[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [Xen-devel] [PATCH v2 3/4] VMX: Assign the right value to 'NDST' field in a concern case
> -----Original Message----- > From: dunlapg@xxxxxxxxx [mailto:dunlapg@xxxxxxxxx] On Behalf Of George > Dunlap > Sent: Thursday, June 23, 2016 2:01 AM > To: Wu, Feng <feng.wu@xxxxxxxxx> > Cc: xen-devel@xxxxxxxxxxxxx; Tian, Kevin <kevin.tian@xxxxxxxxx>; Keir Fraser > <keir@xxxxxxx>; Andrew Cooper <andrew.cooper3@xxxxxxxxxx>; Dario Faggioli > <dario.faggioli@xxxxxxxxxx>; Jan Beulich <jbeulich@xxxxxxxx> > Subject: Re: [Xen-devel] [PATCH v2 3/4] VMX: Assign the right value to 'NDST' > field in a concern case > > On Thu, May 26, 2016 at 2:39 PM, Feng Wu <feng.wu@xxxxxxxxx> wrote: > > Normally, in vmx_cpu_block() 'NDST' filed should have the same > > value with 'dest' or 'MASK_INSR(dest, PI_xAPIC_NDST_MASK)' depending > > on whether x2apic is enabled. However, in the following scenario, > > 'NDST' has different value: > > > > 'vcpu_block' hook gets assigned in vmx_pi_hooks_assign(), but all > > other three PI hooks have not been assigned or not been excuted yet. > > And during this interval, we are running in vmx_vcpu_block(), then > > 'NDST' may have different value. > > > > This patch fix this concern case. > > > > Signed-off-by: Feng Wu <feng.wu@xxxxxxxxx> > > I agree with Jan that a cleaner solution here would be making sure > that all the appropriate state is actually set up for all vcpus before > leaving vmx_pi_hooks_assign(). With the patch you propose, the > following sequence of events is possible: > > * vcpu 0 starts running on a pcpu > * a device is assigned, causing the hooks to be set > * an interrupt from the device is routed to vcpu 0, but it is not > actually delivered properly, since ndst is not pointing to the right > processor. > > One option would be to pause all vcpus before setting the hooks and > then un-pause them; this would force all the vcpus to go through > vmx_pi_switch_to() before vmx_vcpu_block(). Another would be to grab > the scheduler lock for each pcpu and write the vcpu's ndst with the > appropriate value before setting the hooks. That sounds a great idea. Besides that, maybe we can also pause/unpause the domain before/after unsetting the hooks, then we don't need to care about the race condition when vmx_pi_hooks_deassign() and vmx_vcpu_block() get called at the same time. After unpause the domain, we can safely remove the vCPUs from the per-cpu blocking list if needed. Thanks, Feng > > -George _______________________________________________ Xen-devel mailing list Xen-devel@xxxxxxxxxxxxx http://lists.xen.org/xen-devel
|
Lists.xenproject.org is hosted with RackSpace, monitoring our |