Current implementation will disallow P2P DMA if the participating
devices belong to different root complexes. Implementation allows
this default behavior to be overridden for whitelisted devices. The
patch adds an AMD host bridge to be whitelisted
Signed-off-by: Ramesh Errabolu <[email protected]>
---
drivers/pci/p2pdma.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
index 196382630363..7003bb9faf23 100644
--- a/drivers/pci/p2pdma.c
+++ b/drivers/pci/p2pdma.c
@@ -305,6 +305,8 @@ static const struct pci_p2pdma_whitelist_entry {
{PCI_VENDOR_ID_INTEL, 0x2032, 0},
{PCI_VENDOR_ID_INTEL, 0x2033, 0},
{PCI_VENDOR_ID_INTEL, 0x2020, 0},
+ /* AMD Host Bridge Devices */
+ {PCI_VENDOR_ID_AMD, 0x1480, 0},
{}
};
--
2.31.1
On Wed, Aug 11, 2021 at 3:11 PM Ramesh Errabolu <[email protected]> wrote:
>
> Current implementation will disallow P2P DMA if the participating
> devices belong to different root complexes. Implementation allows
> this default behavior to be overridden for whitelisted devices. The
> patch adds an AMD host bridge to be whitelisted
Why do we need this? cpu_supports_p2pdma() should return true for all
AMD Zen CPUs.
Alex
>
> Signed-off-by: Ramesh Errabolu <[email protected]>
> ---
> drivers/pci/p2pdma.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
> index 196382630363..7003bb9faf23 100644
> --- a/drivers/pci/p2pdma.c
> +++ b/drivers/pci/p2pdma.c
> @@ -305,6 +305,8 @@ static const struct pci_p2pdma_whitelist_entry {
> {PCI_VENDOR_ID_INTEL, 0x2032, 0},
> {PCI_VENDOR_ID_INTEL, 0x2033, 0},
> {PCI_VENDOR_ID_INTEL, 0x2020, 0},
> + /* AMD Host Bridge Devices */
> + {PCI_VENDOR_ID_AMD, 0x1480, 0},
> {}
> };
>
> --
> 2.31.1
>
Am 2021-08-11 um 3:29 p.m. schrieb Alex Deucher:
> On Wed, Aug 11, 2021 at 3:11 PM Ramesh Errabolu <[email protected]> wrote:
>> Current implementation will disallow P2P DMA if the participating
>> devices belong to different root complexes. Implementation allows
>> this default behavior to be overridden for whitelisted devices. The
>> patch adds an AMD host bridge to be whitelisted
> Why do we need this? cpu_supports_p2pdma() should return true for all
> AMD Zen CPUs.
This is part of our on-going work to get P2P support upstream. We want
to use pci_p2pdma_distance_many to determine whether P2P is possible
between a pair of devices. This whitelist is used in this function. This
will affect the P2P links reported in the topology and it will be
double-checked in the BO mapping function to ensure a peer mapping is legal.
I think this change is a bit free of context at the moment, as we're
still working on a few other loose ends for P2P support in our internal
branch. I'm hoping we'll have a bigger patch series for upstreamable KFD
P2P support ready in a few weeks. I also think we'll probably want to
add a few more PCI IDs for other supported AMD root complexes.
Regards,
Felix
>
> Alex
>
>> Signed-off-by: Ramesh Errabolu <[email protected]>
>> ---
>> drivers/pci/p2pdma.c | 2 ++
>> 1 file changed, 2 insertions(+)
>>
>> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
>> index 196382630363..7003bb9faf23 100644
>> --- a/drivers/pci/p2pdma.c
>> +++ b/drivers/pci/p2pdma.c
>> @@ -305,6 +305,8 @@ static const struct pci_p2pdma_whitelist_entry {
>> {PCI_VENDOR_ID_INTEL, 0x2032, 0},
>> {PCI_VENDOR_ID_INTEL, 0x2033, 0},
>> {PCI_VENDOR_ID_INTEL, 0x2020, 0},
>> + /* AMD Host Bridge Devices */
>> + {PCI_VENDOR_ID_AMD, 0x1480, 0},
>> {}
>> };
>>
>> --
>> 2.31.1
>>
On Wed, Aug 11, 2021 at 4:50 PM Felix Kuehling <[email protected]> wrote:
>
>
> Am 2021-08-11 um 3:29 p.m. schrieb Alex Deucher:
> > On Wed, Aug 11, 2021 at 3:11 PM Ramesh Errabolu <[email protected]> wrote:
> >> Current implementation will disallow P2P DMA if the participating
> >> devices belong to different root complexes. Implementation allows
> >> this default behavior to be overridden for whitelisted devices. The
> >> patch adds an AMD host bridge to be whitelisted
> > Why do we need this? cpu_supports_p2pdma() should return true for all
> > AMD Zen CPUs.
>
> This is part of our on-going work to get P2P support upstream. We want
> to use pci_p2pdma_distance_many to determine whether P2P is possible
> between a pair of devices. This whitelist is used in this function. This
> will affect the P2P links reported in the topology and it will be
> double-checked in the BO mapping function to ensure a peer mapping is legal.
>
> I think this change is a bit free of context at the moment, as we're
> still working on a few other loose ends for P2P support in our internal
> branch. I'm hoping we'll have a bigger patch series for upstreamable KFD
> P2P support ready in a few weeks. I also think we'll probably want to
> add a few more PCI IDs for other supported AMD root complexes.
We don't need to keep adding AMD root complexes. You must be using an
older kernel or something. All root complexes on all Zen platforms
support P2P DMA. See:
commit dea286bb71baded7d2fb4f090e3b9fd2c1ccac58
Author: Logan Gunthorpe <[email protected]>
Date: Wed Jul 29 17:18:44 2020 -0600
PCI/P2PDMA: Allow P2PDMA on AMD Zen and newer CPUs
Allow P2PDMA if the CPU vendor is AMD and family is 0x17 (Zen) or greater.
