This is the mail archive of the cygwin mailing list for the Cygwin project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Cygwin1.dll 1.7.0-5x: RSYNC failures in close() system call on pipe file descriptors

cygwin <karl <at>> writes:

> -               if (dup2(to_child_pipe[0], STDIN_FILENO) < 0 ||
> -                   close(to_child_pipe[1]) < 0 ||
> -                   close(from_child_pipe[0]) < 0 ||
> -                   dup2(from_child_pipe[1], STDOUT_FILENO) < 0) {
> -                       rsyserr(FERROR, errno, "Failed to dup/close");
> +               if (dup2(to_child_pipe[0], STDIN_FILENO) < 0) {
> +                       rsyserr(FERROR, errno, "Failed to dup2(to-0)");
>                         exit_cleanup(RERR_IPC);
>                 }
> -               if (to_child_pipe[0] != STDIN_FILENO)
> -                       close(to_child_pipe[0]);
> -               if (from_child_pipe[1] != STDOUT_FILENO)
> -                       close(from_child_pipe[1]);

Ouch.  rsync has a bug even without your patch: if you run rsync with stdout 
closed (as opposed to redirected), then to_child_pipe[0] can be 1, and you end 
up closing the just-duplicated from_child_pipe[1] that was copied to the new 
stdout.  To be correct, rsync must compare to_child_pipe[0] against both 0 and 
1, not just 0 (likewise for from_child_pipe[1] against both 0 and 1, not just 
1).  Amazing how often this type of bug reappears, too - historical cygwin has 
had it in the past with newlib's popen implementation (I fixed it 2006-08-22), 
as well as m4 1.4.13-1 having it due to code from gnulib (I fixed it 2009-07-

As an aside, it's interesting to note that on mingw, dup2(1,1) succeeds, but 
has the result of corrupting the fd, such that if you later do dup2(1,2), it 
hangs.  But the fact that mishandled attempts to duplicate Windows handles can 
lead to weird bugs sure lends credence to the idea that BLODA may be making the 
matter worse when we duplicate an open socket handle.

Eric Blake

Problem reports:
Unsubscribe info:

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]