Re: [Qemu-devel] [Spice-devel] [PATCH] server: don't call reds_stream_free from worker thread context

[Yonit Halperin]

On 09/02/2011 06:19 PM, Gerd Hoffmann wrote:
> reds_stream_free() may call the channel_event callback which is not
> supposed to be callsed from worker thread context.  This patch moves
> the reds_stream_free call for the display channel from the worker to
> the dispatcher to fix this issue.
>
> [ Note: not tested yet, against 0.8 branch, sending out for review&
>          comments nevertheless ]
>
> Signed-off-by: Gerd Hoffmann<address@hidden>
> ---
>   server/red_dispatcher.c |    5 +++++
>   server/red_worker.c     |    3 +--
>   2 files changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/server/red_dispatcher.c b/server/red_dispatcher.c
> index f74b13e..801a575 100644
> --- a/server/red_dispatcher.c
> +++ b/server/red_dispatcher.c
> @@ -51,6 +51,7 @@ struct RedDispatcher {
>       int y_res;
>       int use_hardware_cursor;
>       RedDispatcher *next;
> +    RedsStream *stream;
>       RedWorkerMessage async_message;
>       pthread_mutex_t  async_lock;
>       QXLDevSurfaceCreate surface_create;
> @@ -81,6 +82,7 @@ static void red_dispatcher_set_peer(Channel *channel, 
> RedsStream *stream, int mi
>
>       red_printf("");
>       dispatcher = (RedDispatcher *)channel->data;
> +    dispatcher->stream = stream;
>       RedWorkerMessage message = RED_WORKER_MESSAGE_DISPLAY_CONNECT;
>       write_message(dispatcher->channel,&message);
>       send_data(dispatcher->channel,&stream, sizeof(RedsStream *));
> @@ -93,6 +95,9 @@ static void red_dispatcher_shutdown_peer(Channel *channel)
>       red_printf("");
>       RedWorkerMessage message = RED_WORKER_MESSAGE_DISPLAY_DISCONNECT;
>       write_message(dispatcher->channel,&message);
> +    read_message(dispatcher->channel,&message);
> +    ASSERT(message == RED_WORKER_MESSAGE_READY);
> +    reds_stream_free(dispatcher->stream);

Hi,
RED_WORKER_MESSAGE_DISPLAY_DISCONNECT is not the only place that 
triggers red_disconnect_channel (and as a result, 
reds_stream_free(dispatcher->stream)). red_disconnect_channel is called 
also when there is an error upon receive/send and also when timeouts 
related to the client occur (e.g., in flush_display_commands).

We probably better make the dispatcher bi-directional, i.e., not only 
push messages to the worker, but also listen.

Cheers,
Yonit.

>   }
>
>   static void red_dispatcher_migrate(Channel *channel)
> diff --git a/server/red_worker.c b/server/red_worker.c
> index 5f07803..f77b0f2 100644
> --- a/server/red_worker.c
> +++ b/server/red_worker.c
> @@ -8486,8 +8486,6 @@ static void red_disconnect_channel(RedChannel *channel)
>   {
>       channel_release_res(channel);
>       red_pipe_clear(channel);
> -    reds_stream_free(channel->stream);
> -    channel->stream = NULL;
>       channel->send_data.blocked = FALSE;
>       channel->send_data.size = channel->send_data.pos = 0;
>       spice_marshaller_reset(channel->send_data.marshaller);
> @@ -10060,6 +10058,7 @@ static void handle_dev_input(EventListener *listener, 
> uint32_t events)
>       case RED_WORKER_MESSAGE_CURSOR_DISCONNECT:
>           red_printf("cursor disconnect");
>           red_disconnect_cursor((RedChannel *)worker->cursor_channel);
> +        write_ready = 1;
>           break;
>       case RED_WORKER_MESSAGE_CURSOR_MIGRATE:
>           red_printf("cursor migrate");