/* called at system boot time */
int
rf_BootRaidframe()
{
int rc;
if (raidframe_booted)
return (EBUSY);
raidframe_booted = 1;
#if RF_DEBUG_ATOMIC > 0
rf_atent_init();
#endif /* RF_DEBUG_ATOMIC > 0 */
rf_setup_threadid();
rf_assign_threadid();
rc = rf_mutex_init(&configureMutex);
if (rc) {
RF_ERRORMSG3("Unable to init mutex file %s line %d rc=%d\n", __FILE__,
__LINE__, rc);
RF_PANIC();
}
configureCount = 0;
isconfigged = 0;
globalShutdown = NULL;
return (0);
}
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: 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 RF_ACC_TRACE > 0
if (dag->tracerec) {
RF_ETIMER_START(dag->tracerec->timer);
}
#endif
#if DEBUG
#if RF_DEBUG_VALIDATE_DAG
if (rf_engineDebug || rf_validateDAGDebug) {
if (rf_ValidateDAG(dag))
RF_PANIC();
}
#endif
#endif
#if RF_DEBUG_ENGINE
if (rf_engineDebug) {
printf("raid%d: Entering DispatchDAG\n", raidPtr->raidid);
}
#endif
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);
}
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 each succedent recursively check branch */
for (i = 0; i < node->numSuccedents; i++)
if (!BranchDone(node->succedents[i]))
return RF_FALSE;
return RF_TRUE; /* node and all succedent branches aren't in
* fired state */
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);
default:
/* illegal node status */
RF_PANIC();
break;
}
}
/* 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_DEBUG_ENGINE
if (rf_engineDebug) {
printf("raid%d: Firing node 0x%lx (%s)\n",
node->dagHdr->raidPtr->raidid,
(unsigned long) node, node->name);
}
#endif
if (node->flags & RF_DAGNODE_FLAG_YIELD) {
#if defined(__NetBSD__) && 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_DEBUG_ENGINE
if (rf_engineDebug) {
printf("raid%d: Firing (undo) node 0x%lx (%s)\n",
node->dagHdr->raidPtr->raidid,
(unsigned long) node, node->name);
}
#endif
if (node->flags & RF_DAGNODE_FLAG_YIELD)
#if defined(__NetBSD__) && 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;
}
}
/*
* 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)
{
RF_AccTraceEntry_t *tracerec = &desc->tracerec;
RF_Etimer_t timer;
RF_DagHeader_t *dag_h;
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.
*/
RF_ETIMER_START(timer);
/* SelectAlgorithm returns one or more dags. */
selectStatus = rf_SelectAlgorithm(desc,
desc->flags | RF_DAG_SUPPRESS_LOCKS);
if (rf_printDAGsDebug)
for (i = 0; i < desc->numStripes; i++)
rf_PrintDAGList(desc->dagArray[i].dags);
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);
desc->status = 0; /* Good status. */
if (selectStatus) {
/* Failed to create a dag. */
/*
* This happens when there are too many faults or incomplete
* dag libraries.
*/
printf("[Failed to create a DAG]\n");
RF_PANIC();
} else {
/* Bind dags to desc. */
for (i = 0; i < desc->numStripes; i++) {
dag_h = desc->dagArray[i].dags;
while (dag_h) {
dag_h->bp = (struct buf *) desc->bp;
dag_h->tracerec = tracerec;
dag_h = dag_h->next;
}
}
desc->flags |= RF_DAG_DISPATCH_RETURNED;
desc->state++; /* Next state should be rf_State_ExecuteDAG. */
}
return RF_FALSE;
}
/*
* 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)) {
/* crossed commit barrier */
node->dagHdr->status = rf_rollForward;
#if RF_DEBUG_ENGINE
if (rf_engineDebug) {
printf("raid%d: node (%s) returned fail, rolling forward\n", raidPtr->raidid, node->name);
}
#endif
} else {
/* never reached commit barrier */
node->dagHdr->status = rf_rollBackward;
#if RF_DEBUG_ENGINE
if (rf_engineDebug) {
printf("raid%d: node (%s) returned fail, rolling backward\n", raidPtr->raidid, node->name);
}
#endif
}
