xref: /netbsd-src/sys/dev/raidframe/rf_dagfuncs.c (revision c7fb772b85b2b5d4cfb282f868f454b4701534fd)
1 /*	$NetBSD: rf_dagfuncs.c,v 1.35 2021/08/07 16:19:15 thorpej Exp $	*/
2 /*
3  * Copyright (c) 1995 Carnegie-Mellon University.
4  * All rights reserved.
5  *
6  * Author: Mark Holland, William V. Courtright II
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  * dagfuncs.c -- DAG node execution routines
31  *
32  * Rules:
33  * 1. Every DAG execution function must eventually cause node->status to
34  *    get set to "good" or "bad", and "FinishNode" to be called. In the
35  *    case of nodes that complete immediately (xor, NullNodeFunc, etc),
36  *    the node execution function can do these two things directly. In
37  *    the case of nodes that have to wait for some event (a disk read to
38  *    complete, a lock to be released, etc) to occur before they can
39  *    complete, this is typically achieved by having whatever module
40  *    is doing the operation call GenericWakeupFunc upon completion.
41  * 2. DAG execution functions should check the status in the DAG header
42  *    and NOP out their operations if the status is not "enable". However,
43  *    execution functions that release resources must be sure to release
44  *    them even when they NOP out the function that would use them.
45  *    Functions that acquire resources should go ahead and acquire them
46  *    even when they NOP, so that a downstream release node will not have
47  *    to check to find out whether or not the acquire was suppressed.
48  */
49 
50 #include <sys/cdefs.h>
51 __KERNEL_RCSID(0, "$NetBSD: rf_dagfuncs.c,v 1.35 2021/08/07 16:19:15 thorpej Exp $");
52 
53 #include <sys/param.h>
54 #include <sys/ioctl.h>
55 
56 #include "rf_archs.h"
57 #include "rf_raid.h"
58 #include "rf_dag.h"
59 #include "rf_layout.h"
60 #include "rf_etimer.h"
61 #include "rf_acctrace.h"
62 #include "rf_diskqueue.h"
63 #include "rf_dagfuncs.h"
64 #include "rf_general.h"
65 #include "rf_engine.h"
66 #include "rf_dagutils.h"
67 
68 #include "rf_kintf.h"
69 
70 #if RF_INCLUDE_PARITYLOGGING > 0
71 #include "rf_paritylog.h"
72 #endif				/* RF_INCLUDE_PARITYLOGGING > 0 */
73 
74 void     (*rf_DiskReadFunc) (RF_DagNode_t *);
75 void     (*rf_DiskWriteFunc) (RF_DagNode_t *);
76 void     (*rf_DiskReadUndoFunc) (RF_DagNode_t *);
77 void     (*rf_DiskWriteUndoFunc) (RF_DagNode_t *);
78 void     (*rf_RegularXorUndoFunc) (RF_DagNode_t *);
79 void     (*rf_SimpleXorUndoFunc) (RF_DagNode_t *);
80 void     (*rf_RecoveryXorUndoFunc) (RF_DagNode_t *);
81 
82 /*****************************************************************************
83  * main (only) configuration routine for this module
84  ****************************************************************************/
85 int
rf_ConfigureDAGFuncs(RF_ShutdownList_t ** listp)86 rf_ConfigureDAGFuncs(RF_ShutdownList_t **listp)
87 {
88 	RF_ASSERT(((sizeof(long) == 8) && RF_LONGSHIFT == 3) ||
89 		  ((sizeof(long) == 4) && RF_LONGSHIFT == 2));
90 	rf_DiskReadFunc = rf_DiskReadFuncForThreads;
91 	rf_DiskReadUndoFunc = rf_DiskUndoFunc;
92 	rf_DiskWriteFunc = rf_DiskWriteFuncForThreads;
93 	rf_DiskWriteUndoFunc = rf_DiskUndoFunc;
94 	rf_RegularXorUndoFunc = rf_NullNodeUndoFunc;
95 	rf_SimpleXorUndoFunc = rf_NullNodeUndoFunc;
96 	rf_RecoveryXorUndoFunc = rf_NullNodeUndoFunc;
97 	return (0);
98 }
99 
100 
101 
102 /*****************************************************************************
103  * the execution function associated with a terminate node
104  ****************************************************************************/
105 void
rf_TerminateFunc(RF_DagNode_t * node)106 rf_TerminateFunc(RF_DagNode_t *node)
107 {
108 	RF_ASSERT(node->dagHdr->numCommits == node->dagHdr->numCommitNodes);
109 	node->status = rf_good;
110 	rf_FinishNode(node, RF_THREAD_CONTEXT);
