Re: [Qemu-devel] [RFC] vhost: Can we change synchronize_rcu to call_rcu in vhost_set_memory() in vhost kernel module?

[Paolo Bonzini]

Il 12/05/2014 14:12, Michael S. Tsirkin ha scritto:
> On Mon, May 12, 2014 at 01:46:19PM +0200, Paolo Bonzini wrote:
> > Il 12/05/2014 13:07, Michael S. Tsirkin ha scritto:
> > > On Mon, May 12, 2014 at 12:25:35PM +0200, Paolo Bonzini wrote:
> > > > Il 12/05/2014 12:18, Michael S. Tsirkin ha scritto:
> > > > > On Mon, May 12, 2014 at 12:14:25PM +0200, Paolo Bonzini wrote:
> > > > > > Il 12/05/2014 12:08, Michael S. Tsirkin ha scritto:
> > > > > > > On Mon, May 12, 2014 at 11:57:32AM +0200, Paolo Bonzini wrote:
> > > > > > > > Perhaps we can check for cases where only the address is changing,
> > > > > > > > and poke at an existing struct kvm_kernel_irq_routing_entry without
> > > > > > > > doing any RCU synchronization?
> > > > > > >
> > > > > > > I suspect interrupts can get lost then: e.g. if address didn't match any
> > > > > > > cpus, now it matches some. No?
> > > > > >
> > > > > > Can you explain the problem more verbosely? :)
> > > > > >
> > > > > > Multiple writers would still be protected by the mutex, so you
> > > > > > cannot have an "in-place update" writer racing with a "copy the
> > > > > > array" writer.
> > > > >
> > > > > I am not sure really.
> > > > > I'm worried about reader vs writer.
> > > > > If reader sees a stale msi value msi will be sent to a wrong
> > > > > address.
> > > >
> > > > That shouldn't happen on any cache-coherent system, no?
> > > >
> > > > Or at least, it shouldn't become any worse than what can already
> > > > happen with RCU.
> > >
> > > Meaning guest must do some synchronization anyway?
> >
> > Yes, I think so.  The simplest would be to mask the interrupt around
> > MSI configuration changes.  Radim was looking at a similar bug.
>
> Was this with classic assignment or with vfio?

No, it was with QEMU emulated devices (AHCI).  Not this path, but it did 
involve MSI configuration changes while the interrupt was unmasked in 
the device.

> > He couldn't replicate on bare metal,
> > but I don't see why it shouldn't
> > be possible there too.
>
> I doubt linux does something tricky here, so
> I'm not talking about something linux guest does,
> I'm talking about an abstract guarantee of the
> APIC.

APIC or PCI?  The chipset is involved here, not the APICs (the APICs 
just see interrupt transactions).

> If baremetal causes a synchronisation point
> and we don't, at some point linux will use it and
> this will bite us.

Yes, and I agree.

It's complicated; on one hand I'm not sure we can entirely fix this, on 
the other hand it may be that the guest cannot have too high expectations.

The invalid scenario is an interrupt happening before the MSI 
reconfiguration and using the new value.  It's invalid because MSI 
reconfiguration is a non-posted write and must push previously posted 
writes.  But we cannot do that at all!  QEMU has the big lock that 
implicitly orders transactions, but this does not apply to 
high-performance users of irqfd (vhost, VFIO, even QEMU's own virtio-blk 
dataplane).  Even if we removed all RCU uses for the routing table, we 
don't have a "bottleneck" that orders transactions and removes this 
invalid scenario.

At the same time, it's obviously okay if an interrupt happens after the 
MSI reconfiguration and uses the old value.  That's a posted write 
passing an earlier non-posted write, which is allowed.  This is the main 
reason why I believe that masking/unmasking is required on bare metal. 
In addition, if the masking/unmasking is done with a posted write (MMIO 
BAR), the guest needs to do a read for synchronization before it unmasks 
the interrupt.

In any case, whether writes synchronize with RCU or bypass it doesn't 
change the picture.  In either case, writes are ordered against each 
other but not against reads.  RCU does nothing except preventing 
dangling pointer accesses.

Paolo