Annotation of src/sys/dev/raidframe/rf_diskqueue.c, revision 1.26
1.26 ! oster 1: /* $NetBSD: rf_diskqueue.c,v 1.25 2003/12/29 05:22:16 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>
1.26 ! oster 69: __KERNEL_RCSID(0, "$NetBSD: rf_diskqueue.c,v 1.25 2003/12/29 05:22:16 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 int init_dqd(RF_DiskQueueData_t *);
88: static void clean_dqd(RF_DiskQueueData_t *);
89: static void rf_ShutdownDiskQueueSystem(void *);
90:
1.21 oster 91: #ifndef RF_DEBUG_DISKQUEUE
92: #define RF_DEBUG_DISKQUEUE 0
93: #endif
94:
95: #if RF_DEBUG_DISKQUEUE
1.1 oster 96: #define Dprintf1(s,a) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL)
97: #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)
98: #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 99: #else
100: #define Dprintf1(s,a)
101: #define Dprintf2(s,a,b)
102: #define Dprintf3(s,a,b,c)
103: #endif
1.1 oster 104:
1.13 oster 105: /*****************************************************************************
1.1 oster 106: *
1.13 oster 107: * the disk queue switch defines all the functions used in the
108: * different queueing disciplines queue ID, init routine, enqueue
109: * routine, dequeue routine
1.1 oster 110: *
1.13 oster 111: ****************************************************************************/
1.1 oster 112:
1.22 jdolecek 113: static const RF_DiskQueueSW_t diskqueuesw[] = {
1.6 oster 114: {"fifo", /* FIFO */
115: rf_FifoCreate,
116: rf_FifoEnqueue,
117: rf_FifoDequeue,
118: rf_FifoPeek,
1.1 oster 119: rf_FifoPromote},
120:
1.6 oster 121: {"cvscan", /* cvscan */
122: rf_CvscanCreate,
123: rf_CvscanEnqueue,
124: rf_CvscanDequeue,
125: rf_CvscanPeek,
126: rf_CvscanPromote},
127:
128: {"sstf", /* shortest seek time first */
129: rf_SstfCreate,
130: rf_SstfEnqueue,
131: rf_SstfDequeue,
132: rf_SstfPeek,
1.1 oster 133: rf_SstfPromote},
134:
1.6 oster 135: {"scan", /* SCAN (two-way elevator) */
136: rf_ScanCreate,
137: rf_SstfEnqueue,
138: rf_ScanDequeue,
139: rf_ScanPeek,
1.1 oster 140: rf_SstfPromote},
141:
1.6 oster 142: {"cscan", /* CSCAN (one-way elevator) */
143: rf_CscanCreate,
144: rf_SstfEnqueue,
145: rf_CscanDequeue,
146: rf_CscanPeek,
1.1 oster 147: rf_SstfPromote},
148:
149: };
150: #define NUM_DISK_QUEUE_TYPES (sizeof(diskqueuesw)/sizeof(RF_DiskQueueSW_t))
151:
1.24 oster 152: static struct pool rf_dqd_pool;
1.1 oster 153: #define RF_MAX_FREE_DQD 256
154: #define RF_DQD_INC 16
155: #define RF_DQD_INITIAL 64
156:
157: #include <sys/buf.h>
158:
1.6 oster 159: static int
160: init_dqd(dqd)
161: RF_DiskQueueData_t *dqd;
162: {
1.13 oster 163:
1.9 oster 164: dqd->bp = (struct buf *) malloc(sizeof(struct buf),
165: M_RAIDFRAME, M_NOWAIT);
1.1 oster 166: if (dqd->bp == NULL) {
1.6 oster 167: return (ENOMEM);
1.1 oster 168: }
1.6 oster 169: memset(dqd->bp, 0, sizeof(struct buf)); /* if you don't do it, nobody
170: * else will.. */
171: return (0);
1.1 oster 172: }
173:
1.6 oster 174: static void
175: clean_dqd(dqd)
176: RF_DiskQueueData_t *dqd;
1.1 oster 177: {
1.6 oster 178: free(dqd->bp, M_RAIDFRAME);
179: }
180: /* configures a single disk queue */
