On Wed, Nov 21, 2018 at 10:37 PM Max Kirillov <max@xxxxxxxxxx> wrote: > > On Wed, Nov 21, 2018 at 05:04:25PM -0800, Carlo Arenas wrote: > > the last child of its children long gone with an error as shown by : > > > > 9255 1 git-http-backend CALL close(1) > ... > > 9255 1 git-http-backend CALL write(2,0xbfb2a604,0x36) > > 9255 1 git-http-backend GIO fd 2 wrote 54 bytes > > "fatal: request ended in the middle of the gzip stream\n" > > This should be some other test than push_plain, some of the > gzip related ones. Are there other tests failing? it should, but I should note that for test 9 to fail, then either (or both) tests 7 and 8 should first succeed; not that I'd seen any other test fail (after I locally patched the perl path, of course) even when reordering them and while making sure tests 1 and 2 run first to create the dependencies for the rest Peff, could you elaborate on your "load testing" setup? which could give us any hints on what to look for?, FWIW I hadn't been able to reproduce the problem anywhere else (and not for a lack of trying) > > 9255 1 git-http-backend RET write 54/0x36 > > 9255 1 git-http-backend CALL write(1,0xb781f0e0,0x94) > > 9255 1 git-http-backend RET write -1 errno 9 Bad file descriptor > > This is interesting. http-backend for some reason closes its > stdout. Here it then tries to write there something. I have > not seen it in my push_plain run. Maybe it worth redirecting instead > to stderr, to avoid losing some diagnostics? that should help with the garbled output from stderr, AFAIK the process API allows creating a pipe specifically for that with would be better than redirecting stderr into stdout. the fact we got EBADF means that there is a problem somewhere though in the way the previous failure that closed stdout got handled (which should had been most likely in the call to die) Carlo PS. upstreaming the PERL_PATH fix is likely to be good to do soonish as I presume at least all BSD might be affected, let me know if you would rather me do that instead as I suspect we might be deadlocked otherwise ;)
