On Tue, Oct 20, 2026, at 19:37, Bjorn Helgaas wrote:
> On Tue, Oct 20, 2026 at 07:14:47PM +0300, Leon Romanovsky wrote:
>> On Tue, Oct 20, 2026 at 08:13:05AM -0700, Pavel Popov wrote:
>> > Peer-to-Peer DMA (P2PDMA) between two devices requires either a shared
>> > upstream bridge or a host bridge matching pci_p2pdma_whitelist[]. In
>> > VMware guests, passthrough devices appear directly on virtual root
>> > buses without host bridge devices. In Hyper-V guests, each passthrough
>> > device appears in a PCI domain of its own without a host bridge
>> > device. Consequently, P2PDMA checks evaluate endpoint Device IDs
>> > against the host bridge whitelist, causing P2PDMA requests to fail
>> > regardless of host support.
>> >
>> > This affects discrete devices such as GPUs, NICs and NVMe drives.
>> >
>> > Add an explicit list of the Intel Xeon server generations from
>> > Skylake-SP to Diamond Rapids to cpu_supports_p2pdma(). These are the
>> > CPUs whose host bridges are in pci_p2pdma_whitelist[] without
>> > restrictions. AMD Zen and newer are already covered by
>> > cpu_supports_p2pdma(), so only Intel entries are needed.
>> >
>> > Tested with the Intel Level Zero P2P benchmark (ze_peer) on Arc Pro B60
>> > cards in a Linux guest on VMware ESXi 9.0 (Emerald Rapids host).
>> >
>> > Signed-off-by: Pavel Popov <[email protected]>
>> > ---
>> > Changes in v3:
>> > - Drop the hypervisor check and move the Xeon list into
>> > cpu_supports_p2pdma(), next to the AMD check, as suggested by Logan
>> > Gunthorpe and Leon Romanovsky [1]
>> > - Fold the VMware and Hyper-V patches into one, as the check no longer
>> > depends on the hypervisor
>> >
>> > Changes in v2:
>> > - Replace the Skylake-and-newer threshold with an explicit list of Intel
>> > Xeon server generations, from Skylake-SP to Diamond Rapids (Logan
>> > Gunthorpe)
>> > - Move the CONFIG_X86 #ifdef out of hypervisor_supports_p2pdma()
>> > (Logan Gunthorpe)
>> > - Rebased on current mainline
>> >
>> > [1] https://lore.kernel.org/r/20261016153147.GC219304@unreal
>> >
>> > v2:
>> > https://lore.kernel.org/r/[email protected]
>> > v1:
>> > https://lore.kernel.org/r/[email protected]
>> >
>> > drivers/pci/p2pdma.c | 26 ++++++++++++++++++++++++++
>> > 1 file changed, 26 insertions(+)
>> >
>> > diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
>> > index 9334eb314663..2650d605d87d 100644
>> > --- a/drivers/pci/p2pdma.c
>> > +++ b/drivers/pci/p2pdma.c
>> > @@ -21,6 +21,11 @@
>> > #include <linux/seq_buf.h>
>> > #include <linux/xarray.h>
>> >
>> > +#ifdef CONFIG_X86
>> > +#include <asm/cpu_device_id.h>
>> > +#include <asm/intel-family.h>
>> > +#endif
>> > +
>> > struct pci_p2pdma {
>> > struct gen_pool *pool;
>> > bool p2pmem_published;
>> > @@ -519,6 +524,23 @@ static void seq_buf_print_bus_devfn(struct seq_buf
>> > *buf, struct pci_dev *pdev)
>> > seq_buf_printf(buf, "%s;", pci_name(pdev));
>> > }
>> >
>> > +#ifdef CONFIG_X86
>>
>> Do we really need all these #ifdefs? I'm pretty sure they're not
>> needed for the includes.
>
> Apparently we do. I experimentally removed them and cross-compiled
> for arm64:
I'm pretty sure that it's not how asm headers are expected to be consumed.
There are general asm stubs for code outside of arch. Probably general code
shouldn't include ifdefs to access these includes.
Thanks
>
> drivers/pci/p2pdma.c:24:10: fatal error: asm/cpu_device_id.h: No such
> file or directory
> 24 | #include <asm/cpu_device_id.h>
> | ^~~~~~~~~~~~~~~~~~~~~
