Annotation of src/sys/dev/raidframe/rf_diskqueue.c, revision 1.30
1.30 ! oster 1: /* $NetBSD: rf_diskqueue.c,v 1.29 2004/01/01 19:27:35 oster 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>
1.30 ! oster 69: __KERNEL_RCSID(0, "$NetBSD: rf_diskqueue.c,v 1.29 2004/01/01 19:27:35 oster 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:
1.24 oster 150: static struct pool rf_dqd_pool;
1.1 oster 151: #define RF_MAX_FREE_DQD 256
152: #define RF_DQD_INC 16
153: #define RF_DQD_INITIAL 64
154:
155: #include <sys/buf.h>
156:
1.6 oster 157: /* configures a single disk queue */
1.9 oster 158:
1.7 oster 159: int
1.27 oster 160: rf_ConfigureDiskQueue(RF_Raid_t *raidPtr, RF_DiskQueue_t *diskqueue,
161: RF_RowCol_t c, const RF_DiskQueueSW_t *p,
162: RF_SectorCount_t sectPerDisk, dev_t dev,
163: int maxOutstanding, RF_ShutdownList_t **listp,
164: RF_AllocListElem_t *clList)
1.6 oster 165: {
166: diskqueue->col = c;
167: diskqueue->qPtr = p;
168: diskqueue->qHdr = (p->Create) (sectPerDisk, clList, listp);
169: diskqueue->dev = dev;
170: diskqueue->numOutstanding = 0;
171: diskqueue->queueLength = 0;
172: diskqueue->maxOutstanding = maxOutstanding;
173: diskqueue->curPriority = RF_IO_NORMAL_PRIORITY;
174: diskqueue->nextLockingOp = NULL;
175: diskqueue->flags = 0;
176: diskqueue->raidPtr = raidPtr;
1.23 oster 177: diskqueue->rf_cinfo = &raidPtr->raid_cinfo[c];
1.25 oster 178: rf_mutex_init(&diskqueue->mutex);
1.26 oster 179: diskqueue->cond = 0;
1.6 oster 180: return (0);
1.1 oster 181: }
182:
1.6 oster 183: static void
1.27 oster 184: rf_ShutdownDiskQueueSystem(void *ignored)
1.6 oster 185: {
1.24 oster 186: pool_destroy(&rf_dqd_pool);
1.1 oster 187: }
188:
1.30 ! oster 189: int
1.27 oster 190: rf_ConfigureDiskQueueSystem(RF_ShutdownList_t **listp)
1.6 oster 191: {
192:
1.24 oster 193: pool_init(&rf_dqd_pool, sizeof(RF_DiskQueueData_t), 0, 0, 0,
194: "rf_dqd_pl", NULL);
195: pool_sethiwat(&rf_dqd_pool, RF_MAX_FREE_DQD);
196: pool_prime(&rf_dqd_pool, RF_DQD_INITIAL);
197:
1.30 ! oster 198: rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL);
1.24 oster 199:
1.6 oster 200: return (0);
201: }
202:
203: int
1.27 oster 204: rf_ConfigureDiskQueues(RF_ShutdownList_t **listp, RF_Raid_t *raidPtr,
205: RF_Config_t *cfgPtr)
1.6 oster 206: {
1.23 oster 207: RF_DiskQueue_t *diskQueues, *spareQueues;
1.22 jdolecek 208: const RF_DiskQueueSW_t *p;
1.23 oster 209: RF_RowCol_t r,c;
1.6 oster 210: int rc, i;
211:
212: raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs;
213:
214: for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) {
215: if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) {
216: p = &diskqueuesw[i];
217: break;
218: }
219: }
220: if (p == NULL) {
221: RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType);
222: p = &diskqueuesw[0];
223: }
1.10 oster 224: raidPtr->qType = p;
1.23 oster 225:
1.24 oster 226: RF_MallocAndAdd(diskQueues,
227: (raidPtr->numCol + RF_MAXSPARE) *
1.23 oster 228: sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *),
229: raidPtr->cleanupList);
230: if (diskQueues == NULL)
1.6 oster 231: return (ENOMEM);
232: raidPtr->Queues = diskQueues;
1.23 oster 233:
234: for (c = 0; c < raidPtr->numCol; c++) {
235: rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[c],
