rf_paritymap.c revision 1.3.2.3
1 1.3.2.3 sborrill /* $NetBSD: rf_paritymap.c,v 1.3.2.3 2010/03/06 21:05:29 sborrill Exp $ */ 2 1.3.2.2 snj 3 1.3.2.2 snj /*- 4 1.3.2.2 snj * Copyright (c) 2009 Jed Davis. 5 1.3.2.2 snj * All rights reserved. 6 1.3.2.2 snj * 7 1.3.2.2 snj * Redistribution and use in source and binary forms, with or without 8 1.3.2.2 snj * modification, are permitted provided that the following conditions 9 1.3.2.2 snj * are met: 10 1.3.2.2 snj * 1. Redistributions of source code must retain the above copyright 11 1.3.2.2 snj * notice, this list of conditions and the following disclaimer. 12 1.3.2.2 snj * 2. Redistributions in binary form must reproduce the above copyright 13 1.3.2.2 snj * notice, this list of conditions and the following disclaimer in the 14 1.3.2.2 snj * documentation and/or other materials provided with the distribution. 15 1.3.2.2 snj * 16 1.3.2.2 snj * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 17 1.3.2.2 snj * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 18 1.3.2.2 snj * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 19 1.3.2.2 snj * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 20 1.3.2.2 snj * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 21 1.3.2.2 snj * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 22 1.3.2.2 snj * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 23 1.3.2.2 snj * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 24 1.3.2.2 snj * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 25 1.3.2.2 snj * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 26 1.3.2.2 snj * POSSIBILITY OF SUCH DAMAGE. 27 1.3.2.2 snj */ 28 1.3.2.2 snj 29 1.3.2.2 snj #include <sys/cdefs.h> 30 1.3.2.3 sborrill __KERNEL_RCSID(0, "$NetBSD: rf_paritymap.c,v 1.3.2.3 2010/03/06 21:05:29 sborrill Exp $"); 31 1.3.2.2 snj 32 1.3.2.2 snj #include <sys/param.h> 33 1.3.2.2 snj #include <sys/callout.h> 34 1.3.2.2 snj #include <sys/kmem.h> 35 1.3.2.2 snj #include <sys/mutex.h> 36 1.3.2.2 snj #include <sys/rwlock.h> 37 1.3.2.2 snj #include <sys/systm.h> 38 1.3.2.2 snj #include <sys/types.h> 39 1.3.2.2 snj 40 1.3.2.2 snj #include <dev/raidframe/rf_paritymap.h> 41 1.3.2.2 snj #include <dev/raidframe/rf_stripelocks.h> 42 1.3.2.2 snj #include <dev/raidframe/rf_layout.h> 43 1.3.2.2 snj #include <dev/raidframe/rf_raid.h> 44 1.3.2.2 snj #include <dev/raidframe/rf_parityscan.h> 45 1.3.2.2 snj #include <dev/raidframe/rf_kintf.h> 46 1.3.2.2 snj 47 1.3.2.2 snj /* Important parameters: */ 48 1.3.2.2 snj #define REGION_MINSIZE (25ULL << 20) 49 1.3.2.2 snj #define DFL_TICKMS 40000 50 1.3.2.2 snj #define DFL_COOLDOWN 8 /* 7-8 intervals of 40s = 5min +/- 20s */ 51 1.3.2.2 snj 52 1.3.2.2 snj /* Internal-use flag bits. */ 53 1.3.2.2 snj #define TICKING 1 54 1.3.2.2 snj #define TICKED 2 55 1.3.2.2 snj 56 1.3.2.2 snj /* Prototypes! */ 57 1.3.2.2 snj static void rf_paritymap_write_locked(struct rf_paritymap *); 58 1.3.2.2 snj static void rf_paritymap_tick(void *); 59 1.3.2.2 snj static u_int rf_paritymap_nreg(RF_Raid_t *); 60 1.3.2.2 snj 61 1.3.2.2 snj /* Extract the current status of the parity map. */ 62 1.3.2.2 snj void 63 1.3.2.2 snj rf_paritymap_status(struct rf_paritymap *pm, struct rf_pmstat *ps) 64 1.3.2.2 snj { 65 1.3.2.2 snj memset(ps, 0, sizeof(*ps)); 66 1.3.2.2 snj if (pm == NULL) 67 1.3.2.2 snj ps->enabled = 0; 68 1.3.2.2 snj else { 69 1.3.2.2 snj ps->enabled = 1; 70 1.3.2.2 snj ps->region_size = pm->region_size; 71 1.3.2.2 snj mutex_enter(&pm->lock); 72 1.3.2.2 snj memcpy(&ps->params, &pm->params, sizeof(ps->params)); 73 1.3.2.2 snj memcpy(ps->dirty, pm->disk_now, sizeof(ps->dirty)); 74 1.3.2.2 snj memcpy(&ps->ctrs, &pm->ctrs, sizeof(ps->ctrs)); 75 1.3.2.2 snj mutex_exit(&pm->lock); 76 1.3.2.2 snj } 77 1.3.2.2 snj } 78 1.3.2.2 snj 79 1.3.2.2 snj /* 80 1.3.2.2 snj * Test whether parity in a given sector is suspected of being inconsistent 81 1.3.2.2 snj * on disk (assuming that any pending I/O to it is allowed to complete). 82 1.3.2.2 snj * This may be of interest to future work on parity scrubbing. 83 1.3.2.2 snj */ 84 1.3.2.2 snj int 85 1.3.2.2 snj rf_paritymap_test(struct rf_paritymap *pm, daddr_t sector) 86 1.3.2.2 snj { 87 1.3.2.2 snj unsigned region = sector / pm->region_size; 88 1.3.2.2 snj int retval; 89 1.3.2.2 snj 90 1.3.2.2 snj mutex_enter(&pm->lock); 91 1.3.2.2 snj retval = isset(pm->disk_boot->bits, region) ? 