[Skiboot] [RFC 0/3] phb4/capp: Provide user tunable for STQ/DMA engine allocation
andrew.donnellan at au1.ibm.com
Sun Sep 9 23:23:54 AEST 2018
On 09/09/18 21:41, Vaibhav Jain wrote:
> When CAPP is initialized on request of CXL, opal reserves certain number
> of PEC STQ/DMA engines needed for communication with the PSL/XSL on the card.
> Since this allocation comes from the shared pool of PEC resources, it will have
> an impact on the performance of a CAPI or PCI throughput.
> Presently this allocation is statically defined in skiboot with CAPP init code
> only taking into account the link width. However since different cards will
> have will different workload properties hence it might be prudent to let user
> tune these values to get an optimal performance for their CAPI workload.
> Hence this patchset proposes a new scheme called 'capp attribute overrides'
> that lets user set a nvram config to override the the number of DMA/STQ engines
> opal reserves for CAPP. The nvram config is defined following a certain
> convention to address multiple CAPPs present in the system.
While NVRAM attributes might be a good idea for bringup/enablement
environments I think it's not ideal for doing device-specific
configuration in production environments.
Would it be possible to expose this to Linux in the form of an
additional OPAL API call, or does this need to be determined too early
in the process? That would allow the allocation to be determined by the
driver, where it's easier to add device-specific quirks.
Andrew Donnellan OzLabs, ADL Canberra
andrew.donnellan at au1.ibm.com IBM Australia Limited
More information about the Skiboot