rf_diskqueue.c revision 1.23
1/* $NetBSD: rf_diskqueue.c,v 1.23 2003/12/29 02:38:17 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 1.23 2003/12/29 02:38:17 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_freelist.h" 81#include "rf_debugprint.h" 82#include "rf_shutdown.h" 83#include "rf_cvscan.h" 84#include "rf_sstf.h" 85#include "rf_fifo.h" 86#include "rf_kintf.h" 87 88static int init_dqd(RF_DiskQueueData_t *); 89static void clean_dqd(RF_DiskQueueData_t *); 90static void rf_ShutdownDiskQueueSystem(void *); 91 92#ifndef RF_DEBUG_DISKQUEUE 93#define RF_DEBUG_DISKQUEUE 0 94#endif 95 96#if RF_DEBUG_DISKQUEUE 97#define Dprintf1(s,a) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL) 98#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) 99#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) 100#else 101#define Dprintf1(s,a) 102#define Dprintf2(s,a,b) 103#define Dprintf3(s,a,b,c) 104#endif 105 106/***************************************************************************** 107 * 108 * the disk queue switch defines all the functions used in the 109 * different queueing disciplines queue ID, init routine, enqueue 110 * routine, dequeue routine 111 * 112 ****************************************************************************/ 113 114static const RF_DiskQueueSW_t diskqueuesw[] = { 115 {"fifo", /* FIFO */ 116 rf_FifoCreate, 117 rf_FifoEnqueue, 118 rf_FifoDequeue, 119 rf_FifoPeek, 120 rf_FifoPromote}, 121 122 {"cvscan", /* cvscan */ 123 rf_CvscanCreate, 124 rf_CvscanEnqueue, 125 rf_CvscanDequeue, 126 rf_CvscanPeek, 127 rf_CvscanPromote}, 128 129 {"sstf", /* shortest seek time first */ 130 rf_SstfCreate, 131 rf_SstfEnqueue, 132 rf_SstfDequeue, 133 rf_SstfPeek, 134 rf_SstfPromote}, 135 136 {"scan", /* SCAN (two-way elevator) */ 137 rf_ScanCreate, 138 rf_SstfEnqueue, 139 rf_ScanDequeue, 140 rf_ScanPeek, 141 rf_SstfPromote}, 142 143 {"cscan", /* CSCAN (one-way elevator) */ 144 rf_CscanCreate, 145 rf_SstfEnqueue, 146 rf_CscanDequeue, 147 rf_CscanPeek, 148 rf_SstfPromote}, 149 150}; 151#define NUM_DISK_QUEUE_TYPES (sizeof(diskqueuesw)/sizeof(RF_DiskQueueSW_t)) 152 153static RF_FreeList_t *rf_dqd_freelist; 154 155#define RF_MAX_FREE_DQD 256 156#define RF_DQD_INC 16 157#define RF_DQD_INITIAL 64 158 159#include <sys/buf.h> 160 161static int 162init_dqd(dqd) 163 RF_DiskQueueData_t *dqd; 164{ 165 166 dqd->bp = (struct buf *) malloc(sizeof(struct buf), 167 M_RAIDFRAME, M_NOWAIT); 168 if (dqd->bp == NULL) { 169 return (ENOMEM); 170 } 171 memset(dqd->bp, 0, sizeof(struct buf)); /* if you don't do it, nobody 172 * else will.. */ 173 return (0); 174} 175 176static void 177clean_dqd(dqd) 178 RF_DiskQueueData_t *dqd; 179{ 180 free(dqd->bp, M_RAIDFRAME); 181} 182/* configures a single disk queue */ 183 184int 185rf_ConfigureDiskQueue( 186 RF_Raid_t * raidPtr, 187 RF_DiskQueue_t * diskqueue, 188 RF_RowCol_t c, 189 const RF_DiskQueueSW_t * p, 190 RF_SectorCount_t sectPerDisk, 191 dev_t dev, 192 int maxOutstanding, 193 RF_ShutdownList_t ** listp, 194 RF_AllocListElem_t * clList) 195{ 196 int rc; 197 198 diskqueue->col = c; 199 diskqueue->qPtr = p; 200 diskqueue->qHdr = (p->Create) (sectPerDisk, clList, listp); 201 diskqueue->dev = dev; 202 diskqueue->numOutstanding = 0; 203 diskqueue->queueLength = 0; 204 diskqueue->maxOutstanding = maxOutstanding; 205 diskqueue->curPriority = RF_IO_NORMAL_PRIORITY; 206 diskqueue->nextLockingOp = NULL; 207 diskqueue->numWaiting = 0; 208 diskqueue->flags = 0; 209 diskqueue->raidPtr = raidPtr; 210 diskqueue->rf_cinfo = &raidPtr->raid_cinfo[c]; 211 rc = rf_create_managed_mutex(listp, &diskqueue->mutex); 212 if (rc) { 213 rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc); 214 return (rc); 215 } 216 rc = rf_create_managed_cond(listp, &diskqueue->cond); 217 if (rc) { 218 rf_print_unable_to_init_cond(__FILE__, __LINE__, rc); 219 return (rc); 220 } 221 return (0); 222} 223 224static void 225rf_ShutdownDiskQueueSystem(ignored) 226 void *ignored; 227{ 228 RF_FREELIST_DESTROY_CLEAN(rf_dqd_freelist, next, (RF_DiskQueueData_t *), clean_dqd); 229} 230 231int 232rf_ConfigureDiskQueueSystem(listp) 233 RF_ShutdownList_t **listp; 234{ 235 int rc; 236 237 RF_FREELIST_CREATE(rf_dqd_freelist, RF_MAX_FREE_DQD, 238 RF_DQD_INC, sizeof(RF_DiskQueueData_t)); 239 if (rf_dqd_freelist == NULL) 240 return (ENOMEM); 241 rc = rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL); 242 if (rc) { 243 rf_print_unable_to_add_shutdown( __FILE__, __LINE__, rc); 244 rf_ShutdownDiskQueueSystem(NULL); 245 return (rc); 246 } 247 RF_FREELIST_PRIME_INIT(rf_dqd_freelist, RF_DQD_INITIAL, next, 248 (RF_DiskQueueData_t *), init_dqd); 249 return (0); 250} 251 252int 253rf_ConfigureDiskQueues( 254 RF_ShutdownList_t ** listp, 255 RF_Raid_t * raidPtr, 256 RF_Config_t * cfgPtr) 257{ 258 RF_DiskQueue_t *diskQueues, *spareQueues; 259 const RF_DiskQueueSW_t *p; 260 RF_RowCol_t r,c; 261 int rc, i; 262 263 raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs; 264 265 for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) { 266 if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) { 267 p = &diskqueuesw[i]; 268 break; 269 } 270 } 271 if (p == NULL) { 272 RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType); 273 p = &diskqueuesw[0]; 274 } 275 raidPtr->qType = p; 276 277 RF_CallocAndAdd(diskQueues, raidPtr->numCol + RF_MAXSPARE, 278 sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *), 279 raidPtr->cleanupList); 280 if (diskQueues == NULL) 281 return (ENOMEM); 282 raidPtr->Queues = diskQueues; 283 284 for (c = 0; c < raidPtr->numCol; c++) { 285 rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[c], 286 c, p, 287 raidPtr->sectorsPerDisk, 288 raidPtr->Disks[c].dev, 289 cfgPtr->maxOutstandingDiskReqs, 290 listp, raidPtr->cleanupList); 291 if (rc) 292 return (rc); 293 } 294 295 spareQueues = &raidPtr->Queues[raidPtr->numCol]; 296 for (r = 0; r < raidPtr->numSpare; r++) { 297 rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r], 298 raidPtr->numCol + r, p, 299 raidPtr->sectorsPerDisk, 300 raidPtr->Disks[raidPtr->numCol + r].dev, 301 cfgPtr->maxOutstandingDiskReqs, listp, 302 raidPtr->cleanupList); 303 if (rc) 304 return (rc); 305 } 306 return (0); 307} 308/* Enqueue a disk I/O 309 * 310 * Unfortunately, we have to do things differently in the different 311 * environments (simulator, user-level, kernel). 312 * At user level, all I/O is blocking, so we have 1 or more threads/disk 313 * and the thread that enqueues is different from the thread that dequeues. 314 * In the kernel, I/O is non-blocking and so we'd like to have multiple 315 * I/Os outstanding on the physical disks when possible. 316 * 317 * when any request arrives at a queue, we have two choices: 318 * dispatch it to the lower levels 319 * queue it up 320 * 321 * kernel rules for when to do what: 322 * locking request: queue empty => dispatch and lock queue, 323 * else queue it 324 * unlocking req : always dispatch it 325 * normal req : queue empty => dispatch it & set priority 326 * queue not full & priority is ok => dispatch it 327 * else queue it 328 * 329 * user-level rules: 330 * always enqueue. In the special case of an unlocking op, enqueue 331 * in a special way that will cause the unlocking op to be the next 332 * thing dequeued. 