xref: /src/sys/dev/raidframe/rf_dagfuncs.c

/*	$NetBSD: rf_dagfuncs.c,v 1.35 2021/08/07 16:19:15 thorpej Exp $	*/
/*
 * Copyright (c) 1995 Carnegie-Mellon University.
 * All rights reserved.
 *
 * Author: Mark Holland, William V. Courtright II
 *
 * Permission to use, copy, modify and distribute this software and
 * its documentation is hereby granted, provided that both the copyright
 * notice and this permission notice appear in all copies of the
 * software, derivative works or modified versions, and any portions
 * thereof, and that both notices appear in supporting documentation.
 *
 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
 * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
 *
 * Carnegie Mellon requests users of this software to return to
 *
 *  Software Distribution Coordinator  or  Software.Distribution (at) CS.CMU.EDU
 *  School of Computer Science
 *  Carnegie Mellon University
 *  Pittsburgh PA 15213-3890
 *
 * any improvements or extensions that they make and grant Carnegie the
 * rights to redistribute these changes.
 */

/*
 * dagfuncs.c -- DAG node execution routines
 *
 * Rules:
 * 1. Every DAG execution function must eventually cause node->status to
 *    get set to "good" or "bad", and "FinishNode" to be called. In the
 *    case of nodes that complete immediately (xor, NullNodeFunc, etc),
 *    the node execution function can do these two things directly. In
 *    the case of nodes that have to wait for some event (a disk read to
 *    complete, a lock to be released, etc) to occur before they can
 *    complete, this is typically achieved by having whatever module
 *    is doing the operation call GenericWakeupFunc upon completion.
 * 2. DAG execution functions should check the status in the DAG header
 *    and NOP out their operations if the status is not "enable". However,
 *    execution functions that release resources must be sure to release
 *    them even when they NOP out the function that would use them.
 *    Functions that acquire resources should go ahead and acquire them
 *    even when they NOP, so that a downstream release node will not have
 *    to check to find out whether or not the acquire was suppressed.
 */

#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: rf_dagfuncs.c,v 1.35 2021/08/07 16:19:15 thorpej Exp $");

#include <sys/param.h>
#include <sys/ioctl.h>

#include "rf_archs.h"
#include "rf_raid.h"
#include "rf_dag.h"
#include "rf_layout.h"
#include "rf_etimer.h"
#include "rf_acctrace.h"
#include "rf_diskqueue.h"
#include "rf_dagfuncs.h"
#include "rf_general.h"
#include "rf_engine.h"
#include "rf_dagutils.h"

#include "rf_kintf.h"

#if RF_INCLUDE_PARITYLOGGING > 0
#include "rf_paritylog.h"
#endif				/* RF_INCLUDE_PARITYLOGGING > 0 */

void     (*rf_DiskReadFunc) (RF_DagNode_t *);
void     (*rf_DiskWriteFunc) (RF_DagNode_t *);
void     (*rf_DiskReadUndoFunc) (RF_DagNode_t *);
void     (*rf_DiskWriteUndoFunc) (RF_DagNode_t *);
void     (*rf_RegularXorUndoFunc) (RF_DagNode_t *);
void     (*rf_SimpleXorUndoFunc) (RF_DagNode_t *);
void     (*rf_RecoveryXorUndoFunc) (RF_DagNode_t *);

/*****************************************************************************
 * main (only) configuration routine for this module
 ****************************************************************************/
int
rf_ConfigureDAGFuncs(RF_ShutdownList_t **listp)
{
	RF_ASSERT(((sizeof(long) == 8) && RF_LONGSHIFT == 3) ||
		  ((sizeof(long) == 4) && RF_LONGSHIFT == 2));
	rf_DiskReadFunc = rf_DiskReadFuncForThreads;
	rf_DiskReadUndoFunc = rf_DiskUndoFunc;
	rf_DiskWriteFunc = rf_DiskWriteFuncForThreads;
	rf_DiskWriteUndoFunc = rf_DiskUndoFunc;
	rf_RegularXorUndoFunc = rf_NullNodeUndoFunc;
	rf_SimpleXorUndoFunc = rf_NullNodeUndoFunc;
	rf_RecoveryXorUndoFunc = rf_NullNodeUndoFunc;
	return (0);
}



