1 /* $NetBSD: rf_dagdegwr.c,v 1.37 2023/10/15 18:15:19 oster Exp $ */
2 /*
3 * Copyright (c) 1995 Carnegie-Mellon University.
4 * All rights reserved.
5 *
6 * Author: Mark Holland, Daniel Stodolsky, 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 * rf_dagdegwr.c
31 *
32 * code for creating degraded write DAGs
33 *
34 */
35
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: rf_dagdegwr.c,v 1.37 2023/10/15 18:15:19 oster Exp $");
38
39 #include <dev/raidframe/raidframevar.h>
40
41 #include "rf_raid.h"
42 #include "rf_dag.h"
43 #include "rf_dagutils.h"
44 #include "rf_dagfuncs.h"
45 #include "rf_debugMem.h"
46 #include "rf_general.h"
47 #include "rf_dagdegwr.h"
48 #include "rf_map.h"
49
50
51 /******************************************************************************
52 *
53 * General comments on DAG creation:
54 *
55 * All DAGs in this file use roll-away error recovery. Each DAG has a single
56 * commit node, usually called "Cmt." If an error occurs before the Cmt node
57 * is reached, the execution engine will halt forward execution and work
58 * backward through the graph, executing the undo functions. Assuming that
59 * each node in the graph prior to the Cmt node are undoable and atomic - or -
60 * does not make changes to permanent state, the graph will fail atomically.
61 * If an error occurs after the Cmt node executes, the engine will roll-forward
62 * through the graph, blindly executing nodes until it reaches the end.
63 * If a graph reaches the end, it is assumed to have completed successfully.
64 *
65 * A graph has only 1 Cmt node.
66 *
67 */
68
69
70 /******************************************************************************
71 *
72 * The following wrappers map the standard DAG creation interface to the
73 * DAG creation routines. Additionally, these wrappers enable experimentation
74 * with new DAG structures by providing an extra level of indirection, allowing
75 * the DAG creation routines to be replaced at this single point.
76 */
77
78 static
RF_CREATE_DAG_FUNC_DECL(rf_CreateSimpleDegradedWriteDAG)79 RF_CREATE_DAG_FUNC_DECL(rf_CreateSimpleDegradedWriteDAG)
80 {
81 rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp,
82 flags, allocList, 1, rf_RecoveryXorFunc, RF_TRUE);
83 }
84
85 void
rf_CreateDegradedWriteDAG(RF_Raid_t * raidPtr,RF_AccessStripeMap_t * asmap,RF_DagHeader_t * dag_h,void * bp,RF_RaidAccessFlags_t flags,RF_AllocListElem_t * allocList)86 rf_CreateDegradedWriteDAG(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap,
87 RF_DagHeader_t *dag_h, void *bp,
88 RF_RaidAccessFlags_t flags,
89 RF_AllocListElem_t *allocList)
90 {
91
92 RF_ASSERT(asmap->numDataFailed == 1);
93 dag_h->creator = "DegradedWriteDAG";
94
95 /*
96 * if the access writes only a portion of the failed unit, and also
97 * writes some portion of at least one surviving unit, we create two
98 * DAGs, one for the failed component and one for the non-failed
99 * component, and do them sequentially. Note that the fact that we're
100 * accessing only a portion of the failed unit indicates that the
101 * access either starts or ends in the failed unit, and hence we need
102 * create only two dags. This is inefficient in that the same data or
103 * parity can get read and written twice using this structure. I need
104 * to fix this to do the access all at once.
105 */
106 RF_ASSERT(!(asmap->numStripeUnitsAccessed != 1 &&
107 asmap->failedPDAs[0]->numSector !=
108 raidPtr->Layout.sectorsPerStripeUnit));
109 rf_CreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags,
110 allocList);
111 }
112
113
114
115 /******************************************************************************
116 *
117 * DAG creation code begins here
118 */
119 #define BUF_ALLOC(num) \
120 RF_MallocAndAdd(rf_RaidAddressToByte(raidPtr, num), allocList)
121
122
123
124 /******************************************************************************
125 *
126 * CommonCreateSimpleDegradedWriteDAG -- creates a DAG to do a degraded-mode
127 * write, which is as follows
128 *
129 * / {Wnq} --\
130 * hdr -> blockNode -> Rod -> Xor -> Cmt -> Wnp ----> unblock -> term
131 * \ {Rod} / \ Wnd ---/
132 * \ {Wnd} -/
133 *
134 * commit nodes: Xor, Wnd
135 *
136 * IMPORTANT:
137 * This DAG generator does not work for double-degraded archs since it does not
138 * generate Q
139 *
140 * This dag is essentially identical to the large-write dag, except that the
141 * write to the failed data unit is suppressed.
