xref: /netbsd/sys/dev/raidframe/rf_stripelocks.c (revision bf9ec67e)
1 /*	$NetBSD: rf_stripelocks.c,v 1.12 2002/01/19 22:20:48 oster Exp $	*/
2 /*
3  * Copyright (c) 1995 Carnegie-Mellon University.
4  * All rights reserved.
5  *
6  * Authors: Mark Holland, Jim Zelenka
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  * stripelocks.c -- code to lock stripes for read and write access
31  *
32  * The code distinguishes between read locks and write locks. There can be
33  * as many readers to given stripe as desired. When a write request comes
34  * in, no further readers are allowed to enter, and all subsequent requests
35  * are queued in FIFO order. When a the number of readers goes to zero, the
36  * writer is given the lock. When a writer releases the lock, the list of
37  * queued requests is scanned, and all readersq up to the next writer are
38  * given the lock.
39  *
40  * The lock table size must be one less than a power of two, but HASH_STRIPEID
41  * is the only function that requires this.
42  *
43  * The code now supports "range locks". When you ask to lock a stripe, you
44  * specify a range of addresses in that stripe that you want to lock. When
45  * you acquire the lock, you've locked only this range of addresses, and
46  * other threads can concurrently read/write any non-overlapping portions
47  * of the stripe. The "addresses" that you lock are abstract in that you
48  * can pass in anything you like.  The expectation is that you'll pass in
49  * the range of physical disk offsets of the parity bits you're planning
50  * to update. The idea behind this, of course, is to allow sub-stripe
51  * locking. The implementation is perhaps not the best imaginable; in the
52  * worst case a lock release is O(n^2) in the total number of outstanding
53  * requests to a given stripe.  Note that if you're striping with a
54  * stripe unit size equal to an entire disk (i.e. not striping), there will
55  * be only one stripe and you may spend some significant number of cycles
56  * searching through stripe lock descriptors.
57  */
58 
59 #include <sys/cdefs.h>
60 __KERNEL_RCSID(0, "$NetBSD: rf_stripelocks.c,v 1.12 2002/01/19 22:20:48 oster Exp $");
61 
62 #include <dev/raidframe/raidframevar.h>
63 
64 #include "rf_raid.h"
65 #include "rf_stripelocks.h"
66 #include "rf_alloclist.h"
67 #include "rf_general.h"
68 #include "rf_freelist.h"
69 #include "rf_debugprint.h"
70 #include "rf_driver.h"
71 #include "rf_shutdown.h"
72 
73 #ifdef DEBUG
74 
75 #define Dprintf1(s,a)         rf_debug_printf(s,(void *)((unsigned long)a),NULL,NULL,NULL,NULL,NULL,NULL,NULL)
76 #define Dprintf2(s,a,b)       rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),NULL,NULL,NULL,NULL,NULL,NULL)
77 #define Dprintf3(s,a,b,c)     rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),NULL,NULL,NULL,NULL,NULL)
78 #define Dprintf4(s,a,b,c,d)   rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),NULL,NULL,NULL,NULL)
79 #define Dprintf5(s,a,b,c,d,e) 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)
80 #define Dprintf6(s,a,b,c,d,e,f) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),(void *)((unsigned long)e),(void *)((unsigned long)f),NULL,NULL)
