xref: /netbsd-src/sys/dev/raidframe/rf_parityloggingdags.c (revision 3b01aba77a7a698587faaae455bbfe740923c1f5)
1 /*	$NetBSD: rf_parityloggingdags.c,v 1.4 2000/01/07 03:41:04 oster Exp $	*/
2 /*
3  * Copyright (c) 1995 Carnegie-Mellon University.
4  * All rights reserved.
5  *
6  * Author: 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 #include "rf_archs.h"
30 
31 #if RF_INCLUDE_PARITYLOGGING > 0
32 
33 /*
34   DAGs specific to parity logging are created here
35  */
36 
37 #include "rf_types.h"
38 #include "rf_raid.h"
39 #include "rf_dag.h"
40 #include "rf_dagutils.h"
41 #include "rf_dagfuncs.h"
42 #include "rf_debugMem.h"
43 #include "rf_paritylog.h"
44 #include "rf_memchunk.h"
45 #include "rf_general.h"
46 
47 #include "rf_parityloggingdags.h"
48 
49 /******************************************************************************
50  *
51  * creates a DAG to perform a large-write operation:
52  *
53  *         / Rod \     / Wnd \
54  * H -- NIL- Rod - NIL - Wnd ------ NIL - T
55  *         \ Rod /     \ Xor - Lpo /
56  *
57  * The writes are not done until the reads complete because if they were done in
58  * parallel, a failure on one of the reads could leave the parity in an inconsistent
59  * state, so that the retry with a new DAG would produce erroneous parity.
60  *
61  * Note:  this DAG has the nasty property that none of the buffers allocated for reading
62  *        old data can be freed until the XOR node fires.  Need to fix this.
63  *
64  * The last two arguments are the number of faults tolerated, and function for the
65  * redundancy calculation. The undo for the redundancy calc is assumed to be null
66  *
67  *****************************************************************************/
68 
69 void
70 rf_CommonCreateParityLoggingLargeWriteDAG(
71     RF_Raid_t * raidPtr,
72     RF_AccessStripeMap_t * asmap,
73     RF_DagHeader_t * dag_h,
74     void *bp,
75     RF_RaidAccessFlags_t flags,
76     RF_AllocListElem_t * allocList,
77     int nfaults,
78     int (*redFunc) (RF_DagNode_t *))
79 {
80 	RF_DagNode_t *nodes, *wndNodes, *rodNodes = NULL, *syncNode, *xorNode,
81 	       *lpoNode, *blockNode, *unblockNode, *termNode;
82 	int     nWndNodes, nRodNodes, i;
83 	RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
84 	RF_AccessStripeMapHeader_t *new_asm_h[2];
85 	int     nodeNum, asmNum;
86 	RF_ReconUnitNum_t which_ru;
87 	char   *sosBuffer, *eosBuffer;
88 	RF_PhysDiskAddr_t *pda;
89 	RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout), asmap->raidAddress, &which_ru);
90 
91 	if (rf_dagDebug)
92 		printf("[Creating parity-logging large-write DAG]\n");
93 	RF_ASSERT(nfaults == 1);/* this arch only single fault tolerant */
94 	dag_h->creator = "ParityLoggingLargeWriteDAG";
95 
96 	/* alloc the Wnd nodes, the xor node, and the Lpo node */
97 	nWndNodes = asmap->numStripeUnitsAccessed;
98 	RF_CallocAndAdd(nodes, nWndNodes + 6, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList);
99 	i = 0;
100 	wndNodes = &nodes[i];
101 	i += nWndNodes;
102 	xorNode = &nodes[i];
103 	i += 1;
104 	lpoNode = &nodes[i];
105 	i += 1;
106 	blockNode = &nodes[i];
107 	i += 1;
108 	syncNode = &nodes[i];
109 	i += 1;
110 	unblockNode = &nodes[i];
111 	i += 1;
112 	termNode = &nodes[i];
113 	i += 1;
114 
115 	dag_h->numCommitNodes = nWndNodes + 1;
116 	dag_h->numCommits = 0;
117 	dag_h->numSuccedents = 1;
118 
119 	rf_MapUnaccessedPortionOfStripe(raidPtr, layoutPtr, asmap, dag_h, new_asm_h, &nRodNodes, &sosBuffer, &eosBuffer, allocList);
120 	if (nRodNodes > 0)
121 		RF_CallocAndAdd(rodNodes, nRodNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList);
122 
123 	/* begin node initialization */
