap_write(raidPtr->parity_map); |
| 706 | } |
| 707 | } |
| 708 | |
| 709 | /* Alter labels in-core to reflect the current view of things. */ |
| 710 | for (col = 0; col < raidPtr->numCol; col++) { |
| 711 | if (RF_DEAD_DISK(raidPtr->Disks[col].status)) |
| 712 | continue; |
| 713 | clabel = raidget_component_label(raidPtr, col); |
| 714 | |
| 715 | if (pm_use) |
| 716 | flags = RF_PMLABEL_VALID | RF_PMLABEL_WASUSED; |
| 717 | else |
| 718 | flags = RF_PMLABEL_VALID | RF_PMLABEL_DISABLE; |
| 719 | |
| 720 | clabel->parity_map_flags = flags; |
| 721 | clabel->parity_map_tickms = g_tickms; |
| 722 | clabel->parity_map_ntick = g_ntick; |
| 723 | clabel->parity_map_regions = g_regions; |
| 724 | raidflush_component_label(raidPtr, col); |
| 725 | } |
| 726 | /* Note that we're just in 'attach' here, and there won't |
| 727 | be any spare disks at this point. */ |
| 728 | } |
| 729 | |
| 730 | /* |
| 731 | * For initializing the parity-map fields of a component label, both on |
| 732 | * initial creation and on reconstruct/copyback/etc. */ |
| 733 | void |
| 734 | rf_paritymap_init_label(struct rf_paritymap *pm, RF_ComponentLabel_t *clabel) |
| 735 | { |
| 736 | if (pm != NULL) { |
| 737 | clabel->parity_map_flags = |
| 738 | RF_PMLABEL_VALID | RF_PMLABEL_WASUSED; |
| 739 | clabel->parity_map_tickms = pm->params.tickms; |
| 740 | clabel->parity_map_ntick = pm->params.cooldown; |
| 741 | /* |
| 742 | * XXXjld: If the number of regions is changed on disk, and |
| 743 | * then a new component is labeled before the next configure, |
| 744 | * then it will get the old value and they will conflict on |
| 745 | * the next boot (and the default will be used instead). |
| 746 | */ |
| 747 | clabel->parity_map_regions = pm->params.regions; |
| 748 | } else { |
| 749 | /* |
| 750 | * XXXjld: if the map is disabled, and all the components are |
| 751 | * replaced without an intervening unconfigure/reconfigure, |
| 752 | * then it will become enabled on the next unconfig/reconfig. |
| 753 | */ |
| 754 | } |
| 755 | } |
| 756 | |
| 757 | |
| 758 | /* Will the parity map be disabled next time? */ |
| 759 | int |
| 760 | rf_paritymap_get_disable(RF_Raid_t *raidPtr) |
| 761 | { |
| 762 | RF_ComponentLabel_t *clabel; |
| 763 | RF_RowCol_t col; |
| 764 | int dis; |
| 765 | |
| 766 | dis = 0; |
| 767 | for (col = 0; col < raidPtr->numCol; col++) { |
| 768 | if (RF_DEAD_DISK(raidPtr->Disks[col].status)) |
| 769 | continue; |
| 770 | clabel = raidget_component_label(raidPtr, col); |
| 771 | if (clabel->parity_map_flags & RF_PMLABEL_DISABLE) |
| 772 | dis = 1; |
| 773 | } |
| 774 | for (col = 0; col < raidPtr->numSpare; col++) { |
| 775 | if (raidPtr->Disks[raidPtr->numCol+col].status != rf_ds_used_spare) |
| 776 | continue; |
| 777 | clabel = raidget_component_label(raidPtr, raidPtr->numCol+col); |
| 778 | if (clabel->parity_map_flags & RF_PMLABEL_DISABLE) |
| 779 | dis = 1; |
| 780 | } |
| 781 | |
| 782 | return dis; |
| 783 | } |
| 784 | |
| 785 | /* Set whether the parity map will be disabled next time. */ |
| 786 | void |
| 787 | rf_paritymap_set_disable(RF_Raid_t *raidPtr, int dis) |
| 788 | { |
| 789 | RF_ComponentLabel_t *clabel; |
| 790 | RF_RowCol_t col; |
| 791 | |
| 792 | for (col = 0; col < raidPtr->numCol; col++) { |
| 793 | if (RF_DEAD_DISK(raidPtr->Disks[col].status)) |
| 794 | continue; |
| 795 | clabel = raidget_component_label(raidPtr, col); |
| 796 | if (dis) |
| 797 | clabel->parity_map_flags |= RF_PMLABEL_DISABLE; |
| 798 | else |
| 799 | clabel->parity_map_flags &= ~RF_PMLABEL_DISABLE; |
| 800 | raidflush_component_label(raidPtr, col); |
| 801 | } |
| 802 | |
| 803 | /* update any used spares as well */ |
| 804 | for (col = 0; col < raidPtr->numSpare; col++) { |
| 805 | if (raidPtr->Disks[raidPtr->numCol+col].status != rf_ds_used_spare) |
| 806 | continue; |
| 807 | |
| 808 | clabel = raidget_component_label(raidPtr, raidPtr->numCol+col); |
| 809 | if (dis) |
| 810 | clabel->parity_map_flags |= RF_PMLABEL_DISABLE; |
| 811 | else |
| 812 | clabel->parity_map_flags &= ~RF_PMLABEL_DISABLE; |
| 813 | raidflush_component_label(raidPtr, raidPtr->numCol+col); |
| 814 | } |
| 815 | } |
| 816 | |