142 *
143 * IMPORTANT: this dag does not work in the case where the access writes only
144 * a portion of the failed unit, and also writes some portion of at least one
145 * surviving SU. this case is handled in CreateDegradedWriteDAG above.
146 *
147 * The block & unblock nodes are leftovers from a previous version. They
148 * do nothing, but I haven't deleted them because it would be a tremendous
149 * effort to put them back in.
150 *
151 * This dag is used whenever a one of the data units in a write has failed.
152 * If it is the parity unit that failed, the nonredundant write dag (below)
153 * is used.
154 *****************************************************************************/
155
156 void
rf_CommonCreateSimpleDegradedWriteDAG(RF_Raid_t * raidPtr,RF_AccessStripeMap_t * asmap,RF_DagHeader_t * dag_h,void * bp,RF_RaidAccessFlags_t flags,RF_AllocListElem_t * allocList,int nfaults,void (* redFunc)(RF_DagNode_t *),int allowBufferRecycle)157 rf_CommonCreateSimpleDegradedWriteDAG(RF_Raid_t *raidPtr,
158 RF_AccessStripeMap_t *asmap,
159 RF_DagHeader_t *dag_h, void *bp,
160 RF_RaidAccessFlags_t flags,
161 RF_AllocListElem_t *allocList,
162 int nfaults,
163 void (*redFunc) (RF_DagNode_t *),
164 int allowBufferRecycle)
165 {
166 int nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum,
167 rdnodesFaked;
168 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *termNode;
169 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0)
170 RF_DagNode_t *wnqNode;
171 #endif
172 RF_DagNode_t *wndNodes, *rrdNodes, *xorNode, *commitNode;
173 RF_DagNode_t *tmpNode, *tmpwndNode, *tmprrdNode;
174 RF_SectorCount_t sectorsPerSU;
175 RF_ReconUnitNum_t which_ru;
176 char *xorTargetBuf = NULL; /* the target buffer for the XOR
177 * operation */
178 char overlappingPDAs[RF_MAXCOL];/* a temporary array of flags */
179 RF_AccessStripeMapHeader_t *new_asm_h[2];
180 RF_PhysDiskAddr_t *pda, *parityPDA;
181 RF_StripeNum_t parityStripeID;
182 RF_PhysDiskAddr_t *failedPDA;
183 RF_RaidLayout_t *layoutPtr;
184
185 layoutPtr = &(raidPtr->Layout);
186 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress,
187 &which_ru);
188 sectorsPerSU = layoutPtr->sectorsPerStripeUnit;
189 /* failedPDA points to the pda within the asm that targets the failed
190 * disk */
191 failedPDA = asmap->failedPDAs[0];
192
193 #if RF_DEBUG_DAG
194 if (rf_dagDebug)
195 printf("[Creating degraded-write DAG]\n");
196 #endif
197
198 RF_ASSERT(asmap->numDataFailed == 1);
199 dag_h->creator = "SimpleDegradedWriteDAG";
200
201 /*
202 * Generate two ASMs identifying the surviving data
203 * we need in order to recover the lost data.
204 */
205 /* overlappingPDAs array must be zero'd */
206 memset(overlappingPDAs, 0, RF_MAXCOL);
207 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h,
208 &nXorBufs, NULL, overlappingPDAs, allocList);
209
210 /* create all the nodes at once */
211 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is
212 * generated for the
213 * failed pda */
214
215 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) +
216 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0);
217 /*
218 * XXX
219 *
220 * There's a bug with a complete stripe overwrite- that means 0 reads
221 * of old data, and the rest of the DAG generation code doesn't like
222 * that. A release is coming, and I don't wanna risk breaking a critical
223 * DAG generator, so here's what I'm gonna do- if there's no read nodes,
224 * I'm gonna fake there being a read node, and I'm gonna swap in a
225 * no-op node in its place (to make all the link-up code happy).
