[SRU][N/P][PATCH 0/1] kvmppc_set_passthru_irq_hv: Could not assign IRQ map traces are seen when pci device is attached to kvm guest when "xive=off" is set (LP: #2109951)

[Massimiliano Pellizzer]

BugLink: https://bugs.launchpad.net/bugs/2109951

[Impact]

KVM: PPC: Book3S HV: Fix IRQ map warnings with XICS on pSeries KVM Guest

The commit 9576730d0e6e ("KVM: PPC: select IRQ_BYPASS_MANAGER") enabled
IRQ_BYPASS_MANAGER when CONFIG_KVM was set. Subsequently, commit
c57875f5f9be ("KVM: PPC: Book3S HV: Enable IRQ bypass") enabled IRQ
bypass and added the necessary callbacks to create/remove the mappings
between host real IRQ and the guest GSI.

The availability of IRQ bypass is determined by the arch-specific
function kvm_arch_has_irq_bypass(), which invokes
kvmppc_irq_bypass_add_producer_hv(). This function, in turn, calls
kvmppc_set_passthru_irq_hv() to create a mapping in the passthrough IRQ
map, associating a host IRQ to a guest GSI.

However, when a pSeries KVM guest (L2) is booted within an LPAR (L1)
with the kernel boot parameter `xive=off`, it defaults to using emulated
XICS controller. As an attempt to establish host IRQ to guest GSI
mappings via kvmppc_set_passthru_irq() on a PCI device hotplug
(passhthrough) operation fail, returning -ENOENT. This failure occurs
because only interrupts with EOI operations handled through OPAL calls
(verified via is_pnv_opal_msi()) are currently supported.

These mapping failures lead to below repeated warnings in the L1 host:

 [  509.220349] kvmppc_set_passthru_irq_hv: Could not assign IRQ map for
(58,4970)
 [  509.220368] kvmppc_set_passthru_irq (irq 58, gsi 4970) fails: -2
 [  509.220376] vfio-pci 0015:01:00.0: irq bypass producer (token
0000000090bc635b) registration fails: -2
 ...
 [  509.291781] vfio-pci 0015:01:00.0: irq bypass producer (token
000000003822eed8) registration fails: -2

Fix this by restricting IRQ bypass enablement on pSeries systems by
making the IRQ bypass callbacks unavailable when running on pSeries
platform.

[Fix]

Backport from mainline:
- ccdb36cbe65f KVM: PPC: Book3S HV: Fix IRQ map warnings with XICS on pSeries
KVM Guest

[Test Plan]

The issue can be tested executing the following steps:

1. Start a guest with kernel
2. unbind the pci device from host using below command

root at ltcrain8og-lp6:~/NET# virsh nodedev-detach pci_0028_01_00_0
Device pci_0028_01_00_0 detached
root at ltcrain8og-lp6:~/NET# virsh nodedev-detach pci_0028_01_00_1
Device pci_0028_01_00_1 detached

3. Now attach the pci device to guest using virsh attach-device command. The xml
snippet used is given below

root at ltcrain8og-lp6:~/NET# vi pci2.xml
<hostdev mode='subsystem' type='pci' managed='yes'>
  <source>
    <address domain='0x0028' bus='0x01' slot='0x00' function='0x0'/>
  </source>
</hostdev>

root at ltcrain8og-lp6:~/NET# virsh attach-device vm pci2.xml
Device attached successfully

4. Verify that on guest side the pci device gets attached successfully, and the
interface comes up. Moreover, check that the kernel ring buffer does not show
any error of the following type:

[57906.347794] kvmppc_set_passthru_irq_hv: Could not assign IRQ map for
(100,4900)
[57906.347812] kvmppc_set_passthru_irq (irq 100, gsi 4900) fails: -2
[57906.347817] vfio-pci 0028:01:00.0: irq bypass producer (token
00000000c84b35a9) registration fails: -2

[Where Problems Could Occur]

The fix affects the IRQ bypass infrastructure in the PowerPC KVM hypervisor,
specifically the initialization of architecture-specific callback handlers for
IRQ passthrough. An issue with this fix may introduce problems such as the VFIO
subsystem failing to handle IRQ bypass even on capable PowerNV systems.

[Other Info]

Oracular not patched because EOL.