Re: [Qemu-devel] [PATCH v2] accel: forbid early use of kvm_enabled() and friends

[Eduardo Habkost <ehabkost@redhat.com>]

On Fri, Jun 29, 2018 at 11:39:10AM +0100, Daniel P. Berrangé wrote:
> On Fri, Jun 29, 2018 at 12:35:09PM +0200, Paolo Bonzini wrote:
> > On 29/06/2018 12:29, Greg Kurz wrote:
> > > It is unsafe to rely on *_enabled() helpers before the accelerator has
> > > been initialized, ie, accel_init_machine() has succeeded, because they
> > > always return false. But it is still possible to end up calling them
> > > indirectly by inadvertance, and cause QEMU to misbehave.
> > > 
> > > This patch causes QEMU to abort if we try to check for an accelerator
> > > before it has been set up. This will help to catch bugs earlier.
> > > 
> > > Signed-off-by: Greg Kurz <groug@kaod.org>
> > > Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
> > > ---
> > > 
> > > This patch was motivated by an regression we're currently fixing in
> > > spapr because of an early use of kvm_enabled(). David suggested to
> > > post this patch separately:
> > > 
> > > https://lists.nongnu.org/archive/html/qemu-ppc/2018-06/msg01136.html
> > > 
> > > v2: - dropped change in qom/cpu.c (useless header inclusion)
> > >     - only #include "sysemu/kvm.h" if we actually need it
> > >     - added David's R-b from v1 because changes in v2 are minor
> > 
> > This adds a function call on possibly hot paths.  Can you make it inline?
> > 
> > Also asserting current_machine != NULL is not necessary, since you're
> > immediately dereferencing it.
> 
> Is there a practical way to simply initialize the accelerators earlier
> in startup sequence, so we just remove or at least reduce, the liklihood
> of accessing it too early ?

We can reduce it, but not eliminate it: it would still be
possible to use the *_enabled() macros too early, before we parse
all command-line options (on QOM class_init, for example).

-- 
Eduardo