Re: [PATCH] rdma: don't make pages writeable if not requiested

From: Jason Gunthorpe
Date: Thu Mar 21 2013 - 13:11:46 EST


On Thu, Mar 21, 2013 at 11:39:47AM +0200, Michael S. Tsirkin wrote:
> On Thu, Mar 21, 2013 at 02:13:38AM -0700, Roland Dreier wrote:
> > On Thu, Mar 21, 2013 at 1:51 AM, Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > >> In that case, no, I don't see any reason for LOCAL_WRITE, since the
> > >> only RDMA operations that will access this memory are remote reads.
> > >
> > > What is the meaning of LOCAL_WRITE then? There are no local
> > > RDMA writes as far as I can see.
> >
> > Umm, it means you're giving the local adapter permission to write to
> > that memory. So you can use it as a receive buffer or as the target
> > for remote data from an RDMA read operation.
>
> Well RDMA read has it's own flag, IB_ACCESS_REMOTE_READ.
> I don't see why do you need to give adapter permission

The access flags have to do with what independent access remote nodes
get. There are four major cases:

access = IBV_ACCESS_REMOTE_READ says the adaptor will let remote nodes
read the memory.

access = 0 (ie IBV_ACCESS_LOCAL_READ) says that only the adaptor, under
the direct control of the application, can read this memory. Remote
nodes are barred.

access = IBV_ACCESS_REMOTE_WRITE|IBV_ACCESS_LOCAL_WRITE says the adaptor
will let remote nodes write the memory

access = IBV_ACCESS_LOCAL_WRITE bars remote nodes from writing to that
memory. Only the adaptor, under the direct control of the application,
can write the memory.

The fact LOCAL_READ/REMOTE_READ exists makes it possible to do what
you want - it guarentees the adaptor will never write to this memory
under any circumstances, so you can leave the page COW'd. If
LOCAL_WRITE was implied then you'd have to COW everything..

Would it be better to drive the COW break decision off the region's MM
flags? Ie if the memory is mapped read only into the process then you
can keep the COW at the RDMA layer, otherwise you should break
it. That seems more natural than a new flag?

Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/