Re: [Qemu-devel] [PATCH v3.1 00/31] NUMA series, and hostmem improvements

[Hu Tao]

On Fri, May 09, 2014 at 11:13:50PM +0200, Paolo Bonzini wrote:
> Il 09/05/2014 19:54, Eduardo Habkost ha scritto:
> >On Fri, May 09, 2014 at 04:29:49PM +0800, Hu Tao wrote:
> >>On Thu, May 08, 2014 at 04:51:56PM +0200, Paolo Bonzini wrote:
> >>>Il 06/05/2014 11:27, Hu Tao ha scritto:
> >>>>This series includes work on QOMifying the memory backends.
> >>>>the idea is to delegate all properties of the memory backend to
> >>>>a new QOM class hierarchy, in which the concrete classes
> >>>>are hostmem-ram and hostmem-file.  The backend is passed to the
> >>>>machine via "-numa node,memdev=foo" where "foo" is the id of the
> >>>>backend object.
> >>>
> >>>Hello,
> >>>
> >>>I noticed now that if you have the host-nodes property set Linux
> >>>requires you to set a policy other than "default" too.  If you don't,
> >>>the mbind system call fails.
> >>>
> >>>What about squashing something like this?
> >>>
> >>>Paolo
> >>>
> >>>diff --git a/backends/hostmem.c b/backends/hostmem.c
> >>>index d3f8476..a0a3111 100644
> >>>--- a/backends/hostmem.c
> >>>+++ b/backends/hostmem.c
> >>>@@ -299,12 +299,23 @@ host_memory_backend_memory_init(UserCreatable *uc, 
> >>>Error **errp)
> >>>
> >>> #ifdef CONFIG_NUMA
> >>>     unsigned long maxnode = find_last_bit(backend->host_nodes, MAX_NODES);
> >>>+    unsigned policy = backend->policy;
> >>>+
> >>>+    /* Linux does not accept MPOL_DEFAULT with nonzero bitmap, but
> >>>+     * "-object memory-ram,size=128M,hostnodes=0,policy=bind" is a
> >>>+     * bit of a mouthful.  So if the host_nodes bitmap is nonzero,
> >>>+     * pick the BIND policy.
> >
> >Are we sure MPOL_BIND is a better default than MPOL_INTERLEAVE or
> >MPOL_PREFERRED?
> 
> Better than MPOL_PREFERRED, yes.  Better than MPOL_INTERLEAVE, I am
> not sure.
> 
> > * If policy=default is set, it is always going to be MPOL_DEFAULT.
> > * If policy=bind is set, it is always going to be MPOL_BIND.
> > * if policy=preferred is set, it is always going to be MPOL_PREFERRED.
> > * If policy is omitted, it will be be MPOL_DEFAULT is host-nodes is
> >   unset, and MPOL_BIND if host_nodes is set.
> 
> That's possible.  Or we can just detect the case in
> host_memory_backend_memory_init and provide a better error message
> than just "Invalid argument" (aka EINVAL).  We're always free to
> take an error situation and make it non-error in the future, but the
> reverse would be hard to do.

This one looks better. If we decide to turn implict MPOL_BIND into
error in the future, it may cause a long-run guest with policy ommitted
fail to start, which sounds like a compatibility problem.

Regards,
Hu Tao