While multi-MSI appears to work with pci-hyperv.c, there was a concern about
how linux was doing the ITRE allocations.  Patch 2 addresses the concern.

However, patch 2 exposed an issue with how compose_msi_msg() was freeing a
previous allocation when called for the Nth time.  Imagine a driver using
pci_alloc_irq_vectors() to request 32 MSIs.  This would cause compose_msi_msg()
to be called 32 times, once for each MSI.  With patch 2, MSI0 would allocate
the ITREs needed, and MSI1-31 would use the cached information.  Then the driver
uses request_irq() on MSI1-17.  This would call compose_msi_msg() again on those
MSIs, which would again use the cached information.  Then unmask() would be
called to retarget the MSIs to the right VCPU vectors.  Finally, the driver
calls request_irq() on MSI0.  This would call conpose_msi_msg(), which would
free the block of 32 MSIs, and allocate a new block.  This would undo the
retarget of MSI1-17, and likely leave those MSIs targeting invalid VCPU vectors.
This is addressed by patch 1, which is introduced first to prevent a regression.

Jeffrey Hugo (2):
  PCI: hv: Reuse existing ITRE allocation in compose_msi_msg()
  PCI: hv: Fix interrupt mapping for multi-MSI

 drivers/pci/controller/pci-hyperv.c | 76 ++++++++++++++++++++++++++++---------
 1 file changed, 59 insertions(+), 17 deletions(-)

-- 
2.7.4

On Wed, May 11, 2022 at 09:22:11AM -0600, Jeffrey Hugo wrote:
> While multi-MSI appears to work with pci-hyperv.c, there was a concern about
> how linux was doing the ITRE allocations.  Patch 2 addresses the concern.
> 
> However, patch 2 exposed an issue with how compose_msi_msg() was freeing a
> previous allocation when called for the Nth time.  Imagine a driver using
> pci_alloc_irq_vectors() to request 32 MSIs.  This would cause compose_msi_msg()
> to be called 32 times, once for each MSI.  With patch 2, MSI0 would allocate
> the ITREs needed, and MSI1-31 would use the cached information.  Then the driver
> uses request_irq() on MSI1-17.  This would call compose_msi_msg() again on those
> MSIs, which would again use the cached information.  Then unmask() would be
> called to retarget the MSIs to the right VCPU vectors.  Finally, the driver
> calls request_irq() on MSI0.  This would call conpose_msi_msg(), which would
> free the block of 32 MSIs, and allocate a new block.  This would undo the
> retarget of MSI1-17, and likely leave those MSIs targeting invalid VCPU vectors.
> This is addressed by patch 1, which is introduced first to prevent a regression.
> 
> Jeffrey Hugo (2):
>   PCI: hv: Reuse existing ITRE allocation in compose_msi_msg()
>   PCI: hv: Fix interrupt mapping for multi-MSI

Applied this version to hyperv-next. Thanks.

On 5/11/2022 11:51 AM, Wei Liu wrote:
> On Wed, May 11, 2022 at 09:22:11AM -0600, Jeffrey Hugo wrote:
>> While multi-MSI appears to work with pci-hyperv.c, there was a concern about
>> how linux was doing the ITRE allocations.  Patch 2 addresses the concern.
>>
>> However, patch 2 exposed an issue with how compose_msi_msg() was freeing a
>> previous allocation when called for the Nth time.  Imagine a driver using
>> pci_alloc_irq_vectors() to request 32 MSIs.  This would cause compose_msi_msg()
>> to be called 32 times, once for each MSI.  With patch 2, MSI0 would allocate
>> the ITREs needed, and MSI1-31 would use the cached information.  Then the driver
>> uses request_irq() on MSI1-17.  This would call compose_msi_msg() again on those
>> MSIs, which would again use the cached information.  Then unmask() would be
>> called to retarget the MSIs to the right VCPU vectors.  Finally, the driver
>> calls request_irq() on MSI0.  This would call conpose_msi_msg(), which would
>> free the block of 32 MSIs, and allocate a new block.  This would undo the
>> retarget of MSI1-17, and likely leave those MSIs targeting invalid VCPU vectors.
>> This is addressed by patch 1, which is introduced first to prevent a regression.
>>
>> Jeffrey Hugo (2):
>>    PCI: hv: Reuse existing ITRE allocation in compose_msi_msg()
>>    PCI: hv: Fix interrupt mapping for multi-MSI
> 
> Applied this version to hyperv-next. Thanks.

Thanks for picking it up.  Sorry about the confusion.