Re: [PATCH 1/1] vfio/pci: add PCIe TPH device ioctl

[fengchengwen <fengchengwen@huawei.com>]

We also hope to support this TPH feature in the Kunpeng (ARM server) user-space, and hope that this patch can continue.

@Wathsala, I don't have the previous email, could you please send v2 again? I reviewed v2 [1] and found a few issues that I'd like to discuss.

[1] https://lkml.org/lkml/2025/6/2/1116

Thanks

On 1/29/2026 10:06 PM, Wathsala Vithanage wrote:
> Hi Alex,
> 
> Just checking back on the VFIO PCI TPH patch below. You’d mentioned wanting more time to evaluate the implications, so I wanted to see if you had any remaining concerns or if you’d like me to rework this in a different direction.
> 
> Thanks,
> Wathsala
> 
> On 11/6/25 17:19, Wathsala Vithanage wrote:
>>
>> On 11/5/25 13:15, Alex Williamson wrote:
>>> On Mon, 27 Oct 2025 09:33:33 -0500
>>> Wathsala Vithanage <wathsala.vithanage@arm.com> wrote:
>>>
>>>> On 10/16/25 16:41, Jeremy Linton wrote:
>>>>> Hi,
>>>>>
>>>>> On 10/13/25 11:35 AM, Wathsala Vithanage wrote:
>>>>>> TLP Processing Hints (TPH) let a requester provide steering hints that
>>>>>> can enable direct cache injection on supported platforms and PCIe
>>>>>> devices. The PCIe core already exposes TPH handling to kernel drivers.
>>>>>>
>>>>>> This change adds the VFIO_DEVICE_PCI_TPH ioctl and exposes TPH control
>>>>>> to user space to reduce memory latency and improve throughput for
>>>>>> polling drivers (e.g., DPDK poll-mode drivers). Through this interface,
>>>>>> user-space drivers can:
>>>>>>     - enable or disable TPH for the device function
>>>>>>     - program steering tags in device-specific mode
>>>>>>
>>>>>> The ioctl is available only when the device advertises the TPH
>>>>>> Capability. Invalid modes or tags are rejected. No functional change
>>>>>> occurs unless the ioctl is used.
>>>>>>
>>>>>> Signed-off-by: Wathsala Vithanage <wathsala.vithanage@arm.com>
>>>>>> ---
>>>>>>    drivers/vfio/pci/vfio_pci_core.c | 74 ++++++++++++++++++++++++++++++++
>>>>>>    include/uapi/linux/vfio.h        | 36 ++++++++++++++++
>>>>>>    2 files changed, 110 insertions(+)
>>>>>>
>>>>>> diff --git a/drivers/vfio/pci/vfio_pci_core.c
>>>>>> b/drivers/vfio/pci/vfio_pci_core.c
>>>>>> index 7dcf5439dedc..0646d9a483fb 100644
>>>>>> --- a/drivers/vfio/pci/vfio_pci_core.c
>>>>>> +++ b/drivers/vfio/pci/vfio_pci_core.c
>>>>>> @@ -28,6 +28,7 @@
>>>>>>    #include <linux/nospec.h>
>>>>>>    #include <linux/sched/mm.h>
>>>>>>    #include <linux/iommufd.h>
>>>>>> +#include <linux/pci-tph.h>
>>>>>>    #if IS_ENABLED(CONFIG_EEH)
>>>>>>    #include <asm/eeh.h>
>>>>>>    #endif
>>>>>> @@ -1443,6 +1444,77 @@ static int vfio_pci_ioctl_ioeventfd(struct
>>>>>> vfio_pci_core_device *vdev,
>>>>>>                      ioeventfd.fd);
>>>>>>    }
>>>>>>    +static int vfio_pci_tph_set_st(struct vfio_pci_core_device *vdev,
>>>>>> +                   const struct vfio_pci_tph_entry *ent)
>>>>>> +{