111 }
112 
113 void
rf_TerminateUndoFunc(RF_DagNode_t * node)114 rf_TerminateUndoFunc(RF_DagNode_t *node)
115 {
116 }
117 
118 
119 /*****************************************************************************
120  * execution functions associated with a mirror node
121  *
122  * parameters:
123  *
124  * 0 - physical disk address of data
125  * 1 - buffer for holding read data
126  * 2 - parity stripe ID
127  * 3 - flags
128  * 4 - physical disk address of mirror (parity)
129  *
130  ****************************************************************************/
131 
132 void
rf_DiskReadMirrorIdleFunc(RF_DagNode_t * node)133 rf_DiskReadMirrorIdleFunc(RF_DagNode_t *node)
134 {
135 	/* select the mirror copy with the shortest queue and fill in node
136 	 * parameters with physical disk address */
137 
138 	rf_SelectMirrorDiskIdle(node);
139 	rf_DiskReadFunc(node);
140 }
141 
142 #if (RF_INCLUDE_CHAINDECLUSTER > 0) || (RF_INCLUDE_INTERDECLUSTER > 0) || (RF_DEBUG_VALIDATE_DAG > 0)
143 void
rf_DiskReadMirrorPartitionFunc(RF_DagNode_t * node)144 rf_DiskReadMirrorPartitionFunc(RF_DagNode_t *node)
145 {
146 	/* select the mirror copy with the shortest queue and fill in node
147 	 * parameters with physical disk address */
148 
149 	rf_SelectMirrorDiskPartition(node);
150 	rf_DiskReadFunc(node);
151 }
152 #endif
153 
154 void
rf_DiskReadMirrorUndoFunc(RF_DagNode_t * node)155 rf_DiskReadMirrorUndoFunc(RF_DagNode_t *node)
156 {
157 }
158 
159 
160 
161 #if RF_INCLUDE_PARITYLOGGING > 0
162 /*****************************************************************************
163  * the execution function associated with a parity log update node
164  ****************************************************************************/
165 void
rf_ParityLogUpdateFunc(RF_DagNode_t * node)166 rf_ParityLogUpdateFunc(RF_DagNode_t *node)
167 {
168 	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
169 	void *bf = (void *) node->params[1].p;
170 	RF_ParityLogData_t *logData;
171 #if RF_ACC_TRACE > 0
172 	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
173 	RF_Etimer_t timer;
174 #endif
175 
176 	if (node->dagHdr->status == rf_enable) {
177 #if RF_ACC_TRACE > 0
178 		RF_ETIMER_START(timer);
179 #endif
180 		logData = rf_CreateParityLogData(RF_UPDATE, pda, bf,
181 		    (RF_Raid_t *) (node->dagHdr->raidPtr),
182 		    node->wakeFunc, node,
183 		    node->dagHdr->tracerec, timer);
184 		if (logData)
185 			rf_ParityLogAppend(logData, RF_FALSE, NULL, RF_FALSE);
186 		else {
187 #if RF_ACC_TRACE > 0
188 			RF_ETIMER_STOP(timer);
189 			RF_ETIMER_EVAL(timer);
190 			tracerec->plog_us += RF_ETIMER_VAL_US(timer);
191 #endif
192 			(node->wakeFunc) (node, ENOMEM);
193 		}
194 	}
195 }
196 
197 
198 /*****************************************************************************
199  * the execution function associated with a parity log overwrite node
200  ****************************************************************************/
201 void
rf_ParityLogOverwriteFunc(RF_DagNode_t * node)202 rf_ParityLogOverwriteFunc(RF_DagNode_t *node)
203 {
204 	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
205 	void *bf = (void *) node->params[1].p;
206 	RF_ParityLogData_t *logData;
207 #if RF_ACC_TRACE > 0
208 	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
209 	RF_Etimer_t timer;
210 #endif
211 
212 	if (node->dagHdr->status == rf_enable) {
213 #if RF_ACC_TRACE > 0
214 		RF_ETIMER_START(timer);
215 #endif
216 		logData = rf_CreateParityLogData(RF_OVERWRITE, pda, bf,
217 (RF_Raid_t *) (node->dagHdr->raidPtr),
218 		    node->wakeFunc, node, node->dagHdr->tracerec, timer);
219 		if (logData)
220 			rf_ParityLogAppend(logData, RF_FALSE, NULL, RF_FALSE);
221 		else {
222 #if RF_ACC_TRACE > 0
223 			RF_ETIMER_STOP(timer);
224 			RF_ETIMER_EVAL(timer);
