[Intel-xe] [RFC PATCH 0/2] Implement vma madvise ioctl

[Thomas Hellström]

On 5/24/23 22:12, Nirmoy Das wrote:
>
> On 5/24/2023 8:30 PM, Matthew Brost wrote:
>> On Wed, May 24, 2023 at 02:36:46PM +0200, Nirmoy Das wrote:
>>> Sending this initial RFC patch series for vma madvise ioctl
>>> to gether feedback if this the correct way to do that.
>>>
>>> I am adding two expected options for userspace to pass
>>>
>>> DRM_XE_VMA_MADVISE_WILLNEED:
>>>     * Set ttm priority to normal/high(if cap permits)
>>>     * Make sure VMAs are in allowed placement and bound.
>>>
>>> DRM_XE_VMA_MADVISE_DONTNEED:
>>>     * Set ttm priority to low so the BO belong to the vma
>>>     become early target for eviction.
>>>     * Make sure VMAs are not bound.
>>>
>>> Questions:
>>> Should this be part of DRM_IOCTL_XE_VM_MADVISE rather than creating 
>>> new ioctl?
>>>
>> Def not a new IOCTL. Let's take a step back, what are you trying to
>> implement that the current DRM_IOCTL_XE_VM_MADVISE IOCTL / VM bind IOCTL
>> does not support?
>
> AFAIU at this moment:
>
> MADVISE_WILLNEED  ==  XE_VM_BIND_OP_PREFETCH + DRM_XE_VM_MADVISE_PRIORITY
> MADVISE_DONTNEED  ==  XE_VM_BIND_OP_UNMAP + DRM_XE_VM_MADVISE_PRIORITY
>
> So unless we need explicit madvise ioctl or vm_madvise ioctl options 
> for UMD,
> I think we can have madvise equivalent with above vm bind and vm 
> madvise ioctl.
>
> Hi Thomas, Joonas,
>
> What do you think ?

Hi, Nirmoy,

the functionality we need for DONTNEED and WILLNEED would, based on i915 
and the move to vma-based would IMO be something along the lines of:

DONTNEED
1) if userptr, unbind (or perhaps -EINVAL)
2) If bo, mark the vma as dontneed. If all other vmas of the bo are 
marked dontneed, mark the bo as dontneed and adjust its priority.
3) If a dontneed bo is marked for eviction, unbind its vmas, kill its 
storage and mark it as purged. Don't put on rebind list.

WILLNEED
1) If userptr, bind (or perhaps -EINVAL)
2) If bo, remove dontneed marker. If bo was purged, notify user-space. 
(Need feedback from UMD whether they want an error message or just a 
fresh backing store). Adjust its priority, put vma on rebind list.

  So in particular, if no eviction / shrinking happens between DONTNEED 
and WILLNEED they will essentially be NOOPs.

It sounds like this is a bit different than what can be achieved with 
prefetch/unmap/priority.

/Thomas

>
>
> Regards,
>
> Nirmoy
>
>>
>> Matt
>>
>>> Cc: Thomas Hellström <thomas.hellstrom at linux.intel.com>
>>> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
>>> Cc: Matthew Brost <matthew.brost at intel.com>
>>>
>>> Nirmoy Das (2):
>>>    drm/xe: Expose vma bind-unbind functions
>>>    drm/xe: Implement madvise ioctl for vma
>>>
>>>   drivers/gpu/drm/xe/Makefile         |   1 +
>>>   drivers/gpu/drm/xe/xe_device.c      |   2 +
>>>   drivers/gpu/drm/xe/xe_vm.c          |  52 +++----
>>>   drivers/gpu/drm/xe/xe_vm.h          |   3 +
>>>   drivers/gpu/drm/xe/xe_vma_madvise.c | 223 
>>> ++++++++++++++++++++++++++++
>>>   drivers/gpu/drm/xe/xe_vma_madvise.h |  15 ++
>>>   include/uapi/drm/xe_drm.h           |  28 ++++
>>>   7 files changed, 296 insertions(+), 28 deletions(-)
>>>   create mode 100644 drivers/gpu/drm/xe/xe_vma_madvise.c
>>>   create mode 100644 drivers/gpu/drm/xe/xe_vma_madvise.h
>>>
>>> -- 
>>> 2.39.0
>>>