/**
* low-level thread entry point.
*
* @param[in] rock opaque pointer to thread worker object
*
* @return opaque return pointer from pool entry function
*
* @internal
*/
static void *
_afs_tp_worker_run(void * rock)
{
struct afs_thread_pool_worker * worker = rock;
struct afs_thread_pool * pool = worker->pool;
/* register worker with pool */
MUTEX_ENTER(&pool->lock);
queue_Append(&pool->thread_list, worker);
pool->nthreads++;
MUTEX_EXIT(&pool->lock);
/* call high-level entry point */
worker->ret = (*pool->entry)(pool, worker, pool->work_queue, pool->rock);
/* adjust pool live thread count */
MUTEX_ENTER(&pool->lock);
osi_Assert(pool->nthreads);
queue_Remove(worker);
pool->nthreads--;
if (!pool->nthreads) {
CV_BROADCAST(&pool->shutdown_cv);
pool->state = AFS_TP_STATE_STOPPED;
}
MUTEX_EXIT(&pool->lock);
_afs_tp_worker_free(worker);
return NULL;
}
/**
* wakeup all threads waiting in dequeue.
*
* @param[in] list list object
*
* @return operation status
* @retval 0 success
*
* @internal
*/
static int
_afs_wq_node_list_shutdown(struct afs_work_queue_node_list * list)
{
int ret = 0;
struct afs_work_queue_node *node, *nnode;
MUTEX_ENTER(&list->lock);
list->shutdown = 1;
for (queue_Scan(&list->list, node, nnode, afs_work_queue_node)) {
_afs_wq_node_state_change(node, AFS_WQ_NODE_STATE_ERROR);
queue_Remove(node);
node->qidx = AFS_WQ_NODE_LIST_NONE;
node->queue = NULL;
if (node->detached) {
/* if we are detached, we hold the reference on the node;
* otherwise, it is some other caller that holds the reference.
* So don't put the node if we are not detached; the node will
* get freed when someone else calls afs_wq_node_put */
afs_wq_node_put(node);
}
}
CV_BROADCAST(&list->cv);
MUTEX_EXIT(&list->lock);
return ret;
}
/*
* thread to combine salvager child logs
* back into the main salvageserver log
*/
static void *
SalvageLogCleanupThread(void * arg)
{
struct log_cleanup_node * cleanup;
assert(pthread_mutex_lock(&worker_lock) == 0);
while (1) {
while (queue_IsEmpty(&log_cleanup_queue)) {
assert(pthread_cond_wait(&log_cleanup_queue.queue_change_cv, &worker_lock) == 0);
}
while (queue_IsNotEmpty(&log_cleanup_queue)) {
cleanup = queue_First(&log_cleanup_queue, log_cleanup_node);
queue_Remove(cleanup);
assert(pthread_mutex_unlock(&worker_lock) == 0);
SalvageLogCleanup(cleanup->pid);
free(cleanup);
assert(pthread_mutex_lock(&worker_lock) == 0);
}
}
assert(pthread_mutex_unlock(&worker_lock) == 0);
return NULL;
}
/*
* thread to combine salvager child logs
* back into the main salvageserver log
*/
static void *
SalvageLogCleanupThread(void * arg)
{
struct log_cleanup_node * cleanup;
MUTEX_ENTER(&worker_lock);
while (1) {
while (queue_IsEmpty(&log_cleanup_queue)) {
CV_WAIT(&log_cleanup_queue.queue_change_cv, &worker_lock);
}
while (queue_IsNotEmpty(&log_cleanup_queue)) {
cleanup = queue_First(&log_cleanup_queue, log_cleanup_node);
queue_Remove(cleanup);
MUTEX_EXIT(&worker_lock);
SalvageLogCleanup(cleanup->pid);
free(cleanup);
MUTEX_ENTER(&worker_lock);
}
}
MUTEX_EXIT(&worker_lock);
return NULL;
}
/*
* Process all events that have expired relative to the current clock time
* (which is not re-evaluated unless clock_NewTime has been called).
* The relative time to the next event is returned in the output parameter
* next and the function returns 1. If there are is no next event,
* the function returns 0.
