Annotation of src/sys/dev/raidframe/rf_diskqueue.c, revision 1.41.2.4
1.41.2.4! yamt 1: /* $NetBSD: rf_diskqueue.c,v 1.41.2.3 2007/09/03 14:38:17 yamt Exp $ */
1.1 oster 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:
1.13 oster 29: /****************************************************************************
1.1 oster 30: *
31: * rf_diskqueue.c -- higher-level disk queue code
32: *
33: * the routines here are a generic wrapper around the actual queueing
1.6 oster 34: * routines. The code here implements thread scheduling, synchronization,
1.1 oster 35: * and locking ops (see below) on top of the lower-level queueing code.
36: *
1.13 oster 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.
1.1 oster 65: *
1.13 oster 66: ****************************************************************************/
1.15 lukem 67:
68: #include <sys/cdefs.h>
1.41.2.4! yamt 69: __KERNEL_RCSID(0, "$NetBSD: rf_diskqueue.c,v 1.41.2.3 2007/09/03 14:38:17 yamt Exp $");
1.1 oster 70:
1.14 oster 71: #include <dev/raidframe/raidframevar.h>
72:
1.1 oster 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"
1.11 oster 85: #include "rf_kintf.h"
1.1 oster 86:
87: static void rf_ShutdownDiskQueueSystem(void *);
88:
1.21 oster 89: #ifndef RF_DEBUG_DISKQUEUE
90: #define RF_DEBUG_DISKQUEUE 0
91: #endif
92:
93: #if RF_DEBUG_DISKQUEUE
1.1 oster 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)
1.21 oster 97: #else
98: #define Dprintf1(s,a)
99: #define Dprintf2(s,a,b)
100: #define Dprintf3(s,a,b,c)
101: #endif
1.1 oster 102:
1.13 oster 103: /*****************************************************************************
1.1 oster 104: *
1.13 oster 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
1.1 oster 108: *
1.13 oster 109: ****************************************************************************/
1.1 oster 110:
1.22 jdolecek 111: static const RF_DiskQueueSW_t diskqueuesw[] = {
1.6 oster 112: {"fifo", /* FIFO */
113: rf_FifoCreate,
114: rf_FifoEnqueue,
115: rf_FifoDequeue,
116: rf_FifoPeek,
1.1 oster 117: rf_FifoPromote},
118:
1.6 oster 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,
1.1 oster 131: rf_SstfPromote},
132:
1.6 oster 133: {"scan", /* SCAN (two-way elevator) */
134: rf_ScanCreate,
135: rf_SstfEnqueue,
136: rf_ScanDequeue,
137: rf_ScanPeek,
1.1 oster 138: rf_SstfPromote},
139:
1.6 oster 140: {"cscan", /* CSCAN (one-way elevator) */
141: rf_CscanCreate,
142: rf_SstfEnqueue,
143: rf_CscanDequeue,
144: rf_CscanPeek,
1.1 oster 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
1.31 oster 151: #define RF_MIN_FREE_DQD 64
1.1 oster 152:
153: #include <sys/buf.h>
154:
1.6 oster 155: /* configures a single disk queue */
1.9 oster 156:
1.40 perry 157: int
1.27 oster 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)
1.6 oster 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->nextLockingOp = NULL;
173: diskqueue->flags = 0;
174: diskqueue->raidPtr = raidPtr;
1.23 oster 175: diskqueue->rf_cinfo = &raidPtr->raid_cinfo[c];
1.25 oster 176: rf_mutex_init(&diskqueue->mutex);
1.26 oster 177: diskqueue->cond = 0;
1.6 oster 178: return (0);
1.1 oster 179: }
180:
1.40 perry 181: static void
1.27 oster 182: rf_ShutdownDiskQueueSystem(void *ignored)
1.6 oster 183: {
1.32 oster 184: pool_destroy(&rf_pools.dqd);
1.1 oster 185: }
186:
1.30 oster 187: int
1.27 oster 188: rf_ConfigureDiskQueueSystem(RF_ShutdownList_t **listp)
