[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] [PATCH RFC v1 61/74] xen/pvshim: support vCPU hotplug
On Tue, Jan 09, 2018 at 03:16:38AM -0700, Jan Beulich wrote: > >>> On 04.01.18 at 14:06, <wei.liu2@xxxxxxxxxx> wrote: > > @@ -1303,22 +1320,20 @@ long do_vcpu_op(int cmd, unsigned int vcpuid, > > XEN_GUEST_HANDLE_PARAM(void) arg) > > > > break; > > > > - case VCPUOP_up: { > > - bool_t wake = 0; > > - domain_lock(d); > > - if ( !v->is_initialised ) > > - rc = -EINVAL; > > Shouldn't this check remain here? I realize this will complicate > locking (luckily the domain lock is a recursive one, so it shouldn't > be too bad), but I don't think pv_shim_cpu_up() can tolerate failing > because of vcpu_up() failing. > > I also think that the use of "long" for return types and values isn't > really warranted here, and there's also no visible to me reason to > special case CPU0 here. But for simplicity reasons I can see why > you've chosen that option; otoh the locking issue above that you'll > need to solve might be easier to deal with if you didn't switch CPUs > for hypercall processing (without dropping the use of > continue_hypercall_on_cpu()). Right, I'm not sure why bringing a CPU up is required to happen on CPU0, but that's what the current code in arch_do_sysctl does. I'm not sure I'm following the last part of your reply, if for CPU bringup there's no need to switch to CPU0, why would I want to keep the continue_hypercall_on_cpu for in that case? Thanks, Roger. _______________________________________________ Xen-devel mailing list Xen-devel@xxxxxxxxxxxxxxxxxxxx https://lists.xenproject.org/mailman/listinfo/xen-devel
|
Lists.xenproject.org is hosted with RackSpace, monitoring our |