1 /* $NetBSD: rf_dagfuncs.c,v 1.7 2001/02/03 12:51:10 mrg Exp $ */ 2 /* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: Mark Holland, William V. Courtright II 7 * 8 * Permission to use, copy, modify and distribute this software and 9 * its documentation is hereby granted, provided that both the copyright 10 * notice and this permission notice appear in all copies of the 11 * software, derivative works or modified versions, and any portions 12 * thereof, and that both notices appear in supporting documentation. 13 * 14 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 15 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 16 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 17 * 18 * Carnegie Mellon requests users of this software to return to 19 * 20 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 21 * School of Computer Science 22 * Carnegie Mellon University 23 * Pittsburgh PA 15213-3890 24 * 25 * any improvements or extensions that they make and grant Carnegie the 26 * rights to redistribute these changes. 27 */ 28 29 /* 30 * dagfuncs.c -- DAG node execution routines 31 * 32 * Rules: 33 * 1. Every DAG execution function must eventually cause node->status to 34 * get set to "good" or "bad", and "FinishNode" to be called. In the 35 * case of nodes that complete immediately (xor, NullNodeFunc, etc), 36 * the node execution function can do these two things directly. In 37 * the case of nodes that have to wait for some event (a disk read to 38 * complete, a lock to be released, etc) to occur before they can 39 * complete, this is typically achieved by having whatever module 40 * is doing the operation call GenericWakeupFunc upon completion. 41 * 2. DAG execution functions should check the status in the DAG header 42 * and NOP out their operations if the status is not "enable". However, 43 * execution functions that release resources must be sure to release 44 * them even when they NOP out the function that would use them. 45 * Functions that acquire resources should go ahead and acquire them 46 * even when they NOP, so that a downstream release node will not have 47 * to check to find out whether or not the acquire was suppressed. 48 */ 49 50 #include <sys/param.h> 51 #include <sys/ioctl.h> 52 53 #include "rf_archs.h" 54 #include "rf_raid.h" 55 #include "rf_dag.h" 56 #include "rf_layout.h" 57 #include "rf_etimer.h" 58 #include "rf_acctrace.h" 59 #include "rf_diskqueue.h" 60 #include "rf_dagfuncs.h" 61 #include "rf_general.h" 62 #include "rf_engine.h" 63 #include "rf_dagutils.h" 64 65 #include "rf_kintf.h" 66 67 #if RF_INCLUDE_PARITYLOGGING > 0 68 #include "rf_paritylog.h" 69 #endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 70 71 int (*rf_DiskReadFunc) (RF_DagNode_t *); 72 int (*rf_DiskWriteFunc) (RF_DagNode_t *); 73 int (*rf_DiskReadUndoFunc) (RF_DagNode_t *); 74 int (*rf_DiskWriteUndoFunc) (RF_DagNode_t *); 75 int (*rf_DiskUnlockFunc) (RF_DagNode_t *); 76 int (*rf_DiskUnlockUndoFunc) (RF_DagNode_t *); 77 int (*rf_RegularXorUndoFunc) (RF_DagNode_t *); 78 int (*rf_SimpleXorUndoFunc) (RF_DagNode_t *); 79 int (*rf_RecoveryXorUndoFunc) (RF_DagNode_t *); 80 81 /***************************************************************************************** 82 * main (only) configuration routine for this module 83 ****************************************************************************************/ 84 int 85 rf_ConfigureDAGFuncs(listp) 86 RF_ShutdownList_t **listp; 87 { 88 RF_ASSERT(((sizeof(long) == 8) && RF_LONGSHIFT == 3) || ((sizeof(long) == 4) && RF_LONGSHIFT == 2)); 89 rf_DiskReadFunc = rf_DiskReadFuncForThreads; 90 rf_DiskReadUndoFunc = rf_DiskUndoFunc; 