rf_raid5.c revision 1.4.6.2 1 1.4.6.2 nathanw /* $NetBSD: rf_raid5.c,v 1.4.6.2 2001/11/14 19:15:52 nathanw Exp $ */
2 1.1 oster /*
3 1.1 oster * Copyright (c) 1995 Carnegie-Mellon University.
4 1.1 oster * All rights reserved.
5 1.1 oster *
6 1.1 oster * Author: Mark Holland
7 1.1 oster *
8 1.1 oster * Permission to use, copy, modify and distribute this software and
9 1.1 oster * its documentation is hereby granted, provided that both the copyright
10 1.1 oster * notice and this permission notice appear in all copies of the
11 1.1 oster * software, derivative works or modified versions, and any portions
12 1.1 oster * thereof, and that both notices appear in supporting documentation.
13 1.1 oster *
14 1.1 oster * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
15 1.1 oster * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
16 1.1 oster * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
17 1.1 oster *
18 1.1 oster * Carnegie Mellon requests users of this software to return to
19 1.1 oster *
20 1.1 oster * Software Distribution Coordinator or Software.Distribution (at) CS.CMU.EDU
21 1.1 oster * School of Computer Science
22 1.1 oster * Carnegie Mellon University
23 1.1 oster * Pittsburgh PA 15213-3890
24 1.1 oster *
25 1.1 oster * any improvements or extensions that they make and grant Carnegie the
26 1.1 oster * rights to redistribute these changes.
27 1.1 oster */
28 1.1 oster
29 1.1 oster /******************************************************************************
30 1.1 oster *
31 1.1 oster * rf_raid5.c -- implements RAID Level 5
32 1.1 oster *
33 1.1 oster *****************************************************************************/
34 1.4.6.2 nathanw
35 1.4.6.2 nathanw #include <sys/cdefs.h>
36 1.4.6.2 nathanw __KERNEL_RCSID(0, "$NetBSD: rf_raid5.c,v 1.4.6.2 2001/11/14 19:15:52 nathanw Exp $");
37 1.1 oster
38 1.4.6.1 nathanw #include <dev/raidframe/raidframevar.h>
39 1.4.6.1 nathanw
40 1.1 oster #include "rf_raid.h"
41 1.1 oster #include "rf_raid5.h"
42 1.1 oster #include "rf_dag.h"
43 1.1 oster #include "rf_dagffrd.h"
44 1.1 oster #include "rf_dagffwr.h"
45 1.1 oster #include "rf_dagdegrd.h"
46 1.1 oster #include "rf_dagdegwr.h"
47 1.1 oster #include "rf_dagutils.h"
48 1.1 oster #include "rf_general.h"
49 1.1 oster #include "rf_map.h"
50 1.1 oster #include "rf_utils.h"
51 1.1 oster
52 1.1 oster typedef struct RF_Raid5ConfigInfo_s {
53 1.3 oster RF_RowCol_t **stripeIdentifier; /* filled in at config time and used
54 1.3 oster * by IdentifyStripe */
55 1.3 oster } RF_Raid5ConfigInfo_t;
56 1.3 oster
57 1.3 oster int
58 1.3 oster rf_ConfigureRAID5(
59 1.3 oster RF_ShutdownList_t ** listp,
60 1.3 oster RF_Raid_t * raidPtr,
61 1.3 oster RF_Config_t * cfgPtr)
62 1.3 oster {
63 1.3 oster RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
64 1.3 oster RF_Raid5ConfigInfo_t *info;
65 1.3 oster RF_RowCol_t i, j, startdisk;
66 1.3 oster
67 1.3 oster /* create a RAID level 5 configuration structure */
68 1.3 oster RF_MallocAndAdd(info, sizeof(RF_Raid5ConfigInfo_t), (RF_Raid5ConfigInfo_t *), raidPtr->cleanupList);
69 1.3 oster if (info == NULL)
70 1.3 oster return (ENOMEM);
71 1.3 oster layoutPtr->layoutSpecificInfo = (void *) info;
72 1.3 oster
73 1.3 oster RF_ASSERT(raidPtr->numRow == 1);
74 1.3 oster
75 1.3 oster /* the stripe identifier must identify the disks in each stripe, IN
76 1.3 oster * THE ORDER THAT THEY APPEAR IN THE STRIPE. */
77 1.3 oster info->stripeIdentifier = rf_make_2d_array(raidPtr->numCol, raidPtr->numCol, raidPtr->cleanupList);
78 1.3 oster if (info->stripeIdentifier == NULL)
79 1.3 oster return (ENOMEM);
80 1.3 oster startdisk = 0;
81 1.3 oster for (i = 0; i < raidPtr->numCol; i++) {
82 1.3 oster for (j = 0; j < raidPtr->numCol; j++) {
83 1.3 oster info->stripeIdentifier[i][j] = (startdisk + j) % raidPtr->numCol;
84 1.3 oster }
85 1.3 oster if ((--startdisk) < 0)