1.9 oster 181:
1.7 oster 182: int
1.9 oster 183: rf_ConfigureDiskQueue(
184: RF_Raid_t * raidPtr,
185: RF_DiskQueue_t * diskqueue,
186: RF_RowCol_t c,
1.22 jdolecek 187: const RF_DiskQueueSW_t * p,
1.9 oster 188: RF_SectorCount_t sectPerDisk,
189: dev_t dev,
190: int maxOutstanding,
191: RF_ShutdownList_t ** listp,
192: RF_AllocListElem_t * clList)
1.6 oster 193: {
194: diskqueue->col = c;
195: diskqueue->qPtr = p;
196: diskqueue->qHdr = (p->Create) (sectPerDisk, clList, listp);
197: diskqueue->dev = dev;
198: diskqueue->numOutstanding = 0;
199: diskqueue->queueLength = 0;
200: diskqueue->maxOutstanding = maxOutstanding;
201: diskqueue->curPriority = RF_IO_NORMAL_PRIORITY;
202: diskqueue->nextLockingOp = NULL;
203: diskqueue->numWaiting = 0;
204: diskqueue->flags = 0;
205: diskqueue->raidPtr = raidPtr;
1.23 oster 206: diskqueue->rf_cinfo = &raidPtr->raid_cinfo[c];
1.25 oster 207: rf_mutex_init(&diskqueue->mutex);
1.26 ! oster 208: diskqueue->cond = 0;
1.6 oster 209: return (0);
1.1 oster 210: }
211:
1.6 oster 212: static void
213: rf_ShutdownDiskQueueSystem(ignored)
214: void *ignored;
215: {
1.24 oster 216: pool_destroy(&rf_dqd_pool);
1.1 oster 217: }
218:
1.6 oster 219: int
220: rf_ConfigureDiskQueueSystem(listp)
221: RF_ShutdownList_t **listp;
222: {
223: int rc;
224:
1.24 oster 225: pool_init(&rf_dqd_pool, sizeof(RF_DiskQueueData_t), 0, 0, 0,
226: "rf_dqd_pl", NULL);
227: pool_sethiwat(&rf_dqd_pool, RF_MAX_FREE_DQD);
228: pool_prime(&rf_dqd_pool, RF_DQD_INITIAL);
229:
1.6 oster 230: rc = rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL);
231: if (rc) {
1.18 oster 232: rf_print_unable_to_add_shutdown( __FILE__, __LINE__, rc);
1.6 oster 233: rf_ShutdownDiskQueueSystem(NULL);
234: return (rc);
235: }
1.24 oster 236:
1.6 oster 237: return (0);
238: }
239:
240: int
241: rf_ConfigureDiskQueues(
242: RF_ShutdownList_t ** listp,
243: RF_Raid_t * raidPtr,
244: RF_Config_t * cfgPtr)
245: {
1.23 oster 246: RF_DiskQueue_t *diskQueues, *spareQueues;
1.22 jdolecek 247: const RF_DiskQueueSW_t *p;
1.23 oster 248: RF_RowCol_t r,c;
1.6 oster 249: int rc, i;
250:
251: raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs;
252:
253: for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) {
254: if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) {
255: p = &diskqueuesw[i];
256: break;
257: }
258: }
259: if (p == NULL) {
260: RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType);
261: p = &diskqueuesw[0];
262: }
1.10 oster 263: raidPtr->qType = p;
1.23 oster 264:
1.24 oster 265: RF_MallocAndAdd(diskQueues,
266: (raidPtr->numCol + RF_MAXSPARE) *
1.23 oster 267: sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *),
268: raidPtr->cleanupList);
269: if (diskQueues == NULL)
1.6 oster 270: return (ENOMEM);
271: raidPtr->Queues = diskQueues;
1.23 oster 272:
273: for (c = 0; c < raidPtr->numCol; c++) {
274: rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[c],
275: c, p,
276: raidPtr->sectorsPerDisk,
277: raidPtr->Disks[c].dev,
278: cfgPtr->maxOutstandingDiskReqs,
279: listp, raidPtr->cleanupList);
280: if (rc)
281: return (rc);
1.6 oster 282: }