91 rf_DiskWriteFunc = rf_DiskWriteFuncForThreads; 92 rf_DiskWriteUndoFunc = rf_DiskUndoFunc; 93 rf_DiskUnlockFunc = rf_DiskUnlockFuncForThreads; 94 rf_DiskUnlockUndoFunc = rf_NullNodeUndoFunc; 95 rf_RegularXorUndoFunc = rf_NullNodeUndoFunc; 96 rf_SimpleXorUndoFunc = rf_NullNodeUndoFunc; 97 rf_RecoveryXorUndoFunc = rf_NullNodeUndoFunc; 98 return (0); 99 } 100 101 102 103 /***************************************************************************************** 104 * the execution function associated with a terminate node 105 ****************************************************************************************/ 106 int 107 rf_TerminateFunc(node) 108 RF_DagNode_t *node; 109 { 110 RF_ASSERT(node->dagHdr->numCommits == node->dagHdr->numCommitNodes); 111 node->status = rf_good; 112 return (rf_FinishNode(node, RF_THREAD_CONTEXT)); 113 } 114 115 int 116 rf_TerminateUndoFunc(node) 117 RF_DagNode_t *node; 118 { 119 return (0); 120 } 121 122 123 /***************************************************************************************** 124 * execution functions associated with a mirror node 125 * 126 * parameters: 127 * 128 * 0 - physical disk addres of data 129 * 1 - buffer for holding read data 130 * 2 - parity stripe ID 131 * 3 - flags 132 * 4 - physical disk address of mirror (parity) 133 * 134 ****************************************************************************************/ 135 136 int 137 rf_DiskReadMirrorIdleFunc(node) 138 RF_DagNode_t *node; 139 { 140 /* select the mirror copy with the shortest queue and fill in node 141 * parameters with physical disk address */ 142 143 rf_SelectMirrorDiskIdle(node); 144 return (rf_DiskReadFunc(node)); 145 } 146 147 int 148 rf_DiskReadMirrorPartitionFunc(node) 149 RF_DagNode_t *node; 150 { 151 /* select the mirror copy with the shortest queue and fill in node 152 * parameters with physical disk address */ 153 154 rf_SelectMirrorDiskPartition(node); 155 return (rf_DiskReadFunc(node)); 156 } 157 158 int 159 rf_DiskReadMirrorUndoFunc(node) 160 RF_DagNode_t *node; 161 { 162 return (0); 163 } 164 165 166 167 #if RF_INCLUDE_PARITYLOGGING > 0 168 /***************************************************************************************** 169 * the execution function associated with a parity log update node 170 ****************************************************************************************/ 171 int 172 rf_ParityLogUpdateFunc(node) 173 RF_DagNode_t *node; 174 { 175 RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p; 176 caddr_t buf = (caddr_t) node->params[1].p; 177 RF_ParityLogData_t *logData; 178 RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec; 179 RF_Etimer_t timer; 180 181 if (node->dagHdr->status == rf_enable) { 182 RF_ETIMER_START(timer); 183 logData = rf_CreateParityLogData(RF_UPDATE, pda, buf, 184 (RF_Raid_t *) (node->dagHdr->raidPtr), 185 node->wakeFunc, (void *) node, 186 node->dagHdr->tracerec, timer); 187 if (logData) 188 rf_ParityLogAppend(logData, RF_FALSE, NULL, RF_FALSE); 189 else { 190 RF_ETIMER_STOP(timer); 191 RF_ETIMER_EVAL(timer); 192 tracerec->plog_us += RF_ETIMER_VAL_US(timer); 193 (node->wakeFunc) (node, ENOMEM); 194 } 195 } 196 return (0); 197 } 198 199 200 /***************************************************************************************** 201 * the execution function associated with a parity log overwrite node 202 ****************************************************************************************/ 203 int 204 rf_ParityLogOverwriteFunc(node) 205 RF_DagNode_t *node; 206 { 207 RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p; 208 caddr_t buf = (caddr_t) node->params[1].p; 