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);
}
/* force the array into reconfigured mode without doing reconstruction */
int
rf_SetReconfiguredMode(RF_Raid_t *raidPtr, int col)
{
if (!(raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE)) {
printf("Can't set reconfigured mode in dedicated-spare array\n");
RF_PANIC();
}
RF_LOCK_MUTEX(raidPtr->mutex);
raidPtr->numFailures++;
raidPtr->Disks[col].status = rf_ds_dist_spared;
raidPtr->status = rf_rs_reconfigured;
rf_update_component_labels(raidPtr, RF_NORMAL_COMPONENT_UPDATE);
/* install spare table only if declustering + distributed sparing
* architecture. */
if (raidPtr->Layout.map->flags & RF_BD_DECLUSTERED)
rf_InstallSpareTable(raidPtr, col);
RF_UNLOCK_MUTEX(raidPtr->mutex);
return (0);
}
int
rf_State_Map(RF_RaidAccessDesc_t *desc)
{
RF_Raid_t *raidPtr = desc->raidPtr;
RF_AccTraceEntry_t *tracerec = &desc->tracerec;
RF_Etimer_t timer;
RF_ETIMER_START(timer);
if (!(desc->asmap = rf_MapAccess(raidPtr, desc->raidAddress,
desc->numBlocks, desc->bufPtr, RF_DONT_REMAP)))
RF_PANIC();
RF_ETIMER_STOP(timer);
RF_ETIMER_EVAL(timer);
tracerec->specific.user.map_us = RF_ETIMER_VAL_US(timer);
desc->state++;
return RF_FALSE;
}
/*
* This function is really just for debugging user-level stuff: it
* frees up all memory, other RAIDframe resources which might otherwise
* be kept around. This is used with systems like "sentinel" to detect
* memory leaks.
*/
int
rf_UnbootRaidframe()
{
int rc;
RF_LOCK_MUTEX(configureMutex);
if (configureCount) {
RF_UNLOCK_MUTEX(configureMutex);
return (EBUSY);
}
raidframe_booted = 0;
RF_UNLOCK_MUTEX(configureMutex);
rc = rf_mutex_destroy(&configureMutex);
if (rc) {
RF_ERRORMSG3("Unable to destroy mutex file %s line %d rc=%d\n", __FILE__,
__LINE__, rc);
RF_PANIC();
}
#if RF_DEBUG_ATOMIC > 0
rf_atent_shutdown();
#endif /* RF_DEBUG_ATOMIC > 0 */
return (0);
}
int
rf_PQDoubleRecoveryFunc(RF_DagNode_t *node)
{
int np = node->numParams;
RF_AccessStripeMap_t *asmap =
(RF_AccessStripeMap_t *) node->params[np - 1].p;
RF_Raid_t *raidPtr = (RF_Raid_t *) node->params[np - 2].p;
RF_RaidLayout_t *layoutPtr = (RF_RaidLayout_t *) & (raidPtr->Layout);
int d, i;
unsigned coeff;
RF_RaidAddr_t sosAddr, suoffset;
RF_SectorCount_t len, secPerSU = layoutPtr->sectorsPerStripeUnit;
int two = 0;
RF_PhysDiskAddr_t *ppda, *ppda2, *qpda, *qpda2, *pda, npda;
char *buf;
int numDataCol = layoutPtr->numDataCol;
RF_Etimer_t timer;
RF_AccTraceEntry_t *tracerec = node->dagHdr->tracerec;
RF_ETIMER_START(timer);
if (asmap->failedPDAs[1] &&
(asmap->failedPDAs[1]->numSector +
asmap->failedPDAs[0]->numSector < secPerSU)) {
RF_ASSERT(0);
ppda = node->params[np - 6].p;
ppda2 = node->params[np - 5].p;
qpda = node->params[np - 4].p;
qpda2 = node->params[np - 3].p;
d = (np - 6);
two = 1;
} else {
ppda = node->params[np - 4].p;
qpda = node->params[np - 3].p;
d = (np - 4);
}
for (i = 0; i < d; i++) {
pda = node->params[i].p;
buf = pda->bufPtr;
suoffset = rf_StripeUnitOffset(layoutPtr, pda->startSector);
len = pda->numSector;
coeff = rf_RaidAddressToStripeUnitID(layoutPtr,
pda->raidAddress);
/* Compute the data unit offset within the column. */
coeff = (coeff % raidPtr->Layout.numDataCol);
/* See if pda intersects a recovery pda. */
rf_applyPDA(raidPtr, pda, ppda, qpda, node->dagHdr->bp);
if (two)
rf_applyPDA(raidPtr, pda, ppda, qpda, node->dagHdr->bp);
}
/*
* Ok, we got the parity back to the point where we can recover. We
* now need to determine the coeff of the columns that need to be
* recovered. We can also only need to recover a single stripe unit.