225 			tracerec->plog_us += RF_ETIMER_VAL_US(timer);
226 #endif
227 			(node->wakeFunc) (node, ENOMEM);
228 		}
229 	}
230 }
231 
232 void
rf_ParityLogUpdateUndoFunc(RF_DagNode_t * node)233 rf_ParityLogUpdateUndoFunc(RF_DagNode_t *node)
234 {
235 }
236 
237 void
rf_ParityLogOverwriteUndoFunc(RF_DagNode_t * node)238 rf_ParityLogOverwriteUndoFunc(RF_DagNode_t *node)
239 {
240 }
241 #endif				/* RF_INCLUDE_PARITYLOGGING > 0 */
242 
243 /*****************************************************************************
244  * the execution function associated with a NOP node
245  ****************************************************************************/
246 void
rf_NullNodeFunc(RF_DagNode_t * node)247 rf_NullNodeFunc(RF_DagNode_t *node)
248 {
249 	node->status = rf_good;
250 	rf_FinishNode(node, RF_THREAD_CONTEXT);
251 }
252 
253 void
rf_NullNodeUndoFunc(RF_DagNode_t * node)254 rf_NullNodeUndoFunc(RF_DagNode_t *node)
255 {
256 	node->status = rf_undone;
257 	rf_FinishNode(node, RF_THREAD_CONTEXT);
258 }
259 
260 
261 /*****************************************************************************
262  * the execution function associated with a disk-read node
263  ****************************************************************************/
264 void
rf_DiskReadFuncForThreads(RF_DagNode_t * node)265 rf_DiskReadFuncForThreads(RF_DagNode_t *node)
266 {
267 	RF_DiskQueueData_t *req;
268 	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
269 	void *bf = (void *) node->params[1].p;
270 	RF_StripeNum_t parityStripeID = (RF_StripeNum_t) node->params[2].v;
271 	unsigned priority = RF_EXTRACT_PRIORITY(node->params[3].v);
272 	unsigned which_ru = RF_EXTRACT_RU(node->params[3].v);
273 	RF_IoType_t iotype = (node->dagHdr->status == rf_enable) ? RF_IO_TYPE_READ : RF_IO_TYPE_NOP;
274 	RF_DiskQueue_t *dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues;
275 
276 	req = rf_CreateDiskQueueData(iotype, pda->startSector, pda->numSector,
277 	    bf, parityStripeID, which_ru, node->wakeFunc, node,
278 #if RF_ACC_TRACE > 0
279 	     node->dagHdr->tracerec,
280 #else
281              NULL,
282 #endif
283 	    (void *) (node->dagHdr->raidPtr), 0, node->dagHdr->bp);
284 
285 	node->dagFuncData = (void *) req;
286 	rf_DiskIOEnqueue(&(dqs[pda->col]), req, priority);
287 }
288 
289 
290 /*****************************************************************************
291  * the execution function associated with a disk-write node
292  ****************************************************************************/
293 void
rf_DiskWriteFuncForThreads(RF_DagNode_t * node)294 rf_DiskWriteFuncForThreads(RF_DagNode_t *node)
295 {
296 	RF_DiskQueueData_t *req;
297 	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
298 	void *bf = (void *) node->params[1].p;
299 	RF_StripeNum_t parityStripeID = (RF_StripeNum_t) node->params[2].v;
300 	unsigned priority = RF_EXTRACT_PRIORITY(node->params[3].v);
301 	unsigned which_ru = RF_EXTRACT_RU(node->params[3].v);
302 	RF_IoType_t iotype = (node->dagHdr->status == rf_enable) ? RF_IO_TYPE_WRITE : RF_IO_TYPE_NOP;
303 	RF_DiskQueue_t *dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues;
304 
305 	/* normal processing (rollaway or forward recovery) begins here */
306 	req = rf_CreateDiskQueueData(iotype, pda->startSector, pda->numSector,
307 	    bf, parityStripeID, which_ru, node->wakeFunc, node,
308 #if RF_ACC_TRACE > 0
309 	    node->dagHdr->tracerec,
310 #else
311 	    NULL,
312 #endif
313 	    (void *) (node->dagHdr->raidPtr),
314 	    0, node->dagHdr->bp);
315 
316 	node->dagFuncData = (void *) req;
317 	rf_DiskIOEnqueue(&(dqs[pda->col]), req, priority);
318 }
319 /*****************************************************************************
320  * the undo function for disk nodes
321  * Note:  this is not a proper undo of a write node, only locks are released.