209 RF_ParityLogData_t *logData; 210 RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec; 211 RF_Etimer_t timer; 212 213 if (node->dagHdr->status == rf_enable) { 214 RF_ETIMER_START(timer); 215 logData = rf_CreateParityLogData(RF_OVERWRITE, pda, buf, (RF_Raid_t *) (node->dagHdr->raidPtr), 216 node->wakeFunc, (void *) node, node->dagHdr->tracerec, timer); 217 if (logData) 218 rf_ParityLogAppend(logData, RF_FALSE, NULL, RF_FALSE); 219 else { 220 RF_ETIMER_STOP(timer); 221 RF_ETIMER_EVAL(timer); 222 tracerec->plog_us += RF_ETIMER_VAL_US(timer); 223 (node->wakeFunc) (node, ENOMEM); 224 } 225 } 226 return (0); 227 } 228 #else /* RF_INCLUDE_PARITYLOGGING > 0 */ 229 230 int 231 rf_ParityLogUpdateFunc(node) 232 RF_DagNode_t *node; 233 { 234 return (0); 235 } 236 int 237 rf_ParityLogOverwriteFunc(node) 238 RF_DagNode_t *node; 239 { 240 return (0); 241 } 242 #endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 243 244 int 245 rf_ParityLogUpdateUndoFunc(node) 246 RF_DagNode_t *node; 247 { 248 return (0); 249 } 250 251 int 252 rf_ParityLogOverwriteUndoFunc(node) 253 RF_DagNode_t *node; 254 { 255 return (0); 256 } 257 /***************************************************************************************** 258 * the execution function associated with a NOP node 259 ****************************************************************************************/ 260 int 261 rf_NullNodeFunc(node) 262 RF_DagNode_t *node; 263 { 264 node->status = rf_good; 265 return (rf_FinishNode(node, RF_THREAD_CONTEXT)); 266 } 267 268 int 269 rf_NullNodeUndoFunc(node) 270 RF_DagNode_t *node; 271 { 272 node->status = rf_undone; 273 return (rf_FinishNode(node, RF_THREAD_CONTEXT)); 274 } 275 276 277 /***************************************************************************************** 278 * the execution function associated with a disk-read node 279 ****************************************************************************************/ 280 int 281 rf_DiskReadFuncForThreads(node) 282 RF_DagNode_t *node; 283 { 284 RF_DiskQueueData_t *req; 285 RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p; 286 caddr_t buf = (caddr_t) node->params[1].p; 287 RF_StripeNum_t parityStripeID = (RF_StripeNum_t) node->params[2].v; 288 unsigned priority = RF_EXTRACT_PRIORITY(node->params[3].v); 289 unsigned lock = RF_EXTRACT_LOCK_FLAG(node->params[3].v); 290 unsigned unlock = RF_EXTRACT_UNLOCK_FLAG(node->params[3].v); 291 unsigned which_ru = RF_EXTRACT_RU(node->params[3].v); 292 RF_DiskQueueDataFlags_t flags = 0; 293 RF_IoType_t iotype = (node->dagHdr->status == rf_enable) ? RF_IO_TYPE_READ : RF_IO_TYPE_NOP; 294 RF_DiskQueue_t **dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues; 295 void *b_proc = NULL; 296 297 if (node->dagHdr->bp) 298 b_proc = (void *) ((struct buf *) node->dagHdr->bp)->b_proc; 299 300 RF_ASSERT(!(lock && unlock)); 301 flags |= (lock) ? RF_LOCK_DISK_QUEUE : 0; 302 flags |= (unlock) ? RF_UNLOCK_DISK_QUEUE : 0; 303 304 req = rf_CreateDiskQueueData(iotype, pda->startSector, pda->numSector, 305 buf, parityStripeID, which_ru, 306 (int (*) (void *, int)) node->wakeFunc, 307 node, NULL, node->dagHdr->tracerec, 308 (void *) (node->dagHdr->raidPtr), flags, b_proc); 309 if (!req) { 310 (node->wakeFunc) (node, ENOMEM); 311 } else { 312 node->dagFuncData = (void *) req; 313 rf_DiskIOEnqueue(&(dqs[pda->row][pda->col]), req, priority); 314 } 315 return (0); 316 } 317 318 319 /***************************************************************************************** 320 * the execution function associated with a disk-write node 321 ****************************************************************************************/ 