*/
if (asmap->failedPDAs[1] == NULL) { /*
* Only a single stripe unit
* to recover.
*/
pda = asmap->failedPDAs[0];
sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr,
asmap->raidAddress);
/* Need to determine the column of the other failed disk. */
coeff = rf_RaidAddressToStripeUnitID(layoutPtr,
pda->raidAddress);
/* Compute the data unit offset within the column. */
coeff = (coeff % raidPtr->Layout.numDataCol);
for (i = 0; i < numDataCol; i++) {
npda.raidAddress = sosAddr + (i * secPerSU);
(raidPtr->Layout.map->MapSector) (raidPtr,
npda.raidAddress, &(npda.row), &(npda.col),
&(npda.startSector), 0);
/* Skip over dead disks. */
if (RF_DEAD_DISK(raidPtr->Disks[npda.row][npda.col]
.status))
if (i != coeff)
break;
}
RF_ASSERT(i < numDataCol);
RF_ASSERT(two == 0);
/*
* Recover the data. Since we need only to recover one
* column, we overwrite the parity with the other one.
*/
if (coeff < i) /* Recovering 'a'. */
rf_PQ_recover((unsigned long *) ppda->bufPtr,
(unsigned long *) qpda->bufPtr,
(unsigned long *) pda->bufPtr,
(unsigned long *) ppda->bufPtr,
rf_RaidAddressToByte(raidPtr, pda->numSector),
coeff, i);
else /* Recovering 'b'. */
rf_PQ_recover((unsigned long *) ppda->bufPtr,
(unsigned long *) qpda->bufPtr,
(unsigned long *) ppda->bufPtr,
(unsigned long *) pda->bufPtr,
rf_RaidAddressToByte(raidPtr, pda->numSector),
i, coeff);
} else
RF_PANIC();
RF_ETIMER_STOP(timer);
RF_ETIMER_EVAL(timer);
if (tracerec)
tracerec->q_us += RF_ETIMER_VAL_US(timer);
rf_GenericWakeupFunc(node, 0);
return (0);
}
static void
DAGExecutionThread(RF_ThreadArg_t arg)
{
RF_DagNode_t *nd, *local_nq, *term_nq, *fire_nq;
RF_Raid_t *raidPtr;
int ks;
int s;
raidPtr = (RF_Raid_t *) arg;
#if RF_DEBUG_ENGINE
if (rf_engineDebug) {
printf("raid%d: Engine thread is running\n", raidPtr->raidid);
}
#endif
s = splbio();
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);
splx(s);
kthread_exit(0);
}
/* 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 */
/* put node on
a list to
be fired
after we
unlock */
s->next = firelist;
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 */
/* put node on a list to be
fired after we unlock */
a->next = firelist;
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;
}
}
void
rf_CreateNonredundantDAG(
RF_Raid_t *raidPtr,
RF_AccessStripeMap_t *asmap,
RF_DagHeader_t *dag_h,
void *bp,
RF_RaidAccessFlags_t flags,
RF_AllocListElem_t *allocList,
RF_IoType_t type
)
{
RF_DagNode_t *nodes, *diskNodes, *blockNode, *commitNode, *termNode;
RF_PhysDiskAddr_t *pda = asmap->physInfo;
int (*doFunc) (RF_DagNode_t *), (*undoFunc) (RF_DagNode_t *);
int i, n, totalNumNodes;
char *name;
n = asmap->numStripeUnitsAccessed;
dag_h->creator = "NonredundantDAG";
RF_ASSERT(RF_IO_IS_R_OR_W(type));
switch (type) {
case RF_IO_TYPE_READ:
doFunc = rf_DiskReadFunc;
undoFunc = rf_DiskReadUndoFunc;
name = "R ";
if (rf_dagDebug)
printf("[Creating non-redundant read DAG]\n");
break;
case RF_IO_TYPE_WRITE:
doFunc = rf_DiskWriteFunc;
undoFunc = rf_DiskWriteUndoFunc;
name = "W ";
if (rf_dagDebug)
printf("[Creating non-redundant write DAG]\n");
break;
default:
RF_PANIC();
}
/*
* For reads, the dag can not commit until the block node is reached.
