Annotation of src/sys/dev/raidframe/rf_diskqueue.c, revision 1.3
1.3 ! oster 1: /* $NetBSD: rf_diskqueue.c,v 1.2 1998/12/03 14:58:24 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:
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 locking op
38: * is dispatched to the lower levels of the driver, the queue is locked, and no further
39: * I/Os are dispatched until the queue receives & completes a corresponding "unlocking
40: * operation". This code relies on the higher layers to guarantee that a locking
41: * op will always be eventually followed by an unlocking op. The model is that
42: * the higher layers are structured so locking and unlocking ops occur in pairs, i.e.
43: * an unlocking op cannot be generated until after a locking op reports completion.
44: * There is no good way to check to see that an unlocking op "corresponds" to the
45: * op that currently has the queue locked, so we make no such attempt. Since by
46: * definition there can be only one locking op outstanding on a disk, this should
47: * not be a problem.
48: *
49: * In the kernel, we allow multiple I/Os to be concurrently dispatched to the disk
50: * driver. In order to support locking ops in this environment, when we decide to
51: * do a locking op, we stop dispatching new I/Os and wait until all dispatched I/Os
52: * have completed before dispatching the locking op.
53: *
54: * Unfortunately, the code is different in the 3 different operating states
55: * (user level, kernel, simulator). In the kernel, I/O is non-blocking, and
56: * we have no disk threads to dispatch for us. Therefore, we have to dispatch
57: * new I/Os to the scsi driver at the time of enqueue, and also at the time
58: * of completion. At user level, I/O is blocking, and so only the disk threads
59: * may dispatch I/Os. Thus at user level, all we can do at enqueue time is
60: * enqueue and wake up the disk thread to do the dispatch.
61: *
62: ***************************************************************************************/
63:
64: /*
65: * :
66: *
67: * Log: rf_diskqueue.c,v
68: * Revision 1.50 1996/08/07 21:08:38 jimz
69: * b_proc -> kb_proc
70: *
71: * Revision 1.49 1996/07/05 20:36:14 jimz
72: * make rf_ConfigureDiskQueueSystem return 0
73: *
74: * Revision 1.48 1996/06/18 20:53:11 jimz
75: * fix up disk queueing (remove configure routine,
76: * add shutdown list arg to create routines)
77: *
78: * Revision 1.47 1996/06/14 14:16:36 jimz
79: * fix handling of bogus queue type
80: *
81: * Revision 1.46 1996/06/13 20:41:44 jimz
82: * add scan, cscan, random queueing
83: *
84: * Revision 1.45 1996/06/11 01:27:50 jimz
85: * Fixed bug where diskthread shutdown would crash or hang. This
86: * turned out to be two distinct bugs:
87: * (1) [crash] The thread shutdown code wasn't properly waiting for
88: * all the diskthreads to complete. This caused diskthreads that were
89: * exiting+cleaning up to unlock a destroyed mutex.
90: * (2) [hang] TerminateDiskQueues wasn't locking, and DiskIODequeue
91: * only checked for termination _after_ a wakeup if the queues were
92: * empty. This was a race where the termination wakeup could be lost
93: * by the dequeueing thread, and the system would hang waiting for the
94: * thread to exit, while the thread waited for an I/O or a signal to
95: * check the termination flag.
96: *
97: * Revision 1.44 1996/06/10 11:55:47 jimz
98: * Straightened out some per-array/not-per-array distinctions, fixed
99: * a couple bugs related to confusion. Added shutdown lists. Removed
100: * layout shutdown function (now subsumed by shutdown lists).
101: *
102: * Revision 1.43 1996/06/09 02:36:46 jimz
103: * lots of little crufty cleanup- fixup whitespace
104: * issues, comment #ifdefs, improve typing in some
105: * places (esp size-related)
106: *
107: * Revision 1.42 1996/06/07 22:26:27 jimz
108: * type-ify which_ru (RF_ReconUnitNum_t)
109: *
110: * Revision 1.41 1996/06/07 21:33:04 jimz
111: * begin using consistent types for sector numbers,
112: * stripe numbers, row+col numbers, recon unit numbers
113: *
114: * Revision 1.40 1996/06/06 17:28:04 jimz
115: * track sector number of last I/O dequeued
116: *
117: * Revision 1.39 1996/06/06 01:14:13 jimz
118: * fix crashing bug when tracerec is NULL (ie, from copyback)
119: * initialize req->queue
120: *
121: * Revision 1.38 1996/06/05 19:38:32 jimz
122: * fixed up disk queueing types config
123: * added sstf disk queueing
124: * fixed exit bug on diskthreads (ref-ing bad mem)
125: *
126: * Revision 1.37 1996/06/05 18:06:02 jimz
127: * Major code cleanup. The Great Renaming is now done.
128: * Better modularity. Better typing. Fixed a bunch of
129: * synchronization bugs. Made a lot of global stuff
130: * per-desc or per-array. Removed dead code.
