Linux kernel expertise needed for Bug 1941063

[Alexandre Ferrieux]

Hi,

The new [ Bug 1941063 ] "test exec-9.7 fails randomly", reported for
current CVS HEAD on bi-pro Ubuntu 7.0.4, displays a strange "unmoving
seek pointer" issue:

https://sourceforge.net/tracker/?func=detail&atid=110894&aid=1941063&group_id=10894

One possible interpretation of the thing happens to be a kernel bug
(!), characterized by the following:

- a regular file in /tmp is opened for writing
- the fd (CLOEXEC cleared) is inherited by two children
- the two children *simultaneously* write to this fd
- in theory, they should share a single seek pointer, and the
two writes should be in sequence
BUT
- here (two processors), it looks like *sometimes* (==race
condition) each processor+process goes with its own local copy of the
seek pointer, both starting at zero, and both writing at the same
offsets. When both children write the same string, the file ends up
with one single copy of this string ! Just like a missing master
lock...

Of course, it sounds too big to be true, considering the cumulative
experience of the Linux community.
But could it be a recent regression, due to the global effort for
lower contention in the kernel ?

Any insight, dear Linux kernel hackers ?

-Alex