Annotation of src/sys/dev/raidframe/rf_diskqueue.c, revision 1.20
1.20 ! oster 1: /* $NetBSD: rf_diskqueue.c,v 1.19 2002/09/15 21:19:50 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>
1.20 ! oster 69: __KERNEL_RCSID(0, "$NetBSD: rf_diskqueue.c,v 1.19 2002/09/15 21:19:50 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_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"
1.11 oster 86: #include "rf_kintf.h"
1.1 oster 87:
88: static int init_dqd(RF_DiskQueueData_t *);
89: static void clean_dqd(RF_DiskQueueData_t *);
90: static void rf_ShutdownDiskQueueSystem(void *);
91:
92: #define Dprintf1(s,a) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL)
93: #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)
94: #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)
95:
1.13 oster 96: /*****************************************************************************
1.1 oster 97: *
1.13 oster 98: * the disk queue switch defines all the functions used in the
99: * different queueing disciplines queue ID, init routine, enqueue
100: * routine, dequeue routine
1.1 oster 101: *
1.13 oster 102: ****************************************************************************/
1.1 oster 103:
104: static RF_DiskQueueSW_t diskqueuesw[] = {
1.6 oster 105: {"fifo", /* FIFO */
106: rf_FifoCreate,
107: rf_FifoEnqueue,
108: rf_FifoDequeue,
109: rf_FifoPeek,
1.1 oster 110: rf_FifoPromote},
111:
1.6 oster 112: {"cvscan", /* cvscan */
113: rf_CvscanCreate,
114: rf_CvscanEnqueue,
115: rf_CvscanDequeue,
116: rf_CvscanPeek,
117: rf_CvscanPromote},
118:
119: {"sstf", /* shortest seek time first */
120: rf_SstfCreate,
121: rf_SstfEnqueue,
122: rf_SstfDequeue,
123: rf_SstfPeek,
1.1 oster 124: rf_SstfPromote},
125:
1.6 oster 126: {"scan", /* SCAN (two-way elevator) */
127: rf_ScanCreate,
128: rf_SstfEnqueue,
129: rf_ScanDequeue,
130: rf_ScanPeek,
1.1 oster 131: rf_SstfPromote},
132:
1.6 oster 133: {"cscan", /* CSCAN (one-way elevator) */
134: rf_CscanCreate,
135: rf_SstfEnqueue,
136: rf_CscanDequeue,
137: rf_CscanPeek,
1.1 oster 138: rf_SstfPromote},
139:
140: };
141: #define NUM_DISK_QUEUE_TYPES (sizeof(diskqueuesw)/sizeof(RF_DiskQueueSW_t))
142:
143: static RF_FreeList_t *rf_dqd_freelist;
144:
145: #define RF_MAX_FREE_DQD 256
146: #define RF_DQD_INC 16
147: #define RF_DQD_INITIAL 64
148:
149: #include <sys/buf.h>
150:
1.6 oster 151: static int
152: init_dqd(dqd)
153: RF_DiskQueueData_t *dqd;
154: {
1.13 oster 155:
1.9 oster 156: dqd->bp = (struct buf *) malloc(sizeof(struct buf),
157: M_RAIDFRAME, M_NOWAIT);
1.1 oster 158: if (dqd->bp == NULL) {
1.6 oster 159: return (ENOMEM);
1.1 oster 160: }
1.6 oster 161: memset(dqd->bp, 0, sizeof(struct buf)); /* if you don't do it, nobody
162: * else will.. */
163: return (0);
1.1 oster 164: }
165:
1.6 oster 166: static void
167: clean_dqd(dqd)
168: RF_DiskQueueData_t *dqd;
1.1 oster 169: {
1.6 oster 170: free(dqd->bp, M_RAIDFRAME);
171: }
172: /* configures a single disk queue */
1.9 oster 173:
1.7 oster 174: int
1.9 oster 175: rf_ConfigureDiskQueue(
176: RF_Raid_t * raidPtr,
177: RF_DiskQueue_t * diskqueue,
178: RF_RowCol_t r, /* row & col -- debug only. BZZT not any
