Re: [PATCH V4 10/11] vfio/pci: Support dynamic MSI-X

[Alex Williamson <alex.williamson@redhat.com>]

On Fri, 5 May 2023 08:10:33 +0000
"Tian, Kevin" <kevin.tian@intel.com> wrote:

> > From: Chatre, Reinette <reinette.chatre@intel.com>
> > Sent: Saturday, April 29, 2023 2:35 AM
> > 
> > Hi Kevin,
> > 
> > On 4/27/2023 11:50 PM, Tian, Kevin wrote:  
> > >> From: Chatre, Reinette <reinette.chatre@intel.com>
> > >> Sent: Friday, April 28, 2023 1:36 AM
> > >>
> > >> pci_msix_alloc_irq_at() enables an individual MSI-X interrupt to be
> > >> allocated after MSI-X enabling.
> > >>
> > >> Use dynamic MSI-X (if supported by the device) to allocate an interrupt
> > >> after MSI-X is enabled. An MSI-X interrupt is dynamically allocated at
> > >> the time a valid eventfd is assigned. This is different behavior from
> > >> a range provided during MSI-X enabling where interrupts are allocated
> > >> for the entire range whether a valid eventfd is provided for each
> > >> interrupt or not.
> > >>
> > >> The PCI-MSIX API requires that some number of irqs are allocated for
> > >> an initial set of vectors when enabling MSI-X on the device. When
> > >> dynamic MSIX allocation is not supported, the vector table, and thus
> > >> the allocated irq set can only be resized by disabling and re-enabling
> > >> MSI-X with a different range. In that case the irq allocation is
> > >> essentially a cache for configuring vectors within the previously
> > >> allocated vector range. When dynamic MSI-X allocation is supported,
> > >> the API still requires some initial set of irqs to be allocated, but
> > >> also supports allocating and freeing specific irq vectors both
> > >> within and beyond the initially allocated range.
> > >>
> > >> For consistency between modes, as well as to reduce latency and improve
> > >> reliability of allocations, and also simplicity, this implementation
> > >> only releases irqs via pci_free_irq_vectors() when either the interrupt
> > >> mode changes or the device is released.  
> > >
> > > It improves the reliability of allocations from the calling device p.o.v.
> > >
> > > But system-wide this is not efficient use of irqs and not releasing them
> > > timely may affect the reliability of allocations for other devices.  
> > 
> > Could you please elaborate how other devices may be impacted?  
> 
> the more this devices reserves the less remains for others, e.g. irte entries.
> 
> >   
> > > Should this behavior be something configurable?  
> > 
> > This is not clear to me and I look to you for guidance here. From practical
> > side it looks like configuration via module parameters is supported but
> > whether it should be done is not clear to me.
> > 
> > When considering this we need to think about what the user may expect
> > when
> > turning on/off the configuration. For example, MSI-X continues to allocate a
> > range of interrupts during enabling. These have always been treated as a
> > "cache" (interrupts remain allocated, whether they have an associated
> > trigger
> > or not). If there is new configurable behavior, do you expect that the
> > driver needs to distinguish between the original "cache" that the user is
> > used to and the new dynamic allocations? That is, should a dynamic MSI-X
> > capable device always free interrupts when user space removes an eventfd
> > or should only interrupts that were allocated dynamically be freed
> > dynamically?  
> 
> That looks tricky. Probably that is why Alex suggested doing this simple
> scheme and it is on par with the old logic anyway. So I'll withdraw this
> comment.

Don't forget we're also releasing the irq reservations when the guest
changes interrupt mode, ex. reboot, so the "caching" is really only
within a session of the guest/userspace driver where it would be
unusual to have an unused reservation for an extended period.  Thanks,

Alex