Message-ID: <KU1P153MB01201EEC711EC37D30687940BF330@KU1P153MB0120.APCP153.PROD.OUTLOOK.COM>
Date:   Wed, 30 Sep 2020 00:38:04 +0000
From:   Dexuan Cui <decui@...rosoft.com>
To:     Lorenzo Pieralisi <lorenzo.pieralisi@....com>,
        "maz@...nel.org" <maz@...nel.org>
CC:     "wei.liu@...nel.org" <wei.liu@...nel.org>,
        KY Srinivasan <kys@...rosoft.com>,
        Haiyang Zhang <haiyangz@...rosoft.com>,
        Stephen Hemminger <sthemmin@...rosoft.com>,
        "bhelgaas@...gle.com" <bhelgaas@...gle.com>,
        "linux-hyperv@...r.kernel.org" <linux-hyperv@...r.kernel.org>,
        "linux-pci@...r.kernel.org" <linux-pci@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Michael Kelley <mikelley@...rosoft.com>,
        Jake Oshins <jakeo@...rosoft.com>
Subject: RE: [PATCH v2] PCI: hv: Fix hibernation in case interrupts are not
 re-created

> From: Lorenzo Pieralisi <lorenzo.pieralisi@....com>
> Sent: Monday, September 28, 2020 3:43 AM
>
> [+MarcZ - this patch needs IRQ maintainers vetting]

Sure. Hi MarkZ, please also review the patch. Thanks!

> On Tue, Sep 08, 2020 at 04:17:59PM -0700, Dexuan Cui wrote:
> > Hyper-V doesn't trap and emulate the accesses to the MSI/MSI-X
> > registers, and we must use hv_compose_msi_msg() to ask Hyper-V to
> > create the IOMMU Interrupt Remapping Table Entries. This is not an issue
> > for a lot of PCI device drivers (e.g. NVMe driver, Mellanox NIC drivers),
> > which destroy and re-create the interrupts across hibernation, so
> > hv_compose_msi_msg() is called automatically. However, some other
> > PCI device drivers (e.g. the Nvidia driver) may not destroy and re-create
> > the interrupts across hibernation, so hv_pci_resume() has to call
> > hv_compose_msi_msg(), otherwise the PCI device drivers can no longer
> > receive MSI/MSI-X interrupts after hibernation.
>
> This looks like drivers bugs and I don't think the HV controller
> driver is where you should fix them.

IMHO this is not a PCI device driver bug, because I think a PCI device driver
is allowed to keep and re-use the MSI/MSI-X interrupts across hibernation,
otherwise we would not have pci_restore_msi_state() in pci_restore_state().

The in-tree open-source Nvidia GPU driver drivers/gpu/drm/nouveau is such
a PCI device driver that re-uses the MSI-X interrupts across hibernation.
The out-of-tree proprietary Nvidia GPU driver also does the same thing.
It looks some other in-tree PCI device drivers also do the same thing, though
I don't remember their names offhand.

IMO it's much better to change the pci-hyperv driver once and for all, than
to change every such existing (and future?) PCI device driver.

pci_restore_msi_state() directly writes the MSI/MSI-X related registers
in __pci_write_msi_msg() and msix_mask_irq(). On a physical machine, this
works perfectly, but for a Linux VM running on a hypervisor, which typically
enables IOMMU interrupt remapping, the hypervisor usually should trap and
emulate the write accesses to the MSI/MSI-X registers, so the hypervisor
is able to create the necessary interrupt remapping table entries in the
IOMMU, and the MSI/MSI-X interrupts can work in the VM. Hyper-V is different
from other hypervisors in that it does not trap and emulate the write
accesses, and instead it uses a para-virtualized method, which requires
the VM to call hv_compose_msi_msg() to notify the hypervisor of the info
that would be passed to the hypervisor in the case of the trap-and-emulate
method.

I mean this is a Hyper-V specific problem, so IMO we should fix the pci-hyperv
driver rather than change the PCI device drivers, which work perfectly on a
physical machine and on other hypervisors. Also it can be difficult or 
impossible to ask the authors of the aforementioned PCI device drivers to
destry and re-create MSI/MSI-X acorss hibernation, especially for the
out-of-tree driver(s).

> Regardless, this commit log does not provide the information that
> it should.

Hi Lozenzo, I'm happy to add more info. Can you please let me know
what extra info I should provide?

> > Fixes: ac82fc832708 ("PCI: hv: Add hibernation support")
> > Signed-off-by: Dexuan Cui <decui@...rosoft.com>
> > Reviewed-by: Jake Oshins <jakeo@...rosoft.com>
> >
> > ---
> >
> > Changes in v2:
> >     Fixed a typo in the comment in hv_irq_unmask. Thanks to Michael!
> >     Added Jake's Reviewed-by.
> >
> >  drivers/pci/controller/pci-hyperv.c | 44 +++++++++++++++++++++++++++++
> >  1 file changed, 44 insertions(+)
> >
> > diff --git a/drivers/pci/controller/pci-hyperv.c
> b/drivers/pci/controller/pci-hyperv.c
> > index fc4c3a15e570..dd21afb5d62b 100644
> > --- a/drivers/pci/controller/pci-hyperv.c
> > +++ b/drivers/pci/controller/pci-hyperv.c
> > @@ -1211,6 +1211,21 @@ static void hv_irq_unmask(struct irq_data *data)
> >  	pbus = pdev->bus;
> >  	hbus = container_of(pbus->sysdata, struct hv_pcibus_device, sysdata);
> >
> > +	if (hbus->state == hv_pcibus_removing) {
> > +		/*
> > +		 * During hibernation, when a CPU is offlined, the kernel tries
> > +		 * to move the interrupt to the remaining CPUs that haven't
> > +		 * been offlined yet. In this case, the below hv_do_hypercall()
> > +		 * always fails since the vmbus channel has been closed, so we
> > +		 * should not call the hypercall, but we still need
> > +		 * pci_msi_unmask_irq() to reset the mask bit in desc->masked:
> > +		 * see cpu_disable_common() -> fixup_irqs() ->
> > +		 * irq_migrate_all_off_this_cpu() -> migrate_one_irq().
> > +		 */
> > +		pci_msi_unmask_irqpci_msi_unmask_irq(data);
>
> This is not appropriate - it looks like a plaster to paper over an
> issue with hyper-V hibernation code sequence. Fix that issue instead
> of papering over it here.
>
> Thanks,
> Lorenzo

