1 /* $NetBSD: rf_reconbuffer.c,v 1.5 2001/01/27 20:10:49 oster Exp $ */ 2 /* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: Mark Holland 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 * 31 * rf_reconbuffer.c -- reconstruction buffer manager 32 * 33 ***************************************************/ 34 35 #include "rf_raid.h" 36 #include "rf_reconbuffer.h" 37 #include "rf_acctrace.h" 38 #include "rf_etimer.h" 39 #include "rf_general.h" 40 #include "rf_debugprint.h" 41 #include "rf_revent.h" 42 #include "rf_reconutil.h" 43 #include "rf_nwayxor.h" 44 45 #define Dprintf1(s,a) if (rf_reconbufferDebug) printf(s,a) 46 #define Dprintf2(s,a,b) if (rf_reconbufferDebug) printf(s,a,b) 47 #define Dprintf3(s,a,b,c) if (rf_reconbufferDebug) printf(s,a,b,c) 48 #define Dprintf4(s,a,b,c,d) if (rf_reconbufferDebug) printf(s,a,b,c,d) 49 #define Dprintf5(s,a,b,c,d,e) if (rf_reconbufferDebug) printf(s,a,b,c,d,e) 50 51 /***************************************************************************** 52 * 53 * Submit a reconstruction buffer to the manager for XOR. We can only 54 * submit a buffer if (1) we can xor into an existing buffer, which 55 * means we don't have to acquire a new one, (2) we can acquire a 56 * floating recon buffer, or (3) the caller has indicated that we are 57 * allowed to keep the submitted buffer. 58 * 59 * Returns non-zero if and only if we were not able to submit. 60 * In this case, we append the current disk ID to the wait list on the 61 * indicated RU, so that it will be re-enabled when we acquire a buffer 62 * for this RU. 63 * 64 ****************************************************************************/ 65 66 /* 67 * nWayXorFuncs[i] is a pointer to a function that will xor "i" 68 * bufs into the accumulating sum. 69 */ 70 static RF_VoidFuncPtr nWayXorFuncs[] = { 71 NULL, 72 (RF_VoidFuncPtr) rf_nWayXor1, 73 (RF_VoidFuncPtr) rf_nWayXor2, 74 (RF_VoidFuncPtr) rf_nWayXor3, 75 (RF_VoidFuncPtr) rf_nWayXor4, 76 (RF_VoidFuncPtr) rf_nWayXor5, 77 (RF_VoidFuncPtr) rf_nWayXor6, 78 (RF_VoidFuncPtr) rf_nWayXor7, 79 (RF_VoidFuncPtr) rf_nWayXor8, 80 (RF_VoidFuncPtr) rf_nWayXor9 81 }; 82 83 int 84 rf_SubmitReconBuffer(rbuf, keep_it, use_committed) 85 RF_ReconBuffer_t *rbuf; /* the recon buffer to submit */ 86 int keep_it; /* whether we can keep this buffer or we have 87 * to return it */ 88 int use_committed; /* whether to use a committed or an available 89 * recon buffer */ 90 { 91 RF_LayoutSW_t *lp; 92 int rc; 93 94 lp = rbuf->raidPtr->Layout.map; 95 rc = lp->SubmitReconBuffer(rbuf, keep_it, use_committed); 96 return (rc); 97 } 98 99 int 100 rf_SubmitReconBufferBasic(rbuf, keep_it, use_committed) 101 RF_ReconBuffer_t *rbuf; /* the recon buffer to submit */ 102 int keep_it; /* whether we can keep this buffer or we have 103 * to return it */ 104 int use_committed; /* whether to use a committed or an available 105 * recon buffer */ 106 { 107 RF_Raid_t *raidPtr = rbuf->raidPtr; 108 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 109 RF_ReconCtrl_t *reconCtrlPtr = raidPtr->reconControl[rbuf->row]; 110 RF_ReconParityStripeStatus_t *pssPtr; 111 RF_ReconBuffer_t *targetRbuf, *t = NULL; /* temporary rbuf 112 * pointers */ 113 caddr_t ta; /* temporary data buffer pointer */ 114 RF_CallbackDesc_t *cb, *p; 115 int retcode = 0, created = 0; 116 117 RF_Etimer_t timer; 118 119 /* makes no sense to have a submission from the failed disk */ 120 RF_ASSERT(rbuf); 121 RF_ASSERT(rbuf->col != reconCtrlPtr->fcol); 122 123 Dprintf5("RECON: submission by row %d col %d for psid %ld ru %d (failed offset %ld)\n", 124 rbuf->row, rbuf->col, (long) rbuf->parityStripeID, rbuf->which_ru, (long) rbuf->failedDiskSectorOffset); 125 126 RF_LOCK_PSS_MUTEX(raidPtr, rbuf->row, rbuf->parityStripeID); 127 128 RF_LOCK_MUTEX(reconCtrlPtr->rb_mutex); 129 130 pssPtr = rf_LookupRUStatus(raidPtr, reconCtrlPtr->pssTable, rbuf->parityStripeID, rbuf->which_ru, RF_PSS_NONE, &created); 131 RF_ASSERT(pssPtr); /* if it didn't exist, we wouldn't have gotten 132 * an rbuf for it */ 133 134 /* check to see if enough buffers have accumulated to do an XOR. If 135 * so, there's no need to acquire a floating rbuf. Before we can do 136 * any XORing, we must have acquired a destination buffer. If we 137 * have, then we can go ahead and do the XOR if (1) including this 138 * buffer, enough bufs have accumulated, or (2) this is the last 139 * submission for this stripe. Otherwise, we have to go acquire a 140 * floating rbuf. */ 141 142 targetRbuf = (RF_ReconBuffer_t *) pssPtr->rbuf; 143 if ((targetRbuf != NULL) && 144 ((pssPtr->xorBufCount == rf_numBufsToAccumulate - 1) || (targetRbuf->count + pssPtr->xorBufCount + 1 == layoutPtr->numDataCol))) { 145 pssPtr->rbufsForXor[pssPtr->xorBufCount++] = rbuf; /* install this buffer */ 146 Dprintf3("RECON: row %d col %d invoking a %d-way XOR\n", rbuf->row, rbuf->col, pssPtr->xorBufCount); 147 RF_ETIMER_START(timer); 148 rf_MultiWayReconXor(raidPtr, pssPtr); 149 RF_ETIMER_STOP(timer); 150 RF_ETIMER_EVAL(timer); 151 raidPtr->accumXorTimeUs += RF_ETIMER_VAL_US(timer); 152 if (!keep_it) { 153 raidPtr->recon_tracerecs[rbuf->col].xor_us = RF_ETIMER_VAL_US(timer); 154 RF_ETIMER_STOP(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 155 RF_ETIMER_EVAL(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 156 raidPtr->recon_tracerecs[rbuf->col].specific.recon.recon_return_to_submit_us += 157 RF_ETIMER_VAL_US(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 158 RF_ETIMER_START(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 159 160 rf_LogTraceRec(raidPtr, &raidPtr->recon_tracerecs[rbuf->col]); 161 } 162 rf_CheckForFullRbuf(raidPtr, reconCtrlPtr, pssPtr, layoutPtr->numDataCol); 163 164 /* if use_committed is on, we _must_ consume a buffer off the 165 * committed list. */ 166 if (use_committed) { 167 t = reconCtrlPtr->committedRbufs; 168 RF_ASSERT(t); 169 reconCtrlPtr->committedRbufs = t->next; 170 rf_ReleaseFloatingReconBuffer(raidPtr, rbuf->row, t); 171 } 172 if (keep_it) { 173 RF_UNLOCK_PSS_MUTEX(raidPtr, rbuf->row, rbuf->parityStripeID); 174 RF_UNLOCK_MUTEX(reconCtrlPtr->rb_mutex); 175 rf_FreeReconBuffer(rbuf); 176 return (retcode); 177 } 178 goto out; 179 } 180 /* set the value of "t", which we'll use as the rbuf from here on */ 181 if (keep_it) { 182 t = rbuf; 183 } else { 184 if (use_committed) { /* if a buffer has been committed to 185 * us, use