86 1.3 oster startdisk = raidPtr->numCol - 1;
87 1.3 oster }
88 1.1 oster
89 1.3 oster /* fill in the remaining layout parameters */
90 1.3 oster layoutPtr->numStripe = layoutPtr->stripeUnitsPerDisk;
91 1.3 oster layoutPtr->bytesPerStripeUnit = layoutPtr->sectorsPerStripeUnit << raidPtr->logBytesPerSector;
92 1.3 oster layoutPtr->numDataCol = raidPtr->numCol - 1;
93 1.3 oster layoutPtr->dataSectorsPerStripe = layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit;
94 1.3 oster layoutPtr->numParityCol = 1;
95 1.3 oster layoutPtr->dataStripeUnitsPerDisk = layoutPtr->stripeUnitsPerDisk;
96 1.1 oster
97 1.3 oster raidPtr->totalSectors = layoutPtr->stripeUnitsPerDisk * layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit;
98 1.1 oster
99 1.3 oster return (0);
100 1.1 oster }
101 1.1 oster
102 1.3 oster int
103 1.3 oster rf_GetDefaultNumFloatingReconBuffersRAID5(RF_Raid_t * raidPtr)
104 1.1 oster {
105 1.3 oster return (20);
106 1.1 oster }
107 1.1 oster
108 1.3 oster RF_HeadSepLimit_t
109 1.3 oster rf_GetDefaultHeadSepLimitRAID5(RF_Raid_t * raidPtr)
110 1.1 oster {
111 1.3 oster return (10);
112 1.1 oster }
113 1.1 oster #if !defined(__NetBSD__) && !defined(_KERNEL)
114 1.1 oster /* not currently used */
115 1.3 oster int
116 1.3 oster rf_ShutdownRAID5(RF_Raid_t * raidPtr)
117 1.1 oster {
118 1.3 oster return (0);
119 1.1 oster }
120 1.1 oster #endif
121 1.1 oster
122 1.3 oster void
123 1.3 oster rf_MapSectorRAID5(
124 1.3 oster RF_Raid_t * raidPtr,
125 1.3 oster RF_RaidAddr_t raidSector,
126 1.3 oster RF_RowCol_t * row,
127 1.3 oster RF_RowCol_t * col,
128 1.3 oster RF_SectorNum_t * diskSector,
129 1.3 oster int remap)
130 1.1 oster {
131 1.3 oster RF_StripeNum_t SUID = raidSector / raidPtr->Layout.sectorsPerStripeUnit;
132 1.3 oster *row = 0;
133 1.3 oster *col = (SUID % raidPtr->numCol);
134 1.3 oster *diskSector = (SUID / (raidPtr->Layout.numDataCol)) * raidPtr->Layout.sectorsPerStripeUnit +
135 1.3 oster (raidSector % raidPtr->Layout.sectorsPerStripeUnit);
136 1.1 oster }
137 1.1 oster
138 1.3 oster void
139 1.3 oster rf_MapParityRAID5(
140 1.3 oster RF_Raid_t * raidPtr,
141 1.3 oster RF_RaidAddr_t raidSector,
142 1.3 oster RF_RowCol_t * row,
143 1.3 oster RF_RowCol_t * col,
144 1.3 oster RF_SectorNum_t * diskSector,
145 1.3 oster int remap)
146 1.1 oster {
147 1.3 oster RF_StripeNum_t SUID = raidSector / raidPtr->Layout.sectorsPerStripeUnit;
148 1.3 oster
149 1.3 oster *row = 0;
150 1.3 oster *col = raidPtr->Layout.numDataCol - (SUID / raidPtr->Layout.numDataCol) % raidPtr->numCol;
151 1.3 oster *diskSector = (SUID / (raidPtr->Layout.numDataCol)) * raidPtr->Layout.sectorsPerStripeUnit +
152 1.3 oster (raidSector % raidPtr->Layout.sectorsPerStripeUnit);
153 1.1 oster }
154 1.1 oster
155 1.3 oster void
156 1.3 oster rf_IdentifyStripeRAID5(
157 1.3 oster RF_Raid_t * raidPtr,
158 1.3 oster RF_RaidAddr_t addr,
159 1.3 oster RF_RowCol_t ** diskids,
160 1.3 oster RF_RowCol_t * outRow)
161 1.1 oster {
162 1.3 oster RF_StripeNum_t stripeID = rf_RaidAddressToStripeID(&raidPtr->Layout, addr);
163 1.3 oster RF_Raid5ConfigInfo_t *info = (RF_Raid5ConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo;
164 1.1 oster
165 1.3 oster *outRow = 0;
166 1.3 oster *diskids = info->stripeIdentifier[stripeID % raidPtr->numCol];
167 1.1 oster }
168 1.1 oster
169 1.3 oster void
170 1.3 oster rf_MapSIDToPSIDRAID5(
171 1.3 oster RF_RaidLayout_t * layoutPtr,
172 1.3 oster RF_StripeNum_t stripeID,
173 1.3 oster RF_StripeNum_t * psID,
174 1.3 oster RF_ReconUnitNum_t * which_ru)
175 1.1 oster {
176 1.3 oster *which_ru = 0;
177 1.3 oster *psID = stripeID;
178 1.1 oster }
179 1.1 oster /* select an algorithm for performing an access. Returns two pointers,
180 1.1 oster * one to a function that will return information about the DAG, and
181 1.1 oster * another to a function that will create the dag.