322 int 323 rf_DiskWriteFuncForThreads(node) 324 RF_DagNode_t *node; 325 { 326 RF_DiskQueueData_t *req; 327 RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p; 328 caddr_t buf = (caddr_t) node->params[1].p; 329 RF_StripeNum_t parityStripeID = (RF_StripeNum_t) node->params[2].v; 330 unsigned priority = RF_EXTRACT_PRIORITY(node->params[3].v); 331 unsigned lock = RF_EXTRACT_LOCK_FLAG(node->params[3].v); 332 unsigned unlock = RF_EXTRACT_UNLOCK_FLAG(node->params[3].v); 333 unsigned which_ru = RF_EXTRACT_RU(node->params[3].v); 334 RF_DiskQueueDataFlags_t flags = 0; 335 RF_IoType_t iotype = (node->dagHdr->status == rf_enable) ? RF_IO_TYPE_WRITE : RF_IO_TYPE_NOP; 336 RF_DiskQueue_t **dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues; 337 void *b_proc = NULL; 338 339 if (node->dagHdr->bp) 340 b_proc = (void *) ((struct buf *) node->dagHdr->bp)->b_proc; 341 342 /* normal processing (rollaway or forward recovery) begins here */ 343 RF_ASSERT(!(lock && unlock)); 344 flags |= (lock) ? RF_LOCK_DISK_QUEUE : 0; 345 flags |= (unlock) ? RF_UNLOCK_DISK_QUEUE : 0; 346 req = rf_CreateDiskQueueData(iotype, pda->startSector, pda->numSector, 347 buf, parityStripeID, which_ru, 348 (int (*) (void *, int)) node->wakeFunc, 349 (void *) node, NULL, 350 node->dagHdr->tracerec, 351 (void *) (node->dagHdr->raidPtr), 352 flags, b_proc); 353 354 if (!req) { 355 (node->wakeFunc) (node, ENOMEM); 356 } else { 357 node->dagFuncData = (void *) req; 358 rf_DiskIOEnqueue(&(dqs[pda->row][pda->col]), req, priority); 359 } 360 361 return (0); 362 } 363 /***************************************************************************************** 364 * the undo function for disk nodes 365 * Note: this is not a proper undo of a write node, only locks are released. 366 * old data is not restored to disk! 367 ****************************************************************************************/ 368 int 369 rf_DiskUndoFunc(node) 370 RF_DagNode_t *node; 371 { 372 RF_DiskQueueData_t *req; 373 RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p; 374 RF_DiskQueue_t **dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues; 375 376 req = rf_CreateDiskQueueData(RF_IO_TYPE_NOP, 377 0L, 0, NULL, 0L, 0, 378 (int (*) (void *, int)) node->wakeFunc, 379 (void *) node, 380 NULL, node->dagHdr->tracerec, 381 (void *) (node->dagHdr->raidPtr), 382 RF_UNLOCK_DISK_QUEUE, NULL); 383 if (!req) 384 (node->wakeFunc) (node, ENOMEM); 385 else { 386 node->dagFuncData = (void *) req; 387 rf_DiskIOEnqueue(&(dqs[pda->row][pda->col]), req, RF_IO_NORMAL_PRIORITY); 388 } 389 390 return (0); 391 } 392 /***************************************************************************************** 393 * the execution function associated with an "unlock disk queue" node 394 ****************************************************************************************/ 395 int 396 rf_DiskUnlockFuncForThreads(node) 397 RF_DagNode_t *node; 398 { 399 RF_DiskQueueData_t *req; 400 RF_PhysDiskAddr_t *pda = (RF_PhysDiskAddr_t *) node->params[0].p; 401 RF_DiskQueue_t **dqs = ((RF_Raid_t *) (node->dagHdr->raidPtr))->Queues; 402 403 req = rf_CreateDiskQueueData(RF_IO_TYPE_NOP, 404 0L, 0, NULL, 0L, 0, 405 (int (*) (void *, int)) node->wakeFunc, 406 (void *) node, 407 NULL, node->dagHdr->tracerec, 408 (void *) (node->dagHdr->raidPtr), 409 RF_UNLOCK_DISK_QUEUE, NULL); 410 if (!req) 411 (node->wakeFunc) (node, ENOMEM); 412 else { 413 node->dagFuncData = (void *) req; 414 rf_DiskIOEnqueue(&(dqs[pda->row][pda->col]), req, RF_IO_NORMAL_PRIORITY); 415 } 416 417 return (0); 418 } 419 /***************************************************************************************** 420 * Callback routine for DiskRead and DiskWrite nodes. When the disk op completes, 421 * the routine is called to set the node status and inform the execution engine that 422 * the node has fired. 423 ****************************************************************************************/ 424 int 425 rf_GenericWakeupFunc(node, status) 426 RF_DagNode_t *node; 427 int status; 428 { 429 switch (node->status) { 430 case rf_bwd1: 431 node->status = rf_bwd2; 432 if (node->dagFuncData) 433 rf_FreeDiskQueueData((RF_DiskQueueData_t *) node->dagFuncData); 434 return (rf_DiskWriteFuncForThreads(node)); 435 break; 436 case rf_fired: 437 if (status) 438 node->status = rf_bad; 439 else 440 node->status = rf_good; 441 break; 442 case rf_recover: 443 /* probably should never reach this case */ 444 if (status) 445 node->status = rf_panic; 446 else 447 node->status = rf_undone; 448 break; 449 default: 450 printf("rf_GenericWakeupFunc:"); 451 printf("node->status is %d,", node->status); 452 printf("status is %d \n", status); 453 RF_PANIC(); 454 break; 455 } 456 if (node->dagFuncData) 457 rf_FreeDiskQueueData((RF_DiskQueueData_t *) node->dagFuncData); 458 return (rf_FinishNode(node, RF_INTR_CONTEXT)); 459 } 460 461 462 /***************************************************************************************** 463 * there are three distinct types of xor nodes 464 * A "regular xor" is used in the fault-free case where the access spans a complete 465 * stripe unit. It assumes that the result buffer is one full stripe unit in size, 466 * and uses the stripe-unit-offset values that it computes from the PDAs to determine 467 * where within the stripe unit to XOR each argument buffer. 468 * 469 * A "simple xor" is used in the fault-free case where the access touches only a portion 470 * of one (or two, in some cases) stripe unit(s). It assumes that all the argument 471 * buffers are of the same size and have the same stripe unit offset. 472 * 473 * A "recovery xor" is used in the degraded-mode case. It's similar to the regular 474 * xor function except that it takes the failed PDA as an additional parameter, and 475 * uses it to determine what portions of the argument buffers need to be xor'd into 476 * the result buffer, and where in the result buffer they should go. 477 ****************************************************************************************/ 478 479 /* xor the params together and store the result in the result field. 480 * assume the result field points to a buffer that is the size of one SU, 481 * and use the pda params to determine where within the buffer to XOR 482 * the input buffers. 483 */ 484 int 485 rf_RegularXorFunc(node) 486 RF_DagNode_t *node; 487 { 488 RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p; 489 RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec; 490 RF_Etimer_t timer; 491 int i, retcode; 492 493 retcode = 0; 494 if (node->dagHdr->status == rf_enable) { 495 /* don't do the XOR if the input is the same as the output */ 496 RF_ETIMER_START(timer); 497 for (i = 0; i < node->numParams - 1; i += 2) 498 if (node->params[i + 1].p != node->results[0]) { 499 retcode = rf_XorIntoBuffer(raidPtr, (RF_PhysDiskAddr_t *) node->params[i].p, 500 (char *) node->params[i + 1].p, (char *) node->results[0], node->dagHdr->bp); 501 } 502 RF_ETIMER_STOP(timer); 503 RF_ETIMER_EVAL(timer); 504 tracerec->xor_us += RF_ETIMER_VAL_US(timer); 505 } 506 return (rf_GenericWakeupFunc(node, retcode)); /* call wake func 507 * explicitly since no 508 * I/O in this node */ 509 } 510 /* xor the inputs into the result buffer, ignoring placement issues */ 511 