124 	rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nRodNodes + 1, 0, 0, 0, dag_h, "Nil", allocList);
125 	rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWndNodes + 1, 0, 0, dag_h, "Nil", allocList);
126 	rf_InitNode(syncNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nWndNodes + 1, nRodNodes + 1, 0, 0, dag_h, "Nil", allocList);
127 	rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList);
128 
129 	/* initialize the Rod nodes */
130 	for (nodeNum = asmNum = 0; asmNum < 2; asmNum++) {
131 		if (new_asm_h[asmNum]) {
132 			pda = new_asm_h[asmNum]->stripeMap->physInfo;
133 			while (pda) {
134 				rf_InitNode(&rodNodes[nodeNum], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rod", allocList);
135 				rodNodes[nodeNum].params[0].p = pda;
136 				rodNodes[nodeNum].params[1].p = pda->bufPtr;
137 				rodNodes[nodeNum].params[2].v = parityStripeID;
138 				rodNodes[nodeNum].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
139 				nodeNum++;
140 				pda = pda->next;
141 			}
142 		}
143 	}
144 	RF_ASSERT(nodeNum == nRodNodes);
145 
146 	/* initialize the wnd nodes */
147 	pda = asmap->physInfo;
148 	for (i = 0; i < nWndNodes; i++) {
149 		rf_InitNode(&wndNodes[i], rf_wait, RF_TRUE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList);
150 		RF_ASSERT(pda != NULL);
151 		wndNodes[i].params[0].p = pda;
152 		wndNodes[i].params[1].p = pda->bufPtr;
153 		wndNodes[i].params[2].v = parityStripeID;
154 		wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
155 		pda = pda->next;
156 	}
157 
158 	/* initialize the redundancy node */
159 	rf_InitNode(xorNode, rf_wait, RF_TRUE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 1, 2 * (nWndNodes + nRodNodes) + 1, 1, dag_h, "Xr ", allocList);
160 	xorNode->flags |= RF_DAGNODE_FLAG_YIELD;
161 	for (i = 0; i < nWndNodes; i++) {
162 		xorNode->params[2 * i + 0] = wndNodes[i].params[0];	/* pda */
163 		xorNode->params[2 * i + 1] = wndNodes[i].params[1];	/* buf ptr */
164 	}
165 	for (i = 0; i < nRodNodes; i++) {
166 		xorNode->params[2 * (nWndNodes + i) + 0] = rodNodes[i].params[0];	/* pda */
167 		xorNode->params[2 * (nWndNodes + i) + 1] = rodNodes[i].params[1];	/* buf ptr */
168 	}
169 	xorNode->params[2 * (nWndNodes + nRodNodes)].p = raidPtr;	/* xor node needs to get
170 									 * at RAID information */
171 
172 	/* look for an Rod node that reads a complete SU.  If none, alloc a
173 	 * buffer to receive the parity info. Note that we can't use a new
174 	 * data buffer because it will not have gotten written when the xor
175 	 * occurs. */
176 	for (i = 0; i < nRodNodes; i++)
177 		if (((RF_PhysDiskAddr_t *) rodNodes[i].params[0].p)->numSector == raidPtr->Layout.sectorsPerStripeUnit)
178 			break;
179 	if (i == nRodNodes) {
180 		RF_CallocAndAdd(xorNode->results[0], 1, rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit), (void *), allocList);
181 	} else {
182 		xorNode->results[0] = rodNodes[i].params[1].p;
183 	}
184 
185 	/* initialize the Lpo node */
186 	rf_InitNode(lpoNode, rf_wait, RF_FALSE, rf_ParityLogOverwriteFunc, rf_ParityLogOverwriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Lpo", allocList);
187 
188 	lpoNode->params[0].p = asmap->parityInfo;
189 	lpoNode->params[1].p = xorNode->results[0];
190 	RF_ASSERT(asmap->parityInfo->next == NULL);	/* parityInfo must
191 							 * describe entire
192 							 * parity unit */
193 
194 	/* connect nodes to form graph */
195 
196 	/* connect dag header to block node */
197 	RF_ASSERT(dag_h->numSuccedents == 1);
198 	RF_ASSERT(blockNode->numAntecedents == 0);
199 	dag_h->succedents[0] = blockNode;
200 
201 	/* connect the block node to the Rod nodes */
202 	RF_ASSERT(blockNode->numSuccedents == nRodNodes + 1);
