int ConnectionDescriptor::_SendRawOutboundData (const char *data, int length)
{
/* This internal method is called to schedule bytes that
* will be sent out to the remote peer.
* It's not directly accessed by the caller, who hits ::SendOutboundData,
* which may or may not filter or encrypt the caller's data before
* sending it here.
*/
// Highly naive and incomplete implementation.
// There's no throttle for runaways (which should abort only this connection
// and not the whole process), and no coalescing of small pages.
// (Well, not so bad, small pages are coalesced in ::Write)
if (IsCloseScheduled())
//if (bCloseNow || bCloseAfterWriting)
return 0;
if (!data && (length > 0))
throw std::runtime_error ("bad outbound data");
char *buffer = (char *) malloc (length + 1);
if (!buffer)
throw std::runtime_error ("no allocation for outbound data");
memcpy (buffer, data, length);
buffer [length] = 0;
OutboundPages.push_back (OutboundPage (buffer, length));
OutboundDataSize += length;
_UpdateEvents(false, true);
return length;
}
void ConnectionDescriptor::SetNotifyWritable(bool writable)
{
if (!bWatchOnly)
throw std::runtime_error ("notify_writable must be on 'watch only' connections");
bNotifyWritable = writable;
_UpdateEvents(false, true);
}
void ConnectionDescriptor::SetNotifyReadable(bool readable)
{
if (!bWatchOnly)
throw std::runtime_error ("notify_readable must be on 'watch only' connections");
bNotifyReadable = readable;
_UpdateEvents(true, false);
}
bool ConnectionDescriptor::Resume()
{
if (bWatchOnly)
throw std::runtime_error ("cannot pause/resume 'watch only' connections, set notify readable/writable instead");
bool old = bPaused;
bPaused = false;
_UpdateEvents();
return old == true;
}
void ConnectionDescriptor::_WriteOutboundData()
{
/* This is a helper function called by ::Write.
* It's possible for a socket to select writable and then no longer
* be writable by the time we get around to writing. The kernel might
* have used up its available output buffers between the select call
* and when we get here. So this condition is not an error.
*
* 20Jul07, added the same kind of protection against an invalid socket
* that is at the top of ::Read. Not entirely how this could happen in
* real life (connection-reset from the remote peer, perhaps?), but I'm
* doing it to address some reports of crashing under heavy loads.
*/
int sd = GetSocket();
//assert (sd != INVALID_SOCKET);
if (sd == INVALID_SOCKET) {
assert (!bWriteAttemptedAfterClose);
bWriteAttemptedAfterClose = true;
return;
}
LastActivity = MyEventMachine->GetCurrentLoopTime();
size_t nbytes = 0;
#ifdef HAVE_WRITEV
int iovcnt = OutboundPages.size();
// Max of 16 outbound pages at a time
if (iovcnt > 16) iovcnt = 16;
#ifdef CC_SUNWspro
struct iovec iov[16];
#else
struct iovec iov[ iovcnt ];
#endif
for(int i = 0; i < iovcnt; i++){
OutboundPage *op = &(OutboundPages[i]);
#ifdef CC_SUNWspro
iov[i].iov_base = (char *)(op->Buffer + op->Offset);
#else
iov[i].iov_base = (void *)(op->Buffer + op->Offset);
#endif
iov[i].iov_len = op->Length - op->Offset;
nbytes += iov[i].iov_len;
}
#else
char output_buffer [16 * 1024];
while ((OutboundPages.size() > 0) && (nbytes < sizeof(output_buffer))) {
OutboundPage *op = &(OutboundPages[0]);
if ((nbytes + op->Length - op->Offset) < sizeof (output_buffer)) {
memcpy (output_buffer + nbytes, op->Buffer + op->Offset, op->Length - op->Offset);
nbytes += (op->Length - op->Offset);
op->Free();
OutboundPages.pop_front();
}
else {
int len = sizeof(output_buffer) - nbytes;
memcpy (output_buffer + nbytes, op->Buffer + op->Offset, len);
op->Offset += len;
nbytes += len;
}
}
#endif
// We should never have gotten here if there were no data to write,
// so assert that as a sanity check.
