/* $NetBSD: rf_states.c,v 1.50 2016/01/03 08:17:24 mlelstv Exp $ */ /* * Copyright (c) 1995 Carnegie-Mellon University. * All rights reserved. * * Author: Mark Holland, William V. Courtright II, Robby Findler * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. */ #include __KERNEL_RCSID(0, "$NetBSD: rf_states.c,v 1.50 2016/01/03 08:17:24 mlelstv Exp $"); #include #include "rf_archs.h" #include "rf_threadstuff.h" #include "rf_raid.h" #include "rf_dag.h" #include "rf_desc.h" #include "rf_aselect.h" #include "rf_general.h" #include "rf_states.h" #include "rf_dagutils.h" #include "rf_driver.h" #include "rf_engine.h" #include "rf_map.h" #include "rf_etimer.h" #include "rf_kintf.h" #include "rf_paritymap.h" #ifndef RF_DEBUG_STATES #define RF_DEBUG_STATES 0 #endif /* prototypes for some of the available states. States must: - not block. - either schedule rf_ContinueRaidAccess as a callback and return RF_TRUE, or complete all of their work and return RF_FALSE. - increment desc->state when they have finished their work. */ #if RF_DEBUG_STATES static char * StateName(RF_AccessState_t state) { switch (state) { case rf_QuiesceState:return "QuiesceState"; case rf_MapState: return "MapState"; case rf_LockState: return "LockState"; case rf_CreateDAGState: return "CreateDAGState"; case rf_ExecuteDAGState: return "ExecuteDAGState"; case rf_ProcessDAGState: return "ProcessDAGState"; case rf_CleanupState: return "CleanupState"; case rf_LastState: return "LastState"; case rf_IncrAccessesCountState: return "IncrAccessesCountState"; case rf_DecrAccessesCountState: return "DecrAccessesCountState"; default: return "!!! UnnamedState !!!"; } } #endif void rf_ContinueRaidAccess(RF_RaidAccessDesc_t *desc) { int suspended = RF_FALSE; int current_state_index = desc->state; RF_AccessState_t current_state = desc->states[current_state_index]; #if RF_DEBUG_STATES int unit = desc->raidPtr->raidid; #endif do { current_state_index = desc->state; current_state = desc->states[current_state_index]; switch (current_state) { case rf_QuiesceState: suspended = rf_State_Quiesce(desc); break; case rf_IncrAccessesCountState: suspended = rf_State_IncrAccessCount(desc); break; case rf_MapState: suspended = rf_State_Map(desc); break; case rf_LockState: suspended = rf_State_Lock(desc); break; case rf_CreateDAGState: suspended = rf_State_CreateDAG(desc); break; case rf_ExecuteDAGState: suspended = rf_State_ExecuteDAG(desc); break; case rf_ProcessDAGState: suspended = rf_State_ProcessDAG(desc); break; case rf_CleanupState: suspended = rf_State_Cleanup(desc); break; case rf_DecrAccessesCountState: suspended = rf_State_DecrAccessCount(desc); break; case rf_LastState: suspended = rf_State_LastState(desc); break; } /* after this point, we cannot dereference desc since * desc may have been freed. desc is only freed in * LastState, so if we renter this function or loop * back up, desc should be valid. */ #if RF_DEBUG_STATES if (rf_printStatesDebug) { printf("raid%d: State: %-24s StateIndex: %3i desc: 0x%ld %s\n", unit, StateName(current_state), current_state_index, (long) desc, suspended ? "callback scheduled" : "looping"); } #endif } while (!suspended && current_state != rf_LastState); return; } void rf_ContinueDagAccess(RF_DagList_t *dagList) { #if RF_ACC_TRACE > 0 RF_AccTraceEntry_t *tracerec = &(dagList->desc->tracerec); RF_Etimer_t timer; #endif RF_RaidAccessDesc_t *desc; RF_DagHeader_t *dag_h; int i; desc = dagList->desc; #if RF_ACC_TRACE > 0 timer = tracerec->timer; RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); tracerec->specific.user.exec_us = RF_ETIMER_VAL_US(timer); RF_ETIMER_START(tracerec->timer); #endif /* skip to dag which just finished */ dag_h = dagList->dags; for (i = 0; i < dagList->numDagsDone; i++) { dag_h = dag_h->next; } /* check to see if retry is required */ if (dag_h->status == rf_rollBackward) { /* when a dag fails, mark desc status as bad and allow * all other dags in the desc to execute to * completion. then, free all dags and start over */ desc->status = 1; /* bad status */ #if 0 printf("raid%d: DAG failure: %c addr 0x%lx " "(%ld) nblk 0x%x (%d) buf 0x%lx state %d\n", desc->raidPtr->raidid, desc->type, (long) desc->raidAddress, (long) desc->raidAddress, (int) desc->numBlocks, (int) desc->numBlocks, (unsigned long) (desc->bufPtr), desc->state); #endif } dagList->numDagsDone++; rf_ContinueRaidAccess(desc); } int rf_State_LastState(RF_RaidAccessDesc_t *desc) { void (*callbackFunc) (RF_CBParam_t) = desc->callbackFunc; RF_CBParam_t callbackArg; callbackArg.p = desc->callbackArg; /* * We don't support non-async IO. */ KASSERT(desc->async_flag); /* * The parity_map hook has to go here, because the iodone * callback goes straight into the kintf layer. */ if (desc->raidPtr->parity_map != NULL && desc->type == RF_IO_TYPE_WRITE) rf_paritymap_end(desc->raidPtr->parity_map, desc->raidAddress, desc->numBlocks); /* printf("Calling raiddone on 0x%x\n",desc->bp); */ raiddone(desc->raidPtr, desc->bp); /* access came through ioctl */ if (callbackFunc) callbackFunc(callbackArg); rf_FreeRaidAccDesc(desc); return RF_FALSE; } int rf_State_IncrAccessCount(RF_RaidAccessDesc_t *desc) { RF_Raid_t *raidPtr; raidPtr = desc->raidPtr; /* Bummer. We have to do this to be 100% safe w.r.t. the increment * below */ rf_lock_mutex2(raidPtr->access_suspend_mutex); raidPtr->accs_in_flight++; /* used to detect quiescence */ rf_unlock_mutex2(raidPtr->access_suspend_mutex); desc->state++; return RF_FALSE; } int rf_State_DecrAccessCount(RF_RaidAccessDesc_t *desc) { RF_Raid_t *raidPtr; raidPtr = desc->raidPtr; rf_lock_mutex2(raidPtr->access_suspend_mutex); raidPtr->accs_in_flight--; if (raidPtr->accesses_suspended && raidPtr->accs_in_flight == 0) { rf_SignalQuiescenceLock(raidPtr); } rf_unlock_mutex2(raidPtr->access_suspend_mutex); desc->state++; return RF_FALSE; } int rf_State_Quiesce(RF_RaidAccessDesc_t *desc) { #if RF_ACC_TRACE > 0 RF_AccTraceEntry_t *tracerec = &desc->tracerec; RF_Etimer_t timer; #endif RF_CallbackDesc_t *cb; RF_Raid_t *raidPtr; int suspended = RF_FALSE; int need_cb, used_cb; raidPtr = desc->raidPtr; #if RF_ACC_TRACE > 0 RF_ETIMER_START(timer); RF_ETIMER_START(desc->timer); #endif need_cb = 0; used_cb = 0; cb = NULL; rf_lock_mutex2(raidPtr->access_suspend_mutex); /* Do an initial check to see if we might need a callback structure */ if (raidPtr->accesses_suspended) { need_cb = 1; } rf_unlock_mutex2(raidPtr->access_suspend_mutex); if (need_cb) { /* create a callback if we might need it... and we likely do. */ cb = rf_AllocCallbackDesc(); } rf_lock_mutex2(raidPtr->access_suspend_mutex); if (raidPtr->accesses_suspended) { cb->callbackFunc = (void (*) (RF_CBParam_t)) rf_ContinueRaidAccess; cb->callbackArg.p = (void *) desc; cb->next = raidPtr->quiesce_wait_list; raidPtr->quiesce_wait_list = cb; suspended = RF_TRUE; used_cb = 1; } rf_unlock_mutex2(raidPtr->access_suspend_mutex); if ((need_cb == 1) && (used_cb == 0)) { rf_FreeCallbackDesc(cb); } #if RF_ACC_TRACE > 0 RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); tracerec->specific.user.suspend_ovhd_us += RF_ETIMER_VAL_US(timer); #endif #if RF_DEBUG_QUIESCE if (suspended && rf_quiesceDebug) printf("Stalling access due to quiescence lock\n"); #endif desc->state++; return suspended; } int rf_State_Map(RF_RaidAccessDesc_t *desc) { RF_Raid_t *raidPtr = desc->raidPtr; #if RF_ACC_TRACE > 0 RF_AccTraceEntry_t *tracerec = &desc->tracerec; RF_Etimer_t timer; RF_ETIMER_START(timer); #endif if (!