Linux Virtualization - Sep 2015 - [PATCH v7] pci: quirk to skip msi disable on shutdown

On Mon, Sep 21, 2015 at 05:10:43PM -0500, Bjorn Helgaas
wrote:
> On Mon, Sep 21, 2015 at 10:42:13PM +0300, Michael S. Tsirkin wrote:
> > On Mon, Sep 21, 2015 at 01:21:47PM -0500, Bjorn Helgaas wrote:
> > > On Sun, Sep 06, 2015 at 06:32:35PM +0300, Michael S. Tsirkin
wrote:
> > > > On some hypervisors, virtio devices tend to generate
spurious interrupts
> > > > when switching between MSI and non-MSI mode.  Normally,
either MSI or
> > > > non-MSI is used and all is well, but during shutdown, linux
disables MSI
> > > > which then causes an "irq %d: nobody cared"
message, with irq being
> > > > subsequently disabled.
> > > 
> > > My understanding is:
> > > 
> > >   Linux disables MSI/MSI-X during device shutdown.  If the device
> > >   signals an interrupt after that, it may use INTx.
> > > 
> > > This INTx interrupt is not necessarily spurious.  Using INTx to
signal an
> > > interrupt that occurs when MSI is disabled seems like reasonable
behavior
> > > for any PCI device.
> > > And it doesn't seem related to switching between MSI and
non-MSI mode.
> > > Yes, the INTx happens *after* disabling MSI, but it is not at all
> > > *because* we disabled MSI.  So I wouldn't say "they
generate spurious
> > > interrupts when switching between MSI and non-MSI."
> > > 
> > > Why doesn't virtio-pci just register an INTx handler in
addition to an MSI
> > > handler?
> > 
> > The handler causes an expensive exit to the hypervisor,
> > and the INTx lines are shared with other devices.
> 
> Do we care?  Is this a performance path?  I thought we were in a kexec
> shutdown path.
Yes but the handler would always have to be registered, right?
> > Seems silly to slow them down just so we can do something
> > that triggers the device bug.  The bus master is disabled by that
time,
> > if linux can just desist from touching MSI enable device won't
> > send either INTx (because MSI is on) or MSI
> > (because bus master is on) and all will be well.
> 
> It would also be silly to put special-purpose code in the PCI core
> if there's a reasonable way to handle this in a driver.
> 
> Can you describe exactly what the device bug is?  Apparently you're
> saying that if we shut down MSI, it triggers the bug?  And I guess
> you're talking about a virtio device as implemented in qemu or other
> hypervisors?
Yes. Basically depending on an internal device state, disabling MSI
sometimes wedges it.  The most easy to debug effect is if it starts
sending INTx interrupts, for which there's no handler currently.
Full system reset always gets us out of the bad state.

> If we leave MSI enabled (as your patch does), then the device has MSI
> enabled and Bus Master disabled.  I can see these possibilities:
> 
>   1) the device never recognizes an interrupt condition
> ? 2) the device sets the pending bit but doesn't issue the MSI write,
>      so the OS doesn't see the interrupt unless it polls for it
> ? 3) the device signals MSI and we still have?an MSI handler
>      registered, so we silently handle it
>   4) the device signals INTx
> You seem to suggest that if we leave MSI enabled (as your patch does),
> we're in case 1.  But I doubt that disabling MSI causes the device to
> interrupt.
> 
> Case 2 seems more likely to me: the device recognized an interrupt
> condition, e.g., an event occurred, and the OS simply doesn't see the
> interrupt because the device can't issue the MSI message.
> 
> Case 3 does seem like it would be a device bug, because the device
> shouldn't do an MSI write when Bus Master is disabled.  I don't see
> this case mentioned explicitly in the PCI spec, but PCIe r3.0 spec sec
> 7.5.1.1 does make it clear that disabling Bus Master also disables MSI
> messages.
> 
> I don't know whether case 4 would be legal or not.  But apparently it
> doesn't happen with the virtio device anyway, so it's not really a
> concern here.
> 
> Bjorn
It's case 2. Except it doesn't actually set the pending
bit, because PCI spec has language like

	"While a vector is masked, the function is prohibited from sending the
	associated message, and the function must set the associated Pending bit"

but doesn't actually require to set pending if bus master enable
prevents sending MSI.