81 #define Dprintf7(s,a,b,c,d,e,f,g) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),(void *)((unsigned long)e),(void *)((unsigned long)f),(void *)((unsigned long)g),NULL)
82 #define Dprintf8(s,a,b,c,d,e,f,g,h) rf_debug_printf(s,(void *)((unsigned long)a),(void *)((unsigned long)b),(void *)((unsigned long)c),(void *)((unsigned long)d),(void *)((unsigned long)e),(void *)((unsigned long)f),(void *)((unsigned long)g),(void *)((unsigned long)h))
83 
84 #else /* DEBUG */
85 
86 #define Dprintf1(s,a) {}
87 #define Dprintf2(s,a,b) {}
88 #define Dprintf3(s,a,b,c) {}
89 #define Dprintf4(s,a,b,c,d) {}
90 #define Dprintf5(s,a,b,c,d,e) {}
91 #define Dprintf6(s,a,b,c,d,e,f) {}
92 #define Dprintf7(s,a,b,c,d,e,f,g) {}
93 #define Dprintf8(s,a,b,c,d,e,f,g,h) {}
94 
95 #endif /* DEBUG */
96 
97 #define FLUSH
98 
99 #define HASH_STRIPEID(_sid_)  ( (_sid_) & (rf_lockTableSize-1) )
100 
101 static void AddToWaitersQueue(RF_LockTableEntry_t * lockTable, RF_StripeLockDesc_t * lockDesc, RF_LockReqDesc_t * lockReqDesc);
102 static RF_StripeLockDesc_t *AllocStripeLockDesc(RF_StripeNum_t stripeID);
103 static void FreeStripeLockDesc(RF_StripeLockDesc_t * p);
104 static void PrintLockedStripes(RF_LockTableEntry_t * lockTable);
105 
106 /* determines if two ranges overlap.  always yields false if either start value is negative  */
107 #define SINGLE_RANGE_OVERLAP(_strt1, _stop1, _strt2, _stop2)                                     \
108   ( (_strt1 >= 0) && (_strt2 >= 0) && (RF_MAX(_strt1, _strt2) <= RF_MIN(_stop1, _stop2)) )
109 
110 /* determines if any of the ranges specified in the two lock descriptors overlap each other */
111 #define RANGE_OVERLAP(_cand, _pred)                                                  \
112   ( SINGLE_RANGE_OVERLAP((_cand)->start,  (_cand)->stop,  (_pred)->start,  (_pred)->stop ) ||    \
113     SINGLE_RANGE_OVERLAP((_cand)->start2, (_cand)->stop2, (_pred)->start,  (_pred)->stop ) ||    \
114     SINGLE_RANGE_OVERLAP((_cand)->start,  (_cand)->stop,  (_pred)->start2, (_pred)->stop2) ||    \
115     SINGLE_RANGE_OVERLAP((_cand)->start2, (_cand)->stop2, (_pred)->start2, (_pred)->stop2) )
116 
117 /* Determines if a candidate lock request conflicts with a predecessor lock req.
118  * Note that the arguments are not interchangeable.
119  * The rules are:
120  *      a candidate read conflicts with a predecessor write if any ranges overlap
121  *      a candidate write conflicts with a predecessor read if any ranges overlap
122  *      a candidate write conflicts with a predecessor write if any ranges overlap
123  */
124 #define STRIPELOCK_CONFLICT(_cand, _pred)                                        \
125   RANGE_OVERLAP((_cand), (_pred)) &&                                        \
126   ( ( (((_cand)->type == RF_IO_TYPE_READ) && ((_pred)->type == RF_IO_TYPE_WRITE)) ||                      \
127       (((_cand)->type == RF_IO_TYPE_WRITE) && ((_pred)->type == RF_IO_TYPE_READ)) ||                      \
128       (((_cand)->type == RF_IO_TYPE_WRITE) && ((_pred)->type == RF_IO_TYPE_WRITE))                         \
129     )                                                                            \
130   )
131 
132 static RF_FreeList_t *rf_stripelock_freelist;
133 #define RF_MAX_FREE_STRIPELOCK 128
134 #define RF_STRIPELOCK_INC        8
135 #define RF_STRIPELOCK_INITIAL   32
