xref: /netbsd-src/sys/dev/raidframe/rf_diskqueue.c (revision 181254a7b1bdde6873432bffef2d2decc4b5c22f)
1 /*	$NetBSD: rf_diskqueue.c,v 1.58 2020/06/19 19:32:03 jdolecek Exp $	*/
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
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.
65  *
66  ****************************************************************************/
67 
68 #include <sys/cdefs.h>
69 __KERNEL_RCSID(0, "$NetBSD: rf_diskqueue.c,v 1.58 2020/06/19 19:32:03 jdolecek Exp $");
70 
71 #include <dev/raidframe/raidframevar.h>
72 
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"
85 #include "rf_kintf.h"
86 
87 static void rf_ShutdownDiskQueueSystem(void *);
88 
89 #ifndef RF_DEBUG_DISKQUEUE
90 #define RF_DEBUG_DISKQUEUE 0
91 #endif
92 
93 #if RF_DEBUG_DISKQUEUE
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)
97 #else
98 #define Dprintf1(s,a)
99 #define Dprintf2(s,a,b)
100 #define Dprintf3(s,a,b,c)
101 #endif
102 
103 /*****************************************************************************
104  *
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
108  *
109  ****************************************************************************/
110 
111 static const RF_DiskQueueSW_t diskqueuesw[] = {
112 	{"fifo",		/* FIFO */
113 		rf_FifoCreate,
114 		rf_FifoEnqueue,
115 		rf_FifoDequeue,
116 		rf_FifoPeek,
117 	rf_FifoPromote},
118 
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,
131 	rf_SstfPromote},
132 
133 	{"scan",		/* SCAN (two-way elevator) */
134 		rf_ScanCreate,
135 		rf_SstfEnqueue,
136 		rf_ScanDequeue,
137 		rf_ScanPeek,
138 	rf_SstfPromote},
139 
140 	{"cscan",		/* CSCAN (one-way elevator) */
141 		rf_CscanCreate,
142 		rf_SstfEnqueue,
143 		rf_CscanDequeue,
144 		rf_CscanPeek,
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
151 #define RF_MIN_FREE_DQD  64
152 
153 #include <sys/buf.h>
154 
155 /* configures a single disk queue */
156 
157 static void
158 rf_ShutdownDiskQueue(void *arg)
159 {
160 	RF_DiskQueue_t *diskqueue = arg;
161 
162 	rf_destroy_mutex2(diskqueue->mutex);
163 }
164 
165 int
166 rf_ConfigureDiskQueue(RF_Raid_t *raidPtr, RF_DiskQueue_t *diskqueue,
167 		      RF_RowCol_t c, const RF_DiskQueueSW_t *p,
168 		      RF_SectorCount_t sectPerDisk, dev_t dev,
169 		      int maxOutstanding, RF_ShutdownList_t **listp,
170 		      RF_AllocListElem_t *clList)
171 {
172 	diskqueue->col = c;
173 	diskqueue->qPtr = p;
174 	diskqueue->qHdr = (p->Create) (sectPerDisk, clList, listp);
175 	diskqueue->dev = dev;
176 	diskqueue->numOutstanding = 0;
177 	diskqueue->queueLength = 0;
178 	diskqueue->maxOutstanding = maxOutstanding;
179 	diskqueue->curPriority = RF_IO_NORMAL_PRIORITY;
180 	diskqueue->flags = 0;
181 	diskqueue->raidPtr = raidPtr;
182 	diskqueue->rf_cinfo = &raidPtr->raid_cinfo[c];
183 	rf_init_mutex2(diskqueue->mutex, IPL_VM);
184 	rf_ShutdownCreate(listp, rf_ShutdownDiskQueue, diskqueue);
185 	return (0);
186 }
187 
188 static void
189 rf_ShutdownDiskQueueSystem(void *ignored)
190 {
191 	pool_destroy(&rf_pools.dqd);
192 }
193 
194 int
195 rf_ConfigureDiskQueueSystem(RF_ShutdownList_t **listp)
196 {
197 
198 	rf_pool_init(&rf_pools.dqd, sizeof(RF_DiskQueueData_t),
199 		     "rf_dqd_pl", RF_MIN_FREE_DQD, RF_MAX_FREE_DQD);
200 	rf_ShutdownCreate(listp, rf_ShutdownDiskQueueSystem, NULL);
201 
202 	return (0);
203 }
204 
205 int
206 rf_ConfigureDiskQueues(RF_ShutdownList_t **listp, RF_Raid_t *raidPtr,
207 		       RF_Config_t *cfgPtr)
208 {
209 	RF_DiskQueue_t *diskQueues, *spareQueues;
210 	const RF_DiskQueueSW_t *p;
211 	RF_RowCol_t r,c;
212 	int     rc, i;
213 
214 	raidPtr->maxQueueDepth = cfgPtr->maxOutstandingDiskReqs;
215 
216 	for (p = NULL, i = 0; i < NUM_DISK_QUEUE_TYPES; i++) {
217 		if (!strcmp(diskqueuesw[i].queueType, cfgPtr->diskQueueType)) {