1.6 oster 189: {
190:
1.32 oster 191: rf_pool_init(&rf_pools.dqd, sizeof(RF_DiskQueueData_t),
192: "rf_dqd_pl", RF_MIN_FREE_DQD, RF_MAX_FREE_DQD);
1.30 oster 193: rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL);
1.24 oster 194:
1.6 oster 195: return (0);
196: }
197:
1.40 perry 198: int
1.27 oster 199: rf_ConfigureDiskQueues(RF_ShutdownList_t **listp, RF_Raid_t *raidPtr,
200: RF_Config_t *cfgPtr)
1.6 oster 201: {
1.23 oster 202: RF_DiskQueue_t *diskQueues, *spareQueues;
1.22 jdolecek 203: const RF_DiskQueueSW_t *p;
1.23 oster 204: RF_RowCol_t r,c;
1.6 oster 205: int rc, i;
206:
207: raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs;
208:
209: for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) {
210: if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) {
211: p = &diskqueuesw[i];
212: break;
213: }
214: }
215: if (p == NULL) {
216: RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType);
217: p = &diskqueuesw[0];
218: }
1.10 oster 219: raidPtr->qType = p;
1.23 oster 220:
1.40 perry 221: RF_MallocAndAdd(diskQueues,
1.24 oster 222: (raidPtr->numCol + RF_MAXSPARE) *
1.40 perry 223: sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *),
1.23 oster 224: raidPtr->cleanupList);
225: if (diskQueues == NULL)
1.6 oster 226: return (ENOMEM);
227: raidPtr->Queues = diskQueues;
1.23 oster 228:
229: for (c = 0; c < raidPtr->numCol; c++) {
230: rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[c],
231: c, p,
1.40 perry 232: raidPtr->sectorsPerDisk,
1.23 oster 233: raidPtr->Disks[c].dev,
1.40 perry 234: cfgPtr->maxOutstandingDiskReqs,
1.23 oster 235: listp, raidPtr->cleanupList);
236: if (rc)
237: return (rc);
1.6 oster 238: }
239:
1.23 oster 240: spareQueues = &raidPtr->Queues[raidPtr->numCol];
1.6 oster 241: for (r = 0; r < raidPtr->numSpare; r++) {
1.9 oster 242: rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r],
1.23 oster 243: raidPtr->numCol + r, p,
244: raidPtr->sectorsPerDisk,
245: raidPtr->Disks[raidPtr->numCol + r].dev,
246: cfgPtr->maxOutstandingDiskReqs, listp,
247: raidPtr->cleanupList);
1.6 oster 248: if (rc)
249: return (rc);
250: }
251: return (0);
252: }
1.1 oster 253: /* Enqueue a disk I/O
254: *
255: * Unfortunately, we have to do things differently in the different
256: * environments (simulator, user-level, kernel).
257: * At user level, all I/O is blocking, so we have 1 or more threads/disk
258: * and the thread that enqueues is different from the thread that dequeues.
259: * In the kernel, I/O is non-blocking and so we'd like to have multiple
260: * I/Os outstanding on the physical disks when possible.
261: *
262: * when any request arrives at a queue, we have two choices:
263: * dispatch it to the lower levels
264: * queue it up
265: *
266: * kernel rules for when to do what:
267: * locking request: queue empty => dispatch and lock queue,
268: * else queue it
269: * unlocking req : always dispatch it
270: * normal req : queue empty => dispatch it & set priority
271: * queue not full & priority is ok => dispatch it
272: * else queue it
273: *
274: * user-level rules:
275: * always enqueue. In the special case of an unlocking op, enqueue
276: * in a special way that will cause the unlocking op to be the next
277: * thing dequeued.
278: *
279: * simulator rules:
280: * Do the same as at user level, with the sleeps and wakeups suppressed.
