[openib-general] Re: [PATCH repost] PROT_DONTCOPY: ifiniband uverbs fork support

[Hugh Dickins]

On Wed, 10 Aug 2005, Gleb Natapov wrote:
> On Tue, Aug 09, 2005 at 07:13:33PM +0100, Hugh Dickins wrote:
> > Even more I'd prefer one of these two solutions below, which sidestep
> > that uncleanliness - but both of these would be in mmap only, no clean
> > way to change afterwards (except by munmap or mmap MAP_FIXED):
> > 
> > 1.  Use the standard mmap(NULL, len, PROT_READ|PROT_WRITE,
> >     MAP_SHARED|MAP_ANONYMOUS, -1, 0) which gives you a memory object
> >     shared with children, so write-protection and COW won't come into it.
> > 
> > or if there's good reason why that's no good,
> > 
> > 2.  Define a MAP_DONTCOPY to mmap: we have a fine tradition of MAP_flags
> >     to achieve this or that effect, adding one more would be cleaner than
> >     now corrupting mprotect or madvise.
> 
> They are both relying on the way user allocates memory for RDMA. The idea
> behind Michael's propose it to let library (MPI for instance) to tell to the
> kernel that the pages are used for RDMA and it is not safe to copy them now. 
> The pages may be anywhere in the process address space bss, text, stack
> whatever.

That's a nice aim, but I don't think it can quite be done in the face of
the fork issue - one way or another, we have to change the behaviour of a
forked RDMA area slightly, which might interfere with common assumptions.

Your stack example is a good one: if we end up setting VM_DONTCOPY on
the user stack, then I don't think fork's child will get very far without
hitting a SIGSEGV.

Hugh