322  *        old data is not restored to disk!
323  ****************************************************************************/
324 void
rf_DiskUndoFunc(RF_DagNode_t * node)325 rf_DiskUndoFunc(RF_DagNode_t *node)
326 {
327 	RF_DiskQueueData_t *req;
328 	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
329 	RF_DiskQueue_t *dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues;
330 
331 	req = rf_CreateDiskQueueData(RF_IO_TYPE_NOP,
332 	    0L, 0, NULL, 0L, 0, node->wakeFunc, node,
333 #if RF_ACC_TRACE > 0
334 	     node->dagHdr->tracerec,
335 #else
336 	     NULL,
337 #endif
338 	    (void *) (node->dagHdr->raidPtr),
339 	    0, NULL);
340 
341 	node->dagFuncData = (void *) req;
342 	rf_DiskIOEnqueue(&(dqs[pda->col]), req, RF_IO_NORMAL_PRIORITY);
343 }
344 
345 /*****************************************************************************
346  * Callback routine for DiskRead and DiskWrite nodes.  When the disk
347  * op completes, the routine is called to set the node status and
348  * inform the execution engine that the node has fired.
349  ****************************************************************************/
350 void
rf_GenericWakeupFunc(void * v,int status)351 rf_GenericWakeupFunc(void *v, int status)
352 {
353 	RF_DagNode_t *node = v;
354 
355 	switch (node->status) {
356 	case rf_fired:
357 		if (status)
358 			node->status = rf_bad;
359 		else
360 			node->status = rf_good;
361 		break;
362 	case rf_recover:
363 		/* probably should never reach this case */
364 		if (status)
365 			node->status = rf_panic;
366 		else
367 			node->status = rf_undone;
368 		break;
369 	default:
370 		printf("rf_GenericWakeupFunc:");
371 		printf("node->status is %d,", node->status);
372 		printf("status is %d \n", status);
373 		RF_PANIC();
374 		break;
375 	}
376 	if (node->dagFuncData)
377 		rf_FreeDiskQueueData((RF_DiskQueueData_t *) node->dagFuncData);
378 	rf_FinishNode(node, RF_INTR_CONTEXT);
379 }
380 
381 
382 /*****************************************************************************
383  * there are three distinct types of xor nodes:
384 
385  * A "regular xor" is used in the fault-free case where the access
386  * spans a complete stripe unit.  It assumes that the result buffer is
387  * one full stripe unit in size, and uses the stripe-unit-offset
388  * values that it computes from the PDAs to determine where within the
389  * stripe unit to XOR each argument buffer.
390  *
391  * A "simple xor" is used in the fault-free case where the access
392  * touches only a portion of one (or two, in some cases) stripe
393  * unit(s).  It assumes that all the argument buffers are of the same
394  * size and have the same stripe unit offset.
395  *
396  * A "recovery xor" is used in the degraded-mode case.  It's similar
397  * to the regular xor function except that it takes the failed PDA as
398  * an additional parameter, and uses it to determine what portions of
399  * the argument buffers need to be xor'd into the result buffer, and
400  * where in the result buffer they should go.