333 * 334 * simulator rules: 335 * Do the same as at user level, with the sleeps and wakeups suppressed. 336 */ 337void 338rf_DiskIOEnqueue(queue, req, pri) 339 RF_DiskQueue_t *queue; 340 RF_DiskQueueData_t *req; 341 int pri; 342{ 343 RF_ETIMER_START(req->qtime); 344 RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector); 345 req->priority = pri; 346 347#if RF_DEBUG_DISKQUEUE 348 if (rf_queueDebug && (req->numSector == 0)) { 349 printf("Warning: Enqueueing zero-sector access\n"); 350 } 351#endif 352 /* 353 * kernel 354 */ 355 RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue"); 356 /* locking request */ 357 if (RF_LOCKING_REQ(req)) { 358 if (RF_QUEUE_EMPTY(queue)) { 359 Dprintf2("Dispatching pri %d locking op to c %d (queue empty)\n", pri, queue->col); 360 RF_LOCK_QUEUE(queue); 361 rf_DispatchKernelIO(queue, req); 362 } else { 363 queue->queueLength++; /* increment count of number 364 * of requests waiting in this 365 * queue */ 366 Dprintf2("Enqueueing pri %d locking op to c %d (queue not empty)\n", pri, queue->col); 367 req->queue = (void *) queue; 368 (queue->qPtr->Enqueue) (queue->qHdr, req, pri); 369 } 370 } 371 /* unlocking request */ 372 else 373 if (RF_UNLOCKING_REQ(req)) { /* we'll do the actual unlock 374 * when this I/O completes */ 375 Dprintf2("Dispatching pri %d unlocking op to c %d\n", pri, queue->col); 376 RF_ASSERT(RF_QUEUE_LOCKED(queue)); 377 rf_DispatchKernelIO(queue, req); 378 } 379 /* normal request */ 380 else 381 if (RF_OK_TO_DISPATCH(queue, req)) { 382 Dprintf2("Dispatching pri %d regular op to c %d (ok to dispatch)\n", pri, queue->col); 383 rf_DispatchKernelIO(queue, req); 384 } else { 385 queue->queueLength++; /* increment count of 386 * number of requests 387 * waiting in this queue */ 388 Dprintf2("Enqueueing pri %d regular op to c %d (not ok to dispatch)\n", pri, queue->col); 389 req->queue = (void *) queue; 390 (queue->qPtr->Enqueue) (queue->qHdr, req, pri); 391 } 392 RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue"); 393} 394 395 396/* get the next set of I/Os started, kernel version only */ 397void 398rf_DiskIOComplete(queue, req, status) 399 RF_DiskQueue_t *queue; 400 RF_DiskQueueData_t *req; 401 int status; 402{ 403 int done = 0; 404 405 RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete"); 406 407 /* unlock the queue: (1) after an unlocking req completes (2) after a 408 * locking req fails */ 409 if (RF_UNLOCKING_REQ(req) || (RF_LOCKING_REQ(req) && status)) { 410 Dprintf1("DiskIOComplete: unlocking queue at c %d\n", queue->col); 411 RF_ASSERT(RF_QUEUE_LOCKED(queue)); 412 RF_UNLOCK_QUEUE(queue); 413 } 414 queue->numOutstanding--; 415 RF_ASSERT(queue->numOutstanding >= 0); 416 417 /* dispatch requests to the disk until we find one that we can't. */ 418 /* no reason to continue once we've filled up the queue */ 419 /* no reason to even start if the queue is locked */ 420 421 while (!done && !RF_QUEUE_FULL(queue) && !RF_QUEUE_LOCKED(queue)) { 422 if (queue->nextLockingOp) { 423 req = queue->nextLockingOp; 424 queue->nextLockingOp = NULL; 425 Dprintf2("DiskIOComplete: a pri %d locking req was pending at c %d\n", req->priority, queue->col); 426 } else { 427 req = (queue->qPtr->Dequeue) (queue->qHdr); 428 if (req != NULL) { 429 Dprintf2("DiskIOComplete: extracting pri %d req from queue at c %d\n", req->priority, queue->col); 430 } else { 431 Dprintf1("DiskIOComplete: no more requests to extract.