/*****************************************************************************
 * the execution function associated with a terminate node
 ****************************************************************************/
void
rf_TerminateFunc(RF_DagNode_t *node)
{
	RF_ASSERT(node->dagHdr->numCommits == node->dagHdr->numCommitNodes);
	node->status = rf_good;
	rf_FinishNode(node, RF_THREAD_CONTEXT);
}

void
rf_TerminateUndoFunc(RF_DagNode_t *node)
{
}


/*****************************************************************************
 * execution functions associated with a mirror node
 *
 * parameters:
 *
 * 0 - physical disk address of data
 * 1 - buffer for holding read data
 * 2 - parity stripe ID
 * 3 - flags
 * 4 - physical disk address of mirror (parity)
 *
 ****************************************************************************/

void
rf_DiskReadMirrorIdleFunc(RF_DagNode_t *node)
{
	/* select the mirror copy with the shortest queue and fill in node
	 * parameters with physical disk address */

	rf_SelectMirrorDiskIdle(node);
	rf_DiskReadFunc(node);
}

#if (RF_INCLUDE_CHAINDECLUSTER > 0) || (RF_INCLUDE_INTERDECLUSTER > 0) || (RF_DEBUG_VALIDATE_DAG > 0)
void
rf_DiskReadMirrorPartitionFunc(RF_DagNode_t *node)
{
	/* select the mirror copy with the shortest queue and fill in node
	 * parameters with physical disk address */

	rf_SelectMirrorDiskPartition(node);
	rf_DiskReadFunc(node);
}
#endif

void
rf_DiskReadMirrorUndoFunc(RF_DagNode_t *node)
{
}



#if RF_INCLUDE_PARITYLOGGING > 0
/*****************************************************************************
 * the execution function associated with a parity log update node
 ****************************************************************************/
void
rf_ParityLogUpdateFunc(RF_DagNode_t *node)
{
	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
	void *bf = (void *) node->params[1].p;
	RF_ParityLogData_t *logData;
#if RF_ACC_TRACE > 0
	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
	RF_Etimer_t timer;
#endif

	if (node->dagHdr->status == rf_enable) {
#if RF_ACC_TRACE > 0
		RF_ETIMER_START(timer);
#endif
		logData = rf_CreateParityLogData(RF_UPDATE, pda, bf,
		    (RF_Raid_t *) (node->dagHdr->raidPtr),
		    node->wakeFunc, node,
		    node->dagHdr->tracerec, timer);
		if (logData)
			rf_ParityLogAppend(logData, RF_FALSE, NULL, RF_FALSE);
		else {
#if RF_ACC_TRACE > 0
			RF_ETIMER_STOP(timer);
			RF_ETIMER_EVAL(timer);
			tracerec->plog_us += RF_ETIMER_VAL_US(timer);
#endif
			(node->wakeFunc) (node, ENOMEM);
		}
	}
}


/*****************************************************************************
 * the execution function associated with a parity log overwrite node
 ****************************************************************************/
void
rf_ParityLogOverwriteFunc(RF_DagNode_t *node)
{
	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
	void *bf = (void *) node->params[1].p;
	RF_ParityLogData_t *logData;
#if RF_ACC_TRACE > 0
	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
	RF_Etimer_t timer;
#endif

	if (node->dagHdr->status == rf_enable) {
#if RF_ACC_TRACE > 0
		RF_ETIMER_START(timer);
#endif
		logData = rf_CreateParityLogData(RF_OVERWRITE, pda, bf,
(RF_Raid_t *) (node->dagHdr->raidPtr),
		    node->wakeFunc, node, node->dagHdr->tracerec, timer);
		if (logData)
			rf_ParityLogAppend(logData, RF_FALSE, NULL, RF_FALSE);
		else {
#if RF_ACC_TRACE > 0
			RF_ETIMER_STOP(timer);
			RF_ETIMER_EVAL(timer);
			tracerec->plog_us += RF_ETIMER_VAL_US(timer);
#endif
			(node->wakeFunc) (node, ENOMEM);
		}
	}
}

void
rf_ParityLogUpdateUndoFunc(RF_DagNode_t *node)
{
}

void
rf_ParityLogOverwriteUndoFunc(RF_DagNode_t *node)
{
}
#endif				/* RF_INCLUDE_PARITYLOGGING > 0 */