401  ****************************************************************************/
402 
403 /* xor the params together and store the result in the result field.
404  * assume the result field points to a buffer that is the size of one
405  * SU, and use the pda params to determine where within the buffer to
406  * XOR the input buffers.  */
407 void
rf_RegularXorFunc(RF_DagNode_t * node)408 rf_RegularXorFunc(RF_DagNode_t *node)
409 {
410 	RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p;
411 #if RF_ACC_TRACE > 0
412 	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
413 	RF_Etimer_t timer;
414 #endif
415 	int     i, retcode;
416 
417 	retcode = 0;
418 	if (node->dagHdr->status == rf_enable) {
419 		/* don't do the XOR if the input is the same as the output */
420 #if RF_ACC_TRACE > 0
421 		RF_ETIMER_START(timer);
422 #endif
423 		for (i = 0; i < node->numParams - 1; i += 2)
424 			if (node->params[i + 1].p != node->results[0]) {
425 				retcode = rf_XorIntoBuffer(raidPtr, (RF_PhysDiskAddr_t *) node->params[i].p,
426 							   (char *) node->params[i + 1].p, (char *) node->results[0]);
427 			}
428 #if RF_ACC_TRACE > 0
429 		RF_ETIMER_STOP(timer);
430 		RF_ETIMER_EVAL(timer);
431 		tracerec->xor_us += RF_ETIMER_VAL_US(timer);
432 #endif
433 	}
434 	rf_GenericWakeupFunc(node, retcode);	/* call wake func
435 						 * explicitly since no
436 						 * I/O in this node */
437 }
438 /* xor the inputs into the result buffer, ignoring placement issues */
439 void
rf_SimpleXorFunc(RF_DagNode_t * node)440 rf_SimpleXorFunc(RF_DagNode_t *node)
441 {
442 	RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p;
443 	int     i, retcode = 0;
444 #if RF_ACC_TRACE > 0
445 	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
446 	RF_Etimer_t timer;
447 #endif
448 
449 	if (node->dagHdr->status == rf_enable) {
450 #if RF_ACC_TRACE > 0
451 		RF_ETIMER_START(timer);
452 #endif
453 		/* don't do the XOR if the input is the same as the output */
454 		for (i = 0; i < node->numParams - 1; i += 2)
455 			if (node->params[i + 1].p != node->results[0]) {
456 				retcode = rf_bxor((char *) node->params[i + 1].p, (char *) node->results[0],
457 				    rf_RaidAddressToByte(raidPtr, ((RF_PhysDiskAddr_t *) node->params[i].p)->numSector));
458 			}
459 #if RF_ACC_TRACE > 0
460 		RF_ETIMER_STOP(timer);
461 		RF_ETIMER_EVAL(timer);
462 		tracerec->xor_us += RF_ETIMER_VAL_US(timer);
463 #endif
464 	}
465 	rf_GenericWakeupFunc(node, retcode);	/* call wake func
466 						 * explicitly since no
467 						 * I/O in this node */
468 }
469 /* this xor is used by the degraded-mode dag functions to recover lost
470  * data.  the second-to-last parameter is the PDA for the failed
471  * portion of the access.  the code here looks at this PDA and assumes
472  * that the xor target buffer is equal in size to the number of
473  * sectors in the failed PDA.  It then uses the other PDAs in the
474  * parameter list to determine where within the target buffer the
475  * corresponding data should be xored.  */
476 void
rf_RecoveryXorFunc(RF_DagNode_t * node)477 rf_RecoveryXorFunc(RF_DagNode_t *node)
478 {
479 	RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p;
480 	RF_RaidLayout_t *layoutPtr = (RF_RaidLayout_t *) & raidPtr->Layout;
481 	RF_PhysDiskAddr_t *failedPDA = (RF_PhysDiskAddr_t *) node->params[node->numParams - 2].p;
482 	int     i, retcode = 0;
483 	RF_PhysDiskAddr_t *pda;
484 	int     suoffset, failedSUOffset = rf_StripeUnitOffset(layoutPtr, failedPDA->startSector);
485 	char   *srcbuf, *destbuf;