136 
137 static void rf_ShutdownStripeLockFreeList(void *);
138 static void rf_RaidShutdownStripeLocks(void *);
139 
140 static void
141 rf_ShutdownStripeLockFreeList(ignored)
142 	void   *ignored;
143 {
144 	RF_FREELIST_DESTROY(rf_stripelock_freelist, next, (RF_StripeLockDesc_t *));
145 }
146 
147 int
148 rf_ConfigureStripeLockFreeList(listp)
149 	RF_ShutdownList_t **listp;
150 {
151 	unsigned mask;
152 	int     rc;
153 
154 	RF_FREELIST_CREATE(rf_stripelock_freelist, RF_MAX_FREE_STRIPELOCK,
155 	    RF_STRIPELOCK_INITIAL, sizeof(RF_StripeLockDesc_t));
156 	rc = rf_ShutdownCreate(listp, rf_ShutdownStripeLockFreeList, NULL);
157 	if (rc) {
158 		RF_ERRORMSG3("Unable to add to shutdown list file %s line %d rc=%d\n",
159 		    __FILE__, __LINE__, rc);
160 		rf_ShutdownStripeLockFreeList(NULL);
161 		return (rc);
162 	}
163 	RF_FREELIST_PRIME(rf_stripelock_freelist, RF_STRIPELOCK_INITIAL, next,
164 	    (RF_StripeLockDesc_t *));
165 	for (mask = 0x1; mask; mask <<= 1)
166 		if (rf_lockTableSize == mask)
167 			break;
168 	if (!mask) {
169 		printf("[WARNING:  lock table size must be a power of two.  Setting to %d.]\n", RF_DEFAULT_LOCK_TABLE_SIZE);
170 		rf_lockTableSize = RF_DEFAULT_LOCK_TABLE_SIZE;
171 	}
172 	return (0);
173 }
174 
175 RF_LockTableEntry_t *
176 rf_MakeLockTable()
177 {
178 	RF_LockTableEntry_t *lockTable;
179 	int     i, rc;
180 
181 	RF_Calloc(lockTable, ((int) rf_lockTableSize), sizeof(RF_LockTableEntry_t), (RF_LockTableEntry_t *));
182 	if (lockTable == NULL)
183 		return (NULL);
184 	for (i = 0; i < rf_lockTableSize; i++) {
185 		rc = rf_mutex_init(&lockTable[i].mutex);
186 		if (rc) {
187 			RF_ERRORMSG3("Unable to init mutex file %s line %d rc=%d\n", __FILE__,
188 			    __LINE__, rc);
189 			/* XXX clean up other mutexes */
190 			return (NULL);
191 		}
192 	}
193 	return (lockTable);
194 }
195 
196 void
197 rf_ShutdownStripeLocks(RF_LockTableEntry_t * lockTable)
198 {
199 	int     i;
200 
201 	if (rf_stripeLockDebug) {
202 		PrintLockedStripes(lockTable);
203 	}
204 	for (i = 0; i < rf_lockTableSize; i++) {
205 		rf_mutex_destroy(&lockTable[i].mutex);
206 	}
207 	RF_Free(lockTable, rf_lockTableSize * sizeof(RF_LockTableEntry_t));
208 }
209 
210 static void
211 rf_RaidShutdownStripeLocks(arg)
212 	void   *arg;
213 {
214 	RF_Raid_t *raidPtr = (RF_Raid_t *) arg;
215 	rf_ShutdownStripeLocks(raidPtr->lockTable);
216 }
217 
218 int
219 rf_ConfigureStripeLocks(
220     RF_ShutdownList_t ** listp,
221     RF_Raid_t * raidPtr,
222     RF_Config_t * cfgPtr)
223 {
224 	int     rc;
225 
226 	raidPtr->lockTable = rf_MakeLockTable();
227 	if (raidPtr->lockTable == NULL)
228 		return (ENOMEM);
229 	rc = rf_ShutdownCreate(listp, rf_RaidShutdownStripeLocks, raidPtr);
230 	if (rc) {
231 		RF_ERRORMSG3("Unable to add to shutdown list file %s line %d rc=%d\n",
232 		    __FILE__, __LINE__, rc);
233 		rf_ShutdownStripeLocks(raidPtr->lockTable);
234 		return (rc);
235 	}
236 	return (0);
237 }
238 /* returns 0 if you've got the lock, and non-zero if you have to wait.
239  * if and only if you have to wait, we'll cause cbFunc to get invoked
240  * with cbArg when you are granted the lock.  We store a tag in *releaseTag
241  * that you need to give back to us when you release the lock.