-- 
MST

On Tue, Sep 22, 2015 at 02:29:03PM +0300, Michael S. Tsirkin
wrote:
> On Mon, Sep 21, 2015 at 05:10:43PM -0500, Bjorn Helgaas wrote:
> > On Mon, Sep 21, 2015 at 10:42:13PM +0300, Michael S. Tsirkin wrote:
> > > On Mon, Sep 21, 2015 at 01:21:47PM -0500, Bjorn Helgaas wrote:
> > > > On Sun, Sep 06, 2015 at 06:32:35PM +0300, Michael S. Tsirkin
wrote:
> > > > > On some hypervisors, virtio devices tend to generate
spurious interrupts
> > > > > when switching between MSI and non-MSI mode.  Normally,
either MSI or
> > > > > non-MSI is used and all is well, but during shutdown,
linux disables MSI
> > > > > which then causes an "irq %d: nobody cared"
message, with irq being
> > > > > subsequently disabled.
> > > > 
> > > > My understanding is:
> > > > 
> > > >   Linux disables MSI/MSI-X during device shutdown.  If the
device
> > > >   signals an interrupt after that, it may use INTx.
> > > > 
> > > > This INTx interrupt is not necessarily spurious.  Using INTx
to signal an
> > > > interrupt that occurs when MSI is disabled seems like
reasonable behavior
> > > > for any PCI device.
> > > > And it doesn't seem related to switching between MSI and
non-MSI mode.
> > > > Yes, the INTx happens *after* disabling MSI, but it is not
at all
> > > > *because* we disabled MSI.  So I wouldn't say "they
generate spurious
> > > > interrupts when switching between MSI and non-MSI."
> > > > 
> > > > Why doesn't virtio-pci just register an INTx handler in
addition to an MSI
> > > > handler?
> > > 
> > > The handler causes an expensive exit to the hypervisor,
> > > and the INTx lines are shared with other devices.
> > 
> > Do we care?  Is this a performance path?  I thought we were in a kexec
> > shutdown path.
> 
> Yes but the handler would always have to be registered, right?
The pci_device_shutdown() path you're modifying calls drv->shutdown()
immediately before disabling MSI, so I suppose you could register a
handler in a virtio shutdown method.
> > > Seems silly to slow them down just so we can do something
> > > that triggers the device bug.  The bus master is disabled by that
time,
> > > if linux can just desist from touching MSI enable device
won't
> > > send either INTx (because MSI is on) or MSI
> > > (because bus master is on) and all will be well.
> > 
> > It would also be silly to put special-purpose code in the PCI core
> > if there's a reasonable way to handle this in a driver.
> > 
> > Can you describe exactly what the device bug is?  Apparently
you're
> > saying that if we shut down MSI, it triggers the bug?  And I guess
> > you're talking about a virtio device as implemented in qemu or
other
> > hypervisors?
> 
> Yes. Basically depending on an internal device state, disabling MSI
> sometimes wedges it.  The most easy to debug effect is if it starts
> sending INTx interrupts, for which there's no handler currently.
> Full system reset always gets us out of the bad state.
If disabling MSI causes the device to use INTx interrupts, that sounds
perfectly normal to me.

If disabling MSI causes the device to hang, *that* sounds like a bug.
Since this is virtio, we should be able to figure out exactly where
that happens.  Do you have a pointer to a virtio bug report, or even a
QEMU commit that fixes this virtio bug?

I understand that even if there is a virtio fix in QEMU, we want a
solution that works even with an old QEMU that doesn't contain the
fix.  But a pointer to a QEMU fix would really help understand and
document the Linux fix.