283:
1.23 oster 284: spareQueues = &raidPtr->Queues[raidPtr->numCol];
1.6 oster 285: for (r = 0; r < raidPtr->numSpare; r++) {
1.9 oster 286: rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r],
1.23 oster 287: raidPtr->numCol + r, p,
288: raidPtr->sectorsPerDisk,
289: raidPtr->Disks[raidPtr->numCol + r].dev,
290: cfgPtr->maxOutstandingDiskReqs, listp,
291: raidPtr->cleanupList);
1.6 oster 292: if (rc)
293: return (rc);
294: }
295: return (0);
296: }
1.1 oster 297: /* Enqueue a disk I/O
298: *
299: * Unfortunately, we have to do things differently in the different
300: * environments (simulator, user-level, kernel).
301: * At user level, all I/O is blocking, so we have 1 or more threads/disk
302: * and the thread that enqueues is different from the thread that dequeues.
303: * In the kernel, I/O is non-blocking and so we'd like to have multiple
304: * I/Os outstanding on the physical disks when possible.
305: *
306: * when any request arrives at a queue, we have two choices:
307: * dispatch it to the lower levels
308: * queue it up
309: *
310: * kernel rules for when to do what:
311: * locking request: queue empty => dispatch and lock queue,
312: * else queue it
313: * unlocking req : always dispatch it
314: * normal req : queue empty => dispatch it & set priority
315: * queue not full & priority is ok => dispatch it
316: * else queue it
317: *
318: * user-level rules:
319: * always enqueue. In the special case of an unlocking op, enqueue
320: * in a special way that will cause the unlocking op to be the next
321: * thing dequeued.
322: *
323: * simulator rules:
324: * Do the same as at user level, with the sleeps and wakeups suppressed.
325: */
1.6 oster 326: void
327: rf_DiskIOEnqueue(queue, req, pri)
328: RF_DiskQueue_t *queue;
329: RF_DiskQueueData_t *req;
330: int pri;
331: {
332: RF_ETIMER_START(req->qtime);
333: RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector);
334: req->priority = pri;
335:
1.21 oster 336: #if RF_DEBUG_DISKQUEUE
1.6 oster 337: if (rf_queueDebug && (req->numSector == 0)) {
338: printf("Warning: Enqueueing zero-sector access\n");
339: }
1.21 oster 340: #endif
1.6 oster 341: /*
342: * kernel
343: */
344: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
345: /* locking request */
346: if (RF_LOCKING_REQ(req)) {
347: if (RF_QUEUE_EMPTY(queue)) {
1.23 oster 348: Dprintf2("Dispatching pri %d locking op to c %d (queue empty)\n", pri, queue->col);
1.6 oster 349: RF_LOCK_QUEUE(queue);
350: rf_DispatchKernelIO(queue, req);
351: } else {
352: queue->queueLength++; /* increment count of number
353: * of requests waiting in this
354: * queue */
1.23 oster 355: Dprintf2("Enqueueing pri %d locking op to c %d (queue not empty)\n", pri, queue->col);
1.6 oster 356: req->queue = (void *) queue;
357: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
358: }
359: }
360: /* unlocking request */
361: else
362: if (RF_UNLOCKING_REQ(req)) { /* we'll do the actual unlock
363: * when this I/O completes */
1.23 oster 364: Dprintf2("Dispatching pri %d unlocking op to c %d\n", pri, queue->col);
1.6 oster 365: RF_ASSERT(RF_QUEUE_LOCKED(queue));