(desc->asmap = rf_MapAccess(raidPtr, desc->raidAddress, desc->numBlocks, desc->bufPtr, RF_DONT_REMAP))) RF_PANIC(); #if RF_ACC_TRACE > 0 RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); tracerec->specific.user.map_us = RF_ETIMER_VAL_US(timer); #endif desc->state++; return RF_FALSE; } int rf_State_Lock(RF_RaidAccessDesc_t *desc) { #if RF_ACC_TRACE > 0 RF_AccTraceEntry_t *tracerec = &desc->tracerec; RF_Etimer_t timer; #endif RF_Raid_t *raidPtr = desc->raidPtr; RF_AccessStripeMapHeader_t *asmh = desc->asmap; RF_AccessStripeMap_t *asm_p; RF_StripeNum_t lastStripeID = -1; int suspended = RF_FALSE; #if RF_ACC_TRACE > 0 RF_ETIMER_START(timer); #endif /* acquire each lock that we don't already hold */ for (asm_p = asmh->stripeMap; asm_p; asm_p = asm_p->next) { RF_ASSERT(RF_IO_IS_R_OR_W(desc->type)); if (!rf_suppressLocksAndLargeWrites && asm_p->parityInfo && !(desc->flags & RF_DAG_SUPPRESS_LOCKS) && !(asm_p->flags & RF_ASM_FLAGS_LOCK_TRIED)) { asm_p->flags |= RF_ASM_FLAGS_LOCK_TRIED; /* locks must be acquired hierarchically */ RF_ASSERT(asm_p->stripeID > lastStripeID); lastStripeID = asm_p->stripeID; RF_INIT_LOCK_REQ_DESC(asm_p->lockReqDesc, desc->type, (void (*) (struct buf *)) rf_ContinueRaidAccess, desc, asm_p, raidPtr->Layout.dataSectorsPerStripe); if (rf_AcquireStripeLock(raidPtr->lockTable, asm_p->stripeID, &asm_p->lockReqDesc)) { suspended = RF_TRUE; break; } } if (desc->type == RF_IO_TYPE_WRITE && raidPtr->status == rf_rs_reconstructing) { if (!(asm_p->flags & RF_ASM_FLAGS_FORCE_TRIED)) { int val; asm_p->flags |= RF_ASM_FLAGS_FORCE_TRIED; val = rf_ForceOrBlockRecon(raidPtr, asm_p, (void (*) (RF_Raid_t *, void *)) rf_ContinueRaidAccess, desc); if (val == 0) { asm_p->flags |= RF_ASM_FLAGS_RECON_BLOCKED; } else { suspended = RF_TRUE; break; } } else { #if RF_DEBUG_PSS > 0 if (rf_pssDebug) { printf("raid%d: skipping force/block because already done, psid %ld\n", desc->raidPtr->raidid, (long) asm_p->stripeID); } #endif } } else { #if RF_DEBUG_PSS > 0 if (rf_pssDebug) { printf("raid%d: skipping force/block because not write or not under recon, psid %ld\n", desc->raidPtr->raidid, (long) asm_p->stripeID); } #endif } } #if RF_ACC_TRACE > 0 RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); tracerec->specific.user.lock_us += RF_ETIMER_VAL_US(timer); #endif if (suspended) return (RF_TRUE); desc->state++; return (RF_FALSE); } /* * the following three states create, execute, and post-process dags * the error recovery unit is a single dag. * by default, SelectAlgorithm creates an array of dags, one per parity stripe * in some tricky cases, multiple dags per stripe are created * - dags within a parity stripe are executed sequentially (arbitrary order) * - dags for distinct parity stripes are executed concurrently * * repeat until all dags complete successfully -or- dag selection fails * * while !done * create dag(s) (SelectAlgorithm) * if dag * execute dag (DispatchDAG) * if dag successful * done (SUCCESS) * else * !done (RETRY - start over with new dags) * else * done (FAIL) */ int rf_State_CreateDAG(RF_RaidAccessDesc_t *desc) { #if