226 * This should be fixed at some point. --jimz
227 */
228 if (nRrdNodes == 0) {
229 nRrdNodes = 1;
230 rdnodesFaked = 1;
231 } else {
232 rdnodesFaked = 0;
233 }
234
235 blockNode = rf_AllocDAGNode(raidPtr);
236 blockNode->list_next = dag_h->nodes;
237 dag_h->nodes = blockNode;
238
239 commitNode = rf_AllocDAGNode(raidPtr);
240 commitNode->list_next = dag_h->nodes;
241 dag_h->nodes = commitNode;
242
243 unblockNode = rf_AllocDAGNode(raidPtr);
244 unblockNode->list_next = dag_h->nodes;
245 dag_h->nodes = unblockNode;
246
247 termNode = rf_AllocDAGNode(raidPtr);
248 termNode->list_next = dag_h->nodes;
249 dag_h->nodes = termNode;
250
251 xorNode = rf_AllocDAGNode(raidPtr);
252 xorNode->list_next = dag_h->nodes;
253 dag_h->nodes = xorNode;
254
255 wnpNode = rf_AllocDAGNode(raidPtr);
256 wnpNode->list_next = dag_h->nodes;
257 dag_h->nodes = wnpNode;
258
259 for (i = 0; i < nWndNodes; i++) {
260 tmpNode = rf_AllocDAGNode(raidPtr);
261 tmpNode->list_next = dag_h->nodes;
262 dag_h->nodes = tmpNode;
263 }
264 wndNodes = dag_h->nodes;
265
266 for (i = 0; i < nRrdNodes; i++) {
267 tmpNode = rf_AllocDAGNode(raidPtr);
268 tmpNode->list_next = dag_h->nodes;
269 dag_h->nodes = tmpNode;
270 }
271 rrdNodes = dag_h->nodes;
272
273 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0)
274 if (nfaults == 2) {
275 wnqNode = rf_AllocDAGNode(raidPtr);
276 wnqNode->list_next = dag_h->nodes;
277 dag_h->nodes = wnqNode;
278 } else {
279 wnqNode = NULL;
280 }
281 #endif
282
283 /* this dag can not commit until all rrd and xor Nodes have completed */
284 dag_h->numCommitNodes = 1;
285 dag_h->numCommits = 0;
286 dag_h->numSuccedents = 1;
287
288 RF_ASSERT(nRrdNodes > 0);
289 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
290 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList);
291 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
292 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList);
293 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
294 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList);
295 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc,
296 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList);
297 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1,
298 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList);
299
300 /*
301 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as
302 * the failed buffer, save a pointer to it so we can use it as the target
303 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if
304 * a buffer is the same size as the failed buffer, it must also be at the
305 * same alignment within the SU.
306 */
307 i = 0;
308 tmprrdNode = rrdNodes;
309 if (new_asm_h[0]) {
310 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo;
311 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed;
312 i++, pda = pda->next) {
313 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc,
314 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList);
315 RF_ASSERT(pda);
316 tmprrdNode->params[0].p = pda;
317 tmprrdNode->params[1].p = pda->bufPtr;
318 tmprrdNode->params[2].v = parityStripeID;
319 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru);
320 tmprrdNode = tmprrdNode->list_next;
321 }
322 }
323 /* i now equals the number of stripe units accessed in new_asm_h[0] */
324 /* Note that for tmprrdNode, this means a continuation from above, so no need to
325 assign it anything.. */
326 if (new_asm_h[1]) {
327 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo;
328 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed;
329 j++, pda = pda->next) {
330 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc,
331 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList);
332 RF_ASSERT(pda);
333 tmprrdNode->params[0].p = pda;
334 tmprrdNode->params[1].p = pda->bufPtr;
335 tmprrdNode->params[2].v = parityStripeID;
336 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru);
337 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector))
338 xorTargetBuf = pda->bufPtr;
339 tmprrdNode = tmprrdNode->list_next;
340 }
341 }
342 if (rdnodesFaked) {
343 /*
344 * This is where we'll init that fake noop read node
345 * (XXX should the wakeup func be different?)
346 */
347 /* node that rrdNodes will just be a single node... */
348 rf_InitNode(rrdNodes, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
349 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList);
350 }
351 /*
352 * Make a PDA for the parity unit. The parity PDA should start at
353 * the same offset into the SU as the failed PDA.
354 */
355 /* Danner comment: I don't think this copy is really necessary. We are
356 * in one of two cases here. (1) The entire failed unit is written.