int 512 rf_SimpleXorFunc(node) 513 RF_DagNode_t *node; 514 { 515 RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p; 516 int i, retcode = 0; 517 RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec; 518 RF_Etimer_t timer; 519 520 if (node->dagHdr->status == rf_enable) { 521 RF_ETIMER_START(timer); 522 /* don't do the XOR if the input is the same as the output */ 523 for (i = 0; i < node->numParams - 1; i += 2) 524 if (node->params[i + 1].p != node->results[0]) { 525 retcode = rf_bxor((char *) node->params[i + 1].p, (char *) node->results[0], 526 rf_RaidAddressToByte(raidPtr, ((RF_PhysDiskAddr_t *) node->params[i].p)->numSector), 527 (struct buf *) node->dagHdr->bp); 528 } 529 RF_ETIMER_STOP(timer); 530 RF_ETIMER_EVAL(timer); 531 tracerec->xor_us += RF_ETIMER_VAL_US(timer); 532 } 533 return (rf_GenericWakeupFunc(node, retcode)); /* call wake func 534 * explicitly since no 535 * I/O in this node */ 536 } 537 /* this xor is used by the degraded-mode dag functions to recover lost data. 538 * the second-to-last parameter is the PDA for the failed portion of the access. 539 * the code here looks at this PDA and assumes that the xor target buffer is 540 * equal in size to the number of sectors in the failed PDA. It then uses 541 * the other PDAs in the parameter list to determine where within the target 542 * buffer the corresponding data should be xored. 543 */ 544 int 545 rf_RecoveryXorFunc(node) 546 RF_DagNode_t *node; 547 { 548 RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[node->numParams - 1].p; 549 RF_RaidLayout_t *layoutPtr = (RF_RaidLayout_t *) & raidPtr->Layout; 550 RF_PhysDiskAddr_t *failedPDA = (RF_PhysDiskAddr_t *) node->params[node->numParams - 2].p; 551 int i, retcode = 0; 552 RF_PhysDiskAddr_t *pda; 553 int suoffset, failedSUOffset = rf_StripeUnitOffset(layoutPtr, failedPDA->startSector); 554 char *srcbuf, *destbuf; 555 RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec; 556 RF_Etimer_t timer; 557 558 if (node->dagHdr->status == rf_enable) { 559 RF_ETIMER_START(timer); 560 for (i = 0; i < node->numParams - 2; i += 2) 561 if (node->params[i + 1].p != node->results[0]) { 562 pda = (RF_PhysDiskAddr_t *) node->params[i].p; 563 srcbuf = (char *) node->params[i + 1].p; 564 suoffset = rf_StripeUnitOffset(layoutPtr, pda->startSector); 565 destbuf = ((char *) node->results[0]) + rf_RaidAddressToByte(raidPtr, suoffset - failedSUOffset); 566 retcode = rf_bxor(srcbuf, destbuf, rf_RaidAddressToByte(raidPtr, pda->numSector), node->dagHdr->bp); 567 } 568 RF_ETIMER_STOP(timer); 569 RF_ETIMER_EVAL(timer); 570 tracerec->xor_us += RF_ETIMER_VAL_US(timer); 571 } 572 return (rf_GenericWakeupFunc(node, retcode)); 573 } 574 /***************************************************************************************** 575 * The next three functions are utilities used by the above xor-execution functions. 576 ****************************************************************************************/ 577 578 579 /* 580 * this is just a glorified buffer xor. targbuf points to a buffer that is one full stripe unit 581 * in size. srcbuf points to a buffer that may be less than 1 SU, but never more. When the 582 * access described by pda is one SU in size (which by implication means it's SU-aligned), 583 * all that happens is (targbuf) <- (srcbuf ^ targbuf). When the access is less than one 584 * SU in size the XOR occurs on only the portion of targbuf identified in the pda. 585 */ 586 587 int 588 rf_XorIntoBuffer(raidPtr, pda, srcbuf, targbuf, bp) 589 RF_Raid_t *raidPtr; 590 RF_PhysDiskAddr_t *pda; 591 char *srcbuf; 592 char *targbuf; 593 void *bp; 594 { 595 char *targptr; 