Re: how to force C++ objects to an address? (ABI issue B-7)

[Ian Lance Taylor <ian at wasabisystems dot com>]

Anthony Heading <aheading@jpmorgan.com> writes:

> Clutching at straws, if a .so file was already mmap'd into a processes
> address space, is there any chance that a subsequent dlopen would
> decide to use that existing copy?  (Hmm - that sounds a bit unlikely)

That sounds pretty tough to me.  The dynamic linker has to apply the
dynamic relocations when dlopen is called on a shared library.  If a
the library is already mapped into memory, dlopen would have to start
modifying it.  That would seem to violate natural expectations of how
mmap should work.

> Alternatively, if one were to think about changing Linux to support
> this - e.g. creating a variant on dlopen that took an address, or
> maybe better using the (unused?) physical address field in the
> program headers to indicate where a library should be loaded, or
> perhaps less intrusively an extra section to supply that information,
> is it realistic to hope this might be acceptable for the wider world?

It seems like a reasonable extension to me.  I think the natural
approach would be a new version of dlopen.  Using the physical address
field doesn't sound right to me.  Adding a new program header type to
carry the required address ought to work, and would be useful since it
wouldn't require dlopen.  The place to change would be elf/dl-load.c
in glibc.

Ian