357 * Then asmap->parityInfo will describe the entire parity. (2) We are
358 * only writing a subset of the failed unit and nothing else. Then the
359 * asmap->parityInfo describes the failed unit and the copy can also
360 * be avoided. */
361
362 parityPDA = rf_AllocPhysDiskAddr(raidPtr);
363 parityPDA->next = dag_h->pda_cleanup_list;
364 dag_h->pda_cleanup_list = parityPDA;
365 parityPDA->col = asmap->parityInfo->col;
366 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU)
367 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU);
368 parityPDA->numSector = failedPDA->numSector;
369
370 if (!xorTargetBuf) {
371 xorTargetBuf = rf_AllocBuffer(raidPtr, dag_h, rf_RaidAddressToByte(raidPtr, failedPDA->numSector));
372 }
373 /* init the Wnp node */
374 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc,
375 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList);
376 wnpNode->params[0].p = parityPDA;
377 wnpNode->params[1].p = xorTargetBuf;
378 wnpNode->params[2].v = parityStripeID;
379 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru);
380
381 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0)
382 /* fill in the Wnq Node */
383 if (nfaults == 2) {
384 {
385 parityPDA = RF_MallocAndAdd(sizeof(*parityPDA), allocList);
386 parityPDA->col = asmap->qInfo->col;
387 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU)
388 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU);
389 parityPDA->numSector = failedPDA->numSector;
390
391 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc,
392 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList);
393 wnqNode->params[0].p = parityPDA;
394 xorNode->results[1] = BUF_ALLOC(failedPDA->numSector);
395 wnqNode->params[1].p = xorNode->results[1];
396 wnqNode->params[2].v = parityStripeID;
397 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru);
398 }
399 }
400 #endif
401 /* fill in the Wnd nodes */
402 tmpwndNode = wndNodes;
403 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) {
404 if (pda == failedPDA) {
405 i--;
406 continue;
407 }
408 rf_InitNode(tmpwndNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc,
409 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList);
410 RF_ASSERT(pda);
411 tmpwndNode->params[0].p = pda;
412 tmpwndNode->params[1].p = pda->bufPtr;
413 tmpwndNode->params[2].v = parityStripeID;
414 tmpwndNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru);
415 tmpwndNode = tmpwndNode->list_next;
416 }
417
418 /* fill in the results of the xor node */
419 xorNode->results[0] = xorTargetBuf;
420
421 /* fill in the params of the xor node */
422
423 paramNum = 0;
424 if (rdnodesFaked == 0) {
425 tmprrdNode = rrdNodes;
426 for (i = 0; i < nRrdNodes; i++) {
427 /* all the Rrd nodes need to be xored together */
428 xorNode->params[paramNum++] = tmprrdNode->params[0];
429 xorNode->params[paramNum++] = tmprrdNode->params[1];
430 tmprrdNode = tmprrdNode->list_next;
431 }
432 }
433 tmpwndNode = wndNodes;
434 for (i = 0; i < nWndNodes; i++) {
435 /* any Wnd nodes that overlap the failed access need to be
436 * xored in */
437 if (overlappingPDAs[i]) {
438 pda = rf_AllocPhysDiskAddr(raidPtr);
439 memcpy((char *) pda, (char *) tmpwndNode->params[0].p, sizeof(RF_PhysDiskAddr_t));
440 /* add it into the pda_cleanup_list *after* the copy, TYVM */
441 pda->next = dag_h->pda_cleanup_list;
442 dag_h->pda_cleanup_list = pda;
443 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0);
444 xorNode->params[paramNum++].p = pda;
445 xorNode->params[paramNum++].p = pda->bufPtr;
446 }
447 tmpwndNode = tmpwndNode->list_next;
448 }
449
450 /*
451 * Install the failed PDA into the xor param list so that the
452 * new data gets xor'd in.
453 */
454 xorNode->params[paramNum++].p = failedPDA;
455 xorNode->params[paramNum++].p = failedPDA->bufPtr;
456
457 /*
458 * The last 2 params to the recovery xor node are always the failed
459 * PDA and the raidPtr. install the failedPDA even though we have just
460 * done so above. This allows us to use the same XOR function for both
461 * degraded reads and degraded writes.
