RE: [PATCH] PCI: hv: Fix a __local_bh_enable_ip warning in hv_compose_msi_msg()

From: Dexuan Cui
Date: Sat Jun 30 2018 - 14:14:05 EST


> From: Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx>
> Sent: Friday, June 29, 2018 02:39
> To: Dexuan Cui <decui@xxxxxxxxxxxxx>
> On Wed, Jun 13, 2018 at 10:50:05PM +0000, Dexuan Cui wrote:
> > > From: Bjorn Helgaas <helgaas@xxxxxxxxxx>
> > > Sent: Wednesday, June 13, 2018 15:15
> > > > ...
> > > > It looks Lorenzo's pci.git tree has not been updated for 3+ weeks.
> > > > I guess Lorenzo may be on vacation.
> > > >
> > > > @Bjorn, can this patch go through your tree?
> > > > Should I resubmit it?
> > >
> > > No need to resubmit it, Lorenzo has been out for a bit, but I'm sure
> > > he'll pick this up as he catches up.
> > OK, I see. Thanks!
> >
> > > You might, however, fix the commit log:
> > >
> > > This is not an issue because hv_pci_onchannelcallback() is not slow,
> > > and it not a hot path.
> > >
> > > This has at least one typo (I think you mean "and *is* not a hot
> > > path").
> > Sorry -- yes, it's a typo. I hope Lorenzo can help to fix this, or I can
> > resubmit it if Lorenzo or you want me to do it.
> >
> > > I also don't understand the sentence as a whole because the
> > > hv_pci_onchannelcallback() comment says it's called whenever the host
> > > sends a packet to this channel, and that *does* sound like a hot path.
> > Sorry for not making it clear.
> > The host only sends a packet into the channel of the guest when there
> > is a change of device configuration (i.e. hot add or remove a device), or
> > the host is responding to the guest's request.
> >
> > The change of device configuration is only triggered on-demand by the
> > administrator on the host, and the guest's requests are one-off when
> > the device is probed.
> >
> > So IMO the callback is not a hot path.
> >
> > > I also don't understand the "hv_pci_onchannelcallback() is not slow"
> > > part. In other words, you're saying hv_pci_onchannelcallback() is
> > > fast and it's not a hot path. And apparently this has something to do
> > > with the difference between local_bh_disable() and local_irq_save()?
> > >
> > > Bjorn
> > Actually in my original internal version of the patch, I did use
> > local_irq_save/restore().
> >
> > hv_pci_onchannelcallback() itself runs fast, but here since it's in a
> > loop (i.e. the while (!try_wait_for_completion(&comp.comp_pkt.host_event)
> > loop), IIRC I was asked if I really need local_irq_save/restore(),
> > and I answered "not really", so later I switched to
> local_bh_disable()/enable().
> >
> > However, recently I found that if we enable CONFIG_PROVE_LOCKING=y,
> > the local_bh_enable() can trigger a warning because the function
> > hv_compose_msi_msg() can be called with local IRQs disabled (BTW,
> > hv_compose_msi_msg() can also be called with local IRQS enabled in
> > another code path):
> >
> > IRQs not enabled as expected
> > WARNING: CPU: 0 PID: 408 at kernel/softirq.c:162 __local_bh_enable_ip
> >
> > Despite the warning, the code itself can still work correctly, but IMO we'd
> > better switch back to local_irq_save/restore(), and hence I made the patch.
> >
> > I hope the explanation sounds reasonable. :-)
>
> Sorry for the delay in replying. I need to understand if you are
> preventing a spurious lockdep warning or you are fixing a kernel
> bug. From your commit log, I assume the former option but I do
> not think that's what you are really doing.

Now my understanding is:
1) When hv_compose_msi_msg() is called with local irq ENABLED by the upper
level irq code, the current code is good and the lockdep warning is not triggered.

2) When hv_compose_msi_msg() is called with local irq DISABLED by the upper
level irq code, the current code *is* buggg: local_bh_enable() can potentially call
do_softirq(), which is not supposed to run when local irq is DISABLED.
I think the lockdep warning is triggered for this reason.

In summary, now I realized the warning is not spurious, and here at the first
place I should not use local_bh_disable()/enable(), which are not supposed to
be used when local irq can be DISABLED.

> Apart from the commit log typos fixes I would like a log that
> explains *why* this is not a kernel bug fix rather than a harmless
> lockdep warning prevention.
>
> Lorenzo

Now I realized there *is* a bug.
I'm going to send a v2 with a new changelog, though the changed code
will remain the same.

Thanks,
-- Dexuan