*/
int
rxevent_RaiseEvents(struct clock * next)
{
struct rxevent *qe;
struct clock now;
#ifdef RXDEBUG
if (Log)
fprintf(Log, "rxevent_RaiseEvents(%ld.%ld)\n", now.sec, now.usec);
#endif
/*
* Events are sorted by time, so only scan until an event is found that
* has not yet timed out
*/
while (queue_IsNotEmpty(&rxevent_queue)) {
clock_GetTime(&now);
qe = queue_First(&rxevent_queue, rxevent);
if (clock_Lt(&now, &qe->eventTime)) {
*next = qe->eventTime;
clock_Sub(next, &now);
return 1;
}
queue_Remove(qe);
rxevent_nPosted--;
qe->func(qe, qe->arg, qe->arg1);
queue_Append(&rxevent_free, qe);
rxevent_nFree++;
}
return 0;
}
static void cleanup_pthread_cache(void) {
thread_p cur = NULL, next = NULL;
if (pthread_cache_done) {
for(queue_Scan(&active_Q, cur, next, thread)) {
queue_Remove(cur);
}
for(queue_Scan(&cache_Q, cur, next, thread)) {
queue_Remove(cur);
}
pthread_mutex_destroy(&active_Q_mutex);
pthread_mutex_destroy(&cache_Q_mutex);
pthread_cache_done = 0;
}
}
/**
* remove a node from a list.
*
* @param[in] node node object
* @param[in] next_state node state following successful dequeue
*
* @return operation status
* @retval 0 success
* @retval AFS_WQ_ERROR in any of the following conditions:
* - node not associated with a work queue
* - node was not on a linked list (e.g. RUNNING state)
* - we raced another thread
*
* @pre node->lock held
*
* @post node removed from node list
*
* @note node->lock may be dropped internally
*
* @internal
*/
static int
_afs_wq_node_list_remove(struct afs_work_queue_node * node,
afs_wq_work_state_t next_state)
{
int code, ret = 0;
struct afs_work_queue_node_list * list = NULL;
_afs_wq_node_state_wait_busy(node);
if (!node->queue) {
ret = AFS_WQ_ERROR;
goto error;
}
switch (node->qidx) {
case AFS_WQ_NODE_LIST_READY:
list = &node->queue->ready_list;
break;
case AFS_WQ_NODE_LIST_BLOCKED:
list = &node->queue->blocked_list;
break;
case AFS_WQ_NODE_LIST_DONE:
list = &node->queue->done_list;
break;
default:
ret = AFS_WQ_ERROR;
}
if (list) {
code = MUTEX_TRYENTER(&list->lock);
if (!code) {
/* contended */
_afs_wq_node_state_change(node,
AFS_WQ_NODE_STATE_BUSY);
MUTEX_EXIT(&node->lock);
MUTEX_ENTER(&list->lock);
MUTEX_ENTER(&node->lock);
if (node->qidx == AFS_WQ_NODE_LIST_NONE) {
/* raced */
ret= AFS_WQ_ERROR;
goto done_sync;
}
}
queue_Remove(node);
node->qidx = AFS_WQ_NODE_LIST_NONE;
_afs_wq_node_state_change(node, next_state);
done_sync:
MUTEX_EXIT(&list->lock);
}
error:
return ret;
}
/**
* unlink work nodes from a dependency node.
*
* @param[in] dep dependency node
*
* @return operation status
* @retval 0 success
*
* @pre
* - dep->parent and dep->child are either locked, or are not referenced
* by anything else
* - caller holds ref on dep->child
* - dep->child and dep->parent in quiescent state
*
* @internal
*/
static int
_afs_wq_dep_unlink_r(struct afs_work_queue_dep_node *dep)
{
struct afs_work_queue_node *child = dep->child;
queue_Remove(&dep->parent_list);
dep->child = NULL;
dep->parent = NULL;
return _afs_wq_node_put_r(child, 0);
}
/**
* low-level interface to remove an entry from the hash table.
*
* Does not alter the refcount or worry about the children lists or
* anything like that; just removes the hash table entry, frees it, and
* that's all. You probably want @see _VVGC_hash_entry_del instead.
*
* @param[in] hent hash table entry
*
* @pre VOL_LOCK held
*
* @return operation status
* @retval 0 success
*
* @internal
*/
static int
_VVGC_hash_entry_unlink(VVGCache_hash_entry_t * hent)
{
int code;
queue_Remove(hent);
hent->entry = NULL;
hent->volid = 0;
code = _VVGC_hash_entry_free(hent);
return code;
}
/**
* delete a VGC entry from the partition's to-delete list.
*
* When a VGC scan is ocurring, and a volume is removed, but then created
* again, we need to ensure that it does not get deleted from being on the
* dlist. Call this function whenever adding a new entry to the VGC during
* a VGC scan to ensure it doesn't get deleted later.