// Don't bother to make sure nbytes is less than output_buffer because
// if it were we probably would have crashed already.
assert (nbytes > 0);
assert (GetSocket() != INVALID_SOCKET);
#ifdef HAVE_WRITEV
int bytes_written = writev (GetSocket(), iov, iovcnt);
#else
int bytes_written = write (GetSocket(), output_buffer, nbytes);
#endif
bool err = false;
if (bytes_written < 0) {
err = true;
bytes_written = 0;
}
assert (bytes_written >= 0);
OutboundDataSize -= bytes_written;
if (ProxiedFrom && MaxOutboundBufSize && (unsigned int)GetOutboundDataSize() < MaxOutboundBufSize && ProxiedFrom->IsPaused())
ProxiedFrom->Resume();
#ifdef HAVE_WRITEV
if (!err) {
unsigned int sent = bytes_written;
deque<OutboundPage>::iterator op = OutboundPages.begin();
for (int i = 0; i < iovcnt; i++) {
if (iov[i].iov_len <= sent) {
// Sent this page in full, free it.
op->Free();
OutboundPages.pop_front();
sent -= iov[i].iov_len;
} else {
// Sent part (or none) of this page, increment offset to send the remainder
op->Offset += sent;
break;
}
// Shouldn't be possible run out of pages before the loop ends
assert(op != OutboundPages.end());
*op++;
}
}
#else
if ((size_t)bytes_written < nbytes) {
int len = nbytes - bytes_written;
char *buffer = (char*) malloc (len + 1);
if (!buffer)
throw std::runtime_error ("bad alloc throwing back data");
memcpy (buffer, output_buffer + bytes_written, len);
buffer [len] = 0;
OutboundPages.push_front (OutboundPage (buffer, len));
}
#endif
_UpdateEvents(false, true);
if (err) {
#ifdef OS_UNIX
if ((errno != EINPROGRESS) && (errno != EWOULDBLOCK) && (errno != EINTR))
#endif
#ifdef OS_WIN32
if ((errno != WSAEINPROGRESS) && (errno != WSAEWOULDBLOCK))
#endif
Close();
}
}
void ConnectionDescriptor::Write()
{
/* A socket which is in a pending-connect state will select
* writable when the disposition of the connect is known.
* At that point, check to be sure there are no errors,
* and if none, then promote the socket out of the pending
* state.
* TODO: I haven't figured out how Windows signals errors on
* unconnected sockets. Maybe it does the untraditional but
* logical thing and makes the socket selectable for error.
* If so, it's unsupported here for the time being, and connect
* errors will have to be caught by the timeout mechanism.
*/
if (bConnectPending) {
int error;
socklen_t len;
len = sizeof(error);
#ifdef OS_UNIX
int o = getsockopt (GetSocket(), SOL_SOCKET, SO_ERROR, &error, &len);
#endif
#ifdef OS_WIN32
int o = getsockopt (GetSocket(), SOL_SOCKET, SO_ERROR, (char*)&error, &len);
#endif
if ((o == 0) && (error == 0)) {
if (EventCallback)
(*EventCallback)(GetBinding(), EM_CONNECTION_COMPLETED, "", 0);
// 5May09: Moved epoll/kqueue read/write arming into SetConnectPending, so it can be called
// from EventMachine_t::AttachFD as well.
SetConnectPending (false);
}
else
ScheduleClose (false);
//bCloseNow = true;
}
else {
if (bNotifyWritable) {
if (EventCallback)
(*EventCallback)(GetBinding(), EM_CONNECTION_NOTIFY_WRITABLE, NULL, 0);
_UpdateEvents(false, true);
return;
}
assert(!bWatchOnly);
/* 5May09: Kqueue bugs on OSX cause one extra writable event to fire even though we're using
EV_ONESHOT. We ignore this extra event once, but only the first time. If it happens again,
we should fall through to the assert(nbytes>0) failure to catch any EM bugs which might cause
::Write to be called in a busy-loop.
*/
#ifdef HAVE_KQUEUE
if (MyEventMachine->UsingKqueue()) {
if (OutboundDataSize == 0 && !bGotExtraKqueueEvent) {
bGotExtraKqueueEvent = true;
return;
} else if (OutboundDataSize > 0) {
bGotExtraKqueueEvent = false;
}
}
#endif
_WriteOutboundData();
}
}
void ConnectionDescriptor::SetWatchOnly(bool watching)
{
bWatchOnly = watching;
_UpdateEvents();
}
void ConnectionDescriptor::SetConnectPending(bool f)
{
bConnectPending = f;
_UpdateEvents();
}
void ConnectionDescriptor::_UpdateEvents()
{
_UpdateEvents(true, true);
}
