Re: [PATCH v1 0/7] Remove in-tree usage of MAP_DENYWRITE

From: Florian Weimer
Date: Fri Aug 13 2021 - 16:51:41 EST

* Eric W. Biederman:

> Florian Weimer, would it be possible to get glibc's implementation to use
> MAP_SHARED? Just so people reading the code know what to expect of the
> kernel? As far as I can tell there is not a practical difference
> between a read-only MAP_PRIVATE and a read-only MAP_SHARED.

Some applications use mprotect to change page protections behind glibc's
back. Using MAP_SHARED would break fork pretty badly.

Most of the hard-to-diagnose crashes seem to come from global data or
relocations because they are wiped by truncation. And we certainly
can't use MAP_SHARED for those. Code often seems to come back unchanged
after the truncation because the overwritten file hasn't actually
changed. File attributes don't help because the copying is an
adminstrative action in the context of the application (maybe the result
of some automation).

I think avoiding the crashes isn't the right approach. What I'd like to
see is better diagnostics. Writing mtime and ctime to the core file
might help. Or adding a flag to the core file and /proc/PID/smaps that
indicates if the file has been truncated across the mapping since the
mapping was created.

A bit less conservative and even more obvious to diagnose would be a new
flag for the mapping (perhaps set via madvise) that causes any future
access to the mapping to fault with SIGBUS and a special si_code value
after the file has been truncated across the mapping. I think we would
set that in the glibc dynamic loader. It would make the crashes much
less weird.