*
* @param[in] dp the partition from whose dlist we are deleting
* @param[in] parent the parent volumeID of the VGC entry
* @param[in] child the child volumeID of the VGC entry
*
* @return operation status
* @retval 0 success
* @retval ENOENT the specified VGC entry is not on the dlist
*
* @pre VVGCache.part[dp->index].state == VVGC_PART_STATE_UPDATING
*
* @internal VGC use only
*
* @see _VVGC_dlist_add_r
*/
int
_VVGC_dlist_del_r(struct DiskPartition64 *dp, VolumeId parent,
VolumeId child)
{
VVGCache_dlist_entry_t *ent;
ent = _VVGC_dlist_lookup_r(dp, parent, child);
if (!ent) {
return ENOENT;
}
queue_Remove(ent);
free(ent);
return 0;
}
/**
* delete all of the entries in the dlist from the VGC.
*
* Traverses the to-delete list for the specified partition, and deletes
* the specified entries from the global VGC. Also deletes the entries from
* the dlist itself as it goes along.
*
* @param[in] dp the partition whose dlist we are flushing
*/
static void
_VVGC_flush_dlist(struct DiskPartition64 *dp)
{
int i;
VVGCache_dlist_entry_t *ent, *nent;
for (i = 0; i < VolumeHashTable.Size; i++) {
for (queue_Scan(&VVGCache.part[dp->index].dlist_hash_buckets[i],
ent, nent,
VVGCache_dlist_entry)) {
_VVGC_entry_purge_r(dp, ent->parent, ent->child);
queue_Remove(ent);
free(ent);
}
}
}
/**
* dequeue a node from a list object.
*
* @param[in] list list object
* @param[out] node_out address in which to store node object pointer
* @param[in] state new node state
* @param[in] block permit blocking on cv if asserted
*
* @return operation status
* @retval 0 success
* @retval EWOULDBLOCK block not asserted and nothing to dequeue
* @retval EINTR blocking wait interrupted by list shutdown
*
* @post node object returned with node lock held and new state set
*
* @internal
*/
static int
_afs_wq_node_list_dequeue(struct afs_work_queue_node_list * list,
struct afs_work_queue_node ** node_out,
afs_wq_work_state_t state,
int block)
{
int ret = 0;
struct afs_work_queue_node * node;
MUTEX_ENTER(&list->lock);
if (list->shutdown) {
*node_out = NULL;
ret = EINTR;
goto done_sync;
}
if (!block && queue_IsEmpty(&list->list)) {
*node_out = NULL;
ret = EWOULDBLOCK;
goto done_sync;
}
while (queue_IsEmpty(&list->list)) {
if (list->shutdown) {
*node_out = NULL;
ret = EINTR;
goto done_sync;
}
CV_WAIT(&list->cv, &list->lock);
}
*node_out = node = queue_First(&list->list, afs_work_queue_node);
MUTEX_ENTER(&node->lock);
queue_Remove(node);
node->qidx = AFS_WQ_NODE_LIST_NONE;
_afs_wq_node_state_change(node, state);
done_sync:
MUTEX_EXIT(&list->lock);
return ret;
}
/**
* look through log_watch_queue, and if any processes are not still
* running, hand them off to the SalvageLogCleanupThread
*
* @param log_watch_queue a queue of PIDs that we should clean up if
* that PID has died
*/
static void
ScanLogs(struct rx_queue *log_watch_queue)
{
struct log_cleanup_node *cleanup, *next;
assert(pthread_mutex_lock(&worker_lock) == 0);
for (queue_Scan(log_watch_queue, cleanup, next, log_cleanup_node)) {
/* if a process is still running, assume it's the salvage process
* still going, and keep waiting for it */
if (kill(cleanup->pid, 0) < 0 && errno == ESRCH) {
queue_Remove(cleanup);
queue_Append(&log_cleanup_queue, cleanup);
assert(pthread_cond_signal(&log_cleanup_queue.queue_change_cv) == 0);
}
}
assert(pthread_mutex_unlock(&worker_lock) == 0);
}
/**
* look through log_watch_queue, and if any processes are not still