RF_ACC_TRACE > 0 RF_AccTraceEntry_t *tracerec = &desc->tracerec; RF_Etimer_t timer; #endif RF_DagHeader_t *dag_h; RF_DagList_t *dagList; struct buf *bp; int i, selectStatus; /* generate a dag for the access, and fire it off. When the dag * completes, we'll get re-invoked in the next state. */ #if RF_ACC_TRACE > 0 RF_ETIMER_START(timer); #endif /* SelectAlgorithm returns one or more dags */ selectStatus = rf_SelectAlgorithm(desc, desc->flags | RF_DAG_SUPPRESS_LOCKS); #if RF_DEBUG_VALIDATE_DAG if (rf_printDAGsDebug) { dagList = desc->dagList; for (i = 0; i < desc->numStripes; i++) { rf_PrintDAGList(dagList->dags); dagList = dagList->next; } } #endif /* RF_DEBUG_VALIDATE_DAG */ #if RF_ACC_TRACE > 0 RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); /* update time to create all dags */ tracerec->specific.user.dag_create_us = RF_ETIMER_VAL_US(timer); #endif desc->status = 0; /* good status */ if (selectStatus || (desc->numRetries > RF_RETRY_THRESHOLD)) { /* failed to create a dag */ /* this happens when there are too many faults or incomplete * dag libraries */ if (selectStatus) { printf("raid%d: failed to create a dag. " "Too many component failures.\n", desc->raidPtr->raidid); } else { printf("raid%d: IO failed after %d retries.\n", desc->raidPtr->raidid, RF_RETRY_THRESHOLD); } desc->status = 1; /* bad status */ /* skip straight to rf_State_Cleanup() */ desc->state = rf_CleanupState; bp = (struct buf *)desc->bp; bp->b_error = EIO; bp->b_resid = bp->b_bcount; } else { /* bind dags to desc */ dagList = desc->dagList; for (i = 0; i < desc->numStripes; i++) { dag_h = dagList->dags; while (dag_h) { dag_h->bp = (struct buf *) desc->bp; #if RF_ACC_TRACE > 0 dag_h->tracerec = tracerec; #endif dag_h = dag_h->next; } dagList = dagList->next; } desc->flags |= RF_DAG_DISPATCH_RETURNED; desc->state++; /* next state should be rf_State_ExecuteDAG */ } return RF_FALSE; } /* the access has an list of dagLists, one dagList per parity stripe. * fire the first dag in each parity stripe (dagList). * dags within a stripe (dagList) must be executed sequentially * - this preserves atomic parity update * dags for independents parity groups (stripes) are fired concurrently */ int rf_State_ExecuteDAG(RF_RaidAccessDesc_t *desc) { int i; RF_DagHeader_t *dag_h; RF_DagList_t *dagList; /* next state is always rf_State_ProcessDAG important to do * this before firing the first dag (it may finish before we * leave this routine) */ desc->state++; /* sweep dag array, a stripe at a time, firing the first dag * in each stripe */ dagList = desc->dagList; for (i = 0; i < desc->numStripes; i++) { RF_ASSERT(dagList->numDags > 0); RF_ASSERT(dagList->numDagsDone == 0); RF_ASSERT(dagList->numDagsFired == 0); #if RF_ACC_TRACE > 0 RF_ETIMER_START(dagList->tracerec.timer); #endif /* fire first dag in this stripe */ dag_h = dagList->dags; RF_ASSERT(dag_h); dagList->numDagsFired++; rf_DispatchDAG(dag_h, (void (*) (void *)) rf_ContinueDagAccess, dagList); dagList = dagList->next; } /* the DAG will always call the callback, even if there was no * blocking, so we are always suspended in this state */ return RF_TRUE; } /* rf_State_ProcessDAG is entered when a dag completes. * first, check to all dags in the access have completed * if not, fire as many dags as possible */ int rf_State_ProcessDAG(RF_RaidAccessDesc_t *desc) { RF_AccessStripeMapHeader_t *asmh = desc->asmap; RF_Raid_t *raidPtr = desc->raidPtr; RF_DagHeader_t *dag_h; int i, j, done = RF_TRUE; RF_DagList_t *dagList, *temp; /* check