[Pkg-gauche-devel] threads and fork on machine with VIPT-WB cache

[James Bottomley]

On Fri, 2010-04-02 at 12:48 +0900, NIIBE Yutaka wrote:
> Thanks for your quick reply.
> 
> James Bottomley wrote:
> > In COW breaking, the page table entry is copied, so A and B no longer
> > have page table entries at the same physical location.  If the COW is
> > intact, A and B have the same physical page, but it's also accessed by
> > the same virtual address, hence no aliasing.
> 
> Let me explain more.
> 
> In the scenario, I assume:
> 
> 	No aliasing between A and B.
> 	We have aliasing between kernel access and user access.
> 
> Before COW breaking A and B share same data (with no aliasing same
> space identifier color), and B sees data in cache, while memory has
> stale data.
> 
> At COW breaking, kernel copies the memory, it doesn't see new data
> in cache because of aliasing.
> 
> Isn't it possible?

So your theory is that the data the kernel sees doing the page copy can
be stale because of dirty cache lines in userspace (which is certainly
possible in the ordinary way)?  By design that shouldn't happen: the
idea behind COW breaking is that before it breaks, the page is read
only ... this means that processes can have clean cache copies of it,
but never dirty cache copies (because writes are forbidden).  As soon as
one or other process tries to write to the page, it gets a memory
protection trap long before the data it's trying to write goes into the
cache.  By the time the write is allowed to complete (and the cache
becomes dirty), the process will have the new copy of the page which
belongs exclusively to it.

James