1.6 oster 179: * more... */
1.9 oster 180: RF_RowCol_t c,
181: RF_DiskQueueSW_t * p,
182: RF_SectorCount_t sectPerDisk,
183: dev_t dev,
184: int maxOutstanding,
185: RF_ShutdownList_t ** listp,
186: RF_AllocListElem_t * clList)
1.6 oster 187: {
188: int rc;
189:
190: diskqueue->row = r;
191: diskqueue->col = c;
192: diskqueue->qPtr = p;
193: diskqueue->qHdr = (p->Create) (sectPerDisk, clList, listp);
194: diskqueue->dev = dev;
195: diskqueue->numOutstanding = 0;
196: diskqueue->queueLength = 0;
197: diskqueue->maxOutstanding = maxOutstanding;
198: diskqueue->curPriority = RF_IO_NORMAL_PRIORITY;
199: diskqueue->nextLockingOp = NULL;
200: diskqueue->numWaiting = 0;
201: diskqueue->flags = 0;
202: diskqueue->raidPtr = raidPtr;
203: diskqueue->rf_cinfo = &raidPtr->raid_cinfo[r][c];
204: rc = rf_create_managed_mutex(listp, &diskqueue->mutex);
205: if (rc) {
1.18 oster 206: rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc);
1.6 oster 207: return (rc);
208: }
209: rc = rf_create_managed_cond(listp, &diskqueue->cond);
210: if (rc) {
1.18 oster 211: rf_print_unable_to_init_cond(__FILE__, __LINE__, rc);
1.6 oster 212: return (rc);
213: }
214: return (0);
1.1 oster 215: }
216:
1.6 oster 217: static void
218: rf_ShutdownDiskQueueSystem(ignored)
219: void *ignored;
220: {
221: RF_FREELIST_DESTROY_CLEAN(rf_dqd_freelist, next, (RF_DiskQueueData_t *), clean_dqd);
1.1 oster 222: }
223:
1.6 oster 224: int
225: rf_ConfigureDiskQueueSystem(listp)
226: RF_ShutdownList_t **listp;
227: {
228: int rc;
229:
230: RF_FREELIST_CREATE(rf_dqd_freelist, RF_MAX_FREE_DQD,
231: RF_DQD_INC, sizeof(RF_DiskQueueData_t));
232: if (rf_dqd_freelist == NULL)
233: return (ENOMEM);
234: rc = rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL);
235: if (rc) {
1.18 oster 236: rf_print_unable_to_add_shutdown( __FILE__, __LINE__, rc);
1.6 oster 237: rf_ShutdownDiskQueueSystem(NULL);
238: return (rc);
239: }
240: RF_FREELIST_PRIME_INIT(rf_dqd_freelist, RF_DQD_INITIAL, next,
241: (RF_DiskQueueData_t *), init_dqd);
242: return (0);
243: }
244:
245: int
246: rf_ConfigureDiskQueues(
247: RF_ShutdownList_t ** listp,
248: RF_Raid_t * raidPtr,
249: RF_Config_t * cfgPtr)
250: {
251: RF_DiskQueue_t **diskQueues, *spareQueues;
252: RF_DiskQueueSW_t *p;
253: RF_RowCol_t r, c;
254: int rc, i;
255:
256: raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs;
257:
258: for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) {
259: if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) {
260: p = &diskqueuesw[i];
261: break;
262: }
263: }
264: if (p == NULL) {
265: RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType);
266: p = &diskqueuesw[0];
267: }
1.10 oster 268: raidPtr->qType = p;
1.6 oster 269: RF_CallocAndAdd(diskQueues, raidPtr->numRow, sizeof(RF_DiskQueue_t *), (RF_DiskQueue_t **), raidPtr->cleanupList);
270: if (diskQueues == NULL) {
271: return (ENOMEM);
272: }
273: raidPtr->Queues = diskQueues;
274: for (r = 0; r < raidPtr->numRow; r++) {
1.7 oster 275: RF_CallocAndAdd(diskQueues[r], raidPtr->numCol +
276: ((r == 0) ? RF_MAXSPARE : 0),
277: sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *),
278: raidPtr->cleanupList);