1 : 0; 92 1.3.2.2 snj mutex_exit(&pm->lock); 93 1.3.2.2 snj return retval; 94 1.3.2.2 snj } 95 1.3.2.2 snj 96 1.3.2.2 snj /* To be called before a write to the RAID is submitted. */ 97 1.3.2.2 snj void 98 1.3.2.2 snj rf_paritymap_begin(struct rf_paritymap *pm, daddr_t offset, daddr_t size) 99 1.3.2.2 snj { 100 1.3.2.2 snj unsigned i, b, e; 101 1.3.2.2 snj 102 1.3.2.2 snj b = offset / pm->region_size; 103 1.3.2.2 snj e = (offset + size - 1) / pm->region_size; 104 1.3.2.2 snj 105 1.3.2.2 snj for (i = b; i <= e; i++) 106 1.3.2.2 snj rf_paritymap_begin_region(pm, i); 107 1.3.2.2 snj } 108 1.3.2.2 snj 109 1.3.2.2 snj /* To be called after a write to the RAID completes. */ 110 1.3.2.2 snj void 111 1.3.2.2 snj rf_paritymap_end(struct rf_paritymap *pm, daddr_t offset, daddr_t size) 112 1.3.2.2 snj { 113 1.3.2.2 snj unsigned i, b, e; 114 1.3.2.2 snj 115 1.3.2.2 snj b = offset / pm->region_size; 116 1.3.2.2 snj e = (offset + size - 1) / pm->region_size; 117 1.3.2.2 snj 118 1.3.2.2 snj for (i = b; i <= e; i++) 119 1.3.2.2 snj rf_paritymap_end_region(pm, i); 120 1.3.2.2 snj } 121 1.3.2.2 snj 122 1.3.2.2 snj void 123 1.3.2.2 snj rf_paritymap_begin_region(struct rf_paritymap *pm, unsigned region) 124 1.3.2.2 snj { 125 1.3.2.2 snj int needs_write; 126 1.3.2.2 snj 127 1.3.2.2 snj KASSERT(region < RF_PARITYMAP_NREG); 128 1.3.2.2 snj pm->ctrs.nwrite++; 129 1.3.2.2 snj 130 1.3.2.2 snj /* If it was being kept warm, deal with that. */ 131 1.3.2.2 snj mutex_enter(&pm->lock); 132 1.3.2.2 snj if (pm->current->state[region] < 0) 133 1.3.2.2 snj pm->current->state[region] = 0; 134 1.3.2.2 snj 135 1.3.2.2 snj /* This shouldn't happen unless RAIDOUTSTANDING is set too high. */ 136 1.3.2.2 snj KASSERT(pm->current->state[region] < 127); 137 1.3.2.2 snj pm->current->state[region]++; 138 1.3.2.2 snj 139 1.3.2.2 snj needs_write = isclr(pm->disk_now->bits, region); 140 1.3.2.2 snj 141 1.3.2.2 snj if (needs_write) { 142 1.3.2.2 snj KASSERT(pm->current->state[region] == 1); 143 1.3.2.2 snj rf_paritymap_write_locked(pm); 144 1.3.2.2 snj } 145 1.3.2.2 snj 146 1.3.2.2 snj mutex_exit(&pm->lock); 147 1.3.2.2 snj } 148 1.3.2.2 snj 149 1.3.2.2 snj void 150 1.3.2.2 snj rf_paritymap_end_region(struct rf_paritymap *pm, unsigned region) 151 1.3.2.2 snj { 152 1.3.2.2 snj KASSERT(region < RF_PARITYMAP_NREG); 153 1.3.2.2 snj 154 1.3.2.2 snj mutex_enter(&pm->lock); 155 1.3.2.2 snj KASSERT(pm->current->state[region] > 0); 156 1.3.2.2 snj --pm->current->state[region]; 157 1.3.2.2 snj 158 1.3.2.2 snj if (pm->current->state[region] <= 0) { 159 1.3.2.2 snj pm->current->state[region] = -pm->params.cooldown; 160 1.3.2.2 snj KASSERT(pm->current->state[region] <= 0); 161 1.3.2.2 snj mutex_enter(&pm->lk_flags); 162 1.3.2.2 snj if (!(pm->flags & TICKING)) { 163 1.3.2.2 snj pm->flags |= TICKING; 164 1.3.2.2 snj mutex_exit(&pm->lk_flags); 165 1.3.2.2 snj callout_schedule(&pm->ticker, 166 1.3.2.2 snj mstohz(pm->params.tickms)); 167 1.3.2.2 snj } else 168 1.3.2.2 snj mutex_exit(&pm->lk_flags); 169 1.3.2.2 snj } 170 1.3.2.2 snj mutex_exit(&pm->lock); 171 1.3.2.2 snj } 172 1.3.2.2 snj 173 1.3.2.2 snj /* 174 1.3.2.2 snj * Updates the parity map to account for any changes in current activity 175 1.3.2.2 snj * and/or an ongoing parity scan, then writes it to disk with appropriate 176 1.3.2.2 snj * synchronization. 177 1.3.2.2 snj */ 178 1.3.2.2 snj void 179 1.3.2.2 snj rf_paritymap_write(struct rf_paritymap *pm) 180 1.3.2.2 snj { 181 1.3.2.2 snj mutex_enter(&pm->lock); 182 1.3.2.2 snj rf_paritymap_write_locked(pm); 183 1.3.2.2 snj mutex_exit(&pm->lock); 184 1.3.2.2 snj } 185 1.3.2.2 snj 186 1.3.2.2 snj /* As above, but to be used when pm->lock is already held. */ 187 1.3.2.2 snj static void 188 1.3.2.2 snj rf_paritymap_write_locked(struct rf_paritymap *pm) 189 1.3.2.2 snj { 190 1.3.2.2 snj char w, w0; 191 1.3.2.2 snj int i, j, setting, clearing; 192 1.3.2.2 snj 193 1.3.2.2 snj setting = clearing = 0; 194 1.3.2.2 snj for (i = 0; i < RF_PARITYMAP_NBYTE; i++) { 195 1.3.2.2 snj w0 = pm->disk_now->bits[i]; 196 1.3.2.2 snj w = pm->disk_boot->bits[i]; 197 1.3.2.2 snj 198 1.3.2.2 snj for (j = 0; j < NBBY; j++) 199 1.3.2.2 