/**
* 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;
}
/**
* link a dependency node to a parent and child work node.
*
* This links a dependency node such that when the 'parent' work node is
* done, the 'child' work node can proceed.
*
* @param[in] dep dependency node
* @param[in] parent parent node in this dependency
* @param[in] child child node in this dependency
*
* @return operation status
* @retval 0 success
*
* @pre
* - parent->lock held
* - child->lock held
* - parent and child in quiescent state
*
* @internal
*/
static int
_afs_wq_dep_link_r(struct afs_work_queue_dep_node *dep,
struct afs_work_queue_node *parent,
struct afs_work_queue_node *child)
{
int ret = 0;
/* Each dep node adds a ref to the child node of that dep. We do not
* do the same for the parent node, since if the only refs remaining
* for a node are deps in node->dep_children, then the node should be
* destroyed, and we will destroy the dep nodes when we free the
* work node. */
ret = _afs_wq_node_get_r(child);
if (ret) {
goto error;
}
/* add this dep node to the parent node's list of deps */
queue_Append(&parent->dep_children, &dep->parent_list);
dep->child = child;
dep->parent = parent;
error:
return ret;
}
/*
* 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 *
SalvageChildReaperThread(void * args)
{
int slot, pid, status;
struct log_cleanup_node * cleanup;
assert(pthread_mutex_lock(&worker_lock) == 0);
/* loop reaping our children */
while (1) {
/* wait() won't block unless we have children, so
* block on the cond var if we're childless */
while (current_workers == 0) {
assert(pthread_cond_wait(&worker_cv, &worker_lock) == 0);
}
assert(pthread_mutex_unlock(&worker_lock) == 0);
cleanup = (struct log_cleanup_node *) malloc(sizeof(struct log_cleanup_node));
while (Reap_Child("salvageserver", &pid, &status) < 0) {
/* try to prevent livelock if something goes wrong */
sleep(1);
}
VOL_LOCK;
for (slot = 0; slot < Parallel; slot++) {
if (child_slot[slot] == pid)
break;
}
assert(slot < Parallel);
child_slot[slot] = 0;
VOL_UNLOCK;
SALVSYNC_doneWorkByPid(pid, status);
assert(pthread_mutex_lock(&worker_lock) == 0);
if (cleanup) {
cleanup->pid = pid;
queue_Append(&log_cleanup_queue, cleanup);
assert(pthread_cond_signal(&log_cleanup_queue.queue_change_cv) == 0);
}
/* ok, we've reaped a child */
current_workers--;
assert(pthread_cond_broadcast(&worker_cv) == 0);
}
return NULL;
}
/**
* add an entry to the hash table.
*
* @param[in] dp disk partition object
* @param[in] volid volume id
* @param[in] ent volume group object
* @param[out] hash_out address in which to store pointer to hash entry
*
* @pre VOL_LOCK held
*
* @return operation status
* @retval 0 success
* @retval EEXIST hash entry for volid already exists, and it points to
* a different VG entry
*
* @internal
*/
static int
_VVGC_hash_entry_add(struct DiskPartition64 * dp,
VolumeId volid,
VVGCache_entry_t * ent,
VVGCache_hash_entry_t ** hash_out)
{
int code = 0;
VVGCache_hash_entry_t * hent;
int hash = VVGC_HASH(volid);
VVGCache_entry_t *nent;
code = _VVGC_lookup(dp, volid, &nent, hash_out);
if (!code) {
if (ent != nent) {
ViceLog(0, ("_VVGC_hash_entry_add: tried to add a duplicate "
" nonmatching entry for vol %lu: original "
"(%"AFS_PTR_FMT",%lu) new (%"AFS_PTR_FMT",%lu)\n",
afs_printable_uint32_lu(volid),
nent, afs_printable_uint32_lu(nent->rw),
ent, afs_printable_uint32_lu(ent->rw)));
return EEXIST;
}
ViceLog(1, ("_VVGC_hash_entry_add: tried to add duplicate "
"hash entry for vol %lu, VG %lu",
afs_printable_uint32_lu(volid),
afs_printable_uint32_lu(ent->rw)));
/* accept attempts to add matching duplicate entries; just
* pretend we added it */
return 0;
}
code = _VVGC_hash_entry_alloc(&hent);
if (code) {
goto done;
}
hent->entry = ent;
hent->dp = dp;
hent->volid = volid;
queue_Append(&VVGCache_hash_table.hash_buckets[hash],
hent);
done:
if (hash_out) {
*hash_out = hent;
}
return code;
}
/**
* append to a node list object.