462 */
463 xorNode->params[paramNum++].p = failedPDA;
464 xorNode->params[paramNum++].p = raidPtr;
465 RF_ASSERT(paramNum == 2 * nXorBufs + 2);
466
467 /*
468 * Code to link nodes begins here
469 */
470
471 /* link header to block node */
472 RF_ASSERT(blockNode->numAntecedents == 0);
473 dag_h->succedents[0] = blockNode;
474
475 /* link block node to rd nodes */
476 RF_ASSERT(blockNode->numSuccedents == nRrdNodes);
477 tmprrdNode = rrdNodes;
478 for (i = 0; i < nRrdNodes; i++) {
479 RF_ASSERT(tmprrdNode->numAntecedents == 1);
480 blockNode->succedents[i] = tmprrdNode;
481 tmprrdNode->antecedents[0] = blockNode;
482 tmprrdNode->antType[0] = rf_control;
483 tmprrdNode = tmprrdNode->list_next;
484 }
485
486 /* link read nodes to xor node */
487 RF_ASSERT(xorNode->numAntecedents == nRrdNodes);
488 tmprrdNode = rrdNodes;
489 for (i = 0; i < nRrdNodes; i++) {
490 RF_ASSERT(tmprrdNode->numSuccedents == 1);
491 tmprrdNode->succedents[0] = xorNode;
492 xorNode->antecedents[i] = tmprrdNode;
493 xorNode->antType[i] = rf_trueData;
494 tmprrdNode = tmprrdNode->list_next;
495 }
496
497 /* link xor node to commit node */
498 RF_ASSERT(xorNode->numSuccedents == 1);
499 RF_ASSERT(commitNode->numAntecedents == 1);
500 xorNode->succedents[0] = commitNode;
501 commitNode->antecedents[0] = xorNode;
502 commitNode->antType[0] = rf_control;
503
504 /* link commit node to wnd nodes */
505 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes);
506 tmpwndNode = wndNodes;
507 for (i = 0; i < nWndNodes; i++) {
508 RF_ASSERT(tmpwndNode->numAntecedents == 1);
509 commitNode->succedents[i] = tmpwndNode;
510 tmpwndNode->antecedents[0] = commitNode;
511 tmpwndNode->antType[0] = rf_control;
512 tmpwndNode = tmpwndNode->list_next;
513 }
514
515 /* link the commit node to wnp, wnq nodes */
516 RF_ASSERT(wnpNode->numAntecedents == 1);
517 commitNode->succedents[nWndNodes] = wnpNode;
518 wnpNode->antecedents[0] = commitNode;
519 wnpNode->antType[0] = rf_control;
520 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0)
521 if (nfaults == 2) {
522 RF_ASSERT(wnqNode->numAntecedents == 1);
523 commitNode->succedents[nWndNodes + 1] = wnqNode;
524 wnqNode->antecedents[0] = commitNode;
525 wnqNode->antType[0] = rf_control;
526 }
527 #endif
528 /* link write new data nodes to unblock node */
529 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults));
530 tmpwndNode = wndNodes;
531 for (i = 0; i < nWndNodes; i++) {
532 RF_ASSERT(tmpwndNode->numSuccedents == 1);
533 tmpwndNode->succedents[0] = unblockNode;
534 unblockNode->antecedents[i] = tmpwndNode;
535 unblockNode->antType[i] = rf_control;
536 tmpwndNode = tmpwndNode->list_next;
537 }
538
539 /* link write new parity node to unblock node */
540 RF_ASSERT(wnpNode->numSuccedents == 1);
541 wnpNode->succedents[0] = unblockNode;
542 unblockNode->antecedents[nWndNodes] = wnpNode;
543 unblockNode->antType[nWndNodes] = rf_control;
544
545 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0)
546 /* link write new q node to unblock node */
547 if (nfaults == 2) {
548 RF_ASSERT(wnqNode->numSuccedents == 1);
549 wnqNode->succedents[0] = unblockNode;
550 unblockNode->antecedents[nWndNodes + 1] = wnqNode;
551 unblockNode->antType[nWndNodes + 1] = rf_control;
552 }
553 #endif
554 /* link unblock node to term node */
555 RF_ASSERT(unblockNode->numSuccedents == 1);
556 RF_ASSERT(termNode->numAntecedents == 1);
557 RF_ASSERT(termNode->numSuccedents == 0);
558 unblockNode->succedents[0] = termNode;
559 termNode->antecedents[0] = unblockNode;
560 termNode->antType[0] = rf_control;
561 }
562 #define CONS_PDA(if,start,num) \
563 pda_p->col = asmap->if->col; \
564 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \
565 pda_p->numSector = num; \
566 pda_p->next = NULL; \