203 	for (i = 0; i < nRodNodes; i++) {
204 		RF_ASSERT(rodNodes[i].numAntecedents == 1);
205 		blockNode->succedents[i] = &rodNodes[i];
206 		rodNodes[i].antecedents[0] = blockNode;
207 		rodNodes[i].antType[0] = rf_control;
208 	}
209 
210 	/* connect the block node to the sync node */
211 	/* necessary if nRodNodes == 0 */
212 	RF_ASSERT(syncNode->numAntecedents == nRodNodes + 1);
213 	blockNode->succedents[nRodNodes] = syncNode;
214 	syncNode->antecedents[0] = blockNode;
215 	syncNode->antType[0] = rf_control;
216 
217 	/* connect the Rod nodes to the syncNode */
218 	for (i = 0; i < nRodNodes; i++) {
219 		rodNodes[i].succedents[0] = syncNode;
220 		syncNode->antecedents[1 + i] = &rodNodes[i];
221 		syncNode->antType[1 + i] = rf_control;
222 	}
223 
224 	/* connect the sync node to the xor node */
225 	RF_ASSERT(syncNode->numSuccedents == nWndNodes + 1);
226 	RF_ASSERT(xorNode->numAntecedents == 1);
227 	syncNode->succedents[0] = xorNode;
228 	xorNode->antecedents[0] = syncNode;
229 	xorNode->antType[0] = rf_trueData;	/* carry forward from sync */
230 
231 	/* connect the sync node to the Wnd nodes */
232 	for (i = 0; i < nWndNodes; i++) {
233 		RF_ASSERT(wndNodes->numAntecedents == 1);
234 		syncNode->succedents[1 + i] = &wndNodes[i];
235 		wndNodes[i].antecedents[0] = syncNode;
236 		wndNodes[i].antType[0] = rf_control;
237 	}
238 
239 	/* connect the xor node to the Lpo node */
240 	RF_ASSERT(xorNode->numSuccedents == 1);
241 	RF_ASSERT(lpoNode->numAntecedents == 1);
242 	xorNode->succedents[0] = lpoNode;
243 	lpoNode->antecedents[0] = xorNode;
244 	lpoNode->antType[0] = rf_trueData;
245 
246 	/* connect the Wnd nodes to the unblock node */
247 	RF_ASSERT(unblockNode->numAntecedents == nWndNodes + 1);
248 	for (i = 0; i < nWndNodes; i++) {
249 		RF_ASSERT(wndNodes->numSuccedents == 1);
250 		wndNodes[i].succedents[0] = unblockNode;
251 		unblockNode->antecedents[i] = &wndNodes[i];
252 		unblockNode->antType[i] = rf_control;
253 	}
254 
255 	/* connect the Lpo node to the unblock node */
256 	RF_ASSERT(lpoNode->numSuccedents == 1);
257 	lpoNode->succedents[0] = unblockNode;
258 	unblockNode->antecedents[nWndNodes] = lpoNode;
259 	unblockNode->antType[nWndNodes] = rf_control;
260 
261 	/* connect unblock node to terminator */
262 	RF_ASSERT(unblockNode->numSuccedents == 1);
263 	RF_ASSERT(termNode->numAntecedents == 1);
264 	RF_ASSERT(termNode->numSuccedents == 0);
265 	unblockNode->succedents[0] = termNode;
266 	termNode->antecedents[0] = unblockNode;
267 	termNode->antType[0] = rf_control;
268 }
269 
270 
271 
272 
273 /******************************************************************************
274  *
275  * creates a DAG to perform a small-write operation (either raid 5 or pq), which is as follows:
276  *
277  *                                     Header
278  *                                       |
279  *                                     Block
280  *                                 / |  ... \   \
281  *                                /  |       \   \
282  *                             Rod  Rod      Rod  Rop
283  *                             | \ /| \    / |  \/ |
284  *                             |    |        |  /\ |
285  *                             Wnd  Wnd      Wnd   X
286  *                              |    \       /     |
287  *                              |     \     /      |
288  *                               \     \   /      Lpo
289  *                                \     \ /       /
290  *                                 +-> Unblock <-+
291  *                                       |
292  *                                       T
293  *
294  *
295  * R = Read, W = Write, X = Xor, o = old, n = new, d = data, p = parity.