281: */
1.40 perry 282: void
1.27 oster 283: rf_DiskIOEnqueue(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int pri)
1.6 oster 284: {
285: RF_ETIMER_START(req->qtime);
286: RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector);
287: req->priority = pri;
288:
1.21 oster 289: #if RF_DEBUG_DISKQUEUE
1.6 oster 290: if (rf_queueDebug && (req->numSector == 0)) {
291: printf("Warning: Enqueueing zero-sector access\n");
292: }
1.21 oster 293: #endif
1.6 oster 294: /*
295: * kernel
296: */
297: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
298: /* locking request */
299: if (RF_LOCKING_REQ(req)) {
300: if (RF_QUEUE_EMPTY(queue)) {
1.23 oster 301: Dprintf2("Dispatching pri %d locking op to c %d (queue empty)\n", pri, queue->col);
1.6 oster 302: RF_LOCK_QUEUE(queue);
303: rf_DispatchKernelIO(queue, req);
304: } else {
305: queue->queueLength++; /* increment count of number
306: * of requests waiting in this
307: * queue */
1.23 oster 308: Dprintf2("Enqueueing pri %d locking op to c %d (queue not empty)\n", pri, queue->col);
1.6 oster 309: req->queue = (void *) queue;
310: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
311: }
312: }
313: /* unlocking request */
314: else
315: if (RF_UNLOCKING_REQ(req)) { /* we'll do the actual unlock
316: * when this I/O completes */
1.23 oster 317: Dprintf2("Dispatching pri %d unlocking op to c %d\n", pri, queue->col);
1.6 oster 318: RF_ASSERT(RF_QUEUE_LOCKED(queue));
319: rf_DispatchKernelIO(queue, req);
320: }
321: /* normal request */
322: else
323: if (RF_OK_TO_DISPATCH(queue, req)) {
1.23 oster 324: Dprintf2("Dispatching pri %d regular op to c %d (ok to dispatch)\n", pri, queue->col);
1.6 oster 325: rf_DispatchKernelIO(queue, req);
326: } else {
327: queue->queueLength++; /* increment count of
328: * number of requests
329: * waiting in this queue */
1.23 oster 330: Dprintf2("Enqueueing pri %d regular op to c %d (not ok to dispatch)\n", pri, queue->col);
1.6 oster 331: req->queue = (void *) queue;
332: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
333: }
334: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
1.1 oster 335: }
1.6 oster 336:
1.1 oster 337:
338: /* get the next set of I/Os started, kernel version only */
1.40 perry 339: void
1.27 oster 340: rf_DiskIOComplete(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int status)
1.6 oster 341: {
342: int done = 0;
343:
344: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
345:
346: /* unlock the queue: (1) after an unlocking req completes (2) after a
347: * locking req fails */
348: if (RF_UNLOCKING_REQ(req) || (RF_LOCKING_REQ(req) && status)) {
1.23 oster 349: Dprintf1("DiskIOComplete: unlocking queue at c %d\n", queue->col);
1.20 oster 350: RF_ASSERT(RF_QUEUE_LOCKED(queue));
1.6 oster 351: RF_UNLOCK_QUEUE(queue);
352: }
353: queue->numOutstanding--;
354: RF_ASSERT(queue->numOutstanding >= 0);
355:
356: /* dispatch requests to the disk until we find one that we can't. */
357: /* no reason to continue once we've filled up the queue */
358: /* no reason to even start if the queue is locked */
359:
360: while (!done && !RF_QUEUE_FULL(queue) && !RF_QUEUE_LOCKED(queue)) {
361: if (queue->nextLockingOp) {
362: req = queue->nextLockingOp;
363: queue->nextLockingOp = NULL;
1.23 oster 364: Dprintf2("DiskIOComplete: a pri %d locking req was pending at c %d\n", req->priority, queue->col);
1.6 oster 365: } else {
366: req = (queue->qPtr->Dequeue) (queue->qHdr);
367: if (req != NULL) {
1.23 oster 368: Dprintf2("DiskIOComplete: extracting pri %d req from queue at c %d\n", req->priority, queue->col);
1.6 oster 369: } else {
370: Dprintf1("DiskIOComplete: no more requests to extract.\n", "");
371: }
372: }
373: if (req) {
374: queue->queueLength--; /* decrement count of number
375: * of requests waiting in this
376: * queue */
377: RF_ASSERT(queue->queueLength >= 0);
378: }
379: if (!req)
380: done = 1;
381: else
382: if (RF_LOCKING_REQ(req)) {
383: if (RF_QUEUE_EMPTY(queue)) { /* dispatch it */
1.23 oster 384: Dprintf2("DiskIOComplete: dispatching pri %d locking req to c %d (queue empty)\n", req->priority, queue->col);
1.6 oster 385: RF_LOCK_QUEUE(queue);
386: rf_DispatchKernelIO(queue, req);
387: done = 1;
388: } else { /* put it aside to wait for
389: * the queue to drain */
1.23 oster 390: Dprintf2("DiskIOComplete: postponing pri %d locking req to c %d\n", req->priority, queue->col);
1.6 oster 391: RF_ASSERT(queue->nextLockingOp == NULL);
392: queue->nextLockingOp = req;
393: done = 1;
394: }
395: } else
396: if (RF_UNLOCKING_REQ(req)) { /* should not happen:
397: * unlocking ops should
398: * not get queued */
399: RF_ASSERT(RF_QUEUE_LOCKED(queue)); /* support it anyway for
400: * the future */
1.23 oster 401: Dprintf2("DiskIOComplete: dispatching pri %d unl req to c %d (SHOULD NOT SEE THIS)\n", req->priority, queue->col);
1.6 oster 402: rf_DispatchKernelIO(queue, req);
403: done = 1;
404: } else
405: if (RF_OK_TO_DISPATCH(queue, req)) {
1.23 oster 406: Dprintf2("DiskIOComplete: dispatching pri %d regular req to c %d (ok to dispatch)\n", req->priority, queue->col);
1.6 oster 407: rf_DispatchKernelIO(queue, req);
408: } else { /* we can't dispatch it,
409: * so just re-enqueue
410: * it. */
411: /* potential trouble here if
412: * disk queues batch reqs */
1.23 oster 413: Dprintf2("DiskIOComplete: re-enqueueing pri %d regular req to c %d\n", req->priority, queue->col);
1.6 oster 414: queue->queueLength++;
415: (queue->qPtr->Enqueue) (queue->qHdr, req, req->priority);
416: done = 1;
417: }
418: }
419:
420: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
1.1 oster 421: }
422: /* promotes accesses tagged with the given parityStripeID from low priority
423: * to normal priority. This promotion is optional, meaning that a queue
424: * need not implement it. If there is no promotion routine associated with
425: * a queue, this routine does nothing and returns -1.