182 1.1 oster */
183 1.3 oster void
184 1.3 oster rf_RaidFiveDagSelect(
185 1.3 oster RF_Raid_t * raidPtr,
186 1.3 oster RF_IoType_t type,
187 1.3 oster RF_AccessStripeMap_t * asmap,
188 1.3 oster RF_VoidFuncPtr * createFunc)
189 1.1 oster {
190 1.3 oster RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
191 1.3 oster RF_PhysDiskAddr_t *failedPDA = NULL;
192 1.3 oster RF_RowCol_t frow, fcol;
193 1.3 oster RF_RowStatus_t rstat;
194 1.3 oster int prior_recon;
195 1.3 oster
196 1.3 oster RF_ASSERT(RF_IO_IS_R_OR_W(type));
197 1.3 oster
198 1.3 oster if (asmap->numDataFailed + asmap->numParityFailed > 1) {
199 1.3 oster RF_ERRORMSG("Multiple disks failed in a single group! Aborting I/O operation.\n");
200 1.3 oster /* *infoFunc = */ *createFunc = NULL;
201 1.3 oster return;
202 1.3 oster } else
203 1.3 oster if (asmap->numDataFailed + asmap->numParityFailed == 1) {
204 1.3 oster
205 1.3 oster /* if under recon & already reconstructed, redirect
206 1.3 oster * the access to the spare drive and eliminate the
207 1.3 oster * failure indication */
208 1.3 oster failedPDA = asmap->failedPDAs[0];
209 1.3 oster frow = failedPDA->row;
210 1.3 oster fcol = failedPDA->col;
211 1.3 oster rstat = raidPtr->status[failedPDA->row];
212 1.3 oster prior_recon = (rstat == rf_rs_reconfigured) || (
213 1.3 oster (rstat == rf_rs_reconstructing) ?
214 1.3 oster rf_CheckRUReconstructed(raidPtr->reconControl[frow]->reconMap, failedPDA->startSector) : 0
215 1.3 oster );
216 1.3 oster if (prior_recon) {
217 1.3 oster RF_RowCol_t or = failedPDA->row, oc = failedPDA->col;
218 1.3 oster RF_SectorNum_t oo = failedPDA->startSector;
219 1.3 oster
220 1.3 oster if (layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) { /* redirect to dist
221 1.3 oster * spare space */
222 1.3 oster
223 1.3 oster if (failedPDA == asmap->parityInfo) {
224 1.3 oster
225 1.3 oster /* parity has failed */
226 1.3 oster (layoutPtr->map->MapParity) (raidPtr, failedPDA->raidAddress, &failedPDA->row,
227 1.3 oster &failedPDA->col, &failedPDA->startSector, RF_REMAP);
228 1.3 oster
229 1.3 oster if (asmap->parityInfo->next) { /* redir 2nd component,
230 1.3 oster * if any */
231 1.3 oster RF_PhysDiskAddr_t *p = asmap->parityInfo->next;
232 1.3 oster RF_SectorNum_t SUoffs = p->startSector % layoutPtr->sectorsPerStripeUnit;
233 1.3 oster p->row = failedPDA->row;
234 1.3 oster p->col = failedPDA->col;
235 1.3 oster p->startSector = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, failedPDA->startSector) +
236 1.3 oster SUoffs; /* cheating:
237 1.3 oster * startSector is not
238 1.3 oster * really a RAID address */
239 1.3 oster }
240 1.3 oster } else
241 1.3 oster if (asmap->parityInfo->next && failedPDA == asmap->parityInfo->next) {
242 1.3 oster RF_ASSERT(0); /* should not ever
243 1.3 oster * happen */
244 1.3 oster } else {
245 1.3 oster
246 1.3 oster /* data has failed */
247 1.3 oster (layoutPtr->map->MapSector) (raidPtr, failedPDA->raidAddress, &failedPDA->row,
248 1.3 oster &failedPDA->col, &failedPDA->startSector, RF_REMAP);
249 1.3 oster
250 1.3 oster }
251 1.3 oster
252 1.3 oster } else { /* redirect to dedicated spare
253 1.3 oster * space */
254 1.3 oster
255 1.3 oster failedPDA->row = raidPtr->Disks[frow][fcol].spareRow;
