Re: mremap() use is racy

[Hugh Dickins <hugh@veritas.com>]

On Tue, 23 Aug 2005, Ulrich Drepper wrote:
> Linus Torvalds wrote:
> > Actually, it should be pretty much as valid as using mremap - ie it works 
> > on Linux. 
> > 
> > Especially if you use MAP_SHARED, you don't even need to mprotect 
> > anything: you'll get a nice SIGBUS if you ever try to access past the last 
> > page that maps the file.
> 
> If you guarantee this (and test for this) it's fine with me.  The POSIX
> spec explicitly leaves this undefined and requiring to use mremap()
> would be a nice way to work around this without allowing the
> introduction of undefined behavior into programs.  I probably would
> prefer to use mremap() since this makes it clear what should happen but
> I can live with using the too-large mapping.

I don't have spec to hand to quote, but at least 10 years ago the SIGBUS
beyond end of file was required behaviour, on both MAP_SHARED and more
surprisingly MAP_PRIVATE mappings of a file.  Well, you don't get SIGBUS
on any part of the page that covers the end of the file: perhaps what
you read as undefined in the spec is addressing that tail page issue.

We guarantee it: it's a kernel bug if it doesn't work.

Hugh