486 #if RF_ACC_TRACE > 0
487 	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
488 	RF_Etimer_t timer;
489 #endif
490 
491 	if (node->dagHdr->status == rf_enable) {
492 #if RF_ACC_TRACE > 0
493 		RF_ETIMER_START(timer);
494 #endif
495 		for (i = 0; i < node->numParams - 2; i += 2)
496 			if (node->params[i + 1].p != node->results[0]) {
497 				pda = (RF_PhysDiskAddr_t *) node->params[i].p;
498 				srcbuf = (char *) node->params[i + 1].p;
499 				suoffset = rf_StripeUnitOffset(layoutPtr, pda->startSector);
500 				destbuf = ((char *) node->results[0]) + rf_RaidAddressToByte(raidPtr, suoffset - failedSUOffset);
501 				retcode = rf_bxor(srcbuf, destbuf, rf_RaidAddressToByte(raidPtr, pda->numSector));
502 			}
503 #if RF_ACC_TRACE > 0
504 		RF_ETIMER_STOP(timer);
505 		RF_ETIMER_EVAL(timer);
506 		tracerec->xor_us += RF_ETIMER_VAL_US(timer);
507 #endif
508 	}
509 	rf_GenericWakeupFunc(node, retcode);
510 }
511 /*****************************************************************************
512  * The next three functions are utilities used by the above
513  * xor-execution functions.
514  ****************************************************************************/
515 
516 
517 /*
518  * this is just a glorified buffer xor.  targbuf points to a buffer
519  * that is one full stripe unit in size.  srcbuf points to a buffer
520  * that may be less than 1 SU, but never more.  When the access
521  * described by pda is one SU in size (which by implication means it's
522  * SU-aligned), all that happens is (targbuf) <- (srcbuf ^ targbuf).
523  * When the access is less than one SU in size the XOR occurs on only
524  * the portion of targbuf identified in the pda.  */
525 
526 int
rf_XorIntoBuffer(RF_Raid_t * raidPtr,RF_PhysDiskAddr_t * pda,char * srcbuf,char * targbuf)527 rf_XorIntoBuffer(RF_Raid_t *raidPtr, RF_PhysDiskAddr_t *pda,
528 		 char *srcbuf, char *targbuf)
529 {
530 	char   *targptr;
531 	int     sectPerSU = raidPtr->Layout.sectorsPerStripeUnit;
532 	int     SUOffset = pda->startSector % sectPerSU;
533 	int     length, retcode = 0;
534 
535 	RF_ASSERT(pda->numSector <= sectPerSU);
536 
537 	targptr = targbuf + rf_RaidAddressToByte(raidPtr, SUOffset);
538 	length = rf_RaidAddressToByte(raidPtr, pda->numSector);
539 	retcode = rf_bxor(srcbuf, targptr, length);
540 	return (retcode);
541 }
542 /* it really should be the case that the buffer pointers (returned by
543  * malloc) are aligned to the natural word size of the machine, so
544  * this is the only case we optimize for.  The length should always be
545  * a multiple of the sector size, so there should be no problem with
546  * leftover bytes at the end.  */
547 int
rf_bxor(char * src,char * dest,int len)548 rf_bxor(char *src, char *dest, int len)
549 {
550 	unsigned mask = sizeof(long) - 1, retcode = 0;
551 
552 	if (!(((unsigned long) src) & mask) &&
553 	    !(((unsigned long) dest) & mask) && !(len & mask)) {
554 		retcode = rf_longword_bxor((unsigned long *) src,
555 					   (unsigned long *) dest,
556 					   len >> RF_LONGSHIFT);
557 	} else {
558 		RF_ASSERT(0);
559 	}
560 	return (retcode);
561 }
562 
563 /* When XORing in kernel mode, we need to map each user page to kernel
564  * space before we can access it.  We don't want to assume anything
565  * about which input buffers are in kernel/user space, nor about their
566  * alignment, so in each loop we compute the maximum number of bytes
567  * that we can xor without crossing any page boundaries, and do only
568  * this many bytes before the next remap.
