I send 2 messages to bug-coreutils@gnu.org in last 2 weeks but that never
reach the list.
I don't know why as this had work in the past and I receive message from
that same address and was able to reply. Isn't the antispam too good? So I
am a bit forced to try with a direct coreutils message.

Anyway, here is the issue. I compile the same code on different machines and
that's made in a chroot a la LFS. With sparse-fiemap-test.patch, all tests
are ok on three machines but randomly fail on the fourth (all x86 32-bits).
The only difference at this stage between the different machines should be
the running kernel (same glibc-2.11.3 is compiled before coreutils).
This work on ubuntu 2.6.32-26-generic, 2.6.27.57 vanilia kernel, Centos-5.5
2.6.18-194.32.1.el5 but fail in debian v5 2.6.26-2-686

Tested as root with
for i in {1..20}; do make -s check -C tests TESTS=cp/sparse-to-pipe
2>/dev/null;done | grep '1 test'
show this result on the debian machine
1 of 1 test failed
1 of 1 test failed
1 test passed
1 test passed
1 of 1 test failed
1 of 1 test failed
1 of 1 test failed
1 of 1 test failed
1 test passed
1 test passed
1 of 1 test failed
1 test passed
1 test passed
1 of 1 test failed
1 of 1 test failed
1 of 1 test failed
1 test passed
1 test passed
1 of 1 test failed
1 of 1 test failed

FS is ext3
failure is
+ require_sparse_support_
+ test 0 = 0
+ t=sparse.17085
+ dd bs=1 seek=128K of=sparse.17085
++ du -sk sparse.17085
+ set x 0 sparse.17085
+ kb_size=0
+ rm -f sparse.17085
+ test 0 -ge 128
+ mkfifo_or_skip_ pipe
+ test 1 = 1
+ mkfifo pipe
+ timeout 10 cat pipe
+ truncate -s1M sparse
+ cp sparse pipe
+ cmp sparse copy
cmp: EOF on copy
+ fail=1
+ Exit 1

The fun is that if I try to strace the cmp, this change the cmp test to
reliabily work

If I add a dd if=/dev/null of=pipe conv=notrunc,fdatasync, test reliabily
work too.
So I added that in attached patch.

Gilles