242  */
243 int
244 rf_AcquireStripeLock(
245     RF_LockTableEntry_t * lockTable,
246     RF_StripeNum_t stripeID,
247     RF_LockReqDesc_t * lockReqDesc)
248 {
249 	RF_StripeLockDesc_t *lockDesc;
250 	RF_LockReqDesc_t *p;
251 #ifdef DEBUG
252 	int     tid = 0;
253 #endif
254 	int     hashval = HASH_STRIPEID(stripeID);
255 	int     retcode = 0;
256 
257 	RF_ASSERT(RF_IO_IS_R_OR_W(lockReqDesc->type));
258 
259 	if (rf_stripeLockDebug) {
260 		if (stripeID == -1) {
261 			Dprintf1("[%d] Lock acquisition supressed (stripeID == -1)\n", tid);
262 		} else {
263 			Dprintf8("[%d] Trying to acquire stripe lock table 0x%lx SID %ld type %c range %ld-%ld, range2 %ld-%ld hashval %d\n",
264 			    tid, (unsigned long) lockTable, stripeID, lockReqDesc->type, lockReqDesc->start,
265 			    lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2);
266 			Dprintf3("[%d] lock %ld hashval %d\n", tid, stripeID, hashval);
267 			FLUSH;
268 		}
269 	}
270 	if (stripeID == -1)
271 		return (0);
272 	lockReqDesc->next = NULL;	/* just to be sure */
273 
274 	RF_LOCK_MUTEX(lockTable[hashval].mutex);
275 	for (lockDesc = lockTable[hashval].descList; lockDesc; lockDesc = lockDesc->next) {
276 		if (lockDesc->stripeID == stripeID)
277 			break;
278 	}
279 
280 	if (!lockDesc) {	/* no entry in table => no one reading or
281 				 * writing */
282 		lockDesc = AllocStripeLockDesc(stripeID);
283 		lockDesc->next = lockTable[hashval].descList;
284 		lockTable[hashval].descList = lockDesc;
285 		if (lockReqDesc->type == RF_IO_TYPE_WRITE)
286 			lockDesc->nWriters++;
287 		lockDesc->granted = lockReqDesc;
288 		if (rf_stripeLockDebug) {
289 			Dprintf7("[%d] no one waiting: lock %ld %c %ld-%ld %ld-%ld granted\n",
290 			    tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2);
291 			FLUSH;
292 		}
293 	} else {
294 
295 		if (lockReqDesc->type == RF_IO_TYPE_WRITE)
296 			lockDesc->nWriters++;
297 
298 		if (lockDesc->nWriters == 0) {	/* no need to search any lists
299 						 * if there are no writers
300 						 * anywhere */
301 			lockReqDesc->next = lockDesc->granted;
302 			lockDesc->granted = lockReqDesc;
303 			if (rf_stripeLockDebug) {
304 				Dprintf7("[%d] no writers: lock %ld %c %ld-%ld %ld-%ld granted\n",
305 				    tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2);
306 				FLUSH;
307 			}
308 		} else {
309 
310 			/* search the granted & waiting lists for a conflict.
311 			 * stop searching as soon as we find one */
312 			retcode = 0;
313 			for (p = lockDesc->granted; p; p = p->next)
314 				if (STRIPELOCK_CONFLICT(lockReqDesc, p)) {
315 					retcode = 1;
316 					break;
317 				}
318 			if (!retcode)
319 				for (p = lockDesc->waitersH; p; p = p->next)
320 					if (STRIPELOCK_CONFLICT(lockReqDesc, p)) {
321 						retcode = 2;
322 						break;
323 					}
324 			if (!retcode) {
325 				lockReqDesc->next = lockDesc->granted;	/* no conflicts found =>
326 									 * grant lock */
327 				lockDesc->granted = lockReqDesc;
328 				if (rf_stripeLockDebug) {
329 					Dprintf7("[%d] no conflicts: lock %ld %c %ld-%ld %ld-%ld granted\n",
330 					    tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop,
331 					    lockReqDesc->start2, lockReqDesc->stop2);
332 					FLUSH;
333 				}
334 			} else {
335 				if (rf_stripeLockDebug) {
336 					Dprintf6("[%d] conflict: lock %ld %c %ld-%ld hashval=%d not granted\n",