* running, hand them off to the SalvageLogCleanupThread
*
* @param log_watch_queue a queue of PIDs that we should clean up if
* that PID has died
*/
static void
ScanLogs(struct rx_queue *log_watch_queue)
{
struct log_cleanup_node *cleanup, *next;
MUTEX_ENTER(&worker_lock);
for (queue_Scan(log_watch_queue, cleanup, next, log_cleanup_node)) {
/* if a process is still running, assume it's the salvage process
* still going, and keep waiting for it */
if (kill(cleanup->pid, 0) < 0 && errno == ESRCH) {
queue_Remove(cleanup);
queue_Append(&log_cleanup_queue, cleanup);
CV_SIGNAL(&log_cleanup_queue.queue_change_cv);
}
}
MUTEX_EXIT(&worker_lock);
}
/* rxi_FillReadVec
*
* Uses packets in the receive queue to fill in as much of the
* current iovec as possible. Does not block if it runs out
* of packets to complete the iovec. Return true if an ack packet
* was sent, otherwise return false */
int
rxi_FillReadVec(struct rx_call *call, afs_uint32 serial)
{
int didConsume = 0;
int didHardAck = 0;
unsigned int t;
struct rx_packet *rp;
struct rx_packet *curp;
struct iovec *call_iov;
struct iovec *cur_iov = NULL;
curp = call->currentPacket;
if (curp) {
cur_iov = &curp->wirevec[call->curvec];
}
call_iov = &call->iov[call->iovNext];
while (!call->error && call->iovNBytes && call->iovNext < call->iovMax) {
if (call->nLeft == 0) {
/* Get next packet */
if (queue_IsNotEmpty(&call->rq)) {
/* Check that next packet available is next in sequence */
rp = queue_First(&call->rq, rx_packet);
if (rp->header.seq == call->rnext) {
afs_int32 error;
struct rx_connection *conn = call->conn;
queue_Remove(rp);
#ifdef RX_TRACK_PACKETS
rp->flags &= ~RX_PKTFLAG_RQ;
#endif
#ifdef RXDEBUG_PACKET
call->rqc--;
#endif /* RXDEBUG_PACKET */
/* RXS_CheckPacket called to undo RXS_PreparePacket's
* work. It may reduce the length of the packet by up
* to conn->maxTrailerSize, to reflect the length of the
* data + the header. */
if ((error =
RXS_CheckPacket(conn->securityObject, call, rp))) {
/* Used to merely shut down the call, but now we
* shut down the whole connection since this may
* indicate an attempt to hijack it */
MUTEX_EXIT(&call->lock);
rxi_ConnectionError(conn, error);
MUTEX_ENTER(&conn->conn_data_lock);
rp = rxi_SendConnectionAbort(conn, rp, 0, 0);
MUTEX_EXIT(&conn->conn_data_lock);
rxi_FreePacket(rp);
MUTEX_ENTER(&call->lock);
return 1;
}
call->rnext++;
curp = call->currentPacket = rp;
#ifdef RX_TRACK_PACKETS
call->currentPacket->flags |= RX_PKTFLAG_CP;
#endif
call->curvec = 1; /* 0th vec is always header */
cur_iov = &curp->wirevec[1];
/* begin at the beginning [ more or less ], continue
* on until the end, then stop. */
call->curpos =
(char *)curp->wirevec[1].iov_base +
call->conn->securityHeaderSize;
call->curlen =
curp->wirevec[1].iov_len -
call->conn->securityHeaderSize;
/* Notice that this code works correctly if the data
* size is 0 (which it may be--no reply arguments from
* server, for example). This relies heavily on the
* fact that the code below immediately frees the packet
* (no yields, etc.). If it didn't, this would be a
* problem because a value of zero for call->nLeft
* normally means that there is no read packet */
call->nLeft = curp->length;
hadd32(call->bytesRcvd, curp->length);
/* Send a hard ack for every rxi_HardAckRate+1 packets
* consumed. Otherwise schedule an event to send
* the hard ack later on.