256 1.3 oster failedPDA->col = raidPtr->Disks[frow][fcol].spareCol;
257 1.3 oster
258 1.3 oster /* the parity may have two distinct
259 1.3 oster * components, both of which may need
260 1.3 oster * to be redirected */
261 1.3 oster if (asmap->parityInfo->next) {
262 1.3 oster if (failedPDA == asmap->parityInfo) {
263 1.3 oster failedPDA->next->row = failedPDA->row;
264 1.3 oster failedPDA->next->col = failedPDA->col;
265 1.3 oster } else
266 1.3 oster if (failedPDA == asmap->parityInfo->next) { /* paranoid: should
267 1.3 oster * never occur */
268 1.3 oster asmap->parityInfo->row = failedPDA->row;
269 1.3 oster asmap->parityInfo->col = failedPDA->col;
270 1.3 oster }
271 1.3 oster }
272 1.3 oster }
273 1.3 oster
274 1.3 oster RF_ASSERT(failedPDA->col != -1);
275 1.3 oster
276 1.3 oster if (rf_dagDebug || rf_mapDebug) {
277 1.4 oster printf("raid%d: Redirected type '%c' r %d c %d o %ld -> r %d c %d o %ld\n",
278 1.4 oster raidPtr->raidid, type, or, oc,
279 1.4 oster (long) oo, failedPDA->row,
280 1.4 oster failedPDA->col,
281 1.4 oster (long) failedPDA->startSector);
282 1.3 oster }
283 1.3 oster asmap->numDataFailed = asmap->numParityFailed = 0;
284 1.3 oster }
285 1.3 oster }
286 1.3 oster /* all dags begin/end with block/unblock node therefore, hdrSucc &
287 1.3 oster * termAnt counts should always be 1 also, these counts should not be
288 1.3 oster * visible outside dag creation routines - manipulating the counts
289 1.3 oster * here should be removed */
290 1.3 oster if (type == RF_IO_TYPE_READ) {
291 1.3 oster if (asmap->numDataFailed == 0)
292 1.3 oster *createFunc = (RF_VoidFuncPtr) rf_CreateFaultFreeReadDAG;
293 1.3 oster else
294 1.3 oster *createFunc = (RF_VoidFuncPtr) rf_CreateRaidFiveDegradedReadDAG;
295 1.1 oster } else {
296 1.1 oster
297 1.1 oster
298 1.3 oster /* if mirroring, always use large writes. If the access
299 1.3 oster * requires two distinct parity updates, always do a small
300 1.3 oster * write. If the stripe contains a failure but the access
301 1.3 oster * does not, do a small write. The first conditional
302 1.3 oster * (numStripeUnitsAccessed <= numDataCol/2) uses a
303 1.3 oster * less-than-or-equal rather than just a less-than because
304 1.3 oster * when G is 3 or 4, numDataCol/2 is 1, and I want
305 1.3 oster * single-stripe-unit updates to use just one disk. */
306 1.3 oster if ((asmap->numDataFailed + asmap->numParityFailed) == 0) {
307 1.3 oster if (rf_suppressLocksAndLargeWrites ||
308 1.3 oster (((asmap->numStripeUnitsAccessed <= (layoutPtr->numDataCol / 2)) && (layoutPtr->numDataCol != 1)) ||
309 1.3 oster (asmap->parityInfo->next != NULL) || rf_CheckStripeForFailures(raidPtr, asmap))) {
310 1.3 oster *createFunc = (RF_VoidFuncPtr) rf_CreateSmallWriteDAG;
311 1.3 oster } else
312 1.3 oster *createFunc = (RF_VoidFuncPtr) rf_CreateLargeWriteDAG;
313 1.3 oster } else {
314 1.3 oster if (asmap->numParityFailed == 1)
315 1.3 oster *createFunc = (RF_VoidFuncPtr) rf_CreateNonRedundantWriteDAG;
316 1.3 oster else
317 1.3 oster if (asmap->numStripeUnitsAccessed != 1 && failedPDA->numSector != layoutPtr->sectorsPerStripeUnit)
318 1.3 oster *createFunc = NULL;
319 1.3 oster else
320 1.3 oster *createFunc = (RF_VoidFuncPtr) rf_CreateDegradedWriteDAG;
321 1.3 oster }
322 1.1 oster }
323 1.1 oster }
324