426: */
1.40 perry 427: int
1.27 oster 428: rf_DiskIOPromote(RF_DiskQueue_t *queue, RF_StripeNum_t parityStripeID,
429: RF_ReconUnitNum_t which_ru)
1.6 oster 430: {
431: int retval;
432:
433: if (!queue->qPtr->Promote)
434: return (-1);
435: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
436: retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru);
437: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
438: return (retval);
439: }
440:
441: RF_DiskQueueData_t *
1.27 oster 442: rf_CreateDiskQueueData(RF_IoType_t typ, RF_SectorNum_t ssect,
1.41.2.3 yamt 443: RF_SectorCount_t nsect, void *bf,
1.27 oster 444: RF_StripeNum_t parityStripeID,
445: RF_ReconUnitNum_t which_ru,
446: int (*wakeF) (void *, int), void *arg,
1.37 oster 447: RF_AccTraceEntry_t *tracerec, RF_Raid_t *raidPtr,
1.38 oster 448: RF_DiskQueueDataFlags_t flags, void *kb_proc,
449: int waitflag)
1.6 oster 450: {
451: RF_DiskQueueData_t *p;
452:
1.38 oster 453: p = pool_get(&rf_pools.dqd, waitflag);
454: if (p == NULL)
455: return (NULL);
456:
1.34 oster 457: memset(p, 0, sizeof(RF_DiskQueueData_t));
1.41.2.1 yamt 458: if (waitflag == PR_WAITOK) {
1.41.2.4! yamt 459: p->bp = getiobuf(NULL, true);
1.41.2.1 yamt 460: } else {
1.41.2.4! yamt 461: p->bp = getiobuf(NULL, false);
1.41.2.1 yamt 462: }
1.28 oster 463: if (p->bp == NULL) {
1.32 oster 464: pool_put(&rf_pools.dqd, p);
1.38 oster 465: return (NULL);
1.24 oster 466: }
1.6 oster 467:
468: p->sectorOffset = ssect + rf_protectedSectors;
469: p->numSector = nsect;
470: p->type = typ;
1.41 christos 471: p->buf = bf;
1.6 oster 472: p->parityStripeID = parityStripeID;
473: p->which_ru = which_ru;
474: p->CompleteFunc = wakeF;
475: p->argument = arg;
1.39 oster 476: p->next = NULL;
1.6 oster 477: p->tracerec = tracerec;
478: p->priority = RF_IO_NORMAL_PRIORITY;
479: p->raidPtr = raidPtr;
480: p->flags = flags;
481: p->b_proc = kb_proc;
482: return (p);
483: }
484:
1.40 perry 485: void
1.27 oster 486: rf_FreeDiskQueueData(RF_DiskQueueData_t *p)
1.1 oster 487: {
1.41.2.2 yamt 488: int s;
489: s = splbio(); /* XXX protect only pool_put, or neither? */
1.41.2.1 yamt 490: putiobuf(p->bp);
1.32 oster 491: pool_put(&rf_pools.dqd, p);
1.41.2.2 yamt 492: splx(s);
1.1 oster 493: }
CVSweb <webmaster@jp.NetBSD.org>