236: c, p,
237: raidPtr->sectorsPerDisk,
238: raidPtr->Disks[c].dev,
239: cfgPtr->maxOutstandingDiskReqs,
240: listp, raidPtr->cleanupList);
241: if (rc)
242: return (rc);
1.6 oster 243: }
244:
1.23 oster 245: spareQueues = &raidPtr->Queues[raidPtr->numCol];
1.6 oster 246: for (r = 0; r < raidPtr->numSpare; r++) {
1.9 oster 247: rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r],
1.23 oster 248: raidPtr->numCol + r, p,
249: raidPtr->sectorsPerDisk,
250: raidPtr->Disks[raidPtr->numCol + r].dev,
251: cfgPtr->maxOutstandingDiskReqs, listp,
252: raidPtr->cleanupList);
1.6 oster 253: if (rc)
254: return (rc);
255: }
256: return (0);
257: }
1.1 oster 258: /* Enqueue a disk I/O
259: *
260: * Unfortunately, we have to do things differently in the different
261: * environments (simulator, user-level, kernel).
262: * At user level, all I/O is blocking, so we have 1 or more threads/disk
263: * and the thread that enqueues is different from the thread that dequeues.
264: * In the kernel, I/O is non-blocking and so we'd like to have multiple
265: * I/Os outstanding on the physical disks when possible.
266: *
267: * when any request arrives at a queue, we have two choices:
268: * dispatch it to the lower levels
269: * queue it up
270: *
271: * kernel rules for when to do what:
272: * locking request: queue empty => dispatch and lock queue,
273: * else queue it
274: * unlocking req : always dispatch it
275: * normal req : queue empty => dispatch it & set priority
276: * queue not full & priority is ok => dispatch it
277: * else queue it
278: *
279: * user-level rules:
280: * always enqueue. In the special case of an unlocking op, enqueue
281: * in a special way that will cause the unlocking op to be the next
282: * thing dequeued.
283: *
284: * simulator rules:
285: * Do the same as at user level, with the sleeps and wakeups suppressed.
286: */
1.6 oster 287: void
1.27 oster 288: rf_DiskIOEnqueue(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int pri)
1.6 oster 289: {
290: RF_ETIMER_START(req->qtime);
291: RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector);
292: req->priority = pri;
293:
1.21 oster 294: #if RF_DEBUG_DISKQUEUE
1.6 oster 295: if (rf_queueDebug && (req->numSector == 0)) {
296: printf("Warning: Enqueueing zero-sector access\n");
297: }
1.21 oster 298: #endif
1.6 oster 299: /*
300: * kernel
301: */
302: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
303: /* locking request */
304: if (RF_LOCKING_REQ(req)) {
305: if (RF_QUEUE_EMPTY(queue)) {
1.23 oster 306: Dprintf2("Dispatching pri %d locking op to c %d (queue empty)\n", pri, queue->col);
1.6 oster 307: RF_LOCK_QUEUE(queue);
308: rf_DispatchKernelIO(queue, req);
309: } else {
310: queue->queueLength++; /* increment count of number
311: * of requests waiting in this
312: * queue */
1.23 oster 313: Dprintf2("Enqueueing pri %d locking op to c %d (queue not empty)\n", pri, queue->col);
1.6 oster 314: req->queue = (void *) queue;
315: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
316: }
317: }
318: /* unlocking request */
319: else
320: if (RF_UNLOCKING_REQ(req)) { /* we'll do the actual unlock
321: * when this I/O completes */
1.23 oster 322: Dprintf2("Dispatching pri %d unlocking op to c %d\n", pri, queue->col);