>>>>>> +    int ret, mem_type;
>>>>>> +    u16 st;
>>>>>> +    u32 cpu_id = ent->cpu_id;
>>>>>> +
>>>>>> +    if (cpu_id >= nr_cpu_ids || !cpu_present(cpu_id))
>>>>>> +        return -EINVAL;
>>>>>> +
>>>>>> +    if (!cpumask_test_cpu(cpu_id, current->cpus_ptr))
>>>>>> +        return -EINVAL;
>>>>>> +
>>>>>> +    switch (ent->mem_type) {
>>>>>> +    case VFIO_TPH_MEM_TYPE_VMEM:
>>>>>> +        mem_type = TPH_MEM_TYPE_VM;
>>>>>> +        break;
>>>>>> +    case VFIO_TPH_MEM_TYPE_PMEM:
>>>>>> +        mem_type = TPH_MEM_TYPE_PM;
>>>>>> +        break;
>>>>>> +    default:
>>>>>> +        return -EINVAL;
>>>>>> +    }
>>>>>> +    ret = pcie_tph_get_cpu_st(vdev->pdev, mem_type,
>>>>>> topology_core_id(cpu_id),
>>>>>> +                  &st);
>>>>>> +    if (ret)
>>>>>> +        return ret;
>>>>>> +    /*
>>>>>> +     * PCI core enforces table bounds and disables TPH on error.
>>>>>> +     */
>>>>>> +    return pcie_tph_set_st_entry(vdev->pdev, ent->index, st);
>>>>>> +}
>>>>>> +
>>>>>> +static int vfio_pci_tph_enable(struct vfio_pci_core_device *vdev,
>>>>>> int mode)
>>>>>> +{
>>>>>> +    /* IV mode is not supported. */
>>>>>> +    if (mode == PCI_TPH_ST_IV_MODE)
>>>>>> +        return -EINVAL;
>>>>>> +    /* PCI core validates 'mode' and returns -EINVAL on bad values. */
>>>>>> +    return pcie_enable_tph(vdev->pdev, mode);
>>>>>> +}
>>>>>> +
>>>>>> +static int vfio_pci_tph_disable(struct vfio_pci_core_device *vdev)
>>>>>> +{
>>>>>> +    pcie_disable_tph(vdev->pdev);
>>>>>> +    return 0;
>>>>>> +}
>>>>>> +
>>>>>> +static int vfio_pci_ioctl_tph(struct vfio_pci_core_device *vdev,
>>>>>> +                  void __user *uarg)
>>>>>> +{
>>>>>> +    struct vfio_pci_tph tph;
>>>>>> +
>>>>>> +    if (copy_from_user(&tph, uarg, sizeof(struct vfio_pci_tph)))
>>>>>> +        return -EFAULT;
>>>>>> +
>>>>>> +    if (tph.argsz != sizeof(struct vfio_pci_tph))
>>>>>> +        return -EINVAL;
>>>>>> +
>>>>>> +    switch (tph.op) {
>>>>>> +    case VFIO_DEVICE_TPH_ENABLE:
>>>>>> +        return vfio_pci_tph_enable(vdev, tph.mode);
>>>>>> +    case VFIO_DEVICE_TPH_DISABLE:
>>>>>> +        return vfio_pci_tph_disable(vdev);
>>>>>> +    case VFIO_DEVICE_TPH_SET_ST:
>>>>>> +        return vfio_pci_tph_set_st(vdev, &tph.ent);
>>>>>> +    default:
>>>>>> +        return -EINVAL;
>>>>>> +    }
>>>>>> +}
>>>>>> +
>>>>>>    long vfio_pci_core_ioctl(struct vfio_device *core_vdev, unsigned
>>>>>> int cmd,
>>>>>>                 unsigned long arg)
>>>>>>    {
>>>>>> @@ -1467,6 +1539,8 @@ long vfio_pci_core_ioctl(struct vfio_device
>>>>>> *core_vdev, unsigned int cmd,
>>>>>>            return vfio_pci_ioctl_reset(vdev, uarg);
>>>>>>        case VFIO_DEVICE_SET_IRQS:
>>>>>>            return vfio_pci_ioctl_set_irqs(vdev, uarg);
>>>>>> +    case VFIO_DEVICE_PCI_TPH:
>>>>>> +        return vfio_pci_ioctl_tph(vdev, uarg);
>>>>>>        default:
>>>>>>            return -ENOTTY;
>>>>>>        }
>>>>>> diff --git a/include/uapi/linux/vfio.h b/include/uapi/linux/vfio.h