131: *
132: * Revision 1.36 1996/05/30 23:22:16 jimz
133: * bugfixes of serialization, timing problems
134: * more cleanup
135: *
136: * Revision 1.35 1996/05/30 12:59:18 jimz
137: * make etimer happier, more portable
138: *
139: * Revision 1.34 1996/05/30 11:29:41 jimz
140: * Numerous bug fixes. Stripe lock release code disagreed with the taking code
141: * about when stripes should be locked (I made it consistent: no parity, no lock)
142: * There was a lot of extra serialization of I/Os which I've removed- a lot of
143: * it was to calculate values for the cache code, which is no longer with us.
144: * More types, function, macro cleanup. Added code to properly quiesce the array
145: * on shutdown. Made a lot of stuff array-specific which was (bogusly) general
146: * before. Fixed memory allocation, freeing bugs.
147: *
148: * Revision 1.33 1996/05/27 18:56:37 jimz
149: * more code cleanup
150: * better typing
151: * compiles in all 3 environments
152: *
153: * Revision 1.32 1996/05/24 22:17:04 jimz
154: * continue code + namespace cleanup
155: * typed a bunch of flags
156: *
157: * Revision 1.31 1996/05/24 01:59:45 jimz
158: * another checkpoint in code cleanup for release
159: * time to sync kernel tree
160: *
161: * Revision 1.30 1996/05/23 21:46:35 jimz
162: * checkpoint in code cleanup (release prep)
163: * lots of types, function names have been fixed
164: *
165: * Revision 1.29 1996/05/23 00:33:23 jimz
166: * code cleanup: move all debug decls to rf_options.c, all extern
167: * debug decls to rf_options.h, all debug vars preceded by rf_
168: *
169: * Revision 1.28 1996/05/20 16:14:29 jimz
170: * switch to rf_{mutex,cond}_{init,destroy}
171: *
172: * Revision 1.27 1996/05/18 19:51:34 jimz
173: * major code cleanup- fix syntax, make some types consistent,
174: * add prototypes, clean out dead code, et cetera
175: *
176: * Revision 1.26 1996/05/16 19:21:49 wvcii
177: * fixed typo in init_dqd
178: *
179: * Revision 1.25 1996/05/16 16:02:51 jimz
180: * switch to RF_FREELIST stuff for DiskQueueData
181: *
182: * Revision 1.24 1996/05/10 16:24:14 jimz
183: * new cvscan function names
184: *
185: * Revision 1.23 1996/05/01 16:27:54 jimz
186: * don't use ccmn bp management
187: *
188: * Revision 1.22 1995/12/12 18:10:06 jimz
189: * MIN -> RF_MIN, MAX -> RF_MAX, ASSERT -> RF_ASSERT
190: * fix 80-column brain damage in comments
191: *
192: * Revision 1.21 1995/12/01 15:59:59 root
193: * added copyright info
194: *
195: * Revision 1.20 1995/11/07 16:27:20 wvcii
196: * added Peek() function to diskqueuesw
197: * non-locking accesses are never blocked (assume clients enforce proper
198: * respect for lock acquisition)
199: *
200: * Revision 1.19 1995/10/05 18:56:52 jimz
201: * fix req handling in IOComplete
202: *
203: * Revision 1.18 1995/10/04 20:13:50 wvcii
204: * added asserts to monitor numOutstanding queueLength
205: *
206: * Revision 1.17 1995/10/04 07:43:52 wvcii
207: * queue->numOutstanding now valid for user & sim
208: * added queue->queueLength
209: * user tested & verified, sim untested
210: *
211: * Revision 1.16 1995/09/12 00:21:19 wvcii
212: * added support for tracing disk queue time
213: *
214: */
215:
216: #include "rf_types.h"
217: #include "rf_threadstuff.h"
218: #include "rf_threadid.h"
219: #include "rf_raid.h"
220: #include "rf_diskqueue.h"
221: #include "rf_alloclist.h"
222: #include "rf_acctrace.h"
223: #include "rf_etimer.h"
224: #include "rf_configure.h"
225: #include "rf_general.h"
226: #include "rf_freelist.h"
227: #include "rf_debugprint.h"
228: #include "rf_shutdown.h"
229: #include "rf_cvscan.h"
230: #include "rf_sstf.h"
231: #include "rf_fifo.h"
232:
233: #ifdef SIMULATE
234: #include "rf_diskevent.h"
235: #endif /* SIMULATE */
236:
237: #if !defined(__NetBSD__)
238: extern struct buf *ubc_bufget();
239: #endif
240:
241: static int init_dqd(RF_DiskQueueData_t *);
242: static void clean_dqd(RF_DiskQueueData_t *);
243: static void rf_ShutdownDiskQueueSystem(void *);