337 					    tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop,
338 					    hashval);
339 					Dprintf3("[%d] lock %ld retcode=%d\n", tid, stripeID, retcode);
340 					FLUSH;
341 				}
342 				AddToWaitersQueue(lockTable, lockDesc, lockReqDesc);	/* conflict => the
343 											 * current access must
344 											 * wait */
345 			}
346 		}
347 	}
348 
349 	RF_UNLOCK_MUTEX(lockTable[hashval].mutex);
350 	return (retcode);
351 }
352 
353 void
354 rf_ReleaseStripeLock(
355     RF_LockTableEntry_t * lockTable,
356     RF_StripeNum_t stripeID,
357     RF_LockReqDesc_t * lockReqDesc)
358 {
359 	RF_StripeLockDesc_t *lockDesc, *ld_t;
360 	RF_LockReqDesc_t *lr, *lr_t, *callbacklist, *t;
361 #ifdef DEBUG
362 	int     tid = 0;
363 #endif
364 	int     hashval = HASH_STRIPEID(stripeID);
365 	int     release_it, consider_it;
366 	RF_LockReqDesc_t *candidate, *candidate_t, *predecessor;
367 
368 	RF_ASSERT(RF_IO_IS_R_OR_W(lockReqDesc->type));
369 
370 	if (rf_stripeLockDebug) {
371 		if (stripeID == -1) {
372 			Dprintf1("[%d] Lock release supressed (stripeID == -1)\n", tid);
373 		} else {
374 			Dprintf8("[%d] Releasing stripe lock on stripe ID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
375 			    tid, stripeID, lockReqDesc->type, lockReqDesc->start, lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2, lockTable);
376 			FLUSH;
377 		}
378 	}
379 	if (stripeID == -1)
380 		return;
381 
382 	RF_LOCK_MUTEX(lockTable[hashval].mutex);
383 
384 	/* find the stripe lock descriptor */
385 	for (ld_t = NULL, lockDesc = lockTable[hashval].descList; lockDesc; ld_t = lockDesc, lockDesc = lockDesc->next) {
386 		if (lockDesc->stripeID == stripeID)
387 			break;
388 	}
389 	RF_ASSERT(lockDesc);	/* major error to release a lock that doesn't
390 				 * exist */
391 
392 	/* find the stripe lock request descriptor & delete it from the list */
393 	for (lr_t = NULL, lr = lockDesc->granted; lr; lr_t = lr, lr = lr->next)
394 		if (lr == lockReqDesc)
395 			break;
396 
397 	RF_ASSERT(lr && (lr == lockReqDesc));	/* major error to release a
398 						 * lock that hasn't been
399 						 * granted */
400 	if (lr_t)
401 		lr_t->next = lr->next;
402 	else {
403 		RF_ASSERT(lr == lockDesc->granted);
404 		lockDesc->granted = lr->next;
405 	}
406 	lr->next = NULL;
407 
408 	if (lockReqDesc->type == RF_IO_TYPE_WRITE)
409 		lockDesc->nWriters--;
410 
411 	/* search through the waiters list to see if anyone needs to be woken
412 	 * up. for each such descriptor in the wait list, we check it against
413 	 * everything granted and against everything _in front_ of it in the
414 	 * waiters queue.  If it conflicts with none of these, we release it.
415 	 *
416 	 * DON'T TOUCH THE TEMPLINK POINTER OF ANYTHING IN THE GRANTED LIST HERE.
417 	 * This will roach the case where the callback tries to acquire a new
418 	 * lock in the same stripe.  There are some asserts to try and detect
419 	 * this.
420 	 *
421 	 * We apply 2 performance optimizations: (1) if releasing this lock
422 	 * results in no more writers to this stripe, we just release
423 	 * everybody waiting, since we place no restrictions on the number of
424 	 * concurrent reads. (2) we consider as candidates for wakeup only
425 	 * those waiters that have a range overlap with either the descriptor
426 	 * being woken up or with something in the callbacklist (i.e.