/*****************************************************************************
 * the execution function associated with a NOP node
 ****************************************************************************/
void
rf_NullNodeFunc(RF_DagNode_t *node)
{
	node->status = rf_good;
	rf_FinishNode(node, RF_THREAD_CONTEXT);
}

void
rf_NullNodeUndoFunc(RF_DagNode_t *node)
{
	node->status = rf_undone;
	rf_FinishNode(node, RF_THREAD_CONTEXT);
}


/*****************************************************************************
 * the execution function associated with a disk-read node
 ****************************************************************************/
void
rf_DiskReadFuncForThreads(RF_DagNode_t *node)
{
	RF_DiskQueueData_t *req;
	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
	void *bf = (void *) node->params[1].p;
	RF_StripeNum_t parityStripeID = (RF_StripeNum_t) node->params[2].v;
	unsigned priority = RF_EXTRACT_PRIORITY(node->params[3].v);
	unsigned which_ru = RF_EXTRACT_RU(node->params[3].v);
	RF_IoType_t iotype = (node->dagHdr->status == rf_enable) ? RF_IO_TYPE_READ : RF_IO_TYPE_NOP;
	RF_DiskQueue_t *dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues;

	req = rf_CreateDiskQueueData(iotype, pda->startSector, pda->numSector,
	    bf, parityStripeID, which_ru, node->wakeFunc, node,
#if RF_ACC_TRACE > 0
	     node->dagHdr->tracerec,
#else
             NULL,
#endif
	    (void *) (node->dagHdr->raidPtr), 0, node->dagHdr->bp);

	node->dagFuncData = (void *) req;
	rf_DiskIOEnqueue(&(dqs[pda->col]), req, priority);
}


/*****************************************************************************
 * the execution function associated with a disk-write node
 ****************************************************************************/
void
rf_DiskWriteFuncForThreads(RF_DagNode_t *node)
{
	RF_DiskQueueData_t *req;
	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
	void *bf = (void *) node->params[1].p;
	RF_StripeNum_t parityStripeID = (RF_StripeNum_t) node->params[2].v;
	unsigned priority = RF_EXTRACT_PRIORITY(node->params[3].v);
	unsigned which_ru = RF_EXTRACT_RU(node->params[3].v);
	RF_IoType_t iotype = (node->dagHdr->status == rf_enable) ? RF_IO_TYPE_WRITE : RF_IO_TYPE_NOP;
	RF_DiskQueue_t *dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues;

	/* normal processing (rollaway or forward recovery) begins here */
	req = rf_CreateDiskQueueData(iotype, pda->startSector, pda->numSector,
	    bf, parityStripeID, which_ru, node->wakeFunc, node,
#if RF_ACC_TRACE > 0
	    node->dagHdr->tracerec,
#else
	    NULL,
#endif
	    (void *) (node->dagHdr->raidPtr),
	    0, node->dagHdr->bp);

	node->dagFuncData = (void *) req;
	rf_DiskIOEnqueue(&(dqs[pda->col]), req, priority);
}
/*****************************************************************************
 * the undo function for disk nodes
 * Note:  this is not a proper undo of a write node, only locks are released.
 *        old data is not restored to disk!
 ****************************************************************************/
void
rf_DiskUndoFunc(RF_DagNode_t *node)
{
	RF_DiskQueueData_t *req;
	RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p;
	RF_DiskQueue_t *dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues;

	req = rf_CreateDiskQueueData(RF_IO_TYPE_NOP,
	    0L, 0, NULL, 0L, 0, node->wakeFunc, node,
#if RF_ACC_TRACE > 0
	     node->dagHdr->tracerec,
#else
	     NULL,
#endif
	    (void *) (node->dagHdr->raidPtr),
	    0, NULL);

	node->dagFuncData = (void *) req;
	rf_DiskIOEnqueue(&(dqs[pda->col]), req, RF_IO_NORMAL_PRIORITY);
}