Bjorn

On Tue, Sep 22, 2015 at 07:36:40AM -0500, Bjorn Helgaas
wrote:
> On Tue, Sep 22, 2015 at 02:29:03PM +0300, Michael S. Tsirkin wrote:
> > On Mon, Sep 21, 2015 at 05:10:43PM -0500, Bjorn Helgaas wrote:
> > > On Mon, Sep 21, 2015 at 10:42:13PM +0300, Michael S. Tsirkin
wrote:
> > > > On Mon, Sep 21, 2015 at 01:21:47PM -0500, Bjorn Helgaas
wrote:
> > > > > On Sun, Sep 06, 2015 at 06:32:35PM +0300, Michael S.
Tsirkin wrote:
> > > > > > On some hypervisors, virtio devices tend to
generate spurious interrupts
> > > > > > when switching between MSI and non-MSI mode. 
Normally, either MSI or
> > > > > > non-MSI is used and all is well, but during
shutdown, linux disables MSI
> > > > > > which then causes an "irq %d: nobody
cared" message, with irq being
> > > > > > subsequently disabled.
> > > > > 
> > > > > My understanding is:
> > > > > 
> > > > >   Linux disables MSI/MSI-X during device shutdown.  If
the device
> > > > >   signals an interrupt after that, it may use INTx.
> > > > > 
> > > > > This INTx interrupt is not necessarily spurious.  Using
INTx to signal an
> > > > > interrupt that occurs when MSI is disabled seems like
reasonable behavior
> > > > > for any PCI device.
> > > > > And it doesn't seem related to switching between
MSI and non-MSI mode.
> > > > > Yes, the INTx happens *after* disabling MSI, but it is
not at all
> > > > > *because* we disabled MSI.  So I wouldn't say
"they generate spurious
> > > > > interrupts when switching between MSI and
non-MSI."
> > > > > 
> > > > > Why doesn't virtio-pci just register an INTx
handler in addition to an MSI
> > > > > handler?
> > > > 
> > > > The handler causes an expensive exit to the hypervisor,
> > > > and the INTx lines are shared with other devices.
> > > 
> > > Do we care?  Is this a performance path?  I thought we were in a
kexec
> > > shutdown path.
> > 
> > Yes but the handler would always have to be registered, right?
> 
> The pci_device_shutdown() path you're modifying calls
drv->shutdown()
> immediately before disabling MSI, so I suppose you could register a
> handler in a virtio shutdown method.
I guess we could. Not sure what we'd do e.g. if that fails.
> > > > Seems silly to slow them down just so we can do something
> > > > that triggers the device bug.  The bus master is disabled by
that time,
> > > > if linux can just desist from touching MSI enable device
won't
> > > > send either INTx (because MSI is on) or MSI
> > > > (because bus master is on) and all will be well.
> > > 
> > > It would also be silly to put special-purpose code in the PCI
core
> > > if there's a reasonable way to handle this in a driver.
> > > 
> > > Can you describe exactly what the device bug is?  Apparently
you're
> > > saying that if we shut down MSI, it triggers the bug?  And I
guess
> > > you're talking about a virtio device as implemented in qemu
or other
> > > hypervisors?
> > 
> > Yes. Basically depending on an internal device state, disabling MSI
> > sometimes wedges it.  The most easy to debug effect is if it starts
> > sending INTx interrupts, for which there's no handler currently.
> > Full system reset always gets us out of the bad state.
> 
> If disabling MSI causes the device to use INTx interrupts, that sounds
> perfectly normal to me.
> 
> If disabling MSI causes the device to hang, *that* sounds like a bug.
> Since this is virtio, we should be able to figure out exactly where
> that happens.  Do you have a pointer to a virtio bug report, or even a
> QEMU commit that fixes this virtio bug?
> 
> I understand that even if there is a virtio fix in QEMU, we want a
> solution that works even with an old QEMU that doesn't contain the
> fix.  But a pointer to a QEMU fix would really help understand and
> document the Linux fix.
> 
> Bjorn
I'm not sure we ever understood it completely.

I think some of it has to do with the way the whole virtio 0
device register layout changes when you enable/disable MSI.  So should
be ok when using the modern virtio 1 model since we fixed this thing.

I was hoping that since disabling MSI in pci core is only useful as a
work-around (for devices with a broken bus master enable - even though I
don't think we know what these are exactly), a flag for not disabling it
won't be held to such a high standard.

-- 
MST