|
We can always reproduce it Greg. It takes a few
minutes, but it happens on Raspberry. We are not aware of this
happening on Linux/PC (and we have many servers on production
transferring lots of GB on a daily basis - big binaries is one of
the strengths of Plastic compared to Git... that's why). Thanks, pablo On 9/13/2016 1:59, Greg Young wrote:
We dealt with an issue similar to this long ago that was due to threadpool requests being dropped occasionally. How reproducible is your issue?On Mon, Sep 12, 2016 at 11:45 PM, psant...@codicesoftware.com <psant...@codicesoftware.com> wrote:Hi Alan, I'll try to create a small test case. We are experiencing this running a Plastic SCM server on raspberry, and a client on a different machine. It will take me a few hours to isolate a test case taking just the required parts from the plastic network protocol. Meanwhile: do you have any suggestion to try to check whether my assumption is correct? I mean, I suspect somehow the BeginReceive never awakes. I guess it is using epoll down the stairs. Any log or something I can build to make sure the issue is on my code and not on Mono? Thanks, pablo On 9/12/2016 11:37, Alan McGovern wrote: Can you share the code to repro the issue? Sent from my iPhone On 12 Sep 2016, at 09:39, "psant...@codicesoftware.com" <psant...@codicesoftware.com> wrote: Thanks for the hint Chris, We are actually enqueuing the request and doing the actual read on a different threadpool. That being said, we are reproducing this issue with a single client, I mean, single thread attending on the server. Not sure what can go wrong, tcp connection is still there, client can send, it simply looks like the server never wakes up to attend it, randomly. Any other hints would be appreciated. Thanks, pablo On 9/9/2016 18:11, Chris Swiedler wrote: >From what I understand it’s dangerous to do blocking reads on the callback from something like BeginReceive. The callback will happen on a threadpool thread, and the blocking reads could then cause the threadpool to be exhausted. I don’t know if that’s causing your specific problem (it could, if the reads really do block and you have enough of them) but it’s something to watch out for. chris From: Mono-devel-list [mailto:mono-devel-list-boun...@lists.dot.net] On Behalf Of psant...@codicesoftware.com Sent: Friday, September 09, 2016 2:42 AM To: mono-devel-list@lists.dot.net Cc: dude <rdea...@codicesoftware.com> Subject: [Mono-dev] Potential issue with async sockets in Raspberry ARM – Mono 4.4.0 Hi, I’m posting here hoping someone can throw some light into the problem. I'm a little bit lost now. Our server code (Plastic SCM) running on Raspberry on Mono 4.4.0 (exact version below) ends up not awaking from socket.BeginReceive after a while. I mean, client connects and requests data in 4MB chunks, and depending on the run, it can transfer a few GB but it ends up not awaking. The client just sits waiting on a “socket recv” but the server doesn’t answer. Connection is established (can be checked at OS level). The code could be simplified as follows: mSocket.BeginReceive(buffer, 0, 0, SocketFlags.None, OnMessageReceived, null); _OnMessageReceived_ => does the EndReceive and then reads data and enqueues the request for a threadpool to attend it. Once the request is attented and the response sent, BeginReceive is invoked again. Important: all “BeginReceive()” calls are done from the same thread which NEVER dies. All we use the BeginReceive for is to decouple socket from thread, so we do not have a 1-1. You see we do pass “zero” as bytes to read, because all we want to achieve is to get awakened when new data is received, then just read using blocking calls (no async). I’m asking if there could be something about Mono because I read 4.6 release notes (http://www.mono-project.com/docs/about-mono/releases/4.6.0/) and the “atomic fixes for ARM64”. Could it be related somehow. The same code runs on Windows and Linux PCs (even Macs) without issues. We use the same code on all our production servers and even our Cloud roles, and we are not aware of issues. Now we are testing a new faster storage and using Raspberry to check how fast it goes on slower hardware. Complete Mono version: Mono JIT compiler version 4.4.0 (tarball Tue Jun 14 13:44:06 UTC 2016) Copyright (C) 2002-2014 Novell, Inc, Xamarin Inc and Contributors. www.mono-project.com TLS: __thread SIGSEGV: normal Notifications: epoll Architecture: armel,vfp+hard Disabled: none Misc: softdebug LLVM: supported, not enabled. GC: sgen . Thanks! Pablo Santos www.plasticscm.com _______________________________________________ Mono-devel-list mailing list Mono-devel-list@lists.dot.net http://lists.dot.net/mailman/listinfo/mono-devel-list _______________________________________________ Mono-devel-list mailing list Mono-devel-list@lists.dot.net http://lists.dot.net/mailman/listinfo/mono-devel-list |
_______________________________________________ Mono-devel-list mailing list Mono-devel-list@lists.dot.net http://lists.dot.net/mailman/listinfo/mono-devel-list