567 pda_p->bufPtr = BUF_ALLOC(num)
568 #if (RF_INCLUDE_RAID6 > 0) || (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0)
569 void
rf_WriteGenerateFailedAccessASMs(RF_Raid_t * raidPtr,RF_AccessStripeMap_t * asmap,RF_PhysDiskAddr_t ** pdap,int * nNodep,RF_PhysDiskAddr_t ** pqpdap,int * nPQNodep,RF_AllocListElem_t * allocList)570 rf_WriteGenerateFailedAccessASMs(
571 RF_Raid_t * raidPtr,
572 RF_AccessStripeMap_t * asmap,
573 RF_PhysDiskAddr_t ** pdap,
574 int *nNodep,
575 RF_PhysDiskAddr_t ** pqpdap,
576 int *nPQNodep,
577 RF_AllocListElem_t * allocList)
578 {
579 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
580 int PDAPerDisk, i;
581 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit;
582 int numDataCol = layoutPtr->numDataCol;
583 int state;
584 unsigned napdas;
585 RF_SectorNum_t fone_start, ftwo_start = 0;
586 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1];
587 RF_PhysDiskAddr_t *pda_p;
588 RF_RaidAddr_t sosAddr;
589
590 /* determine how many pda's we will have to generate per unaccess
591 * stripe. If there is only one failed data unit, it is one; if two,
592 * possibly two, depending whether they overlap. */
593
594 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector);
595
596 if (asmap->numDataFailed == 1) {
597 PDAPerDisk = 1;
598 state = 1;
599 *pqpdap = RF_MallocAndAdd(2 * sizeof(**pqpdap), allocList);
600 pda_p = *pqpdap;
601 /* build p */
602 CONS_PDA(parityInfo, fone_start, fone->numSector);
603 pda_p->type = RF_PDA_TYPE_PARITY;
604 pda_p++;
605 /* build q */
606 CONS_PDA(qInfo, fone_start, fone->numSector);
607 pda_p->type = RF_PDA_TYPE_Q;
608 } else {
609 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector);
610 if (fone->numSector + ftwo->numSector > secPerSU) {
611 PDAPerDisk = 1;
612 state = 2;
613 *pqpdap = RF_MallocAndAdd(2 * sizeof(**pqpdap),
614 allocList);
615 pda_p = *pqpdap;
616 CONS_PDA(parityInfo, 0, secPerSU);
617 pda_p->type = RF_PDA_TYPE_PARITY;
618 pda_p++;
619 CONS_PDA(qInfo, 0, secPerSU);
620 pda_p->type = RF_PDA_TYPE_Q;
621 } else {
622 PDAPerDisk = 2;
623 state = 3;
624 /* four of them, fone, then ftwo */
625 *pqpdap = RF_MallocAndAdd(4 * sizeof(*pqpdap),
626 allocList);
627 pda_p = *pqpdap;
628 CONS_PDA(parityInfo, fone_start, fone->numSector);
629 pda_p->type = RF_PDA_TYPE_PARITY;
630 pda_p++;
631 CONS_PDA(qInfo, fone_start, fone->numSector);
632 pda_p->type = RF_PDA_TYPE_Q;
633 pda_p++;
634 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector);
635 pda_p->type = RF_PDA_TYPE_PARITY;
636 pda_p++;
637 CONS_PDA(qInfo, ftwo_start, ftwo->numSector);
638 pda_p->type = RF_PDA_TYPE_Q;
639 }
640 }
641 /* figure out number of nonaccessed pda */
642 napdas = PDAPerDisk * (numDataCol - 2);
643 *nPQNodep = PDAPerDisk;
644
645 *nNodep = napdas;
646 if (napdas == 0)
647 return; /* short circuit */
648
649 /* allocate up our list of pda's */
650
651 pda_p = RF_MallocAndAdd(napdas * sizeof(*pda_p), allocList);
652 *pdap = pda_p;
653
654 /* linkem together */
655 for (i = 0; i < (napdas - 1); i++)
656 pda_p[i].next = pda_p + (i + 1);
657
658 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress);
659 for (i = 0; i < numDataCol; i++) {
660 if ((pda_p - (*pdap)) == napdas)
661 continue;
662 pda_p->type = RF_PDA_TYPE_DATA;
663 pda_p->raidAddress = sosAddr + (i * secPerSU);
664 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0);
665 /* skip over dead disks */
666 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->col].status))
667 continue;
668 switch (state) {
669 case 1: /* fone */
670 pda_p->numSector = fone->numSector;
671 pda_p->raidAddress += fone_start;
672 pda_p->startSector += fone_start;
673 pda_p->bufPtr = BUF_ALLOC(pda_p->numSector);
674 break;
675 case 2: /* full stripe */
676 pda_p->numSector = secPerSU;