* For writes, the dag commits immediately.
*/
dag_h->numCommitNodes = 1;
dag_h->numCommits = 0;
dag_h->numSuccedents = 1;
/*
* Node count:
* 1 block node
* n data reads (or writes)
* 1 commit node
* 1 terminator node
*/
RF_ASSERT(n > 0);
totalNumNodes = n + 3;
RF_CallocAndAdd(nodes, totalNumNodes, sizeof(RF_DagNode_t),
(RF_DagNode_t *), allocList);
i = 0;
diskNodes = &nodes[i];
i += n;
blockNode = &nodes[i];
i += 1;
commitNode = &nodes[i];
i += 1;
termNode = &nodes[i];
i += 1;
RF_ASSERT(i == totalNumNodes);
/* Initialize nodes. */
switch (type) {
case RF_IO_TYPE_READ:
rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, n, 0, 0, 0, dag_h, "Nil",
allocList);
rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, 1, n, 0, 0, dag_h, "Cmt",
allocList);
rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc,
rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm",
allocList);
break;
case RF_IO_TYPE_WRITE:
rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, 1, 0, 0, 0, dag_h, "Nil",
allocList);
rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, n, 1, 0, 0, dag_h, "Cmt",
allocList);
rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc,
rf_TerminateUndoFunc, NULL, 0, n, 0, 0, dag_h, "Trm",
allocList);
break;
default:
RF_PANIC();
}
for (i = 0; i < n; i++) {
RF_ASSERT(pda != NULL);
rf_InitNode(&diskNodes[i], rf_wait, RF_FALSE, doFunc, undoFunc,
rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList);
diskNodes[i].params[0].p = pda;
diskNodes[i].params[1].p = pda->bufPtr;
/* Parity stripe id is not necessary. */
diskNodes[i].params[2].v = 0;
diskNodes[i].params[3].v =
RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, 0);
pda = pda->next;
}
/*
* Connect nodes.
*/
/* Connect hdr to block node. */
RF_ASSERT(blockNode->numAntecedents == 0);
dag_h->succedents[0] = blockNode;
if (type == RF_IO_TYPE_READ) {
/* Connecting a nonredundant read DAG. */
RF_ASSERT(blockNode->numSuccedents == n);
RF_ASSERT(commitNode->numAntecedents == n);
for (i = 0; i < n; i++) {
/* Connect block node to each read node. */
RF_ASSERT(diskNodes[i].numAntecedents == 1);
blockNode->succedents[i] = &diskNodes[i];
diskNodes[i].antecedents[0] = blockNode;
diskNodes[i].antType[0] = rf_control;
/* Connect each read node to the commit node. */
RF_ASSERT(diskNodes[i].numSuccedents == 1);
diskNodes[i].succedents[0] = commitNode;
commitNode->antecedents[i] = &diskNodes[i];
commitNode->antType[i] = rf_control;
}
/* Connect the commit node to the term node. */
RF_ASSERT(commitNode->numSuccedents == 1);
RF_ASSERT(termNode->numAntecedents == 1);
RF_ASSERT(termNode->numSuccedents == 0);
commitNode->succedents[0] = termNode;
termNode->antecedents[0] = commitNode;
termNode->antType[0] = rf_control;
} else {
/* Connecting a nonredundant write DAG. */
/* Connect the block node to the commit node. */
RF_ASSERT(blockNode->numSuccedents == 1);
RF_ASSERT(commitNode->numAntecedents == 1);
blockNode->succedents[0] = commitNode;
commitNode->antecedents[0] = blockNode;
commitNode->antType[0] = rf_control;
RF_ASSERT(commitNode->numSuccedents == n);
RF_ASSERT(termNode->numAntecedents == n);
RF_ASSERT(termNode->numSuccedents == 0);
for (i = 0; i < n; i++) {
/* Connect the commit node to each write node. */
RF_ASSERT(diskNodes[i].numAntecedents == 1);
commitNode->succedents[i] = &diskNodes[i];
diskNodes[i].antecedents[0] = commitNode;
diskNodes[i].antType[0] = rf_control;
/* Connect each write node to the term node. */
RF_ASSERT(diskNodes[i].numSuccedents == 1);
diskNodes[i].succedents[0] = termNode;
termNode->antecedents[i] = &diskNodes[i];
termNode->antType[i] = rf_control;
}
}
}