1 /* $NetBSD: rf_reconbuffer.c,v 1.17 2004/01/02 11:52:15 hannken 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 <sys/cdefs.h> 36 __KERNEL_RCSID(0, "$NetBSD: rf_reconbuffer.c,v 1.17 2004/01/02 11:52:15 hannken Exp $"); 37 38 #include "rf_raid.h" 39 #include "rf_reconbuffer.h" 40 #include "rf_acctrace.h" 41 #include "rf_etimer.h" 42 #include "rf_general.h" 43 #include "rf_revent.h" 44 #include "rf_reconutil.h" 45 #include "rf_nwayxor.h" 46 47 #ifdef DEBUG 48 49 #define Dprintf1(s,a) if (rf_reconbufferDebug) printf(s,a) 50 #define Dprintf2(s,a,b) if (rf_reconbufferDebug) printf(s,a,b) 51 #define Dprintf3(s,a,b,c) if (rf_reconbufferDebug) printf(s,a,b,c) 52 #define Dprintf4(s,a,b,c,d) if (rf_reconbufferDebug) printf(s,a,b,c,d) 53 #define Dprintf5(s,a,b,c,d,e) if (rf_reconbufferDebug) printf(s,a,b,c,d,e) 54 55 #else /* DEBUG */ 56 57 #define Dprintf1(s,a) {} 58 #define Dprintf2(s,a,b) {} 59 #define Dprintf3(s,a,b,c) {} 60 #define Dprintf4(s,a,b,c,d) {} 61 #define Dprintf5(s,a,b,c,d,e) {} 62 63 #endif 64 65 /***************************************************************************** 66 * 67 * Submit a reconstruction buffer to the manager for XOR. We can only 68 * submit a buffer if (1) we can xor into an existing buffer, which 69 * means we don't have to acquire a new one, (2) we can acquire a 70 * floating recon buffer, or (3) the caller has indicated that we are 71 * allowed to keep the submitted buffer. 72 * 73 * Returns non-zero if and only if we were not able to submit. 74 * In this case, we append the current disk ID to the wait list on the 75 * indicated RU, so that it will be re-enabled when we acquire a buffer 76 * for this RU. 77 * 78 ****************************************************************************/ 79 80 /* 81 * nWayXorFuncs[i] is a pointer to a function that will xor "i" 82 * bufs into the accumulating sum. 83 */ 84 static const RF_VoidFuncPtr nWayXorFuncs[] = { 85 NULL, 86 (RF_VoidFuncPtr) rf_nWayXor1, 87 (RF_VoidFuncPtr) rf_nWayXor2, 88 (RF_VoidFuncPtr) rf_nWayXor3, 89 (RF_VoidFuncPtr) rf_nWayXor4, 90 (RF_VoidFuncPtr) rf_nWayXor5, 91 (RF_VoidFuncPtr) rf_nWayXor6, 92 (RF_VoidFuncPtr) rf_nWayXor7, 93 (RF_VoidFuncPtr) rf_nWayXor8, 94 (RF_VoidFuncPtr) rf_nWayXor9 95 }; 96 97 /* 98 * rbuf - the recon buffer to submit 99 * keep_it - whether we can keep this buffer or we have to return it 100 * use_committed - whether to use a committed or an available recon buffer 101 */ 102 int 103 rf_SubmitReconBuffer(RF_ReconBuffer_t *rbuf, int keep_it, int use_committed) 104 { 105 const RF_LayoutSW_t *lp; 106 int rc; 107 108 lp = rbuf->raidPtr->Layout.map; 109 rc = lp->SubmitReconBuffer(rbuf, keep_it, use_committed); 110 return (rc); 111 } 112 113 /* 114 * rbuf - the recon buffer to submit 115 * keep_it - whether we can keep this buffer or we have to return it 116 * use_committed - whether to use a committed or an available recon buffer 117 */ 118 int 119 rf_SubmitReconBufferBasic(RF_ReconBuffer_t *rbuf, int keep_it, 120 int use_committed) 121 { 122 RF_Raid_t *raidPtr = rbuf->raidPtr; 123 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 124 RF_ReconCtrl_t *reconCtrlPtr = raidPtr->reconControl; 125 RF_ReconParityStripeStatus_t *pssPtr; 126 RF_ReconBuffer_t *targetRbuf, *t = NULL; /* temporary rbuf 127 * pointers */ 128 caddr_t ta; /* temporary data buffer pointer */ 129 RF_CallbackDesc_t *cb, *p; 