296  * When the access spans a stripe unit boundary and is less than one SU in size, there will
297  * be two Rop -- X -- Wnp branches.  I call this the "double-XOR" case.
298  * The second output from each Rod node goes to the X node.  In the double-XOR
299  * case, there are exactly 2 Rod nodes, and each sends one output to one X node.
300  * There is one Rod -- Wnd -- T branch for each stripe unit being updated.
301  *
302  * The block and unblock nodes are unused.  See comment above CreateFaultFreeReadDAG.
303  *
304  * Note:  this DAG ignores all the optimizations related to making the RMWs atomic.
305  *        it also has the nasty property that none of the buffers allocated for reading
306  *        old data & parity can be freed until the XOR node fires.  Need to fix this.
307  *
308  * A null qfuncs indicates single fault tolerant
309  *****************************************************************************/
310 
311 void
312 rf_CommonCreateParityLoggingSmallWriteDAG(
313     RF_Raid_t * raidPtr,
314     RF_AccessStripeMap_t * asmap,
315     RF_DagHeader_t * dag_h,
316     void *bp,
317     RF_RaidAccessFlags_t flags,
318     RF_AllocListElem_t * allocList,
319     RF_RedFuncs_t * pfuncs,
320     RF_RedFuncs_t * qfuncs)
321 {
322 	RF_DagNode_t *xorNodes, *blockNode, *unblockNode, *nodes;
323 	RF_DagNode_t *readDataNodes, *readParityNodes;
324 	RF_DagNode_t *writeDataNodes, *lpuNodes;
325 	RF_DagNode_t *unlockDataNodes = NULL, *termNode;
326 	RF_PhysDiskAddr_t *pda = asmap->physInfo;
327 	int     numDataNodes = asmap->numStripeUnitsAccessed;
328 	int     numParityNodes = (asmap->parityInfo->next) ? 2 : 1;
329 	int     i, j, nNodes, totalNumNodes;
330 	RF_ReconUnitNum_t which_ru;
331 	int     (*func) (RF_DagNode_t * node), (*undoFunc) (RF_DagNode_t * node);
332 	int     (*qfunc) (RF_DagNode_t * node);
333 	char   *name, *qname;
334 	RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout), asmap->raidAddress, &which_ru);
335 	long    nfaults = qfuncs ? 2 : 1;
336 	int     lu_flag = (rf_enableAtomicRMW) ? 1 : 0;	/* lock/unlock flag */
337 
338 	if (rf_dagDebug)
339 		printf("[Creating parity-logging small-write DAG]\n");
340 	RF_ASSERT(numDataNodes > 0);
341 	RF_ASSERT(nfaults == 1);
342 	dag_h->creator = "ParityLoggingSmallWriteDAG";
343 
344 	/* DAG creation occurs in three steps: 1. count the number of nodes in
345 	 * the DAG 2. create the nodes 3. initialize the nodes 4. connect the
346 	 * nodes */
347 
348 	/* Step 1. compute number of nodes in the graph */
349 
350 	/* number of nodes: a read and write for each data unit a redundancy
351 	 * computation node for each parity node a read and Lpu for each
352 	 * parity unit a block and unblock node (2) a terminator node if
353 	 * atomic RMW an unlock node for each data unit, redundancy unit */
354 	totalNumNodes = (2 * numDataNodes) + numParityNodes + (2 * numParityNodes) + 3;
355 	if (lu_flag)
356 		totalNumNodes += numDataNodes;
357 
358 	nNodes = numDataNodes + numParityNodes;
359 
360 	dag_h->numCommitNodes = numDataNodes + numParityNodes;
361 	dag_h->numCommits = 0;
362 	dag_h->numSuccedents = 1;
363 
364 	/* Step 2. create the nodes */
365 	RF_CallocAndAdd(nodes, totalNumNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList);
366 	i = 0;