*
* @param[in] list list object
* @param[in] node node object
* @param[in] state new node state
*
* @return operation status
* @retval 0 success
* @retval AFS_WQ_ERROR raced to enqueue node
*
* @pre
* - node lock held
* - node is not on a list
* - node is either not busy, or it is marked as busy by the calling thread
*
* @post
* - enqueued on list
* - node lock dropped
*
* @internal
*/
static int
_afs_wq_node_list_enqueue(struct afs_work_queue_node_list * list,
struct afs_work_queue_node * node,
afs_wq_work_state_t state)
{
int code, ret = 0;
if (node->qidx != AFS_WQ_NODE_LIST_NONE) {
/* raced */
ret = AFS_WQ_ERROR;
goto error;
}
/* deal with lock inversion */
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);
/* assert state of the world (we set busy, so this should never happen) */
osi_Assert(queue_IsNotOnQueue(node));
}
if (list->shutdown) {
ret = AFS_WQ_ERROR;
goto error_unlock;
}
osi_Assert(node->qidx == AFS_WQ_NODE_LIST_NONE);
if (queue_IsEmpty(&list->list)) {
/* wakeup a dequeue thread */
CV_SIGNAL(&list->cv);
}
queue_Append(&list->list, node);
node->qidx = list->qidx;
_afs_wq_node_state_change(node, state);
error_unlock:
MUTEX_EXIT(&node->lock);
MUTEX_EXIT(&list->lock);
error:
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);
}
/**
* add a VGC entry to the partition's to-delete list.
*
* This adds a VGC entry (a parent/child pair) to a list of VGC entries to
* be deleted from the VGC at the end of a VGC scan. This is necessary,
* while a VGC scan is ocurring, volumes may be deleted. Since a VGC scan
* scans a partition in VVGC_SCAN_TBL_LEN chunks, a VGC delete operation
* may delete a volume, only for it to be added again when the VGC scan's
* table adds it to the VGC. So when a VGC entry is deleted and a VGC scan
* is running, this function must be called to ensure it does not come
* back onto the VGC.
*
* @param[in] dp the partition to whose dlist we are adding
* @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 ENOMEM memory allocation error
*
* @pre VVGCache.part[dp->index].state == VVGC_PART_STATE_UPDATING
*
* @internal VGC use only
*/
int
_VVGC_dlist_add_r(struct DiskPartition64 *dp, VolumeId parent,
VolumeId child)
{
int bucket = VVGC_HASH(child);
VVGCache_dlist_entry_t *entry;
entry = malloc(sizeof(*entry));
if (!entry) {
return ENOMEM;
}
entry->child = child;
entry->parent = parent;
queue_Append(&VVGCache.part[dp->index].dlist_hash_buckets[bucket],
entry);
return 0;
}
/**
* Thread to look for SalvageLog.$pid files that are not from our child
* worker salvagers, and notify SalvageLogCleanupThread to clean them
* up. This can happen if we restart during salvages, or the
* salvageserver crashes or something.
*
* @param arg unused
*
* @return always NULL
*/
static void *
SalvageLogScanningThread(void * arg)
{
struct rx_queue log_watch_queue;
queue_Init(&log_watch_queue);
{
DIR *dp;
struct dirent *dirp;
char prefix[AFSDIR_PATH_MAX];
size_t prefix_len;
afs_snprintf(prefix, sizeof(prefix), "%s.", AFSDIR_SLVGLOG_FILE);
prefix_len = strlen(prefix);
dp = opendir(AFSDIR_LOGS_DIR);
assert(dp);
while ((dirp = readdir(dp)) != NULL) {
pid_t pid;
struct log_cleanup_node *cleanup;
int i;
if (strncmp(dirp->d_name, prefix, prefix_len) != 0) {
/* not a salvage logfile; skip */
continue;
}
errno = 0;
pid = strtol(dirp->d_name + prefix_len, NULL, 10);
if (errno != 0) {
/* file is SalvageLog.<something> but <something> isn't
* a pid, so skip */
continue;
}
VOL_LOCK;
for (i = 0; i < Parallel; ++i) {
if (pid == child_slot[i]) {
break;
}
}
VOL_UNLOCK;
if (i < Parallel) {
/* this pid is one of our children, so the reaper thread
* will take care of it; skip */
continue;
}
cleanup =
(struct log_cleanup_node *) malloc(sizeof(struct log_cleanup_node));
cleanup->pid = pid;
queue_Append(&log_watch_queue, cleanup);
}
closedir(dp);
}
ScanLogs(&log_watch_queue);
while (queue_IsNotEmpty(&log_watch_queue)) {
sleep(SALVAGE_SCAN_POLL_INTERVAL);
ScanLogs(&log_watch_queue);
}
return NULL;
}
/* 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;
}
int
rxi_WritevAlloc(struct rx_call *call, struct iovec *iov, int *nio, int maxio,
int nbytes)
{
struct rx_connection *conn = call->conn;
struct rx_packet *cp = call->currentPacket;
int requestCount;
int nextio;
/* Temporary values, real work is done in rxi_WritevProc */
int tnFree;
unsigned int tcurvec;
char *tcurpos;
int tcurlen;
requestCount = nbytes;
nextio = 0;
/* 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);
}
if (call->mode != RX_MODE_SENDING) {
if ((conn->type == RX_SERVER_CONNECTION)
&& (call->mode == RX_MODE_RECEIVING)) {
call->mode = RX_MODE_SENDING;
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
call->nLeft = 0;
call->nFree = 0;
}
} else {
return 0;
}
}
/* Set up the iovec to point to data in packet buffers. */
tnFree = call->nFree;
tcurvec = call->curvec;
tcurpos = call->curpos;
tcurlen = call->curlen;
do {
int t;
if (tnFree == 0) {
/* current packet is full, allocate a new one */
MUTEX_ENTER(&call->lock);
cp = rxi_AllocSendPacket(call, nbytes);
MUTEX_EXIT(&call->lock);
if (cp == NULL) {
/* out of space, return what we have */
*nio = nextio;
return requestCount - nbytes;
}
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_IOVQ;
#endif
queue_Append(&call->iovq, cp);
#ifdef RXDEBUG_PACKET
call->iovqc++;
#endif /* RXDEBUG_PACKET */
tnFree = cp->length;
tcurvec = 1;
tcurpos =
(char *)cp->wirevec[1].iov_base +
call->conn->securityHeaderSize;
tcurlen = cp->wirevec[1].iov_len - call->conn->securityHeaderSize;
}
if (tnFree < nbytes) {
/* try to extend the current packet */
int len, mud;
len = cp->length;
mud = rx_MaxUserDataSize(call);
if (mud > len) {
int want;
want = MIN(nbytes - tnFree, mud - len);
rxi_AllocDataBuf(cp, want, RX_PACKET_CLASS_SEND_CBUF);
if (cp->length > (unsigned)mud)
cp->length = mud;
tnFree += (cp->length - len);
if (cp == call->currentPacket) {
call->nFree += (cp->length - len);
}
}
}
/* fill in the next entry in the iovec */
t = MIN(tcurlen, nbytes);
t = MIN(tnFree, t);
iov[nextio].iov_base = tcurpos;
iov[nextio].iov_len = t;
nbytes -= t;
tcurpos += t;
tcurlen -= t;
tnFree -= t;
nextio++;
if (!tcurlen) {
/* need to get another struct iov */
if (++tcurvec >= cp->niovecs) {
/* current packet is full, extend it or move on to next packet */
tnFree = 0;
} else {
tcurpos = (char *)cp->wirevec[tcurvec].iov_base;
tcurlen = cp->wirevec[tcurvec].iov_len;
}
}
} while (nbytes && nextio < maxio);
*nio = nextio;
return requestCount - nbytes;
}
int
rxi_WriteProc(struct rx_call *call, char *buf,
int nbytes)
{
struct rx_connection *conn = call->conn;
struct rx_packet *cp = call->currentPacket;
unsigned int t;
int 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);
}
if (call->mode != RX_MODE_SENDING) {
if ((conn->type == RX_SERVER_CONNECTION)
&& (call->mode == RX_MODE_RECEIVING)) {
call->mode = RX_MODE_SENDING;
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
call->nLeft = 0;
call->nFree = 0;
}
} else {
return 0;
}
}
/* Loop condition is checked at end, so that a write of 0 bytes
* will force a packet to be created--specially for the case where
* there are 0 bytes on the stream, but we must send a packet
* anyway. */
do {
if (call->nFree == 0) {
MUTEX_ENTER(&call->lock);
#ifdef AFS_GLOBAL_RXLOCK_KERNEL
rxi_WaitforTQBusy(call);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
cp = call->currentPacket;
if (call->error)
call->mode = RX_MODE_ERROR;
if (!call->error && cp) {
/* Clear the current packet now so that if
* we are forced to wait and drop the lock
* the packet we are planning on using
* cannot be freed.
*/
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
call->currentPacket = (struct rx_packet *)0;
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);
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_TQ;
#endif
queue_Append(&call->tq, cp);
#ifdef RXDEBUG_PACKET
call->tqc++;
#endif /* RXDEBUG_PACKET */
cp = (struct rx_packet *)0;
if (!