1.6 oster 323: RF_ASSERT(RF_QUEUE_LOCKED(queue));
324: rf_DispatchKernelIO(queue, req);
325: }
326: /* normal request */
327: else
328: if (RF_OK_TO_DISPATCH(queue, req)) {
1.23 oster 329: Dprintf2("Dispatching pri %d regular op to c %d (ok to dispatch)\n", pri, queue->col);
1.6 oster 330: rf_DispatchKernelIO(queue, req);
331: } else {
332: queue->queueLength++; /* increment count of
333: * number of requests
334: * waiting in this queue */
1.23 oster 335: Dprintf2("Enqueueing pri %d regular op to c %d (not ok to dispatch)\n", pri, queue->col);
1.6 oster 336: req->queue = (void *) queue;
337: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
338: }
339: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
1.1 oster 340: }
1.6 oster 341:
1.1 oster 342:
343: /* get the next set of I/Os started, kernel version only */
1.6 oster 344: void
1.27 oster 345: rf_DiskIOComplete(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int status)
1.6 oster 346: {
347: int done = 0;
348:
349: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
350:
351: /* unlock the queue: (1) after an unlocking req completes (2) after a
352: * locking req fails */
353: if (RF_UNLOCKING_REQ(req) || (RF_LOCKING_REQ(req) && status)) {
1.23 oster 354: Dprintf1("DiskIOComplete: unlocking queue at c %d\n", queue->col);
1.20 oster 355: RF_ASSERT(RF_QUEUE_LOCKED(queue));
1.6 oster 356: RF_UNLOCK_QUEUE(queue);
357: }
358: queue->numOutstanding--;
359: RF_ASSERT(queue->numOutstanding >= 0);
360:
361: /* dispatch requests to the disk until we find one that we can't. */
362: /* no reason to continue once we've filled up the queue */
363: /* no reason to even start if the queue is locked */
364:
365: while (!done && !RF_QUEUE_FULL(queue) && !RF_QUEUE_LOCKED(queue)) {
366: if (queue->nextLockingOp) {
367: req = queue->nextLockingOp;
368: queue->nextLockingOp = NULL;
1.23 oster 369: Dprintf2("DiskIOComplete: a pri %d locking req was pending at c %d\n", req->priority, queue->col);
1.6 oster 370: } else {
371: req = (queue->qPtr->Dequeue) (queue->qHdr);
372: if (req != NULL) {
1.23 oster 373: Dprintf2("DiskIOComplete: extracting pri %d req from queue at c %d\n", req->priority, queue->col);
1.6 oster 374: } else {
375: Dprintf1("DiskIOComplete: no more requests to extract.\n", "");
376: }
377: }
378: if (req) {
379: queue->queueLength--; /* decrement count of number
380: * of requests waiting in this
381: * queue */
382: RF_ASSERT(queue->queueLength >= 0);
383: }
384: if (!req)
385: done = 1;
386: else
387: if (RF_LOCKING_REQ(req)) {
388: if (RF_QUEUE_EMPTY(queue)) { /* dispatch it */
1.23 oster 389: Dprintf2("DiskIOComplete: dispatching pri %d locking req to c %d (queue empty)\n", req->priority, queue->col);
1.6 oster 390: RF_LOCK_QUEUE(queue);
391: rf_DispatchKernelIO(queue, req);
392: done = 1;
393: } else { /* put it aside to wait for
394: * the queue to drain */
1.23 oster 395: Dprintf2("DiskIOComplete: postponing pri %d locking req to c %d\n", req->priority, queue->col);
1.6 oster 396: RF_ASSERT(queue->nextLockingOp == NULL);
397: queue->nextLockingOp = req;
398: done = 1;
399: }
400: } else
401: if (RF_UNLOCKING_REQ(req)) { /* should not happen:
402: * unlocking ops should