569  *
570  * len - is in longwords
571  */
572 int
rf_longword_bxor(unsigned long * src,unsigned long * dest,int len)573 rf_longword_bxor(unsigned long *src, unsigned long *dest, int len)
574 {
575 	unsigned long *end = src + len;
576 	unsigned long d0, d1, d2, d3, s0, s1, s2, s3;	/* temps */
577 	unsigned long *pg_src, *pg_dest;   /* per-page source/dest pointers */
578 	int     longs_this_time;/* # longwords to xor in the current iteration */
579 
580 	pg_src = src;
581 	pg_dest = dest;
582 	if (!pg_src || !pg_dest)
583 		return (EFAULT);
584 
585 	while (len >= 4) {
586 		longs_this_time = RF_MIN(len, RF_MIN(RF_BLIP(pg_src), RF_BLIP(pg_dest)) >> RF_LONGSHIFT);	/* note len in longwords */
587 		src += longs_this_time;
588 		dest += longs_this_time;
589 		len -= longs_this_time;
590 		while (longs_this_time >= 4) {
591 			d0 = pg_dest[0];
592 			d1 = pg_dest[1];
593 			d2 = pg_dest[2];
594 			d3 = pg_dest[3];
595 			s0 = pg_src[0];
596 			s1 = pg_src[1];
597 			s2 = pg_src[2];
598 			s3 = pg_src[3];
599 			pg_dest[0] = d0 ^ s0;
600 			pg_dest[1] = d1 ^ s1;
601 			pg_dest[2] = d2 ^ s2;
602 			pg_dest[3] = d3 ^ s3;
603 			pg_src += 4;
604 			pg_dest += 4;
605 			longs_this_time -= 4;
606 		}
607 		while (longs_this_time > 0) {	/* cannot cross any page
608 						 * boundaries here */
609 			*pg_dest++ ^= *pg_src++;
610 			longs_this_time--;
611 		}
612 
613 		/* either we're done, or we've reached a page boundary on one
614 		 * (or possibly both) of the pointers */
615 		if (len) {
616 			if (RF_PAGE_ALIGNED(src))
617 				pg_src = src;
618 			if (RF_PAGE_ALIGNED(dest))
619 				pg_dest = dest;
620 			if (!pg_src || !pg_dest)
621 				return (EFAULT);
622 		}
623 	}
624 	while (src < end) {
625 		*pg_dest++ ^= *pg_src++;
626 		src++;
627 		dest++;
628 		len--;
629 		if (RF_PAGE_ALIGNED(src))
630 			pg_src = src;
631 		if (RF_PAGE_ALIGNED(dest))
632 			pg_dest = dest;
633 	}
634 	RF_ASSERT(len == 0);
635 	return (0);
636 }
637 
638 #if 0
639 /*
640    dst = a ^ b ^ c;
641    a may equal dst
642    see comment above longword_bxor
643    len is length in longwords
644 */
645 int
646 rf_longword_bxor3(unsigned long *dst, unsigned long *a, unsigned long *b,
647 		  unsigned long *c, int len, void *bp)
648 {
649 	unsigned long a0, a1, a2, a3, b0, b1, b2, b3;
650 	unsigned long *pg_a, *pg_b, *pg_c, *pg_dst;	/* per-page source/dest
651 								 * pointers */
652 	int     longs_this_time;/* # longs to xor in the current iteration */
653 	char    dst_is_a = 0;
654 
655 	pg_a = a;
656 	pg_b = b;
657 	pg_c = c;
658 	if (a == dst) {
659 		pg_dst = pg_a;
660 		dst_is_a = 1;
661 	} else {
662 		pg_dst = dst;
663 	}
664 
665 	/* align dest to cache line.  Can't cross a pg boundary on dst here. */
666 	while ((((unsigned long) pg_dst) & 0x1f)) {
667 		*pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++;
668 		dst++;
669 		a++;
670 		b++;
671 		c++;
672 		if (RF_PAGE_ALIGNED(a)) {
673 			pg_a = a;
674 			if (!pg_a)
675 				return (EFAULT);
676 		}