[bhelgaas: commit log, simplify #if/#else/#endif]
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Logan Gunthorpe <[email protected]>
Signed-off-by: Bjorn Helgaas <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Cc: Christian König <[email protected]>
Cc: Huang Rui <[email protected]>
Alex
>
> Regards,
> Felix
>
>
> >
> > Alex
> >
> >> Signed-off-by: Ramesh Errabolu <[email protected]>
> >> ---
> >> drivers/pci/p2pdma.c | 2 ++
> >> 1 file changed, 2 insertions(+)
> >>
> >> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
> >> index 196382630363..7003bb9faf23 100644
> >> --- a/drivers/pci/p2pdma.c
> >> +++ b/drivers/pci/p2pdma.c
> >> @@ -305,6 +305,8 @@ static const struct pci_p2pdma_whitelist_entry {
> >> {PCI_VENDOR_ID_INTEL, 0x2032, 0},
> >> {PCI_VENDOR_ID_INTEL, 0x2033, 0},
> >> {PCI_VENDOR_ID_INTEL, 0x2020, 0},
> >> + /* AMD Host Bridge Devices */
> >> + {PCI_VENDOR_ID_AMD, 0x1480, 0},
> >> {}
> >> };
> >>
> >> --
> >> 2.31.1
> >>
I will investigate it further. I am using DKMS-5.11 branch. The codebase I am using to build has the right definition i.e. allow P2PDMA for Zen CPU's.
Regards,
Ramesh
-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Wednesday, August 11, 2021 4:07 PM
To: Kuehling, Felix <[email protected]>
Cc: Errabolu, Ramesh <[email protected]>; amd-gfx list <[email protected]>; Maling list - DRI developers <[email protected]>; Bjorn Helgaas <[email protected]>; Linux PCI <[email protected]>; LKML <[email protected]>
Subject: Re: [PATCH] Whitelist AMD host bridge device(s) to enable P2P DMA
[CAUTION: External Email]
On Wed, Aug 11, 2021 at 4:50 PM Felix Kuehling <[email protected]> wrote:
>
>
> Am 2021-08-11 um 3:29 p.m. schrieb Alex Deucher:
> > On Wed, Aug 11, 2021 at 3:11 PM Ramesh Errabolu <[email protected]> wrote:
> >> Current implementation will disallow P2P DMA if the participating
> >> devices belong to different root complexes. Implementation allows
> >> this default behavior to be overridden for whitelisted devices. The
> >> patch adds an AMD host bridge to be whitelisted
> > Why do we need this? cpu_supports_p2pdma() should return true for
> > all AMD Zen CPUs.
>
> This is part of our on-going work to get P2P support upstream. We want
> to use pci_p2pdma_distance_many to determine whether P2P is possible
> between a pair of devices. This whitelist is used in this function.
> This will affect the P2P links reported in the topology and it will be
> double-checked in the BO mapping function to ensure a peer mapping is legal.
>
> I think this change is a bit free of context at the moment, as we're
> still working on a few other loose ends for P2P support in our
> internal branch. I'm hoping we'll have a bigger patch series for
> upstreamable KFD P2P support ready in a few weeks. I also think we'll
> probably want to add a few more PCI IDs for other supported AMD root complexes.
We don't need to keep adding AMD root complexes. You must be using an older kernel or something. All root complexes on all Zen platforms support P2P DMA. See:
commit dea286bb71baded7d2fb4f090e3b9fd2c1ccac58
Author: Logan Gunthorpe <[email protected]>
Date: Wed Jul 29 17:18:44 2020 -0600
PCI/P2PDMA: Allow P2PDMA on AMD Zen and newer CPUs
Allow P2PDMA if the CPU vendor is AMD and family is 0x17 (Zen) or greater.
[bhelgaas: commit log, simplify #if/#else/#endif]
Link: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fr%2F20200729231844.4653-1-logang%40deltatee.com&data=04%7C01%7CRamesh.Errabolu%40amd.com%7C7ffc96d473d4471889f808d95d0bf57b%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637643128220735445%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=vmwJP2VwKPrXwl3ngUfwN%2BJu3JWeMP0ZZMDh29evW%2Bg%3D&reserved=0
Signed-off-by: Logan Gunthorpe <[email protected]>
Signed-off-by: Bjorn Helgaas <[email protected]>
Reviewed-by: Alex Deucher <[email protected]>
Cc: Christian K?nig <[email protected]>
Cc: Huang Rui <[email protected]>
Alex
>
> Regards,
> Felix
>
>
> >
> > Alex
> >
> >> Signed-off-by: Ramesh Errabolu <[email protected]>
> >> ---
> >> drivers/pci/p2pdma.c | 2 ++
> >> 1 file changed, 2 insertions(+)
> >>
> >> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c index
> >> 196382630363..7003bb9faf23 100644
> >> --- a/drivers/pci/p2pdma.c
> >> +++ b/drivers/pci/p2pdma.c
> >> @@ -305,6 +305,8 @@ static const struct pci_p2pdma_whitelist_entry {
> >> {PCI_VENDOR_ID_INTEL, 0x2032, 0},
> >> {PCI_VENDOR_ID_INTEL, 0x2033, 0},
> >> {PCI_VENDOR_ID_INTEL, 0x2020, 0},
> >> + /* AMD Host Bridge Devices */
> >> + {PCI_VENDOR_ID_AMD, 0x1480, 0},
> >> {}
> >> };
> >>
> >> --
> >> 2.31.1
> >>