>>>>>> index 75100bf009ba..cfdee851031e 100644
>>>>>> --- a/include/uapi/linux/vfio.h
>>>>>> +++ b/include/uapi/linux/vfio.h
>>>>>> @@ -873,6 +873,42 @@ struct vfio_device_ioeventfd {
>>>>>>      #define VFIO_DEVICE_IOEVENTFD        _IO(VFIO_TYPE, VFIO_BASE + 16)
>>>>>>    +/**
>>>>>> + * VFIO_DEVICE_PCI_TPH - _IO(VFIO_TYPE, VFIO_BASE + 22)
>>>>>> + *
>>>>>> + * Control PCIe TLP Processing Hints (TPH) on a PCIe device.
>>>>>> + *
>>>>>> + * Supported operations:
>>>>>> + * - VFIO_DEVICE_TPH_ENABLE: enable TPH in no-steering-tag (NS) or
>>>>>> + *   device-specific (DS) mode. IV mode is not supported via this ioctl
>>>>>> + *   and returns -EINVAL.
>>>>>> + * - VFIO_DEVICE_TPH_DISABLE: disable TPH on the device.
>>>>>> + * - VFIO_DEVICE_TPH_SET_ST: program an entry in the device TPH
>>>>>> Steering-Tag
>>>>>> + *   (ST) table. The kernel derives the ST from cpu_id and mem_type;
>>>>>> the
>>>>>> + *   value is not returned to userspace.
>>>>>> + */
>>>>>> +struct vfio_pci_tph_entry {
>>>>>> +    __u32 cpu_id;            /* CPU logical ID */
>>>>>> +    __u8  mem_type;
>>>>>> +#define VFIO_TPH_MEM_TYPE_VMEM        0   /* Request volatile memory
>>>>>> ST */
>>>>>> +#define VFIO_TPH_MEM_TYPE_PMEM        1   /* Request persistent
>>>>>> memory ST */
>>>>>> +    __u8  rsvd[1];
>>>>>> +    __u16 index;            /* ST-table index */
>>>>>> +};
>>>>>> +
>>>>>> +struct vfio_pci_tph {
>>>>>> +    __u32 argsz;            /* Size of vfio_pci_tph */
>>>>>> +    __u32 mode;            /* NS and DS modes; IV not supported */
>>>>>> +    __u32 op;
>>>>>> +#define VFIO_DEVICE_TPH_ENABLE        0
>>>>>> +#define VFIO_DEVICE_TPH_DISABLE        1
>>>>>> +#define VFIO_DEVICE_TPH_SET_ST        2
>>>>>> +    struct vfio_pci_tph_entry ent;
>>>>>> +};
>>>>>> +
>>>>>> +#define VFIO_DEVICE_PCI_TPH    _IO(VFIO_TYPE, VFIO_BASE + 22)
>>>>> A quick look at this, it seems its following the way the existing vfio
>>>>> IOCTls are defined, yet two of them (ENABLE and DISABLE) won't likely
>>>>> really change their structure, or don't need a structure in the case
>>>>> of disable. Why not use IOW() and let the kernel error handling deal
>>>>> with those two as independent ioctls?
>>>>>
>>>>>
>>>>> Thanks,
>>>>
>>>> It will require two IOCTLs. I’m ok with having two IOCTLs for this
>>>> feature if the maintainers are fine with it.
>>> TBH, I'm not sure why we didn't use a DEVICE_FEATURE for this. Seems
>>> like we could implement a SET operation that does enable/disable and
>>
>> Thanks Alex, it was implemented as a DEVICE_FEATURE in RFC v1,
>> except it had a GET operation to get the tag to the user; which we
>> decided to drop.
>>
>>> another for steering tags.  I still need to fully grasp the
>>> implications of this support though.  Thanks,
>> This is now same as the already merged RDMA TPH feature.
>> https://lore.kernel.org/linux-rdma/cover.1751907231.git.leon@kernel.org/
>>
>> --wathsala
>>
>>
>>
> 
>