677 		if (RF_PAGE_ALIGNED(b)) {
678 			pg_b = a;
679 			if (!pg_b)
680 				return (EFAULT);
681 		}
682 		if (RF_PAGE_ALIGNED(c)) {
683 			pg_c = a;
684 			if (!pg_c)
685 				return (EFAULT);
686 		}
687 		len--;
688 	}
689 
690 	while (len > 4) {
691 		longs_this_time = RF_MIN(len, RF_MIN(RF_BLIP(a), RF_MIN(RF_BLIP(b), RF_MIN(RF_BLIP(c), RF_BLIP(dst)))) >> RF_LONGSHIFT);
692 		a += longs_this_time;
693 		b += longs_this_time;
694 		c += longs_this_time;
695 		dst += longs_this_time;
696 		len -= longs_this_time;
697 		while (longs_this_time >= 4) {
698 			a0 = pg_a[0];
699 			longs_this_time -= 4;
700 
701 			a1 = pg_a[1];
702 			a2 = pg_a[2];
703 
704 			a3 = pg_a[3];
705 			pg_a += 4;
706 
707 			b0 = pg_b[0];
708 			b1 = pg_b[1];
709 
710 			b2 = pg_b[2];
711 			b3 = pg_b[3];
712 			/* start dual issue */
713 			a0 ^= b0;
714 			b0 = pg_c[0];
715 
716 			pg_b += 4;
717 			a1 ^= b1;
718 
719 			a2 ^= b2;
720 			a3 ^= b3;
721 
722 			b1 = pg_c[1];
723 			a0 ^= b0;
724 
725 			b2 = pg_c[2];
726 			a1 ^= b1;
727 
728 			b3 = pg_c[3];
729 			a2 ^= b2;
730 
731 			pg_dst[0] = a0;
732 			a3 ^= b3;
733 			pg_dst[1] = a1;
734 			pg_c += 4;
735 			pg_dst[2] = a2;
736 			pg_dst[3] = a3;
737 			pg_dst += 4;
738 		}
739 		while (longs_this_time > 0) {	/* cannot cross any page
740 						 * boundaries here */
741 			*pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++;
742 			longs_this_time--;
743 		}
744 
745 		if (len) {
746 			if (RF_PAGE_ALIGNED(a)) {
747 				pg_a = a;
748 				if (!pg_a)
749 					return (EFAULT);
750 				if (dst_is_a)
751 					pg_dst = pg_a;
752 			}
753 			if (RF_PAGE_ALIGNED(b)) {
754 				pg_b = b;
755 				if (!pg_b)
756 					return (EFAULT);
757 			}
758 			if (RF_PAGE_ALIGNED(c)) {
759 				pg_c = c;
760 				if (!pg_c)
761 					return (EFAULT);
762 			}
763 			if (!dst_is_a)
764 				if (RF_PAGE_ALIGNED(dst)) {
765 					pg_dst = dst;
766 					if (!pg_dst)
767 						return (EFAULT);
768 				}
769 		}
770 	}
771 	while (len) {
772 		*pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++;
773 		dst++;
774 		a++;
775 		b++;
776 		c++;
777 		if (RF_PAGE_ALIGNED(a)) {
778 			pg_a = a;
779 			if (!pg_a)
780 				return (EFAULT);
781 			if (dst_is_a)
782 				pg_dst = pg_a;
783 		}
784 		if (RF_PAGE_ALIGNED(b)) {
785 			pg_b = b;
786 			if (!pg_b)
787 				return (EFAULT);
788 		}
789 		if (RF_PAGE_ALIGNED(c)) {
790 			pg_c = c;
791 			if (!pg_c)
792 				return (EFAULT);
793 		}
794 		if (!dst_is_a)
795 			if (RF_PAGE_ALIGNED(dst)) {
796 				pg_dst = dst;
797 				if (!pg_dst)
798 					return (EFAULT);
799 			}
800 		len--;
801 	}
802 	return (0);
803 }
804 
805 int
806 rf_bxor3(unsigned char *dst, unsigned char *a, unsigned char *b,
807 	 unsigned char *c, unsigned long len, void *bp)
808 {
809 	RF_ASSERT(((RF_UL(dst) | RF_UL(a) | RF_UL(b) | RF_UL(c) | len) & 0x7) == 0);
810 
811 	return (rf_longword_bxor3((unsigned long *) dst, (unsigned long *) a,
812 		(unsigned long *) b, (unsigned long *) c, len >> RF_LONGSHIFT, bp));
813 }
814 #endif
815