1.6 oster 279: if (diskQueues[r] == NULL)
280: return (ENOMEM);
281: for (c = 0; c < raidPtr->numCol; c++) {
1.9 oster 282: rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[r][c],
283: r, c, p,
284: raidPtr->sectorsPerDisk,
285: raidPtr->Disks[r][c].dev,
286: cfgPtr->maxOutstandingDiskReqs,
287: listp, raidPtr->cleanupList);
1.6 oster 288: if (rc)
289: return (rc);
290: }
291: }
292:
293: spareQueues = &raidPtr->Queues[0][raidPtr->numCol];
294: for (r = 0; r < raidPtr->numSpare; r++) {
1.9 oster 295: rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r],
1.6 oster 296: 0, raidPtr->numCol + r, p,
297: raidPtr->sectorsPerDisk,
298: raidPtr->Disks[0][raidPtr->numCol + r].dev,
299: cfgPtr->maxOutstandingDiskReqs, listp,
300: raidPtr->cleanupList);
301: if (rc)
302: return (rc);
303: }
304: return (0);
305: }
1.1 oster 306: /* Enqueue a disk I/O
307: *
308: * Unfortunately, we have to do things differently in the different
309: * environments (simulator, user-level, kernel).
310: * At user level, all I/O is blocking, so we have 1 or more threads/disk
311: * and the thread that enqueues is different from the thread that dequeues.
312: * In the kernel, I/O is non-blocking and so we'd like to have multiple
313: * I/Os outstanding on the physical disks when possible.
314: *
315: * when any request arrives at a queue, we have two choices:
316: * dispatch it to the lower levels
317: * queue it up
318: *
319: * kernel rules for when to do what:
320: * locking request: queue empty => dispatch and lock queue,
321: * else queue it
322: * unlocking req : always dispatch it
323: * normal req : queue empty => dispatch it & set priority
324: * queue not full & priority is ok => dispatch it
325: * else queue it
326: *
327: * user-level rules:
328: * always enqueue. In the special case of an unlocking op, enqueue
329: * in a special way that will cause the unlocking op to be the next
330: * thing dequeued.
331: *
332: * simulator rules:
333: * Do the same as at user level, with the sleeps and wakeups suppressed.
334: */
1.6 oster 335: void
336: rf_DiskIOEnqueue(queue, req, pri)
337: RF_DiskQueue_t *queue;
338: RF_DiskQueueData_t *req;
339: int pri;
340: {
341: RF_ETIMER_START(req->qtime);
342: RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector);
343: req->priority = pri;
344:
345: if (rf_queueDebug && (req->numSector == 0)) {
346: printf("Warning: Enqueueing zero-sector access\n");
347: }
348: /*
349: * kernel
350: */
351: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
352: /* locking request */
353: if (RF_LOCKING_REQ(req)) {
354: if (RF_QUEUE_EMPTY(queue)) {
355: Dprintf3("Dispatching pri %d locking op to r %d c %d (queue empty)\n", pri, queue->row, queue->col);
356: RF_LOCK_QUEUE(queue);
357: rf_DispatchKernelIO(queue, req);
358: } else {
359: queue->queueLength++; /* increment count of number
360: * of requests waiting in this
361: * queue */
362: Dprintf3("Enqueueing pri %d locking op to r %d c %d (queue not empty)\n", pri, queue->row, queue->col);
363: req->queue = (void *) queue;
364: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
365: }
366: }
367: /* unlocking request */
368: else