/*****************************************************************************
 * Callback routine for DiskRead and DiskWrite nodes.  When the disk
 * op completes, the routine is called to set the node status and
 * inform the execution engine that the node has fired.
 ****************************************************************************/
void
rf_GenericWakeupFunc(void *v, int status)
{
	RF_DagNode_t *node = v;

	switch (node->status) {
	case rf_fired:
		if (status)
			node->status = rf_bad;
		else
			node->status = rf_good;
		break;
	case rf_recover:
		/* probably should never reach this case */
		if (status)
			node->status = rf_panic;
		else
			node->status = rf_undone;
		break;
	default:
		printf("rf_GenericWakeupFunc:");
		printf("node->status is %d,", node->status);
		printf("status is %d \n", status);
		RF_PANIC();
		break;
	}
	if (node->dagFuncData)
		rf_FreeDiskQueueData((RF_DiskQueueData_t *) node->dagFuncData);
	rf_FinishNode(node, RF_INTR_CONTEXT);
}


/*****************************************************************************
 * there are three distinct types of xor nodes:

 * A "regular xor" is used in the fault-free case where the access
 * spans a complete stripe unit.  It assumes that the result buffer is
 * one full stripe unit in size, and uses the stripe-unit-offset
 * values that it computes from the PDAs to determine where within the
 * stripe unit to XOR each argument buffer.
 *
 * A "simple xor" is used in the fault-free case where the access
 * touches only a portion of one (or two, in some cases) stripe
 * unit(s).  It assumes that all the argument buffers are of the same
 * size and have the same stripe unit offset.
 *
 * A "recovery xor" is used in the degraded-mode case.  It's similar
 * to the regular xor function except that it takes the failed PDA as
 * an additional parameter, and uses it to determine what portions of
 * the argument buffers need to be xor'd into the result buffer, and
 * where in the result buffer they should go.
 ****************************************************************************/

/* xor the params together and store the result in the result field.
 * assume the result field points to a buffer that is the size of one
 * SU, and use the pda params to determine where within the buffer to
 * XOR the input buffers.  */
void
rf_RegularXorFunc(RF_DagNode_t *node)
{
	RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p;
#if RF_ACC_TRACE > 0
	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
	RF_Etimer_t timer;
#endif
	int     i, retcode;

	retcode = 0;
	if (node->dagHdr->status == rf_enable) {
		/* don't do the XOR if the input is the same as the output */
#if RF_ACC_TRACE > 0
		RF_ETIMER_START(timer);
#endif
		for (i = 0; i < node->numParams - 1; i += 2)
			if (node->params[i + 1].p != node->results[0]) {
				retcode = rf_XorIntoBuffer(raidPtr, (RF_PhysDiskAddr_t *) node->params[i].p,
							   (char *) node->params[i + 1].p, (char *) node->results[0]);
			}
#if RF_ACC_TRACE > 0
		RF_ETIMER_STOP(timer);
		RF_ETIMER_EVAL(timer);
		tracerec->xor_us += RF_ETIMER_VAL_US(timer);
#endif
	}
	rf_GenericWakeupFunc(node, retcode);	/* call wake func
						 * explicitly since no
						 * I/O in this node */
}
/* xor the inputs into the result buffer, ignoring placement issues */
void
rf_SimpleXorFunc(RF_DagNode_t *node)
{
	RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p;
	int     i, retcode = 0;
#if RF_ACC_TRACE > 0
	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
	RF_Etimer_t timer;
#endif