snj if (pm->current->state[i * NBBY + j] != 0) 200 1.3.2.2 snj w |= 1 << j; 201 1.3.2.2 snj 202 1.3.2.2 snj if (w & ~w0) 203 1.3.2.2 snj setting = 1; 204 1.3.2.2 snj if (w0 & ~w) 205 1.3.2.2 snj clearing = 1; 206 1.3.2.2 snj 207 1.3.2.2 snj pm->disk_now->bits[i] = w; 208 1.3.2.2 snj } 209 1.3.2.2 snj pm->ctrs.ncachesync += setting + clearing; 210 1.3.2.2 snj pm->ctrs.nclearing += clearing; 211 1.3.2.2 snj 212 1.3.2.2 snj /* 213 1.3.2.2 snj * If bits are being set in the parity map, then a sync is 214 1.3.2.2 snj * required afterwards, so that the regions are marked dirty 215 1.3.2.2 snj * on disk before any writes to them take place. If bits are 216 1.3.2.2 snj * being cleared, then a sync is required before the write, so 217 1.3.2.2 snj * that any writes to those regions are processed before the 218 1.3.2.2 snj * region is marked clean. (Synchronization is somewhat 219 1.3.2.2 snj * overkill; a write ordering barrier would suffice, but we 220 1.3.2.2 snj * currently have no way to express that directly.) 221 1.3.2.2 snj */ 222 1.3.2.2 snj if (clearing) 223 1.3.2.2 snj rf_sync_component_caches(pm->raid); 224 1.3.2.2 snj rf_paritymap_kern_write(pm->raid, pm->disk_now); 225 1.3.2.2 snj if (setting) 226 1.3.2.2 snj rf_sync_component_caches(pm->raid); 227 1.3.2.2 snj } 228 1.3.2.2 snj 229 1.3.2.2 snj /* Mark all parity as being in need of rewrite. */ 230 1.3.2.2 snj void 231 1.3.2.2 snj rf_paritymap_invalidate(struct rf_paritymap *pm) 232 1.3.2.2 snj { 233 1.3.2.2 snj mutex_enter(&pm->lock); 234 1.3.2.2 snj memset(pm->disk_boot, ~(unsigned char)0, 235 1.3.2.2 snj sizeof(struct rf_paritymap_ondisk)); 236 1.3.2.2 snj mutex_exit(&pm->lock); 237 1.3.2.2 snj } 238 1.3.2.2 snj 239 1.3.2.2 snj /* Mark all parity as being correct. */ 240 1.3.2.2 snj void 241 1.3.2.2 snj rf_paritymap_forceclean(struct rf_paritymap *pm) 242 1.3.2.2 snj { 243 1.3.2.2 snj mutex_enter(&pm->lock); 244 1.3.2.2 snj memset(pm->disk_boot, (unsigned char)0, 245 1.3.2.2 snj sizeof(struct rf_paritymap_ondisk)); 246 1.3.2.2 snj mutex_exit(&pm->lock); 247 1.3.2.2 snj } 248 1.3.2.2 snj 249 1.3.2.2 snj /* 250 1.3.2.2 snj * The cooldown callout routine just defers its work to a thread; it can't do 251 1.3.2.2 snj * the parity map write itself as it would block, and although mutex-induced 252 1.3.2.2 snj * blocking is permitted it seems wise to avoid tying up the softint. 253 1.3.2.2 snj */ 254 1.3.2.2 snj static void 255 1.3.2.2 snj rf_paritymap_tick(void *arg) 256 1.3.2.2 snj { 257 1.3.2.2 snj struct rf_paritymap *pm = arg; 258 1.3.2.2 snj 259 1.3.2.2 snj mutex_enter(&pm->lk_flags); 260 1.3.2.2 snj pm->flags |= TICKED; 261 1.3.2.2 snj mutex_exit(&pm->lk_flags); 262 1.3.2.2 snj wakeup(&(pm->raid->iodone)); /* XXX */ 263 1.3.2.2 snj } 264 1.3.2.2 snj 265 1.3.2.2 snj /* 266 1.3.2.2 snj * This is where the parity cooling work (and rearming the callout if needed) 267 1.3.2.2 snj * is done; the raidio thread calls it when woken up, as by the above. 268 1.3.2.2 snj */ 269 1.3.2.2 snj void 270 1.3.2.2 snj rf_paritymap_checkwork(struct rf_paritymap *pm) 271 1.3.2.2 snj { 272 1.3.2.2 snj int i, zerop, progressp; 273 1.3.2.2 snj 274 1.3.2.2 snj mutex_enter(&pm->lk_flags); 275 1.3.2.2 snj if (pm->flags & TICKED) { 276 1.3.2.2 snj zerop = progressp = 0; 277 1.3.2.2 snj 278 1.3.2.2 snj pm->flags &= ~TICKED; 279 1.3.2.2 snj mutex_exit(&pm->lk_flags); 280 1.3.2.2 snj 281 1.3.2.2 snj mutex_enter(&pm->lock); 282 1.3.2.2 snj for (i = 0; i < RF_PARITYMAP_NREG; i++) { 283 1.3.2.2 snj if (pm->current->state[i] < 0) { 284 1.3.2.2 snj progressp = 1; 285 1.3.2.2 snj pm->current->state[i]++; 286 1.3.2.2 snj if (pm->current->state[i] == 0) 287 1.3.2.2 snj zerop = 1; 288 1.3.2.2 snj } 289 1.3.2.2 snj } 290 1.3.2.2 snj 291 1.3.2.2 snj if (progressp) 292 1.3.2.2 snj callout_schedule(&pm->ticker, 293 1.3.2.2 snj mstohz(pm->params.tickms)); 294 1.3.2.2 snj else { 295 1.3.2.2 snj mutex_enter(&pm->lk_flags); 296 1.3.2.2 snj pm->flags &= ~TICKING; 297 1.3.2.2 snj mutex_exit(&pm->lk_flags); 298 1.3.2.2 snj } 299 1.3.2.2 snj 300 1.3.2.2 snj if (zerop) 301 1.3.2.2 snj rf_paritymap_write_locked(pm); 302 1.3.2.2 snj mutex_exit(&pm->lock); 303 1.3.2.2 snj } else 304 1.3.2.2 snj mutex_exit(&pm->lk_flags); 305 1.3.2.2 snj } 306 1.3.2.2 snj 307 