403: * not get queued */
404: RF_ASSERT(RF_QUEUE_LOCKED(queue)); /* support it anyway for
405: * the future */
1.23 oster 406: Dprintf2("DiskIOComplete: dispatching pri %d unl req to c %d (SHOULD NOT SEE THIS)\n", req->priority, queue->col);
1.6 oster 407: rf_DispatchKernelIO(queue, req);
408: done = 1;
409: } else
410: if (RF_OK_TO_DISPATCH(queue, req)) {
1.23 oster 411: Dprintf2("DiskIOComplete: dispatching pri %d regular req to c %d (ok to dispatch)\n", req->priority, queue->col);
1.6 oster 412: rf_DispatchKernelIO(queue, req);
413: } else { /* we can't dispatch it,
414: * so just re-enqueue
415: * it. */
416: /* potential trouble here if
417: * disk queues batch reqs */
1.23 oster 418: Dprintf2("DiskIOComplete: re-enqueueing pri %d regular req to c %d\n", req->priority, queue->col);
1.6 oster 419: queue->queueLength++;
420: (queue->qPtr->Enqueue) (queue->qHdr, req, req->priority);
421: done = 1;
422: }
423: }
424:
425: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
1.1 oster 426: }
427: /* promotes accesses tagged with the given parityStripeID from low priority
428: * to normal priority. This promotion is optional, meaning that a queue
429: * need not implement it. If there is no promotion routine associated with
430: * a queue, this routine does nothing and returns -1.
431: */
1.6 oster 432: int
1.27 oster 433: rf_DiskIOPromote(RF_DiskQueue_t *queue, RF_StripeNum_t parityStripeID,
434: RF_ReconUnitNum_t which_ru)
1.6 oster 435: {
436: int retval;
437:
438: if (!queue->qPtr->Promote)
439: return (-1);
440: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
441: retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru);
442: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
443: return (retval);
444: }
445:
446: RF_DiskQueueData_t *
1.27 oster 447: rf_CreateDiskQueueData(RF_IoType_t typ, RF_SectorNum_t ssect,
448: RF_SectorCount_t nsect, caddr_t buf,
449: RF_StripeNum_t parityStripeID,
450: RF_ReconUnitNum_t which_ru,
451: int (*wakeF) (void *, int), void *arg,
452: RF_DiskQueueData_t *next,
453: RF_AccTraceEntry_t *tracerec, void *raidPtr,
454: RF_DiskQueueDataFlags_t flags, void *kb_proc)
1.6 oster 455: {
456: RF_DiskQueueData_t *p;
457:
1.24 oster 458: p = pool_get(&rf_dqd_pool, PR_WAITOK);
1.28 oster 459: p->bp = pool_get(&bufpool, PR_NOWAIT); /* XXX: make up our minds here.
460: WAITOK, or NOWAIT?? */
461:
462: if (p->bp == NULL) {
1.24 oster 463: /* no memory for the buffer!?!? */
464: pool_put(&rf_dqd_pool, p);
465: return(NULL);
466: }
1.6 oster 467:
1.28 oster 468: memset(p->bp, 0, sizeof(struct buf));
1.6 oster 469: p->sectorOffset = ssect + rf_protectedSectors;
470: p->numSector = nsect;
471: p->type = typ;
472: p->buf = buf;
473: p->parityStripeID = parityStripeID;
474: p->which_ru = which_ru;
475: p->CompleteFunc = wakeF;
476: p->argument = arg;
477: p->next = next;
478: p->tracerec = tracerec;
479: p->priority = RF_IO_NORMAL_PRIORITY;
480: p->raidPtr = raidPtr;
481: p->flags = flags;
482: p->b_proc = kb_proc;
483: return (p);
484: }
485:
486: void
1.27 oster 487: rf_FreeDiskQueueData(RF_DiskQueueData_t *p)
1.1 oster 488: {
1.28 oster 489: pool_put(&bufpool, p->bp);
1.24 oster 490: pool_put(&rf_dqd_pool, p);
1.1 oster 491: }
CVSweb <webmaster@jp.NetBSD.org>