366: rf_DispatchKernelIO(queue, req);
367: }
368: /* normal request */
369: else
370: if (RF_OK_TO_DISPATCH(queue, req)) {
1.23 oster 371: Dprintf2("Dispatching pri %d regular op to c %d (ok to dispatch)\n", pri, queue->col);
1.6 oster 372: rf_DispatchKernelIO(queue, req);
373: } else {
374: queue->queueLength++; /* increment count of
375: * number of requests
376: * waiting in this queue */
1.23 oster 377: Dprintf2("Enqueueing pri %d regular op to c %d (not ok to dispatch)\n", pri, queue->col);
1.6 oster 378: req->queue = (void *) queue;
379: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
380: }
381: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
1.1 oster 382: }
1.6 oster 383:
1.1 oster 384:
385: /* get the next set of I/Os started, kernel version only */
1.6 oster 386: void
387: rf_DiskIOComplete(queue, req, status)
388: RF_DiskQueue_t *queue;
389: RF_DiskQueueData_t *req;
390: int status;
391: {
392: int done = 0;
393:
394: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
395:
396: /* unlock the queue: (1) after an unlocking req completes (2) after a
397: * locking req fails */
398: if (RF_UNLOCKING_REQ(req) || (RF_LOCKING_REQ(req) && status)) {
1.23 oster 399: Dprintf1("DiskIOComplete: unlocking queue at c %d\n", queue->col);
1.20 oster 400: RF_ASSERT(RF_QUEUE_LOCKED(queue));
1.6 oster 401: RF_UNLOCK_QUEUE(queue);
402: }
403: queue->numOutstanding--;
404: RF_ASSERT(queue->numOutstanding >= 0);
405:
406: /* dispatch requests to the disk until we find one that we can't. */
407: /* no reason to continue once we've filled up the queue */
408: /* no reason to even start if the queue is locked */
409:
410: while (!done && !RF_QUEUE_FULL(queue) && !RF_QUEUE_LOCKED(queue)) {
411: if (queue->nextLockingOp) {
412: req = queue->nextLockingOp;
413: queue->nextLockingOp = NULL;
1.23 oster 414: Dprintf2("DiskIOComplete: a pri %d locking req was pending at c %d\n", req->priority, queue->col);
1.6 oster 415: } else {
416: req = (queue->qPtr->Dequeue) (queue->qHdr);
417: if (req != NULL) {
1.23 oster 418: Dprintf2("DiskIOComplete: extracting pri %d req from queue at c %d\n", req->priority, queue->col);
1.6 oster 419: } else {
420: Dprintf1("DiskIOComplete: no more requests to extract.\n", "");
421: }
422: }
423: if (req) {
424: queue->queueLength--; /* decrement count of number
425: * of requests waiting in this
426: * queue */
427: RF_ASSERT(queue->queueLength >= 0);
428: }
429: if (!req)
430: done = 1;
431: else
432: if (RF_LOCKING_REQ(req)) {
433: if (RF_QUEUE_EMPTY(queue)) { /* dispatch it */
1.23 oster 434: Dprintf2("DiskIOComplete: dispatching pri %d locking req to c %d (queue empty)\n", req->priority, queue->col);
1.6 oster 435: RF_LOCK_QUEUE(queue);
436: rf_DispatchKernelIO(queue, req);
437: done = 1;
438: } else { /* put it aside to wait for
439: * the queue to drain */
1.23 oster 440: Dprintf2("DiskIOComplete: postponing pri %d locking req to c %d\n", req->priority, queue->col);
1.6 oster 441: RF_ASSERT(queue->nextLockingOp == NULL);
442: queue->nextLockingOp = req;
443: done = 1;
444: }
445: } else