	if (node->dagHdr->status == rf_enable) {
#if RF_ACC_TRACE > 0
		RF_ETIMER_START(timer);
#endif
		/* don't do the XOR if the input is the same as the output */
		for (i = 0; i < node->numParams - 1; i += 2)
			if (node->params[i + 1].p != node->results[0]) {
				retcode = rf_bxor((char *) node->params[i + 1].p, (char *) node->results[0],
				    rf_RaidAddressToByte(raidPtr, ((RF_PhysDiskAddr_t *) node->params[i].p)->numSector));
			}
#if RF_ACC_TRACE > 0
		RF_ETIMER_STOP(timer);
		RF_ETIMER_EVAL(timer);
		tracerec->xor_us += RF_ETIMER_VAL_US(timer);
#endif
	}
	rf_GenericWakeupFunc(node, retcode);	/* call wake func
						 * explicitly since no
						 * I/O in this node */
}
/* this xor is used by the degraded-mode dag functions to recover lost
 * data.  the second-to-last parameter is the PDA for the failed
 * portion of the access.  the code here looks at this PDA and assumes
 * that the xor target buffer is equal in size to the number of
 * sectors in the failed PDA.  It then uses the other PDAs in the
 * parameter list to determine where within the target buffer the
 * corresponding data should be xored.  */
void
rf_RecoveryXorFunc(RF_DagNode_t *node)
{
	RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p;
	RF_RaidLayout_t *layoutPtr = (RF_RaidLayout_t *) & raidPtr->Layout;
	RF_PhysDiskAddr_t *failedPDA = (RF_PhysDiskAddr_t *) node->params[node->numParams - 2].p;
	int     i, retcode = 0;
	RF_PhysDiskAddr_t *pda;
	int     suoffset, failedSUOffset = rf_StripeUnitOffset(layoutPtr, failedPDA->startSector);
	char   *srcbuf, *destbuf;
#if RF_ACC_TRACE > 0
	RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
	RF_Etimer_t timer;
#endif

	if (node->dagHdr->status == rf_enable) {
#if RF_ACC_TRACE > 0
		RF_ETIMER_START(timer);
#endif
		for (i = 0; i < node->numParams - 2; i += 2)
			if (node->params[i + 1].p != node->results[0]) {
				pda = (RF_PhysDiskAddr_t *) node->params[i].p;
				srcbuf = (char *) node->params[i + 1].p;
				suoffset = rf_StripeUnitOffset(layoutPtr, pda->startSector);
				destbuf = ((char *) node->results[0]) + rf_RaidAddressToByte(raidPtr, suoffset - failedSUOffset);
				retcode = rf_bxor(srcbuf, destbuf, rf_RaidAddressToByte(raidPtr, pda->numSector));
			}
#if RF_ACC_TRACE > 0
		RF_ETIMER_STOP(timer);
		RF_ETIMER_EVAL(timer);
		tracerec->xor_us += RF_ETIMER_VAL_US(timer);
#endif
	}
	rf_GenericWakeupFunc(node, retcode);
}
/*****************************************************************************
 * The next three functions are utilities used by the above
 * xor-execution functions.
 ****************************************************************************/


/*
 * this is just a glorified buffer xor.  targbuf points to a buffer
 * that is one full stripe unit in size.  srcbuf points to a buffer
 * that may be less than 1 SU, but never more.  When the access
 * described by pda is one SU in size (which by implication means it's
 * SU-aligned), all that happens is (targbuf) <- (srcbuf ^ targbuf).
 * When the access is less than one SU in size the XOR occurs on only
 * the portion of targbuf identified in the pda.  */

int
rf_XorIntoBuffer(RF_Raid_t *raidPtr, RF_PhysDiskAddr_t *pda,
		 char *srcbuf, char *targbuf)
{
	char   *targptr;
	int     sectPerSU = raidPtr->Layout.sectorsPerStripeUnit;
	int     SUOffset = pda->startSector % sectPerSU;
	int     length, retcode = 0;