1.3.2.2 snj /* 308 1.3.2.2 snj * Set parity map parameters; used both to alter parameters on the fly and to 309 1.3.2.2 snj * establish their initial values. Note that setting a parameter to 0 means 310 1.3.2.2 snj * to leave the previous setting unchanged, and that if this is done for the 311 1.3.2.2 snj * initial setting of "regions", then a default value will be computed based 312 1.3.2.2 snj * on the RAID component size. 313 1.3.2.2 snj */ 314 1.3.2.2 snj int 315 1.3.2.2 snj rf_paritymap_set_params(struct rf_paritymap *pm, 316 1.3.2.2 snj const struct rf_pmparams *params, int todisk) 317 1.3.2.2 snj { 318 1.3.2.2 snj int cooldown, tickms; 319 1.3.2.2 snj u_int regions; 320 1.3.2.2 snj RF_RowCol_t col; 321 1.3.2.2 snj RF_ComponentLabel_t *clabel; 322 1.3.2.2 snj RF_Raid_t *raidPtr; 323 1.3.2.2 snj 324 1.3.2.2 snj cooldown = params->cooldown != 0 325 1.3.2.2 snj ? params->cooldown : pm->params.cooldown; 326 1.3.2.2 snj tickms = params->tickms != 0 327 1.3.2.2 snj ? params->tickms : pm->params.tickms; 328 1.3.2.2 snj regions = params->regions != 0 329 1.3.2.2 snj ? params->regions : pm->params.regions; 330 1.3.2.2 snj 331 1.3.2.2 snj if (cooldown < 1 || cooldown > 128) { 332 1.3.2.2 snj printf("raid%d: cooldown %d out of range\n", pm->raid->raidid, 333 1.3.2.2 snj cooldown); 334 1.3.2.2 snj return (-1); 335 1.3.2.2 snj } 336 1.3.2.2 snj if (tickms < 10) { 337 1.3.2.2 snj printf("raid%d: tick time %dms out of range\n", 338 1.3.2.2 snj pm->raid->raidid, tickms); 339 1.3.2.2 snj return (-1); 340 1.3.2.2 snj } 341 1.3.2.2 snj if (regions == 0) { 342 1.3.2.2 snj regions = rf_paritymap_nreg(pm->raid); 343 1.3.2.2 snj } else if (regions > RF_PARITYMAP_NREG) { 344 1.3.2.2 snj printf("raid%d: region count %u too large (more than %u)\n", 345 1.3.2.2 snj pm->raid->raidid, regions, RF_PARITYMAP_NREG); 346 1.3.2.2 snj return (-1); 347 1.3.2.2 snj } 348 1.3.2.2 snj 349 1.3.2.2 snj /* XXX any currently warm parity will be used with the new tickms! */ 350 1.3.2.2 snj pm->params.cooldown = cooldown; 351 1.3.2.2 snj pm->params.tickms = tickms; 352 1.3.2.2 snj /* Apply the initial region count, but do not change it after that. */ 353 1.3.2.2 snj if (pm->params.regions == 0) 354 1.3.2.2 snj pm->params.regions = regions; 355 1.3.2.2 snj 356 1.3.2.2 snj /* So that the newly set parameters can be tested: */ 357 1.3.2.2 snj pm->ctrs.nwrite = pm->ctrs.ncachesync = pm->ctrs.nclearing = 0; 358 1.3.2.2 snj 359 1.3.2.2 snj if (todisk) { 360 1.3.2.2 snj raidPtr = pm->raid; 361 1.3.2.2 snj for (col = 0; col < raidPtr->numCol; col++) { 362 1.3.2.3 sborrill if (RF_DEAD_DISK(raidPtr->Disks[col].status)) 363 1.3.2.3 sborrill continue; 364 1.3.2.3 sborrill 365 1.3.2.2 snj clabel = raidget_component_label(raidPtr, col); 366 1.3.2.2 snj clabel->parity_map_ntick = cooldown; 367 1.3.2.2 snj clabel->parity_map_tickms = tickms; 368 1.3.2.2 snj clabel->parity_map_regions = regions; 369 1.3.2.3 sborrill 370 1.3.2.3 sborrill /* Don't touch the disk if it's been spared */ 371 1.3.2.3 sborrill if (clabel->status == rf_ds_spared) 372 1.3.2.3 sborrill continue; 373 1.3.2.3 sborrill 374 1.3.2.2 snj raidflush_component_label(raidPtr, col); 375 1.3.2.2 snj } 376 1.3.2.3 sborrill 377 1.3.2.3 sborrill /* handle the spares too... */ 378 1.3.2.3 sborrill for (col = 0; col < raidPtr->numSpare; col++) { 379 1.3.2.3 sborrill if (raidPtr->Disks[raidPtr->numCol+col].status == rf_ds_used_spare) { 380 1.3.2.3 sborrill clabel = raidget_component_label(raidPtr, raidPtr->numCol+col); 381 1.3.2.3 sborrill clabel->parity_map_ntick = cooldown; 382 1.3.2.3 sborrill clabel->parity_map_tickms = tickms; 383 1.3.2.3 sborrill clabel->parity_map_regions = regions; 384 1.3.2.3 sborrill raidflush_component_label(raidPtr, raidPtr->numCol+col); 385 1.3.2.3 sborrill } 386 1.3.2.3 sborrill } 387 1.3.2.2 snj } 388 1.3.2.2 snj return 0; 389 1.3.2.2 snj } 390 1.3.2.2 snj 391 1.3.2.2 snj /* 392 1.3.2.2 snj * The number of regions may not be as many as can fit into the map, because 393 1.3.2.2 snj * when regions are too small, the overhead of setting parity map bits 394 1.3.2.2 snj * becomes significant in comparison to the actual I/O, while the 395 1.3.2.2 snj * corresponding gains in parity verification time become negligible. Thus, 396 1.3.2.2 snj * a minimum region size (defined above) is imposed. 