it */ 186 t = reconCtrlPtr->committedRbufs; 187 RF_ASSERT(t); 188 reconCtrlPtr->committedRbufs = t->next; 189 t->next = NULL; 190 } else 191 if (reconCtrlPtr->floatingRbufs) { 192 t = reconCtrlPtr->floatingRbufs; 193 reconCtrlPtr->floatingRbufs = t->next; 194 t->next = NULL; 195 } 196 } 197 198 /* If we weren't able to acquire a buffer, append to the end of the 199 * buf list in the recon ctrl struct. */ 200 if (!t) { 201 RF_ASSERT(!keep_it && !use_committed); 202 Dprintf2("RECON: row %d col %d failed to acquire floating rbuf\n", rbuf->row, rbuf->col); 203 204 raidPtr->procsInBufWait++; 205 if ((raidPtr->procsInBufWait == raidPtr->numCol - 1) && (raidPtr->numFullReconBuffers == 0)) { 206 printf("Buffer wait deadlock detected. Exiting.\n"); 207 rf_PrintPSStatusTable(raidPtr, rbuf->row); 208 RF_PANIC(); 209 } 210 pssPtr->flags |= RF_PSS_BUFFERWAIT; 211 cb = rf_AllocCallbackDesc(); /* append to buf wait list in 212 * recon ctrl structure */ 213 cb->row = rbuf->row; 214 cb->col = rbuf->col; 215 cb->callbackArg.v = rbuf->parityStripeID; 216 cb->callbackArg2.v = rbuf->which_ru; 217 cb->next = NULL; 218 if (!reconCtrlPtr->bufferWaitList) 219 reconCtrlPtr->bufferWaitList = cb; 220 else { /* might want to maintain head/tail pointers 221 * here rather than search for end of list */ 222 for (p = reconCtrlPtr->bufferWaitList; p->next; p = p->next); 223 p->next = cb; 224 } 225 retcode = 1; 226 goto out; 227 } 228 Dprintf2("RECON: row %d col %d acquired rbuf\n", rbuf->row, rbuf->col); 229 RF_ETIMER_STOP(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 230 RF_ETIMER_EVAL(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 231 raidPtr->recon_tracerecs[rbuf->col].specific.recon.recon_return_to_submit_us += 232 RF_ETIMER_VAL_US(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 233 RF_ETIMER_START(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 234 235 rf_LogTraceRec(raidPtr, &raidPtr->recon_tracerecs[rbuf->col]); 236 237 /* initialize the buffer */ 238 if (t != rbuf) { 239 t->row = rbuf->row; 240 t->col = reconCtrlPtr->fcol; 241 t->parityStripeID = rbuf->parityStripeID; 242 t->which_ru = rbuf->which_ru; 243 t->failedDiskSectorOffset = rbuf->failedDiskSectorOffset; 244 t->spRow = rbuf->spRow; 245 t->spCol = rbuf->spCol; 246 t->spOffset = rbuf->spOffset; 247 248 ta = t->buffer; 249 t->buffer = rbuf->buffer; 250 rbuf->buffer = ta; /* swap buffers */ 251 } 252 /* the first installation always gets installed as the destination 253 * buffer. subsequent installations get stacked up to allow for 254 * multi-way XOR */ 255 if (!pssPtr->rbuf) { 256 pssPtr->rbuf = t; 257 t->count = 1; 258 } else 259 pssPtr->rbufsForXor[pssPtr->xorBufCount++] = t; /* install this buffer */ 260 261 rf_CheckForFullRbuf(raidPtr, reconCtrlPtr, pssPtr, layoutPtr->numDataCol); /* the buffer is full if 262 * G=2 */ 263 264 out: 265 RF_UNLOCK_PSS_MUTEX(raidPtr, rbuf->row, rbuf->parityStripeID); 266 RF_UNLOCK_MUTEX(reconCtrlPtr->rb_mutex); 267 return (retcode); 268 } 269 270 int 271 rf_MultiWayReconXor(raidPtr, pssPtr) 272 RF_Raid_t *raidPtr; 273 RF_ReconParityStripeStatus_t *pssPtr; /* the pss descriptor for this 274 * parity stripe */ 275 { 276 int i, numBufs = pssPtr->xorBufCount; 277 int numBytes = rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit * raidPtr->Layout.SUsPerRU); 278 RF_ReconBuffer_t **rbufs = (RF_ReconBuffer_t **) pssPtr->rbufsForXor; 279 RF_ReconBuffer_t *targetRbuf = (RF_ReconBuffer_t *) pssPtr->rbuf; 280 281 RF_ASSERT(pssPtr->rbuf != NULL); 282 RF_ASSERT(numBufs > 0 && numBufs < RF_PS_MAX_BUFS); 283 #ifdef _KERNEL 284 #ifndef __NetBSD__ 285 thread_block(); /* yield the processor before doing a big XOR */ 286 #endif 287 #endif /* _KERNEL */ 288 /* 289 * XXX 290 * 291 * What if more than 9 bufs? 292 */ 293 nWayXorFuncs[numBufs] (pssPtr->rbufsForXor, targetRbuf, numBytes / sizeof(long)); 294 295 /* release all the reconstruction buffers except the last one, which 296 * belongs to the disk whose submission caused this XOR to take place */ 297 for (i = 0; i < numBufs - 1; i++) { 298 if (rbufs[i]->type == RF_RBUF_TYPE_FLOATING) 299 rf_ReleaseFloatingReconBuffer(raidPtr, rbufs[i]->row, rbufs[i]); 300 else 301 if (rbufs[i]->type == RF_RBUF_TYPE_FORCED) 302 rf_FreeReconBuffer(rbufs[i]); 303 else 304 RF_ASSERT(0); 305 } 306 targetRbuf->count += pssPtr->xorBufCount; 307 pssPtr->xorBufCount = 0; 308 return (0); 309 } 310 /* removes one full buffer from one of the full-buffer lists and returns it. 311 * 312 * ASSUMES THE RB_MUTEX IS UNLOCKED AT ENTRY. 313 */ 314 RF_ReconBuffer_t * 315 rf_GetFullReconBuffer(reconCtrlPtr) 316 RF_ReconCtrl_t *reconCtrlPtr; 317 { 318 RF_ReconBuffer_t *p; 319 320 RF_LOCK_MUTEX(reconCtrlPtr->rb_mutex); 321 322 if ((p = reconCtrlPtr->priorityList) != NULL) { 323 reconCtrlPtr->priorityList = p->next; 324 p->next = NULL; 325 goto out; 326 } 327 if ((p = reconCtrlPtr->fullBufferList) != NULL) { 328 reconCtrlPtr->fullBufferList = p->next; 329 p->next = NULL; 330 goto out; 331 } 332 out: 333 RF_UNLOCK_MUTEX(reconCtrlPtr->rb_mutex); 334 return (p); 335 } 336 337 338 /* if the reconstruction buffer is full, move it to the full list, 339 * which is maintained sorted by failed disk sector offset 340 * 341 * ASSUMES THE RB_MUTEX IS LOCKED AT ENTRY. */ 342 int 343 rf_CheckForFullRbuf(raidPtr, reconCtrl, pssPtr, numDataCol) 344 RF_Raid_t *raidPtr; 345 RF_ReconCtrl_t *reconCtrl; 346 RF_ReconParityStripeStatus_t *pssPtr; 347 int numDataCol; 348 { 349 RF_ReconBuffer_t *p, *pt, *rbuf = (RF_ReconBuffer_t *) pssPtr->rbuf; 350 351 if (rbuf->count == numDataCol) { 352 raidPtr->numFullReconBuffers++; 353 Dprintf2("RECON: rbuf for psid %ld ru %d has filled\n", 354 (long) rbuf->parityStripeID, rbuf->which_ru); 355 if (!reconCtrl->fullBufferList || (rbuf->failedDiskSectorOffset < reconCtrl->fullBufferList->failedDiskSectorOffset)) { 356 Dprintf2("RECON: rbuf for psid %ld ru %d is head of list\n", 357 (long) rbuf->parityStripeID, rbuf->which_ru); 358 rbuf->next = reconCtrl->fullBufferList; 359 reconCtrl->fullBufferList = rbuf; 360 } else { 361 for (pt = reconCtrl->fullBufferList, p = pt->next; p && p->failedDiskSectorOffset < rbuf->failedDiskSectorOffset; pt = p, p = p->next); 362 rbuf->next = p; 363 pt->next = rbuf; 364 Dprintf2("RECON: rbuf for psid %ld ru %d is in list\n", 365 (long) rbuf->parityStripeID, rbuf->which_ru); 366 } 367 #if 0 368 pssPtr->writeRbuf = pssPtr->rbuf; /* DEBUG ONLY: we like 369 * to be able to find 370 * this rbuf while it's 371 * awaiting write */ 372 #else 373 rbuf->pssPtr = pssPtr; 374 #endif 375 pssPtr->rbuf = NULL; 376 rf_CauseReconEvent(raidPtr, rbuf->row, rbuf->col, NULL, RF_REVENT_BUFREADY); 377 } 378 return (0); 379 } 380 381 382 /* release a floating recon buffer for someone else to use. 383 * assumes the rb_mutex is LOCKED at entry 384 */ 385 void 386 rf_ReleaseFloatingReconBuffer(raidPtr, row, rbuf) 387 RF_Raid_t *raidPtr; 388 RF_RowCol_t row; 389 RF_ReconBuffer_t *rbuf; 390 { 391 RF_ReconCtrl_t *rcPtr = raidPtr->reconControl[row]; 392 RF_CallbackDesc_t *cb; 393 394 Dprintf2("RECON: releasing rbuf for psid %ld ru %d\n", 395 (long) rbuf->parityStripeID, rbuf->which_ru); 396 397 /* if anyone is waiting on buffers, wake one of them up. They will 398 * subsequently wake up anyone else waiting on their RU */ 399 if (rcPtr->bufferWaitList) { 400 rbuf->next = rcPtr->committedRbufs; 401 rcPtr->committedRbufs = rbuf; 402 cb = rcPtr->bufferWaitList; 403 rcPtr->bufferWaitList = cb->next; 404 rf_CauseReconEvent(raidPtr, cb->row, cb->col, (void *) 1, RF_REVENT_BUFCLEAR); /* arg==1 => we've 405 * committed a buffer */ 406 rf_FreeCallbackDesc(cb); 407 raidPtr->procsInBufWait--; 408 } else { 409 rbuf->next = rcPtr->floatingRbufs; 410 rcPtr->floatingRbufs = rbuf; 411 } 412 } 413 /* release any disk that is waiting on a buffer for the indicated RU. 414 * assumes the rb_mutex is LOCKED at entry 415 */ 416 void 417 rf_ReleaseBufferWaiters(raidPtr, pssPtr) 418 RF_Raid_t *raidPtr; 419 RF_ReconParityStripeStatus_t *pssPtr; 420 { 421 RF_CallbackDesc_t *cb1, *cb = pssPtr->bufWaitList; 422 423 Dprintf2("RECON: releasing buf waiters for psid %ld ru %d\n", 424 (long) pssPtr->parityStripeID, pssPtr->which_ru); 425 pssPtr->flags &= ~RF_PSS_BUFFERWAIT; 426 while (cb) { 427 cb1 = cb->next; 428 cb->next = NULL; 429 rf_CauseReconEvent(raidPtr, cb->row, cb->col, (void *) 0, RF_REVENT_BUFCLEAR); /* arg==0 => we haven't 430 * committed a buffer */ 431 rf_FreeCallbackDesc(cb); 432 cb = cb1; 433 } 434 pssPtr->bufWaitList = NULL; 435 } 436 /* when reconstruction is forced on an RU, there may be some disks waiting to 437 * acquire a buffer for that RU. Since we allocate a new buffer as part of 438 * the forced-reconstruction process, we no longer have to wait for any 439 * buffers, so we wakeup any waiter that we find in the bufferWaitList 440 * 441 * assumes the rb_mutex is LOCKED at entry 442 */ 443 void 444 rf_ReleaseBufferWaiter(rcPtr, rbuf) 445 RF_ReconCtrl_t *rcPtr; 446 RF_ReconBuffer_t *rbuf; 447 { 448 RF_CallbackDesc_t *cb, *cbt; 449 450 for (cbt = NULL, cb = rcPtr->bufferWaitList; cb; cbt = cb, cb = cb->next) { 451 if ((cb->callbackArg.v == rbuf->parityStripeID) && (cb->callbackArg2.v == rbuf->which_ru)) { 452 Dprintf2("RECON: Dropping row %d col %d from buffer wait list\n", cb->row, cb->col); 453 if (cbt) 454 cbt->next = cb->next; 455 else 456 rcPtr->bufferWaitList = cb->next; 457 rf_CauseReconEvent((RF_Raid_t *) rbuf->raidPtr, cb->row, cb->col, (void *) 0, RF_REVENT_BUFREADY); /* arg==0 => no 458 * committed buffer */ 459 rf_FreeCallbackDesc(cb); 460 return; 461 } 462 } 463 } 464