Re: [Chromium-dev] glFinish can hang parallel tilesort?

[Alan H. <al...@re...>]

David,

I'll have to read your email a few times to completely understand
what's happening....But....

Just as a test - try the multithread-1-branch, we break out
of crNetGetNextMessage now to process multiple clients.

Alan.

On Thu, Mar 28, 2002 at 02:44:56PM -0800, David Thompson wrote:
> I have been having problems with parallel tilesort.
> If I insert a glFinish() anywhere in the application,
> it will hang with both servers and clients in
> crNetRecv. I think I've figured out the problem, but
> I wanted to run it by the list...
> 
> Basically, it looks like the crserver can get stuck in
> crGetNextMessage (called from crServerSerializeRemoteStreams).
> Because crGetNextMessage is only checking one client
> connection's buffer but is calling crNetRecv (which
> will accept buffers from any connection), the server
> will never get around to a writeback command issued from
> some other client.
> 
> For example, consider two client machines A & B and a
> tilesort app with two tiles. Crserver processes run on
> machines C & D. If both A & B call glFinish() then two
> packets get sent to C and two get sent to D. All
> four packets contain a writeback command. If the server on
> machine C is running the inner loop of crServerSerializeRemoteStreams
> with run_queue set to client A's connection, it will continue to
> call crGetNextMessage for client A's buffer. Once it has processed
> the writeback request, it continues to process requests for ONLY
> client A. But client A may not be done receiving writebacks because
> server D may be in crServerSerializeRemoteStreams and only
> processing messages for client B. This means both A&B are
> waiting on writebacks from different machines, neither of which
> will answer until the other client dies or sends something
> that causes the connection to get blocked.
> 
> So, is it
> 1. I am missing something,
> 2. This is dreadful and should get fixed,
> 3. I am a moron for calling glFinish from more than one node... ?
> 
> 		David T.
> 
> 
> 
> _______________________________________________
> Chromium-dev mailing list
> Chr...@li...
> https://lists.sourceforge.net/lists/listinfo/chromium-dev