Re: Intermittent Issues with RCS binary co

[Paul Eggert]

"Meinert, Debora L." <address@hidden> writes:

> The filesystems are VXFS (journaled filesystems).

Have you applied all the relevant operating system patches?
Historically, VxFS has had several issues with mmap on HP-UX 11.
(I don't use HP-UX so I'm not really an expert here.)

> Are the RCS files themselves OK?  That is, do they contain all the
> data in question?
> Yes, as far as we can tell.  The filesystem that contains all our RCS files
> is .../.../archives and from what we can see the files are in tack there.
> They get corrupted after the checkout command. (above)

When you say "they get corrupted" do you mean that the working files
are corrupted, but the RCS files are OK?  The command that you wrote:

  co -r1.0 "XXGLPMCA.rdf" > "XXGLPMCA.rdf.out" 2>&1

should not modify the RCS file at all.  If it is modifying the RCS
file, there is something seriously wrong.  So I'll assume that it's
the working files that are incorrect.

> Because it is intermittent, it is not something we can test at this time.
> When it happens we can generally reproduce it using the same files.

This suggests that it's a VxFS bug.  Possibly VxFS is getting confused
with mmap, and is caching the wrong data.

> Any other suggestions while we wait.  

Here's what I'd do.  I'd build a separate copy of RCS in which mmap is
turned off.  (Manually modify 'has_mmap' to be 0, remove all the .o
files, and then type 'make'.)  Then, when the problem recurs, try
using this specially modified copy of RCS.  If it works correctly, but
the original version of RCS does not, that strongly suggests an
operating system or file system bug, in the area of mmap.

Or, you can simply switch to the specially modified copy of RCS.  If
the problem never recurs, then it's most likely an mmap-related bug.