427 	 * something we've just now woken up). This allows us to avoid the
428 	 * long evaluation for some descriptors. */
429 
430 	callbacklist = NULL;
431 	if (lockDesc->nWriters == 0) {	/* performance tweak (1) */
432 		while (lockDesc->waitersH) {
433 
434 			lr = lockDesc->waitersH;	/* delete from waiters
435 							 * list */
436 			lockDesc->waitersH = lr->next;
437 
438 			RF_ASSERT(lr->type == RF_IO_TYPE_READ);
439 
440 			lr->next = lockDesc->granted;	/* add to granted list */
441 			lockDesc->granted = lr;
442 
443 			RF_ASSERT(!lr->templink);
444 			lr->templink = callbacklist;	/* put on callback list
445 							 * so that we'll invoke
446 							 * callback below */
447 			callbacklist = lr;
448 			if (rf_stripeLockDebug) {
449 				Dprintf8("[%d] No writers: granting lock stripe ID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
450 				    tid, stripeID, lr->type, lr->start, lr->stop, lr->start2, lr->stop2, (unsigned long) lockTable);
451 				FLUSH;
452 			}
453 		}
454 		lockDesc->waitersT = NULL;	/* we've purged the whole
455 						 * waiters list */
456 
457 	} else
458 		for (candidate_t = NULL, candidate = lockDesc->waitersH; candidate;) {
459 
460 			/* performance tweak (2) */
461 			consider_it = 0;
462 			if (RANGE_OVERLAP(lockReqDesc, candidate))
463 				consider_it = 1;
464 			else
465 				for (t = callbacklist; t; t = t->templink)
466 					if (RANGE_OVERLAP(t, candidate)) {
467 						consider_it = 1;
468 						break;
469 					}
470 			if (!consider_it) {
471 				if (rf_stripeLockDebug) {
472 					Dprintf8("[%d] No overlap: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
473 					    tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
474 					    (unsigned long) lockTable);
475 					FLUSH;
476 				}
477 				candidate_t = candidate;
478 				candidate = candidate->next;
479 				continue;
480 			}
481 			/* we have a candidate for release.  check to make
482 			 * sure it is not blocked by any granted locks */
483 			release_it = 1;
484 			for (predecessor = lockDesc->granted; predecessor; predecessor = predecessor->next) {
485 				if (STRIPELOCK_CONFLICT(candidate, predecessor)) {
486 					if (rf_stripeLockDebug) {
487 						Dprintf8("[%d] Conflicts with granted lock: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
488 						    tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
489 						    (unsigned long) lockTable);
490 						FLUSH;
491 					}
492 					release_it = 0;
493 					break;
494 				}
495 			}
496 
497 			/* now check to see if the candidate is blocked by any
498 			 * waiters that occur before it it the wait queue */
499 			if (release_it)
500 				for (predecessor = lockDesc->waitersH; predecessor != candidate; predecessor = predecessor->next) {
501 					if (STRIPELOCK_CONFLICT(candidate, predecessor)) {
502 						if (rf_stripeLockDebug) {
503 							Dprintf8("[%d] Conflicts with waiting lock: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
504 							    tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
505 							    (unsigned long) lockTable);
506 							FLUSH;
507 						}
508 						release_it = 0;
509 						break;
510 					}
511 				}
512 
513 			/* release it if indicated */
514 			if (release_it) {
515 				if (rf_stripeLockDebug) {
516 					Dprintf8("[%d] Granting lock to candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
517 					    tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
518 					    (unsigned long) lockTable);
519 					FLUSH;
520 				}
521 				if (candidate_t) {
522 					candidate_t->next = candidate->next;
523 					if (lockDesc->waitersT == candidate)
524 						lockDesc->waitersT = candidate_t;	/* cannot be waitersH
525 											 * since candidate_t is
526 											 * not NULL */
527 				} else {
528 					RF_ASSERT(candidate == lockDesc->waitersH);
529 					lockDesc->waitersH = lockDesc->waitersH->next;
530 					if (!lockDesc->waitersH)
531 						lockDesc->waitersT = NULL;
532 				}
533 				candidate->next = lockDesc->granted;	/* move it to the
534 									 * granted list */
535 				lockDesc->granted = candidate;
536 
537 				RF_ASSERT(!candidate->templink);
538 				candidate->templink = callbacklist;	/* put it on the list of
539 									 * things to be called
