On Mon, Aug 24, 2009 at 09:51:07PM +0100, Daniel P. Berrange wrote:
* qemud/stream.c: Handle incoming stream data packets, queuing until stream becomes writable. Handle stream completion handshake * po/POTFILES.in: Add qemud/stream.c [...] +/* + * Callback that gets invoked when a stream becomes writable/readable + */ +static void +remoteStreamEvent(virStreamPtr st, int events, void *opaque) +{ + struct qemud_client *client = opaque; + struct qemud_client_stream *stream; + + /* XXX sub-optimal - we really should be taking the server lock + * first, but we have no handle to the server object + * We're lucky to get away with it for now, due to this callback + * executing in the main thread, but this should really be fixed + */ + virMutexLock(&client->lock);
Shouldn't the stream stucture carry a reference to the connection/server ? [...]
static int -remoteStreamFilter(struct qemud_client *client ATTRIBUTE_UNUSED, - struct qemud_client_message *msg ATTRIBUTE_UNUSED, - void *opaque ATTRIBUTE_UNUSED) +remoteStreamFilter(struct qemud_client *client, + struct qemud_client_message *msg, void *opaque) { + struct qemud_client_stream *stream = opaque; + + if (msg->hdr.serial == stream->serial && + msg->hdr.proc == stream->procedure && + msg->hdr.type == REMOTE_STREAM) { + DEBUG("Incoming rx=%p serial=%d proc=%d status=%d", + stream->rx, msg->hdr.proc, msg->hdr.serial, msg->hdr.status); + + /* If there are queued packets, we need to queue all further + * messages, since they must be processed strictly in order. + * If there are no queued packets, then OK/ERROR messages + * should be processed immediately. Data packets are still + * queued to only be processed when the stream is marked as + * writable. + */
Since we are already separating different classes of packets, why not extend this to provide 2-3 classes for the data too: BACKGROUND/NORMAL/URGENT that can be added later using the current flags for creating a stream. ACK, though this need much testing, ideally we can automate this in TCK, especially about Abort effect at various point in time of a transfer. This is hard to test by the application writer so we should try to make the API as foolproof as prossible. Daniel -- Daniel Veillard | libxml Gnome XML XSLT toolkit http://xmlsoft.org/ daniel@veillard.com | Rpmfind RPM search engine http://rpmfind.net/ http://veillard.com/ | virtualization library http://libvirt.org/