Re: [PATCH v4] KVM: Resize kvm_io_range array dynamically

[Alex Williamson <alex.williamson@redhat.com>]

On Thu, 2012-03-01 at 12:14 +0200, Sasha Levin wrote:
> On Thu, Mar 1, 2012 at 9:01 AM, Amos Kong <akong@redhat.com> wrote:
> > This patch makes the kvm_io_range array can be resized dynamically.
> > Set an upper bounds for kvm_io_range to limit userspace.
> > 1000 is a very large limit and not bloat the typical user.
> 
> Whats the reason for making everything dynamic? Memory savings there
> aren't that significant.

We're currently looking at about 14k for these arrays with 300 entries
since we have two of them.  If we're going to double it or triple it to
handle the maximum use case, why impose that on the typical VM?  It may
not be multiple megabytes, but I wouldn't say it's insignificant either.

> 
> If you want to make it more efficient just define:
> static struct kvm_io_bus io_bus[2];
> 
> somewhere in kvm_main.c, and just switch between them when you need to
> do insertion and removal of devices. You get the benefit of zero slub
> usage, no allocations in any of the paths, and much simpler logic.

It's updated via rcu.  The change Amos is proposing is fairly trivial;
allocate the necessary size and memcpy instead of memdup.  Maybe it can
be optimized further, but this seems like a step in the right direction
to handle worst case use and, if anything, benefit the typical user too.
Thanks,

Alex