Source code needs to be compiled with -D_REENTRANT

[wereHamster]

Is there a bug somewhere or can that error be silenced without any side effects? FYI, errno 11 is Resource temporarily unavailable

[charles]

wereHamster: it looks like we could silence this DHT error without any side-effects. cc'ing jch for a second opinion...

[charles]

maybe we could log the first error, but not log subsequent identical errors... jch, werehamster, how does this look to you?

[wereHamster]

Simply hiding the error will not fix the underlying bug, event_callback() will still be called dozens of times each millisecond. It looks like dht_periodic() can fail with EAGAIN (errno 11), in that case try again right away. With my patch the web UI feels much more responsive because the libevent loop isn't overloaded with events from DHT sockets.

[jch]

I've just checked the code (in third-party/dht/dht.c), and I don't see how dht_periodic can possibly return EAGAIN. (It's certainly not supposed to -- "call me again soon" is signalled by returning 1 with tosleep set to 0.)

Which version of the DHT is that? (Check dht/CHANGES.)

--Juliusz

[wereHamster]

15 December 2009: dht-0.13

[jch]

I don't see any way this version could return -1 with errno being EAGAIN.

Charles, any ideas?

--Juliusz

[wereHamster]

Don't forget, this is on OpenSolaris?, and that OS likes to screw with its users in thousand different ways. I'll add a few printf()'s to see where the errno comes from.

[charles]

wereHamster: I don't see any way for dht_periodic() to return -1 with errno set to EAGAIN. Good luck with the printf()s..

I'd like to get this ticket resolved before 1.80... if the two of you can suss it out in the next day or two, that would be great.

[wereHamster]

The friendly folks (*cough*) in #opensolaris suggested that it might be because third-party/dht was compiled without -D_REENTRANT. I've modified the Makefile there and the daemon has been running then whole night now without dht_periodic not even once returning with errno set to EAGAIN.

diff --git a/third-party/dht/Makefile.am b/third-party/dht/Makefile.am
index c9e6477..f69a1d7 100644
--- a/third-party/dht/Makefile.am
+++ b/third-party/dht/Makefile.am
@@ -1,3 +1,6 @@
+
+AM_CFLAGS = @PTHREAD_CFLAGS@
+
 noinst_LIBRARIES = libdht.a
 libdht_a_SOURCES = dht.c
 noinst_HEADERS = dht.h

[wereHamster]

I think the optimization in the dht event callback (if -> while in case dht_periodic returns with EINTR) is still useful, so I added it to the new patch.

[charles]

wereHamster: should the other third-party libraries under Transmission's configure script (libnatpmp, miniupnpc) also have @PTHREAD_CFLAGS@ in their AM_CFLAGS?

[wereHamster]

Likely yes. It can't hurt anyway.

[charles]

It looks harmless to me, both in code and from testing it out on Linux yesterday & overnight. I think it's safe to commit this despite the freeze.

Reassigning Version from 1.76+ to 1.70 since this is the first time we've compiled DHT with @PTHREAD_CFLAGS@... this behavior should have existed on Solaris prior to 1.76+

[charles]

Fixed in trunk for 1.80 by r9942