/* $FreeBSD$ */ /* $NetBSD: rf_engine.c,v 1.10 2000/08/20 16:51:03 thorpej Exp $ */ /* * Copyright (c) 1995 Carnegie-Mellon University. * All rights reserved. * * Author: William V. Courtright II, Mark Holland, Rachad Youssef * * 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. */ /**************************************************************************** * * * engine.c -- code for DAG execution engine * * * * Modified to work as follows (holland): * * A user-thread calls into DispatchDAG, which fires off the nodes that * * are direct successors to the header node. DispatchDAG then returns, * * and the rest of the I/O continues asynchronously. As each node * * completes, the node execution function calls FinishNode(). FinishNode * * scans the list of successors to the node and increments the antecedent * * counts. Each node that becomes enabled is placed on a central node * * queue. A dedicated dag-execution thread grabs nodes off of this * * queue and fires them. * * * * NULL nodes are never fired. * * * * Terminator nodes are never fired, but rather cause the callback * * associated with the DAG to be invoked. * * * * If a node fails, the dag either rolls forward to the completion or * * rolls back, undoing previously-completed nodes and fails atomically. * * The direction of recovery is determined by the location of the failed * * node in the graph. If the failure occured before the commit node in * * the graph, backward recovery is used. Otherwise, forward recovery is * * used. * * * ****************************************************************************/ #include #include #include #include #include #include #include #include #include #include static void DAGExecutionThread(RF_ThreadArg_t arg); #define DO_INIT(_l_,_r_) { \ int _rc; \ _rc = rf_create_managed_mutex(_l_,&(_r_)->node_queue_mutex); \ if (_rc) { \ return(_rc); \ } \ _rc = rf_create_managed_cond(_l_,&(_r_)->node_queue_cond); \ if (_rc) { \ return(_rc); \ } \ } /* synchronization primitives for this file. DO_WAIT should be enclosed in a while loop. */ /* * XXX Is this spl-ing really necessary? */ #define DO_LOCK(_r_) \ do { \ ks = splbio(); \ RF_LOCK_MUTEX((_r_)->node_queue_mutex); \ } while (0) #define DO_UNLOCK(_r_) \ do { \ RF_UNLOCK_MUTEX((_r_)->node_queue_mutex); \ splx(ks); \ } while (0) #define DO_WAIT(_r_) \ RF_WAIT_COND((_r_)->node_queue, (_r_)->node_queue_mutex) #define DO_SIGNAL(_r_) \ RF_BROADCAST_COND((_r_)->node_queue) /* XXX RF_SIGNAL_COND? */ static void rf_ShutdownEngine(void *); static void rf_ShutdownEngine(arg) void *arg; { RF_Raid_t *raidPtr; raidPtr = (RF_Raid_t *) arg; raidPtr->shutdown_engine = 1; DO_SIGNAL(raidPtr); } int rf_ConfigureEngine( RF_ShutdownList_t ** listp, RF_Raid_t * raidPtr, RF_Config_t * cfgPtr) { int rc; DO_INIT(listp, raidPtr); raidPtr->node_queue = NULL; raidPtr->dags_in_flight = 0; rc = rf_init_managed_threadgroup(listp, &raidPtr->engine_tg); if (rc) return (rc); /* we create the execution thread only once per system boot. no need * to check return code b/c the kernel panics if it can't create the * thread. */ if (rf_engineDebug) { printf("raid%d: Creating engine thread\n", raidPtr->raidid); } if (RF_CREATE_THREAD(raidPtr->engine_thread, DAGExecutionThread, raidPtr,"raid")) { RF_ERRORMSG("RAIDFRAME: Unable to create engine thread\n"); return (ENOMEM); } if (rf_engineDebug) { printf("raid%d: Created engine thread\n", raidPtr->raidid); } RF_THREADGROUP_STARTED(&raidPtr->engine_tg); /* XXX something is missing here... */ #ifdef debug printf("Skipping the WAIT_START!!