397 1.3.2.2 snj * 398 1.3.2.2 snj * Note that, if the number of regions is less than the maximum, then some of 399 1.3.2.2 snj * the regions will be "fictional", corresponding to no actual disk; some 400 1.3.2.2 snj * parts of the code may process them as normal, but they can not ever be 401 1.3.2.2 snj * written to. 402 1.3.2.2 snj */ 403 1.3.2.2 snj static u_int 404 1.3.2.2 snj rf_paritymap_nreg(RF_Raid_t *raid) 405 1.3.2.2 snj { 406 1.3.2.2 snj daddr_t bytes_per_disk, nreg; 407 1.3.2.2 snj 408 1.3.2.2 snj bytes_per_disk = raid->sectorsPerDisk << raid->logBytesPerSector; 409 1.3.2.2 snj nreg = bytes_per_disk / REGION_MINSIZE; 410 1.3.2.2 snj if (nreg > RF_PARITYMAP_NREG) 411 1.3.2.2 snj nreg = RF_PARITYMAP_NREG; 412 1.3.2.2 snj 413 1.3.2.2 snj return (u_int)nreg; 414 1.3.2.2 snj } 415 1.3.2.2 snj 416 1.3.2.2 snj /* 417 1.3.2.2 snj * Initialize a parity map given specific parameters. This neither reads nor 418 1.3.2.2 snj * writes the parity map config in the component labels; for that, see below. 419 1.3.2.2 snj */ 420 1.3.2.2 snj int 421 1.3.2.2 snj rf_paritymap_init(struct rf_paritymap *pm, RF_Raid_t *raid, 422 1.3.2.2 snj const struct rf_pmparams *params) 423 1.3.2.2 snj { 424 1.3.2.2 snj daddr_t rstripes; 425 1.3.2.2 snj struct rf_pmparams safe; 426 1.3.2.2 snj 427 1.3.2.2 snj pm->raid = raid; 428 1.3.2.2 snj pm->params.regions = 0; 429 1.3.2.2 snj if (0 != rf_paritymap_set_params(pm, params, 0)) { 430 1.3.2.2 snj /* 431 1.3.2.2 snj * If the parameters are out-of-range, then bring the 432 1.3.2.2 snj * parity map up with something reasonable, so that 433 1.3.2.2 snj * the admin can at least go and fix it (or ignore it 434 1.3.2.2 snj * entirely). 435 1.3.2.2 snj */ 436 1.3.2.2 snj safe.cooldown = DFL_COOLDOWN; 437 1.3.2.2 snj safe.tickms = DFL_TICKMS; 438 1.3.2.2 snj safe.regions = 0; 439 1.3.2.2 snj 440 1.3.2.2 snj if (0 != rf_paritymap_set_params(pm, &safe, 0)) 441 1.3.2.2 snj return (-1); 442 1.3.2.2 snj } 443 1.3.2.2 snj 444 1.3.2.2 snj rstripes = howmany(raid->Layout.numStripe, pm->params.regions); 445 1.3.2.2 snj pm->region_size = rstripes * raid->Layout.dataSectorsPerStripe; 446 1.3.2.2 snj 447 1.3.2.2 snj callout_init(&pm->ticker, CALLOUT_MPSAFE); 448 1.3.2.2 snj callout_setfunc(&pm->ticker, rf_paritymap_tick, pm); 449 1.3.2.2 snj pm->flags = 0; 450 1.3.2.2 snj 451 1.3.2.2 snj pm->disk_boot = kmem_alloc(sizeof(struct rf_paritymap_ondisk), 452 1.3.2.2 snj KM_SLEEP); 453 1.3.2.2 snj pm->disk_now = kmem_alloc(sizeof(struct rf_paritymap_ondisk), 454 1.3.2.2 snj KM_SLEEP); 455 1.3.2.2 snj pm->current = kmem_zalloc(sizeof(struct rf_paritymap_current), 456 1.3.2.2 snj KM_SLEEP); 457 1.3.2.2 snj 458 1.3.2.2 snj rf_paritymap_kern_read(pm->raid, pm->disk_boot); 459 1.3.2.2 snj memcpy(pm->disk_now, pm->disk_boot, sizeof(*pm->disk_now)); 460 1.3.2.2 snj 461 1.3.2.2 snj mutex_init(&pm->lock, MUTEX_DEFAULT, IPL_NONE); 462 1.3.2.2 snj mutex_init(&pm->lk_flags, MUTEX_DEFAULT, IPL_SOFTCLOCK); 463 1.3.2.2 snj 464 1.3.2.2 snj return 0; 465 1.3.2.2 snj } 466 1.3.2.2 snj 467 1.3.2.2 snj /* 468 1.3.2.2 snj * Destroys a parity map; unless "force" is set, also cleans parity for any 469 1.3.2.2 snj * regions which were still in cooldown (but are not dirty on disk). 470 1.3.2.2 snj */ 471 1.3.2.2 snj void 472 1.3.2.2 snj rf_paritymap_destroy(struct rf_paritymap *pm, int force) 473 1.3.2.2 snj { 474 1.3.2.2 snj int i; 475 1.3.2.2 snj 476 1.3.2.2 snj callout_halt(&pm->ticker, NULL); /* XXX stop? halt? */ 477 1.3.2.2 snj callout_destroy(&pm->ticker); 478 1.3.2.2 snj 479 1.3.2.2 snj if (!force) { 480 1.3.2.2 snj for (i = 0; i < RF_PARITYMAP_NREG; i++) { 481 1.3.2.2 snj /* XXX check for > 0 ? */ 482 1.3.2.2 snj if (pm->current->state[i] < 0) 483 1.3.2.2 snj pm->current->state[i] = 0; 484 1.3.2.2 snj } 485 1.3.2.2 snj 486 1.3.2.2 snj rf_paritymap_write_locked(pm); 487 1.3.2.2 snj } 488 1.3.2.2 snj 489 1.3.2.2 snj mutex_destroy(&pm->lock); 490 1.3.2.2 snj mutex_destroy(&pm->lk_flags); 491 1.3.2.2 snj 492 1.3.2.2 snj kmem_free(pm->disk_boot, sizeof(struct rf_paritymap_ondisk)); 493 1.3.2.2 snj kmem_free(pm->disk_now, sizeof(struct rf_paritymap_ondisk)); 494 1.3.2.2 snj kmem_free(pm->current, sizeof(struct rf_paritymap_current)); 