1 /* $NetBSD: rf_diskqueue.c,v 1.52 2009/03/23 18:38:54 oster Exp $ */ 2 /* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: Mark Holland 7 * 8 * Permission to use, copy, modify and distribute this software and 9 * its documentation is hereby granted, provided that both the copyright 10 * notice and this permission notice appear in all copies of the 11 * software, derivative works or modified versions, and any portions 12 * thereof, and that both notices appear in supporting documentation. 13 * 14 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 15 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 16 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 17 * 18 * Carnegie Mellon requests users of this software to return to 19 * 20 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 21 * School of Computer Science 22 * Carnegie Mellon University 23 * Pittsburgh PA 15213-3890 24 * 25 * any improvements or extensions that they make and grant Carnegie the 26 * rights to redistribute these changes. 27 */ 28 29 /**************************************************************************** 30 * 31 * rf_diskqueue.c -- higher-level disk queue code 32 * 33 * the routines here are a generic wrapper around the actual queueing 34 * routines. The code here implements thread scheduling, synchronization, 35 * and locking ops (see below) on top of the lower-level queueing code. 36 * 37 * to support atomic RMW, we implement "locking operations". When a 38 * locking op is dispatched to the lower levels of the driver, the 39 * queue is locked, and no further I/Os are dispatched until the queue 40 * receives & completes a corresponding "unlocking operation". This 41 * code relies on the higher layers to guarantee that a locking op 42 * will always be eventually followed by an unlocking op. The model 43 * is that the higher layers are structured so locking and unlocking 44 * ops occur in pairs, i.e. an unlocking op cannot be generated until 45 * after a locking op reports completion. There is no good way to 46 * check to see that an unlocking op "corresponds" to the op that 47 * currently has the queue locked, so we make no such attempt. Since 48 * by definition there can be only one locking op outstanding on a 49 * disk, this should not be a problem. 50 * 51 * In the kernel, we allow multiple I/Os to be concurrently dispatched 52 * to the disk driver. In order to support locking ops in this 53 * environment, when we decide to do a locking op, we stop dispatching 54 * new I/Os and wait until all dispatched I/Os have completed before 55 * dispatching the locking op. 56 * 57 * Unfortunately, the code is different in the 3 different operating 58 * states (user level, kernel, simulator). In the kernel, I/O is 59 * non-blocking, and we have no disk threads to dispatch for us. 60 * Therefore, we have to dispatch new I/Os to the scsi driver at the 61 * time of enqueue, and also at the time of completion. At user 62 * level, I/O is blocking, and so only the disk threads may dispatch 63 * I/Os. Thus at user level, all we can do at enqueue time is enqueue 64 * and wake up the disk thread to do the dispatch. 