Re: [Qemu-devel] [RFC PATCH] vl: only display available accelerators

[Eduardo Habkost]

On Wed, Nov 08, 2017 at 03:03:34PM -0300, Philippe Mathieu-Daudé wrote:
> On 11/08/2017 02:25 PM, Eduardo Habkost wrote:
> > On Wed, Nov 08, 2017 at 05:06:29PM +0000, Daniel P. Berrange wrote:
> >> On Wed, Nov 08, 2017 at 02:59:05PM -0200, Eduardo Habkost wrote:
> >>> On Wed, Nov 08, 2017 at 01:21:33PM -0300, Philippe Mathieu-Daudé wrote:
> >>>> On 11/08/2017 10:28 AM, Daniel P. Berrange wrote:
> >>>>> On Mon, Oct 30, 2017 at 09:20:29AM +0100, Eduardo Habkost wrote:
> >>>>>> On Mon, Oct 30, 2017 at 01:00:56AM -0300, Philippe Mathieu-Daudé wrote:
> >>>>>>> examples configuring with '--enable-kvm --disable-tcg'
> >>>>>>>
> >>>>>>> - before
> >>>>>>>
> >>>>>>>   $ qemu-system-x86_64 -accel help
> >>>>>>>   Possible accelerators: kvm, xen, hax, tcg
> >>>>>>>
> >>>>>>>   $ qemu-system-x86_64 -accel tcg
> >>>>>>>   qemu-system-x86_64: -machine accel=tcg: No accelerator found
> >>>>>>>
> >>>>>>>   # qemu-system-x86_64 -accel hax
> >>>>>>>   qemu-system-x86_64: -machine accel=hax: No accelerator found
> >>>>>>>
> >>>>>>>   # qemu-system-x86_64 -accel xen
> >>>>>>>   xencall: error: Could not obtain handle on privileged command 
> >>>>>>> interface: No such file or directory
> >>>>>>>   xen be core: xen be core: can't open xen interface
> >>>>>>>   can't open xen interface
> >>>>>>>   qemu-system-x86_64: failed to initialize Xen: Operation not 
> >>>>>>> permitted
> >>>>>>>
> >>>>>>> - after
> >>>>>>>
> >>>>>>>   $ qemu-system-x86_64 -accel help
> >>>>>>>   Possible accelerators: kvm
> >>>>>>>
> >>>>>>> Signed-off-by: Philippe Mathieu-Daudé <address@hidden>
> >>>>>>> ---
> >>>>>>> RFC because:
> >>>>>>>     - I don't think this is the nicest way, too much #ifdef'fery in 
> >>>>>>> main()
> >>>>>>
> >>>>>> I suggest using object_class_get_list(TYPE_ACCEL, false).
> >>>>>
> >>>>> And check the result of the available() method on the returned classes
> >>>>> too, to filter the results.
> >>>>
> >>>> Good idea! I'll use that.
> >>>
> >>> It looks like QTest is the only accelerator that implements
> >>> ->available(), and its return value is a build-time constant that
> >>> depends only on CONFIG_POSIX.
> >>>
> >>> I wonder why we don't simply avoid compiling the qtest class if
> >>> CONFIG_POSIX is unset, making the ->available() method
> >>> unnecessary.
> >>
> >> Yeah that does seem simpler, though I'm surprised that Xen does not
> >> implement the available method. Xen is an accel I'd expect to see
> >> compiled into an x86 build, but is only available if the host is
> >> actually booted under a Xen hypervisor.  Likewise shouldn't kvm
> >> only report itself as available if the /dev/kvm actually exists.
> >> But maybe that's not the kind of semantics code using available()
> >> expects ?
> > 
> > Currently the only caller of ->available() calls ->init_machine()
> > immediately afterwards, so for the current code it doesn't matter
> > if the check is inside ->available() or ->init_machine().
> > 
> > That said, I'm not sure we should look for /dev/kvm or check for
> > the Xen hypervisor when handling "-accel help".  I expect help
> > text to tell the user what the QEMU binary supports, not what the
> > current host supports.
> 
> Yes I prefer that too, I'll write something up such:
> 
> static bool kvm_available(void)
> {
>     return access("/dev/kvm", R_OK|W_OK) == 0;
> }
> 
> However I wonder, if an user is not in the kvm group but is in sudoers
> and run "qemu-system-aarch64 -accel help" he won't see KVM as
> available... (I tend to not use sudo when looking for -help output).

If we do that, I think we should include kvm in the list anyway,
but just add a column indicating that it doesn't seem to be
available on the host.

In either case, I suggest implementing it as a separate patch,
and just ignore ->available() in the first version.

> 
> same with:
> 
> static bool hax_available(void)
> {
>     return access("/dev/HAX", R_OK|W_OK) == 0;
> }
> 
> ok for:
> 
> static bool xen_available(void)
> {
>     return access("/sys/hypervisor/properties/features", F_OK) == 0;
> }
> 
> On POWER7 we have:
> 
> /* On ppc64, the test only works with kvm-hv, but not with kvm-pr */
> accel = access("/sys/module/kvm_hv", F_OK) ? "tcg" : "kvm:tcg";
> 
> Isn't it cleaner to register 2 accelerators, KVM-PR and KVM-HV and have
> KVM being an alias?

I believe "-accel kvm,kvm-type=HV" and "-accel kvm,kvm-type=PR"
would be cleaner ways to represent the two modes, instead of
having two different accelerator classes.  This is not supported
by the current "<accel1>:<accel2>" syntax, but I think we should
replace that syntax with something better anyway.

-- 
Eduardo