244: /* From rf_kintf.c */
245: int rf_DispatchKernelIO(RF_DiskQueue_t *,RF_DiskQueueData_t *);
246:
247:
248: #define Dprintf1(s,a) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL)
249: #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)
250: #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)
251: #define Dprintf4(s,a,b,c,d) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),NULL,NULL,NULL,NULL)
252: #define Dprintf5(s,a,b,c,d,e) if (rf_queueDebug) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),(void *)((unsigned long)e),NULL,NULL,NULL)
253:
254: #if !defined(KERNEL) && !defined(SIMULATE)
255:
256: /* queue must be locked before invoking this */
257: #define SIGNAL_DISK_QUEUE(_q_,_wh_) \
258: { \
259: if ( (_q_)->numWaiting > 0) { \
260: (_q_)->numWaiting--; \
261: RF_SIGNAL_COND( ((_q_)->cond) ); \
262: } \
263: }
264:
265: /* queue must be locked before invoking this */
266: #define WAIT_DISK_QUEUE(_q_,_wh_) \
267: { \
268: (_q_)->numWaiting++; \
269: RF_WAIT_COND( ((_q_)->cond), ((_q_)->mutex) ); \
270: }
271:
272: #else /* !defined(KERNEL) && !defined(SIMULATE) */
273:
274: #define SIGNAL_DISK_QUEUE(_q_,_wh_)
275: #define WAIT_DISK_QUEUE(_q_,_wh_)
276:
277: #endif /* !defined(KERNEL) && !defined(SIMULATE) */
278:
279: /*****************************************************************************************
280: *
281: * the disk queue switch defines all the functions used in the different queueing
282: * disciplines
283: * queue ID, init routine, enqueue routine, dequeue routine
284: *
285: ****************************************************************************************/
286:
287: static RF_DiskQueueSW_t diskqueuesw[] = {
288: {"fifo", /* FIFO */
289: rf_FifoCreate,
290: rf_FifoEnqueue,
291: rf_FifoDequeue,
292: rf_FifoPeek,
293: rf_FifoPromote},
294:
295: {"cvscan", /* cvscan */
296: rf_CvscanCreate,
297: rf_CvscanEnqueue,
298: rf_CvscanDequeue,
299: rf_CvscanPeek,
300: rf_CvscanPromote },
301:
302: {"sstf", /* shortest seek time first */
303: rf_SstfCreate,
304: rf_SstfEnqueue,
305: rf_SstfDequeue,
306: rf_SstfPeek,
307: rf_SstfPromote},
308:
309: {"scan", /* SCAN (two-way elevator) */
310: rf_ScanCreate,
311: rf_SstfEnqueue,
312: rf_ScanDequeue,
313: rf_ScanPeek,
314: rf_SstfPromote},
315:
316: {"cscan", /* CSCAN (one-way elevator) */
317: rf_CscanCreate,
318: rf_SstfEnqueue,
319: rf_CscanDequeue,
320: rf_CscanPeek,
321: rf_SstfPromote},
322:
323: #if !defined(KERNEL) && RF_INCLUDE_QUEUE_RANDOM > 0
324: /* to make a point to Chris :-> */
325: {"random", /* random */
326: rf_FifoCreate,
327: rf_FifoEnqueue,
328: rf_RandomDequeue,
329: rf_RandomPeek,
330: rf_FifoPromote},
331: #endif /* !KERNEL && RF_INCLUDE_QUEUE_RANDOM > 0 */
332: };
333: #define NUM_DISK_QUEUE_TYPES (sizeof(diskqueuesw)/sizeof(RF_DiskQueueSW_t))
334:
335: static RF_FreeList_t *rf_dqd_freelist;
336:
337: #define RF_MAX_FREE_DQD 256
338: #define RF_DQD_INC 16
339: #define RF_DQD_INITIAL 64
340:
341: #ifdef __NetBSD__
342: #ifdef _KERNEL
343: #include <sys/buf.h>
344: #endif
345: #endif
346:
347: static int init_dqd(dqd)
348: RF_DiskQueueData_t *dqd;
349: {
350: #ifdef KERNEL
351: #ifdef __NetBSD__
352: /* XXX not sure if the following malloc is appropriate... probably not quite... */
353: dqd->bp = (struct buf *) malloc( sizeof(struct buf), M_DEVBUF, M_NOWAIT);
354: /* XXX */
355: /* printf("NEED TO IMPLEMENT THIS BETTER!\n"); */
356: #else
357: dqd->bp = ubc_bufget();
358: #endif
359: if (dqd->bp == NULL) {
360: return(ENOMEM);
361: }
1.3 ! oster 362: #ifdef __NetBSD__
! 363: memset(dqd->bp,0,sizeof(struct buf)); /* if you don't do it, nobody else will.. */
! 364: #endif
1.1 oster 365: #endif /* KERNEL */
366: return(0);
367: }
368:
369: static void clean_dqd(dqd)
370: RF_DiskQueueData_t *dqd;
371: {
372: #ifdef KERNEL
373: #ifdef __NetBSD__
374: /* printf("NEED TO IMPLEMENT THIS BETTER(2)!\n"); */
375: /* XXX ? */
376: free( dqd->bp, M_DEVBUF );
377: #else
378: ubc_buffree(dqd->bp);
379: #endif
380:
381: #endif /* KERNEL */
382: }
383:
384: /* configures a single disk queue */
385: static int config_disk_queue(
386: RF_Raid_t *raidPtr,
387: RF_DiskQueue_t *diskqueue,
388: RF_RowCol_t r, /* row & col -- debug only. BZZT not any more... */
389: RF_RowCol_t c,
390: RF_DiskQueueSW_t *p,
391: RF_SectorCount_t sectPerDisk,
392: dev_t dev,
393: int maxOutstanding,
394: RF_ShutdownList_t **listp,
395: RF_AllocListElem_t *clList)
396: {
397: int rc;
398:
399: diskqueue->row = r;
400: diskqueue->col = c;
401: diskqueue->qPtr = p;
402: diskqueue->qHdr = (p->Create)(sectPerDisk, clList, listp);
403: diskqueue->dev = dev;
404: diskqueue->numOutstanding = 0;
405: diskqueue->queueLength = 0;
406: diskqueue->maxOutstanding = maxOutstanding;
407: diskqueue->curPriority = RF_IO_NORMAL_PRIORITY;
408: diskqueue->nextLockingOp = NULL;
409: diskqueue->unlockingOp = NULL;
410: diskqueue->numWaiting=0;
411: diskqueue->flags = 0;
412: diskqueue->raidPtr = raidPtr;
413: #if defined(__NetBSD__) && defined(_KERNEL)
414: diskqueue->rf_cinfo = &raidPtr->raid_cinfo[r][c];
415: #endif
416: rc = rf_create_managed_mutex(listp, &diskqueue->mutex);
417: if (rc) {
418: RF_ERRORMSG3("Unable to init mutex file %s line %d rc=%d\n", __FILE__,
419: __LINE__, rc);
420: return(rc);
421: }
422: rc = rf_create_managed_cond(listp, &diskqueue->cond);
423: if (rc) {
424: RF_ERRORMSG3("Unable to init cond file %s line %d rc=%d\n", __FILE__,
425: __LINE__, rc);
426: return(rc);
427: }
428: return(0);
429: }
430:
431: static void rf_ShutdownDiskQueueSystem(ignored)
432: void *ignored;
433: {
434: RF_FREELIST_DESTROY_CLEAN(rf_dqd_freelist,next,(RF_DiskQueueData_t *),clean_dqd);
435: }
436:
437: int rf_ConfigureDiskQueueSystem(listp)
438: RF_ShutdownList_t **listp;
439: {
440: int rc;
441:
442: RF_FREELIST_CREATE(rf_dqd_freelist, RF_MAX_FREE_DQD,
443: RF_DQD_INC, sizeof(RF_DiskQueueData_t));
444: if (rf_dqd_freelist == NULL)
445: return(ENOMEM);
446: rc = rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL);
447: if (rc) {
448: RF_ERRORMSG3("Unable to add to shutdown list file %s line %d rc=%d\n",
449: __FILE__, __LINE__, rc);
450: rf_ShutdownDiskQueueSystem(NULL);
451: return(rc);
452: }
453: RF_FREELIST_PRIME_INIT(rf_dqd_freelist, RF_DQD_INITIAL,next,
454: (RF_DiskQueueData_t *),init_dqd);
455: return(0);
456: }
457:
458: #ifndef KERNEL
459: /* this is called prior to shutdown to wakeup everyone waiting on a disk queue
460: * and tell them to exit
461: */
462: void rf_TerminateDiskQueues(raidPtr)
463: RF_Raid_t *raidPtr;
464: {
465: RF_RowCol_t r, c;
466:
467: raidPtr->terminate_disk_queues = 1;
468: for (r=0; r<raidPtr->numRow; r++) {
469: for (c=0; c<raidPtr->numCol + ((r==0) ? raidPtr->numSpare : 0); c++) {
470: RF_LOCK_QUEUE_MUTEX(&raidPtr->Queues[r][c], "TerminateDiskQueues");
471: RF_BROADCAST_COND(raidPtr->Queues[r][c].cond);
472: RF_UNLOCK_QUEUE_MUTEX(&raidPtr->Queues[r][c], "TerminateDiskQueues");
473: }
474: }
475: }
476: #endif /* !KERNEL */
477:
478: int rf_ConfigureDiskQueues(
479: RF_ShutdownList_t **listp,
480: RF_Raid_t *raidPtr,
481: RF_Config_t *cfgPtr)
482: {
483: RF_DiskQueue_t **diskQueues, *spareQueues;
484: RF_DiskQueueSW_t *p;
485: RF_RowCol_t r, c;
486: int rc, i;
487:
488: raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs;
489:
490: for(p=NULL,i=0;i<NUM_DISK_QUEUE_TYPES;i++) {
491: if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) {
492: p = &diskqueuesw[i];
493: break;
494: }
495: }
496: if (p == NULL) {
497: RF_ERRORMSG2("Unknown queue type \"%s\". Using %s\n",cfgPtr->diskQueueType, diskqueuesw[0].queueType);
498: p = &diskqueuesw[0];
499: }
500:
501: RF_CallocAndAdd(diskQueues, raidPtr->numRow, sizeof(RF_DiskQueue_t *), (RF_DiskQueue_t **), raidPtr->cleanupList);
502: if (diskQueues == NULL) {
503: return(ENOMEM);
504: }
505: raidPtr->Queues = diskQueues;
506: for (r=0; r<raidPtr->numRow; r++) {
507: RF_CallocAndAdd(diskQueues[r], raidPtr->numCol + ((r==0) ? raidPtr->numSpare : 0), sizeof(RF_DiskQueue_t), (RF_DiskQueue_t *), raidPtr->cleanupList);