65 * 66 ****************************************************************************/ 67 68 #include <sys/cdefs.h> 69 __KERNEL_RCSID(0, "$NetBSD: rf_diskqueue.c,v 1.52 2009/03/23 18:38:54 oster Exp $"); 70 71 #include <dev/raidframe/raidframevar.h> 72 73 #include "rf_threadstuff.h" 74 #include "rf_raid.h" 75 #include "rf_diskqueue.h" 76 #include "rf_alloclist.h" 77 #include "rf_acctrace.h" 78 #include "rf_etimer.h" 79 #include "rf_general.h" 80 #include "rf_debugprint.h" 81 #include "rf_shutdown.h" 82 #include "rf_cvscan.h" 83 #include "rf_sstf.h" 84 #include "rf_fifo.h" 85 #include "rf_kintf.h" 86 87 static void rf_ShutdownDiskQueueSystem(void *); 88 89 #ifndef RF_DEBUG_DISKQUEUE 90 #define RF_DEBUG_DISKQUEUE 0 91 #endif 92 93 #if RF_DEBUG_DISKQUEUE 94 #define Dprintf1(s,a) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL) 95 #define Dprintf2(s,a,b) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),NULL,NULL,NULL,NULL,NULL,NULL) 96 #define Dprintf3(s,a,b,c) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),NULL,NULL,NULL,NULL,NULL) 97 #else 98 #define Dprintf1(s,a) 99 #define Dprintf2(s,a,b) 100 #define Dprintf3(s,a,b,c) 101 #endif 102 103 /***************************************************************************** 104 * 105 * the disk queue switch defines all the functions used in the 106 * different queueing disciplines queue ID, init routine, enqueue 107 * routine, dequeue routine 108 * 109 ****************************************************************************/ 110 111 static const RF_DiskQueueSW_t diskqueuesw[] = { 112 {"fifo", /* FIFO */ 113 rf_FifoCreate, 114 rf_FifoEnqueue, 115 rf_FifoDequeue, 116 rf_FifoPeek, 117 rf_FifoPromote}, 118 119 {"cvscan", /* cvscan */ 120 rf_CvscanCreate, 121 rf_CvscanEnqueue, 122 rf_CvscanDequeue, 123 rf_CvscanPeek, 124 rf_CvscanPromote}, 125 126 {"sstf", /* shortest seek time first */ 127 rf_SstfCreate, 128 rf_SstfEnqueue, 129 rf_SstfDequeue, 130 rf_SstfPeek, 131 rf_SstfPromote}, 132 133 {"scan", /* SCAN (two-way elevator) */ 134 rf_ScanCreate, 135 rf_SstfEnqueue, 136 rf_ScanDequeue, 137 rf_ScanPeek, 138 rf_SstfPromote}, 139 140 {"cscan", /* CSCAN (one-way elevator) */ 141 rf_CscanCreate, 142 rf_SstfEnqueue, 143 rf_CscanDequeue, 144 rf_CscanPeek, 145 rf_SstfPromote}, 146 147 }; 148 #define NUM_DISK_QUEUE_TYPES (sizeof(diskqueuesw)/sizeof(RF_DiskQueueSW_t)) 149 150 #define RF_MAX_FREE_DQD 256 151 #define RF_MIN_FREE_DQD 64 152 153 #include <sys/buf.h> 154 155 /* configures a single disk queue */ 156 157 int 158 rf_ConfigureDiskQueue(RF_Raid_t *raidPtr, RF_DiskQueue_t *diskqueue, 159 RF_RowCol_t c, const RF_DiskQueueSW_t *p, 160 RF_SectorCount_t sectPerDisk, dev_t dev, 161 int maxOutstanding, RF_ShutdownList_t **listp, 162 RF_AllocListElem_t *clList) 163 { 164 diskqueue->col = c; 165 diskqueue->qPtr = p; 166 diskqueue->qHdr = (p->Create) (sectPerDisk, clList, listp); 167 diskqueue->dev = dev; 168 diskqueue->numOutstanding = 0; 169 diskqueue->queueLength = 0; 170 diskqueue->maxOutstanding = maxOutstanding; 171 diskqueue->curPriority = RF_IO_NORMAL_PRIORITY; 172 diskqueue->flags = 0; 173 diskqueue->raidPtr = raidPtr; 174 diskqueue->rf_cinfo = &raidPtr->raid_cinfo[c]; 175 rf_mutex_init(&diskqueue->mutex); 176 diskqueue->cond = 0; 177 return (0); 178 } 179 180 static void 181 rf_ShutdownDiskQueueSystem(void *ignored) 182 { 183 pool_destroy(&rf_pools.dqd); 184 } 185 186 int 187 rf_ConfigureDiskQueueSystem(RF_ShutdownList_t **listp) 188 { 189 190 rf_pool_init(&rf_pools.dqd, sizeof(RF_DiskQueueData_t), 191 "rf_dqd_pl", RF_MIN_FREE_DQD, RF_MAX_FREE_DQD); 192 rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL); 