369: if (RF_UNLOCKING_REQ(req)) { /* we'll do the actual unlock
370: * when this I/O completes */
371: Dprintf3("Dispatching pri %d unlocking op to r %d c %d\n", pri, queue->row, queue->col);
372: RF_ASSERT(RF_QUEUE_LOCKED(queue));
373: rf_DispatchKernelIO(queue, req);
374: }
375: /* normal request */
376: else
377: if (RF_OK_TO_DISPATCH(queue, req)) {
378: Dprintf3("Dispatching pri %d regular op to r %d c %d (ok to dispatch)\n", pri, queue->row, queue->col);
379: rf_DispatchKernelIO(queue, req);
380: } else {
381: queue->queueLength++; /* increment count of
382: * number of requests
383: * waiting in this queue */
384: Dprintf3("Enqueueing pri %d regular op to r %d c %d (not ok to dispatch)\n", pri, queue->row, queue->col);
385: req->queue = (void *) queue;
386: (queue->qPtr->Enqueue) (queue->qHdr, req, pri);
387: }
388: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
1.1 oster 389: }
1.6 oster 390:
1.1 oster 391:
392: /* get the next set of I/Os started, kernel version only */
1.6 oster 393: void
394: rf_DiskIOComplete(queue, req, status)
395: RF_DiskQueue_t *queue;
396: RF_DiskQueueData_t *req;
397: int status;
398: {
399: int done = 0;
400:
401: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
402:
403: /* unlock the queue: (1) after an unlocking req completes (2) after a
404: * locking req fails */
405: if (RF_UNLOCKING_REQ(req) || (RF_LOCKING_REQ(req) && status)) {
406: Dprintf2("DiskIOComplete: unlocking queue at r %d c %d\n", queue->row, queue->col);
1.20 ! oster 407: RF_ASSERT(RF_QUEUE_LOCKED(queue));
1.6 oster 408: RF_UNLOCK_QUEUE(queue);
409: }
410: queue->numOutstanding--;
411: RF_ASSERT(queue->numOutstanding >= 0);
412:
413: /* dispatch requests to the disk until we find one that we can't. */
414: /* no reason to continue once we've filled up the queue */
415: /* no reason to even start if the queue is locked */
416:
417: while (!done && !RF_QUEUE_FULL(queue) && !RF_QUEUE_LOCKED(queue)) {
418: if (queue->nextLockingOp) {
419: req = queue->nextLockingOp;
420: queue->nextLockingOp = NULL;
421: Dprintf3("DiskIOComplete: a pri %d locking req was pending at r %d c %d\n", req->priority, queue->row, queue->col);
422: } else {
423: req = (queue->qPtr->Dequeue) (queue->qHdr);
424: if (req != NULL) {
425: Dprintf3("DiskIOComplete: extracting pri %d req from queue at r %d c %d\n", req->priority, queue->row, queue->col);
426: } else {
427: Dprintf1("DiskIOComplete: no more requests to extract.\n", "");
428: }
429: }
430: if (req) {
431: queue->queueLength--; /* decrement count of number
432: * of requests waiting in this
433: * queue */
434: RF_ASSERT(queue->queueLength >= 0);
435: }
436: if (!req)
437: done = 1;
438: else
439: if (RF_LOCKING_REQ(req)) {
440: if (RF_QUEUE_EMPTY(queue)) { /* dispatch it */
441: Dprintf3("DiskIOComplete: dispatching pri %d locking req to r %d c %d (queue empty)\n", req->priority, queue->row, queue->col);
442: RF_LOCK_QUEUE(queue);
443: rf_DispatchKernelIO(queue, req);
444: done = 1;
445: } else { /* put it aside to wait for
446: * the queue to drain */
447: Dprintf3("DiskIOComplete: postponing pri %d locking req to r %d c %d\n", req->priority, queue->row, queue->col);
