On Sat, Nov 27, 2021 at 02:18:34AM +0100, Thomas Gleixner wrote:
> The [PCI] MSI code has gained quite some warts over time. A recent
> discussion unearthed a shortcoming: the lack of support for expanding
> PCI/MSI-X vectors after initialization of MSI-X.
>
> PCI/MSI-X has no requirement to setup all vectors when MSI-X is enabled in
> the device. The non-used vectors have just to be masked in the vector
> table. For PCI/MSI this is not possible because the number of vectors
> cannot be changed after initialization.
>
> The PCI/MSI code, but also the core MSI irq domain code are built around
> the assumption that all required vectors are installed at initialization
> time and freed when the device is shut down by the driver.
>
> Supporting dynamic expansion at least for MSI-X is important for VFIO so
> that the host side interrupts for passthrough devices can be installed on
> demand.
>
> This is the first part of a large (total 101 patches) series which
> refactors the [PCI]MSI infrastructure to make runtime expansion of MSI-X
> vectors possible. The last part (10 patches) provide this functionality.
>
> The first part is mostly a cleanup which consolidates code, moves the PCI
> MSI code into a separate directory and splits it up into several parts.
>
> No functional change intended except for patch 2/N which changes the
> behaviour of pci_get_vector()/affinity() to get rid of the assumption that
> the provided index is the "index" into the descriptor list instead of using
> it as the actual MSI[X] index as seen by the hardware. This would break
> users of sparse allocated MSI-X entries, but non of them use these
> functions.
I don't know all the irqdomain stuff all that well anymore, but I read
through all the patches and only noticed a small spello
[patch 02/22] PCI/MSI: Fix pci_irq_vector()/pci_irq_get_attinity()
^^^^ ff
It all seems good, I especially like the splitting of msi.c and
removal of ops..
Reviewed-by: Jason Gunthorpe <jgg@xxxxxxxxxx>
Thanks,
Jason
