Re: dd: reblock failure -- sporadic failures

[Jim Meyering]

Bob Proulx <address@hidden> wrote:
> This seems very strange.  Periodically I have seen the dd reblock test
> fail.  It isn't a hard failure and that seems like a critical clue to
> me.  But it also seems quite strange.
>
>   http://buildbot.proulx.com:9000/i686-gnu-linux-full/builds/990/step-test/0
>
>   + dd bs=3 ibs=1 obs=1
>   ...
>   + diff -u err exp-no-reblock
>   --- err         2008-11-29 07:07:22.000000000 -0700
>   +++ exp-no-reblock   2008-11-29 07:07:21.000000000 -0700
>   @@ -1,3 +1,3 @@
>   -1+1 records in
>   -1+1 records out
>   +0+2 records in
>   +0+2 records out
>    4 bytes (4 B) copied
>   + fail=1

Hi Bob,

This should fix it in most cases.
There will always be a race when the consumer can start
far later than the producer.

I demonstrated the failure locally by using a shorter sleep interval:

    while :; do printf .;(echo x; sleep .01; echo y) \
      | dd bs=3 ibs=1 obs=1 2>&1 |grep -F '1+1 records in' && break; done
    ..........................................................................\
    ..........................................................................\
    ..........................................................................\
    ..............................................1+1 records in

Here's the fix in my queue:

>From 149a55b6ff4137668b9a945fa6cab30bdd7933ea Mon Sep 17 00:00:00 2001
From: Jim Meyering <address@hidden>
Date: Sun, 30 Nov 2008 18:43:55 +0100
Subject: [PATCH] tests: avoid occasional dd/reblock failure

* tests/dd/reblock: Sleep longer to avoid a race condition.
Reported by Bob Proulx.
---
 tests/dd/reblock |    5 ++++-
 1 files changed, 4 insertions(+), 1 deletions(-)

diff --git a/tests/dd/reblock b/tests/dd/reblock
index 542529a..ef9e036 100755
--- a/tests/dd/reblock
+++ b/tests/dd/reblock
@@ -44,7 +44,10 @@ sed 's/,.*//' err > k && mv k err
 compare err exp-reblock || fail=1

 # Demonstrate that bs=N supersedes even following ibs= and obs= settings.
-(echo x; sleep .1; echo y) | dd bs=3 ibs=1 obs=1 > out 2> err || fail=1
+# Choosing a delay of 0.1 would result in an occasional lost race where
+# the consumer's first read would consume 3 bytes rather than the expected 2.
+# Not wanting to sleep a full second, I'll raise that to 0.3.
+(printf ab; sleep .3; printf cd) | dd bs=3 ibs=1 obs=1 > out 2> err || fail=1
 sed 's/,.*//' err > k && mv k err
 compare err exp-no-reblock || fail=1

--
1.6.0.4.1101.g642f8