On Wed, Jan 07, 2015 at 11:48:59AM +0000, Finucane, Stephen wrote: > > We've heard similar reports before but it's challenging to find the > > problem. I can't reproduce the problem on my 32-bit Debian system by > > just, for example, switching to GCC 4.6. > > From [here]( http://openvswitch.org/pipermail/dev/2014-December/049833.html), > you'll see I have a number of boards failing and one passing. I'll take > config files from one of each (note - the kernels of the failing boards have > been updated) > > > What's in config.h and config.log? > > The 'config.h' files are identical for both machines. The > 'config.log' is slightly different, but this seems to be due to SSL > support being available on one board but not the other. I've > attached (as attachments due the size of these files - hope this is > ok?) both version irregardless.
I mainly wanted to see whether one of them had the GCC 4.x atomics completely disabled, but both of them define HAVE_GCC4_ATOMICS in their config.h, so that's not the problem. > > How many cores does the system running the build have? > WORKING: 8 cores w/ 8 threads, dual socket (Intel(R) Xeon(R) CPU E5-2660 0). > > BROKEN: 10 cores w/ 10 threads, dual socket (Intel(R) Xeon(R) CPU E5-2680 v2). With a single core, the test-atomic program might run very slowly because of constant switching between threads (although I haven't personally observed this, someone else told me that it happens), but that's obviously not a problem with that many cores. Jarno's idea of getting a disassembly is a good idea, but it might be easier if you just pass along test-atomic.o from a broken machine. Then we can just disassemble it and otherwise examine it for ourselves. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