596 int sectPerSU = raidPtr->Layout.sectorsPerStripeUnit; 597 int SUOffset = pda->startSector % sectPerSU; 598 int length, retcode = 0; 599 600 RF_ASSERT(pda->numSector <= sectPerSU); 601 602 targptr = targbuf + rf_RaidAddressToByte(raidPtr, SUOffset); 603 length = rf_RaidAddressToByte(raidPtr, pda->numSector); 604 retcode = rf_bxor(srcbuf, targptr, length, bp); 605 return (retcode); 606 } 607 /* it really should be the case that the buffer pointers (returned by malloc) 608 * are aligned to the natural word size of the machine, so this is the only 609 * case we optimize for. The length should always be a multiple of the sector 610 * size, so there should be no problem with leftover bytes at the end. 611 */ 612 int 613 rf_bxor(src, dest, len, bp) 614 char *src; 615 char *dest; 616 int len; 617 void *bp; 618 { 619 unsigned mask = sizeof(long) - 1, retcode = 0; 620 621 if (!(((unsigned long) src) & mask) && !(((unsigned long) dest) & mask) && !(len & mask)) { 622 retcode = rf_longword_bxor((unsigned long *) src, (unsigned long *) dest, len >> RF_LONGSHIFT, bp); 623 } else { 624 RF_ASSERT(0); 625 } 626 return (retcode); 627 } 628 /* map a user buffer into kernel space, if necessary */ 629 #define REMAP_VA(_bp,x,y) (y) = (x) 630 631 /* When XORing in kernel mode, we need to map each user page to kernel space before we can access it. 632 * We don't want to assume anything about which input buffers are in kernel/user 633 * space, nor about their alignment, so in each loop we compute the maximum number 634 * of bytes that we can xor without crossing any page boundaries, and do only this many 635 * bytes before the next remap. 636 */ 637 int 638 rf_longword_bxor(src, dest, len, bp) 639 unsigned long *src; 640 unsigned long *dest; 641 int len; /* longwords */ 642 void *bp; 643 { 644 unsigned long *end = src + len; 645 unsigned long d0, d1, d2, d3, s0, s1, s2, s3; /* temps */ 646 unsigned long *pg_src, *pg_dest; /* per-page source/dest 647 * pointers */ 648 int longs_this_time;/* # longwords to xor in the current iteration */ 649 650 REMAP_VA(bp, src, pg_src); 651 REMAP_VA(bp, dest, pg_dest); 652 if (!pg_src || !pg_dest) 653 return (EFAULT); 654 655 while (len >= 4) { 656 longs_this_time = RF_MIN(len, RF_MIN(RF_BLIP(pg_src), RF_BLIP(pg_dest)) >> RF_LONGSHIFT); /* note len in longwords */ 657 src += longs_this_time; 658 dest += longs_this_time; 659 len -= longs_this_time; 660 while (longs_this_time >= 4) { 661 d0 = pg_dest[0]; 662 d1 = pg_dest[1]; 663 d2 = pg_dest[2]; 664 d3 = pg_dest[3]; 665 s0 = pg_src[0]; 666 s1 = pg_src[1]; 667 s2 = pg_src[2]; 668 s3 = pg_src[3]; 669 pg_dest[0] = d0 ^ s0; 670 pg_dest[1] = d1 ^ s1; 671 pg_dest[2] = d2 ^ s2; 672 pg_dest[3] = d3 ^ s3; 673 pg_src += 4; 674 pg_dest += 4; 675 longs_this_time -= 4; 676 } 677 while (longs_this_time > 0) { /* cannot cross any page 678 * boundaries here */ 679 *pg_dest++ ^= *pg_src++; 680 longs_this_time--; 681 } 682 683 /* either we're done, or we've reached a page boundary on one 684 * (or possibly both) of the pointers */ 685 if (len) { 686 if (RF_PAGE_ALIGNED(src)) 687 REMAP_VA(bp, src, pg_src); 688 if (RF_PAGE_ALIGNED(dest)) 689 REMAP_VA(bp, dest, pg_dest); 690 if (!pg_src || !pg_dest) 691 return (EFAULT); 692 } 693 } 694 while (src < end) { 695 *pg_dest++ ^= *pg_src++; 696 src++; 697 dest++; 698 len--; 699 if (RF_PAGE_ALIGNED(src)) 700 REMAP_VA(bp, src, pg_src); 701 if (RF_PAGE_ALIGNED(dest)) 702 REMAP_VA(bp, dest, pg_dest); 703 } 704 RF_ASSERT(len == 0); 705 return (0); 706 } 707 708 709 /* 710 dst = a ^ b ^ c; 711 a may equal dst 712 see comment above longword_bxor 713 */ 714 int 715 rf_longword_bxor3(dst, a, b, c, len, bp) 