446: if (RF_UNLOCKING_REQ(req)) { /* should not happen:
447: * unlocking ops should
448: * not get queued */
449: RF_ASSERT(RF_QUEUE_LOCKED(queue)); /* support it anyway for
450: * the future */
1.23 oster 451: Dprintf2("DiskIOComplete: dispatching pri %d unl req to c %d (SHOULD NOT SEE THIS)\n", req->priority, queue->col);
1.6 oster 452: rf_DispatchKernelIO(queue, req);
453: done = 1;
454: } else
455: if (RF_OK_TO_DISPATCH(queue, req)) {
1.23 oster 456: Dprintf2("DiskIOComplete: dispatching pri %d regular req to c %d (ok to dispatch)\n", req->priority, queue->col);
1.6 oster 457: rf_DispatchKernelIO(queue, req);
458: } else { /* we can't dispatch it,
459: * so just re-enqueue
460: * it. */
461: /* potential trouble here if
462: * disk queues batch reqs */
1.23 oster 463: Dprintf2("DiskIOComplete: re-enqueueing pri %d regular req to c %d\n", req->priority, queue->col);
1.6 oster 464: queue->queueLength++;
465: (queue->qPtr->Enqueue) (queue->qHdr, req, req->priority);
466: done = 1;
467: }
468: }
469:
470: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
1.1 oster 471: }
472: /* promotes accesses tagged with the given parityStripeID from low priority
473: * to normal priority. This promotion is optional, meaning that a queue
474: * need not implement it. If there is no promotion routine associated with
475: * a queue, this routine does nothing and returns -1.
476: */
1.6 oster 477: int
478: rf_DiskIOPromote(queue, parityStripeID, which_ru)
479: RF_DiskQueue_t *queue;
480: RF_StripeNum_t parityStripeID;
481: RF_ReconUnitNum_t which_ru;
482: {
483: int retval;
484:
485: if (!queue->qPtr->Promote)
486: return (-1);
487: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
488: retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru);
489: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
490: return (retval);
491: }
492:
493: RF_DiskQueueData_t *
494: rf_CreateDiskQueueData(
495: RF_IoType_t typ,
496: RF_SectorNum_t ssect,
497: RF_SectorCount_t nsect,
498: caddr_t buf,
499: RF_StripeNum_t parityStripeID,
500: RF_ReconUnitNum_t which_ru,
501: int (*wakeF) (void *, int),
502: void *arg,
503: RF_DiskQueueData_t * next,
504: RF_AccTraceEntry_t * tracerec,
505: void *raidPtr,
506: RF_DiskQueueDataFlags_t flags,
507: void *kb_proc)
508: {
509: RF_DiskQueueData_t *p;
510:
1.24 oster 511: p = pool_get(&rf_dqd_pool, PR_WAITOK);
512: if (init_dqd(p)) {
513: /* no memory for the buffer!?!? */
514: pool_put(&rf_dqd_pool, p);
515: return(NULL);
516: }
1.6 oster 517:
518: p->sectorOffset = ssect + rf_protectedSectors;
519: p->numSector = nsect;
520: p->type = typ;
521: p->buf = buf;
522: p->parityStripeID = parityStripeID;
523: p->which_ru = which_ru;
524: p->CompleteFunc = wakeF;
525: p->argument = arg;
526: p->next = next;
527: p->tracerec = tracerec;
528: p->priority = RF_IO_NORMAL_PRIORITY;
529: p->raidPtr = raidPtr;
530: p->flags = flags;
531: p->b_proc = kb_proc;
532: return (p);
533: }
534:
535: void
536: rf_FreeDiskQueueData(p)
537: RF_DiskQueueData_t *p;
1.1 oster 538: {
1.24 oster 539: clean_dqd(p);
540: pool_put(&rf_dqd_pool, p);
1.1 oster 541: }
CVSweb <webmaster@jp.NetBSD.org>