[PATCH] t9010: Open backflow FIFO once to work around kernel race condition

[PATCH] t9010: Open backflow FIFO once to work around kernel race condition

[Anders Kaseorg]

Linux’s implementation of open() on FIFOs suffers from a race
condition (see http://marc.info/?l=linux-kernel&m=134071285509470 )
that causes t9010-svn-fe.sh to occasionally fail with
  t9010-svn-fe.sh: 27: eval: cannot create backflow: Interrupted system call
Although I can’t reproduce this locally, it happens particularly often
on the Launchpad PPA builders.

Sidestep this problem by opening the backflow FIFO once for
read+write.  Also, replace the stream FIFO with a shell pipe so we
don’t have to do manual process management.

Signed-off-by: Anders Kaseorg <andersk@mit.edu>
---
 t/t9010-svn-fe.sh | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/t/t9010-svn-fe.sh b/t/t9010-svn-fe.sh
index b7eed24..c52a7de 100755
--- a/t/t9010-svn-fe.sh
+++ b/t/t9010-svn-fe.sh
@@ -11,9 +11,9 @@ reinit_git () {
 		return 127
 	fi
 	rm -fr .git &&
-	rm -f stream backflow &&
+	rm -f backflow &&
 	git init &&
-	mkfifo stream backflow
+	mkfifo backflow
 }
 
 try_dump () {
@@ -22,10 +22,9 @@ try_dump () {
 	maybe_fail_fi=${3:+test_$3} &&
 
 	{
-		$maybe_fail_svnfe test-svn-fe "$input" >stream 3<backflow &
-	} &&
-	$maybe_fail_fi git fast-import --cat-blob-fd=3 <stream 3>backflow &&
-	wait $!
+		$maybe_fail_svnfe test-svn-fe "$input" |
+		$maybe_fail_fi git fast-import --cat-blob-fd=3
+	} 3<>backflow
 }
 
 properties () {
-- 
1.7.11.1

Re: [PATCH] t9010: Open backflow FIFO once to work around kernel race condition

[Jonathan Nieder]

Hi,

Quick first impressions:

Anders Kaseorg wrote:

> Sidestep this problem by opening the backflow FIFO once for
> read+write.

Is that portable?

>             Also, replace the stream FIFO with a shell pipe so we
> don’t have to do manual process management.

Doesn't this mean we wouldn't notice if test-svn-fe crashes?

Maybe it would be simpler to make test-svn-fe launch fast-import as a
child, avoiding the fifos altogether.

Thanks again for tracking this down, by the way.
Jonathan

Re: [PATCH] t9010: Open backflow FIFO once to work around kernel race condition

[Anders Kaseorg]

On Tue, 26 Jun 2012, Jonathan Nieder wrote:
> Anders Kaseorg wrote:
> > Sidestep this problem by opening the backflow FIFO once for
> > read+write.
> 
> Is that portable?

Presumably; it’s POSIX, at least.
http://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap02.html#tag_18_07_07

> Doesn't this mean we wouldn't notice if test-svn-fe crashes?

Oh hmm, good point.

Anders

Re: [PATCH] t9010: Open backflow FIFO once to work around kernel race condition

[Junio C Hamano]

Jonathan Nieder <jrnieder@gmail.com> writes:

> Hi,
>
> Quick first impressions:
>
> Anders Kaseorg wrote:
>
>> Sidestep this problem by opening the backflow FIFO once for
>> read+write.
>
> Is that portable?

If you mean [n]<>word, it probably is, even though it is somewhat
unusual and I wouldn't be surprised if implementations get wrong.

>>             Also, replace the stream FIFO with a shell pipe so we
>> don’t have to do manual process management.
>
> Doesn't this mean we wouldn't notice if test-svn-fe crashes?

Very good point.

Re: [PATCH] t9010: Open backflow FIFO once to work around kernel race condition

[Stefano Lattarini]

[Adding bug-autoconf]

On 06/27/2012 12:40 AM, Jonathan Nieder wrote:
> Hi,
> 
> Quick first impressions:
> 
> Anders Kaseorg wrote:
> 
>> Sidestep this problem by opening the backflow FIFO once for
>> read+write.
> 
> Is that portable?
>
According to the Autoconf manual, no:

    Some shells, like ash, don't recognize bi-directional redirection (‘<>’).
    And even on shells that recognize it, it is not portable to use on fifos:
    Posix does not require read-write support for named pipes, and Cygwin
    does not support it:

       $ mkfifo fifo
       $ exec 5<>fifo
       $ echo hi >&5
       bash: echo: write error: Communication error on send

But while the issue about Cygwin might still be relevant, the one about
ash seems to be out-of-date: I've verified that the "exec 5<>fifo"
command works with both dash 0.5.2 and dash 0.5.5.1 (that's why I'm
CC:ing bug-autoconf).

Regards,
  Stefano