segfault running julialang/julia#3775

[ViralBShah]

I get a segfault running this from JuliaLang/julia#3775

@sprofile let 
              tic()
              n = 10000;
              m = 100;
              l = 10;
              A = zeros(n,m);
              for i=1:n
                  for j=1:m
                      c = rand(10,10);
                      A[i,j] = c[2,5];
                  end
                  h = find(A[i,:].>0.5);
                  for k=1:l
                      c = fft(A[:,l]);
                      d = min(abs(c));
                  end
              end
              toc()
              end;

Program received signal EXC_BAD_ACCESS, Could not access memory.
Reason: KERN_INVALID_ADDRESS at address: 0x0000000000000f78
0x00007fff8d327610 in libunwind::CompactUnwinder_x86_64<libunwind::LocalAddressSpace>::stepWithCompactEncodingRBPFrame ()
(gdb) bt
#0  0x00007fff8d327610 in libunwind::CompactUnwinder_x86_64<libunwind::LocalAddressSpace>::stepWithCompactEncodingRBPFrame ()
#1  0x00007fff8d32758f in libunwind::CompactUnwinder_x86_64<libunwind::LocalAddressSpace>::stepWithCompactEncoding ()
#2  0x00007fff8d32750e in libunwind::UnwindCursor<libunwind::LocalAddressSpace, libunwind::Registers_x86_64>::step ()
#3  0x00000001002ccbb6 in rec_backtrace (data=0x104037b88, maxsize=971470) at task.c:533
#4  0x00000001011fcd18 in sprofile_bt ()
#5  <signal handler called>
#6  0x00007fff92bf8831 in free ()
#7  0x000000010ba03dc8 in hinsert ()

[timholy]

Did not segfault for me on Kubuntu 12.04. Are you on OSX? If so, this is presumably the same as #14.

[ViralBShah]

Yes, I am on OS X, and I suspect it is the same thing.

[timholy]

OK, closing. If you get evidence that this is somehow different, feel free to reopen.

[ViralBShah]

Sure. I just wanted it recorded here, so that it is yet another test case, if we do ever find a way to fix #14. I am hoping that @loladiro 's recent work on OS X backtraces might help.

[Keno]

@timholy You might want to try if you can make this work for OS X: http://web.mit.edu/darwin/src/modules/xnu/osfmk/man/task_sample.html . The basic implementation should be similar to my recent pull request against julia. If you need help figuring this out, feel free to ping me.

[Keno]

There's also http://web.mit.edu/darwin/src/modules/xnu/osfmk/man/thread_sample.html, which might actually be better in this case.

[timholy]

Thanks @loladiro. I may give this a go at some point, though I'll need willing testers on the Mac.

From another perspective, what's probably happening is that Profile is revealing a bug in Julia's backtraces. For what may be a separate problem, that eventually became clear in JuliaLang/julia#3469. So in a sense maybe it's better not to fix it by working around the problem specifically in Profile, but to focus our attention on the Julia side.

[ViralBShah]

I am happy to test!

[timholy]

:-)

[vtjnash]

It's interesting to walk the stack for this in gdb -- the frame before it faults has the instruction pointer off by one. I'm not sure how the sigtramp function would cause this (it appears to be is approximately apple's version of sigreturn).