367 	blockNode = &nodes[i];
368 	i += 1;
369 	unblockNode = &nodes[i];
370 	i += 1;
371 	readDataNodes = &nodes[i];
372 	i += numDataNodes;
373 	readParityNodes = &nodes[i];
374 	i += numParityNodes;
375 	writeDataNodes = &nodes[i];
376 	i += numDataNodes;
377 	lpuNodes = &nodes[i];
378 	i += numParityNodes;
379 	xorNodes = &nodes[i];
380 	i += numParityNodes;
381 	termNode = &nodes[i];
382 	i += 1;
383 	if (lu_flag) {
384 		unlockDataNodes = &nodes[i];
385 		i += numDataNodes;
386 	}
387 	RF_ASSERT(i == totalNumNodes);
388 
389 	/* Step 3. initialize the nodes */
390 	/* initialize block node (Nil) */
391 	rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nNodes, 0, 0, 0, dag_h, "Nil", allocList);
392 
393 	/* initialize unblock node (Nil) */
394 	rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nNodes, 0, 0, dag_h, "Nil", allocList);
395 
396 	/* initialize terminatory node (Trm) */
397 	rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList);
398 
399 	/* initialize nodes which read old data (Rod) */
400 	for (i = 0; i < numDataNodes; i++) {
401 		rf_InitNode(&readDataNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, nNodes, 1, 4, 0, dag_h, "Rod", allocList);
402 		RF_ASSERT(pda != NULL);
403 		readDataNodes[i].params[0].p = pda;	/* physical disk addr
404 							 * desc */
405 		readDataNodes[i].params[1].p = rf_AllocBuffer(raidPtr, dag_h, pda, allocList);	/* buffer to hold old
406 												 * data */
407 		readDataNodes[i].params[2].v = parityStripeID;
408 		readDataNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, lu_flag, 0, which_ru);
409 		pda = pda->next;
410 		readDataNodes[i].propList[0] = NULL;
411 		readDataNodes[i].propList[1] = NULL;
412 	}
413 
414 	/* initialize nodes which read old parity (Rop) */
415 	pda = asmap->parityInfo;
416 	i = 0;
417 	for (i = 0; i < numParityNodes; i++) {
418 		RF_ASSERT(pda != NULL);
419 		rf_InitNode(&readParityNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, nNodes, 1, 4, 0, dag_h, "Rop", allocList);
420 		readParityNodes[i].params[0].p = pda;
421 		readParityNodes[i].params[1].p = rf_AllocBuffer(raidPtr, dag_h, pda, allocList);	/* buffer to hold old
422 													 * parity */
423 		readParityNodes[i].params[2].v = parityStripeID;
424 		readParityNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
425 		readParityNodes[i].propList[0] = NULL;
426 		pda = pda->next;
427 	}
428 
429 	/* initialize nodes which write new data (Wnd) */
430 	pda = asmap->physInfo;
431 	for (i = 0; i < numDataNodes; i++) {
432 		RF_ASSERT(pda != NULL);
433 		rf_InitNode(&writeDataNodes[i], rf_wait, RF_TRUE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, nNodes, 4, 0, dag_h, "Wnd", allocList);
434 		writeDataNodes[i].params[0].p = pda;	/* physical disk addr
435 							 * desc */
436 		writeDataNodes[i].params[1].p = pda->bufPtr;	/* buffer holding new
437 								 * data to be written */
438 		writeDataNodes[i].params[2].v = parityStripeID;
439 		writeDataNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
440 
441 		if (lu_flag) {
442 			/* initialize node to unlock the disk queue */
443 			rf_InitNode(&unlockDataNodes[i], rf_wait, RF_FALSE, rf_DiskUnlockFunc, rf_DiskUnlockUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Und", allocList);