193 194 return (0); 195 } 196 197 int 198 rf_ConfigureDiskQueues(RF_ShutdownList_t **listp, RF_Raid_t *raidPtr, 199 RF_Config_t *cfgPtr) 200 { 201 RF_DiskQueue_t *diskQueues, *spareQueues; 202 const RF_DiskQueueSW_t *p; 203 RF_RowCol_t r,c; 204 int rc, i; 205 206 raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs; 207 208 for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) { 209 if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) { 210 p = &diskqueuesw[i]; 211 break; 212 } 213 } 214 if (p == NULL) { 215 RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType); 216 p = &diskqueuesw[0]; 217 } 218 raidPtr->qType = p; 219 220 RF_MallocAndAdd(diskQueues, 221 (raidPtr->numCol + RF_MAXSPARE) * 222 sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *), 223 raidPtr->cleanupList); 224 if (diskQueues == NULL) 225 return (ENOMEM); 226 raidPtr->Queues = diskQueues; 227 228 for (c = 0; c < raidPtr->numCol; c++) { 229 rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[c], 230 c, p, 231 raidPtr->sectorsPerDisk, 232 raidPtr->Disks[c].dev, 233 cfgPtr->maxOutstandingDiskReqs, 234 listp, raidPtr->cleanupList); 235 if (rc) 236 return (rc); 237 } 238 239 spareQueues = &raidPtr->Queues[raidPtr->numCol]; 240 for (r = 0; r < raidPtr->numSpare; r++) { 241 rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r], 242 raidPtr->numCol + r, p, 243 raidPtr->sectorsPerDisk, 244 raidPtr->Disks[raidPtr->numCol + r].dev, 245 cfgPtr->maxOutstandingDiskReqs, listp, 246 raidPtr->cleanupList); 247 if (rc) 248 return (rc); 249 } 250 return (0); 251 } 252 /* Enqueue a disk I/O 253 * 254 * In the kernel, I/O is non-blocking and so we'd like to have multiple 255 * I/Os outstanding on the physical disks when possible. 256 * 257 * when any request arrives at a queue, we have two choices: 258 * dispatch it to the lower levels 259 * queue it up 260 * 261 * kernel rules for when to do what: 262 * unlocking req : always dispatch it 263 * normal req : queue empty => dispatch it & set priority 264 * queue not full & priority is ok => dispatch it 265 * else queue it 266 */ 267 void 268 rf_DiskIOEnqueue(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int pri) 269 { 270 RF_ETIMER_START(req->qtime); 271 RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector); 272 req->priority = pri; 273 274 #if RF_DEBUG_DISKQUEUE 275 if (rf_queueDebug && (req->numSector == 0)) { 276 printf("Warning: Enqueueing zero-sector access\n"); 277 } 278 #endif 279 RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue"); 280 if (RF_OK_TO_DISPATCH(queue, req)) { 281 Dprintf2("Dispatching pri %d regular op to c %d (ok to dispatch)\n", pri, queue->col); 282 rf_DispatchKernelIO(queue, req); 283 } else { 284 queue->queueLength++; /* increment count of number of requests waiting in this queue */ 285 Dprintf2("Enqueueing pri %d regular op to c %d (not ok to dispatch)\n", pri, queue->col); 286 req->queue = (void *) queue; 287 (queue->qPtr->Enqueue) (queue->qHdr, req, pri); 288 } 289 RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue"); 290 } 291 292 293 /* get the next set of I/Os started */ 294 void 295 rf_DiskIOComplete(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int status) 296 { 297 int done = 0; 298 299 RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete"); 300 queue->numOutstanding--; 301 RF_ASSERT(queue->numOutstanding >= 0); 302 303 /* dispatch requests to the disk until we find one that we can't. */ 304 /* no reason to continue once we've filled up the queue */ 305 /* no reason to even start if the queue is locked */ 306 307 while (!done && !RF_QUEUE_FULL(queue)) { 308 req = (queue->qPtr->Dequeue) (queue->qHdr); 309 if (req) { 310 Dprintf2("DiskIOComplete: extracting pri %d req from queue at c %d\n", req->priority, queue->col); 311 queue->queueLength--; /* decrement count of number of requests waiting in this queue */ 312 RF_ASSERT(queue->queueLength >= 0); 313 if (RF_OK_TO_DISPATCH(queue, req)) { 314 Dprintf2("DiskIOComplete: dispatching pri %d regular req to c %d (ok to dispatch)\n", req->priority, queue->col); 315 rf_DispatchKernelIO(queue, req); 316 } else { 317 /* we can't dispatch it, so just re-enqueue it. 318 potential trouble here if disk queues batch reqs */ 319 Dprintf2("DiskIOComplete: re-enqueueing pri %d regular req to c %d\n", req->priority, queue->col); 320 queue->queueLength++; 321 (queue->qPtr->Enqueue) (queue->qHdr, req, req->priority); 322 done = 1; 323 } 324 } else { 325 Dprintf1("DiskIOComplete: no more requests to extract.\n", ""); 326 done = 1; 327 } 328 } 329 330 RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete"); 331 } 332 /* promotes accesses tagged with the given parityStripeID from low priority 333 * to normal priority. This promotion is optional, meaning that a queue 334 * need not implement it. If there is no promotion routine associated with 335 * a queue, this routine does nothing and returns -1. 336 */ 337 int 338 rf_DiskIOPromote(RF_DiskQueue_t *queue, RF_StripeNum_t parityStripeID, 339 RF_ReconUnitNum_t which_ru) 340 { 341 int retval; 342 343 if (!queue->qPtr->Promote) 344 return (-1); 345 RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote"); 346 retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru); 347 RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote"); 348 return (retval); 349 } 350 351 RF_DiskQueueData_t * 352 rf_CreateDiskQueueData(RF_IoType_t typ, RF_SectorNum_t ssect, 353 RF_SectorCount_t nsect, void *bf, 354 RF_StripeNum_t parityStripeID, 355 RF_ReconUnitNum_t which_ru, 356 int (*wakeF) (void *, int), void *arg, 357 RF_AccTraceEntry_t *tracerec, RF_Raid_t *raidPtr, 358 RF_DiskQueueDataFlags_t flags, void *kb_proc, 359 int waitflag) 360 { 361 RF_DiskQueueData_t *p; 362 363 p = pool_get(&rf_pools.dqd, waitflag); 364 if (p == NULL) 365 return (NULL); 366 367 memset(p, 0, sizeof(RF_DiskQueueData_t)); 368 if (waitflag == PR_WAITOK) { 369 p->bp = getiobuf(NULL, true); 370 } else { 371 p->bp = getiobuf(NULL, false); 372 } 373 if (p->bp == NULL) { 374 pool_put(&rf_pools.dqd, p); 375 return (NULL); 376 } 377 SET(p->bp->b_cflags, BC_BUSY); /* mark buffer busy */ 378 379 p->sectorOffset = ssect + rf_protectedSectors; 380 p->numSector = nsect; 381 p->type = typ; 382 p->buf = bf; 383 p->parityStripeID = parityStripeID; 384 p->which_ru = which_ru; 385 p->CompleteFunc = wakeF; 386 p->argument = arg; 387 p->next = NULL; 388 p->tracerec = tracerec; 389 p->priority = RF_IO_NORMAL_PRIORITY; 390 p->raidPtr = raidPtr; 391 p->flags = flags; 392 p->b_proc = kb_proc; 393 return (p); 394 } 395 396 void 397 rf_FreeDiskQueueData(RF_DiskQueueData_t *p) 398 { 399 int s; 400 s = splbio(); /* XXX protect only pool_put, or neither? */ 401 putiobuf(p->bp); 402 pool_put(&rf_pools.dqd, p); 403 splx(s); 404 } 405