	RF_ASSERT(pda->numSector <= sectPerSU);

	targptr = targbuf + rf_RaidAddressToByte(raidPtr, SUOffset);
	length = rf_RaidAddressToByte(raidPtr, pda->numSector);
	retcode = rf_bxor(srcbuf, targptr, length);
	return (retcode);
}
/* it really should be the case that the buffer pointers (returned by
 * malloc) are aligned to the natural word size of the machine, so
 * this is the only case we optimize for.  The length should always be
 * a multiple of the sector size, so there should be no problem with
 * leftover bytes at the end.  */
int
rf_bxor(char *src, char *dest, int len)
{
	unsigned mask = sizeof(long) - 1, retcode = 0;

	if (!(((unsigned long) src) & mask) &&
	    !(((unsigned long) dest) & mask) && !(len & mask)) {
		retcode = rf_longword_bxor((unsigned long *) src,
					   (unsigned long *) dest,
					   len >> RF_LONGSHIFT);
	} else {
		RF_ASSERT(0);
	}
	return (retcode);
}

/* When XORing in kernel mode, we need to map each user page to kernel
 * space before we can access it.  We don't want to assume anything
 * about which input buffers are in kernel/user space, nor about their
 * alignment, so in each loop we compute the maximum number of bytes
 * that we can xor without crossing any page boundaries, and do only
 * this many bytes before the next remap.
 *
 * len - is in longwords
 */
int
rf_longword_bxor(unsigned long *src, unsigned long *dest, int len)
{
	unsigned long *end = src + len;
	unsigned long d0, d1, d2, d3, s0, s1, s2, s3;	/* temps */
	unsigned long *pg_src, *pg_dest;   /* per-page source/dest pointers */
	int     longs_this_time;/* # longwords to xor in the current iteration */

	pg_src = src;
	pg_dest = dest;
	if (!pg_src || !pg_dest)
		return (EFAULT);

	while (len >= 4) {
		longs_this_time = RF_MIN(len, RF_MIN(RF_BLIP(pg_src), RF_BLIP(pg_dest)) >> RF_LONGSHIFT);	/* note len in longwords */
		src += longs_this_time;
		dest += longs_this_time;
		len -= longs_this_time;
		while (longs_this_time >= 4) {
			d0 = pg_dest[0];
			d1 = pg_dest[1];
			d2 = pg_dest[2];
			d3 = pg_dest[3];
			s0 = pg_src[0];
			s1 = pg_src[1];
			s2 = pg_src[2];
			s3 = pg_src[3];
			pg_dest[0] = d0 ^ s0;
			pg_dest[1] = d1 ^ s1;
			pg_dest[2] = d2 ^ s2;
			pg_dest[3] = d3 ^ s3;
			pg_src += 4;
			pg_dest += 4;
			longs_this_time -= 4;
		}
		while (longs_this_time > 0) {	/* cannot cross any page
						 * boundaries here */
			*pg_dest++ ^= *pg_src++;
			longs_this_time--;
		}

		/* either we're done, or we've reached a page boundary on one
		 * (or possibly both) of the pointers */
		if (len) {
			if (RF_PAGE_ALIGNED(src))
				pg_src = src;
			if (RF_PAGE_ALIGNED(dest))
				pg_dest = dest;
			if (!pg_src || !pg_dest)
				return (EFAULT);
		}
	}
	while (src < end) {
		*pg_dest++ ^= *pg_src++;
		src++;
		dest++;
		len--;
		if (RF_PAGE_ALIGNED(src))
			pg_src = src;
		if (RF_PAGE_ALIGNED(dest))
			pg_dest = dest;
	}
	RF_ASSERT(len == 0);
	return (0);
}

#if 0
/*
   dst = a ^ b ^ c;
   a may equal dst
   see comment above longword_bxor
   len is length in longwords
*/
int
rf_longword_bxor3(unsigned long *dst, unsigned long *a, unsigned long *b,
		  unsigned long *c, int len, void *bp)
{
	unsigned long a0, a1, a2, a3, b0, b1, b2, b3;
	unsigned long *pg_a, *pg_b, *pg_c, *pg_dst;	/* per-page source/dest
								 * pointers */
	int     longs_this_time;/* # longs to xor in the current iteration */
	char    dst_is_a = 0;

	pg_a = a;
	pg_b = b;
	pg_c = c;
	if (a == dst) {
		pg_dst = pg_a;
		dst_is_a = 1;
	} else {
		pg_dst = dst;
	}

