On 20.02.2017 01:56, Steve Cotton wrote: > On Sun, Feb 19, 2017 at 09:38:12PM +0100, Steve Cotton wrote: >> Sorry, but it's turned out that my patch either doesn't completely >> avoid the bug, or doesn't avoid another bug which gives the same error >> message. The build has failed on arm64. > > I've tried to reproduce this locally (on amd64, not arm64). With my > patch, I can't replicate the failure. Santiago, please would you test > how it fares on your autobuilders? > > Testing by removing my patch and trying to debug the root cause, I > haven't found the cause yet. But I think it will need a complex patch > to either libsdl1.2 or xvfb, and I don't think this bug justifies any > complex patch during the freeze. > > It seems that one of the XOpenDisplay calls in SDL's X11VideoInit fails. > Having a GCC breakpoint at the start of X11VideoInit, or running MakeDat > under strace, makes the bug unreproducible. Adding a long sleep to the > last point in the ri-li package's code before SDL_Init doesn't help. > Running another SDL programe first in the same xvfb server doesn't > change anything, so it seems that the server can't be primed in advance. > > Markus, I'm really sorry that what looked like a risk-free patch has > caused a failed rebuild. Depending on the TC's ruling, does it sound > sensible to say that 2.0.1+ds-4 is in Stretch, and -5 doesn't affect -4?
Hi Steve, no worries and thanks for providing a helping hand here. I've also tried a couple of different options in the past hours. I think your initial patch wasn't completely wrong and the underlying issue has something to do how SDL initializes its subsystems but I have reverted it for now. I have read about issues when using SDL in virtual machines but I have found only one clue namely to manually link with -lX11 to avoid this kind of error. I have tried this solution at least 20 times on asachi.debian.org (arm64 porterbox) and couldn't reproduce the build failure anymore. I have uploaded another revision a few minutes ago. -5 and -6 don't affect Stretch. I don't know about a TC ruling but since it is obvious that the claim of "99 %" build failures is not true I stand by my opinion that this is not release critical. We could also stop building the data from source but this isn't something I would call an improvement over the current situation. Regards, Markus
signature.asc
Description: OpenPGP digital signature