495 1.3.2.2 snj } 496 1.3.2.2 snj 497 1.3.2.2 snj /* 498 1.3.2.2 snj * Rewrite parity, taking parity map into account; this is the equivalent of 499 1.3.2.2 snj * the old rf_RewriteParity, and is likewise to be called from a suitable 500 1.3.2.2 snj * thread and shouldn't have multiple copies running in parallel and so on. 501 1.3.2.2 snj * 502 1.3.2.2 snj * Note that the fictional regions are "cleaned" in one shot, so that very 503 1.3.2.2 snj * small RAIDs (useful for testing) will not experience potentially severe 504 1.3.2.2 snj * regressions in rewrite time. 505 1.3.2.2 snj */ 506 1.3.2.2 snj int 507 1.3.2.2 snj rf_paritymap_rewrite(struct rf_paritymap *pm) 508 1.3.2.2 snj { 509 1.3.2.2 snj int i, ret_val = 0; 510 1.3.2.2 snj daddr_t reg_b, reg_e; 511 1.3.2.2 snj 512 1.3.2.2 snj /* Process only the actual regions. */ 513 1.3.2.2 snj for (i = 0; i < pm->params.regions; i++) { 514 1.3.2.2 snj mutex_enter(&pm->lock); 515 1.3.2.2 snj if (isset(pm->disk_boot->bits, i)) { 516 1.3.2.2 snj mutex_exit(&pm->lock); 517 1.3.2.2 snj 518 1.3.2.2 snj reg_b = i * pm->region_size; 519 1.3.2.2 snj reg_e = reg_b + pm->region_size; 520 1.3.2.2 snj if (reg_e > pm->raid->totalSectors) 521 1.3.2.2 snj reg_e = pm->raid->totalSectors; 522 1.3.2.2 snj 523 1.3.2.2 snj if (rf_RewriteParityRange(pm->raid, reg_b, 524 1.3.2.2 snj reg_e - reg_b)) { 525 1.3.2.2 snj ret_val = 1; 526 1.3.2.2 snj if (pm->raid->waitShutdown) 527 1.3.2.2 snj return ret_val; 528 1.3.2.2 snj } else { 529 1.3.2.2 snj mutex_enter(&pm->lock); 530 1.3.2.2 snj clrbit(pm->disk_boot->bits, i); 531 1.3.2.2 snj rf_paritymap_write_locked(pm); 532 1.3.2.2 snj mutex_exit(&pm->lock); 533 1.3.2.2 snj } 534 1.3.2.2 snj } else { 535 1.3.2.2 snj mutex_exit(&pm->lock); 536 1.3.2.2 snj } 537 1.3.2.2 snj } 538 1.3.2.2 snj 539 1.3.2.2 snj /* Now, clear the fictional regions, if any. */ 540 1.3.2.2 snj rf_paritymap_forceclean(pm); 541 1.3.2.2 snj rf_paritymap_write(pm); 542 1.3.2.2 snj 543 1.3.2.2 snj return ret_val; 544 1.3.2.2 snj } 545 1.3.2.2 snj 546 1.3.2.2 snj /* 547 1.3.2.2 snj * How to merge the on-disk parity maps when reading them in from the 548 1.3.2.2 snj * various components; returns whether they differ. In the case that 549 1.3.2.2 snj * they do differ, sets *dst to the union of *dst and *src. 550 1.3.2.2 snj * 551 1.3.2.2 snj * In theory, it should be safe to take the intersection (or just pick 552 1.3.2.2 snj * a single component arbitrarily), but the paranoid approach costs 553 1.3.2.2 snj * little. 554 1.3.2.2 snj * 555 1.3.2.2 snj * Appropriate locking, if any, is the responsibility of the caller. 556 1.3.2.2 snj */ 557 1.3.2.2 snj int 558 1.3.2.2 snj rf_paritymap_merge(struct rf_paritymap_ondisk *dst, 559 1.3.2.2 snj struct rf_paritymap_ondisk *src) 560 1.3.2.2 snj { 561 1.3.2.2 snj int i, discrep = 0; 562 1.3.2.2 snj 563 1.3.2.2 snj for (i = 0; i < RF_PARITYMAP_NBYTE; i++) { 564 1.3.2.2 snj if (dst->bits[i] != src->bits[i]) 565 1.3.2.2 snj discrep = 1; 566 1.3.2.2 snj dst->bits[i] |= src->bits[i]; 567 1.3.2.2 snj } 568 1.3.2.2 snj 569 1.3.2.2 snj return discrep; 570 1.3.2.2 snj } 571 1.3.2.2 snj 572 1.3.2.2 snj /* 573 1.3.2.2 snj * Detach a parity map from its RAID. This is not meant to be applied except 574 1.3.2.2 snj * when unconfiguring the RAID after all I/O has been resolved, as otherwise 575 1.3.2.2 snj * an out-of-date parity map could be treated as current. 576 1.3.2.2 snj */ 577 1.3.2.2 snj void 578 1.3.2.2 snj rf_paritymap_detach(RF_Raid_t *raidPtr) 579 1.3.2.2 snj { 580 1.3.2.2 snj if (raidPtr->parity_map == NULL) 581 1.3.2.2 snj return; 582 1.3.2.2 snj 583 1.3.2.2 snj simple_lock(&(raidPtr->iodone_lock)); 584 1.3.2.2 snj struct rf_paritymap *pm = raidPtr->parity_map; 585 1.3.2.2 snj raidPtr->parity_map = NULL; 586 1.3.2.2 snj simple_unlock(&(raidPtr->iodone_lock)); 587 1.3.2.2 snj /* XXXjld is that enough locking? Or too much? */ 588 1.3.2.2 snj rf_paritymap_destroy(pm, 0); 589 1.3.2.2 snj kmem_free(pm, sizeof(*pm)); 590 1.3.2.2 snj } 591 1.3.2.2 snj 592 1.3.2.2 snj /* 593 1.3.2.2 snj * Attach a parity map to a RAID set if appropriate. Includes 594 1.3.2.2 snj * configure-time processing of parity-map fields of component label. 