[Skiboot] [RESEND][PATCH] phb4, doc: Make GPU-Direct bandwidth optimizations Witherspoon specific
stewart at linux.ibm.com
Wed Aug 1 16:27:59 AEST 2018
Vaibhav Jain <vaibhav at linux.ibm.com> writes:
> Patch 5690c5a8980f("phb4: Reallocate PEC2 DMA-Read engines to improve
> GPU-Direct bandwidth") introduced allocation of extra DMA-read engines
> for improving Mellanox CX5 GPU-Direct bandwidth. At present CX5 is the
> only card thats using these optimizations so these changes will only
> impact Witherspoon systems.
> However hardware team has raised the possibility of other
> non-witherspoon systems in future that may be using a similar card,
> where these optimizations wont be needed. So they have asked us to
> make these changes Witherspoon specific.
> Hence this patch updates the phb4_init_capp_regs() &
> enable_capi_mode() to configure the extra DMA-read engine allocation
> if and only if skiboot is running on Witherspoon platform.
> Cc: stable #6.0.6+
> Fixes: 5690c5a8980f("phb4: Reallocate PEC2 DMA-Read engines to improve
> GPU-Direct bandwidth")
> Signed-off-by: Vaibhav Jain <vaibhav at linux.ibm.com>
> Resend -> Updated the request for merge to stable from 5.0.6+ to 6.0.6+
> .../opal-pci-set-phb-capi-mode-93.rst | 8 ++--
> hw/phb4.c | 47 ++++++++++++-------
> 2 files changed, 35 insertions(+), 20 deletions(-)
> diff --git a/doc/opal-api/opal-pci-set-phb-capi-mode-93.rst b/doc/opal-api/opal-pci-set-phb-capi-mode-93.rst
> index 6a8d2be8..09ecd69c 100644
> --- a/doc/opal-api/opal-pci-set-phb-capi-mode-93.rst
> +++ b/doc/opal-api/opal-pci-set-phb-capi-mode-93.rst
> @@ -44,10 +44,10 @@ CAPP-PSL transactions.
> -* If PHB is in PEC2 then requesting mode `OPAL_PHB_CAPI_MODE_DMA_TVT1` will
> - allocate extra 16/8 dma read engines to the PHB depending on its stack
> - (stack 0/ stack 1). This is needed to improve the Direct-GPU DMA read
> - performance for the Mellanox CX5 card.
> +* On a Witherspoon system if PHB is in PEC2 then requesting mode
> + `OPAL_PHB_CAPI_MODE_DMA_TVT1` will allocate extra 16/8 dma read engines to the
> + PHB depending on its stack (stack 0/ stack 1). This is needed to improve the
> + Direct-GPU DMA read performance for the Mellanox CX5 card.
> * Mode `OPAL_PHB_CAPI_MODE_PCIE` not yet supported on Power-9.
> * Requesting mode `OPAL_PHB_CAPI_MODE_CAPI` on Power-9 will disable fast-reboot.
> * Modes `OPAL_PHB_CAPI_MODE_DMA`, `OPAL_PHB_CAPI_MODE_SNOOP_OFF` are
> diff --git a/hw/phb4.c b/hw/phb4.c
> index a3aa8b80..ee238109 100644
> --- a/hw/phb4.c
> +++ b/hw/phb4.c
> @@ -148,6 +148,9 @@ static void phb4_init_hw(struct phb4 *p);
> #define PHB4_CAN_STORE_EOI(p) \
> (XIVE_STORE_EOI_ENABLED && ((p)->rev >= PHB4_REV_NIMBUS_DD20))
> +/* Are we running on a Witherspoon system */
> +#define IS_WITHERSPOON() (strcmp(platform.name, "Witherspoon") == 0)
I really don't want us detecting platforms this way and doing platform
specific behaviour outside of the platform/ code.
So, there'll instead have to be some kind of platform call for working
out what to do.
I am wondering the wiseness of having this opal call being overloaded in
such a way though.
OPAL Architect, IBM.
More information about the Skiboot