508: if (diskQueues[r] == NULL)
509: return(ENOMEM);
510: for (c=0; c<raidPtr->numCol; c++) {
511: rc = config_disk_queue(raidPtr, &diskQueues[r][c], r, c, p,
512: raidPtr->sectorsPerDisk, raidPtr->Disks[r][c].dev,
513: cfgPtr->maxOutstandingDiskReqs, listp, raidPtr->cleanupList);
514: if (rc)
515: return(rc);
516: }
517: }
518:
519: spareQueues = &raidPtr->Queues[0][raidPtr->numCol];
520: for (r=0; r<raidPtr->numSpare; r++) {
521: rc = config_disk_queue(raidPtr, &spareQueues[r],
522: 0, raidPtr->numCol+r, p,
523: raidPtr->sectorsPerDisk,
524: raidPtr->Disks[0][raidPtr->numCol+r].dev,
525: cfgPtr->maxOutstandingDiskReqs, listp,
526: raidPtr->cleanupList);
527: if (rc)
528: return(rc);
529: }
530: return(0);
531: }
532:
533: /* Enqueue a disk I/O
534: *
535: * Unfortunately, we have to do things differently in the different
536: * environments (simulator, user-level, kernel).
537: * At user level, all I/O is blocking, so we have 1 or more threads/disk
538: * and the thread that enqueues is different from the thread that dequeues.
539: * In the kernel, I/O is non-blocking and so we'd like to have multiple
540: * I/Os outstanding on the physical disks when possible.
541: *
542: * when any request arrives at a queue, we have two choices:
543: * dispatch it to the lower levels
544: * queue it up
545: *
546: * kernel rules for when to do what:
547: * locking request: queue empty => dispatch and lock queue,
548: * else queue it
549: * unlocking req : always dispatch it
550: * normal req : queue empty => dispatch it & set priority
551: * queue not full & priority is ok => dispatch it
552: * else queue it
553: *
554: * user-level rules:
555: * always enqueue. In the special case of an unlocking op, enqueue
556: * in a special way that will cause the unlocking op to be the next
557: * thing dequeued.
558: *
559: * simulator rules:
560: * Do the same as at user level, with the sleeps and wakeups suppressed.
561: */
562: void rf_DiskIOEnqueue(queue, req, pri)
563: RF_DiskQueue_t *queue;
564: RF_DiskQueueData_t *req;
565: int pri;
566: {
567: int tid;
568:
569: RF_ETIMER_START(req->qtime);
570: rf_get_threadid(tid);
571: RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector);
572: req->priority = pri;
573:
574: if (rf_queueDebug && (req->numSector == 0)) {
575: printf("Warning: Enqueueing zero-sector access\n");
576: }
577:
578: #ifdef KERNEL
579: /*
580: * kernel
581: */
582: RF_LOCK_QUEUE_MUTEX( queue, "DiskIOEnqueue" );
583: /* locking request */
584: if (RF_LOCKING_REQ(req)) {
585: if (RF_QUEUE_EMPTY(queue)) {
586: Dprintf3("Dispatching pri %d locking op to r %d c %d (queue empty)\n",pri,queue->row, queue->col);
587: RF_LOCK_QUEUE(queue);
588: rf_DispatchKernelIO(queue, req);
589: } else {
590: queue->queueLength++; /* increment count of number of requests waiting in this queue */
591: Dprintf3("Enqueueing pri %d locking op to r %d c %d (queue not empty)\n",pri,queue->row, queue->col);
592: req->queue = (void *)queue;
593: (queue->qPtr->Enqueue)(queue->qHdr, req, pri);
594: }
595: }
596: /* unlocking request */
597: else if (RF_UNLOCKING_REQ(req)) { /* we'll do the actual unlock when this I/O completes */
598: Dprintf3("Dispatching pri %d unlocking op to r %d c %d\n",pri,queue->row, queue->col);
599: RF_ASSERT(RF_QUEUE_LOCKED(queue));
600: rf_DispatchKernelIO(queue, req);
601: }
602: /* normal request */
603: else if (RF_OK_TO_DISPATCH(queue, req)) {
604: Dprintf3("Dispatching pri %d regular op to r %d c %d (ok to dispatch)\n",pri,queue->row, queue->col);
605: rf_DispatchKernelIO(queue, req);
606: } else {
607: queue->queueLength++; /* increment count of number of requests waiting in this queue */
608: Dprintf3("Enqueueing pri %d regular op to r %d c %d (not ok to dispatch)\n",pri,queue->row, queue->col);
609: req->queue = (void *)queue;
610: (queue->qPtr->Enqueue)(queue->qHdr, req, pri);
611: }
612: RF_UNLOCK_QUEUE_MUTEX( queue, "DiskIOEnqueue" );
613:
614: #else /* KERNEL */
615: /*
616: * user-level
617: */
618: RF_LOCK_QUEUE_MUTEX( queue, "DiskIOEnqueue" );
619: queue->queueLength++; /* increment count of number of requests waiting in this queue */
620: /* unlocking request */
621: if (RF_UNLOCKING_REQ(req)) {
622: Dprintf4("[%d] enqueueing pri %d unlocking op & signalling r %d c %d\n", tid, pri, queue->row, queue->col);
623: RF_ASSERT(RF_QUEUE_LOCKED(queue) && queue->unlockingOp == NULL);
624: queue->unlockingOp = req;
625: }
626: /* locking and normal requests */
627: else {
628: req->queue = (void *)queue;
629: Dprintf5("[%d] enqueueing pri %d %s op & signalling r %d c %d\n", tid, pri,
630: (RF_LOCKING_REQ(req)) ? "locking" : "regular",queue->row,queue->col);
631: (queue->qPtr->Enqueue)(queue->qHdr, req, pri);
632: }
633: SIGNAL_DISK_QUEUE( queue, "DiskIOEnqueue");
634: RF_UNLOCK_QUEUE_MUTEX( queue, "DiskIOEnqueue" );
635: #endif /* KERNEL */
636: }
637:
638: #if !defined(KERNEL) && !defined(SIMULATE)
639: /* user-level only: tell all threads to wake up & recheck the queue */
640: void rf_BroadcastOnQueue(queue)
641: RF_DiskQueue_t *queue;
642: {
643: int i;
644:
645: if (queue->maxOutstanding > 1) for (i=0; i<queue->maxOutstanding; i++) {
646: SIGNAL_DISK_QUEUE(queue, "BroadcastOnQueue" );
647: }
648: }
649: #endif /* !KERNEL && !SIMULATE */
650:
651: #ifndef KERNEL /* not used in kernel */
652:
653: RF_DiskQueueData_t *rf_DiskIODequeue(queue)
654: RF_DiskQueue_t *queue;
655: {
656: RF_DiskQueueData_t *p, *headItem;
657: int tid;
658:
659: rf_get_threadid(tid);
660: RF_LOCK_QUEUE_MUTEX( queue, "DiskIODequeue" );
661: for (p=NULL; !p; ) {
662: if (queue->unlockingOp) {
663: /* unlocking request */
664: RF_ASSERT(RF_QUEUE_LOCKED(queue));
665: p = queue->unlockingOp;
666: queue->unlockingOp = NULL;
667: Dprintf4("[%d] dequeueing pri %d unlocking op r %d c %d\n", tid, p->priority, queue->row,queue->col);
668: }
669: else {
670: headItem = (queue->qPtr->Peek)(queue->qHdr);
671: if (headItem) {
672: if (RF_LOCKING_REQ(headItem)) {
673: /* locking request */
674: if (!RF_QUEUE_LOCKED(queue)) {
675: /* queue isn't locked, so dequeue the request & lock the queue */
676: p = (queue->qPtr->Dequeue)( queue->qHdr );
677: if (p)
678: Dprintf4("[%d] dequeueing pri %d locking op r %d c %d\n", tid, p->priority, queue->row, queue->col);
679: else
680: Dprintf3("[%d] no dequeue -- raw queue empty r %d c %d\n", tid, queue->row, queue->col);
681: }
682: else {
683: /* queue already locked, no dequeue occurs */
684: Dprintf3("[%d] no dequeue -- queue is locked r %d c %d\n", tid, queue->row, queue->col);
685: p = NULL;
686: }
687: }
688: else {
689: /* normal request, always dequeue and assume caller already has lock (if needed) */
690: p = (queue->qPtr->Dequeue)( queue->qHdr );
691: if (p)
692: Dprintf4("[%d] dequeueing pri %d regular op r %d c %d\n", tid, p->priority, queue->row, queue->col);
693: else
694: Dprintf3("[%d] no dequeue -- raw queue empty r %d c %d\n", tid, queue->row, queue->col);
695: }
696: }
697: else {
698: Dprintf3("[%d] no dequeue -- raw queue empty r %d c %d\n", tid, queue->row, queue->col);
699: }
700: }
701:
702: if (queue->raidPtr->terminate_disk_queues) {
703: p = NULL;
704: break;
705: }
706: #ifdef SIMULATE
707: break; /* in simulator, return NULL on empty queue instead of blocking */
708: #else /* SIMULATE */
709: if (!p) {
710: Dprintf3("[%d] nothing to dequeue: waiting r %d c %d\n", tid, queue->row, queue->col);
711: WAIT_DISK_QUEUE( queue, "DiskIODequeue" );
712: }
713: #endif /* SIMULATE */
714: }
715:
716: if (p) {
717: queue->queueLength--; /* decrement count of number of requests waiting in this queue */
718: RF_ASSERT(queue->queueLength >= 0);
719: queue->numOutstanding++;
720: queue->last_deq_sector = p->sectorOffset;
721: /* record the amount of time this request spent in the disk queue */
722: RF_ETIMER_STOP(p->qtime);
723: RF_ETIMER_EVAL(p->qtime);
724: if (p->tracerec)
725: p->tracerec->diskqueue_us += RF_ETIMER_VAL_US(p->qtime);