716 unsigned long *dst; 717 unsigned long *a; 718 unsigned long *b; 719 unsigned long *c; 720 int len; /* length in longwords */ 721 void *bp; 722 { 723 unsigned long a0, a1, a2, a3, b0, b1, b2, b3; 724 unsigned long *pg_a, *pg_b, *pg_c, *pg_dst; /* per-page source/dest 725 * pointers */ 726 int longs_this_time;/* # longs to xor in the current iteration */ 727 char dst_is_a = 0; 728 729 REMAP_VA(bp, a, pg_a); 730 REMAP_VA(bp, b, pg_b); 731 REMAP_VA(bp, c, pg_c); 732 if (a == dst) { 733 pg_dst = pg_a; 734 dst_is_a = 1; 735 } else { 736 REMAP_VA(bp, dst, pg_dst); 737 } 738 739 /* align dest to cache line. Can't cross a pg boundary on dst here. */ 740 while ((((unsigned long) pg_dst) & 0x1f)) { 741 *pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++; 742 dst++; 743 a++; 744 b++; 745 c++; 746 if (RF_PAGE_ALIGNED(a)) { 747 REMAP_VA(bp, a, pg_a); 748 if (!pg_a) 749 return (EFAULT); 750 } 751 if (RF_PAGE_ALIGNED(b)) { 752 REMAP_VA(bp, a, pg_b); 753 if (!pg_b) 754 return (EFAULT); 755 } 756 if (RF_PAGE_ALIGNED(c)) { 757 REMAP_VA(bp, a, pg_c); 758 if (!pg_c) 759 return (EFAULT); 760 } 761 len--; 762 } 763 764 while (len > 4) { 765 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); 766 a += longs_this_time; 767 b += longs_this_time; 768 c += longs_this_time; 769 dst += longs_this_time; 770 len -= longs_this_time; 771 while (longs_this_time >= 4) { 772 a0 = pg_a[0]; 773 longs_this_time -= 4; 774 775 a1 = pg_a[1]; 776 a2 = pg_a[2]; 777 778 a3 = pg_a[3]; 779 pg_a += 4; 780 781 b0 = pg_b[0]; 782 b1 = pg_b[1]; 783 784 b2 = pg_b[2]; 785 b3 = pg_b[3]; 786 /* start dual issue */ 787 a0 ^= b0; 788 b0 = pg_c[0]; 789 790 pg_b += 4; 791 a1 ^= b1; 792 793 a2 ^= b2; 794 a3 ^= b3; 795 796 b1 = pg_c[1]; 797 a0 ^= b0; 798 799 b2 = pg_c[2]; 800 a1 ^= b1; 801 802 b3 = pg_c[3]; 803 a2 ^= b2; 804 805 pg_dst[0] = a0; 806 a3 ^= b3; 807 pg_dst[1] = a1; 808 pg_c += 4; 809 pg_dst[2] = a2; 810 pg_dst[3] = a3; 811 pg_dst += 4; 812 } 813 while (longs_this_time > 0) { /* cannot cross any page 814 * boundaries here */ 815 *pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++; 816 longs_this_time--; 817 } 818 819 if (len) { 820 if (RF_PAGE_ALIGNED(a)) { 821 REMAP_VA(bp, a, pg_a); 822 if (!pg_a) 823 return (EFAULT); 824 if (dst_is_a) 825 pg_dst = pg_a; 826 } 827 if (RF_PAGE_ALIGNED(b)) { 828 REMAP_VA(bp, b, pg_b); 829 if (!pg_b) 830 return (EFAULT); 831 } 832 if (RF_PAGE_ALIGNED(c)) { 833 REMAP_VA(bp, c, pg_c); 834 if (!pg_c) 835 return (EFAULT); 836 } 837 if (!dst_is_a) 838 if (RF_PAGE_ALIGNED(dst)) { 839 REMAP_VA(bp, dst, pg_dst); 840 if (!pg_dst) 841 return (EFAULT); 842 } 843 } 844 } 845 while (len) { 846 *pg_dst++ = *pg_a++ ^ *pg_b++ ^ *pg_c++; 847 dst++; 848 a++; 849 b++; 850 c++; 851 if (RF_PAGE_ALIGNED(a)) { 852 REMAP_VA(bp, a, pg_a); 853 if (!pg_a) 854 return (EFAULT); 855 if (dst_is_a) 856 pg_dst = pg_a; 857 } 858 if (RF_PAGE_ALIGNED(b)) { 859 REMAP_VA(bp, b, pg_b); 860 if (!pg_b) 861 return (EFAULT); 862 } 863 if (RF_PAGE_ALIGNED(c)) { 864 REMAP_VA(bp, c, pg_c); 865 if (!pg_c) 866 return (EFAULT); 867 } 868 if (!dst_is_a) 869 if (RF_PAGE_ALIGNED(dst)) { 870 REMAP_VA(bp, dst, pg_dst); 871 if (!pg_dst) 872 return (EFAULT); 873 } 874 len--; 875 } 876 return (0); 877 } 878 879 int 880 rf_bxor3(dst, a, b, c, len, bp) 881 unsigned char *dst; 882 unsigned char *a; 883 unsigned char *b; 884 unsigned char *c; 885 unsigned long len; 886 void *bp; 887 { 888 RF_ASSERT(((RF_UL(dst) | RF_UL(a) | RF_UL(b) | RF_UL(c) | len) & 0x7) == 0); 889 890 return (rf_longword_bxor3((unsigned long *) dst, (unsigned long *) a, 891 (unsigned long *) b, (unsigned long *) c, len >> RF_LONGSHIFT, bp)); 892 } 893