677 pda_p->bufPtr = BUF_ALLOC(secPerSU);
678 break;
679 case 3: /* two slabs */
680 pda_p->numSector = fone->numSector;
681 pda_p->raidAddress += fone_start;
682 pda_p->startSector += fone_start;
683 pda_p->bufPtr = BUF_ALLOC(pda_p->numSector);
684 pda_p++;
685 pda_p->type = RF_PDA_TYPE_DATA;
686 pda_p->raidAddress = sosAddr + (i * secPerSU);
687 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0);
688 pda_p->numSector = ftwo->numSector;
689 pda_p->raidAddress += ftwo_start;
690 pda_p->startSector += ftwo_start;
691 pda_p->bufPtr = BUF_ALLOC(pda_p->numSector);
692 break;
693 default:
694 RF_PANIC();
695 }
696 pda_p++;
697 }
698
699 RF_ASSERT(pda_p - *pdap == napdas);
700 return;
701 }
702 #define DISK_NODE_PDA(node) ((node)->params[0].p)
703
704 #define DISK_NODE_PARAMS(_node_,_p_) \
705 (_node_).params[0].p = _p_ ; \
706 (_node_).params[1].p = (_p_)->bufPtr; \
707 (_node_).params[2].v = parityStripeID; \
708 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru)
709
710 void
rf_DoubleDegSmallWrite(RF_Raid_t * raidPtr,RF_AccessStripeMap_t * asmap,RF_DagHeader_t * dag_h,void * bp,RF_RaidAccessFlags_t flags,RF_AllocListElem_t * allocList,const char * redundantReadNodeName,const char * redundantWriteNodeName,const char * recoveryNodeName,void (* recovFunc)(RF_DagNode_t *))711 rf_DoubleDegSmallWrite(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap,
712 RF_DagHeader_t *dag_h, void *bp,
713 RF_RaidAccessFlags_t flags,
714 RF_AllocListElem_t *allocList,
715 const char *redundantReadNodeName,
716 const char *redundantWriteNodeName,
717 const char *recoveryNodeName,
718 void (*recovFunc) (RF_DagNode_t *))
719 {
720 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
721 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode,
722 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode;
723 RF_PhysDiskAddr_t *pda, *pqPDAs;
724 RF_PhysDiskAddr_t *npdas;
725 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i;
726 RF_ReconUnitNum_t which_ru;
727 int nPQNodes;
728 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru);
729
730 /* simple small write case - First part looks like a reconstruct-read
731 * of the failed data units. Then a write of all data units not
732 * failed. */
733
734
735 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \
736 * / -------PQ----- / \ \ Wud Wp WQ \ | /
737 * --Unblock- | T
738 *
739 * Rrd = read recovery data (potentially none) Wud = write user data
740 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q
741 * (could be two)
742 *
743 */
744
745 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList);
746
747 RF_ASSERT(asmap->numDataFailed == 1);
748
749 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed);
750 nReadNodes = nRrdNodes + 2 * nPQNodes;
751 nWriteNodes = nWudNodes + 2 * nPQNodes;
752 nNodes = 4 + nReadNodes + nWriteNodes;
753
754 nodes = RF_MallocAndAdd(nNodes * sizeof(*nodes), allocList);
755 blockNode = nodes;
756 unblockNode = blockNode + 1;
757 termNode = unblockNode + 1;
758 recoveryNode = termNode + 1;
759 rrdNodes = recoveryNode + 1;
760 rpNodes = rrdNodes + nRrdNodes;
761 rqNodes = rpNodes + nPQNodes;
762 wudNodes = rqNodes + nPQNodes;
763 wpNodes = wudNodes + nWudNodes;
764 wqNodes = wpNodes + nPQNodes;
765
766 dag_h->creator = "PQ_DDSimpleSmallWrite";
767 dag_h->numSuccedents = 1;
768 dag_h->succedents[0] = blockNode;
769 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList);
770 termNode->antecedents[0] = unblockNode;
771 termNode->antType[0] = rf_control;
772
773 /* init the block and unblock nodes */
774 /* The block node has all the read nodes as successors */
775 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList);
776 for (i = 0; i < nReadNodes; i++)