IMO this patch is fixing this Hyper-V specific problem. :-)

The probem is unique to Hyper-V because chip->irq_unmask() may fail in a
Linux VM running on Hyper-V. 

chip->irq_unmask() can not fail on a physical machine. I guess this is why 
the return value of irq_unmask() is defined as "void" in include/linux/irq.h:
struct irq_chip {
...
        void            (*irq_mask)(struct irq_data *data);
        void            (*irq_unmask)(struct irq_data *data);

As I described in the comment, in a VM on Hyper-V, chip->irq_unmask()
fails during the suspending phase of hibernation because it's called
when the non-boot CPUs are being offlined, and at this time all the devices,
including Hyper-V VMBus devices, have been "frozen" -- this is part of
the standard Linux hibernation workflow. Since Hyper-V thinks the VM has
frozen the pci-hyperv VMBus device at this moment (i.e. closed the VMBus
channel of the VMBus device), it fails chip->irq_unmask(), i.e.
hv_irq_unmask() -> hv_do_hypercall().

On a physical machine, unmasking an MSI/MSI-X register just means an MMIO
write, which I think can not fail here.

So I think this patch is the correct fix, considering Hyper-V's unique
implementation of the MSI "chip" (i.e. Hyper-V does not trap and emulate
the MSI/MSI-X register accesses, and uses a para-virtualized method as I
explained above), and the fact that I shouldn't and can't change the
standard Linux hibernation workflow.

In hv_irq_unmask(), when I skip the hypercall in the case of
hbus->state == hv_pcibus_removing, I still need the pci_msi_unmask_irq(),
because of the sequences in kernel/irq/cpuhotplug.c:

static bool migrate_one_irq(struct irq_desc *desc)
{
...
        if (maskchip && chip->irq_mask)
                chip->irq_mask(d);
...
        err = irq_do_set_affinity(d, affinity, false);
...
        if (maskchip && chip->irq_unmask)
                chip->irq_unmask(d);

Here if hv_irq_unmask does not call pci_msi_unmask_irq(), the desc->masked
remains "true", so later after hibernation, the MSI interrupt line always
reamins masked, which is incorrect.

Here the slient failure of hv_irq_unmask() does not matter since all the
non-boot CPUs are being offlined (meaning all the devices have been
frozen). Note: the correct affinity info is still updated into the
irqdata data structure in migrate_one_irq() -> irq_do_set_affinity() ->
hv_set_affinity(), so when the VM resumes, hv_pci_resume() ->
hv_pci_restore_msi_state() is able to correctly restore the irqs with
the correct affinity.

I hope the explanation can help clarify things. I understand this is not
as natual as tht case that Linux runs on a physical machine, but due to
the unique PCI pass-through implementation of Hyper-V, IMO this is
the only viable fix for the problem here. BTW, this patch is only confined
to the pci-hyperv driver and I believe it can no cause any regression.

Thanks,
-- Dexuan


> > +		return;
> > +	}
> > +
> >  	spin_lock_irqsave(&hbus->retarget_msi_interrupt_lock, flags);
> >
> >  	params = &hbus->retarget_msi_interrupt_params;
> > @@ -3372,6 +3387,33 @@ static int hv_pci_suspend(struct hv_device
> *hdev)
> >  	return 0;
> >  }
> >
> > +static int hv_pci_restore_msi_msg(struct pci_dev *pdev, void *arg)
> > +{
> > +	struct msi_desc *entry;
> > +	struct irq_data *irq_data;
> > +
> > +	for_each_pci_msi_entry(entry, pdev) {
> > +		irq_data = irq_get_irq_data(entry->irq);
> > +		if (WARN_ON_ONCE(!irq_data))
> > +			return -EINVAL;
> > +
> > +		hv_compose_msi_msg(irq_data, &entry->msg);
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> > +/*
> > + * Upon resume, pci_restore_msi_state() -> ... ->  __pci_write_msi_msg()
> > + * re-writes the MSI/MSI-X registers, but since Hyper-V doesn't trap and
> > + * emulate the accesses, we have to call hv_compose_msi_msg() to ask
> > + * Hyper-V to re-create the IOMMU Interrupt Remapping Table Entries.
> > + */
> > +static void hv_pci_restore_msi_state(struct hv_pcibus_device *hbus)
> > +{
> > +	pci_walk_bus(hbus->pci_bus, hv_pci_restore_msi_msg, NULL);
> > +}
> > +
> >  static int hv_pci_resume(struct hv_device *hdev)
> >  {
> >  	struct hv_pcibus_device *hbus = hv_get_drvdata(hdev);
> > @@ -3405,6 +3447,8 @@ static int hv_pci_resume(struct hv_device *hdev)
> >
> >  	prepopulate_bars(hbus);
> >
> > +	hv_pci_restore_msi_state(hbus);
> > +
> >  	hbus->state = hv_pcibus_installed;
> >  	return 0;
> >  out:

Powered by blists - more mailing lists