595 1.3.2.2 snj */ 596 1.3.2.2 snj void 597 1.3.2.2 snj rf_paritymap_attach(RF_Raid_t *raidPtr, int force) 598 1.3.2.2 snj { 599 1.3.2.2 snj RF_RowCol_t col; 600 1.3.2.2 snj int pm_use, pm_zap; 601 1.3.2.2 snj int g_tickms, g_ntick, g_regions; 602 1.3.2.2 snj int good; 603 1.3.2.2 snj RF_ComponentLabel_t *clabel; 604 1.3.2.2 snj u_int flags, regions; 605 1.3.2.2 snj struct rf_pmparams params; 606 1.3.2.2 snj 607 1.3.2.2 snj if (raidPtr->Layout.map->faultsTolerated == 0) { 608 1.3.2.2 snj /* There isn't any parity. */ 609 1.3.2.2 snj return; 610 1.3.2.2 snj } 611 1.3.2.2 snj 612 1.3.2.2 snj pm_use = 1; 613 1.3.2.2 snj pm_zap = 0; 614 1.3.2.2 snj g_tickms = DFL_TICKMS; 615 1.3.2.2 snj g_ntick = DFL_COOLDOWN; 616 1.3.2.2 snj g_regions = 0; 617 1.3.2.2 snj 618 1.3.2.2 snj /* 619 1.3.2.2 snj * Collect opinions on the set config. If this is the initial 620 1.3.2.2 snj * config (raidctl -C), treat all labels as invalid, since 621 1.3.2.2 snj * there may be random data present. 622 1.3.2.2 snj */ 623 1.3.2.2 snj if (!force) { 624 1.3.2.2 snj for (col = 0; col < raidPtr->numCol; col++) { 625 1.3.2.3 sborrill if (RF_DEAD_DISK(raidPtr->Disks[col].status)) 626 1.3.2.3 sborrill continue; 627 1.3.2.2 snj clabel = raidget_component_label(raidPtr, col); 628 1.3.2.2 snj flags = clabel->parity_map_flags; 629 1.3.2.2 snj /* Check for use by non-parity-map kernel. */ 630 1.3.2.2 snj if (clabel->parity_map_modcount 631 1.3.2.2 snj != clabel->mod_counter) { 632 1.3.2.2 snj flags &= ~RF_PMLABEL_WASUSED; 633 1.3.2.2 snj } 634 1.3.2.2 snj 635 1.3.2.2 snj if (flags & RF_PMLABEL_VALID) { 636 1.3.2.2 snj g_tickms = clabel->parity_map_tickms; 637 1.3.2.2 snj g_ntick = clabel->parity_map_ntick; 638 1.3.2.2 snj regions = clabel->parity_map_regions; 639 1.3.2.2 snj if (g_regions == 0) 640 1.3.2.2 snj g_regions = regions; 641 1.3.2.2 snj else if (g_regions != regions) { 642 1.3.2.2 snj pm_zap = 1; /* important! */ 643 1.3.2.2 snj } 644 1.3.2.2 snj 645 1.3.2.2 snj if (flags & RF_PMLABEL_DISABLE) { 646 1.3.2.2 snj pm_use = 0; 647 1.3.2.2 snj } 648 1.3.2.2 snj if (!(flags & RF_PMLABEL_WASUSED)) { 649 1.3.2.2 snj pm_zap = 1; 650 1.3.2.2 snj } 651 1.3.2.2 snj } else { 652 1.3.2.2 snj pm_zap = 1; 653 1.3.2.2 snj } 654 1.3.2.2 snj } 655 1.3.2.2 snj } else { 656 1.3.2.2 snj pm_zap = 1; 657 1.3.2.2 snj } 658 1.3.2.2 snj 659 1.3.2.2 snj /* Finally, create and attach the parity map. */ 660 1.3.2.2 snj if (pm_use) { 661 1.3.2.2 snj params.cooldown = g_ntick; 662 1.3.2.2 snj params.tickms = g_tickms; 663 1.3.2.2 snj params.regions = g_regions; 664 1.3.2.2 snj 665 1.3.2.2 snj raidPtr->parity_map = kmem_alloc(sizeof(struct rf_paritymap), 666 1.3.2.2 snj KM_SLEEP); 667 1.3.2.2 snj if (0 != rf_paritymap_init(raidPtr->parity_map, raidPtr, 668 1.3.2.2 snj ¶ms)) { 669 1.3.2.2 snj /* It failed; do without. */ 670 1.3.2.2 snj kmem_free(raidPtr->parity_map, 671 1.3.2.2 snj sizeof(struct rf_paritymap)); 672 1.3.2.2 snj raidPtr->parity_map = NULL; 673 1.3.2.2 snj return; 674 1.3.2.2 snj } 675 1.3.2.2 snj 676 1.3.2.2 snj if (g_regions == 0) 677 1.3.2.2 snj /* Pick up the autoconfigured region count. */ 678 1.3.2.2 snj g_regions = raidPtr->parity_map->params.regions; 679 1.3.2.2 snj 680 1.3.2.2 snj if (pm_zap) { 681 1.3.2.2 snj good = raidPtr->parity_good && !force; 682 1.3.2.2 snj 683 1.3.2.2 snj if (good) 684 1.3.2.2 snj rf_paritymap_forceclean(raidPtr->parity_map); 685 1.3.2.2 snj else 686 1.3.2.2 snj rf_paritymap_invalidate(raidPtr->parity_map); 687 1.3.2.2 snj /* This needs to be on disk before WASUSED is set. */ 688 1.3.2.2 snj rf_paritymap_write(raidPtr->parity_map); 689 1.3.2.2 snj } 690 1.3.2.2 snj } 691 1.3.2.2 snj 692 1.3.2.2 snj /* Alter labels in-core to reflect the current view of things. */ 693 1.3.2.2 snj for (col = 0; col < raidPtr->numCol; col++) { 694 1.3.2.3 sborrill if (RF_DEAD_DISK(raidPtr->Disks[col].status)) 695 1.3.2.3 sborrill continue; 696 1.3.2.2 snj clabel = raidget_component_label(raidPtr, col); 697 1.3.2.2 snj 698 1.3.2.2 snj if (pm_use) 699 1.3.2.2 snj flags = RF_PMLABEL_VALID | RF_PMLABEL_WASUSED; 700 1.3.2.2 snj else 701 1.3.2.2 snj flags = RF_PMLABEL_VALID | RF_PMLABEL_DISABLE; 702 1.3.2.2 snj 703 1.3.2.2 snj clabel->parity_map_flags = flags; 704 1.3.2.2 snj clabel->parity_map_tickms = g_tickms; 705 1.3.2.2 snj clabel->parity_map_ntick = g_ntick; 706 