448: RF_ASSERT(queue->nextLockingOp == NULL);
449: queue->nextLockingOp = req;
450: done = 1;
451: }
452: } else
453: if (RF_UNLOCKING_REQ(req)) { /* should not happen:
454: * unlocking ops should
455: * not get queued */
456: RF_ASSERT(RF_QUEUE_LOCKED(queue)); /* support it anyway for
457: * the future */
458: Dprintf3("DiskIOComplete: dispatching pri %d unl req to r %d c %d (SHOULD NOT SEE THIS)\n", req->priority, queue->row, queue->col);
459: rf_DispatchKernelIO(queue, req);
460: done = 1;
461: } else
462: if (RF_OK_TO_DISPATCH(queue, req)) {
463: Dprintf3("DiskIOComplete: dispatching pri %d regular req to r %d c %d (ok to dispatch)\n", req->priority, queue->row, queue->col);
464: rf_DispatchKernelIO(queue, req);
465: } else { /* we can't dispatch it,
466: * so just re-enqueue
467: * it. */
468: /* potential trouble here if
469: * disk queues batch reqs */
470: Dprintf3("DiskIOComplete: re-enqueueing pri %d regular req to r %d c %d\n", req->priority, queue->row, queue->col);
471: queue->queueLength++;
472: (queue->qPtr->Enqueue) (queue->qHdr, req, req->priority);
473: done = 1;
474: }
475: }
476:
477: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
1.1 oster 478: }
479: /* promotes accesses tagged with the given parityStripeID from low priority
480: * to normal priority. This promotion is optional, meaning that a queue
481: * need not implement it. If there is no promotion routine associated with
482: * a queue, this routine does nothing and returns -1.
483: */
1.6 oster 484: int
485: rf_DiskIOPromote(queue, parityStripeID, which_ru)
486: RF_DiskQueue_t *queue;
487: RF_StripeNum_t parityStripeID;
488: RF_ReconUnitNum_t which_ru;
489: {
490: int retval;
491:
492: if (!queue->qPtr->Promote)
493: return (-1);
494: RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
495: retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru);
496: RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
497: return (retval);
498: }
499:
500: RF_DiskQueueData_t *
501: rf_CreateDiskQueueData(
502: RF_IoType_t typ,
503: RF_SectorNum_t ssect,
504: RF_SectorCount_t nsect,
505: caddr_t buf,
506: RF_StripeNum_t parityStripeID,
507: RF_ReconUnitNum_t which_ru,
508: int (*wakeF) (void *, int),
509: void *arg,
510: RF_DiskQueueData_t * next,
511: RF_AccTraceEntry_t * tracerec,
512: void *raidPtr,
513: RF_DiskQueueDataFlags_t flags,
514: void *kb_proc)
515: {
516: RF_DiskQueueData_t *p;
517:
518: RF_FREELIST_GET_INIT(rf_dqd_freelist, p, next, (RF_DiskQueueData_t *), init_dqd);
519:
520: p->sectorOffset = ssect + rf_protectedSectors;
521: p->numSector = nsect;
522: p->type = typ;
523: p->buf = buf;
524: p->parityStripeID = parityStripeID;
525: p->which_ru = which_ru;
526: p->CompleteFunc = wakeF;
527: p->argument = arg;
528: p->next = next;
529: p->tracerec = tracerec;
530: p->priority = RF_IO_NORMAL_PRIORITY;
531: p->raidPtr = raidPtr;
532: p->flags = flags;
533: p->b_proc = kb_proc;
534: return (p);
535: }
536:
537: void
538: rf_FreeDiskQueueData(p)
539: RF_DiskQueueData_t *p;
1.1 oster 540: {
1.6 oster 541: RF_FREELIST_FREE_CLEAN(rf_dqd_freelist, p, next, clean_dqd);
1.1 oster 542: }
CVSweb <webmaster@jp.NetBSD.org>