	/* align dest to cache line.  Can't cross a pg boundary on dst here. */
	while ((((unsigned long) pg_dst) & 0x1f)) {
		*pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++;
		dst++;
		a++;
		b++;
		c++;
		if (RF_PAGE_ALIGNED(a)) {
			pg_a = a;
			if (!pg_a)
				return (EFAULT);
		}
		if (RF_PAGE_ALIGNED(b)) {
			pg_b = a;
			if (!pg_b)
				return (EFAULT);
		}
		if (RF_PAGE_ALIGNED(c)) {
			pg_c = a;
			if (!pg_c)
				return (EFAULT);
		}
		len--;
	}

	while (len > 4) {
		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);
		a += longs_this_time;
		b += longs_this_time;
		c += longs_this_time;
		dst += longs_this_time;
		len -= longs_this_time;
		while (longs_this_time >= 4) {
			a0 = pg_a[0];
			longs_this_time -= 4;

			a1 = pg_a[1];
			a2 = pg_a[2];

			a3 = pg_a[3];
			pg_a += 4;

			b0 = pg_b[0];
			b1 = pg_b[1];

			b2 = pg_b[2];
			b3 = pg_b[3];
			/* start dual issue */
			a0 ^= b0;
			b0 = pg_c[0];

			pg_b += 4;
			a1 ^= b1;

			a2 ^= b2;
			a3 ^= b3;

			b1 = pg_c[1];
			a0 ^= b0;

			b2 = pg_c[2];
			a1 ^= b1;

			b3 = pg_c[3];
			a2 ^= b2;

			pg_dst[0] = a0;
			a3 ^= b3;
			pg_dst[1] = a1;
			pg_c += 4;
			pg_dst[2] = a2;
			pg_dst[3] = a3;
			pg_dst += 4;
		}
		while (longs_this_time > 0) {	/* cannot cross any page
						 * boundaries here */
			*pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++;
			longs_this_time--;
		}

		if (len) {
			if (RF_PAGE_ALIGNED(a)) {
				pg_a = a;
				if (!pg_a)
					return (EFAULT);
				if (dst_is_a)
					pg_dst = pg_a;
			}
			if (RF_PAGE_ALIGNED(b)) {
				pg_b = b;
				if (!pg_b)
					return (EFAULT);
			}
			if (RF_PAGE_ALIGNED(c)) {
				pg_c = c;
				if (!pg_c)
					return (EFAULT);
			}
			if (!dst_is_a)
				if (RF_PAGE_ALIGNED(dst)) {
					pg_dst = dst;
					if (!pg_dst)
						return (EFAULT);
				}
		}
	}
	while (len) {
		*pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++;
		dst++;
		a++;
		b++;
		c++;
		if (RF_PAGE_ALIGNED(a)) {
			pg_a = a;
			if (!pg_a)
				return (EFAULT);
			if (dst_is_a)
				pg_dst = pg_a;
		}
		if (RF_PAGE_ALIGNED(b)) {
			pg_b = b;
			if (!pg_b)
				return (EFAULT);
		}
		if (RF_PAGE_ALIGNED(c)) {
			pg_c = c;
			if (!pg_c)
				return (EFAULT);
		}
		if (!dst_is_a)
			if (RF_PAGE_ALIGNED(dst)) {
				pg_dst = dst;
				if (!pg_dst)
					return (EFAULT);
			}
		len--;
	}
	return (0);
}

int
rf_bxor3(unsigned char *dst, unsigned char *a, unsigned char *b,
	 unsigned char *c, unsigned long len, void *bp)
{
	RF_ASSERT(((RF_UL(dst) | RF_UL(a) | RF_UL(b) | RF_UL(c) | len) & 0x7) == 0);

	return (rf_longword_bxor3((unsigned long *) dst, (unsigned long *) a,
		(unsigned long *) b, (unsigned long *) c, len >> RF_LONGSHIFT, bp));
}
#endif