130 int retcode = 0, created = 0; 131 132 RF_Etimer_t timer; 133 134 /* makes no sense to have a submission from the failed disk */ 135 RF_ASSERT(rbuf); 136 RF_ASSERT(rbuf->col != reconCtrlPtr->fcol); 137 138 Dprintf4("RECON: submission by col %d for psid %ld ru %d (failed offset %ld)\n", 139 rbuf->col, (long) rbuf->parityStripeID, rbuf->which_ru, (long) rbuf->failedDiskSectorOffset); 140 141 RF_LOCK_PSS_MUTEX(raidPtr, rbuf->parityStripeID); 142 143 RF_LOCK_MUTEX(reconCtrlPtr->rb_mutex); 144 145 pssPtr = rf_LookupRUStatus(raidPtr, reconCtrlPtr->pssTable, rbuf->parityStripeID, rbuf->which_ru, RF_PSS_NONE, &created); 146 RF_ASSERT(pssPtr); /* if it didn't exist, we wouldn't have gotten 147 * an rbuf for it */ 148 149 /* check to see if enough buffers have accumulated to do an XOR. If 150 * so, there's no need to acquire a floating rbuf. Before we can do 151 * any XORing, we must have acquired a destination buffer. If we 152 * have, then we can go ahead and do the XOR if (1) including this 153 * buffer, enough bufs have accumulated, or (2) this is the last 154 * submission for this stripe. Otherwise, we have to go acquire a 155 * floating rbuf. */ 156 157 targetRbuf = (RF_ReconBuffer_t *) pssPtr->rbuf; 158 if ((targetRbuf != NULL) && 159 ((pssPtr->xorBufCount == rf_numBufsToAccumulate - 1) || (targetRbuf->count + pssPtr->xorBufCount + 1 == layoutPtr->numDataCol))) { 160 pssPtr->rbufsForXor[pssPtr->xorBufCount++] = rbuf; /* install this buffer */ 161 Dprintf2("RECON: col %d invoking a %d-way XOR\n", rbuf->col, pssPtr->xorBufCount); 162 RF_ETIMER_START(timer); 163 rf_MultiWayReconXor(raidPtr, pssPtr); 164 RF_ETIMER_STOP(timer); 165 RF_ETIMER_EVAL(timer); 166 raidPtr->accumXorTimeUs += RF_ETIMER_VAL_US(timer); 167 if (!keep_it) { 168 raidPtr->recon_tracerecs[rbuf->col].xor_us = RF_ETIMER_VAL_US(timer); 169 RF_ETIMER_STOP(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 170 RF_ETIMER_EVAL(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 171 raidPtr->recon_tracerecs[rbuf->col].specific.recon.recon_return_to_submit_us += 172 RF_ETIMER_VAL_US(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 173 RF_ETIMER_START(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 174 175 rf_LogTraceRec(raidPtr, &raidPtr->recon_tracerecs[rbuf->col]); 176 } 177 rf_CheckForFullRbuf(raidPtr, reconCtrlPtr, pssPtr, layoutPtr->numDataCol); 178 179 /* if use_committed is on, we _must_ consume a buffer off the 180 * committed list. */ 181 if (use_committed) { 182 t = reconCtrlPtr->committedRbufs; 183 RF_ASSERT(t); 184 reconCtrlPtr->committedRbufs = t->next; 185 rf_ReleaseFloatingReconBuffer(raidPtr, t); 186 } 187 if (keep_it) { 188 RF_UNLOCK_PSS_MUTEX(raidPtr, rbuf->parityStripeID); 189 RF_UNLOCK_MUTEX(reconCtrlPtr->rb_mutex); 190 rf_FreeReconBuffer(rbuf); 191 return (retcode); 192 } 193 goto out; 194 } 195 /* set the value of "t", which we'll use as the rbuf from here on */ 196 if (keep_it) { 197 t = rbuf; 198 } else { 199 if (use_committed) { /* if a buffer has been committed to 200 * us, use it */ 201 t = reconCtrlPtr->committedRbufs; 202 RF_ASSERT(t); 203 reconCtrlPtr->committedRbufs = t->next; 204 t->next = NULL; 205 } else 206 if (reconCtrlPtr->floatingRbufs) { 207 t = reconCtrlPtr->floatingRbufs; 208 reconCtrlPtr->floatingRbufs = t->next; 209 t->next = NULL; 210 } 211 } 212 213 /* If we weren't able to acquire a buffer, append to the end of the 214 * buf list in the recon ctrl struct. */ 215 if (!t) { 216 RF_ASSERT(!keep_it && !use_committed); 217 Dprintf1("RECON: col %d failed to acquire floating rbuf\n", rbuf->col); 218 219 raidPtr->procsInBufWait++; 220 if ((raidPtr->procsInBufWait == raidPtr->numCol - 1) && (raidPtr->numFullReconBuffers == 0)) { 221 printf("Buffer wait deadlock detected. Exiting.