Seg fault upon interrupting disk io in progress

[amitmurthy]

To simulate it, run the following code, interrupt once and run the function again.

julia> function file_io(secs)
           t1 = time()
           wstr = "Hello World"
           fn = tempname()
       #    fn = "/home/amitm/tmp/ec2/" * split(tempname(), "/")[3]
           while((time() - t1) < secs)
               fio = open(fn, "w+")
               write(fio, wstr)
               flush(fio)
               close(fio)
               rm(fn)
           end
       end
file_io (generic function with 1 method)

julia> file_io(60)
^CERROR: interrupt
 in unlink at fs.jl:93

julia> file_io(60)
Segmentation fault

NOTE: Came across this while I was simulating access to a slow filesystem (by using sshfs (via fuse) to mount an EC2 directory locally).

When I tried with the sshfs mountpoint - /home/amitm/tmp/ec2, julia segfaults the first time it is interrupted itself.

[amitmurthy]

Though I don't see it while testing with the local file system now, I still see it on a 'slow' filesystem simulated by sshfs mounting a directory on an EC2 instance.

valgrind output is here - https://gist.github.com/amitmurthy/8431825

[JeffBezanson]

Could you try putting JL_SIGATOMIC_BEGIN() / JL_SIGATOMIC_END() around jl_uv_writecb (in jl_uv.c)?

[amitmurthy]

It improved the issue in the sense that it took me 4 counts of running file_io(60) and interrupting it before the seg fault happended, but it did nevertheless.

This was the change I tried with :

DLLEXPORT void jl_uv_writecb(uv_write_t* req, int status)
{
    JL_SIGATOMIC_BEGIN();
    JULIA_CB(writecb, req->handle->data, 2, CB_PTR, req, CB_INT32, status)
    free(req);
    JL_SIGATOMIC_END();
    (void)ret;
}

Any other function to wrap with SIGATOMIC and try ?

[JeffBezanson]

I would try adding sigatomic_begin(), sigatomic_end() around calls to ios_file, ios_write, and ios_close in io.jl.

But, this may be something of a lost cause as there is no general way to guarantee that everything is interrupt-safe. #2622 suggests wrapping all ccalls, but I'm kind of against this since it adds overhead and probably still won't make everything work.