\n", ""); 432 } 433 } 434 if (req) { 435 queue->queueLength--; /* decrement count of number 436 * of requests waiting in this 437 * queue */ 438 RF_ASSERT(queue->queueLength >= 0); 439 } 440 if (!req) 441 done = 1; 442 else 443 if (RF_LOCKING_REQ(req)) { 444 if (RF_QUEUE_EMPTY(queue)) { /* dispatch it */ 445 Dprintf2("DiskIOComplete: dispatching pri %d locking req to c %d (queue empty)\n", req->priority, queue->col); 446 RF_LOCK_QUEUE(queue); 447 rf_DispatchKernelIO(queue, req); 448 done = 1; 449 } else { /* put it aside to wait for 450 * the queue to drain */ 451 Dprintf2("DiskIOComplete: postponing pri %d locking req to c %d\n", req->priority, queue->col); 452 RF_ASSERT(queue->nextLockingOp == NULL); 453 queue->nextLockingOp = req; 454 done = 1; 455 } 456 } else 457 if (RF_UNLOCKING_REQ(req)) { /* should not happen: 458 * unlocking ops should 459 * not get queued */ 460 RF_ASSERT(RF_QUEUE_LOCKED(queue)); /* support it anyway for 461 * the future */ 462 Dprintf2("DiskIOComplete: dispatching pri %d unl req to c %d (SHOULD NOT SEE THIS)\n", req->priority, queue->col); 463 rf_DispatchKernelIO(queue, req); 464 done = 1; 465 } else 466 if (RF_OK_TO_DISPATCH(queue, req)) { 467 Dprintf2("DiskIOComplete: dispatching pri %d regular req to c %d (ok to dispatch)\n", req->priority, queue->col); 468 rf_DispatchKernelIO(queue, req); 469 } else { /* we can't dispatch it, 470 * so just re-enqueue 471 * it. */ 472 /* potential trouble here if 473 * disk queues batch reqs */ 474 Dprintf2("DiskIOComplete: re-enqueueing pri %d regular req to c %d\n", req->priority, queue->col); 475 queue->queueLength++; 476 (queue->qPtr->Enqueue) (queue->qHdr, req, req->priority); 477 done = 1; 478 } 479 } 480 481 RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete"); 482} 483/* promotes accesses tagged with the given parityStripeID from low priority 484 * to normal priority. This promotion is optional, meaning that a queue 485 * need not implement it. If there is no promotion routine associated with 486 * a queue, this routine does nothing and returns -1. 487 */ 488int 489rf_DiskIOPromote(queue, parityStripeID, which_ru) 490 RF_DiskQueue_t *queue; 491 RF_StripeNum_t parityStripeID; 492 RF_ReconUnitNum_t which_ru; 493{ 494 int retval; 495 496 if (!queue->qPtr->Promote) 497 return (-1); 498 RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote"); 499 retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru); 500 RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote"); 501 return (retval); 502} 503 504RF_DiskQueueData_t * 505rf_CreateDiskQueueData( 506 RF_IoType_t typ, 507 RF_SectorNum_t ssect, 508 RF_SectorCount_t nsect, 509 caddr_t buf, 510 RF_StripeNum_t parityStripeID, 511 RF_ReconUnitNum_t which_ru, 512 int (*wakeF) (void *, int), 513 void *arg, 514 RF_DiskQueueData_t * next, 515 RF_AccTraceEntry_t * tracerec, 516 void *raidPtr, 517 RF_DiskQueueDataFlags_t flags, 518 void *kb_proc) 519{ 520 RF_DiskQueueData_t *p; 521 522 RF_FREELIST_GET_INIT(rf_dqd_freelist, p, next, (RF_DiskQueueData_t *), init_dqd); 523 524 p->sectorOffset = ssect + rf_protectedSectors; 525 p->numSector = nsect; 526 p->type = typ; 527 p->buf = buf; 528 p->parityStripeID = parityStripeID; 529 p->which_ru = which_ru; 530 p->CompleteFunc = wakeF; 531 p->argument = arg; 532 p->next = next; 533 p->tracerec = tracerec; 534 p->priority = RF_IO_NORMAL_PRIORITY; 535 p->raidPtr = raidPtr; 536 p->flags = flags; 537 p->b_proc = kb_proc; 538 return (p); 539} 540 541void 542rf_FreeDiskQueueData(p) 543 RF_DiskQueueData_t *p; 544{ 545 RF_FREELIST_FREE_CLEAN(rf_dqd_freelist, p, next, clean_dqd); 546} 547