218 			p = &diskqueuesw[i];
219 			break;
220 		}
221 	}
222 	if (p == NULL) {
223 		RF_ERRORMSG2("Unknown queue type \"%s\".  Using %s\n", cfgPtr->diskQueueType, diskqueuesw[0].queueType);
224 		p = &diskqueuesw[0];
225 	}
226 	raidPtr->qType = p;
227 
228 	diskQueues = RF_MallocAndAdd(
229 	    (raidPtr->numCol + RF_MAXSPARE) * sizeof(*diskQueues),
230 	    raidPtr->cleanupList);
231 	if (diskQueues == NULL)
232 		return (ENOMEM);
233 	raidPtr->Queues = diskQueues;
234 
235 	for (c = 0; c < raidPtr->numCol; c++) {
236 		rc = rf_ConfigureDiskQueue(raidPtr, &diskQueues[c],
237 					   c, p,
238 					   raidPtr->sectorsPerDisk,
239 					   raidPtr->Disks[c].dev,
240 					   cfgPtr->maxOutstandingDiskReqs,
241 					   listp, raidPtr->cleanupList);
242 		if (rc)
243 			return (rc);
244 	}
245 
246 	spareQueues = &raidPtr->Queues[raidPtr->numCol];
247 	for (r = 0; r < raidPtr->numSpare; r++) {
248 		rc = rf_ConfigureDiskQueue(raidPtr, &spareQueues[r],
249 					   raidPtr->numCol + r, p,
250 					   raidPtr->sectorsPerDisk,
251 					   raidPtr->Disks[raidPtr->numCol + r].dev,
252 					   cfgPtr->maxOutstandingDiskReqs, listp,
253 					   raidPtr->cleanupList);
254 		if (rc)
255 			return (rc);
256 	}
257 	return (0);
258 }
259 /* Enqueue a disk I/O
260  *
261  * In the kernel, I/O is non-blocking and so we'd like to have multiple
262  * I/Os outstanding on the physical disks when possible.
263  *
264  * when any request arrives at a queue, we have two choices:
265  *    dispatch it to the lower levels
266  *    queue it up
267  *
268  * kernel rules for when to do what:
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 void
275 rf_DiskIOEnqueue(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int pri)
276 {
277 	RF_ETIMER_START(req->qtime);
278 	RF_ASSERT(req->type == RF_IO_TYPE_NOP || req->numSector);
279 	req->priority = pri;
280 
281 #if RF_DEBUG_DISKQUEUE
282 	if (rf_queueDebug && (req->numSector == 0)) {
283 		printf("Warning: Enqueueing zero-sector access\n");
284 	}
285 #endif
286 	RF_LOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
287 	if (RF_OK_TO_DISPATCH(queue, req)) {
288 		Dprintf2("Dispatching pri %d regular op to c %d (ok to dispatch)\n", pri, queue->col);
289 		rf_DispatchKernelIO(queue, req);
290 	} else {
291 		queue->queueLength++;	/* increment count of number of requests waiting in this queue */
292 		Dprintf2("Enqueueing pri %d regular op to c %d (not ok to dispatch)\n", pri, queue->col);
293 		req->queue = (void *) queue;
294 		(queue->qPtr->Enqueue) (queue->qHdr, req, pri);
295 	}
296 	RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOEnqueue");
297 }
298 
299 
300 /* get the next set of I/Os started */
301 void
302 rf_DiskIOComplete(RF_DiskQueue_t *queue, RF_DiskQueueData_t *req, int status)
303 {
304 	int     done = 0;
305 
306 	RF_LOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
307 	queue->numOutstanding--;
308 	RF_ASSERT(queue->numOutstanding >= 0);
309 
310 	/* dispatch requests to the disk until we find one that we can't. */
311 	/* no reason to continue once we've filled up the queue */
312 	/* no reason to even start if the queue is locked */
313 
314 	while (!done && !RF_QUEUE_FULL(queue)) {
315 		req = (queue->qPtr->Dequeue) (queue->qHdr);
316 		if (req) {
317 			Dprintf2("DiskIOComplete: extracting pri %d req from queue at c %d\n", req->priority, queue->col);
318 			queue->queueLength--;	/* decrement count of number of requests waiting in this queue */
319 			RF_ASSERT(queue->queueLength >= 0);
320 			if (RF_OK_TO_DISPATCH(queue, req)) {
321 				Dprintf2("DiskIOComplete: dispatching pri %d regular req to c %d (ok to dispatch)\n", req->priority, queue->col);
322 				rf_DispatchKernelIO(queue, req);
323 			} else {
324 				/* we can't dispatch it, so just re-enqueue it.
