Re: [PATCH] cowlinks v2

[Jamie Lokier <jamie@shareable.org>]

Pavel Machek wrote:
> > > > Here's a tricky situation:
> > > > 
> > > >    1. A file is cowlinked.  Then each cowlink is mmap()'d, one per process.
> > > > 
> > > >    2. At this point both mappings share the same pages in RAM.
> > > > 
> > > >    3. Then one of the cowlinks is written to...
> > > 
> > > Using mmap with PROT_WRITE on a cowlink must preemptively
> > > break the link.
> > 
> > I forget to mention, they are PROT_READ shared mappings.
> 
> I'm not mm guru, but... with rmap, we should be able to find all the
> users of that shared memory, and unmap their pages, right?

Yes.  I bring it up only because it's tricky, and the simple cowlink
implementations so far don't deal with it.

A page can only exist in one address_space.  So if pages are shared
before the cow is broken, the address_space must be of the shared
cowid object, not an individual address_space per cowlink.
Afterwards, the copied pages are in the non-shared cowlink object's
address_space.

> Until copy is done, we don't do anything, because write is not allowed
> to progress until copy is done. When copy is done we should unmap all
> the pages that still point to "old" copy, let write progress, and make
> users fault in.

I agree.

(Ross suggested using COW pages.  While technically possible, that
would be pretty complicated to implemented as it implies pages shared
among more than one address_space, and the facility for write() to
break COW sharing in the page cache, and update page tables when that
happens.)

-- Jamie