Re: [PATCH v6 2/4] Add NVMM accelerator: x86 CPU support

[Paolo Bonzini]

On 01/04/21 14:32, Reinoud Zandijk wrote:
> On Thu, Apr 01, 2021 at 10:35:40AM +0200, Paolo Bonzini wrote:
> > On 31/03/21 22:07, Reinoud Zandijk wrote:
> > > +void nvmm_vcpu_kick(CPUState *cpu);
> >
> > Not defined anywhere.
>
> Hmmm, indeed. I think its a leftover of the former patch. Good catch.
>
> > > +{
> > > +#if NVMM_USER_VERSION == 1
> > > +    struct sigaction sigact;
> > > +    sigset_t set;
> > > +
> > > +    /* Install the IPI handler. */
> > > +    memset(&sigact, 0, sizeof(sigact));
> > > +    sigact.sa_handler = nvmm_ipi_signal;
> > > +    sigaction(SIG_IPI, &sigact, NULL);
> > > +
> > > +    /* Allow IPIs on the current thread. */
> > > +    sigprocmask(SIG_BLOCK, NULL, &set);
> > > +    sigdelset(&set, SIG_IPI);
> > > +    pthread_sigmask(SIG_SETMASK, &set, NULL);
> > > +#else
> > > +    /*
> > > +     * We use the nvmm_vcpu_stop() mechanism, and don't use signals.
> > > +     * Nothing to do.
> > > +     */
> > > +#endif
> >
> > Since nvmm_vcpu_stop is very similar to KVM's immediate_exit mechanism, I
> > think you still need to have a dummy signal handler to kick the VM out of
> > the run loop *if it is in the kernel*.  The signal handler however can just
> > do nothing.
>
> Are you worried the in-kernel thread will somehow get stuck or halt on exit of
> Qemu and left as a zombie?

No, you need all three of these:

- the signal to interrupt a thread that is running the VM

- the exit_request field to interrupt a thread that is running QEMU code

- nvmm_vcpu_stop() to interrupt a thread that is running kernel code but 
has not yet started running the VM.

> > Also, can you just drop support for NVMM_USER_VERSION == 1?
>
> Now thats a good suggestion. We could add support for it in the pkgsrc
> package. When 9.0 gets retired, we could then retire it there without the need
> to patch Qemu again.

If it's still in use in the wild I have no problem keeping it.  The 
difference is small; I was just asking.

> > > diff --git a/target/i386/nvmm/meson.build b/target/i386/nvmm/meson.build
> > > new file mode 100644
> > > index 0000000000..c154e78014
> > > --- /dev/null
> > > +++ b/target/i386/nvmm/meson.build
> > > @@ -0,0 +1,4 @@
> > > +i386_softmmu_ss.add(when: 'CONFIG_NVMM', if_true: files(
> > > +  'nvmm-all.c',
> > > +  'nvmm-accel-ops.c',
> > > +))
> >
> > The nvmm library should be added here.
>
> I am not sure what you mean by that. You provided a patch for the meson.build
> file, will that not suffice?

That patch dropped nvmm from meson.build, here is where it should be 
added (so that it is conditional on CONFIG_NVMM and qemu-system-arm does 
not link to libnvmm).

Paolo