444 			unlockDataNodes[i].params[0].p = pda;	/* physical disk addr
445 								 * desc */
446 			unlockDataNodes[i].params[1].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, lu_flag, which_ru);
447 		}
448 		pda = pda->next;
449 	}
450 
451 
452 	/* initialize nodes which compute new parity */
453 	/* we use the simple XOR func in the double-XOR case, and when we're
454 	 * accessing only a portion of one stripe unit. the distinction
455 	 * between the two is that the regular XOR func assumes that the
456 	 * targbuf is a full SU in size, and examines the pda associated with
457 	 * the buffer to decide where within the buffer to XOR the data,
458 	 * whereas the simple XOR func just XORs the data into the start of
459 	 * the buffer. */
460 	if ((numParityNodes == 2) || ((numDataNodes == 1) && (asmap->totalSectorsAccessed < raidPtr->Layout.sectorsPerStripeUnit))) {
461 		func = pfuncs->simple;
462 		undoFunc = rf_NullNodeUndoFunc;
463 		name = pfuncs->SimpleName;
464 		if (qfuncs) {
465 			qfunc = qfuncs->simple;
466 			qname = qfuncs->SimpleName;
467 		}
468 	} else {
469 		func = pfuncs->regular;
470 		undoFunc = rf_NullNodeUndoFunc;
471 		name = pfuncs->RegularName;
472 		if (qfuncs) {
473 			qfunc = qfuncs->regular;
474 			qname = qfuncs->RegularName;
475 		}
476 	}
477 	/* initialize the xor nodes: params are {pda,buf} from {Rod,Wnd,Rop}
478 	 * nodes, and raidPtr  */
479 	if (numParityNodes == 2) {	/* double-xor case */
480 		for (i = 0; i < numParityNodes; i++) {
481 			rf_InitNode(&xorNodes[i], rf_wait, RF_TRUE, func, undoFunc, NULL, 1, nNodes, 7, 1, dag_h, name, allocList);	/* no wakeup func for
482 																	 * xor */
483 			xorNodes[i].flags |= RF_DAGNODE_FLAG_YIELD;
484 			xorNodes[i].params[0] = readDataNodes[i].params[0];
485 			xorNodes[i].params[1] = readDataNodes[i].params[1];
486 			xorNodes[i].params[2] = readParityNodes[i].params[0];
487 			xorNodes[i].params[3] = readParityNodes[i].params[1];
488 			xorNodes[i].params[4] = writeDataNodes[i].params[0];
489 			xorNodes[i].params[5] = writeDataNodes[i].params[1];
490 			xorNodes[i].params[6].p = raidPtr;
491 			xorNodes[i].results[0] = readParityNodes[i].params[1].p;	/* use old parity buf as
492 											 * target buf */
493 		}
494 	} else {
495 		/* there is only one xor node in this case */
496 		rf_InitNode(&xorNodes[0], rf_wait, RF_TRUE, func, undoFunc, NULL, 1, nNodes, (2 * (numDataNodes + numDataNodes + 1) + 1), 1, dag_h, name, allocList);
497 		xorNodes[0].flags |= RF_DAGNODE_FLAG_YIELD;
498 		for (i = 0; i < numDataNodes + 1; i++) {
499 			/* set up params related to Rod and Rop nodes */
500 			xorNodes[0].params[2 * i + 0] = readDataNodes[i].params[0];	/* pda */
501 			xorNodes[0].params[2 * i + 1] = readDataNodes[i].params[1];	/* buffer pointer */
502 		}
503 		for (i = 0; i < numDataNodes; i++) {
504 			/* set up params related to Wnd and Wnp nodes */
505 			xorNodes[0].params[2 * (numDataNodes + 1 + i) + 0] = writeDataNodes[i].params[0];	/* pda */
506 			xorNodes[0].params[2 * (numDataNodes + 1 + i) + 1] = writeDataNodes[i].params[1];	/* buffer pointer */
507 		}
508 		xorNodes[0].params[2 * (numDataNodes + numDataNodes + 1)].p = raidPtr;	/* xor node needs to get
509 											 * at RAID information */