*/
call->nHardAcks++;
didConsume = 1;
continue;
}
}
break;
}
/* It's possible for call->nLeft to be smaller than any particular
* iov_len. Usually, recvmsg doesn't change the iov_len, since it
* reflects the size of the buffer. We have to keep track of the
* number of bytes read in the length field of the packet struct. On
* the final portion of a received packet, it's almost certain that
* call->nLeft will be smaller than the final buffer. */
while (call->iovNBytes && call->iovNext < call->iovMax && curp) {
t = MIN((int)call->curlen, call->iovNBytes);
t = MIN(t, (int)call->nLeft);
call_iov->iov_base = call->curpos;
call_iov->iov_len = t;
call_iov++;
call->iovNext++;
call->iovNBytes -= t;
call->curpos += t;
call->curlen -= t;
call->nLeft -= t;
if (!call->nLeft) {
/* out of packet. Get another one. */
#ifdef RX_TRACK_PACKETS
curp->flags &= ~RX_PKTFLAG_CP;
curp->flags |= RX_PKTFLAG_IOVQ;
#endif
queue_Append(&call->iovq, curp);
#ifdef RXDEBUG_PACKET
call->iovqc++;
#endif /* RXDEBUG_PACKET */
curp = call->currentPacket = (struct rx_packet *)0;
} else if (!call->curlen) {
/* need to get another struct iov */
if (++call->curvec >= curp->niovecs) {
/* current packet is exhausted, get ready for another */
/* don't worry about curvec and stuff, they get set somewhere else */
#ifdef RX_TRACK_PACKETS
curp->flags &= ~RX_PKTFLAG_CP;
curp->flags |= RX_PKTFLAG_IOVQ;
#endif
queue_Append(&call->iovq, curp);
#ifdef RXDEBUG_PACKET
call->iovqc++;
#endif /* RXDEBUG_PACKET */
curp = call->currentPacket = (struct rx_packet *)0;
call->nLeft = 0;
} else {
cur_iov++;
call->curpos = (char *)cur_iov->iov_base;
call->curlen = cur_iov->iov_len;
}
}
}
}
/* If we consumed any packets then check whether we need to
* send a hard ack. */
if (didConsume && (!(call->flags & RX_CALL_RECEIVE_DONE))) {
if (call->nHardAcks > (u_short) rxi_HardAckRate) {
rxevent_Cancel(call->delayedAckEvent, call,
RX_CALL_REFCOUNT_DELAY);
rxi_SendAck(call, 0, serial, RX_ACK_DELAY, 0);
didHardAck = 1;
} else {
struct clock when, now;
clock_GetTime(&now);
when = now;
/* Delay to consolidate ack packets */
clock_Add(&when, &rx_hardAckDelay);
if (!call->delayedAckEvent
|| clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
rxevent_Cancel(call->delayedAckEvent, call,
RX_CALL_REFCOUNT_DELAY);
MUTEX_ENTER(&rx_refcnt_mutex);
CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
MUTEX_EXIT(&rx_refcnt_mutex);
call->delayedAckEvent =
rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
}
}
}
return didHardAck;
}
/* rxi_ReadProc -- internal version.
*
* LOCKS USED -- called at netpri
*/
int
rxi_ReadProc(struct rx_call *call, char *buf,
int nbytes)
{
struct rx_packet *cp = call->currentPacket;
struct rx_packet *rp;
int requestCount;
unsigned int t;
/* XXXX took out clock_NewTime from here. Was it needed? */
requestCount = nbytes;
/* Free any packets from the last call to ReadvProc/WritevProc */
if (queue_IsNotEmpty(&call->iovq)) {
#ifdef RXDEBUG_PACKET
call->iovqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &call->iovq);
}
do {
if (call->nLeft == 0) {
/* Get next packet */
MUTEX_ENTER(&call->lock);
for (;;) {
if (call->error || (call->mode != RX_MODE_RECEIVING)) {
if (call->error) {
call->mode = RX_MODE_ERROR;
MUTEX_EXIT(&call->lock);
return 0;
}
if (call->mode == RX_MODE_SENDING) {
MUTEX_EXIT(&call->lock);
rxi_FlushWrite(call);
MUTEX_ENTER(&call->lock);
continue;
}
}
if (queue_IsNotEmpty(&call->rq)) {
/* Check that next packet available is next in sequence */
rp = queue_First(&call->rq, rx_packet);
if (rp->header.seq == call->rnext) {
afs_int32 error;
struct rx_connection *conn = call->conn;
queue_Remove(rp);
#ifdef RX_TRACK_PACKETS
rp->flags &= ~RX_PKTFLAG_RQ;
#endif
#ifdef RXDEBUG_PACKET
call->rqc--;
#endif /* RXDEBUG_PACKET */
/* RXS_CheckPacket called to undo RXS_PreparePacket's
* work. It may reduce the length of the packet by up
* to conn->maxTrailerSize, to reflect the length of the
* data + the header. */
if ((error =
RXS_CheckPacket(conn->securityObject, call,
rp))) {
/* Used to merely shut down the call, but now we
* shut down the whole connection since this may
* indicate an attempt to hijack it */
MUTEX_EXIT(&call->lock);
rxi_ConnectionError(conn, error);
MUTEX_ENTER(&conn->conn_data_lock);
rp = rxi_SendConnectionAbort(conn, rp, 0, 0);
MUTEX_EXIT(&conn->conn_data_lock);
rxi_FreePacket(rp);
return 0;
}
call->rnext++;
cp = call->currentPacket = rp;
#ifdef RX_TRACK_PACKETS
call->currentPacket->flags |= RX_PKTFLAG_CP;
#endif
call->curvec = 1; /* 0th vec is always header */
/* begin at the beginning [ more or less ], continue
* on until the end, then stop. */
call->curpos =
(char *)cp->wirevec[1].iov_base +
call->conn->securityHeaderSize;
call->curlen =
cp->wirevec[1].iov_len -
call->conn->securityHeaderSize;
/* Notice that this code works correctly if the data
* size is 0 (which it may be--no reply arguments from
* server, for example). This relies heavily on the
* fact that the code below immediately frees the packet
* (no yields, etc.). If it didn't, this would be a
* problem because a value of zero for call->nLeft
* normally means that there is no read packet */
call->nLeft = cp->length;
hadd32(call->bytesRcvd, cp->length);
/* Send a hard ack for every rxi_HardAckRate+1 packets
* consumed. Otherwise schedule an event to send
* the hard ack later on.