to see if this is the last dag */ dagList = desc->dagList; for (i = 0; i < desc->numStripes; i++) { if (dagList->numDags != dagList->numDagsDone) done = RF_FALSE; dagList = dagList->next; } if (done) { if (desc->status) { /* a dag failed, retry */ /* free all dags */ dagList = desc->dagList; for (i = 0; i < desc->numStripes; i++) { rf_FreeDAG(dagList->dags); temp = dagList; dagList = dagList->next; rf_FreeDAGList(temp); } desc->dagList = NULL; rf_MarkFailuresInASMList(raidPtr, asmh); /* note the retry so that we'll bail in rf_State_CreateDAG() once we've retired the IO RF_RETRY_THRESHOLD times */ desc->numRetries++; /* back up to rf_State_CreateDAG */ desc->state = desc->state - 2; return RF_FALSE; } else { /* move on to rf_State_Cleanup */ desc->state++; } return RF_FALSE; } else { /* more dags to execute */ /* see if any are ready to be fired. if so, fire them */ /* don't fire the initial dag in a list, it's fired in * rf_State_ExecuteDAG */ dagList = desc->dagList; for (i = 0; i < desc->numStripes; i++) { if ((dagList->numDagsDone < dagList->numDags) && (dagList->numDagsDone == dagList->numDagsFired) && (dagList->numDagsFired > 0)) { #if RF_ACC_TRACE > 0 RF_ETIMER_START(dagList->tracerec.timer); #endif /* fire next dag in this stripe */ /* first, skip to next dag awaiting execution */ dag_h = dagList->dags; for (j = 0; j < dagList->numDagsDone; j++) dag_h = dag_h->next; dagList->numDagsFired++; rf_DispatchDAG(dag_h, (void (*) (void *)) rf_ContinueDagAccess, dagList); } dagList = dagList->next; } return RF_TRUE; } } /* only make it this far if all dags complete successfully */ int rf_State_Cleanup(RF_RaidAccessDesc_t *desc) { #if RF_ACC_TRACE > 0 RF_AccTraceEntry_t *tracerec = &desc->tracerec; RF_Etimer_t timer; #endif RF_AccessStripeMapHeader_t *asmh = desc->asmap; RF_Raid_t *raidPtr = desc->raidPtr; RF_AccessStripeMap_t *asm_p; RF_DagList_t *dagList; int i; desc->state++; #if RF_ACC_TRACE > 0 timer = tracerec->timer; RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); tracerec->specific.user.dag_retry_us = RF_ETIMER_VAL_US(timer); /* the RAID I/O is complete. Clean up. */ tracerec->specific.user.dag_retry_us = 0; RF_ETIMER_START(timer); #endif /* free all dags */ dagList = desc->dagList; for (i = 0; i < desc->numStripes; i++) { rf_FreeDAG(dagList->dags); dagList = dagList->next; } #if RF_ACC_TRACE > 0 RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); tracerec->specific.user.cleanup_us = RF_ETIMER_VAL_US(timer); RF_ETIMER_START(timer); #endif for (asm_p = asmh->stripeMap; asm_p; asm_p = asm_p->next) { if (!rf_suppressLocksAndLargeWrites && asm_p->parityInfo && !(desc->flags & RF_DAG_SUPPRESS_LOCKS)) { RF_ASSERT_VALID_LOCKREQ(&asm_p->lockReqDesc); rf_ReleaseStripeLock(raidPtr->lockTable, asm_p->stripeID, &asm_p->lockReqDesc); } if (asm_p->flags & RF_ASM_FLAGS_RECON_BLOCKED) { rf_UnblockRecon(raidPtr, asm_p); } } #if RF_ACC_TRACE > 0 RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); tracerec->specific.user.lock_us += RF_ETIMER_VAL_US(timer); RF_ETIMER_START(timer); #endif rf_FreeAccessStripeMap(asmh); #if RF_ACC_TRACE > 0 RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); tracerec->specific.user.cleanup_us += RF_ETIMER_VAL_US(timer); RF_ETIMER_STOP(desc->timer); RF_ETIMER_EVAL(desc->timer); timer = desc->tracerec.tot_timer; RF_ETIMER_STOP(timer); RF_ETIMER_EVAL(timer); desc->tracerec.total_us = RF_ETIMER_VAL_US(timer); rf_LogTraceRec(raidPtr, tracerec); #endif desc->flags |= RF_DAG_ACCESS_COMPLETE; return RF_FALSE; }