\n"); #endif #if 1 printf("Waiting for DAG engine to start\n"); RF_THREADGROUP_WAIT_START(&raidPtr->engine_tg); #endif /* engine thread is now running and waiting for work */ if (rf_engineDebug) { printf("raid%d: Engine thread running and waiting for events\n", raidPtr->raidid); } rc = rf_ShutdownCreate(listp, rf_ShutdownEngine, raidPtr); if (rc) { RF_ERRORMSG3("Unable to add to shutdown list file %s line %d rc=%d\n", __FILE__, __LINE__, rc); rf_ShutdownEngine(NULL); } return (rc); } static int BranchDone(RF_DagNode_t * node) { int i; /* return true if forward execution is completed for a node and it's * succedents */ switch (node->status) { case rf_wait: /* should never be called in this state */ RF_PANIC(); break; case rf_fired: /* node is currently executing, so we're not done */ return (RF_FALSE); case rf_good: for (i = 0; i < node->numSuccedents; i++) /* for each succedent */ if (!BranchDone(node->succedents[i])) /* recursively check * branch */ return RF_FALSE; return RF_TRUE; /* node and all succedent branches aren't in * fired state */ break; case rf_bad: /* succedents can't fire */ return (RF_TRUE); case rf_recover: /* should never be called in this state */ RF_PANIC(); break; case rf_undone: case rf_panic: /* XXX need to fix this case */ /* for now, assume that we're done */ return (RF_TRUE); break; default: /* illegal node status */ RF_PANIC(); break; } } static int NodeReady(RF_DagNode_t * node) { int ready; switch (node->dagHdr->status) { case rf_enable: case rf_rollForward: if ((node->status == rf_wait) && (node->numAntecedents == node->numAntDone)) ready = RF_TRUE; else ready = RF_FALSE; break; case rf_rollBackward: RF_ASSERT(node->numSuccDone <= node->numSuccedents); RF_ASSERT(node->numSuccFired <= node->numSuccedents); RF_ASSERT(node->numSuccFired <= node->numSuccDone); if ((node->status == rf_good) && (node->numSuccDone == node->numSuccedents)) ready = RF_TRUE; else ready = RF_FALSE; break; default: printf("Execution engine found illegal DAG status in NodeReady\n"); RF_PANIC(); break; } return (ready); } /* user context and dag-exec-thread context: * Fire a node. The node's status field determines which function, do or undo, * to be fired. * This routine assumes that the node's status field has alread been set to * "fired" or "recover" to indicate the direction of execution. */ static void FireNode(RF_DagNode_t * node) { switch (node->status) { case rf_fired: /* fire the do function of a node */ if (rf_engineDebug) { printf("raid%d: Firing node 0x%lx (%s)\n", node->dagHdr->raidPtr->raidid, (unsigned long) node, node->name); } if (node->flags & RF_DAGNODE_FLAG_YIELD) { #if defined(__NetBSD__) || defined(__FreeBSD__) && defined(_KERNEL) /* thread_block(); */ /* printf("Need to block the thread here...\n"); */ /* XXX thread_block is actually mentioned in * /usr/include/vm/vm_extern.h */ #else thread_block(); #endif } (*(node->doFunc)) (node); break; case rf_recover: /* fire the undo function of a node */ if (rf_engineDebug) { printf("raid%d: Firing (undo) node 0x%lx (%s)\n", node->dagHdr->raidPtr->raidid, (unsigned long) node, node->name); } if (node->flags & RF_DAGNODE_FLAG_YIELD) #if defined(__NetBSD__) || defined(__FreeBSD__) && defined(_KERNEL) /* thread_block(); */ /* printf("Need to block the thread here...\n"); */ /* XXX thread_block is actually mentioned in * /usr/include/vm/vm_extern.h */ #else thread_block(); #endif (*(node->undoFunc)) (node); break; default: RF_PANIC(); break; } } /* user context: * Attempt to fire each node in a linear array. * The entire list is fired atomically. */ static void FireNodeArray( int numNodes, RF_DagNode_t ** nodeList) { RF_DagStatus_t dstat; RF_DagNode_t *node; int i, j; /* first, mark all nodes which are ready to be fired */ for (i = 0; i < numNodes; i++) { node = nodeList[i]; dstat = node->dagHdr->status; RF_ASSERT((node->status == rf_wait) || (node->status == rf_good)); if (NodeReady(node)) { if ((dstat == rf_enable) || (dstat == rf_rollForward)) { RF_ASSERT(node->status == rf_wait); if (node->commitNode) node->dagHdr->numCommits++; node->status = rf_fired; for (j = 0; j < node->numAntecedents; j++) node->antecedents[j]->numSuccFired++; } else { RF_ASSERT(dstat == rf_rollBackward); RF_ASSERT(node->status == rf_good); RF_ASSERT(node->commitNode == RF_FALSE); /* only one commit node * per graph */ node->status = rf_recover; } } } /* now, fire the nodes */ for (i = 0; i < numNodes; i++) { if ((nodeList[i]->status == rf_fired) || (nodeList[i]->status == rf_recover)) FireNode(nodeList[i]); } } /* user context: * Attempt to fire each node in a linked list. * The entire list is fired atomically. */ static void FireNodeList(RF_DagNode_t * nodeList) { RF_DagNode_t *node, *next; RF_DagStatus_t dstat; int j; if (nodeList) { /* first, mark all nodes which are ready to be fired */ for (node = nodeList; node; node = next) { next = node->next; dstat = node->dagHdr->status; RF_ASSERT((node->status == rf_wait) || (node->status == rf_good)); if (NodeReady(node)) { if ((dstat == rf_enable) || (dstat == rf_rollForward)) { RF_ASSERT(node->status == rf_wait); if (node->commitNode) node->dagHdr->numCommits++; node->status = rf_fired; for (j = 0; j < node->numAntecedents; j++) node->antecedents[j]->numSuccFired++; } else { RF_ASSERT(dstat == rf_rollBackward); RF_ASSERT(node->status == rf_good); RF_ASSERT(node->commitNode == RF_FALSE); /* only one commit node * per graph */ node->status = rf_recover; } } } /* now, fire the nodes */ for (node = nodeList; node; node = next) { next = node->next; if ((node->status == rf_fired) || (node->status == rf_recover)) FireNode(node); } } } /* interrupt context: * for each succedent * propagate required results from node to succedent * increment succedent's numAntDone * place newly-enable nodes on node queue for firing * * To save context switches, we don't place NIL nodes on the node queue, * but rather just process them as if they had fired. Note that NIL nodes * that are the direct successors of the header will actually get fired by * DispatchDAG, which is fine because no context switches are involved. * * Important: when running at user level, this can be called by any * disk thread, and so the increment and check of the antecedent count * must be locked. I used the node queue mutex and locked down the * entire function, but this is certainly overkill. */ static void PropagateResults( RF_DagNode_t * node, int context) { RF_DagNode_t *s, *a; RF_Raid_t *raidPtr; int i, ks; RF_DagNode_t *finishlist = NULL; /* a list of NIL nodes to be * finished */ RF_DagNode_t *skiplist = NULL; /* list of nodes with failed truedata * antecedents */ RF_DagNode_t *firelist = NULL; /* a list of nodes to be fired */ RF_DagNode_t *q = NULL, *qh = NULL, *next; int j, skipNode; raidPtr = node->dagHdr->raidPtr; DO_LOCK(raidPtr); /* debug - validate fire counts */ for (i = 0; i < node->numAntecedents; i++) { a = *(node->antecedents + i); RF_ASSERT(a->numSuccFired >= a->numSuccDone); RF_ASSERT(a->numSuccFired <= a->numSuccedents); a->numSuccDone++; } switch (node->dagHdr->status) { case rf_enable: case rf_rollForward: for (i = 0; i < node->numSuccedents; i++) { s = *(node->succedents + i); RF_ASSERT(s->status == rf_wait); (s->numAntDone)++; if (s->numAntDone == s->numAntecedents) { /* look for NIL nodes */ if (s->doFunc == rf_NullNodeFunc) { /* don't fire NIL nodes, just process * them */ s->next = finishlist; finishlist = s; } else { /* look to see if the node is to be * skipped */ skipNode = RF_FALSE; for (j = 0; j < s->numAntecedents; j++) if ((s->antType[j] == rf_trueData) && (s->antecedents[j]->status == rf_bad)) skipNode = RF_TRUE; if (skipNode) { /* this node has one or more * failed true data * dependencies, so skip it */ s->next = skiplist; skiplist = s; } else /* add s to list of nodes (q) * to execute */ if (context != RF_INTR_CONTEXT) { /* we only have to * enqueue if we're at * intr context */ s->next = firelist; /* put node on a list to * be fired after we * unlock */ firelist = s; } else { /* enqueue the node for * the dag exec thread * to fire */ RF_ASSERT(NodeReady(s)); if (q) { q->next = s; q = s; } else { qh = q = s; qh->next = NULL; } } } } } if (q) { /* xfer our local list of nodes to the node queue */ q->next = raidPtr->node_queue; raidPtr->node_queue = qh; DO_SIGNAL(raidPtr); } DO_UNLOCK(raidPtr); for (; skiplist; skiplist = next) { next = skiplist->next; skiplist->status = rf_skipped; for (i = 0; i < skiplist->numAntecedents; i++) { skiplist->antecedents[i]->numSuccFired++; } if (skiplist->commitNode) { skiplist->dagHdr->numCommits++; } rf_FinishNode(skiplist, context); } for (; finishlist; finishlist = next) { /* NIL nodes: no need to fire them */ next = finishlist->next; finishlist->status = rf_good; for (i = 0; i < finishlist->numAntecedents; i++) { finishlist->antecedents[i]->numSuccFired++; } if (finishlist->commitNode) finishlist->dagHdr->numCommits++; /* * Okay, here we're calling rf_FinishNode() on nodes that * have the null function as their work proc. Such a node * could be the terminal node in a DAG. If so, it will * cause the DAG to complete, which will in turn free * memory used by the DAG, which includes the node in * question. Thus, we must avoid referencing the node * at all after calling rf_FinishNode() on it. */ rf_FinishNode(finishlist, context); /* recursive call */ } /* fire all nodes in firelist */ FireNodeList(firelist); break; case rf_rollBackward: for (i = 0; i < node->numAntecedents; i++) { a = *(node->antecedents + i); RF_ASSERT(a->status == rf_good); RF_ASSERT(a->numSuccDone <= a->numSuccedents); RF_ASSERT(a->numSuccDone <= a->numSuccFired); if (a->numSuccDone == a->numSuccFired) { if (a->undoFunc == rf_NullNodeFunc) { /* don't fire NIL nodes, just process * them */ a->next = finishlist; finishlist = a; } else { if (context != RF_INTR_CONTEXT) { /* we only have to enqueue if * we're at intr context */ a->next = firelist; /* put node on a list to * be fired after we * unlock */ firelist = a; } else { /* enqueue the node for * the dag exec thread * to fire */ RF_ASSERT(NodeReady(a)); if (q) { q->next = a; q = a; } else { qh = q = a; qh->next = NULL; } } } } } if (q) { /* xfer our local list of nodes to the node queue */ q->next = raidPtr->node_queue; raidPtr->node_queue = qh; DO_SIGNAL(raidPtr); } DO_UNLOCK(raidPtr); for (; finishlist; finishlist = next) { /* NIL nodes: no need to * fire them */ next = finishlist->next; finishlist->status = rf_good; /* * Okay, here we're calling rf_FinishNode() on nodes that * have the null function as their work proc. Such a node * could be the first node in a DAG. If so, it will * cause the DAG to complete, which will in turn free * memory used by the DAG, which includes the node in * question. Thus, we must avoid referencing the node * at all after calling rf_FinishNode() on it. */ rf_FinishNode(finishlist, context); /* recursive call */ } /* fire all nodes in firelist */ FireNodeList(firelist); break; default: printf("Engine found illegal DAG status in PropagateResults()\n"); RF_PANIC(); break; } } /* * Process a fired node which has completed */ static void ProcessNode( RF_DagNode_t * node, int context) { RF_Raid_t *raidPtr; raidPtr = node->dagHdr->raidPtr; switch (node->status) { case rf_good: /* normal case, don't need to do anything */ break; case rf_bad: if ((node->dagHdr->numCommits > 0) || (node->dagHdr->numCommitNodes == 0)) { node->dagHdr->status = rf_rollForward; /* crossed commit * barrier */ if (rf_engineDebug || 1) { printf("raid%d: node (%s) returned fail, rolling forward\n", raidPtr->raidid, node->name); } } else { node->dagHdr->status = rf_rollBackward; /* never reached commit * barrier */ if (rf_engineDebug || 1) { printf("raid%d: node (%s) returned fail, rolling backward\n", raidPtr->raidid, node->name); } } break; case rf_undone: /* normal rollBackward case, don't need to do anything */ break; case rf_panic: /* an undo node failed!!! */ printf("UNDO of a node failed!!!/n"); break; default: printf("node finished execution with an illegal status!!!