325 				   potential trouble here if disk queues batch reqs */
326 				Dprintf2("DiskIOComplete: re-enqueueing pri %d regular req to c %d\n", req->priority, queue->col);
327 				queue->queueLength++;
328 				(queue->qPtr->Enqueue) (queue->qHdr, req, req->priority);
329 				done = 1;
330 			}
331 		} else {
332 			Dprintf1("DiskIOComplete: no more requests to extract.\n", "");
333 			done = 1;
334 		}
335 	}
336 
337 	RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOComplete");
338 }
339 /* promotes accesses tagged with the given parityStripeID from low priority
340  * to normal priority.  This promotion is optional, meaning that a queue
341  * need not implement it.  If there is no promotion routine associated with
342  * a queue, this routine does nothing and returns -1.
343  */
344 int
345 rf_DiskIOPromote(RF_DiskQueue_t *queue, RF_StripeNum_t parityStripeID,
346 		 RF_ReconUnitNum_t which_ru)
347 {
348 	int     retval;
349 
350 	if (!queue->qPtr->Promote)
351 		return (-1);
352 	RF_LOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
353 	retval = (queue->qPtr->Promote) (queue->qHdr, parityStripeID, which_ru);
354 	RF_UNLOCK_QUEUE_MUTEX(queue, "DiskIOPromote");
355 	return (retval);
356 }
357 
358 RF_DiskQueueData_t *
359 rf_CreateDiskQueueData(RF_IoType_t typ, RF_SectorNum_t ssect,
360 		       RF_SectorCount_t nsect, void *bf,
361 		       RF_StripeNum_t parityStripeID,
362 		       RF_ReconUnitNum_t which_ru,
363 		       void (*wakeF) (void *, int), void *arg,
364 		       RF_AccTraceEntry_t *tracerec, RF_Raid_t *raidPtr,
365 		       RF_DiskQueueDataFlags_t flags, const struct buf *mbp,
366 		       int waitflag)
367 {
368 	RF_DiskQueueData_t *p;
369 
370 	p = pool_get(&rf_pools.dqd, waitflag | PR_ZERO);
371 	if (p == NULL)
372 		return (NULL);
373 
374 	if (waitflag == PR_WAITOK) {
375 		p->bp = getiobuf(NULL, true);
376 	} else {
377 		p->bp = getiobuf(NULL, false);
378 	}
379 	if (p->bp == NULL) {
380 		pool_put(&rf_pools.dqd, p);
381 		return (NULL);
382 	}
383 	SET(p->bp->b_cflags, BC_BUSY);	/* mark buffer busy */
384 	if (mbp) {
385 		SET(p->bp->b_flags, mbp->b_flags & rf_b_pass);
386 		p->bp->b_proc = mbp->b_proc;
387 	}
388 
389 	p->sectorOffset = ssect + rf_protectedSectors;
390 	p->numSector = nsect;
391 	p->type = typ;
392 	p->buf = bf;
393 	p->parityStripeID = parityStripeID;
394 	p->which_ru = which_ru;
395 	p->CompleteFunc = wakeF;
396 	p->argument = arg;
397 	p->next = NULL;
398 	p->tracerec = tracerec;
399 	p->priority = RF_IO_NORMAL_PRIORITY;
400 	p->raidPtr = raidPtr;
401 	p->flags = flags;
402 	return (p);
403 }
404 
405 void
406 rf_FreeDiskQueueData(RF_DiskQueueData_t *p)
407 {
408 	putiobuf(p->bp);
409 	pool_put(&rf_pools.dqd, p);
410 }
411