777 blockNode->succedents[i] = rrdNodes + i;
778
779 /* The unblock node has all the writes as successors */
780 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList);
781 for (i = 0; i < nWriteNodes; i++) {
782 unblockNode->antecedents[i] = wudNodes + i;
783 unblockNode->antType[i] = rf_control;
784 }
785 unblockNode->succedents[0] = termNode;
786
787 #define INIT_READ_NODE(node,name) \
788 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \
789 (node)->succedents[0] = recoveryNode; \
790 (node)->antecedents[0] = blockNode; \
791 (node)->antType[0] = rf_control;
792
793 /* build the read nodes */
794 pda = npdas;
795 for (i = 0; i < nRrdNodes; i++, pda = pda->next) {
796 INIT_READ_NODE(rrdNodes + i, "rrd");
797 DISK_NODE_PARAMS(rrdNodes[i], pda);
798 }
799
800 /* read redundancy pdas */
801 pda = pqPDAs;
802 INIT_READ_NODE(rpNodes, "Rp");
803 RF_ASSERT(pda);
804 DISK_NODE_PARAMS(rpNodes[0], pda);
805 pda++;
806 INIT_READ_NODE(rqNodes, redundantReadNodeName);
807 RF_ASSERT(pda);
808 DISK_NODE_PARAMS(rqNodes[0], pda);
809 if (nPQNodes == 2) {
810 pda++;
811 INIT_READ_NODE(rpNodes + 1, "Rp");
812 RF_ASSERT(pda);
813 DISK_NODE_PARAMS(rpNodes[1], pda);
814 pda++;
815 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName);
816 RF_ASSERT(pda);
817 DISK_NODE_PARAMS(rqNodes[1], pda);
818 }
819 /* the recovery node has all reads as precedessors and all writes as
820 * successors. It generates a result for every write P or write Q
821 * node. As parameters, it takes a pda per read and a pda per stripe
822 * of user data written. It also takes as the last params the raidPtr
823 * and asm. For results, it takes PDA for P & Q. */
824
825
826 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL,
827 nWriteNodes, /* succesors */
828 nReadNodes, /* preds */
829 nReadNodes + nWudNodes + 3, /* params */
830 2 * nPQNodes, /* results */
831 dag_h, recoveryNodeName, allocList);
832
833
834
835 for (i = 0; i < nReadNodes; i++) {
836 recoveryNode->antecedents[i] = rrdNodes + i;
837 recoveryNode->antType[i] = rf_control;
838 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i);
839 }
840 for (i = 0; i < nWudNodes; i++) {
841 recoveryNode->succedents[i] = wudNodes + i;
842 }
843 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0];
844 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr;
845 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap;
846
847 for (; i < nWriteNodes; i++)
848 recoveryNode->succedents[i] = wudNodes + i;
849
850 pda = pqPDAs;
851 recoveryNode->results[0] = pda;
852 pda++;
853 recoveryNode->results[1] = pda;
854 if (nPQNodes == 2) {
855 pda++;
856 recoveryNode->results[2] = pda;
857 pda++;
858 recoveryNode->results[3] = pda;
859 }
860 /* fill writes */
861 #define INIT_WRITE_NODE(node,name) \
862 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \
863 (node)->succedents[0] = unblockNode; \
864 (node)->antecedents[0] = recoveryNode; \
865 (node)->antType[0] = rf_control;
866
867 pda = asmap->physInfo;
868 for (i = 0; i < nWudNodes; i++) {
869 INIT_WRITE_NODE(wudNodes + i, "Wd");
870 DISK_NODE_PARAMS(wudNodes[i], pda);
871 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i);
872 pda = pda->next;
873 }
874 /* write redundancy pdas */
875 pda = pqPDAs;
876 INIT_WRITE_NODE(wpNodes, "Wp");
877 RF_ASSERT(pda);
878 DISK_NODE_PARAMS(wpNodes[0], pda);
879 pda++;
880 INIT_WRITE_NODE(wqNodes, "Wq");
881 RF_ASSERT(pda);
882 DISK_NODE_PARAMS(wqNodes[0], pda);
883 if (nPQNodes == 2) {
884 pda++;
885 INIT_WRITE_NODE(wpNodes + 1, "Wp");
886 RF_ASSERT(pda);
887 DISK_NODE_PARAMS(wpNodes[1], pda);
888 pda++;
889 INIT_WRITE_NODE(wqNodes + 1, "Wq");
890 RF_ASSERT(pda);
891 DISK_NODE_PARAMS(wqNodes[1], pda);
892 }
893 }
894 #endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */
895