726: }
727:
728: if (p && RF_LOCKING_REQ(p)) {
729: RF_ASSERT(!RF_QUEUE_LOCKED(queue));
730: Dprintf3("[%d] locking queue r %d c %d\n",tid,queue->row,queue->col);
731: RF_LOCK_QUEUE(queue);
732: }
733: RF_UNLOCK_QUEUE_MUTEX( queue, "DiskIODequeue" );
734:
735: return(p);
736: }
737:
738: #else /* !KERNEL */
739:
740: /* get the next set of I/Os started, kernel version only */
741: void rf_DiskIOComplete(queue, req, status)
742: RF_DiskQueue_t *queue;
743: RF_DiskQueueData_t *req;
744: int status;
745: {
746: int done=0;
747:
748: RF_LOCK_QUEUE_MUTEX( queue, "DiskIOComplete" );
749:
750: /* unlock the queue:
751: (1) after an unlocking req completes
752: (2) after a locking req fails
753: */
754: if (RF_UNLOCKING_REQ(req) || (RF_LOCKING_REQ(req) && status)) {
755: Dprintf2("DiskIOComplete: unlocking queue at r %d c %d\n", queue->row, queue->col);
756: RF_ASSERT(RF_QUEUE_LOCKED(queue) && (queue->unlockingOp == NULL));
757: RF_UNLOCK_QUEUE(queue);
758: }
759:
760: queue->numOutstanding--;
761: RF_ASSERT(queue->numOutstanding >= 0);
762:
763: /* dispatch requests to the disk until we find one that we can't. */
764: /* no reason to continue once we've filled up the queue */
765: /* no reason to even start if the queue is locked */
766:
767: while (!done && !RF_QUEUE_FULL(queue) && !RF_QUEUE_LOCKED(queue)) {
768: if (queue->nextLockingOp) {
769: req = queue->nextLockingOp; queue->nextLockingOp = NULL;
770: Dprintf3("DiskIOComplete: a pri %d locking req was pending at r %d c %d\n",req->priority,queue->row, queue->col);
771: } else {
772: req = (queue->qPtr->Dequeue)( queue->qHdr );
1.2 oster 773: if (req != NULL) {
774: Dprintf3("DiskIOComplete: extracting pri %d req from queue at r %d c %d\n",req->priority,queue->row, queue->col);
775: } else {
776: Dprintf1("DiskIOComplete: no more requests to extract.\n","");
777: }
1.1 oster 778: }
779: if (req) {
780: queue->queueLength--; /* decrement count of number of requests waiting in this queue */
781: RF_ASSERT(queue->queueLength >= 0);
782: }
783: if (!req) done=1;
784: else if (RF_LOCKING_REQ(req)) {
785: if (RF_QUEUE_EMPTY(queue)) { /* dispatch it */
786: Dprintf3("DiskIOComplete: dispatching pri %d locking req to r %d c %d (queue empty)\n",req->priority,queue->row, queue->col);
787: RF_LOCK_QUEUE(queue);
788: rf_DispatchKernelIO(queue, req);
789: done = 1;
790: } else { /* put it aside to wait for the queue to drain */
791: Dprintf3("DiskIOComplete: postponing pri %d locking req to r %d c %d\n",req->priority,queue->row, queue->col);
792: RF_ASSERT(queue->nextLockingOp == NULL);
793: queue->nextLockingOp = req;
794: done = 1;
795: }
796: } else if (RF_UNLOCKING_REQ(req)) { /* should not happen: unlocking ops should not get queued */
797: RF_ASSERT(RF_QUEUE_LOCKED(queue)); /* support it anyway for the future */
798: Dprintf3("DiskIOComplete: dispatching pri %d unl req to r %d c %d (SHOULD NOT SEE THIS)\n",req->priority,queue->row, queue->col);
799: rf_DispatchKernelIO(queue, req);
800: done = 1;
801: } else if (RF_OK_TO_DISPATCH(queue, req)) {
802: Dprintf3("DiskIOComplete: dispatching pri %d regular req to r %d c %d (ok to dispatch)\n",req->priority,queue->row, queue->col);
803: rf_DispatchKernelIO(queue, req);
804: } else { /* we can't dispatch it, so just re-enqueue it. */
805: /* potential trouble here if disk queues batch reqs */
806: Dprintf3("DiskIOComplete: re-enqueueing pri %d regular req to r %d c %d\n",req->priority,queue->row, queue->col);
807: queue->queueLength++;
808: (queue->qPtr->Enqueue)(queue->qHdr, req, req->priority);
809: done = 1;
810: }
811: }
812:
813: RF_UNLOCK_QUEUE_MUTEX( queue, "DiskIOComplete" );
814: }
815: #endif /* !KERNEL */
816:
817: /* promotes accesses tagged with the given parityStripeID from low priority
818: * to normal priority. This promotion is optional, meaning that a queue
819: * need not implement it. If there is no promotion routine associated with
820: * a queue, this routine does nothing and returns -1.