*/
call->nHardAcks++;
if (!(call->flags & RX_CALL_RECEIVE_DONE)) {
if (call->nHardAcks > (u_short) rxi_HardAckRate) {
rxevent_Cancel(call->delayedAckEvent, call,
RX_CALL_REFCOUNT_DELAY);
rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
} else {
struct clock when, now;
clock_GetTime(&now);
when = now;
/* Delay to consolidate ack packets */
clock_Add(&when, &rx_hardAckDelay);
if (!call->delayedAckEvent
|| clock_Gt(&call->delayedAckEvent->
eventTime, &when)) {
rxevent_Cancel(call->delayedAckEvent,
call,
RX_CALL_REFCOUNT_DELAY);
MUTEX_ENTER(&rx_refcnt_mutex);
CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
MUTEX_EXIT(&rx_refcnt_mutex);
call->delayedAckEvent =
rxevent_PostNow(&when, &now,
rxi_SendDelayedAck, call,
0);
}
}
}
break;
}
}
/*
* If we reach this point either we have no packets in the
* receive queue or the next packet in the queue is not the
* one we are looking for. There is nothing else for us to
* do but wait for another packet to arrive.
*/
/* Are there ever going to be any more packets? */
if (call->flags & RX_CALL_RECEIVE_DONE) {
MUTEX_EXIT(&call->lock);
return requestCount - nbytes;
}
/* Wait for in-sequence packet */
call->flags |= RX_CALL_READER_WAIT;
clock_NewTime();
call->startWait = clock_Sec();
while (call->flags & RX_CALL_READER_WAIT) {
#ifdef RX_ENABLE_LOCKS
CV_WAIT(&call->cv_rq, &call->lock);
#else
osi_rxSleep(&call->rq);
#endif
}
cp = call->currentPacket;
call->startWait = 0;
#ifdef RX_ENABLE_LOCKS
if (call->error) {
MUTEX_EXIT(&call->lock);
return 0;
}
#endif /* RX_ENABLE_LOCKS */
}
MUTEX_EXIT(&call->lock);
} else
/* osi_Assert(cp); */
/* MTUXXX this should be replaced by some error-recovery code before shipping */
/* yes, the following block is allowed to be the ELSE clause (or not) */
/* It's possible for call->nLeft to be smaller than any particular
* iov_len. Usually, recvmsg doesn't change the iov_len, since it
* reflects the size of the buffer. We have to keep track of the
* number of bytes read in the length field of the packet struct. On
* the final portion of a received packet, it's almost certain that
* call->nLeft will be smaller than the final buffer. */
while (nbytes && cp) {
t = MIN((int)call->curlen, nbytes);
t = MIN(t, (int)call->nLeft);
memcpy(buf, call->curpos, t);
buf += t;
nbytes -= t;
call->curpos += t;
call->curlen -= t;
call->nLeft -= t;
if (!call->nLeft) {
/* out of packet. Get another one. */
#ifdef RX_TRACK_PACKETS
call->currentPacket->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
} else if (!call->curlen) {
/* need to get another struct iov */
if (++call->curvec >= cp->niovecs) {
/* current packet is exhausted, get ready for another */
/* don't worry about curvec and stuff, they get set somewhere else */
#ifdef RX_TRACK_PACKETS
call->currentPacket->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
call->nLeft = 0;
} else {
call->curpos =
(char *)cp->wirevec[call->curvec].iov_base;
call->curlen = cp->wirevec[call->curvec].iov_len;
}
}
}
if (!nbytes) {
/* user buffer is full, return */
return requestCount;
}
} while (nbytes);
return requestCount;
}
/* Add the indicated event (function, arg) at the specified clock time */
