Bug #14434: IO#reopen fails after EPIPE

• Status: Feedback
• Assignee: -
• Target version: -
• Backport: 2.3: UNKNOWN, 2.4: UNKNOWN, 2.5: UNKNOWN
• Tags: core, patch

[nobu (Nobuyoshi Nakada)]

Consider the following code, which emulates yes | head -1.

IO.popen("head -1", "w") do |f|
  f.sync = false
  stdout = STDOUT.dup
  STDOUT.reopen(f)
  loop{puts "y"} rescue break $!
ensure
  STDOUT.reopen(stdout)
end

This fails with Errno::EPIPE in IO#reopen.

Traceback (most recent call last):
	4: from -:1:in `<main>'
	3: from -:1:in `popen'
	2: from -:7:in `block in <main>'
	1: from -:7:in `ensure in block in <main>'
-:7:in `reopen': Broken pipe (Errno::EPIPE)

BTW, this "broken pipe" IO can't even close.

r, w = IO.pipe
w.sync = false # make buffered
w.print("foo")
r.close
# w.reopen(STDOUT) rescue p $! #=> #<Errno::EPIPE: Broken pipe>
w.close rescue p $! #=> #<Errno::EPIPE: Broken pipe>

This is caused by flushing buffered data.
I think there is no way to recover "broken pipe" FD, so nothing can be done for the remained data and should be discarded gently.

diff --git i/io.c w/io.c
index 0a4e66f5ed..5eca6762f2 100644
--- i/io.c
+++ w/io.c
@@ -7150,8 +7150,7 @@ io_reopen(VALUE io, VALUE nfile)
 	}
     }
     if (fptr->mode & FMODE_WRITABLE) {
-        if (io_fflush(fptr) < 0)
-            rb_sys_fail(0);
+        fptr_finalize_flush(fptr, TRUE, FALSE);
     }
     else {
 	io_tell(fptr);
@@ -7177,7 +7176,8 @@ io_reopen(VALUE io, VALUE nfile)
     if (fd != fd2) {
 	if (IS_PREP_STDIO(fptr) || fd <= 2 || !fptr->stdio_file) {
 	    /* need to keep FILE objects of stdin, stdout and stderr */
-	    if (rb_cloexec_dup2(fd2, fd) < 0)
+	    fd = (fd < 0) ? rb_cloexec_dup(fd2) : rb_cloexec_dup2(fd2, fd);
+	    if (fd < 0)
 		rb_sys_fail_path(orig->pathv);
             rb_update_max_fd(fd);
 	}

[jeremyevans0 (Jeremy Evans)]

I tried updating this patch for the current master branch, but it breaks test_reopen_inherit:

  1) Failure:
TestIO#test_reopen_inherit [/ruby/test/ruby/test_io.rb:2385]:
<"outerr"> expected but was
<"">.

@nobu (Nobuyoshi Nakada), do you think there is a better way to fix this? I'm not sure the current behavior is a bug, if the pipe is broken, it seems reasonable for IO#reopen to raise an exception.

Here's the patch I used:

diff --git a/io.c b/io.c
index 0d6e217857..e41ca74db7 100644
--- a/io.c
+++ b/io.c
@@ -7372,8 +7372,7 @@ io_reopen(VALUE io, VALUE nfile)
        }
     }
     if (fptr->mode & FMODE_WRITABLE) {
-        if (io_fflush(fptr) < 0)
-            rb_sys_fail_on_write(fptr);
+        fptr_finalize_flush(fptr, TRUE, FALSE, 0);
     }
     else {
         flush_before_seek(fptr);
@@ -7399,7 +7398,8 @@ io_reopen(VALUE io, VALUE nfile)
     if (fd != fd2) {
        if (IS_PREP_STDIO(fptr) || fd <= 2 || !fptr->stdio_file) {
            /* need to keep FILE objects of stdin, stdout and stderr */
-           if (rb_cloexec_dup2(fd2, fd) < 0)
+            fd = (fd < 0) ? rb_cloexec_dup(fd2) : rb_cloexec_dup2(fd2, fd);
+            if (fd < 0)
                rb_sys_fail_path(orig->pathv);
             rb_update_max_fd(fd);
        }

[nobu (Nobuyoshi Nakada)]

I think this is a bug definitely, but non-critical.

[mame (Yusuke Endoh)]

• Status changed from Open to Feedback

I asked @nobu (Nobuyoshi Nakada) about the current status of this ticket, and he said "the current behavior is definitely a bug, but I am not able to think of a good way to fix it now." I'd like to set this as feedback for a good fix.