821: */
822: int rf_DiskIOPromote(queue, parityStripeID, which_ru)
823: RF_DiskQueue_t *queue;
824: RF_StripeNum_t parityStripeID;
825: RF_ReconUnitNum_t which_ru;
826: {
827: int retval;
828:
829: if (!queue->qPtr->Promote)
830: return(-1);
831: RF_LOCK_QUEUE_MUTEX( queue, "DiskIOPromote" );
832: retval = (queue->qPtr->Promote)( queue->qHdr, parityStripeID, which_ru );
833: RF_UNLOCK_QUEUE_MUTEX( queue, "DiskIOPromote" );
834: return(retval);
835: }
836:
837: RF_DiskQueueData_t *rf_CreateDiskQueueData(
838: RF_IoType_t typ,
839: RF_SectorNum_t ssect,
840: RF_SectorCount_t nsect,
841: caddr_t buf,
842: RF_StripeNum_t parityStripeID,
843: RF_ReconUnitNum_t which_ru,
844: int (*wakeF)(void *,int),
845: void *arg,
846: RF_DiskQueueData_t *next,
847: RF_AccTraceEntry_t *tracerec,
848: void *raidPtr,
849: RF_DiskQueueDataFlags_t flags,
850: void *kb_proc)
851: {
852: RF_DiskQueueData_t *p;
853:
854: RF_FREELIST_GET_INIT(rf_dqd_freelist,p,next,(RF_DiskQueueData_t *),init_dqd);
855:
856: p->sectorOffset = ssect + rf_protectedSectors;
857: p->numSector = nsect;
858: p->type = typ;
859: p->buf = buf;
860: p->parityStripeID= parityStripeID;
861: p->which_ru = which_ru;
862: p->CompleteFunc = wakeF;
863: p->argument = arg;
864: p->next = next;
865: p->tracerec = tracerec;
866: p->priority = RF_IO_NORMAL_PRIORITY;
867: p->AuxFunc = NULL;
868: p->buf2 = NULL;
869: #ifdef SIMULATE
870: p->owner = rf_GetCurrentOwner();
871: #endif /* SIMULATE */
872: p->raidPtr = raidPtr;
873: p->flags = flags;
874: #ifdef KERNEL
875: p->b_proc = kb_proc;
876: #endif /* KERNEL */
877: return(p);
878: }
879:
880: RF_DiskQueueData_t *rf_CreateDiskQueueDataFull(
881: RF_IoType_t typ,
882: RF_SectorNum_t ssect,
883: RF_SectorCount_t nsect,
884: caddr_t buf,
885: RF_StripeNum_t parityStripeID,
886: RF_ReconUnitNum_t which_ru,
887: int (*wakeF)(void *,int),
888: void *arg,
889: RF_DiskQueueData_t *next,
890: RF_AccTraceEntry_t *tracerec,
891: int priority,
892: int (*AuxFunc)(void *,...),
893: caddr_t buf2,
894: void *raidPtr,
895: RF_DiskQueueDataFlags_t flags,
896: void *kb_proc)
897: {
898: RF_DiskQueueData_t *p;
899:
900: RF_FREELIST_GET_INIT(rf_dqd_freelist,p,next,(RF_DiskQueueData_t *),init_dqd);
901:
902: p->sectorOffset = ssect + rf_protectedSectors;
903: p->numSector = nsect;
904: p->type = typ;
905: p->buf = buf;
906: p->parityStripeID= parityStripeID;
907: p->which_ru = which_ru;
908: p->CompleteFunc = wakeF;
909: p->argument = arg;
910: p->next = next;
911: p->tracerec = tracerec;
912: p->priority = priority;
913: p->AuxFunc = AuxFunc;
914: p->buf2 = buf2;
915: #ifdef SIMULATE
916: p->owner = rf_GetCurrentOwner();
917: #endif /* SIMULATE */
918: p->raidPtr = raidPtr;
919: p->flags = flags;
920: #ifdef KERNEL
921: p->b_proc = kb_proc;
922: #endif /* KERNEL */
923: return(p);
924: }
925:
926: void rf_FreeDiskQueueData(p)
927: RF_DiskQueueData_t *p;
928: {
929: RF_FREELIST_FREE_CLEAN(rf_dqd_freelist,p,next,clean_dqd);
930: }
CVSweb <webmaster@jp.NetBSD.org>