(call->
flags & (RX_CALL_FAST_RECOVER |
RX_CALL_FAST_RECOVER_WAIT))) {
rxi_Start(0, call, 0, 0);
}
} else if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
}
/* Wait for transmit window to open up */
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;
#ifdef RX_ENABLE_LOCKS
if (call->error) {
call->mode = RX_MODE_ERROR;
MUTEX_EXIT(&call->lock);
return 0;
}
#endif /* RX_ENABLE_LOCKS */
}
if ((cp = rxi_AllocSendPacket(call, nbytes))) {
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_CP;
#endif
call->currentPacket = cp;
call->nFree = cp->length;
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;
}
if (call->error) {
call->mode = RX_MODE_ERROR;
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
call->currentPacket = NULL;
}
MUTEX_EXIT(&call->lock);
return 0;
}
MUTEX_EXIT(&call->lock);
}
if (cp && (int)call->nFree < nbytes) {
/* Try to extend the current buffer */
int len, mud;
len = cp->length;
mud = rx_MaxUserDataSize(call);
if (mud > len) {
int want;
want = MIN(nbytes - (int)call->nFree, mud - len);
rxi_AllocDataBuf(cp, want, RX_PACKET_CLASS_SEND_CBUF);
if (cp->length > (unsigned)mud)
cp->length = mud;
call->nFree += (cp->length - len);
}
}
/* If the remaining bytes fit in the buffer, then store them
* and return. Don't ship a buffer that's full immediately to
* the peer--we don't know if it's the last buffer yet */
if (!cp) {
call->nFree = 0;
}
while (nbytes && call->nFree) {
t = MIN((int)call->curlen, nbytes);
t = MIN((int)call->nFree, t);
memcpy(call->curpos, buf, t);
buf += t;
nbytes -= t;
call->curpos += t;
call->curlen -= (u_short)t;
call->nFree -= (u_short)t;
if (!call->curlen) {
/* need to get another struct iov */
if (++call->curvec >= cp->niovecs) {
/* current packet is full, extend or send it */
call->nFree = 0;
} else {
call->curpos = (char *)cp->wirevec[call->curvec].iov_base;
call->curlen = cp->wirevec[call->curvec].iov_len;
}
}
} /* while bytes to send and room to send them */
/* might be out of space now */
if (!nbytes) {
return requestCount;
} else; /* more data to send, so get another packet and keep going */
} while (nbytes);
return requestCount - nbytes;
}
/* 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;
}
/* Flush any buffered data to the stream, switch to read mode
* (clients) or to EOF mode (servers)
*
* LOCKS HELD: called at netpri.
*/
void
rxi_FlushWrite(struct rx_call *call)
{
struct rx_packet *cp = NULL;
/* 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);
}
if (call->mode == RX_MODE_SENDING) {
call->mode =
(call->conn->type ==
RX_CLIENT_CONNECTION ? RX_MODE_RECEIVING : RX_MODE_EOF);
#ifdef RX_KERNEL_TRACE
{
int glockOwner = ISAFS_GLOCK();
if (!glockOwner)
AFS_GLOCK();
afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
__FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
call);
if (!glockOwner)
AFS_GUNLOCK();
}
#endif
MUTEX_ENTER(&call->lock);
#ifdef AFS_GLOBAL_RXLOCK_KERNEL
rxi_WaitforTQBusy(call);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
if (call->error)
call->mode = RX_MODE_ERROR;
cp = call->currentPacket;
if (cp) {
/* cp->length is only supposed to be the user's data */
/* cp->length was already set to (then-current)
* MaxUserDataSize or less. */
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
cp->length -= call->nFree;
call->currentPacket = (struct rx_packet *)0;
call->nFree = 0;
} else {
cp = rxi_AllocSendPacket(call, 0);
if (!cp) {
/* Mode can no longer be MODE_SENDING */
return;
}
cp->length = 0;
cp->niovecs = 2; /* header + space for rxkad stuff */
call->nFree = 0;
}
/* The 1 specifies that this is the last packet */
hadd32(call->bytesSent, cp->length);
rxi_PrepareSendPacket(call, cp, 1);
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_TQ;
#endif
queue_Append(&call->tq, cp);
#ifdef RXDEBUG_PACKET
call->tqc++;
#endif /* RXDEBUG_PACKET */
if (!
(call->
flags & (RX_CALL_FAST_RECOVER | RX_CALL_FAST_RECOVER_WAIT))) {
rxi_Start(0, call, 0, 0);
}
MUTEX_EXIT(&call->lock);
}
}
/* 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;
}