510 		xorNodes[0].results[0] = readParityNodes[0].params[1].p;
511 	}
512 
513 	/* initialize the log node(s) */
514 	pda = asmap->parityInfo;
515 	for (i = 0; i < numParityNodes; i++) {
516 		RF_ASSERT(pda);
517 		rf_InitNode(&lpuNodes[i], rf_wait, RF_FALSE, rf_ParityLogUpdateFunc, rf_ParityLogUpdateUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Lpu", allocList);
518 		lpuNodes[i].params[0].p = pda;	/* PhysDiskAddr of parity */
519 		lpuNodes[i].params[1].p = xorNodes[i].results[0];	/* buffer pointer to
520 									 * parity */
521 		pda = pda->next;
522 	}
523 
524 
525 	/* Step 4. connect the nodes */
526 
527 	/* connect header to block node */
528 	RF_ASSERT(dag_h->numSuccedents == 1);
529 	RF_ASSERT(blockNode->numAntecedents == 0);
530 	dag_h->succedents[0] = blockNode;
531 
532 	/* connect block node to read old data nodes */
533 	RF_ASSERT(blockNode->numSuccedents == (numDataNodes + numParityNodes));
534 	for (i = 0; i < numDataNodes; i++) {
535 		blockNode->succedents[i] = &readDataNodes[i];
536 		RF_ASSERT(readDataNodes[i].numAntecedents == 1);
537 		readDataNodes[i].antecedents[0] = blockNode;
538 		readDataNodes[i].antType[0] = rf_control;
539 	}
540 
541 	/* connect block node to read old parity nodes */
542 	for (i = 0; i < numParityNodes; i++) {
543 		blockNode->succedents[numDataNodes + i] = &readParityNodes[i];
544 		RF_ASSERT(readParityNodes[i].numAntecedents == 1);
545 		readParityNodes[i].antecedents[0] = blockNode;
546 		readParityNodes[i].antType[0] = rf_control;
547 	}
548 
549 	/* connect read old data nodes to write new data nodes */
550 	for (i = 0; i < numDataNodes; i++) {
551 		RF_ASSERT(readDataNodes[i].numSuccedents == numDataNodes + numParityNodes);
552 		for (j = 0; j < numDataNodes; j++) {
553 			RF_ASSERT(writeDataNodes[j].numAntecedents == numDataNodes + numParityNodes);
554 			readDataNodes[i].succedents[j] = &writeDataNodes[j];
555 			writeDataNodes[j].antecedents[i] = &readDataNodes[i];
556 			if (i == j)
557 				writeDataNodes[j].antType[i] = rf_antiData;
558 			else
559 				writeDataNodes[j].antType[i] = rf_control;
560 		}
561 	}
562 
563 	/* connect read old data nodes to xor nodes */
564 	for (i = 0; i < numDataNodes; i++)
565 		for (j = 0; j < numParityNodes; j++) {
566 			RF_ASSERT(xorNodes[j].numAntecedents == numDataNodes + numParityNodes);
567 			readDataNodes[i].succedents[numDataNodes + j] = &xorNodes[j];
568 			xorNodes[j].antecedents[i] = &readDataNodes[i];
569 			xorNodes[j].antType[i] = rf_trueData;
570 		}
571 
572 	/* connect read old parity nodes to write new data nodes */
573 	for (i = 0; i < numParityNodes; i++) {
574 		RF_ASSERT(readParityNodes[i].numSuccedents == numDataNodes + numParityNodes);
575 		for (j = 0; j < numDataNodes; j++) {
576 			readParityNodes[i].succedents[j] = &writeDataNodes[j];
577 			writeDataNodes[j].antecedents[numDataNodes + i] = &readParityNodes[i];
578 			writeDataNodes[j].antType[numDataNodes + i] = rf_control;
579 		}
580 	}
581 
582 	/* connect read old parity nodes to xor nodes */
583 	for (i = 0; i < numParityNodes; i++)
584 		for (j = 0; j < numParityNodes; j++) {
585 			readParityNodes[i].succedents[numDataNodes + j] = &xorNodes[j];
586 			xorNodes[j].antecedents[numDataNodes + i] = &readParityNodes[i];
587 			xorNodes[j].antType[numDataNodes + i] = rf_trueData;