\n"); 222 rf_PrintPSStatusTable(raidPtr); 223 RF_PANIC(); 224 } 225 pssPtr->flags |= RF_PSS_BUFFERWAIT; 226 cb = rf_AllocCallbackDesc(); /* append to buf wait list in 227 * recon ctrl structure */ 228 cb->col = rbuf->col; 229 cb->callbackArg.v = rbuf->parityStripeID; 230 cb->callbackArg2.v = rbuf->which_ru; 231 cb->next = NULL; 232 if (!reconCtrlPtr->bufferWaitList) 233 reconCtrlPtr->bufferWaitList = cb; 234 else { /* might want to maintain head/tail pointers 235 * here rather than search for end of list */ 236 for (p = reconCtrlPtr->bufferWaitList; p->next; p = p->next); 237 p->next = cb; 238 } 239 retcode = 1; 240 goto out; 241 } 242 Dprintf1("RECON: col %d acquired rbuf\n", rbuf->col); 243 RF_ETIMER_STOP(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 244 RF_ETIMER_EVAL(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 245 raidPtr->recon_tracerecs[rbuf->col].specific.recon.recon_return_to_submit_us += 246 RF_ETIMER_VAL_US(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 247 RF_ETIMER_START(raidPtr->recon_tracerecs[rbuf->col].recon_timer); 248 249 rf_LogTraceRec(raidPtr, &raidPtr->recon_tracerecs[rbuf->col]); 250 251 /* initialize the buffer */ 252 if (t != rbuf) { 253 t->col = reconCtrlPtr->fcol; 254 t->parityStripeID = rbuf->parityStripeID; 255 t->which_ru = rbuf->which_ru; 256 t->failedDiskSectorOffset = rbuf->failedDiskSectorOffset; 257 t->spCol = rbuf->spCol; 258 t->spOffset = rbuf->spOffset; 259 260 ta = t->buffer; 261 t->buffer = rbuf->buffer; 262 rbuf->buffer = ta; /* swap buffers */ 263 } 264 /* the first installation always gets installed as the destination 265 * buffer. subsequent installations get stacked up to allow for 266 * multi-way XOR */ 267 if (!pssPtr->rbuf) { 268 pssPtr->rbuf = t; 269 t->count = 1; 270 } else 271 pssPtr->rbufsForXor[pssPtr->xorBufCount++] = t; /* install this buffer */ 272 273 rf_CheckForFullRbuf(raidPtr, reconCtrlPtr, pssPtr, layoutPtr->numDataCol); /* the buffer is full if 274 * G=2 */ 275 276 out: 277 RF_UNLOCK_PSS_MUTEX(raidPtr, rbuf->parityStripeID); 278 RF_UNLOCK_MUTEX(reconCtrlPtr->rb_mutex); 279 return (retcode); 280 } 281 /* pssPtr - the pss descriptor for this parity stripe */ 282 int 283 rf_MultiWayReconXor(RF_Raid_t *raidPtr, RF_ReconParityStripeStatus_t *pssPtr) 284 { 285 int i, numBufs = pssPtr->xorBufCount; 286 int numBytes = rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit * raidPtr->Layout.SUsPerRU); 287 RF_ReconBuffer_t **rbufs = (RF_ReconBuffer_t **) pssPtr->rbufsForXor; 288 RF_ReconBuffer_t *targetRbuf = (RF_ReconBuffer_t *) pssPtr->rbuf; 289 290 RF_ASSERT(pssPtr->rbuf != NULL); 291 RF_ASSERT(numBufs > 0 && numBufs < RF_PS_MAX_BUFS); 292 #ifdef _KERNEL 293 #ifndef __NetBSD__ 294 thread_block(); /* yield the processor before doing a big XOR */ 295 #endif 296 #endif /* _KERNEL */ 297 /* 298 * XXX 299 * 300 * What if more than 9 bufs? 301 */ 302 nWayXorFuncs[numBufs] (pssPtr->rbufsForXor, targetRbuf, numBytes / sizeof(long)); 303 304 /* release all the