540 									 * after we release the
541 									 * mutex */
542 				callbacklist = candidate;
543 
544 				if (!candidate_t)
545 					candidate = lockDesc->waitersH;
546 				else
547 					candidate = candidate_t->next;	/* continue with the
548 									 * rest of the list */
549 			} else {
550 				candidate_t = candidate;
551 				candidate = candidate->next;	/* continue with the
552 								 * rest of the list */
553 			}
554 		}
555 
556 	/* delete the descriptor if no one is waiting or active */
557 	if (!lockDesc->granted && !lockDesc->waitersH) {
558 		RF_ASSERT(lockDesc->nWriters == 0);
559 		if (rf_stripeLockDebug) {
560 			Dprintf3("[%d] Last lock released (table 0x%lx): deleting desc for stripeID %ld\n", tid, (unsigned long) lockTable, stripeID);
561 			FLUSH;
562 		}
563 		if (ld_t)
564 			ld_t->next = lockDesc->next;
565 		else {
566 			RF_ASSERT(lockDesc == lockTable[hashval].descList);
567 			lockTable[hashval].descList = lockDesc->next;
568 		}
569 		FreeStripeLockDesc(lockDesc);
570 		lockDesc = NULL;/* only for the ASSERT below */
571 	}
572 	RF_UNLOCK_MUTEX(lockTable[hashval].mutex);
573 
574 	/* now that we've unlocked the mutex, invoke the callback on all the
575 	 * descriptors in the list */
576 	RF_ASSERT(!((callbacklist) && (!lockDesc)));	/* if we deleted the
577 							 * descriptor, we should
578 							 * have no callbacks to
579 							 * do */
580 	for (candidate = callbacklist; candidate;) {
581 		t = candidate;
582 		candidate = candidate->templink;
583 		t->templink = NULL;
584 		(t->cbFunc) (t->cbArg);
585 	}
586 }
587 /* must have the indicated lock table mutex upon entry */
588 static void
589 AddToWaitersQueue(
590     RF_LockTableEntry_t * lockTable,
591     RF_StripeLockDesc_t * lockDesc,
592     RF_LockReqDesc_t * lockReqDesc)
593 {
594 	if (!lockDesc->waitersH) {
595 		lockDesc->waitersH = lockDesc->waitersT = lockReqDesc;
596 	} else {
597 		lockDesc->waitersT->next = lockReqDesc;
598 		lockDesc->waitersT = lockReqDesc;
599 	}
600 }
601 
602 static RF_StripeLockDesc_t *
603 AllocStripeLockDesc(RF_StripeNum_t stripeID)
604 {
605 	RF_StripeLockDesc_t *p;
606 
607 	RF_FREELIST_GET(rf_stripelock_freelist, p, next, (RF_StripeLockDesc_t *));
608 	if (p) {
609 		p->stripeID = stripeID;
610 	}
611 	return (p);
612 }
613 
614 static void
615 FreeStripeLockDesc(RF_StripeLockDesc_t * p)
616 {
617 	RF_FREELIST_FREE(rf_stripelock_freelist, p, next);
618 }
619 
620 static void
621 PrintLockedStripes(lockTable)
622 	RF_LockTableEntry_t *lockTable;
623 {
624 	int     i, j, foundone = 0, did;
625 	RF_StripeLockDesc_t *p;
626 	RF_LockReqDesc_t *q;
627 
628 	RF_LOCK_MUTEX(rf_printf_mutex);
629 	printf("Locked stripes:\n");
630 	for (i = 0; i < rf_lockTableSize; i++)
631 		if (lockTable[i].descList) {
632 			foundone = 1;
633 			for (p = lockTable[i].descList; p; p = p->next) {
634 				printf("Stripe ID 0x%lx (%d) nWriters %d\n",
635 				    (long) p->stripeID, (int) p->stripeID, p->nWriters);
636 
637 				if (!(p->granted))
638 					printf("Granted: (none)\n");
639 				else
640 					printf("Granted:\n");
641 				for (did = 1, j = 0, q = p->granted; q; j++, q = q->next) {
642 					printf("  %c(%ld-%ld", q->type, (long) q->start, (long) q->stop);
643 					if (q->start2 != -1)
644 						printf(",%ld-%ld) ", (long) q->start2,
645 						    (long) q->stop2);
646 					else
647 						printf(") ");
648 					if (j && !(j % 4)) {
649 						printf("\n");
650 						did = 1;
651 					} else
652 						did = 0;
653 				}
654 				if (!did)
655 					printf("\n");
656 
657 				if (!(p->waitersH))
658 					printf("Waiting: (none)\n");
659 				else
660 					printf("Waiting:\n");
661 				for (did = 1, j = 0, q = p->waitersH; q; j++, q = q->next) {
662 					printf("%c(%ld-%ld", q->type, (long) q->start, (long) q->stop);
663 					if (q->start2 != -1)
664 						printf(",%ld-%ld) ", (long) q->start2, (long) q->stop2);
665 					else
666 						printf(") ");
667 					if (j && !(j % 4)) {
668 						printf("\n         ");
669 						did = 1;
670 					} else
671 						did = 0;
672 				}
673 				if (!did)
674 					printf("\n");
675 			}
676 		}
677 	if (!foundone)
678 		printf("(none)\n");
679 	else
680 		printf("\n");
681 	RF_UNLOCK_MUTEX(rf_printf_mutex);
682 }
683