588 		}
589 
590 	/* connect xor nodes to write new parity nodes */
591 	for (i = 0; i < numParityNodes; i++) {
592 		RF_ASSERT(xorNodes[i].numSuccedents == 1);
593 		RF_ASSERT(lpuNodes[i].numAntecedents == 1);
594 		xorNodes[i].succedents[0] = &lpuNodes[i];
595 		lpuNodes[i].antecedents[0] = &xorNodes[i];
596 		lpuNodes[i].antType[0] = rf_trueData;
597 	}
598 
599 	for (i = 0; i < numDataNodes; i++) {
600 		if (lu_flag) {
601 			/* connect write new data nodes to unlock nodes */
602 			RF_ASSERT(writeDataNodes[i].numSuccedents == 1);
603 			RF_ASSERT(unlockDataNodes[i].numAntecedents == 1);
604 			writeDataNodes[i].succedents[0] = &unlockDataNodes[i];
605 			unlockDataNodes[i].antecedents[0] = &writeDataNodes[i];
606 			unlockDataNodes[i].antType[0] = rf_control;
607 
608 			/* connect unlock nodes to unblock node */
609 			RF_ASSERT(unlockDataNodes[i].numSuccedents == 1);
610 			RF_ASSERT(unblockNode->numAntecedents == (numDataNodes + (nfaults * numParityNodes)));
611 			unlockDataNodes[i].succedents[0] = unblockNode;
612 			unblockNode->antecedents[i] = &unlockDataNodes[i];
613 			unblockNode->antType[i] = rf_control;
614 		} else {
615 			/* connect write new data nodes to unblock node */
616 			RF_ASSERT(writeDataNodes[i].numSuccedents == 1);
617 			RF_ASSERT(unblockNode->numAntecedents == (numDataNodes + (nfaults * numParityNodes)));
618 			writeDataNodes[i].succedents[0] = unblockNode;
619 			unblockNode->antecedents[i] = &writeDataNodes[i];
620 			unblockNode->antType[i] = rf_control;
621 		}
622 	}
623 
624 	/* connect write new parity nodes to unblock node */
625 	for (i = 0; i < numParityNodes; i++) {
626 		RF_ASSERT(lpuNodes[i].numSuccedents == 1);
627 		lpuNodes[i].succedents[0] = unblockNode;
628 		unblockNode->antecedents[numDataNodes + i] = &lpuNodes[i];
629 		unblockNode->antType[numDataNodes + i] = rf_control;
630 	}
631 
632 	/* connect unblock node to terminator */
633 	RF_ASSERT(unblockNode->numSuccedents == 1);
634 	RF_ASSERT(termNode->numAntecedents == 1);
635 	RF_ASSERT(termNode->numSuccedents == 0);
636 	unblockNode->succedents[0] = termNode;
637 	termNode->antecedents[0] = unblockNode;
638 	termNode->antType[0] = rf_control;
639 }
640 
641 
642 void
643 rf_CreateParityLoggingSmallWriteDAG(
644     RF_Raid_t * raidPtr,
645     RF_AccessStripeMap_t * asmap,
646     RF_DagHeader_t * dag_h,
647     void *bp,
648     RF_RaidAccessFlags_t flags,
649     RF_AllocListElem_t * allocList,
650     RF_RedFuncs_t * pfuncs,
651     RF_RedFuncs_t * qfuncs)
652 {
653 	dag_h->creator = "ParityLoggingSmallWriteDAG";
654 	rf_CommonCreateParityLoggingSmallWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList, &rf_xorFuncs, NULL);
655 }
656 
657 
658 void
659 rf_CreateParityLoggingLargeWriteDAG(
660     RF_Raid_t * raidPtr,
661     RF_AccessStripeMap_t * asmap,
662     RF_DagHeader_t * dag_h,
663     void *bp,
664     RF_RaidAccessFlags_t flags,
665     RF_AllocListElem_t * allocList,
666     int nfaults,
667     int (*redFunc) (RF_DagNode_t *))
668 {
669 	dag_h->creator = "ParityLoggingSmallWriteDAG";
670 	rf_CommonCreateParityLoggingLargeWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList, 1, rf_RegularXorFunc);
671 }
672 #endif				/* RF_INCLUDE_PARITYLOGGING > 0 */
673