Re: [PATCH v2 1/4] pci: disable msi/msix at probe time

[Bjorn Helgaas <bhelgaas@google.com>]

On Mon, Mar 23, 2015 at 08:22:39PM +0100, Michael S. Tsirkin wrote:
> On Mon, Mar 23, 2015 at 01:50:06PM -0500, Bjorn Helgaas wrote:
> > Hi Michael,
> > 
> > On Thu, Mar 19, 2015 at 07:57:52PM +0100, Michael S. Tsirkin wrote:
> > > commit d52877c7b1afb8c37ebe17e2005040b79cb618b0
> > >     pci/irq: let pci_device_shutdown to call pci_msi_shutdown v2
> > > 
> > > attempted to address the problem of kexec getting
> > > started after linux enabled msi/msix for a device,
> > > and drivers being confused by msi being enabled,
> > > by disabling msi at shutdown.
> > > 
> > > But arguably, it's better to disable msi/msix when kexec
> > > starts - for example, kexec might run after a crash (kdump)
> > > and shutdown callbacks are not always invoked in that case.
> > > 
> > > Cc: Yinghai Lu <yhlu.kernel.send@gmail.com>
> > > Cc: Ulrich Obergfell <uobergfe@redhat.com>
> > > Cc: Fam Zheng <famz@redhat.com>
> > > Cc: Rusty Russell <rusty@rustcorp.com.au>
> > > Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > > ---
> > >  drivers/pci/pci-driver.c | 6 ++++++
> > >  1 file changed, 6 insertions(+)
> > > 
> > > diff --git a/drivers/pci/pci-driver.c b/drivers/pci/pci-driver.c
> > > index 3cb2210..2ebd2a8 100644
> > > --- a/drivers/pci/pci-driver.c
> > > +++ b/drivers/pci/pci-driver.c
> > > @@ -305,6 +305,12 @@ static long local_pci_probe(void *_ddi)
> > >  	 */
> > >  	pm_runtime_get_sync(dev);
> > >  	pci_dev->driver = pci_drv;
> > > +	/*
> > > +	 * When using kexec, msi might be left enabled by the previous kernel,
> > > +	 * this breaks things as some drivers assume msi/msi-x is off at boot.
> > > +	 * Fix this by forcing msi off at startup.
> > > +	 */
> > > +	pci_msi_off(pci_dev);
> > 
> > I think this makes sense, but I have a few questions.  This is a device
> > initialization thing, so it seems like a better fit for the enumeration
> > path, e.g,. pci_msi_init_pci_dev(), than for the driver binding path.
> > 
> > But when CONFIG_PCI_MSI=y, pci_msi_init_pci_dev() already does basically
> > the same thing, so we shouldn't need this change unless CONFIG_PCI_MSI is
> > not set in the kdump kernel.
> > 
> > If this is a problem just with kexeced kernels that don't have
> > CONFIG_PCI_MSI=y, I think I would prefer to fix this by moving
> > pci_msi_init_pci_dev() outside the #ifdef so it works regardless of
> > CONFIG_PCI_MSI.  That would also be nice because we could clean up the
> > duplication between pci_msi_off() and pci_msi_init_pci_dev().  It would
> > also make the starting machine state less dependent on the new kernel,
> > which seems like a good thing.
> 
> What you say above makes sense.
> OK so the simplest fix is something like below then.
> Fixes the duplication and kexec for CONFIG_PCI_MSI=n.
> Acceptable? Pls let me know, if yes I'll test and
> resubmit properly.
> 
> > Are there any bugzillas we could reference here?
> 
> I'll check this point. Maybe not - the real bugfix is
> patch 2/4, this was just found by reading code,
> but it's a dependency to make sure 2/4 does not
> introduce regressions.

OK, can you add the bugzilla link to that patch, if there is one?

> diff --git a/drivers/pci/msi.c b/drivers/pci/msi.c
> index 0e037af..2ab59d4 100644
> --- a/drivers/pci/msi.c
> +++ b/drivers/pci/msi.c
> @@ -1062,18 +1062,8 @@ EXPORT_SYMBOL(pci_msi_enabled);
>  void pci_msi_init_pci_dev(struct pci_dev *dev)
>  {
>  	INIT_LIST_HEAD(&dev->msi_list);
> -
> -	/* Disable the msi hardware to avoid screaming interrupts
> -	 * during boot.  This is the power on reset default so
> -	 * usually this should be a noop.
> -	 */
>  	dev->msi_cap = pci_find_capability(dev, PCI_CAP_ID_MSI);
> -	if (dev->msi_cap)
> -		msi_set_enable(dev, 0);
> -
>  	dev->msix_cap = pci_find_capability(dev, PCI_CAP_ID_MSIX);
> -	if (dev->msix_cap)
> -		msix_clear_and_set_ctrl(dev, PCI_MSIX_FLAGS_ENABLE, 0);
>  }
>  
>  /**
> diff --git a/drivers/pci/probe.c b/drivers/pci/probe.c
> index 8d2f400..c455501 100644
> --- a/drivers/pci/probe.c
> +++ b/drivers/pci/probe.c
> @@ -1485,6 +1485,12 @@ static struct pci_dev *pci_scan_device(struct pci_bus *bus, int devfn)
>  
>  static void pci_init_capabilities(struct pci_dev *dev)
>  {
> +	/* Disable the msi hardware to avoid screaming interrupts
> +	 * during boot.  This is the power on reset default so
> +	 * usually this should be a noop.
> +	 */
> +	pci_msi_off(dev);
> +
>  	/* MSI/MSI-X list */
>  	pci_msi_init_pci_dev(dev);

Could we move pci_msi_init_pci_dev() from msi.c to pci.c and make it look
something like the following?

  void pci_msi_off(struct pci_dev *dev)
  {
    if (dev->msi_cap) {
      ...
    }
    if (dev->msix_cap) {
      ...
    }
  }

  void pci_msi_init_pci_dev(struct pci_dev *dev)
  {
  #ifdef CONFIG_PCI_MSI
    INIT_LIST_HEAD(&dev->msi_list);
  #endif

    dev->msi_cap = pci_find_capability(dev, PCI_CAP_ID_MSI);
    dev->msix_cap = pci_find_capability(dev, PCI_CAP_ID_MSIX);
    pci_msi_off(dev);
  }

Then I think we could remove pci_msi_off() calls from a couple quirks as
well.  And we'd only have one MSI-related callout from
pci_init_capabilities().

Bjorn