1.3.2.2 snj clabel->parity_map_regions = g_regions; 707 1.3.2.2 snj raidflush_component_label(raidPtr, col); 708 1.3.2.2 snj } 709 1.3.2.3 sborrill /* Note that we're just in 'attach' here, and there won't 710 1.3.2.3 sborrill be any spare disks at this point. */ 711 1.3.2.2 snj } 712 1.3.2.2 snj 713 1.3.2.2 snj /* 714 1.3.2.2 snj * For initializing the parity-map fields of a component label, both on 715 1.3.2.3 sborrill * initial creation and on reconstruct/copyback/etc. */ 716 1.3.2.2 snj void 717 1.3.2.2 snj rf_paritymap_init_label(struct rf_paritymap *pm, RF_ComponentLabel_t *clabel) 718 1.3.2.2 snj { 719 1.3.2.2 snj if (pm != NULL) { 720 1.3.2.2 snj clabel->parity_map_flags = 721 1.3.2.2 snj RF_PMLABEL_VALID | RF_PMLABEL_WASUSED; 722 1.3.2.2 snj clabel->parity_map_tickms = pm->params.tickms; 723 1.3.2.2 snj clabel->parity_map_ntick = pm->params.cooldown; 724 1.3.2.2 snj /* 725 1.3.2.2 snj * XXXjld: If the number of regions is changed on disk, and 726 1.3.2.2 snj * then a new component is labeled before the next configure, 727 1.3.2.2 snj * then it will get the old value and they will conflict on 728 1.3.2.2 snj * the next boot (and the default will be used instead). 729 1.3.2.2 snj */ 730 1.3.2.2 snj clabel->parity_map_regions = pm->params.regions; 731 1.3.2.2 snj } else { 732 1.3.2.2 snj /* 733 1.3.2.2 snj * XXXjld: if the map is disabled, and all the components are 734 1.3.2.2 snj * replaced without an intervening unconfigure/reconfigure, 735 1.3.2.2 snj * then it will become enabled on the next unconfig/reconfig. 736 1.3.2.2 snj */ 737 1.3.2.2 snj } 738 1.3.2.2 snj } 739 1.3.2.2 snj 740 1.3.2.2 snj 741 1.3.2.2 snj /* Will the parity map be disabled next time? */ 742 1.3.2.2 snj int 743 1.3.2.2 snj rf_paritymap_get_disable(RF_Raid_t *raidPtr) 744 1.3.2.2 snj { 745 1.3.2.2 snj RF_ComponentLabel_t *clabel; 746 1.3.2.2 snj RF_RowCol_t col; 747 1.3.2.2 snj int dis; 748 1.3.2.2 snj 749 1.3.2.2 snj dis = 0; 750 1.3.2.2 snj for (col = 0; col < raidPtr->numCol; col++) { 751 1.3.2.3 sborrill if (RF_DEAD_DISK(raidPtr->Disks[col].status)) 752 1.3.2.3 sborrill continue; 753 1.3.2.2 snj clabel = raidget_component_label(raidPtr, col); 754 1.3.2.2 snj if (clabel->parity_map_flags & RF_PMLABEL_DISABLE) 755 1.3.2.2 snj dis = 1; 756 1.3.2.2 snj } 757 1.3.2.3 sborrill for (col = 0; col < raidPtr->numSpare; col++) { 758 1.3.2.3 sborrill if (raidPtr->Disks[raidPtr->numCol+col].status != rf_ds_used_spare) 759 1.3.2.3 sborrill continue; 760 1.3.2.3 sborrill clabel = raidget_component_label(raidPtr, raidPtr->numCol+col); 761 1.3.2.3 sborrill if (clabel->parity_map_flags & RF_PMLABEL_DISABLE) 762 1.3.2.3 sborrill dis = 1; 763 1.3.2.3 sborrill } 764 1.3.2.2 snj 765 1.3.2.2 snj return dis; 766 1.3.2.2 snj } 767 1.3.2.2 snj 768 1.3.2.2 snj /* Set whether the parity map will be disabled next time. */ 769 1.3.2.2 snj void 770 1.3.2.2 snj rf_paritymap_set_disable(RF_Raid_t *raidPtr, int dis) 771 1.3.2.2 snj { 772 1.3.2.2 snj RF_ComponentLabel_t *clabel; 773 1.3.2.2 snj RF_RowCol_t col; 774 1.3.2.2 snj 775 1.3.2.2 snj for (col = 0; col < raidPtr->numCol; col++) { 776 1.3.2.3 sborrill if (RF_DEAD_DISK(raidPtr->Disks[col].status)) 777 1.3.2.3 sborrill continue; 778 1.3.2.2 snj clabel = raidget_component_label(raidPtr, col); 779 1.3.2.2 snj if (dis) 780 1.3.2.2 snj clabel->parity_map_flags |= RF_PMLABEL_DISABLE; 781 1.3.2.2 snj else 782 1.3.2.2 snj clabel->parity_map_flags &= ~RF_PMLABEL_DISABLE; 783 1.3.2.2 snj raidflush_component_label(raidPtr, col); 784 1.3.2.2 snj } 785 1.3.2.3 sborrill 786 1.3.2.3 sborrill /* update any used spares as well */ 787 1.3.2.3 sborrill for (col = 0; col < raidPtr->numSpare; col++) { 788 1.3.2.3 sborrill if (raidPtr->Disks[raidPtr->numCol+col].status != rf_ds_used_spare) 789 1.3.2.3 sborrill continue; 790 1.3.2.3 sborrill 791 1.3.2.3 sborrill clabel = raidget_component_label(raidPtr, raidPtr->numCol+col); 792 1.3.2.3 sborrill if (dis) 793 1.3.2.3 sborrill clabel->parity_map_flags |= RF_PMLABEL_DISABLE; 794 1.3.2.3 sborrill else 795 1.3.2.3 sborrill clabel->parity_map_flags &= ~RF_PMLABEL_DISABLE; 796 1.3.2.3 sborrill raidflush_component_label(raidPtr, raidPtr->numCol+col); 797 1.3.2.3 sborrill } 798 1.3.2.2 snj } 799
Indexes created Thu Oct 02 01:09:59 GMT 2025