\n"); RF_PANIC(); break; } /* enqueue node's succedents (antecedents if rollBackward) for * execution */ PropagateResults(node, context); } /* user context or dag-exec-thread context: * This is the first step in post-processing a newly-completed node. * This routine is called by each node execution function to mark the node * as complete and fire off any successors that have been enabled. */ int rf_FinishNode( RF_DagNode_t * node, int context) { /* as far as I can tell, retcode is not used -wvcii */ int retcode = RF_FALSE; node->dagHdr->numNodesCompleted++; ProcessNode(node, context); return (retcode); } /* user context: * submit dag for execution, return non-zero if we have to wait for completion. * if and only if we return non-zero, we'll cause cbFunc to get invoked with * cbArg when the DAG has completed. * * for now we always return 1. If the DAG does not cause any I/O, then the callback * may get invoked before DispatchDAG returns. There's code in state 5 of ContinueRaidAccess * to handle this. * * All we do here is fire the direct successors of the header node. The * DAG execution thread does the rest of the dag processing. */ int rf_DispatchDAG( RF_DagHeader_t * dag, void (*cbFunc) (void *), void *cbArg) { RF_Raid_t *raidPtr; raidPtr = dag->raidPtr; if (dag->tracerec) { RF_ETIMER_START(dag->tracerec->timer); } if (rf_engineDebug || rf_validateDAGDebug) { if (rf_ValidateDAG(dag)) RF_PANIC(); } if (rf_engineDebug) { printf("raid%d: Entering DispatchDAG\n", raidPtr->raidid); } raidPtr->dags_in_flight++; /* debug only: blow off proper * locking */ dag->cbFunc = cbFunc; dag->cbArg = cbArg; dag->numNodesCompleted = 0; dag->status = rf_enable; FireNodeArray(dag->numSuccedents, dag->succedents); return (1); } /* dedicated kernel thread: * the thread that handles all DAG node firing. * To minimize locking and unlocking, we grab a copy of the entire node queue and then set the * node queue to NULL before doing any firing of nodes. This way we only have to release the * lock once. Of course, it's probably rare that there's more than one node in the queue at * any one time, but it sometimes happens. * * In the kernel, this thread runs at spl0 and is not swappable. I copied these * characteristics from the aio_completion_thread. */ static void DAGExecutionThread(RF_ThreadArg_t arg) { RF_DagNode_t *nd, *local_nq, *term_nq, *fire_nq; RF_Raid_t *raidPtr; int ks; raidPtr = (RF_Raid_t *) arg; if (rf_engineDebug) { printf("raid%d: Engine thread is running\n", raidPtr->raidid); } mtx_lock(&Giant); RF_THREADGROUP_RUNNING(&raidPtr->engine_tg); DO_LOCK(raidPtr); while (!raidPtr->shutdown_engine) { while (raidPtr->node_queue != NULL) { local_nq = raidPtr->node_queue; fire_nq = NULL; term_nq = NULL; raidPtr->node_queue = NULL; DO_UNLOCK(raidPtr); /* first, strip out the terminal nodes */ while (local_nq) { nd = local_nq; local_nq = local_nq->next; switch (nd->dagHdr->status) { case rf_enable: case rf_rollForward: if (nd->numSuccedents == 0) { /* end of the dag, add to * callback list */ nd->next = term_nq; term_nq = nd; } else { /* not the end, add to the * fire queue */ nd->next = fire_nq; fire_nq = nd; } break; case rf_rollBackward: if (nd->numAntecedents == 0) { /* end of the dag, add to the * callback list */ nd->next = term_nq; term_nq = nd; } else { /* not the end, add to the * fire queue */ nd->next = fire_nq; fire_nq = nd; } break; default: RF_PANIC(); break; } } /* execute callback of dags which have reached the * terminal node */ while (term_nq) { nd = term_nq; term_nq = term_nq->next; nd->next = NULL; (nd->dagHdr->cbFunc) (nd->dagHdr->cbArg); raidPtr->dags_in_flight--; /* debug only */ } /* fire remaining nodes */ FireNodeList(fire_nq); DO_LOCK(raidPtr); } while (!raidPtr->shutdown_engine && raidPtr->node_queue == NULL) DO_WAIT(raidPtr); } DO_UNLOCK(raidPtr); RF_THREADGROUP_DONE(&raidPtr->engine_tg); RF_THREAD_EXIT(0); }