Created attachment 1127783 [details] docker-strace Description of problem: Running docker in a fully up-to-date rawhide vm on F23 host, docker is segfaulting very early in startup on sched_getaffinity(0, 8192, [0]) = 8 Full strace is attached Version-Release number of selected component (if applicable): docker-1.10.1-3.git49805e4.fc24.x86_64 How reproducible: Always Steps to Reproduce: 1. run /usr/bin/docker Actual results: seg fault Expected results: no seg fault Additional info: strace is attached
If you run an older version of docker are you seeing this?
Created attachment 1127986 [details] docker-coredump.tar.gz
Aye, it's still a thing if I downgrade to the last 1.10.0 build in koji. I went ahead reupgraded and grabbed the abrt output.
Running gdb: (gdb) bt #0 _dl_lookup_symbol_x (undef_name=0x4056e9 "mmap", undef_map=0x7f6ae3f5b128, ref=ref@entry=0x7ffd6b3c8920, symbol_scope=0x7f6ae3f5b480, version=0x7f6ae3e9c450, type_class=type_class@entry=1, flags=1, skip_map=0x0) at dl-lookup.c:809 #1 0x00007f6ae3d43ea4 in _dl_fixup (l=<optimized out>, reloc_arg=<optimized out>) at ../elf/dl-runtime.c:111 #2 0x00007f6ae3d4c2af in _dl_runtime_resolve_sse () at ../sysdeps/x86_64/dl-trampoline.h:112 This is starting to look like a bug in glibc.
Does docker work if you set LD_BIND_NOW=1?
(In reply to Florian Weimer from comment #5) > Does docker work if you set LD_BIND_NOW=1? Gave that a try, same result.
(In reply to John Dulaney from comment #4) > Running gdb: > > (gdb) bt > #0 _dl_lookup_symbol_x (undef_name=0x4056e9 "mmap", > undef_map=0x7f6ae3f5b128, ref=ref@entry=0x7ffd6b3c8920, > symbol_scope=0x7f6ae3f5b480, version=0x7f6ae3e9c450, > type_class=type_class@entry=1, flags=1, skip_map=0x0) at dl-lookup.c:809 > #1 0x00007f6ae3d43ea4 in _dl_fixup (l=<optimized out>, reloc_arg=<optimized > out>) at ../elf/dl-runtime.c:111 > #2 0x00007f6ae3d4c2af in _dl_runtime_resolve_sse () at > ../sysdeps/x86_64/dl-trampoline.h:112 > > This is starting to look like a bug in glibc. I can reproduce this.
There is a known golang ABI issue, see bug 1304591 comment 7. I'm surprised LD_BIND_NOW=1 doesn't work around this. Can you set a breakpoint on x_cgo_mmap and see if you can get a longer backtrace?
(In reply to Florian Weimer from comment #8) > There is a known golang ABI issue, see bug 1304591 comment 7. > > I'm surprised LD_BIND_NOW=1 doesn't work around this. Can you set a > breakpoint on x_cgo_mmap and see if you can get a longer backtrace? Using LD_BIND_NOW=1 resolves the issue. In comment 7 (of this issue) I only meant to say that I could reproduce the crash on rawhide, but had not don anything further. The crash backtrace is is exactly as in bug 1304591 and it is indeed an incorrect stack alignment issue with the hand-written Go assembly. *** This bug has been marked as a duplicate of bug 1304591 ***