reconstruction buffers except the last one, which 305 * belongs to the disk whose submission caused this XOR to take place */ 306 for (i = 0; i < numBufs - 1; i++) { 307 if (rbufs[i]->type == RF_RBUF_TYPE_FLOATING) 308 rf_ReleaseFloatingReconBuffer(raidPtr, rbufs[i]); 309 else 310 if (rbufs[i]->type == RF_RBUF_TYPE_FORCED) 311 rf_FreeReconBuffer(rbufs[i]); 312 else 313 RF_ASSERT(0); 314 } 315 targetRbuf->count += pssPtr->xorBufCount; 316 pssPtr->xorBufCount = 0; 317 return (0); 318 } 319 /* removes one full buffer from one of the full-buffer lists and returns it. 320 * 321 * ASSUMES THE RB_MUTEX IS UNLOCKED AT ENTRY. 322 */ 323 RF_ReconBuffer_t * 324 rf_GetFullReconBuffer(RF_ReconCtrl_t *reconCtrlPtr) 325 { 326 RF_ReconBuffer_t *p; 327 328 RF_LOCK_MUTEX(reconCtrlPtr->rb_mutex); 329 330 if ((p = reconCtrlPtr->fullBufferList) != NULL) { 331 reconCtrlPtr->fullBufferList = p->next; 332 p->next = NULL; 333 } 334 RF_UNLOCK_MUTEX(reconCtrlPtr->rb_mutex); 335 return (p); 336 } 337 338 339 /* if the reconstruction buffer is full, move it to the full list, 340 * which is maintained sorted by failed disk sector offset 341 * 342 * ASSUMES THE RB_MUTEX IS LOCKED AT ENTRY. */ 343 int 344 rf_CheckForFullRbuf(RF_Raid_t *raidPtr, RF_ReconCtrl_t *reconCtrl, 345 RF_ReconParityStripeStatus_t *pssPtr, int numDataCol) 346 { 347 RF_ReconBuffer_t *p, *pt, *rbuf = (RF_ReconBuffer_t *) pssPtr->rbuf; 348 349 if (rbuf->count == numDataCol) { 350 raidPtr->numFullReconBuffers++; 351 Dprintf2("RECON: rbuf for psid %ld ru %d has filled\n", 352 (long) rbuf->parityStripeID, rbuf->which_ru); 353 if (!reconCtrl->fullBufferList || (rbuf->failedDiskSectorOffset < reconCtrl->fullBufferList->failedDiskSectorOffset)) { 354 Dprintf2("RECON: rbuf for psid %ld ru %d is head of list\n", 355 (long) rbuf->parityStripeID, rbuf->which_ru); 356 rbuf->next = reconCtrl->fullBufferList; 357 reconCtrl->fullBufferList = rbuf; 358 } else { 359 for (pt = reconCtrl->fullBufferList, p = pt->next; p && p->failedDiskSectorOffset < rbuf->failedDiskSectorOffset; pt = p, p = p->next); 360 rbuf->next = p; 361 pt->next = rbuf; 362 Dprintf2("RECON: rbuf for psid %ld ru %d is in list\n", 363 (long) rbuf->parityStripeID, rbuf->which_ru); 364 } 365 rbuf->pssPtr = pssPtr; 366 pssPtr->rbuf = NULL; 367 rf_CauseReconEvent(raidPtr, rbuf->col, NULL, RF_REVENT_BUFREADY); 368 } 369 return (0); 370 } 371 372 373 /* release a floating recon buffer for someone else to use. 374 * assumes the rb_mutex is LOCKED at entry 375 */ 376 void 377 rf_ReleaseFloatingReconBuffer(RF_Raid_t *raidPtr, RF_ReconBuffer_t *rbuf) 378 { 379 RF_ReconCtrl_t *rcPtr = raidPtr->reconControl; 380 RF_CallbackDesc_t *cb; 381 382 Dprintf2("RECON: releasing rbuf for psid %ld ru %d\n", 383 (long) rbuf->parityStripeID, rbuf->which_ru); 384 385 /* if anyone is waiting on buffers, wake one of them up. They will 386 * subsequently wake up anyone else waiting on their RU */ 387 if (rcPtr->bufferWaitList) { 388 rbuf->next = rcPtr->committedRbufs; 389 rcPtr->committedRbufs = rbuf; 390 cb = rcPtr->bufferWaitList; 391 rcPtr->bufferWaitList = cb->next; 392 rf_CauseReconEvent(raidPtr, cb->col, (void *) 1, RF_REVENT_BUFCLEAR); /* arg==1 => we've 393 * committed a buffer */ 394 rf_FreeCallbackDesc(cb); 395 raidPtr->procsInBufWait--; 396 } else { 397 rbuf->next = rcPtr->floatingRbufs; 398 rcPtr->floatingRbufs = rbuf; 399 } 400 } 401