[Xen-devel] [PATCH] PCI/MSI: don't disable AMD IOMMU MSI on Xen dom0 (was: Re: [PATCH V2] amd iommu: re-enable iommu msi if dom0 disabled it)

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 14.06.12 at 17:15, Wei Wang <wei.wang2@xxxxxxx> wrote:
> Am 14.06.2012 16:18, schrieb Jan Beulich:
>> Have you at all considered getting this fixed on the kernel side?
>> As I don't have direct access to any AMD IOMMU capable
>> system - can one, other than by enumerating the respective
>> PCI IDs or reading ACPI tables, reasonably easily identify the
>> devices in question (e.g. via vendor/class/sub-class or some
>> such)? That might permit skipping those in the offending kernel
>> code...
> 
> AMD IOMMUs (both v1 and v2) uses class id 08 (System Base Peripheral) 
> and sub class id 06 (IOMMU). Combined with PCI_VENDEOR_ID_AMD, this 
> should be enough to identify amd iommu device. I could send you a kernel 
> patch for review using this approach. I would believe that fixing this 
> issue in 4.2, no matter how, is really important for amd iommu.

As you didn't come forward with anything, here's my first
take on this:

Commit d5dea7d95c48d7bc951cee4910a7fd9c0cd26fb0 ("PCI: msi: Disable msi
interrupts when we initialize a pci device") caused MSI to get disabled
on Xen Dom0 despite it having got turned on purposefully by the
hypervisor. As an immediate band aid, suppress the disabling in this
specific case.

I don't think, however, that either the original change or this fix are
really appropriate. The original fix, besides leaving aside the
presence of a hypervisor like Xen, doesn't deal with all possible
cases (in particular it has no effect if the secondary kernel was built
with CONFIG_PCI_MSI unset). And taking into account a hypervisor like
Xen, the logic like this should probably be skipped altogether (i.e. it
should be entirely left to the hypervisor, being the entity in control
of IRQs).

Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
Cc: stable@xxxxxxxxxx 

---
 drivers/pci/msi.c       |    7 +++++++
 include/linux/pci_ids.h |    1 +
 2 files changed, 8 insertions(+)

--- 3.5-rc3/drivers/pci/msi.c
+++ 3.5-rc3-xen-pci-msi-no-iommu-disable/drivers/pci/msi.c
@@ -20,6 +20,7 @@
 #include <linux/errno.h>
 #include <linux/io.h>
 #include <linux/slab.h>
+#include <xen/xen.h>
 
 #include "pci.h"
 #include "msi.h"
@@ -1022,7 +1023,13 @@ void pci_msi_init_pci_dev(struct pci_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.
+        * But on a Xen host don't do this for IOMMUs which the hypervisor
+        * is in control of (and hence has already enabled on purpose).
         */
+       if (xen_initial_domain()
+           && (dev->class >> 8) == PCI_CLASS_SYSTEM_IOMMU
+           && dev->vendor == PCI_VENDOR_ID_AMD)
+               return;
        pos = pci_find_capability(dev, PCI_CAP_ID_MSI);
        if (pos)
                msi_set_enable(dev, pos, 0);
--- 3.5-rc3/include/linux/pci_ids.h
+++ 3.5-rc3-xen-pci-msi-no-iommu-disable/include/linux/pci_ids.h
@@ -75,6 +75,7 @@
 #define PCI_CLASS_SYSTEM_RTC           0x0803
 #define PCI_CLASS_SYSTEM_PCI_HOTPLUG   0x0804
 #define PCI_CLASS_SYSTEM_SDHCI         0x0805
+#define PCI_CLASS_SYSTEM_IOMMU         0x0806
 #define PCI_CLASS_SYSTEM_OTHER         0x0880
 
 #define PCI_BASE_CLASS_INPUT           0x09




_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel