Handling MSI-X Vector Shortages

Handling MSI-X Vector Shortages

[Tang, CQ]

Hi,
              I am writing a driver to deal with MSI-X vector shortage. PCI local bus specification 6.8.3.2 clearly says that, software controlled aliasing can be used to solve the problem, where multiple MSI-X table entries are configured with the same vector.
              pci_enable_msix() call from driver only configure the entries specified in the argument with different vectors. Does Linux provide help function to copy MSI-X table entries? I search the pci/msi source code and don't find an exported function to do that.   
       If there are such function, please tell the function name. if not, can we add such help function?
       Or is driver allowed to access the MSI-X table entries directly, etc, read from one entry (configured by PCI subsystem), and write to another unused entry?

              Thank you very much for the help.

--CQ Tang

Re: Handling MSI-X Vector Shortages

[Bjorn Helgaas]

[+cc Christoph, Alexander]

Best to wrap your emails to fit in 70 columns or so.

On Wed, Jun 15, 2016 at 07:03:09PM +0000, Tang, CQ wrote:
> Hi,
>               I am writing a driver to deal with MSI-X vector shortage. PCI local bus specification 6.8.3.2 clearly says that, software controlled aliasing can be used to solve the problem, where multiple MSI-X table entries are configured with the same vector.
>               pci_enable_msix() call from driver only configure the entries specified in the argument with different vectors. Does Linux provide help function to copy MSI-X table entries? I search the pci/msi source code and don't find an exported function to do that.   

I don't think so.

>        If there are such function, please tell the function name. if not, can we add such help function?
>        Or is driver allowed to access the MSI-X table entries directly, etc, read from one entry (configured by PCI subsystem), and write to another unused entry?

I don't think drivers should fiddle with MSI-X table entries directly.
That sounds like a way to confuse the PCI core.  I added Christoph
and Alexander because they're working on similar issues.

Bjorn