struct rxevent *
rxevent_Post(struct clock * when, void (*func)(), void *arg, void *arg1)
/* when - When event should happen, in clock (clock.h) units */
{
struct rxevent *ev, *qe, *qpr;
#ifdef RXDEBUG
if (Log) {
struct clock now;
clock_GetTime(&now);
fprintf(Log, "%ld.%ld: rxevent_Post(%ld.%ld, %p, %p)\n",
now.sec, now.usec, when->sec, when->usec, func, arg);
}
#endif
#if defined(AFS_SGIMP_ENV)
ASSERT(osi_rxislocked());
#endif
/*
* If we're short on free event entries, create a block of new ones and
* add them to the free queue
*/
if (queue_IsEmpty(&rxevent_free)) {
int i;
#if defined(AFS_AIX32_ENV) && defined(KERNEL)
ev = (struct rxevent *) rxi_Alloc(sizeof(struct rxevent));
queue_Append(&rxevent_free, &ev[0]), rxevent_nFree++;
#else
ev = (struct rxevent *) osi_Alloc(sizeof(struct rxevent) *
rxevent_allocUnit);
xsp = xfreemallocs;
xfreemallocs = (struct xfreelist *) ev;
xfreemallocs->next = xsp;
for (i = 0; i < rxevent_allocUnit; i++)
queue_Append(&rxevent_free, &ev[i]), rxevent_nFree++;
#endif
}
/* Grab and initialize a new rxevent structure */
ev = queue_First(&rxevent_free, rxevent);
queue_Remove(ev);
rxevent_nFree--;
/* Record user defined event state */
ev->eventTime = *when;
ev->func = func;
ev->arg = arg;
ev->arg1 = arg1;
rxevent_nPosted += 1; /* Rather than ++, to shut high-C up
* regarding never-set variables */
/*
* Locate a slot for the new entry. The queue is ordered by time, and we
* assume that a new entry is likely to be greater than a majority of the
* entries already on the queue (unless there's very few entries on the
* queue), so we scan it backwards
*/
for (queue_ScanBackwards(&rxevent_queue, qe, qpr, rxevent)) {
if (clock_Ge(when, &qe->eventTime)) {
queue_InsertAfter(qe, ev);
return ev;
}
}
/* The event is to expire earlier than any existing events */
queue_Prepend(&rxevent_queue, ev);
if (rxevent_ScheduledEarlierEvent)
(*rxevent_ScheduledEarlierEvent) (); /* Notify our external
* scheduler */
return ev;
}
/* rxi_WritevProc -- internal version.
*
* Send buffers allocated in rxi_WritevAlloc.
*
* LOCKS USED -- called at netpri.
*/
int
rxi_WritevProc(struct rx_call *call, struct iovec *iov, int nio, int nbytes)
{
struct rx_packet *cp = NULL;
#ifdef RX_TRACK_PACKETS
struct rx_packet *p, *np;
#endif
int nextio;
int requestCount;
struct rx_queue tmpq;
#ifdef RXDEBUG_PACKET
u_short tmpqc;
#endif
requestCount = nbytes;
nextio = 0;
MUTEX_ENTER(&call->lock);
if (call->error) {
call->mode = RX_MODE_ERROR;
} else if (call->mode != RX_MODE_SENDING) {
call->error = RX_PROTOCOL_ERROR;
}
#ifdef AFS_GLOBAL_RXLOCK_KERNEL
rxi_WaitforTQBusy(call);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
cp = call->currentPacket;
if (call->error) {
call->mode = RX_MODE_ERROR;
MUTEX_EXIT(&call->lock);
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
cp->flags |= RX_PKTFLAG_IOVQ;
#endif
queue_Prepend(&call->iovq, cp);
#ifdef RXDEBUG_PACKET
call->iovqc++;
#endif /* RXDEBUG_PACKET */
call->currentPacket = (struct rx_packet *)0;
}
#ifdef RXDEBUG_PACKET
call->iovqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &call->iovq);
return 0;
}
/* Loop through the I/O vector adjusting packet pointers.
* Place full packets back onto the iovq once they are ready
* to send. Set RX_PROTOCOL_ERROR if any problems are found in
* the iovec. We put the loop condition at the end to ensure that
* a zero length write will push a short packet. */
nextio = 0;
queue_Init(&tmpq);
#ifdef RXDEBUG_PACKET
tmpqc = 0;
#endif /* RXDEBUG_PACKET */
do {
if (call->nFree == 0 && cp) {
clock_NewTime(); /* Bogus: need new time package */
/* The 0, below, specifies that it is not the last packet:
* there will be others. PrepareSendPacket may
* alter the packet length by up to
* conn->securityMaxTrailerSize */
hadd32(call->bytesSent, cp->length);
rxi_PrepareSendPacket(call, cp, 0);
queue_Append(&tmpq, cp);
#ifdef RXDEBUG_PACKET
tmpqc++;
#endif /* RXDEBUG_PACKET */
cp = call->currentPacket = (struct rx_packet *)0;
/* The head of the iovq is now the current packet */
if (nbytes) {
if (queue_IsEmpty(&call->iovq)) {
MUTEX_EXIT(&call->lock);
call->error = RX_PROTOCOL_ERROR;
#ifdef RXDEBUG_PACKET
tmpqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &tmpq);
return 0;
}
cp = queue_First(&call->iovq, rx_packet);
queue_Remove(cp);
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_IOVQ;
#endif
#ifdef RXDEBUG_PACKET
call->iovqc--;
#endif /* RXDEBUG_PACKET */
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_CP;
#endif
call->currentPacket = cp;
call->nFree = cp->length;
call->curvec = 1;
call->curpos =
(char *)cp->wirevec[1].iov_base +
call->conn->securityHeaderSize;
call->curlen =
cp->wirevec[1].iov_len - call->conn->securityHeaderSize;
}
}
if (nbytes) {
/* The next iovec should point to the current position */
if (iov[nextio].iov_base != call->curpos
|| iov[nextio].iov_len > (int)call->curlen) {
call->error = RX_PROTOCOL_ERROR;
MUTEX_EXIT(&call->lock);
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
queue_Prepend(&tmpq, cp);
#ifdef RXDEBUG_PACKET
tmpqc++;
#endif /* RXDEBUG_PACKET */
cp = call->currentPacket = (struct rx_packet *)0;
}
#ifdef RXDEBUG_PACKET
tmpqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &tmpq);
return 0;
}
nbytes -= iov[nextio].iov_len;
call->curpos += iov[nextio].iov_len;
call->curlen -= iov[nextio].iov_len;
call->nFree -= iov[nextio].iov_len;
nextio++;
if (call->curlen == 0) {
if (++call->curvec > cp->niovecs) {
call->nFree = 0;
} else {
call->curpos = (char *)cp->wirevec[call->curvec].iov_base;
call->curlen = cp->wirevec[call->curvec].iov_len;
}
}
}
} while (nbytes && nextio < nio);
/* Move the packets from the temporary queue onto the transmit queue.
* We may end up with more than call->twind packets on the queue. */
#ifdef RX_TRACK_PACKETS
for (queue_Scan(&tmpq, p, np, rx_packet))
{
p->flags |= RX_PKTFLAG_TQ;
}
#endif
if (call->error)
call->mode = RX_MODE_ERROR;
queue_SpliceAppend(&call->tq, &tmpq);
if (!(call->flags & (RX_CALL_FAST_RECOVER | RX_CALL_FAST_RECOVER_WAIT))) {
rxi_Start(0, call, 0, 0);
}
/* Wait for the length of the transmit queue to fall below call->twind */
while (!call->error && call->tnext + 1 > call->tfirst + (2 * call->twind)) {
clock_NewTime();
call->startWait = clock_Sec();
#ifdef RX_ENABLE_LOCKS
CV_WAIT(&call->cv_twind, &call->lock);
#else
call->flags |= RX_CALL_WAIT_WINDOW_ALLOC;
osi_rxSleep(&call->twind);
#endif
call->startWait = 0;
}
/* cp is no longer valid since we may have given up the lock */
cp = call->currentPacket;
if (call->error) {
call->mode = RX_MODE_ERROR;
call->currentPacket = NULL;
MUTEX_EXIT(&call->lock);
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
}
return 0;
}
MUTEX_EXIT(&call->lock);
return requestCount - nbytes;
}